1. Introduction
In Bennett’s paper [
1], a number of interesting discrete inequalities and their consequences are studied. These results are the starting point for several papers (see e.g., [
2,
3,
4,
5,
6,
7,
8,
9]).
The origin of the following inequality also goes back to Bennett’s paper [
1]. It is an abstract version of an inequality obtained by Bennett.
We say that the real intervals , , and are nonoverlapping if any two have at most one point in common.
Theorem 1 (
Theorem of 1 [
8])
. Suppose that C is an interval carrying a positive Borel measure μ and , , and are three nonoverlapping compact subintervals of C of a positive measure. Then and provide a set of necessary and sufficient conditions under which every convex function f defined on C verifies the inequality A similar result is found in [
2] for six disjointed compact subintervals of
C and for three-convex functions at a point.
Theorem 1 is generalized in [
8] by using the concepts of Steffensen–Popoviciu and dual Steffensen–Popoviciu measures, which play an important role in the theory of convex functions.
The -algebra of Borel sets on an interval is denoted by .
Definition 1. Let be an interval with a nonempty interior, and let μ be a finite signed measure on .
(a) μ is called a Steffensen–Popoviciu measure on C if andfor every nonnegative, convex, and μ-integrable function . (b) μ is called a dual Steffensen–Popoviciu measure on C if and inequality (2) holds for every nonnegative, concave, and μ-integrable function . The generalization of Theorem 1 is as follows.
Theorem 2 (
Theorem 4 of [
8])
. The statement of Theorem 1 remains valid if μ is a finite signed Borel measure on C and , , and are three nonoverlapping subintervals of C such that the restriction of μ to each of the intervals and is a Steffensen–Popoviciu measure and the restriction of μ to is a dual Steffensen–Popoviciu measure. In this paper, we first extend Theorem 1 in a more general framework. Instead of inequality (
1), we consider the new inequality
where
is a measurable space with a finite signed measure
on
,
is a
-integrable function,
,
, and
are three nonoverlapping and nonempty subintervals of the real interval
C, and
is a convex function for which
is
-integrable
. It is worth stressing that the intervals can be of any type, not necessarily compact. Starting from a generalization of a majorization-type integral inequality from the recent paper [
10], we give necessary and sufficient conditions for inequality (
3) to be satisfied. The result includes Theorems 1 and 2 as special cases.
As an application, we first give a simple proof of the integral Jensen and Lah–Ribarič inequalities for finite signed measures which results come from [
10]. The formulation of the statements is also sharper than in [
10]. As a second application, a common generalization of some of the main results of [
5] is obtained. Our result uses finite signed measures instead of the classical Lebesgue measure and gives a joint proof of the theorems considered from [
5], which are proven by specific methods there. Finally, we give interesting refinements of the Hermite–Hadamard–Fejér inequality. It is worth noting that the result includes an inequality that can also be used to estimate the weighted arithmetic mean over a bounded open interval. This is a novel extension of the Hermite–Hadamard–Fejér inequality.
2. Preliminary Results
Let be a measurable space ( always means a -algebra of subsets of X). If is either a measure or a signed measure on , then the real vector space of -integrable real functions on X is denoted by .
The interior of an interval is denoted by .
The following generalization of Theorem 6 of [
10] plays a crucial role in proving the main result of this paper. It is interesting in itself and well applicable.
Theorem 3. Let and be measurable spaces, let be finite signed measures on , and let be finite signed measures on . Let be an interval with a nonempty interior, and let , be functions such that and .
Then, for every convex function for which and , the inequalityholds if and only ifand Proof. We can follow the proof of Theorem 6 of [
10], with appropriate modifications.
The proof is complete. □
Remark 1. Analogously to Theorem 6 in [10], the previous statement could be formulated for (nonnegative) increasing, (nonnegative) decreasing, or nonnegative convex functions only. We now give the characterization of the Steffensen–Popoviciu and the dual Steffensen–Popoviciu measures.
A characterization of Steffensen–Popoviciu measures is presented in [
10]. Characterization of Steffensen–Popoviciu measures for compact intervals occurred much earlier, independently due to Popoviciu [
11] and Fink [
12].
The identity function on an interval is denoted by .
Theorem 4 (
see Theorem 11 of [
10])
. Let be an interval with a nonempty interior, and let μ be a finite signed measure on such that . Then(a) For every nonnegative, convex and μ-integrable function the inequalityholds if and only ifand (b) μ is a Steffensen–Popoviciu measure if and only if and (6) is satisfied. Characterization of dual Steffensen–Popoviciu measures for compact intervals and continuous concave functions can be found in a recent paper [
13]. We generalize the result slightly by dropping the continuity condition.
