Almost sure null bankruptcy of testing-by-betting strategies
Pith reviewed 2026-05-16 05:18 UTC · model grok-4.3
The pith
Betting strategies such as universal portfolios and Krichevsky-Trofimov reach zero wealth with probability one under any non-degenerate null distribution.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
These betting strategies produce non-negative wealth martingales that converge to zero almost surely under non-degenerate null distributions, because the partial sums of their logarithmic increments diverge to infinity with probability one.
What carries the argument
The multiplicative wealth update rules that turn capital into a non-negative martingale whose logarithmic increments sum to a divergent series almost surely.
If this is right
- Under the null, wealth converges to zero almost surely for each listed strategy.
- Any non-bankrupt betting strategy is improvable by some modification.
- Sums of terms that are O_p(n^{-1}) diverge almost surely, a fact used in the proofs.
- The asymptotic behavior is settled on almost all paths, complementing existing pathwise regret bounds.
Where Pith is reading between the lines
- In extended sequential testing one may need to restart or reset the wealth process at regular intervals to counteract the almost-sure bankruptcy.
- Comparable multiplicative-update algorithms in other online-learning settings are likely to exhibit the same almost-sure zero-wealth behavior under stationary nulls.
- The result suggests that effective information accumulation for these procedures eventually plateaus when the null is true.
Load-bearing premise
The null distribution is non-degenerate and the strategies use the specific multiplicative updates that make the wealth process a martingale with divergent log-increments.
What would settle it
Generate a long sequence of i.i.d. fair coin tosses, run the Krichevsky-Trofimov betting strategy on it, and check whether the wealth process hits exactly zero almost surely or remains bounded away from zero on a positive-probability set of paths.
read the original abstract
The bounded mean betting procedure serves as a crucial interface between the domains of (1) sequential, anytime-valid statistical inference, and (2) online learning and portfolio selection algorithms. While recent work in both domains has established the exponential wealth growth of numerous betting strategies under any alternative distribution, the tightness of the inverted confidence sets, and the pathwise minimax regret bounds, little has been studied regarding the asymptotics of these strategies under the null hypothesis. Under the null, a strategy induces a wealth martingale converging to some random variable that can be zero (bankrupt) or non-zero (non-bankrupt, e.g. when it eventually stops betting). In this paper, we show the conceptually intuitive but technically nontrivial fact that these strategies (universal portfolio, Krichevsky-Trofimov, GRAPA, hedging, etc.) all go bankrupt with probability one, under any non-degenerate null distribution. Part of our analysis is based on the subtle almost sure divergence of various sums of $\sum_n O_p(n^{-1})$ type, a result of independent interest. We also demonstrate the necessity of null bankruptcy by showing that non-bankrupt strategies are all improvable in some sense. Our results significantly deepen our understanding of these betting strategies as they qualify their behavior on "almost all paths", whereas previous results are usually on "all paths" (e.g. regret bounds) or "most paths" (e.g. concentration inequalities and confidence sets).
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proves that common testing-by-betting strategies (universal portfolio, Krichevsky-Trofimov, GRAPA, hedging, etc.) produce wealth processes that converge to zero almost surely under any non-degenerate null distribution. The argument proceeds by representing log-wealth as a martingale whose conditional increments have negative drift of order O_p(n^{-1}), showing that the sum of these drifts diverges almost surely, and supplementing this with a necessity result that any non-bankrupt strategy is improvable.
Significance. If the almost-sure divergence holds, the result supplies a sharp pathwise characterization under the null that complements existing minimax regret bounds and concentration inequalities. The auxiliary lemma on almost-sure divergence of sums of O_p(n^{-1}) terms is flagged as potentially of independent interest and strengthens the paper's contribution beyond the specific betting procedures.
major comments (2)
- [Main divergence argument (likely §3–4)] The load-bearing step converting the O_p(n^{-1}) conditional drift into almost-sure divergence of the sum (the 'subtle' argument highlighted in the abstract) is not yet verified pathwise for every listed strategy. If the running estimator stabilizes faster than 1/sqrt(n) on a positive-probability set of null paths, the realized bet size can be o(1/n) and standard Borel-Cantelli or ergodic upgrades may fail, leaving P(W_∞ > 0) > 0 for at least one procedure.
