📊 statistics

Regularized e-processes: anytime valid inference with knowledge-based efficiency gains

This contribution presents a regularized e-process framework that leverages incomplete prior knowledge to achieve efficiency gains while ensuring anytime-valid inference via a generalized Ville inequality, thereby enabling possibilistic uncertainty quantification with strong frequentist calibration and Bayesian properties.

Original authors: Ryan Martin

Published 2026-05-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ryan Martin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Stop-and-Go" Problem

Imagine you are a detective solving a case. In traditional statistics, the rules were strict: you had to specify exactly how many pieces of evidence you would gather before starting your search. If you planned to examine 100 pieces of evidence, you had to review all 100 before drawing a conclusion. If you stopped early because you found the answer, or continued because you were confused, the mathematics stated that your conclusion might be wrong.

However, in the real world, we rarely stick to a fixed plan. We look at data as it comes in. If a new drug looks fantastic after just 10 patients, we might stop the study early to save lives. If it looks terrible, we might stop to save money. This is called dynamic data collection.

The problem is that standard statistical tools fail with this approach. They cannot handle "stop-and-go" without risking false alarms.

The Solution: The "E-Process" (The Unstoppable Alarm)

To fix this, statisticians recently invented something called an E-process. Imagine an E-process as a special kind of alarm clock or counter.

How it works: Every time you receive a new piece of data, the counter goes up or down.
The Safety Rule: If the hypothesis you are testing is actually true (e.g., "the drug does nothing"), this counter is mathematically guaranteed to stay low. It is like a leaky bucket that can never fill beyond a certain point if the water source is fake.
The Trigger: If the counter suddenly shoots up and crosses a high line, you know the hypothesis is likely false.
The Superpower: Due to its construction, you can check this counter at any time, whenever you want. Whether you stop after 5 pieces of evidence or after 5,000, the alarm remains reliable. It never gives a false alarm simply because you stopped early.

The Missing Puzzle Piece: Ignoring Your Brain

Here is the catch with standard E-processes: they are "data-only." They treat your prior knowledge as if it does not exist.

Imagine you are a detective who, based on years of experience, knows with near certainty that the suspect is left-handed. Yet, the standard E-process ignores this. It treats every possibility (left-handed, right-handed, ambidextrous) as equally likely until the data proves otherwise. This makes the investigation slower and less efficient, as the detective wastes time checking "right-handed" suspects who are highly unlikely.

The author of this paper asks: Can we build an E-process that respects what we already know without breaking the safety rules?

The Innovation: The "Regularized" E-Process

The paper proposes a new method called the Regularized E-process. Imagine giving the detective an intelligent filter or a knowledge-based lens.

The Filter (Regularizer): Before you even look at the data, you feed your "prior knowledge" into the system. Perhaps you know the suspect is likely left-handed. The system creates a "filter" that makes the "left-handed" path easier to traverse and the "right-handed" path harder.
The Boost: When the data begins to point toward a "left-handed" suspect, the regularized counter shoots up faster than a standard counter. You find the answer sooner.
The Safety Net: The tricky part is that usually, when you tweak a statistical tool to make it faster, you destroy its safety guarantees. The author proves a new mathematical rule (a generalized version of the "Ville inequality") which states: As long as you carefully define your "knowledge," you can speed up the process without breaking the safety alarm.

How It Handles "Uncertain" Knowledge

The paper acknowledges that we rarely know things with 100% certainty. We might say, "I am 90% sure the suspect is left-handed," or "I am fairly sure the suspect is not a giant."

Instead of forcing you to pick a single precise number (as a standard Bayesian statistician might), this method uses imprecise probabilities.

Analogy: Imagine a map with a "fuzzy" zone. Instead of saying "The suspect is exactly at point X," you say "The suspect is somewhere in this gray area."
The method allows you to use this fuzzy, partial knowledge to speed up your investigation. It does not demand that you be a psychic; it only asks that you be honest about what you know and what you do not.

The Result: Faster, Safer, and Smarter

By combining the "anytime valid" safety of E-processes with a "regularizer" that uses your prior knowledge, the paper achieves three things:

Efficiency: You can recognize the truth faster. If your prior knowledge is good, you need fewer data points to be confident in your conclusion.
Safety: Even with this speed boost, the method is mathematically proven to be safe. You will not be fooled by random noise, no matter when you stop the experiment.
Flexibility: It works with vague knowledge (like "I am 90% sure") rather than forcing you to invent fake precision.

