Gaussian-efficient testing by betting on the mean of bounded data

arXiv:2608.21694 · stat.ME · Submitted 2026-08-22 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Gaussian-efficient testing by betting on the mean of bounded data".

Jane: Here's the gist of the paper. The proposed method, GE-betting, provides a novel nonasymptotic confidence interval for a constant conditional mean mu in sequences of

0, 1: -valued random variables X 1,, X n where E

X i F i-1: = mu.

Tom: The core advance lies in designing betting fractions that track the conditional rejection probability of the most powerful terminal test in a limiting Gaussian experiment (GE-betting), and then mimicking this tracking within finite samples.

Jane: Key Properties and Guarantees: 1.

Tom: Martingale Validity: The GE-betting update is shown to be martingale-valid under the constant-conditional-mean model, ensuring that the inference remains valid even when the data exhibit martingale dependence.

Paper summary: Tom: So looking at the title, Gaussian-efficient testing by betting on the mean of bounded data, it really sounds like they're tackling a specific problem in statistics related to estimating a parameter from data that is constrained between zero and one.

Jane: And what I take away from this paper is that they developed this GE-betting method which uses betting fractions designed to track terminal rejection probabilities from a limiting Gaussian experiment.

Lu: The authors show that when the data are independent and identically distributed with mean mu between zero and one, the resulting confidence interval width attains the Central Limit Theorem benchmark almost surely.

Meng: It’s about showing that you can get an efficient, tight interval even under martingale dependence, which is a significant step for real-world deployment where data rarely comes from perfectly independent sources.

Jane: They proved this using four steps: first establishing that upper and lower tail fractions are predictable and their caps keep wealth multipliers non-negative.

Tom: And then they showed that the deterministic inversion over the candidate means produces an interval for every sequence, which is computationally better than searching for an arbitrary subset of values.

Lu: They also proved the extension to sampling without replacement, where it becomes a Brownian bridge experiment with a square root correction in the interval width.

Meng: It sounds like they're providing a concrete tool that handles both the iid case and some more complex dependent scenarios efficiently enough for practical use at scale.

Jane: They conclude by showing that this method achieves the tightest results in simulations and demonstrates that martingale-valid inference doesn't have to be inefficient compared to independence-based benchmarks.

Conclusion: Tom: So we've been looking at this paper, "Gaussian-efficient testing by betting on the mean of bounded data." Essentially, they're showing a way to test something about a mean when your data is stuck between zero and one.

Jane: That’s right. The authors are using something called GE-betting to create a confidence interval that works even when the data isn't independent or identically distributed.

Lu: The core idea is this tracking system they build by mimicking what happens in a perfect Gaussian experiment, but adapting it for these real-world bounded numbers.

Meng: From an engineering standpoint, I’m interested in how they handle the scaling. If this method works on millions of data points, does the complexity explode?

Tom: They show that this update process remains martingale valid even when there's dependence in the data sequence. That’s a big deal for real-world scenarios where things aren't perfectly clean.

Jane: The result is pretty concrete when we look at the efficiency. In cases where the data are independent, they manage to get an interval width that matches what you expect from standard central limit theorem tests almost surely.

Lu: And they even extend this to sampling without replacement, which means we’re talking about a square root correction in the width of that confidence interval.

Meng: A square root correction is nice, but I need to know how much better it is than just running a standard test on the same data. Does this method save significant computation time?

Tom: It’s less about speed and more about accuracy under dependence. The paper suggests this approach avoids the usual penalty you take when you have to assume independence when your data isn't independent.

Jane: So, it means we can build more reliable confidence intervals for bounded data without needing a perfect model of how that data is generated.

Lu: It’s pushing the boundary on what we thought was possible with these types of sequential inference methods. It shows that the martingale validity holds up under these specific constraints.

Meng: If this translates well to production environments where you can’t control the exact dependence structure, it could be very useful for setting operational bounds on AI models.

Tom: Exactly. It’s moving inference away from overly strict independence assumptions toward something more robust for data that actually looks like real-world information.

Jane: We'll look at how this new method changes what we consider efficient when testing bounded parameters next.

Diego Martinez-Taboada, Aaditya Ramdas

Department of Statistics & Data Science, Carnegie Mellon University · Department of Statistics, Stanford University

stat.ME

Submitted: 2026-08-22

Updated: 2026-10-30

Code: https://github.com/DMartinezT/betting

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 98/100

The gist: The proposed method, GE-betting, provides a novel nonasymptotic confidence interval for a constant conditional mean mu in sequences of [0, 1] -valued random variables X 1…

Key concepts

GE-betting
A method that designs betting fractions based on tracking the conditional rejection probability of the most powerful terminal test in a limiting Gaussian experiment. This allows for inference on the constant conditional mean $\mu$ in finite samples, ensuring martingale validity.
Martingale Validity
A property ensuring that the GE-betting update remains valid even when data exhibit martingale dependence. This guarantees that the resulting statistical inference is sound despite non-independent observations.
Gaussian Efficiency
When data are independent and identically distributed (i.i.d.), this method produces confidence intervals whose limiting width matches the benchmark set by the Central Limit Theorem, outperforming previous betting intervals.

