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 5139–5238 of 6559. Source text is displayed literally.

5139(\cref{sec:thompson}). Applied to bisimilar types, it separates interim correlated rationalizability from5140Bayesian equilibrium (\cref{ex:redundant}). Applied to equivalent continuation games, it is the Markov5141restriction, and it shows that the folk theorem's cooperation is a focal structure on histories5142(\cref{prop:markov}). Four literatures, developed separately over thirty years, are one requirement and5143four choices of type.5144 5145\emph{One modality.} Knowledge, common knowledge, distributed knowledge and subgame perfection are all of5146the form $p^*\forall_p$ or its relatives for a suitable map $p$ (\cref{def:K,thm:aumann-ck,prop:box}), and5147sequential rationality is their composite.5148 5149\emph{One obstruction theory.} The existence of a common prior for given posteriors, and the existence of5150incentive-compatible payments for a given allocation, are both the triviality of a 1-cocycle; uniqueness in5151both is connectedness; and the local-to-global theorems of mechanism design are the statement that the5152obstruction lives in loops (\cref{thm:H1,thm:rochet}). The impossibility theorems of Holmstr\"om, of Green5153and Laffont, and of Myerson and Satterthwaite are computations with the resulting potentials.5154 5155\emph{Two universal properties.} Finite-horizon games are elements of an initial algebra and5156infinite-horizon games of a terminal coalgebra. Backward induction, the one-shot deviation principle,5157self-generation, the automaton representation of strategies, recursive contracts, the universal type space5158and rationalizability are instances of recursion and corecursion (\cref{ch:perfection,ch:repetition,ch:belief}).5159Whether the two agree is governed by whether the relevant operator is a contraction, and this one5160distinction separates the theory's uniqueness results from its multiplicity results (\cref{sec:banach}).5161 5162\emph{One algebra of composition.} Normal-form games, sequential games, principal--agent problems,5163rational-expectations pricing and signal jamming are composites of the same few boxes in different orders,5164and the forward/backward discipline of optics is what distinguishes a deviation from a change of5165conjecture (\cref{ch:open,ch:signalling}).5166 5167\section{What was said that could not be said before}5168 5169Less than one might like, and it is as well to be exact.5170 5171Three statements in this book are, as far as we know, not in the literature in the form given: that the5172homotopy colimit of the type maps is a graph whose $\pi_0$ is common knowledge and whose $H^1$ obstructs5173common priors, with Hellman--Samet tightness as its contractibility (\cref{prop:hocolim,thm:H1}); that5174incentive-compatible payment rules are the copresheaves on an enriched category determined by the5175allocation, with revenue equivalence as a groupoid condition (\cref{thm:rochet}); and that the revelation5176principle is an initiality statement resting on the contravariance of the equilibrium property5177(\cref{thm:revelation}). Each is a repackaging of known theorems, with a short proof from them. None5178enlarges the set of facts about games. Their value, if they have any, is in suggesting generalizations,5179and the problems below are meant to test that.5180 5181The same must be said of the word ``higher''. The homotopy types that appeared are graphs. The categories5182that appeared are 1-categories, a bicategory and a double category of open games, and categories enriched5183in posets. Where genuinely higher structure is visible---the homotopy types of equilibrium sets, which by5184Datta's theorem are arbitrary; the simplicial models of many-agent knowledge; the $\infty$-toposes in which5185smooth and probabilistic games ought to live---this book has pointed and not built. A reader who came for5186$(\infty,1)$-categories has been given the 1-truncation of a programme.5187 5188\section{What the language does not do}5189 5190\emph{It does not solve games.} Composition of open games composes the equilibrium check, not the search5191(\cref{ch:open}). Computing equilibria is hard for reasons no notation removes (\cref{ch:mixing}).5192 5193\emph{It does not select equilibria.} The invariance principle says what a selection must respect, and5194thereby that many proposed selections are not well defined; it does not supply one. The structure theorem5195says there is no continuous one.5196 5197\emph{It does not adjudicate modelling choices.} Whether the order of moves is part of the game, whether5198histories may serve as frames, whether types are hierarchies or something finer: the type-theoretic5199discipline forces these choices to be declared. It does not make them.5200 5201\emph{It does not replace the examples.} Every structural statement in this book was extracted from models5202that economists built for particular purposes, and the worked examples carry content that the general5203statements do not.5204 5205\section{Problems}5206 5207The following seem to us well posed, open, and of a kind that the language of this book makes natural.5208 5209\begin{problem}[A univalent type of extensive games]\Spg5210\Cref{sec:thompson} defined $\EFG_{\mathrm{T}}$ as a quotient. Find a notion of structure, intrinsic rather5211than presented, whose structure-preserving equivalences are exactly the Thompson equivalences, so that5212invariance holds by the structure identity principle rather than by decree. The set-theoretic5213reformulations of the extensive form in terms of plays and choices \cite{AFR16} are a natural starting5214point; the reduced strategic form is the trivial answer, and the problem is to find one that retains5215enough dynamic structure to state sequential rationality.5216\end{problem}5217 5218\begin{problem}[Perfection as an operation on open games]\Spg5219The modality $\Box$ of \cref{prop:box} was defined for closed extensive games. Define it on the category of5220open games, as an operation on equilibrium predicates, so that $\Box(\mathcal{H}\circ\mathcal{G})$ is5221computed from $\Box\mathcal{H}$ and $\Box\mathcal{G}$, recovering \cite{Hed16,GKLF18} as special cases, and5222extend it to information sets with beliefs so as to obtain a compositional sequential equilibrium for the5223Bayesian open games of \cite{BHZ19}.5224\end{problem}5225 5226\begin{problem}[The exit-path category of the space of games]\Spg5227The equilibrium correspondence is semialgebraic, so the space $\mathcal{U}$ of games with given action sets5228admits a stratification over each stratum of which the projection5229$\mathrm{pr}\colon\mathcal{E}\to\mathcal{U}$ of \cref{sec:KM} is a fibre bundle. The fibres then assemble into a5230functor from the exit-path $\infty$-category of the stratification to spaces. Describe this functor for the5231smallest nontrivial cases. Characterize the essential components and the stable sets of \cite{KM86,GW05} as5232subfunctors with a universal property. \Cref{rem:truncated} suggests that a refinement in the proper sense5233is a homotopy-invariant of this functor; make that a definition and compare it with the axiomatic5234characterizations.5235\end{problem}5236 5237\begin{problem}[Higher cohomology of information]\Spg5238For $n\ge3$ agents the simplicial model $\mathcal{C}$ of \cref{ch:knowledge} has higher homotopy that the