Independent scholarship · Open accessThe monthly digest

Indexed text attachment

evolutionary-game-theory-in-higher-categories.tex

Source for Evolutionary Game Theory in Higher Categories

View paper · Download original file

Lines 2581–2680 of 3949. Source text is displayed literally.

2581transported function of its possible inputs. Existence of a2582maximizing strategy is again an additional hypothesis.2583 2584Stochastic parallel composition requires a decision about joint2585laws. Tensoring two kernels using independent randomness gives2586the product law conditional on their inputs. That tensor does2587not have the universal property of a cartesian product in2588stochastic maps. A joint distribution is generally not determined2589by its two marginals. For example, a nondegenerate finite law2590paired with itself by sharing its sample and the product of2591two independent copies of the law have the same marginals.2592The continuation may detect their dependence. Consequently2593continual substitution at the level of marginal expected2594payoffs alone can lose information relevant for later decisions.2595Bayesian open games supply additional structure for stochastic2596play, hidden information and best responses2597\cite{bolt-bayesian}. The calculation above describes only the2598forward-kernel and expected-return part of that framework.2599 2600\section{An equilibrium interface and a revision rule}2601 2602Response relations are evaluated at a context rather than at a2603time. To extract a trajectory one supplies a protocol for when2604revisions occur, which states agents observe, and how allowed2605revisions become a state update. For a fixed context, let2606\(B(\sigma)\) denote the set of responses to \(\sigma\).2607If it is nonempty at every incumbent, a deterministic revision2608rule can choose a single \(U(\sigma)\in B(\sigma)\) at each2609incumbent. A fixed point of \(U\) is an equilibrium because its2610selected response is itself. The converse holds if the rule2611retains the incumbent whenever \(\sigma\in B(\sigma)\).2612The relation alone leaves this decision open, as well as2613the choice among other responses and the possible regularity2614of the selector. A protocol can instead attach a probability2615law to responses, which adds another layer of dynamical data.2616 2617In continuous time a smooth population2618vector field \(F\) and its positive rescaling \(aF\) by a smooth2619function \(a\) have the same rest points. On any interval over2620which a solution of one field exists and the scalar remains2621positive, its trajectory can be reparametrized in time to a2622solution of the other. This follows from the chain rule by2623solving for the derivative of the new time parameter to equal2624\(a\) along the trajectory. Quantities measured at specified2625times can nevertheless change. A contextual equilibrium2626criterion does not specify the clock, the distribution of2627samples, or the size of a discrete update. These data enter2628the dynamical interface in the next chapter.2629\par2630\chapter{Open dynamics and network wiring}2631\label{ch:open-dynamics}2632 2633An open dynamical system specifies how its current state changes2634under an input from its surroundings. Its output can supply another2635system with an input. A wiring diagram determines which values2636are passed and, when species or states are identified, where their2637rates of change are combined. Different choices of interface give2638different composition rules.2639 2640\section{Input and output ports}2641 2642Consider a smooth state space \(M\), an input space \(A\), and2643an output space \(B\). Write a module as2644\[2645       v:M\times A\longrightarrow TM,\qquad h:M\to B,2646       \qquad v(x,a)\in T_xM .2647\]2648The state equation is \(\dot x=v(x,a)\) and its output is2649\(h(x)\). A single time-independent wiring of modules indexed2650by \(i\) can be expressed by maps2651\[2652  w:A_{\mathrm{ext}}\times\prod_i B_i\to\prod_i A_i,2653  \qquad2654  r:\prod_i B_i\to B_{\mathrm{ext}} .2655\]2656At state \(\boldsymbol x=(x_i)_i\), put2657\(\boldsymbol b=(h_i(x_i))_i\) and2658\(\boldsymbol a=w(a_{\mathrm{ext}},\boldsymbol b)\).2659The new state velocity has components \(v_i(x_i,a_i)\) and2660the output is \(r(\boldsymbol b)\). A feedback wire can carry2661\(h_i(x_i)\) to the same or another module. As the output depends2662only on the current state, closing such a wire requires no2663simultaneous algebraic solution for the inputs.2664 2665For nested networks the rule is associative: to compute one2666component \(v_i\), evaluate the elementary outputs, substitute2667them into the wiring maps of each enclosing network, and apply2668the resulting input to \(v_i\). Flattening the network substitutes2669the same functions into one composite. Associativity of function2670composition and the canonical comparison of product state spaces2671give equality of the component velocities. The same calculation2672applies to outputs. This is the readout form of the wiring algebra2673developed in \cite{vagner-wiring}. If a readout is also allowed to2674depend on its instantaneous input, a directed loop may require2675a separate condition guaranteeing that its algebraic equations2676have a unique solution.2677 2678\section{Replicator modules}2679 2680For a population on a strategy simplex, expose a state readout