Simon's stable-infimum conjecture

What Lean checks

Complete proposed proof of the full parameterwise convex order, including the Brownian endpoint. The independent review checks the prior Mellin transform, increasing logarithmic Lévy measures, exponential integrability and conditional Jensen argument. No numerical or finite-case inference is used.

Status and original statement

This is a complete proposed proof of the convex-order conjecture in Remark 6 of Thomas Simon, Comparing Fréchet and positive stable laws, Electronic Journal of Probability 19 (2014), no. 16, DOI 10.1214/EJP.v19-3058. The remark is on printed page 13 of the final journal article, and on printed pages 13–14 of arXiv:1310.1888v2, dated 27 January 2014. Both versions were checked directly. Independent mathematical audit passed; the separate prior-resolution search remains the basis for the bounded priority assessment. No external peer review is claimed.

For each α∈[1/2,1)\alpha\in[1/2,1), let X(α)X^{(\alpha)} be the spectrally negative strictly 1/α1/\alpha-stable Lévy process with normalization

𝔼eλXt(α)=etλ1/α(t,λ≥0),\mathbb E e^{\lambda X_t^{(\alpha)}}=e^{t\lambda^{1/\alpha}} \qquad(t,\lambda\ge0),

and put

Iα=−inf0≤t≤1Xt(α),Vα=Γ(1+α)Iα.I_\alpha=-\inf_{0\le t\le1}X_t^{(\alpha)}, \qquad V_\alpha=\Gamma(1+\alpha)I_\alpha.

The conjecture is

Vβ≼cxVα(12≤β<α<1).V_\beta\preceq_{\mathrm{cx}}V_\alpha \qquad\left(\frac12\le\beta<\alpha<1\right). \tag{1}

Here convex order means 𝔼ϕ(Vβ)≤𝔼ϕ(Vα)\mathbb E\phi(V_\beta)\le\mathbb E\phi(V_\alpha) for every convex function for which the expectations exist. This is the full parameter range of the stated conjecture. The direction is opposite to the corresponding ordering of the suprema. The proof below gives the stronger factorization

Vα=dVβRβ,α,Rβ,α>0,𝔼Rβ,α=1,V_\alpha\ \stackrel d=\ V_\beta R_{\beta,\alpha}, \qquad R_{\beta,\alpha}>0,\qquad \mathbb E R_{\beta,\alpha}=1, \tag{2}

where the factors on the right are independent and logRβ,α\log R_{\beta,\alpha} is infinitely divisible.

1. Mellin transform of the infimum

We use the established factorization recorded in Simon's Remark 6:

Iα=dMαYα,Mα=Zα−α,𝔼e−λZα=e−λα,I_\alpha\stackrel d=M_\alpha Y_\alpha, \qquad M_\alpha=Z_\alpha^{-\alpha},\qquad \mathbb E e^{-\lambda Z_\alpha}=e^{-\lambda^\alpha}, \tag{3}

with independent factors. For 1/2<α<11/2<\alpha<1, the density of YαY_\alpha is

fα(t)=−sin(π/α)πt1/α−2(1+t)t2/α−2t1/αcos(π/α)+1,t>0.f_\alpha(t)= -\frac{\sin(\pi/\alpha)}\pi \frac{t^{1/\alpha-2}(1+t)} {t^{2/\alpha}-2t^{1/\alpha}\cos(\pi/\alpha)+1}, \qquad t>0. \tag{4}

The factorization (3), including (4), is prior work, not a claim of this manuscript. The Mellin formulas (6) and (8) below also appear, in equivalent forms with reciprocal parameter, in Lemma 3.3 and Proposition 3.6 of P. Patie and T. Simon, Intertwining certain fractional derivatives, Potential Analysis 36 (2012), 569–587, arXiv:1008.4713. We supply their elementary calculation for completeness. The comparison of the resulting Lévy measures in Sections 2–3 is the step used here to resolve the conjecture. At α=1/2\alpha=1/2, the infimum has the same half-Gaussian law as M1/2M_{1/2}.

Set θ=2π−π/α∈(0,π)\theta=2\pi-\pi/\alpha\in(0,\pi). The elementary integral

Jθ(r)=∫0∞ur−1u2−2ucosθ+1du=πsin((1−r)(π−θ))sinθsin(πr)(0<r<2)J_\theta(r)=\int_0^\infty \frac{u^{r-1}}{u^2-2u\cos\theta+1}\,du =\frac{\pi\sin((1-r)(\pi-\theta))} {\sin\theta\sin(\pi r)} \quad(0\lt r<2) \tag{5}

is understood by continuity at r=1r=1. For completeness, partial fractions at the roots eiθ,e−iθe^{i\theta},e^{-i\theta}, followed by Euler's integral ∫0∞ur−1/(u+z)du=πzr−1/sin(πr)\int_0^\infty u^{r-1}/(u+z)\,du=\pi z^{r-1}/\sin(\pi r), proves it first for 0<r<10\lt r<1; the substitution u↦1/uu\mapsto1/u gives 1<r<21\lt r<2. Use principal powers with |argz|<π|\arg z|<\pi.

