Searcharxiv⌕ Search

arXiv subjects

Jianhang Ai

Publications and source records attributed to Jianhang Ai.

5 recordsLinked to original sources

Removing the floor from Hoeffding and Bernstein inequalities for finite-population U-statistics

Let $C$ be a uniformly sampled $n$-element subset of a finite population $D$ of size $N$, and let $U_C$ be the complete U-statistic of order $k$ built from a bounded kernel on the $k$-subsets of $D$. We prove a product inequality for the uniform distribution on $\binom{D}{n}$ and deduce from it the comparison $ \log \mathbb{E}\left[e^{θ(U_C-\mathbb{E}U_C)}\right] \le \frac{n}{k}\log \mathbb{E}\left[e^{\frac{kθ}{n}(Y-\mathbb{E}Y)}\right]$, where $Y$ is a single kernel value taken at a uniformly sampled $k$-subset of $D$. This yields Hoeffding, Bennett and Bernstein inequalities for $U_C$ with effective sample size $n/k$ in place of the $\lfloor n/k\rfloor$ available so far. The gain is negligible when $n$ is much larger than $k$, but it approaches a factor of two in the exponent when $k$ is close to $n$.

math.ST↗

On the probability that a discrete IFR random variable exceeds its mean

Let $X$ be an integer-valued random variable with $\mathbb{E}(|X|)<\infty$ whose survival probabilities $S_j=\mathbb{P}(X\ge j), j\in\mathbb{Z}$, form a log-concave sequence. Suppose that for some integer $m$ and some $γ\in (0, 1/2]$ it holds $\mathbb{E}(X)\ge m$ and $\mathbb{P}(X=m+1)\geγ\cdot\mathbb{P}(X=m)> 0$. It is shown that \[ \mathbb{P}(X \ge m+1) \,\ge\, \fracγ{1+2γ} \, , \] with the bound being best possible. As a consequence, it is shown that if $H \sim \text{Hyp}(n, i, k)$ is a hypergeometric random variable for which both $ik/n$ and $(n-i)(n-k)/n$ are at least $\log(4/3)$, then it holds $\mathbb{P}(H >\mathbb{E}(H)) \ge 1/4$ when $\mathbb{E}(H)$ is not an integer, and $\mathbb{P}(H >\mathbb{E}(H)) \ge 1/6$ when $\mathbb{E}(H)$ is an integer, both bounds being best possible.

math.PR↗

An entropic analogue of the MMS conjecture

Let $P=\{x_1,\ldots,x_n\}$ be a multiset consisting of $n\ge 2$ real numbers such that $\sum_{i=1}^{n}x_i=0$ and $\sum_{i=1}^{n}|x_i|>0$, and let $k <n$ be a positive integer. We sample $k$ elements from $P$ without replacement and set $X_P$ be the sum of the elements in our sample. It is shown that the Shannon entropy of $X_P$ satisfies \[ \mathbf{H}(X_P) \ge \mathbf{H}(\text{Ber}(k/n)) \, , \] where $\text{Ber}(k/n)$ is a Bernoulli random variable of mean $k/n$. The result is sharp, and may be seen as an entropic analogue of the Manickam-Miklós-Singhi (MMS) conjecture.

math.CO↗

On lower bounds for hypergeometric tails

Let $n,k$ be positive integers such that $n\geq k$, and let $H$ be a hypergeometric random variable counting the number of black marbles in a sample without replacement of size $k$ from an urn that contains $i\in \{1,\ldots, n\}$ black and $n - i$ white marbles. It is shown that \[ \mathbb{P}(H \ge \mathbb{E}(H)) \ge k/n\, , \, \text{when} \,\, n\ge 8k \, . \] Furthermore, provided that $1\le \mathbb{E}(H)\le \min\{i,k\}-1$ as well as that $\frac{(n-i)(n-k)}{n}>1$, it is shown that \[ \mathbb{P}(H\ge \mathbb{E}(H)) \,\ge\, \frac{e^{-1/12}}{4\sqrt{2}} \cdot \sqrt{\frac{n-1}{n}} \cdot\frac{ \sqrt{\text{Var}(H)} }{1 + \sqrt{1+ \frac{n-1}{n-k}\cdot\text{Var}(H)}}\, . \] Auxiliary results which may be of independent interest include an upper bound on the tail conditional expectation and a lower bound on the mean absolute deviation of the hypergeometric distribution.

math.PR↗

Concentration inequalities for the sum in sampling without replacement: an approach via majorization

Let $P=(x_1,\ldots,x_n)$ be a population consisting of $n\ge 2$ real numbers whose sum is zero, and let $k <n$ be a positive integer. We sample $k$ elements from $P$ without replacement and denote by $X_P$ the sum of the elements in our sample. In this article, using ideas from the theory of majorization, we deduce non-asymptotic lower and upper bounds on the probability that $X_P$ exceeds its expected value.

math.PR↗