Source
Omidshafiei, Papadimitriou, Piliouras, Tuyls, Rowland, Lespiau, Czarnecki, Lanctot, Pérolat & Munos, LIT-1516, read in NOTE-1156 from arXiv:1903.01373v4: eqs. 2–10 (pp. 6–7), Theorem 2.1.2 (p. 8, proof pp. 42–43), Definition 2.4.23 and Theorem 2.4.24 (p. 14, proof p. 44), Theorem 2.5.1 (p. 15, proof p. 45).
What was actually shown
- Uniqueness. Every fixation probability under Fermi imitation with finite payoffs is positive. Any profile reaches any other by changing one population at a time, so the chain is irreducible, with a unique stationary distribution.
- The limit of the transitions. For payoff advantage Δ > 0, the fixation probability (1 − e^{−αΔ})/(1 − e^{−mαΔ}) tends to 1. For Δ < 0 it tends to 0, and for Δ = 0 it is 1/m. So at α = ∞ the chain moves only along weakly better deviations, and its closed classes are the sink strongly connected components of the response graph, the MCCs. The reader re-derived the limits.
- The link to dynamics. On each edge of the profile polytope, where one population mixes two strategies and the rest are pure, the replicator flow stays on the edge and runs to the better end, or rests when the payoffs tie. So every weakly better reply is reachable by an ε-chain: an ε-push onto the edge, or small hops along a line of rest points. Within a sink strongly connected component every profile reaches every other, so an MCC is chain transitive and lies in one chain component. The paper’s converse, that every asymptotically stable sink chain component contains an MCC, is argued through a volume-preserving conjugation of the interior flow. That argument is asserted, not carried out, and is THEORY-1233.
What could have come out otherwise: a sink of the better-response graph could have straddled two chain components, or an attracting component could have avoided the pure profiles altogether.
What this does not say
- That α-Rank’s ranking has a limit, or which. At α = ∞ the chain is reducible whenever there are several MCCs, as the authors say. At large finite α the split among sinks depends on payoff differences and m, and no result characterizes it.
- That every MCC is in an asymptotically stable component, or the converse. Neither is proved here. The paper claims the converse, that every asymptotically stable sink chain component contains an MCC, and that is THEORY-1233.
- That the ranking resists padding. In the symmetric single-population model, duplicating scissors in rock-paper-scissors moves rock to first place at every α tried (NOTE-1156).
- That its masses are evolutionary strength in a population-dependent sense. The single-population fitness is local: payoff against the one opponent met, regardless of the population’s composition.
Connections
The discrete form of its Conley reading is the condensation of a win digraph, acyclic and so always carrying an ordinal potential, with MCCs as its sinks. That bears on how TERM-077 and CLAIM-451 treat cyclic records. In a game with THEORY-896’s finite improvement property and no payoff ties, every MCC is a single profile, a pure equilibrium (NOTE-1156).
Note of 2026-10-30: the re-derivation, written out
The second check at the status review (ADR-065 §5) followed each step.
-
Uniqueness. With finite payoffs, the fixation probability (1 − e^(−αΔ))/(1 − e^(−mαΔ)), or 1/m at Δ = 0, is positive. So every move that changes one population’s strategy has positive probability η·ρ. Any profile reaches any other by changing populations one at a time, so the finite chain is irreducible and its stationary distribution is unique and positive.
-
The limit. As α → ∞ the fixation probability tends to 1 for Δ > 0. For Δ < 0 it behaves as e^((m−1)αΔ) and tends to 0. At Δ = 0 it is 1/m, which is also the formula’s limit as Δ → 0. The transition matrix converges entrywise to the matrix with η on weakly better strict moves, η/m on neutral moves and 0 otherwise. A closed class of a finite chain is a sink strongly connected component of its transition graph, and here that graph is the weakly-better-response graph. On a sink the limit chain is the paper’s MCC.
-
Containment in one chain component. Fix an edge of the polytope: population p mixes strategies s and s′, and the others are pure. The payoff difference Δ between s′ and s does not depend on population p’s own mix, so the flow on the edge is ẋ = x(1 − x)Δ. Faces are invariant. For Δ > 0 the orbit runs from s to s′, so an ε-chain goes from s to s′ by a push of size ε onto the edge. For Δ = 0 the edge is rest points, so ⌈1/ε⌉ hops of size ε join s and s′. Pure profiles are rest points and so chain recurrent. In a sink strongly connected component every profile reaches every other by weakly better moves, so they are pairwise chain equivalent, and they lie in one chain component. This step does not use the conjugation.
-
What the converse would need (THEORY-1233). In log-ratio coordinates u = log(x_{i+1}/x_1) for each population, population p’s payoff differences do not depend on its own coordinates, so the interior vector field has divergence 0. That is the conjugation the paper mentions, and it is the easy part. What the paper does not carry out is the rest. A volume-preserving interior flow has no interior attractor. An attractor’s trace on a face is again an attractor there, so the induction reaches a vertex. And the vertex’s weakly-better closure lies in the attractor.
A script (numpy, in the session’s scratchpad, s942_alpharank.py) checked steps 1 to 4 on 400 random games with 2 or 3 players, 2 or 3 strategies each, and integer payoffs from 0 to 2, so that ties occur. At α = 0.1 and m = 20, each chain had every one-population move positive, a transition matrix of rank n − 1 (so a unique stationary distribution), and a positive stationary vector. At α = 200 the transition matrix was within 10⁻¹² of the limit matrix. The limit matrix’s closed classes equalled the sink SCCs of the weakly-better-response graph, computed separately from the payoffs, in all 400 games. Integrating the full replicator flow from the midpoints of 932 polytope edges, the flow stayed on the edge to 10⁻¹² and ended at the better end, or did not move when payoffs tied. The interior log-ratio divergence of two-population flows was 0 at 100 random points. It had two controls, each of which had to find a difference. Dropping neutral moves from the graph changed the sink structure in 359 of 400 games, which shows the ties matter. And a one-population replicator, whose payoffs depend on its own mix, had divergence up to 1.23, which shows the divergence check can detect a non-zero value.
Note of 2026-10-30: the clause the narrowing drops
The old title said that, as the intensity grows, the chain’s transitions “become those of Markov chains confined to the sink strongly connected components”. The limit chain also has transitions out of profiles that are not in any sink. It moves along weakly better deviations everywhere. It is the MCC only on the sinks, which are its closed classes. The old wording was imprecise rather than false in substance, so it is recorded here only (ADR-047’s kind 3). The paper’s Theorem 2.5.1 says “the macro-model’s transition probabilities become the MCC transitions” (NOTE-1156), with the same ellipsis.