Indexed text attachment
evolutionary-game-theory-in-higher-categories.tex
Source for Evolutionary Game Theory in Higher Categories
View paper · Download original file
Lines 1201–1300 of 3949. Source text is displayed literally.
1201Set \(U_p=\{x\in\Delta(S):x_i>0\text{ whenever }p_i>0\}\).1202The quantity1203\[1204 D(p\Vert x)=\sum_{i:p_i>0}p_i\log\frac{p_i}{x_i}1205\]1206is finite on \(U_p\), including when \(p\) lies in a proper1207face. The bound \(-\log z\ge1-z\) at \(z=x_i/p_i\) yields1208\[1209 D(p\Vert x)\ge1-\sum_{i:p_i>0}x_i\ge0.1210\]1211Equality demands \(x_i=p_i\) on the support of \(p\), which1212also forces the remaining entries of \(x\) to vanish.1213The derivative along \eqref{eq:replicator} is1214\begin{equation}1215 \frac{d}{dt}D(p\Vert x(t))1216 = x(t)\cdot F(x(t))-p\cdot F(x(t)).1217 \label{eq:kl-dissipation}1218\end{equation}1219The cancellation uses \(\sum_i p_i=1\). Support preservation1220makes the logarithms legitimate for all finite times1221whenever they are defined initially.1222 1223\begin{theorem}[Local stability from superiority]1224\label{thm:superiority-stability}1225For the replicator equation of Chapter~\ref{ch:replicator},1226a locally superior state \(p\) is a locally asymptotically1227stable rest point relative to \(\Delta(S)\).1228\end{theorem}1229 1230\begin{proof}1231For any \(q\in\Delta(S)\), approach \(p\) along the segment1232\(x=p+\epsilon(q-p)\). Dividing the superiority inequality1233by \(\epsilon>0\) and taking its limit shows1234\((q-p)\cdot F(p)\le0\). Thus \(p\) is Nash and is a rest1235point of the replicator field. Take a closed relative1236ball around \(p\) contained in \(U_p\) and in the neighborhood1237where superiority holds. Taking the ball small enough1238is possible even when \(p\) lies on the boundary because1239only its positive coordinates need to stay positive.1240If the simplex is a singleton, stability is immediate.1241Otherwise, for a sufficiently small radius the relative1242sphere is nonempty, and the minimum there of \(D(p\Vert-)\)1243is positive by compactness. Any trajectory starting inside1244with entropy below this minimum stays inside. By taking the1245radius arbitrarily small we obtain Lyapunov stability.1246 1247For such a trajectory, the entropy is nonincreasing and1248bounded below. If a point \(z\ne p\) were a limit point,1249its entropy derivative would be strictly negative.1250The trajectory has bounded speed on the closed ball.1251It would therefore spend intervals of a fixed positive1252duration where the derivative is bounded above by a1253negative constant, on arbitrarily late visits near \(z\).1254Taking disjoint visits contradicts convergence of the1255entropy values. The limit points are consequently all \(p\).1256The trajectory lies in a compact set and hence converges1257to \(p\).1258\end{proof}1259 1260The same argument supplies a global conclusion on \(U_p\)1261if \(p\) is a Nash state and1262\[1263 (x-p)\cdot\bigl(F(x)-F(p)\bigr)<01264 \qquad(x\ne p)1265\]1266on the whole simplex. The Nash inequality implies1267\((x-p)\cdot F(x)<0\). A finite sublevel set of1268\(D(p\Vert-)\) stays away from zeros of the required1269coordinates, so the compact-set argument applies there.1270For completeness, each term with \(p_i>0\) is bounded below1271by \(p_i\log p_i\), while the term for \(x_i\) tends to1272infinity at \(x_i=0\). This gives positive lower bounds1273for the required coordinates on any one sublevel set.1274This strict inequality is a special sufficient condition.1275Hofbauer and Sandholm establish broader geometric criteria1276for convergence in stable population games1277\cite{hofbauer-sandholm-stable}.1278 1279\section{Selection as a metric gradient}1280 1281At an interior state \(x\), tangent vectors are vectors \(u\)1282with \(\sum_i u_i=0\). The Shahshahani metric is1283\[1284 g_x(u,v)=\sum_i\frac{u_i v_i}{x_i}.1285\]1286The same definition applies to the relative interior of1287each face by restricting the coordinates to its support1288\cite{shahshahani,hofbauer-sigmund}. Define the fitness1289one-form on a face by \(\alpha_x(u)=\sum_iF_i(x)u_i\).1290If \(R_F\) denotes the replicator vector field, its1291metric pairing with \(u\) is1292\[1293 g_x(R_F(x),u)=\sum_i(F_i(x)-\phi(x))u_i=\alpha_x(u).1294\]1295The field is the metric dual of the payoff one-form.1296Adding a state-dependent baseline to all payoffs leaves1297the one-form unchanged, because all tangent vectors1298have zero sum.1299 1300Pointwise, every tangent covector has a representative