Indexed text attachment
evolutionary-game-theory-in-higher-categories.tex
Source for Evolutionary Game Theory in Higher Categories
View paper · Download original file
Lines 281–380 of 3949. Source text is displayed literally.
281target state can be represented by lifting its mass to selected282types above its support. Such a lift need not be unique.283Expected payoff agrees for all lifts by284Proposition~\ref{prop:exact-fitness}. A selection rule using285the labels of the lifts can still see the difference. Passing286to aggregated population states must therefore be justified for287the particular rules one wants to transport.288 289\section{Parameterized populations}290 291Suppose payoffs vary continuously with an environmental292parameter \(s\) in a space \(S\), giving an array \(A(s)\).293Evaluation \((s,p,q)\mapsto u_{A(s)}(p,q)\) is continuous294by its finite-sum formula. Substituting a map \(r:S'\to S\)295pulls back the family to \(A(r(s'))\), and evaluations agree296after the same substitution. If local descriptions use different297names for the types, a transition is a bijection of the298corresponding type sets. An exact transition then gives the299comparison in Figure~\ref{fig:exact-payoff-map}, and these300comparisons compose because pushforward does. The simplex301may consequently vary as a bundle with transition maps302permuting its coordinates. Expected payoff is then a single303continuous function on pairs of states over the same parameter,304as long as the local payoff matrices obey the transition305comparisons. The groupoid of labels keeps track of both the306state and its change of presentation.307 308This setup will support two further comparisons. Static309stability asks which distributions do well against populations310with a small mutant component. A dynamics asks where311trajectories go after an actual change in the population.312The payoff operation is common to the two questions, and313additional assumptions on the trajectory rule will enter later.314 315The exact payoff condition is closed under316pullback of the type labels. Given \(B\) on \(J\)317and any \(f:I\to J\), there is a unique matrix on318\(I\) for which \(f\) is exact. Its entries are319\(B_{f(i),f(j)}\). Pulling back twice gives the same320matrix as pulling back along the composite. Thus the321payoff spaces form a contravariant system over type322sets, whereas the population spaces form a covariant323system. Their comparison in the payoff square explains324why treating both as covariant under a quotient of325labels can change the model.326\par327\chapter{Evolutionary stability and invasion}328\label{ch:ess}329 330Fix a population simplex \(\Delta(I)\) and an expected payoff331\(u_A\) as in Chapter~\ref{ch:population}. We use the symmetric332Nash condition333\[334 u_A(p,p)\geq u_A(q,p)\qquad(q\in\Delta(I)).335\]336It says that no alternative earns more against \(p\) than a337resident earns against \(p\). An evolutionary stability338test also asks what happens when the alternative changes339the population against which these payoffs are evaluated.340Maynard Smith and Price introduced the stability idea in341the study of animal conflict \cite{maynard-price}. Alexander342develops its different formulations in \cite[Ch.~2]{alexander}.343 344\section{Best replies and the resident population}345 346Let \(M_p=\max_i (Ap)_i\). The value \(u_A(p,p)\) is the347\(p\)-weighted average of the numbers \((Ap)_i\). A pure maximizer348earns \(M_p\) against \(p\), while the average cannot exceed349\(M_p\). Hence \(p\) satisfies the symmetric Nash condition350if and only if \(u_A(p,p)=M_p\). Equivalently, every type in351the support of \(p\) attains \(M_p\). If any type with352positive weight earned less, the weighted average would be353strictly smaller than \(M_p\).354 355This calculation also identifies the complete best-reply356correspondence at \(p\). It is the face spanned by those pure357types at which \(Ap\) reaches its maximum. Strict symmetric358Nash means all other population states \(q\ne p\) earn less359against \(p\). In the mixed strategy simplex this forces360the best-reply face to consist of a single point. In361particular, a resident supported on several types can362satisfy the Nash condition without satisfying its strict363version.364 365\section{The invasion test and the two inequalities}366 367For a resident \(p\), a competitor \(q\ne p\), and a mutant368share \(0<\varepsilon<1\), put369\[370 x_{\varepsilon}=(1-\varepsilon)p+\varepsilon q.371\]372The pair \(p,q\) may consist of mixed strategies. The373parameter \(\varepsilon\) specifies the share of the374competitor in the population. It does not describe the375probabilities within either of those two strategies.376Define the payoff advantage in the resulting population by377\[378 D_{p,q}(\varepsilon)379 =u_A(p,x_{\varepsilon})-u_A(q,x_{\varepsilon}).380\]