Applied Mathematics E-Notes, 23(2023), 528-536 c
Available free at mirror sites of http://www.math.nthu.edu.tw/ amen/
ISSN 1607-2510
Generalizations Of Kantorovich’s Inequality Induced By
Majorization Theory
Huan-Nan Shiy, Dong-Sheng Wangz, Chun-Ru Fux
Received 7 September 2022
Abstract
In this paper, we apply Karamata’s theorem combined with majorization theory to establish a new
inequality for the upper bound of the product of two …nite sums of convex functions. As applications,
we derive some new generalizations of Kantorovich’s inequality.
1
Introduction and Preliminaries
Let us start with some fundamental notations or de…nitions needed in this paper. The symbols R and N will
denote the set of real numbers and the set of positive integers, respectively. For convenience, let
Rn = R
|
and
R
R
}
{z
n
Rn++ = fx = (x1 ; : : : ; xn ) 2 Rn : xi > 0; i = 1; : : : ; ng:
In particular, R1++ , simply denoted by R++ , is R++ := (0; 1).
Rn is called convex if
De…nition 1 ([14, 15, 16]) A set
( x1 + y1 ; : : : ; xn + yn ) 2
for any
;
2 [0; 1] with
+
= 1 and x = (x1 ; : : : ; xn ); y = (y1 ; : : : ; yn ) 2 .
De…nition 2 ([14, 15, 16]) Let x = (x1 ; : : : ; xn ) and y = (y1 ; : : : ; yn ) 2 Rn . x is said to be majorized by
y (in symbols x y) if
k
k
X
X
x[i]
y[i] for 1 k n 1,
i=1
and
i=1
n
X
i=1
where x[1]
x[n] and y[1]
xi =
n
X
yi ,
i=1
y[n] are rearrangements of x and y in descending order.
De…nition 3 ([2]) Let f : I ! R and g : I ! R be two real-valued functions on an interval I. f and g are
said to be
Mathematics Subject Classi…cations: 26D15, 26A51, 26A45.
of Electronic Information, Teacher’s College, Beijing Union University, Beijing City, 100011, P.R.China
z Basic courses department, Beijing Polytechnic, Beijing 100176, China
x Institute of Mathematics and Physics, Beijing Union University, Beijing 100101, China, Corresponding author.
y Department
528
529
Shi et al.
(i) similarly ordered if
(f (x)
f (y))(g(x)
g(y))
0 for every x; y 2 I;
(f (x)
f (y))(g(x)
g(y))
0 for every x; y 2 I
(f (x)
f (y))(g(y)
g(x))
0 for every x; y 2 I:
(ii) oppositely ordered if
or,
The estimation of the upper bound of the product of two …nite sum is a long lasting mathematical subject.
Pólya-Szegö established a famous inequality as follows:
Theorem 1 (Pólya-Szegö [1, 2]) Let 0 < m1
n
X
a2k
k=1
n
X
1
4
b2k
k=1
r
M1 and 0 < m2
ak
M1 M2
+
m1 m2
r
m1 m 2
M1 M 2
!2
n
X
k=1
bk
ak bk
M2 (k = 1; : : : ; n). Then
!2
(1)
:
In [3], an upper bound inequality equivalent to inequality (1) was introduced:
Theorem 2 (Kantorovich inequality ) Let fxk g, k = 1; : : : ; n be any real number sequence. If 0 < m
xk M , i = 1; : : : ; n, then
!
!
n
n
(M + m)2
1X
1X 1
:
xk
n
n
xk
4M m
k=1
k=1
Kantorovich inequality is a well-known inequality, this inequality is useful in numerical analysis and
statistics, especially in the method of steepest descent. Therefore, it is valuable for its generalization and
application. Over the years, various variations and extensions of this inequality have been investigated by
many authors in several contexts. Reference [3]–[13] has many forms of generalizations and applications.
In 2005, Xu [5] proved the following generalized Kantorovich inequality.
Theorem 3 ([5]) Let
> 0. If 0 < m xk M , i = 1; : : : ; n, then
!
!
n
n
1X
1X 1
(M +1 m
(M
m )(M m)
xi
n i=1
n i=1 xi
4(M m)
+1 2
)
:
The following lemmas are important and will be used for proving our main results.
Lemma 1 ([16, 17]) Let m xi M; i = 1; : : : ; n; n
and unique integer k 2 f0; 1; : : : ; ng such that
n
X
xi = (n
2, and m 6= M . Then there is a unique l 2 [m; M )
1)m + l + kM;
k
i=1
where l, k is determined by
(x1 ; : : : ; xn )
(M; : : : ; M ; l; m; : : : ; m):
| {z } | {z }
k
Remark 1
(i) Because l =
So we can determine k.
