Inverse of a fuzzy matrix of fuzzy numbers

Mehdi Dehghana,∗, Mehdi Ghateea , Behnam Hashemia,b

a Department of Applied Mathematics, Faculty of Mathematics and Computer Science,
Amirkabir University of Technology, No.424, Hafez Avenue, Tehran 15914, Iran
b Department of Mathematics, Faculty of Basic Sciences,
Shiraz University of Technology, Modarres Boulevared, Shiraz, Iran

September 11, 2007

Abstract
The aim of this paper is to extend the concept of inverse of a matrix with fuzzy
numbers as its elements, which may be used to model uncertain and imprecise aspects
of real world problems. We pursue two main ideas based on employing real scenarios
and arithmetic operators. In each case, exact and inexact strategies are provided. In
the first idea, we give some necessary and sufficient conditions for invertibility of fuzzy
matrices based on regularity of their scenarios. And then Zadeh’s extension principle
and interpolation on Rohn’s approach for inverting interval matrices are followed to
compute fuzzy inverse. In the second idea, Dubois and Prade’s arithmetic operators
will be employed for the same purpose. But with respect to the inherent difficulties
which are derived from the positivity restriction on spreads of fuzzy numbers, the con-
cept of ε-inverse of a fuzzy matrix and its relaxation are generalized and some useful
theorems will be revealed. Finally fuzzifying the defuzzified version of the original
problem for introducing fuzzy inverse, which can be followed by each idea, will be pre-
sented.

Keywords: Fuzzy Matrix, Degree of Regularity, Extension Principle, Fuzzy Arith-

metic Operators, Natural Cubic Splines, Fuzzy Inverse, ε-Inverse, Interval Matrix.

1 Introduction
Physical systems may be modeled by mathematical descriptions. In many practical situ-
ations, to measure the value of physical quantities we use regular measuring instruments.
Measurements are commonly not absolutely accurate. So, instead of the actual value x of a
physical quantity, we only know the measurement result x0 . For a measuring instrument the
manufacturer must provide it with some guaranteed error bounds δ. Otherwise, whatever
∗
Corresponding author. E-mail Addresses: mdehghan@aut.ac.ir (M. Dehghan), ghatee@aut.ac.ir, gha-
tee@gawab.com (M. Ghatee), hashemi am@aut.ac.ir, hoseyn.hashemi@gawab.com (B. Hashemi).

1
for the value x0 we measure, the actual value x can be arbitrarily far from x0 . If we know
the bound δ, and the result of the measurement is x0 , then the actual value must belong to
the interval [x0 − δ, x0 + δ]. In many cases, in addition to the bounds on the measurement
error, we also know the probability distribution of the measurement error (usually, normal
with known standard deviation). In this case, traditional statistical methods work well.
However, there are many practical situations when we do not know these probabilities. In
these situations, the only information we have is the intervals of possible values of the actual
quantities, so we have to use techniques of interval computations.
Interval computations has various real-life applications from robotics, automatic control,
image processing, astrophysics, traffic control and expert systems to ergonomics, social sci-
ences and economics [27]. Further applications have also introduced in [26]. A website that
listed all the interval softwares is: http : //www.cs.utep.edu/intervalcomp/ intsof t.html.
On the other hand, in many practical situations, in addition to the intervals that are
guaranteed by the manufacturer, it is reasonable to consider subintervals that the manufac-
turer cannot guarantee, but which the experts (designers or producers of the manufacturing
instrument) claim to be true. To explain the practical usefulness of fuzzy sets, let us recall
a real life example where such subintervals are useful [35].
Equations that describe how AIDS is transferred are known. To solve these equations,
we must know the initial conditions and the numerical values of the parameters (such as the
death rate among AIDS patients, etc). The majority of these parameters can be determined
more or less accurately from the known statistical data. However, there is one important
parameter about which it is very difficult to get a reliable statistical estimate: the number of
intravenous drug abusers. So, in the absence of a reliable statistical estimate, we may use the
expert estimates. Also, note that different experts can have different ideas of which values
are possible and which values are not [35]. So, we have nested intervals for describing the
expert knowledge. Fuzzy methodology provides us with a way to represent expert estimates
in a form understandable by a computer: namely, we represent an experts statement by a
function (called membership function) that assigns to every real number x a degree of belief
that x satisfies the expert estimate. Therefore, fuzzy numbers constitute a very interesting
alternative to intervals that can represent those subintervals with their degrees of belief
(the membership function, in fuzzy notion). A fuzzy number may emerge in a practical
application as a description of the variation in our knowledge about the correct value of
some measurements when the level of our confidence in that knowledge varies. Fuzzy systems
including fuzzy set theory and fuzzy logic have many successful applications. Sophisticated
fuzzy set theoretic methods have been applied to various areas ranging from fuzzy topological
spaces to quantum optics, medicine and so on.
The term ”fuzzy matrix”, that is the main concept of this paper, has at least two different
meanings in the literature. In the first class A = (aij )m×n is called a fuzzy matrix, if
aij ∈ [0, 1], (i = 1, 2, ..., m; j = 1, 2, ..., n). They appear with fuzzy relations and have been
first described in [28] in detail. After that many attentions have been interested in this
theme [21, 4, 37]. For example, Hashimoto [21] used the Gödel-implication operator and
showed some properties of sub-inverses of fuzzy matrices of the first class. Also, Cho [4]
introduced regularity properties of them in 1999. On the other hand, a matrix with fuzzy
number entries is called a fuzzy matrix, too [3, 13, 20, 23]. With respect to their complicated
arithmetic structure, investigation on the second class is neglected. In this article we shall

2
focus our attention on this class of fuzzy matrices.
Also note that there is a strong connection between interval and fuzzy matrices, because
each α−cut of a fuzzy matrix of fuzzy numbers is an interval matrix. The problems of check-
ing invertibility and finding the inverse of square interval matrices have been warm topics
in the recent researches. For example, Rohn [44, 45], Rex and Rohn [39], Rump [46] and
Kuttler [32] have examined the singularity and regularity of interval matrices with presenting
several useful conditions employing spectral radius, eigenvalues and positive definiteness of
some related real matrices. Jansson [24] introduced a method which can utilized for checking
regularity of interval matrices, based on the topological and graph theoretical properties of
the solution set of linear interval systems. Also, an algorithm is presented in [25] for checking
regularity of interval matrices. In addition, Rohn [40] utilized an infinite series to create an
analytic scheme for getting inverse of real scenarios for the class of inverse-positive interval
matrices. Also, Herzberger and Bethke [22] and Rohn [41, 43] defined the narrowest inter-
val matrix containing the set of real scenarios inverse, as the inverse of an interval matrix.
NP-hardness of checking regularity of interval matrices was propounded by Poljak and Rohn
[38] in 1988, too.
In this paper we pursue two ideas for finding inverse of a fuzzy matrix: In the first idea,
which is called ”scenario-based”, we will consider real matrix A in the support of fuzzy
matrix A e (which is named a scenario in this paper), and define fuzzy inverse with respect to
scenarios. In the second idea, which we name ”arithmetic-based”, we will try to find a fuzzy
matrix B e as the inverse of fuzzy matrix A, e such that Ae⊗B e = I,
e where Ie is fuzzy identity
matrix.
The rest of this paper is organized as follows. In Section 2 some basic definitions and
results on fuzzy numbers, their arithmetic operators and basic notation are given. In Section
3 we define the concept of inverse of a fuzzy matrix. In Section 4 we present necessary and
sufficient conditions for the invertibility of fuzzy matrices in the scenario-based idea. Then
we use Zadeh’s extension principle and Rohn’s scheme to find fuzzy inverse. In Section
5 Dubois and Prade’s arithmetic operators are utilized to introduce fuzzy inverse in the
arithmetic-based idea. Then, we mention the inherent difficulties for traditional inverse
problem in fuzzy nature and introduce the concept of ε-inverse and its relaxation version
with some theorems and properties. Section 6 ends this paper with a discussion on fuzzifying
the defuzzified problem as another useful method with a brief conclusion and suggestions for
future works.