Real Example from the Paper

The paper tests this on a medical study comparing two treatments:

Treatment A (Old): Had a very high mortality rate in the past.
Treatment B (New): Showed promising results in early, small studies.

The Dilemma: Doctors want to stop the bad treatment (A) as quickly as possible, but they need proof. Standard tests might take too long to prove B is better, or they might be unreliable if doctors stop the study early because they see how well B is working.

The Paper's Approach:

Researchers feed the "prior knowledge" (Treatment A is likely bad, Treatment B is likely good) into the regularized E-process.
The system quickly gathers evidence that Treatment B is superior.
Because the method is "anytime valid," doctors can stop the study early, confident that the result is real and not a fluke, saving lives by switching patients to the better treatment immediately.

Summary

This paper introduces a way to make statistical investigations faster by using what you already know, while keeping them safe so you can stop looking for answers at any time. It bridges the gap between "blind data crunching" and "human intuition," ensuring that your gut feelings (when supported by evidence) help you find the truth sooner without compromising reliability.

Technical Summary: Regularized e-Processes: Anytime Valid Inference with Knowledge-Based Efficiency Gains

1. Problem Statement

Classical statistical methods typically rely on the assumption that sample sizes are fixed in advance. However, in many practical applications, data collection is dynamic, with stopping rules determined by the data itself (e.g., sequential monitoring). Applying fixed-sample-size methods in such "anytime" scenarios compromises theoretical guarantees and contributes to reliability issues and the replication crisis.

While e-processes have emerged as a solution, offering "anytime valid" inference (reliability independent of the stopping rule) via the Ville inequality, they face a limitation: their growth rate is bounded. If an e-process cannot grow faster through algorithmic adjustments, efficiency stagnates (the speed at which false hypotheses are rejected).

Furthermore, standard methods for anytime valid inference are often "off-the-shelf" and ignore a researcher's partial prior knowledge regarding the quantity of interest $\Theta$ . Traditional Bayesian approaches require precise prior distributions, which may not be justifiable with incomplete information. Conversely, ignoring such knowledge leads to inefficiency. The article addresses the challenge of incorporating incomplete, qualitative prior knowledge into anytime valid inference to achieve efficiency gains without sacrificing reliability.

2. Methodology

The proposed framework integrates imprecise probability (specifically possibility theory) with e-processes via a regularization mechanism.

2.1. Imprecise Probabilistic Regularization

The researcher's prior knowledge is not encoded as a precise probability distribution but as imprecise probability, represented by a credal set $\mathcal{Q}$ (a collection of compatible probability measures). A special case of this is possibility theory, where uncertainty is quantified by a contour function $q(\theta) \in [0, 1]$ .

Regularizer ( $\rho$ ): A function $\rho: \mathcal{T} \to [0, \infty)$ is defined to represent the weight of evidence against a hypothesis $\Theta = \theta$ based on prior knowledge. To be valid, $\rho$ must satisfy a boundedness condition relative to the prior: its expected upper value under the credal set must be $\leq 1$ ( $\overline{\mathbb{E}}_{\mathcal{Q}}[\rho] \leq 1$ ).
Construction: For possibility priors with contour $q$ , a regularizer is constructed by calibrating the inverse of the contour: $\rho(\theta) = \{ \gamma \circ q(\theta) \}^{-1}$ , where $\gamma$ is a "calibration function" ensuring the expectation bound holds. This allows $\rho(\theta)$ to be large (boosting the e-process) for values of $\theta$ deemed implausible by the prior.

2.2. The Regularized e-Process

The core innovation is the definition of a regularized e-process ( $e_{\text{reg}}$ ) by multiplying a standard data-driven e-process ( $e_\theta$ ) with the regularizer:
$e_{\text{reg}}(z_n, \theta) = \rho(\theta) \times e_\theta(z_n)$
This multiplication boosts the e-process for hypotheses incompatible with prior knowledge, theoretically enabling faster rejection of false hypotheses (efficiency gain).

2.3. Generalized Anytime Validity

A critical technical challenge is that manipulating an e-process typically destroys its anytime validity. The article resolves this by redefining the validity criterion as knowledge-dependent.

