Property of μ | Characterization | Reference |
β-convex | −F is completely monotone | Theorem 6 a) |
β-concave | −F is completely monotone | Theorem 6 b) |
We were interested in Bernstein and Lévy measures having certain convexity-type properties. The convexity-type properties were an extension of the harmonic convexity property considered in [
Citation: Wissem Jedidi, Hristo S. Sendov, Shen Shan. Classes of completely monotone and Bernstein functions defined by convexity properties of their spectral measures[J]. AIMS Mathematics, 2024, 9(5): 11372-11395. doi: 10.3934/math.2024558
[1] | Feng Qi, Kottakkaran Sooppy Nisar, Gauhar Rahman . Convexity and inequalities related to extended beta and confluent hypergeometric functions. AIMS Mathematics, 2019, 4(5): 1499-1507. doi: 10.3934/math.2019.5.1499 |
[2] | Chuan-Yu Cai, Qiu-Ying Zhang, Ti-Ren Huang . Properties of generalized $ (p, q) $-elliptic integrals and generalized $ (p, q) $-Hersch-Pfluger distortion function. AIMS Mathematics, 2023, 8(12): 31198-31216. doi: 10.3934/math.20231597 |
[3] | Moquddsa Zahra, Dina Abuzaid, Ghulam Farid, Kamsing Nonlaopon . On Hadamard inequalities for refined convex functions via strictly monotone functions. AIMS Mathematics, 2022, 7(11): 20043-20057. doi: 10.3934/math.20221096 |
[4] | Feng Qi . Completely monotonic degree of a function involving trigamma and tetragamma functions. AIMS Mathematics, 2020, 5(4): 3391-3407. doi: 10.3934/math.2020219 |
[5] | Li Xu, Lu Chen, Ti-Ren Huang . Monotonicity, convexity and inequalities involving zero-balanced Gaussian hypergeometric function. AIMS Mathematics, 2022, 7(7): 12471-12482. doi: 10.3934/math.2022692 |
[6] | Muhammad Aslam Noor, Khalida Inayat Noor . Higher order strongly general convex functions and variational inequalities. AIMS Mathematics, 2020, 5(4): 3646-3663. doi: 10.3934/math.2020236 |
[7] | Qing-Bo Cai, Melek Sofyalıoğlu, Kadir Kanat, Bayram Çekim . Some approximation results for the new modification of Bernstein-Beta operators. AIMS Mathematics, 2022, 7(2): 1831-1844. doi: 10.3934/math.2022105 |
[8] | Muhammad Amer Latif . Fejér type inequalities for harmonically convex functions. AIMS Mathematics, 2022, 7(8): 15234-15257. doi: 10.3934/math.2022835 |
[9] | Shaoyuan Xu, Li Fan, Yan Han . Fixed points of generalized $ \varphi $-concave-convex operators with mixed monotonicity and applications. AIMS Mathematics, 2024, 9(11): 32442-32462. doi: 10.3934/math.20241555 |
[10] | Shuang-Shuang Zhou, Ghulam Farid, Chahn Yong Jung . Convexity with respect to strictly monotone function and Riemann-Liouville fractional Fejér-Hadamard inequalities. AIMS Mathematics, 2021, 6(7): 6975-6985. doi: 10.3934/math.2021409 |
We were interested in Bernstein and Lévy measures having certain convexity-type properties. The convexity-type properties were an extension of the harmonic convexity property considered in [
A function f:(0,∞)→[0,∞) is called completely monotone, if it is C∞ and satisfies
(−1)nf(n)(x)≥0, for all x>0 and n∈N. |
Completely monotone (CM) functions find many applications in analysis and probability and an excellent introduction into their properties can be found in the monographs [7,11]. A function g:(0,∞)→[0,∞) is called Bernstein, if it is C∞ and satisfies
(−1)n−1g(n)(x)≥0, for all x>0 and n∈N. |
We see from the definition that if g is a Bernstein function (BF), then g′ is CM. These two classes of functions have classic integral representations, which are useful for our developments.
Theorem 1 (Bernstein). A function f is CM, if and only if it can be expressed as a Laplace transform
f(x)=∫[0,∞)e−xtμ(dt), | (1.1) |
where μ is a Radon measure on [0,∞), such that the integral converges for all x>0.
The measure μ in the Bernstein representation will be called the Bernstein measure of f.
Theorem 2 (de Finetti-Lévy-Khintchine). A function g is BF if and only if it can be represented as
g(x)=a+bx+∫(0,∞)(1−e−xt)ν(dt) | (1.2) |
for a Radon measure ν on (0,∞) and some constants a,b≥0. The measure ν satisfies
∫(0,∞)(1∧t)ν(dt)<∞. | (1.3) |
The triplet (a,b,ν) uniquely determines the Bernstein function g, and vice-versa.
The measure ν in this representation is usually called the Lévy measure of the Bernstein function g, and (a,b,ν) is called the Lévy triplet of g. The constants a and b≥0 are called the killing rate and the drift term respectively.
Recently, research has focused on different subclasses of CM and BFs. In [8], the authors investigated CM and BFs with measures that satisfy certain convexity properties. A measure μ on [0,∞) is called harmonically convex if x↦xμ[0,x] is a convex function on (0,∞). A measure ν on (0,∞) is said to have harmonically concave tail if x↦xν(x,∞) is a concave function on (0,∞). Among the main results in [8] are the following:
Theorem 3. For any CM function f and a number α∈(0,2/3], there exists a unique harmonically convex measure μα on [0,∞), such that
f(xα)=∫[0,∞)e−xtμα(dt). |
Theorem 4. For any Bernstein function g and a number α∈(0,2/3], there exists a unique triplet (a,b,να), such that
g(xα)=a+bx+∫(0,∞)(1−e−xt)να(dt), |
where a,b≥0 are constants, and να is a measure on (0,∞) with harmonically concave tail. The measure να satisfies the integrability condition
∫(0,∞)(1∧t)να(dt)<∞. |
One of the open problems formulated in [8] was to find the largest possible value of r for which Theorems 3 and 4 hold for all values of α in the interval (0,r]. This question was successfully answered in [2]. It was shown there, see [2, Theorem 6.3], that Theorems 3 and 4 hold for all α∈(0,α∗], where
α∗:=infx>0(log(1+ex−e−x)−log(2−e−x)x)≈0.717461058844... |
and α∗ is the largest value for which Theorems 3 and 4 hold. Theorem 3 suggests that it is natural to consider the set, denoted HCM, of all BFs h, such that the composition f∘h is a CM function having a harmonically convex measure for any CM function f. Analogously, Theorem 4 suggests to consider the set, denoted HBF, of all BFs h, such that the composition g∘h is a BF with measure that has harmonically concave tail for any BF g. In this way, the results in [2] show that {xα:α∈[0,α∗]}⊂HCM∩HBF, so the latter two sets are non-empty. These two sets of functions have surprising properties, see Section 7 in [8]:
(1) We have HCM=HBF;
(2) A BF g∈HBF if and only if x↦1−e−tg(x)(1+txg′(x)) is a BF for all t≥0;
(3) For any Bernstein function f and g∈HBF, one has f∘g∈HBF;
(4) The set HBF is closed with respect to point-wise convergence.
Apart from these properties, very little is known about the set HBF. Is it a convex set? What are its generators? (A function f∈HBF is called a generator for the class HBF if it cannot be represented as a composition g∘h for some non-affine BF g and some h∈HBF.)