2 Preliminary Concepts
Fuzzy numbers are one way to describe the vagueness and lack of precision of data. They
are developed based on the theory of fuzzy sets which Zadeh [50] was introduced in 1965.
In order to keep this paper self-contained, we give some concepts that will be used later.

Definition 2.1 Let X be a universal set and e

a be a fuzzy set with the membership function
µea . The set
a) = {x ∈ X | µea (x) > 0},
Supp(e
a. We also use Supp(e
is called the support of e a) as the closure of Supp(e
a).

3
Definition 2.2 The α−cut of a fuzzy set e
a is the following ordinary set

a]α = {x|µea (x) ≥ α, α ∈ [0, 1]}.

[e

We separately define [e
a]0 = Supp(e
a).

Definition 2.3 Let f : X → Y be a mapping from a set X to set Y . Then the extension
principle allows us to define the fuzzy set eb in Y induced by the fuzzy set e
a in X through f
as follows:
eb = {(y, µe(y))|y = f (x), x ∈ X},
b

with
supy=f (x) µea (x), f −1 (y) 6= φ,

µeb (y) = µf (ea) (y) =
0, f −1 (y) = φ,

where f −1 (y) is the image inverse of y.

Definition 2.4 A fuzzy number is a convex normalized fuzzy set of the real line R1 whose
membership function is piecewise continuous [47]. Fuzzy number e a is called positive (nega-
tive), denoted by e
a > 0 (e
a < 0), if its membership function µea (x) satisfies µea (x) = 0, for
each x < 0 (x > 0). Also F(R) is shown the class of all fuzzy sets on real numbers. It is
a is an interval number denoted by [a, a].
well-known that the α−cut of a fuzzy number e

Let ea and eb be two fuzzy numbers with the membership functions µea (x) and µeb (x),
respectively. Then according to the extension principle of Zadeh, the binary operation ”.”
in Rn can be extended to the binary operation of fuzzy numbers e a and eb as follows:

µea eb (z) = sup min(µea (x), µeb (y)).

z=x.y

To get fast computation formulas for the operations of fuzzy numbers, Dubois and Prade
introduced the concept of LR fuzzy numbers [11] as follows:

Definition 2.5 A fuzzy number e a is said to be an LR fuzzy number if

L( a−x

α
), x ≤ a, α > 0,
µea (x) = x−a
R( β ), x ≥ a, β > 0,

where a is the mean value of ea and α and β are the left and right spreads, respectively.
L(.) and R(.) are symmetric and non-increasing on [0, ∞) respect to the left and right shape
functions which L(0) = R(0) = 1. Such LR fuzzy number e a is symbolically written

a = (a, α, β)LR .
e

Definition 2.6 For the same shape functions L and R, two LR fuzzy numbers e a = (a, α, β)LR
and b = (b, γ, δ)LR are said to be equal, if a = b, α = γ and β = δ. e
e a is said to be a subset
of b, if a − α ≥ b − γ and a + β ≤ b + δ.
e

4
Definition 2.7 The following formulas for the exact addition, scaler multiplication, sub-
traction and approximate multiplication and division are presented by Dubois and Prade
[13]:
(1) (a, α, β)LR ⊕ (b, γ, δ)LR = (a + b, α + γ, β + δ)LR .

(λa, λα, λβ)LR , λ > 0,
(2) λ ⊗ (a, α, β)LR =
(λa, −λβ, −λα)RL , λ < 0.

(3) a
e eb = (a, α, β)LR (b, γ, δ)RL = (a − b, α + δ, β + γ)LR .

(4) (a, α, β)LR ⊗ (b, γ, δ)LR ∼

= (ab, aγ + bα, aδ + bβ)LR ,

where e
a > 0 and eb > 0.
a δa + αb γa + βb
(5) (a, α, β)LR (b, γ, δ)RL ∼
=( , , )LR ,
b b2 b2
where e
a > 0 and eb > 0.

Similar approximate formulas hold when e a < 0 and eb > 0 or e a < 0 and eb < 0, [47]. Also,
Wagenknecht et al. [48] defined more accurate but slightly difficult operators for the approx-
imate formulas of ⊗ and , which may be utilized in place of Dubois and Prade’s operators.

In addition, many techniques are proposed in the literature for comparing two fuzzy
numbers (see e.g Wang and Kerre [49] for a review). The following simple ordering, based
on subset property, is a direct extension of the same definition for intervals [33].

a α eb if and only if
Definition 2.8 Let ã and b̃ be two fuzzy numbers. We say e
(
max[ea]α ≤ max[eb]α ,
a]α ≥ min[eb]α .
min[e

a  eb when for each α ∈ [0, 1], we have e

Also we write e a α eb.
To introduce the properties of this ordering, one can check the reasonable properties those
proposed by Wang and Kerre [49] as follows.

Proposition 2.9 Consider an arbitrary finite subset A of fuzzy numbers. Let e

a, eb ∈ A. The
following are true.
• e
a α e
a.

• If e
a  eb & eb  e a then ea ' eb, (in which ' denotes equality for fuzzy numbers).
Note that this assertion is not generally true for α .

• e
a α eb & eb α e a α e
c then e c.

• Since α is not a complete order, inf supp(e

a) < sup supp(eb), dose not imply that
a α eb.
e

5
 • Let S and S 0 be two arbitrary finite sets of fuzzy quantities in which α can be applied
a, eb are in S ∩ S 0 . We obtain the ranking order e
and e a α eb on S 0 if e
a α eb on S.

• Let e a⊕e
a, eb, e c, eb ⊕ e a α eb on {e
c be elements of S. If e a⊕e
a, eb}, then e c α eb ⊕ e
c on
{e
a⊕e c, b ⊕ e
e c}.

• Let e a⊗e
a, eb, e c, eb ⊗ e a α eb on {e
c be elements of S. If e c ≥ 0, then e
a, eb} and e a⊗e
c α eb ⊗ e
c
on {e
a⊗e c, b ⊗ e
e c}.
Proof. Straightforward. 
Definition 2.10 Ă is called an interval matrix, if its elements are interval numbers [18,
19, 30]. Similarly Ae is called a fuzzy matrix, if its elements are fuzzy numbers [13, 12]. A e
will be nonnegative (non-positive) and denoted by A e ≥ 0 (A e ≤ 0) if its elements are all
m×n
nonnegative (non-positive) fuzzy numbers. Also F(R) is shown the set of all m × n fuzzy
matrices of fuzzy numbers. Note that the α−cut of fuzzy matrix A e is an interval matrix
which contains interval numbers [f aij ]α = [aij (α), aij (α)]. So, we can denote the α−cut of A e
by [A]
e α = [A(α), A(α)].

