Confidence sequences with informative, bounded-influence priors

Fran\c{c}ois Caron; Stefano Cortinovis; Valentin Kilian

arxiv: 2506.22925 · v3 · submitted 2025-06-28 · 📊 stat.ME · math.ST· stat.TH

Confidence sequences with informative, bounded-influence priors

Stefano Cortinovis , Valentin Kilian , Fran\c{c}ois Caron This is my paper

Pith reviewed 2026-05-19 07:18 UTC · model grok-4.3

classification 📊 stat.ME math.STstat.TH

keywords confidence sequencesinformative priorsmethod of mixturesVille's inequalityprior misspecificationGaussian observationsbounded influencesequential inference

0 comments

The pith

Informative priors with polynomial or exponential tails produce confidence sequences that tighten when correct yet stay bounded under any misspecification for Gaussian data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper constructs confidence sequences that incorporate prior information to tighten coverage intervals over time while preserving validity. Standard informative priors risk making sequences vacuous if they mismatch the data. The authors combine mixture martingales with global priors having polynomial or exponential tails and apply an extended form of Ville's inequality. This yields sequences sharper than non-informative versions when the prior aligns with the observations, yet the intervals remain finite and valid for Gaussian data with known variance even under complete prior misspecification. Illustrations use several classical priors.

Core claim

By combining the method of mixtures with a global informative prior whose tails are polynomial or exponential and the extended Ville's inequality, we construct confidence sequences that are sharper than their non-informative counterparts whenever the prior is well specified, yet remain bounded under arbitrary misspecification.

What carries the argument

Global informative priors with polynomial or exponential tails, deployed via the method of mixtures together with the extended Ville's inequality, to control the width of time-uniform confidence sequences.

If this is right

The sequences tighten relative to non-informative baselines when the prior matches the true parameter.
The sequences remain finite and valid even under arbitrary prior misspecification.
The same construction applies directly to several classical priors such as normal and Cauchy.
The approach balances informativeness and robustness in sequential parameter estimation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The tail-control technique may extend to other location-scale families while preserving the bounded-influence property.
In online experiments the method could allow cautious use of historical data without risking invalid sequential tests.
Choosing the prior scale parameter from a small pilot sample might preserve the robustness guarantees if the tail condition is maintained.

Load-bearing premise

The observations follow a Gaussian distribution with known variance.

What would settle it

Generate Gaussian observations with a mean far from any prior mass and check whether the resulting sequence width stays finite as the sample size grows to infinity.

read the original abstract

Confidence sequences are collections of confidence regions that simultaneously cover the true parameter for every sample size at a prescribed confidence level. Tightening these sequences is of practical interest and can be achieved by incorporating prior information through the method of mixture martingales. However, confidence sequences built from informative priors are vulnerable to misspecification and may become vacuous when the prior is poorly chosen. We study this trade-off for Gaussian observations with known variance. By combining the method of mixtures with a global informative prior whose tails are polynomial or exponential and the extended Ville's inequality, we construct confidence sequences that are sharper than their non-informative counterparts whenever the prior is well specified, yet remain bounded under arbitrary misspecification. The theory is illustrated with several classical priors.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a construction for informative yet robust confidence sequences in the Gaussian known-variance case by using priors with controlled tails.

read the letter

The main thing to know is that this paper shows how to fold informative priors into confidence sequences for Gaussian data with known variance without the sequences turning vacuous when the prior is misspecified. They do this by choosing global priors with polynomial or exponential tails, then combining the method of mixtures with an extended Ville's inequality to keep the sequences bounded under arbitrary misspecification while still gaining sharpness when the prior fits well. That targeted trade-off looks like the actual new piece, since earlier mixture-martingale work did not focus on this bounded-influence property for informative cases. The abstract frames the problem clearly and points to illustrations with classical priors, which suggests they are thinking about concrete use rather than just theory. The soft spot is that only the abstract is available, so there is no way to check the derivations, the size of the constants, or how the bounds behave in examples. The known-variance Gaussian restriction is also a real limit on scope, though it is a reasonable place to start. This is for people already working with anytime-valid inference or sequential analysis who want to add prior information without the usual robustness penalty. A reader comfortable with martingale methods would get the most from it and could judge whether the extension is worth adopting. It deserves a serious referee to verify the technical claims and see the illustrations in full.