Terminology

Summary

The proposed method, GE-betting, provides a novel nonasymptotic confidence interval for a constant conditional mean mu in sequences of [0, 1] -valued random variables X 1,, X n where E[X i F i-1] = mu. p. 0 The core advance lies in designing betting fractions that track the conditional rejection probability of the most powerful terminal test in a limiting Gaussian experiment (GE-betting), and then mimicking this tracking within finite samples. p. 0

Key Properties and Guarantees:

  1. Martingale Validity: The GE-betting update is shown to be martingale-valid under the constant-conditional-mean model, ensuring that the inference remains valid even when the data exhibit martingale dependence.

  2. Interval Geometry: The construction provides an interval on every sample path, which is achieved by sharing a predictable variance estimator across candidate means.

  3. Gaussian Efficiency (IID Case): When observations are independent and identically distributed (i.i.d.) with mu in (0, 1) and variance sigma squared > 0, the limiting width of the confidence interval matches that of the Central Limit Theorem benchmark: len(I n) = 2 sigma z 1-delta/2/sqrt n + o(n-1/2) almost surely (Theorem 5. p. 01). This surpasses previous, less efficient betting intervals. p. 0

  4. Extension to Sampling Without Replacement: The method is extended to sampling without replacement, where the Gaussian experiment becomes a Brownian bridge, leading to a square-root correction in the interval width. It achieves state-of-the-art empirical performance and rapidly approaches the Gaussian finite-population benchmark under this setting.

Methodology Overview:

The GE-betting update is defined by computing single mean and variance estimates iteratively:

mu bt = 1/2 + sum from j=1 to t+1 X j t, v bt = 1/4 + sum from j=1 to t+1(X j - mu bj-1) squared

This update is derived from an auxiliary Gaussian experiment to track the conditional probability of the optimal terminal Gaussian rejection event. The method's implementation involves deriving a betting fraction from this auxiliary experiment (Section 5.1), transferring it to a bounded-data update (Section 5. p. 32), and finally stating the formal validity and efficiency result (Theorem 5. p. 43).

Theoretical Derivations:

The proof of Theorem 7.1, concerning validity under sampling without replacement, follows a four-step structure:

  • Step 1 (Finite-Sample Validity): Establishes that upper and lower-tail fractions are predictable and their caps keep wealth multipliers non-negative, leading to an interval intersection.

  • Step 2 (Bridge Tracking): Utilizes strong preterminal bridge tracking to control estimator errors, showing almost sure bounds on the standardized Doob martingale.

  • Step 3 (Terminal Polarization): Proves summable terminal polarization using exponential decay bounds related to the rejection threshold.

  • Step 4 (Interval Endpoints): Connects grid decisions to interval endpoints using monotonicity, ultimately proving that the resulting confidence interval width is bounded by z + epsilon eventually almost surely, as shown by intersecting over positive rational epsilon 0.

The authors conclude that GE-betting demonstrates that martingale-valid inference need not incur an empirical width penalty relative to independence-based benchmarks. The research suggests promising avenues for broader efficiency theory concerning conditional e-values and extensions beyond bounded observations. p. 8

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Predictable Betting Strategy for Bounded Data: AI systems can implement a confidence interval by using predictable betting fractions derived from inverting terminal e-values generated by Gaussian-efficient betting (GE-betting). This allows the system to maintain martingale validity even when data are not independent and identically distributed.

  2. Gaussian Efficiency for IID Data: For iid data, the AI can achieve a confidence interval width that matches the central limit theorem benchmark, as its construction tracks the conditional rejection probability of the most powerful terminal test in a limiting Gaussian experiment.

  3. Sampling Without Replacement Inference: AI systems can perform inference on finite populations drawn without replacement by using a Brownian-bridge design calculation to account for the uncertainty removed as the population is revealed, resulting in bounds that are competitive with empirical benchmarks like WSR running intersection.

  4. Anytime Validity for Fixed Horizons: The system can be designed to be horizon-free within the planned window, meaning it can generate a confidence sequence where coverage is guaranteed at every time point up to a known maximum sample size n, rather than just at the final observation time.

  5. Adaptive Interval Geometry: The AI's confidence interval endpoints are derived from an interval on every sample path property when using a shared predictable variance estimator across candidate means, ensuring the resulting interval is convex and easy to compute by finding its two endpoints.

Sources

Related papers