Stochastic dynamics in finite populations

Stochastic differential equations (SDE) provide a general framework to describe the evolutionary dynamics of an arbitrary number of types in finite populations, which results in demographic noise, and to incorporate mutations. For large, but finite populations this allows to include demographic noise without requiring explicit simulations. Instead, the population size only rescales the amplitude of the noise. Moreover, this framework admits the inclusion of mutations between different types, provided that mutation rates, $\mu$, are not too small compared to the inverse population size $1/N$. This ensures that all types are almost always represented in the population and that the occasional extinction of one type does not result in an extended absence of that type. For $\mu N\ll 1$ this limits the use of SDE’s, but in this case well established alternative approximations are available based on time scale separation. We illustrate our approach by a Rock-Scissors-Paper game with mutations, where we demonstrate excellent agreement with simulation based results for sufficiently large populations. In the absence of mutations the excellent agreement extends to small population sizes.

This tutorial complements a series of research articles by Arne Traulsen, Jens Christian Claussen & Christoph Hauert

Rock-Paper-Scissors game

Value of the element ${\mathcal {C}}_{xx}(x,y,z)$ of the noise matrix ${\mathcal {C}}(\mathbf {x} )$ for $d=3$ strategies and $\mu =0$. ${\mathcal {C}}_{xx}(x,y,z)$ determines how the noise in the $x$-direction affects the $x$-coordinate. In the case of $\mu =0$, this noise vanishes for $x\to 0$. For $y\to 0$ and $z\to 0$ we recover the usual multiplicative noise from one-dimensional evolutionary processes.

Comparisons between the deterministic dynamics in infinite populations, the stochastic dynamics in finite populations and individual based simulations focus on the Rock-Paper-Scissors game with a generic payoff

${\begin{matrix}~&{\begin{matrix}\ \ R\quad &S\quad &P\quad \end{matrix}}\\{\begin{matrix}R\\S\\P\end{matrix}}&{\begin{pmatrix}0&{\frac {s}{2}}&-1\\-1&0&2+s\\{\frac {1+s}{3}}&-1&0\end{pmatrix}}\end{matrix}}.$

According to the replicator equation the game exhibits saddle node fixed points at $x=1,y=1$, and $z=1-x-y=1$ as well as an interior fixed point at $\textstyle {\hat {\mathbf {x} }}=\left({\frac {1}{2}},{\frac {1}{3}},{\frac {1}{6}}\right)$ independent of the parameter $s$. For $s>1$, ${\hat {x}}$ is a stable focus and an unstable focus for $s<1$. In the non-generic case $s=1$ the dynamics exhibits closed orbits.

Deterministic Dynamics

Replicator dynamics

In the limit $N\to \infty$ with $s=1.4$ and without mutations, $\mu =0$, ${\hat {x}}$ is an attractor of the replicator dynamics. The figure shows a sample trajectory that spirals towards the interior fixed point ${\hat {x}}$.

Replicator-Mutator dynamics

Including mutations in the replicator dynamics gives rise to the replicator-mutator dynamics. Mutations destabilizes the interior fixed point ${\hat {x}}$ an can give rise to stable limit cycles. The image shows a sample trajectory for $\mu =0.002$.

Stochastic Dynamics

Stochastic differential equations

Stochastic differential equations, with mutations

Individual Based Simulations

Individual based simulations

Individual based simulations, with mutations

From finite to infinite populations

Performance comparison of individual based simulations (IBS) versus stochastic differential equations (SDE). a ratio of the CPU times $CPU_{\text{SDE}}/CPU_{\text{IBS}}$ as a function of the population size, $N$, and the number of strategic types, $d$. The bold contour indicates equal performance. For small $N$ and large $d$ IBS are faster (red region), but for larger $N$ and smaller Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle d} SDE are faster (blue region). Each contour indicates a performance difference of one order of magnitude. b computational time with $d=10$ as a function of $N$ for IBD (red) and SDE (blue). As a reference for the scaling $N^{2}$ (red) and a constant (blue) are shown. c computational time with $N=5000$ as a function of $d$ for IBD (red) and SDE (blue). As a reference for the scaling $d^{1/2}$ (red) and $d^{3}$ (blue) are shown. For a proper scaling argument much larger $d$ are required but already $d=100$ far exceeds typical evolutionary models and hence is only of limited relevance in the current context. All comparisons use a constant payoff matrix and the local update process (such that ${\mathcal {A}}_{k}(\mathbf {x} )=0$ and $\gamma (\pi _{j},\pi _{k})=1/2$), a mutation rate of $1/N$ and are based on at least $1000$ time steps as well as at least one minute running time. CPU time is measured in milliseconds required to calculate $1000$ time steps. The time increment for the SDE is $dt=0.01$.

In unstructured, finite populations of constant size, $N$, consisting of $d$ distinct strategic types and with a mutation rate, $\mu$, evolutionary changes can be described by the following class of birth-death processes: In each time step, one individual of type $j$ produces a single offspring and displaces another randomly selected individual of type $k$. With probability $1-\mu$, no mutation occurs and $j$ produces an offspring of the same type. But with probability $\mu$, the offspring of an individual of type $i$ ($i\neq j$) mutates into a type $j$ individual. This results in two distinct ways to increase the number of $j$ types by one at the expense of decreasing the number of $k$ types by one, hence keeping the population size constant. Biologically, keeping $N$ constant implies that the population has reached a stable ecological equilibrium and assumes that this equilibrium remains unaffected by trait frequencies. The probability for the event of replacing a type $k$ individual with a type $j$ individual is denoted by $T_{kj}$ and is a function of the state of the population $\mathbf {X} =(X_{1},X_{2},\ldots X_{d})$, with $X_{n}$ indicating the number of individuals of type $n$ such that $\textstyle \sum _{n=1}^{d}X_{n}=N$.

