we-solve (beta)

Econ-73 · economics · computer science · Computational economics

Non-equilibrium learning attractors

LLM status Open unverified

Statement

Original subject group: Game theory and strategic interaction

Audit classification: Explicit published open question. The v4 section 4 Algorithmic problems item 1 explicitly asks for a local-source converse. The revised statement fixes an attraction-based interpretation, with the uniform convention supported by Biggar--Shames (2023); Lyapunov stability alone is distinguished.

Definitions and precise research target

Model and research target. Let a finite two-player game have rational payoff matrices $A,B\in\mathbb Q^{r\times s}$, mixed-strategy space $X=\Delta_r\times\Delta_s$, and replicator flow $\varphi_t$ defined by \[ \dot x_a=x_a((Ay)_a-x^TAy),\qquad \dot y_b=y_b((x^TB)_b-x^TBy). \] Here $\Delta_k$ is the probability simplex on $k$ actions. The preference graph has pure profiles as vertices and an arc for each unilateral weak payoff improvement. Let $H$ be a strongly connected component with no outgoing arcs. Its content $K_H$ consists of mixed profiles whose product distribution is supported entirely on $H$. A compact invariant set $S\subseteq X$ is attracting if some positively invariant neighborhood $U$ of $S$ satisfies $\sup_{z\in U}\operatorname{dist}(\varphi_t(z),S)\to0$; it is an attractor if it has no proper compact invariant attracting subset. Apply these definitions to $S=K_H$. This is the attraction-based interpretation of the source question; Lyapunov stability alone means that every sufficiently close initial condition stays close for all future times and does not imply attraction. A subgame $Y$ is a product of faces of the two simplexes. A local source of $H$ is a point $z\in K_H\cap Y$, with $Y\not\subseteq K_H$, that is a quasi-strict Nash equilibrium of the negated-payoff game restricted to $Y$: its best responses are exactly its supported actions. Determine whether every sink component whose content fails to be an attractor necessarily has such a local source in some subgame. Prove the implication for all finite bimatrix games in the stated domain, or supply a counterexample and identify the alternative obstruction.

Definitions and maintained assumptions (editorial unless a source definition is named). Use $\Delta_r=\{x\ge0:\sum_a x_a=1\}$. The preference graph has an arc $(a,b)\to(a',b)$ when $A_{a'b}\ge A_{ab}$, and similarly for player two; endpoints are distinct. A sink is a strongly connected component with no outgoing arc. Define $K_H=\{(x,y):\operatorname{supp}(x)\times\operatorname{supp}(y)\subseteq H\}$. The polynomial replicator field defines a global flow on compact $X$. To make the audit’s convention explicit, an attracting set is nonempty compact invariant $S$ with a relative positively invariant neighborhood $U$ and $\sup_{z\in U}\operatorname{dist}(\varphi_t(z),S)\to0$. Allow $S=U=X$, otherwise requiring a strictly larger neighborhood would contradict the source’s whole-space examples. An attractor is minimal among such attracting sets. A subgame is a product of faces. Quasi-strict Nash in the negated subgame means every supported action minimizes its original expected payoff, all supported actions tie, and each unsupported action in that subgame has strictly larger original payoff. The finite rational bimatrix restriction is a specialization of the source’s broader domain. For a sink content, attracting implies minimal attracting: an attracting subset contains a pure vertex of $H$, hence all of $H$; the chain-component structure then forces it to contain $K_H$. This uses the 2023 reference's Lemma 3.5 and Theorem 5.2, together with the fact that an attracting set that meets a chain component contains that component. Thus minimality does not introduce a distinct sink-content obstruction.

Variants and assumption changes

  1. Many players (source question, editorial formalization). For $N$ finite players replace product supports and the replicator equations accordingly. Ask whether every nonattracting sink content admits a local source in some product-face subgame.
  2. Weaker convergence (source research direction). Require only $\operatorname{dist}(\varphi_t(z),K_H)\to0$ for every interior $z$, without uniformity or attraction of boundary histories. Determine the appropriate witness criterion; this is distinct from the preceding minimal-attractor question.

