Mathematical Analysis and Its Applications 2015

Springer Proceedings in Mathematics & Statistics
P.N. Agrawal
R.N. Mohapatra
Uaday Singh
H.M. Srivastava Editors
Mathematical
Analysis and its
Applications
Roorkee, India, December 2014
Volume 143
This book series features volumes composed of selected contributions from

workshops and conferences in all areas of current research in mathematics and
statistics, including operation research and optimization. In addition to an overall
evaluation of the interest, scientific quality, and timeliness of each proposal at the
hands of the publisher, individual contributions are all refereed to the high quality
standards of leading journals in the field. Thus, this series provides the research
community with well-edited, authoritative reports on developments in the most
exciting areas of mathematical and statistical research today.
More information about this series at http://www.springer.com/series/10533

P.N. Agrawal R.N. Mohapatra
•
Uaday Singh H.M. Srivastava

•
Editors
Mathematical Analysis
and its Applications
Roorkee, India, December 2014
123
Editors
P.N. Agrawal Uaday Singh
Department of Mathematics Department of Mathematics
Indian Institute of Technology Roorkee Indian Institute of Technology Roorkee
Roorkee, Uttarakhand Roorkee, Uttarakhand
India India
R.N. Mohapatra H.M. Srivastava

Department of Mathematics Department of Mathematics and Statistics
University of Central Florida University of Victoria
Orlando, FL Victoria, BC
USA Canada
ISSN 2194-1009 ISSN 2194-1017 (electronic)

ISBN 978-81-322-2484-6 ISBN 978-81-322-2485-3 (eBook)
DOI 10.1007/978-81-322-2485-3
Library of Congress Control Number: 2015942480
Mathematics Subject Classification: 34H05, 41A25, 41A36, 65M12, 68T10, 70J25, 76M10
Springer New Delhi Heidelberg New York Dordrecht London

© Springer India 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.
Printed on acid-free paper
Springer (India) Pvt. Ltd. is part of Springer Science+Business Media

(www.springer.com)
Preface
Welcome to the Proceedings of the International Conference on Recent Trends in

Mathematical Analysis and its Applications 2014 (ICRTMAA-2014). With a view
to bringing together experts and young researchers working on various areas of
Mathematical Analysis and its Applications, the Department of Mathematics at the
Indian Institute of Technology, Roorkee, decided to organize this international
conference from December 21 to 23, 2014. We invited many mathematicians of
repute to deliver invited lectures on topics of their interest and also young
researchers to present their research during the conference. The conference pro-
vided an opportunity to exchange ideas and share cutting-edge research.
The conference had invited speakers and participants from United States of
America, Germany, South Africa, the Sultanate of Oman, and reputed institutions
from India. When we called for papers, we received 94 papers. Each of these was
sent to two carefully chosen referees, who went through the papers and gave their
recommendations on whether the paper should be included in the proceedings
of the conference. Based on the recommendations, we selected 60 papers for
inclusion in the proceedings.
Mathematical Analysis is an interesting subject which has applications in many
different areas of pure and applied mathematics. The papers included in this volume
demonstrate the versatility and inherent beauty of Analysis. The areas included are
Compressive sensing, Approximation Theory, Solitons and nonlinear waves,
Galerkin methods, singularly perturbed differential and difference equations, fractal
interpolations and surface fitting, equilibrium problems, optimization using various
techniques, different methods to approach pricing European options, Fractional,
functional and other types of differential equations, stochastic integro-differential
equations, mathematical models from Biology, wavelet frames, Frames in
semi-inner product spaces, and Frames in Hilbert C* module.
The main theme in each of these papers is to employ analytical techniques to
solve the problems at hand and wherever possible to find their numerical solutions.
The large collection of papers in Approximation theory shows the use of operators,
q-integers, and splines. The authors of these papers have carefully described the
problem and showed appropriate methods to obtain the solution. For a problem in
v
vi Preface
mathematical biology one may need to validate the model by showing that the
model is a realistic representation of the problem, and its solution may yield
valuable insights. In these proceedings, the papers have been classified into two
parts namely, “Mathematical Analysis” and “Applications.”
The trend in research is constantly changing due to developments of new tools,
and a conference of this type is a valuable means for knowledge transfer. This
volume which contains the research papers and a few carefully selected survey
papers will provide the readers with an opportunity to see how mathematical
analysis can be applied in various contexts to solve problems.
As the Conference Chairs of ICRTMAA 2014, we are thankful to all the funding
agencies for their grant support for the successful completion of this international
conference. The conference was supported partially by the following funding
agencies:
1. Department of Science and Technology (DST), Government of India, New
Delhi
2. Uttarakhand State Council for Science and Technology (UCOST), India
3. Quality Improvement Programme Centre, Indian Institute of Technology
(IIT) Roorkee, India
4. International Society for Analysis, Applications and Computation (ISAAC)
We would like to express our thanks to Prof. Pradipta Banerji, Director IIT
Roorkee, Roorkee India, for his constant encouragement, motivation, and support.
We also extend our profound thanks to all the authorities of IIT Roorkee as well as
the faculty members and research scholars of the Department of Mathematics, IIT
Roorkee, Roorkee.
We are grateful to the Chairs and Members of the Screening Committee,
Registration Committee, Publication Committee, Academic Programme Committee,
Catering Committee, Cultural Committee, Finance Committee, and Advisory
Committee which worked as a team by investing their invaluable time and hard work
to make this event a success.
We extend our hearty thanks to the keynote speakers who kindly accepted our
invitation. Especially, we would like to thank the following experts:
1. Prof. R.A. Zalik, Auburn University, USA
2. Prof. R.N. Mohapatra, University of Central Florida, USA
3. Prof. Zhiseng Shuai, University of Central Florida, USA
4. Prof. Margareta Heilmann, Germany
5. Prof. Elena Berdysheva, Muskat, Oman
6. Prof. M.C. Joshi, IIT Gandhinagar, India
7. Prof. R.K. Mohanty, South Asian University, India
A total of 71 subject experts from around the world contributed to the
peer-review process. We express our sincere gratitude to the reviewers for spending
their valuable time to review the papers and for sorting out the papers for pre-
sentation at the conference.
Preface vii
Our goal will be accomplished if the readers find this volume useful and
informative for their future research. We are thankful to Springer for publishing the
proceedings of the conference.
February 2015 P.N. Agrawal

R.N. Mohapatra
Uaday Singh
H.M. Srivastava
International Program Committee
Prof. R.P. Agarwal, Texas A&M University, USA

Prof. T. Amarnath, Central University of Hyderabad, India
Prof. D. Bahuguna, IIT Kanpur, India
Prof. Ismail Naci Cangul, Uludag University, Turkey
Prof. B.K. Dass, University of Delhi, India
Prof. Octay Duman, TOBB University, Turkey
Prof. Zoltan Finta, Babes-Bolyai University, Romania
Prof. Raju K. George, IIST, Thiruvananthapuram, India
Prof. N.K. Govil, Auburn University, USA
Prof. Vijay Gupta, NSIT New Delhi, India
Prof. Nurhayat Ispir, Gazi University, Turkey
Prof. G.P. Kapoor, IIT Kanpur, India
Prof. Harun Karsli, Abant Izzet Baysal University, Turkey
Prof. Imdad Khan, AMU, India
Prof. N.I. Mahmudov, Eastern Mediterranean University, Turkey
Prof. S.K. Mishra, BHU, Varanasi, India
Prof. R.N. Mohapatra, University of Central Florida, USA
Prof. A.J. Lopez Moreno, Universidad de Jaen, Spain
Prof. M. Mursaleen, AMU, India
Prof. M.T. Nair, IIT Madras, India
Prof. A.K. Nandakumaran, IISc Bangalore, India
Prof. A.K. Pani, IIT Bombay, India
Prof. R.K.S. Rathore, IIT Kanpur, India
Prof. H.M. Srivastava, University of Victoria, Canada
Prof. P.D. Srivastava, IIT Kharagpur, India
Prof. Richard A. Zalik, Auburn University, USA
ix
x International Program Committee
Local Organizing Committee
Patron: Prof. Pradipta Banerji, Director IIT Roorkee

President: Prof. R.C. Mittal, Head, Department of Mathematics, IIT Roorkee
Conveners: Prof. P.N. Agrawal, Department of Mathematics, IIT Roorkee
Dr. Uaday Singh, Department of Mathematics, IIT Roorkee
Members:
Prof. S.P. Sharma, Prof. Rama Bhargava, Prof. V.K. Katiyar, Prof. N. Sukavanam,
Prof. Kusum Deep, Prof. Tanuja Srivastava, Dr. A. Swaminathan, Dr. Sanjeev
Kumar, Dr. D.N. Pandey
Sub-committees:
Academic Program Committee Catering Committee

1. Prof. Tanuja Srivastava 1. Prof. Rama Bhargava
2. Dr. D.N. Pandey 2. Prof. V.K. Katiyar
3. Dr. Ankik Giri 3. Dr. D.N. Pandey
Publication Committee Purchase Committee
1. Prof. R.C. Mittal 1. Prof. Roshan Lal
2. Prof. Rama Bhargava 2. Prof. P.N. Agrawal
3. Prof. N. Sukavanam 3. Prof. Sunita Gakkhar
4. Dr. A. Swaminathan 4. Dr. Sanjeev Kumar
5. Dr. P. Bera
Transport Committee Accommodation Committee
1. Dr. Sandip Banerjee 1. Prof. V.K. Katiyar
2. Dr. Uaday Singh 2. Prof. M.L. Mittal
3. Dr. S.K. Gupta 3. Prof. Kusum Deep
4. Dr. Ram Jiwari 4. Dr. R.K. Pandey
Cultural Program Committee Registration Committee
1. Prof. Kusum Deep 1. Prof. S.P. Sharma
2. Dr. Madhu Jain 2. Prof. S.P. Yadav
3. Dr. Pratibha 3. Dr. Maheshanand
4. Dr. A.K. Nayak 4. Dr. Aditi Gangopadhyay
Finance Committee
1. Prof. S.P. Sharma
2. Prof. P.N. Agrawal
Contents
Part I Mathematical Analysis
Convergence Analysis of Legendre Spectral Galerkin Method

for Volterra-Fredholm-Hammerstein Integral Equations . . . . . . . . . . . 3
Payel Das and Gnaneshwar Nelakanti
Analysis of an Eco-Epidemiological Model with Migrating

and Refuging Prey. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Shashi Kant and Vivek Kumar
Stability Analysis of an Integro Differential Equation Model

of Ring Neural Network with Delay . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Swati Tyagi, Syed Abbas and Rajendra K. Ray
Approximation of Solutions of a Stochastic Differential Equation. . . . . 51

Sanjukta Das, D.N. Pandey and N. Sukavanam
Reconstruction of Multiply Generated Splines from Local

Average Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
P. Devaraj and S. Yugesh
Approximation of Periodic Functions Belonging

to WðLr ; nðtÞ; ðb 0ÞÞ-Class By ðC1 TÞ Means of Fourier Series . . . . . 73
Smita Sonker
Modified Baskakov-Szász Operators Based on q-Integers . . . . . . . . . . 85

P.N. Agrawal and Arun Kajla
Approximation of Functions of Class Lipða; pÞ in Lp -Norm . . . . . . . . . 109

M.L. Mittal and Mradul Veer Singh
xi
xii Contents
A New Genuine Durrmeyer Operator. . . . . . . . . . . . . . . . . . . . . . . . . 121

Vijay Gupta
Degree of Approximation by Certain Genuine Hybrid Operators. . . . . 131

Meenu Goyal and P.N. Agrawal
Frames in Semi-inner Product Spaces. . . . . . . . . . . . . . . . . . . . . . . . . 149

N.K. Sahu and Ram N. Mohapatra
Applications of Generalized Monotonicity to Variational-Like

Inequalities and Equilibrium Problems . . . . . . . . . . . . . . . . . . . . . . . . 159
N.K. Mahato and R.N. Mohapatra
Simultaneous Approximation Properties of q-Modified

Beta Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Asha Ram Gairola, Girish Dobhal and Karunesh Kumar Singh
Nonlinear Mixed Variational-Like Inequality with Respect

to Weakly Relaxed g a Monotone Mapping in Banach Spaces . . . . . 185
Gayatri Pany and Sabyasachi Pani
Pál Type (0; 1)-Interpolation on Mixed Tchebycheff Abscissas-II . . . . . 197

Neha Mathur and Pankaj Mathur
Degree of Approximation of f ‰ L½0; ‘Þ by Means

of Fourier–Laguerre Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Soshal Saini and Uaday Singh
Sparse Approximation of Overdetermined Systems

for Image Retrieval Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
M. Srinivas and R. Ramu Naidu
kth Order Kantorovich Modification of Linking

Baskakov-Type Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Margareta Heilmann and Ioan Raşa
Rate of Convergence of Modified Schurer-Type q-Bernstein

Kantorovich Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
Manjari Sidharth and P.N. Agrawal
Operators of Durrmeyer Type with Respect to Arbitrary Measure . . . 255

Elena E. Berdysheva
Contents xiii
Part II Applications
Construction of Sparse Binary Sensing Matrices Using

Set Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
R. Ramu Naidu
Topological and Nontopological 1-Soliton Solution

of the Generalized KP-MEW Equation . . . . . . . . . . . . . . . . . . . . . . . . 275
Amiya Das and Asish Ganguly
Mild Solutions for Impulsive Functional Differential Equations

of Order a ‰ ð1; 2Þ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Ganga Ram Gautam and Jaydev Dabas
An OðN 1 lnNÞ4 Parameter Uniform Difference Method

for Singularly Perturbed Differential-Difference Equations . . . . . . . . . 299
Komal Bansal and Kapil K. Sharma
Almost Periodicity of a Modified Leslie–Gower Predator–Prey

System with Crowley–Martin Functional Response . . . . . . . . . . . . . . . 309
Jai Prakash Tripathi and Syed Abbas
Existence of a Mild Solution for Impulsive Neutral Fractional

Differential Equations with Nonlocal Conditions . . . . . . . . . . . . . . . . . 319
Alka Chadha and Dwijendra N. Pandey
Numerical Solution of Highly Oscillatory Nonlinear Integrals

Using Quasi-Monte Carlo Methods. . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Nageswara Rao Narni
Approximate Controllability of Semilinear Stochastic System

with State Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Anurag Shukla, N. Sukavanam and D.N. Pandey
Fourth-Order Derivative-Free Optimal Families

of King’s and Ostrowski’s Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 359
Ramandeep Behl, S.S. Motsa, Munish Kansal and V. Kanwar
Existence of Solution for Fractional Stochastic

Integro-Differential Equation with Impulsive Effect. . . . . . . . . . . . . . . 373
Mohd Nadeem and Jaydev Dabas
xiv Contents
Singularly Perturbed Convection-Diffusion Turning Point

Problem with Shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
Pratima Rai and Kapil K. Sharma
High-Order Compact Finite Difference Method

for Black–Scholes PDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
Kuldip Singh Patel and Mani Mehra
On Stability of Steady-States for a Two-Dimensional

Network Model of Ferromagnetic Nanowires . . . . . . . . . . . . . . . . . . . 405
Sharad Dwivedi and Shruti Dubey
Fractional Functional Impulsive Differential Equation

with Integral Boundary Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
Vidushi Gupta and Jaydev Dabas
Controllability of Nonlinear Fractional Neutral Stochastic

Dynamical Systems with Poisson Jumps . . . . . . . . . . . . . . . . . . . . . . . 429
T. Sathiyaraj and P. Balasubramaniam
Efficient Meshfree Method for Pricing European

and American Put Options on a Non-dividend Paying Asset . . . . . . . . 439
Kailash C. Patidar and Abdelmgid O.M. Sidahmed
A Laplace Transform Approach for Pricing European Options . . . . . . 451

Edgard Ngounda and Kailash C. Patidar
Quintic Hermite Fractal Interpolation in a Strip: Preserving

Copositivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
A.K.B. Chand and S.K. Katiyar
Existence Result for Semilinear Fractional Stochastic

Evolution Inclusions Driven by Poisson Jumps . . . . . . . . . . . . . . . . . . 477
P. Tamilalagan and P. Balasubramaniam
A Numerical Investigation of Blood Flow in an Arterial

Segment with Periodic Body Acceleration . . . . . . . . . . . . . . . . . . . . . . 489
Mamata Parida and Ameeya K. Nayak
Improving R-Order Convergence of Derivative Free

with Memory Method by Two Self-accelerator Parameters . . . . . . . . . 501
Anuradha Singh and J.P. Jaiswal
Contents xv
Numerical Solutions of Differential Equations Using Modified

B-spline Differential Quadrature Method . . . . . . . . . . . . . . . . . . . . . . 509
R.C. Mittal and Sumita Dahiya
Complete Controllability of a Delayed Semilinear Stochastic

Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
Urvashi Arora and N. Sukavanam
On Mixed Integro-Differential Inequalities and Applications . . . . . . . . 535

S.D. Kendre and S.G. Latpate
Buckling and Vibration of Functionally Graded Circular

Plates Resting on Elastic Foundation . . . . . . . . . . . . . . . . . . . . . . . . . 545
Neha Ahlawat and Roshan Lal
Discrete Fourier Transform and Extended Modified

Hermite Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
R.A. Malekar
A Numerical Simulation Based on Modified Keller Box Scheme

for Fluid Flow: The Unsteady Viscous Burgers’ Equation . . . . . . . . . . 565
B. Mayur Prakash, Ashish Awasthi and S. Jayaraj
A Novel Approach to Surface Interpolation: Marriage

of Coons Technique and Univariate Fractal Functions. . . . . . . . . . . . . 577
A.K.B. Chand, P. Viswanathan and K.M. Reddy
Constrained 2D Data Interpolation Using Rational

Cubic Fractal Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593
A.K.B. Chand and K.R. Tyada
Transverse Vibrations of Nonhomogeneous Rectangular

Kirchhoff Plates of Variable Thickness . . . . . . . . . . . . . . . . . . . . . . . . 609
Roshan Lal and Renu Saini
Compartmental Disease Models with Heterogeneous

Populations: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 619
R.N. Mohapatra, Donald Porchia and Zhisheng Shuai
Radially Symmetric Vibrations of Exponentially Tapered

Clamped Circular Sandwich Plate Using Harmonic
Differential Quadrature Method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
Rashmi Rani and Roshan Lal
xvi Contents
Hybrid Projective Synchronization of Fractional-Order

Neural Networks with Time Delays . . . . . . . . . . . . . . . . . . . . . . . . . . 645
G. Velmurugan and R. Rakkiyappan
Approximations of Solutions of a Class of Neutral Differential

Equations with a Deviated Argument . . . . . . . . . . . . . . . . . . . . . . . . . 657
Pradeep Kumar, Dwijendra N. Pandey and D. Bahuguna
Approximation of Solutions to Fractional Integro-Differential

Equations with Finite Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
Renu Chaudhary and Dwijendra N. Pandey
An Efficient Hybrid Approach for Simulating MHD Nanofluid

Flow over a Permeable Stretching Sheet . . . . . . . . . . . . . . . . . . . . . . . 701
Rama Bhargava, Mania Goyal and Pratibha
Some Advanced Finite Element Techniques for the Problems

of Mechanics: A Review. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
Rama Bhargava and Rangoli Goyal
Duality for a Nondifferentiable Multiobjective Second-Order

Fractional Programming Problem Involving
ðF; a; q; dÞ—V-type-I Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 729
Ramu Dubey and S.K. Gupta
Consistency and Unconditional Stability of a Positive Upwind

Scheme for One-Dimensional Species Transport Equation . . . . . . . . . . 743
S. Prabhakaran and L. Jones Tarcius Doss
Numerical Solution for a Coupled System of Singularly

Perturbed Initial Value Problems with Discontinuous
Source Term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753
S. Chandra Sekhara Rao and Sheetal Chawla
Editors and Contributors
About the Editors
P.N. Agrawal is Professor at the Department of Mathematics, Indian Institute of

Technology Roorkee, India. He obtained his Ph.D. degree from Indian Institute of
Technology Kanpur, India, in 1980. Professor Agrawal has 110 research papers to
his credit published in several journals of repute as well as guided many post-
graduate and Ph.D. students. He has presented research papers at several interna-
tional conferences in India and abroad and also delivered invited lectures.
Ram N. Mohapatra is professor of Mathematics, University of Central Florida,
Orlando, USA. He received his Ph.D. degree from the University of Jabalpur, India,
in 1968. Earlier, he taught at Sambalpur University in India, American University in
Beirut, Lebanon, University of Alberta, and York University, Canada, prior to
coming to Orlando. His area of research is Mathematical Analysis and he is the
author of two books, two edited monographs, and over 120 research papers to his
credit. He referees articles for professional journals and serves as a member of the
editorial board of a number of journals.
Uaday Singh is assistant professor at the Department of Mathematics, Indian
Institute of Technology Roorkee, India. Earlier, Dr. Singh worked as lecturer at
Banaras Hindu University, Varanasi. His areas of research are fourier approxima-
tion and summability theory. He has more than 20 publications to his credit that
include research papers, monograph and book chapters. He is actively involved in
teaching undergraduate and postgraduate students as well as Ph.D. students. He is a
member of many professional societies such as American Mathematical Society
(AMS), International Society for Augmentative and Alternative Communication
(ISAAC), International Association of Engineers (IAENG), and Indian
Mathematical Society (IMS).
H.M. Srivastava is professor emeritus at the Department of Mathematics and
Statistics, University of Victoria, Canada. He joined the University of Victoria in 1969,
first as associate professor (1969–1974), and then as full professor (1974–2006).
xvii
xviii Editors and Contributors
Professor Srivastava has held numerous visiting positions including (for example) those
at West Virginia University in the USA (1967–1969), Universite Laval in Canada
(1975), and the University of Glasgow in the UK (1975–1976), and indeed also at many
other universities and research institutes in different parts of the world. Professor
Srivastava has published 21 books (monographs and edited volumes), 30 book chapters,
45 papers in international conference proceedings, and over 1000 scientific research
journal articles on various topics of mathematical analysis and applicable mathematics.
In addition, he has written forewords to several books by other authors and to several
special issues of scientific journals. He has also edited (and contributed to) many
volumes, which are dedicated to the memories of famous mathematical scientists.
Citations of his research contributions can be found in many books and monographs,
Ph.D. and D.Sc. theses, and scientific journal articles, much too numerous to be
recorded here. Currently, he is actively associated editorially (that is, as editor, honorary
editor, senior editor, associate editor or editorial board member) with over 200 inter-
national scientific research journals. His biographical sketches (many of which are
illustrated with his photograph) have appeared in various issues of more than 50
international biographies, directories, and Who’s Who’s.
Contributors
Syed Abbas School of Basic Sciences, Indian Institute of Technology Mandi,

Mandi, Himachal Pradesh, India
P.N. Agrawal Department of Mathematics, Indian Institute of Technology
Roorkee, Roorkee, India
Neha Ahlawat Department of Mathematics, Indian Institute of Technology
Urvashi Arora Department of Mathematics, Indian Institute of Technology
Ashish Awasthi Department of Mathematics, National Institute of Technology
Calicut, Kozhikode, India
D. Bahuguna Department of Mathematics and Statistics, Indian Institute of
Technology Kanpur, Kanpur, India
P. Balasubramaniam Gandhigram Rural Institute-Deemed University,
Gandhigram, Tamil Nadu, India
Komal Bansal Department of Mathematics, Panjab University, Chandigarh, India
Ramandeep Behl School of Mathematics, Statistics and Computer Science,
University of KwaZulu-Natal, Pietermaritzburg, South Africa
Elena E. Berdysheva Department of Mathematics and Science, German
University of Technology in Oman, Muscat, Sultanate of Oman
Editors and Contributors xix
Rama Bhargava Department of Mathematics, Indian Institute of Technology

Alka Chadha Department of Mathematics, Indian Institute of Technology
A.K.B. Chand Department of Mathematics, Indian Institute of Technology
Madras, Chennai, India
S. Chandra Sekhara Rao Department of Mathematics, Indian Institute of
Technology Delhi, Hauz Khas, New Delhi, India
Renu Chaudhary Department of Mathematics, Indian Institute of Technology
Roorkee, Roorkee, Uttarakhand, India
Sheetal Chawla Department of Mathematics, Indian Institute of Technology
Delhi, Hauz Khas, New Delhi, India
Jaydev Dabas Department of Applied Science and Engineering, IIT Roorkee,
Saharanpur Campus, Saharanpur, India
Sumita Dahiya Department of Mathematics, IIT Roorkee, Roorkee, Uttarakhand,
India
Amiya Das IIT Kharagpur, Kharagpur, India
Payel Das Mathematics Department, IIT Kharagpur, Kharagpur, India
Sanjukta Das Department of Mathematics, IIT Roorkee, Roorkee, India
P. Devaraj Department of Mathematics, College of Engineering Guindy Campus,
Anna University, Chennai, India
Girish Dobhal University of Petroleum and Energy Studies, Dehradun, India
L. Jones Tarcius Doss College of Engineering Guindy, Anna University,
Chennai, India
Ramu Dubey Department of Mathematics, Indian Institute of Technology
Shruti Dubey Department of Mathematics, Indian Institute of Technology
Sharad Dwivedi Department of Mathematics, Indian Institute of Technology
Asha Ram Gairola University of Petroleum and Energy Studies, Dehradun, India
Asish Ganguly IIT Kharagpur, Kharagpur, India
Ganga Ram Gautam Department of Applied Science and Engineering, IIT
Roorkee, Saharanpur Campus, Saharanpur, India
xx Editors and Contributors
Mania Goyal Department of Mathematics, Indian Institute of Technology

Meenu Goyal Department of Mathematics, Indian Institute of Technology
Rangoli Goyal Department of Mathematics, Indian Institute of Technology
S.K. Gupta Department of Mathematics, Indian Institute of Technology Roorkee,
Roorkee, India
Vidushi Gupta Department of Applied Science and Engineering, IIT Roorkee,
Saharanpur, India
Vijay Gupta Department of Mathematics, Netaji Subhas Institute of Technology,
New Delhi, India
Margareta Heilmann Faculty of Mathematics and Natural Sciences, University of
Wuppertal, Wuppertal, Germany
J.P. Jaiswal Department of Mathematics, Maulana Azad National Institute of
Technology, Bhopal, Madhya Pradesh, India
S. Jayaraj Department of Mechanical Engineering, National Institute of
Technology Calicut, Kozhikode, India
Arun Kajla Department of Mathematics, Indian Institute of Technology Roorkee,
Roorkee, India
Munish Kansal University Institute of Engineering and Technology, Panjab
University, Chandigarh, India
Shashi Kant Department of Applied Mathematics, Delhi Technological
University, Delhi, India
V. Kanwar University Institute of Engineering and Technology, Panjab
University, Chandigarh, India
S.K. Katiyar Department of Mathematics, Indian Institute of Technology Madras,
Chennai, India
S.D. Kendre Department of Mathematics, Savitribai Phule Pune University, Pune
(MS), India
Pradeep Kumar Department of Mathematics and Statistics, Indian Institute of
Technology Kanpur, Kanpur, India
Vivek Kumar Department of Applied Mathematics, Delhi Technological
University, Delhi, India
Roshan Lal Department of Mathematics, Indian Institute of Technology Roorkee,
Roorkee, India
Editors and Contributors xxi
S.G. Latpate Department of Mathematics, Nowrosjee Wadia College of Arts and

Science, Pune, India
N.K. Mahato Indian Institute of Information Technology, Design and
Manufacturing Jabalpur, Jabalpur, India
R.A. Malekar National Defence Academy, Pune, India
Neha Mathur Department of Mathematics, Career Convent College, Lucknow,
India
Pankaj Mathur Department of Mathematics and Astronomy, Lucknow
University, Lucknow, India
Mani Mehra Indian Institute of Technology, Delhi, India
M.L. Mittal Department of Mathematics, Indian Institute of Technology Roorkee,
Roorkee, India
R.C. Mittal Department of Mathematics, IIT Roorkee, Roorkee, Uttarakhand,
India
R.N. Mohapatra Department of Mathematics, University of Central Florida,
Orlando, FL, USA
S.S. Motsa School of Mathematics, Statistics and Computer Science, University of
KwaZulu-Natal, Pietermaritzburg, South Africa
Mohd Nadeem Department of Applied Science and Engineering, IIT Roorkee,
Saharanpur, India
R. Ramu Naidu Department of Mathematics, IIT Hyderabad, Hyderabad, India
Nageswara Rao Narni Rajiv Gandhi University of Knowledge Technologies,
Hyderabad, Andhra Pradesh, India
Ameeya K. Nayak Department of Mathematics, Indian Institute of Technology
Gnaneshwar Nelakanti Mathematics Department, IIT Kharagpur, Kharagpur,
India
Edgard Ngounda Department of Mathematics and Applied Mathematics,
University of the Western Cape, Bellville, South Africa
Dwijendra N. Pandey Department of Mathematics, Indian Institute of
Technology Roorkee, Roorkee, India
Sabyasachi Pani IIT Bhubaneswar, Odisha, India
Gayatri Pany IIT Bhubaneswar, Odisha, India
Mamata Parida Department of Mathematics, Utkal University, Bhubaneswar,
India
xxii Editors and Contributors
Kuldip Singh Patel Indian Institute of Technology Delhi, Delhi, India

Kailash C. Patidar Department of Mathematics and Applied Mathematics,
University of the Western Cape, Bellville, South Africa
Donald Porchia Department of Mathematics, University of Central Florida,
Orlando, FL, USA
S. Prabhakaran College of Engineering Guindy, Anna University, Chennai, India
B. Mayur Prakash Department of Mechanical Engineering, National Institute of
Technology Calicut, Kozhikode, India
Pratibha Department of Mathematics, Indian Institute of Technology Roorkee,
Roorkee, India
Pratima Rai Department of Mathematics, University of Delhi, Delhi, India
R. Rakkiyappan Department of Mathematics, Bharathiar University, Coimbatore,
Tamil Nadu, India
Rashmi Rani Indian Institute of Technology Roorkee, Roorkee, India
Rajendra K. Ray School of Basic Sciences, Indian Institute of Technology,
Mandi, Himachal Pradesh, India
Ioan Raşa Department of Mathematics, Technical University, Cluj-Napoca,
Romania
K.M. Reddy Department of Mathematics, Indian Institute of Technology Madras,
Chennai, India
N.K. Sahu Dhirubhai Ambani Institute of Information and Communication
Technology, Gandhinagar, India
Renu Saini Department of Mathematics, Indian Institute of Technology Roorkee,
Roorkee, India
Soshal Saini Department of Mathematics, Indian Institute of Technology Roorkee,
Roorkee, India
T. Sathiyaraj Gandhigram Rural Institute-Deemed University, Gandhigram,
Tamil Nadu, India
Kapil K. Sharma Department of Mathematics, South Asian University (SAU),
Chanakyapuri, New Delhi, India
Zhisheng Shuai Department of Mathematics, University of Central Florida,
Orlando, FL, USA
Anurag Shukla Department of Mathematics, Indian Institute of Technology
Editors and Contributors xxiii
Abdelmgid O.M. Sidahmed Department of Mathematics and Applied

Mathematics, University of the Western Cape, Bellville, South Africa
Manjari Sidharth Department of Mathematics, Indian Institute of Technology
Anuradha Singh Department of Mathematics, Maulana Azad National Institute of
Technology, Bhopal, Madhya Pradesh, India
Karunesh Kumar Singh Government Polytechnic, Rampur, India
Mradul Veer Singh Department of Mathematics, Indian Institute of Technology
Uaday Singh Department of Mathematics, Indian Institute of Technology
Smita Sonker Department of Mathematics, National Institute of Technology
Kurukshetra, Kurukshetra, India
M. Srinivas Department of Computer Science and Engineering, IIT Hyderabad,
Hyderabad, India
N. Sukavanam Department of Mathematics, Indian Institute of Technology
P. Tamilalagan Department of Mathematics, Gandhigram Rural Institute—
Deemed University, Gandhigram, Tamil Nadu, India
Jai Prakash Tripathi School of Basic Sciences, Indian Institute of Technology
Mandi, Mandi, Himachal Pradesh, India
K.R. Tyada Department of Mathematics, Indian Institute of Technology Madras,
Chennai, India
Swati Tyagi School of Basic Sciences, Indian Institute of Technology, Mandi,
Himachal Pradesh, India
G. Velmurugan Department of Mathematics, Bharathiar University, Coimbatore,
Tamil Nadu, India
P. Viswanathan Department of Mathematics, Indian Institute of Technology
S. Yugesh Department of Mathematics, College of Engineering Guindy Campus,
Anna University, Chennai, India
Part I
Mathematical Analysis
Convergence Analysis of Legendre
Spectral Galerkin Method for
Volterra-Fredholm-Hammerstein
Integral Equations
Payel Das and Gnaneshwar Nelakanti
Abstract In this paper, we analyze the Legendre spectral Galerkin method for a class
of nonlinear Volterra-Fredholm mixed-type integral equations. Existence and con-
vergence of the approximate and iterated approximate solutions to the exact solution
are discussed and the rates of convergence are obtained. We prove that the iterated
approximate solution improves over the approximate solution for Volterra-Fredholm-
Hammerstein integral equations with smooth kernels. Also, we obtain optimal order
of convergence for the iterated Legendre Galerkin method.
Keywords Volterra-Fredholm-Hammerstein integral equation · Spectral method ·

Legendre Galerkin · Convergence rates
1 Introduction
In this section, we consider the following Volterra-Fredholm-Hammerstein integral

equations
t 1
x(t) − k1 (t, s)ψ1 (s, x(s))ds − k2 (t, s)ψ2 (s, x(s))ds = f (t), − 1 ≤ t ≤ 1,
−1 −1
(1)
where k1 , k2 , f , ψ1 and ψ2 are known functions and x is the unknown solution to be
found in a Banach space X. The existence and uniqueness of the solution of Eq. (1)
are discussed in [1, 2].
The Volterra-Fredholm integral equation of type (1) arises from parabolic bound-
ary value problems and in various other physical and biological models (see [3–6]).
The essential features of these models are of wide applicability.
Several numerical methods for approximating the solution of nonlinear integral
equations of type (1) are available in the literature. Classical projection methods
P. Das (B) · G. Nelakanti

Mathematics Department, IIT Kharagpur, Kharagpur, India
e-mail: dasPayel2@gmail.com
G. Nelakanti
e-mail: gnanesh@maths.iitkgp.ernet.in
© Springer India 2015 3
P.N. Agrawal et al. (eds.), Mathematical Analysis and its Applications,
Springer Proceedings in Mathematics & Statistics 143,
DOI 10.1007/978-81-322-2485-3_1
4 P. Das and G. Nelakanti
such as spline-based collocation and Galerkin methods are applied to investigate the
approximate solutions of nonlinear Volterra-Fredholm integral equations of type (1)
(see [2, 7–9]). However, it is necessary to increase the number of partitioning points,
to obtain more accuracy in spline-based projection methods. This leads to solve a
large system of nonlinear equations, which is computationally very expensive. Since
use of global polynomials implies smaller nonlinear system, we will use orthog-
onal projection method using global polynomial basis functions to overcome the
difficulties in terms of computational work encountered in the existing techniques.
In this paper, we apply Galerkin method to the Eq. (1) using Legendre polyno-
mial basis functions, which can be generated recursively with ease and possess nice
property of orthogonality. We prove that the approximated solution of the Legendre
1
Galerkin method converges to the exact solution with the order O(n 2 −r ) in infinity
norm, and the iterated Legendre Galerkin solution converges with the order O(n −r )
in infinity norm, n being the highest degree of Legendre polynomial employed in the
approximation and r being the smoothness of the kernels.
We organize this paper as follows. In Sect. 2, we discuss the Legendre spectral
Galerkin method for the equation of type (1). In Sect. 3, we obtain the existence
and convergence results for the approximate and iterated approximate solutions.
Throughout this paper, we assume that c is a generic constant.
2 Legendre Spectral Galerkin Method:

Volterra-Fredholm-Hammerstein Integral Equations
In this section, we describe the Galerkin method for approximating the solution
of Volterra-Fredholm mixed type Hammerstein integral equations using Legendre
polynomial basis functions.
Let X = C[−1, 1] and consider the following Volterra-Fredholm-Hammerstein
integral equation:
t 1
x(t)− k1 (t, s)ψ1 (s, x(s)) ds − k2 (t, s)ψ2 (s, x(s)) ds = f (t), −1 ≤ t ≤ 1,
−1 −1
(2)
where k1 , k2 , f , ψ1 and ψ2 are known functions and x is the unknown function to be
determined.
Throughout the paper, the following assumptions are made on f , k1 (., .), k2 (., .),
ψ1 (., x(.)) and ψ2 (., x(.)):
(i) f ∈ C[−1, 1].
(ii) lim ki (t, .) − ki (t , .)∞ = 0, t, t ∈ [−1, 1], i = 1, 2.
t→t
Convergence Analysis of Legendre Spectral Galerkin Method … 5
(iii) For i = 1, 2, ki (., .) ∈ C r ([−1, 1] × [−1, 1]), r ≥ 1. Let Mi = ki ∞ =

sup |ki (t, s)| < ∞, i = 1, 2, and
t,s∈[−1,1]
i+j
∂
kl r,∞ = max i j kl (t, s) , l = 1, 2.
0≤i, j≤r ∂t ∂s
t,s∈[−1,1]
(iv) The nonlinear functions ψi (., x(.)) are continuous in s ∈ [−1, 1] and are
Lipschitz continuous in x, i.e., for any x1 , x2 ∈ R, there exist constants
li > 0, i = 1, 2 such that
|ψi (s, x1 ) − ψi (s, x2 )| ≤ li |x1 − x2 |, ∀s ∈ [−1, 1].
(0,1)
(v) The partial derivatives ψi (s, x) of ψi (s, x) with respect to the second variable
(0,1) (0,1)
exist and ψi (., x(.)) ∈ C([−1, 1] × R). The functions ψi (., x(.)) are
Lipschitz continuous in x, i.e., for any x1 , x2 ∈ R, there exist constants ci >
0, i = 1, 2 such that
(0,1) (0,1)
|ψi (s, x1 ) − ψi (s, x2 )| ≤ ci |x1 − x2 |, ∀s ∈ [−1, 1].
For convenience, we define the operators K1 , K2 on X as

t 1
(K1 ψ1 )(x)(t) = k1 (t, s)ψ1 (s, x(s)) ds ; (K2 ψ2 )(x)(t) = k2 (t, s)ψ2 (s, x(s)) ds.
−1 −1
Using K1 and K2 , Eq. (2) can be written as
x − K1 ψ1 (x) − K2 ψ2 (x) = f. (3)
Letting T (x) := f + K1 ψ1 (x) + K2 ψ2 (x), x ∈ X, Eq. (3) takes the form
x = T x. (4)
Using similar technique given in Theorem 2.4 of [10], it can be easily proved that
T has a unique fixed point in X. We assume x0 to be a isolated solution of Eq. (4) in
X. We denote di = sup |ψi(0,1) (s, x0 (s))|.
s∈[−1,1]
Using Leibniz rule, we have
[(K1 ψ1 ) (x0 )x](1) ∞

∂ t
= sup k1 (t, s)ψ1(0,1) (s, x0 (s))x(s)ds
t∈[−1,1] ∂t −1
t
∂
≤ sup |k1 (t, t)||ψ (0,1) (t, x0 (t))||x(t)| + k1 (t, s) ψ1(0,1) (s, x0 (s))x(s)ds
t∈[−1,1] −1 ∂t

(0,1)
≤ sup |k1 (t, t)||ψ1 (t, x0 (t))||x(t)|
t∈[−1,1]

∂ t
+ sup k1 (t, s) sup |ψ (0,1) (s, x0 (s))||x(s)| ds
∂t 1
t,s∈[−1,1] s∈[−1,1] −1
≤ M1 d1 x∞ + 2k1 1,∞ d1 x∞ < ∞. (5)
And for j = 0, 1, 2, ..., r , we have

∂j 1
(0,1)
[(K2 ψ2 ) (x0 )x]( j) ∞ = sup j k2 (t, s)ψ2 (s, x0 (s))x(s)ds
t∈[−1,1] ∂t −1

∂j
(0,1)
≤ 2 sup j k2 (t, s) sup |ψ2 (s, x0 (s))||x(s)|
t,s∈[−1,1] ∂t s∈[−1,1]
≤ 2k2 j,∞ d2 x∞ < ∞. (6)
Next we will apply Legendre Galerkin method to the Eq. (2). To do this, we let Xn =
span{φ0 , φ1 , φ2 , . . ., φn } be the sequence of Legendre polynomial subspaces of X
of degree ≤ n, where {φ0 , φ1 , φ2 , . . ., φn } forms an orthonormal set and φi ’s are
given by
2i + 1
φi (s) = L i (s), i = 0, 1, ..., n, (7)
2
where L i ’s are the Legendre polynomials of degree ≤ i. These Legendre polynomials

can be generated by the following three-term recurrence relation:
L 0 (s) = 1, L 1 (s) = s, s ∈ [−1, 1], (8)
and for i = 1, 2, …, n − 1
(i + 1)L i + 1 (s) = (2i + 1)s L i (s) − i L i−1 (s), s ∈ [−1, 1]. (9)
Orthogonal Projection Operator: Let X = C[−1, 1] and let the operator Pn :

X → Xn be the orthogonal projection defined as

n

Pn x = x, φ j φ j , x ∈ X, (10)
j=0
1
where x, φ j = −1 x(t)φ j (t)dt.
We quote some crucial properties of Pn from Canuto et al. [11] (pp. 283–287).
Lemma 1 Let Pn : X → Xn denote the orthogonal projection defined by (10). Then
the projection Pn satisfies the following properties:
(i) {Pn : n ∈ N} is uniformly bounded in L 2 -norm.
(ii) There exists a constant c > 0 such that for any n ∈ N and u ∈ X,
Pn u − u L 2 ≤ c inf u − φ L 2 → 0, as n → ∞. (11)

φ∈Xn
Hence there exists a constant p1 ≥ 0, independent of n, such that
Pn u L 2 ≤ p1 u∞ , u ∈ X. (12)
Lemma 2 Let Pn be the orthogonal projection defined as (10). Then for any u ∈
C r [−1, 1], there hold
u − Pn u L 2 ≤ cn −r u (r ) L 2 , (13)
3
u − Pn u∞ ≤ cn 4 −r u (r ) L 2 , (14)
1
u − Pn u∞ ≤ cn 2 −r V (u (r )
), (15)
where c is a constant independent of n and V (u (r ) ) denotes the total variation of u (r ) .

The Legendre spectral Galerkin method for Eq. (3) is seeking an approximate
solution xn ∈ Xn , such that
xn − Pn K1 ψ1 (xn ) − Pn K2 ψ2 (xn ) = Pn f, (16)
where Pn : X → Xn is the orthogonal projection operator.

Let Tn be the operator defined as
Tn (u) := Pn f + Pn K1 ψ1 (u) + Pn K2 ψ2 (u), u ∈ X. (17)
Then the Eq. (16) can be written as xn = Tn xn .

To obtain greater accuracy in approximate solution, we further consider the iter-
ated Legendre Galerkin method for (3). To this end, we define the iterated solution
as
x̃n = f + K1 ψ1 (xn ) + K2 ψ2 (xn ). (18)
Since Pn x̃n = xn , it follows that the iterated approximate solution x̃n satisfies
x̃n = f + K1 ψ1 (Pn x̃n ) + K2 ψ2 (Pn x̃n ). (19)
Letting Tn (u) := f + K1 ψ1 (Pn u) + K2 ψ2 (Pn u), u ∈ X, the Eq. (19) can be written
as x̃n = Tn x̃n .
3 Convergence Results
In this section, we analyze the existence and convergence of the approximate solutions
in the Legendre Spectral Galerkin method for the Eq. (2). To do this, we will use the
well-known Vainikko’s Theorem (Theorem 2 in [12]). We first give the following

lemma, which we need for our analysis.
Lemma 3 For any x, y ∈ X, the following hold for i = 1, 2,

√
(Ki ψi )(x0 ) − (Ki ψi )(x)∞ ≤ 2Mi li x0 − x L 2 , (20)
[(Ki ψi ) (x0 ) − (Ki ψi ) (x)]y∞ ≤ Mi ci x0 − x L 2 y L 2 . (21)
Proof Using Lipschitz continuity of ψi (., x(.)), ψi(0,1) (., x(.)) and Cauchy-Schwarz
inequality, the proof of the above lemma follows.
In the following theorem, we give the error bounds for the approximate solution
xn to x0 .
Theorem 1 Let x0 ∈ C r [−1, 1], r ≥ 1, be an isolated solution of the Eq. (3).

Assume that 1 is not an eigenvalue of the linear operator T (x0 ), where T (x0 )
denotes the Frechet derivative of T (x) at x0 . Then the Eq. (16) has a unique solution
xn ∈ B(x0 , δ) = {x : x − x0 ∞ < δ} for some δ > 0 and for sufficiently large n.
Moreover, there exists a constant 0 < q < 1, independent of n such that
αn αn
≤ xn − x0 ∞ ≤ ,
1+q 1−q
where αn = (I − Tn (x0 ))−1 (Tn (x0 ) − T (x0 ))∞ . Further, we obtain
1
xn − x0 ∞ = O n 2 −r .
Proof Using estimates (5), (6) and (14), we have

[Tn (x0 ) − T (x0 )]x∞
= [Pn (K1 ψ1 ) (x0 ) + Pn (K2 ψ2 ) (x0 ) − (K1 ψ1 ) (x0 ) − (K2 ψ2 ) (x0 )]x∞
≤ (Pn − I)((K1 ψ1 ) (x0 )x)∞ + (Pn − I)((K2 ψ2 ) (x0 )x)∞
1 3
≤ cn − 4 [(K1 ψ1 ) (x0 )x](1) ∞ + cn 4 −r [(K2 ψ2 ) (x0 )x](r ) ∞
1 3

≤ c n − 4 (M1 d1 + 2k1 1,∞ d1 ) + 2n 4 −r k2 r,∞ d2 x∞ . (22)
Since r ≥ 1, it follows that
Tn (x0 ) − T (x0 )∞ → 0, as n → ∞.
This shows that Tn (x0 ) is norm convergent to T (x0 ). Since 1 is not an eigen-
value of the linear operator T (x0 ), we have (I − T (x0 ))−1 is invertible on X.
Hence by a result from Ahues et al. [13], we have, for some sufficiently large n,
(I − Tn (x0 ))−1 exists and uniformly bounded on X, i.e., there exists some A1 > 0,
such that (I − Tn (x0 ))−1 ∞ ≤ A1 < ∞.
For any x ∈ B(x0 , δ), we have

Tn (x0 ) − Tn (x)∞
= Pn (K1 ψ1 ) (x0 ) + Pn (K2 ψ2 ) (x0 ) − Pn (K1 ψ1 ) (x) − Pn (K2 ψ2 ) (x)∞
≤ Pn (K1 ψ1 ) (x0 ) − Pn (K1 ψ1 ) (x)∞ + Pn (K2 ψ2 ) (x0 ) − Pn (K2 ψ2 ) (x)∞ .
(23)
(0,1)
Using Lipschitz continuity of ψ1 (., x(.)) and Leibniz rule, we obtain
[((K1 ψ1 ) (x0 ) − (K1 ψ1 ) (x))y](1) ∞
t
∂ (0,1) (0,1)
= sup k (t, s) ψ (s, x (s)) − ψ (s, x(s)) y(s)ds
∂t 1 1 0 1
t∈[−1,1] −1
t
∂
≤ c1 sup |k1 (t, t)||(x0 − x)(t)||y(t)| + k1 (t, s) (x0 − x)(s)y(s)ds
t∈[−1,1] −1 ∂t
≤ c1 sup [|k1 (t, t)||(x0 − x)(t)||y(t)|]
t∈[−1,1]
t
∂
+ c1 sup k1 (t, s) sup [|(x0 − x)(s)|y(s)|] ds
∂t
t,s∈[−1,1] s∈[−1,1] −1
≤ c1 M1 x0 − x∞ y∞ + 2c1 k1 1,∞ x0 − x∞ y∞ . (24)
From estimates (14), (21), (24), for the first term of (23), we have
[Pn (K1 ψ1 ) (x0 ) − Pn (K1 ψ1 ) (x)]y∞

≤ (Pn − I)[(K1 ψ1 ) (x0 ) − (K1 ψ1 ) (x)]y∞ + [(K1 ψ1 ) (x0 ) − (K1 ψ1 ) (x)]y∞
1
≤ cn − 4 [((K1 ψ1 ) (x0 ) − (K1 ψ1 ) (x))y](1) ∞ + [(K1 ψ1 ) (x0 ) − (K1 ψ1 ) (x)]y∞
1
≤ cn − 4 (c1 M1 + 2c1 k1 1,∞ )x0 − x∞ y∞ + 2M1 c1 x − x0 ∞ y∞
1

≤ cn − 4 (c1 M1 + 2c1 k1 1,∞ ) + 2M1 c1 δy∞ . (25)
(0,1)
Using Lipschitz continuity of ψ2 (., x(.)), we get
[((K2 ψ2 ) (x0 ) − (K2 ψ2 ) (x0 ))y](r ) ∞
r 1
∂ (0,1) (0,1)
= sup r k2 (t, s) ψ2 (s, x0 (s)) − ψ2 (s, x(s)) y(s)ds
t∈[−1,1] ∂t −1
r 1
∂

≤ c2 sup r k2 (t, s) sup |(x0 − x)(s)| |y(s)|ds
t,s∈[−1,1] ∂t s∈[−1,1] −1
≤ 2c2 k2 r,∞ x0 − x∞ y∞ ≤ 2c2 k2 r,∞ δy∞ . (26)
Now using estimates (14), (21), (26), for the second term of (23), we obtain
[Pn (K2 ψ2 ) (x0 ) − Pn (K2 ψ2 ) (x)]y∞

≤ (Pn − I)[(K2 ψ2 ) (x0 ) − (K2 ψ2 ) (x)]y∞ + [(K2 ψ2 ) (x0 ) − (K2 ψ2 ) (x)]y∞

3
≤ cn 4 −r [((K2 ψ2 ) (x0 ) − (K2 ψ2 ) (x))y]r ∞ + [(K2 ψ2 ) (x0 ) − (K2 ψ2 ) (x)]y∞
3
≤ [2c2 cn 4 −r k2 r,∞ + 2M2 c2 ]δy∞ . (27)
Combining estimates (23), (25) and (27), we have

1 3

Tn (x0 ) − Tn (x)∞ ≤ cn − 4 (c1 M1 + 2c1 k1 1,∞ ) + 2M1 c1 + 2c2 cn 4 −r k2 r,∞ + 2M2 c2 δ.
Hence, we have
−1
sup (I − Tn (x0 )) (Tn (x0 ) − Tn (x))∞
x−x 0 ∞ ≤δ
1 3

≤ A1 cn − 4 (c1 M1 + 2c1 k1 1,∞ ) + 2M1 c1 + 2c2 cn 4 −r k2 r,∞ + 2M2 c2 δ ≤ q (say).
We choose δ in such a way that q ∈ (0, 1). This proves the estimate (4.4) of Theorem
2 in [12].
Since r ≥ 1, taking use of estimates (3) and (15), we have
αn = (I − Tn (x0 ))−1 (Tn (x0 ) − T (x0 ))∞

≤ A1 Tn (x0 ) − T (x0 )∞
= A1 Pn K1 ψ1 (x0 ) + Pn K2 ψ2 (x0 ) + Pn f − K1 ψ1 (x0 ) − K2 ψ2 (x0 ) − f ∞
= A1 (Pn − I)(K1 ψ1 (x0 ) + K2 ψ2 (x0 ) + f )∞

1 (r )
= A1 (Pn − I)x0 ∞ ≤ A1 cn 2 −r V x0 → 0, as n → ∞. (28)
By choosing n large enough such that αn ≤ δ(1 − q), the Eq. (4.5) of Theorem 2 in
[12] is satisfied. Hence applying Theorem 2 of [12], we obtain
αn αn
≤ xn − x0 ∞ ≤ ,
1+q 1−q
where αn = (I − Tn (x0 ))−1 (Tn (x0 ) − T (x0 ))∞ .

And the estimate (28) implies
1
xn − x0 ∞ = O n 2 −r .
This completes the proof.
Next, we discuss the existence and convergence of the iterated approximate solu-
tion x̃n to x0 .
Theorem 2 Let x0 ∈ C[−1, 1], be an isolated solution of the Eq. (3). Assume that 1
is not an eigenvalue of T (x0 ). Then for sufficiently large n, the operator I − Tn (x0 )
is invertible on X and there exist constants A2 > 0 independent of n such that

(I − Tn (x0 ))−1 ∞ ≤ A2 .
Proof Consider
|Tn (x0 )x(t)| = |[(K1 ψ1 ) (Pn x0 ) + (K2 ψ2 ) (Pn x0 )]Pn x(t)|

≤ |(K1 ψ1 ) (Pn x0 )Pn x(t)| + |(K2 ψ2 ) (Pn x0 )Pn x(t)|. (29)
We have
t
(0,1)
(K1 ψ1 ) (x0 )Pn x∞ = sup
k 1 (t, s)ψ 1 (s, x 0 (s))Pn x(s)ds

t∈[−1,1] −1
1
(0,1)
≤ sup |k1 (t, s)| sup |ψ1 (s, x0 (s))| |Pn x(s)|ds
t,s∈[−1,1] s∈[−1,1] −1
√
≤ 2M1 d1 Pn x L 2 . (30)
Using estimates (12), (21) and (30), we have
(K1 ψ1 ) (Pn x0 )Pn x∞

≤ [(K1 ψ1 ) (Pn x0 ) − (K1 ψ1 ) (x0 )](Pn x)∞ + (K1 ψ1 ) (x0 )(Pn x)∞
√
≤ M1 c1 Pn x0 − x0 L 2 Pn x L 2 + 2M1 d1 Pn x L 2
√
≤ (M1 c1 p1 Pn x0 − x0 L 2 + 2M1 p1 d1 )x∞ . (31)
In a similar manner, we obtain

√
(K2 ψ2 ) (Pn x0 )Pn x∞ ≤ (M2 c2 p1 Pn x0 − x0 L 2 + 2M2 p1 d2 )x∞ . (32)
Combining estimates (11), (29), (31) and (32), we get

√
Tn (x0 )∞ ≤ (M1 c1 + M2 c2 ) p1 Pn x0 − x0 L 2 + (M1 d1 + M2 d2 ) 2 p1 < ∞. (33)
This shows that Tn (x0 )∞ is uniformly bounded.

Next we consider
|Tn (x0 )x(t) − Tn (x0 )x(t )| ≤ |(K1 ψ1 ) (Pn x0 )Pn x(t) − (K1 ψ1 ) (Pn x0 )Pn x(t )|
+ |(K2 ψ2 ) (Pn x0 )Pn x(t) − (K2 ψ2 ) (Pn x0 )Pn x(t )|.
(34)
We can write
|(K1 ψ1 ) (Pn x0 )Pn x(t) − (K1 ψ1 ) (Pn x0 )Pn x(t )| ≤ T1 + T2 + T3 , (35)
where
T1 = |(K1 ψ1 ) (Pn x0 )Pn x(t) − (K1 ψ1 ) (x0 )Pn x(t)|,

T2 = |(K1 ψ1 ) (x0 )Pn x(t) − (K1 ψ1 ) (x0 )Pn x(t )|,
T3 = |(K1 ψ1 ) (x0 )Pn x(t ) − (K1 ψ1 ) (Pn x0 )Pn x(t )|.
Now using the estimates (11), (12), and (21), we have
T1 = |[(K1 ψ1 ) (Pn x0 ) − (K1 ψ1 ) (x0 )]Pn x(t)|

≤ [(K1 ψ1 ) (Pn x0 ) − (K1 ψ1 ) (x0 )]Pn x∞
≤ M1 c1 (I − Pn )x0 L 2 Pn x L 2
≤ M1 c1 p1 Pn x0 − x0 L 2 x∞ → 0, as n → ∞, (36)
and similarly
T3 = |[(K1 ψ1 ) (Pn x0 ) − (K1 ψ1 ) (x0 )]Pn x(t )|

≤ M1 c1 p1 Pn x0 − x0 L 2 x∞ → 0, as n → ∞. (37)
Since k1 (t, s) ∈ C([−1, 1]×[−1, 1]), k1 (t, s) is uniformly continuous in first variable
t. Hence for any > 0, however small, there exists some number δ > 0 such that
|k1 (t, s) − k1 (t , s)| < , whenever |t − t | < δ.
Using estimate (12), we get

t
(0,1)
T2 = [k1 (t, s) − k1 (t
, s)]ψ1 (s, x0 (s))Pn x(s)ds
−1
1
≤ sup |k1 (t, s) − k1 (t , s)| sup |ψ1(0,1) (s, x0 (s))| |Pn x(s)|ds
−1≤s≤1 s∈[−1,1] −1
√ √
≤ d1 2Pn x L 2 ≤ 2d1 p1 x∞ → 0, as t → t . (38)
Hence from estimates (35), (36), (37) and (38), we have
|(K1 ψ1 ) (Pn x0 )Pn x(t) − (K1 ψ1 ) (Pn x0 )Pn x(t )| → 0, as t → t and n → ∞. (39)
On similar lines, it can be proved that
|(K2 ψ2 ) (Pn x0 )Pn x(t) − (K2 ψ2 ) (Pn x0 )Pn x(t )| → 0, as t → t and n → ∞. (40)
Hence combining estimates (34), (39) and (40), we have
|Tn (x0 )x(t) − Tn (x0 )x(t )| → 0, as t → t and n → ∞. (41)

This implies {Tn (x0 )}∞

n=1 is a family of collectively compact operators. Since T (x 0 )
−1
is compact and (I − T (x0 )) exists, it follows from the theory of collectively
compact operators (see Anselone [14]) that (I − In (x0 ))−1 exists and is uniformly
bounded for sufficiently large n, i.e., ∃ some A2 > 0, independent of n such that
(I − Tn (x0 ))−1 ∞ ≤ A2 .
Theorem 3 Let x0 ∈ C r [−1, 1], r ≥ 1 be an isolated solution of the Eq. (3). Assume
that 1 is not an eigenvalue of T (x0 ), then for sufficiently large n, the iterated solution
x̃n defined by (19) is the unique solution in the sphere B(x0 , δ) = {x : x − x0 ∞ <
δ}. Moreover, there exists a constant 0 < q < 1, independent of n such that
βn βn
≤ x̃n − x0 ∞ ≤ ,
1+q 1−q
where βn = (I − Tn (x0 ))−1 (Tn (x0 ) − T (x0 ))∞ . Also, we obtain
x̃n − x0 ∞ = O(n −r ).
Proof From Theorem 2, we have (I − Tn (x0 ))−1 exists and it is uniformly bounded
on X for sufficiently large n, i.e., there exists a constant A2 > 0, such that
(I − Tn (x0 ))−1 ∞ ≤ A2 .
For any x ∈ B(x0 , δ), consider
[Tn (x) − Tn (x0 )]y∞

= [(K1 ψ1 ) (Pn x) + (K2 ψ2 ) (Pn x) − (K1 ψ1 ) (Pn x0 ) − (K2 ψ2 ) (Pn x0 )]Pn y∞
≤ [(K1 ψ1 ) (Pn x0 ) − (K1 ψ1 ) (Pn x)]Pn y∞ + [(K2 ψ2 ) (Pn x) − (K2 ψ2 ) (Pn x0 )]Pn y∞ .
(42)
Using estimates (12) and (21), we get
[(K1 ψ1 ) (Pn x0 ) − (K1 ψ1 ) (Pn x)]Pn y∞ ≤ M1 c1 Pn (x0 − x) L 2 Pn y L 2
≤ M1 c1 p12 x − x0 ∞ y∞
≤ M1 c1 p12 δy∞ . (43)
Again following similar lines, we can deduce that
[(K2 ψ2 ) (Pn x0 ) − (K2 ψ2 ) (Pn x)]Pn y∞ ≤ M2 c2 p12 δy∞ . (44)
Thus we obtain
−1
sup (I − Tn (x0 )) (Tn (x) − Tn (x0 ))∞ ≤ A2 (M1 c1 + M2 c2 ) p12 δ ≤ q (say).
x−x0 ∞ ≤δ
We choose δ in such a way that 0 < q < 1. Hence the estimate (4.4) of Theorem 2
in [12] is proved.
Next using estimates (13) and (20), we have
Tn (x0 ) − T (x0 )∞ = K1 ψ1 (Pn x0 ) + K2 ψ2 (Pn x0 ) − K1 ψ1 (x0 ) − K2 ψ2 (x0 )∞
≤ K1 ψ1 (Pn x0 ) − K1 ψ1 (x0 )∞ + K2 ψ2 (Pn x0 ) − K2 ψ2 (x0 )∞
√
≤ 2(M1 l1 + M2 l2 )Pn x0 − x0 L 2
√ (r )
≤ 2(M1 l1 + M2 l2 )cn −r x0 ∞ . (45)
This implies
βn = (I − Tn (x0 ))−1 (Tn (x0 ) − T (x0 ))∞

≤ (I − Tn (x0 ))−1 ∞ Tn (x0 ) − T (x0 )∞
√
≤ A2 2(M1l1 + M2 l2 )cn −r x0(r ) ∞ → 0, as n → ∞. (46)
Choose n large enough such that βn ≤ δ(1 − q). Thus Eq. (4.5) of Theorem 2 in [12]
is satisfied. Hence by Theorem 2 of [12],
βn βn
xn − x0 ∞ ≤
≤ ,
1+q 1−q
where
−1
βn = (I − Tn (x0 )) (Tn (x0 ) − T (x0 ))∞ .
Hence from estimate (46), we have
x̃n − x0 ∞ = O(n −r ).

Remark 1 From Theorems 1 and 3, we observe that the Legendre Galerkin 1 solution

and the iterated Legendre Galerkin solution converges with the orders O n 2 −r and
O(n −r ) in the infinity norm, respectively. This shows that iterated Legendre Galerkin
approximation improves over the Legndre Galerkin approxiamtion and optimal con-
vergence rate is obtained in case of the iterated Legendre Galerkin approximate
solution.
References
1. Pachpatte, B.G.: On a nonlinear Volterra-Fredholm integral equation. Sarajevo J. Math. 16,

61–71 (2008)
2. Brunner, H.: On the numerical solution of nonlinear Volterra-Fredholm integral equation by
collocation methods. SIAM J. Numer. Anal. 27, 987–1000 (1990)
3. Wazwaz, A.M.: A reliable treatment for mixed Volterra-Fredholm integral equations. Appl.
Math. Comput. 127, 405–414 (2002)
4. Bakodah, H.O., Darwish, M.A.: On discrete Adomian decomposition method with Chebyshev
abscissa for nonlinear integral equations of Hammerstein Type. Adv. Pure Math. 2, 310–313
(2012)
5. Darwish, M.A.: Fredholm-Volterra integral equation with singular kernel. Korean J. Comput.
Appl. Math. 6, 163–174 (1999)
6. Darwish, M.A.: Note on stability theorems for nonlinear mixed integral equations. J. Appl.
Math. Comput. 6, 633–637 (1999)
7. Maleknejad, K., Karami, M., Rabbani, M.: Using the Petrov-Galerkin elements for solving
Hammerstein integral equations. Appl. Math. Comput. 172, 831–845 (2006)
8. Maleknejad, K., Derili, H.: The collocation method for Hammerstein equations by Daubechies
wavelets. Appl. Math. Comput. 172, 846–864 (2006)
9. Delves, L.M., Mohamed, J.L.: Computational Methods for Integral Equations. Cambridge
University Press, Cambridge, UK (1985)
10. Kaneko, H., Noren, R.D., Xu, Y.: Regularity of the solution of Hammerstein equations with
weakly singular kernels. Integr. Eqn. Oper. Theory 13, 660–670 (1990)
11. Canuto, C., Hussaini, M.Y., Quarteroni, A., Zang, T.A.: Spectral Methods, Fundamentals in
Single Domains. Springer-Verlag, Berlin (2006)
12. Vainikko, G.M.: A perturbed Galerkin method and the general theory of approximate methods
for nonlinear equations. USSR Comput. Math. Math. Phys. 7, 1–41 (1967)
13. Ahues, M., Largillier, A., Limaye, B.: Spectral Computations for Bounded Operators. CRC
Press, Boca Raton (2010)
14. Anselone, P.M.: Collectively Compact Operator Approximation Theory and Application to
Integral equations. Prentice Hall, Englewood cliffs, NJ (1971)
Analysis of an Eco-Epidemiological Model
with Migrating and Refuging Prey
Shashi Kant and Vivek Kumar
Abstract This paper concerns a predator–prey system with migrating and refuging
prey with disease infection. Analysis of the model regarding stability has been per-
formed. The effect of time delay on the above system is also studied. By assuming
the time delay a bifurcation parameter, the stability of the positive equilibrium, and
Hopf-bifurcation is studied. Further, the directions of Hopf-bifurcation and the sta-
bility of bifurcated periodic solutions are calculated using the famous normal form
theory, Riesz representation theorem and central manifold theorem. This is not a case
study, hence real data is not available. However, to verify our theoretical predictions,
some numerical simulations are also included.
Keywords Predator–prey model · Stability · Hopf-bifurcation · Migration ·

Refuge · Delay
1 Introduction
The dynamic relation between prey and predator has been studied extensively in the
literature. At first sight, prey–predator dynamics may seem very simple mathemati-
cally, but they are, in fact very difficult and challenging. The classical Lotka-Volterra
model is a first stepping stone in the study of prey–predator dynamics and inter-
actions [1, 23]. In mathematical ecology, this model is extensively used and cited
and proved a milestone in the progress of mathematical ecology. On the other hand,
the famous work of Kermack-Mckendric [25] in epidemiological studies received
much attention among applied mathematicians, scientists, and ecologists. After the
work of [1, 23, 25], many mathematical models have been published for reference
S. Kant (B) · V. Kumar

Department of Applied Mathematics, Delhi Technological University,
Delhi 110042, India
e-mail: onlineskmishra@gmail.com
V. Kumar
e-mail: vivekkumar.ag@gmail.com

DOI 10.1007/978-81-322-2485-3_2
18 S. Kant and V. Kumar
(see [2, 6, 7, 13–17, 21, 24, 27, 29], etc., and references therein). Combined and/or
overlapping study of ecology and epidemiology is termed as Eco-epidemiology. Eco-
epidemiological models are gaining popularity day-by-day. The present study also
falls under the purview of Eco-epidemiology.
To study the environmental impact on prey–predator models, the ‘time delay’ has
been investigated by the researchers. A good number of papers are available in the
literature for instance (see [4, 8, 18]). In these papers, most of the authors investigate
the ‘time delay’ as a game changing. Time delay may cause changes in stability,
occurrence for limit cycle, bifurcation, etc.
Further, migration of species especially prey is also evolved and few references
are available in the literature, for reference we can refer [11, 19, 20]. Prey-refuge
in prey-predator models also play an important role in dynamical nature. Further if
prey-refuge is more than outbreak of the prey population occurs. To understand a
role of prey-refuge in mathematical ecology few publications are available. At this
juncture we may refer readers to ([9, 22, 24, 26, 29] and references therein).
Pal and Samanta [3] proposed the following mathematical model by incorporating
prey-refuge in the model proposed of Xiao and Chen [28]:
⎧
⎪
⎨ dt = r1 S(1 − k ) − S I β,
dS S+I
dt = S I β − cI − aY +I ,
dI bI Y (1)
⎪
⎩ dY pbI Y
dt = −dY + aY +I .
Motivated by the model of Samanta [18] and model in [12], Hu and Li [10] proposed
the following model:
⎧
⎪
⎨ dt = r S 1 − k − S I β − p1 SY,
dS S+I
dt = −cI + S I β − p2 I Y,
dI (2)
⎪
⎩ dY
dt = −dY + qp1 S(t − τ )Y (t − τ ) + qp2 I (t − τ )Y (t − τ ).
In order to study the influence of prey-refuge, migration, and disease on the Prey-
predator system, in this paper, we concentrate on an eco-epidemiological prey-
predator system consisting of three species as in [10]. Motivated by the models in
[10] and [3], we propose a mathematical model in which prey is migrating and refug-
ing with disease in both species. We present stability and Hopf-bifurcation analysis
of the mathematical model. Detailed assumptions for model formulation are listed
in the next section.
The rest of the paper is structured as follows: In the next section, we formulate our
main mathematical model with the help of biological and ecological assumptions.
In Sect. 3, we consider the model without delay. In Sect. 4, we discuss the stability
of mathematical model with delay. In Sect. 5, we discuss the direction and stability
of Hopf-bifurcation using the normal form theory, Riesz representation theorem and
central manifold theorem as in [10]. Numerical simulations have been done in Sect. 6
followed by discussion in the last Sect. 7.
Analysis of an Eco-Epidemiological Model with Migrating and Refuging Prey 19
2 The Model
In this paper, we propose to study a prey-predator system by means of mathematical

modeling. To formulate the model and in view of simplicity we make the following
assumptions:
• In the absence of disease and predation the healthy (susceptible) prey population
has logistic growth with growth rate r and carrying capacity k, i.e.,

dN N
= rN 1 − . (3)
dt k
• Disease is spreading in both populations. After disease prey population is divided

into two parts susceptible prey(S) and infected prey (I). Thus total biomass of prey
population is S(t) + I (t).
• Due to mathematical complexity, the bifurcation of predator population and the
detailed dynamics of the disease infection in the predator population is omitted.
Further, it is also assumed that disease infection in predator occurs due to eating
of the infected prey and not due to outside infection. In other words, it is easy to
understand that the disease infection starts from prey and then carries forward to
predator. For example H1N1, H5N1, etc., may be pointed out here to understand
the physical phenomenon better. Thus total biomass of predator population is Y.
• Infected prey population does not become immune as well as they have no
reproduction rate. However, infected prey population contributes the carrying
capacity k.
• Predator population consumes both susceptible as well as infected prey population.
• Due to environmental and fear factors, we consider out migration in prey popula-
tion, i.e., once prey migrated they will not return. Let m 1 and m 2 be the migration
rates of susceptible and infected prey respectively. Further, healthy prey population
is more active compared to infected one. Hence, healthy prey can migrate more
easily than infected prey before their predation. Hence, by using this ecological
information, we can impose the mathematical condition m 1 > m 2 .
• Let d2 and d3 be natural death rates for infected prey and predator population
respectively.
• Death (mortality) rate due to disease for infected prey population and predator
population are denoted by c and d4 respectively.
• The coefficient for S-prey and I-prey to predator are denoted by q1 and q2 respec-
tively. The relationship between q1 and q2 is established later.
• Let a refuge protecting m 3 S of healthy prey and m 4 I that of infected prey, where
m 3 , m 4 ∈ [0, 1). Hence, (1 − m 3 )S and (1 − m 4 )I of healthy and infected prey,
respectively, are available to the predator for predation.
Based on these assumptions, model takes the following form:

⎧
⎪
⎨ dt = r S 1 − k − S I β − p1 (1 − m 3 )SY − m 1 S,
dS S+I
dt = S I β − p2 (1 − m 4 )I Y − d2 I − m 2 I − cI,
dI
⎪
⎩ dY
dt = q1 p1 (1 − m 3 )S(t − τ )Y (t − τ ) + q2 p2 (1 − m 4 )I (t − τ )Y (t − τ ) − d3 Y − d4 Y.
(4)
The initial conditions are
⎧
⎪
⎪ S(t) = φ1 (t) > 0,
⎪
⎪
⎪
⎪
⎨ I (t) = φ2 (t) > 0,
Y (t) = φ3 (t) > 0, (5)
⎪
⎪
⎪
⎪ 3 ,
(φ1 (t), φ2 (t), φ3 (t)) ∈ C = C [−τ , 0], R+
⎪
⎪
⎩ R 3 = {(x, y, z)|x ≥ 0, y ≥ 0, z ≥ 0},
+
where,
β : Disease Contact Rate

p1 , p2 : Predation Coefficients of Susceptible (S) and Infected (I) Prey
τ : Gestation period (delay).
Ecological and biological assumptions suggests the following relationship

between q1 and q2 :
q2 = q1 and 0 < q1 ≤ 1,
q2 > q1 and 0 < q2 ≤ 1.
3 Analysis of the Model Without Delay
In this section, model (4) is investigated under the condition τ = 0. Before going to
main analysis, we state two lemmas for our model without proof.
Lemma 1 Each solution of the system (4) without delay with the initial conditions
(5) are strictly positive for all t ≥ 0.
Lemma 2 Solutions of the system (4) without delay with the initial conditions (5)
3.
are eventually bounded, i.e., uniformity bounded in R+
3.1 Equilibrium Points
System of ODEs under consideration has the following equilibrium points:

(i) E 1 (0, 0, 0) = (0, 0, 0).

(ii) E 2 (
S, 0, 0). where
S = (r − m 1 ) rk .
∗ ∗
(iii) E 3 (S , 0, Y ), where
⎧
⎨ S ∗ = q pd3(1+−d4m ) ,
1 1 3
⎩Y ∗ = kp1 (1 − m 3 )q1 (r − m 1 ) − r (d3 + d4 )

kq1 (1 − m 3 ) p12
.
(iv) E 4 (S, I , 0), where

⎧
⎨S = c + d2 + m 2
β ,
⎩I = {(r − m 1 )kβ − r (c + d2 + m 2 )}
.
(β(r + kβ))
(v) E 5 (
S, ), where
I,Y
⎧ (r − m 1 )q2 (1 − m 4 ) p2 k + (c + d2 + m 2 )q2 p1 (1 − m 3 )(1 − m 4 ) p2 k − (r + kβ)(d3 + d4 )
⎪
⎪S= ,
⎪ rq2 p2 (1 − m 4 ) − q1 p1 (1 − m 3 )(r + kβ) + q2 (1 − m 4 ) p2 βk
⎨
(d3 + d4 ) − q1 p1 (1 − m 3 )S
I = q2 p2 (1 − m 4 ) ,
⎪
⎪
⎪
⎩Y βS − (c + d2 + m 2 )
= .
p2 (1 − m 4 )
3.1.1 Existence Conditions
We have the following existence conditions:

(i) Trivial equilibrium E 1 always exists.
(ii) E 2 exists provided (r − m 1 ) > 0 or r > m 1 . Physical meaning implies that
existence of E 2 is independent of other parameters and depends only on growth
rate and migration of S, viz., r and m 1 . E 2 exists if growth rate of S is greater
than migration of itself.
(iii) E 3 exists provided dr3−m+d4 > kq1 p1 (1−m 3 ) . This is the case when no disease
1 r
infection occurs in the prey population.

r −m 1
(iv) E 4 exists provided c+d 2 +m 2
> kβ
r
. This is the case when predator does not
survive.
(v) E 5 exists provided the following conditions are satisfied:
⎧
⎪
⎪ d3 + d4 > q1 p1 (1 − m 3 ) S,
⎪
⎨β S > (c + d2 + m 2 ),
⎪
⎪ (r − m 1 )q2 p2 (1 − m 4 )k + (c + d2 + m 2 )q2 p2 (1 − m 4 ) p1 (1 − m 3 )k > (r + kβ)(d3 + d4 ),
⎪
⎩
(rq2 p2 (1 − m 4 ) + q2 p1 (1 − m 3 )βk) > q1 p1 (1 − m 3 )(r + kβ).
This equilibrium point is very important, since it provides the coexistence of all the
three populations.
3.2 Stability Analysis
Jacobian matrix of the system is given by

⎛ ⎞
2r S rI
(r − m 1 − K − k − β I − p1 (1 − m 3 )Y ) − rkS − β S (− p1 (1 − m 3 )S)
⎜ ⎟
J =⎝ (β I ) (β S − p2 (1 − m 4 )Y − c − d2 − m 2 ) (− p2 (1 − m 4 )I ) ⎠,
(q1 (1 − m 3 ) p1 Y ) (q2 p2 (1 − m 4 )Y ) (q1 p1 (1 − m 3 )S+q2 p2 (1 − m 4 )I − d3 − d4 )
(6)
with this matrix stability analysis is carried out. We will focus on the non zero
equilibrium point.
After a little calculation we see that trivial equilibrium point is locally stable if
r < (m 1 ). Equilibrium (E 2 ) is locally asymptotically stable provided the following
conditions are satisfied:

(β
S − c − d2 − m 2 ) = βk(r r−m 1 ) < 0,
S − d3 − d4 ) = q1 p1 (1 − m 3 ) k(r −m
(q1 p1 (1 − m 3 )
r
1)
− d3 − d4 < 0.
Equilibrium (E 3 ) is locally asymptotically stable if

⎧
⎪
⎪ β(d3 +d4 )
− p2 (1−m 4 )(r −m 1 )
+ p2 (1−m 3 )r (d3 +d4 )
− (c + + ) < 0;
⎪ q1 (1−m 3 ) p1
⎪ p1 (1−m 3 ) kq1 (1−m 3 ) p12
d2 m 2
⎪
⎪
⎪
⎨
Quadratic equation (λ2 − ξλ + ζ) have roots with negative real parts, where
⎪
⎪ξ = −(d3 +d4 )[kq1 p1 (1−m 3 )+r ] ,
⎪
⎪
⎪
⎪ kq1 (1−m 3 ) p1
⎪
⎩ (d3 +d4 )[kq1 p1 (1−m 3 )(r −m 1 )−(d3 +d4 )r ]
ζ= kq1 p1 (1−m 3 ) .
Equilibrium (E 4 ) is locally asymptotically stable if

⎧
⎪
⎪(q1 p1 S + q2 p2 I − d3 − d4 ) = −(d3 + d4 ) + q1 p1 (c+d
β
2 +m 2 )
+ q2 p2 [(r −m 1β(r
)kβ−r (c+d2 +m 2 )]
+kβ) < 0;
⎪
⎪
⎨Equation (λ2 − Bλ + C) have roots with negative real parts, where
−3r (c+d2 +m 2 )
⎪B =
⎪
⎪ kβ ,

⎪
⎩C = (c + d + m ) (r − m ) − r (c+d2 +m 2 ) .
2 2 1 kβ
Remark 1 In this case no infection occurs in the system, hence ecologically mortality
due to infection in predator population may be omitted. Similarly, parameter c may
also be deleted. Jacobian matrix at E 3 is now reduced to
⎡ ∗
∗
⎤
r − m 1 − 2rkS − p1 (1 − m 3 )Y ∗ − r Sk − β S (− p1 (1 − m 3 )S ∗ )
⎢ ⎥
J =⎣ 0 (β S ∗ − p2 (1 − m 4 )Y ∗ − d2 − m 2 ) 0 ⎦,
(q1 p1 (1 − m 3 )Y ∗ ) (q2 p2 (1 − m 4 )Y ∗ ) (q1 p1 (1 − m 3 )S ∗ − d3 )
(7)
where S∗ = d3
q1 p1 (1−m 3 ) and Y∗ = kq1 p1 (1−m 3 )(r −m 1 )−r (d3 )
kq1 (1−m 3 ) p12
.
Characteristic equation is given by (λ − λ1 )(λ2 − ξλ + ζ) = 0,
β(d3 ) p2 (1−m 4 )(r −m 1 )

where λ1 = (β S ∗ − p2 (1 − m 4 )Y ∗ − d2 − m 2 ) = q1 p1 (1−m 3 ) − p1 (1−m 3 ) +
p2 (1−m 4 )r (d3 )
kq1 (1−m 3 ) p12
− (d2 + m 2 ),
ξ = −(d3 )[kq 1 p1 (1−m 3 )+r ]

kq1 (1−m 3 ) p1 , ζ = (d3 )[kq1 p1 kq
(1−m 3 )(r −m 1 )−(d3 )r ]
1(1−m 3 ) p1
.
Thus E 3 is locally asymptotically stable if the following conditions are satisfied:
⎧
⎨ β(d3 )
− p2 (1−m 4 )(r −m 1 )
+ p2 (1−m 4 )r (d3 )
− (d2 + m 2 ) < 0,
q1 (1−m 3 ) p1 p1 (1−m 3 ) kq1 (1−m 3 ) p12
⎩
Equation (λ2 − ξλ + ζ) have roots with negative real parts.
3.2.1 Positive Equilibrium
In this case, populations of all the three species exists simultaneously. As promised,
we will furnish the detail of the stability of the positive equilibrium point. For the
stability of the positive equilibrium E 5 , we state the following theorem:
Theorem 1 System (4) with τ = 0 is locally asymptotically stable at E 5 if the

following conditions are satisfied:
(i)
S + + < 0.
I + Y
(ii) A1 A2 + A3 > 0,
where = (q1 p1 (1 − m 3 ) + β − 2rk ),
= (−( rk + β) + q2 (1 − m 4 ) p2 ),
= −( p1 (1 − m 3 ) + p2 (1 − m 4 )),
= (r − m 1 − m 2 − d2 − d3 − d4 − c),
A1 = S + I + Y + ,
A2 = S [βq1 p1 (1 − m 3 ) − 2rq1 (1 −
2
k
m 3 ) p1
− 2rkβ ] + Y 2 [ p1 (1 − m 3 )(1 −
m 4 ) p2 ] + I 2 [−( k + β)q2 p2 (1 − m 4 )] +
r
S I [−( k + β)q1 p1 (1 − m 3 ) + βq2 p2
r
2rq1 p1 (1 − m 3 )
(1 − m 4 ) − k ] + Y I [−q p
2 2 (1 − m 4 )(1 − m 3 ) p1 + p2 (1 − m 4 )( rk +
β)]+ SY[−q1 p2 p1 (1 − m 4 )(1 − m 3 )+ 2r p2 (1 − m 4 ) − p1 (1 − m 3 )β]+ S[−β(d3 +d4 )
k
− (c + m 2 + d2 )q1 p1 (1 − m 3 ) + (r − m 1 )q1 p1 (1 − m 3 ) + 2r (d3k+d4 ) + β(r −
m 1 ) + 2r (c + dk2 + m 2 ) ] + I [−(c + d2 + m 2 )(1 − m 4 )q2 p2 + (r − m 1 )q2 (1 −
m 4 ) p2 + ( rk + β)(d3 + d4 ) + ( rk + β)(c + d2 + m 2 )] + Y [ p1 (1 − m 3 )(d3 + d4 )
− p2 (r − m 1 )(1 − m 4 ) + p1 (1 − m 3 )(c + m 2 + d2 )] + [ p2 (1 − m 4 )(d3 + d4 ) +
(c + m 2 + d2 )(d3 + d4 ) − (d3 + d4 )(r − m 1 ) − (r − m 1 )(c + m 2 + d2 )],
A3 = S 3 [− 2r βq1 p1k(1 − m 3 ) ] + [ 2rq1 p1 p2 (1 − m 4 )(1 − m 3 ) ] +
S2Y k I 2S[q2 p2 (1 − m 4 )
βq (1 − )
β( k + β)] + S I [
r 2 2r p
2 2
k
m 4
] + S [(r − m 1 )βq1 (1 − m 3 ) p1 + 2r β(dk3 + d4 ) +
2
2rq1 p1 (1 − m 3 )(c + m 2 + d2 )
k ]+ I 2[( rk +β)q2 (1 − m 4 ) p2 (c+d2 +m 2 )]+ S [−2βq2 p1 p2
IY

(1 − m 4 )(1 − m 3 ) + q1 p1 p2 (1 − m 4 )(1 − m 3 )( + β)] + S I [(r − m 1 )βq2 p2 (1 −
r
k
m 4 ) + ( rk + β)(c + m 2 + d2 )q1 p1 (1 − m 3 ) + 2rk (c + m 2 + d2 )q2 p2 (1 − m 4 )] +
[β p1 (1 − m 3 )(d3 + d4 )] +
SY [(c + m 2 + d2 )q2 p1 p2 (1 − m 4 )(1 − m 3 )] +
IY
S[(r − m 1 ){−β(d3 + d4 ) − (c + m 2 + d2 )q1 p1 (1 − m 3 )} − 2r p2 (1 − mk4 )(d3 +d4 ) −

2r (c+m 2 +d2 )(d3 +d4 )
k ]+
I [−(r − m 1 )(c + m 2 + d2 )q2 p2 (1 − m 4 ) − ( rk + β)(c + m 2 +
d2 )(d3 + d4 )] + [ p2 (1 − m 4 )(d3 + d4 )(r − m 1 ) + (r − m 1 )(c + m 2 + d2 )(d3 + d4 )].
Proof 1 Jacobian matrix at E 5 is given by
⎛ 2r ⎞
(r − k
S
− f racr
I k − β )
I − p1 (1 − m 3 )Y (− rkS − β
S) (− p1 (1 − m 3 )
S)
J =⎝ (β
I) (β − c − d2 − m 2 )
S− p2 (1 − m 4 )Y (− p2 (1 − m 4 )
I) ⎠.
)
(q1 p1 (1 − m 3 )Y )
(q2 p2 (1 − m 4 )Y (q1 p1 (1 − m 3 )
S + q2 p2 (1 − m 4 )
I − d3 − d4 )
(8)
The characteristic equation is given by λ3 + A1 λ2 + A2 λ + A3 = 0, where A1 ,
A2 and A3 are the same as defined in statement of the theorem. By Routh-Hurwitz
criteria the theorem follows.
4 Stability Analysis of the Model with Time Delay
In this section, model (4) with τ = 0 is considered. It is also important to mention that
we will consider the positive equilibrium (E ∗ ) only. At any point, jacobian matrix
of system (4) is given by
⎡ ⎤
r 1 − S+I
k − rkS − β I − p1 (1 − m 3 )Y − m 1 − rkS − β S (− p1 (1 − m 3 )S)
⎢ ⎥
J =⎣ (β I ) (β S − p2 (1 − m 4 )Y − c − d2 − m 2 ) (− p2 (1 − m 4 )I )⎦
0 0 (−d3 − d4 )
⎡ ⎤
0 0 0
+⎣ 0 0 0 ⎦ (e−λτ ),
(q1 p1 (1 − m 3 )Y ) (q2 p2 (1 − m 4 )Y ) (q1 p1 (1 − m 3 )S + q2 p2 (1 − m 4 )I )
here λ being a complex number. Now we state two more lemmas from [10, 27];
Lemma 3 Let λ = (A + i B) A > 0, B > 0 then,
• if A < B , all roots of the −λτ = 0 have positive real parts
−1
A equation λ + A − Be
for τ < √ 2 2 cos
1
B .
B −A
• if A > B, all roots of the equation λ + A − Be−λτ = 0 have negative real parts
for any τ .
Lemma 4 Let the polynomial, h(z) = z 3 + p0 z 2 + q0 z + r0 = 0

(i) if r0 < 0, then this equation has at least one positive root;
(ii) if r0 ≥ 0 and = ( p02 − 3q0 ) ≤ 0 then this equation has no positive roots;
(iii) if r0 ≥ 0 and =√( p02 − 3q0 ) > 0, then this equation has positive roots if and
− p0 +
only if z 1∗ = 3 and h(z 1∗ ) ≤ 0.
Let us study nonzero equilibrium E ∗ = (S∗ , I∗ , Y∗ ). The jacobian matrix at E ∗ =

(S∗ , I∗ , Y∗ ) is given as
J (E ∗ ) =
⎛ S ⎞
r 1− ∗ +I∗ − r S∗
− p1 (1 − m 3 )Y∗ − β I∗ − m 1 − r Sk∗ − β S∗ (− p1 (1 − m 3 )S∗ )
k k
⎜ β I∗ (β S∗ − p2 (1 − m 4 )Y∗ − c − d2 − m 2 ) p2 (1 − ,m 4 )I∗ ⎟
⎝ ⎠.
(q1 p1 (1 − m 3 )Y∗ e−λτ ) (q2 p2 (1 − m 4 )Y∗ e−λτ ) (q1 p1 (1 − m 3 )S∗
+ q2 p2 (1 − m 4 )I∗ )e −λτ − d3 − d4 )
The characteristics equation is given as (λ3 + M2 λ2 + M1 λ + M0 ) + (n 2 λ2 + n 1 λ +

n 0 )e−λτ = 0, where,

2r p1 (1 − m 3 )β q1 p1 (1 − m 3 )β
M2 = + − (1 − m 4 )S∗
k p2 (1 − m 4 ) q 2 p2

r (d3 + d4 ) p1 (1 − m 3 )(c + d2 + m 2 )
+ + d3 + d4 − −r ,
q2 p2 (1 − m 4 )K p2 (1 − m 4 )
M1 = ({(d3 + d4 )(c + d2 + m 2 ) − (r − m 1 )(c + d2 + m 2 )} + S∗ {−(d3 + d4 )β +

K 3 + d4 ) + (r − m 1 )β + k (c + d2 + m 2 )} + I∗ {( k + β)(d3 + d4 + c + d2 +
(d
2r 2r r
m 2 )} + Y∗ {(d3 + d4 ) p2 (1 − m 4 ) + (d3 + d4 ) p1 (1 − m 3 ) + (c + d2 + m 2 ) p1 (1 − m 3 ) +
(r − m 1 ) p2 (1 − m 4 )} + S∗2 (− 2rkβ ) + Y∗2 ( p1 p2 (1 − m 3 )(1 − m 4 )) + S∗ Y∗ (−β p1 (1 −
m 3 ) + 2rk p2 (1 − m 4 )) + I∗ Y∗ ( rk + β) p2 (1 − m 4 )),
M0 = ({−(r − m 1 )(c + d2 + m 2 )(d3 + d4 ) + S∗ {(d3 + d4 )β(r − m 1 ) + 2rk (d3 +
d4 )}(c + d2 + m 2 )} + I∗ {(d3 + d4 )(c + d2 + m 2 )( rk + β)} + Y∗ {(d3 + d4 )(c + d2 +
m 2 ) p1 (1−m 3 )−(r −m 1 )(d3 +d4 ) p2 }+ S∗2 (− 2rk (d3 +d4 )β)+Y∗2 (d3 +d4 ) p2 p1 (1−
m 3 )(1−m 4 ))+S∗ Y∗ {−(d3 +d4 )β p1 (1−m 3 )− 2rk (d3 +d4 ) p2 (1−m 4 )}+I∗ Y∗ {( rk +β)
p2 (1 − m 4 )(d3 + d4 )}),
n 2 = −(d3 + d4 ),
n 1 = (S∗2 {q1 p1 (1 − m 3 )β − −2r k q1 p1 (1 − m 3 )} + I∗ {−( k + β)q2 p2 (1 − m 4 )} +
2 r
S∗ Y∗ {−q1 p2 p1 (1 − m 3 )(1 − m 4 )} + S∗ I∗ {q2 p2 (1 − m 4 )β + 2rk q2 p2 (1 − m 4 )} +

Y∗ I∗ {−q2 p2 p1 (1 − m 3 )(1 − m 4 )} + S∗ {{−(c + d2 + m 2 ) − (r − m 1 )}q1 p1 (1 −
m 3 )} + I∗ {−(c + d2 + m 2 )q2 p2 (1 − m 4 ) + (r − m 1 )q2 p2 (1 − m 4 )}),
n 0 = −(S∗2 {q1 p1 (1 − m 3 )β(r − m 1 − 2rk ) + 2r K (c + d2 + m 1 )q1 p1 (1 − m 3 ) −
q1 { p1 (1 − m 3 )}2 β} + I∗2 {( rk + β)(c + d2 + m 2 )q2 p2 (1 − m 4 )} + S∗2 Y∗ { 2rk q1 p1 p2 (1 −
m 3 )(1 − m 3 ) + q1 { p1 (1 − m 3 )}2 β} + S∗2 I∗ {− 2rk q2 p2 (1 − m 4 )β} + S∗ I∗ Y∗ {( rk +
β)q1 p1 p2 (1 − m 3 )(1 − m 4 ) − 2βq2 p1 p2 (1 − m 3 )(1 − m 4 )} + S∗ Y∗ {{−(r − m 1 ) +
( rk + β)}q1 p1 p2 (1 − m 3 )(1 − m 4 )} + S∗ I∗ {β(r − m 1 )q2 p2 (1 − m 4 ) + 2rk (c +
d2 + m 2 )q2 p2 (1 − m 4 ) + (c + d2 + m 2 )( rk + β)q1 p1 (1 − m 3 )} + Y∗ I∗ {(c + d2 +
m 2 )q2 p2 p1 (1−m 3 )(1−m 4 )}+ S∗ {(c +d2 +m 2 )(r −m 1 )q1 p1 (1−m 3 )}+ I∗ {−(c +
d2 + m 2 )(r − m 1 )q2 p2 (1 − m 4 )}).
Now we put λ = iω (ω > 0) we get
Real Part: {n 2 ω 2 + n 0 } cos ωτ + {n 1 ω sin ωτ − M2 ω 2 + M0 }, Imaginary Part:
n 1 ω cos ωτ −(−n 2 ω 2 +n 0 ) sin ωτ +M1 ω−ω 3 (Real Part)2 +(Imaginary Part)2 =
ω 6 + p0 ω 4 + q0 ω 2 + r0 . Hence, we have ω 6 + p0 ω 4 + q0 ω 2 + r0 = 0, where
p0 = (M22 − 2M1 − n 22 ) q0 = (M12 − 2M2 M0 + 2n 2 n 0 − n 21 ) r0 = (M02 − n 20 ). If
we put z = ω 2 , then we have the equation z 3 + p0 z 2 + q0 z + r0 = 0. If M02 ≥ n 20 ,
then we will have r0 ≥ 0, we have two situations for (i) = ( p02 − 3q0 ) ≤ 0.
(ii) = ( p02 − 3q0 ) > 0.
In situation (i) we have to say that E ∗ is absolutely stable if r0 ≥ 0 and =
( p02 − 3q0 ) ≤ 0. Also, if we have and r0 ≥ 0 = ( p02 − 3q0√) > 0 then equation
− p0 +
has negative roots if and only if h(z 1∗ ) > 0 where z 1∗ = 3 thus we have the
following theorem for the stability of E ∗ .
Theorem 2 E ∗ (S∗ , I∗ , Y∗ ) is absolutely stable if one of the following conditions
holds:
(i) = ( p02 − 3q0 ) ≤ 0. √
− p0 +
(ii) = ( p02 − 3q0 ) > 0 and z 1∗ = < 0.
√3
− p0 +
(iii) = ( p02 − 3q0 ) > 0, z 1∗ = 3 > 0 and h(z 1∗ ) > 0 provided r0 ≥ 0.
Next, if we consider the case when r0 < 0 or {r0 ≥ 0, = ( p02 − 3q0 ) > 0,
z 1∗
> 0, h(z 1∗ ) < 0}. Then, according to lemma , equation will have one positive
root say ω0 that is the characteristic equation has a pair of purely imaginary roots
say ±iω0 . Now assume that iω0 , ω0 > 0 is a root of h(z), then we have real and
imaginary parts as under.
Real Part = {n 2 ω 2 + n 0 } cos ωτ + {n 1 ω sin ωτ − M2 ω 2 + M0 } = 0.
Imaginary Part = n 1 ω cos ωτ − (−n 2 ω 2 + n 0 ) sin ωτ + M1 ω − ω 3 = 0.
Solving the above equation for τ , we have (by eliminating sin ωτ between these
equations)
n 1 ω02 {ω0 −M1 }−{M2 ω02 −M0 }{n 2 ω02 −n 0 }
τ= 1
ω0 cos−1 n 21 ω02 +n 2 ω02 −n 0
+ 2kπ
ω0 , (k = 0, 1, 2, . . .)
We call it as a ’critical value’ and may be denoted as τk = ω10 cos−1

n 1 ω02 {ω0 −M1 }−{M2 ω02 −M0 }{n 2 ω02 −n 0 }
n 2 ω 2 +n ω 2 −n
+ 2kπ
ω0 , (k = 0, 1, 2, . . .). This is correspond-
1 0 2 0 0
ing to the characteristic equation as it has purely imaginary roots ±iω, which is
a result similar to that of Hu et al. 2012 [10]. Differentiating the characteristics
equation w.r.t. τ , we get ( dλ −1 = (3λ2 +2M2 λ+M1 )eλτ + 2n 2 λ+n 1
− λτ or
dτ ) (λ2 n +λn +n )λ (λ2 n +λn +n )λ
2 1 0 2 1 0
λτ
= (3λ +2M2 λ+M(λ1 )e + (2n 2 λ+n 1 ) − τ (2n 2 λ+n 1 )
2
( dλ −1
dτ ) 2 n +λn +n )λ
2 1 0
. As proved in [10], it is easy to
d(Reλ)
prove the transversality condition at τk e.g. dτ = 0. τk is used as a point for
direction of Hopf Bifurcation as in the next section.
Remark 2 The equilibrium points E 5 of model (4) with τ = 0 and E ∗ of model (4)
with τ = 0 are ecologically similar. Both convey the message that all the species
exist simultaneously.
5 Direction and Stability of the Hopf-Bifurcation
With the symbols used in [10] and procedure explained in [5], we have the following
system of functional differential equation, u̇(t) = L μ (μt )+ F(μ, u t ), where u t (θ) =
u(t + θ) ∈ R 3 and L μ : R × C → R3 and F : R × C → R3 are given as
⎡ r S∗ ⎤
− k − rk + β S∗ (− p1 (1 − m 3 )S∗ )
L μ φ = (τk + μ) ⎣ β I∗ (β S∗ − p2 (1 − m 4 )Y∗ − c − d2 − m 2 ) − p2 (1 − m 4 )I∗ ⎦ × φ(0)
0 0 −d3 − d4 )
⎡ ⎤
0 0 0
+ (τk + μ) ⎣ 0 0 0 ⎦ × φ(−1),
q1 p1 (1 − m 3 )Y∗ q2 p2 (1 − m 4 )Y∗ q1 p1 (1 − m 3 )S∗ + q2 p2(1−m 4 ) I∗
and
⎛ ⎞
− rk φ21 (0) − rk + β φ1 (0)φ2 (0) − p1 (1 − m 3 )φ1 (0)φ3 (0)
F(μ, θ) = ⎝ βφ1 (0)φ2 (0) − p2 (1 − m 4 )φ2 (0)φ3 (0) ⎠,
q1 p1 (1 − m 3 )φ1 (−1)φ3 (−1) + q2 p2 (1 − m 4 )φ1 (−1)φ2 (−1)
φ(0) ≡ (φ1 (0), φ1 (0), φ1 (0))T ∈ C i.e.
⎡ r S∗ ⎤ ⎛ ⎞
− k − rk + β S∗ (− p1 (1 − m 3 )S∗ ) φ1 (0)
L μ φ = (τk + μ) ⎣ β I∗ (β S∗ − p2 (1 − m 4 )Y∗ − c − d2 − m 2 ) − p2 (1 − m 4 )I∗ ⎦ × ⎝ φ2 (0) ⎠
0 0 −d3 − d4 φ3 (0)
⎡ ⎤ ⎛ ⎞
0 0 0 φ1 (−1)
+ (τk +μ) ⎣ 0 0 0 ⎦ ⎝
× φ2 (−1) ⎠ ,
q1 p1 (1 − m 3 )Y∗ q2 (1 − m 4 ) p2 Y∗ q1 p1 (1 − m 3 )S∗ + q2 (1 − m 4 ) p2 I∗ φ3 (−1)
we have considered, τ = (τk + μ), μ = 0 gives the hopf bifurcation value for
the mathematical model with delay as promised in the previous section. Normal-
izing delay τ by timescaling t → τt the model is written in the Banach Space
C ≡ C([−1, 0], R3 ). By the Riesz representation theorem, we found that there
exists a matrix function whose components are bounded variation function η(θ, μ)
in θ ∈ [−1,0] such that
L μ φ = dη(θ, μ)φ(θ),φ ∈ C, ∈ [−1, 0).
We can choose
⎡ r S∗ ⎤
− k − rk + β S∗ (− p1 (1 − m 3 )S∗ )
η(θ, μ) = (τk + μ) ⎣ β I∗ (β S∗ − p2 (1 − m 4 )Y∗ − c − d2 − m 2 ) − p2 (1 − m 4 )I∗ ⎦ × δ(θ)
0 0 −d3 − d4 )
⎡ ⎤
0 0 0
− (τk + μ) ⎣ 0 0 0 ⎦ × δ(θ + 1).
q1 p1 (1 − m 3 )Y∗ q2 p2 (1 − m 4 )Y∗ q1 (1 − m 3 ) p1 S∗ + q2 (1 − m 4 ) p2 I∗
where δ(θ) denotes the dirac delta function, viz.,


0, θ = 0
δ(θ) = .
1, θ = 0,
for φ ∈ C1 ([−1, 0], R3 ), define
dφ(θ)
dθ θ ∈ [−1, 0)
A(μ)φ(θ) = 0 .
−1 dη(θ, μ)φ(θ) θ = 0
dφ(θ)
dθ , −1 ≤ θ < 0
or A(μ)φ(θ) = 0
−1 dη(θ, μ)φ(θ), θ = 0
and
0, θ ∈ [−1, 0) 0, −1 ≤ θ < 0
R(μ)φ(θ) = . = .
F(μ, φ), θ = 0 F(μ, φ), θ = 0
˙ = L μ (μt ) + F(μ, μt ) may be written as
with these symbols, u(t)
˙ = Aμ(μt ) + R(μ)μt
u(t) (9)
which is an abstract differential equation. Where u t (θ) = u(t

+ θ), −1 ≤ θ < 0.
Now we come to operator theory, for ψ ∈ C1 [0, 1], (R3 )∗ we define A∗ , adjoint
operator of A,
− dψ(S) S ∈ (0, 1]
A∗ ψ(S) = 0 d S T
−1 dη (S, μ)ψ(−S) S = 0.
0θ T
And a bilinear product < ψ(S), φ(θ) >= ψ(0)φ(0) − 1 ξ=0 ψ (ξ − θ)dη(θ)
φ(ξ)dξ where η(θ) = η(θ, 0). Then A(0) and A∗ are adjoint operators. Now ±iω0 τk
are eigen values of A(0). Hence they are eigenvalues of A∗ also. To determine the
poincare normal form of the operator A, we first need to evaluate the eigenvectors of
A(0) and A∗ corresponding to iω0 τk and −iω0 τk respectively. Suppose that q(θ) =
(1, α1 , α2 )T exp(iω0 τk θ) is the eigen vector of A(0) corresponding to iω0 τk , then
we have A(0)q(θ) = iω0 q(θ) from the definition of A(0),we have
⎡⎡ r S∗ ⎤
− k − rk + β S∗ (− p1 (1 − m 3 )S∗ )
⎣⎣ β I∗ (β S∗ − p2 (1 − m 4 )Y∗ − c − d2 − m 2 ) − p2 (1 − m 4 )I∗ ⎦
0 0 −d3 − d4
⎡ ⎤ ⎤⎛ ⎞
0 0 0 1
+⎣ 0 0 0 ⎦ × exp(iω0 τk )⎦ ⎝ α1 ⎠
q1 p1 (1 − m 3 )Y∗ q2 p2 (1 − m 4 )Y∗ q1 p1 (1 − m 3 )S∗ + q2 (1 − m 4 ) p2 I∗ α2
⎛ ⎞
1
= iω0 ⎝ α1 ⎠ ,
α2
⎡⎡ ⎤ ⎛ ⎞
− r kS∗ − rk + β S∗ (− p1 (1 − m 3 )S∗ ) 1
⎣⎣ β I∗ (β S∗ − p2 (1 − m 4 )Y∗ − c − d2 − m 2 ) − p2 (1 − m 4 )I∗ ⎦ × ⎝ α1 ⎠ +
0 0 −d3 − d4 α2
⎡ ⎤ ⎤⎛ ⎞
0 0 0 exp(−iω0 τk )
⎣ 0 0 0 ⎦ × exp(iω0 τk )⎦ ⎝ α1 exp(−iω0 τk ) ⎠
q1 p1 (1 − m 3 )Y∗ q2 p2 (1 − m 4 )Y∗ q1 p1 (1 − m 3 )S∗ + q2 p2 (1 − m 4 )I∗ α2 exp(−iω0 τk )
⎛ ⎞
1
= iω0 α1 ⎠ .
⎝
α2
− p2 (1−m 4 )I∗ (iω0 + r Sk∗ )− p1 (1−m 3 )β S∗ I∗
We obtain, α1 = p2 (1−m 4 )( rk +β)S∗ I∗ − p1 (1−m 3 )S∗ (iω0 −β S∗ +c+d2 +m 2 + p2 (1−m 4 )Y∗ )
,
q1 p1 (1−m 3 )Y∗ exp(−iω0 τk +q2 p2 (1−m 4 )Y∗ exp(−iω0 τk )
α2 = iω0 +d3 +d4 −q1 p1 (1−m 3 )S∗ +q2 p2 (1−m 4 )I∗ .
Next, suppose that q∗ (s) = B(1, α1∗ , α2∗ ) exp(iω0 τk s) is the eigen vector of A∗
corresponding to −iω0 τ k similarly, we have,
− p1 (1−m 3 )( rk +β)S∗ − p2 (1−m 4 )(iω0 − r Sk∗ )
α1∗ = p2 (1−m 4 )β I∗ − p1 (1−m 3 )(iω0 +β S∗ −c−d2 −m 2 − p2 (1−m 4 )Y∗ ) ,
− p1 (1−m 3 )S∗ − p2 (1−m 4 )I∗ α1∗
α2∗ = −iω0 +d3 +d4 −(q1 p1 (1−m 3 )S∗ +q2 p2 (1−m 4 )I∗ ) exp(−iω0 τk ) ,
where B has to be calculated. We have the conditions
< q ∗ , q(θ) >= 1
which may be verified.
< q ∗ , q(θ) >= 0
0 ∞
< q ∗ , q(θ) >= q ∗ (0)q(0) − ∗ T (ξ − θ)dη(θ)q(ξ)dξ
−1 ξ=0 q
0 ∞
= B(1, α1 ∗ , α2 ∗ )(1, α1 , α2 )T − B(1, α1 ∗ , α2 ∗ ) exp(−iω0 τk (ξ − θ))
−1 ξ=0
0
dη(θ) × (1, α1 , α2 )T exp(iω0 τk ξ)dξ = B{1 + α1 α1 ∗ + α2 α2 ∗ − −1 (1, α1 ∗ , α2 ∗ )
exp(iω0 τk )dη(θ)(1, α1 , α2 )T } = B{1 + α1 α1 ∗ + α2 α2 ∗ + τk [q2 p2 (1 − m 4 )α2 ∗ Y∗ +
q2 p2 (1 − m 4 )α1 α2 ∗ Y∗ + (q1 p1 (1 − m 3 )S∗ + q2 p2 (1 − m 4 )I∗ )α2 α2 ∗ ] exp(−iω0 τk )}
which gives:
1
B = {1 + α1 α1 ∗ + α2 α2 ∗ + τk [q2 p2 (1 − m 4 )α2 ∗ Y∗ + q2 p2 (1 − m 4 )α1 α2 ∗ Y∗
+(q1 p1 (1 − m 3 )S∗ + q2 p2 (1 − m 4 )I∗ )α2 α2 ∗ ] exp(−iω0 τk )}.
5.1 Stability of Bifurcated Periodic Solutions

We first compute the coordinates to describe the Center Manifold C0 at μ = 0. Let
u t be the solution of u̇(t) = L μ (u t ) + F(μ, μt ) and define, z(t) =< q ∗ , u t >, q ∗
being the eigenvalue of A∗ . And W (t, θ) = u t (θ) − 2Re{z(t)q(θ)} on the Center
Manifold C0 , we have,W (t, θ) = W (z(t), z(t), θ), where,
2 2 z3
W (z, z, θ) = W20 (θ) z2 + W02 (θ) z2 + W11 (θ)zz + W30 3 + ···
In fact, z and z are local coordinates for the Center Manifold C0 in the direction
of q ∗ and q ∗ respectively. The existence of C0 will provide an opportunity to reduce
the system u̇(t) = L μ (u t ) + F(μ, μt ) into an Ordinary Differential Equation ODE(
in a single complex variable z) on C0 which is very interesting. u t is the solution of
system under consideration. u t ∈ C0 , we have
ż(t) = < q ∗ , u˙t >
= < q ∗ , A(u t ) + R(u t ) >
= < q ∗ , A(u t ) > + < q ∗ , R(u t ) >

= < A∗ q ∗ , (u t ) > + < q ∗ , R(u t ) >
= iω0 τ z + q ∗ · F(0, W (t, 0) + 2Re[z(t)q(θ)])
2 2
Rewrite it as ż(t) = iω0 τ z + g(z, z), where g(z, z) = g20 (θ) z2 + g02 (θ) z2 +
2
g11 (θ)zz + g21 zz3 + · · ·
The above two equations give us g(z, z) = (q ∗ )T F(z, z)
⎛ ⎞
− rk u 21 (t) − ( rk + β)u 1 (t)u 2 (t) − p1 (1 − m 3 )u 1 (t)u 3 (t)
= τk B(1, α1∗ , α2∗ ) ⎝ βu 1 (t)u 2 (t) − p2 (1 − m 4 )u 2 (t)u 3 (t) ⎠
p1 q1 (1 − m 3 )u 1 (t − 1)(t)u 3 (t − 1) + p2 q2 (1 − m 4 )u 1 (t − 1)u 2 (t − 1)
Further,
u(t + θ) = W (t, θ) + z(t)q(θ) + z(t)q(θ),
u 1 (t) = z + z + W (1) (t, 0),
u 2 (t) = α1 z + α1 z + W (2) (t, 0),
u 3 (t) = α2 z + α2 z + W (3) (t, 0),
u 1 (t − 1) = z exp(−iω0 τk ) + z exp(iω0 τk ) + W (1) (t, −1),
u 2 (t − 1) = α1 z exp(−iω0 τk ) + α1 z exp(iω0 τk ) + W (2) (t, −1),
u 3 (t − 1) = α2 z exp(−iω0 τk ) + α2 z exp(iω0 τk ) + W (3) (t, −1). Hence, g(z, z) =
τk B[− rk u 21 (t) − ( rk + β)u 1 (t)u 2 (t) − p1 (1 − m 3 )u 1 (t)u 3 (t) + α1 ∗ {βu 1 (t)u 2 (t) −
p2 (1 − m 4 )u 2 (t)u 3 (t)} + α2 ∗ { p1 (1 − m 3 )q1 u 1 (t − 1)(t)u 3 (t − 1) + p2 (1 − m 4 )q2 u 1
(t − 1)u 2 (t − 1)}].
Putting the values of u 1 , u 2 , u 3 , u 1 (t − 1), u 2 (t − 1), u 3 (t − 1) etc. in g(z, z), we
get
g(z, z) = τk B − rk [z + z + W (1) (t, 0)]2 − ( rk + β)[z + z + W (1) (t, 0)][α1 z +
α1 z + W (2) (t, 0)] − p1 [z + z + W (1) (t, 0)][α2 z + α2 z + W (3) (t, 0)] + α1 ∗ (β[z + z +
W (1) (t, 0)][α1 z+α1 z+W (2) (t, 0)]− p2 (1−m 4 )[α1 z+α1 z+W (2) (t, 0)][α2 z+α2 z+
W (3) (t, 0)])+α2 ∗ ( p1 (1−m 3 )q1 [z exp(−iω0 τk )+ z exp(iω0 τk )+ W (1) (t, −1)][α2 z
exp(−iω0 τk ) + α2 z exp(iω0 τk ) + W (3) (t, −1)] + p2 (1 − m 4 )q2 [z exp(−iω0 τk ) +
z exp(iω0 τk ) + W (1) (t, −1)][α1 z exp(−iω0 τk ) + α1 z exp(iω0 τk ) + W (2) (t, −1)]) .
From this equation we can find the values of the coefficients g20 (θ), g02 (θ),g11 (θ),
g21 (θ), etc., by comparing the same powers of z, we have
g20 = 2τk B{− rk − ( rk + β)α1 − p1 (1 − m 3 )α2 + βα1 ∗ α1 − α1 ∗ α1 α2 p2
(1 − m 4 ) + α2 ∗ ( p1 (1 − m 3 )q1 α2 + p2 (1 − m 4 )q2 α1 ) exp(−2iω0 τk )},
g11 = τk B(− 2rk + (α1 ∗ )β + α2 ∗ p2 (1 − m 4 )q2 − rk + β)(α1 + α1 ) + (α2 ∗ p1
(1 − m 3 )q1 − p1 (1 − m 3 ))(α2 + α2 ) − α1 ∗ p2 (α2 α1 + α1 α2 )),
g02 = 2τk B{− rk − ( rk + β)α1 − p1 (1 − m 3 )α2 + βα1 ∗ α1 − α1 ∗ α1 α2 p2 (1 −
m 4 ) + α2 ∗ ( p1 (1 − m 3 )q1 α2 + p2 q2 (1 − m 4 )α1 ) exp(2iω0 τk )},
(1) (1) (2) (1)
g21 = 2τk B(− rk (2W11 (0) + W20 (0)) − ( rk + β)[W11 (0) + α1 W11 (0) +
(1) (1) (3) (3) (1)
2 α1 W20 (0) + 2 W20 (0)] − p1 (1 − m 3 )[W11 (0) + 2 W20 (0) + α2 W11 (0) +
1 1 1
1 (1) ∗ (2) (2) (1) (1)
2 W20 (0)] + α1 β[W11 (0) + 2 W20 (0) + α1 W11 (0) + 2 α1 W20 (0)]
1 1
These coefficients are used in calculating C0 , etc. Now we need to calculate

W20 (θ) and W11 (θ). Now u̇ t = A(μ)u t + R(μ)u t and z(t) = < q ∗ , u t >,W (t, θ) =
u t (θ) − 2Re{z(t)q(θ)} gives us
Ẇ = u˙t − zq − żq
AW − 2Req ∗ (0)F0 q(θ), for − 1 ≤ θ < 0
=
AW − 2Req ∗ (0)F0 q(θ) + F0 , forθ = 0.
Rewrite the above equation as, Ẇ = AW + H (z, z, θ), where,
2 2 2
H (z, z, θ) = H20 (θ) z2 + H11 (θ)zz + H02 (θ) z2 + H21 (θ) z2z + · · · Near to the
origin on C0 , Ẇ = Wz ż + Wz z (A − 2iω0 τk )W20 (θ) = −H20 (θ) and AW11 (θ) =
−H11 (θ) hence for −1 ≤ θ < 0 we have, H (z, z, θ) = −2Re(q ∗ (0)F0 q(θ)) =
−g(z, z)q(θ) − g(z, z)q(θ), by comparing the coefficients of z, we have H20 (θ) =
−g20 q(θ) − g 02 q(θ) and H11 (θ) = −g11 q(θ) − g 11 q(θ),
Ẇ20 (θ) = 2iω0 τk W20 (θ) + g20 q(θ) + g 02 q(θ),
Ẇ11 (θ) = g11 q(θ) + g 11 q(θ).
Integrating, we have
ig 20 q(0)
W20 (θ) = ωig020
τk q(0) exp(iω0 τk θ) + 3ω0 τk exp(−iω0 τk θ) + E 1 exp(2iω0 τk θ),
W11 (θ) = iωg21
0 τk
q(0) exp(iω0 τk θ) + igω110q(0)
τk exp(−iω0 τk θ) + E 2 .
where E 1 and E 2 are to be determined. From definitions of A and (A − 2iω0 τk )
W20 (θ) = −H20 (θ) 0
(A − 2iω0 τk )W20 (θ) = −H20 (θ) gives us −1 dη(θ)W20 (θ) = 2iω0 τk W20 (0) −
H20 (0) which gives us H20 (0) = −g20 q(0) − g 02 q(0)
− rk − ( rk + β)α1 − p1 (1 − m 3 )α2
+2τk βα1 − p2 (1 − m 4 )α1 α2 .
(q1 p1 (1 − m 3 )α2 + q2 (1 − m 4 ) p2 α1 ) exp(−2iω0 τk )
0
Now, (iω0 τk I − −1 exp(iω0 τk θ)dη(θ))q(0) = 0

0
−iω0 τk I − −1 exp(−iω0 τk θ)dη(θ) q(0) = 0

0
And we have 2iω0 τk I − −1 exp(iω0 τk θ)dη(θ)
⎛ ⎞
− rk − ( rk + β)α1 − p1 (1 − m 3 )α2
E 1 = 2τk ⎝ βα1 − p2 (1 − m 4 )α1 α2 ⎠ , which leads
(q1 p1 (1 − m 3 )α2 + q2 p2 (1 − m 4 )α1 ) exp(−2iω0 τk )
to
⎛ ⎞
2iω0 + r Sk∗ S∗ ( rk +β) p1 (1−m 3 )S∗
⎝ −β I∗ 2iω0 −β S∗ +c+d2 +m 2 + p2 (1−m 4 )Y∗ p2 (1−m 4 )I∗ ⎠ ×E 1
−q1 (1−m 3 ) p1 Y∗ exp(−2iω0 τk ) −q2 p2 (1−m 4 )Y∗ exp(−2iω0 τk ) 2iω0 +d3 +d4 −(q1 p1 (1−m 3 )S∗
⎛ +q2 p2 (1−m 4⎞ )I∗ ) exp(−2iω0 τk )
− rk − ( rk + β)α1 − p1 (1 − m 3 )α2
= 2⎝ βα1 − p2 (1 − m 4 )α1 α2 ⎠.
(q1 p1 (1 − m 3 )α2 + q2 p2 (1 − m 4 )α1 ) exp(−2iω0 τk )
0
E 1 can be calculated from this equation. Now, −1 dη(θ)W11 (θ) = −H11 (0)
⎛ ⎞
− rk − ( rk + β)Re(α1 ) − p1 (1 − m 3 )Re(α2 )
⎜ β Re(α1 ) − p2 (1 − m 4 )Re(α1 α2 ) ⎟
H11 (0) = −g11 q(0) − g 11 q(0) + 2τk ⎝ ⎠,
(q1 p1 (1 − m 3 )Re(α2 ) + q2 p2 (1 − m 3 )Re(α1 )
⎛ ⎞
r S∗
S∗ ( rk + β) p1 (1 − m 3 )S∗
⎜ k ⎟
⎝ −β I∗ −β S∗ + c + d2 + m 2 + p2 (1 − m 4 )Y∗ p2 (1 − m 4 )I∗ ⎠
−q1 p1 (1 − m 3 )Y∗ −q2 p2 (1 − m 4 )Y∗ d3 + d4 − (q1 p1 (1 − m 3 )S∗ + q2 p2 (1 − m 4 )I∗ )
⎛ ⎞
− rk − ( rk + β)Re(α1 ) − p1 (1 − m 3 )Re(α2 )
⎜ ⎟
× E2 = 2 ⎝ β Re(α1 ) − p2 (1 − m 4 )Re(α1 α2 ) ⎠.
(q1 p1 (1 − m 3 )Re(α2 ) + q2 p2 (1 − m 4 )Re(α1 )
E 2 can be obtained from this equation. By putting values of E 1 and E 2 we can

obtain W20 (θ) and W11 (θ) and hence g20 ,g11 , g02 , g21 etc. Hence as stated in [5, 10],
we can obtain the following values;
⎧
c1 (0) = 2ωi0 τk (g11 g20 − 2 | g11 |2 − |g023 | ) + g221 ,
2
⎪
⎪
⎪
⎪
⎨μ = − Re(c 1 (0)) ,
2
Re(λ (τk )) (10)
⎪
⎪β = 2Re(c (0)),
⎪
⎪ 2 1
⎩
T2 = − ω01τk [I m(c1 (0)) + μ2 I m(λ (τk ))],
which determine the direction and stability of the model with delay at the critical
value τk . Now, we state the following theorem due to [5, 10, 21], which is the main
result of this section:
Theorem 3 (i) The sign of μ2 determined the direction of Hopf bifurcation: if μ2 >
0(μ2 < 0), then the Hopf bifurcation is supercritical (subcritical).
(ii)The stability of bifurcated periodic solutions is determined by β2 : the periodic
solutions are stable if β2 < 0 and unstable if β2 > 0.
(iii)The period of bifurcated periodic solutions is determined by T2 : the period
increases if T2 > 0 and decreases if T2 < 0.
From part (i) of this theorem, it is clear that Hopf bifurcation is supercritical if

either Re(c1 (0)) < 0 or Re(λ (τk )) < 0. Similarly, Hopf bifurcation is subcritical

if Re(λ (τk )) > 0 and Re(c1 (0)) > 0.
6 Numerical Simulation
In this section, we consider a hypothetical set of parameters P1 = {r = 0.8,

k = 1, β = 1, p1 = 0.12, p2 = 6, m 1 = 0.02, m 2 = 0.06, m 3 = 0.5, m 4 =
0.2, d2 = 0.05, d3 = 0.6, d4 = 0.5, c = 0.025, q1 = 0.75, q2 = 0.75}. We will
focus on positive equilibrium. Calculation shows that S = .2339, I = .2749, Y=
.0487, thus model has the positive equilibrium E 5 (.2339, .2749, 0.0487). Also, =
−0.5550, = 1.8, = −4.860, = −0.380, A1 = .3698, A2 = .1647, A3 =
.0217, therefore
S + + = −0.0335 < 0 and A1 A2 + A3 = .0826 > 0,
I + Y
Computed solution for the Phase plane plot of susceptible prey (S) Phase plane plot of infected prey (I)
interval [0,1000] and infected prey(I) and predator (Y)
1 1 0.7
S
I 0.6
0.8 Y 0.8
0.5
0.6 0.6 0.4
S,I,Y
Y
I
0.4 0.4 0.3
0.2
0.2 0.2
0.1
0 0 0
0 200 400 600 800 1000 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
time t S I
Phase plane plot of Prey (S)
and Predator (Y) 3D Plot
0.2
0.8
0.15
0.6
0.4
Y
Y
0.1
0.2
0.05 0
1
1
0.5 0.5
0
0 0.2 0.4 0.6 0.8 1 I 0 0
S
S
Fig. 1 Solution of system (4) for initial function S(0) = 0.6, I (0) = 0.2, Y (0) = 0.2 with
parameter set P1 , τ = 15.14 < τ0 , the positive equilibrium point is stable
hence E 5 (.2339, .2749, 0.0487) is stable. Indeed, we also have the jacobian matrix
at E 5 ;
⎛ ⎞
−0.1349 −0.4210 −0.0140
⎝ 0.2749 −0.1349 −1.3195 ⎠ ,
0.0022 0.0029 −0.1
this has the characteristics equation λ3 + 0.3698λ2 + 0.1647λ + 0.0217. It has three
roots, viz.,
⎧
⎨ −0.1020 + 0.3471i,
−0.1020 − 0.3471i,
⎩
−0.1658,
hence E 5 (0.2339, 0.2749, 0.0487) is stable. It is also calculated that n 0 = −0.1941,

n 2 = −1.1, n 1 = 0.6895, m 0 = 0.2158, m 1 = −0.5248, m 2 = 1.4698. Therefore,
p0 = 3.1633, q0 = −0.4073, r0 = 0.0089, h(z) = z 3 + 3.1633z 2 − 0.4073z +
0.0089, p02 − 3q0 = 11.2284 > 0 and z 1∗ = 0.063. From this h(z 1∗ ) = 11.20998528,
hence E ∗ is stable. Further ω0 = 0.6382 and τ0 = 33.14. Thus, Hopf bifurcation
occurs as the τ passes through τ0 which is depicted by numerical simulation in
Figs. 1 and 2.
Computed solution for the Phase plane plot of susceptible prey (S) Phase plane plot of infected prey (I)
interval [0,1000] and infected prey(I) and predator (Y)
1.2 1.2 0.7
S
1 I 1 0.6
Y
0.8 0.8 0.5
S,I,Y
0.6 0.6 0.4
Y
I
0.4 0.4 0.3
0.2 0.2 0.2
0 0 0.1
−0.2 −0.2 0
0 200 400 600 800 1000 0 0.2 0.4 0.6 0.8 1 −0.2 0 0.2 0.4 0.6 0.8 1 1.2
time t S I
Phase plane plot of Prey (S)

and Predator (Y) 3D Plot
0.2
0.8
0.15
0.6
0.4
Y
Y
0.1
0.2
0.05 0
1
0.5 1
0 0 0.5
0 0.2 0.4 0.6 0.8 1
I
−0.5 0 S
S
Fig. 2 Solution of system (4) for initial function S(0) = 0.6, I (0) = 0.2, Y (0) = 0.2 with
parameter set P1 , τ = 60.30 > τ0 , the positive equilibrium point is unstable
7 Discussion
In this paper, we have considered a delayed prey–predator system with infection.

Migration has been allowed among prey population only. It is also considered that
prey population has self-defence in the form of prey refuge. This decreases the
availability of prey population for predation to predators. For instance, only (1−m 3 )S
of sound prey are available for predation. Similarly, (1 − m 4 )I of infected prey are
available for predation. Stability results have been investigated.
Similar to the study of [10], in this paper the time delay τ is the gestation period of
predator. In our analysis this is found to be the bifurcation parameter. It is proved that
beyond some specific value of τ , Hopf-bifurcation occurs. The direction of Hopf-
bifurcation and stability of bifurcated periodic solutions have been derived using the
central manifold reduction technique and normal form theory.
In this paper, bifurcation of predator into two parts, viz., healthy predator and
infected predator has been ignored. The same may be done in the future. Further, for
simplification, parameters are taken as time independent. In real-life the parameters
are time dependent, this may also considered in the future.
The main issue in applied mathematical modeling is to identify the real parame-
ters. The present study is not a case study, hence real parameters are not available.
Hence, the main scope of this study is to study a real eco-system and to identify the
real/experimental parameters.
References
1. Lotka, A.: Elements of Physical Biology. Williams and Wilkins, Baltimore (1925)
2. Bate, A.M., Hilker, F.M.: Predator-prey oscillations can shift when disease become endemic.
J. Theor. Biol. 316, 1–8 (2013)
3. Pal, A.K., Samanta, G.P.: Stability analysis of an eco-epidimiolgical model incorporating a
prey refuge. Nonlinear Anal.: Model. Control 15, 473–491 (2010)
4. Mukhopadhyaya, B., Bhattacharyya, B.: Dynamics of a delay-diffusion prey-predator model
with disease in the prey. J. Appl. Math Comput. 17(12), 361–377 (2005)
5. Celik Karaaslanli, C.: Bifurcation analysis and its applications, Chapter 1. In: Andriychuk, M.
(ed.) Numerical Simulation: From Theory to Industry. INTECH, Vienna (2012)
6. Greenhalgh, D., Haque, M.: A predator-prey model with disease in the prey species only. Math.
Methods Appl. Sci. 30, 911–929 (2007)
7. Hilker, F.M., Schmitz, K.: Disease-induced stabilization of predator prey oscillations. J. Theor.
Biol. 255, 299–306 (2008)
8. Wang, F., et al.: Stability and bifurcation of a stage-structured predator-prey model with both
discete and distributed delays. Chaos Solitons Fractals 46, 19–27 (2013)
9. Faina, S., et al.: Role of prey dispersal and refuges on predator-prey dynamics. SIAM J. Appl.
Math. 70(6), 1821–1839 (2010)
10. Hu, G.-P., Li, X.-L.: Stability and Hopf bifurcation for a delayed predator-prey model with
disease in the prey. Chaos Solitons Fractals 45, 229–237 (2012)
11. Sun, G.-Q., et al.: Dynamical complexity of a spatial predator-prey model with migration. Ecol.
Modell. 219, 248–255 (2008)
12. Chattopadhyay, J., Arino, O.: A predator-prey model with disease in prey. Non-Linear Anal.
36, 747–766 (1999)
13. Berthier, K., et al.: Dynamics of a feline virus with two transmission Medels within exponen-
tially growing host populations. Proc. R. Soc. Lond. 267, 2049–2056 (2000)
14. Maoxing, L., et al.: An Impulsive predator-prey model with communicable disease in the prey
species only. Non-Linear Anal.: Real World Appl. 10, 3098–3111 (2009)
15. Haque, M., et al.: An eco-epidemiological predator-prey model with standard disease incidence.
Math. Methods Appl. Sci. (2008). doi:10.1002/mma.1071
16. Haque, M., Venturino, E.: Increase of the prey may decrease the healthy prtedator population
in presence of a disease in predator. HERMISS 7, 38–59 (2006)
17. Naji, R.K.: The dynamics of prey-predator model with disease in the prey. J. Math. Comput.
Sci. 2(4), 1052–1072 (2012)
18. Samanta, G.P.: Analysis of a delay nonautonomous predator-prey system with disease in the
prey. Non-Linear Analy.: Model. Control 15(1), 97–108 (2010)
19. Sinha, S., et al.: Modelling a predator-prey system with infected prey in polluted environment.
Appl. Math. Model. 34, 1861–1872 (2010)
20. Bhattacharyya, R., Mukhopadhyay, B.: Spatial dynamics of nonlinear prey-predator models
with prey migration and predator switching. Ecol. Complex. 3, 160–169 (2006)
21. Jana, S., Kar, T.K.: Modeling and analysis of a prey-predator system with disease in the prey.
Chaos Solitons Fractals 47, 42–53 (2013)
22. Kar, T.K.: Stability Analysis of a prey-predator model incorporating a prey refuge. Commun.
Nonlinear Sci. Numer. Simul. 10, 681–691 (2005)
23. Volterra, V.: Variazioni e fluttauazionidel numero d individui in species animals conviventii.
Mem Acd. Linciei 2, 31–33 (1926)
24. Krivan, V.: On the Gauss predator-prey model with a refuge: a fresh look at the history. J.
Theor. Biol. 274, 67–73 (2011)
25. Kermack, W.O., Mckendrick, A.G.: Contributons to the mathematical theory of epidimics, Part
1. Proc. R. Soc. A 115(5), 700–721 (1927)
26. Yu, X., Sun, F.: Global dynamics of a predator-prey model incorporating a constant prey-refuge.
Electron. J. Differ. Equ. 4, 1–8 (2013)
27. Xiao, Y., Chen, L.: Modeling and analysis of a predator-prey model with disease in the prey.
Math. Biosci. 171, 59–82 (2001)
28. Xiao, Y., Chen, L.: A ratio-dependent predator-prey model with disease in the prey. Appl. Math.
Comput. 131, 397–414 (2002)
29. Ma, Z., et al.: Effect of prey refuges on a predator-prey model with a class of functional
responses: the role of refuges. Math. Biosci. 218, 73–79 (2009)
Stability Analysis of an Integro Differential
Equation Model of Ring Neural Network
with Delay
Swati Tyagi, Syed Abbas and Rajendra K. Ray
Abstract In this paper we present and study a ring neural network model with delays.
We study existence and uniqueness of equilibrium point and global stability of the
model system. At the end few examples have been given to illustrate the analytical
findings.
Keywords Ring network · Neural network · Time delay · Equilibrium point ·

Existence and uniqueness · Global asymptotic stability
AMS Subject Classification: 92B20 · 34D23 · 34K20 · 35A35

The term neural networks denote the collection of interconnected, interacting neu-
rons, which can be biological/artificial. A system of connected nodes constitutes an
artificial neural network. Arranging the nodes in different configurations yields dis-
tinct artificial neural networks with characteristic properties. For performing parallel
computation the model of neural networks are also very much promising. Several
authors studied the dynamics of neural network [1–7] theoretically and numerically.
The integration and communication delays are present everywhere in all physical
systems, due to which several questions arise in our mind about their effects on the
dynamic and various other properties of neural networks. While implementing artifi-
cial neural systems, due to the finite switching speed of neurons and amplifiers, time
delays are ineluctable and so we cannot avoid them. Over some recent years, these
effects of delays on almost all physical dynamical networks and the stability analy-
sis of time delayed system have received a remarkable attention. Recurrent neural
networks have been extensively used to study dynamical varying data and have been
used in many practical applications [8]. In 1994, Baldi and Atiya studied the effects
of delays on neural dynamics by using additive neural network model given as [9]:
S. Tyagi (B) · S. Abbas · R.K. Ray

School of Basic Sciences, Indian Institute of Technology, Mandi 175001, H.P., India
e-mail: styagi99@gmail.com
S. Abbas
e-mail: sabbas.iitk@gmail.com

DOI 10.1007/978-81-322-2485-3_3
38 S. Tyagi et al.
du i ui
=− + Ti j f j (u j ) + Ii (i = 1, 2, ..., n),
dt τi
j
Ti j are synaptic connection strengths, τi are time constants, Ii are external inputs, f i
are input/output transfer functions. Recently Feng and Plamondon [10] studied the
following recurrent neural network. They considered a ring of neurons connected
cyclically with delayed interactions, which was modelled as:
du i (t) 1
= − u i (t) + wi,i−1 f i−1 u i−1 (t − τi,i−1 ) (i mod n), (1)
dt ri
They discussed the conditions under which a ring network could be exploited as an
oscillatory pattern generator. Using Chafee’s theory, they guaranteed the existence of
permanent oscillations in a time delayed neural network model. Motivated by their
work, in the present paper, we will be particularly studying stability of the following
delayed neural ring network model,

du i (t) 1 t
= − u i (t) + wi,i−1 f i−1 t, u i−1 (t − τi,i−1 ), ki−1 (t − s)u i−1 (s)ds
dt ri −∞
(i mod n), (2)
with the passive decay rates given by − r1i > 0, τi,i−1 ≥ 0 are the time delays
present in the system, wi,i−1 = 0 (i = 1, 2, ..., n) are the weights. By taking the
transformation u = (u 1 , u 2 , ..., u n )T the system (2) can be rewritten as:
t
du(t)
= Ru(t) + W f t, u(t − τ ), K (t − s)u(s)ds , (3)
dt −∞
where R is a diagonal matrix given by, R = diag(− r11 , − r12 , ..., − r1n ), and the weight
matrix W is given by,
⎛ ⎞
0 0 ··· ··· ··· w1,n
⎜w2,1 0 ··· ··· 0 0 ⎟
⎜ ⎟
W =⎜
⎜ ··· ··· ··· ··· ··· ··· ⎟⎟,
⎝ 0 0 · · · wn−1,n−2 0 0 ⎠
0 0 ··· ··· wn,n−1 0
The kernel matrix K is given by,

⎛ ⎞
0 0 ··· ··· 0 kn (t − s)
⎜k1 (t − s) 0 ··· ··· 0 0 ⎟
⎜ ⎟
K (t − s) = ⎜
⎜ ··· ··· ··· ··· ··· ··· ⎟ ⎟,
⎝ 0 0 · · · kn−2 (t − s) 0 0 ⎠
0 0 ··· ··· kn−1 (t − s) 0
Stability Analysis of an Integro Differential Equation … 39
and the activation functions are given by,

t
f (t, u(t − τ ), K (t − s)u(s)ds)
−∞
t
= ( f 1 (t, u 1 (t − τ2,1 ), k1 (t − s)u 1 (s)ds),
−∞
t
f 2 (t, u 2 (t − τ3,2 ), k2 (t − s)u 2 (s)ds),
−∞
t
...., f n−1 (t, u n−1 (t − τn,n−1 ), kn−1 (t − s)u n−1 (s)ds),
−∞
t
f n (t, u n (t − τ1,n ), kn (t − s)u n (s)ds))T .
−∞
For any matrix A = (ai j )n×n , we denote |A| = (|ai j |)n×n and A > 0 denotes √ A to
be a positive definite matrix, A = max{λ} : where λ is an eigen value of A T A
n 1
and for x = (x1 , x2 , ..., xn )T , x = ( i=1 xi2 ) 2 . The remaining of this paper is
organized as follows. In Sect. 1, some definitions and lemmas are given, which help
us in deriving the existence and uniqueness of the equilibrium point, which is derived
in Sect. 3. Global stability of the solution of the system (2) is studied in Sect. 4, and
in Sect. 5, some examples have been given to validate the theoretical findings.
Remark 1 In this paper, the model that we are studying is a generalization of

[10, 11], and after using suitable transformations and assumptions, it becomes similar
to [10, 11].
1 Preliminaries
Assumptions
1. The activation functions f i (u i )(i = 1, 2, ..., n) are bounded, continuous and non
linear Lipschitz functions:
f : I × X ×C → X
f (t, x, φ) − f (t, y, ψ) ≤ c1 x − ya + c2 φ − ψb ,

or f (t, x, φ) − f (t, y, ψ) ≤ ci x − y, (4)
t
where x = u(t − τ ), y = v(t − τ ), φ = −∞ ki−1 (t − s)u i−1 (s)ds, ψ =
t
−∞ ki−1 (t − s)vi−1 (s)ds, x, y ∈ R , c1 , c2 > 0, ci = max ci (i = 1, 2, ..., n)
n
40 S. Tyagi et al.
and f i (0) = 0 (i = 1, 2, ..., n), where a norm is defined on C([−r, 0], X ) and
b norm is defined on ((−∞, 0], X ).
Remark 2 To prove result (4), we have,
t t
φ − ψ = ki−1 (t − s)u i−1 (s)ds − ki−1 (t − s)vi−1 (s)ds,
−∞ −∞
t
= ki−1 (t − s)[u i−1 (s) − vi−1 (s)]ds,
−∞
t
≤ ki−1 (t − s)dsu i−1 − vi−1 ,
−∞
∞
= ki−1 (s)dsu i−1 − vi−1 , k ∈ L 1 (0, ∞),
0
≤ c2 u i−1 − vi−1 .
Since we have assumed that k ∈ L 1 (0, ∞), so it is bounded by some constant, say
c2 , and hence we get the above result.
2. In the neighbourhood of zero point, the function f i (u i ) (i = 1, 2, ..., n) are
differentiable.
Definition 1 [10] A continuous mapping H : R n → R n is called a homeomor-
phism, if it satisfies the following properties:
1. H is bijection (one-one and onto).
2. The inverse map H −1 is also continuous.
Lemma 1 [10]: A continuous mapping H (x) : R n → R n is a homeomorphism if

it satisfies the following properties:
1. H(x) is injective.
1/2
n
2. limx→∞ H (x) → ∞, where x = 2
i=1 x i .
2 Existence and Uniqueness
Lemma 2 Suppose that there is a positive diagonal matrix D = diag(d1 , d2 ,

..., dn ), di > 0 (i = 1, 2, ..., n) such that,
−1
r1 d1 |w1,n |ln dn−1 < 1 and ri di |wi,i−1 |li−1 di−1 < 1, (5)

where li s are constants with 0 < li ≤ ci (i = 2, 3, ..., n), then the system (2)
has a unique equilibrium point v∗ , where v∗ = [v1∗ , v2∗ , ..., vn∗ ]T and is equal to
(0, 0, ..., 0)T.
Proof Let we define a map corresponding to (2):
H (u) = (H1 (u), H2 (u), ..., Hn (u))T : R n → R n, (6)
where,
t
1
H1 (u) = − u 1 + w1,n f n t, u n , kn (t − s)u n (s)ds ,
r1 −∞
t
1
Hi (u) = − u i + wi,i−1 f i−1 t, u i−1 , ki−1 (t − s)u i−1 (s)ds ,
ri −∞
(i = 2, 3, ..., n). (7)
Let if v∗ be an equilibrium point of (2), then it must satisfy H (v∗ ) = 0. Now we show
the existence and uniqueness of the equilibrium point v∗ . For this, we need to show
that H (v) is an homeomorphism. First we show that H (v) is injective on R n . Assume
on the contrary that, there exist two vectors u, v ∈ R n where u = [u 1 , u 2 , ..., u n ]T
and v = [v1 , v2 , ..., vn ]T with u = v such that H (u) = H (v), then we get,
1
(v1 − u 1 ) = w1,n [ f n (t, vn , ψn ) − f n (t, u n , φn )],
r1
1
(vi − u i ) = wi,i−1 [ f i−1 (t, vi−1 , ψi−1 ) − f i−1 (t, u i−1 , φi−1 )],
ri
(i = 2, 3, ..., n) (8)
t t
where φi = −∞ ki (t − s)u i (s)ds and ψi = −∞ ki (t − s)vi (s)ds, (i = 1, 2, ..., n).
Since we have ri > 0 (i = 1, 2, ..., n), we get,
1
|v1 − u 1 | = w1,n | f n (t, vn , ψn ) − f n (t, u n , φn )|,
r1
1
|vi − u i | = wi,i−1 | f i−1 (t, vi−1 , ψi−1 ) − f i−1 (t, u i−1 , φi−1 )|
ri
(i = 2, 3, ..., n). (9)
From (4), there exist 0 ≤ li ≤ ci such that

f n (t, vn , φ) − f n (t, u n , ψ) = ln v − u + φ − ψ ,

f i−1 (t, vi−1 , φ) − f i−1 (t, u i−1 , ψ) = li−1 vi−1 − u i−1 + φ − ψ
(i = 2, 3, ..., n), (10)
42 S. Tyagi et al.
Using (10) in (9), we get,

v1 − u 1 = r1 |w1,n |ln vn − u n + φ − ψ ,

= r1 |w1,n |ln vn − u n + Mn vn − u n ,
= r1 |w1,n |ln (1 + Mn )vn − u n ,
= r1 |w1,n |ln vn − u n , (where ln = ln (1 + Mn )),
= r1 |w1,n |ln · rn |wn,n−1 |ln−1 vn−1 − u n−1 ,
= · · · = r1 |w1,n |ln · rn |wn,n−1 |ln−1 · rn−1 |wn−1,n−2 |ln−2
· · · r2 |w2,1 |l1 v1 − u 1 ,
r1 d1 |w1,n |ln rn dn |wn,n−1 |ln−1
= ·
dn dn−1
rn−1 dn−1 |wn−1,n−2 |ln−2 r2 d2 |w2,1 |l1
· ··· v1 − u 1
dn−2 d1
< v1 − u 1 ,
which is a contradiction, Thus we have v1 = u 1 . Similarly for i = 2, 3, ..., n, using

similar above calculations, we have
vi − u i < vi − u i ,
which is again a contradiction. From this we can say that vi = u i (i = 2, 3, ...n).

This implies that v = u, which is a contradiction. Hence map H is injective. Also
f (u) is bounded from our assumption, which clearly implies that as u → ∞, each
Hi (u) → ∞(i = 1, 2, ..., n), So H (u) → ∞ and H is a homeomorphism,
and thus system (2) has only one equilibrium point. As we assumed f i (0) = 0,
(i = 1, 2, ..., n), we can say that (0, 0, ..., 0)T is the unique equilibrium point.
Note 1 If the activation function is monotone increasing or a monotone decreasing

bounded continuous function, then (4) changes to
0 < f i (t, u, φ) − f i (t, v, ψ) ≤ li (u − v) (u > v), i = 1, 2, ..., n, (11)
or
− li (u − v) ≤ f i (t, u, φ) − f i (t, v, ψ) < 0, (u > v), i = 1, 2, ..., n. (12)
Note 2 In justification of the above remark, for example, functions such as tanh(u),
ar ctan(u) satisfy condition (11), while for functions like 1+ex1p(u) , condition (12) is
satisfied.
Now, we have the following lemma:

Lemma 3 Assume that condition (11) holds, and
r1 w1,n ln · r2 w2,1 l1 · r3 w3,2 l2 · · · rn wn,n−1ln−1 < 1, (13)
or that the condition (12) holds, and
|r1 w1,n (−ln ) · r2 w2,1 (−l1 ) · r3 w3,2 (−l2 ) · · · rn wn,n−1 (−ln−1 )| > 1. (14)
Then system (2) has unique equilibrium point.
Proof Doing the similar calculations as done in Lemma (1), if u, v be any two
equilibrium points and condition (11) holds, then we have,
1
(v1 − u 1 ) ≤ w1,n ln (vn − u n ),
r1
1
(vi − u i ) ≤ wi,i−1 li−1 (vi−1 − u i−1 ), (i = 2, 3, ..., n). (15)
ri
Thus if v1 = u 1 , then we have
v1 − u 1 ≤ r1 w1,n ln (vn − u n ),
≤ r1 w1,n ln · rn wn,n−1ln−1 (vn−1 − u n−1 ),
≤ · · · ≤ r1 w1,n ln · rn wn,n−1ln−1
·rn−1 wn−1,n−2 ln−2 · · · r2 w2,1 l1 (v1 − u 1 ),
= r1 w1,n ln · r2 w2,1 l1
· · · rn−1 wn−1,n−2 ln−2 · rn wn,n−1ln−1 (v1 − u 1 ),
< v1 − u 1 .
This is a contradiction. So we have v1 = u 1 . Similarly for i = 2, 3, ..., n, we get,
vi − u i < vi − u i ,
which is also a contradiction. So vi = u i (i = 2, 3, ..., n). This implies that v = u,

from which we can conclude that the system (2) has a unique equilibrium point.
Similarly, If condition (12) holds, and if v1 = u 1 , then we have,
v1 − u 1 ≥ r1 w1,n (−ln )(vn − u n ),

≥ r1 w1,n (−ln ) · rn wn,n−1 (−ln−1 )(vn−1 − u n−1 ),
≥ · · · ≥ r1 w1,n (−ln ) · rn wn,n−1 (−ln−1 )
·rn−1 wn−1,n−2 (−ln−2 ) · · · r2 w2,1 (−l1 )(v1 − u 1 ).
44 S. Tyagi et al.
From (12), it implies that |v1 − u 1 | > |v1 − u 1 |, which is a contradiction. Hence
v1 = u 1 and in the similar way, we can conclude that vi = u i (i = 2, 3, ..., n). Thus
the system (2) has a unique equilibrium point.
Lemma 4 Suppose that the condition (11) (or condition (12)) holds. If there

exists some constants 0 < di ≤ ci (or − ci ≤ di < 0 (i = 1, 2, ..., n)) for
the system (3) such that the matrix R + W D is a non-singular matrix, where
D = diag(d1 , d2 , ..., dn )T , then the given system (3) has a unique equilibrium
point.
Proof Let we define a map associated with system (3) given by:
H (u) = Ru + W f (t, u, φ). (16)
Assume on the contrary that there exist two vectors v, u ∈ R n with v = u, where
v = [v1 , v2 , ..., vn ]T and u = [u 1 , u 2 , ..., u n ]T such that H (v) = H (u), then we get,
R(v − u) + W ( f (t, v, ψ) − f (t, u, φ)) = 0. (17)
Since (11) (or (12)) holds, so there exists 0 < di ≤ ki and − ki ≤ di < 0
(i = 1, 2, ..., n) such that
f i (t, vi , ψi ) − f i (t, u i , φi ) = di (vi − u i ), (i = 1, 2, ..., n). (18)
Using (18) in (17), we obtain
(R + W D)(v − u) = 0, (19)
where
⎛ ⎞
− r11 0 0 0 ··· ··· dn w1,n
⎜w d − r12 0 ··· ··· 0 ⎟
⎜ 2,1 1 0 ⎟
⎜ ⎟
⎜ ··· ··· ··· ··· ··· ··· ··· ⎟
R +WD = ⎜ ⎟.
⎜ ··· ··· ··· ··· ··· ··· ··· ⎟
⎜ ⎟
⎝ 0 0 0 · · · wn−1,n−2 dn−2 − rn−1
1
0 ⎠
0 0 0 ··· ··· wn,n−1 dn−1 − r1n
Since it is given that (R + W D) is a non-singular matrix, so det(R + W D) = 0.

Hence from (19), v = u, which is a contradiction. So H (u) is injective on R n . Also
since each f i (u i ) is a bounded continuous function, so f (u) is bounded, and it is
clear from the defined mapping that H (u) → ∞, as u → ∞. Thus H is a
homeomorphism and hence system (3) has a unique equilibrium point.
3 Global Stability
Theorem 1 Under the Assumptions (1)–(2) , there exist a unique equilibrium point
for the system (3). Furthermore, system (3) is globally asymptotically stable.
Proof For the given neural ring network model, existence and uniqueness of the
equilibrium point has already been proved in Sect. 1. In this section, We prove the
global stability of system (3) by constructing suitable Lyapunov
t function. Throughout
the calculations, for simplicity, we denote f (u, u(t −τ ), −∞ K (t −s)u(s)ds) = f .
Consider the Lyapunov function given as:
V = V1 + V2 + V3 + V4 , (20)
where
t 2
V1 = u(t) − R u(s)ds ,
t−τ
t
V2 = 2 u(s) W f ds,
t−τ
t t
V3 = −2RW f (θ − t + τ )u(θ )dθ − 2R 2 (θ − t + τ )u(θ )dθ u(t − τ ),
t−τ t−τ
t t t
V4 = τ R 2 u 2 (s)ds − R u 2 (s)ds − 2 u(s)ds WM. (21)
t−τ t−τ t−τ
Taking derivative with respect to t, we get
V̇ = V˙1 + V˙2 + V˙3 + V˙4 . (22)
Differentiating each Vi with respect to t, we get,

t
V˙1 = 2 u(t) − R W f + Ru(t − τ ) ,
u(s)ds
t−τ
t
= 2 u(t)W f + u(t)Ru(t − τ ) − R u(s)ds W f
t−τ
t
−R 2
u(s)ds u(t − τ ) . (23)
t−τ
For V2 , taking M as bound of f (because we have assumed earlier that the activation
functions are monotonic, bounded and Lipschitz continuous), we have
46 S. Tyagi et al.

V˙2 ≤ 2 u(t) − u(t − τ ) W M, (24)
t
V˙3 = −2τ RW f u(t) + 2RW f u(s)ds − 2τ R 2 u(t − τ )u(t)
t−τ
t
+ 2R 2 u(s)ds u(t − τ ), (25)
t−τ
V˙4 = τ R 2 u 2 (t) − τ R 2 u 2 (t − τ ) − Ru 2 (t) + Ru 2 (t − τ ) − 2u(t)W M + 2u(t − τ )W M.
(26)
Using (24), (25) in (26) in (22) and taking modulus with use of inequality −2ab ≤
(a 2 + b2 ), we obtain
V̇ ≤ −|R|u 2 (t) + 2u(t)|W |M − 2τ |R||W || f |u(t) − 2τ |R 2 ||u(t − τ )||u(t)|

+ τ |R 2 |u 2 (t) − τ |R 2 |u 2 (t − τ ) − |R|u 2 (t),
≤ −2|R|u 2 (t) + 2u(t)|W |M − 2τ |R||W || f |u(t) − 2τ |R 2 ||u(t − τ )||u(t)|
+ τ |R 2 |u 2 (t) − τ |R 2 |u 2 (t − τ ),
≤ −2|R|u 2 (t) + 2Mu(t)|W | − 2τ M|R||W |u(t) + τ |R 2 |u 2 (t)
+ τ |R 2 |u 2 (t − τ ) + τ |R 2 |u 2 (t) − τ |R 2 |u 2 (t − τ ),
= −2|R|u 2 (t) + 2Mu(t)|W | − 2τ M|R||W |u(t) + 2τ |R 2 |u 2 (t),
≤ −2|R|u T (t)u(t),
≤ 0. (27)
Hence the given neural ring network model (3) is globally asymptotically stable.
Remark 3 In proof of global stability, throughout the calculations, u 2 (t) =

u T (t)u(t).
4 Examples
Example 1 Consider the following system, in which the activation function is taken
to be tanh(u) with value of delay kernel term to be 0, so our system takes the
following form as:
u 1 (t) = −0.3u 1 (t) + w1,2 · tanh(u 2 (t − τ1,2 )),

u 2 (t) = −0.5u 2 (t) + w2,1 · tanh(u 1 (t − τ2,1 )), (28)
where R = diag(−0.3, −0.5). Since the activation function we considered here

tanh(u) is a monotone increasing function. Selecting li s to be equal to 1, and taking
w1,2 = 1.4 and w2,1 = −1.2, we get that r1 w1,2 l2 · r2 w2,1l1 = −0.252 < 1 or
|r1 w1,2 l2 ·r2 w2,1 l1 | > 1. Thus by Lemma (2), we get that the system (2) has a unique
equilibrium point.
Example 2 Consider the following three-node system, with tanh(u) as the activation
function.
t
u 1 (t) = −0.3u 1 (t) + w1,3 · f 3 t, u 3 (t − τ1,3 ), k3 (t − s)u 3 (s)ds ,
−∞
t

u 2 (t) = −0.4u 2 (t) + w2,1 · f 1 t, u 1 (t − τ2,1 ), k1 (t − s)u 1 (s)ds ,
−∞
t
u 3 (t) = −0.2u 2 (t) + w3,2 · f 2 t, u 2 (t − τ3,2 ), k2 (t − s)u 2 (s)ds . (29)
−∞
Selecting the integral term in activation function in the following manner,

t
f i t, u i (t − τi,i−1 ), ki−1 (t − s)u i−1 (s)ds
−∞
t
= tanh(u i (t − τi,i−1 )) + ki−1 (t − s)u i−1 (s)ds,
−∞
∞
= tanh(u i (t − τi,i−1 )) + ki−1 (s)u i−1 (t − s)ds.
0
Approximating the integral term with summation term, we have

∞

= tanh(u i (t − τi,i−1 )) + ki−1 (s)u i−1 (t − s).
0

Now to simplify the term ∞ 0 ki−1 (s)u i−1 (t − s), let we choose kernel term as
ki−1 (s) = es , which can be solved as:
∞
∞

ki−1 (s)u i−1 (t − s) = es u i−1 (t − s),
s=0 s=0
= u i−1 (t) + e1 u i−1 (t − 1) + e2 u i−1 (t − 2)
+ e3 u i−1 (t − 3) + · · · ,
≤ u i−1 (t) + e1 u i−1 (t) + e2 u i−1 (t) + e3 u i−1 (t) + · · · ,
(because we have assumed activation functions to be monotonic increasing in this

example)
48 S. Tyagi et al.
= u i−1 (t)(1 + e1 + e2 + e3 + · · · ),
u i−1 (t)
= .
1 − e(1)
Substituting all these values in our example, our system reduces to following form:
u 3 (t)
u 1 (t) ≈ −0.3u 1 (t) + w1,3 · tanh(u 3 (t − τ1,3 )) + ,
1 − e1
u 1 (t)
u 2 (t) ≈ −0.4u 2 (t) + w2,1 · tanh(u 1 (t − τ2,1 )) + ,
1 − e1
u 2 (t)
u 3 (t) ≈ −0.2u 2 (t) + w3,2 · tanh(u 2 (t − τ3,2 )) + , (30)
1 − e1
where R = diag(−0.3, −0.4, −0.2), w1,3 = 1.2, w2,1 = −3.5, w3,2 = −2.6.
Taking d1 = 1, d2 = 1/2, d3 = 1, we get det (R + W D) = 5.4630 = 0, so by
Lemma (3), the system (2) has a unique equilibrium point.
5 Discussion
In this paper, a ring neural network model (2) with delay has been studied and various
results for the existence and uniqueness of the equilibrium point are obtained. Global
stability of the equilibrium point is also studied with the help of suitable Lyapunov
function. At the end, some examples are also given to validate our results.
References
1. Pan, L., Cao, J.: Anti-periodic solution for delayed cellular neural networks with impulsive
effects. Nonlinear Anal. Real World Appl. 12(6), 3014–3027 (2011)
2. Xiang, H., Cao, J.: Almost periodic solutions of recurrent neural networks with continuously
distributed delays. Nonlinear Anal. 71(12), 6097–6108 (2009)
3. Yang, Y., Cao, J.: Stability and periodicity in delayed cellular neural networks with impulsive
effects. Nonlinear Anal. Real World Appl. 8(1), 362–374 (2007)
4. Cao, J., Liang, J.: Boundedness and stability for Cohen-Grossberg neural network with time-
varying delays. J. Math. Anal. Appl. 296(2), 665–685 (2004)
5. Abbas, S., Xia, Y.: Existence and attractivity of k-almost automorphic sequence solution of a
model of cellular neural networks with delay. Acta Math. Sci. Ser. B Engl. Ed. 33(1), 290–302
(2013)
6. Abbas, S.: Existence and attractivity of k-pseudo almost automorphic sequence solution of a
model of bidirectional neural networks. Acta Appl. Math. 119, 57–74 (2012)
7. Abbas, S.: Pseudo almost periodic sequence solutions of discrete time cellular neural networks.
Nonlinear Anal. Model. Control 14(3), 283–301 (2009)
8. Gao, B., Zhang, W.: Equilibria and their bifurcations in a recurrent network involving iterates
of a transcendental function. IEEE Trans. Neural Netw. 19(5), 782–794 (2008)
9. Baldi, P., Atiya, A.F.: How delays affect neural dynamics and learning. IEEE Trans. Neural
Netw. 5(4), 612–621 (1994)
10. Feng, C., Plamondon, R.: An oscillatory criterion for a delayed neural ring network model.
Neural Netw. 29–30, 70–79 (2012)
11. Wei, J., Zhang, C.: Bifurcation analysis of a class of neural networks with delays. Nonlinear
Anal: Real World Appl. 9, 2234–2252 (2008)
Approximation of Solutions
of a Stochastic Differential
Equation
Sanjukta Das, D.N. Pandey and N. Sukavanam
Abstract The existence, uniqueness, and convergence of approximate solutions of

a stochastic differential equation with deviated argument is studied using analytic
semigroup theory and fixed point method. Then we consider Faedo-Galerkin approx-
imation of solution and prove some convergence results. We also study an example
to illustrate our result.
Keywords Analytic semigroup · Deviated argument · Stochastic Fractional differ-

ential equation
Mathematics Subject Classification (2010): 34G10, 34G20, 34K30, 35K90,

47D20
1 Introduction
Fractional differential equations appear abundantly in the theory of fractals, visco-

elasticity, seismology, polymers, etc. Stochastic evolution equations are natural gen-
eralizations of ordinary differential equations incorporating the random noise which
causes fluctuations in deterministic models. For details refer [1]. In certain real-world
problems, delay depends not only on the time but also on the unknown quantity as
we can see in [2]. Das et al. [3, 4] can be referred for related work with deviated argu-
ment. Bahuguna et. al. [5] discussed the Faedo-Galerkin approximation of solution.
S. Das (B) · D.N. Pandey · N. Sukavanam

Department of Mathematics, IIT Roorkee, Roorkee, India
e-mail: sanjukta_das44@yahoo.com
D.N. Pandey
e-mail: nsukvfma@iitr.ernet.in
N. Sukavanam
e-mail: nsukvfma@iitr.ernet.in

DOI 10.1007/978-81-322-2485-3_4
52 S. Das et al.
So far the Faedo-Galerkin approximation of solution stochastic fractional differential

equation with deviated argument is neglected in the literature. In an attempt to fill this
gap we study the following stochastic fractional differential equation with deviated
argument in a separable Hilbert space (H, (., .)).
β dw(t)
c
Dt u(t) + Au(t) = f (u(t), u(h(u(t)))) , t ∈ [0, T ]
dt
u(0) = u 0 ∈ H (1)
β
where 0 < β < 1 and 0 < T < ∞. c Dt denotes the Caputo fractional derivative of
order β and A : D(A) ⊂ X → H is a linear operator. A and the functions f, h are
defined in the hypotheses (H 1) − (H 3) of Sect. 2.
2 Preliminaries
Here we deal with two separable Hilbert spaces H and K .

(H1) A is a closed, densely defined, self-adjoint operator with pure point spectrum
0 ≤ λ0 ≤ λ1 ≤ · · · ≤ λm ≤ · · · with λm → ∞ and m → ∞ and
corresponding complete orthonormal system of eigenfunctions φ j such that
Aφ j = λ j φ j and < φi , φ j >= δi, j
(H2) The function f : [O, T ] × Hα × Hα−1 → L(K , H ) is continuous and ∃

constant L f such that
f (u, u 1 ) − f (v, v1 )2Q ≤ L f [+u − vα + u 1 − v1 α−1 ]
(H3) The map h : Hα × R+ → R+ satisfies h(u, ) − h(v, ) ≤ L h (u − vα )

If (H 1) is satisfied then −A is the infinitesimal generator of an analytic semigroup
{e−t A : t ≥ 0} in H . We also note that ∃ constant C such that S(t) ≤ Ceωt and
di −1
constants Ci ’s such that dt i S(t) ≤ C i , t > 0, i = 1, 2. Also AS(t) ≤ Ct
α
and A S(t) ≤ Cα t . −α
We define the space Hα as D(Aα ) endowed with the norm .α . Let (Ω, F, P)
be a complete probability space endowed with complete family of right continu-
ous increasing sub σ—algebras {Ft , t ∈ J } such that Ft ⊂ F. A H —valued ran-
dom variable is a F—measurable process. We also assume that W is a Wiener
process on K with covariance operator Q. Suppose Q is symmetric, positive, lin-
1
ear and bounded operator with T r Q < ∞. Let K 0 = Q 2 (K ). The space L 02 =
1
L 2 (K 0 , Hα ) is a separable Hilbert space with norm ψ L 0 = ψ Q 2 L 2 (K ,Hα ) . Let
2
L 2 (Ω, F, P; Hα ) ≡ L 2 (Ω; Hα ) be the Banach space of all strongly measur-
able, square integrable, Hα —valued random variables equipped with the norm
Approximation of Solutions of a Stochastic Differential Equation 53
u(.)2L 2 = Eu(.; w)2Hα . C Tα denotes the Banach space of all continuous

maps from J = (0, T ] into L 2 (Ω; Hα ) which satisfy sup t∈J Eu(t)C
2
α < ∞.
L 2 (Ω, Hα ) = { f ∈ L 2 (Ω, Hα ) : f is F0 − measurable} denotes an important

0
subspace. For 0 ≤ α < 1 define
C Tα−1 = {u ∈ C Tα : u(t) − u(s)α−1 ≤ L|t − s|, ∀t, s ∈ [0, T ]}.
Now let us define mild solution of (1):

Definition 1 The mild solution of (1) is a continuous Ft adapted stochastic process
u ∈ C Tα ∩ C Tα−1 which satisfies the following:
1. u(t) ∈ Hα has C àdl àg paths on t ∈ [0, T ].
2. ∀t ∈ [0, T ], u(t) is the solution of the integral equation
t
u(t) = Tβ (t)u 0 + (t − s)β−1 Sβ (t − s) f (u(s), u(h(u(s), s)))dw(s), t ∈ [0, T ]
0
(2)
∞ ∞
where Sβ (t) = 0 ζβ (θ)S(t β θ)dθ; and Tβ (t) = q 0 θζβ (θ)S(t β θ)dθ; ζβ is a
probability density function defined on (0, ∞), i.e. ζβ (θ) ≥ 0, θ ∈ (0, ∞) and
∞ βC α
0 ζβ (θ)dθ = 1. Also Tβ (t)u ≤ Cu, Sβ (t)u ≤ Γ (1+β) u, A Sβ (t)u ≤
βCα Γ (2−α) −αβ
Γ (1+β(1−α)) t u.
Lemma 1 Let f : J × Ω × Ω → L 02 be a strongly measurable mapping with
T p
0 E f (t) L 0 dt < ∞. Then
2
t t
p
E f (s)dw(s) ≤ ls
p
E f (s) L 0 ds
0 0 2
∀t ∈ [0, T ] and p ≥ 2 where ls is a constant containing p and T.

ls is incorporated into the constants in the following sections.
3 Existence and Uniqueness of Approximate Solutions
In this section we consider a sequence of approximate integrals and establish the

existence and uniqueness of solution for each of the approximate integral equa-
tions. For 0 ≤ α < 1 and u ∈ C Tα0 , the hypotheses (H 2) − (H 3), imply that
f (u(s), u(h(u(s), s))) is continuous on [0, T0 ]. Therefore, ∃ a positive constant
N = 2L f [T0θ1 + 2R(1 + L L h ) + L L h T0θ2 ] + 2N0 , N0 = E f (u 0 , u 0 )2
such that f (s, u(s), u(h(u(s), s))) ≤ N , t ∈ [0, T ]. Choose T0 , 0 < T0 ≤ T

such that
54 S. Das et al.
2 β(1−α)−1
βCα Γ (2 − α) T0 R
N ≤ ,
Γ (1 + β(1 − α)) β(1 − α) − 1 4
2 β(1−α)−1
βCα Γ (2 − α) T0
D = 2L f ≤1 (3)
Γ (1 + β(1 − α)) 2β(1 − α) − 1
Let
B R = {u ∈ CTα0 ∩ CTα−1
0
: u(0) = u 0 , u − u 0 T0 ,α ≤ R}
It is easy to see that B R is a closed and bounded subset of CTα−1

0
and complete. Let
us define the operator Fn : B R :→ B R by
t
(Fn u)(t) = Tβ (t)u 0 + (t −s)β−1 Sβ (t −s) f n (u(s), u(h(u(s), s)))dw(s). (4)
0
Theorem 1 If the hypotheses (H 1), (H 2) and (H 3) are satisfied and u 0 ∈ L 02 (Ω,

X α ), 0 ≤ α < 1, then ∃ a unique u n ∈ B R such that Fn u n = u n , ∀ n = 0, 1, 2, · · · ,
i.e., u n satisfies the approximate integral equation
t
u n (t) = Tβ (t)u 0 + (t − s)β−1 Sβ (t − s) f n (s, u n (s),u n (h(u n (s), s)))dw(s),
0
t ∈ [0, T ] (5)
Proof Step1 : We need to show that Fn u ∈ CTα−1

0
, ∀u ∈ CTα−1
0
. It is easy to check
α α α−1
that Fn : CT → CT . If u ∈ CT0 , 0 < t1 < t2 < T0 and 0 ≤ α < 1 then
EFn u(t2 ) − Fn u(t1 )2α−1

≤ 3E[Tβ (t2 ) − Tβ (t1 )]u 0 2α−1
t2
+ 3E (t2 − s)β−1 Aα−1 Sβ (t2 − s) f n (u(s), u(h(u(s)))dw(s)2Q
t1
t1
+ 3E A[(t2 − s)β−1 Sβ (t2 − s) − (t1 − s)β−1 Sβ (t1 − s)]
0
Aα−2 × f n (u(s), u(h(u(s))))dw(s) Q

β 2 C 2 Γ 2 (2 − α) t2
≤ 3E[Tβ (t2 ) − Tβ (t1 )]u 0 2α−1 +3 2 α (t2 − s)2β(1−α)−2
Γ (1 + β(1 − α)) t1
× A−1 2 E f n (u(s), u(h(u(s), )))2 ds
t1
+3 A[(t2 − s)β−1 Sβ (t2 − s) − (t1 − s)β−1 Sβ (t1 − s)]
0
× Aα−2 2 E f n (u(s), u(h(u(s))))2 ds (6)
∀u ∈ H, we can write
t2 t2
β β d
[S(t2 θ) − S(t1 θ)]u = S(t β θ)udt = θβt β−1 AS(t β θ)dt.
t1 dt t1
The first term of (6) can be estimated as follows:

∞ 2
β β
[Tβ (t2 ) − Tβ (t1 )]u 0 2α−1 ≤ ζβ (θ)S(t2 θ) − S(t1 θ)Aα−1 u 0 dθ
0
∞ t2 2
d
≤ ζβ (θ)[ S(t β θ)dt]u 0 α dθ
0 t1 dt
≤ C12 u 0 2α−1 (t2 − t1 )2 (7)
For the second term of (6) we get the following estimate

t2
(t2 − s)2β(1−α)−2 E f n (u(s), u(h(u(s))))2 ds
t1
N (t2 − t1 )2β(1−α)−1
≤ (8)
2β(1 − α) − 1
For the third term we will use the following estimate

t1
A[(t2 − s)β−1 Sβ (t2 − s) − (t1 − s)β−1 Sβ (t1 − s)]2
0
× Aα−2 2 E f n (u(s), u(h(u(s))))2 ds
t1 ∞ 2
d d
≤ ζβ (θ)[ S((t − s)β θ)|t=t2 − S((t − s)β θ)|t=t1 ]dθ
0 0 dt dt
× E f (u(s), u(h(u(s))))2 ds
t1 ∞ t2 2
d2
≤ ζβ (θ)[ Aα−2 2 S((t − s)β θ)dt]dθ N ds
0 0 t1 dt
≤ C22 Aα−2 2 (t2 − t1 )2 N T0 (9)
Hence from inequalities (7)–(9) we see that the map Fn : CTα−1 0

→ CTα−1
0
is well-
defined. Now we prove that Fn : B R → B R . So for t ∈ [0, T0 ] and u ∈ B R .
E(Fn u)(t) − u 0 2α

≤ 2E(Tβ (t) − I )u 0 2α
t
+ 2E (t − s)β−1 Sβ (t − s) f (u(s), u(h(u(s))))dw(s)2Q
0
56 S. Das et al.
2
βCα Γ (2 − α) t
≤ 2E(Tβ (t) − I )u 0 2α +2 (t2 − s)2β(1−α)−2 2
Γ (1 + β(1 − α)) 0
× E f n (u(s), u(h(u(s))))2 ds
β(1−α)−1
R βCα Γ (2 − α) 2 T R R
≤ +2 N 0 ≤ + =R
2 Γ (1 + β(1 − α)) β(1 − α) − 1 2 2
Now we show that Fn is a contraction map by using (3) in last but one inequality.
∀u, v ∈ B R
t
E(Fn u)(t) − (Fn v)(t)2α = E (t − s)β−1 Aα Sβ (t − s)
0
× [ f (u(s), u(h(u(s)))) − f (s, v(s), v(h(v(s), s)))dw(s)]2Q

βCα Γ (2 − α) 2 t
≤ (t2 − s)2β(1−α)−2
Γ (1 + β(1 − α)) 0
× E f (u(s), u(h(u(s)))) − f (v(s), v(h(v(s))))2 ds

βCα Γ (2 − α) 2 T 2 β(1 − α) − 1
≤ 2L f (1 + 2L Lh)u − v2α
Γ (1 + β(1 − α)) 2β(1 − α) − 1
≤ u − v2α .
This implies that ∃ a unique fixed point u n of Fn . Thus there a unique mild approx-
imate solution of (1)
Lemma 2 Let (H 1) − (H 3) hold. If u 0 ∈ L 02 (Ω, D(Aα )), ∀0 < α < η < 1, then
u n (t) ∈ D(Aγ ) for all t ∈ [0, T0 ] with 0 < γ < η < 1. Also if u 0 ∈ D(A), then
u n (t) ∈ D(Aγ ) ∀t ∈ [0, T0 ], where 0 < γ < η < 1.
Proof By Theorem (1) we get the existence of a unique u n ∈ B R , satisfying (5).

Theorem 2.6.13 of [6] implies for t > 0, 0 ≤ γ < 1, S(t) : H → D(Aγ ) and for
0 ≤ γ < η < 1, D(Aη ) ⊂ D(Aγ ). It is easy to see that Holder continuity of u n can
be proved using the similar arguments from (6) to (9). Also from Theorem 1.2.4 in
[6], we have S(t)u ∈ D(A) if u ∈ D(A). The result follows from these facts and
that D(A) ⊂ D(Aγ ) for 0 ≤ γ < 1.
Lemma 3 Let (H 1) − (H 3) hold and u 0 ∈ L 02 (Ω, X α ). Then for any t0 ∈ (0, T0 ]

∃ a constant Ut0 , independent of n such that Eu n (t)2γ ≤ Ut0 ∀t ∈ [t0 , T0 ], n =
1, 2, · · · . Also if u 0 ∈ L 02 (Ω, D(A)) then ∃ constant U0 independent of n such that
Eu n (t)2γ ≤ U0 ∀t ∈ [t0 , T0 ], n = 1, 2, · · · , ∀ 0 < γ ≤ 1.
Proof Let u 0 ∈ L 02 (Ω, Hα ). Applying Aγ on both sides of (4)

Eu n (t)2γ
t
≤ 2ETβ (t)u 0 2γ + 2 (t − s)β−1 Sβ (t − s) f n (u(s), u(h(u(s))))dw(s)2Q
0
2
−2γβ βCγ Γ (2 − γ) N (T0 )2β(1−γ)−1
≤ 2Cγ2 t0 u 0 2 + = U t0 .
Γ (1 + β(1 − γ)) 2β(1 − γ) − 1
Also if u 0 ∈ L 02 (Ω, D(A)), then we have that u 0 ∈ L 02 (Ω, D(Aγ )) for 0 ≤ γ < 1.
Hence,
Eu n (t)2γ
t
≤ 2ETβ (t)u 0 2γ + 2 (t − s)β−1 Sβ (t − s) f n (u(s), u(h(u(s))))dw(s)2Q
0
2
βCγ Γ (2 − γ) N (T0 )2β(1−γ)−1
≤ 2C 2 u 0 2 + = U0 .
Γ (1 + β(1 − γ)) 2β(1 − γ) − 1
Hence proved.
4 Convergence of Solutions
In this section the convergence of the solution u n ∈ Hα of the approximate integral

equation (5) to a unique solution u of (2), is discussed.
Theorem 2 Let the hypotheses (H 1) − (H 3) hold and if u 0 ∈ L 02 (Ω, Hα ) then
∀t0 ∈ (0, T ],
lim sup u n (t) − u m (t)α = 0.
m→∞ {n≥M,t ≤t≤T }
0 0
Proof Let 0 < α < γ < η. For t0 ∈ (0, T0 ]
E f n (u n (t), u n (h(u n (t)))) − f m (t, u m (t), u m (h(u m (t))))2

≤ 2E f n (u n (t), u n (h(u n (t)))) − f n (t, u m (t), u m (h(u m (t))))2
≤ 2E f n (u m (t), u m (h(u m (t)))) − f m (t, u m (t), u m (h(u m (t))))2
≤ 2(2L f (1 + 2L L h )[Eu n − u m 2α + E(P n − P m )u m (t)2α ]) (10)
Now,
1
E(P n − P m )u m (t)2 ≤ EAα−γ (P n − P m )Aγ u m (t)2 ≤ 2(γ−α)
EAγ u m (t)2
λm
58 S. Das et al.
Then we have
E f n (t, u n (t), u n (h(u n (t)))) − f m (t, u m (t), u m (h(u m (t))))2

1 γ
≤ 2 2L f (1 + 2L L h ) Eu n − u m α + 2(γ−α) EA u m (t)
2 2
λm
For 0 < t0 < t0

t
t0
Eu n (t) − u m (t)2α ≤2 + (t − s)β−1 Aα Sβ (t − s)2
0 t0
× E f n (u n (t), u n (h(u n (t)))) − f m (u m (t), u m (h(u m (t))))2 ds (11)
The estimate of first integral of the above inequality is
Eu n (t) − u m (t)2α

t
0
≤ (t − s)β−1 Aα Sβ (t − s)2
0
× E f n (u n (t), u n (h(u n (t)))) − f m (u m (t), u m (h(u m (t))))2 ds

βCγ Γ (2 − γ) 2 2N (t0 − δ1 t0 )2β(1−γ)−2
≤ t0 , 0 < δ < 1 (12)
Γ (1 + β(1 − γ)) 2β(1 − γ) − 1
The estimate of second integral is

t
Eu n (t) − u m (t)2α ≤ (t − s)β−1 Aα Sβ (t − s)2
t0
× E f n (u n (t), u n (h(u n (t)))) − f m (u m (t), u m (h(u m (t))))2 ds

βCγ Γ (2 − γ) 2 t
≤ (t − s)2β(α−1)−2
Γ (1 + β(1 − γ)) t0

EAγ u m (s)2
× 4L f (1 + 2L L h ) Eu n − u m α + 2
ds
λ2 (γ − α)
t
βCγ Γ (2 − γ) 2
≤ 4L f (1 + 2L L h ) [ (t − s)2β(α−1)−2
Γ (1 + β(1 − γ)) t0
2β(1−α)−1
T U t0
× Eu n − u m 2α ds + 2(γ−α) 0 ] (13)
λm 2β(1 − α) − 1
Substituting inequalities (12), (13) into (11) we get

Eu n (t) − u m (t)2α

βCγ Γ (2 − γ) 2 4N (t0 − δ1 t0 )2β(1−γ)−2
≤ t0
Γ (1 + β(1 − γ)) 2β(1 − γ) − 1
t
βCγ Γ (2 − γ) 2
+ 8L f (1 + 2L L h ) [ (t − s)2β(α−1)−2
Γ (1 + β(1 − γ)) t0
2β(1−α)−1
U t0 T0
× Eu n − u m 2α ds + 2(γ−α) 2β(1 − α) − 1
]
λm
By using Gronwall’s inequality, ∃ a constant D such that

βCγ Γ (2 − γ) 2 4N (t0 − δ1 t0 )2β(1−γ)−2
Eu n (t) − u m (t)2α≤ t0
Γ (1 + β(1 − γ)) 2β(1 − γ) − 1
2β(1−α)−1

βCγ Γ (2 − γ) 2 Ut0 T0
+ 8L f (1 + 2L L h ) ×D
Γ (1 + β(1 − γ)) 2(γ−α) 2β(1 − α) − 1
λm
Let m → ∞. Taking supremum over [t0 , T0 ] we get the following inequality:

2
βCγ Γ (2 − γ) 4N (t0 − δ1 t0 )2β(1−γ)−2
Eu n (t) − u m (t)2α ≤ t0 × D
Γ (1 + β(1 − γ)) 2β(1 − γ) − 1
Since t0 is arbitrary, the right-hand side can be made infinitesimally small by choosing
t0 sufficiently small. Thus the lemma is proved.
Corollary 1 If u 0 ∈ D(A), then lim sup Eu n (t) − u m (t)2α = 0

m→∞ {n≥m, 0≤t≤T }
0
Proof By using Lemmas (2) and (3) we can take t0 = 0 in the proof of Theorem (2)
and hence the corollary follows.
Theorem 3 Let us assume that (H 1) − (H 3) are satisfied and suppose u 0 ∈

L 02 (Ω, X α ). Then for t ∈ [0, T0 ], ∃ a unique function u n ∈ B R where
t
u n (t) = Tβ u 0 + 0 (t − s)β−1 Sβ (t − s) f n (u n (s), u n (h n (u n (s))))dw(s),
and u(t) ∈ B R , where
u(t) = Tβ u 0 + 0 (t − s)β−1 Sβ (t − s) f (u(s), u(h(u(s))))dw(s), t ∈ [0, T0 ], such
t
that u n → u as n → ∞ in B R and u satisfies (2) on [0, T0 ].
Proof By using the above Corollary, Theorems 1 and 2 it is to see that ∃ u(t) ∈ B R
such that
limn→∞ Eu n (t) − u(t)2α = 0 on [0, T0 ]. Now
60 S. Das et al.
t
Eu n (t) − Tβ u 0 + (t − s)β−1 Sβ (t − s) f n (u n (s), u n (h n (u n (s))))dw(s)2
t0
t0
≤ E (t − s)β−1 Sβ (t − s) f n (u n (s), u n (h n (u n (s))))dw(s)2
0
2 2β−2
βC T0
≤ N t0 (14)
Γ (1 + β) 2β − 2
Let n → ∞ then
Eu n (t) − Tβ u 0 + t0 (t − s)β−1 Sβ (t − s) f n (u n (s), u n (h n (u n (s))))dw(s)2
t
2 2β−2
βC T0
≤ Γ (1+β) N 2β−2 t0 and since t0 is arbitrary we conclude u(t) satisfies (2).
Uniqueness follows easily from Theorems 1, 2 and Gronwall’s inequality.
4.1 Faedo-Galerkin Approximations
We know from the previous sections that for any 0 ≤ T0 ≤ T , we have a unique
u ∈ C Tα0 satisfying the integral equation
t
u(t) = Tβ u 0 + 0 (t − s)β−1 Sβ (t − s) f (u(s), u(h(u(s))))dw(s), t ∈ [0, T0 ] Also,
∃ a unique solution u n ∈ C Tα0 of the approximate integral equation
t
u n (t) = Tβ u 0 + 0 (t − s)β−1 Sβ (t − s) f n (u n (s), u n (h(u n (s))))dw(s), t ∈ [0, T0 ].
Faedo-Galerkin approximation ū n = P n u n is given by
P n u n (t) = ū n (t) = Tβ (t)P n u 0
t
+ 0 (t − s)β−1 Sβ (t − s)P n f (u n (s), u n (h(u n (s))))dw(s), t ∈ [0, T0 ]. If
the solution u(t) to (2) exists on [0, T0 ] then it has the representation
∞
u(t) = αi (t)φi , where αi (t) = (u(t), φi ) for i = 0, 1, 2, 3, · · · and
i=0
n
ū n (t) = αin (t)φi , where αin (t) = (ū n (t), φi ) for i = 0, 1, 2, 3, · · · .
i=0
As a consequence of Theorems 1 and 2, we have the following result.
Theorem 4 Let us assume that (H 1) − (H 3) are satisfied and suppose u 0 ∈
L 02 (Ω, X α ). Then for t ∈ [0, T0 ], ∃ a unique function u n ∈ B R where
t
u n (t) = Tβ P n u 0 + 0 (t − s)β−1 Sβ (t − s)P n f n (u n (s), u n (h(u n (s))))dw(s),
and u(t) ∈ B R , where
u(t) = Tβ u 0 + 0 (t − s)β−1 Sβ (t − s) f (u(s), u(h(u(s))))dw(s), t ∈ [0, T0 ], such
t
that u n → u as n → ∞ in B R and u satisfies (2) on [0, T0 ].

Now the convergence of αin (t) → αi (t) is shown. It is easily seen that
∞

n

Aα [u(t) − ū n (t)] = Aα {αi (t) − αin (t)}φi + Aα αi (t)φi
i=0 i=n+1

n ∞

= λiα {αi (t) − αin (t)}φi + λiα αi (t)φi . Thus we have
i=0
n i=n+1
EAα [u(t) − ū n (t)2 ≥ i=0 λi E|αi (t) − αi (t)| .
2α n 2
Theorem 5 Let us assume (H 1) − (H 3)hold.

n
(i) If u 0 ∈ L 02 (Ω, X α ) then lim sup λi (t)2α Eαi (t) − αin (t)2 = 0
n→∞ t∈[t ,T ]
0 0

i=0

n
(ii) If u 0 ∈ L 02 (Ω, D(A)) then lim sup λi (t)2α Eαi (t) − αin (t)2 = 0
n→∞ t∈[0,T ]
0 i=0
Theorem 5 follows from the facts mentioned above the theorem.

Corollary 2 Let us assume (H 1) − (H 3) hold.
(i) If u 0 ∈ L 02 (Ω, X α ) then lim sup EAα [ū n (t) − ū m (t)]2 = 0
n→∞ t∈[t ,T ],n≥m
0 0
(ii) If u 0 ∈ L 02 (Ω, D(A)) then lim sup EAα [ū n (t) − ū m (t)]2 = 0
n→∞ t∈[0,T ],n≥m
0
Proof
EAα [ū n (t) − ū m (t)]2 = EP n u n (t) − P m u m (t)2α

≤ 2EP n [u n (t) − u m (t)]2α + 2E(P n − P m )ym (t)2α
1
≤ 2E[u n (t) − u m (t)]2α + 2 γ−α EAγ u m (t)2
λm
Then the result (i) follows from Theorem 2 and result (ii) follows from Corollary 1.
5 Example
Suppose for t ≥ 0, x ∈ (0, 1), 0 < β ≤ 1
dw(t)
c
D β vt (t, x) = vx x (t, x) + F(v(t, x), v(h(t, v(x)))) ,
dt
v(t, x) = v0 , t = 0, x ∈ (0, 1) and v(t, 0) = v(t, 1) = 0, t ≥ 0 (15)
Let F be an appropriate Holder continuous function satisfying (H 2) in

L 02 (K , (0, 1)). w is a standard L 2 (0, 1) valued Weiner process. Let us define A =
2
− ddx 2 , f := F, v(x) = u(t) and let D(A) = H01 (0, 1) ∩ H 2 (0, 1), D(A1/2 ) =
H01 (0, 1). Then (15) can be reformulated into (1). Now from Theorems (1), (2) we
can similarly prove the existence, uniqueness, and approximation of the mild solution
of (15).
62 S. Das et al.
References
1. Xua, D., Yanga, Z., Huang, Y.: Existence uniqueness and continuation theorems for stochastic
functional differential equations. J. Differ. Equ. 245(6), 1681–1703 (2008)
2. Das, S., Pandey, D.N., Sukavanam, N.: Approximate controllability of a second-order neutral
differential equation with state dependent delay. Differ. Equa. Dyn. Syst. 218, (2014). doi:10.
1007/s12591-014-0218-6
3. Das, S., Pandey, D.N., Sukavanam, N.: Approximate controllability of a functional differential
equation with deviated argument. Nonlinear Dyn. Syst. Theory 14(3), 265–277 (2014)
4. Das, S., Pandey, D.N., Sukavanam, N.: Exact controllability of an impulsive semilinear system
with deviated argument in a banach space. J. Differ. Equ. 461086, 6 (2014). doi:10.1155/2014/
461086
5. Bahuguna, D., Muslim, M.: Approximation of solutions to retarded differential equations with
applications to population dynamics. J. Appl. Math. Stochast. Anal. 01, 1–11 (2005)
6. Pazy, A.: Semigroups of Linear Operators and Applications to Partial Differential Equations.
Springer (1983)
Reconstruction of Multiply Generated
Splines from Local Average Samples
P. Devaraj and S. Yugesh
Abstract We analyze the following average sampling

problem: Let h be a non-
negative measurable function supported in − 21 , 21 . Given a sequence of samples
{yn }n∈Z ∈ RZ of polynomial growth, find a multiply generated spline f of polyno-
1
mial growth such that 2 1 f (n − t)h(t)dt = yn , n ∈ Z. It is shown that the solution
−2
to this problem is unique over certain subspaces of the multiply generated spline
space of polynomial growth.
Keywords Interpolation · Multiply generated splines · Average sampling
1 Introduction
The sampling theorem is one of the widely used results in the signal processing field.
The well-known Shannon sampling theorem states that, any bandlimited signal f
is completely determined by its samples [4, 8]. Although the Shannon sampling
theorem is very useful, it has a number of problems when using it for practical appli-
cations. The bandlimited functions have analytic continuation to the entire complex
plane and hence they are of infinite duration which is not always realistic. On the
other hand, the sinc function has a very slow decay. Further, the measured samples
are not exact in practical problems and they are the average of the signal around
the sampling point and the averaging function depends on the aperture device used
for capturing the samples. For these reasons, sampling and local average sampling
have been investigated in several other classes of signals. In general, spline spaces
yield many advantages in their generation and numerical treatment so that there are
P. Devaraj · S. Yugesh (B)

Department of Mathematics, College of Engineering Guindy Campus,
Anna University, Chennai 600 025, India
e-mail: devaraj@annauniv.edu
S. Yugesh
e-mail: mathsyugesh@gmail.com

DOI 10.1007/978-81-322-2485-3_5
64 P. Devaraj and S. Yugesh
many practical applications in signal, image processing, and communication the-

ory. In the literature [1–8] many authors have investigated the generalized sampling
technique for multiply generated shift-invariant spaces and spline subspaces. The
multiply generated spline space is defined in [5, 6] as

r
S = f : f = ai (n)βdi (t − n)
i=1 n∈Z
with suitable coefficients ai (n), where βdi is the cardinal central B-spline of degree
di and is defined by,
βdi = χ[− 1 , 1 ] χ[− 1 , 1 ] · · · χ[− 1 , 1 ] (di + 1 terms),

2 2 2 2 2 2
where represents the convolution (The convolution of two functions f and g is

defined as f g(n) = f (t)g(n − t)dt). We consider the following subspace of the
multiply generated spline space:

r
S N := f : f = an βdi (t − n)
n∈Z i=1
If M = max{d1 , d2 , . . . , dr } and m = min{d1 , d2 , . . . , dr }, then f ∈ S N provided

that f (x) ∈ C m−1 (R) and that the restriction of f (x) to any interval between con-
secutive knotsis identical with a polynomial of degree not exceeding M. If di ’s are
distinct, then ri=1 βdi (. − n), n ∈ Z are globally linearly independent.
We consider the following local average sampling problem:
Problem: Let {yn }n∈Z be a given sequence of real numbers. Find a spline

f ∈ SN
such that f h(n) = yn , n ∈ Z, where h ∈ L 1 (R) and h ri=1 βdi (n) = 0, for
some n ∈ Z and
1
0 2
1 1
supp(h) ⊆ − , , h(t) ≥ 0, t ∈ R, 0 < h(t)dt < ∞, 0 < h(t)dt < ∞.
2 2
− 12 0
(1)
We show that this problem has infinitely many solutions. The uniqueness of so-
lution is obtained by imposing the following growth conditions on the samples and
the splines as that of Schoenberg [9]:

SN ,γ = f (t) ∈ S N : f (t) = O(|t|γ ) as t → ±∞
and
Dγ = {yn } : yn = O(|n|γ ) as n → ±∞ .
Reconstruction of Multiply Generated Splines … 65
This problem over the singly generated spline space is analyzed in [10]. It is
shown in [10] that the local average sampling problem has a unique solution for
d ≤ 4 when the spline space is generated by a single central B-spline. For d > 4 the
same problem has been posed as an open problem. The same authors have analyzed
the problem for d ≥ 5 by reducing the support of h. They have shown in [11] that
the local average sampling
problem for singly generated spline has a unique solution
when h is supported in − 2l , 2l , l < 1.
r
Lemma 1 Let ψ(x) = i=1 βdi (x) and let A be the greatest integer such that
h ψ(n) = 0, ∀n < A, and let N1 be the smallest nonnegative integer such that
h ψ(n) = 0, ∀n > A+ N1 . Then the solutions of the problem form a linear manifold
in S N of dimension N1 . Moreover, N1 = M + 1, if M is odd and N1 = M, if M is
even.
Proof When N1 = 0 this problem has a unique solution. For N1 > 0, we consider
the linear map from CZ to S N defined by

{an }n∈Z
−→ an ψ(t − n).
n∈Z
Since the integer translates of ψ are globally linearly independent, this map is an
isomorphism from CZ onto S N . Therefore h f (n) = yn in CZ if and only if,

N1
h ψ(A + k)an−A−k = yn , ∀n ∈ Z.
k=0
This forms a linear difference equation of order N1 with constant coefficients. Hence
the solution space is an N1 dimensional manifold in S N .
2 Local Average Sampling Theorems
Theorem 1 (Main Theorem) Let di ≤ 4 and let h(t) be an integrable function

satisfying condition (1). Then for a given sequence of numbers {yn }n∈Z ∈ Dγ , there
exists a unique f ∈ S N ,γ such that
f h(n) = yn , n ∈ Z. (2)
We define the function

r
G(z) := G i (z)
i=1
where
G i (z) := h βdi (n)z n .
n∈Z
The exponential Euler spline is defined as

ϒz,di (t) = z n βdi (n − t).
n∈Z
21
In terms of the exponential Euler spline we can write G i (z) = h(t)ϒz,di (t)dt.
− 21
Hence
1
2
G(z) = h(t)ϒz (t)dt,
− 21

where ϒz (t) = ri=1 ϒz,di (t) = n∈Z z n ri=1 βdi (n − t).
We need some properties of ϒz (t).
Lemma 2 For d ∈ N, n ∈ Z and z ∈ C \ {0}, we have:
(i) ϒz −1 (−t) = ϒz (t),
(ii) ϒz (t + n) = (z)n ϒ z (t),
(iii) dt (ϒz,di +1 (t)) = 1 − 1z ϒz,di t + 21 ,

d

(iv) ϒ−1,di 21 = 0 and ϒ−1,di (t) > 0 for t ∈ − 21 , 21 .
Proof (i)

r
ϒz −1 (−t) = z −n βdi (n + t)
n∈Z i=1
r
r
= z −n βdi (−n − t) = zn βdi (n − t) = ϒz (t).
n∈Z i=1 n∈Z i=1
(ii)

r
r
ϒz (t + n) = zk βdi (k − t − n) = z n zk βdi (k − t) = z n ϒz (t).
k∈Z i=1 k∈Z i=1
(iii)
d d
(ϒz,di +1 (t)) = z n (βdi +1 (n − t))
dt dt
n∈Z

1

1

= z n βdi n − t + − βdi n − t −
2 2
n∈Z

1 1
= ϒz,di t+ − z n−1 βdi n − 1 − t +
2 2
n∈Z

1 1 n 1
= ϒz,di t + − z βdi n − t +
2 z 2
n∈Z

1 1 1
= ϒz,di t + − ϒz,di t +
2 z 2

1 1
= 1− ϒz,di t +
z 2
(iv)
1 1
ϒ−1,di = (−1) βdi n −
n
= 0.
2 2
n∈Z
We shall show that ϒ−1,di (t) > 0 for t ∈ − 21 , 21 , by using induction on di . For
di = 1 by simple manipulation we get ϒ−1,1 (t) > 0 for t ∈ (− 21 , 21 ). Assume that
it is true for di and using (iii) we get

d 1 1
(ϒ−1,di +1 (t)) = 2ϒ−1,di t+ > 0, t ∈ − , 0 .
dt 2 2

Using ϒ−1,di +1 − 21 = 0 and ϒ−1,di +1 and being an even function, we obtain that

ϒ−1,di (t) > 0 for t ∈ − 21 , 21 .
2.1 Uniqueness Theorem
Theorem 2 Let = { f ∈ S N : f h(n) = 0, n ∈ Z} and z 1 , z 2 , . . . , zl be the

roots of G(z). If the roots of G(z) are simple, then the set of functions ϒz −1 , where
j
j = 1, 2, . . . , l form a basis of .
Proof By Lemma 1, is a l = N1 dimensional subspace of S N .
Using Lemma 2, we get
1
2
h ϒz −1 (n) = h(t)ϒz −1 (n − t)dt
j − 21 j
1
2
= z nj h(t)ϒz −1 (−t)dt
− 21 j
1
2
= z nj h(t)ϒz j (t)dt
− 21
= z nj G(z j )
Therefore, ϒz −1 ∈ for j = 1, 2, . . . , l.
j
Next, we have to prove that the elements of are linearly independent.

l l r
c j ϒz −1 (t) = 0 ⇔ cj z −n
j βdi (n − t) = 0
j
j=1 j=l n∈Z i=1
⎡ ⎤

l
r
⇔ ⎣ c j z −n ⎦ βdi (n − t) = 0.
j
n∈Z j=1 i=1
r
As i=1 βdi (n − t) are linearly independent, we obtain

l
c j z −n
j = 0.
j=1
This is a linear system of equation in the variable c1 , c2 , . . . , cl with coefficient

matrix, the Vandermonde’s determinant. Therefore c j = 0.
Hence, the functions ϒz −1 (t), j = 1, 2, . . . , l form a basis of .
j
Theorem 3 Let di ∈ N and h(t) be an integrable function satisfying condition (1).

If the roots of G(z) are simple and no roots on the unit circle |z| = 1, then for a given
sequence of numbers {yn }n∈Z ∈ Dγ , there exists a unique f ∈ S N ,γ , such that
f h(n) = yn , n ∈ Z. (3)
Moreover, the solution can be written as

f (t) = yn L(t − n),
n∈Z
r r
where the reconstruction function L is given by L(t) := L i (t) :=
i=1 i=1
n∈Z cn βdi (t − n) and cn are the coefficients of the Laurent series expansion of
G(z)−1 . Further the reconstruction function L is of exponential decay.

Proof Let C(z)
= G −1 (z) = n∈Z cn z n . Then there exist μ ∈ (0, 1) such that
cn = O μ|n| . As βdi has compact support, we obtain that O(L) = O(μ|t| ). Now
for |t| > 2, we have
γ μ|t−n|
γ μ|[t]−n|−1
n∈Z |n| n∈Z |n|
≤
(|t| + 1)γ (|[t]|)γ

(|[t] − n|)γ μ|n|−1
= n∈Z
(|[t]|)γ

≤ (1 + |n|)γ μ|n|−1
n∈Z
< ∞.
As a consequence of the above inequality we obtain that

f (t) = yn L(t − n) = O(|t|γ ),
n∈Z
as t → ±∞. Since yn L(t − n) = O(|n|γ μ|t−n| ), it is easy to see that the series

yn L(t − n)
n∈Z
converges uniformly and absolutely in every finite interval. Also,

f (t) = yn L(t − n)
n∈Z

r
= yn ck βdi (t − n − k)
n∈Z i=1 k∈Z

r
= yn ck−n βdi (t − k).
k∈Z n∈Z i=1
Therefore f ∈ S N ,γ .
Using C(z)G(z) = 1, we obtain that

r
(h L)(n) = ck h βdi (n − k) = δ0 (n).
i=1 k∈Z

Hence f (t) = n∈Z yn L(t − n) satisfies
(h f )(n) = yn , n ∈ Z. (4)
Now we shall show the uniqueness. Assume that f, g ∈ S N ,γ satisfy (4). Then
f − g ∈ . Using Theorem 2, there exist a constant c j such that

l
f (t) − g(t) = c j ϒz −1 .
j
j=1
γ ).
As f, g ∈ S N ,γ , we get f (t) − g(t) =O(|t|
Using Lemma 2 and the behavior of ϒz −1 (t) at ±∞, we get c j = 0 and hence
j
f = g.

For di = 1, 2, 3, 4 we shall show that the roots of G(z) are simple and not on the
unit circle |z| = 1.
Proof (Main Theorem) As a consequence of Theorem 1 it is sufficient to prove

that, all the roots of G(z) are simple and not on the unit circle |z| = 1 for distinct
di = 1, 2, 3, 4.
We have G(z) = ri=1 G i (z). We can write

r
−li
G(z) = z 2 Pi (z)
i=1

di + 1 if di is odd
where li := and Pi (z) is a polynomial of degree li . Therefore,
di if di is even
−m
r
m−li −m
G(z) = z 2 z 2 Pi (z) = z 2 P(z),
i=1
where P(z) is a polynomial of degree m = max(l1 , l2 , . . . , lr ).