A characterization of the BFs, having Lévy measure with harmonically concave tail was proven in [8, Lemma 6.1]. It states that a BF g has Lévy measure ν with harmonically concave tail if and only if g(x)−xg′(x) is a BF. The feat in [2] was accomplished by relaxing this property and considering the class BFs of all BFs g, such that sg(x)−xg′(x) is Bernstein, for some s>0. Then, the set HBF was extended to BF∗s, the later being the set of all BFs g, such that 1−e−tg(x)∈BFs, for all t>0. See Definition 1.4 in [2], where for technical reasons the killing rate and the drift term are removed from g. In particular, [2, Theorem 6.3], shows that
e−λα(1+αλα) is completely monotone if and only if α∈[0,α∗]. |
The latter is related to a problem on the unimodality of reciprocal positive stable distributions raised by Simon in [10].
In the current work, we hope to shed more light into these classes of CM and BFs, by relaxing the notion of harmonic convexity, see Definition 1. For a value of a parameter β∈[0,1], we say that a function h:(0,∞)→R is β-convex (β-concave) if xβh(x) is convex (concave) on (0,∞). Thus, we consider Bernstein measures that are β-convex and Lévy measures with β-concave tail, see Definition 2. The main results may be succinctly summarized as follows and they parallel those in [8,9].
Suppose f is CM with measure μ and define
F(x):=β(β−1)f(x)x−2(β−1)f′(x)+xf″(x)−β(β−1)μ({0})x. |
Then, as shown in Table 1, we have the following characterization of β-convexity (β-concavity) of the measure μ:
Property of μ | Characterization | Reference |
β-convex | −F is completely monotone | Theorem 6 a) |
β-concave | −F is completely monotone | Theorem 6 b) |
Similarly, suppose g is Bernstein with Lévy triplet (a,b,ν). Define
G(x):=β(β−1)g(x)x−2(β−1)g′(x)+xg″(x)−β(β−1)ax−(β−1)(β−2)b. |
Then, as shown in Table 2, we have the following characterization of β-convexity (β-concavity) of the tail of the measure ν:
Property of ν | Characterization | Reference |
β-convex tail | −G is completely monotone | Theorem 7 a) |
β-concave tail | −G is completely monotone | Theorem 7 b) |
The paper is organized as followed: Section 2 introduces the background concepts, notions, and some useful preliminary results. In Section 3, we characterize the CM functions having a β-convex (β-concave) measure and the BFs having a Lévy measure with β-convex (β-concave) tail. Section 4 contains several corollaries from the results in Section 3. Finally, the Appendix collects several classical properties of the Lebesgue-Stieltjes integral that are difficult to find in the formulation that we need.
A function h:I→R is convex on the convex interval I if
h(αx+(1−α)y)≤αh(x)+(1−α)h(y),for x,y∈I, and α∈[0,1]. |
The function h is concave if the opposite inequality holds. If h is twice differentiable in an open interval I, then h is convex on I if and only if its second order derivative is non-negative on I. Convex functions are continuous (in fact locally Lipschitz) on the interior of their domain. The directional derivatives exist (both left and right, in the extended sense) for every x∈I. The right directional derivative, denoted h′+(x), is right-continuous, while the left directional derivative, denoted h′−(x), is left-continuous. When h is convex, then both h′+(x) and h′−(x) are non-decreasing functions in x, see [6, Theorem 24.1]. Moreover, for any x,y in the interior of I we have
h(y)−h(x)=∫(x,y)h′+(t)dt=∫(x,y)h′−(t)dt, | (2.1) |
see [6, Corollary 24.2.1] for details. In addition, if h is convex and y>x, then
h′+(x)≤f(y)−f(x)y−x. | (2.2) |
Definition 1. Let β∈[0,1]. A function h:(0,∞)→R is called β-convex (β-concave) if xβh(x) is convex (concave) on (0,∞).
We consider β∈[0,1] in the following content without further notice. A function h is 0-convex if it is convex; and it is 1-convex precisely when h(1/x) is convex. The latter equivalence follows from Lemma 2.2 in [5], that we state for completeness.
Lemma 1. A function h:(0,∞)→R is convex (concave) if and only if xh(1/x) is convex (concave).
When h(1/x) is convex, we say that h is harmonically convex, since it satisfies the inequality
h(21/x+1/y)≤h(x)+h(y)2 |
for every x,y>0. Such functions are also called reciprocally convex in [5]. Thus, β-convexity connects the notions of convexity and harmonic/reciprocal convexity.
The following equivalence is an immediate consequence from Lemma 1.
Corollary 1. A function h is β-convex (β-concave) precisely when h(1/x) is (1−β)-convex ((1−β)-concave).
If h:(0,∞)→R is β-convex, then the directional derivatives of h(x) exist for all x>0. More precisely, it can be shown that
h′+(x)=−βx−1h(x)+x−β(xβh(x))′+andh′−(x)=−βx−1h(x)+x−β(xβh(x))′−. |
The cumulative distribution function for measure μ on [0,∞) is denoted by
Fμ(x):=μ[0,x], |
while the tail of measure ν on (0,∞) is denoted by
ˉν(x):=ν(x,∞). |
Note that ˉν(x) is non-increasing and a right-continuous function.
Definition 2. Let μ and ν be measures on [0,∞) and (0,∞), respectively.
a) We say that μ is β-convex (β-concave), if Fμ is β-convex (β-concave) on [0,∞).
b) We say that ν has β-convex (β-concave) tail, if ˉν is β-convex (β-concave) on (0,∞).
The next examples illustrate this concept.
Example 1. a) Consider the CM function f(x)=x−α for α∈(0,1). Its measure μ has cumulative distribution function
Fμ(x)=xααΓ(α). |
Since x↦xβFμ(x)=xα+β/(αΓ(α)), μ is a β-convex measure if 1−α≤β≤1, and it is β-concave if 0≤β≤1−α.
b) Consider the BF g(x)=xα for α∈(0,1). Its Lévy measure ν has a tail given by
ˉν(x)=x−αΓ(1−α). |
Since x↦xβˉν(x)=xβ−α/Γ(1−α), ν is a measure with a β-convex tail if 0≤β≤α, and it is a measure with a β-concave tail if α≤β≤1.
A CM function f uniquely determines its Bernstein measure μ. Indeed, for all k∈N and t>0, define the operator
Lk(f(x);t):=(−1)kxk+1f(k)(x)|x=k/t. | (2.3) |
The following is an inversion formula for the Laplace-Stieltjes integrals, see [11, Chapter Ⅶ, Theorems 6a and 7a].
Theorem 5 (Inversion formula). Suppose f is a CM function with measure μ.
a) For every t>0,
limk→∞∫(0,t]Lk(f(x);u)du=μ[0,t]+μ[0,t)2−μ({0}). | (2.4) |
b) If measure μ has density u(t), then for every t>0 in the Lebesgue set of u(t),
limk→∞Lk(f(x);t)=u(t). |
Note that the Lebesgue set of a function contains its points of continuity. If t>0 is a point of continuity of Fμ, the right-hand side of (2.4) becomes Fμ(t)−Fμ(0).