References and source audit

  1. Definition2.1; §4 Algorithmic problems item1, printed p.12. Explicit attraction-converse question checked; the formal statement uses attraction rather than interpreting every occurrence of "stable" as Lyapunov stability. Added domain restrictions are editorial. The uniform-attraction convention has explicit support in the 2023 reference. https://arxiv.org/pdf/2502.07975v3
  2. Definition2.1; §4 Algorithmic problems item1, printed p.12. Explicit attraction-converse question checked; the formal statement uses attraction rather than interpreting every occurrence of "stable" as Lyapunov stability. Added domain restrictions are editorial. The uniform-attraction convention has explicit support in the 2023 reference. https://arxiv.org/pdf/2502.07975v4
  3. Biggar--Shames (2023), Definition3.2 verifies the uniform-attraction convention; Lemma3.5 and Theorem5.2 give the sink/chain structure. https://proceedings.mlr.press/v201/biggar23a/biggar23a.pdf
  1. Oliver Biggar; Christos H. Papadimitriou. 2025. Sink equilibria and the attractors of learning in games. Definition2.1; §4 Algorithmic problems, item1, printed p.12, version3 (26October2025). https://arxiv.org/pdf/2502.07975v3.
  2. Oliver Biggar; Christos H. Papadimitriou. 2026. Sink equilibria and the attractors of learning in games. §4 Algorithmic problems, item1, printed p.12, version4 (4March2026). https://arxiv.org/pdf/2502.07975v4.
  3. Oliver Biggar; Iman Shames. 2023. The Replicator Dynamic, Chain Components and the Response Graph. Definition3.2, Lemma3.5, Theorem5.2; Proceedings of Machine Learning Research201,1–22. https://proceedings.mlr.press/v201/biggar23a/biggar23a.pdf.

Common conventions: Source mathematical conventions

These conventions apply to each entry unless explicitly replaced.

Models and identification. A primitive model $m$ specifies all probability laws, feasible choices, information, equilibrium and measurement rules used by its target. The admissible class $\mathcal M$ is fixed independently of outcomes. If $P_O(m)$ is its observable probability law and $\tau(m)$ the target, its identified set at observable law $P$ is \[ \mathcal I_\tau(P)=\{\tau(m):m\in\mathcal M,\ P_O(m)=P\}. \] Point identification means that this set is a singleton. An empty set signals incompatibility of data and assumptions. Coordinate bounds need not describe the joint set. Bounds are sharp only if every retained target is attainable or arbitrarily approachable by compatible models. Finite bounds require explicit restrictions on unobserved quantities; unrestricted confounding, hidden wealth, extrapolation or arbitrary equilibrium selection can yield uninformative sets.

Causal interventions. A potential outcome $Y_i(p)$ is the outcome under a complete policy regime $p$, including timing, financing, information and market spillovers. The target population and units are fixed before treatment. With interference, use the complete assignment vector or a justified exposure mapping; individual no-interference notation alone cannot identify an economy-wide intervention. A difference of mean potential outcomes does not require the cross-world joint distribution. Conditioning on post-treatment migration, survival, enrollment or employment changes the estimand and needs separate justification.

Statistical statements. An estimator, confidence set, forecast or test specifies a sampling law, model class and loss/coverage criterion. Uniform $(1-\alpha)$-coverage means $\inf_{m\in\mathcal M}\Pr_m\{\tau(m)\in C_n\}\geq1-\alpha$, or an explicitly asymptotic version. The whole parameter space trivially has coverage, so informativeness requires a stated width, risk or power objective. A forecasting improvement is relative to a named benchmark, information set and evaluation population.

Equilibrium and optimization. An equilibrium comprises feasible plans, optimality, beliefs where needed, budget constraints and all market/resource clearing. The primitives or a stated benchmark define each correspondence $\mathcal E(p;m)$. Nonemptiness is assumed or proved, never inferred just from compact policy sets. A selected-equilibrium target uses a declared selection rule; a robust target quantifies over all permitted equilibria. Use suprema or infima unless attainment is established. An $\varepsilon$-optimal policy is feasible and within $\varepsilon$ of the finite optimum value. Report an empty feasible set separately.

