Indexed text attachment
evolutionary-game-theory-in-higher-categories.tex
Source for Evolutionary Game Theory in Higher Categories
View paper · Download original file
Lines 1949–2048 of 3949. Source text is displayed literally.
1949\[1950 P^\epsilon(x,y)1951 =c_{xy}\epsilon^{r_{xy}}+o(\epsilon^{r_{xy}}),1952 \qquad c_{xy}>0,\quad r_{xy}\geq 0 .1953\]1954Call \(r_{xy}\) the resistance. Omit identically zero edges and set1955\[1956 \gamma_i=\min_{\tau\text{ rooted at }i}1957 \sum_{(x,y)\in\tau}r_{xy}.1958\]1959For each root define \(a_i>0\) by summing \(\prod_{(x,y)\in\tau}c_{xy}\)1960over its minimizers. Each tree product has exponent given by the sum of its edge resistances. The tree sum is finite and its coefficients at the smallest exponent are positive. Hence1961\[1962 W_i(P^\epsilon)1963 =\epsilon^{\gamma_i}(a_i+o(1)).1964\]1965Writing \(\gamma_*=\min_i\gamma_i\), the stationary mass tends to zero for \(\gamma_i>\gamma_*\) and tends to \(a_i/\sum_{\gamma_j=\gamma_*}a_j\) for \(\gamma_i=\gamma_*\). A state with positive limiting mass is called stochastically stable. Thus the minimum resistance determines the stable set, and the coefficients determine the limiting weights within it. The leading powers depend on the revision and mutation mechanism. An ordering of payoffs alone does not provide them.1966 1967This argument is a direct finite-state counterpart to resistance and stochastic potential in Young's theory of regular perturbed Markov processes \cite{young-conventions}. Mutation mechanisms with weights decaying faster than all powers may require a different asymptotic scale. Aperiodicity is needed for general convergence from arbitrary initial laws along discrete time. The stationary weights in the tree theorem are defined for every irreducible kernel.1968 1969\section{Eliminating transient excursions}1970 1971The state set of a population chain can be large. An exact reduction separates states recurrent under no mutation from transient excursion states. Suppose \(P^\epsilon\to P^0\) and write \(E=K\amalg T\), where \(K\) is the set of recurrent states of \(P^0\) and \(T\) is the set of transient states. Define successive visits to \(K\) by counting each ordinary time step spent in \(K\), including consecutive time steps there. With matrix blocks indexed by the starting and ending sets, this trace kernel is1972\begin{equation}1973 \widehat P^\epsilon1974 = P^\epsilon_{KK}1975 + P^\epsilon_{KT}(I-P^\epsilon_{TT})^{-1}P^\epsilon_{TK}.1976 \label{eq:trace-kernel-egt}1977\end{equation}1978If \(T\) is empty, only the first term is needed. The geometric series for the inverse sums incomplete excursions according to the number of steps that stay in \(T\). For each \(m\geq0\) the term with \(m\) such steps follows the block sequence1979\[1980\begin{tikzcd}[column sep=large]1981 K \arrow[r,"P^\epsilon_{KT}"{above}] &1982 T \arrow[r,"(P^\epsilon_{TT})^m"{above}] &1983 T \arrow[r,"P^\epsilon_{TK}"{above}] &1984 K .1985\end{tikzcd}1986\]1987Transience makes the series converge at \(\epsilon=0\), since the probability of staying in \(T\) decays geometrically in a finite chain. The inverse therefore exists and is continuous for sufficiently small \(\epsilon\).1988 1989The stationary equation on \(T\) is1990\(\pi_T^\epsilon=\pi_K^\epsilon P^\epsilon_{KT}1991 +\pi_T^\epsilon P^\epsilon_{TT}\).1992Solving it and substituting into the equation on \(K\) gives1993\(\pi_K^\epsilon\widehat P^\epsilon=\pi_K^\epsilon\).1994Thus the trace stationary distribution is the normalized restriction of \(\pi^\epsilon\). The trace is irreducible for positive \(\epsilon\), since an original positive-probability path between visited states decomposes into consecutive trace steps. Moreover, \(P^0_{KT}=0\), because each recurrent class is closed at zero mutation. The same stationary identity then gives \(\pi_T^\epsilon\to0\). This permits computing stochastic stability using only the trace, when its transitions satisfy the requisite asymptotics. A trace records return events in the original chain. Its stationary probabilities describe fractions of visits, and the normalization accounts for the time omitted in excursions.1995 1996The error in dropping transient mass can be estimated before computing the trace stationary law. For a nonnegative matrix \(B\), let \(\|B\|_\infty\) be its maximum row sum. The stationary equation gives1997\[1998 \pi^\epsilon(T)1999 \leq \|P^\epsilon_{KT}\|_\infty2000 \big\|(I-P^\epsilon_{TT})^{-1}\big\|_\infty .2001\]2002The second factor bounds the expected duration of an excursion from any state in \(T\), and remains bounded as \(\epsilon\to 0\) at fixed population size. For example, when the probability of leaving \(K\) in one update is uniformly \(O(\epsilon)\), the total stationary time outside \(K\) is \(O(\epsilon)\).2003 2004\section{Rare steps and fixation}2005 2006If the recurrent classes of \(P^0\) are singleton absorbing states and \(P^\epsilon=P^0+\epsilon R+o(\epsilon)\) entrywise, formula~\eqref{eq:trace-kernel-egt} gives2007\[2008 \widehat P^\epsilon=I+\epsilon2009 \bigl(R_{KK}+R_{KT}(I-P^0_{TT})^{-1}P^0_{TK}\bigr)2010 +o(\epsilon).2011\]2012The off-diagonal first-order terms are direct changes between absorbing states and changes that enter \(T\) once at order \(\epsilon\), then reach an absorbing state according to \(P^0\). The latter factor contains the unperturbed fixation probabilities. If several rare steps are needed, first order alone cannot recover all resistances. More generally the trace need not lump several recurrent states into a single class at positive mutation. One can either keep all states of \(K\), check an exact lumpability criterion as in Chapter~\ref{ch:aggregation}, or invoke a separate asymptotic class reduction with its hypotheses.2013 2014Call the first-order coefficient \(G\). Nonnegativity of off-diagonal trace probabilities and normalization of their rows imply that \(G\) is a continuous-time generator on \(K\). If it is irreducible, the normalized trace stationary laws converge to its unique stationary law. Indeed, divide their stationary equations by \(\epsilon\) after subtracting the identity term. Every accumulation point annihilates \(G\). Irreducibility makes that accumulation point unique, and compactness gives convergence. This argument identifies the scope of the first-order approximation.2015 2016These conclusions address a fixed population size. Chapter~\ref{ch:mean-field} treats a different limiting operation, in which population size increases over a bounded time interval. Neither proof interchanges large population, long time, and vanishing mutation. Such an interchange requires additional bounds, because absorption and mixing time can themselves depend strongly on population size and on mutation.2017\par2018\chapter{Mean-field limits and interaction structure}2019\label{ch:mean-field}2020 2021A deterministic evolution can describe the leading motion of a stochastic population over a bounded time interval. The connection is a convergence statement about processes, with a specified scale for the frequency increments and the update rate. It requires bounds on the microscopic drift and on fluctuations. Kurtz gave a general framework for the approximation of differential equations by jump processes \cite{kurtz-limit}. Bena\"im and Weibull studied the approximation and time scales for evolutionary games \cite{benaim-weibull}. The result below isolates assumptions that can be checked by evaluating the local update law.2022 2023\section{A generator on the frequency simplex}2024 2025Let \(D\subseteq\mathbb R^d\) be a compact convex set and let \(D_N\subseteq D\) be a finite state set. The simplex lattice \(C_N(S)/N\) is the main application. Fix a finite set \(\mathcal V\) of possible jump directions. Write the generator as2026\[2027 \mathcal L_N\varphi(x)2028 =N\sum_{v\in\mathcal V}\beta_v^N(x)2029 [\varphi(x+v/N)-\varphi(x)] .2030\]2031Each intensity \(\beta_v^N\) is nonnegative and is zero whenever the indicated transition is infeasible. Terms of zero intensity are omitted. Suppose2032\[2033 B=\sup_{N,x\in D_N}\sum_v\beta_v^N(x)\|v\|^2<\infty,2034 \qquad2035 F_N(x)=\sum_v\beta_v^N(x)v .2036\]2037The vector \(F_N(x)\) is the instantaneous conditional drift. The rate of quadratic variation from jumps is of order \(N^{-1}\) since each jump squares the factor \(N^{-1}\) and the generator contributes \(N\). If \(F_N\) converges uniformly on its state sets to a field \(F\), Taylor's formula gives \(\mathcal L_N\varphi\to D\varphi[F]\) for a smooth \(\varphi\) with bounded second derivative on a neighborhood of \(D\). The Taylor remainder is at most \(B\sup\|D^2\varphi\|/(2N)\). The convergence of generators on such functions identifies the candidate equation.2038 2039\begin{theorem}[Bounded-time approximation]2040\label{thm:egt-fluid}2041Assume \(F\) is Lipschitz on \(D\) with constant \(L\), and set2042\(\delta_N=\sup_{x\in D_N}\|F_N(x)-F(x)\|\). Suppose \(\delta_N\to0\) and \(X_N(0)\to x_0\) in probability. If the solution of \(\dot x=F(x)\) from \(x_0\) remains in \(D\) on \([0,T]\), then2043\[2044 \sup_{0\leq t\leq T}\|X_N(t)-x(t)\|\longrightarrow02045 \quad\text{in probability}.2046\]2047\end{theorem}2048