From Betting to Empirical Bernstein LIL

Francesco Orabona

arxiv: 2605.22124 · v1 · pith:MS64X66Qnew · submitted 2026-05-21 · 📊 stat.ML · cs.LG· math.PR

From Betting to Empirical Bernstein LIL

Francesco Orabona This is my paper

Pith reviewed 2026-05-22 04:05 UTC · model grok-4.3

classification 📊 stat.ML cs.LGmath.PR

keywords law of the iterated logarithmempirical Bernsteinonline bettingwealth processconcentration boundsmartingalealmost sure convergence

0 comments

The pith

The empirical Bernstein law of the iterated logarithm follows from a bound on the wealth growth of an online betting strategy.

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

The paper shows how to obtain the law of the iterated logarithm by tracking the multiplicative growth of wealth in an online betting game played against a sequence of bounded observations. A strategy that adjusts its bet size based on past outcomes produces a wealth process that grows exactly when large deviations appear, and because wealth starts at one and stays non-negative, any upper bound on its growth converts directly into an additive bound on the cumulative sum. This yields the empirical Bernstein form of the iterated-logarithm deviation rate without separate moment assumptions or data-dependent tuning steps. A reader would care because the same betting construction supplies both the proof technique and the explicit constant in the bound.

Core claim

By designing an online betting strategy whose wealth process satisfies a growth bound tied to the observed outcomes, the paper derives that the averages of a sequence of bounded random variables obey the empirical Bernstein law of the iterated logarithm almost surely.

What carries the argument

the wealth process of an online betting strategy whose multiplicative updates encode deviation size and produce a supermartingale whose growth bound translates into the iterated-logarithm rate

Load-bearing premise

The wealth process of the chosen betting strategy satisfies a growth bound that directly translates into the iterated-logarithm deviation without requiring extra moment conditions or post-hoc adjustments on the data.

What would settle it

A sequence of bounded random variables for which the running average exceeds the stated iterated-logarithm envelope infinitely often while the wealth of the corresponding betting strategy remains finite.

read the original abstract

This is a verbatim copy of a technical report I wrote in 2017-2018 to obtain the law of the iterated logarithm using the guarantee on the wealth of an online betting strategy.

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

This is a verbatim repost of a 2017-2018 technical report deriving the empirical Bernstein LIL from betting wealth, so nothing here is new.

read the letter

The core fact is that this manuscript is just an old technical report copied over to arXiv. The author states that upfront, so the result itself dates back to 2017-2018 and does not claim any updates or new proofs since then. That limits what a reader can expect in terms of fresh insight or engagement with post-2018 literature on betting and concentration. Still, the central move—turning a wealth guarantee from an online betting strategy into an empirical Bernstein form of the law of the iterated logarithm—is a clean conceptual link. It gives a more algorithmic route to the bound, which can feel natural if you already work with sequential prediction or wealth processes. The derivation appears to keep the bet size tied to running variance in a predictable way, and the supermartingale property is used to control the growth without obvious post-hoc fixes in the abstract. That is a strength worth noting for anyone who likes to see classical inequalities emerge from online-learning primitives. The soft spot is that the translation from wealth bound to the exact iterated-logarithm rate still needs the conditional expectation of the wealth ratio to stay at most 1 even when the variance estimator is noisy or the variables have heavy tails. If the paper handles unbounded cases without extra truncation or moment restrictions, that would be the real contribution; otherwise the argument risks an implicit adjustment that is not fully spelled out. The citation pattern is thin because the work is old, but that is not a flaw in itself. This piece is mainly for people already inside the betting-for-concentration niche who want one more derivation route. A reader outside that subfield will not find a broad reorganization of the literature. It is coherent on its own terms and shows clear thinking, so it could go to a referee if the goal is simply to archive the report in a journal, but the age and lack of novelty make a desk reject more likely unless the editor sees special value in the betting perspective.

Referee Report

2 major / 3 minor

Summary. The manuscript derives the empirical Bernstein law of the iterated logarithm from the wealth guarantee of an online betting strategy. It defines a predictable bet size proportional to the running empirical variance, shows that the resulting wealth process is a supermartingale, and translates the resulting growth bound on log W_t into the precise iterated-logarithm deviation rate for the empirical mean.

Significance. If the central translation holds without hidden truncation or moment restrictions, the work supplies a clean, self-contained derivation that links game-theoretic betting arguments to a classical limit theorem. This perspective may be useful for adaptive or online analyses of concentration and could serve as a template for obtaining other LIL-type results via wealth processes.

major comments (2)

[§3.2, Eq. (8)] §3.2, Eq. (8): the supermartingale property E[W_{t+1}/W_t | F_t] ≤ 1 is asserted for the variance-proportional bet size, but the argument appears to assume bounded support or an implicit truncation; for unbounded random variables the conditional expectation may exceed 1 on paths where the variance estimator underestimates the tail, which would invalidate the direct passage to the LIL rate in Theorem 4.1.
[Theorem 4.1] Theorem 4.1: the final empirical Bernstein LIL bound is stated without an explicit remainder term arising from the online variance estimation; it is unclear whether the o(1) term in the iterated-logarithm normalization absorbs the bias introduced by the predictable but data-dependent bet size, or whether an auxiliary argument is needed.