Welfare and accounting. Normative utility, interpersonal normalization, social weights, numeraire, discounting and the use of public receipts are primitives. Behavioral utility need not coincide with the welfare criterion. Household welfare includes ownership income and rents once; adding profits again to compensating variation generally double-counts them. A monetary equivalent is defined by an explicit transfer and price convention, with monotonicity and range conditions. Average effects, mechanism contributions, transfers and welfare losses are different targets. Infinite sums require convergence and dynamic feasible plans require terminal or no-Ponzi conditions.

Source versions. A working-paper date, revision date and journal publication year are distinguished. A DOI for a journal version is not automatically the DOI of an earlier manuscript. Locators refer to the stated version and printed pages where specified. A metadata-only check does not verify a claim about a full-text conclusion. Every entry's source subsection narrows the provenance claim accordingly.

Statement licence: CC-BY-4.0 · Source file

Contributions 2

Econ-73-A2 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-73, version 2

Non-equilibrium learning attractors: Algebraic attraction of nonrectangular cross sinks

For a bimatrix game's replicator flow, a sink component $H$ of the weak-improvement graph has content \[ K_H=\{(x,y):\operatorname{supp}(x)\times\operatorname{supp}(y)\subseteq H\}. \] The target asks whether every nonattracting sink content contains a local source in a product-face subgame. Version 2 adds a genuinely nonrectangular cross-shaped class with a uniform algebraic attraction bound and a direct minimality proof. It also treats all $2\times2$ sink contents, including payoff ties. The rectangular exponential certificate from Version 1 is retained. Qualitative subgame and chain-component results used below are credited explicitly to the literature. No conclusion for arbitrary nonrectangular contents is claimed.

An attracting set here is nonempty, compact, invariant, and uniformly attracts a relative positively invariant neighborhood. An attractor in the catalogue is minimal among such sets. In contrast, Definition 3.2 of Biggar--Shames [chain] calls the uniformly attracting set itself an attractor; when applying their results we use that convention, not an extra minimality assertion.

Theorem (Rectangular sinks contract transversely): Let $A,B$ be payoff matrices with entries of absolute value at most $M$, and let a nonempty sink component of the weak-improvement graph be a rectangle $H=I\times J$. If this is the whole pure-profile set, then $K_H=X$. Otherwise define $\delta>0$ to be the minimum of all existing numbers \[ A_{ab}-A_{a'b}\quad(a\in I,a'\notin I,b\in J), \qquad B_{ab}-B_{ab'}\quad(a\in I,b\in J,b'\notin J). \] At least one of these lists is nonempty. Put \[ \eta=\min\left\{\frac12,\frac{\delta}{2\delta+4M}\right\}, \quad p=\sum_{a\notin I}x_a,\quad q=\sum_{b\notin J}y_b. \] The relative neighborhood $U=\{p<\eta,q<\eta\}$ is positively invariant, and every trajectory starting there satisfies \[ p(t)+q(t)\leq(p(0)+q(0))e^{-\delta t/4}. \] Consequently $K_H$ is uniformly attracting. No point of $K_H$ is a local source of $H$ in the sense of the question.
Proof: If $H$ is the whole pure-profile set, $K_H=X$ is invariant and uniformly attracts $U=X$ with distance identically zero; there is no subgame outside $K_H$ in which to test a local source. Assume henceforth that $H$ is proper. There is no outgoing weak-improvement arc from any vertex of $H$. Thus every displayed payoff difference is strictly positive. Finiteness gives $\delta>0$. For any inside row $a\in I$ and outside row $a'\notin I$, and any opponent distribution with outside mass $q$, the expected inside-minus-outside payoff difference is at least \[ \delta(1-q)-2Mq=\delta-(\delta+2M)q\geq\delta/2 \] when $q\leq\eta$. The analogous estimate for columns holds whenever $p\leq\eta$.