- [Necessity section] The necessity claim that non-bankrupt strategies are 'improvable in some sense' requires an explicit definition of improvement and a uniform argument that applies to all listed procedures without extra assumptions on the null.
minor comments (2)
- [Abstract] The abstract's reference to 'subtle almost sure divergence of various sums of ∑_n O_p(n^{-1}) type' would benefit from a one-sentence pointer to the key technical device (e.g., a uniform integrability or ergodic condition).
- [Notation and preliminaries] Notation for the wealth process W_n and the conditional drift c_n should be introduced once and used consistently; occasional switches between multiplicative and additive forms are distracting.
Simulated Author's Rebuttal
We thank the referee for the careful reading and constructive comments on our manuscript. The points raised help us strengthen the presentation of the pathwise arguments and the necessity result. We address each major comment below.
read point-by-point responses
-
Referee: The load-bearing step converting the O_p(n^{-1}) conditional drift into almost-sure divergence of the sum (the 'subtle' argument highlighted in the abstract) is not yet verified pathwise for every listed strategy. If the running estimator stabilizes faster than 1/sqrt(n) on a positive-probability set of null paths, the realized bet size can be o(1/n) and standard Borel-Cantelli or ergodic upgrades may fail, leaving P(W_∞ > 0) > 0 for at least one procedure.
Authors: We thank the referee for highlighting this potential subtlety. Lemma 3.2 in the manuscript establishes almost-sure divergence for any sequence whose conditional drifts are O_p(n^{-1}) with a uniform negative lower bound in expectation; this lemma is applied uniformly because each listed strategy (universal portfolio, Krichevsky-Trofimov, GRAPA, hedging) produces bet sizes driven by estimators that converge at rate 1/sqrt(n) almost surely under non-degenerate nulls by the strong law of large numbers for martingales. Nevertheless, to rule out the faster-stabilization scenario explicitly and make the pathwise verification transparent for every procedure, we will insert a new subsection in Section 4 that checks the bet-size lower bound on a pathwise basis for each strategy. revision: yes
-
Referee: The necessity claim that non-bankrupt strategies are 'improvable in some sense' requires an explicit definition of improvement and a uniform argument that applies to all listed procedures without extra assumptions on the null.
Authors: We agree that greater precision is needed. In the revision we will supply an explicit definition of improvability: a strategy is improvable if there exists another valid betting rule that yields strictly larger expected log-wealth under the null while preserving the same growth guarantees under alternatives. We will then give a uniform argument, based solely on the wealth-martingale property and the assumption that the null is non-degenerate, showing that any non-bankrupt strategy admits such an improvement; the argument does not rely on procedure-specific features beyond those already stated in the paper. revision: yes
Circularity Check
No circularity: direct martingale proof with independent divergence result
full rationale
The derivation relies on standard martingale convergence theorems applied to the wealth process under the null, combined with a separate argument establishing almost-sure divergence of sums of the form ∑ O_p(n^{-1}). This divergence step is explicitly flagged as a result of independent interest and is not obtained by fitting, self-definition, or reduction to prior self-citations. No equations are shown to be equivalent to their inputs by construction, and the central claim (W_n → 0 a.s.) follows from the stated assumptions on non-degenerate nulls and the multiplicative update rules without circular renaming or imported uniqueness theorems. The paper is self-contained against external benchmarks.
Axiom & Free-Parameter Ledger
axioms (2)
- standard math Martingale convergence theorem: a non-negative martingale converges almost surely to a finite limit.
- domain assumption Almost-sure divergence of sums of order O_p(n^{-1}) under non-degenerate distributions.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.