As di ’s are distinct, we can take d1 = 1, d2 = 2, d3 = 3, and d4 = 4. Therefore
m = 4 and we obtain
P(z) = z 2 G(z)

= z 4 h βd4 (2) + h βd3 (2) + z 3 h βd4 (1) + h βd3 (1) + h βd2 (1)

+ h βd1 (1) + z 2 h βd4 (0) + h βd3 (0) + h βd2 (0) + h βd1 (0)

+ z h βd4 (−1) + h βd3 (−1) + h βd2 (−1) + h βd1 (−1) + h βd4 (−2)

+ h βd3 (−2) .
Hence P(0) > 0 and P(1) > 0.

We can write
4
1
2
P(z) = z 2 h(t)ϒz,di (t)dt. (5)
i=1 − 21
Using Lemma 2 and Eq. (5) we get
4
1
2
P(−1) = h(t)ϒ−1,di (t)dt > 0.
i=1 − 21
Since lim z−→∞ P(z) = ∞, It is suffices to find z 0 ∈ (−1, 0) such that

4
1 1
ϒz 0 ,di (t) < 0, for all t ∈ − , , (6)
2 2
i=1
since for such a z 0 , we have
4
1
2
P(z 0 ) = z 02 h(t)ϒz 0 ,di (t)dt < 0, z 0 ∈ (−1, 0)
i=1 − 21
4 1 4 1
1 1 2 1 2
and P = 2 h(t)ϒz −1 ,di (t)dt = 2 h(t)ϒz 0 ,di (−t)dt < 0
z0 z 0 i=1 − 21 0 z 0 i=1 − 21
4
and z 0−1 ∈ (−∞, −1). By solving i=1 ϒz 0 ,di 21 = 0, we get a unique z 0 ∈

(−1, 0).

4
1 1 1 1 1
Now ϒz 0 ,di = 0 ⇔ ϒz 0 ,1 + ϒz 0 ,2 + ϒz 0 ,3 + ϒz 0 ,4 =0
2 2 2 2 2
i=1

3 3 1 1 1 1
⇔ z 03 β4 + β3 + z 02 β4 + β3 + β2 + β1
2 2 2 2 2 2

1 1 1 1
+ z 0 β4 − + β3 − + β2 − + β1 −
2 2 2 2

3 3
+ β4 − + β3 − =0
2 2
3 93 93 3
⇔ z 03 + z 02 + z0 + = 0.
48 48 48 48
√ √
The possible solutions of z√ 0 are −1, −15 − 4 14, −15 + 4 14. The unique solution
z 0 ∈ (−1, 0) is −15 + 4 14. For this z 0 value

4
1 1
ϒz 0 ,di (t) < 0, for all t ∈ − , .
2 2
i=1
Thus we can conclude that all the roots of G(z) are simple and not on the unit circle
|z| = 1 for di = 1, 2, 3, 4.
Remark 1 The condition that the zeros of the Laurent polynomial G(z) are simple
and do not lie on the unit circle |z| = 1 is a sufficient condition for uniqueness of
solution for the local average sampling problem.
3 Conclusion
We proved local average sampling theorem over certain subspaces of the multiply
generated spline spaces of polynomial growth. Let h(t) be an integrable function
satisfying condition (1). We have shown that if the roots of G(z) are simple and no
roots on the unit circle |z| = 1, then for a given sequence of numbers {yn }n∈Z ∈ Dγ ,
there exists a unique f ∈ S N ,γ such that f h(n) = yn , n ∈ Z, for the distinct
di ≤ 4. Also, we have shown that the roots of G(z) are simple and not on the unit
circle |z| = 1, for di ≤ 4. We could not find a proof for di ≥ 5.
Acknowledgments The authors thank Anna University, Chennai-25, India for providing the Anna
Centenary Research fellowship to the second author to carry out this research work.
References
1. Aldroubi, A., Sun, Q., Wai-Shing, T.: Nonuniform average sampling and reconstruction in
multiply generated shift-invariant spaces. Constr. Approx. 20, 173–189 (2004)
2. Garcia, A.G., Hernandez-Medina, M.A., Perez-Villalon, G.: Generalized sampling in shift-
invariant spaces with multiple stable generators. J. Math. Anal. Appl. 337(1), 69–84 (2008)
3. de Boor, C., Devore, R., Ron, A.: The structure of finitely generated shift invariant spaces in
L 2 (R d ). J. Funct. Anal. 119, 37–78 (1994)
4. Gröcheni: Reconstruction algorithms in irregular samplings. Math. Comput. 59, 181–194
(1992)
5. Xian, J., Zhu, J., Wang, J.: Local sampling and reconstruction in spline signal spaces and multi-
ply generated spline signal spaces. International Joint Conference on Computational Sciences
and Optimization, 2000
6. Xian, J., Sun, W.: Local sampling and reconstruction in shift-invariant spaces and their appli-
cations in spline subspaces. Numer. Funct. Anal. Optim. 31(3), 366–386 (2010)
7. Xian, J., Qiang, X.: Non-uniform sampling and reconstruction in weighted finitely generated
shift-invariant spaces. Far East. J. Math. Sci. 8(3), 281–293 (2003)
8. Butzer, P.L., Engels, W., Ries, S., Stens, R.L.: The shannon sampling series and the recon-
struction of signals in terms of linear, quadratic and cubic splines. SIAM J. Appl. Math. 46(2),
299–323 (1986)
9. Schoenberg, I.J.: Cardinal spline interpolation. SIAM Regional Conference Series in Applied
Mathematics, 1973
10. Perez-Villalon, G., Portal, A.: Reconstruction of splines from local average samples. Appl.
Math. Lett. 25, 1315–1319 (2012)
11. Perez-Villalon, G., Portal, A.: Sampling in shift-invariant spaces of functions with polynomial
growth. Appl. Anal. 92(12), 2536–2546 (2013)
Approximation of Periodic Functions
Belonging to W (L r , ξ(t), (β ≥ 0))-Class
By (C 1 · T ) Means of Fourier Series
Smita Sonker
Abstract Various investigators such as Khan [3], Qureshi [8–10], Qureshi and Nema
[11], Leindler [6] and Chandra [1] have determined the degree of approximation of
functions belonging to the classes W (L r , ξ(t)), Li p(ξ(t), r ), Li p(α, r ) and Li pα
using different summability methods with monotonocity conditions. Recently, Lal
[5] has determined the degree of approximation of the functions belonging to Li pα
and W (L r , ξ(t)) classes by using Cesàro-Nörlund (C 1 · N p )—summability with non-
increasing weights { pn }. In this paper, we shall determine the degree of approximation
of 2π -periodic function f belonging to the function classes Li pα and W (L r , ξ(t))
by C 1 · T —means of Fourier series of f . Our theorems generalize the results of Lal
[5], and we also improve these results in the light of [7, 12, 13]. From our results,
we derive some corollaries also.
Keywords Trigonometric fourier series · W (L r , ξ(t), (β ≥ 0))-class · Fourier

series · Matrix means · Signals · Trigonometric polynomials
1 Introduction
For a given signal f ∈ L r := L r [0, 2π ], r ≥ 1, let
a0
n n
sn ( f ) := sn ( f ; x) = + (ak cos kx + bk sin kx) = u k ( f ; x), (1)
2
k=1 k=0
denote the partial sums, called trigonometric polynomial of degree (or order) n, of the
first (n + 1) terms of the Fourier series of f . The matrix means of (1) are defined by
S. Sonker (B)
Department of Mathematics, National Institute of Technology Kurukshetra,
Kurukshetra 136119, India
e-mail: smita.sonker@gmail.com

DOI 10.1007/978-81-322-2485-3_6
74 S. Sonker

n
tn ( f ) := tn ( f ; x) = an,k sk , n = 0, 1, 2, ...,
k=0
where T ≡ (an,k ) isa lower triangular matrix with non-negative entries such that
an,−1 = 0, An,k = rn=k an,r so that An,0 = 1, ∀n ≥ 0. The Fourier series of f is
said to be T -summable to s, if tn ( f ) → s as n → ∞.
By superimposing C 1 summability upon T summability, we get the C 1 · T sum-
mability. Thus the C 1 · T means of {sn ( f )} denoted by tnC ·T ( f ) are given by
1
n
r
tnC ·T ( f ) −1
1
:= (n + 1) ar,k sk ( f ) .
r =0 k=0
If tnC ·T → s1 as n → ∞, then the Fourier series of f is said to be C 1 · T —summable

1
to the sum s1 . The regularity of methods C 1 and T implies regularity of method C 1 ·T.
A function f ∈ Li pα if | f (x + t)− f (x)| = O(|t|α ), for 0 < α ≤ 1, f ∈ Li p(α, r )
1/r
= O(|t|α ), 0 < α ≤ 1, r ≥ 1,
2π
if 0 | f (x + t) − f (x)|r d x
1/r
2π
f ∈ Li p(ξ(t), r ) if 0 | f (x + t) − f (x)|r d x = O(ξ(t)) and
1/r
f ∈ W (L r , ξ(t)) if 0 |( f (x + t) − f (x)) sinβ (x/2)|r d x
2π
= O(ξ(t)),
β ≥ 0, r ≥ 1, where ξ(t) is a positive increasing function of t.
If β = 0, W (L r , ξ(t), ) ≡ Li p(ξ(t), r ) and for ξ(t) = t α (α > 0), Li p(ξ(t), r ) ≡
Li p(α, r ). Li p(α, r ) → Li pα as r → ∞. Thus
Li pα ⊆ Li p(α, r ) ⊆ Li p(ξ(t), r ) ⊆ W (L r , ξ(t)).
The L r -norm of f ∈ L r [0, 2π ] is defined by

1/r
2π

f
r = 2π 1
0 | f (x)| d x
r (1 ≤ r < ∞) and
f
∞ = sup | f (x)|.
x∈[0,2π ]
The degree of approximation of f ∈ L r denoted by E n ( f ) is given by
E n ( f ) = min
f (x) − Tn (x)
r ,
Tn
in terms of n , where Tn (x) is a trigonometric polynomial of degree n.

This method of approximation is called trigonometric Fourier approximation.
We also write
φ(t) = f (x + t) + f (x − t) − 2 f (x),
1 n r
sin(r − k + 1/2)t
(C · T )n (t) =
1
ar,r −k ,
2π(n + 1) sin(t/2)
r =0 k=0
and τ = [1/t], the integral part of 1/t.

Approximation of Periodic Functions … 75
2 Known Results
Various investigators such as Khan [3], Qureshi [8–10], Qureshi and Nema [11],
Leindler [6] and Chandra [1] have determined the degree of approximation of func-
tions belonging to the classes W (L r , ξ(t)), Li p(ξ(t), r ), Li p(α, r ) and Li pα with
r ≥ 1 and 0 < α ≤ 1 using different summability methods with monotonocity con-
ditions on the rows of summability matrices. Recently, Lal [5] has determined the
degree of approximation of the functions belonging to Li pα and W (L r , ξ(t)) classes
by using Cesáro-Nörlund (C 1 · N p )—summability with non-increasing weights { pn }.
He proved:
Theorem 1 Let N p be a regular Nörlund method defined by a sequence { pn } such

that
n
Pτ Pv−1 = O(n + 1). (2)
v=τ
Let f ∈ L 1 [0, 2π ] be a 2π -periodic function belonging to Li p α (0 < α ≤ 1), then

the degree of approximation of f by C 1 · N p means of its Fourier series is given by

O((n + 1)−α ), 0 < α < 1,
sup |tnC N (x) − f (x)| =
tnC N − f
∞ =
0≤x≤2π O (log(n + 1)π e/(n + 1)) , α = 1.
Theorem 2 If f is a 2π -periodic function and Lebesgue integrable on [0, 2π ] and

is belonging to W (L r , ξ(t)) class then its degree of approximation by C 1 · N p means
of its Fourier series is given by

tnC N − f
r = O (n + 1)β+1/r ξ (1/(n + 1)) ,
provided ξ(t) satisfies the following conditions:
{ξ(t)/t} be a decreasing sequence, (3)

1/r
π/(n+1) r
β
t|φ(t)| sin (t)/ξ(t) dt = O((n + 1)−1 ), (4)
0
π r 1/r
t −δ |φ(t)|/ξ(t) dt = O((n + 1)δ ), (5)
π/(n+1)
where δ is an arbitrary number such that s(1 − δ) − 1 > 0, r −1 + s −1 = 1, r ≥ 1,

conditions (4) and (5) hold uniformly in x.
The improved version of above theorems with their generalization to non-monotone

weights { pn } can be seen in [13].
76 S. Sonker
3 Main Results
In this paper, we generalize Theorems 1 and 2 by replacing matrix N p with matrix

T in the light of Remarks 2.3 and 2.4 of [13, pp. 3–4]. More precisely, we prove:
Theorem 3 If T ≡ (an,k ) is a lower triangular regular matrix with non-negative

and non-decreasing (with respect to k) entries which satisfy

n
Ar,r −τ = O(n + 1), (6)
r =τ
hold uniformly in τ = [1/t], then the degree of approximation of a 2π -periodic

function f ∈ Li pα (0 < α ≤ 1) ⊂ L 1 [0, 2π ] by C 1 · T means of its Fourier series
is given by

1 ·T O((n + 1)−α ), 0 < α < 1,

tnC ( f ) − f (x)
∞ = (7)
O ((log(n + 1))/(n + 1)) , α = 1.
Theorem 4 If T ≡ (an,k ) be a lower triangular regular matrix with non-negative

and non-decreasing (with respect to k) entries which satisfy condition (6), then the
degree of approximation of a 2π -periodic function with r > 1 and 0 < β s < 1 by
C 1 · T means of its Fourier series is given by
1 ·T

tnC ( f ) − f (x)
r = O (n + 1)β+1/r ξ (1/(n + 1)) , (8)
provided positive increasing function ξ(t) satisfies the conditions:
ξ(t)/t be a decreasing function, (9)

1/r
π/(n+1) r
β
|φ(t)| sin (t/2)/ξ(t) dt = O((n + 1)−1/r ), (10)
0
π 1/r
−δ r
t |φ(t)| sinβ (t/2)/ξ(t) dt = O((n + 1)δ−1/r ), (11)
π/(n+1)
where δ is a real number such that β + 1/r > δ > r −1 , r −1 + s −1 = 1, r > 1. Also,
conditions (10) and (11) hold uniformly in x.
Remark1 If we take an,k = pn−k /Pn for k ≤ n and an,k = 0 for k > n such that
Pn (= nk=0 pk = 0) → ∞ as n → ∞ and P−1 = 0 = p−1 , then C 1 · T means
reduce to C 1 · N p means and

n
n
r
n
r
n
n
Ar,r −τ = ar,k = ( pr −k /Pr ) = (Pτ /Pr ) = Pτ Pr−1 .
r =τ r =τ k=r −τ r =τ k=r −τ r =τ r =τ
Therefore, condition (6) reduces to condition (2) and tnC ·T means reduce to tnC N
1
means. Hence our Theorems 3 and 4 are generalization of Theorems 1 and 2,

respectively.
4 Lemmas
We need the following lemmas for the proof of our theorems.
Lemma 1 Let {ar,k } be a non-negative sequence of real numbers, then
(C 1 · T )n (t) = O(n + 1), for 0 < t ≤ π/(n + 1).
Proof Using sin nt ≤ nt and sin(t/2) ≥ t/π for 0 < t ≤ π/(n + 1), we have
n r

1
(C · T )n (t) = (2π(n + 1))−1 ar,r −k sin((r − k + 1/2)t)/ sin(t/2)

r =0 k=0

n
r
= (2π(n + 1))−1 ar,r −k |sin((r − k + 1/2)t)/ sin(t/2)|
r =0 k=0
n r
≤ (2π(n + 1))−1 ar,r −k (r − k + 1/2)t/(t/π )
r =0 k=0

n r
≤ (4(n + 1))−1 ar,r −k (2r − 2k + 1)
r =0 k=0
n
r
≤ (4(n + 1))−1 (2r + 1) ar,r −k
r =0 k=0
n
= (4(n + 1))−1 (2r + 1)Ar,0 = O(n + 1).
r =0
Lemma 2 [4] If {ar,k } is a non-negative and non-decreasing (with respect to k)

sequence, then for 0 ≤ a < b ≤ ∞, 0 < t ≤ π and for every r
b

ar,r −k ei(r −k)t = O(Ar,r −τ ).

k=a
78 S. Sonker
Lemma 3 If {ar,k } is non-negative and non-decreasing (with respect to k) sequence,

then for 0 < t ≤ π
n r n

i(r −k)t −1
ar,r −k e = O(t ) + O Ar,r −τ ,

r =0 k=0 r =τ
holds uniformly in τ = [1/t].
Proof For 0 < t ≤ π, we have

n r τ −1 r
n r
i(r −k)t i(r −k)t i(r −k)t
ar,r −k e ≤ ar,r −k e + ar,r −k e
r =τ k=0

r =0 k=0 r =0 k=0

τ −1
r n r

≤ ar,r −k |ei(r −k)t | + ar,r −k ei(r −k)t

r =0 k=0 r =τ k=0

n

τ −1
r r

≤ ar,r −k + ar,r −k ei(r −k)t

r =τ k=0

r =0 k=0
τ −1
n
n
≤ 1+ O(Ar,r −τ ) = (τ − 1 + 1) + O Ar,r −τ
r =0 r =τ r =τ
n

= O(t −1 ) + O Ar,r −τ ,
r =τ
in view of Lemma 2.
Lemma 4 If {ar,k } is non-negative and non-decreasing (with respect to k) sequence

and satisfies the condition (6), then

|(C 1 · T )n (t)| = O t −2 /(n + 1) + O(t −1 ), for π/(n + 1) < t ≤ π.
Proof Using sin(t/2) ≥ t/π , for π/(n + 1) < t ≤ π and Lemma 3, we have
n r

−1
|(C · T )n (t)| =
1
(2π(n + 1)) ar,r −k sin((r − k + 1/2)t)/ sin(t/2)

r =0 k=0
n r

≤ (2π(n + 1))−1 ar,r −k sin((r − k + 1/2)t)/(t/π )

r =0 k=0
n r

= (2t (n + 1))−1 ar,r −k sin(r − k + 1/2)t

r =0 k=0
n r

≤ (2t (n + 1))−1 ar,r −k ei(r −k+1/2)t

r =0 k=0

n r

= (2t (n + 1))−1 eit/2 ar,r −k ei(r −k)t

r =0 k=0
n r

= (2t (n + 1))−1 ar,r −k ei(r −k)t

r =0 k=0
n

−1 −1
= (2t (n + 1)) O(t ) + O Ar,r −τ = O t −2 /(n + 1) + O(t −1 ),

r =τ
in view of condition (6).
5 Proof of Theorem 3
Following Titchmarsh [14], we have

π
1
sn ( f ; x) − f (x) = φ(t)(sin(n + 1/2)t/ sin(t/2))dt
2π 0
1 ·T
Denoting C 1 · T means of {sn ( f ; x)} by tnC ( f ), we write
π
n
r
1 ·T
tnC ( f ) − f (x) = φ(t)(2π(n + 1))−1 ar,r −k sin((r − k + 1/2)t)/ sin(t/2)dt
0 r =0 k=0
π/(n+1) π
= φ(t)(C 1 · T )n (t)dt + φ(t)(C 1 · T )n (t)dt
0 π/(n+1)
= I1 + I2 , say. (12)
Using Lemma 1 and the fact that f ∈ Li p α ⇒ φ ∈ Li p α {[2], Lemma 5.27}, we

have
π/(n+1) π/(n+1)
|I1 | ≤ |φ(t)||(C 1 · T )n (t)|dt = O(n + 1) t α dt
0 0
= O(n + 1)((n + 1)−α−1 ) = O((n + 1)−α ). (13)
Now, using Lemma 4 and the fact that f ∈ Li p α ⇒ φ ∈ Li p α,

π π
|I2 | ≤ |φ(t)| (C 1 · T )n (t) dt ≤ |φ(t)|O (t −2 /(n + 1)) + t −1 dt
π/(n+1) π/(n+1)
= O(I21 ) + O(I22 ), say, (14)
where
π
O((n + 1)−α ), 0 < α < 1,
I21 = (n + 1)−1 t α−2 dt = (15)
π/(n+1) O (log(n + 1)/(n + 1)) , α = 1.
80 S. Sonker
and π
I22 = O t α−1 dt = O((n + 1)−α ). (16)
π/(n+1)
Collecting (12)–(16), we get

O((n + 1)−α ), 0 < α < 1,
tnC ·T ( f ) −
1
f (x) =
O(log(n + 1)/(n + 1)), α = 1.
Thus

1 ·T 1 ·T O((n + 1)−α ), 0 < α < 1,

tnC ( f ) − f
∞ = sup {|tnC (x) − f (x)|} =
0≤x≤2π O((log(n + 1))/(n + 1)), α = 1.
6 Proof of Theorem 4
Following the proof of Theorem 3,

π/(n+1) π
1 ·T
tnC ( f ) − f (x) = φ(t)(C 1 · T )n (t)dt + φ(t)(C 1 · T )n (t)dt
0 π/(n+1)

= I1 + I2 , say. (17)
Using Hölder’s inequality, φ(t) ∈ W (L r , ξ(t)), condition (10), Lemma 1 and mean
value theorem for integrals, we have
π/(n+1)

β β
|I1 | = lim (φ(t) sin (t/2)/ξ(t)) · (ξ(t)(C · T )n (t))/(sin (t/2)) dt
1
ε→0 ε
1/r
π/(n+1) r
≤ |φ(t)| sinβ (t/2)/ξ(t) dt
0
1/s
π/(n+1) s
β
· lim ξ(t)|(C · T )n (t)|/(sin (t/2)
1
dt
ε→0 ε
1/s
π/(n+1)
= O((n + 1) −1/r
) lim ξ(t)(n + 1)/(sin (t/2))s dt
β
ε→0 ε
1/s
π/(n+1)
−βs
= O(n + 1) 1−1/r
(ξ(π/(n + 1)) lim t dt
ε→0 ε
= O(ξ(1/(n + 1)(n + 1)β+1−1/r −1/s ) = O((n + 1)β ξ(1/(n + 1)), (18)

in view of condition (9), i.e. (ξ(π/(n + 1))/(π/(n + 1))) ≤ (ξ(1/(n + 1))/

(1/(n + 1))).
Using Lemma 4, we have
π

|I2 | = |φ(t)| O t /(n + 1) + O t −1 dt
−2
π/(n+1)
π π
−2 −1
=O t |φ(t)|/(n + 1)dt + O t |φ(t)|dt
π/(n+1) π/(n+1)

= O(I21 ) + O(I22 ). (19)
Using Hölder’s inequality, | sin t| ≤ 1, sin(t/2) ≥ (t/π ) and condition (11), we

have
π

|I21 | = (n + 1)−1 (t −δ |φ(t)| sinβ (t/2)/ξ(t)) · (ξ(t)/(t −δ+2 sinβ (t/2))) dt
π/(n+1)
1/r
π −δ
≤ ((n + 1)−1 ) t |φ(t)| sinβ (t/2)/ξ(t)r dt
π/(n+1)
π s 1/s

· ξ(t)/ t −δ+2 sinβ (t/2) dt
π/(n+1)
1/r
π −δ
= O((n + 1)−1 ) t |φ(t)| sinβ (t/2)/ξ(t)r dt
π/(n+1)
π s 1/s
−δ+2 β
· ξ(t)/ t sin (t/2) dt
π/(n+1)
π s 1/s

= O((n + 1)−1 )O (n + 1)δ−1/r ξ(t)/ t −δ+2 sinβ (t/2) dt
π/(n+1)
π s 1/s
= O((n + 1)δ−1−1/r ) ξ(t)/t −δ+2+β dt
π/(n+1)
π 1/s
= O((n + 1)δ−1/r )ξ(π/(n + 1)) t −(−δ+1+β)s dt
π/(n+1)

= O((n + 1)δ−1/r )ξ(π/(n + 1)) (n + 1)(−δ+1+β)−1/s dt
= O(ξ(1/(n + 1)(n + 1)β ) (20)
in view of decreasing nature of ξ(t)/t and r −1 + s −1 = 1.

Similarly, as above, we have

π π −δ
|I22 | = t −1 |φ(t)|dt = t |φ(t)| sinβ (t/2)/ξ(t) ξ(t)/(t 1−δ sinβ (t/2)) dt
π/(n+1) π/(n+1)
1/r s 1/s
π −δ π

≤ t |φ(t)| sinβ (t/2)/ξ(t)r dt ξ(t)/ t 1−δ sinβ (t/2) dt
π/(n+1) π/(n+1)
π s 1/s
= O (n + 1)δ−1/r ξ(t)/t 1−δ+β dt
π/(n+1)
82 S. Sonker
π 1/s
= O (n + 1)δ+1−1/r ξ(1/(n + 1)) t (δ−β)s dt
π/(n+1)

δ+1−1/r
= O (n + 1) ξ(1/(n + 1))(n + 1)(−δ+β)−1/s
= O(ξ(1/(n + 1))(n + 1)β+1−1/r −1/s

= O(ξ(1/(n + 1))(n + 1)β . (21)
Collecting (17)–(21), we have

1 ·T
|tnC ( f ) − f (x)| = O (n + 1)β ξ(1/(n + 1)) .
Hence,
1/r
1 ·T
2π 1 ·T

tnC ( f ) − f (x)
r = 1/2π |tnC ( f ) − f (x)|r d x = O (n + 1)β ξ (1/(n + 1) .
0
Remark 2 The proof of Theorem 3, for r = 1, i.e. s = ∞ can be written by using

sup norm while using Hölder’s inequality.
7 Corollaries
The following corollaries can be derived from Theorem 4

1. If β = 0 , then for f ∈ Li p(ξ(t), r ),
tnC ·T ( f ) − f (x)
r = O (ξ(1/n)) .
1
2. If β = 0, ξ(t) = t α (0 < α ≤ 1), then for f ∈ Li p(α, r )(α > 1/r ),

1 ·T

tnC ( f ) − f (x)
r = O n −α . (22)
3. If r → ∞ in Corollary 2, then for f ∈ Li pα(0 < α < 1), (22) gives

1 ·T

tnC ( f ) − f (x)
∞ = O(n −α ).
Remark 3 In view of Remark 2, corollaries of Lal [5, p. 350] are particular cases of
our Corollaries 2 and 3, respectively.
References
1. Chandra, P.: Trigonometric approximation of functions in L p -norm. J. Math. Anal. Appl.

275(1), 13–26 (2002)
2. Faddeen, L.M.: Absolute Nörlund summability. Duke Math. J. 9, 168–207 (1942)
3. Khan, H.: On the degree of approximation of functions belonging to the class Li p(α, p),
Indian. J. Pure Appl. Math. 5(2), 132–136 (1974)
4. Kishore, N., Hotta, G.C.: On absolute matrix summability of Fourier series Indian. J. Math.
13(2), 99–110 (1971)
5. Lal, S.: Approximation of functions belonging to the generalized Lipschitz Class by C 1 .N p
summability method of Fourier series. Appl. Math. Comput. 209, 346–350 (2009)
6. Leindler, L.: Trigonometric approximation in L p -norm. J. Math. Anal. Appl. 302(1), 129–136
(2005)
7. Lenski, W., Szal, B.: Approximation of functions belonging to the class L p (w)β by linear
operators. Acta ET Commentationes Universitatis Tartuensis De Mathematica 13, 11–24 (2009)
8. Qureshi, K.: On the degree of approximation of a periodic function f by almost Nörlund means.
Tamkang J. Math. 12(1), 35–38 (1981)
9. Qureshi, K.: On the degree of approximation of a function belonging to the Class Li pα, Indian.
J. Pure Appl. Math. 13(8), 560–563 (1982)
10. Qureshi, K.: On the degree of approximation of a function belonging to weighted W (L p , ξ(t))
- class. Indian J. Pure Appl. Math. 13(4), 471–475 (1982)
11. Qureshi, K., Nema, H.K.: On the degree of approximation of functions belonging to weighted
class. Ganita 41(1–2), 17–22 (1990)
12. Rhoades, B.E., Ozkoklu, K., Albayrak, I.: On the degree of approximation of functions belong-
ing to a Lipschitz class by Hausdorff means of its Fourier series. Appl. Math. Comput. 217,
6868–6871 (2011)
13. Singh, U., Mittal, M.L., Sonker, S.: Trigonometric approximation of signals (functions) belong-
ing to W (L p , ξ(t))−class by matirx (C 1 .N p ) operator. Int. J. Math. Math. Sci. 2012, 1–11
(2012)
14. Titchmarsh, E.C.: The Theory of Functions, pp. 402–403. Oxford University Press, Oxford
(1939)
Modified Baskakov-Szász Operators
Based on q-Integers
P.N. Agrawal and Arun Kajla
Abstract In the present paper we introduce the Stancu variant of certain q-modified
Baskakov Szász operators. We estimate the moments of the operators and obtain some
direct results in terms of the modulus of continuity. Then, we study the Voronovskaja
type theorem and the rate of convergence of these operators in terms of the weighted
modulus of continuity. Further, we discuss the point-wise estimation using the Lip-
schitz type maximal function. Finally, we investigate the rate of statistical conver-
gence of these operators using weighted modulus of continuity.
Keywords q-Baskakov-Szasz operators · q-integers · Modulus of smoothness ·

Point-wise estimates · Statistical convergence
Mathematics Subject Classification (2010): 26A15 · 40A35
1 Introduction
In recent years, the most interesting area of research in approximation theory is

the application of q-calculus. In 1997, Phillips [20] first considered a modification
of Bernstein polynomials based on q-integers. He studied the rate of convergence
and Voronovskaja-type asymptotic formula for these operators. Very recently, Gupta
and Kim [14] considered q-Baskakov operators and they obtained some direct local
results and the degree of approximation in terms of modulus of continuity. Subse-
quently, several researchers have considered the different types of operators in this
direction and studied their approximation properties.
Let α and β be any two real numbers satisfying the condition that 0 ≤ α ≤ β,
Stancu [21] defined in the following operators:
P.N. Agrawal · A. Kajla (B)

Department of Mathematics, Indian Institute of Technology Roorkee, Roorkee 247667, India
e-mail: rachitkajla47@gmail.com
P.N. Agrawal
e-mail: pna_iitr@yahoo.co.in
DOI 10.1007/978-81-322-2485-3_7
86 P.N. Agrawal and A. Kajla

n
k +α
Snα,β (f , x) = pn,k (x) , 0 ≤ x ≤ 1,
n+β
k=0
where pn,k (x) is the Bernstein basis function.

Recently, Büyükyazici [7] considered the Stancu–Chlodowsky polynomials and
investigated their convergence. In 2012, Verma et al. [22] introduced a Stancu type
generalization of certain q-Baskakov Durrmeyer operators and discussed some local
direct results of these operators. For some other research papers where Stancu type
operators have been considered, we refer to [1, 3, 4, 13, 15], etc.
Now, we give some basic definitions and concepts of q-calculus [6, 17]. For any
real number q > 0, the q-integer [n]q and q-factorial [n]q ! are defined as
⎧
⎨ (1 − q )
n
, if q = 1
[n]q = (1 − q)
⎩ n, if q = 1
and

[n]q [n − 1]q . . . 1, n = 1, 2, . . .
[n]q ! =
1, n = 0.
The q-Pochhammer symbol is defined as

(1 + x)(1 + qx) . . . (1 + qn−1 x), n = 1, 2, . . .
(x; q)n =
1, n = 0.
The q-binomial coefficients are given by

n [n]q !
= , 0 ≤ k ≤ n.
k q [k]q ![n − k]q !
The q-derivative Dq of a function f is given by
f (x) − f (qx)
(Dq f )(x) = , if x = 0.
(1 − q)x
The q-Jackson integrals and q-improper integrals are defined as

a ∞

f (x)dq (x) = (1 − q)a f (aqn )qn , a > 0,
0 n=0
and
∞/A ∞

qn qn
f (x)dq (x) = (1 − q) f , A > 0.
0 n=−∞
A A
Modified Baskakov-Szász Operators Based on q-Integers 87
The q-Beta integral is defined by

1
1−q
Γq (t) = x t−1 Eq (−qx)dq x, t > 0 (1)
0
which satisfies the following functional equation:
Γq (t + 1) = [t]q Γq (t), Γq (1) = 1.
To approximate Lebesgue integrable functions on the interval [0, ∞), Agrawal and
Mohammad [2] introduced the following operators:
∞
∞
Mn (f (t); x) = n pn,v (x) qn,v−1 (t)f (t)dt + (1 + x)−n f (0). (2)
v=1 0
where

n + v−1 v
pn,v (x) = x (1 + x)−(n + v) , x ∈ [0, ∞)
v
and
e−nt (nt)v
qn,v (t) = , ∀ t ∈ [0, ∞).
v!
In [2], Agrawal et al. studied the asymptotic approximation and error estimates in
terms of modulus of continuity in simultaneous approximation by (2).
In [16], Gupta and Srivastava considered a sequence of positive linear operators
combining the Baskakov and Szász basis functions. Deo [8] studied the simultaneous
approximation by Lupas operators with the weight functions of Szász operators.
Definition 1 For f ∈ Cγ [0, ∞) := {f ∈ C[0, ∞) : f (t) = O(eγ t ) as t → ∞

for some γ > 0} and each positive integer n, the q-Baskakov operators [5] are
defined as
∞
n + k−1 k(k−1) xk [k]q
Vn,q (f ; x) = q 2 f (3)
k=0
k q (1 + x)qn + k qk−1 [n]q
∞
q [k]q
= pn,k (x)f k−1 .
q [n]q
k=0
Remark 1 The first three moments of the q-Baskakov operators (see [5]) are
given by

x x
Vn,q (1; x) = 1, Vn,q (t; x) = x, Vn,q (t ; x) = x + 1+ 2
. 2
[n]q q
Definition 2 For f ∈ Cγ [0, ∞), 0 < q < 1 and each positive integer n, the
q-Baskakov Szász operators defined as
∞
q
(1−qn ) t
q−k−1 sn,k (t)f
q q
Bn,q (f ; x) = [n]q pn,k (x) dq t, (4)
0 qk
k=0

q n + k − 1 k(k−1) xk
where pn,k (x) = q 2
k (1 + x)(n
q
+ k)
q ([n]q t)k
and sn,k (t) = Eq (−[n]q t) (5)
[k]q !
have been considered by Gupta [12].
2 Construction of Operators
For f ∈ Cγ [0, ∞), 0 < q < 1 and each positive integer n, the Stancu-type general-
ization of the operators (2) based on q-integers is defined as follows:
∞
q
(α,β) (1−qn ) [n]q tq−k + α
q−k sn,k−1 (t)f
q q
Mn,q (f ; x) = [n]q pn,k (x) dq t
0 [n]q + β
k=1

α q
+f pn,0 (x), (6)
[n]q + β
q q
where pn,k (x) and sn,k (t) are as defined in (5).
If α = β = 0 and q → 1−, the operators (6) reduce to the operators (2), which
is a modification of the operator given by (4) where the value of the function at zero
is considered explicitly. The aim of this paper is to study some direct results and
asymptotic formula for the operators (6). We also discuss the rate of convergence
and point-wise estimation. Lastly, we study the statistical approximation properties
of these operators.
3 Basic Results
3.1 Moment Estimates

(α,β)
For α = β = 0, we denote the operator Mn,q by Mn,q .
Lemma 1 For the operators Mn,q (f ; x), the following equalities hold:
(i) Mn,q (1; x) = 1;
(ii) Mn,q (t; x) = x;
1 [2]q x
(iii) Mn,q (t 2 ; x) = x 2 1 + + .
q[n]q [n]q
Proof First, for f (t) = 1, we have

∞ q
(1−qn )
q−k sn,k−1 (t)dq t + pn,0 (x).
q q
Mn,q (1; x) = [n]q k=1 pn,k (x) 0
Substituting [n]q t = qy and using (1)
∞
1
1−q (qy)k−1 Eq [−qy]
q−k + 1
q q
Mn,q (1; x) = [n]q pn,k (x) dq y + pn,0 (x)
0 [k − 1]q ! [n]q
k=1
∞
q Γq k q
= pn,k (x) + pn,0 (x)
[k − 1]q !
k=1
∞
q
= pn,k (x)
k=0
= Vn,q (1; x) = 1, in view of Remark 1.
Next, let f (t) = t, we have

∞
q
(1−qn ) ([n]q t)k−1
q−2k t k Eq (−[n]q t)
q
Mn,q (t; x) = [n]q pn,k (x) dq t.
0 [k − 1]q !
k=1
Again, substituting [n]q t = qy and using (1)
∞
1
1−q (qy)k
q−2k + 1 Eq (−qy)
q
Mn,q (t; x) = [n]q pn,k (x) dq y
0 [k − 1]q !([n]q )2
k=1
∞
1
q 1 1−q
= pn,k (x) Eq (−qy)yk dq y
[n]q [k − 1]q !qk−1 0
k=1
∞
q Γq (k + 1)
= pn,k (x)
[n]q [k − 1]q !qk−1
k=1
∞
q [k]q
= pn,k (x)
[n]q qk−1
k=1
∞
q [k]q
= pn,k (x) = Vn,q (t; x) = x, on applying Remark 1.
[n]q qk−1
k=0
Finally, we give the second moment as follows:

∞
q 2
(1−qn ) t ([n]q t)k−1
q−k
q
Mn,q (t 2 ; x) = [n]q pn,k (x) Eq (−[n]q t)dq t.
0 qk [k − 1]q !
k=1
Again, substituting [n]q t = qy, using (1) and [k + 1]q = [k]q + qk , we have
∞
1
Eq (−qy) (qy)2 (qy)k−1 q
1−q
q−k
q
Mn,q (t 2 ; x) = [n]q pn,k (x) dq y
0 q2k ([n]q )2 [k − 1]q ![n]q
k=1
∞
1
1 1−q
Eq (−qy)yk + 1 dq y
q
= pn,k (x) 2k−2
q ([n]q ) [k − 1]q ! 0
2
k=1
∞
q 1
= pn,k (x) Γ (k + 2)q
q2k−2 ([n] q ) [k − 1]q !
2
k=1
∞
q 1
= pn,k (x) [k]q ([k]q + qk )
q2k−2 ([n]q )2
k=1
∞
∞
q ([k]q )2 q q [k]q
= pn,k (x) + pn,k (x) k−1
([n]q ) q
2 2k−2 [n]q q [n]q
k=1 k=1
q
= Vn,q (t 2 ; x) + Vn,q (t; x)
[n]q

x x qx
=x +
2
1+ +
[n]q q [n]q

1 [2] q x
= x2 1 + + , on using Remark 1.
q[n]q [n]q

(α,β)
Lemma 2 For Mn,q (t m ; x), m = 0, 1, 2 we have
(α,β)
(i) Mn,q (1; x) = 1;
(α,β) [n]q x + α
(ii) Mn,q (t; x) = ;
[n]q + β
(α,β) [n]q (1 + q[n]q )x 2 [n]q ([2]q + 2α)x α2
(iii) Mn,q (t 2 ; x) = + + .
q([n]q + β) 2 ([nq ] + β) 2 ([n]q + β)2
Proof Using Lemma 1, we estimate the moments as follows:
For f (t) = 1, we have
∞ qn
(α,β)
q−k sn,k−1 (t)dq t + pn,0 (x) = Mn,q (1; x) = 1.
q 1−q q q
Mn,q (1; x) = [n]q pn,k (x)
k=1 0
Next, we obtain the first-order moment

∞ q
(α,β) 1−qn [n]q tq−k + α α
q−k sn,k−1 (t)
q q q
Mn,q (t; x) = [n]q pn,k (x) dq t + pn,0 (x)
0 [n]q + β [n]q + β
k=1
[n]q α
= Mn,q (t; x) + Mn,q (1; x)
[n]q + β [n]q + β
[n]q α
= x +
[n]q + β ([n]q + β)
[n]q x + α
= .
[n]q + β
Finally, for f (t) = t 2 we obtain

∞
q 2
(α,β) 2 1−qn [n]q tq−k + α 2 α
q−k sn,k−1 (t)
q q q
Mn,q (t ; x) = [n]q pn,k (x) dq t + pn,0 (x)
0 [n]q + β [n]q + β
k=1
([n]q )2 2[n]q α α2
= Mn,q (t 2 ; x) + Mn,q (t, x) + Mn,q (1; x)
([n]q + β)2 ([n]q + β)2 ([n]q + β)2

[n]2q 1 x(1 + q) 2[n]q α α2
= x2 1 + + + x +
([n]q + β)2 q[n]q [n]q ([n]q + β)2 ([n]q + β)2
[n]q (1 + q[n]q ) 2 [n]q ([2]q + 2α) α2
= x + x + .
q([n]q + β) 2 ([n]q + β) 2 ([n]q + β)2
Hence, the proof is completed.
Remark 2 By simple computation, we have
(α,β) α − βx
Mn,q ((t − x); x) = ,
[n]q + β
(α,β) x 2 ([n]q + qβ 2 ) x([2]q [n]q − 2αβ) α2
Mn,q ((t − x)2 ; x) = + + .
q([n]q + β)2 ([n]q + β)2 ([n]q + β)2
Lemma 3 For every q ∈ (0, 1) we have

(α,β) [2]q (1 + β 2 ) 2 1
Mn,q ((t − x)2 ; x) ≤ φ (x) + ,
q([n]q + β) ([n]q + β)
√
where φ(x) = x(1 + x), x ∈ [0, ∞).
Proof
(α,β) x 2 ([n]q + qβ 2 ) x([2]q [n]q − 2αβ) α2

Mn,q ((t − x)2 ; x) = + +
q([n]q + β)2 ([n]q + β)2 ([n]q + β)2
([2]q [n]q + β 2 ) 2 α2
≤ (x + x) +
q([n]q + β) 2 ([n]q + β)2
[2]q ([n]q + β 2 ) 2 α2
≤ (x + x) +
q([n]q + β)2 ([n]q + β)2
[2]q [n]q (1 + β 2 ) 2 α2
≤ φ (x) +
q([n]q + β) 2 ([n]q + β)2
[2]q (1 + β 2 ) 2 α2
≤ φ (x) +
q([n]q + β) ([n]q + β)2

[2]q (1 + β 2 ) 2 1
≤ φ (x) + .
q([n]q + β) [n]q + β
4 Main Results
If q = {qn } be a sequence in (0, 1) satisfying the following conditions:
lim qn = 1 and lim qnn = c, (0 ≤ c < 1). (7)

n→∞ n→∞
(α,β)
Our first result is a basic convergence theorem for the operators Mn,qn .
Theorem 1 Let qn ∈ (0, 1) and lim qnn = c, (0 ≤ c < 1). Then the sequence
n→∞
(α,β)
Mn,qn (f ; x) converges to f uniformly on [0, A], A > 0, for each f ∈ Cγ [0, ∞) if
and only if lim qn = 1.
n→∞
(α,β)
Remark 3 If lim qn = 1, then in view of Remark 2, Mn,qn ((t − x)2 ; x) → 0 uni-
n→∞
formly on [0, A] as n → ∞. Therefore, the well-known Korovkin theorem implies
(α,β)
that {Mn,qn (f ; x)} converges to f uniformly on [0, A] for each f ∈ Cγ [0, ∞). The
converse part follows on proceeding in a manner similar to the proof of [3], Theo-
rem 1.
4.1 Direct Theorem
Let CB [0, ∞) be the space of all continuous and bounded functions f defined on the
interval [0, ∞), endowed with the norm . on the space given by
f = sup | f (x) | . (8)

0≤x<∞
If δ > 0 and W 2 = {g ∈ CB [0, ∞) : g

, g
∈ CB [0, ∞)}, then the K-functional is

defined as
K2 (f , δ) = inf{ f − g + δ g
: g ∈ W 2 }. (9)
By ([9], p. 177, Theorem

√ 2, 4) there exists an absolute constant C > 0 such that
K2 (f , δ) ≤ Cω2 (f , δ),
where second order modulus of the smoothness of f ∈ CB [0, ∞) is defined as
√
ω2 (f , δ) = sup√
sup | f (x + 2h) − 2f (x + h) + f (x) | .
0<h≤ δ 0≤x<∞
The first-order modulus of continuity is defined as
ω(f , δ) = sup
√
sup | f (x + h) − f (x) | .
0<h≤ δ 0≤x<∞
(α,β)
The next result is a direct local approximation theorem for the operators Mn,q .
Theorem 2 Let f ∈ CB [0, ∞) and let {qn } be sequence satisfying the conditions
(7). Then, for every x ∈ [0, ∞) we have

(α,β) 4(1 + β 2 ) 1 | α − βx |
| Mn,q (f ; x) − f (x) | ≤ Cω2 f , φ 2 (x) + + ω f, .
q([n]q + β) ([n]q + β) [n]q + β
(α,β)
Proof We introduce auxiliary operator Ln,q as follows:

(α,β) (α,β) α − βx
Ln,q (f ; x) = Mn,q (f ; x) − f x + + f (x). (10)
([n]q + β)
These operators are linear and preserve the linear functions. Hence, we have
(α,β)
Ln,q (t − x; x) = 0. (11)
Let g ∈ W 2 . From the Taylor’s expansion of g, we get

t
g(t) = g(x) + g
(x)(t − x) + (t − u)g
(u)du, t ∈ [0, ∞).

x
In view of (10), we get

t
(α,β) (α,β)
Ln,q (g; x) = g(x) + Ln,q (t − u)g
(u)du; x
x
t

(α,β)
| Ln,q (g; x) − g(x) | = Ln,q
(α,β)
(t − u)g
(u)du; x
x
t
(α,β)
≤ Mn,q (t − u)g
(u)du; x
x
α−βx
x+ ([n]q + β) α − βx
+ x + − u g
(u)du
x [n]q + β
t
(α,β)

≤ Mn,q (t − u)g
(u)du ; x

x
x + α−βx
([n]q + β)
+ (x + α − βx − u |g
(u)|du
([n]q + β)
x
2
(α,β) (α − βx)
≤ Mn,q ((t − x)2 ; x) + g
.
([n]q + β)
(12)
2
α − βx (α 2 − 2αβx + β 2 x 2 ) α 2 + 2αβx + β 2 x 2 β 2 (1 + 2x + x 2 )
= ≤ ≤
([n]q + β) ([n]q + β)2 ([n]q + β)2 ([n]q + β)2
2
2(1 + β ) 1
≤ x(1 + x) +
q([n]q + β) ([n]q + β)

2(1 + β 2 ) 2 1
= φ (x) + . (13)
q([n]q + β) ([n]q + β)
On the other hand, from (6), (10) and Lemma 2, we have

(α,β) (α,β) (α,β)
|Ln,q (f ; x)| ≤ |Mn,q (f , x)| + 2 f ≤ f Mn,q (1; x) + 2 f ≤ 3 f .
From (12) and (13), we have

(α,β) 4(1 + β 2 ) 1
| Ln,q (g; x) − g(x) | ≤ φ 2 (x) + g
.
q([n]q + β) ([n]q + β)
Hence
(α,β) (α,β) (α,β)
| Mn,q (f ; x) − f (x) | ≤| Ln,q (f − g; x) − (f − g)(x) | + | Ln,q (g; x) − g(x) |

α − βx
+ f x + − f (x)
([n]q + β)

4(1 + β 2 ) 1
≤ 4 f − g + φ 2 (x) + g

q([n]q + β) ([n]q + β)

| α − βx |
+ ω f, .
([n]q + β)
Now, taking infimum on the right-hand side over all g ∈ W 2 , we get

(α,β) 4(1 + β 2 ) 1
| Mn,q (f ; x) − f (x) | ≤ Cω2 f , φ (x) +
2
q([n]q + β) ([n]q + β)

| α − βx |
+ ω f, .
[n]q + β
4.2 Rate of Convergence
Let Bx2 [0, ∞) be the space of all functions defined on [0, ∞) and satisfying the
condition |f (x)| ≤ Mf (1 + x 2 ), where Mf is a constant depending on f . Let Cx2 [0, ∞)
be the subspace of all continuous functions belonging to Bx2 [0, ∞). Also, Cx∗2 [0, ∞)
f (x)
is the subspace of all functions f ∈ Cx2 [0, ∞), for which lim is finite. The
x→∞ 1 + x
2
|f (x)|
norm on Cx∗2 [0, ∞) is defined as f x2 := sup 1 + x2
. For any positive number a,
x∈[0,∞)
the usual modulus of continuity is defined as
ωa (f , δ) = sup |f (t) − f (x)|.

|t−x|≤δ, x,t∈[0,a]
We observe that for a function f ∈ Cx2 [0, ∞), the modulus of continuity ωa (f , δ)
tends to zero as δ → 0. Now we give a rate of convergence theorem for the operator
(α,β)
Mn,qn .
Theorem 3 Let f ∈ Cx2 [0, ∞), qn ∈ (0, 1) such that qn → 1 as n → ∞ and ωa + 1
be its modulus of continuity on the finite interval [0, a + 1] ⊂ [0, ∞), where a > 0,
then we have the following inequality:

(α,β) K 1
|Mn,q (f ; x) − f (x)| ≤ φ 2 (x) +
n
qn ([n]qn + β) ([n]qn + β)

2(1 + β 2 ) 1
+ 2ωa + 1 f , φ (x) +
2 ,
qn ([n]qn + β) [n]qn + β
where K = 8Mf (1 + a2 )(1 + β 2 ).
Proof For x ∈ [0, a] and t > a + 1, since t − x > 1, we have
|f (t) − f (x)| ≤ Mf (2 + x 2 + t 2 ) ≤ Mf (2 + 3x 2 + 2(t − x)2 )

≤ Mf (t − x)2 (2 + 3x 2 + 2) ≤ Mf (t − x)2 (4 + 3a2 )
|f (t) − f (x)| ≤ 4Mf (1 + a2 )(t − x)2 . (14)
For x ∈ [0, a] and t ≤ a + 1, we have

|t − x|
|f (t) − f (x)| ≤ ωa + 1 (f , |t − x|) ≤ 1 + ωa + 1 (f , δ), with δ > 0.
δ
(15)
From (14) and (15), for all t ∈ [0, ∞) and x ∈ [0, a] we can write

|t − x|
|f (t) − f (x)| ≤ 4Mf (1 + a2 )(t − x)2 + 1+ ωa + 1 (f , δ). (16)
δ
Hence, using Schwarz inequality,

(α,β) (α,β)
|Mn,q n
(f ; x) − f (x)| ≤ Mn,q n
(|f (t) − f (x)|; x)
(α,β)
≤ 4Mf (1 + a2 )Mn,q ((t − x)2 ; x)
n
1 (α,β) 1
+ ωa + 1 (f , δ) 1 + {Mn,qn ((t − x) ; x)} .
2 2
δ
In view of Lemma 3, for x ∈ [0, a]

(α,β) 8Mf (1 + a2 )(1 + β 2 ) 2 1
|Mn,q (f ; x) − f (x)| ≤ φ (x) +
n
qn ([n]qn + β) ([n]qn + β)
1
1 2(1 + β )
2 1 2
+ ωa + 1 (f , δ) 1 + φ 2 (x) + .
δ qn ([n]qn + β) ([n]qn + β)

2(1 + β 2 ) 1
Now, by choosing δ = φ 2 (x) + , we get the
qn ([n]qn + β) ([n]qn + β)
desired result.
4.3 Voronovskaja Type Theorem
In this section we establish a Voronovskaja type asymptotic formula for the operators
(α,β)
Mn,q .
Lemma 4 Assume that qn ∈ (0, 1), qn → 1 as n → ∞. Then, for every x ∈ [0, ∞)

there hold
(α,β)
lim [n]qn Mn,q n
(t − x; x) = α − βx
n→∞
and
(α,β)
lim [n]qn Mn,q n
((t − x)2 ; x) = x 2 + 2x.
n→∞
In view of Remark 2, the proof of this Lemma easily follows. Hence the details are
omitted.
Theorem 4 Let 0 < qn < 1 and qn → 1 as n → ∞. Then, for all f ∈ Cx2 [0, ∞)
we have
(α,β)
lim Mn,q n
(f ) − f x2 = 0.
n→∞
Proof Using [11], it is sufficient to verify the following conditions:

(α,β) m
lim Mn,q n
(t ; x) − x m x2 = 0, for m = 0, 1, 2. (17)
n→∞
(α,β)
Since Mn,qn (1; x) = 1, for m = 0, (17) holds. By Lemma 2, we have
(α,β)
(α,β) |Mn,qn (t; x) − x|
Mn,qn (t; x) − x x2 = sup
x∈[0,∞) (1 + x 2 )

[n]qn x + α
− x
([n]qn + β)
≤ sup
x∈[0,∞) 1 + x2
β x α 1
≤ sup + sup
([n]qn + β) x∈[0,∞) (1 + x 2 ) ([n]qn + β) x∈[0,∞) 1 + x 2
α +β
≤ = o(1) as n → ∞.
([n]qn + β)
Hence, the condition (17) holds for m = 1.

Again, by Lemma 2, we obtain

(α,β)
(α,β) |Mn,qn (t 2 ; x) − x 2 |
Mn,qn (t 2 ; x) − x 2 x2 = sup
x∈[0,∞) (1 + x 2 )

[n]qn (1 + qn [n]qn )x2 [n]qn (1 + qn + 2α)x α2

+ + − x 2
qn ([n]qn + β)2 ([n]qn + β)2 ([n]qn + β)2
= sup
x∈[0,∞) 1 + x2
([n]qn (1 + 2qn β) + β 2 ) x2
≤ sup
qn ([n]qn + β)2 x∈[0,∞) (1 + x )
2
[n]qn (1 + qn + 2α) x
+ sup
([n]qn + β)2 x∈[0,∞) (1 + x )
2
α2 1
+ sup = o(1) as n → ∞,
([n]qn + β)2 x∈[0,∞) 1 + x 2
which implies that the condition (17) holds for m = 2. This completes the proof.
Theorem 5 Assume that qn ∈ (0, 1), qn → 1 as n → ∞. Then, for any f ∈

Cx∗2 [0, ∞) such that f
, f
∈ Cx∗2 [0, ∞) we have
(α,β) 1
lim [n]qn (Mn,q n

(f ; x) − f (x)) = (α − βx)f
(x) + f (x)(x 2 + 2x),
n→∞ 2
uniformly in x ∈ [0, A], A > 0.
Proof Let f , f
, f
∈ Cx∗2 [0, ∞) and x ∈ [0, A] be fixed. By Taylor’s expansion, we

may write
f (t) = f (x) + f
(x)(t − x) + f (x)(t − x)2 + r(t, x)(t − x)2 , (18)
2
where r(t, x) is Peano form of the remainder, r(., x) ∈ Cx∗2 [0, ∞) and
lim r(t, x) = 0.
t→x
(α,β)
Applying Mn,qn to the above Eq. (18) we obtain
(α,β)
[n]qn (Mn,q n
(f ; x) − f (x)) = f
(x)[n]qn Mn,q
(α,β)
(t − x; x)
1
+ f
(x)[n]qn Mn,q (α,β)

((t − x)2 ; x)
2
(α,β)
+ [n]qn Mn,q r(t, x)(t − x)2 ; x .
By Cauchy Schwarz inequality, we have

(α,β) (α,β) (α,β)
Mn,qn r(t, x)(t − x) ; x ≤ Mn,qn r (t, x); x
2 2 Mn,qn (t − x) ; x .
4
(19)
We observe that r 2 (x, x) = 0 and r 2 (., x) ∈ Cx∗2 [0, ∞)). Then, it follows from
Theorem 3 that
(α,β) 2
lim [n]qn (Mn,q n
(r (t, x), x) = r 2 (x, x) = 0, (20)
n→∞
uniformly with respect to x ∈ [0, A]. Now, from (19)–(20) and in view of the fact
that
2
(α,β) 1
Mn,q ((t − x) ; x) = O
4
n
[n]qn
we obtain
(α,β)
lim [n]qn Mn,q n
(r(t, x)(t − x)2 , x) = 0,
n→∞
uniformly in x ∈ [0, A]. Thus, we obtain

(α,β)
(α,β)
lim [n]qn Mn,q n
(f , x) − f (x) = lim [n] qn f (x)Mn,qn ((t − x); x)
n→∞ n→∞
1
(α,β)
+ f (x)Mn,q n
((t − x)2 ; x)
2
(α,β)
+ Mn,q n
(r(t, x)(t − x)2 , x)
= (α − βx)f
(x) + f (x)(x 2 + 2x),
2
uniformly in x ∈ [0, A].
Corollary 1 Let q = qn satisfy 0 < qn < 1 and let qn → 1 as n → ∞. For each

f ∈ Cx2 [0, ∞) and p > 0, we have
(α,β)
|Mn,qn (f ; x) − f (x)|
sup = 0.
x∈[0,∞) (1 + x 2 )1 + p
Proof For any fixed x0 > 0

α,β (α,β) (α,β)
|Mn,qn (f ; x) − f (x)| |Mn,qn (f ; x) − f (x)| |Mn,qn (f ; x) − f (x)|
sup ≤ sup + sup
x∈[0,∞) (1 + x 2 )1 + p x≤x0 (1 + x 2 )1 + p x≥x0 (1 + x 2 )1 + p
(α,β)
(α,β) Mn,qn (1 + t 2 , x)
≤ Mn,q (f ) − f C[0,x0 ] + f x2 sup
n
x≥x0 (1 + x 2 )1 + p
|f (x)|
+ sup 2 1+p
. (21)
x≥x0 (1 + x )
Since |f (x)| ≤ Mf (1 + x 2 ), we have
|f (x)| Mf Mf
sup +
≤ sup ≤ .
x≥x0 (1 + x )
2 1 p
x≥x0 (1 + x )
2 p (1 + x02 )p
Let ε > 0 be arbitrary. Then, we can choose x0 to be so large that
Mf ε
< (22)
(1 + x02 )p 3
and in view of Theorem 4, we obtain

(α,β)
Mn,qn (1 + t 2 , x) (1 + x 2 ) f x2 f x2 f x2 ε
f x2 lim = = ≤ < . (23)
n→∞ (1 + x 2 )1 + p (1 + x 2 )1 + p (1 + x 2 )p (1 + x0 )
2 p 3
Using Theorem 3, we see that the first term of inequality (21) implies that
(α,β) ε
Mn,q n
(f ) − f C[0,x0 ] < as n → ∞. (24)
3
Combining (22)–(24), we get the desired result.
4.4 Point-Wise Estimates
Now, we establish some pointwise estimates of the rate of convergence of the oper-
ators (6). First, we give the relationship between the local smoothness of f and local
approximation.
We know that a function f ∈ CB [0, ∞) is in LipM γ on D, γ ∈ (0, 1], D ⊂ [0, ∞)
if it satisfies the condition
|f (t) − f (x)| ≤ M|t − x|γ , t ∈ [0, ∞) and x ∈ D,
where M is a constant depending only on γ and f .


Theorem 6 Let f ∈ CB [0, ∞) LipM γ , γ ∈ (0, 1], and D be any bounded subset
of the interval [0, ∞). Then, for each x ∈ [0, ∞) we have
γ
(α,β) [2]q (1 + β 2 ) 2 1 2
|Mn,q (f ; x) − f (x)| ≤ M φ (x) + + 2(d(x, D))γ ,
q([n]q + β) ([n]q + β)
where d(x, D) represents the distance between x and D.
Proof Let D be the closure of the set D in [0, ∞). Then, there exists at least one
point x0 ∈ D such that
d(x, D) = |x − x0 |.
By the definition of LipM γ , we get

(α,β) (α,β) (α,β)
|Mn,q (f ; x) − f (x)| ≤ Mn,q (|f (t) − f (x0 )|; x) + Mn,q (|f (x0 ) − f (x)|; x))

(α,β) γ γ
≤ M Mn,q (|t − x0 | ; x) + |x0 − x|

(α,β) γ γ
≤ M Mn,q (|t − x| , x) + 2|x − x0 | .
Now, by Holder’s inequality with p = 2

γ and 1
q = 1 − 1p , we have
1 1
(α,β) (α,β)
(|t − x|γ p ; x) (α,β) q
+ 2(d(x, D))γ
p q
|Mn,q (f ; x) − f (x)| ≤ M Mn,q Mn,q (1 , x)
γ
(α,β)
+ 2(d(x, D))γ
2
≤M Mn,q (|t − x|2 ; x)
γ
[2]q (1 + β 2 ) 2 1 2
≤M φ (x) + + 2(d(x, D))γ .
q([n]q + β) ([n]q + β)

(α,β)
Now, we give local direct estimate for the operators Mn,q using the Lipschitz
type maximal function of order γ studied by Lenze [18]
|f (t) − f (x)|
ωγ (f , x) =
sup , x ∈ [0, ∞) and γ ∈ (0, 1]. (25)
t=x,t∈[0,∞) |t − x|γ
Theorem 7 Let γ ∈ (0, 1] and f ∈ CB [0, ∞). Then, for all x ∈ [0, ∞), we have
γ
(α,β) [2]q (1 + β 2 ) 2 1 2
|Mn,q (f ; x) − f (x)| ≤
ωγ (f , x) φ (x) + .
q([n]q + β) ([n]q + β)
Proof From (25), we have
ωγ (f , x)|t − x|γ
|f (t) − f (x)| ≤
and hence
(α,β) (α,β) (α,β)
|Mn,q (f ; x) − f (x)| ≤ Mn,q (|f (t) − f (x)|; x) ≤
ωγ (f , x)Mn,q (|t − x|γ ; x).
Now, applying Holder’s inequality with p = 2

γ and 1
q = 1 − 1p , we have
γ
(α,β) (α,β)
|Mn,q (f ; x) − f (x)| ≤
ωγ (f , x)Mn,q ((t − x)2 ; x) 2 .
On using Lemma 3, we have our assertion.
4.5 Statistical Approximation
A sequence (xn )n is said to be statistically convergent to a number L denoted by

st − lim xn = L if for every ε > 0,
n
δ{n ∈ N : |xn − L| ≥ ε} = 0,
where
1
n
δ(K) = lim χK (j)
n n
j=1
is the natural density of K ⊆ N and χK is the characteristic function of K. We note

that every convergent sequence is statistically convergent, but the converse need not
be true.
For example, let

log10 n, n ∈ {10k , k ∈ N}
xn =
1, otherwise.
It follows that the sequence {xn } converges statistically to 1, but lim xn does not
n
exit.
Theorem 8 For any f ∈ Cx∗2 [0, ∞) and a sequence (qn )n in (0, 1) such that
1
st − lim qn = 1, st − lim(qn )n = a, (0 ≤ a < 1), st − lim = 0, (26)
n n n [n]qn
(α,β)
the operator Mn,q (f ; x) statistically converges to f (x), that is
(α,β)
st − lim Mn,q (f ) − f x2 = 0.
n
Proof Let us define ei (x) = x i , i = 0, 1, 2. It is sufficient to prove that

(α,β)
st − lim Mn,q n
(ei ) − ei x2 = 0, for i = 0, 1, 2. It is clear that
n
(α,β)
st − lim Mn,q n
(e0 ; .) − e0 x2 = 0.
n
From Lemma 2
(α,β)
(α,β) |Mn,qn (e1 ; x) − e1 (x)|
Mn,q (e1 ; .) − e1 x2 = sup
n
x∈[0,∞) (1 + x 2 )

[n]qn x + α

([n]qn + β) − x
≤ sup
x∈[0,∞) 1 + x2
α β
≤ e0 x2 + e1 x2
([n]qn + β) ([n]qn + β)
α β
≤ + . (27)
([n]qn + β) ([n]qn + β)
Since, by the conditions (26), we get

α
st − lim =0
n ([n]qn + β)
and
β
st − lim = 0.
n ([n]qn + β)
For ε > 0, let us define the following sets:

(α,β)
E := n ∈ N : Mn,q n
(e1 ; .) − e 1 x 2≥ ε ,

α ε
E1 := n ∈ N : ≥ ,
([n]qn + β) 2

β ε
E2 := n ∈ N : ≥ .
([n]qn + β) 2

By (27), it is clear that E ⊆ E1 E2 which implies that δ(E) ≤ δ(E1 ) + δ(E2 ) = 0,
and hence
(α,β)
st − lim Mn,q n
(e1 ; .) − e1 x2 = 0.
n
Similarly, we can estimate

(α,β)
(α,β) |Mn,qn (e2 ; x) − e2 (x)|
Mn,q (e2 ; .) − e2 x2 = sup
n
x∈[0,∞) (1 + x 2 )

[n]qn (1 + qn [n]qn )x2 [n] n (1 + qn + 2α)x
+ q([n] + α2
− x 2
q ([n] + β)2
n qn qn + β)2 ([n]qn + β)2
= sup
x∈[0,∞) 1 + x2
([n]qn (1 + 2qn β) + β 2 ) [n]qn (1 + qn + 2α)
≤ e2 x2 + e1 x2
qn ([n]qn + β)2 ([n]qn + β)2
α2
+ e0 x2
([n]qn + β)2
([n]qn (1 + 2qn β) + β 2 ) [n]qn (1 + qn + 2α) α2
≤ + + .
qn ([n]qn + β)2 ([n]qn + β)2 ([n]qn + β)2
(28)
Again, using (26), we get
([n]qn (1 + 2qn β) + β 2 )
st − lim = 0,
n qn ([n]qn + β)2
[n]qn (1 + qn + 2α)
st − lim = 0,
n ([n]qn + β)2
α2
st − lim = 0.
n ([n]qn + β)2
For a given ε > 0, we consider the following sets:

(α,β)
F := n ∈ N : Mn,q n
(e2 ; .) − e2 x 2 ≥ ε ,

[n]qn (1 + 2qn β) + β 2 ε
F1 := n ∈ N : ≥ ,
qn ([n]qn + β)2 3

[n]qn (1 + qn + 2α) ε
F2 := n ∈ N : ≥ ,
([n]qn + β)2 3

α2 ε
F3 := n ∈ N : ≥ .
([n]qn + β)2 3

Consequently, by (28) we obtain F ⊆ F1 F2 F3 , which implies that δ(F) ≤
δ(F1 ) + δ(F2 ) + δ(F3 ) = 0. Hence, we get
(α,β)
st − lim Mn,q n
(e2 ; .) − e2 x2 = 0.
n
This completes the proof of the theorem.

4.5.1 Rate of Statistical Convergence
For f ∈ Cx∗2 [0, ∞), following Freud [10], the weighted modulus of continuity of f
is defined as
|f (x + h) − f (x)|
Ω2 (f , δ) = sup .
x≥0,0<h≤δ 1 + (x + h)
2
Lemma 5 [19]. Let f ∈ Cx∗2 [0, ∞). Then,

(i) Ω2 (f , δ) is a monotone increasing function of δ,
(ii) lim Ω2 (f , δ) = 0,
δ→0
(iii) For any λ ∈ [0, ∞), Ω2 (f , λδ) ≤ (1 + λ)Ω2 (f , δ).
Theorem 9 Let f ∈ Cx∗2 [0, ∞) and (qn )n be a sequence satisfying (26). Then, for
sufficiently large n.
(α,β)
|Mn,q n
(f ; x) − f (x)| ≤ KΩ2 (f , δn )(1 + x 2 + λ ), x ∈ [0, ∞),

[2]qn (1 + β 2 )
where λ ≥ 1, δn = qn ([n]qn + β) and K is a positive constant independent f and n.
Proof
(α,β) (α,β)
|Mn,qn (f ; x) − f (x)| ≤ Mn,qn (|f (t) − f (x)|; x)

(α,β) |t − x|
≤ Mn,qn (1 + (x + |t − x|)2 ) 1 + ; x Ω2 (f , δ)
δ

(α,β) |t − x|
≤ Mn,qn (1 + (t + 2x)2 ) 1 + ; x Ω2 (f , δ)
δ

(α,β) 1 (α,β)
≤ Mn,qn (μx (t); x) + Mn,qn (μx (t)ψx (t); x) Ω2 (f , δ),
δ
where μx (t) = 1 + (t + 2x)2 and ψx (t) = |t − x|.
Now, using Cauchy–Schwarz inequality to the second term on the right-hand side,
we obtain

(α,β) (α,β) 1 (α,β) 2 ; x) M (α,β) (μ2 ; x) Ω (f , δ).
|Mn,q (f ; x) − f (x)| ≤ M n,q (μx ; x) + M n,q (ψx n,q x 2
n n
δ n n
(29)
From Lemma 2

(α,β) [n]qn (1 + qn [n]qn ) 2 [n]qn ([2]q + 2α) α2
Mn,q (1 + t 2 ; x) = 1 + x + x + ,
n
qn ([n]qn + β) 2 ([n]qn + β) 2 ([n]qn + β) 2
which implies that there exists a constant C1 > 0 such that

1 (α,β) 1 [n]qn (1 + qn [n]qn ) x 2

Mn,q (1 + t 2
; x) = +
1 + x2 n
1 + x2 qn ([n]qn + β)2 1 + x 2
[n]qn ([2]q + 2α) x
+
([n]qn + β)2 1 + x 2
α2 1
+ ,
([n]qn + β) 1 + x 2
2
≤ (1 + C1 ), for sufficiently large n . (30)
We have
μx (t) = 1 + (2x + t)2 ≤ 1 + 2(4x 2 + 2t 2 ). (31)
From (30) and (31), there is a positive constant K1 , such that

(α,β)
Mn,q n
(μx (t); x) ≤ K1 (1 + x 2 ), for sufficiently large n.
Similarly, from Lemma 2

(α,β) 2 (α,β)
Mn,q n
(μx (t); x) = Mn,q (1 + (2x + t) ) ; x ,
n
2 2

(α,β)
≤ Mn,qn (1 + 2(4x + 2t )) ; x ,
2 2 2

(α,β) (α,β)
≤ 64 Mn,q n
(1 + t 4 ; x) + (1 + x 2 )Mn,q n
(1 + t 2 ; x)

(α,β)
+ (1 + x )Mn,qn (1; x) .
2
Since
1
M (α,β) (1 + t 4 ; x) ≤ (1 + C2 ), for some constant C2 > 0 when n is sufficiently large ,
1 + x 4 n,qn
there exists a positive constant K2 such that

(α,β)
Mn,qn (μ2x (t); x) ≤ K2 (1 + x 2 ), for sufficiently large n.
Also, from Lemma 3 we have
(α,β) [2]qn (1 + β 2 ) 2 [2]qn (1 + β 2 )

Mn,q (ψx2 (t); x) ≤ φ (x) + .
n
qn ([n]qn + β) qn ([n]qn + β)2
Now from (29), we have

(α,β)
|Mn,qn (f ; x) − f (x)| ≤ Ω2 (f , δ)

2 2 1 [2]qn (1 + β 2 ) 2 [2]qn (1 + β 2 )
K1 (1 + x ) + K2 (1 + x ) φ (x) + .
δ qn ([n]qn + β) qn ([n]qn + β)2

[2]qn (1 + β 2 )
Choosing δ = qn ([n]qn + β) = δn , we obtain
(α,β)
|Mn,q n
(f ; x) − f (x)| ≤ Ω2 (f , δn )(1 + x 2 )(K1 + K2 1 + φ 2 (x)), for sufficiently large n.
Hence, for sufficiently large n

(α,β)
|Mn,q n
(f ; x) − f (x)| ≤ KΩ2 (f , δn )(1 + x 2 + λ ), x ∈ [0, ∞),
where λ ≥ 1 and K is a positive constant. This completes the proof of the theo-
rem.
Acknowledgments The authors are extremely grateful to the reviewers for careful reading of the
manuscript and for making valuable suggestions leading to better presentation of the paper. The
last author is thankful to the “University Grants Commission” India, for financial support to carry
out the above research work.
References
1. Agrawal, P.N., Gupta, V., Kumar, A.S.: On q-analogue of Bernstein-Schurer-Stancu operators.

Appl. Math. Comput. 219, 7754–7764 (2013)
2. Agrawal, P.N., Mohammad, A.J.: On simultaneous approximation by a linear combination of
a new sequence of linear positive operators. Turkish J. Math. 28(4), 353–365 (2004)
3. Agrawal, P.N., Kumar, A.S., Sinha, T.A.K.: Stancu type generalization of modified Schurer
operators based on q-intgers. Appl. Math. Comput. 226, 765–776 (2014)
4. Aral, A., Gupta, V.: On the q-analogue of Stancu-Beta operators. Appl. Math. Lett. 25(1),
67–71 (2012)
5. Aral, A., Gupta, V.: Generalized q-Baskakov operators. Mathematica Slovaca 61(4), 619–634
(2011)
6. Aral, A., Gupta, V., Agrawal, R.P.: Applications of q-Calculus in Operator Theory. Springer,
New York (2013)
7. Büyükyazici, I.: Approximation by Stancu-Chlodowsky polynomials. Comput. Math. Appl.
59(1), 274–282 (2010)
8. Deo, N.: Simultaneous approximation by Lupas modified operators with weighted function of
Sazsz operators, JIPAM. J. Inequal Pure Appl. Math. 5(4) Article 113 (2004) 5pp
9. DeVore, R.A., Lorentz, G.G.: Constructive Approximation. Springer, Berlin (1993)
10. Freud, G.: Investigations on weighted approximation by polynomials. Studia Sci. Math. Hungar
8, 285–305 (1973)
11. Gadzjiev, A.D.: Therems of the type of P. P. Korovkin type theorems, Math. Zametki. 20(5),
781–786 (1976)
12. Gupta, V.: A note on q-Baskakov-Szasz operators. Lobachevskii J. Math. 31(4), 359–366 (2010)
13. Gupta, V., Agarwal, R.P.: Convergence Estimates in Approximation Theory. Springer, Berlin
(2014)
14. Gupta, V., Kim, T.: On a q-analogue of Baskakov basis functions. Russ. J. Math. Phys. 20(3),
276–282 (2013)
15. Gupta, V., Karsli, H.: Some approximation properties by q-Szasz-Mirakyan-Baskakov-Stancu
operators. Lobachevskii J. Math. 33(2), 175–182 (2012)
16. Gupta, V., Srivastava, G.S.: Simultaneous approximation by Baskakov-Szász type operators,
Bull. Math. Soc. Sci. Math. Roumanie (N.S.), 37(85), 73–85 (1993)
17. Kac, V., Cheung, P.: Quantum Calculus (Universitext). Springer, New York (2002)
18. Lenze, B.: On Lipschitz-type maximal functions and their smoothness spaces. Nederl. Akad.
Wetensch. Indag. Math. 50(1), 53–63 (1988)
19. Lopez-Moreno, A.J.: Weighted simultaneous approximation with Baskakov type operators.
Acta Math. Hungar. 104(1–2), 143–151 (2004)
20. Phillips, G.M.: Bernstein polynomials based on the q-integers. Ann. Numer. Math. 4, 511–518
(1997)
21. Stancu, D.D.: Approximation of functions by a new class of linear polynomial operators. Rev.
Roumaine Math. Pures Appl. 13, 1173–1194 (1968)
22. Verma, D.K., Agrawal, P.N.: Approximation by Baskakov-Durrmeyer-Stancu operators based
on q-integers. Lobachevskii J. Math. 34(2), 187–196 (2013)
Approximation of Functions of Class
Lip(α, p) in L p -Norm
M.L. Mittal and Mradul Veer Singh
Abstract Mittal and Rhoades (Int. J. Math. Game Theory Algebra 9(4), 259–267,
1999 [9]; J. Comput. Anal. Appl. 2(1) 1–10, 2000 [10]) and Mittal et al. (J. Math.
Anal. Appl. 326(1) 667–676, 2007 [7]; Appl. Math. Comput. 217(9), 4483–4489,
2011 [8]) initiated the studies of error estimates E n ( f ) through trigonometric-Fourier
approximation (tfa) for situations in which the summability matrix T does not have
monotone rows. In this paper, we extend the results of Mittal et al. (Appl. Math.
Comput. 217(9), 4483–4489, 2011 [8]) to a more general Cλ -method in view of
Armitage and Maddox (Analysis 9, 195–204, 1989 [1]), which in turn generalizes
the several previous known results due to Mittal and Singh (Int. J. Math. Math. Sci.,
Art. ID 267383, 1–6, 2014 [11]), Deg̃er et al. (Proc. Jangjeon Math. Soc. 15(2), 203–
213, 2012 [4]), Leindler (J. Math. Anal. Appl. 302, 129–136, 2005 [6]), Chandra (J.
Math. Anal. Appl. 275, 13–26, 2002 [3]) and Quade (Duke Math. J. 3(3), 529–543,
1937 [15]).
Keywords Trigonometric Fourier approximation · Cλ -method · L p -norm · Class

Li p(α, p)
1 Introduction
For a given function f ∈ L p := L p [0, 2π ], p ≥ 1, let
a0
n n
sn ( f ) := sn ( f ; x) = + (ak cos kx + bk sin kx) = u k ( f ; x) (1)
2
k=1 k=0
M.L. Mittal · M.V. Singh (B)

Department of Mathematics, Indian Institute of Technology Roorkee,
Roorkee 247667, India
e-mail: mradul.singh@gmail.com
M.L. Mittal
e-mail: mlmittal_iit@yahoo.co.in

DOI 10.1007/978-81-322-2485-3_8
110 M.L. Mittal and M.V. Singh
denote the partial sums, called trigonometric polynomials of degree (or order) n, of
the first (n + 1) terms of the Fourier series of f at a point x.
A positive sequence c := {cn } is called almost monotone decreasing (increasing)
if there exists a constant K := K (c), depending on the sequence c only, such that for
all n ≥ m, cn ≤ K cm (K cn ≥ cm ). Such sequences will be denoted by c ∈ AMDS
and c ∈ AMIS respectively. A sequence which is either AMDS or AMIS is called
almost monotone sequence and will be denoted by c ∈ AMS.
Let F be an infinite subset of N and F the range of strictly increasing sequence of
positive integers, say F = {λ(n)}∞ n=1 . The Cesàro submethod C λ is defined as
λ(n)
1
(Cλ x)n = xk , (n = 1, 2, 3, ...),
λ(n)
k=1
where {xk } is a sequence of real or complex numbers. Therefore, the Cλ -method

yields a subsequence of the Cesàro method C1 , and hence it is regular for any λ.
Matrix-Cλ is obtained by deleting a set of rows from Cesàro matrix. The basic
properties of Cλ -method can be found in [1, 14].
Define
λ(n)

τnλ ( f ) = τnλ ( f ; x) = aλ(n),k sk ( f ; x), ∀n ≥ 0.
k=0
The trigonometric Fourier series of the signal f is said to be T λ -summable to s if

τnλ ( f ) → s as n → ∞.
Throughout T ≡ (an,k ), a linear operator, will denote an infinite lower triangular
matrix with nonnegative entries and row sums 1. Such a matrix T is said to have
monotone rows if , ∀n, {an,k } is either nonincreasing or nondecreasing in k, 0 ≤ k ≤
n. A linear operator T is said to be regular if it is limit-preserving over the space of
convergent sequences.
We write
2π
1
sn ( f ; x) = f (x + t)Dn (t) dt, Dn (t) = (sin(n + 1/2)t)/2 sin(t/2),
π 0
λ(n)

Aλ(n),k = aλ(n),r , Aλ(n),0 ≡ 1, ∀n ≥ 0.
r =k
The notation [x] means the greatest integer contained in x.
2 Known Results
Chandra [3] proved three theorems on the trigonometric approximation using Nör-
lund and Riesz matrices. Some of them give sharper estimates than the results proved
by Quade [15], Mohapatra and Russell [12] and himself earlier [2]. Similar results
Approximation of Functions of Class Lip(α, p) in L p -Norm 111
were proved by Khan [5] for generalized N p -mean and Mohapatra et al. [13] for Tay-
lor mean. Leindler [6] extended the results of Chandra [3] without the assumption of
monotonicity on the generating sequence { pn }. Leindler [6] proved the following:
Theorem 1 ([6]) If f ∈ Li p(α, p) and { pn } be positive. If one of the conditions
(i) p > 1, 0 < α < 1 and { pn } ∈ AMDS,
(ii) p > 1, 0 < α < 1 and { pn } ∈ AMIS and
(n + 1) pn = O(Pn ) holds, (2)

(iii) p > 1, α = 1 and n−1 k| pk | = O(Pn ),
n−1 k=1
(iv) p > 1, α = 1, k=0 | pk | = O(Pn /n) and (2) holds,

(v) p = 1, 0 < α < 1 and n−1k=−1 | pk | = O(Pn /n),
maintains, then
|| f − Nn ( f )|| p = O(n −α ). (3)
f ∈ Li p(α, 1), 0 < α < 1. If the positive { pn } satisfies

Theorem 2 ([6]) Let
conditions (2) and n−1
k=0 | pk | = O(Pn /n) hold, then
|| f − Rn ( f )||1 = O(n −α ).
Mittal et al. [7, 8] extended the work of Chandra to general matrices. Mittal et al. [8]
proved the following:
Theorem 3 ([8]) Let f ∈ Li p(α, p) and let T = (an,k ) be an infinite regular
triangular matrix.
(i) If p > 1, 0 < α < 1, {an,k } ∈ AMS in k and satisfies
(n + 1)max{an,0 , an,r } = O(1). (4)
where r := [n/2] then

|| f − τn ( f )|| p = O(n −α ). (5)

(ii) If p > 1, α = 1 and n−1
k=0 (n − k)| k an,k | = O(1), or
(iii)If p > 1, α = 1 and nk=0| k an,k | = O(an,0 ), or
(iv) If p = 1, 0 < α < 1 and nk=0 | k an,k | = O(an,0 ),
and also (n + 1)an,0 = O(1), holds then (5) is satisfied.
Recently, Deg̃er et al. [4] extended the results of Chandra [3] to more general Cλ -
method in view of Armitage and Maddox [1]. Deg̃er et al. [4] proved:
Theorem 4 ([4]) Let f ∈ Li p(α, p) and { pn } be positive such that
(λ(n) + 1) pλ(n) = O(Pλ(n) ), (6)
If either (i) p > 1, 0 < α ≤ 1 and { pn } is monotonic or (ii) p = 1, 0 < α < 1 and
{ pn } is nondecreasing then
|| f − Nnλ ( f )|| p = O(n −α ).
Theorem 5 ([4]) Let f ∈ Li p(α, 1), 0 < α < 1. If the positive { pn } satisfies
condition (6) and nondecreasing, then || f − Rnλ ( f )||1 = O(n −α ).
Very recently, in [11], the authors of this paper generalized two theorems of Deg̃er
et al. [4], by dropping the monotonicity on the elements of the matrix rows. These
results also generalize the results of Leindler [6] to more general Cλ -method.
Theorem 6 ([11]) If f ∈ Li p(α, p) and { pn } be positive. If one of the following
conditions
(i) p > 1, 0 < α < 1 and { pn } ∈ AMDS,
(ii) p > 1, 0 < α < 1 and { pn } ∈ AMIS and (6) holds,
λ(n)−1
(iii) p > 1, α = 1 and k=1 k| pk | = O(P λ(n) ),
λ(n)−1 Pλ(n)
(iv) p > 1, α = 1, k=0 | pk | = O λ(n) and (6) holds,
λ(n)−1
Pλ(n)
(v) p = 1, 0 < α < 1 and k=−1 | pk | = O λ(n) ,
maintains, then
|| f − Nnλ ( f )|| p = O (λ(n))−α . (7)
Theorem 7 ([11]) Let f ∈ Li p(α, 1),0 < α < 1. If the positive { pn } satisfies (6)
λ(n)−1 Pλ(n)
and the condition k=0 | pk | = O λ(n) holds, then

|| f − Rnλ ( f )||1 = O (λ(n))−α . (8)
3 Main Results
Mittal and Rhoades [9, 10] initiated the studies of error estimates through
trigonometric-Fourier approximation (tfa) for situations in which the summability
matrix T does not have monotone rows. In continuation of Mittal and Singh [11],
in this paper, we generalize Theorem 3 of Mittal et al. [8] using more general Cλ -
method. We prove the following:
Theorem 8 Let f ∈ Li p(α, p) and let T = (an,k ) be an infinite regular triangular
matrix.
(i) If p > 1, 0 < α < 1, {an,k } ∈ AMS in k and satisfies
(λ(n) + 1)max{aλ(n),0 , aλ(n),r } = O(1), (9)
where r := [λ(n)/2] then
|| f − τnλ ( f )|| p = O((λ(n))−α ). (10)
(ii) If p > 1, α = 1 and

λ(n)−1

(λ(n) − k)| k aλ(n),k | = O(1), or (11)
k=0
(iii) If p > 1, α = 1 and
λ(n)

| k aλ(n),k | = O(aλ(n),0 ), or (12)
k=0
(iv) If p = 1, 0 < α < 1 and
λ(n)

| k aλ(n),k | = O(aλ(n),0 ), (13)
k=0
and also
(λ(n) + 1)aλ(n),0 = O(1), (14)
holds then (10) is satisfied.

Remarks (1) If λ(n) = n, then our Theorem 8 generalizes Theorem 3.
(2) If T ≡ (an,k ) is a Nörlund N p (or weighted R p ) matrix then-
(a) If λ(n) = n, then condition (9) (or (14)) reduces to (2) while the conditions (11),
(12), (13) reduce to conditions in (iii), (iv) and (v) of Theorem 1 respectively. Thus
our Theorem 8 generalizes Theorems 1 and 2.
(b) Deg̃er et al. [4] used the monotone sequences { pn } in Theorems 4 and 5 while our
Theorem 8 claims less than the requirement of their theorems. For example, condition
(11) of Theorem 8 is automatically satisfied if { pn } is nonincreasing sequence, i.e.,
L.H.S. of (11) gives
λ(n)−1
λ(n)−1

k pλ(n)−k 1

(λ(n) − k) = (λ(n) − k)| pλ(n)−k − pλ(n)−k−1 |
Pλ(n) Pλ(n)
k=0 k=0
Pλ(n)−1 − λ(n) pλ(n)
= = O(1) = R.H.S.,
Pλ(n)
while the condition (12) is always satisfied if { pn } is nondecreasing, i.e.,
λ(n)

λ(n)

k pλ(n)−k
= 1 | pλ(n)−k − pλ(n)−k−1 |
P P
k=0 λ(n) λ(n) k=0
1
= [ pλ(n) − pλ(n)−1 + pλ(n)−1 − pλ(n)−2 + ... + p0 − p−1 ]
Pλ(n)

pλ(n)
=O .
Pλ(n)
Further, condition (9) (or (14)) of Theorem 8 reduces to (6) of Theorem 4. Thus
our Theorem 8 generalizes the Theorems 4 and 5 of Deg̃er et al. [4] under weaker
assumptions and gives sharper estimate because all the estimates of Deg̃er et al. [4]
are in terms of n while our estimates are in terms of λ(n) and (λ(n))−α ≤ n −α for
0 < α ≤ 1.
(c) Also, Theorem 8 extends Theorems 6 and 7 of Mittal, Singh [11] where two
theorems of Deg̃er et al. [4] were generalized by dropping the monotonicity on the
elements of matrix rows.
4 Lemmas
We shall use the following lemmas in the proof of our Theorem:

Lemma 1 ([15]) If f ∈ Li p(1, p), for p > 1 then
||σn ( f ) − sn ( f )|| p = O(n −1 ), ∀n > 0.
Lemma 2 ([15]) If f ∈ Li p(α, p), for 0 < α ≤ 1 and p > 1. Then
|| f − sn ( f )|| p = O(n −α ), ∀n > 0.
Note: We are using sums upto λ(n) in the nth partial sums sn and σn and writing
these sums snλ and σnλ , respectively, in the above lemmas for our purpose in this paper.
Lemma 3 Let T have AMS rows and satisfy (4). Then, for 0 < α < 1,
λ(n)

aλ(n),k (k + 1)−α = O (λ(n) + 1)−α .
k=0
Proof Suppose that the rows of T are AMDS. Then there exists a K > 0 such that
λ(n)
λ(n)
λ(n)

aλ(n),k (k + 1)−α = K aλ(n),0 (k + 1)−α = K aλ(n),0 (k + 1)−α
k=0 k=0 k=0
= O(aλ(n),0 (λ(n) + 1) 1−α
) = O((λ(n) + 1)−α ).
A similar result can be proved if the rows of T are AMIS.

5 Proof of the Theorem 8
Case I. p > 1, 0 < α < 1. We have

λ(n)
λ(n)

τnλ ( f ) − f = aλ(n),k sk ( f ) − f = aλ(n),k (sk ( f ) − f ) (15)
k=0 k=0
Thus in view of Lemmas 2 and 3 we have

λ(n)
λ(n)

||τnλ ( f ) − f || p ≤ aλ(n),k ||sk ( f ) − f || p = aλ(n),k O((k + 1)−α )
k=0 k=0

= O (λ(n) + 1)−α .
Case III. p > 1, α = 1. We have
||τnλ ( f ) − f || p ≤ ||τnλ ( f ) − snλ ( f )|| p + ||snλ ( f ) − f || p .
Again using the Lemma 2, we get

||τnλ ( f ) − f || p ≤ ||τnλ ( f ) − snλ ( f )|| p + O (λ(n))−1 . (16)
So, it remains to show that

||τnλ ( f ) − snλ ( f )|| p = O (λ(n))−1 . (17)
Since Aλ(n),0 = 1, we have
λ(n)
λ(n)

Aλ(n),k − Aλ(n),0
τnλ ( f ) − snλ ( f ) = (Aλ(n),k − Aλ(n),0 )u k ( f ) = (ku k ( f )).
k
k=1 k=1
Thus using Abel’s transformation, we get

λ(n)−1

k
||τnλ ( f ; x) − snλ ( f ; x)|| p ≤ k .|| ju j ( f )|| p
k
k=1 j=1
λ(n)
Aλ(n),λ(n) − Aλ(n),0

+ .|| ju j ( f )|| p . (18)
λ(n)
j=1
Let σn (s) denote the nth term of the (C, 1) transform of the sequence s, then
λ(n)

1
snλ ( f ) − σnλ ( f ) = ju j ( f ).
(λ(n) + 1)
j=1
λ(n)

|| ju j || p = (λ(n) + 1)||snλ ( f ) − σnλ ( f )|| p = (λ(n) + 1)O (λ(n))−1 = O(1).
j=1
(19)
Note that

Aλ(n),0 − Aλ(n),λ(n)
≤ (λ(n))−1 Aλ(n),0 = O (λ(n))−1 .
λ(n)
Thus
λ(n)
Aλ(n),0 − Aλ(n),λ(n)
.|| ju ( f )|| = O (λ(n)) −1
. (20)
λ(n) j p
j=1
Now

Aλ(n),k − Aλ(n),0 1 Aλ(n),k+1 − Aλ(n),0
k = k (Aλ(n),k − Aλ(n),0 ) +
k k k(k + 1)
⎡ ⎤
λ(n)
λ(n)

1 ⎣(k + 1) k Aλ(n),k +
= aλ(n),r − aλ(n),r ⎦
k(k + 1)
r =k+1 r =0
⎡ ⎤
1 k
= ⎣(k + 1)aλ(n),k − aλ(n),r ⎦ . (21)
k(k + 1)
r =0
Next we claim that ∀k ∈ N ,

k
k−1
| aλ(n),r − (k + 1)aλ(n),k | ≤ (r + 1)|aλ(n),r − aλ(n),r +1 |, (22)
r =0 r =0
If k = 1, then the inequality (22) reduces to

1
| aλ(n),r − 2aλ(n),1 | = |aλ(n),0 − aλ(n),1 |.
r =0
Thus (22) holds for k = 1. Now let us assume that (22) is true for k = m, i.e.,

m
m−1
| aλ(n),r − (k + 1)aλ(n),m | ≤ (r + 1)|aλ(n),r − aλ(n),r +1 |. (23)
r =0 r =0
Let k = m + 1, using (23), we get

m+1
| aλ(n),r − (m + 2)aλ(n),m+1 |
r =0

m
=| aλ(n),r − (m + 1)aλ(n),m + (m + 1)aλ(n),m − (m + 1)aλ(n),m+1 |
r =0

m−1
≤ (r + 1)|aλ(n),r − aλ(n),r +1 | + (m + 1)|aλ(n),m − aλ(n),m+1 |
r =0
(m+1)−1

= (r + 1)|aλ(n),r − aλ(n),r +1 |.
r =0
Thus (22) is true ∀k. Using (12), (14), (21), (22), we get

λ(n)
λ(n)

k
1
| k |= (k + 1)aλ(n),k − aλ(n),r
k k(k + 1)
k=1 k=1 r =0
λ(n)

k−1
1
≤ (m + 1)|aλ(n),m − aλ(n),m+1 |
k(k + 1)
k=1 m=0
λ(n)

k
1
= m|aλ(n),m−1 − aλ(n),m |
k(k + 1)
k=1 m=1
λ(n)
∞
1
≤ m| m aλ(n),m−1 |
k(k + 1)
m=1 k=m
λ(n)−1

= | k aλ(n),k | = O(aλ(n),0 ) = O (λ(n))−1 .
k=0
(24)
Combining (18), (19), (20) and (24) yields (17). From (17) and (16), we get

||τnλ ( f ) − f || p = O (λ(n))−1 .
λ(n)−1
Case II. p > 1, α = 1. For this we first prove that the condition k=0 (λ(n) −
k)| k aλ(n),k | = O(1) implies that
λ(n)

k = O (λ(n))−1 . (25)
k
k=1
As in case (iii), using (22) and taking r := [λ(n)/2] throughout the case, we have

λ(n)

λ(n)
Aλ(n),k − Aλ(n),0 1

k

k = (k + 1)aλ(n),k − aλ(n),m
k(k + 1)
k=1
k
k=1 m=0
λ(n)

k−1
1
= (m + 1)|aλ(n),m − aλ(n),m+1 |
k(k + 1)
k=1 m=0
⎛ ⎞
r λ(n)

k
=⎝ + ⎠ k −1 (k + 1)−1 m| m aλ(n),m−1 |
k=1 k=r +1 m=1
:= B1 + B2 , say.
Now interchanging the order of summation and using (11), we get

r
k
r ∞

B1 = k −1 (k + 1)−1 m| m aλ(n),m−1 | ≤ m| m aλ(n),m−1 | k −1 (k + 1)−1
k=1 m=1 m=1 k=m

r λ(n)

= | m aλ(n),m−1 | = | λ(n)−m aλ(n),λ(n)−m |
m=1 m=λ(n)−r +1
λ(n)

m
= | λ(n)−m aλ(n),λ(n)−m |.
r −1
m=r −1
λ(n) λ(n)−1
1 1
≤ m| λ(n)−m aλ(n),λ(n)−m | = (λ(n) − k)| k aλ(n),k |
r −1 r −1
m=1 k=0
1
= O(1) = O (λ(n))−1 . (26)
r −1
λ(n)

k
Now B2 = k −1 (k + 1)−1 m| m aλ(n),m−1 |
k=r m=1
⎡⎛ ⎞ ⎤
λ(n)

r
k
≤ k −1 (k + 1)−1 ⎣⎝ + ⎠ m| m aλ(n),m−1 |⎦ := B21 + B22 , say.
k=r m=1 m=r
Furthermore, using again our assumption, we get
λ(n)

r
B21 = k −1 (k + 1)−1 m| m aλ(n),m−1 |
k=r m=1
λ(n)
λ(n)

≤ r −1 (k + 1)−1 m| λ(n)−m aλ(n),λ(n)−m |
k=r m=1
λ(n)
λ(n)−1

= r −1 (k + 1)−1 (λ(n) − k)| k aλ(n),k |
k=r k=0
λ(n)

= O(r −1 ) (k + 1)−1 = O (λ(n))−1 . (27)
k=r
Again interchanging the order of summation and using (11), we get
λ(n)

k λ(n)

k
B22 = k −1 (k + 1)−1 m| m aλ(n),m−1 | ≤ (k + 1)−1 | m aλ(n),m−1 |
k=r m=r k=r m=r
λ(n)
λ(n)
λ(n)
λ(n)

≤ | m aλ(n),m−1 | (k + 1)−1 ≤ (r + 1)−1 | m aλ(n),m−1 | 1
m=r k=m m=r k=m
λ(n)

= (r + 1)−1 (λ(n) − m + 1)| m aλ(n),m−1 |
m=r
λ(n)−1

−1
= (r + 1) (λ(n) − k)| k aλ(n),k |
k=r −1

= (r + 1)−1 O(1) = O (λ(n))−1 . (28)
Summing up our partial results (26), (27), (28) we verified (25). Thus (16), (18),
(19), (25) and Lemma 2, again yield

|| f − τnλ ( f )|| p = O (λ(n))−1 .
Case IV. p = 1, 0 < α < 1.

Using Abel’s transformation, conditions (13), (14), convention an,n+1 = 0 and the
result of Quade [15], we obtain
λ(n)
λ(n)

||τnλ ( f ) − f ||1 = || aλ(n),k sk ( f ) − f ||1 = || aλ(n),k (sk ( f ) − f )||1
k=0 k=0
λ(n)−1
λ(n)

k
= || k aλ(n),k (sr ( f ) − f ) + (aλ(n),λ(n) − aλ(n),λ(n)+1 ) (sr ( f ) − f )||1
k=0 r =0 r =0
λ(n)
λ(n)

k
= || k aλ(n),k (sr ( f ) − f )||1 = || k aλ(n),k (k + 1)(σk ( f ) − f )||1
k=0 r =0 k=0
⎛ ⎞
λ(n)
λ(n)

≤ (k + 1)| k aλ(n),k |.||σk ( f ) − f )||1 = O ⎝ (k + 1)1−α | k aλ(n),k |⎠
k=0 k=0
λ(n)

= O λ(n)1−α | k aλ(n),k | = O λ(n)1−α O aλ(n),0 = O (λ(n))−α .
k=0
This completes the proof of case (iv) and hence the proof of Theorem 8 is complete.
References
1. Armitage, D.H., Maddox, I.J.: A new type of Cesáro mean. Analysis 9, 195–204 (1989)
2. Chandra, P.: A note on degree of approximation by Nörlund and Riesz operators. Mat. Vestnik
42, 9–10 (1990)
3. Chandra, P.: Trigonometric approximation of functions in L p -norm. J. Math. Anal. Appl. 275,
13–26 (2002)
4. Deg̃er, U., Dag̃adur, İ., Küçükaslan, M.: Approximation by trigonometric polynomials to func-
tions in L p -norm. Proc. Jangjeon Math. Soc. 15(2), 203–213 (2012)
5. Khan, H.H.: On the degree of approximation of functions belonging to class Lip(α, p). Indian
J. Pure Appl. Math. 5(2), 132–136 (1974)
6. Leindler, L.: Trigonometric approximation in L p -norm. J. Math. Anal. Appl. 302, 129–136
(2005)
7. Mittal, M.L., Rhoades, B.E., Mishra, V.N., Singh, U.: Using infinite matrices to approximate
functions of class Lip(α, p) using trigonometric polynomials. J. Math. Anal. Appl. 326(1),
667–676 (2007)
8. Mittal, M.L., Rhoades, B.E., Sonker, S., Singh, U.: Approximation of signals of class Lip(α, p)
by linear operators. Appl. Math. Comput. 217(9), 4483–4489 (2011)
9. Mittal, M.L., Rhoades, B.E.: On the degree of approximation of continuous functions by using
linear operators on their Fourier series. Int. J. Math. Game Theory Algebra 9(4), 259–267
(1999)
10. Mittal, M.L., Rhoades, B.E.: Degree of approximation to functions in a normed space. J.
Comput. Anal. Appl. 2(1), 1–10 (2000)
11. Mittal, M.L., Singh, M.V.: Approximation of signals (functions) by trigonometric polynomials
in L p -norm. Int. J. Math. Math. Sci. Art. ID 267383, 1–6 (2014)
12. Mohapatra, R.N., Russell, D.C.: Some direct and inverse theorems in approximation of func-
tions. J. Aust. Math. Soc. Ser. A 34(2), 143–154 (1983)
13. Mohapatra, R.N., Holland, A.S.B., Sahney, B.N.: Functions of class Lip(α, p) and their Taylor
mean. J. Approx. Theory 45, 363–374 (1985)
14. Osikiewicz, J.A.: Equivalence results for Cesáro submethods. Analysis 20(1), 35–43 (2000)
15. Quade, E.S.: Trigonometric approximation in the mean. Duke Math. J. 3(3), 529–543 (1937)
A New Genuine Durrmeyer Operator
Vijay Gupta
Abstract The generalization of the Bernstein polynomials based on certain parame-

ter was considered by Stancu (Rew. Roum. Math. Pure. Appl. 13, 1173–1194, 1968
[14]). Recently, Gupta and Rassias (J. Math. Anal. 8(2), 146–155, 2014 [11]) pro-
posed a Durrmeyer-type modification of the Lupaş operators and established some
results. Actually, the genuine operators are important as far as the approximation is
concerned. Here we propose genuine Durrmeyer-type operators, which preserve lin-
ear functions. We establish moments using generalized hypergeometric function and
obtain an asymptotic formula and a direct result in terms of second-order modulus
of continuity. In the end we propose an open problem for the readers.
Keywords Bernstein polynomials · Moments · Asymptotic formula · Genuine oper-

ators · Linear functions
1 Introduction
In the year 1967 Durrmeyer [4] introduced the integral modification of the Bernstein
polynomials as

n 1
Mn ( f, x) = (n + 1) pn,k (x) pn,k (t) f (t)dt, x ∈ 0, 1, (1)
k=0 0
where
n k
pn,k (t) = t (1 − t)n−k .
k
Derriennic [3] first studied these operators in detail and she estimated some results
in ordinary and simultaneous approximation. Later some direct estimates in simulta-
neous approximation for linear combinations were discussed by Agrawal and Gupta
V. Gupta (B)
Department of Mathematics, Netaji Subhas Institute of Technology, Sector 3 Dwarka,
New Delhi 110078, India
e-mail: vijaygupta2001@hotmail.com
DOI 10.1007/978-81-322-2485-3_9
122 V. Gupta
[9]. Also Gupta in [7] and Finta and Gupta in [5] considered the q analogue of
the Bernstein–Durrmeyer operators and established some direct results. The rate of
convergence in simultaneous approximation in some other form has been discussed
in [10].
(α)
For Pn : C[0, 1] → C[0, 1], with a non-negative parameter α, Stancu in [14]
considered a sequence of positive linear operators, which is defined as

n
k (α)
Pn(α) ( f, x) = f pn,k (x), (2)
n
k=0
(α)
where pn,k (x) is the Pólya distribution with density function given by
k−1 n−k−1
(α) n ν=0 (x + vα) μ=0 (1 − x + μα)
pn,k (x) = n−1 , x ∈ [0, 1].
k (1 + λα) λ=0
(0)
As a special case pn,k (x) is the density function of the binomial distribution and
(0)
Pn ( f, x) reduces to the classical Bernstein polynomials. For convergence point of
view Lupaş and Lupaş [12] considered the operators (2) by taking α = 1/n, later in
[13] some approximation properties have been discussed for this case. In an alternate
form such operators can be represented as

2(n!)
n
(1/n) n k
Pn ( f, x) = f (nx)k (n − nx)n−k , (3)
(2n)! k n
k=0
where the Pochhammer symbol is given as (x)i = x(x + 1)(x + 2) . . . (x + i − 1).

Very recently, Gupta and Rassias [11] introduced the Durrmeyer-type integral
modification of the operators (3) and established some direct results, but their opera-
tors preserve only the constant functions. We propose here a genuine Durrmeyer-type
modification of the operators (3), which also preserve linear functions and can be
defined as

n−1 1
(1/n) (1/n)
Dn ( f, x) = (n − 1) pn,k (x) pn−2,k−1 (t) f (t) dt (4)
k=1 0
(1/n) (1/n)
+ pn,0 (x) f (0) + pn,n (x) f (1),
where
(1/n) 2(n!) n
pn,k (x) = (nx)k (n − nx)n−k
(2n)! k
and the Bernstein basis function is defined as in (1).

A New Genuine Durrmeyer Operator 123
In the present paper, we establish the moments of the operators (4) using gener-
alized hypergeometric series. We also establish an asymptotic formula and a direct
result in terms of second-order modulus of continuity.
2 Auxiliary Results
We estimate the moments using the recurrence relation. Such relation for some other
form was given by Greubel [6] in an open problem raised by the author [8]. To make
the paper self-content, we provide the detailed proof below:
(1/n)
Lemma 1 For r ≥ 1, if we denote Tn,r (x) = Dn (er , x), then we have
(r + 1)(2r − nx + 3n − 3) + n(nx + x − 3) + 3
Tn,r +1 (x) = Tn,r (x)
(r + n)2
r (r − 1)(nx + 1 − r − 2n)
+ Tn,r −1 (x)
(n + r − 1)(n + r )2
Proof By definition, we have

n−1 1
(1/n) (1/n)
Tn,r (x) = (n − 1) pn,k (x) pn−2,k−1 (t) t r dt + pn,n (x)
k=1 0
Also, by simple computation, we have

1 (n − 1)!(k + r − 1)!
(n − 1) pn−2,k−1 (t)t r dt = .
0 (n + r − 1)!(k − 1)!
Thus

n−1
(1/n) (n − 1)!(k + r − 1)! (1/n)
Tn,r (x) = pn,k (x) + pn,n (x)
(n + r − 1)!(k − 1)!
k=1
n
(1/n) (n − 1)!(k + r − 1)!
= pn,k (x) .
(n + r − 1)!(k − 1)!
k=1
(1/n)
Substituting the value of pn,k (x) and using

n (−1)k (−n)k (−1)k (a)n
= , (a)n−k = ,0 ≤ k ≤ n
k k! (1 − a − n)k
we can write, by using the identity (k + r − 1)! = (r )k .(r − 1)! in the next step
124 V. Gupta

n
2.n! (−1)k (−n)k (−1)k (n − nx)n (n − 1)!(k + r − 1)!
Tn,r (x) = . (nx)k .
(2n)! k! (1 − 2n + nx)k (n + r − 1)!(k − 1)!
k=1
2.(r − 1)!.(n − 1)!n!(n − nx)n

n
(−n)k (nx)k (r )k 1
= .
(n + r − 1)!(2n)! (k − 1)!(1 − 2n + nx)k k!
k=1
2.(r − 1)!.(n − 1)!n!(n − nx)n (−n)k+1 (nx)k+1 (r )k+1

n
1
= . .
(n + r − 1)!(2n)! k!(1 − 2n + nx)k+1 (k + 1)!
k=0
Next, using (k + 1)! = (2)k , (a)k+1 = a(a + 1)k , we have
2.(r − 1)!.(n − 1)!n!(n − nx)n (−n)(−n + 1)k nx(nx + 1)k r (r + 1)k 1

n
Tn,r (x) = .
(n + r − 1)!(2n)! (2)k (1 − 2n + nx)(2 − 2n + nx)k k!
k=0
2.(r )!.nx(n − nx)n
n
(−n + 1)k (nx + 1)k (r + 1)k 1
= .
2n (2)k (2 − 2n + nx)k k!
(2n − nx − 1)(n + r − 1)! k=0
n
2nx.Γ (r + 1).(n − nx)n
= . 3 F2 (−n + 1, nx + 1, r + 1; 2, 2 − 2n + nx; 1),
2n
(2n − nx − 1)Γ (n + r )
n
where 3 F2 is the Hypergeometric polynomial. Applying the recurrence relations

between the parameters of a 3 F2 series
(a − d)(a − e)3 F2 (a − 1, b, c; d, e; z)
= a(a + 1)(1 − z)3 F2 (a + 2, b, c; d, e; z)
+a[d + e − 3a − 2 + z(2a − b − c + 1)]3 F2 (a + 1, b, c; d, e; z)
+[(2a − d)(2a − e) − a(a − 1) − z(a − b)(a − c)]3 F2 (a, b, c; d, e; z).
Substituting a = r + 1, b = −n + 1, c = nx + 1, d = 2, e = nx − 2n + 2 and
z = 1, the above relation reduces to the following:
(r − 1)(r + 2n − nx − 1)3 F2 (−n + 1, nx + 1, r ; 2, nx − 2n + 2; 1)

= −(r + 1)(n + r )3 F2 (−n + 1, nx + 1, r + 2; 2, nx − 2n + 2; 1)
+φ 3 F2 (−n + 1, nx + 1, r + 1; 2, nx − 2n + 2; 1),
where φ = (r + 1)(2r − nx + 3n − 3) + n(nx + x − 3) + 3. Since,

2n
(2n − nx − 1)
n (n + r )
. Tn,r (x) = 3 F2 (−n + 1, nx + 1, r + 1; 2, nx − 2n + 2; 1),
2nx(n − nx)n (r + 1)
then the result follows after simple computation, i.e.

(r + 1)(2r − nx + 3n − 3) + n(nx + x − 3) + 3
Tn,r +1 (x) = Tn,r (x)
(r + n)2
r (r − 1)(nx + 1 − r − 2n)
+ Tn,r −1 (x).
(n + r − 1)(n + r )2
(1/n)
Remark 1 By definition of operator using Pn (e0 , x) = 1 we have Tn,0 (x) = 1,
by applying Lemma 1, we get
n(n − 1)x 2 + (3n + 1)x

Tn,1 (x) = x, Tn,2 (x) = .
(n + 1)2
(1/n)
Remark 2 If we denote μn,r (x) = Dn ((t − x)r , x), then by Remark 1, we get
(3n + 1)x(1 − x)
μn,1 (x) = 0, μn,2 (x) = .
(n + 1)2
Moreover, we have
μn,1 (x) = O(n −[(m+1)/2] ).

(1/n)
Lemma 2 For f ∈ C [0, 1], we have Dn ( f, x) ≤ f , where ||.|| is the
sup-norm on [0, 1].
Proof From the definition of operator and Remark 1, we get

(1/n) (1/n)
Dn ( f, x) ≤ f Dn (1, x) = f .
Lemma 3 For n ∈ N , we have

3 2
(1/n)
Dn (t − x)2 , x ≤ δ (x),
n+1 n
where δn2 (x) = ϕ 2 (x) + n+1 ,

1
where ϕ 2 (x) = x(1 − x).
Proof By Remark 2, we have

(3n + 1)x(1 − x)
(1/n) 3 1
Dn (t − x) , x =
2
≤ ϕ (x) +
2
,
(n + 1)2 n+1 n+1
which is desired.
126 V. Gupta
3 Convergence Estimates
(1/n)
In this section, we present some convergence estimates of the operators Dn ( f, x) .
Theorem 1 Let f ∈ C[0, 1] and if f exists at a point x ∈ [0, 1], then

3x (1 − x)
(1/n)
lim n Dn ( f, x) − f (x) = f (x).
n→∞ 2
Proof By Taylor’s expansion of f, we have
(t − x)2
f (t) = f (x) + (t − x) f (x) + f (x) + ε(t, x)(t − x)2 ,
2
(1/n)
where ε(t, x) → 0 as t → x. Applying Dn on above Taylor’s expansion and
using Remark 2, we have
(1/n) (1/n) 1 (1/n)

Dn ( f, x) − f (x) = f (x)Dn ((t − x), x) +
f (x)Dn ((t − x)2 , x)
2
(1/n)
+Dn (ε(t, x)(t − x)2 , x).
Thus

(1/n)
lim n Dn ( f, x) − f (x)
n→∞
1 (1/n) (1/n)
= lim n f (x) Dn ((t − x)2 , x) + lim n Dn (ε(t, x)(t − x)2 , x)
n→∞ 2 n→∞
3x (1 − x)
(1/n)
= f (x) + lim n Dn ε (t, x) (t − x)2 , x
2 n→∞
3x (1 − x)
=: f (x) + F.
2
In order to complete the proof, it is sufficient to show that F = 0. By Cauchy–
Schwarz inequality, we have

1/2
1/2
(1/n) (1/n)
F = lim n Dn ε2 (t, x) , x Dn (t − x)4 , x . (5)
n→∞
Furthermore, since ε2 (x, x) = 0 and ε2 (., x) ∈ C[0, 1], it follows that

(1/n)
lim n Dn ε2 (t, x), x = 0, (6)
n→∞
uniformly with respect to x ∈ [0, 1]. Thus from (5), (6) and application of Remark
2, we get

1/2
1/2
(1/n) (1/n)
lim n Dn ε2 (t, x), x Dn (t − x)4 , x = 0.
n→∞
Thus, we have
3x (1 − x)
(1/n)
lim n Dn ( f, x) − f (x) = f (x),
n→∞ 2
which completes the proof.
To prove the next direct result, we

need the following auxiliary function
viz. Peetre’s
K -functional which for W 2 = g ∈ C [0, 1] : g , g ∈ C [0, 1] is defined as:

K 2 ( f, δ) = inf f − g + δ g : g ∈ W 2 (δ > 0),
where . is the uniform norm on C [0, 1] .

(1/n)
Theorem 2 For the operators Dn , there exists a constant C > 0 such that

(1/n)
Dn ( f, x) − f (x) ≤ Cω2 f, (n + 1)−1 δn (x) ,
1/2 √
where f ∈ C [0, 1], δn (x) = ϕ 2 (x) + n+1
1
, ϕ(x) = x(1 − x) and x ∈ [0, 1]
and the second-order modulus of continuity is given by
ω2 ( f, η) = sup sup | f (x + 2h) − 2 f (x + h) + f (x)| .

0<h≤η x,x+2h∈[0,1]
Proof By Taylor’s formula, we can write

t
g (t) = g (x) + (t − x) g (x) + (t − u) g (u) du.
x
Applying the above Taylor’s formula, we have

t
(1/n) (1/n)
Dn (g, x) = g (x) + Dn (t − u) g (u) du, x .
x
Hence
t
(1/n) (1/n)
Dn (g, x) − g (x) ≤ Dn |t − u| g (u) du, x
x

(1/n)
≤ Dn (t − x)2 , x g .
For f ∈ C [0, 1] and g ∈ W 2 , using Lemmas 2 and 3 we have

128 V. Gupta

(1/n) (1/n) (1/n)
Dn ( f, x) − f (x) ≤ Dn ( f − g, x) − ( f − g)(x) + Dn (g, x) − g (x)
3 2
≤ 2 f − g + δ (x) g .
n+1 n
Taking infimum over all g ∈ W 2 , we obtain

(1/n) 1 2
Dn ( f, x) − f (x) ≤ 3K 2 f, δ (x)
n+1 n
Using the inequality due to DeVore and Lorentz [2], there exists a positive constant
C > 0 such that √
K 2 ( f, δ) ≤ Cω2 f, δ ,
we get at once

(1/n)
Dn ( f, x) − f (x) ≤ Cω2 f, (n + 1)−1 δn (x) ,
so the proof is completed.
Remark 3 It is easy to construct operators of summation-integral type, but under

the integral sign in (4), the Pólya basis functions are not possible at this moment.
Also the simultaneous approximation as done in [1] for usual Bernstein–Durrmeyer
polynomials are not analogous for (4). We left it for readers, this can be considered
as on open problem.
References
1. Agrawal, P.N., Gupta, V.: Simultaneous approximation by linear combination of modified

Bernstein polynomials. Bull. Greek Math. Soc. 39, 29–43 (1989)
2. DeVore, R.A., Lorentz, G.G.: Constructive Approximation. Springer, Berlin (1993)
3. Derriennic, M.M.: Sur l’approximation de functionsintegrable sur [0, 1] par des polynomes de
Bernstein modifies. J. Approx. Theory 31, 323–343 (1981)
4. Durrmeyer, J.L.: Une formule d’ inversion de la Transformee de Laplace: applications a la
Theorie des Moments, These de 3e Cycle, Faculte des Sciences de l’ Universite de Paris (1967)
5. Finta, Z., Gupta, V.: Approximation by q-Durrmeyer operators. J. Appl. Math. Comput. 29(1–
2), 401–415 (2009)
6. Greubel, G.C.: Solution of the Problem 130 (Europeon Mathematical Socity Newsletter, March
2014). Eur. Math. Soc. Newsl, Sept (2014)
7. Gupta, V.: Some approximation properties on q-Durrmeyer operators. Appl. Math. Comput.
197(1), 172–178 (2008)
8. Gupta, V.: Europeon mathematical socity newsletter. Problem 130, 64 (2014)
9. Gupta, V., Agarwal R.P.: Convergence Estimates in Approximation Theory. Springer, New
York (2014)
10. Gupta, V., Lopez-Moreno, A.J., Latorre-Palacios, J.-M.: On simultaneous approximation of
the Bernstein Durrmeyer operators. Appl. Math. Comput. 213(1), 112–120 (2009)
11. Gupta, V., Rassias, ThM: Lupaş-Durrmeyer operators based on Polya distribution. Banach J.
Math. Anal. 8(2), 146–155 (2014)
12. Lupaş, L., Lupaş, A.: Polynomials of binomial type and approximation operators, Studia Univ.
Babes-Bolyai. Mathematica 32(4), 61–69 (1987)
13. Miclăuş, D.: The revision of some results for Bernstein Stancu type operators. Carpathian J.
Math. 28(2), 289–300 (2012)
14. Stancu, D.D.: Approximation of functions by a new class of linear polynomial operators. Rew.
Roum. Math. Pure. Appl. 13, 1173–1194 (1968)
Degree of Approximation by Certain
Genuine Hybrid Operators
Meenu Goyal and P.N. Agrawal
Abstract This paper is in continuation of our work on certain genuine hybrid

operators in (Positivity (Under review)) [3]. First, we discuss some direct results
in simultaneous approximation by these operators, e.g. pointwise convergence the-
orem, Voronovskaja-type theorem and an error estimate in terms of the modulus of
continuity. Next, we estimate the rate of convergence for functions having a derivative
that coincides a.e. with a function of bounded variation.
Keywords Rate of convergence · Modulus of continuity · Simultaneous approxi-

mation · Bounded variation
Mathematics Subject Classication (2010): 41A25 · 26A15 · 26A45
1 Introduction
Recently, Gupta and Rassias [5] introduced the Lupaş-Durrmeyer operators based on
Polya distribution and discussed some local and global direct results. Also, Gupta [2]
studied some other hybrid operators of Durrmeyer type. Păltǎnea [11] (see also [10])
considered a Durrmeyer-type modification of the genuine Szász-Mirakjan operators
based on two parameters α, ρ > 0. Inspired by his work, in [3] Gupta et al. introduced
certain genuine hybrid operators as follows:
For c ∈ {0, 1} and f ∈ Cγ [0, ∞) := { f ∈ C[0, ∞) : | f (t)| ≤ M f eγ t , for some
γ > 0, M f > 0}, we define
M. Goyal (B) · P.N. Agrawal

e-mail: meenu.goyaliitr@yahoo.com
P.N. Agrawal

DOI 10.1007/978-81-322-2485-3_10
132 M. Goyal and P.N. Agrawal
∞
∞
ρ
Bαρ ( f, x) = pα,k (x, c) θα,k (t) f (t)dt + pα,0 (x, c) f (0), (1)
k=1 0
∞
= K αρ (x, t) f (t)dt, (2)
0
where
(−x)k (k) ρ αρ −αρt

pα,k (x, c) = φα,c (x), θα,k (t) = e (αρt)kρ−1
k! (kρ)
∞
ρ
and K αρ (x, t) = pα,k (x, c)θα,k (t) + pα,0 (x, c)δ(t); x ∈ (0, ∞).
k=1
ρ
It is observed that the operators Bα ( f, x) are well-defined for αρ > γ . We assume
that

e−αx , for c = 0,
φα,c (x) = −α
(1 + x) , for c = 1.
As shown in paper [3], the operators (1) include several linear positive operators as
special cases. Further, we note that the operators (1) preserve the linear functions.
In [3], we studied some direct results, e.g. Voronovskaja-type theorems in ordinary
and simultaneous approximation for first-order derivatives as well as results in local
and weighted approximation. In this paper, we continue this work by discussing
simultaneous approximation for f (r ) (x), r ∈ N and the rate of convergence of the
operators (1) for the functions with derivatives of bounded variation on each finite
subinterval of (0, ∞). The paper is organized as follows:
In Sect. 2, we discuss some auxiliary results and then in Sect. 3, we obtain the
main results of this paper.
2 Auxiliary Results
For f : [0, ∞) → R, we define

∞

k
Sα ( f ; x) = pα,k (x, c) f (3)
α
k=0
such that (3) makes sense for all x ≥ 0.

Degree of Approximation by Certain Genuine Hybrid Operators 133
For m ∈ N0 = N ∪ {0}, the mth order central moment of the operators Sα is

given by
∞
m
k
υα,m (x) := Sα ((t − x)m ; x) = pα,k (x, c) −x .
α
k=0
Lemma 1 For the function υα,m (x), we have
υα,0 (x) = 1, υα,1 (x) = 0
and

x(1 + cx)[υα,m (x) + mυα,m−1 (x)] = αυα,m + 1 (x).
Thus,
(i) υα,m (x) is a polynomial in x of degree [m/2];
(ii) for each x ∈ [0, ∞), υα,m (x) = O(α −[(m + 1)/2] ) , where [β] denotes the integral
part of β.
Proof For the cases c = 0 and 1, the proof of this lemma can be found in [8, 12]
respectively.
Lemma 2 For the mth order (m ∈ N0 ) moment of the operators (1) defined as
ρ
u α,m (x) := Bα (t m ; x), we have
x 1
u α,0 (x) = 1, u α,1 (x) = x, u α,2 (x) = x +
2 + (1 + cx)
α ρ
and

x(1 + cx)u α,m (x) = αu α,m + 1 (x) − ρ + αx u α,m (x), m ∈ N.
m
Consequently, for each x ∈ (0, ∞) and m ∈ N, u α,m (x) = x m + α −1 ( pm (x, c) +

o(1)),
where pm (x, c) is a rational function of x depending on the parameters m and c.
Lemma 3 [3] For m ∈ N0 , if the mth order central moment μα,m (x) for the oper-
ρ
ators Bα is defined as
∞
∞
ρ ρ
μα,m (x) := Bα ((t − x)m , x) = pα,k (x, c) θα,k (t)(t − x)m dt + pα,0 (x, c)(−x)m ,
k=1 0
then we have the following recurrence relation:

1 m
αμα,m + 1 (x) = x(1 + cx)μα,m (x) + mx + (1 + cx) μα,m−1 (x) + μα,m (x).
ρ ρ
Consequently,
{1 + ρ(1 + cx)}x
(i) μα,0 (x) = 1, μα,1 (x) = 0, μα,2 (x) = ;
αρ
(ii) μα,m (x) is a polynomial in x of degree
atmost m;
(iii) for every x ∈ (0, ∞), μα,m (x) = O α −[(m + 1)/2] ;

1
(iv) the coefficients of α −m in μα,2m (x) and μα,2m−1 (x) are (2m − 1)!! x +

m ρ
(1 + cx)
m−2
(2m − 1)!!(m − 1) m−1 1 1
and x + (1 + cx) (1 + cx) + (1 +
3 ρ

ρ
2 1
2cx) + + (1 + cx) respectively.
ρ ρ
Corollary 1 For x ∈ [0, ∞) and α > 0, it is observed that
λx(1 + cx) 1
μα,2 (x) ≤ , where λ = 1 + > 1.
α ρ
Corollary 2 [3] Let γ and δ be any two positive real numbers and [a, b] ⊂ (0, ∞)
be any bounded interval. Then, for any m > 0 there exists a constant M independent
of α such that

∞ ρ
pα,k (x, c) θα,k (t)eγ t dt ≤ M α −m ,

k=1 |t−x|≥δ
where
.
is the sup-norm over [a, b].
Lemma 4 For every x ∈ (0, ∞) and r ∈ N0 , there exist polynomials qi, j,r (x) in x
independent of α and k such that
dr (qi, j,r (x, c))

p α,k (x, c) = p α,k (x, c) α i (k − αx) j ,
d xr ( p(x, c))r
2i + j≤r
i, j≥0
where p(x, c) = x(1 + cx).
Proof For the cases c = 0, 1, the proof of this lemma can be seen in [8, 12] respec-
tively.
3 Main Results
3.1 Simultaneous Approximation
Throughout this section, we assume that 0 < a < b < ∞.

ρ(r )
In the following theorem, we show that the derivative Bα ( f ; .) is also an approx-
(r
imation process for f . )
Theorem 1 (Basic convergence theorem) Let f ∈ Cγ [0, ∞). If f (r ) exists at a point

x ∈ (0, ∞), then we have
r
d ρ
lim B ( f ; ω) = f (r ) (x). (4)
α→∞ dωr α
ω=x
Further, if f (r ) is continuous on (a − η, b + η), η > 0, then the limit in (4) holds

uniformly in [a, b].
Proof By our hypothesis, we have

r
f (ν) (x)
f (t) = (t − x)ν + ψ(t, x)(t − x)r , t ∈ [0, ∞),
ν!
ν=0
where the function ψ(t, x) → 0 as t → x. From the above equation, we may write
r
dr ρ f (ν) (x) d r ρ ν
B ( f (t); ω) = B (t − x) ; ω)
dωr α ω=x ν! dωr α ω=x
ν=0
r
d
+ B ρ (ψ(t, x)(t − x)r ; ω)
dωr α ω=x
= : I1 + I2 , say.
First, we estimate I1 .
ν
f (ν) (x) d r ν
r
ν− j ρ j
I1 = (−x) Bα (t ; ω)
ν! dωr j ω=x
ν=0 j=0
r ν
r
f (ν) (x) ν d
= (−x)ν− j B ρ j
(t ; ω)
ν! j dωr α ω=x
ν=0 j=0
r −1 (ν)
ν r
f (x) ν ν− j d ρ j
= (−x) B (t ; ω)
ν! j dωr α ω=x
ν=0 j=0
r r
f (r ) (x) r r− j d ρ j
+ (−x) B (t ; ω)
r! j dωr α ω=x
j=0
:= I3 + I4 , say.
First, we estimate I4 .
r −1 r
f (r ) (x) r d ρ j f (r ) (x) d r ρ r
I4 = (−x)r − j Bα (t ; ω) + Bα (t ; ω)
r! j dωr ω=x r! dωr ω=x
j=0
:= I5 + I6 , say.

1 1 1
I6 = f (r ) (x) + O , I3 = O and I5 = O , as α → ∞.
α α α
Combining the above estimates, for each x ∈ (0, ∞) we obtain I1 → f (r ) (x) as
α → ∞.
Next, we estimate I2 . By making use of Lemma 4, we have
∞
∞
pα,k (x, c) ρ
|I2 | ≤ α i |k − αx| j |qi, j,r (x, c)| θα,k (t)|ψ(t, x)||(t − x)r |dt
( p(x, c))r
k=1 2i + j≤r 0
i, j≥0
r
d
+ pα,0 (ω, c) |ψ(0, x)(−x)r |

dωr ω=x
:= I7 + I8 , say.
Since ψ(t, x) → 0 as t → x, for a given ε > 0 there exists a δ > 0 such that
|ψ(t, x)| < ε whenever |t − x| < δ. For |t − x| ≥ δ, |(t − x)r ψ(t, x)| ≤ Meγ t , for
some constant M > 0.
Again, using Lemma 4, we have
∞

|qi, j,r (x, c)| ρ
|I7 | ≤ α |k − αx|
i j
pα,k (x, c) ε θα,k (t)|t − x|r dt
( p(x, c))r
k=1 2i + j≤r |t−x|<δ
i, j≥0

ρ
+M θα,k (t)eγ t dt := I9 + I10 , say.
|t−x|≥δ
|qi, j,r (x, c)|

Let K = sup . By applying the Schwarz inequality, Lemmas 1 and 3,
2i + j≤r ( p(x, c))r
i, j≥0
we get
∞
∞ 1
2
ρ
|I9 | ≤ εK α |k − αx| pα,k (x, c)
i j
θα,k (t)(t − x) dt
2r
k=1 2i + j≤r 0
i, j≥0

∞ 2 j 1
k 2
≤ εK αi + j −x pα,k (x, c)
α
2i + j≤r k=1
i, j≥0

∞ ∞ 1
2
ρ
pα,k (x, c) θα,k (t)(t − x)2r dt
k=1 0
1
2
i+j
≤ εK α υα,2 j (x) − x φα,c (x)
2j
2i + j≤r
i, j≥0
1
2
Bαρ ((t − x) ; x)) − x φα,c (x)
2r 2r

=ε α i + j {O(α − j ) + O(α −s1 )}1/2
2i + j≤r
i, j≥0
×{O(α −r ) + O(α −s2 )}1/2 , for any s1 , s2 > 0.
Choosing s1 , s2 such that s1 > j and s2 > r, we have |I9 | = ε

α i + j O(α − j/2 )O(α −r/2 ) = ε.O(1).
2i + j≤r
i, j≥0
Since ε > 0 is arbitrary, I9 → 0 as α → ∞.

Now, we estimate I10 . By applying Cauchy–Schwarz inequality, Lemma 1 and Corol-
lary 2, we obtain
∞

ρ
|I10 | ≤ M K α i |k − αx| j pα,k (x, c) θα,k (t)eγ t dt
k=1 2i + j≤r |t−x|≥δ
i, j≥0

∞ 2 j 1/2
k
≤ M1 αi + j −x pα,k (x, c)
α
2i + j≤r k=1
i, j≥0

∞ 1/2
ρ
× pα,k (x, c) θα,k (t)e2γ t dt , where M1 = M K
k=1 |t−x|≥δ
1/2
≤ M1 α i + j υα,2 j (x) − x 2 j φα,c (x)
2i + j≤r
i, j≥0

∞ 1/2
ρ
× pα,k (x, c) θα,k (t)e2γ t dt
k=1 |t−x|≥δ

= α i + j {O(α − j ) + O(α −m 1 )}1/2
2i + j≤r
i, j≥0
× {O(α −m 2 )}1/2 , for any m 1 , m 2 > 0.

Choosing m 1 > j, we get

|I10 | = α i + j O(α − j/2 )O(α −m 2 /2 ) = O(α (r −m 2 )/2 ),
2i + j≤r
i, j≥0
which implies that I10 = o(1), as α → ∞, on choosing m 2 > r. Next, we estimate

I8 . We may write
r
d

|I8 | = p (ω, c) |ψ(0, x)|x r
dωr
α,0
ω=x
(r )
= |φα,c (x)| |ψ(0, x)|x r .
(r ) (r ) (−1)r (α)r
Now, we observe that φα,0 (x) = e−αx (−α)r and φα,1 (x) = , which
(1 + x)α + r
implies that I8 = O(α − p ) for any p > 0, in view of the fact that |ψ(0, x)x r | ≤ N1 ,
for some N1 > 0.
By combining the estimates I7 − I10 , we obtain I2 → 0 as α → ∞.
To prove the uniformity assertion, it is sufficient to remark that δ(ε) in the above
proof can be chosen to be independent of x ∈ [a, b] and also that the other estimates
hold uniformity in x ∈ [a, b]. This completes the proof of the theorem.
Next, we establish an asymptotic formula.

Theorem 2 (Voronovskaja type result) Let f ∈ Cγ [0, ∞). If f admits a derivative
of order (r + 2) at a fixed point x ∈ (0, ∞), then we have
+2
r
dr ρ (r )
lim α B ( f ; ω) − f (x) = Q(ν, r, c, a, x) f (ν) (x), (5)
α→∞ dωr α ω=x ν=1
where Q(ν, r, c, a, x) are certain rational functions of x independent of α.

Further, if f (r + 2) is continuous on (a − η, b + η), η > 0, then the limit in (5) holds
uniformly in [a, b].
Proof From the Taylor’s theorem, for t ∈ [0, ∞) we may write
+2
r
f (ν) (x)
f (t) = (t − x)ν + ψ(t, x)(t − x)r + 2 , (6)
ν!
ν=0
where the function ψ(t, x) → 0 as t → x.

Now, from Eq. (6), we have
+2
r
dr ρ f (ν) (x) d r ρ ν
B ( f (t); ω) = (B ((t − x) ; ω)
dωr α ω=x ν! dwr α ω=x
ν=0
r
d
+ B ρ (ψ(t, x)(t − x)r + 2 ; ω)
dωr α ω=x
r+ 2 (ν) ν r
f (x) ν ν− j d ρ j
= (−x) B (t ; ω)
ν! j dωr α ω=x
ν=0 j=0
r
d
+ B ρ (ψ(t, x))(t − x)r + 2 ; ω
dωr α ω=x
:= J1 + J2 , say.
Proceeding in a manner similar to the estimate of I2 in Theorem 1, for each x ∈ (0, ∞)

we get α J2 → 0 as α → ∞.
Next, we estimate J1 .
r −1 (ν)
ν r
f (x) ν d
J1 = (−x)ν− j B ρ j
(t ; ω)
ν! j dωr α ω=x
ν=0 j=0
f (r ) (x)
r
r
r
d

+ (−x)r − j B ρ j
(t ; ω)
r! j dωr α ω=x
j=0
r +1 r
f (r + 1) (x) r + 1 r + 1− j d ρ j
+ (−x) B (t ; ω)
(r + 1)! j dωr α ω=x
j=0
r +2 r
f (r + 2) (x) r +2 d
+ (−x)r + 2− j B ρ j
(t ; ω) .
(r + 2)! j dωr α ω=x
j=0
Making use of Lemma 2, we have

r
+2
J1 = f (r ) (x) + α −1 Q(ν, r, c, a, x) f (ν) (x) + o(1) .
ν=1
Thus, from the estimates of J1 and J2 , the required result follows.

The uniformity assertion follows as in the proof of Theorem 1. This completes the
proof.
ρ(r )
The next result provides an estimate of the degree of approximation in Bα ( f ; x)
(r )
→ f (x), r ∈ N.
Theorem 3 (Degree of approximation) Let r ≤ q ≤ r + 2, f ∈ Cγ [0, ∞) and
f (q) exist and be continuous on (a − η, b + η) where η > 0 is sufficiently small.
Then, for sufficiently large α
r
d
B ρ
( f ; ω)
dωr α
ω=x

− f (r ) (x) ≤ max{C1 α −(q−r )/2 ω f (q) (α −1/2 , (a − η, b + η)), C2 α −1 },
C[a,b]
where C1 = C1 (r, c) and C2 = C2 (r, f, c).
Proof By our hypothesis we have,

q
f (i) (x) f (q) (ξ ) − f (q) (x)
f (t) = (t − x)i + (t − x)q χ (t) + φ(t, x)(1 − χ (t)),
i! q!
i=0
(7)
where ξ lies between t and x and χ (t) is the characteristic function of (a −η, b + η).
The function φ(t, x) for t ∈ [a, b] is bounded by Meγ t for some constant M > 0.
dr ρ
We operate Bα (.; ω) on the equality (7) and break the right-hand side into
dωr
three parts E 1 , E 2 and E 3 , say, corresponding to the three terms on the right-hand
side of Eq. (7).
Now, treating E 1 in a manner similar to the treatment of J1 of Theorem 2, we get
E 1 = f (r ) (x) + O(α −1 ), uniformly in x ∈ [a, b].
Making use of the inequality

(q) (q) |t − x|
|f (ξ ) − f (x)| ≤ 1 + ω f (q) (δ), δ > 0,
δ
and Lemma 4, we get

∞
ω f (q) (δ) α i |k − αx| j |qi, j,r (x, c)|
|E 2 | ≤ pα,k (x, c)
q! ( p(x, c))r
k=1 2i + j≤r
i, j≥0
∞
ρ |t − x|
× θα,k (t) 1+ |t − x|q χ (t)dt
δ
0

x q + 1 (r )
+ x + q
φα,c (x)
δ
= E4 + E5.
Finally, let
|qi, j,r (x, c)|
S1 = sup sup ,
x∈[a,b] 2i + j≤r ( p(x, c))r
i, j≥0
then by applying Schwarz inequality, Lemmas 1 and 3, we obtain

∞ 2 j 1/2
ω f (q) (δ)S1 i+j k
E4 ≤ α −x pα,k (x, c)
q! α
2i + j≤r k=1
i, j≥0

∞ ∞ 1/2
ρ
× pα,k (x, c) θα,k (t)(t − x)2q dt
k=1 0

∞ ∞ 1/2
1 ρ 2q + 2
+ pα,k (x, c) θα,k (t)(t − x) dt
δ
k=1 0
1/2
i+j
≤ ω f (q) (δ)S1 α υα,2 j (x) − x φα,c (x)
2j
2i + j≤r
i, j≥0
1/2
× Bαρ ((t − x) ; x) − x φα,c (x)
2q 2q
1/2
1
+ Bαρ ((t − x)2q + 2 ; x) − x 2q + 2 φα,c (x)
δ

= ω f (q) (δ) α i + j {O(α − j ) + O(α −s1 )}1/2
2i + j≤r
i, j≥0
1
×{(O(α −q ) + O(α −s2 )}1/2 + {(O(α −(q + 1) )
δ
+ O(α −s3 ))}1/2 , f or any s1 , s2 , s3 > 0.
Choosing s1 , s2 , s3 such that s1 > j, s2 > q, s3 > q + 1, we have

1 1 1 1
|E 4 | = ω f (q) (δ) αi + j O O + O .
α j/2 α q/2 δ α (q + 1)/2
2i + j≤r
i, j≥0
Now, on choosing δ = α −1/2 , we get
|E 4 | ≤ C1 α −(q−r )/2 ω f (q) (α −1/2 , (a − η, b + η)).
Next, proceeding in a manner similar to the estimate of I8 in Theorem 1, we have

E 5 = O(α − p ), for any p > 0. Choosing p > 1, we have E 5 = O(α −1 ), as α → ∞.
Finally, proceeding along the lines of the estimate of I10 of Theorem 2, we obtain
E 3 = o(α −1 ) as α → ∞.
On combining the estimates of E 1 − E 5 , we get the required result.
In this section, we shall estimate the rate of convergence for the generalized hybrid
ρ
operators Bα for functions with derivatives of bounded variation. In recent years,
several researchers have obtained results in this direction for different sequences of
linear positive operators. We refer the reader to some of the related papers (cf. [1, 4,
6, 7, 9], etc.).
Let f ∈ D BVγ [0, ∞), γ ≥ 0 be the class of all functions defined on [0, ∞),
having a derivative that coincides, a.e. with a function of bounded variation on every
finite subinterval of [0, ∞) and | f (t)| ≤ Mt γ , ∀ t > 0.
It turns out that for f ∈ D BVγ [0, ∞), we may write
x
f (x) = g(t)dt + f (0),
0
where g(t) is a function of bounded variation on each finite subinterval of [0, ∞).
Lemma 5 For all x ∈ (0, ∞), λ > 1 and α sufficiently large, we have
t
ρ 1 λx(1 + cx)
(i) λα (x, t) = K αρ (x, u)du ≤ , 0 ≤ t < x;
(x − t) 2 α
0
∞
ρ 1 λx(1 + cx)
(ii) 1 − λα (x, z) = K αρ (x, u)du ≤ , x < z < ∞.
(z − x)2 α
z
Proof First we prove (i).

t t 2
x −u
λρα (x, t) = K αρ (x, u)du ≤ K αρ (x, u)du
x −t
0 0
1
≤ B ρ ((u − x)2 ; x)
(x − t)2 α
1 λx(1 + cx)
≤ .
(x − t) 2 α
The proof of (ii) is similar.
Theorem 4 Let f ∈ D BVγ [0, ∞), γ ≥ 0. Then for every x ∈ (0, ∞), r (∈ N) >
2γ and sufficiently large α, we have

f (x + ) − f (x−) λx(1 + cx) 1/2

|Bαρ ( f ; x) − f (x)| ≤

2 α
x+ √x √ x
[α] x +
x λ(1 + cx) m
α

+√ ( fx ) + ( fx )
α x α x
x− √α m=1 x− m

1/2
λx(1 + cx)
+ | f (x + )|
α
λ(1 + cx)
+ | f (2x) − f (x) − x f (x + )|
αx
A(r, x) λ(1 + cx)
+M + | f (x)| ,
α γ /2 αx
where ⎧
⎪
⎨ f (t) − f (x + ), x < t < ∞

f x (t) = 0 t=x ,
⎪
⎩
f (t) − f (x−), 0 ≤ t < x
b (x))
a( f is the total variation of f x on [a, b], A(r, x) is a constant depending on
r and x and M is a constant depending on f and γ .
Proof By the hypothesis, we may write

1
f (t) = f (x + ) + f (x−) + f x (t)
2

1
+ f (x + ) − f (x−) sgn(t − x)
2

1
+ δx (t) f (t) − f (x + ) + f (x−) , (8)
2
where
1 t=x
δx (t) =
0 t = x.
From Eqs. (2) and (8), we have
∞ ∞
ρ ρ ρ
Bα ( f ; x) − f (x) = K α (x, t) f (t)dt − f (x) = ( f (t) − f (x))K α (x, t)dt
0 0
x ∞
ρ ρ
= ( f (t) − f (x))K α (x, t)dt + ( f (t) − f (x))K α (x, t)dt
0 x
x x ∞ t
ρ ρ
=− f (u)du K α (x, t)dt + f (u)du K α (x, t)dt
0 t x x
= I1 (x) + I2 (x), say.
Using Eq. (8), we get
x x
1 1
I1 (x) = f (x + ) + f (x−) + f x (u) + f (x + ) − f (x−) sgn(u − x)
2 2
0 t

1 ρ
+ δx (u) f (u) − f (x + ) + f (x−) du K α (x, t)dt.
2
t
Since δx (u)du = 0, we have
x
x x t
1 ρ ρ
I1 (x) = f (x + ) + f (x−) (x − t)K α (x, t)dt + f x (u)du K α (x, t)dt
2
0 0 x
x
1 ρ
− f (x + ) − f (x−) |x − t|K α (x, t)dt. (9)
2
0
Proceeding similarly, we find that
∞ ∞ t
1 ρ ρ
I2 (x) = f (x + ) + f (x−) (t − x)K α (x, t)dt +
f x (u)du K α (x, t)dt
2
x x x
∞
1 ρ
+ f (x + ) − f (x−) |t − x|K α (x, t)dt. (10)
2
x
By combining (9) and (10), we get
∞
ρ 1 ρ
Bα ( f ; x) − f (x) = f (x + ) + f (x−) (t − x)K α (x, t)dt
2
0
∞
1 ρ
+ f (x + ) − f (x−) |t − x|K α (x, t)dt
2
0
x x ∞ t
ρ ρ
− f x (u)du K α (x, t)dt + f x (u)du K α (x, t)dt.
0 t x x
Hence

f (x + ) + f (x−) ρ
|Bαρ ( f ; x) − f (x)| ≤ |B (t − x; x)| + f (x + ) − f (x−) B ρ (|t − x|; x)
2 α 2 α
x x ∞ t

+ f x (u)du K αρ (x, t)dt + f x (u)du K αρ (x, t)dt .
0 t x x
(11)
On application of Lemma 5 and integration by parts, we obtain
x x x x x
ρ ∂ ρ ρ
f x (u)du K α (x, t)dt = f x (u)du λα (x, t)dt = f x (t)λα (x, t)dt.
∂t
0 t 0 t 0
Thus,
x x x

(u)du K ρ (x, t)dt ≤ ρ

f x α | f x (t)|λα (x, t)dt
0 t 0
x− √xα
x
ρ ρ
≤ | f x (t)|λα (x, t)dt + | f x (t)|λα (x, t)dt.
0 x− √xα
ρ
Since f x (x) = 0 and λα (x, t) ≤ 1, we get
x x x
x
| f x (t)|λρα (x, t)dt = | f x (t) − f x (x)|λρα (x, t)dt ≤ ( f x )dt
t
x− √xα x− √xα x− √xα
x
x
x x
≤ ( f x ) dt = √ ( f x ).
α
x− √xα x− √xα
x
x− √α
Similarly, using Lemma 5 and putting t = x − ux , we get
x− √xα x− √xα

λx(1 + cx) dt
| f x (t)|λρα (x, t)dt ≤ | f x (t)|
α (x − t)2
0 0
x− √xα

x
λx(1 + cx) dt
≤ ( f x )
α t
(x − t)2
0
√ √
α [ α] x
λ(1 + cx)
x
λ(1 + cx)

= ( f x )du ≤ ( f x ).
α x α x
1 x− u m=1 x− m
Consequently,
√
x x [ α] x
x
x
λ(1 + cx)
f x (u)du K αρ (x, t)dt ≤ √ ( f x ) + ( f x ).
α α
x x
0 t x− √α m=1 x− m
(12)
Also, we have
∞ t 2x t
∂
f
(u)du K ρ
(x, t)dt ≤ f
(u)du (1 − λ ρ
(x, t))dt
x α x
∂t α
x x x x
∞ t

+ f x (u)du K αρ (x, t)dt
2x x
∞

≤ ( f (t) − f (x))K α (x, t)dt
ρ
2x
∞

+ | f (x + )| (t − x)K α (x, t)dt
ρ
2x
2x

+ f x (u)du |1 − λρα (x, 2x)|
x
2x
+ | f x (t)|(1 − λρα (x, t))dt.
x
Applying Lemma 5, we get
∞ t ∞ ∞
ρ γ ρ ρ

f x (u)du K α (x, t)dt ≤ M t K α (x, t)dt + | f (x)| K α (x, t)dt

x x 2x 2x

λx(1 + cx) 1/2

+ | f (x + )|
α
λ(1 + cx)
+ | f (2x) − f (x) − x f (x + )|
αx
x+ √x √ x
α [ α] x +
x λ(1 + cx) m
+√ ( fx ) + ( f x ).
α x
α
m=1 x
(13)
We note that we can choose r ∈ N such that 2r > γ .

Since t ≤ 2(t − x) and x ≤ t − x when t ≥ 2x, using Hölder’s inequality and
Lemma 3, we obtain
∞ ∞
M t γ K αρ (x, t)dt + | f (x)| K αρ (x, t)dt
2x 2x
∞ ∞
γ γ ρ | f (x)|
≤ 2 M (t − x) K α (x, t)dt + (t − x)2 K αρ (x, t)dt
x2
2x 2x
∞ γ /2r
γ λ(1 + cx)
≤2 M (t − x)2r K αρ (x, t)dt + | f (x)|
αx
0
A(r, x) λ(1 + cx)
≤ M γ /2 + | f (x)| , where M = 2γ M. (14)
α αx
Using Lemma 3 and combining (11), (12), (13) and (14), we get the required result.
Acknowledgments The authors are extremely grateful to the reviewers for careful reading of the
manuscript and for making valuable comments leading to better presentation of the paper. The first
author is thankful to the “Council of Scientific and Industrial Research” India for financial support
to carry out the above research work.
References
1. Acar, T., Gupta, V., Aral, A.: Rate of convergence for generalized Szasz operators. Bull. Math.
Sci. 1(1), 99–113 (2011)
2. Gupta, V.: A new class of Durrmeyer operators. Adv. Stud. Contemp. Math. (Kyungshang)
23(2), 219–224 (2013)
3. Gupta, V., Agrawal, P.N., Goyal, M.: Approximation by certain genuine hybrid operators.
Positivity (Under review)
4. Gupta, V., Agarwal, R.P.: Convergence Estimates in Approximation Theory. Springer, New
York (2014)
5. Gupta, V., Rassias, ThM: Lupaş Durmeyer operators based on Polya distribution. Banach J.
Math. Anal. 8(2), 146–155 (2014)
6. Ispir, N.: Rate of convergence of generalized rational type Baskakov operators. Math. Comput.
Model. 46, 625–631 (2007)
7. Karsli, H.: Rate of convergence of new gamma type operators for functions with derivatives of
bounded variation. Math. Comput. Model. 45, 617–624 (2007)
8. Kasana, H.S., Prasad, G., Agrawal, P.N., Sahai, A.: Modified Szasz operators. In: Proceedings
of the International Conference on Mathematical Analysis and its Applications, Kuwait (1985)
9. Özarslan, M.A., Duman O., Kaanoǧlu, C.: Rates of convergence of certain King-type operators
for functions with derivative of bounded variation. Math. Comput. Model. 52, 334–345 (2010)
10. Păltǎnea, R.: Modified Szász-Mirakjan operators of integral form. Carpathian J. Math. 24,
378–385 (2008)
11. Păltǎnea, R.: Simultaneous approximation by a class of Szász-Mirakjan operators. J. Appl.
Funct. Anal. 9(3–4), 356–368 (2014)
12. Sinha, R.P., Agrawal, P.N., Gupta, V.: On simultaneous approximation by modified Baskakov
operators. Bull. Soc. Math. Belg. 43, 217–231 (1991)
Frames in Semi-inner Product Spaces
N.K. Sahu and Ram N. Mohapatra
Abstract The objective of this paper is to study the theory of frames in semi-inner
product spaces. Several researchers have studied frames in Banach spaces by using
the bounded linear functionals. Application of semi-inner product is a new approach
to investigate the theory of frames. The notion of semi-frame is introduced in this
new aspect.
Keywords Frames · Semi-frames · Semi-inner product
1 Introduction
The theory of frames plays a fundamental role in signal processing, image processing,
data compression, sampling theory and has found considerable applications in many
more fields. Mathematically, the frame is equivalent to a spanning set in a vector
space, but it may not be minimal. It may have more number of vectors than a basis.
One of the main advantages in using frames in signal transmission over a basis is
that if in the process of transmission, signal along a frame is lost, it is possible to
reconstruct completely due to the built in redundancy which is not possible while
using a basis. In applications one determines “optimal frames with erasers” (see Han
and Sun [12], Pehlivan et al. [15] and the references there in).
The main objective of this paper is to describe Frames in Hilbert space, Banach
space, Hilbert C ∗ -module and then define semi-inner product space, and Frames and
semi-frames in that context. The advantage of using semi-inner product is to facilitate
N.K. Sahu (B)

Dhirubhai Ambani Institute of Information and Communication Technology,
Gandhinagar, India
e-mail: nabindaiict@gmail.com
R.N. Mohapatra
University of Central Florida, Orlando, FL 32816, USA
e-mail: ramm@mail.ucf.edu

DOI 10.1007/978-81-322-2485-3_11
150 N.K. Sahu and R.N. Mohapatra
calculations in uniformly convex smooth Banach spaces and obtain results that pose
difficulties.
Frames in Hilbert Spaces

Frames for a Hilbert space were formally defined by Duffin and Schaeffer [5] in
1952. Frames in Hilbert space have been well investigated. For Hilbert space frames
one can refer to Christensen [3] and the references there in.
Definition 1 Let H be a Hilbert space and I be an index set. A sequence { f i }i∈I of
elements in H is called a Bessel sequence for H if there exists a real constant B > 0
such that

| f, f i |2 ≤ B f 2 for all f ∈ H. (1)
i∈I
B is called the Bessel bound for the Bessel sequence { f i }.
Definition 2 A sequence { f i }i∈I of elements in a Hilbert space H is called a frame

for H , if there exist real constants A, B with 0 < A ≤ B < ∞ such that

A f 2 ≤ | f, f i |2 ≤ B f 2 for all f ∈ H. (2)
i∈I
Here A and B are called lower and upper frame bounds, respectively.
The largest number A and the smallest number B satisfying the frame inequality (2)
for all f ∈ H are called optimal frame bounds. If A = B, we call the frame { f i }, a
tight frame. When A = B = 1, the frame is called a Parseval frame. If all the frame
elements have the same norm, then the frame is called equal norm frame, and if all
the frame elements are of unit norm, then it is called unit norm frame. A frame is
exact if it ceases to be a frame when any one of its element is removed. A frame is
exact if and only if it is a Riesz basis. A non-exact frame is called over complete in the
sense that if at least one vector is removed, the remaining ones still constitute a frame.
Frames in Banach Spaces

While constructing frames in Hilbert space we need the sequence space l 2 . Similarly,
while constructing frames in Banach space one needs a Banach space of scalar-valued
sequences (BK-space).
A Banach space of scalar valued sequences (BK-space) is a linear space of
sequences with a norm which makes it a Banach space and for which the coor-
dinate functionals are continuous. In a BK-space, the unit vectors are defined by
ei ( j) = δi j (Kronecker delta). Gröchenig [7] first generalized the concept of frames
to Banach spaces and called them atomic decompositions.
Definition 3 Let X be a Banach space with norm . X and X d be an associated
BK-space with norm . X d . Let { f i } be a sequence of elements in X ∗ , the dual space
of X , and {xi } be a sequence of elements in X . If
Frames in Semi-inner Product Spaces 151
(i) {(x, f i )} ∈ X d , for all x ∈ X , where (x, f i ) denotes the value of the functional
f i at the point x,
x X and {(x, f i )} X d are equivalent,
(ii) the norms

(iii) x = (x, f i )xi , for all x ∈ X , then the pair { f i }, {xi } is called an atomic
i
decomposition of X with respect to X d .
With a more general setting Gröchenig defined Banach frames as follows:
Definition 4 Let X be a Banach space with norm . X and X d be an associated
BK-space with norm . X d . Let { f i } be a sequence of elements in X ∗ and an operator
S : X d → X be given. If
(i) {(x, f i )} ∈ X d , for all x ∈ X ,
(ii) the norms x X and {(x, f i )} X d are equivalent,
(iii) S is bounded and linear, and S((x, f i )) = x for each x ∈ X , then { f i }, S is a
Banach frame for X with respect to X d .
There is considerable research on frames in Banach spaces and for details on frames
in Banach spaces one may refer to Christensen and Heil [4], Stoeva [17], Casazza
and Christensen [2], Koushik [13].
Frames in Hilbert C ∗ -module
In recent years, many mathematicians generalized the frame theory in Hilbert spaces
to frame theory in Hilbert C ∗ -modules and got significant results which enrich the
theory of frames.
Definition 5 Let A be a unital C ∗ -algebra and J be a finite or countable index set.
A sequence {x j } j∈J of elements in a Hilbert A -module H is said to be a frame if
there exists two real constants A, B > 0 such that

Ax, x ≤ x, x j x j , x ≤ Bx, x (3)
j∈J
for every x ∈ H . The optimal constants (maximal for A and minimal for B) are
called frame bounds.
The frame {x j } j∈J is said to be a tight frame if A = B, and said to be a Parseval
frame if A = B = 1.
Wu Jing in his doctoral dissertation to University of Central Florida gave an
equivalent formulation of modular frames, and derived many interesting results.
Details about these can be found in Han et al. [8–11].
Due to lack of inner product structure in general Banach spaces, people studied
the theory of frames by taking the help of bounded linear functionals, that is by
taking the help of the dual space. Many of the results on classical frame theory have
been generalized to Banach spaces, in this way. The use of arbitrary bounded linear
functionals is not always a convenient way to study these notions. It is also difficult
to construct examples to verify the established theoretical results. So, in this paper
we have put some effort to study the theory of frames in Banach spaces in a different
way. We have seen frames being defined in semi-inner product spaces (see Zhang and
Zhang [18]). We use the notion of semi-inner product to study more into the theory
of frames. In the next section, we give a brief introduction of semi-inner product
space. It is worth mentioning that this approach will result for frames in l p and L p
spaces for 1 < p < ∞.
2 Semi-inner Product Space
Definition 6 (Lumer [14]) Let X be a vector space over the field F of real or complex
numbers. A functional [., .] : X × X → F is called a semi-inner product if it satisfies
the following:
1. [x + y, z] = [x, z] + [y, z], ∀x, y, z ∈ X ;
2. [λx, y] = λ[x, y], ∀λ ∈ F and x, y ∈ X ;
3. [x, x] > 0, for x = 0;
4. |[x, y]|2 ≤ [x, x][y, y].
The pair (X, [., .]) is called a semi-inner product space.
1
We observe that x = [x, x] 2 is a norm on X . Hence every semi-inner product
space is a normed linear space. On the other hand, in a normed linear space, one can
generate semi-inner product in infinitely many different ways. Giles [6] had proved
that if the underlying space X is a uniformly convex smooth Banach space then it is
possible to define a semi-inner product, uniquely. Also the unique semi-inner product
has the following nice properties:
(i) [x, y] = 0 if and only if y is orthogonal to x, that is if and only if y ≤
y + λx, for all scalars λ.
(ii) Generalized Riesz representation theorem: If f is a continuous linear functional
on X then there is a unique vector y ∈ X such that f (x) = [x, y], for all x ∈ X .
(iii) The semi-inner product is continuous, that is for each x, y ∈ X , we have
Re[y, x + λy] → Re[y, x] as λ → 0.
The sequence space l p , p > 1 and the function space L p , p > 1 are uniformly
convex smooth Banach spaces. So one can define semi-inner product on these spaces,
uniquely.
Example 1 The real sequence space l p for 1 < p < ∞ is a semi-inner product
space with the semi-inner product defined by
1
[x, y] = p−2
xi yi |yi | p−2 , x, y ∈ l p .
y p i
Example 2 (Giles [6]) The real Banach space L p (X, μ) for 1 < p < ∞ is a
semi-inner product space with the semi-inner product defined by

1
[ f, g] = p−2
f (x)|g(x)| p−1 sgn(g(x))dμ, f, g ∈ L p .
g p X
3 Frames in Semi-inner Product Spaces
Recently, Zhang and Zhang [18] investigated the theory of frames in Banach spaces
by applying the notion of semi-inner product. They generalized the classical theory
on frames and Riesz bases in this new perspective. They have defined frames in the
following way:
Definition 7 Let X be a Banach space with a compatible semi-inner product [., .]

and norm . X . Let X d be an associated BK -space (sequence space with continuous
coordinate linear functionals) with norm . X d . A sequence of elements { f j } ⊆ X
is called an X d -frame for X if {[ f, f j ]} ∈ X d , for all f ∈ X and there exist two
positive constants A, B such that
A f X ≤ {[ f, f j ]} X d ≤ B f X for all f ∈ X.
They have also defined frames for the dual space X ∗ of the Banach space X .

and norm . X . Let X ∗ be the dual space of X . Let X d be an associated BK -space
with norm . X d , and X d∗ be the dual space of X d . A sequence of elements { f j∗ } ⊆ X ∗
is an X d∗ -frame for X ∗ if {[ f j , f ]} ∈ X d∗ , for all f ∈ X and there exist two positive
constants A, B such that
A f X ≤ {[ f j , f ]} X d∗ ≤ B f X for all f ∈ X.
The notion of frame is too restrictive, in the sense that one cannot satisfy both
upper and lower frame bounds simultaneously. Thus there is a scope for two natural
generalizations, named as upper semi-frame and lower semi-frame. The notion of
semi-frame in Hilbert space was studied by Antoine and Balazs [1]. In this paper we
define the notion of semi-frame in Banach spaces by using the semi-inner product.

and norm . X . Let X d be an associated BK -space with norm . X d . A sequence of
elements { f j } ⊆ X is called upper semi-X d -frame for X if
(i) { f j } is total in X ;
(ii) {[ f, f j ]} ∈ X d , for all f ∈ X ;
(iii) there exists a positive constant B such that
0 ≤ {[ f, f j ]} X d ≤ B f X for all f ∈ X.

and norm . X . Let X d be an associated BK -space with norm . X d . A sequence of
elements { f j } ⊆ X is called lower semi-X d -frame for X if
(i) { f j } is total in X ;
(ii) {[ f, f j ]} ∈ X d , for all f ∈ X ;
(iii) there exists a positive constant A such that
A f X ≤ {[ f, f j ]} X d for all f ∈ X.
Similarly, we define upper semi-X d∗ -frame and lower semi-X d∗ -frame for the dual
space X ∗ .

is upper semi-X d∗ -frame for X ∗ if
(i) { f j∗ } is total in X ∗ ;
(ii) {[ f j , f ]} ∈ X d∗ , for all f ∈ X ;
(iii) there exists a positive constant B such that
0 ≤ {[ f j , f ]} X d∗ ≤ B f X for all f ∈ X.

is lower semi-X d∗ -frame for X ∗ if
(i) { f j∗ } is total in X ∗ ;
(ii) {[ f j , f ]} ∈ X d∗ , for all f ∈ X ;
(iii) there exists a positive constant A such that
A f X ≤ {[ f j , f ]} X d∗ for all f ∈ X.
Zhang and Zhang [18] established the reconstruction property for X d -frame and
X d∗ -frame in a semi-inner space X . They defined the operator (so-called analysis
operator) U : X → X d by U ( f ) = {[ f, f j ]}. They proved that
Theorem 1 ([18]) If { f j } is an X d -frame for X and Rng(U ) has an algebraic

complement in X d , then there exists an X d∗ -frame {g ∗j } for X ∗ such that

f = [ f, f j ]g j for all f ∈ X
j∈I
and
f∗ = [g j , f ] f j∗ for all f ∈ X.
j∈I
Based on the above theorem, we formulate the following definition.

Definition 13 Let { f j } be an X d -frame for X . If there exists an X d∗ -frame {g ∗j } for
X ∗ such that
f = [ f, f j ]g j for all f ∈ X
j∈I
and
f∗ = [g j , f ] f j∗ for all f ∈ X.
j∈I
Then { f j } and {g ∗j } are called dual frame pair.

Now we are in a position to propose the following theorem.
Theorem 2 Let { f j } be an upper semi-X d -frame for X with bound B. If {g ∗j } is a
sequence of element in X ∗ such that { f j } and {g ∗j } are dual frame pair, then {g ∗j } is
a lower semi-X d∗ -frame for X ∗ with bound B1 .
Proof Since { f j } is an upper semi-X d -frame for X with bound B, we have
0 ≤ {[ f, f j ]} X d ≤ B f X for all f ∈ X.
Now

f 2X = [ f, f ] = [ f, f j ]g j , f

= [[ f, f j ]g j , f ]

= [ f, f j ][g j , f ]
≤ {[ f, f j ]} X d {[g j , f ]} X d∗
≤ B f X {[g j , f ]} X d∗
1
⇒ f X ≤ {[g j , f ]} X d∗ for all f ∈ X.
B
That is, {[g j , f ]} is a lower semi-X d∗ -frame for X ∗ .
Similarly, we can easily prove the following theorem.

Theorem 3 Let {g j } be an upper semi-X d∗ -frame for X ∗ with bound B. If { f j } is a
sequence of element in X such that { f j } and {g ∗j } are dual frame pair, then { f j } is a
lower semi-X d -frame for X with bound B1 .
Proof Since {g j } is an upper semi-X d∗ -frame for X ∗ with bound B, we have
0 ≤ {[g j , f ]} X d∗ ≤ B f X for all f ∈ X.
Now for any f ∈ X , we have

f 2X = [ f, f ] = [ f, f j ]g j , f

= [[ f, f j ]g j , f ]

= [ f, f j ][g j , f ]
≤ {[ f, f j ]} X d {[g j , f ]} X d∗
≤ {[ f, f j ]} X d B f X
1
⇒ f X ≤ {[ f, f j ]} X d for all f ∈ X.
B
That is, {[ f, f j ]} is a lower semi-X d -frame for X .
4 Frames for l p (1 < p < ∞) Spaces
In this section, we define frames in l p (1 < p < ∞) spaces.

We know that l p (1 < p < ∞) spaces are uniformly convex smooth Banach
spaces. It is seen that those spaces are semi-inner product spaces with uniquely
defined semi-inner product (see Giles [6]). For the rest of this section, we assume
that X is the real sequence space l p (1 < p < ∞) with norm . p and semi-inner
product [., .]. The following definitions of Bessel sequence and frame can be found
in Sahu and Nahak [16].
∞ ⊆ X is called a Bessel sequence if
Definition 14 A set of elements f = { f i }i=1
there exists a constant B > 0 such that
∞

|[ f i , x]|q ≤ B(x p )q , ∀x ∈ X,
i=1
where 1 < p, q < ∞ and 1

p + 1
q = 1. The number B is called Bessel bound.
∞ in X is called a frame if there exist
Definition 15 A sequence of elements { f i }i=1
positive constants A and B such that
∞

A(x p ) ≤
q
|[ f i , x]|q ≤ B(x p )q , ∀x ∈ X,
i=1
where 1 < p, q < ∞ and 1

p + 1
q = 1. A and B are called lower and upper frame
bounds, respectively.
If A = B then the frame is called a tight frame, and if A = B = 1 then the frame is
called a Parseval frame. A frame is called a normalized frame if each frame element
has unit norm.
∞ ∈ l p , where e = (0, 0, ..., 1, 0, 0..), where 1 is
Example 3 Consider the set {ei }i=1 i
at the ith coordinate and 0 at the other coordinates.
(i) {e1 , 0, e2 , 0, e3 , 0, ....} is a Parseval frame.
{e1 , e1 , e2 , e2 , ......} is a tight frame with bound 2.
(ii)
(iii) √e1 , √e1 , √e2 , √e2 , .... is a tight frame with bound √
2
p .
2 2 2 2 ( 2) p−1
(iv) {nen }∞ is a lower semi-frame but not a frame.

n=1
∞
(v) n1 en n=1 is an upper semi-frame but not a frame.
Some of the classical frame theory results in Hilbert spaces can be generalized to
l p spaces in this new approach. The reconstruction formula naturally holds true for
Parseval frames and tight frames. In this connection, we state the following two
theorems.
∞ is a Parseval frame for X if and only if
Theorem 4 A set of elements { f i }i=1
∞
|[ f i , x]|q−2
x= [ f i , x] f i , ∀x ∈ X. (4)
{[ f i , x]}q−2
i=1
∞ is a tight frame with bound A for X if and

Theorem 5 A set of elements { f i }i=1
only if
∞
1 |[ f i , x]|q−2
x= [ f , x] f i ∀x ∈ X.
q−2 i
(5)
A q {[ f i , x]}
2
i=1
Acknowledgments We thank the referees for their comments which have been used in the final
presentation.
References
1. Antoine, J.P., Balazs, P.: Frames and semi-frames. J. Phys. A: Math. Theor. 44(20), 205201
(2011)
2. Casazza, P.G., Christensen, O.: The reconstruction property in Banach spaces and perturbation
theorem. Canad. Math. Bull. 51, 348–358 (2008)
3. Christensen, O.: An Introduction to Frames and Riesz Bases. Birkhauser, Boston (2002)
4. Christensen, O., Heil, C.: Perturbations of Banach frames and atomic decompositions. Math.
Nachr. 158, 33–47 (1997)
5. Duffin, R.J., Schaeffer, A.C.: A class of nonharmonic Fourier series. Trans. Amer. Math. Soc.
72, 341–366 (1952)
6. Giles, J.R.: Classes of semi-inner product spaces. Trans. Amer. Math. Soc. 129, 436–446 (1967)
7. Gröchenig, K.: Localization of frames, Banach frames, and the invertibility of the frame oper-
ator. J. Four. Anal. Appl. 10, 105–132 (2004)
8. Han, D., Jing, W., Mohapatra, R.N.: Structured Parseval frames in Hilbert C ∗ modules. Con-
temp. Math. (American Math. Soc.) 414, 275–288 (2006)
9. Han, D., Jing, W., Larson, D., Mohapatra, R.N.: Riesz bases and their dual modular frames in
Hilbert C ∗ modules. J. Math. Anal. Appl. 343, 246–256 (2008)
10. Han, D., Jing, W., Mohapatra, R.N.: Perturbation of frames and Riesz bases in Hilbert C ∗
modules. Linear Algebra Appl. 431, 746–759 (2009)
11. Han, D., Jing, W., Larson, D., Li, P., Mohapatra, R.N.: Dilation of dual frame pairs in Hilbert
C ∗ modules. Resultate der Mathematik 63, 241–250 (2013)
12. Han, D., Sun, W.: Reconstruction of signals from frame coefficients with erasures at unknown
locations. IEEE Trans. Inform. Theor. 60(7), 4013–4025 (2014)
13. Kaushik, S.K.: A generalization of frames in Banach spaces. J. Contemp. Math. Anal. 44,
212–218 (2009)
14. Lumer, G.: Semi-inner product spaces. Trans. Amer. Math. Soc. 100, 29–43 (1961)
15. Pehlivan, S., Han, D., Mohapatra, R.N.: Linearly connected sequences and spectrally optimal
dual frames for erasures. J. Funct. Anal. 265(11), 2855–2876 (2013)
16. Sahu, N.K., Nahak, C.: Semi-inner product: application to frame theory and numerical range of
operators. Springer Proceedings in Mathematics & Statistics, pp. 77–90. Springer, New York
(2014)
17. Stoeva, D.T.: On p-frames and reconstruction series in separable Banach spaces. Integ. Trans-
forms Spec. Functions 17, 127–133 (2006)
18. Zhang, H., Zhang, J.: Frames, Riesz bases, and sampling expansions in Banach spaces via
semi-inner products. Appl. Comput. Harmon. Anal. 31, 1–25 (2011)
Applications of Generalized
Monotonicity to Variational-Like
Inequalities and Equilibrium Problems
N.K. Mahato and R.N. Mohapatra
Abstract In this paper, we introduce the concept of relaxed (ρ-θ )-η-invariant

monotonicity to establish the existence of solutions for variational-like inequality
problems in reflexive Banach spaces. Again we introduce the concept of (ρ-θ )-
monotonicity for bifunctions. The existence of solution for equilibrium problem
with (ρ-θ )-monotonicity is established by using the KKM technique.
Keywords Variational-like inequality problem · Relaxed (ρ-θ )-η-invariant

monotonicity · Equilibrium problem · (ρ-θ )-monotonicity · KKM mappping
1 Introduction
Let K be a nonempty subset of a real reflexive Banach space X , and X ∗ be the dual
space of X . Consider the operator T : K → X ∗ and the bifunction η : K × K → X .
Then the variational-like inequality problem (in short, VLIP) is to find x ∈ K , such
that
T x, η(y, x) ≥ 0, ∀y ∈ K , (1)
where ., . denote the pairing between X and X ∗ .

If we take η(x, y) = x − y, then (1) becomes to find x ∈ K , such that
T x, y − x) ≥ 0, ∀y ∈ K , (2)
N.K. Mahato
Indian Institute of Information Technology, Design and Manufacturing Jabalpur,
482 005 Jabalpur, India
e-mail: nihar@iiitdmj.ac.in
R.N. Mohapatra (B)
University of Central Florida, Orlando 32816, FL, USA
e-mail: ramm@mail.ucf.edu

DOI 10.1007/978-81-322-2485-3_12
160 N.K. Mahato and R.N. Mohapatra
which are variational inequality problems (VIP) [1, 2]. Variational inequalities have
been studied by many authors [1–5] in both finite- and infinite-dimensional spaces.
When we deal with variational inequalities, the most common assumption for the
operator T is monotonicity. Recently, many authors have established the existence of
solutions for variational inequalities with various types of generalized monotonicity
assumptions (see [3, 5–8] and the references therein). Fang and Huang [5] defined
the concept of relaxed η-α monotonicity and obtained the existence of solutions for
variational-like inequalities. Bai et al. [3] extended the idea of relaxed η-α monotonic-
ity to relaxed η-α pseudomonotonicity. Yang et al. [9] defined several kinds of invari-
ant monotone maps and generalized invariant monotone maps. Behera et al. [10]
defined various concepts of generalized (ρ-θ )-η-invariant monotonicity to general-
ized concepts of Yang et al. [9]. Very recently, Mahato and Nahak [11] introduced
relaxed (ρ-θ )-η-invariant pseudomonotonicity to study variational-like inequalities
and (ρ-θ )-pseudomonotonicity to study equilibrium problems. But in [11], authors
did not consider the concepts such as relaxed (ρ-θ )-η-invariant monotone mappings,
and (ρ-θ )-monotone bifunctions. Therefore, we organized this article to consider
these monotonicity concepts and study the variational-like inequality problems and
equilibrium problem.
Inspired and motivated by [5, 9–11], in this paper, we introduce the concept of
relaxed (ρ-θ )-η-invariant monotone mappings to establish the existence of solu-
tions for variational-like inequality problems. We also introduce the notion of
(ρ-θ )-monotonicity for bifunctions. By using the KKM technique we have studied
the existence of solutions of equilibrium problem with (ρ-θ )-monotone mappings in
reflexive Banach spaces.
2 Preliminaries
We begin with the definition of relaxed (ρ-θ )-η-invariant monotone mappings. For
this consider the function θ : K × K → R and ρ ∈ R.
Definition 1 The operator T : K → X ∗ is said to be relaxed (ρ-θ )-η-invariant
monotone with respect to θ , if for any pair of distinct points x, y ∈ K , we have
T x, η(y, x) + T y, η(x, y) + ρ|θ (x, y)|2 ≤ 0, wher e θ (x, y) = θ (y, x). (3)
Remark 1 (i) If we take ρ = 0 then from (3) it follows that

T x, η(y, x) + T y, η(x, y) ≤ 0, ∀x, y ∈ K , and T is said to be invariant
monotone, see [9].
(ii) If we take ρ = 0, and η(x, y) = x − y, then (3) reduces to T x − T y, x − y ≥
0, ∀x, y ∈ K , and T is said to be monotone map.
From the above definitions, it is clear that invariant monotonicity ⇒ relaxed
(ρ-θ )-η-invariant monotonicity. However, in general a relaxed (ρ-θ )-η-invariant
monotone map may not be an invariant monotone map.
Applications of Generalized Monotonicity to Variational-Like Inequalities . . . 161
Example 1 Let K = [1, 5] and T : [1, 5] → R be defined by T x = x 2 + 1. Let the

functions η and θ be defined by η(x, y) = x 2 + y 2 , θ (x, y) = (x 2 + y 2 )(x 2 + y 2 +5).
Now, T x, η(y, x) + T y, η(x, y) = (x 2 + y 2 )(x 2 + y 2 + 2), which is not less
than 0. Therefore, T is not invariant monotone. But, T is relaxed (ρ-θ )-η-invariant
monotone with respect to θ for any ρ < 1.
Definition 2 [5] The operator T : K → X ∗ is said to be η-hemicontinuous if for

any fixed x, y ∈ K , the mapping f : [0, 1] → R defined by f (t) = T (x + t
(y − x)), η(y, x) is continuous at 0+ .
3 Relaxed (ρ-θ)-η-Invariant Monotonicity and (VLIP)
In this section, we establish the existence of the solution for (VLIP), using relaxed
(ρ-θ )-η-invariant monotonicity. Consider the following problems:
find x ∈ K such that T y, η(x, y) + ρ|θ (x, y)|2 ≤ 0, ∀y ∈ K . (4)
Theorem 1 Let K be a closed convex subset of a reflexive Banach space X . Assume

that T : K → X ∗ is η-hemicontinuous and relaxed (ρ-θ )-η-invariant monotone
with the following conditions:
(i) η(x, y) + η(y, x) = 0, ∀x, y ∈ K ;

|θ (x, xt )|2
(ii) lim = 0, where xt = t y + (1 − t)x, ∀x, y ∈ K ;
t→0 t
(iii) for a fixed z, y ∈ K , the mapping x → T z, η(x, y) is convex.
Then the Problems (1) and (4) are equivalent.
Proof Let x be a solution of (1). From the definition of relaxed (ρ-θ )-η-invariant
monotonicity of T , we get T y, η(x, y) + ρ|θ (x, y)|2 ≤ −T x, η(y, x) ≤ 0.
Conversely, suppose that x ∈ K is a solution of (4), i.e.,
T y, η(x, y) + ρ|θ (x, y)|2 ≤ 0, ∀y ∈ K . (5)
Choose any point y ∈ K and consider xt = t y + (1 − t)x, t ∈ (0, 1], then xt ∈ K .

Therefore, from (5) we have
T xt , η(x, xt ) + ρ|θ (x, xt )|2 ≤ 0;

⇒T xt , η(xt , x) − ρ|θ (x, xt )|2 ≥ 0;
⇒ T xt , η(xt , x) ≥ ρ|θ (x, xt )|2 . (6)
Now, T xt , η(xt , x) ≤ tT xt , η(y, x) + (1−t)T xt , η(x, x) = tT xt , η(y, x).
(7)
From (6) and (7) we have

T xt , η(y, x) ≥ ρ |θ(x,x t )|
2
t .
Since T is η-hemicontinuous and taking t → 0 we get
T x, η(y, x) ≥ 0, ∀y ∈ K .
Definition 3 Let f : K → 2 X be a set-valued mapping. Then f is said to be KKM

n
mapping if for any {y1 , y2 , . . . , yn } of K we have co{y1 , y2 , . . . , yn } ⊂ f (yi ),
i=1
where co{y1 , y2 , . . . , yn } denotes the convex hull of y1 , y2 , . . . , yn .
Lemma 1 ([12]) Let M be a nonempty subset of a Hausdorff topological vector

space X and let f : M → 2 X be a KKM mapping.
If f (y) is closed in X , for all
y ∈ M and compact for some y ∈ M, then f (y) = ∅.
y∈M
Theorem 2 Let K be a nonempty bounded closed convex subset of a real reflexive

Banach space X . Assume that T : K → X ∗ is η-hemicontinuous and relaxed (ρ-θ )-
η-invariant monotone. Let the following hold:
(i) η(x, y) + η(y, x) = 0, ∀x, y ∈ K ;

|θ (x, xt )|2
(ii) lim = 0, where xt = t y + (1 − t)x, ∀x, y ∈ K ; and θ is lower
t→0 t
semicontinuous in the first argument;
(iii) for a fixed z, y ∈ K , the mapping x → T z, η(x, y) is convex and lower
semicontinuous.
Then the Problem (1) has a solution.
Proof Consider the set-valued mapping F : K → 2 X such that

F(y) = {x ∈ K : T x, η(y, x) ≥ 0}, ∀y ∈ K .
It is easy to see that x ∈ K solves the (VLIP) if and only if x ∈ ∩ y∈K F(y). We
claim that F is a KKM mapping. If possible, let F not be a KKM mapping. Then
there exists {x1 , x2 , . . . , xm } ⊂ K such that co{x1 , x2 , . . . , xm } not contained in
m
∪i=1
m
F(xi ), that means there exists a x0 ∈ co{x1 , x2 , . . . , xm }, x0 = ti xi where
i=1

m
ti ≥ 0, i = 1, 2, . . . , m, ti = 1, but x0 ∈
/ ∪i=1
m
F(xi ).
i=1
Hence, T x0 , η(xi , x0 ) < 0; for i = 1, 2, . . . , m. From (i) and (iii) it follows
that

m
0 = T x0 , η(x0 , x0 ) ≤ ti T x0 , η(xi , x0 ) < 0,
i=1
which is a contradiction. Hence F is a KKM mapping.
Assume G : K → 2 X such that G(y) = {x ∈ K : T y, η(x, y) + ρ|θ (x, y)|2 ≤
0}, ∀y ∈ K .
From the relaxed (ρ-θ )-η-invariant monotonicity of T it follows that F(y) ⊂

G(y), ∀y ∈ K . Therefore, G is also a KKM mapping.
Since K is closed bounded and convex, it is weakly compact. From the assump-
tions, we know that G(y) is weakly closed for all y ∈ K . In fact, because
x → T z, η(x, y) and x → ρ|θ (x, y)|2 are lower semicontinuous. Therefore,
G(y) is weakly compact in K , for each y ∈ K.
Therefore, from Lemma 1 and Theorem 1 it follows that F(y) = G(y) = ∅.
y∈K y∈K
So there exists x ∈ K such that T x, η(y, x) ≥ 0, ∀y ∈ K , i.e., the Problem (1)
has a solution.
Theorem 3 Let K be a nonempty unbounded closed convex subset of a real reflexive

Banach space X . Suppose that T : K → X ∗ is η-hemicontinuous and relaxed
(ρ-θ )-η-invariant monotone. Let the following hold:
(i) η(x, y) + η(y, x) = 0, ∀x, y ∈ K ;

|θ (x, xt )|2
(ii) lim = 0, where xt = t y + (1 − t)x, ∀x, y ∈ K ; and θ is lower
t→0 t
(iii) for a fixed z, y ∈ K , the mapping x → T z, η(x, y) is convex and lower
semicontinuous;
(iv) T is weakly η-coercive, i.e., there exits x0 ∈ K such that T x, η(x, x0 ) > 0,
whenever x → ∞ and x ∈ K .
Then the Problem (1) has solution.
Proof Since the proof of this theorem is very similar to Theorem 3 in [11], hence it
is omitted.
4 (ρ-θ)-Monotonicity and Equilibrium Problem
The equilibrium problem (in short, EP) for the bifunction f : K × K → R is to find
x ∈ K , such that
f (x, y) ≥ 0, ∀y ∈ K . (8)
Problems like (8) were initially studied by Fan [13]. Later on Blum and Oettli [4]
discussed that equilibrium problem contains many problems as particular cases for
example, mathematical programming problems, complementary problems, varia-
tional inequality problems, fixed-point problems, and minimax inequality problems.
Inspired and motivated by [11, 14], we introduced the concept of (ρ-θ )-monotonicity
to establish the existence of solution of equilibrium problem over bounded as well
as unbounded domain.
Let K be a nonempty subset of a real reflexive Banach space X . Consider the

function f : K × K → R and θ : K × K → R and ρ ∈ R.
Definition 4 The function f : K × K → R is said to be (ρ-θ )-monotone with

respect to θ : K × K → R if, for all x, y ∈ K , we have
f (x, y) + f (y, x) ≤ ρ|θ (x, y)|2 .
Remark 2 In the above definition,

(i) for ρ > 0 and θ (x, y) = x − y, f is weakly monotone;
(ii) for ρ = 0, f is monotone;
(iii) for ρ < 0 and θ (x, y) = x − y, f is strongly monotone.
We now give an example to show that (ρ-θ )-monotonicity is a generalization of

monotonicity.
Example 2 Let K = [1, 10]. Let the functions f and θ be defined by
f (x, y) = x 2 + y 2 and θ (x, y) = 2(x 2 + y 2 ) + 4.
f (x, y) + f (y, x) = 2(x 2 + y 2 )

≤ ρ(2x 2 + 2y 2 + 4)2 , for any ρ ≥ 1.
Therefore, f is (ρ-θ )-monotone with respect to θ . But f is not monotone.
Theorem 4 Let K be a nonempty convex subset of a real reflexive Banach space X .

Suppose f : K × K → R is (ρ-θ )-monotone with respect to θ and is hemicontinuous
in the first argument with the following conditions:
(i) f (x, x) = 0, ∀x ∈ K ;
(ii) for fixed z ∈ K , the mapping x → f (z, x) is convex;
|θ (x, xt )|2
(iii) lim = 0, where xt = t y + (1 − t)x, ∀x, y ∈ K .
t→0 t
Then x ∈ K is a solution of (8) if and only if
f (y, x) ≤ ρ|θ (x, y)|2 , ∀y ∈ K . (9)
Proof Let x is a solution of (8), i.e., f (x, y) ≥ 0, ∀y ∈ K . Therefore, from the

definition of (ρ-θ )-monotonicity of f it follows that
f (y, x) ≤ ρ|θ (x, y)|2 − f (x, y) ≤ ρ|θ (x, y)|2 , ∀y ∈ K . (10)
Conversely, suppose x ∈ K satisfying (9), i.e.,
f (y, x) ≤ ρ|θ (x, y)|2 , ∀y ∈ K . (11)

Choose any point y ∈ K and xt = t y + (1 − t)x, t ∈ (0, 1], then xt ∈ K . Therefore,

from (11) we have
f (xt , x) ≤ ρ|θ (x, xt )|2 , ∀y ∈ K . (12)
Now conditions (i) and (ii) imply that,

0 = f (xt , xt ) ≤ t f (xt , y) + (1 − t) f (xt , x)
⇒ t[ f (xt , x) − f (xt , y)] ≤ f (xt , x). (13)
From (12) and (13) we have

f (xt , x) − f (xt , y) ≤ ρ |θ(x,x t )|
2
t , ∀y ∈ K .
Since f is hemicontinuous in the first argument and taking t → 0, it implies that
f (x, y) ≥ 0, ∀y ∈ K . Hence x is a solution of (8).
Theorem 5 Let K be a nonempty bounded convex subset of a real reflexive Banach

space X . Suppose f : K × K → R is (ρ-θ )-monotone with respect to θ and is
hemicontinuous in the first argument with the following conditions:
(i) f (x, x) = 0, ∀x ∈ K ;
(ii) for fixed z ∈ K , the mapping x → f (z, x) is convex and lower semicontunuous;
|θ (x, xt )|2
(iii) lim = 0, where xt = t y + (1 − t)x, ∀x, y ∈ K , and θ is upper
t→0 t
semicontinuous in the first argument.
Then the Problem (8) has a solution.
Proof Consider the two set-valued mappings F : K → 2 X and G : K → 2 X such

that
F(y) = {x ∈ K : f (x, y) ≥ 0}, ∀y ∈ K ,
and
G(y) = {x ∈ K : f (y, x) ≤ ρ|θ (x, y)|2 }, ∀y ∈ K .
is easy to see that x ∈ K solves the equilibrium Problem (8) if and only if
It
x ∈ F(y). First to show that F is a KKM mapping. If possible, let F not be a
y∈K
KKM mapping. Then there exists {x1 , x2 , . . . , xm } ⊂ K such that co{x1 , x2 , . . . , xm }

m
is not contained in F(xi ), that means there exists a x0 ∈ co{x1 , x2 , . . . , xm },
i=1

m
m
m
x0 = ti xi where ti ≥ 0, i = 1, 2, . . . , m, ti = 1, but x0 ∈
/ F(xi ).
i=1 i=1 i=1
Hence, f (x0 , xi ) < 0; for i = 1, 2, . . . , m. From (i) and (ii) it follows that

m
0 = f (x0 , x0 ) ≤ ti f (x0 , xi ) < 0,
i=1
which is a contradiction. Hence F is a KKM mapping.
From the (ρ-θ )-monotonicity of f we will show that F(y) ⊂ G(y), ∀y ∈ K .
For any given y ∈ K , let x ∈ F(y), then
f (x, y) ≥ 0.
From the (ρ-θ )-monotonicity of f , it follows that
f (y, x) ≤ ρ|θ (x, y)|2 − f (x, y) ≤ ρ|θ (x, y)|2 .
Therefore x ∈ G(y), i.e., F(y) ⊂ G(y), ∀y ∈ K . This implies that G is also a KKM
mapping.
Since K is closed bounded and convex, it is weakly compact. From the assump-
tions, we know that G(y) is weakly closed for all y ∈ K . In fact, because x → f (z, x)
is lower semicontinuous and x → ρ|(θ (x, z)|2 is upper semicontinuous. Therefore,
G(y) is weakly compact in K , for each y ∈ K.
Therefore from Lemma 1 and Theorem 4 it follows that F(y) = G(y) =
y∈K y∈K
∅.
So there exists x ∈ K such that f (x, y) ≥ 0, ∀y ∈ K , i.e., (8) has a solution.
Theorem 6 Let K be a nonempty unbounded closed convex subset of a real reflexive

Banach space X . Suppose f : K × K → R is (ρ-θ )-monotone with respect to θ and
is hemicontinuous in the first argument and satisfy the following assumptions:
(i) f (x, x) = 0, ∀x ∈ K ;
(ii) for fixed z ∈ K , the mapping x → f (z, x) is convex and lower semicontinuous;
|θ (x, xt )|2
(iii) lim = 0, where xt = t y + (1 − t)x, ∀x, y ∈ K , and is upper
t→0 t
(iv) f is weakly coercive, that is there exists x0 ∈ K such that f (x, x0 ) < 0,
whenever x → +∞ and x ∈ K .
Then (8) has a solution.
Proof Since the proof of this theorem is very similar to Theorem 4.9. in [11], hence
it is omitted.
5 Application to Fixed-Point Problems
Let X = X ∗ be a Hilbert space. Let T : K → K be a given mapping. Then the

fixed-point problem states that find x ∈ K such that
T x = x.
Now, by the setting f (x, y) = x − T x, y − x we can show that if x solves the

equilibrium problem (8) then x is also a solution of the above fixed-point problem.
Indeed, let x is a solution of the equilibrium problem, i.e., f (x, y) ≥ 0, ∀y ∈ K .

Let us choose y = T x, then
f (x, y) = f (x, T x) = −T x − x ≥ 0 ⇒ T x = x,
which shows that x is a fixed point of T .

In this case, notice that f (x, y) is (ρ-θ )-monotone if and only if T is (ρ-θ )-
monotone. Since by Theorems 5 and 6, the equilibrium problem has solution, hence
by the above result the fixed-point problem also has solution.
6 Conclusions
In this study the existence of solutions for variational-like inequality problems under
a new concept relaxed (ρ-θ )-η-invariant monotone maps in reflexive Banach spaces
have been established. We have also obtained the existence of solutions of variational
inequality and equilibrium problems with (ρ-θ )-monotone mappings. This leads
to the natural question of making sensitivity analysis and obtaining results using
ε-efficiency conditions as in [15, 16]. We plan to pursue these as our subsequent
research works.
References
1. Browder, F.E.: Nonlinear monotone operators and convex sets in banach spaces. Bull. Amer.
Math. Soc 71(5), 780–785 (1965)
2. Lions, J.L., Stampacchia, G.: Variational inequalities. Commun. Pure Appl. Math. 20(3), 493–
519 (1967)
3. Bai, M.R., Zhou, S.Z., Ni, G.Y.: Variational-like inequalities with relaxed η-α pseudomonotone
mappings in Banach spaces. Appl. Math. Lett. 19(6), 547–554 (2006)
4. Blum, E., Oettli, W.: From optimization and variational inequalities to equilibrium problems.
Math. Student-India 63(1), 123–145 (1994)
5. Fang, Y.P., Huang, N.J.: Variational-like inequalities with generalized monotone mappings in
Banach spaces. J. Optim. Theory Appl. 118(2), 327–338 (2003)
6. Lee, B.S., Lee, G.M.: Variational inequalities for (η, θ)-pseumonotone operators in nonreflexive
banach spaces. Appl. Math. Lett. 12(5), 13–17 (1999)
7. Luc, D.T.: Existence results for densely pseudomonotone variational inequalities. J. Math.
Anal. Appl. 254(1), 291–308 (2001)
8. Hadjisavvas, N., Schaible, S.: Quasimonotone variational inequalities in Banach spaces. J.
Optim. Theory Appl. 90(1), 95–111 (1996)
9. Yang, X.M., Yang, X.Q., Teo, K.L.: Generalized invexity and generalized invariant monotonic-
ity. J. Optim. Theory Appl. 117(3), 607–625 (2003)
10. Behera, N., Nahak, C., Nanda, S.: Generalized (ρ, θ)-η-invexity and generalized (ρ, θ)-η-
invariant-monotonicity. Nonlinear Anal. Theory Meth. Appl. 68(8), 2495–2506 (2008)
11. Mahato, N.K., Nahak, C.: Variational-like inequalities and equilibrium eroblems with gener-
alized monotonicity in Banach spaces. Adv. Oper. Res. 2012, 15pp. (2012)
12. Fan, K.: Some properties of convex sets related to fixed point theorems. Mathematische Annalen
266(4), 519–537 (1984)
13. Fan, K.: A minimax inequality and applications. Inequalities 3, 103–113 (1972)
14. Mahato, N.K., Nahak, C.: Mixed equilibrium problems with relaxed α-monotone mapping in
banach spaces. Rendiconti del Circolo Matematico di Palermo (2013). doi: 10.1007/s12,215-
013-0103-0
15. Mohapatra, R.N., Verma, R.U.: Sensitivity analysis for cocoercively monotone variational
inclusions and (a, η)-maximal monotonicity. J. Appl. Math. Comput. 26(1–2), 281–293 (2008)
16. Verma, R.U., Mohapatra, R.N.: The ε-efficiency conditions for multiobjective fractional pro-
gramming problems. Dyn. Contin. Discrete Impuls. Syst. Ser. A-Math. Anal. 19(1), 641–660
(2012)
Simultaneous Approximation Properties
of q-Modified Beta Operators
Asha Ram Gairola, Girish Dobhal and Karunesh Kumar Singh
Abstract We establish some approximation properties in simultaneous approxima-

q
tion for a q-analogue of the q-modified Beta operators Bn ( f, x) introduced by Gupta
and Kim. The convergence properties of the q-derivatives of these operators are dis-
cussed. Using the estimates for q-moments, the rate of approximation in simultaneous
approximation is obtained in terms of modulus of continuity.
Keywords q-Modified beta operator · q-integers · Rate of approximation · Modulus

of smoothness · Convergence
1 Introduction
In [21] Phillips introduced the q-Bernstein polynomials for a continuous function on

[0, 1]. The approximation properties of these operators have been studied by several
authors (see [13, 20–26]). In the last decade the q-analogue of the Meyer-König and
Zeller operators, Baskakov operators, Szász operators, etc. have also been introduced
and studied (see [1, 9–11, 17, 21]). In this paper we study the convergence properties
of the q-derivatives of a q-analogue of modified Beta operators introduced by Gupta
and Kim [12].
First, we provide the notations and definitions of the q-calculus used in this paper.
Let N be the set of positive integers.
For n ∈ N, and q ∈ (0, 1) the q-analogue of a non-negative integer n is defined by
[n]q = 1 + q + q 2 + · · · + q n−1 .
A.R. Gairola (B) · G. Dobhal

University of Petroleum and Energy Studies, Dehradun, India
e-mail: ashagairola@gmail.com
G. Dobhal
e-mail: girish.dobhal@gmail.com
K.K. Singh
Government Polytechnic, Rampur, India
e-mail: kks_iitr@gmail.com

DOI 10.1007/978-81-322-2485-3_13
170 A.R. Gairola et al.

Clearly, [n]q = n for q = 1. The q-factorial [n]q ! is given by [n]q ! = n−1
j=0 [n − j]q
n
for n ∈ N and [0]q ! = 1. The q-binomial coefficients k q are defined by

n [n]q !
= ,0 k n
k q [k]q ![n − k]q !
n−1
and the q-rising product (a + b)qn is defined by (a + b)qn = j=0 (a + q j b). The
q-Jackson integrals and q-improper integrals are given by
a ∞

f (x) dq x = (1 − q) a f (aq n ) q n
0 n=0
and
∞/A
∞

qn qn
f (x) dq x = (1 − q) f , A > 0,
n=−∞
A A
0
respectively. Here the sums are assumed to be absolutely convergent (see [14, 18]).
The q-derivative Dq f (t) of a real function f : R → R is defined as
f (t) − f (qt)
Dq f (t) = if t = 0,
(1 − q)t
Dq f (0) = f (0) provided f (0) exists. The product formula for q-differentiation is
given by
Dq ( f (x)g(x)) = f (q x)Dq (g(x)) + g(x)Dq ( f (x)).
We remark that the q-analogue of the integration by part formula is given by
b b
f (x)dq g(x) = f (b)g(b) − f (a)g(a) − g(q x)dq f (x).
a a
For further details we refer to [2, 15, 22].

q
In [12] Gupta and Kim defined a q-analogue of modified Beta operators Bn ( f, x)
as follows:
∞ ∞/A

q [n − 1]q q q
Bn ( f, x) = bn,k (x) q k pn,k (t) f (t) dq t,
[n]q
k=0 0

q q k(k−1)/2 x k q n + k − 1 q k(k−1)/2 x k
bn,k (x) = , pn,k (x) = .
Bq (k + 1, n)(1 + x)qn+k+1 k n+k
q (1 + x)q
Simultaneous Approximation Properties of q-Modified Beta Operators 171
Here f ∈ C B [0, ∞), the class of the continuous and bounded functions on [0, ∞).
The space C B [0, ∞) is a Banach space with respect to the uniform norm f =
q
sup0x<∞ | f (x)|. The operators Bn ( f, x) are linear, positive and reproduce constant
functions. The modulus of continuity ω( f, δ) for f ∈ C B [0, ∞) is defined by
ω( f, δ) = sup sup | f (x + h) − f (x)|.