For $0<p<1$, summing the replicator equations over the outside rows yields \[ \dot p=p(1-p)(\overline u_{\mathrm{out}}-\overline u_{\mathrm{in}}) \leq-\frac\delta2 p(1-p)\leq-\frac\delta4p \] as long as $p,q\leq\eta$. At $p=0$ the same inequality holds by invariance of faces. The same calculation gives $\dot q\leq-\delta q/4$. At either upper boundary the corresponding derivative points inward; hence $U$ is positively invariant. Integration proves the exponential estimate.

Renormalize the inside coordinates of $x$ and $y$ to obtain a point in $\Delta_I\times\Delta_J=K_H$. Its product-space $\ell^1$ distance from $(x,y)$ is $2p+2q$. Therefore \[ \sup_{z\in U}\operatorname{dist}_1(\varphi_t(z),K_H) \leq4\eta e^{-\delta t/4}, \] which is the required uniform attraction. The face is compact and invariant.

Finally take $z\in K_H$ and a product-face subgame $Y$ containing $z$ but not contained in $K_H$. At least one player has an action available in $Y$ outside its corresponding set $I$ or $J$. Against $z$ that action has strictly lower original payoff than every supported action, by the same sink inequalities with no outside opponent mass. In the negated game it is strictly better than the supported actions. Thus $z$ is not even a Nash equilibrium of the negated subgame, and in particular is not quasi-strict there. ∎

Corollary (Minimality under the target convention): Under the hypotheses of the theorem, $K_H$ is an attractor in the minimal-attracting-set convention of the question.
Proof: We invoke three established facts in the uniformly attracting convention of [chain]: Lemma 3.5 says such a set contains a nonempty attracting set of response-graph nodes; Theorem 5.2 puts the content of a strongly connected component inside the chain component containing its nodes; and Lemma 3.9 implies that an attracting set containing a chain-recurrent point contains its entire chain component. These facts are imported, not proved by the transverse estimate.

Suppose $S\subseteq K_H$ is nonempty, compact, invariant and attracting in the ambient strategy space. The first fact supplies an attracting set of pure nodes inside $S$. All such nodes belong to $H$. Because $H$ is strongly connected, any nonempty subset of its nodes with no outgoing response paths contains every node of $H$. Thus $S$ contains $H$. The second and third facts then imply $K_H\subseteq S$. Hence $S=K_H$, establishing minimality. The same argument covers $K_H=X$ when $H$ is the entire response graph. ∎

New advance: a nonrectangular cross

Number rows and columns from $0$, and suppose each player has at least two actions. Consider the pure-profile set \[ H=(\{0\}\times\{\text{all columns}\}) \ \cup\ (\{\text{all rows}\}\times\{0\}). \tag{1} \] Assume the anchor ties and strict exterior gaps \begin{align} A_{a0}&=A_{00}\quad\text{for every row }a, &B_{0b}&=B_{00}\quad\text{for every column }b,\label{wsref-0-65-71-3a-63-72-6f-73-73-2d-74-69-65-73}\\ A_{0b}&>A_{ab},&B_{a0}&>B_{ab} \quad\text{for every }a\neq0, b\neq0.\label{wsref-0-65-71-3a-63-72-6f-73-73-2d-67-61-70-73} \end{align} Let $\delta>0$ be the minimum of the finitely many strict gaps in (eq:cross-gaps). Unlike a rectangle, (1) has content equal to a union of two faces: \[ K_H=(\{e_0\}\times\Delta_{\rm columns}) \ \cup\ (\Delta_{\rm rows}\times\{e_0\}). \tag{2} \]
Lemma (Geometry and stationarity of the cross): Under (eq:cross-ties)--(eq:cross-gaps), $H$ is a sink strongly connected component of the weak-improvement graph. Its content is (2), is connected, and consists entirely of stationary points of the replicator.
Proof: Every node of the anchor row is joined in both directions to $(0,0)$ by the column player's payoff ties. Every node of the anchor column is joined in both directions to $(0,0)$ by the row player's payoff ties. Thus $H$ is strongly connected. A unilateral move from $H$ to an exterior node $(a,b)$ with $a,b\neq0$ must leave $(0,b)$ by a row move or $(a,0)$ by a column move. Both moves strictly reduce the moving player's payoff by (eq:cross-gaps), so no outgoing weak-improvement arc exists. Consequently no larger strongly connected component contains $H$.