For such processes we can easily derive a Master equation:

$P^{\tau +1}({\mathbf {X} })=P^{\tau }({\mathbf {X} })+\!\sum _{j,k=1}^{d}\!P^{\tau }({\mathbf {X} }_{j}^{k})T_{kj}({\mathbf {X} }_{j}^{k})-P^{\tau }({\mathbf {X} })\ T_{jk}({\mathbf {X} }))$

where $P^{\tau }({\mathbf {X} })$ denotes the probability of being in state $\mathbf {X}$ at time $\tau$ and ${\mathbf {X} }_{j}^{k}=(X_{1},\ldots X_{j}-1,\ldots X_{k}+1,\ldots X_{d})$ represents a state adjacent to $\mathbf {X}$. For large but finite $N$ the Kramers-Moyal expansion yields a convenient approximation in the form of a Fokker-Planck equation:

${\dot {\rho }}({\mathbf {x} })=-\sum _{k=1}^{d-1}{\frac {\partial }{\partial x_{k}}}\rho ({\mathbf {x} }){\mathcal {A}}_{k}({\mathbf {x} })+{\frac {1}{2}}\sum _{j,k=1}^{d-1}{\frac {\partial ^{2}}{\partial x_{k}\partial x_{j}}}\rho ({\mathbf {x} }){\mathcal {B}}_{jk}({\mathbf {x} })$

where ${\mathbf {x} }={\mathbf {X} }/N$ represents the state of the population in terms of frequencies of the different strategic types and $\rho ({\mathbf {x} })$ is the probability density in state $\mathbf {x}$. The drift vector ${\mathcal {A}}_{k}({\mathbf {x} })$ is given by

${\mathcal {A}}_{k}({\mathbf {x} })=\sum _{j=1}^{d}{\Big (}T_{jk}({\mathbf {x} })-T_{kj}({\mathbf {x} }){\Big )}=-1+\sum _{j=1}^{d}T_{jk}({\mathbf {x} })$

For the second equality we have used $\textstyle \sum _{j=1}^{d}T_{kj}({\boldsymbol {x}})=1$, which simply states that a $k$-type individual transitions to some other type (including staying type $k$) with probability one. ${\mathcal {A}}_{k}({\mathbf {x} })$ is bounded in $[-1,d-1]$ because the $T_{jk}$ are probabilities.

The diffusion matrix ${\mathcal {B}}_{jk}({\mathbf {x} })$ is defined as

${\begin{aligned}{\mathcal {B}}_{jk}({\mathbf {x} })&=-{\frac {1}{N}}\left[T_{jk}({\mathbf {x} })+T_{kj}({\mathbf {x} })\right]\quad {\rm {for}}\quad j\neq k\\{\mathcal {B}}_{jj}({\mathbf {x} })&={\frac {1}{N}}\sum _{l=1,l\neq j}^{d}{\Big (}T_{jl}({\mathbf {x} })+T_{lj}({\mathbf {x} }){\Big )}\end{aligned}}$

Note that the diffusion matrix is symmetric, ${\mathcal {B}}_{jk}({\mathbf {x} })={\mathcal {B}}_{kj}({\mathbf {x} })$ and vanishes as $\sim 1/N$ in the limit $N\to \infty$.

The noise arising through demographic changes and mutations is uncorrelated in time and hence the Itô calculus can be applied to derive a Langevin equation

${\dot {x}}_{k}={\mathcal {A}}_{k}({\mathbf {x} })+\sum _{j=1}^{d-1}{{\mathcal {C}}_{kj}({\mathbf {x} })}\xi _{j}(t)$

where the $\xi _{j}(t)$ represent uncorrelated Gaussian white noise with unit variance, $\langle \xi _{k}(t)\xi _{j}(t^{\prime })\rangle =\delta _{kj}\delta (t-t^{\prime })$. The matrix ${{\mathcal {C}}({\mathbf {x} })}$ is defined by ${\mathcal {C}}^{T}({\mathbf {x} }){\mathcal {C}}({\mathbf {x} })={\mathcal {B}}({\mathbf {x} })$ and its off-diagonal elements are responsible for correlations in the noise of different strategic types. In the limit $N\to \infty$ the matrix ${\mathcal {C}}({\mathbf {x} })$ vanishes with $\sim 1/{\sqrt {N}}$ and we recover a deterministic replicator mutator equation.

References

1. Traulsen, A., Claussen, J. C. & Hauert, C. (2012) Stochastic differential equations for evolutionary dynamics with demographic noise and mutations. Phys. Rev. E in print.
2. Traulsen, A., Claussen, J. C. & Hauert, C. (2006) Coevolutionary dynamics in large, but finite populations. Phys. Rev. E 74 011901 doi: 10.1103/PhysRevE.74.011901.
3. Traulsen, A., Claussen, J. C. & Hauert, C. (2005) Coevolutionary Dynamics: From Finite to Infinite Populations. Phys. Rev. Lett. 95 238701 doi: 10.1103/PhysRevLett.95.238701.