In order to deal with the higher order derivatives in the inversion formula we need the following identity which can be proved by induction (see also [1, Lemma 2.7.12] for a more general case). For any integer k≥0 and a Ck+1 function r on (0,∞), the following identity holds:
(xk+1(r(x)x)(k))′=xkr(k+1)(x). | (2.5) |
Suppose f is CM function with measure μ. An integration by parts in (1.1) leads to
f(x)=x∫(0,∞)e−xtFμ(t)dtandlimt→∞e−xtFμ(t)=0, | (2.6) |
for any x>0. A direct consequence of the definition of a Bernstein function is μ({0})=f(∞). With the latter, one can also represent f as
f(x)=μ({0})+∫(x,∞)(−f′(t))dt, |
where −f′(t) is CM. As a non-increasing function, that is integrable at infinity, is o(1/t) as t approaches infinity, we obtain
limx→∞xf′(x)=0. | (2.7) |
Lemma 2. Suppose f is CM with μ({0})=0. Then
limx→∞xk+1(f(x)x)(k)=0, | (2.8) |
for any k≥1.
Proof. The proof uses the inequality
e−xxk≤(k+1)ke−x/(k+1) for all k≥1 and x≥0, |
which follows from
maxx∈(0,∞)xk(k+1)ke−xk/(k+1)=e−k≤1. |
Indeed, by (2.6), we have
|xk+1(f(x)x)(k)|=x∫(0,∞)e−xt(xt)kFμ(t)dt≤(k+1)kx∫(0,∞)e−xt/(k+1)Fμ(t)dt=(k+1)k+1(−e−xt/(k+1)Fμ(t)|∞t=0+∫(0,∞)e−xt/(k+1)dFμ(t))=(k+1)k+1Fμ(0)+(k+1)k+1f(xk+1). |
The fact that μ has no mass at zero implies Fμ(0)=limx→∞f(x) = 0, and (2.8) follows.
Lemma 3. Suppose f is CM. If its measure μ is β-convex (or β-concave), then
limt→0 t2−β(tβFμ(t))′+e−xt=limt→∞t2−β(tβFμ(t))′+e−xt=0, | (2.9) |
for any x>0.
Proof. As a product of two CM functions, f(x)/x is also CM and so is −(f(x)/x)′. By (2.6), we have
−x(f(x)x)′=x∫(0,∞)e−xttFμ(t)dt=−∫(0,∞)tFμ(t)d(e−xt)=−limt→∞tFμ(t)e−xt+limt→0 tFμ(t)e−xt+∫(0,∞)e−xtd(tFμ(t))=∫(0,∞)e−xtd(tβFμ(t)t1−β)=(1−β)∫(0,∞)e−xtFμ(t)dt+∫(0,∞)e−xtt1−βd(tβFμ(t))=(1−β)(f(x)x)+∫(0,∞)e−xtt1−βd(tβFμ(t)), |
where in the penultimate equality we used Lemma 5. This shows that the last integral has to be finite. Since μ is β-convex (or β-concave), using (2.1), we obtain
∫(0,∞)e−xtt1−βd(tβFμ(t))=∫(0,∞)t1−β(tβFμ(t))′+e−xtdt<∞. | (2.10) |
Next, note that since tβFμ(t) is non-decreasing, then (tβFμ(t))′+ is non-negative. Hence
∫(0,1)t1−β(tβFμ(t))′+dt≤ex∫(0,1)t1−β(tβFμ(t))′+e−xtdt<∞, |
since e−x≤e−xt for all t∈(0,1).
If μ is β-convex, then tβFμ(t) is convex and non-decreasing. Thus, (tβFμ(t))′+ is non-decreasing and non-negative, so is t2−β(tβFμ(t))′+. Suppose
limt→0t2−β(tβFμ(t))′+=c≥0. |
We have
∫(0,1)ctdt≤∫(0,1)t2−β(tβFμ(t))′+tdt=∫(0,1)t1−β(tβFμ(t))′+dt<∞. |
Therefore, c has to be zero. The first limit in (2.9) follows.
To see the second limit, for a fixed ϵ>0, use (2.2) to bound
0≤t2−β(tβFμ(t))′+e−xt≤1ϵt2−βe−xt((t+ϵ)βFμ(t+ϵ)−tβFμ(t))=1ϵt2−βe−xt/2(exϵ/2e−x(t+ϵ)/2(t+ϵ)βFμ(t+ϵ)−e−xt/2tβFμ(t)). | (2.11) |
Then, using (2.6) one can see that the last expression converges to zero as t approaches infinity.
If μ is β-concave, then tβFμ(t) is concave and non-decreasing. Thus, (tβFμ(t))′+ is non-increasing and non-negative. Notice
∫(0,1)(tβFμ(t))′+d(t2−β)=(2−β)∫(0,1)t1−β(tβFμ(t))′+dt<∞, |
we conclude (tβFμ(t))′+ is o(1/t2−β) as t approaches zero by Lemma 8. The first limit in (2.9) follows from here.
For the second limit, note that t1−βe−xt is a decreasing function for large enough t. Thus, t1−β(tβFμ(t))′+e−xt is decreasing and we can apply Lemma 7 to the second integral in (2.10).
Condition (1.3) on the measure ν implies
limt→∞ˉν(t)=0. |
Fubini's theorem applied to the Lévy-Khintchine representation (1.2) gives
g(x)=a+bx+x∫(0,∞)e−xtˉν(t)dt, implying that ∫(0,1)ˉν(t)dt<∞. | (2.12) |
A non-increasing function that is integrable at zero is o(1/t) as t approaches zero, thus
limt→0 tˉν(t)=0 and limt→∞ e−xtˉν(t)=0 for any x>0. | (2.13) |
(For more details about (2.12) and (2.13), refer to (3.3) and (3.6) in [7].) Integration by parts, using the facts that ∫∞0te−xtdt=x−2 and ∫∞0e−xtdt=x−1, shows that
g(x)=x2∫(0,∞)e−xtκ(t)dt, | (2.14) |
where κ(t):=at+b+∫t0ˉν(s)ds is positive, non-decreasing and concave, see [7, p.23, (3.4)].
Lemma 4. If the Lévy measure ν, of a BF g, has β-convex (or β-concave) tail, then
limt→0 tβ−1(t1−βˉν(1/t))′+e−x/t=limt→∞ tβ−1(t1−βˉν(1/t))′+e−x/t=0,for any x > 0 . | (2.15) |
Proof. Without loss of generality, we can assume a=b=0. As g is a Bernstein function, g(x)/x is CM, and so is −(g(x)/x)′. By (2.12) and (2.13), we obtain
−x(g(x)x)′=x∫(0,∞)e−xttˉν(t)dt=−∫(0,∞)tˉν(t)d(e−xt)=−limt→∞tˉν(t)e−xt+limt→0 tˉν(t)e−xt+∫(0,∞)e−xtd(tˉν(t))=−∫(0,∞)e−x/td(t1−βˉν(1/t)tβ−2)=(2−β)∫(0,∞)e−x/tt−2ˉν(1/t)dt−∫(0,∞)e−x/ttβ−2d(t1−βˉν(1/t)), |
where we used Lemma 6 in the penultimate equality and Lemma 5 in the last equality. Now, applying Lemma 6 again, we conclude that
−x(g(x)x)′=(2−β)(g(x)x)−∫(0,∞)e−x/ttβ−2d(t1−βˉν(1/t)). |
This shows that the last integral has to be finite for all x>0. Since ν has β-convex (or β-concave) tail, using Corollary 1 and (2.1), we obtain
∫(0,∞)e−x/ttβ−2d(t1−βˉν(1/t))=∫(0,∞)tβ−2(t1−βˉν(1/t))′+e−x/tdt<∞. |
Denote h(t):=(t1−βˉν(1/t))′+. Applying Lemma 6 again, we have
∫(0,∞)t−βh(1/t)e−xtdt=∫(0,∞)tβ−2h(t)e−x/tdt<∞. | (2.16) |
Next, observe that for all x>0, we have
∫(0,1)t−βh(1/t)dt≤ex∫(0,1)t−βh(1/t)e−xtdt≤ex∫(0,∞)t−βh(1/t)e−xtdt<∞. | (2.17) |
If ν has β-convex tail, then t1−βˉν(1/t) is convex and non-decreasing, hence h is non-decreasing and non-negative. By (2.2), for an ϵ>0, we obtain
0≤tβ−1(t1−βˉν(1/t))′+e−x/t≤1ϵtβ−1e−x/t((t+ϵ)1−βˉν(1/(t+ϵ))−t1−βˉν(1/t)). |
Analogously to (2.11), one can see that the right-hand side converges to 0 as t approaches zero, showing the first limit in (2.15).