A product support avoids all exterior nodes if and only if it gives zero mass either to nonanchor rows or to nonanchor columns. This is (2). The two faces are connected and meet at $(e_0,e_0)$, so their union is connected. On the first face the row strategy is pure and cannot change; all column payoffs tie by (eq:cross-ties), so its replicator velocity is zero. The second face is stationary for the analogous reason. ∎

Theorem (Uniform algebraic attraction and minimality): Under the cross assumptions, let $p=1-x_0$, $q=1-y_0$, and $z=pq$. For every $\eta\in(0,1)$, the relative neighborhood $U_\eta=\{z<\eta\}$ of $K_H$ is positively invariant and satisfies \[ \sup_{(x,y)\in U_\eta} \operatorname{dist}_1(\varphi_t(x,y),K_H) \leq \frac{2\sqrt\eta} {1+\delta(1-\sqrt\eta)\sqrt\eta\,t} \quad(t\geq0). \tag{3} \] Thus $K_H$ is uniformly attracting, and is an attractor under the catalogue's minimality convention. No point of $K_H$ is a local source in a product-face subgame not contained in $K_H$.
Proof: For every nonanchor row $a$, its payoff minus the anchor row's payoff against $y$ is \[ (Ay)_a-(Ay)_0 =\sum_{b\neq0}y_b(A_{ab}-A_{0b})\leq-\delta q, \] since the column-$0$ terms cancel. Averaging over nonanchor rows and summing their replicator equations gives \[ \dot p\leq-\delta p(1-p)q. \] The same calculation for columns gives $\dot q\leq-\delta q(1-q)p$. These inequalities also hold when either outside mass is $0$ or $1$, by the replicator equation and invariance of its faces. Therefore \[ \dot z=q\dot p+p\dot q \leq-\delta z(p+q-2pq) \leq-2\delta z^{3/2}(1-\sqrt z), \tag{4} \] where the second inequality uses $p+q\geq2\sqrt{pq}$. For $z<\eta<1$ the right side is nonpositive, so $U_\eta$ is positively invariant. Put $c=\delta(1-\sqrt\eta)>0$. For $z>0$, (4) gives $(z^{-1/2})'\geq c$, whence \[ \sqrt{z(t)}\leq\frac{\sqrt{z(0)}}{1+c\sqrt{z(0)}t} \leq\frac{\sqrt\eta}{1+c\sqrt\eta t}. \] The zero case follows from invariance and satisfies the same bound. The distance in the product $\ell^1$ metric to (2) is $2\min(p,q)\leq2\sqrt{pq}$: moving to the first face costs $2p$, and to the second costs $2q$. This proves (3).

Here minimality can be proved without a chain-component theorem. Suppose $S$ is a nonempty compact invariant attracting subset of $K_H$, with relative neighborhood $W$ in the whole strategy space. Every point of $K_H$ is stationary by Lemma Lemma (Geometry and stationarity of the cross). Any point in $W\cap K_H$ therefore stays put; its distance to $S$ can tend to zero only if it belongs to the closed set $S$. Hence $W\cap K_H\subseteq S$. Since $W$ is a neighborhood of $S$, this makes $S$ relatively open in $K_H$. It is also relatively closed by compactness. Connectedness of $K_H$ forces $S=K_H$.

Finally take a point in $K_H$ and a product-face subgame $Y$ containing it but not contained in $K_H$. Such a $Y$ contains a nonanchor row and a nonanchor column. If $p=0$ and $q>0$, every available nonanchor row has strictly lower original payoff than row $0$, by the first inequality of the proof. It is strictly better in the negated game, ruling out even Nash equilibrium there. If $q=0$ and $p>0$, the column argument is identical. At $p=q=0$, the anchor ties make every such unsupported action tie with the supported action. It is then another best response in the negated subgame, which violates quasi-strictness. These cases exhaust $K_H$. ∎