Referee Report

1 major / 0 minor

Summary. The manuscript develops confidence sequences for the mean parameter under Gaussian observations with known variance. It combines the method of mixtures with global informative priors whose tails are polynomial or exponential, together with an extended form of Ville's inequality, to produce sequences that are sharper than their non-informative counterparts when the prior is well-specified yet remain bounded (non-vacuous) under arbitrary misspecification. The approach is illustrated with several classical priors.

Significance. If the central claims are established with full rigor, the work would supply a concrete mechanism for trading off informativeness against robustness in sequential inference, addressing a recurring practical limitation of prior-informed confidence sequences. The explicit use of tail-controlled priors to enforce bounded influence constitutes a targeted and potentially useful extension of existing mixture-martingale techniques.

major comments (1)

[Abstract] Abstract: the construction is asserted to deliver both sharpness under correct specification and boundedness under arbitrary misspecification, yet the abstract supplies no derivation, proof sketch, or verification that the polynomial/exponential tails together with the extended Ville inequality actually produce the claimed bounded-influence property.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback and positive evaluation of the potential significance of the work. We address the single major comment below.

read point-by-point responses

Referee: [Abstract] Abstract: the construction is asserted to deliver both sharpness under correct specification and boundedness under arbitrary misspecification, yet the abstract supplies no derivation, proof sketch, or verification that the polynomial/exponential tails together with the extended Ville inequality actually produce the claimed bounded-influence property.

Authors: We agree that the abstract, as a concise summary, does not include a derivation or proof sketch. The full manuscript develops the result by constructing mixture martingales from global priors with polynomial or exponential tails and applying the extended Ville inequality; the tail decay ensures that the resulting supermartingale remains integrable (hence yields a non-vacuous bound) even when the prior is arbitrarily misspecified, while still delivering tighter intervals under correct specification. To address the referee's point we will revise the abstract to include a one-sentence indication of how the tail conditions and extended Ville inequality together enforce the bounded-influence property. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

Only the abstract is available, which describes a construction that combines the established method of mixtures with global informative priors having polynomial or exponential tails and the extended Ville's inequality. The resulting confidence sequences are claimed to be sharper than non-informative versions under correct prior specification while remaining bounded under misspecification, all in the Gaussian known-variance setting. No equations, fitted parameters renamed as predictions, self-citations, or other load-bearing reductions to the paper's own inputs are present in the text. The claims follow from the combination of these tools without any step reducing by definition to the result itself.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption of Gaussian observations with known variance and the modeling choice of priors with polynomial or exponential tails; no free parameters or new invented entities are mentioned in the abstract.

axioms (1)

domain assumption Observations are i.i.d. Gaussian with known variance
Explicitly identified in the abstract as the setting for studying the informativeness-robustness trade-off.

pith-pipeline@v0.9.0 · 5629 in / 1314 out tokens · 67866 ms · 2026-05-19T07:18:48.778574+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

By combining the method of mixtures with a global informative prior whose tails are polynomial or exponential and the extended Ville’s inequality, we construct confidence sequences that are sharper than their non-informative counterparts whenever the prior is well specified, yet remain bounded under arbitrary misspecification.
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Theorem 4.3 … CeV α,n(y1:n) − yn → [−σ κ / n ± σ √n √log(n (eg−1 κ(α))²)] as yn → ∞

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Asymptotically Log-Optimal Bayes-Assisted Confidence Sequences for Bounded Means
stat.ML 2026-05 unverdicted novelty 7.0

A Bayesian predictive model adaptively constructs asymptotically log-optimal confidence sequences for bounded means using test martingales.
Asymptotically Log-Optimal Bayes-Assisted Confidence Sequences for Bounded Means
stat.ML 2026-05 unverdicted novelty 7.0

A Bayesian predictive model adaptively selects martingale factors to construct asymptotically log-optimal confidence sequences for bounded means while preserving anytime validity under misspecification.