Now, h(1/t) is non-increasing and non-negative, and so is t−βh(1/t). As a non-increasing function which is integrable near zero, see (2.17), is o(1/u) as u approaches zero, we have
limu→0u1−βh(1/u)=0. |
The second limit in (2.15) follows.
If ν has β-concave tail, then t1−βˉν(1/t) is concave and non-decreasing, hence h is non-increasing and non-negative. The function tβ−2e−x/t is also non-increasing for t close to zero. Hence, from (2.16) and Lemma 7, we conclude that tβ−2h(t)e−x/t is o(1/t) as t approaches zero. This shows the first limit in (2.15).
Next, h(1/u) is non-decreasing and non-negative, and so is u1−βh(1/u). Supposing
limu→0u1−βh(1/u)=c≥0, |
we would have
∫(0,1)cudu≤∫(0,1)u1−βh(1/u)udu=∫(0,1)u−βh(1/u)du<∞. |
Therefore, c has to be zero and the second limit in (2.15) follows.
By replacing t with 1/t, the previous lemma can be reformulated as follows.
Corollary 2. If the Lévy measure ν of a BF g has β-convex (or β-concave) tail, then
limt→0 t1−β(tβ−1ˉν(t))′+e−xt=limt→∞ t1−β(tβ−1ˉν(t))′+e−xt=0,for any x > 0 . |
In this section, we characterize CM and BFs with β-convexity properties on their measures. In Theorem 6, we consider CM functions with β-convex and β-concave measures. In Theorem 7, BFs whose measures have β-concave tail and β-convex tail are considered. These results extend the characterizations in [8,9]. The boundary cases of both Theorems 6 and 7, when β∈{0,1}, are explored in Corollaries 7 to 14.
Theorem 6. Suppose f is a CM function with measure μ. Consider the function
F(x)=β(β−1)f(x)x−2(β−1)f′(x)+xf″(x)−β(β−1)μ({0})x. | (3.1) |
a) The measure μ is β-convex, if and only if F is CM.
b) The measure μ is β-concave, if and only if −F is CM.
Proof. Notice F can be rewritten as
F(x)=β(β−1)f(x)−μ({0})x−2(β−1)f′(x)+xf″(x). |
Thus, without loss of generality, we can assume μ({0})=0.
a) For sufficiency, suppose F is CM. Anticipating the use of the inversion formula in Theorem 5, define
Gk(t):=∫(0,t]Lk(f(x);u)du, | (3.2) |
where Lk is the operator defined in (2.3). We claim that for every k≥2, the function t↦tβGk(t) is convex on the positive reals. Indeed,
G′k(t)=Lk(f(x);t)=(−1)kxk+1f(k)(x)|x=k/t, |
and using (2.5), we have
G″k(t)=(−1)kddx(xk+1f(k)(x))|x=k/t(−kt2)=(−1)k+1xk(xf(x))(k+1)|x=k/t(k2t2)1k=(−1)k+11kxk+2(xf(x))(k+1)|x=k/t. |
Also notice that
ddt(xk(f(x)x)(k−1)|x=k/t)=ddx(xk(f(x)x)(k−1))|x=k/t(−kt2)=−1kxk+1f(k)(x)|x=k/t, |
therefore, by Lemma 2, we obtain
∫(0,t]−1kxk+1f(k)(x)|x=k/udu=xk(f(x)x)(k−1)|x=k/t−limu→0xk(f(x)x)(k−1)|x=k/u=xk(f(x)x)(k−1)|x=k/t. |
So, we have
Gk(t)=∫(0,t](−1)kxk+1f(k)(x)|x=k/udu=(−1)k−1k∫(0,t]−1kxk+1f(k)(x)|x=k/udu=(−1)k−1kxk(f(x)x)(k−1)|x=k/t. |
Putting everything together and after a trivial calculation, using that (xf(x))(k+1)=2f(k)(x)+(xf″(x))(k−1), we obtain
(tβGk(t))″=tβ−2(β(β−1)Gk(t)+2βtG′k(t)+t2G″k(t))=tβ−2xkk(−1)k−1F(k−1)(x)|x=k/t. | (3.3) |
As F is CM, we know that (−1)k−1F(k−1)(x)≥0, for all x>0 and k≥2, which implies (tβGk(t))″≥0. This concludes the claim.
Let Fμ be continuous at x,y>0 and at the convex combination (1−α)x+αy, where α∈[0,1]. Then, (2.4) shows that
limk→∞Gk(t)=Fμ(t)−Fμ(0)=Fμ(t) |
for t∈{x,y,(1−α)x+αy}. As tβGk(t) is convex for all k≥2, we obtain
((1−α)x+αy)βFμ((1−α)x+αy)≤(1−α)xβFμ(x)+αyβFμ(y). |
We will show that Fμ is continuous on (0,∞), thus completing the proof.
Recall that Fμ is a right-continuous, non-decreasing function. Let u>0 be a jump point for Fμ, that is, Fμ(u−)<Fμ(u). Let {yn} be a sequence where Fμ is continuous, that decreases and converges to u. (Recall that the points of discontinuity of Fμ is countable.) Choose a sequence {xn} of points of continuity of Fμ, that converges to u from the left. Synchronized with {xn}, choose a sequence {αn}⊂[1/4,3/4], such that the convex combination (1−αn)xn+αnyn is to the right of u and is a point of continuity of Fμ. By compactness, we may assume that {αn} converges to an α∈[1/4,3/4]. So we have limk→∞Gk(t)=Fμ(t) for every t∈{xn,yn,(1−αn)xn+αnyn,n∈N}. The convexity of the functions tβGk(t), in the limit gives
((1−αn)xn+αnyn)βFμ((1−αn)xn+αnyn)≤(1−αn)xβnFμ(xn)+αnyβnFμ(yn), |
for each n∈N. Letting n approach infinity, the right-continuity of Fμ, shows that
uβFμ(u)≤(1−α)uβFμ(u−)+αuβFμ(u). |
Using that α≠1, and u>0, gives Fμ(u−)≥Fμ(u), which is a contradiction. Therefore, Fμ is continuous on (0,∞).
