Resampling Extinction Dynamics

• Resampling extinction dynamics is a stochastic process where noise in multinomial sampling drives irrecoverable state extinction.
• The study derives a closed-form law for predicting first-extinction times, reducing computational complexity from exponential to linear.
• Empirical validations confirm that the theoretical predictions match simulation results across population genetics, ecology, and iterative model training.

Resampling extinction dynamics refers to the quantitative theory of how extinction events occur in systems whose state evolves via discrete resampling—particularly under multinomial noise—characterized by steps that mimic the probabilistic elimination of rare components. This process is relevant to population genetics (allele loss via genetic drift), ecological modeling, chemical networks, novel applications in iterative machine learning setups, and any system where finite-sample stochasticity may drive states to irreversible extinction. Unlike deterministic decay, resampling extinction describes the stochastic, noise-driven crossing of absorbing boundaries (states vanishing to zero), with both the timing and the pathway determined predominantly by the microscopic sampling process.

1. Stochastic Law for First-Extinction Time

A key contribution of the recent work is the derivation of a closed-form, computationally efficient law for the distribution and mean of the first-extinction time under multinomial resampling (Benati et al., 24 Sep 2025). Consider a system described by a probability vector $p = (p_1, ..., p_M)$, where at each iteration, $N$ samples are drawn multinomially with probabilities $p_i$ and the next $p_i$ is the empirical frequency of state $i$:

$p_i(t+1) = \frac{n_i(t)}{N}$

For large $N$, the increment for state $i$ is approximately a zero-drift (mean zero), square-root-variance diffusion:

$dp_i = \sqrt{\frac{p_i}{N}}\,dW_i(t)$

The stochastic differential equation above is the zero-drift Feller (square-root) diffusion, with the extinction boundary at $p_i = 0$ being absorbing. The first-extinction time for any state corresponds to the first passage of $N$0 to zero.

For an initial state $N$1, the cumulative distribution function (CDF) for the extinction of $N$2 by time $N$3 is

$N$4

and, assuming statistical independence among states (valid in the diffusion approximation), the probability that none of the $N$5 states has yet gone extinct by time $N$6 is

$N$7

The mean time to first extinction is

$N$8

This formula is both computationally tractable (linear in $N$9) and exact in the context of independent square-root diffusions, as justified in the large-$p_i$0 regime where multinomial sampling noise dominates extinction (Benati et al., 24 Sep 2025).

2. Comparison to Classical Wright-Fisher/Baxter Results

Earlier work on extinction times, such as the analysis of the Wright-Fisher process by Baxter et al., produced exact (but computationally expensive) formulas that scale as $p_i$1, due to the need to consider all possible nonempty subsets of the state space. The recent derivation reproduces the Baxter mean extinction time exactly by treating each state as an independent Feller diffusion, reducing computational cost from exponential to linear in the number of states. This equivalence is explicitly proven by direct calculation, giving confidence that the diffusion approximation captures not just the leading-order mean but the full functional behavior in appropriate regimes.

This reconciliation establishes the practical relevance of the diffusion-based method for real-world systems with large state spaces or requiring rapid evaluation of extinction risk.

3. Empirical and Simulation Validation

The closed-form law and its assumptions were validated through high-precision simulations across a range of $p_i$2, $p_i$3, and initial entropy values for the probability distribution. For example, with $p_i$4 samples, $p_i$5 states, and entropy $p_i$6, the empirical mean first-extinction time (approx.\ 45.62) closely matches the theoretical prediction (46.40).

Furthermore, the full cumulative distribution of first-extinction times matches the theory: Kolmogorov--Smirnov test statistics (e.g., $p_i$7, $p_i$8) indicate no significant difference between simulations and theoretical CDFs, confirming both the mean and higher moments are accurately predicted. These validation tests are robust to variation in initial distribution and system parameters (Benati et al., 24 Sep 2025).

4. Applications and Predictive Uses

A key potential of the first-extinction law is its predictive utility in contexts where extinction (or mode collapse) is undesirable, as illustrated in model self-training. In such settings, an iterative process (e.g., a neural network's next output distribution being resampled from its own previous prediction) leads to the potential collapse of some states' probabilities to zero. The iteration at which the first extinction occurs coincides in practice with the theoretically computed first-extinction time based on the starting distribution and sample count. This empirical sufficiency suggests the first-extinction law can serve as a universal predictor of the onset of collapse in iterative sampling or self-training pipelines.

The same law applies to other domains with similar dynamics, such as:

• Diversity loss in population genetics (allele loss via drift)
• Rare species extinction in ecological communities under repeated sampling
• Chemical reaction networks where noise drives species elimination
• Finite-sample induced loss of novelty in innovation models
• Information loss in quantum measurement resampling or Dirichlet process mixtures

5. Unified View Across Domains

By deriving and validating a general closed-form law for first-extinction timing under resampling, the theory unifies disparate phenomena—from genetic drift to AI model collapse and chemical network extinction—under a common stochastic-dynamical law. The underlying process is characterized by independent Feller diffusions for each state, with the extinction probability accumulating as the probability mass stochastically fluctuates to zero. The law's computational efficiency makes it suitable for large-scale application in real systems, while its equivalence to established results ensures theoretical rigor.

A key implication is that the time to first extinction (or model collapse) is not a complicated function of unseen “rare paths,” but is wholly determined, in this regime, by the initial stationary distribution and sample size.

6. Methodological and Theoretical Implications

This framework establishes that the extinction dynamics of high-dimensional resampling processes can be analytically characterized without recourse to high-dimensional numerical simulations or exponential-complexity combinatorics. The independence approximation is justified in the low-abundance limit for rare states, which dominate the earliest extinction events, and thus is widely applicable. The theory is agnostic to the meaning of "states" (alleles, tokens, chemical species, neural network outputs, etc.), making it a versatile quantitative tool.

The unified law for extinction under resampling clarifies the stochastic underpinnings of diversity loss and sets the stage for further work linking sampling noise, system design parameters, and long-term stability in high-dimensional discrete stochastic processes.