We shall use the following notations. Numbers and matrices are denoted by small and
capital letters respectively. The absolute value of a matrix A = (aij ) is denoted by |A| =
(|aij |); the same notation applies to vectors as well. The accents 00 ˘ 00 and 00 e 00 are used
for interval and fuzzy numbers, respectively. Each operation on a fuzzy matrix is defined
componentwise unless other case be imposed. Also, the center and radius matrices of a fuzzy
matrix A e may be defined as follows:

max [A]
e 0 + min [A]
e0
Ac = ,
2
and the radius matrix
e 0 − min [A]
max [A] e0
∆= .
2

Definition 2.11 The multiplication of A e = (faij )m×n and B

e = (bf
ij )n×p may be presented
with the m × p fuzzy matrix C(f
e cij ) and defined as follows
⊕
X
cf
ij = ik ⊗ bkj ,
af f
k=1,...,n

3 What is the inverse of a fuzzy matrix?

To answer this question we pursue two ideas for finding inverse of a fuzzy matrix: ”scenario-
based” and ”arithmetic-based” ideas. In the first idea, we will consider real matrix A (which
is derived from fuzzy matrix A
e and is named a scenario in this paper), and define fuzzy inverse
with respect to scenarios. In addition, we will show that checking regularity (invertibility)
of a fuzzy matrix is NP-hard, as the same result satisfies for interval matrices. Nevertheless,

6
we will extend some necessary and sufficient conditions for invertibility of fuzzy matrices.
In this step, the base of our work is Zadeh’s extension principle [13, 50]. But we show that
this terminology may not necessarily result a fuzzy matrix. In the next step, we will accept
some approximations and introduce the approximate inverse fuzzy matrix. Here, firstly we
use Rohn’s approach for finding the inverse of an interval matrix [41, 42], on the α−cuts of a
fuzzy matrix. Then, we use interpolation methods to get a fuzzy matrix of which its entries
are fuzzy numbers. This idea is detailed in Section 4.
In the second idea, which we name ”arithmetic-based”, we will try to find a fuzzy matrix
B as the inverse of fuzzy matrix A,
e e such that A e⊗B e = I,
e where Ie is fuzzy identity matrix.
In this case, we employ fuzzy approximate arithmetic operators [48, 12] and define fuzzy
inverse for nonnegative fuzzy matrices with nonnegative fuzzy inverse. We show that this
terminology may produce some fuzzy numbers with negative spreads. But since we need
positive spreads for fuzzy numbers, again we adhere some approximations and defined the
concept of ε-inverse of a fuzzy matrix, which allow us to define fuzzy inverse by tolerance ε.
We expand this idea in Section 5.

4 Scenario-based idea
Consider the fuzzy matrix Ae and a real matrix A which is derived from A e and is named a
scenario in this paper. We define fuzzy inverse with respect to scenarios.

4.1 Regularity of a fuzzy matrix

This section is devoted to the problem of checking invertibility of fuzzy matrices. The
approximate inverses of uncertain matrices has been introduced in [14]. Also, the singularity
and regularity of interval matrices are defined as follows (See e.g Rohn [44, 42] and Kreinovich
et al. [30]).

Definition 4.1 A square interval matrix Ă is said to be regular if each A ∈ Ă, which named
as scenario, be nonsingular and it is called singular if it contains a singular matrix.

Rohn [44], Rex and Rohn [39] and Rump [46] presented several useful conditions for
clarifying singularity and regularity of interval matrices by using spectral radius, eigenvalues
and positive definiteness of some real matrices which are related to the original interval
matrix.
Naturally, regularity and singularity may be generalized for fuzzy matrices as follows.

Definition 4.2 Let A e = (f aij ) be a n × n fuzzy matrix. Ae is called α−regular, if each

A ∈ [A]α is nonsingular. Also for α < β in [0, 1], A e is named (α, β)−regular if each A in
[A]α and out of [A]β is regular, i.e. denote [A]α = [A(α), A(α)], [A]β = [A(β), A(β)], we say
Ae is (α, β)−regular if each A thatA(α) ≤ A ≤ A(β) or A(β) ≤ A ≤ A(α) is nonsingular.
Finally, Ae is called regular if each A ∈ Supp(A)e is nonsingular. Similarly, we may define
α−singular and singular fuzzy matrices.

7
Remark 4.3 Note that A e is α−regular if and only if A e is (α, 1)−regular and is regular if
and only if it is (0, 1)−regular, i.e. 0−regular. Similarly, A e is regular if and only if the
e α is regular (as defined in [44]), for all α ∈ [0, 1].
interval matrix [A]

Example 4.4 Consider the following three fuzzy matrices

 
(1, 0.1, 0.1) (0, 0, 0)
Ae1 ,
(0, 0, 0) (1, 0.1, 0.1)
 
(1, 1, 2) (0, 0, 1)
A
e2 ,
(0, 0, 0) (1, 2, 3)
 
(1, 0, 2) (0, 0, 3)
A
e3 ,
(0, 0, 1) (0, 1, 2)

where each element is a LR fuzzy number with linear functions L and R, which may be named
as triangular numbers. By using Theorem (4.7) and Remark (4.8), we can show that A e1 is
regular, while A
e2 is 0.49−regular but not regular (because the lower bound of its 0−cut is a
singular matrix). Also A e3 is nor regular neither α−regular for any α ∈ [0, 1], because the
center matrix of Ae3 which exist in each α−cut, is singular. But A e3 is (0.2, 0.5)−regular.
The following theorem presents necessary and sufficient conditions for singularity of square
fuzzy matrices.

e is α−regular, then for each α0 ∈ [α, 1] the fuzzy matrix A

Remark 4.5 If A e is also α0 −regular.

Theorem 4.6 A square fuzzy matrix A

e is singular, if and only if

(6) |Ac x| ≤ ∆|x|,

has a nontrivial solution.

Proof. Let Ae be singular. Thus there exist a singular matrix A∗ ∈ Supp(A).

e So we can
choose a α ∈ (0, 1] that

A∗ ∈ [A]
e α = [A(α), A(α)] ⊆ [A]
e 0,

be singular. According to Proposition 10 in [44], the system |Ac x| ≤ ∆|x| has a nontrivial
solution.
Conversely let Equation (6) be satisfied where x 6= 0. So [A]
e 0 is singular and the proof is
complete. 

Lemma 1 Let Ă ⊆ B̆ be two arbitrary interval matrices.

1) If B̆ is regular, then Ă is also regular. (The converse statement is not hold).
2) If Ă is singular, then B̆ is also singular. (But the converse is not necessarily true).

Proof. Straightforward.

8
Theorem 4.7 Consider the fuzzy matrix A. e Assume α < β are given in [0, 1]. Each of the
following conditions implies that A
e is (α, β)−regular.

(7) 1) max { ρ(|A−1 −1

c,1 |∆1 ), ρ(|Ac,2 |∆2 ) } < 1.

(8) 2) σmax (∆1 ) < σmin (Ac,1 ) and σmax (∆2 ) < σmin (Ac,2 ),

where
1 1
Ac,1 = (A(α) + A(β)), Ac,2 = (A(α) + A(β)),
2 2
and
1 1
∆1 = (A(β) − A(α)), ∆2 = (A(α) − A(β)),
2 2
with [A]
e α = [A(α), A(α)] and [A]
e β = [A(β), A(β)].

