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 1341–1440 of 3949. Source text is displayed literally.

1341For matrix payoffs, write \(A=S+K\) with \(S\) symmetric1342and \(K\) skew-symmetric. When \(K=0\), one can take1343\(V(x)=x^{\mathsf T}Sx/2\). More generally \(A\) defines a1344potential on the interior precisely when the bilinear form1345\(u^{\mathsf T}Av\) is symmetric on tangent vectors.1346Indeed, the exterior derivative of the payoff one-form1347on constant tangent vectors is1348\(v^{\mathsf T}Au-u^{\mathsf T}Av\). This must vanish1349for a potential, and it suffices because the open1350simplex is convex. Thus a skew part matters for1351integrability only through its restriction to1352tangent directions.1353 1354\section{Dissipation and a conservative obstruction}1355 1356At an interior rest point \(p\) of a matrix game,1357\(Ap=c\mathbf{1}\). Consequently1358\[1359 (x-p)\cdot Ax=(x-p)^{\mathsf T}S(x-p).1360\]1361The right side depends only on the symmetric part1362because \(u^{\mathsf T}Ku=0\). Strict negative1363definiteness of \(S\) on the tangent hyperplane gives1364strict entropy dissipation. The skew part can1365nevertheless change the direction of the velocity.1366For a purely skew field \(A=K\) with \(Kp=0\), one1367instead has \(\phi(x)=0\) and \(p^{\mathsf T}Kx=0\)1368everywhere. Equation~\eqref{eq:kl-dissipation} makes1369\(D(p\Vert x)\) a first integral on the interior.1370Every trajectory from a distinct interior state has1371positive constant entropy, so it cannot converge to1372\(p\). This obstructs asymptotic stability regardless1373of whether individual level sets carry periodic motion.1374 1375Suppose that \(p\) is interior and that there is a strict1376uniform bound of the form1377\((x-p)\cdot F(x)\le -c\lVert x-p\rVert^2\)1378near \(p\), with \(c>0\). The entropy Hessian in the1379ambient positive orthant is1380\(\operatorname{diag}(p_i/x_i^2)\). This Hessian is1381bounded above and below by positive definite diagonal1382matrices on a small closed neighborhood of \(p\). The first1383derivative of entropy at \(p\) vanishes on the tangent1384space. Taylor's theorem therefore gives positive constants1385\(a,b\) such that1386\[1387 a\lVert x-p\rVert^2\le D(p\Vert x)1388       \le b\lVert x-p\rVert^2 .1389\]1390An initial state close enough stays in this neighborhood1391by Lyapunov stability. Then \eqref{eq:kl-dissipation}1392implies exponential decay at rate at least \(c/b\) for1393the entropy. This argument uses the full support of the1394resident. At a boundary resident, redistributing mass to1395an absent strategy can give a linear term in the expansion.1396For a fixed \(q\) with mass \(m\) outside the support of1397\(p\), direct differentiation yields1398\[1399 D(p\Vert p+\epsilon(q-p))1400      =\epsilon m+O(\epsilon^2)1401       \qquad(\epsilon\downarrow 0).1402\]1403The general stability implication accommodates this boundary1404term through the entropy sign and compactness.1405 1406These statements concern the vector field associated1407with a given payoff rule. Equation~\eqref{eq:potential-rate}1408uses the exact derivative of a potential and1409\eqref{eq:kl-dissipation} uses the comparison with a1410resident. They give separate ways to turn a strategic1411inequality into a dynamical conclusion, with distinct1412hypotheses at the boundary and on the rest set.1413\par1414\chapter{Revision protocols and response dynamics}1415\label{ch:revision}1416 1417\section{Response as a differential inclusion}1418 1419A payoff assignment gives current incentives. A revision1420protocol determines which of those incentives are used when1421a strategy changes. This extra datum affects boundary motion,1422the identification of rest points, and convergence.1423Alexander and Sandholm develop the passage from revision1424protocols to population dynamics1425\cite{alexander,sandholm-population}.1426 1427Write1428\[1429 B(x)=\operatorname*{arg\,max}_{y\in\Delta(S)} y\cdot F(x).1430\]1431Because the objective is linear in \(y\), this is the1432nonempty convex face spanned by the maximizers of1433\(F_i(x)\). Best-response revision uses the differential1434inclusion1435\[1436  \dot x(t)\in B(x(t))-x(t).1437\]1438The solution convention is an absolutely continuous path1439satisfying the inclusion almost everywhere. If \(F\) is1440continuous, the graph of \(B\) is closed. In fact, taking a