Substituting u=t1/αu=t^{1/\alpha} into (4) and applying (5) gives, for 1−1/α<s<1/α1-1/\alpha\lt s<1/\alpha,

𝔼Yαs=αsinθπ[Jθ(1−α+αs)+Jθ(1+αs)]=α[sin(π(1−α)(1−s))sin(πα(1−s))+sin(π(1−α)s)sin(παs)]=αsin(πα)sin(πs)sin(παs)sin(πα(1−s)).\begin{aligned} \mathbb E Y_\alpha^s &=\frac{\alpha\sin\theta}{\pi} \big[J_\theta(1-\alpha+\alpha s) +J_\theta(1+\alpha s)\big]\\ &=\alpha\left[ \frac{\sin(\pi(1-\alpha)(1-s))}{\sin(\pi\alpha(1-s))} +\frac{\sin(\pi(1-\alpha)s)}{\sin(\pi\alpha s)} \right]\\ &=\frac{\alpha\sin(\pi\alpha)\sin(\pi s)} {\sin(\pi\alpha s)\sin(\pi\alpha(1-s))}. \end{aligned} \tag{6}

Apparent singularities at s=0,1s=0,1 are removable. The elementary sine addition identity gives the last equality. Both endpoints of the stated interval are integrability boundaries in (4).

The stable fractional-moment formula is

𝔼Mαs=Γ(1+s)Γ(1+αs)(s>−1).\mathbb E M_\alpha^s =\frac{\Gamma(1+s)}{\Gamma(1+\alpha s)} \qquad(s>-1). \tag{7}

Multiplying (6) and (7), and applying the reflection and recurrence formulas for the gamma function, yields

𝔼Iαs=Γ(1−αs)Γ(1+α−αs)Γ(1−α+αs)Γ(1+α)Γ(1−α)Γ(2−s).\boxed{\displaystyle \mathbb E I_\alpha^s =\frac{ \Gamma(1-\alpha s) \Gamma(1+\alpha-\alpha s) \Gamma(1-\alpha+\alpha s)} {\Gamma(1+\alpha)\Gamma(1-\alpha)\Gamma(2-s)}.} \tag{8}

For 1/2<α<11/2<\alpha<1, formula (8) holds throughout the complex strip 1−1/α<ℜs<1/α1-1/\alpha<\Re s<1/\alpha, by analytic continuation from the real interval. At α=1/2\alpha=1/2, the duplication formula reduces its right side to Γ(1+s)/Γ(1+s/2)\Gamma(1+s)/\Gamma(1+s/2), so (8) holds there too, at least in the common strip −1<ℜs<2-1<\Re s<2. In particular,

𝔼Iα=1Γ(1+α),𝔼Vα=1.\mathbb E I_\alpha=\frac1{\Gamma(1+\alpha)}, \qquad \mathbb E V_\alpha=1. \tag{9}

2. The logarithm has increasing Lévy measures

Define the analytic cumulant Kα(s)=log𝔼VαsK_\alpha(s)=\log\mathbb E V_\alpha^s, choosing the logarithm equal to zero at s=0s=0. Formula (8) gives

Kα(s)=slogΓ(1+α)+logΓ(1−αs)+logΓ(1+α−αs)+logΓ(1−α+αs)−logΓ(1+α)−logΓ(1−α)−logΓ(2−s).\begin{aligned} K_\alpha(s)={}&s\log\Gamma(1+\alpha) +\log\Gamma(1-\alpha s) +\log\Gamma(1+\alpha-\alpha s)\\ &+\log\Gamma(1-\alpha+\alpha s) -\log\Gamma(1+\alpha) -\log\Gamma(1-\alpha) -\log\Gamma(2-s). \end{aligned} \tag{10}

We use the gamma integral, valid for a>0a>0 and ℜ(a+bs)>0\Re(a+bs)>0,

logΓ(a+bs)−logΓ(a)−bsψ(a)=∫0∞(e−bsu−1+bsu)e−auu(1−e−u)du.\log\Gamma(a+bs)-\log\Gamma(a)-bs\psi(a) =\int_0^\infty (e^{-bsu}-1+bsu)\frac{e^{-au}}{u(1-e^{-u})}\,du. \tag{11}

It also follows directly by differentiating twice in ss, using the integral for ψ′\psi', and fixing the value and first derivative at zero. Thus no infinite-divisibility hypothesis is being assumed in (11).

