Indexed text attachment
evolutionary-game-theory-in-higher-categories.tex
Source for Evolutionary Game Theory in Higher Categories
View paper · Download original file
Lines 741–840 of 3949. Source text is displayed literally.
741to preserve the payoff gaps to all absent types.742Any Nash state sufficiently near \(p\) for a sufficiently743small perturbation still has positive coordinates on744\(S\), and the same payoff gaps prevent extra positive745coordinates outside \(S\). Its uniqueness therefore follows746from the restricted affine equation. Here the strict gaps747and the curvature on the face control different aspects of748the same continuation.749\par750\chapter{Payoff equivalence and spaces of games}751\label{ch:payoff-equivalence}752 753The payoff array contains information about the absolute score754of each encounter. The static comparisons in Chapter~\ref{ch:ess}755and the usual replicator field use only relative scores against756the same population. This difference suggests an equivalence757relation on games with fixed type labels. Once labels are also758allowed to vary, the resulting space of games has a groupoid759presentation. Its arrows identify numerical presentations of760the same relative-fitness model.761 762\section{The quotient by common fitness increments}763 764Fix a nonempty finite type set \(I\), put \(n=|I|\), and765write \(\mathbf 1\) for its constant vector. Let766\[767 \mathcal A_I=\operatorname{End}(\mathbb R^I),768 \qquad W_I=\{\mathbf 1 b^{\mathsf T}:b\in\mathbb R^I\}.769\]770If \(B=A+\mathbf 1 b^{\mathsf T}\), then771\[772 Bx=Ax+(b^{\mathsf T}x)\mathbf 1 .773\]774The increment is common to every row type, though its775size can depend on the opponent distribution \(x\).776For \(p,q,x\) in the simplex we have777\((p-q)^{\mathsf T}(B-A)x=0\), since \(p,q\) have the778same total mass. Thus comparing \(p\) and \(q\) against779\(x\) gives identical advantage in the two games.780This includes both comparisons in the ESS test.781 782Define the zero-mean projection783\(\Pi_I v=v-\mathbf 1(\sum_i v_i)/n\). Its kernel784is the span of \(\mathbf 1\) and its image is \(T_I\).785The map \(A\mapsto\Pi_I A\) has kernel \(W_I\):786each column is killed precisely when that column is787constant across the rows. It maps onto the linear maps788\(\mathbb R^I\to T_I\), since the projection acts789as the identity on their image. It follows that790\[791 \mathcal Q_I=\mathcal A_I/W_I792 \simeq\operatorname{Hom}(\mathbb R^I,T_I).793\]794Each class has a canonical representative with zero795column sums. Alternatively, knowing the relative scores796against every population state determines this class.797Indeed the relative scores give all row differences798in \((A-B)x\). If these vanish for all \(x\), evaluation799at each pure state shows \(A-B\in W_I\).800 801\begin{proposition}[The quotient determined by replicator motion]802\label{prop:replicator-payoff-quotient}803Write \(X_A\) for the polynomial vector field on the804simplex given by805\[806 (X_A(x))_i=x_i\bigl((Ax)_i-x^{\mathsf T}Ax\bigr).807\]808Then \(X_A=X_B\) on the whole simplex if and only if809\(A-B\in W_I\). If \(B=\lambda A+\mathbf 1b^{\mathsf T}\)810for constant \(\lambda>0\), then \(X_B=\lambda X_A\)811and the static stability predicates agree.812\end{proposition}813 814\begin{proof}815The common fitness increment cancels in the expression816for \(X_A\) because the weights of \(x\) add to one.817Scaling every fitness by \(\lambda\) scales both terms818in the difference and every payoff comparison by the819same positive number. This proves the forward formulas.820 821Conversely, take \(x\) in the relative interior of the822simplex. Each \(x_i>0\), so equality of the fields823implies \(((A-B)x)_i=x^{\mathsf T}(A-B)x\) for all \(i\).824Thus \(\Pi_I(A-B)x=0\) on the relative interior.825Continuity gives zero on its closure, and in particular826on every pure state. The column argument then places827\(A-B\) in \(W_I\).828\end{proof}829 830If a solution curve for \(X_A\) is parameterized by831\(t\), changing its parameter to \(\lambda t\) gives a832solution for \(X_B\) on the corresponding time interval.833Positive scaling therefore gives a comparison of time834units. Keeping time fixed gives the sharper quotient835\(\mathcal Q_I\). Neither equivalence identifies the836unadjusted absolute payoffs. Those retain a role if837another process, such as a baseline contribution to838fitness, is sensitive to their absolute levels.839 840\section{Relabeling as a groupoid action}