Now we show necessity. Suppose μ is a β-convex measure, we prove that F is CM. First, by (2.6) and [4, Theorem A.5.2], we have
f′(x)=∫(0,∞)e−xtFμ(t)dt−x∫(0,∞)e−xttFμ(t)dt |
and
f″(x)=−2∫(0,∞)e−xttFμ(t)dt+x∫(0,∞)e−xtt2Fμ(t)dt. |
To simplify the notation, denote
An(x):=∫(0,∞)e−xttnFμ(t)dtandBm(x):=∫(0,∞)e−xttmd(tβFμ(t)). |
It is not difficult to see that both are well defined. With that notation, we can rewrite
f(x)x=A0(x),f′(x)=A0(x)−xA1(x),xf″(x)=−2xA1(x)+x2A2(x). |
By (2.6) and Lemma 3, using Fubini's theorem, we have
xA1(x)=x∫(0,∞)e−xttFμ(t)dt=−∫(0,∞)tFμ(t)d(e−xt)=−tFμ(t)e−xt|∞t=0+∫(0,∞)e−xtd(tβFμ(t)t1−β)=∫(0,∞)e−xtt1−βd(tβFμ(t))+(1−β)∫(0,∞)e−xtFμ(t)dt=B1−β(x)+(1−β)A0(x) |
and
x2A2(x)=x2∫(0,∞)e−xtt2Fμ(t)dt=−xt2Fμ(t)e−xt|∞t=0+x∫(0,∞)e−xtd(tβFμ(t)t2−β)=(2−β)x∫(0,∞)e−xttFμ(t)dt+x∫(0,∞)e−xtt2−βd(tβFμ(t))=(2−β)xA1(x)+x∫(0,∞)e−xtt2−β(tβFμ(t))′+dt=(2−β)xA1(x)−t2−β(tβFμ(t))′+e−xt|∞t=0+∫(0,∞)e−xtd(t2−β(tβFμ(t))′+)=(2−β)xA1(x)+∫(0,∞)e−xtt2−βd(tβFμ(t))′++(2−β)∫(0,∞)e−xtt1−β(tβFμ(t))′+dt=(2−β)xA1(x)+∫(0,∞)e−xtt2−βd(tβFμ(t))′++(2−β)B1−β(x). |
(Note that the above equations also hold if μ is β-concave.) To summarize, we have
xA1(x)=B1−β(x)+(1−β)A0(x),x2A2(x)=2(2−β)B1−β(x)+(2−β)(1−β)A0(x)+∫(0,∞)e−xtt2−βd(tβFμ(t))′+. |
Therefore, it can be shown that
F(x)=β(β−1)f(x)x−2(β−1)f′(x)+xf″(x)=β(β−1)A0(x)−2(β−1)(A0(x)−xA1(x))−2xA1(x)+x2A2(x)=∫(0,∞)e−xtt2−βd(tβFμ(t))′+. |
The convexity of tβFμ(t) implies that (tβFμ(t))′+ is right-continuous and non-decreasing. Thus, the last integral is the Laplace transform of t2−βd(tβFμ(t))′+ on (0,∞). By the Bernstein representation theorem, F(x) is CM.
b) The proof is very much analogous to the proof for part a), so we will only address the differences.
For sufficiency, suppose −F is CM. Define Gk(t) as (3.2). Without any further assumptions, analogously to (3.3), we have
(tβGk(t))″=tβ−2kxk(−1)k−1F(k−1)(x)|x=k/t. |
As −F is CM, we know that tβGk(t) is concave. Analogous proof by contradiction applies to verify the continuity of Fμ. Therefore, the measure μ is β-concave, as Gk(t) converges to Fμ(t) for all t>0 and Gk is concave for all k≥2.
For necessity, suppose μ is β-concave, we prove −F is CM. Using the notation An and Bm from part a), we have
f(x)x=A0(x),f′(x)=A0(x)−xA1(x),xf″(x)=−2xA1(x)+x2A2(x), |
xA1(x)=B1−β(x)+(1−β)A0(x), |
x2A2(x)=2(2−β)B1−β(x)+(2−β)(1−β)A0(x)+∫(0,∞)e−xtt2−βd(tβFμ(t))′+. |
Thus, we obtain
−F(x)=∫(0,∞)e−xtt2−βd(−(tβFμ(t))′+). |
The concavity of tβFμ(t) implies that −(tβFμ(t))′+ is right-continuous and non-decreasing. The last integral is the Laplace transform of t2−βd(−(tβFμ(t))′+) on (0,∞). Hence, −F is a CM function.
Theorem 7. Suppose g is BF with Lévy triplet (a,b,ν). Consider the function
G(x):=β(β−1)g(x)x−2(β−1)g′(x)+xg″(x)−β(β−1)ax−(β−1)(β−2)b. | (3.4) |
a) The measure ν has β-convex tail, if and only if G is CM.
b) The measure ν has β-concave tail, if and only if −G is CM.
Proof. Without loss of generality, we can assume a=b=0. By (2.12) we have
g(x)=x∫(0,∞)e−xtˉν(t)dt. | (3.5) |
a) We show sufficiency first. Suppose G is CM. Anticipating the use of the inversion formula in Theorem 5, define
Gk(t):=Lk(g(x)x;t)=(−1)kxk+1(g(x)x)(k)|x=k/t, | (3.6) |
where the operator Lk is defined in (2.3). We claim that tβGk(t) is convex on (0,∞) for every k≥1. Notice that by (2.5),
G′k(t)=(−1)kddx(xk+1(g(x)x)(k))|x=k/t(−kt2)=(−1)k+11kxk+2g(x)(k+1)|x=k/t, |
and
G″k(t)=(−1)k+11kddx(xk+2g(x)(k+1))|x=k/t(−kt2)=(−1)k+21k2xk+3(xg(x))(k+2)|x=k/t. |
So we have
(tβGk(t))″=tβ−2(β(β−1)Gk(t)+2βtGk(t)′+t2Gk(t)″)=tβ−2xk+1(−1)kG(k)(x)|x=k/t. |
As G is CM, we know (−1)kG(k)(x)≥0 for all x>0 and k≥1, which implies tβGk(t) is convex. This closes the claim.
Let x and y belong to the interval (0,∞) and let the function ˉν be continuous at these points as well as at the convex combination (1−α)x+αy, where α is in [0,1]. By Theorem 5, as k approaches infinity, the limit of Gk equals ˉν(t) for any t taken from the set {x,y,(1−α)x+αy}. Furthermore, since tβGk(t) is convex for all k≥1, we have the inequality:
((1−α)x+αy)βˉν((1−α)x+αy)≤(1−α)xβˉν(x)+αyβˉν(y). |
To conclude the proof, we need to establish the continuity of ˉν on (0,∞).
Recall that ˉν is non-increasing and right-continuous. Suppose there exists a point u>0, where the function exhibits a jump, that is ˉν(u−)>ˉν(u). Let {yn} be a sequence where ˉν is continuous, that decreases and converges to u. Choose a sequence {xn} of points of continuity of ˉν that converges to u from the left. Together with it, choose a sequence {αn} in the interval [1/4,3/4], such that the convex combination (1−αn)xn+αnyn is to the left of u and is a point of continuity of ˉν. By compactness, we may assume that {αn} converges to an α in [1/4,3/4]. Thus, we can conclude that the limit of limk→∞Gk(t) equals ˉν for every t in the set {xn,yn,(1−αn)xn+αnyn,n∈N}. Given the convexity of the functions tβGk(t) for every k≥1 and each n∈N, we obtain:
((1−αn)xn+αnyn)βˉν((1−αn)xn+αnyn)≤(1−αn)xβnˉν(xn)+αnyβnˉν(yn). |
Letting n approach infinity, the right-continuity of ˉν implies that
uβˉν(u−)≤(1−α)uβˉν(u−)+αuβˉν(u). |
Using that α≠0 and u>0, we arrive at ˉν(u−)≤ˉν(u), which is a contradiction. Therefore, ˉν is continuous on (0,∞).