Proof. Since (7) we have ρ(|A−1 −1

c,1 |∆1 ) < 1 and ρ(|Ac,2 |∆2 ) < 1. From Theorem (12) in
[44] the intervals [A(α), A(β)] and [A(β), A(α)] are regular. Thus the fuzzy matrix A e is
(α, β)−regular. The proof of the next part is similar.

Remark 4.8 Let A e = (A, M, N ) be a fuzzy matrix with LR fuzzy numbers as its entries and
for α ∈ [0, 1], set
1
Ac (α) = A + (R−1 (α)N − L−1 (α)M ),
2
and
1
∆(α) = (R−1 (α)N + L−1 (α)M ),
2
as the parametric center and radius matrices. So it is clear that if

ρ(|A−1
c (α)|∆(α)) < 1, or σmax (∆(α)) < σmin (Ac (α)),

then Ae will be α−regular. Moreover, for α = 0 each of the above assumptions imply the
regularity of A.
e

Theorem 4.9 Each of the following conditions implies α−singularity of fuzzy matrix A.
e

(9) 1) max (∆(α)(|A−1

c (α)|)jj ≥ 1.
j

(10) 2) (∆(α) − |Ac (α)|)−1 ≥ 0.

Proof. By using Theorem (12) of [44] the proof is clear.

Please note that according to Remark (4.5) the α−regularity is not a unique property.
To define the degree of regularity for a fuzzy matrix we also need the following results.

9
Definition 4.10 Consider the following set of ±1−vectors

Y = {y ∈ Rn | |yj | = 1 f or j = 1, ..., n},

and for each vector q ∈ Y the n × n diagonal matrix defined by (Tq )ii = qi and (Tq )ij = 0 for
i 6= j, i, j = 1, ..., n. Denote Ayz by

Ayz = Ac − Ty ∆Tz ,

for y, z ∈ Y.

Kreinovich [31] has noted that the optimal number of these special vertex matrices for
checking regularity of an n × n interval matrix is 22n−1 .

Lemma 2 [44] An interval matrix Ă is regular if and only if determinants of all the matrices
Ayz , y, z ∈ Z are nonzero and of the same sign.

Let us define the degree of regularity of a fuzzy matrix A e denoted by α∗ as 1 − β ∗ where

β ∗ is a degree of belief that A e is not regular. Hence in order to define α∗ , we must define β ∗ .
The degree of belief that aij is a possible value of afij is described by the membership function
µafij (aij ). Also, for given real numbers aij the regularity of a scenario matrix A = (aij ) is a
crisp statement, thus its degree of belief is 1 (=true) or 0 (=false). Therefore we arrive at
the following formula:

β ∗ = max{min[µag
11 (a11 ), µa
g 12 (a12 ), ..., µa
g nn (ann ), β(a11 , a12 , ..., ann )]},
aij

where β(a11 , a12 , ..., ann ) is equal to 1 if the corresponding scenario matrix A = (aij ) is not
regular and to 0 if it has the regularity and max is taken over all possible scenario matrices
A. In addition, note that when β = 0, then min = 0, thus we do not need to consider these
values when we compute max. Also, when β = 1, then since min(x, 1) = x, for every degree
of belief x ∈ [0, 1], we do not need to consider this term in the min. Therefore, we can
simplify the above formula into the following one:

(11) β ∗ = max{min[µag
11 (a11 ), µa
g 12 (a12 ), ..., µa
g nn (ann )]},
aij

where max is taken over all aij for which its corresponding scenario matrix is not regular.
From (11), the degree of regularity of a fuzzy matrix can be defined by α∗ = 1 − β ∗ .
Here we utter an algorithm to determine the degree of regularity of A
e = (f
aij ) as similar as
one that has proposed in [5] for computation of the degree of P-property for a fuzzy matrix.
This algorithm consists of k iterations. On the j-th iteration, we have an interval [h− , h+ ]
that contains β ∗ . After k iterations, we can take any endpoint of the resulting interval as
the desired estimate for β ∗ , and compute α∗ = 1 − β ∗ .

Algorithm 1
Step 1. Set h− = 0, h+ = 1, N = 1, and ε  1.
Step 2. Compute h = 21 (h+ + h− ).
Step 3. Find the real (crisp) values aij and aij for which µafij (aij ) = µafij (aij ) = h. So, the

10
interval matrix  = [aij , aij ] is formed, that corresponds to the h-cut of the fuzzy matrix A.
e
Step 4. Use Lemma 2 to investigate the regularity of the interval matrix Â.
Step 5. If  is a regular interval matrix, set h+ = h. Else if  is not a regular interval
matrix, set h− = h.
Step 6. Take β ∗,N = h+ and compute α∗,N = 1 − β ∗,N . The accuracy is 2−N .
Step 7. If the accuracy 2−N ≤ ε, then α∗ = α∗,N . Otherwise set N = N + 1 and return to
Step two.

On each iteration, we detect the regularity of an interval matrix (Step 4). Also, as noted
by Kreinovich [31] by the use of Lemma 2, we need to check the regularity of 22n−1 real
crisp matrices. Totally we have k(22n−1 ) crisp matrices to check the regularity. So, the most
time-consuming part of this algorithm is checking weather an interval matrix is regular, i.e.
Step 4.

4.2 Using Zadeh’s extension principle

Now we use Zadeh’s extension principle [13, 50] to introduce the inverse of a fuzzy matrix.
A similar method utilized for obtaining fuzzy rank in [23]. Let Ă be a square regular interval
matrix. The concept of inverse of Ă is introduced in the literature (see e.g., [18]) as follows:
Set
B̆ = {A−1 |AA−1 = I, A ∈ Ă}.
Thus the set B̆ is called the inverse of interval matrix Ă.
Note that
• B̆ is not necessarily an interval matrix [34]. But if each element of B̆ is an interval
number, it will be an interval matrix.
• Regularity is a necessary condition for the validation of the above definition [39, 46].
Otherwise Ă has no inverse.
Now we use a similar way to introduce the inverse of a square fuzzy matrix. Suppose A
e is
regular, else consider its biggest invertible subset. Set
Ω(α) = {A−1 |AA−1 = I, A = (aij )ij , aij ∈ [f
aij ]α },
for α ∈ [0, 1].
e ∈ F(R)n×n by
Define the membership function of B
µBe (B) = sup{α | B ∈ Ω(α)},

for B ∈ Ω(0) and define µBe (B) = 0 when B ∈

/ Ω(0). Then B e is the inverse of fuzzy matrix
A.
e Note that B
e does not represent a fuzzy matrix, unless all of its entries be fuzzy numbers.

Theorem 4.11 Consider the fuzzy matrix A. e The inverse of A

e is the fuzzy matrix B
e if and
only if its membership function defined as follows
e −1
µBe (B) = sup {α | B ∈ [A]α }.
06α61

11
 Proof. From definition of Rohn [40, 43] for the inverse of an interval matrix, the proof is
clear.

Proposition 4.12 Checking regularity of a fuzzy matrix, is NP-hard.

Proof. Since the same result is satisfied for interval matrices, (See e.g. [44, 30]), and with
Theorem (4.11), the proof is straightforward. 

