A remark on the comparison of the sum and the maximum of positive random variables

Kazuki Okamura · 2026

Read A remark on the comparison of the sum and the maximum of positive random variables by Kazuki Okamura (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-10395.txt.

Introduction

Okamura disproves a guess about sums and peaks of half normal random draws. Arnold and Villasenor had claimed a link between the sum and the largest value in such samples. His proof breaks that claim for draws that are free and share one half normal law. The same line of proof also reaches a wider class of broad gamma laws. The short note thus closes the guess and hands later work a more general tool.

Catalog record

CREATORKazuki Okamura
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE13621
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.10395

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A remark on the comparison of the sum and the maximum of positive random variables Kazuki Okamura Abstract . We disprove a conjecture stated in a recent paper by Arnold and Villasenor concerning the sum and the maximum of independent and identically distributed half-normal random variables. Our method is applicable to generalized gamma distributions. 2020 Mathematics Subject Classification 60E05; 62E10 1. Introduction Recently, Arnold and Villasenor [ 2 ] provided two proofs that the distributions of the sum and the maximum of two independent half-normal random variables are of the same type. More specifically, it holds that X_{1}+X_{2}=\sqrt{2}\max\{X_{1},X_{2}\}\quad\textup{ in distribution} (1.1) if \displaystyle X_{1} and \displaystyle X_{2} are independent and have a common half-normal distribution. They also showed that under some regularity conditions, ( 1.1 ) implies that \displaystyle X_{1} has a half-normal distribution. In [ 2 , Section 4] , they conjectured that for \displaystyle n\geq 3 , X_{1}+\cdots+X_{n}=(n!)^{1/n}\max\{X_{1},\dots,X_{n}\}\quad\textup{ in distribution} (1.2) if \displaystyle X_{1},\dots,X_{n} are independent and have a common half-normal distribution. The purpose of this note is to disprove the conjecture. Our arguments are applicable to a class of distributions containing all generalized gamma distributions, which are originally introduced by [ 4 ] . Assume that \displaystyle X_{i},i\geq 1 , are independent and identically distributed random variables supported on \displaystyle(0,\infty) . Let \displaystyle S_{n}\coloneqq\sum_{i=1}^{n}X_{i} and \displaystyle M_{n}\coloneqq\max_{1\leq i\leq n}X_{i} for \displaystyle n\geq 1 . Let \displaystyle\Gamma be the gamma function and define \displaystyle h(x)\coloneqq\frac{\log\Gamma(x+1)}{x},x>0 . Theorem 1.1 . Assume that \displaystyle X_{1} has a strictly positive density \displaystyle f supported on \displaystyle(0,\infty) and that there exist four positive constants \displaystyle c_{1},c_{2},\alpha , and \displaystyle\beta such that \lim_{x\to+0}\frac{f(x)}{x^{\alpha-1}}=c_{1}. (1.3) \lim_{x\to\infty}\frac{-\log f(x)}{x^{\beta}}=c_{2}. (1.4) Assume that \displaystyle