Now we show necessity. Suppose measure ν has β-convex tail. As a result, function tβˉν(t) is convex and s1−βˉν(1/s) is also convex by Corollary 1. We prove that G is CM. By (3.5) and change of variable s=1/t,
g(x)x=∫(0,∞)e−xtˉν(t)dt=∫(0,∞)e−x/ss−2ˉν(1/s)ds. |
Therefore, by [4, Theorem A.5.2], we have
g′(x)=∫(0,∞)e−x/ss−2ˉν(1/s)ds−x∫(0,∞)e−x/ss−3ˉν(1/s)ds,g′(x)=−2∫(0,∞)e−x/ss−3ˉν(1/s)ds+x∫(0,∞)e−x/ss−4ˉν(1/s)ds, |
To simplify the notation, denote
Cn(x)=∫(0,∞)e−x/ss−2−nˉν(1/s)ds,Dm(x)=∫(0,∞)e−x/ss−2+md(s1−βˉν(1/s)). |
With these notations, we can rewrite
g(x)x=C0(x),g′(x)=C0(x)−xC1(x),xg″(x)=−2xC1(x)+x2C2(x). |
By (2.13) and Lemma 4, we have
xC1(x)=x∫(0,∞)e−x/ss−3ˉν(1/s)ds=∫(0,∞)s−1ˉν(1/s)d(e−x/s)=s−1ˉν(1/s)e−x/s|∞s=0−∫(0,∞)e−x/sd(s1−βˉν(1/s)sβ−2)=−∫(0,∞)e−x/ssβ−2d(s1−βˉν(1/s))+(2−β)∫(0,∞)e−x/ss−2ˉν(1/s)ds=−Dβ(x)+(2−β)C0(x), |
and
x2C2(x)=x2∫(0,∞)e−x/ss−4ˉν(1/s)ds=xs−2ˉν(1/s)e−x/s|∞s=0−x∫(0,∞)e−x/sd(s1−βˉν(1/s)sβ−3)=(3−β)x∫(0,∞)e−x/ss−3ˉν(1/s)ds−x∫(0,∞)e−x/ssβ−3d(s1−βˉν(1/s))=(3−β)xC1(x)−x∫(0,∞)e−x/ssβ−3(s1−βˉν(1/s))′+ds=(3−β)xC1(x)−sβ−1(s1−βˉν(1/s))′e−x/s|∞s=0+∫(0,∞)e−x/sd(sβ−1(s1−βˉν(1/s))′+)=(3−β)xC1(x)+∫(0,∞)e−x/ssβ−1d(s1−βˉν(1/s))′++(β−1)∫(0,∞)e−x/ssβ−2d(s1−βˉν(1/s))=(3−β)xC1(x)+∫(0,∞)e−x/ssβ−1d(s1−βˉν(1/s))′++(β−1)Dβ. |
(Note that the above equations also hold if ν has β-concave tail.) To summarize,
xC1(x)=−Dβ(x)+(2−β)C0(x),x2C2(x)=2(β−2)Dβ+(3−β)(2−β)C0(x)+∫(0,∞)e−x/ssβ−1d(s1−βˉν(1/s))′+. |
Therefore, it can be shown that
G(x)=β(β−1)g(x)x−2(β−1)g′(x)+xg″(x)=β(β−1)C0(x)−2(β−1)(C0(x)−xC1(x))−2xC1(x)+x2C2(x)=∫(0,∞)e−x/ssβ−1d(s1−βˉν(1/s))′+. |
As s1−βˉν(1/s) is convex, (s1−βˉν(1/s))′+ is non-decreasing. It defines a Radon measure. One can see G is CM by definition.
b) The proof is very much analogous to the proof for part a), so we will only address the difference.
For sufficiency, suppose −G is CM. Define Gk as (3.6). Without any further assumption,
(tβGk(t))″=tβ−2xk+1(−1)kG(k)(x)|x=k/t. |
As −G is CM, we know that tβGk(t) is concave. Analogous proof by contradiction applies to verify the continuity of ˉν(t). As Gk(t) converges to ˉν(t) for all t>0, and as Gk has β-concave tail for all k≥1, the tail of ν is β-concave.
To show the necessity, suppose that the tail of ν is β-concave, we prove −G is CM. Following the notation Cn and Dm in part a), we also have
g(x)x=C0(x),g′(x)=C0(x)−xC1(x),xg″(x)=−2xC1(x)+x2C2(x), |
xC1(x)=−Dβ(x)+(2−β)C0(x), |
x2C2(x)=(3−β)xC1(x)+∫(0,∞)e−x/ssβ−1d(s1−βˉν(1/s))′++(β−1)Dβ. |
Thus, we obtain
−G(x)=∫(0,∞)e−x/ssβ−1d(−(s1−βˉν(1/s))′). |
As sβˉν(s) is concave, s1−βˉν(1/s) is concave by Corollary 1, implying −(s1−βˉν(1/s))′+ is non-decreasing. It defines a Radon measure and −G is CM by definition.
This section contains several corollaries of the main results.
Corollary 3. The CM function F in Theorem 6 has the representation
F(x)=∫∞0e−xsr′β(s)ds, | (4.1) |
where r′β(s)=s2−β(sβFμ(s))″ for almost all s>0.
Proof. We omit the proof since it is almost identical to the one for the next corollary. One just needs to replace ˉν(s) with Fμ(s) and ν(ds) with μ(ds).
Corollary 4. The CM function G in Theorem 7 has the representation
G(x)=∫∞0e−xsr′β(s)ds, | (4.2) |
where r′β(s)=s2−β(sβˉν(s))″ for almost all s>0.
Proof. Using Fubini's theorem, several times, function G can be rewritten as
G(x)=β(β−1)g(x)−a−bxx−2(β−1)(g′(x)−b)+xg″(x)=β(β−1)∫∞0e−xtˉν(t)dt−2(β−1)∫(0,∞)e−xttν(dt)−x∫(0,∞)e−xtt2ν(dt)=β(β−1)x∫∞0ˉν(t)(∫∞te−xsds)dt−2(β−1)x∫(0,∞)(∫∞te−xsds)tν(dt)−x∫(0,∞)e−xtt2ν(dt)=β(β−1)x∫∞0e−xs(∫s0ˉν(t)dt)ds−2(β−1)x∫(0,∞)e−xs(∫s0tν(dt))ds−x∫(0,∞)e−xss2ν(ds)=x∫∞0e−xsρβ(ds), | (4.3) |
where
ρβ(ds):=[β(β−1)∫s0ˉν(t)dt−2(β−1)∫s0tν(dt)]ds−s2ν(ds)=[β(β−1)∫s0ˉν(t)dt−2(β−1)(∫s0ˉν(t)dt−sˉν(s))]ds−s2ν(ds)=(1−β)[(2−β)∫s0ˉν(t)dt−2sˉν(s)]ds−s2ν(ds). | (4.4) |
Comparing (4.3) with (2.6) we make the following observation: G (resp. −G) is CM precisely when ρβ (resp. −ρβ) has a non-negative, non-decreasing density function rβ. In that case, solving for ν(ds) in (4.4) shows that necessarily ν has a density functionand and without loss of generality we may assume that it is of the form m(s)/sβ+1,s>0. Substituting it in (4.4), we obtain
rβ(s)=(1−β)[(2−β)∫s0ˉν(t)dt−2sˉν(s)]−s1−βm(s)=(1−β)(2−β)∫s0ˉν(t)dt−(2−β)sˉν(s)+s2−β(sβˉν(s))′+ | (4.5) |
=(1−β)(2−β)∫s0ˉν(t)dt−(1−β)sˉν(s)+s3−β(sβ−1ˉν(s))′+. | (4.6) |
Using that sβˉν(s) is convex (resp. concave), the derivative of rβ exists almost everywhere and differentiating (4.5) gives
r′β(s)=s2−β(sβˉν(s))″. | (4.7) |
Using L'Hopital's rule, (2.13), Corollary 2, and (4.6), one sees that
lims→0+rβ(s)=0=lims→∞e−xsrβ(s),x>0. |
An application of Fubini's theorem finally gives
G(x)=x∫∞0e−xsρβ(ds)=∫∞0e−xsr′β(s)ds, | (4.8) |
which completing the proof.
