Optimal-Degree Polynomial Approximations for Exponentials and Gaussian Kernel Density Estimation

TL;DR

This paper derives precise asymptotics for polynomial degrees needed to approximate exponential functions, impacting algorithms for Gaussian kernel density estimation in high dimensions, especially regarding runtime depending on data diameter.

Contribution

It provides exact asymptotic bounds for polynomial approximation degrees of exponentials, and analyzes their implications for the complexity of Gaussian KDE algorithms based on data diameter.

Findings

Asymptotic bounds for $d_{B; \delta}(e^{-x})$ and $d_{B; \delta}(e^{x})$ are established.

Algorithms for Gaussian KDE achieve near-linear time when data diameter is small.

Runtime lower bounds are proved assuming SETH for large data diameters.

Abstract

For any real numbers $B \ge 1$ and $\delta \in (0, 1)$ and function $f: [0, B] \rightarrow \mathbb{R}$, let $d_{B; \delta} (f) \in \mathbb{Z}_{> 0}$ denote the minimum degree of a polynomial $p(x)$ satisfying $\sup_{x \in [0, B]} \big| p(x) - f(x) \big| < \delta$. In this paper, we provide precise asymptotics for $d_{B; \delta} (e^{-x})$ and $d_{B; \delta} (e^{x})$ in terms of both $B$ and $\delta$, improving both the previously known upper bounds and lower bounds. In particular, we show $$d_{B; \delta} (e^{-x}) = \Theta\left( \max \left\{ \sqrt{B \log(\delta^{-1})}, \frac{\log(\delta^{-1}) }{ \log(B^{-1} \log(\delta^{-1}))} \right\}\right), \text{ and}$$ $$d_{B; \delta} (e^{x}) = \Theta\left( \max \left\{ B, \frac{\log(\delta^{-1}) }{ \log(B^{-1} \log(\delta^{-1}))} \right\}\right).$$ Polynomial approximations for $e^{-x}$ and $e^x$ have applications to the design of algorithms for many problems, and our degree bounds show both the power and limitations of these algorithms. We focus in particular on the Batch Gaussian Kernel Density Estimation problem for $n$ sample points in $\Theta(\log n)$ dimensions with error $\delta = n^{-\Theta(1)}$. We show that the running time one can achieve depends on the square of the diameter of the point set, $B$, with a transition at $B = \Theta(\log n)$ mirroring the corresponding transition in $d_{B; \delta} (e^{-x})$: - When $B=o(\log n)$, we give the first algorithm running in time $n^{1 + o(1)}$. - When $B = \kappa \log n$ for a small constant $\kappa>0$, we give an algorithm running in time $n^{1 + O(\log \log \kappa^{-1} /\log \kappa^{-1})}$. The $\log \log \kappa^{-1} /\log \kappa^{-1}$ term in the exponent comes from analyzing the behavior of the leading constant in our computation of $d_{B; \delta} (e^{-x})$. - When $B = \omega(\log n)$, we show that time $n^{2 - o(1)}$ is necessary assuming SETH.