n\geq 2 and \beta<\frac{\log n}{\log n+h(\alpha)-h(n\alpha)}. (1.5) Then there is no constant \displaystyle C such that \displaystyle S_{n}=CM_{n} in distribution. If \displaystyle X_{1} has the standard half-normal distribution, then ( 1.3 ) and ( 1.4 ) hold with \displaystyle c_{1}=\sqrt{2/\pi},c_{2}=1/2,\alpha=1 , and \displaystyle\beta=2 . For \displaystyle\alpha=1 and \displaystyle\beta=2 , condition ( 1.5 ) is equivalent to \displaystyle n\geq 3 ; this can be verified by induction on \displaystyle n . For every \displaystyle\beta\in(0,1] , ( 1.5 ) holds for every \displaystyle n\geq 2 . We remark that \displaystyle\log n+h(\alpha)-h(n\alpha)>0 for every \displaystyle n\geq 2 by Proposition 2.3 (i) below. We show Theorem 1.1 by contradiction. The main idea of the proof is to first compare the asymptotic behavior of \displaystyle P(S_{n}\leq x) and \displaystyle P(M_{n}\leq x) as \displaystyle x\to+0 and then compare the asymptotic behavior of \displaystyle P(S_{n}>x) and \displaystyle P(M_{n}>x) as \displaystyle x\to+\infty . 2. Proof First, we compare the asymptotic behavior of \displaystyle P(S_{n}\leq x) and \displaystyle P(M_{n}\leq x) as \displaystyle x\to+0 . Let \displaystyle C_{\alpha,n}\coloneqq\frac{\Gamma(n\alpha+1)^{1/(n\alpha)}}{\Gamma(\alpha+1)^{1/\alpha}} . Lemma 2.1 . If there exists a constant \displaystyle C such that \displaystyle S_{n}=CM_{n} in distribution, then \displaystyle C=C_{\alpha,n} . Proof. We see that for every \displaystyle x>0 , P(S_{n}\leq x)=\int_{D_{1}}\prod_{i=1}^{n}f(x_{i})dx_{1}\cdots dx_{n}, where \displaystyle D_{1}\coloneqq\{(x_{1},\dots,x_{n})\in(0,\infty)^{n}\colon x_{1}+\cdots+x_{n}<x\} . By the change of variables \displaystyle x_{i}=xu_{i} , we obtain that for every \displaystyle x>0 , P(S_{n}\leq x)=x^{\alpha n}\int_{D_{2}}\prod_{i=1}^{n}\frac{f(xu_{i})}{(xu_{i})^{\alpha-1}}\prod_{i=1}^{n}u_{i}^{\alpha-1}du_{1}\cdots du_{n}, where \displaystyle D_{2}\coloneqq\{(u_{1},\dots,u_{n})\in(0,\infty)^{n}\colon u_{1}+\cdots+u_{n}<1\} . By ( 1.3 ), we can apply the dominated convergence theorem and obtain that \lim_{x\to+0}\int_{D_{2}}\left|\prod_{i=1}^{n}\frac{f(xu_{i})}{(xu_{i})^{\alpha-1}}-c_{1}^{n}\right|\prod_{i=1}^{n}u_{i}^{\alpha-1}du_{1}\cdots du_{n}=0. Using this together with the identity \displaystyle\int_{D_{2}}\prod_{i=1}^{n}u_{i}^{\alpha-1}du_{1}\cdots du_{n}=\frac{\Gamma(\alpha)^{n}}{\Gamma(n\alpha+1)} , we obtain that \lim_{x\to+0}\frac{P(S_{n}\leq x)}{x^{\alpha n}}=c_{1}^{n}\frac{\Gamma(\alpha)^{n}}{\Gamma(n\alpha+1)}. (2.1) We see that for every \displaystyle x>0 , P(M_{n}\leq x)=P(X_{1}\leq x)^{n}=x^{\alpha n}\left(\int_{0}^{1}\frac{f(xu)}{(xu)^{\alpha-1}}u^{\alpha-1}du\right)^{n}. By ( 1.3 ), we can apply the dominated convergence theorem and obtain that \lim_{x\to+0}\frac{P(M_{n}\leq x)}{x^{\alpha n}}=\frac{c_{1}^{n}}{\alpha^{n}}. (2.2) By ( 2.1 ) and ( 2.2 ), we see that \displaystyle C=C_{\alpha,n} . ∎ Second, we compare the asymptotic behavior of \displaystyle P(S_{n}>x) and \displaystyle P(M_{n}>x) as \displaystyle x\to+\infty . By Lemma 2.1 , \displaystyle S_{n}=C_{\alpha,n}M_{n} in