Instead of demanding validity against all possible data-generating distributions, the framework demands validity against every joint distribution of (Data, $\Theta$ ) that is compatible with both the statistical model and the researcher's credal set $\mathcal{Q}$ .
This is formalized via an induced joint upper probability $\overline{\mathbb{P}}$ , constructed by taking the supremum over all priors in $\mathcal{Q}$ and their corresponding projections onto the parameter space.

3. Main Contributions and Theoretical Results

3.1. Generalized Ville Inequality

The article establishes a generalized Ville inequality for the regularized e-process.

Theorem 1: If $\rho$ is a valid regularizer relative to $\mathcal{Q}$ , then for any stopping time $N$ :
$\overline{\mathbb{P}}\left( e_{\text{reg}}(Z_N, f(\Omega)) > \alpha^{-1} \right) \leq \alpha$
This ensures that Type I error control is maintained, with the error probability measured relative to the researcher's prior knowledge (via $\overline{\mathbb{P}}$ ) rather than as a worst-case uniform bound.

3.2. e-Possibilistic Inference Models (IMs)

The framework extends beyond hypothesis testing to full uncertainty quantification.

Possibilistic IMs: The regularized e-process is transformed into a possibility contour $\pi_{\text{reg}}(z_n, \theta) = 1 \wedge e_{\text{reg}}(z_n, \theta)^{-1}$ .
Theorem 2 & 3: The resulting e-possibilistic IM is strongly anytime valid. It satisfies:
$\overline{\mathbb{P}}\left( \pi_{\text{reg}}(Z_N, f(\Omega)) \leq \alpha \right) \leq \alpha$
This implies that the IM avoids "false confidence" (systematically assigning high confidence to false hypotheses), even under dynamic stopping rules.

3.3. Decision-Theoretic Properties

The article demonstrates that the e-possibilistic IM satisfies several desirable properties:

Likelihood Principle: Inference depends only on the likelihood function.
No Sure Loss: The pricing system derived from the IM is coherent and prevents a "Dutch Book" scenario where an agent is guaranteed to lose money.
Decision Making: The framework supports formal decision making by minimizing the expected upper loss (a Choquet integral) and provides a rule robust to the uncertainty quantified by the IM.

4. Results and Illustrations

The article validates the methodology through theoretical proofs and numerical illustrations:

Efficiency Gains: In a normal distribution mean model, incorporating partial prior information (e.g., "95% sure the mean is positive") via the regularizer significantly reduces the number of samples required to reject a false hypothesis compared to the unregularized e-process. The gain is most pronounced when data are slightly inconsistent with the prior, yet the prior is strong.
Application in Clinical Trials: The framework is applied to data from the Ware (1989) clinical study on ECMO vs. CMT.
- Prior knowledge (historical survival rates) was encoded as a possibility measure.
- The regularized e-process produced tighter confidence sets for the ECMO survival rate (where prior knowledge was strong) and slightly wider sets for the CMT rate (where prior knowledge was weak/incompatible with the data), demonstrating a conservative yet efficient adjustment.
- The conclusion that ECMO is superior was reached with stronger justification than with the standard Fisher's Exact Test (which lacks anytime validity) and aligned with generalized Bayesian analyses.

5. Significance and Claims

The article claims to provide a rigorous, non-asymptotic framework for utilizing partial prior knowledge in dynamic data scenarios.

Bridging the Gap: It bridges the gap between the strict reliability of frequentist anytime valid methods and the efficiency of Bayesian methods, without requiring the strong assumption of a precise prior distribution.
Reliability with Efficiency: It is proven that efficiency gains through regularization do not come at the cost of reliability, provided the validity definition is appropriately relaxed to account for the specific prior knowledge used.
Practical Utility: By using possibility theory and imprecise probabilities, the method enables researchers to utilize qualitative or "surprise-based" knowledge (e.g., "I would be very surprised if $\Theta = c$ ") without forcing them to fabricate a precise probability distribution.
Robustness: The resulting inference is "safe" against adversarial or data-dependent stopping rules and avoids the systematically misleading conclusions associated with "false confidence" in traditional probabilistic quantification.

The author notes that while the theory is established, the practical elicitation of priors (converting qualitative knowledge into a possibility contour) remains a psychological and methodological challenge for future research. Additionally, the current theory focuses on fixed-dimensional parameters, leaving adaptation to high-dimensional cases as an open question.