4.3 Using Rohn’s scheme

Using Zadeh’s extension principle for introducing fuzzy inverse, may not necessarily result a
fuzzy matrix. As a first computational attempt at finding the inverse of fuzzy matrix A,
e we
follow the idea of Rohn [41, 42]. Let Ă = [A, A] be an n × n regular interval matrix. Rohn
defined the inverse of Ă as the narrowest interval matrix containing the set {A−1 |A ∈ Ă},
i.e. as the interval matrix B̆ = [B, B] whose bounds are given by

(12) B ij = min{(A−1 )ij |A ∈ Ă},

(13) B ij = max{(A−1 )ij |A ∈ Ă}.

In addition, he obtained some important theorems in order to find B and B [41, 42]. From
the immediate result of the Theorem 1.1 in [42] the following formulas were obtained

(14) B ij = min{(A−1
yz )ij | y, z ∈ Y },

(15) B ij = max{(A−1
yz )ij | y, z ∈ Y },

where i, j = 1, ..., n.
Now we follow the same strategy to find the inverse of fuzzy matrix A. e Fix α ∈ [0, 1].
We can find bounds for the inverse of [A]
e α using the above method as B(α) = [B(α), B(α)],
where B(α) = (B ij (α))i,j and B(α) = (B ij (α))i,j . Our aim is to find the functions Lij and
Rij which satisfy

Lij (α) = B ij (α),

Rij (α) = B ij (α),

for α ∈ [0, 1]. Thus we could approximate the inverse of fuzzy matrix A e with B e = (Beij )
where Beij is a Lij Rij fuzzy number, if and only if for every α1 , α2 ∈ [0, 1], that α1 < α2 we
have [B(α2 ), B(α2 )] ⊆ [B(α1 ), B(α1 )]. The following theorem, guarantees satisfying of this
condition.

Theorem 4.13 Let [A] e α1 and [A]

e α2 be two α−cut matrices of fuzzy matrix A
e such that
α1 < α2 . Then we have

[B(α2 ), B(α2 )] ⊆ [B(α1 ), B(α1 )].

12
 Proof. We have [A] e α1 = [A(α1 ), A(α1 )] and [A]
e α2 = [A(α2 ), A(α2 )]. Since α1 < α2 , we
have [A(α2 ), A(α2 )] ⊆ [A(α1 ), A(α1 )]. So, we get

{A−1 : A ∈ [A]
e α2 } ⊆ {A−1 : A ∈ [A]
e α1 }.

Thus using Equations 14 and 15, it is clear that B(α1 ) < B(α2 ) and B(α1 ) > B(α2 ). So we
have [B(α2 ), B(α2 )] ⊆ [B(α1 ), B(α1 )]. 
We use the matrix form of natural cubic splines. Suppose we have the following data
table
α α0 α1 ... αm
B B0 B1 ... Bm
in which αi and Bi (i = 0, 1, , ..., m) are m + 1 scalars in [0, 1] and n × n real matrices,
respectively. We construct the matrix
 
S11 (α) S12 (α) . . . S1n (α)
 S21 (α) S22 (α) . . . S2n (α) 
S(α)  ,
 
.. .. .. ..
 . . . . 
Sn1 (α) Sn2 (α) . . . Snn (α)

such that Sij (α) : [0, 1] → R is a natural cubic spline which interpolates the following table
of data
α α0 α1 ... αm
(B)ij (B0 )ij (B1 )ij ... (Bm )ij
The interested reader is referred to the book by Kincaid and Cheney [29] for more details
about natural cubic splines. Similarly, we construct the n × n matrix S(α) = Sij (α) which
interpolates the following table
α α0 α1 ... αm
B B0 B1 ... Bm

So, we get the matrix S(α) in which Sij (α) = (Sij (α), Sij (α)) approximates the matrix
Bij (α).
Remark 4.14 We may use the Lagrange interpolation in place of splines. But for large
number of data points, the Lagrange interpolation leads to a polynomial of large degree and
the Runge’s phenomenon shows that this maybe dangerous [1]. So we prefer to use splines.

Example 4.15 Consider the following fuzzy matrix

 
(80, 4, 8) (25, 5, 5)
A=
e ,
(50, 10, 6) (120, 6, 10)
with positive LR entries in which L(x) = R(x) = 1 − |x|, i.e. a fuzzy matrix of triangular
fuzzy numbers. We choose m = 4 and α ∈ {0, 0.25, 0.50, 0.75, 1}, to get five interval matrices
 
(76, 88) (20, 30)
Ă0 = ,
(40, 56) (114, 130)

13
  
(77, 86) (21.25, 28.75)
Ă1 = ,
(42.5, 54.5) (115.5, 127.5)
 
(78, 84) (22.5, 27.5)
Ă2 = ,
(45, 53) (117, 125)
 
(79, 82) (23.75, 26.25)
Ă3 = ,
(47.5, 51.5) (118.5, 122.5)
 
(80, 80) (25, 25)
Ă4 = .
(50, 50) (120, 120)

Using Rohn’s scheme, we get the following interval matrices:

 
(0.0114, 0.0164) (−0.0043, −0.0014)
B̆0 = ,
(−0.0081, −0.0028) (0.0077, 0.0109)
 
(0.0122, 0.0158) (−0.0039, −0.0018)
B̆1 = ,
(−0.0075, −0.0037) (0.0082, 0.0105)
 
(0.0129, 0.0153) (−0.0036, −0.0022)
B̆2 = ,
(−0.0069, −0.0045) (0.0087, 0.0102)
 
(0.0137, 0.0148) (−0.0033, −0.0026)
B̆3 = ,
(−0.0064, −0.0053) (0.0092, 0.0099)
 
(0.0144, 0.0144) (−0.0030, −0.0030)
B̆4 = .
(−0.0060, −0.0060) (0.0096, 0.0096)

Now utilizing natural cubic splines, for example, for B11 we get,


 0.0114 + 0.0033α − 0.0006α3 ; 0 ≤ α ≤ 0.25
2 3
0.0113 + 0.0035α − 0.0007α + 0.0003α ; 0.25 ≤ α ≤ 0.50

B 11 (α) =

 0.0114 + 0.0031α + 0.0001α2 − 0.0002α3 ; 0.50 ≤ α ≤ 0.75
0.0111 + 0.0045α − 0.0018α2 + 0.0006α3 ; 0.75 ≤ α ≤ 1,


and


 0.0164 − 0.0023α + 0.0007α3 ; 0 ≤ α ≤ 0.25
0.0164 − 0.0025α + 0.0008α2 − 0.0003α3 ; 0.25 ≤ α ≤ 0.50

B 11 (α) =

 0.0163 − 0.0021α − 0.0001α2 + 0.0003α3 ; 0.50 ≤ α ≤ 0.75
0.0167 − 0.0037α + 0.0020α2 − 0.0007α3 ; 0.75 ≤ α ≤ 1,


Figure 1 shows the plot of natural cubic splines of each of the elements of matrix B(α).
In each part Bij (α) is shown with dashed-line, while Bij (α) is shown by boxes. Each element
of matrix B(α) is a fuzzy number, as you can see in the Figure 1.

14
5 Arithmetic-based idea for permuted diagonal fuzzy
matrices
In this section, we try to find a fuzzy matrix B e as the inverse of fuzzy matrix A,
e such that
Ae⊗B e = I,e in which Ie = (I, 0, 0) is the fuzzy identity matrix for multiplication of positive
fuzzy matrices. Here, we employ fuzzy approximate arithmetic operators [13]. The use
of Dubois and Prade’s approximate arithmetic operators is restricted to the case that the
sign of numbers is known. We define the fuzzy inverse for nonnegative fuzzy matrices with
nonnegative fuzzy inverses.

