Ferreira–Simon's stochastic-order conjecture

What Lean checks

Complete proposed proof for every pair of real parameters in the source interval. The Poisson-series scale and strict sequence-norm comparison give a simultaneous coupling, both endpoint limits and the associated Mittag–Leffler inequalities. Independent internal review passed.

Date: 21 November 2026. Status: complete proposed proof, independently checked within the research campaign. The source audit records the prior-resolution search and its limits. The note is unrefereed.

Statement and source

Rui A. C. Ferreira and Thomas Simon, Mittag-Leffler functions and convex ordering, arXiv:2512.04940v1 (4 December 2025), Remark 7(b), printed page 18, conjecture that

α↦MαΓ(1−α)\alpha\longmapsto \frac{M_\alpha}{\Gamma(1-\alpha)}

is non-increasing in the usual stochastic order on (0,1)(0,1). Here ZαZ_\alpha is the standard positive stable random variable with Laplace transform 𝔼e−tZα=e−tα\mathbb E e^{-tZ_\alpha}=e^{-t^\alpha}, and Mα=Zα−αM_\alpha=Z_\alpha^{-\alpha} has moment-generating function EαE_\alpha, the classical Mittag–Leffler function. Primary source: https://arxiv.org/pdf/2512.04940 .

The usual stochastic order is X≤stYX\leq_{\mathrm{st}}Y if ℙ(X>u)≤ℙ(Y>u)\mathbb P(X>u)\leq\mathbb P(Y>u) for every real uu. We prove the conjecture by a simultaneous almost surely strictly decreasing coupling.

Theorem

There is a family (Xα)0<α<1(X_\alpha)_{0<\alpha<1} on one probability space such that

Xα=dMαΓ(1−α)X_\alpha\stackrel d=\frac{M_\alpha}{\Gamma(1-\alpha)}

for every α\alpha, and, on one probability-one event,

Xα>Xβ(0<α<β<1).X_\alpha>X_\beta\qquad(0<\alpha<\beta<1).

Moreover, on that event, Xα→T1X_\alpha\to T_1 as α↓0\alpha\downarrow0, where T1T_1 has the unit exponential distribution, and Xα→0X_\alpha\to0 as α↑1\alpha\uparrow1.

Proof

Let Tn=ξ1+⋯+ξnT_n=\xi_1+\cdots+\xi_n, where the ξj\xi_j are independent unit exponential random variables. Thus 0<T1<T2<⋯0<T_1<T_2<\cdots are the arrival times of a Poisson process of rate one. By the strong law, Tn/n→1T_n/n\to1 almost surely. Work on this probability-one event. For every 0<α<10<\alpha<1, the series

Wα=∑n=1∞Tn−1/αW_\alpha=\sum_{n=1}^{\infty}T_n^{-1/\alpha}

converges: eventually Tn≥n/2T_n\geq n/2, and 1/α>11/\alpha>1. The same event works for all parameters. Define

Xα=Wα−α.X_\alpha=W_\alpha^{-\alpha}.

First identify the marginal law. The Laplace functional of a rate-one Poisson process gives, for t≥0t\geq0,

𝔼e−tWα=exp[−∫0∞(1−e−ts−1/α)ds].\mathbb E e^{-tW_\alpha} =\exp\left[-\int_0^\infty(1-e^{-t s^{-1/\alpha}})\,ds\right].

For completeness, this functional follows first for nonnegative step functions with bounded support from independence of Poisson counts: 𝔼e−cN(A)=exp(|A|(e−c−1))\mathbb E e^{-cN(A)}=\exp(|A|(e^{-c}-1)). Approximation by simple functions, followed by monotone convergence, gives the displayed expression. The integral is finite because its integrand is bounded by one near zero and by ts−1/αt s^{-1/\alpha} near infinity.