Pn
i=1
xi (n k 1)m
Pn
nm
i=1 xi
1
M m
n k 1
kM 2 [m; M ), we see that
Pn
nm
i=1 xi
k
:
M m
530
Generalizations of Kantorovich’s Inequality
(ii) According to the proof of Lemma 1 in reference [17], we know that l is a variable that depends on
x1 ; : : : ; xn and m l < M .
Lemma 2 ([16, 17]) Let x = (x1 ; : : : ; xn ) 2 Rn , y = (y1 ; : : : ; yn ) 2 Rn and g is convex function on I
If x y, then
n
n
X
X
g(yi ):
g(xi )
R.
i=1
i=1
Kontenovich inequality is a famous inverse inequality of the famous Cauchy-Schwarz inequality, from
Theorem 2 and Theorem 3 we observe that the functions corresponding to the two sequences are respectively:
x and x1 , x and x1 ; ( > 0). They have the common oppositely ordered pproperty, and x1 , x1 is convex
function. So this tells us, for two general functions, when they are oppositely ordered and convex, there may
be results similar to Theorem 2 and Theorem 3.
2
New Inequalities for Di¤erentiable Convex Functions
In this section, we establish the following new inequalities which will be applied to established new generalizations of Kantorovich’s inequality and other new results.
Theorem 4 Let 0 < m xi
M for i = 1; : : : ; n. Let f and g be two nonnegative convex functions and
have second derivatives on [m, M ]. Suppose that
(H1) (f (M )
f (m))(g(m)
g(M ))
0, and
(H2) [kf (M ) + (n k 1)f (m) + f (x)]g 00 (x) + [kg(M ) + (n
m x M and 1 k n 1.
k
1)g(m) + g(x)]f 00 (x) + 2f 0 (x)g 0 (x)
0 for
Then
n
X
i=1
!
f (xi )
n
X
i=1
!
g(xi ) (f (M )
f (m))(g(m)
g(M ))
n [f (M )g(m) f (m)g(M )]
2
2
:
(2)
Proof. In order to prove our conclusion, we consider the following two possible cases:
Case 1. If m = M , then (2) is obvious.
Case 2. Suppose m < M . If (f (M ) f (m))(g(m) g(M )) = 0, then the conclusion also holds immediately.
Hence we may assume that (f (M ) f (m))(g(m) g(M )) > 0. By Lemma 1, there exists k 2 N with
1 k n 1, such that
(x1 ; : : : ; xn ) (M; : : : ; M ; l; m; : : : ; m):
| {z } | {z }
n k 1
k
According to the Remark 1, we know that l is a variable that depends on x1 ; : : : ; xn and m
Because f and g are nonnegative convex functions, by Lemma 2, we have
n
X
f (xi )
kf (M ) + (n
k
1)f (m) + f (l);
kg(M ) + (n
k
1)g(m) + g(l):
i=1
and
n
X
g(xi )
i=1
Let
ha (b) := [af (M ) + (n
a
1)f (m) + f (b)] [ag(M ) + (n
a
1)g(m) + g(b)]
l
M.
531
Shi et al.
for (a; b) 2 N
[m; M ]. So, we get
n
X
f (xi )
n
X
g(xi )
hk (l):
i=1
i=1
Clearly, for any …xed a, ha is a function of the variable b. Taking the derivative of ha with respect to the
variable b, we have
h0a (b) = f 0 (b)[ag(M ) + (n
a
1)g(m) + g(b)] + g 0 (b)[af (M ) + (n
1)f (m) + f (b)];
a
and
h00a (b) = f 00 (b)[ag(M ) + (n a 1)g(m) + g(b)]
+g 00 (b)[af (M ) + (n a 1)f (m) + f (b)] + 2f 0 (b)g 0 (b)
0:
Thus for any a, ha is a convex function on [m; M ]. Since l 2 [m; M ], we see that hk (l)
hk (l) hk (M ), where
hk (m) = [kf (M ) + (n
= [kf (M ) + (n
k 1)f (m) + f (m)][kg(M ) + (n
k)f (m)] [kg(M ) + (n k)g(m)]
= f (M )g(M )k 2 + k(n
= '(k);
k)f (M )g(m) + k(n
k
hk (m) or
1)g(m) + g(m)]
k)2 f (m)g(m)
k)f (m)g(M ) + (n
and
'(k) = (f (M )
g(m))k 2 + n(f (M )g(m) + f (m)g(M )
f (m))(g(M )
2f (m)g(m))k + n2 f (m)g(m):
Next, we will …nd the maximum value of '. (f (M ) f (m))(g(m) g(M )) 0, ' has a maximum value.