Apply (11) to each term of (10), and change variables t=αut=\alpha u in the three terms with slopes of magnitude α\alpha. This gives

Kα(s)=dαs+∫0∞(est−1−st)kα(t)tdt+∫0∞(e−st−1+st)hα(t)tdt,K_\alpha(s)=d_\alpha s +\int_0^\infty(e^{st}-1-st)\frac{k_\alpha(t)}t\,dt +\int_0^\infty(e^{-st}-1+st)\frac{h_\alpha(t)}t\,dt, \tag{12}

where dα=Kα′(0)d_\alpha=K_\alpha'(0) is finite and

kα(t)=1+e−tet/α−1−e−2t1−e−t,hα(t)=etet/α−1.\boxed{\displaystyle k_\alpha(t)=\frac{1+e^{-t}}{e^{t/\alpha}-1} -\frac{e^{-2t}}{1-e^{-t}}, \qquad h_\alpha(t)=\frac{e^t}{e^{t/\alpha}-1}.} \tag{13}

For every fixed t>0t>0, both functions are strictly increasing in α\alpha. Moreover,

k1/2(t)=1+e−te2t−1−e−2t1−e−t=0,k_{1/2}(t)= \frac{1+e^{-t}}{e^{2t}-1} -\frac{e^{-2t}}{1-e^{-t}}=0, \tag{14}

and hα(t)>0h_\alpha(t)>0. Consequently kα,hα≥0k_\alpha,h_\alpha\ge0 for every α∈[1/2,1)\alpha\in[1/2,1).

These functions define legitimate Lévy measures. Near zero,

kα(t)=2α−1t+O(1),hα(t)=αt+O(1).k_\alpha(t)=\frac{2\alpha-1}{t}+O(1), \qquad h_\alpha(t)=\frac\alpha t+O(1). \tag{15}

At infinity, kα(t)=O(e−t/α)k_\alpha(t)=O(e^{-t/\alpha}) and hα(t)=O(e−(1/α−1)t)h_\alpha(t)=O(e^{-(1/\alpha-1)t}). At the endpoint α=1/2\alpha=1/2, the first function vanishes identically, which is consistent with these estimates. It follows that the measure with density kα(t)/tk_\alpha(t)/t at positive jumps tt and density hα(t)/th_\alpha(t)/t at negative jumps −t-t has finite integral against 1∧t21\wedge t^2; its large jumps also have finite first absolute moment. The compensation ∓st\mp st over the whole positive half-line in (12) is therefore valid. Since 1/α>11/\alpha>1, its positive jumps have an exponential moment at parameter one.

For imaginary ss, formula (12) is the Lévy–Khintchine representation of logVα\log V_\alpha. The same integrals converge for every real ss in a neighborhood of [0,1][0,1].

3. Independent mean-one increments and convex order

Fix 1/2≤β<α<11/2\le\beta<\alpha<1. Both differences

kα(t)−kβ(t)>0,hα(t)−hβ(t)>0(t>0)k_\alpha(t)-k_\beta(t)>0, \qquad h_\alpha(t)-h_\beta(t)>0 \qquad(t>0) \tag{16}

are nonnegative Lévy densities after division by tt, with the same integrability properties just checked. Thus

Kα(s)−Kβ(s)K_\alpha(s)-K_\beta(s) \tag{17}

is the cumulant of an infinitely divisible real random variable Lβ,αL_{\beta,\alpha}. Specifically its drift in the compensation convention (12) is dα−dβd_\alpha-d_\beta, and its Lévy measure is the difference of the two Lévy measures in (13). The exponential moment at one is finite, and (9) gives

𝔼eLβ,α=exp(Kα(1)−Kβ(1))=1.\mathbb E e^{L_{\beta,\alpha}} =\exp(K_\alpha(1)-K_\beta(1))=1. \tag{18}

Take Lβ,αL_{\beta,\alpha} independent of VβV_\beta and let Rβ,α=eLβ,αR_{\beta,\alpha}=e^{L_{\beta,\alpha}}. Comparing characteristic functions of logarithms gives (2). The construction is a martingale coupling because

𝔼[VβRβ,α∣Vβ]=Vβ.\mathbb E[V_\beta R_{\beta,\alpha}\mid V_\beta]=V_\beta. \tag{19}

Conditional Jensen's inequality now proves (1). This argument covers every pair 1/2≤β<α<11/2\le\beta<\alpha<1 and every convex test function in the definition of convex order, so it is not a moment-only comparison or a restricted-parameter case. For convex functions with a possibly infinite positive expectation, subtract a supporting affine function first; monotone approximation or the extended conditional Jensen inequality then gives the same conclusion.

Scope and dependence

The only specific stable-process input is the previously established factorization (3)–(4), as recorded in the original source. The remaining steps are an explicit Mellin computation, the gamma integral, Lévy–Khintchine, and conditional Jensen. The proof does not rely on numerical evidence, a finite collection of moments, or Lean verification. The separate literature record determines whether this is eligible to be counted as a new resolution.