跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2601.09564 · v3

A Spectral Representation Of The Simple Hypothesis Testing Problem

首次提交 2026-01-14 23:33
版本更新 2026-11-08 01:15
雷达观测 2026-11-10 11:14
arXiv ↗PDF ↗

摘要

The minimum Type II error probability (or volume) of simple hypothesis tests with randomized detectors is expressed as a Riemann integral of the (cumulative) distribution function of the likelihood ratio for all non-negative Type I error probability (or volume) values: β(ε)=∫0∞∣F(1/τ)−ε∣+dτβ(\varepsilon)=\int_{0}^{\infty}|F(1/τ)-\varepsilon|^{+}dτ for all ε≥0\varepsilon\geq0. The derivation relies on convex conjugation (the Legendre transform) and level set integration, and is sufficiently general to extend to tests between σσ-finite measures. Measure change identities relating the β(⋅)β(\cdot) functions of different hypothesis testing problems are established. Approximations for Type II and Type I volumes are derived under two different hypotheses concerning the Kolmogorov distance to normality for the log-likelihood ratio distributions. In both the Central Limit Theorem and the large deviations regimes, the resulting non-asymptotic expressions recover and extend state-of-the-art characterizations of optimal performance in the memoryless case, and improve and generalize them in the Markovian case. Finally, The distinction between the implications of the Central Limit Theorem and the Berry--Esseen Theorem for the asymptotic simple hypothesis testing problem is clarified.

作者备注:30 pages. Submitted to IEEE Transactions on Information Theory

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v3,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v3 · 2026-11-08 01:15 · 最新观测版本

A Spectral Representation Of The Simple Hypothesis Testing Problem

The minimum Type II error probability (or volume) of simple hypothesis tests with randomized detectors is expressed as a Riemann integral of the (cumulative) distribution function of the likelihood ratio for all non-negative Type I error probability (or volume) values: β(ε)=∫0∞∣F(1/τ)−ε∣+dτβ(\varepsilon)=\int_{0}^{\infty}|F(1/τ)-\varepsilon|^{+}dτ for all ε≥0\varepsilon\geq0. The derivation relies on convex conjugation (the Legendre transform) and level set integration, and is sufficiently general to extend to tests between σσ-finite measures. Measure change identities relating the β(⋅)β(\cdot) functions of different hypothesis testing problems are established. Approximations for Type II and Type I volumes are derived under two different hypotheses concerning the Kolmogorov distance to normality for the log-likelihood ratio distributions. In both the Central Limit Theorem and the large deviations regimes, the resulting non-asymptotic expressions recover and extend state-of-the-art characterizations of optimal performance in the memoryless case, and improve and generalize them in the Markovian case. Finally, The distinction between the implications of the Central Limit Theorem and the Berry--Esseen Theorem for the asymptotic simple hypothesis testing problem is clarified.

查看此版本解读 →查看此版本原文 ↗