Distribution Testing Meets Sum Estimation

Abstract

We study the problem of estimating the sum of $n$ elements, each with weight $w(i)$, in a structured universe. Our goal is to estimate $W = \sum_{i=1}^n w(i)$ within a $(1 \pm \epsilon)$ factor using a sublinear number of samples, assuming weights are non-increasing, i.e., $w(1) \geq w(2) \geq \dots \geq w(n)$. The sum estimation problem is well-studied under different access models to the universe $U$. However, to the best of our knowledge, nothing is known about the sum estimation problem using non-adaptive conditional sampling. In this work, we explore the sum estimation problem using non-adaptive conditional weighted and non-adaptive conditional uniform samples, assuming that the underlying distribution ($D(i)=w(i)/W$) is monotone. We also extend our approach to to the case where the underlying distribution of $U$ is unimodal. Additionally, we consider support size estimation when $w(i) = 0$ or $w(i) \geq W/n$, using hybrid sampling (both weighted and uniform) to access $U$. We propose an algorithm to estimate $W$ under the non-increasing weight assumption, using $O(\frac{1}{\epsilon^3} \log{n} + \frac{1}{\epsilon^6})$ non-adaptive weighted conditional samples and $O(\frac{1}{\epsilon^3} \log{n})$ uniform conditional samples. Our algorithm matches the $\Omega(\log{n})$ lower bound by \cite{ACK15}. For unimodal distributions, the sample complexity remains similar, with an additional $O(\log{n})$ evaluation queries to locate the minimum weighted point in the domain. For estimating the support size $k$ of $U$, where weights are either $0$ or at least $W/n$, our algorithm uses $O\big( \frac{\log^3(n/\epsilon)}{\epsilon^8} \cdot \log^4 \frac{\log(n/\epsilon)}{\epsilon} \big)$ uniform samples and $O\big( \frac{\log(n/\epsilon)}{\epsilon^2} \cdot \log \frac{\log(n/\epsilon)}{\epsilon} \big)$ weighted samples to output $\hat{k}$ satisfying $k - 2\epsilon n \leq \hat{k} \leq k + \epsilon n$.