Since
0
' (k) = 2(f (M ) f (m)) (g(M ) g(m))k + n(f (M )g(m) + f (m)g(M ) 2f (m)g(m)),
0
if ' (w) = 0, then we obtain
w=
2nf (m)g(m) n(f (M )g(m) + f (m)g(M ))
A
=
.
2(f (M ) f (m))(g(M ) g(m))
2B
Hence ' has maximum value
'(w) = '
A
2B
=B
A
2B
2
A
A
4BC A2
+C =
;
2B
4B
where
A = 2nf (m)g(m)
B = (f (M )
n(f (M )g(m) + f (m)g(M ));
f (m))(g(M )
g(m));
2
C = n f (m)g(m):
On the other hand, since
hk (M ) = [kf (M ) + (n
k
= [(k + 1)f (M ) + (n
= hk+1 (m);
1)f (m) + f (M )] [kg(M ) + (n
k
1)g(m) + g(M )]
k
1)f (m)] [(k + 1)g(M ) + (n
k
the maximum value of hk (M ) is the same as that of hk (m). So, we have
n
X
i=1
f (xi )
n
X
i=1
g(xi )
hk (l)
hk (m)
4BC A2
:
4B
1)g(m)]
532
Generalizations of Kantorovich’s Inequality
Since
4BC = 4(f (M )g(M )
f (M )g(m)
= 4n2 f (m)g(m)f (M )g(M )
f (m)g(M ) + f (m)g(m))n2 f (m)g(m)
4n2 f (m)f (M )g 2 (m)
4n2 f 2 (m)g(m)g(M ) + 4n2 f 2 (m)g 2 (m);
and
A2 = 4n2 f 2 (m)g 2 (m)
4n2 f (m)f (M )g 2 (m)
4n2 f 2 (m)g(m)g(M )
+ n2 f 2 (M )g 2 (m) + n2 f 2 (m)g 2 (M ) + 2n2 f (m)f (M )g(m)g(M );
we get
4BC
and hence
A2 =
n2 (f (M )g(m)
f (m)g(M ))2
n2 (f (M )g(m) f (m)g(M ))2
4BC A2
=
4B
4(f (M ) f (m))(g(M ) g(m))
0:
Therefore we show that
n
X
f (xi )
n
X
g(xi )
i=1
i=1
n2 (f (M )g(m) f (m)g(M ))2
;
4(f (M ) f (m))(g(M ) g(m))
or, equivalence,
n
X
i=1
!
f (xi )
n
X
!
g(xi ) (f (M )
i=1
f (m))(g(m)
n (f (M )g(m) f (m)g(M ))
2
g(M ))
2
:
The proof is completed.
As a consequence of Theorem 4, we can obtain the following generalized Kantorovich’s inequality.
Corollary 1 Let 0 < m xi
M for i = 1; : : : ; n, and let f and g be two nonnegative convex functions
and have second derivatives on [m, M ]. If f; g are oppositely ordered on [m, M ] and (f g)00 0 on [m, M ],
then
!
! n
n
2
X
X
n (f (M )g(m) f (m)g(M ))
:
g(xi ) (f (M ) f (m))(g(m) g(M ))
f (xi )
2
i=1
i=1
Proof. Since f; g are oppositely ordered on [m, M ], we have (f (M ) f (m))(g(m) g(M )) 0 and condition
(H1) as in Theorem 4 is proved. We conclude that condition (H2) as in Theorem 4 holds. In fact note that
(f g)00
0 on [m; M ]
if and only if
f (x)g 00 (x) + f 00 (x)g(x) + 2f 0 (x)g 0 (x)
For m
x
M and 1
k
n
0 for x 2 [m; M ].
1, we obtain
[kf (M ) + (n
k
1)f (m) + f (x)]g 00 (x)
+[kg(M ) + (n k 1)g(m) + g(x)]f 00 (x) + 2f 0 (x)g 0 (x)
= [kf (M ) + (n k 1)f (m)]g 00 (x) + [kg(M ) + (n k 1)g(m)]f 00 (x)
+f (x)g 00 (x) + g(x)f 00 (x) + 2f 0 (x)g 0 (x)
0:
Therefore all the conditions of Theorem 4 are satis…ed and the conclusion follows from Theorem 4 immediately.
