🔢 mathematics

Absorption and redistribution in a poker tournament: an exponential limit

This paper proves that in a Markovian poker tournament model where particles merge or redistribute mass, the empirical distribution of normalized masses converges to an exponential law as the population shrinks, provided the initial distribution has finite second moment and relative entropy, by combining nonlinear kinetic equation analysis with finite-population approximation techniques.

Original authors: Alexander Dvorsky

Published 2026-09-24

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Alexander Dvorsky

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the study of how large groups of things change over time, scientists often look for simple patterns hidden within complex chaos. Imagine a room filled with people, each holding a different amount of money. If these people randomly pair up and exchange their funds according to specific rules, the distribution of wealth in the room will shift. Sometimes, the money might concentrate in fewer hands; other times, it might spread out more evenly. This kind of system, where individual parts interact and change the whole, is a central topic in probability and physics. Researchers are particularly interested in whether, after enough time and enough interactions, the system settles into a predictable shape, regardless of how the money was distributed at the very beginning. This question matters because it helps us understand how order can emerge from randomness in everything from the flow of traffic to the spread of ideas, and even the financial dynamics of a poker tournament.

A mathematician named Alexander Dvorsky has recently explored this question using the specific rules of a poker tournament as a model. In a tournament, players start with a stack of chips. As the game progresses, they are eliminated one by one, but the total number of chips in play remains constant. When two players go all-in, their combined chips are either split between the winner and the loser, or the loser is eliminated and their chips are transferred entirely to the winner. Dvorsky asked a precise question: if you start with a large group of players with random chip stacks, and you watch the tournament as players are eliminated, does the distribution of the remaining stacks eventually look the same, no matter what the starting stacks were? Specifically, if you measure the size of a player's stack relative to the average stack size at that moment, does that ratio settle into a single, universal pattern?

The answer, according to the study, is yes. The research proves that as the number of players drops, the relative sizes of the remaining stacks converge toward a specific, well-known pattern called the exponential distribution. This pattern is famous for its "memoryless" property, meaning that the chance of having a very large stack does not depend on how much you already have; the odds of having a stack twice the average are always the same, regardless of your current size. The study shows that this pattern emerges naturally from the mechanics of the game, provided the initial stacks are not all exactly the same size. The researchers demonstrated that this convergence happens reliably when the number of remaining players is large enough to smooth out random fluctuations, but small enough that the game has progressed significantly from the start.

To reach this conclusion, the team combined two different approaches. First, they analyzed a simplified, theoretical version of the game where the number of players is effectively infinite. In this idealized world, they showed that the system is drawn toward the exponential pattern, much like a ball rolling down a hill to the bottom of a valley. They proved that the process of merging stacks and redistributing chips acts as a mechanism that constantly reduces the "disorder" or unpredictability of the system, pushing it toward this stable shape. Second, they had to account for the fact that real tournaments have a finite number of players, which introduces random noise. They developed a method to measure how closely a real, finite tournament follows the ideal, infinite one. They found that the difference between the real game and the ideal pattern becomes negligible as long as the number of remaining players is not too small compared to the starting number.

The researchers tested these findings with computer simulations involving up to a million starting players. They ran thousands of tournaments with different starting conditions, including cases where everyone started with nearly the same amount and cases where the starting amounts varied wildly. In every case, as the number of players shrank, the distribution of the remaining stacks moved closer to the predicted exponential pattern. The simulations confirmed that the theoretical predictions hold true even in the messy reality of a finite group. However, the study also identified a limit to this behavior. If the number of remaining players becomes too small relative to the starting number, the random fluctuations of the specific game take over, and the pattern begins to break down. The researchers calculated that the pattern holds firm as long as the number of survivors is large enough to keep these random errors in check, specifically when the number of survivors is greater than the starting number raised to a specific power.

