Comparison on the criticality parameters for two supercritical branching processes in random environments

Xiequan Fan, Haijuan Hu, Hao Wu, Yinna YearXiv 2022 · DOI · arXiv

TL;DR

This paper establishes the first theoretical framework for comparing the growth rates (criticality parameters) of two independent branching processes evolving in random environments. By proving non-uniform Berry-Esseen bounds and Cramér’s moderate deviations, the authors provide the mathematical tools necessary to construct rigorous confidence intervals and perform hypothesis testing on the difference between these growth rates.

What Problem Does This Paper Address?

Branching processes in random environments (BPRE) are stochastic models used to describe population growth (like bacteria, genes, or even the spread of information) where the "rules" of reproduction change randomly at each generation. In a "supercritical" process, the population is expected to grow to infinity, and its long-term growth rate is defined by a criticality parameter, $\mu$.

While mathematicians have well-established limit theorems for a single population—knowing, for instance, that the log of the population size divided by the number of generations converges to $\mu$—there was a significant gap in the literature regarding the comparison of two such populations. If you have two different species in two different changing environments, how do you statistically prove one grows faster than the other? This paper treats this as a "two-sample" problem, providing the convergence rates and deviation probabilities needed for robust statistical inference.

Methods

The authors employ several sophisticated techniques from probability theory to analyze the difference between the normalized log-populations of two BPREs:

  1. Martingale Decomposition: They decompose the log-population size into two parts: a sum of independent and identically distributed (iid) random variables representing the environmental fluctuations, and a martingale term representing the inherent randomness of the branching process itself.
  2. Non-uniform Berry-Esseen Bounds: Standard Berry-Esseen theorems provide a global error bound for how fast a distribution converges to a Normal distribution. The authors prove a non-uniform version, which provides much tighter error bounds in the "tails" of the distribution (values far from the mean).
  3. Cramér’s Moderate Deviations: They establish expansions that describe the asymptotic behavior of tail probabilities. This is crucial for understanding the likelihood of rare, extreme growth differences between the two populations.
  4. Laplace Transform Analysis: To handle the "noise" introduced by the branching mechanism, the authors perform a deep dive into the Laplace transforms of the martingale limits, establishing new bounds on their harmonic moments.

Datasets

Not discussed in the paper. This is a theoretical mathematics paper; the results are proven analytically rather than through empirical data or simulation.

Key Insights and Ideas

  • Environment Dominates: A key "aha" moment in BPRE theory, reinforced here, is that for large populations, the randomness of the environment (the changing growth conditions) eventually outweighs the randomness of the branching (who has how many kids). This allows the difference between two processes to be treated similarly to a sum of independent variables.
  • The Two-Sample U-Statistic View: The authors successfully frame the comparison of growth rates as a two-sample $U$-statistic problem, a standard tool in classical statistics, but adapted here for the much more complex world of dependent branching structures.
  • Refining the Error Factor: Their non-uniform Berry-Esseen bound improves upon previous single-sample results by adding a decaying factor $(1+|x|^{1+\delta'})^{-1}$, which significantly sharpens the accuracy of the normal approximation in the tails.

Key Results

  • Central Limit Theorem (Theorem 2.1): Formally proves that the difference between the estimated growth rates of two populations, when properly scaled, converges to a standard Normal distribution.
  • Convergence Rates (Theorem 2.2): Establishes that the error in this Normal approximation vanishes at a rate of $(m \wedge n)^{-\delta/2}$, where $n$ and $m$ are the generations of the two processes.
  • Moderate Deviation Principle (Theorem 2.3): Provides a precise formula for the probability that the difference between the two growth rates deviates significantly from the mean, valid for a wide range of "moderate" deviations.
  • Statistical Application: The paper concludes by providing explicit formulas for constructing confidence intervals for the difference $\mu_1 - \mu_2$ (Theorem 3.1) and for the variance of the environmental fluctuations (Theorem 3.2).

Limitations and Future Work

  • Supercriticality Requirement: The results are strictly for supercritical processes (where extinction is not the certain outcome). The behavior of subcritical or critical processes in this two-sample context remains an open question.
  • Moment Conditions: The proofs rely on specific "moment conditions" (Assumptions A1-A4), which essentially require that the offspring distributions aren't too wild or heavy-tailed.
  • Independence of Environments: The paper assumes the two environments are independent (or have a specific correlation structure). Future work could explore more complex dependencies between the environments of the two populations being compared.

Summary updated

Abstract

Let $\{Z_{1,n} , n\geq 0\}$ and $\{Z_{2,n}, n\geq 0\}$ be two supercritical branching processes in different random environments, with criticality parameters $μ_1$ and $μ_2$ respectively. It is known that $\frac{1}{n} \ln Z_{1,n} \rightarrow μ_1$ and $\frac{1}{m} \ln Z_{2,m} \rightarrow μ_2$ in probability as $m, n \rightarrow \infty.$ In this paper, we are interested in the comparison on the two criticality parameters. To this end, we prove a non-uniform Berry-Esseen's bound and Cramér's moderate deviations for $\frac{1}{n} \ln Z_{1,n} - \frac{1}{m} \ln Z_{2,m}$ as $m, n \rightarrow \infty.$ An application is also given for constructing confidence interval for $μ_1-μ_2$.
BibTeX
@misc{fan2022comparison,
  author = {Xiequan Fan and Haijuan Hu and Hao Wu and Yinna Ye},
  title = {{Comparison on the criticality parameters for two supercritical branching processes in random environments}},
  year = {2022},
  eprint = {2205.09551},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.09551},
}
Figure