For t>0t>0, substitute u=ts−1/αu=t s^{-1/\alpha} and integrate by parts:

∫0∞(1−e−ts−1/α)ds=αtα∫0∞(1−e−u)u−α−1du=tα∫0∞e−uu−αdu=Γ(1−α)tα.\begin{aligned} \int_0^\infty(1-e^{-t s^{-1/\alpha}})\,ds &=\alpha t^\alpha\int_0^\infty(1-e^{-u})u^{-\alpha-1}\,du\\ &=t^\alpha\int_0^\infty e^{-u}u^{-\alpha}\,du\\ &=\Gamma(1-\alpha)t^\alpha. \end{aligned}

Both boundary terms vanish since 0<α<10<\alpha<1. Uniqueness of Laplace transforms now gives

Wα=dΓ(1−α)1/αZα,Xα=dZα−αΓ(1−α).W_\alpha\stackrel d=\Gamma(1-\alpha)^{1/\alpha}Z_\alpha, \qquad X_\alpha\stackrel d=\frac{Z_\alpha^{-\alpha}}{\Gamma(1-\alpha)}.

It remains to compare the coupled variables. Set an=Tn−1>0a_n=T_n^{-1}>0. Then

Xα=1∥a∥1/α,∥a∥p=(∑n=1∞anp)1/p.X_\alpha=\frac{1}{\|a\|_{1/\alpha}}, \qquad \|a\|_p=\left(\sum_{n=1}^\infty a_n^p\right)^{1/p}.

For 1<p<q1<p<q, put A=∥a∥pA=\|a\|_p. Every an/Aa_n/A lies strictly between zero and one because infinitely many terms are positive. Hence

∑n(an/A)q<∑n(an/A)p=1,\sum_n(a_n/A)^q<\sum_n(a_n/A)^p=1,

and therefore ∥a∥q<∥a∥p\|a\|_q<\|a\|_p. If α<β\alpha<\beta, then 1/β<1/α1/\beta<1/\alpha, so Xα>XβX_\alpha>X_\beta. This proves the conjecture for all parameter pairs, with no parameter discretization.

For the first endpoint, write a1=maxnana_1=\max_n a_n. Fix p0>1p_0>1. For p≥p0p\geq p_0,

a1≤∥a∥p≤a1(∑n(an/a1)p0)1/p.a_1\leq\|a\|_p \leq a_1\left(\sum_n(a_n/a_1)^{p_0}\right)^{1/p}.

Thus ∥a∥p→a1\|a\|_p\to a_1 as p→∞p\to\infty, giving Xα→T1X_\alpha\to T_1 as α↓0\alpha\downarrow0.

For the second endpoint, Tn/n→1T_n/n\to1 implies ∑nan=∞\sum_n a_n=\infty. For each integer NN,

liminfp↓1∥a∥p≥limp↓1(∑n=1Nanp)1/p=∑n=1Nan.\liminf_{p\downarrow1}\|a\|_p\geq\lim_{p\downarrow1}\left(\sum_{n=1}^N a_n^p\right)^{1/p}=\sum_{n=1}^N a_n.

Letting N→∞N\to\infty proves ∥a∥p→∞\|a\|_p\to\infty, hence Xα→0X_\alpha\to0 as α↑1\alpha\uparrow1. This completes the proof.

Mittag–Leffler consequence

The Laplace transform of MαM_\alpha is Eα(−x)E_\alpha(-x). For x>0x>0, the coupled inequalities and the strict decrease of u↦e−xuu\mapsto e^{-xu} give

11+x<Eα(−xΓ(1−α))<Eβ(−xΓ(1−β))<1(0<α<β<1).\frac{1}{1+x} <E_\alpha\left(-\frac{x}{\Gamma(1-\alpha)}\right) <E_\beta\left(-\frac{x}{\Gamma(1-\beta)}\right) <1 \qquad(0<\alpha<\beta<1).

The first inequality follows also directly from Xα<T1X_\alpha<T_1, since ∥a∥1/α>a1\|a\|_{1/\alpha}>a_1. At x=0x=0, all four terms equal one. These are exactly the inequalities proposed as a consequence in Remark 7(b), with strict inequalities for positive xx.

Scope and attribution

The Poisson series representation of positive stable variables is classical. No novelty is claimed for that representation, the Poisson Laplace functional, or monotonicity of sequence norms. The proposed contribution is their application to the full stochastic-order assertion explicitly left open by Ferreira and Simon in December 2025. No earlier resolution was found in the sources checked in the separate priority audit; that bounded search does not certify exclusive priority.

This proof is analytic and probabilistic. It does not depend on numerical tests, computer algebra, or Lean verification.