0<h δ x∈[0,∞)
We have denoted the set of non-negative integers by the symbol N0 . Let λ 0. We

define the class
Cλr [0, ∞) = { f |Dq1 f, Dq2 f, ...Dqr f ∈ C 1 [0, ∞), Dqr f (t) = O(t λ ) as t → ∞},
where C 1 [0, ∞) is the class of continuously differentiable functions on [0, ∞). It

q
was shown in [12] that the operators Bn ( f, x) do not converge to any arbitrary real
function f (x) in case of a fixed q in (0, 1). In this paper we will show that this
property is inherited to the q-derivatives of these operators, i.e. the q-derivatives
q
of the operators Bn ( f, x) do not converge to the corresponding q-derivatives of the
function f for a fixed q. In what follows, we shall use the notations ϕ 2 (x) = x(1+ x)
and throughout this paper C is a constant independent of f and n but may depend
on q. Further, C is not necessarily same at each occurrence.
The paper is organized as follows:
Section 2 contains some estimates of the moments and expressions for the r th
q
q-derivative of theoperatorsBn . Finally, Sect. 3 studies convergence properties of
q
the derivatives Dq Bn ( f, x) and the error estimates for the functions of specified
r
smoothness in terms of the modulus of continuity ω( f, δ).
2 Moment Estimates
Remark 1 For q-differentiation, we have the formula (see [3])
xα x α−1 xα
Dq β
= [α] β+1
− ([β] − [α]) β+1
, (1)
(1 + x)q (1 + x)q (1 + x)q
where Dq denotes the q-derivative operator and α, β are arbitrary real numbers.
q
Using (1) for the weight pn,k (x) we obtain

q n+k−1 k(k−1)/2 [k]x k−1 ([n + k] − [k])x k
Dq [ pn,k (x)] = q −
k q (1 + x)qn+k+1 (1 + x)qn+k+1

n+k−1 [k] (q x)k

= q k(k−1)/2 − ([n + k] − [k]) .
k q x q k (1 + x)(1 + q x)qn+k
Since, we have [n + k] − [k] = q k [n] we get

q n+k−1 k(k−1)/2 [k] (q x)k

Dq [ pn,k (x)] = q − q [n]
k
k q x q k (1 + x)(1 + q x)qn+k

q
[k] pn,k (q x)
= − [n] .
q k ϕ 2 (x)
Thus, we get the identity

q q
q k ϕ 2 (x)Dq [ pn,k (x)] = [k]q − q k [n]q x pn,k (q x). (2)
Similarly, we have

q q
q k ϕ 2 (x)Dq [bn,k (x)] = [k]q − q k [n + 1]q x bn,k (q x). (3)
q
Lemma 1 For the functions Sn,m (x) = Bn ((t − x)qm , x) we have
[n + 1]q x 1
Sn,0 (x) = 1, Sn,1 (x) = +
q [n − 2]q
2 q[n − 2]q
and there holds the recurrence relation

q 2 ϕ 2 (x) Dq Sn,m (x) + [m]q Sn,m−1 (x) + q m+2 [m]q x(1 + [2]q xq m+2 )Sn,m−1 (q x)

= [n]q q m+2 − [n + 1]q q 2 x − [m + 1]q q(1 + [2]q xq m+2 ) Sn,m (q x)
+([n]q q − [m + 2]q )Sn,m+1 (q x).
Proof Sn,0 (x) and Sn,1 (x) follow from Lemma 3, [12]. Now, using product formula
for q-differentiation and Remark 1, we get
q
Dq (t − x)qm bn,k (x)
q q
= −[m]q (t − x)qm−1 bn,k (x) + (t − q x)qm Dq bn,k (x)
q (t − q x)qm ([k]q − q k [n + 1]q x) q

= −[m]q (t − x)qm−1 bn,k (x) + bn,k (q x)
q k ϕ 2 (x)
Hence,

ϕ 2 (x) Dq Sn,m (x) + [m]q Sn,m−1 (x)
∞ ∞/A

[n − 1]q ([k]q − q k [n]q t/q) q q
= bn,k (q x) q k pn,k (t)(t − q x)qm dq t
[n]q qk
k=0 0
∞ ∞/A

[n − 1]q (q k [n]q t/q − q k [n + 1]q x) q q
+ b (q x) q k pn,k (t)(t − q x)qm dq t
[n]q qk n,k
k=0 0
= J1 + J2 , say.
Using (2), with the transformation t = qu, we get
∞ ∞/A

[n − 1]q q q
J1 = bn,k (q x) ϕ 2 (u)Dq ( pn,k (u))(qu − q x)qm q du
[n]q
k=0 0
We split ϕ 2 (u) as
1 1
ϕ 2 (u) = q m x 1 + [2]q xq m+2 + 1 + [2]q xq m+2 qu − q m+1 x + 2 qu − q m+1 x qu − q m+2 x
q q
and use in J1 . This yields
∞ ∞/A

[n − 1]q q
J1 = bn,k (q x) q m x 1 + [2]q xq m+2 (qu − q x)qm
[n]q
k=0 0
∞/A

1
+ 1 + [2]q xq m+2 (qu − q x)qm+1
q
0
∞/A

1 q
+ 2
(qu − q x)q
m+2
Dq ( pn,k (u))q du
q
0
= K 1 + K 2 + K 3 , say.
Now, integration by parts gives
∞

m+2 [n − 1]q q
K 1 = −[m]q m+1
x 1 + [2]q xq bn,k (q x)
[n]q
k=0
∞/A

q
× pn,k (qu)(qu − q x)qm−1 dq u
0
[n − 1] ∞
q q
= −[m]q m x 1 + [2]q xq m+2 bn,k (q x)
[n]q
k=0
∞/A

q
× pn,k (t)(t − q x)qm−1 dq u
0

= −[m]q m x 1 + [2]q xq m+2 Sn,m−1 (q x).
Similarly, we obtain
[m + 1]q
K2 = − 1 + [2]q xq m+2 Sn,m (q x)
q
and
[m + 2]q
K3 = − Sn,m+1 (q x).
q2
By simple calculations we obtain
[n]q
J2 = ([n]q q m − [n + 1])x)Sn,m (q x) + Sn,m+1 (q x).
q
Combining J1 and J2 the required relation follows.
Corollary 1 For the functions Sn,m (x) we have (i) Sn,m (x) are polynomials
in x
of degree exactly m; (ii) there holds the order Sn,m (x) = O [n]q−m+1 , for all x ∈
[0, ∞).
Lemma 2 [8] Let m ∈ N0 , 0 < q < 1. There exists a constant C = C(q, m) > 0
independent of x and n such that

1
q
Bn (t − x)qm , x C (m+1)/2
.
[n]q
Following is a Lorentz-type lemma (see [4, p. 112]) for the q-differentiation of

q
bn,k (x).
q
Lemma 3 For the functions bn,k (x) there holds
r j
q r −1 x r 1 + q n+k+1 x Dqr bn,k (x) =
q q
α j (x)[n + 1]iq [k]q − q k [n + 1]q x bn,k (x),
q
2i+ j r
i, j 0
where αi (x) are polynomials in x independent of [n]q .
Proof The proof follows by an induction on r.
Lemma 4 For the functions Q r,l (x) defined by

∞
r
q
Q r,l (x) = [k]q − q k+l [n + 1]q x bn,k (x),
k=0

there holds the order Q 2i,l (x) = O [n]q2i .
∞ q k(k − 1)/2 t k [n + k]q !

Proof The identity = 1 will be frequently used. We
k=0 [k]q ![n − 1]q !(1 + t)n+k+1
q
2
shall apply an induction on i. For i = 1 we write [k]q − q k+l [n + 1]q x = [k]q2 −
2[k]q [n + 1]q q k+l x + q 2k+2l [n + 1]q2 x 2 and obtain three terms, namely D1 , D2 and
D3 . Now,
∞
q
D1 = [k]q2 bn,k (x)
k=0
∞
(1 + q[k − 1]q )q k(k−1)/2 x k [n + k]q !
=
k=0 [k − 1]q ![n − 1]q !(1 + x)qn+k+1
∞ ∞
q k(k−1)/2 x k [n + k]q ! q k(k−1)/2 x k [n + k]q !
= n+k+1
+ q n+k+1
k=0 [k − 1]q ![n − 1]q !(1 + x)q k=0 [k − 2]q ![n − 1]q !(1 + x)q
k
∞ q k(k−1)/2 q 2 x [n + k + 2]q !
[n]q x 2 2
= +x q
1+x 2 n+k+1
k=0 [k]q ![n − 1]q !(1 + x)(1 + q x) 1 + q x q
k
∞ k(k−1)/2 q 2 x [n + k + 2]q !
[n]q x [n + 1]q [n]q x 2 q 2 q n+k+2 x)2 .
= + n+k+3 (1 + q q
1+x (1 + x)(1 + q x) [k]q ![n − 1]q ! 1 + q x2
k=0 q
This gives
∞
[n]q x [n + 1]q [n]q x 2 q 2 (1 + x)2 q k(k−1)/2 (q 2 x)k [n + k + 2] !
q
|D1 | = +
1+x (1 + x)(1 + q x) [k]q ![n − 1]q !(1 + q 2 x)qn+k+3
k=0
[n]q x [n + 1]q [n]q x 2 q 2 (1 + x)
+ .
1+x (1 + q x)
Next, we have
∞
q k q k(k−1)/2 x k [n + k]q !
D2 = −2xq l [n + 1]q
k=0
[k − 1]q ![n − 1]q !(1 + x)qn+k+1
∞
−2xq l [n + 1]q [n]q q 2k q k(k−1)/2 x k+1 [n + k + 1]q !
= (1 + q n+k+1 x)q3
(1 + x)(1 + q x) [k]q ![n]q !(1 + q 2 x)qn+k+2
k=0
2q l+1 x[n + 1]q [n]q (1 + x)2

Therefore, |D2 | (1 + q x) . Similarly, we have
∞
(q l [n + 1]q x)2 q 2k q k(k−1)/2 x k [n + k]q !
n+k+1 2
|D3 | = (1 + q x)q
(1 + x)(1 + q x) [k]q ![n − 1]q !(1 + q 2 x)qn+k+1
k=0
(q l [n + 1]q x) (1 + x)
2
.
(1 + q x)
Combining the estimates for D1 –D3 , we obtain

Q 2,l (x) [n]q x + [n + 1]q [n]q x q (1 + x) + 2q x[n + 1]q [n]q (1 + x)
2 2 l+1 2
1+x (1 + q x) (1 + q x)
(q [n + 1]q x) (1 + x)
l 2
+ = O [n]q2 , for all x ∈ (0, ∞).
(1 + q x)
Let the lemma be true for a certain i. By q-differentiation we get

∞
2i−1
q
Dq Q 2i,l (x) = −[2i]q [n]q q l q k [k]q − q k+l+1 [n + 1]q x bn,k (q x)
k=0
∞

2i
[k]q − q k+l+1 [n + 1]q x q
+ b (q x) [k]q − q k
[n + 1]q x
q k ϕ ( x) n,k
k=0
Rearrangement of the terms gives
∞

2i+1
[k]q − q k+l+1 [n + 1]q x q
bn,k (q x)
qk
k=0
= ϕ 2 (x)Dq Q 2i,l (x)
∞
2i−1
q
+ ϕ 2 (x)[2i]q [n]q q l q k [k]q − q k+l+1 [n + 1]q x bn,k (q x)
k=0
∞
2i
q
− [n + 1]q x q l − 1 [k]q − q k+l+1 [n + 1]q x bn,k (q x). (4)
k=0
Therefore, from the definition of Q 2i+1,l (x) and (4) we get

∞ 2i+1
[k]q − q k+l+1 [n + 1]q x

Q 2i+1,l (x) b
q
(q x)
qk n,k
k=0
C[n]q2i + C [n]q2i+1 = C[n]q2i+1 .
Lemma 5 Let f ∈ Cλr [0, ∞). Then

r
∞ ∞/A

[n + m − 1]q q
q q
Dqr Bn ( f, x) = bn+r,k (x) q k pn−r,k+r (q r t)Dqr f (t) dq t.
[n − m]q
m=0 k=0 0
Proof First, we prove the lemma for r = 1. By product rule for q-differentiation we
obtain
xl [l]q x l−1 [l + m]q (q x)l
Dq l+m
= l+m
− .
(1 + x)q (1 + x)q (1 + x)l+m+1
q
This gives
q q q
Dq bn,k (x) = [n]q q k−1 bn+1,k−1 (x) − q bn+1,k (x)
q q q
Dq pn,k (t) = [n]q q k−1 pn+1,k−1 (t) − q pn+1,k (t) .
Next q-integration by parts gives

q
Dq Bn ( f, x)
∞

q q
= [n − 1]q q k−1 bn+1,k−1 (x) − bn+1,k (x)
k=0
∞/A

q
× q k pn,k (t) f (t) dq t
0
∞/A

∞

q q q
= [n − 1]q q k bn+1,k (x) q k qpn,k+1 (t) − pn,k (t) f (t) dq t
k=0 0
∞/A

∞

q q
= q k bn+1,k (x) −Dq pn−1,k+1 (t) f (t) dq t
k=0 0
∞ ∞/A

q q
= q k bn+1,k (x) pn−1,k+1 (t)Dq ( f (t)) dq t.
k=0 0
Hence, the result holds for r = 1. Now the lemma follows by induction on r and
straightforward calculations.
For f ∈ Cλr [0, ∞) we define the operators

Bn,q,r ( f, x) by
∞ ∞/A

Bn,q,r ( f, x) =
q
bn+r,k (x)
q
q k pn−r,k+r (q r t) f (t) dq t.
k=0 0
Lemma 6 For the functions Un,m,r (x) =

Bn,q,r (t m , x), we have
1
Un,0,r (x) = ,
q 2r [n − r − 1]q
q −1 [n + r + 1]q x + [2]q q r −2 + [r ]q
Un,1,r (x) =
[n − r + 1]q q 2r [n − r + 1]q − [2]q q −2
and there holds the relation

q 2r [n − r + 1]q − [m + 2]q q −m−2 Un,m+1,r (q x) (5)
= ϕ 2 (x)Dq Un,m,r (x) + [n + r + 1]q xUn,m,r (q x)

+ [m + 1]q q r −m−1 + [r ]q Un,m,r (q x).
Proof We have
∞ ∞/A

q q
Un,m,r (x) = bn+r,k (x) q k pn−r,k+r (q r t)t m dq t.
k=0 0
Using (2) we get
ϕ 2 (x)Dq Un,m,r (x) + [n + r + 1]q xUn,m,r (q x)

∞ ∞/A

−k q q
= q [k]q bn+r,k (q x) q k pn−r,k+r (q r t)t m dq t (6)
k=0 0

We write q −k [k]q = q r q −k−r [k + r ]q − [n − r + 1]q tq r +[n −r +1]q tq r −[r ]q
and substitute in (6). This gives three terms I1 , I2 and I3 . Using the transformation
t → u/q r we obtain
∞/A

∞
1
q −k−r [k + r ]q − [n − r + 1]q u pn−r,k+r (u) dq u
q q
I1 = q k bn+r,k (q x)
q mr
k=0 0
∞/A

∞
1
q q
= q k bn+r,k (q x) u m+1 + u m+2 pn−r,k+r (u) dq u
q mr
k=0 0
= J1 + J2 , say.
Now, integration by parts and the inverse transformation u → tq r give

∞ ∞/A

−[m + 1]q k q q um
J1 = mr
q bn+r,k (q x) pn−r,k+r (u) m+1 dq u
q q
k=0 0
r −m−1
= −[m + 1]q q Un,m,r (q x).
Similarly, J2 = −[m + 2]q q 2r −m−1 Un,m+1,r (q x). From definition we get I2 =

q 2r [n−r +1]q Un,m+1,r (q x) and I3 = −[r ]q Un,m,r (q x). Combining these estimates,
we obtain the required recurrence relation.

Corollary 2 Since Bn,q,r (t − x)2 , x) is a quadratic polynomial, we write

Bn,q,r (t − x)2 , x) = α0 + α1 x + α2 x 2 . Using the recurrence relation (5) the
coefficients αi are given as
2
[2]q q r−2 + [r ]q
α0 = ,
q 4r [n − r + 1]q − [2]q q −2 [n − r + 1]q − [3]q q −3
−1
q + 2 [2]q q r−2 + [r ]q [n + r + 1]q − 2q 2r [2]q q r−2 + [r ]q [n − r + 1]q − [3]q q −3
α1 =
q [n − r + 1]q − [2]q q
4r −2 [n − r + 1]q − [3]q q −3
and

[n + r + 1]q q −2 + [n + r + 1]q − 2q 2r−1 [n − r + 1]q − [3]q q −3
α2 =
q 4r [n − r + 1]q − [2]q q −2 [n − r + 1]q − [3]q q −3

q [n − r + 1]q − [2]q q −2 [n − r + 1]q − [3]q q −3
2r
+ 4r −2

−3
.
q [n − r + 1]q − [2]q q [n − r + 1]q − [3]q q
Now, it follows from straightforward calculations that |α0 | C

[n − r + 1]q2
, |α1 |
−1

= C(r, a) and |α | (1+q −2q )
2r 2r
C
[n − r + 1]q where C, C 2 q 4r
+ O [n − r + 1]q . Con-
1
−1

Bn,q,r (t − x)2 , x C (1+q q−2q ) 2
2r 2r
sequently, 4r x + O [n−r1+1]q , for all x ∈
[0, ∞).
3 Simultaneous Approximation Using q-Moments
Theorem 1 If f ∈ Cλr [0, ∞) and qn be a sequence in (0, 1)

with qn ↑ 1, then there
q
exists a number 0 < qˆn < 1 such that the sequence Dqr n Bn n ( f, x) converges to

Dqr n f (x) pointwise for qn ∈ (qˆn , 1).
Proof There exists qˆn ∈ (0, 1) (see [7]) such that for all qn ∈ (qˆn , 1) we have

r
(Dql n f )(x) (Dqr n+1 f )(ξ )
f (t) = (t − x)lq + (t − x)rq+1 .
[k]qn ! [r + 1]qn !
l=0

q
Since, Bn n (t − x)qmn , x are polynomials of degree exactly m, (see [8]), we obtain
q
Dqr n Bn n ( f, x) − Dqr n f (x)
1
(Bn n )(r ) (Dqr n+1 f )(ξ )(t − x)rq+1
q
= , x
[r + 1]qn ! n
[n − 1]qn
∞ j
qn
= α j (x)[n + 1]iqn [k]qn − qnk [n + 1]q x bn,k (x)
[n]qn
k=0
2i+ jr
i, j0
∞/A

(Dqr n+1 f )(ξ )
(t − x)rq+1
q
× qnk pn,k
n
(t) dqn t.
[r + 1]qn ! n
0
Let Ti, j be a typical term of the sum over i, j. Using Hölder’s inequality first for
integration and then for summation we obtain
∞
j
[n − 1]qn qn
|Ti, j | CDqr n+1 f [n + 1]iqn [k]qn − qnk [n + 1]q x bn,k (x)
[n]qn 2i+ jr k=0
i, j0

∞/A

k qn r +1
× qn pn,k (t)(t − x)q dqn t

0
[n − 1]qn j
∞
qn
CDqr n+1 f [n + 1]iqn [k]qn − qnk [n + 1]q x bn,k (x)
[n]qn
2i+ jr k=0
i, j0
⎛ ⎞ 1
∞/A
r +1

r
⎜ qn j r +1 ⎟
× k
qn pn,k (t)|t − qn x|
⎝ dqn t ⎠
j=0 0
⎛ ⎞1/2
∞

CDqr n+1 f [n + 1]iqn ⎝ (x)⎠
2 j q
[k]qn − qnk [n + 1]q x bn,k n
2i+ jr k=0

i, j0
⎛ ⎛ ⎞2/r +1 ⎞1/2
∞ ∞/A

⎜ [n − 1]qn r
⎜ ⎟ ⎟
×⎜ (t)|t − qn x|r +1 dqn t ⎠ ⎟
qn qn j
⎝ [n]q bn,k (x) ⎝ qnk pn,k ⎠
n k=0 j=0 0
⎛ ⎞1/2
∞

CDqr n+1 f [n + 1]iqn ⎝ (x)⎠
q
([k]qn − qnk+s [n + 1]q x)2 j bn,k
n
2i+ jr k=0

i, j0
⎛⎛ ⎞ 1 ⎞1/2
∞ ∞/A
r +1
⎜
r
⎜ ⎜ [n − 1]qn j 2r +2 ⎟ ⎟
dqn t ⎠ ⎟
qn k qn
×⎝ ⎝ bn,k (x) qn pn,k (t) t − qn x ⎠ .
[n]qn
j=0 k=0 0
Now, we use the relation (see [19])

2r +2
2r +2−s
2r +2 1 − qnn
(t − x) = α2r +2,s x 2r +2−s (t − x)qs ,
[n]qn
s=1
where the constants α2r +2,s are independent of x, qn and n. Using Corollary 1 in
above relation we get
2r
+2 n
2r +2−s

qn
Bn ((t − x)2r +2 , x) α2r +2,s 1 − qn q
x 2r +2−s Bn n (t − x)qs , x
[n]qn
s=1

2r +2 n
2r +2−s
(s+1)

α2r +2,s 1 − qn 1 2
x 2r +2−s
[n]qn [n]qn
s=1

r +1
1
= O .
[n]qn
2r +2 2r +2 2r +2−l

Therefore, using (t − qn x)2r +2 =
j l x(1 − q j )
l=0 l (t − x) n and
Hölder’s inequality we obtain
+2

2r 2r + 2
qn j 2r +2−l 1
Bn ((t − qn x)2r +2 , x)
j
x(1 − qn ) l/2
.
l
l=0 [n]qn

j
Let (1 − qn ) = O [n]1ρ , ρ 0, for all j. This implies (1 − qnr ) = O [n]1ρ .
qn qn
Consequently, we get
⎛ ⎞1/2
2r+2 r +1
2r + 2

1

⎜
r
i+ j j 2r+2−l 1 ⎟
|Ti, j | CDqr+1
n
f [n + 1]qn ⎝ x(1 − qn ) l/2 ⎠
2i+ j r j=0 l=0
l [n]q n
i, j 0
⎛ 1 ⎞1/2
r +1

r (2r+2)(ρ−1/2)
[n]qn
[n + 1]qn ⎝ ⎠
i+ j
CDqr+1
n
f (2r+2)ρ
2i+ j r j=0 [n]qn
i, j 0
1
CDqr+1
n
f 1/2
.
[n]qn
q
Since, this inequality is independent of ρ, it follows that Dqr n Bn n ( f, x) converges

pointwise to Dqr n f (x) as n → ∞ for qn ∈ (qˆn , 1).
Theorem 2 Let f ∈ Cλr [0, ∞) and q ∈ (0, 1). Then there exists C > 0 independent
of f and n such that
r

[n − m]q q
2r
q [n − r − 1]q Dq Bn ( f, x) − Dq ( f (x))
r r
[n + m − 1]q
m=0

Cω Dqr f, [n − r − 1]q (1 + q 2r − 2q 2r −1 )x ,
for all x ∈ [0, ∞).
Proof We have
∞ ∞/A

[n − r − 1]q q q
bn+r,k (x) q k pn−r,k+r (q r t) dq t = 1.
q −2r
k=0 0
This gives
r

[n − m]q
2r q
q [n − r − 1]q Dqr Bn ( f, x) − Dqr ( f (x))
[n + m − 1]q
m=0
∞/A

∞

q q
q 2r [n − r − 1]q bn+r,k (x) q k pn−r,k+r (q r t) Dqr f (t) − Dqr f (x) dq t
k=0 0
∞ ∞/A

|t − x|
q q
q 2r [n − r − 1]q ω Dqr f, δ bn+r,k (x) q k pn−r,k+r (q r t) 1 + dq t
δ
k=0 0
= T1 + T2 , say.

We have T1 = ω Dqr f, δ . Using Schwarz’s inequality and Corollary 2, we obtain
⎛ ⎞1/2
∞ ∞/A

q 2r [n − r − 1]q ⎜ q q ⎟
T2 ⎝ bn+r,k (x) q k pn−r,k+r (q r t)(t − x)2 dq t ⎠
δ
k=0 0

q 2r [n − r − 1]q (1 + q 2r − 2q 2r −1 ) 2 1
C x +O .
δ q 4r [n − r + 1]q

Now, Choosing δ = [n − r − 1]q (1 + q 2r − 2q 2r −1 )x we get
r

[n − m]q q
2r
q [n − r − 1]q Dq Bn ( f, x) − Dq ( f (x))
r r
[n + m − 1]q
m=0

Cω Dqr f, [n − r − 1]q (1 + q 2r − 2q 2r −1 )x .
Remark 2 Let (qn ) be a sequence in (0, 1) such that qn ↑ 1 as n → ∞ with the rate
γ γ
1 − qn = O(1/[n]q ), γ > 2. Then it follows that (1 + q 2r − 2q 2r −1 ) = O(1/[n]q ).
Consequently, we obtain
r

[n − m]q
2r q
q [n − r − 1]q Dqr Bn ( f, x) − Dqr ( f (x))
[n + m − 1]q
m=0

x
Cω Dqr f, (γ −2)/2 .
[n]q
The right-hand side tends to 0 on every finite compact subinterval of [0, ∞).
q
Remark 3 From Theorem 1 it follows that Dqr Bn ( f ) do not converge to Dqr f
unless we take a sequence (qn ) in (0, 1) such that qn ↑ 1. Similarly in Theorem 2,
we observe that for a fixed q the right-hand side do not tend to 0 as n → ∞. Therefore,
q
the operators Bn ( f, x) fail to possess simultaneous approximation properties.
Acknowledgments The authors are indebted to anonymous referees for many valuable comments
especially in connection with Lemma 1 which helped to improve the paper.
References
1. Aral, A., Gupta, V.: The q-derivative and applications to q-Szász Mirakyan operators. CAL-
COLO 43, 151–170 (2006)
2. Aral, A., Gupta, V., Agarwal, R.P.: Applications of q-Calculus in Operator Theory. Springer,
New York (2013)
3. De Sole, A., Kac, V.G.: On integral representations of q-gamma and q-beta functions. Rend.
Mat. Acc. Lincei, s. 9. 16, 11–29 (2005)
4. DeVore, R.A, Lorentz, G.G.: Constructive Approximation, Springer-Verlag, Berlin, New York
(1993)
5. Doğru, O., Duman, O.: Statistical approximation of Meyer-König and Zeller operators based
on the q-integers. Publ. Math. Debrecen 68, 199–214 (2006)
6. Durrmeyer, J.L.: Une formula d’inversion, de la transformee de Laplace: Application a la
theorie des Moments, These de 3e Cycle, Faculte des Sciences de l’universite de Paris (1967)
7. Ernst, T.: A new notation for q-calculus and a new q-Taylor formula, Uppsala University
Report. Depart. Math. 1–28 (1999)
8. Gairola, A.R., Agrawal, P.N., Dobhal, G., Singh, K.K.: Moments of a q-Baskakov-beta oper-
ators in case 0 < q < 1. J. Classical Anal. 2(1), 9–22 (2013)
9. Gupta, V.: Some approximation properties of q-Durrmeyer operators. Appl. Math. Comput.
197(1), 172–178 (2008)
10. Gupta, V.: On certain q-Durrmeyer type operators. Appl. Math. Comput. 209, 415–420 (2009)
11. Gupta, V., Heping, W.: The rate of convergence of q-Durrmeyer operators for 0 < q < 1. Math.
Meth. Appl. Sci. 31, 1946–1955 (2008)
12. Gupta, V., Kim, T.: On the rate of approximation by q-modified Beta operators. J. Math. Anal.
Appl. 377, 471–480 (2011)
13. II’inskii, A., Ostrovska, S.: Convergence of generalized Bernstein polynomials. J. Approx.
Theor. 116(1), 100–112 (2002)
14. Jackson, F.H.: On a q-definite integrals. Quarterly J. Pure Appl. Math. 41, 193–203 (1910)
15. Kac, V.G., Cheung, P.: Quantum Calculus. Universitext, Springer-Verlag, New York (2002)
16. Kantorovich, L.V.: Sur certain developments suivant les polynomes de la forms de S. Bernstein,
Dokl. Akad. Nauk. SSSR. 1–2, 563–568, 595–600 (1930)
17. Karsli, H., Gupta, V.: Some approximation properties of q-Chlodowsky operators. Appl. Math.
Comput. 195, 220–229 (2008)
18. Koornwinder, T.H.: q-Special Functions, a tutorial. In: Gerstenhaber, M., Stasheff, J. (eds.)
Deformation and Qauntum Groups with Applications of Mathematical Physics, Contemp.
Math., 134 (1992), Amer. Math. Soc. (1992)
19. Mahmudov, N.: The moments for q-Bernstein operators in the case 0 < q < 1. Numer. Algor.
43–450 (2010)
20. Ostrovska, S.: q-Bernstein polynomials and their iterates. J. Approx. Theor. 123(2), 232–255
(2003)
21. Phillips, G.M.: Bernstein polynomials based on the q-integers. Ann. Numer. Math. 4, 511–518
(1997)
22. Phillips, G.M.: Interpolation and Approximation by Polynomials, CMS Books in Mathematics,
vol. 14. Springer, Berlin (2003)
23. Wang, H., Meng, F.: The rate of convergence of q-Bernstein polynomials for 0 < q < 1. J.
Approx. Theor. 136(2), 151–158 (2005)
24. Wang, H., Wu, X.: Saturation of convergence of q-Bernstein polynomials in the case q > 1. J.
Math. Anal. Appl. 337(1), 744–750 (2008)
25. Wang, H.: Voronovskaya-type formulas and saturation of convergence for q-Bernstein poly-
nomials for 0 < q < 1. J. Approx. Theor. 145, 182–195 (2007)
26. Wang, H.: Properties of convergence for q-Bernstein polynomials. J. Math. Anal. Appl. 340(2),
1096–1108 (2008)
Nonlinear Mixed Variational-Like Inequality
with Respect to Weakly Relaxed η − α
Monotone Mapping in Banach Spaces
Gayatri Pany and Sabyasachi Pani
Abstract In this paper, we have studied Nonlinear Mixed Variational-like inequal-

ity with respect to weakly relaxed η − α monotone mapping, involving a nonlinear
bifunction, in Banach space. Significance of weakly relaxed η − α monotonicity
is illustrated through an example. Existence of the solution to the problem is estab-
lished using KKM (Knaster, Kuratowski and Mazurkiewicz) technique. Also we have
proposed an iterative algorithm using auxiliary principle technique, which involves
formulation of an auxiliary minimizing problem and then characterizing it by an aux-
iliary variational inequality problem. Solvability of the auxiliary variational inequal-
ity problem is established. Finally convergence of the iterates to the exact solution
is proved.
Keywords Weakly relaxed η − α monotone mapping · KKM technique · Auxiliary

principle technique · Iterative algorithm
1 Introduction
Study of variational inequality problem (in short VIP) mainly involves twofold aspect,
namely qualitative and numerical. Qualitative aspect includes study of existence and
uniqueness of the solution of the corresponding problem, while constructing iterative
algorithm to find approximate solutions to the actual solution, study of convergence
criteria, obtaining error bounds come under the numerical aspect. These two aspects
of VIP provide an elegant framework of study in various fields like optimization, eco-
nomics, transportation, oceanography, fluid flow through porous media, pure, applied
and engineering sciences. These applications lead to generalizations of variational
inequality theory in various directions. Variational-like inequalities problems (VLIP),
G. Pany (B) · S. Pani

IIT Bhubaneswar, Odisha 751007, India
e-mail: gp11@iitbbs.ac.in
S. Pani
e-mail: spani@iitbbs.ac.in

DOI 10.1007/978-81-322-2485-3_14
186 G. Pany and S. Pani
generalized mixed variational-like inequalities are some of the importantgeneraliza-

tion of variational inequality. Generalized mixed variational-like inequalities find
applications in the areas like optimization theory (see [10, 12]), structural analysis
[11], and economics [7].
The concept of monotonicity plays an important role for proving existence of
solutions of VIP and VLIP. Chen [2] introduced semimonotonicity, using the combi-
nation of compactness and monotonicity and studied the corresponding variational
inequality in Banach space. Huang and Deng [8] studied set-valued strongly nonlin-
ear mixed variational-like inequality under strongly η − α monotonicity in Hilbert
space. They have provided an iterative algorithm using auxiliary principle technique.
There are substantial number of results on existence and uniqueness of variational
inequalities under Hilbert space setting. These concepts were also generalized into
Banach space setting. Fang and Huang [6] studied variational-like inequalities with
respect to relaxed η −α monotone mapping in reflexive Banach space by introducing
a new concept of relaxed η −α monotonicity. Later, this work was extended by Bai et
al. [1], introducing the concept of relaxed η−α pseudomonotonicity. Recently, Kutbi
and Sintunavarat [9] proved some existence results for variational-like inequalities
involving a single operator using weakly relaxed η − α monotone mapping. But
there was no discussion on the numerical aspect. Motivated by these works, we have
extended these concepts to nonlinear mixed variational-like inequality with respect
to weakly relaxed η − α monotone mapping, given by
Find w ∈ K , N (w, y), η(v, w) + b(w, v) − b(w, w) ≥ 0, ∀y, v ∈ K ,
where N : E × E → E ∗ is η-hemicontinuous and b : E × E → R is a convex

lower semicontinuous function in second variable. Carrying out our study on the
numerical aspect we have proposed an iterative algorithm to approximate the exact
solution using auxiliary principle technique which is due to Glowinski et al. [11].
Following the ideas of Ding [4], we have discussed the covergence criteria for the
proposed iterative algorithm.
2 Preliminaries
Definition 1 If T : K → E ∗ and η : K × K → E ∗ , then T is η−hemicontinuous

if f (t) = T (x + t (y − x)), η(y, x) is continuous at 0, where f : [0, 1] →
(−∞, +∞).
Definition 2 N and η are said to have 0−diagonally concave relation, if the function
φ : K × K → R defined by
φ(w, v) = N (w, v), η(w, v)

Nonlinear Mixed Variational-Like Inequality … 187
is 0−diagonally concave in v, i.e., for any finite set {v1 , · · · , vm } ⊂ K and for any

m
convex combination of vi , λi φ(w, vi ) ≤ 0. N and η are said to have 0−diagonally
i=1
convex relation on K if −N and η have 0−diagonally concave relation.
Definition 3 N is η−monotone with respect to first argument if
N (w1 , v) − N (w2 , v), η(w, v) ≥ 0, ∀w, v ∈ K .
Definition 4 N is η−antimonotone with respect to second argument if
N (w, v1 ) − N (w, v2 ), η(w, v) ≤ 0, ∀w, v ∈ K .
Definition 5 F : K → E ∗ is KKM mapping if, for any {x1 , · · · , xn } ⊂ K ,

n
co{x1 , · · · , xn } ⊂ i=1 F(xi ).
n
n
Definition 6 If S = {x1 , · · · , xn }, co{x1 , · · · , xn } = αi xi : αi =
i=1 i=1
1, ∀αi ≥ 0 .
Definition 7 f : K → (−∞, +∞] is lower semicontinuous at x0 if f (xo ) ≤

lim infx→x0 f (x).
Definition 8 T : K → E ∗ is Lipschitz continuous if ∃ α > 0, such that T x −

T y ≤ α x − y , ∀ x, y ∈ K .
Definition 9 T is η−coercive with respect to a proper function f : K →

(−∞, +∞] if there exists x0 ∈ K such that
T x − T x0 , η(x, x0 ) + f (x) − f (x0 )

→ ∞.
η(x, x0 )
Definition 10 If T : K → E ∗ , η : K × K → E, α : E → R, t > 0, z ∈ E and

p > 1 is a constant, then T is
• weakly relaxed η − α monotone if T x − T y, η(y, x) ≥ α(x − y), ∀x, y ∈ K ,
d
where lim α(t z) = 0, lim α(t z) = 0.
t→0 t→0 dt
• relaxed η − α monotone if α(t z) = t p α(z) and T x − T y, η(y, x) ≥ α(x −
y), ∀x, y ∈ K .
• strongly η − α monotone if T x − T y, η(y, x) ≥ c x − y 2 , ∀x, y ∈ K . Here
α(t z) = c z 2 ; c > 0 is a constant,
Remark 1 From the definitions it is obvious that

• Strongly η − α monotonicity is a special case of relaxed η − α monotonicity.
• Relaxed η − α monotonicity implies weakly relaxed η − α monotonicity.
But weakly relaxed η − α monotonicity does not always imply relaxed η − α

monotonicity.

π
Example 1 If T : 0, → R, T x = sin x, η(x, y) = x − y, α(z) = − sin2 z,
2
d
then lim α(t z) = 0 and lim α(t z) = 0.
t→0 t→0 dt
T x − T y, η(y, x) = sin x − sin y, x − y ≥ − sin2 (x − y) = α(x − y).
So T is weakly relaxed η − α monotone, but not relaxed η − α monotone as α(t z) =

t p α(z).
The following Lemma is used to prove the existence result for our problem.
Lemma 1 ([5]) If M is a nonempty subset of a Hausdorff topological vector space
a KKM mapping, F(x) is closed in X, ∀x ∈ K and compact for
X, F : M → 2 X is
some x ∈ K , then x∈M F(x) = φ.
Another useful result that we have used to prove the existence of solution for the
auxiliary variational inequality is the following:
Lemma 2 ([3]) Let K be a nonempty convex subset of a topological vector space
and let φ : K × K → R be such that:
1. for each x ∈ K , y → φ(x, y) is lower semicontinuous on each nonempty compact
subset of K ,

m
2. for each nonempty finite set {x1 , · · · , xm } ⊂ K and for each y = λi xi , (λi ≥
i=1

m
0, = 1), min1≤i≤m φ(xi , y) ≤ 0,
i=1
3. there exists nonempty compact convex subset X 0 of K and a nonempty compact
subset D of K such that for each y ∈ K /D, there is an x ∈ co(X 0 ∪ {y}) with
φ(x, y) > 0.
Then there exists an ŷ ∈ D such that φ(x, ŷ) ≤ 0, for all x ∈ K .
3 Results
3.1 Existence Result
We have first established an equivalence between the problems (1) and (2) given
below. Next we have shown that the set-valued mapping F : K → 2 E is a KKM
mapping. Then applying Lemma 1, solvability is proved. In Theorem 3 adding an
extra condition of η−coercivity to the mapping N , solvability is established in the
case, where K is unbounded.
Theorem 1 Let K be a nonempty compact convex subset of a real reflexive Banach

space E and E ∗ be the dual space of E. Let N : E × E → E ∗ be η−hemicontinuous
and weakly relaxed η − α monotone and b : E × E → R be a convex lower
semicontinuous function in second variable, such that,
1. η(w, w) = 0, ∀w ∈ K ,
2. v → N (w, y), η(v, w) is convex for any w, y ∈ K .
Then the following problems are equivalent:
w ∈ K , N (w, y), η(v, w) + b(w, v) − b(w, w) ≥ 0, ∀v ∈ K , (1)
w ∈ K , N (v, y), η(v, w) + b(w, v) − b(w, w) ≥ α(w − v), ∀v ∈ K . (2)
Proof Let w ∈ K be a solution of (1). As N is weakly relaxed η − α monotone, we

have,
N (v, y), η(v, w) + b(w, v) − b(w, w) ≥ N (w, y), η(v, w) + α(w − v)
+ b(w, v) − b(w, w) ≥ α(w − v), ∀v ∈ K .
So w is a solution of (2). Conversely let w ∈ K be a solution of (2). Let vt =

(1 − t)w + tv, t ∈ (0, 1). So vt ∈ K . As w ∈ K is a solution of (2), we have,

N (vt , y), η(vt , w) + b(w, vt ) − b(w, w) ≥ α t (w − v) , and
b(w, vt ) − b(w, w) ≤ t (b(w, v) − b(w, w)).
Using these results, we get,

α t (w − v)
N (w + t (v − w), y), η(v, w) + b(w, v) − b(w, w) ≥ , ∀v ∈ K .
t
d
Since N is η−hemicontinuous and lim α(t z) = 0, letting t → 0 and applying
t→0 dt
L’Hospital’s rule, we get,
N (w, y), η(v, w) + b(w, v) − b(w, w) ≥ 0, ∀v ∈ K .

Theorem 2 If N : E × E → E ∗ is η− hemicontinuous and weakly relaxed η − α
monotone and b : E × E → R is a convex lower semicontinuous function in second
variable and the following assumptions
1. η(v, v) = 0, ∀y ∈ K ,
2. v → N (w, y), η(v, w) is convex and lower semicontinuous for any w, y ∈ K ,
3. For any vβ , vβ converging to v, α(v) ≤ lim inf α(vβ ),
hold, then problem (1) is solvable.
Proof Let F, G : K → 2 E be defined by,
F(v) = {w ∈ K , N (w, y), η(v, w) + b(w, v) − b(w, w) ≥ 0}, ∀v ∈ K ,
G(v) = {w ∈ K , N (v, y), η(v, w) + b(w, v) − b(w, w) ≥ α(w − v)}, ∀v ∈ K .
We claim that F is a KKM mapping, if not, then there exists {v1 , · · · , vn } ⊂ K and

n
n n
ti > 0, i = 1, 2, · · · , n, such that, ti = 1, v = ti vi ∈
/ i=1 F(vi ). Then by
i=1 i=1
definition of F, N (v, y), η(vi , v) + b(w, vi ) − b(w, v) < 0, for i = 1, 2 · · · , n. By
our assumption
0 = N (v, y), η(v, v)

n n
=⇒ N (v, y), η ti vi , v ≤ ti N (v, y), η(vi , v)
i=1 i=1
n
< ti (b(w, v) − b(w, vi ))
i=1

n
= b(w, v) − ti b(w, vi )
i=1
≤ b(w, v) − b(w, v) = 0.
This is a contradiction. So F is a KKM mapping. We now claim that F(v) ⊂

G(v), ∀v ∈ K . Let w ∈ F(v), then
N (w, y), η(v, w) + b(w, v) − b(w, w) ≥ 0, for any v ∈ K .
Using weakly relaxed η − α monotonicity of N , we have w ∈ G(v). So F(v) ⊂

G(v), ∀v ∈ K . So G is a KKM mapping. As α is weakly lower semicontinuous, G(v)
is weakly closed. v → N (w, y), η(v, w) and b are convex and lower semicontinu-
ous and hence weakly lower semicontinuous. As K is bounded closed and convex in
the reflexive Banach space E, it is weakly compact. Since G(v) is weakly closed for
all v ∈ K , it isweakly compact and hence the family {G(v)} has finite intersection
property, i.e., v∈K G(v) = . So conditions of Lemma 1 are satisfied, hence we
have,
F(v) = G(v) = .
v∈K v∈K
So there exists w ∈ K such that
N (w, y), η(v, w) + b(w, v) − b(w, w) ≥ 0, ∀v ∈ K .

Next we prove the existence result for an unbounded set K .

Theorem 3 If N : E × E → E ∗ is η−hemicontinuous and weakly relaxed η − α
monotone, b : E × E → R is a convex lower semicontinuous function in second
variable, K is a nonempty closed convex unbounded subset of E and the following
assumptions
1. for w0 ∈ K , [N (w, y) − N (w0 , y), η(v, w0 ) + b(w, v) − b(w0 , v)]/ η(v,
w0 ) → ∞, as v → ∞; i.e., N is η−coercive with respect to b in the second
variable.
2. η(w, v) + η(v, w) = 0, ∀w, v ∈ K ,
4. for any vβ , vβ converging to v, α(v) ≤ lim inf α(vβ ),
hold, then problem (1) is solvable.
Proof Consider the problem, find wr ∈ K ∩ Br such that
N (wr , y), η(v, wr ) + b(wr , v) − b(wr , wr ) ≥ 0, ∀v ∈ K ∩ Br (3)
where Br = {v ∈ E : v ≤ r }.
By Theorem 2 (3) has a solution wr ∈ K ∩ Br . Choosing w0 < r, we can put
w0 in place of v in (3), so we have,
N (wr , y), η(w0 , wr ) + b(wr , w0 ) − b(wr , wr ) ≥ 0.
Now,
N (wr , y), η(w0 , wr ) + b(wr , w0 ) − b(wr , wr )

= −N (wr , y), η(wr , w0 ) + N (w0 , y), η(w0 , wr ) + N (w0 , y), η(wr , w0 )
+ b(wr , w0 ) − b(wr , wr )
= −N (wr , y) − N (w0 , y), η(wr , w0 ) + b(wr , w0 ) − b(wr , wr ) + N (w0 , y), η(w0 , wr )

−N (wr , y) − N (w0 , y), η(wr , w0 ) + b(wr , wr ) − b(wr , w0 )
≤ η(wr , w0 ) + N (w0 , y) .
η(wr , w0 )
If wr = r and r → ∞, then by η−coercivity of N with respect to b in the second

variable, the above inequality reduces to
N (wr , y), η(w0 , wr ) + b(wr , w0 ) − b(wr , wr ) < 0.
This is a contradiction as N (wr , y), η(w0 , wr ) + b(wr , w0 ) − b(wr , wr ) ≥ 0. So

wr < r. Now for any v ∈ K , we choose 0 < ε < 1, such that,
wr + ε(v − wr ) ∈ K ∩ Br .
By assumption 3 and convexity of b, we have from Eq. (3),
0 ≤ N (wr , y), η(wr + ε(v − wr ), wr ) + b(wr + ε(v − wr )) − b(wr , wr )

≤ (1 − ε)N (wr , y), η(wr , wr ) + εN (wr , y), η(v, wr ) + (1 − ε)b(wr , wr ) + εb(v, wr )
− b(wr , wr )
= εN (wr , y), η(v, wr ) + εb(wr , v) − εb(wr , wr ).
So,
N (wr , y), η(v, wr ) + b(wr , v) − b(wr , wr ) ≥ 0, ∀y ∈ K
and wr ∈ K is a solution of problem (1). Hence problem (1) is solvable.
3.2 Iterative Algorithm and Convergence Analysis
In this section, we have obtained an iterative algorithm for finding approximate solu-
tions to the nonlinear mixed variational-like inequality problem (1), using auxiliary
principle technique. This technique involves first formulating an auxiliary minimiz-
ing problem and then characterizing it by an auxiliary variational inequality problem.
In the formulation of the auxiliary minimizing problem, the differentiable convex
functional α : E → R is considered as an auxiliary differentiable convex functional.
The auxiliary minimizing problem is defined as follows:
minw∈K {α(w) + ρ N (v, y), η(w, v) − α (v), w + ρb(v, w)}, (4)
where w ∈ E, v ∈ K and ρ is a positive constant. If w → N (v, y), η(w, v) is

convex, then (4) is equivalent to following auxiliary variational inequality problem
in the sense that solution to both the problems are same. The auxiliary variational
inequality problem is given by,
α (w) − α (v), u − w ≥ − ρN (v, y), η(u, w) + ρb(v, w) − ρb(v, u), for all u ∈ K . (5)
Note 1 If w = v, then v is a solution of (1).

Keeping in view these results we propose an iterative algorithm as follows:
1. Let v0 be the initial approximation for n = 0.
2. At the nth step solve the auxiliary minimizing or auxiliary variational inequality
problem with v∗ = vn . Let vn+1 be the solution.
3. If vn+1 − vn ≤ ε; ε > 0, stop, otherwise repeat 2.
Next, we prove the result that gurantees the existence of solution of (4) or (5).
Theorem 4 Let E be a reflexive Banach space with dual space E ∗ , N : E × E →
E ∗ , b : E × E → R be η−hemicontinuous, convex lower semicontinuous function
in second variable, linear in the first argument, bounded and b(u, v) − b(u, w) ≤
b(u, v − w) and α : E → R be a differentiable convex functional, such that
2. N is weakly lower semicontinuous, weakly relaxed η − α monotone and
η−convex with respect to first argument for any w, y ∈ K ,
3. N is strongly η−monotone in first argument, η−antimonotone and weakly
relaxed monotone in second argument,
4. N is Lipschitz continuous in first and second argument with respect to constants
σ1 and σ2 respectively,
5. η(v, w) = η(v, u) + η(u, w), for any u, v, w ∈ K ,
6. η(v, u) + η(u, w) = 0, for any u, v, w ∈ K ,
7. η is Lipschitz continuous with respect to constant δ > 0,
8. N and η have 0−diagonally convex relation with respect to first argument,
9. w → α (w) is continuous from weak to strong topology and α is strongly
monotone,
10. α = 2μ, (σ1 + σ2 )δ + μ > 0 and 0 < ρ < ρ(σα−2μ1 +σ2 )δ
.
Then there exists a solution w ∈ K of the problem (1) and for each ρ > 0, there
exists a solution wn+1 ∈ K of problems (4) or (5) and the approximate solutions
converge strongly to the exact solution.
Proof As N is η−convex with respect to first argument for any w, y ∈ K , it is easy

to check that condition 2 of Theorem 2 is satisfied. By condition 6, η(v, v) = 0.
Hence condition 1 of Theorem 2 is satisfied. So all the conditions of Theorem 2 are
satisfied. Hence solution to problem (1) exists. Now to prove the second part of the
conclusion, we have to show that all the conditions of Lemma 2 are satisfied. For
this purpose we define φ : K × K → R, by,
φ(u, w) = α (vn ) − α (w), u − w − ρN (vn , yn ), η(u, w) + ρb(vn , w) − ρb(vn , u).
As w → α (w) is continuous from weak to strong topology, the function w →

α (w), w is weak continuous on K . So w → φ(u, w) is weakly lower semicon-
tinuous. So condition 1 is satisfied. To prove the second condition we assume the
contrary. So there exists {u 1 , · · · , u n } ⊂ K and w which is a convex combination of
u i , such that φ(u i , w) > 0. From this we get,

n
n
λi α (vn ) − α (w), u i − w − ρN (vn , yn ), η(u i , w) + ρb(vn , w) − ρ λi b(vn , u i ) > 0.
i=1 i=1
As b is convex in the second argument we have

n
λi α (vn ) − α (w), u i − w − ρN (vn , yn ), η(u i , w) > 0.
i=1
This contradicts condition 8. So condition 2 of Lemma 2 holds. Now considering

a set D = {v ∈ K : v − u ∗ ≤ θ }, where θ = α1 [μ u ∗ ] + δ N (u ∗ , y) and
using condition 2, 3, 6, 8 and conditions on b, condition 3 is proved. Hence all the
conditions of Lemma reflem2 are satisfied. So there exists w0 ∈ K such that
α (w0 ) − α (vn ), u − w0 ≥ −ρN (vn , yn ), η(u, w0 ) + ρb(vn , w0 ) − ρb(vn , u) (6)
for all u ∈ K . Hence there exists a solution.

Now for convergence analysis we consider the following functional Γ : K →
(−∞, +∞], defined by,
Γ (v) = α(v0 ) − α(v) − α (v), v0 − v,
where v0 is assumed to be the unique solution of problem (1). By strong monotonicity

of α , we have,
σ
Γ (v) = α(v0 ) − α(v) − α (v), v0 − v ≥ v − v0 2 .
2
Putting w0 = vn+1 , u = v0 in (6) and by antisymmetricity of η, strong monotonicity
of α , we get,
σ
Γ (vn ) − Γ (vn+1 ) ≥ vn − vn+1 2 + ρN (vn , yn ), η(vn+1 , v0 ) + ρb(vn , vn+1 )
2
σ
− ρb(vn , v0 ) = vn − vn+1 2 + ρN (vn , yn ) − N (v0 , y0 ), η(vn+1 , v0 )
2
+ ρN (v0 , y0 ), η(vn+1 , v0 ) + ρb(vn , vn+1 ) − ρb(vn , v0 ).
v0 being a solution of (1), it follows that,

σ
Γ (vn ) − Γ (vn+1 ) ≥ vn − vn+1 2 + ρN (vn , yn ) − N (v0 , y0 ), η(vn+1 , v0 )
2
+ ρ[b(v0 , v0 ) − b(v0 , vn+1 ) + b(vn , vn+1 ) − b(vn , v0 )]
σ
= vn − vn+1 2 + M.
2
Now using the conditions on b and the conditions 3–5 and 7, we get,
M = ρN (vn , yn ) − N (v0 , y0 ), η(vn+1 , v0 )

− ρ[b(vn − v0 , v0 ) − b(vn − v0 , vn+1 ) + b(vn − v0 , vn ) − b(vn − v0 , vn )]
≥ ρ[N (vn , yn ) − N (v0 , y0 ), η(vn+1 , vn ) + N (vn , yn ) − N (v0 , y0 ), η(vn , v0 )]
− ρ[b(vn − v0 , v0 − vn ) + b(vn − v0 , vn − vn+1 )]
≥ ρ[N (vn , yn ) − N (v0 , yn ), η(vn , v0 )
+ N (v0 , yn ) − N (v0 , y0 ), η(vn , v0 )
+ N (vn , yn ) − N (v0 , yn ), η(vn+1 , vn )
+ N (v0 , yn ) − N (v0 , y0 ), η(vn+1 , vn )]
− ρμ[ vn − v0 2 + vn − v0 vn − vn+1 ]
≥ ρα vn − v0 2 − ρσ1 δ vn − v0 vn+1 − vn
− ρσ2 δ vn − v0 vn+1 − vn − ρμ[ vn − v0 2 + vn − v0 vn − vn+1 ].
From this we have,
Γ (vn ) − Γ (vn+1 ) ≥ ρ[α − (2μ + (σ1 + σ2 )δ)] vn − v0 2 .
Condition 10 implies that {Γ (vn )} is a strictly decreasing sequence and it is nonneg-

ative by the strong monotonicity property and hence converges. So {vn } converges
to v0 strongly as n → ∞. This completes the proof.
4 Concluding Remarks
In this work, we have studied the existence of the solution of nonlinear mixed
variational-like inequality with respect to weakly realaxed η − α monotone map-
ping in case of both bounded and unbounded sets. We have obtained an iterative
algorithm using auxiliary principle technique and we have shown that the iterates
approximate to the exact solution strongly. Further we are trying to frame this prob-
lem for nonconvex setting using hemivariational inequality concept.
Acknowledgments The first author wishes to thank DST-INSPIRE (Code No. IF110762) for the
grant of research fellowship and to IIT Bhubaneswar for providing the research facilities.
References
1. Bai, M.R., Zhou, S.Z., Ni, G.Y.: Variational-like inequalities with relaxed pseudomonotone
mappings in banach spaces. Appl. Math. Lett. 19(6), 547–554 (2006)
2. Chen, Y.Q.: On the semi-monotone operator theory and applications. J. Math. Anal. Appl.
231(1), 177–192 (1999)
3. Ding, X., Tan, K.: A minimax inequality with applications to existence of equilibrium point
and fixed point theorems. Colloq. Math. 63, 233–247 (1992)
4. Ding, X.P.: General algorithm for nonlinear variational-like inequalities in reflexive banach
spaces. Indian J. Pure Appl. Math. 29(2), (1998)
5. Fan, K.: Some properties of convex sets related to fixed point theorems. Math. Ann. 266(4),
519–537 (1984)
6. Fang, Y., Huang, N.: Variational-like inequalities with generalized monotone mappings in
banach spaces. J. Optim. Theory Appl. 118(2), 327–338 (2003)
7. Huang, N.J.: On the generalized implicit quasivariational inequalities. J. Math. Anal. Appl.
216(1), 197–210 (1997)
8. Huang, N.J., Deng, C.X.: Auxiliary principle and iterative algorithms for generalized set-valued
strongly nonlinear mixed variational-like inequalities. J. Math. Anal. Appl. 256(2), 345–359
(2001)
9. Kutbi, M.A., Sintunavarat, W.: On the solution existence of variational-like inequalities prob-
lems for weakly relaxed monotone mapping. In: Abstract and Applied Analysis, vol. 2013.
Hindawi Publishing Corporation (2013)
10. Tian, G.: Generalized quasi-variational-like inequality problem. Math. Oper. Res. 18(3), 752–
764 (1993)
11. Tremolieres, R., Lions, J.L., Glowinski, R.: Numerical analysis of variational inequalities.
Elsevier (2011)
12. Yao, J.C.: The generalized quasi-variational inequality problem with applications. J. Math.
Anal. Appl. 158(1), 139–160 (1991)
Pál Type (0; 1)-Interpolation
on Mixed Tchebycheff Abscissas-II
Neha Mathur and Pankaj Mathur
Abstract In this paper, we have considered the interpolation problem when function
values are prescribed on the zeros of (n-1)th Tchebycheff polynomial of second
kind and weighted first derivatives are prescribed on the zeros of nth Tchebycheff
polynomial of first kind. It has been shown that such an interpolation exists when
n is even, the explicit representation of which has been obtained. The convergence
theorem for the interpolatory polynomial has also been dealt with.
Keywords Interpolation · Tchebycheff polynomials · Inter-scaled zeros
1 Introduction
n−1
Let {x2i,2n+1 }i=1
n and {x2i+1,2n+1 }i=1 be two distinct point systems in the interval
[−1, 1], which are interscaled such that
− 1 = x2n+1,2n+1 < x2n,2n+1 < · · · < x1,2n+1 = 1 (1)
where the points {x2i,2n+1 }i=1

n are the distinct zeros of Tn (x) the nth Tchebycheff
n−1
polynomial of first kind and {x2i+1,2n+1 }i=1 are the distinct zeros of Un−1 (x) the
2n+1
nth Tchebycheff polynomial of second kind. Further, let {αi,2n+1 }i=1 be arbitrary
given numbers. We seek to find a polynomial Sn (x) of degree ≤ 2n satisfying the
conditions:
Sn (x2i+1,2n+1 ) = α2i+1,2n+1 ; i = 0, 1, 2, . . . , n (2)

1 − x 2 Sn (x2i,2n+1 ) = α2i,2n+1 ; i = 1, 2, . . . , n (3)
N. Mathur
Department of Mathematics, Career Convent College, Lucknow, India
e-mail: neha_mathur13@yahoo.com
P. Mathur (B)
Department of Mathematics and Astronomy, Lucknow University, Lucknow, India
e-mail: pankaj_mathur14@yahoo.co.in

DOI 10.1007/978-81-322-2485-3_15
198 N. Mathur and P. Mathur
In 1975, Pál [11] introduced an interpolation process on an interscaled set of points
− ∞ < xn < yn−1 < · · · < x2 < y1 < x1 < ∞. (4)

n−1
where {xi }i=1
n and {yi }i=1 are the distinct real zeros of (say)

n
Wn (x) = (x − xi ).
l=1
and of Wn (x), respectively. Pál proved that for given arbitrary numbers {αi∗ }i=1
n
and {βi∗ }i=1 there exists a unique polynomial of degree ≤ 2n − 1 satisfying the
n−1
conditions:
Rn (xi ) = αi∗ , i = 1, 2, . . . , n, (5)

Rn (yi ) = βi∗ , i = 1, 2, . . . , n − 1 (6)
and an initial condition
Rn (a) = 0
where a is a given point, different from the nodal points (4). After which many
mathematicians have taken up this problem on different sets of nodes. For more
details, one is referred to [1–3, 5–10, 12–15, 17, 18], etc.
In this paper, we have considered the converse problem considered in [8]. We
have shown that for n even, there exist a unique polynomial Sn (x) of degree ≤ 2n
satisfying the conditions (2)–(3). The explicit representation of Sn (x) is obtained and
the estimates of the fundamental polynomials leading to the convergence theorem
has also been dealt with.
In Sect. 2, we give preliminaries. Existence, uniqueness, and the explicit repre-
sentation of the interpolatory polynomials have been dealt with in Sect. 3. Section 4
is devoted to the estimation of the fundamental polynomials and the proof of the
convergence theorem.
2 Preliminaries
We characterize the points

1 π
x2i = cos i − , i = 1(1)n (7)
2 n
as the zeros of Tn (x) = cos nθ, x = cosθ(−1 < x < 1), nth Tchebycheff polyno-
mial of first kind and
Pál Type (0; 1)-Interpolation on Mixed Tchebycheff Abscissas-II 199

iπ
x2i+1 = cos , i = 1(1)n − 1 (8)
n
as the zeros of Un−1 (x) the (n − 1)th Tchebycheff polynomial of second kind.
n−1
Obviously, x = 0 either in {x2i }i=1
n or in {x2i+1 }i=1 according as n is even or odd.
Also xi = −x2n+2−i , i = 1(1)n. The differential equation satisfied by Tn (x) [16] is
(1 − x 2 )Tn (x) − x Tn (x) + n 2 Tn (x) = 0 (9)
and that Un−1 (x) is

(1 − x 2 )Un−1 (x) − 3xUn−1 (x) + (n 2 − 1)Un−1 (x) = 0. (10)
We have that for i = 1, 2, . . . , n
2
n
Tn (x) 1
2i (x) = = + Tr (x2i )Tr (x) (11)
(x − x2i Tn (x2i )) n n
r =1
and for i = 1, 2, . . . , n − 1
Un−1 (x)
2i+1 (x) = (x
(x − x2i+1 Un−1 2i+1 ))
2(1 − x2i+1
2 )
n−1
= Ur (x2i+1 )Ur (x). (12)
n
r =1
Also,

n
|2i (x)| = O(log n), (13)
i=1

n−1
|2i+1 (x)| = O(n), (14)
i=1
n
−m
O(n log n), m = 1/2
1 − xk2 = (15)
O(n 2 ), m=1
k=1
and

k k
c1 ≤ 1 − xk2 ≤ c2 (16)
n n
where xk s are the zeros of Tn (x) or Un−1 (x).

3 Existence, Uniqueness, and Explicit Representation

of the Interpolatory Polynomials
We shall prove the following:

Theorem 1 Let n be even and the (2n + 1) points in [−1, 1] be given by Eq. (1)
together with (7) and (8), then there exist a unique polynomial Sn (x) of degree ≤ 2n
satisfying the conditions (2) and (3). But if n is odd then there exists, in general, no
polynomial of degree ≤ 2n and if the polynomials exist, they are infinitely many.
For n even, the interpolatory polynomial Sn (x) satisfying the conditions (2) and (3)
can be represented as:

n
n
Sn (x) = α2i+1 A2i+1 (x) + α2i B2i (x) (17)
i=0 i=1
where {A2i+1 (x)}i=0

n and {B (x)}n are called the fundamental polynomials of first
2i i=1
and second kind, respectively, each of degree ≤ 2n, which are uniquely determined
by the following conditions: for i = 0, 1, 2, . . . , n

A (x )=δ , j = 0, 1, 2, . . . , n
√ 2i+1 2 j+1 i j (18)
1 − x 2 A2i+1 (x2 j ) = 0, j = 1, 2, . . . , n
and for i = 1, 2, . . . , n

B2i (x2 j+1 ) = 0, j = 0, 1, 2, . . . , n
√ (19)
1 − x B2i (x2 j ) = δi j ,
2 j = 1, 2, . . . , n
The explicit forms of fundamental polynomials are given in the following:

Theorem 2 For n even, the fundamental polynomials {B2i (x)}i=1
n satisfying the
conditions (19) can be represented as

Un−1 (x) x x
B2i (x) =
2i (x)d x + c2i Tn (x)d x (20)
1 − x2i −1 −1
n−1 (x 2i )
2U
where 2i (x)are defined by (11) and

1 −1 1
c2i = − Tn (x)d x 2i (x)d x.
−1 −1
Theorem 3 For n even, the fundamental polynomials {A2i+1 (x)}i=0 n satisfying the
conditions (18) can be represented as for i = 1, 2, . . . , n − 1
(1 − x 2 )Tn (x)2i+1 (x)

A2i+1 (x) = (21)
(1 − x2i+1
2 )Tn (x2i+1 )

Un−1 (x) x (1 − x 2 )Tn (x)
−
dx
(1 − x2i+1 )Tn (x2i+1 )Un−1 (x2i+1 )
2
−1 x − x2i+1
x
+ c3i Tn (x)d x
−1
where 2i+1 (x) are given by (12) and

−1
1 1 (1 − x 2 )Tn (x)
c3i = − Tn (x)d x d x.
−1 −1 x − x2i+1
For i = 0, n;
x
A1 (x) = c3 Un−1 (x) Tn (x)d x
−1
where
1 −1
c3 = n Tn (x)d x .
−1
Similarly we have
x
A2n+1 (x) = c4 Un−1 (x) Tn (x)d x
1
where c4 = −c3 .
The polynomial Sn (x), for n even satisfies the following quantitative estimate:
Theorem 4 Let f (x) have a continuous derivative in [−1, 1], then

n
n
Sn ( f, x) = f (x2i+1 )A2i+1 (x) + f (x2i )B2i (x) (22)
i=0 i=1
satisfies the relation:

|Sn (x) − f (x)| = O(1) (1 − x 2 )|Tn (x)| + |Un−1 (x)| log n
√ √
1 − x2 1 − x2
+ ω f ,
n n
√
1−y 2
where δn (y) = n and cr (r = 1, 2, 3) are constants, independent of f, n and x.
We will prove only our main Theorem 4 in this chapter, as the proof of other
Theorems is quite similar to that of theorems in [5]. In order to prove the theorem,
we shall need the estimates of the fundamental polynomials.
4 Estimation of the Fundamental Polynomials
Lemma 1 For n even and x ∈ [−1, 1], we have

x

T (x)d x ≤ 2(n 2 − 1)−1 (23)
n
−1
and
x log n
2i (x)d x ≤ O (24)
n
−1
where 2i (x) are given by (11).
Proof On integrating the left-hand side of (23), we get

x 1 Tn−1 (x) Tn+1 (x) 2
Tn (x)d x = − + 2
−1 2 n−1 n+1 n −1
from which (23) follows. For (24), we have by (11)

x x n
1 2
2i (x)d x = + Tr (x2i )Tr (x) d x
−1 n n
−1 r =1
n
2 1 x

≤ + Tr (x2i ) Tr (x)d x
n n −1
r =1
2
1 |x − 1| cos(r − 1)θ
n
2 cos(r + 1)θ 2
≤ + + r −1 − +
n 2n 2 r +1 r2 − 1
r =2

log n
=O .
n
Thus (24) follows.
Lemma 2 For n even and x ∈ [−1, 1], we have

x (1 − x 2 )Tn (x) 4n log n
dx ≤
. (25)
x − x2i+1
−1 1 − x2i+1
2
Proof Since Tn (x) = nUn−1 (x), we have

x (1 − x 2 )Tn (x) x

d x ≤ n Un−1 (x2i+1 ) (1 − x )2i+1 (x)d x .
2
(26)
x − x2i+1
−1 −1
By (12), we have
n−1 x
x 2 1 − x2i+1
2

1 − x 2 2i+1 (x)d x ≤ Ur (x2i+1 ) 1 − x 2 Ur (x)d x (27)
−1 n −1
r =0
Since

x Tr −1 (x) Tr +3 (x) Tr +1 (x)
(1 − x 2 )Ur (x)d x = + − (28)
−1 4(r − 1) 4(r + 3) 2(r + 1)

(−1)r 2r 2 + 8r + 1
+
4 (r − 1)(r + 3)(r + 1)
hence by using (27) and (28) in (26) the lemma follows.

Lemma 3 For n even, i = 1, 2, 3, . . . , n and x ∈ [−1, 1], we have
log n
|B2i (x)| ≤ |Un−1 (x)|
n
Proof The proof of this Lemma follows by (20) and Lemma 1.
Lemma 4 For n even, i = 1, 2, 3, . . . , n − 1 and x ∈ [−1, 1], we have

1 − x 2 Tn (x)2i+1 (x) 4 |Un−1 (x)| log n
|A2i+1 (x)| ≤ +
1 − x2
2i+1 1 − x2i+1
2
Proof Using (23) and Lemma 2 in (21) the Lemma follows.
5 Proof of the Main Theorem
In order to prove our main Theorem 4, we need the following important result of
Gopengaus [4]: Let f ∈ C r [−1, 1], then for n ≥ 4r + 5, there exists a polynomial
Q n (x) of degree atmost n such that for all x ∈ [−1, 1] and for k = 0, 1, . . . , r

(k)
f (x) − Q (k)
n ((x) ≤ ck (δ n (x))r −k
ω f (r )
, δ n (x) , (29)
√
where δn (x) = 1−x 2
n and ck s are constants independent of f, n, and x.
Proof of Theorem 4 From the uniqueness of Sn (x) in (22), it follows that every
polynomial Q n (x) of degree ≤ 2n with the property (29) satisfies the relation

n
n
Q n (x) = Q n (x2i+1 )A2i+1 (x) + Q n (x2i )B2i (x).
i=0 i=1
Therefore,
|Sn (x) − f (x)| ≤ |Sn (x) − f (x)| + |Sn (x) − f (x)|

n
≤ | f (x2i+1 ) − Q n (x2i+1 )| |A2i+1 (x)| (30)
i=0
n
+ | f (x2i+1 ) − Q n (x2i+1 )||B2i (x)| + |Q n (x) − f (x)|
i=1
≡ 1 + 2 + |Q n (x) − f (x)| (31)
Now by (29) and Lemma 4, we have

⎛
⎞

8 1 − x 2i+1 1 − x2i+1
2 2

n−1 c
⎠ |(1 − x )Tn (x)2i+1 (x)|
2
1 ≤ ω ⎝ f ,
i=1
n n |(1 − x2i+1
2 )|

4 |Un−1 (x)| log n
+
1 − x2i+1
2

Using the property of the modulus of continuity ω f , λδ ≤ (1 + λ)ω f , δ , we
have
⎛

⎞

n−1 2
1 − x2i+1 2
c8 1 − x2
⎝1 + ⎠ |(1 − x
)Tn (x)2i+1 (x)|
1 ≤ ω
f ,
n n 1 − x2 2
1 − x2i+1
i=1

+ 4|(1 − x 2 )Un−1 (x)| log n

c8 1 − x2 2i+1 (x)
n−1
≤ ω f , | 1 − x 2 Tn (x)|
+ 4n|Un−1 (x)| log n
n n 2
i=1 1 − x 2i+1

4|Un−1 (x)| log n
n−1 n−1
+ | (1 − x 2 )Tn (x)| 2i+1 (x) + 2
1 − x2i+1
i=1 1 − x2 i=1
Thus by Schwartz inequality, (14), (15) and (16), we have

√
1 − x2

1 ≤ c9 ω f , (1 − x 2 )Tn (x) + c10 |Un−1 (x)| log n (32)
n
Also by (29) and Lemma 3, we have

⎛
⎞

n 1 − x2i
2
2 ≤ c11 ω ⎝ f , ⎠ |Un−1 (x)| log n .
n n
i=1
Again by the property of the modulus of continuity, we have
n−1 ⎛
⎞
√
log n 1− x2 1 − x2i
2
2 ≤ c11 ω f , ⎝1 + √ ⎠
n n
i=1 1 − x 2
Thus by (15), we have

√
1 − x2
2 ≤ c12 |Un−1 (x)| log nω f , . (33)
n
Hence using (29), (32) and (33) in (30), Theorem 4 follows.
References
1. Akhlagi, M.R., Sharma, A.: Some Pál Type Interpolation Problems, Approx., Optimization and
Computing Theory and Applications (A.G. Law and C.L. Wang Ceds.), pp. 37–40. Elsevier
Science Publisher B.V. (North Holland), IMACS (1990)
2. Datta, S., Mathur, P.: On Pál type weighted (0,2;0)-interpolation on infinite interval (−∞, ∞).
Approx. Theory Appl. China 17(4), 1–10 (2001)
3. Datta, S., Mathur, P.: Modified weighted (0; 0, 2)-interpolation on infinite interval (−∞, ∞),
Ann. Univ. Sci. Budapest 44, 39–52 (2001)
4. Gopengauz, I.E.: On the Theorem of A.F. Timan on approximation of continuous functions on
a line segment. Math. Zametski 2, 163–172 (1967)
5. Joó, I.: On Pál interpolation. Ann. Univ. Sci. Budapest, Sect. Math. 37, 247–262 (1994)
6. Joó, I., Pál, L.G.: Lacunary (0; 0, 1)-interpolation on the roots of Jacobi polynomials and their
derivatives, respectively I (existence, explicit formulae, unicity). Anal. Math. 22, 289–298
(1996)
7. Mathur, P.: (0, 1; 0)-interpolation on infinite interval (−∞, +∞). Anal. Theory Appl. 22(2),
105–113 (2006)
8. Mathur, N., Mathur, P.: Pál type (0; 1)-interpolation on mixed Tchebycheff abscissas-I. Int. J.
Math. Arch. 4(8), 225–235 (2013)
9. Mathur, K.K., Srivastava, R.: Pál type interpolation on infinite interval. J. Math. Anal. Appl.
192, 346–359 (1995)
10. Muneer, Y.E.: On the convergence of (0, 1, 2) interpolation. Serdica Math. J. 31, 87–94 (2005)
11. Pál, L.G.: A new modification of H-F interpolation. Anal. Math. I, 197–205 (1975)
12. Pál, L.G.: A general Lacunary (0; 0, 1)-interpolation process. Anal. Univ. Budapest, Sect.
Comp. 16, 291–301 (1996)
13. Sebestyen, Z.F.: Pál type interpolation on the roots of Hermite polynomials. Pure Math. Appl.
9(3–4), 429–439 (1998)
14. Srivastava, R., Mathur, K.K.: Weighted (0; 0, 2)-interpolation on the roots of Hermite polyno-
mials. Acta. Math. Hung. 70(1–2), 57–73 (1996)
15. Szabados, J., Varma, A.K.: On a convergent Pál-type (0,2)-interpolation process. Acta. Math.
Hungar. 66(4), 301–326 (1995)
16. Szegö, G.: Orthogonal polynomials. Am. Math. Soc. Coll. Publ. (1959)
17. Szili, L.: A convergence theorem for the Pál, method of interpolation on the roots of Hermite
polynomials. Anal. Math. 11, 75–84 (1985)
18. Xie, T.F.: On Pás problem. Chinese Quart. J. Math. 7, 48–52 (1992)
Degree of Approximation of f ∈ L[0, ∞)
by Means of Fourier–Laguerre Series
Soshal Saini and Uaday Singh
Abstract In this paper, we determine the degree of approximation of functions

belonging to L[0, ∞) by the Hausdorff means of its Fourier–Laguerre series at
x = 0. Our theorem extends some of the recent results of Nigam and Sharma
[A study on degree of approximation by (E, 1) summability means of the Fourier–
Laguerre expansion, Int. J. Math. Math. Sci. (2010), Art. ID 351016, 7], Krasniqi
[On the degree of approximation of a function by (C, 1)(E, q) means of its
Fourier–Laguerre series, International Journal of Analysis and Applications 1 (2013),
33–39] and Sonker [Approximation of Functions by (C, 2)(E, q) means of its
Fourier–Laguerre series, Proceeding of ICMS-2014 ISBN 978-93-5107-261-4:125–
128.] in the sense that the summability methods used by these authors have been
replaced by the Hausdorff matrices.
Keywords Degree of approximation · Hausdorff means · Fourier–Laguerre series
1 Introduction
Let f be a function belonging to L[0, ∞) in the sense that f is Labesgue integrable

in the interval [0, ∞). The Fourier–Laguerre expansion of f is given by
∞

f (x) ∼ an L (α)
n (x), (1)
n=0
S. Saini (B) · U. Singh

e-mail: sainisoshal25@gmail.com
U. Singh
e-mail: usingh2280@yahoo.co.in

DOI 10.1007/978-81-322-2485-3_16
208 S. Saini and U. Singh
where
∞
n+α
Γ (α + 1) an = e−x x α f (x)L (α)
n (x)d x (2)
n 0
(α)
and L n (x) denotes the nth Laguerre polynomial of order α > −1, defined by the
generating function
∞

−xω
L (α)
n (x)ω = (1 − ω)
n −α−1
exp . (3)
1−ω
n=0
When x = 0,
n+α
L (α)
n (0) = [9].
n
The nth partial sums of (1) are defined by

n
(α)
sn ( f ; x) = ak L k (x). (4)
k=0
The Cesàro means of order λ of the Fourier–Laugerre series are defined by

n

1 λ+n−k−1
Cnλ ( f ; x) = sk ( f ; x).
n + λ k=0 n−k
n
The Euler means of the Fourier–Laugerre series are defined by

n

q 1 n
E n ( f ; x) = q n−k sk ( f ; x), q > 0.
(1 + q)n k
k=0
The Hausdorff matrix H ≡ (h n,k ) is an infinite lower triangular matrix defined by

⎧
⎨ n
n−k μk , 0 ≤ k ≤ n,
h n,k = k
⎩
0 , k > n,
where is the forward difference operator defined by μn = μn − μn+1 and

k+1 μn = k (μn ). If H is regular, then {μn }, known as moment sequence, has
the representation
1
μn = u n dγ (u),
0
Degree of Approximation of f ∈ L[0, ∞) … 209
where γ (u), known as mass function, is continuous at u = 0 and belongs to BV[0, 1]

such that γ (0) = 0, γ (1) = 1; and for 0 < u < 1, γ (u) = [γ (u + 0) + γ (u −
0)]/2 [11].
The Hausdorff means of the Fourier–Laugerre series are defined by

n
Hn ( f ; x) := h n,k sk ( f ; x), n = 0, 1, 2, ... (5)
k=0
The Fourier–Laugerre series is said to be summable to s by the Hausdorff means, if

Hn ( f ; x) → s as n → ∞, [3].
For the examples of Hausdorff matrices, one can see [7, 8, 11] and references
therein.
In this paper, the class of all regular Hausdorff matrices with moment sequence
{μn } associated with mass function γ (u) having constant derivative, is denoted by
H1 .
We also write
e−y y α ( f (y) − f (0))
ϕ(y) = ,
Γ (α + 1)
and
n
n
g(u, y) = u k (1 − u)n−k L k(α+1) (y).
k
k=0
2 Known Results
Gupta [2] obtained the degree of approximation of f ∈ L[0, ∞) by Cesàro means

of order k of the Fourier–Laguerre series at the point x = 0, where k > α + 1/2.
Nigam and Sharma [5] have used (E, 1) means of the Fourier–Laguerre series for
−1 < α < 1/2 which is more appropriate range from the application point of view.
The authors have proved the following result:
Theorem A If
n
1 n
E n1 = n s → ∞ as n → ∞,
2 k k
k=0
then the degree of approximation of Fourier–Laguerre expansion at the point x = 0

by (E, 1) means E n1 is given by
E n1 (0) − f (0) = o(ξ(n)), (6)
provided that
t
Φ(t) = |ϕ(y)|dy = o t α+1 ξ(1/t) , t → 0, (7)

0
n
e y/2 y −((2α+3)/4) |ϕ(y)|dy = o n −((2α+1)/4) ξ(n) , (8)

δ
∞
e y/2 y −1/3 |ϕ(y)|dy = o(ξ(n)), n → ∞, (9)
n
where δ is a fixed positive constant and α ∈ (−1, −1/2), and ξ(t) is a positive
monotonic increasing function of t such that ξ(n) → ∞ as n → ∞.
Following, Nigam and Sharma [5], Krasniqi [4] has used the (C, 1)(E, q) means of
the Fourier–Laguerre series to obtain the degree of approximation of f ∈ L[0, ∞)
at point x = 0 and has proved the following result:
Theorem B The degree of approximation of the Fourier–Laguerre expansion at the
point x = 0 by the [(C, 1)(E, q)]n means is given by
[(C, 1)(E, q)]n (0) − f (0) = o(ξ(n)), (10)
provided that the conditions (7)–(9) given in Theorem A are satisfied.

Recently, Sonker [10] has also proved the same result using [(C, 2)(E, q)]n means
of the Fourier–Laguerre series for the same range of α as follows:
Theorem C The degree of approximation of the Fourier–Laguerre expansion at the
point x = 0 by the [(C, 2)(E, q)]n means is given by
[(C, 2)(E, q)]n (0) − f (0) = o(ξ(n)), (11)
provided that the conditions (7)–(9) given in Theorem A are satisfied.

v v
Remark 1 We observe that Krasniqi [4, p. 37] has optimized k=0 q k k (2α+1)/4
k
by its maximum value (1 + q)v v(2α+1)/4 . This is possible only when α > −1/2. But
the author has used −1 < α < 1/2 [4, p. 35, Theorem 2.1]. Similar error can also
be seen in [5, 10].
3 Main Results
In this paper, we extend the above results using the Hausdorff means, which is a
more general summability method, for an appropriate range of α. More precisely,
we prove the following:
Theorem 1 The degree of approximation of f ∈ L[0, ∞) at the point x = 0 by the

Hausdorff means of the Fourier–Laguerre series generated by H ∈ H1 is given by
Hn ( f ; 0) − f (0) = o(ξ(n)) (12)
where ξ(t) is a positive increasing function such that ξ(t) → ∞ as t → ∞ and

satisfies the following conditions
t
Φ(y) = |ϕ(y)|dy = o t α+1 ξ(1/t) , t → 0, (13)

0
n
e y/2 y −((2α+3)/4) |ϕ(y)|dy = o n −((2α+1)/4) ξ(n) , (14)

δ
and
∞
e y/2 y −1/3 |ϕ(y)|dy = o(ξ(n)), n → ∞, (15)
n
where δ is a fixed positive constant and α > −1/2.
For the proof of our theorem, we need the following lemmas:

Lemma 1 [9, p. 177]. Let α be an arbitrary real number, c and δ be fixed positive
constants. Then

(α)
O n (α) , if 0 ≤ x ≤ nc ,
L n (x) = −(2α+1)/4 (2α−1)/4 (16)
O x n , if nc ≤ x ≤ δ,
as n → ∞.
Lemma 2 [9, p. 240]. Let α be an arbitrary real number, δ > 0 and 0 < η < 4.
Then

−x/2 (α/2+1/4) (α)
O n (α/2−1/4) , if δ ≤ x ≤ (4 − η)n,
max e x |L n (x)| =
O n (α/2−1/12) , if x ≥ δ,
(17)
as n → ∞.
Lemma 3 For 0 < u < 1 and 0 ≤ y ≤ δ,

(α+1)
1 O n , if 0 ≤ y ≤ n1 ,
g(u, y)dγ (u) = −(2α+3)/4 (2α+1)/4 (18)

0 O y n , if n1 ≤ y ≤ δ,
as n → ∞.
Proof The g(u, y) can be written as

n
k
n u (α+1)
g(u, y) = (1 − u)n Lk (y).
k 1−u
k=0
Then
1 n k

1 n u (α+1)

g(u, y)dγ (u) = (1 − u) n
Lk (y)dγ (u)
0 k 1−u
0 k=0

1 n k
n u (α+1)
= M (1 − u)n Lk (y)du
0 k 1−u
k=0
Now, using Lemma 1 for 0 ≤ y ≤ 1

n (taking α + 1 for α and c = 1), we have
1 n k
1 n u
g(u, y)dγ (u) = (1 − u)n O(k α+1 )du
k 1−u
0 0 k=0
1 n k
n u
α+1
=O n (1 − u) n
du
0 k 1−u
k=0
1
= O n α+1 (1 − u)n du

0
α+1
=O n . (19)
Again, using Lemma 1 for 1

n ≤ y ≤ δ, we have
n k
1 1
n u
g(u, y)dγ (u) = (1 − u) n O y −(2α+3)/4 k (2α+1)/4 du
0 0 k 1−u
k=0
⎛ ⎞
1 n k
−(2α+3)/4 (2α+1)/4 n u
= O ⎝y n (1 − u) n du ⎠
0 k 1−u
k=0

= O y −(2α+3)/4 n (2α+1)/4 . (20)
Collecting (19) and (20), the proof of Lemma 3 is completed.
Lemma 4 For 0 < u < 1,

1 y/2 −(2α+3)/4 (2α+1)/4
O e y n , if δ ≤ y ≤ n,
g(u, y)dγ (u) = y/2 −(3α+5)/6 (α+1)/2 (21)
O e y n , if y ≥ δ,
0
as n → ∞.
Proof Following the Lemma 3, we have

1 n
k

1 n u (α+1)
g(u, y)dγ (u) = (1 − u) n
Lk (y)du
0 k 1−u
0 k=0
Now, using Lemma 2 for δ ≤ y ≤ n (taking α + 1 for α and η = 3), we have

n k
1 1
n u −(y/2) (2α+3)/4 (α+1)
g(u, y)dγ (u) = e (y/2) −(2α+3)/4
y (1 − u)n
e y Lk (y)du
0 0 k 1−u
k=0
1 n
k
n u
= e y/2 y −(2α+3)/4 (1 − u)n O k (2α+1)/4 du
0 k 1−u
k=0

= O e y/2 y −(2α+3)/4 n (2α+1)/4 . (22)
Again, using Lemma 2 for y ≥ n, we have

n k
1 1
n u

g(u, y)dγ (u) = e(y/2) y −(3α+5)/6 (1 − u)n e−(y/2) y (3α+5)/6 L k(α+1) (y)du
0 0 k 1−u
k=0
1 n
k
n u
= e y/2 y −(3α+5)/6 (1 − u)n O k (α+1)/2 du
0 k 1−u
k=0

y/2 −(3α+5)/6 (α+1)/2

= O e y n . (23)
Collecting (22) and (23), the proof of Lemma 4 is completed.
Proof of Theorem 1 We have

n
(α)
sn (0) = ak L k (0)
k=0
n ∞
1 (α) (α)
= e−y y α f (y)L k (y)dy L k (0)
k=0 Γ (α + 1)
n + α 0
n
∞ n
1 (α)
= e−y y α f (y) L k (y)dy
Γ (α + 1) 0
k=0
∞
1
= e−y y α f (y)L n(α+1) (y)dy,
Γ (α + 1) 0
so that

n
Hn ( f ; 0) = h n,k sk (0)
k=0
n ∞
n 1 (α+1)
= n−k
Δ μk e−y y α f (y)L k (y)dy.
k Γ (α + 1) 0
k=0
Thus
n
∞
n 1 (α+1)
Hn ( f ; 0) − f (0) = Δn−k μk e−y y α f (y)L k (y)dy − f (0)
k Γ (α + 1) 0
k=0
n
n ∞
1 (α+1)
= Δn−k μk e−y y α ( f (y) − f (0))L k (y)dy
k Γ (α + 1) 0
k=0
n ∞
n (α+1)
= Δn−k μk ϕ(y)L k (y)dy
k 0
k=0
∞ n
n (α+1)
= ϕ(y) Δ μk L k
n−k
(y) dy
0 k
k=0
∞ n
n 1
(α+1)
= ϕ(y) u (1 − u) dγ (u)L k
k n−k
(y) dy
0 k 0
k=0
∞ n

1
n n−k (α+1)
= ϕ(y) u (1 − u)
k
Lk (y)dγ (u) dy
0 0 k=0 k
∞ 1
= ϕ(y) g(u, y)dγ (u) dy
0 0
and

∞ 1

|Hn ( f ; 0) − f (0)| = ϕ(y) g(u, y)dγ (u) dy
0 0
∞
1

≤ |ϕ(y)| g(u, y)dγ (u) dy
0 0
n ∞
1/n δ 1

= + + + |ϕ(y)| g(u, y)dγ (u) dy
0 1/n δ n 0
= I1 + I2 + I3 + I4 . (24)
Now, using Lemma 3 for 0 ≤ y ≤ n1 , we have

1/n 1
I1 = |ϕ(y)| g(u, y)dγ (u) dy
0 0

1/n
= O n α+1 |ϕ(y)|dy
0

(α+1) 1 α+1
= O(n )o ξ(n)
n
= o(ξ(n)), (25)

Further, using Lemma 3 for 1

n ≤ y ≤ δ, we have,
δ
I2 = |ϕ(y)|O y −(2α+3)/4 n (2α+1)/4 dy

1/n

δ
(2α+1)/4 −(2α+3)/4
= O n y |ϕ(y)|dy .
1/n
Following [5, p. 6], we have

I2 = o(ξ(n)), (26)

Now, using Lemma 4 for δ ≤ y ≤ n, we have
1
n
I3 =
|ϕ(y)| g(u, y)dγ (u) dy
δ 0
n
= O e y/2 y −((2α+3)/4) n (2α+1)/4 |ϕ(y)|dy

δ

n
(2α+1)/4 y/2 −((2α+3)/4)
= O n e y |ϕ(y)|dy

δ
= O n (2α+1)/4 o (n −(2α+1)/4 )ξ(n)

= o(ξ(n)), (27)

Further, using Lemma 4, we have

∞ 1
I4 = |ϕ(y)| g(u, y)dγ (u) dy
n 0
∞
= |ϕ(y)|O e y/2 y −(3α+5)/6 n (α+1)/2 dy

n

∞ e y/2 y −1/3 |ϕ(y)|
(α+1)/2
= O n dy
n y (α+1)/2

= o (ξ(n))n (α+1)/2 n −(α+1)/2

= o(ξ(n)), (28)

Collecting (24)–(28), we have
Hn ( f ; 0) − f (0) = o(ξ(n)).
Hence the proof of Theorem 1 is completed.

4 Corollaries
The following corollaries can be derived from our Theorem 1.
Corollary 1 As discussed in [7, p. 306, Lemma 1] and [11, p. 38], if we take the
mass function γ (u) given by

0, 0 ≤ u ≤ a,
γ (u) =
1, a ≤ u ≤ 1,
where a = (1+q)
1
, q > 0, the Hausdorff matrix H reduces to Euler matrix (E, q), q >
0 and defines the corresponding (E, q) means given by
n

1 n
E qn ( f ; x) = q n−k sk ( f ; x), q > 0.
(1 + q)n k
k=0
Hence the Theorem 1 reduces to Theorem A (result proved by Nigam and Sharma
[5, p. 3, Theorem 2.1]).
Corollary 2 As discussed in [1, p. 400] and [6, p. 2747], the Cesàro matrix of order
λ, is also a Hausdorff matrix obtained by mass function γ (u) = 1 − (1 − u)λ and
the corresponding Cesàro means are given by
n

1 λ+n−k−1
Cnλ ( f ; x) = sk ( f ; x).
n + λ k=0 n−k
n
Further, Rhoades [7, p. 308] and Rhoades et al. [8, p. 6869] has mentioned that the
product of two Hausdorff matrices is again a Hausdorff matrix. Hence the Theorem
B and Theorem C (results proved by Krasniqi [4, p. 35, Theorem 2.1] and Sonker
[10, p. 126, Theorem 1]) are also particular cases of our Theorem 1.
Remark 2 This is an open problem to associate the above discussed results with the
L p -spaces.
Acknowledgments The authors express their sincere gratitude to the reviewers for their valuable
suggestions for improving the paper. This research is supported by the Council of Scientific and
Industrial Research (CSIR), New Delhi, India (Award No.- 09/143(0821)/2012-EMR-I) in the form
of fellowship to the first author.
References
1. Garabedian, H.L.: Hausdorff matrices. Am. Math. Monthly 46(7), 390–410

2. Gupta, D.P.: Degree of approximation by Cesàro means of Fourier-Laguerre expansions. Acta
Sci. Math. (Szeged) 32, 255–259 (1971)
3. Hardy, G.H.: Divergent Series. Oxford at the Clarendon Press (1949)
4. Krasniqi, X.Z.: On the degree of approximation of a function by (C, 1)(E, q) means of its
Fourier-Laguerre series. Int. J. Anal. Appl. 1, 33–39 (2013)
5. Nigam, H.K, Ajay, S.: A study on degree of approximation by (E, 1) summability means of
the Fourier-Laguerre expansion. Int. J. Math. Math. Sci. (Art. ID 351016), 7 (2010)
6. Rhoades, B.E.: Commutants for some classes of Hausdorff matrices. Proc. Am. Math. Soc.
123(9), 2745–2755 (1995)
7. Rhoades, B.E.: On the degree of approximation of functions belonging to the weighted
(L p , ξ(t)) class by Hausdorff means. Tamkang J. Math. 32(4), 305–314 (2001)
8. Rhoades, B.E., Kevser, O., Albayrak Inc.: On the degree of approximation of functions belong-
ing to a Lipschitz class by Hausdorff means of its Fourier series. Appl. Math. Comput. 217(16),
6868–6871 (2011)
9. Szegö, G.: Orthogonal polynomials. Am. Math. Soc. Colloquium Publ. 23, (1939)
10. Sonker, S.: Approximation of Functions by (C, 2)(E, q) means of its Fourier-Laguerre series.
Proceeding of ICMS-2014, (2014). ISBN:978-93-5107-261-4:125–128
11. Singh, U., Sonker, S.: Trigonometric approximation of signals (functions) belonging to
weighted (L p , ξ(t))-class by Hausdorff means. J. Appl. Funct. Anal. 8(1), 37–44 (2013)
Sparse Approximation of Overdetermined
Systems for Image Retrieval Application
M. Srinivas and R. Ramu Naidu
Abstract The recent developments in the field of compressed sensing (CS) have
been shown to have tremendous potential for applications such as content-based
image retrieval. The underdetermined framework present in CS requires some
implicit assumptions on the image database or needs the projection (or downsam-
pling) of database members into lower dimensional space. The present work, how-
ever, poses the problem of image retrieval in overdetermined setting. The main feature
of the proposed method is that it does not require any downsampling operation or
implicit assumption on the databases. Our experimental results demonstrate that our
method has potential for such applications as content-based image retrieval.
Keywords Overdetermined Systems · K-SVD · Image retrieval · LASSO ·

Underdetermined System
1 Introduction
Content-based image retrieval (CBIR) from large image databases has been an active
area of research for long due to its applications in various fields like satellite imaging,
medicine, etc. CBIR systems extract features from the raw images and calculate an
associative measure (similarity or dissimilarity) between a query image and database
images based on these features. Several CBIR systems based on wavelets, Gabor
transform have been proposed in the literature ([1] and the references therein).
In recent years, sparse representations have received a lot of attention from the
signal- and image-processing communities. Sparse coding involves representation
of an image as a linear combination of a few atoms of a given dictionary [2]. It is a
M. Srinivas (B)
Department of Computer Science and Engineering, IIT Hyderabad, Hyderabad, India
e-mail: cs10p002@iith.ac.in
R.R. Naidu
Department of Mathematics, IIT Hyderabad, Hyderabad, India
e-mail: ma11p003@iith.ac.in

DOI 10.1007/978-81-322-2485-3_17
220 M. Srinivas and R.R. Naidu
powerful tool for efficiently processing data in unconventional ways. This is mainly
due to the fact that signals and images of interest admit sparse representations in some
dictionary. The dictionaries can be composed of wavelet or Fourier basis functions or
can be learned from data. It has been observed that the dictionaries learned directly
[2–4] from data provide better representation, and hence improve the performance
on many practical applications such as classification. Several algorithms for learning
dictionaries have been developed for example, the K -SVD [5] and the method of
optimal directions (MOD) [6]. These techniques are used in many applications such
as image restoration, denoising, and texture classification.
The image retrieval or classification based on the sparse approximations typically
works under some assumptions implicitly on the databases. One often encounters
the situations such as:
• the database may not be big enough so that the sparsity promoting underdetermined
setting could be efficiently deployed
• when the classification of images is unsupervised, there is no guarantee that a
cluster has enough members, and consequently the dictionary learning involving
underdetermined setting may not be useful effectively.
One way to overcome the stated problems is to downsample the images (or project
them to lower dimensional spaces) or to extract some relevant features so that the
classification problem could be addressed in underdetermined setting.
The present work aims at proposing a novel CBIR algorithm by posing the problem
in the form of an overdetermined framework. The salient features of our method are
twofold: (1) even if the database is relatively smaller in size and images are bigger
in size, the method could be useful and (2) the method per se does not require
downsampling of images.
We realize our objective using the LASSO [7] at two stages. In the first step, we
identify the most relevant clusters of the database by finding sparse approximation to
the system q ≈ Φx. Where q is query image and Φ is the matrix whose columns are
the cluster centers of the data. While in the second step, we obtain the desired retrieval
performance by obtaining sparse approximation to the overdetermined system q ≈
Ψ y, where Ψ is the matrix whose columns are the images belonging to the relevant
clusters.
The paper is organized into several sections. In Sects. 2 and 3, we present, respec-
tively, sparse approximations to overdetermined systems and motivation for the
present work in more detail. While in Sects. 4 and 5, we present image retrieval
through overdetermined systems and simulation results, respectively. In the last
section, we provide our concluding remarks.
2 Sparse Approximation to Overdetermined Linear System
The focus in this paper is on systems which can be represented in the form y ≈ Ax,
where the dimensions of the objects A, y, x are m ×n, m ×1, n ×1, respectively. It is
Sparse Approximation of Overdetermined Systems for Image Retrieval Application 221
assumed that m > n, which is typically the case in our image retrieval problem. When
y comes from query image and A is generated using database members consisting of
members relevant to y, one may suspect that the system y has sparse representation
in A. We can obtain such a sparse approximation using the following optimization
problem:
Au k − y2 = inf{Ax − y2 | x ∈ R n , x0 ≤ k},
where x0 = |{i|xi = 0}|, the number of nonzero components in x. The vector
u k is called as a best k-sparse approximation to y ≈ Ax and it contains atmost k
nonzero terms.
A simple but computationally costly approach to solve this problem is by using
brute force search, that is, randomly pick (n − k) elements of x to be zero and find
the remaining using the standard least squares method. It can be easily seen that this
method becomes intractable for high values of n.
The main culprit in the problem is the l0 norm which makes the problem compu-
tationally costly. Tibshirani [7] made use of the l1 norm in the place of the l0 norm
and solved the following (modified) problem (for a fixed λ):-
find xsuch that Ax − y22 + λx1 is minimized.
The main ideas behind using the l1 norm are:

• convex optimization methods can be used to solve the above-modified problem
• minimizing l1 norm typically provides sparse solutions
The minimizing algorithm called the LASSO is obtained by solving a series of
quadratic programming problems [7]. Usually, the parameter λ is obtained by cross-
validation.
3 Motivation for Present Work
The recent sparsity-based methods are found to be useful for applications in image
processing [2, 3, 8]. In this section, we quickly review the relevant methods and
present our motivation for the current work.
Given sufficient training samples of ith class, Ai (whose columns are samples)
for i = 1, 2, . . . K , it is shown in [8] that the class label of unknown object y may
be obtained by solving
y = [A1 |A2 | . . . |A K ] x0 (1)

A
for sparse solution, which is obtained from
x0 = arg min α1 subject to Aα = y. (2)

α
The index î, defined by
î = arg min y − Aδi (x0 )22 , (3)

i=1,2,··· ,K
is taken as the class label of y. Here, δi is a characteristic function that selects the
coefficients associated with the ith labeled samples. This method being supervised
provides excellent classification1 provided the labeled data for each class are suffi-
cient. The dictionary-based methods, however, train class-specific dictionaries using
labeled data and then assign each testing image to the class for which the best recon-
struction is obtained [2], which rely on the premise that two signals belonging to
the same cluster have decomposition in terms of similar atoms. After identifying the
class label, one may retrieve most relevant images from the ith class using some
similarity metric.
These methods are effective when the database and labeled data for each class are
sufficiently big enough to accommodate underdetermined framework. In the absence
of the database or class-specific labeled data being sufficiently big enough; however,
one may need to consider features of images or downsampled images or random
projections of images into lower dimensional spaces meeting the theoretical restric-
tions posed by the Lemma [9]. The reduced dimension of data may have bearing on
the retrieval performance. In addition, when the classification is unsupervised, sparse
recovery based on underdetermined setting may not always be useful. This is because
a cluster may not have enough members. Motivated by these considerations, we con-
sider proposing an unsupervised method that involves the sparse approximations to
overdetermined systems. We believe the overdetermined setting could involve very
little restriction on the database being used.
4 Overdetermined Setting for Image Retrieval
Inspired by the ideas from [8], one may pose the retrieval problem in overdetermined
setting as
q ≈ [I1 I2 . . . I N ] α0 , (4)

Φ
where, for j = 1, 2, . . . , N , I j is vectorized version of jth database image and q is

query image. The problem of retrieving the L best matches of query image q may
be obtained from
α0L = arg min {Φx − q2 | x0 ≤ L}, (5)

x∈R N
1 Viewingthe problem of image retrieval as an extension of classification, in this paper, we use the
words “classification” and “retrieval” synonymous. This is of course a slight abuse of convention.
which could be solved using LASSO. For large databases consisting of somewhat
bigger images, Φ has bigger size and the computational cost in using Φ could be
more, making thereby the method less useful. In view of this problem, instead of
searching for the relevance in the entire database, we search for the relevance of q to
few clusters, which is followed by further search within the relevant clusters through
LASSO again.
After dividing the database into K clusters, namely C1 , C2 , . . . , C K with cluster
centers e1 , e2 , . . . , e K , we identify the K 1 most relevant clusters from
e K 1 = arg min { [e1 e2 . . . e K ] x K − q2 | x K 0 ≤ K 1 }, (6)

x K ∈R K
Ψ
The number K 1 may be heuristically determined as the significant number of

cluster centers in whose spanning space q results in minimum misfit. Suppose
Ci1 , Ci2 , . . . , Ci K 1 ∈ {Ci }i=1
K are the clusters whose indices are those of the nonzero
entries of e K 1 . We then search for the L most relevant images within the pruned
clusters from
θ L = arg min { [Ci1 Ci2 . . . Ci K 1 ] θ − q2 | θ 0 = L}, (7)

θ∈R L 1
Γ
K1
In the above equation, L 1 := l=1 |Cil |, which is the column size of Γ . We determine
the degree of relevance of retrieved images by computing the error:
errorrel = min q − Γ δi (θ L )22 . (8)

i∈support{θ L }
The images that result in small errors are considered to be more relevant. The
distribution of data into several clusters is in general unknown. If the images within
each cluster are closer to cluster center and the cluster centers are wide apart, the
proposed method is expected to work very well. The block diagram of the method
is shown in Fig. 1.
Fig. 1 Block diagram of proposed method

5 Simulation Results
In this section, we demonstrate the performance of the proposed overdetermined

(OD) method on Image Retrieval in Medical Applications (IRMA) medical image
database2 and compare our results to the ones obtained by the underdetermined
(UD) setting as stated in (4). In implementing the UD-based method, we use the
orthogonal matching pursuit (OMP) [3] algorithm. We analyze the performances of
both methods using the standard precision and recall, which are defined as
Number of relevant images retrieved

Precision =
Total number of images retrieved
Number of relevant images retrieved

Recall =
Total number of relevant images
To begin with, we execute our method on a small medical database wherein each
image is of size 120×120. This database consists of 311 images of skull, breast, chest,
hand, etc. Of them, we consider 11 as testing images. As in (4), if we form Φ with
database members as columns, we get a matrix whose size is 14,400 × 300, a very
tall and slim matrix that occupies huge memory space on computer and involves high
computational cost. One may reduce the row dimension of this matrix by projecting
the images to lower dimension spaces. But the dimension of reduced size in OD
case is not dictated by the resulting size as it is in the UD case. To be able to apply
the powerful theory of UD framework, one needs to downsample (or project) the
database members to a space of dimension less than 300, which can impact retrieval
performance. The dimension of projected space could be further small if one wishes
to use dictionary-based approach [4].
Using K-means algorithm, we divided the database into 10 clusters (that is, K =
10). It is to be mentioned here that the bigger value for K might slightly increase
computational complexity, but it can have no significant bearing on performance. This
is because we consider more than one cluster when searching for relevant images.
We form the matrix system (6) and obtain sparse approximation with K 1 = 3. This
selection is based on the observation that the choice of K 1 = 3 is good enough to
result relevant images in three clusters, which is supported by the plots in Figs. 2 and
3. In our simulation work, in order to deal with overdetermined matrix of reasonable
size, we have downsampled the database members to 60 × 60. Therefore, the size
of Φ becomes 3600 × 300. In UD setting, we have projected images to a space of
lower dimension 256 (the associated matrix is of size 256 × 300). As stated already,
this downsampling operation is not at all mandatory in OD case while it is necessary
in UD case. The plots in Figs. 2 and 3 show that the average performance by OD
method is better than that given by UD method.
2 www.irma-project.org.
Fig. 2 Precision–Recall graph with K 1 = 3 on the medical database
Fig. 3 Precision–Recall graph with K 1 = 5 on the medical database
Using the clusters whose indices correspond to the nonzero components of e K 1

(defined in (6)), we form the matrix Γ , and solve (7) for θ L . From the nonzero
locations of θ L , we obtain the best matches from the database for a given query
image. The performance of our method is shown in Fig. 4 for several query images.
The precision–recall plots for the cases of K 1 = 3 and K 1 = 5 are shown in Figs. 2
Fig. 4 Retrieval performance of the proposed method. The first image on each row is the query
image and the remaining correspond to those retrieved by the method. The upper and lower parts of
the figure, respectively, correspond to the performances of over- and underdetermined frameworks
and 3, respectively. In Fig. 4, on every row, the first image refers to the query image,
the next five images correspond to the retrieval performance of OD setting, while the
remaining five correspond to the retrieval by UD-based framework.
6 Conclusions
The present work has proposed an overdetermined framework for CBIR. The moti-
vation for the present work comes from the fact that the sparsity promoting clas-
sification methods that involve the use of underdetermined matrix equations work
on some implicit assumptions on the databases or project data into lower dimen-
sion spaces to accommodate the ideas from the theory of compressed sensing. The
present work, however, does not need any such requirement. The preliminary simu-
lation results reported in this paper demonstrate that the overdetermined framework
has potential for image retrieval problems. The medical database members contain
some scale variation. The retrieval performance on these databases can be improved
by incorporating rotation and scale invariant features in the retrieval process. Our
future efforts shall address this aspect.
Acknowledgments Authors are thankful to Dr. C. Krishna Mohan, Dr. Phanindra Jampana and Dr.
C.S. Sastry for fruitful discussions. Authors would like to thank Dr. T.M. Deserno, Department of
Medical Informatics, RWTH Aachen, Germany for making the original IRMA Database available
for research purposes.
References
1. Chen, Y., Sastry, C.S., Patel, V., Philips, J., Chellappa, R.: In-Plane rotation and scale invariant
simultaneous dictionary learning and clustering. IEEE Trans. Image Process. 22(6), 2166–2180
(2013)
2. Patel, V.M., Wu, T., Biswas, S., Philips, P.J., Chellappa, R.: Dictionary-based face recognition
under variable lighting and pose. IEEE Trans. Inform. Forensics Security 7(3), 954–965 (2012)
3. Elad, M.: Sparse and Redundant Representations. Springer, New York (2010)
4. Sprechmann, P., Sapiro, G.: Dictionary learning and sparse coding for unsupervised clustering.
In Proceedings of ICASSP (2010)
5. Aharon, M., Elad, M., Bruckstein, A.M.: The k-svd: an algorithm for designing of over-complete
dictionaries for sparse representation. IEEE Trans. Signal Process. 54(11), 4311–4322 (2006)
6. K-Delgado, K., Murray, J.F., Rao, B.D., Engan, K., Lee, T.W., Sejnowski, T.J.: Dictionary
learning algorithms for sparse representation. Neural Comput. 15, 349–396 (2003)
7. Tibshirani, R.: Regression shrinkage and selection via the LASSO. J. Statis. Soft. 33(1), 1 (2010)
8. Wright, J., Yang, A.Y., Ganesh, A., Sastry, S.S., Ma, Y.: Robust face recognition via sparse
representation. IEEE Trans. Pattern Anal. Mach. Intell. 31(2), 210–227 (2009)
9. Baraniuk, R., Davenport, M., DeVore, R., Wakin, M.: A simple proof of the restricted isometry
property for random matrices. Constr. Approximation 28(3), 253–263 (2008)
kth Order Kantorovich Modification
of Linking Baskakov-Type Operators
Margareta Heilmann and Ioan Raşa
Abstract In 1957 Baskakov introduced a general method for the construction of

positive linear operators depending on a real parameter c. The so-called genuine
Baskakov–Durrmeyer-type operators form a class of operators reproducing the lin-
ear functions, interpolating at (finite) endpoints of the interval, and having other
nice properties. In this paper we consider a nontrivial link between Baskakov-type
operators and genuine Baskakov–Durrmeyer-type operators. We establish explicit
representations for the images of monomials and for the moments; they are useful,
e.g., in studying asymptotic formulas.
Keywords Baskakov-and-Durrmeyer-type operators · Linking operators ·

Kantorovich-type modifications · Moments · Images of monomials
2010 AMS Subject Classification: 41A36 · 41A10 · 41A60
1 Introduction and Definition of the Operators
In 1957 Baskakov [1] introduced a general method for the construction of positive
linear operators depending on a real parameter c including the classical Bernstein,
Szász-Mirakjan, and Baskakov operators as special cases. All these Baskakov-type
operators preserve linear functions and interpolate at (finite) endpoints of the cor-
responding interval. The so-called Bernstein–Durrmeyer operators were introduced
by Durrmeyer in [2] and independently developed by Lupaş [9]. Afterwards, this
construction was carried over to many other classical operators; for instance see
M. Heilmann (B)
Faculty of Mathematics and Natural Sciences, University of Wuppertal,
Gaußstraße 20, Wuppertal 42119, Germany
e-mail: heilmann@math.uni-wuppertal.de
I. Raşa
Department of Mathematics, Technical University,
Str. Memorandumului 28, Cluj-napoca 400114, Romania
e-mail: Ioan.Rasa@math.utcluj.ro
DOI 10.1007/978-81-322-2485-3_18
230 M. Heilmann and I. Raşa
[10, 16] and in the general setting for so-called Baskakov–Durrmeyer-type opera-
tors [6]. These operators have a lot of nice properties; they commute, they commute
with certain differential operators, they are self-adjoint but they only reproduce con-
stants.
The consideration of so-called genuine Baskakov–Durrmeyer-type operators
leads to a class of operators again reproducing the linear functions and interpo-
lating at (finite) endpoints of the corresponding interval. These operators are related
to the Baskakov–Durrmeyer-type operators in the same way as the Baskakov-type
operators to their corresponding Kantorovich variants.
In [11, 12] Păltănea introduces operators depending on a parameter ρ ∈ R+ , which
constitute a nontrivial link between the Bernstein and Szász-Mirakjan operators,
respectively, and their genuine Durrmeyer modifications. Further results can also be
found in [3, 4, 13].
In this paper we consider a nontrivial link between Baskakov-type operators
and genuine Baskakov–Durrmeyer-type operators. Moreover, we investigate the kth
order Kantorovich modification of them; for k = 1 this means a link between the
Kantorovich modification of Baskakov-type and Baskakov–Durrmeyer-type opera-
tors.
In what follows for c ∈ R we use the notations

j−1
j−1
a c, j := (a + cl), a c, j := (a − cl), j ∈ N; a c,0 = a c,0 := 1
l=0 l=0
which can be considered as a generalization of rising and falling factorials. Note that
a −c, j = a c, j and a c, j = a −c, j . This notation enables us to state the results for the
different operators in a unified form.
In a recent paper [8] we already considered the linking operators between the
kth order Kantorovich modification of the Bernstein and the genuine Bernstein–
Durrmeyer operators. Comparison of the results in [8] with the outcomes of the
present paper shows that all the representations for the moments and the images of
monomials are also valid for the Bernstein case by setting c = −1 in the subsequent
theorems.
In the following definitions of the operators we omit the parameter c in the nota-
tions in order to reduce the necessary sub- and superscripts.
Let c ∈ R, c ≥ 0, n ∈ R, n > c, ρ ∈ R+ , j ∈ N0 , x ∈ [0, ∞). Then the basis
functions are given by
nj
j! x j e−nx , c = 0,
pn, j (x) = n c, j −( nc + j )
j! x j (1 + cx) , c > 0.
In the following definition we assume that f : [0, ∞) −→ R is given in such a way

that the corresponding integrals and series are convergent.
kth Order Kantorovich Modification … 231
Definition 1 The operators of Baskakov type are defined by

∞

j
Bn ( f, x) = pn, j (x) f , (1)
n
j=0
the genuine Baskakov–Durrmeyer-type operators are denoted by

∞
∞
Bn,1 ( f, x) = f (0) pn,0 (x) + pn, j (x) pn+2c, j−1 (t) f (t)dt, (2)
j=1 0
and for ρ ∈ R+ the linking operators are given by

∞
ρ
Bn,ρ ( f, x) = Fn, j ( f ) pn, j (x) (3)
j=0
∞
∞
ρ
= f (0) pn,0 (x) + pn, j (x)(n + c) μn, j (t) f (t)dt, (4)
j=1 0
where
⎧
⎨ (nρ) jρ jρ−1 −nρt
, c = 0,
ρ Γ ( jρ) t e
μn, j (t) =
(1 + ct)−( c + j )ρ−1
c jρ n
⎩ t jρ−1 , c > 0.
B ( jρ, nc ρ+1)
Setting c = 0 in (2) leads to the Phillips operators [14], c > 0 was investigated in
[18]. To the best of our knowledge the case c = 0 in (3) was first considered in [12].
As in [8] for the Bernstein case we also consider the kth order Kantorovich
modification of the operators Bn,ρ , i.e.,
(k)
Bn,ρ := D k ◦ Bn,ρ ◦ Ik (5)
where D k denotes the kth order ordinary differential operator and

x (x − t)k−1
Ik f = f, if k = 0, and Ik ( f, x) = f (t)dt, if k ∈ N.
0 (k − 1)!
For k = 0 we omit the superscript (k) as indicated by the definition above.

This general definition contains many known operators as special cases. For
c = 0 we get the linking operators considered in [13]. For ρ = 1 we get the
genuine Baskakov–Durrmeyer-type operators Bn,1 , for ρ = 1, k ∈ N the Baskakov–
(1)
Durrmeyer-type operators Bn,1 (see [6, (1.3)], named Mn+c there) and the auxiliary
(k)
operators Bn,1 considered in [7, (3.5)], (named Mn+c,k−1 there) with the explicit
representation
∞ ∞
(k) n c,k
(Bn,1 f )(x) = p n+ck, j (x) pn−c(k−2), j+k−1 (t) f (t)dt.
n c,k−1 j=0 0
For an arbritrary sequence of linear operators, the images of monomials and

the moments are important, e.g., in studying the asymptotic behavior. In this paper
we establish explicit representations for the images of the monomials and for the
moments of the investigated operators. Corresponding recursion formulas and further
results will be given in a forthcoming paper.
Below we will use the following basic formulas.
∞
ρ n
μn, j (t)dt = B jρ, ρ + 1 , (6)
0 c
∞

pn, j (x) = 1, (7)
j=0
j
pn, j (x) = x pn+c, j−1 (x), (8)
n

x(1 + cx) pn, j (x) = ( j − nx) pn, j (x), (9)
with the convention pn,l (x) = 0, if l < 0. As usual, empty products are defined to
be one.
2 Explicit Formulas for the Images of Monomials
In this section we prove general explicit formulas for the images of the monomials of
(k)
the operators Bn,ρ . In what follows we denote by eν (t) = t ν , ν ∈ N0 , the monomials
and by
l
l
lh f (x) = (−1)l−κ f (x + κh) (10)
κ
κ=0
the lth order forward difference of a function f with step h and define

ν−1
l

pνρ (ξ) := ξ+ , ν ∈ N.
ρ
l=1
This can be rewritten as

ν−1 i ρ
1 pν (1)
i
pνρ (ξ) = (ξ − l), (11)
i!
i=0 l=1
which can be derived by using the Newton representation of the interpolation poly-
ρ
nomial of pν for the equidistant knots 1, 2, . . . , ν.
We first consider the images of the monomials for the case k = 0, i.e., for the
operators Bn,ρ .
Theorem 1 Let n ∈ R, nρ > c(ν − 1), ρ ∈ R+ , ν ∈ N0 , ν ≤ n. Then
(Bn,ρ e0 )(x) = 1, (12)

ρν
ν
n c,i
ρ
(Bn,ρ eν )(x) = i−1 p ν (1) x i , ν ∈ N. (13)
(nρ)c,ν (i − 1)! 1
i=1
Proof (12) follows immediately from (6) and (7).

In order to prove (13) we take into account that for c = 0
ν−1
(nρ) jρ ∞
ν jρ−1 −nρt 1 Γ ( jρ + ν) 1 l
t t e dt = · = ν j+
Γ ( jρ) 0 (nρ)ν Γ ( jρ) n ρ
l=0
and for c > 0

∞
c jρ
t ν t jρ−1 (1 + ct)−( c + j )ρ−1 dt
n

B jρ, nc ρ + 1 0
n ν−1
−ν Γ ( jρ + ν)Γ ρ+1−ν ρν l
=c nc = j+ .
Γ ( jρ)Γ cρ + 1
(nρ)c,ν ρ
l=0
Thus we get for ν ≥ 1 with (8) and (11)
ρν
∞
ν−1
l

(Bn,ρ eν )(x) = p n, j (x) j + (14)
(nρ)c,ν ρ
j=1 l=0
∞
ρν
= nx pn+c, j−1 (x) pνρ ( j)
(nρ) c,ν
j=1
∞ i pν (1) ν−1 ρ i
ρν
= nx p n+c, j−1 (x) 1
( j − l)
(nρ)c,ν i!
j=1 i=0 l=1
ν−1
ρ ∞

i
ρν i1 pν (1)
= nx pn+c, j−1 (x) ( j − l).
(nρ)c,ν i!
i=0 j=i+1 l=1
Applying (8) for j ≥ i + 1 we have

i
i
pn+c, j−1 (x) ( j − l) = pn+c(i+1), j−i−1 (x)x i (n + cl).
l=1 l=1
Hence with (7)
ρν n c,i i−1 ρ i
ν
(Bn,ρ eν )(x) = 1 pν (1) x .
(nρ)c,ν (i − 1)!
i=1
Remark 1 Using (10), the representation (13) can be rewritten as
ν
ρν c,i i
i−1
1
(Bn,ρ eν )(x) = n x (−1)i−1−κ p ρ (1 + κ).
(nρ) c,ν κ!(i − 1 − κ)! ν
i=1 κ=0
Now we consider the special cases ρ = 1 and ρ → ∞.

ρ = 1: Then with [5, (3.48)] (see [8, p. 323])

ν
i−1
1 pν (1) = (ν − 1)!
1
.
i
Thus
ν
1 (ν − 1)! ν i
(Bn,1 eν )(x) = n c,i x ,
n c,ν (i − 1)! i
i=1
which coincides with the formula given in [18, Lemma 1.11] and [7, (4.3)] with
s = −1 and taking n + c instead of n there.
ρν 1
ρ → ∞: Then → ν , and (see [8, p. 323])
(nρ)c,ν n
∞
i−1
1 pν (1) = (i − 1)!σν ,
i
j
where σν denote the Stirling numbers of second kind. Thus
ν
1 c,i i i
(Bn,∞ eν )(x) = n σν x ,
nν
i=1
which coincides with the corresponding result for the classical Baskakov-type oper-
ators which can be calculated directly from the definition of the operators by
using (8).
Next, we consider the images of the monomials for the case k ∈ N.
Theorem 2 Let n ∈ R, k ∈ N, ρ ∈ R+ , ν ∈ N0 , nρ > c(ν + k − 1). Then

(k)
(Bn,ρ eν )(x)
ν!
ν
ρν+k n c,i+k
i+k−1 ρ
= (i + k) p ν+k (1) xi . (15)
(ν + k)! (nρ)c,ν+k i! 1
i=0
(k) ν!
Proof By using Bn,ρ eν = k
(ν+k)! D Bn,ρ eν+k we get from (13) for k ∈ N
(k)
(Bn,ρ eν )(x)
ρν+k n c,i i−1 ρ
ν+k
ν! i!
= p ν+k (1) x i−k
(ν + k)! (nρ) c,ν+k (i − 1)! 1
(i − k)!
i=k
ν! ρν+k ν
n c,i+k
ρ
= (i + k) i+k−1 p ν+k (1) xi .
(ν + k)! (nρ)c,ν+k i! 1
i=0
Remark 2 Using again (10), the representation (15) can be rewritten as
ν
(k) ν! ρν+k c,i+k (i + k)! i
(Bn,ρ eν )(x) = n x
(ν + k)! (nρ)c,ν+k i!
i=0

i+k−1
1 ρ
× (−1)i+k−1−κ p (1 + κ).
κ!(i + k − 1 − κ)! ν+k
κ=0
Again we consider the special cases ρ = 1 and ρ → ∞.

ρ = 1: Then again with [5, (3.48)]

ν+k
i+k−1 1
pν+k (1) = (ν + k − 1)! .
1
i +k
Thus
ν

(k) 1 (ν + k − 1)! ν i
(Bn,1 eν )(x) = n c,i+k x .
n c,ν+k (i + k − 1)! i
i=0
This coincides with the corresponding result in [7, Satz 4.2] for the auxiliary operators
(k)
with the notation Bn,ρ = Mn+c,k−1 there.
ρν+k 1
ρ → ∞: Then → ν+k and
(nρ) c,ν+k n
∞
i+k−1
1 pν+k i+k
(1) = (i + k − 1)!σν+k .
Thus
ν
(k) ν! 1 n c,i+k i+k i
(Bn,∞ eν )(x) = (i + k)!σν+k x .
(ν + k)! n ν+k i!
i=0
From the explicit representations of the images of the monomials we can deduce the
(k)
following result concerning the limit of the operators Bn,ρ when ρ → ∞.
Corollary 1 For each polynomial p we have

(k)
lim Bn,ρ p(x) = Bn(k) p(x)
ρ→∞
uniformly on every compact subinterval of [0, ∞).

(k)
For the evaluation of Bn,ρ eν , k ∈ N, for special values of ν, we use the represen-
tation
ν+k−1

ρ
pν+k (ξ) = ρ−l σl (1, 2, . . . , ν + k − 1)ξ ν+k−1−l ,
l=0
with the notation σ j (x0 , x1 , . . . , xn ), j ∈ N, for the symmetric function which

is the sum of all products of j distinct values from the set {x0 , x1 , . . . xn } and
σ0 (x0 , x1 , . . . , xn ) := 1.
For the monomial em , it is known (see, e.g., [15, Theorem 1.2.1]) that

j+k−1 0, m < j + k − 1,
1 em (1) =
( j + k − 1)!τm−( j+k−1) (1, 2, . . . , j + k), 0 ≤ j + k − 1 ≤ m,
with the complete symmetric function τ j (x0 , x1 , . . . , xn ) which is the sum of all
products of x0 , x1 , . . . , xn of total degree j, j ∈ N, and τ0 (x0 , x1 , . . . , xn ) := 1.
(k)
Thus we can rewrite (Bn,ρ eν ) as
ν
(k) ν! ρν+k n c,i+k (i + k)! i
(Bn,ρ eν )(x) = x (16)
(ν + k)! (nρ)c,ν+k i!
i=0
ν−i

× ρ−l σl (1, 2, . . . , ν + k − 1)τν−l− j (1, 2, . . . , i + k).
l=0
As a corollary we present the results for ν = 0, 1, 2.

Corollary 2 For k ∈ N0 the images for the first monomials are given by
(k) ρk
(Bn,ρ e0 )(x) = · n c,k ,
(nρ)c,k

(k) ρk+1 c,k 1 1
(Bn,ρ e1 )(x) = · n k 1 + + (n + ck)x ,
(nρ)c,k+1 2 ρ

(k) ρk+2 c,k 1 3k + 1 k + 1 3k + 5
(Bn,ρ e2 )(x) = ·n k + +
(nρ)c,k+2 2 6 ρ 6ρ2

1
+(n + ck) (k + 1) 1 + x + (n + c(k + 1))x 2
.
ρ
Proof For k = 0 the identities follow from Theorem 1. For k ∈ N we derive the
proposition by using the representation (16) and the fact that for m ∈ N
σ0 (1, . . . , m) = τ0 (1, . . . , m) = 1,
1
σ1 (1, . . . , m) = τ1 (1, . . . , m) = m(m + 1),
2
1
σ2 (1, . . . , m) = (m − 1)m(m + 1)(3m + 2),
24
1
τ2 (1, . . . , m) = m(m + 1)(m + 2)(3m + 1).
24
(k)
In the following theorem we state a representation of Bn,ρ eν in terms of the images
(k)
of monomials of the operators Bn . This underlines the close relationship beween
(k)
the linking operators Bn,ρ and the kth order Kantorovich modification of the classical
operators Bn .
(k)
Theorem 3 The images of the monomials under Bn,ρ can be expressed as
ν

(k) ν! 1 i+k (i + k)! (k)
(Bn,ρ eν )(x) = sν+k (ρn)i+k (Bn ei )(x), k ∈ N0 ,
(ν + k)! (nρ)c,ν+k i!
i=0
i+k
where sν+k denote the Stirling numbers of first kind.
Proof For ν ∈ N and k = 0 we derive from (14)
∞ ν−1
1
(Bn,ρ eν )(x) = p n, j (x) ( jρ + l)
(nρ)c,ν
j=1 l=0
ν
∞
i
1 j
= sνi (ρn)i pn, j (x)
(nρ)c,ν n
i=0 j=1
ν

1
= sνi (ρn)i (Bn ei )(x).
(nρ)c,ν
i=0
(k) ν!
For k ∈ N the conclusion follows by using (Bn,ρ eν ) = (ν+k)! D (Bn,ρ eν+k )
k and
(k)
D k (Bn ei ) = (i−k)! (Bn ei−k ),
i!
respectively.
For the case k = 0 a corrresponding result for the Bernstein operators can be
found in [17, Theorem 3.2.1].
3 Explicit Formulas for the Moments
(k)
Next, we consider the moments of Bn,ρ and Bn,ρ . For abbreviation, we use the
notation
(k) (k)
Mn,ρ,m (x) = Bn,ρ (e1 − xe0 )m (x), m ∈ N0 , x ∈ [0, ∞) (17)
where we again omit the superscript (k) in case k = 0. We use the fact that
m
(k) m (k)
Mn,ρ,m (x) = (−x)m−ν (Bn,ρ eν )(x).
ν
ν=0
Again, we first treat the case k = 0.
Theorem 4 Let n ∈ R, ρ ∈ R+ , m ∈ N0 , nρ > c(m − 1). Then
Mn,ρ,0 (x) = 1, (18)

Mn,ρ,1 (x) = 0, (19)

m
i
ρν+m−i ν m
Mn,ρ,m (x) = (−x)m + (−x)i (−1) (20)
(nρ)c,ν+m−i i −ν
i=1 ν=1
n c,ν ρ
× ν−1 pν+m−i (1), m ≥ 2.
(ν − 1)! 1
Proof Equations (18) and (19) follow immediately from Corollary 2.

In order to prove (20) we apply Theorem 1. With the index transform i → i − m +
ν, changing the order of summation and applying the index transform ν → ν +m −i,
we derive
Mn,ρ,m (x)
m
ρν n c,i i−1 ρ i
ν
m
= (−x) +
m
(−x) m−ν
1 pν (1) x
ν (nρ)c,ν (i − 1)!
ν=1 i=1
m
m ρν
= (−x)m + (−1)m−ν
ν (nρ)c,ν
ν=1

m
n c,i−m+ν
ρ
× i−m+ν−1 p ν (1) xi
(i − m + ν − 1)! 1
i=m−ν+1

m
m
m ρν
= (−x) + m
x i
(−1)m−ν
ν (nρ)c,ν
i=1 ν=m+1−i
n c,i−m+ν
ρ
× i−m+ν−1 p ν (1)
(i − m + ν − 1)! 1

m i

m ρν+m−i
= (−x) +
m
(−x)i
(−1)ν
i −ν (nρ)c,ν+m−i
i=1 ν=1
n c,ν
ρ
× ν−1 p ν+m−i (1) .
(ν − 1)! 1
Remark 3 Analogously as for the images of monomials, (20) can be rewritten as

m
i
ρν+m−i m
Mn,ρ,m (x) = (−x)m + (−x)i n c,ν
(nρ)v,ν+m−i i −ν
i=1 ν=1
ν−1
1 ρ
× (−1)κ+1 p (1 + κ).
κ!(ν − 1 − κ)! ν+m−i
κ=0
Next, we consider the special cases ρ = 1 and ρ → ∞.

ρ = 1: With [5, (3.48)]

ν +m −i
ν−1 1
pν+m−i (1) = (ν + m − i − 1)! .
1
ν
we get
m
m!
i
n c,ν
Mn,ρ,m (x) = (−x)m + (−x)i (−1)ν c,ν+m−i
i! n
i=1 ν=1

i ν +m −i −1
× ,
ν ν−1
which coincides with the result in [18, Korollar 1.12] and with [7, Korollar 4.4] with
s = −1 and n + c instead of n there.
ρν+m−i 1
ρ → ∞: Then → ν+m−i and
(nρ) c,ν+m−i n
ν−1
1
∞
pν+m−i ν
(1) = (ν − 1)!σν+m−i .
Thus

m
i
n c,ν m
Mn,∞,m (x) = (−x)m + (−x)i (−1)ν σν+m−i
ν
.
n ν+m−i i −ν
i=1 ν=1
In our next theorem, we evaluate the moments for the case k ∈ N.
Theorem 5 Let n ∈ R, ρ ∈ R+ , k ∈ N, m ∈ N0 , nρ > c(m + k − 1). Then

m
i
(k) ρν+m−i+k ν m
Mn,ρ,m (x) = (−x)i (−1) (21)
(nρ)c,ν+m−i+k i −ν
i=0 ν=0
(ν + m − i)! (ν + k) c,k+ν ν+k−1 ρ
× n 1 pν+m−i+k (1).
(ν + m − i + k)! ν!
Proof The result can be proved by using Theorem 2 and carrying out the same steps
as in the proof of Theorem 4.
Remark 4 With (10), we can rewrite the representation (21) as
(k)
Mn,ρ,m (x)

ρν+m−i+k
m i
m (ν + m − i)!
= (−x)i
i=0 ν=0
(nρ)c,ν+m−i+k i − ν (ν + m − i + k)!
ν+k−1
(ν + k)! 1 ρ
× n c,ν+k (−1)k+1+κ p (1 + κ).
ν! κ!(ν + k − 1 − κ)! ν+m−i+k
κ=0
From Theorem 5 we derive the following identity for the special cases ρ = 1 and
ρ → ∞.
ρ = 1: With [5, (3.48)] we have

ν +m −i +k
1ν+k−1 pν+m−i+k
1
(1) = (ν + m − i + k − 1)! .
ν+k
Thus

m!
m i
(k) n c,ν+k i ν +m −i +k −1
Mn,1,m (x) = (−x)i (−1)ν c,ν+m−i+k .
i! n ν ν+k−1
i=0 ν=0
This coincides with the result [7, Korollar 4.4] for the moments of the auxiliary
operators named Mn+c,k−1 there.
ρν+m−i+k 1
ρ → ∞: Then → ν+m−i+k and
(nρ) c,ν+m−i+k n
1ν+k−1 pν+m−i+k
∞ ν+k
(1) = (ν + k − 1)!σν+m−i+k .
Thus
(k)
Mn,∞,m (x)

m
i
n c,ν+k m (ν + m − i)!(ν + k)! ν+k
= (−x)i (−1)ν σ .
n ν+m−i+k i − ν (ν + m − i + k)!ν! ν+m−i+k
i=0 ν=0
With the same notations and arguments used for Corollary 2, the moments (20)
and (21) can be computed by using
ρ
1ν+k−1 pν+m− j+k (1)
j
m−
= (ν + k − 1)! ρ−l σl (1, 2, . . . , ν + m − j + k − 1)τm− j−l (1, 2, . . . , ν + k).
l=0
Corollary 3 For k ∈ N0 the first moments are given by

(k) ρk (k) ρk+1 c,k 1 1
Mn,ρ,0 (x) = n c,k
, M (x) = n k 1 + (1 + 2cx),
(nρ)c,k n,ρ,1
(nρ)c,k+1 2 ρ

(k) ρk+2 1 1
Mn,ρ,2 (x) = n c,k
1 + n + c 1 + k(k + 1) x(1 + cx)
(nρ)c,k+2 ρ ρ

k 1 3k + 5
+ (3k + 1) 1 + + .
12 ρ ρ
References
1. Baskakov, V.A.: An instance of a sequence of positive linear operators in the space of continuous
functions. Dokl. Akad. Nauk SSSR 113(2), 249–251 (1957)
2. Durrmeyer, J.L.: Une formule d’inversion de la transformée de Laplace: applications à la théorie
des moments, Thèse de 3e cycle, Faculté des Sciences de l’Université de Paris, 1967
3. Gonska, H., Păltănea, R.: Simultaneous approximation by a class of Bernstein-Durrmeyer
operators preserving linear functions. Czechoslovak Math. J. 60(135), 783–799 (2010)
4. Gonska, H., Păltănea, R.: Quantitative convergence theorems for a class of Bernstein-
Durrmeyer operators preserving linear functions. Ukrainian Math. J. 62(7), 1061–1072 (2010)
5. Gould, H.W.: Combinatorial Identities: A Standardized Set of Tables Listing 500 Binomial
Coefficient Summations. Henry W. Gould, Morgantown (1972)
6. Heilmann, M.: Direct and converse results for operators of Baskakov-Durrmeyer operators.
Approx. Theory Appl. 5(1), 105–127 (1989)
7. Heilmann, M.: Erhöhung der Konvergenzgeschwindigkeit bei der Approximation von Funk-
tionen mit Hilfe von Linearkombinationen spezieller positiver linearer Operatoren, Habilita-
tionschrift Universität Dortmund (1992)
ρ
8. Heilmann, M., Raşa, I.: k-th order Kantorovich type modification of the operators Un . J. Appl.
Funct. Anal. 9(3–4), 320–334 (2014)
9. Lupaş, A.: Die Folge der Betaoperatoren, Dissertation, Universität Stuttgart (1972)
10. Mazhar, S.M., Totik, V.: Approximation by modified Szász operators. Acta Sci. Math. 49,
257–269 (1985)
11. Păltănea, R.: A class of Durrmeyer type operators preserving linear functions. Ann. Tiberiu
Popoviciu Semin. Funct. Equ. Approx. Convexity (Cluj-Napoca) 5, 109–117 (2007)
12. Păltănea, R.: Modified Szász-Mirakjan operators of integral form. Carpathian J. Math. 24(3),
378–385 (2008)
13. Păltănea, R.: Simultaneous approximation by a class of Szász-Mirakjan operators. J. Appl.
Funct. Anal. 9(3–4), 356–368 (2014)
14. Phillips, R.S.: An inversion formula for Laplace transforms and semi-groups of linear operators.
Ann. Math. 59(2), 325–356 (1954)
15. Phillips, G.M.: Interpolation and Approximation by Polynomials. Springer (2003)
16. Sahai, A., Prasad, G.: On simultaneous approximation by modified Lupas operators. J. Approx.
Theory 45, 122–128 (1985)
17. Stanila, E.D.: On Bernstein-Euler-Jacobi operators, PhD Thesis University of Duisburg-Essen
(2014)
18. Wagner, M.: Quasi-Interpolanten zu genuinen Baskakov-Durrmeyer-Typ Operatoren, Dissser-
tation Universität Wuppertal (2013)
Rate of Convergence of Modified
Schurer-Type q-Bernstein Kantorovich
Operators
Manjari Sidharth and P.N. Agrawal
Abstract Lin (J. Inequal. Appl. 465, 2014 [10]) introduced a new modified Schurer-
type q-Bernstein Kantorovich operators and discussed a local approximation theorem
and the statistical convergence of these operators. In this paper we study the rate
of convergence by means of the first-order modulus of continuity, Lipschitz class
function, the modulus of continuity of the first-order derivative and the Voronovskaja-
type theorem.
Keywords Schurer type q-Bernstein Kantorovich operators · Rate of convergence ·

Modulus of continuity · Lipschitz class function
1 Introduction
In 1987, q-analogue of classical Bernstein polynomials was introduced by Lupas

[11]. After a decade, Phillips [13] proposed another generalization of these polynomi-
als based on q-integers and discussed the rate of convergence and Voronovskaja-type
asymptotic formula. The q-analogue of Bernstein polynomials due to Phillips was
studied by several researchers, e.g. Ostrovska [14, 15], Kim [9], Wang [17], etc. Sub-
sequently, some other generalizations based on q-integers of the other well-known
positive linear operators were proposed and studied.
In 2005, Derriennic [5] introduced the q-analogue of Bernstein Durmeyer polyno-
mials with Jacobi weights and studied some approximation properties. Later, Gupta
[7] introduced the q-analogue of the Bernstein Durmeyer operators which was inves-
M. Sidharth (B) · P.N. Agrawal

e-mail: manjarisidharth93@gmail.com
P.N. Agrawal

DOI 10.1007/978-81-322-2485-3_19
244 M. Sidharth and P.N. Agrawal
tigated later by Finta and Gupta [6] and several other researchers. Dalmanoglu [4]
introduced the Kantorovich-type modification of q-Bernstein polynomials and estab-
lished some approximation results. Subsequently, Radu investigated the statistical
convergence results of these operators.
Muraru [12] introduced Bernstein–Schurer polynomials based on q-integers
and established the rate of convergence in terms of modulus of the continuity.
Agrawal et al. [1] considered the Stancu varient of these operators and discussed
some local and global direct results. Later, Agrawal et al. [2] proposed Durmeyer-
type modification of these operators and discussed some local direct results and
studied the rate of convergence of modified limit q-Bernstein–Schurer-type opera-
tors.
Very recently, Lin [10] introduced a new kind of modified Schurer-type q-
Bernstein Kantorovich operators as follows:
Let p ∈ N0 (the set of non-negative integers) be arbitrary but fixed and α, β be
integers satisfying 0 ≤ α ≤ β. For f ∈ C[0, 1 + p], he defined
n+
p 1
(α,β) t q[k + α]q
K n,q ( f ; x) = p̄n,k (q; x) f + dq t, x ∈ [0, 1]
0 [n + 1 + β]q [n + 1 + β]q
k=0
n+
p+k−1
n (α,β)
where p̄n,k (q; x) = xk (1 − q s x). It is clear that K n,q ( f ; x) is a
k q
s=0
linear positive operator. It is remarked that when α = β = 0, it reduces to the
operator discussed in [16].
In the present paper, we continue the work done by Lin by discussing the rate of
convergence in terms of the modulus of continuity, elements of Lipschitz-type space
and Voronovskaja-type theorem. Throughout this paper, we consider 0 < q < 1. For
the properties of the q-calculus, we refer to [3, 8].
2 Preliminaries
In this section, we give some basic results which will be used in the sequel.
(α,β)
Lemma 1 ([10]) For K n,q (t m ; x), m = 0, 1, 2, we have
(α,β)
(i) K n,q (1; x) = 1,
(α,β) [n + p]q 1 1
(ii) K n,q (t; x) = q α+1 x + + q[α]q ,
[n + 1 + β]q [n + 1 + β]q [2]q
(α,β) 2 [n + p]q [n + p − 1]q 2α+3 2
(iii) K n,q (t ; x) = q x
[n + 1 + β]q2
Rate of Convergence of Modified Schurer-Type … 245

[n+ p] α+1
+ [n+1+β]q 2 2
[2]q q + q 2+α (2[α]q + q α ) x
q

2q[α]q
+ 1
[n+1+β]q2
1
[3]q + [2]q + q [α]q .
2 2
Remark 1 For the modified Schurer-type q-Bernstein Kantorovich operators, we

have

(α,β) 1 + 2α
(i) lim [n]qn (K n,qn ((t − x); x) = − (α + 1)x ,
n→∞ 2
(α,β)
(ii) lim [n]qn (K n,q n
((t − x)2 ; x) = x(1 − x).
n→∞
3 Main Results
The first modulus of continuity of f ∈ C[0, 1 + p] for δ > 0, is given by
ω( f, δ) = max | f (x + h) − f (x)|.
0<|h|<δ,x,x+h∈[0,1+ p]
We observe that for all f ∈ C[0, 1 + p], we have
lim ω( f, δ) = 0
δ→0+
and for any δ > 0,

|x − y|
| f (x) − f (y)| ≤ ω( f, δ) +1 . (1)
δ
Theorem 1 For f ∈ C[0, 1 + p] , we have

(α,β) (α,β)
|K n,q ( f ; x) − f (x)| ≤ 2ω f ; δn,q ,
(α,β) (α,β)
where ω( f, .) is the modulus of continuity of f and δn,q := K n,q ((t − x)2 ; x).
(α,β)
Proof Using the linearity and positivity of the operator K n,q ( f ; x) in view of (1),
we get
n+ p 1
(α,β) t q[k + α]
|K n,q ( f ; x) − f (x)| = p̄n,k (q; x) f + − f (x) dq t
0 [n + 1 + β]q [n + 1 + β]q
k=0
n+
p 1
t q[k + α]q
≤ f
p̄n,k (q; x) + − f (x)dq t
[n + 1 + β]q [n + 1 + β]q
k=0 0

t +
q[k+α]q
− x
n+
p 1 [n+1+β]q [n+1+β]q
≤ p̄n,k (q; x) + 1 ω( f, δ)dq t
0 δ
k=0
n+
p n+
p 1
ω( f, δ) t
≤ ω( f, δ) p̄n,k (q; x) + p̄n,k (q; x)
δ [n + 1 + β]
0 q
k=0 k=0

q[k + α]q
+ − x dq t .
[n + 1 + β]q
On applying the Cauchy–Schwarz inequality, we have

1

t q[k + α]q

[n + 1 + β] + [n + 1 + β] − x dq t
0 q q
1 2
1/2
t q[k + α]q
≤ + − x dq t
0 [n + 1 + β]q [n + 1 + β]q

(α,β)
= an,k (x).
Hence,
ω( f, δ) (α,β) 1/2
n+ p
(α,β)
|K n,q ( f ; x) − f (x)| ≤ ω( f, δ) + {an,k } p̄n,k (q; x).
δ
k=0
Again applying the Cauchy–Schwarz inequality, we get

(α,β)
|K n,q ( f ; x) − f (x)|
n+ p
1/2 n+
1/2
ω( f, δ) (α,β) p
≤ ω( f, δ) + an,k p̄n,k (q; x) p̄n,k (q; x)
δ
k=0 k=0
n+
p 1 2
1/2
ω( f, δ) t q[k + α]q
= ω( f, δ) + p̄n,k (q; x) + −x dq t
δ 0 [n + 1 + β]q [n + 1 + β]q
k=0

1/2
ω( f, δ) (α,β)
= ω( f, δ) + K n,q ((t − x)2 ; x) .
δ
(α,β) (α,β)
Choosing δ := δn,q = K n,q ((t − x)2 ; x), we have

(α,β) (α,β)
|K n,q ( f ; x) − f (x)| ≤ 2ω f, δn,q (x) .
Hence, we get the desired result.

Corollary 1 Let f ∈ Li p M (ξ ) for 0 < ξ ≤ 1, then

ξ/2
(α,β) (α,β)
|K n,q ( f ; x) − f (x)| ≤ 2M δn,q (x) ,
(α,β) (α,β)
where δn,q (x) = K n,q ((t − x)2 ; x).
Proof Since f ∈ Li p M (ξ ), we have ω( f, δ) ≤ Mδ ξ for any δ > 0.

Hence the result follows from Theorem 1.
Theorem 2 If f (x) has a continuous derivative f (x) and ω( f , δ) is the modulus

of continuity of f (x) on [0, 1 + p], then

(α,β) (α,β)
|K n,q ( f ; x) − f (x)| ≤ M |μ(α,β)
n,q, p | + ω( f
, δ) 1 + δ n,q (x) ,
where M is a positive constant such that | f (x)| ≤ M and

q α+1 [n + p]q 1 1
μ(α,β)
n,q, p (x) = −x x+ + q[α]q . (2)
[n + 1 + β]q [n + 1 + β]q [2]q
Proof On applying the mean value theorem, we get

t q[k + α]q t q[k + α]q
f + − f (x) = + − x f (ξ )
[n + 1 + β]q [n + 1 + β]q [n + 1 + β]q [n + 1 + β]q

t q[k + α]q
= + − x f (x)
[n + 1 + β]q [n + 1 + β]q

t q[k + α]q
+ + − x ( f (ξ ) − f (x)),
[n + 1 + β]q [n + 1 + β]q

t q[k + α]q
where ξ lies between + and x.
[n + 1 + β]q [n + 1 + β]q
Hence, we get

(α,β) p 1
n+
t q[k + α]q

|K n,q ( f ; x) − f (x)| = f (x) + −x
0 [n + 1 + β]q [n + 1 + β]q
k=0
n+ p+k−1

n
× xk (1 − q s x)dq t
k q
s=0
p 1
n+
t q[k + α]q

+ + −x
[n + 1 + β]q [n + 1 + β]q
k=0 0
n+ p+k−1
n
× ( f (ξ ) − f (x)) xk (1 − q s x)dq t
k q
s=0

(α,β) p 1
n+
t q[k + α]q
≤ | f (x)||K n,q ((t − x); x)| +
[n + 1 + β] + [n + 1 + β] − x
0 q q
k=0
n+ p+k−1

n
× | f (ξ ) − f (x)|
xk (1 − q s x)dq t
k q
s=0
⎛
⎞
t +
q[k+α]q
− x
(α,β)
n+
p 1 ⎜ [n+1+β]q [n+1+β]q ⎟
≤ M|μn,q, p | + ω( f , δ) ⎜
⎝ + 1⎟
⎠
0 δ
k=o
n+ p+k−1
t q[k + α]q n
× + − x xk (1 − q s x)dq t
[n + 1 + β]q [n + 1 + β]q k q s=0

(α,β) p 1
n+
t q[k + α]q
≤ M |μn,q, p | + ω( f , δ)
[n + 1 + β] + − x

q [n + 1 + β]q
k=o 0
n+ p+k−1

n
k q
s=0
n+
p 1 2
ω( f , δ) t t
+ + −x
δ [n + 1 + β]q [n + 1 + β]q
k=0 0
n+ p+k−1

n
× xk (1 − q s x)dq t,
k q
s=0
(α,β)
where μn,q, p is given by (2).
Now, applying Cauchy–Schwarz inequality in second term of the right side of the
inequality, we have
(α,β)
|K n,q ( f ; x) − f (x)|
n+
p 1 2
t q[k + α]q
≤ M |μ(α,β)
n,q, p | + ω( f
, δ) + −x
0 [n + 1 + β]q [n + 1 + β]q
k=0
n+
p+k−1 1/2
n
k q
s=0
n+ p 2
ω( f , δ) 1
n+
p+k−1
t t n
+ + −x xk (1 − q s x)dq t,
δ 0 [n + 1 + β]q [n + 1 + β]q k q
k=0 s=0
, δ)
(α,β) ω( f
≤ M |μ(α,β)
n,q, p | + ω( f , δ) K n,q ((t − x) ; x) +
2 (α,β)
K n,q ((t − x)2 ; x).
δ
(α,β) (α,β)
Choosing δ := δn,q = K n,q ((t − x)2 ; x), we have

(α,β) (α,β)
|K n,q ( f ; x) − f (x)| ≤ M|μ(α,β) (α,β)
n,q, p | + ω( f , δn,q (x))(1 + δn,q (x)).

3.2 Asymptotic Result
Theorem 3 Let f ∈ C[0, 1 + p], 0 < qn < 1 be a sequence such that qn → 1 and

[n]q → 0, as n → ∞. Suppose that f (x) exist at a point x ∈ [0, 1], then we have
1
n

(α,β) 1 + 2α 1
lim [n]qn K n,q n
( f ; x) − f (x) = − (α + 1)x f (x) + x(1 − x) f (x).
n→∞ 2 2
Proof By Taylor’s expansion we have
1
f (t) = f (x) + (t − x) f (x) + f (x)(t − x)2 + r (t, x)(t − x)2 , (3)
2
where r (t, x) is the Peano form of the remainder and lim r (t, x) = 0.
t→x
(α,β)
On applying K n,qn (.; x) on both sides of Eq. (3), we get
(α,β) 1
K n,q n
( f ; x) − f (x) = f (x)K n,q
(α,β)
n
((t − x); x) + (α,β)
f (x)K n,q n
((t − x)2 ; x)
2
(α,β)
+ K n,q n
((t − x)2 r (t, x); x).
Taking the limit as n → ∞ on both sides of the above equation, we get

(α,β) (α,β)
lim [n]qn (K n,qn ( f ; x) − f (x)) = lim [n]qn f (x) K n,qn ((t − x); x)
n→∞ n→∞

f (x) (α,β)
+ lim [n]qn (K n,qn ((t − x)2 ; x
n→∞ 2
(α,β)
+ lim [n]qn K n,qn ((t − x)2 r (t, x); x). (4)
n→∞
From the Remark 1, we have

(α,β) 1 + 2α
lim [n]qn (K n,q n
((t − x); x) = − (α + 1)x , (5)
n→∞ 2
uniformly in [0,1], and

(α,β)
lim [n]qn (K n,q n
((t − x)2 ; x) = x(1 − x) , uniformly in [0, 1]. (6)
n→∞
Hence in order to prove the result it is sufficient to show that

(α,β)
[n]qn K n,q n
((t − x)2 r (t, x); x) → 0 as n → ∞ , uniformly in [0, 1].
By using the Cauchy–Schwarz inequality, we have

(α,β) (α,β) (α,β)
K n,q n
((t − x)2 r (t, x); x) ≤ K n,qn (r 2 (t, x); x) K n,qn ((t − x)4 ; x). (7)
We observe that r 2 (x, x) = 0, and from the Basic convergence theorem [10], we
have
(α,β) 2
lim K n,q n
(r (t, x); x) = r 2 (x, x) = 0. (8)
n→∞
Hence, from (7) and (8), we get

(α,β)
lim [n]qn K n,q n
((t − x)2 r (t, x); x) = 0, uniformly in [0, 1], (9)
n→∞
in view of the fact that

(α,β) 1
K n,q ((t − x)4 ; x) = O as n → ∞, uniformly in [0, 1].
n
n2
Now, combining (4)–(6) and (9), we get the required result.

Now, we consider the following two-parameter Lipschitz-type space:

(a,b)
Li p M (r ) := f ∈ C[0, 1 + p] : | f (t) − f (x)|

|t − x|r
≤M ; x, ∈ (0, 1], t ∈ [0, 1 + p] ,
(t + ax + bx 2 )r/2
where M is a positive constant and r ∈ (0, 1].

Theorem 4 Let f ∈ Li p (a,b)
M (r ). Then ∀x ∈ (0, 1), we have
(α,β) r/2
(α,β) δn,q (x)
|K n,q ( f ; x) − f (x)| ≤ M ,
ax + bx 2
(α,β) (α,β)
where δn,q (x) = K n,q ((t − x)2 ; x).
Proof First, we prove the result for r = 1.

p
n+ 1 t q[k + α]q
(α,β)
|K n,q ( f ; x) − f (x)| ≤ p̄n,k (q; x)f + − f (x)dq t
[n + 1 + β]q [n + 1 + β]q
k=0 0

1 q[k+α]q
[n+1+β]q + [n+1+β]q − x
t
p
n+
≤ M p̄n,k (q; x) dq t.
q[k+α]q
[n+1+β]q + [n+1+β]q + ax + bx
0 t 2
k=0
Since
1 1
<√ ,
t
+
q[k+α]q
+ ax + bx 2 ax + bx 2
[n+1+β]q [n+1+β]q
the last inequality implies that
p
n+ 1
M t q[k + α]q
(α,β)
|K n,q ( f ; x) − f (x)| ≤ √ p̄n,k (q; x) + − x dq t
[n + 1 + β] [n + + β]
ax + bx k=0
2 0 q 1 q
M (α,β)
= √ K n,q (|t − x|; x)
ax + bx 2

(α,β)
δn,q (x)
= M ,
ax + bx 2
on applying the Cauchy–Schwarz inequality. Hence, the result is true for r = 1.

Now, we prove the result for r ∈ (0, 1). Applying the Hölder’s inequality with
1 1
p = and q = , we get
r 1−r

p
n+ 1 t q[k + α]q
(α,β)
|K n,q ( f ; x) − f (x)| ≤ p̄n,k (q; x) f + − f (x)dq t
[n + 1 + β]q [n + 1 + β]q
k=0 0
n+ 1/r
r
p 1 t q[k + α]q
≤ p̄n,k (q; x) f + − f (x)dq t .
[n + 1 + β]q [n + 1 + β]q
k=0 0
Again applying Hölder’s inequality with p = 1

r and q = 1
1−r , we get
n+
p
1
(α,β) f t q[k + α]q
|K n,q ( f ; x) − f (x)| ≤ p̄n,k (q; x) +
0 [n + 1 + β]q [n + 1 + β]q
k=0
1/r
r

− f (x) dq t .
Since f ∈ Li p (a,b)
M (r ), we have

n+ 1 t q[k+α]q
r
(α,β) p [n+1+β]q + [n+1+β]q − x
|K n,q ( f ; x) − f (x)| ≤ M p̄n,k (q; x) dq t
0 t q[k+α]q
[n+1+β]q + [n+1+β]q + ax + bx
k=0 2
n+
p 1
r
M t q[k + α]q
≤ p̄n,k (q; x)
r [n + 1 + β] + [n + 1 + β] − x dq t
ax + bx 2 k=0 0 q q
M r
(α,β)
≤
r
K n,q (|t − x|; x) .
ax + bx 2
Thus, on applying Cauchy–Schwarz inequality, we have

(α,β) r/2
(α,β) δn,q (x)
|K n,q ( f ; x) − f (x)| ≤ M .
ax + bx 2
Acknowledgments The authors are extremely grateful to the reviewers for their valuable comments
leading to a better presentation of the paper.
References
1. Agrawal, P.N., Gupta, V., Sathish, A.: Kumar, On q-analogue of Bernstein-Schurer-Stancu

operators. Appl. Math. Comput. 219(14), 7754–7764 (2013)
2. Agrawal, P.N., Sathish, A., Kumar, Sinha, T.A.K.: Stancu type generalization of modified
Schurer operators based on q-integers. Appl. Math. Comput. 226, 765–776 (2014)
3. Aral, A., Gupta, V., Agarwal, R.P.: Applications of q-Calculus in Operator Theory. Springer
(2013)
4. Dalmanoglu, Ö., Dogru, O.: On statistical approximation properties of Kantorovich type q-
Bernstein operators. Math. Comput. Modell. 52, 760–771 (2010)
5. Derriennic, M.M.: Modified bernstein polynomials and jacobi polynomials in q-calculas. Rend.
Circ. Mat. Palermo Serie II (Suppl. 76), 269–290 (2005)
6. Finta, Z., Gupta, V.: Approximation properties of q-Baskakov operators. Cent. Eur. J. Math.
8(1), 199–211 (2010)
7. Gupta, V.: Some approximation properties of q-Durmeyer operators. Appl. Math. Comput.
197(1), 172–178 (2008)
8. Kac, V., Cheung, P.: Quantum Calculus. Universitext, Springer, New York (2002)
9. Kim, T.: A note on q-Bernstein polynomials. Russ. J. Math. Phys. 18(1), 73–82 (2011)
10. Lin, Q.: Statistical approximation of modified Schurer type q-Bernstein Kantorovich operators.
J. Inequal. Appl. 465, (2014)
11. Lupas, A.: A q-analogue of the Bernstein operators. Semin. Numer. Stat. Calc. University of
Cluj-Napoca 9, 85–92 (1987)
12. Muraru, C.V.: Note on q-Bernstein-Schurer operators. Stud. Univ. Babes-Bolyai Math. 56,
489–495 (2011)
13. Phillips, G.M.: Bernstein polynomials based on q-integers. Ann. Numer. Math. 4(1–4), 511–
518 (1997)
14. Ostrovska, S.: q-Bernstein polynomials and their iterates. J. Approx. Theory 123(2), 232–255
(2003)
15. Ostrovska, S.: On Lupas q-analogue of the Bernstein operator. Rocky Mountain J. Math. 36(5),
1615–1629 (2006)
16. Vedi, T.: Some Schurer Type q-Bernstein operators. Dissertation. Eastern Mediterranean Uni-
versity (2011)
17. Wang, H.: Properties of convergence for q Bernstein polynomials. J. Math. Anal. Appl. 340(2),
1096–1108 (2008)
Operators of Durrmeyer Type with Respect
to Arbitrary Measure
Elena E. Berdysheva
Abstract In this paper, we give an overview of operators of Durrmeyer type with

respect to arbitrary measure. Our construction includes the Bernstein–Durrmeyer
operator, the Szász–Mirakjan–Durrmeyer operator, and the Baskakov–Durrmeyer
operator with respect to arbitrary measure. We are particularly interested in the con-
vergence of the operators. We discuss the uniform and the pointwise convergence as
well as convergence in the corresponding weighted L p -spaces. A new result is the
statement on the L p -convergence of the Szász–Mirakjan–Durrmeyer operator and
the Baskakov–Durrmeyer operator without additional restrictions on the measure.
Keywords Bernstein–Durrmeyer operator · Szász–Mirakjan–Durrmeyer operator ·

Baskakov–Durrmeyer operator · Uniform convergence · Pointwise convergence ·
L p -convergence
2010 AMS Subject Classification: 41A36
1 Introduction
In this paper, we consider a class of positive linear operators of Durrmeyer

type. Let
c ∈ R. Depending on the value of c, we consider the intervals Ic = 0, − 1c for
c < 0 and Ic = [0, ∞) for c ≥ 0. For n > 0, k ∈ N0 := N ∪ {0} and x ∈ Ic we
define the basis functions by the formulae

[c] − nc n
pn,k (x) := (−1)k (cx)k (1 + cx)− c −k , c = 0,
k (1)
[0] [c] (nx)k −nx
pn,k (x) := lim pn,k (x) = e , c = 0.
c→0 k!
E.E. Berdysheva (B)

Department of Mathematics and Science, German University of Technology in Oman,
PC 130 Athaibah, P.O. Box 1816 Muscat, Sultanate of Oman
e-mail: elena.berdysheva@gutech.edu.om

DOI 10.1007/978-81-322-2485-3_20
256 E.E. Berdysheva
[c]
The functions pn,k satisfy the property
∞
[c]
pn,k (x) = 1.
k=0
The basis functions are nonnegative on Ic , i.e.,
[c]
pn,k (x) ≥ 0, x ∈ Ic , for all k ∈ N0 ,
if the following conditions are fulfilled: n > 0 if c ≥ 0, and n = −c, ∈ N, if

[c]
c < 0. In the latter case pn,k ≡ 0 for k > − nc , k ∈ N0 . It is not difficult to see that
[c]
pn,k (x) = p [1]
n (cx),
,k
c > 0,
c
[c] [−1]
pn,k (x) = p− n (−cx), c < 0.
c ,k
Thus, there are only three significantly different cases, namely, c = −1, c = 0, and
c = 1. We will restrict our consideration to these three cases.
The basis functions (1) are traditionally used to define positive linear operators for
functions on Ic . The first and the most well known of these operators is the operator
of the form
∞
[c] k [c]
Bn f := f pn,k , (2)
n
k=0
where f is a continuous function on Ic . This is a positive linear operator that repro-

duces linear functions. In the case c = −1, the operator is the famous Bernstein
operator. It was introduced by Bernstein [8] to give a constructive proof of the
Weierstrass Approximation Theorem. The work of Bernstein initiated a study of
numerous modifications and generalizations of this operator by many authors. We
will call operator (2) when c = 0 the Szász–Mirakjan operator. It was considered
independently by several authors, including Mirakjan [18], Favard [12], and Szász
[22]. In the case c = 1, operator (2) was first defined by Baskakov [1], who also
introduced a general frame which includes the operators Bn[c] with c ∈ R.
Aiming to have a similar construction for integrable functions, the so-called
Durrmeyer variant of operator (2) was introduced as
∞
[c]
f (t) pn,k (t) dt [c]
Mn[c]
Ic
f := [c]
pn,k , (3)
k=0 Ic pn,k (t) dt
where f ∈ L p (Ic ), 1 ≤ p < ∞, or f ∈ C(Ic ). This is a positive linear operator that

reproduces constant functions. The Bernstein–Durrmeyer operator (that corresponds
to the case c = −1) was defined independently by Durrmeyer [11] and Lupaş [16]
Operators of Durrmeyer Type with Respect to Arbitrary Measure 257
and became known due to Derriennic (e.g., [9]). It can be naturally generalized for
functions of several variables defined on the d-dimensional simplex. The Szász–
Mirakjan–Durrmeyer operator (case c = 0) is due to Mazhar and Totik [17]. The
Baskakov–Durrmeyer operator (case c = 1) was defined by Sahai and Prasad [21]
and, independently, by Heilmann [13].
In what follows, we will use the following function spaces and norms. For a
compact set A, we denote by C(A) the space of continuous functions on A with the
norm

f
C(A) := max | f (x)|.
x∈A
The space L ∞ (Ic , ρ) is the space of essentially bounded functions on Ic with

respect to a measure ρ with the norm

f
L ∞ (Ic ,ρ) := ess supx∈Ic | f (x)|.
The spaces L p (Ic , ρ), 1 ≤ p < ∞, are the spaces of functions f for which | f | p
is integrable on Ic with the norm
1
p

f
L p (I c ,ρ) := | f (x)| dρ(x)
p
.
Ic
2 Durrmeyer-Type Operators with Respect

to Arbitrary Measure
We generalize the operators of Durrmeyer-type (3) in the following way: the new
operator has the same form, but the integration is taken not with respect to the
Lebesgue measure d x but with respect to some measure dρ(x). The exact definition
of the operator is as follows. Let ρ be a nonnegative locally bounded Borel measure on
Ic . Then, in particular, ρ is regular (being a nonnegative bounded Borel measure on a
metric space), and thus polynomials are dense in the spaces L p (Ic , ρ), 1 ≤ p < ∞,
and in C(Ic ). Furthermore, we suppose that
supp ρ = ∂ Ic (4)
[c]
(where ∂ Ic denotes the boundary of Ic ). This condition guarantees that Ic pn,k (t)
dρ(t) = 0 for all n and k. In the cases c = 0 and c = −1, we need to additionally
take care about the convergence. This requirement implies the following further
conditions on the measure ρ: let
∞
e−γt dρ(t) < ∞ for some γ > 0 if c = 0 (5)
0
258 E.E. Berdysheva
and ∞
(1 + t)−γ dρ(t) < ∞ for some γ > 0 if c = 1. (6)
0
Definition 1 Let ρ be a nonnegative locally bounded Borel measure on Ic that satis-

fies (4), (5), and (6). Let f ∈ L p (Ic , ρ), 1 ≤ p ≤ ∞. The Durrmeyer-type operator
with respect to the measure ρ is defined by the formula
∞
[c]
f (t) pn,k (t) dρ(t) [c]
[c] Ic
Mn,ρ f := [c]
pn,k (7)
k=0 Ic pn,k (t) dρ(t)
for n ∈ N if c = −1, n > γ p−1 p−1

p if c = 0, and n ≥ γ p if c = 1.
Please note that the Bernstein–Durremer operator (c = −1) can also be similarly
designed for functions of several variables defined on the d-dimensional simplex.
All results presented in this paper remain valid also in this case. However, we restrict
our presentation to the one-dimensional case, for the sake of simplicity.
The Bernstein–Durrmeyer operator (c = −1) with respect to an arbitrary measure
was for the first time studied in full generality in [5], to our knowledge. However,
the Bernstein–Durrmeyer operator in a special case of a measure ρ different from
the Lebesgue measure, namely, for the Jacobi measure dρ(x) = x α (1 − x)β d x,
α, β > −1, is well known and very well studied. It was introduced by Păltănea [19],
see also paper [7] by Berens and Xu. The multidimensional case was considered, for
example, in [10]. A more general operator than the Bernstein–Durrmeyer operator
with respect to Jacobi measure was considered by Păltănea in [20, Section 5.2]:
he studied Bernstein–Durrmeyer operators of the form (7) with dρ(x) = x α (1 −
x)β h(x) d x, where h ∈ C[0, 1], h(t) > 0 for all t ∈ [0, 1], α, β > −1. The Szász–
Mirakjan–Durrmeyer operators (case c = 0) and the Baskakov–Durrmeyer operators
(case c = 1) with respect to arbitrary measure were introduced and studied in [4].
We introduced an arbitrary measure in the construction having in mind appli-
cations in learning theory. Indeed, Jetter and Zhou [14] applied the Bernstein–
Durrmeyer operator with respect to arbitrary measure, in the one-dimensional case,
to bias-variance estimates for vector support machine classifiers. Li [15] used the
Bernstein–Durrmeyer operators with respect to arbitrary measure in the multidimen-
sional case in study of learning rates of least-squares regularized linear regression
with polynomial kernels.
Just to give the reader a feeling how the Bernstein–Durrmeyer operators can be
applied in the frames of learning theory, we give a short description of the problem
considered by Li in [15]. Let X be a compact set in Rd , Y = R, and σ be a Borel
probability measure on Z := X × Y . We denote by σ X the marginal distribution of
σ on X . For a function f : X → Y , the least-squares error is

E( f ) := ( f (x) − y)2 dσ.
Z
The function
f σ (x) := y dσ(y|x),
Y
where σ(y|x) is the conditional probability induced by σ, minimizes the least-squares

error. It is called the regression function. The aim is to find a good approximation of
the regression function based on a random sample z := {(xi , yi )}i=1 m ⊂ Z of size m.
For x, t ∈ X , put K n (x, t) := (1 + x · t) , where x · t = x1 t1 + · · · + xd td . This

n
function is a Mercer kernel, and we denote by H K n the corresponding reproducing

kernel Hilbert space, which is in this case the space of algebraic polynomials of
degree at most n endowed with the corresponding norm. The least-squares regularized
regression algorithm with the polynomial kernel K n is the minimization problem

1
m
f z,n,λ := arg min ( f (xi ) − yi ) + λ
f
H K ,
2 2
f ∈H K n m n
i=1
where λ > 0 is the regularization parameter. One usually takes λ = λ(m) with
limm→∞ λ(m) = 0. We expect that f z,n,λ gives a good approximation of the regres-
sion function f σ .
Li considered the case when X is the d-dimensional simplex, and, in particular,
gave an estimate for the rate of convergence
f z,n,λ − f σ
L 2 (X,σ X ) of type O(m −β )
with some β > 0. One of the important steps in deriving this estimate is approxi-
[−1]
mating of f σ by a function from H K n . This approximation is realized by Mn,ρ fσ
(more exactly, by its multivariate version) with ρ = σ X . The method employs the
[−1]
estimates for the rate of convergence of the operator Mn,ρ as given in Theorem 6
below.
Following this description of an application of operators (7), we return to dis-
cussing their properties. Obviously, operators (7) are linear positive operators that
reproduce constant functions. Our first result is about boundedness of the operators.
Theorem 1 Let 1 ≤ p ≤ ∞. Let ρ and n be as in Definition 1. Then the operator

[c]
Mn,ρ : L p (Ic , ρ) → L p (Ic , ρ)
is well-defined. Moreover,
[c]

Mn,ρ
L p (Ic ,ρ)→L p (Ic ,ρ) = 1.
In what follows, we concentrate on the question of convergence of operators (7).

For other properties of the operators under consideration see, e.g., [5].
260 E.E. Berdysheva
3 Uniform and Pointwise Convergence
Uniform and pointwise convergence of the Bernstein–Durrmeyer operators (c = −1)

with respect to arbitrary measure, in the multivariate case, was studied in [2, 3]. The
methods of these papers were adopted to the case of an unbounded interval in [4],
where corresponding results for the Szász–Mirakjan–Durrmeyer operators (c = 0)
and the Baskakov–Durrmeyer operators (c = 1) with respect to arbitrary measure
were obtained. Below we summarize the results on the uniform and the pointwise
convergence (in the one-dimensional case).
Recall that a measure ρ on Ic is called strictly positive if ρ(A ∩ Ic ) > 0 for every
open set A ⊂ R such that A∩ Ic = ∅. This is equivalent to the fact that supp (ρ) = Ic .
Our first statement gives necessary and sufficient conditions for the uniform con-
vergence of the Bernstein–Durrmeyer operator with respect to a measure ρ on the
compact set I−1 = [0, 1] for every function f ∈ C([0, 1]).
Theorem 2 Let c = −1, and ρ and n be as in Definition 1. Then

[−1]
lim
f − Mn,ρ f
C([0,1]) = 0 for every f ∈ C([0, 1])
n→∞
if and only if ρ is strictly positive on [0, 1].
A statement about pointwise convergence can be proved in all three cases c = −1,
c = 0, or c = 1.
Theorem 3 Let c = −1, c = 0, or c = 1, and ρ and n be as in Definition 1 with

p = ∞. Let x ∈ supp ρ. Let f ∈ L ∞ (Ic , ρ) and continuous at x. Then

[c]
lim f (x) − Mn,ρ f (x) = 0.
n→∞
Moreover, the convergence is uniform on every compact subset of the interior of

supp ρ. For a set B, we denote by B ◦ the interior of B. We understand the interior
relatively to the set Ic , i.e., the boundary points of Ic may belong to it.
Theorem 4 Let c = −1, c = 0, or c = 1, and ρ and n be as in Definition 1 with

p = ∞. Let A be a compact set, A ⊂ (supp ρ)◦ . Let f ∈ L ∞ (Ic , ρ) and continuous
on A. Then
[c]
lim
f − Mn,ρ f
C(A) = 0.
n→∞
A very interesting open question is to obtain estimates for rates of uniform or point-
wise convergence of operators of Durrmeyer type with respect to arbitrary measure.
4 Convergence in the Weighted L p -spaces
The first result on the convergence of operators (7) in the spaces L p (Ic , ρ),
1 ≤ p < ∞, was obtained for the Bernstein–Durrmeyer operator (c = −1) with
respect to arbitrary measure, in the multidimensional case, by Li [15]. She also
obtained estimates for the rate of convergence in terms of a K-functional that are
good in the cases when p = 1 or p = 2. Using the same idea, the estimates for other
values of p (1 ≤ p < ∞) were improved in [6].
The methods of Li’s paper [15] were transferred to the cases of the Szász–
Mirakjan–Durrmeyer operators (c = 0) and the Baskakov–Durrmeyer operators
(c = 1) in [4]. Also estimates for the rate of convergence were proved. However,
results in [4] were obtained under a very restrictive assumption that the measure ρ
on [0, ∞) has finite moments up to a certain order, and, in particular, is bounded on
[0, ∞) (see Theorem 6 below). Note that this condition is not satisfied in the classical
case of the Lebesgue measure.
Here we present a new result. Namely, we prove that the Szász–Mirakjan–
Durrmeyer operators (c = 0) and the Baskakov–Durrmeyer operators (c = 1)
with respect to arbitrary measure converge in L p (Ic , ρ), 1 ≤ p < ∞, for every
f ∈ L p (Ic , ρ), 1 ≤ p < ∞, without additional assumptions on the measure ρ. Our
method is a further development of the method of Li from [15]. This development
allows to overcome difficulties arising when we work on an infinite interval. The
result formulated below includes Li’s result [15] for c = −1 and is new in the cases
c = 0 and c = 1.
Theorem 5 Let 1 ≤ p < ∞. Let c = −1, c = 0, or c = 1, and ρ and n be as in
Definition 1. Then
[c]
lim
f − Mn,ρ f
L p (Ic ,ρ) = 0
n→∞
for each f ∈ L p (Ic , ρ).

We finish the paper by giving a statement about rates of convergence. This
result was proved in [6] for c = −1 and in [4] for c = 0 and c = 1. Denote by
C 1 (Ic ) the set of continuously differentiable functions g on Ic such that
g
C(Ic ) :=
supx∈Ic |g (x)| < ∞. Consider the K-functional

K( f, t) p,ρ := inf
f − g
L p (Ic ,ρ) + t
g
C(Ic ) , 1 ≤ p < ∞.
g∈C 1 (Ic )
Theorem 6 Let 1 ≤ p < ∞. Let c = −1, c = 0, or c = 1, and ρ and n be as in

Definition 1. Assume, in addition, that

x s dρ(x) < ∞ for some even s. (8)
Ic
Let 1 ≤ p ≤ s, and f ∈ L p (Ic , ρ). Then there is a constant C p,ρ that depends only
on p and the measure ρ such that
262 E.E. Berdysheva

[c] C p,ρ

f − Mn,ρ f
L p (Ic ,ρ) ≤ 2 K f, √ .
n p,ρ
Note that condition (8) is automatically satisfied for all s if c = −1 since in

this case I−1 = [0, 1] is a bounded interval. In the cases c = 0 and c = −1,
when Ic = [0, ∞), this condition is quite restrictive. Condition (8) guarantees, in
particular, that the set of continuously differentiable functions with finite support is
dense in the spaces L p (Ic , ρ), 1 ≤ p < ∞. Thus, K( f, t) p,ρ → 0 as t → 0 for all
f ∈ L p (Ic , ρ), 1 ≤ p < ∞.
An open question is to obtain estimates for the rates of convergence of the Szász–
Mirakjan–Durrmeyer operators (c = 0) and the Baskakov–Durrmeyer operators
(c = 1) with respect to arbitrary measure in L p (Ic , ρ), 1 ≤ p < ∞, in the general
case.
References
1. Baskakov, V.A.: An instance of a sequence of linear positive operators in the space of continuous
functions. Dokl. Akad. Nauk SSSR 113, 249–251 (1957) [in Russian]
2. Berdysheva, E.E.: Uniform convergence of Bernstein-Durrmeyer operators with respect to
arbitrary measure. J. Math. Anal. Appl. 394, 324–336 (2012)
3. Berdysheva, E.E.: Bernstein-Durrmeyer operators with respect to arbitrary measure, II: point-
wise convergence. J. Math. Anal. Appl. 418, 734–752 (2014)
4. Berdysheva, E.E., Al-Aidarous, E.: Szász-Mirakjan-Durrmeyer and Baskakov-Durrmeyer
operators with respect to arbitrary measure (submitted)
5. Berdysheva, E.E., Jetter, K.: Multivariate Bernstein-Durrmeyer operators with arbitrary weight
functions. J. Approx. Theory 162, 576–598 (2010)
6. Berdysheva, E.E., Li, B.-Z.: On L p -convergence of Bernstein-Durrmeyer operators with respect
to arbitrary measure. Publ. Inst. Math. Nouv. Sér. 96(110), 23–29 (2014)
7. Berens, H., Xu, Y.: On Bernstein-Durrmeyer polynomials with Jacobi-weights. In: Chui, C.K.
(ed.) Approximation Theory and Functional Analysis, pp. 25–46. Academic Press, Boston
(1991)
8. Bernstein, S.: Démonstration du théorème de Weierstrass, fondée sur le calcul des probabilités.
Commun. Soc. Math. Kharkow 13(2), 1–2 (1912–13)
9. Derriennic, M.-M.: Sur l’approximation de fonctions intégrables sur [0, 1] par des polynômes
de Bernstein modifiés. J. Approx. Theory 31, 325–343 (1981)
10. Ditzian, Z.: Multidimensional Jacobi-type Bernstein-Durrmeyer operators. Acta Sci. Math.
(Szeged) 60, 225–243 (1995)
11. Durrmeyer, J.-L.: Une formule d’inversion de la transformée de Laplace: Applications à la
théorie des moments, Thèse de 3e cycle, Faculté des Sciences de l’Université de Paris (1967)
12. Favard, J.: Sur le multiplicateurs d’interpolation. J. Math. Pures Appl. IX 23, 219–247 (1944)
13. Heilmann, M.: Direct and converse results for operators of Baskakov-Durrmeyer type. Approx.
Theory Appl. 5(1), 105–127 (1989)
14. Jetter, K., Zhou, D.-X.: Approximation with polynomial kernels and SVM classifiers. Adv.
Comput. Math. 25, 323–344 (2006)
15. Li, B.-Z.: Approximation by multivariate Bernstein-Durrmeyer operators and learning rates of
least-square regularized regression with multivariate polynomial kernels. J. Approx. Theory
173, 33–55 (2013)
16. Lupaş, A.: Die Folge der Betaoperatoren, Dissertation, Universität Stuttgart (1972)
17. Mazhar, S.M., Totik, V.: Approximation by modified Szász operators. Acta Sci. Math. 49,
257–269 (1985)
18. Mirakjan, G.M.: Approximation des fonctions continues au moyen de polynômes de la forme

mn
e−nx Ck,n x k , C. R. (Dokl.) Acad. Sc. URSS 31, 201–205 (1941)
k=0
19. Păltănea, R.: Sur un opérateur polynomial defini sur l’ensemble des fonctions intégrables,
“Babeş-Bolyai” Univ. Fac. Math. Res. Semin. 2, 101–106 (1983)
20. Păltănea, R.: Approximation Theory Using Positive Linear Operators. Birkhäuser-Verlag,
Boston (2004)
21. Sahai, A., Prasad, G.: On simultaneous approximation by modified Lupas operators. J. Approx.
Theory 45, 122–128 (1985)
22. Szász, O.: Generalization of S. Bernstein’s polynomials to the infinite interval. J. Res. Nat. Bur.
Stand. 45, 239–245 (1950)
Part II
Applications
Construction of Sparse Binary Sensing
Matrices Using Set Systems
R. Ramu Naidu
Abstract Recent developments at the intersection of algebra and optimization

theory—by the name of compressed sensing (CS)—aim at providing linear sys-
tems with sparse descriptions. The deterministic construction of the sensing matri-
ces is now an active directions in CS. The sparse sensing matrix contributes to fast
processing with low computational complexity. The present work attempts to relate
the notion of set systems to CS. In particular, we show that the set system theory
may be adopted to designing a binary CS matrix of high sparsity from the existing
binary CS matrices.
Keywords Compressed sensing · Restricted isometry property · Deterministic

construction · Set systems
1 Introduction
In recent years, sparse representations have become a powerful tool for efficiently
processing data in nontraditional ways. Compressed sensing (CS) is an emerging area
potential for sparsity-based representations. Since the problem of sparse recovery
through l0 norm minimization is generally NP-hard, Donoho et al. [1], Candes [2]
and Cohen et al. [3] have made several pioneering contributions and have reposed
the problem as an l1 -minimization problem. It is known that restricted isometry
property (RIP) is a sufficient condition to ensure the equivalence between l0 and l1
norm problems. As verifying RIP is computationally hard, there is much interest in
construction of RIP matrices.
Of late, the deterministic construction of binary CS matrices has attracted signif-
icant attention. Devore [4], Li et al. [5], Amini et al. [6], Indyk [7] have constructed
deterministic binary sensing matrices using ideas from algebra, graph theory, and
R.R. Naidu (B)

Department of Mathematics, Indian Institute of Technology, Medak,
Hyderabad 502205, A.P., India
e-mail: ma11p003@iith.ac.in

DOI 10.1007/978-81-322-2485-3_21
268 R.R. Naidu
coding theory. Devore [4] has been first to construct deterministic binary sensing
matrix of size p 2 × pr +1 with p number of ones in each column and coherence
being at most rp , for every fixed r and prime power p such that r < p. In the present
work, using the results from set systems we construct a binary sensing matrix from
a given binary sensing matrix in such a way that the resulting matrix is more sparser
than the input matrix. Consequently, the new matrix has potential for resulting in fast
algorithms.
The paper is organized into several sections. In Sect. 2, we present basic CS theory
and the conditions that ensure the equivalence between l0 -norm problem and l1 -norm
problem. In Sect. 3, we use the ideas from the set system theory and construct binary
sensing matrices of higher sparsity from the existing ones. We present our concluding
remarks in the last section.
2 Sparse Recovery from Linear Measurements
CS refers to the problem of reconstruction of an unknown vector u ∈ R M from

the linear measurements y = (u, φ1 , . . . , u, φ M ) ∈ R m with u, φ j being the
innerproduct between u and φ j . The basic objective in CS is to design a recovery pro-
cedure based on the sparsity assumption on u when the number of measurements m
is much small compared to M. Sparse representations seem to have merit for various
applications in areas such as image/signal processing and numerical computation.
A vector u ∈ R M is said to be k-sparse, if it has at most k nonzero coordinates.
One can find the sparse vector from its linear measurements by solving the following
l0 -norm optimization problem:
min v0 subject to φv = y. (1)

v
Here, v0 = |{i | vi = 0}| . The l0 -norm problem (1) is an NP-hard problem [2].
Candes et al. [2] have proposed the following l1 -norm minimization problem instead
of l0 -norm problem, making it computationally tractable LPP problem:
min v1 subject to φv = y. (2)

v
Here, v1 denotes the l1 -norm of the vector v ∈ R M .

Donoho et al. [1] and Kashin et al. [8], have provided the conditions under which
the solution to l0 -norm problem (1) is the same as that of l1 -norm problem (2). For
later use, we denote the solution to l1 -norm problem by f φ (y) and solution to l0 -norm
problem by u 0φ (y) ∈ R M .
Construction of Sparse Binary Sensing Matrices Using Set Systems 269
2.1 Equivalence Between l 0 -Norm and l 1 -Norm Problems
Definition 1 The mutual coherence μ(φ) of a given matrix φ is the largest absolute
normalized innerproduct between the pairs of columns of φ, that is,
| φiT φ j |
μ(φ) = max , (3)
1≤ i, j≤ M, i= j φi 2 φ j 2
where φi is the ith column of φ. It is known [1] that for μ-coherent matrices φ,
one has
u 0φ (y) = f φ (y) = u, (4)
provided u is k-sparse with k < 21 (1 + μ1 ).

Candes et al. ([2] and the references therein) have introduced the following isome-
try condition on matrices φ and have established its important role in CS. An m × M
matrix φ is said to satisfy the restricted isometry property (RIP) of order k with
constant δk if for all k-sparse vectors x ∈ R M , we have
(1 − δk ) xl22 ≤ φxl22 ≤ (1 + δk ) xl22 . (5)
The following proposition [9] relates the RIP constant δk and μ.

Proposition 1 Suppose that φ1 , . . . , φ M are the unit-normed columns of the matrix
φ with coherence μ. Then φ satisfies the RIP of order k with constant δk = (k − 1)μ.
Candes [2] has shown that whenever φ satisfies RIP of order 3k with δ3k < 1, the
CS reconstruction error satisfies the following estimate

u − f φ (φu) M ≤ Ck −1
2 σk (u) M , (6)
l 2
l 1
where σk (u)l M denotes the l1 error of the best k—term approximation, and the con-
1
stant C depends only on δ3k . This implies that the bigger the value of k for which
we can verify the RIP then better the guarantee we have on the performance of φ.
One of the important problems in CS theory deals with constructing CS matrices
that satisfy the RIP for the largest possible range of k. It is known that the widest
range possible is k ≤ C mM [4, 10–12]. However, the only known matrices that
log( m )
satisfy the RIP for this range are based on random constructions [10]. To the best
of our knowledge, designing the good deterministic constructions of RIP matrices is
still an open problem.
Since the sparsity of the sensing matrix is key to minimizing the computational
complexity associated with the matrix vector multiplication, it is desirable that the
CS matrix has smaller density. The sparse sensing matrix may contribute to fast
processing with low computational complexity in compressed sensing [13].
270 R.R. Naidu
Definition 2 [14] The density of a matrix is the ratio of the number of its nonzero
entries to the total number of its entries.
It may be noted that the density of the sensing matrix constructed by Devore [4] is
p . The sensing matrix constructed by Li et al. [5] have q as density. This matrix,
1 1
a generalization of [4], is of size |P|q × q L (G) , where q is any prime power and
P is the set of all rational points on algebraic curve X over finite field Fq . Amini
et al. [6] have constructed binary sensing matrices using OOC codes. The density of
this matrix is mλ , where m is row size and λ is the number of ones in each column.
Many data mining tasks can be concerned with identifying a small number of inter-
esting items from a tremendously large group without exceeding certain resource con-
straints. Specific examples [15] include the sketching and monitoring of heavy hitters
in high-volume data streams, source localization in sensor networks, multiplier-less
data compression and tomography. Note that, all these applications naturally cor-
respond to binary matrices. Furthermore, binary matrices with small density are
generally better. Thus, we focus on designing sparse binary matrices herein.
The present work attempts to address the deterministic construction of new binary
sensing matrix of smaller density from a given binary sensing matrix. Suppose φ is
a binary CS matrix of size m × M with m(m+1) 2 < M. In the next section, using the
results from set systems, we construct a binary sensing matrix ψ from φ in such a
way that the resulting matrix ψ is more sparse compared to the given matrix φ.
3 Construction of Binary CS Matrix of Smaller Density

from Existing Binary Matrix
Before presenting the main result, we discuss the definitions and results [16] relevant
to our construction methodology. Let V = {v1 , v2 , . . . , vm } be a set of m elements
(treated as “universe”). A set system S on V is simply some subset chosen from all
of the subsets of V , that is, S ⊂ 2V , the power set. A hypergraph is a collection of
several subsets of V , where some subsets may be present with a multiplicity greater
than 1. A set system may, however, contain each subset of V at most once.
Definition 3 Let H = {H1 , H2 , . . . , HM } be a hypergraph of M sets over the
universe V , and let φ = {φi j } be the m × M binary sensing incidence matrix of
hypergraph H , that is, the columns of φ correspond to the sets of H. The characteristic
vector on each H j gives the j t h column in φ, that is, φi j = 1 if xi ∈ H j otherwise
φi j = 0.
Definition 4 Let A = {ai j } and B = {bi j } be the two m × M matrices over a ring
R. Their dream product is an m × M matrix C = {ci j }, denoted by A B, and is
defined as ci j = ai j bi j for 1 ≤ i ≤ m, 1 ≤ j ≤ M.

Definition 5 Let f (x1 , x2 , . . . , xm ) = I ⊆{1,2,...,m} a I x I be a multilinear poly-
nomial,
where x I = i∈I x i . Let w( f ) = |{a I : a I = 0}| and let L 1 ( f ) =
I ⊆{1,2,...,m} |a I |.
Definition 6 Let H be a set system on the universe X with m × M incidence matrix

φ. Let f (x1 , x2 , . . . , xm ) = I ⊆{1,2,...,m} a I x I be a multilinear polynomial with
nonnegative integer coefficients or coefficients from Zr . Then f (Hφ ) is a hypergraph
on the L 1 ( f )-element vertex set, and its incidence matrix is the L 1 ( f )× M matrix ψ.
The rows of ψ correspond to x I ’s of f ; there are a I identical rows of ψ corresponding
to the same x I . The row, corresponding to x I is defined as the dream product of those
rows of φ that correspond to vi , i ∈ I .
Lemma 1 [16] Suppose in the Definition 6, the coefficients of x1 , x2 , . . . , xm are

nonzeros in f . Then the resulting Hypergraph f (Hφ ) is a set system [16].
The most remarkable property of f (Hφ ) is given by the following theorem:

Theorem 1 [16] Let H = {H1 , H2 , . . . , HM } be a set system and φ its
m × M incidence matrix. Let f be a multilinear polynomial with nonnegative
1 , H
integer coefficients or coefficients from Zr . Let f (Hφ ) = { H 2 , . . . , H
M }.
Then for any 1 ≤ k ≤ M and for any 1 ≤ i 1 < i 2 < · · · < i k ≤ M :
i1 ∩ H
f (Hi1 ∩ Hi2 ∩ . . . ∩ Hik ) = | H i2 ∩ . . . H
ik |.
Following theorem discusses the construction of a new set system from a given set
system using the Definition 6, Lemma 1 and Theorem 1.
Theorem 2 [16] Let f be an m—variable symmetric polynomial with nonnegative
integer coefficients, and H a set system of size M on the m element universe with
m × M incidence matrix φ. Suppose that
L(H ) = {|Hi ∩ H j |, Hi = H j , Hi , H j ∈ H } = {l1 , l2 , . . . , ls }.
Then one may construct in O(L 1 ( f )m M) time a hypergraph f (Hφ ) of size M on

the L 1 ( f )—vertex universe such that the sizes of the pairwise intersections of the
sets of f (Hφ ) are
f (l1 ), f (l2 ), . . . , f (ls ).
3.1 Set Systems for Designing CS Matrices
Using the afore-stated results [16] from set system theory, we construct a new binary
sensing matrix from a given binary sensing matrix. The new matrix has small density
as compared to the given one.

Theorem 3 Suppose f (x1 , x2 , . . . , xm ) = x1 + x2 + · · · + xm + i< j xi x j is a
symmetric polynomial. Let φ be a binary sensing matrix of size m × M such that
m(m+1)
2 < M with the coherence being at most rk . Here k represents the number
of nonzero elements that each column of φ has. Then there exists a binary sensing
r +(r2)
matrix ψ of size m(m+1) × M whose coherence is at most .
2 k+(k2)
272 R.R. Naidu
Proof Define H = {Hi : 1 ≤ i ≤ M, Hi = supp(φi ), where φi is ith column of φ}.

