pith. sign in

arxiv: 2602.08927 · v3 · pith:ADZZUETWnew · submitted 2026-02-09 · 📊 stat.ML · cs.LG· stat.ME

Online monotone density estimation and log-optimal calibration

Rohan Hore , Ruodu Wang , Aaditya Ramdas This is my paper

Pith reviewed 2026-05-25 06:44 UTC · model grok-4.3

classification 📊 stat.ML cs.LGstat.ME
keywords online density estimationmonotone densityGrenander estimatorp-to-e calibratorlog-likelihood regretsequential hypothesis testingonline learningexpert aggregation
0
0 comments X

The pith

Online monotone density estimators achieve an O(n^{1/3}) expected cumulative log-likelihood gap to the true density.

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

The paper studies how to estimate a monotone density from data arriving one point at a time, without knowing future observations in advance. It introduces two concrete procedures: an online version of the classical Grenander estimator and an aggregation method that combines multiple candidate estimators with exponential weights. In the setting where the true density is monotone, both procedures keep the total shortfall in log-likelihood relative to the true density at order n to the one-third. The same machinery is then used to construct p-to-e calibrators that are optimal for sequential hypothesis testing.

Core claim

In the well-specified stochastic setting where the underlying density is monotone, the expected cumulative log-likelihood gap between the online estimators and the true density admits an O(n^{1/3}) bound. The problem of constructing log-optimal p-to-e calibrators for sequential hypothesis testing can be formulated as an online monotone density estimation problem, and the proposed estimators can be adapted to produce empirically adaptive calibrators with matching optimality guarantees. A pathwise regret bound of order sqrt(n log n) also holds for the aggregation estimator relative to the best offline monotone estimator chosen in hindsight.

What carries the argument

The online analogue of the Grenander estimator together with the expert aggregation estimator based on exponential weighting, both operating under the monotone-density constraint.

If this is right

  • The estimators remain competitive with the best possible offline monotone estimator chosen after seeing the whole sequence.
  • The same procedures yield log-optimal p-to-e calibrators that adapt to the observed data.
  • The O(n^{1/3}) bound applies directly to the cumulative performance of the resulting calibrators in sequential testing.
  • The pathwise sqrt(n log n) regret guarantee holds under only minimal regularity assumptions on the observed sequence.

Where Pith is reading between the lines

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

  • The reduction suggests that shape-constrained online estimation techniques may transfer to other calibration tasks that admit a density-estimation view.
  • The monotone constraint could be relaxed to other shape restrictions while preserving similar regret rates if analogous estimators exist.
  • In practice the aggregation method may be preferable when the monotone assumption is only approximate.

Load-bearing premise

The data are generated from a fixed monotone density in a well-specified stochastic model.

What would settle it

A sequence of observations drawn from a monotone density for which either online estimator produces cumulative log-likelihood shortfall that grows faster than order n to the one-third.

read the original abstract

We study the problem of online monotone density estimation, where density estimators must be constructed in a predictable manner from sequentially observed data. We propose two online estimators: an online analogue of the classical Grenander estimator, and an expert aggregation estimator inspired by exponential weighting methods from the online learning literature. In the well-specified stochastic setting, where the underlying density is monotone, we show that the expected cumulative log-likelihood gap between the online estimators and the true density admits an $O(n^{1/3})$ bound. We further establish a $\sqrt{n\log{n}}$ pathwise regret bound for the expert aggregation estimator relative to the best offline monotone estimator chosen in hindsight, under minimal regularity assumptions on the observed sequence. As an application of independent interest, we show that the problem of constructing log-optimal p-to-e calibrators for sequential hypothesis testing can be formulated as an online monotone density estimation problem. We adapt the proposed estimators to build empirically adaptive p-to-e calibrators and establish their optimality. Numerical experiments illustrate the theoretical results.

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.

Referee Report

0 major / 1 minor

Summary. The paper proposes two online estimators for monotone density estimation—an online analogue of the Grenander estimator and an expert aggregation estimator based on exponential weighting. In the well-specified stochastic setting with monotone true density, it claims an O(n^{1/3}) bound on the expected cumulative log-likelihood gap to the true density. It further claims a sqrt(n log n) pathwise regret bound for the aggregation estimator relative to the best offline monotone estimator in hindsight, under minimal regularity. The work formulates log-optimal p-to-e calibration for sequential testing as an instance of this problem, adapts the estimators to produce empirically adaptive calibrators with optimality guarantees, and includes numerical experiments.

Significance. If the stated rates hold, the work provides online estimators whose cumulative performance matches the classical Grenander rate under monotonicity and connects online learning techniques to a new application in p-to-e calibration. The pathwise regret result under weak assumptions on the sequence is a positive feature, as is the explicit reduction of calibration to monotone density estimation.

minor comments (1)
  1. The abstract states the O(n^{1/3}) and sqrt(n log n) bounds without derivation outlines; the introduction or a dedicated section should briefly indicate the main proof ingredients (online-learning regret plus Grenander analysis) to improve readability.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary, significance assessment, and recommendation of minor revision. No specific major comments were raised in the report.

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external classical results

full rationale

The paper's core claims—the O(n^{1/3}) cumulative log-likelihood bound in the well-specified monotone setting and the reduction of log-optimal p-to-e calibration to online monotone density estimation—are presented as following from standard online learning techniques combined with classical Grenander estimator analysis. No equations or steps in the provided abstract or reader summary reduce a claimed prediction or uniqueness result to a self-defined quantity, fitted parameter, or self-citation chain internal to the paper. The derivation chain is therefore self-contained against external benchmarks, consistent with the reader's assessment of no obvious circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on the domain assumption that the density is monotone and on standard stochastic i.i.d. sampling; no free parameters or invented entities are introduced in the abstract.

axioms (2)
  • domain assumption The underlying density is monotone
    Invoked in the well-specified stochastic setting to obtain the O(n^{1/3}) bound.
  • domain assumption Data are generated i.i.d. from the true density
    Implicit in the stochastic setting used for the expected cumulative gap.

pith-pipeline@v0.9.0 · 5708 in / 1328 out tokens · 41467 ms · 2026-05-25T06:44:50.175570+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Shape-constrained density estimation with Wasserstein projection

    math.ST 2026-03 unverdicted novelty 7.0

    Wasserstein projection onto displacement-convex sets yields a convex optimization approach to shape-constrained univariate density estimation for non-increasing and log-concave cases, with structural properties and a ...