1. Introduction
In mathematical analysis and applications, analytic inequalities serve as fundamental tools for comparison, approximation, and optimization. The classical Chebyshev sum inequality states that for any real tuples
and
both are increasing or decreasing, we have
This inequality was generalized by Matharu and Aujla [
1] to the case of positive semidefinite matrices involving the Hadamard (entrywise) product ∘: for any matrices
and
, and any positive numbers
, we have
To discuss a continuous version of this inequality in a compact form, let us denote for each integrable function
,
The latter is called the general Chebyshev functional (see [
2]). The Chebyshev functional (
4) has many applications in numerical quadrature, probability and statistics, and existence for solutions to certain differential equations. It was obtained in [
3] that if such
f and
g are synchronous on
, that is,
for all
, then
The opposite inequality of (
6) holds if both
f and
g are opposite-synchronous on
. Operator extensions of this inequality were presented by Moslehian and Bakherad [
4]. They generalized the Chebyshev integral inequality (
6) to the case of continuous fields of Hilbert space operators with a bounded measurable function involving Hadamard products by using the notion of synchronous Hadamard property. They proved that if two continuous fields
of operators, parametrized by a compact Hausdorff space
equipped with a Radon measure
, have the synchronous Hadamard property, then
where
is a bounded measurable function. Moreover, they gave some Chebyshev-type inequalities concerning operator means and Hadamard products.
A complement of (
6) was introduced by Grüss [
5], providing an estimate of the difference between the integral of the product and the product of the integrals for two functions. For any integrable functions
satisfying the conditions
,
for all
and
are real constants, we have
We can apply Grüss inequalities to estimate error bounds for some integral means and numerical quadrature rules; see e.g., [
6,
7]. In [
8], Gonska, Raşa and Rusu used the terminology Chebyshev–Grüss inequalities referring to Gr
ss inequalities for (special cases of) generalized Chebyshev functionals
which have a general form
where
E is an expression in terms of certain properties of
and some kind of oscillations of
f and
g. They also established new Chebyshev–Grüss inequalities via discrete oscillations.
On the other hand, in the theory of operator product, the notion of tensor product for Hilbert space operators was extended to the Tracy–Singh product for such operators [
9]. Algebraic, order, and analytic properties of the Tracy–Singh product for operator were discussed in [
9,
10]. The notion of tensor product was also generalized to the Khatri–Rao product for Hilbert space operators in [
11]. The work [
11] shows that the Khatri–Rao product and the Tracy–Singh product are related via isometric selection operators.
In this paper, we establish new several integral inequalities of Chebyshev-type for continuous fields of self-adjoint operators involving Khatri–Rao products and operator means. In
Section 2, we give preliminaries on Khatri–Rao and Tracy–Singh products for operators, and Bochner integrability of continuous field of operators on a locally compact Hausdorff space. In
Section 3, we provide Chebyshev-type inequalities involving Khatri–Rao products of operators under the assumption of synchronous Khatri–Rao property. In
Section 4, we establish Chebyshev integral inequalities concerning weighted Pythagorean means and Khatri–Rao products under the assumption of synchronous monotone property. We prove Chebyshev–Grüss inequalities via oscillations for continuous fields of operators in
Section 5. Our results generalize the matrix result [
1] and the result for integrable functions [
8]. Moreover, our results include the results for tensor products of operators, and Khatri–Rao/Kronecker/Hadamard products of matrices, which have been not investigated in the literature. Finally, we summarize our work in
Section 6.
2. Preliminaries
Throughout this paper, let and be complex Hilbert spaces. The symbol stands to the Banach space of bounded linear operators from into itself. The vector space of self-adjoint operators on is denoted by . Denote the set of all positive (positive invertible, respectively) operators on by (, resp.). For any , the situation (, resp.) means that (, resp.).
Through this paper, we apply the projection theorem to decompose
where all
and
are Hilbert spaces. Each operator
and
can be expressed uniquely as operator matrices
where
and
for each
.
2.1. Khatri–Rao Product and Tracy–Singh Product for Operators
Recall that the tensor product of
and
is a unique bounded linear operator
such that for all
and all
,
Fix a countable orthonormal basis
of
. Recall that the Hadamard product of
is defined to be the operator
such that for all
,
It is known that the Hadamard product of
can be expressed as
where
is the isometry defined by
for all
(see e.g., [
12]).
From the previous setting, we define the Khatri–Rao product of
A and
B to be the bounded linear operator from
into itself represented by an operator matrix
We define the Tracy–Singh product of
A and
B to be the bounded linear operator from
into itself represented by an operator matrix
The maps and are bilinear. Moreover, we have:
Lemma 1 ([
9,
10,
11]).
Let be compatible operators. Then- 1.
If and , then .
- 2.
If and , then .
- 3.
.
- 4.
.
- 5.
If A and B are invertible, then .
- 6.
If A and B are positive, then for any .