Since all columns of φ are distinct, H is a set system. Let L(H ) = {|Hi ∩ H j |, Hi =
H j , Hi , H j ∈ H } = {l1 , l2 , . . . , ls }. Since the coherence of the matrix φ is at most
r
k , the cardinality of overlap between the supports of any two columns is at most r .
Consequently, li ≤ r for all i. Let X = {1, 2, . . . , m}.
Since f (x1 , x2 , . . . , xm ) = x1 + x2 + · · · + xm + i< j xi x j is a symmetric

polynomial and H is a set system, we have L 1 ( f ) = m + m2 and f (Hφ ) is a set
system of size M on L 1 ( f )-element universe, from Theorem 2 and Lemma 1.
Define (vi j )m×1 to be the characteristic vector on Hi ∩ H j in the universe
X . Since
each ls = |Hi ∩ H j | for some i = j, f (ls ) = f ((vi j )) ≤ f (r ) = r + r2 . It follows

that f (li ) ≤ r + r2 for all i. Therefore, the sizes of the pairwise intersections of the

sets of f (Hφ ) is at most r + r2 . The incidence matrix ψ of the set system f (Hφ )

is of size L 1 ( f ) × M, that is, (m + m2 ) × M. From the hypothesis of the theorem

m + m2 < M, so it is an underdetermined system and its first m rows are same

as φ and remaining m2 rows are the dream products of first m rows. Each column
k
in ψ contains k + 2 number of ones. The cardinality of overlap between any two

columns is at most r + r2 . It follows that coherence of the matrix ψ is μ(ψ), which
r +(2)r
is at most equal to .
k+(k2)
The following theorem concludes that the matrix ψ0 = 1 ψ defined is RIP

k+(2k )
compliant.

r +(r2)
Theorem 4 The matrix ψ0 = 1 ψ has the RIP with δ = (k − 1)
k+(2k ) k+(k2)
k+( k
)
whenever k − 1 < r +(
2
.
r
2 )
Proof Proof follows from the Proposition 1 and Theorem 3
k+( k
)
Remark 1 The density of the new matrix ψ is m+(
2
, which is smaller than k
m, the
m
2 )
density of φ.
As the sensing matrices of higher sparsity (or lower density) have potential for fast
processing, the construction of such matrices is of relevance. In the present work,
we have used the ideas from the set system theory and have showed that a CS matrix
of higher sparsity can be generated from a given binary CS matrix.
Acknowledgments The author gratefully acknowledges the support (Ref No. 20-6/2009(i)EU-IV)
that he receives from UGC, Govt of India. Author is thankful to Dr. C. S. Sastry and Dr. Phanindra
Varma for fruitful discussions.
References
1. Donoho, D.L., Elad, M., Temlyakov, V.N.: Stable recovery of sparse overcomplete represen-
tations in the presence of noise. IEEE Trans. Inf. Theory 52, 6–18 (2006)
2. Candes, E.: The restricted isometry property and its implications for compressed sensing.
Comptes Rendus Mathematique 346, 589–592 (2008)
3. Cohen, A., Dahmen, W., DeVore, R.: Compressed sensing and best k-term approximation. J.
Amer. Math. Soc. 22(1) 211–231 (2009)
4. DeVore, R.A.: Deterministic constructions of compressed sensing matrices. J. Complex. 23,
918–925 (2007)
5. Li, S., Gao, F., Ge, G., Zhang, S.: Deterministic construction of compressed sensing matrices
via algebraic curves. IEEE Trans. Inf. Theory 58, 5035–5041 (2012)
6. Amini, A., Marvasti, F.: Deterministic construction of binary, bipolar and ternary compressed
sensing matrices. IEEE Trans. Inf. Theory 57, 2360–2370 (2011)
7. Indyk, P.: Explicit constructions for compressed sensing matrices. In: Proceedings 19th Annual
ACM-SIAM Symposium Discrete Algorithms, pp. 30–33 (2008)
8. Kashin, B.S., Temlyakov, V.N.: A remark on compressed sensing. Math. Notes 82(6), 748–755
(2007)
9. Bourgain, J., Dilworth, S., Ford, K., Konyagin, S., Kutzarova, D.: Explicit constructions of RIP
matrices and related problems. Duke Math. J. 159, 145–185 (2011)
10. Baraniuk, R., Davenpor, M., De Vore, R., Wakin, M.: A simple proof of the restricted isometry
property for random matrices. Constr. Approx. 28(3), 253–263 (2008)
11. Garnaev, A., Gluskin, E.: The widths of a Euclidean ball. Dokl. Akad. Nauk USSR 277, 1048–
1052 (1984); English transl. in Soviet Math. Dokl. 30, 200–204 (1984)
12. Kashin, B.S.: Widths of certain finite-dimensional sets and classes of smooth functions. Izv.
Akad. Nauk SSSR, Ser. Mat. 41, 334–351 (1977). English transl. in Math. USSR IZV. 11,
317–333 (1978)
13. Gilbert, A., Indyk, P.: Sparse recovery using sparsematrices. Proc. IEEE 98(6), 937–947 (2010)
14. Fickus, M., Mixon, D.G., Tremain, J.C.: Steiner equiangular tight frames. Lin. Alg. Applicat.
436(5), 1014–1027 (2012)
15. Iwen, M.: Compressed sensing with sparse binary matrices: instance optimal error guarantees
in near-optimal time. J. Complex. 30(1), 1–15 (2014)
16. Grolmusz, V.: Constructing set-systems with prescribed intersection sizes. J. Algorithms 44,
321–337 (2001)
Topological and Nontopological
1-Soliton Solution of the Generalized
KP-MEW Equation
Amiya Das and Asish Ganguly
Abstract In this paper, we obtain the topological and nontopological 1-soliton

solution of the generalized Kadomtsev–Petviashvili modified equal width
(KP-MEW) equation. The use of solitary wave ansatz method in context of dou-
bly periodic Jacobi elliptic functions is done, which leads to the exact topological
and nontopological soliton solutions. The Jacobi elliptic function solution degen-
erates into solitary wave solution in the limiting case of the modulus parameter.
We derive the power law nonlinearity parameter domain for the existence of soliton
solution, which is different for the topological and nontopological soliton. Also we
identify the parametric restriction on the coefficients for the existence of solitary
wave solutions. Finally, the remarkable features of such solitons are demonstrated in
several interesting figures.
Keywords Topologican soliton · Nontopological soliton · KP-MEW equation ·

Jacobi elliptic functions
1 Introduction
Theory of nonlinear evolution equations (NLEE’s) has a remarkable interest in the

area of science and engineering, particularly in fluid dynamics, nonlinear optics, bio-
chemistry, geophysics, etc [1–3]. The studies of these nonlinear evolution equations
have attracted a great deal of attention as their closed form analytical solutions are
necessary to carry out further investigation to examine the physical properties of
these solutions. A bunch of powerful techniques have been introduced to carry out
A. Das (B) · A. Ganguly

IIT Kharagpur, Kharagpur 721302, India
e-mail: amiya620@gmail.com
A. Ganguly
e-mail: aganguly@maths.iitkgp.ernet.in

DOI 10.1007/978-81-322-2485-3_22
276 A. Das and A. Ganguly
the integration of various NLEE’s. A few of these techniques are tanh–sech method,
sine–cosine method, homogeneous balance method, Jacobi elliptic function method,
F-expansion method, (G /G)-expansion method [4–11], etc.
In 1965, Zabusky and Kruskal [12] discovered that the Korteweg de-Vries (KdV)
equation has a pulse-like solitary wave solution which interacts “elastically” with
another such solution. They termed such type of solution as soliton. In mathematics
and physics, a soliton is a self-reinforcing solitary wave (a wave packet or pulse) that
preserves its shapes while traveling at constant speed.
The delicate balance between nonlinearity and dispersion effects in the medium
causes envelope soliton which are stable nonlinear wave packets that maintains their
shapes during the propagation in a nonlinear dispersive medium [13]. Two different
types of envelope solitons, cnoidal (nontopological) and snoidal (topological) can
propagate in nonlinear dispersive media. The nontopological soliton is a pulse on
a zero intensity background which has no phase change for large spatial distance,
whereas the topological soliton appears as a intensity dip in an infinitely extended
constant background [14].
The prototype example of NLEE in the area of theoretical physics and applied
mathematics is the Korteweg de-Vries (KdV) equation
qt + aqqx + qx x x = 0, (a is a constant parameter) . (1)
A two-dimensional generalization of the KdV equation is Kadomtsev–Petviashvili

(KP) equation
(qt + aqqx + qx x x )x + q yy = 0, (2)
which depicts the evolution of quasi-one-dimensional shallow-water waves when

effects of the surface tension and the viscosity are negligible [15]. The modified
equal width (MEW) equation, which appears in many physical applications [16, 17]
is of the form
qt + a q 3 − bqx xt = 0 . (3)
x
In this work, our motivation is to seek the topological and nontopological soliton
solutions of the gKP-MEW equation, which is a generalized form of the MEW
equation in the KP sense along with the power law nonlinearity [5]

qt + a q n x + bqx xt x + cq yy = 0, (4)
where a, b and c are real-valued constants. The first term in (4) depicts the evolution
term, while the second term stands as the nonlinear term with the power law indicated
by the exponent n, and the third term utters the dispersion in the x-direction. The
fourth term, i.e., the y-dependence term is considered as a weak dependence on the
y-coordinate. The index n indicating the power law nonlinearity is a positive real
Topological and Nontopological 1-Soliton Solution … 277
number. In [18], the single peak solitary wave solutions of the generalized KP-MEW
(2, 2) equation is studied under an inhomogeneous boundary condition and different
form of solutions like peakons, compactons, cuspons, loop solitons, and smooth
solitons are analyzed by phase portrait analysis.
The rest of the paper is organized as follows. In Sect. 2, the nontopological soliton
solution in context of the doubly periodic Jacobi elliptic functions are obtained for
the gKP-MEW equation. In Sect. 3, the topological soliton solution in terms of the
doubly periodic Jacobi elliptic functions is obtained. We conclude in Sect. 4.
2 Nontopological Soliton
Nontopological solitons are also known as bell-shaped solitons or cnoidal waves

as there is no phase change for large spatial distance . The nontopological solitons
are used to impart loads of information across trans-continental and trans-oceanic
distances. In order to find the nontopological 1-soliton solution of (4), the solitary
wave ansatze is assumed in the form of Jacobi elliptic functions as follows [19–23]:
q(x, y, t) = A cn p (B1 x + B2 y − vt) (5)
where snτ ≡ sn(τ ; k), cnτ ≡ cn(τ ; k), dnτ ≡ dn(τ ; k) are three Jacobian elliptic
functions of real modulus k 2 (0 < k 2 < 1) and k 2 = 1 − k 2 is the complementary
modulus. Here A indicates the soliton amplitude, with B1 and B2 as the inverse widths
of the soliton along the x- and y- direction, respectively. Also v depicts the velocity
of the soliton and the unknown exponent p will be calculated during tracking the
solution of (4). To make the structure of the soliton more general, the inverse widths
of the soliton in the x- and y-directions should be taken different, namely B1 = B2 ,
in general. From (5), using the notation
τ = B1 x + B2 y − vt, (6)
it is possible to obtain qxt , (q n )x x , qx x xt , q yy and substituting them in the KPMEW

equation (4) yields

AB1 v − p( p − 1) 1 − k 2 cn p−2 − p 2 2k 2 − 1 cn p + p( p + 1)k 2 cn p+2

+ a An B12 np(np − 1) 1 − k 2 cnnp−2 + n 2 p 2 2k 2 − 1 cnnp − np(np + 1)k 2 cnnp+2
2
+ b AB13 v p( p − 1)( p − 2)( p − 3) 1 − k 2 cn p−4 − 2 p( p − 1) p 2 − 2 p + 2
2
× 1−k 2
2k − 1 cn
2 p−2
− 12 pk p − p + 1 k − 1 + p 2k − 1
2 2 2 4 2
cn p

+ 2 p( p + 1) p 2 + 2 p + 2 k 2 2k 2 − 1 cn p+2 − p( p + 1)( p + 2)( p + 3)k 4 cn p+4

+ c AB22 p( p − 1) 1 − k 2 cn p−2 + p 2 2k 2 − 1 cn p − p( p + 1)k 2 cn p+2 = 0. (7)
The value of p can be obtained by equating the exponents np and p + 2 as
2
p= , (8)
n−1
which can also be achieved by equating the exponents np + 2 and p + 4. Finally,

equating the coefficients of the linearly independent functions cn p+ j τ for j = 0, 2, 4
value of v and A are obtained as follows:

c(n − 1)2 2k 2 − 1 B22
v= , (9)
B1 (n − 1)2 2k 2 − 1 + 4bB12 + bB12 k 2 k 2 − 1 5n 3 − 32n 2 + 65n − 26

(n − 1)2 ck 2 B22 (n + 1) − 2 2k 2 − 1 an 2 B12 An−1
v= , (10)
B1 (n + 1)k 2 (n − 1)2 + 4bB12 n 2 + 1 2k 2 − 1
and
1
bcB22 k 2 (n + 1) k 2 k 2 − 1 5n 3 − 48n 2 − 42 + 65n − 4n 2 n−1
A= .
2an 2 (n − 1) 2k − 1 + 4bB1 + bB1 k k − 1 5n − 32n + 65n − 26
2 2 2 2 2 2 3 2
(11)
Now from (9) to (11), it is clear that the restrictions that must be imposed on the
parameters for the formation of the nontopological soliton solution are c > 0, and
ab < 0. It is required to be noticed that equating the coefficient of cn p+4 τ leads
to the relation between the amplitude and the inverse widths of the soliton given
in (11) and the two values of the velocity of the soliton are calculated equating the
coefficients of cn p τ and cn p+2 τ consequently. Now it is interesting to note that the
relation between amplitude and the inverse widths of the soliton (11) can also be
obtained by equating the two velocities of the soliton given by (9) and (10), which
sustains the consistency of the method of solution. Thus the elliptic periodic solution
of the gKP-MEW equation with power law nonlinearity is given by
2
q(x, y, t) = A cn n−1 (B1 x + B2 y − vt), (12)
where the amplitude A is given by (11) dependent on the inverse widths B1 and B2
and the velocity v of the soliton, given by (9) or (10).
Finally, consider the limits k → 1, cnτ → sechτ , the Jacobi elliptic periodic
solution (12) degenerate into the nontopological 1-soliton solution
A
q(x, y, t) = 2
, (13)
cosh n−1 (B1 x + B2 y − vt)
where the relation between the amplitude and inverse widths and the velocity of the
soliton becomes
c(n − 1)2 B22

v= , (14)
B1 (n − 1)2 + 4bB12

(n − 1)2 cB22 (n + 1) − 2an 2 B12 An−1
v= , (15)
B1 (n + 1) (n − 1)2 + 4bB12 (n 2 + 1)
and
1
2bc(n + 1)B22 n−1
A= − . (16)
a (n − 1)2 + 4bB12
The numerical simulation of the solution (13) with particular choice of parameters
is shown in Fig. 1.
Fig. 1 Profile of the nontopological solitary wave at a t = −10, b t = 0, c t = 10 for B1 = 1,

B2 = −1, a = 0.2, b = −0.1, c = 0.5, and n = 2
3 Topological Soliton
The topological solitons which are also known as snoidal waves are supported by
the gKP-MEW equation under certain restictions. In case of topological solitons,
the phase changes for large spatial distance. One of the essential difference between
nontopological and topological solitons consists of the existence of multiple bound
states that can form nontopological solitons in clear contrast with topological soli-
tons [25]. To start off finding the topological soliton solution of the gKP-MEW equa-
tion, the starting hypothesis is taken in context of Jacobi elliptic function as follows
[22–24]:
q(x, y, t) = A sn p (B1 x + B2 y − vt) (17)
where A, B1 and B2 represent the soliton amplitude and the inverse widths of the soli-
ton, respectively, and v depicts the velocity of the soliton. The value of the unknown
exponent p will be determined during the course of derivation of the topological
soliton solution of (4). From (17), by utilizing the notation
τ = B1 x + B2 y − vt (18)
one can easily obtain qxt , (q n )x x , qx x xt , q yy and substituting them into (4) gives

AB1 v − p( p − 1)sn p−2 + p 2 1 + k 2 sn p − p( p + 1)k 2 sn p+2

+ a An B12 np(np − 1)snnp−2 − n 2 p 2 1 + k 2 snnp + np(np + 1)snnp+2

+ b AB13 v − p( p − 1)( p − 2)( p − 3)sn p−4 + 2 1 + k 2 p( p − 1) p 2 − 2 p + 2 sn p−2

− p 2 p 2 + 4 p 2 k 2 + p 2 k 4 + 10k 2 sn p + 2k 2 1 + k 2 p( p + 1) p 2 + 2 p + 2 sn p+2

− k 4 p( p + 1)( p + 2)( p + 3) + c AB12 − p 2 1 + k 2 sn p + p( p − 1)sn p−2

+ p( p + 1)k 2 sn p+2 = 0 (19)
Equating the coefficients np and p + 2 or np + 2 and np + 4, the value of p is

obtained as
2
p= . (20)
n−1
The same value of p can also be obtained by equating the exponents np and p + 2
and the exponent pairs np − 2 and p. Note that, there are five linearly independent
functions are there in (19), which are namely, sn p+ j τ for j = −4, −2, 0, 2, 4. Hence,
each of the coefficients of these linearly independent functions of (19) must be zero.
It is to be noticed that the function sn p−4 τ stands alone linearly independent and its
coefficient must vanish. Also, the common factor of all the coefficients of sn p−2 τ is
( p − 1). Hence, the coefficients of these two linearly independent functions produces
p = 1, (21)
and
n = 3. (22)
Thus from (22), it can be concluded that for the gKP-MEW equation, topological
soliton exist only for n = 3, which is the KP-MEW equation. Hence, if n = 3,
topological soliton do not exist for gKP-MEW equation, an important observation.
Finally, setting the coefficients of the other linearly independent functions namely
sn p+ j τ for j = 0, 2, 4 to zero, we obtain

cB22 1 + k 2 − 6a A2 B12
v = 2 , (23)
B1 1 + k 2 − bB13 1 + 14k 2 + k 4

9a A2 B12 1 + k 2 − 2ck 2 B22
v= , (24)
20bB13 k 2 1 + k 2 − 2k 2 B1
a A2
v= . (25)
2bB1 k 4
and the value of A is obtained as

2bck 2 B22
A= , (26)
a 1 − bB12 − bB12 k 2
by equating the values of v in (23) and (24), which shows the relation between
the amplitude and the inverse widths of the soliton. From (26), it is clear that the
constraint relation
ab > 0 and c > 0 (27)
has to be sustain for the existence of the Jacobi elliptic periodic solution. Hence the
doubly periodic Jacobi elliptic solution for the KP-MEW equation
(qt + 3aq 2 qx + bqx xt )x + cq yy = 0 (28)
is given by
q(x, y, t) = A sn(B1 x + B2 y − vt) (29)
where the velocity of the soliton is given by (23) or (24) or (25). The relation between
the free parameter and the inverse widths of the soliton is shown in (26) and the
restriction on the coefficients of the dispersion term in x-direction and the weak
dependence term along the y-coordinate for the existence of the doubly periodic
Jacobi elliptic solution of the KP-MEW equation is conveyed in (27).
Finally, consider the limits k → 1, snτ → tanhτ , the elliptic periodic solution
(29) degenerate into the topological 1-soliton solution
q(x, y, t) = A tanh(B1 x + B2 y − vt), (30)
with velocity as
cB22 − 3a A2 B12
v= , (31)
B1 − 8bB13
9a A2 B12 − cB22
v= , (32)
20bB13 − B1
Fig. 2 Profile of the topological solitary wave at a t = −10, b t = 0, c t = 10 for B1 = 1,

B2 = −1, a = 0.2, b = 0.1, c = 0.3
Fig. 3 Contour plot of the a nontopological and b topological soliton with same parameters at
t =0
and
a A2
v= . (33)
2bB1
Now, equating the three values of the velocity given by (31), (32) and (33), the
relation between the free parameters A and B1 , B2 is obtained as follows

2bcB22
A= . (34)
a 1 − 2bB12
The existence of the soliton solution requires the same constraint relation (27). The
numerical simulation of the solution (30) with particular choice of parameters is
shown in Fig. 2. The contour plot corresponding to the nontopological and topological
soliton solutions are demonstrated in Fig. 3.
4 Conclusions
In this paper, the solitary wave ansatz method is exploited in the context of doubly
periodic Jacobi elliptic functions to carry out the 1-soliton solution of the gKP-MEW
equation. The elliptic function solution degenerates into the solitary wave solution
in the limiting case of the elliptic modulus parameter. Both topological (snoidal) and
nontopological (cnoidal) solitons are studied. An interesting fact is observed that
the topological soliton for the gKP-MEW equation exist only for n = 3, or in other
words for the KP-MEW equation only, which is reported in the literature for the first
time as far as we know. In future, further aspects of this problem can be studied; such
as addition of perturbation term, self-steeping term, soliton–soliton interaction and
others.
Acknowledgments The authors acknowledge anonymous referees for critical comments and men-
tioning some useful references.
Appendix
Here we will furnish a primary introduction about elliptic functions (for more details,
see [26, 27]). Three Jacobian elliptic functions are defined as
sn(x, k) = sin ϕ, cn(x, k) = cos ϕ, dn(x, k) = dϕ/dx (35)
where the amplitude function ϕ(z, k) is defined by the integral

ϕ dτ
z(ϕ, k) = (36)
0 1 − k 2 sin2 τ
The square of the real number k is called elliptic modulus parameter and k 2 ∈ (0, 1).
Also k 2 = 1−k 2 is called complementary modulus parameter. To avoid the complex-
ity, in the text we inhibit the explicit modular dependence and write snx, cnx, dnx etc.
These are doubly periodic functions of periods 4K , 2i K ; 4K , 4i K , and 2K , 4i K ,
respectively, where the quarter-periods K and K are the real numbers given by
K (k) ≡ K = z(π/2, k), K (k) ≡ K = K (k ) (37)
K is called complete elliptic integral of second kind. Some useful relations are

sn2 x + cn2 x = 1, dn2 x + k 2 sn2 x = 1, k 2 cn2 x − 1 = dn2 x − 1 (38)
the rules of differentiation are
sn x = cnx dnx, cn x = −snx dnx, dn x = −k 2 snx cnx (39)
and

k→1 tanhx k→1 sechx k→1 sechx
sn(x, k) −−−→ , cn(x, k) −−−→ , dn(x, k) −−−→ (40)
k→0 sinx k→0 cosx k→0 1
References
1. Pelinovsky, D.: Intermediate nonlinear Schrödinger equation for internal waves in a fluid of
finite depth. Phys. Lett. A 197, 401–406 (1995)
2. Hasegawa, A.: Plasma Instabilities and Nonlinear Effects. Springer, Berlin (1975)
3. Gray, P., Scott, S.: Chemical Oscillations and Instabilities. Clarendon, Oxford (1990)
4. Malfliet, W., Hereman, W.: The tanh method. I: exact solutions of nonlinear evolution and wave
equations. Phys. Scr. 54, 563–568 (1996)
5. Wazwaz, A.M.: The tanh method for travelling wave solutions of nonlinear equations. Appl.
Math. Comput. 154, 713–723 (2004)
6. Fan, E., Zhang, H.: A note on the homogeneous balance method. Phys. Lett. A 246, 403–406
(1998)
7. Fan, E., Zhang, J.: Application of the Jacobi elliptic function method to special-type nonlinear
equations. Phys. Lett. A 305, 383–392 (2002)
8. Bagchi, B., Das, S., Ganguly, A.: New exact solutions of a generalized shallow water wave
equation. Phys. Scr. 82, 025003 (2010)
9. Wang, M.L., Li, X.Z., Zheng, J.L.: The G /G -expansion method and travelling wave solu-
tions of nonlinear evolution equations in mathematical physics. Phys. Lett. A 372, 417–423
(2008)
10. Das, A., Ganguly, A.: Some new exact solutions
of the (2+1) dimensional potential Kadomstev-
Petviashvili equation by the extended G /G -expansion method. Int. J. Nonlinear Sci. 14,
86–94 (2012)
11. Das, A., Ganguly, A.: A new variation of the (G /G)-expansion method: traveling wave solu-
tions to nonlinear equations. Int. J. Nonlinear Sci. 17, 268–280 (2014)
12. Zabusky, N.J., Kruskal, M.D.: Interactions of solitons in a collisionless plasma and the recur-
rence of initial states, Phys. Rev. Lett. 15, 240–243 (1965)
13. Bagada, A.V., Melkov, G.A., Serga, A.A., Slavin, A.N.: Parametric interaction of dipolar spin
wave solitons with localized electromagnetic pumping. Phys. Rev. Lett. 79, 21372140 (1997)
14. Scott, M.M., Kostylev, M.P., Kalinikos, B.A., Patton, C.E.: Excitation of bright and dark enve-
lope solitons for magnetostatic waves with attractive nonlinearity. Phys. Rev. B 71, 1–4 (2005)
15. Kadomtsev, B.B., Petviashvili, V.I.: On the stability of solitary waves in weakly dispersive
media. Soviet. Phys. Dokl. 15, 539–541 (1970)
16. Biswas, A.: Soliton perturbation theory for the generalized equal width equation. Pac. J. Appl.
Math. 1, 99–103 (2008)
17. Wazwaz, A.M.: The tanh and sine-cosine methods for a reliable treatment of the modified equal
width equation and its variants. Commun. Nonlinear Sci. Num. Simul. 11, 148–160 (2006)
18. Wei, M., Tang, S., Fu, H., Chen, G.: Single peak solitary wave solutions for the generalized
KP-MEW (2, 2) equation under boundary condition. Appl. Math. Comput. 219, 8979–8990
(2013)
19. Ismail, M.S., Petkovic, M.D., Biswas, A.: 1-soliton solution of the generalized KP equation
with generalized evolution. Appl. Math. Comput. 216, 2220–2225 (2010)
20. Biswas, A.: 1-soliton solution of the generalized camassaHolm Kadomtsev-Petviashvili equa-
tion. Commun. Nonlinear Sci. Num. Simul. 14, 2524–2527 (2009)
21. Biswas, A., Ranasinghe, A.: 1-soliton solution of Kadomtsev-Petviashvili equation with power
law nonlinearity. Appl. Math. Comput. 214, 645–647 (2009)
22. Ebadi, G., Fard, N.Y., Bhrawy, A.H., Kumar, S., Triki, H., Yildirim, A., Biswas, A.: Solitons
and other solutions to the (3 + 1)-dimensional extended Kadomtsev-Petviashvili equation with
power law nonlinearity. Rom. Rep. Phys. 65, 27–62 (2013)
23. Biswas, A., Triki, H., Labidi, M.: Bright and dark solitons of the Rosenau-Kawahara equation
with power law nonlinearity. Phys. Wave Phenom. 19, 24–29 (2011)
24. Biswas, A., Ranasinghe, A.: Topological 1-soliton solution of Kadomtsev-Petviashvili equation
with power law nonlinearity. Appl. Math. Comput. 217, 1771–1773 (2010)
25. Palacios, S.L., Fernandez-Diaz, J.M.: Black optical solitons for media with parabolic nonlin-
earity law in the presence of fourth order dispersion. Opt. Commun. 178, 457–460 (2000)
26. Gradshteyn, I.S., Ryzhik, I.M.: Tables of Integrals, Series and Products. Academic Press,
New York (1963)
27. Abramowitz, M., Stegun, I.A. (eds.): Handbook of Mathematical Functions. Dover Publications
INC, New York (1965)
Mild Solutions for Impulsive Functional
Differential Equations of Order α ∈ (1, 2)
Ganga Ram Gautam and Jaydev Dabas
Abstract In this research paper, first we develop the definition of mild solutions
for impulsive fractional differential equations of order α ∈ (1, 2). Second, we study
the uniqueness result of mild solutions for impulsive fractional differential equation
with state-dependent delay by applying fixed point theorem and solution operator. At
last, we present an example to illustrate the uniqueness result using fractional partial
derivatives.
Keywords Fractional order differential equation · Functional differential equa-

tions · Impulsive conditions · Fixed point theorem
1 Introduction
In this research paper, we consider the following impulsive fractional differential

equation with state-dependent delay of the form
C
Dtα u(t) = Au(t) + f (t, u ρ(t,u t ) ), t ∈ J = [0, T ], t = tk , (1)
u(t) = φ(t), t ∈ (−∞, 0], u (0) = u 1 ∈ X, (2)
Δu(tk ) = Ik (u(tk− )), Δu (tk ) = Q k (u(tk− )), k = 1, 2, ...m, (3)
where C Dtα is the Caputo’s fractional derivative of order α ∈ (1, 2), u is ordinary
derivative with respect to t and J is operational interval. A : D(A) ⊂ X → X
is the sectorial operator defined on a complex Banach space X . The functions
G.R. Gautam (B) · J. Dabas

Department of Applied Science and Engineering, IIT Roorkee,
Saharanpur Campus, 247001 Saharanpur, India
e-mail: gangaiitr11@gmail.com
J. Dabas
e-mail: jay.dabas@gmail.com

DOI 10.1007/978-81-322-2485-3_23
288 G.R. Gautam and J. Dabas
f : J × Bh → X, ρ : J × Bh → (−∞, T ] and φ ∈ Bh are given and

satisfies some assumptions, where Bh is introduced in Sect. 2. The history func-
tion u t : (−∞, 0] → X is defined by u t (θ ) = u(t + θ ), θ ∈ (−∞, 0] belongs
to Bh . Here 0 = t0 < t1 < · · · < tm < tm+1 = T < ∞ and the functions
Ik , Q k ∈ C(X, X ), k = 1, 2, ...m, are bounded. We have u(tk ) = u(tk+ ) − u(tk− )
where u(tk+ ) and u(tk− ) represent the right- and left-hand limits of u(t) at t = tk ,
also we take u(tk− ) = u(tk ). Furthermore, u (tk ) = u (tk+ ) − u (tk− ) where u (tk+ )
and u (tk− ) represent the right- and left-hand limits of u (t) at t = tk , also we take
u (tk− ) = u (tk ), respectively.
Impulsive differential equations with fractional order (see for fractional calcu-
lus [15, 16, 18–20]) are paying attention by many researchers because the model
processes which are subjected to abrupt changes cannot described by ordinary dif-
ferential equations, so such type equations are modeled in term of impulse. The most
important applications of these equations are in the ecology, mechanics, electrical,
and medicine biology. On the other hand, functional differential equations originate
in several branches of engineering, applied mathematics, and science. Recently, frac-
tional functional differential equations with state-dependent delay seems frequently
in many fields as modeling of equations, panorama of natural phenomena, and porous
media. See for more details of the relevant development theory in the cited papers
[1, 2, 4–9, 11, 13].
In our survey, we found that Feckan et al. [12] gave the new concept of solu-
tion for impulsive nonlinear fractional differential equation order α ∈ (0, 1). Wang
et al. [22] defined the definition of mild solution using the probability density func-
tion for the impulsive fractional evolution equation of order α ∈ (0, 1). By Motivated
work [22], Dabas and Chauhan [10] defined the mild solution for neutral impulsive
fractional functional differential equation of order α ∈ (0, 1) using analytic operator
theory. Wang et al. [23] extended the problem, consider in paper [12] for of order
α ∈ (1, 2). Shu et al. [21] introduced the definition of mild solution for fractional
differential equations with nonlocal conditions of order α ∈ (1, 2) without impulse.
We found that there is no literature available on mild solution for impulsive fractional
functional differential equation of order α ∈ (1, 2).
To fill this gap and inspired by the above-mentioned work [10, 12, 21–23], we
develop the definition of mild solution for the problem (1)–(3) and show the existence
result. For further details, this work has four sections, Sect. 2 provides some basic
definitions, preliminaries, theorems, and lemmas. The Sect. 3 is equipped with main
results for the considered problem (1)–(3) and in Sect. 4 an example is considered.
2 Preliminaries and Background Martials
Let (X, · X ) be a complex Banach space of functions with the norm u X =

supt∈J {|u(t)| : u ∈ X } and L(X ) denotes the Banach space of bounded linear
operators from X into X equipped with norm is denoted by · L(X ) .
Mild Solutions for Impulsive Functional Differential … 289
For the analysis of the infinite delay, we shall use abstract phase space Bh as
defined in [14] details are as follow: 0
Let h : (−∞, 0] → (0, ∞) be a continuous function with l = −∞ h(s)ds
< ∞, s ∈ (−∞, 0]. For any a > 0, we define space
B = {ψ : [−a, 0] → X such that ψ(t) is bounded and measurable} ,
equipped with the norm ψ [−a,0] = sups∈[−a,0] ψ(s) X , ∀ ψ ∈ B. Let us define

abstract space as
0
Bh = ψ : (−∞, 0] → X, s.t. for any a ≥ c > 0, ψ |[−c,0] ∈ B h(s) ψ [s,0] ds < ∞ .
−∞
0
If Bh is endowed with the norm ψ Bh = −∞ h(s) ψ [s,0] ds, ∀ ψ ∈ Bh , then
it is clear that (Bh , · Bh ) is a complete Banach space. Let
Ct1 ([0, T ], X ) = C 1 ([0, t]; X ), 0 < t ≤ T < ∞,
be a Banach space of all functions u : [0, T ] → X such that u is continuously

differentiable on [0, T ] endowed with the norm
⎧ ⎫
⎨ 1 ⎬
u C 1 = sup u j (t) X , u ∈ Ct1 .
t∈[0,T ] ⎩ ⎭
t
j=0
To use the impulsive condition with infinite delay, we consider a Banach space
Bh := PC 1 ((−∞, T ]; X ), T < ∞,
formed by all functions u : (−∞, T ] → X such that u is continuously differentiable

on [0, T ] except for a finite number of points ti ∈ (0, T ), i = 1, 2, . . . , N, at which

u (ti+ ) and u (ti− ) = u (ti ) exist and endowed with the seminorm · B in Bh
h

u B = sup{ u C 1 : 0 ≤ t ≤ T } + φ Bh , u ∈ Bh .
h t
For a function u ∈ Bh and i ∈ {0, 1, . . . , N }, we introduce the function

ū i ∈ C 1 ((ti , ti+1 ]; X ) given by

u (t), for t ∈ (ti , ti+1 ],
ū i (t) =
u (ti+ ), for t = ti .
Let u : (−∞, T ] → X be the function such that u 0 = φ, u | Jk ∈ C 1 (Jk , X ) then for

all t ∈ Jk , the following conditions hold:
(C1 ) u t ∈ Bh .
(C2 ) u(t) X ≤ H u t Bh .
(C3 ) u t Bh ≤ K (t) sup { u(s) : 0 ≤ s ≤ t} + M(t) φ Bh , where H > 0 is
constant; K , M : [0, ∞) → [0, ∞), K (·) is continuous, M(·) is locally
bounded and K , M are independent of u(t).
(C4φ ) The function t → φt is well-defined and continuous from the set
(ρ − ) = {ρ(s, ψ) : (s, ψ) ∈ [0, T ] × Bh }
into Bh and there exist a continuous and bounded function J φ : (ρ − ) →

(0, ∞) such that φt Bh ≤ J φ (t) φ Bh for every t ∈ (ρ − ).
Lemma 1 ([5]) Let u : (−∞, T ] → X be function such that u 0 = φ, u | Jk ∈

C 1 (Jk , X ) and if (C4φ ) hold, then

u s Bh ≤ (Mb + J φ ) φ Bh + K b sup u(θ) ; θ ∈ [0, max{0, s}] , s ∈ (ρ − ) ∪ Jk ,
where J φ = supt∈ (ρ − ) J φ (t), Mb = sups∈[0,T ] M(s) and K b = sups∈[0,T ] K (s).
Definition 1 Caputo’s derivative of order α > 0 with lower limit a, for a function
f : [a, ∞) → R such that f ∈ C n ([a, ∞), R) is defined as
t
C α 1
a Dt f (t) = (t − s)n−α−1 f (n) (s)ds =a Jtn−α f (n) (t),
Γ (n − α) a
where a ≥ 0, n − 1 < α < n, n ∈ N.
Definition 2 The Riemann–Liouville fractional integral operator of order α > 0

with lower limit a, for a continuous function f : [a, ∞) → R such that f ∈
1 ([a, ∞), R) is defined by
L loc
t
α 1
a Jt f (t) = f (t),
0
a Jt f (t) = (t − s)α−1 f (s)ds, t > 0,
Γ (α) a
where a ≥ 0 and Γ (·) is the Euler gamma function.
Definition 3 ([21]) Let A : D(A) ⊆ X → X be a densely defined, closed, and

linear operator in X . A is said to be sectorial of the type (M, θ, α, μ) if there exist
μ ∈ R, θ ∈ ( π2 , π ), M > 0, such that such that the α-resolvent of A exists outside
the sector and following two conditions are satisfied
(1) μ + Sθ = {μ + λα : λ ∈ C , |Arg(−λα )| < θ },
(2) (λα I − A)−1 L(X ) ≤ |λαM−μ| , λ ∈
/ μ + Sθ ,
where X is the complex Banach space with norm denoted · X .
Definition 4 ([19]) A two parameter function of the Mittag-Leffler type is defined

by the series expansion and integral form
∞

yk 1 μα−β eμ
E α,β (y) = = dμ, α, β > 0, y ∈ C,
Γ (αk + β) 2π ι c μα − y
k=0
1
where c is a contour which starts and ends at −∞ and encircles the disk |μ| ≤ |y| α
counter clockwise.
The Laplace integral of this function given by
∞ λα−β 1
e−λt t β−1 E α,β (ωt α )dt = α
, Reλ > ω α , ω > 0.
0 λ −ω
From paper [17], putting β = 1, ω = A and using the sign ÷ for the juxtaposition
of a function depending on t with its Laplace transform depending on λ, we get the
following Laplace transform pairs
λα−1 1
Sα (t) = E α (At α ) ÷ α
, Reλ > A α .
λ I−A
More general Laplace transform pairs with integral
j λα− j−1
0 Jt Sα (t) ÷ , j = 0, 1.
λα I − A
Definition 5 ([2]) Let A be a closed and linear operator with the domain D(A)
defined in a Banach space X and α > 0. We say that A is the generator of a solution
operator if there exist ω ≥ 0 and a strongly continuous function Sα : R+ → L(X ),
such that {λα : Reλ > ω} ⊂ ρ(A) and
∞
λα−1
x= eλt Sα (t)xdt, Reλ > ω, x ∈ X.
λα I − A 0
In this case, Sα (t) is called the solution operator generated by A.
Definition 6 ([3]) Let A be a closed and linear operator with domain D(A) defined
on a Banach space X. Let ρ(A) be the resolvent set of A, we call A is the generator
of an α-resolvent family if there exists ω ≥ 0 and a strongly continuous function
Tα : R+ → L(X ) such that {λα : Reλ > ω} ⊂ ρ(A) and
∞
α −1
(λ I − A) x= e−λt Tα (t)xdt, Reλ > ω, x ∈ X.
0
In this case, Tα (t) is called α-resolvent family generated by A.

Lemma 2 Let f be a continuous function and A be a sectorial operator of the type

(M, θ, α, μ). Consider following differential equation of order α ∈ (1, 2)
C
Dtα u(t) = Au(t) + f (t), t ∈ J = [0, T ], t = tk , (4)

u(0) = u 0 ∈ X, u (0) = u 1 ∈ X, (5)
Δu(tk ) = Ik (u(tk− )), Δu (tk ) = Q k (u(tk− )), t = tk , k = 1, 2, ...m. (6)
Then a function u(t) ∈ PC 1 ([0, T ], X ) is a solution of the system (4)–(6) if it satisfies

following integral equation
⎧ t t
⎪
⎨ Sα (t)u 0 + u 1 0 Sα (s)ds + 0 Tα (t − s) f (s)ds, t ∈ (0, t1 ]
k
u(t) = Sα (t)u 0 + K α (t)u 1 + i=1 Sα (t − ti )Ii (u(ti− ))
⎪
⎩ + k Q u(t − ) t t
i=1 i i ti Sα (s − ti )ds + 0 Tα (t − s) f (s)ds, t ∈ (tk , tk+1 ],
where Sα (t) and Tα (t) are operators generated by A and defined as

1 1
Sα (t) = eλt λα−1 (λα I − A)−1 dλ; Tα (t) = eλt (λα I − A)−1 dλ,
2πi Γ 2πi Γ
and Γ is a suitable path such that λα ∈

/ μ + Sθ for λ ∈ Γ.
Proof If t ∈ (0, t1 ], we have following problem

C
Dtα u(t) = Au(t) + f (t), (7)
u(0) = u 0 , u (0) = u 1 . (8)
By Lemma 3.1 in [23], the solution of Eqs. (7)–(8), we get

t (t − s)α−1 t (t − s)α−1
u(t) = u 0 + u 1 t + Au(s)ds + f (s)ds. (9)
0 Γ (α) 0 Γ (α)
If t ∈ (tk , tk+1 ], k = 1, 2, ...m, we have the following problem

C
Dtα u(t) = Au(t) + f (t), (10)
u(tk+ ) = u(tk− ) + Ik (u(tk− )), (11)
u (tk+ ) = u (tk− ) + Q k (u(tk− )). (12)
By Lemma 3.1 in [23] the solution of Eqs. (10)–(12), we get

k
k
u(t) = u 0 + u 1 t + Ii (u(ti− )) + Q i (u(ti− ))(t − ti )
i=1 i=1
t
t (t − s)α−1 (t − s)α−1
+ Au(s)ds + f (s)ds. (13)
0 Γα 0 Γ (α)
Summarizing Eqs. (9) and (13) to t ∈ (0, T ], we get

m
m
u(t) = u 0 + u 1 t + χti (t)Ii (u(ti− )) + χti (t)Q i (u(ti− ))(t − ti )
i=1 i=1

t (t − s)α−1 t (t − s)α−1
+ Au(s)ds + f (s)ds, (14)
0 Γα 0 Γ (α)
where

0 t ≤ ti
χti (t) =
1 t > ti .
By taking the Laplace transformation on Eq. (14), we have
u 1 e−λti e−λti
m m
u0
L{u(t)} = + 2+ Ii (u(ti− )) + Q i (u(ti− ))
λ λ λ λ2
i=1 i=1
A 1
+ α L{u(t)} + α L{ f (t)}. (15)
λ λ
On simplifying Eq. (15), we get
λα−1 (u 0 ) λα−2 (u 1 ) λα−1 m

L{u(t)} = α
+ α + e−λti Ii (u(ti− ))
(λ I − A) (λ I − A) (λα I − A)
i=1

m
λα−2 1
+ e−λti Q i (u(ti− )) + L{ f (t)}. (16)
(λα I − A) (λα I − A)
i=1
Now, taking the inverse Laplace transformation of Eq. (16), we have

t
m
u(t) = Sα (t)u 0 + u 1 Sα (s)ds + χti (t)Ii (u(ti− ))Sα (t − ti )
0 i=1

m t t
+ χti (t)Q i (u(ti− )) Sα (s − ti )ds + Tα (t − s) f (s)ds, t ∈ J.
i=1 ti 0
This complete the proof of the lemma.
Now, we state the definition of mild solutions of problem (1)–(3) by Lemma 2.

Definition 7 A function u : (−∞, T ] → X such that u ∈ Bh , u(0) = φ(0),
u (0) = u 1 , is called a mild solution of problem (1)–(3) if it satisfies the following
integral equation
⎧ t t
⎪
⎨ Sα (t)φ(0) + u 1 0 Sα (s)ds + 0 Tα (t − s) f (s, u ρ(s,u s ) )ds, t ∈ (0, t1 ]
Ii (u(ti− ))Sα (t − ti )
t k
u(t) = Sα (t)φ(0) + u 1 0 Sα (s)ds + i=1
⎩ + k Q (u(t − )) t S (s − t )ds + t T (t − s) f (s, u
⎪
i=1 i i ti α i 0 α ρ(s,u s ) )ds, t ∈ (tk , tk+1 ].
3 Uniqueness Result of Mild Solution
In this section, we prove the existence of mild solutions for the problem (1)–(3) with
a non-convex valued right-hand side . If A sectorial operator of the type (M, θ, α, μ)
then the strongly continuous functions Sα (t) ≤ M, Tα (t) ≤ M. To prove our
results, we shall assume the function ρ is continuous. Our result is based on contrac-
tion fixed point theorem, for this we have following assumptions
(H1 ) The function f is continuous and there exists l f ∈ L 1 (J, R+ ) such that
( f (t, ψ) − f (t, ξ )) X ≤ l f (t) ψ − ξ Bh for every ψ, ξ ∈ Bh .
(H2 ) The functions Ik , Q k are continuous and there exist li , l j ∈ L 1 (J, R+ ) such
that
Ik (x) − Ik (y) X ≤ li (t) x − y X ; Q k (x) − Q k (y) X ≤ l j (t) x − y X ,
for all x, y ∈ X and k = 1, . . . , m.
Theorem 1 Let the assumption (H1 ) and (H2 ) hold and the constant
T
Δ = M m li L 1 (J,R+ ) + mT l j L 1 (J,R+ ) + K b l f (s)ds < 1.
0
Then problem (1)–(3) has a unique mild solutions u on J.

Proof We convert the problem (1)–(3) in to fixed point problem. Let φ̄ : (−∞, T ) →
X be the extension of φ to (−∞, T ] such that φ(t) ¯ = φ(0) on J. Consider the

space Banach Bh = u ∈ Bh : u(0) = φ(0), u (0) = u 1 and define the operator
P : Bh → Bh as
⎧ t t
⎪
⎨ Sα (t)φ(0) + u 1 0 Sα (s)ds + 0 Tα (t − s) f (s, ū ρ(s,ū s ) )ds, t ∈ (0, t1 ]
Ii (ū(ti− ))Sα (t − ti )
t k
Pu(t) = Sα (t)φ(0) + u 1 0 Sα (s)ds + i=1
⎩ + k Q (ū(t − )) t S (s − t )ds + t T (t − s) f (s, ū
⎪
i=1 i i ti α i 0 α ρ(s,ū s ) )ds, t ∈ (tk , tk+1 ],
¯ = φ and ū = u on J. It is clear that u

where ū : (−∞, T ] → X is such that u(0)
is unique mild solution of the problem (1)–(3) if and only if u is a solution of the
operator equation Pu = u. Let u, u ∗ ∈ Bh , for t ∈ (0, t1 ] we have

t
Pu − Pu ∗ X ≤ Tα (t − s) L(X ) f s, ū ρ(s,ū s ) − f s, ū ∗ρ(s,ū ∗s ) X ds
0
T
∗
Pu − Pu Bh ≤ M K b
l f (s)ds u − u ∗ Bh .
0
Now, without lose of generality we consider the subinterval (tk , tk+1 ] to prove our
result. Let u, u ∗ ∈ Bh for (tk , tk+1 ], we have

k

Pu − Pu ∗ X ≤ Sα (t − ti ) L(X ) Ii ū ti− − Ii ū ∗ ti− X
i=1
k
t
+ Sα (s − ti ) L(X ) ds Q i ū ti− − Q i ū ∗ ti− X
i=1 ti

t
+ Tα (t − s) L(X ) f s, ū ρ(s,ū s ) − f s, ū ∗ρ (s,ū ∗ ) X ds
s
0
T
Pu − Pu ∗ Bh ≤ M m li L 1 (J,R+ ) + mT l j L 1 (J,R+ ) + K b l f (s)ds u − u ∗ Bh
0
≤ Δ u − u ∗ Bh .
Since Δ < 1, which implies that P is contraction map. Hence P has a unique fixed
point, which is the mild solutions of problem (1)–(3) on J. This completes the proof
of the theorem.
4 Application
Consider the following impulsive fractional partial differential equation of the form
t
∂α ∂2 u(s − ρ1 (s)ρ2 ( u ), x) 1
α
u(t, x) = 2 u(t, x) + e2(s−t) ds, t = , (17)
∂t ∂y −∞ 81 2
u(t, 0) = u(t, π ) = 0; u (t, 0) = u (t, π ) = 0 t ≥ 0, (18)
u(t, x) = φ(t, x), u (t, x) = 0, t ∈ (−∞, 0], x ∈ [0, π ], (19)
− −
u 12 u 12
Δu|t= 1 = − , Δu |t= 1 = − , (20)
2
36 + u 12 2
49 + u 12
α
where ∂t∂ α is Caputo’s fractional derivative of order α ∈ (1, 2), 0 < t1 < t2 < · · · <
tn < T are prefixed numbers and φ ∈ Bh . Let X = L 2 [0, π ] and define the operator
A : D(A) ⊂ X → X by Aw = w with the domain D(A) := {w ∈ X : w, w are
absolutely continuous, w ∈ X, w(0) = 0 = w(π )}. Then
∞

Aw = n 2 (w, wn )wn , w ∈ D(A),
n=1

where wn (x) = π2 sin(nx), n ∈ N is the orthogonal set of eigenvectors of A. It is
well known that A is the infinitesimal generator of an analytic semigroup {T (t)}t≥0
in X given by
∞

e−n t (ω, ωn )ωn , for all ω ∈ X, and every t > 0.
2
T (t)ω =
n=1
By subordination principle of solution operator, we have Sα (t) L(X ) ≤ M for t ∈ J.

0
Let h(s) = e2s , s < 0 then l = −∞ h(s)ds = 21 < ∞, for t ∈ (−∞, 0] and define
0
φ Bh = h(s) sup φ(θ ) L 2 ds.
−∞ θ∈[s,0]
Hence for (t, φ) ∈ [0, 1] × Bh , where φ(θ )(x) = φ(θ, x), (θ, x) ∈ (−∞, 0] ×
[0, π ]. We assume that ρi : [0, ∞) → [0, ∞), i = 1, 2, are continuous functions.
Set u(t)(x) = u(t, x), and ρ(t, φ) = ρ1 (t)ρ2 ( φ(0) ), we have
φ u u
f (t, φ)(x) = , Ik (u) = , Jk (u) = ,
81 36 + u 49 + u
then with these settings the problem (17)–(20) can be written in the abstract form of
Eqs. (1)–(3). It is obvious that the maps f, Ik , Jk following the assumption H1 , H2 .
This implies that there exists a unique mild solutions of problem (17)–(20) on J.
References
1. Abbas, S., Benchohra, M.: Impulsive partial hyperbolic functional differential equations of
fractional order with state-dependent delay. An Int. J. Theory Appl. (2010). ISSN 1311–0454
2. Agarwal, R.P., Andrade, B.D.: On fractional integro-differential equations with state-dependent
delay. Comput. Math. Appl. 62, 1143–1149 (2011)
3. Araya, D., Lizama, C.: Almost automorphic mild solutions to fractional differential equations.
Nonlinear Anal. TMA 69, 3692–3705 (2008)
4. Benchohra, M., Litimein, S., N’Guerekata, G.: On fractional integro-differential inclusions
with state-dependent delay in Banach spaces. Appl. Anal. 2011, 1–16 (2011)
5. Benchohra, M., Berhoun, F.: Impulsive fractional differential equations with state dependent
delay. Commun. Appl. Anal. 14(2), 213–224 (2010)
6. Carvalho dos Santos, J.P., Arjunan, M.M.: Existence results for fractional neutral integro-
differential equations with state-dependent delay. Comp. and Math. with Appl. 62, 1275–1283
(2011)
7. Chauhan, A., Dabas, J., Kumar, M.: Integral boundary-value problem for impulsive fractional
functional integro-differential equation with infinite delay. Electr. J. Diff. Eqn. 2012(229), 1–13
(2012)
8. Chauhan, A., Dabas, J.: Existence of mild solutions for impulsive fractional order semilinear
evolution equations with nonlocal conditions. Electr. J. Diff. Eqn. 2011(107), 1–10 (2011)
9. Chauhan, A., Dabas, J.: Local and global existence of mild solution to an impulsive frac-
tional functional integro-differential equation with nonlocal condition. Commun. Nonlinear
Sci. Num. Simul. 19, 821–829 (2014)
10. Dabas, J., Chauhan, A.: Existence and uniqueness of mild solution for an impulsive neutral
fractional integro-differential equations with infinity delay. Math. Comput. Mod.57, 754–763
(2013)
11. Dabas, J., Gautam, G.R.: Impulsive neutral fractional integro-differential equation with state
dependent delay and integral boundary condition. Electr. J. Diff. Eqn. 2013(273), 1–13 (2013)
12. Feckan, M., Zhou, Y., Wang, J.: On the concept and existence of solution for impulsive fractional
differential equations. Commun. Nonlinear Sci. Numer. Simul. 17, 3050–3060 (2012)
13. Fu, X., Huang, R.: Existence of solutions for neutral integro-differential equations with state-
dependent delay. Appl. Math. Comput. 224, 743–759 (2013)
14. Hale, J.K., Kato, J.: Phase space for retarded equations with infinite delay. Funkcialaj Ekvacioj
21, 11–41 (1978)
15. Kilbas, A.A., Srivastava, H.M., Trujillo, J.J.: Theory and Applications of Fractional Differential
Equations, vol. 204. North-Holland Mathematics Studies, Elsevier Science, Amsterdam (2006)
16. Lakshmikantham, V., Leela, V.S., Vasundhara Devi, J.: Theory of Fractional Dynamic Systems.
Cambridge Scientific Publishers, Cambridge (2009)
17. Mainardi, F., Gorenflo, R.: On Mittag-Leffer-type functions in fractional evolution processes.
J. Comput.Appl. Math. 118, 283–299 (2000)
18. Miller, K.S., Ross, B.: An Introduction to the Fractional Calculus and Differential Equations.
Wiley, New York (1993)
19. Podlubny, I.: Fractional Differential Equation. Academic Press, San Diego (1999)
20. Samko, S.G., Kilbas, A.A., Marichev, O.I.: Fractional Integrals and Derivatives Theory and
Applications. Gordon and Breach, Yverdon (1993)
21. Shu, X., Wang, Q.: The existence and uniqueness of mild solutions for fractional differential
equations with nonlocal conditions of order 1 < α < 2. Comput. Math. Appl. 64, 2100–2110
(2012)
22. Wang, J., Fekckan, M., Zhou, Y.: On the new concept of solutions and existence results for
mpulsive fractional evolution equations. Dyn. PDE 8(4), 345–361 (2011)
23. Wang, J., Li, X., Wei, W.: On the natural solution of an impulsive fractional differential equation
of order q ∈ (1, 2). Commun. Nonlinear. Sci. Numer. Simul. 17, 4384–4394 (2012)
An O(N −1 l nN)4 Parameter Uniform
Difference Method for Singularly Perturbed
Differential-Difference Equations
Komal Bansal and Kapil K. Sharma
Abstract In this paper, we propose an O(N −1ln N )4 parameter uniform numerical

scheme for singularly perturbed differential-difference equations (SPDDE) having
both delay and advance arguments in reaction term. These types of problems are
ubiquitous in many mathematical models of physical and biological phenomena.
Piecewise uniform fitted mesh with improved fourth-order numerov method is used.
A parameter uniform error estimate of order O(N −1 ln N )4 is proved. We calculate
numerical solution of some examples using the proposed method to establish the
higher order parameter uniform estimates.
Keywords Singular perturbation · Parameter uniform error estimate · Differential-

difference equations · Piecewise uniform fitted mesh · Numerov method
AMS subject classifications: 34K26 · 65L12 · 34K28
1 Introduction
Mathematical models for many real-life phenomena involve differential equations

having nonsmooth solution with singularities related to boundary layer. In 1994,
Lange and Miura [7] proposed a mathematical model for first-exit time problem in
the modeling of the activation of neuronal variability. The problem for expected first-
exit time, given initial membrane potential, can develop as a general boundary-value
problem for linear second-order differential-difference equation [11, 12].
K. Bansal
Department of Mathematics, Panjab University, Chandigarh 160014, India
e-mail: bansalkomal87@gmail.com
K.K. Sharma (B)
Department of Mathematics, South Asian University (SAU), Akbar Bhawan,
Chanakyapuri, New Delhi 110021, India
e-mail: kapil.sharma@sau.ac.in

DOI 10.1007/978-81-322-2485-3_24
300 K. Bansal and K.K. Sharma
Consider a linear singularly perturbed differential-difference equation of mixed

type
εy (x) + α(x)y(x − δ) + ω(x)y(x) + β(x)y(x + η) = f (x), x ∈ Ω = (0, 1) (1)
and 0 < ε 1 subject to the interval and boundary conditions
y(x) = φ(x) on − δ ≤ x ≤ 0, y(x) = γ(x) on 1 ≤ x ≤ 1 + η, (2)
where α(x), ω(x), β(x), f(x), φ(x), and γ(x) are smooth functions, δ and η are the
small shifting parameters of o(ε). It is assumed that (α(x) + β(x) + ω(x)) satisfies
the condition
(α(x) + β(x) + ω(x)) ≤ −b∗ < 0, ∀x ∈ Ω = [0, 1] (3)
where b∗ is a positive constant. Because of the assumption (3) solution of the

boundary-value problem considered here has no oscillation and will exhibit boundary
layer behavior.
In 2002, Kadalbajoo and Sharma [4, 6] initiated the numerical study of such type of
boundary-interval value problems, since than a lot of work had been carried out. Rao
and Chakravarthy [1, 10] proposed some higher order methods but their methods
are not parameter uniform. In [5] Kadalbajoo and Sharma established parameter
uniform numerical methods but the order of convergence is less than one. In [9],
Patidar and Sharma constructed nonstandard finite difference methods (NSFDMs)
which are parameter uniform with second order of convergence. In this paper, we
develop and analyze a parameter uniform difference method with O(N −1ln N )4
based on numerov finite difference scheme which utilizes a piecewise uniform mesh
condensed in the boundary layer regions to capture the singular behavior of the
solution in the layer regions.
Notations and Terminology: We used standard notations and symbols. Further, C
will denote a positive constant which may take different values in different equations
and inequalities and is always independent of ε and step size h. Here || .||∞ represent
the standard supremum norm.
2 Some Deducible Estimates
Let us consider Taylor series expansion of the terms y(x − δ) and y(x + η) in (1),
we have
y(x − δ) ≈ y(x) − δ y (x), (4)
y(x + η) ≈ y(x) + η y (x), (5)

An O(N −1 ln N )4 Parameter Uniform Difference Method … 301
using (4)–(5) in (1)–(2), we obtain
− εy (x) + Aε (x)y (x) + B(x)y(x) ≈ F(x), (6)
y(0) ≈ φ(0), y(1) ≈ γ(1); (7)
where Aε (x) ≡ Aε,δ,η (x) = α(x)δ − β(x)η,

B(x) = −(α(x) + w(x) + β(x)) and F(x) = − f (x).
Since (6)–(7) is an approximation of (1)–(2), it is important to have the knowledge
of the error which may occur in neglecting the higher order terms in (4)–(5). As
delay (δ) and advance (η) are sufficiently small, as a result the solution u of the new
problem (resulting from (6)–(7))
− εu (x) + Aε (x)u (x) + B(x)u(x) = F(x), (8)
u(0) = φ(0) = φ0 , u(1) = γ(1) = γ1 , (9)
which differ from the original problem (1)–(2) by O(δ 2 u , η 2 u ) terms will give a
good approximation to the solution y of problem (1)–(2).
Remark 1 In the differential equations of the type considered in this paper when
both the delay and the advance arguments are present, and the fact that δ and η are of
the order of ε then to find the layer behavior of the solution, one has to observe how
large is the coefficient of convection term as compared to ε, which solely depends on
the values of ε. If ε is not very small, say, e.g., ε ∈ [10−2 , 1], then Aε (x) contributes
a lot and we have only one boundary layer in the solution. Boundary layer will be on
the left- or right side of the domain depends upon the sign of Aε (x). When ε is very
small then Aε (x) does not contribute significantly, and therefore contribution of the
coefficient of the reaction term has much influence which results into two boundary
layers one at each end.
As we are interested in developing the numerical scheme in the case when ε is

very small, therefore, in view of the above remark, we dropout the term Aε (x) from
(8) and consider the problem
L∗ u ≡ −ε(x)u (x) + B(x)u(x) = F(x), (10)
u(0) = φ0 , u(1) = γ1 . (11)
Operator of the continuous problem (10)–(11) will satisfy continuous maximum

principle, this implies that the solution is unique and since the problem under con-
sideration is linear, the existence of the solution is implied by its uniqueness.
Lemma 1 The solution u ε (x) of the constant coefficient problem corresponding to

(10)–(11) satisfies
|u ε | ≤ C(1 + S(x, b∗ ))
where
S(x, b∗ ) = exp −x b∗ /ε + exp −(1 − x) b∗ /ε
and
|u (k)
ε | ≤ Cε
−(k/2)
S(x, b∗ ) ∀ k≥1
Lemma 2 The solution u(x) of (10)–(11) has the decomposition
u(x) := vε (x) + w L (x) + w R (x)
where the regular(smooth) component vε (x) satisfies
|vε (x)| ≤ C(1 + S(x, b∗ ))

(k)
vε (x) ≤ C 1 + ε−(k−2)/2 S(x, b∗ ) , ∀ k ≥ 1.
The singular components w L , w R satisfy

(k)
w L (x) ≤ Cε−(k/2) exp −x b∗ /ε , ∀k ≥0

(k)
w R (x) ≤ Cε−(k/2) exp −(1 − x) b∗ /ε , ∀k ≥0
Note: For the proofs of Lemma 1 and Lemma 2, the reader can refer to [5, 9].
Remark 2 Lemma 2 and the fact that the shift arguments are sufficiently small
implies that the Eq. (8) is a good approximation to Eq. (1). This is because of the
fact that
S(ξ, b∗ )
u (ξ) ≤ C 1 +
ε
with δ = o(ε), η = o(ε) and therefore

∗)
2 S(ξ, b
δ u ≈C δ +δ
2 2
.
ε

Similarly, we can approximate η 2 u (ξ) and S(ξ, b∗ ) → 0 as ε → 0 which leads to

the fact that the terms δ 2 u , η 2 u are exponentially small. By similar arguents, one
can show that higher order terms of Taylor series (4) and (5) are exponentially small.
3 The Discrete Problem
The fitted piecewise uniform mesh on the interval (0,1) is constructed by partitioning
the interval in to three subintervals (0,ν), (ν, 1 − ν), and (1 − ν,1). Assuming that
N = 2l with l ≥ 3 the intervals (0,ν) and (1 − ν,1) are each divided in to N/4 equal
mesh elements while the interval (ν,1 − ν) is divided into N/2 equal mesh elements
which guarantee that there is at least one point in the boundary layer region. The
resulting piecewise uniform mesh ΩνN depends on just one parameter ν where the
transition parameter ν is defined as

√
ε
ν = min 1/4, ∗ ln(N ) . (12)
b
We consider the following differential equation
εU (x) + (β(x)η − α(x)δ)U (x) + (α(x) + ω(x) + β(x))U (x) = f (x), x ∈ Ω = (0, 1)
(13)
εU (x) = f (x) − (β(x)η − α(x)δ)U (x) − (α(x) + ω(x) + β(x))U (x), (14)
εU (x) = g(x, U, U ). (15)
Now we consider fourth-order numerov method and fitted mesh [3, 8] to solve the
Eq. (14) and this equation is approximated by the following scheme

2ε Ui+1 − Ui Ui − Ui−1 1
− = [gi+1 + 10gî + gi−1 ] (16)
h i+1 + h i h i+1 hi 12
where gî = g(xi , Ui , Ûi )
),
gi±1 = g(xi±1 , Ui±1 , Ui±1 (17)
Ui+1 − Ui−1
Ui = , (18)
h i+1 + hi

Ui+1 − Ui−1 2h i+1 Ui+1 − Ui Ui − Ui−1
Ui+1 = + − , (19)
h i+1 + hi h i+1 + h i h i+1 hi

Ui+1 − Ui−1 2h i Ui+1 − Ui Ui − Ui−1
Ui−1 = − − , (20)
h i+1 + hi h i+1 + h i h i+1 hi
h i+1 + hi
Ûi = Ui − (gi+1 − gi−1 ), (21)
40
From the proposed scheme (16), we get the corresponding discrete operator L εN .
Lemma 3 Assume that the mesh function χi satisfies χ0 ≥ 0, χ N ≥ 0. Then for the
discrete operator L εN , if L εN χi ≥ 0 for 1 ≤ i ≤ N − 1 implies that χi ≥ 0 for all
0 ≤ i ≤ N.
Lemma 4 If Z i is any mesh function such that Z 0 = Z n = 0. Then

1 N
|Z i | ≤ max Lε Z j , ∀ 0 ≤ i ≤ N .
b∗ 1≤ j≤N −1
Theorem 1 The fitted mesh numerov method (16) with the fourth-order numerov
method and the piecewise uniform fitted mesh ΩνN , condensing at the boundary
points x = 0 and x = 1, is ε-uniform for the problem (1)–(2) provided that ν is chosen
to satisfy condition (12) above. Moreover, the solution u ε of (1)–(2) and the solution
Uε of (15) satisfy the following ε-uniform error estimate
sup ||Uε − u ε ||Ω N ≤ C N −4 (ln N )4

0<ε≤1 ν
where C is constant independent of ε and mesh size h.

Solution: The solution Uε of the discrete problem are decomposed in a similar manner
to the decomposition of the solution u ε of (1)–(2). Thus
Uε = Vε + Wε , (22)
where Vε is the solution of the following nonhomogeneous problem
L εN Vε = f, Vε (0) = v(0), Vε (1) = v(1) (23)
and Wε is the solution of the following homogeneous problem
L εN Wε = 0, Wε (0) = w(0), Wε (1) = w(1). (24)
We can write the error in the form
Uε − u ε = (Vε − vε ) + (Wε − wε ) (25)
Now the singular and smooth components of the error can be calculated separately.
We use the following classical argument [3, 8] to estimate the smooth component.
From the differential and difference equations
L εN (Vε − vε )(xi ) = f − L εN vε (26)

= (L ε − L εN )vε (27)
h4 6
= v (ξ)ε. (28)
240 ε
√
(6) ε
Now by using the estimate of vε and the fact that 1
4 ≤ β ln(N ) we get

N
L ε (Vε − vε )(xi ) ≤ C (lnN )2 N −4 . (29)
By using Lemma 4 we get
|(Vε − vε )(xi )| ≤ C (lnN )2 N −4 . (30)
Estimates of the singular component of the error depends upon whether ν = 1

or
√ 4
ε
ν= β ln(N ).
Case 1: √
ε
In the case when ν = 1
4 the mesh is uniform and 1
4 ≤ β ln(N ). By using the same
√
(6) ε
classical argument as for Vε − vε estimates of wε and the fact that 4 ≤ β ln(N )
1
leads to
N
L ε (Wε − wε )(xi ) ≤ C (lnN )4 N −4 . (31)
Case 2: √
ε
In the case when 41 > β ln(N ). In this case, mesh will be piecewise uniform with the
2(1−2ν)
mesh spacing N in the subinterval [ν, 1 − ν] and 4νN in each of the subintervals
[0, ν] and [1 − ν, 1]
Subcase 1:
Estimates of the error in (0, ν) and (1 − ν, 1)
L εN (Wε − wε )(xi ) = (L ε − L εN )wε (32)

h4
= wε6 (ξ)ε. (33)
240
(6)
Now by using the estimate of wε and the fact that step size i.e. h = 4ν
and
√ N
ε
ν= b∗ ln(N ) we get

N
L ε (Wε − wε )(xi ) ≤ C (ln N )4 N −4 . (34)
Subcase 2:
Estimates of the error in (ν, 1 − ν)
L εN (Wε − wε )(xi ) = (L ε − L εN )wε (35)

h4
= wε6 (ξ)ε. (36)
240
In the interval (ν, 1 − ν) solution is smooth, this implies

N
L ε (Wε − wε )(xi ) ≤ C N −4 . (37)
Combining (31), (34), (37) gives

N
L ε (Wε − wε )(xi ) ≤ C N −4 (ln N )4 , (38)
for all xi ∈ (0, 1). Applying Lemma 4 to the mesh function Wε − wε leads the
required estimate of the error in the singular component of the solution
|(Wε − wε )(xi )| ≤ C (ln N )4 N −4 . (39)
This concludes the proof of the theorem.
4 Numerical Experiments
Example 1 Consider the problem
εy (x) + 0.25y(x − δ) − y(x) + 0.25y(x + η) = 1, x ∈ Ω = (0, 1)
subject to the interval and boundary conditions
y(x) = 1, −δ ≤ x ≤ 0; y(x) = 0, 1 ≤ x ≤ 1 + η.
Example 2 Consider the problem
εy (x) − y(x − δ) + y(x) − 0.5y(x + η) = 0, x ∈ Ω = (0, 1)
y(x) = 1, −δ ≤ x ≤ 0; y(x) = 1, 1 ≤ x ≤ 1 + η.
We use the double-mesh principle [2] to calculate the maximum absolute error
and order of convergence of the numerical scheme. The following estimates for
maximum point wise error E εN and order of convergence Pεn are computed using the
double-mesh principle
(E εN /E ε2N )

E εN = max YiN − Y2i2N , Pεn = (40)
0≤i≤n−1 log2
The calculated maximum absolute error and order of convergence for Examples 1
and 2 are arranged in the form of Tables 1 and 2 repectively.
Table 1 Maximum pointwise error E εN and numerical order of convergence PεN for each value of
ε and N are given, respectively, for example 1 for δ = 0.0001, η = 0.0001
ε N =8 N = 16 N = 32 N = 64
2−0 9.764395E-008 6.105717E-009 3.816677E-010 2.390149E-011
3.999 3.999 3.997
2−2 3.755846E-006 2.351924E-007 1.470671E-008 9.194143E-010
3.997 3.999 3.999
2−4 6.193481E-005 3.899850E-006 2.442115E-007 1.527083E-008
3.989 3.997 3.999
2−6 7.038173E-004 4.509816E-005 2.839569E-006 1.778154E-007
3.964 3.989 3.997
2−8 7.773966E-003 6.939454E-004 4.446212E-005 2.799506E-006
3.485 3.964 3.989
2−10 8.666641E-003 2.138141E-003 3.463243E-004 4.446165E-005
2.019 2.626 2.961
2−12 9.412269E-003 2.150507E-00 3.467433E-004 4.624679E-005
2.129 2.632 2.906
2−14 9.194118E-003 2.150605E-003 3.470122E-004 6.159023E-005
2.095 2.631 2.494
2−16 8.772140E-003 2.146675E-003 3.469423E-004 8.252287E-005
2.030 2.629 2.071
2−18 8.635118E-003 2.143429E-003 3.467774E-004 7.052199E-005
2.010 2.627 2.297
2−20 8.608975E-003 2.141461E-003 3.466508E-004 5.029886E-005
2.007 2.627 2.784
Table 2 Maximum pointwise error E εN and numerical order of convergence PεN for each value of
ε and N are given, respectively, for example 2 for δ = 0.00005, η = 0.0001
ε N =8 N = 16 N = 32 N = 64
2−0 1.345542e-008 8.412097e-010 5.263479e-011 3.377632e-012
3.999 3.998 3.961
2−4 1.751296e-005 1.098727e-006 6.873708e-008 4.297450e-009
3.994 3.998 3.999
2−8 2.535968e-003 1.671284e-004 1.115936e-005 7.001167e-007
3.923 3.904 3.994
2−12 9.214721e-003 2.089910e-003 3.358351e-004 4.485627e-005
2.140 2.637 2.904
2−16 9.134400e-003 2.092756e-003 3.619246e-004 8.178389e-005
2.125 2.531 2.145
2−20 8.521945e-003 2.085891e-003 3.360410e-004 6.383956e-005
2.030 2.633 2.396
5 Conclusion
We construct a numerical scheme using shishkin mesh technique with the improved
numerov method to solve singularly perturbed problem having both delay and
advance arguments in reaction terms. We used priori estimates on the solution and its
derivatives to prove a parameter uniform error estimate. We divide the domain in to
two regions, namely interior and outer regions, which are dealt separately. We made
a matlab program of the proposed numerical scheme to validate the computational
efficiency, consistency, and stability of the scheme.
References
1. Chakravarthy, P.: Pramod, Rao, R. Nageshwar: A modified Numerov method for solving sin-
gularly perturbed differential-difference equations arising in science and engineering. Results
Phys. 2, 100–103 (2012)
2. Doolan, E.P., Miller, J.J.H., Schilders, W.H.A.: Uniform Numerical Methods for Problems with
Initial and Boundary Layers. Boole Press, Dublin (1980)
3. Jain, M.K., Lynger, S.R.K., Jain, R.K.: Numerical Methods for Scientific and Engineering
Computation. New Age International Publishers, New Delhi (1984)
4. Kadalbajoo, M.K., Sharma, K.K.: Numerical analysis of boundary-value problems for sin-
gularly perturbed differential-difference equations with small shifts of mixed type. J. Optim.
Theory Appl. 115, 145–163 (2002)
5. Kadalbajoo, M.K., Sharma, K.K.: ε-Uniform fitted mesh method for singularly perturbed
differential-difference equations: Mixed type of shifts with layer behavior. Int. J. Comput.
Math. 81, 49–62 (2004)
6. Kadalbajoo, M.K., Sharma, K.K.: Numerical treatment of boundary value problems for second
order singularly perturbed delay differential equations. Comput. Appl. Math. 24, 151–172
(2005)
7. Lange, C.G., Miura, R.M.: Singular perturbation analysis of boundary-value problems for
differential-difference equations. V. Small shifts with layer behavior. SIAM J. Appl. Math. 54,
72–249 (1994)
8. Miller, J.J.H., Riordan, E.O’., Shishkin, G.I.: Fitted numerical methods for singular perturba-
tion problems. In: Error estimates in the maximum norm for linear problems in one and two
dimensions, World Scientific Publishing Co., Inc, River Edge, NJ (1996)
9. Patidar, Kailash C., Sharma, Kapil K.: ε-Uniformly convergent non-standard finite difference
methods for singularly perturbed differential-difference equations with delay and advance. Int.
J. Numer. Methods Eng. 66, 272–296 (2006)
10. Rao, R.: Nageshwar, Chakravarthy, P. Pramod: A finite difference method for singularly per-
turbed differential-difference equations arising from a model of neuronal variability. J. Taibah
Univ. Sci. 7, 128–136 (2013)
11. Stein, R.B.: A theoretical analysis of neuronal variability. Biophys. J. 5, 173–194 (1965)
12. Stein, R.B.: Some models of neuronal variability. Biophys. J. 7, 37–68 (1967)
Almost Periodicity of a Modified
Leslie–Gower Predator–Prey System
with Crowley–Martin Functional Response
Jai Prakash Tripathi and Syed Abbas
Abstract In this paper, we discuss a modified Leslie–Gower Lotka–Volterra system

with Crowley–Martin type functional response. Crowley–Martin functional response
is similar to the Beddington–DeAngelis functional response but contains an extra
term describing mutual interference by predators at high prey density. The rates
are assumed to be almost periodic, which generalizes the concept of periodicity.
We discuss the permanence, existence, uniqueness, and asymptotic stability of an
almost periodic solution of the model under consideration by applying comparison
theorem of differential equations and constructing a suitable Lyapunov functional.
The analytical results obtained in this paper are illustrated with the help of a numerical
example.
Keywords Almost-periodicity · Asymptotic stability · Crowley-Martin

· Permanence · Predator-prey · Lyapunov functional
1 Introduction
The rate of prey consumption by an average predator per unit time, i.e., preda-
tor’s functional response is one of important feature of predator–prey relation-
ship [1, 2]. The Crowley–Martin [3, 4] type functional response is given by,
aX
p(x, y) = . The parameters a and b stand for the effects
1 + bX + cY + bcX Y
of capture rate and handling time, respectively. Parameter c describes the magnitude
of interference among predators. Crowley–Martin functional response predicts that
interference affects on feeding rate remain important at high prey abundance.
J.P. Tripathi (B) · S. Abbas

School of Basic Sciences, Indian Institute of Technology Mandi,
Mandi 175001, H.P, India
e-mail: jtripathi85@gmail.com
S. Abbas
e-mail: sabbas.iitk@gmail.com

DOI 10.1007/978-81-322-2485-3_25
310 J.P. Tripathi and S. Abbas
A model must becomes nonautonomous when the temporal inhomogeneity of

the environment is taken into account in the model. The consideration of periodic
and almost periodic coefficients can be found in case of nonautonomous systems. A
function g : R → R is called almost periodic if g(x + τ ) = g(x) is satisfied with
an arbitrary degree of accuracy by infinitely many values of τ, those values being
spread over the whole range from −∞ to +∞ in such a way as not to leave empty
intervals of arbitrary great length.
Almost periodic solutions of ecological models have received significant attention
from researchers during last few decades [5]. Recently, Abbas et al. [6] discussed the
existence of a unique almost periodic solution of a delayed phytoplankton model.
In theoretical ecology, there are several papers [5, 7, 8] on permanence and almost
periodic solutions of Lotka–Volterra system. However, no work has been done for
modified Leslie–Gower type predator–prey model with Crowley–Martin functional
response.
Motivated by above, in this work, we consider the following nonautonomous mod-
ified Leslie–Gower [9] type predator–prey model with Crowley–Martin functional
response
d x(t) b2 (t)y(t)
= x(t) a1 (t) − b1 (t)x(t) − ,
dt a(t) + b(t)x(t) + c(t)y(t) + d(t)x(t)y(t)
dy(t) e(t)y(t)
= y(t) a2 (t) − , (1)
dt x(t) + k(t)
where x(t) and y(t) denote the density of prey and predator at time t, respectively;
a1 , a2 , b1 , b2 , a, b, c, d, e, k ∈ C(R, R + ) are nonnegative almost periodic functions
of t.
Functions ai (t), bi (t), a(t), b(t), c(t), d(t), e(t), k(t), (i = 1, 2) are continuous,
bounded by positive numbers. Our main objective is to obtain sufficient conditions for
the existence of a unique globally attractive almost periodic solution of the system (1).
2 Boundedness and Permanence
Let g(t) be a continuous and bounded function on R. Let gl and g u denote inf g(t)
t∈R
and supg(t), respectively. Then we have
t∈R
min {ail , bil , a l , bl , cl , d l , el , k l , }>0, max {aiu , biu , a u , bu , cu , d u , eu , k u , } < ∞.
i=1,2 i=1,2
du
Lemma 1 [10] If p > 0, q > 0 and ≤ (≥) u(t) q − pu(t) , u(0) > 0,
dt
then we have q q
lim sup u(t) ≤ lim inf u(t) ≥ .
t→+∞ p t→+∞ p
Lemma 2 The positive cone is positively invariant with respect to the model sys-
tem (1).
Almost Periodicity of a Modified Leslie–Gower Predator–Prey System … 311
Proof The proof is similar to the proof given in [1].
Theorem 1 If a1l a l > b2u M2ε , then

κε := (x, y) ∈ R 2 |m ε1 ≤ x ≤ M1ε , m ε2 ≤ x ≤ M2ε ,
is positively invariant with respect to the model system (1). Here
a1u a1l a l − b2u M2ε

M1ε := + ε, m ε1 := − ε,
b1l b1u a l
a2u (M1ε + k u ) a2l (m ε1 + k l )
M2ε := , m ε2 := .
el eu
Proof First equation of the model system (1) gives
d x(t)
≤ x(t)(a1u − b1l S(t)) (2)
dt
Using Lemma 1, Eq. (2) implies that
au
lim sup x(t) ≤ 1l ≡ M1 .
t→∞ b1
Thus for sufficiently small ε > 0, ∃ a positive real number T1 such that
x(t) ≤ M1 + ε ≡ M1ε , ∀ t ≥ T1 .
From the second equation of the model system (1), we obtain
dy(t) el y(t)
≤ y(t) a2u − ε . (3)
dt M1 + k u
Again according to Lemma 1 and Eq. (3), one has

a u (M ε + k u )
lim sup y(t) ≤ 2 1 l ≡ M2 .
t→∞ e
Hence, for the above ε, ∃ a positive real number T2 ≥ T1 ≥ 0 such that
a u (M ε + k u )
y(t) ≤ 2 1 l + ε ≡ M2ε ∀ t ≥ T2 .
e
From the first equation of the model system (1), we have
d x(t) bu M ε
≤ x(t) a1l − b1u x(t) − 2 l 2 . (4)
dt a
Using Lemma 1, for ε > 0 ∃ a positive real number T3 ≥ T2 ≥ 0 such that
a l a l − bu M ε
x(t) ≥ 1 u l2 2 − ε ≡ m ε1 .
b1 a
In the similar fashion, from the second equation of the model system (1), one can
also find that
dy(t) eu y(t)
≤ y(t) a2l − ε . (5)
dt m 1 + kl
Applying the Lemma 1, for ε > 0 ∃ a positive real number T4 ≥ T3 ≥ 0 such that
a l (m ε + k l )
y(t) ≥ 2 1u − ε ≡ m ε2 .
e
Thus we arrive at the following result:
Theorem 2 If a1l a l > b2u M2ε , then the model system (1) is permanent and the set κε
with ε > 0 is an ultimately bounded region of the model system (1).
Let (S) be the collection of all solutions X (t) = (x(t), y(t))T of (1) on R.
Theorem 3 (S) is nonempty.
Proof Proof is similar to the proof given in [8].
3 Existence of a Unique Almost Periodic Solution
Definition 1 [7] A function g(t, x), where g is an m-vector, t is a real scalar and x is
an n− vector, is said to be almost periodic in t uniformly with respect to x ∈ X ⊂ R n ,
if g(t, x) is continuous in t ∈ R and x ∈ X, and if for any ε > 0, it is possible to
find a constant l(ε) > 0 such that in any interval of length l(ε) there exist a τ such
that the inequality

m
||g(t + τ, x) − g(t, x)|| = |gi (t + τ, x) − gi (t, x)| < ε
i=1
is satisfied for all t ∈ R, x ∈ X. The number τ is called an ε-translation number of

g(t, x).
Definition 2 [11] A bounded positive solution X (t) = (x̂(t), ŷ(t)) of the model
system (1) with X (0) > 0 is said to be globally attractive (globally asymptotically
stable), if any other solution Y (t) = (x(t), y(t)) of the system (1) with Y (0) > 0,
satisfies lim |X (t) − Y (t)| = 0
t→+∞
Definition 3 [8] The upper right Dini derivative for a function G : R → R is

defined as
G(t + h) − G(t)
D + G(t) = lim sup
h→0 + h
Theorem 4 If the condition of Theorem 2 and if
b2u M2 (bu + d u M2 ) e u M2
b1l > +
(a l + bl m 1 + cl m 2 + d l m 1 m 2 )2 (m 1 + k l )2
e l b2u b2u M2 (cu + d u M2 )
> +
M1 + k u (a l + bl m 1 + cl m 2 + d l m 1 m 2 ) (a l + bl m 1 + cl m 2 + d l m 1 m 2 )2
hold. Then any two positive solutions X (t) = (x(t), y(t)) and Y ∗ (t) = (x ∗ (t),
y ∗ (t)) of the model system (1) satisfies
lim |X (t) − Y ∗ (t)| = 0
t→∞
Proof Consider any two positive solutions X (t) = (x(t), y(t))T and Y ∗ (t) =
(x ∗ (t), y ∗ (t)) of the model system (1). Theorem 2 gives that for an enough small
ε > 0 ∃ a T > 0 such that
s1 − ε < x(t) < S1 + ε, s2 − ε < y(t) < S2 + ε,

s1 − ε < x ∗ (t) < S1 + ε, s2 − ε < y ∗ (t) < S2 + ε, (6)
for all t ≥ T. Define Δ(t, x(t), y(t)) = a(t) + b(t)x(t) + c(t)y(t) + d(t)x(t)y(t).
Let S1 (t) = | ln x(t) − ln x ∗ (t)|.
The Dini derivative of S1 (t) along the solution of (1) gives
ẋ(t) x˙∗ (t)
D + G 1 (t) = sgn(x(t) − x ∗ (t)) − ∗
x(t) x (t)
y(t)
= sgn(x(t) − x (t)) − b1 (t)(x(t) − x ∗ (t) − b2 (t)
∗
Δ(t, x(t), y(t))
y ∗ (t)
−
Δ(t, x ∗ (t), y ∗ (t))
y(t) y ∗ (t)
= −b1 (t)|x(t) − x ∗ (t)| − sgn(x(t) − x ∗ (t))b2 (t) −
Δ(t, x(t), y(t)) Δ(t, x(t), y(t))
y ∗ (t) y ∗ (t)
+ − ∗ ∗
Δ(t, x(t), y(t)) Δ(t, x (t), y (t))

∗ |y(t) − y ∗ (t)|
≤ −b1 (t)|x(t) − x (t) + b2 (t)
Δ(t, x(t), y(t))

y ∗ (t) (b(t) + d(t)y ∗ (t))|x(t) − x ∗ (t)| + (c(t) + d(t)x(t))|y(t) − y ∗ (t)|
+
Δ(t, x ∗ (t), y ∗ (t)).Δ(t, x(t), y(t))
Furthermore, consider S2 (t) = | ln y(t) − ln y ∗ (t)|.

The upper right derivative of S2 (t) is given by
y˙∗ (t)
ẏ(t)
D + G 2 (t) = sgn(y(t) − y ∗ (t)) −
y(t) y ∗ (t)
y ∗ (t) y(t)
= sgn(y(t) − y ∗ (t))e(t) ∗ −
x (t) + k(t) x(t) + k(t)
y ∗ (t)(x(t) − x ∗ (t)) y ∗ (t) − y(t)
= sgn(y(t) − y ∗ (t))e(t) ∗
+
(x (t) + k(t))(x(t) + k(t)) x(t) + k(t)
e(t)y ∗ (t)|x(t) − x ∗ (t)| |y(t) − y ∗ (t)|
≤ ∗ − e(t)
(x (t) + k(t))(x(t) + k(t)) x(t) + k(t)
Combining the two functions G i (t), i = 1, 2, we obtain G(t) = G 1 (t) + G 2 (t). For
t ≥ T, we have

b2 (t)(S2 + ε) b(t) + d(t)(S2 + ε) e(t)(S2 + ε)
D + G(t) ≤ − b1 (t) − − |x(t) − x ∗ (t)|
Δ2 (t, s1 − ε, m − ε) (s1 − ε + k(t))2

e(t) b2 (t) b2 (t)(S2 + ε) e(t) + d(t)(S2 + ε)
− − −
S1 + ε + k(t) Δ(t, s1 − ε, m − ε) Δ2 (t, s1 − ε, m − ε)
|y(t) − y ∗ (t)|.
b2u S2 (bu + d u S2 ) eu S2 el
Define ρ = min b1l − − ,
(a l + bl s1 + cl s2 + d l s1 s2 )2 (s + k l )2 S1 + k u
b2u b2u S2 (cu + d u S2 ) 1
− l − . Choosing ε → 0,
(a + bl s1 + cl s2 + d l s1 s2 ) (a l + bl s1 + cl s2 + d l s1 s2 )2
the above inequality takes the following form:

D + G(t) ≤ −ρ |x(t) − x ∗ (t)| + |y(t) − y ∗ (t)| . (7)
Integrating the above relation (7) from T to t, we obtain

t
G(t) + ρ |x(s) − x ∗ (s)| + |y(s) − y ∗ (s)| ds < G(t) < +∞,
T
which gives
t G(t)
lim sup |x(s) − x ∗ (s)| + |y(s) − y ∗ (s)| ds < < +∞.
t→∞ T ρ
One can easily observe that |x(t)−x ∗ (t)| and |y(t)− y ∗ (t)| are uniformly continuous
on [T, +∞). Thus we have
lim |x(t) − x ∗ (t)| = 0, lim |y(t) − y ∗ (t)| = 0.

t→∞ t→∞
Hence the solution of the model system (1) is globally attractive.
Theorem 5 Under the conditions of the Theorem 4, the model system (1) has a
unique almost periodic solution.
Proof Theorem 3, implies that ∃ a bounded positive solution u(t) = (u 1 (t), u 2 (t))
of the model system (1). Thus ∃ a sequence {tn }, tn → ∞ as n → ∞, such that
(u 1 (t + tn ), u 2 (t + tn ))T satisfies
d x(t)
= x(t) a1 (t + tn ) − b1 (t + tn )x(t)
dt
b2 (t + tn )y(t)
− ,
a(t + tn ) + b(t + tn )x(t) + c(t + tn )y(t) + d(t + tn )x(t)y(t)
dy(t) e(t + tn )y(t)
= y(t) a2 (t + tn ) − + k(t + tn ) . (8)
dt x(t)
Thus {u i (t +tn )} for i = 1, 2 and {u˙i (t +tn )} for i = 1, 2 are uniformly bounded and
equicontinuous. Hence the Ascoli’s theorem implies that ∃ a uniformly convergent
subsequence {u i (t + tm )} ⊂ {u i (t + tn )} and for any ε > 0 there exist k(ε) > 0 such
that
|u i (t + tk ) − u i (t + tm )| < ε, i = 1, 2. (9)
Thus one can deduce that u i (t) for i = 1, 2 are asymptotically almost periodic.
Hence {u i (t + tm )} can be written as sum of an almost periodic function u i1 (t + tm )
and a continuous function u i2 (t + tm ) (i = 1, 2) defined on R, and we have
u i (t + tm ) = u i2 (t + tm ) + u i1 (t + tm ) ∀ t ∈ R.
We also have that lim u i2 (t + tm ) = 0, lim u i1 (t + tm ) = u i1 (t), u i1 is an

m→∞ m→∞
almost periodic function. Thus we obtain that lim u i (t + tm ) = u i1 (t) (i = 1, 2).
m→∞
Furthermore,
u i (t + tm + h) − u i (t + tm ) u i (t + tm + h) − u i (t + tm )
lim u̇ i (t + tm ) = lim lim = lim lim
m→∞ m→∞ h→0 h h→0 m→∞ h
u i1 (t + h) − u i1 (t)
lim . (10)
h→0 h
Thus one can deduce that u̇ i1 exist for i = 1, 2. We have a sequence {tn } such that
tn → ∞ as n → ∞ for which
a1 (t + tn ) → ai (t), bi (t + tn ) → bi (t), a(t + tn ) → a(t), b(t + tn ) → b(t),

c(t + tn ) → c(t), d(t + tn ) → d(t), e(t + tn ) → e(t), k(t + tn ) → k(t) (i = 1, 2).
u̇ 11 = lim u̇ 1 (t + tn )
n→∞

= lim u 1 (t + tn ) a1 (t + tn ) − b1 (t + tn )u 1 (t + tn )
n→∞
b2 (t + tn )u 2 (t + tn )
−
a(t + tn ) + b(t + tn )u 1 (t + tn ) + c(t + tn )u 2 (t + tn ) + d(t + tn )u 1 (t + tn )u 2 (t + tn )
b2 (t)u 21 (t)
= u 11 (t) a1 (t) − b1 (t)u 11 (t) − ,
a(t) + b(t)u 11 (t) + c(t)u 21 (t) + d(t)u 11 (t)u 21 (t)
u̇ 21 = lim u̇ 2 (t + tn )
n→∞
e(t + tn )u 2 (t + tn )
= lim u 2 (t + tn ) a2 (t + tn ) −
n→∞ u 1 (t + tn ) + k(t + tn )
e(t)u 21 (t)
= u 21 (t) a2 (t) − .
u 11 + k(t)
Thus Theorem 4 implies that system (1) possess a unique positive almost periodic
solution.
4 Numerical Example
Consider the following predator–prey model system:
d x(t) √
= x(t) 9.9 + sin 5t − 10.9x(t)
dt
√
(0.3 + 0.19 sin 5t)y(t)
− √ √ ,
8 + cos 11t + (10 + sin 3t)x(t) + 5y(t) + 0.1x(t)y(t)
√
dy(t) √ (12 + 0.2 sin 13t)y(t)
= y(t) 0.5 + 0.29 sin 3t − . (11)
dt x(t) + 2
Here a1l = 8.9, a1u = 10.9, b1l = b1u = 10.9, b2l = 0.11, b2u = 0.49, a l = 7,
a u = 9, bl = 9, bu = 11, cl = cu = 5, d l = d u = 0.1, el = 11.79,
eu = 12.19, k l = k u = 2. And so S1 = 1, M2 = 0.2034, s1 = 0.8167, s2 = 0.0462.
bu M2 (bu + d u M2 )
Hence we have b1l = 10.9 > l 2 l
(a + b s1 + cl s2 + d l s1 s2 )2
e u M2 el
+ = 0.0057 + 0.9300 = 0.9357 and = 3.9333 >
(s1 + k )l 2 S1 + k u
b2u
(a + b s1 + cl s2 + d l s1 s2 )
l l
bu M2 (cu + d u M2 )
+ l 2l = 0.3428 + 0.0027 = 0.3455. This confirms that
(a + b s1 + cl s2 + d l s1 s2 )2
the parametric values involved with the model system (11) satisfy all the sufficient
conditions for the global attractivity of solutions obtained in Theorem 4. Hence the
model system (11) admits a unique, globally attractive, positive, and almost periodic
solution. The almost periodic coexistence have been depicted in Fig. 1.
1.5
x y
Solutions
0.5
0
0 50 100 150 200
time t
Fig. 1 Almost periodic solution of the nonautonomous system (11)

In this present work, a Lotka–Volterra model with modified Leslie–Gower and

Crowley–Martin functional response is considered. By constructing a suitable
Lyapunov function, a set of sufficient conditions for the existence of a unique, globally
attractive, almost periodic solution of the model system (1) is obtained. Furthermore,
an example is given to substantiate our analytical findings. Our results and example
generalize the results obtained in [8] and indicate that the dynamic behavior of the
considered model (global attractivity of almost periodic solution) depend upon the
mutual interference parameter d, i.e., the mutual interference among predators also
affects the global attractivity of solution at high prey density. Another important task
will be to consider similar type of predator–prey system with mutual interference
among prey species.
Acknowledgments The research work of first author (J.P. Tripathi) is supported by the Council of
Scientific and Industrial Research (CSIR) (No. 09/1058(0001)/2011-EMR-1), India.
References
1. Tripathi, J.P., Abbas, S., Thakur, M.: Local and global stability analysis of two prey one predator
model with help. Commun. Nonlinear Sci. Numer. Simulat. 19, 3284–3297 (2014)
2. Tripathi, J.P., Abbas, S., Thakur, M.: Stability analysis of two prey one predator model. AIP
Conf. Proc. 1479, 905–909 (2012)
3. Crowley, P.H., Martin, E.K.: Functional responses and interference within and between year
classes of a dragonfly population. J. North Amer. Benth. Soc. 8, 21–211 (1989)
4. Tripathi, J.P., Abbas, S.: Dynamical analysis of a prey predator model with Beddington DeAn-
gelis type function response incorporating a prey refuge. Nonlinear Dyn. 21, 99–111 (2014).
doi:10.1007/s11071-014-1859-2
5. Chen, F., Shi, C.: Global attractivity in an almost periodic multi-species nonlinear ecological
model. Appl. Math. Comput. 180, 376–392 (2006)
6. Abbas, S., Sen, M., Banerjee, M.: Almost periodic solution of a non-autonomous model of
phytoplankton allelopathy. Nonlinear Dynamics 67, 203–214 (2012)
7. Chen, F., Cao, X.: Existence of almost periodic solution in a ratio-dependent Leslie system
with feedback controls. J. Math. Anal. Appl. 341, 1399–1412 (2008)
8. Lin, X., Chen, F.: Almost periodic solution for a Volterra model with mutual interference and
Beddington-DeAngelis functional response. Appl. Math. Comput. 214, 548–556 (2009)
9. Leslie, P.H.: Some further notes on the use of matrices in population mathematics. Biometrica
35, 213–245 (1948)
10. Tripathi, J.P., Abbas, S., Thakur, M.: A density dependent delayed predator-prey model with
Beddington-DeAngelis type Function Response incorporating a prey refuge. Commun. Non-
linear Sci. Numer. Simulat. (2014). doi:10.1016/j.cnsns.2014.08.018
11. Fan, M., Kuang, Y.: Dynamics of a nonautonomous predator-prey system with Beddington-
DeAngelis functional response. J. Math. Anal. Appl. 295, 15–39 (2004)
Existence of a Mild Solution for Impulsive
Neutral Fractional Differential Equations
with Nonlocal Conditions
Alka Chadha and Dwijendra N. Pandey
Abstract In the present work, we study the existence of a mild solution of a

fractional-order differential equation with impulsive conditions in a Banach space
X. We establish the existence and uniqueness of the mild solution by using some
fixed-point theorems and resolvent semigroup theory.
Keywords Fractional calculus · Caputo derivative · Impulsive conditions ·

Resolvent operator · Neutral fractional differential equation
2010 Mathematics Subject Classification: 26A33 · 34K37 · 34K40 · 34K45 ·

35R11 · 45J05 · 45K05
1 Introduction
In recent few decades, fractional calculus has received much attention of researchers
mainly due to its demonstrated applications in widespread fields of science and
engineering, e.g., fluid flow, rheology dynamical, mechanics, electrical engineer-
ing, modeling of many physical phenomena, and so on. Fractional calculus has
been available and applicable to deal with real system characterized by power laws,
anomalous diffusion process, etc. The nonlinear oscillations of earthquake are one
of such important models. The deficiency of continuum traffic flow can be charac-
terized by the fractional derivative. For more details on fractional calculus, we refer
to [15–17, 20].
On the other hand, impulsive differential equations have played an important role
in real-world problems for describing a process which at certain moments changes
A. Chadha (B) · D.N. Pandey

e-mail: alkachadda23@gmail.com
D.N. Pandey
e-mail: dwij.iitk@gmail.com

DOI 10.1007/978-81-322-2485-3_26
320 A. Chadha and D.N. Pandey
their state rapidly and which cannot be described by using the classical differen-
tial problem. Such process is investigated in various fields such as biology, physics,
control theory, population dynamics, medicine, and so on. Impulsive differential
equations are an appropriate model to hereditary phenomena for which a delay argu-
ment arises in the modeling equations. For the general theory of such equations, we
refer to monographs [1, 2] and papers [3–5, 8, 11].
In this paper, our main objective is to establish the existence and uniqueness of
a solution for the fractional order neutral differential equation in a Banach space X
with norm · X ,
η
c
Dt [u(t) − F(t, u(h1 (t)))] = A[u(t) − F(t, u(h1 (t)))] + G(t, u(h2 (t))),
t ∈ [0, T ], t = tk 0 < T < ∞, (1.1)
u(tk ) = Ik (u(tk− )), k = 1, 2, . . . m, (1.2)
u(0) = u0 + g(u) ∈ X, (1.3)
η
where c Dt is the Caputo fractional derivative of order 0 < η < 1 and A : D(A) ⊂
X → X is a closed linear operator with dense domain D(A) in a Banach space X
and Ik : X → X, 0 = t0 < t1 < · · · < tm < tm+1 = T , u|t=tk = u(tk+ ) − u(tk− ),
and u(tk+ ) = limh→0+ u(tk + h) and u(tk− ) = limh→0− u(tk + h) denote the right
and left limits of u(t) at t = tk , respectively. The functions F, G, h1 , h2 , and g are
appropriate continuous functions to be specified later.
The organization of the paper is as follows: Sect. 2 gives some basic definitions,
Lemmas, and Theorems as preliminaries as these are useful for proving our results.
Section 3 focuses on proving the existence result of mild solution to problem (1.1)–
(1.3). Section 4 provides an example to illustrate the theory.
2 Preliminaries and Assumptions
In this section, we discuss some definitions and notations about sectorial opera-
tors, solution operator, and analytic solution operators required for establishing our
results.Throughout this paper, X is a complex Banach space equipped with the norm
· X . The symbol C([0, T ]; X) stands for the Banach space of all continuous func-
tions from [0, T ] into X with supremum norm, i.e., y[0,T ] = supt∈[0,T ] y(t). The
notation L(X, Y ) denotes the Banach spaces of all bounded linear operators from X
into Y with the operator norm denoted by · L(X,Y ) and when X = Y then we write
simply L(X) and · L(X) . In addition, PC([0, T ], X) represents the Banach space
of all the piecewise continuous functions from [0, T ] into X with the norm
uPC = max{ sup u(t + 0)X , sup u(t − 0)X },

t∈[0,T ] t∈[0,T ]
and Br (x, X) denotes a closed ball with center at x and radius r in X.

Existence of a Mild Solution for Impulsive Neutral Fractional Differential Equations … 321
To set the structure for our primary existence results, we recall the following
definitions.
Definition 1 The definition of one parameter Mittag-Leffler function is given by

∞
zk
Eα (z) = ,
Γ (αk + 1)
k=0
it is the standard definition of Mittag-Leffler function of one parameter and two

parameter function of Mittag-Leffler type is defined by
∞

zk 1 μα−β eμ
Eα,β (z) = = dμ, α, β > 0, z ∈ C,
Γ (αz + β) 2π i C μa − z
k=0
where C is a contour, which starts and ends at −∞ and encircles the disk |μ| ≤ |z|1/α
counterclockwise. The Laplace transform of the Mittag-Leffler is defined as
λα−β
L(t β−1 Eα, β (−ρ α t α )) = , Re λ > ρ 1/α , ρ > 0.
λα+ ρα
For more details we refer to [15].
Definition 2 The Riemann–Liouville fractional integral operator J of order η > 0

is defined by t
η 1
Jt F(t) = (t − s)η−1 F(s)ds, (2.1)
Γ (η) 0
where F ∈ L 1 ((0, T ); X).
Definition 3 The Riemann–Liouville fractional derivative is given by

η m−η
Dt F(t) = Dtm Jt F(t), m − 1 < η < m, m ∈ N, (2.2)
m m−η
where Dtm = dtd m , F ∈ L 1 ((0, T ); X), Jt F ∈ W m,1 ((0, T ); X). Here the notation
W m,1 ((0, T ); X) stands for the Sobolev space defined by

m−1
tk
W m,1 ((0, T ); X) = {y ∈ X : ∃z ∈ L 1 ((0, T ); X) : y(t) = dk (2.3)
k!
k=0
t m−1
+ ∗ z(t), t ∈ (0, T )}.
(m − 1)!
Note that z(t) = ym (t), dk = yk (0).

Definition 4 The Caputo fractional derivative is given by

t
η 1
c
Dt F(t) = (t − s)m−η−1 F m (s)ds, m − 1 < η < m, (2.4)
Γ (m − η) 0
where F ∈ C m−1 ((0, T ); X) ∩ L 1 ((0, T ); X) and the following holds
η η
tk
m−1
Jt (c Dt F(t)) = F(t) − F k (0). (2.5)
k!
k=0
The Laplace transform of the Caputo derivative of order η > 0 is given by
η

m−1
η
L[ c
Dt u(t); λ] = λ L[u(t)] − λη−k−1 uk (0), m − 1 < η < m. (2.6)
k=0
Definition 5 [3] An operator A, which is closed and linear, is called sectorial operator
if there are constants ω ∈ R, θ ∈ [π/2, π ], M > 0 such that the following two
conditions are satisfied:

(1) ρ(A) ⊃ (θ, ω) = {λ ∈ C : λ = ω, |arg(λ − ω)| < θ },

(2) R(λ, A)L(X) ≤ |λ−ω|M
, ω ∈ (θ, ω) ,
where ρ(A) be the resolvent set of A.
For more details we refer to [13]. Consider the following Cauchy problem for the
fractional evolution equation
η
c
Dt u(t) = Au(t), t > 0; u(0) = x, uk (0) = 0, k = 1, . . . , m − 1, (2.7)
where η > 0 and m = η.
Definition 6 [13] A family {Sη (t)}t≥0 ⊂ L(X) is called a solution operator for (2.7)
if the following conditions are satisfied:
(a) Sη (t) is strongly continuous for t ≥ 0 and Sη (0) = I;
(b) Sη (t)D(A) ⊂ D(A) and ASη (t)x = Sη (t)Ax, for all x ∈ D(A), t ≥ 0;
(c) Sη (t)x is a solution of (2.7), for all x ∈ D(A), t ≥ 0.
The solution operator Sη (t) of (2.7) is also defined by (see [13])
∞
λη−1 (λη I − A)−1 x = e−λt Sη (t)xdt, Re λ > ω, x ∈ X, (2.8)
0
where ω ≥ 0 and {λη : Reλ > ω} ⊂ ρ(A).

An operator A is said to belong to C η (X; M, ω), or C η (M, ω) if the problem

a solution operator Sη (t) satisfying ωt
(2.7) has
η η η
Sηη (t) ≤ Me , t ≥ 0. Denote
C (ω) = {C (M, ω); M ≥ 1}, or C = {C (ω; ω ≥ 0)} Bazhlekova [13].
Definition 7 [13] A solution operator Sη (t) of (2.7) is said to be analytic if Sη (t)
admits an analytic extension to a sector θ0 for some θ0 ∈ (0, π/2].
An analytic solution operator Sη (t) is said to be of analyticity type (θ0 , ω0 ) if
for each θ < θ0 and ω > ω0 there exists
a positive constant M = M(θ, ω) such
that Sη (t) ≤ Meω Re t , for t ∈ θ = {t ∈ C/{0} : |arg t| < θ }. Denote
A η (θ0 , ω0 ) = {A ∈ C η ; A generates analytic solution operator Sη (t) of type
(θ0 , ω0 )}.
Lemma 1 [13, 14] Let η ∈ (0, 2). A linear closed densely defined operator A
belongs to A η (θ0 , ω0 ) if and only if λη ∈ ρ(A) for each λ ∈ θ0 +π/2 (ω0 ), and for
any ω > ω0 , θ < θ0 , there exists a constant C = C(θ, ω) such that
C
λη−1 R(λη , A) ≤ , λ∈ (ω). (2.9)
|λ − ω|
θ+π/2
Now, we have following result for mild solution of Eqs. (2.12) and (2.13).
Theorem 1 Suppose A is a sectorial operator and f satisfies the uniform Hölder
condition with exponent β ∈ (0, 1], then
t
u(t) = Sη (t)x0 + Tη (t − s)f (s)ds, t ∈ [0, T ], (2.10)
0
where

1
Sη (t) = eλt λη−1 R(λη , A)dλ,
2π i Γ
(2.11)

1
Tη (t) = eλt R(λη , A)dλ,
2π i Γ
is the mild solution for fractional Cauchy problem

η
c
Dt u(t) = Au(t) + f (t), 0 < η < 1, t ∈ [0, T ], (2.12)
u(0) = x0 ∈ X, (2.13)

where Γ is a suitable path lying on θ, ω . For 0 < η < 1, Tη (t) is the η-resolvent
family and Sη (t) is the solution operator, generated by A.
If η ∈ (0, 1) and A ∈ Aη (θ0 , ω0 ), then for any x ∈ X and t > 0, we have
Sη (t)x ∈ D(A) and
Sη (t)L(X) ≤ Meωt , Tη (t)L(X) ≤ Ceωt (1 + t η−1 ), t > 0, ω > ω0 .

Let
S = sup Sη (t)L(X) , M
M T = sup Ceωt (1 + t 1−η ).
0≤t≤T 0≤t≤T
Thus we have
S , Tη (t)L(X) ≤ t η−1 M
Sη (t)L(X) ≤ M T .
For more details about solution operators, we refer to [6, 14], and references cited
in these papers.
Consider the set of functions
PC(I, X) = {u : I → X : u ∈ C((tk , tk+1 ], X), k = 0, 1, . . . , m and ∃

u(tk+ ) and u(tk− ), k = 1, . . . , m with u(tk− ) = u(tk )}, (2.14)
equipped with the norm

uPC = sup u(t)X ,
t∈I
which is a Banach space (PC(I, X), · PC ).

Now, we assume following assumptions on F, G, h1 , h2 , and Ik which will be
used later to establish main result.
(A1) For 0 < β < 1, the function Aβ F : [0, T ] × X → X is continuous and there
exists a constant LF > 0 such that
Aβ F(t, x) − Aβ F(s, y) ≤ LF [|t − s| + x − yX ], (2.15)
and
Aβ F(t, x) ≤ L1 x + L2 , (2.16)
for every x, y ∈ X and t, s ∈ [0, T ] and L1 and L2 .

(A2) The function G : [0, T ] × X → X is continuous and there exists a constant
LG > 0 such that
G(t, x) − G(t, y) ≤ LG x − yX , (2.17)
for every x, y ∈ X and t, ∈ [0, T ].

(A3) Ik : X → X, where k = 1, . . . , m are continuous functions and there exists a
constant L > 0 such that
Ik (x) − Ik (y) ≤ Lx − yX (2.18)
for each x, y ∈ X.
(A4) There exists a constant Lg > 0 such that
g(x) − g(y) ≤ Lg x − yX , x, y ∈ X, (2.19)

and
g(x) ≤ C1 x + C2 , x ∈ X. (2.20)
(A5) hi ∈ C([0, T ], [0, T ]), i = 1, 2.
Next, we are giving the definition of the mild solution for the problem (1.1)–(1.3).
Definition 8 The function u : [0, T ] → X is said to be a mild solution of Eqs. (1.1)–
(1.3) if u(·) ∈ PC([0, T ], X) satisfying the following integral equation
u(t) = Sη (t)[u0 + g(u) − F(0, u(h1 (0)))] + F(t, u(h1 (t)))

t
+ Tη (t − s)G(s, u(h2 (s)))ds
0

m
+ Sη (t − ti )Ii (u(ti )), (2.21)
i=1
for each t ∈ [0, T ]. and also satisfies the following impulsive conditions u|t=ti =
Ii (u(ti− )), i = 1, . . . , m.
3 Existence Results
In this section, we establish the existence and uniqueness of a mild solution of

Eqs. (1.1)–(1.3) using Banach fixed-point theorem.
Theorem 2 Let (A1)–(A5) holds and
T LG
S (Lg + LF ) + LF A−β + M Tη S L < 1.
R=M + mM (3.1)
η
Then impulsive problem (1.1)–(1.3) has a unique mild solution u ∈ X.
Proof Let u0 ∈ X be fixed. Define a mapping Q : PC([0, T ]; X) → PC([0, T ]; X)

such that
(Qu)(t) = Sη (t)[u0 + g(u) − F(0, u(h1 (0)))] + F(t, u(h1 (t)))

t
+ Tη (t − s)G(s, u(h2 (s)))ds
0

m
+ Sη (t − ti )Ii (u(ti )), (3.2)
i=1
for each t ∈ [0, T ]. Since F and G are continuous functions and Sη (t), t ≥ 0
and Rη (t), t ≥ 0 are compact, thus it is easy to show that the map Q is well
defined on PC([0, T ]; X). To establish the result, it is sufficient to show that the
mapping Q is a contraction mapping on PC([0, T ]; X). To this end, let t ∈ [0, T ] and
u∗ , u∗∗ ∈ PC([0, T ]; X). Thus, we obtain
(Qu∗ )(t) − (Qu∗∗ )(t)
≤ Sη (t)[g(u∗ ) − g(u∗∗ ) − (F(0, u∗ (h1 (0))) − F(0, u∗∗ (h1 (0))))]
+ F(t, u∗ (h1 (t))) − F(t, u∗∗ (h1 (t)))
t
+ Tη (t − s) × G(s, u∗ (h2 (s))) − G(s, u∗∗ (h2 (s)))ds
0

m
+ Sη (t − ti )[Ii (u∗ (ti )) − Ii (u∗∗ (ti ))],
i=1
t
S [Lg + LF ] + LF A−β )u∗ − u∗∗ X + M
≤ (M T (t − s)η−1 LG u∗ − u∗∗ X ds
0
S Lu∗ − u∗∗ X ,
+ mM
η
≤ [M T LG T + m M
S (Lg + LF ) + LF A−β + M S L]u∗ − u∗∗ X . (3.3)
η
Since R < 1 by the inequality (3.1), it indicates that the map Q is a strict contraction
on PC([0, T ]; X). Hence, by Banach fixed-point theorem, there exists a unique fixed
u ∈ PC([0, T ]; X) such that Qu(t) = u(t) which is a mild solution of the problem
(1.1)–(1.3). Therefore, the proof of the theorem is completed.
Our second result is based on the Schaefer’s fixed-point theorem. Schaefer’s the-
orem is a special case of the far-reaching Leray–Schauder theorem which was dis-
covered earlier by Juliusz Schauder and Jean Leray. The statement is as follows:
Theorem 3 Let Q : X → X be a continuous and a compact map such that the set
{x ∈ X : x = λPx for some 0 ≤ λ ≤ 1} is bounded, then Q has a fixed point.
For this second result, we need to assume a new set of assumptions on G, F & Ik ’s,
where k = 1, 2, . . . , m.
Assumptions
(A6) G : [0, T ] × X → X is a continuous function and there exists a continu-
ous function mG : [0, T ] → (0, ∞) and continuous nondecreasing function
W : [0, ∞) → (0, ∞) such that
G(t, x)X ≤ mG (t)W ( x), (t, x) ∈ [0, T ] × X, (3.4)

∞
and 0 Wds(s) < ∞.
(A7) The functions Ik : X → X are completely continuous and there exist Ω > 0
such that
Ω= max { Ik (x)X }. (3.5)
1≤k≤m, x∈X
(A8) The function F is compact satisfying hypotheses (A1).

(A9) The operator families Sη (t), t ≥ 0 and Tη (t), t ≥ 0 are compacts.
S ) < 1 and
Theorem 4 Suppose (A4)–(A9) holds. If A ∈ A η (θ0 , ω0 ) and LF (1 + M
T T η
M T ∞ ds
η(1−L (1+M
1 )) 0 mG (s)ds < ω1 W (s) , where
S
MS S m
M 1
ω1 = u0 + Ω(1 + LF ) + S )
L (1 + M
S )
1 − L1 (1 + M S )
1 − L1 (1 + M S ) 2
1 − L1 (1 + M
holds, then there exist at least one mild solution of the impulsive problem (1.1)–(1.3)
on [0,T].
Proof Consider an operator Q : PC([0, T ]; X) → PC([0, T ]; X) as in Theorem 2.

It can be easily proved that map Q is well defined on PC([0, T ]; X).
Step 1: We show the continuity of the map Q.

To prove the continuity, let un be sequence in PC([0, T ]; X) such that limn→∞ un
(t) = u(t), i.e., un → u as n → ∞ in PC([0, T ]; X). Since G and F are continuous.
Therefore, by the continuity of G, F, g, and Ii , i = 1, . . . , m, we deduce that
G(t, un (h2 (t))) → G(t, u(h2 (t))), as n → ∞, (3.6)

F(t, un (h1 (t))) → F(t, u(h1 (t))), as n → ∞, (3.7)
g(un ) → g(u), as n → ∞, (3.8)
Ii (un (ti )) → Ii (u(ti )), as n → ∞. (3.9)
Now for every t ∈ [0, T ], we have
(Qun )(t) − (Qu)(t) ≤ Sη (t)[g(un ) − g(u) − (F(0, un (h1 (0))) − F(0, u(h1 (0))))]
+ F(t, un (h1 (t))) − F(t, u(h1 (t)))
t
+ Tη (t − s) · G(s, un (h2 (s))) − G(s, u(h2 (s)))ds
0

m
+ Sη (t − ti )[Ii (u(ti )) − Ii (u(ti ))],
i=1
S [g(un ) − g(u) + F(0, un (h1 (0))) − F(0, u(h1 (0)))]
≤M
+ F(t, un (h1 (t))) − F(t, u(h1 (t)))
t
+MT (t − s)η−1 G(s, un (h2 (s))) − G(s, u(h2 (s)))ds
0

m
S
+M [Ii (u(ti )) − Ii (u(ti ))]. (3.10)
i=1
Then by the continuity of G, F, g, and Ii and dominated convergence theorem, we

get that Qun (t) converges to Qu(t) in X, i.e., limn→∞ Qn u(t) = Qu(t) in X for each
t ∈ [0, T ]. Hence this proves the continuity of the map Q.
Step 2: Second, we show that Q maps bounded sets into bounded sets in
PC([0, T ]; X). To prove the result, it is enough to show that for any r > 0 there
exists γ > 0 such that QuPC ≤ γ for each u ∈ Br (PC) = {u ∈ PC([0, T ]; X) :
uPC ≤ r}. Let G1 = sup G(t, u(h2 (t))), then for any u ∈ Br (PC),
t∈I, u∈Br
t ∈ [0, T ], we have
Qu(t)X ≤ S [ u0 + g(u) + L1 r + L2 ] +
M sup F(t, u(h1 (t)))
0≤t≤T , u∈Br

M η
T T G1

m
+ + Sη (t − ti )Ii (u(ti )),
η
i=1
η
≤MS [ u0 + C1 r + C2 + L1 r + L2 ] + L1 r + L2 + MT T G1 + mM
S Ω,
η
= γ. (3.11)
Thus, we conclude that Qu(t) ≤ γ .
Step 3: Q maps bounded sets into equicontinuous sets of PC([0, T ]; X).

To this end, we show that Q(Br ) is equicontinuous. Take 0 ≤ τ < t ≤ T and
u ∈ C([0, T ]; X), we have
Qu(t) − Qu(τ ) ≤ [Sη (t) − Sη (τ )](u0 + g(u) − F(0, u(h1 (0))))

+ F(t, u(h1 (t))) − F(τ, u(h1 (τ )))
t τ
+ Tη (t − s)G(s, u(h2 (s)))ds − Tη (τ − s)G(s, u(h2 (s)))ds,
0 0

m
+ [Sη (t − ti ) − Sη (τ − ti )]Ii (u(ti )),
i=1
≤ [Sη (t) − Sη (τ )][ u0 + F(0, u(h1 (0)))] + LF (|t − τ |)

t
+M T (t − s)η−1 G(s, u(h2 (s)))ds,
τ
t
+M T [(t − s)η−1 − (τ − s)η−1 ] G(s, u(h2 (s)))ds
0

m
+ Sη (t − ti ) − Sη (τ − ti )Ii (u(ti )). (3.12)
i=1
Since Sη (t), t > 0 and Tη (t), t > 0 are compact, therefore Qu(t) − Qu(τ ) → 0
as t → τ . Therefore, Qu is equicontinuous on [0, T ]. Hence we conclude that Qu(t)
is equicontinuous on [0, T ]
Step 4: Q maps Br into a compact set in X.

For this, we decompose Q into Q1 and Q2 , where
(Q1 u)(t) = Sη (t)[g(u) − F(0, u(h1 (0)))] + F(t, u(h1 (t)))

t
+ Tη (t − s)G(s, u(h2 (u(s))))ds, t ∈ [0, T ], (3.13)
0

m
(Q2 u)(t) = Sη (t − ti )Ii (u(ti )), t ∈ [0, T ]. (3.14)
i=1
We now prove that {Q1 u(t), u ∈ Br } is relatively compact on X, for all t ∈ [0, T ].
It is obvious that the set {Q1 u(t), u ∈ Br } is relatively compact in X for t = 0. Let
0 < t ≤ T be fixed and 0 < ε < t. For u ∈ Br define an operator Q1,ε by
Q1,ε u(t) = Sη (t)[g(u) − F(0, u(h1 (0)))] + F(t, u(h1 (t)))

t−ε
+ Tη (t − s)G(s, u(h2 (u(s))))ds, t ∈ [0, T ]. (3.15)
0
Since Sη (t) and Tη (t) are compact, the set {Q1,ε u(t), u ∈ Br } is relatively compact
in X for every ε, 0 < ε < t. Moreover, for every u ∈ Br , we have
t
Q1 u(t) − Q1,ε u(t) ≤ Tη (t − s) G(s, u(h2 (s)))ds, (3.16)
t−ε
therefore, taking ε → 0 we can easily see that there are relatively compact sets arbi-
trarily close to the set {Q1 u(t), u ∈ Br }. Hence the set {Q1 u(t), u ∈ Br } is relatively
compact in X and by (A8) we conclude that Q1 is compact for all t ∈ [0, T ]. Next,
we show that {Q2 u(t), u ∈ B r } is relatively compact in X, for all t ∈ [0, T ]. For
t ∈ [0, T ] we have Q2 u(t) = m i=1 Sη (t − ti )Ii (u(ti )) which is equicontinuous and
bounded, by (A9) it follows that {Q2 u(t), u ∈ Br } is relatively compact subset of
X, for all t ∈ [0, T ]. Hence, by (A6)–(A8) and Arzela–Ascoli theorem, we conclude
that Q2 is compact for all t ∈ [0, T ]. Therefore, Q = Q1 + Q2 is compact.
Step 5: (A priori bounds) We prove that the set E = {u ∈ PC([0, T ]; X) such that
u = λQu for some 0 < λ < 1} is bounded.
Let u ∈ E with u(t) = λQu(t) for some 0 < λ < 1. Then for each t ∈ [0, T ],
S [ u0 + C1 u(t) + C2 ] + M
u(t)X ≤ λ[M S (L1 u(t) + L2 ) + L1 u(t) + L2
t
+M T (t − s)η−1 G(s, u(h2 (s)))ds + mMS Ω],
0
S u0 + [L1 (M
≤ λ[M S + 1) + MS C1 ] u(t) + L2 (1 + M
S ) + M
S C2
η t
M TT S Ω].
+ mG (s)W ( u(s))ds + mM (3.17)
η 0
Therefore, for all t ∈ [0, T ], by the Young inequality ([13], p. 6), we get
S
M
MS C2
u(t)X ≤ u0 +
1 − L1 (1 + MS ) − MS C1 S ) − M
1 − L1 (1 + M S C1
L2 (1 + MS ) + mM S Ω
+

1 − L1 (1 + MS ) − M S C1
T T η t
M
+ m (s)W ( u(s))ds,
η(1 − L1 (1 + MS )) − MS C1 0 G
T T η t
M
≤ ω1 + m (s)W ( u(s))ds
η(1 − L1 (1 + M S ) − MS C1 ) 0 G
S
M
M S C2 S ) + mM
L2 (1+M S Ω
where ω1 = u0 + + . Then
S ) − M
1−L1 (1+M S C1 S ) − M
1−L1 (1+M S C1 S ) − M
1−L1 (1+M S C1
for all t ∈ [0, T ],
T T η t
M
u(t) ≤ βλ (t) ω1 + m (s)W ( u(s))ds.
S C1 ) 0 G
S ) − M
η(1 − L1 (1 + M
Calculating βλ (t) for t ∈ [0, T ], we obtain
T T η
M
βλ (t) ≤ m (t)W ( u(t)).
S C1 ) G
S ) − M
η(1 − L1 (1 + M
Thus we have
dβλ (t) dβλ (t) T T η

M
≤ ≤ m (t)dt. (3.18)
W ( βλ (t)) W ( u(t)) η(1 − L1 (1 + M S C1 ) G
S ) − M
Since W (s) is positive and nondecreasing. Integrating both sides, we get

βλ (t) T T η T ∞
ds M ds
≤ mG (s)ds < , (3.19)
ω1 W (s)
η(1 − L1 (1 + MS ) − MS C1 ) 0 ω1 W (s)
where, we have βλ (0) = ω1 , βλ (t) is positive and nondecreasing. Hence, from the
above inequality, we obtain that the set of functions {βλ : λ ∈ (0, 1)} is bounded.
This implies that set {u ∈ PC([0, T ]; X) : u = λQu, 0 < λ < 1} is bounded in X.
Hence by Schaefer’s fixed-point theorem, we get that Q has a fixed point on [0, T ].
4 Application
We consider the following fractional-order impulsive partial functional differential

system of the form

t

∂η ∂
z(t, x) + b(t, ξ, x) z(sin t, ξ ) + z(sin t, ξ ) dξ
∂t η 0 ∂ξ

t

∂2 ∂ ∂
= 2 z(t, x) + b(t, ξ, x) z(sin t, ξ ) + z(sin t, ξ ) dξ + χ t, z(sin t, x) ,
∂x 0 ∂ξ ∂x
0 ≤ t ≤ 1, 0 ≤ x ≤ π, (4.1)
z(t, 0) = z(t, π ) = 0, (4.2)

z(0, x) = z0 (x), 0 ≤ x ≤ π, (4.3)
z|tk = z(tt+k ) − z(tt−k ) = Ik (z(tt−k )), k = 1, . . . , m, (4.4)
where 0 < η < 1 and 0 < t1 < t2 < · · · < tm < 1 and b : [0, 1] × [0, π ] × [0, π ]
→ R and χ : [0, 1] × R → R are continuous functions. Take X = L 2 [0, π ] and let
an operator A such that
Af = f (4.5)
with the domain
D(A) = H 2 ([0, π ]) = {f (·) ∈ X : f , f ∈ X and f (0) = f (π ) = 0}. (4.6)
It implies that A generates a strongly continuous semigroup T (·) which is analytic

∞
e−n t (f , zn )zn . On the other hand, A
2
and semi-adjoint and is given by T (t)f =
n=1
has a discrete spectrum, the eigenvalues are −n2 , n ∈ N with the corresponding
normalized eigenvectors zn (x) = π2 sin(nx). We have that if f ∈ D(A) then Af =
∞
n2 (f , zn )zn , for all f ∈ X and t > 0.
n=1
Now we assume following assumptions:
(i) b : [0, 1] × [0, π ] × [0, π ] → R is continuously differentiable with b(t, ξ, 0) =
b(t, ξ, π ) = 0.
(ii) The function b is measurable and
π π
sup b2 (t, ξ, x)dξ dx < ∞, (4.7)
0≤t≤1 0 0
∂2
and function ∂x 2
is measurable and
2 1/2
π π ∂2
K1 = sup b(t, ξ, x) dξ dx < ∞. (4.8)
0≤t≤1 0 0 ∂x 2
(iii) χ : [0, 1]×R → R is Lipschitz continuous with respect to the second argument
and there exists positive constant a0 such that
χ (t, x1 ) − χ (t, x2 ) ≤ a0 x1 − x2 , (4.9)
for t ∈ [0, 1], x1 , x2 ∈ R.

(iv) Ik ∈ C(X, X), k = 1, 2, . . . , m such that
Ik (x) ≤ ψk ( x), (4.10)
for x ∈ X, where ψk ∈ (J, R+ ) is nondecreasing function.

π
Let h1 (t) = h2 (t) = sin t, F(t, z)(x) = 0 b(t, ξ, x)[z(ξ ), z (ξ )]dξ, and G(t, z)
(x) = χ (t, z (ξ )).
Therefore, the Eqs. (4.1) and (4.2) can be reformulated as
dη
[u(t) + F(t, u(h1 (t)))] = A[u(t) + F(t, u(h1 (t)))] + G(t, u(h2 (t))), 0 ≤ t ≤ 1,
dt η
u(0) = u0 ,
u|tk = Ik (u(tt−k )), k = 1, 2, . . . , m. (4.11)
It is easy to verify that F and G satisfy the condition A1 and A2, respectively, and
from (ii) it is clear that F(t, z) is bounded linear operator on R. Thus from Theorem 2,
the system (4.11)–(4.11) admits a mild solution [0, T ] as well as (4.1)–(4.2).
Acknowledgments The authors would like to thank the referee for valuable comments and sug-
gestions. The work of the first author is supported by the University Grants Commission (UGC),
Government of India, New Delhi.
References
1. Benchohra, M., Henderson, J., Ntouyas, S.K.: Impulsive differential equations and inclusions.
Contemporary Mathematics and Its Applications, vol. 2. Hindawi Publishing Corporation,
New York (2006)
2. Lakshmikantham, V., Baǐnov, D., Simeonov, P.S.: Theory of Impulsive Differential Equations.
World Scientific, Singapore-London (1989)
3. Shu, X.-B., Lai, Y., Chen, Y.: The existence of mild solutions for impulsive fractional partial
differential equations. Nonlinear Anal.: TMA 74, 2003–2011 (2011)
4. Zhang, X., Huang, X., Liu, Z.: The existence and uniqueness of mild solutions for impulsive
fractional equations with nonlocal conditions and infinite delay. Noninear Anal. Hybrid Syst.
4, 775–781 (2010)
5. Balachandran, K., Samuel, F.P.: Existence of mild solutions for quasilinear integrodifferential
equations with impulsive conditions. Elect. J. Diff. Equ. 84, 1–9 (2009)
6. Lizama, C.: Regularized solutions for abstract Volterra equations. J. Math. Anal. Appl. 243,
278–292 (2000)
7. Araya, D., Lizama, C.: Almost automorphic mild solutions to fractional differential equations.
Nonlinear Anal.: TMA. 69, 3692–3705 (2008)
8. Abbas, M.I.: Existence for fractional order impulsive integrodifferential inclusions with non-
local initial conditions. Int. J. Math. Anal. 6, 1813–1828 (2012)
9. Ezzinbia, K., Fu, X., Hilal, K.: Existence and regularity in the α-norm for some neutral partial
differential equations with nonlocal conditions. Nonlinear Anal.: TMA. 67, 1613–1622 (2007)
10. Ezzinbia, K., Fu, X.: Existence and regularity of solutions for some neutral partial differential
equations with nonlocal conditions. Nonlinear Anal.: TMA. 57, 1029–1041 (2004)
11. Mophou, M.G.: Existence and uniqueness of mild solutions to impulsive fractional differential
equations. Nonlinear Analy.: TMA. 72, 1604–1615 (2010)
12. Pazy, A.: Semi-groups of Linear Operator and Applications of Partial Differential Equations.
Springer verlag, New York (1983)
13. Bazhlekova, E.: Fractional evolution equations in Banach spaces, Ph.D. Thesis. Eindhoven
University of Technology (2001)
14. Prüss, J.: Evolutionary Integral Equations and Applications (Monographs Math), vol. 87.
Birkhauser-Verlag, New York (1993)
15. Podlubny, I.: Fractional Differential Equations. Academic press, New York (1993)
16. Miller, K.S., Ross, B.: An Introduction to the Fractional Calculus and Fractional Differential
Equations. John Wiley and Sons Inc, New York (1993)
17. Samko, S.G., Kilbas, A.A., Marichev, O.I.: Fractional Integrals and Derivatives: Theory and
Applications. Gordon and Breach Science Publisher, Yverdon (1993)
fractional integro-differential equation with infinite delay. Math. Comp. Model. 57, 754–763
(2013)
19. Li, K., Peng, J., Jia, J.: Cauchy problems for fractional differential equations with Riemann-
Liouville fractional derivatives. J. Fract. Anal. 263, 476–510 (2012)
Equations. Elsevier, Amsterdam (2006)
21. Wang, J., Fec̆kan, M., Zhou, Y.: On the new concept of solution and existence results for
impulsive fractional evolution equations. Dyn. PDE 8, 345–361 (2011)
22. Fec̆kan, M., Zhou, Y., Wang, J.: On the concept and existence of solution for impulsive fractional
differential equations. Commun. Nonlinear Sci. Numer. Simulat. 17, 3050–3060 (2012)
23. Wang, J., Li, X., Wei, W.: On the natural solution of an impulsive fractional differential equation
of order q∈(1,2). Commun. Nonlinear Sci. Numer. Simulat. 17, 4384–4394 (2012)
24. Granas, A., Dugundji, J.: Fixed Point Theory. Springer-Verlag, New York (2003)
Numerical Solution of Highly Oscillatory
Nonlinear Integrals Using Quasi-Monte
Carlo Methods
Nageswara Rao Narni
Abstract Highly oscillating integrals occur in many engineering applications. This

paper discusses the quasi-Monte Carlo methods for calculation of the highly oscillat-
ing integrals using a low discrepancy sequence. We evaluated the highly oscillating
integrals using a low discrepancy sequence known as Vander Corput sequence. The
theoretical error bounds are calculated and are compared with analytical results.
The reliability of the quasi-Monte Carlo methods is compared with He’s homotopy
perturbation method.
Keywords Highly oscillatory integrals · Quasi Monte-Carlo methods · Low dis-

crepancy sequence
1 Introduction
Highly oscillatory functions arise in wide range of applications in science and engi-
neering. The integration of high oscillatory functions is a challenging task from
several years. Most of the techniques or analysis for integration of highly oscillatory
functions are problem-oriented or technique-oriented. For example, integration of
these functions occurs in solving the problems modeling of wave phenomena like
diffraction of light, scattering of acoustic waves [8], scattering of electromagnetic
waves [11], etc. The boundary element method also requires the evaluation of highly
oscillatory integrals [3]. Explicit solution exists only for a few cases. So one needs
to go for numerical methods.
The main goal of the present paper is on the analysis and computation of the
integrals of the form
N.R. Narni (B)

Rajiv Gandhi University of Knowledge Technologies,
RK Valley, YSR Dist., AP 516329, India
e-mail: narninrao@gmail.com

DOI 10.1007/978-81-322-2485-3_27
336 N.R. Narni
Fig. 1 Error in Gaussian quadrature with f (x) = cos(x), g(x) = x 2 for the quadrature points 5,
10, and 16

I [ f, Ω] = f (x)eiωg(x) d V, (1)
Ω
where Ω ⊂ Rn is bounded and open domain with piecewise smooth boundary. The
functions f, g ∈ C ∞ are smooth. For large values of |ω|, the integral (1) oscillates
rapidly as a function of ω.
A classical technique to compute (1) is Gaussian quadrature [6] method. If the
integrand oscillates rapidly (for large values of ω), the Gaussian quadrature methods
are not appropriate. For example, let us consider the following integral
1 2
cos(x)eiωx d x. (2)
0
The integral is evaluated using Gauss–Legendre quadrature rule at different

quadrature points. One may observe that the Gauss–Legendre quadrature rule gives
good results for small values of ω. As ω becomes large in comparison with quadra-
ture points, high oscillations set in and the error becomes O(1). The absolute error
with respect to ω is plotted in the Fig. 1.
There exists few techniques to calculate the highly oscillating integrals. Among
all Asymptotic expansion methods, Filon-type methods, and Levin-type methods are
most popular. The Asymptotic method in a straightforward manner is nothing but
Numerical Solution of Highly Oscillatory Nonlinear Integrals … 337
repeatedly applying the integration by parts. But the accuracy of the asymptotic
expansion is limited due to the divergence of the series.
An even better approach is Filon [2]-type method. In this method instead of
approximating whole integral, we approximate f (x) of Eq. (1) at a set of quadrature
nodes c1 , . . . , cν , by a polynomial f˜. Evaluation of the moments makes Filon-type
methods difficult to certain type of applications.
In Levin [9]-type method we collocate the integrand at specific points. The Levin-
type method is advantageous over Filon-type method and it is due to the fact that
Levin-type method works easily on all types of domains and nonlinear oscillators.
Most of the methods used in the current research are either Filon- or Levin-type
methods or a modified form of these methods [1, 14]. He’s homotopy Perturbation
Method (HPM) is used in [10] for the numerical solution of the highly oscillating
integrals.
In spite of all these methods, new applications continuously give rise to situations
where straightforward application of these formulas are either inefficient or simply
not possible. For example, if the function to be integrated contains critical point, then
both Filon-type method and Levin collocation method are not accurate. So there is a
need to device new methods for the integration of highly oscillatory functions.
The Monte Carlo method can be used to approximate the definite integral. This
method gives the accuracy O √1
n
, which is not at all competitive with good algo-
rithms, such as the Romberg method [6]. The present paper proposes the applica-
tion of quasi-Monte Carlo methods for the numerical integration of highly oscilla-
tory functions. In these methods selection of abscissas are based on Vander Corput
sequence [7], which is a low discrepancy sequence.
In Sect. 2, an introduction to quasi-Monte Carlo methods, low discrepancy
sequences, and Vander Corput sequence are presented. Section 3 gives the error
bounds for quasi-Monte Carlo integration of the highly oscillating integrals. In
Sect. 4, the quasi-Monte Carlo method with a Vander Corput sequence is applied
to various problems. The efficiency of the QMC method is compared with other
methods. Conclusions are drawn and are discussed in Sect. 5.
2 Quasi-Monte Carlo Methods
The only difference between the Monte Carlo and quasi-Monte Carlo methods is the
selection of abscissa set {xi } (grid points). In Monte Carlo methods the abscissa are
generated as a set of random number, whereas in quasi-Monte Carlo methods the
quadrature nodes are calculated from deterministic algorithms.
338 N.R. Narni
2.1 Low Discrepancy Sequences
Now we introduce a quantity (the so-called discrepancy of the sequence) that mea-
sures the deviation of the sequence from an ideal distribution. This measure enables
us to distinguish between good and bad sequences.
Definition 1 [7] For a nonempty set M of measurable subsets of C N+ , the

discrepancy DnM of the finite sequence x1 , x2 , . . . , xn ∈ C N+ with respect to M
is defined by

A(E; n)

DnM (x1 , x2 , . . . , xn ) := sup − C E (x)d x .
E∈M
n +
CN
n
where A(E; n) := i=1 C E (x i ) counts the number of points xi ∈ E and E ⊆ C N+ .
2.2 Vander Corput Sequence
In this subsection, we are going to describe a low discrepancy sequence known

as Vander Corput sequence [7]. In fact, this is the only infinite sequence having a
uniformly smaller discrepancy than any other sequence exists upto now. Therefore,
we have chosen this sequence for our numerical integration.
We define thes so-called Vander Corput sequence {xn } as follows: For n ≥ 1,
let n − 1 = a
j=0 j 2 j be the dyadic expansion of n − 1. Then we set x =
n
s − j−1 . The sequence {x } is then clearly contained in the unit interval. The
j=0 a j 2 n
following theorem gives the bounds for discrepancy of the Vander Corput sequence.
Theorem 1 The discrepancy D N (ζ ) of the Vander Corput sequence ζ = {x N }

satisfies
ln(N + 1)
D N (ζ ) ≤
N ln 2
for N grid points.
Proof See [7].
3 Error Bounds for Quasi-Monte Carlo Methods
The selection of a numerical scheme is generally based on its accuracy (error bound),
convergence, and computational cost. In this section, we are going to analyze these
characters for quasi-Monte Carlo methods.
3.1 Variation of a Function
The variation of a univariate real function f : [a, b] → R characterizes the regularity

of f on the interval [a, b]. For a partition P of the interval [a, b] into n subintervals,
P : {xi : a = x0 < x1 < · · · < x N −1 < x N = b},
the sum

n
V ( f ; P) := | f (xi ) − f (xi−1 ) |
i=1
measures the discrete variation of f with respect to the specific partition P. The
continuous variation of f can be characterized by the supremum of all such discrete
variations V ( f ; P).
Definition 2 [12] Variation of a univariate function. The variation of a univariate
function f : [a, b] → R is defined as

n
V ( f ) := sup {V ( f ; P)} = sup | f (xi ) − f (xi−1 ) | .
P i=1
If V ( f ) is finite, f is said to be of bounded variation on [a, b]. If f is continuously

differentiable then the relationship holds.
b
V( f ) = | f
(x) | d x
a
The variation of nonlinear oscillator is obtained as follows:

Proposition 1 Suppose f (x), g(x) are two continuously differentiable real valued
functions with a finite bounded variation on [a, b] and let
b
I = f (x)eiωg(x) d x,
a
then the bounded variation of the integrand is calculated as

b
BV f (x)e iωg(x)
≤ iωg
(x) f (x) + f
(x) d x
a
Proof Let the integrand be denoted as
φ(x) = f (x)eiωg(x)
Since φ(x) is piecewise smooth then the bounded variation is calculated as

340 N.R. Narni
b
BV (φ) = |φ
(x)|d x
a
b

= iωg
(x) f (x)eiωg(x) + f
(x)eiωg(x) d x
a
b
∴ BV (φ) ≤ iωg
(x) f (x) + f
(x) eiωg(x) d x
a
b
≤ iωg
(x) f (x) + f
(x) d x
a
Corollary 1 Suppose g(x) = Constant, then

b

BV f (x)eiωg(x) ≤ f (x) d x
a
which corresponds to normal integration.
Corollary 2 Suppose g(x) = x, i.e., linear oscillator, then

b
BV f (x)eiωg(x) ≤ iω f (x) + f
(x) d x
a
But it is not always possible to calculate the variation of the functions.
3.2 Error Bounds
Now we discuss the error bounds for quasi-Monte Carlo approximation for more
general integration domains. All these bounds depend on the variation of the inte-
grand which involves the oscillatory parameter ω. A classical result is the following
inequality of Koksma [7].
Theorem 2 If f has bounded variation V ( f ) on [0, 1], then, for any sequence
x1 , x2 , . . . , x N ∈ [0, 1], we have
1
1
N

f (xn ) − f (u)du ≤ V ( f )D N (x1 , x2 , . . . , x N ). (3)
N 0
n=1
Proof We can assume that x1 ≤ x2 ≤ · · · ≤ x N . Put x0 = 0 and x N +1 = 1. Using

summation by parts and integration by parts, we obtain
1 1
1
N N
n
f (xn ) − f (u)du = − ( f (xn+1 ) − f (xn )) + ud f (u)
N 0 N 0
n=1 n=0
N xn+1
n
= u− d f (u).
N
n=0 xn
For fixed n with 0 ≤ n ≤ N , we have

n

u − ≤ D N (x1 , x2 , . . . , x N ) for xn ≤ u ≤ xn+1
N
1 N xn+1
1
N

∴ f (xn ) − f (u)du ≤ D N (x1 , x2 , . . . , x N ) |d f (u)|
N 0 xn
n=1 n=0

N
≤ D N (x1 , x2 , . . . , x N ) | f (xn+1 ) − f (xn )|
n=1
≤ D N (x1 , x2 , . . . , x N )V ( f ).
Hence we get the desired inequality.
The Koksma’s inequality is applicable for C ∞ functions also.
In this section, we consider two different example problems. We evaluated the inte-
grals using quasi-Monte Carlo methods with Vander Corput sequence and error
bounds are calculated using Koksma’s inequality.
Example 1 In this example, we consider the highly oscillating integrals of the form
b
I = eiωg(x) d x
a
where g
(0) = g
(0) = · · · = g (r −1) (0) = 0 and g (r ) (x) = 0, for all x ∈ [0, 1].

This oscillator is known as irregular oscillator, where g is real.
In particular let us consider the integral
√ √
1
iωx 2 erf( −iω) π
e dx = √
0 2 −iω
342 N.R. Narni
It can be observed that the above integral has a unique critical point 0 of g(x) in
[0, 1]. Therefore, quasi-Monte Carlo method can be applied to evaluate the integral
by using Vander Corput sequence.
Now we calculate the error bound for the given integral as follows: As we know
f (x) = 1, and g(x) = x 2 and is continuous and differentiable. Therefore, the
variation of φ(x) can be calculated as
1 1
2
V (φ) = |φ
(x)|d x = 2iωxeiωx d x
0 0
1
iωx 2
∴ V (φ) = 2ω xe dx
0
≤ 2ω
By Koksma inequality (3), we get

1
1
N

f (xn ) − f (u)du ≤ ωD N (x1 , x2 , . . . , x N ).
N 0
n=1
From Theorem 1, we know that
ln(N + 1)
D N (x1 , x2 , . . . , x N ) ≤
N ln 2
Therefore, the error bound for this integral is obtained as
ln(N + 1)
Error Bound ≤ ω (4)
N ln 2
From this Eq. (4), we can observe that the error bound is dependent on both oscillating
parameter ω and number of quadrature points N . The absolute error of the numerical
scheme is plotted in Fig. 2 for N = ω. We can observe that the error is of order
ln(N ).
Example 2 Consider the highly oscillatory integral [13]:

1
I( f ) = f (x)eiω sin x d x (5)
0
where f (x) = cos(sin x) cos(x) and g(x) = sin x. The bounded variation of the
above integral is obtained as follows:
1
BV ( f ) ≤ iωg
(x) f (x) + f
(x) d x
0
1
≤ |iω cos(x) cos(sin x) cos(x) − cos(x) sin(sin(x)) cos(x)
0
− cos(sin(x)) sin(x)| d x ≤ ω
The exact solution of the above integral (5) is obtained as
eiω sin 1 1
I (Q) = [iω cos(sin 1) + sin(sin 1)] − iω.
1 + ω2 1 + ω2
The approximate solution of the above integral (5) is calculated using homotopy
perturbation method (HPM) in [10]. It is given as
Fig. 2 Absolute error for N = ω and 0 ≤ ω ≤ 100
Fig. 3 The Absolute error of the integral (5) using HPM and QMC methods
344 N.R. Narni
1
I10 (Q) = iω(457 + 37ω2 + 5ω4 + ω6 + ω8 ) + e0.8414709848iω (968.8821733
ω 10
+169.63521533iω + 37.55809703ω2 + 17.513719298iω3

−2.7632852366ω4 + 2.3069869759iω5 − 1.4293570493ω6

+0.4328909146iω7 − 0.4028624431ω8 − 0.6663667454iω9
and the absolute error between exact and HPM is plotted in Fig. 3.The numerical
solution of the above integral (5) is calculated using quasi-Monte Carlo method and
the corresponding absolute error with respect to the exact integral is plotted in Fig. 3.
We can observe that the numerical solution is same as the exact solution with very
small difference as ω increases. The relative absolute error for the two methods are
calculated. Corresponding to HPM method we have relative error 0.002522606 and
corresponding quasi-Monte Carlo method gives the relative error 0.068065001. The
relative error corresponding to HPM is less compared to quasi-Monte Carlo method
due to the fact that HPM is a semiquantitive method and is applicable to few specific
problems. This shows that the quasi-Monte Carlo methods are reliable and can be
applied to a wide range of problems.
5 Conclusions
In this paper we are able to find the numerical solutions of highly nonlinear oscil-
latory integrals using quasi-Monte Carlo Methods. It is observed that the numerical
solution satisfies the analytical error bounds. This shows the good agreement of
results between analytical and numerical calculations. The work is under progress
for the application of quasi-Monte Carlo methods to higher dimensional problems.
This is due to the fact that as dimension of the problem increases the computational
cost of traditional Gaussian-type methods increases. Therefore, Monte Carlo and
quasi-Monte Carlo methods are the best choices.
References
1. Asheim, S.: A combined Filon/Asymptotic quadrature method for highly oscillatory problems.
BIT Numer. Math. 48, 425–448 (2008)
2. Filon, L.N.G.: On a quadrature formula for trigonometric integrals. Proc. Roy. Soc. Edinburgh
49, 38–47 (1928)
3. Hsiao, G.C., Wendland, W.L.: Boundary Integral Equations. Springer-verlag, Berlin (2008)
4. Iserles, A., Norsett, S.P., Efficient quadrature of highly oscillatory integrals using derivatives,
Proc. Roy. Soc. Edinburgh, A. 461,1383–1399 (2005)
5. Iserles, A., Norsett, S.P., Olver, S., Highly oscillating quadrature: The story so far
6. Jain, M.K., Iyengar, S.R.K., Jain, R.K.: Introduction to Numerical Analysis. Narosa Publishers,
New Delhi (1993)
7. Kuipers, L., Niederreiter, H.: Uniform Distribution of Sequences. John-Wiley & Sons, Newyork
(1974)
8. Kunik, M., Skrzpacz, P.: Diffraction of light revisited. Math. Methods. Appl. Sci. 31(7), 793–
820 (2007)
9. Levin, D.: Procedures for computing one-and-two dimensional integrals of functions with rapid
irregular oscillations. Math. Comp. 38, 531–538 (1982)
10. Molabahrami, A., Khani. F.: Numerical solutions of highly oscillatory integrals. Appl. Math.
Comput. 198, 657–664 (2008)
11. Oseledets, I.V., Stavtsev, S.L., Tyrtyshnikov, E.E.: Integration of oscillating functions in a
quasi-three-dimensional electrodynamic problem. Comput. Math. Math. Phys. 49(2), 292–303
(2009)
12. Walter R.: Real and complex analysis, 3rd edition, McGraw-Hill Company, New Delhi, 1987
13. Wu, X.Y., Xia, J.L.: Two low accuracy methods for stiff systems. Appl. Math. Comput. 123,
141–153 (2001)
14. Xiang, S.: Efficient quadrature for highly oscillatory integrals involving critical points. J. Com-
put. Appl. Math. 206, 688–698 (2006). doi:10.1016/j.cam.2006.08.018
Approximate Controllability of Semilinear
Stochastic System with State Delay
Anurag Shukla, N. Sukavanam and D.N. Pandey
Abstract The objective of this paper is to present some sufficient conditions for
approximate controllability of semilinear stochastic system with state delay. Suf-
ficient conditions are obtained by separating the given semilinear system into two
systems namely a semilinear deterministic system and a linear stochastic system.
To prove our results, the Schauder fixed-point theorem is applied. At the end, an
example is given to show the effectiveness of the result.
Keywords Approximate controllability · State delay · Stochastic system · Schauder

fixed point theorem
1 Introduction
Controllability concepts play a vital role in deterministic control theory. It is well

known that controllability of deterministic equation is widely used in many fields
of science and technology. But in many practical problems such as fluctuating stock
prices or physical system subject to thermal fluctuations, population dynamics, etc.,
some randomness appear, so the system should be modelled stochastic form.
In setting of deterministic systems: Kalman [1] introduced the concept of control-
lability for finite-dimensional deterministic linear control systems. Then Barnett [2]
and Curtain [3] introduced the concepts of deterministic control theory in finite and
infinite-dimensional spaces. Naito [4] established sufficient conditions for approx-
imate controllability of deterministic semilinear control system dominated by the
A. Shukla (B) · N. Sukavanam · D.N. Pandey

e-mail: anuragshukla259@gmail.com
N. Sukavanam
e-mail: nsukavanam@gmail.com
D.N. Pandey

DOI 10.1007/978-81-322-2485-3_28
348 A. Shukla et al.
linear part using Schauder’s fixed-point theorem. In [5, 6], Wang extended the results
of [4] and established sufficient conditions for delayed deterministic semilinear sys-
tems using same Schauder’s fixed-point theorem. In [7] author provided more appli-
cations of Schauder’s fixed-point theorem in nonlinear controllability problems.
In setting of stochastic systems: In [8, 9] Mahmudov established some results
for controllability of linear stochastic systems in finite-dimensional and infinite-
dimensional spaces, respectively. Sukavanam et al. in [10] obtained some sufficient
conditions for s-controllability of an abstract first-order semilinear control system
using Schauder’s fixed-point theorem. Recently, Anurag et al. [11] obtained some
sufficient conditions for approximate controllability of retarded semilinear stochastic
system with nonlocal conditions using Banach fixed-point theorem.
The present paper is generalized form of the system taken in [10]. In this paper
system is taken with finite delay in state which is not discussed up to now in the
literature in best of my knowledge. The technique is adopted similar to discussed in
[10, 12] with suitable modifications.
Let X and U be the Hilbert spaces and Z = L 2 [0, b; X ], Z h = L 2 [−h, b; X ],
0 < h < b, and Y = L 2 [0, b; U ] be function spaces. Rk denotes k-dimensional real
Euclidean space. Let (Ω, ζ, P) be the probability space with a probability measure P
on Ω and a filtration {ζt |t ∈ [0, b]} generated by a Wiener Process {ω(s) : 0 ≤ s ≤ t}.
We consider the semilinear stochastic control system of the form:
d x(t) = [Ax(t) + Bu(t) + f (t, xt )]dt + dω(t), t > 0

x(t) = ξ(t), t ∈ [−h, 0] (1)
where the state function x ∈ Z ; A : D(A) ⊆ X → X is a closed linear operator

which generates a strongly continuous semigroup S(t); B : Y → Z is a bounded
linear operator; function f : [0, b] × X → X is a nonlinear operator such that, f
is measurable with respect to t, for all x ∈ Z and continuous with respect to x for
almost all t ∈ [0, b]; xt ∈ L 2 ([−h, 0], X ) = C (let)-valued stochastic processes
and defined as xt (s) = {x(t + s)| − h ≤ s ≤ 0|}; Control u(t) takes values in U for
each t ∈ [0, b].
By splitting the system (1), we get the following pair of coupled systems
dy(t)
= [Ay(t) + Bv(t) + f (t, (y + z)t ]; 0 ≤ t ≤ b
dt
y(t) = ψ(t), t ∈ [−h, 0] (2)
and
dz(t) = [Az(t) + Bw(t)]dt + dω(t); 0 ≤ t ≤ b

z(t) = ξ(t) − ψ(t), t ∈ [−h, 0] (3)
The system represented by (3) is linear stochastic system and for each realization z(t)
of system (3), the system given by (2) is a deterministic system. Thus the solution y(t)
Approximate Controllability of Semilinear Stochastic System . . . 349
of the semilinear system (2) depends on the solution z(t) of linear stochastic system
(3). The functions v and w are Y -valued control function, such that u = v + w.
It can be easily seen that, the solution x(t) of the semilinear stochastic system (1)
is given by y(t) + z(t) where y(t) and z(t) are the solutions of the systems (2) and
(3), respectively.
2 Preliminaries
In this section, some definitions are discussed which will be used in proof of main
results.
The mild solution of the systems (1) can be written as
⎧ t t
⎪
⎨ S(t)ξ(0) + S(t − s){Bu(s) + f (s, xs )}ds + S(t − s)dω(s), t > 0
x(t) = 0 0
⎪
⎩ ξ(t) −h ≤t ≤0
(4)
the mild solution of the semilinear system (2) can be written as
⎧ t
⎪
⎨ S(t)ψ(0) + S(t − s){Bv(s) + f (s, (y + z)s }ds, t > 0
y(t) = 0 (5)
⎪
⎩ ψ(t) −h ≤t ≤0
and the mild solution of the linear stochastic system (3) can be written as
⎧ t t
⎪
⎨ S(t)(ξ(0) − ψ(0)) + S(t − s)Bw(s)ds + S(t − s)dω(s), t > 0
z(t) = 0 0
⎪
⎩ ξ(t) − ψ(t) −h ≤t ≤0
(6)
Consider the linear system corresponding to the system (2) given by
dp(t)
= Ap(t) + Br (t), t > 0
dt
p(t) = ψ(t) t ∈ [−h, 0] (7)
The mild solution of the above linear system is expressed as

⎧ t
⎪
⎨ S(t)ψ(0) + S(t − s)Br (s)ds t > 0
p(t) = 0 (8)
⎪
⎩ ψ(t) −h ≤t ≤0
Definition 1 The set given by K T ( f ) = {x(T ) ∈ X : x ∈ Z h } where x is a mild

solution of (1) corresponding to control u ∈ Y is called Reachable set of the
system (1).
Definition 2 The system (1) is said to be approximately controllable if K T ( f ) is

dense in X , means K T ( f ) = X .
3 Basic Assumptions
In this section, some basic conditions and lemmas are assumed and discussed for
obtaining the main results. Throughout this paper D(A), R(A), and N0 (A) denote
the domain, range, and null space of operator A, respectively.
The following conditions are assumed:
(H1 ) For every p ∈ Z there exists a q ∈ R(B) such that L p = Lq where the operator
L : Z → X is defined as
b
Lx = S(b − s)x(s)ds
0
(H2 ) The semigroup {S(t), t ≥ 0} generated by A is compact on X and there is a

constant M ≥ 0 such that ||S(t)|| ≤ M.
(H3 ) f (t, x) satisfies Lipschitz continuity on Z . i.e
|| f (t, x1 ) − f (t, x2 )|| ≤ l p ||x1 − x2 ||, l p > 0
(H4 ) f (t, x) satisfies linear growth condition, that is,
|| f (t, x)|| ≤ a1 + b1 ||x||,
where a1 and b1 are constants.

(H5 ) Mbb1 (1 + c) < 1
where the constants b and b1 appear in the above conditions. The constant c is defined
in Lemma 1.
Let G : N0⊥ (L) → R(B) be an operator defined as follows:
Ga = a0
where a ∈ N0⊥ (L) and a0 is the unique minimum norm element in the set

{a + N0 (L)} R(B)} satisfying the following condition

||Ga|| = ||a0 || = min ||e|| : e ∈ {a + N0 (L)} R(B)} (9)
The operator G is well defined, linear, and continuous (see [4], Lemma 1). From
continuity of G, it follows that ||Ga|| ≤ c||a|| Z , for some constant c ≥ 0.
Since Z = N0 (L) + R(B) as is evident from condition (H 1), any element z ∈ Z
can be expressed as
z = n + q : n ∈ N0 (L), q ∈ R(B)
Lemma 1 In [12], for z ∈ Z and n ∈ N0 (L), the following inequality holds
||n|| Z ≤ (1 + c)||z|| Z (10)
where c is such that ||G|| ≤ c.
Let us introduce some operators in the following way:

K : Z → Z defined by
t
(K z)(t) = S(t − s)z(s)ds
0
Now, let M0 be the subspace of Z h (see [11]) such that

m ∈ Z h : m(t) = (K n)(t), n ∈ N0 (L) 0 ≤ t ≤ b
M0 =
m(t) = 0, −h ≤ t ≤ 0
It can be noted that m(b) = 0, for all m ∈ M0 .

For each solution p(t) of the system (7) with control r and for each realization
z(t) of the system (3), define the random operator f p : M0 → M0 as

K n, 0 < t < b
fp = (11)
0, −h ≤ t ≤ 0
where n is given by the unique decomposition
F( p + z + m) = n + q: n ∈ N0 (L), q ∈ R(B), (12)
where F: L 2 ([0, b], C) → X given by
(F x)(t) = f (t, xt (.)); 0≤t ≤b
It is easy to see that F satisfies Lipschitz continuity (H3 ) and linear growth conditions
(H4 ).
4 Main Results
In this section, approximate controllability of systems (2), (3) is proved. System (1)
is splitted in systems (2), (3), so if systems (2), (3) are approximately controllable
then system (1) is also approximately controllable.
The linear system (7) corresponding to system (2) is approximately controllable

under the condition (H1 ) (see [5]).
For approximate controllability of (3)
dz(t) = [Az(t) + Bw(t)]dt + dω(t); 0 ≤ t ≤ b

z(t) = ξ(t) − ψ(t), t ∈ [−h, 0] (13)
The mild solution of above system is

⎧ t t
⎨ S(t)(ξ(0) − ψ(0)) + S(t − s)Bw(s)ds + S(t − s)dω(s), t > 0
z(t) = 0 0
⎩
ξ(t) − ψ(t) −h ≤t ≤0
(14)
Define the operator L b0 : L 2 [0, b; U ] → L 2 [Ω, ζt , X ], the controllability operator
Πsb : L 2 [Ω, ζt , X ] → L 2 [Ω, ζt , X ] associated with (14), and the controllability
operator Γsb : X → X associated with the corresponding deterministic system of
(14) as
b
L b0 = S(b − s)Bw(s)ds (15)
0
b
Πsb {.} = S(b − t)B B ∗ S ∗ (b − t)E{.|ζt }dt (16)
s
b
Γsb = S(b − t)B B ∗ S ∗ (b − t)dt (17)
s
It is easy to see that the operators L b0 , Πsb , Γsb are linear-bounded operators, and the
adjoint (L b0 )∗ : L 2 [Ω, ζt , X ] → L 2 [0, b; U ] of L b0 is defined by
(L b0 )∗ = B ∗ S ∗ (b − t)E{z|ζt }Π0b = L b0 (L b0 )∗ .
Before studying the approximate controllability of system (3), let us first investigate
the relation between Πsb and Γsb ; s ≤ r < b and resolvent operator R(λ, Πsb ) =
(λI + Πsb )−1 and R(λ, Γrb ) = (λI + Γrb )−1 , s ≤ r < b for λ > 0, respectively.
ζ
Lemma 2 For every z ∈ L 2 [Ω, ζt , X ] there exists ϕ(.) ∈ L 2 (0, b; L(Rk , X )) such
that
t
1. E{z|ζt } = E{z} + 0 ϕ(s)dω(s),
b
2. Πsb z = Γsb Ez + s Γrb ϕ(r )dω(r ),
b
3. R(λ, Πsb )z = R(λ, Γsb )E{z|ζt } + s Γrb ϕ(r )dω(r ).
Proof The proof is straightforward adaption of the proof of [10, Lemma 2.3].

Theorem 1 The control system (3) is approximately controllable on [0, b] if and
only if one of the following conditions holds.
1. Π0b > 0.
2. λR(λ, Π0b ) converges to the zero operator as λ → 0+ in the strong operator
topology.
3. λR(λ, Π0b ) converges to the zero operator as λ → 0+ in the weak operator
topology.
Proof The proof is straightforward adaption of the proof of [9, Theorem 2].
Lemma 3 Under the conditions (H2 ), (H4 ), and (H5 ), the operator f p has a fixed
point m 0 ∈ M0 for each realization z(t) of the system (3).
Proof From the compactness of S(t) the integral operator K is compact and hence
f p is compact for each p, (see [1]). Now let ||m|| ≤ r̃ . Then from the condition (H4 )
and from the inequality (10) and (12), we have
t 2

|| f p (m)|| 2
≤ S(t − s)n(s)ds
0
b t 2

≤ S(t − s)n(s)ds dt

0 0
≤ M b (1 + c)2 ||F( p + z + m||2Z
2 2
≤ M 2 b2 (1 + c)2 {a1 + b1 || p + z + m|| Z }2

≤ M 2 b2 (1 + c)2 {a1 + b1 || p + z|| + b1r̃ }2 (18)
Using Schauder’s fixed-point theorem, it is clear from the compactness of f p and

(18) that f p has a fixed point in M0 in a ball of radius r̃ > 0, if
Mb(1 + c)(a1 + b1 || p + z||)

r̃ >
1 − Mb(1 + c)b1
Thus f p (m 0 ) = m 0
The approximate controllability of the semilinear system (2) is proved in following

manner using the above lemma.
Lemma 4 For each realization z(t) of the system (3), the semilinear control system
(2) is approximately controllable under the conditions (H1 )–(H4 ).
Proof From the Eq. (12), we have
F( p + z + m) = n + q
Operating K on both the sides at m = m 0 (fixed point of f p ) and using (11), we get
KF( p + z + m 0 ) = Kn + Kq
= m 0 + Kq
Adding p on both sides, we get
p + KF( p + z + m 0 ) = p + m 0 + Kq
Let p + m 0 = y ∗ , then the above equation is equivalent to
p + K F(y ∗ + z) = y ∗ + K q
Since, from the Eq. (8)

p = S(t)ψ(0) + KBr
we have
S(t)ψ(0) + K Br + K F(y ∗ + z) = y ∗ + K q
S(t)ψ(0) + K (Br − q) + K F(y ∗ + z) = y ∗
Thus, it follows that y ∗ (t) is a solution of the semilinear system
dy ∗ (t)
= Ay ∗ (t) + f (t, (y ∗ + z)t ) + Br (t) − q(t),
dt
y ∗ (0) = ψ(0) (19)
with control (Br − q).

Moreover, since y ∗ (t) = p(t) + m 0 (t), it follows that
y ∗ (b) = p(b) + m 0 (b),
as m 0 (b) = 0 it follows that

y ∗ (b) = p(b) (20)
From the Eqs. (19) and (20), it is clear that the reachable set of (19) is a superset of
the reachable set of the system (7), which is dense in X .
Further q ∈ R(B) implies that for any given ε1 > 0, there exists v1 ∈ Y such that
||q − Bv1 || ≤ ε1 .
Now consider the equation
dy(t)
= Ay(t) + f (t, (y + z)t ) + B(r (t) − v1 (t)),
dt
y(0) = ψ(0) (21)
Let y(t) be the solution of the system (21), corresponding to control v = r −v1 . Then
||y ∗ (b) − y(b)|| can be made arbitrary small by choosing a suitable v1 , which implies
that the reachable set of the system (21) is dense in the reachable set of the system
(19), which in turn is dense in X . This proves that the system (2) is approximately
controllable.
5 Example
Consider the stochastic control system with delay governed by the semilinear heat
equation

∂ 2 y(t, x)
∂ y(t, x) = + Bu(t, x) + f (t, y(t + v, x)) ∂t + ∂ω(t)
∂x2
for 0 < t < τ ; v ∈ [−h, 0]; 0 < x < π
with conditions y(t, 0) = y(t, π ) = 0, 0 ≤ t ≤ τ
y(t, x) = ξ(t, x), −h ≤ t ≤ 0, 0 ≤ x ≤ π (22)
The system (22) can be written in the abstract form (1),

2
by2
setting X = L 2 (0, π )
and A = ddx 2 , with domain consisting of all y ∈ X with dd x y2 ∈ X and y(0) = 0 =
y(π ). Take φ( x) = (2/π )1/2 sin(nx), 0 ≤ x ≤ π, n = 1, 2, 3, ..., then {φn (x)} is an
orthonormal basis for X and φn ia an eigenfunction corresponding to the eigenvalue
λn = −n 2 of the operator A, n = 1, 2, 3, .... Then the C0 -semigroup T (t) generated
by A has eλn t as the eigenvalues and φn as their corresponding eigenfunctions.
Define an infinite-dimensional space U by
∞
∞

U = u:u= u n φn with u 2n < ∞
n=2 n=2
The norm defined by

∞ 1/2
||u||U = u 2n
n=2
ξ(t, x) is known function.

Let B be a continuous linear operator from U to X defined as
∞
∞

Bu = 2u 2 φ1 + u n φn , u = u n φn ∈ U
n=2 n=2
The nonlinear operator f is assumed to satisfy conditions (H3 ) and (H4 ).

The approximate controllability of the corresponding semilinear deterministic
heat equation of (22) was considered by Naito [4] and proved under the conditions
(H1 )–(H4 ). Here approximate controllability of the stochastic semilinear heat control
system (22) is considered.
The system (22) can be associated with two control systems under the initial and
boundary conditions, as given below
∂ y(t, x) ∂ 2 y(t, x)
= + y(t − h, x)
∂t ∂x2
+Bv(t, x) + f (t, y(t − h, x) + z(t − h, x)) t ∈ [0, b] x ∈ [0, π ] (23)
y(t, x) = ξ(t, x), −h ≤ t ≤ 0, 0 ≤ x ≤ π

2
∂ z(t, x)
∂z(t, x) = + z(t − h, x) + Bw(t) ∂t + ∂ω(t) (24)
∂x2
The system (24) is a linear stochastic system and for each realization z(t) of the
system (24), the system (23) is a deterministic system.
From Lemma 4 and using the conditions (H1 )–(H4 ), it is clear that for each real-
ization z(t) of the system (24), the system (23) is approximately controllable. The
linear stochastic system (24) is approximately controllable from Lemma 3 corre-
sponding to (23) and linear system corresponding to system (23) is approximately
controllable from [4].
suggestions for improving the paper. This research is supported by the Council of Scientific and
Industrial Research (CSIR) in the form of fellowship to the first author.
References
1. Kalman, R.E.: Controllability of linear systems. Contribut. Differ. Equ. 1, 190–213 (1963)
2. Barnett, S.: Introduction to Mathematical Control Theory. Clarendon Press, Oxford (1975)
3. Curtain, R.F., Zwart, H.J.: An Introduction to Infinite Dimensional Linear Systems Theory.
Springer, New York (1995)
4. Naito, K.: Controllability of semilinear control systems dominated by the linear part. SIAM J.
Control Optim. 25, 715–722 (1987)
5. Wang, L.W.: Approximate controllability of integrodifferential equations with multiple delays.
J. Optim. Theory Appl. 143, 185–206 (2009)
6. Wang, L.W.: Approximate controllability of delayed semilinear control of control system. J.
Appl. Math. Stoch. Anal. 1, 67–76 (2005)
7. Klamka, J.: Schuder fixed point theorem in nonlinear controllability problems. Control Cyber-
net. 29(1), 153–165 (2000)
8. Mahmudov, N.I., Denker, A.: On controllability of linear stochastic systems. Int. J. Control
73(2), 144–151 (2000)
9. Mahmudov, N.I.: Controllability of linear stochastic systems in hilbert spaces. J. Math. Anal.
Appl. 259(1), 64–82 (2001)
10. Sukavanam, N., Kumar, M.: S-controllability of an abstract first order semilinear control sys-
tem. Numer. Funct. Anal. Optim. 31(7–9), 1023–1034 (2010)
11. Shukla, A., Arora, U., Sukavanam, N.: Approximate controllability of retarded semilinear
stochastic system with non local conditions. J. Appl. Math. Comput. doi:10.1007/s12190-014-
0851-9
12. Sukavanam, N.: Approximate controllability of semilinear control systems with growing
nonlinearit. In: Mathematical Theory of Control Proceedings of International Conference,
pp. 353–357. Marcel Dekker, New York (1993)
Fourth-Order Derivative-Free Optimal
Families of King’s and Ostrowski’s Methods
Ramandeep Behl, S.S. Motsa, Munish Kansal and V. Kanwar
Abstract In this paper, we present several new fourth-order optimal schemes that
do not require any derivative evaluation for solving nonlinear equations, numerically.
The presented approach of deriving these families is based on approximating deriva-
tives by finite difference and weight function approach. The fourth-order derivative-
free optimal families of King’s and Ostrowski’s methods are the main findings of the
present work. Further, we have also shown that the families of fourth-order meth-
ods proposed by Petković et al., Appl Math Comput 217:1887–1895 (2010) [12]
and Kung-Traub, J ACM 21:643–651 (1974) [8] are special cases of our proposed
schemes. The proposed methods are compared with their closest competitors in a
series of numerical experiments. All the methods considered here are found to be
more effective to similar robust methods available in the literature.
Keywords Nonlinear equation · Newton’s method · Order of convergence ·

Ostrowski’s method · Simple root
1 Introduction
Multipoint iterative methods for solving nonlinear equations play a significant role
in the field of iterative processes since they circumvent the drawbacks of one-point
R. Behl (B) · S.S. Motsa

School of Mathematics, Statistics and Computer Science,
University of KwaZulu-Natal, Private Bag X01, Pietermaritzburg 3209, South Africa
e-mail: ramanbehl87@yahoo.in
S.S. Motsa
e-mail: sandilemotsa@gmail.com
M. Kansal · V. Kanwar
University Institute of Engineering and Technology, Panjab University,
Chandigarh 160 014, India
e-mail: mkmaths@gmail.com
V. Kanwar
e-mail: vmithil@yahoo.co.in

DOI 10.1007/978-81-322-2485-3_29
360 R. Behl et al.
iterations, such as Newton’s method. Such constructions occasionally possess a better

order of convergence and efficiency index for finding the simple or multiple roots. In
past and recent years, many fourth-order optimal multipoint iterative methods have
been proposed and analyzed by King [7], Kanwar et al. [6], Behl et al. [1] and for
detailed explanation of such methods, one can refer the excellent text books written
by Traub [14], Ostrowski [9], and Petković et al. [11].
King’s family [7], Jarratt’s method [1], and Ostrowski’s method [1, 6, 9] are one
of the most efficient fourth-order multipoint methods known to date. However, all
these multipoint methods require the evaluation of first-order derivative of function
at each step. But, there are many practical situations in which the calculations of
derivatives are expensive and/or it requires a great deal of time for them to be given
or calculated. Therefore, derivative-free family of King’s methods and Ostrowski’s
method are still needed.
In the last few years, many derivative-free fourth-order multipoint methods were
proposed and analyzed by Petković et al. [12], Cordero et al. [4], Soleymani et al.
[13], Cordero et al. [2, 3, 5], Zheng et al. [15], Peng et al. [10], and references
cited therein. Some attempts have been made by Cordero et al. [4] to develop an
optimal family of Ostrowski’s method free from derivative. They obtained a fourth-
and sixth-order multipoint family of Ostrowski’s method in which the derivative is
not required. But these proposed methods are not optimal in the sense of Kung-Traub
conjecture.
Therefore, the construction of an optimal derivative-free family of King’s method
or Ostrowski’s method having biquadratic convergence is an open and challeng-
ing problem in computational mathematics. With this aim, we proposed an optimal
scheme of King’s family in which there is no need to find the derivatives of functions.
All the proposed methods considered here are found to be effective and comparable
to the classical Ostrowski’s method, King’s method, and recently developed robust
methods, respectively.
2 Construction of Novel Techniques Without Memory
In this section, we intend to develop derivative-free optimal families of King’s method

and Ostrowski’s method. For this purpose, we consider the following scheme:
f (xn )
yn = xn − ,
f (xn )
(1)
f (yn )
xn+1 = yn − .
f (yn )
This method has a fourth-order of convergence. But, the scheme (1) has two major
drawbacks as it requires four functional evaluations per iteration and first-order
derivative is computed at every iteration. Therefore, we consider here some suit-
able approximations of f (xn ) by forward approximation and f (yn ) by similar to
Fourth-Order Derivative-Free Optimal Families of King’s and Ostrowski’s Methods 361
King’s approximation [7], which are given by
f (xn ) ≈ f [xn , u n ], where u n = xn + α f (xn ), and α ∈ R\{0},

and
(2)
f (xn ) + a f (yn )
f (yn ) = f [xn , u n ] , where a, β ∈ R,
f (xn ) + β f (yn )
respectively.
Using the above approximations of derivatives f (xn ) and f (yn ) in (1), we get
⎧
⎪ f (xn )
⎪
⎨ yn = xn − , u n = xn + α f (xn ), where α ∈ R\{0},
f [xn , u n ]
(3)
⎪
⎪ f (yn ) f (xn ) + β f (yn )
⎩ xn+1 = yn − , where a, β ∈ R.
f [xn , u n ] f (xn ) + a f (yn )
This is a modified family of King’s method and satisfies the following error equation:

a − β + 2 + α(a − β + 1)c1
en+1 = (1 + αc1 )c23 en3 + O(en4 ), (4)
c13
(n)
where en = xn − r, cn = f n!(r ) , n = 1, 2, 3, . . .
But, according to the Kung-Traub conjecture [8], the above scheme is not optimal.
Therefore, we shall now make use of weight function approach to build our optimal
scheme based on (3) by a simple change in its second step. Therefore, we consider
⎧
⎪ f (xn )
⎪
f [xn , u n ]

(5)
⎪
⎪ f (yn ) f (xn ) + β f (yn ) f (yn )
⎩ xn+1 = yn − Q ,
f [xn , u n ] f (xn ) + γ f (yn ) f (u n )

(yn )
where γ and β are two free disposable parameters and Q ff (u n ) ∈ R is a real-
valued weight function. Theorem 1 indicates that under what conditions on the weight
function and disposable parameters in (5), the order of convergence will reach at the
optimal level four.
3 Convergence Analysis
Theorem 1 Let f : I ⊆ R → R be sufficiently differentiable function defined on an

open interval I , enclosing a simple zero of f (x) (say x = r ∈ I ). Assume that initial
guess x = x0 is sufficiently close to r ∈ I . Then, our proposed scheme (5) free from
362 R. Behl et al.
derivatives, has an optimal fourth-order convergence when
γ = β − 1, Q(0) = 1, Q (0) = 1. (6)
It satisfies the following error equation

β + 3 + 2α(β − 1)c1 + α 2 βc12
en+1 = (1 + αc1 )c23 en4 + O(en5 ), (7)
c13
where en and cn are already defined in Eq. (4).
Proof Let x = r be a simple zero of f (x). Expanding f (xn ) and f (xn ) about x = r
by the Taylor’s series expansion, we have
f (xn ) = (c1 en + c2 en2 ) + O(en3 ), (8)
and
f (xn +α f (xn )) = (1+αc1 )c1 en +(1+3αc1 +α 2 c12 )c2 en2 +2α(αc1 +1)c22 en3 +O(en4 ),
(9)
respectively.
From Eqs. (8) and (9), we have

α f 2 (xn ) 1 2 + 2αc 1 + α 2 c2
= en − α + c2 en2 + 1
c22 en3
f (xn + α f (xn )) − f (xn ) c1 c12

4 + 5αc1 + 3α 2 c12 + α 3 c13 3 4
− c2 en + O(en5 ),
c13
(10)
and

α f 2 (xn )
f (yn ) = f xn − ,
f (xn + α f (xn )) − f (xn )

1 + αc1 2 + 2αc1 + 2α 2 c12 5 + 7αc1 + 4α 2 c12 + α 3 c13
= c2 en2 − c22 en3 + c23 en4 + O(en5 ).
c1 c1 c12
(11)
Furthermore, we have

α f (xn ) f (yn )( f (xn ) + β f (yn )) 1 + αc1
= c2 en2
( f (u n ) − f (x))( f (xn ) + γ f (yn )) c1

4 + γ − β + α(5 + 2γ − 2β)c1 + α 2 (γ + 2 − β)c12
− c22 en3 + O(en4 ),
c12
(12)
and

f (yn ) c2 2αc1 + 3 2 2 8 + 8αc1 + 3α 2 c12 3 3
= en − c2 en + c2 en + O(en4 ).
f (u n ) c1 c12 c13
(13)

(yn )
Since, it is clear from (13) that ff (u n)
is of order en . Hence, we can consider the
Taylor’s expansion of the weight function Q in the neighborhood of zero. Therefore,
we have

f (yn ) f (yn )
Q = Q(0) + Q (0) + O(en3 ). (14)
f (u n ) f (u n )
Using (11), (12), (13) and (14) in scheme (5), we have the following error equation:

f (xn ) f (yn ) f (xn ) + β f (yn ) f (yn )
en+1 = en − − Q ,
f [xn , u n ] f [xn , u n ] f (xn ) + γ f (yn ) f (u n )

αc1 + 1 1
= −(Q(0) − 1) c2 en2 + 2 − 2 − Q (0) + Q(0)(γ − β + 4) − α(2 + Q (0)
c1 c1

1
+ Q(0)(2β − 2γ − 5))c1 + α 2 (−1 + Q (0)(γ − β + 2))c12 c22 en3 − 3 − 4 − Q (0)(γ + 7)
c1
+ Q(0)(13 + γ 2 − γ (β − 7) − 7β) + β Q (0) + α(−5 + Q(0)(20 + 3γ 2 − 3γ (β − 5) − 15β)
− 2Q (0)(5 + γ − β))c1 + α 2 (−3 + Q(0)(12 + 3γ 2 + γ (11 − 3β) − 11β) + Q (0)(−4 − γ + β))c12

+ α 3 (−1 + Q(0)(3 + γ 2 − γ (β − 3) − 3β))c13 c23 en4 + O(en5 ).
(15)
For obtaining an optimal general class of fourth-order methods, the coefficients of

en2 and en3 in the error Eq. (15) must be zero simultaneously. After simplifying the
Eq. (15), we have the following equations involving γ , Q(0) and Q (0)
⎧
⎪ Q(0) = 1,
⎪
⎪
⎨ − Q (0) + Q(0)(γ − β + 4) = 2,
(16)
⎪
⎪ (2 + Q (0) + Q(0)(2β − 2γ − 5)) = 0,
⎪
⎩
(−1 + Q (0)(γ − β + 2)) = 0,
respectively.
After simplifying the Eq. (16), we get the following values of γ , Q(0) and Q (0)
γ = β − 1, Q(0) = 1, Q (0) = 1. (17)
Using the above conditions in scheme (5), we shall get the following error
equation:
364 R. Behl et al.

β + 3 + 2α(β − 1)c1 + α 2 βc12
en+1 = (1 + αc1 )c23 en4 + O(en5 ), (18)
c13
where α, β ∈ R are two free disposable parameters.

This completes the proof of the Theorem 1.
Remark 1 From computational point of view, the family of methods (5) which is
totally derivative-free reaches the highest possible convergence and efficiency index
using only three functional evaluations viz., f (xn ), f (u n ), f (yn ) per full iteration.
From the application point of view, methods in which there are derivative evaluations
per full cycle are restricted, when the problem considered a bear massive time or
load for computing the √ derivatives. For example, the nonlinear function h(x) =
tan(ln x) + cos(x 4 ) × (1/(2x) , has a first derivative that is hard to write. In fact in
such cases, the derivative evaluation is expensive and/or occasionally takes a great
deal of time. Such shortcomings lead us to study optimal iterative methods which
are totally derivative-free per full iteration.
4 Special Cases
In this section, we shall consider some particular cases of our proposed scheme (5)
and mention some weight functions Q(x) that satisfy all the conditions of Theorem 1,
as follow:
Case 1 Let us consider the following weight function:
Q(x) = x + 1. (19)
Using the above weight function in scheme (5), we obtain

⎧
⎪ f (xn )
⎪
f [xn , u n ]
(20)
⎪
⎪ α f (xn ) f (yn )( f (u n ) + f (yn ))( f (xn ) + β f (yn ))
⎩ xn+1 = yn − ,
f (u n )( f (u n ) − f (xn ))( f (xn ) + (β − 1) f (yn ))
where β ∈ R.
This is a new modified derivative-free optimal general class of fourth-order King’s
method and one can easily get many new families of methods by choosing different
values of the disposable parameters α and β.
Subspecial Cases of Optimal Family (20)
(i) For β = 21 , family (20) reads as
⎧
⎪ f (xn )
⎪
⎨ yn = xn − , u n = xn + α f (xn ),
f [xn , u n ]
(21)
⎪
⎪ α f (xn ) f (yn )( f (u n + f (yn ))(2 f (xn ) + f (yn ))
⎩ xn+1 = yn − .
f (u n )( f (u n ) − f (xn ))(2 f (xn ) − f (yn ))
This is a new modified derivative-free optimal family of King’s method.

(ii) For β = 0, family (20) reads as
⎧ f (xn )
⎪
⎪ yn = xn − , u n = xn + α f (xn ),
⎨ f [xn , u n ]
(22)
⎪
⎪ α f 2 (xn ) f (yn )( f (u n + f (yn ))
⎩ xn+1 = yn − .
f (u n )( f (u n ) − f (xn ))( f (xn ) − f (yn ))
This is a new modified derivative-free optimal family of Ostrowski’s method.

Case 2 Let us consider the following weight function:
1
Q(x) = . (23)
1−x
Using the above weight function in scheme (5), we obtain

⎧
⎪ f (xn )
⎪
f [xn , u n ]
⎪
⎪ α f (xn ) f (u n ) f (yn )( f (xn ) + β f (yn ))
⎩ xn+1 = yn − ,
( f (xn ) − f (u n ))( f (u n ) − f (yn ))( f (xn ) + (β − 1) f (yn ))
(24)
where β ∈ R.
This is another new modified general class of fourth-order optimal King’s method.
Subspecial Cases of Optimal Family (24)
(i) For β = 1, family (24) reads as
⎧
⎪ f (xn )
⎪
⎨ yn = xn − , u n = xn + α f (xn ),
f [xn , u n ]
(25)
⎪
⎪ α f (u n ) f (yn )( f (xn ) + f (yn ))
⎩ xn+1 = yn − .
( f (xn ) − f (u n ))( f (u n ) − f (yn ))
This is a fourth-order optimal family of derivative-free methods independently

derived by Petković et al. [12].
(ii) For β = 0, family (24) reads as
⎧ f (xn )
⎪
⎪ yn = xn − , u n = xn + α f (xn ),
⎨ f [xn , u n ]
(26)
⎪
⎪ α f 2 (xn ) f (u n ) f (yn )
⎩ xn+1 = yn − .
( f (xn ) − f (u n ))( f (u n ) − f (yn ))( f (xn ) − f (yn ))
366 R. Behl et al.
This is another fourth-order optimal family of multipoint methods independently

derived by Kung and Traub [8].
Remark 2 The first most striking feature of this contribution is that we have devel-
oped optimal derivative-free families of King’s method and Ostrowski’s method for
the first time which will converge even though the guess is far from zero or the deriv-
ative is small in the vicinity of the required root. Therefore, these techniques can be
used as an alternative to King’s and Ostrowski’s techniques or in the cases where
King’s and Ostrowski’s techniques are not successful.
In this section, we shall check the effectiveness of new optimal methods. We employ
the present methods, namely (21), method (22), and family (20) for
β= 1
100 and β = 101
100 denoted by MKM 14 , MOM 4 , MKM 24 , and MKM 34 , respec-
tively, with |β = 1| to solve nonlinear equations given in Table 1. We compare them
with existing King’s method for β = 21 (KM 4 ), Ostrowski’s method (OM 4 ), Cordero
et al. method (4) (CM 4 ) [4], Petković et al. method (12) (PM 4 ) [12], and Kung-Traub
method (KT 4 ) [8], respectively. For better comparison of our proposed methods, we
have given three comparison tables in each example: one is corresponding to absolute
error value of given nonlinear functions (with the same total number of functional
evaluations = 12), the second is with respect to the number of iterations taken by each
method to obtain the root correct up to 35 significant digits, and the last one is corre-
sponding to computational order of convergence in Table 2, 3 and 4, respectively. All
computations have been performed using the programming package Mathematica
9 with multiple precision arithmetic. We use ε = 10−34 as a tolerance error. The
following stopping criteria are used for computer programs:
(i) |xn+1 − xn | < ε and (ii) | f (xn+1 )| < ε.
Table 1 Test functions

f (x) r [a, b]
f 1 (x) = sin x 0.0000000000000000000000000000000000 [−1.51, 1.51]
f 2 (x) = e−x + sin x 3.1830630119333635919391869956363946 [1.3, 4]
f 3 (x) = sin2 x − x 2 + 1 1.4044916482153412260350868177868681 [0.5, 1.8]
f 4 (x) = tan−1 x 0.0000000000000000000000000000000000 [−1, 1]
f 5 (x) = (x − 1)3 − 1 2.0000000000000000000000000000000000 [1.5, 3]
f 6 (x) = x 3 − cos x + 2 −1.1725779647539700126733327148688486 [−3, 0]
f 7 (x) = √tan(log x) + 0.443260783556767073513472596321 [0.38, 0.50]
cos(x 3 ) × 1/2x
Table 2 Comparison of different fourth-order methods with the same total number of functional evaluation (TNFE = 12)
f (x) x0 K M4 M K M41 O M4 M O M4 C M4 P M4 M K M42 K T4 M K M43
1. −1.51 C 1.1.e−552 C 4.9e−557 C 1.4e−801 3.6e−800 4.9e−557 4.0e−917
−1.31 C 5.4e−1118 C 7.4e−1126 C 5.9e−1397 2.3e−1396 7.4e−1126 7.3e−1452
1.31 C 5.4e−1118 C 7.4e−1126 C 5.9e−1397 2.3e−1396 7.4e−1126 7.3e−1452
1.51 C 1.1.e−552 C 4.9e−557 C 1.4e−801 3.6e−800 4.9e−557 4.0e−917
2. 1.5 C 2.1e−184 C 5.9e−183 C 4.0e−174 4.0e−174 5.9e−183 3.2e−180
2.5 5.3e−182 3.6e−530 6.8e−187 3.7e−530 8.8e−153 1.1e−586 1.1e−586 3.7e−530 1.1e−586
3.5 1.4e−274 3.5e−568 6.8e−278 3.5e−568 4.0e−245 9.2e−598 9.2e−598 3.5e−568 9.6e−598
4 1.2e−128 1.1e−317 6.4e−140 9.0e−318 4.1e−127 2.5e−344 2.5e−344 9.0e−318 5.8e−345
3. 0.5 2.6e+0 1.8e−73 1.6e−16 1.9e−64 9.4e−130 2.1e−130 7.5e−131 1.9e−64 6.9e−107
1.2 1.7e−160 1.4e−220 3.4e−197 4.5e−184 3.9e−256 1.5e−224 1.1e−225 4.5e−184 3.7e−169
1.8 3.6e−141 9.5e−117 1.1e−159 6.2e−77 1.0e−195 1.0e−56 4.3e−51 6.2e−77 5.3e−78
4. −1 1.4e−102 5.7e−1485 1.5e−134 1.2e−1490 7.0e−147 8.1e−1773 1.8e−1772 1.2e−1490 1.2e−1804
1 1.4e−102 5.7e−1485 1.5e−134 1.2e−1490 7.0e−147 88.1e−1773 1.8e−1772 1.2e−1490 1.2e−1802
5. 1.5 3.7e−10 9.5e−125 9.7e−60 1.2e−61 8.7e−93 1.6e−136 1.5e−137 1.2e−61 1.6e−48
2.5 1.4e−100 2.1e−43 1.2e−122 3.6e−35 D 5.6e−83 3.9e−81 1.4e−31 2.7e−34
6. −1.3 5.9e−219 6.0e−231 1.2e−248 5.1e−146 1.4e−174 9.4e−184 4.4e−181 5.1e−146 2.5e−126
−1 6.7e−153 1.3e−225 1.4e−194 1.8e−142 2.6e−173 8.6e−183 7.7e−181 1.8e−142 1.4e−134
0 F 8.1e+0 F 6.2e+1 D 7.2e+0 7.1e+0 6.2e+1 6.4e−6
Fourth-Order Derivative-Free Optimal Families of King’s and Ostrowski’s Methods
7. 0.44 3.8e−191 1.5e−130 0.1e−343 1.6e−345 C 2.3e−101 6.4e−105 1.6e−142 2.1e−100

0.45 9.0e−396 1.0e−342 1.6e−560 1.0e−346 9.2e−317 1.92e−312 1.0e−330 1.0e−346 6.8e−312
C Stands for converge to undesired root, D Stands for divergent, F Stands for failure
367
368
Table 3 Comparison of different fourth-order methods with respect to number of iterations

1. −1.51 C 4 C 4 C 4 4 4 4
−1.31 C 4 C 4 C 4 4 4 4
1.31 C 4 C 4 C 4 4 4 4
1.51 C 4 C 4 C 4 4 4 4
2. 1.5 C 4 C 4 C 4 4 4 4
2.5 4 4 4 4 4 3 3 4 3
3.5 4 3 4 3 4 3 3 3 3
4 5 4 4 4 5 4 4 4 4
3. 0.5 9 5 6 5 5 5 5 5 5
1.2 4 4 4 4 4 4 4 4 4
1.8 4 5 4 5 4 5 5 5 5
4. −1 5 4 5 4 5 4 4 4 4
1 5 4 5 4 5 4 4 4 4
5. 1.5 6 5 5 5 5 4 4 5 5
2.5 5 6 5 5 D 5 5 6 5
6. −1.3 4 4 4 4 4 4 4 4 5
−1 4 4 4 4 4 4 4 4 4
0 F 15 F 10 D 11 13 10 7
7. 0.40 4 4 4 4 C 5 4 4 4
0.45 4 4 4 4 4 4 4 4 4
C Stands for converge to undesired root, D Stands for divergent, F Stands for failure
R. Behl et al.
Table 4 Computational order of convergence of different fourth-order methods
1. −1.51 na 5.0000 na 5.0000 na 4.0009 4.0009 4.0067 4.0004
−1.31 na 4.0001 na 4.0001 na 4.0000 4.0000 4.0001 4.0000
1.31 na 4.0001 na 4.0001 na 4.0000 4.0000 4.0001 4.0000
1.5 na 5.0000 na 5.0000 na 4.0009 4.0009 9.0067 4.0004
2. 1.5 na 3.8024 na 3.8147 na 4.0026 4.0026 3.8147 4.0029
2.5 4.0031 4.0000 4.0025 4.0000 4.0000 4.0000 4.0000 4.0000 4.0000
3.5 3.9999 4.0000 3.9999 4.0000 3.9997 4.0000 4.0000 4.0000 4.0000
4 4.0000 4.0000 4.0000 4.0000 4.0000 4.0000 4.0000 4.0000 4.0000
3. 0.5 4.0000 4.0000 4.0000 4.0000 4.0000 4.0000 4.0000 4.0000 4.0000
1.2 3.9994 4.0001 3.9999 4.0002 4.0025 4.0001 4.0001 4.0002 4.0003
1.8 3.9985 4.0000 3.9995 4.0000 3.8975 4.0000 4.0001 4.0000 4.0000
4. −1 5.0000 4.0000 5.0000 4.0000 5.0000 4.0000 4.0000 4.0000 4.0000
1 5.0000 4.0000 5.0000 4.0000 5.0000 4.0000 4.0000 4.0000 4.0000
5. 1.5 4.0000 4.0000 4.0000 4.0000 4.0000 4.0016 4.0016 4.0000 4.0001
2.5 4.0000 4.0000 4.0000 3.9998 nd 4.0000 4.0000 4.0000 4.0013
6. −1.3 4.0000 4.0000 4.0000 4.0008 3.9997 3.9998 3.9998 4.0007 4.0000
−1 3.9992 4.0001 3.9999 4.0007 3.9997 3.9998 3.9998 4.0008 4.0014
0 * 4.0000 * 3.9997 nd 4.0000 4.0000 3.9997 4.0000
Fourth-Order Derivative-Free Optimal Families of King’s and Ostrowski’s Methods
7. 0.40 4.000 3.9999 4.0000 4.0001 na 3.9999 4.0333 4.0001 4.0000

0.45 4.000 4.0000 4.0000 4.0001 4.000 4.0000 4.0000 4.0000 4.0000
na Not applicable in the case of converge to undesired root, nd Not defined in the case of divergent, * In the case of failure
369
370 R. Behl et al.
Example 1 sin x = 0.
This equation has an infinite number of roots. It can be seen that King’s method
(for β = 21 ), Ostrowski’s method, and Cordero et al. method do not necessar-
ily converge to the root that is nearest to the starting value. For example, King’s
method (for β = 21 ), Ostrowski’s method, and Cordero et al. method with initial
guess x0 = −1.51 converge to 6.2831 . . . , 9.4247 . . . , and −25.1327 . . . , respec-
tively, far away from the required root zero. Similarly, King’s method (for β = 21 ),
Ostrowski’s method, and Cordero et al. method with initial guess x0 = −1.31 con-
verge to 28.2743 . . . , −6.2831 . . . , and 3.1415 . . . , respectively, and so on. Our
methods do not exhibit this type of behavior.
Example 2 e−x + sin x = 0.

Again, this equation has an infinite number of roots. It can be seen that King’s method
(for β = 21 ), Ostrowski’s method, and Cordero et al. method do not necessarily con-
verge to the root that is nearest to the starting value. For example, King’s method (for
β = 21 ) and Ostrowski’s method with initial guess x0 = 1.5 converge to 9.4248 . . . ,
while Cordero et al. method converges to 25.1327 . . . far away from the required root
3.1830 . . . . Similarly, King’s method (for β = 21 ), Ostrowski’s method, and Cordero
et al. method with initial guess x0 = 1.6 converge to 12.566 . . . , 18.8495 . . . , and
6.2813 . . . , respectively, and so on. Our methods do not exhibit this type of behav-
ior. Similarly, we can check the behavior of all the methods on the text problems
mentioned in Table 1.
6 Conclusions
In this paper, we have proposed several new modified derivative-free families of

King’s method, Ostrowski’s method for solving nonlinear scalar equations. The main
advantage of these methods is that they do not use the evaluation of any derivatives but
an optimal order of convergence is nonetheless maintained. All the proposed families
of methods require three functional evaluations, viz. f (xn ), f (u n ), and f (yn ). In
order to obtain an assessment of the efficiency index of our proposed families of
methods, we shall make use of the efficiency index√[14]. For our proposed iteration
schemes, we find p = 4 and d = 3, yielding E = 3 4 ∼ = 1.587, which is better than
those of most third-order methods E ≈ 1.442, Newton’s method and Steffensen’s
method E ≈ 1.414. Furthermore, the numerical examples considered here show
that in many cases all our proposed methods are efficient alternative to the existing
methods.
Acknowledgments The authors wish to thank the anonymous referees for their valuable sugges-
tions and comments.
References
1. Behl, R., Kanwar, V., Sharma, K.K.: Optimal equi-scaled families of Jarratt’s method. Int. J.
Comput. Math. 90, 408–422 (2013)
2. Cordero, A., Hueso, J.L., Martínez, E., Torregrosa, J.R.: Generating optimal derivative free
iterative methods for nonlinear equations by using polynomial interpolation. Math. Comput.
Model. 57, 1950–1956 (2013)
3. Cordero, A., Hueso, J.L., Martínez, E., Torregrosa, J.R.: A new technique to obtain derivative-
free optimal iterative methods for solving nonlinear equations. J. Comput. Appl. Math. 252,
95–102 (2013)
4. Cordero, A., Hueso, J.L., Martínez, E., Torregrosa, J.R.: Steffensen type methods for solving
nonlinear equations. J. Comput. Appl. Math. 236, 3058–3064 (2012)
5. Cordero, A., Torregrosa, J.R.: A class of Steffensen type methods with optimal order of con-
vergence. Appl. Math. Comput. 217, 7653–7659 (2011)
6. Kanwar, V., Behl, R., Sharma, K.K.: Simply constructed family of a Ostrowski’s method with
optimal order of convergence. Comput. Math. Appl. 62, 4021–4027 (2011)
7. King, R.F.: A family of fourth order methods for nonlinear equations. SIAM J. Numer. Anal.
10, 876–879 (1973)
8. Kung, H.T., Traub, J.F.: Optimal order of one-point and multi-point iteration. J. ACM 21,
643–651 (1974)
9. Ostrowski, A.M.: Solutions of Equations and System of Equations. Academic Press, New York
(1960)
10. Peng, Y., Feng, H., Li, Q., Zhang, X.: A fourth-order derivative-free algorithm for nonlinear
equations. J. Comput. Appl. Math. 235, 2551–2559 (2011)
11. Petković, M.S., Neta, B., Petković, L.D., Dzunić, J.: Multipoint Methods for Solving Nonlinear
Equations. Academic Press, New York (2012)
12. Petković, M.S., Ilić, S., Džunić, J.: Derivative free two-point methods with and without memory
for solving nonlinear equations. Appl. Math. Comput. 217, 1887–1895 (2010)
13. Soleymani, F., Sharma, R., Li, X., Tohidi, E.: An optimized derivative-free form of the Potra-
Pták method. Math. Comp. Model. 56, 97–104 (2012)
14. Traub, J.F.: Iterative Methods for the Solution of Equations. Prentice-Hall, Englewood Cliffs,
NJ (1964)
15. Zheng, Q., Zhao, P., Huang, F.: A family of fourth-order Steffensen-type methods with the
applications on solving nonlinear ODEs. Appl. Math. Comput. 217, 8196–8203 (2011)
Existence of Solution for Fractional
Stochastic Integro-Differential Equation
with Impulsive Effect
Mohd Nadeem and Jaydev Dabas
Abstract This paper is concerned with the existence and uniqueness of the solution
for an impulsive fractional stochastic integro-differential equation. The existence and
uniqueness results are shown using the fixed point technique on a Hilbert space.
Keywords Fractional order differential equation · Stochastic functional differential

equations · Existence results · Impulsive conditions
1 Introduction
It is well known that the fractional calculus is a classical mathematical notion and is
a generalization of ordinary differentiation and integration to arbitrary order. Nowa-
days, studying fractional calculus has become an active area of research field as
it has gained considerable importance due to its numerous applications in various
fields, such as physics, chemistry, viscoelasticity, engineering sciences, etc. For more
details, one can see the cited papers [1–8, 14] and reference therein.
The deterministic models often fluctuate due to environmental noise. A natural
extension of a deterministic model is stochastic model, where relevant parameters are
modeled as suitable stochastic processes. Due to this fact that, most of the problems in
a practical life situation are modeled by stochastic equations rather than deterministic.
Therefore, it is of great significance to introduce stochastic effects in the investigation
of differential equations [13]. For more details on stochastic differential equations
see [10–12] and references therein.
However, it is known that the impulsive effects exist widely in different areas of
real world such as mechanics, electronics, telecommunications, finance, economics,
M. Nadeem (B) · J. Dabas

Saharanpur Campus, Saharanpur 247001, India
e-mail: mohdnadeem.jmi@gmail.com
J. Dabas

DOI 10.1007/978-81-322-2485-3_30
374 M. Nadeem and J. Dabas
etc., for more detail see [9]. Due to this fact, the states of many evolutionary processes
are often subject to instantaneous perturbations and experience abrupt changes at
certain moments of time. The duration of these changes is very short and negligible
in comparison with the duration of the process considered, and can be thought of
as impulses. Therefore, it is important to consider the effect of impulses in the
investigation of stochastic differential equations.
Wang et al. [16] considered the following impulsive fractional differential equation
for order q ∈ (1, 2)
Dt u(t) = f (t, u(t)), t ∈ J = [0, T ], q ∈ (1, 2),

c q
Δu(tk ) = Ik (u(tk− )), Δu (tk ) = Jk (u(tk− )), k = 1, 2, . . . , m,

u(0) = u0 , u (0) = u0 ,
and discussed the existence and uniqueness of solutions with the help of Banach
fixed point theorem and Krasnoselskii fixed point theorem.
Sakthivel et al. [15] considered the following impulsive fractional stochastic dif-
ferential equations with infinite delay in the form
⎧
⎨ Dtα x(t) = Ax(t) + f (t, xt , B1 x(t)) + σ (t, xt , B2 x(t)) dw(t)
dt , t ∈ [0, T ], t = tk ,
Δx(tk ) = Ik (x(tk )), k = 1, 2, . . . , m,
⎩
x(t) = φ(t), φ(t) ∈ Bh ,
and discussed the existence of mild solutions using Banach contraction principle,
Krasnoselskii’s fixed point theorem.
Motivated by the mentioned work [15, 16], in this article, we are concerned with
the existence and uniqueness of solution for impulsive fractional functional integro-
differential equation of the form:
t
c
Dtα x(t)
= f t, x(t), xt , K(t, s)x(s)ds
0
t
dw(t)
+ g t, x(t), xt , K(t, s)x(s)ds , t ∈ J = [0, T ], t = tk , (1)
0 dt
Δx(tk ) = Ik (x(tk− )), Δx (tk ) = Qk (x(tk− )), k = 1, 2, . . . , m, (2)
x(t) = φ(t), x (0) = x1 , t ∈ [−d, 0], (3)
where J is an operational interval and c Dtα denotes the Caputo’s fractional derivative
of order α ∈ (1, 2) and x(·) takes the value in the real separable Hilbert space H ;
f : J × H × PCL 0 × H → H and g : J × H × PC 0 × H → L (K , H )
L
and Ik , Qk : H → H are appropriate functions; φ(t) is F0 -measurable H -valued
random variables independent of w. Here let 0 = t0 < t1 < · · · < tm < tm+1 = T ,
Δx(tk ) = x(tk+ ) − x(tk− ), Δx (tk ) = x (tk+ ) − x (tk− ), x(tk+ ) and x(tk− ) denote the
right and left limits of x at tk . Similarly, x (tk+ ) and x (tk− ) denote the right and left
limits of x at tk , respectively.
Existence of Solution for Fractional Stochastic Integro-Differential Equation ... 375
For further details, this work has three sections. Second section provides some
basic definitions, preliminaries, theorems, and lemmas. Third section is equipped
with main results for the considered problem (1)–(3).
2 Preliminaries
Let H , K be two real separable Hilbert spaces and L (K , H ) be the space of

bounded linear operators from K into H . For convenience, we will use the same
notation · to denote the norms in H , K and L (K , H ), and use (·, ·) to denote
the inner product of H and K without any confusion. Let (Ω, F , {Ft }t≥0 , P) be
a complete filtered probability space satisfying that F0 contains all P-null sets of
F . An H -valued random variable is an F -measurable function x(t) : Ω → H
and a collection of random variables S = {x(t, ω) : Ω → H \ t ∈ J} is called
stochastic process. Usually we write x(t) instead of x(t, ω) and x(t) : J → H in the
space of S. W = (Wt )t≥0 be a Q-Wiener process defined on (Ω, F , {Ft }t≥0 , P)
with the covariance operator Q such that TrQ < ∞. We assume that there exists a
complete orthonormal system {ek }k≥1 in K , a bounded sequence of nonnegative real
numbers λk such that Qek = λk ek , k = 1, 2, . . . , and a sequence of independent
Brownian motions {βk }k≥1 such that
∞

(w(t), e)K = λk (ek , e)K βk (t), e ∈ K , t ≥ 0.
k=1
1
Let L02 = L 2 (Q 2 K , H ) be the space of all Hilbert Schmidt operators from
1
Q 2 K to H with the inner product < ϕ, ψ >L 2 = Tr[ϕQψ∗].
0
The collection of all strongly measurable, square integrable, H -valued random
variables, denoted by L 2 (Ω, F , {Ft }t≥0 , P; H ) = L 2 (Ω; H ), is a Banach
space equipped with norm x(·)2L 2 = Ex(·, w)2H , where E denotes expectation

defined by E(h) = Ω h(w)dP. An important subspace is given by L02 (Ω; H ) =

{f ∈ L 2 (Ω, H ) : f is F0 - is measurable}.
Let PCL 0 = C([−d, 0], L 2 (Ω; H )) be a Banach space of all continuous map
from [−d, 0] into L 2 (Ω; H ) satisfying the condition sup Eφ(t)2 < ∞ with
norm

φPC 0 = sup Eφ(t)H , φ ∈ PCL 0
.
L
t∈[−d,0]
Consider C 2 (J, L 2 (Ω; H )) be a Banach space of all continuously differentiable

map from J into L 2 (Ω; H ) satisfying the condition sup Ex(t)2 < ∞ with norm
defined
1

x2C 2 = sup Ex j (t)2H , x ∈ C 2 (J, L 2 (Ω; H )) .
t∈J j=0
To study the impulsive conditions, we consider

2
PCL = PC 2 ([−d, T ], L 2 (Ω; H ))
a Banach space of all such continuous functions x : [−d, T ] → L 2 (Ω; H ), which

are continuously differentiable on [0, T ] except for a finite number of points ti ∈
(0, T ), i = 1, 2, . . . , N , at which x (ti+ ) and x (ti− ) = x (ti ) exist and are endowed
with the norm
1
x2PC 2 = sup Ex j (t)2H , x ∈ PCL 2
.
L t∈J j=0
Definition 1 The Reimann–Liouville fractional integral operator for order α > 0,

of a function f : R + → R and f ∈ L 1 (R + , X) is defined by
t
1
Jt0 f (t) = f (t), Jtα f (t) = (t − s)α−1 f (s)ds, α > 0, t > 0,
Γ (α) 0
where Γ (·) is the Gamma function.
Definition 2 Caputo’s derivative of order α > 0 for a function f : [0, ∞) → R is

defined as
t
1
Dtα f (t) = (t − s)n−α−1 f (n) (s)ds = J n−α f (n) (t),
Γ (n − α) 0
for n − 1 < α < n, n ∈ N. If 0 < α < 1, then

t
1
Dtα f (t) = (t − s)−α f (1) (s)ds.
Γ (1 − α) 0
Obviously, Caputo’s derivative of a constant is equal to zero.

Lemma 1 A measurable Ft -adapted stochastic process x : [−d, T ] → H such
that x ∈ PCL 2 is called a mild solution of the system (1)–(3) if x(0) = φ(0)
and x (0) = x1 on [−d, 0], Δx|t=tk = Ik (x(tk− )) and Δx |t=tk = Qk (x(tk− )),

k = 1, 2, · · · , m the restriction of x(·) to the interval [0, T )\t1 , · · · , tm is continuous
and x(t) satisfies the following fractional integral equation
⎧
t
t
⎪ α−1
⎪φ(0) + x1 t + Γ (α) 0 (t − s) f s, x(s), xs , 0 K(s, t)x(s)ds ds
1
⎪
⎪
⎪
⎪
t
α−1 g s, x(s), x , t K(s, t)x(s)ds dw(s),

⎪+ Γ (α)
⎪
1
0 (t − s) t ∈ (0, t1 ],
⎪
⎪
s 0
⎪
⎪φ(0) + x1 t + I1 (x(t − )) + Q1 (x(t − ))(t − t1 )
⎪
⎪ 1
⎪
⎪
t 1

⎪
⎪ + 1
(t − α−1 f s, x(s), x , t K(s, t)x(s)ds ds
⎪
⎨ Γ (α) 0 s) s

t
0
x(t) = + 1 (t − s) α−1 g s, x(s), x , t K(s, t)x(s)ds dw(s), t ∈ (t1 , t2 ],
⎪
⎪ Γ (α) 0 s 0
⎪
⎪
⎪
⎪ · · ·
⎪
⎪
⎪
⎪ φ(0) + x1 t + ki=1 Ii(x(ti− )) + Qi (x(ti− ))(t − ti )
⎪
⎪

⎪
⎪ t α−1 f s, x(s), x , t K(s, t)x(s)ds ds
⎪+ Γ (α)
⎪
1
0 (t − s)
⎪
⎪

s 0

t
⎪
⎩+ 1 t α−1
Γ (α) 0 (t − s) g s, x(s), xs , 0 K(s, t)x(s)ds dw(s), t ∈ (tk , tk+1 ].
Further, we introduce the following assumptions to establish our results:

(H1) The nonlinear maps f and g are continuous and there exit constants μ1 , μ2 ,
μ3 , v1 , v2 , v3 > 0 such that
Ef (t, x, ϕ, u) − f (t, y, ψ, v)2H ≤ μ1 x − y2H + μ2 ϕ − ψPC 0 + μ3 u − v2H ,

L
Eg(t, x, ϕ, u) − g(t, y, ψ, v)2H ≤ v1 x − y2H + v2 ϕ − ψPC 0 + v3 u − v2H

L
for all x, y, u, v ∈ H , t ∈ J and ϕ, ψ ∈ PC 0

L.
(H2) The functions Ik , Qk are continuous and there exists LI , LQ > 0, such that
EIk (x) − Ik (y)2H ≤ LI Ex − y2H ,

EQk (x) − Qk (y)2H ≤ LQ Ex − y2H
for all x, y ∈ H and k = 1, 2, · · · , m.
3 Existence and Uniqueness Results
This result is based on Banach contraction fixed point theory.