One of the most striking aspects of the work is that it applies to a wide variety of starting conditions. Whether the players begin with stacks that are all very similar or stacks that vary greatly, the system forgets its starting point and settles into the same final shape. The only exception is if every single player starts with the exact same number of chips, a scenario that is practically impossible in a real tournament and was excluded from the study. The research also highlights a special case: if the starting stacks are already distributed in a specific way known as the Dirichlet distribution, the system maintains this distribution perfectly throughout the entire tournament, without needing to evolve toward the exponential shape. This provides a rare example of a system that preserves its initial structure, offering a useful benchmark for checking the accuracy of the simulations.

The findings have a direct connection to how poker tournaments are actually run. In many modern tournaments, the size of the mandatory bets, known as blinds, increases as players are eliminated. This rule is designed to keep the average stack size in a specific range relative to the blinds. The study suggests that under these conditions, the distribution of stack sizes relative to the average will naturally tend toward the exponential pattern. This means that a player with a stack twice the average size will always have roughly the same probability of being in that position, regardless of how many players are left or how the game started. The researchers noted that this pattern predicts that about 13.5 percent of the remaining players will have stacks larger than twice the average, a figure that remains stable as the tournament progresses.

While the study focuses on the mechanics of chip distribution, it offers a broader lesson about how complex systems behave. It shows that even when individual interactions are random and the total number of participants is shrinking, a large group can still settle into a highly predictable state. The work does not claim to predict the outcome of any specific hand or the winner of a tournament, but rather describes the statistical landscape that all players inhabit as the game unfolds. By proving that the system converges to a single, universal shape, the research provides a clear mathematical description of how wealth or resources are redistributed in a competitive, eliminating environment. The results are supported by rigorous mathematical proofs and extensive computer simulations, giving a high degree of confidence in the conclusion that the exponential pattern is the inevitable destination for such systems, provided the group remains large enough to smooth out the noise of chance.

Technical Summary: Absorption and Redistribution in a Poker Tournament: An Exponential Limit

Problem Statement
The paper investigates a conservative binary-exchange process with a decreasing population, motivated by the mechanics of two-player all-in poker confrontations. In this model, a finite collection of positive masses (or "stacks") interacts in uniformly chosen pairs. Upon an encounter between masses $a < b$ , the system updates according to a fair coin flip:

Redistribution (Probability 1/2): The pair becomes $(2a, b-a)$ . Both particles survive, and total mass is conserved.
Merging/Absorption (Probability 1/2): The pair becomes $(a+b)$ . One particle is eliminated, and its mass is transferred to the survivor.

The total mass $M_N$ is conserved throughout the process. As the population $n$ decreases from an initial size $N$ to 1, the mean surviving mass $M_N/n$ increases. The central question is whether the empirical distribution of the normalized masses (divided by the current mean) converges to a limiting distribution independent of the initial law as $N \to \infty$ and $n/N \to 0$ .

Methodology
The author employs a combination of kinetic theory, entropy methods, and quantitative finite-population analysis.

Kinetic Equation: The process is approximated by a nonlinear kinetic equation describing the evolution of the probability density $f_s$ in "kinetic time" $s = \log(N/n)$ . The equation accounts for the gain terms from redistribution (maps $A$ and $B$ ) and merging (map $C$ ), as well as loss and dilation terms due to the decreasing population and mean normalization.
Entropy Dissipation: The proof of convergence relies on a Lyapunov argument using relative entropy $H(f)$ $H (f)$ with respect to the mean-one exponential law $E$ $E$ .
- Merging: The merging component is shown to dissipate relative entropy at a rate determined by a functional $I(f)$ , which measures the failure of the memoryless property.
- Redistribution: Using a two-particle entropy method (adapted from Apenko, 2013), it is proven that the redistribution map cannot increase the relative entropy.
- Global Attraction: These inequalities, combined with a second-moment bound, establish that the solution to the kinetic equation converges to the exponential law in the first Wasserstein distance ( $W_1$ ) as $s \to \infty$ .
Finite-Population Approximation: To bridge the gap between the kinetic limit and the finite stochastic system, the author utilizes a backward branching construction. This couples the history of sampled survivors in the finite system with independent branching trees corresponding to the kinetic solution. The error is controlled by bounding the probability of "history intersections" (collisions in the backward tree).
Exact Dirichlet Family: A specific case is analyzed where initial chip proportions follow a uniform Dirichlet distribution ( $D_N$ ). It is shown that this family is preserved under the dynamics (up to reordering), providing an exact finite-population benchmark and a transport bound of order $O(n^{-1/2})$ .