Lemma 2 ([
11]).
There is an isometry Z such that for any operators and . 2.2. Bochner Integration for Operator-Valued Maps
Throughout this paper, let be a locally compact Hausdorff space endowed with a finite Radon measure . A continuous map is called a continuous field of operators in parametrized by , denoted by . For convenience, for each , we may write instead of . The field A is said to be bounded if there is a constant such that for all . The set of all bounded continuous fields of operators in parametrized by is denoted by . If is such that the norm function is Lebergue integrable on (e.g., ), then we can form the Bochner integral .
Lemma 3 (see e.g., [
13]).
Let and be Banach spaces, and a bounded linear operator. For any Bochner integrable function , the composition is also Bochner integrable and Proposition 1. For any and , we have Proof. The map
is Bochner integrable on
because it is continuous. Since the map
is bounded linear operator, we have by Lemma 3 that the map
is Bochner integrable on
and (
13) holds. □
3. Chebyshev-Type Inequalities Involving Khatri–Rao Products of Operators
We introduce the following property, and prove Chebyshev-type inequalities involving Khatri–Rao products of operators.
Definition 1. The fields A and B of operators parametrized by Ω are said to have the synchronous Khatri–Rao property if, for all , They are said to have the opposite-synchronous Khatri–Rao property if the reverse of (
14)
holds for all . The following result is an extension of the Chebyshev integral inequality (
6) to the case of operators involving Khatri–Rao products.
Theorem 1. Let and , and let be a bounded measurable function.
- 1.
If A and B have the synchronous Khatri–Rao property, then - 2.
If A and B have the opposite-synchronous Khatri–Rao property, then then the reverse of (
15)
holds.
Proof. By using Lemma 1 and Proposition 1, we have
Here, we have used Fubini’s Theorem [
14] to interchange the order of integrals. For the case 1, we have
and thus (
15) holds. For another case, we get the reverse of (
16) and, thus, the reverse of (
15) holds. □
For the case
, i.e.,
and
are not decomposed, the synchronous Khatri–Rao property in Definition 1 reduces to the synchronous tensor property:
If two fields
A and
B of operators parametrized by
have the synchronous tensor property, then
A and
B have the synchronous Hadamard property ([
4], Definition 2.1), i.e.,
for all
. The following result gives Chebyshev-type inequalities involving tensor products and Hadamard products.
Corollary 1. Let and let be a bounded measurable function.
- 1.
If A and B have the synchronous tensor property, then - 2.
If A and B have the opposite-synchronous tensor property, then the reverses of (
17)
and (
18)
holds.
Proof. For the case
, the Khatri–Rao product in Theorem 1 reduces to the tensor product. Assume that
A and
B have the synchronous tensor property. Using the fact that the Hadamard product is expressed as the deformation of tensor product via the isometry
U defined in (
10), we obtain
Case 2 for Hadamard products can be similarly treated. □
We can see that the inequality (
18) is the same as (
7), but they hold under different hypothesis. The next corollary is a discrete version of Theorem 1.
Corollary 2. Let and where , and is a nonnegative number for each .
- 1.
If A and B have the synchronous Khatri–Rao property, then - 2.
If A and B have the opposite-synchronous Khatri–Rao property, then the reverse of (
19)
holds.
Proof. From the previous theorem, consider the finite space equipped with the counting measure and for all . □
This corollary generalizes Chebyshev sum inequalities for the case of real numbers in inequality (
1) and for Hadamard product of matrices in [
1].
Next, we illustrate Chebyshev-type inequalities for bounded linear operators induced from matrices. Recall that with each
one can naturally associate a bounded linear operator
For any complex matrices
and
partitioned in block-matrix form, we have [
11]
Example 1. Consider , , and , where First, we check the hypothesis of Corollary 2. Since , we have . Similarly, . By the positivity of the Khatri–Rao product, we get i.e., the fields and have the synchronous Khatri–Rao product property.
Now, we can check that the following matrix is positive semidefinite: Passing through the induced linear maps, we get Finally, applying the property (
20)
, we have i.e., the inequality (
19)
in Corollary 2 holds. 4. Chebyshev Integral Inequalities Concerning Weighted Pythagorean Means of Operators
We start this section by introducing order assumptions on continuous fields and supplying preliminaries on weighted arithmetic/geometric/harmonic means of operators. The main part is to establish Chebyshev-type inequalities involving Khatri–Rao products concerning such operator means and order assumptions.
Throughout this section, the space is equipped with a total ordering ≼. Consider the following definitions:
Definition 2. We say that a field A is increasing (decreasing, resp.) whenever implies (, respectivley).
Definition 3. Two ordered pairs and of self-adjoint operators are said to have the synchronous property if either The pairs and are said to have the opposite-synchronous property if either Definition 4. Let A, B, C, D be continuous fields of self-adjoint operators parametrized by Ω. Two ordered pairs and are said to have the synchronous monotone property if and have the synchronous property for all . The pairs and are said to have the opposite-synchronous monotone property if and have the opposite-synchronous property for all .
