Independent scholarship · Open accessThe monthly digest

Indexed text attachment

games-and-information-in-higher-categories.tex

Source for Games and Information in Higher Categories

View paper · Download original file

Lines 3119–3218 of 6559. Source text is displayed literally.

3119subgames replaced by information sets, each equipped with the belief that turns it into a decision3120problem: perfection and knowledge, composed.3121 3122\begin{example}[Where subgame perfection is silent]\label{ex:seq-entry}3123An entrant chooses \emph{Out}, with payoffs $(0,4)$, or enters at a \emph{small} or a \emph{large} scale. The3124incumbent observes entry but not its scale, and fights or accommodates. Accommodating small entry gives3125$(1,2)$, fighting it $(-1,3)$; accommodating large entry gives $(2,1)$, fighting it $(-2,-1)$. There is no3126proper subgame. Let $m$ be the incumbent's belief that entry is small. Fighting is optimal iff3127$3m-(1-m)\ge2m+(1-m)$, that is $m\ge\tfrac23$. There are two kinds of sequentially rational assessment: the3128entrant enters large, the incumbent accommodates, $m=0$; and the entrant stays out, the incumbent would3129fight, sustained by any belief $m\ge\tfrac23$ at an information set that is never reached. Had fighting3130been worse than accommodating against both scales, no belief would support it, and the second equilibrium3131would disappear although it remains subgame perfect.3132\end{example}3133 3134On the equilibrium path beliefs are determined by Bayes' rule. Off the path, the conditioning event has3135probability zero and Bayes' rule is silent; in the example any $m$ is allowed. The various definitions of3136\term{perfect Bayesian equilibrium} \cite{FT91b} add requirements of the form ``update by Bayes' rule wherever3137possible''. Kreps and Wilson's definition is the cleanest.3138 3139\begin{definition}\Scl[\cite{KW82a}]3140An assessment $(\sigma,\mu)$ is \term{consistent} if it is the limit of assessments $(\sigma^k,\mu^k)$ in3141which $\sigma^k$ gives positive probability to every action and $\mu^k$ is derived from $\sigma^k$ by3142Bayes' rule. A \term{sequential equilibrium} is a consistent, sequentially rational assessment.3143\end{definition}3144 3145\begin{theorem}\Scl[\cite{KW82a,Sel75}]3146Every finite extensive game with perfect recall has a sequential equilibrium. Every trembling-hand perfect3147equilibrium is sequential, and for generic payoffs the two concepts coincide. Every sequential equilibrium3148is subgame perfect.3149\end{theorem}3150 3151\section{The geometry of conditioning}\label{sec:cps}3152 3153Consistency is a closure condition. On the open set of fully mixed strategies, beliefs are a rational3154function of strategies; the consistent assessments are the closure of the graph of that function. Taking3155the closure of the graph of a rational map is a classical operation in geometry---it is a \emph{blow-up}---and3156looking at it this way explains what off-path beliefs are.3157 3158Consider the simplest case: a finite set $\Omega$ and the map sending a fully supported $p\in\Dist\Omega$ to3159its family of conditionals $\bigl(p(\cdot\mid S)\bigr)_{S}$, one for each nonempty $S\subseteq\Omega$.3160 3161\begin{definition}\Scl[\cite{Mye86,Ren55}]3162A \term{conditional probability system} on $\Omega$ is a family $(p_S)_{\emptyset\neq S\subseteq\Omega}$ with3163$p_S\in\Dist(S)$ and $p_T(R)=p_S(R)\,p_T(S)$ for $R\subseteq S\subseteq T$.3164\end{definition}3165 3166\begin{proposition}[Conditional systems are a blow-up]\label{prop:cps}\Srf[\cite{McL89a,Mye86,BBD91}]3167The space of conditional probability systems on a finite set $\Omega$ is the