Remark 5.1 We denote the class of all nonnegative matrices with nonnegative inverses by
C(N M N I). A square crisp matrix belongs to C(N M N I) if and only if it is the product of a
permutation matrix by a diagonal matrix [10]. In other words, a square crisp matrix belongs
to C(N M N I) if and only if its entries are all zero except for a single positive entry in each
row and column. We call such matrices the permuted diagonal matrices.

Also we can utilize approximate arithmetic operators such as the operators of Wa-
genknecht et al. [48], Giachetti et. al [15, 16] or the method of Oussalah and De Schutter
[36] which permit us to mention the error analysis.

5.1 Using Dubois and Prade’s operators

Definition 5.2 Consider two fuzzy matrices A e = (f
aij ) and Be = (bf
ij ) of the same size. We
say A ⊆ B (A ⊆α B) if and only if af
e e e e ij  bij for every i, j (f
f aij α bij for a fixed α).
f

Lemma 3 Let A e = (faij ), B

e = (bf
ij ) and C = (f
e cij ) be three square positive fuzzy matrices
and for a fixed α ∈ [0, 1], A(α) and A(α) belong to C(N M N I). Then, we have

(16) e⊗B
A e ⊆α A
e ⊗ C,
e if f e ⊆α C.
B e

e⊗B
Proof. Let A e ⊆α A
e⊗C
e for α ∈ [0, 1]. We can write

e ⊗ B]
[A e α ⊆ [A
e ⊗ C]
e α ⇒ [A] e α ⊆ [A]
e α [B] e α [C]
e α,

or equivalently

[A(α), A(α)].[B(α), B(α)] ⊆ [A(α), A(α)].[C(α), C(α)].

Using product of intervals [33] and positivity property, we get

[A(α).B(α), A(α).B(α)] ⊆ [A(α).C(α), A(α).C(α)].

On the other hand, with the interpretation of subset property, we must have

 A(α).B(α) ≥ A(α).C(α),

A(α).B(α) ≤ A(α).C(α).


15
Since A(α) and A(α) belong to C(N M N I), by multiplying two sides of the above inequalities
in A(α)−1 ≥ 0 and A(α) −1 ≥ 0, respectively, we get

 B(α) ≥ C(α),

B(α) ≤ C(α).


Again with employing the subset property, we obtain

e ⊆α C.
B e

We can use the same strategy to prove the converse statement. 

Consider a positive fuzzy matrix Ae = (A, M, N ). Now we introduce a fuzzy matrix
0 0 0
B
e = (A , M , N ) which satisfies

(A0 , M 0 , N 0 ) ⊗ (A, M, N ) = (I, 0, 0).

The simplest way to analyze the above statement is to get B

e as a positive fuzzy matrix. In
this case based on formula (4), we may propose

(17) e = (A−1 , −A−1 M A−1 , −A−1 N A−1 ).

B

By this terminology, consider

 
(4, 3, 2) (5, 2, 1)
A=
e .
(7, 4, 3) (10, 6, 5)
Using the above formula we have
 
(2.00, −6.80, −6.20) (−1.00, 3.60, 3.40)
B=
e .
(−1.40, 4.80, 4.44) (0.80, −2.60, −2.48)

But B e1,1 and B

e2,2 are not valid fuzzy numbers, because they have negative spreads. Even
if they have positive spreads, since Be1,2 and B
e2,1 are negative fuzzy numbers, we are not
allowed to use the Dubois and Prade’s multiplication formula (4). Thus, in these cases the
formula (17) is not valid.

Remark 5.3 Consider the fuzzy matrix A e = (A, M, N ). If the left and right spread matrices
(M and N) are zero matrices and the aim is to find A e−1 , the problem will be reduced to finding
the inverse of the crisp matrix A and in this case the right and left spreads of the inverse
matrix will be zero, i.e. we have (A−1 , 0, 0) as the inverse of (A, 0, 0).

5.2 ε-Inverse of a fuzzy matrix

Now we introduce the concept of ε-inverse of a fuzzy matrix and sufficient conditions for its
existence. Then, we propose a relaxed version of ε-inverse of a fuzzy matrix.

16
Definition 5.4 Let ε be a positive crisp number. A fuzzy matrix (Ā, M̄ , N̄ ), is said to be
the left ε-inverse of fuzzy matrix à = (A, M, N ), if
(18) (Ā, M̄ , N̄ ) ⊗ (A, M, N ) = (I, εI, εI).
Similarly a fuzzy matrix (Ā, M̄ , N̄ ), is said to be the right ε-inverse of fuzzy matrix Ã, if
(19) (A, M, N ) ⊗ (Ā, M̄ , N̄ ) = (I, εI, εI).

Theorem 5.5 Let à = (A, M, N ) be a positive fuzzy matrix and the inverse of the center
matrix A−1 , be nonnegative. Moreover, suppose
(20) ai,j ≥ max{mi,j + ni,j , 2mi,j }.
Then, there exists a nonnegative fuzzy matrix as the left ε-inverse of Ã, if
mi,j ni,j ai,j − mi,j
(21) max{ , } ≤ ε ≤ min{ }.
i,j ai,j ai,j i,j ai,j
Proof. Firstly as a simple mathematical exercise one can check that if (20) is not satisfied,
there is no  fulfilling (21). Moreover, the positive fuzzy matrix (Ā, M̄ , N̄ ) is the left ε-inverse
of Ã, if and only if
(Ā, M̄ , N̄ ) ⊗ (A, M, N ) = (ĀA, ĀM + M̄ A, ĀN + N̄ A) = (I, εI, εI).
e−1 = (Ā, M̄ , N̄ ) ≥ 0 we can write
So if A

 ĀA = I,
ĀM + M̄ A = εI,
ĀN + N̄ A = εI.


Thus, we get

 Ā = A−1 ,
(22) M̄ = (εI − A−1 M )A−1 ,
N̄ = (εI − A−1 N )A−1 .


Since
mi,j ni,j
ε ≥ max{ , },
i,j ai,j ai,j
we have (εI − A−1 M ) ≥ 0 and (εI − A−1 N ) ≥ 0. Moreover due to the given assumption,
A−1 ≥ 0, thus M̄ ≥ 0 and N̄ ≥ 0. So, the inverse matrix (Ā, M̄ , N̄ ) is the left ε-inverse of Ã.
On the other hand, with the right inequality of (21), we have ε ≤ 1. Also, since M is the
left spread of a positive fuzzy matrix, it must be positive. So
A−1 ((1 − ε)I + M A−1 ) ≥ 0.
Also, note that
Ā − M̄ = A−1 − (εI − A−1 M )A−1 A−1 ((1 − ε)I + M A−1 ).
Thus, it is clear that Ā − M̄ ≥ 0. A e−1 = (Ā, M̄ , N̄ ) is a nonnegative fuzzy matrix and the
above assumption on A e−1 is satisfied. 