Theorem 1 Suppose that the assumptions (H1) and (H2) hold and

4T 2α 1 ∗ 1 ∗
Θ = 4(mLI + mT 2 LQ ) + (μ1 + μ2 + μ3 K ) + (v1 + v 2 + v 3 K ) < 1,
Γ (α) α 2 T (2α − 1)

t
where K ∗ = supt∈[0,t] 0 K(t, s)ds < ∞. Then the system (1)–(3) has a unique
solution.
Proof We convert the problem (1)–(3) into fixed point problem. We consider an
operator N : PCL
2 → PC 2 defined by
L
⎧
t

⎪ φ(0) + + 1
(t − α−1 f s, x(s), x , t K(s, t)x(s)ds ds
⎪
⎪ x 1 t Γ (α) s) s
⎪
⎪
t
0
t
0
⎪
⎪ α−1
⎪+ Γ (α) 0 (t − s) g s, x(s), xs , 0 K(s, t)x(s)ds dw(s), t ∈ (0, t1 ],
1
⎪
⎪
⎪
⎪ − −
⎪
⎪φ(0) +
x1 t + I1 (x(t1 )) + Q1 (x(t1
))(t − t1 )
⎪
⎪
⎪
⎪ t α−1 t
⎪+ Γ (α) 0 (t − s) f s, x(s), xs , 0 K(s, t)x(s)ds ds
1
⎨

(Nx)(t) = + 1 0t (t − s)α−1 g s, x(s), xs , 0t K(s, t)x(s)ds dw(s), t ∈ (t1 , t2 ],

⎪
⎪ Γ (α)
⎪
⎪· · ·
⎪
⎪
⎪
⎪
⎪
⎪φ(0) + x1 t + ki=1 Ii(x(ti− )) + Qi (x(ti− ))(t − ti )
⎪
⎪

⎪
⎪ t α−1 f s, x(s), x , t K(s, t)x(s)ds ds
⎪+ Γ (α)
⎪
1
0 (t − s)
⎪
⎪

s 0

t
⎪
⎩+ 1 t α−1
Γ (α) 0 (t − s) g s, x(s), xs , 0 K(s, t)x(s)ds dw(s), t ∈ (tk , tk+1 ].
Now we show that N is a contraction map. For this we take two points x, x ∗ such
that for t ∈ (0, t1 ]
t t
∗ 1 α−1
E(Nx)(t) − (Nx )(t)2H ≤ 2E (t − s) [f s, x(s), xs , K(s, t)x(s)ds
Γ (α) 0 0
t
−f s, x ∗ (s), xs∗ , K(s, t)x ∗ (s)ds ds2H
0
t t
1
+ 2E (t − s)α−1 [g(s, x(s), xs , K(s, t)x(s)ds)
Γ (α) 0 0
t
− g s, x ∗ (s), xs∗ , K(s, t)x ∗ (s)ds dw(s)2H
0

2T 2α 1 ∗
≤ (μ 1 + μ 2 + μ 3 K
Γ (α) α 2
1
+ (v1 + v2 + v3 K ∗ )x − x ∗ 2PC 2 .
T (2α − 1) L
When t ∈ (t1 , t2 ],
E(Nx)(t) − (Nx ∗ )(t)2H ≤ 4EI1 (x(t1− )) − I1 (x ∗ (t1− ))2H

+ 4E Q1 (x(t1− ))(t − t1 ) − Q1 (x ∗ (t1− ))(t − t1 )2H
t t
1
+ 4E (t − s)α−1 [f s, x(s), xs , K(s, t)x(s)ds
Γ (α) 0 0
t
− f s, x ∗ (s), xs∗ , K(s, t)x ∗ (s)ds ]ds2H
0
t t
1
+ 4E (t − s)α−1 [g(s, x(s), xs , K(s, t)x(s)ds)
Γ (α) 0 0
t
− g s, x ∗ (s), xs∗ , K(s, t)x ∗ (s)ds ]dw(s)2H
0

4T 2α 1
≤ 4(LI + T 2 LQ ) + (μ1 + μ2 + μ3 K ∗ )
Γ (α) α 2

1
+ (v1 + v2 + v3 K ∗ ) x − x ∗ 2PC 2 .
T (2α − 1) L
Similarly for t ∈ (tk , tk+1 ], k = 2, 3, . . . , m,

4T 2α 1
E(Nx)(t) − (Nx ∗ )(t)2H ≤ 4(mLI + mT 2 LQ ) + (μ1 + μ2 + μ3 K ∗ )
Γ (α) α 2

1
(v1 + v2 + v3 K ∗ ) x − x ∗ 2PC 2
T (2α − 1) L
= Θx − x ∗ 2PC 2 .
L
Since Θ < 1, by the condition given in Theorem 1, N is a contraction map and

therefore it has a unique fixed point x ∈ PCL2 which is a solution of our equation
(1)–(3) on J. This completes the proof of the theorem.
References
1. Chauhan, A., Dabas, J.: Existence of mild solutions for impulsive fractional-order semilinear
evolution equations with nonlocal conditions. Electron. J. Differ. Equ. 2011(107), 1–10 (2011)
2. Chauhan, A., Dabas, J.: Local and global existence of mild solution to an impulsive frac-
tional functional integro-differential equation with nonlocal condition. Commun. Nonlinear.
Sci. Numer. Simulat. 19, 821–829 (2014)
3. Chauhan, A., Dabas, J., Kumar, M.: Integral boundary-value problem for impulsive fractional
functional differential equations with infinite delay. Electron. J. Differ. Equ. 2012(229), 1–13
(2012)
fractional integro-differential equation with infinite delay. Math. Comput. Model. 57, 754–763
(2013)
5. Dabas, J., Chauhan, A., Kumar, M.: Existence of the mild solutions for impulsive fractional
equations with infinite delay. Intern. J. Differ. Equ. 793023, 20 (2011)
6. Dabas, J., Gautam, G.R.: Impulsive neutral fractional integro-differential equations with state
dependent delays and integral conditions. Electron. J. Differ. Equ. 2013(273), 1–13 (2013)
Equations. Elsevier Science B.V, Amsterdam (2006)
8. Lakshmikantham, V.: Theory of fractional differential equations. Nonlinear Anal. 69, 3337–
3343 (2008)
9. Lakshmikantham, V., Bainov, D., Simeonov, P.: Theory of Impulsive Differential Equations.
World Scientific Press, Singapore (1989)
10. Li, C., Sun, J., Sun, R.: Stability analysis of a class of stochastic differential delay equations
with nonlinear impulsive effects. J. Franklin Inst. 347, 1186–1198 (2010)
11. Lin, A., Ren, Y., Xia, N.: On neutral impulsive stochastic integro-differential equations with
infinite delays via fractional operators. Math. Comput. Model. 51, 413–424 (2010)
12. Longa, S., Teng, L., Xu, D.: Global attracting set and stability of stochastic neutral partial
functional differential equations with impulses. Stat. Probab. Lett. 82, 1699–1709 (2012)
13. Oksendal, B.: Stochastic Differential Equations. Springer, Berlin, Heidelberg (2005)
14. Podlubny, I.: Fractional Differential Equations. Acadmic Press, New York, USA (1993)
15. Sakthivel, R., Revathi, P., Ren, Y.: Existence of solutions for nonlinear fractional stochastic
differential equations. Nonlilear Anal. 81, 70–86 (2013)
16. Wang, J.R., Li, X., Wei, W.: On the natural solution of an impulsive fractional differential
equation of order α ∈ (1, 2). Commun. Nonlinear Sci. Numer. Simulat. 17, 4384–4394 (2012)
Singularly Perturbed Convection-Diffusion
Turning Point Problem with Shifts
Pratima Rai and Kapil K. Sharma
Abstract In this paper, a class of singularly perturbed turning point problem with
shifts (i.e., delay as well as advance) is considered. Presence of turning point results
into twin boundary layers in the solution of the problem under consideration. For
the numerical approximation of the problem, a finite difference scheme is proposed
on a uniform mesh. Interpolation is used to tackle the terms containing shifts and to
deal with the difficulty arising due to presence of the turning point a combination
of backward and forward difference is used in the first derivative term. Convergence
analysis is given for the proposed numerical scheme. Numerical results are presented
which illustrate the theoretical results and depict the effect of shifts on the layer
behavior of the solution.
Keywords Singular perturbation · Turning point · Positive shifts · Negative shifts ·

Finite difference scheme · Boundary layer
AMS Subject Classifications: 34K26 · 65L12 · 34K28
1 Introduction
In the recent past, numerical treatment of singularly perturbed differential equations

with positive/negative shifts have attracted the attention of many researchers. These
type of differential equations model a wide range of real-life phenomena including
variety of models of physiological processes or diseases [5], first exit problem in neu-
P. Rai
Department of Mathematics, University of Delhi, Delhi 110007, India
e-mail: prai@maths.du.ac.in
K.K. Sharma (B)
Department of Mathematics, South Asian University (SAU), Akbar Bhawan, Chanakyapuri
110021, New Delhi, India
e-mail: kapil.sharma@sau.ac.in

DOI 10.1007/978-81-322-2485-3_31
382 P. Rai and K.K. Sharma
robiology [18, 19], population ecology, materials with thermal memory, variational
problem in control theory [2, 4], etc.
In this paper, we consider a boundary value problem for singularly perturbed
differential-difference equation of the form
εu (x) + a(x)u (x) − b(x)u(x) + c(x)u(x − δ) + d(x)u(x + η) = f (x) x ∈ Ω = (−1, 1)

(1)
u(x) = ϕ(x), x ∈ Ω0 ; u(x) = γ (x), x ∈ Ω4 (2)
where 0 < ε 1, Ω̄ = [−1, 1], Ω = Ω1 ∪ Ω2 ∪ Ω3 , Ω1 = (−1, −1 + δ], Ω2 =

(−1 + δ, 1 − η), Ω3 = [1 − η, 1), Ω0 = [−1 − δ, −1], Ω4 = [1, 1 + η], δ, η are
the shift arguments independent of ε; a(x), b(x), c(x), d(x), ϕ(x), γ (x), f (x)
are sufficiently smooth functions satisfying
a(0) = 0, a (0) ≤ 0, (3)
|a (x)| ≥ |a (0)/2| ∀ x ∈ Ω̄, (4)
b(x) > 0, b(x) − c(x) − d(x) ≥ k > 0, c(x) > 0, d(x) > 0 ∀ x ∈ Ω. (5)
For u(x) to be smooth solution of the problem (1)–(2), it should satisfy boundary
conditions, be continuous on Ω̄ and differentiable on Ω. When shift terms are zero
(i.e., δ, η = 0), the above problem reduces to boundary value problem for singularly
perturbed differential equation with turning point. In this case, as the perturbation
parameter tends to zero there may be boundary or interior layer depending upon the
argument β = b(0)/a (0). For β > 0, the solution of the problem exhibit interior
layer around the turning point whereas for β < 0 there are twin boundary layers of
exponential type in the solution of the problem.
For questions on existence and uniqueness of differential equations with shifts,
one can see [6, 10–12] whereas numerical analysis is given in [3, 13]. Investigation of
boundary value problems for the second-order differential equations with shifts was
initiated by Lange and Miura [7–9] who extended the method of matched asymptotic
expansions developed for ODEs. Numerical study of singularly perturbed differential
equations with small shifts was considered in [14–16] where authors approximated
the terms containing shifts by Taylor series expansion. Taylor’s series holds good
when shifts are small but give bad approximation when shifts are large. In [1, 17],
authors constructed numerical schemes to deal with such type of problems when
shifts are large but, their study is limited to the case when the convection coefficient
has same sign throughout the domain, i.e., non-turning point case. In this paper, we
study the turning point case with both positive (advance) as well as negative (delay)
shifts.
Singularly Perturbed Convection-Diffusion Turning Point Problem with Shifts 383
An outline of the paper is as follows. In the next section, we discuss some proper-
ties of the exact solution. In Sect. 3, numerical scheme is described and convergence
properties of the proposed scheme are analyzed. Numerical experiments are pre-
sented in Sect. 4 which validate the theoretical results computationally. The paper
ends with a summary of the main conclusions.
We adopt certain conventions that throughout the paper C denotes a generic pos-
itive constant independent of N and ε, ||.|| is maximum norm.
2 Some Properties of the Exact Solution
In this section, we analyze some properties of the solution and its derivatives for
the problem (1)–(2) which are needed later on for the convergence analysis of the
proposed numerical method.
Let L ε denotes the differential operator occurring in problem (1)–(2) which is
defined as
⎧
⎪
⎨εu (x) + a(x)u (x) − b(x)u(x) + d(x)u(x + η) = f (x) − c(x)ϕ(x − δ) x ∈ Ω1
L ε u(x) = εu (x) + a(x)u (x) − b(x)u(x) + c(x)u(x − δ) + d(x)u(x + η) = f (x) x ∈ Ω2
⎪
⎩
εu (x) + a(x)u (x) − b(x)u(x) + c(x)u(x − δ) = f (x) − d(x)γ (x + η) x ∈ Ω3 .
(6)
L ε satisfy following minimum principle on Ω̄

Lemma 1 Let Ψ be a smooth function satisfying Ψ (−1) ≥ 0, Ψ (1) ≥ 0 and
L ε Ψ (x) ≤ 0, ∀ x ∈ Ω. Then, Ψ (x) ≥ 0 ∀ x ∈ Ω̄.

Immediate consequence of the above minimum principle is the following stability
estimate.
Lemma 2 If u(x) is solution of the problem (1)–(2) then for some positive constant
C we have
||u(x)|| ≤ C || f ||/k + max{||ϕ||, ||γ ||} . (7)

Next, we divide the domain Ω̄ into three regions, D1 = [−1, −μ], D2 =
(−μ, μ), D3 = [μ, 1], 0 < μ ≤ 1/2. Following theorem gives us bound on
the derivatives of the solution u(x) of the problem (1)–(2) in the intervals D1 , D3 .
Theorem 3 Let a(x), b(x), c(x), d(x), f (x) ∈ C j (Ω̄), ϕ(x) ∈ C j (Ω0 ), γ (x) ∈
C j (Ω4 ), ||a|| = M, |a(x)| ≥ α > 0, x ∈ D1 ∪ D3 . Then, there exist a positive
constant C such that the solution u(x) of the problem (1)–(2) satisfies

(i) C 1 + ε−i e−α(1−x)/ε , i = 1, . . . j + 1, x ∈ D3

|u (x)| ≤
C 1 + ε−i e−α(x+1)/ε i = 1, . . . j + 1, x ∈ D1 .

Theorem 3 gives us bound on the derivatives of the solution outside the turning
point region. Therefore, we are left with obtaining the bound on the derivatives of
the solution in the region D2 which is given by the following theorem.
Theorem 4 Let a(x), b(x), c(x), d(x), f (x) ∈ C j (Ω̄), ϕ(x) ∈ C j (Ω0 ), γ (x) ∈
C j (Ω4 ) and conditions (3)–(5) holds. Then there exist a positive constant C such
that the solution u(x) of the problem (1)–(2) satisfies
|u (i) (x)| ≤ C, i = 1, . . . j, x ∈ D2 . (8)
3 Discretization and Convergence
In this section, we describe a upwind finite difference scheme on uniform mesh. Let
w = xi = −1 + i h, where i = 1, 2, . . . , N − 1; h = 2/N be uniform mesh on
Ω, w̄ = w ∪ {−1, 1} and x = 0 be the turning point. To tackle the terms containing
positive/negative shifts, we use interpolation. Earlier, authors [1, 17] constructed
numerical schemes in which they considered a special type of mesh in which the
term containing shifts lie at the nodal point. But this mesh selection has a drawback
that it put restriction on the number of mesh generations. To overcome this drawback,
we propose a scheme which works equally well in both the cases, i.e., whether the
terms containing shifts lie at the node or not. If xi −δ, xi +η, i = 0, . . . , N are not the
nodal points then there exist 0 < m 0 , m 1 < N such that m 0 h < δ < (m 0 + 1)h
and m 1 h < η < (m 1 + 1)h. In our algorithm, interpolation is used to approximate
the value of xi − δ and xi + η in terms of neighboring nodal points.
We introduce certain notations for the mesh functions. For any mesh function
g(x), we have
gi = g(xi ), D+ gi = (gi+1 − gi )/ h, D− gi = (gi − gi−1 )/ h

||g|| ≡ ||g||∞ = max0≤i≤N |gi |, D+ D− gi = (gi+1 − 2gi + gi−1 )/ h 2
si = xi − δ, ri = xi + η

D+ gi , ai > 0
D∗ gi =
D− gi , ai < 0.
The difference scheme for the boundary value problem (1)–(2) is given by
ci
L N Ui = ε D+ D− Ui + ai D∗ Ui − bi Ui + (si − xi−m 0 −1 )Ui−m 0 + (xi−m 0 − si )Ui−m 0 −1
h
di
+ (ri − xi+m 1 )Ui+m 1 +1 + (xi+1+m 1 − ri )Ui+m 1 = f i (9)
h
Ui = ϕi , i = −m 0 , . . . , 0
Ui = γi , i = N , . . . , N + m 1 .
Lemma 5 Suppose Ψ0 ≥ 0 and Ψ N ≥ 0. Then, L ε Ψi ≤ 0 for all i = 1, 2, . . . , N −

1 implies Ψi ≥ 0 for all i = 0, 1, . . . , N .
Lemma 6 Let Ui be the solution of the problem (9). Then

||U || ≤ C || f ||/K + max{||ϕ||, ||γ ||} . (10)
where K = min1≤i≤N −1 (bi − ci − di ).
Theorem 7 Assume h < ε. Then the error ei = L N (u i − Ui ) between the solution

u(xi ) of the continuous problem (1)–(2) and the solution Ui of the discrete problem
(9) satisfies the estimate
h
|ei | ≤ C , i = 0, . . . N . (11)
h+ε

Lemma 6 and Theorem 7 gives us following theorem
Theorem 8 Let u be solution of (1) and (2) and U be solution of (9). Then
h
||U − u|| ≤ C .
h+ε
4 Numerical Results
In this section, we computationally verify the theoretical results obtained in the

previous section. For this we study the performance of the proposed scheme (9)
applied to the following test examples. For δ, η = 0 the solution of the boundary
value problem exhibit twin boundary layers. We also illustrate the effect of shift
on the solution behavior. Since exact solutions are not known for the considered
examples we use double mesh principle to estimate the accuracy in the maximum
norm
E N = max0≤i≤N |UiN − U2i2N |,
and the convergence rate

EN
R N = log2 .
E 2N
Example 1
εu (x) + 4(1 − 2x)u (x) − 4u(x) + u(x − δ) + 2u(x + η) = 0, x ∈ (0, 1) (12)

u(x) = 1, −δ ≤ x ≤ 0, u(x) = 1, 1 ≤ x ≤ 1 + η. (13)
5 Conclusion and Discussion
In this paper, a class of linear second-order singularly perturbed convection-diffusion

turning point problem with shifts is considered. The solution of such type of differ-
ential equations have boundary or interior layers depending upon the sign of the
convection and the reaction coefficients. Here, we considered the case where the
presence of turning point results into twin boundary layers. There are also terms
containing positive and negative shifts.
Difficulty arising due to the presence of the turning point is tackled using com-
bination of forward and backward difference in the numerical approximation of the
first derivative term. Interpolation is used to deal with the terms containing shifts. The
Tables 1, 2 and 3 gives maximum pointwise error E N and the rate of convergence
R N for the considered example for different nonzero values of δ, η. Table 4 gives
maximum point wise error for the solution when δ = 0, i.e., when we have only pos-
itive shifts, whereas Table 5 gives maximum point wise error for the solution when
η = 0, i.e., when only negative shift is there. Graphs are plotted (Figs. 1, 2 and 3) for
the considered example to illustrate the effect of shifts on the layer behavior of the
solution. It is observed that as the value of the shift argument increases the thickness
of the boundary layers increases and steepness decreases. Moreover, the magnitude
of the shift depends upon the value of the positive/negative shifts and the coefficient
of the terms containing shifts.
Table 1 Maximum pointwise error E N and rate for convergence R N for δ = 0.15, η = 0.15
N = 128 N = 256 N = 512 N = 1024 N = 2048
1 7.56580E − 4 3.78626E − 4 1.12511E − 4 5.61834E − 5 4.73650E −5
0.99 1.7 1.0 0.25 1.0
2−1 3.89010E − 3 1.97371E − 3 6.22136E − 4 3.11787E − 4 2.49418E −4
0.98 1.7 1.0 0.32 1.0
2−2 1.40237E − 2 7.20224E − 3 2.49091E − 3 1.25406E − 3 9.17788E −4
0.96 1.5 0.99 0.45 1.0
2−3 2.29938E − 2 1.20099E − 2 5.11759E − 3 2.59958E − 3 1.56537E −3
0.94 1.2 0.98 0.73 0.99
2−4 3.46149E − 2 1.93660E − 2 9.90240E − 3 5.13310E − 3 2.69769E −3
0.84 0.97 0.95 0.93 0.99
2−5 5.30166E − 2 3.31337E − 2 1.87340E − 2 1.00508E − 2 5.24686E −3
0.68 0.82 0.9 0.94 0.97
2−6 8.01397E − 2 5.30070E − 2 3.31428E − 2 1.88765E − 2 1.01488E −2
0.6 0.68 0.81 0.9 0.95
2−7 8.60812E − 2 7.98999E − 2 5.30411E − 2 3.32542E − 2 1.89585E −2
1.1 0.59 0.67 0.81 0.9
ε↓ N = 128 N = 256 N = 512 N = 1024 N = 2048
1 7.55930E − 4 1.71292E − 4 8.54595E − 5 6.45943E − 5 4.23081E −5
2.1 1.0 0.4 0.6 2.1
2−1 3.44352E − 3 8.93709E − 4 4.49362E − 4 4.38537E − 4 2.19673E −4
1.9 0.99 0.35 1.0 1.9
2−2 1.12767E − 2 3.20727E − 3 1.62803E − 3 1.46869E − 3 7.37259E −4
1.8 0.97 0.15 0.99 1.8
2−3 2.09117E − 2 7.08495E − 3 3.66956E − 3 2.83638E − 3 1.42862E −3
1.6 0.95 0.37 1.0 1.6
2−4 3.39450E − 2 1.48829E − 2 7.98753E − 3 5.12619E − 3 2.60610E −3
1.2 0.9 0.64 0.98 1.3
2−5 5.03637E − 2 2.78583E − 2 1.58758E − 2 9.46715E − 3 4.89823E −3
0.85 0.81 0.75 0.95 1.1
2−6 7.59403E − 2 4.57304E − 2 2.86893E − 2 1.72498E − 2 9.26065E −3
0.73 0.67 0.73 0.9 1.0
2−7 8.26705E − 2 6.93906E − 2 4.64092E − 2 2.99831E − 2 1.70699E −2
0.25 0.58 0.63 0.81 0.93
ε↓ N = 128 N = 256 N = 512 N = 1024 N = 2048
1 4.79304E − 4 2.38424E − 4 1.74441E − 4 8.71994E − 5 2.98139E −5
1.0 0.45 1.0 1.5 1.0
2−1 2.45619E − 3 1.23898E − 3 8.91549E − 4 4.47194E − 4 1.56942E −4
0.99 0.47 0.99 1.5 1.0
2−2 9.00776E − 3 4.62390E − 3 3.20011E − 3 1.61050E − 3 5.94638E −4
0.96 0.53 0.99 1.4 1.0
2−3 1.75028E − 2 9.23944E − 3 5.64799E − 3 2.86161E − 3 1.21974E −3
0.92 0.71 0.98 1.2 0.99
2−4 3.06080E − 2 1.74018E − 2 1.00675E − 2 5.20123E − 3 2.46531E −3
0.81 0.79 0.95 1.1 0.99
2−5 4.97544E − 2 3.11371E − 2 1.84108E − 2 9.87295E − 3 4.94305E −3
0.68 0.76 0.9 1.0 0.97
2−6 7.57999E − 2 5.02539E − 2 3.21188E − 2 1.82812E − 2 9.64407E −3
0.59 0.65 0.81 0.92 0.95
2−7 8.16723E − 2 7.60033E − 2 5.11203E − 2 3.20309E − 2 1.80823E −2
1.0 0.57 0.67 0.82 0.9
Table 4 Maximum pointwise error E N for δ = 0, ε = .01

η↓ N = 128 N = 256 N = 512 N = 1024 N = 2048
0.15 9.06105E −2 7.49621E −2 4.76055E −2 2.88584E −2 1.59083E −2
0.25 9.09313E −2 7.27323E −2 4.61588E −2 2.77859E −2 1.52839E −2
0.35 2.48565E −1 5.79099E −1 7.31938E −1 9.10935E −1 9.15949E −1
0.45 5.47839E −2 5.15808E −2 3.26882E −2 1.83858E −2 1.01365E −2
Table 5 Maximum pointwise error E N for η = 0, ε = .01

δ↓ N = 128 N = 256 N = 512 N = 1024 N = 2048
0.15 9.04784E −2 7.49711E −2 4.77148E −2 2.89146E −2 1.59173E −2
0.25 9.00035E −2 7.43240E −2 4.72919E −2 2.86399E −2 1.57649E −2
0.35 8.47844E −2 6.99147E −2 4.44824E −2 2.69308E −2 1.48238E −2
0.45 7.83714E −2 6.36476E −2 4.04636E −2 2.44278E −2 1.34427E −2
Fig. 1 The numerical 1.4

solution for Example 1 when δ=0, η=0
δ and η are nonzero (ε = 0.1) 1.2
δ=0.1, η=0.4
δ=0.42, η=0.25
δ=0.2, η=0.25
1
0.8
u
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
x

solution for Example 1 when η=0.15
δ = 0 (ε = 0.1) 1.2 η=0.25
η=0.35
1 η=0.45
0.8
u
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
x

solution for Example 1 when δ =0.15
η = 0 (ε = 0.1) 1.2 δ =0.25
δ =0.35
1 δ =0.45
0.8
u
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
x
Acknowledgments The first author wish to acknowledge Delhi University, India, 110007 for pro-
viding financial support for the above work.
References
1. Amiraliyev, G.M., Cimen, E.: Numerical method for a singularly perturbed convection-
diffusion problem with delay. Appl. Math. Comput. 216, 2351–2359 (2010)
2. Bellman, R., Cooke, K.L.: Differential Difference Equations. Academic Press, New York
(1991)
3. Cryer, C.W.: Numerical methods for functional differential equations, in delay and functional
differential equations. In: Schmitt, K. (ed.) Proceeding of Conference held at Park city, Utah,
1972
4. E’lsgolts, L.E.: Qualitative Methods in Mathematical Analysis, Translations of Mathematical
Monographs, vol. 12. American Mathematical Society, Providence (1964)
5. Glass, L., Mackey, M.C.: Oscillation and chaos in physiological control system. Science 197,
287–289 (1977)
6. Henderson, J.: Boundary Value Problems for Functional Differential Equations. World Scien-
tific (1995)
7. Lange, C.G., Miura, R.M.: Singular perturbation analysis of boundary value problems for
differential-difference equations. SIAM J. Appl. Math. 42, 502–531 (1982)
differential-difference equations II. Rapid oscillations and resonance. SIAM J. Appl. Math. 45,
687–707 (1985)
differential-difference equations V. Small shifts with layer behavior. SIAM J. Appl. Math. 54,
249–272 (1994)
10. Lee, J.W., O’Regan, D.: Existence results for differential delay equations-II. Non-linear Anal.
17, 683–702 (1991)
11. Lee, J.W., O’Regan, D.: Existence results for differential delay equations-I. J. Diff. Eqn. 102,
342–359 (1993)
12. Ntouyas, S.K., Sficas, Y.G., Tsamatos, P.: An existence principle for boundary value problems
for second order functional differential equations. Nonlinear Anal. 20, 215–222 (1993)
13. Norkin, S.B.: Differential Equations of the Second Order with Retarded Arguments, Trans-
lations of Mathematical Monographs, vol. 31. American Mathematical Society, Providence
(1972)
14. Ramos, J.I.: Exponential methods for singularly perturbed ordinary differential-difference
equations. Appl. Math. Comput. 182, 1528–1541 (2006)
15. Kadalbajoo, M.K., Kumar, D.: Fitted mesh B-spline collocation method for singularly perturbed
differential-difference equations with small delay. Appl. Math. Comput. 204, 90–98 (2008)
16. Kadalbajoo, M.K., Sharma, K.K.: Parameter- uniform fitted mesh method for singularly per-
turbed delay differential equations with layer behaviour. Elect. Trans. Numer. Anal. 23, 180–201
(2006)
17. Kadalbajoo, M.K., Sharma, K.K.: A numerical method based on finite difference for boundary
value problems for singularly perturbed delay differential equations. Appl. Math. Comput. 197,
692–707 (2008)
18. Stein, R.B.: A theoretical analysis of neuronal variability. Biophys. J. 5, 173–194 (1965)
19. Tuckwell, H.C.: On the first exit time problem for temporally homogeneous Markov process.
J. Appl. Probab. 13, 39–46 (1976)
High-Order Compact Finite Difference
Method for Black–Scholes PDE
Kuldip Singh Patel and Mani Mehra
Abstract In this paper, Black–Scholes PDE is solved for European option

pricing by high-order compact finite difference method using polynomial interpola-
tion. Numerical results obtained are compared with standard finite difference method
and error with the analytic solution is discussed.
Keywords Option pricing · European options · Black-Scholes PDE · Compact

finite difference methods
1 Introduction
The seminal work in the area of option pricing was done by Black and Scholes
in 1973 [1] for pricing European option by solving a parabolic partial differential
equation (PDE), (commonly known as Black–Scholes PDE). Many computational
techniques [2] such as finite difference methods, spectral methods, Fast Fourier trans-
form techniques [3] have been extensively used for solving Black–Scholes PDE. In
the era of advanced computational techniques, compact finite difference method [4]
is highly recommended. Various compact finite difference techniques have already
been studied for Black–Scholes PDE for pricing the European and American option
[5, 6] and their convergence has also been studied [7].
An option, in finance, is a contract that gives its owner the right (but not the
obligation) to buy or sell a prescribed amount of particular asset from the writer of
the option for a prescribed fixed price (called the strike price) on or before the certain
date (called maturity date). For various purpose, there are many kinds of options,
such as, vanilla options (European call or put option, American call or put option),
Asian option, Bermudan option, exotic option, look-back option, barrier option, etc.
K.S. Patel (B) · M. Mehra

Indian Institute of Technology, Delhi, India
e-mail: kuldip@maths.iitd.ac.in
M. Mehra
e-mail: mmehra@maths.iitd.ac.in

DOI 10.1007/978-81-322-2485-3_32
394 K.S. Patel and M. Mehra
[8, 9]. Options that can be exercised only on the maturity date are called European
option, while options that can be exercised at any time up to the maturity date are
called American option. If the option is to buy the asset it is a call option, if to sell
the asset it is a put option.
Various finite difference methods have been studied for Black–Scholes PDE. A
major disadvantage of the finite difference approach is the widening of the compu-
tational stencil as the order of the approximation is increased. High-order compact
finite difference schemes [4] consider the value of the function and its first/higher
derivatives as unknowns at each discretization point. Compared to standard explicit
finite difference schemes, these schemes are implicit and give a higher order of
accuracy for the same number of grid points and also provide high resolution char-
acteristics [10]. This feature brings them closer to the spectral methods, while the
freedom in choosing the mesh geometry and the boundary conditions is maintained.
The rest of the paper is organized as follows. In Sect. 1.1, some introduction is
given about Black–Scholes PDE. In Sect. 1.2, a very short review of finite difference
method is given. In Sect. 1.3, compact finite difference scheme for first and second
derivative is discussed. In Sect. 2.1, analytical solution of Black–Scholes PDE is
given. In Sect. 2.2, solution of Black–Scholes PDE by compact finite difference
method is given and error with the analytic solution is discussed.
1.1 Black–Scholes PDE
It is not always easy to determine the value of option because of the stochastic nature
of financial markets. Option pricing theory has made a great leap forward since the
development of the Black–Scholes option pricing model by Black and Scholes in
1973 [1] and previously by Merton in 1973 [11]. The famous Black–Scholes PDE
can be written as:
∂V 1 ∂2V ∂V
+ σ 2 S2 2 + r S − r V = 0. (1)
∂t 2 ∂S ∂S
The solution of above famous Black–Scholes PDE provides both an option pricing
for European call and put option and a hedging portfolio that replicates the contingent
claim under the following assumptions [12]:
• The asset price S follows geometric Brownian motion, i.e. S satisfies the following
stochastic differential equation:
d S = μS dt + σ S dW,
where μ is the drift rate, σ is the volatility and dW is the increment of a standard
Brownian motion.
• The drift μ (which measures the average rate of growth of the asset price),
the volatility σ (which measures the standard deviation of the returns) and the
High-Order Compact Finite Difference Method for Black–Scholes PDE 395
risk-free interest rate r are constant for 0 ≤ t ≤ T and no dividends are paid in
that time period.
• The market is frictionless.
• There are no arbitrage opportunities.
Under the assumption discussed above the market is complete. The completeness
of market implies that any derivative and any asset can be replicated or hedged with
a portfolio of other assets in the market [13]. The parabolic PDE given in Eq. (1)
can be transformed into the heat equation and solved analytically to price the option
[14].
1.2 Finite Difference Method
In general, finite difference (FD) methods are used to numerically approximate the
solutions of certain ordinary and partial differential equations. In the case of a bivari-
ate, parabolic PDE, such as Eq. (1), we start by establishing a rectangular solution
domain in the two variables, S and t. We then form finite difference approximations
to each of the derivative terms in the PDE. Perhaps the most popular FD methods
used in computational finance are:
• Explicit Euler,
• Implicit Euler and
• Crank–Nicolson method.
Using each of these three methods has its advantages and disadvantages. The
easiest scheme among the above three methods to implement is the explicit Euler
method. The main disadvantage in using explicit Euler is that it is unstable for
certain choices of domain discretization. Though Implicit Euler and Crank–Nicolson
methods involve solving linear systems of equations at each time step, they are each
unconditionally stable with respect to the domain discretization. Crank–Nicolson
exhibits the greatest accuracy among the above three methods for a given domain
discretization.
1.3 Compact Finite Difference Method
High-order compact finite difference schemes [15] are developed conventionally

using method of undetermined coefficient [4, 16]. Polynomial interpolation has also
been used to derive arbitrarily high-order compact schemes for the first and second
derivatives on non-uniform grids [17]. Boundary and near boundary schemes of the
same order as the interior have also been developed using polynomial interpolation
Table 1 Compact finite difference scheme for first derivative

Index i In , Im Uniform grid,
xi = x1 + h(i − 1)
1 (3, 4), (1, 2) f 1 + 3 f 2
= −17 6h f 1 + 2h f 2 + 2h f 3 − 6h f 4
3 3 1
1 1
2, 3, .., N − 1 (i − 1, i + 1), (i) 4 f i−1 + f i + 4 f i+1
= 4h 3
( f i+1 − f i−1 )
N (N − 2, N − 3), (N , N − 1) f N + 3 f N −1 = 6h
17
fN
− 2h f N −1 − 2h f N −2 + 6h
3 3 1
f N −3
Table 2 Compact finite difference scheme for second derivative

Index i Uniform grid, xi = x1 + h(i − 1)
1 f 1 + 44 f 2 =13
h2 1
f − h272 f 2 + h152 f 3 − 1
f
h2 4
1 1
2, 3, .., N − 1 10 f i−1 + f i + 10 f i+1
= 5h 2 ( f i+1 − f i−1 ) − 5h
6 12
2 fi
N
f N + 44 f N −1
= h132 f N − h272 f N −1 + h152 f N −2 − h12 f N −3
[17]. It has been proved that polynomial interpolation is more efficient than the
conventional method of undetermined coefficients [16] for finding coefficients of the
scheme.
Consider a set of n points In on which values of the function and its first derivative
have been specified and another set of m points Im on which only function values
have been specified. The independent variable representing the points is xi , i being
the index of the node and the function values are given by f i = f (xi ). First derivative
is given by f i = f (xi ) and second derivative is given by f i = f (xi ). Compact
finite difference scheme for first and second derivative on uniform grid with step size
h are given in Tables 1 and 2 respectively.
2 Numerical Results
2.1 Closed Form Solution of Black–Scholes PDE
Black–Scholes PDE can be written as
∂V 1 ∂2V ∂V
+ σ 2 S2 2 + r S − r V = 0. (2)
∂t 2 ∂S ∂S
By the definition of the European option, it is clear that at expiry date T, the value
of European option V (S, t) (also called as pay-off function) is given by

max(S-X,0) for a European call,
V (S, T ) = (3)
max(X-S,0) for a European put,
and the solution V (S, t) of the Black–Scholes PDE (Eq. (2)) with the above final
condition (Eq. (3)) is given by

S N (d1) − X N (d2)e−r (T −t) for a European call,
V (S, t) = (4)
X N (−d2)e−r (T −t) − S N (−d1) for a European put,
where X is the strike price, r is the interest rate,

x −y 2
1 2
N (x) = √ e dy,
2π −∞

log(S/ X ) + r + 21 σ 2 (T − t)
d1 = √ ,
σ T −t

log(S/ X ) + r − 21 σ 2 (T − t)
d2 = √ ,
σ T −t
this implies √
d2 = d1 − σ T − t.
For example price of European put option (Fig. 1) for a non-dividend paying stock
by the above formula is
V = 1.6306, (5)
for the given stock price S = 50, strike price X = 50, interest rate r = 0.1, volatility
σ = 0.2 and time to maturity T = 5/12.
Fig. 1 Price of European

put options calculated by
60
closed form solution for
X = 50, S = 50, r = 0.10,
40
σ = 0.20 and T = 5/12
V(S,t)
20
0
1
150
0.5 100
50
t 0 0 S
2.2 Compact Finite Difference Method for Black–Scholes PDE
In the following, Black–Scholes PDE is solved by compact finite difference method

using polynomial interpolation to get the option price and the price of option obtained
by closed form solution of Black–Scholes PDE will serve as benchmark for compar-
ison. The initial and boundary conditions for a European call option are
V (S, T ) = max(S − X, 0),
V (0, t) = 0,
lim V (S, t) = S,
S→∞
while the initial and boundary conditions for a European put option are
V (S, T ) = max(X − S, 0),
V (0, t) = X e−r (T −t) ,
lim V (S, t) = 0,
S→∞
with X as the exercise price.

Black–Scholes PDE (Eq. (1)) can be written as:
1 ∂2 ∂
Vt = L V where L ≡ − σ 2 S 2 2 − r S + r I. (6)
2 ∂S ∂S
If we write ∂∂S ≡ D and ∂∂S 2 ≡ D 2 (where D and D 2 are compact finite difference
2
differentiation matrix for first and second derivative respectively), then
1
L ≡ − σ 2 S 2 D 2 − r S D + r I,
2
where D = A−1 B and D 2 = P −1 Q. Now time discretization can be done for Eq. (6)
according to implicit Euler, explicit Euler and Crank–Nikolson method. Matrices A,
B and P, Q (constructed from Tables 1 and 2 respectively) are given as:

60
finite difference method for
X = 50, S = 50, T = 5/12,
40
r = 0.10, σ = 0.20,
V(S,t)
Smin = 0, Smax = 150,
20
M = 100 and N = 500
0
1
150
0.5 100
50
t 0 0 S

60
compact finite difference
method for X = 50, S = 50,
40
T = 5/12, r = 0.10,
V(S,t)
σ = 0.20, Smin = 0,
20
Smax = 150, M = 100 and
N = 500
0
1
150
0.5 100
50
t 0 0 S
⎛ ⎞ ⎛ −17 3 3 −1
⎞
1 3 0 0 ... ... 0 6h 2h 2h 6h ... ... 0
⎜1 .. ⎟ ⎜ .. ⎟
⎜ 1 1 0 ⎜ −3 ⎟
.⎟ .
3
⎜4 ⎟ ⎜ 4h 0 4h 0 ⎟
⎜ ..
4
. .. ⎟ ⎜ .. .. .. ⎟
A=⎜. .. .⎟ B=⎜
⎜ . . . ⎟
⎟ (7)
⎜ ⎟ ⎜ ⎟
⎜. .. 1 ⎟ ⎜ .. .. ⎟
⎝ .. . 4 1 14 ⎠ ⎝ . −3
.
4h 0 4h
3
⎠
0 ... ... 3 1 0 ... ... 1 −3 −3 17
6h 2h 2h 6h
⎛ ⎞ ⎛ 13 −27 15 −1 ⎞
1 44 0 0 . . . . . . 0 h2 h2 h2 h2
... ... 0
⎜ 1 .. ⎟ ⎜ .. ⎟
⎜ ⎜ 6 −12 6
. ⎟
1
⎜ 10 1 10 0 . ⎟
⎟ ⎜ 5h 2 5h 2 5h 2
0 ⎟
⎜ .. ⎜ .. .. ⎟
P=⎜ . . .. .. ⎟
.
⎟ Q=⎜
⎜ .
..
. . ⎟
⎟
⎜ ⎟ ⎜ ⎟
⎜ . .. 1 ⎟ ⎜ .. .. ⎟
⎝ .. . 10 1 10 1 ⎠
⎝ . 6 −12 6
. ⎠
5h 2 5h 2 5h 2
0 ... ... 44 1 0 ... ... −1 15 −27 13
h2 h2 h2 h2
(8)
Price of European put option by closed form solution (Fig. 1), by standard finite
difference method (Fig. 2) and by compact finite difference method (Fig. 3) for a
non-dividend paying stock are given in Table 3. It can be noticed that explicit (Euler)
Table 3 Price of European put option for parameters S = 50, strike price X = 50, interest rate
r = 0.1, volatility σ = 0.2, time to maturity T = 5/12, number of time steps M = 100 number of
space steps N = 500, Smin = 0 and Smax = 150
Method Standard finite Compact finite Closed form solution
difference method difference method
Explicit Euler −7.2447e90 6.1154e106
Implicit Euler 1.6277 1.6279 1.6306
Crank–Nicolson 1.6304 1.6307
method
Fig. 4 Error versus number −2

10
of space steps in finite ||error||∞
difference method by O(N−2)

Crank–Nicolson method for
X = 0.2, T = 5/12,
r = 0.1, σ = 0.4, Smin = 0,
−3
10
Smax = 1, and M = 200
||error||∞
−4
10
−5
10
10 20 30 40 50 60 70 80 90
No. of space steps
method gives unstable results and Crank–Nicolson method gives the better results
for the same parameter.
Error with the analytic solution is plotted for finite difference method (Fig. 4)
and for compact finite difference method (Fig. 5) versus number of grid points. It
can be seen from Figs. 4 and 5 that method presented in this paper has more order
of convergence (O(h 3 )) as compared to standard finite difference method (O(h 2 )).
From Table 4, it can be concluded that high order of accuracy is obtained by using
compact finite difference method (up to 10−6 for N = 200) as compared to standard
finite difference method (up to 10−4 for N = 200) for the same parameters.
In Figs. 6 and 7, the results for various time steps are compared and it is observed
that three standard finite difference methods converge to identical precision for lager
number of time steps. Errors are plotted as the difference between analytic Black–
Scholes price and the finite difference price. In Figs. 8 and 9, the results for various
time steps are compared and it is observed that three compact finite difference meth-
ods converge to identical precision for larger number of time steps. Errors are plotted
as the difference between analytic Black–Scholes price and the compact finite dif-
ference price. It can be noticed that error is largest at the money.
Fig. 5 Error versus number −2

10
of space steps in compact ||error||
∞
−3
finite difference method by O(N )
Crank–Nicolson method for −3
X = 0.2, T = 5/12, 10
r = 0.1, σ = 0.4, Smin = 0,

Smax = 1, and M = 200
∞
||error||
−4
10
−5
10
−6
10
10 20 30 40 50 60
No. of space steps
Table 4 Comparison of errors for parameters S = 50, strike price X = 50, interest rate r = 0.3,
volatility σ = 0.05, time to maturity T = 5/12, number of time steps M = 100, Smin = 0 and
Smax = 150
Number of grid points Error in standard finite Error in compact finite
difference method difference method difference method
N = 50 0.2712 0.0444
N = 100 0.0461 0.0045
N = 150 0.0033 8.8943e-06
N = 200 3.1888e-04 1.0864e-06
Fig. 6 Error versus strike x 10

−3
price with 30 time steps in 2.5

Crank−Nicolson
finite difference method for Explicit
2
X = 10, S = 10, T = Implicit
0.1, r = 0.05, σ = 1.5
0.10, Smin = 5, Smax =
Error
15, N = 100, M = 30 1
0.5
−0.5
5 10 15
Strike price

−4
20
price with 300 time steps in Crank−Nicolson
finite difference method for Explicit
X = 10, S = 10, T = 15 Implicit
0.1, r = 0.05, σ =
0.10, Smin = 5, Smax = 10
Error
15, N = 100, M = 300
5
−5
5 10 15
Strike price

−3
1.4
price with 30 time steps in Crank−Nicolson
compact finite difference 1.2 Implicit
method for X = 10, S = Explicit
10, T = 0.1, r = 0.05, σ = 1
0.05, Smin = 5, Smax = 15
0.8
Error
0.6
0.4
0.2
0
5 10 15
Strike price

−3
price with 60 time steps in 1.2

Crank−Nicolson
compact finite difference Implicit
1
method for X = 10, S = Explicit
10, T = 0.1, r = 0.05, σ =
0.8
0.05, Smin = 5, Smax = 15
Error
0.6
0.4
0.2
0
5 10 15
Strike price
3 Discussion and Conclusion
In this paper, compact finite difference method using polynomial interpolation is

used to solve Black–Scholes PDE for pricing European put option. It can be noticed
from Table 4 that high-order accuracy is obtained by using compact finite difference
method (up to 10−6 for N = 200) as compared to standard finite difference method
(up to 10−4 for N = 200) for the same parameters. It is shown in Figs. 4 and 5 that
more order of convergence is obtained by using compact finite difference method
(O(h 3 )) as compared to standard finite difference method (O(h 2 )) for the same
parameters. Also the results for various time steps are compared (Figs. 8 and 9 for
compact finite difference method and Figs. 6 and 7 for standard finite difference
method) and it is observed that the three method converge to identical precision for
larger number of time steps.
References
1. Black, F., Scholes, M.: Pricing of options and corporate liabilities. J. Polit. Econ. 81, 637–654
(1973)
2. Achdou, Y., Pironneau, O.: Computational Methods for Options Pricing. SIAM, Philadelphia
(2005)
3. Carr, P., Madan, D.B., Smith, R.H.: Option valuation using the fast fourier transform. J. Comput.
Finan. 2, 61–73 (1999)
4. Lele, S.K.: Compact finite difference schemes with spectral-like resolution. J. Comput. Phys.
103, 16–42 (1992)
5. Tangman, D.Y., Gopaul, A., Bhuruth, M.: A fast high-order finite difference algorithm for
pricing american options. J. Comput. Appl. Math. 222, 17–29 (2008)
6. Cen, Zhongdi, Le, Anbo: A robust and accurate finite difference method for a generalized
black-scholes equation. J. Comput. Appl. Math. 235, 3728–3733 (2011)
7. During, Bertram, Fournie, Michel, Jungel, Ansgar: Convergence of a high-order compact finite
difference scheme for a non-linear black-scholes equation. Math. Model. Numer. Anal. 38,
359–369 (2004)
8. Hull, J.C.: Options, Futures and Other Derivatives. Prentice Hall, Upper Saddle River (2006)
9. Ugur, O.: An introduction to Mathematical Finance. Imperial College Press, London (2008)
10. Adam, Y.: Highly accurate compact implicit methods and boundary conditions. J. Comput.
Phys. 24, 10–22 (1977)
11. Merton, R.C.: Theory of rational option pricing. Bell J. Econ. 4, 141–183 (1973)
12. Seydel, R.: Tools for Computational Finance, 2nd edn. Springer, Berlin (2004)
13. Tavella, D., Randall, C.: Pricing Financial Instruments: The Finite Difference Method. Wiley,
New York (2000)
14. Willmott, P., Howison, S., Dewynne, J.: Option Pricing: Mathematical Models and Computa-
tion. Oxford Financial Press, Oxford (1993)
15. Goedheer, W.J., Potters, J.H.H.M.: A compact finite difference scheme on a non-equidistant
mesh. J. Comput. Phys. 61, 269–279 (1985)
16. Gamet, L., Ducros, F., Nicoud, F., Poinsot, T.: Compact finite difference schemes on non-
uniform meshes, application to direct numerical simulation of compressible flows. Int. J. Numer.
Meth. Fluids 29, 159–191 (1999)
17. Shukla, R.K., Zhong, X.: Derivation of high-order compact finite difference schemes for non-
uniform grid using polynomial interpolation. J. Comput. Phys. 204, 404–429 (2005)
On Stability of Steady-States
for a Two-Dimensional Network
Model of Ferromagnetic Nanowires
Sharad Dwivedi and Shruti Dubey
Abstract This article concerns with the mathematical study of stability properties
of steady-states for a two-dimensional network model of ferromagnetic nanowires.
We consider the finite network model of ferromagnetic nanowires of semi-infinite
length. We derive a sufficient condition independent of the size of the network under
which the relevant configurations (steady-states) of magnetization are shown to be
asymptotically stable. To be precise, we establish the result under certain condition
on the length between the two consecutive nanowires. We use perturbation technique
and energy method to derive the result.
Keywords Ferromagnetic material · Landau–Lifschitz equation · Stability · Domain

walls · Micromagnetics
AMS Subject Classification: Primary: 35B35 · 34D05 · 35Q60 · Secondary:

35K55
1 Introduction
Experimentally, it has been observed that below a critical temperature, ferromagnetic

materials have a tendency to split up into a small uniformly magnetized regions
called domains separated by a thin transition layer known as domain walls. Over
the period of time, study of formation and motions of domain walls gained a lot of
attention and became one of the most fascinating topic among researchers. It is due
to the fact that the ferromagnetic materials are used on a wide scale in magnetic
storage industry. In particular, ferromagnetic nanowires play a very dominant role
S. Dwivedi (B) · S. Dubey

Department of Mathematics, Indian Institute of Technology Madras,
Chennai 600036, India
e-mail: sharadiitm@gmail.com
S. Dubey
e-mail: sdubey@iitm.ac.in

DOI 10.1007/978-81-322-2485-3_33
406 S. Dwivedi and S. Dubey
in nanoelectronic devices in which the information is encoded as magnetic domains

separated by domain walls along the wire. For example, in case of racetrack memory,
we obtain a three-dimensional storage device by using U-shaped nanowires normal
to the plane of silicon wafer (see [1]). For the rigorous treatment of domains and its
characteristics, we refer the reader to [2] and the references therein.
The model used to describe the magnetic behavior of ferromagnetic material is
called micromagnetism and was introduced by Brown [3]. The evolution of mag-
netization inside the ferromagnetic medium is triggered by the Landau–Lifschitz
equation which is parabolic and nonlinear. The relevant configurations of magne-
tization are minimizers of an energy functional, consisting of several components.
We shall see that these relevant configurations of magnetization coincide with the
steady-states of Landau–Lifschitz equation.
The general framework of the ferromagnetism is as follows. We consider a finite
homogeneous ferromagnetic material which occupies a domain Ω ⊂ R3 . The time-
varying magnetic moment u of a ferromagnetic material is a solution of the Landau–
Lifschitz equation (LLE)
∂u
= −u × He f f − u × u × He f f , (1)
∂t
with the physical saturation constraint
|u(t, .)| = 1 for (t, .) ∈ R+ × R3 a.e., (2)
where the abbreviation a.e. stands for almost everywhere. The total effective field
He f f = −∇E is derived from the micromagnetism energy E given by

A 1
E (u) = |∇u|2 + |Hd (u)|2 − Ha · u, (3)
2 2
Ω R3 Ω
where the term represents the exchange, stray field and external energy contribution
respectively. The constant A > 0 is called the exchange constant. Also, Ha denotes
an applied magnetic field and Hd (u) is the stray field which is characterized by the
Maxwell equations:
⎧
⎪
⎨ curl Hd (u) = 0 in R ,
3
div (Hd (u) + ū) = 0 in R3 ,

⎪
⎩
Hd (u) vanishes at infinity.
where ū is the extension of u in R3 by 0 outside of Ω. We obtain that,
He f f = u + Hd (u) + Ha .
On Stability of Steady-States for a Two-Dimensional Network Model . . . 407
We take the scalar product of (1) with He f f and integrate in time (assuming time
invariant applied field). Using He f f = −∇E , we obtain (see [4, 5])

d
E (u(t)) = − |He f f (u) − He f f (u)) · u u|2 , (4)
dt
Ω
this denotes the dissipation of energy which is mainly due to the second term appears
on the right hand side of (1). Furthermore, steady-states of (1) satisfy u × He f f = 0
in domain Ω, which is exactly the Euler–Lagrange equations of the minimization
problem for (3). Therefore, minimizers of (3), i.e., relevant physical configurations
of the magnetization are nothing but the steady-state solutions of (1) under the
constraint (2).
Existence results of weak solutions for the Landau–Lifschitz equation have been
discussed in [6–8], whereas the strong solutions are considered in [9, 10] and known
to exist locally in time. Numerical aspects of ferromagnetic materials have been
investigated in [11, 12] and the references therein. Stability and controllability results
related with ferromagnetic nanowires are studied in [5, 13, 14]. Higher dimensional
models and network models of such materials can be found in [15, 16].
In the present article, we consider a two-dimensional finite network model of fer-
romagnetic nanowires of semi-infinite length. We assume the relevant configuration
of the magnetization of the network is of the form u ∗ = μe1 where μ = (μi )i∈I
with μi = {−1, +1}. We prove that these relevant configurations are asymptotically
stable in a long time behavior under certain condition on the distance between the
consecutive nanowires. The organization of this article is as follows:
In Sect. 2, we present the schematics of the considered model and introduce the
problem related to the stability of the steady-states in the absence of external mag-
netic field. In Sect. 3, we give the statement of the main result and establish some
preliminary estimates to derive the Theorem.
2 Modeling of a Network Model
In this section, we present a schematics and modeling of a network model under

consideration. We consider a two-dimensional finite network model of ferromag-
netic nanowires of semi-infinite length. In which nanowires are supposed to have
homogeneous geometry and to be placed on the plane (e1 , e2 ), where (e1 , e2 , e3 ) is
the canonical basis of R3 . We represent the distance between the two consecutive
nanowires by > 0. Since, we consider the finite framework of a network model
therefore the index i takes it values in the finite set I = {0, 1, 2, . . . , N }. We denote
the coordinates of a point on the ith nanowire by (xi , i), where 0 ≤ xi < ∞ with
i ∈ I (see Fig. 1). We use the following notations:
Fig. 1 Schema of network model
⎧ I
⎪
⎪ R3 = u = (u i )i∈I , such that ∀ i ∈ I, u i ∈ R3 ,
⎪
⎪
⎪
⎪
⎪
⎨ I
S2 = u = (u i )i∈I ∈ (R3 ) I , such that ∀ i ∈ I, |u i | = 1 ,

⎪
⎪
⎪
⎪
⎪
⎪
⎪ u = sup|u i |, where i ∈ I and | · | is the euclidean norm in R3 .
⎩
i
where S2 represents the unit sphere in R3 .

We assume that the magnetization on each nanowire is constant in the space
variable, i.e., we deal with the ordinary differential model of micromagnetism. This
can be justified by the assumption that the radius of the nanowires are very small
as compared to . We denote u i = u i (t) the magnetization at any point on the ith
nanowire. Therefore, the unknown u = (u 0 , . . . , u N ) is defined as u: R+ → (S2 ) I ,
i.e., u = u(t) = (u i (t))i∈I . Exchange field vanishes due to the aforementioned
assumption renders the only contribution of demagnetizing (stray) field in total effec-
tive field.
We recall that the stray energy is connected with the magnetic field generated by
the medium itself. We calculate the stray field for the entire network in the follow-
ing fashion. On a fixed nanowire say j0 , we represent its stray field as Hd (u)( j0 )
which consist of two parts: the stray field generated on j0 th nanowire by its own
magnetization, i.e., by u j0 , denoted by Hdint (u)( j0 ), and the field generated by
the magnetization of other nanowires, denoted by Hdext (u)( j0 ). We write the stray
field as:
Hd (u)( j0 ) = Hdint (u)( j0 ) + Hdext (u)( j0 ).

The demagnetizing field on j0 th nanowire due to its own magnetization is given by

(see [13, 17])
Hdint (u)( j0 ) = −u 2j0 e2 − u 3j0 e3 . (5)
The stray field generated by the i 0 th nanowire on the j0 th nanowire is given by

(see [18])
∞ ∞
1 u i0 3 (ζ − η)
Hi0 , j0 (u i0 )(ζ ) = − dy + u i · (ζ − η)dy.
4π 0 |ζ − η|3 4π 0 |ζ − η|5 0
with ζ = (x, j0 ) and η = (y, i 0 ), where x and y belongs to [0, ∞).
On calculating the values of these integrals, we obtain:
Hi0 , j0 (u i0 )(ζ ) = (Hi01, j0 , Hi02, j0 , Hi03, j0 ),
where,

1 x2 |i 0 − j0 |2 3 |i 0 − j0 |3
Hi01, j0 = − u +
1
u .
2
4π 2 |i 0 − j0 |2 (x 2 + 2 |i 0 − j0 |2 )3/2 i0 (x 2 + 2 |i 0 − j0 |2 )3/2 i0

1 3 |i 0 − j0 |3
Hi02, j0 = u1
4π 2 |i 0 − j0 |2 (x 2 + 2 |i 0 − j0 |2 )3/2 i0

x
+ 1+ 2 (x 2
+ 2 2
|i 0 − j0 |2
) u i20 .
(x + 2 |i 0 − j0 |2 )3/2

1 x
Hi03, j0 = − 1+ u i30 .
4π 2 |i 0 − j0 |2 (x 2 + 2 |i 0 − j0 |2 )1/2
Therefore, the total network exterior field at the j0 th nanowire is given by:
⎛ ⎞
Ψ 1 (u 1 )( j0 ) + Ψ 2 (u 2 )( j0 )
Hdext (u)( j0 ) = Hi0 , j0 (u(i 0 )) = ⎝ Ψ 2 (u 1 )( j0 ) + Ψ 3 (u 2 )( j0 ) ⎠ , (6)
i 0 = j0 Ψ 4 (u 3 )( j0 )
where (u 1 , u 2 , u 3 ) are the coordinates of u and for k = {1, . . . , 4}, the linear operators
Ψ k : (R3 ) I → (R3 ) I is defined as, for all u = (u i )i∈I in (R3 ) I ,

1 1 x2 | j − j0 |2
Ψ 1 (u)( j0 )(ζ ) = − u( j),
4π2 | j − j0 |2 (x 2 + 2 | j − j0 |2 )3/2
j= j0

1 1 3 | j − j0 |3
Ψ 2 (u)( j0 )(ζ ) = u( j),
4π2 | j − j0 |2 (x 2 + 2 | j − j0 |2 )3/2
j= j0

1 1 x
Ψ 3 (u)( j0 )(ζ ) = 1 + (x 2
+ 2 2
| j − j0 | 2
) u( j),
4π2 | j − j0 |2 (x 2 + 2 | j − j0 |2 )3/2
j= j0

1 1 x
Ψ 4 (u)( j0 )(ζ ) = − 1 + u( j).
4π2 | j − j0 |2 (x 2 + 2 | j − j0 |2 )1/2
j= j0
Hence, we study the following system:
du i
= −u i × (He f f (u))(i) − u i × (u i × (He f f (u))(i)) (7)
dt
He f f (u))(i) = Hdint (u)(i) + Hdext (u)(i)
for i ∈ I and t ∈ R+ with u i : R+ → S2 .

We assume the relevant steady-states configurations of the magnetization distri-
bution as:
u i∗ = μi e1 with μi = {−1, +1} , ∀ i ∈ I. (8)
Experimentally, we relate these relevant configurations to the memory state in a mag-

netic storage device, where μi = 1 corresponds to a bit 1 and μi = −1 corresponds
to a bit 0. Next, we introduce the problem related to stability of the relevant config-
urations under consideration.
Asymptotic stability of any relevant configuration.
In the absence of an external applied field, for any initial conditions in a vicinity
of a given relevant configuration, the solution of the Landau–Lifschitz equation (7)
converges to the relevant configuration.
We give the mathematical statement of the result in the following section:
3 Main Result
In order to state the result, we need to introduce the following notations. We observe
that for ρ ∈ S2 , if 0 < ρ1 < 1 (resp. −1 < ρ1 < 0), the quantity ρ22 + ρ32 exhibits
the distance between ρ and +e1 (resp. −e1 ). We have:
1
|ρ − e1 |2 ≤ ρ22 + ρ32 ≤ |ρ − e1 |2 .
2
For α > 0 sufficiently small, we define D+1 (α) and D−1 (α) by

D+1 (α) = ρ ∈ S2 , ρ1 > 0 and ρ22 + ρ32 < α 2 ,

D−1 (α) = ρ ∈ S2 , ρ1 < 0 and ρ22 + ρ32 < α 2 .
For μ = (μi )i∈I with μi ∈ {−1, +1}, we denote, for α > 0,

Dμ (α) = u ∈ (S2 ) I , ∀ i ∈ I, u i ∈ Dμi (α) . (9)
Our main result about the asymptotic stability of any relevant position is the
following:
Theorem 1 Suppose u is the solution of the Landau–Lifschitz equation (7) with
initial condition u(0) = u init , where u init satisfies the saturation condition (2).
There exists β, a positive constant independent of the size of the network such that if
1
≤ β, (10)
2
then there exist α0 > 0 and κ > 0, such that for all relevant configurations u ∗
(i.e., u i∗ = μi e1 for all i ∈ I ), for all u init ∈ Dμ (α0 ), u satisfies:
u(t) − u ∗ → 0 as t → ∞.
Proof We derive the stability result of a relevant configuration for the Landau–
Lifschitz equation without an external magnetic source. For this we analyze the
following system with unknown u defined as u: R+ → (S2 ) I ,
du
= −u × He f f (u) − u × (u × He f f (u)) (11)
dt
He f f (u) = Hdint (u) + Hdext (u)
The existence and uniqueness of a solution of (11) for any initial condition fol-
lows from the Cauchy–Lipschitz theorem. We assume that u ∗ be a fixed relevant
configuration satisfies the saturation constraint (2), i.e, u ∗ ∈ (S2 ) I such that
u i∗ = μi e1 , with μi ∈ {−1, +1} , ∀ i ∈ I.
Because of the physical saturation constraint (2), we only deal with perturbations u
of u ∗ satisfying:
|u i (t)| = 1, ∀ i ∈ I and ∀ t ≥ 0.
We consider u as a small perturbation of u ∗ and describe it as:
u i = μi e1 + γ (ωi )μi e1 + ωi2 e2 + ωi3 e3 , ∀ i ∈ I (12)

with ωi = ωi2 , ωi3 and γ : (R2 ) I → (R) I is a smooth map defined as γ (ωi ) =

1 − |ωi |2 − 1.
To obtain the transformed system of (11) in new variable ω ∈ C 1 R+ ; (R2 ) I ,
we use the perturbation (12) of u ∗ . We substitute (12) in (11) and take the projection
of the obtained expression along the direction of e2 and e3 .
After a lengthy algebraic computations, it yields that u given by (12) satisfies (11)
if and only if ω = (ω2 , ω3 ) verifies the following system:

dω −1 −μ
= ω + A (μ) + B(ω) + C (ω), (13)
dt μ −1
where
Ψ 2 (μ)
A (μ) = ,
−μΨ 2 (μ)
The linear term B(ω) is given by:

⎛ ⎞
−(ω3 + μω2 )Ψ 1 (μ) + Ψ 2 (μγ )
⎜ +Ψ 3 (ω2 ) + μΨ 4 (ω3 ) ⎟
⎜ ⎟
⎜
B(ω) = ⎜ ⎟,
⎟
⎝ (ω2 − μω3 )Ψ 1 (μ) − μΨ 2 (μγ ) ⎠
−μγ Ψ 2 (μ) − μΨ 3 (ω2 ) + Ψ 4 (ω3 )
The nonlinear term C (ω) is given by:

⎛ ⎞
−μγ ω3 − (ω3 + μω2 + μγ ω2 )Ψ 1 (μγ )
⎜ −μγ ω2 Ψ 1 (μ) − (ω3 + μω2 + μγ ω2 )Ψ 2 (ω2 ) ⎟
⎜ ⎟
⎜ −(Ψ 2 (μ) + Ψ 2 (μγ ))(ω2 )2 − (ω2 )2 Ψ 3 (ω2 ) ⎟
⎜ ⎟
⎜ +(μγ − ω2 ω3 )Ψ 4 (ω3 ) + ((ω2 )2 + (ω3 )2 )ω2 ⎟
⎜ ⎟
⎜ ⎟
C (ω) = ⎜
⎜
⎟.
⎟
⎜ μγ ω + (ω − μω − μγ ω )Ψ (μγ )
2 2 3 3 1
⎟
⎜ −μγ ω3 Ψ 1 (μ) + (ω2 − μω3 − μγ ω3 )Ψ 2 (ω2 ) ⎟
⎜ ⎟
⎜ −(Ψ 2 (μ) + Ψ 2 (μγ ))ω2 ω3 − μγ Ψ 2 (μγ ) ⎟
⎜ ⎟
⎝ −(μγ + ω2 ω3 )Ψ 3 (ω2 ) − (ω3 )2 Ψ 4 (ω3 ) ⎠
+((ω ) + (ω ) )ω
2 2 3 2 3
Our objective is to analyze the stability behavior of a relevant configuration u ∗ for

LLE (11). Evidently, both the forms of Landau–Lifschitz equation, (11) and (13) are
equivalent and the stability of zero solution for (13) renders the stability of u ∗ for
(11). We state this in the following Proposition.

Proposition 1 Let u ∈ C 1 R+ ; (S2 ) I with |u| = 1 and verifies (11). Let ω ∈

C 1 R+ ; (R2 ) I defined by:
u = μe1 + γ (ω)μe1 + ω2 e2 + ω3 e3
Then u is a solution to Landau–Lifschitz equation (11) if and only if ω is a solution to

(13). Moreover, u ∗ is asymptotically stable for (11) if and only if 0 is asymptotically
stable for (13).
Proof We follow the similar technique used in partial differential equation frame-
work in [13–15]. It is apparent that by taking the projection on both e2 and e3 axis, if
u satisfies (11) then ω verifies (13). For the converse part, we write (11) on the form
du
= F (u).
dt
Furthermore u · F (u) = 0. Since ω satisfies (13), we have

du
− F (u) · ek = 0, ∀ k ∈ {2, 3} .
dt
du
Using the constraint |u| = 1, which renders u · = 0. we obtain
dt

du
μ (1 + γ ) − F (u) · e1 = 0,
dt
with μ = 0 and γ = −1, implies that u satisfies (11). This completes the proof of
Proposition 1.
Now we study the stability of zero solution for the transformed Landau–Lifschitz
equation (13). First, we establish some preliminary estimates. We estimate the
linear operators Ψ k : (R3 ) I → (R3 ) I in the following fashion, we obtain for all
k = {1, . . . , 4}
⎛ ⎞
K 1 ⎝
1
Ψ k (u) ≤ ⎠ u , ∀ u = (u i )i∈I ∈ (R3 ) I . (14)
π 2 | j|2
j=0
Using (14), the operators A , B and C appear on the right hand side of (13) are
estimated with straightforward arguments in the following lemmas.
Lemma 1 There exists a constant K 2 such that, for all ω ∈ (R3 ) I with ω < 1,
we have
K2
B(ω) ≤ 2 ω .

1
Lemma 2 We assume that ≤ 1. There exist constants K 3 and K 4 such that, for
2
all ω ∈ (R ) with ω < 1, we have
3 I
A (μ) ≤ K 3 and C (ω) ≤ K 4 ω 2 .
It is worth to mention that the constants K 2 , K 3 and K 4 neither depend on nor

notice that γ (ω) ≤ ω whenever ω < 1.
on the size of the network. We
We have ω ∈ C 1 R+ ; (R2 ) I with,
1
2
|ωi (t)| = (ωi2 (t))2 + (ωi3 (t))2
We notice that u ∈ Dμ (α) if and only if |ω| <α (see (9)).
Taking the inner product of (13) with ωi2 , ωi3 , we obtain, for all i ∈ I ,

2 d 2 3 d 3
ωi ωi + ωi ωi + (ωi2 )2 + (ωi3 )2 = ((A (μ))i + (B(ω))i
dt dt

+ (C (ω))i ) · ωi2 , ωi3 ,
Using Lemmas 1 and 2, we have,
1 d K2
|ωi |2 + |ωi |2 ≤ K 3 ω + 2 ω 2 + K 4 ω 3 . (15)
2 dt
We define β by,
1
β= (16)
K2
Our goal is to show that zero solution is asymptotically stable for (13). We set the
1
distance between the nanowires in such a way so that 2 remains less than β.

Multiplying (15) by e2t and integrate from 0 to t. We get, for all i ∈ I ,
t
|ωi (t)|2 e2t ≤ ω(0) 2 + 2K 3 ω(v) e2v dv
0
t
K2 t
+2 2 ω(v) 2 e2v dv + 2K 4 ω(v) 3 e2v dv
0 0
We take the supremum on i ∈ I and obtain that

t
ω(t) e2 2t
≤ ω(0) +2K 3
2
ω(v) e2v dv
0
t
K2 t
+2 2 ω(v) e dv + 2K 4
2 2v
ω(v) 3 e2v dv
0 0
K2 1
We denote κ = 1 − . Equation (16) together with condition 2 < β implies
2
κ −2v κ
κ > 0. Now while ω(v) ≤ e ≤ , we have
2K 4 2K 4
t
K3
ω(t) e 2 2t
≤ ω(0) + 2
κt + (2 − κ) ω(v) 2 e2v dv, ∀ t ≥ 0.
K4 0
κ
Using Gronwall lemma, while ω(v) ≤ , we obtain
2K 4

K3
ω(t) ≤ ω(0) +
2 2
κt e−κt
K4
It is evident that the term te−κt → 0 as t → ∞. Therefore, whenever ω(0) ≤

κ
, we obtain
2K 4
ω(t) − 0 → 0 as t → ∞.
κ
We set α0 = , and it shows that zero solution is asymptotically stable for
2K 4
the perturbed LLE (13) which in turn reflect the asymptotic behavior of relevant
configurations for LLE (11). This completes the proof of Theorem 1.
References
1. Parkin, S.S.P., Hayashi, M., Thomos, L.: Magnetic domain-wall racetrack memory. Science
320, 190–194 (2008)
2. Hubert, A., Schäfer, R.: Magnetic Domains: The Analysis of Magnetic Microstructures.
Springer, Berlin (1998)
3. Brown, W.F.: Micromagnetics. Wiley, New York (1963)
4. Alouges, F.: Mathematical models in micromagnetism: An introduction. ESAIM Proc. 22,
114–117 (2007)
5. Labbé, S., Privat, Y., Trélat, E.: Stability properties of steady-states for a network of ferromag-
netic nanowires. J. Diff. Eqn. 253, 1709–1728 (2012)
6. Alouges, F., Soyeur, A.: On global weak solutions for Landau-Lifshitz equations: Existence
and nonuniqueness. Nonlinear Anal. Theory Meth. Appl. 18, 1071–1084 (1992)
7. Carbou, G., Fabrie, P.: Time average in micromagnetism. J. Differ. Eqn. 147, 383–409 (1998)
8. Visintin, A.: On Landau Lifschitz equation for ferromagnetism. Japan J. Appl. Math. 1, 69–84
(1985)
9. Carbou, G., Fabrie, P.: Regular solutions for Landau-Lifschitz equation in a bounded domain.
Differ. Integral Eqn. 14, 213–229 (2001)
10. Carbou, G., Fabrie, P.: Regular solutions for Landau-Lifschitz equation in R3 . Commun. Appl.
Anal. 5, 17–30 (2001)
11. Ban̆as, L., Bartels, S., Prohl, A.: A convergent implicit finite element discretization of the
Maxwell-Landau-Lifshitz-Gilbert equation, SIAM J. Numer. Anal. 46, 1399–1422 (2008)
12. Labbé, S.: Fast computation for large magnetostatic systems adapted for micromagnetism.
SIAM J. Sci. Comput. 26, 2160–2175 (2005)
13. Carbou, G., Labbé, S.: Stability for static walls in ferromagnetic nanowires. Discrete Contin.
Dynam. Syst. Ser. B 6, 273–290 (2006)
14. Carbou, G., Labbé, S., Trélat, E.: Control of travelling walls in a ferromagnetic nanowire.
Discrete Contin. Dynam. Systems Ser. S 1, 51–59 (2008)
15. Carbou, G.: Stability of static walls for a three-dimensional model of ferromagnetic material.
J. Math. Pures Appl. 93, 183–203 (2010)
16. Agarwal, S., Carbou, G., Labbé, S., Prieur, C.: Control of a network of magnetic ellipsoidal
samples. Math. Control Relat. Fields 1(2), 129–147 (2011)
17. Sanchez, D.: Behaviour of the LandauLifschitz equation in a ferromagnetic wire. Math. Meth.
Appl. Sci. 32(2), 167–205 (2009)
18. Griffiths, D.J.: Introduction to Electrodynamics, 3rd edn. PearsonBenjamin Cummings, San
Francisco (2008)
Fractional Functional Impulsive Differential
Equation with Integral Boundary Condition
Vidushi Gupta and Jaydev Dabas
Abstract In this article, we discuss the existence and uniqueness of solution for
fractional order differential equation with integral boundary condition and fractional
impulsive conditions. In our problem delay also include with finite domain. Some
important fixed point theorems are the main tools to establish the existence and
uniqueness results for the solution of the problem.
Keywords Fractional order differential equation · Impulsive conditions · Boundary

value problem · Fixed point theorems
1 Introduction
The present work is related to study the existence and uniqueness of solution for
impulsive fractional differential equations with some special boundary conditions
given as follows:
c
Dtα x(t) = f (t, x(t), x(t − τ )), t ∈ [0, T ], t = tk (1)
Δx(tk ) = Ik (x(tk− )), k = 1, 2, . . . , m, (2)
Δ(c D q x(tk− )) = Jk (x(tk− )), q ∈ (0, 1), k = 1, 2, . . . , m, (3)
x(t) = φ(t), t ∈ [−τ, 0] (4)
T
x(0) − x(T ) = p(x(s))ds, (5)
0
V. Gupta (B) · J. Dabas

Saharanpur Campus, 247001 Saharanpur, India
e-mail: rinkygupta00@gmail.com
J. Dabas

DOI 10.1007/978-81-322-2485-3_34
418 V. Gupta and J. Dabas
where c Dtα is Caputo’s derivative and α ∈ (1, 2), T < ∞. The function f : [0, T ]
× X × PC0 → X and p: X → X are given continuous functions and satisfied some
assumptions. The Eq. (2)–(3) are impulsive conditions having some properties, with
0 = t0 < t1 < · · · < tm < tm+1 = T, Ik , Jk ∈ C(X, X ), (k = 1, 2, . . . , m),
are bounded functions. We have Δx(tk ) = x(tk+ ) − x(tk− ) and Δ(c D q x(tk )) =
(c D q x(tk+ )) − (c D q x(tk− )), x(tk+ ) = lim h→0 x(tk + h) and x(tk− ) = limh→0 x(tk −
h) represents the right and left-hand limits of x(t) at t = tk respectively with x(ti− ) =
x(ti ).
Recently, the study of differential equations of the type of non-integer order has
been an important tool in the area of research in mathematics. Its useful applica-
tions included mathematical modeling in many engineering and science discipline
like physics, chemistry, biophysics, biology, etc. Its nonlocal behavior is the vital
characteristic that makes it vary from its rival in classical calculus. For more details
one can refer the books [1–3] and papers [4, 5] and the references therein.
In recent years, the theory of impulsive differential equations for integer order
comes in various applications of mathematical modeling of phenomena and practical
situations. For instance, the impulsive differential equations captured from real-world
problems describe the dynamics of processes in which sudden, discontinuous jumps
occurs. For more details one can see the papers [4, 7–9] and references therein.
Problems with integral boundary conditions arise naturally in thermal conduction
problems, semiconductor problems, hydrodynamic problems, etc. However, periodic
boundary value problems with impulsive fractional evolution equations have not been
studied extensively. For more details one can see the papers [6, 7, 10–14, 20] and
the references therein.
Differential equations with time delay are often used to model phenomena in
economics, biology, medicine, ecology, and other fields of sciences. They take into
account that in many applications, some time elapses between causes and their effects.
These concerns, for instance, investments in economics and finance, typically yield-
ing returns only after some time lag. Delay reaction also transpire in population
dynamics, where individuals always need some time to mature, or in medicine,
where contagious diseases have cross-infection. Presently, many authors [12, 15–
20] are currently working on field of fractional delay differential equation with finite
domain.
In [21] Ravichadran et al. proved the existence and uniqueness of mild solutions
for the following impulsive fractional functional differential equations of the form:
t
α
D x(t) = Ax(t) + f (t, xt , h(t, s, xs )ds), t ∈ [0, T ], t = tk ,
0
x(t) = φ(t), t ∈ [−d, 0], Δx(tk ) = Q k (x(tk− )), k = 1, 2, . . . , m,
Authors established their results by using some fixed point theorems.

Fractional Functional Impulsive Differential Equation . . . 419
In [20] Dabas et al. considered the following impulsive neutral fractional integro-
differential equation with state-dependent delay subject to integral boundary condi-
tion
t
(t − s)α−1
Dtα x(t) + g(s, xρ(s,xs ) )ds = f (t, xρ(t,xt ) , B(x)(t)), t ∈ [0, T ], t = tk
0 Γ (α)
Δx(tk ) = Ik (x(tk− )), Δx (tk ) = Q k (x(tk− )), k = 1, 2, . . . , m,
T
x(t) = φ(t), t ∈ [−d, 0], ax (0) + bx (T ) = q(x(s))ds.
0
The existence results are proved by applying the classical fixed point theorems.
In [8] Xi Fu et al. concerned with the fractional separated boundary value prob-
lem with fractional impulsive conditions. By using the Schaefer fixed point theo-
rem, Banach fixed point theorem, and nonlinear alternative of Leray–Schauder type
authors obtained the existence results. Chouhan et al. [12] studied the solution for the
system with infinite delay and by using Banach contraction and Krasnoselkii fixed
point theorems authors established the existence and uniqueness results. In [14]
Yu et al. concerned periodic fractional impulsive boundary value problem. The exis-
tence and boundedness of piecewise continuous mild solutions and design parameter
drift for periodic motion of linear problems are presented. Furthermore, the authors’
established existence results of piecewise continuous mild solutions for semilinear
impulsive periodic problems are shown by using the Schauder’s fixed point theorem.
To the best of the authors’ knowledge, no one has studied the existence and
uniqueness of solutions for fractional boundary value problem (1)–(5) by using the
Caputo derivative. The purpose of this paper is to fill in this gap. The organization
of the paper is as follows. In Sect. 2, we give some basic preliminaries concerning
the fractional integral, fractional derivative, and fixed point theorems. In Sect. 3, we
present our main results.
2 Preliminary
Let (X, · X ) be a complex Banach space. In order to defined the solution of

the problem (1), we consider the following space: PC T = PC([−τ , T ]; X ) =
{φ: [−τ , T ] → X : φ(t) is continuous every where except for a finite number of
points ti , i = 1, 2, . . . , m, at which φ(ti+ ) and φ(ti− ) exist and φ(ti ) = φ(ti− )}. The
space PCt = PC([−τ , t]: X ) for t ∈ [−τ, T ] is the Banach space of all continuous
functions except for a finite number of points ti , i = 1, 2, . . . , m, at which φ(ti+ )
and φ(ti− ) exist and φ(ti ) = φ(ti− ) endowed with the norm
φt = sup {φ(s) X , φ ∈ PCt },

−τ ≤s≤t
where · X is the norm in X . For delay we consider the Banach space PC0 =
C([−τ , 0]: X ) endowed with the above sup-norm.
Definition 1 ([4]) The fractional integral of order α with lower limit zero for a
function f : [0, ∞) → R of order is defined as
t (t − s)α−1
Itα f (t) = f (s)ds, t > 0, α > 0, (6)
0 Γ (α)
provided the right side is pointwise defined on [0, ∞), where Γ is the gamma func-
tion.
Definition 2 ([4]) The Riemann–Liouville derivative of order α with the lower limit
zero for a function f : [0, ∞) → R can be written as
d n t
1
L
Dtα f (t) = (t − s)n−α−1 f (s)ds, t > 0, n − 1 < α < n. (7)
Γ (n − α) dt 0
Definition 3 ([4]) The Caputo’s derivative of order α for a function f : [0, ∞) → R

can be written as
t (k)
n−1 k
c
Dtα f (t) = L Dtα f (t) − f (0) , t > 0, n − 1 < α < n. (8)
k!
k=0
Remark 1 ([4]) If f (t) ∈ C n [0, ∞), for order n − 1 < α < n then

1 t f (n) (s)
c
Dtα f (t) = ds = Itn−α f (n) (t), t > 0. (9)
Γ (n − α) 0 (t − s)α+1−n
The Caputo derivative of constant is equal to zero.
Lemma 1 ([5]) Let α > 0, then the differential equation

c
D α h(t) = 0 (10)
has solutions h(t) = c0 + c1 t + c2 t 2 + · · · + cn−1 t n−1 and I α D α h(t) = h(t) + c0 +

c1 t + c2 t 2 + · · · + cn−1 t n−1 , where ci ∈ R, i = 0, 1, . . . , n − 1, n = [α] + 1.
Lemma 2 Let α ∈ (1, 2) and f : [0, T ] × X × PC0 → X be continuously differen-

tiable function. A piecewise continuous differential function x(t): (−τ , T ] → X is
a solution of system (1)–(5) on [0, T ] iff x(t) satisfied following integral equation,
⎧
⎪
⎪φ(t), t ∈ [−τ, 0],
⎪
⎪ t (t−s)α−1
⎪
⎪
⎪ 0 Γ (α) f (s, x(s), x(s − τ ))ds
⎪
+ φ(0)

⎪
⎪ T
⎪
⎪− t
p(x(s))ds + T m Γ (2−q)
J (x(t −
))
⎪
⎪ T 0 i=1 ti 1−q i i
⎪
⎪ T (T −s)α−1
⎪
⎪+ f (s, x(s), x(s − τ ))ds , t ∈ [0, t1 ),
⎪
⎨ 0 Γ (α)
x(t) = . . .
⎪
⎪ t (t−s)α−1 k
⎪
⎪
⎪ 0 Γ (α) f (s, x(s), x(s − τ ))ds + i=1 Ii (x(ti− )) + φ(0)
⎪
⎪
⎪ T m Γ (2−q) −
⎪− Tt 0 p(x(s))ds + T i=1
⎪ Ji (x(t ))
⎪
⎪ ti
1−q i
⎪ T (T −s)α−1
⎪
⎪
⎪+ Γ (α) f (s, x(s), x(s − τ ))ds
⎪
⎪
0
⎪
⎩+ k (t − ti ) Γ (2−q) Ji (x(t − )) ,
i=1 t 1−q i t ∈ (tk , tk+1 ].
i
Proof For t ∈ [0, t1 ], then
Dtα x(t) = f (t, x(t), x(t − τ )), (11)

x(t) = φ(t), t ∈ [−τ, 0].
Taking the Riemann–Liouville fractional integral on (11) and using the Lemma (1),
then
t (t − s)α−1
x(t) = f (s, x(s), x(s − τ ))ds − c0 − c1 t, (12)
0 Γ (α)
using the condition x(0) = φ(0) we compute c0 = −φ(0), then we have

t (t − s)α−1
x(t) = f (s, x(s), x(s − τ ))ds + φ(0) − c1 t. (13)
0 Γ (α)
Same way for t ∈ (t1 , t2 ], then

Dtα x(t) = f (t, x(t), x(t − τ )),
(14)
Δx(t1 ) = I1 (x(t1− )), Δ(c D q x(t1− )) = J1 (x(t1− )),
Again applying the Riemann–Liouville fractional integral operator and using the
Lemma (1), then
t (t − s)α−1
x(t) = f (s, x(s), x(s − τ ))ds − c2 − c3 t, (15)
0 Γ (α)
on applying first impulsive condition Δx(t1 ) = I1 (x(t1− )), we get
− c2 = I1 (x(t1− )) + c3 t1 + φ(0) − c1 t1 . (16)

Substituting the value of c2 in (15), we obtain

(t − s)α−1
t
x(t) = f (s, x(s), x(s − τ ))ds + I1 (x(t1− ))
0 Γ (α)
+φ(0) − c1 t1 + c3 (t1 − t). (17)
From (17) and (13), we get

t
1 t 1−q
D q x(t) = (t − s)α−q−1 f (s, x(s), x(s − τ ))ds − c3 , (18)
Γ (α − q) 0 Γ (2 − q)
t
1 t 1−q
D x(t) =
q
(t − s)α−q−1 f (s, x(s), x(s − τ ))ds − c1 . (19)
Γ (α − q) 0 Γ (2 − q)
Using the second impulsive condition Δ(D q x(t1 )) = J1 (x(t1− )), then we have
Γ (2 − q)
c3 = − J1 (x(t1− )) + c1 . (20)
t1 1−q
On putting the value of c3 in (17), we prevail

(t − s)α−1
t
x(t) = f (s, x(s), x(s − τ ))ds + I1 (x(t1− ))
0 Γ (α)
Γ (2 − q)
+ φ(0) + (t − t1 ) J1 (x(t1− )) − c1 t. (21)
t1 1−q
Similarly for t ∈ (tk , tk+1 ], then
k
t (t − s)α−1
x(t) = f (s, x(s), x(s − τ ))ds + Ii (x(ti− )) + φ(0) − c1 t
0 Γ (α)
i=1

k
Γ (2 − q)
+ (t − ti ) Ji (x(ti− )) . (22)
ti 1−q
i=1
T
Now using the boundary condition x(0) − x(T ) = 0 p(x(s))ds, we compute the
following value of the constant c1 given as:
m
1 T Γ (2 − q)
c1 = p(x(s))ds + T Ji (x(ti− ))
T 0 ti 1−q
i=1

T (T − s)α−1
+ f (s, x(s), x(s − τ ))ds . (23)
0 Γ (α)
So that for t ∈ (tk , tk+1 ], then we get
k
t(t − s)α−1
x(t) = f (s, x(s), x(s − τ ))ds + Ii (x(ti− )) + φ(0)
0 Γ (α)
i=1
T m
t Γ (2 − q) −
− p(x(s))ds + T J i (x(t i ))
T 0 ti 1−q
i=1
T
(T − s)α−1
+ f (s, x(s), x(s − τ ))ds
0 Γ (α)
k
Γ (2 − q) −
+ (t − ti ) Ji (x(ti )) . (24)
ti 1−q
i=1
This is complete proof.
3 Existence and Uniqueness Results
Our first result is based on Banach contraction principle.

Theorem 1 Suppose the following conditions holds:
1. There exist the positive constants L f 1 , L f 2 , L p such that
f (t, x, ψ) − f (t, y, χ ) X ≤ L f 2 x − y X + L f 1 ψ − χ X ,
p(x) − p(y) X ≤ L p x − y X , t ∈ [0, T ], ∀ψ, χ , x, y ∈ X.
2. The bounded continuous function Ik , Jk satisfied the condition for constant

L I , L J such that
Ik (x) − Ik (y) X ≤ L I x − y X , Jk (x) − Jk (y) X ≤ L J x − y X , ∀x, y ∈ X.

2T α
3. And assume that Γ (α+1) (L f 1 + L f2 ) + m L I + T L p + 2T q mΓ (2 − q)L J < 1.
Then system (1)–(5) has a unique solution.
Proof Let the space PC T is closed convex set, invested with the uniform topology
and the operator P : PC T → PC T is defined by
⎧
⎪
⎪ φ(t) t ∈ [−τ, 0],
⎪
⎪ t (t−s)α−1
⎪
⎪
⎪ 0 Γ (α) f (s, x(s), x(s − τ ))ds
⎪
+ φ(0)

⎪
⎪ T
⎪
⎪ − t
p(x(s))ds + T m Γ (2−q)
J (x(t −
))
⎪
⎪ T 0 i=1 ti 1−q i i
⎪
⎪ T (T −s)α−1
⎪
⎪ + f (s, x(s), x(s − τ ))ds , t ∈ [0, t1 ),
⎪
⎨ 0 Γ (α)
P x(t) = . . . , (25)
⎪
⎪ t (t−s)α−1 k
⎪
⎪
⎪ 0 Γ (α) f (s, x(s), x(s − τ ))ds + i=1 Ii (x(ti− )) + φ(0)
⎪
⎪
⎪ T m Γ (2−q) −
⎪− Tt 0 p(x(s))ds + T i=1
⎪ Ji (x(t ))
⎪
⎪ ti
1−q i
⎪ T (T −s)α−1
⎪
⎪
⎪ + Γ (α) f (s, x(s), x(s − τ ))ds
⎪
⎪
0
⎪
⎩+ k (t − ti ) Γ (2−q) Ji (x(t − )) ,
i=1 t 1−q i t ∈ (tk , tk+1 ].
i
Let x, x ∗ ∈ PC T and t ∈ [0, t1 ).

(t − s)α−1
t
P(x) − P(x ∗ ) X ≤ f (s, x(s), x(s − τ )) − f (s, x ∗ (s), x ∗ (s − τ )) X ds
0 Γ (α)
m
|t| T Γ (2 − q) − ∗ −
+ p(x(s)) − p(x ∗ (s)) X ds + T J i (x(ti )) − Ji (x (ti )) X
T 0 i=1
|ti |1−q
T
(T − s)α−1
+ f (s, x(s), x(s − τ )) − f (s, x ∗ (s), x ∗ (s − τ )) X ds
0 Γ (α)
Using the given conditions, we get

2T α
P(x) − P(x ∗ ) PC T ≤ (L f1 + L f2 ) + L p T + T q mΓ (2 − q)L J x − x ∗ PC T .
Γ (α + 1)
For t ∈ (tk , tk+1 ], we have
P(x) − P(x ∗ ) X ≤
t
(t − s)α−1
f (s, x(s), x(s − τ )) − f (s, x ∗ (s), x ∗ (s − τ )) X ds
0 Γ (α)

|t| T
k
+ Ii (x(ti− )) − Ii (x ∗ (ti− )) X + p(x(s)) − p(x ∗ (s)) X ds
T 0
i=1

Γ (2 − q)
m
− ∗ −
+T Ji (x(ti )) − Ji (x (ti )) X
i=1
|ti |1−q
T
(T − s)α−1
0 Γ (α)
k
Γ (2 − q) − ∗ −
+ (|t − ti |) J i (x(ti )) − Ji (x (ti )) X
i=1
|ti |1−q
Again in the same way using the given assumptions of the Theorem (1), we have
P(x) − P(x ∗ ) PC T
2T α
≤ (L f1 + L f2 ) + m L I + T L p + 2T q mΓ (2 − q)L J x − x ∗ PC T
Γ (α + 1)
≤ Δx − x ∗ PC T .
Since Δ < 1, implies that the map P is a contraction map and has a unique fixed point
x ∈ PC T , which is a solution of the system (1)–(5) on [−τ , T ]. This is complete
proof of theorem.
Our second result is based on Krasnoselkii’s fixed point theorem [20].

Theorem 2 Let Ik , Jk and p be the continuous functions and satisfy
Ik (x) X ≤ C1 , Jk (x) X ≤ C2 , p(x) X ≤ C3 , C1 > 0, C2 > 0, C3 > 0.
Further, f : [0, T ] × X × PC0 → X is continuous function for every t ∈ [0, T ],

and satisfy the condition
f (t, x, ψ) − f (t, y, χ ) X ≤ L f 2 x − y X + L f 1 ψ − χ X , L f 1 , L f 2 > 0.
Then system (1)–(5) has at least one solution.

α
Proof Let r ≥ [mC1 + φ(0) + 2T q mC2 + Γ (α+1) 2T
(L f1 + L f2 )r + C3 T ].
Consider the space PC T = {x ∈ PC T : x PC T ≤ r }, then PC Tr is a bounded,
r
closed convex subset in PC T . For t ∈ (tk , tk+1 ], the operators N : PC Tr → PC Tr

and P : PC Tr → PC Tr are defined as

k m

Γ (2 − q)
N (x) = Ii (x(ti− )) + φ(0) − t Ji (x(ti− ))
ti 1−q
i=1 i=1

k
Γ (2 − q)
+ (t − ti ) Ji (x(ti− )) (26)
ti 1−q
i=1

− s)α−1
t (t t T
P(x) = f (s, x(s), x(s − τ ))ds − p(x(s))ds
0 Γ (α) T 0
T
(T − s)α−1
+ f (s, x(s), x(s − τ ))ds . (27)
0 Γ (α)
Now the proof of the Theorem (2) is given in form of following steps:
Step 1. Let x, x ∗ ∈ PC Tr then
N (x) + P(x ∗ ) X

k m

Γ (2 − q)
≤ Ii (x(ti− )) X + φ(0) + |t| Ji (x(ti− )) X
|ti |1−q
i=1 i=1

k
Γ (2 − q)
+ (|t − ti |) Ji (x(ti− )) X
i=1
|ti |1−q
t(t − s)α−1
+ f (s, x ∗ (s), x ∗ (s − τ )) X ds
0 Γ (α)
T
|t| T ∗ (T − s)α−1
+ p(x (s)) X ds + f (s, x ∗ (s), x ∗ (s − τ )) X ds
T 0 0 Γ (α)
2T α
≤ mC1 + φ(0) + 2T q mC2 + (L f1 + L f2 )r + C3 T
Γ (α + 1)
≤ r.
Which implies that N (x)+ P(x ∗ ) X ≤ r. which means that N (x)+ P(x ∗ ) ∈ PC Tr .
Step 2. Let xn → x be sequence in PC Tr then,
N (xn ) − N (x) X

k
≤ Ii (xn (ti− )) − Ii (x(ti− )) X + φ(0)
i=1
m

Γ (2 − q)
+ |t| Ji (xn (ti− )) − Ji (x(ti− )) X
i=1
|ti |1−q
k
Γ (2 − q) − −
+ (|t − ti |) Ji (xn (ti )) − Ji (x(ti )) X .
i=1
|ti |1−q
Since the functions Ik and Jk , k = 1, 2, . . . , m, are continuous, hence
N (xn ) − N (x) → 0.
Which implies that the mapping N is continuous on PC Tr .

Step 3. For each t ∈ (tk , tk+1 ], k = 0, 1, . . . , m for each x ∈ PC Tr , we have
k m
− Γ (2 − q) −
N (x) X ≤ Ii (x(ti )) X + φ(0) + |t| Ji (x(ti )) X
i=1 i=1
|ti |1−q
k
Γ (2 − q) −
+ |(t − ti )| J i (x(t i ))
i=1
|ti |1−q
≤ mC1 + φ(0) + 2T q mC2 .
Which implies that the mapping N is uniformly bounded.

Step 4. Now to show that N is equicontinuous. Let l1 , l2 ∈ (tk , tk+1 ], tk ≤ l1 <
l2 ≤ tk+1 , k = 1, 2, . . . , m, x ∈ PC Tr , we have
N (x)(l2 ) − N (x)(l1 ) X
m k
Γ (2 − q) − Γ (2 − q) −
≤ (l2 − l1 ) Ji (x(ti )) X + (l 2 − l 1 ) Ji (x(ti )) X
i=1
|ti |1−q i=1
|ti |1−q
As l2 → l1 , then N (x)(l2 )− N (x)(l1 ) → 0. Which show that N is equicontinuous

map. Combing Step 2 to Step 4 together with the Ascol’s theorem, we conclude that
the operator N is a compact.
Step 5. Now we show that P is a contraction mapping. Let x, x ∗ ∈ PC Tr and
t ∈ (tk , tk+1 ], k = 1, 2, . . . , m, we have
P(x) − P(x ∗ ) X
t
(t − s)α−1
≤ f (s, x(s), x(s − τ )) − f (s, x ∗ (s), x ∗ (s − τ )) X ds
0 Γ (α)

|t| T
+ p(x(s)) − p(x ∗ (s)) X ds
T 0
T
(T − s)α−1
0 Γ (α)
2T α
P(x) − P(x ∗ ) PC Tr ≤ (L f1 − L f2 ) + T L p x − x ∗ PC Tr
Γ (α + 1)
As Δ < 1, therefore P is a contraction map. Thus all the assumptions of the Kras-
noselkii’s theorem are satisfied, which implies that the system (1)–(5) has at least
one solution on (τ, T ]. This completes the proof of the theorem.
References
Equations. In: North-Holland Mathematical Studies, vol. 204. Elsevier, Amsterdam (2006)
Equations. Wiley, New York (1993)
3. Podlubny, I.: Fractional Differential Equations. Academic Press, New York (1999)
4. Feckan, M., Zhou, Y., Wang, J.R.: On the concept and existence of solution for impulsive solu-
tion for impulsive fractional differential equations. Commun. Nonlinear Sci. Numer. Simulat.
17, 3050–3060 (2012)
5. Agarwal, R.P., Benchohra, M., Hamani, S.: A survey on existence results for boundary value
problems of nonlinear fractional differential equations and inclusions. Acta Applicandae Math-
ematicae 109, 973–1033 (2010)
6. Fu, X., Liu, X.: Existence results for fractional differential equations with separated boundary
conditions and fractional impulsive conditions. Abstr. Appl. Anal. Article ID 785078, 9 (2013)
7. Ahmad, B., Sivasundaram, S.: Existence of solutions for impulsive integral boundary value
problems of fractional order. Nonlinear Anal. Hybrid Syst. 4, 134–141 (2010)
8. Guo, T.L., Jiang, W.: Impulsive fractional functional differential equations. Comput. Math.
Appl. 64(10), 3414–3424 (2012)
9. Feckan, M., Zhou, Y., Wang, J.: Response to "Comments on the concept of existence of solution
of existence of solution for impulsive fractional differential equations [Commun Nonlinear Sci
Numer Simul 2014;19:401–3.]”, Commun. Nonlinear Sci. Numer. Simulat. 19, 4213–4215
(2014)
10. Ahmad, B., Nieto, J.J.: Existence results for nonlinear boundary value problems of fractional
integrodifferential equations with integral boundary conditions. Bound. Value Probl. 10(Article
ID 708576), 11 (2009)
11. Boucherif, A.: Second-order boundary value problems with integral boundary conditions. Non-
linear Anal. 70, 364–371 (2009)
12. Chauhan, A., Dabas, J.: Integral boundary-value problem for impulsive fractional functional
integrodifferential equation with infinite delay. Electron. J. Differ. Equ. 2012, 1–13 (2012)
13. Wang, J., Zhou, Y., Feckan, M.: On recent developments in the theory of boundary value
problems for impulsive fractional differential equations. Comp. Math. Appl. 64, 3008–3020
(2012)
14. Yu, X., Wang, J.: Periodic BVPs for fractional order impulsive evolution equations. Bound.
Value Probl. (2014)
15. Agarwal, R.P., Andrade, B.D.: On fractional integro-differential equations with state-dependent
delay. Comput. Math. Appl. 62, 1143–1149 (2011)
16. Belmekki, M., Mekhalfi, K., Ntouyas, S.K.: Semilinear functional differential equations with
fractional order and finite delay. Malaya J. Matematika 1(1), 73–81 (2012)
17. Benchohra, M., Berhoun, F.: Impulsive fractional differential equations with state dependent
delay. Commun. Appl. Anal. 14(2), 213–224 (2010)
18. Benchohra, M., Litimein, S., NGuerekata, G.: On fractional integro-differential inclusions with
state-dependent delay in Banach spaces. Appl. Anal. 1–16 (2011)
19. Carvalho, J.P., Santos, D., Cuevas, C., Andrade, B.D.: Existence Results for a fractional equation
with state-dependent delay. Adv. Differ. Equ. Article ID 642013, (2011)
20. Dabas, J., Gautam, G.R.: Impulsive neutral fractional integro-differential equations with state
dependent delays and integral condition. Electron. J. Differ. Equ. 273, 1–13 (2013)
21. Ravicharndran, C., Arjunan, M.M.: Existence result for impulsive fractional semilinear func-
tional integrodifferential equations in Banach spaces. J. Fractional Calc. Appl. 1–11, 3 (2012)
Controllability of Nonlinear Fractional
Neutral Stochastic Dynamical Systems
with Poisson Jumps
T. Sathiyaraj and P. Balasubramaniam
Abstract This paper is concerned with the controllability of fractional neutral

stochastic dynamical systems with Poisson jumps in the finite dimensional space.
Sufficient conditions for controllability results are obtained by using Krasnoselskii’s
fixed point theorem. The controllability Grammian matrix is defined by Mittag-
Leffler matrix function.
Keywords Controllability · Fractional differential equation · Mittag-Leffler func-

tion · Neutral stochastic system · Poisson jumps
MSC: 93B05 · 26A33 · 34A08 · 34K50 · 60J65
1 Introduction
Fractional differential equations have recently been proved to be valuable tools in the
modeling of many phenomena in various fields of science and engineering. It draws a
great application in nonlinear oscillations of earthquakes, many physical phenomena
such that seepage flow in porous media and in fluid dynamic traffic model. There
has been a significant development in fractional differential equations in recent years
(see [6, 9, 11, 12]).
It is well known that the concept of controllability plays an important role in
engineering and control theory. The controllability results for linear and nonlinear
integral order dynamical systems in finite-dimensional space have discussed exten-
sively (see [4]). Local null controllability of nonlinear functional differential systems
T. Sathiyaraj · P. Balasubramaniam (B)

Gandhigram Rural Institute-Deemed University,
Gandhigram 624 302, Tamil Nadu, India
e-mail: balugru@gmail.com
T. Sathiyaraj
e-mail: sathiyaraj133@gmail.com

DOI 10.1007/978-81-322-2485-3_35
430 T. Sathiyaraj and P. Balasubramaniam
in Banach space has been studied in [1]. Approximate controllability of fractional

order semilinear systems with bounded delay has been studied (see [8]).
In recent years, the controllability problems for stochastic differential equations
have become a field of increasing interest (see [2, 7, 10] and references therein).
Stochastic differential equations have many applications in ecology, finance, and
economics. The extensions of deterministic controllability concepts to stochastic
system have been discussed only in a limited number of publications.
The Poisson jumps have become very popular in recent years, because it is exten-
sively used to model many of the phenomena arising in areas such as economics,
finance, physics, biology, medicine, and other science. For example, if a system
jumps from a “normal state” to a “bad state,” the strength of systems is random. It
is natural and necessary to include a jump term in any dynamical system to make
more realistic sysrems. Complete controllability of stochastic evolution equations
with jumps has been studied in [13].
However, to the best of authors’ knowledge, there are no relevant reports on
the controllability of fractional neutral stochastic dynamical systems with Poisson
jumps in the finite-dimensional space. Motivated by the above, in this article the
controllability of fractional neutral stochastic dynamical systems is studied with
Poisson jumps in finite-dimensional spaces. Sufficient conditions for controllability
results are obtained by using Krasnoselskii’s fixed point theorem with a Grammian
matrix defined by Mittag-Leffler matrix function.
The paper is organized as follows: In Sect. 2, some well-known fractional opera-
tors and the solution representation of linear fractional stochastic differential equation
with Poisson jumps are discussed. In Sect. 3, the linear and nonlinear fractional neu-
tral stochastic differential equation with Poisson jumps are considered and the con-
trollability conditions are established by using the controllability Grammian matrix
which is defined by means of the Mittag-Leffler matrix function. Finally, concluding
remarks are given in Sect. 4.
2 Preliminaries
Let p and q are some positive constants satisfying n − 1 < q < n, n − 1 < p < n
and n ∈ N. Let Rm be the m-dimensional Euclidean space. The following notations
and definitions are well known, for a suitable function f ∈ L 1 (R+ ), R+ = [0, ∞)
for more details, (see [6]).
(a) Riemann–Liouville fractional operator:
x
q 1
(I0+ f )(x) = (x − t)q−1 f (t)dt
Γ (q) 0
(b) Mittag-Leffler Function:

The most interesting properties of the Mittag-Leffler function are associated with
Controllability of Nonlinear Fractional Neutral Stochastic … 431
their Laplace integral

∞ s q− p
e−st t p−1 E q, p (±at q )dt = ,
0 (s q ∓ a)
That is,
s q− p
L {t p−1 E q, p (±at q )}(s) = ,
(s q ∓ a)
(see [12]) for more details.

(c) Solution representation:
Consider the linear fractional stochastic differential equation with Poisson jumps
represented in the following form:
t +∞
1−q
d Jt (x(t) − x0 ) = Ax(t) + Bu(t) + σ (s)dw(s) dt + h(t, η)λ(dt, dη),
0 −∞
s, t ∈ J := [0, T ],
x(0) = x0 , (1)
1−q
where 0 < q < 1, Jt is the (1 − q)− order Riemann–Liouville fractional integral
operator x ∈ Rn , u ∈ Rm , A, B are matrices of dimensions n×n, n×m respectively
and σ : J −→ Rn×n , h : J × J −→ Rn are given functions.
+∞ t, η ∈ J } is a centered Poisson random measure with parameter
Let {λ(dt, dη),
π(dη)dt. Let −∞ π(dη) < ∞ and λ(dt, dη) = λ(dt, dη) − π(dη)dt is compen-
sated Poisson random measure which is independent of w(s).
Now applying the Riemann–Liouville fractional integral operator on both sides, we
get
t t
1 1
x(t) = x0 + (t − s)q−1
Ax(s)ds + (t − s)q−1 Bu(s)ds
Γ (q) 0 Γ (q) 0
t s
1
+ (t − s) q−1
σ (θ )dw(θ )ds
Γ (q) 0 0
t +∞
1
+ (t − s)q−1 h(s, η)λ(ds, dη).
Γ (q) 0 −∞
Taking the Laplace Transformation on both sides, we obtain
1 1 1 1 1
x̂(s) = x0 + q A x̂(s) + q B û(s) + q σ̂ (s) + q ĥ(s).
s s s s s
Taking inverse Laplace Transformation on both sides, we get

t s
x(t) = E q,1 (At )x0 +
q
(t − s) E q,q (A(t − s) ) Bu(s) +
q−1 q
σ (θ )dw(θ ) ds
0 0
t +∞
+ (t − s)q−1
E q,q (A(t − s) )
q
h(s, η)λ(ds, dη). (2)
0 −∞
Let (Ω, F , P) be the complete probability space with a probability measure P on

Ω and w(t) = (w1 (t), w2 (t), . . . , wn (t))T be an n−dimensional Wiener process
defined on the probability space. Let {Ft |t ∈ J } be the filtration generated by
{w(s) : 0 ≤ s ≤ t} defined on the probability space (Ω, F , P). Let L 2 (Ω, FT , Rn )
denotes the Hilbert space of all FT measurable square integrable random variables
with values in Rn . Let L F
2 (J, R ) be the Hilbert space of all square integrable and
n
Ft -measurable processes with values in Rn . Let B is the Banach space of all square
integrable and Ft -adapted process x(t) with norm
x2 = sup{Ex(t)2 },
t∈J
where E(·) denotes the mathematical expectation operator of stochastic process with
respect to the given probability measure P. Let L (Rn , Rm ) be the space of all
linear transformation from Rn to Rm . Further, we assume that the set of admissible
controls Uad := L F2 (J, R ). Now let us introduce the following operators and sets.
m
The linear bounded operator
L ∈ L (L F
2 (J, R ), L 2 (Ω, Ft , R ))
m n
is defined by
T
Lu = (T − s)q−1 E q,q (A(T − s)q )Bu(s)ds
0
and its adjoint linear bounded operator
L∗ : L 2 (Ω, FT , Rn ) −→ L F
2 (J, R )
m
is defined by
(L∗ z)(t) = B ∗ E q,q (A∗ (T − t)q )E{z|Ft },
and the set of all states attainable from x0 in time t > 0 using admissible controls is
defined by
Rt (Uad ) = {x(t; x0 , u) ∈ L 2 (Ω, FT , Rn ) : u(·) ∈ Uad }.

The linear controllability operator W0T ∈ L (L 2 (Ω, FT , Rn ), L 2 (Ω, FT , Rn ))

which is associated with the operator L is defined by
T
W0T = LL∗ {·} = (T − τ )q−1 [E q,q (A(T − τ )q )B][E q,q (A(T − τ )q )B]∗ E{(·)|Ft }dτ,
0
and the deterministic matrix ΓsT ∈ L (Rn , Rn ) is

T
ΓsT = (T − τ )q−1 [E q,q (A(T − τ )q )B][E q,q (A(T − τ )q )B]∗ dτ, s ∈ J.
s
Definition 1 The system (1) is said to be controllable on J if for every x0 , x1 ∈ Rn

there exists a stochastic control u(t) ∈ Uad such that the solution of x(t) of system
(1) satisfies the conditions x(0) = x0 and x(T ) = x1 .
Definition 2 The system (1) is completely controllable on J if
RT (x0 ) = L 2 (Ω, FT , Rn ),
that is, all points in L 2 (Ω, FT , Rn ) can be exactly reached from an arbitrary initial
condition x0 ∈ L 2 (Ω, FT , Rn ) at time T.
3 Controllability Results
In this section, we discuss the controllability criteria of linear and nonlinear stochastic
system with Poisson jumps.
Lemma 1 ([10]) If the linear system (1) is completely controllable, then for some
γ > 0,
E W0T z, z
≥ γ Ez2 ,
for all z ∈ L 2 (Ω, Ft , Rn )

and, consequently,
1
E(W0T )−1 2 ≤ = l2 .
γ
Lemma 2 ([10]) Assume that the operator W0T is invertible. Then, for arbitrary
x1 ∈ L 2 (Ω, FT , Rn ), the control
T
u(t) = B ∗ E q,q (A∗ (T − t)q )E (W0T )−1 x1 − E q,1 (AT q )x0 − (T − s)q−1 E q,q (A(T − s)q )
0
s
T +∞

× σ (θ)dw(θ) ds − (T − s)q−1 E q,q (A(T − s)q ) h(s, η)λ(ds, dη) Ft
0 0 −∞
transfers the system (1) form x0 ∈ Rn to x1 ∈ Rn at time T.

Proof Substituting the control u(t) into the solution x(t) in (2) and substituting
t = T, one can easily verify that the control u(t) steers the linear system x(t) from
x0 to x1 .
Let us consider the nonlinear fractional neutral stochastic dynamical systems with
Poisson jumps represented in the following form

1−q
d Jt (x(t) − g(t, x(t)) − x0 − g(0, x0 )) = A x(t) − g(t, x(t)) + Bu(t)
t
1−q
+ Jt f (t, x(t)) + σ (s, x(s))dw(s) dt
0
+∞
+ h(t, x(t), η)λ(dt, dη), s, t ∈ J,
−∞
x(0) = x0 , (3)
1−q
where 0 < q < 1, Jt is the (1 − q)–order Riemann–Liouville fractional integral
operator, A, B are the matrices of dimensions n × n, n × m respectively and f :
J × Rn −→ Rn , σ : J × Rn −→ Rn×n and h : J × Rn × R −→ Rn , are given
functions. Then the solution (3) is given by (see [3, 5])
t
x(t) = E q,1 (At )[x0 + g(0, x0 )] + g(t, x(t)) +
q
E q,1 (A(t − s)q ) f (s, x(s))ds
0
t s
+ (t − s)q−1
E q,q (A(t − s) ) Bu(s) +
q
σ (θ, x(θ ))dw(θ ) ds
0 0
t +∞
+ (t − s)q−1 E q,q (A(t − s)q ) h(s, x(s), η)λ(ds, dη).
0 −∞
Lemma 3 (Krasnoselskii’s fixed point theorem) Let E be a Banach space, let B

be a bounded closed and convex subset of E and let Φ1 , Φ2 be maps of B into E
such that Φ1 x, Φ2 y ∈ B for every pair x, y ∈ B. If Φ1 is a contraction and Φ2 is
completely continuous, then the equation Φ1 x + Φ2 x = x has a solution of B.
In order to prove the main results we assume the following conditions hold:
(H1) The functions g, f, σ and h satisfy the following Lipschitz conditions and
there exist some positive constants K , L , M and N such that
(i) g(t, x) − g(t, y)2 ≤ K x − y2
(ii) f (t, x) − f (t, y)2 ≤ Lx − y2
(iii) σ
+∞(t, x) − σ (t, y)2 ≤ Mx − y2
−∞ h(t, x, η) − h(t, y, η) λ(dη) ≤ N x − y
(iv) 2 2
(H2) The functions g, f, σ and h are continuous and satisfy the following linear
growth conditions. That is, there exist some positive constants K , L, M and
N such that
(i) g(t, x)2 ≤ K (1 + x2 )

(ii) f (t, x)2 ≤ L(1 + x2 )
(iii) σ
+∞(t, x)2 ≤ M(1 + x2 )
−∞ h(t, x, η) λ(dη) ≤ N (1 + x )
(iv) 2 2
(H3) The linear system (1) is completely controllable on J.

Now, define the nonlinear operator Φ from B to B as follows
t
(Φx)(t) = E q,1 (At q )[x0 + g(0, x0 )] + g(t, x(t)) + E q,1 (A(t − s)q ) f (s, x(s))ds
0
t s
+ (t − s)q−1 E q,q (A(t − s)q ) Bu(s) + σ (θ, x(θ))dw(θ) ds
0 0
t +∞
+ (t − s)
E q,q (A(t − s) )
q−1
h(s, x(s), η)λ(ds, dη)
q
0 −∞

E(Φx)(t)2 ≤ Δ := 36l1 l2 x1 2 + 6(1 + 6l1 l2 ) 2S1 (x0 2 + g(0, x0 )2 )
T 2q+1 T 2q
+ K + T 2 S2 L + 2
S3 Mσ M + 2 S3 N (1 + Ex2 )
q q
where

u x (t) = B ∗ E q,q (A∗ (T − t)q )E (W0T )−1 x1 − E q,1 (AT q )[x0 + g(0, x0 ) − g(T, x(T ))
T T
− E q,1 (A(T − s)q ) f (s, x(s))ds − (T − s)q−1 E q,q (A(T − s)q )
0 0
s T
× σ (θ, x(θ))dw(θ) ds − (T − s)q−1 E q,q (A(T − s)q )
0 0
+∞

× h(s, x(s), η)λ(ds, dη) Ft .
−∞
Applying Lemma 3 we need to construct two mapping Φ1 and Φ2 such that
(Φx)(t) = (Φ1 x)(t) + (Φ2 x)(t)
where
t t
(Φ1 x)(t) = E q,1 (A(t − s)q ) f (s, x(s))ds + (t − s)q−1 E q,q (A(t − s)q ) Bu(s)
0 0
s t
+ σ (θ, x(θ))dw(θ) ds + (t − s)q−1 E q,q (A(t − s)q )
0 0
+∞
× h(s, x(s), η)λ(ds, dη),
−∞
and
(Φ2 x)(t) = E q,1 (At q )[x0 + g(0, x0 )] + g(t, x(t)).
For convenience, let us introduce the following notations:
l1 = max{ΓsT 2 }, S1 = E q,1 (At q )2 , S2 = E q,1 (A(T − s)q )2 , S3 = E q,q (A(T − s)q )2 .
Theorem 1 Assume that the conditions (H1)–(H3) are hold and if Δ < 1 are
satisfied, then the nonlinear system (3) is completely controllable on J.
Proof In order to make more clear presentations, we divide the proof into the fol-
lowing three several steps.
Step I: For t ∈ J and any x, y ∈ B, we have
2 t
t
E(Φ1 x)(t)2 ≤ 4E
E q,1 (A(t − s)q ) f (s, x(s))ds
+ 4E (t − s)
q−1
E q,q (A(t − s)q )
0 0
2 2
t s
× Bu x (s)ds + 4E
(t − s)q−1
E q,q (A(t − s)q
) σ (θ, x(θ))dw(θ)ds

0 0
t 2
+∞
+ 4E
(t − s)q−1
E q,q (A(t − s)q
) h(s, x(s), η)λ(ds, dη)
.
0 −∞
Now, we have the following estimate

2
t
E
(t − s)q−1 E q,q (A(t − s)q )Bu x (s)ds
≤ 6l1 l2 x1 + 2S1 (x0 + g(0, x0 ) )
2 2 2
0
T 2q+1 T 2q
+ K + T 2 S2 L + 2
S3 Mσ M + 2 S3 N
q q

× (1 + Ex ) .
2
Thus

E(Φ1 x)(t)2 ≤ 4 T 2 S2 E f (t, x(t))2 + 6l1 l2 x1 2 + 2S1 (x0 2 + g(0, x0 )2 )
T 2q+1 T 2q
+ K + T 2 S2 L + 2
S3 Mσ M + 2 S3 N (1 + Ex2 )
q q
+∞
T 2q+1 2+ T 2q
2 λ(dη)
+ S3 M σ Eσ (t, x(t)) S 3 Eh(t, x(t), η)
q2 q2 −∞

2 2 2
≤ 4 6l1 l2 [x1 + 2S1 (x0 + g(0, x0 ) )] + 6l1 l2 K + (1 + 6l1 l2 )
T 2q+1 T 2q
× T 2 S2 L + S3 M σ M + S3 N (1 + Ex2)
q2 q2
and
E(Φ2 y)(t)2 ≤ 2E q,1 (At q )[x0 + g(0, x0 )]2 + 2Eg(t, y(t))2
≤ 4S1 [x0 2 + g(0, x0 )2 ] + 2K (1 + Ey2 ).
By the condition Δ < 1, we can find a r > 0 such that
x, y ∈ Br = {x ∈ B : Ex2 ≤ r }, EΦ1 x + Φ2 y2 ≤ r
that is Φ1 x + Φ2 y ∈ Br .
Step II: Φ1 is a contraction mapping on Br . For any x, y ∈ Br and t ∈ J, we have
t 2

E(Φ1 x)(t) − (Φ1 y)(t)2 ≤ 4E
E q,1 (A(t − s) )[ f (s, x(s)) − f (s, y(s))]ds
q

0
t 2

+ 4 E (t − s) q−1 E q,q (A(t − s) )B[u x (s) − u y (s)]ds
q

0
t

+ 4 E (t − s)q−1 E q,q (A(t − s)q )
0
s
2
× [σ (θ, x(θ)) − σ (θ, y(θ))]dw(θ) ds
0
t

+4 E (t − s)q−1 E q,q (A(t − s)q )
0
+∞
2
× [h(s, x(s), η) − h(s, y(s), η)]λ(ds, dη)
−∞

T 2q+1 T 2q
≤ 4 4l1 l2 K + (1 + 4l1 l2 ) T 2 S2 L + S M
3 σ M + S3 N
q2 q2
× Ex(t) − y(t)2 =: Υ Ex(t) − y(t)2 .
From the condition Δ < 1, we obtain Υ < 1, which implies that Φ1 is a contraction
mapping.
Step III: Φ2 is a completely continuous operator.
Due to continuity of A and continuity of g, the operator is Φ2 is continuous.
Next, we will show that {Φ2 x, x ∈ Br } is relatively compact. It suffices to show
that the family of function {Φ2 x, x ∈ Br } is uniformly bounded and equicontinuous
for any t ∈ J and {(Φ2 x)(t), x ∈ Br } is relatively compact. For any x ∈ Br , we
have EΦ2 x2 ≤ r which implies that {Φ2 x, x ∈ Br } is uniformly bounded. In the
following, we will show that {Φ2 x, x ∈ Br } is a family of equicontinuous functions.
For any x ∈ Br and 0 ≤ t1 < t2 ≤ T, we have
q q
E(Φ2 x)(t2 ) − (Φ2 x)(t1 )2 ≤ 4E q,1 (At2 ) − E q,1 (At1 )2 (x0 2 + g(0, x0 )2 )
+ 2Eg(t2 , x(t2 )) − g(t1 , x(t1 ))2 .
The right side of the above equation is independently of x ∈ Br as (t2 − t1 ) −→ 0

which means that {Φ2 x, x ∈ Br } is equicontinuous. Therefore {Φ2 x, x ∈ Br } is
relatively compact by Arzela–Ascoli theorem. The continuity of Φ2 and relative
compactness of {Φ2 x, x ∈ Br } imply that Φ2 is a completely continuous operator.

By using Krasnoseskii’s fixed point theorem we obtain that Φ1 +Φ2 has a fixed point
on Br . Therefore the system (3) has atleast one fixed point on J.
4 Conclusion
This paper deal with the controllability of fractional neutral stochastic dynamical
systems with Poisson jumps in the finite-dimensional space. Sufficient conditions
for controllability results have been obtained by using Krasnoseskii’s fixed point
theorem . The controllability Grammian matrix is defined by Mittag-Leffler matrix
function.
Acknowledgments The work of authors are supported by Council of Scientific and Indus-
trial Research, Extramural Research Division, Pusa, New Delhi, India under the grant No.
25/(0217)/13/EMR-II.
References
1. Balachandran, K., Balasubramaniam, P., Dauer, J.: Local null controllability of nonlinear func-
tional differential systems in Banach space. J. Optim. Theory Appl. 88(1), 61–75 (1996)
2. Balasubramaniam, P., Vembarasan, V., Senthilkumar, T.: Approximate controllability of impul-
sive fractional integro-differential systems with nonlocal conditions in Hilbert space. Numer.
Func. Anal. Opt. 35(2), 177–197 (2014)
3. Chikriy, A.A., Matichin, I.I.: Presentation of solutions of linear systems with fractional deriv-
atives in the sense of Riemann-Liouville, Caputo and Miller-Ross. J. Automat. Informat. Sc.
40(6), 1–11 (2008)
4. Karthikeyan, S., Balachandran, K.: Constrained controllability of nonlinear stochastic impul-
sive systems. Int. J. Appl. Math. Comput. Sci. 21(2), 307–316 (2011)
5. Kexue, L., Jigen, P.: Controllability of fractional neutral stochastic functional differential sys-
tem. Z. Angew. Math. Phys. 1–19 (2013). doi:10.1007/s00033-013-0369-2
6. Kilbas, A.A.A., Srivastava, H.M., Trujillo, J.J.: Theory and Applications of Fractional Differ-
ential Equations. Elsevier Science Limited (2006)
7. Klamka, J.: Stochastic controllability of linear systems with delay in control. Tech. Sci. 55(1),
23–29 (2007)
8. Kumar, S., Sukavanam, N.: Approximate controllability of fractional order semilinear systems
with bounded delay. J. Differ. Equ. 252(11), 6163–6174 (2012)
9. Lakshmikantham, V., Leela, S., Devi, J.V.: Theory of Fractional Dynamic Systems. Scientific
Publishers, Cambridge (2009)
10. Mahmudov, N., Zorlu, S.: Controllability of nonlinear stochastic systems. Int. J. Control. 76(2),
95–104 (2003)
Equations. Wiley, New York (1993)
12. Podlubny, I.: Fractional Differential Equations: An Introduction to Fractional Derivatives, Frac-
tional Differential Equations, to Methods of Their Solution and Some of Their Applications.
Academic Press (1998)
13. Sakthivel, R., Ren, Y.: Complete controllability of stochastic evolution equations with jumps.
Rep. Math. Phys. 68(2), 163–174 (2011)
Efficient Meshfree Method for Pricing
European and American Put Options
on a Non-dividend Paying Asset
Kailash C. Patidar and Abdelmgid O.M. Sidahmed
Abstract We develop efficient meshfree method based on radial basis functions

(RBFs) to solve European and American option pricing problems arising in compu-
tational finance. The application of RBFs leads to system of differential equations
which are then solved by a time integration θ -method. The main difficulty in pricing
the American options lies in the fact that these options are allowed to be exercised at
any time before their expiry. Such an early exercise right purchased by the holder of
the option results into a free boundary problem. Following the approach of Nielsen
et al. [B.F. Nielsen, O. Skavhaug and A. Tveito, Penalty methods for the numerical
solution of American multi-asset option problems. J. Comput. Appl. Math. 222, 3–16
(2008)], we use a small penalty term to remove the free boundary. The method is
analyzed for stability. Numerical results describing the payoff functions and option
values are also present. We also compute the two important Greeks, delta and gamma,
of these options.
Keywords American and European Options · Meshfree Methods · Free Boundary

Value Problems · Stability Analysis
1 Introduction
Options are frequently priced by means of partial differential equations (PDEs).

These options can be categorized into standard and nonstandard options. The present
work deals with the standard options (like European and American options). A large
amount of work has already been done to solve the PDEs representing European
options. However, the same for American options is not fully explored. For some
K.C. Patidar (B) · A.O.M. Sidahmed

Department of Mathematics and Applied Mathematics, University of the Western Cape,
Bellville 7535, South Africa
e-mail: kpatidar@uwc.ac.za
A.O.M. Sidahmed
e-mail: abdelmgid@gmail.com

DOI 10.1007/978-81-322-2485-3_36
440 K.C. Patidar and A.O.M. Sidahmed
historical developments, readers may refer to classical works of Black and Scholes
[1] and Merton [12, 13].
Researchers have attempted to solve these problems using a variety of techniques,
see e.g., adaptive θ -methods for solving American options (Khaliq et al. [11]), com-
pact finite difference methods (Zhao et al. [17]), generalized trapezoidal schemes
(Chawla et al. [3]), etc. On the other hand, methods based on meshfree approxima-
tions have been used a lot for problems in other domains of science and engineering,
see e.g., [8, 9, 14]. One of these popular meshfree methods are those based on the
radial basis functions (RBFs). Wua and Hon [16] used such an approximation for
solving diffusion-type problems under free boundary condition. In their work, the
numerical solution of the Black–Scholes equation for pricing American options,
which is a classical heat diffusion equation under free boundary value condition, is
obtained and compared with the traditional binomial method for numerical verifica-
tion.
In this work, we construct a meshfree method based on RBFs to solve European
and American option pricing problems. For American put option we use the penalty
method to remove the free boundary by adding a small penalty term. The basic idea
behind the use of RBFs is to use interpolation with a linear combination of basis
functions of the same type. A variety of RBFs are found in the literature. The two
RBFs that we will use in this paper are Gaussian and Multiquadratic.
The rest of the paper is organized as follows. Two option pricing problems are
described in Sect. 2. In Sect. 3 we discuss the application of radial basis functions to
solve these problems. The stability analysis of the numerical methods is presented
in Sect. 4. Finally, some numerical results along with a discussion on them are given
in Sect. 5.
2 Problem Description
In this paper, we consider the mathematical models for pricing European and Amer-
ican options. While a European option can only be exercised on the expiration date,
the American option can be exercised at any time before the expiration date.
The European option satisfies the following Black–Scholes equation
∂V 1 ∂2V ∂V
+ σ 2 S2 2 + r S − r V = 0, (1)
∂t 2 ∂S ∂S
where r is the risk-free interest rate, σ is the volatility of the stock price, and V (S, t)
is the option value at time t for the stock’s price S.
The initial condition is given by the terminal payoff

max(X − S, 0) for put
V (S, T ) = (2)
max(S − X, 0) for call
Efficient Meshfree Method for Pricing European … 441
whereas the boundary conditions are given by

V (0, t) = X e−r (T −t) , V (S, t) → 0 as S → ∞ for put
V (S, T ) =
V (0, t) = 0, V (S, t) → S as S → ∞ for call,
(3)
where T is the maturity time and X is the strike price of the option.
The exact solution of Eq. (1) with the initial condition (2) and the boundary con-
ditions (3) is given by [15]

X e−r (T −t) N (−d2 ) − S N (−d1 ) for put
V (S, T ) = (4)
S N (d1 ) − X e−r (T −t) N (d2 ) for call
where N (·) is the cumulative distribution function of the standard normal distribution
with

log(S/ X ) + r + 21 σ 2 (T − t)
d1 = √ (5)
σ T −t
and

log(S/ X ) + r − 21 σ 2 (T − t)
d2 = √ . (6)
σ T −t
On the other hand, the American option problem takes the form of a free boundary
problem. The early exercise constraint leads to the following model for the value
P(S, t) of an American put to sell the underlying asset [10]:
∂P
⎫
+ 21 σ 2 S 2 ∂∂ SP2 + r S ∂∂ PS − r P = 0, S > S f (t), 0 ≤ t < T ⎪
2
∂t ⎪
⎪
P(S, T ) = max(E − S, 0), S ≥ 0, ⎪
⎬
∂P
∂S (S f , t) = −1; P(S f (t), t) = E − S f (t), ⎪
(7)
lim S→∞ P(S, t) = 0, ⎪
⎪
⎪
⎭
S f (T ) = E; P(S, t) = E − S, 0 ≤ S < S f (t),
where S f (t) represents the free boundary, σ is the volatility of the underlying asset,
r is the risk-free interest rate, and E is the exercise price of the option. Since early
exercise is permitted, the value P of the option must satisfy
P(S, t) ≥ max(E − S, 0), S ≥ 0, 0 ≤ t ≤ T. (8)

3 Use of Radial Basis Functions in Pricing Options
3.1 European Options
We approximate the unknown function V (the value of the European option) using
the radial basis functions as

N
V (S, t) ≈ a j (t)φ(|S − x j |), (9)
j=1
where a j are unknown coefficients and φ(|S − x j |) are the RBFs. We will use the
following Gaussian radial basis functions for this problem
φ(S) = e−|S−x j |
2 /c2
, (10)
where c is a positive parameter.

Collocating at the N points x j ( j = 1, 2, . . . , N ), Eq. (1) becomes
∂ V (xi , t) 1 2 2 ∂ 2 V (xi , t) ∂ V (xi , t)

+ σ Si + r Si − r V (xi , t) = 0. (11)
∂t 2 ∂ S2 ∂S
Differentiating (9), we obtain
∂ V (xi , t) da j (t)
N
= φ(|S − x j |), (12)
∂t dt
j=1
∂ V (xi , t)
N
∂φ(|S − x j |)
= a j (t) , (13)
∂S ∂S
j=1
∂ 2 V (xi , t)
N
∂ 2 φ(|S − x j |)
= a j (t) . (14)
∂ S2 ∂ S2
j=1
Using (10)–(11), and simplifying, we obtain
da
Φ + Ra = 0, (15)
dt
where
Φi j = e−xi −x j
2 /c2
(16)
and

1 2 2 4(xi − x j )2 − 2c2 −2(xi − x j )
Ri j = σ xi Φi j + r xi Φi j − r Φi j . (17)
2 c4 c2
To solve the system described by Eq. (15), we use a θ -method
a n+1 − a n
Φ + θ Ra n+1 + (1 − θ )Ra n = 0, (18)
Δt
with the initial condition given by the first part of Eq. (2) and boundary conditions
given by the first part of Eq. (3).
We can rewrite Eq. (18) as
[Φ − (1 − θ )ΔtR]a n = [Φ + θ ΔtR]a n+1 . (19)
⇒ a n = [Φ − (1 − θ )ΔtR]−1 [Φ + θ ΔtR]a n+1 . (20)
Furthermore, Eq. (9) applied at all collocation points can be written in the matrix
form as
V = Φa. (21)
Using Eq. (21), Eq. (20) can be written as
V n = Φ −1 [Φ − (1 − θ )ΔtR]−1 [Φ + θ ΔtR]ΦV n+1 . (22)
The above equation is solved along with (2) and the first part of Eq. (9) to obtain
the numerical solution. Also the form of this equation should be read in context to
the computing process because in the problems like those considered in this paper,
we usually have a final boundary value problem rather than an initial-boundary
value problem. To this end, note that the scheme given by (19) corresponding to
θ = 0, 0.5, and 1 are implicit Euler, Crank–Nicolson and explicit Euler methods,
respectively.
3.2 American Options
To solve the American option problem (7), which is a free boundary problem, we
approximate the model by adding a penalty term. This leads to a nonlinear partial
differential equation on a fixed domain. More precisely, we consider the initial-
boundary value problem
∂ Pε 1 ∂ 2 Pε ∂ Pε εC
+ σ 2 S2 + rS − r Pε + = 0, (23)
∂t 2 ∂ S2 ∂S Pε + ε − q(S)
with the initial condition as the first part of Eq. (2) and the boundary conditions as
Pε (0, t) = E, lim Pε (S, t) = 0, (24)

S→∞
where C ≥ r E, q(S) = E − S, and 0 < ε

1.
Using multiquadric radial basis functions, we find
∂φ(xi − x j ) (xi − x j )
= (25)
∂S (xi − x j )2 + c2
and
∂ 2 φ(xi − x j ) c2
= . (26)
∂ S2 ((xi − x j )2 + c2 )3
Proceeding in the similar manner as in the previous case, we obtain
da
Φ + Ra + Q(a) = 0, (27)
dt
where

Φi j = (xi − x j )2 + c2 , i, j = 1, . . . , N , (28)
εC
Q(a) = , i = 1, . . . , N
Φi a + ε − q(xi )
and
⎛ ⎞ ⎛ ⎞
1 c 2 (x − x )
i j = σ 2 xi2 ⎝
R ⎠ + r xi ⎝ i j ⎠ − r Φi j . (29)
2 ((xi − x j ) + c )
2 2 3 (xi − x j ) + c
2 2
Using θ -method, Eq. (27) becomes
a n+1 − a n n+1 + (1 − θ )Ra

n + θ Q(a n+1 ) + (1 − θ )Q(a n ) = 0. (30)
Φ + θ Ra
Δt
Consequently, the nonlinear penalty term gives rise to a nonlinear system of equations
whose solution is typically found by a modified Newton method. By replacing a n in
the penalty term by a n+1 (as in [10]), the linearly implicit scheme corresponding to
Eq. (30) is given by
a n+1 − a n n + Q(a n+1 ) = 0,

n+1 + (1 − θ )Ra
Φ + θ Ra (31)
Δt
with the initial condition given by the first part of Eq. (2) and boundary conditions
given by Eq. (3).
4 Stability Analysis
To proceed with the stability analysis, let us define the error at the nth time level by
en = Vexact
n
− Vapp
n
, (32)
n
where Vexact n are the exact and numerical solutions obtained by either (18)
and Vapp
or (31), respectively.
For the scheme given by (21) the error equation at (n + 1)th level can be written
as
en = Ben+1 , (33)
where B is the amplification matrix is given by
B = Φ −1 [Φ + θ ΔtR][Φ − (1 − θ )ΔtR]−1 Φ.
The numerical method will be stable if ρ(B) ≤ 1, where ρ(B) is the spectral radius
of B.
Substituting the value of B in Eq. (33) and simplifying, we obtain
[Φ − (1 − θ )ΔtR]Φ −1 en = [Φ + θ ΔtR]Φ −1 en+1 . (34)
This implies
[I − (1 − θ )Δt M]en = [I + θ Δt M]en+1 (35)
where M = RΦ −1 and I is an N × N identity matrix.

It is clear from Eq. (35) that the numerical scheme is stable if all the eigenvalues
of the matrix [I − (1 − θ )Δt M]−1 [I + θ Δt M] are less than unity, which means that
1 + θ Δtλ M

≤ 1, (36)
1 − (1 − θ )Δtλ M
where λ M represent the eigenvalues of the matrix M.

To check the above, we consider different cases. Firstly, when θ = 1, we have

explicit Euler method. The above condition for stability in this case becomes
|1 + Δtλ M | ≤ 1. (37)
Hence the explicit Euler method will be stable if Δt ≥ −2/λ M and λ M ≤ 0.

Secondly, when θ = 0, we have implicit Euler method which is unconditionally
stable as can be seen from (36) because λ M ≤ 0. Finally, when θ = 0.5, we have
the Crank–Nicolson’s method. Even in this case, the inequality (36) will hold as
long as λ M ≤ 0 and this does happen. Therefore, the Crank–Nicholson’s method is
unconditionally stable. Note that the stability analysis for (31) can be done along the
similar lines.
5 Numerical Results and Discussion
Using the RBF approach, the resulting problems for European and American put
options are solved via Crank–Nicolson’s method (i.e., θ = 0.5) with Δt = 0.01.
Results are presented in Table 1.
The parameters used for the simulations for European put option problem are:
r = 0.05, σ = 0.2, D = 0, E = 10, t0 = 0, T = 0.5, S0 = 0 and Smax = 30.
The first column in this table represents values of the asset price S, the second column
represents the exact solution and the other three columns indicate the numerical
values of the European put option that we obtain using the radial basis function
approach with 21, 41 and 101 nodes, respectively.
For the American put options, we choose r = 0.1, σ = 0.2, D = 0, E =
1, t0 = 0, T = 1, ε = 0.01, S0 = 0, and Smax = 2. We again use the Crank–
Nicolson method with Δt = 0.01. Using the multiquadratic radial basis function
Table 1 Values of European put option using radial basis functions

S Exact RBF21 RBF41 RBF101
6 3.7532 3.7528 3.7529 3.7532
7 2.7568 2.7659 2.7594 2.7572
8 1.7987 1.8510 1.8080 1.8003
9 0.9880 1.0079 0.9908 0.9886
10 0.4420 0.5280 0.4628 0.4454
11 0.1606 0.2087 0.1754 0.1629
12 0.0483 0.0499 0.0504 0.0486
13 0.0124 0.0206 0.0147 0.0127
14 0.0028 0.0040 0.0035 0.0029
Table 2 Values of American put option using radial basis functions

S RBF21 RBF41 RBF101
0.6 4.0000e-01 4.0000e-01 4.0000e-01
0.7 3.0011e-01 3.0011e-01 3.0012e-01
0.8 2.0198e-01 2.0202e-01 2.0204e-01
0.9 1.1657e-01 1.1687e-01 1.1695e-01
1.0 5.9688e-02 6.0169e-02 6.0295e-02
1.1 2.8756e-02 2.9196e-02 2.9330e-02
1.2 1.3656e-02 1.4005e-02 1.4147e-02
1.3 6.7494e-03 7.0219e-03 7.2231e-03
1.4 3.6229e-03 3.9109e-03 4.2547e-03
10 1
9 0.9
8 0.8
7 0.7
6 0.6
5 0.5
P
V
4 0.4
3 0.3
2 0.2
1 0.1
0 0
0 5 10 15 20 25 30 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
S S
Fig. 1 Left plot Value of the European put at t0 using 101 points and r = 0.05, σ = 0.2, E =
10, t0 = 0, T = 0.5, S0 = 0 and Smax = 30. The curve with ‘*’ shows payoff whereas the
solid curve represents the value of the option. Right plot Value of an American put at t0 using 101
points and r = 0.1, σ = 0.2, E = 1, T = 1, ε = 0.01. The curve with ‘*’ shows payoff whereas
the solid curve represents the value of the option
√
r 2 + c2 , we obtain reasonably accurate results in the sense that they are very close
to those obtained by Fasshauer in [4]. This can be seen from Table 2.
Finally, in Fig. 1, we depict some special cases for European and American options
as indicated in the figure caption.
The accuracy of the solution obtained by using meshfree methods depends on
the choice of the shape parameter c. The choice of the optimal value of this para-
meter is still an open problem. Many researchers have chosen it as c = 2h, where
h = (Smax − S0 )/(N − 1). After some numerical experiments, we found the optimal
value of this shape parameter using Gaussian RBFs as approximately 0.79.
Since the radial basis functions are infinitely differentiable, the computations
of the derivatives of options are readily available from the derivatives of the basis
functions. In Table 3, we present values of delta for European put options and compare
them with their exact values. It is clear from the results presented in these tables
Table 3 Values of option’s delta for European put

S Analytic values of option’s Δ Numerical values of option’s Δ
8.0000 −0.9083 −0.90665
9.0000 −0.6906 −0.6902
10.0000 −0.4023 −0.4031
11.0000 −0.1784 −0.1798
12.0000 −0.0622 −0.0625
Table 4 Comparison of delta for American put option

S LUBA EXP QFK RBFs
80 −1.0000 −1.0000 −1.0000 −0.9997
90 −0.6173 −0.6207 −0.6212 −0.6220
100 −0.3588 −0.3582 −0.3581 −0.3602
110 −0.2108 −0.2109 −0.2108 −0.2129
120 −0.1256 −0.1257 −0.1256 −0.1280
Table 5 Values of option’s gamma for European put

S Analytic values of option’s Γ Numerical values of option’s
Γ
6.0000 0.0016 0.0014
7.0000 0.0303 0.0315
8.0000 0.1455 0.1461
9.0000 0.2770 0.2767
10.0000 0.2736 0.2722
11.0000 0.1677 0.1678
12.0000 0.0722 0.0723
that the numerical values of the option’s delta lie between −1 and 0 which is in
agreement with what is mentioned in Hull [5]. Note that analytical solution for the
Δ for American option is not available and therefore in Table 4, we compare them
with some of those seen in the literature. We also calculate the Γ of a portfolio of
options on an underlying asset which is the rate of change of the portfolio’s delta
with respect to the price of the underlying asset. It is the second partial derivative of
the portfolio with respect to the asset price. If the absolute value of Γ is large, Δ is
highly sensitive to the price of the underlying asset. Table 5 gives the values of Γ
for European put options. The first column in this table represents the values of the
asset price S, the second column represents the analytical values of option’s Γ and
the third column represents the numerical values of it using the proposed approach.
In Table 4, the acronyms LUBA, EXP, QFK and RBF, respectively, stand for
Lower and Upper Bound Approximations [2], multipiece Exponention [6], Quadra-
ture Formula of Kim equations [7], and Radial Basis Function approach proposed in
this paper.
6 Conclusions
In this paper, we presented a meshfree method based on radial basis function approx-
imations to solve European and American style option pricing problems. While the
approach for European option problems was straightforward, we have to use a penalty
approach to solve the problems for pricing American options. Proposed method is
analyzed for stability. Numerous comparative results are present. It may be noted
that the calculation of Greeks from our method was free of any spurious oscillations.
Furthermore, the methods can be used to solve the problems where asset pays a
dividend because the only difference in that case would be the fact that the asset
can be less than the payoff which will not affect the performance of the method.
Another important feature of the proposed meshfree method is the local adaptivity
of the radial basis functions which allows for its possible extensions to multi-asset
problems.
Acknowledgments We thank the anonyms referees for their valuable comments and suggestions.
The research of KCP was supported by the South African National Research Foundation. AOMS
acknowledges the financial support of AL-Neelain University, Sudan.
References
1. Black, F., Scholes, M.: Pricing of options and corporate liabilities. J. Pol. Eco. 81(3), 637–654
(1973)
2. Broadie, M., Detemple, J.: American option valuation: new bounds, approximations, and a
comparison of existing methods. Rev. Fin. Stud. 9, 1211–1250 (1996)
3. Chawla, M.M., AL-Zanaidi, M.A., Evans, D.J.: Generralized trapezoidal formulas for valuing
American options. Int. J. Comp. Math. 81(3), 375–381 (2004)
4. Fasshauer, G.E., Khaliq, A.Q.M., Voss, D.A.: Using meshfree approximation for multi-asset
American option. J. Chin. Inst. Eng. 27(4), 563–571 (2004)
5. Hull, J.C.: Options, Futures, and Other Derivatives. Prentice Hall, Upper Saddle River (2009)
6. Ju, N.: Pricing an American option by approximating its early exercise boundary as a multipiece
exponential function. Rev. Fin. Stud. 11, 627–646 (1998)
7. Kallast, S., Kivinukk, A.: Pricing and hedging American options using approximations by
Kim integral equations. Eur. Fin. Rev. 7(361–383), 361–383 (2003)
8. Kansa, E.J.: Multiquadrics—a scatered data approximation scheme with applications to com-
putional Fluid-Dynamics-I. Comput. Math. Appl. 19, 127–145 (1990)
9. Kansa, E.J.: Multiquadrics—a scatered data approximation scheme with applications to com-
putional Fluid-Dynamics-II. Comput. Math. Appl. 19, 147–161 (1990)
10. Khaliq, A.Q.M., Voss, D.A., Kazmi, S.H.K.: A linearly implicit predictor-corrector scheme
for pricing American options using a penalty method approach. J. Bank. Fin. 30, 489–502
(2006)
11. Khaliq, A.Q.M., Voss, D.A., Kazmi, S.H.K.: Adaptive θ-methods for pricing American
options. J. Comput. Appl. Math. 222, 210–227 (2008)
12. Merton, R.C.: Theory of rational option pricing. Bell J. Eco. Man. Sci. 4, 141–183 (1973)
13. Merton, R.C.: Option pricing when the underlying stocks are discontinuous. J. Fin. Eco. 5,
125–144 (1976)
14. Tatari, M., Dehghan, M.: A method for solving partial differential equations via radial basis
functions: application to the heat equation. Eng. Anal. Bound. Elem. 34, 206–212 (2010)
15. Wilmott, P., Howison, S., Dewynne, J.: The Mathematical Financial Derivatives: A Student
Introduction. Cambridge University Press, Oxford (1995)
16. Wua, Z., Hon, Y.C.: Convergence error estimate in solving free boundary diffusion problem
by radial basis functions method. Eng. Anal. Bound. Elem. 27, 73–79 (2003)
17. Zhao, J., Davison, M., Corless, R.M.: Compact finite difference method for American option
pricing. J. Comput. Appl. Math. 206, 306–321 (2007)
A Laplace Transform Approach
for Pricing European Options
Edgard Ngounda and Kailash C. Patidar
Abstract In this paper we investigate two efficient numerical methods for solving
the Black–Scholes equation for pricing European options. We use spectral methods
to discretize the associated partial differential equation with respect to space (asset
direction) and generate a system of ordinary differential equations in time. This
system is then solved by applying the numerical inversion of the Laplace transform
which is based on the Talbot’s method [A. Talbot, The accurate numerical inversion
of Laplace transforms, IMA J. Appl. Math. 23(1), 97–120 (1979)]. This involves an
application of trapezoidal rule to approximate a Bromwich integral. Using Cauchy’s
integral theorem, we deform the Bromwich line into a contour which starts and
ends in the left half plane. Comparative numerical results obtained by this and other
three methods (Exponential Time Differencing Runge–Kutta Methods of order 4,
MATLAB solver ode15s and Crank-Nicholson’s method) are presented.
Keywords Option Pricing · Contour Integrals · Spectral Methods · Exponential

Time Differencing Runge-Kutta Methods
1 Introduction
Since its development in the 1970s by F. Black and M. Scholes, the Black–Scholes
equation has become a fundamental model for pricing financial derivatives [1]. A
derivative security is a financial instrument whose value depends on the values of
some other underlying variables, e.g. stocks, foreign currency. Among the most
popular derivatives, options are actively traded on different financial markets over
the world. An option gives its holder the right without any obligation to buy (call
E. Ngounda · K.C. Patidar (B)

Department of Mathematics and Applied Mathematics, University of the Western Cape,
Bellville 7535, South Africa
e-mail: kpatidar@uwc.ac.za
E. Ngounda
e-mail: edgardngounda@gmail.com

DOI 10.1007/978-81-322-2485-3_37
452 E. Ngounda and K.C. Patidar
option) or to sell (put option) the underlying asset by a certain date (maturity date) for
a certain price (strike price). The European options can only be exercised at maturity.
The Black–Scholes partial differential equation can be used to model different
types of options. However, a closed form solution cannot always be found and we
must therefore resort to numerical methods to solve such a PDE. Some of the most
popular methods used in the past to tackle these type of problems are those based on
Monte Carlo simulations [2], binomial trees [3] and finite difference methods [7].
Finite difference methods are classical methods for solving PDEs and have been
used extensively to price options since the advent of the financial mathematics. The
authors in [9] used a grid stretching in combination with backward difference method
of fourth order in time to solve the European options. In [11], Tangman et al. used
a method based on the grid stretching to generate a high-order compact scheme to
improve on the well-known second-order Crank–Nicolson method for solving these
problems. In spite of the popularity of these time marching methods, one of their
critical drawback is that they usually require as many time steps as spatial meshes to
maintain their stability.
In this paper, we consider the application of Laplace transform which has recently
been investigated by some researchers and is considered to be a valuable alternative
method to finite differences methods for solving parabolic PDEs [4, 10, 15]. This
has led to great applications in the financial world.
The rest of the paper is organized as follows. In Sect. 2 we give a full description
of the Black–Scholes equation which is used to model the European put and call
options. In Sect. 3, we introduce the spectral discretization method. Application of
the Laplace integration method is discussed in Sect. 4. Finally, in Sect. 2, we present
comparative numerical results.
2 Problem Description
We consider the following Black–Scholes (BS) equation to price European options
∂V 1 ∂2 V ∂V
+ σ2 S2 2 + r S − r V = 0, S ∈ (0, ∞), t ∈ (0, T ). (1)
∂t 2 ∂S ∂S
Final and boundary conditions are given by
⎧
⎨ max(S − K , 0) for call
V (S, T ) = (2)
⎩
max(K − S, 0) for put
and
A Laplace Transform Approach for Pricing European Options 453
⎫
V (0, t) = 0, V (S, t) → S − K e−r (T −t) as S → ∞ for call, ⎬
(3)
⎭
V (0, t) = K e−r t , V (S, t) → 0 as S → ∞ for put.
In the above, V (S, t) is the price of a call/put option for the underlying asset whose
price is S at time t up to the expiry date T , r is the interest rate, σ is the volatility of
the underlying asset, and K is the strike price.
We set τ = T − t to transform the backward formulation (1)–(3) to the following
forward equation:
∂V 1 ∂2 V ∂V
− σ2 S2 2 − r S + r V = 0, (4)
∂τ 2 ∂S ∂S
The initial condition is given by the terminal payoff
⎧
⎨ max(S − K , 0) for call
V (S, 0) = (5)
⎩
max(K − S, 0) for put
and the boundary conditions are given by

⎫
V (0, τ ) = 0, V (S, τ ) → S − K e−r τ as S → ∞ for call, ⎬
(6)
⎭
V (0, τ ) = K e−r τ , V (S, τ ) → 0 as S → ∞ for put.
3 Spectral Discretization
To semi-discretize the PDE (1), we consider a spectral method. The basic idea behind
the spectral methods is as follows. For a given set of points, we interpolate the
unknown solution and differentiate the interpolating polynomial at these grid points.
This discretization process leads to a system of equations which can then be solved
using any state-of-the-art solvers.
The discretization using spectral method (in this paper) is based on the Chebyshev
polynomial interpolation [13]. Methods such as finite elements or finite differences
divide the domain into subdomains and use local polynomials of low degree. By
contrast, spectral methods use global representations of high degree over the entire
domain.
The implementation of spectral methods can be divided into three categories,
namely, the Galerkin, tau, and the collocation (or pseudospectral) methods. The first
two of these methods use the expansion coefficients of the global approximation
and the latter can be viewed as a method of finding numerical approximations to
derivatives at collocation points. In a manner similar to finite difference or finite
element methods, the equation to be solved is satisfied in space at the collocation

points. In this paper, we use the third one, i.e., the spectral collocation method.
The spectral process involves seeking the solution to a differential equation by
polynomial interpolation. In order to review the concept of polynomial interpolation,
we consider interpolating an arbitrary function f (x) at N + 1 distinct nodes {xk }k=0
N
in [−1, 1].
N
Given a set of grid points x j j=0 , an interpolating approximation to a function
f (x) is a polynomial f N (x) of degree N , determined by the requirement that the
N
interpolant agrees with f (x) at the set of interpolation points x j j=0 , i.e.,
f N (xi ) = f (xi ), i = 0, 1, ..., N . (7)
We define by L k (x), the Lagrange polynomial of degree N ,

N
x − xj
L k (x) = , k = 0, 1, ..., N .
j=0
xk − x j
j =k
Note that L k (x) satisfies L j (xk ) = δ jk , where δ jk is the Kronecker delta function.
The interpolation polynomial f N (x) is then given by

N
f N (x) = f (xk )L k (x). (8)
k=0
In this paper, we use the Chebyshev points as the grid points. These are given by

2 j+1
Chebyshev zeros: x j = cos 2(N +1)
π , j = 0, ..., N ,
and
Chebyshev extrema: x j = cos jπN , j = 0, ..., N .
The Chebyshev points are often defined as the projection onto the interval [−1, 1]
of the roots of unity along the unit circle |z| = 1 in the complex plane [13]. For Euro-
pean options, since the payoff is nonsmooth, a direct application of the Chebyshev
points for discretization leads to low-order approximation. To regain a high-order
accuracy an alternative approach was proposed by Tangman [12]. The basic idea is
to modify the Chebyshev points as
x = [xk , x ]T , (9)
where

K − Smin 2πk N
xk = Smin + 1 − cos , k = 0, 1, ..., , (10)
2 N 2

Smax − K 2πl N
x = K + 1 − cos , = 1, 2, ..., . (11)
2 N 2
for N even. This discretization clusters grid nodes at the boundaries located at Smin
and Smax as well as at the strike price K where the discontinuity of the payoff occurs.
As we show in Sect. 5, it follows that local grid refinement improve accuracy of the
spectral method at the payoff. Another advantage of this strategy is that it applies
directly to the Eq. (4) without the need for transforming into the interval [−1, 1].
Differentiation Matrices
The concept of collocation derivatives is associated with the interpolation polynomial
f N (x) as described above. These are the derivatives of f N (x) at the collocation points
{xk }k=0
N . Using (8), we can see that the mth order collocation derivative of f (x) is
N
given by
d m f N (x)
N
d m L k (x)
= f (x k ) . (12)
dxm dxm
k=0
Nodal representation yields
d m f N (x j )
N
d m L k (x j )
= f (x k ) , j = 0, ..., N , (13)
dxm dxm
k=0
which can be expressed by the matrix formula
f (m) N = D (m)
N fN , (14)
where ⎡ ⎤ ⎡ (m) ⎤
f N (x0 ) f N (x0 )
⎢ .. ⎥ (m) ⎢ .. ⎥
fN = ⎣ . ⎦, f N = ⎣ . ⎦,
f N (x N ) (m)
f N (x N )
(m)
and D N is the (N + 1) × (N + 1) differentiation matrix of order m with entries

(m) (m)
DN = L k (x j ), j, k = 0, ..., N . (15)
j,k
The computation of these differentiation matrices for an arbitrary order m has been
considered in [6, 13]. Following the approach in [16], Weideman and Reddy [14]
developed a MATLAB algorithm (DMSUITE package) that computes the Chebyshev
grid points as well as the differentiation matrices of an arbitrary order. The suite
contains a function chebdif that computes the extreme points of the Chebyshev
(m)
polynomial TN (x) and the differentiation matrix D N . The code takes as input the
size of the differentiation matrix N and the highest derivative order m and produces
()
matrices D N of order = 1, 2, ..., m.
(1)
Formulas for the computation of the entries of D N , N ≥ 1, let i, j = 0, 1, ...N ,
are (as given in [13]):
2N 2 + 1 2N 2 + 1
(1) (1)
DN = , DN = , (16)
00 6 NN 6
−x j
D (1)
N = , j = 1, ..., N − 1, (17)
jj 2(1 − x 2j )
ci (−1)i+ j
(1)
DN = , i = j, i, j = 0, ..., N , (18)
ij c j (xi − x j )
where ⎧
⎨ 2, i = 0 or N
ci =
⎩
1, otherwise.
Higher order derivatives are evaluated by recursions at a cost of O(N 2 ) operations

[14, 16]. This turns out to be cost-effective as compared to O(N 3 ) if higher derivatives
are obtained by taking powers of the first derivative [14].
Using the differentiation matrices as described above, we can rewrite (4) in matrix
form as
∂V 1 2
− σ P D (2) V − r Q D (1) V + r V = 0, (19)
∂τ 2
where P and Q are the diagonal matrices with entries on the main diagonals as
(xk + 1)2 and (xk + 1), respectively, for k = 0, ..., N .
We will solve Eq. (19) using several time integration methods as indicated in the
next two sections.
4 Application of the Laplace Transform to Price

the European Call and Put Options
Applying the Laplace transform to Eq. (4), we obtain
1 ∂ 2 V̄ ∂ V̄
z V̄ − σ 2 S 2 2 − r S + r V̄ = V0 . (20)
2 ∂S ∂S
The boundary conditions are given by
⎫
V̄ (0, z) = 0, V̄ (S, z) = Smax
z − K
(z+r ) for call, ⎪
⎬
(21)
⎪
⎭
V̄ (0, z) = (z + r ) ,
K
V̄ (S, z) = 0 for put.
The Eq. (19), therefore, becomes
1
z V̄ − σ 2 P D (2) (1)
N V̄ + r Q D N V̄ − r V̄ = V0 ,
2

1 2 (2) (1)
z k I − σ P D N + r Q D N − r I V̄k = V0 k = 0, ..., N − 1. (22)
2
A straight forward application of the Laplace inversion formula [8] yields

∞
h
V(t) = e z()t V̄z ()d. (23)
2πi −∞
Using the symmetry, the trapezoidal approximation yields
h zk t
M−1
V M (t) = e V̄k z k , (24)
π
k=0
where
V̄k = (z k I − A)−1 V0 , k = 0, 1, ..., N − 1, (25)
and
1 2 (2) (1)
A= σ P D N − rQD N + r I. (26)
2
(1) (2)
Now since the differentiation matrices D N and D N are not sparse, the Eq. (25)
indicates the bulk of the computation in the trapezoidal rules (24). To speed up this
computation, an Hessenberg decomposition can be computed once at the beginning
as follows:
A = MHMT , (27)
where H = (h i j ) is an upper Hessenberg matrix, i.e., h i j = 0, i > j + 1, and M an

orthogonal matrix. Then for each z k , k = 0, 1, ..., M − 1, the Eq. (25) becomes
(z k I − MHMT )Vk = V0 . k = 0, 1, ..., N − 1. (28)
From this we have
(z k I − H )Uk = M T V0 k = 0, 1, ..., N − 1, (29)

Table 1 Parameters used in the contour over an interval [t0 , Λt0 ]

Λ α A(α) μΛt0 /M B(α)
1 1.1721 1.0818 4.4921 2.3157
5 1.0791 2.4578 1.5013 1.2570
10 1.0236 3.3744 0.8871 1.0888
50 0.9381 5.5582 0.3452 0.7152
The right column B(α) shows the convergence rate over the contour for each set parameters
where Uk = M T Vk , so that
Vk = MUk , k = 0, 1..., N − 1. (30)
The solution Vk for each z k , is obtained by the computation of an almost triangular

system (29) and combining the result in (30) at only O(N 2 ) operations [5]. During
this process, the Hessenberg reduction (27) is only computed once beforehand.
For numerical implementation, we considered the following contour parameters
defined over an interval [t0 , Λt0 ] (as defined in [15])
z =
μ(1 + sin (iw − α)), (31)
where
(π − 2α) Λ − π + 4α
A(α) = cosh−1
(4α − π) sin α
and
A(α) 4απ − π 2 M
h= ,
μ= ,
M A(α) Λt0
with Λ ∈ N and M is the number of points in the trapezoidal rule. The convergence

rate of the Laplace method on these contour is given by O e−B(α)M where
π 2 − 2πα
B(α) = .
(π − 2α)Λ + 4α − π
cosh−1 (4α − π) sin(α)
Values of the above parameters are given in Table 1.
We compare the results obtained by using our Laplace transform method with
those obtained by simulations that we perform using ETDRK-4 (Exponential Time
Differencing Runge–Kutta Method of order 4) as well as the more conventional
Table 2 Comparison of the errors defined by (32), for the Crank–Nicolson’s method, ETDRK4
and the Laplace inversion approach applied for a European call option
ode15s Crank–Nicolson ETDRK4 Laplace inversion method
N Time (s) Error Time (s) Error Time (s) Error Time (s) Error
20 12.5E-2 8.2E-2 6.0E-2 7.1E-3 5.2E-2 7.4E-3 1.1E-2 7.4E-3
30 14.8E-2 6.97E-4 4.47E-2 1.3E-3 5.4E-2 1.00E-3 5.1E-3 1.0E-3
40 19.3E-2 6.59E-5 10.9E-2 1.93E-4 7.5E-2 1.22E-4 7.0E-3 1.18E-4
50 22.3E-2 7.86E-6 13.3E-2 9.67E-5 9.5E-2 4.85E-5 9.1E-3 1.07E-5
60 24.2E-2 4.63E-6 1.63E-2 9.73E-5 12.6E-2 4.86E-5 1.2E-4 3.52E-6
80 31.0E-2 1.86E-5 26.1E-2 9.80E-5 21.8E-2 4.89E-5 2.0E-4 5.80E-7
25 10
20 8
15 6
4
V
10
V
5 2
0 0
−5 −2
0 5 10 15 20 25 30 0 5 10 15 20 25 30
S S
0.2
0 0.4
0.35
−0.2 0.3
−0.4 0.25
0.2
Δ
−0.6 0.15
−0.8 0.1
0.05
−1 0
−1.2 −0.05
0 5 10 15 20 25 30 0 5 10 15 20 25 30
S S
Fig. 1 Top figures Europeans call option (left), put (right). Bottom figures Δ (left) and Γ (right)
for European put option. K = 10, r = 0.05, σ = 0.2, Smax = 3K , T = 0.25 N = 80
time-marching methods such as Crank-Nicholson’s method (with stepsize 2.5e − 3)

and the well-known MATLAB solver ode15s. These results are presented in Table 2.
For the numerical simulations, we fix spatial variable S at Smax = 3K to reduce the
domain truncation error. Other parameters are chosen as follows K = 15, σ = 0.2,
r = 0.05, T = 0.25. Maximum absolute errors are calculated using the formula
error = max |V(t) − V M (t)|, (32)

t∈[0,T ]
where V(t) is the analytical solution obtained by using the Black–Scholes formula
and V M (t) is the numerical solution obtained by any of the three methods as indicated
in Table 2.
In Fig. 1, we plot values for Europeans call (and put) options as well as the Greeks
Δ and Γ . We notice that both Greeks are free of oscillations.
It is worth mentioning here that even though in practice, the use of spectral methods
for boundary value problems may be troublesome because the presence of bound-
aries often introduces stability conditions that are both highly restrictive and often
difficult to analyze, one should note that for smooth solutions the results using spec-
tral methods are of a degree of accuracy that local approximation methods cannot
produce. For such solutions spectral methods can often achieve an exponential con-
vergence rate as compared to the algebraic convergence rate of finite difference or
finite element methods.
One may also think that the matrices in spectral methods are neither sparse nor
symmetric, in contrast to the situation in finite differences or finite elements where the
sparsity structure of the matrices simplifies the computation. However, the number
of discretization points required to achieve the expected accuracy using the spectral
method is much less than those required in finite difference or finite element methods,
and therefore the spectral method is still very efficient as compared to these other
two methods.
Acknowledgments E. Ngounda acknowledges the Agence National des Bourses du Gabon for
the financial support. Patidar’s research was supported by the South African National Research
Foundation.
References
1. Black, F., Scholes, M.: The pricing of options and corporate liabilities. J. Pol. Eco. 81, 637–659
(1973)
2. Boyle, P.: Options: a Monte-Carlo approach. J. Fin. Eco. 4, 323–338 (1973)
3. Cox, J.C., Rubinstein, M.: Options Markets. Prentice-Hall, Englewood Cliffs (1985)
4. Duffy, D.G.: Transform Methods for Solving Partial Differential Equations, 2nd edn. Chapman
& Hall/CRC, Boca Raton (2004)
5. Golub, G.H., van Loan, C.F.: Matrix Computations, 3rd edn. Johns Hopkins University Press,
Baltimore (1996)
6. Huang, W.Z., Sloan, M.D.: The pseudospectral method for solving differential eigenvalue
problems. J. Comput. Phys. 111, 399–409 (1994)
7. Hull, J.C.: Options Futures and Others Derivatives, 4th edn. Prentice-Hall International, New
Jersey (2000)
8. Ngounda, E., Patidar, K.C., Pindza, E.: Contour integral method for European options with
jumps. Commun. Nonlinear Sci. Numer. Sim. 18, 478–492 (2013)
9. Oosterlee, C.W., Leentvaar, C.C.W., Huang, X.: Accurate american option pricing by grid
stretching and high-order finite differences. Delft University of Technology, The Netherlands,
Technical Report (2005)
10. Talbot, A.: The accurate numerical inversion of Laplace transforms. IMA J. Appl. Math. 23(1),
97–120 (1979)
11. Tangman, D.Y., Gopaul, A., Bhuruth, M.: Numerical pricing of option using high-order compact
finite difference schemes. J. Comput. Appl. Math. 218(2), 270–280 (2008)
12. Tangman, D.Y., Gopaul, A., Bhuruth, M.: Exponential time integration and Chebyshev dis-
cretization schemes for fast pricing of options. Appl. Numer. Math. 58(9), 1309–1319 (2008)
13. Trefeten, L.N.: Spectral Method in MATLAB. SIAM, Philadelphia (2000)

14. Weideman, J.A.C., Reddy, S.C.: A MATLAB differentiation matrix suit. ACM Trans. Math.
Soft. 26, 465–519 (2000)
15. Weideman, J.A.C., Trefethen, L.N.: Parabolic and hyperbolic contours for computing the
Bromwich integral. Math. Comput. 76, 1341–1356 (2007)
16. Welfert, D.B.: Generation of pseudospectral differentiation matrices. I. SIAM J. Numer. Anal.
34, 1640–1657 (1997)
Quintic Hermite Fractal Interpolation
in a Strip: Preserving Copositivity
A.K.B. Chand and S.K. Katiyar
Abstract The notion of fractal interpolation provides a general framework which

includes traditional nonrecursive splines as special cases. In this paper, we describe
a procedure for the construction of quintic Hermite FIFs as α-fractal function cor-
responding to the classical quintic Hermite interpolant. In contrast to traditional
piecewise nonrecursive quintic Hermite interpolant, its fractal version has a sec-
ond derivative which is differentiable in a finite or dense subset of the interpolation
interval. This scheme offers an additional freedom over the classical quintic Her-
mite interpolants due to the presence of scaling factors. The elements of the iterated
function system are identified so that the class of α-fractal function f α reflects the
fundamental shape properties such as positivity, monotonicity, and convexity in addi-
tion to the regularity of f in the given interval. Using this general theory, an algorithm
for positivity of quintic Hermite FIF is presented. Finally, the algorithm for a quintic
Hermite fractal interpolants copositive with a given data set is prescribed.
Keywords Fractals · Iterated function system · Quintic Hermite fractal function ·

Positivity
MSC 28A80 · 41A20 · 65D10 · 26A48
1 Preamble
It is not ideal to use (piecewise) smooth interpolant with a desired precision when the
data has very irregular structure, for instance, real-world signal presented by climate
data, time series, financial series, and bioelectric recordings. Barnsley [1] introduced
A.K.B. Chand · S.K. Katiyar (B)

e-mail: sbhkatiyar@gmail.com
A.K.B. Chand
e-mail: chand@iitm.ac.in

DOI 10.1007/978-81-322-2485-3_38
464 A.K.B. Chand and S.K. Katiyar
fractal interpolation function (FIF) based on the theory of iterated function system
(IFS). A FIF is obtained as a fixed point of a suitable map defined on a space of
continuous functions. Fractal interpolation function (FIF) captures the irregularity
of data very effectively in comparison with the classical interpolants and provides an
effective tool for modeling data sampled from real-world signals which are usually
difficult using classical approach. FIFs are used to approximate naturally occurring
functions which show some sort of self-similarity under magnification. The fractal
continuation of an analytic function and fractal tiling from the attractor of one IFS
to the attractor of another are studied recently in [3, 4], respectively.
By imposing appropriate conditions on the scaling factors, Barnsley et al. [2]
observed that if the problem is of differentiable type, then the elements of the IFS may
be suitably chosen so that the corresponding FIF is smooth. Smooth FIF constitutes
an advance in the technique of approximation, since the classical methods of real data
interpolation can be generalized by means of smooth fractal techniques. However, it
is difficult to get all types of boundary conditions for fractal splines in this iterative
construction. Fractal splines with general boundary conditions have been studied in
[5–7] in simpler ways. By using suitable iterated function system (IFS), Barnsley
and Navascués have provided a method to perturb a continuous function so as to
yield a class of continuous functions f α , where α is a free parameter, called scaling
vector. For suitable values of scale vector α, the fractal functions f α simultaneously
interpolate and approximate f (see, for instance, [12–15, 19]). Further, the parameter
α can be used to modify or preserve properties of f .
Apart from suitable degree of smoothness, it may be desirable or even necessary
that the interpolant possesses some properties inherent in the data, depending on
practical background of the problem. The problem of searching a sufficiently smooth
function that preserves the qualitative shape property inherent in the data is generally
referred to as shape preserving interpolation/approximation. The shape properties are
mathematically expressed in terms of conditions such as positivity, monotony, and
convexity. Recently, Chand and collaborators have developed the shape preserving
aspects of the cubic Hermite fractal interpolation function and several rational cubic
FIFs with shape parameters (see, for instance, [7–9, 18]). Though these FIFs can
render shape preserving interpolants, the order of continuity is only C 1 . For enhanc-
ing the order of continuity to C 2 , the IFS parameters are to be selected so as to
satisfy a linear system governing the global C 2 -continuity. It is not known whether
the solution of this system is compatible with the shape preserving conditions. In the
current article we construct quintic Hermite FIF which is C 2 , and then constrain the
parameters (treating slopes and moments as parameters) so as to obtain positivity.
By a slight modification of the proposed method we shall also obtain C 2 -continuous
fractal spline copositive with given data.
Quintic Hermite Fractal Interpolation in a Strip … 465
2 Basic Facts
In this section we briefly recall requisite general material for our study. For a detailed
exposition reader may refer to [1, 2, 10, 16].
2.1 Quintic Hermite Interpolant
For r ∈ N, let Nr denote the subset {1, 2, . . . , r } of N. Let a set of data points D =
{(xi , yi ) ∈ I × R : i ∈ N N } satisfying x1 < x2 < · · · < x N , where I = [x1 , x N ] be
given. The local mesh spacing is h i = xi+1 −xi , and the slope of the linear interpolant
between the data points is i = yi+1h i−yi . A quintic Hermite function Q ∈ C 2 (I )

is uniquely determined by yi , di , and Di , where Q(xi ) = yi , Q (xi ) = di , and

Q (xi ) = Di , i ∈ N N = {1, 2, . . . , N }. The quintic Hermite interpolation defined
over the subinterval Ii = [xi , xi+1 ] has the form:

Di+1 − Di −3di − 3di+1 + 6i
Q i (x) = + (x − xi )5
2h i3 h i4

−2Di+1 + 3Di 8di + 7di+1 − 15i
+ + (x − xi )4
2h i2 h i3

Di+1 − 3Di −6di − 4di+1 + 10i
+ + (x − xi )3
2h i h i2
Di
+ (x − xi )2 + di (x − xi ) + yi . (1)
2
2.2 Nonnegativity/Nonpositivity of Quintic Hermite

Interpolant
The quintic Hermite interpolant preserves nonnegativity or positivity if

−5τi yi 5τi yi 8di 20yi −8di 20yi
≤ τi di ≤ and τi Di ≥ τi max − 2 , − 2 , (2)
hi h i−1 h i−1 h i−1 h i hi
where τi = sgn(yi ), i = 2, 3, . . . , N − 1, and endpoint derivatives are calculated

by arithmetic mean method and it is addressed in [10].
2.3 IFS for Fractal Functions
Let a set of data points D = {(xi , yi ) ∈ R2 : i ∈ N N } satisfying x1 < x2 < · · · <

x N , N > 2, be given. Set I = [x1 , x N ], Ii = [xi , xi+1 ] for i ∈ N N −1 . Suppose
L i : I → Ii , i ∈ N N −1 be contraction homeomorphisms such that
L i (x1 ) = xi , L i (x N ) = xi+1 . (3)
Let 0 < ri < 1, i ∈ N N −1 , and X := I × R. Let N − 1 continuous mappings

Fi : X → R be given satisfying:
Fi (x1 , y1 ) = yi , Fi (x N , y N ) = yi+1 , |Fi (x, y) − Fi (x, y ∗ )| ≤ ri |y − y ∗ |, (4)
where (x, y), (x, y ∗ ) ∈ X . Define functions wi : X → Ii × R, wi (x, y) =

(L i (x), Fi (x, y)) ∀ i ∈ N N −1 . It is known [1] that there exists a metric on R2 ,
equivalent to the Euclidean metric, with respect to which wi , i ∈ N N −1 , are con-
tractions. The collection I = {X ; wi : i ∈ N N −1 } is called an Iterated Function
System (IFS). Associated with the IFS I , there is a set valued Hutchinson map
N −1
W : H (X ) → H (X ) defined by W (B) = ∪ wi (B) for B ∈ H (X ) , where H (X )
i=1
is the set of all nonempty compact subsets of X endowed with the Hausdorff metric
h. The Hausdorff metric h completes H (X ). Further, W is a contraction map on the
complete metric space (H (X ), h). By the Banach Fixed Point Theorem, there exists
a unique set G ∈ H (X ) such that W (G) = G. This set G is called the attractor
or deterministic fractal corresponding to the IFS I . For any choices of L i and Fi
satisfying the conditions prescribed in (3)–(4) , the following theorem holds.
Proposition 1 (Barnsley [1]) The IFS {X ; wi : i ∈ N N −1 } defined above admits a
unique attractor G, and G is the graph of a continuous function g : I → R which
obeys g(xi ) = yi for i ∈ N N .
Definition 1 The aforementioned function g whose graph is the attractor of an IFS

is called a Fractal Interpolation Function (FIF) or a Self-referential function
corresponding to the IFS {X ; wi : i ∈ N N −1 }.
The above fractal interpolation function g is obtained as the fixed point of the Read-
Bajraktarević (RB) operator T on a complete metric space (G , ρ) defined as

(T h)(x) = Fi L i−1 (x), h ◦ L i−1 (x) ∀ x ∈ Ii , i ∈ N N −1 ,
where G := {h : I → R | h is continuous on I, h(x1 ) = y1 , h(x N ) = y N }

equipped with the metric ρ(h, h ∗ ) = max{|h(x) − h ∗ (x)| : x ∈ I } for h, h ∗ ∈ G .
It can be seen that T is a contraction mapping on (G , ρ) with a contraction factor
r ∗ := max{ri : i ∈ N N −1 } < 1. The fixed point of T is the FIF g corresponding to
the IFS I . Therefore, g satisfies the functional equation:

g(x) = Fi L i−1 (x), g ◦ L i−1 (x) , x ∈ Ii , i ∈ N N −1 , (5)
The most extensively studied FIFs in theory and applications so far are defined by
the mappings:
L i (x) = ai x + bi , Fi (x, y) = αi y + qi (x), i ∈ N N −1 . (6)
Here −1 < αi < 1 and qi : I → R are suitable continuous functions satisfying

(4). The parameter αi is called a scaling factor of the transformation wi , and α =
(α1 , α2 , . . . , α N −1 ) is the scale vector corresponding to the IFS. Let f ∈ C (I ) be a
continuous function and consider the case:
qi (x) = f ◦ L i (x) − αi b(x). (7)
Here b : I → R is a continuous map that fulfills the conditions b(x1 ) = f (x1 ),

b(x N ) = f (x N ), and b = f . This case is proposed by Barnsley [1] and Navascués
[11] as generalization of any continuous function. Here the interpolation data set is
{(xi , f (xi )) : i ∈ N N }. We define the α-fractal function corresponding to f in the
following:
Definition 2 The continuous function f α : I → R whose graph is the attractor of
the IFS defined by (6)–(7) is referred to as α-fractal function associated with f ,
with respect to b and the partition D.
According to (5), f α satisfies the functional equation:
f α (x) = f (x) + αi [( f α − b) ◦ L i−1 (x)] ∀x ∈ Ii , i ∈ N N −1 . (8)
Note that for α = 0, f α = f . Thus the aforementioned equation may be treated as

an entire family of functions f α with f as its germ. By this method one can define
fractal analogues of any continuous function.
2.4 Differentiable FIFs (Fractal Splines)
For a prescribed data set, a FIF with C r-continuity is obtained as the fixed point of
IFS (6), where the scaling factors αi and the functions qi are chosen according to the
following proposition.
Proposition 2 (Barnsley and Harrington [2]) Let {(xi , yi ) : i ∈ N N } be given
interpolation data with strictly increasing abscissae. Let L i (x) = ai x+bi , i ∈ N N −1 ,
satisfy (3) and Fi (x, y) = αi y + qi (x), i ∈ N N −1 , satisfy (4). Suppose that for some
integer r ≥ 0, |αi | ≤ κair , 0 < κ < 1, and qi ∈ C r (I ), i ∈ N N −1 . Let
(k) (k) (k)
αi y+qi (x) q1 (x1 ) q N −1 (x N )
Fi,k (x, y) = , y1,k = , y N ,k = , k = 1, 2, . . . , r.
aik a1k −α1 a N −1 −α N −1
k
If Fi−1,k (x N , y N ,k ) = Fi,k (x1 , y1,k ) for i = 2, 3, . . . , N − 1 and k = 1, 2, . . . , r ,

then the IFS {X ; (L i (x), Fi (x, y)) : i ∈ N N−1 } determines a FIF
g ∈ C r [x1 ,x N ],
and g (k) is the FIF determined by the IFS X ; L i (x), Fi,k (x, y) : i ∈ N N −1 for
k = 1, 2, . . . , r .
The equality proposed in the above proposition demands the resolution of systems of
equations. Sometimes the system has no solution, mainly whenever some boundary
conditions are imposed on the function (see [2]). However, for the special class of
IFS used to construct α-fractal function f α , the procedure can be easily carried out.
Assume |αi | < air , for all i ∈ N N −1 . Then to obtain a fractal perturbation f α ∈
C (I ) corresponding to a given f ∈ C r (I ), it is enough to find the conditions on
r
function b so that the IFS defined by (6) fulfills the conditions of the above theorem.
Assuming a uniform partition, Navascués and Sebastián have undertaken this in Ref.
[17] which is improved by Viswanathan et al. in Ref. [20] by allowing nonuniform
partition and unequal scale factors in different subintervals.
Proposition 3 Let f ∈ C r (I ). Suppose D = {x1 , x2 , . . . , x N } be an arbitrary
partition on I satisfying x1 < x2 < · · · < x N . Let |αi | < air , for all i ∈ N N −1 .
Further suppose that b ∈ C r (I ) fulfills b(k) (x1 ) = f (k) (x1 ), b(k) (x N ) = f (k) (x N )
for k = 0, 1, . . . , r . Then the corresponding fractal function f α is r -smooth, and
( f α )(k) (xi ) = f (k) (xi ) for i ∈ N N and k = 0, 1, . . . , r .
3 C 2 Quintic Hermite FIF Preserving Copositivity
In this section, first we shall find the strip condition for r th derivative of α-fractal
function, then we develop an algorithm for copositivity of C 2 -quintic Hermite FIF.
3.1 Strip Condition for r t h Derivative of α-Factal Function
In this subsection we shall provide conditions on the parameters so as to ensure that

r th derivative of a C r -continuous fractal function lies in a rectangle whenever its
classical counterpart f does so. Our proof is a modification of that given in [20]
where we take into account new requirements. However, we work with a slightly
more general obstacle for the r th derivative where r ∈ N ∪ {0} is arbitrary, whereas
[20] deals with constraining graph of f α or its first two derivatives within a rectangle
I × [0, M]. For a succinct presentation of the theorem, let us introduce the following
notation for a continuous function g defined on a compact interval J :
m(g; J ) = min {g(x) : x ∈ J }, M(g; J ) = max {g(x) : x ∈ J }.

Theorem 1 Let f ∈ C r (I ) be such that M1 ≤ f (r ) (x) ≤ M2 , for all x ∈ I , some

suitable constants M1 and M2 . The α-fractal function f α (cf. (5)) corresponding
to f satisfies M1 ≤ ( f α )(r ) (x) ≤ M2 , for all x ∈ I provided the base function
b ∈ C r (I ) satisfies b(k) (x1 ) = f (k) (x1 ), b(k) (x N ) = f (k) (x N ) for k = 0, 1, . . . , r,
and the scaling factors |αi | < air , for all i ∈ N N −1 obey the following additional
conditions:

air [M1 − m( f (r ) ; Ii )] −air [M2 − M( f (r ) ; Ii )]
max , ≤ αi
M2 − m(b(r ) ; I ) M(b(r ) ; I ) − M1

air [m( f (r ) ; Ii ) − M1 ] air [M2 − M( f (r ) ; Ii )]
≤ min , .
M(b(r ) ; I ) − M1 M2 − m(b(r ) ; I )
Proof With the stated conditions on the scale factors and the function b we can
ensure from Proposition 3 that corresponding fractal function f α is r -smooth. Note
that ( f α )(r ) is a fractal function corresponding to the IFS {X ; (L i (x), Fi,r (x, y)) :
i ∈ N N −1 } (see Proposition (3)), ( f α )(r ) (xi ) = f (r ) (xi ) and ( f α )(r ) is constructed
iteratively using the following functional equation:
αi α (r )
( f α )(r ) (L i (x)) = Fi,r (x, ( f α )(r ) (x)) = f (r ) (L i (x)) + r ( f ) − b(r ) (x).
ai
Therefore, to prove M1 ≤ ( f α )(r ) (x) ≤ M2 , for all x ∈ I , by the property of

the attractor of the IFS, it is enough to prove that M1 ≤ y ≤ M2 implies M1 ≤
Fi,r (x, y) ≤ M2 ∀i ∈ N N −1 .
First, let 0 ≤ αi < air . We note that M1 ≤ y ≤ M2 implies aαri M1 + f (r ) (L i (x))−
i
αi (r ) αi αi (r ) αi (r ) (L (x)) − αi b(r ) (x).
air b (x) ≤ air y + f (r ) (L i (x)) −
air b (x) ≤ air M2 + f i air
αi (r ) αi (r )
Therefore, our target M1 ≤ ar y + f (L i (x)) − ar b (x) ≤ M2 is achieved if
i i
M1 (1 − aαri ) ≤ f (r ) (L i (x)) − aαri b(r ) (x) ≤ M2 (1 − aαri ).
i i i
Note that f (r ) (L i (x)) ≥ m( f (r ) ; Ii ) and b(r ) (x) ≤ M(b(r ) ; I ) is true for all
x ∈ I . Therefore, M1 (1 − aαri ) ≤ f (r ) (L i (x)) − aαri b(r ) (x) holds if M1 (1 − aαri ) ≤
i i i
(r ) αi (r )
air [m( f (r ) ; Ii ) − M1 ]
m( f ; Ii ) − ar M(b ; I ). This inequality holds if αi ≤ .
i M(b(r ) ; I ) − M1
Similarly, f (r ) (L i (x)) ≤ M( f (r ) ; Ii ) and b(r ) (x) ≥ m(b(r ) ; I ) is true for all x ∈ I .
Therefore,
αi (r ) αi αi
f (r ) (L i (x)) − air b (x) ≤ M2 (1 − air ) holds if M( f (r ) ; Ii ) − air m(b(r ) ; I ) ≤
αi air [M2 − M( f (r ) ; Ii )]
M2 (1 − air ), which in turn holds if αi ≤ .
M2 − m(b(r ) ; I )
Now assume −ai < αi ≤ 0. In this case, M1 ≤ y ≤ M2 implies that
r
αi (r ) (L (x)) − αi b(r ) (x) ≤ αi y + f (r ) (L (x)) − αi b(r ) (x) ≤ αi M +
a r M 2 + f i ar ar i ar ar 1
i i i i i
f (r ) (L i (x)) − aαri b(r ) (x). Consequently, for M1 ≤ Fi,r (x, y) ≤ M2 , it is sufficient to
i
verify M1 − aαri M2 ≤ f (r ) (L i (x))− aαri b(r ) (x) ≤ M2 − aαri M1 . By appropriately using

i i i
the definition of m(b(r ) ; I ), M(b(r ) ; I ), m( f (r ) ; Ii ), M( f (r ) ; Ii ) on lines similar to
the first part, we get
a r [M1 − m( f (r ) ; Ii )] −air [M2 − M( f (r ) ; Ii )]
αi ≥ i and αi ≥ . Combination of the
M2 − m(b(r ) ; I ) M(b(r ) ; I ) − M1
obtained conditions on the scale factors completes the proof.
Consequence. If f : I = [a, b] → R is a C r -continuous r -convex function (i.e.,
f (r ) ≥ 0), then using the foregoing theorem we can select a scale vector α such
that f α ∈ C r (I ) and f α preserves r -convexity of f . Note that for r = 0, 1, 2,
r -convexity reduces to positivity, monotonicity, and convexity, respectively.
Remark 1 Let us confine to the positivity case (i.e., r = 0 and M1 = 0). If it is
enough to consider the nonnegative scale factors, then the following condition on
f ;Ii )
the scale factors ensures the nonnegativity of f α : 0 ≤ αi ≤ m(
M(b;I ) , where |αi | < 1
is assumed.
Remark 2 If f ∈ C (I ) is nonpositive (i.e., f (x) ≤ 0, for all x ∈ I ), then
we may construct f α satisfying f α (x) ≤ 0, for all x ∈ I by employing Theo-
rem 1. Taking r = 0 and M2 = 0 then the condition for M1 ≤ f α ≤ 0 can
M1 − m( f ; Ii ) −M( f ; Ii )
be obtained as: |αi | < 1 and max − , ≤ αi ≤
m(b; I ) M1 − M(b; I )
M1 − m( f ; Ii ) M( f ; Ii )
min , .
M1 − M(b; I ) m(b; I )
Remark 3 The aforementioned fractal scheme can be modified and extended to
produce piecewise-defined α-fractal function which is copositive with the given
f ∈ C (I ). For this, the interval I has to be subdivided into subintervals, say
I j , j = 1, 2, . . . , r in such a way that the function f is positive or negative through-
out the subinterval I j . In each of these subintervals I j , we take base function b j , and
a scaling factor α j so as to meet the specification in Theorem 1. Consequently, we
α
can get the fractal functions f j j that retain the nonpositivity/nonnegativity of the
functions f j = f |I j , j = 1, 2, . . . , r . Denoting α to be the r -rowed matrix whose
rows are the scaling factors α j , we define f α in a piecewise manner as follows:
f α |I j = f jα .
j
3.2 Quintic Hermite FIF as α-Fractal Function
Consider a set of data points D = {(xi , yi , di , Di ) : i ∈ N N }, where yi denotes the

function value, di , Di denote the first derivative and second derivative value of an
unknown function at the knot point xi , respectively. To construct the C 2 -quintic
Hermite FIF corresponding to D, one may employ the general theory given in Sect. 2.
The traditional nonrecursive C 2 -quintic Hermite interpolant corresponding to D can
be represented as:

h i2 (Di+1 − Di ) − 6h i (di + di+1 ) + 12(yi+1 − yi ) x − x1 5
Q(L i (x)) =
2 x N − x1

h i (−2Di+1 + 3Di ) + h i (16di + 14di+1 ) − 30(yi+1 − yi )
2
x − x1 4
+
2 x N − x1

h i (Di+1 − 3Di ) − h i (12di + 8di+1 ) + 20(yi+1 − yi )
2
x − x1 3
+
2 x N − x1

h 2 Di x − x1 2 x − x1
+ i + h i di + yi . (9)
2 x N − x1 x N − x1
L −1 (x)−x x−xi
For x ∈ [xi , xi+1 ], using i x N −x1 1 = xi+1 −xi , one can see that the above expression
coincides with the classical quintic Hermite interpolant (cf. (1)). According to pre-
scription in Theorem 1, we have to select the function b so as to obtain a C 2 -quintic
Hermite fractal perturbation for Q ∈ C 2 (I ). A natural choice of b is the two-point
quintic Hermite interpolants (with knots at x1 and x N ) corresponding to Q. That is,
(x N − x1 )2 (D N − D1 ) − 6(x N − x1 )(d1 + d N ) + 12(y N − y1 ) x − x1 5

b(x) = ( )
2 x N − x1

(x N − x1 )2 (−2D N + 3D1 ) + (x N − x1 )(16d1 + 14d N ) − 30(y N − y1 ) x − x1 4
+
2 x N − x1

2
(x N − x1 ) (D N − 3D1 ) − (x N − x1 )(12d1 + 8d N ) + 20(y N − y1 ) x − x1 3
+
2 x N − x1
D1
+ (x − x1 )2 + d1 (x − x1 ) + y1 . (10)
2
Using (9)–(10), we obtain the fractal function Q α ∈ C 2 (I ) corresponding to Q ∈

C 2 (I ) as:
Q α (L i (x)) = αi Q α (x) + Q(L i (x)) − αi b(x),

5 4 3
= αi Q α (x) + A2i xx−x 1
−x + Bi
2 x
x−x1
−x + Ci
2 x
x−x1
−x (11)
2 N 1 N 1 N 1
Di x−x1 x−x1
+ 2 x N −x1 + E i x N −x1 + Fi , ∀x ∈ I, i ∈ N N −1 ,
where
Ai = h i2 (Di+1 − Di ) − 6h i (di + di+1 ) + 12(yi+1 − yi )

− αi [(x N − x1 )2 (D N − D1 ) − 6(x N − x1 )(d1 + d N ) + 12(y N − y1 )],
Bi = h i2 (−2Di+1 + 3Di ) + h i (16di + 14di+1 ) − 30(yi+1 − yi )
− αi [(x N − x1 )2 (−2D N + 3D1 )+(x N − x1 )(16d1 +14d N )−30(y N − y1 )],
Ci = h i2 (Di+1 − 3Di ) − h i (12di + 8di+1 ) + 20(yi+1 − yi )
− αi [(x N − x1 )2 (D N − 3D1 ) − (x N − x1 )(12d1 + 8d N ) + 20(y N − y1 )],

Di = h i2 Di − αi D1 (x N − x1 )2 , E i = h i di − αi d1 (x N − x1 ), Fi = yi − αi y1 .
Note that the function Q α : I → R enjoys the interpolation conditions Q α (xi ) = yi ,

(Q α ) (xi ) = di and (Q α ) (xi ) = Di .
Remark 4 When αi = 0, for all i ∈ J , Q α reduces to the classical C 2 -quintic
Hermite interpolant Q given in (9). The diversity of options α in Q α allows us to
choose the best if a problem combined with approximation and optimization is to be

approached. The function (Q α ) : I → R may be nondifferentiable in a dense subset
of I or a finite subset of I depending on the values of α. Thus the perturbation allows

new geometric possibility: The graph of (Q α ) owns a fractal dimension which may
be used as an index for experimental signals.
3.3 Algorithm for Positivity of C 2 -Quintic Hermite FIF
Here we shall find conditions for the positivity of C 2 -quintic Hermite FIF. Recall
that we have viewed C 2 -quintic Hermite FIF as α-fractal function Q α corresponding
to quintic Hermite interpolant Q. Therefore, it is not hard to see that Theorem 1 in
conjunction with positivity condition for Q gives the following algorithm.
An Algorithm for positive C 2 -quintic Hermite FIF
Step 1: Compute the approximate derivative values di , Di , i ∈ N N and check if they
satisfy conditions given in (2).
Step 2: To get positive quintic Hermite interpolant by the modified derivative val-
ues obtained in Step 1. If not, modify according to −5y 5yi
h i ≤ di ≤ h i−1 and Di ≥
i

i −8di
max h8d i
i−1
− h20y
2 , hi − 20y
h2
i
for i = 2, 3, . . . , N − 1, and endpoint derivatives
i−1 i
are calculated by arithmetic mean method.
Step 3: Denote the derivative values obtained at the end of Step 2 by di , Di for
i ∈ N N . For Q and b, compute the constants m(b; I ) = min b(x), M(b; I ) =
x∈I
max b(x), m( f ; Ii ) = min f (x), M( f ; Ii ) = max f (x). Choose |αi | < ai and
x∈I x∈Ii x∈Ii

M1 − m( f ; Ii ) M2 − M( f ; Ii ) m( f ; Ii ) − M1 M2 − M( f ; Ii )
max ,− ≤ αi ≤ min ,
M2 − m(b; I ) M(b; I ) − M1 M(b; I ) − M1 M2 − m(b; I )
according to the prescription in Theorem 1.
Step 4: Input these derivative values chosen in Step 2 and the scaling parameters as
prescribed by Step 3 in the functional equation represented by (11) whereupon the
points of the graph of Q α are computed.
Remark 5 On similar lines we can get an algorithm for a nonpositive C 2 -quintic
Hermite FIF.
Next, we shed some light on to get interpolation of positive/negative data in
adjacent intervals by considering Remark 3. Let us discuss this with an example.
Consider a continuous function Φ defined on data set {(xi , yi ) : i = 1, 2, . . . , 7}
having zeros at the points x3 , x6 . It is nonnegative on [x1 , x3 ], nonpositive on [x3 , x6 ],

and nonnegative on [x6 , x7 ]. To get the desired interpolant, subdivide the interpolation
domain into three subintervals I1 = [x1 , x3 ], I2 = [x3 , x6 ], and I3 = [x6 , x7 ]
for our convenience. We can apply the developed algorithm to obtain positivity
preserving quintic Hermite FIF Q α1 on I1 . With proper renaming of the data points
1
if necessary, the negative quintic Hermite FIF algorithm can be applied to obtain a
negative quintic Hermite FIF Q α2 on I2 . Since I3 does not contain sufficient number
2
of knot points, iterations of the IFS code cannot produce any new points. To overcome
this problem, we introduce a new node say, (x6∗ , y6∗ ) in such a way x6 < x6∗ < x7 .
We apply the developed positivity preserving quintic Hermite FIF algorithm with
an arbitrary but shape consistent extra node to obtain a positive quintic Hermite FIF
Q α3 . Define a quintic Hermite FIF Q α in a piecewise manner such that Q α |I j = Q αj
3 j
α
for j = 1, 2, 3. Then Q is copositive with the original data.
4 Numerical Illustration
In this section, we will illustrate positivity preserving C 2 -quintic Hermite FIF with
some simple examples. Let us take a set of positive data D = {(0, 0.1), (0.4, 1),
(0.75, 2), (1, 5)}. Note that for the implementation of the C 2 -quintic Hermite FIF
one requires in input the values of the derivatives at the knot points. Therefore, in
the absence of other conditions/information, estimates of derivatives are necessary.
Values (rounded off to two decimal places) of di , Di , i = 1, 2, 3, 4 estimated using
the arithmetic mean method are d1 = 1.92, d2 = 2.57, d3 = 8.19, d4 = 15.80,
D1 = −6.07, D2 = 9.31, D3 = 8.19, and D4 = 36.48. The nonpositive C 2 -quintic
Hermite FIF Q α (L i (x)) is displayed in Fig. 1a. This illustrates the importance of the
positivity preserving C 2 -quintic Hermite FIF algorithm developed in the previous
section. Now improve the derivative values as prescribed in (2), i.e., d2 = 10.50,
d3 = 20.57, D2 = −73.60, D3 = −139.31, and endpoint derivatives as in Fig. 1a.
(a) (b) (c)

6 5 6
Data Data Data
Nonpositive quintic Hermite FIF 4.5 Positive quintic Hermite Interpolant Positive quintic Hermite FIF
5 5
4
4 3.5
4
3
3
2.5 3
2 2
1.5 2
1
1 1
0 0.5
−1 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 1 C 2 -Quintic Hermite FIF (the interpolating data points are given by the circles and the
relevant C 2 -quintic Hermite FIF by the solid lines)
We generate positive C 2 -quintic Hermite interpolant displayed in Fig. 1b by taking

these modified derivative values and assuming the values of scaling factors αi = 0,
for all i ∈ {1, 2, 3}. To illustrate our algorithm, we first construct the positive C 2 -
quintic Hermite interpolant Q as in Fig. 1b. After computing the maximum and
minimum value of Q and b, we select scaling factors according to the prescription
in Theorem 1. Input the derivative values and parameters values in the functional
equation represented by (11) whereupon the points of the graph of Q α are computed
in Fig. 1c.
In this paper, we have developed methods to identify the elements of the IFS so that
the α-fractal function f α retains fundamental shape property, namely, positivity, and
order of continuity inherent in the function f . For a data with prescribed or estimated
slopes and moments at knot points, the quintic Hermite FIF is constructed, which
generalizes the classical quintic Hermite interpolant. The considerable flexibility and
diversity offered by quintic Hermite FIF (see Remark 2) can be exploited in fields
of applications such as CAE, computer graphics, animation, visual simulation, and
image processing.
Acknowledgments The first author is thankful to the Department of Science & Technology, India
for the SERC DST Project No. SR/S4/MS: 694/10
References
1. Barnsley, M.F.: Fractal functions and interpolation. Constr. Approx. 2(4), 303–329 (1986)
2. Barnsley, M.F., Harrington, A.N.: The calculus of fractal functions. J. Approx. Theory 57(1),
14–34 (1989)
3. Barnsley, M., Vince, A.: Fractal continuation. Constr. Approx. 38(1), 311–337 (2013)
4. Barnsley, M., Vince, A.: Fractal tilings from iterated function system. Discrete Comput. Geom.
51(1), 729–752 (2014)
5. Chand, A.K.B., Kapoor, G.P.: Generalized cubic spline fractal interpolation functions. SIAM
J. Numer. Anal. 44(2), 655–676 (2006)
6. Chand, A.K.B., Navascués, M.A.: Generalized Hermite fractal interpolation. Rev. R. Acad. de
ciencias. Zaragoza 64(2), 107–120 (2009)
7. Chand, A.K.B., Viswanathan, P.: A constructive approach to cubic Hermite fractal interpolation
function and its constrained aspects. BIT Numer. Math. 53(4), 841–865 (2013)
8. Chand, A.K.B., Katiyar, S.K., Saravana Kumar, G.: A new class of rational fractal function for
curve fitting. Proceeding of Computer Aided Engineering CAE 2013, pp. 78–93, ISBN:No-
80689-17-3 (2013)
9. Chand, A.K.B., Vijender, N., Navascués, M.A.: Shape preservation of scientific data through
rational fractal splines. Calcolo 51, 329–362 (2013)
10. Dougherty, R.L., Edelman A.S., Hyman, J.M.: Nonnegativity-, monotonicity-, or convexity-
preserving cubic and quintic Hermite interpolation. Math. Comput. 52, 471–494 (1989)
11. Navascués, M.A.: Fractal polynomial interpolation. Z. Anal. Anwend. 25(2), 401–418 (2005)
12. Navascués, M.A.: Fractal approximation. Complex Anal. Oper. Theory 4, 953–974 (2010)
13. Navascués, M.A.: Reconstruction of sampled signals with fractal functions. Acta. Appl. Math.
110, 1199–1210 (2010)
14. Navascués, M.A.: Fractal Haar system. Nonlinear Anal. 74, 4152–4165 (2011)
15. Navascués, M.A., Chand, A.K.B.: Fundamental sets of fractal functions. Acta Appl. Math. 100,
247–261 (2008)
16. Navascués, M.A., Chand, A.K.B., Viswanathan, P., Sebastián, M.V.: Fractal interpolation func-
tions: a short survey. Appl. Math. 5, 1834–1841 (2014)
17. Navascués, M.A., Sebastián, M.V.: Smooth fractal interpolation. J. Inequal. Appl. (Article ID
78734), 20 (2004)
18. Viswanathan, P., Chand, A.K.B.: A fractal procedure for monotonicity preserving interpolation.
Appl. Math. Comput. 247, 190–204 (2014)
19. Viswanathan, P., Chand, A.K.B.: Fractal rational functions and their approximation properties.
J. Approx. Theory 185, 31–50 (2014)
20. Viswanathan, P., Chand, A.K.B., Navascués, M.A.: Fractal perturbation preserving fundamen-
tal shapes: bounds on the scale factors. J. Math. Anal. Appl. 419, 804–817 (2014)
Existence Result for Semilinear Fractional
Stochastic Evolution Inclusions Driven
by Poisson Jumps
P. Tamilalagan and P. Balasubramaniam
Abstract In this manuscript, the sufficient conditions are established for the exis-
tence of mild solutions of semilinear fractional stochastic evolution inclusions driven
by Poisson jumps in a Hilbert space. The results are obtained by using a fixed point
theorem for condensing multivalued map due to Martelli.
Keywords Fixed point theorem · Multivalued map · Mild solution · Poisson jump
MSC 26A33, 34A08, 34A60, 47H10
1 Introduction
The theory and applications of fractional differential equations [FDEs] have notable
contributions during the last few decades. Since the FDEs are the most powerful
toll for describing the real-life phenomena more preciously, thus there is a rapid
development in its applications [14]. In neurophysiology, the behavior of voltage
potentials of spatially extended neurons has been described by stochastic differential
equations [SDEs] driven by Poisson jump. SDEs driven by a Poisson process has
applications in various fields such as storage systems, queuing systems, economic
systems, and neurophysiology system. The study of SDE driven by a Poisson jump
has considerable attentions (see [9, 12, 15]). Differential inclusions serve as an
efficient tool in analysis of uncertain, nonlinear, and hybrid as well as switching and
time-variant systems. Random differential and integral inclusions play an important
role in characterizing many physical, biological, social, and engineering problems
[3, 4]. The applicability of fractional differential inclusions [FDIs] in modeling of
P. Tamilalagan · P. Balasubramaniam (B)

Department of Mathematics, Gandhigram Rural Institute—Deemed University,
Gandhigram 624 302, Tamil Nadu, India
e-mail: balugru@gmail.com
P. Tamilalagan
e-mail: mathematicstamil@gmail.com

DOI 10.1007/978-81-322-2485-3_39
478 P. Tamilalagan and P. Balasubramaniam
various practical and engineering systems insists its necessity. Thus very few authors
investigated the existence of solutions of FDIs [1, 5, 8]. However, to the best of
authors’ knowledge there is no work reported on semilinear fractional stochastic
differential inclusions driven by Poisson jump.
In this manuscript, we study the existence of mild solution for the following semi-
linear fractional stochastic evolution inclusions driven by Poisson jumps described by

d Jt1−α (x(t) − x(0)) ∈ [Ax(t) + F(x(ρ(t)))]dt +
L(t, x(t−), z)N(dt, dz), t ∈ J := [0, b]
Z
x(t) = φ(t), t ∈ J0 = [−r, 0], r > 0 (1)
where 0 < α < 1, Jt1−α is the (1 − α) order Riemann–Liouville fractional integral

operator, A : D(A) ⊂ H → H is the infinitesimal generator of analytic semigroup
of bounded linear operators {T (t), t ≥ 0} on a separable Hilbert space H with inner
product ·, · and norm · . ρ : [0, ∞) → [−r, ∞), r ≥ 0 be a suitable delay
function, x : [−r, b] → H and φ : J0 → H is the cadlag function such that φ(t) is
0
F0 measurable for all t ∈ J0 , Eφ(0)p < ∞ and −r Eφ(s)p ds < ∞, p ≥ 2.
Let K be another separable Hilbert space with inner product ·, · and norm ·K . Let

N(dt, dz) be a compensated Poisson random measure associated with the Poisson
point process k(·). Let L(K, H) be the space of all bounded linear operators from K
into H with norm · . Let F : H → 2H and L : J × H × (K − {0}) → H are the
appropriate functions to be defined later.
2 Preliminaries
In this section, we furnish some basic preliminaries, definitions, notations and lem-
mas, which are required to establish the main result. Let (Ω, F , {Ft }t≥0 , P) be a
complete probability space with the filtration {Ft }t≥0 , satisfying the usual conditions,
that is, right continuous and F0 containing all P-null sets. Let Lp (Ω, F , P; H) ≡
Lp (Ω, H) be the space of all p-integrable random variables with values in H, that are
measurable with respect to {Ft , t ∈ J}. Let J1 = [−r, b] and C = C(J1 , H) denote
the family of continuous H-valued stochastic processes {ξ(t) : t ∈ J1 } which are
Ft -measurable and ξ < ∞, where
1
ξ = ξ C = sup Eξ(t)p p .
t∈J1
Let Bσ (H) denotes the Borel σ -algebra of H. Let k(t), t ≥ 0 be a stationary

Ft -adapted and K-valued Poisson point process. The counting measure Nk defined by

Nk ((t1 , t2 ] × Λ)(ω) := IΛ (k(s)),
t1 < s≤t2
Existence Result for Semilinear Fractional Stochastic Evolution Inclusions Driven … 479
for any Λ ∈ Bσ (K) is called the Poisson random measure associated to the Poisson
by
point process k. Then define the measure N

N(dt, dz) := Nk (dt, dz) − π(dt, dz),
where π(dt, dz) is the compensator. We assume that k is σ -finite and stationary, there
exists a characteristic measure λ such that π(dt, dz) = dtλ(dz).
Definition 1 [14] The Riemann–Liouville fractional integral of order α > 0 for the
function x : J → H is defined by
t
1
Jtα x(t) = (t − s)α−1 x(s)ds
Γ (α) 0
and the Laplace transform of the Riemann–Liouville’s fractional integral is given by

∞
1
L{Jtα x(t)} = x̂(λ), where x̂(λ) = e−λt x(t)dt, Re(λ) > w.
λα 0
Definition 2 [10] The Riemann–Liouville fractional derivative of order 0 < α < 1

for the function x : J → H can be defined as
d 1−α
Dtα x(t) = J x(t).
dt t
Definition 3 [11] The Caputo fractional derivative of order 0 < α < 1 for the
function x can be defined in terms of Riemann–Liouville fractional derivative as
follows:
c α
Dt x(t) = Dtα (x(t) − x(0))
the Laplace transform of the Caputo fractional derivative is given by
L{c Dtα x(t)} = λα x̂(λ) − λα−1 x(0).
Definition 4 [11] The Mainardi’s function is defined by

∞
(−z)n
Mα (z) = , 0 < α < 1, z ∈ C,
n!Γ (1 − αn − α)
n=0
∞
it is clear that 0 Mα (r)dr = 1, 0 < α < 1 on the otherhand, Mα (z) satisfies the
following equality:
∞

α 1 α
Mα e−λr dr = e−λ
0 rα + 1 rα
∞ Γ (δ + 1)
and r δ Mα (r)dr = , δ > −1, 0 < α < 1.
0 Γ (αδ + 1)
Lemma 1 [4] Let H be a Hilbert space and Φ : H → BCC(H) be a u.s.c and

condensing map. If the set
U = {x ∈ H : λ̂x ∈ Φx for some λ̂ > 1}
is bounded, then Φ has a fixed point.
The following lemma is crucial in the proof of our main result.
Lemma 2 [4] Let J be a compact interval and H be a Hilbert space. Let F be a

multivalued map which is measurable for each u ∈ H, upper semicontinuous with
respect to u and for each fixed u ∈ H the set NF,u = {f ∈ L p (H) : f (t) ∈ F(u)
for a.e t ∈ J is nonempty. Also let Π be a linear continuous mapping from L p (J, H)
to C(J, H), then the operator
Π ◦ NF : C(J, H) → BCC(C(J, H)), x → (Π ◦ NF )(x) = Π (NF,x )
is a closed graph operator in C(J, H) × C(J, H).
Lemma 3 [2] For any p ≥ 2 there exists cp > 0 such that

t p
2p
t
E sup

H(s, z)N(ds, dz)
≤ cp E H(s, z) λ(dz)ds
2
0≤s≤t 0 Z 0 Z
t
+E H(s, z) λ(dz)ds .
p
0 Z
Motivated by [10, 11], we present the following definition of mild solution for (1).
Definition 5 An H-valued stochastic process x(t) satisfying Eφp < ∞ and f ∈

L(H) is a selection of F(x(ρ(t))) is called a mild solution of (1), if the following
conditions are satisfied
(i) x(t) is Ft -adapted and Cadlag,
(ii) x(t) satisfies the following integral equations
t
x(t) = Sα (t)φ(0) + (t − s)α−1 Tα (t − s)f (s)ds
0
t
+ α−1
(t − s) Tα (t − s)
L(s, x(s−), z)N(ds, dz)
0 Z
where
∞
Sα (t)x = Mα (r)T (t α r)xdr, t ≥ 0, x ∈ H (2)
0
∞
and Tα (t)x = αrMα (r)T (t α r)xdr (3)
0
The existence of solution for (1) is derived under the following assumptions:
H1 The analytic semigroup T (t) generated by A is compact for t > 0 and there
exists M > 0 such that
sup T (t) ≤ M, t ≥ 0 (4)

t≥0
from (2) and (4) we have Sα (t) ≤ M and from (3) and (4), we have Tα (t) ≤
M
Γ (α) (for details, see [10]).
H2 ρ : [0, ∞) → [−r, ∞), r ≥ 0 is a continuous function such that −r ≤ ρ(t) ≤
t for all t ≥ 0.
H3 F : H → BCC(H); u → F(u) is measurable for each u ∈ H, u.s.c with respect
to u and for each fixed u ∈ H the set
NF,u = {f ∈ L p (H) : f (t) ∈ F(u)for a.e t ∈ J}
is nonempty.
H4 EF(u)p = sup{Evp : v ∈ F(u)} ≤ η(t)ψ(Eup ) for almost all t ∈ J
and u ∈ H, where η ∈ L p (J, R+ ) and ψ : R+ → (0, ∞) is continuous and
increasing with
t
1
(t − s)(α−1)p η(s)ψ(Ex(ρ(s))p )ds = Λ.
r 0
H5 There exist positive constant ML such that, for x ∈ H

EL(t, x(t−), z)p λ(dz) ≤ ML x(t)p .
Z
For more details on this section, the reader can refer [6, 7, 13].
3 Main Result
Theorem 1 Assume that the hypotheses H1 –H5 are hold, then the initial value prob-
lem (1) has at least one mild solution on J1 = [−r, b], provide that
⎧ ⎛ p ⎞⎫
⎨ M p bp−1 M pc M b
p
2 αp− 2
+ M b αp−p + 1 ⎬
3p−1 Λ+ p
p ⎝ L L ⎠ < 1. (5)
⎩ Γ p (α) Γ (α) αp − p + 1 ⎭
Proof Consider the multivalued map Φ : C → 2C , defined by

⎧ ⎧
⎨ ⎨ φ(t); t ∈ [−r,0]
(Φx)(t) = h ∈ C : h(t) = Sα (t)φ(0) + 0 (t − s)α−1 Tα (t − s)f (s)ds,
t
⎩ ⎩ t α−1

+ 0 (t − s) Tα (t − s) Z L(s, x(s−), z)N(ds, dz), t ∈ J = [0, b]
where f ∈ NF,x = {f ∈ L p (H) : f (t) ∈ F(x(ρ(s))) for a.e t ∈ J}.