533
Shi et al.
3
New generalizations of Kantorovich’s inequality
Applying Theorem 4 or Corollary 1, we can obtain new exponential generalizations of Kantorovich’s inequality.
Theorem 5 Let 0 < m
(i)
= 1 and
(ii)
> 1 and
xi
M for i = 1; : : : ; n. If one of the following conditions is satis…ed:
< 0;
< 0 with
+
1 or
+
0,
then
n
1X
x
n i=1 i
!
n
1X
x
n i=1 i
!
(M
m )(m
M m
M )
m M
2
:
2
(3)
Proof. (i) Suppose that = 1 and < 0. Let f (x) = x and g(x) = x for x > 0. Then f (x) is an increasing
convex function and g(x) is a decreasing convex function. Clearly, (f (M ) f (m))(g(m) g(M )) 0. Let
m x M and 1 k n 1. Set u = kM + (n k 1)m. Thus u M . Since
[kf (M ) + (n k 1)f (m) + f (x)]g 00 (x)
+[kg(M ) + (n k 1)g(m) + g(x)]f 00 (x) + 2f 0 (x)g 0 (x)
=
[kM + (n
=
=
(u + x) (
1)x
2
x
[(u + x)(1
=
and m
x
M
2
x
[u(
k
1)m + x] (
2
1)l
+2 x
) 2x]
+ 1) + (
2
+2 x
1
1
1)x]
u, we get
[kf (M ) + (n k 1)f (m) + f (x)]g 00 (x)
+ [kg(M ) + (n k 1)g(m) + g(x)]f 00 (x) + 2f 0 (x)g 0 (x)
0.
Hence all the conditions of Theorem 4 are satis…ed. By Theorem 4, we have
!
!
n
n
M m
m M
1X
1X
m )(m
M )
x
x (M
n i=1 i
n i=1 i
2
2
:
00
(ii) Let '(x) = x for x > 0. Then ' (x) = (
1)x 2 for x > 0. It is easy to see that when
1 or
< 0, h(x) is a convex function on R++ . Suppose that > 1and < 0 with +
1 or +
0. Let
f (x) = x and g(x) = x for x > 0. Then f (x) is an increasing convex function and g(x) is a decreasing
convex function. Obviously, (f (M ) f (m))(g(m) g(M )) 0. Direct calculation gives
(f (x)g(x))00 = f (x)g 00 (x) + g(x)f 00 (x) + 2f 0 (x)g 0 (x)
= (
1)x
+
= ( + )( +
2
+ (
1)x
1)x
+
2
+
2
+2
x
+
2
0:
Therefore all the conditions of Corollary 1 are satis…ed and the conclusion follows immediately from Corollary
1.
The proof is completed.
Remark 2
(a) Taking
= 1 and
< 0 in Theorem 5, we can get Theorem 3.
534
Generalizations of Kantorovich’s Inequality
(b) Let 0 < m
M for i = 1; : : : ; n. For any
xi
An (x)(M
> 0, by Theorem 5, we obtain
m )
m)(M
(M
+1
m
4M m
+1 2
)
Hn (x );
where An (x) is the arithmetic mean of x1 ;
; xn and Hn (x ) is the harmonic mean of x1 ;
In particular, if we take = 1 in last inequality, then
(M + m)2
Hn (x):
4M m
An (x)
Theorem 6 Let 0 < m
M for i = 1; : : : ; n. If one of the following conditions is satis…ed:
xi
2M
p 1;
(i) p > 1 and
(ii) p > 1 and
(iii) p
; xn .
0;
m<
1 and
0,
then
n
1X p
x
n i=1 i
!
n
1X
1
n i=1 ( + xi )p
!
