Indexed text attachment
evolutionary-game-theory-in-higher-categories.tex
Source for Evolutionary Game Theory in Higher Categories
View paper · Download original file
Lines 3221–3320 of 3949. Source text is displayed literally.
3221\chapter{Selection under stochastic compression}3222\label{ch:stochastic-compression}3223 3224An observer can group strategies by a classification rule.3225A deterministic grouping sends each strategy to one class,3226whereas a randomized classifier sends it to a distribution of3227classes. These two changes of description have different effects3228on replicator dynamics. When fitness is pulled back along the3229classifier, the difference is given by a conditional covariance.3230 3231\section{Columns as conditional distributions}3232 3233Let $I,J$ be nonempty finite sets, and let $A$ be a3234column-stochastic matrix from $I$ to $J$. Its column $a_i$3235is the conditional distribution of the label $j\in J$ given3236a strategy $i\in I$. Write $x\in\Delta(I)$ and $z=Ax$.3237The pushforward of a distribution is $Ax$. The corresponding3238pullback of a payoff vector $\gamma\in\mathbb R^J$ is3239$A^\top\gamma$, since3240\[3241 \langle x,A^\top\gamma\rangle=\langle Ax,\gamma\rangle.3242\]3243This is expectation along a Markov kernel, expressed in3244dual coordinates \cite{fritz-markov}. If $g:\Delta(J)\to\mathbb3245R^J$ is the observed fitness, consider the fine fitness3246$f(x)=A^\top g(Ax)+c(x)\mathbf 1_I$, where $c$ is a3247common payoff offset. This assumption controls the dependence3248of a fine type's payoff on all its possible labels. It must3249be checked in a proposed aggregation.3250 3251For a probability vector $w$, set3252$M(w)=\operatorname{diag}(w)-ww^\top$. The replicator3253field is $R_f(w)=M(w)f(w)$. This form includes the3254constraint that shifts by a common payoff have zero field,3255since $M(w)\mathbf 1=0$. For a channel $A$ define3256\[3257 C_A(x)=\operatorname{diag}(Ax)-A\operatorname{diag}(x)A^\top .3258\]3259 3260\begin{theorem}[Covariance of a compressed replicator field]3261\label{thm:channel-defect}3262Under the payoff hypothesis above,3263\[3264 R_g(Ax)-A R_f(x)=C_A(x)\,g(Ax).3265\]3266The matrix $C_A(x)$ is positive semidefinite. At a given $x$3267the defect vanishes exactly when $g(Ax)$ is constant on the3268support of every column $a_i$ with $x_i>0$.3269In particular, exact pushforward for all $x$ and all3270coarse fitness maps holds precisely when $A$ represents3271a deterministic function $I\to J$.3272\end{theorem}3273 3274\begin{proof}3275The common offset disappears in $M(x)f(x)$. Using3276$Axx^\top A^\top=(Ax)(Ax)^\top$ gives3277\[3278 M(Ax)-A M(x)A^\top=C_A(x).3279\]3280Multiplication by $g(Ax)$ gives the claimed difference.3281For an arbitrary test vector $u\in\mathbb R^J$,3282\begin{align*}3283 C_A(x)&=\sum_i x_i3284 \bigl(\operatorname{diag}(a_i)-a_i a_i^\top\bigr),\\3285 u^\top C_A(x)u&=\sum_i x_i3286 \left(\sum_j (a_i)_j u_j^2-3287 \left(\sum_j(a_i)_j u_j\right)^2\right).3288\end{align*}3289The summands are conditional variances and are nonnegative.3290A variance is zero precisely when $u$ is constant on the3291support of the corresponding column. For a positive3292semidefinite real symmetric matrix $C$, $u^\top C u=0$3293is equivalent to $Cu=0$ by diagonalization.3294Taking $u=g(Ax)$ proves the equality criterion. Every3295deterministic column has zero covariance. Conversely choose3296$x$ concentrated on any specified $i$ and allow a constant3297target fitness taking an arbitrary vector value. Universal3298vanishing then forces every coordinate indicator to have zero3299variance under $a_i$, so every coordinate probability is either3300zero or one. Their sum is one, proving determinism.3301\end{proof}3302 3303Deterministic compression groups types without spreading any3304type across distinct observed labels. Its exactness is the3305coordinate expression of summing replicator derivatives3306over the fibers of a function, when the payoffs agree within3307fibers. In stochastic compression, equality can still hold3308for particular payoffs or particular support faces. It is3309then a condition on the conditional distribution and the3310payoff together.3311 3312\section{Composition of two compressions}3313 3314For a second channel $B:J\to K$ the defects obey3315\[3316 C_{BA}(x)=C_B(Ax)+B C_A(x)B^\top.3317\]3318Indeed, first add and subtract3319$B\operatorname{diag}(Ax)B^\top$ in the expression for3320$C_{BA}(x)$. The two terms are the covariance from the