Key Contributions and Results

Convergence Theorem (Theorem 1.1): Under assumptions of independent initial masses with mean 1, finite second moment, and finite relative entropy to the exponential law, the expected first Wasserstein distance between the normalized empirical distribution $\mu_{N,n}$ and the exponential law $E$ vanishes in the joint limit:
$\frac{n}{N} \to 0 \quad \text{and} \quad \frac{n^6}{N^5} \to \infty.$
The error bound is given by:
$E W_1(\mu_{N,n}, E) \leq W_1(f_{\log(N/n)}, E) + C \left( \frac{1}{n} + \frac{N^5}{n^6} \right)^{1/4} + \frac{\sqrt{m_0^2 - 1}}{\sqrt{N}}.$
The first term represents kinetic relaxation, the second controls the finite-population approximation error (dependent on history intersections), and the third accounts for initial mass fluctuations.
Entropy and Compactness: The paper establishes that the set of probability measures with bounded second moments and bounded relative entropy is compact in $W_1$ . It proves that the functional $I(f)$ vanishes if and only if the distribution is exponential, ensuring the uniqueness of the limit.
Exact Preservation of Dirichlet Law: Proposition 5.1 demonstrates that if the initial proportions are uniformly distributed on the simplex (Dirichlet $D_N$ ), the proportions at any subsequent population size $n$ remain Dirichlet distributed ( $D_n$ ). This yields a specific convergence rate of $O(n^{-1/2})$ independent of the $N/n$ ratio, provided $n \to \infty$ .
Numerical Validation: Monte Carlo simulations with up to $N = 2^{20}$ particles confirm the theoretical predictions.
- Non-exponential initial ensembles (Uniform, Gamma, Lognormal) rapidly approach the exponential profile (relaxation) before finite-population fluctuations dominate at very low survivor counts.
- The observed $W_1$ distances align with the theoretical fluctuation scale of the exact Dirichlet reference, even in regimes slightly outside the strict $n^6/N^5 \to \infty$ condition.

Significance and Claims
The paper claims to provide the first rigorous proof of an exponential scaling limit for a conservative exchange process with particle absorption, specifically modeling the "all-in" mechanic of poker tournaments.

Mechanism of Convergence: The work highlights that the convergence is driven by the dissipation of relative entropy through merging, a process that redistribution cannot reverse. This distinguishes the model from pure coagulation or pure exchange models.
Quantitative Finite-Size Effects: Unlike many mean-field limits that assume infinite populations, this work provides explicit error bounds for finite $N$ and $n$ , identifying the critical threshold $n \gg N^{5/6}$ required for the kinetic approximation to hold.
Poker Application: The results offer a theoretical justification for the observation that tournament chip stacks, when normalized by the average stack size, tend toward an exponential distribution as the field shrinks, provided the initial distribution is not pathological (e.g., deterministic equal stacks).
Limitations: The author modestly notes that the current proof requires the condition $n^6/N^5 \to \infty$ due to the coupling method used for history intersections. Extending the result to the full survivor regime ( $n/N \to 0$ for any $n \to \infty$ ) would require controlling correlations after history intersections, a task left for future work. Additionally, the results assume continuous initial densities to avoid exact equal-stack encounters, which would alter the elimination probabilities.

More like this