Indexed text attachment
evolutionary-game-theory-in-higher-categories.tex
Source for Evolutionary Game Theory in Higher Categories
View paper · Download original file
Lines 2029–2128 of 3949. Source text is displayed literally.
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 2049\begin{proof}2050The coordinate martingales of the generator give the vector decomposition2051\[2052 X_N(t)=X_N(0)+\int_0^t F_N(X_N(s))\,ds+M_N(t).2053\]2054For the squared Euclidean norm of the martingale, the compensator of the sum of squared jumps contributes at most \(BT/N\) in expectation up to time \(T\). In detail, the jump \(v/N\) contributes \(\|v\|^2/N^2\) at rate \(N\beta_v^N(X_N(s))\). Their integrated sum bounds \(\mathbb E\|M_N(T)\|^2\). Thus Doob's quadratic inequality, applied to the submartingale \(\|M_N(t)\|\), gives2055\[2056 \mathbb E\sup_{0\leq t\leq T}\|M_N(t)\|^2\leq 4BT/N .2057\]2058Subtracting the deterministic integral equation and using the Lipschitz bound yields, by Gr\"onwall's inequality,2059\[2060 \sup_{0\leq t\leq T}\|X_N(t)-x(t)\|2061 \leq e^{LT}2062 \left(\|X_N(0)-x_0\|+T\delta_N+2063 \sup_{0\leq t\leq T}\|M_N(t)\|\right).2064\]2065Every term in parentheses tends to zero in probability.2066\end{proof}2067 2068The estimate is quantitative once drift and initial-state errors are bounded. The factor \(e^{LT}\) also locates the difficulty of extending the comparison to long time. The theorem controls complete paths on each fixed interval. It does not compare stationary distributions, which involve behavior on a time scale that can increase with \(N\).2069 2070For deterministic initial states with error \(a_N=\|X_N(0)-x_0\|\), choose any \(\eta>e^{LT}(a_N+T\delta_N)\). The proof gives2071\[2072 \mathbb P\!\left(\sup_{t\leq T}\|X_N(t)-x(t)\|>\eta\right)2073 \leq2074 \frac{4BT}{N\bigl(e^{-LT}\eta-a_N-T\delta_N\bigr)^2}.2075\]2076The expression separates the initial discrepancy and the systematic drift error from the probability of a fluctuation. Different microscopic kernels can share the same limit \(F\) while differing in \(B\) or in the conditional covariance from Chapter~\ref{ch:markov}.2077 2078For an update chain with transition matrix \(P_N\), an independent Poisson clock of rate \(N\) gives the generator \(N(P_N-I)\). The chain is evaluated at the number of clock arrivals by time \(t\), so a unit of macroscopic time corresponds on average to \(N\) updates. A self-loop consumes a clock arrival and produces a zero displacement. This convention makes the drift of the continuous-time chain \(N\) times the mean increment in one discrete update. It also preserves stationary laws because \(\pi N(P_N-I)=0\) is equivalent to \(\pi P_N=\pi\).2079 2080\section{Recovering deterministic selection}2081 2082Use the birth--death update of Chapter~\ref{ch:markov} at rate \(N\). Its mean field is \(c(n)-x\) by Equation~\eqref{eq:birth-death-drift}. Suppose the weights of present types converge uniformly to bounded strictly positive functions \(f_i(x)\), with a common positive lower bound. The limiting offspring fractions and field are2083\[2084 c_k(x)=\sum_i2085 \frac{x_i f_i(x)}{\overline f(x)}M_{ik},2086 \qquad2087 F_k(x)=c_k(x)-x_k,\qquad2088 \overline f(x)=\sum_i x_i f_i(x).2089\]2090If the weights are Lipschitz, the field is Lipschitz. Its components sum to zero, and \(c_k(x)\geq0\) at a face with \(x_k=0\). A direct form of the integral equation is2091\[2092 x_k(t)=e^{-t}x_k(0)+2093 \int_0^t e^{-(t-s)}c_k(x(s))\,ds .2094\]2095This exhibits the role of the offspring distribution. Invariance of the simplex also