Example (Algebraic behavior is present, not only a loose bound): Take the $2\times2$ game \[ A=B=\begin{pmatrix}0&0\\0&-1\end{pmatrix}. \] Its three profiles other than $(1,1)$ form the cross sink and $\delta=1$. For symmetric initial masses $p(0)=q(0)=r_0\in(0,1)$, symmetry and uniqueness of the ODE give $p(t)=q(t)=r(t)$ with $\dot r=-r^2(1-r)$. In particular \[ \frac1{r_0^{-1}+t}\leq r(t) \leq\frac1{r_0^{-1}+(1-r_0)t}. \] Indeed, $r$ decreases and $(1/r)'=1-r\in[1-r_0,1]$; integration gives the bounds. The distance to $K_H$ is $2r(t)$, so this trajectory has algebraic rather than exponential decay. The explicit rate in (3) addresses a phenomenon absent from the rectangular estimate.
Corollary (Every sink content in a $2\times2$ game is an attractor): For every bimatrix game with two actions per player, allowing ties, the content of every sink strongly connected component is an attractor under the catalogue convention. Consequently the local-source implication holds throughout this subclass: its nonattracting-sink premise never occurs.
Proof: The underlying pure-profile adjacency graph is a square. A one-vertex sink is rectangular. A two-vertex strongly connected component must be an edge, since opposite vertices are not adjacent and have no path within that pair; it too is rectangular. The retained rectangular theorem and its minimality corollary apply to both cases.

Three vertices form a path with a central corner and two endpoints. A strongly connected directed graph on a path must contain each of its edges in both directions. For the weak-improvement graph this means payoff ties for the moving player on both edges. Relabel the central corner as $(0,0)$. The no-outgoing-arc condition from the endpoints to the fourth vertex gives exactly the two strict exterior gaps in (eq:cross-gaps), and the bidirectional edges give (eq:cross-ties). Theorem Theorem (Uniform algebraic attraction and minimality) applies.

If the component has four vertices, it is the whole pure-profile set and its content is the whole strategy space. It uniformly attracts that space with zero distance. Its minimality follows from the retained corollary, whose imported chain-component facts expressly cover this case. This exhausts all possible component sizes. ∎

What the certificate excludes.
The rectangular theorem uses uniform transverse payoff gaps. The cross extension replaces a linear distance estimate by the product $pq$, relying on exact anchor ties and strict losses on the missing rectangle. General nonrectangular contents may have neither geometry. Their overlapping faces need not consist of stationary points and need not admit this product contraction. The full local-source converse in larger games remains unresolved.
Source relationship.
Biggar and Papadimitriou [source], Definition 2.1 and Section 4, pose the local-source witness question. The qualitative attracting-subgame case is already covered by Biggar and Shames [chain], Lemma 5.4. The retained rectangular result makes the strict payoff margin, neighborhood and exponential estimate explicit. The cross calculation and its algebraic rate are proved directly here, as is cross minimality. The $2\times2$ corollary combines these arguments with the cited established chain-component facts; no literature-wide novelty claim is made.

Completion Estimate

30%

This is a heuristic estimate for the full catalogue target. The cross theorem handles a nonrectangular union of stationary faces with a proved algebraic rate, and the case analysis covers every two-action bimatrix game, including ties. The source question concerns all finite bimatrix games, and general nonrectangular contents remain outside these arguments.

References

Go to contribution →

Econ-73-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-73

Non-equilibrium learning attractors: An explicit transverse contraction certificate for rectangular sinks

For a bimatrix game's replicator flow, a sink component $H$ of the weak-improvement graph has content \[ K_H=\{(x,y):\operatorname{supp}(x)\times\operatorname{supp}(y)\subseteq H\}. \] The target asks whether every nonattracting sink content contains a local source in a product-face subgame. This note isolates a class in which nonattraction is impossible and gives an explicit attraction rate. The qualitative attracting-subgame principle is known; the purpose here is a quantitative certificate and an exact account of its scope.

