Independent scholarship · Open accessThe monthly digest

Editorially reviewed

Evolutionary Game Theory in Higher Categories

A categorical treatment of J. McKenzie Alexander's Evolutionary Game Theory (2023). The replicator equation is a functor on games and payoff-preserving maps; stable sets, not stable points, survive duplication of strategies; index, Conley index and Hodge decomposition; open systems; finite populations; coverings explain local-interaction patterns. 59 pp., 28 solved exercises, 42 checks.

Anonymous← All research

Evolutionary Game Theory in Higher Categories

Stability, replicators, recurrence, finite populations and local interaction

Evolutionary game theory replaces the rational player by a population: strategies are inherited or imitated, and the question changes from "what should one do?" to "what persists?". J. McKenzie Alexander's Evolutionary Game Theory (Cambridge Elements in Decision Theory and Philosophy, 2023) is a compact map of the subject, written for philosophers and candid about where the models are fragile. This volume retells that material with one question in mind: which constructions are natural, in the sense that they commute with the evident maps between games?

Size. 59 pages; 8 chapters in four parts (Statics, Dynamics, Structure, Assessment); four appendices (toolkit, section-level concordance with the source, computations, solutions); 28 exercises with solutions to all of them; a companion script with 42 computational checks. Every numbered statement carries a status tag: classical, literature, reformulation, observation or programme. The book reproduces none of the source's text, figures or tables; where a numerical example of the source is recomputed, this is said.

The organizing statement

Symmetric games $(S,A)$ form a category $\mathsf{Game}$ whose morphisms are payoff-preserving maps $f\colon S\to T$, $B_{f(s)f(s')}=A_{ss'}$. They are generated by inclusions of subgames, clone maps (surjections that duplicate strategies) and symmetries. The replicator flow $\varphi^A$ on the simplex $\Delta(S)$ commutes with pushforward:

$$\begin{CD} \Delta(S) @>{\varphi^A_t}>> \Delta(S) \\ @V{f_*}VV @VV{f_*}V \\ \Delta(T) @>>{\varphi^B_t}> \Delta(T) \end{CD}$$

So $\mathrm{Rep}\colon\mathsf{Game}\to\mathsf{Dyn}$ is a functor. Invariance of faces, the lumping of duplicate strategies and the action of symmetries are its three generating cases. Most of the book consists of asking the same question of each construction in the source.

Results, with their status

Stability (Ch. 2). The source's four definitions of an evolutionarily stable strategy are one quadratic form, $Q_x(y)=\pi(x\mid y)-\pi(y\mid y)$, positive near $x$ (Maynard Smith–Price; Hofbauer–Schuster–Sigmund). Observation: the pullback of an evolutionarily stable set along a clone map is an evolutionarily stable set, while an evolutionarily stable strategy is destroyed by duplicating a strategy in its support. The set-valued refinements that the source surveys (Thomas; Balkenborg–Schlag) are what the point-valued notion is forced to become.

The replicator equation (Ch. 3). Quotient of the growth equation by scaling; functoriality as above; Shahshahani gradient for symmetric payoffs with $\frac{d}{dt}(\text{mean payoff})=2\,\mathrm{Var}$; $\frac{d}{dt}D(x\,\Vert\,y(t))=-Q_x(y(t))$, so stability is the statement that a nearby population gains information about $x$; the folk theorem; elimination of dominated strategies with its exponential rate; conservation of relative entropy in zero-sum games; eigenvalues $(b-a)/6\pm i\sqrt3(a+b)/6$ for generalized Rock–Paper–Scissors; the time-average theorem.

Index, recurrence, decomposition (Ch. 4). For a regular game the indices of the Nash equilibria sum to $1$, the Euler characteristic of the simplex (Hofbauer–Sigmund; Poincaré–Hopf), so two stable conventions force an unstable third. Behind the number is the Conley index, a homotopy type invariant under continuation; behind recurrent behaviour, Conley's decomposition into chain components. Every zero-sum game splits orthogonally into rating differences and a cyclic part (Jiang–Lim–Yao–Ye; Candogan et al.); "closed implies exact" is the vanishing of $H^1$ of a simplex, and a harmonic part appears when comparisons are incomplete.

Revision protocols and open systems (Ch. 5). The replicator equation is the mean dynamic of pairwise proportional imitation (Schlag; Sandholm). Observation: every imitative protocol is natural for all maps of games, while the Brown–von Neumann–Nash and Smith dynamics are natural only for symmetries. Imitation counts individuals; a direct protocol counts names. The Hofbauer–Sandholm theorem that dominated strategies can survive (the source's Theorem 8) is the visible cost of giving up naturality in exchange for Nash stationarity. A population is an open system, payoffs in and state out; the replicator box is an integrator followed by softmax; a game is a wiring of boxes.