distribution. Without loss of generality, we can let \displaystyle c_{2}=1 . By the assumption ( 1.4 ), for every \displaystyle\epsilon\in(0,1/2) , it holds that for large \displaystyle x , \displaystyle\exp\left(-(1+\epsilon)^{\beta}x^{\beta}\right)\leq f(x)\leq\exp\left(-(1-\epsilon)^{\beta}x^{\beta}\right) . By integration by parts, it holds that for every \displaystyle\beta>0 and \displaystyle z>0 , \int_{z}^{\infty}\exp(-t^{\beta})dt=\frac{\exp(-z^{\beta})}{\beta}z^{1-\beta}-\frac{\beta-1}{\beta}\int_{z}^{\infty}t^{-\beta}\exp(-t^{\beta})dt. Hence we see that for large \displaystyle x , P(X_{1}>x)\geq\frac{1}{2}\frac{\exp(-(1+\epsilon)^{\beta}x^{\beta})}{\beta(1+\epsilon)^{\beta}x^{\beta-1}} (2.3) and P(X_{1}>x)\leq 2\frac{\exp(-(1-\epsilon)^{\beta}x^{\beta})}{\beta(1-\epsilon)^{\beta}x^{\beta-1}}. (2.4) We give a lower bound for \displaystyle P(M_{n}>x) . Since \displaystyle f is strictly positive on \displaystyle(0,\infty) , \displaystyle P(X_{1}>x)>0 for every \displaystyle x>0 . We see that \lim_{x\to\infty}\frac{P(M_{n}>x)}{P(X_{1}>x)}=n. (2.5) By ( 2.3 ) and ( 2.5 ), we see that for large \displaystyle x , P(M_{n}>x)\geq\frac{n\exp(-(1+\epsilon)^{\beta}x^{\beta})}{4\beta(1+\epsilon)^{\beta}x^{\beta-1}}. (2.6) We give an upper bound for \displaystyle P(S_{n}>x) . We first consider the case where \displaystyle\beta\geq 1 . Since \displaystyle\beta\geq 1 , by convexity, we see that \displaystyle\sum_{i=1}^{n}X_{i}^{\beta}\geq n^{1-\beta}S_{n}^{\beta} . By this and the exponential Chebyshev inequality, we see that for every \displaystyle\epsilon\in(0,1) , \displaystyle x>0 and \displaystyle n\geq 1 , \displaystyle\displaystyle P(S_{n}>x) \displaystyle\displaystyle\leq P\left(\sum_{i=1}^{n}X_{i}^{\beta}\geq n^{1-\beta}x^{\beta}\right) \displaystyle\displaystyle\leq\exp\left(-(1-\epsilon)^{\beta}n^{1-\beta}x^{\beta}\right)E\left[\exp\left((1-\epsilon)^{\beta}X_{1}^{\beta}\right)\right]^{n}. (2.7) By ( 1.4 ), \displaystyle E\left[\exp\left((1-\epsilon)^{\beta}X_{1}^{\beta}\right)\right]<\infty . By ( 1.5 ), it holds that \displaystyle C_{\alpha,n}^{-\beta}<n^{1-\beta} . Now we take a sufficiently small \displaystyle\epsilon>0 such that \displaystyle(1+\epsilon)^{\beta}C_{\alpha,n}^{-\beta}<(1-\epsilon)^{\beta}n^{1-\beta} . By this, ( 2.6 ) and ( 2 ), we obtain that \lim_{x\to\infty}\frac{P\left(M_{n}>x/C_{\alpha,n}\right)}{P(S_{n}>x)}=\infty. This contradicts \displaystyle S_{n}=C_{\alpha,n}M_{n} in distribution. The proof of Theorem 1.1 in the case where \displaystyle\beta\geq 1 is completed. We next consider the case where \displaystyle 0<\beta<1 . By ( 1.4 ), \displaystyle E[\exp(tX_{1})]=\infty for every \displaystyle t>0 . By [ 3 , Theorem 2] , \liminf_{x\to\infty}\frac{P(S_{n}>x)}{P(M_{n}>x)}<\infty. (2.8) By Lemma 2.1 , if there exists a constant \displaystyle C such that \displaystyle S_{n}=CM_{n} in distribution, then \displaystyle C=C_{\alpha,n} . By ( 2.8 ) and ( 2.5 ), we obtain that \limsup_{x\to\infty}\frac{P(X_{1}>x)}{P(X_{1}>x/C_{\alpha,n})}=\limsup_{x\to\infty}\frac{P(M_{n}>x)}{P(M_{n}>x/C_{\alpha,n})}>0. (2.9) By