Corollary 5. The function G, defined in (3.4), can never be a Bernstein function.
Proof. We use the notation and representations from the proof of Corollary 4. Comparing (4.3) with (2.12) we observe that: G is a BF precisely when ρβ has a non-negative, non-increasing density function rβ. Assuming the latter, then (4.7) shows that ν has β-concave tail. Then, Theorem 7, part b) shows that −G is CM, which is a contradiction.
A similar proof shows the next analogous corollary.
Corollary 6. The function F, defined in (3.1), can never be a BF.
Standard facts about BFs imply that if xG(x) is a BF, then G is completely monotone.
The next corollaries deal with special cases of the main results. Some of them re-derive several of the results in [9].
Corollary 7. Suppose f is CM with measure μ. Then, μ is harmonically convex precisely when f(x)−xf′(x) is CM.
Proof. By Theorem 6 part a), with β=1, μ is harmonically convex precisely when xf″(x) is CM. We show this condition is equivalent to f(x)−xf′(x) being CM. If f(x)−xf′(x) is CM, then
xf″(x)=−(f(x)−xf′(x))′ |
is CM. Conversely, if xf″(x) is CM, then, to see f(x)−xf′(x) is CM, it suffices to show its non-negativity. This is trivial, because f(x)≥0 and f′(x)≤0 for all x>0.
Corollary 8. Suppose f is CM with measure μ. Then, μ is convex precisely when x(f(x)−μ({0})) is CM.
Proof. By Theorem 6 part a), applied to the shifted function f(x)−μ({0}) with β=0, the measure μ is convex precisely when 2f′(x)+xf″(x) is CM. We show this condition is equivalent to x(f(x)−μ({0})) being CM. If x(f(x)−μ({0})) is CM, then
2f′(x)+xf″(x)=(x(f(x)−μ({0})))″ |
is completely monotone.
Conversely, suppose 2f′(x)+xf″(x) is CM. To see x(f(x)−μ({0})) is CM, we only have to show
x(f(x)−μ({0}))≥0andxf′(x)+f(x)−μ({0})≤0. |
The first inequality holds because f(x)−μ({0})≥0. For the second inequality, as 2f′(x)+xf″(x)≥0, we know xf′(x)+f(x)−μ({0}) is non-decreasing. By (2.7), we obtain
limx→∞xf′(x)+f(x)−μ({0})=0. |
The second inequality follows from here.
Corollary 9. Suppose f is CM with measure μ. Then, μ is harmonically concave precisely when f(x)=μ({0}).
Proof. Consider the shifted function f(x)−μ({0}). By Theorem 6 part b), with β=1, the measure μ is harmonically concave precisely when −xf″(x) is CM. We show this condition is equivalent to f(x)=μ({0}). If f(x)=μ({0}), then −xf″(x)=0, which is CM. Conversely, if −xf″(x) is CM, then so is −xf″(x)(1/x)=−f″(x). Thus, we obtain f″(x)≤0. Notice that f″(x)≥0, because f is CM. Therefore, we have f″(x)=0, and f(x)=μ({0}).
Corollary 10. Suppose f is CM with measure μ. Then, μ is concave precisely when f(x)+xf′(x) is CM.
Proof. Without loss of generality, we can assume μ has no mass at zero. By Theorem 6 part b), with β=0, the measure μ is concave, if and only if −2xf′(x)−xf″(x) is CM. We show this condition is equivalent to f(x)+xf′(x) being CM. If f(x)+xf′(x) is CM, then
−2xf′(x)−xf″(x)=−(f(x)+xf′(x))′ |
is CM. Conversely, if −2xf′(x)−xf″(x) is CM, to see f(x)+xf′(x) is CM, it suffices to show it is non-negative. As its derivative is non-positive, f(x)+xf′(x) is non-increasing. By (2.7), we obtain
limx→∞f(x)+xf′(x)=0. |
So f(x)+xf′(x)≥0. This completes the proof.
Corollary 11. Suppose g is a Bernstein function with Lévy triplet (a,b,ν). Then, ν has harmonically convex tail precisely when g(x)=a+bx.
Proof. By Theorem 7 part a), applied to the shifted function g(x)−a−bx with β=1, the measure ν has harmonically convex tail, if and only if xg″(x) is completely monotone. We show this condition is equivalent to g(x)=a+bx. If g(x)=a+bx, then xg″(x)=0, which is CM. Conversely, if xg″(x) is CM, then so is xf″(x)(1/x)=f″(x), that is g″(x)≥0. Because g is a Bernstein function, g″(x)≤0. Thus, we obtain g″(x)=0, which implies g(x)=a+bx.
Corollary 12. Suppose g is a Bernstein function with Lévy triplet (a,b,ν). Then, ν has convex tail precisely when g(x)+xg′(x) is a Bernstein function.
Proof. By Theorem 7 part a), applied to the shifted BF g(x)−a−bx with β=0, the measure ν has convex tail precisely when 2g′(x)+xg″(x) is CM. We show this condition is equivalent to g(x)+xg′(x) being a Bernstein function. If g(x)+xg′(x) is a Bernstein function, then
2g′(x)+xg″(x)=(g(x)+xg′(x))′, |
is CM. (Note that 2g′(x)+xg″(x)≥0.) Conversely, if 2g′(x)+xg″(x) is CM, then it suffices to show g(x)+xg′(x)≥0 to see that g(x)+xg′(x) is a Bernstein function. This is trivial, because g(x)≥0 and g′(x)≥0.
Corollary 13. Suppose g is a Bernstein function with Lévy triplet (a,b,ν). Then, ν has harmonically concave tail precisely when g(x)−xg′(x) is a Bernstein function.
Proof. By Theorem 7 part b), applied to the shifted BF g(x)−a−bx with β=1, the measure ν has harmonically concave tail, if and only if −xg″(x) is CM. We show this condition is equivalent to g(x)−xg′(x) being a Bernstein function. If g(x)−xg′(x) is a Bernstein function, then
−xg″(x)=(g(x)−xg′(x))′, |
is CM. Conversely, if −xg″(x) is CM, then, to show g(x)−xg′(x) is a Bernstein function, it suffices to show it is non-negative. As its derivative is non-negative, g(x)−xg′(x) is non-decreasing. Noticing that limx→0xg′(x)=0, see [8, (2.11)], we obtain
limx→0g(x)−xg′(x)=a≥0. |
So g(x)−xg′(x)≥0, and this completes the proof.
Corollary 14. Suppose g is a Bernstein function with Lévy triplet (a,b,ν). Then, ν has concave tail precisely when g(x)=a+bx.
Proof. Consider the shifted BF g(x)−a−bx. By Theorem 7 part b), with β=0, the measure ν has concave tail precisely when −2g′(x)−xg″(x) is CM. We show this condition is equivalent to g(x)=0. If g(x)=0, then −2g′(x)−xg″(x)=0, which is CM. Conversely, if −2g′(x)−xg″(x) is CM, then, g(x)+xg′(x) is non-increasing, as
(g(x)+xg′(x))′=2g′(x)+xg″(x)≤0. |
Since
limx→0g(x)+xg′(x)=0, |
we obtain that g(x)+xg′(x)≤0, and thus its anti-derivative xg(x) is non-increasing. Because g(x) approaches zero as x approaches zero, we know xg(x)≤0. However, g is a Bernstein function, indicating g(x)=0. This concludes the proof.