Finite populations (Ch. 6). Moran process, fixation formula, one-third law (Nowak et al.), mean-field limit (Kurtz). The source's table of fixation probabilities is reproduced and its limit $1-a/c=\tfrac14$ derived. The source concludes that the replicator dynamics cannot be the limit of the birth–death process; the book shows that this holds for one order of limits and fails for the other, and that the two descriptions compose: a mutant lineage escapes a boundary layer with probability $1-a/c$ and is then carried by the flow.

Local interaction (Ch. 7). Observation: imitate-the-best is natural for coverings of the interaction graph, $U_{\tilde\Gamma}\circ p^*=p^*\circ U_\Gamma$, and equivariant under symmetries of graph and game. The rotating clusters of Rock–Paper–Scissors on a ring of nine are a relative equilibrium fixed by a twisted symmetry (rotation by three composed with a cyclic relabelling); of the 27 configurations with that symmetry, nine form the wave and the other eighteen join it in one round. For the Prisoner's Dilemma on the $1,2$-circulant, a cluster of cooperators grows if $3R+S>2T+2P$, is frozen if $3R+S<2T+2P<4R$, and is invaded if $2T+2P>4R$. Isothermal theorem, amplifiers and suppressors (Lieberman–Hauert–Nowak).

Quantitative results

StatementValue
Hawk–Dove invasion form$Q=C(p-q)^2/2$
Escape from a neutrally stable state, $0.05\to\tfrac12$time $20.94$; order $1/x_0$ against $\log(1/x_0)$ under selection
Index theorem, 400 random gamesindices sum to $1$; all 590 stable strategies found have index $+1$
Dominated twin under the Smith dynamictime-averaged share $0.116$; extinction at rate $0.1$ under the replicator
Defect of naturality under cloningreplicator $4\times10^{-17}$; Smith $0.555$; Brown–von Neumann–Nash $0.380$
Source's fixation table, $N=10,\dots,10^4$$0.34263$, $0.25879$, $0.25088$, $0.25009$ reproduced (no self-interaction); limit $\tfrac14$
Mean-field fluctuations, $N=50,200,800$$0.158$, $0.084$, $0.040$: order $N^{-1/2}$
Star, well-mixed, directed line ($N=6$, $r=1.5$)$0.4432$, $0.3654$, $0.1667$
Circulant thresholds ($R=1$, $P=S=0$, $T=b$)$b=\tfrac32$ and $b=2$; source's run reproduced, ending in the cycle $1,5,3$ of defectors

What is and is not claimed

Nothing here is a new theorem of evolutionary game theory. The statements tagged observation are short, and some are likely known to people who work with these models in other forms: the pullback theorem for stable sets; the functoriality of the replicator equation stated as such; the naturality of imitative protocols and its failure for direct ones; the covering theorem and the twisted-symmetry account of the ring; the thresholds on the circulant; the reading of the two limits. The closing chapter assesses how much of this is "higher": most of it is ordinary category theory; homotopy enters through the index and the Conley index; cohomology enters in degree one; the compositional structure of open systems is two-dimensional.

Not treated: chaos in the replicator equation (cited only), Bomze's classification (cited only), the source's centipede simulation (not reproduced). One point of difference with the source is recorded: with the label-free tie rule used here, Rock–Paper–Scissors on a ring has a frozen configuration, $RPS\,RPS\,RPS$, besides cycling and uniformity; no tie rule was found in the source, and its remark may be correct for the rule it uses.

The source was read from a text extraction of its PDF: its numbered definitions and theorems were located by search and read in context, and the remaining text was read selectively, section by section. Figures were consulted only through captions and the surrounding discussion.

Files

A companion paper, Evolutionary Game Theory in Machine Learning and AI: Naturality, Regret, and Recurrence, applies the framework to learning rules, adversarial training, the evaluation of agents and populations of learners.

Status and authorship

Contribution. Exposition and reorganization of known material in categorical language, with a small number of observations and a reproducible set of computations.

Limitations. As listed above; in addition, proofs marked as sketches are sketches, the bibliography was assembled from memory except for one checked entry, and neither proofs nor bibliographic details have been independently refereed. The script's checks are evidence that examples are stated correctly and are not proofs. Corrections are welcome.

AI assistance. The text was drafted and typeset with a large language model (Anthropic's Claude) under the direction of the submitting author, who is responsible for it. Generative AI assisted with drafting, literature recall, mathematical consistency checks, computation and typesetting.

Reproducibility artifacts

evolutionary-game-theory-in-higher-categories.tex · 122,734 bytes · Read in browser · Download
Text indexing selected · View indexed source
SHA-256 fddf16cd534dab2b4271fc4220b71705904d521fa785742502b03a49004e9231

evolutionary-game-theory-in-higher-categories.pdf · 432,530 bytes · Download
SHA-256 d94c922417ffaf5f38f365bd9c17e4ab3e939fab08c5e3a277eff723d9ea7c13

Discussion

Open a careful, constructive conversation about this note.

Comments appear after editorial approval.