Independent scholarship · Open accessThe monthly digest

Indexed text attachment

evolutionary-game-theory-in-higher-categories.tex

Source for Evolutionary Game Theory in Higher Categories

View paper · Download original file

Lines 981–1080 of 3949. Source text is displayed literally.

981For a matrix \(A\), the assignment \(F(x)=Ax\) evaluates the982expected payoff of each strategy against the population. The983derivation below applies to nonaffine assignments as well.984Taylor and Jonker give an early dynamical treatment, and985Alexander explains the abundance and revision interpretations986of the equation \cite{taylor-jonker,alexander}.987 988Suppose \(n_i\) is the abundance of strategy \(i\) and that989\(\dot n_i=n_iF_i(x)\), where \(x_i=n_i/\sum_j n_j\).990Writing \(M=\sum_j n_j\) and \(\phi(x)=\sum_jx_j F_j(x)\),991one obtains \(\dot M=M\phi(x)\). The quotient rule gives992\begin{equation}993  \dot x_i = x_i\bigl(F_i(x)-\phi(x)\bigr).994  \label{eq:replicator}995\end{equation}996The subtraction of mean fitness is thus the correction due to997normalization. Conversely, if \(x\) solves \eqref{eq:replicator},998then \(M(t)=M(0)\exp(\int_0^t\phi(x(s))\,ds)\) recovers a999positive total abundance and \(n_i=Mx_i\) solves the abundance1000equation. The abundance formulation uses independence of the1001fitness assignment from the total. If the rates depend on total1002abundance, the proportions generally retain that dependence1003and need an additional state variable.1004 1005There are two simple symmetries of \eqref{eq:replicator}.1006Adding \(c(x)\) to every \(F_i(x)\) adds \(c(x)\) to \(\phi(x)\)1007and leaves the field fixed. Relabeling \(S\) transports \(x\)1008and \(F\) together. For an affine assignment, adding the same1009row vector to each row of the payoff matrix is one instance1010of the first symmetry. Equality of vector fields records a1011specific observational equivalence of fitness data.1012 1013\section{The simplex and its faces}1014 1015Assume the fitness assignment extends to a locally Lipschitz1016map on a neighborhood of \(\Delta(S)\). On that neighborhood1017form \(\phi(x)=\sum_j x_j F_j(x)\) without requiring \(\sum_jx_j=1\).1018For a local solution \(x\) set \(m(t)=\sum_jx_j(t)\).1019Then \(\dot m=(1-m)\phi(x)\). If \(m(0)=1\), the integrating1020factor for this scalar equation implies \(m(t)=1\) throughout1021its interval. Independently, each coordinate satisfies1022\begin{equation}1023 x_i(t)=x_i(0)\exp\left(1024     \int_0^t\bigl(F_i(x(s))-\phi(x(s))\bigr)\,ds1025     \right).1026 \label{eq:coordinate-factor}1027\end{equation}1028This identity includes coordinates starting at zero, because1029the scalar linear equation along a given solution has the1030zero solution for such an initial value.1031 1032\begin{proposition}[Existence and restriction to faces]1033For the locally Lipschitz assignment just specified,1034\eqref{eq:replicator} has a unique solution for every initial1035state of \(\Delta(S)\), defined for all real times. Throughout1036any finite time interval its support equals the initial1037support. On the face indexed by a nonempty subset \(T\subseteq S\),1038the solution is the replicator solution for the restricted1039assignment on \(\Delta(T)\).1040\end{proposition}1041 1042\begin{proof}1043Local existence and uniqueness follow from the locally1044Lipschitz field. Conservation of mass and1045\eqref{eq:coordinate-factor} keep a solution in the simplex1046throughout its interval of existence, in either direction1047of time. A solution confined to this compact subset of the1048domain can be continued at each finite endpoint. Support1049preservation follows from the positivity of the exponential.1050On the face the terms \(x_jF_j(x)\) with \(j\notin T\) vanish1051from the mean, and each missing coordinate has zero velocity.1052Uniqueness identifies the restricted trajectory with the1053solution calculated directly in the face.1054\end{proof}1055 1056In particular, if \(j_T:\Delta(T)\to\Delta(S)\) extends a1057state by zero, and \(F_T\) restricts \(F\circ j_T\) to the1058coordinates in \(T\), the solution operators obey the1059commutative square1060\[1061\begin{tikzcd}[column sep=large]1062\Delta(T)\arrow[r,"\Phi_t^{F_T}"]\arrow[d,hook] &1063\Delta(T)\arrow[d,hook]\\1064\Delta(S)\arrow[r,"\Phi_t^F"] & \Delta(S).1065\end{tikzcd}1066\]1067Here both vertical maps are \(j_T\). The group law1068\(\Phi_{s+t}=\Phi_s\circ\Phi_t\) follows by uniqueness applied1069after a change of time origin. A relabeling yields the same1070square with the vertical maps replaced by permutations.1071These comparisons fix the sense in which the construction1072respects the data of a game. A map between strategy sets that1073identifies distinct types needs separate hypotheses to carry1074the differential equation through the induced map on states.1075 1076A rest state has equal fitness \(\phi(x)\) on its support.1077This condition is sufficient because the velocity of every1078absent type is already zero. A Nash state also bounds the1079payoffs of absent types above by \(\phi(x)\). The distinction1080is caused by support preservation. In the interior the two