Ferreira–Simon's Wright log-concavity conjecture

What Lean checks

The actual infinite Wright series has a continuous beta-zero counterexample family, with admissible alpha approaching one. Strict compilation, the complete 25-declaration owned audit including eight private declarations, and independent stock-kernel replay passed. The positive-beta refinements in the written proof and the combined repository gate remain separate and pending.

Status: complete analytic disproof. Separate internal reviews by the root and algebra agents found no mathematical gap. See issue 100 and NOVELTY.md for the bounded prior-resolution audit. The proof gives intervals of counterexamples and does not infer a theorem from a finite numerical calculation.

1. Source and precise assertion

Ferreira and Simon define

ϕ(ρ,β,z)=∑n=0∞znn!Γ(β+ρn),ρ>−1.\phi(\rho,\beta,z) =\sum_{n=0}^{\infty}\frac{z^n}{n!\Gamma(\beta+\rho n)}, \qquad \rho>-1.

The reciprocal gamma function is interpreted by its entire extension, including its zeros at the nonpositive integers. For 0<α<10\lt \alpha\lt 1 and β≥0\beta\geq0, their random variable Mα,βM_{\alpha,\beta} has density

pα,β(x)=Γ(α+β)ϕ(−α,β,−x),x>0.p_{\alpha,\beta}(x) =\Gamma(\alpha+\beta)\phi(-\alpha,\beta,-x), \qquad x>0.

For 0<β<10\lt \beta\lt 1, their Theorem D defines α*(β)\alpha_*(\beta) as the unique root in (0,1)(0,1) of

1Γ(β−α)2−1Γ(β)Γ(β−2α)=0.\frac{1}{\Gamma(\beta-\alpha)^2} -\frac{1}{\Gamma(\beta)\Gamma(\beta-2\alpha)}=0. \tag{1}

Their proof in §5.1 shows that the left side is positive for 0<α<α*(β)0\lt \alpha\lt \alpha_*(\beta) and negative for α*(β)<α<1\alpha_*(\beta)\lt \alpha\lt 1. They set α*(0)=1\alpha_*(0)=1, and also set α*(β)=1\alpha_*(\beta)=1 for β≥1\beta\geq1.

The unnumbered conjecture in §6.4 asserts, for all admissible parameters,

α≤α*(β)⇔Mα,β has a log-concave density⇔t↦1Eα,α+β(−t) is convex on ℝ+.\alpha\leq\alpha_*(\beta) \quad\Longleftrightarrow\quad M_{\alpha,\beta}\text{ has a log-concave density} \quad\Longleftrightarrow\quad t\longmapsto\frac{1}{E_{\alpha,\alpha+\beta}(-t)} \text{ is convex on }\mathbb R_+. \tag{2}

Here Eα,γ(z)=∑n≥0zn/Γ(γ+αn)E_{\alpha,\gamma}(z)=\sum_{n\geq0}z^n/\Gamma(\gamma+\alpha n).

Source: R. A. C. Ferreira and T. Simon, On the log-concavity of the Wright function, Constructive Approximation 60 (2024), 309–338, DOI. The accessible primary arXiv v2 has Theorem D on printed page 5, its relevant proof on pages 15–16, and the conjecture on pages 23–24.

2. Counterexample families

Theorem. Both of the following statements hold.

  1. For every sufficiently small t>0t>0, the parameters

    α=1−t,β=0\alpha=1-t,\qquad\beta=0 \tag{3}

    satisfy α<α*(0)=1\alpha\lt \alpha_*(0)=1, but pα,0p_{\alpha,0} is not log concave.

  2. For every sufficiently small t>0t>0, the parameters

    α=1−3t,β=t\alpha=1-3t,\qquad\beta=t \tag{4}

    satisfy α<α*(β)\alpha\lt \alpha_*(\beta), but pα,βp_{\alpha,\beta} is not log concave.

In both cases the log-concavity inequality fails at x=1/2x=1/2. Thus the first implication in (2) is false, including at interior parameters with β>0\beta>0.