We shall prove that Φ is a completely continuous multivalued map u.s.c with
convex closed values. The proof will be given in several steps.
Step 1 Φx is convex for each x ∈ C. Indeed, if h1 and h2 belong to Φx, then there
exists f1 , f2 ∈ NF,x such that for each t ∈ J, we have
t
hi (t) = Sα (t)φ(0) + (t − s)α−1 Tα (t − s)fi (s)ds
0
t
+ α−1
(t − s) Tα (t − s)
L(s, x(s−), z)N(ds, dz) , i = 1, 2.
0 Z
Let 0 ≤ v ≤ 1, then for each t ∈ J, we have

t
(vh1 + (1 − v)h2 )(t) = Sα (t)φ(0) + (t − s)α−1 Tα (t − s)(vf1 (s) + (1 − v)f2 (s))ds
0
t
+ (t − s) α−1 Tα (t − s)
0 Z
since NF,x is convex, thus we have (vh1 + (1 − v)h2 ) ∈ Φx.

Step 2 Φ maps bounded sets into bounded sets in C. Indeed it is enough to show
that there exists a positive constant l such that for each h ∈ Φx, x ∈ Bq = {x ∈ C :
xp ≤ q} one has hp ≤ l. If h ∈ Φx, then there exists f ∈ NF,x such that for
each t ∈ J, by H1 –H5 and Lemma 3, we have
t p

α−1 T (t − s)f (s)ds
Eh(t)p ≤ 3p−1 ESα (t)φ(0)p + E (t − s) α
0
t p
α−1 T (t − s)

+E (t − s) α
0 Z

M p bp−1 t
≤ 3p−1 M p Eφ(0)p + (t − s)(α−1)p Ef (s)p ds
Γ p (α) 0

t 2 p
Mp α−1 2
+ p cp E (t − s) L(s, x(s−), z) λ(dz)ds
Γ (α) 0 Z

t p
Mp
+ p cp E (t − s)α−1 L(s, x(s−), z) λ(dz)ds
Γ (α) 0 Z

M p bp−1 t
≤ 3p−1 M p Eφ(0)p + (t − s)(α−1)p η(s)ψ E x(ρ(s))p ds
Γ p (α) 0
p ⎫
M p cp ML bαp−p + 1 q ⎬
p
M p cp ML2 bαp− 2 q
+ p + p
Γ (α)(αp − p + 1) Γ (α)(αp − p + 1) ⎭
t
M p bp−1
≤ 3p−1 M p Eφ(0)p + sup ψ(Ex(t) p) (t − s)(α−1)p η(s)ds
Γ p (α) t∈J 0
⎛ p ⎞⎫
M p cp q ML2 bαp− 2 + ML bαp−p + 1 ⎬
p
+ p ⎝ ⎠
Γ (α) αp − p + 1 ⎭
then for each h ∈ Φ(Bq ), we have

t
M p bp−1
hp ≤ 3p−1 M p Eφ(0)p + sup ψ(Ex(t)p) (t − s)(α−1)p η(s)ds
Γ p (α) t∈J 0
⎛ p ⎞⎫
M p cp q ML2 bαp− 2 + ML bαp−p + 1 ⎬
p
+ p ⎝ ⎠ := l.
Γ (α) αp − p + 1 ⎭
Step 3 Φ maps bounded sets into equicontinuous sets of C. For each x ∈ Bq and
h ∈ Φx there exists f ∈ NF,x , we have
Eh(t2 ) − h(t1 )p

t2

= E Sα (t2 )φ(0) + (t2 − s)α−1 Tα (t2 − s)f (s)ds
0
t2
+ (t2 − s)α−1 Tα (t2 − s)
0 Z
t1
− Sα (t1 )φ(0) − (t1 − s)α−1 Tα (t1 − s)f (s)ds
0
t1 p

− (t1 − s)α−1 Tα (t1 − s)
0 Z

≤ 11p−1 E [Sα (t2 ) − Sα (t1 )]φ(0)p
t1 −ε p

+ E (t2 − s) α−1
[Tα (t2 − s) − T α (t1 − s)]f (s)ds

0
t 1 p

+ E [(t2 − s)α−1 − (t1 − s)α−1 ]Tα (t1 − s)f (s)ds
t1 −ε
t 1 p

+ E (t2 − s)α−1 [Tα (t2 − s) − Tα (t1 − s)]f (s)ds
t1 −ε
t −ε p
1
+ E [(t2 − s) α−1
− (t 1 − s) α−1
]T α (t1 − s)f (s)ds

0
t 2 p

+ E
(t2 − s) α−1
T (t
α 2 − s)f (s)ds

t1
t1 −ε p

+ E
(t2 − s) α−1
[T α (t2 − s) − T α (t1 − s)] L(s, x(s−), z)
N(ds, dz)

0 Z
t 1 p

+ E
[(t2 − s)α−1 − (t1 − s)α−1 ]Tα (t1 − s)
t1 −ε Z
t 1 p

+ E
(t2 − s) α−1
[T (t
α 2 − s) − T (t
α 1 − s)] L(s, x(s−), z)
N(ds, dz)

t1 −ε Z
t1 −ε p

+ E
[(t2 − s) α−1
− (t1 − s) α−1
]T (t
α 1 − s) L(s, x(s−), z)
N(ds, dz)

0 Z
t
2 p
+E
(t2 − s)α−1 Tα (t2 − s)
L(s, x(s−), z)N(ds, dz) .
t1 Z
As t2 → t1 the right-hand side of the above inequality tends to zero, since the
compactness of T (t) for t > 0 implies the continuity in the uniform operator topology,
the equicontinuity for the cases t1 < t2 ≤ 0 and t1 ≤ 0 ≤ t2 are obvious. As a
consequence of the steps 2, 3 together with the Arzela Ascoli theorem it is concluded
that Φ : C → 2C is a compact multivalued map and therefore a condensing map.
Step 4 Φ has a closed graph.
Let xn → x∗ , hn ∈ Φxn and hn → h∗ . We shall prove that h∗ ∈ Φx∗ , hn ∈ Φxn
means that there exists fn ∈ NF,xn such that
t
hn (t) = Sα (t)φ(0) + (t − s)α−1 Tα (t − s)fn (s)ds
0
t
+ α−1
(t − s) Tα (t − s)
L(s, xn (s−), z)N(ds, dz) , t ∈ J.
0 Z
We must prove that there exists f∗ ∈ NF,x∗ such that

t
h∗ (t) = Sα (t)φ(0) + (t − s)α−1 Tα (t − s)f∗ (s)ds
0
t
+ α−1
(t − s) Tα (t − s)
L(s, x∗ (s−), z)N(ds, dz) , t ∈ J.
0 Z
Now, for every t ∈ J, L is continuous, we have

t

hn (t) − Sα (t)φ(0) − (t − s)α−1 Tα (t − s)
L(s, xn (s−), z)N(ds, dz)

0 Z

t

− h∗ (t) − Sα (t)φ(0) − α−1
(t − s) Tα (t − s)
L(s, x∗ (s−), z)N(ds, dz) →0
0 Z
as n → ∞.
Consider the linear continuous operator Π : L p (H) → C(J, H),

t
f → Π (f )(t) = (t − s)α−1 Tα (t − s)f (s)ds
0
from Lemma 2, it follows that Π ◦ NF is a closed graph operator. Moreover, we have

t
hn (t) − Sα (t)φ(0) − (t − s)α−1 Tα (t − s)
L(s, xn (s−), z)N(ds, dz) ∈ Π (NF,xn ).
0 Z
Since xn → x∗ , it follows from Lemma 2 that

t
h∗ (t) − Sα (t)φ(0) −
(t − s)α−1 Tα (t − s) L(s, x∗ (s−), z)N(ds, dz)
0 Z
t
= (t − s)α−1 Tα (t − s)f∗ (s)ds
0
for some f∗ ∈ NF,x∗ . Therefore Φ is completely continuous multivalued map, u.s.c

with convex closed values. In order to prove that Φ has a fixed point, we need one
more step.
Step 5 Φ has a solution, if the set U = {x ∈ C : λ̂x ∈ Φx, for some λ̂ > 1} is
bounded, such that for x ∈ U, Ex(t)p ≤ r. Assume that it is not true. Let x ∈ U
be a solution for λ̂x ∈ Φx, for some λ̂ > 1 with Ex(t)p > r. Then there exists
f ∈ NF,x , we have
t
x(t) = λ̂−1 Sα (t)φ(0) + λ̂−1 (t − s)α−1 Tα (t − s)f (s)ds
0
t
+ λ̂ −1 α−1
(t − s) Tα (t − s)
L(s, x(s−), z)N(ds, dz) , t ∈ J.
0 Z
By H1 –H2 and H4 –H5 , we have

p p
−1 −1 t
Ex(t) ≤ 3
p
E λ̂ Sα (t)φ(0) + E λ̂
p−1 (t − s) Tα (t − s)f (s)ds
α−1

0
t p
−1
+E
λ̂ (t − s)α−1 Tα (t − s)

0 Z

M p bp−1 t
r ≤ 3p−1 M p Eφ(0)p + (t − s)(α−1)p η(s)ψ(Ex(ρ(s))p )ds
Γ p (α) 0
⎛ p ⎞⎫
M p cp r ML2 bαp− 2 + ML bαp−p + 1 ⎬
p
+ p ⎝ ⎠
Γ (α) αp − p + 1 ⎭
by diving both sides of the above inequality by r and taking r → ∞, we obtain

⎧ ⎛ p ⎞⎫
⎨ M p bp−1M pc M b
p
2 αp− 2
+ M b αp−p + 1 ⎬
p ⎝ L L ⎠
1 ≤ 3p−1 Λ+ p
⎩ Γ p (α) Γ (α) αp − p + 1 ⎭
which is contradiction to our assumption (5), thus Φ has a solution if U is bounded.

4 Conclusion
In this manuscript, the Poisson jumps are incorporated with the stochastic differential
inclusions lead to new systems in the world of fractional calculus. The existence of
mild solutions for semilinear fractional stochastic evolution inclusions driven by
Poisson jumps has been studied in a Hilbert space by means of fractional calculus
and the fixed point theorem for condensing multivalued map due to Martelli.
Acknowledgments The work of authors are supported by Council of Scientific and

Industrial Research, Extramural Research Division, Pusa, New Delhi, India under the grant No.
25(0217)/13/EMR-II.
References
1. Ait Dads, E., Benchohra, M., Hamani, S.: Impulsive fractional differential inclusions involving
the Caputo fractional derivative. Fract. Calc. Appl. Anal. 12, 15–38 (2009)
2. Applebaum, D.: Levy Processes and Stochastic Calculus. Cambridge University Press,
Cambridge (2009)
3. Balasubramaniam, P., Ntouyas, S.K.: Controllability for neutral stochastic functional differen-
tial inclusions with infinite delay in abstract space. J. Math. Anal. Appl. 324, 161–176 (2006)
4. Balasubramaniam, P., Ntouyas, S.K., Vinayagam, D.: Existence of solutions of semilinear
stochastic delay evolution inclusions in a Hilbert space. J. Math. Anal. Appl. 305, 438–451
(2005)
5. Chang, Y.K., Nieto, J.J.: Some new existence results for fractional differential inclusions with
boundary conditions. Math. Comput. Model. 49, 605–609 (2009)
6. Da Prato, G., Zabczyk, J.: Stochastic Equations in Infinite Dimensions. Cambridge University
Press, Cambridge (1992)
7. Deimling, K.: Multivalued Differential Equations. De Gruyter, Berlin/New York (1992)
8. El-Sayed, A.M.A., Ibrahim, A.G.: Multivalued fractional differential equations. Appl. Math.
Comput. 68, 15–25 (1995)
9. Hausenblas, E.: SPDEs driven by Poisson random measure with non Lipschitz coefficients:
existence results. Probab. Theory Relat. Fields 137, 161–200 (2007)
10. Li, K., Peng, J.: Controllability of fractional neutral stochastic functional differential systems.
Z. Angew. Math. Phys. 65, 941–959 (2014)
11. Li, K., Peng, J., Gao, J.: Existence results for semilinear fractional differential equations via
Kuratowski measure of noncompactness. Fract. Calc. Appl. Anal. 15, 591–610 (2012)
12. Luo, J., Taniguchi, T.: The existence and uniqueness for non-Lipschitz stochastic neutral delay
evolution equations driven by Poisson jumps. Stoch. Dyn. 9, 135–152 (2009)
Springer-Verlag, Berlin (1983)
14. Podlubny, I.: Fractional Differential Equations. Academic Press, San Diego (1998)
15. Taniguchi, T.: The existence and asymptotic behaviour of solutions to non-Lipschitz stochastic
functional evolution equations driven by Poisson jumps. Stochastic 82, 339–363 (2010)
A Numerical Investigation of Blood Flow
in an Arterial Segment with Periodic
Body Acceleration
Mamata Parida and Ameeya K. Nayak
Abstract A fluid convection flow driven by a periodic body acceleration with ther-
mal stratification in an arterial cross section filled with an incompressible Newtonian
fluid (blood) is studied. A two-dimensional computational visualization technique
is used to study the steady flow behavior of the viscous electrically conducting
fluid flow. The driving force is generated by putting an external magnetic field in the
transverse direction of the flow. A numerical method based on the pressure correction
iterative algorithm (SIMPLE) is adopted to compute the flow field and temperature
along the arterial cross section. Variation over a wide range of parameters such as
Prandtl number, Hartmann number, and Womersley number have been investigated
for the flow and heat transfer characteristics.
Keywords Body acceleration · Magnetic field · Finite volume method
1 Introduction
In recent years, the study of magnetohydrodynamic (MHD) flow of blood has gained
the attention of many researchers because of its wide range of physiological applica-
tions. Magnetohydrodynamic fluids are those physiological fluids which are electri-
cally conducting. The hydrodynamic property of blood flow through arteries plays
a vital role for understanding the function of the cardiovascular system under nor-
mal and diseased conditions. Several studies analytical as well as experimental have
been carried out to analyze the flow of blood through arteries. It was observed by
some investigators [2, 12, 15] that under certain conditions blood exhibit viscoelastic
M. Parida (B)
Department of Mathematics, Utkal University, Bhubaneswar 751 004, India
e-mail: mamatakmru@gmail.com
A.K. Nayak
Roorkee 247 667, India
e-mail: ameeyanayak@gmail.com

DOI 10.1007/978-81-322-2485-3_40
490 M. Parida and A.K. Nayak
behavior which may be due to the viscoelastic properties of the individual red cells
and the internal structures formed by cellular interactions. Fung et al. [3] have stud-
ied blood flow mechanics in the arteries of different sizes. An experimental work
was done by Taylor and Draney [14] for quantifying the blood flow velocity and
pressure field in human artery. Blood as a dilute suspension of spherical particles
which, though rigid, were free to move with the fluid and rotate under the influence
of shearing force was studied by Jones [5]. A theoretical and experimental study was
conducted by Berger et al. [1] to give an idea of the pressure drop and heat exchange
in the fluid when it is subjected to move along a curved path. Several attempts have
been made [8, 10, 16] to study the effect of magnetic field on the blood flow in the
arteries in various physiological conditions. Heat transfer and fluid flow character-
istics of blood in multistenosed arteries with the effect of magnetic field have been
investigated in [13].
In the present paper, we have adopted a numerical method to study the character-
istics of blood flow and heat transfer through a rectangular duct under the influence
of periodic body acceleration and in the presence of transverse magnetic field since it
has large-scale applications in biomedical devices [7]. Blood is considered to be New-
tonian, viscous, incompressible, and electrically conducting fluid. Blood behaves as
a non-Newtonian fluid in very narrow arteries, whereas its behavior is Newtonian in
most of the arteries. The result of computation thus obtained for the physical quan-
tities velocity, pressure, temperature, and stream function are presented graphically.
The effects of magnetic field and body acceleration on axial blood flow, temperature,
and pressure have been studied.
2 Problem Formulation and Numerical Method
2.1 Physical Configuration
We consider the mechanically driven flow of an incompressible electrically conduct-

ing Newtonian fluid within a long rectangular channel of length L, width w, and
height h shown in Fig. 1a. The flow is assumed to be laminar and the binary fluid is
assumed to be incompressible and the flow may be driven by stretching of the channel
walls. Density variation of the fluid follows Boussinesq’s assumption and changes
with temperature only. Initially, the binary mixture is considered to be at rest with a
uniform temperature T∞ and constant density everywhere. An uniformly distributed
Fig. 1 Schematic of the flow (a) (b)

configuration of the artery
h
w y
L x
A Numerical Investigation of Blood Flow ... 491
external magnetic field of strength B0 is applied along the transverse direction of

the flow. The length, width, and height of the channel are along z, x and y direc-
tions, respectively. For computational purposes we have considered a channel cross
section, where height is considered to be equal length as of width of the channel to
study the flow properties and it is represented by a two-dimensional plane as shown
in Fig. 1b.
2.2 Model Equations
As the fluid is assumed to be Newtonian and its density is supposed to be constant,

except the gravitational force term in the Navier–Stokes equation, where it varies
linearly with the local temperature fraction [4] and is given by,
ρ(T ) = ρ0 [1 − βT (T − T∞ )]
where ρ0 is the density of the undisturbed fluid and βT is the volumetric coefficient
of thermal expansion. Initially, the electrically conducting fluid is considered to be
at rest. The flow is assumed to have periodic body acceleration given by
G (t ) = a cos(ωb t + φg ). (1)
where a , ωb and φg denote the amplitude, frequency, and phase difference of body
acceleration. We now introduce the nondimensional variables defined by
x y u v p T − Tw
x= , y = ,u = ,v = ,p= , t = t ω, θ = (2)
h h ωh ωh μω T∞ − Tw
Here ω is the frequency of pulse and Tw is the temperature of the wall. The governing
Navier–Stokes equations and heat transport equations in nondimensional form with
the Boussinesq-fluid assumption are given by
∂u ∂v
+ =0 (3)
∂x ∂y
2
∂u ∂u ∂u 1 ∂p 1 ∂ u ∂2u H2 1
+u +v =− 2 + 2 + − u + 2 G(t) (4)
∂t ∂x ∂y α ∂x α ∂x 2 ∂y 2 α 2 α

∂v ∂v ∂v 1 ∂p 1 ∂2v ∂2v
+u +v =− 2 + 2 + 2 (5)
∂t ∂x ∂y α ∂y α ∂x 2 ∂y
2
∂θ ∂θ ∂θ 1 ∂ θ ∂2θ H 2 Ec 2
+u +v = 2 + + u (6)
∂t ∂x ∂y α Pr ∂x 2 ∂y 2 α2

where α = h ωρμ , is the Womersley number, μ is the coefficient of viscosity, H =
μC
σ
h B0 μ is the Hartmann number, σ is the electrical conductivity, Pr = κ0p is
the Prandtl number, C p is the specific heat at constant pressure, κ0 is the thermal
conductivity, Ec = C p (Tωwh−T∞ ) is the Eckert number and G(t) = a cos(bt + φg ),
2 2
ρh
a = ωμ a , b = ωωb . Here in Eq. 4 the last two terms are due to the effect of applied
magnetic field and the body acceleration. Initially (t = 0), the fluid is considered to
be at rest and with an uniform temperature θ = 0.
∂θ
Boundary conditions: t > 0, u = 0, v = 0, ∂x = 0.
On the sidewalls (x = 0; x = 1)u = v = 0; θ = 0 on the lower lid (y = 0)
u = v = 0; θ = 1 on the upper lid (y = 1).
The local Nusselt number of the upper and lower plates of the channel is obtained
by calculating the temperature gradient on the plates from the relation
h ∂θ
Nu = − | y=0 (7)
Tw − T∞ ∂ y
The stream function ψ has been computed from the velocity components by using
the definition
∂ψ ∂ψ
u= v=− (8)
∂y ∂x
2.3 Numerical Methods
In order to tackle the model nonlinear partial differential equations numerically, the
method of Newton’s linearization technique is applied, that is, when the values of
the dependent variables at the nth iteration are known, the corresponding values of
variables at the next iteration can be obtained by applying the Newton’s linearization
method as:
Vi n+1 = Vi n + ΔVi n
where V stands for u, v, and θ; ΔVi n represents the error at the nth iteration and i is
the grid index. To find the numerical solution of the governing fluid flow equations
together with the specified boundary conditions we opt a numerical method using
control volume approach. This method involves integrating the continuity, momen-
tum, and energy equations over a specified control volume on a staggered grid. In
the staggered grid arrangement, the velocity components are stored at the midpoints
of the cell faces to which they are normal and the physical quantities such as the
pressure and temperature are placed at the cell center. The discretized form of the
governing equations is obtained by integrating over each of the control volumes
using the finite volume method. The u-momentum equation after integration over
the u-control volume becomes
Fig. 2 Schematic diagram w e

of u control volume for
uW uP uE
u [6, 11]
Fe u e − Fw u w + Fn u n − Fs u s = b (9)
where Fe is the nonlinear coefficient of u e and b contains the source terms and time-
derivative terms. The convective term at any interface is estimated by a quadratic
interpolation of u. For example, at the east face (Fig. 2) we have

3 3 1
ue = u E + u P − uW i f Fe > 0 (10)
8 4 8

3 3 1
ue = u E + u P − u EE i f Fe < 0 (11)
4 8 8
which can be summarized as

3 3 1 3 3 1
Fe u e = u E + u P − u W [[Fe , 0]] − u E + u P − u EE [[−Fe , 0]] (12)
8 4 8 4 8 8
The v-momentum equation is calculated in a similar manner. A third-order accu-

rate QUICK (quadratic upstream interpolation for convective kinematics) scheme is
employed to discretize the convective terms in the Navier–Stokes equations. A third-
level implicit scheme is used for discretization of time derivatives. The pressure
correction-based iterative algorithm SIMPLE (Semi-Implicit Method for Pressure
Linked Equations) is used for solving the discretized equations. We have obtained
a time-independent numerical solution which is convergent by advancing the flow
field variables through a sequence of shorter time step Δt = 0.001. For the range
of parameter values considered here, the flow field achieves a steady state after a
transient state, and this steady state is independent of the initial conditions. In all
the numerical computations that have been carried out and presented in the form of
figures we have adopted the steady state condition considering time t = 0.5. The
discretization of the governing Navier–Stokes equations and heat transfer equation
results in a system of algebraic equations of the form
Aφ φi−1,
n+1 φ n+1 φ n+1
j + B φi, j + C φi+1, j = D
φ
(13)
where φ denotes u, v, and θ; Aφ , B φ and C φ gives the coefficient matrix and D φ

gives the pressure and source terms for φ at time level n and time derivative terms.
This system of algebraic equations can thus be written in matrix form with the
coefficient matrix as a tridiagonal matrix. Due to coupling of energy equation with
the momentum equations, the system of algebraic equations is solved through a block
elimination method. Convergence criteria is employed of the form
|φi, j n+1 − φi, j n | < ε (14)

N uU
|1.0 − |<δ (15)
NuL
Here i and j denote the cell indices, n is the time level, φ stands for u, v, or θ and
N u U and N u L are the area averaged Nusselt number on the upper and lower lids,
respectively, and the value of ε is considered to be 10−4 and that of δ is 10−3 .
3 Results and Discussion
We have investigated the blood flow phenomenon and heat transfer through an arterial
segment under the effect of body acceleration as well as an external magnetic field.
The physiological applicable data used for computation of numerical results are
collected from the existing literatures [9] and are listed as: α = 3.0, H = 1.0,
Pr = 21.0, b = 1.0, a = 1.0, φg = 0.0, h = 1.0, Ec = 0.0002, T∞ = 310.0 K,
ρ = 1050.0 kg/m3 , σ = 0.8 s/m. For computation purposes we have taken these
values, however, any deviation from the listed values has been mentioned inside the
figures. In order to validate the present mathematical model we have compared our
results with that of Misra et al. [8] produced in Fig. 3. For the purpose of comparison,
both the studies have been naturally brought to the same platform.
The distribution of axial velocity is presented in Fig. 4. Figure 4a depicts that the
magnetic field parameter brings quantitative as well as qualitative changes in velocity
profile. It can be observed that the velocity decreases as the magnetic strength para-
meter increases. Figure 4b includes the axial profile for various values of Womersley
number. The variation of flow velocity can be determined from the inset figure. As
we move toward the core region from the inflow region we find that the velocity
decreases with increasing Womersley number. However, as the flow velocity attains
a saturation point in the core of the channel the velocity increases as α increases.
The nondimensional pressure distribution is presented in Fig. 5 for different values
of Hartmann number and Womersley number. We observe that during the change of
Hartmann number Fig. 5a the pressure variation is almost constant along the outflow
region of the artery. However for H = 1, initially the pressure shows a negative
variation as the reverse flow profile is observed. The pressure variation is found to
be optimum in case of α = 2. Figure 6 shows a variation in the dimensionless tem-
perature along the flow axis. For different Prandtl numbers and Womersley numbers,
we find that the temperature remains almost invariant close to the channel walls.
Figure 6a presents the temperature variation for different values of Prandtl number
along the cross section of an artery. For large values of Pr the effects of temperature
70
Present result
Results due to Misra and Shit (2007)
60
50
40
u
30
20
10
0
0 0.25 0.5 0.75 1
x
Fig. 3 Comparison of axial velocity profile for present solution in the absence of any magnetic
field with that of Misra and Shit [8] when α = 4 and Re = 90
(a) 2.5 (b) 2.5

α=2
2 α=3
2 α=4
1.5 2.038
1.5
1
u
1 2.03
H=1 0.5
H=3
0.5 H=5 0
2.022
0.5
0 -0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X X
Fig. 4 Axial distribution of velocity profiles along the centerline symmetry, a for various values
of H when Pr = 21 and α = 3, b for various values of α when Pr = 21 and H = 1
variation decreases gradually. In Fig. 6b we find that when the Womersley number
decreases, there is an increase in the variation of temperature profile. In Figs. 7, 8,
and 9 we present the average temperature variation or rate of heat transfer (N u)
along the lower and upper wall of the artery. It is observed from Fig. 7 that the rate of
heat transfer decreases with the increase of the Prandtl number Pr. The rate of heat
transfer decreases with the increase of Hartmann number H as observed from Fig. 8.
It is interesting to mention here that at the lower boundary the N u rapidly decreases
with the increase of H upto the core and beyond which, no significant change is
noticed. Hence the rate of heat transfer can be increased whenever necessary by the
(a) 0.06 (b) 0.04

H=1 α=2
0.05 0.03 α=3
H=3
0.04 H=5 α=4
0.02
0.03
0.02 0.01
p
p
0.01 0
0
-0.01
-0.01
-0.02 -0.02
-0.03 -0.03
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X X
Fig. 5 Distribution of dimensionless pressure along x-axis a for various values of H when, Pr = 21
and α = 3, and b for Pr = 21, H = 1, and α = 2, 3, 4
(a) 1 (b) 1
0.9 Pr=21 α=2
0.8 Pr=14 0.8 α=3
0.7 Pr=7 α=4
0.6 0.6
0.5
θ
0.4 0.4
0.3
0.2 0.2
0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X X
Fig. 6 Distribution of dimensionless temperature along x-axis a for various values of Pr when,
H = 1 and α = 3, and b for Pr = 21, H = 1, and α = 2, 3, 4
(a) 12 (b) 0.4

11 Pr=21 0.35
Pr=14 Pr=7
10
Pr=7 0.3 Pr=14
9 Pr=21
8 0.25
Nu
Nu
7 0.2
6 0.15
5
0.1
4
3 0.05
2 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
X X
Fig. 7 The average heat transfer rate versus x-axis, where H = 1, α = 3, and Pr is chosen to be,
7, 14, 21, a along lower boundary and b along upper boundary
application of the certain magnetic field strength. In Fig. 9a, b the rate of heat transfer
for various values of Womersley numbers are presented for fixed values of Pr and
H . The heat transfer rate is increasing with the increase of Womersley number at
(a) 12 3.65
(b) 0.4
H=5 H=5
11 0.35
3.55 H=3 H=3
10 3.45 H=1 0.3 H=1
9 3.35 0.25
8 3.25
Nu
Nu
0.2
7 3.15
6
0.2 0.3 0.4 0.15
5 0.1
4 0.05
3 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
X X
Fig. 8 The average heat transfer rate versus x-axis, for different values of H , where Pr is chosen
to be =21 and α is 3, a along lower boundary and b along upper boundary
(a) 14 (b) 0.4

α=2
12 0.35 α=3
α=2 0.3 α=4
10 α=3
α=4 0.25
8
Nu
Nu
0.2
6 0.15
4 0.1
0.05
2
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
X X
Fig. 9 The heat transfer rate along the arterial wall for different values of α, a along the lower
boundary and b along the upper boundary, when Pr = 21 and H = 1
(a) (b) 0.6

0.9 H=1
0.8 0.5 H=3
H=5
0.7 -0.0942531
0.4
0.6 -0.0754025
0.5
y
0.3
ψ
-0.0565519
0.4 -0.0377013
0.3 -0.0251342 0.2

0.2 -0.0125671
0.1
0.1 -0.00293735
00
16
01
75
-0.0
0 0
0 0.25 0.5 0.75 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x X
Fig. 10 a Contour plots, b surface plot of stream function along the cross section of the artery for
H = 1, when Pr = 21 and α = 3
the lower arterial wall, but the upper wall variation is large at the initial position but
at the outflow region the variation is almost constant. In Fig. 10 we have presented
the stream function variation for constant parametric values of Hartmann number,
Womersley number, and Prandtl number. From Fig. 10 a we can observe that the
stream function is symmetric about the midpoint of the channel where the maximum
value of the function is attained.
4 Conclusion
A numerical approach for simulation of the blood flow in an artery under the com-
bined effect of periodic acceleration and an external magnetic field has been studied.
We have made an attempt to examine the effect of the Prandtl number, the Hartmann
number, and the Womersley number on the flow and heat transport characteristic of
blood. Some graphical presentations of the computed results have been performed.
The study provides the fact that the axial velocity is largely influenced by the mag-
netic field parameter. This is due to the fact that when the biomagnetic fluid (blood)
is subjected to a magnetic field, the action of magnetization introduces an orienta-
tion of the blood charged ions with the magnetic field. The Prandtl number and the
Womersley number also play a major role on the velocity as well as on the pressure
and temperature distribution.
References
1. Berger, S.A., Talbot, L., Yao, L.S.: Flow in curved pipes. Annu. Rev. Fluid Mech. 15, 461–512
(1983)
2. Fukada, E., Kaibara, M.: Viscoelastic study of aggregation of red blood cells. Biorheology 17,
17782 (1980)
3. Fung, Y.C., Zweifach, B.W.: Microcirculation, mechanics of blood flow in capillaries. Annu.
Rev. Fluid Mech. 3, 189–210 (1971)
4. Gebhart, B., Pera, L.: The nature of vertical natural convection flows resulting from the com-
bined buoyancy effects of thermal and mass diffusion. Int. J. Heat Mass Transf. 14, 2025–2050
(1971)
5. Jones, R.T.: Blood flow. Annu. Rev. Fluid Mech. 1, 223–244 (1969)
6. Leonard, B.P.: A stable and accurate convective modeling procedure based on quadratic
upstream interpolation. Comput. Meth. Appl. Mech. Eng. 19, 59–98 (1979)
7. Lima, R., Imai, Y., Ishikawa, T., Takeda, M., Tanaka, S., Tsubota, K., Wada, S., Yamaguchi,
T.: In vitro blood flow in a rectangular PDMS microchannel experimental observations using
a confocal micro-PIV system. Biomed. Microdevices 10(2), 153–167 (2008)
8. Misra, J.C., Shit, G.C.: Effect of magnetic field on blood flow through an artery: A numerical
model. Comput. Technol. 12, 3–16 (2007)
9. Misra, J.C., Chandra, S., Kundu, P.K., Shit, G.C.: Hydromagnetic flow and heat transfer of a
second-grade viscoelastic fluid in a channel with oscillatory stretching walls. J. Eng. Math. 69,
91–100 (2011)
10. Ogulu, A., Amos, E.: Modeling pulsatile blood flow within a homogeneous porous bed in the
presence of a uniform magnetic field and time-dependent suction. Int. Commun. Heat Mass
Transf. 34, 989–995 (2007)
11. Shyy, W., Thakur, S., Wright, J.: Second-order upwind and central difference schemes for
recirculating computation. AIAA J. 30, 923–932 (1992)
12. Stoltz, J.F., Lucius, M.: Viscoelasticity and thixotropy of human blood. Biorheology. 18, 45373
(1981)
13. Tashtoush, B., Magableh, A.: Magnetic field effect on heat transfer and fluid flow characteristics
of blood flow in multi-stenosis arteries. Heat Mass Transf. 44, 297–304 (2008)
14. Taylor, C.A., Draney, M.T.: Experimental and computational methods in cardiovascular fluid
mechanics. Annu. Rev. Fluid Mech. 36, 197–231 (2004)
15. Thurston, G.B.: Viscoelasticity of human blood. Biophys. J. 12, 120517 (1972)
16. Tzirtzilakis, E.E.: A mathematical model for blood flow in magnetic field. Phys. Fluids. 17,
077103 (2005)
Improving R-Order Convergence
of Derivative Free with Memory Method
by Two Self-accelerator Parameters
Anuradha Singh and J.P. Jaiswal
Abstract The object of the present paper is to improve the R-order convergence
of with memory method proposed by Eftekhari (Int J Differ Eqn 2014:6, 2014) [1].
To achieve this goal, one more iterative parameter is introduced, which is calculated
with the help of Newton’s interpolatory polynomial of degree five. It is shown that the
R-order convergence of the proposed method is increased from 11.2915 to 13.4031
without any extra evaluation. Smooth as well as nonsmooth examples are presented
to confirm theoretical result and superiority of the new scheme.
Keywords R-order convergence · Self accererating parameter · Computational

efficiency
1 Introduction
Finding zeros of a scalar function f has importance among the most significant prob-
lems in not only the theory and practical of applied mathematics, but also of many
branches of engineering sciences, physics, computer science, finance, to mention
only some fields. These problems lead to a rich blend of mathematics, numerical
analysis, and computational science. To solve these types of nonlinear equations,
iterative methods such as Newton’s method and its modification are usually used.
During the last few years, multipoint methods have drawn the attention of many
researchers. Multipoint iterative methods are defined as methods that require evalua-
tion of functions and its derivatives at a number of values of the independent variable.
The main goal and motivation in the construction of new methods is to achieve the
A. Singh · J.P. Jaiswal (B)

Department of Mathematics, Maulana Azad National Institute of Technology,
Bhopal 462051, Madhya Pradesh, India
e-mail: asstprofjpmanit@gmail.com
A. Singh
e-mail: singhanuradha87@gmail.com

DOI 10.1007/978-81-322-2485-3_41
502 A. Singh and J.P. Jaiswal
highest computational efficiency; in other words, it is desirable to attain as high as

possible convergence order with a fixed number of function evaluations per iteration.
In our paper, the order of convergence and efficiency index have been improved
in the neighborhood of a simple root. The efficiency index is defined [2] as I =
1
p d , where p is the order of convergence and d is the total number of function
evaluations per iteration. By Kung and Traub conjecture [3], a multipoint method
without memory performing n + 1 function evaluations per iteration can have at
most convergence order 2n . The iterative methods that agree with this conjecture are
known as so-called optimal methods. In [4], Petkovic et al. have presented a large
collection of without memory multipoint methods for solving nonlinear equations.
In the recent past, researchers have focused to improve the existing methods without
additional evaluation of function and derivative in such a way it agrees with Kung
and Traub conjecture. As a result, the methods give higher computational efficiency.
The main motive in constructing iterative algorithms for solving nonlinear equa-
tions is to achieve as high as possible convergence rate with a fixed number of function
evaluations per iteration. To find this aim, the idea of with memory method, which
use the information from current and previous iterations, is first initiated by Traub in
1964. The order of convergence of new multipoint methods with memory is greater
than the order of convergence of the corresponding optimal multipoint methods with-
out memory. Accelerated convergence is obtained by variation of self-accelerating
parameters, which are recursively calculated as the iteration proceed using infor-
mation from the current and previous iterations. The improved convergence rate,
attained without extra evaluations, is a significant advantage of multipoint methods
with memory. In this paper, we have presented modified and efficient version of
existing with memory method.
Rest of the paper is organized as follows: Sect. 2 is devoted to the development
and theoretical proof of the improved three-points with memory method. It is shown
that the proposed scheme has greater convergence order without extra evaluation. As
a result, it shows high computational efficiency. Finally, one smooth and nonsmooth
examples are presented to justify the significance of the present work.
2 New Derivative-Free Iterative Method with Memory
In the convergence analysis of the new method, we employ the notation used in
Traub’s book [5]: if m k and n k are null sequences and m k /n k → C, where C is
a nonzero constant, we shall write m k = O(n k ) or m k ∼ Cn k . We also use the
concept of R-order of convergence introduced by Ortega and Rheinboldt [6]. Let
xk be a sequence of approximations generated by an iterative method (IM). If this
sequence converges to a zero ξ of function f with the R-order O R ((I M), ξ) ≥ r ,
we will write
ek+1 ∼ Dk,r ekr ,
Improving R-Order Convergence of Derivative Free with Memory Method … 503
where Dk,r tends to the asymptotic error constant Dr of the iterative method (IM)
when k → ∞.
In [7] Thukral constructed the following derivative-free without memory method
wk = xk + β f (xk ), k = 0, 1, 2, ...,

f (xk )
yk = xk − ,
f [xk , wk ]

f [xk , wk ] f (yk )
z k = yk − ,
f [wk , yk ] f [xk , yk ]

f (z k ) −1 f (yk )3 f [xk , yk ] f (z k )
xk+1 = z k − 1 − 1− ,
f (wk ) f (wk )2 f (xk ) f [yk , z k ] f [xk , z k ]
(1)
where β ∈ R + and f [., .] denotes the usual divided difference. The author showed
that it has optimal eighth-order of convergence. Very recently, Eftekhari [1] first
replaced parameter β in the above method by iterative parameter βk . This itera-
tive parameter is also present in the coefficient of first term of the error expression.
To achieve higher order convergence without extra evaluation, the author approx-
imated the iterative parameter by Newton interpolatory polynomial of third and
fourth degree, respectively. In fact the maximum R-order convergence of his pro-
posed method is 11.2915. In this work, we discuss the modified version of the same
with memory, which has more higher R-order of convergence as well as high com-
putational efficiency. For this purpose, we introduce one more parameter and thus
our proposed method is given by
wk = xk + β f (xk ), k = 0, 1, 2, ...,

f (xk )
yk = xk − ,
f [xk , wk ] + α f (wk )

f [xk , wk ] f (yk )
z k = yk − ,
f [wk , yk ] f [xk , yk ] + α f (yk )

f (z k ) −1 f (yk )3 f [xk , yk ] f (z k )
xk+1 = z k − 1 − 1− ,
f (wk ) f (wk )2 f (xk ) f [yk , z k ] f [xk , z k ] + α f (z k )
(2)
where α, β ∈ R + . The error expression of the above method is
ek+1 = M8,1 (1 + βc1 )3 (αc1 + c2 )2 ek8 + O(ek9 ). (3)
(i)
where M8,1 is asymptotic constant, ci = f i!(ξ) and ξ is the exact root. Since the
above error equation contains both the parameters, which can be approximated in
such a way that they increase the local convergence order. For this purpose, first we
replace the parameters α and β by iterative parameters αk and βk , respectively and
then approximation of these parameters are given by the following way
1 1 1
βk = − ≈− =− ,
c1 c˜1 Ñ4 (xk )

c2 c2 Ñ (wk )
αk = − ≈ − = − 5 , (4)
c1 c1 2 Ñ5 (wk )
where Ñ4 (t) = Ñ4 (t; xk , z k−1 , yk−1 , xk−1 , wk−1 ),

Ñ5 (t) = Ñ5 (t; xk , wk , z k−1 , yk−1 , xk−1 , wk−1 ) are Newton’s interpolatory polyno-
mial of degree four and five, respectively. Before going to prove the main result,
we state the following lemma which can be obtained using the error of Newton’s
interpolation, in the same manner as in [8]:
Ñ5 (wk )
Lemma 1 If βk = − 1
and αk = − , then the estimates
Ñ4 (xk ) 2 Ñ5 (wk )
c5
(i) 1 + βk c1 ∼ − ek−1,z ek−1,y ek−1,w ek−1 ,
c1
(ii) αk c1 + c2 ∼ c6 ek−1,z ek−1,y ek−1,w ek−1 .
The theoretical proof of the order of convergence of the proposed method is given
by the following theorem:
Theorem 1 If an initial approximation, x0 is sufficiently close to a simple zero ξ of
f (x) = 0 and the parameters βk and αk in the iterative scheme (2) is recursively
calculated by the forms given in (4).√Then the R-order of convergence of with memory
scheme (2) with (4) is at least 7 + 41 = 13.4031.
Proof First we assume that the R-orders of convergence of sequences xk , wk , yk , z k
are at least r , l, m and n, respectively. Hence
2
ek+1 ∼ Dk,r ekr ∼ Dk,r (Dk−1,r ek−1
r
)r ∼ Dk,r Dk−1,r
r r
ek−1 . (5)
and
ek,w ∼ Dk,l ekl ∼ Dk,l (Dk−1,r ek−1

r
)l ∼ Dk,l Dk−1,r
l rl
ek−1 . (6)
Similarly
ek,y ∼ Dk,m Dk−1,r

m rm
ek−1 , (7)
ek,z ∼ Dk,n Dk−1,r

n rn
ek−1 . (8)
By virtue of the above equations and Lemma 1, we have

c5 n+m+l+1
1 + βk c1 ∼ − (Dk−1,n )(Dk−1,m )(Dk−1,l )ek−1 , (9)
c1
and
n+m+l+1
αk c1 + c2 ∼ c6 (Dk−1,n )(Dk−1,m )(Dk−1,l )ek−1 . (10)
For the scheme (2), it can be derived that
ek,w ∼ (1 + βk c1 )ek , (11)
ek,y ∼ M2,1 (1 + βk c1 )(αk c1 + c2 )ek2 , (12)
ek,z ∼ M4,1 (1 + βk c1 )2 (αk c1 + c2 )ek4 , (13)
and
ek+1 ∼ M8,1 (1 + βk c1 )3 (αk c1 + c2 )2 ek8 , (14)
where M2,1 , M4,1 and M8,1 are asymptotic constants. Using (9) in the Eq. (11) and
then simplifying we obtain

c5 n+m+l+r +1
ek,w ∼ − (Dk−1,n )(Dk−1,m )(Dk−1,l )(Dk−1,r )ek−1 .
c1
(15)
Similarly by virtue of (9) and (10), the Eqs. (12), (13) and (14), respectively, become

−c5 c6 2(n+l+m+1)+2r
ek,y ∼ M2,1 (Dk−1,n
2
)(Dk−1,m
2
)(Dk−1,l
2
)(Dk−1,r
2
)ek−1 ,
c1
(16)

c52 c6 3(n+l+m+1)+4r
ek,z ∼ M4,1 (Dk−1,n
3
)(Dk−1,m
3
)(Dk−1,l
3
)(Dk−1,r
4
)ek−1 ,
c12
(17)
and

−c53 c62 5(n+l+m+1)+8r
ek+1 ∼ M8,1 .(Dk−1,n
5
)(Dk−1,m
5
)(Dk−1,l
5
)(Dk−1,r
8
)ek−1 .
c13
(18)
Now comparing the equal powers of ek−1 in Eqs. (6)–(15), (7)–(16), (8)–(17) and
(5)–(18), we find the following system of nonlinear equations:
rl − r − (n + l + m + 1) = 0,
r m − 2r − 2(n + l + m + 1) = 0,
r n − 4r − 3(n + l + m + 1) = 0,
r 2 − 8r − 5(n + l + m + 1) = 0.

√
√
Solving these equations, we get l = 15 4 + 41 , m = 25 4 + 41 , n =

√ √
1
5 17 + 3 41 , r = 7 + 41. And thus we proved the result.
Note 1: The efficiency index of the proposed method (2) with (4) is (13.4031)1/4 =
1.9134 which is more than (11.2915)1/4 = 1.8831 of method proposed by
Eftekhari [1].
3 Numerical Results and Conclusions
In this section, the new method is applied to solve some nonlinear equations (smooth
as well as nonsmooth) and compared with several with memory derivative-free meth-
ods. The absolute errors in the first three iterations are given in Tables 1 and 2, where
the exact roots are computed with 1,000 significant digits. The computational order
of convergence (COC) is defined by
ln(| f (xk )/ f (xk−1 )|)

COC = .
ln(| f (xk−1 )/ f (xk−2 )|)
To test the performance of new method consider, the following two nonlinear func-
tions (which are taken from [1, 9]):
1. f 1 (x) = 10(x 4 + x), x < 0

= −10(x 3 + x), x ≥ 0.
2 +x
2. f 2 (x) = e x cos(x)−1 + sin(πx) + x log(x sin(x) + 1).
The effectiveness of the new scheme with memory method (2) with (4) (NWM) is
confirmed by comparing this with some recently established with memory methods.
Specifically, we consider the 12th-order method (31) (LWMa), (32) (LWMb), (33)
(LWMc), and (34) (LWMd) introduced by Lotfi and Tavakoli in [10], 12th-order
method (12)–(8) for φ1 (EWMa), (12)–(8) for φ2 (EWMb), and (12)–(8) for φ3
(EWMc), introduced by Eftekhari in [1] and 14th-order method (22) (LSWMa) and
(23) (LSWMb) introduced by Lotfi et al. in [11]. In the present scenario, high-order
Table 1 Numerical results for f 1 (x)

Method |x1 − ξ| |x2 − ξ| |x3 − ξ| COC
x0 = −0.8 γ0 = 0.01 α0 = 0.01 ξ = −1
LWMa 0.18233e+1 0.99985e+0 1.00000e+0 2.2130
LWMb 0.86418e−1 0.52042e−9 0.45465e−109 11.824
LWMc 0.12019e−1 0.36933e−6 0.74211e−73 12.343
LWMd 0.36954e+0 0.36954e+0 0.36954e+0 1.0042
EWMa 0.57265e−1 0.10604e−9 0.32534e−106 11.117
EWMb 0.14034e+2 0.47011e+1 0.11965e+1 1.0065
EWMc 0.57265e−1 0.10604e−9 0.32534e−106 11.117
LSWMa 0.41716e−3 0.45375e−45 0.1.4833e−632 14.000
LSWMb 0.32191e−2 0.49928e−31 0.27724e−435 13.996
NWM 0.55140e−1 0.31086e−14 0.40744e−200 14.075
Table 2 Numerical results for f 2 (x)

Method |x1 − ξ| |x2 − ξ| |x3 − ξ| COC
x0 = −0.8 γ0 = 0.01 α0 = 0.01 ξ = −1
LWMa 0.12710e−2 0.88275e−32 0.20559e−384 12.094
LWMb 0.32541e−2 0.12598e−29 056465e−356 11.906
LWMc 0.26347e−2 0.45753e−30 0.28758e−361 11.931
LWMd 0.19011e−2 0.1.3528e30 0.33386e−370 12.066
EWMa 0.25117e−2 0.18296e−26 0.93718e−294 11.073
EWMb 0.25118e−2 0.22661e−24 0.81372e−246 10.045
EWMc 0.25117e−2 0.18296e−26 0.93718e−294 11.073
LSWMa 0.28748e−2 0.46101e−36 0.58786e−519 14.290
LSWMb 0.15835e−2 0.86423e−39 0.72387e−532 13.598
NWM 0.24395e−2 0.77037e−33 0.32310e−464 14.143
methods are important because numerical applications use high precision in their
computations; for this reason, numerical tests have been carried out using variable
precision arithmetic in MATHEMATICA 8 with 1,000 significant digits.
The numerical results showed in Tables 1 and 2 are in concordance with the theory
developed in this paper. From the results displayed in Tables 1 and 2, we can conclude
that the order of convergence of the derivative-free method without memory can be
made more higher by the method with memory by imposing one more parameter
without any additional calculations and the computational efficiency of the with
memory method is very high. The R-order of convergence is increased from 11.2915
to 13.4031 in accordance with the quality of the applied accelerating method given
by (2) with (4). We can see that the self-accelerating parameters play a key role in
increasing the order of convergence of the iterative method.
Acknowledgments The authors are grateful to editor and reviewers for their significant suggestions
which improved of the quality of the paper.
References
1. Eftekhari, T.: On some iterative methods with memory and high efficiency index for solving
nonlinear equations. Int. J. Differ. Eqn. 2014(Article ID 495357), 6 (2014)
2. Ostrowski, A.M.: Solution of Equations and System of Equations. Prentice-Hall, Englewood
Cliffs (1964)
3. Kung, H.T., Traub, J.F.: Optimal order of one point and multipoint iteration. J. Assoc. Comput.
Math. 21, 643–651 (1974)
4. Petkovic, M.S., Neta, B., Petkovic, L.D., Dzunic, J.: Multipoint Methods for Solving Nonlinear
Equation. Academic Press, Elsevier (2012)
5. Traub, J.F.: Iterative Methods for Solution of Equations. Prentice Hall, New York (1964)
6. Ortega, J.M., Rheinboldt, W.C.: Iterative Solution of Nonlinear Equations in Several Variables.
Academic Press, New York (1970)
7. Thukral, R.: A family of three-point derivative-free methods of eighth-order for solving non-
linear equations, J. Mod. Meth. Numer. Math. 3(2), 11–21 (2012)
8. Dzunic, J.: On efficient two-parameter methods for solving nonlinear equations. Numer. Algo-
rithm 63, 549–569 (2013)
9. Hafiz, M.A., Bahgat, M.S.M.: Solving non-smooth equations using family of derivative-free
optimal methods. J. Egypt. Math. Soc. 21, 38–43 (2013)
10. Lotfi, T., Tavakoli, E.: On a new efficient Steffenssen-like iterative class by applying a suitable
self-accelerator parameter. Sci. World J. 2014(Article ID 769758), 9 (2014)
11. Lotfi, T., Soleymani, F., Shateyi, S., Assari, P., Haghani, F.K.: New mono- and bi-accelerator
iterative methods with memory for nonlinear equations. Abstr. Appl. Anal. 2014(Article ID
705674), 8 (2014)
Numerical Solutions of Differential
Equations Using Modified B-spline
Differential Quadrature Method
R.C. Mittal and Sumita Dahiya
Abstract In this article, a modified cubic B-spline differential quadrature method

(MCB-DQM) is proposed to solve some of the basic differential equations. Here we
have considered an ordinary differential equation of order two along with heat equa-
tion and one- and two-dimensional wave equations. A nonlinear ordinary differential
equation of order two is also considered. The ordinary differential equation is reduced
to a system of nonhomogeneous linear equations which is then solved by using the
Gauss elimination method, whereas the heat equation and the one-dimensional and
two-dimensional heat and wave equations are reduced to a system of ordinary dif-
ferential equations. The system is then solved by the optimal four-stage three-order
strong stability preserving time stepping Runge–Kutta (SSP-RK43) scheme. The
reliability and efficiency of the method have been tested on six examples.
Keywords Ordinary differential equation · Heat equation · Wave equation cubic

B-spline functions · Modified cubic B-spline quadrature method · System of ordi-
nary differential equations · Gauss elimination method · Runge–Kutta fourth-order
method
1 Introduction
To describe change, the most accurate way is to use differentials and derivatives, that
is why differential equation arises in many different contexts. In this article, we have
taken a second-order ordinary differential equation with boundary conditions of the
form
d 2u du
2
+A + Bu = f (x), x ∈ [a, b] (1)
dx dx
R.C. Mittal · S. Dahiya (B)

Department of Mathematics, IIT Roorkee, Roorkee 247667, Uttarakhand, India
e-mail: sumita.iitr@gmail.com
R.C. Mittal
e-mail: rcmmmfma@iitr.ernet.in

DOI 10.1007/978-81-322-2485-3_42
510 R.C. Mittal and S. Dahiya
with the boundary conditions given as u(a) = g1 (x) and u(b) = g2 (x) where
g1 (x), g2 (x), A, and B are constants or functions of x.
Second, we have taken the wave equation. The wave is an important second-order
partial differential equation which describes the nature of waves such as light
waves, water waves, and sound waves. It arises in the fluid dynamics and elec-
tromagnetic fields. The problem of a vibrating string was earlier studied by Jean
le Rond [1], d’Alembert [2], Leonhard Euler, Daniel Bernoulli, and Joseph-Louis
Lagrange. The one-dimensional wave equation was discovered by d’Alembert in
1746, and then within 10 years Euler discovered the three-dimensional wave equa-
tion. Here we will restrict our results to one- and two-dimensional wave equations.
The wave equation is a hyperbolic partial differential equation. It is time-dependent
and concerns a time variable t, spatial variables x1 , x2 , . . . , xn , and a scalar function
u = u(x1 , x2 , . . . , xn ; t) which can model the displacement of a wave. The wave
equation is then given by
∂ 2u
= c2 ∇ 2 u (2)
∂t 2
where ∇ 2 is the spatial Laplacian and where c is a fixed constant. The equation
alone does not specify a solution; a unique solution is obtained by setting a problem
with initial conditions or boundary conditions. The one-dimensional wave equation
is given by
∂ 2u 2∂ u
2
= c (3)
∂t 2 ∂x2
and the two-dimensional wave equation is given by
2
∂ 2u 2 ∂ u ∂ 2u
=c + 2 (4)
∂t 2 ∂x2 ∂y
Third, we have considered the heat equation. The heat equation is a parabolic equation
that describes the variation in temperature or distribution of heat in a given region
over time. Generally in a coordinate system, heat equation is given by
∂u
− α∇ 2 u = 0 (5)
∂t
where α is a positive constant, and ∇ is the Laplacian operator. The heat equation
is of fundamental importance in diverse scientific fields. In mathematics, it is the
prototypical parabolic partial differential equation. In probability theory, the heat
equation is connected with the study of Brownian motion via the Fokker–Planck
equation. In financial mathematics, it is used to solve the Black–Scholes partial
differential equation. The diffusion equation, a more general version of the heat
equation, arises in connection with the study of chemical diffusion and other related
processes. The heat equation is derived from the Fourier’s Law and conservation of
energy.
Numerical Solutions of Differential Equations … 511
2 Description of Modified Cubic B-Spline Differential

Quadrature Method (MCB-DQM)
Differential quadrature method [3] is a numerical technique to solve ordinary and

partial differential equations. With this method, the spatial derivatives of unknown
functions are approximated at any grid point using weighted sum of all the functional
values at certain point in the whole computational domain. In two-dimensional DQM,
first, we discretize the domain D = {(x, y) : a ≤ x ≤ b; c ≤ y ≤ d} as D 1 =
{(xi , y j ), i = 1, 2, . . . .N ; j = 1, 2, . . . , M} by taking step length x = xi − xi+1
in x-axis direction and y = y j − y j−1 in y-axis direction. According to DQM, the
approximation of the first-order partial derivative with respect to x of the dependent
function u(x, y, t), keeping y j fixed, at point xi is given as follows:

N
(1)
u x (xi , y j , t) = aik u(xk , y j , t), i = 1, 2, ...., N (6)
k=1
Similarly, the approximation of the first-order partial derivative of the dependent

function u(x, y, t) with respect to y, keeping the point xi fixed, at point y j is given
as follows:

N
(1)
u y (xi , y j , t) = ā jk u(xi , yk , t), j = 1, 2, . . . , M (7)
k=1
(1) (1)
where ai j and ā jk are unknown, representing the weighting coefficients of the
first-order partial derivatives with respect to x and y. There are many approaches to
calculate these weighting coefficients such as Shu’s approach [4], Quan and Chang’s
approach [5, 6], and Bellman’s approach [3]. In recent years, most of the differential
quadrature method using various test functions such as Lagrange interpolation poly-
nomial, Legendre polynomials, Lagrange interpolation cosine functions, spline func-
tions, etc. are based on Shu’s approach. Nowadays, most frequently used quadrature
methods are based on sine–cosine expansion and Lagrange interpolation. Korkmaz
and Dag [7, 8] have used cosine expansion-based differential quadrature method and
sine differential quadrature method for many nonlinear partial differential equations.
Mittal et al. [9–12] proposed polynomial-based differential quadrature method for
numerical solutions of nonlinear partial differential equations. Here, in this article
an approach based on modified cubic B-spline functions has been proposed to find
the weighting coefficients of differential quadrature method. Computed results show
that reported results are accurate.
2.1 Modified Cubic B-Spline Functions
In this method, a modification of cubic B-spline functions is used to find the weighting
coefficients ai(1) (1)
j and ā jk . The cubic B-spline functions are defined as follows:
⎧
⎪
⎪ (x − xm−2 )3 x ∈ [xm−2 , xm−1 )
⎪
⎪
⎨ (x − xm−2 )3 − 4(x − xm−1 )3 x ∈ [xm−1 , xm )
1
ϕm (x) = 3 (xm+2 − x)3 − 4(xm+1 − x)3 x ∈ [xm , xm+1 ) m = 0, 1, . . . , N + 1
h ⎪⎪
⎪
⎪ (x − x)3 x ∈ [xm+1 , xm+2 )
⎩ m+2
0 otherwise
(8)
where {ϕ0 (x), ϕ1 (x), . . . , ϕ N (x)} forms a basis over the domain interval [a, b]. The
values of cubic B-splines and their derivatives at the nodal points are given in Table 1.
The modification in cubic B-spline functions is done in such a way, so that the result-
ing matrix becomes diagonally dominant. Modified cubic B-spline basis functions
at the knots are defined as follows [13]:
φ1 (x) = ϕ1 (x) + 2ϕ0 (x),

φ2 (x) = ϕ2 (x) − ϕ0 (x),
φl (x) = ϕl (x), l = 3, 4, . . . , N − 2,
φ N −1 = ϕ N −1 (x) − ϕ N +1 (x)
φ N (x) = ϕ N (x) + 2ϕ N +1 (x) (9)
The function φl (x), l = 1, 2, . . . , N again form a basis over the interval [a, b].
2.2 To Compute Weighting Coefficients
(1)
Keeping y-axis fixed in Eq. (6), we find the weighting coefficients aik . Putting the
functions φm (x), m = 1, 2, . . . N in Eq. (6), we get

N
(1)
φl (xi , y j ) = aik φl (xk , y j ), j = 1, 2, . . . , M (10)
k=1
Table 1 Values of ϕm (x) and its derivatives at the nodal points

xm−2 xm−1 xm xm+1 xm+2
ϕm (x) 0 1 4 1 0

ϕm (x) 0 3/ h 0 −3/ h 0

ϕm (x) 0 6/h 2 −12/h 2 6/h 2 0
For any arbitrary choice of l, we get the following algebraic system of equations
(1) ⎤⎡ ⎡ ⎤
⎡ ai1⎤ φ1,i
φ1,1 φ1,2 ⎢ a (1) ⎥ ⎢ φ ⎥
⎢φ2,1 φ2,2 φ2,3 ⎥⎢ i2 ⎥ ⎢ 2,i ⎥
⎢ ⎥⎢ .. ⎥ ⎢ . ⎥
⎢ φ φ φ ⎥⎢ . ⎥ ⎢
⎢ ⎥
⎢ .. ⎥
⎥
⎢ 3,1 3,2 3,3 ⎥⎢ =
⎢ ... ... ... ⎥ ⎢ .. ⎥ ⎢ . ⎥
⎥ ⎢
⎢ ⎥ ⎢ . ⎥ ⎢ .. ⎥ ⎥
⎣ ... φ N −1,N −2 φ N −1,N −1 φ N −1,N ⎦ ⎢ (1) ⎥ ⎢ ⎥
... φ N ,N −1 φ N ,N ⎣ai N −1 ⎦ ⎣φ N −1,i ⎦
(1)
ai N φ N ,i
(11)
Here the Eq. (11) gives the systems of tridiagonal algebraic system of equations
for each i. These tridiagonal system of equations can be easily solved by “Thomas
(1)
Algorithm” providing the weighting coefficients of first-order derivatives aik . For
i = 1, we have the following tridiagonal system of equations:
⎡ (1) ⎤ ⎡ ⎤
⎡ ⎤ a11 −6
6 1 ⎢ a (1) ⎥ ⎢ h6 ⎥
⎢0 4 1 ⎥⎢ 12 ⎥ ⎢h ⎥
⎢ ⎥⎢ . ⎥ ⎢ .. ⎥
⎢ 1 4 1 ⎢
⎥ ⎢ .. ⎥ ⎥ ⎢ . ⎥
⎢ ⎥⎢ =⎢ ⎥
⎢ ... ... ... ⎥ ⎢ .. ⎥ ⎢ ⎥ (12)
⎢ ⎥ ⎢ . ⎥ ⎢ ... ⎥
⎥
⎣ ... 1 4 0 ⎦ ⎢ (1) ⎥ ⎢ ⎣0⎦
⎥
. . . 1 6 ⎣a1N −1 ⎦
(1) 0
a1N
The solution of Eq. (12) by “Thomas algorithm” gives the weighting coefficients
(1) (1) (1)
a11 , a12 , . . . , a1N . In the same way, we can also find out the weighting coefficients
for i = 2, 3, . . . , N . Using these coefficients we are able to find out the first-order
partial derivatives. The higher order derivatives can be calculated by the following
recurrence relation:
(r −1)
(r ) (1) (r −1)
ai j
ai j = r ai j aii − , for i = j
xi − x j
i, j = 1, 2, . . . , N ; r = 2, 3, . . . , N − 1 (13)
(r )

N
(r )
aii = − ai j , for i = j (14)
j=1, j=i
Here ai(rj −1) and ai(rj ) are the weighting coefficients of (r − 1)th- and (r )th-order
partial derivatives w.r.t. x.
(1)
We can find out the weighting coefficients ā jk of first-order partial derivatives with
respect to y (keeping x-axis fixed) in the same manner by putting modified cubic B-
spline functions in Eq. (7). The second and higher order derivatives can be calculated
in the same way by the recurrence formulae
(r −1)
(r ) (1) (r −1)
āi j
āi j = r āi j āii − , for i = j
yi − y j
i, j = 1, 2, ...., N ; r = 2, 3, ...., N − 1 (15)
(r )

N
(r )
āii = − āi j , for i = j (16)
j=1, j=i
(r −1) (r )
Here āi j and āi j are the weighting coefficients of (r −1)th- and (r )th-order partial
derivatives w.r.t. y. In one-dimensional DQM, one can assume uniformly distributed
N knots: a = x1 ≤ x2 , . . . , x N −1 ≤ x N = b such that h = xi − xi−1 on the real axis.
By using modified cubic B-spline functions, the first-order derivative approximation
is given as follows:

N
(1)
φl (xi ) = ai j φl (x j ), for i = 1, 2, . . . , N ; k = 1, 2, . . . , N (17)
j=1
which provide the weighted coefficients ai j for i = 1, 2, . . . , N ; j = 1, 2, . . . , N .

The second derivatives can be evaluated in the same way by the recurrence formulas

N
(2) (1) (1) 1 (2) (1)
ai j = 2ai j aii − , for i = j, and aii = − ai j (18)
xi − x j
i=1,i= j
3 Numerical Scheme Based on Modified Cubic B-Spline

Differential Quadrature Method
Discretizing the spatial derivatives by applying the modified cubic B-spline differen-
tial quadrature method, Eq. (1), is reduced to the following system of linear equation

N
(2)

N
(1)
aik u(xk , t) + A aik u(xk , t) + Bu(xk , t) = f (xk ), (19)
k=1 k=1
This is a system of N equations in N − 2 unknowns. After using the boundary

conditions and doing certain manipulations we will solve a system of N −2 equations
in N − 2 unknowns. Gauss Elimination method will be used for this purpose.
For the one-dimensional wave equation, introducing an auxiliary function w, Eq. (3)
reduces to a system of partial difference equations given as follows:
∂w ∂ 2u
= 2,
∂t ∂x
∂u
w= , x ∈ [a, b], t ∈ (0, T ] (20)
∂t
The initial value and boundary value conditions on w are given as
w(x, 0) = w0 (x), x ∈ [a, b], (21)
w(x, t) = g(x, t), x ∈ [a, b], t ∈ (0, T ] (22)
Discretizing the spatial derivatives by applying the modified cubic B-spline differ-
ential method, Eq. (3) reduces to
dwi, j (2)
N
= aik u(xk , t),
dt
k=1
du i j
= w(xk , t) (23)
dt
with initial conditions
u(xi , 0) = u 0 (xi ), w(xi , 0) = w0 (xi ) (24)
and boundary conditions
u(xi , t) = f (xi ), w(xi , t) = g(xi ) (25)
In the same way, an auxiliary function is introduced in the two-dimensional wave

equation, and the Eq. (4) is reduced to
∂w ∂ 2u ∂ 2u
= 2 + 2,
∂t ∂x ∂y
∂u
w= , (x, y) ∈ [a, b], t ∈ (0, T ] (26)
∂t
The initial value and boundary value conditions on w are given as
w(x, y, 0) = w0 (x, y), (x, y) ∈ [a, b], (27)
w(x, y, t) = g(x, y, t), (x, y) ∈ [a, b], t ∈ (0, T ] (28)
Discretizing the spatial derivatives by applying the modified cubic B-spline differ-
ential method, Eq. (3) reduces to
dwi, j (2)
N (2) N
= aik u(xk , y j , t) + ā jk u(xi , yk , t),
dt
k=1 k=1
du i j
= w(xi , y j , t) (29)
dt
u(xi , y j , 0) = u 0 (xi , y j ), w(xi , y j , 0) = w0 (xi , y j ) (30)
u(xi , y j , t) = f (xi , y j ), w(xi , y j , t) = g(xi , y j ) (31)
For the heat equation given by Eq. (5), the same procedure is applied.
There are various methods to solve this system of ordinary differential equations.
We used the optimal four-stage, order three strong stability preserving time stepping
Runge–Kutta [SSP-RK43] scheme [14] to solve the system of ordinary differential
equations. In this scheme the system in Eqs. (20) and (23) along with the initial and
boundary conditions is integrated from time t0 to t0 + t through the following
operations:
t
u (1) = u m + L(u m )
2
t
u (2) = u (1) + L(u (1) )
2
2 u(2) t
u (3) = um + + L(u (2) )
3 3 6
t
u (m+1) = u (3) + L(u (3) ) (32)
2
and consequently the solution u at a particular time level is completely known.
Stability of the proposed system is guaranteed unconditionally as the eigenvalues of
corresponding matrix will be real and negative.
Table 2 Comparison between exact values and calculated values with stepsize h = π/40
x MC B-D Q M Exact
0.07 8.89041 8.89048
0.15 8.56967 8.56974
0.24 8.05311 8.05317
0.31 7.36956 7.36062
0.39 6.51566 6.51571
0.47 5.54507 5.54511
0.55 4.47771 4.47773
0.63 3.34387 3.34389
0.71 2.17444 2.17444
In this section, the numerical solutions by the proposed method (MCB-DQM) are
evaluated for some examples of hyperbolic problem.The computational work is done
with the help of DEV C++. Existence of analytical solutions help to measure the
accuracy of numerical methods. In the present study, the accuracy and efficiency of
this method are measured for various numerical examples. The performance of the
MCB-DQM method is measured by the maximum absolute error k which is defined
as k = |u exact − u mcb−dqm |.
Problem 1 Consider the following second-order one-dimensional ordinary differen-

tial equation
d 2u
+ 4u = 8 sin(2x), x ∈ (0, π/4) (33)
dx2
with the boundary conditions given as
u(0) = 9 and u(π/4) = 1 (34)
The exact solution is given by
u(x) = 9 cos(2x) + sin(2x) − 2x cos(2x)
We have compared the computed values by using differential quadrature method

with modified B-splines with the exact results. The result in terms of absolute error
is shown in Table 2.
Table 3 The absolute error at intermediate knots for time t = 0.1, 0.5 and 1
x t = 0.1 t = 0.5 t =1
0.1 6.14E-04 6.59E-04 4.86E-04
0.2 1.44E-04 4.37E-04 4.25E-04
0.3 4.28E-05 3.38E-04 4.22E-04
0.4 5.19E-05 2.53E-04 4.04E-04
0.5 2.90E-05 2.25E-04 3.99E-04
0.6 5.19E-05 2.52E-04 4.04E-04
0.7 4.28E-05 3.38E-04 4.22E-04
0.8 1.44E-04 4.37E-04 4.25E-04
0.9 6.14E-04 6.59E-04 4.86E-04
Problem 2 Consider the one-dimensional heat equation given by

∂u 1 ∂ 2u
= 2 , x ∈ (0, 1) (35)
∂t π ∂x2

u(x, 0) = sin(π x)

u(0, t) = 0, u(1, t) = 0
u(x, t) = sin(π x) exp(−t), x ∈ [0, 1]
The calculated values are compared with the exact values for time t = 0.1, t = 0.5,
and t = 1. The results are shown in Table 3.
Problem 3 Consider the two-dimensional heat equation given by
∂u ∂ 2u ∂ 2u
= 2 + 2 , (x, y) ∈ (0, 1) × (0, 1) (36)
∂t ∂x ∂y

u(x, y, 0) = sin(π x) sin(π y)
u(0, y, t) = 0, u(1, y, t) = 0
u(x, 0, t) = 0, u(x, 1, t) = 0 (37)

Table 4 Comparison of exact values and computed values at intermediate points for time t = 0.1
and t = 1
x, y t = 0.1 t =1
MCB-DQM Exact MCB-DQM Exact
0.1, 0.1 0.013378 0.013264 2.63842E-010 2.54963E-010
0.2, 0.2 0.048180 0.047992 9.50204E-010 9.22467E-010
0.3, 0.3 0.911731 0.090918 1.79811E-009 1.74755E-009
0.4, 0.4 0.125932 0.125646 2.48362E-009 2.41505E-009
0.5, 0.5 0.139208 0.138911 2.74544E-009 2.67001E-009
0.6, 0.6 0.125932 0.125646 2.48362E-009 2.41505E-009
0.7, 0.7 0.911731 0.090918 1.79811E-009 1.74755E-009
0.8, 0.8 0.048180 0.047992 9.50204E-010 9.22467E-010
0.9, 0.9 0.013378 0.013264 2.63842E-010 2.54963E-010
0.14
0.12
0.1
0.08
0.06
0.04
0.02 0
0 0.5
0 0.2 0.4 0.6 0.8 1 1
Fig. 1 Plots of numerical solution at time level t = 0.1 for Example 4
u(x, y, t) = sin(π x) sin(π y) exp(−2π 2 t), (x, y) ∈ [0, 1] × [0, 1]
The calculated values are compared with the exact values for time t = 0.1 and t =
1. The results are shown in Table 4. The surface plots of numerical solutions at time
level t = 0.1 is depicted in Fig. 1.
Problem 4 Consider the one-dimensional wave equation
Table 5 The absolute error at intermediate knots for time t = 0.01, 0.1 and 1
x t = 0.01 t = 0.1 t =1
0.1 2.0178E-013 4.7239E-009 6.8554E-007
0.2 3.4674E-013 3.1219E-011 1.3029E-006
0.3 4.7734E-013 4.2971E-011 1.7932E-006
0.4 5.6104E-013 5.0515E-011 2.1077E-006
0.5 5.8990E-013 5.3116E-011 2.2165E-006
0.6 5.6103E-013 5.0516E-011 2.1077E-006
0.7 4.7722E-013 4.2972E-011 1.7932E-006
0.8 3.4673E-013 2.9860E-011 1.3029E-006
0.9 2.0178E-013 4.7239E-009 6.8554E-007
∂ 2u ∂ 2u
= , x ∈ (0, 1) (38)
∂t 2 ∂x2
along with the boundary conditions given by u(0, t) = 0 and u(1, t) = 0

and the initial conditions given by u(x, 0) = 0 and ∂u∂t (x, 0) = π sin(π x).
The exact solution is given by u(x, t) = sin(π x) sin(π t). The computed results are
compared with the exact results for different values of time t, taking t = 0.001. The
absolute error at intermediate knots for time t = 0.01, 0.1, and 1 is shown in Table 5.
Problem 5 Consider the two-dimensional wave equation given by

∂ 2u 1 ∂ 2u ∂ 2u
= + , (x, y) ∈ (0, 1) × (0, 1) (39)
∂t 2 2 ∂x2 ∂ y2

u(x, y, 0) = 0
u(0, y, t) = 0, u(1, y, t) = 0
u(x, 0, t) = 0, u(x, 1, t) = 0 (40)
u(x, y, t) = sin(π x) sin(π y) sin(π t), (x, y) ∈ [0, 1] × [0, 1]
The calculated values are compared with the exact values for time t = 0.1 and t = 1.
The results are shown in Table 6. The surface plots of the numerical solutions at time
level t = 1 is depicted in Fig. 2.
Table 6 Comparison of exact values and computed values at intermediate points for time t = 0.1
and t = 1
x, y t = 0.1 t =1
Exact MCB-DQM Exact MCB-DQM
0.1, 0.1 0.029508 0.029532 −0.029508 −0.029546
0.2, 0.2 0.106763 0.106740 −0.106763 −0.106311
0.3, 0.3 0.202254 0.202242 −0.202254 −0.200769
0.4, 0.4 0.279508 0.279484 −0.279508 −0.276865
0.5, 0.5 0.309017 0.308992 −0.309017 −0.305909
0.6, 0.6 0.279508 0.279484 −0.279508 −0.276865
0.7, 0.7 0.202254 0.202242 −0.202254 −0.200769
0.8, 0.8 0.106763 0.106741 −0.106763 −0.106311
0.9, 0.9 0.029508 0.029532 −0.029508 −0.029546
−0.05
−0.1
−0.15
−0.2
−0.25
−0.3
−0.35
1
0.8 1
0.5 0.6
0.2 0.4
0 0
Fig. 2 Plots of numerical solution at time level t = 1 for Example 5
Problem 6 Consider the following nonlinear second-order one-dimensional partial

differential equation
∂u ∂ 2u ∂u
= 2 − u , x ∈ (0, 1) (41)
∂t ∂x ∂x
with the boundary conditions given as
u(0) = 0 and u(1) = 0 (42)
and initial condition given as

Table 7 Comparison between exact values and calculated values

x t MCB-DQM Exact
0.25 0.4 0.317526 0.31752
0.8 0.199558 0.19956
1.0 0.165601 0.16560
3.0 0.027761 0.02775
0.50 0.4 0.584541 0.58454
0.8 0.367406 0.36740
1.0 0.298352 0.29834
3.0 0.041069 0.04106
0.75 0.4 0.645641 0.64562
0.8 0.385369 0.38534
1.0 0.295885 0.29586
3.0 0.030443 0.03044
u(x, 0) = 4x(1 − x); (43)
We have compared the computed values by using differential quadrature method with
modified B-splines with the exact results. The description of the numerical solutions
of this example for different values of t is shown in Table 7.
5 Conclusion
In this paper, we have proposed a modified cubic B-spline differential quadrature

method (MCB-DQM) to solve second-order ordinary differential equation, wave
equation, and heat equation. The numerical examples show that this scheme can
produce highly accurate solutions. The main outcomes are as follows:
1. A technique based on modified cubic B-spline is proposed to find the weighting
coefficients rather than using the traditional method of Lagrange interpolation
[7].
2. Modifications in cubic B-spline are done in such a way that matrix size and
complexity get reduced when applied with differential quadrature method.
3. The method is easy to implement and economical in data complexity, which
results in less errors. We require less number of grid points, that means low
memory storage, which can be counted as an advantage of (MCB-DQM). This
method can be implemented easily to solve two-dimensional nonlinear partial
differential equations.
References
1. Cannon, J.T., Dostrovsky, S.: The evolution of dynamics, vibration theory from 1687 to 1742.
In: Studies in the History of Mathematics and Physical Sciences, vol. 6, pp. ix + 184. Springer,
New York (1981). ISBN:0-3879-0626-6
2. Gerard, F.W.: The vibrating string controversy. Am. J. Phys. 55(1), 33–37 (1987)
3. Bellman, R., Kashef, B.G., Casti, J.: Differential quadrature: a technique for the rapid solution
of nonlinear partial differential equations. J. Comput. Phys 10, 40–52 (1972)
4. Shu, C.: Differential Quadrature and its Application in Engineering. Athenaeum Press Ltd.,
Great Britain (2000)
5. Quan, J.R., Chang, C.T.: Newinsights in solving distributed system equations by the quadrature
methods- I. Comput. Chem. Eng. 13, 779–788 (1989)
6. Quan, J.R., Chang, C.T.: Newinsights in solving distributed system equations by the quadrature
methods- II. Comput. Chem. Eng. 13, 1017–1024 (1989)
7. Korkmaz, A., Dag, I.: Shock wave simulations using Sinc differential quadrature method. Eng.
Comput. 28(6), 654–674 (2011)
8. Korkmaz, A., Dag, I.: A differential quadrature algorithm for nonlinear Schrdinger equation.
Nonlinear Dyn. 56(12), 6983 (2009)
9. Mittal, R.C., Jiwari, R.: Numerical study of two-dimensional reactiondiffusion Brusselator
system. Appl. Math. Comput. 217(12), 5404–5415 (2011)
10. Mittal, R.C., Jiwari, R.: A differential quadrature method for solving Burgers-type equation.
Int. J. Numer. Methods Heat Fluid Flow 22(7), 880–895 (2012)
11. Jiwari, R., Pandit, S., Mittal, R.C.: Numerical simulation of two-dimensional sine-Gordon
solitons by differential quadrature method. Comput. Phys. Commun. 183(3), 600–616 (2012)
12. Jiwari, R., Pandit, S., Mittal, R.C.: A differential quadrature algorithm to solve the two dimen-
sional linear hyperbolic telegraph equation with Dirichlet and Neumann boundary conditions.
Appl. Math. Comput. 218(13), 7279–7294 (2012)
13. Mittal, R.C., Jain, R.K.: Numerical solutions of nonlinear Burgers equation with modified cubic
Bsplines collocation method. Appl. Math. Comput. 218, 7839–7855 (2012)
14. Spiteri, J.R., Ruuth, S.J.: A new class of optimal high-order strongstability-preserving time-
stepping schemes. SIAM J. Numer. Anal. 40(2), 469–491 (2002)
Complete Controllability of a Delayed
Semilinear Stochastic Control System
Urvashi Arora and N. Sukavanam
Abstract In this paper, complete controllability of a delayed semilinear stochastic

system is considered under some basic and readily verified conditions. A fixed-point
approach is employed for achieving the required result. At the end, an example is
given to show the effectiveness of the result.
Keywords Complete controllability · Delayed system · Stochastic control system ·

Banach fixed point theorem
1 Introduction
Controllability concepts play a vital role in deterministic control theory. It is well

known that controllability of deterministic equations is widely used in many fields
of science and technology. Kalman [1] introduced the concept of controllability for
finite-dimensional deterministic linear control systems. The basic concepts of con-
trol theory in finite- and infinite-dimensional spaces have been introduced in [2], and
[3]. However, in many cases, some kind of randomness can appear in the problem,
so that the system should be modeled by a stochastic form. Only few authors have
studied the extensions of deterministic controllability concepts to stochastic control
systems. Klamka et al. [4–7] studied the controllability of linear stochastic systems in
finite-dimensional spaces with delay and without delay in control as well as in state.
In [8–14], Mahmudov et al. established results for controllability of linear and semi-
linear stochastic systems in Hilbert Spaces. Shen and Sun [15] studied the control-
lability of stochastic first-order nonlinear systems with delay in control in finite-
dimensional as well as in infinite-dimensional spaces. Sakthivel et al. [16] studied
U. Arora (B) · N. Sukavanam

e-mail: urvashiaroraiitr@gmail.com
N. Sukavanam
e-mail: nsukavanam@gmail.com

DOI 10.1007/978-81-322-2485-3_43
526 U. Arora and N. Sukavanam
the controllability of nonlinear stochastic systems in finite-dimensional spaces using

Banach fixed-point theorem. Throughout this paper, we use the following standard
notations:
(a) (Ω, , P): Let be the σ algebra on the nonempty set Ω and P: → [0, 1] be
the probability measure. Then the triple (Ω, , P) is called a probability space.
(b) {t |t ∈ [0, T ]}: the filtration generated by an n-dimensional Wiener process
{ω(s): 0 ≤ s ≤ t} defined on the probability space.
(c) L 2 (Ω, T , Rn ): the Hilbert space of all T -measurable square-integrable vari-
ables with values in Rn .
(d) L 2 ([0, T ], Rn ): the Hilbert space of all square-integrable and t -measurable
processes with values in Rn .
(e) H2 : the Banach space of all square-integrable and t -adapted processes ϕ(t)
with norm
||ϕ||2 = sup E||ϕ(t)||2
t∈[0,T ]
where E denotes the Expected value.

(f) L(X, Y ): the space of all linear bounded operators from a Banach space X into
a Banach space Y .
(g) Uad = L 2 ([0, T ], Rm ): the set of admissible controls.
The problem of controllability of a linear stochastic system of the form
d x(t) = [Ax(t) + Bu(t)]dt + σ̃ dω(t) t ∈ [0, b]

x(0) = x0
has been studied by various authors [7, 8, 14], where σ̃ : [0, b] → Rn×n .
Also, the complete controllability of linear stochastic system with state delay
d x(t) = [A0 x(t) + A1 x(t − h) + B0 u(t)]dt + σ dω(t)
given the initial condition as a random function has been studied by Klamka [4].
In this paper, we examine the complete controllability of delayed semilinear sto-

chastic control system given as:

d x(t) = [Ax(t) + Bu(t) + f (t, x(t − h))]dt + σ (t, x(t − h))dω(t), t ∈ (0, T ]
(1)
x(t) = ψ(t)for t ∈ [−h, 0], x(0) = x0 = ψ(0)(say).
where the state x(t) ∈ Rn and the control u(t) ∈ Rm , A is an n × n constant matrix,
B is an n × m constant matrix. f : [0, T ] × Rn → Rn , σ : [0, T ] × Rn → Rn×n are
nonlinear functions, ω is an n-dimensional Wiener process and h > 0 is a constant
point delay.
Complete Controllability of a Delayed Semilinear Stochastic Control System 527
2 Preliminaries
It is well known that for given initial condition, any admissible control u ∈ Uad ,
for t ∈ [−h, T ] and suitable nonlinear functions f (t, x(t)) and σ (t, x(t)), the mild
solution of the semilinear stochastic differential state equation (1) can be represented
as
⎧ t
⎪
⎪
⎪
⎪ exp(At)x 0 + exp(A(t − s))[Bu(s) + f (s, x(s − h))]ds
⎨ 0
t
x(t; x0 , u) = (2)
⎪
⎪ + exp(A(t − s))σ (s, x(s − h))dω(s) f or t > 0
⎪
⎪
⎩ 0
ψ(t) f or t ∈ [−h, 0]
Let us introduce the following operators and sets. (see [16])

L T ∈ L(Uad , L 2 (Ω, T , Rn )) defined by
T
LT u = exp(A(T − s))Bu(s)ds
0
Then its adjoint operator L ∗T : L 2 (Ω, T , Rn ) → Uad is given by
L ∗T z = B ∗ exp(A∗ (T − t))E{z|t }
The set of all states reachable in time T from initial state x(0) = x0 ∈ L 2 (Ω, 0 , Rn ),
using admissible controls is defined as
RT (Uad ) = {x(T ; x0 , u) ∈ L 2 (Ω, T , Rn ) : u ∈ Uad }

T
wher e x(T ; x0 , u) = exp(AT )x0 + exp(A(T − s))Bu(s)ds
0
T
+ exp(A(T − s))( f (s, x(s − h))ds + σ (s, x(s − h))dω(s))
0
Let us introduce the linear controllability operator

Π0T ∈ L(L 2 (Ω, T , Rn ), L 2 (Ω, T , Rn )) as follows:
Π0T {.} = L T (L T )∗ {.}

T
= exp(A(T − t))B B ∗ exp(A∗ (T − t))E{.|t }dt
0
The corresponding controllability operator for deterministic model is:
sT = L T (s)L ∗T (s)

T
= exp(A(T − t))B B ∗ exp(A∗ (T − t))dt
s
Definition 1 The stochastic dynamic system (1) is said to be completely controllable

on [0, T ] if
RT (Uad ) = L 2 (Ω, T , Rn )
i.e., all the points in L 2 (Ω, T , Rn ) can be reached from the point x0 in time T.
Lemma 1 [17] Let G: [0, T ] × Rn → Rn×n be a strongly measurable mapping

T
such that E||G(t)|| p dt < ∞. Then
0
t p t

E G(s)dω(s) ≤ L G E||G(s)|| p ds, (3)
0 0
for all t ∈ [0, T ] and p ≥ 2, where L G is the constant involving p and T .
Lemma 2 Schwartz inequality: Let ψ1 (x) and ψ2 (x) be any two square-integrable
real functions in [a, b] then
b
2 b b
ψ1 (x)ψ2 (x)d x ≤ [ψ1 (x)] d x
2
[ψ2 (x)]2 d x
a a a
3 Main Result
Lemma 3 Assume that the operator (Π0T ) is invertible. Then for arbitrary x T ∈
L 2 (Ω, T , Rn ), f (.) ∈ L 2 ([0, T ], Rn ),σ (.) ∈ L 2 ([0, T ], Rn×n ), the control defined
as:
u(t) = B ∗ exp(A∗ (T − t))E{(Π0T )−1 p(x)|t } (4)
where
T
p(x) = x T − exp(At)x0 − exp(A(T − s)) f (s, x(s − h)ds
0
T
− exp(A(T − s))σ (s, x(s − h))dω(s))
0
transfers the system (1) from x0 ∈ Rn to the final state x T at time T , provided the
system (1) has a solution.
Proof By substituting (4) in (2), we can easily obtain the following

t
x(t; x0 , u) = exp(At)x0 + exp(A(t − s))B B ∗ exp(A∗ (T − s))E{(Π0T )−1 p(x)|s }ds
0
t
+ exp(A(t − s))( f (s, x(s − h)) + σ (s, x(s − h))dω(s))
0
Hence, for a given final time t = T, we simply have the following equality:
T
x(T ; x0 , u) = exp(AT )x0 + exp(A(T − s))(B B ∗ exp(A∗ (T − s))E (Π0T )−1 ×
0
T
x T − exp(AT )x0 − exp(A(T − s))( f (s, x(s − h))ds
0
T

+σ (s, x(s − h))dω(s)) s ds + exp(A(T − s)) f (s, x(s − h))ds
0
T
+ exp(A(T − s))σ (s, x(s − h))dω(s)
0
Thus, taking into account the form of the operator 0T , we have

x(T ; x0 , u) = exp(AT )x0 + (Π0T )(Π0T )−1 x T − exp(AT )x0
T
− exp(A(T − s))( f (s, x(s − h))ds + σ (s, x(s − h))dω(s))
0
T
+ exp(A(T − s))( f (s, x(s − h))ds + σ (s, x(s − h))dω(s))
0
= xT
Therefore, we see that the control u(t) transfers the system (1) from the initial state
x0 ∈ L 2 (Ω, T , Rn ) to the final state x T ∈ L 2 (Ω, T , Rn ) at time T.
Now we assume the following hypotheses:

(i) f and σ satisfy the Lipschitz condition with respect to x. i.e.,
|| f (t, x1 ) − f (t, x2 )||2 ≤ L 1 ||x1 − x2 ||2

||σ (t, x1 ) − σ (t, x2 )||2 ≤ L 2 ||x1 − x2 ||2
(ii) ( f, σ ) is continuous on [0, T ] × Rn and satisfies

|| f (t, x)||2 + ||σ (t, x)||2 ≤ L(||x||2 + 1)
(iii) The linear system corresponding to (1) is completely controllable.
Let us define the nonlinear operator S: H2 → H2 for t ∈ [−h, T ] as follows:

⎧
⎪ ψ(t) f or t ∈ t[−h, 0]
⎪
⎪ t
⎪
⎨
exp(At)x0 + exp(A(t − s))Bu(s)ds + exp(A(t − s)) f (s, x(s − h))ds
(Sx)(t) =
⎪
⎪ t 0 0
⎪
⎪
⎩+ exp(A(t − s))σ (s, x(s − h))dω(s) for t ∈ [0, T ]
0
From Lemma 3, the control u(t) transfer the system (1) from the initial state x0 to
the final state x T provided that the operator S has a fixed point. So, if the operator S
has a fixed point then the system (1) is completely controllable.
Now for convenience, let us introduce the notation
l1 = max{|| exp(At)||2 : t ∈ [0, T ]}, l2 = ||B||2

l3 = E||x T ||2 , M = max ||Π0T ||2
Lemma 4 [12] For every z ∈ L 2 (Ω, T , Rn ), there exists a process

ϕ(.) ∈ L 2 ([0, T ], Rn×n ) such that
T
z = Ez + ϕ(s)dω(s)
0
T
Π0T z = Γ0T Ez + ΓsT ϕ(s)dω(s)
0
Moreover,
E||Π0T z||2 ≤ ME||E{z|T }||2

≤ ME||z||2 , z ∈ L 2 (Ω, T , Rn )
Note that if the assumption (iii) holds, then for some γ > 0
E Π0T z, z
≥ γ E||z||2 , for all z ∈ L 2 (Ω, T , Rn )
(see Mahmudov [8]) and consequently
1
E||(Π0T )−1 ||2 ≤ = l4
γ
Theorem 1 Assume that the conditions (i), (ii), and (iii) hold. In addition if the
inequality
4l1 (Ml1l4 + 1)(L 1 T + L 2 L σ )T < 1 (5)

holds, then the system (1) is completely controllable.
Proof As mentioned above, to prove the complete controllability it is enough to show

that S has a fixed point in H2 . To do this, we use the contraction mapping principle.
To apply the contraction mapping principle, first we show that S maps H2 into itself.
Now by Lemmas 1 and 2, we have

E||(Sx)(t)||2 = Eψ(t) + exp(At)x0 + Π0t exp(A∗ (T − t))( 0T )−1 ×
T
x T − exp(AT )x0 − exp(A(T − s)) f (s, x(s − h))ds
0
T
− exp(A(T − s))σ (s, x(s − h))dω(s)

0
t 2

+ exp(A(t − s))( f (s, x(s − h))ds + σ (s, x(s − h))dω(s))
0

≤ 5||ψ||2 + 5l1 ||x0 ||2 + 5E t0 exp(A∗ (T − t))( 0T )−1 ×
T
x T − exp(AT )x0 − exp(A(T − s)) f (s, x(s − h))ds
0
T
2

− exp(A(T − s))σ (s, x(s − h))dω(s)
0
t
+ 5t || exp(A(t − s))||2 E|| f (s, x(s − h)||2 ds
0
t
+ 5L σ || exp(A(t − s))||2 E||σ (s, x(s − h))||2 ds
0

≤ 5||ψ||2 + 5l1 ||x0 ||2 + 20Ml1 l4 l3 + l1 ||x0 ||2
T T
+ T l1 E|| f (s, x(s − h))||2 ds + l1 L σ E||σ (s, x(s − h))||2 ds
0 0
t
+ 5l1 (T E|| f (s, x(s − h))||2 + L σ E||σ (s, x(s − h))||2 )ds
0
t
≤ B1 + B2 (T E|| f (s, x(s − h))||2 + L σ E||σ (s, x(s − h))||2 )ds
0
where B1 > 0 and B2 > 0 are suitable constants. It follows from the above and the
condition (ii) that there exists C1 > 0 such that
T
E||(Sx)(t)||2 ≤ C1 1 + E||x(r − h)||2 dr
0

≤ C1 1 + T sup E||x(t)||2
−h≤t≤T
for all t ∈ [−h, T ]. Therefore, S maps H2 into itself. Second, we show that S is a
contraction mapping, indeed.
2

E(Sx)(t) − (Sy)(t) = EΠ0t exp(A∗ (T − t))(Π0T )−1
T
× exp(A(T − s))( f (s, y(s − h)) − f (s, x(s − h)))ds
0
T
+ exp(A(T − s))(σ (s, y(s − h)) − σ (s, x(s − h)))dω(s)

0
t
+ exp(A(t − s))( f (s, x(s − h)) − f (s, y(s − h)))ds
0
2
t
+ exp(A(t − s))(σ (s, y(s − h)) − σ (s, x(s − h)))dω(s)
0
T
≤ 4Ml12 l4 T E|| f (s, x(s − h)) − f (s, y(s − h))||2 ds
0
T
+ Lσ E||σ (s, x(s − h)) − σ (s, y(s − h))||2 ds
0
t
+ 4l1 T E|| f (s, x(s − h)) − f (s, y(s − h))||2 ds
0
t
+ Lσ E||σ (s, x(s − h)) − σ (s, y(s − h))||2 ds
0
T
= 4Ml12 l4 (L 1 T + L 2 L σ ) E||x(s − h) − y(s − h)||2 ds
0
t
+ 4l1 (L 1 T + L 2 L σ ) E||x(s − h) − y(s − h)||2 ds
0
T
≤ 4l1 (Ml1 l4 + 1)(L 1 T + L 2 L σ ) E||x(s − h) − y(s − h)||2 ds
0
It results that
sup E||(Sx)(t) − (Sy)(t)||2 ≤ 4l1 (Ml1 l4 + 1)(L 1 T + L 2 L σ )T sup E||x(t) − y(t)||2

t∈[−h,T ] t∈[−h,T ]
Therefore, S is a contraction mapping if the inequality (5) holds. Then the mapping
S has a unique fixed point x(.) in H2 which is the solution of the equation. Thus the
system (1) is completely controllable. So, the theorem is proved.
Remark 1 If we consider the time-varying semilinear stochastic differential equation
of the form
⎫
d x(t) = [A(t)x(t) + B(t)u(t) + f (t, x(t − h)])dt + σ (t, x(t − h))dω(t), ⎬
t ∈ (0, T ] (6)
⎭
x(t) = ψ(t), for t ∈ [−h, 0], x(0) = x0 .
where A(t) and B(t) are the matrices of n × n and n × m, respectively, and f, σ are
defined as previously. The solution of the above equation for t > 0 is
t t
x(t; x0 , u) = φ(t, 0)x0 + φ(t, s)B(s)u(s)ds + φ(t, s) f (s, x(s − h))ds
0 0
t
+ φ(t, s)σ (s, x(s − h))dω(s) (7)
0
If the functions f, σ satisfy Lipschitz conditions and linear growth conditions and
the condition (iii) is also satisfied, then by suitably applying the above theorem, one
can show that the system (6) is completely controllable.
4 Example
Consider a two-dimensional semilinear stochastic system with delay in state
d x(t) = [Ax(t) + Bu(t) + f (t, x(t − h))]dt + σ (t, x(t − h))dω(t), t ∈ [0, T ] (8)
with initial condition x0 ∈ R2 as a random function.

Here ω(t) is a two-dimensional Wiener process and

−1 1 10
A= , B=
−1 −1 01

1 sin x(t − h) 1 x(t − h) 0

f (t, x(t − h)) = , σ (t, x(t − h)) =
a x(t − h) b 0 cos x(t − h)
Take the final point x(T ) ∈ R2 . For this system, the controllability matrix
T
Γ0T = exp(−At)B B ∗ exp(−A∗ t)dt
0

1 10
= (exp(2b) − 1)
2 01
is nonsingular if T > 0.
Moreover
2
|| f (t, x(t − h) − f (t, y(t − h))||2 ≤ ||x(t − h) − y(t − h)||2 and
a2
2
||σ (t, x(t − h) − σ (t, y(t − h))||2 ≤ 2 ||x(t − h) − y(t − h)||2
b
So, L 1 = a22 and L 2 = b22

√
Also, here ||A|| = 2, ||B|| = 2
We can see that the conditions of Theorem 1 are satisfied. So the system (8) is
completely controllable.
suggestions for improving the paper. This research is supported by the Ministry of Human Resource
and Development (MHRD) in the form of fellowship to the first author.
References
1. Kalman, R.E.: Controllability of linear systems. Contribut. Differ. Equ. 1, 190–213 (1963)
2. Curtain, R.F., Zwart, H.: An introduction to infinite-dimensional linear systems theory. Texts
in Applied Mathematics, p. 21. Springer, New York (1995)
3. Barnett, S.: Introduction to Mathematical Control Theory. Clarendon Press, Oxford (1975)
4. Klamka, J.: Stochastic controllability of linear systems with state delays. Int. J. Appl. Math.
Comput. Sci. 17(1), 5–13 (2007)
5. Klamka, J.: Stochastic controllability of linear systems with delay in control. Bullet. Polish
Acad. Sci. 55, 23–29 (2007)
6. Klamka, J., Socha, L.: Some remarks about stochastic controllability. IEEE Trans. Automatic
Control AC 22(5), 880–881 (1977)
7. Klamka, J., Socha, L.: Some remarks about stochastic controllability for delayed linear systems.
Internat. J. Control 32(3), 561–566 (1980)
8. Mahmudov, N.I.: Controllability of linear stochastic systems. IEEE Trans. Automat. Control
46(5), 724–731 (2001)
9. Mahmudov, N.I.: Approximate controllability of semilinear deterministic and stochastic evo-
lution equations in abstract spaces. SIAM J. Control Optim. 42(5), 1604–1622 (2003)
10. Mahmudov, N.I., Zorlu, S.: Controllability of nonlinear stochastic systems. J. Control 76, 95–
104 (2003)
11. Mahmudov, N.I.: On controllability of semilinear stochastic systems in Hilbert spaces. IMA J.
Math. Control Inform. 19(4), 363–376 (2002)
12. Mahmudov, N.I.: Controllability of linear stochastic systems in Hilbert spaces. J. Math. Anal.
Appl. 259(1), 64–82 (2001)
13. Mahmudov, N.I., Semi, N.: Approximate controllability of semilinear control systems in Hilbert
spaces. TWMS J. App. Eng. Math. 2(1), 67–74 (2012)
14. Mahmudov, N.I., Denker, A.: On controllability of linear stochastic systems. Internat. J. Control
73(2), 144–151 (2000)
15. Shen, L., Sun, J.: Relative controllability of stochastic nonlinear systems with delay in control.
Nonlinear Anal. Real World Appl. 13(6), 2880–2887 (2012)
16. Sakthivel, R., Kim, J.-H., Mahmudov, N.I.: On controllability of nonlinear stochastic systems.
Rep. Math. Phys. 58(3), 433–443 (2006)
17. Da Prato, G., Zabczyk, J.: Stochastic equations in infinite dimensions. Encyclopedia of Math-
ematics and its Applications. Cambridge University Press, Cambridge (1992)
On Mixed Integro-Differential Inequalities
and Applications
S.D. Kendre and S.G. Latpate
Abstract In this paper, we establish mixed integro-differential inequalities which

can be used as handy tools to study properties of solutions of certain mixed integro-
differential and differential equations.
Keywords Integrodifferential inequality · Integral equations · Explicit bound
1 Introduction
Integral inequalities involving functions and their derivatives have played a signifi-
cant role in the development of various branches of analysis. In the past few years
with the development of the theory of nonlinear differential and integral equations,
many authors have established several integral and integro-differential inequalities,
see [1, 2, 4, 6–12, 14]. These inequalities play an important role in the study of some
properties of differential, integral, and integro-differential equations. Existence of so-
lutions of a certain mixed integral and integro-differential equations were studied in
[3, 13] by M.B. Dhakne and H.L. Tidke.
In this paper, we establish mixed integro-differential inequalities which provide
an explicit bound on unknown function. In particular, we extend the result established
by B.G. Pachpatte in [12]. Some applications are also given to convey the importance
of our results.
Before proceeding with the statement of our main result, we state some important
integral inequalities that will be used in further discussion.
S.D. Kendre (B)

Department of Mathematics, Savitribai Phule Pune University, Pune (MS) 411007, India
e-mail: sdkendre@yahoo.com
S.G. Latpate
Department of Mathematics, Nowrosjee Wadia College of Arts and Science, Pune 411001, India
e-mail: sglatpate@gmail.com

DOI 10.1007/978-81-322-2485-3_44
536 S.D. Kendre and S.G. Latpate
Lemma 1 (Fangcui Jiang and Fanwei Meng [5]). Assume that a ≥ 0, p ≥ q ≥ 0,

and p = 0, then
q q q−p p p − q qp
ap ≤ k a+ k , f or any k > 0. (1)
p p
Theorem 1 (Pachpatte [12]) Let u(t), a(t), b(t), c(t) ∈ C(I = [α, β], R+ ), a(t)
be continuously differentiable on I, a (t) ≥ 0 and
t β
u(t) ≤ a(t) + b(s)u(s)ds + c(s)u(s)ds, t ∈ I. (2)
α α
β s
If p = c(s) exp b(σ )dσ ds < 1, then
α α
t t s

u(t) ≤ M exp b(s)ds + a (s) exp b(σ )dσ ds, t ∈ I, (3)
α α α
where
β s s
1
M= a(α) + c(s) a (τ ) exp b(σ )dσ dτ ds , t ∈ I. (4)
1− p α α τ
2 Main Results
In this section, we establish some nonlinear mixed integro-differential inequalities,

which can be used in the analysis of various problems in the theory of nonlinear
integrodifferential and Volterra-Fredholm integral equations.
Theorem 2 Let u(t), u (t), f (t), g(t), c(t), c (t) ∈ C(I, R+ ) and u(α) = 0. If
t β
[u (t)] p ≤ c(t) + f (s)u q (s)ds + g(s)[u (s)] p ds (5)
α α
β s
and Q̄ 5 = g(s) exp (σ − α) n 1 f (σ )dσ ds < 1, then
q
α α
β s
s
c(α) + α g(s) α c (τ ) + (τ − α) n 2 f (τ ) exp τ (σ − α) n 1 f (σ )dσ dτ ds
q q
[u (t)] p ≤
1 − Q̄ 5
t
× exp (s − α) n 1 f (s)ds
q
α

t
t
+ c (s) + (s − α)q n 2 f (s) exp (s − α)q n 1 f (s)ds ds, (6)
α s
On Mixed Integro-Differential Inequalities and Applications 537
q q−p p p − q qp
where k > 0, n 1 = k , n2 = k and p, q are the same as defined in
p p
Lemma 1.
Theorem 3 Let u(t), u (t), g(t), c(t), c (t) ∈ C(I, R+ ), f (t, s), f t (t, s)
∈ C(D, R+ ), f (t, s) be nondecreasing in t ∈ I, for each s ∈ I and u(α) = 0.
If
t β
[u (t)] p ≤ c(t) + f (t, s)u q (s)ds + g(s)[u (s)] p ds (7)
α α
β s
and Q̄ 6 = g(s) exp n 1 Ā(σ )dσ ds < 1, then
α α
β s
s
c(α) + α g(s) α c (τ ) + n 2 Ā(τ ) exp τ n 1 Ā(σ )dσ dτ ds
[u (t)] p ≤
1 − Q̄ 6
t t t

× exp n 1 Ā(s)ds + c (s) + n 2 Ā(s) exp n 1 Ā(σ )dσ ds, (8)

α α s
t
where Ā(t) = (t − α) f (t, t) + q
(s − α)q f t (t, s)ds and p, q, n 1 , n 2 are as same
α
defined in Theorem 2.
Theorem 4 Let u(t), u (t), c(t), c (t) ∈ C(I, R+ ), f (t, s), g(t, s), f t (t, s), gt (t, s)
∈ C(D, R+ ), f (t, s), g(t, s) be nondecreasing in t ∈ I, for each s ∈ I and
u(α) = 0. If
t β

[u (t)] ≤ c(t) +p
f (t, s)u (s)ds +
q
g(t, s)[u (s)] p ds (9)
α α
β s
and Q̄ 7 = g(α, s) exp n 1 B̄(σ )dσ ds < 1, then
α α

p c(α) + αβ g(α, s) αs c (τ ) + n 2 Ā(τ ) exp τs n 1 B̄(σ )dσ dτ ds
[u (t)] ≤
1 − Q̄ 7
t t t

× exp n 1 B̄(s)ds + c (s) + n 2 Ā(s) exp n 1 B̄(σ )dσ ds, (10)

α α s
where p, q, n 1 , n 2 are the same as defined in Theorem 2, Ā(t) is the same as defined in
t β
1
Theorem 3 and B̄(t) = (t −α) f (t, t) +
q
(s −α) f t (t, s)ds +
q
gt (t, s)ds.
α n1 α
2.1 Proofs of the Theorems 2–4
The proofs of the Theorems 2–4 resemble one another. Therefore, we give the detailed
proof of Theorem 4 only and the proofs of Theorems 2 and 3, can be completed by
closely looking at the proof of Theorem 4.
Proof Define a function z(t) by the right-hand side of (9)

t β
z(t) = c(t) + f (t, s)u q (s)ds + g(t, s)[u (s)] p ds,
α α
t 1
then u(t) ≤ α z p (s), z(t) is a nondecreasing,
β
z(α) = c(α) + g(α, s)[u (s)] p (s)ds (11)
α
and
t
q

z (t) ≤ c (t) + (t − α) f (t, t) +
q
(s − α) f t (t, s)ds z p (t)
q
α
β
+ gt (t, s)ds z(t). (12)
α
Applying Lemma 1 to (12), we have

t q
β
z (t) ≤ c (t) + (t − α)q f (t, t) + (s − α)q f t (t, s)ds z p (t) + gt (t, s)ds z(t)
α α
t
β
≤ c (t) + (t − α)q f (t, t) + (s − α)q f t (t, s)ds [n 1 z(t) + n 2 ] + gt (t, s)ds z(t)
α α
t β
1
= c (t) + n 1 (t − α)q f (t, t) + (s − α)q f t (t, s)ds + gt (t, s)ds z(t)
α n1 α
t
+ n 2 (t − α)q f (t, t) + (s − α)q f t (t, s)ds
α
= c (t) + n 1 B̄(t)z(t) + n 2 Ā(t),
or, equivalently
⎡ ⎤
t
⎣ z(t)
⎦ ≤ [c (t) + n 2 Ā(t)] exp − n 1 B̄(s)ds . (13)
t α
exp α n 1 B̄(s)ds
By integrating (13), we get

z(t) t
s
≤ z(α) + c (s) + n 2 Ā(s) exp − n 1 B̄(s)ds ds,
t α α
exp α n 1 B̄(s)ds
i.e.
t
z(t) ≤ z(α) exp n 1 B̄(s)ds
α
t t

+ c (s) + n 2 Ā(s) exp n 1 B̄(σ )dσ ds. (14)

α s
As [u (t)] p ≤ z(t) from (14), we have

t

[u (t)] ≤ z(α) exp
p
n 1 B̄(s)ds
α
t t

+ c (s) + n 2 Ā(s) exp n 1 B̄(σ )dσ ds. (15)

α s
Now from (11) and (15), we have

β s
z(α) ≤ c(α) + z(α) g(α, s) exp n 1 B̄(σ )dσ ds
α α
β
s
s
+ g(α, s) c (τ ) + n 2 Ā(τ ) exp n 1 B̄(σ )dσ dσ ds,
α α τ
and hence

c(α) + αβ g(α, s) αs c (τ ) + n 2 Ā(τ ) exp τs n 1 B̄(σ )dσ dτ ds
z(α) ≤ . (16)
1 − Q̄ 7
The required inequality (10) follows, from inequalities (15) and (16). This completes
the proof.
Theorem 5 Let u(t), u (t) ∈ C(I, R+ ), f (t, s), g(t, s), h t (t, s) ∈ C(D, R+ ),
u(α) = 0, f (t, s), g(t, s) be nondecreasing in t ∈ I, for each s ∈ I and c ≥ 0, be a
constant. If
t s β
[u (t)] p ≤ c + h(t, s) u q (s) + f (s, σ )u q (σ )dσ ds + g(t, s)[u (s)] p ds,
α α α
and s
β
Q̄ 8 = g(α, s) exp n 1 B̄1 (σ )dσ ds < 1, then
α α
β s s
c+ α g(α, s) α n 2 Ā1 (τ ) exp τ n 1 B̄1 (σ )dσ dτ ds
[u (t)] ≤ p
1 − Q̄ 8
t t t
× exp n 1 B̄1 (σ )dσ + n 2 Ā1 (s) exp n 1 B̄1 (σ )dσ , (17)
α α s
where
t s
Ā1 (t) = h t (t, s) (s − α)q + (σ − α q ) f (s, σ )dσ ds
α α
t
+ h(t, t) (t − α) + q
(σ − α) f (t, σ )dσ ds ,
q
α
t s
B̄1 (t) = h t (t, s) (s − α) +
q
(σ − α ) f (s, σ )dσ ds
q
α α
t β
1
+ h(t, t) (t − α) +
q
(σ − α) f (t, σ )dσ ds +
q
gt (t, s)ds
α n1 α
and p, q, n 1 , n 2 are the same as defined in Theorem 2.
Proof Let us define a function z(t) by

t s β
z(t) = c + h(t, s) u q (s) + f (s, σ )u q (σ )dσ ds + g(t, s)[u (s)] p ds,
α α α
t 1
then u(t) ≤ α z p (s),
β
z(α) = c + g(α, s)[u (s)] p ds (18)
α
and
t s
z (t) ≤ h t (t, s) (s − α)q + (σ − α q ) f (s, σ )dσ ds
α α
t
q β
+ h(t, t) (t − α)q + (σ − α)q f (t, σ )dσ zp (t) + gt (t, s)ds z(t). (19)
α α
Applying Lemma 1 to (19), we have

t s
z (t) ≤ h t (t, s) (s − α)q + (σ − α q ) f (s, σ )dσ ds
α α
t β
+ h(t, t) (t − α)q + (σ − α)q f (t, σ )dσ [n 1 z(t) + n 2 ] + gt (t, s)ds z(t)
α α
t s
≤ n1 h t (t, s) (s − α) + q
(σ − α ) f (s, σ )dσ ds
q
α α
t β
1
+ h(t, t) (t − α)q + (σ − α)q f (t, σ )dσ + gt (t, s)ds z(t)
α n1 α
t s
+ n2 h t (t, s) (s − α) +
q
(σ − α ) f (s, σ )dσ ds
q
α α
t
+ h(t, t) (t − α)q + (σ − α)q f (t, σ )dσ
α
≤ n 1 B̄1 (t)z(t) + n 2 Ā1 (t),
or, equivalently,
⎡ ⎤
t
⎣ z(t) ⎦
≤ n 2 Ā1 (t) exp − n 1 B̄1 (σ )dσ . (20)
t
exp α n 1 B̄1 (σ )dσ α
By integrating (20), we obtain an estimate

t t t
z(t) ≤ z(α) exp n 1 B̄1 (σ )dσ + n 2 Ā1 (s) exp n 1 B̄1 (σ )dσ ds. (21)
α α s
As [u (t)] p ≤ z(t) from (21), we have

t t t
[u (t)] p ≤ z(α) exp n 1 B̄1 (σ )dσ + n 2 Ā1 (s) exp n 1 B̄1 (σ )dσ . (22)
α α s
Now from Eq. (18) and (22), we have

β s s
c+ α g(α, s) α n 2 Ā1 (τ ) exp τ n 1 B̄1 (σ )dσ dτ ds
z(α) ≤ . (23)
1 − Q̄ 8
From (22) and (23), we get the required inequality (17) and hence the proof.
3 Applications
One of the main motivations for the study of different type inequalities given in
the previous sections is to apply them as tools in the study of various classes of
integral equations. In the following section we give application of some theorems
of previous sections. In fact we discuss the boundedness behavior of solutions of a
nonlinear mixed integral equations.
Example 1 Consider the following general mixed nonlinear integro-differential
equation
β

t
[y (t)] = x(t) +
p
F t, s, y q (s) ds + G t, s, [y (s)] p ds, (24)
α α
for t ∈ I, where p ≥ q ≥ 0 with p = 0,y(t) is unknown function, x ∈

C(I, R ), F, G ∈ C(D × R , R ), I = [α, β], D = (t, s) ∈ I : α ≤ s ≤ t ≤ β , R n is
n n n 2
n dimensional Euclidean space with norm |.|.

We assume that every solution y(t) of Eq. (24) exists on I and functions x, y, F, G
involved in the Eq. (24) satisfy the following conditions:
|x(t)| ≤ c(t), (25)

|F(t, s, y )| ≤ f (t, s) |y| ,
q q
(26)
|G(t, s, y p )| ≤ g(t, s)|y | p , (27)
where c, f, g are the same as defined in Theorem 4. From Eqs. (24) and (25)–(27),
we obtain
t β
|y (t)| p ≤ c(t) + f (t, s)|y(s)|q ds + g(t, s)|y (s)| p ds. (28)
α α
Applying Theorem 4 to (28), we get the following explicit known bound:

β s
s
c(α) + α g(α, s) α c (τ ) + n 2 Ā(τ ) exp τ n 1 B̄(σ )dσ dτ ds
|y (t)| ≤
p
1 − Q̄ 7
t
t t
× exp n 1 B̄(s)ds + c (s) + n 2 Ā(s) exp n 1 B̄(σ )dσ ds,

α α s
provided Q̄ 7 < 1, where Ā, B̄, Q̄ 7 , n 1 , n 2 are the same as defined in Theorem 4.
Example 2 We calculate the explicit bound on a solution of the following nonlinear

integral equation:
t 1
1 2 2
[u (t)] = 4 + 3
u (s)ds + s[u (s)]3 ds, (29)
1
3
s 1
3
where u(t) is defined

as in Theorem 2 and we assume that every solution u(t) of (29)
1 1
exists on I = , . Also, here
3 2
1 s
2 2 1 3 1√ 3k
5
√
3k
√
3
Q̄ 5 = s exp √ σ dσ ds = e 27 e 108 −1 k < 1, for k > 0.
1
3
3 3k 0 2
Hence, by Theorem 2 and Eq. (29), we get

⎡ 1
2 1 t σ dσ 2 1
2/3 s τ exp 2 √ 1 s σ dσ dτ ds + 4
⎢ exp 3 √3 1 1 3s k 1 3 k
3 τ
⎢ k 3
u (t) ≤ ⎢ 3 3

⎣ 21 s
1 − 1 s exp 23 √ 3
1
0 σ dσ ds
3 k
t 13
1 t 2 1
+ k 2/3 s exp √ σ dσ ds
3 1 3 3k s
3
⎡
2
⎤ 13
2 t2 − 18
1
⎢ ⎛ ⎞
5
√ √ ⎥
⎢ 3 3k 5k + 4
− 1 k 4/3 − 144
3 ⎥
⎢ t 2 − 91 √ 4 e 108 e 3 k ⎥
⎢ 1 2/3 ⎜ 3 √ √ 3 3
k ⎟ ⎥
= ⎢ ⎥
3 3
⎢3k ⎝ ke 3 k − ⎠+ ⎥
⎢ 2 2 1
√
3
5
√
3k √ ⎥
⎢ 1 − 23 e 27 k e 108 −1 3
k ⎥
⎣ ⎦
By integrating the above inequality we get the desired bound for u(t).
References
1. Abdeldaim, A., Yakout, M.: On some new inequalities of Gronwall-Bellman-Pachpatte type.

Appl. Math. Comput. 217, 7887–7899 (2011)
2. Bellman, R.: The stability of solutions of linear differential equations. Duke Math. J. 10, 643–
647 (1943)
3. Dhakne, M.B., Lamb, G.B.: On globle existance of abstrct nonlinear integrodifferential equa-
tions. Indian J. Pure Appl. Math. 33(5), 665–676
4. El-Owaidy, H., Ragab, A., Abdeldaim, A.: On some new integral inequalities of Growall-
Bellman type. Appl. Math. Comput. 106, 289–303 (1999)
5. Jiang, F., Meng, F.: Explicit bounds on some new nonlinear integral inequality with delay. J.
Comput. Appl. Math. 205, 479–486 (2007)
6. Meng, F.W., Li, W.N.: On some new integral inequalitise and their applications. Appl. Math.
Comput. 148, 381–392 (2004)
7. Gronwall, H.T.: Note on the derivatives with respect to a parameter of the solutions of a system
of differential equations. Ann. Math. 2, 292–296 (1918–1919)
8. Li, L., Meng, F., He, L.: Some generalized integral inequalities and their applications. J. Math.
Anal. Appl. 372, 339–349 (2010)
9. Pachpatte, B.G.: A note on Gronwall type integral and integro differential inequalities. Tamkang
J. Math. 8, 53–59 (1977)
10. Pachpatte, B.G.: On some fundamental integro differential inequalities for differential equa-
tions. Chinese J. Math. 6(1), 17–23 (1978)
11. Pachpatte, B.G.: Inequalities for Differential and Integral Equations. Academic Press, New
York and London (1998)
12. Pachpatte, B.G.: Integral and finite difference inequalities and applications. In: North-Holland
Mathematics Studies, vol. 205. (2006)
13. Tidke, H.L., Dhakne, M.B.: Existance of solutions for nolinear mixed type integrodifferential
equation of second order. Surv. Math. Appl. 5, 61–72 (2010)
14. Young-Ho, K.: Gronwall-Bellman and Pachpatte type integral inequalities with applications.
Nonlinear Anal. 71, e2641–e2656 (2009)
Buckling and Vibration of Functionally
Graded Circular Plates Resting on Elastic
Foundation
Neha Ahlawat and Roshan Lal
Abstract The axisymmetric vibrations of functionally graded circular plates sub-

jected to uniform in-plane force resting on elastic foundation have been analysed on
the basis of classical plate theory. The material properties, i.e. Young’s modulus and
density vary continuously through the thickness of the plate and obey a power law
distribution of the volume fraction of the constituents. Differential transform method
was employed to solve the differential equation governing the motion of simply sup-
ported plates. The effect of various plate parameters was studied on the first three
modes of vibration. By allowing the frequency to approach zero, the critical buckling
loads for the plate were computed. A comparison of results with those available in
the literature has been presented.
Keywords Functionally graded circular plates · Buckling · Differential transform ·

Foundation · Axisymmetric vibrations
1 Introduction
Many researches dealing with vibration characteristics of functionally graded ma-

terial (FGM) plates have appeared in the literature due to their wide application in
nuclear energy reactors, solar energy generators, space shuttle, etc. FGMs are usu-
ally made from a mixture of ceramic and metal as they are able to withstand high
temperature gradient environments while maintaining their structural integrity.
Numerous studies on the static/dynamic behaviour of FGM plates of various
geometries have been made and reported in Refs. [1–6], to mention a few. Of these,
Ref. [1] is an excellent review of the work upto 2012, on the deformation, stress, vibra-
tion and stability problems of FG plates. In Ref. [2], Feldman and Aboudi analysed the
bifurcational buckling of FG rectangular plates employing a combination of micro-
mechanical and structural approach. Abrate [3] studied the free vibrations, buckling
N. Ahlawat (B) · R. Lal

Department of Mathematics, Indian Institute of Technology Roorkee, Roorkee 247667, India
e-mail: ahlawatneha@gmail.com
R. Lal
e-mail: rlatmfma@iitr.ac.in
DOI 10.1007/978-81-322-2485-3_45
546 N. Ahlawat and R. Lal
and static deflections of FG plates. The technique of differential transform developed

by Zhou [4] and Chen and Ho [5] for solving the initial value problems was employed
by Malik and Dang [6] in studying the free vibration of continuous systems, particu-
larly for thin beams. Later, this method was extended to study the free axisymmetric
vibrations of isotropic/ FGM circular plates of uniform/non-uniform thickness with
various boundary conditions/constraints in Refs. [7, 8]. Kumar and Lal [9] predicted
the natural frequencies for axisymmetric vibrations of two-directional FG annu-
lar plates resting on Winkler foundation using differential quadrature method and
Chebyshev collocation technique.
In the present study, the effect of in-plane force on the axisymmetric vibrations of
FGM circular plates resting on elastic foundation using differential transformation
method (DTM) has been studied. The material properties, i.e. Young’s modulus and
density are assumed to be graded in the thickness direction and these properties
vary according to a power-law in terms of volume fractions of the constituents.
The natural frequencies are obtained for simply supported boundary condition with
different values of volume fraction index, in-plane force parameter and foundation
parameter. A comparison of results is given.
2 Mathematical Formulation
Consider an FGM circular plate of radius a, thickness h, mass density ρ subjected

to uniform in-plane tensile force N0 , resting on elastic foundation of modulus k f
and referred to a cylindrical polar coordinate system (R, θ, z), z = 0 being the
middle plane of the plate. The top and bottom surfaces are z = +h/2 and z = −h/2,
respectively. The line R = 0 is the axis of the plate. The equation of motion governing
transverse axisymmetric vibration of the present model (Fig. 1) is given by [13]:
2 1 1
Dw,R R R R + Dw,R R R − 2 D + R 2 N0 w,R R + 3 D − R 2 N0 w,R + k f w + ρhw,tt = 0
R R R
(1)
where w is the transverse deflection, D the flexural rigidity and ν the Poisson’s ratio.
Here, a comma followed by a suffix denotes the partial derivative with respect to that
variable.
For a harmonic solution, the deflection w can be expressed as
w(R, t) = W (R)eiωt (2)
where ω is the radian frequency. Equation (1) reduces to
2 1 1 N0
DW,R R R R + DW,R R R − 2 DW,R R + 3 DW,R − N0 W R R − W,R + k f W − ρhω2 W = 0
R R R R
(3)
Assuming that the top and bottom surfaces of the plate are ceramic and metal-rich,
respectively, for which the variations of the Young’s modulus E(z) and the density
Buckling and Vibration of Functionally Graded Circular Plates ... 547
Fig. 1 Functionally graded circular plate under uniform tensile load N0
ρ(z) in the thickness direction are taken as

g
z 1
E(z) = (E c − E m ) + + Em (4)
h 2
g
z 1
ρ(z) = (ρc − ρm ) + + ρm (5)
h 2
where E c , ρc and E m , ρm denote the Young’s modulus and the density of ceramic
and metal constituents, respectively, and g is the volume fraction index.
The flexural rigidity and mass density are given by
h/2
1
D = E(z)z 2 dz (6)
1 − ν2 −h/2
h/2
1
ρ = ρ(z)dz (7)
h −h/2
Substituting Eq. (4) and Eq. (5) into Eq. (6) and Eq. (7), we obtain

h3 g2 + g + 2 Em
D = (E c − E m ) + (8)
1 − v2 4(g + 1)(g + 2)(g + 3) 12
ρc + ρm g
ρ = (9)
g+1
Using non-dimensional variables, r = R/a, f = W/a, Eq. (3) now becomes
2 1 1 N0 2
D f ,rrrr + D f ,rrr − 2 D f ,rr + 3 D f ,r − N0 a 2 f ,rr − a f ,r + k f f = ρa 4 Ω 2 f h (10)
r r r r
Substituting the values of D and ρ from Eq. (8) and Eq. (9), the Eq. (10) can be
written as

ρc + ρm g
r 3 B f ,rrrr + 2r 2 B f ,rrr − r B f ,rr + B f ,r − Nr 3 f ,rr − Nr 2 f ,r + k f r 3 f = r 3 Ω 2 f
ρc (g + 1)
(11)
ρc ha 4 2 Ec h 3 kf a4
where D = D ∗ B, N = N0 2
D∗ a , Ω2 = D∗ ω , D∗ = 12(1−ν 2 )
, Kf = D∗

Em g2 + g + 2 Em
B = 3 1− +
E c (g + 1)(g + 2)(g + 3) Ec
2.1 Boundary Condition
The relations which should be satisfied for a simply-supported plate at the edge are
2

d f 1 df
f (1) = 0, Mr |r =1 = −D +v = 0 (12)
dr 2 r dr r =1
where Mr is the radial bending moment.
2.2 Regularity conditions
For the axisymmetric boundary conditions, the regularity conditions at the centre
(r = 0) of the circular plate can be defined as

df d3 f 1 d2 f 1 df
|r =0 = 0, Q r |r =0 = 3
+ 2
− 2 = 0 (13)
dr dr r dr r dr r =0
3 Method of Solution: Description of the Method
The differential transform of the kth derivative of f (r ) is given by

1 d k f (r )
Fk = (14)
k! dr k r = r0
where f (r ) is the original function and Fk is the transformed function.

The inverse transformation of the function Fk is defined as
∞

f (r ) = (r − r0 )k Fk
k=0
In actual applications, the function f (r ) is expressed by a finite series. So, the above
expression may be written as

n
f (r ) = (r − r0 )k Fk (15)
k=0
The convergence of the natural frequencies decides the value of n. Some basic the-
orems which are frequently used in practical problems are given in Table 1.
3.1 Transformation of the Governing Differential Equation
Applying the transformation rules given in Table 1, the transformed form of the
governing differential equation (11) around r0 = 0 can be written as

ρc + ρm g
B[(k 2 − 1)2 ]Fk+1 + K f Fk−3 − N (k − 1)2 Fk−1 = Ω 2 Fk−3
ρc (g + 1)
(16)
3.2 Transformation of the Boundary/Regularity Conditions
By applying transformations rules given in Table 1, the Eq. (12) and Eq. (13) becomes

n
n
Fk = 0, [k(k − 1) + vk]Fk = 0 (17)
k=0 k=0
F1 = F3 = F5 = F7 = · · · = F4k+1 = F4k+3 = 0 (18)
Table 1 Transformation Original functions Transformed functions

rules for one-dimensional
DTM f (r ) = g(r ) ± h(r ) Fk = G k ± Hk
f (r ) = λg(r ) Fk = λG k
k
f (r ) = g(r )h(r ) Fk = l=0 G l Hk−l
d n g(r ) (k + n)!
f (r ) = Fk = k! G k + n
dr n
1, k = n
f (r ) = rn Fk = δ(k − n) =
0, k = n
4 Frequency Equations
The use of regularity condition (18) in the equation of motion, i.e. Eq. (16), gives

1 2 ρc + ρm g
F2k = . 4N (k − 1)2
F2k−2 + Ω − K f F2k−4 (19)
16Bk 2 (k − 1)2 ρc (g + 1)
Also, the use of regularity condition (18) in the relations (17) given for simply
supported edge condition leads to

n
n
F2k = 0, [2k(2k − 1) + 2vk]F2k = 0 (20)
k=0 k=0
Applying the boundary condition (20) on the resulting F2k expressions, we get the
following equations:
(m) (m)
Φ11 (Ω)F0 + Φ12 (Ω)F2 = 0
(m) (m)
Φ21 (Ω)F0 + Φ22 (Ω)F2 = 0 (21)
(m) (m) (m) (m)

where Φ11 ,Φ12 ,Φ21 and Φ21 are polynomials in Ω of degree m where m = 2n.
Equation (21) can be expressed in matrix form as follows

(m) (m)
Φ11 (Ω) Φ12 (Ω) F0 0
(m) (m) = (22)
Φ21 (Ω) Φ22 (Ω) F2 0
For a non-trivial solution of Eq. (22), the frequency determinant must vanish and
hence
(m) (m)
Φ11 (Ω) Φ12 (Ω)
(m) (m) =0 (23)
Φ21 (Ω) Φ22 (Ω)
The frequency Eq. (23) provides the values of the frequency parameter Ω. The lowest
three roots of this equation have been obtained using MATLAB. The values of elastic
constants are
Table 2 Convergence study for first three modes of vibration for N = 20, K f = 100, g = 5
No. of terms First mode Second mode Third mode
10 17.2528 37.6839 68.9160
12 17.2528 37.6838 71.5054
14 17.2528 37.6838 71.6026
16 17.2528 37.6838 71.6037
18 17.2528 37.6838 71.6037
90
80
70
60
50
Ω
40
30
20
10
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
g
Fig. 2 Frequency parameter Ω for K f = 50; _ _ _ _ _, N = 20; ,N =

0; .........., N = −10, Box, firstmode; , secondmode; , thirdmode
Youngs modulus and density for aluminium: E m = 70 Gpa, ρm = 2, 702 kg/m 3

and for alumina: E c = 380 Gpa, ρc = 3, 800 kg/m 3
The values of the volume fraction index g are taken as: 0, 1, 3, 5; the in-plane
force parameter N : −20, −10, 0, 10, 20, foundation parameter K f : 10, 50, 100
and v = 0.3. In order to choose an appropriate value of the number of terms ‘n’
in Eq. (20), a computer program has been developed and run for various values of
g, N and K f . The convergence of frequency parameter for the first three modes
of vibration for a specified plate taking g = 5, N = 20 and K f = 100 is shown
in Table 2, as maximum deviations were observed for this data. The value of n has
been fixed as 18. The values of frequency parameter Ω for different values of plate
parameters has been given in Table 3. Figure 2 shows the behaviour of Ω verses g for
N = −10, 0, 20 and K f = 50 for all the three modes of vibration. The value of Ω
90
80
70
60
50
Ω
40
30
20
10
0
−80 −70 −60 −50 −40 −30 −20 −10 0 10 20
Ncr
Fig. 3 Critical buckling loads Ncr for K f = 50; _ _ _ _ _, g = 5; ,g =

0; Box, first mode; , second mode; , third mode
increases with the increasing value of N , whatever be the value of g and decreases
vice versa. The value of Ω is found to decrease with increasing value of g for both
tensile and compressive values of N except in fundamental mode of vibration. In
this case, Ω is found to increase continuously for tensile in-plane forces. The graph
for critical buckling load for the three mode of vibration for g = 0 and 5 for the plate
has been plotted in Fig. 3. It can be seen that the values of critical buckling loads Ncr
for an isotropic plate (g = 0) are higher than that for an FGM plate (g = 5) for all the
three modes of vibration. Figure 4 shows the effect of foundation parameter K f on
Ω for g = 5 and g = 0 for all the modes. It is clear that Ω increases with increasing
value of K f . The values of critical buckling load parameter Ncr in compression for
different values of volume fraction index g = 0, 1, 3, 5 for K f = 50 are reported
in Table 4. The values of Ncr decrease with the increasing value of g. The results
for Ncr for an isotropic plate have been compared in Table 5 with Gupta and Ansari
[10] obtained by using Ritz method and Vol’mir [11] exact solutions. An excellent
agreement among the results has been noticed.
The results for Ω have been compared in Table 6 with Leissa [12] obtained by
series 151 solution and Wu et al. [13] obtained by generalized differential quadra-
ture method. A close agreement of the results shows the versatility of the present
technique.
Table 3 Frequency parameter Ω for FGM circular plate
g N K f = 10 K f = 50 K f = 100
↓ ↓ I II III I II III I II III
0 -20 * 16.8557 63.3356 * 18.0032 63.6506 2.8229 19.3420 64.0422
-10 * 24.2607 68.9946 4.0369 25.0716 69.2838 8.1423 26.0496 69.6437
0 5.8614 29.8878 74.2235 8.6230 30.5496 74.4924 11.1515 31.3573 74.8273
10 9.6085 34.6126 79.1076 11.5031 35.1857 79.3600 13.5027 35.8892 79.6744
20 12.2575 38.7661 83.7072 13.7930 39.2786 83.9458 15.4999 39.9100 84.2431
1 -20 * * 45.4645 * * 45.9758 * 4.8035 46.6071
-10 * 16.3378 54.2428 2.7105 17.7110 54.6720 8.1111 19.2905 55.2039
0 5.3428 24.9600 61.7867 8.6776 25.8796 62.1639 11.5648 26.9852 62.6323
10 9.8124 31.2926 68.5052 11.9598 32.0310 68.8456 14.1944 32.9306 69.2687
20 12.8046 36.5442 74.6212 14.5160 37.1784 74.9339 16.4060 37.9562 75.3228
3 -20 * * 37.3981 * * 38.0747 * * 38.9040
-10 * 12.3957 48.5235 2.0263 14.3080 49.0469 8.2425 16.3876 49.6934
0 5.2319 23.2909 57.5376 8.8566 24.3625 57.9797 11.9278 25.6392 58.5276
Buckling and Vibration of Functionally Graded Circular Plates ...
10 10.0671 30.5213 65.3197 12.3456 31.3467 65.7094 14.7053 32.3489 66.1934

20 13.2382 36.3408 72.2687 15.0438 37.0367 72.6212 17.0338 37.8887 73.0593
5 -20 * * 32.6063 * * 33.4045 * * 34.3761
-10 * 9.8802 45.2698 1.5612 12.2599 45.8480 8.2641 14.7025 46.5607
0 5.1535 22.3306 55.0973 8.9020 23.4807 55.5734 12.0459 24.8435 56.1628
10 10.1426 30.0014 63.4202 12.4723 30.8670 63.8342 14.8801 31.9160 64.3480
20 13.3833 36.0767 70.7711 15.2250 36.7996 71.1424 17.2528 37.6838 71.6037
*denotes the fact that frequency parameter does not exist because of buckling
553
90
80
70
60
Ω
50
40
30
20
10
10 20 30 40 50 60 70 80 90 100
Kf
Fig. 4 Frequency parameter

for , g = 5 N = 20; _._._._._, g = 0 N =
20; and K f = 10, 50, 100; Box, first mode; , second mode; , third mode
Table 4 The critical buckling loads in compression for K f = 50

Modes g=0 g=1 g=3 g=5
I 12.8013 11.0752 10.5465 10.3115
II 30.7213 18.8872 15.3517 13.8384
III 74.1492 44.1788 35.1538 31.3409
Table 5 Comparison of critical buckling load parameter Ncr for isotopic plate (g = 0, K f = 0)
Reference First mode Second mode Third mode
Present 4.1978 29.0452 73.4768
Gupta and Ansari [10] 4.1978 29.0452 73.4768
Vol’mir [11] 4.1978 29.0452 73.4768
Table 6 Comparison of frequency parameter Ω for N = 0, g = 0, K f = 0

Reference First mode Second mode Third mode
Present 4.9351 29.7200 74.1561
Leissa [12] 4.977 29.76 74.20
Wu et al. [13] 4.935 29.72 74.156
6 Conclusions
Numerical results show that the values of the frequency parameter Ω for an isotropic
plate are higher than that for FGM plate for the same set of values of other parameters.
The value of the frequency parameter Ω increases with increasing value of in-plane
force parameter, whatever be the value of g. The value of the frequency parameter
increases in presence of the foundation, whatever be the value of g and N .
References
1. Jha, D.K., Kant, T., Singh, R.K.: A critical review of recent research on functionally graded
plates. Compos. Struct. 96, 833–849 (2013)
2. Feldman, E., Aboudi, J.: Buckling analysis of functionally graded plates subjected to uniaxial
loading. Compos. Struct. 38(1–4), 29–36 (1997)
3. Abrate, S.: Free vibration, buckling, and static deflections of functionally graded plates. Com-
pos. Sci. Technol. 66(14), 2383–2394 (2006)
4. Zhou, J.K.: Differential transformation and its application for electrical circuits. Huazhong
University Press, Wuhan, China (1986)
5. Chen, C.K., Ho, S.H.: Solving partial differential equations by two dimensional differential
transform. Appl. Math. Comput. 106(2–3), 171–179 (1999)
6. Malik, M.: Huy Dang, H.: Vibration analysis of continuous systems by differential transfor-
mation. Appl. Math. Comput. 96(1), 17–26 (1998)
7. Yalcin, H.S., Arikoglu, A., Ozkol, I.: Free vibration analysis of circular plates by differential
transformation method. Appl. Math. Comput. 212(2), 377–386 (2009)
8. Shariyat, M., Alipour, M.M.: A differential transform approach for modal analysis of variable
thickness two-directional FGM circular plates on elastic foundations. Iran. J. Mech. Eng. 11,
15–38 (2010)
9. Kumar, Y., Lal, R.: Prediction of frequencies of free axisymmetric vibration of two-directional
functionally graded annular plates on Winkler foundation. Eur. J. Mech. A. Solids 42, 219–228
(2013)
10. Gupta, U.S., Ansari, A.H.: Asymmetric vibrations and elastic stability of polar orthotropic
circular plates of linearly varying profile. J. Sound Vib. 215(2), 231–250 (1998)
11. Vol’mir, A.S.: Stability of elastic systems. FTD-MT-64-335, WP-FB, Ohio (1966)
12. Leissa, A.W.: Vibration of plates. NASA SP. 160, Washington (1969)
13. Wu, T.Y., Wang, Y.Y., Liu, G.R.: Free vibration analysis of circular plates using generalized
differential quadrature rule. Comput. Meth. Appl. Mech. Eng. 191, 5365–5380 (2002)
Discrete Fourier Transform and Extended
Modified Hermite Polynomials
R.A. Malekar
Abstract Linear combinations of the solutions for modified second-order linear or-
dinary differential equations are related to eigenfunctions of discrete Fourier trans-
form. This leads to in particular linear combinations of extended modified Hermite
polynomials as eigenfunctions of discrete Fourier transform.
Keywords Discrete Fourier transform · Modified second order linear ordinary

differential equation · Extended modified Hermite polynomials
1 Introduction
The Hemite functions are well-known eigenfunctions of the continuous Fourier trans-
form. One of the problem associated with FT is to find a wider class of eigenfunctions.
There has been work in this direction but the same has not been the case for the Dis-
crete Fourier transform (DFT). The major work in this direction is based on the
eigenfunctions of the continuous Fourier transform (FT) [1, 2]. The eigenfunctions
of DFT are expressed in terms of derivatives of Jacobi theta functions [3]. Some
of the well-known classical theta function identities are derived using the DFT [4,
5]. The recent work also uses the classical approach of constructing eigenfunctions
of DFT in terms of Hermite and Gaussian functions, which are eigenfunctions of
FT [6].
The eigenfunctions of DFT was generalized by Matveev [7] in terms of absolutely
summable series. We extend the work of Matveev in this paper to generate eigenfunc-
tions in terms of solutions to the modified second-order linear ordinary differential
equations. This is illustrated in particular as a periodic extension of Hermite poly-
nomials so that their linear combinations are eigenfunctions of DFT. This leads to
another important class of eigenfunctions of DFT which comes from the solutions
of modified second-order linear ordinary differential equation.
R.A. Malekar (B)

National Defence Academy, Khadakwasala, Pune 411023, India
e-mail: rajanish1973@gmail.com

DOI 10.1007/978-81-322-2485-3_46
558 R.A. Malekar
The well-known basic facts about DFT are discussed in Sect. 2. In Sect. 3 the linear
combinations of solutions to modified second-order ODE are related as eigenfunc-
tions of DFT. In Sect. 4 various properties of extended modified Hermite polynomials
are discussed. In Sect. 5 linear combinations of extended modified Hermite polyno-
mials are related to eigenfunctions of the DFT.
2 Preliminary Results
The matrix Φ(n) corresponding to the discrete Fourier transform of size n is given
by
1 2πi
Φ jk (n) = √ q jk , j, k = 0, · · · , n − 1, q=e n . (1)
n
Definition 1 For f = ( f 0 , · · · , f n−1 )t ∈ C n we define the discrete Fourier trans-

form f˜ ∈ C n by f˜ = Φ f = ( f˜0 , f˜1 , · · · , fñ−1 ), where
1
n−1
2πi jk
f˜k = √ fj e n .
n
j=0
It is clear that Φ 4 = I . Then any sequence (vector) see [7], f j for j = 0, 1, 2, · · · ,

n − 1 generates eigenvector v(k) of Φ : Φv = i k v via formula
v j (k) = f j + (−1)k f˜j + (−1)k f n− j + (−i)3k f˜− j . (2)
The multiplicities of the eigenvalues of the DFT Φ(n) are given by
n = 4m + 2 ⇒ m 1 = m, m 2 = m + 1, m 3 = m, m 0 = m + 1.
n = 4m ⇒ m 1 = m, m 2 = m, m 3 = m − 1, m 0 = m + 1.
n = 4m + 1 ⇒ m 1 = m, m 2 = m, m 3 = m, m 0 = m + 1.
n = 4m + 3 ⇒ m 1 = m + 1, m 2 = m + 1, m 3 = m, m0 = m + 1
where m k is the multiplicity of i k see [8].
3 Modified Second-Order Linear ODE and DFT
The well-known special functions are solutions of second-order linear ODEs which
are expressed as a power series. The fundamental existence and uniqueness theorem
of ODE asserts that the linear second-order ODE of the form
Discrete Fourier Transform and Extended Modified Hermite Polynomials 559
d 2w dw
+ P(z) + Q(z)w(z) = 0 (3)
dz 2 dz

with P(z), Q(z) are analytic and initial conditions w(0) = c0 and w (0) = c1 has a
unique analytic solution in some region |z| < R see [9].
Let
gm = cm z m = w(z)
m ≥0 m ≥0
be a summable power series which is a solution of (3). In order to extend the sum-
mation over Z, define
gm = 0 if m < 0,
= cm z m if m ≥ 0.

gm = gm = cm z m = w(z).
m∈Z m ≥0 m ≥0
2πi
Let q = e n . Define a continuous periodic function w(z, t) of t ∈ R, and a contin-
uous periodic sequence η j+n = η j by

w(z, t) = gm q mt and η j = gnm+ j .
m∈Z m∈Z
We have w(z, t) = w(q t z). Then by the change of variable in (3) the modified function
w(z, t) is the solution of the modified homogeneous linear differential equation given
by
1 d 2 w(z, t) 1 dw(z, t)
2t 2
+ P(q t z) t + Q(q t z)w(z, t) = 0. (4)
q dw q dz
We give the relationship between solutions of (4) and the eigenfunctions of DFT.
Theorem 1 Let w(z, t) be an analytic solution of modified linear second order
homogeneous differential equation (4) with the initial conditions w(0, t) = c0 ,
w (0, t) = c1 and analytic coefficients. Then the vector v(z, k) with the components
v j (z, k) given by
1 (−i)k
n−1
v j (z, k) = w(z, l) + (−1)k w(z, −l) q − jl + √ w(z, j) + (−1)k w(z, − j)
n n
l=0
(5)
is an eigenvector of DFT: Φv = i k v.
Proof We note the following:

560 R.A. Malekar

n−1
q jl η j = w(z, l).
j=0
Applying the inversion formula we get
n−1
1
gnm+ j = η j = w(z, l)q − jl , j = 0, 1, 2 · · · , n − 1
n
m ≥0 l =0
Consider in (5) the following term:
(−i)k
√ w(z, j) + (−1)k w(z, − j)
n
(−i)k (−i)3k
= √ w(q j z) + √ w(q − j z)
n n
(−i) k (−i)3k
= √ gm q m j + √ gm q −m j
n n
m ≥0 m ≥0
n−1 n−1
(−i)k
(kn + r ) j (−i)3k
= √ gkn + r q + √ gkn + r q −(kn + r ) j
n n
r =0 k ≥0 r =0 k ≥0
n−1 n−1
(−i)k (−i)3k
= √ ηr q r j + √ ηr q −r j
n n
r =0 r =0
Therefore (5) becomes
v j (z, k) = η j + (−1)k η− j + (−i)k η̃ j + (−i)3k η̃− j . (6)
From (2) it is clear that v j (z, k) is a component of the eigenvector for DFT.
The particular cases of the differential equation (4) give rise to modified form of
well-known special functions. The extended modified Hermite polynomials arise in
this manner. We show in the next section that the linear combinations of the Hermite
polynomials, the modified Hermite polynomials are eigenfunctions of DFT. These
statements are verified in the following section. The properties of these functions are
also studied.
4 Extended Modified Hermite Polynomials

as Eigenfunctions of DFT
The classical orthogonal polynomials are solutions of second-order ODE. From The-
orem 1 we have a linear combinations of all classical orthogonal polynomials such as
Legendre, Hermite, Laguarre and Chebyshev in their modified form as an eigenfunc-
tions of DFT. In this section we apply the result to the extended modified Hermite
polynomials (EMHP) as eigenfunctions of DFT. The extended modified Hermite
polynomials (EMHP) are defined by
n

2
(−1)r n! 2πi
Hn (z, t) = (2z)n−2r q r t and t ∈ R, q=e n
r ! (n − 2r )!
r =0
At t = 0 we get the classical Hermite polynomials. EMHP are related to Hermite

polynomials by t nt
Hn (z, t) = Hn q − 2 z q 2 . (7)
t
At q − 2 = i, (7) reduces to modified Hermite polynomials given by
Hn (z, t) = Hn (i z)i −n . (8)
The modified polynomials of the type (8) are discussed in [10]. The paper [10]
gives combinatorial interpretation of modified polynomials in terms of combinatorial
probability defined on compound urn model. A real extension of classical Hermite
polynomials in the modified form is discussed in [11]. They have been identified in
the spectral approximation of boundary layer problems.
It is clear from (4), the functions Hn (z, t) for a particular value of t is the solution
of modified Hermite differential equation given by
d2 y dy
qt 2
− 2z + n(n + 1)y = 0. (9)
dz dz
The Rodrigues formula for EMHP in the modified form is given by
−t x 2 d n −(q −t x 2 )
Hn (z, t) = (−1)n q nt eq e . (10)
dxn
Simple calculations show that
H0 (z, t) = 1, H1 (z, t) = 2z, H2 (z, t) = 4z 2 − 2q t ,

H3 (z, t) = 8z 3 − 12zq t , H4 (z, t) = 16z 4 − 48q t x 2 + 12q 2t .
562 R.A. Malekar
In general, for all n ≥ 1 the following relation holds (see [12]):

Hn (z, t) = 2z Hn−1 (z, t˘) − q t Hn−1 (z, t˘) (11)
where t˘ = (n−1)t
n .
The following result follows from of Theorem 1:
Proposition 1 The vector Hn (z, k, t) with components
1
n−1
v j (z, k) = Hn (z, l) + (−1)k Hn (z, −l) q − jl
n
l=0
(−i)k
+ √ Hn (z, j) + (−1)k Hn (z, − j)
n
is an eigenvector of DFT Φv = i k v.
It is interesting to see the explicit form of eigenvectors of DFT Φ(n) in terms
of EMHP for some particular values of n. For n = 3 eigenvector corresponding to
eigenvalue +1 is given by
√ √
t
1 3
√ 3 3−1 3 3−1
v(z, 0) = √ 8z ( 3 + 1) − 12z, 8z − 12z, 8z − 12z .
3 2 2
The eigenvector corresponding to eigenvalue −1 is given by

√ √
t
1 3
√ 3 3+1 3 3+1
v(z, 2) = √ 8z ( 3 − 1) + 12z, 8z − 3z, 8z − 12z .
3 2 2
It is clear that all eigenvectors are real for real values of z though the polynomials
used are with imaginary argument. It is possible to construct eigenvectors of DFT for
further values of n applying the same techniques. As this construction is independent
of the orthogonal polynomials involved, the same construction can be done for other
orthogonal polynomials like the Laguerre and Chebyshev and Legender polynomials.
5 Conclusion
The construction of eigenvectors gives new class of eigenfunctions in terms of modi-

fied form of well-known special functions. This is an interesting example of the new
way of generating the eigenvectors of DFT, which is now known in the literature.
However, the EMHP are not orthogonal, they could be made orthogonal with suitable
weight function. The further progress in this direction may be possible. The periodic
extensions of these special functions should be of some interest in special function
theory.
References
1. Grunbaum, F.A.: The eigenvectors of the discrete Fourier transform: a version of the Hermite
functions. J. Math. Anal. Appl. 88, 355–363 (1982)
2. Mehta, M.L.: Eigen values and eigenvectors of finite Fourier transform. J. Math. Phys. 28(4),
781–785 (1987)
3. Ruzzi, M.: Jacobi theta functions and discrete Fourier transform. J. Math. Phys. 47, 063507
(2006)
4. Malekar, R.A., Bhate, H.: Discrete Fourier transform and Jacobi theta function identities. J.
Math. Phys. 51, 023511 (2010)
5. Malekar, R.A., Bhate, H.: Discrete Fourier transform and Riemann identities for θ functions.
Appl. Math. Lett. 25, 1415–1419 (2012)
6. Hanna, M.T., et al.: Hermite Gausssian Like eigenvectors of the DFT Matrix based on the direct
utilization of orthogonal projection matrices on its eigenspaces. IEEE Trans. Signal Process.
54, (2006)
7. Matveev, V.B.: Intertwining relations between Fourier transform and discrete Fourier transform,
the related functional identities and beyond. Inverse Prob. 17, 633–657 (2001)
8. Auslander, L., Tolimieri, R.: Is computing with the finite Fourier transform pure or applied
mathematics ? Bulletin American Math Soc I , 847–97 (1979)
9. Coddington, E.A., Levinson, N.: Theory of Ordinary Differential Equations. Tata McGraw-Hill
Publishing Company Ltd, New Delhi (1992)
10. Charalambides, C.A.: Combinatorial probability interpretation of certain modified orthogonal
polynomials. Eur. J. Comb. 29, 1704–1716 (2008)
11. Adzic, N.: Modified Hermite polynomials in the spectral approximation for boundary layer
problems. Bull. Austral. Math. Soc. 65L10 45, 267–276 (1992)
12. Folland, G.B.: Fourier Analysis and Its Applications. Wordsworth and Brooks/Cole Mathe-
matics Series (1992)
A Numerical Simulation Based on Modified
Keller Box Scheme for Fluid Flow:
The Unsteady Viscous Burgers’ Equation
B. Mayur Prakash, Ashish Awasthi and S. Jayaraj
Abstract In this paper the numerical solution of unsteady viscous Burgers’ equation
is presented. A combination of Modified Keller Box difference scheme and Hopf-
Cole transformation is proposed to solve the Burgers’ equation. The proposed scheme
is an implicit scheme with second-order accuracy in space and time. Two test prob-
lems are considered to validate the proposed algorithm. Numerical results which are
calculated for various values of kinematic viscosity and time steps are matching with
the exact solution. It is also observed that, the proposed method yields satisfactory
results for all the cases considered.
Keywords Burgers’ equation · Hopf-Cole transformation · Finite difference

method · Box method · Kinematic viscosity
1 Introduction
The Navier–Stokes equation is considered to be a cornerstone in fluid mechanics

which, when expressed in its originality is a set of unsteady, nonlinear, second-
order partial differential equations. Burgers’ equation whose exact solution is well
known, can be considered as a simplified form of the one-dimensional Navier–Stokes
equation. Burgers’ model is suitable for analysis in gas dynamics, shock wave theory,
B.M. Prakash (B) · S. Jayraj

Department of Mechanical Engineering, National Institute of Technology Calicut,
Kozhikode, India
e-mail: bonkile.mayur@gmail.com
S. Jayraj
e-mail: sjayaraj@nitc.ac.in
A. Awasthi
Department of Mathematics, National Institute of Technology Calicut,
Kozhikode, India
e-mail: aawasthi@nitc.ac.in

DOI 10.1007/978-81-322-2485-3_47
566 B.M. Prakash et al.
cosmology, and traffic flow. The application of this model in various similar important
fields, always require the solution of basic Burgers’ equation.
In 1915, Harry Bateman (1882–1946) [1], an English mathematician, introduced
the following Eq. (1) in his paper along with its initial (2) and boundary conditions (3):
∂u ∂u ∂ 2u
+u = v 2 , 0 < x < L, 0 < t < τ (1)
∂t ∂x ∂x
u(x, 0) = f (x), 0 < x < L (2)
u(0, t) = g1 (t), u(L , t) = g2 (t), 0 < t < τ (3)
where u, x, t and v are the velocity, spatial coordinate, time, and kinematic viscosity,
respectively. The f , g1 , and g2 are prescribed functions of variables depending upon
the specific conditions for the problem to be solved. Later in 1948 Johannes Martinus
Burgers (1895–1981) [4], a Dutch physicist, explained the mathematical modeling
of turbulence with the help of Eq. (1). In order to honor the contributions of Burgers,
this equation is well known as the “Burgers’ equation.” The simultaneous presence of
nonlinear convective term (u ∂∂ux ) and diffusive term (v ∂∂ xu2 ) adds an additional feature
2
to the Burgers’ equation. When v approaches zero, Eq. (1) becomes inviscid Burgers’
equation which is a model for nonlinear wave propagation. When u approaches zero,
Eq. (1) becomes the heat equation.
Julian David Cole (1925–1999) [6] and E. Hopf (1902–1983) [8] independently
introduced a transformation which converts Burgers’ equation into linear heat equa-
tion and is solved exactly for an arbitrary initial condition. Hence, the transformation
is famously known as Hopf-Cole transformation (4).
θx
u(x, t) = −2v , (4)
θ
where, θ satisfies the following heat equation:
∂θ ∂ 2θ
=v 2 (5)
∂t ∂x
Benton and Platman [2] were given 35 distinct analytical solutions of Burgers’ equa-
tion with different initial conditions. Rodin [12] studied some approximate and exact
solution of boundary value problem for Burgers’ equation with the help of Hopf-Cole
transformation. Kutluay et al. [11] used Hopf-cole transformation to convert Burgers’
equation to heat equation. The transformed heat equation with the insulated boundary
conditions was solved by explicit and exact-explicit finite difference method. Burns
et al. [5] considered Burgers’ equation with zero-Neumann boundary conditions to
show that for moderate value of viscosity, numerical solution approaches noncon-
stant shock-type stationary solution. Based on Hopf-Cole linearization, Brander and
Hedenfalk [3] solved Burgers’ equation in one-space dimension for an arbitrary inci-
A Numerical Simulation Based on Modified Keller Box Scheme for Fluid Flow … 567
dent pulse of finite length. Restrictive Pade approximation classical implicit finite
difference method was implemented by Gulsu [7], whose accuracy was demostrated
by the two-test problem. In this paper, we consider the Modified Keller Box method
[13]. Equation (1) is converted to linear heat equation (5) by the Hopf-Cole trans-
formation (4), as explained by Kadalbajoo and Awasthi [9]. The present method
has accuracy of second order in space and time. The accuracy and reliability of the
present method is verified by performing several numerical experiments.
2 Difference Scheme
The solution domain of Eq. (5) is discretized with uniform mesh. The space interval
[0,1] is divided into N equal subinterval. The time interval [0, τ ] is divided into M
equal subintervals. Assuming Δx = 1/N as the mesh width in space and xi is set as
xi = iΔx for i = 0, 1, ..., N . Assuming Δt = τ/M as the mesh width in time and
t n is set as t n = nΔt for n = 0, 1, ..., M.
2.1 Keller Box Method
In Keller Box Method [10], second and higher derivatives of parabolic partial differ-
ential equation are replaced by first derivatives through the introduction of additional
variables which result in a system of first-order equations. Equation (5) is written as
a system of two first-order equations:
∂θ
=T (6)
∂x
∂θ ∂T
=v (7)
∂t ∂x
We use only central differences about (n + 21 , i − 21 ), making use of four points at

the corners of a “box” (Fig. 1a). The resulting difference equations for Eqs. (6) and
(7) are
θin+1 − θi−1
n+1
= T n+1 (8)
Δx 1
i− 2
⎛ ⎞
θ n+11 − θ n n+ 21
Ti
n+ 1
− Ti−1 2
i− 2 i− 12
= v⎝ ⎠ (9)
Δt Δx
(a) (b) (c)
Fig. 1 a Grid for box scheme; b difference molecule for evaluation of T n+1
1 ; c difference molecule
i− 2
for Eq. (9)
1
The discretized terms containing subscript or superscript 2 in Eqs. (8) and (9) are
defined as averages, for example,
θin+1 + θi−1
n+1
θ n+11 = (10)
i− 2 2
n+ 21 Tin + Tin+1
Ti = (11)
2
Averaged expressions (10) and (11) are substituted into Eqs. (8) and (9). The resulting
difference equations become
θin+1 − θi−1
n+1
Tin+1 + Ti−1
n+1
= (12)
Δx 2

θin+1 + θi−1
n+1
Tin − Ti−1
n θin + θi−1
n Tin+1 − Ti−1
n+1
=v + +v (13)
Δt Δx Δt Δx
2.1.1 Modified Box Method
In Eqs. (12) and (13) T’s can be expressed in terms of θ s. Substituting Eq. (12) into
n+1
Eq. (13), Ti−1 is eliminated. Equation (12) is evaluated at time level n to eliminate
n
Ti−1 . Accordingly,
n+1
θin+1 + θi−1
n+1
Tin + Tin+1 θ n + θi−1
n θi n+1
− θi−1 Tin+1 θin − θi−1
n
Tin
=v + i −v − +2 −
Δt Δx Δt (Δx)2 Δx (Δx)2 Δx
(14)
Equations (12) and (13) are rewritten with the i index advanced by 1.
n+1
θi+1 − θin+1 n+1
Ti+1 + Tin+1
= (15)
Δx 2
n+1
θi+1 + θin+1 n − Tn
Ti+1 i θi+1
n + θn
i
n+1
Ti+1 − Tin+1
=v + +v (16)
Δt Δx Δt Δx
n+1 n , Eq. (15) is simply substituted into Eq. (16). The result is
To eliminate Ti+1 and Ti+1
n+1 n+1
θi+1 + θin+1 n − θn)
(θi+1 i Tn θ n + θin (θi+1 − θin+1 ) Tin+1
= 2v − i + i+1 + 2v −
Δt (Δx)2 Δx Δt (Δx)2 Δx
(17)
Adding Eqs. (14) and (17) and after rearranging, the final expression is

2vΔt n+1 4vΔt n+1 2vΔt
1− θ + 2 + θ + 1− θ n+1
(Δx)2 i−1 (Δx)2 i (Δx)2 i+1

2vΔt 4vΔt 2vΔt
= 1+ θn + 2 − θn + 1 + θn (18)
(Δx)2 i−1 (Δx)2 i (Δx)2 i+1
2vΔt
Introducing l = in this equation the result is written in the tridiagonal for-
(Δx)2
mat as
n+1
ai θi−1 + bi θin+1 + ci θi+1
n+1
= di (19)
where,
2vΔt
ai = 1 − = (1 − l) (20)
(Δx)2
4vΔt
bi = 2 + = 2(1 + l) (21)
(Δx)2
2vΔt
ci = 1 − = (1 − l) (22)
(Δx)2

2vΔt 4vΔt 2vΔt
di = 1 + θ n
+ 2 − θ n
+ 1 + θn
(Δx)2 i−1 (Δx)2 i (Δx)2 i+1
= (1 + l)(θi−1
n
+ θi+1
n
) + 2(1 − l)θin (23)
After assembling the entire system of equations and applying boundary conditions,
the general form is Fθ = d, where F and θ are the matrices of order N×N and N
×1 respectively, given by
⎛ ⎞ ⎛ n+1 ⎞
2(1 + l) 2(1 − l) θ1
⎜ (1 − l) 2(1 + l) (1 − l) ⎟ ⎜ θ n+1 ⎟
⎜ ⎟ ⎜ 2 ⎟
⎜ . . . ⎟ ⎜ . ⎟
⎜ ⎟ ⎜ ⎟
F =⎜
⎜ . . . ⎟ θ =⎜ . ⎟
⎟ ⎜ ⎟
⎜ . . . ⎟ ⎜ . ⎟
⎜ ⎟ ⎜ n+1 ⎟
⎝ (1 − l) 2(1 + l) (1 − l) ⎠ ⎝θ ⎠
N −1
2(1 − l) 2(1 + l) θNn+1
The approximate solution of Burgers’ equation (1) in terms of the approximate solu-
tion of heat equation using the Hopf-Cole transformation (4) is given by

n − θn
θi+1 i−1
u in = −v (24)
Δxθin
The proposed scheme is unconditionally stable based on [14].
3 Numerical Experiments and Discussion
This section contain the results of two examples to validate the theoretical results
obtained. Hopf-Cole transformation is used to find the exact solution of these exam-
ples. The fluid properties considered include lubricating oil at 40 ◦ C (v = 1 cm2 /s),
saturated fatty acid methyl esters at 40 ◦ C (v = 0.1 cm2 /s), water at 20 ◦ C (v =
0.01 cm2 /s), water at 50 ◦ C (v = 0.005 cm2 /s) for numerical experiments.
Example 1 Burgers’ equation (1) with initial condition and homogeneous boundary
conditions
u(x, 0) = sin(π x), 0 < x < 1,
u(0, t) = u(1, t) = 0, 0 ≤ t ≤ τ
By using Hopf-Cole transformation (4) Eq. (1) is transformed to the linear heat
equation (5) with initial condition (25) and boundary conditions (no heat transfer) (26)

1
θ (x, 0) = ex p − [1 − cos(π x)] , 0 < x < 1 (25)
2π v
θx (0, t) = θx (1, t) = 0, 0 ≤ t ≤ τ. (26)
Example 2 Burgers’ equation (1) with the following initial condition and boundary
conditions:
u(x, 0) = 4x(1 − x), 0 < x < 1,

u(0, t) = 0 = u(1, t), 0 ≤ t ≤ τ.
Using Hopf-Cole transformation (4) Eq. (1) is transformed to linear heat equation
(5) with initial condition and with boundary conditions (no heat transfer)

1 4x 3
θ (x, 0) = ex p − 2x −
2
, 0<x <1 (27)
2v 3
θx (0, t) = θx (1, t) = 0, 0 ≤ t ≤ τ. (28)
Exact solution was elaborated in [9] for both Examples (1) and (2).
Computed results are displayed in Tables 1, 2, 3 and 4 at different nodal points for
diverse values of kinematic viscosity. It is found that the computed results show better
agreement with the exact solution as the mesh size is refined. These results show that
Table 1 Comparison of the numerical solution with exact solution at different space points of
Example 1 at T = 0.1 for v = 1 and Δt = 0.0001
x N = 10 N = 20 N = 40 N = 80 Exact solution
0.1 0.10596 0.10871 0.10940 0.10958 0.10953
0.2 0.20298 0.20822 0.20954 0.20988 0.20979
0.3 0.28249 0.28973 0.29156 0.29202 0.29189
0.4 0.33683 0.34538 0.34754 0.34808 0.34792
0.5 0.35987 0.36891 0.37118 0.37175 0.37157
0.6 0.34787 0.35651 0.35868 0.35922 0.35904
0.7 0.30037 0.30775 0.30960 0.31006 0.30990
0.8 0.22087 0.22625 0.22760 0.22793 0.22781
0.9 0.11702 0.11986 0.12057 0.12075 0.12068
Table 2 Comparison of the numerical solution with the exact solution at different space points of
0.1 0.11099 0.11378 0.11449 0.11466 0.11461
0.2 0.21127 0.21659 0.21793 0.21826 0.21816
0.3 0.29113 0.29845 0.30029 0.30075 0.30061
0.4 0.34274 0.35135 0.35352 0.35406 0.35389
0.5 0.36097 0.37002 0.37230 0.37287 0.37269
0.6 0.34386 0.35247 0.35464 0.35518 0.35501
0.7 0.29293 0.30026 0.30210 0.30257 0.30242
0.8 0.21307 0.21840 0.21974 0.22008 0.21997
0.9 0.11210 0.11490 0.11561 0.11578 0.11573
0.1 0.10919 0.11203 0.11275 0.11293 0.11289
0.2 0.20921 0.21462 0.21600 0.21634 0.21625
0.3 0.29126 0.29873 0.30062 0.30109 0.30097
0.4 0.34742 0.35624 0.35846 0.35902 0.35886
0.5 0.37135 0.38067 0.38301 0.38360 0.38342
0.6 0.35914 0.36804 0.37028 0.37084 0.37066
0.7 0.31023 0.31784 0.31975 0.32022 0.32007
0.8 0.22821 0.23375 0.23514 0.23549 0.23537
0.9 0.12094 0.12387 0.12460 0.12478 0.12472
0.1 0.11453 0.11742 0.11814 0.11833 0.11827
0.2 0.21802 0.22351 0.22489 0.22524 0.22513
0.3 0.30044 0.30799 0.30989 0.31037 0.31022
0.4 0.35372 0.36259 0.36483 0.36539 0.36521
0.5 0.37254 0.38187 0.38422 0.38481 0.38463
0.6 0.35491 0.36378 0.36602 0.36658 0.36640
0.7 0.30235 0.30991 0.31181 0.31229 0.31214
0.8 0.21993 0.22543 0.22681 0.22716 0.22705
0.9 0.11571 0.11860 0.11933 0.11951 0.11946
Fig. 2 Numerical solution of Example 1 at several times for N = 80 with distinguishable liquids
having different values of v and Δt, a saturated fatty acid methyl esters at 40 ◦ C with v = 0.1, Δt =
0.0001; b lubricating oil at 40 ◦ C with v = 1, Δt = 0.0001; c water at 20 ◦ C with v = 0.01, Δt =
0.01; d water at 50 ◦ C with v = 0.005, Δt = 0.01
Fig. 3 Numerical solution profiles of Example 1 at Δx = 0.0125 for, a lubricating oil at 40 ◦ C

with v = 1, Δt = 0.0001; b Saturated fatty acid methyl esters at 40 ◦ C with v = 0.1, Δt = 0.001;
c Water at 20 ◦ C with v = 0.01, Δt = 0.001; d Water at 50 ◦ C with v = 0.005, Δt = 0.01
Fig. 4 Numerical solution of Example 2 at several times for N = 80 with distinguishable liquids
having different values of v and Δt, a saturated fatty acid methyl esters at 40 ◦ C with v = 0.1, Δt =
0.0001; b lubricating oil at 40 ◦ C with v = 1, Δt = 0.0001; c water at 20 ◦ C with v = 0.01, Δt =
0.01; d water at 50 ◦ C with v = 0.005, Δt = 0.01
the scheme is consistent and accurate of order two in space and time. Figures 2 and 4
show the graphs for computed and exact solution at different times for various values
of kinematic viscosity. From these graphs it is observed that the proposed method
gives accurate results for any value of time step Δt. In order to show the physical
behavior of the given problem, surface plots of the computed solutions are presented
(Figs. 3 and 5) for different liquids with distinct values of kinematic viscosity. When
the kinematic viscosity is very large and the fluid velocity is very slow, Reynold
number (Re) becomes very much less compared to unity (Re << 1). This type of
flow is known as creeping flow which occurs in the case of some paints, MEMS,
viscous polymers, lubrication, etc. For the creeping flow this method is proved to
give very accurate results. When Re >> 1, the inertia force is more dominating than
Fig. 5 Numerical solution profiles of Example 2 at Δx = 0.0125 for, a Lubricating oil at 40 ◦ C

with v = 1, Δt = 0.0001; b Saturated fatty acid methyl esters at 40 ◦ C with v = 0.1, Δt = 0.001;
c Water at 20 ◦ C with v = 0.01, Δt = 0.01; d Water at 50 ◦ C with v = 0.005, Δt = 0.1
2
viscous force and diffusion term v ∂∂ xu2 tends to zero. Even if Re >> 1, it is possible
that the inertia and viscous forces are of comparable magnitude, especially in the
neighborhood of a solid boundary. This thin region is called the boundary layer and
the proposed method works quite satisfactorily in this region as well.
4 Conclusions
The Modified Keller-Box difference scheme, which is a highly accurate and efficient
method coupled with Hopf-Cole transformation, is used to study the properties of the
solution of Burgers’ equation for a wide range of kinematic viscosity. The present
numerical experiments have confirmed that the present method is unconditionally
stable. It is second order accurate in space and time. There is no requirement with
respect to mesh size restriction. The results are in good agreement with the exact
solution for modest values of the kinematic viscosity. The physical behavior of the
solution is explained and it is concluded that the numerically calculated values are
in close agreement with the exact solution.
Acknowledgments The authors thank the anonymous referees for their valuable time, effort, and
extensive comments which helped to improve the quality of the paper.
References
1. Bateman, H.: Some recent researches on the motion of fluids. Mon. Weather Rev. 43, 163–170
(1915)
2. Benton, E.R., Platzman, G.W.: A table of solutions of the one-dimensional Burgers’ equations.
Quart. Appl. Math. 30, 195–212 (1972)
3. Brander, O., Hedenfalk, J.: A new formulation of the general solution to burgers equation.
Wave Motion 28, 319–332 (1998)
4. Burgers, J.: A Mathematical Model Illustrating the Theory of Turbulence. Advances in Applied
Mechanics, pp. 171–199. Academic Press, New York (1948)
5. Burns, J., Balogh, A., Gilliam, D., Shubol, V.: Numerical stationary solutions for a viscous
Burgers’ equation. J. Math. Syst. Estim. Control 8, 1–16 (1998)
6. Cole, J.: On a quaslinear parabolic equations occurring in aerodynamics. Quart. Appl. Math.
9, 225–236 (1951)
7. Gulsu, M.: A finite difference approach for solution of Burgers’ equation. Appl. Math. Comput.
175, 1245–1255 (2006)
8. Hopf, E.: The partial differential equation u t + uu x = vu x x . Comm. Pure Appl. Math. 3,
201–230 (1950)
9. Kadalbajoo, M., Awasthi, A.: A numerical method based on crank-nicolson scheme for Burgers’
equation. Appl. Math. Comput. 182, 1430–1442 (2006)
10. Keller, H.B.: A new difference scheme for parabolic problems. In: Numerical Solutions of
Partial Difference Equations, vol. 2. Academic Press, New York (1970)
11. Kutluay, S., Bahadir, A., Ozdes, A.: Numerical solution of one-dimensional burgers equation:
explicit and exact-explicit finite difference methods. J. Comput. Appl. Math. 103, 251–261
(1999)
12. Rodin, E.Y.: On some approximate and exact solutions of boundary value problems for Burgers
equation. J. Math. Anal. Appl. 30, 401–414 (1970)
13. Tannehill, J.C., Anderson, D.A., Pletcher, R.H.: Computational Fluid Mechanics and Heat
Transfer. Taylor and Francis, Philadelphia (1997)
14. Vajiravelu, K., Prasad, K.V.: Keller-Box Method and Its Application. Walter de Gruyter and
Co., Orlando (2014)
A Novel Approach to Surface Interpolation:
Marriage of Coons Technique and Univariate
Fractal Functions
A.K.B. Chand, P. Viswanathan and K.M. Reddy
Abstract The current article is intended to demonstrate that the theory of fractal
functions when applied in conjunction with methods in the classical numerical analy-
sis can supply new solution techniques that supplement and subsume the existing
ones. To this end, in the first part of the paper, we review a C 1 -continuous ratio-
nal cubic fractal interpolation function (FIF) introduced recently [Viswanathan and
Chand, Elec. Trans. Numer. Anal. 41 (2014), pp. 420–442]. We carry out the con-
vergence analysis of this univariate rational FIF and determine suitable values of the
derivative parameters so that its global smoothness enhances to C 2 . In the subsequent
part of the article, we apply Coons technique of transfinite interpolation in order to
construct a new kind of C 1 -continuous bivariate fractal interpolation surface.
Keywords Fractal interpolation function · Convergence · Coons technique · Blend-

ing function · Fractal interpolation surface
1 Prologue
Fractal function proposed by Barnsley [1] is a powerful tool for the approximation of
natural data sets and some complicated experimental variables. Fractal interpolation
function (FIF) provides a totally new interpolation method which has proved to be
advantageous over the traditional nonrecursive interpolation techniques. Traditional
interpolation techniques such as polynomial, rational, trigonometric, and spline inter-
polation always render interpolating curves that are infinitely smooth except possibly
at a finite number of points corresponding to the given interpolation data. FIF pos-
sesses the novelty of providing one of the very few methods of non-differentiable
A.K.B. Chand (B) · P. Viswanathan · K.M. Reddy

Department of Mathematics, Indian Institute of Technology Madras, Chennai, India
P. Viswanathan
e-mail: amritaviswa@gmail.com
K.M. Reddy
e-mail: mahipalnitw@gmail.com

DOI 10.1007/978-81-322-2485-3_48
578 A.K.B. Chand et al.
interpolants. Nonsmooth version of FIF remains well-suited to approximate images

of natural objects such as profiles of mountain ranges, tops of clouds, and stalactite
hung roofs of caves, wherein regularity is almost missing [2]. Graph of the fractal
interpolant possesses a non-integer dimension which can be used as a quantifier for
the complexity of the underlying phenomenon.
Barnsley and Harrington [3] initiated the construction of smooth FIFs, and
unfolded the striking relationship between fractal functions and splines. Smooth FIFs
can be applied to generalize the classical interpolation techniques (see, for instance,
[5–7, 10, 13]). Therefore, fractal interpolation offers the flexibility of choosing either
a smooth or a nonsmooth approximant depending on the modeling problem at hand.
Apart from modeling a prescribed set of data with an interpolant having a suitable
degree of smoothness, preserving fundamental shape properties such as positivity,
monotonicity, and convexity inherent in the data is extremely important in many
practical applications. There is a myriad of research articles in the field of clas-
sical approximation theory and numerical analysis targeting the problem of shape
preserving interpolation, often directed at specific application domains. For a com-
pendium of some of these, the reader may refer to the survey article [12]. However,
the suitability of FIFs as a bona fide technique for shape preservation has not been
fully exploited hitherto. Recently, the authors have initiated shape preserving fractal
interpolation and approximation through their researches that appeared in [8–10,
15–17].
In this article, first we shall revisit (see Sect. 3) the C 1 -continuous rational cubic
FIF with linear denominator introduced in [16]. Note that here we shall adopt a
constructional approach and also that the FIF provides a fractal generalization of
the rational cubic spline introduced in [14]. The convergence analysis of this cubic
rational FIF is carried out in detail. Suitable values of derivatives are determined
treating them as parameters so that global smoothness of the rational FIF enhances
to C 2 . These findings when applied with suitable choice of shape parameters involved
in the structure of the FIF provides an alternative to the standard moment construction
of C 2 -cubic FIF discussed elaborately in [5].
In contrast to univariate fractal interpolants, the volume of literature available on
bivariate fractal interpolants is less extensive, though its scope is wide. To build on
the literature and to demonstrate that the cooperation of fractal theory with the tools
in numerical analysis can produce new solution techniques, we apply the Coons’ idea
of transfinite interpolation [4] on these rational cubic FIFs. This culminates with a
bivariate interpolant (surface) whose boundaries are cubic rational FIFs. Unlike the
traditional interpolating surfaces, these fractal surfaces have partial derivatives that
are smooth or nonsmooth depending on the choice of parameters, a fact that can find
potential applications in various nonlinear and nonequilibrium phenomena.
2 Basics of Fractal Interpolation
This section targets to equip a novice reader with the preliminaries of fractal interpo-
lation required for the current study. These materials are collected from well-known
treatises [1–3].
A Novel Approach to Surface Interpolation … 579
We use the following notation throughout the article. For m ∈ N, we denote by

Nm , the subset of N containing first m natural numbers.
For n > 2, let x1 < x2 < · · · < xn be real numbers and I = [x1 , xn ] be the
closed interval that contains them. Let the prescribed set of interpolation data be
{(xi , yi ) ∈ I × R : i ∈ Nn }. For i ∈ Nn−1 , set Ii = [xi , xi+1 ] and let L i : I → Ii
be contraction homeomorphisms that obey the endpoint conditions
L i (x1 ) = xi , L i (xn ) = xi+1 . (1)
Let Fi : I × R → R be continuous maps satisfying

Fi (x1 , y1 ) = yi , Fi (xn , yn ) = yi+1 , i ∈ Nn−1 ,
|Fi (x, y) − Fi (x, y )| ≤ ri |y − y |, x ∈ I ; y, y ∈ R; for some 0 ≤ ri < 1.
(2)

For i ∈ Nn−1 , define functions ωi (x, y) = L i (x), Fi (x, y) . Consider the collection
{I ×R; ωi : i ∈ Nn−1 } referred to as an iterated function system (IFS). The following
is the most fundamental theorem in the field of fractal interpolation.
Theorem 1 ([1]) The IFS {I × R; ωi : i ∈ Nn−1 } defined above admits a unique
attractor G, and G is the graph of a continuous function f : I → R which obeys
f (xi ) = yi for i ∈ Nn .
The function which has made its debut in the above theorem is termed a fractal
interpolation function (FIF), and it satisfies the functional equation:

f (x) = Fi L i−1 (x), f ◦ L i−1 (x) , x ∈ Ii , i ∈ Nn−1 . (3)
The most extensively studied FIFs in theory and applications hitherto are defined
through the system of maps
L i (x) = ai x + bi , Fi (x, y) = αi y + qi (x), (4)
where αi are constants satisfying 0 ≤ |αi | < 1 and qi are continuous functions so
that the “join-up conditions” in (2) imposed on the bivariate functions Fi are satis-
fied. The multiplier αi is called a vertical scaling factor for the transformation ωi and
the vector α = (α1 , α2 , . . . , αn−1 ) ∈ (−1, 1)n−1 is called a scale vector of the IFS.
The prescriptions in (1) uniquely determine the constants ai and bi appearing in the
affine map L i .
Let us recall that the function f determined by the IFS in (4), which takes the
form
f (x) = αi f L i−1 (x) + qi L i−1 (x) , (5)
is, in general, nonsmooth. The following theorem determines the conditions on αi

and functions qi so that the FIF is C p -continuous.
Theorem 2 ([3]) Let {(xi , yi ) : i ∈ Nn } be a given data set with strictly increasing
abscissae. Let L i (x) = ai x + bi satisfy (1) and Fi (x, y) = αi y + qi (x) obey (2)
p
for i ∈ Nn−1 . Suppose that for some integer p ≥ 0, |αi | < ai and qi ∈ C p (I ),
i ∈ Nn−1 . Let
(k) (k) (k)

αi y + qi (x) q1 (x1 ) qn−1 (xn )
Fi,k (x, y) = , y1,k = , yn,k = k , k = 1, 2, . . . , p.
aik a1k − α1 an−1 − αn−1
(xn , yn,k ) = Fi,k (x1 ,y1,k ) for i = 2, 3, . . . , n − 1 and k ∈p N p , then the

If Fi−1,k
(k)
IFS I × R; L i (x), Fi (x, y) : i ∈ Nn−1 determines a FIF f ∈ C (I ), and f
is the FIF determined by I × R; L i (x), Fi,k (x, y) : i ∈ Nn−1 for k ∈ N p .
3 Rational Cubic FIF with Linear Denominator Revisited
For n ∈ N with n > 2, let {(xi , yi ) : i ∈ Nn } be a given set of interpolation data

with strictly increasing abscissae. Let di be the derivative value (given or estimated
by some standard methods) at the knot point xi .
3.1 Construction of C 1 -Rational Cubic FIF
Consider the IFS in (4) where qi ∈ C 1 (I ) is chosen such that
Pi (θ ) Ai (1 − θ )3 + Bi θ (1 − θ )2 + Ci θ 2 (1 − θ ) + Di θ 3
qi (x) ≡ qi∗ (θ ) = = ,
Q i (θ ) (1 − θ )ri + θ ti
(6)
where θ := xx−x 1
n −x 1
and ri > 0, ti > 0 are free shape parameters. The “join-up
conditions” on the bivariate function Fi are equivalent to qi (x1 ) = yi − αi y1 and
qi (x N ) = yi+1 −αi y N . This observation when coupled with (5) yields the conditions
f (xi ) = yi and f (xi+1 ) = yi+1 , which in turn determines the following coefficients:
Ai = ri (yi − αi y1 ), Di = ti (yi+1 − αi yn ). (7)
In accordance with the principle of construction of a C 1 -spline FIF (see Theorem 2),
we impose the condition |αi | < ai for i ∈ Nn−1 on the scaling factors. Adhering to
the notation in Theorem 2, let us denote y1,1 = d1 , yn,1 = dn , Fi,1 (x1 , d1 ) = di , and
Fi,1 (xn , dn ) = di+1 for i ∈ Nn−1 . Then, f is C 1 -continuous,
the derivative
f is the
fractal function corresponding to the IFS {I × R; L i (x), Fi,1 (x, y) : i ∈ Nn−1 },
and f interpolates the data set {(xi , di ) : i ∈ Nn }. Further, f satisfies the functional
equation:
αi f (x) + qi (x)

f L i (x) = Fi,1 x, f (x) = . (8)
ai
Imposing the conditions noted previously, namely Fi,1 (x1 , d1 ) = di , and Fi,1 (xn ,
dn ) = di+1 on (8), routine algebraic manipulations provide the following coeffi-
cients:
Bi = (2ri + ti )yi + ri h i di − αi [(2ri + ti )y1 + ri (xn − x1 )d1 ],

(9)
Ci = (ri + 2ti )yi+1 − ti h i di+1 − αi [(ri + 2ti )yn − ti (xn − x1 )dn ],
where h i = xi+1 − xi is the length of the subinterval Ii . In view of (6), (7) and (9),
the desired FIF with rational qi takes the form
Pi (x)
f L i (x) = αi f (x) + , (10)
Q i (x)

Pi (x) ≡ Pi∗ (θ) = ri {yi − αi y1 }(1 − θ)3 + ti {yi+1 − αi yn }θ 3 + (2ri + ti )yi + ri h i di

−αi [(2ri + ti )y1 + ri (xn − x1 )d1 ] θ(1 − θ)2 + (ri + 2ti )yi+1

−ti h i di+1 − αi [(ri + 2ti )yn − ti (xn − x1 )dn ] θ 2 (1 − θ),
x − x1
Q i (x) ≡ Q i∗ (θ) = (1 − θ)ri + θ ti , i ∈ Nn−1 , θ =
xn − x1
The aforementioned rational FIF depends on the choice of scaling vector and may
be denoted by f α for clarity. The reader is invited to refer [16] for construction of
the aforementioned rational cubic spline FIF using a different approach via α-fractal
functions.
Remark 3 If the scaling factors αi = 0 for all i ∈ Nn−1 , then the rational cubic FIF
f coincides with the C 1 -rational cubic spline C ∈ C 1 (I ) introduced in [14]. Let us
note that C is defined in a piecewise manner such that for x ∈ Ii , i ∈ Nn−1
−1
C(x) = (1 − φ)ri + φti ri yi (1 − φ)3 + [(2ri + ti )yi + ri h i di ]φ(1 − φ)2

+ [(ri + 2ti )yi+1 − ti h i di+1 ]φ 2 (1 − φ) + ti yi+1 φ 3 ,

(11)
L −1 (x)−x
where φ = i xn −x1 1 = x−x i
h i . If the shape parameters pertaining to each subinterval
is taken according to ri = ti , then the rational cubic spline FIF reduces to the C 1 -
cubic Hermite spline FIF studied in [10], which again with the zero scaling on each
Ii reduces to the traditional nonrecursive C 1 -cubic Hermite interpolant.
3.2 Rational Cubic FIF with C 2 -Continuity
Assuming the derivative values di , i ∈ Nn to be free parameters, we derive suitable

conditions on di so that the rational cubic spline f ∈ C 2 (I ). In view of principle
of construction of a C 2 -FIF (cf. Theorem 2), we take the scaling factors such that
|αi | ≤ κai2 ∀ i ∈ Nn−1 , where 0 ≤ κ < 1.
The functional equation of f (2) with x = x1 and x = xn applied on the affine
homeomorphisms L i and L i−1 produce
αi (2) 1 3 αi
f (2) (xi+ ) = f (x1 ) + 3 (2ri + 4ri2 ti )[Δi − (yn − y1 )]
ai2 ri h i hi

(12)
αi αi
− (2ri3 + 2ri2 ti )[di − d1 (xn − x1 )] − 2ri2 ti [di+1 − dn (xn − x1 )] ,
hi hi
αi−1 (2) 1 3 αi−1

f (2) (xi− ) = 2
f (xn ) + 3 (2ti−1 + 2ri−1 ti−1
2
)[di − dn (xn − x1 )]
ai−1 ti−1 h i−1 h i−1

(13)
αi−1 αi−1
+ 2ri−1 ti−1
2
[di−1 − d1 (xn − x1 )] − (2ti−1
3
+ 4ri−1 ti−1
2
)[Δi−1 − (yn − y1 )] ,
h i−1 h i−1
where Δi := yi+1h i−yi denotes the secant slope. Continuity of f (2) at the interior
points xi , i = 2, 3, . . . , n − 1 demands f (2) (xi+ ) = f (2) (xi− ), which in view of
(12)–(13) yields
− αi ai−1
2
ri ti−1 h i h i−1 f (2) (x1 ) + 2ri−1 ai2 ai−1
2
ri h i di−1 + 2ai2 ai−1 2
ti−1 h i−1 ti di+1
αi α i−1
− [2ai2 ai−1
2
ti−1 h i−1 (xn − x1 )(ri + ti ) + 2ai2 ai−1 2
ri ri−1 h i (xn − x1 )]d1
hi h i−1
+ [2ai2 ai−1
2
ti−1 h i−1 (ri + ti ) + 2ai2 ai−1
2
ri h i (ri−1 + ti−1 )]di
α i αi−1
− [2ai2 ai−1
2
ti−1 ti h i−1 (xn − x1 ) + 2ai2 ai−1 2
ri h i (xn − x1 )(ti−1 + ri−1 )]dn
hi h i−1
+ αi−1 ai2 ri ti−1 h i h i−1 f (2) (xn )
αi
= 2ai2 ai−1
2
ti−1 h i−1 [Δi − (yn − y1 )](ri + 2ti ) + 2ai2 ai−1 2
ri h i ×
hi
αi−1
[Δi−1 − (yn − y1 )](2ri−1 + ti−1 ), i = 2, 3, . . . , n − 1.
h i−1
(14)
Next, from the functional equation for f (2) we get
α1
r1 h 1 (a12 − α1 ) f (2) (x1 ) + 2a12 [1 −
(xn − x1 )](r1 + t1 )d1 + 2t1 a12 d2
h1
α1 α1 (15)
− 2t1 a12 (xn − x1 ) dn = 2[Δ1 − (yn − y1 )]a12 (r1 + 2t1 ).
h1 h1
Similarly,
αn−1
− tn−1 h n−1 (an−1
2
− αn−1 ) f (2) (xn ) − 2rn−1 an−1
2
(xn − x1 )d1
h n−1
αn−1
+ 2rn−1 an−1
2
dn−1 + 2an−1
2
[1 − (xn − x1 )](rn−1 + tn−1 )dn (16)
h n−1
αn−1
= 2[Δn−1 − (yn − y1 )]an−1
2
(tn−1 + 2rn−1 ).
h n−1
Now, for i = 2, 3, . . . , n − 1, let us introduce the following notation:
ri−1 ri h i αi ri μi h i
λi = , μi = 1 − λi , Ai = − ,
ti−1 ti h i−1 + ri−1 ri h i 2a 2 ti
i
αi−1 ri h i μi ∗ αi μi (xn − x1 )(ri + ti ) αi−1 λi (xn − x1 )
Bi = 2 t
, Ai = − − ,
2ai−1 i h i ti h i−1
μi (ri + ti ) λi (ri−1 + ti−1 )
Ci = + ,
ti ri−1
αi μi (xn − x1 ) αi−1 λi (ri−1 + ti−1 )(xn − x1 )
Bi∗ = − − ,
hi h i−1 ri−1
αi−1
μi (ri + 2ti )[Δi − αh i (yn − y1 )] λi (2ri−1 + ti−1 )[Δi−1 − h i−1 (yn − y1 )]
βi = i
+ .
ti ri−1
Using the above notation, the continuity condition (14) can be reformulated as
Ai∗ d1 + Ai f (2) (x1 ) + λi di−1 + Ci di + μi di+1 + Bi f (2) (xn ) + Bi∗ dn = βi . (17)
Letting
α1
A∗1 = 2a12 [1 − (xn − x1 )](r1 + t1 ), A1 = r1 h 1 (a12 − α1 ), μ1 = 2a12 t1 , λn = 2rn−1 an−1
2
,
h1

α1 α1 αn−1
B1∗ = −μ1 (xn − x1 ), β1 = 2a12 (r1 + 2t1 ) Δ1 − (yn − y1 ) , A∗n = −λn (xn − x1 )
h1 h1 h n−1

αn−1
Bn = −tn−1 h n−1 (an−1
2
− αn−1 ), Bn∗ = 2an−1 2
(rn−1 + tn−1 ) 1 − (xn − x1 ) ,
h n−1

αn−1
βn = 2an−1
2
(tn−1 + 2rn−1 ) Δn−1 − (yn − y1 ) ,
h n−1
(15)–(16) can be rewritten as

A∗1 d1 + A1 f (2) (x1 ) + μ1 d2 + B1∗ dn = β1 ,
(18)
A∗n d1 + λn dn−1 + Bn f (2) (xn ) + Bn∗ dn = βn .
The matrix form that defines Eqs. (17) and (18) is

⎡ ⎤⎡ ⎤ ⎡ ⎤
A∗1 A1 μ1 · · · 0 0 B1∗ d1 β1
⎢ A ∗ + λ2 A2 C 2 · · · 0 B2∗ ⎥⎢ f (2) (x1 ) ⎥ ⎢ ⎥
⎢ 2 ∗ B2 ⎥⎢ ⎥ ⎢ β2 ⎥
⎢ A A 3 λ3 · · · 0 B3 B3∗ ⎥⎢ d2 ⎥ ⎢ β3 ⎥
⎢ 3 ⎥⎢ ⎥ ⎢ ⎥
⎢ .. .. .. .. .. .. .. ⎥⎢ .. ⎥ ⎢ .. ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥.
⎢ ∗. . . . . . . ⎥⎢ . ⎥ ⎢ . ⎥
⎢ A An−2 0 · · · μn−2 Bn−2 ∗
Bn−2 ⎥⎢ ⎢
dn−1 ⎥ ⎢ βn−2 ⎥
⎥
⎢ n−2 ⎥⎢ ⎥
⎣ A∗ An−1 0 · · · Cn−1 Bn−1 Bn−1 + μn−1 ⎦ ⎣
∗
f (xn ) ⎦ ⎣ βn−1 ⎦
(2)
n−1
A∗n 0 0 · · · λn Bn Bn∗ dn βn
Hence, the linear system governing the C 2 -continuity of the rational cubic FIF con-
sists of a coefficient matrix of order N × (N + 2). By prescribing suitable boundary
conditions similar to that of the classical rational cubic spline [14], the rectangular
system reduces to a square system.
If we take zero scaling factor in each interval, then the continuity equation (17)
reduces to
ti−1 h i−1 (ri + 2ti )Δi + ri h i (2ri−1 + ti−1 )Δi−1
λi di−1 + Ci di + μi di+1 = ,
ti−1 ti h i−1 + ri−1ri h i
which coincides with the linear tridiagonal system corresponding to the C 2 -continuity
of the classical rational cubic spline C (see Eq. (4.1) in [14] or Eq. (2) in [11]). The
feasibility of the above construction depends on the existence and uniqueness of the
solution of the linear system corresponding to a fixed choice of the scaling factors and
the shape parameters, which follows from the existence and uniqueness of the fixed
point of Read-Bajraktarévic operator (see [5]). Using suitable boundary conditions,
and solving the corresponding square system of equations, the values di , i ∈ Nn are
determined.
Remark 4 For the shape parameters satisfying ri = ti ∀ i ∈ Nn−1 , our fractal rational
cubic FIF reduces to the C 1 -cubic Hermite FIF. Consequently, the above elements of
C 2 theory renders, in particular, the construction of a C 2 -cubic spline FIF through
slopes di . For the construction of C 2 -cubic spline FIF through moments, the reader
is invited to refer [5].
The goal in the current section is to establish theorem that can eventually handle the
convergence of the developed rational cubic FIF f under the assumption that the
original function Φ ∈ C 1 (I ). Convergence result for the rational cubic spline FIF
f is derived using the convergence result for its classical nonrecursive counterpart
C (cf. (12)) and an upper bound for the uniform distance between f and C. The
following lemma establishes an upper bound for the uniform distance between the
classical rational cubic spline C = f 0 and the perturbed fractal function f = f α .
Proof follows on lines similar to that of Lemma 4.1 (cf. [5]), and is hence omitted.
Lemma 5 Let f and C, respectively, be the rational cubic spline FIF and the
classical rational cubic spline interpolant for the data {(xi , yi ) : i ∈ Nn }.
i (αi ,x)
Let the rational functions qi (αi , x) = PQ i (x)
involved in the FIF f satisfy
∂qi
(ξi , x) ≤ D0 ∀ (ξi , x) ∈ (−ai , ai ) × Ii and i ∈ 1, 2, . . . , n − 1. Then with
∂αi
|α|∞ := max{|αi | : i ∈ Nn−1 },
|α|∞
f − C∞ ≤ (C∞ + D0 ).
1 − |α|∞
With this preparation, an upper bound for the L ∞ -norm of the error in approximating
the data generating function Φ ∈ C 1 (I, R) with the rational cubic spline FIF f
(cf. (10)) is deduced in the following theorem. For simplicity of presentation, we
introduce the following notation: |y|∞ = max{|yi | : 1 ≤ i ≤ n}, |d|∞ = max{|di | :
1 ≤ i ≤ n}, |I | = xn − x1 , h = max{h i : i ∈ Nn−1 }, γ = max{γi : i ∈ Nn−1 }, and
δ = min{δi : i ∈ Nn−1 }. We shall assume that the shape parameters are selected so
as to prevent the situation: quantities γ and δ approach to zero as the norm of the
partition approaches zero.
Theorem 6 Let Φ ∈ C 1 (I, R) be the function generating the data {(xi , yi ) : i ∈
Nn } and let f be the corresponding rational cubic spline FIF defined through the
functional equation (10). Then
γ |α|∞
Φ − f ∞ ≤ (4ω( f ; h) + h|d|∞ ) + |y|∞ + max{|y1 |, |yn |}
4δ 1 − |α|∞
1
+ h|d|∞ + |I | max{|d1 |, |dn |} .