Theorem 5. Let be a compact interval with a nonempty interior, and let μ be a finite signed measure defined on . Then,
(a) For every nonnegative concave function , the inequalityholds if and only ifwhere the first term is taken to be 0 for , while the third term is taken to be 0 for . (b) μ is a dual Steffensen–Popoviciu measure if and only if and (7) is satisfied. Proof. (a)If
f is continuous, we can apply Theorem 3.4 of [
13].
If f is not continuous, then there exists an increasing sequence of nonnegative concave functions such that is continuous and converges pointwise to f on C, and therefore the result follows from the first part of the proof and B. Levi’s theorem.
(b) It is proven in [
13] that (
7) implies
.
The proof is complete. □
We often use the following simple statement in this paper.
Lemma 1. Let be a measurable space, and let μ be a finite signed measure on . Let be an interval with a nonempty interior, and let be a function such that .
(a) Suppose that and t is not the left endpoint of C. Ifthen (b) Suppose that and t is not the right endpoint of C. Ifthen Proof. Only (a) is proved, and (b) can be treated similarly.
Let
be an increasing sequence in
converging to
t. Then, the sequence of functions
is increasing and converges pointwise to the function
Since
is finite, the monotone convergence theorem implies that
The proof is complete. □
We need the following result. This is a simplified version of Lemma 4 of [
14].
Lemma 2 (
see Lemma 4 of [
14])
. Let with , and let μ be a finite signed measure on such that Finally, for completeness, we give the Hermite–Hadamard inequality for measures and its special case, the Hermite–Hadamard–Fejér inequality.
Theorem 6 (see [
15])
. Let a, , .(a) Let μ be a finite measure on with . Then,for every convex function , whererepresents the barycenter of μ. (b) Let λ be the Lebesgue measure on . Let the weight function be nonnegative, Lebesgue-integrable, and symmetric about the vertical line , that is, Thenfor every convex function . 3. Main Results
The following result is a common generalization of Theorems 1 and 2.
Theorem 7. Let be an interval with a nonempty interior. Let , , and in this order be nonoverlapping and nonempty subintervals of C. Let be the endpoints of . Let be a measurable space, and let be a finite signed measure on . Let be a function such that .
Then, for every convex function for which , the inequalityholds if and only ifandand Proof. Theorem 3 is obviously applicable, and only the actual form of conditions (
4) and (
5) needs to be given.
It is easy to see that the two conditions in (
4) are exactly (
8) and (
9) in the current situation.
In the present case, inequality (
5) is equivalent to
It is sufficient to show that the previous inequality is satisfied exactly if inequalities (
10)–(
12) hold. We will justify this by taking into account the position of
w. The cases considered may include empty cases, depending on the intervals.
In the following, let be fixed.
(i) If
, then (
13) takes the following form:
which is obviously satisfied by (
8) and (
9).
(ii) Now suppose that
. Then, (
13) is
and this is the same, because of (
8) and (
9), as
For
, (
13) means that
and according to (
8) and (
9), it is the same as
These two cases give (
10).
(iii) Assume
. Then, (
13) has the form
which is identical to (
11).
(iv) Now, let
. Then, (
13) is equivalent to
If
, then (
13) can be rewritten as follows:
(v) Finally, if
, then (
13) is obviously satisfied.
The proof is complete. □
Remark 2. If is a measure in the previous theorem , then conditions (10)–(12) are automatically met. Conditions (10) and (12) are obviously satisfied, and only (11) needs to be justified. Let be fixed. Condition (11) is proved if the following inequalities are true: The first inequality comes from the easily verifiable facts that the function in the second integral is nonnegative and conditions (8) and (9) imply the second inequality. A simple calculation shows thatwhich yields the third inequality. If we assume that
in Theorem 7, then conditions (
10)–(
12) are further simplified. They can be formulated using only the interior of the intervals
,
and
.
Proposition 1. Suppose that conditions of Theorem 7 are satisfied with the additional constraint that . Then, (10)–(12) are equivalent to the following conditions:and Proof. We only need to show two things.
(i) We first verify that (
14) implies
Since
and
, it is enough to show that (
16) holds for
.
If
, then the integral in (
16) is zero for
.
If
, then Lemma 1 (a) can be applied by using (
14).