3. A locally uniform limit

Fix a>0a>0 and b≥0b\geq0. For t>0t>0 sufficiently small, put

ft(x)=ϕ(−(1−at),bt,−x)=∑n=0∞(−1)nxnn!Γ(−n+(an+b)t).f_t(x)=\phi(-(1-at),bt,-x) =\sum_{n=0}^{\infty} \frac{(-1)^n x^n}{n!\Gamma(-n+(an+b)t)}. \tag{5}

We prove that, on every closed disk |x|≤r<1|x|\leq r\lt 1, ft/tf_t/t and each of its fixed-order derivatives converge uniformly to those of

ha,b(x)=∑n=0∞(an+b)xn=b+(a−b)x(1−x)2.h_{a,b}(x)=\sum_{n=0}^{\infty}(an+b)x^n =\frac{b+(a-b)x}{(1-x)^2}. \tag{6}

Write un=(an+b)tu_n=(an+b)t. For n≥1n\geq1, the classical reflection formula (DLMF 5.5.3) gives

(−1)ntn!Γ(−n+un)=Γ(n+1−un)n!sin(πun)πt.\frac{(-1)^n}{t\,n!\Gamma(-n+u_n)} =\frac{\Gamma(n+1-u_n)}{n!} \frac{\sin(\pi u_n)}{\pi t}. \tag{7}

This formula also holds when −n+un-n+u_n is an integer, by continuity of 1/Γ1/\Gamma. Choose 0<t<1/(a+b)0\lt t\lt 1/(a+b). Since n≥1n\geq1,

0<un<n,1<n+1−un<n+1.0\lt u_n\lt n,\qquad 1\lt n+1-u_n\lt n+1.

Gamma is positive and log convex on the positive half-line; see the Bohr–Mollerup characterization. Its values at the endpoints of [1,n+1][1,n+1] are 11 and n!n!. Hence

0<Γ(n+1−un)≤n!.0\lt \Gamma(n+1-u_n)\leq n!.

Using |sinv|≤|v||\sin v|\leq|v| in (7), we obtain the bound

|(−1)ntn!Γ(−n+(an+b)t)|≤an+b,n≥1.\left|\frac{(-1)^n}{t\,n!\Gamma(-n+(an+b)t)}\right| \leq an+b, \qquad n\geq1. \tag{8}

For each fixed n≥1n\geq1, the right side of (7) tends to an+ban+b as t↓0t\downarrow0. At n=0n=0, the coefficient equals

1tΓ(bt)=bΓ(1+bt)(b>0),\frac{1}{t\Gamma(bt)}=\frac{b}{\Gamma(1+bt)} \quad (b>0),

which tends to bb and stays bounded near t=0t=0. It is identically zero if b=0b=0.

For any fixed derivative order jj, the differentiated terms with n≥max(1,j)n\geq\max(1,j) are dominated on |x|≤r|x|\leq r by

(an+b)n(n−1)⋯(n−j+1)rn−j.(an+b)n(n-1)\cdots(n-j+1)r^{n-j}.

The sum of these bounds is finite for every 0<r<10\lt r\lt 1. Dominated convergence for the power series proves the assertion, in particular convergence of the first two derivatives. No expansion uniform in nn is assumed; inequality (8) supplies the required uniform domination.

4. The curvature calculation

For a positive twice differentiable function ff, log concavity requires

Ψf(x):=f′(x)2−f(x)f″(x)≥0,\Psi_f(x):=f'(x)^2-f(x)f''(x)\geq0,

because (logf)″=−Ψf/f2(\log f)''=-\Psi_f/f^2. By the limit just proved,

Ψft(x)t2→Ψha,b(x)\frac{\Psi_{f_t}(x)}{t^2} \longrightarrow \Psi_{h_{a,b}}(x)

locally uniformly for |x|<1|x|\lt 1. Direct differentiation of (6) gives

Ψha,b(x)=a2−2ab−b2−2(a2−b2)x−(a−b)2x2(1−x)6.\Psi_{h_{a,b}}(x) =\frac{a^2-2ab-b^2-2(a^2-b^2)x-(a-b)^2x^2}{(1-x)^6}. \tag{9}