4
In particular, f converges to Φ uniformly, as the norm of the partition approaches
zero.
Proof For the data generating function Φ, let C and f , respectively, be the classi-
cal rational cubic spline and the rational cubic spline FIF with respect to the data
{(xi , yi ) : i ∈ Nn }. We will use the following triangle inequality for our calculations:
Φ − f ∞ ≤ Φ − C∞ + C − f ∞ . (19)
By Lemma 5,
|α|∞
f − C∞ ≤ C∞ + D0 (20)
1 − |α|∞
With a careful observation of the expressions of functions C and qi , and with rigorous
calculations, the following bounds for C∞ and D0 can be obtained:
h
C∞ ≤ |y|∞ + |d|∞ . (21)
4
On similar lines D0 can be selected as a constant such that
1
D0 ≥ max{|y1 |, |yn |} + |I | max{|d1 |, |dn |}. (22)
4
Substituting the bounds (21)–(22) in (20) we see that
|α|∞ 1
f − C∞ ≤ |y|∞ + max{|y1 |, |yn |} + (h|d|∞ + |I | max{|d1 |, |dn |}) . (23)

1 − |α|∞ 4
Next we shall turn our attention to the first summand appearing in (19), which
corresponds to the error bound for the classical rational cubic spline C (cf. (11)).
1
C(x) − Φ(x) = (1 − θ)3 ri yi − Φ(x) + θ(1 − θ)2 (2ri + ti )(yi − Φ(x)) + ri h i di
Q i (θ)

+ θ 2 (1 − θ) (ri + 2ti )(yi+1 − Φ(x)) − ti h i di+1 + θ 3 ti yi+1 − Φ(x) .
Thus, we infer
γi hi
|C(x) − Φ(x)| ≤ ω(Φ; h i ) + max{|di |, |di+1 |} ,

δi 4
which implies
γ
Φ − C∞ ≤ (4ω(Φ; h) + h|d|∞ ). (24)
4δ
Now substituting (23) and (24) in (19), we obtain
γ |α|∞
Φ − f ∞ ≤ (4ω(Φ; h) + h|d|∞ ) + |y|∞ + max{|y1 |, |yn |}
4δ 1 − |α|∞
1
+ (h|d|∞ + |I | max{|d1 |, |dn |}) .

4
Recalling the fact that the modulus of continuity ω(Φ; h) of a continuous function
Φ on a closed interval I approaches zero as h → 0 and noting that |α|∞ < |Ih | , we
can deduce the uniform convergence of f to Φ. This finishes the proof.
5 Construction of Bicubic Partially Blended Rational FIS
Let Δ = {(xi , y j , z i, j ) : i ∈ N M , j ∈ N N } be a set of bivariate interpolation data,

where x1 < x2 < · · · < xi < xi+1 < · · · < x M and y1 < y2 < · · · < y j < y j+1 <
· · · < y N . Let Ri, j := Ii × J j = [xi , xi+1 ] × [y j , y j+1 ]; i ∈ N M−1 , j ∈ N N −1 ,
be the generic subrectangular region and let us put R := I × J = [x1 , x M ] ×
[y1 , y N ]. Following the one-dimensional scheme, let us denote by h i = xi+1 −
xi , h ∗j = y j+1 − y j , i ∈ N M−1 , j ∈ N N −1 . Let z i,x j and z i, j be the x-partials and
y
y-partials at the point (xi , y j ). In this section, we wish to construct a C 1 -continuous

bivariate function Ψ : R → R such that Ψ (xi , y j ) = z i, j , ∂Ψ ∂ x (x i , y j ) = z i, j , and
x
∂Ψ y
∂ y (x i , y j ) = z i, j for i ∈ Nm and j ∈ Nn . This is achieved by blending the univariate
rational cubic FIFs using the partially bicubic Coons technique [4].
y
For each i ∈ N M , let us consider the univariate data set Δxi := {(xi , y j , z i, j , z i, j ) :
j ∈ N N }, which is the data associated with the i-th grid line parallel to y-axis.
Consider the affine maps L ∗j : [y1 , y N ] → [y j , y j+1 ] defined by L ∗j (y) = c j y + d j
satisfying L ∗j (y1 ) = y j and L ∗j (y N ) = y j+1 , j ∈ N N −1 . For a fixed i ∈ N M , let αi,∗ j
be the scaling factors along the vertical grid line x = xi such that |αi,∗ j | < c j < 1
and let the shape parameters be selected so as to satisfy ri,∗ j > 0 and ti,∗ j > 0 for all
j ∈ N N −1 . Following the functional equation for the univariate rational cubic FIF
given in (10), we infer that the rational cubic FIF S ∗ corresponding to Δxi , i ∈ N M
enjoys the functional equation:
Pi,∗ j (φ)
S ∗ (xi , y) = αi,∗ j S ∗ xi , L ∗j −1 (y) + ∗ , (25)
Q i, j (φ)
where

Pi,∗ j (φ) = ri,∗ j (z i, j − αi,∗ j z i,1 )(1 − φ)3 + ti,∗ j (z i, j+1 − αi,∗ j z i,N )φ 3 + (2ri,∗ j + ti,∗ j )z i, j

y
+ ri,∗ j h ∗j z i, j − αi,∗ j (2ri,∗ j + ti,∗ j )z i,1 + ri,∗ j (y N − y1 )z i,1 (1 − φ)2 φ + (2ti,∗ j + ri,∗ j )z i, j+1
y
y
− ti,∗ j h ∗j z i, j+1 − αi,∗ j (2ti,∗ j + ri,∗ j )z i,N − ti,∗ j (y N − y1 )z i,N (1 − φ)φ 2 ,
y
Q i,∗ j (φ) = (1 − φ)ri,∗ j + φti,∗ j ,

L ∗j −1 (y) − y1 y − yj
and φ = = , y ∈ J j , j ∈ N N −1 .
y N − y1 h ∗j
Next, let us consider the univariate data sets obtained by taking sections of R with
the lines y = y j , j ∈ N N , namely Δ y j := {(xi , y j , z i, j , z i,x j ) : i ∈ N M }. Let
L i : [x1 , x M ] → [xi , xi+1 ] be affine maps L i (x) = ai x + bi satisfying L i (x1 ) = xi
and L i (x M ) = xi+1 , i ∈ N M−1 . For each fixed j ∈ N N , let the scaling factors obey
|αi, j | < ai < 1 and the shape parameters fulfill ri, j > 0 and ti, j > 0 for i ∈ N M−1 .
Modifying and adapting the expression for the univariate rational cubic FIF given in
the preceding section, we obtain the fractal function S interpolating Δ y j for j ∈ N N :
Pi, j (θ )
S(x, y j ) = αi, j S L i −1 (x), y j + , (26)
Q i, j (θ )
wherein

Pi, j (θ) = ri, j (z i, j − αi, j z 1, j )(1 − θ)3 + ti, j (z i+1, j − αi, j z M, j )θ 3 + (2ri, j + ti, j )z i, j + ri, j h i z i,x j

− αi, j (2ri, j + ti, j )z 1, j + ri, j (x M − x1 )z 1, x
j (1 − θ) θ + (ri, j + 2ti, j )z i+1, j
2

− ti, j h i z i+1,
x
j − αi, j (ri, j + 2ti, j )z M, j − ti, j (x M − x 1 )z M, j (1 − θ)θ ,
x 2
Q i, j (θ) = (1 − θ)ri, j + θti, j ,

L i −1 (x) − x1 x − xi
and θ = = , x ∈ Ii , i ∈ N M−1 .
x M − x1 hi
In what follows, we shall blend these univariate FIFs given in (25)–(26) using well-
known bicubic partially blended Coons patch to obtain the desired surface. Consider
the network of FIFs S(x, y j ), S(x, y j+1 ), S ∗ (xi , y), and S ∗ (xi+1 , y) for i ∈ N M−1
and j ∈ N N −1 . Consider the cubic Hermite functions b0,3 i (x) = (1 − θ )2 (1 + 2θ ),
j
i (x) = θ 2 (3 − 2θ ), b (y) = (1 − φ)2 (1 + 2φ), and b (y) = φ 2 (3 − 2φ). j
b3,3 0,3 3,3
These functions are called the blending functions, because their effect is to blend
together four separate boundary curves to provide a single well-defined surface. On
each individual patch Ri, j = Ii × J j , i ∈ N M−1 and j ∈ N N −1 , we define the surface
S(x, y )
i (x) i (x)

S ∗ (xi , y) j j j
Ψ (x, y) = b0,3 b3,3 + b0,3 (y) b3,3 (y)
S ∗ (xi+1 , y) S(x, y j+1 )
j
i z i, j z i, j+1 b (y)
− b0,3 (x) b3,3
i (x) 0,3 ,
z i+1, j z i+1, j+1 b j (y)
3,3
:= Ψ1 (x, y) + Ψ2 (x, y) − Ψ3 (x, y).
(27)
It is plain to see that Ψ1 provides a surface which is cubic blending of two bound-
aries S ∗ (xi , y) and S ∗ (xi+1 , y) and Ψ2 gives a surface which is cubic blending of
the remaining opposite pair of edge curves. In this process, the corners will be added
twice and Ψ3 provides the “correction” so that the successive substitution of x = xi ,
x = xi+1 , y = y j , and y = y j+1 quickly confirms that the surface patch has the four
original curves as its boundaries. The word bicubic partially blended is chosen to
suggest that only two of the cubic Hermite basis functions are used for the blending
process, and the blended surface is a fractal surface in the sense that the boundaries
consist of FIFs.
The following theorem is a direct consequence of the properties of the univariate

FIFs forming the boundaries of Ψ and the blending functions.
Theorem 7 The bicubic partially blended rational fractal surface Ψ (cf. (27))
satisfies the interpolation conditions Ψ (xi , y j ) = z i, j , ∂Ψ ∂ x (x i , y j ) = z i, j and
x
∂Ψ y
∂ y (x i , y j ) = z i, j , for i ∈ N M , j ∈ N N , and possesses C -continuity.
1
6 Numerical Examples
The aim of this section is to illustrate the rational cubic fractal interpolation surface
(FIS) with some examples. Consider the positive surface interpolation data (see
Table 1) with 16 points taken at random. Let us note that in Table 1, the 1st, 2nd,
Table 1 Interpolation data for positive blending rational cubic FISs

↓ x/y → 0.5 1.5 2.5 4
1 (1.1, 0.1, 4.3) (3.7, 0.7, 2.1) (219.1, 0.6, 5.8) (812.1, 0.11, 4)
2 (213.6, 1.7, 3) (493.2, 3.8, 2.9) (11.2, 2.1, 9.1) (819.9, 1, 0.1)
3 (736.2, 2.8, 3) (312.9, 1.55, (18.2, 1, 12.8) (27.1, 2, 3.2)
11.1)
4 (4, 3.7, 2.7) (7.9, 8.99, 3.9) (9.2, 3.22, 8.5) (11.88, 1.7, 9.22)
(a) (b)
1000 1000
800 800
600 600
400 400
200 200
0 0
−200 −200
4 4
3 4 3 4
2 3 2 3
1 2 1 2
0 1 0 1
(c) (d)
1000 1000
800 800
600 600
400 400
200 200
0 0
−200 −200
4 4
3 4 3 4
2 3 2 3
1 2 1 2
0 1 0 1
(e) (f)
1500 1000
1000 800
600
500
400
0
200
−500 0
4 4
3 4 3 4
2 3 2 3
1 2 1 2
0 1 0 1
Fig. 1 Partially bicubically blended rational FIS. a Rational cubic FIS. b Effect of change in α
in Fig. 1a. c Effect of change in α* in Fig. 1a. d Effect of change in r and t in Fig. 1a. e Effect of
change in r* and t* in Fig. 1a. f Classical rational cubic FIS
and 3rd components of (., ., .) represent the function value, the first order partial
derivatives in x-direction, and y-direction at (xi , y j ), where i, j are in {1, 2, 3, 4}.
For simplicity of presentation, let us represent the parameters, namely the (hori-
zontal) scaling factors |αi, j | < ai < 1 and the (horizontal) shape parameters ri, j > 0,
ti, j > 0 with the aid of matrices of order (M − 1) × N as follows: α = [αi, j ],
r = [ri, j ], and t = [ti, j ]. Similarly, the parameters along the vertical grid lines may
be represented using matrices of order M × (N − 1): α ∗ = [αi,∗ j ], r ∗ = [ri,∗ j ], and
t ∗ = [ti,∗ j ]. We shall refer to the parameter matrices α, α ∗ as scaling matrices and r ,
t, r ∗ , t ∗ as shape matrices.
The details of the scaling and the shape parameters used in the construction of
Fig. 1a–f are provided in Tables 2 and 3. By using the matrices of the scaling and the
shape parameters (see Tables 2 and 3), a rational cubic FIS is generated in Fig. 1a,
which is taken as a reference surface. We construct a blending cubic rational FIS in
Fig. 1b by changing the scaling matrix α with respect to the IFS matrices of Fig. 1a.
The effects of perturbed α ∗ in Fig. 1a are captured in Fig. 1c. Similarly, by changing
the shape matrices r , t, r ∗ and t ∗ we obtain Fig. 1d, e. We retrieve the classical rational
cubic surface plotted in Fig. 1f by taking α = [0]3×4 and α ∗ = [0]4×3 . We selected
the parameters at random except for the mild conditions imposed in the construction
of the cubic spline FIS. Among various values of the scaling factors and the shape
parameters satisfying the conditions imposed for C 1 -continuity, we can choose the
values so that the constructed interpolant satisfies certain additional conditions, for
instance, preserving shape inherent in the data set. This is reserved for a future work.
Let us conclude the paper with some remarks. By using the convergence of the
univariate FIFs forming the boundaries, it can be shown that the fractal surface Ψ
converges to the original function Φ generating the bivariate data. It is shown that
by appropriate choices of derivative parameters, the global continuity of the uni-
variate rational cubic spline can be made C 2 . By using bicubically blended Coons
patch (instead of partially bicubically blended scheme) on a network of C 2 boundary
curves, we will be able to obtain a C 2 -continuous surface. From the shape preserv-
ing properties of the networks of boundary curves that follow from [16], the shape
preserving properties of transfinite interpolating surface Ψ can be deduced. These
topics will be the focal point of the articles that follow.
Table 2 Scaling matrices in the construction of blending rational cubic FISs in Fig. 1a–f
Scaling matrices in x-direction Figure Scaling matrices in y-direction Figure
⎡ ⎤
⎡ ⎤ −0.28 −0.27 −0.28
−0.31 −0.31 −0.3 −0.2 ⎢ ⎥
⎢ ⎥ ⎢ 0.26 0.25 0.24 ⎥
α = ⎣ −0.3 −0.29 −0.28 −0.25 ⎦ 1a, c–e, f α∗ = ⎢
⎢ 0.27 0.28 0.29 ⎥
⎥ 1a, b, d, e
⎣ ⎦
0.19 0.29 0.31 0.30
−0.27 −0.28 −0.28
⎡ ⎤
⎡ ⎤ 0.28 0.27 0.279
0.3100 0.3100 0.3300 0.2000 ⎢ ⎥
⎢ ⎥ ⎢ −0.26 −0.25 −0.24 ⎥
α = ⎣ −0.3010 −0.3290 −0.3280 −0.3250 ⎦ 1b α∗ = ⎢
⎢ 0.27 −0.28 0.285 ⎥
⎥ 1c
⎣ ⎦
0.3190 0.3285 0.3090 0.3000
0.2690 0.279 0.28
Table 3 Scaling matrices in the construction of blending rational cubic FISs in Fig. 1a–f
Matrices of shape parameters in Figure Matrices of shape parameters in Figure
x-direction y-direction
⎡ ⎤
⎡ ⎤ 3.5 20 11
2 8.5 13.7 56 ⎢ 4.5 10 23 ⎥
⎢ ⎥ ⎢ ⎥
r = ⎣ 40 23 98 45 ⎦ 1a–c, e, f r∗ = ⎢ ⎥ 1a–d, f
⎣ 8.9 40 54 ⎦
75 35.5 17 8
27 30 19
⎡ ⎤
⎡ ⎤ 4 10 23
40 2 98 4 ⎢ 8 40 54 ⎥
⎢ ⎥ ⎢ ⎥
r = ⎣ 75 35 17 8 ⎦ 1d r∗ = ⎢ ⎥ 1e
⎣ 3 20 10 ⎦
2 34 14 16
27 3 19
⎡ ⎤
⎡ ⎤ 10 10 10
1111 ⎢ 10 10 10 ⎥
⎢ ⎥ ⎢ ⎥
t = ⎣1 1 1 1⎦ 1a–c, e, f t∗ = ⎢ ⎥ 1a–d, f
⎣ 10 10 10 ⎦
1111
10 10 10
⎡ ⎤
⎡ ⎤ 111
1111 ⎢1 1 1⎥
⎢ ⎥ ⎢ ⎥
t = ⎣2 2 2 2⎦ 1d t∗ = ⎢ ⎥ 1e
⎣1 1 1⎦
1111
111
Acknowledgments The authors are thankful to the Department of Science & Technology, India
for the SERC DST Project No. SR/S4/MS: 694/10.
References
1. Barnsley, M.F.: Fractal functions and interpolations. Constr. Approx. 2, 303–329 (1986)
2. Barnsley, M.F.: Fractals Everywhere. Academic Press, Orlando (1988)
3. Barnsley, M.F., Harrington, A.N.: The calculus of fractal interpolation functions. J. Approx.
Theory 57(1), 14–34 (1989)
4. Böhm, W.: A survey of curve and surface methods in CAGD. Comput. Aided Geom. Des. 1,
1–60 (1984)
J. Numer. Anal. 44(2), 655–676 (2006)
6. Chand, A.K.B., Navascués, M.A.: Natural bicubic spline fractal interpolation. Nonlinear Anal.
TMA 69, 3679–3691 (2008)
7. Chand, A.K.B., Navascués, M.A.: Generalized Hermite fractal interpolation. Rev. R. Acad.
Cienc. Zaragoza 64(2), 107–120 (2009)
8. Chand, A.K.B., Vijender, N.: Monotonicity preserving rational quadratic fractal interpolation
functions. Adv. Num. Anal. 2014(Art ID 504825), 1–17 (2014)
9. Chand, A.K.B., Vijender, N., Agarwal, R.P.: Rational iterated function system for posi-
tive/monotonic shape preservation. Adv. Differ. Equ. 2014(30), 1–19 (2014)
10. Chand, A.K.B., Viswanathan, P.: A constructive approach to cubic Hermite Fractal Interpolation
Function and its constrained aspects. BIT Numer. Math. 53, 841–865 (2013)
11. Duan, Q., Chen, T.Z., Djidjeli, K., Price, W.G., Twizell, E.H.: Convexity control and approxi-
mation properties of interpolating curves. Korean J. Comput. Appl. Math. 7(2), 97–405 (2000)
12. Greiner, H.: A survey on univariate data interpolation and approximation by splines of given
shape. Math. Comput. Model. 15(10), 97–108 (1991)
13. Navascués, M.A.: Fractal polynomial interpolation. Z. Anal. Anwend. 24(2), 401–418 (2005)
14. Shrivastava, M., Joseph, J.: C 2 -rational cubic spline involving tension parameters. Proc. Indian
Acad. Sci. (Math. Sci.) 110(3), 305–314 (2000)
15. Viswanathan, P., Chand, A.K.B.: A C 1 rational cubic fractal interpolation function: convergence
and associated parameter identification problem. Acta Appl. Math. 136(1), 19–41 (2015)
16. Viswanathan, P., Chand, A.K.B.: α-fractal rational splines for constrained interpolation. Elec-
tron. Tran. Numer. Anal. 41, 420–442 (2014)
tal shapes: bounds on the scale factors. J. Math. Anal. Appl. 419(2), 804–817 (2014)
Constrained 2D Data Interpolation Using
Rational Cubic Fractal Functions
A.K.B. Chand and K.R. Tyada
Abstract In this paper, we construct the C 1 -rational cubic fractal interpolation func-
tion (RCFIF) and its application in preserving the constrained nature of a given data
set. The C 1 -RCFIF is the fractal design of the traditional rational cubic interpolant
(θ)
of the form qpii (θ) , where pi (θ) and qi (θ) are the cubic polynomials with three tension
parameters. We derive the uniform error bound between the RCFIF with the origi-
nal function in C 3 [x1 , xn ]. When the data set is constrained between two piecewise
straight lines, we deduce the sufficient conditions on the parameters of the RCFIF
so that it lies between those two lines. Numerical examples are given to support that
our method is interactive and smooth.
Keywords Fractal interpolation · Rational function · Rational fractal interpolation ·

Convergence · Constrained data · Positivity.
1 Introduction
In the literature a wide range of development on shape preserving classical spline

interpolation techniques have been discussed, see [1–3, 13–18]. Although the clas-
sical splines interpolate the data smoothly, certain derivatives of the classical inter-
polants are either piecewise smooth or globally smooth in nature. Therefore, the
classical interpolants are not suitable to approximate functions that have irregular
nature or fractality in their first-order derivatives, see for instance [9–11, 19, 20]. On
the other hand, fractal interpolation is an ideal tool in such scenario as well. Since
the classical polynomial spline interpolant representation available in the literature
is unique for given data, and it simply depends on the data points, it is difficult to
A.K.B. Chand · K.R. Tyada (B)

e-mail: kurmaths86@gmail.com
A.K.B. Chand
DOI 10.1007/978-81-322-2485-3_49
594 A.K.B. Chand and K.R. Tyada
preserve all the hidden shape properties of the given data, for example, data between
two piecewise straight lines, positivity, monotonicity, or convexity. For this reason a
user needs an interactive shape preserving smooth curve representation of interpola-
tion data. In these cases, the rational interpolation functions with shape parameters
are preferred over various shape preserving interpolating techniques, see [15].
Fractal interpolation is a modern and advance technique to analyze various sci-
entific data. Barnsley [4] coined the term fractal interpolation function which was
constructed based on the theory of iterated functions system (IFS). An IFS ensures an
attractor which is the graph of a continuous function that interpolates the given data
points. FIFs are the fixed points of Read-Bajraktaverić operator [4], which is defined
on suitable function spaces. Barnsley and Harrington [5] introduced the construc-
tion of k- times differentiable polynomial spline FIF with a fixed type of boundary
conditions. The polynomial spline FIFs with general body conditions were studied
recently in [6–8]. Dalla and Drakopoulos [12] introduced polar fractal interpolation
functions and developed the range restriction concept for a FIF. A specific feature
of spline FIF is that its certain derivative can be used to capture the irregularity
associated with the original function from where the interpolation data is obtained.
The use of fractal splines for constrained curve interpolation has been extensively
investigated in the literature, see [1, 3, 14, 18], and references therein. Abbas [1] con-
structed a C 1 -piecewise rational cubic function to preserve the shape of constrained
2D and 3D data. Awang [3] developed a C 2 -rational cubic function to 2D constrained
data interpolation. Duan [14] constructed a kind of rational spline based on the func-
tion values to constrain the interpolating curve to lies between two piecewise straight
lines. Hussain and collaborators [16–18] used different types of C 1 -piecewise ratio-
nal cubic functions to preserve the shape of various constrained data.
The shape preservation of scientific data through different types of smooth rational
FIFs are studied recursively in [8–11, 19, 20]. Inspired by the work of Duan in
constrained interpolation, we have constructed the smooth RCFIF so that it can be
used for shape preservation. In particular, when the interpolation data set lies in
between the two given piecewise straight lines, the IFS parameters of the proposed
RCFIF are restricted so that it lies between these straight lines.
The paper is organized as follows: In Sect. 2, the general theory of FIF for a given
data set is reviewed. The construction of C 1 RCFIFs passing through a set of data
points is discussed in Sect. 3. In Sect. 4, we have deduced the error estimation of
the RCFIF with an original function for convergence results. In Sect. 5, the range
of scaling factors and shape parameters are restricted according to the sufficient
conditions so that the developed RCFIF lies between two piecewise straight lines
followed by conclusions in Sect. 6.
2 Basics of FIF Theory
Let P : {x1 , x2 , . . . , xn } be a partition of the real compact interval I = [x1 , xn ],

where x1 < x2 < · · · < xn . Denote Λ := {1, 2, . . . , n − 1} and Λ∗ := {1, 2, . . . , n}.
Constrained 2D Data Interpolation Using Rational Cubic Fractal Functions 595
Let a set of data points {(x j , f j ) ∈ I × K : j ∈ Λ∗ } be given, where K is a compact

set in R. Let Ii = [xi , xi+1 ] and L i : I → Ii , i ∈ Λ be contractive homeomorphisms
such that L i (x1 ) = xi , L i (xn ) = xi+1 for i ∈ Λ, and
|L i (x) − L i (x ∗ )| ≤ li |x − x ∗ | ∀ x, x ∗ ∈ I, 0 < li < 1. (1)
Let C = I × K , and consider n − 1 continuous mappings Fi : C → K satisfying

Fi (x1 , f 1 ) = f i , Fi (xn , f n ) = f i+1 , i ∈ Λ,
|Fi (x, f ) − Fi (x, f˜)| ≤ |λi || f − f˜| ∀ x ∈ I, ∀ f, f˜ ∈ K for 0 ≤ |λi | < 1. (2)
Now, define functions ωi : C → Ii × K such that ωi (x, f ) = (L i (x), Fi (x, f ))

∀ i ∈ Λ.
Proposition 1 (Barnsley and Harrington [4]) The IFS {C; ωi , i ∈ Λ} defined above
admits a unique attractor G such that G is the graph of a continuous function
g : I → K which obeys g(x j ) = f j for j ∈ Λ∗ .
The above function g is called a FIF associated with the IFS {I × K ; ωi (x, f ) =
(L i (x), Fi (x, f )), i ∈ Λ}. The functional representation of g follows from the fixed
point of the Read-Bajraktarević operator T [4]. The FIF g satisfies the following
functional equation:
T g(x) ≡ Fi (L i−1 (x), g ◦ L i−1 (x)) = g(x), x ∈ Ii , i ∈ Λ. (3)
The following IFS has been studied extensively in the literature of FIF theory:
{C; ωi (x, f ) = (L i (x), Fi (x, f )), i ∈ Λ}, (4)
where L i (x) = ai x + bi , Fi (x, f ) = λi f + Mi (x) with Mi : I → R are suitable

continuous functions such that (2) is satisfied. The multiplier λi is called a scaling
factor of the transformation ωi , and λ = (λ1 , λ2 , . . . , λn−1 ) is the scale vector
associated with the IFS (4). The scaling factors give an additional degree of freedom
to FIFs over its counterparts in classical interpolation and allow us to modify their
shape preserving properties. The existence of a spline FIF is given by Barnsley and
Harrington [5] based on the calculus of fractal functions, and that result has been
extended for the existence of rational spline FIF in the following theorem [11]:
Theorem 1 Let {(x j , f j ) : j ∈ Λ∗ } be the given data set such that x1 < x2 <
−xi
· · · < xn . Suppose that L i (x) = ai x + bi , where ai = xxi+1
n −x 1
, bi = xn xxi −x 1 xi+1
n −x 1
pi (x)
and Fi (x, f ) = αi f + Mi (x), Mi (x) = qi (x) , pi (x) and qi (x) are chosen poly-
nomials of degree r and s, respectively, and qi (x) = 0 ∀ x ∈ [x1 , xn ] for i ∈
p
Λ. Suppose for some integer p ≥ 0, |αi | < ai , i ∈ Λ. Let Fi,m (x, f ) =
(m) (m) (m)
λi f +Mi (x) M (x ) M (xn ) (m)
, f 1,m = a1m −λ11 , f n,m = a m n−1 , m = 1, 2, ..., p, where Mi (x)
aim 1 n−1 −λn−1
represents the mth derivative of Mi (x) with respect to x. If Fi,m (xn , f n,m ) =
Fi+1,m (x1 , f 1,m ), i = 1, 2, . . . , n − 2, m = 1, 2, . . . , p, then the IFS {I ×

K ; ωi (x, f ) = (L i (x), Fi (x, f )), i ∈ Λ} determines a rational FIF Φ ∈ C p [x1 , xn ]
such that Φ(L i (x)) = λi Φ(x) + Mi (x), and Φ (m) is the FIF determined by
{I × K ; ωi,m (x, f ) = (L i (x), Fi,m (x, f )), i = 1, . . . , n − 1} for m = 1, 2, . . . , p.
3 C 1 -Rational Cubic Fractal Interpolation Function
In this section, we construct the RCFIF with three shape parameters in each subinter-
val with the help of Theorem 1. Let {(x j , f j ), j ∈ Λ∗ } be a given set of interpolation
data for an original function Ψ such that x1 < x2 < · · · < xn . Consider the IFS
{I × K ; ωi (x, f ) = (L i (x), Fi (x, f )), i ∈ Λ}, where L i (x) = ai x + bi and
(x)
Fi (x, f ) = λi f (x) + Mi (x), Mi (x) = qpii (x) , where pi (x) and qi (x) are cubic
(1)
polynomials, qi (x) = 0 ∀ x ∈ [x1 , xn ], and |λi | < ai , i ∈ Λ. Let Fi (x, d) =
(1)
λi d+Mi (x) (1)
ai , where Mi (x) is the first order derivative of Mi (x), x ∈ [x1 , xn ].
Fi (x, f ) satisfying the following join-up conditions:
Fi (x1 , f 1 ) = f i , Fi (xn , f n ) = f i+1 , Fi(1) (x1 , d1 ) = di , Fi(1) (xn , dn ) = di+1 , (5)
where di denote the first-order derivative of Ψ with respect to x at knot xi . The

attractor of the above IFS will be the graph of a C 1 -rational cubic FIF. From (3) one
can observe that our FIF can be written as
pi (θ)
Φ(L i (x)) = λi Φ(x) + Mi (x) = λi Φ(x) + , (6)
qi (θ)
where
pi (θ) = (1 − θ)3 Ui + θ(1 − θ)2 Vi + θ2 (1 − θ)Wi + θ3 X i ,
qi (θ) = (1 − θ)3 αi + θ(1 − θ)γi + θ3 βi ,
θ = x−xl , l = x n − x 1 , x ∈ I , and αi , βi and γi are positive real shape parameters.

1
To ensure that the rational cubic FIF is C 1 -continuous, the following interpolation
conditions are imposed:
Φ(L i (x1 )) = f i , Φ(L i (xn )) = f i+1 , Φ (L i (x1 )) = di , Φ (L i (xn )) = di+1 . (7)
From (6) and (7), it is clear that at x = x1 we get
Ui
Φ(L i (x1 )) = f i ⇒ f i = λi f 1 + ⇒ Ui = αi ( f i − λi f 1 ).
αi
Similarly, at x = xn we obtain
Xi
Φ(L i (xn )) = f i+1 ⇒ f i+1 = λi f n + ⇒ X i = βi ( f i+1 − λi f n ).
βi
Again from (6) and (7), at x = x1 we observe that
Vi − γi ( f i − λi f 1 )
Φ (L i (x1 )) = di ⇒ ai di = λi d1 +
αi
⇒ Vi = γi ( f i − λi f 1 ) + αi (ai di − λi d1 ).
Similarly, at x = xn we notice that
γi ( f i+1 − λi f n ) − Wi
Φ (L i (xn )) = di+1 ⇒ ai di+1 = λi dn +
βi
⇒ Wi = γi ( f i+1 − λi f n ) − βi (ai di+1 − λi dn ).
Now substituting Ui , Vi , Wi and X i in (6), we get the required C 1 -RCFIF with the
numerator,
pi (θ) = αi ( f i − λi f 1 )(1 − θ)3 + {γi ( f i − λi f 1 ) + αi (ai di − λi d1 )}θ(1 − θ)2

+ {γi ( f i+1 − λi f n ) − βi (ai di+1 − λi dn )}θ2 (1 − θ) + βi ( f i+1 − λi f n )θ3 .
In most applications, the derivatives d j ( j ∈ Λ∗ ) are not given, and hence must be
calculated either from the given data or by some numerical methods. In this paper we
have calculated d j , j ∈ Λ∗ from the given data using the arithmetic mean method.
Remark 1 If λi = 0 for all i ∈ Λ, the RCFIF Φ becomes the classical rational cubic
interpolation function S(x)(say) that is defined in [2] on each subinterval [xi , xi+1 ]
as
pi (z)
S(x) = , x ∈ [xi , xi+1 ], (8)
qi (z)
where z = x−x h i , h i = x i+1 − x i ,

i
pi (z) = αi f i (1 − z)3 + (γi f i + h i αi di )z(1 − z)2 + (γi f i+1 − h i βi di+1 )z 2 (1 − z) +

βi f i+1 z 3 , and qi (z) = αi (1 − z)3 + γi z(1 − z) + βi z 3 .
Due to the implicit expression of the RCFIF Φ, it is not easy to compute the uniform
error bound Φ − Ψ ∞ by using any standard numerical analysis techniques. Hence
we derive an upper bound of the error by using the classical counterpart S of Φ with
the help of
Φ − Ψ ∞ ≤ Φ − S ∞ + S − Ψ ∞ , (9)
where S is given by (8).

Now the error estimation between the original function Ψ and the classical rational
cubic function S in an arbitrary subinterval Ii = [xi , xi+1 ] can be found by using
the Peano-Kernel theorem and the details are given in [2].
Proposition 2 The error between the classical rational cubic function defined in (8)
and the original function Ψ ∈ C 3 [x1 , xn ] is
|Ψ (x) − S(x)| ≤ Ψ (3) h i3 ci , x ∈ [xi , xi+1 ], (10)
ci = max0≤z≤1 Θ(αi , βi , z),

max Θ1 (αi , βi , z) for 0 ≤ z ≤ z ∗ ,
Θ(αi , βi , z) =
max Θ2 (αi , βi , z) for z ∗ ≤ z ≤ 1,
where z ∗ = 1 − (γi β−β

i
i)
and Θ1 (αi , βi , γi , z), and Θ2 (αi , βi , γi , z) are obtained
from the proof of Theorem 3.1 in [2].
Theorem 2 Let Φ be the C 1 continuous RCFIF and Ψ ∈ C 3 [x1 , xn ] is the data

generating function with respect to the given data {(x j , f j ), j ∈ Λ∗ }. Let d j be the
bounded first order derivative at the knot x j , j ∈ Λ∗ . Let |λ|∞ = max{|λi |, i ∈ Λ}
and the shape parameters αi > 0, βi > 0 and γi > max{αi , βi } for i ∈ Λ. Then
|λ|∞
Ψ − Φ ∞ ≤ Ψ (3) ∞ h 3 c + (E(h) + E ∗ (h)), (11)
1 − |λ|∞
where h = max1≤i≤n−1 {h i }, E(h) = Ψ ∞ + 2h E 1 , E ∗ (h) = F + 2h E 2 with

E 1 = max1≤ j≤n {|d j |}, F = max{| f 1 |, | f n |}, E 2 = max{|d1 |, |dn |}, and c =
max1≤i≤n−1 {ci }, ci is defined as in Proposition 2.
Proof Consider the space F ∗ = {g ∈ C 1 (I, R) | g(x1 ) = f 1 , g(xn ) =

f n , g (x1 ) = d1 , g (xn ) = dn }. From (2) and (6), the Read-Bajraktarević opera-
tor Tλ∗ : F ∗ → F ∗ for the RCFIF can be written as

pi L i−1 (x), λi

Tλ∗ g(x) = λi g L i−1 (x) + , x ∈ Ii , i ∈ Λ. (12)
qi L i−1 (x)
Note that Φ is the fixed point of Tλ∗ with λ = 0 and S is the fixed point of T0∗ . Since
Tλ∗ is a contractive operator with the contraction factor |λ|∞ , we have
Tλ∗ Φ − Tλ∗ S ∞ ≤ |λ|∞ Φ − S ∞ . (13)
From (12), we have

−1
⎛
⎞

∂ pi L i (x),τi
⎜ −1 ⎟
⎜ qi L i (x) ⎟
|Tλ∗ S(x) − T0∗ S(x)| ≤ |λ|∞ ⎜
⎜ S ∞ +

⎟ , |τi | ∈ (0, λi ),
⎟
⎝ ∂λi ⎠

(14)
where the mean value theorem for function of several variables is used in this cal-
culation. Now we wish to find out the error bounds of the terms on the right sides
of (14). From the classical rational cubic trigonometric function (8), it is easy to
observe that
S(x) = σ1 (αi , βi , γi , z) f i + σ2 (αi , βi , γi , z) f i+1 + σ3 (αi , βi , γi , z)di − σ4 (αi , βi , γi , z)di+1 ,

(15)
where σ1 (αi , βi , γi , z) = qi 1(z) αi (1 − z)3 + γi z(1 − z)2 ≥ 0,

σ2 (αi , βi , γi , z) = qi 1(z) γi z 2 (1 − z) + βi z 3 ≥ 0,

σ3 (αi , βi , γi , z) = qih(z)
i
αi z(1 − z)2 ≥ 0,
2
σ4 (αi , βi , γi , z) = qih(z)
i
βi z (1 − z) ≥ 0.
It is easy to verify that σ1 (αi , βi , γi , z) + σ2 (αi , βi , γi , z) = 1.
Also, for αi > 0, βi > 0, γi > 0 and choosing γi > max{αi , βi } we obtain the
following inequality:
hi
σ3 (αi , βi , γi , z) + σ4 (αi , βi , γi , z) = αi z(1 − z)2 + βi z 2 (1 − z)
qi (z)

αi βi
≤ hi + ≤ 2h i .
γi γi
Thus, |S(x)| ≤ max j=i,i+1 {| f j |}+2h i max j=i,i+1 {|d j |}. Since the above estimation
is true for i ∈ Λ, we get the following estimation:
S ∞ ≤ E(h) := Ψ ∞ + 2h E 1 , (16)
Since qi (x) is independent of λi , from the first term in the right side of (14),

pi (L i−1 (x),τi )
∂
qi (L i−1 (x))
= σ1 (αi , βi , γi , z) f 1 + σ2 (αi , βi , γi , z) f n + σ3 (αi , βi , γi , z)d1
∂λi
− σ4 (αi , βi , γi , z)dn .
Now by applying a similar argument, the following estimate can be obtained:

p (L −1 (x),τ )
∂ i i −1 i
qi (L i (x))
≤ E ∗ (h) := F + 2h E 2 . (17)
∂λi

Substituting (16) and (17) into (14), we have
|Tλ∗ S(x) − T0∗ S(x)| ≤ |λ|∞ (E(h) + E ∗ (h)), x ∈ [x1 , xn ].
Consequently, we obtain
Tλ∗ S − T0∗ S ∞ ≤ |λ|∞ (E(h) + E ∗ (h)). (18)
Using (13) and (18) in

Φ − S ∞ = Tλ∗ Φ − T0∗ S ∞ ≤ Tλ∗ Φ − Tλ∗ S ∞ + Tλ∗ S − T0∗ S ∞ ,
we have the following estimation:
|λ|∞ (E(h) + E ∗ (h))

Φ − S ∞ ≤ . (19)
1 − |λ|∞
From Proposition 2, we have |Ψ (x) − S(x)| ≤ |Ψ (3) |h i3 ci which gives that
Ψ − S ∞ ≤ Ψ (3) ∞ h 3 c, (20)
where h and c are defined in the statement of theorem. Substituting (19) and (20) in
(9), we obtain the desired upper bound in (11).
Convergence Result: Assume that max1≤ j≤n {|d j |} is bounded for every partition of
the domain I . Since |λi | < ai , i ∈ Λ ⇒ |λ|∞ < h , and Theorem 2 proves that the
RCFIF Φ converges uniformly to the original function Ψ as h → 0. Additionally, if
h i3
|λi | < ai3 = 3
for i ∈ Λ, then Ψ − Φ ∞ = O(h 3 ) as h → 0.
5 Constrained C 1 -RCFIF
In this section, we discuss on the construction of a constrained RCFIFs whose graph

lie in between two piecewise straight lines ‘L u ’ and ‘L b ’ when the given interpolation
data are distributed between ‘L u ’ and ‘L b ’. In general, a RCFIF may not lie in
between ‘L u ’ and ‘L b ’ with an arbitrary choice of IFS parameters. In order to avoid
this circumstance, it is required to deduce sufficient data dependent restrictions on
the scaling factor λi and on the shape parameters αi , βi and γi so that the RCFIF
preserves the shape of the constrained data.
Suppose that the line ‘L u ’ is defined piecewise over [xi , xi+1 ] such that L u (x j ) =
f j ∀ j ∈ Λ∗ . Similarly, ‘L b ’ is defined piecewise over [xi , xi+1 ] such that L b (x j ) =
u

f jb ∀ j ∈ Λ∗ . The IFSs for ‘L u ’ and ‘L b ’ over I are given by R; (L i (x), Fiu (x)),

i ∈ Λ}, R; (L i (x), Fib (x)), i ∈ Λ where Fib (x) = (1 − θ)μi + θηi , and Fiu (x) =
(1 − θ)μi∗ + θηi∗ where θ = xx−x 1
n −x 1
with μi = m i xi + ci , ηi = m i xi+1 + ci and
∗ ∗ ∗ ∗ ∗ ∗
μi = m i xi + ci , ηi = m i xi+1 + ci , i ∈ Λ.
Theorem 3 Let Φ be the RCFIF (6) defined over the interval [x1 , xn ] with respect
to the given data {(x j , y j ), j ∈ Λ∗ }. Further, assume that the data points lie above
the straight line ‘L u ’ and below the straight line ‘L b ’. Then the RCFIF Φ lies in
between the straight line ‘L u ’ and ‘L b ’ if the following conditions are satisfied for
all i ∈ Λ:
(i) Select the scaling factors as
0 < λi < min{λiu , λib }, (21)
(ii) Select the shape parameters as
αi > 0, βi > 0 and γi > max{γiu , γib }, (22)
where λiu , γiu , λib and γib are defined in (29)–(32) respectively.
Proof Let {(x j , f j ) : j ∈ Λ∗ } be the given set of data points lying in between the
straight lines L u and L b . Then,
m i x j + ci = f jb < f j < f ju = m i∗ x j + ci∗ ∀ i ∈ Λ j ∈ Λ∗ .

Since ‘L u ’ and ‘L b ’ are FIFs associated with the IFSs R; (L i (x), Fiu (x)), i ∈ Λ

and R; (L i (x), Fib (x)), i ∈ Λ respectively, then the functional equations of ‘L u ’
and ‘L b ’ are
L b (L i (x)) = m i L i (x) + ci = μi (1 − θ) + ηi θ = ri (θ) (say),

(23)
L u (L i (x)) = m i∗ L i (x) + ci∗ = μi∗ (1 − θ) + ηi∗ θ = ri∗ (θ) (say),
−xi
where L i (x) = ai x + bi with ai = xxi+1 n −x 1
and bi = xn xxi −x 1 xi+1
n −x 1
, θ = x−x
,
1
= xn − x1 .
Note that at x = x1 , μi = m i xi + ci , μi∗ = m i∗ xi + ci∗ and at x = xn , ηi =
m i xi+1 + ci , ηi∗ = m i∗ xi+1 + ci∗ . Thus the curve will lie in between the straight lines
‘L u ’ and ‘L b ’ if the C 1 -RCFIF Φ satisfies the following conditions:
L b (L i (x)) < Φ(L i (x)) < L u (L i (x)) ∀ x ∈ [x1 , xn ], i ∈ Λ. (24)

x −x ∗j
Let θ j = xnj −x11 and let ri = ri (θ j ), ri = ri∗ (θ j ). Assume that λi ∈ [0, ai ), i ∈ Λ
j
as Φ ∈ C 1 [x1 , xn ]. In order to the RCFIF Φ lies between the piecewise straight lines
‘L u ’ and ‘L b ’, it is clear from (24) that for the next generation of interpolation points
should satisfy the following inequalities:
∗j
ri (θ j ) < Φ(L i (x j )) < ri∗ (θ j ) ⇒ ri < Φ(L i (x j )) < ri ,
j
j pi (θ j ) pi (θ j ) ∗j pi (θ j ) (25)
⇒ λi ri + < λi f j + < λi ri + .
qi (θ j ) qi (θ j ) qi (θ j )
j pi (θ j ) ∗j
For the validity of ri < λi f j + qi (θ j ) < ri , we need to impose the following
conditions from (25):
j j pi (θ j ) ∗j pi (θ j ) ∗j
ri < λi ri + , and λi ri + < ri .
qi (θ j ) qi (θ j )
Therefore, the RCFIF lies in between the straight lines ‘L u ’ and ‘L b ’ if
pi (θ j )
≥ 0 ∀ θ ∈ [0, 1], i ∈ Λ j ∈ Λ∗ ,
j
Ω1,i (θ j ) := (λi − 1)ri + (26)
qi (θ j )
∗j pi (θ j )
Ω2,i (θ j ) := (λi − 1)ri + ≤ 0 ∀ θ ∈ [0, 1], for every i ∈ Λ j ∈ Λ∗ . (27)
qi (θ j )
After some algebraic simplifications, Ω1,i (θ) is reformulated as
pi∗ (θ j )
Ω1,i (θ j ) = > 0, (28)
qi (θ j )
where pi∗ (θ j ) = (1 − θ j )3 Ui∗ + θ j (1 − θ j )2 Vi∗ + θ2j (1 − θ j )Wi∗ + θ3j X i∗ , with

Ui∗ = Ui + αi (λi − 1)ri , Vi∗ = Vi + γi (λi − 1)ri , Wi∗ = Wi + γi (λi − 1)ri , and
j j j
X i∗ = X i + βi (λi − 1)ri .
j
It is clear that the shape parameters αi > 0, βi > 0 and γi > 0 guarantee that
the denominator in (28) is positive. Thus the RCFIF preserves the constrained aspect
of the constrained data if (28) holds for all i ∈ Λ i.e., if the numerator pi∗ (θ j ) is
positive. Therefore, pi∗ (θ j ) > 0 if each Ui∗ , Vi∗ , Wi∗ and X i∗ are positive.
Since αi > 0 and Ui∗ = Ui +αi (λi −1)ri = αi ( f i −λi f 1 +(λi −1)ri ), j ∈ Λ∗ ,
j j
the choice of
j

f i − ri ∗
λi < Ξi := min j
: j ∈Λ yields Ui∗ > 0.
f 1 − ri
Similarly, since βi > 0 and X i∗ = X i + βi (λi − 1)ri = βi ( f i+1 − λi f n +

j
(λi − 1)ri ), j ∈ Λ∗ , the selection of

j
f i+1 − ri ∗
λi < i := min j
: j ∈Λ ensures X i∗ > 0.
f n − ri
Consider Vi∗ = Vi + γi (λi − 1)ri = γi ( f i − λi f 1 + (λi − 1)ri ) + αi (ai di − λi d1 ).

j j
Then for ai di − λi d1 > 0, arbitrary αi > 0 and γi > 0 provide Vi∗ > 0. Otherwise
for αi > 0, the choice of

−αi (ai di − λi d1 ) ∗
γi > ϒi := max j
: j ∈Λ results Vi∗ > 0.
f i − λi f 1 + (λi − 1)ri

Similarly consider Wi∗ = Wi + γi (λi − 1)ri = γi f i+1 − λi f n + (λi − 1)ri −
j j
βi (ai di+1 − λi dn ). Then for (ai di+1 − λi dn ) < 0, arbitrary βi > 0 and γi > 0
provide X i∗ > 0. Otherwise for βi > 0, the selection of

βi (ai di+1 − λi dn ) ∗

γi > ℵi := max j
: j ∈Λ produces Wi∗ > 0.
f i+1 − λi f n + (λi − 1)ri
Hence Ω1,i (θ j ) > 0 ∀ i ∈ Λ, j ∈ Λ∗ when

(i) the scaling factors are chosen as
λi < λiu := min{ai , Ξi , i } (29)
(ii) the shape parameters are chosen as αi > 0, βi > 0 and
γi > γiu := max{0, ϒi , ℵi }. (30)
Using the similar argument as above, we deduce that Ω2,i (θ j ) < 0 ∀ θ ∈ [0, 1],
i ∈ Λ j ∈ Λ∗ , i.e., the RCFIF Φ lies below the straight line ‘L u ’ when
(i) the scaling factors are selected as
λi < λib := min{Ξi∗ , i∗ } (31)
(ii) the shape parameters are selected as αi > 0, βi > 0 and
γi > γib := max{ϒi∗ , ℵi∗ }, (32)

∗j
∗j
ri − f i ri − f i+1
where Ξi∗ := min ∗j : j∈ Λ∗ ∗
, i := min ∗j
∗
: j ∈ Λ , ϒi∗ =
ri − f 1 ri − f n

−αi (ai di −λi d1 ) ∗ and ℵ∗ = max βi (ai di+1 −λi dn ) ∗
max ∗j : j ∈ Λ i ∗j : j ∈ Λ .
f i −λi f 1 +(λi −1)ri f i+1 −λi f n +(λi −1)ri
Thus the RCFIF preserves the constraining nature of the data and lies between the
straight lines if the IFS parameters are selected according to (21) and (22).
Remark 2 It is clear that positivity preserving interpolation is a special case of the

j
above developed constrained interpolation setting. By considering ri = 0 in (26) and
∗j
ri = ∞ in (27) for i ∈ Λ, j ∈ Λ∗ , the RCFIF (6) preserves the positivity feature
of the given data with respect to the restricted IFS parameters calculated from the
∗j
Theorem 3. Since ri = ∞, there is no need of captivating the RCFIF from above
by a piecewise straight line.
5.1 Numerical Example
We present a numerical example to illustrate the construction of the C 1 -RCFIFs

and related constrained interpolation problem discussed in the previous section. For
this we considered the interpolating data set {(0, 1), (0.5, 0.2), (1, 0.8), (1.5, 0.4)}
which is constrained in between the two straight lines taken from [14]:
⎧ 8 ⎧ 8
⎨ − 5 x + 1.07, 0 ≤ x ≤ 0.5,
⎪ ⎨ − 5 x + 0.93, 0 ≤ x ≤ 0.5,
⎪
L u = 65 x − 0.33, 0.5 ≤ x ≤ 1.0, , L b = 65 x − 0.47, 0.5 ≤ x ≤ 1.0,
⎪
⎩ 4 ⎪
⎩ 4
− 5 x + 1.67, 1.0 ≤ x ≤ 1.5, − 5 x + 1.53, 1.0 ≤ x ≤ 1.5.
(33)
The derivative values at the knots are calculated using the arithmetic mean method.
The constrained C 1 -RCFIF are generated iteratively using the IFS parameters given
in Table 1. For simplicity, we have fixed two of the shape parameters αi = 1 and
βi = 1, i = 1, 2, 3. For arbitrary choice of rational IFS parameters, the RCFIF Φ1
may not preserve the constrained nature of the given data, see for instance Fig. 1a.
So by using Theorem 3, we have calculated the restrictions on the IFS parameters
that satisfy the constrained inequalities (21) and (22), so that the RCFIF (6) must be
C 1 -continuous in [0, 1.5] and bounded between the upper straight line ‘L u ’ and the
lower straight line ‘L b ’. The choice of scaling factors and shape parameters as per
Theorem 3 are shown in Table 1. Figure 1b is generated as the graph of such RCFIF
Φ2 which preserves the constrained nature of given data for a particular restricted IFS
parameters. The constrained RCFIF Φ3 in Fig. 1c is generated with a perturbation
in λ2 , and it has major effects in second subinterval, while the changes in third
subinterval are also noticeable in comparison with Φ2 . This illustrates the global
effect of scaling parameter in a rational FIF. These effects are distributed according
Table 1 Scaling factors and shape parameters used in the RCFIFs

Figure Scaling factors (λ) Shape parameters (γ)
1a λ = (−0.1, 0.1323, 0.1324) γ = (8.4148, 19.1938, 12.0329)
1b λ = (0.0123, 0.1123, 0.1323) γ = (20, 100, 50)
1c λ = (0.0123, 0.01, 0.1323) γ = (20, 100, 50)
1d λ = (0.0123, 0.01, 0.01) γ = (20, 100, 50)
1e λ = (0.0123, 0.1123, 0.01) γ = (50, 100, 50)
1f λ = (0, 0, 0) γ = (50, 100, 50)
1.5 1.5
data points
(a) UnConstrained−RCFIF
(b) data points
Constrained−RCFIF
1 1
φ φ
Lu Lu
0.5 0.5
b
L
Lb
0 0
0 0.5 1 1.5 0 0.5 1 1.5
1.5 1.5
data points data points
(c) Constrained−RCFIF
(d) Constrained−RCFIF
1 1
φ
u
L
0.5 0.5
b
L
0 0
0 0.5 1 1.5 0 0.5 1 1.5
1.5 1.5
data points
(e) Constrained−RCFIF
(f) data points
Constrained−RCFIF
1 1
φ
Lu Lu φ
0.5 0.5
b
L Lb
0 0
0 0.5 1 1.5 0 0.5 1 1.5
Fig. 1 Rational cubic fractal interpolation functions. a Unconstrained RCFIF Φ1 . b Constrained

RCFIF Φ2 . c Constrained RCFIF Φ3 , effects of λ2 . d Constrained RCFIF Φ4 , effects of λ3 (λ2 ).
e Constrained RCFIF Φ5 , effects of λ3 , γ1 . f Constrained classical rational spline S
to the code space related with map L 2 in the given domain. Next, we modify only
λ3 with respect to IFS parameters of Φ3 to generate Φ4 . The perturbation effects of
scaling parameter(s) on the shape of Φ4 are worth to be noted in comparison with
the shape of Φ3 (Φ2 ). By changing the scaling factor λ3 and shape parameter γ1 , we
have constructed the constrained RCFIF Φ5 with pleasing effects in [x1 , x2 ] and in
[x3 , x4 ] whose graph is shown in Fig. 1e. We observe that in [x1 , x2 ] and [x3 , x4 ], the
RCFIF looks like converging to straight lines. The effect of fractality or irregularity
in some portions of RCFIF can be restricted by setting the associated scaling factors
to zero therein. Finally by setting all the scaling factors to zero, we have generated
the graph of classical rational cubic interpolant S in Fig. 1f.
6 Conclusions
In this paper we have constructed C 1 -RCFIF to preserve the constrained aspect of

given data. The RCFIF reduces to the traditional rational cubic interpolant by setting
all scaling factors to zero. The developed RCFIF converges uniformly to the data
generating original function as h → 0, and additionally if |λi | < ai3 , then the order of
convergence is O(h 3 ). We have developed the sufficient data-dependent conditions
on the rational IFS parameters to preserve the shape of the given data in such away
that the RCFIF lies between two piecewise straight lines. The effects of the rational
IFS parameters on the shape of the RCFIFs are illustrated. The developed RCFIF can
be used for the visualization of data with/without slopes at the knots. Applications
of the proposed RCFIF in geometric modeling problems are under investigation.
Acknowledgments The first author is thankful to the Department of Science & Technology, India
for the SERC DST Project No. SR/S4/MS: 694/10.
References
1. Abbas, M., Jamal, E., Ali, J.M.: Shape preserving constrained data visualization using spline
functions. Int. J. Appl. Math. Stat. 29(5), 34–50 (2012)
2. Abbas., M, Majid., A.A, Awang., M.N.H, Ali, J.M.: Local convexity shape-preserving data
visualization by spline function. ISRN Math. Anal. 2012(Article ID: 174048), 1–14 (2012)
3. Awang, M.N.H., Abbas. M., Majid, A.A., Ali, J.M.: Data visualization for constrained data
using C 2 rational cubic spline. In: Proceedings of the World Congress on Engineering and
Computer Science, vol. 1 (2013)
4. Barnsley, M.F.: Fractal functions and interpolation. Constr. Approx. 2, 303–329 (1986)
5. Barnsley, M.F., Harrington, A.N.: The calculus of fractal interpolation functions. J. Approx.
Theory 57(1), 14–34 (1989)
J. Numer. Anal. 44(2), 655–676 (2006)
7. Chand, A.K.B., Navascués, M.A.: Generalized Hermite fractal interpolation. Rev. R. Acad.
Cienc. Zaragoza 64(2), 107–120 (2009)
8. Chand, A.K.B., Viswanathan, P.: Cubic Hermite and cubic spline fractal interpolation functions.
AIP Conf. Proc. 1479, 1467–1470 (2012)
9. Chand, A.K.B., Viswanathan, P.: A constructive approach to cubic Hermite fractal interpolation
function and its constrained aspects. BIT Numer. Math. 53(4), 841–865 (2013)
10. Chand, A.K.B., Vijender, N., Agarwal, R.P.: Rational iterated function system for posi-
tive/monotonic shape preservation. Adv. Differ. Eqn. 2014(30), 1–19 (2014)
11. Chand, A.K.B., Vijender, N., Navascués, M.A.: Shape preservation of scientific data through
rational fractal splines. Calcolo 51(2), 329–362 (2014)
12. Dalla, L., Drakapoulos, V.: On the parameter identification problem in the plane and polar
fractal interpolation functions. J. Approx. Theory 101, 289–302 (1999)
13. Duan, Q., Xu, G., Liu, A., Wang, X., Cheng, F.: Constrained interpolation using rational cubic
spline with linear denominators. Korean J. Comput. Appl. Math. 6, 203–215 (1999)
14. Duan, Q., Wang, L., Twizell, E.H.: A new C 2 rational interpolation based on function values
and constrained control of the interpolant curves. J. Appl. Math. Comput. 61, 311–322 (2005)
15. Gregory, J.A., Delbourgo, R.: Shape preserving piecewise rational interpolation. SIAM J. Stat.
Comput. 6(4), 967–976 (1985)
16. Hussian, M.Z., Sarfraz, M., Hussain, M.: Scientific data visualization with shape preserving
C 1 -rational cubic interpolation. Eur. J. Pure Appl. Math. 3(2), 194–212 (2010)
17. Sarfraz, M., Hussain, M.Z.: Data visualization using rational spline interpolation. J. Comp.
Appl. Math 189, 513–525 (2006)
18. Tahira, S.S., Sarfraz, M., Hussain., M.Z.: Shape preserving constrained data visualization using
rational functions. J. Prime Res. Math. 7, 35–51 (2011)
19. Viswanathan, P., Chand, A.K.B.: A C 1 -rational cubic fractal interpolation function: conver-
gence and associated parameter identification problem. Acta. Appl. Math. 136(1), 19–41 (2015)
tal shapes: Bounds on the scale factors. J. Math. Anal. Appl. 419(2), 804–817 (2014)
Transverse Vibrations of Nonhomogeneous
Rectangular Kirchhoff Plates of Variable
Thickness
Roshan Lal and Renu Saini
Abstract In this paper the free transverse vibrations of nonhomogeneous rectangular

Kirchhoff plates of linearly varying thickness along one direction have been stud-
ied using generalized differential quadrature method. The nonhomogeneity of the
plate material is assumed to arise due to linear variations in Young’s modulus and
density of the plate material with the in-plane coordinates. Numerical results have
been computed when the plate is clamped at all the four edges. The effect of various
plate parameters such as nonhomogeneity parameters, density parameters, thickness
parameter, and aspect ratio on the frequencies has been investigated for the first
two modes of vibration. Three-dimensional mode shapes for a specified plate have
been plotted. A comparison of results with those available in the literature has been
presented.
Keywords Rectangular · Nonhomogeneity · GDQ · Variable thickness
1 Introduction
Nowadays, the study of nonhomogeneous materials is of great interest due to their

wide applications in various fields of engineering such as aerospace, mechani-
cal, nuclear, marine, structural engineering, etc. Plywood, timber, fiber-reinforced
plastic, etc., are good examples of nonhomogeneous materials. The mechanical
properties of such materials display spatial variations. In this regard, some of the
high-strength light-weight nonhomogeneous/composite materials fabricated by mix-
ing two or more materials, e.g., carbon fiber and epoxies are being used for aerospace
applications and in high performance sporting goods. The nonhomogeneity of a struc-
ture is characterized by a number of factors governing its structural features. For
plate-type structure these features are geometrical imperfections, including foreign
R. Lal · R. Saini (B)

e-mail: renusaini189@gmail.com
R. Lal
DOI 10.1007/978-81-322-2485-3_50
610 R. Lal and R. Saini
materials and reinforcements of various types [1–3]. Sometimes, plate-type structural

elements have to work under high temperature environment which causes nonhomo-
geneity in the material, particularly in aerospace industry, modern missile technology,
and microelectronics. Further, these plates with appropriate thickness variation have
significantly greater efficiency for vibration compared to plates of uniform thickness
and also provide the advantage of material saving and hence cost requirement. Thus
their design requires an accurate analysis for their vibration characteristic. In partic-
ular, rectangular plates are key components in ocean structures and in the aerospace
industry. In this regard, numerous studies dealing with the vibration of rectangu-
lar plates of nonuniform thickness have been carried out and reported in references
[4–7], to mention a few. Various models for nonhomogeneity of the plate material
have been proposed in the literature [8–12]. In these papers, it is considered that
nonhomogeneity of the plate material arises due to change in only one space variable
except in references [11, 12]. The present study analyzes the effect of nonhomo-
geneity on the free transverse vibration of thin rectangular plates of varying thick-
ness employing generalized differential quadrature (GDQ) method when the plate is
clamped at all the four edges. Nonhomogeneity of the plate material is assumed to
arise due to linear variation in Young’s modulus and density of the plate material with
the in-plane coordinates while the thickness is varying linearly along one direction.
The effect of various parameters on the natural frequencies has been investigated for
the first two modes of vibration. A comparison of results has been presented.
Configuration of a nonhomogeneous isotropic rectangular plate of length a, breadth

b, thickness h, and density ρ is shown in Fig. 1. The x- and y-axes are taken along the
edges of the plate, the axis of z is perpendicular to the x y-plane. The middle surface
being z = 0 and origin is at the one of the corners of the plate. The differential
equation governing the transverse vibration of such plates [12], is given by
2 2
∂ 2 D ∂ 2w ∂ D ∂ w ∂ 2 D ∂ 2w ∂ 2w
∇ 2 (D∇ 2 w)−(1−ν) − 2 + −ρh 2 = 0 (1)
∂x ∂y
2 2 ∂ x∂ y ∂ x∂ y ∂y ∂x
2 2 ∂t
where ∇ 2 is Laplacian operator, D = Eh 3 /12(1 − ν 2 ) is the flexural rigidity, w(x, y, t) is

the transverse displacement E is the Young’s modulus, ν is the Poisson ratio, and t is the
time.
For a harmonic solution, the displacement w is assumed to be
w(x, y, t) = w̄(x, y, t)eiωt (2)
where ω is the circular frequency in radians.

Transverse Vibrations of Nonhomogeneous Rectangular Kirchhoff Plates … 611
Fig. 1 Geometry of the

rectangular plate
b CCCCC
H
h0
x
a
Using Eq. (2), Eq. (1) reduces to

∂ 4 w̄ ∂ 4 w̄ ∂ 4 w̄ ∂ D ∂ 3 w̄ ∂ 3 w̄ ∂ D ∂ 3 w̄ ∂ 3 w̄
D + 4 + 2D 2 2 + 2 + +2 + 2
∂x4 ∂y ∂x ∂y ∂x ∂x3 ∂ x∂ y 2 ∂ y ∂ y3 ∂x ∂y

2
∂ w̄ ∂ D 2 2
∂ D 2
∂ w̄ 2
∂ D 2
∂ D
+ +ν 2 + 2 ν 2 +
∂x2 ∂x2 ∂y ∂y ∂x ∂ y2
∂ 2 D ∂ 2 W̄
+ 2(1 − ν) − ρhω2 w̄ = 0 (3)
∂ x∂ y ∂ x∂ y
Taking the following nondimensional variables X = x/a, Y = y/b, H = h/a,

W = w̄/a and assuming that Young’s modulus and density of the plate material vary
with the in-plane coordinates by the relations
E(X, Y ) = E 0 (1 + α1 X + α2 Y )
ρ(X, Y ) = ρ0 (1 + β1 X + β2 Y ) (4)
and thickness of the plate varies linearly in X -direction, given by
H (X ) = h 0 (1 + γ X ) (5)
where E 0 , ρ0 and h 0 are the Young’s modulus, density, and thickness of the plate
at X = 0, Y = 0, γ is the thickness parameter, α1 , α2 are the nonhomogeneity
parameters, and β1 , β2 are the density parameters, respectively.
Equation (3) now reduces to

∂4W 2 ∂ W
4
4∂ W
4 ∂3W 2 ∂ W
3
A0 + 2λ +λ + A1 +λ
∂ X4 ∂ X 2 ∂Y 2 ∂Y 4 ∂ X3 ∂ X ∂Y 2
2
∂ W
3 ∂ W
3 ∂ W ∂ W
2 ∂ W
2
+A2 λ4 + λ2 + A3 + λ2 + A4 − A5 W = 0
∂Y 3 ∂ X ∂Y
2 ∂X 2 ∂Y 2 ∂ X ∂Y
(6)
where A0 = (1 + α1 X + α2 Y )(1 + γ X )2
A1 = (6γ (1 + α1 X + α2 Y ) + 2α1 (1 + γ X ))(1 + γ X )
A2 = 2α2 (1 + γ X )2 , A3 = 6γ 2 (1 + α1 X + α2 Y ) + 6α1 γ (1 + γ X )
A4 = 6(1 − ν)λ2 γ α2 (1 + γ X ), A5 = Ω 2 (1 + β1 X + β2 Y )
λ = a/b, Ω 2 = 12ρ(1 − ν 2 )ω2 /a E 0 h 20
Equation (6) is a fourth-order partial differential equation of variable coefficients
with respect to X and Y . The clamped boundary condition (i.e., CCCC) has been
considered in the present paper and the relation that should satisfied at clamped edges
given by:
dW dW
W = = 0, W = = 0, at X = 0 or X = 1, and Y = 0 or Y = 1, respectively.
dX dY
3 Method of Solution: Generalized Differential Quadrature

(GDQ) Method
The computational domain of a rectangular plate is 0 ≤ X ≤ 1, 0 ≤ Y ≤ 1. If N

and M are the number of grid points in X and Y directions, respectively, then the
total number of function values in the whole domain is N × M. According to GDQ
method [13], the nth and mth order derivatives of W (X, Y ) with respect to X , Y and
its mixed derivative with respect to X and Y are given by
∂ n W (X, Y ) (n)
N
= ai j W (X l , Y j )
∂ Xn
l=1
∂ m W (X, Y ) (m)
M
= bi j W (X i , Yl ) (7)
∂Y m
l=1
∂ m+n W (X, Y )
N
M
(n) (m)
= ail1 b jl2 W (X l1 , Yl2 )
∂ X n ∂Y m
l1 =1 l2 =1
i = 1, 2, . . . , N ; j = 1, 2, . . . , M; n = 1, 2, . . . , N − 1; m = 1, 2, . . . , M − 1
(n) (m)
where ai j and bi j are the weighting coefficients associated with nth and mth order
derivatives with respect to X and Y respectively. The weighting coefficient of first-
order derivative are determined as
⎧
⎨ P (1) (X i )
, j = i
ai(1) = (X i −X j )P (1) (X j )
⎩ − N
j (1)
(8)
j=1, j=i ai j , j = i,
for i, j = 1, 2, . . . , N

where P (1) (X i ) = Nj=1, j=i (X i − X j )
Similarly, for the second and higher order derivatives the recurrence relationships
are obtained as follows:
⎧ (n−1)

⎨ n a (n−1 a (1) − ai j
(n)
ai j = ii ij (X i −X j ) , j = i (9)
⎩ N (n)
− j=1, j=i ai j , j = i,
for i, j = 1, 2, . . . , N , n = 2, 3, . . . , N − 1.
(m)
The corresponding coefficients bi j associated with derivatives with respect to Y
required can be similarly determined [13].
Discretizing Eq. (6) at the internal grid points (X i , Y j ), with 3 ≤ i ≤ N − 2 and
3 ≤ j ≤ M − 2, it reduces to
⎛ ⎞

N
(4)

N
M
(2) (2)

M
(m)
A0 (i, j) ⎝ ail Wl, j + 2λ2 ail1 b jl2 Wl1 ,l2 + λ4 bi j Wi,l ⎠
l=1 l1 =1 l2 =1 l=1
⎛ ⎞

N
(3)

N
M
(1) (2)
+ A1 (i, j) ⎝ ail Wl, j + λ2 ail1 b jl2 Wl1 ,l2 ⎠
l=1 l1 =1 l2 =1
⎛ ⎞

M
(3)

M
(2) (1)
+ A2 (i, j) ⎝λ4 bi j Wi,l + λ2 ail1 b jl2 Wl1 ,l2 ⎠ (10)
l=1 l2 =1
N
(2)

M
(2)

N
M
(1) (1)
+ A3 (i, j) ail Wl, j + νλ 2
bi j Wi,l + A4 (i, j) ail1 b jl2 Wl1 ,l2
l=1 l=1 l1 =1 l2 =1
− A5 (i, j)Wi, j = 0
Similarly, the boundary conditions can be non-dimensionalized and then discretized

using GDQ. Here, the grid points chosen for collocation are the zeroes of shifted
Chebyshev polynomials and are given by

1 2i − 1 π 1 2j − 1 π
X i+1 = 1 + cos , Yi+1 = 1 + cos (11)
2 N −2 2 2 M −2 2
i = 1, 2, . . . , N − 2 j = 1, 2, . . . , M − 2
4 Numerical Results and Discussions
Equation (10) together with boundary condition form a eigenvalue problem [13],
which has been solve numerically using GDQ. The values of various plate parameters
are taken as follows: Nonhomogeneity parameters α1 , α2 = (−0.5(0.1)0.5), density
parameters β1 , β2 = (−0.5(0.1)0.5), thickness parameter γ = (−0.5(0.1)0.5),
aspect ratio a/b = (0.25(0.25)2.0) and Poisson ratio ν = 0.3.
The values of grid points N and M have been fixed as 15, since further increase in
the values of grid points, frequency parameter remain constant at the fourth place of
decimals. The convergence of frequency parameter Ω for a particular set α1 = α2 =
β1 = β2 = γ = 0.5, a/b = 1 is shown in Table 1.
Figure 2 shows the behavior of frequency parameter Ω with nonhomogeneity
parameter α1 for α2 = ±0.5, γ = 0.5, β1 = ±0.5, β2 = 0.5 and a/b = 1 for the
first two modes of vibration. It is observed that the value of frequency parameter
Ω increases with increasing values of nonhomogeneity parameter α1 . Further, it is
increases with increasing values of α2 while it decreases with increasing values of
β1 keeping all other parameters fixed.
The effect of thickness parameter on the frequency parameter Ω for α1 =
0.5, α2 = ±0.5, β1 = β2 = 0.5 and a/b = 1 for the first two mode of vibration has
Table 1 Convergence study for frequency parameter Ω for the first three modes of vibration
No. of terms I mode II mode III mode
N = M =8 44.5957 89.7378 90.5676
N = M = 10 44.5942 90.5409 90.9843
N = M = 12 44.5940 90.5478 90.9795
N = M = 13 44.5940 90.5477 90.9795
N = M = 14 44.5940 90.5477 90.9795
N = M = 15 44.5940 90.5477 90.9795
Fig. 2 Frequency parameter 110

Ω for CCCC square plate for 100
β2 = γ = 0.5. I mode:—;II 90
mode: ....; , α2 = 0.5,
80
β1 = 0.5; , α2 = 0.5
β1 = −0.5;, α2 = −0.5, 70
β1 = 0.5; , α2 = −0.5, Ω
60
β1 = −0.5 50
40
30
20
-0.5 -0.3 -0.1 0.1 0.3 0.5
α1

α2 = β2 = 0.5, I mode:—;II 90
mode: ....; α1 = 0.5, 80
β1 = 0.5; α1 = 0.5 70
β1 = −0.5;, α1 = −0.5, Ω 60
β1 = 0.5; , α1 = −0.5, 50
β1 = −0.5 40
30
20
-0.5 -0.3 -0.1 0.1 0.3 0.5
γ1

α2 = 0.5, γ = 0.5. I
mode:—;II mode: ....; 120
α1 = 0.5, β2 = 0.5;
100
α1 = 0.5 β2 = −0.5;,
α1 = −0.5, β2 = 0.5; , Ω 80
α1 = −0.5, β2 = −0.5
60
40
20
-0.5 -0.3 -0.1 0.1 0.3 0.5
β1
been shown in Fig. 3. It is seen that the frequency parameter Ω increases with the
increasing values of whatever be the value of other plate parameters.
The value of frequency parameter Ω increases with increasing values of α1 , while
it decreases with increasing values of β1 . The rate of increase of frequency parameter
Ω with γ is higher for second mode than that of first mode.
Figure 4 depicts the behavior of the frequency parameter Ω with the density
parameter β1 for α1 = ±0.5, β2 = ±0.5, α2 = 0.5, γ = 0.5 and a/b = 1 for the
first two modes of vibration. It is found that the frequency parameter Ω decreases with
increasing values of density parameter β1 . The value of Ω increases with increasing
values of α1 . The rate of decrease of Ω with β1 increases with increasing values of
α1 , while it is decreases with increasing values of β1 . This rate is higher in the second
mode compared to the first mode.
Figure 5 illustrates the behavior of frequency parameter Ω with increasing values
of aspect ratio a/b for α1 = β1 = ±0.5, α2 = β2 = 0.5 and γ = 0.5 for the first
two modes of vibration. It is clear that the frequency parameter Ω increases with
increasing values of a/b. The rate of increase of Ω with a/b increases with increase
200
180
140
Ω 100
60
20
0.25 0.75 1.25 1.75 2.0
a/b
Fig. 5 Frequency parameter Ω for CCCC square plate for α2 = β2 = γ = 0.5. I mode:—;II
mode: ....; , α1 = 0.5, β1 = 0.5; α1 = 0.5 β1 = −0.5;, α1 = −0.5, β1 = 0.5; , α1 = −0.5,
β1 = −0.5
0.2
0.2
0.15
0.1
0.1
0
0.05
-0.1
0
1 -0.2
0.8 1
1
0.6 1
0.4 0.5 0.5
0.2 0.5
0 0
0 0
I Mode II Mode
Fig. 6 Three dimensional mode shapes of CCCC square plate; for α1 = α2 = β1 = β2 = γ = 0.5
in the number of modes. This rate of increase is much higher for a/b > 1 compared
to a/b < 1.
Three-dimensional mode shapes for a specified plate have been shown in Fig. 6
for the first three modes of vibration.
A comparison of frequency parameter for homogeneous square plate with the

results available in the literature and obtained by other methods is presented in
Table 2. A close agreement of results is obtained.
Table 2 Comparison of frequency parameter Ω for homogeneous (α1 = α2 = β1 = β2 = 0)

plate
References γ I mode II mode III mode
[14] 0.0 35.992 73.413 73.413
[15] 0.0 35.986 73.395 73.395
[16] 0.0 35.986 73.395 73.395
[17] 0.0 35.99 73.41 −
[18] 0.0 35.99 73.419 73.419
[19] 0.0 35.45 72.03 72.03
[20] 0.0 35.989 73.399 73.399
Present 0.0 35.9852 73.3938 73.3938
[16] 0.2 39.5097 80.5201 80.5857
[20] 0.2 39.513 80.525 80.591
Present 0.2 39.5094 80.5184 80.5842
[15] 0.4 42.9088 87.2835 87.5259
[20] 0.4 42.913 87.901 87.901
Present 0.4 42.9084 87.2822 87.5233
[20] 0.5 44.574 90.564 90.926
Present 0.5 44.5694 90.5543 90.9158
[20] −0.4 28.377 57.530 57.890
Present −0.4 28.3737 57.5225 57.8837
5 Conclusion
The effect of nonhomogeneity and thickness variation on the vibration characteristics

of isotropic rectangular plates has been studied on the basis of classical plate theory
using generalized differential quadrature (GDQ) method. The thickness of the plate is
taken linear along one direction. The nonhomogeneity of the plate material is assumed
to arise due to the linear variations in Young’s modulus and density of the plate
material with both the in-plane coordinates. It is found that the values of frequency
parameter Ω increases with increasing values of nonhomogeneity parameters α1
and α2 aspect ratio, while it decreases with increasing values of density parameters
β1 and β2 keeping other plate parameters fixed. The frequency parameter Ω also
increases with increasing values of thickness parameter γ . The present analysis will
be helpful to design engineers dealing with nonhomogeneity in obtaining the desired
frequency by taking one or more plate parameters considered here.
Acknowledgments The authors wish to express their sincere thanks to the learned reviewers for
their constructive comments in improving the paper. One of the authors, Renu Saini, is thankful
to the Ministry of Human Resources and Development (MHRD), India for the financial support to
carry out this research work.
References
1. Jones, R.M.: Mechanics of Composite Materials, 2nd edn. Taylor and Francis, Philadelphia
(1999)
2. Hudramovich, V.S.: Features of nonlinear deformation and critical states of shell systems with
geometrical imperfections. Int. Appl. Mech. 43(12), 1323–1355 (2006)
3. Hudramovich, V.S.: Contact mechanics of shell structures under local loading. Int. Appl. Mech.
45(7), 708–729 (2009)
4. Grigorenko, A.Y., Tregubenko, T.V.: Numerical and experimental analysis of natural vibrations
of rectangular plates with variable thickness. Int. Appl. Mech. 36(2), 131–134 (2000)
5. Kang, S.W., Kim, S.H.: Vibration analysis of simply supported rectangular plates with unidi-
rectionally, arbitrarily varying thickness. J. Sound Vib. 312, 551–562 (2008)
6. Civalek, O.: Fundamental frequency of isotropic and orthotropic rectangular plates with linearly
varying thickness by discrete singular convolution method. Appl. Math. Mod. 33, 3825–3835
(2009)
7. Manna, M.C.: Free vibration of tapered isotropic rectangular plates. J. Vib. Control 8(1), 76–91
(2011)
8. Lal, R., Dhanpati: Transverse vibrations of non-homogeneous orthotropic rectangular plates
of variable thickness: a spline technique. J. Sound Vib. 306, 203–214 (2007)
9. Lal, R., Dhanpati: Quintic splines in the study of buckling and vibration of non-homogeneous
orthotropic rectangular plates with variable thickness. Int. J. Appl. Math. Mech. 3, 18–35 (2007)
10. Lal, R., Saini, R.: Buckling and vibration of non-homogeneous rectangular plates subjected to
linearly varying in-plane force. Shock Vib. 20, 879–894 (2013)
11. Lal, R., Kumar, Y., Gupta, U.S.: Transverse vibrations of nonhomogeneous rectangular plates
of uniform thickness using boundary characteristic orthogonal polynomials. Int. J. Appl. Math.
Mech. 6, 93–109 (2010)
12. Saini, R., Lal, R.: Vibration analysis of non-homogeneous rectangular plates using GDQ. Int.
J. Adv. Sci. Tech. ISSN 2348–5426, 166–172 (2014)
13. Shu, C.: Differential Quadrature and Its Application in Engineering. Springer-Verlag, London
(2000)
14. Leissa, A.W.: Free vibrations of rectangular plates. J. Sound Vib. 31, 257–293 (1973)
15. Bhat, R.B.: Natural frequencies of rectangular plates using orthogonal polynomials in the
Rayleigh-Ritz method. J. Sound Vib. 102, 493–499 (1985)
16. Bhat, R.B., Laura, P.A.A., Gutierrez, R.G., Cortinez, V.H.: Numerical experiment on the deter-
mination of natural frequencies of transverse vibrations of rectangular plates of non-uniform
thickness. J. Sound Vib. 139, 205–219 (1990)
17. Liew, K.M., Lam, K.Y., Chow, S.T.: Free vibration analysis of rectangular plate using orthog-
onal plate function. Compost. Struct. 34, 75–89 (1990)
18. Bardell, N.S.: Free vibration analysis of a flat plate using the hierarchical finite element method.
J. Sound Vib. 151, 263–289 (1991)
19. Kerboua, Y., Lakis, A.A., Thomas, M., Marcouiller, L.: Hybrid method for vibration analysis
of rectangular plates. Nucl. Eng. Des. 237, 791–801 (2007)
20. Singh, B., Saxena, V.: Transverse vibration of a rectangular plate with bidirectional thickness
variation. J. Sound Vib. 198, 51–65 (1996)
Compartmental Disease Models
with Heterogeneous Populations: A Survey
R.N. Mohapatra, Donald Porchia and Zhisheng Shuai
Abstract Compartmental models for infectious disease transmission among hetero-

geneous host populations are surveyed. Mathematical methodologies for analyzing
heterogeneous disease models are reviewed. Specifically, three methods are pro-
vided to establish the global stability of the endemic equilibrium for a multigroup
SIS model. The survey is concluded with several open problems.
Keywords Infectious disease · Mathematical model · Global stability · Lyapunov

function · Comparison metho
1 Introduction
Heterogeneity is one of the most important characteristics in modeling infectious

diseases due to the variability of both the pathogen and the host population. Ignoring
these heterogeneous structures in the disease transmission can fail in prediction of
disease outbreak and misevaluate disease control and intervention strategies. For
example, in [1] Hethcote and Yorke point out that 60 % of gonorrhea infections in
the United States was caused by less than 2 % of the human population, showing
that heterogeneity in individual behavior significantly affects the disease dynamics.
Superspreading and the effect of individual variation have been highlighted in recent
disease outbreaks such as the 2002–2004 severe acute respiratory syndrome (SARS)
outbreak [2, 3]. Heterogeneity in the host population can be the result from factors
such as age, gender, and genetic heterogeneity of host individuals, and geographical
locations such as individuals living in the same city, community, or country. It may
R.N. Mohapatra · D. Porchia · Z. Shuai (B)

Department of Mathematics, University of Central Florida, Orlando, FL 32816, USA
e-mail: Zhisheng.Shuai@ucf.edu
R.N. Mohapatra
e-mail: Ram.Mohapatra@ucf.edu
D. Porchia
e-mail: Donald.Porchia@ucf.edu

DOI 10.1007/978-81-322-2485-3_51
620 R.N. Mohapatra et al.
also be from disease-dependent host parameters such as susceptibility to the disease,

disease transmission rate, or recovery rate.
Multigroup models have been widely used in the literature to capture and model
the effect of host heterogeneity on disease transmission. The pioneering multigroup
model of SIS type has been proposed by Lajmanovich and Yorke [4] in 1976, describ-
ing the transmission of gonorrhea among heterogeneous host populations that have
different levels of sexual activities. They successfully establish the global dynamics
of the multigroup SIS model by constructing suitable Lyapunov functions. Specif-
ically, the global stability of the endemic equilibrium is proved by analyzing the
intersection of invariant sets in which Lyapunov functions do not vary along solu-
tions of the model. Their results in [4] have been widely cited and used to establish the
global dynamics for other multigroup models; however, the technique of using mul-
tiple Lyapunov functions to show the global stability, which is reviewed in Sect. 3.3,
has seen less application in the literature.
Following the pioneering work [4], studies on various multigroup models have
been conducted; see, e.g., [5–15] and the references therein. However, due to the large
scale and complexity of multigroup models, progress in the mathematical analysis of
their global dynamics has been slow. In particular, the questions of uniqueness and
global stability of the endemic equilibrium, when the basic reproduction number R0
is greater than 1, remain largely open. Hethcote [9] established the global stability of
the endemic equilibrium for multigroup SIR models without vital dynamics. Beretta
and Capasso [6] derived sufficient conditions for global stability with constant group
population sizes. Thieme [14] proved global stability of the endemic equilibrium
of multigroup SEIRS models under certain restrictions. For a class of SIRS models
with constant group sizes, Lin and So [12] proved that the endemic equilibrium is
globally asymptotically stable if the cross-group contact rates are small or if the
recovery rates in each group are small. On the other hand, results in the opposite
direction also exist in the literature. For a multigroup SIR model with proportionate
incidence, uniqueness of endemic equilibria may not hold [11, 15] when R0 is greater
than 1.
In 2006, Guo et al. [16] established the complete global dynamics for a class of
multigroup SIR models with varying group sizes. In particular, it was proved that
when R0 is greater than 1, the endemic equilibrium of the model is unique and
globally asymptotically stable. This completely resolved the open question on the
uniqueness and global stability of the endemic equilibrium. The proof relies on the
use of a class of global Lyapunov functions and Kirchhoff’s Matrix Tree Theorem
(see Theorem A in Apendix). Lyapunov functions of this type have previously been
used in the ecological literature (e.g., see [17–19]) and was rediscovered for several
classes of epidemic models (see [20–22]). The key to the analysis was in using
graph theory to provide a complete description of the complicated patterns exhibited
in the derivative of the Lyapunov function. This graph-theoretic approach to the
construction of Lyapunov functions has been further developed in the series of papers
[23–25], and has now become a standard tool for analyzing heterogeneous disease
models. For example, the graph-theoretic approach has been applied to heterogeneous
models incorporating nonlinear incidence functions in [24, 26, 27], discrete and
Compartmental Disease Models with Heterogeneous Populations: A Survey 621
distributed time delays for disease latency in [28–31], disease progression/relapse

in [32–34], vaccine strategies in [35–37], and random fluctuations in [38–40]. The
approach has also been applied to heterogeneous models for specific diseases, such
as, cholera in [41, 42] and dengue in [43, 44].
Besides Lyapunov function methods, other methods such as comparison argu-
ments [45] and the theory of monotone dynamical systems [46] have also been
applied to establish global dynamics for heterogeneous models. In Sect. 2, we revisit
the classic multigroup SIR model [4], and in Sect. 3 review mathematical methods
that can be applied to the model analysis. Specifically, we provide three methods
to prove the global stability of the endemic equilibrium when R0 > 1: the original
proof by Lajmanovich and Yorke in [4], the graph-theoretic approach as shown in
[25, Sec. 6], and the comparison method in [45]. We conclude with discussions and
several open problems in Sect. 4.
2 Revisit the Multigroup SIS Model
In this section, we revisit the classical multigroup SIS model [4] and summarize by
results in the model analysis.
The total host population is divided into n host groups; the population size for
each group k is denoted as Nk . The host group Nk is further categorized into two
compartments: the compartment containing susceptible individuals and the compart-
ment containing infectious individuals, with population sizes Sk and Ik , respectively.
Thus, Nk = Sk + Ik . The new infections in host group k are due to within-group
transmission βkk Sk Ik and all between-group transmissions (cross infection) βk j Sk I j
for j = k. Thus the multigroup SIS model, which was first proposed in [4] for the
transmission of gonorrhea in a heterogeneous host population, takes the following
form:

n
Sk = Λk − βk j Sk I j − μk Sk + γk Ik ,
j=1
(2.1)

n
Ik = βk j Sk I j − (μk + γk )Ik , k = 1, 2, . . . , n.
j=1
Here Λk represents the new input (e.g., birth, migration) into group k, μk represents
the mortality rate in group k, γk represents the recovery rate of infectious individuals
in group k, and βk j represents the transmission coefficients between susceptible
individuals in group k and infectious individuals in group j. Parameters γk , βk j are
assumed to be nonnegative and Λk , μk positive. In addition, the contact matrix [βk j ]
is assumed to be irreducible; biologically, this is the same as assuming that any two
groups have a direct or indirect transmission route.
Adding the two equations in (2.1) gives Nk = Λk − μk Nk , and thus limt→∞
Nk (t) = Λk /μk . By the theory of asymptotically autonomous systems [47, 48], it is
sufficient to study the long time behaviors of the reduced system

n Λ
Ik =
k
βk j − Ik I j − (μk + γk )Ik , k = 1, 2, . . . , n. (2.2)
μk
j=1
The multigroup model (2.1) (or the reduced model (2.2)) also includes as a special
case other models in the literature. For example, a network-based SIS model in [49,
50] takes the following form:
n
j P( j)ρ j
ρk
j=1
= −ρk + λk(1 − ρk ) n , k = 1, 2, . . . , n, (2.3)
j=1 j P( j)
where ρk represents the relative density of infected nodes with degree k (i.e., the
probability that a node with k links is infected), P(k) ≥ 0 represents the density of
nodes with degree k, λ represents the effective disease spreading rate, and n is the
maximum degree of all nodes. Model (2.2) is equivalent to (2.3) if we set
λk j P( j)
βk j = n , Λk = μk , γk = 1 − μk . (2.4)
l=1 l P(l)
Now we start to analyze model (2.1) within a feasible region.

Λk
Lemma 1 The feasible region Γ = (S1 , I1 , . . . , Sn , In ) ∈ R2n
+ | Sk + Ik = μk
is positively invariant with respect to (2.1).
Model (2.1) always admits a disease-free equilibrium P0 = (S10 , 0, . . . , Sn0 , 0) in

Γ with Sk0 = Λ ∗ ∗ ∗ ∗ ∗
μk . There might be an endemic equilibrium P = (S1 , I1 , . . . , Sn , In )
k
with Sk∗ , Ik∗ > 0, which lies in int(Γ ), the interior of Γ .

Following the next generation matrix approach [51, 52], define two n ×n matrices
F = [βk j Sk0 ] and V = diag{μ1 + γ1 , . . . , μn + γn }, representing the disease inci-
dence matrix and disease transfer matrix, respectively. Then the basic reproduction
number is defined as the spectral radius of the next generation matrix FV−1 ; that is

βk j Sk0
R0 = ρ F V −1 = ρ . (2.5)
μj + γj
For the network SIS model (2.3), using (2.4), the disease-free equilibrium and
the basic reproduction number can be evaluated as P0 = (1, 0, . . . , 1, 0) and
R0 = ρ([βk j ]) = n λl P(l) ρ([k j P( j)]), respectively. Since the matrix [k j P( j)]
l=1 n 2
has rank 1, the spectral radius can be calculated as l=1 l P(l). Therefore, the basic
n
reproduction number for (2.3) becomes R0 = λ <k
2>
<k> with <k > =
j j
l=1 l P(l),
agreeing with the threshold value in [49, 50].
The basic reproduction number R0 completely determines the disease dynamics
of (2.1) as shown in the following sharp threshold result.
Theorem 1 Assume that contact matrix [βk j ] is irreducible.
(a) The disease-free equilibrium P0 is globally asymptotically stable in Γ if R0 ≤ 1
and becomes unstable if R0 > 1.
(b) If R0 > 1, then model (2.1) is uniformly persistent, and there exists a unique
endemic equilibrium P ∗ that is globally asymptotically stable in int(Γ ).
3 Stability of the Multigroup SIS Model
In the following we provide the proof for the sharp threshold result (Theorem 1) in
Sect. 2. In particular, three different methods are provided for the global stability of
the endemic equilibrium.
3.1 Global Stability of the Disease-Free Equilibrium
In this subsection, we prove the global stability of the disease-free equilibrium when
R0 ≤ 1. Let x = (I1 , . . . , In )T denote the disease compartment in (2.1) and F, V
defined as above. It follows from Lemma 1 that x ≤ (F −V )x. Following the matrix-
theoretic method in [25, Sect. 2], a Lyapunov function L can be constructed for (2.1);
that is, L = wT V −1 x, where wT is a left eigenvector of the nonnegative matrix V −1 F
corresponding to the eigenvalue ρ(V −1 F) = ρ(F V −1 ) = R0 . Differentiating L
along (2.1) gives
L = wT V −1 x ≤ wT V −1 (F − V )x = (R0 − 1)wT x. (3.1)
Hence, L ≤ 0 if R0 ≤ 1. Furthermore, L = 0 implies that Sk = Sk0 for all k. By

the first equation of (2.1), it follows that Ik = 0 for all k. Hence the largest invariant
set where L = 0 is the singleton {P0 }. By LaSalle’s invariance principle [53], P0 is
globally asymptotically stable in Γ when R0 ≤ 1.
If R0 > 1, then by (3.1), L = (R0 − 1)wT x > 0 when Sk = Sk0 and Ik > 0,
for all k. Continuity thus shows that L > 0 in a neighborhood of P0 . Solutions in
the positive cone sufficiently close to P0 move away from P0 , implying that P0 is
unstable. Using a uniform persistence result from [54] and an argument as in the
proof of Proposition 3.3 of [55], it can be shown that, when R0 > 1, instability of P0
implies uniform persistence of (2.1). Uniform persistence and the positive invariance
of the compact set Γ imply the existence of an endemic equilibrium of (2.1) (see
Theorem D.3 in [56] or Theorem 2.8.6 in [57]).
3.2 Global Stability of the Endemic Equilibrium: Lyapunov

Function
In this subsection we apply the graph-theoretic method to construct a Lyapunov

function to prove the global stability of the endemic equilibrium P ∗ when R0 > 1.
Let Dk = Ik − Ik∗ − Ik∗ ln IIk∗ . Differentiating and using the equilibrium equation
k
gives
⎛ ⎞
∗ n n
I − I k ⎝ Λ Λ I
Dk = − Ik∗ I ∗j ∗ ⎠
k k k k
βk j − Ik I j − βk j
Ik μk μk Ik
j=1 j=1
∗
Ik∗ I j
n n
(Ik − Ik ) 2
Λk ∗ ∗ Ik Ij
=− βk j I j + βk j − Ik I j 1 − ∗ + ∗ −
Ik μk Ik Ij Ik I ∗j
j=1 j=1

Ik∗ I j
n
Λk ∗ ∗ Ik Ij
≤ βk j − Ik I j 1 − ∗ + ∗ −
μk Ik Ij Ik I ∗j
j=1
n
n
Λk ∗ ∗ Ij Ij Ik Ik
≤ βk j − Ik I j − ln − + ln := ak j G k j ,
μk I ∗j I ∗j Ik∗ Ik∗
j=1 j=1

Λk Ij Ij
with ak j = βk j μk − Ik∗ I ∗j ≥ 0 and G k j = I ∗j − ln I ∗j − Ik
Ik∗ + ln Ik
Ik∗ . The second
inequality follows from the fact that 1−x ≤ − ln x for all x > 0. Let Hk = IIk∗ −ln IIk∗ ,
k k
then G k j = H j − Hk . A weighted digraph G can be constructed to associate with
the weight matrix A = [ak j ]; see Appendix for more details. Notice that along any
directed cycle C of (G , A),

Gr s = (Hs − Hr ) = 0.
(s,r )∈E (C ) (s,r )∈E (C )
Here E (C ) denotes the arc set of C . Since all assumptions of Theorem B in

Appendix hold, let ci be as given in Theorem A in Appendix, then by Theorem B,
D = nk=1 ck Dk is a Lyapunov function for (2.2). Using this Lyapunov function,
the irreducibility of [βi j ], and LaSalle’s Invariance Principle, it can be proved that
{(I1∗ , . . . , In∗ )} is the largest invariant set for (2.2), and thus if R0 > 1, then the
positive equilibrium (I1∗ , . . . , In∗ ) is globally asymptotically stable for (2.2). As a
consequence, P ∗ is globally asymptotically stable and thus unique in int(Γ ) for
(2.1).
3.3 Global Stability of the Endemic Equilibrium: Multiple

Lyapunov Functions
In [4] Lajmanovich and Yorke construct two Lyapunov functions for (2.1), and are
able to prove the global stability of the endemic equilibrium by analyzing the inter-
section of invariant sets where Lyapunov functions do not vary along its solutions.
We summarize their proof in this subsection.
Let D1 = max{M − 1, 0} and D2 = max{1 − m, 0}, where M = maxk IIk∗ and
k
m = mink { IIk∗ }. Without loss of generality, assume M = II1∗ and m = I2
I2∗ in some
k 1
time interval I . Thus, if M ≥ 1 in I , then I1 ≥ I1∗ , and by (2.2)

n Λ I I1
D1 1 j
=M = β1 j − I1 − (μ1 + γ1 )
μ1 I1∗ I1∗
j=1

n Λ I I1
− I1∗
1 j
≤ β1 j − (μ1 + γ1 )
μ1 I1∗ I1∗
j=1

n Λ I ∗ I1 I1
I1∗
1 j
≤ β1 j − − (μ1 + γ1 ) = 0.
μ1 (I1∗ )2 I1∗
j=1
I1 Ij
The second inequality follows from the fact that I1∗ ≥ I ∗j for all j and the last equality
follows from the equilibrium equation. Similarly, if m ≤ 1 in I , then

n Λ2 I j I2
D2 = −m = β2 j I2 − ∗ + (μ2 + γ2 ) ∗
μ2 I2 I2
j=1

n Λ2 I j I2
≤ β2 j I2∗ − ∗ + (μ2 + γ2 ) ∗
μ2 I2 I2
j=1
n Λ2 I j I2
∗
I2
≤ β2 j I2∗ − ∗ + (μ2 + γ2 ) ∗ = 0.
μ2 (I2 ) 2 I2
j=1
The invariant sets where D1 = 0 and D2 = 0 are E 1 = {(I1 , . . . , In ) ∈ Rn+ | Ik ≤
Ik∗ , k = 1, . . . , n} and E 2 = {(I1 , . . . , In ) ∈ Rn+ | Ik ≥ Ik∗ , k = 1, . . . , n}, respec-
tively. Notice that E 1 ∩ E 2 = {(I1∗ , . . . , In∗ )}. Therefore, by LaSalle’s Invariance
Principle, (I1∗ , . . . , In∗ ) is globally asymptotically stable in int(Γ ) if R0 > 1.
3.4 Global Stability of the Endemic Equilibrium: Comparison

Method
In this subsection we provide another proof for the global stability of the endemic
equilibrium for (2.1), which follows the comparison argument for the network SIS
model (2.3) used in [45] and the fact that (2.2) is a cooperative system [46]. In the
following we show that lim inf t→∞ Ik (t) = lim supt→∞ Ik (t) for all k. Specifically,
we construct two sequences: one is a decreasing sequence bounded from below
(upper solutions) and the other an increasing sequence bounded from above (lower
solutions).
(1) (1)
Let u k = Λ μk for all k, and by Lemma 1, Ik (t) ≤ u k for all t ≥ 0. Define the
k
upper solution sequence

n Λk (m)
(m+1) j=1 βk j μk u j
uk = , 1 ≤ k ≤ n, m = 1, 2, . . . (3.2)
μk + γk + nj=1 βk j u (m)
j
(m)
It can be shown that Ik (t) ≤ u k for all m = 2, 3, . . . and t ≥ 0 when applying the
reduction argument to differential inequalities
⎛ ⎞

n
n
n
Λk (m) Λk (m) (m)
Ik ≤ βk j − Ik u j − (μk + γk )Ik = βk j u − Ik ⎝μk + γk + βk j u j ⎠ .
μk μk j
j=1 j=1 j=1
(2) Λk (1)
In addition, it follows from (3.2) that u k < μk = u k , and induction argument
shows that for m = 2, 3, . . .
n Λk (m) n Λk (m−1)
(m+1) j=1 βk j μk u j j=1 βk j μk u j (m)
uk = (m)
< (m−1)
= uk .
μk + γk + nj=1 βk j u j μk + γk + nj=1 βk j u j
(m)
Hence, the sequence {u k } is decreasing so its limit exists. Denote u k = limm→∞
(m)
u k . It follows from taking the limit on both sides of (3.2) that

n Λ
k
βk j − u k u j − (μk + γk )u k = 0, (3.3)
μk
j=1
implying that (u 1 , . . . u n ) is the equilibrium of (2.2).

By Theorem 1, model (2.1) is uniformly persistent when R0 > 1; thus there exists
0 < ε 1 such that lim inf t→∞ Ik (t) ≥ ε for all k. We keep ε small enough such
that
(R0 − 1)(μk + γk )
ε < min n . (3.4)
k j=1 βk j

βk j Sk0 βk j Sk0 βk j Sk0
Since matrices μj +γj and μk + γk are similar, it follows that ρ μk + γk =
R0 > 1. Let (η1 , . . . , ηn )T with 0 < ηk < ε, 1 ≤ k ≤ n, be the Perron eigenvector
βk j Sk0
for μk + γk corresponding to the eigenvalue R0 ; that is,

n
βk j Sk0
η j = R0 η k . (3.5)
μk + γk
j=1

(m) (1)
Now consider the lower solutions lk with lk = ηk and
n Λk (m)
(m+1) j=1 βk j μk l j
lk = (m)
, 1 ≤ k ≤ n, m = 1, 2, . . . (3.6)
μk + γk + nj=1 βk j l j
Choose T large enough such that Ik (t) ≥ ε for all t > T and k = 1, . . . , n. It follows
from (3.4) and (3.5) that
n Λk n
j=1 βk j μk η j j=1 βk j Sk η j
0
(2) (1)
lk = > = ηk = l k .
μk + γk + nj=1 βk j η j R0 (μk + γk )
(m)
Reduction can further be used to show that Ik (t) ≥ lk for all m = 2, 3, . . . and
(m) (m)
t ≥ T , and that {lk } is increasing so its limit exists. Denote lk = limm→∞ lk ,
and lk > 0 satisfies the same equation as u k in (3.3). Therefore, (l1 , . . . , ln ) is the
nontrivial equilibrium of (2.2).
Given the uniqueness of the endemic equilibrium of (2.2) (see Sects. 3.2 or 3.3),
the inequalities lk ≤ Ik (t) ≤ u k imply that all solutions of (2.2) approach to the
unique endemic equilibrium.
4 Discussions and Open Problems
The classic multigroup SIR model (2.1), which was first proposed in [4], is revisited
in Sect. 2. The methods that can be applied to analyze this kind of heterogeneous
disease models are reviewed in Sect. 3. The graph-theoretic method in Sect. 3.2 can
be used to guide the construction of Lyapunov functions for heterogeneous models,
which requires individual Lyapunov functions for homogeneous models as build-
ing blocks. The method using multiple Lyapunov functions as shown in [4] (see
Sect. 3.3) requires strong techniques in combining the invariant sets derived from
each Lyapunov functions, which makes it hard to apply for other heterogeneous
models. Comparison method as in [45] (also see Sect. 3.4) requires certain monotone
properties on the flow generated by the model. Due to these limitations, the global
stability of the endemic equilibrium for several multigroup infectious disease models
remains open. For example, in model (2.1), each group population size Nk either is
a constant or approaches a constant, which allows the reduction to model (2.2) con-
sisting of only infectious compartments. If disease-induced mortality is introduced
to each group (i.e., adding a term αk Ik at the end of the equation for Ik in (2.1)), then
such a reduction does not hold any more. It remains open whether the endemic equi-
librium is unique and whether it is globally asymptotically stable when R0 > 1. The
global dynamics of the SEIRS model has recently been resolved by Cheng and Yang
[58], based on previous results using the theory of compound differential equations
in [59], but no complete studies on global dynamics of the multigroup SEIRS model
have been done yet.
Multigroup models can also be used to model spatial heterogeneity, in which
between-group transmissions are due to pathogen and/or host movements; see, for
example, [60, 61]. In [61], the multigroup model is also called the Lagrangian model,
in comparison with Eulerian (multipatch) models that explicitly incorporate pathogen
and/or host movements. Demographic and movement data can be used to derive the
contact matrix [βk j ] for this situation; see, for example, the gravity model used in
[62] to derive between-group transmission for the cholera outbreak in Haiti.
Acknowledgments The authors would like to acknowledge the financial support from the College
of Sciences and the Department of Mathematics at the University of Central Florida.
Appendix: Graph-Theoretic Method to the Construction

of Lyapunov Functions
Given a weighted digraph G with n vertices, define the n ×n weight matrix A = [ai j ]
with entry ai j > 0 equal to the weight of arc ( j, i) if it exists, and 0 otherwise. We
denote such a weighted digraph by (G , A). A digraph G is strongly connected if, for
any pair of distinct vertices i, j, there exists a directed path from i to j (and also from
j to i). A weighted digraph (G , A) is strongly connected if and only if the weight
matrix A is irreducible [63]. The Laplacian matrix L = [i j ] of (G , A) is defined as

−ai j
for i = j,
i j = (1)
k=i ik for i = j.
a
The following result gives a graph-theoretic description of the cofactors of the diag-
onal entries of L. We refer the reader to [64] for its proof.
Theorem A (Kirchhoff’s Matrix Tree Theorem) Assume n ≥ 2 and let ci be the

cofactor of ii in L. Then

ci = w(T ), i = 1, 2, . . . , n, (2)
T ∈Ti
where Ti is the set of all spanning trees T of (G , A) that are rooted at vertex i, and
w(T ) is the weight of T . If (G , A) is strongly connected, then ci > 0 for 1 ≤ i ≤ n.
Let U be an open set in Rm . Consider a differential equation system
z k = f k (z 1 , z 2 , . . . , z m ), k = 1, 2, . . . , m, (3)
with z = (z 1 , z 2 , . . . , z m ) ∈ U . The following result can be used to construct

Lyapunov functions for (3), and its proof can be found in [24, Sect. 3] or [25, Sect. 3].
Theorem B Suppose that the following assumptions are satisfied.
(1) There exist functions Di : U → R, G i j : U → R, and constants ai j ≥ 0 such
that, for every 1 ≤ i ≤ n, Di = Di |(3) ≤ nj=1 ai j Gi j (z) for z ∈ U .
(2) For A = [ai j ], each directed cycle C of (G , A) has (s,r )∈E (C ) G r s (z) ≤ 0 for
z ∈ U , where E (C ) denotes the arc set of the directed cycle C .
n
Then, the function D(z) = i=1 ci Di (z), with constants ci ≥ 0 as given in
Theorem A, satisfies D = D |(3) ≤ 0; that is, D is a Lyapunov function for (3).
References
1. Hethcote, H.W., Yorke, J.A.: Gonorrhea Transmission Dynamics and Control. Lecture Notes
in Biomath, vol. 56. Springer, Berlin (1984)
2. Galvani, A.P., May, R.M.: Dimensions of superspreading. Nature 438, 293–295 (2005)
3. Lloyd-Smith, J.O., Schreiber, S.J., Kopp, P.E., Getz, W.M.: Superspreading and the effect of
individuals variation on disease emergence. Nature 438, 355–359 (2005)
4. Lajmanovich, A., Yorke, J.A.: A deterministic model for gonorrhea in a nonhomogeneous
population. Math. Biosci. 28, 221–236 (1976)
5. Aronsson, G., Mellander, I.: A deterministic model in biomathematics, asymptotic behavior
and threshold conditions. Math. Biosci. 49, 207–222 (1980)
6. Beretta, E., Capasso, V.: Global stability results for a multigroup SIR epidemic model. In: Hal-
lam, T.G., Gross, L.J., Levin, S.A. (eds.) Mathematical Ecology, pp. 317–342. World Scientific,
Singapore (1986)
7. Hethcote, H.W.: Mathematical models for the spread of infectious diseases. In: Ludwig, D.,
Cooke, K.L. (eds.) Epidemiology, pp. 122–131. SIAM, Philadelphia (1975)
8. Hethcote, H.W.: Qualitative analyses of communicable disease models. Math. Biosci. 28, 335–
356 (1976)
9. Hethcote, H.W.: An immunization model for a heterogeneous population. Theor. Popu. Biol.
14, 338–349 (1978)
10. Hethcote, H.W., Thieme, H.R.: Stability of the endemic equilibrium in epidemic models with
subpopulations. Math. Biosci. 75, 205–227 (1985)
11. Huang, W., Cooke, K.L., Castillo-Chavez, C.: Stability and bifurcation for a multiple-group
model for the dynamics of HIV/AIDS transmission. SIAM J. Appl. Math. 52, 835–854 (1992)
12. Lin, X., So, J.W.H.: Global stability of the endemic equilibrium and uniform persistence in
epidemic models with subpopulations. J. Austral. Math. Soc. Ser. B 34, 282–295 (1993)
13. Rass, L., Radcliffe, J.: Global asymptotic convergence results for multitype models. Int. J.
Appl. Math. Comput. Sci. 10, 63–79 (2000)
14. Thieme, H.R.: Local stability in epidemic models for heterogeneous populations. In: Capasso,
V., Grosso, E., Paveri-Fontana, S.L. (eds.) Mathematics in Biology and Medicine. Lecture
Notes in Biomath, vol. 57, pp. 185–189. Springer, Berlin (1985)
15. Thieme, H.R.: Mathematics in Population Biology. Princeton University Press, Princeton
(2003)
16. Guo, H., Li, M.Y., Shuai, Z.: Global stability of the endemic equilibrium of multigroup SIR
epidemic models. Can. Appl. Math. Q. 14, 259–284 (2006)
17. Freedman, H.I., So, J.W.H.: Global stability and persistence of simple food chains. Math.
Biosci. 76, 69–86 (1985)
18. Goh, B.S.: Global stability in many-species systems. Am. Nat. 111, 135–143 (1977)
19. Hsu, S.B.: On global stability of a predator-prey system. Math. Biosci. 39, 1–10 (1978)
20. Guo, H., Li, M.Y.: Global dynamics of a staged progression model for infectious diseases.
Math. Biosci. Eng. 3, 513–525 (2006)
21. Korobeinikov, A., Maini, P.K. A Lyapunov function and global properties for SIR and SEIR
epidemiological models with nonlinear incidence, Math. Biosci. Eng. 1, 57–60 (2004)
22. Korobeinikov, A., Wake, G.C.: Lyapunov functions and global stability for SIR, SIRS, and SIS
epidemiological models. Appl. Math. Lett. 15, 955–960 (2002)
23. Guo, H., Li, M.Y., Shuai, Z.: A graph-theoretic approach to the method of global Lyapunov
functions. Proc. Am. Math. Soc. 136, 2793–2802 (2008)
24. Li, M.Y., Shuai, Z.: Global-stability problems for coupled systems of differential equations on
networks. J. Differ. Eqn. 248, 1–20 (2010)
25. Shuai, Z., van den Driessche, P.: Global stability of infectious disease models using Lyapunov
functions. SIAM J. Appl. Math. 73, 1513–1532 (2013)
26. Sun, R.: Global stability of the endemic equilibrium of multigroup SIR models with nonlinear
incidence. Comput. Math. Appl. 60, 2286–2291 (2010)
27. Sun, R., Shi, J.: Global stability of multigroup epidemic model with group mixing and nonlinear
incidence rates. Appl. Math. Comput. 218, 280–286 (2011)
28. Li, M.Y., Shuai, Z., Wang, C.: Global stability of multi-group epidemic models with distributed
delays. J. Math. Anal. Appl. 361, 38–47 (2010)
29. Shu, H., Fan, D., Wei, J.: Global stability of multi-group SEIR epidemic models with distributed
delays and nonlinear transmission. Nonlinear Anal. Real World Appl. 13, 1581–1592 (2012)
30. Shuai, Z., van den Driessche, P.: Impact of heterogeneity on the dynamics of an SEIR epidemic
model. Math. Biosci. Eng. 9, 393–411 (2012)
31. Yuan, Z., Wang, L.: Global stability of epidemiological models with group mixing and nonlinear
incidence rates. Nonlinear Anal. Real World Appl. 11, 995–1004 (2010)
32. Guo, H., Li, M.Y., Shuai, Z.: Global dynamics of a general class of multi-stage models for
infectious diseases. SIAM J. Appl. Math. 72, 261–279 (2012)
33. Wang, J., Pang, J., Liu, X.: Modelling diseases with relapse and nonlinear incidence of infection:
a multi-group epidemic model. J. Biol. Dyn. 8, 99–116 (2014)
34. Yuan, Z., Zou, X.: Global threshold property in an epidemic model for disease with latency
spreading in a heterogeneous host population. Nonlinear Anal. Real World Appl. 11, 3479–
3490 (2010)
35. Ding, D., Ding, X.: Global stability of multi-group vaccination epidemic models with delays.
Nonlinear Anal. Real World Appl. 12, 1991–1997 (2011)
36. Kuniya, T.: Global stability of a multi-group SVIR epidemic model. Nonlinear Anal. Real
World Appl. 14, 1135–1143 (2013)
37. Wang, J., Takeuchi, Y., Liu, S.: A multi-group SVEIR epidemic model with distributed delay
and vaccination. Int. J. Biomath. 5(3), 235–258 (2012)
38. Ji, C., Jiang, D., Shi, N.: Multigroup SIR epidemic model with stochastic perturbation. Phys.
A 390, 1747–1762 (2011)
39. Ji, C., Jiang, D., Yang, Q., Shi, N.: Dynamics of a multigroup SIR epidemic model with
stochastic perturbation. Automatica 48, 121–131 (2012)
40. Yang, Q., Mao, X.: Extinction and recurrence of multi-group SEIR epidemic models with
stochastic perturbations. Nonlinear Anal. Real World Appl. 14, 1434–1456 (2013)
41. Eisenberg, M.C., Shuai, Z., Tien, J.H., van den Driessche, P.: A cholera model in a patchy
environment with water and human movement. Math. Biosci. 246, 105–112 (2013)
42. Shuai, Z., van den Driessche, P.: Global dynamics of cholera models with differential infectivity.
Math. Biosci. 234, 118–126 (2011)
43. Ding, D., Wang, X. Ding, X.: Global stability of multigroup dengue disease transmission model,
J. Appl. Math. 2012, no. 342472
44. Huang, G., Wang, J., Zu, J.: Global dynamics of multi-group dengue disease model with latency
distributions, Math. Meth. Appl. Sci. (2014) (in press)
45. Wang, L., Dai, G.-Z.: Global stability of virus spreading in complex heterogeneous networks.
SIAM J. Appl. Math. 68, 1495–1502 (2008)
46. Smith, H.L.: Monotone Dynamical Systems. An Introduction to the Theory of Competitive
and Cooperative Systems. American Mathematical Society, Providence (1995)
47. Markus, L.: Asymptotically autonomous differential systems. In: Lefschetz, S. (ed.) Contribu-
tions to the Theory of Nonlinear Oscillations III. Annals of Mathematics Studies, vol. 36, pp.
17–29. Princeton University Press, Princeton (1956)
48. Mischaikow, K., Smith, H., Thieme, H.R.: Asymptotically autonomous semiflows: chain recur-
rence and Lyapunov functions. Trans. Am. Math. Soc. 347, 1669–1685 (1995)
49. Pastor-Satorras, R., Vespignani, A.: Epidemic spreading in scale free networks. Phys. Rev. Lett.
86, 3200–3203 (2001)
50. Pastor-Satorras, R., Vespignani, A.: Epidemic dynamics in finite size scale-free networks. Phys.
Rev. E 65, 035108 (2002)
51. Diekmann, O., Heesterbeek, J.A.P., Metz, J.A.J.: On the definition and the computation of the
basic reproduction ratio R0 in models for infectious diseases in heterogeneous populations. J.
Math. Biol. 28, 365–382 (1990)
52. van den Driessche, P., Watmough, J.: Reproduction numbers and sub-threshold endemic equi-
libria for compartments models of disease transmission. Math. Biosci. 180, 29–48 (2002)
53. LaSalle, J.P.: The Stability of Dynamical Systems, Regional Conference Series in Applied
Mathematics. SIAM, Philadelphia (1976)
54. Freedman, H.I., Ruan, S., Tang, M.: Uniform persistence and flows near a closed positively
invariant set. J. Dyn. Differ. Eqn. 6, 583–600 (1994)
55. Li, M.Y., Graef, J.R., Wang, L., Karsai, J.: Global dynamics of a SEIR model with varying
total population size. Math. Biosci. 160, 191–213 (1999)
56. Smith, H.L., Waltman, P.: The Theory of the Chemostat: Dynamics of Microbial Competition.
Cambridge University Press, Cambridge (1995)
57. Bhatia, N.P., Szegö, G.P.: Dynamical Systems: Stability Theory and Applications. Lecture
Notes in Mathematics, vol. 35. Springer, Berlin (1967)
58. Cheng, Y., Yang, X.: On the global stability of SEIRS models in epidemiology. Can. Appl.
Math. Q. 20, 115–133 (2012)
59. Li, M.Y., Muldowney, J.S., van den Driessche, P.: Global stability of the SEIRS model in
epidemiology. Can. Appl. Math. Q. 7, 409–425 (1999)
60. Lloyd, A.L., May, R.M.: Spatial heterogeneity in epidemic models. J. Theor. Biol. 179, 1–11
(1996)
61. Cosner, C., Beier, J.C., Cantrell, R.S., Impoinvil, D., Kapitanski, L., Potts, M.D., Troyo, A.,
Ruan, S.: The effects of human movement on the persistence of vector-borne diseases. J. Theory
Biol. 258, 550–560 (2009)
62. Tuite, A.R., Tien, J.H., Eisenberg, M.C., Earn, D.J.D., Ma, J., Fisman, D.N.: Cholera epidemic
in Haiti, 2010: using a transmission model to explain spatial spread of disease and identify
optimal control interventions. Ann. Intern. Med. 154, 593–601 (2011)
63. Berman, A., Plemmons, R.J.: Nonnegative Matrices in the Mathematical Sciences. Academic
Press, New York (1979)
64. Moon, J.W.: Counting Labelled Trees. William Clowes and Sons, London (1970)
Radially Symmetric Vibrations
of Exponentially Tapered Clamped Circular
Sandwich Plate Using Harmonic Differential
Quadrature Method
Rashmi Rani and Roshan Lal
Abstract In the present paper, axisymmetric vibrations of a circular sandwich

plate with relatively stiff core of exponentially varying thickness have been inves-
tigated. The face sheets are treated as membranes of constant thickness and the
core is assumed to be solid as well as moderately thick. The equations of motion
have been derived using Hamilton’s energy principle. The frequency equation for
clamped boundary condition is obtained by employing harmonic differential quadra-
ture method. The lowest three roots of this equation have been reported as the fre-
quencies for the first three modes of vibration. The effect of various plate parameters
on the natural frequencies has been studied. Three-dimensional mode shapes for
a specified sandwich plate have been illustrated. A comparison of the results with
published work has been made.
Keywords Sandwich plate · HDQ method · Mode shapes
1 Introduction
A sandwich essentially consists of two thin faces sandwiching a light core between
them. It is one of the most useful forms of composite structures which has wide
applications in aerospace and many other industries. Sandwich construction provides
several key benefits over the conventional structures, such as very high bending
stiffness, low weight, cost effectiveness, durability together with a very high stiffness-
to-weight ratio and high bending strength. Due to these extra ordinary features,
sandwich plates are used for both interior and exterior components of aircraft (e.g.,
overhead bins, floor panels, radome, aerodynamic fairings), space vehicles, trains,
ships, boats, cargo containers, and in residential construction [1, 2]. In many practical
situations, particularly in the design of aerospace vehicles such as wings, control
R. Rani (B) · R. Lal

Indian Institute of Technology Roorkee, Roorkee, India
e-mail: rani.msrashmi9@gmail.com
R. Lal

DOI 10.1007/978-81-322-2485-3_52
634 R. Rani and R. Lal
surfaces (e.g., ailerons, elevators, and rudders), and rotor blades of helicopter, it
becomes essential to use tapered sandwich construction for grater structural and
aerodynamic efficiency. It has necessitated to study the dynamic behavior of sandwich
plates of nonuniform thickness with a fair amount of accuracy. Until now, various
theories for multilayered structures, particularly for composite and sandwich plates,
have been developed and given in references [3, 4], to mention a few. The work
up to 2008 has been reported by Carrera and Brichetto in their excellent survey
article [5] on numerical assessment of classical and refined theories for the analysis
of sandwich plates. Very recently, solution of the static buckling for a uniformly
compressed rectangular sandwich plate having two parallel edges simply supported
using the generalized Galerkin method has been given by Lopatin and Morzov [6].
Khalili et al. [7] used finite element procedure based on second-order Lagrangian
elements and Galerkin-type formulation for the analysis of rectangular multilayered
and sandwich plates. In the recent time, harmonic differential quadrature method has
emerged as a powerful technique to solve a variety of problems in engineering and
physical sciences and gives highly accurate solution with minimal computational
effort [8].
In the present work, harmonic differential quadrature (HDQ) method has been
employed to study the axisymmetric behavior of a circular sandwich plate with
isotropic core of exponentially varying thickness in radial direction using a refined
theory. The face sheets are isotropic and treated as membranes of constant thickness.
The effect of transverse shear deformation and rotatory inertia is retained in the core.
The frequency equation for clamped boundary condition has been obtained. The
lowest three roots of these frequency equations have been reported as the natural
frequencies for the first three modes of vibration. Effect of various plate parameters
has been studied on natural frequencies for clamped boundary condition. Comparison
of results with those available in the literature has been presented.
Consider a circular sandwich plate of radius a and thickness 2(h c + h f ) referred to

cylindrical polar coordinate (r, θ, z) being the middle surface of the plate and also
the plane of symmetry. The line r = 0 is the axis of the plate. A cross-sectional view
of the plate with exponentially varying core thickness h c (r ), the facing thickness
h f ( h c ), and facing slope φ is shown in Fig. 1. Any location in the lower or upper
facing is identified by its r -coordinate or by its φ-coordinate, where φ = φ(r ). The
variables φ and r are connected by the relations
r = Rθ sin φ and dr = Rφ cos φdφ (1)
where Rφ is the radius of curvature of the core-facing interface and Rθ is the length
of the normal between any point on the core-facing interface and the axis of sandwich
plate. The thickness variation of the core with radial distance is given by:
Radially Symmetric Vibrations of Exponentially Tapered Clamped Circular … 635
Fig. 1 Cross-section of
circular sandwich plate with
core of exponentially varying
thickness, i.e., h c = h o eαx
dh c
r = − tan φ, h c = h o eαx
d
and
dφ d 2hc
= − cos2 φ 2
dr dr
where α is taper parameter and h o is the thickness of the core at the center of the
plate. The differential equations governing the axisymmetric vibration of such plate
[9], are given by
d 2ψ dψ d2W dW
Uo 2
+ U 1 + (U 2 − Ω 2
P2 ) + U 3 2
+ U4 = 0, (2)
dx dx dx dx
d 2ψ dψ d2W dW
U5 2
+ U 6 + U 7 ψ + U 8 2
+ U9 − Ω 2 P10 W = 0, (3)
dx dx dx dx
where

d Hc 2 d Hc
Uo = + 3Rc Hc
Rc Hc2 x x + 3R f H f x cos φ Hc + 2
3
dx dx
dφ
− 9R f H f Hc x 2 cos2 φ sin φ , U3 = −3R f H f x 2 sin φ cos2 φ,
dx

d Hc d Hc
U2 = 3Rc vc Hc x − Rc Hc2 + 3R f H f x cos3 φ − Hc sec φ − 3ks x 2
dx dx
d 2 Hc dφ d Hc dφ
− 3R f H f cos3 φx 2 − 3R f H f Hc v f sin φ x − 9R f H f Hc x 2 cos2 φ sin φ ,
dx2 dx dx dx
dφ
U4 = 3R f H f x sin φ(v f − cos2 φ) − 3ks x 2 − 3R f H f x 2 cos φ(cos2 φ − 2 sin2 φ)
dx
U5 = R f H f Hc x sin φ cos2 φ, P2 = −x 2 Hc (Hc + 3H f Rρ sec φ),
d Hc
U6 = ks Hc x R f H f x sin φ(2x cos2 φ + v f Hc + Hc cos2 φ
dx
dφ
+ R f H f Hc x cos φ(cos2 φ − 2 sin2 φ) ,
dx

d Hc d Hc
U7 = R f H f sin φ sin φ cos φ − v f + ks Hc + x
dx dx
dφ
+ R f H f (x sin φ(2 cos2 φ − sin2 φ) − v f Hc cos φ) ,
dx
Uo = (Rc Hc2 + 3R f H f Hc cos3 φ)x 2 ,
P10 = −x(Hc + H f Rρ sec φ), U8 = x(ks Hc + R f H f cos φ sin2 φ),
ρc a 2 Ω 2 ρf
Ω2 = and Rρ =
μc ρc
The solution of Eqs. (2) and (3) together with regularity condition Ψ = Q r = 0 [10]
at the center x = 0 and clamped boundary condition at the edge x = 1 gives rise
to a two point boundary value problem with variable coefficients whose closed form
solution is not possible. Keeping this in view, an approximate solution is obtained
by employing HDQ method.
3 Method of Solution
Let xi , 1 = 1, 2, ..., m be the m grid points in the applicability range [0, 1] of the
plate. According to HDQ method [8], the nth-order derivatives of W (x) and Ψ (x)
with respect to x at the ith-point xi are given by

m
(n)
(Wxn , Ψxn ) = Ci j (W (x j )), (Ψ (x j )), i = 1, 2, ..., m (4)
j=1
where Cinj are the weighting coefficients and the first-order weighting coefficients,
i.e., Cinj for n = 1 have been given as follows:
(1) (π/2)M 1 (xi )

Ci j =
M 1 (x j ) sin[(xi − x j )/2]
where

m
xi − x j
M (xi ) =
1
sin π , i, j = 1, 2, ..., m but, j = i (5)
2
( j=1
j =i )
and the second-order coefficients are generated from the recurrence relation

xi − x j
Ci(2)
j = Ci(1)
j
(1)
Ci j − π ctg π , i, j = 1, 2, ..., m, but j = i,
2
(6)
with

m
Cii(n) = − Ci(n)
j , i = 1, 2, ..., m and n = 1 or 2. (7)
( j=1
)
j =i
Now, discretizing Eqs. (2) and (3) at the grid points x = xi , i = 2, 3, ..., (m − 1),
and substituting the values of first two derivatives of W and Ψ from Eq. (4), we get

m
(2) (1)

m
(2) (1)
(U0,i Ci, j + U1,i Ci, j )Ψ j + (U3,i Ci, j + U4,i Ci, j )W j + (U2,i − Ω 2 P2,i )Ψi = 0,
j=1 j=1
(8)

m
(2) (1)

m
(2) (1)
(U5,i Ci, j + U6,i Ci, j )Ψ j + (U8,i Ci, j + U9,i Ci, j )W j + U7,i Ψi − Ω 2 P10,i Wi = 0,
j=1 j=1
(9)
The satisfaction of Eqs. (8) and (9) at (m−2) internal grid points xi , i = 2, ..., (m-1)
together with the regularity condition: Ψ = Q r = 0 (Wu et al. [10]), at the center
of the plate provides a set of (2m −2) equations in terms of unknowns W j = W (x j )
and Ψ j = Ψ (x j ), j = 1, 2, ..., m. The resulting system of equations can be
written in matrix form as
[U ][C] = [0], (10)

where U and C are the matrices of orders (2m − 2) × 2m and (2m × 1), respectively.
The above (m − 2) internal grid points chosen for collocation are the zeros of shifted
Chebyshev polynomial of order (m − 2) with orthogonality range (0, 1), given by

1 2k − 1 π
xk+1 = 1 + cos .
2 m−2 2
4 Boundary Condition and Frequency Equation
By satisfying the relation: (i) Ψ = W = 0 : for clamped edge (C-plate), a set of two
homogeneous equations is obtained. These equations together with the field Eq. (10)
give a complete set of 2m equations in terms of 2m unknowns which can be written
as
U
[C] = [0], (11)
UC
where U C is a matrix of order 2 × 2m. For a nontrivial solution of Eq. (11), the
frequency determinant must vanish and hence

U
C = 0. (12)
U
The frequency Eq. (12) provides the values of the frequency parameter Ω and solved
using MATLAB for various values of plate parameters. The numerical values of
the lowest three roots have been reported as the first three natural frequencies to
investigate the influence of the taper parameters, core thickness at the center, and
face thickness for clamped boundary condition. In the work reported here, the values
of various plate parameters are taken as follows:
α = −0.5(0.1)0.5, Ho = 0.05(0.05)0.30, H f = 0.0025(0.0025)0.02.
The material for the core and the facings are taken to be PVC (Poly vinyl chloride)
and aluminum, respectively, for which the various constants are, Rc = 2.85, R f =
1232.21, Rρ = 20.76, vc = 0.3 and v f = 0.3 and ks = 1 from Ref. [8].
To choose the appropriate number of grid points m, a computer program used to
evaluate the frequencies, was run for m = 5(1)20 for different sets of plate para-
meters for clamped boundary condition. The numerical values showed a consistent
improvement with the increase in the number of grid points m. In all the computa-
tions, the number of grid points has been taken as 12, since further increase in m
does not improve the result even at the fourth place of decimal. The convergence of
Table 1 Convergence study for the first three frequencies

Boundary condition Clamped plate
m/mode I II III
10 3.0664 4.6264 6.8181
11 3.0664 4.6264 6.8170
12 3.0664 4.6264 6.8171
13 3.0664 4.6264 6.8171
14 3.0664 4.6264 6.8171
15 3.0664 4.6264 6.8171
16 3.0664 4.6264 6.8171
17 3.0664 4.6264 6.8171
18 3.0664 4.6264 6.8171
19 3.0664 4.6264 6.8171
20 3.0664 4.6264 6.8171
Table 2 Comparison of results for isotropic circular plate

Boundary condition Clamped plate
Method/mode I II III
Exact [11] 10.2158 39.7711 89.1041
Finite element [12] 10.2159 39.7766 89.1708
Rayleigh-Ritz [13] 10.2160 39.7710 89.1030
Receptance [14] 10.2160 39.7710 89.1041
Ritz method [15] 10.2158 39.7711 –
symplectic method 10.2160 39.7710 –
[16]
Present 10.2158 39.7711 89.1041
frequency parameter Ω for a particular set α = 0.5, Ho = 0.3, H f = 0.02 is shown

in Table 1.
A comparison of results for specified uniform isotropic circular plate (α =
0, H f = 0, vc = 0.3, Ho = 0.001, ks = 1) with exact solution [11] and obtained
by Finite element method [12], Rayleigh–Ritz method [13], receptance method [14],
Ritz method [15], symplectic method [16] has been presented in Table 2. An excellent
agreement of the frequencies shows the versatility of the present method.
The numerical results are given in Figs. 2, 3 and 4. Figure 2 depicts the effect of
taper parameter α on the frequency parameter Ω for H f = 0.005 and two values of
Ho = 0.1, 0.2, for all the three modes of vibration for clamped plate. It is observed
that frequency parameter Ω increases with the increasing values of taper parameter
α. The rate of increase of frequency parameter Ω with taper parameter α decreases
Fig. 2 Taper parameter α

versus frequency parameter
for H f = 0.005;
Ho = 0.1; , Ho = 0.2; .
First mode:—— ; Second
mode:· · · · · · · · · ; Third
mode:− − −
Fig. 3 Core thickness at the

center Ho versus frequency
parameter
for α = 0.5:
H f = 0.005; ,
H f = 0.010; . First
mode:—— ; Second
mode:· · · · · · · · · ; Third
mode:− − −
Fig. 4 Face thickness H f

versus frequency parameter
for α = 0.5; Ho = 0.1; ,

Ho = 0.1; . First
mode:—— ; Second
mode:· · · · · · · · · ; Third
mode:− − −
with the increase in the number of modes. The effect of Ho is more pronounced for
α = −0.5 as compared to α = 0.5.
Figure 3 demonstrates the effect of core thickness at the center Ho on the fre-
quency parameter Ω for α = 0.5 and two different values of face thickness
H f = 0.005, 0.01, for all the three modes for clamped plate. It is observed that the
frequency parameter Ω increases with the increasing values of Ho for all the three
modes. This effect is more pronounced for Ho = 0.05 as compared to Ho = 0.3 . Fur-
ther, the rate of increase in the values of Ω is comparatively higher for Ho ≤ 0.175 as
compared to Ho > 0.175. This rate increases with the increase in number of modes
with the increasing values of H f .
Figure 4 depicts the behavior of facing thickness H f on the frequency parameter
Ω for α = 0.5 and two different values of Ho = 0.1, 0.2, for all the three modes of
vibration for clamped plate. It can be seen that the frequency parameter Ω decreases
with the increasing values of face thickness H f for all the three modes. This effect is
more pronounced for H f = 0.02 as compared to H f = 0.0025 for the increasing val-
ues of Ho from 0.1 to 0.2. This effect increases with the increase in number of modes.
The rate of decrease in the values of frequency parameter Ω with increasing values of
H f is found to increase with increasing number of modes. Three-dimensional mode
shapes for specified clamped circular sandwich plate taking Ho = 0.1, H f = 0.005,
and α = 0.5 have been plotted and shown in Fig. 5.
Fig. 5 First three mode shapes for clamped circular sandwich plate
6 Conclusion
The radially symmetric vibration of clamped circular sandwich plates with core of
exponentially varying thickness has been analyzed employing HDQ method. The
effect of transverse shear deformation and rotatory inertia has been retained in the
core and the face sheets are treated as membrane of constant thickness. It is observed
that the frequency parameter Ω,
(i) increases with the increasing values of taper parameter α and core thickness at
the center Ho ,
(ii) decreases with the increasing values of face thickness parameter H f .
The present analysis will be of great help to the design engineers dealing with sand-
wich structures in obtaining the desired frequency by varying one or more plate
parameters involved in the present model.
Acknowledgments One of the authors, Rashmi Rani is thankful to Ministry of Human Resources
and Development (MHRD), India for the financial support to carry out this research work.
References
1. Vel, S.S., Caccese, V., Zhao, H.: Elastic coupling effects in tapered sandwich panels with
laminated anisotropic composite facings. J. Compos. Mater. 39, 2161–2183 (2005)
2. Mustapha, S., Ye, L.: Damage identification and assessment in tapered sandwich structures
using guided waves. Key Eng. Mater. 558, 25–38 (2013)
3. Altenbach, H.: Theories for laminated and sandwich plates. Mech. Compos. Mater. 34, 333–348
(1998)
4. Carrera, E., Ciuffreda, A.: Bending of composites and sandwich plates subjected to localized
lateral loadings: a comparison of various theories. Compos. Struct. 68, 185–202 (2005)
5. Carrera, E., Brischetto, S.: A survey with numerical assessment of classical and refined theories
for the analysis of sandwich plates. Appl. Mech. Rev. 62, 010803–010819 (2009)
6. Lopatin, A.V., Morozov, E.V.: Buckling of a uniformly compressed rectangular SSCF com-
posite sandwich plate. Compos. Struct. 105, 108–115 (2013)
7. Khalili, S.M.R., Shariyat, M., Rajabi, I.: A finite element based global local theory for static
analysis of rectangular sandwich and laminated composite plates. Compos. Struct. 107, 177–
189 (2014)
8. Civalek, O., Catal, H.H.: Linear static and vibration analysis of circular and annular plates by
the harmonic differential quadrature (HDQ) method. Eng. Arch. Fac. Osmangazi University,
XVII(1), (2003)
9. Lal, R., Rani, R.: Mode shapes and frequencies of radially symmetric vibrations of annular
sandwich plates of variable thickness. Acta Mech. 225, 1565–1580 (2014)
10. Wu, T.Y., Wang, Y.Y., Liu, G.R.: Free vibration analysis of circular plates using generalized
differential quadrature rule. Comput. Methods Appl. Mech. Eng. 191, 5365–5380 (2002)
11. Leissa, A.W.: Vibration of Plates. U.S. Government Printing Office (NASA SP 160), Wash-
ington, DC (1969)
12. Pardoen, G.C.: Axisymmetric vibration and stability of circular plate. Comput. struct. 9, 89–95
(1978)
13. Kim C.S., Dickinson S.M.: On the free, transverse vibration of annular and circular, thin,
sectorial plates subject to certain complicating effects. J. Sound Vib. 134, 407–421 (1989)
14. Azimi, S.: Free vibration of circular plates with elastic edge supports using the receptance
method. J. Sound Vib. 120(1), 19–35 (1998)
15. Ansari A.H.: Vibration of plates of variable thickness. Ph.D. Thesis, University of Roorkee,
India (2000)
16. Zhou Z.H., Wong K.W., Xu X.S., Leung, A.Y.T.: Natural vibration of circular and annular thin
plates by Hamiltonian approach. J. Sound Vib. 330, 1005–1017 (2011)
Hybrid Projective Synchronization
of Fractional-Order Neural Networks
with Time Delays
G. Velmurugan and R. Rakkiyappan
Abstract In this paper, the problem of hybrid projective synchronization of

fractional-order neural networks with time delay is extensively investigated. The
fractional-order neural networks with hub structure and time delay is considered. By
using stability theorem of linear fractional order systems with multiple time delays
and constructing an appropriate linear feedback control, some new sufficient con-
ditions are derived to ensure projective synchronization of fractional-order neural
networks with time delays. It means that the response system can be synchronized
with the drive system based on the choice of a scaling matrix. Finally, a numerical
example is provided to demonstrate the effectiveness of our theoretical results.
Keywords Hybrid projective synchronization · Fractional-order · Neural

networks · Time delays
1 Introduction
Fractional calculus has become an active area of research in recent years due to their
widespread applications in various fields of science and engineering, such as dielec-
tric polarization, viscoelasticity, heat conduction, biology, etc. [1–3]. Generally, most
of the real-world problems are modeled by fractional-order dynamical systems rather
than integer-order ones. That is fractional-order systems provide more accurate result
than the integer-order systems. In [4], the authors pointed out that fractional deriva-
tives provide an excellent tool for the description of memory and hereditary properties
G. Velmurugan (B) · R. Rakkiyappan

Department of Mathematics, Bharathiar University, Coimbatore 641046,
Tamil Nadu, India
e-mail: gvmuruga@gmail.com
R. Rakkiyappan
e-mail: rakkigru@gmail.com

DOI 10.1007/978-81-322-2485-3_53
646 G. Velmurugan and R. Rakkiyappan
of various processes. Recently, many of the researchers have focused their interest
and attention to analysis the fractional-order dynamical systems and many good
results have been reported in the existing literature [5–7].
In the past few decades, the dynamical analysis of neural networks has received
increasing interest and hot topic of research because of their potential applications
in numerous fields, such as pattern recognition, associative memory and combinato-
rial optimization, etc. [8, 9]. In fact, fractional-order systems have infinite memory.
According to that feature, the incorporation of a memory term into a neural network
model is an extremely important improvement. Therefore, it is necessary to investi-
gate the dynamical analysis of fractional-order neural networks (FNNs). As we know
that, time delay is an unavoidable factor in the practical applications. It follows that,
many authors extensively analysis the FNNs with time delays and some remarkable
results have been proposed in the literature [10–14].
On the other hand, synchronization of fractional-order chaotic systems and FNNs
have received much attention in the area of nonlinear science and it has been applied
many fields such as image processing, secure communication and ecological sys-
tem. In [15], the authors have been introduced synchronization of chaotic systems.
In the literature, there are many types of synchronization have been exposed and
investigated, such as complete synchronization, anti-phase synchronization, projec-
tive synchronization, etc. [15–19]. There are several methods have been provided for
the synchronization based on linear feedback control, adaptive control, sliding mode
control, etc. Moreover, complete synchronization and anti-phase synchronization are
the special case of projective synchronization. However, projective synchronization
was first introduced in [20] and it has been providing faster communication with
its proportional feature. This feature can be used to extend binary digital to M-nary
digital communication for achieving fast communication in [21]. Thus, the analysis
of projective synchronization is very important in both theoretical and application
point of view. Recently, several important results have been derived for projective
synchronization in the literature [22–26]. In [22], the authors extensively studied the
modified projective synchronization of time-delayed fractional-order chaotic sys-
tems. Some new sufficient conditions were obtained to realize projective synchro-
nization of FNNs with open loop control and adaptive control in [23]. To the best of
our knowledge, there are few results of the projective synchronization of FNNs.
Motivated by the above discussion, the problem of hybrid projective synchroniza-
tion of FNNs with time delay is studied in this paper. Some new sufficient condi-
tions are derived to ensure the hybrid projective synchronization of FNNs with time
delays by using linear feedback control. Here, we use the Adams-Bashforh-moulton
predictor-corrector method [27] to solve FNNs by numerically.
This paper organized as follows. In Sect. 2, some basic definitions of fractional
calculus are given. Some new sufficient criteria for hybrid projective synchronization
of FNNs with time delays are obtained in Sect. 3. In Sect. 4, a numerical example is
provided to show the effectiveness of our main results. The conclusion of this paper
is given in Sect. 5.
Hybrid Projective Synchronization of Fractional-Order Neural Networks . . . 647
2 Preliminaries
In this section, we provide some basic definitions of fractional calculus. Throughout

this paper, we use the Caputo fractional-order derivative.
Definition 1 [1] The fractional integral of order α for a function g is defined as
t
1
I α g(t) = (t − τ )α−1 g(τ )dτ, (1)
Γ (α) 0
where t ≥ t0 and α > 0, Γ (·) is the gamma function defined as Γ (α) =

∞ α−1 −t
0 t e dt.
Definition 2 [1] The Caputo fractional derivative of order α for a function g(t) is

α 1 t g (n) (τ )
C Dt g(t) = dτ, (2)
Γ (n − α) 0 (t − τ )α−n+1
where t > 0 and n is a positive integer such that n − 1 < α < n ∈ Z + .
Definition 3 [14] The Laplace transform of the Caputo fractional-order

derivatives is

n−1
L{C Dtα g(t); s} = s α G(s) − s α−k−1 g (k) (0), n − 1 < α ≤ n,
k=0
where G(s) is the Laplace transform of g(t), g k (0) = 0, k = 1, 2, · · · , n, are the

initial conditions.
In this paper, consider the FNNs with time delays as drive system is described as

n
n
D α xi (t) = −ci xi (t) + ai j f j (x j (t)) + bi j f j (x j (t − τ )), i = 1, 2, · · · , n, (3)
j=1 j=1
where n corresponds to the number of units. 0 < α < 1, xi (t) is the state vector
of the ith unit. f j (·) denotes the nonlinear activation function. ai j and bi j denotes
the connection weight matrices without delay and with delay. ci > 0 is the self-
feedback connection weight matrix and τ is the constant time delay. Equation (3)
can be rewritten as in the vector form as follows
f (x(t)) +
D α x(t) = −C x(t) + A B f (x(t − τ )), (4)
where x(t) = (x1 (t), x2 (t), · · · , xn (t))T ∈ R n , C = diag(c1 , c2 , · · · , cn ) ∈ R n×n ,

= (ai j )n×n ∈ R n×n ,
A B = (bi j )n×n ∈ R n×n , f (x(t)) = ( f 1 (x1 (t)), f 2 (x2 (t)),
· · · , f n (xn (t)))T and f (x(t − τ )) = ( f 1 (x1 (t − τ )), f 2 (x2 (t − τ )), · · · , f n (xn

(t − τ )))T . System (4) can be linearize as follows
J x(t) + x̄(t − τ ),
D α x(t) = −C x(t) + A (5)

where J is the Jacobian matrix of f (x(t)) and x̄(t − τ ) = ( nj=1 b1 j q1 j x j

(t − τ ), · · · , nj=1 bn j qn j x j (t − τ ))T is the linearization vector of B f (x(t − τ )) at
the equilibrium point. Also, denote Ā = A J and B̄ = (bi j qi j )n×n , then (5) can be
rewritten as
D α x(t) = −C x(t) + Āx(t) + B̄x(t − τ ). (6)
The corresponding response system is given by
D α y(t) = −C y(t) + Āy(t) + B̄ y(t − τ ) + ψ(t), (7)
where ψ(t) = (ψ1 (t), ψ2 (t), · · · , ψn (t))T is the control input. Here, consider the
linear feedback control to realize synchronization between the derive system (6) and
response system (7). The controller ψ(t) is defined as
ψ(t) = Γ (y(t) − βx(t)), (8)
where Γ = diag(γ1 , γ2 , · · · , γn ) ∈ R n×n is a feedback gain matrix. The initial

conditions associated with the system (6) and (7) is x(t) = x(0) and y(t) = y(0), t ∈
[−τ, 0].
Definition 4 If there exists a real scaling matrix β ∈ R n×n , such that for any two
solutions x(t) and y(t) of system (6) and system (7) with different initial values
denoted by x(0) and y(0), one has
lim y(t) − βx(t) = 0, (9)

t→∞
then, drive system (6) and response system (7) are said to be globally hybrid projec-
tively synchronized.
3 Main Results
In this section, some new sufficient criteria has been derived to ensure that system
(6) and (7) is projectively synchronized under linear feedback control.
Let us define w(t) = y(t) − βx(t) be the synchronization errors. From (6) and
(7), the error system as follows
D α w(t) = −Cw(t) + Āw(t) + B̄w(t − τ ) + Γ w(t), (10)
with initial conditions w(t) = δ(0), t ∈ [−τ, 0].

Taking the Laplace transform, we have
s α E 1 (s) − s α−1 δ1 (0) = (−c1 + γ1 + ā11 )E 1 (s) + ā12 E 2 (s) + · · · + ā1n E n (s)
0
+ b̄11 e−sτ E 1 (s) + e−st δ1 (t)dt
−τ
0
+ b̄12 e−sτ E 2 (s) + e−st δ2 (t)dt
−τ
0
−sτ
+ · · · + b̄1n e E n (s) + e−st δn (t)dt ,
−τ
s α E 1 (s) − s α−1 δ1 (0) = (−c1 + γ1 + ā11 + b̄11 e−sτ )E 1 (s) + (ā12 + b̄12 e−sτ )E 2 (s) + · · ·
0
+ (ā1n + b̄1n e−sτ )E n (s) + b̄11 e−sτ e−st δ1 (t)dt
−τ
0 0
+ b̄12 e−sτ e−st δ2 (t)dt + · · · + b̄1n e−sτ e−st δn (t)dt,
−τ −τ
s α E 2 (s) − s α−1 δ2 (0) = (−c2 + γ2 + ā22 + b̄22 e−sτ )E 2 (s) + (ā21 + b̄21 e−sτ )E 1 (s) + · · ·
0
+ (ā2n + b̄2n e−sτ )E n (s) + b̄21 e−sτ e−st δ1 (t)dt
−τ
0 0
+ b̄22 e−sτ e−st δ2 (t)dt + · · · + b̄2n e−sτ e−st δn (t)dt,
−τ −τ
···
···
α α−1
s E n (s) − s δn (0) = (−cn + γn + ānn + b̄nn e−sτ )E n (s) + (ān1 + b̄n1 e−sτ )E 1 (s) + · · ·
0
+ (ān,n−1 + b̄n,n−1 e−sτ )E n−1 (s) + b̄n1 e−sτ e−st δ1 (t)dt
−τ
0 0
−sτ −st −sτ
+ b̄n2 e e δ2 (t)dt + · · · + b̄nn e e−st δn (t)dt, (11)
−τ −τ
where E(s) is the Laplace transform of w(t) with E(s) = L(w(t)). The vector form
of (11) is
Δ(s) · E(s) = k(s), (12)
where Δ(s) is the characteristic matrix of system (11) and k(s) is the remainder
nonlinear part of system (11), such as
⎛ ⎞
s α + d1 − b̄11 e−sτ −ā12 − b̄12 e−sτ · · · −ā1n − b̄1n e−sτ
⎜ −ā21 − b̄21 e−sτ s α + d2 − b̄22 e−sτ · · · −ā2n − b̄2n e−sτ ⎟
Δ(s) = ⎜
⎝
⎟ (13)
⎠
··· ··· ··· ···
−ān1 − b̄n1 e−sτ −ān2 − b̄n2 e−sτ · · · s α + dn − b̄nn e−sτ
with di = ci − γi − āii , (i = 1, 2, · · · , n) and

0 0
α−1 −sτ −st −sτ
k1 (s) = s δ1 (0) + b̄11 e e δ1 (t)dt + b̄12 e e−st δ2 (t)dt + · · ·
−τ −τ
0
+ b̄1n e−sτ e−st δn (t)dt,
−τ
0 0
k2 (s) = s α−1 δ2 (0) + b̄21 e−sτ e−st δ1 (t)dt + b̄22 e−sτ e−st δ2 (t)dt + · · ·
−τ −τ
0
+ b̄2n e−sτ e−st δn (t)dt,
−τ
···
0 0
kn (s) = s α−1 δn (0) + b̄n1 e−sτ e−st δ1 (t)dt + b̄n2 e−sτ e−st δ2 (t)dt + · · ·
−τ −τ
0
+ b̄nn e−sτ e−st δn (t)dt.
−τ
Now, τ = 0, the systems (10) becomes
D α w(t) = (−C + Γ + Ā + B̄)w(t) = M w(t), (14)
where,
⎛ ⎞
−c1 + γ1 + ā11 + b̄11 ā12 + b̄12 ··· ā1n + b̄1n
⎜ ā + b̄ −c + γ2 + ā22 + b̄22 ··· ā2n + b̄2n ⎟
M =⎜
⎝
21 21 2 ⎟.
⎠
··· ··· ··· ···
ān1 + b̄n1 ān2 + b̄n2 · · · −cn + γn + ānn + b̄nn
(15)
Suppose C = 0, Ā = 0 and Γ = 0, then (10) becomes the system in [6] and we

have the following results.
Theorem 1 ([6]) If all the roots of the characteristic equation det(Δ(s)) = 0 have
negative real parts, then the zero solution of (10) is Lyapunov asymptotically stable.
Theorem 2 ([6]) If α ∈ (0, 1), C = 0, Ā = 0, Γ = 0, all the eigenvalues of M

satisfy |arg(λ)| > απ
2 and the characteristic equation det(Δ(s)) = 0 has no pure
imaginary roots for τ > 0, then the zero solution of (10) is Lyapunov asymptotically
stable.
If α ∈ (0, 1), C = 0, Ā = 0, Γ = 0, then Theorem 2 is not valid to study the

stability of (10). Therefore, we consider the following Theorem.
Theorem 3 ([14]) If α ∈ (0, 1), all the eigenvalues of M satisfy |arg(λ)| > π2 and
the characteristic equation det(Δ(s)) = 0 has no pure imaginary roots for τ > 0,
then the zero solution of (10) is Lyapunov asymptotically stable.
3.1 The FNNs with Hub Structure and Time Delays
Hub structure is a common feature of neural networks, which is used to understand

the mechanism of complex recurrent networks. Now, consider the FNNs with hub
structure and time delays

D α x1 (t) = −c1 x1 (t) + nj=1 a1 j f j (x j (t)) + b1 f 1 (x1 (t − τ )),
D α xi (t) = −ci xi (t) + ai1 f 1 (x1 (t)) + aii f i (xi (t)) + b f i (xi (t − τ )), i = 2, · · · , n,
(16)
where ci > 0, the first neuron is the center of the hub and all the other i − 1 neurons
are connected directly only to the central neuron and to themselves.
The response system defined as

D α y1 (t) = −c1 y1 (t) + nj=1 a1 j f j (y j (t)) + b1 f 1 (y1 (t − τ )) + ψ1 (t),
D α yi (t) = −ci yi (t) + ai1 f 1 (y1 (t)) + aii f i (yi (t)) + b f i (yi (t − τ )) + ψi (t), i = 2, · · · , n.
(17)
The linear forms of equations (16) and (17) are
D α x(t) = −C x(t) + Āx(t) + B̄x(t − τ ), (18)

D α y(t) = −C y(t) + Āy(t) + B̄ y(t − τ ) + ψ(t). (19)
From (18) and (19), the error system as
D α w(t) = −Cw(t) + Āw(t) + B̄w(t − τ ) + Γ w(t), (20)

⎛ ⎞ ⎛ ⎞
ā11 ā12 ā13 ··· ā1n b̄1 0 0 ··· 0
⎜ ā21 ā22 0 ··· 0 ⎟ ⎜ ··· 0 ⎟
⎜
where C = diag(c1 , · · · , cn ), Ā = ⎝ ⎟ , B̄ = ⎜ 0 b̄ 0 ⎟.
··· ··· ··· ··· ··· ⎠ ⎝··· ··· ··· ··· ···⎠
ān1 0 0 ··· ānn 0 0 0 ··· b̄
Theorem 4 When α ∈ (0, 1), c̄2 − b̄ > 0, c̄1 + c̄2 − b̄1 − b̄ > 0, (c̄1 − b̄1 )(c̄2 −
n− Φ > 0, where c̄1 = c1 − γ1 − ā11 , c̄2 = ci − γi − āii , (i = 2, · · · , n), Φ =
b̄)
i=2 (āi1 ā1i ).
(i) if Φ = 0 and b̄2 − c̄22 sin2 απ 2 απ
2 < 0 and b̄1 − c̄1 sin 2 < 0, then the zero
2 2
solution of (20) is Lyapunov asymptotically stable;

(ii) if Φ = 0 and b̄2 − c̄22 sin2 απ
2 < 0, then the zero solution of (20) is Lyapunov
asymptotically stable.
Proof Taking the Laplace transform of equations (20) and using the same method
of finding for Δ(s) in the preliminaries, we have
⎛ ⎞
s α + c̄1 − b̄1 e−sτ −ā12 −ā13 ··· −ā1n
⎜ − ā s α + c̄ − b̄e−sτ 0 ··· 0 ⎟
⎜ 21 2 ⎟
Δ(s) = ⎜
⎜ −ā31 0 s α + c̄2 − b̄e−sτ ··· 0 ⎟ . (21)
⎟
⎝ ··· ··· ··· ··· ··· ⎠
−ān1 0 0 · · · s α + c̄2 − b̄e−sτ
Hence, Δ(s) is n × n matrix (n ≥ 3 in hub structure). Generally, characteristic

equation det(Δ(s)) = 0 satisfies
α n−2
det(Δ(s)) = s + c̄2 − b̄e−sτ

× (s α + c̄1 − b̄1 e−sτ )(s α + c̄2 − b̄e−sτ ) − Φ = 0. (22)

From (22), if Φ = 0 then s α + c̄2 − b̄e−sτ = 0 or s α + c̄1 − b̄1 e−sτ = 0, where
n
Φ = i=2 āi1 ā1i .
Next prove that det(Δ(s)) = 0 has no pure imaginary roots for any τ > 0.
Suppose that there exists s = ηi = |η|(cos π2 +i sin(± π2 ), that is a pure imaginary
root of s α + c̄2 − b̄e−sτ = 0, where η is a real number. If η > 0, s = ηi =
|η|(cos π2 + i sin( π2 )) and if η < 0, s = ηi = |η|(cos π2 − i sin( π2 )). Substituting
s = ηi = |η|(cos π2 + i sin(± π2 )) into s α + c̄2 − b̄e−sτ = 0 which gives
απ απ
|η|α cos + i sin(± ) + c̄2 − b̄(cos ητ − i sin ητ ) = 0. (23)
2 2
From (23), we separate the real and imaginary parts
απ
|η|α cos + c̄2 = b̄ cos ητ (24)
2
and
απ
|η|α sin(± ) = −b̄ sin ητ. (25)
2
Squaring and adding equations (24) and (25), one can obtain
απ
|η|2α + 2|η|α c̄2 cos + c̄22 − b̄2 = 0. (26)
2
Obviously, when 0 < α < 1 and b̄2 − c̄22 sin2 απ 2 < 0, the above equation (26) has
no real solutions, i.e., det(Δ(s)) = 0 has no pure imaginary roots for any τ > 0.
Similarly, if the s α + c̄1 − b̄1 e−sτ = 0, we have b̄12 − c̄12 sin2 απ
2 < 0. If Φ = 0 then
α
from (22), we have s +c̄2 −b̄e −sτ = 0 and (s +c̄1 −b̄1 e )(s α +c̄2 −b̄e−sτ )−Φ =
α −sτ
0. Clearly, we also derived b̄ − c̄2 sin2 απ

2 2
2 < 0. Hence, the conditions (i) and (ii) of
Theorem 4 is easily derived from the above results.
π
Moreover, prove that all the eigenvalues of M satisfy |arg(λ)| > 2. The coeffi-
cient matrix M of system (20) satisfies
⎛ ⎞
−c̄1 + b̄1 ā12 ā13 · · · ā1n
⎜ ā21 −c̄2 + b̄ 0 ··· 0 ⎟
M =⎜
⎝ ···
⎟.
··· ··· ··· ··· ⎠
ān1 0 0 · · · −c̄2 + b̄
Choose c̄2 − b̄ > 0, c̄1 + c̄2 − b̄1 − b̄ > 0, (c̄1 − b̄1 )(c̄2 − b̄)−Φ > 0, the eigenvalues
of M have negative real parts, i.e., all the eigenvalues of M satisfy |arg(λ)| > π2 .
Thus, the proof of the Theorem 4 is completed.
4 Numerical Example
In this section, a numerical example is given to show the effectiveness of our results.
Example 1 Consider the FNNs with hub structure and time delays drive system (16)
and response system (17) with the following parameter values, such as α = 0.95, τ =
0.03, c1 = 2, c2 = 10, c3 = 1, c4 = 2, a11 = 2, a22 = 1, a33 = 2, a44 =
5, a12 = −2, a21 = 0, a13 = 3, a31 = 0, a14 = 1, a41 = 0, b1 = −5, b = −12,
x(0) = (−0.5, 3.5, −0.3, 0.1)T and y(0) = (0.6, −2.0, 0.5, −0.3)T . The condition
(i) of Theorem 4 is satisfied for given parameter values. Thus, hybrid projective
synchronization between drive system (16) and response system (17) can be achieved
with Γ = diag(−4, −20, −10, −8). The convergence behavior of the error system
(20) and the state trajectories of system (16) and (17) are shown in Fig. 1 with the
scaling matrix β = diag(2.5, 2.5, 2.5, 2.5).
1 4
2
u (t), v (t)
0.5
u (t), v (t)
2 1.5
1
0
1
0
1
−0.5
e (t),e (t),e (t),e (t)
0.5
4
−1 −2 0
0 1 2 3 4 5 0 1 2 3 4 5
3
t t −0.5
2
0.5 0.4 −1
−1.5
1
u (t), v (t)
u (t), v (t)
0.2
3
−2
0 0
−2.5
3
−0.2
−3
−0.5 −0.4 0 1 2 3 4 5
0 1 2 3 4 5 0 1 2 3 4 5
t
t t
Fig. 1 Time responses and state trajectories of FNNs (16) and (17). The error state curves
of the hybrid projective synchronization between (16) and (17) with α = 0.95, Γ =
diag(−4, −20, −10, −8), τ = 0.03 and scaling matrix β = diag(2.5, 2.5, 2.5, 2.5)
5 Conclusion
In this paper, the hybrid projective synchronization of fractional-order neural net-

works with time delays have been extensively studied. First, we considered the
fractional-order neural networks with hub structure and time delays. Some new suffi-
cient conditions for hybrid projective synchronization of fractional-order neural net-
works with hub structure and time delays have been derived by using linear feedback
control, stability theorem of linear fractional order systems with multiple time delays
and appropriate scaling matrix. In addition, the scaling matrix β = I (β = −I ) where
I is the Identity matrix, then hybrid projective synchronization becomes complete
synchronization (anti-phase synchronization) of the considered network. Finally, a
numerical example is given to show the effectiveness of our main results.
Acknowledgments This work was supported by NBHM research project No. 2/48(7)/2012/
NBHM(R.P.)/R and D-II/12669.
References
1. Podlubny, I.: Fractional Differential Equations. Academic Press, New York (1999)
2. Koeller, R.C.: Application of fractional calculus to the theory of viscoelasticity. J. Appl. Mech.
51, 294–298 (1984)
3. Heaviside, O.: Electromagnetic Theory. Chelsea, New York (1971)
4. Petras, I.: A note on the fractional-order cellular neural networks. In: International Joint Con-
ference on Neural Networks, pp. 1021–1024 (2006)
5. Li, Y., Chen, Y., Podlubny, I.: Mittag-Leffler stability of fractional order nonlinear dynamic
systems. Automatica 45, 1965–1969 (2009)
6. Deng, W., Li, C.: Lü, J.: Stability analysis of linear fractional differential system with multiple
time delays. Nonlinear Dynam. 48, 409–416 (2007)
7. Shen, J., Lam, J.: Non-existence of finite-time stable equilibria in fractional-order nonlinear
systems. Automatica 50, 547–551 (2014)
8. Carpenter, G.A.: Neural network models for pattern recognition and associative memory. Neural
Netw. 2, 243–257 (1989)
9. Cochocki, A., Rolf, U.: Neural Networks for Optimization and Signal Processing, 1st edn.
Wiley, New York
10. Lundstrom, B., Higgs, M., Spain, W., Fairhall, A.: Fractional differentiation by neocortical
pyramidal neurons. Nature Neurosci. 11, 1335–1342 (2008)
11. Boroomand, A., Menhaj, M.: Fractional-order Hopfield neural networks. Lect. Notes Comput.
Sci. 5506, 883–890 (2009)
12. Chen, L., Chai, Y., Wu, R., Ma, T., Zhai, H.: Dynamic analysis fo a class of fractional-order
neural networks with delay. Neurocomputing 111, 190–194 (2013)
13. Chen, J., Zeng, Z., Jiang, P.: Global Mittag-Leffler stability and synchronization of memristor-
based fractional-order neural networks. Neural Netw. 51, 1–8 (2014)
14. Wang, H., Yu, Y., Wen, G.: Stability analysis of fractional-order Hopfield neural networks with
time delays. Neural Netw. 55, 98–109 (2014)
15. Perora, L.M., Carroll, T.L.: Synchronization in chaotic systems. Phys. Rev. Lett. 64, 821–824
(1990)
16. Zhu, H., He, Z.S., Zhou, S.B.: Lag synchronization of the fractional-order system via nonlinear
observer. Int. J. Mod. Phys. B 25, 3951–3964 (2011)
17. Taghvafard, H., Erjaee, G.H.: Phase and anti-phase synchronization of fractional order chaotic
systems via active control. Commun. Nonlinear Sci. Numer. Simul. 16, 4079–4088 (2011)
18. Wang, X.Y., He, Y.J.: Projective synchronization of fractional order chaotic system based on
linear separation. Phys. Lett. A 372, 435–441 (2008)
19. Kuntanapreeda, S.: Robust synchronization of fractional-order unified chaotic systems via
linear control. Comput. Math. Appl. 63, 183–190 (2012)
20. Mainieri, R., Rehacek, J.: Projective synchronization in three-dimensional chaotic systems.
Phys. Rev. Lett. 82, 3024–3045 (1999)
21. Chee, C.Y., Xu, D.: Chaos-based M-nary digital communication technique using controller
projective synchronization. IEE Proc. G (Circuits, Devices and Systems) 153, 357–360 (2006)
22. Wang, S., Yu, Y., Wen, G.: Hybrid projective synchronization of time-delayed fractional-order
chaotic systems. Nonlinear Anal. Hybrid Syst. 11, 129–138 (2014)
23. Yu, J., Hu, C., Jiang, H., Fan, X.: Projective synchronization for fractional neural networks.
Neural Netw. 49, 87–95 (2014)
24. Wang, S., Yu, Y.G., Diao, M.: Hybrid projective synchronization of chaotic fractional order
systems with different dimensions. Phys. A 389, 4981–4988 (2010)
25. Zhou, P., Zhu, W.: Function projective synchronization for fractional-order chaotic systems.
Nonlinear Anal. Real World Appl. 12, 811–16 (2011)
26. Wang, X.Y., Zhang, X.P., Ma, C.: Modified projective synchronization of fractional-order
chaotic systems via active sliding mode control. Nonlinear Dynam. 69, 511–17 (2012)
27. Bhalekar, S., Daftardar-Gejji, V.: A predictor-corrector scheme for solving nonlinear delay
differential equations of fractional order. J. Fract. Calc. Appl. 1, 1–8 (2011)
Approximations of Solutions of a Class
of Neutral Differential Equations
with a Deviated Argument
Pradeep Kumar, Dwijendra N. Pandey and D. Bahuguna
Abstract In this paper, we study the approximations of solutions to a class of non-

linear neutral differential equations with a deviated argument in a Hilbert space.
We consider an associated integral equation corresponding to the given problem
and a sequence of approximate integral equations. We establish the existence and
uniqueness of solutions to every approximate integral equation using the fixed point
theory. Then, we prove the convergence of the solutions of the approximate integral
equations to the solution of the associated integral equation. Next, we consider the
Faedo–Galerkin approximations of solutions and prove some convergence results.
Keywords Abstract differential equation with a deviated argument · Banach fixed

point theorem · Analytic semigroup · Faedo–Galerkin approximations
MSC: 34K30 · 34G20 · 47H06
1 Introduction
In the present study, we are concerned with the approximations of solutions to the
following class of neutral differential equation with a deviated argument in a sepa-
rable Hilbert space (H, ., (., .)):
P. Kumar (B) · D. Bahuguna

Department of Mathematics and Statistics, Indian Institute of Technology Kanpur,
Kanpur 208016, India
e-mail: prdipk@gmail.com
D. Bahuguna
D.N. Pandey
e-mail: dhiren@iitk.ac.in

DOI 10.1007/978-81-322-2485-3_54
658 P. Kumar et al.
d
[u(t) + g(t, u(t))] + A[u(t) + g(t, u(t))] = f (t, u(t), u(h(u(t), t))),
dt
u(0) = u 0 , 0 < t ≤ T < ∞. (1)
where A : D(A) ⊂ H → H is a close, densely defined, positive definite, self-

adjoint linear operator. Functions f, g, and h are suitably defined satisfying certain
conditions to be stated later.
Haloi et al. [1] studied the following neutral differential equation with a deviated
argument
d
[u(t) + g(t, u(a(t)))] + A(t)[u(t) + g(t, u(a(t)))] = f (t, u(t), u([h(u(t), t)])), t > 0,
dt
u(0) = u 0 . (2)
A(t), for each t ≥ 0, generates an analytic semigroup of bounded linear operators

on X. The nonlinear functions f, g, and h satisfy suitable growth conditions in their
arguments and a : [0, T ] → [0, T ] satisfies the delay property. In this paper, we
study Eq. (2) for an autonomous case, i.e., A(t) = A.
Ezzinibi et al. [2] have established the existence and stability of solution of the fol-
lowing nonlinear partial neutral functional differential equations with infinite delay:
d
[u(t) − g(t, u t )] = A[u(t) − g(t, u t )] + f (t, u t ), t ≥ 0,
dt
u 0 = ϕ ∈ C0 ,
where the operator A is the Hille–Yosida operator not necessarily densely defined
on the Banach space B. The functions g and f are continuous from [0, ∞) × C0
into B.
In the present work we are interested in the Faedo–Galerkin approximations of
solutions to (1). This technique basically uses the idea of finite-dimensional projec-
tions of solutions which gives rise to a sequence of approximate solutions. These
approximate solutions are then required to be shown to converge to the solution of
the problem under consideration. Initially, the Faedo–Galerkin approximations of
solutions to the particular case of (1) where g, h ≡ 0 and f (t, u) = M(u) has
been considered by Milleta [3]. For a nice introduction and related study of various
problems in this direction, we refer to (see [4–10]) and reference cited therein.
For the earlier works on existence and uniqueness of solutions to the differential
equations with deviating arguments, we refer to (see [1, 11–15]) and references cited
therein.
The plan of the paper is as follows: In the second section, we provide some of
the notations, notions, and results required for later sections. In the third section,
we consider an integral equation associated with (1) and then consider a sequence
of approximate integral equations and establish the existence and uniqueness of a
solution to each of the approximate integral equation. Also, we prove the convergence
of the solutions of the approximate integral equations and show that the limiting
Approximations of Solutions of a Class of Neutral Differential Equations … 659
function satisfies the associated integral equation. In the fourth section, we consider
the Faedo–Galerkin approximations of solutions and prove some convergence results
for such approximations. In the last section, we have given an example to show some
of the applications of the results obtained in the earlier sections.
In this section, we shall provide the assumptions, notations, notions, and related
results needed for the subsequent sections. We assume that the operator A satisfies
the following:
(H1) A is a closed, positive definite, self-adjoint, linear operator from the domain
D(A) ⊂ H of A into H such that D(A) is dense in H . Also we assume that
A has the pure point spectrum
0 < λ0 ≤ λ1 ≤ λ2 ≤ · · · , λm ≤ · · ·,
where λm → ∞ as m → ∞ and a corresponding complete orthonormal

system of eigenfunctions {u i }, i.e., Au i = λi u i and (u i, u j ) = δi j , where
δi j = 1 if i = j and zero otherwise.
These assumptions on A guarantee that −A generates an analytic semigroup, denoted
by S(t), t ≥ 0.
We mention some notions and preliminaries essential for our purpose. It is well
known that there exist constants M̃ ≥ 1 and ω ≥ 0 such that
S(t) ≤ M̃eωt , t ≥ 0.
Without loss of generality we may assume that S(t) is uniformly bounded by M,

i.e., S(t) ≤ M and 0 ∈ ρ(−A). In this case it is possible to define the fractional
power Aα for 0 ≤ α ≤ 1 as closed linear operator with domain D(Aα ) ⊆ H (cf. Pazy
[16], pp. 69–75 and p. 195). Furthermore, D(Aα ) is dense in H and the expression
xα = Aα x,
defines a norm on D(Aα ). Henceforth, we represent by Hα the space D(Aα ) endowed

with the norm · α . Also, for each α > 0, we define H−α = (Hα )∗ , the dual space
of Hα is a Banach space endowed with the norm x−α = A−α x.
Lemma 1 [16] Suppose that −A is the infinitesimal generator of an analytic semi-
group S(t), t ≥ 0 with S(t) ≤ M for t ≥ 0 and 0 ∈ ρ(−A). Then we have the
following properties:
(i) Hα is a Banach space for 0 ≤ α ≤ 1.
(ii) For 0 < δ ≤ α < 1, the embedding Hα → Hδ is continuous.
660 P. Kumar et al.
(iii) Aα commutes with S(t) and there exists a constant Cα > 0 depending on
0 ≤ α ≤ 1 such that
Aα S(t) ≤ Cα t −α , t > 0.
It can be seen easily that Ctα = C([0, t]; Hα ), for all t ∈ [0, T ], is a Banach space
endowed with the supremum norm,
ψt,α := sup ψ(η)α , ψ ∈ Ctα .

0≤η≤t
We set,
CTα−1 = C([0, T ]; Hα−1 ) = {y ∈ CTα : y(t) − y(s)α−1 ≤ L|t − s|, ∀ t, s ∈ [0, T ]},
where L is a suitable positive constant to be specified later.

We assume the following conditions:
(H2): Let U1 ⊂ Dom(f) is an open subset of R+ × Hα × Hα−1 and for each (t, u, v) ∈ U1
there is a neighborhood V1 ⊂ U1 of (t, u, v). The nonlinear map f : R+ × Hα × Hα−1 → H
satisfies the following condition:
f (t, x, ψ) − f (s, y, ψ̃) ≤ L f [|t − s|θ1 + x − yα + ψ − ψ̃α−1 ],
where 0 < θ1 ≤ 1, 0 ≤ α < 1, L f > 0 is a constant, (t, x, ψ), (s, y, ψ̃) ∈ V1 .

(H3): Let U2 ⊂ Dom(h) is an open subset of Hα × R+ and for each (x, t) ∈ U2 there is
a neighborhood V2 ⊂ U2 of (x, t). The map h : Hα × R+ → R+ satisfies the following
condition:
|h(x, t) − h(y, s)| ≤ L h [x − yα + |t − s|θ2 ],
where 0 < θ2 ≤ 1, 0 ≤ α < 1, L h > 0 is a constant, (x, t), (y, s) ∈ V2 and h(., 0) = 0.
(H4): Let U3 ⊂ Dom(g) is an open subset of [0, T ] × Hα−1 and for each (t, x) ∈ U3 there
is a neighborhood V3 ⊂ U3 of (x, t). There exist positive constants 0 < α < β < 1, such
that the function Aβ g is continuous for (t, u) ∈ [0, T0 ] × Hα−1 such that
Aβ g(t, x) − Aβ g(s, y) ≤ L g {|t − s| + x − yα−1 }, and

4L g Aα−β−1 = η < 1
where L g , η > 0 are positive constants and (x, t), (y, s) ∈ V3 .
3 Approximate Solutions and Convergence
The existence of a solution to (1) is closely related to the following integral equation
(3):
Definition 1 A continuous function u : [0, T ] → H is said to be a mild solution of
Eq. (1) if u is the solution of the following integral equation:
u(t) = S(t)[u(0) + g(0, u 0 )] − g(t, u(t))

t
+ S(t − s) f (s, u(s), u(h(u(s), s)))ds, t ∈ [0, T ] (3)
0
and satisfies the initial condition u(0) = u 0 .
Definition 2 By a solution of the problem (1), we mean a function u : [0, T ] → H

satisfying the following three conditions:
(i) u(.) + g(., u(.)) ∈ CTα−1 ∩ C 1 ((0, T ), H ) ∩ C([0, T ], H ),
(ii) u(t) ∈ D(A), and (t, u(t), u(h(u(t), t))) ∈ U1 ,
d
(iii) dt [u(t) + g(t, u(t))] + A[u(t) + g(t, u(t))] = f (t, u(t), u(h(u(t), t))), for all
t ∈ (0, T ],
(iv) u(0) = u 0 .
Let Hn ⊆ H denote the finite-dimensional subspace spanned by {u 0 , u 1 , . . . , u n }

and let P n : H −→ Hn be the corresponding projection operator for n = 0, 1, 2, . . ..
We define
gn : R+ × H −→ H as gn (t, u(t)) = g(t, P n u(t)). (4)
Also, we define
f n : R+ × H × H −→ H
given by
f n (s, u(s), u(h(u(s), s))) = f (s, P n u(s), P n u[h n (u(s), s)]). (5)
For a fixed R > 0, we choose 0 < T0 = T0 (α, β, u 0 ) ≤ T such that
T01−α
Cα L f [2 + L L h ] ≤1−η (6)
1−α
where η = 4L g Aα−β−1 < 1,
1
R 1−α
T0 ≤ (1 − α)(Cα [2 + L L h ]L f )−1 ,
4
and satisfying the following
R
(S(t) − I )Aα [u 0 + gn (0, u 0 )] + Aα−β L g [T0 + R] ≤ , (7)
2
for all t ∈ [0, T0 ].
662 P. Kumar et al.
T01−α R
Cα N ≤ . (8)
1−α 2
For more details of choosing such a T0 , we refer Theorem 2.2 of [15].

We set
W = {u ∈ CTα0 ∩ CTα−1
0
: u(0) = u 0 , u − u 0 T0 ,α ≤ R}.
Clearly, W is a closed and bounded subset of CTα−1

0
.
We define a map Fn : W → W given by
t
(Fn u)(t) = S(t)[u 0 + gn (0, u 0 )] − gn (t, u(t)) + S(t − s)
0
× f n (s, u(s), u(h(u(s), s)))ds, t ∈ [0, T0 ].
Theorem 1 Let us assume that the assumptions (H1)–(H4) are satisfied and u 0 ∈
D(Aα ) for 0 ≤ α < 1. Then there exists a unique u n ∈ CTα−1 0
∩ CTα0 such that
Fn u n = u n for each n = 0, 1, 2, . . . , i.e., u n satisfies the approximate integral
equation
u n (t) = S(t)[u 0 + gn (0, u 0 )] − gn (t, u(t))

t
+ S(t − s) f n (s, u(s), u(h(u(s), s)))ds, t ∈ [0, T0 ].
0
Proof In order to prove this theorem first we need to show that Fn u ∈ CTα−1
0
for any
α−1 α α
u ∈ CT0 . Clearly, Fn : CT0 → CT0 .
If u ∈ CTα−1
0
, T0 > t2 > t1 > 0, and 0 ≤ α < 1, then we get
(Fn u)(t2 ) − (Fn u)(t1 )α−1

≤ (S(t2 ) − S(t1 ))(u 0 + gn (0, u 0 ))α−1
+ Aα−1−β Aβ gn (t2 , u(t2 )) − Aβ gn (t1 , u(t1 ))
t1
+ (S(t2 − s) − S(t1 − s))Aα−1 f n (s, u(s), u(h(u(s), s)))ds
0
t2
+ S(t2 − s)Aα−1 f n (s, u(s), u(h(u(s), s)))ds. (9)
t1
The continuity of u ∈ CTα0 and the assumptions (H2)–(H3) will imply that the function
f (s, u(s), u(h(u(s), s))) is continuous on [0, T0 ]. Therefore, we can easily show that
there exists a positive constant N such that
f (s, u(s), u(h(u(s), s))) ≤ N = L f [T0 θ1 + R(1 + L L h ) + L L h T0θ2 ] + N0 ,
where N0 = f (0, u 0 , u 0 ).

Similarly, we have
Aβ g(t, u(t)) ≤ L g [T0 + R] + g(0, u 0 )β = N1 .
For the first part of right-hand side of (9), we have,
(S(t2 ) − S(t1 ))(u 0 + gn (0, u 0 ))α−1

t2
≤ Aα−1 S (s)(u 0 + gn (0, u 0 ))ds
t1
t2
= Aα S(s)(u 0 + gn (0, u 0 ))ds
t1
t2
≤ S(s)[u 0 α + Aα−β gn (0, u 0 )β ]ds
t1
≤ C1 (t2 − t1 ), (10)
where C1 = [u 0 α + Aα−β gn (0, u 0 )β ]M.

For the second part of right-hand side of (9), we can see that
Aα−β−1 Aβ gn (t2 , u(t2 )) − Aβ gn (t1 , u(t1 ))

≤ Aα−β−1 L g [(t2 − t1 ) + u(t2 ) − u(t1 )α−1 ]
≤ Aα−β−1 [L g (1 + L)](t2 − t1 )
≤ C2 (t2 − t1 ). (11)
where C2 = Aα−β−1 [L g (1 + L)]. To handle the third part of the right-hand side
of (9), observe that,
t2 −t1
(S(t2 − s) − S(t1 − s))α−1 ≤ Aα−1 S (l)S(t1 − s)dl
0
t2 −t1
≤ S(l)Aα S(t1 − s)dl
0
≤ MCα (t2 − t1 )(t1 − s)−α . (12)
664 P. Kumar et al.
Now we use the inequality (12) to get the bound for third part as given below,
t1
(S(t2 − s) − S(t1 − s))Aα−1 f n (s, u(s), u(h(u(s), s)))ds
0
≤ C3 (t2 − t1 ), (13)
T01−α
where C3 = NMCα 1−α .
For the bound for fourth part, we have,
t2
S(t2 − s)Aα−1 f n (s, u(s), u(h(u(s), s)))ds ≤ C4 (t2 − t1 ), (14)
t1
where C4 = Aα−1 M N .
We use the inequalities (10), (11), (13)–(14) in inequality (9) to get the following
inequality:
(Fn u)(t2 ) − (Fn u)(t1 )α−1 ≤ L|t2 − t1 |,
where, L = max{Ci , i = 1, 2, . . . 4}. Hence, Fn : CTα−1

0
→ CTα−1
0
follows.
Our next task is to show that Fn : W → W . Now, for t ∈ [0, T0 ] and u ∈ W ,
we have
(Fn u)(t) − u 0 α
≤ (S(t) − I )Aα [u 0 + gn (0, u 0 )]
+ Aα−β Aβ gn (s, u(s))) − Aβ gn (0, u(0))
t
+ S(t − s)Aα f n (s, u(s), u(h(u(s), s))])ds
0
≤ (S(t) − I )Aα [u 0 + gn (0, u 0 )] + Aα−β L g [T0 + R]
T01−α
+ Cα N .
1−α
Hence, from inequalities (7) and (8), we get
Fn u − u 0 T0 ,α ≤ R.
Therefore, Fn : W → W .
Now, if t ∈ [0, T0 ] and u, v ∈ W , then
(Fn u)(t) − (Fn v)(t)α

≤ Aα−β Aβ gn (t, u(s)) − Aβ gn (t, v(s))
t
+ S(t − s)Aα f n (s, u(s), u(h[u(s), s]))
0
− f n (s, v(s), v(h[v(s), s])))ds. (15)
We have the following inequalities,
Aβ gn (s, u(s))) − Aβ gn (t, v(t)) ≤ L g A−1 u − vT0 ,α , (16)
f n (s, u(s), u(h(u(s), s))) − f n (s, v(s), v[h(v(s), s)])

≤ L f [2 + L L h ]u − vT0 ,α . (17)
We use the inequalities (16) and (17) in the inequality (15) and get
β−α
T0
(Fn u)(t) − (Fn v)(t)α ≤ L g Aα−β−1 + C1+α−β L g )
β −α

T 1−α
+ Cα L f [2 + L L h ] 0 u − vT0 ,α .
1−α
Hence from inequality (6), we get the following inequality given below
Fn u − Fn vT0 ,α < u − vT0 ,α .
Therefore, the map Fn has a unique fixed point u n ∈ W which is given by,
u n (t) = S(t)[u 0 + gn (0, u 0 )] − gn (t, u(t))

t
+ S(t − s) f n (s, u(s), u(h(u(s), s)))ds t ∈ [0, T0 ]. (18)
0
Hence, the mild solution u n of Eq. (1) is given by the Eq. (18) and belongs to W ,
hence, the theorem is proved.
Lemma 2 Let (H1)–(H4) hold. If u 0 ∈ D(Aα ) then u n (t) ∈ D(Aϑ ), for all t ∈
(0, T ] where 0 ≤ ϑ ≤ β < 1. Furthermore, if u 0 ∈ D(A) then u n (t) ∈ D(Aϑ ), for
all t ∈ [0, T ] where 0 ≤ ϑ ≤ β < 1.
666 P. Kumar et al.
Proof From Theorem 1, we have the existence of a unique u n ∈ CTα0 ∩ CTα−10

sat-
isfying (18). Part (a) of Theorem 2.6.13 in Pazy [16] implies that for t > 0 and
0 ≤ ϑ < 1, S(t) : H → D(Aϑ ) and for 0 ≤ ϑ ≤ β < 1, D(Aβ ) ⊆ D(Aϑ ).
(H2)–(H4) implies that the map t → Aβ g(t, u n (t) is Hölder continuous on [0, T ]
with the exponent ρ = min{γ , ϑ} since the Hölder continuity of u n can be easily
established using the similar arguments from (9) to (13). Also from Theorem 1.2.4
in Pazy [16], we have S(t)x ∈ D(A) if x ∈ D(A). The required result follows from
these facts and the fact that D(A) ⊆ D(Aϑ ) for 0 ≤ ϑ ≤ 1.
Lemma 3 Let (H1) and (H2) hold. If u 0 ∈ D(Aα ) and t0 ∈ (0, T0 ] then
u n (t)ϑ ≤ Ut0 , α < ϑ < β, t ∈ [t0 , T0 ], n = 1, 2, . . . ,
for some constant Ut0 , dependent of t0 and
u n (t)ϑ ≤ U0 , 0 < ϑ ≤ α, t ∈ (0, T0 ], n = 1, 2, . . . ,
for some constant U0 . Moreover, if u 0 ∈ D(A), then there exists a constant U0 , such
that
u n (t)ϑ ≤ U0 , 0 < ϑ < β, t ∈ [0, T0 ], n = 1, 2, . . . .
Proof First, we assume that u 0 ∈ D(Aα ). Applying Aϑ on both the sides of (18)
and using (iii) of Lemma 1, for t ∈ [t0 , T ] and α < ϑ < β, we have
u n (t)ϑ ≤ Aϑ S(t)(u 0 + gn (0, u 0 ) + Aϑ−β Aβ gn (t, u n (t))

t
+ S(t − s)Aϑ f n (s, u n (s), u n [h(u n (s), s)])ds
0
≤ Cϑ t0−ϑ (u 0 + gn (0, u 0 ) + Aϑ−β N1
T 1−ϑ
+ Cϑ N ≤ U t0 .
1−ϑ
Again, for t ∈ (0, T0 ] and 0 < ϑ ≤ α, u 0 ∈ D(Aϑ ) and
u n (t)ϑ ≤ M(Aϑ u 0 + gn (0, u˜0 ϑ ) + Aϑ−β N1

T 1−ϑ
+ Cϑ N ≤ U0 .
1−ϑ
Furthermore, If u 0 ∈ D(A) then u 0 ∈ D(Aϑ ) for 0 < ϑ ≤ 1 and we can easily get
the required estimate. This completes the proof of the proposition.
In this section, we establish the convergence of the solution u n ∈ X α (T ) of the

approximate integral equation (18) to a unique solution u of (3).
Theorem 2 Let (H1)–(H4) hold. If u 0 ∈ D(Aα ), then for any t0 ∈ (0, T0 ],

m→∞ {n≥m, t ≤t≤T }
0
Proof Let 0 < α < ϑ < β. For n ≥ m, we have
f n (t, u n (t), u n [h(u n (t), t)]) − f m (t, u m (t), u m [h(u m (t), t)])
≤ f n (t, u n (t), u n [h(u n (t), t)]) − f n (t, u m (t), u m [h(u m (t), t)])
+ f n (t, u m (t), u m [h(u m (t), t)]) − f m (t, u m (t), u m [h(u m (t), t)])
≤ L f (2 + L L h )u n (t) − u m (t)α + L f [(P n − P m )u m (t)α
+ A−1 (P n − P m )u m (h(u m (t), t))α ].
Also,
1
(P n − P m )u m (t)α ≤ Aα−ϑ (P n − P m )Aϑ u m (t) ≤ Aϑ u m (t).
λϑ−α
m
Thus, we have
f n (t, u n (t), u n [h(u n (t), t)]) − f m (t, u m (t), u m [h(u m (t), t)])
1
≤ L f (2 + L L h )u n (t) − u m (t)α + L f ϑ−α Aϑ u m (t)
λm
A −1
+ ϑ−α Aϑ u m (h(u m (t), t)) .

λm
Similarly,
Aβ gn (t, u n (t)) − Aβ gm (t, u m (t))

≤ Aβ gn (t, u n (t)) − Aβ gn (t, u m (t)) + Aβ gn (t, u m (t)) − Aβ gm (t, u m (t))

−1 1 ϑ
≤ L g A u n (t) − u m (t)α + ϑ−α A u m (t) .
λm
Now, for 0 < t0 < t0 , we may write

668 P. Kumar et al.
u n (t) − u m (t)α
≤ S(t)Aα (gn (0, u 0 − gm (0, u 0 ) + Aα−β Aβ gn (t, u n (t)) − Aβ gm (t, u m (t))
t0 t
+ + Aα S(t − s) f n (s, u n (s), u n [h(u n (s), s)])
0 t0
− f m (s, u m (s), u m [h(u m (s), s)])ds.
We estimate the first term as
S(t)Aα (gn (0, u 0 − gm (0, u 0 ) ≤ MAα−β Aβ g(0, P n u 0 ) − Aβ g(0, P m u 0 )

≤ MAα−β−1 L g (P n − P m )Aα u 0 .
The first and the third integrals are estimated as
t0
Aα S(t − s) f n (s, u n (s),u n [h(u n (s), s)]) − f m (s, u m (s), u m [h(u m (s), s)])ds
0
≤ 2Cα N (t0 − t0 )−α t0 .
For the second and the fourth integrals, we have

t
Aα S(t − s) f n (s, u n (s), u n [h(u n (s), s)]) − f m (s, u m (s), u m [h(u m (s), s)])ds
t0
t 1
≤ Cα L f (t − s)−α (2 + L L h )u n (s) − u m (s)α + ϑ−α
Aϑ u m (s)
t0 λm
A−1
+ ϑ−α
Aϑ u m (h(u m (t), t)) ds
λm
Ut T0 1−α t
≤ Cα L f (1 + A−1 ) ϑ−α0 + (2 + L L h ) (t − s)−α u n (s) − u m (s)α ds .
λm (1 − α) t0
Therefore,
u n (t) − u m (t)α ≤MAα−β−1 L g (P n − P m )Aα u 0

α−β−1
Ut0
+ A L g u n (t) − u m (t)α + ϑ−α
λm

Cα N U t
+2 t + Dα ϑ−α 0
(t0 − t0 )α 0 λm
t
Cα L f (2 + L L h )
+ u n (s) − u m (s)α ds,
t0 (t − s)α
where
T0 1−α
Dα = Cα L f (1 + A−1 ) .
1−α
Since Aα−β−1 L g < 1, we have
1 Ut
u n (t) − u m (t)α ≤ M(P n − P m )Aα u + Aα−β−1 L 0
0 g
(1 − Aα−β−1 L g ) λϑ−α
m

Cα N Ut
+2 α
t0 + Dα ϑ−α0
(t0 − t0 ) λm
t
Cα L f (2 + L L h )
+ u n (s) − u m (s)α ds.
t0 (t − s)α
Lemma 5.6.7 in [16] implies that there exists a constant C such that
u n (t) − u m (t)α
1 Ut0
≤ M(P n − P m )Aα u 0 + (Aα−β−1 L g + Dα )
(1 − Aα−β−1 L g ) λϑ−α
m

Cα N
+2 t C.
(t0 − t0 )α 0
Taking supremum over [t0 , T ] and letting m → ∞, we obtain
lim sup u n (t) − u m (t)α

m→∞ {n≥m,t∈[t ,T ]}
0
2
Cα N
≤ C.
(1 − Aα−β−1 L g ) (t0 − t0 )α
As t0 is arbitrary, the right-hand side may be made as small as desired by taking t0
sufficiently small. This completes the proof of the proposition.
Corollary 1 If u 0 ∈ D(A) then

m→∞ {n≥m, 0≤t≤T }
With the help of Theorems 1 and 2, the convergence of the solutions u n (t) of the
approximate integral equations (18) follows from the next result.
Theorem 3 Let (H1)–(H4) hold and let u 0 ∈ D(Aα ). Then there exists a unique
function u n ∈ W
670 P. Kumar et al.
u n (t) = S(t)[u(0) + gn (0, u 0 )] − gn (t, u(t))

t
+ S(t − s) f n (s, u n (s), u n [h n (u n (s), s)])ds, t ∈ [0, T0 ]
0
and u ∈ W
u(t) = S(t)[u(0) + g(0, u 0 )] − g(t, u(t))

t
+ S(t − s) f (s, u(s), u(h(u(s), s)))ds, t ∈ [0, T0 ]
0
such that u n → u as n → ∞ in W and u satisfies (3) on [0, T0 ].
5 Faedo–Galerkin Approximations
In this section, we will discuss the Faedo–Galerkin Approximations of solutions and

prove the convergence results for such approximations.
For any 0 < t < T0 , we have a unique u ∈ W satisfying the integral equation
u(t) = S(t)[u(0) + g(0, u 0 )] − g(t, u(t))

t
+ S(t − s) f (s, u(s), u(h(u(s), s)))ds.
0
Then it has the representation

∞

u(t) = αi (t)φi , αi (t) = (u(t), φi ), i = 0, 1, . . . (19)
i=0
where φi’ s are defined in (H 1).