(ii) Second, we confirm that (
15) yields
Since
and
, all we need to prove is that (
17) holds for
.
If
, then the integral in (
17) is zero for
.
If
, then Lemma 1 (b) can be applied by using (
15).
The proof is complete. □
The following statement is a simple but important consequence of Theorem 7. This is the version that can be easily compared with Theorem 1. We consider only the case related to Proposition 1.
Theorem 8. Let be an interval with a nonempty interior. Let , , and in this order be nonoverlapping and nonempty subintervals of C. Let be a finite signed measure on such that . Assume and .
Then, for every convex function for which , the inequalityholds if and only ifandand Proof. Since is necessarily bounded, the function and the restriction of f to are also bounded, and therefore the finiteness of implies that and are -integrable. The result then follows directly from Proposition 1.
The proof is complete. □
As a special case of the previous theorem, we obtain an extension of Theorem 2.
The following notations are used for some special functions defined on an interval
:
Corollary 1. Let be an interval with a nonempty interior. Let , , and in this order be nonoverlapping subintervals of C, whose interior is not empty. Let , be the endpoints of . Let be a finite signed measure on such that and .
If (19) and (20) are satisfied, and are Steffensen–Popoviciu measures, and is a dual Steffensen–Popoviciu measure, then the inequality (18) holds for every convex function for which . Proof. According to Theorem 8, we need to show that conditions (
21)–(
23) are satisfied.
Since the function
is nonnegative and convex for all
and the measure
is a Steffensen–Popoviciu measure, (
21) follows.
Similarly, since the function
is nonnegative and convex for all
and the measure
is also a Steffensen–Popoviciu measure, (
23) follows too.
For every
, let
be the affine function joining to the points
and
. Then,
is nonpositive,
is concave and nonnegative, and
. Using this, the properties of the measures, and (
19) and (
20), we obtain
which gives (
22).
The proof is complete. □
Remark 3. For a finite signed measure, we can use Theorems 4 and 5 to decide whether it is a Steffensen–Popoviciu measure or a dual Steffensen–Popoviciu measure.
4. Applications
Theorem 12 of [
10] is an extension of the classical integral Jensen and Lah–Ribarič inequalities to finite signed measures. As a first application, we give a sharper formulation and simpler proof of these results.
Theorem 9. Let be a measurable space, and let μ be a finite signed measure on such that . Let be a μ-integrable function. Let be the narrowest interval containing the range of φ, and let be the endpoints of . Then,
(a) Letand assume . The inequalityholds for every convex function for which if and only ifand (b) If andthen . If andthen . (c) The inequality (24) holds for every convex function for which if and only if (27) and (28) are satisfied. (d) Assume a, and . For every convex function , the inequalityholds if and only if Proof. If is a constant function, then the statements are obvious.
In the following, we suppose that is not a constant function, that is, the interior of the interval is nonempty.
(a) Define the interval
by
and the intervals
and
by
when
, and
when
.
Then, , , and are nonempty and nonoverlapping intervals.
Then, and are nonempty and measurable subsets of X. Let denote the restriction of to , and the restriction of to . Let , consider the measure on , and let , .
With these notations, the conditions of Theorem 7 are satisfied and inequality (
24) can be written as
According to Theorem 7, it is sufficient to show that (
8) and (
9) are satisfied and (
10)–(
12) are equivalent to (
25) and (
26). All of these are easy to check.
(b) The proof of Theorem 12 (a) in [
10] can be copied.
(c) If the inequality (
24) holds for every convex function
for which
, then
, and by (a), conditions (
25) and (
26) are satisfied. We have to prove that (
27) and (
28) are also satisfied.
Let
. Then, by (
25),
If
, (
27) follows from (
26) and Lemma 1 (b).
We can prove similarly that if
, then (
28) holds.
Conversely, if (
27) and (
28) are satisfied, then (
25) and (
26) are obviously satisfied, and by (b),
.
(d) Define the nonoverlapping intervals
,
, and
by
The measurable spaces
with the finite signed measures
on
are defined by
and
The functions
are defined by
With these notations, the conditions of Theorem 7 are satisfied and inequality (
29) can be written as
Now, Theorem 7 implies the result if we show that (
8) and (
9) are satisfied and (
10)–(
12) are equivalent to (
30). All can be obtained by simple calculations.
The proof is complete. □
Remark 4. (a) If there exists a point such that the inequalityholds for every convex function for which , then necessarily. (b) Assume C is an interval such that .
