Indexed text attachment
evolutionary-game-theory-in-higher-categories.tex
Source for Evolutionary Game Theory in Higher Categories
View paper · Download original file
Lines 161–260 of 3949. Source text is displayed literally.
161coordinates at positive \(t\). It is sufficient for small \(t\)162because only finitely many of the positive coordinates can decrease.163 164A map of type sets \(f:I\to J\) pushes a distribution forward by165\[166 (f_*x)_j=\sum_{f(i)=j}x_i.167\]168Its linear extension uses the same formula on all vectors.169If \(g:J\to K\), collecting the terms in the sums indexed by each170fiber of \(g f\) gives \((gf)_*=g_*f_*\). The identity map gives171the identity on distributions. Thus simplex formation is a functor172from finite sets to convex sets and affine maps. It extends on all173sets to \(\mathsf D I\), the set of finitely supported probability174distributions. A map on Dirac states extends uniquely to an affine175map out of \(\mathsf D I\), since every element is a finite176combination \(\sum_i x_i\delta_i\) and its coefficient function177is uniquely determined by \(x\). Accordingly a stochastic178assignment of an output distribution to each type is determined179by its values on types and then acts affinely on populations.180 181There is also a canonical way to combine mixtures of populations.182For \(w\in\mathsf D\mathsf D I\), its barycenter has coordinates183\(\mu_I(w)_i=\sum_x w_x x_i\). Taking the barycenter of a Dirac184distribution on a population returns that population. Mixing185Dirac distributions on individual types returns their original186coefficient distribution. The two orders of flattening three187layers of mixtures agree because in either order the coefficient188of \(i\) is obtained by adding the same products of three weights.189This gives the unit and multiplication of the finite-distribution190monad. The construction uses only finite sums even when \(I\)191is infinite. It differs from fixing a population size, since192the weights may be any nonnegative reals adding to one.193 194\section{Random matching and expected fitness}195 196A symmetric two-player game on \(I\) assigns the row player's197payoff \(A_{ij}\) when the row type is \(i\) and its opponent is198\(j\). The opponent receives \(A_{ji}\) in that encounter.199Symmetry is the exchange of player roles in this description.200No symmetry of the entries of \(A\) is required.201If the players draw types independently with distributions \(p,q\),202the expected payoff is203\[204 u_A(p,q)=\sum_{i,j}p_i A_{ij}q_j=p^{\mathsf T}Aq .205\]206Its two separate affine laws follow by collecting the coefficients207of the product distribution. In particular, a randomly sampled208opponent from a population \(x\) gives the fitness vector \(Ax\),209while the population average is \(\overline u_A(x)=x^{\mathsf T}Ax\).210A correlated matching prescription instead supplies a conditional211distribution of opponents for each row type. If these distributions212are \(q(i,x)\), the corresponding fitness is213\(F_i(x)=\sum_j A_{ij}q(i,x)_j\). General matching rules214therefore need not produce a linear fitness map, even when the215encounter payoffs are fixed.216 217The pairing in \(u_A\) is functorial with respect to probability218aggregation. It uses the transpose of the aggregation map to219transport a score. For a score \(b\in\mathbb R^J\), define220\((f^*b)_i=b_{f(i)}\). Then221\[222 \sum_i x_i(f^*b)_i=\sum_j (f_*x)_j b_j.223\]224This follows by collecting the left-hand terms over each fiber of225\(f\). The formula distinguishes the direction in which226population weights move from the direction in which a fitness227score is read back.228 229\section{Exact maps between games}230 231Consider the category whose objects are pairs \((I,A)\) and whose232maps \((I,A)\to(J,B)\) are type maps \(f:I\to J\) satisfying233\(A_{ij}=B_{f(i),f(j)}\) for every \(i,j\).234Identity maps satisfy the same equation. If \(f\) and \(g\) are235payoff-preserving, the comparison for \(gf\) follows by applying236the two equalities successively. This category describes exact237games with strategy maps. There are useful coarser notions of238comparison, including maps that preserve only payoff239differences, developed in Chapter~\ref{ch:payoff-equivalence}.240 241\begin{proposition}[Fitness under exact aggregation]242\label{prop:exact-fitness}243If \(f:(I,A)\to(J,B)\) is payoff-preserving, then244\[245 Ax=f^*(Bf_*x),\qquad246 u_A(p,q)=u_B(f_*p,f_*q).247\]248Conversely, either the fitness identity at all population states249or the payoff identity at all pairs of distributions implies250payoff preservation.251\end{proposition}252 253\begin{proof}254Regroup the terms of \((Ax)_i=\sum_k B_{f(i),f(k)}x_k\)255over the fibers of \(f\). This gives the claimed fitness identity.256The identity for payoffs then follows from the pairing between257aggregation and score pullback. Conversely, evaluating the first258identity at \(\delta_j\) gives equality in each component \(i\).259Evaluating the second identity at \((\delta_i,\delta_j)\) also260gives equality of entries.