The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.
The work of the first author was supported by the "Research Supporting Project number (RSP2024R162), King Saud University, Riyadh, Saudi Arabia". The work of the second author was partially supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada, project number RGPIN-2020-06425.
The authors declare no conflict of interest.
Lemma 5. Suppose f is continuous with bounded variation on (0,∞) and g is right-continuous with bounded variation on (0,∞). Then,
∫(0,∞)m(x)d(f(x)g(x))=∫(0,∞)m(x)f(x)dg(x)+∫(0,∞)m(x)g(x)df(x), | (A.1) |
where m is right-continuous and non-negative on (0,∞).
Proof. The sketch of proof is provided below.
Step 1: Show that (A.1) holds for increasing g on the closed interval [a,b]⊂(0,∞). For partition a=x0<x1<⋯<xn=b, we have
∫[a,b]m(x)d(f(x)g(x))=limmesh→0n−1∑i=0m(xi)[f(xi+1)g(xi+1)−f(xi)g(xi)]=limmesh→0n−1∑i=0m(xi)f(xi)[g(xi+1)−g(xi)]+limmesh→0n−1∑i=0m(xi)g(xi)[f(xi+1)−f(xi)]+limmesh→0n−1∑i=0m(xi)[f(xi+1)−f(xi)][g(xi+1)−g(xi)]=∫[a,b]m(x)f(x)dg(x)+∫[a,b]m(x)g(x)df(x)+limmesh→0n−1∑i=0m(xi)[f(xi+1)−f(xi)][g(xi+1)−g(xi)]. |
Notice that f is continuous, thus uniformly continuous, on [a,b]. For any ϵ>0, there exists δ>0, such that for any |t−s|<δ, we have |f(t)−f(s)|<ϵ. For any partition whose mesh is small, we obtain
limmesh→0|n−1∑i=0m(xi)[f(xi+1)−f(xi)][g(xi+1)−g(xi)]|≤ϵlimmesh→0n−1∑i=0|m(xi)[g(xi+1)−g(xi)]|=ϵ∫[a,b]m(x)dg(x). |
So this limit can be arbitrarily small, which indicates
∫[a,b]m(x)d(f(x)g(x))=∫[a,b]m(x)f(x)dg(x)+∫[a,b]m(x)g(x)df(x). |
Step 2: (A.1) holds for g with bounded variation on [a,b], as such g can be represented as the difference of two increasing functions.
Step 3: (A.1) holds on (0,∞), as the equation holds when taking the limit of a→0 and b→∞.
The following result is Theorem 6.2.2 in [3] or Theorem A.1 in [8]. It gives integration by parts for Lebesgue-Stieltjes integrals on finite intervals. It can be extended to the interval (0,∞) by taking limits.
Theorem 8. Let f,g:[a,b]→R be continuous and right-continuous functions, respectively, with bounded variation. Then
∫[a,b]fdg+∫[a,b]gdf=f(b)g(b)−f(a−)g(a−),∫(a,b]fdg+∫(a,b]gdf=f(b)g(b)−f(a)g(a),∫(a,b)fdg+∫(a,b)gdf=f(b−)g(b−)−f(a)g(a). |
The following lemma is a particular case of the change of variable formula for Lebesgue-Stieltjes integrals, see [11, Theorem 11a].
Lemma 6. Suppose f is continuous on (0,∞), and g has bounded variation on (0,∞). Then
∫(0,∞)f(x)dg(x)=−∫(0,∞)f(1/t)dg(1/t). |
Lemma 7. Suppose f(t)≥0 is non-increasing. If ∫(0,∞)f(t)dt<∞, then f(t) is o(1/t) as t approaches zero or infinity.
Lemma 8. Suppose f(t)≥0 is non-increasing. If ∫(0,1)f(t)d(tp)<∞ for some p>0. Then f(t) is o(1/tp) as t→0.
Proof. After a change of variable by s=tp, we obtain ∫(0,1)f(s1/p)d(s)<∞. Because f(s1/p) is non-increasing for any p>0, we conclude that f(s1/p) is o(1/s) as s→0. This implies f(t) is o(1/tp) as t→0.
Lemma 9. Suppose f(t)≥0 is non-increasing and g is strictly increasing with g(0)=0. If ∫(0,1)f(t)d(g(t))<∞, then f(t) is o(1/g(t)) as t→0.
Proof. As g is strictly increasing with g(0)=0, its inverse function g−1(t) is also strictly increasing with g−1(0)=0. Change of variable by setting t=g−1(s), we have f(g−1(s)) is non-increasing ∫(0,g(1))f(g−1(s))ds<∞. So f(g−1(s)) is o(1/s) as s→0, which implies limt→0f(t)g(t)=0.
[1] | W. Arendt, C. Batty, M. Hieber, F. Neubrander, Vector-valued Laplace transforms and Cauchy problems, Basel: Birkhäuser, 2011. http://dx.doi.org/10.1007/978-3-0348-0087-7 |
[2] |
S. Bridaa, W. Jedidi, H. Sendov, Generalized unimodality and subordinators, with applications to stable laws and to the Mittag-Leffler function, J. Theor. Probab., 37 (2024), 1–42. http://dx.doi.org/10.1007/s10959-023-01242-z doi: 10.1007/s10959-023-01242-z
![]() |
[3] | M. Carter, B. Brunt, The Lebesgue-Stieltjes integral: a practical introduction, New York: Springer-Verlag, 2000. http://dx.doi.org/10.1007/978-1-4612-1174-7 |
[4] | R. Durrett, Probability: theory and examples, 4 Eds., Cambridge: Cambridge University Press, 2010. |
[5] |
M. Merkle, Reciprocally convex functions, J. Math. Anal. Appl., 293 (2004), 210–218. http://dx.doi.org/10.1016/j.jmaa.2003.12.021 doi: 10.1016/j.jmaa.2003.12.021
![]() |
[6] | R. Rockafellar, Convex analysis, Princeton: Princeton University Press, 1970. |
[7] | R. Schilling, R. Song, Z. Vondracek, Bernstein functions theory and applications, 2 Eds., Berlin: De Gruyter, 2012. |
[8] |
H. Sendov, S. Shan, New representation theorems for completely monotone and Bernstein functions with convexity properties on their measures, J. Theor. Probab., 28 (2015), 1689–1725. http://dx.doi.org/10.1007/s10959-014-0557-9 doi: 10.1007/s10959-014-0557-9
![]() |
[9] | H. Sendov, S. Shan, Properties of completely monotone and Bernstein functions related to the shape of their measures, J. Convex Anal., 23 (2016), 981–1015. |
[10] |
T. Simon, On the unimodality of power transformations of positive stable densities, Math. Nachr., 285 (2012), 497–506. http://dx.doi.org/10.1002/mana.201000062 doi: 10.1002/mana.201000062
![]() |
[11] | D. Widder, The Laplace transform, Princeton: Princeton University Press, 1941. |
Property of μ | Characterization | Reference |
β-convex | −F is completely monotone | Theorem 6 a) |
β-concave | −F is completely monotone | Theorem 6 b) |
Property of ν | Characterization | Reference |
β-convex tail | −G is completely monotone | Theorem 7 a) |
β-concave tail | −G is completely monotone | Theorem 7 b) |