Indexed text attachment
evolutionary-game-theory-in-higher-categories.tex
Source for Evolutionary Game Theory in Higher Categories
View paper · Download original file
Lines 1481–1580 of 3949. Source text is displayed literally.
1481 \label{eq:pairwise-drift}1482\end{equation}1483Each individual transfer contributes with opposite signs1484to the two relevant coordinates. Summing therefore gives1485zero, and the rate at an absent coordinate is nonnegative.1486If \(F\) is locally Lipschitz, the incoming-term variation1487argument of Chapter~\ref{ch:replicator} shows that the1488resulting equation has a unique global forward trajectory1489from each simplex state.1490 1491\begin{proposition}[Rest states under comparison]1492A state is a rest point of \eqref{eq:pairwise-drift}1493if and only if it is Nash.1494\end{proposition}1495 1496\begin{proof}1497Reindexing the sum of the outgoing terms gives1498\[1499 F(x)\cdot W(x)=1500 \sum_{i,j} x_j\rho(F_i(x)-F_j(x))1501 (F_i(x)-F_j(x))\ge0.1502\]1503Every summand is nonnegative, and it is positive1504exactly when \(x_j>0\) and \(F_i(x)>F_j(x)\).1505At a rest point no such pair exists. Thus every used1506strategy maximizes payoff, so \(x\in B(x)\).1507Conversely, if \(x\in B(x)\), every transition away1508from a used strategy has rate zero. Both flux terms1509vanish.1510\end{proof}1511 1512This is the positive-correlation argument for pairwise1513comparison dynamics \cite{sandholm-pairwise}. It also yields1514a strict increase of the potential along nonstationary1515paths when the payoff one-form is exact. In a related1516protocol, sample the candidate from the current population1517and use \([F_i-F_j]_+=\max(F_i-F_j,0)\) as the switch1518rate conditional on the candidate. The flux from \(j\) to1519\(i\) is then \(x_j x_i[F_i-F_j]_+\).1520Combining it with its reverse flux and using1521\([a]_+-[-a]_+=a\) recovers \eqref{eq:replicator}.1522This latter sampling convention cannot create an absent1523type. The state-independent availability of candidates is1524the hypothesis that makes the preceding characterization1525include the boundary Nash inequalities.1526 1527\section{Logit revision and regularization}1528 1529Fix a noise parameter \(\eta>0\). The logit response is1530\[1531 B^\eta_i(x)=\frac{\exp(F_i(x)/\eta)}1532 {\sum_j\exp(F_j(x)/\eta)} .1533\]1534The smooth response map from fitness vectors to probabilities1535preserves common payoff baselines and is equivariant under1536strategy relabeling. For a bijection \(S\to S'\) with1537permutation map \(P\) the latter statement is1538\[1539\begin{tikzcd}[column sep=large]1540\mathbb R^S \arrow[r,"\operatorname{softmax}_{\eta}"]\arrow[d] &1541 \Delta(S)\arrow[d]\\1542\mathbb R^{S'}\arrow[r,"\operatorname{softmax}_{\eta}"]&1543 \Delta(S').1544\end{tikzcd}1545\]1546The two vertical maps apply \(P\). The result follows because1547the normalization is a sum over the strategy set. A1548many-to-one map need not commute with the same response,1549since several copied entries contribute several terms to1550that sum.1551 1552In detail, let \(q\colon S'\twoheadrightarrow S\) be a1553surjection and pull a payoff vector \(f\) back to \(f\circ q\).1554The mass assigned to the fiber of \(i\) is proportional to1555\(\lvert q^{-1}(i)\rvert\exp(f_i/\eta)\). The effective1556coarse payoff includes a shift of \(\eta\log\lvert1557q^{-1}(i)\rvert\), which cancels from the response when1558the fiber sizes agree. More generally, using positive1559reference weights \(w_j\) on \(S'\) makes the mass of a fiber1560proportional to1561\(\exp(f_i/\eta)\sum_{j:q(j)=i} w_j\).1562The weighted response therefore commutes with lumping for1563pulled-back fitness vectors when one pushes the reference1564weights forward. To extend this calculation to a1565semiconjugacy of the revision equations, the fine fitness1566must be constant within each fiber and must be computed1567from the coarse state by the same rule at every instant.1568 1569One can derive the logit rule by regularizing the response1570optimization. Let \(H(y)=-\sum_i y_i\log y_i\), using the1571continuous extension \(0\log0=0\). For fixed \(x\), write1572\(Z=\sum_j\exp(F_j(x)/\eta)\). Substituting the definition1573of \(B^\eta\) gives1574\[1575 y\cdot F(x)+\eta H(y)1576 =\eta\log Z-\eta D(y\Vert B^\eta(x)).1577\]1578The identity is valid also for \(y\) on a proper face because1579the second argument of entropy has full support. By the1580nonnegativity established in Chapter~\ref{ch:lyapunov}, the