Theorem (Rectangular sinks contract transversely): Let $A,B$ be payoff matrices with entries of absolute value at most $M$, and let a nonempty sink component of the weak-improvement graph be a rectangle $H=I\times J$. If this is the whole pure-profile set, then $K_H=X$. Otherwise define $\delta>0$ to be the minimum of all existing numbers \[ A_{ab}-A_{a'b}\quad(a\in I,a'\notin I,b\in J), \qquad B_{ab}-B_{ab'}\quad(a\in I,b\in J,b'\notin J). \] At least one of these lists is nonempty. Put \[ \eta=\min\left\{\frac12,\frac{\delta}{2\delta+4M}\right\}, \quad p=\sum_{a\notin I}x_a,\quad q=\sum_{b\notin J}y_b. \] The relative neighborhood $U=\{p<\eta,q<\eta\}$ is positively invariant, and every trajectory starting there satisfies \[ p(t)+q(t)\leq(p(0)+q(0))e^{-\delta t/4}. \] Consequently $K_H$ is uniformly attracting. No point of $K_H$ is a local source of $H$ in the sense of the question.
Proof: There is no outgoing weak-improvement arc from any vertex of $H$. Thus every displayed payoff difference is strictly positive. Finiteness gives $\delta>0$. For any inside row $a\in I$ and outside row $a'\notin I$, and any opponent distribution with outside mass $q$, the expected inside-minus-outside payoff difference is at least \[ \delta(1-q)-2Mq=\delta-(\delta+2M)q\geq\delta/2 \] when $q\leq\eta$. The analogous estimate for columns holds whenever $p\leq\eta$.

For $0<p<1$, summing the replicator equations over the outside rows yields \[ \dot p=p(1-p)(\overline u_{\mathrm{out}}-\overline u_{\mathrm{in}}) \leq-\frac\delta2 p(1-p)\leq-\frac\delta4p \] as long as $p,q\leq\eta$. At $p=0$ the same inequality holds by invariance of faces. The same calculation gives $\dot q\leq-\delta q/4$. At either upper boundary the corresponding derivative points inward; hence $U$ is positively invariant. Integration proves the exponential estimate.

Renormalize the inside coordinates of $x$ and $y$ to obtain a point in $\Delta_I\times\Delta_J=K_H$. Its product-space $\ell^1$ distance from $(x,y)$ is $2p+2q$. Therefore \[ \sup_{z\in U}\operatorname{dist}_1(\varphi_t(z),K_H) \leq4\eta e^{-\delta t/4}, \] which is the required uniform attraction. The face is compact and invariant.

Finally take $z\in K_H$ and a product-face subgame $Y$ containing $z$ but not contained in $K_H$. At least one player has an action available in $Y$ outside its corresponding set $I$ or $J$. Against $z$ that action has strictly lower original payoff than every supported action, by the same sink inequalities with no outside opponent mass. In the negated game it is strictly better than the supported actions. Thus $z$ is not even a Nash equilibrium of the negated subgame, and in particular is not quasi-strict there. ∎

Corollary (Minimality under the target convention): Under the hypotheses of the theorem, $K_H$ is an attractor in the minimal-attracting-set convention of the question.
Proof: Use the established replicator facts of Biggar and Shames [chain]: an attracting set contains an attracting set of pure profiles (Lemma 3.5), and the content of a strongly connected response component lies in its chain component (Theorem 5.2). An attracting set meeting a chain component contains that component. If a nonempty compact invariant attracting subset of $K_H$ existed, its pure profiles would include the whole sink component $H$, since $H$ is strongly connected and contains all pure profiles of $K_H$. It would then contain $K_H$ by the chain-component fact. Thus no proper such subset exists. These are cited structural results; the transverse rate above was proved directly. ∎
What the certificate excludes.
Rectangularity matters: an outside row is uniformly inferior against every inside column. For a nonrectangular $H$, a row can be inside over some columns and outside over others; one cannot define fixed masses $p,q$ with these uniform payoff gaps. The proof supplies no contraction function across the union of faces making up a nonrectangular content. It therefore does not establish the local-source converse.
Source relationship.
Biggar and Papadimitriou [source], Definition 2.1 and Section 4, pose the local-source witness question. The qualitative attracting-subgame case is already covered by Biggar and Shames [chain], Lemma 5.4. The present result makes the strict payoff margin, neighborhood, and decay estimate explicit; it is not presented as a new qualitative resolution of that settled case.

References

Completion Estimate

15%

Go to contribution →