Also, we have a unique solution u n ∈ Hα (T0 ) from the approximate integral
equations
u n (t) = S(t)[u(0) + gn (0, u 0 )] − gn (t, u n (t))

t
+ S(t − s) f n (s, u n (s), u n (h(u n (s), s)))ds. (20)
0
Let P n u n (t) = û n (t) is the orthogonal projection of (20) on the first n elements of
{φi } satisfying the following equations:
P n u n (t) = S(t)[P n u(0) + P n gn (0, u 0 )] − P n gn (t, u n (t))

t
+ S(t − s)P n f n (s, u n (s), u n (h(u n (s), s)))ds,
0
by using (4) and (5), we get the following
û n (t) = S(t)[P n u(0) + P n g(0, u 0 )] − P n g(t, û n (t))

t
+ S(t − s)P n f (s, û n (s), û n (h(û n (s), s)))dθ ds. (21)
0
The solution û n of (21) has the following representation:

n
û n (t) = αin (t)φi , αin (t) = (û n (t), φi ), i = 0, 1, . . . (22)
i=0
Then we get a system of equations from (21) and (22)
dβ n
α (t) + Hin (t, α0n , α1n , . . . , αnn ) + λi [αin (t) + Hin (t, α0n , α1n , . . . , αnn )]
dt β i
= Fin (t, α0n , α1n , · · · , αnn , τ0n , τ1n , · · · , τnn )
αin (0) = u i ,
where

n
n
Fin = f t, αin φi , τin φi , φi ,
i=0 i=0

n
Hin = g t, αin φi , φi ,
i=0

τin = αin h α0n , α1n , · · · , αnn , t
and u i = (u 0 , φi ) for i = 1, 2, · · · , n.
Convergence of αin (t) → αi (t) follows from following theorem and the fact that:
∞
∞

α α
A [u(t) − û n (t)] = A (αi (t) − αin (t))φi = λiα (αi (t) − αin (t))φi .
i=0 i=0
Thus, we have

n
α
A [u(t) − û n (t)] ≥ 2
λi2α (αi (t) − αin (t))2 .
i=0
672 P. Kumar et al.
Theorem 4 Let (H1) and (H2) hold. Then we have the following:
(a) If u 0 ∈ D(Aα ), then for any 0 < t0 ≤ T0 ,

n
lim sup λi2α (αi (t) − αin (t))2 = 0.
n→∞ t ≤t≤T
0 0 i=0
(b) If u 0 ∈ D(A), then

n

n→∞ 0≤t≤T
0 i=0
The assertion of this theorem follows from the facts mentioned above and the fol-
lowing result:
Proposition 1 Let (H1) and (H2) hold and let T be any number such that 0 < T <
tmax , then we have the following:
(a) If u 0 ∈ D(Aα ), then for any 0 < t0 ≤ T0 ,
lim sup Aα [û n (t) − û m (t)] = 0.

n→∞ {n≥m,t ≤t≤T }
0 0
(b) If u 0 ∈ D(A), then
lim sup Aα [û n (t) − û m (t)] = 0.

n→∞ {n≥m,0≤t≤T }
0
Proof For n ≥ m, we have
Aα [û n (t) − û m (t)] = Aα [P n u n (t) − P m u m (t)]

≤ P n [u n (t) − u m (t)]α + (P n − P m )u m α
1
≤ u n (t) − u m (t)α + ϑ−α Aϑ u m .
λm
If u 0 ∈ D(Aα ) then the result in (a) follows from Proposition 2. If u 0 ∈ D(A), (b)
follows from Corollary 1.
Theorem 5 Let (H1)–(H4) hold and let u 0 ∈ D(Aα ). Then there exists a unique
function û n ∈ W
û n (t) = S(t)[u(0) + gn (0, u 0 )] − gn (t, û n (t))

t
+ S(t − s) f n (s, û n (s), û n [h n (û n (s), s)])ds, t ∈ [0, T0 ]
0
and u ∈ W
u(t) = S(t)[u(0) + g(0, u 0 )] − g(t, u(t))

t
+ S(t − s) f (s, u(s), u(h(u(s), s)))ds, t ∈ [0, T0 ]
0
such that û n → u as n → ∞ in W and u satisfies (3) on [0, T0 ].
6 Examples
Let X = L 2 (0, 1). We consider the following partial differential equations with a
deviated argument,
⎧
⎪
⎪ ∂ [w(t, x) + ∂x f 1 (t, w(t, x))] − ∂x2 [w(t, x) + ∂x f 1 (t, w(t, x))]
⎨ t
= f 2 (x, w(t, x)), + f 3 (t, x, w(t, x)), x ∈ (0, 1), t > 0,
(23)
⎪ w(t, 0) = w(t, 1) = 0, t ∈ [0, T ], 0 < T < ∞,
⎪
⎩
w(0, x) = u 0 , x ∈ (0, 1),
where
x
f 2 (x, w(t, x)) = K (x, s)w(s, h(t)(a1 |w(s, t)| + b1 |ws (s, t)|))ds.
0
The function f 3 : R+ ×[0, 1]×R → R is measurable in x, locally Hölder continuous

in t, locally Lipschitz continuous in u and uniformly in x. Further, we assume that
a1 , b1 ≥ 0, (a1 , b1 ) = (0, 0), h : R+ → R+ is locally Hölder continuous in t with
h(0) = 0 and K : [0, 1] × [0, 1] → R.
We define an operator A as follows:
Au = −u with u ∈ D(A) = {u ∈ H01 (0, 1) ∩ H 2 (0, 1) : u ∈ X }. (24)
Here clearly the operator A is self-adjoint with compact resolvent and is the
infinitesimal generator of an analytic semigroup S(t). Now we take α = 1/2,
D(A1/2 ) = H01 (0, 1) is the Banach space endowed with the norm,
x1/2 := A1/2 x, x ∈ D(A1/2 )
and we denote this space by X 1/2 . Also, for t ∈ [0, T ], we denote
1/2
Ct = C([0, t]; D(A1/2 )),
674 P. Kumar et al.
endowed with the sup norm
1/2
ψt,1/2 := sup ψ(η)α , ψ ∈ Ct .
0≤η≤t
We observe some properties of the operators A and A1/2 defined by (24). For u ∈
D(A) and λ ∈ R, with Au = −u = λu, we have Au, u = λu, u; that is,

−u , u = |u |2L 2 = λ|u|2L 2
so λ > 0. A solution u of Au = λu is of the form

√ √
u(x) = C cos( λx) + D sin( λx)
and the conditions u(0) = u(1) = 0 imply that C = 0 and λ = λn = n 2 π 2 , n ∈ N.

Thus, for each n ∈ N, the corresponding solution is given by

u n (x) = D sin( λn x).
√
We have u n , u m = 0 for n = m and u n , u n = 1 and hence D = 2. For
u ∈ D(A), there exists a sequence of real numbers {αn } such that

u(x) = αn u n (x), (αn )2 < +∞ and (λn )2 (αn )2 < +∞.
n∈N n∈N n∈N
We have
A1/2 u(x) = λn αn u n (x)
n∈N

with u ∈ D(A1/2 ); that is, n∈N λn (αn )
< +∞. X − 1 = H 1 (0, 1) is a Sobolev
2
2
space of negative index with the equivalent norm .− 1 = ∞ n=1 |., u n | . For more
2
2
details on the Sobolev space of negative index, we refer to Gal [15].
The Eq. (23) can be reformulated as the following abstract equation in X =
L 2 (0, 1):
d
[u(t) + g(t, u(t))] + A[u(t) + g(t, u(t))] = f (t, u(t), u(h(u(t), t))) t > 0,
dt
u(0) = u 0 , (25)
where u(t) = w(t, .) that is u(t)(x) = w(t, x), x ∈ (0, 1). The function g : R+ ×
X 1/2 → X , such that g(t, u(t))(x) = ∂x f 1 (t, w(t, x)) and the operator A is same
as in Eq. (24).
The function f : R+ × X 1/2 × X −1/2 → X , is given by
f (t, ψ, ξ )(x) = f 2 (x, ξ ) + f 3 (t, x, ψ), (26)
where f 2 : [0, 1] × X → H01 (0, 1) is given by

x
f 2 (t, ξ ) = K (x, y)ξ(y)dy, (27)
0
and f 3 : R × [0, 1] × H 2 (0, 1) → H01 (0, 1) satisfies the following:
f 3 (t, x, ψ) ≤ Q(x, t)(1 + ψ H 2 (0,1) ) (28)
with Q(., t) ∈ X and Q is continuous in its second argument. We can easily verify
that the function f satisfied the assumptions (H1)–(H4). For more details see [15].
Acknowledgments We highly appreciate the valuable comments and suggestions of the referees
on our manuscript, which helped to considerably improve the quality of the manuscript. The first
author would like to thank Prof. Shobha Madan, HOD, Mathematics and Statistics, IIT Kanpur for
providing financial support to attend the conference. Also, the first and third authors would like to
acknowledge the financial aid from the Department of Science and Technology, New Delhi, under
its research project SR/S4/MS:796/12.
References
1. Haloi, R., Pandey, D.N., Bahuguna, D.: Existence of solutions to a non-autonomous abstract
neutral differential equation with deviated argument. J. Nonl. Evol. Equ. Appl. 5, 75–90 (2011)
2. Adimy, M., Bouzahir, H., Ezzinbi, K.: Existence and stability for some partial neutral functional
differential equations with infinite delay. J. Math. Anal. Appl. 294, 438–461 (2004)
3. Miletta, P.D.: Approximation of solutions to evolution equations. Math. Meth. Appl. Sci. 17,
753–763 (1994)
4. Bazley, N.W.: Approximation of wave equations with reproducing nonlinearities. Nonlinear
Anal. 3, 539–546 (1979)
5. Bazley, N.W.: Approximation of operators with reproducing nonlinearities. Manuscripta Math.
18, 353–369 (1976)
6. Bahuguna, D., Muslim, M.: A study of nonlocal history-valued retarded differential equations
using analytic semigroups. Nonlinear Dyn. Syst. Theory 6, 63–75 (2006)
7. Bahuguna, D., Shukla, R.: Approximations of solutions to nonlinear sobolev type evolution
equation. Elec. J. Diff. Equ. 31, 1–16 (2003)
8. Bahuguna, D., Srivastava, S.K., Singh, S.: Approximations of solutions to semilinear integrod-
ifferential equations. Numer. Funct. Anal. Optim. 22, 487–504 (2001)
9. Bahuguna, D., Muslim, M.: Approximation of solutions to retarded differential equations with
applications to population dynamics. J. Appl. Math. Stoch. Anal. 1, 1–11 (2005)
10. Muslim, M.: Approximation of solutions to History-valued neutral functional differential equa-
tions. Comput. Math. Appl. 51, 537–550 (2006)
11. El’sgol’ts, L.E., Norkin, S.B.: Introduction to the Theory of Differential Equations with Devi-
ating Arguments. Academic Press (1973)
676 P. Kumar et al.
12. Driver, R.D.: A two-body problem of classical electrodynamics: the one-dimensional case.
Ann. Phys. 21, 122–142 (1963)
13. Oberg, R.J.: On the local existence of solutions of certain functional-differential equations.
Proc. Amer. Math. Soc. 20, 295–302 (1969)
14. Ezzinbi, K., Xianlong, Fu, Hilal, K.: Existence and regularity in the α-norm for some neutral
partial differential equations with nonlocal conditions. Nonlinear Anal. 67, 1613–1622 (2007)
15. Gal, C.G.: Nonlinear abstract differential equations with deviated argument. J. Math. Anal.
Appl. 333, 177–189 (2007)
Springer (1983)
Approximation of Solutions
to Fractional Integro-Differential Equations
with Finite Delay
Renu Chaudhary and Dwijendra N. Pandey
Abstract In the present work, we study the approximation of solutions to fractional

integro-differential equations with finite delay in an arbitrary separable Hilbert space
H . We consider an associated sequence of approximate integral equations. We estab-
lish the existence and uniqueness of the solutions to each approximate integral equa-
tion using the fixed-point arguments. Then we prove the convergence of the solutions
of the approximate integral equations to the solution of the given fractional integro-
differential equation. Finally, we consider the Faedo–Galerkin approximations of the
solutions and prove some convergence results.
Keywords Analytic semigroup · Banach fixed point theorem · Faedo–Galerkin

approximations · Fractional integro-differential equation · Mild solution
1 Introduction
The investigation of fractional differential equations, that is, calculus of derivatives

of any arbitrary real or complex order, has gained importance and popularity during
the past three decades. In various problems of physics, mechanics and engineering,
fractional differential equations have been proved to be a valuable tool in the mod-
elling of many phenomena. These differential equations are also very important to
describe the memory and hereditary properties of various materials and phenomenon.
More details on the theory and its applications we refer to Kilbas and Trujillo [1],
Lakshmikantham [2], Miller and Ross [3], Podlubny [4], and Kilbas and Samko [5].
R. Chaudhary (B) · D.N. Pandey

Roorkee 247667, Uttarakhand, India
e-mail: rrenu94@gmail.com
D.N. Pandey

DOI 10.1007/978-81-322-2485-3_55
678 R. Chaudhary and D.N. Pandey
In the present work we are concerned with the approximation of solution to fol-
lowing fractional integro-differential equation with finite delay in a separable Hilbert
space (H, . , (, ))
q
Dt [u(t) + f (t, u t )] + Au(t) = g(t, u t ) + K (u)(t), t ∈ (0, T ], (1)
u(t) = φ(t), t ∈ [−h, 0], 0 < h < ∞, (2)
t
where K is the nonlinear Volterra operator K (u)(t) = 0 M(t − s)k(u)(s)ds, 0 <
q
T < ∞, Dt is Caputo fractional derivative, where 0 < q < 1. A : D(A) ⊂
H → H is closed, positive definite and self-adjoint linear operator with densely
defined domain D(A) which is the infinitesimal generator of an analytic semigroup
of bounded linear operator on H . The nonlinear operator k is defined on D(Aα )
for some α. The map M is a real-valued continuous function defined on R+ . The
nonlinear functions f and g are defined from [0, T ] × C0 into H and φ ∈ C0 . Here,
C0 : = C([−h, 0]; H ) be the Banach space of all continuous functions from [−h, 0]
into H endowed with the supremum norm u0 := sup u(s), u ∈ C0 . Further,
−h≤s≤0
for t ∈ [0, T ], Ct := C([−h, t]; H ) be the Banach space of all continuous functions
from [−h, t] into H endowed with supremum norm ut := sup u(s), u ∈ Ct .
−h≤s≤t
For u ∈ C T and 0 ≤ t ≤ T , u t ∈ C0 be the function defined by u t (θ ) = u(t +
θ ) for θ ∈ [−h, 0].
Initial studies concerning existence, uniqueness and finite time blow-up of solu-
tions for the following equation:
u (t) + Au(t) = M(u(t)), u(0) = φ, (3)
have been considered by Murakami [6] and Segal [7].

Although Faedo–Galerkin method is useful for convergence yet the convergence
behaviour in many cases is not known. Bazely in [8] and [9] showed the uniform
convergence of the approximations to solutions of the nonlinear wave equation
u (t) + Au(t) + M(u(t)) = 0, u(0) = φ, u (0) = ψ, (4)
on any closed subinterval [0, T ] of existence of the solution. Miletta [10] has estab-
lished the convergence of Faedo–Galerkin approximate solutions to (3).
In [11] Bahuguna and Srivastava extended the results of Miletta [10] and con-
sidered the Faedo–Galerkin approximations of the solutions to following functional
integro-differential equation:
t
du(t)
= −Au(t) + M(u(t)) + g(t − s)k(u(s))ds, (5)
dt 0
in a separable real Hilbert space H .

Approximation of Solutions to Fractional Integro-Differential Equations … 679
The case of (1) in which q = 1, K = 0 and u t (θ ) = u(t), i.e. θ = 0 have been

studied by Bahuguna [12], the case in which u t (θ ) = u(t), i.e. θ = 0 have been
studied by Chaddha [13] and the case in which q = 1 and K = 0 have been studied
by Dubey [14]. Thus this paper is the generalization of [12–14]. For more details
to the existence of an approximate solution to different differential equations see
[15, 16].
The manuscript is organized as follows: In Sect. 2, we recall some necessary
preliminaries, lemmas, theorems and assumptions. In Sect. 3, we show the existence
and uniqueness of approximate solutions. In Sect. 4, the convergence of approximate
integral equation to the associated integral equation is established. In Sect. 5, we
consider the Faedo–Galerkin approximate solutions and prove the convergence of
such approximations.
In this section, we have some basic definitions, assumptions and properties of frac-
tional calculus which will be used further in this paper.
Definition 1 The Riemann–Liouville integral of order q with the lower limit zero
for a function f ∈ L 1 ((0, T ); H ) is defined by
t
q 1
It f (t) = (t − s)q−1 f (s)ds, t > 0, q > 0.
Γ (q) 0
Here q is called the order of fractional integration.
Definition 2 The Riemann–Liouville fractional derivative of order q for a function

f ∈ L 1 ((0, T ); H ) is defined by

q dm 1 t f (s)
Dt f (t) = [ ds],
dt Γ (m − q)
m
0 (t − s)q+1−m
where m − 1 < q < m and m ∈ N.
Definition 3 The Caputo fractional derivative of order q for a function f ∈

C m−1 ((0, T ); H ) ∩ L 1 ((0, T ); H ) is defined by
t
q 1
Dt f (t) = (t − s)m−q−1 f m (s)ds,
Γ (m − q) 0
where m − 1 < q < m and m ∈ N.
We shall use the following assumptions:

(H1) Since A is a closed, positive definite, self-adjoint linear operator from the
domain D(A) ⊂ H of A into H such that D(A) is dense in H . Therefore,
we can assume that A has the pure point spectrum 0 < λ0 ≤ λ1 ≤ λ2 ≤
· · · ≤ λm ≤ · · · with λm → 0 and a corresponding complete orthonormal
systemof eigenfunctions {φ j }, i.e. Aφ j = λ j φ j and < φi , φ j >= δi j , where
1, if i = j;
δi j =
0, otherwise.
These assumptions on A guarantee that −A is the infinitesimal generator of an
analytic semigroup Q(t) then for c > 0, large enough, −(A + cI ) is invertible and
generates a bounded analytic semigroup. This allows us to reduce the general case
in which −A is the infinitesimal generator of an analytic semigroup to the case in
which the semigroup is bounded and the generator is invertible.
Hence without loss of generality we can assume that Q(t) ≤ C for t > 0 and
0 ∈ ρ(−A), where ρ(−A) is the resolvent set of −A, i.e. we can define the positive
fractional power Aα as closed linear operator with domain D(Aα ) for α ∈ (0, 1].
Moreover, D(Aα ) is dense in H with the norm uα = Aα u.
Here, we signify the space D(Aα ) by Hα endowed with the α − nor m( · α ).
Also, we have that Hk
→ Hα for 0 < α < 1 and therefore, the embedding is
continuous. For study on the fractional powers of closed linear operators, we refer
to book by Pazy [17].
(H2) The function φ(t) ∈ D(A), for all t ∈ [−h, 0] and φ is locally H ölder
continuous on [−h, 0] and define

(t) = φ(t), t ∈ [−h, 0];
φ
φ(0), t ∈ [0, T ].
(H3) For 0 < α < 1, the nonlinear map g defined from [0, T ]×C0α into H is contin-
uous and there exists a non-decreasing function G R from [0, ∞) into [0, ∞)
depending on R > 0 such that
g(t, u 1 ) − g(t, u 2 ) ≤ G R (t)u 1 − u 2 0,α
and
g(t, u) ≤ G R (t),
), where B R (Z , z 0 ) = {z ∈
for (t, u 1 ), (t, u 2 ) and (t, u) in [0, T ] × B R (C0α , φ
Z |z − z 0 Z ≤ R} for any Banach space Z with its norm . Z .
(H4) There exist positive constants 0 < α < β < 1 such that the function Aβ f
β
defined from [0, T ] × C0α into C0 is continuous and there exist constants L
and 0 < γ ≤ 1 such that
Aβ f (t, u 1 ) − Aβ f (s, u 2 ) ≤ L f {| t − s |γ +u 1 − u 2 0,α }
).
and 2L f Aα−β < 1, for all (t, u 1 ), (s, u 2 ) in [0, T ] × B R (C0α , φ
(H5) For 0 < α < 1, the nonlinear map k defined on D(Aα ) into H is continuous
and there exists constant Ck > 0 such that
k(u 1 )(t) − k(u 2 )(t) ≤ Ck (t)Aα (u 1 − u 2 )
and
k(u)(t) ≤ Ck (t),
for u 1 , u 2 ∈ D(Aα ).
Definition 4 A continuous function u : [−h, T ] → H is said to be a mild solution
for the problem (1)–(2) if u(·) satisfies the following integral equation:
⎧
⎪
⎪ φ(t), t ∈ [−h, 0];
⎨ S (t)[φ(0) + f (0, φ )] − f (t, u )
q t
u(t) = t (6)
⎪
⎪ + 0 (t − s)q−1 Q q (t − s)A f (s, u s )ds
⎩ t
+ 0 (t − s)q−1 Q q (t − s)[g(s, u s ) + K (u)(s)]ds, t ∈ [0, T ].
The operators Sq (t) and Q q (t) are defined as follows:

∞
Sq (t) = ςq (ξ )Q(t q ξ )dξ,
0
∞
Q q (t) = q ξ ςq (ξ )Q(t q ξ )dξ,
0
1
1 1− q − q1
where ςq (ξ ) = qξ × ψq (ξ
) is a probability density function defined on
∞

∞
(0, ∞), i.e. ςq (ξ ) ≥ 0, 0 ςq (ξ )dξ = 1 and ψq (ξ ) = π 1
(−1)n−1 ξ −nq−1
n=1
Γ (nq + 1)
sin(nπq), 0 < ξ < ∞. For further details on mild solution see [18].
n!
Lemma 1 (Pazy [17]) Let −A be the infinitesimal generator of an analytic semi-
group {Q(t) : t ≥ 0} such that Q(t) ≤ C, for t ≥ 0 and 0 ∈ ρ(−A). Then,
1. For 0 < α ≤ 1, Hα is a Hilbert space.
2. The operator Aα Q(t) is bounded for every t > 0 and
AQ(t) ≤ Ct −1 ,
Aα Q(t) ≤ Cα t −α .
Lemma 2 (Zhou [19]) The operators {Sq (t), t ≥ 0} and {Q q (t), t ≥ 0} are bounded
linear operators such that
qCα Γ (2−α)t −qα
(i) Sq (t)z ≤ Cz, Q q (t)z ≤ qC
Γ (1+q) z and Aα Q q (t)z ≤ Γ (1+q(1−α) z, for any
z ∈ H.
(ii) The families {Sq (t) : t ≥ 0} and {Q q (t) : t ≥ 0} are strongly continuous.
(iii) If Q(t) is compact, then Sq (t) and Q q (t) are compact operators for any t > 0.
3 Approximate Solutions and Convergence

In this section, we show the existence of a mild solution of the Eqs. (1)–(2) on [0, T ]
for some T , 0 < T < ∞. Let Hn denote the finite-dimensional subspace of the
Hilbert space H spanned by {u 0 , u 1 , · · ·, u n } and let P n : H → Hn for n = 1, 2, · · ·,
be the corresponding projection operators.
Let 0 < T0 < T be an arbitrarily fixed constant and let
),
A1 = max Aβ f (t, φ
0≤t≤T
M0 = sup M(t).
0≤t≤T0
β
For each n, we define f n : [0, T ] × C0α → C0 and gn : [0, T ] × C0α → H ,
respectively, by f n (t, u t ) = f (t, P n u t ), gn (t, u t ) = g(t, P n u t ) nd k : D(Aα ) → H
and K by
kn (u) = k(P n u),

t
K n (u)(t) = M(t − s)kn (u(s))ds.
0
We choose T0 , 0 < T0 ≤ T sufficiently small such that
R
(Sq (t) − I )Aα (φ(0) + f n (0, φ
) ≤ (1 − μ) ,
3
q(β−α)
γ Γ (1 − (α − β)) T
Aα−β L f T0 + C1+α−β (L f R + A1 ) 0
Γ (1 + q(β − α)) (β − α)
q(1−α)
Cα Γ (2 − α) T R
+ C(T ) 0 < (1 − μ) ,
Γ (1 + q(1 − α)) (1 − α) 6
q(β−α) q(1−α)
C1+α−β Γ (1 − (α − β)) T Cα Γ (2 − α) T
Lf 0 + C(T ) 0 < (1 − μ),
Γ (1 + q(β − α)) (β − α) Γ (1 + q(1 − α)) 1−α

where C(T ) = G R (T ) + T0 M0 Ck (T ), R = 0,α 2 ), Cα , C1+α−β
2(R 2 + 3φ
and μ = Aα−β L f < 1, are constants.
Let Aα : Ctα → Ct be defined as (Aα ψ)(t) = Aα (ψ(t)) and (P n u t )(s) =
P (u(t + s)), for all s ∈ [−h, 0], t ∈ [0, T0 ]. Now, we consider
n
) = {u ∈ C Tα : u − φ
B R = B R (C Tα0 , φ α ≤ R}
0
and define the operator Q n on B R as follows:

⎧
⎪ (t),
φ t ∈ [−h, 0];
⎪
⎨ S (t)[φ
q (0) + f n (0,
φ )] − f n (t, u t )
(Q n u)(t) = t
⎪
⎪ + 0 (t − s)q−1 Q q (t − s)A f n (s, u s )ds
⎩ t
+ 0 (t − s)q−1 Q q (t − s)[gn (s, u s ) + K (u)(s)]ds, t ∈ [0, T0 ].
(7)
Theorem 1 Suppose that conditions (H 1)–(H 4) are fulfilled and φ(t) ∈ D(Aα ),
for all t ∈ [−h, 0]. Then, there exists a unique fixed point u n ∈ B R of the map Q n ,
i.e. Q n u n = u n for each n = 0, 1, 2, · · ·; that is, u n satisfies the approximate integral
equation
⎧
⎪ (t),
φ t ∈ [−h, 0];
⎪
⎨ S (t)[φ(0) + f n (0, φ
)] − f n (t, (u n )t )+
u n (t) = t
q
⎪
⎪ (t − s)q−1 Q q (t − s)A f n (s, (u n )s )ds
⎩ 0 t
+ 0 (t − s)q−1 Q q (t − s)[gn (s, (u n )s ) + K n (u n )(s)]ds, t ∈ [0, T0 ].
(8)
Proof To prove the theorem, we first show that map t → (Q n u)(t) is continuous
)
from [−h, T0 ] into D(Aα ) with respect to norm · α . Thus, for any u ∈ B R (C Tα , φ
and t ∈ [−h, 0], we have
(t + h) − φ
(Q n u)(t + h) − (Q n u)(t) = φ (t),
for sufficiently small h > 0. Further, for t ∈ (0, T0 ] and sufficiently small h > 0, we
have
Aα [(Q n u)(t + h) − (Q n u)(t)] =
[Sq (t + h) − Sq (t)]Aα (φ ) − Aα−β [Aβ f n (t + h, u t+h ) − Aβ f n (t, u t )]

(0) + f n (0, φ
t
+ [(t + h − s)q−1 − (t − s)q−1 ]A1+α−β Q q (t + h − s)Aβ f n (s, u s )ds
0
t
+ (t − s)q−1 A1+α−β [Q q (t + h − s) − Q q (t − s)]Aβ f n (s, u s )ds
0
t+h
+ (t + h − s)q−1 A1+α−β Q q (t + h − s)Aβ f n (s, u s )ds
t
t
+ [(t + h − s)q−1 − (t − s)q−1 ]Aα Q q (t + h − s)[gn (s, u s ) + K n (u n )(s)]ds
0
t
+ (t − s)q−1 Aα [Q q (t + h − s) − Q q (t − s)][gn (s, u s ) + K n (u n )(s)]ds
0
t+h
+ (t + h − s)q−1 Aα Q q (t + h − s)[gn (s, u s ) + K n (u n )(s)]ds
t

8
= Ji . (9)
i=1
For z ∈ H , we have
t+h t+h
d
[Q(t + h) ξ ) − Q(t ξ )]z =
q q
Q(τ q ξ )zdτ = qξ(τ )q−1 AQ(τ q ξ )zdτ.
t dτ t
Here
∞
J1 ≤ (0) + f n (0, φ
ςq (ξ )Q((t + h)q ξ ) − Q(t q ξ )Aα (φ )dξ
0
∞ t+h
≤ ςq (ξ ) (0) + f n (0, φ
qξ τ q−1 Aα Q(τ q ξ )dτ A(φ )dξ
0 t
∞ t+h
≤ qCα ξ 1−α ςq (ξ ) (0) + f n (0, φ
τ q(1−α)−1 A(φ )dτ dξ
0 t
Γ (1 − q)
≤ Cα A(φ (0) + f n (0, φ))[(t + h)q(1−α) − t q(1−α) ]
Γ (1 + q(1 − α))
qCα Γ (2 − α)
≤ A(φ (0) + f n (0, φ
))h[t + θ (h)]q(1−α)−1
Γ (1 + q(1 − α))
qCα A(φ (0) + f n (0, φ
))Γ (2 − α) q(1−α)−1 q(1−α)
≤ θ (h) , (10)
Γ (1 + q(1 − α))
where 0 < θ < 1 [ see [18, 20].

For J2 we have
J2 ≤ Aα−β Aβ f n (t + h, u t+h ) − Aβ f n (t, u t )

≤ L f Aα−β [| h |γ +P n u t+h − P n u t α ]
≤ L f Aα−β [| h |γ +u t+h − u t α ] (11)
and
(0)) + Aβ f n (t, φ

Aβ f n (t, u t ) ≤ Aβ f n (t, u t ) − Aβ f n (t, φ (0))
≤ L f P u t − φ
n (0)α + A1
≤ L f R + A1 . (12)
Also,
t
qC1+α−β Γ (1 − (α − β))
J3 ≤ (L f R + A1 ) (t − s)m 1 −1 (t + h − s)−m 1 k1 − (t − s)−m 1 k1 ds
Γ (1 + q(β − α)) 0
qC1+α−β Γ (1 − (α − β))
≤ (L f R + A1 )k1 ψ k1 −1 (1 − d)−m 1 (1−k1 )−1 h m 1 (1−k1 ) , (13)
Γ (1 + q(β − α))
(1−q)
where m 1 = 1 − q(1 − α − β), k1 = 1−q(1+α−β) , d = (1 − ( mk11 )1/k1 m 1 ) and
0 < ψ ≤ 1 (see [16, 20]).

qC1+α−β Γ (1 − (α − β)) t
J4 ≤ (L f R + A1 ) (t − s)q−1
Γ (1 + q(β − α)) 0
×[(t − s)−q(1+α−β) − (t + h − s) −q(1+α−β)
]ds
qC1+α−β Γ (1 − (α − β)) (L f R + A1 ) α−β
≤ d (1 − d3 )−q(β−α)−1 h q(β−α) ,
Γ (1 + q(β − α)) 1+α−β 2
(14)
where 0 < d2 ≤ 1 and d3 = (1 − (1 + α − β/q))1/q(1+α−β) .

qC1+α−β Γ (1 − (α − β)) t+h
J5 ≤ (L f R + A1 ) (t + h − s)q(β−α)−1 ds
Γ (1 + q(β − α)) t
C1+α−β Γ (1 − (α − β)) h q(β−α)

≤ (L f R + A1 ) . (15)
Γ (1 + q(β − α)) (β − α)
Similarly,
t
qCα Γ (2 − α)
J6 ≤ C(T ) (t − s)m 2 −1 [(t + h − s)−m 1 k2 − (t − s)−m 2 k2 ]ds
Γ (1 + q(1 − α)) 0
qCα Γ (2 − α)
≤ C(T )k2 ψ1k2 −1 (1 − d4 )−m 2 (1−k2 )−1 h m 2 (1−k2 ) , (16)
Γ (1 + q(1 − α))
1/k2 m 2
1−q k2
where m 2 = 1 − qα, k2 = 1−qα , 0 < θ1 ≤ 1 and d4 = (1 − m2 ).
qCα Γ (2 − α)
J7 ≤ C(T )θ2α−1 (1 − d5 )−q(1−α)−1 (h)q(1−α) , (17)
Γ (1 + q(1 − α))
where d5 = (1 − (α/q)1/qα ) and 0 < θ2 ≤ 1.
Cα Γ (2 − α) (h)q(1−α)
J8 ≤ C(T ) . (18)
Γ (1 + q(1 − α)) (1 − α)
from (9)–(18) we have the map Q n which is H ölder continuous from [−h, T0 ] into
D(Aα ) with α−norm.
) for any u ∈ B R (C α , φ
Now we show that Q n (u) ∈ B R (C Tα , φ ). Thus, for any
T
t ∈ [−h, 0],
(Q n u)(t) − φ(t)α = 0.

Further, for t ∈ (0, T0 ]
α ≤ (Sq (t) − I )Aα (φ

(Q n u)(t) − φ (0) + f n (0, φ
))
+ A α−β β ) − A f n (t, u t )
A f n (0, φ β
t
+ (t − s)q−1 A1+α−β Q q (t − s)Aβ f n (s, u s )ds
0
t
+ (t − s)q−1 Aα Q q (t − s)[gn (s, u s ) + K (u)(s)]ds
0
R γ
≤ (1 − μ) + Aα−β L f [T0 + u t − φα ]
3
q(β−α)
C1+α−β Γ (1 − (α − β)) T
+ (L f R + A1 ) 0
Γ (1 + q(1 − α)) (β − α)
q(1−α)
Cα Γ (2 − α) T
+ C(T ) 0
Γ (1 + q(1 − α)) (1 − α)
R R
≤ (1 − μ) + (1 − μ) + μR ≤ R.
3 6
Thus
T0 ,α ≤ R.
Q n u − φ
Hence Q n maps B R (C Tα0 , φ ) into B R (C α , φ

). Now we show that Q n is a strict
T0
) and t ∈ [−h, 0], we have
contraction on B R (C T0 , φ ). For u, v ∈ B R (C Tα0 , φ
α
(Q n u)(t) − (Q n v)(t)α = 0.
), then we have

Again, if t ∈ (0, T0 ] and u, v ∈ B R (C Tα0 , φ
(Q n u)(t) − (Q n v)(t)α ≤ Aα−β Aβ f n (t, u t ) − Aβ f n (t, vt )α

t
+ (t − s)q−1 A1+α−β Q q (t − s)Aβ f n (s, u s ) − Aβ f n (s, vs )ds
0
t
+ (t − s)q−1 Aα Q q (t − s)[gn (s, u s ) − gn (s, vs )
0
+ K n (u)(s) − K n (v)(s)]ds. (19)
Now,
Aβ f n (t, u t ) − Aβ f n (t, vt ) ≤ L f u − vT0 ,α , (20)
gn (s, u s ) − gn (s, vs ) ≤ G R (T )u − vT0 ,α (21)

and
s
K n (u)(s) − K n (v)(s)α ≤ M0 kn (u)(τ ) − kn (v)(τ )α dτ
0
≤ M0 T0 Ck (T )u − vT0 ,α . (22)
Using (20), (21) and (22) in (19), we get
q(β−α)
C1+α−β Γ (1 − (α − β)) T
(Q n u)(t) − (Q n v)(t)α ≤ [Aα−β L f + Lf 0
Γ (1 + q(β − α)) (β − α)
q(1−α)
Cα Γ (2 − α) T
+ C(T ) 0 ] × u − vT0 ,α .
Γ (1 + q(1 − α)) (1 − α)
Taking supremum on t over [−h, T0 ], we get

q(β−α)
C1+α−βΓ (1−(α−β)) T
(Q n u)(t) − (Q n v)(t)T0 ,α ≤ Aα−β L f + Lf 0
Γ (1 + q(β − α)) (β − α)
q(1−α)

Cα Γ (2 − α) T
+ C(T ) 0 × u − vT0 ,α .
Γ (1 + q(1 − α)) (1 − α)
). Hence,
The above estimate implies that Q n is a strict contraction on B R (C Tα0 , φ
α
there exists a unique u n ∈ B R (C T0 , φ ) such that Q n u n = u n . Clearly, u n satisfies
(8) on [0, T0 ].
Lemma 3 Let the hypotheses (H 1)–(H 4) hold. If φ(t) ∈ D(A) for t ∈ [−h, 0],
then u n (t) ∈ D(Aη ), for all t ∈ [−h, T0 ] with 0 < η ≤ β < 1.
Proof For all t ∈ [−h, 0] it is obvious. For t ∈ (0, t0 ], from Theorem (1) we
) satisfying (8). Theorem (1.2.4) in
have the existence of a unique u n ∈ B R (C Tα0 , φ
Pazy [17] implies that for t > 0 and 0 ≤ η < 1, τ (t) : H → D(Aη ) and for
0 ≤ η ≤ β < 1, D(Aβ ) ⊆ D(Aη ). (H4) implies that the map t → Aβ f (t, (u n )t )
is H ölder continuous on [0, T0 ]. Similarly, H ölder continuity of u n can be easily
established using (9)–(18). It follows that (see Theorem 4.3.2 in [17])
t
(t − s)q−1 Q q (t − s)Aβ f n (s, (u n )s )ds ∈ D(A).
0
Also, from Theorem (1.2.4) in Pazy [17], we have τ (t)u ∈ D(A) if u ∈ D(A).
The required result follows from these facts and the fact that D(A) ⊆ D(Aη ) for
0 ≤ η ≤ 1.
Proposition 1 Let the hypotheses (H 1)–(H 4) hold. If φ(t) ∈ D(A) for t ∈ [−h, 0],
then there exists a constant U0 such that
u n (t)μ ≤ U0 , 0 ≤ μ < β < 1, t ∈ [−h, T0 ], n = 1, 2, ...
Proof If φ(t) ∈ D(A), for all t ∈ [−h, 0], hence, φ(t) ∈ D(Aβ ), for all t ∈ [−h, 0]
and 0 ≤ μ ≤ β < 1. Now applying Aμ on both the sides of (8), then for all
t ∈ [−h, 0], we have
u n (t)μ = φ(t)μ ≤ φ0,μ = U0 .
Now for t ∈ (0, T0 ], we have
(0) + f n (0, φ
Aμ u n (t) ≤ Aμ Sq (t)(φ )) + Aμ−β Aβ f n (t, (u n )t )
t
+ (t − s)q−1 A1+μ−β Q q (t − s)Aβ f n (s, (u n )s )ds
0
t
+ (t − s)q−1 Aμ Q q (t − s)[gn (s, (u n )s ) + K n (u n )(s)ds.
0
(23)
Thus we have the required result for t ∈ [−h, T0 ].
In this section, we establish the convergence of the solution u n ∈ C Tα0 of the approx-
imate integral equation (8) to a unique solution u(·) ∈ C Tα0 of (6) on [−h, T0 ].
Theorem 2 Let the hypotheses (H 1)–(H 4) hold. If φ(t) ∈ D(Aα ) for t ∈ [−h, 0],
then
lim sup u n (t) − u k (t)α = 0, (24)

k→∞ {n≥k,t0 ≤t≤T0 }
for all t0 ∈ (0, T0 ].
Proof Let n ≥ k ≥ n 0 , where n 0 is large enough and n, k, n 0 ∈ N . For −h ≤ t0 ≤ 0,

we have
Aα [u n (t) − u k (t)] = φ(t) − φ(t)α = 0.
For t0 ∈ (0, T0 ] and 0 < α < η < β < 1,
gn (t, (u n )t ) − gk (t, (u k )t ) ≤ gn (t, (u n )t − gn (t, (u k )t ) + gn (t, (u k )t ) − gk (t, (u k )t )

≤ G R (T )[(u n )t − (u k )t α + (P n − P k )(u k )t α ].
Also,
(P n − P k )(u k )t α ≤ Aα−μ (P n − P k )Aμ (u k )t

1
≤ μ−α (u k )t μ .
λk
Thus
1
gn (t, (u n )t ) − gk (t, (u k )t ) ≤ G R (T )[(u n )t − (u k )t α + μ−α (u k )t μ ]
λk
U0
≤ G R (T )[u n − u k t,α + μ−α ].
λk
Similarly,

β β U0
A f n (t, (u n )t ) − A f k (t, (u k )t ) ≤ L f u n − u k t,α + μ−α
λk
and

1
kn (u n )(t) − kk (u k )(t) ≤ Ck (T ) u n (t) − u k (t)α + μ−α u k (t)μ .
λk
Therefore,
t
K n (u n )(t) − K k (u k )(t) ≤ M0 kn (u n )(τ ) − kk (u k )(τ )dτ
0
U0
≤ T0 M0 Ck (T )[u n − u k t,α + μ−α ].
λk
We choose t0 such that 0 < t0 < t < T0 , we have

) − f k (0, φ
u n (t) − u k (t)α ≤ Sq (t)Aα ( f n (0, φ )
+ Aα−β Aβ f n (t, (u n )t ) − Aβ f k (t, (u k )t )

t0 t
+ + (t − s)q−1 A1+α−β Q q (t − s)Aβ f n (s, (u n )s )
0 t0
− Aβ f k (s, (u k )s )ds

t0 t
+ + (t − s)q−1 Aα Q q (t − s)gn (s, (u n )s ) − gk (s, (u k )s )ds
0 t0
t
t0
+ + (t − s)q−1 Aα Q q (t − s)K n (u n )(s) − K k (u k )(s)ds.
0 t0
(25)
The first term of above inequality can be evaluated as
) − f k (0, φ
Sq (t)Aα ( f n (0, φ ) ≤ CAα−β Aβ f n (0, φ
) − Aβ f k (0, φ
)
α−β k
≤ CA L f (P − P )φ 0,α .
n
First, third and fifth integral of the inequality (25) can be evaluated as
t0
(t − s)q−1 A1+α−β Q q (t − s)[Aβ f n (s, (u n )s ) − Aβ f k (s, (u k )s )]ds
0
2qC1+α−β Γ (1 − (α − β))
≤ (L f R + A1 )(T0 − t0 )q(β−α)−1 t0 .
Γ (1 + q(β − α))
t0
(t − s)q−1 Aα Q q (t − s)[gn (s, (u n )s ) − gk (s, (u k )s )]ds
0
2qCα Γ (2 − α)
≤ G R (T )(T0 − t0 )q(1−α)−1 t0 .
Γ (1 + q(1 − α))
t0
(t − s)q−1 Aα Q q (t − s)[K n (u n )(s) − kk (u k )(s)]ds
0
2qCα Γ (2 − α)
≤ T0 M0 Ck (T )(T0 − t0 )q(1−α)−1 t0 .
Γ (1 + q(1 − α))
We estimate the second, forth and sixth integral of the inequality (25) as
t
(t − s)q−1 A1+α−β Q q (t − s)[Aβ f n (s, (u n )s ) − Aβ f k (s, (u k )s )]ds
t0

qC1+α−β Γ (1 − (α − β)) t 1
≤ Lf (t − s)q(β−α)−1 [(u n )s − (u k )s α + μ−α (u k )s μ ]ds
Γ (1 + q(β − α)) t0 λk
q(β−α)
qC1+α−β Γ (1 − (α − β)) U0 T0 t
≤ Lf[ μ−α + (t − s)q(β−α)−1 u n − u k s,α ds].
Γ (1 + q(β − α)) q(β − α)λk t0
t
(t − s)q−1 Aα Q q (t − s)gn (s, (u n )s ) − gk (s, (u k )s )ds
t0
t
qCα Γ (2 − α) 1
≤ G R (T ) (t − s)q(1−α)−1 [(u n )s − (u k )s + μ−α (u k )s μ ds]
Γ (1 + q(1 − α)) t0 λk
q(1−α) t
qCα Γ (2 − α) U0 T0 q(1−α)−1 u − u ds].
≤ G R (T )[ μ−α + (t − s) n k s,α
Γ (1 + q(1 − α)) q(1 − α)λk t0
t
(t − s)q−1 Aα Q q (t − s)[K n (u n )(s) − K k (u k )(s)]ds
t0

qCα Γ (2 − α) t 1
≤ T0 M0 Ck (T ) (t − s)q(1−α)−1 [(u n )(s) − (u k )(s)α + μ−α (u k )s μ ]ds
Γ (1 + q(1 − α)) t0 λk
q(1−α)
qCα Γ (2 − α) U0 T0 t
≤ T0 M0 Ck (T )[ μ−α + (t − s)q(1−α)−1 u n − u k s,α ds].
Γ (1 + q(1 − α)) q(1 − α)λk t0
Thus,

U0
u n (t) − u k (t)α ≤ CA α−β 0,α + A
L f (P − P )φ
n k α−β
L f u n − u k t,α + μ−α
λk

qC1+α−β Γ (1 − (α − β))
+2 (L f R + A1 )
(T0 − t0 )−q(β−α)+1 Γ (1 + q(β − α))
qCα Γ (2 − α) U0
+ C(T )]t0 + K α,β μ−α
(T0 − t0 )−q(1−α)+1 Γ (1 + q(1 − α)) λk
t
qC1+α−β Γ (1 − (α − β))
+ [ L f (t − s)q(β−α)−1

t0 Γ (1 + q(β − α))

qCα Γ (2 − α)
+ C(T )(t − s)q(1−α)−1 u n − u k s,α ds, (26)
Γ (1 + q(1 − α))
where
q(β−α) q(1−α)
qC1+α−β Γ (1 − (α − β)) T qCα Γ (2 − α) T
K α,β = Lf 0 + C(T ) 0 .
Γ (1 + q(β − α)) q(β − α) Γ (1 + q(1 − α)) q(1 − α)
and further, we have
B3
u n (t) − u k (t)α ≤ B1 + Aα−β L f (u n − u k t,α ) + B2 t0 + μ−α
λk
t
+ B4 (t − s)q(1−α)−1 u n − u k s,α ds, (27)
t0
where
B1 = CAα−β L f (P n − P k )φ 0,α ,

qC1+α−β Γ (1 − (α − β))
B2 = 2 (L f R + A1 )
(T0 − t0 )−q(β−α)+1 Γ (1 + q(β − α))

qCα Γ (2 − α)
+ C(T ) ,
(T0 − t0 )−q(1−α)+1 Γ (1 + q(1 − α))

B3 = Aα−β L f + K α,β U0 ,
qC1+α−β Γ (1 − (α − β)) qCα Γ (2 − α)

B4 = Lf + C(T ).
Γ (1 + q(β − α)) Γ (1 + q(1 − α))
Replace t by t + θ in above inequality where t0 − t ≤ θ ≤ 0, we get
B3
u n (t + θ ) − u k (t + θ )α ≤B1 + Aα−β L f (u n − u k t,α ) + B2 t0 + μ−α
λk
t+θ
+ B4 (t + θ − s)q(1−α)−1 u n − u k s,α ds.
t0
Now put s − θ = γ in the integral term of above inequality and we get
B3
u n (t + θ ) − u k (t + θ )α ≤ B1 + Aα−β L f (u n − u k t,α ) + B2 t0 + μ−α
λk
t
+ B4 (t − γ )q(1−α)−1 u n − u k γ +θ,α dγ
t0 −θ
B3
≤ B1 + Aα−β L f (u n − u k t,α ) + B2 t0 + μ−α
λk
t
+ B4 (t − γ )q(1−α)−1 u n − u k γ ,α dγ
t0 −θ
B3
≤ B1 + Aα−β L f (u n − u k t,α ) + B2 t0 + μ−α
λk
t
+ B4 (t − γ )q(1−α)−1 u n − u k γ ,α dγ .
t0
Thus,
B3
sup u n (t + θ ) − u k (t + θ)α ≤ B1 + Aα−β L f (u n − u k t,α ) + B2 t0 + μ−α
t0 −t≤θ≤0 λk
t
+ B4 (t − γ )q(1−α)−1 u n − u k γ ,α dγ . (28)
t0
Since u n (t + θ ) = φ(t + θ ) for t + θ ≤ 0 and for all n ≥ n 0 , thus, we have
sup u n (t + θ ) − u k (t + θ )α ≤ sup u n (t + θ ) − u k (t + θ )α

−h−t≤θ≤0 0≤θ+t≤t0
+ sup u n (t + θ ) − u k (t + θ )α .
t0 −t≤θ≤0
(29)
For 0 < t ≤ t0 , we have from (27)

B5
u n (t) − u k (t)α ≤ B1 + Aα−β L f (u n − u k t,α ) + B2 t0 + μ−α , (30)
λk
where B5 = Aα−β LU0 . Using inequalities (28) and (30) in (29), we get
sup u n (t + θ ) − u k (t + θ )α ≤2B1 + 2Aα−β L f (u n − u k t,α ) + 2B2 t0

−h≤t+θ ≤t

B3 + B5 t
+ μ−α + B4 (t − γ )q(1−α)−1 u n − u k γ ,α dγ .
λk t0
Since we have 2L f Aα−β < 1. We have
1 B3 + B5
u n (t) − u k (t)α ≤ [2B1 + 2B2 t0 + μ−α
(1 − 2L f Aα−β ) λk
t
+ B4 (t − γ )q(1−α)−1 u n − u k γ ,α ]dγ .
t0
Lemma (5.6.7) in [17] implies that there exists a constant M such that

1 B3 + B5
u n (t) − u k (t)α ≤ 2B1 + 2B2 t0 + μ−α M.
(1 − 2L f Aα−β ) λk
Letting k → ∞ and since t0 is arbitrary, the right-hand side may be made as small
as desired by taking t0 sufficiently small, which gives the required result.
Theorem 3 Let the hypotheses (H 1)–(H 4) hold and let φ(t) ∈ D(Aα ) for t ∈
[−h, 0]. Then there exists a unique function u ∈ C Tα0 such that u n → u as n → ∞
in C Tα0 and u satisfies (6) on [−h, T0 ].
Proof Let φ(t) ∈ D(Aα ) for t ∈ [−h, 0]. since for 0 < t ≤ T0 , Aα u n (t) con-
verges to Aα u(t) as n → ∞ and u n (t) = u(t) = φ(t), for all n and t ∈ [−h, 0],
we have, for −h = t = T , Aα u n (t) converges to Aα u(t) in H as n → ∞. Since
), it follows that u ∈ B R (C α , φ
u n ∈ B R (C Tα0 , φ ) and for any 0 < t0 ≤ T ,
T0
lim sup u n (t) − u(t)α = 0.

n→∞ t ≤t≤T
0 0
Also, we have
sup gn (t, (u n )t ) − g(t, u t ) ≤ G R (T )(u n − uT0 ,α + (P n − I )uT0 ,α ) → 0,

t0 ≤t≤T0
sup Aβ f n (t, (u n )t )− Aβ f (t, u t ) ≤ L f (u n −uT0 ,α +(P n − I )uT0 ,α ) → 0,

t0 ≤t≤T0
sup kn (u n )(s) − kk (u k )(s) ≤ Ck (T )(u n − uT,α + (P n − I )uT0 ) → 0.

t0 ≤t≤T0
As n → ∞. For t0 ∈ (0, t0 ), the integral equation (8) can be written as
(0) + f n (0, φ
(u n )(t) = Sq (t)[φ )] − f n (t, (u n )t )
t0 t
+ + (t − s)q−1 Q q (t − s)A f n (s, (u n )s )ds
0 t
t0 t
0
+ + (t − s)q−1 Q q (t − s)gn (s, (u n )s )ds

0 t0
t0 t
+ + (t − s)q−1 Q q (t − s)K n (u n )(s)ds.
0 t0
First, third and fifth integral can be evaluated as

t0 t0
(t − s)q−1 Q q (t − s)A f n (s, (u n )s )ds ≤ (t − s)q−1 A1−β Q q (t − s)Aβ f n (s, (u n )s )ds
0 0
qC1−β Γ (1 + β) qβ−1
≤(L f R + A1 )T0 t0 .
Γ (1 + qβ)
t0
C q−1
(t − s)q−1 Q q (t − s)gn (s, (u n )s )ds ≤ G R (T )T0 t0 .
0 Γ (q)
t0
C q−1
(t − s)q−1 Q q (t − s)K n (u n )(s)ds ≤ T0 M0 Ck (t)T0 t0 .
0 Γ (q)
Thus, we have
t
(0) + f n (0, φ
u n (t) − Sq (t)[φ )] + f n (t, (u n )t ) − (t − s)q−1 Q q (t − s)A f n (s, (u n )s )ds
t0
t
− (t − s)q−1 Q q (t − s)[gn (s, (u n )s ) + K n (u n )(s)]ds
t0

qC1−β Γ (1 + β) qβ−1 C q−1
≤ (L f (R) + A1 )T0 + C(T )T0 t0 .
Γ (1 + qβ) Γ (q)
Let n → ∞, we get
t
u(t) − Sq (t)[φ(0) + f (0, φ)] + f (t, u t ) − (t − s)q−1 Q q (t − s)A f (s, u s )ds
t0
t t
− (t − s)q−1 Q q (t − s)g(s, u s )ds − (t − s)q−1 Q q (t − s)K (u)(s)ds
t0 t0

qC1−β Γ (1 + β) qβ−1 C q−1
≤ (L f (R) + A1 )T0 + C(T )T0 t0 .
Γ (1 + qβ) Γ (q)
Since t0 ∈ (0, T0 ] is arbitrary, hence we conclude that u is the solution of the

integral equation (6).
Now, we show the uniqueness of the solutions to (6). Let u 1 , u 2 be two solutions
on the interval [−h, T0 ]. For t ∈ [−h, 0] uniqueness is obvious and for t ∈ (0, T0 ],
we have
u 1 (t) − u 2 (t)α ≤ Aα−β Aβ f (t, (u 1 )t ) − Aβ f (t, (u 2 )t )

t
+ (t − s)q−1 A1+α−β Q q (t − s)[Aβ f (s, (u 1 )s ) − Aβ f (s, (u 2 )s )]ds
0
t
+ (t − s)q−1 Aα Q q (t − s)[g(s, (u 1 )s ) − g(s, (u 2 )s )]ds
0
t
+ (t − s)q−1 Aα Q q (t − s)[K (u 1 )(s) − K (u 2 )(s)]ds
0
qC1+α−β Γ (1 − (α − β))
≤ Aα−β L f (u 1 )t − (u 2 )t )α + Lf
Γ (1 + q(β − α))
t
× (t − s)q(β−α)−1 (u 1 )s − (u 2 )s α ds
0
t
qCα Γ (2 − α)
+ G R (t) (t − s)q(1−α)−1 (u 1 )s − (u 2 )s α ds
Γ (1 + q(1 − α)) 0
t
qCα Γ (2 − α)
+ T0 M0 Ck (t) (t − s)q(1−α)−1 u 1 (s) − u 2 (s)α ds
Γ (1 + q(1 − α)) 0
≤ Aα−β L f sup u 1 (t + θ) − u 2 (t + θ)α
−h≤θ ≤0
t
+ C (t − s)q(1−α)−1 sup u 1 (s + θ) − u 2 (s + θ)α ds
0 −h≤θ ≤0

qCα Γ (2 − α) t
+ T0 M0 Ck (t) (t − s)q(1−α)−1 u 1 (s) − u 2 (s)α ds,
Γ (1 + q(1 − α)) 0
qC Γ (1−(α−β)) α Γ (2−α)
where C = 1+α−β
Γ (1+q(β−α)) L f + ΓqC (1+q(1−α)) G R (t).
Let θ̄ ∈ [−t, 0] and t ∈ [0, T0 ] and suppose T0 ≤ h, hence, we have 0 ≤ t ≤ h.
For t ≤ −θ̄ , we have u 1 (t + θ̄ ) = u 2 (t + θ̄ ). For t ≥ −θ̄ , we have
u 1 (t + θ̄) − u 2 (t + θ̄)α ≤ Aα−β L f sup u 1 (t + θ̄ + θ) − u 2 (t + θ̄ + θ)α

−h≤θ≤0
t+θ̄
+ C (t + θ̄ − s)q(1−α)−1 sup u 1 (s + θ) − u 2 (s + θ)α ds
0 −h≤θ≤0
t
qCα Γ (2 − α)
+ T0 M0 Ck (t) (t − s)q(1−α)−1 u 1 (s) − u 2 (s)α ds.
Γ (1 + q(1 − α)) 0
Substitute s = η + θ̄ , we get
u 1 (t + θ̄) − u 2 (t + θ̄)α ≤ Aα−β L f sup u 1 (t + θ̄ + θ) − u 2 (t + θ̄ + θ)α

−h≤θ ≤0
t
+ C (t − η)q(1−α)−1 sup u 1 (η + θ̄ + θ) − u 2 (η + θ̄ + θ)α dη
−θ̄ −h≤θ ≤0

Γ (1 + q(1 − α)) 0
Let θ = γ − θ̄ , then we get
u 1 (t + θ̄) − u 2 (t + θ̄)α ≤ Aα−β L f sup u 1 (t + γ ) − u 2 (t + γ )α

−h+θ̄ ≤γ ≤0
t
+ C (t − η)q(1−α)−1 sup u 1 (η + γ ) − u 2 (η + γ )α dη
−θ̄ −h+θ̄ ≤γ ≤0

Γ (1 + q(1 − α)) 0
Since u 1 (η + γ ) = u 2 (η + γ ) on [−h + θ̄ , −h] then
(u 1 )t (θ̄) − (u 2 )t (θ̄)α ≤ Aα−β L f sup (u 1 )t (γ ) − (u 2 )t (γ )α

−h≤γ ≤0
t
+ C (t − η)q(1−α)−1 sup (u 1 )η (γ ) − (u 2 )η (γ )α dη
−θ̄ −h≤γ ≤0

+ T0 M0 Ck (t) (t − s)q(1−α)−1 u 1 (s) − u 2 (s)α ds
Γ (1 + q(1 − α)) 0
≤ Aα−β L f sup (u 1 )t (γ ) − (u 2 )t (γ )α
−h≤γ ≤0
t
+ C (t − η)q(1−α)−1 sup (u 1 )η (γ ) − (u 2 )η (γ )α dη
0 −h≤γ ≤0

Γ (1 + q(1 − α)) 0
Taking supremum on θ̄ over [−h, 0], we get

t
(u 1 )t − (u 2 )t 0,α ≤ Aα−β L f (u 1 )t − (u 2 )t 0,α + C (t − η)q(1−α)−1 (u 1 )η − (u 2 )η 0,α dη
0

Γ (1 + q(1 − α)) 0
Since Aα−β L f < 1, we have

C t
(u 1 )t − (u 2 )t 0,α ≤ (t − η)q(1−α)−1 (u 1 )η − (u 2 )η 0,α dη.
1 − Aα−β L f 0
(31)
qCα Γ (2−α)
where C = C + Γ (1+q(1−α)) T0 M0 C k (t). Using Lemma (5.6.7) in Pazy [17], we
get
(u 1 )t − (u 2 )t 0,α = 0,
for all t ∈ [0, T0 ]. and from the fact that
1
u 1 (t) − u 2 (t) ≤ (u 1 )t − (u 2 )t α
λα0
it implies that u 1 = u 2 on [0, T0 ].
5 Faedo–Galerkin Approximations
In this section, we discuss the Faedo–Galerkin Approximation of solution and prove

the convergence result for such approximation.
For any 0 < T0 < T , we have a unique u ∈ C Tα0 satisfying the following integral
equation:
⎧
⎪ φ(t), t ∈ [−h, 0];
⎨
S (t)[φ(0) + f (0, )] − f (t, u t ) + t (t − s)q−1 Q q (t − s)A f (s, u s )ds
φ
u(t) = q t 0
⎪
⎩ + 0 (t − s)q−1 Q q (t − s)[g(s, u s ) + K (u)(s)]ds, t ∈ [0, T0 ].
(32)
Also, we have a unique solution u n ∈ C Tα0 of the approximate integral equation

⎧
⎪ φ (t), t ∈ [−h, 0];
⎨ t

u n (t) = Sq (t)[φ(0) + f n (0,φt )] − f n (t, (u n )t ) + 0 (t − s)
q−1 Q (t − s)
q
⎪
⎩ A f n (s, (u n )s )ds + 0 (t − s)q−1 Q q (t − s)[gn (s, (u n )s ) + K n (u n )(s)]ds, t ∈ (0, T0 ].
(33)
If we project (33) onto Hn , we get the Faedo–Galerkin approximation

u n (t) =
P n u n (t) satisfying
⎧ n
⎪ P φ (t), t ∈ [−h, 0];
⎪
⎪
⎨ Sq (t)[P n φ(0) + P n f (0, P n φ)] − P n f (t, P n (u n )t )

u n (t) = P n u n (t) = + t (t − s)q−1 AQ q (t − s)P n f (s, P n ( u n )t )ds
⎪
⎪ 0
⎪
⎩ + 0t (t − s)q−1 Q q (t − s)[P n g(s, P n (
u n )t ) + P n K ( u n )(s)]ds, t ∈ [0, T0 ].
(34)
The solutions u of (32) and

u n of (34), have the representation
∞
u(t) = i=0 αi (t)u i , αi (t) = (u(t), u i ), i = 0, 1, · · ·; (35)
∞ n

u n (t) = i=0 αi (t)u i , αin (t) = (
u n (t), u i ), i = 0, 1, · · · (36)
Using (36) in (34), we get the following system of first-order functional differential
equations
dq n
(α (t) + Fin (t, (α0n )t , · · ·, (αin )t )) + λi αin (t) = G in (t, (α0n )t , ..., (αnn )t )
dt q i
t
+ M(t − s)kin (α0n (s), · · ·, αnn (s))ds, (37)
0
αin (0) = φi ,
here

n
Fin (t, (α0n )t , · · ·, (αin )t )) = f (t, (αin )t u i ), u i ,
i=0

n
G in (t, (α0n )t , · · ·, (αnn )t) = g(t, (αin )t u i ), u i ,
i=0

n
kin (α0n (t), · · ·, αnn (t)) = k( (αin )(t)u i ), u i
i=0
and φi = (φ, u i ) for i = 1, 2, . . . n.
The system (37) determines the αin (t) s. Now, we shall prove the convergence of
αin (t)
→ αi (t). We have
∞ ∞

α α
A [u(t) −
u (t)] = A (αi (t) − αin (t))u i = λiα (αi (t) − αin (t))u i .
i=0 i=0
Therefore, we have
∞

α
A [u(t) −
u (t)] ≥ 2
λi2α (αi (t) − αin (t))2 .
i=0
Theorem 4 Let the hypotheses (H 1)–(H 4) hold. If φ(0) ∈ D(Aα ) for t ∈ [−h, 0],
then ∞


n→∞ t ≤t≤T
0 0 i=0
The statement of this theorem follows from the facts mentioned above and the
following result:
Proposition 2 Let the hypotheses (H 1)–(H 4) hold. If φ(0) ∈ D(Aα ) for t ∈

[−h, 0], then
lim sup Aα [u n (t) −
u m (t)] = 0.
n→∞ n≥m,t ≤t≤T
0 0
Proof For n = m, we have
Aα [ u m (t)] = Aα [P n u n (t) − P m u m (t)]

u n (t) −
≤ P n [u n (t) − u m (t)]α + (P n − p m )u m α
1
≤ u n (t) − u m (t)α + μ−α Aμ u m .
λm
If φ(t) ∈ D(A) for t ∈ [−h, 0], then the result follows from Theorem (4.1).
References
1. Kilbas, A.A., Srivastava, H.M., Trujillo, J.J.: Theory and applications of fractional differential
equations. North-Holland Mathematics Studies. Elsevier, Amsterdam (2006)
2. Lakshmikantham, V.: Theory of fractional functional differential equations. Nonlinear Anal.
69(10), 3337–3343 (2008)
3. Miller, K.S., Ross, B.: An introduction to the fractional calculus and fractional differential
equations. A Wiley-Interscience Publication, Wiley, New York (1993)
4. Podlubny, I.: Fractional differential equations. Mathematics in Science and Engineering. Aca-
demic Press, San Diego (1999)
5. Samko, S.G., Kilbas, A.A., Marichev, O.I.: Fractional Integrals and Derivatives, Translated
from the 1987 Russian Original. Gordon and Breach, Yverdon (1993)
6. Murakami, H.: On non-linear ordinary and evolution equations. Funkcial. Ekvac. 9, 151–162
(1966)
7. Segal, I.: Non-linear semi-groups. Ann. Math. 78(2), 339–364 (1963)
8. Bazley, N.W.: Approximation of wave equations with reproducing nonlinearities. Nonlinear
Anal. 3(4), 539–546 (1979)
9. Bazley, N.W.: Global convergence of Faedo-Galerkin approximations to nonlinear wave equa-
tions. Nonlinear Anal. 4(3), 503–507 (1980)
10. Miletta, P.D.: Approximation of solutions to evolution equations. Math. Methods Appl. Sci.
17(10), 753–763 (1994)
11. Bahuguna, D., Srivastava, S.K.: Approximation of solutions to evolution integrodifferential
equations. J. Appl. Math. Stochastic Anal. 9(3), 315–322 (1996)
12. Bahuguna, D., Shukla, R.: Approximations of solutions to nonlinear Sobolev type evolution
equations. Electron. J. Differ. Equ. 2003(1), 16pp. (electronic)
13. Chaddha, A., Pandey, D.N.: Approximations of solutions for a Sobolev type fractional order
differential equation. Nonlinear Dyn. Syst. Theory 14(1), 11–29 (2014)
14. Dubey, R.S.: Approximations of solutions to abstract neutral functional differential equation.
Numer. Funct. Anal. Optim. 32(3), 286–308 (2011)
15. Muslim, M., Bahuguna, D.: Existence of solutions to neutral differential equations with deviated
argument, Electron. J. Qual. Theory Differ. Equ. 2008(27), 12 pp. (2008)
16. Muslim, M., Nandakumaran, A.K.: Existence and approximations of solutions to some frac-
tional order functional integral equations. J. Integral Equations Appl. 22(1), 95–114 (2010)
17. Pazy, A.: Semigroups of linear operators and applications to partial differential equations.
Applied Mathematical Sciences. Springer, New York (1983)
18. El-Borai, M.M.: Some probability densities and fundamental solutions of fractional evolution
equations. Chaos Solitons Fractals 14(3), 433–440 (2002)
19. Zhou, Y., Jiao, F.: Existence of mild solutions for fractional neutral evolution equations. Com-
put. Math. Appl. 59(3), 1063–1077 (2010)
20. Abujabal, H.A.S., El-Borai, M.M.: On the Cauchy problem for some abstract nonlinear differ-
ential equations. Korean J. Comput. Appl. Math. 3(2), 279–290 (1996)
An Efficient Hybrid Approach for Simulating
MHD Nanofluid Flow over a Permeable
Stretching Sheet
Rama Bhargava, Mania Goyal and Pratibha
Abstract The problem of magnetohydrodynamics boundary layer flow and heat

transfer on a permeable stretching surface in a nanofluid under the effect of heat gen-
eration and partial slip is simulated using numeric symbolic approach. The Brownian
motion and thermophoresis effects are also considered. The boundary layer equa-
tions governed by the PDEs are transformed into a set of ODEs with the help of
transformations. The differential equations are solved by variational finite element
method as well as hybrid approach. The results obtained by the two approaches
match well. The effects of different controlling parameters on the flow field and heat
transfer characteristics are examined. The comparison confirms excellent agreement.
The efficiency of the hybrid approach is demonstrated through a table. The present
study is of great interest in coating and suspensions, cooling of metallic plate, oils and
grease, paper production, coal water or coal-oil slurries, heat exchangers technology,
materials processing exploiting.
Keywords Nanofluids · MHD · Stretching sheet · FEM · Hybrid approach
1 Introduction
The complicated mathematical models arising in problems of fluid flow and heat
transfer has forced for finding numerical solutions using grid-based methods, e.g.,
finite element methods, finite volume method, etc. However, due to the basic problem
of meshing and remeshing, these methods consume a lot of computational time.
R. Bhargava · M. Goyal (B) · Pratibha

e-mail: goyalmania87@gmail.com
R. Bhargava
e-mail: rbharfma@iitr.ac.in
Pratibha
e-mail: pratifma@iitr.ac.in

DOI 10.1007/978-81-322-2485-3_56
702 R. Bhargava et al.
Further, the occurrence of many integrals by default in the elements as well in the
global matrix lead to many inaccuracies. Presently, symbolic computation is vastly
applied in finite element analysis to solve system of equations, to derive stiffness
matrices, which provide closed-form expressions for integration and can deliver a
drastic savings of CPU time. Symbolic computation provide closed-form expressions
for integration. Kaminski [1] developed the semi-analytical probabilistic version of
the finite element method (FEM) related to the homogenization problem. The hybrid
computational implementation of the system MAPLE with homogenization-oriented
FEM code MCCEFF was invented to provide probabilistic FEM analysis.
Alns and Mardal [2] generated a low-level C++ code based on symbolic expres-
sions to accomplish a high degree of abstraction in the problem definition while
surpassing the run-time efficiency of traditional hand written C++ codes for FEM.
Integrations were performed analytically by Griffith et al. [3] in closed form with
the help of computer algebra software for 3-, 6-, 10- and 15-noded triangles. The
analytical routines ran significantly faster. Videla et al. [4] generated exact expres-
sions for the stiffness matrix of an 8-node plane elastic finite element using computer
algebra software. The reduction in CPU time was over 50 %. Eyheramendy and Saad
[5] recently explored a track in which the symbolic framework is integrated into the
classic numerical one which provides a dynamic binding of symbolically created
formula.
It was observed that a hybrid approach, combining FEM with symbolic compu-
tations has a lot of advantages namely:
1. It increases accuracy.
2. It saves a lot of computational time.
Convective heat transfer can be enhanced by changing the flow geometry, bound-
ary conditions, or by enhancing thermal conductivity of the fluid. Researchers have
also tried to increase the thermal conductivity of base fluids by suspending micro- or
larger-sized solid particles in fluids, since the thermal conductivity of solid is typi-
cally higher than that of liquid. Modern nanotechnology provides new opportunities
to process and produce materials with average crystallite sizes below 50 nm. Fluids
with nanoparticles suspended in them are called nanofluids, a term first proposed by
Choi [6]. Choi et al. [7], and Masuda et al. [8] have shown that a very small amount
of nanoparticles (usually less than 5 %), when dispersed uniformly and suspended
stably in base fluids, can provide dramatic improvements in the thermal conductivity
and in the heat transfer coefficient of the base fluid. Nanofluid is a suspension of
nanoparticles in the base fluid. A comprehensive survey of convective transport in
nanofluids was made by Buongiorno [9]. Khan and Pop [10] have used the model of
Kuznetsov and Nield [11], Makinde and Aziz [12], Khan and Pop [10].
It is now a well-accepted fact that many fluids of industrial and geophysical impor-
tance are non-Newtonian. Due to much attention in many industrial applications, the
research on boundary layer behaviour of a viscoelastic fluid over a continuously
stretching surface keeps going. McCormack and Crane [13] have provided compre-
hensive discussion on boundary layer flow caused by stretching of an elastic flat
sheet moving in its own plane with a velocity varying linearly with distance. Sev-
An Efficient Hybrid Approach for Simulating MHD Nanofluid … 703
eral researchers, viz., Gupta and Gupta [14], Dutta et al. [15], Chen and Char [16]
extended the work of Crane [13] by including the effects of heat and mass transfer
under different situations. Later, Rajagopal et al. [17] and Chang [18] presented an
analysis on flow of viscoelastic fluid over a stretching sheet. The above sources all
utilize the no-slip condition. Wang [19] discussed the partial slip effects on the planar
stretching flow.
A study of utilizing heat source or sink in moving fluids assumes a greater signif-
icance in all situations that deal with exothermic or endothermic chemical reaction
and those concerned with dissociating fluids. Sparrow and Cess [20] investigated
the steady stagnation point flow and heat transfer in the presence of temperature-
dependent heat absorption. Later, Azim et al. [21] discussed the effect of viscous
Joule heating on MHD-conjugate heat transfer for a vertical flat plate in the presence
of heat generation. One of the latest works is the study of the heat transfer charac-
teristic in the mixed convection flow of a nanofluid along a vertical plate with heat
source/sink, studied by Rana and Bhargava [22].
In real situations in nanofluids, the base fluid does not satisfy the properties of
Newtonian fluids, hence it is more justified to consider them as viscoelastic fluids. In
this paper, the base fluid is taken as second grade fluid. To the best of our knowledge,
no studies have so far investigated to analyze the partial slip effect on the boundary
layer flow of viscoelastic nanofluid over a permeable stretching sheet under the
effect of MHD and heat generation. The objective of the present paper is to extend
the work of Noghrehabadi [23] by taking base fluid as second grade fluid with a
new approach. The hybrid technique is used to simulate the heat and mass transfer
characteristics of the flow. Therefore, working with symbolic-numeric environment,
we solve our models using FEM, working with closed-form analytical expressions,
directly transforming complex analytical expressions into numerical tools and drastic
savings of CPU time.
The results so obtained are compared with those of FEM and shown in Table III
which shows a drastic saving in computational time. The effects of flow controlling
parameters on the fluid velocity, temperature, nanoparticle concentration are shown.
The heat transfer rate and the nanoparticle volume fraction rate have been demon-
strated graphically and discussed. Our aim is to reduce, substantially, the CPU time,
which is a subject of concern when dealing with large FEM meshes. The stiffness
terms of elements are computed analytically through symbolic computation which
will provide closed-form expressions for integration and has provided a drastic sav-
ings of CPU time. The analytical integration also ensures accurate results even for
distorted elements.
Consider two-dimensional, steady, incompressible, laminar flow of non-Newtonian

nanofluid past a stretching sheet in a quiescent fluid. The velocity of the stretching
sheet is u w = U = cx. The x-axis is taken along the plate in the vertically upward
direction and the y-axis is taken normal to the plate. A transverse magnetic field of
strength Bo is applied parallel to the y-axis. The surface of plate is maintained at
uniform temperature and concentration, Tw and Cw , respectively, and these values
are assumed to be greater than the ambient temperature and concentration, T∞ and
C∞ , respectively. Moreover, it is assumed that both the fluid phase and nanoparticles
are in thermal equilibrium state. The thermophysical properties of the nanofluid are
assumed to be constant. The pressure gradient and external forces are neglected. The
governing equations are:
∂u ∂v
+ =0 (1)
∂x ∂y

∂u ∂u ∂ 2u σ 2 α1 ∂u ∂ 2 u ∂ 3u ∂u ∂ 2 v ∂ 3u
u +v =ν 2 − Bo u + + u + + v (2)
∂x ∂y ∂y ρf ρ f ∂ x ∂ y2 ∂ x∂ 2 ∂ y ∂ y2 ∂ y3

∂T ∂T ∂2T Qo ∂C ∂ T DT ∂C 2
u +v = αm 2 + + τ DB + (3)
∂x ∂y ∂y (T − T∞ ) f ∂y ∂y T∞ ∂ y
∂C ∂C ∂ 2C DT ∂ 2 T
u +v = DB 2 + (4)
∂x ∂y ∂y T∞ ∂ y 2
The boundary conditions for the velocity, temperature, and concentration fields are
given as follows:
∂u
u = U + κν , v = vw , T = Tw , C = Cw at y = 0 (5)
∂y
u = 0, T = T∞ , C = C∞ as y → ∞ (6)
where u and v are the velocity components along the x and y-directions, respectively,
p is the pressure, ρ f is the density of base fluid, ρ p is the nanoparticle density, μ is
the absolute viscosity of the base fluid, ν is the kinematic viscosity of the base fluid, σ
is the electrical conductivity of the base fluid, α1 is the material fluid parameter, T is
the fluid temperature, αm is the thermal diffusivity, τ = (ρC) p /(ρC) f is the ratio of
effective heat capacity of the nanoparticle material to heat capacity of the fluid, C is
the nanoparticle volume fraction, D B and DT are the Brownian diffusion coefficient
and the thermophoresis diffusion coefficient, T∞ is the free stream temperature, C p
is the specific heat at constant pressure, and g, k are the acceleration due to gravity,
the thermal conductivity of the fluid respectively.
To transform the governing equations into a set of similarity equations, the fol-
lowing dimensionless parameters are introduced:

c √ T − T∞ C − C∞
η= y, u = cx f (η), v = − cν f (η), θ(η) = , φ(η) = . (7)
ν Tw − T∞ Cw − C∞
The transformed momentum, energy, and concentration equations together with the
boundary conditions given by (1)–(4), (5), (6) can be written as

f + f f − f 2 − M f − α f 2 − 2 f f + f f iv = 0 (8)
1
θ + f θ + N bθ φ + N tθ 2 + Qθ = 0 (9)
Pr
N t
φ + Le f φ + θ =0 (10)
Nb
The transformed boundary conditions are
f (0) = s, f (0) = 1 + K f (0), θ (0) = 1, φ(0) = 1 at η = 0 (11)

f (∞) → 0, θ (∞) → 0, φ(∞) → 0 as η → ∞ (12)
where primes denote differentiation with respect to η and the seven parameters
appearing in Eqs. (8–10) are defined as follows:
ν ν σ Bo2 α1 c Qo
Pr = , Le = ,M = ,α = ,Q =
αm DB ρf c μ cρC f
(ρC) p D B (Cw − C∞ ) (ρC) f DT (Tw − T∞ )
Nb = , Nt = (13)
(ρC) f ν (ρC) f T∞ ν
In Eq. (13), Pr, Le, M, α, Q, N b, and N t denote the Prandtl number, the Lewis
number, the magnetic field strength parameter, the viscoelastic parameter, the heat
source/sink parameter, the Brownian motion parameter, and the thermophoresis para-
meter, respectively.
The physical quantities of interest are the local heat flux N u and the local mass
diffusion flux Sh from the vertical moving plate, which are defined as
xqw xh w
Nu = , Sh = (14)
k(Tw − T∞ ) D B (Cw − C∞ )
where τw is the wall skin friction, qw is the surface heat flux, and h w is the wall mass
flux given by
∂T ∂C
qw = −k = 0, h w = −D B =0 (15)
∂y y ∂y y
Using (7) in (14), one can obtain
−1/2 −1/2
Rex N u x = −θ (0) = N ur, Rex Sh x = −φ (0) = Shr, (16)
where Rex = u w (x)x/ν is the local Reynolds number based on the stretching velocity
−1/2 −1/2
u w (x). Kuznetsov and Nield [11] referred Rex N u x and Rex Sh x as the reduced
Nusselt number N ur = −θ (0) and reduced Sherwood number Shr = −φ (0),

respectively.
3 Method of Solution
The problem has been undertaken by FEM and hybrid approach. The Finite Element
Method (FEM) is a numerical and computer-based technique of solving a variety of
practical engineering problems that arise in different fields. The method essentially
consists of assuming the piecewise continuous function for the solution and obtaining
the parameters of the functions in a manner that reduces the error in the solution.
The steps involved in the finite element analysis are as follows:
• Discretization of the domain into set of finite elements.
• Weighted integral formulation of the differential equation.
• Defining an approximate solution over the element.
• Substitution of the approximate solution and the generation of the element equa-
tions.
• Assembly of the stiffness matrices for each element.
• Imposition of the boundary conditions.
• Solution of assembled equations.
The entire flow domain is divided into 10,000 quadratic elements of equal size. Each
element is three-noded and therefore the whole domain contains 20001 nodes. A
system of equations have been obtained which is solved numerically. The code of
the algorithm has been executed in MATLAB. Excellent convergence was achieved
for all the results.
Nonlinear ordinary differential equations (8)–(10) together with the boundary con-
ditions (11) and (12) are solved numerically using FEM and numeric symbolic
approach. The numerical computations have been carried out for different values
of the parameters involved. The aim of the present study is to examine the variations
of different quantities of parameters as given 0 ≤ K ≤ 10, 0 ≤ α ≤ 10, 0 ≤ Pr ≤
70, 0.1 ≤ N t ≤ 0.5, 0.1 ≤ N b ≤ 0.5, and 5 ≤ Le ≤ 30 and to show the efficiency
of FEM with symbolic approach. The computational work is carried out by taking
size of the element η = 0.0001. It is observed that, if the number of elements is
increased or the size of the element is decreased in the same domain, even then the
accuracy is not affected.
Figure 1 demonstrates that the effect of increasing value of slip parameter K is
to shift the streamlines toward stretching boundary and thereby reduce thickness of
the momentum boundary layer. Therefore, the effect of slip parameter K is seen to
decrease the boundary layer velocity while the temperature and concentration are
increased with increase in the slip parameter. Figures 4, 5 and 6 show the effect of
viscoelastic parameter α on the evolution of fluid motion and subsequent on the
1
0.9
hybrid
0.8 FEM
0.7
0.6
f ’(η) 0.5
0.4
0.3
0.2
0.1
0
0 0.5 1 1.5 2 2.5 3 3.5 4
η
Fig. 1 Effect of K on velocity profile for Pr = Le = M = 1.0, N b = N t = 0.5, Q = 0.1, α =

0.5, s = 0.5
1
0.9
hybrid
0.8
FEM
0.7
0.6
θ (η) 0.5
0.4
0.3
0.2
0.1
0
0 1 2 3 4 5 6 7 8 9 10
η
Fig. 2 Effect of K on temperature profile for Pr = Le = M = 1.0, N b = N t = 0.5, Q =

0.1, α = 0.5, s = 0.5
distribution of heat and mass across the sheet as time evolves. From this plot it is
evident that increasing values of viscoelastic parameter α opposes the motion of
the liquid close to the stretching sheet and assists the motion of the liquid far away
from the stretching sheet. Increasing values of second-grade parameter enables the
liquid to flow at a faster rate due to which there is decline in the heat transfer. This is
responsible for the increase in momentum boundary layer, whereas the thermal and
concentration boundary layers reduce when the viscoelastic effects intensify (Figs. 2
and 3).
The variations in velocity field, temperature distribution, and nanoparticle con-
centration profile for various values of M are presented in Figs. 7, 8 and 9. It is clear
from these figures that the velocity decreases, whereas the temperature and concen-
tration increase with the increase in the magnetic field parameter. The hydromagnetic
force in Eq. (7) is a linear Lorentzian body force which acts transverse to the direc-
tion of application, i.e., in the negative x-direction, parallel to the plate surface. It is
1
0.9
hybrid
0.8
FEM
0.7
0.6
φ (η) 0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12
η
Fig. 3 Effect of K on nanoparticle concentration profile for Pr = Le = M = 1.0, N b = N t =

0.5, Q = 0.1, α = 0.5, s = 0.5
0.6
hybrid
0.5
FEM
0.4
f ’(η)
0.3
0.2
0.1
0
0 1 2 3 4 5 6
η
Fig. 4 Effect of α on velocity profile for Pr = Le = M = K = 1.0, N b = N t = 0.5, Q =

0.1, s = 0.5
1
0.9
hybrid
0.8
FEM
0.7
0.6
θ (η) 0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12
η
Fig. 5 Effect of α on temperature profile for Pr = Le = M = K = 1.0, N b = N t = 0.5, Q =

0.1, s = 0.5
1
0.9
hybrid
0.8
FEM
0.7
0.6
φ (η) 0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12
η
Fig. 6 Effect of α on nanoparticle concentration profile for Pr = Le = M = K = 1.0, N b =

N t = 0.5, Q = 0.1, s = 0.5
0.5
hybrid
0.4 FEM
0.3
f ’(η)
0.2
0.1
0
0 1 2 3 4 5 6
η
Fig. 7 Effect of M on velocity profile for Pr = Le = K = 1.0, N b = N t = 0.5, Q = 0.1, s =

0.5, α = 0.5
1
0.9
hybrid
0.8
FEM
0.7
0.6
θ (η) 0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12
η
Fig. 8 Effect of M on temperature profile for Pr = Le = K = 1.0, N b = N t = 0.5, Q =

0.1, s = 0.5, α = 0.5
1
0.9
hybrid
0.8
FEM
0.7
0.6
φ (η) 0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12
η
Fig. 9 Effect of M on nanoparticle concentration profile for Pr = Le = K = 1.0, N b = N t =

0.5, Q = 0.1, s = 0.5, α = 0.5
0.5
0.45
hybrid
0.4
FEM
0.35
0.3
f ’(η) 0.25
0.2
0.15
0.1
0.05
0
1 2 3 4 5 6
η
Fig. 10 Effect of s on velocity profile for Pr = Le = K = M = 1.0, N b = N t = 0.5, Q =

0.1, α = 0.5
1
0.9
hybrid
0.8
FEM
0.7
0.6
φ (η) 0.5
0.4
0.3
0.2
0.1
0
0 1 2 3 4 5 6 7 8 9 10
η
Fig. 11 Effect of s on nanoparticle concentration profile for Pr = Le = K = M = 1.0, N b =

N t = 0.5, Q = 0.1, α = 0.5
Table 1 Comparison of results for N ur when α = K = M = Q = s = 0, N b = N t = 0

Pr Wang [24] Gorla and Sidawi [25] Present results
0.07 0.0656 0.0656 0.0695
0.20 0.1691 0.1691 0.1694
0.70 0.4539 0.5349 0.4541
2.00 0.9114 0.9114 0.9120
7.00 1.8954 1.8905 1.8954
20.0 3.3539 3.3539 3.3539
70.0 6.4622 6.4622 6.4623
Table 2 Variation of N ur and Shr with N b, N t and Le when s = 0.5, K = 1.0, Pr = 1.0, Q =
0.1, M = 1.0, α = 0.5
Nb Nt Le = 1 Le = 10
N ur Shr N ur Shr
0.1 0.1 0.50278 0.25287 0.48513 4.6238
0.2 0.48406 −0.15312 0.45966 4.3467
0.3 0.46576 −0.52569 0.43552 4.1034
0.4 0.44788 −0.86599 0.41262 3.8909
0.5 0.43041 −1.1751 0.39089 3.7066
0.3 0.1 0.44173 0.44173 0.41204 4.8534
0.2 0.42444 0.40848 0.39010 4.7790
0.3 0.40757 0.34848 0.36947 4.7140
0.4 0.39110 0.24442 0.34984 4.6575
0.5 0.37502 0.15688 0.33120 4.6089
0.5 0.1 0.38546 0.62732 0.34938 4.8975
0.2 0.36959 0.56696 0.33066 4.8621
0.3 0.35412 0.51221 0.31292 4.8313
0.4 0.33903 0.46290 0.29611 4.8048
0.5 0.32430 0.41882 0.28015 4.7821
directly proportional to the applied magnetic field, B0 . This force inhibits momentum
development and decelerates the flow. The supplementary work done in dragging the
conducting nanofluid against the action of the magnetic field, B0 , is manifested as
thermal energy. This heats the conducting nanofluid and elevates temperatures. The
warming of the boundary layer therefore also aids in nanoparticle diffusion which
causes a rise in nanoparticle volume fraction, φ.
Figures 10 and 11 depict the effects of suction parameter S on velocity and con-
centration profile. It is noticed that both momentum and concentration boundary
layer thickness decrease with the increase in suction parameter.
In the present study, the local rate of heat transfer N ur and local rate of mass
transfer at the sheet Shr , defined in Eq. (16), are the important characteristics. The
numerical values of N ur and Shr are exhibited in Tables 1 and 2. Table 1 shows
Table 3 Comparison of CPU time in the two approaches (in seconds)

Number of Usual FEM Hybrid approach
elements
Element matrix Assembled Element matrix Assembled
matrix matrix
1 0.07 0.15 0.03 0.08
10 1.31 1.63 0.67 0.88
20 2.15 4.15 1.05 2.02
50 8.89 9.83 4.48 4.87
100 17.82 22.98 8.89 11.45
200 46.85 58.72 24.02 29.58
500 120.08 141.12 61.54 71.32
1000 425.05 503.39 211.43 212.18
that the excellent correlation between the current FEM computations and the earlier
results of Wang [24] and Gorla and Sidawi [25].
5 Conclusions
The problem of MHD boundary-layer flow of a viscoelastic nanofluid past a stretch-

ing sheet has been taken with numeric symbolic approach to exhibit the effect of
partial slip (i.e., Navier’s condition) and heat source/sink on the fluid flow and heat
transfer characteristics. The result can be summarized as follows:
1. With the increase in the second grade parameter α, the velocity and the momen-
tum boundary layer thickness increases, however, the temperature and nanopar-
ticles concentration decrease.
2. There is a decrease in the velocity, but temperature and concentration are found
to increase with increase in velocity slip parameter K .
3. Magnetic field decelerates the flow,as expected and enhances temperatures and
nanoparticle volume fraction (concentration) distributions in the boundary layer.
4. With increase in the slip parameter K , heat transfer rate and mass transfer rate
decrease which can be used for controlling the heat flow e.g., in reactors, etc.
5. With increase of thermophoretic number N t, the effect of velocity slip parameter
K on reduced Nusselt number N ur and reduced Sherwood number Shr increases
and decreases respectively.
6. The reduced Nusselt number and reduced Sherwood number show the effective-
ness of heat flow.
7. The results obtained by FEM and Hybrid approach tally nicely as shown in graphs
and a drastic saving in computational time as shown in Table 3. Obviously it will
be quite effective in more complicated problems. The CPU time which is a
subject of concern when dealing with large FEM meshes, in real-life problems
is expected to curtail drastically which will be of great use.
Acknowledgments Authors 1 and 3 are thankful to SERB, DST for providing financial support
for completing this paper under the project sponsored by them.
References
1. Kaminski, M.: On semi-analytical probabilistic finite element method for homogenization of

the periodic fiber-reinforced composites. Int. J. Numer. Meth. Eng. 86, 1144–1162 (2011)
2. Alns, M.S., Mardal, K.: On the efficiency of symbolic computations combined with code
generation for finite element methods. ACM Trans. Math. Softw. 37, 1–26 (2010)
3. Griffiths, D., Huang, J., Schiermeyer, R.P.: Elastic stiffness of straight-sided triangular finite
elements by analytical and numerical integration. Commun. Numer. Meth. Eng. 25, 247–262
(2009)
4. Videla, L., Baloa, T., Griffiths, D., Cerrolaza, M.: Exact integration of an 8-node plane elastic
finite element by symbolic computation. Numer. Meth. Part. Diff. Eqns. 24, 249–261 (2008)
5. Eyhermandy, D., Saad, R.: A dynamic approach for automating finite element code develop-
ment. In: 11th World Congress on Computational Mechanics (WCCM XI), July 20–25, 2014.
Spain, Barcelona (2014)
6. Choi, S.: Enhancing thermal conductivity of fluids with nanoparticles in developments and
applications of Non-Newtonian flows. ASME FED 231/MD 66, 99–105 (1995)
7. Choi, S., Zhang, Z., Yu, W., Lockwood, F., Grulke, E.: Anomalously thermal conductivity
enhancement in nanotube suspensions. Appl. Phys. Lett. 79, 2252–2254 (2001)
8. Masuda, H., Ebata, A., Teramae, K., Hishinuma, N.: Alteration of thermal conductivity and
viscosity of liquid by dispersing ultra-fine particles. Netsu Bussei 7, 227–33 (1993)
9. Buongiorno, J.: Convective transport in nanofluids. ASME J. Heat Transf. 128, 240–250 (2006)
10. Khan, W., Pop, I.: Boundary-layer flow of a nanofluid past a stretching sheet. Int. J. Heat Mass
Transf. 53, 2477–2483 (2010)
11. Kuznetsov, A., Nield, D.: Natural convective boundary-layer flow of a nanofluid past a vertical
plate. Int. J. Therm. Sci. 49, 243–247 (2010)
12. Makinde, O., Aziz, A.: Boundary layer flow of a nanofluid past a stretching sheet with a
convective boundary condition. Int. J. Therm. Sci. 50, 1326–1332 (2011)
13. McCormack, P., Crane, L.: Physical Fluid Dynamics. Academic Press, New York (1973)
14. Gupta, P., Gupta, A.: Heat and mass transfer on a stretching sheet with suction or blowing.
Can. J. Chem. Eng. 55, 744–746 (1977)
15. Dutta, B., Roy, P., Gupta, A.: Temperature field in the flow over a stretching sheet with uniform
heat flux. Int. Commun. Heat Mass Transf. 12, 89–94 (1985)
16. Chen, C., Char, M.: Heat transfer of a continuous stretching surface with suction or blowing.
J. Math Anal. Appl. 135, 568–580 (1988)
17. Rajagopal, K., Na, T., Gupta, A.: Flow of a viscoelastic fluid over a stretching sheet. Rheol.
Acta 23, 213–221 (1984)
18. Chang, W.: The non-uniqueness of the flow of a viscoelastic fluid over a stretching sheet. Q.
Appl. Math. 47, 365–366 (1989)
19. Wang, C.: Flow due to a stretching boundary with partial slip-an exact solution of the Navier
Stokes equations. Chem. Eng. Sci. 57, 3745–3747 (2002)
20. Sparrow, E., Cess, R.: Temperature dependent heat sources or sinks in a stagnation point flow.
Appl. Sci. Res. 10, 185–197 (1961)
21. Azim, M., Mamun, A., Rahman, M.: Viscous joule heating MHD-conjugate heat transfer for
a vertical flat plate in the presence of heat generation. Int. Commun. Heat Mass Transfer 37,
666–74 (2010)
22. Rana, P., Bhargava, R.: Numerical study of heat transfer enhancement in mixed convection
flow along a vertical plate with heat source/sink utilizing nanofluids. Commun. Nonlinear Sci.
Numer. Simulat. 16, 4318–4334 (2011)
23. Noghrehabadi, A., Pourrajab, R., Ghalambaz, M.: Effect of partial slip boundary condition on
the flow and heat transfer of nanofluids past stretching sheet prescribed constant wall temper-
ature. Int. J. Therm. Sci. 54, 253–261 (2012)
24. Wang, C.: Free convection on a vertical stretching surface. J. Appl. Math. Mech. 69, 418–4334
(1989)
25. Gorla, R., Sidawi, I.: Free convection on a vertical stretching surface with suction and blowing.
Appl. Sci. Res. 52, 247–257 (1994)
Some Advanced Finite Element Techniques
for the Problems of Mechanics: A Review
Rama Bhargava and Rangoli Goyal
Abstract Reduced efficiency and difficult complex domain discretization for grid-
based numerical methods, e.g. FEM/FVM have led to the formulation of advanced
finite element techniques. Propagation of fire, detection of cracks in ice or bones,
bursting of stars, etc. are some such domains where the grid-based methods fail. The
present paper contains a review of some advanced techniques—αFEM and MeshFree
Methods, along with the details about their implementation and development.
Keywords Meshfree Methods · Alpha FEM · EFGM · MLPG · N-SFEM · Paral-

lelization
1 Introduction
A mathematical model is the best approximation to the physical world. Such models
are assembled based on guaranteed conservation principles and/or empirical obser-
vations. Numerical methods are essential for the effective simulation of the mathe-
matical model as the underlying partial differential equations (PDE) usually have to
be approximated. Many numerical methods have been developed to achieve this task.
The grid-based methods introduce a finite number of nodes and can be based on the
principles of weighted residual methods. The three classical choices for grid-based
methods are the finite difference method (FDM), the finite element method (FEM)
and the finite volume method (FVM).
The FDM is the oldest among the grid-based methods. It is based upon the appli-
cation of a local Taylor expansion to approximate the differential equations. It uses a
topologically quadrangular network of lines to build the mesh for discretization of the
R. Bhargava · R. Goyal (B)

e-mail: goyaldma@iitr.ac.in
R. Bhargava
e-mail: rbharfma@iitr.ac.in

DOI 10.1007/978-81-322-2485-3_57
716 R. Bhargava and R. Goyal
PDE. This method fails when we have complex geometries in multiple dimensions.
This drawback encouraged the use of an integral form of the PDEs and consequently
the development of the finite element and finite volume techniques. FEM and FVM
involve subdividing the domain into a large number of finite elements/control vol-
umes and then solving the governing equations of fluid flow. A system of algebraic
equations is formed and solved using various iterative methods. The numerical meth-
ods differ in the definition and derivation of the algebraic equations.
2 Failure of Grid-Based Methods
The Finite Element Method (FEM) [1] and Finite Volume Method (FVM) [2] may be
the most well-known members of these thoroughly developed mesh-based methods.
The grid-based methods need a priori definition of the connectivity of the nodes,
i.e. they rely on a mesh. Element-dependent solution is obtained with mesh-based
methods, i.e. shape quality of elements and element density affect the solution. Dis-
torted or low-quality meshes lead to higher errors and demand re-meshing, a time
and human effort consuming task, which is not guaranteed to be feasible in finite
time for complex three-dimensional geometries. Also, they are not well suited to
treat problems with moving discontinuities or the ones that do not align with ele-
ment edges. Re-meshing is difficult for three dimensions and requires projection of
quantities between consecutive meshes with significant degradation of accuracy.
In contrast, a comparably new class of numerical methods has been developed
which approximates partial differential equations only based on a set of nodes without
the need for an additional mesh. The present paper is devoted to a review of a
few advanced finite element techniques, which are highly efficient for solving the
mathematical models, in particular the problem of mechanics.
3 MeshFree Methods
3.1 Introduction
Definition 1 MeshFree method is a method used to establish system algebraic equa-

tions for the whole problem domain without the use of a predefined mesh for the
domain discretization.
MeshFree methods (MFs) [3, 4] use a set of nodes distributed within the problem
domain as well as sets of nodes distributed on the boundaries of the domain to repre-
sent (not discretize) the problem domain and its boundaries. These sets of distributed
Some Advanced Finite Element Techniques … 717
nodes are called as field nodes. These nodes do not form a mesh, implying that no
priori information about the relationship between the nodes (for the interpolation or
approximation of the unknown functions of field variables) is required.
3.2 Features of MeshFree Methods
We list some of the important features of Meshfree Methods (MFs) while comparing
them with the analogous properties of mesh-based methods.
• Absence of Mesh
In MFs the connectivity of nodes is determined at run-time. There is no sensitivity
regarding the mesh alignment. h-adaptivity is simpler as nodes have to be added,
and their connectivity will be determined at run-time. p-adaptivity is also simpler
as compared to mesh-based methods.
No re-meshing during the solution for the problems with large deformations of
the domain or moving discontinuities is required.
• Continuity of shape functions
The shape functions for MFs can be constructed to have any desired order of
continuity. In problems where the discontinuities can be physically justified such
as cracks, different material properties, etc., the continuity of shape functions and
derivatives is not desirable.
• Convergence
Numerical experiments suggest that the convergence results for MFs are signifi-
cantly better than the results obtained by mesh-based shape functions [5]. However,
theory fails to predict this for higher order of convergence.
• Computational effort
MFs are often substantially more time-consuming than their mesh-based counter-
parts. The shape functions for MFs are complex in nature unlike the polynomial
type shape functions for mesh-based methods. Thus, more number of integration
points are required for sufficiently accurate evaluation of the integrals of weak
form. The resulting global system of equations has in general a larger bandwidth
for MFs than for comparable mesh-based methods [6]. However, the assembly of
the resulting matrix is simpler.
• Essential boundary conditions
The shape functions of most of the MFs do not satisfy the Kronecker delta property,
i.e. φi (x j ) = δij . Thus, the imposition of essential boundary conditions requires
attention and may lower the convergence of the method.
3.3 Comparison—FEM and MeshFree Methods

FEM MeshFree methods
Element Mesh Yes No
Shape function formulation Based on predefined Based on local support
elements domains
Stiffness matrix Symmetric May or may not be depending
on technique used
Imposition of essential boundary Easy and standard Exceptional treatments may be
condition required depending on method
used
Computation speed Fast Slower compared to FEM
depending on method used
Accuracy Accurate compared to Usually more accurate than
FDM FEM
Adaptive analysis Difficult for 3D cases and Easier
complex geometries
Stage of development Well developed Young, with challenging issues
Availability of commercial Many Few
software packages
3.4 Solution Procedure for MeshFree Methods
We now list the steps in a Meshfree Method [3].

1. Domain Representation:
In the MFs, the problem domain and its boundary are modelled and represented by
using sets of nodes distributed in the problem domain and on its boundary. These
nodes carry the values of the field variables in MFs formulation, and are thus called
field nodes. The density of the nodes depends on desired accuracy and existing
resources. The nodal distribution is usually not uniformly relatively fine (coarse)
discretization in regions where a high (low) gradient of strains and/or stresses is
expected. The density can be automatically controlled using adaptive algorithms in
the code.
2. Function interpolation/approximation:
Since there is no mesh in MFs, the field variable (e.g. a component of the displace-
ment) u at any point at x = (x, y, z) within the problem domain is interpolated using
function values at field nodes within a small local support domain of the point at x,
i.e. u(x) = i=1
n φ (x)u , where n is the number of nodes which are included in the
i i
local support domain of x, u i is the nodal field variable at ith node and φi (x) is the
shape function at ith node determined using the nodes included in support domain
of x.
A local support domain of a point x governs the number of nodes to be used

to support or approximate the function value at x. It can have different shapes and
dimensions, usually circular or rectangular. For different points of interest x, there
can be differently dimensionalized and shaped support domains.
3. Formation of system equations
The discrete equations of MFs can be framed using the shape functions and strong- or
weak-form system equations. The partial differential equation for the mathematical
model is said to be the strong form of system equation. The weak form places a
weaker consistency on the approximate function and is achieved by introducing
an integral operation to the system equation based on a mathematical or physical
principle. These equations are often written in nodal matrix form and are assembled
into the global system matrices for the entire problem domain. The discretized system
equations of MFs are similar to those of FEM in terms of bandwidth. They can be
symmetric or asymmetric depending on the method used.
4. Solve the global MFs equations
Solving for symmetric (or asymmetric) system of equations using either the standard
linear algebraic equation solvers (for static problems), such as the Gauss elimination
method, LU decomposition method, etc. or the standard eigenvalue equation solvers
(for free vibration and buckling problems), such as Jacobis method, Givens method,
Inverse iteration, etc.
3.5 Meshfree Interpolation/Approximation Techniques
A good shape function should satisfy certain following basic requirements [3].
• Ideally, the shape function should possess the Kronecker delta function property
making it easier for imposition of essential boundary conditions.
• The algorithm must be stable and computationally efficient.
• The construction of shape functions should satisfy a certain order of consistency
to ensure the convergence of numerical results.
• Smaller and compact domain for field variable approximation so that the banded
system matrix can be handled with good computational efficiency.
3.6 Categories of Meshfree Methods
The idea of meshfree analysis dates back from 1977, with Monaghan and Gingold
[7] developing a Lagrangian method based on the Kernel Estimates method to model
astrophysics problems such as exploding stars and dust clouds that had no boundaries.
This method, named Smoothed Particle Hydrodynamics (SPH), is a particle method
based on the idea of replacing the fluid by a set of moving particles and transforming
the governing partial differential equations into the kernel estimates integrals.
The first meshfree method based on the Galerkin technique was only introduced
over a decade after Monaghan and Gingold first published the SPH method. The
Diffuse Element Method (DEM) was introduced by Nayroles and Touzot in 1991.
Many authors believe that it was only after the DEM that the idea of a meshfree
technique appealed to the research community. The idea behind the DEM was to
replace the FEM interpolation within an element by the Moving Least Square (MLS)
local interpolation.
The three categories of MFs are given in the following table.
Categories MeshFree approximation

techniques
Integral representation Smoothed Particle
Hydrodynamics (SPH)
Reproducing Kernel Particle
Method (RKPM)
Series representation Moving Least Squares (MLS)
Point Interpolation Methods (PIM,
RPIM)
Partition of Unity (PU) methods
Differential representation General Finite difference method
(GFDM)
• In the integral representation method, the function is characterized using its

information in a local domain (smoothing domain or influence domain) with the
help of a weighted integral operation. The consistency is achieved by selecting
an appropriate weight function. It is often used in the smoothed particle hydrody-
namics (SPH).
• The series representation methods have an extensive history. They are well estab-
lished in FEM and are now used in Meshfree methods based on arbitrarily dis-
tributed nodes. The consistency is ensured by ensuring the completeness of the
basis functions. The moving least square (MLS) approximation is the most widely
used method. The point interpolation method (PIM) using radial basis function (or
RPIM) is also often used.
• The differential representation method has also been established and in use for
a long time in the finite difference method (FDM). The finite difference approx-
imation is not globally compatible, and the consistency is ensured by the theory
of Taylor series. Differential representation methods are generally used for estab-
lishing system equations based on strong-form formulations, such as FDM and
the general finite difference method (GFDM).
3.7 Smoothed Particle Hydrodynamics Approach
The SPH method was introduced in 1977 by Lucy. The name SPH comes from the
smoothing character of the particles point properties to the kernel function, therefore
leading to a continuous field. Similar to finite element formulation, u h (x) can be
written as u h (x) = φi (x)u i , where, φi (x) are the SPH shape functions given by
φi (x) = W (x − xi )Vi , where Vi represents the volume of node i. The SPH shape
functions satisfy neither the Kronecker Delta property nor Partition of Unity.
3.8 Reproducing Kernel Particle Method (RKPM)
The reproducing kernel particle method (RKPM) [8] is an advancement of the con-
tinuous SPH approximation near the boundaries. For increasing the order of com-
pleteness of the approximation, a correction function C(x, ξ ) is introduced into the
approximation:
u h (x) = u(ξ )W (x − ξ, h)C(x, ξ )dξ

where C(x, ξ ) is the correction function. An example of the correction function

in one dimension is C(x, ξ ) = c1 (x) + c2 (x)(ξ − x), where, c1 (x) and c2 (x) are
coefficients which are evaluated by enforcing the corrected kernel to reproduce the
function.
A remarkable point to be observed is that if we choose: Vi = 1, the RKPM and
MLS are the same.
3.9 Moving Least Squares Approximation
This method was presented by Shepard in late 1960s for assembling smooth approx-
imations to definite cloud of points. The MLS approximation [3, 9] has two main
features that make it popular:
(1) The approximated field function is continuous and smooth in the complete prob-
lem domain.
(2) It is capable of constructing an approximation with the preferred order of con-
sistency.
The MLS approximation of field variable u(x) is defined as
u h (x) = mj=1 p j (x)a j (x) = p T (x)a(x)

where p(x) is the basis function of coordinates, x T = [x, y] for 2D problem, and
m is number of basis functions. The basis function is constructed using monomials
from the Pascal triangle to guarantee minimum completeness. a(x) is a vector of
coefficients, which can be obtained by minimizing the following weighted discrete
L 2 norm.
J = i=1
n
W (x − xi )[ p t (xi )a(x) − u i ]2
where n is the number of nodes in support domain of x and u i is the nodal parameter
of u at x = xi . The number of nodes, n used in MLS approximation is typically much
greater than number of coefficients, m because of which the approximated function,
u h does not pass through the nodal values.
The stationarity of J with respect to a(x) gives
∂J
=0
∂a
which gives set of linear equations A(x)a(x) = B(x)Us , where, Us is the vector that
has nodal parameters of field function for all nodes in support domain and A(x) is the
weighted moment matrix. The above equation is solved for a(x) = A−1 (x)B(x)Us
in the approximation of field variable u h (x). The shape function for the ith node is
defined by
φi (x) = i=1
m
p j (x)(A−1 (x)B(x)) ji = p T (x)(A−1 B)i
The consistency of MLS approximation depends on the complete order of the mono-
mial taken in the polynomial basis. If the complete order of monomial is k, then the
shape function will possess C k consistency.
3.10 Point Interpolation Method
In this method, approximation is acquired by allowing the interpolation function to

pass through the function values at each scattered node within the defined domain
of support. The formulation of PIM starts with following finite series representation
u h (x, x Q ) = i=1
n
Bi (x)ai (x Q )
where Bi (x) are basis functions defined in coordinate space x T = [x, y, z], n is the
number of nodes in support domain of x Q and ai (x Q ) is the coefficient for basis
function Bi (x).
Two types of PIM have been developed, one that uses polynomial functions and
the other that uses radial basis functions.
3.11 Various MeshFree Methods
Due to brevity of paper, we will be discussing only a few methods. There are various
other methods such as, Diffuse Element Method, Least-squares Meshfree Method,
Local Boundary Integral Equation, Natural Element Method, Hybrid meshfree meth-
ods [10], etc.
3.12 Element-Free Galerkin Method
It mainly deals with FEM shape function being replaced by a nodal-based approxi-
mation. Moving least square (MLS) approximation is used for the construction of the
shape function in EFGM [11, 12]. For the calculation of system matrices, the cells
of the background mesh are used for integration and hence, it is not truly a meshless
method. The typical solution procedure using EFGM are:
• Nodal discretization of the solution area.
• Construct the approximation function using Moving Least Squares Approxima-
tion.
• Build the equivalent form of the physical problems (PDEs) using the Galerkin
Weak form.
• Substitute the approximation and their derivatives into the equivalent form, and
construct the solution matrix.
• Enforcement of essential boundary conditions using Penalty method. In penalty
method, we introduce a penalty factor to correct the difference between the variable
of MLS approximation and the given variable on the essential boundary.
• Solve the solution matrix to obtain the results.
3.13 Meshless Local Petrov–Galerkin
The main idea of MLPG [13, 14] is that the implementation of the integral form
of the weighted residual method is restricted to a very small local subdomain of a
node. This means that weak form is fulfilled at each node in problem domain in local
integral sense and thus, it is more stable. MLS shape functions are used. For building
the equivalent form of the physical problems, the Petrov–Galerkin form is used which
gives us the freedom to choose weight and trail functions individualistically. There
are two major drawbacks of MLPG method.
1. Asymmetry of matrices because of Petrov–Galerkin formulation.
2. For domains intersecting with the boundary of problem domain, the local inte-
gration is tricky.
3.14 Node-Based Smooth FEM (N-SFEM)
The smoothed finite element procedures were developed by joining Meshfree meth-
ods and Finite Element Method. It gives upper bound solutions to problems of solid
mechanics. In this numerical procedure, we do not construct the shape functions
explicitly [15]. We perform point interpolation using nodes within the element that
cloud the point of interest. The procedure for NS-FEM [16, 17] briefly follows:
1. Discretize the domain into a set of elements to find the node coordinates and
connectivity between the elements.
2. Compute the area of cell k associated with node k, find neighbouring cells for
each node.
3. For all the nodes, using the node connectivity, evaluate the stiffness matrix and
force vector associated with the cell of node. Assemble to form the system stiffness
matrix and force vector.
4. Implement essential boundary conditions.
5. Solve system equations to find displacements and hence, value of stress and strain
at nodes of interest.
4 αFEM
The αFEM [17–19] is a novel advancement of FEM in which the gradients of strains
are scaled by a factor α ∈ [0, 1]. The idea is to find a combined model of the
standard FEM and NS-FEM which makes the finest use of the upper bound property
of the N-SFEM and the lower bound property of the standard FEM. The procedure
is similar to standard FEM except that the stiffness matrix and the Jacobi matrix are
substituted by scaled strain matrix and equivalent scaled Jacobi matrix, respectively
[20]. For overestimation problems, exact α approach is used and for underestimation
problems, zero α approach is used.
The area of triangular element (say Ve ) is divided into four parts with scale factor
α. Three quadrilaterals at corners are scaled down by a factor of (1 − α 2 ) with an
equal area of (1−α3 )Ve . The remaining Y-shaped part in the middle has area α 2 Ve .
2
The Y-shaped area is evaluated using FEM while the three quadrilaterals at corners
are evaluated using N-SFEM [19].
Hence,
K IαJF E M = k=1
n
K INJ−S
(k)
FEM
+ k=1
e
K IFJE(k)
M
where n is the total number of nodes and e is the total number of elements in problem
domain.
For 2D problems using triangular elements
k=1
e
K IFJE(k)
M
= B IT M B J α 2 Ve
k=1
e
K INJ−S
(k)
FEM
= (1 − α 2 )B IT M B J V k
where Bi is the strain gradient matrix of the ith element around node k, M is matrix
of material constants, Ve is the volume of element e and V k is volume of cell k.
The result of αFEM is a continuous function of α from results of N-SFEM to that
of standard FEM. According to numerical experiments on different linear problems,
it is recommended to use directly α = 0.6 for 2D problems and α = 0.7 for 3D
problems. This method is variationally consistent and has the same computational
complexity as that of FEM.
5 Hybrid Meshfree Methods
Attempts have also been made to couple Meshfree methods that are formulated
using MLS shape functions and Meshfree methods that are formulated using Point
interpolation method (PIM) shape functions or Finite element (FE) shape functions
[10]. The aim is to simplify the procedure of imposing the boundary conditions.
Some examples of hybrid Meshfree methods are: SPH coupling with FEM, EFG
coupling with BEM, MLPG coupling with BEM or FEM, etc.
6 Parallelization of EFGM
Although, EFGM is very efficient for solving problems of irregular geometry or phase
transition problem, the major disadvantage of EFGM is the increased computational
cost. The additional computational cost of EFGM is from several sources listed as
follows:
1. The need to identify the number of nodes in support domain of a quadrature

point.
2. The relative complexity of the shape functions, which increase the cost of eval-
uating them and their derivatives.
3. The additional expense of dealing with essential boundary conditions. To over-
come this difficulty of high computational time, parallel implementation of
EFGM can be done. In EFGM, the process of forming stiffness matrices at each
quadrature point is totally independent. Therefore, in a parallel computing set-
up, stiffness matrices can be computed at different work stations and finally
assembled at the master server.
7 Related Problems
Fluid flow and thermal effects during phase change (melting and solidification) are
of great interest in a number of manufacturing processes, where a solid material is
formed by the freezing of a liquid. Its main characteristic is that a moving interface
separates two phases with different physical properties [21]. Temperature differences
in the melt give rise to buoyancy forces that produce significant convective flow. Due
to the problem complexity, direct application of numerical methods (FEM) to the
problems of phase change is not an easy task.
As the nature of surface influences the rate of heat transfer, the analysis of natural
convection heat transfer with different types of heat transfer surfaces [22, 23] (e.g.
wavy surfaces or rough surfaces) becomes important, e.g. the transfer of heat gener-
ated due to friction in car tyres. Numerical simulation (FEM and BEM) of fluid flow
problems with complex geometries is a challenge as they require re-meshing at each
stage of simulation.
Therefore, to tackle the above mechanics problems, Meshfree techniques are used.
Li et al. [18] have utilized αFEM for simulation of a phase transition problem
during the cryosurgery treatment of tumour tissue, in which transition occurs between
the healthy blood tissues and tumour tissues. In their study, a comparison has been
shown between obtained FEM and αFEM results, which illustrate that αFEM results
obtained with 291 nodes are in well agreement with FEM results obtained with 12876
nodes. Li et al. [24] applied αFEM to analyze 2D underwater exterior scattering
problems in the unbounded domain.
8 Conclusions
In this paper, an overview of Meshfree methods and FEM has been presented. We have
discussed properties and advantages of Meshfree methods compared to standard finite
elements. The MFs themselves have been explained in detail, taking into account
the different viewpoints and origins of each method. We have concentrated more
on pointing out important characteristic features rather than on explaining how the
method functions. Since standard Meshfree methods do not fulfil the KroneckerDelta
property, essential boundary conditions cannot be imposed as straightforwardly as
in finite element methods. We have summarized on how to incorporate essential
boundary conditions using the penalty method.
From the literature review, it is observed that the results obtained with αFEM are
much more accurate and can be used to improve solutions for non-linear problems
of large deformation.
We hope this paper to be a helpful tool for the reader’s successful work with
advanced finite element techniques.
Acknowledgments The second author would like to thank Department of Science and Technology,
Government of India, for its financial support through the award of a research grant.
References
1. Reddy, J.N.: An Introduction to the Finite Element Method. McGraw-Hill Inc., New York
(2005)
2. Causona, D.M., Mingham, C.G., Qia, L.: Introductory Finite Volume Methods for PDEs. Ventus
publishing (2011)
3. Liu, G.R.: Mesh Free Methods Moving beyond the finite element Method. CRC Press, USA
(2003)
4. Babuska, I., Banerjee, U., Osborn, J.E.: Meshless and generalized finite element methods: a
survey of some major results. Meshfree Methods for Partial Differential Equations, vol. 26.
Springer Verlag, Berlin (2002)
5. Li, S., Liu, W.K.: Meshfree and particle methods and their applications. Appl. Mech. Rev. 55,
1–34 (2002)
6. Fries, T., Matthies, H.: Classification and Overview of Meshfree Methods. Scientific Comput-
ing, Germany (2003)
7. Gingold, R.A., Monaghan, J.J.: Smoothed particle hydrodynamics: theory and application to
non-spherical stars. Mon. Not. R. Astron. Soc. 181, 375389 (1977)
8. Chen, J.S., Yoon, S., Wang, H.P., Liu, W.K.: An improved reproducing kernel particle method
for nearly incompressible finite elasticity. Comp. Methods Appl. Mech. Eng. 181, 117–145
(2000)
9. Breitkopf, P., Touzot, G., Villon, P.: Consistency approach and diffuse derivation in element
free methods based on moving least squares approximation. Comput. Assist. Mech. Eng. Sci.
5, 479501 (1998)
10. Pian, T.H.H., Wu, C.C.: Hybrid and Incompatible Finite Element Methods. CRC Press, Boca
Raton (2006)
11. Chung, H.J., Belytschko, T.: An error estimate in the EFG method. Comput. Mech. 21, 91–100
(1998)
12. Beissel, S., Belytschko, T.: Nodal integration of the element-free Galerkin method. Comp.
Methods Appl. Mech. Eng. 139, 49–74 (1996)
13. Atluri, S.N., Zhu, T.: A new meshless local PetrovGalerkin (MLPG) approach in computational
mechanics. Comput. Mech. 22, 117127 (1998)
14. Gu, T.Y., Liu, G.R.: A meshless local PetrovGalerkin (MLPG) method for free and forced
vibration analyses for solids. Comput. Mech. 27(3), 188198 (2001)
15. Dai, K.Y., Liu, G.R., Nguyen, T.T.: An n-sided polygonal smoothed finite element method
(nSFEM) for solid mechanics. Finite Elem. Anal. Des. 43, 847860 (2007)
16. Liu, G.R., Dai, K.Y., Nguyen, T.T.: A smoothed finite element method for mechanics problems.
Comput. Mech. 39, 859877 (2007)
17. Liu, G.R., Trung, N.T.: Smoothed Finite Element Methods. CRC Press, USA (2010)
18. Li, E., Liu, G.R., Tan, V., He, Z.C.: An efficient algorithm for phase change problem in tumor
treatment using α-FEM. Int. J. Therm. Sci. 49, 1954–1967 (2010)
19. Liu, G.R., Nguyen-Thoi, T., Lam, K.Y.: A novel alpha finite element method (αFEM) for exact
solution to mechanics problems using triangular and tetrahedral elements. Comput. Methods
Appl. Mech. Eng. 197, 38833897 (2008)
20. Trung, N.T., Hung, N.X.: About applying directly the alpha finite element method for solid
mechanics using triangular and tetrahedral elements. Vietnam J. Mech. VAST 32(4), 235–246
(2010)
21. Belhamadia, Y., Kane, A.S., Fortin, A.: An enhanced mathematical model for phase change
problems with natural convection. Int. J. Numer. Anal. Model. 3(2), 192–206 (2012)
22. Nikfar, M., Mahmoodi, M.: Meshless local PetrovGalerkin analysis of free convection of nano
fluid in a cavity with wavy side walls. Eng. Anal. Boundary Elem. 36, 433445 (2012)
23. Das, P.K., Mahmud, S.: Numerical investigation of natural convection inside a wavy enclosure.
Int. J. Therm. Sci. 42, 397406 (2003)
24. Li, W., Li, Y.F., Chai, Y.B.: Calculation of underwater acoustic scattering problems in
unbounded domain using the alpha finite element method, ICCM2014. Cambridge, England
Duality for a Nondifferentiable
Multiobjective Second-Order Fractional
Programming Problem Involving
(F, α, ρ, d)—V -type-I Functions
Ramu Dubey and S.K. Gupta
Abstract In this paper, a nondifferentiable multiobjective fractional programming

problem in which a support function appears in the numerator and denomina-
tor of each objective function as well as in each constraint function. We propose
Schaible type dual for a nondifferentiable second-order fractional programming prob-
lem. Next, we prove appropriate duality results using second-order (F, α, ρ, d)—
V -type-I convexity assumptions.
Keywords Fractional programming problem · Support function · Duality results ·

(F, α, ρ, d)—V -type-I functions
1 Introduction
Mangasarian [1] introduced the concept of second-order duality for nonlinear pro-
gramming problems. By introducing an additional vector p ∈ R n , he formulated the
second-order dual and established duality theorems under convexity assumptions.
One significant practical use of duality is that it provides bounds for the value of the
objective function. Second-order duality may provide tighter bounds than first-order
duality because there are more parameters involved in it. In the dual formulation of
Mangasarian the same vector p appears at three places. In [2], Hanson introduced
the dual by replacing the same vector p by three different vectors p, q, and r ∈ R n
thus providing applicability to a wider class of functions that may give still tighter
bounds than Mangasarian’s dual. The author also established weak and strong duality
theorems under second-order type-I assumptions.
Mond [3] introduced second-order convex functions and established second-order
duality results. Bector and Chandra [4] formulated second-order Mond-Weir type
R. Dubey (B) · S.K. Gupta

e-mail: rdubeyjiya@gmail.com
S.K. Gupta
e-mail: skgiitr@gmail.com

DOI 10.1007/978-81-322-2485-3_58
730 R. Dubey and S.K. Gupta
dual for a nondifferentiable fractional program and established duality results using
the concept of second-order pseudoconvexity and quasiconvexity. Jeyakumar [5] and
Yang [6] also discussed second-order dual formulation under ρ-convexity and its
generalizations. Suneja et al. [7] presented a pair of Mond-Weir type multiobjective
second-order symmetric dual programs and obtained duality results using second-
order η-convex function. Preda [8] defined (F, ρ)-convex functions which generalize
the definition of second-order (F, ρ)-convex functions given in Aghezzaf [9] and
Zhang and Mond [10]. Ahmad and Husain [11] extended the concept to second-order
(F, α, ρ, d)-convex functions and obtained duality results for a second-order Mond-
Weir type multiobjective dual. Recently, Hachimi and Aghezzaf [12] extended the
notion of (F, α, ρ, d)-type I functions to second-order generalized (F, α, ρ, d)-type
I functions and established mixed duality results.
In this paper, we have formulated a dual model for a nondifferentiable multiob-
jective second-order fractional programming problem. In this dual, we have gener-
alized the models already existing in the literature replacing the same vector p ∈ R n
involved in the objective and constraint functions with five different vectors p, q, r, s
and t ∈ R n and thus provide applicability to a wider class of functions and may give
more tighter bounds than Mangasarian dual. Using (F, α, ρ, d)−V -type-I functions,
duality theorems are derived for a Schaible type dual program.
2 Notations and Preliminaries
Definition 1 [13] Let C be a compact convex set in R n . The support function of C

is defined by
S(x|C) = max{x T y : y ∈ C}.
A support function, being convex and everywhere finite, has a subdifferential, that
is, there exists a z ∈ R n such that
S(y|C) ≥ S(x|C) + z T (y − x), ∀x ∈ C.
The subdifferential of S(x|C) is given by
∂ S(x|C) = {z ∈ C : z T x = S(x|C)}.
For a convex set D ⊂ R n , the normal cone to D at a point x ∈ D is defined by
N D (x) = {y ∈ R n : y T (z − x) ≤ 0, ∀z ∈ D}.
When C is a compact convex set, y ∈ NC (x) if and only if S(y|C) = x T y, or

equivalently, x ∈ ∂ S(y|C).
Throughout the paper, we use the index sets K = {1, 2, . . . , k} and M =
{1, 2, . . . , m}.
Duality for a Nondifferentiable Multiobjective Second-Order Fractional … 731
Consider the following multiobjective programming problem:
(P) Minimize f (x) = ( f 1 (x), f 2 (x), . . . , f k (x))

subject to X = {x ∈ X : h j (x) ≤ 0, j ∈ M},
0
where X ⊂ R n is open and f i : X → R and h j : X → R, i ∈ K , j ∈ M are twice

differentiable functions on X.
Definition 2 [14] A solution x 0 ∈ X 0 is said to be an efficient (or Pareto optimal)

solution of (P) if there exists no x ∈ X 0 such that for some i ∈ K , f i (x) < f i (x 0 )
and f j (x) ≤ f j (x 0 ), ∀ j ∈ K .
Definition 3 [12] A functional F : X × X × R n → R is said to be sublinear with

respect to the third variable if for all (x, u) ∈ X × X,
(i) Fx,u (a1 + a2 ) ≤ Fx,u (a1 ) + Fx,u (a2 ), for all a1 , a2 ∈ R n ,
(ii) Fx,u (αa) = αFx,u (a), for all α ∈ R+ and a ∈ R n .
Definition 4 [15] The functions ( f, h) are said to be second-order (F, α, ρ, d) −

V -type-I at u ∈ X if there exist vectors α = (α11 , α21 , . . . , αk1 , α12 , α22 , . . . , αm
2 ),
ρ = (ρ1 , ρ2 , . . . , ρk , ρ1 , ρ2 , . . . , ρm ) and d ∈ R where αi , α j : X × X → R+ \{0},

1 1 1 2 2 2 1 2
ρi1 , ρ2j ∈ R for all i ∈ K , j ∈ M and d : X × X → R such that for each x ∈ X 0

and p, q, r, s, t ∈ R n and for all i ∈ K , j ∈ M, we have

1
f i (x) − f i (u) ≥ Fx,u αi1 (x, u) ∇ f i (u) + ∇ 2 f i (u) p − q T ∇ 2 f i (u)r + ρi1 d 2 (x, u)
2
and

1
−h j (u) ≥ Fx,u α2j (x, u) ∇h j (u) + ∇ 2 h j (u) p − s T ∇ 2 h j (u)t + ρ2j d 2 (x, u).
2
Remark 1 If αi1 = α1 (x, y), α2j = α2 (x, y) and p = q = r = s = t, then the

Definition 4 reduces to Hachimi and Aghezzaf [12].
Definition 5 [15] The functions ( f, h) are said to be second-order semistrictly

(F, α, ρ, d) − V -type-I at u ∈ X if there exist vectors α = (α11 , α21 , . . . , αk1 , α12 , α22 ,
2 ), ρ = (ρ1 , ρ1 , . . . , ρ1 , ρ2 , ρ2 , . . . , ρ2 ) and d ∈ R where α1 , α2 : X × X →
. . . , αm 1 2 k 1 2 m i j
R+ \ {0}, ρi1 , ρ2j ∈ R for all i ∈ K , j ∈ M and d : X × X → R such that for each
x ∈ X 0 and p, q, r, s, t ∈ R n and for all i ∈ K , j ∈ M, we have

1
f i (x) − f i (u) > Fx,u αi1 (x, u) ∇ f i (u) + ∇ 2 f i (u) p − q T ∇ 2 f i (u)r + ρi1 d 2 (x, u)
2
and

1
−h j (u) ≥ Fx,u α2j (x, u) ∇h j (u) + ∇ 2 h j (u) p − r T ∇ 2 h j (u)t + ρ2j d 2 (x, u).
2
Consider the following multiobjective non-differentiable programming problem:

f 1 (x) + S(x|C1 ) f 2 (x) + S(x|C2 ) f k (x) + S(x|Ck )
(PP) Minimize G(x) = , ,...,
g1 (x) − S(x|D1 ) g2 (x) − S(x|D2 ) gk (x) − S(x|Dk )
subject to x ∈ X 0 = {x ∈ X : h j (x) + S(x|E j ) ≤ 0, j ∈ M},
where x ∈ X ⊂ R n , f i , gi : X → R (i ∈ K ) and h j : X → R ( j ∈ M) are

continuously differentiable functions.
f i (.) + S(.|Ci ) ≥ 0 and gi (.) − S(.|Di ) > 0; Ci , Di and E j are compact
convex sets in R n and S(x|Ci ), S(x|Di ) and S(x|E j ) denote the support functions
of compact convex sets, Ci , Di and E j for all i ∈ K , j ∈ M, respectively.
Theorem 1 (Karush-Kuhn-Tucker type Necessary Condition) [16] Assume that x̄

is an efficient solution of (PP) at which the Kuhn-Tucker constraint qualification is
satisfied. Then there exist 0 < λ̄ ∈ R k , 0 ≤ ȳ j ∈ R m , z¯i ∈ R n , v¯i ∈ R n and
w¯ j ∈ R n , i ∈ K , j ∈ M such that

k
m
f i (x̄) + x̄ T z¯i
λ̄i ∇ + y¯j ∇(h j (x̄) + x̄ T w¯ j ) = 0,
gi (x̄) − x̄ T v¯i
i=1 j=1

m
y¯j (h j (x̄) + x̄ T w¯ j ) = 0,
j=1
x̄ T z¯i = S(x̄|Ci ),
x̄ T v¯i = S(x̄|Di ),
x̄ T w¯ j = S(x̄|E j ),
z¯i ∈ Ci , v¯i ∈ Di , w¯ j ∈ E j , i ∈ K , j ∈ M.
3 Duality Model
Consider the following Schaible type dual (DP) of (PP):
(DP) maximize (β1 , β2 , . . . , βk )

subject to

k
m
λi ∇ ( f i (u) + u T z i ) − βi (gi (u) − u T vi ) + y j ∇(h j (u) + u T w j )
i=1 j=1

k
m
+ λi (∇ 2 f i (u) − βi ∇ 2 gi (u)) p + y j ∇ 2 h j (u) p = 0, (1)
i=1 j=1
( f i (u) + u T z i ) − βi (gi (u) − u T vi ) ≥ 0, i ∈ K , (2)

k
λi q T (∇ 2 f i (u) − βi ∇ 2 gi (u))r ≤ 0, (3)
i=1

m
1
y j (h j (u) + u T w j − s T ∇ 2 h j (u)t) ≥ 0, (4)
2
j=1
z i ∈ Ci , vi ∈ Di , w j ∈ E j , i ∈ K , j ∈ M, (5)
y j ≥ 0, βi ≥ 0, λi > 0, i ∈ K , j ∈ M. (6)
We now discuss the weak duality, strong, and strict converse duality results for the
pair (PP) and (DP).
Theorem 2 (Weak Duality Theorem) Let x be a feasible solution for (PP) and
(u, β, z, v, y, λ, w, p, q, r, s, t) be feasible for (DP). Suppose that:

(i) For any i ∈ k, j ∈ M, f i (.) + (.)T z i − βi (gi (.) − (.)T vi ), h j (.) + (.)T w j is
second-order
(F, α, ρ, d) − V -type-I at u,
(ii) αi1 (x, u) = α2j (x, u) = α(x, u), ∀ i ∈ K and j ∈ M,

k
m
(iii) λi ρi1 + y j ρ2j ≥ 0.
i=1 j=1
Then, the following cannot hold:
f i (x) + S(x|Ci )
≤ βi , for all i ∈ K (7)
gi (x) − S(x|Di )
and
fr (x) + S(x|Cr )
< βr , for some r ∈ K . (8)
gr (x) − S(x|Dr )
Proof Let (7) and (8) hold, then using λi > 0, x T z i ≤ S(x|Ci ), x T vi ≤ S(x|Di ),
we have

k
λi ( f i (x) + x T z i − βi (gi (u) − x T vi )) < 0. (9)
i=1
Since ( f i (.) + (.)T z i − βi (gi (.) − (.)T vi ), h j (.) + (.)T w j ) is second-order

(F, α, ρ, d) − V -type- I at u, therefore for i ∈ K and j ∈ M, we have
( f i (x) + x T z i − βi (gi (x) − x T vi )) − ( f i (u) + u T z i − βi (gi (u) − u T vi ))

≥ Fx,u αi1 (x, u){∇( f i (u) + u T z i − βi (gi (u) − u T vi ))

+ (∇ 2 f i (u) − βi ∇ 2 gi (u)) p}
1
− q T (∇ 2 f i (u) − βi ∇ 2 gi (u))r + ρi1 d 2 (x, u) (10)
2
and

−(h j (u) + u T w j ) ≥ Fx,u α2j (x, u) ∇(h j (u) + u T w j ) + ∇ 2 h j (u) p
1
− s T ∇ 2 h j (u)t + ρ2j d 2 (x, u).
2
(11)
Using (2) and multiplying (10) by λi and (11) by y j , summing over i ∈ K and
j ∈ M, we get

k m
λi f i (x) + x T z i − βi (gi (x) − x T vi ) − y j h j (u) + u T w j
i=1 j=1

k
≥ Fx,u λi αi1 (x, u)∇( f i (u) + u T z i − βi (gi (u) − u T vi ))
i=1

m
+ α2j (x, u)y j ∇(h j (u) + u T w j )
j=1

k
m
+ αi1 (x, u)λi (∇ 2 f i (u) − βi ∇ 2 gi (u)) p + α2j (x, u)y j ∇ 2 h j (u) p
i=1 j=1
1
k
− λi q T (∇ 2 f i (u) − βi ∇ 2 gi (u))r
2
i=1

1
m k m
− y j s T ∇ 2 h j (u)t + λi ρi1 + y j ρ2j d 2 (x, u). (12)
2
j=1 i=1 j=1
Further, using αi1 (x, u) = α2j (x, u) = α(x, u) and feasibility conditions (3), (4) in
(12), we get

k
λi f i (x) + x T z i − βi (gi (x) − x T vi )
i=1

k
≥ Fx,u α(x, u) λi ∇( f i (u) + u z i − βi gi (u) − u vi
T T
i=1

m
+ y j ∇(h j (u) + u T w j )
j=1

k
m
+ λi ∇ 2 f i (u) − βi ∇ 2 gi (u) p + y j ∇ 2 h j (u) p
i=1 j=1
k

m
+ λi ρi1 + y j ρ j d 2 (x, u).
2
(13)
i=1 j=1
Using hypothesis (iii), we have

k
λi f i (x) + x z i − βi (gi (x) − x vi )
T T
i=1

k
≥ Fx,u α(x, u) λi ∇( f i (u) + u T z i − βi (gi (u)
i=1

m
− u T vi ))) + y j ∇(h j (u) + u T w j )
j=1

k
m
+ λi ∇ f i (u) − βi ∇ gi (u) p +
2 2
y j ∇ h j (u) p .
2
(14)
i=1 j=1
Using feasibility condition (1) and the result Fx,u (0) = 0, we get

k
λi f i (x) + x T z i − βi (gi (x) − x T vi ) ≥ 0,
i=1
which contradicts (9) and hence the result.
Theorem 3 (Strong Duality Theorem) If ū is an efficient solution of (PP) and let

the Kuhn-Tucker constraint qualification be satisfied. Then there exist λ̄ ∈ R k ,
ȳ ∈ R m , z¯i ∈ R n , v¯i ∈ R n and w¯ j ∈ R n , i ∈ K , j ∈ M, such that
(ū, β̄, z̄, v̄, ȳ, λ̄, w̄, p̄ = 0, q̄ = 0, r̄ , s̄ = 0, t¯) is a feasible solution of (DP)
and the objective function values of (PP)and (DP) are equal. Furthermore, if the
conditions of Theorem 1 hold for all feasible solutions of (PP) and (DP), then
(ū, β̄, z̄, v̄, ȳ, λ̄, w̄, p̄ = 0, q̄ = 0, r̄ , s̄ = 0, t¯) is an efficient solution of (DP).
Proof Since ū is an efficient solution of (PP) at which the Kuhn-Tucker constraint

qualification is satisfied, so by the Karush-Kuhn-Tucker type necessary conditions
(Theorem 1), there exist μ̄ ∈ R k , ȳ ∈ R m , z¯i ∈ R n , v¯i ∈ R n and w¯ j ∈ R n ,
i ∈ K , j ∈ M, such that

k
m
f i (ū) + ū T z¯i
μ̄i ∇ + y¯j ∇(h j (ū) + ū T w¯ j ) = 0, (15)
gi (ū) − ū T v¯i
i=1 j=1

m
y¯j (h j (ū) + ū T w¯ j ) = 0, (16)
j=1
ū T z¯i = S(ū|Ci ), (17)

ū T v¯i = S(ū|Di ), (18)
ū T w¯ j = S(ū|E j ), (19)
z¯i ∈ Ci , v¯i ∈ Di , w¯ j ∈ E j , (20)
μ̄i > 0, y¯j ≥ 0. (21)
Equation (15) can be written as

k
μ̄i f i (ū) + ū T z¯i
∇( f i (ū) + ū z¯i ) −
T
∇(gi (ū) − ū v¯i )
T
gi (ū) − ū T v¯i gi (ū) − ū T v¯i
i=1

m
+ y¯j ∇(h j (ū) + ū T w¯ j ) = 0. (22)
j=1
μ̄i f i (ū) + ū T z¯i

Setting λ̄i = and β̄i = , i ∈ K , we get
gi (ū) − ū T v¯i gi (ū) − ū T v¯i

k
m
λ̄i ∇ f i (ū) + ū T z¯i − β̄i (gi (ū) − ū T v¯i ) + y¯j ∇(h j (ū) + ū T w¯ j ) = 0,
i=1 j=1
(23)
f i (ū) + ū T z¯i − β̄i (gi (ū) − ū T v¯i ) = 0, (24)
λ̄i > 0, β̄i ≥ 0. (25)
Thus, (ū, β̄, z̄, v̄, ȳ, λ̄, w̄, p̄ = 0, q̄ = 0, r̄ , s̄ = 0, t¯) is feasible for (DP) and the
objective function values of (PP) and (DP) are equal.
We now show that (ū, β̄, z̄, v̄, ȳ, λ̄, w̄, p̄ = 0, q̄ = 0, r̄ , s̄ = 0, t¯) is an efficient
solution of (DP). If not, then there exists a feasible solution (u , β , z , v , y , w , λ̄,
p = 0 = q = 0, r = s = 0, t ) of (DP) such that
f i (ū) + ū T z¯i f i (u ) + u T z i
≤ , ∀i ∈ K
gi (ū) − ū v¯i
T gi (u ) − u T vi
and
fr (ū) + ū T z¯r fr (u ) + u T zr
< , for some r ∈ K .
gr (ū) − ū v¯r
T gr (u ) − u T vr
This contradicts the Theorem 2. Hence, (ū, β̄, z̄, v̄, ȳ, λ̄, w̄, p̄ = 0, q̄ = 0, r̄ ,
s̄ = 0, t¯) is an efficient solution of (DP).
Theorem 4 (Strict Converse Duality Theorem) Let x̄ be a feasible solution for (PP)
and (ū, β̄, z̄, v̄, ȳ, λ̄, w̄, p̄, q̄, r̄ , s̄, t¯) be feasible for (DP). Suppose that

k
(i) λ̄i f i (x̄) + x̄ T z̄ i − β̄i (gi (x̄) − x̄ T v̄i ) ≤ 0,
i=1
(ii) For any i ∈ K , j ∈ M, f i (.) + (.)T z̄ i − β̄i (gi (.) − (.)T v̄i ), h j (.) + (.)T w j
is second-order semi-strictly (F, α, ρ, d) − V -type-I at ū,
(iii) αi1 (x̄, ū) = α2j (x̄, ū) = α(x̄, ū), for all i ∈ K and j ∈ M,

k
m
(iv) λ̄i ρi1 + ȳ j ρ2j ≥ 0.
i=1 j=1
Then, x̄ = ū.
Proof We suppose that x̄
= ū and exhibit a contradiction. Since (ū, β̄, z̄, v̄, ȳ, λ̄, w̄,
p̄, q̄, r̄ , s̄, t¯) is feasible solution for (DP), then by the dual constraint (1), we have

k

m
Fx̄,ū λ̄i ∇ ( f i (ū) + ū T z̄ i ) − β̄i (gi (ū) − ū T v̄i ) + ȳ j ∇(h j (ū) + ū T w̄ j )
i=1 j=1

k
m
+ λ̄i (∇ 2 f i (ū) − β̄i ∇ 2 gi (ū)) p̄ + ȳ j ∇ 2 h j (ū) p̄ = 0. (26)
i=1 j=1
By hypothesis (ii), we get
( f i (x̄) + x̄ T z̄ i − β̄i (gi (x̄) − x̄ T vi )) − ( f i (ū) + ū T z̄ i − β̄i (gi (ū) − ū T v̄i ))

> Fx̄,ū αi1 (x̄, ū) ∇( f i (ū) + ū T z̄ i − β̄i (gi (ū) − ū T v̄i )) + (∇ 2 f i (ū) − β̄i ∇ 2 gi (ū)) p̄
1
− q̄ T (∇ 2 f i (ū) − β̄i ∇ 2 gi (ū))r̄ + ρi1 d 2 (x̄, ū) (27)
2
and

−(h j (ū) + ū T w̄ j ) ≥ Fx̄,ū α2j (x̄, ū)[∇(h j (ū) + ū T w̄ j ) + ∇ 2 h j (ū) p̄]
1
− s̄ T ∇ 2 h j (ū)t¯ + ρ2j d 2 (x̄, ū). (28)
2
Using (2) and multiplying (27) by λ̄i and (28) by ȳ j , summing over i ∈ K and
j ∈ M, we have

k
m
λ̄i [ f i (x̄) + x̄ T z̄ i − β̄i (gi (x̄) − x̄ T v̄i )] − ȳ j h j (ū) + ū T w̄ j
i=1 j=1

k
> Fx̄,ū λ̄i αi1 (x̄, ū)∇( f i (ū) + ū T z̄ i − β̄i (gi (ū) + ū T v̄i ))
i=1

m
k
+ α2j (x̄, ū) ȳ j ∇(h j (ū) + ū T w̄ j ) + αi1 (x̄, ū)λ̄i (∇ 2 f i (ū)
j=1 i=1

1
m k
− β̄i ∇ 2 gi (ū)) p̄ + α2j (x̄, ū) ȳ j ∇ 2 h j (ū) p̄ − λ̄i q̄ T (∇ 2 f i (ū) − β̄i ∇ 2 gi (ū))r̄
2
j=1 i=1

m
k
m
1
− y j s̄ T ∇ 2 h j (ū)t¯ + λ̄i ρi1 + y j ρ2j d 2 (x̄, ū). (29)
2
j=1 i=1 j=1
Finally, using αi1 (x̄, ū) = α2j (x̄, ū) = α(x̄, ū) and feasibility conditions (3) and (4)
in (29), we get

k
k
λ̄i [ f i (x̄) + x̄ T z̄ i − β̄i (gi (x̄) − x̄ T v̄i )] > Fx̄,ū α(x̄, ū) λ̄i ∇( f i (ū)
i=1 i=1

m
+ ū T z̄ i − β̄i (gi (ū) − ū T v̄i )) + ȳ j ∇(h j (ū)
j=1

k
m
+ ū w̄ j ) +
T
λ̄i (∇ f i (ū) − β̄i ∇ gi (ū)) p̄ +
2 2
ȳ j ∇ h j (ū) p̄
2
i=1 j=1
k

m
+ λ̄i ρi +
1
ȳ j ρ j d 2 (x̄, ū).
2
(30)
i=1 j=1
Using hypothesis (iii), we get

k
λ̄i [ f i (x̄) + x̄ T z̄ i − β̄i (gi (x̄) − x̄ T v̄i )]
i=1
k

> Fx̄,ū α(x̄, ū) λ̄i ∇( f i (ū) + ū z̄ i − β̄i (gi (ū) − ū v̄i ))
T T
i=1

m
+ ȳ j ∇(h j (ū) + ū T w̄ j )
j=1
⎛ ⎞

k
m
+ λ̄i ⎝(∇ 2 f i (ū) − β̄i ∇ 2 gi (ū)) p̄ + ȳ j ∇ h j (ū) p̄
2 ⎠ .
i=1 j=1
Using Eq. (26), the above inequality implies

k
λ̄i [ f i (x̄) + x̄ T z̄ i − β̄i (gi (x̄) − x̄ T v̄i )] > 0,
i=1
which contradicts hypothesis (i). Hence, x̄ = ū.
4 Special Cases
(i) For p = 0, q = 0, s = 0 and k = 1, then second-order dual (DP) becomes a

first-order problem, given by Husain and Jabeen [17].
(ii) For p = 0, q = 0, s = 0, Ci = {0}, Di = {0}, i ∈ K and E j = {0},
j ∈ M, then (PP) reduces to the problem studied in Egudo [14].
(iii) If k = 1, Ci = {0}, Di = {0}, i ∈ K and E j = {0}, j ∈ M, then (PP)
becomes the problem considered by Mond and Weir [18].
(iv) If Di = {0}, i ∈ K , then (PP) and (DP) reduced to the problems considered in
Gulati and Geeta [19].
Acknowledgments The authors express their sincere gratitude to the reviewer for her/his valuable
suggestions for improving the presentation of the paper. This research is supported by “Ministry of
Human Resource and Development,” New Delhi (India) in the form of fellowship to the first author.
References
1. Mangasarian, O.L.: Second-and higher-order duality in nonlinear programming. J. Math. Anal.

Appl. 51, 607–620 (1975)
2. Hanson, M.A.: Second-order invexity and duality in mathematical programming. Opsearch 30,
313–320 (1993)
3. Mond, B.: Second-order duality for nonlinear programs. Opsearch 11, 90–99 (1974)
4. Bector, C.R., Chandra, S.: First and second order duality for a class of nondifferentiable frac-
tional programming problems. J. Inf. Optim. Sci. 7, 335–348 (1986)
5. Jeyakumar, V.: First and second order fractional programming duality in nonlinear program-
ming. J. Math. Anal. Appl. 51, 607–620 (1975)
6. Yang, X.Q.: Second-order global optimality conditions for convex composite optimization.
Math. Program. 81, 327–347 (1998)
7. Suneja, S.K., Lalitha, C.S., Khurana, S.: Second order symmetric duality in multiobjective
programming. European J. Oper. Res. 144, 492–500 (2003)
8. Preda, V.: On efficiency and duality for multiobjective programs. J. Math. Anal. Appl. 166,
265–377 (1992)
9. Aghezzaf, B.: Second-order mixed type duality in multiobjective programming problems. J.
Math. Anal. Appl. 285, 97–106 (2003)
10. Zhang, J., Mond, B.: Second-order duality for multiobjective nonlinear programming general-
ized convexity. In: Glover, B.M., Craven, B.D., Ralph, D. (eds.) Proceedings of the Optimization
Miniconference III, pp. 79–95. University of Ballarat, Ballarat (1997)
11. Ahmad, I., Husain, Z.: Second-order (F, α, ρ, d)-convexity and duality in multiobjective pro-
gramming. Inform. Sci. 176, 3094–3103 (2006)
12. Hachimi, M., Aghezzaf, B.: Second-order duality in multiobjective programming involving
generalized type I functions. J. Math. Anal. Appl. 296, 382–392 (2004)
13. Gupta, S.K., Kailey, N., Sharma, M.K.: Multiobjective second-order nondifferentiable sym-
metric duality involving (F, α, ρ, d)-convex function. J. Appl. Math. Inform. 28, 1395–1408
(2010)
14. Egudo, R.R.: Multiobjective fractional duality. Bull. Austral. Math. Soc. 37, 367–378 (1988)
15. Jayswal, A., Kumar, D., Kumar, R.: Second order duality for nondifferentiable multiobjec-
tive programming problem involving (F, α, ρ, d)-V -type I functions. Optim. Lett. 4, 211–226
(2010)
16. Miettinen, K.M.: Nonlinear Multiobjective Optimization. Kluwer Academic, Bosten, M. A.
(1999)
17. Husain, I., Jabeen, Z.: On fractional programming containing support functions. J. Appl. Math.
Comput. 18, 361–376 (2005)
18. Mond, B., Weir, T.: Duality for fractional programming with generalized convexity conditions.
J. Inform. Optim. Sci. 3, 105–124 (1982)
19. Gulati, T.R., Geeta, M.: Duality in nondifferentiable multiobjective fractional programming
with generalized invexity. J. Math. Anal. Appl. 35, 103–118 (2011)
Consistency and Unconditional
Stability of a Positive Upwind
Scheme for One-Dimensional
Species Transport Equation
S. Prabhakaran and L. Jones Tarcius Doss
Abstract A positivity preserving upwind scheme for one-dimensional species trans-

port equation is discussed in this article. It is proved that the proposed numerical
scheme is unconditionally stable. Consistency of the scheme is also discussed in
detail. It is shown that the local truncation error is consistent with the advection-
diffusion-reaction equation when Δt → 0 and inconsistent when Δx → 0. Hence,
the numerical approximation converges to exact solution only when Δt → 0.
Keywords Positivity preserving · Stability · Consistency · Advection · Difussion ·

Reaction
1 Introduction
The problem related to one-dimensional unsteady state species transport is given as

follows: Find u(x, t) satisfying the governing equation
∂u ∂ 2u ∂u
R −D 2 + v = −ku 0 < x < ∞, t > 0, (1)
∂t ∂x ∂x
subject to the boundary conditions:
u(0, t) = u 0 t > 0 (2)

lim u(x, t) = 0 t > 0 (3)
x→∞
and initial condition:
S. Prabhakaran · L.J.T. Doss (B)

College of Engineering Guindy, Anna University, Chennai 600 025, India
e-mail: jones@annauniv.edu
S. Prabhakaran
e-mail: karan.shivam@gmail.com

DOI 10.1007/978-81-322-2485-3_59
744 S. Prabhakaran and L.J.T. Doss
u(x, 0) = 0, 0 < x < ∞, (4)
where, u is the property being transported; v the prescribed transport velocity; D the
diffusion coefficient; R the retardation factor; k the first order reaction coefficient.
The above equation is used to model the chemical species transportation with first-
order reaction with the assumption that the degradation reaction occurs in the liquid
phase. This problem is useful to find movement of solute particles with the ground-
water flow beneath the earth surface. These types of advection-diffusion-reaction
equations are also used to model air pollution, exponential traveling wave, bacterial
growth, tumor growth, colonization of Europe by oaks, adsorption of contaminants,
etc.
Analytical solution to this problem has been discussed by researchers Cho [2] and
Clement et al. [3], etc. Analytical solution to (1) with conditions (2), (3) and (4) is
given by (see [3]):
u0 vx mx
Rx − mt

u(x, t) = ex p ex p − er f c √
2 2D 2D 4D Rt
mx
Rx + mt
+ ex p er f c √ , (5)
2D 4D Rt
√
where m = v2 + 4k D.
Positivity preserving finite difference schemes are discussed by Chen-Charpentier
[1] and Karahan [5]. In this article, we have discussed stability and consistency of
positivity preserving scheme through upwind finite volume formulation. Illustrative
figures are given to show the variation in consistency.
2 Mathematical Description
The vector form of (1) is given by
∂u
R − ∇ · (D∇u) + v · ∇(u) = −ku. (6)
∂t
The numerical scheme is derived on the following control volume (CV) (Fig. 1)
Fig. 1 Control volume

Consistency and Unconditional Stability of a Positive Upwind Scheme … 745
In this article, the species transport is considered in the x direction only. Hence,
we assume that the other two dimensions y and z are infinitesimal. Further,
m + 1, m, m − 1 are nodal indices and m − 21 and m + 21 are face indices of control
volume. Let Δx > 0 be the spatial discretization length with nodal points x0 = 0
and xm = x0 + mΔx. Further, let xm + 1 = xm +2xm + 1 and xm− 1 = xm +2xm−1 . Inte-
2 2
grating the governing Eq. (6) over the local control volume CV in the time interval
(t, t + t), we obtain (Versteg and Malalasekara [7])
t + Δt t + Δt t + Δt
∂u
R d V dt = ∇ · (D∇u)d V dt − v · ∇ud V dt
t CV ∂t t CV t CV
t + t
− kud V dt.
t CV
Then, the volume integral is converted to a integral over a boundary surface by

applying Gauss divergence theorem and a suitable numerical approximation is used
at the boundary surface.
t + Δt t + Δt
R (u n + 1 − u n )d V = −
→
n · D∇ud Sdt − −
→
n · (uv)d Sdt
CV t S t S
t + Δt
− kud V dt
t CV
where −
→n is the unit normal to the surface S (Here, the surface is cross sectional
area A). The one-dimensional formulation of above is given by

t + Δt ∂u ∂u
R(u nm+ 1 − u nm ) dV = − DA DA dt
CV t m + 21 ∂ x m− 1∂x
2
t + Δt
− (Auv)m + 1 − (Auv)m− 1 dt
t 2 2
t + Δt
− um kd V dt
t CV
where u m is the average over the control volume. Let us assume that the area of cross
section A is uniform and velocity v is constant. Using the explicit weighed average
over the time interval, we have that
n
n
∂u ∂u
R(Umn + 1 − Umn )AΔx = D A − Δt
∂x m + 1 ∂ x m− 1
2
2
− v A u nm + 1 − u nm− 1 Δt − kUmn AΔxΔt.

2 2
where Umn be the approximation of u(x, t) at the nodal point (xm , tn ). Using the
central difference approximation for the derivative term, we obtain

DΔt n vΔt kUmn Δt
(Umn + 1 − Umn ) = U − 2U n
m + U n
m +1 − u n
− u n
− .
RΔx 2 m−1
1 1
RΔx m + 2 m− 2 R
The upwind approximation in a positive direction to advection term in above is given

by the following:
u nm + 1 = Umn u nm− 1 = Um−1

n
.
2 2
Therefore, we have
DΔt n vΔt n kUmn Δt

(Umn + 1 − Umn ) = 2
Um−1 − 2Umn + Umn + 1 − Um − Um−1
n
− .
RΔx RΔx R
Let us replace Umn with Umn + 1 on the right-hand side of the above equation to get
positivity preserving numerical scheme [1, 5].

2DΔt vΔt kΔt n+1 DΔt vΔt DΔt
1+ + + U m = + U n
m−1 + U n
m + Umn + 1 .
RΔx 2 RΔx R RΔx 2 RΔx RΔx 2
(7)
The above can be written as
Umn + 1 = aUm−1
n
+ bUmn + cUmn + 1 (8)
where
DΔt
RΔx 2
+ RΔx
vΔt
1
a= b=
1+ 2DΔt
RΔx 2
+ RΔx
vΔt
+ kΔt
R 1+ 2DΔt
RΔx 2
+ vΔt
RΔx + kΔt
R
DΔt
RΔx 2
c= .
1+ 2DΔt
RΔx 2
+ RΔx
vΔt
+ kΔt
R
Here, all the coefficients D, R, v and k are positive. Also, the mesh lengths Δt and
Δx are positive. Therefore a, b, c are all positive and hence positivity is preserved
in this upwind scheme.
3 Stability
The general form of an explicit finite difference numerical scheme is given as
Umn + 1 = aUm−1
n
+ bUmn + cUmn + 1 . (9)
Let Umn = Bξ n eimθ . Then using von Neumann stability analysis in (9), we obtain
ξ = ae−iθ + b + ceiθ = b + (a + c)(cos θ ) + i(c − a) sin θ.
The amplification factor ξ should meet the condition |ξ | ≤ 1 in order to get a stable
scheme which is equivalently |ξ |2 ≤ 1 (Smith [6]). Therefore we get
b2 + (a + c)2 cos2 θ + 2b(a + c) cos θ + (c − a)2 sin2 θ ≤ 1

⇔ (a 2 + c2 )(cos2 θ + sin2 θ ) + b2 + 2ac(cos2 θ − sin2 θ ) + 2b(a + c) cos θ ≤ 1
⇔ a 2 + b2 + c2 + 2ac(cos2 θ − sin2 θ ) + 2b(a + c) cos θ ≤ 1
⇔ (a + b + c)2 − 2b(a + c)(1 − cos θ ) − 2ac(1 − cos 2θ ) ≤ 1
θ
⇔ (a + b + c)2 ≤ 1 + 4b(a + c) sin2 + 4ac sin2 θ
2
θ θ θ
⇔ (a + b + c)2 ≤ 1 + 4b(a + c) sin2 + 16ac sin2 cos2
2 2 2
4 θ 2 θ 2 θ
⇔ (a + b + c) + 16ac sin
2
≤ 1 + 4b(a + c) sin + 16ac sin .
2 2 2
Maximizing the trigonometric functions in the above inequality with respect to their
argument θ , we obtain
(a + b + c)2 ≤ 1 + 4b(a + c). (10)
We now establish the stability condition for the generalized explicit scheme (9) as
follows: If the coefficients a, b, and c of (9) satisfy the following conditions:
(i) a ≥ 0, b ≥ 0, and c ≥ 0
(ii) (a + b + c)2 ≤ 1 + 4b(a + c)
then the scheme (9) is stable. This is evident from previous derivation. The scheme
(8) is stable only when it satisfies the above-mentioned stability criteria. Clearly, the
coefficients a, b, and c from (8) are positive. Further, substituting these coefficients
in (10), we have that
2
1+ 2DΔt
RΔx 2
+ RΔx
vΔt 2DΔt
RΔx 2
+ vΔt
RΔx
≤ 1 + 4 2
1+ 2DΔt
RΔx 2
+ RΔx
vΔt
+ kΔt
R 1+ 2DΔt
+ vΔt
+ kΔt
RΔx 2 RΔx R

2DΔt vΔt 2 2DΔt vΔt kΔt 2 2DΔt vΔt
1+ + ≤ 1 + + + + 4 +
RΔx 2 RΔx RΔx 2 RΔx R RΔx 2 RΔx
Let α = 2DΔt
RΔx 2
, β= vΔt
RΔx and γ = kΔt
R , then we get
(1 + α + β)2 ≤ (1 + α + β + γ )2 + 4(α + β)
4α + 4β + 2γ + 2αγ + 2βγ ≥ 0.
The constants α, β are γ all positive, because the coefficients D, R, v, and k and also
the mesh lengths Δt and Δx are positive. Therefore, the above inequality holds for
any choice of parameters and mesh lengths. Hence, the scheme (8) is unconditionally
stable.
4 Truncation Error and Consistency
The truncation error Tm,n for the explicit scheme at interior nodal point (xm , tn ) is
defined by Smith [6]
1
Tm,n = u(xm , tn + 1 ) − Umn + 1
Δt
where u(xm , tn + 1 ) and Umn + 1 are the values of exact and numerical solution at
(xm , tn + 1 ) respectively. Let Umn + 1 = aUm−1
n + bUmn + cUmn + 1 , we have that
Δt Tm,n = u(xm , tn + 1 ) − aUm−1

n
− bUmn − cUmn + 1 .
Following the usual procedure of obtaining the truncation error, we replace numerical
solution by exact solution
Δt Tm,n = u(xm , tn + 1 ) − au(xm−1 , tn ) − bu(xm , tn ) − cu(xm + 1 , tn )

= u(xm , tn + Δt) − au(xm − Δx, tn ) − bu(xm , tn ) − cu(xm + Δx, tn ).
Expanding using Taylor series, we have that

∂u Δt 2 ∂ 2 u ∂u Δx 2 ∂ 2 u
Δt Tm,n = u + Δt + + · · · − a u − Δx + + · · ·
∂t 2 ∂t 2 ∂x 2 ∂x2

∂u Δx ∂ u
2 2
− bu − c u + Δx + + ···
∂x 2 ∂x2 (xm ,tn )

∂u Δt 2 ∂ 2 u ∂u
= Δt + − (a + b + c − 1)u + (a − c)Δx
∂t 2 ∂t 2 ∂x

Δx 2 ∂ 2 u
− (a + c) + ···
2 ∂x2 (xm ,tn )

∂u Δt ∂ 2 u 1 ku v ∂u
= Δt + + − +
∂t 2 ∂t 2 1+ 2DΔt
+ vΔt
+ kΔt R R ∂x
RΔx 2 RΔx R

D ∂ 2u vΔx ∂ 2u
− + + ··· using a, b and c.
R ∂x2 2R ∂ x 2 (xm ,tn )

∂u Δt ∂ 2 u 1 ku v ∂u
Tm,n = + + − +
∂t 2 ∂t 2 1+ 2DΔt
+ vΔt
+ kΔt R R ∂x
RΔx 2 RΔx R

D ∂ 2u vΔx ∂ 2u
− + + ··· (11)
R ∂x 2 2R ∂ x 2 (xm ,tn )
It is clear that the local truncation error is not consistent with the partial differential
equation (1). Suppose that Δt → 0 then the truncation error is consistent with (1)
and the order of truncation error is Δx. Also note that the error is inconsistent when
Δx → 0.
An artificial boundary is fixed at the farther end and allowed to approach infinity
to handle boundary condition (3) in the numerical computation. It is easy to fix the
boundary since the initial condition (4) is zero initial condition (Table 1).
The conclusion from Sect. 3 is that the proposed scheme is unconditionally stable.
Also from Sect. 4, it is shown that the scheme is consistent only when Δt → 0
and inconsistent when Δx → 0. This claim is also supported by the numerical
computation which is shown in Figs. 2 and 3. From Fig. 2 and Table 2, it is evident
that the numerical approximation is consistent with exact solution when Δt → 0.
Hence, the positivity preserving unconditionally stable upwind scheme converges to
exact solution only when Δt → 0. Also, it can be concluded from Fig. 3 that the
numerical scheme is inconsistent when Δx → 0.
Table 1 Parameters used for Parameter Value

computation
u0 1 mg/l
k 0.01 h−1
R 2
v 1 cm h−1
T 50 h
D 0.18
Fig. 2 Convergence of numerical solution when Δt → 0
Table 2 L 1 error when Δt L 1 error

Δx = 0.2 and Δt varies
1 95
0.1 41.9
0.01 5.96
0.001 2.22
Fig. 3 Divergence of numerical solution when Δx → 0
Acknowledgments The authors would like to thank DST-SERB for providing financial support
through the project SR/S4/MS:775/12 to carry out this research work.
References
1. Chen-Charpentier, B.M., Kojouharov, H.V.: An unconditionally positivity preserving scheme

for advection diffusion reaction equations. Math. Comput. Model. 57(910), 21772185 (2013)
2. Cho, C.M.: Convective transport of ammonium with nitrification in soil. Can. J. Soil Sci. 51(3),
339350 (1971)
3. Clement, T.P., Quezada, C.R., Lee, K.-K.: Generalised solution to multi-dimensional multi
species transport equations coupled with a first-order reaction network involving distinct retar-
dation factors. Adv. Water Resour. 27, 507–520 (2004)
4. Eymard, R., Gallout, T., Herbin, R.: Handbook of Numerical Analysis. Addison-Wesley, New
York (2000)
5. Karahan, H.: Unconditional stable explicit finite difference technique for the advection diffusion
equation using spreadsheets. Adv. Eng. Soft. 38(2), 80–86 (2007)
6. Smith, G.D.: Numerical Solutions of Partial Differential Equations: Finite Difference Methods.
Oxford University Press, Oxford (1985)
7. Versteg, H.K., Malalasekara, W.: An Introduction to Computational Fluid Dynamics: A Finite
Volume Method. Addison-Wesley, New York (1996)
Numerical Solution for a Coupled System
of Singularly Perturbed Initial Value
Problems with Discontinuous Source Term
S. Chandra Sekhara Rao and Sheetal Chawla
Abstract In this work, we study a numerical method for a coupled system of

singularly perturbed initial value problems having discontinuous source term. The
leading term of each equation is multiplied by a distinct small positive parameter,
due to which the overlapping initial and interior layers are generated in the solution.
The problem is discretized using backward Euler difference scheme which involves
an appropriate piecewise-uniform variant of Shishkin mesh that is fitted to both the
initial and interior layers. The method is proved to be uniformly almost first-order
accurate with respect to all the parameters. Numerical results are presented in support
of the theory.
Keywords Singular perturbation · Initial layer · Interior layer · Coupled system ·

Discontinuous source term · Uniformly convergent · Shishkin mesh
1 Introduction
Consider a coupled system of singularly perturbed initial value problem with discon-
tinuous source term on the unit interval Ω = (0, 1), and assume a single discontinuity
in the source term at a point d ∈ Ω. Let Ω1 = (0, d) and Ω2 = (d, 1) and the jump
at d in any function is given as [ω](d) = ω(d+) − ω(d−). The corresponding initial
value problem is to find u 1 , . . . , u m ∈ C 0 (Ω) ∩ C 1 (Ω1 ∪ Ω2 ), such that
Lu := Eu + Au = f in Ω1 ∪ Ω2 , (1)
S. Chandra Sekhara Rao (B) · S. Chawla

Department of Mathematics, Indian Institute of Technology Delhi,
Hauz Khas 110016, New Delhi, India
e-mail: scsr@maths.iitd.ac.in
S. Chawla
e-mail: chawlaasheetal@gmail.com

DOI 10.1007/978-81-322-2485-3_60
754 S. Chandra Sekhara Rao and S. Chawla
u(0) = p, (2)
where E = diag(ε1 , . . . , εm ) with small parameters ε1 , . . . , εm are such that 0 <

ε1 ≤ · · · ≤ εm ≤ 1,
A(x) = (ai j (x))m×m and f (x) = ( f j (x))m×1 (3)
are given. We assume that the coupling matrix satisfies the following positivity con-
ditions:

m
aii (x) > |ai j (x)|, for 1 ≤ i ≤ m, and ai j (x) ≤ 0 for i = j, (4)
j=i, j=1
and for some constant α, we have

m
0<α< min (ai j (x)) (5)
x∈Ω,1≤i≤m
j=1
The source terms f 1 (x), . . . , f m (x) are sufficiently smooth on Ω \ {d}. The solution
components u 1 , . . . , u m of the problem (1) and (2) have overlapping initial layers at
x = 0 and have overlapping interior layers to the right side of point of discontinuity
at x = d.
Shishkin [8] laid down the framework for singularly perturbed reaction-diffusion
problems with discontinuous coefficients. Dunne and Riordan [2] discussed the sin-
gularly perturbed initial value problem with discontinuous coefficients in scalar case.
Parameter-robust numerical methods for systems of singularly perturbed differen-
tial equations were analyzed in [7]. A parameter-uniform numerical method was
constructed in [3] for a system of singularly perturbed initial value problem where
all of the singular perturbation parameters are equal. In this case all the solution
components have an initial layer of same width, due to which the analysis is sim-
pler. The case where the small parameter is associated with only one equation was
considered in [4]. The most difficult and general case is that each component of
the solution has its own initial layer that overlaps and interacts with others and this
was considered in [1, 5, 9] . In [9], first order (up to logarithmic factor) uniformly
convergent numerical method was developed. A hybrid finite difference scheme on a
piecewise-uniform Shishkin mesh was considered in [1] and the scheme was almost
second-order accurate, uniformly in both small parameters. In all these works the
source term is smooth. In the present work, the source term is discontinuous, due to
which each solution component has an initial layer as well as interior layer.
This paper is organized as follows. Section 2, presents the properties of the exact
solution. The mesh and the scheme that approximate singularly perturbed initial
value problem with discontinuous source term are constructed in Sect. 3. In Sect. 4,
Numerical Solution for a Coupled System of Singularly Perturbed … 755
it is proved that the numerical approximation computed by finite difference method

is almost first-order accurate in the maximum norm, uniformly with respect to the
parameters ε1 , . . . , εm . In Sect. 5, the results of numerical experiments are presented
for validation of the theoretical results.
Notations We use C to denote a generic positive constant and C = (C, C, . . . , C)T to
denote a generic positive constant vector which are independent of the perturbation
parameters and the discretization parameter N, but may not be the same at each
occurrence. Define v ≤ w if vi ≤ wi , for 1 ≤ i ≤ m. We consider the maximum
norm and denote it by . S , where S is a closed and bounded subset of Ω. We
define v S = max |v(x)| and v S = max{ v1 S , v2 S , . . . , vm S }.
x∈S
2 Properties of the Exact Solution
Theorem 1 The problem and (2) has a solution u = (u 1 , . . . , u m )T with

1 (1)
u 1 , . . . , u m ∈ C(Ω) C (Ω1 Ω2 ).
Proof The result can be proved by following the similar technique considered in [6].
Theorem 2 Suppose that a function u ∈ C 1 (Ω1 ∪ Ω2 )m satisfies u(0) ≥ 0, Lu(x) ≥

0 for all x ∈ Ω1 ∪ Ω2 . Then u(x) ≥ 0, x ∈ Ω.
Proof Let u i ( pi ) = min{u i (x)}, for 1 ≤ i ≤ m. Assume without loss of generality

x∈Ω
u 1 ( p1 ) ≤ u i ( pi ), for 2 ≤ i ≤ m. If u 1 ( p1 ) ≥ 0, then there is nothing to prove.
Suppose that u 1 ( p1 ) < 0, then the proof is completed by showing that this leads to
contradiction. Note that p1 = {0}, so either p1 ∈ Ω1 ∪ Ω2 or p1 = d.
In the first case

m
(Lu)1 ( p1 ) = ε1 u 1 ( p1 ) + a1 j ( p1 )u j ( p1 )
j=1

m
m
m
= ε1 u 1 ( p1 ) + a1 j ( p1 )u j ( p1 ) + a1 j ( p1 )u 1 ( p1 ) − a1 j ( p1 )u 1 ( p1 ) < 0.
j=1 j=2 j=2
In the second case, since u ∈ C(Ω)m and u 1 (d) < 0, there exists a neighborhood
Nh = (d − h, d) such that u 1 (x) < 0 for all x ∈ Nh . Now choose a point x1 =
d, x1 ∈ Nh such that u 1 (x1 ) > u 1 (d). It follows from the mean value theorem that,
for some x2 ∈ Nh , u 1 (x2 ) = u 1 (d)−u
d−x1
1 (x 1 )
< 0, since x2 ∈ Nh .
m
Thus (Lu)1 (x2 ) = ε1 u 1 (x2 ) + a1 j (x2 )u j (x2 ) < 0.
j=1
Lemma 1 Let u be the solution of (1) and (2). Then,

1
uΩ ≤ max u(0), LuΩ1 ∪Ω2 .
α
An immediate consequence of this result is that the solution u is unique.

To derive sharper bounds on the derivatives of the solution, the solution is decom-
posed into a sum, composed of a regular component v and a singular component
w. That is, u = v + w. The regular component v, is defined as the solution of the
following problem:
Lv(x) = f (x), x ∈ Ω1 ∪ Ω2 , v(0) = A−1 (0)f (0), v(d+) = A−1 (d+)f (d+). (6)
The singular component w, is defined as the solution of the following problem:
Lw(x) = 0, x ∈ Ω1 ∪ Ω2 , w(0) = u(0) − v(0), [w](d) = −[v](d). (7)
Theorem 3 Let A(x) satisfy (4) and (5). Then the components vi , 1 ≤ i ≤ m of the
regular component v and its derivatives satisfy the bounds for all x ∈ Ω1 ∪ Ω2 , and
k = 0, 1, 2,
v(k) Ω1 ∪Ω2 ≤ C for k = 0, 1 with

vi Ω1 ∪Ω2 ≤ C εi−1 .
Consider the following layer functions:
Bεli (x) = e−αx/εi ,

Bεri (x) = e−α(x−d)/εi .
Lemma 2 Let A(x) satisfy (4) and (5). Then the components wi , 1 ≤ i ≤ m of the
singular component w and its derivatives satisfy the bounds for all x ∈ Ω1 ∪ Ω2 ,
⎧ m
⎪ Bεlq (x)
⎪
⎪ C , x ∈ Ω1
⎪
⎨ εq
CBεlm (x), x ∈ Ω1
|wi (x)| ≤
q=i
|wi (x)| ≤
CBεrm (x), x ∈ Ω2 , ⎪
⎪
m B (x)
εrq
⎪
⎪ C , x ∈ Ω2 ,
⎩ εq
q=i
⎧ m
⎪ Bεlq (x)
⎪
⎪ C , x ∈ Ω1
⎪
⎨ εq
|εi wi (x)| ≤
q=1
⎪
⎪
m B (x)
εrq
⎪
⎪ C , x ∈ Ω2 .
⎩ εq
q=1
Proof We have u = v+w and by Lemma 1 |w(0)| ≤ C and |w(d+)| ≤ C. Define the
barrier function Υ := CBεlm (x)e, with C chosen sufficiently large such that Υ ≥ |w|
at x = 0, d+,
⎛ ⎞
m
ε
m
ε
LΥ = CBεlm ⎝ α ⎠ ≥ 0 = |Lw|Ω1 .
1 m
a1 j − α, . . . , am j −
εm εm
j=1 j=1
Using continuous maximum principle, we get the required bound on w. Now to

m
bound first-order derivative of wi , consider εi wi + ai j w j = 0, together with the
j=1
bound on w. This implies that

Cεi−1 Bεlm (x), x ∈ Ω1 ,
|wi | ≤
Cεi−1 Bεrm (x), x ∈ Ω2 .
Now to find the sharper bound consider the system of m − 1 equations
Êŵ + Âŵ = g,
where Ê, Â is the matrix obtained by deleting the last row and column from E, A,
respectively, and the components of g are gi = −aim wm , for 1 ≤ i ≤ m − 1.
Using the bounds derived earlier and the decomposition of ŵ = q + r, into regular
and singular component we get the required result. Now to bound second-order

m
derivatives, differentiate εi wi + ai j w j = 0 once and using the estimates of wi ,
j=1
we get the required bounds on singular component w and its derivatives.
Lemma 3 For all i, j such that 1 ≤ i ≤ j ≤ m, there exists a unique point

xi, j ∈ (0, d) such that εi−1 Bεli (xi, j ) = ε−1 −1
j Bεl j (x i, j ). Also, εi Bεri (d + x i, j ) =
ε−1 −1 −1
j Bεr j (d + x i, j ). On [0, x i, j ) we have εi Bεli (x) > ε j Bεl j (x) and on (x i, j , d)
we have εi−1 Bεli (x) < ε−1 −1
j Bεl j (x). Similarly on (d, d + x i, j ) we have εi Bεri (x) >
ε−1 −1 −1
j Bεr j (x) and on (d + x i, j , 1] we have εi Bεri (x) < ε j Bεr j (x).
For the analysis of the convergence, a more precise decomposition of the components
of the singular component w is required.
Theorem 4 The singular component w can be decomposed in this way as follows,

for 1 ≤ i ≤ m :
m
wi (x) = wi,εq (x)
q=1
where
⎧ B (x) ⎧ B (x)
⎨ C εlq , x ∈ Ω ⎨ C εlq , x ∈ Ω
εq 1 εq 1
|wi,εq (x)| ≤ |εi wi,εq (x)| ≤
⎩ C Bεrq (x) , x ∈ Ω , ⎩ C Bεrq (x) , x ∈ Ω .
εq 2 εq 2
Proof Define a function wi,ε1 as follows:

m
wi,ε1 (x) = wi (x) − wi,εq (x)
q=2
and for 1 < q ≤ m, we have

⎧
⎪ 2 [(x−xq−1,q ) ] w(k) (x
k
⎪
⎪ k=0 q−1,q ), x ∈ [0, xq−1,q ),
⎪
⎪
k! i
⎪
⎪ m
⎪
⎪ w (x) − wi,εr (x), x ∈ [xq−1,q , d),
⎪
⎨ i
r =q+1
wi,εq =
⎪
⎪ 2 [(x−d−xq−1,q ) ] w(k) (d
k
+ xq−1,q ), x ∈ (d, d + xq−1,q ),
⎪
⎪ k=0 k! 1
⎪
⎪ m
⎪
⎪
⎪
⎪ w (x) − wi,εr (x), x ∈ [d + xq−1,q , 1].
⎩ i
r =q+1
Now we establish the bounds on the second derivative.

For x ∈ [xm−1,m , d) ∪ [d + xm−1,m , 1].

m B (x)
εlq Bεl (x)
(x)| = |ε w (x)| ≤ C
|εi wi,ε ≤C m .
i i
m εq εm
q=1
For x ∈ [0, xm−1,m ) ∪ (d, d + xm−1,m ).

m B (x
εlq m−1,m ) Bεlm (xm−1,m )
(x)| = |ε w (x
|εi wi,ε i i m−1,m )| ≤ C ≤ C ≤
m εq εm
q=1
Bεlm (x)
C .
εm
Now, for each m − 1 ≤ q ≤ 2, it follows that
(x) = 0.
For x ∈ [xq−1,q , d) ∪ [d + xq−1,q , 1], wi,ε q
For x ∈ [0, xq−1,q ) ∪ (d, d + xq−1,q ].
m B (x
εlq q−1,q ) Bεlq (xq−1,q ) Bεlq (x)
(x)| = |ε w (x
|εi wi,ε )| ≤ C ≤C ≤C .
i i q−1,q
q εq εq εq
q=1
(x) = 0.
For x ∈ [x1,2 , d) ∪ [d + x1,2 , 1], wi,ε 1
For x ∈ [0, x1,2 ) ∪ (d, d + x1,2 ).
m
m B (x)
εlq Bεl (x)
(x)| = |ε w (x) −
|εi wi,ε i i ε w
i i,εq (x)| ≤ C ≤C 1 .
1 εq ε1
q=2 q=1
For the bounds on the first derivatives we have the relation

xq,q+1 xq,q+1
Bεlq (t) B (x)

|wi,ε (x)| = w
(t)dt ≤ C dt ≤ C εlq .
q i,εq εq εq
x x
3 Discretization of the Problem
In this section, we discretize the system of initial value problem (1) and (2) using
a fitted mesh method composed of a backward difference scheme on a piecewise
N
uniform variant of Shishkin mesh with points Ω = {xi : i = 0, . . . , N }. Let N =
2 , k ≥ 6 be a positive integer. Define the transition parameter
k

d εm (1 − d) εm
σεlm := min , ln N , σεrm := min , ln N ,
2 α 2 α
σ σ
εlk+1 εk εrk+1 εk
σεlk := min , ln N , σεrk := min , ln N ,
2 α 2 α
k = m − 1, . . . , 1.
The interior points of the mesh are denoted by

N N
Ω N
= xi : 1 ≤ i ≤ − 1 ∪ xi : + 1 ≤ i ≤ N − 1 = Ω1N ∪ Ω2N .
2 2
Let h i = xi − xi−1 be the ith mesh step and i = h i +h2 i+1 , clearly x N = d.
2
We divide the interval [0, d] into m + 1 subintervals [0, σεl1 ], [σεl1 , σεl2 ], . . . ,
[σεlm−1 , σεlm ], [σεlm , d]. On the subinterval [0, σεl1 ] a uniform mesh of N /2m+1
mesh intervals, on each subinterval [σεlk , σεlk+1 ], 1 ≤ k ≤ m − 1, a uniform
mesh of N /2m−k+2 mesh intervals, and on [σεlm , d] a uniform mesh of N/4
mesh intervals are placed. Similarly, we divide the interval [d, 1] into subintervals
[d, d + σεr1 ], [d + σεr1 , d + σεr2 ], . . . , [d + σεrm−1 , d + σεrm ], [d + σεrm , 1]. On the
subinterval [d, d + σεr1 ] a uniform mesh of N /2m+1 mesh intervals, on each subin-
terval [d + σεrk , d + σεrk+1 ], 1 ≤ k ≤ m − 1, a uniform mesh of N /2m−k+2 mesh
intervals, and on [d + σεrm , 1] a uniform mesh of N /4 mesh intervals are placed. Let
h εl1 and h εr1 be the mesh lengths on [0, σεl1 ] and on [d, d +σεr1 ] respectively. Let H1
and H2 be the mesh lengths on [σεlm , d] and on [d + σεrm , 1] respectively; h εlk and
h εrk be the mesh lengths on [σεlk , σεlk+1 ] and on [d + σεrk , d + σεrk+1 ], k = 2, . . . , m
respectively.
Define the discrete finite difference operator L N as follows:
LN U = f , for all xi ∈ Ω N , U(x0 ) = p, (8)

where
L N := ED− + A,
and at x N /2 = d the scheme is given by

L N U x N = E D− U x N + A x N U x N = f x N −1 ,
2 2 2 2 2
where
Z (xi ) − Z (xi−1 )
D − Z (xi ) = , i = 1, . . . , N .
hi
Lemma 4 Suppose that a mesh function W satisfies W0 ≥ 0 and L N W ≥ 0, for all

N
xi ∈ Ω N , implies that W ≥ 0 for all xi ∈ Ω .
Lemma 5 If U be the numerical solution of (1) and (2), then,

1
UΩ N ≤ max U(0), fΩ N ∪Ω N .
α 1 2
4 Error Analysis
By a Taylor expansion on regular and singular components, we have

d (xi − xi−1 )
|εk − D − vk (xi )| ≤ Cεk |vk |2 ≤ C N −1 , (9)
dx 2
and

d −
Cεk (xi −x2 i−1 ) |wk |2 (10)
|εk −D wk (xi )| ≤ Cε max |wk |, (11)
dx k
[xi−1 ,xi ]
j
where k = 1, . . . , m, i = N2 , | z k | j := max | dd x zj |, ∀ j ∈ N. Now to evaluate
the error estimates for the singular components on different subintervals considered
as follows:
Case (i) For xi ∈ [σεlm , d) ∪ [d + σεrm , 1].
Using (11) and bounds on singular components, we have, for j = 1, . . . , m.

m B (x)
εlq
| ((L N − L)w) j (xi ) | ≤ Cε j
εq
q= j
≤ C Bεlm [xi−1 ,xi ] = Bεlm (xi−1 ) ≤ C N −1 .

Similar arguments prove a similar result for the subinterval [d + σεrm , 1]. Hence,
for xi ∈ [σεlm , d) ∪ [d + σεrm , 1] we have,
| ((L N − L)w) j (xi ) | ≤ C N −1 .
Case (ii) For xi ∈ (0, σεl1 ] ∪ (d, d + σεr1 ].

Using (10) and the bounds on the singular components yields

m B (x)
εlq
| ((L N − L)w) j (xi ) |≤ C(xi − xi−1 )ε j wj ≤ h i ≤ C(N −1 ln N ).
εq
q=1
Case (iii) For xi ∈ (σεlk , σεlk+1 ) ∪ (d + σεrk , d + σεrk+1 ), where 1 ≤ k ≤ m − 1.

Using the decomposition in Theorem 4 of singular components and bounds on
singular components gives

m−1
d d
| ((L N − L)w) j (xi ) |= | εj − D − w j,εq (xi ) + ε j − D − w j,εm (xi )|.
dx dx
q=1
(12)
Consider the first part of (12) and using the bounds on singular components, we
obtain

m−1
m−1
d
| εj − D − w j,εq (xi )| ≤ ε j wj,εq [xi−1 ,xi ] ≤ C Bεlm−1 (xi−1 ) ≤ C N −1 .
dx
q=1 q=1
Using the bounds on singular components for the second part of (12), we have

d hi
|ε j − D − w j,εm (xi )| ≤ ε j wj,εm ≤ C(N −1 ln N ).
dx 2
Case (iv) For xi ∈ {σεlk , d + σεrk }, where 1 ≤ k ≤ m − 1.

Using the decomposition of the singular components and bounds on singular
components defined in Theorem 4 gives

m−1
d d
| ((L N − L)w) j (xi ) |≤ | εj − D − w j,εq (xi ) + ε j − D − w j,εm (xi )|.
dx dx
q=1
(13)
Consider the first part of (13) for the case j ≤ k, and using the definition of point
xi, j we have

m−1
m−1
d
| εj − D − w j,εq (xi )| ≤ ε j wj,εq [xi−1 ,xi ] ≤ C N −1
dx
q=1 q=1
and if, j > k, using the bounds on singular component and the analysis in Case (i),
we have

m−1
m−1
d −
| εj −D w j,εq (xi )| ≤ ε j wj,εq [xi−1 ,xi ] ≤ C N −1 .
dx
q=1 q=1
For the second part of (13), use bounds on singular components defined in Theorem 4,
to obtain

d
|ε j − D − w j,εm (xi )| ≤ Cε j h εk w j,εm ≤ C N −1 ln N .
dx
Now at the point x N /2 = d,

m
(L N (U − u))1 (d) = f 1 (d − H1 ) − ε1 D − u 1 (d) − a1 j (d)u j (d),
j=1
ε1 m
= f 1 (d − H1 ) − (u 1 (d) − u 1 (d − H1 )) − a1 j (d)u j (d),
H1
j=1

m
ε1 d
= f 1 (d − H1 ) − u 1 (s) ds − a1 j (d)u j (d).
H1 t=d−H1 j=1
Using the bounds on derivatives of u, we obtain
|(L N (U − u))1 (d))| ≤ C(N −1 ln N),
and similarly, we can prove for 2 ≤ j ≤ m
|(L N (U − u)) j (d))| ≤ C(N −1 ln N).
We conclude this section with the following main result which follows by using the
error analysis for the regular and singular components, and the discrete maximum
principle.
Theorem 5 Let u be the solution of given problem (1) and (2) and U be the solution
of discrete problem, then
U − uΩ N ≤ C(N −1 ln N ).
5 Numerical Results
Example 1 Consider the following singularly perturbed initial value problem with
discontinuous source term (Table 1):
ε1 u 1 (x) + 4u 1 (x) − u 2 (x) − u 3 (x) = f 1 (x), x ∈ Ω1 ∪ Ω2

ε2 u 2 (x) − u 1 (x) + (4 + x)u 2 (x) − u 3 (x) = f 2 (x), x ∈ Ω1 ∪ Ω2
ε3 u 3 (x) − 2u 1 (x) − u 2 (x) + (5 + x)u 3 (x) = f 3 (x), x ∈ Ω1 ∪ Ω2
u 1 (0) = 0, u 2 (0) = 0, u 3 (0) = 0,
where

x for 0 ≤ x ≤ 0.5 1 for 0 ≤ x ≤ 0.5
f 1 (x) = f 2 (x) = and
2 for 0.5 < x ≤ 1, 4 for 0.5 < x ≤ 1,

1 + x for 0 ≤ x ≤ 0.5
2
f 3 (x) =
3 for 0.5 < x ≤ 1.
The exact solution of the test example is not known. Therefore, we estimate the error
for U by comparing it to the numerical solution U obtained on the mesh x j that
contains the mesh points of the original mesh and their midpoints, that is, x2 j =
x j , j = 0, . . . , N ,
x2 j+1 = (x j + x j+1 )/2, j = 0, . . . , N − 1.
For different values of N and ε1 , ε2 , ε3 , we compute
i ) N .
DεN1 ,ε2 ,ε3 := (U − U)(x Ω
Table 1 Maximum point-wise errors D N , DεN1 ,ε2 ,ε3 with ε2 = 10−5 and ε3 = 10−3 for Example 1
ε1 = 10− j N = 64 N = 128 N = 256 N = 512 N = 1024 N = 2048

j =5 4.94E-02 3.34E-02 2.14E-02 1.29E-02 7.40E-03 4.13e-03
6 4.93E-02 3.34E-02 2.14E-02 1.29E-02 7.40E-03 4.13e-03
7 6.41E-02 4.89E-02 3.30E-02 2.12E-02 1.27E-02 7.29e-03
8 6.93E-02 5.37E-02 3.68E-02 2.40E-02 1.46E-02 8.49e-03
9 6.98E-02 5.42E-02 3.72E-02 2.43E-02 1.48E-02 8.62e-03
10 6.99E-02 5.43E-02 3.72E-02 2.43E-02 1.48E-02 8.64e-03
.. .. .. .. .. .. ..
. . . . . . .
15 6.99E-02 5.43E-02 3.72E-02 2.43E-03 1.48E-02 8.64e-03

DN 6.99E-02 5.43E-02 3.72E-02 2.43E-03 1.48E-02 8.64e-03
pN 4.69E-01 6.74E-01 7.39E-01 8.42E-01 9.01E-01
Then the parameter-uniform error is computed as
D N := max {DεN1 ,ε2 ,ε3 },

Sε1 ,ε2 ,ε3
where the singular perturbation parameters take values in the set

Sε1 ,ε2 ,ε3 = {(ε1 , ε2 , ε3 )|, ε1 = 10− j , 0 ≤ j ≤ 15, ε2 = 10−l , 0 ≤ l ≤ j, ε3 =
10−k , 0 ≤ k ≤ l} and the order of convergence is calculated using the formula
ln(D N ) − ln(D 2N )
pN = .
ln(2 ln N ) − ln(ln(2N ))
References
1. Cen, Z., Xu, A., Le, A.: A second-order hybrid finite difference scheme for a system of singularly
perturbed initial value problems. J. Comput. Appl. Math. 234, 3445–3457 (2010)
2. Dunne, R.K., Riordan, E.O’.: Interior layers arising in linear singularly perturbed differential
equations with discontinuous coefficients. In: Proceedings of the Fourth International Confer-
ence on Finite Difference Methods: Theory and Applications. Lozenetz, Bulgaria, 26–29 August
(2006). Rousse University, Bulgaria, 29–38 (2007)
3. Hemavathi, S., Bhuvaneswari, T., Valarmathi, S., Miller, J.J.H.: A parameter uniform numerical
method for a system of singularly perturbed ordinary differential equations. Appl. Math. Comput.
191, 1–11 (2007)
4. Meenakshi, P.M., Valarmathi, S., Miller, J.J.H.: Solving a partially singularly perturbed initial
value problem on shishkin meshes. Appl. Math. Comput. 215, 3170–3180 (2010)
5. Rao, S.C.S., Kumar, S.: Second order global uniformly convergent numerical method for a
coupled system of singularly perturbed initial value problems. Appl. Math. Comput. 219, 3740–
3753 (2012)
6. Rao, S.C.S., Chawla, S.: Interior layers in coupled system of two singularly perturbed reaction-
diffusion equations with discontinuous source term. NAA 2012, LNCS 8236, 445–453 (2012)
7. Roos, H.G., Stynes, M., Tobiska, L.: Robust numerical methods for singularly perturbed dif-
ferential equations. Springer Series in Computational Mathematics, 2nd edn. Springer, Berlin
(2008)
8. Shishkin, G.I.: Discrete approximation of singularly perturbed elliptic and parabolic equations.
Russian Academy of Sciences, Ural Section, Ekaterinburg (in Russian)
9. Valarmathi, S., Miller, J.J.H.: A parameter-uniform finite difference method for singularly per-
turbed linear dynamical systems. Int. J. Numer. Anal. Model. 7, 535–548 (2010)

Mathematical Analysis and Its Applications 2015

Uploaded by

Copyright:

Available Formats

Mathematical Analysis and Its Applications 2015

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mathematical Analysis and Its Applications 2015

Uploaded by

Copyright:

Available Formats

Springer Proceedings in Mathematics & Statistics

This book series features volumes composed of selected contributions from

More information about this series at http://www.springer.com/series/10533

Uaday Singh H.M. Srivastava

R.N. Mohapatra H.M. Srivastava

ISSN 2194-1009 ISSN 2194-1017 (electronic)

Library of Congress Control Number: 2015942480

Springer New Delhi Heidelberg New York Dordrecht London

Printed on acid-free paper

Springer (India) Pvt. Ltd. is part of Springer Science+Business Media

Welcome to the Proceedings of the International Conference on Recent Trends in

February 2015 P.N. Agrawal

Prof. R.P. Agarwal, Texas A&M University, USA

Local Organizing Committee

Patron: Prof. Pradipta Banerji, Director IIT Roorkee

Academic Program Committee Catering Committee

Part I Mathematical Analysis

Convergence Analysis of Legendre Spectral Galerkin Method

Analysis of an Eco-Epidemiological Model with Migrating

Stability Analysis of an Integro Differential Equation Model

Approximation of Solutions of a Stochastic Differential Equation. . . . . 51

Reconstruction of Multiply Generated Splines from Local

Approximation of Periodic Functions Belonging

Modified Baskakov-Szász Operators Based on q-Integers . . . . . . . . . . 85

Approximation of Functions of Class Lipða; pÞ in Lp -Norm . . . . . . . . . 109

A New Genuine Durrmeyer Operator. . . . . . . . . . . . . . . . . . . . . . . . . 121

Degree of Approximation by Certain Genuine Hybrid Operators. . . . . 131

Frames in Semi-inner Product Spaces. . . . . . . . . . . . . . . . . . . . . . . . . 149

Applications of Generalized Monotonicity to Variational-Like

Simultaneous Approximation Properties of q-Modiﬁed

Nonlinear Mixed Variational-Like Inequality with Respect

Pál Type (0; 1)-Interpolation on Mixed Tchebycheff Abscissas-II . . . . . 197

Degree of Approximation of f ‰ L½0; ‘Þ by Means

Sparse Approximation of Overdetermined Systems

kth Order Kantorovich Modification of Linking

Rate of Convergence of Modified Schurer-Type q-Bernstein

Operators of Durrmeyer Type with Respect to Arbitrary Measure . . . 255

Construction of Sparse Binary Sensing Matrices Using

Topological and Nontopological 1-Soliton Solution

Mild Solutions for Impulsive Functional Differential Equations

An OðN 1 lnNÞ4 Parameter Uniform Difference Method

Almost Periodicity of a Modified Leslie–Gower Predator–Prey

Existence of a Mild Solution for Impulsive Neutral Fractional

Numerical Solution of Highly Oscillatory Nonlinear Integrals

Approximate Controllability of Semilinear Stochastic System

Fourth-Order Derivative-Free Optimal Families

Existence of Solution for Fractional Stochastic

Singularly Perturbed Convection-Diffusion Turning Point

High-Order Compact Finite Difference Method

On Stability of Steady-States for a Two-Dimensional

Fractional Functional Impulsive Differential Equation

Controllability of Nonlinear Fractional Neutral Stochastic

Efficient Meshfree Method for Pricing European

A Laplace Transform Approach for Pricing European Options . . . . . . 451

Quintic Hermite Fractal Interpolation in a Strip: Preserving

Existence Result for Semilinear Fractional Stochastic

A Numerical Investigation of Blood Flow in an Arterial

Improving R-Order Convergence of Derivative Free

Numerical Solutions of Differential Equations Using Modified

(iii) For i = 1, 2, ki (., .) ∈ C r ([−1, 1] × [−1, 1]), r ≥ 1. Let Mi = ki ∞ =

[(K1 ψ1 ) (x0 )x](1) ∞

≤ M1 d1 x∞ + 2k1 1,∞ d1 x∞ < ∞. (5)

Pn u − u L 2 ≤ c inf u − φ L 2 → 0, as n → ∞. (11)

Pn u L 2 ≤ p1 u∞ , u ∈ X. (12)

Tn (x0 ) − T (x0 )∞ → 0, as n → ∞.

[Pn (K1 ψ1 ) (x0 ) − Pn (K1 ψ1 ) (x)]y∞

[Pn (K2 ψ2 ) (x0 ) − Pn (K2 ψ2 ) (x)]y∞

≤ (Pn − I)[(K2 ψ2 ) (x0 ) − (K2 ψ2 ) (x)]y∞ + [(K2 ψ2 ) (x0 ) − (K2 ψ2 ) (x)]y∞

αn = (I − Tn (x0 ))−1 (Tn (x0 ) − T (x0 ))∞

where αn = (I − Tn (x0 ))−1 (Tn (x0 ) − T (x0 ))∞ .

|Tn (x0 )x(t)| = |[(K1 ψ1 ) (Pn x0 ) + (K2 ψ2 ) (Pn x0 )]Pn x(t)|

(K1 ψ1 ) (Pn x0 )Pn x∞