17
Remark 5.6 With the hypotheses of the previous theorem, there exists a nonnegative fuzzy
matrix as the right ε-inverse of Ã, if
mi,j ni,j ai,j − mi,j
(23) max{ , } ≤ ε ≤ min{ }.
i,j ai,j ai,j i,j ai,j
The following theorem gives supplemental information about the second approach of the
previous section:

Theorem 5.7 Consider the nonnegative fuzzy matrix A e = (A, M, N ) and let A ∈ C(N M N I).
Moreover, suppose that ni,j = 0 if ai,j is zero. Then

(A−1 , A−1 (εI − M A−1 ), A−1 (εI − N A−1 )),

is the left ε-inverse of A,

e where ε > 0 is chosen in the following interval
nij mij mij
[ max{ max{ | aij 6= 0}, max{ | aij 6= 0} } , 1 + max{ | aij 6= 0} ].
i,j aij i,j aij i,j aij
Proof. We must prove

(24) (A, M, N ) ⊗ (A−1 , A−1 (εI − M A−1 ), A−1 (εI − N A−1 ) = (I, εI, εI).

Based on the above assumptions on ε, the fuzzy matrix

e = (A−1 , A−1 (εI − M A−1 ), A−1 (εI − N A−1 )),
B

is positive. Thus (24) can be rewritten as

(AA−1 , AA−1 (εI − M A−1 ) + M A−1 , AA−1 (εI − N A−1 ) + N A−1 ) = (I, εI, εI),

or equivalently

 AA−1 = I,
AA−1 (εI − M A−1 ) + M A−1 = εI,
AA−1 (εI − N A−1 ) + N A−1 = εI,


which is a valid system. 

Theorem 5.8 Let A e = (A, M, N ) be a nonnegative fuzzy matrix and A ∈ C(N M N I), such
that if aij = 0 then nij = 0. Moreover, assume that there exist some ε > 0, such that
mij nij
(25) ε ≥ max{ max{ |aij > 0}, max{ |aij > 0} }.
i,j aij aij
e = (A−1 , 0, 0). Then
Set B
e⊗B
A e ⊆ A e−1 .
e⊗A

Moreover, if for a fixed α ∈ [0, 1], A(α) and A(α) belong to C(N M N I), we have
e−1 .
e ⊆α A
B

18
 Proof. We have
e = (A, M, N ) ⊗ (A−1 , 0, 0) = (I, M A−1 , N A−1 ).
e⊗B
A

Using Equation (25) we get

e = (I, M A−1 , N A−1 ) ⊆ (I, εI, εI) = A
e⊗B
A e−1 .
e⊗A

Now suppose that for a fixed α ∈ [0, 1], A(α) and A(α) belong to C(N M N I). Employing
Lemma (3) we get
e−1 . 
e ⊆α A
B

Example 5.9 Let

 
(100, 8, 1) (0, 0, 0)
Ae= ,
(0, 0, 0) (140, 2, 4)
with nonnegative LR entries. Using ε = 0.0800 we have:
−3
 
−1 (0.0100, 0, 0.7000 × 10 ) (0, 0, 0)
Ae = .
(0, 0, 0) (0.0071, 0.4694 × 10−3 , 0.3673 × 10−3 )
So we get:
 
e−1 =
e⊗A (1, 0.0800, 0.0800) (0, 0, 0)
A ,
(0, 0, 0) (1, 0.0800, 0.0800)

which is close to the fuzzy unitary matrix I.

e

5.3 Relaxation of ε-inverse

In this section we relax the Definition (18) employing the subset property in place of strict
equality. Dubois and Prade used this terminology for the concept of fuzzy linear systems in
[12], too. They called it, the system of tolerance constraints.

Definition 5.10 Let ε be a positive number. A fuzzy matrix (Ā, M̄ , N̄ ), is said to be the
left tolerance ε-inverse of fuzzy matrix à = (A, M, N ), if

(26) (Ā, M̄ , N̄ ) ⊗ (A, M, N ) ⊆ (I, εI, εI).

Similarly a fuzzy matrix (Ā, M̄ , N̄ ), is said to be the right tolerance ε-inverse of fuzzy matrix
Ã, if

(27) (A, M, N ) ⊗ (Ā, M̄ , N̄ ) ⊆ (I, εI, εI).

Finally, if there exists a matrix Be which is the left and right tolerance ε-inverses of fuzzy
matrix Ã, we call it the tolerance ε-inverse.
Note that since the left and right ε-inverses of a fuzzy matrix are not unique, the tolerance
ε-inverse of a fuzzy matrix is not unique.

19
Theorem 5.11 Consider A e = (A, M, N ) with A = (aij ), M = (mij ) and N = (nij ) belong
to C(N M N I). The nonnegative matrix B e = (A−1 , M −1 , N −1 ) is the right tolerance ε-inverse
of A,
e if there exist some ε > 0 which fulfill

aij + k p nik A−1 aij + k p mik A−1

P P P P
k,p npj k,p mpj
max{ } ≤ ε ≤ min{ },
i,j nij i,j mij
in which A−1
k,p denotes (k, p)
th
element of A−1 .

Proof. Since A, M and N belong to C(N M N I), we can write

(A, M, N ) ⊗ (A−1 , M −1 , N −1 ) = (AA−1 , AM −1 + M A−1 , AN −1 + N A−1 ),
from the assumption on ε we have
AM −1 + M A−1 ≥ εI,

AN −1 + N A−1 ≤ εI.

Thus the result is obtained. 

6 A Discussion and Conclusion

One can suggest other methods for constructing inverse of a fuzzy matrix. For example
fuzzifying the defuzzified version of the original problem may be presented. Buckley et.al.
utilized a similar method for obtaining fuzzy eigenvalue in [2]. We may implement the
following algorithmic steps:
Step 1. (Deffuzify the fuzzy matrix A e = (A, M, N )). Choose the crisp matrix A, as the
initial matrix.
Step 2. (Find the inverse of crisp matrix). Find the inverse of A by employing traditional
techniques of numerical linear algebra (See e.g. [17]).
Step 3. (Fuzzify the crisp inverse). Define matrix (A−1 , X, Y ) as the fuzzified inverse of
A−1 and finally find matrices X and Y that satisfy in the definition of fuzzy inverse.
The implementation of this method depends on the definition of the fuzzy inverse. For
example if we utilize the scenario-based idea, we need to find two crisp matrices X and Y ,
such that for all  ∈ Supp(A),
e we have:

A−1 − X ≤ Â−1 ≤ A−1 + Y.

Thus, we must solve:
max ||Â−1 − A−1 ||, and max ||Â−1 + A−1 ||.
e
Â∈Supp(A) e
Â∈Supp(A)

In this paper we discussed the problem of finding inverse of a fuzzy matrix of fuzzy
numbers from both theoretical and computational points of view. We used the scenario-
based and arithmetic-based ideas to introduce fuzzy inverse. In the scenario-based idea,
we generalized Rohn’s approach and in the arithmetic-based idea, we present ε-inverse as
computational methods. It may be useful in various nice areas such as the theory of fuzzy
linear system of equations [7, 9, 8, 6], linear programming and modern control.

20
References
[1] M.B. Allen and E.L. Isaacson, Numerical Analysis for Applied Sciences, John Wiley
and Sons, Inc., New York, (1997).

[2] J.J. Buckley, T. Feuring, and Y. Hayashi, Solving fuzzy equations using evaluationary
algorithms and neural nets, Soft Comput. 6 (2002) 116-123.

[3] J.J. Buckley and Y. Qu, Solving systems of linear fuzzy equations, Fuzzy Sets and
Systems 43 (1991) 33-43.

