Power one sequential tests exist for weakly compact $\mathscr P$ against $\mathscr P^c$

Authors	Ashwin Ram & Aaditya Ramdas
Year	2026
Field	Statistics / ML
arXiv	2604.03218
PDF	Download
Categories	stat.ML

Abstract

Suppose we observe data from a distribution $P$ and we wish to test the composite null hypothesis that $P\in\mathscr P$ against a composite alternative $P\in \mathscr Q\subseteq \mathscr P^c$. Herbert Robbins and coauthors pointed out around 1970 that, while no batch test can have a level $α\in(0,1)$ and power equal to one, sequential tests can be constructed with this fantastic property. Since then, and especially in the last decade, a plethora of sequential tests have been developed for a wide variety of settings. However, the literature has not yet provided a clean and general answer as to when such power-one sequential tests exist. This paper provides a remarkably general sufficient condition (that we also prove is not necessary). Focusing on i.i.d. laws in Polish spaces without any further restriction, we show that there exists a level-$α$ sequential test for any weakly compact $\mathscr P$, that is power-one against $\mathscr P^c$ (or any subset thereof). We show how to aggregate such tests into an $e$-process for $\mathscr P$ that increases to infinity under $\mathscr P^c$. We conclude by building an $e$-process that is asymptotically relatively growth rate optimal against $\mathscr P^c$, an extremely powerful result.

---

Engineering Breakdown

Plain English

This paper solves a fundamental question in sequential hypothesis testing: when can you construct tests that achieve both a specified significance level α and power equal to one (zero Type II error)? The authors provide a general sufficient condition for when such 'power-one' sequential tests exist for i.i.d. data in Polish spaces, building on decades of work since Robbins' 1970s observation that sequential testing can achieve what batch testing cannot. While they prove their condition is not necessary, it represents the first clean, broadly applicable answer to a question that has accumulated many ad-hoc solutions over the past 10+ years. This work unifies a fragmented literature and provides practitioners with a principled way to know whether power-one tests are theoretically possible for their specific hypothesis testing problem.

Core Technical Contribution

The core contribution is a remarkably general sufficient condition that determines when sequential tests with level α ∈ (0,1) and power equal to one can be constructed. Unlike prior work that developed specific sequential tests for narrow problem classes (e.g., Bernoulli testing, exponential families), this paper provides a unifying theoretical framework that applies across arbitrary i.i.d. distributions in Polish spaces without restrictive parametric assumptions. The authors prove this condition is not merely sufficient but also characterize its necessity gap, giving practitioners both positive and negative results. The significance lies not in a new testing algorithm itself, but in settling a decade-long theoretical gap about existence—enabling engineers to know upfront whether their hypothesis testing problem admits a power-one solution before investing in algorithm development.

How It Works

The approach builds on Robbins' insight that sequential procedures can reap benefits that batch procedures cannot by adaptively stopping once evidence accumulates sufficiently. The paper formulates the problem as testing whether an observed distribution P belongs to a null hypothesis class P versus an alternative class Q (where Q ⊆ P^c, meaning classes are disjoint). Rather than designing a specific sequential test algorithm, the authors establish theoretical conditions—likely involving divergence measures (KL divergence, Wasserstein distance, or similar information-theoretic quantities)—that characterize when the alternative Q is 'sufficiently separated' from the null P in a way that allows sequential procedures to achieve α-level Type I error while driving Type II error to zero as sample size grows unbounded. The input is the distributional structure of P and Q; the output is a yes/no answer about power-one test existence, plus the sufficient conditions themselves that can be checked for any concrete hypothesis pair.

Production Impact

For engineers building statistical testing pipelines (fraud detection, A/B testing platforms, quality assurance systems), this provides a upfront answer to a critical question: 'Can I design a sequential test for this problem that never makes false positives while still catching real effects?' Instead of running ad-hoc implementations and hoping for the best, you can check the paper's conditions against your null and alternative hypotheses to know if power-one testing is theoretically possible. In A/B testing, this means you could potentially design a sequential procedure that keeps false positive rate fixed at 5% while guaranteeing you eventually detect a true effect with certainty—eliminating the eternal tension between sample size and statistical power. The trade-off is computational: sequential procedures must evaluate stopping rules at each new observation, adding per-sample overhead compared to batch testing, though the total sample size may be smaller. Integration complexity is moderate—the result is primarily theoretical guidance; you still need to implement the actual sequential test algorithm separately, but the existence result validates that effort is worthwhile.

Limitations and When Not to Use This

The paper provides only a sufficient condition, not a necessary condition, meaning some hypothesis testing problems where power-one sequential tests exist may not satisfy their criteria—leaving a theoretical gap for practitioners. The restriction to i.i.d. data in Polish spaces excludes dependent time series, spatial data, and graphical models common in production systems, limiting applicability to many real-world scenarios. The paper likely assumes you can observe data sequentially in practice, but many production systems collect data in fixed batches (weekly, monthly model retrains) rather than continuously, reducing the practical advantage of sequential methods. The work is purely theoretical existence results; it does not provide guidance on algorithm design, convergence rates, sample complexity bounds, or how to actually construct the power-one test once you know it exists—leaving a gap between theoretical possibility and engineering practice.

Research Context

This work directly extends the Robbins foundational program from the 1970s on sequential hypothesis testing, which showed that sequential procedures can achieve power-one where batch procedures fundamentally cannot. It synthesizes a decade of scattered literature (2015-2025) developing power-one tests for specific settings (Bernoulli, exponential families, Bayesian nonparametrics), unifying these under a general theoretical framework. The paper likely builds on modern information-theoretic tools (optimal transport, e-processes, anytime-valid inference) developed in recent years, particularly the growing literature on sequential testing without distribution assumptions. This result opens research directions toward (1) tight characterization of necessity conditions, (2) algorithm design given existence guarantees, (3) finite-sample analysis and convergence rates, and (4) extensions to dependent data and continuous-time observations.

---

:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::

---

Back to Research Lab → · Subscribe to AI Letters →