[Paper Review] Kolmogorov's strong law of large numbers in game-theoretic probability: Reality's side

Vladimir Vovk|arXiv (Cornell University)|Mar 20, 2013

Statistical Mechanics and Entropy1 references3 citations

TL;DR

This paper presents an explicit, constructive strategy for Reality in game-theoretic probability that ensures the strong law of large numbers fails when the variance series ∑vₙ/n² diverges. By strategically setting xₙ = 0 until Skeptic’s capital constraint is nearly breached, then setting xₙ = ±n, Reality forces Skeptic’s capital to grow unboundedly if the variance condition holds, thereby violating the almost-sure convergence of Sₙ/n to zero.

ABSTRACT

The game-theoretic version of Kolmogorov's strong law of large numbers says that Skeptic has a strategy forcing the statement of the law in a game of prediction involving Reality, Forecaster, and Skeptic. This note describes a simple matching strategy for Reality.

Motivation & Objective

• To provide a constructive, explicit strategy for Reality in game-theoretic probability that realizes the existence assertion in Theorem 4.1(2) of [2].
• To demonstrate that Reality can prevent the almost-sure convergence of Sₙ/n to zero when ∑vₙ/n² = ∞, even under Skeptic’s capital constraints.
• To offer a simple, self-contained proof of a key implication of Kolmogorov’s strong law of large numbers in the game-theoretic framework.
• To clarify the role of Reality in counteracting Skeptic’s attempts to enforce the SLLN through capital control.

Proposed method

• Reality uses a threshold-based strategy: it sets xₙ = 0 until Skeptic’s move satisfies 𝒦ₙ₋₁ + fₙ(n) ≤ 1.
• When the threshold is reached, Reality sets xₙ = n or xₙ = -n, ensuring a large capital gain for Reality.
• The strategy relies on the function fₙ(x) = Mₙx + Vₙ(x² - vₙ), which models Skeptic’s capital increase; with Mₙ = 0, the focus is on Vₙ and vₙ.
• Reality exploits the fact that Vₙ > n⁻²(1 - 𝒦ₙ₋₁) when fₙ(n) > 1 - 𝒦ₙ₋₁, leading to increasing losses for Skeptic.
• The proof considers two cases: if the threshold is triggered infinitely often, Sₙ/n does not converge to zero; otherwise, Skeptic’s capital eventually becomes negative.
• The argument holds even under unbounded forecasting, as Reality can win when Vₙ < 0 by choosing large |xₙ|.

Experimental results

Research questions

• RQ1 Can an explicit strategy for Reality be constructed to ensure the failure of the strong law of large numbers under the condition ∑vₙ/n² = ∞?
• RQ2 How can Reality exploit Skeptic’s capital constraints to prevent almost-sure convergence of Sₙ/n to zero?
• RQ3 What is the minimal assumption on the variance sequence vₙ that allows Reality to force divergence of Sₙ/n?
• RQ4 Does the game-theoretic framework allow Reality to counteract Skeptic’s capital control and enforce SLLN failure?

Key findings

• Reality can explicitly construct a strategy that guarantees Sₙ/n does not converge to zero whenever ∑vₙ/n² diverges.
• The strategy ensures Skeptic’s capital remains bounded by 1, satisfying the collateral duty constraint.
• If the threshold condition is triggered infinitely often, Sₙ/n fails to converge to zero, directly satisfying the main implication.
• If the threshold is triggered only finitely often, Skeptic eventually becomes bankrupt due to cumulative losses from Vₙvₙ ≥ vₙn⁻²(1 - 𝒦ₙ₋₁).
• The result holds even under unbounded forecasting, as Reality can win when Vₙ < 0 by choosing |xₙ| large.
• The proof establishes that Reality can effectively enforce the failure of the strong law of large numbers in the game-theoretic probability framework.