Recall that the three classical Pythagorean means are the following symmetric means: the arithmetic mean, the harmonic mean, and the geometric mean. For each
, the
w-weighted versions of such means are respectively defined for any
by
These means can be defined for arbitrary positive operators by the following continuity argument with respect to the strong-operator topology:
For brevity, we write
for
. The Pythagorean means have the following remarkable property: for any
,
, and
, we have
here
is anyone of
.
Lemma 4 (see e.g., [
15]).
The weighted geometric means, weighted arithmetic means and weighted harmonic means for operators are (jointly) monotone in the sense that if and , then where σ is any of . 4.1. Inequalities on Weighted Geometric Means
Recall that a linear map between two operator algebras is said to be positive if it maps positive operators to positive operators.
Lemma 5 ([
16]).
Let and . Then Theorem 2. Let and let be a bounded measurable function. If are either all increasing, or all decreasing, then Proof. Let
. Without loss of generality, assume that
. By applying Lemmas 1 and 5, Proposition 1, and Fubini’s Theorem [
14], we have
If
are all increasing, we have by Lemma 4 that
and
If
are all decreasing, we have
and
By Lemma 1, both cases lead to the same conclusion that
and hence (
22) holds. □
The next corollary is a discrete version of Theorem 1.
Corollary 3. Let and where for each . If are either all increasing, or all decreasing, then Proof. Setting equipped with the counting measure and for all in Theorem 2, we get the result. □
Operator inequality (
23) can be regarded as a generalization of the Chebyshev sum inequality (
1). The next goal is to establish a reverse version of Theorem 2.
Lemma 6. Let be such that and . Denote , and Then for any , we have Proof. Consider a map
, where
Z is the isometry in Lemma 2. Since
is a unital positive linear map, we have by ([
17], Corollary 3.5) that
From ([
16], Theorem 1), we get
Theorem 3. Let with and for all , and a bounded measurable function. Let as in Lemma 6. If either
- 1.
are increasing and are decreasing, or
- 2.
are decreasing and are increasing,
Proof. Let
. Without loss of generality, assume that
. By applying Lemmas 1 and 6, Proposition 1, and Fubini’s Theorem [
14], we have
We have by Lemmas 1 and 4 that
and hence (
25) holds. □
4.2. Inequalities on Weighted Arithmetic Means
Lemma 7 ([
18]).
Let and .- 1.
If and have the synchronous property, then - 2.
Ifandhave the opposite-synchronous property, then the reverse of (
26)
holds.
Theorem 4. Let and let be a bounded measurable function.
- 1.
If have the synchronous monotone property and all of are either increasing or decreasing, then - 2.
If have the opposite-synchronous monotone property and if either
- (a)
are increasing and are decreasing, or
- (b)
are decreasing and are increasing,
then the reverse of (
27)
holds.
Proof. Let
. Without loss of generality, assume that
. First, we consider the case 1. We have by using Lemmas 1 and 7, proposition 1, and Fubini’s Theorem that
By Lemmas 1 and 4, we have
and hence (
27) holds. The case 2 can be similarly treated. □
We can illustrate Theorem 4 for the case of operators induced from matrices as follows.
Example 2. Consider the following pairs of induced bounded linear operators: , , and , where We can check the hypothesis of Theorem 4: (i) all of are increasing, and (ii) have the synchronous monotone property. Set equipped with the counting measure and for all . Let us denote . Now, a direct computation reveals thator equivalently, Passing through the induced linear maps and applying the propertiesand (
20)
, we obtain Thus, the equality (
27)
holds in this case. 4.3. Inequalities on Weighted Harmonic Means
Lemma 8 ([
18]).
Let and . If and have the opposite-synchronous property, then Theorem 5. Let and a bounded measurable function. If and have the opposite-synchronous monotone property and all of are either increasing or decreasing, then Proof. Let
with
. By applying Lemmas 1 and 8, Proposition 1, and Fubini’s theorem, we get
By Lemmas 1 and 4,
and hence (
28) holds. □
Lemma 9 ([
18]).
Let with and . DenoteIf and have the synchronous property, then for any , Theorem 6. Let be such that for all , and . Let be a bounded measurable function. If and have the synchronous monotone property and if either
- 1.
are increasing and are decreasing, or
- 2.
are decreasing and are increasing,
then with the constant λ in (
29)
we have Proof. Let
with
. We have by using Lemmas 1 and 9, Proposition 1, and Fubini’s Theorem that
We have, by Lemmas 1 and 4,
and hence (
30) holds. □
Remark 1. When we set equipped with the counting measure, we get discrete versions of Theorems 3–6. Matrix analogs of our results can be obtained particularly by setting . In this case, our results include Chebyshev-type inequalities for Khatri–Rao products, Kronecker products and Hadamard products of matrices.