If (27) and (28) are satisfied, then by , these inequalities also hold for every . If is a convex function for which , then is also a convex function such that . However, if is a convex function for which , then f may not be extended to a convex function defined on C. It can be seen that the statements (a), (b), and (c) are the sharpest for .
A common generalization of some of the main results of [
5] is obtained in the following statement. It is formulated in such a way that it can be easily comparable with Theorem 2.2 [
5].
Proposition 2. Let be an interval, and let , , , , , such that . Let be a finite signed measure on such that . Let p, q, and r be positive numbers. Then, the inequalityholds for every convex function if and only ifandand Proof. The result is an immediate consequence of Theorem 8.
The proof is complete. □
Remark 5. (a) As mentioned in Remark 2, for finite measures, conditions (33)–(35) are automatically satisfied. Therefore, in this case, it is sufficient to consider only conditions (31) and (32). Theorem 2.2 of [5] is the special case of the previous result, where is the classical Lebesgue measure on . Then, of course, the form of (32) is Theorem 2.1 of [5] is the special case of the previous result, where is the classical Lebesgue measure on andIn this case, (32) can also be written as (36). Proposition 2 is a generalization of the above theorems even for finite measures.
(b) The proofs of Theorem 2.1 and Theorem 2.2 are unique and direct proofs, and Proposition 2 clearly shows the relationship between the two theorems.
Finally, we give interesting refinements of the Hermite–Hadamard–Fejér inequality.
Proposition 3. Let a, , . Let λ be the Lebesgue measure on . Let the weight function be positive, Lebesgue-integrable, and symmetric about the vertical line , that is, Then, there exists a unique positive number such that
(a) inequalitieshold for every convex function , whererepresents the barycenter of on . (b) inequalities (37) also hold for every convex function for which is Lebesgue-integrable. Proof. (a) Let the measure
, that is,
Then,
is absolutely continuous with respect to
. Since
w is positive, the function
is strictly increasing and
. Furthermore,
p is continuous,
and
due to the symmetry property of
w. It now follows that there exists a unique positive number
such that
Again, using the symmetry of
w, we obtain that
We introduce the following intervals
These intervals are nonoverlapping and
By applying Lemma 2 first to the interval
and then to the interval
, we obtain
Since
it follows from (
38) that
From the above, we can apply Theorem 8, from which we obtain the second and the third inequalities using
Applying the Hermite–Hadamard–Fejér inequality to f on , we obtain the first inequality.
Now, we prove the fourth inequality.
Let
be the barycenter of
on
.
Due to the symmetry property of
w and the choice of
,
By the Hermite–Hadamard inequality,
Again using the Hermite–Hadamard inequality and (
39), we obtain
From the previous two inequalities, the fourth inequality is implied.
The fifth inequality remains to be verified, which, with rearrangement, is equivalent to the following
that is,
This inequality follows immediately from the convexity of f.
(b) Using the intervals
we can prove exactly as in (a).
The proof is complete. □
Remark 6. (a) The two inequalities in (37) illustrate very well the phenomenon of concentration convexity to endpoints. In addition, they refine the Hermite–Hadamard–Fejér inequality. (b) From (a) and (b), the inequalities in (37) can be considered as an extension of the Hermite–Hadamard–Fejér inequality for bounded open intervals. This is particularly interesting for convex functions on which are unbounded but μ-integrable on . The restriction of every convex function to is a bounded and convex function. Conversely, any bounded and convex function on can be extended to a convex function defined on . However, there may be unbounded and convex functions on which are μ-integrable on , and inequality (37) can be applied to such functions. A simple consequence of the previous result is the following refinement of the classical Hermite–Hadamard inequality.
Corollary 2. Let a, , . Then,
(a) inequalitieshold for every convex function . (b) The inequalities in (40) also hold for every convex function for which f is Lebesgue-integrable. Proof. In this case, the number is , and therefore the result follows from Proposition 3 with some calculation.
The proof is complete. □
Remark 7. (a) The inequality on the right hand side of (40) is proved in [6] by means of a different method. In the same paper, the inequalityis shown as a lower bound. The left-hand inequality in (40) is sharper, since This follows from the fact that f is a continuous convex function on , and these two intervals have the same midpoint.
(b) If we want to choose the numbers such thatfor every convex function , then according to Theorem 8,andnecessarily. From these conditions, a simple calculation gives that We can see that the inequalities in (40) are satisfied only for the intervals