Conditions for stable evolution of organism traits

Researchers found that genetic variance must stay below a certain limit to prevent traits from overshooting fitness peaks.

The multivariate breeder's equation Δzˉ=G∇ln⁡Wˉ\Delta\bar{z}=G\nabla\ln\bar{W} describes how continuous traits in organisms evolve over discrete time. This system can be written as xk+1=xk+Pk∇f(xk)x_{k+1}=x_k+P_k\nabla f(x_k).

In continuous time, an isolated local maximum of ff is uniformly stable if P(t)P(t) is positive semidefinite. This is true for the GG-matrix. The system is uniformly asymptotically stable if the maximum is an isolated critical point and ∫0∞λ1(P(t)) dt=∞\int_0^\infty\lambda_1(P(t))\,dt=\infty.

The discrete analog of this result is false because iterates can oscillate or diverge even with a constant positive definite PP. Stability is restored if the largest eigenvalue of PkP_k, λn\lambda_n, follows the bound λn(Pk)≤2/L\lambda_n(P_k)\le 2/L. Here, LL is a local Lipschitz constant of ∇f\nabla f.

Under this bound, the maximum is uniformly stable. If a stricter version is met along with ∑kλ1(Pk)=∞\sum_k\lambda_1(P_k)=\infty, the maximum is uniformly asymptotically stable. Biologically, this means genetic variance in every direction must be small relative to the curvature of log mean fitness.

Outside this regime, the mean phenotype can overshoot a fitness peak as hypothesized by Fisher's geometric model. These same conditions also apply to the mean update of Gaussian evolutionary optimization algorithms where the covariance of the search population plays the role of GG.

Key facts

  • The multivariate breeder's equation Δzˉ=G∇ln⁡Wˉ\Delta\bar{z}=G\nabla\ln\bar{W} describes the evolution of continuous traits.
  • Discrete systems can oscillate or diverge if the largest eigenvalue λn\lambda_n exceeds 2/L2/L.
  • Genetic variance must be small relative to the curvature of log mean fitness to avoid overshooting fitness peaks.

All sections

Edition