follows by constructing the solution locally as the fixed point of this integral map on paths in the simplex, where it is a contraction on short intervals and maps such paths to such paths. Short intervals can be concatenated because the simplex is compact and the constants can be chosen uniformly. With no mutation, \(F_k(x)=x_k(f_k(x)-\overline f(x))/\overline f(x)\). Because \(\overline f(x)>0\), this has the replicator trajectories after a positive change of time. The rates with respect to calendar time differ by this common factor. With mutation, normalization and transmission are combined in \(c\) and need not give a particular additive mutation convention on the replicator equation.2096 2097The distinction between finite and limiting payoffs can be bounded explicitly. For \(n_i>0\) the payoff in Chapter~\ref{ch:markov} satisfies2098\[2099 u_i^N(n)-(Ax)_i=\frac{(Ax)_i-A_{ii}}{N-1}.2100\]2101If \(f_i(n)=\phi(u_i^N(n))\) with a Lipschitz positive map \(\phi\) on the interval covered by the payoffs, the difference from the weight \(\phi((Ax)_i)\) is \(O(N^{-1})\) uniformly over present types. The absent types contribute zero to the birth distribution. Positivity on the compact payoff interval gives a uniform lower bound for the normalization, so the corresponding drift approximation has order \(N^{-1}\). The martingale estimate controls the stochastic error separately at order \(N^{-1/2}\) in root mean square over fixed times.2102 2103\section{When counts close over a graph}2104 2105Suppose the population is an \(N\)-vertex weighted graph. For an oriented pair of vertices \(u,v\), let \(\lambda^N_{uv}\geq0\) be a sampling weight with \(\sum_{u,v}\lambda^N_{uv}=1\). A type \(i\) at \(u\) replaces a type \(j\) at \(v\) at rate \(N\lambda^N_{uv}\varphi(i,j)\), for fixed nonnegative bounded biases \(\varphi\). The relevant statistic is the oriented edge frequency2106\[2107 \theta^N_{ij}(z)=\sum_{u,v}\lambda^N_{uv}2108 \mathbf1_{\{z(u)=i,\ z(v)=j\}}.2109\]2110The density drift on a labeled state \(z\) is2111\(\sum_{i,j}\theta^N_{ij}(z)\varphi(i,j)(e_i-e_j)\), where \(e_i\) is the coordinate unit at \(i\). For a general graph the edge frequencies can vary when the vertex frequencies are fixed. A closed equation for vertex frequencies requires additional information about their evolution.2112 2113An explicit sufficient bound exists for regular weighted sampling. Suppose \(\lambda^N_{uv}=W_{uv}/N\), with \(W\) a nonnegative matrix with row sums one, and write \(J_N\) for the matrix with all entries \(1/N\). If \(\mathbf 1_i\) is the vertex indicator of type \(i\), then2114\[2115 \left|\theta^N_{ij}(z)-x_i x_j\right|2116 = \frac1N\left|\mathbf1_i^{\mathsf T}(W-J_N)\mathbf1_j\right|2117 \leq \|W-J_N\|_{2\to2}\sqrt{x_i x_j}.2118\]2119Thus \(\|W-J_N\|_{2\to2}\to0\) is sufficient for uniform closure of the drift. The proof of Theorem~\ref{thm:egt-fluid} still applies to the projected coordinate processes, using the generator on labeled configurations to obtain their compensators. Uniform drift approximation substitutes for the exact Markov property of the projection.2120\[2121 \begin{tikzcd}[column sep=4.65em]2122 \text{labeled state}\arrow[r,"{\text{observe}}"{above}] &2123 (x,\theta^N)\arrow[r,"{\text{retain counts}}"{above}] & x2124 \end{tikzcd}2125\]2126With uniform sampling of distinct ordered vertices, \(W=(N J_N-I)/(N-1)\). Then \(W-J_N=(J_N-I)/(N-1)\) and its norm is \(1/(N-1)\) for \(N\geq2\). For any graph family satisfying the norm bound, the limiting field is2127\[2128 F(x)=\sum_{i,j}x_i x_j\varphi(i,j)(e_i-e_j).