On a population model with memory

Basic Information

• Paper ID: 2501.00306
• Title: On a population model with memory
• Author: Jean Bertoin (Institute of Mathematics, University of Zurich, Switzerland)
• Classification: math.PR (Probability Theory)
• Publication Date: December 31, 2024 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2501.00306

Abstract

This paper considers a population model with memory, which is described by a standard branching process on a finite type space with a given mean reproduction matrix. Motivated by the phenomenon of atavism in evolutionary biology, the author studies a modified dynamics where individuals maintain complete memory of their ancestors, and the reproduction process involves reactivation of genes randomly selected from ancestral lineages. By comparing the spectral radii of two mean reproduction matrices (with and without memory), it is observed that on average, the memory-augmented model always grows at least as fast as the memoryless model. The proof relies on analyzing biased Markov chains on the memory space and establishes the existence of a unique ergodic law through asymptotic coupling.

Research Background and Motivation

Problem Definition

The core question investigated in this paper is: does introducing a hereditary memory mechanism in population models benefit population growth? Specifically, how does the average growth rate of a population change when individuals can "remember" their ancestors' genetic information and randomly reactivate this information during reproduction?

Biological Motivation

The research motivation stems from the phenomenon of atavism in evolutionary biology:

1. Genetic Memory: DNA preserves memories of past events; phenotypically disappeared traits are not completely erased in DNA
2. Gene Dormancy: Certain genes may remain dormant for many generations
3. Trait Reappearance: Traits from distant ancestors may reappear, typically due to mutations or genetic control defects that activate dormant genes

Theoretical Significance

The author points out that even without natural selection and environmental change, merely the possibility of blindly reactivating ancestral genetic material is always beneficial compared to memoryless models. This suggests that the advantage of atavism for populations derives not only from improved adaptability but also from the intrinsic benefits of using previous generation memories for reproduction.

Core Contributions

1. Theoretical Results: Proves that the spectral radius of the branching process with memory is always no less than that of the memoryless model (Theorem 1.1: $\bar{r} \geq r$)
2. Mathematical Framework: Establishes a complete mathematical framework for analyzing population models with memory, including operator theory and Markov chain methods
3. Technical Innovation: Proves unique ergodicity through biased Markov chains and asymptotic coupling techniques
4. Biological Insights: Provides mathematical proof of the intrinsic advantage of genetic memory mechanisms, offering theoretical support for evolutionary biology

Detailed Methodology

Model Setup

Memoryless Model (Baseline Model)

• State Space: Finite type set $S$
• Reproduction Kernel: $\pi(s, \cdot)$, specifying the distribution of offspring types given parental type
• Mean Reproduction Matrix: $m \in \mathbb{R}_+^{S \times S}$, where $m(s,t) = \sum_{n \in \mathbb{N}^S} n_t \pi(s,n)$
• Spectral Radius: $r$ (Perron-Frobenius eigenvalue)

Model with Memory

• Memory Space: $S^{\mathbb{N}} = \{s = (s_0, s_1, s_2, \ldots)\}$, where $s_j$ denotes the type of the ancestor from $j$ generations ago
• Memory Activation: Probability measure $\tau = (\tau(j) : j \geq 0)$ describes the probability of activating the memory of the $j$-th generation ancestor
• Reproduction Mechanism: An individual first selects an ancestral generation $T$ according to $\tau$, then produces offspring according to $\pi(s_T, \cdot)$
• Operator Definition: $\bar{m}f(s) = \sum_{j=0}^{\infty} \tau(j) \sum_{t \in S} m(s_j, t)f(ts)$

Key Technical Components

1. Biased Markov Chain Construction

Introduce the symbol space $\bar{S} = S \times S^{\mathbb{N}}$, where $(s, \mathbf{s})$ records the individual's activated type and complete memory. Construct the transition kernel: $Qf(s, \mathbf{s}) = \sum_{j=0}^{\infty} \tau(j) \sum_{t \in S} \bar{m}(s_j, t)f(s_j, t\mathbf{s})$

where $\bar{m}(s,t) = \frac{m(s,t)h(t)}{rh(s)}$ is the normalized transition matrix.

2. Many-to-One Formula

Establish the connection between operator powers and Markov chains: $\bar{m}^k f(s) = r^k \mathbb{E}_s\left[f(X_k) \times \prod_{j=1}^k \frac{h(Y_j)}{h(X_j)}\right]$

3. Asymptotic Coupling Technique

Since the state space is infinite and the chain maintains complete historical memory, traditional strong Feller properties fail. Employ asymptotic coupling:

• Initialization Phase: When memories have no common prefix, the two chains evolve independently
• Integration Phase: When memories share a common prefix, use the shared portion to activate identical types

Main Theoretical Results

Theorem 1.1 (Main Result)

Statement: $\bar{r} \geq r$, i.e., the spectral radius of the memory-augmented model is no less than that of the memoryless model.

