Indexed text attachment
evolutionary-game-theory-in-higher-categories.tex
Source for Evolutionary Game Theory in Higher Categories
View paper · Download original file
Lines 2989–3088 of 3949. Source text is displayed literally.
2989\[2990 \|q(x(t))-z(t)\|2991 \le e^{Lt}\|q(x(0))-z(0)\|2992 +\begin{cases}2993 \epsilon(e^{Lt}-1)/L,& L>0,\\2994 \epsilon t,& L=0.2995 \end{cases}2996\]2997The integral equations for the two curves, followed by the2998triangle inequality, bound their distance at \(t\) by its2999initial value plus the integral of \(L\) times the current3000distance and \(\epsilon\). The integral form of Gronwall's3001inequality yields the claim. For polynomial fields, formal3002approximate differential equivalences give more specific3003ways to find and certify reductions \cite{cardelli-lumping}.3004 3005If \(h:N\to P\) compares \(G\) to a third field \(J\), the3006chain rule gives3007\[3008 r_{h\circ q}(x)=Dh_{q(x)}r_q(x)+r_h(q(x)).3009\]3010In particular a bound on \(Dh\) controls the magnification3011of the first reduction's defect under composition. The error3012estimate uses a specified time interval and hypotheses about3013the region visited by the comparison trajectories. Long-time3014conclusions need additional information. For example, if the3015target satisfies the contraction inequality3016\[3017 \langle u-v,G(u)-G(v)\rangle\leq -\lambda\|u-v\|^23018 \qquad (\lambda>0)3019\]3020on the comparison region, the error in Euclidean norm satisfies3021\[3022 \|q(x(t))-z(t)\|3023 \leq e^{-\lambda t}\|q(x(0))-z(0)\|3024 +\frac{\epsilon}{\lambda}(1-e^{-\lambda t}).3025\]3026For positive distance, differentiate its square, apply the3027contraction inequality and Cauchy--Schwarz, and divide by3028twice the distance. At distance zero the upper right3029derivative of the distance is at most \(\epsilon\), since3030the difference of velocities is then the defect itself.3031Integrating the resulting differential inequality gives the3032bound. If the two trajectories remain in the region, the3033bound holds on their whole common forward time domain.3034\par3035 3036\bookpart{V}{Descent, compression, and learning}3037\chapter{Fitness systems over varying contexts}3038\label{ch:contextual-games}3039 3040A population in an ecological or strategic context uses a set3041of available types, a state of its interactions, and a rule for3042evaluating prospective behavior. When contexts vary, these data3043can have nontrivial transition maps. A model can then be a family3044of games without having a single global numbering of the types.3045This chapter uses sheaves to state exactly what must agree between3046presentations.3047 3048\section{An internal strategy object}3049 3050Let $B$ be a smooth manifold and let $p:S\to B$ be a smooth3051locally constant finite covering, with all fibers nonempty.3052Its sheets over an open set give types in that context. The3053set of types is a sheaf of finite sets over $B$. Think of an3054open subset as a stage on which type labels may be observed.3055On a trivializing open set $U$, the probability state space3056over $U$ is $U\times\Delta(I_U)$, where $I_U$ indexes the3057local sheets and3058\[3059 \Delta(I_U)=\left\{x\in\mathbb R^{I_U}:3060 x_i\geq0,\quad \sum_i x_i=1\right\}.3061\]3062On an overlap the sheets are related by a locally constant3063permutation $\sigma_{UV}(b):I_V\to I_U$. The transition3064$(b,x)\mapsto (b,\sigma_{UV}(b)x)$ carries faces to faces3065and preserves both the mass and the nonnegativity conditions.3066The transitions assemble a simplex bundle over $B$. Equivalently,3067the simplex is a sheaf of states, using smooth plots on each3068context. The closed fibers have corners. Smoothness of a3069fitness map on such a fiber means that its local components3070extend smoothly to a neighborhood in the affine span.3071 3072A fitness operator is a section of the relative function object3073from the simplex bundle to the bundle of real type-indexed3074vectors. Concretely it is a collection of functions3075$f_U(b,x)\in\mathbb R^{I_U}$ with3076\[3077 f_U(b,\sigma_{UV}(b)x)=\sigma_{UV}(b)f_V(b,x).3078\]3079This is a descent condition on maps between bundles. It can be3080applied equally to restricted subfamilies of fitness operators.3081For instance a collection of pairwise payoff matrices3082$A_U(b)$ gives the operators $A_U(b)x$ when the matrices satisfy3083$A_U(b)\sigma_{UV}(b)=\sigma_{UV}(b)A_V(b)$. A constant affine shift3084of all payoffs leaves the replicator field unchanged, as in3085Chapter~\ref{ch:payoff-equivalence}. It can be convenient to3086take the corresponding quotient of fitness data before imposing3087compatibility.3088