closure of the image of the3168open simplex in $\prod_S\Dist(S)$. It is a compactification of the open simplex in which each face of the closed simplex has3169been replaced by the product of that face with the space of conditional systems on the complementary set3170of states. Its strata are indexed by the ordered partitions $(\Omega_1,\dots,\Omega_k)$ of $\Omega$: a point of that3171stratum is a fully supported distribution on each block, the states of $\Omega_{j+1}$ being infinitely less3172likely than those of $\Omega_j$. Topologically it is a closed ball \cite{McL89a}.3173\end{proposition}3174 3175For three states the picture is \cref{fig:cps}. A sequence of priors converging to the vertex $\delta_1$3176has, along the way, a well-defined ratio $p_2:p_3$, and different sequences converge to different limiting3177ratios. The simplex forgets the ratio; the blow-up remembers it, replacing the vertex by a segment. Each3178corner becomes an edge and the triangle becomes a hexagon. In general the face poset is that of the3179ordered partitions, which is the face poset of the permutohedron.3180 3181\begin{figure}[ht]\centering3182\begin{tikzpicture}[x=1cm,y=1cm]3183  \coordinate (A) at (0,0); \coordinate (B) at (4.4,0); \coordinate (C) at (2.2,3.81);3184  \draw[densely dotted] (A)--(B)--(C)--cycle;3185  \coordinate (a1) at ($(A)!0.2!(B)$); \coordinate (a2) at ($(A)!0.2!(C)$);3186  \coordinate (b1) at ($(B)!0.2!(A)$); \coordinate (b2) at ($(B)!0.2!(C)$);3187  \coordinate (c1) at ($(C)!0.2!(A)$); \coordinate (c2) at ($(C)!0.2!(B)$);3188  \draw[thick,fill=black!5] (a1)--(b1)--(b2)--(c2)--(c1)--(a2)--cycle;3189  \draw[very thick,accent] (a1)--(a2) (b1)--(b2) (c1)--(c2);3190  \node[lab] at (-0.25,-0.2) {$\delta_1$}; \node[lab] at (4.65,-0.2) {$\delta_2$}; \node[lab] at (2.2,4.05) {$\delta_3$};3191  \node[lab] at (2.2,1.25) {$p_1,p_2,p_3>0$};3192  \node[lab,accent,anchor=east] at (0.1,0.75) {$p_2:p_3$};3193  \node[lab,accent,anchor=west] at (4.3,0.75) {$p_1:p_3$};3194  \node[lab,accent,anchor=west] at (2.75,3.45) {$p_1:p_2$};3195\end{tikzpicture}3196\caption{Conditional probability systems on three states. Each vertex of the simplex is blown up to a3197segment recording the limiting ratio of the two vanishing probabilities.}\label{fig:cps}3198\end{figure}3199 3200A consistent system of beliefs is a point of such a blow-up, constrained by the product structure that3201independent trembles of different players impose \cite{McL89b,KR97}. The equivalent descriptions in the3202literature are different coordinate systems on the same space: \emph{lexicographic probability systems}3203\cite{BBD91} list the distributions on the blocks in order; \emph{non-Archimedean probabilities}3204\cite{Ham94,Hal10} are single distributions with values in an ordered field containing infinitesimals, the blocks3205being the orders of magnitude. The last makes the geometry algebraic: a consistent assessment is the3206shadow of a fully mixed strategy profile over a larger ordered field, and sequential rationality is ordinary3207rationality there.3208 3209\begin{remark}[What kind of infinitesimal]3210The infinitesimals needed here are invertible elements of an ordered field, so that ratios of small3211probabilities make sense. They are not the nilpotent infinitesimals of synthetic differential geometry,3212which have no ratios. The appropriate ambient setting is real-closed fields and semialgebraic geometry3213rather than a smooth topos; conversely, the fact that equilibrium conditions are semialgebraic is what3214allows the transfer of results between $\R$ and its non-Archimedean extensions.3215\end{remark}3216 3217\section{Reputation}3218