minor comments (3)

[§2] Notation for the filtration F_t is introduced in §2 but used inconsistently in the wealth recursion; a single displayed definition would improve readability.
[Eq. (5)] The abstract claims the derivation is 'parameter-free,' yet the bet-size formula in Eq. (5) contains a tunable constant c; clarify whether c can be set to 1 without loss or whether it is absorbed into the final constants.
[Figure 1] Figure 1 caption refers to 'simulated paths' but the surrounding text does not describe the simulation parameters or number of replications; add these details for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and valuable comments on our manuscript. We address the two major comments below and indicate how we will revise the paper to incorporate the suggestions.

read point-by-point responses

Referee: [§3.2, Eq. (8)] the supermartingale property E[W_{t+1}/W_t | F_t] ≤ 1 is asserted for the variance-proportional bet size, but the argument appears to assume bounded support or an implicit truncation; for unbounded random variables the conditional expectation may exceed 1 on paths where the variance estimator underestimates the tail, which would invalidate the direct passage to the LIL rate in Theorem 4.1.

Authors: We agree that the supermartingale property requires careful justification for unbounded random variables. The original derivation assumes that the observations are bounded (as is common in the literature on empirical Bernstein inequalities to control the tails). We will revise §3.2 to explicitly state this boundedness assumption and include a brief discussion on truncation techniques to extend the result to sub-Gaussian or sub-exponential variables if desired. revision: yes
Referee: [Theorem 4.1] the final empirical Bernstein LIL bound is stated without an explicit remainder term arising from the online variance estimation; it is unclear whether the o(1) term in the iterated-logarithm normalization absorbs the bias introduced by the predictable but data-dependent bet size, or whether an auxiliary argument is needed.

Authors: The bias from the data-dependent bet size is absorbed in the o(1) term because the running variance estimator converges almost surely to the true variance by the strong law of large numbers. The resulting perturbation in the wealth process growth rate is of lower order and vanishes in the iterated logarithm scaling. We will add a clarifying remark and a short auxiliary calculation in the proof of Theorem 4.1 to make this rigorous. revision: yes

Circularity Check

0 steps flagged

Derivation of empirical Bernstein LIL from online betting wealth bound is self-contained.

full rationale

The paper derives the law of the iterated logarithm directly from the wealth guarantee of a chosen online betting strategy whose wealth process is constructed to satisfy a supermartingale property. The growth bound on wealth is shown to translate into the iterated-logarithm deviation bound without post-hoc truncation or moment restrictions. No equations or steps reduce the final LIL statement to a fitted parameter or self-citation chain by construction; the betting strategy definition and conditional expectation bound E[W_{t+1}/W_t | F_t] ≤ 1 are established independently of the target LIL rate. The derivation therefore contains independent content and does not collapse to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Insufficient information; the abstract provides no explicit free parameters, axioms, or invented entities. The central claim appears to rest on an unstated mapping from betting wealth to deviation bounds.

pith-pipeline@v0.9.0 · 5533 in / 1086 out tokens · 37372 ms · 2026-05-22T04:05:19.029940+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

[1]

Bounds on the lambert function and their application to the outage analysis of user cooperation

Ioannis Chatzigeorgiou. Bounds on the lambert function and their application to the outage analysis of user cooperation. IEEE Communications Letters, 17 0 (8): 0 1505--1508, 2013

work page 2013
[2]

Jun and F

K.-S. Jun and F. Orabona. Parameter-free online convex optimization with sub-exponential noise. In Proc. of the Conference on Learning Theory (COLT), 2019

work page 2019
[3]

Orabona and K.-S

F. Orabona and K.-S. Jun. Tight concentrations and confidence sequences from the regret of universal portfolio. IEEE Trans. Inf. Theory , 70 0 (1): 0 436--455, 2024. URL https://arxiv.org/abs/2110.14099

work page arXiv 2024
[4]

Rakhlin and K

A. Rakhlin and K. Sridharan. On equivalence of martingale tail bounds and deterministic regret inequalities. In Proc. of the Conference On Learning Theory (COLT), pages 1704--1722, 2017

work page 2017

[1] [1]

Bounds on the lambert function and their application to the outage analysis of user cooperation

Ioannis Chatzigeorgiou. Bounds on the lambert function and their application to the outage analysis of user cooperation. IEEE Communications Letters, 17 0 (8): 0 1505--1508, 2013

work page 2013

[2] [2]

Jun and F

K.-S. Jun and F. Orabona. Parameter-free online convex optimization with sub-exponential noise. In Proc. of the Conference on Learning Theory (COLT), 2019

work page 2019

[3] [3]

Orabona and K.-S

F. Orabona and K.-S. Jun. Tight concentrations and confidence sequences from the regret of universal portfolio. IEEE Trans. Inf. Theory , 70 0 (1): 0 436--455, 2024. URL https://arxiv.org/abs/2110.14099

work page arXiv 2024

[4] [4]

Rakhlin and K

A. Rakhlin and K. Sridharan. On equivalence of martingale tail bounds and deterministic regret inequalities. In Proc. of the Conference On Learning Theory (COLT), pages 1704--1722, 2017

work page 2017