In particular,

Ψha,b(0)=a2−2ab−b2,Ψha,b(1/2)=−16(a2+6ab+b2)<0.\Psi_{h_{a,b}}(0)=a^2-2ab-b^2, \qquad \Psi_{h_{a,b}}(1/2)=-16(a^2+6ab+b^2)\lt 0. \tag{10}

For the boundary family (3), take a=1,b=0a=1,b=0. Then

h1,0(x)=x(1−x)2,(h1,0,h1,0′,h1,0″)(1/2)=(2,12,80),h_{1,0}(x)=\frac{x}{(1-x)^2}, \qquad (h_{1,0},h_{1,0}',h_{1,0}'')(1/2)=(2,12,80),

so Ψft(1/2)/t2→−16\Psi_{f_t}(1/2)/t^2\to-16. It follows that Ψft(1/2)<0\Psi_{f_t}(1/2)\lt 0 for all sufficiently small positive tt. The source's extension α*(0)=1\alpha_*(0)=1 makes the threshold condition automatic.

For the interior family (4), take a=3,b=1a=3,b=1. Then

h3,1(x)=1+2x(1−x)2,(h3,1,h3,1′,h3,1″)(0)=(1,4,14),h_{3,1}(x)=\frac{1+2x}{(1-x)^2}, \qquad (h_{3,1},h_{3,1}',h_{3,1}'')(0)=(1,4,14),

and

(h3,1,h3,1′,h3,1″)(1/2)=(8,40,256).(h_{3,1},h_{3,1}',h_{3,1}'')(1/2)=(8,40,256).

Consequently,

Ψft(0)t2→2>0,Ψft(1/2)t2→−448<0.\frac{\Psi_{f_t}(0)}{t^2}\longrightarrow2>0, \qquad \frac{\Psi_{f_t}(1/2)}{t^2}\longrightarrow-448\lt 0. \tag{11}

The first expression is exactly

Ψft(0)=1Γ(β−α)2−1Γ(β)Γ(β−2α).\Psi_{f_t}(0) =\frac{1}{\Gamma(\beta-\alpha)^2} -\frac{1}{\Gamma(\beta)\Gamma(\beta-2\alpha)}. \tag{12}

For sufficiently small t>0t>0, we have 0<β<10\lt \beta\lt 1, 0<α<10\lt \alpha\lt 1, and (12) is strictly positive. By the source's unique-threshold characterization, α<α*(β)\alpha\lt \alpha_*(\beta). The second limit in (11) gives the failure of log concavity at 1/21/2.

Finally, pα,β=Γ(α+β)ftp_{\alpha,\beta}=\Gamma(\alpha+\beta)f_t multiplies Ψft\Psi_{f_t} by the positive number Γ(α+β)2\Gamma(\alpha+\beta)^2. Both sign conclusions therefore hold for the normalized probability density. This proves the theorem.

5. Scope

More generally, every fixed a>0,b≥0a>0,b\geq0 satisfying a2>2ab+b2a^2>2ab+b^2 gives the same counterexample family α=1−at,β=bt\alpha=1-at,\beta=bt for all sufficiently small t>0t>0. If b>0b>0, this condition is a/b>1+2a/b>1+\sqrt2; if b=0b=0, it always holds. Equations (8)–(12) prove this statement without further estimates.

The proof disproves the complete conjectured equivalence by disproving its threshold-to-density implication. It makes no claim about the remaining equivalence between density log concavity and reciprocal Mittag–Leffler convexity at parameters outside the theorem. Every counterexample has 0<α<10\lt \alpha\lt 1; the degenerate limit (1,0)(1,0) is used only to prove the limiting formula. No explicit numerical value of a small parameter is needed: the strict limiting signs prove an interval of admissible counterexamples.