(M p
mp ) [( + M )p
( + m)p ]
2
[M p ( + M )p mp ( + m)p ]
:
4( + m)p ( + M )p
Proof. We only verify (i) and (ii), and a similar argument could be made for proving (iii). Let f (x) = xp
2M
m <
0, then direct
and g(x) = ( + x) p for 0 < x
M . If p > 1 and
p 1 , or p > 1 and
calculation give
f 0 (x) = pxp 1 > 0, g 0 (x) = p( + x) p 1 < 0,
f 00 (x) = p(p
1)xp
2
> 0, g 00 (x) = p(p + 1)( + x)
p 2
> 0,
and
(f (x)g(x))00 = f 00 (x)g(x) + g 00 (x)f (x) + 2f 0 (x)g 0 (x)
= p(p
1)
xp 2
xp
+
p(p
+
1)
( + x)p
( + x)p+2
xp 2
(p 1)( + x)2 + (p + 1)x2
( + x)p+2
xp 2
=p
[(p 1)
2x]:
( + x)p+2
=p
If p > 1 and
m<
So, we get
(f (x)g(x))00 = p
2M
p 1,
xp 1
( + x)p+1
2px( + x)
0, we have
( + x)p+2 > 0, p
If p > 1 and
2p2
since x
(f (x)g(x))00 = p
0 and (p
1)
2x < 0:
xp 2
[(p
( + x)p+2
1)
2x]
0:
M , we obtain
xp 2
[(p
( + x)p+2
1)
2x]
p
xp 2
(2M
( + x)p+2
2x)
0:
535
Shi et al.
On the other hand, since
f (M ) = M p , f (m) = mp , g(M ) =
it is not hard to show that (f (M )
immediately from Corollary 1.
The proof is completed.
Remark 3 In fact, by taking p =
f (m))(g(m)
1
1
and g(m) =
;
p
( + M)
( + m)p
g(M ))
0. Therefore the desired conclusion follows
1 and
= 0 in Theorem 6, we can obtain
!
!
n
n
1X 1
(M + m)2
1X
xk
;
n
n
xk
4M m
k=1
k=1
which is the original Kantorovich inequality. Hence, Theorem 6 is a real generalization of Kantorovich
inequality.
4
Conclusions
In this paper, we use Karamata’s theorem combined with majorization theory to establish an inequality for
the upper bound of the product of two …nite sums about convex function. As applications, we derive some
new generalizations of Kantorovich’s inequality.
Acknowledgements. The authors are indebted to professor Wei-Shih Du and the referees for many
helpful and valuable comments and suggestions.
References should be typed as follows:
References
[1] D. S. Mitrinović, Analytic Inequalities, Springer-Verlag Berlin. Heidelberg. New York, 1970.
[2] J. C. Kuang, Applied Inequalities, 4th ed, Shandong Press of Science and Technology, Jinan, China,
2010 (In Chinese).
[3] P. J. Liu, Kantorovich Inequality, Press of Harbin Industrial University, Harbin, China, 2017 (In Chinese).
[4] W. Greub and W. Rheinboldt, On a generalization of an inequality of L. V. Kantorovich, Proc. Amer.
Math. Soc, 1959, 10: 407–415.
[5] T. Q. Xu, Extension of Kantorovich inequality, Journal of Chengdu University, 24(2005), 81–83.
[6] S. S. Dragomir, M. L.Scholz and J. Sunde, Some upper bounds for relative entropy and applications,
Comput. Math. Appl., 39(2000), 91–100.
[7] Z. Liu, K. Wang and C. Xu, A note on Kantorovich inequality for Hermite matrices, J. Inequal Appl.,
(2011), 1–6.
[8] A. Clausing, Kantorovich-type inequalities, The American Mathematical Monthly, 89(1982), 327–330.
[9] C. G. Khatri and C. R. Rao, Some extensions of the Kantorovich inequality and statistical applications,
J. Multivariate Anal., 11(1981), 498–505.
[10] V. Ptak, The Kantorovich inequality, The American Mathematical Monthly, (102)1995, 820–821.
536
Generalizations of Kantorovich’s Inequality
[11] P. Henrici, Two remarks on the Kantorovich inequality, The American Mathematical Monthly, (68)1961,
904–906.
[12] Z. Liu, On Kantorovich type inequality, Journal of Anshan Institute of Technology, (23)2000, 293–295.
[13] X. N. Lu and Z. G. Xiao, Extension of Kantorovich inequality, Journal of Yueyang University, (10)1997,
33–35.
[14] B. Y. Wang, Foundations of Majorization Inequalities, Beijing Normal Univ. Press, Beijing, China,
1990.
[15] H. N. Shi, Schur Convex Functions and Inequalities, Press of Harbin Industrial University, Harbin,
China, 2017.
[16] A. M. Marshall and I. Olkin, Inequalities:
York:Academies Press, 1979.
Theory of Majorization and its Application, New
[17] H. N. Shi, Schur Convex Functions and Inequalities: Volume 1: Concepts, Properties, and Applications
in Symmetric Function Inequalities, Harbin Institute of Technology PressLtd, Harbin, Heilongjiang and
Walter de Gruyter GmbH, Berlin/Boston, 2019.