An Elementary Proof of the Near Optimality of LogSumExp Smoothing

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Samakhoana, Thabo, Grimmer, Benjamin
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912833025867776
author Samakhoana, Thabo
Grimmer, Benjamin
author_facet Samakhoana, Thabo
Grimmer, Benjamin
contents We consider the design of smoothings of the (coordinate-wise) max function in $\mathbb{R}^d$ in the infinity norm. The LogSumExp function $f(x)=\ln(\sum^d_i\exp(x_i))$ provides a classical smoothing, differing from the max function in value by at most $\ln(d)$. We provide an elementary construction of a lower bound, establishing that every overestimating smoothing of the max function must differ by at least $\sim 0.8145\ln(d)$. Hence, LogSumExp is optimal up to small constant factors. However, in small dimensions, we provide stronger, exactly optimal smoothings attaining our lower bound, showing that the entropy-based LogSumExp approach to smoothing is not exactly optimal.
format Preprint
id arxiv_https___arxiv_org_abs_2512_10825
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle An Elementary Proof of the Near Optimality of LogSumExp Smoothing
Samakhoana, Thabo
Grimmer, Benjamin
Statistics Theory
Machine Learning
Optimization and Control
We consider the design of smoothings of the (coordinate-wise) max function in $\mathbb{R}^d$ in the infinity norm. The LogSumExp function $f(x)=\ln(\sum^d_i\exp(x_i))$ provides a classical smoothing, differing from the max function in value by at most $\ln(d)$. We provide an elementary construction of a lower bound, establishing that every overestimating smoothing of the max function must differ by at least $\sim 0.8145\ln(d)$. Hence, LogSumExp is optimal up to small constant factors. However, in small dimensions, we provide stronger, exactly optimal smoothings attaining our lower bound, showing that the entropy-based LogSumExp approach to smoothing is not exactly optimal.
title An Elementary Proof of the Near Optimality of LogSumExp Smoothing
topic Statistics Theory
Machine Learning
Optimization and Control
url https://arxiv.org/abs/2512.10825