[4] H.H. Cho, Regular fuzzy matrices and fuzzy equations, Fuzzy Sets and Systems, 105
(1999) 445-451.

[5] M. Dehghan and B. Hashemi, Determination of the degrees of P-property and nonneg-
ative invertibility for a fuzzy matrix, International Journal of Approximate Reasoning,
In press (2006).

[6] M. Dehghan, B. Hashemi, M. Ghatee, Solution of the fully fuzzy linear systems using
iterative techniques, Chaos, Solitons & Fractals, In press.

[7] M. Dehghan and B. Hashemi, Iterative solution of fuzzy linear systems, Appl. Math.
Comput., 175 (2006) 645-674.

[8] M. Dehghan, B. Hashemi, Solution of the fully fuzzy linear systems using the decom-
position procedure, Appl. Math. Comput., In press.

[9] M. Dehghan, M. Ghatee and B. Hashemi, Some computations on fuzzy matrices, Sub-
mitted for publication.

[10] R. DeMarr, Nonnegative matrices with nonnegative inverses, Proc. Am. Math. Soc.
(1972) 307-308.

[11] D. Dubois and H. Prade, Operations on fuzzy numbers, Internat. J. Systems Sci. 9
(1978) 613-626.

[12] D. Dubois and H. Prade, Systems of linear fuzzy constraints, Fuzzy Sets and Systems
3 (1980) 37-48.

[13] D. Dubois and H. Prade, Fuzzy Sets and Systems: Theory and Applications, Academic
Press, New York (1980).

[14] L. El Ghaoui, Inversion error, condition number, and approximate inverses of uncertain
matrices, Linear Algebra and its Applications 343-344 (2002) 171-193.

[15] R.E. Giachetti and R.E. Young, Analysis of the error in the standard approximation
used for multiplication of triangular and trapezoidal fuzzy numbers and the development
of a new approximation, Fuzzy Sets and Systems 91 (1997) 1-13.

21
[16] R.E. Giachetti and R.E. Young, A parametric representation of fuzzy numbers and their
arithmetic operators, Fuzzy Sets and Systems 91 (1997) 185-202.
[17] G.H. Golub and C.F. Van Loan, Matrix Computations, Johns Hopkins University Press,
Baltimore, (1989).
[18] E. Hansen, Interval arithmetic in matrix computations, Part I, SIAM J. Numer. Anal.
2 (1965) 308-320.
[19] E. Hansen, Interval arithmetic in matrix computations, Part II, SIAM J. Numer. Anal.
4 (1967) 1-9.
[20] B. Hashemi, Topics in Fuzzy Numerical Linear Algebra, M.Sc. Thesis, Amirkabir Uni-
versity of Technology (Tehran Polytechnic), Tehran, April 2006.
[21] H. Hashimoto, Subinverses of fuzzy matrices, Fuzzy Sets and Systems 12 (1984) 155-168.
[22] J. Herzberger and D. Bethke, On two algrithms for bounding the inverse of and interval
matrix, Interval Computations 1 (1991) 44-53.
[23] H.N. Ismail and N.N. Morsi, Fuzzy rank of a fuzzy matrix, Fuzzy Sets and Systems 41
(1991) 243-249.
[24] C. Jansson, Calculation of exact bounds for the solution set of linear interval systems,
Linear Algebra Appl. 251 (1997) 321-340.
[25] C. Jansson and J. Rohn, An algorithm for checking regularity of interval matrices, SIAM
J. Matrix Anal. Appl. 20 (1999) 756-776.
[26] L. Jaulin, M. Kieffer, O. Didrit, and E.Walter, Applied Interval Analysis, With Exam-
ples in Parameter and State Estimation, Robust Control and Robotics. London, U.K.:
Springer-Verlag, 2001.
[27] R.B. Kearfott and V. Kreinovich, Applications of interval computations, Kluwer Acad-
emic Publishers, Dordrecht/Boston/London (1996).
[28] K.H. Kim and F.W. Roush, Generalized fuzzy matrices, Fuzzy Sets and Systems 4
(1980) 293-315.
[29] D. Kincaid and W. Cheney, Numerical Analysis Mathematics of Scientific Computing,
Brooks/Cole, California (1991).
[30] V. Kreinovich, A. Lakeyev, J. Rohn, and P. Kahl, Computational Complexity and
Feasibility of Data Processing and Interval Computations, Kluwer Academic Publishers,
Netherlands (1998).
[31] V. Kreinovich, Optimal finite characterization of linear problems with inexact data,
Reliable Computing, 11 (2005) 479-489.
[32] J. Kuttler, A fourth-order finite-difference approximation for the fixed membrance eigen-
problem, Math. Comp., 25 (1971) 237-256.

22
[33] R.E. Moore, Method and Application of Interval Analysis, SIAM, Philadelphia (1979).

[34] A. Neumaier, Interval Methods for Systems of Equations, Cambridge University Press,
Cambridge (1990).

[35] H.T. Nguyen and V. Kreinovich, Nested intervals and sets: concepts, relations to fuzzy
sets, and applications. In Applications of Interval Computations, (Edited by R.B. Kear-
fott et al.), pp. 254-290, Kluwer Academic Publishers, Netherlands, (1996).

[36] M. Oussalah, J. De Schutter, Approximated fuzzy LR computations, Inform. Sciences

153 (2003) 155-175.

[37] CT. Pang, Simultaneousley contrallable fuzzy matrices, Computers and Mathematics
with Applications 50 (2005) 1647-1658.

[38] S. Poljak and J. Rohn, Radius of nonsingularity, Research Report, KAM Series, Charles
University, (1988) 88-117.

[39] G. Rex and J. Rohn, Sufficient conditions for regularity and singularity of interval
matrices, SIAM J. Matrix Anal. Appl. 20 (1998) 437-445.

[40] J. Rohn, Inverse-positive interval matrices, Z. Angew. Math. Mech. 67 (1987) 492-493.

[41] J. Rohn, Systesms of linear interval equations, Linear Algebra Appl. 126 (1989) 39-78.

[42] J. Rohn, Interval matrices: Singularity and real eigenvalues, SIAM J. Matrix Anal.
Appl. 14 (1993) 82-91.

[43] J. Rohn, Inverse interval matrix, SIAM J. Numer. Anal. 30 (1993) 864-870.

[44] J. Rohn, Cheking properties of interval matrices, Technical Report 686 of Charles Uni-
versity (1996).

[45] J. Rohn, Regularity of interval matrices and theorems of the alternatives, Reliab. Com-
put. 12 (2006) 99-105.

[46] S. M. Rump, Almost sharp bounds for the componentwise distance to the nearest sin-
gular matrix, Linear and Multilinear Algebra 42 (1998) 93-108.

[47] M. Sakawa, Fuzzy Sets and Interactive Multiobjective Optimization, Plenum Press,
New York and London (1973).

[48] M. Wagenknecht, R. Hampel, and V. Schneider, Computational aspects of fuzzy arith-

metics based on Archimedean t-norms, Fuzzy Sets and Systems 123 (2001) 49-62.

[49] X. Wang and E. Kerre, Reasonalble properties for the ordering of fuzzy quantities(I),
Fuzzy Sets and Systems 118 (2001) 375-385.

[50] L.A. Zadeh, Fuzzy sets, Inform. Control 8 (1965) 338-353.

23