Lemma 2.2 (i) below, \displaystyle h(x) is strictly increasing. Hence \displaystyle C_{\alpha,n}>1 for \displaystyle n\geq 2 . By ( 2.3 ) and ( 2.4 ), \lim_{x\to\infty}\frac{P(X_{1}>x)}{P(X_{1}>x/C_{\alpha,n})}=0. This contradicts ( 2.9 ). The proof of Theorem 1.1 in the case where \displaystyle 0<\beta<1 is completed. The following assertions provide properties of \displaystyle h . Lemma 2.2 . (i) \displaystyle h is strictly increasing on \displaystyle(0,\infty) . (ii) \displaystyle g(x)\coloneqq h(\exp(x)) is strictly convex on \displaystyle\mathbb{R} . Proof. (i) Let \displaystyle\psi be the digamma function. Then \displaystyle\frac{d}{dx}\log\Gamma(x+1)=\psi(x+1) . Since the digamma function is strictly increasing, \displaystyle\log\Gamma(x+1) is strictly convex. Hence, the secant slope \displaystyle\frac{\log\Gamma(x+1)-\log\Gamma(1)}{x}=h(x) is strictly increasing on \displaystyle(0,\infty) . (ii) Since \displaystyle h(t)=\frac{1}{t}\int_{0}^{t}\psi(s+1)ds , we see that th^{\prime}(t)=\psi(t+1)-\frac{1}{t}\int_{0}^{t}\psi(s+1)ds=\frac{1}{t}\int_{0}^{t}s\psi^{\prime}(s+1)ds. Since \displaystyle g^{\prime}(x)=\exp(x)h^{\prime}(\exp(x)) , it suffices to show that \displaystyle q(s)\coloneqq s\psi^{\prime}(s+1) is strictly increasing. By an integral expression of the trigamma function ( [ 1 , Formula 6.4.1, p.260] ), \psi^{\prime}(s+1)=\int_{0}^{\infty}\frac{r\exp(-sr)}{\exp(r)-1}dr. Hence q(s)=\int_{0}^{\infty}\exp(-r)\rho\left(\frac{r}{s}\right)dr, (2.10) where \displaystyle\rho(t)\coloneqq\frac{t}{\exp(t)-1} . Since \displaystyle\rho is strictly decreasing, \displaystyle q is strictly increasing. ∎ Proposition 2.3 . Set \displaystyle a_{n}\coloneqq\frac{h(n\alpha)-h(\alpha)}{\log n} for \displaystyle n\geq 2 . Then (i) \displaystyle a_{n}<1 for every \displaystyle n and \displaystyle\lim_{n\to\infty}a_{n}=1 . (ii) \displaystyle(a_{n})_{n} is strictly increasing. (iii) There exists \displaystyle N(\alpha,\beta)\geq 2 such that ( 1.5 ) holds if and only if \displaystyle n\geq N(\alpha,\beta) . Proof. (i) By ( 2.10 ), \displaystyle q(s)<1 for every \displaystyle s>0 . Hence g^{\prime}(x)=\frac{1}{\exp(x)}\int_{0}^{\exp(x)}q(s)ds<1. Since \displaystyle a_{n}=\frac{g(\log(n\alpha))-g(\log\alpha)}{\log(n\alpha)-\log\alpha} , \displaystyle a_{n}<1 . By Stirling’s formula, h(x)=\frac{\log\Gamma(x+1)}{x}=\log x-1+O\left(\frac{\log x}{x}\right),\quad x\to\infty. Hence \displaystyle\lim_{n\to\infty}a_{n}=1 . (ii) Let \displaystyle b_{n}\coloneqq h((n+1)\alpha)-h(n\alpha) and \displaystyle c_{n}\coloneqq\log(n+1)-\log n for \displaystyle n\geq 1 . Since \displaystyle a_{n}=\frac{\sum_{i=1}^{n-1}b_{i}}{\sum_{i=1}^{n-1}c_{i}} , it suffices to show that \displaystyle\left(\frac{b_{n}}{c_{n}}\right)_{n} is strictly increasing. We see that \frac{b_{n}}{c_{n}}=\frac{g(\log((n+1)\alpha))-g(\log(n\alpha))}{\log((n+1)\alpha)-\log(n\alpha)}. By the mean value theorem and Lemma 2.2 (ii), \displaystyle\left(\frac{b_{n}}{c_{n}}\right)_{n} is strictly increasing. (iii) This follows from (i) and (ii). ∎ \displaystyle\alpha\backslash\beta 1.0

Read the complete text