Proof Strategy:

1. Transform the problem into analyzing the product function $\prod_{j=1}^k \frac{h(Y_j)}{h(X_j)}$ via the many-to-one formula
2. Apply Birkhoff's ergodic theorem, requiring proof that: $\liminf_{k \to \infty} \frac{1}{k} \sum_{j=1}^k (\log h(Y_j) - \log h(X_j)) \geq 0$
3. The key lies in proving unique ergodicity of the biased chain and properties of the invariant measure

Proposition 3.2 (Unique Ergodicity)

The biased Markov chain $Q$ is uniquely ergodic, i.e., there exists a unique invariant probability measure $\sigma$.

Lemma 3.3 (Marginal Distribution Properties)

The marginal distributions of the invariant measure $\sigma$ satisfy: $\sigma(\{(s, \mathbf{s}) : s = t\}) = \sigma(\{(s, \mathbf{s}) : s_j = t\}) = \varrho(t)h(t)$

Technical Innovation and Challenges

Main Technical Challenges

1. Infinite State Space: The memory space $S^{\mathbb{N}}$ is infinite
2. Failure of Strong Feller Property: Traditional analytical tools are inapplicable due to complete memory retention
3. Absence of Irreducibility: Standard assumptions of classical Markov chain theory are not satisfied

Innovative Solutions

1. Asymptotic Coupling: Construct two trajectories that eventually become infinitely close but do not coincide completely
2. Compactification Technique: Make the state space compact through pseudo-distance $d(s,s') = a_{|s \wedge s'|}$
3. Multi-level Analysis: Combine operator theory, ergodic theory, and coupling techniques

Concrete Examples

Example 1.2

Consider $S = \{a,b\}$, $m = \begin{pmatrix} 1 & 1 \\ 1 & 2 \end{pmatrix}$, $\tau(0) = u$, $\tau(1) = 1-u$.

• Spectral radius of memoryless model: $r = \frac{3+\sqrt{5}}{2} \approx 2.618$
• Spectral radius of memory-augmented model: $\bar{r}(u)$ is a function of $u$, and $\bar{r}(u) \geq r$ for all $u \in (0,1)$

The paper provides a graph showing how $\bar{r}(u)$ varies with $u$, demonstrating that the inequality is typically strict.

Example 3.1 (Balanced Case)

When the mean reproduction matrix $m$ is balanced (i.e., $\sum_{t \in S} m(s,t)$ is independent of $s$), we have $\bar{r} = r$, in which case memory provides no additional advantage.

Related Work

The paper connects this work to the following areas:

1. Reinforced Galton-Watson Processes: Similar to recent work by Bertoin and Mallein 3,4
2. Penalized Markov Processes: Related to work by Champagnat et al. 7
3. Multi-type Branching Processes: Extension of classical theory 1,12

Conclusions and Discussion

Main Conclusions

1. Genetic Memory is Always Beneficial: Even without selection pressure, the memory mechanism can increase the population growth rate
2. Mathematical Mechanism: The advantage arises from the genetic material of prolific ancestors being present in a larger portion of the current population, making it more likely to be reactivated
3. Theoretical Completeness: Provides a complete mathematical framework for analyzing branching processes with memory

Limitations

1. Model Simplification: Does not account for environmental change and natural selection
2. Memory Activation Fairness: Assumes activation probability is independent of memory content
3. Technical Limitations: Certain classical results (such as the complete Perron-Frobenius theorem) are difficult to establish in the unbounded support case

Future Directions

1. Consider activation mechanisms dependent on memory content
2. Introduce environmental change and selection pressure
3. Study optimization problems with finite memory length
4. Explore applications in other biological contexts

In-Depth Evaluation

Strengths

1. Theoretical Rigor: Proof techniques are sophisticated, particularly the application of asymptotic coupling
2. Biological Relevance: Provides mathematical theoretical support for the phenomenon of atavism
3. Technical Innovation: Overcomes difficulties of traditional Markov chain theory in the infinite memory case
4. Clear Results: Main theorems are concise and possess intuitive biological interpretations

Weaknesses

1. Limited Practical Application: The model is overly idealized with significant gaps from real biological systems
2. Computational Complexity: For practical computation, the complexity introduced by infinite memory may be prohibitive
3. Parameter Sensitivity: Lacks guidance on choosing the memory activation distribution $\tau$

Impact

1. Theoretical Contribution: Adds a new dimension to branching process theory in probability
2. Interdisciplinary Value: Connects mathematics and evolutionary biology
3. Methodological Significance: Asymptotic coupling techniques may apply to other stochastic processes with infinite memory

Applicable Scenarios

1. Theoretical modeling in evolutionary biology
2. Analysis of random systems with historical dependence
3. Theoretical foundation research for genetic algorithms and evolutionary computation

References

The paper cites 14 related references covering multiple fields including branching process theory, Markov chain theory, coupling techniques, and evolutionary biology, providing a solid theoretical foundation for the research.