{"id":"60fa92f6-af64-46ac-9994-7ab9b79d347a","arxiv_id":"2412.19216","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Fictitious play is claimed to converge in every 3x3 game without an internal indifferent point, via a new projection to a planar dynamical system.","lead":"This paper claims fictitious play, a classic learning rule in game theory, converges to a Nash equilibrium in every 3x3 game without an internal indifferent point, using a new projection that collapses the dynamics to a plane. The theorem as stated is contradicted by the paper's own degenerate example, and the proof leaves key cases and a reduction step unproven.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1 is contradicted by the paper's own Example 4.2: a 3x3 game without IIP whose CFP is said to approach the NE continuum but never any equilibrium point.","rationale":"The reader's overall rationale already mentions Example 4.2 as a contradiction, and I agree that this is the decisive problem: the paper's own example refutes the central theorem if 'converges' is used in the paper's defined sense. However, the reader's labelled weakest_assumption is Proposition 3.3, the projection reduction, which is a proof-gap concern about nondegenerate games; my identified load-bearing concern is instead the internal counterexample that kills the theorem as stated. The projection issue is secondary: even if the reduction were repaired, Theorem 3.1 would still be false for the degenerate game in Example 4.2 unless the statement is amended. Thus the correct verdict remains rejection, but for a more fundamental reason than the Proposition 3.3 gap.","tokens_in":22956,"tokens_out":3965,"duration_ms":37364,"concrete_test":"Independently simulate CFP for Example 4.2 at k = -2/3 from a generic interior initial condition, recording x(t), y(t) and the distance to the equilibrium continuum. If the trajectory does not converge pointwise to some (x*, y*) — for instance, if a coordinate has positive oscillation or the distance to every candidate NE fails to tend to 0 — then Theorem 3.1 is false for that game and must be restricted or reformulated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 3.1: 'Every CFP approaches equilibrium in every 3x3 game without IIP.' Under the paper's own Section 2.2, convergence means pointwise convergence to a strategy profile (x*, y*). Example 4.2 (Section 4.2, k = -2/3) is a 3x3 game explicitly classified as without IIP, and the authors state that degeneracy does not affect the claimed convergence and that the same projection method applies. Yet they describe its CFP dynamics as one that 'surrounds and approaches the set of NEs, but will never converge to any point in the set' (Figure 9). That is exactly failure of pointwise convergence. Therefore Theorem 3.1 as stated is internally inconsistent with the paper's own example. The scope is also ambiguous: Section 3's proof, e.g. Proposition 3.2, assumes nondegenerate games, while Theorem 3.1 and Section 4.2 claim to cover degenerate games. If Theorem 3.1 is intended only for nondegenerate games, the theorem, the abstract, and the Section 4.2 claim that degeneracy does not affect convergence must be revised. If it is intended as stated, then the example must actually converge pointwise, which the text denies.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies continuous-time fictitious play (CFP) in two-player 3x3 games without an internal indifferent point (IIP). It introduces a projection from the four-dimensional strategy space to a planar system, defines saddle and sink Nash equilibria for the resulting non-smooth best-response dynamics, and claims Theorem 3.1: every CFP approaches equilibrium in every 3x3 game without IIP. The paper also claims that the projection method extends to degenerate games and to some higher-dimensional games, and it compares the proposed class with quasi-supermodular games.","tokens_in":23165,"tokens_out":4726,"duration_ms":45083,"significance":"If the main theorem were correct, the paper would identify a substantial new class of games with the fictitious play property and would offer a geometric reduction that could be useful for other game classes. The paper contains concrete examples and simulations, and the geometric projection idea is genuinely interesting. However, the central theorem is contradicted by the paper's own degenerate Example 4.2, and key proof steps are only sketched for a 'typical case' or handled by assertion of similarity. The contribution is therefore not established as it stands.","major_comments":[{"comment":"Section 2.2 defines convergence of CFP as pointwise convergence to a strategy profile (x*, y*). Theorem 3.1 asserts convergence in every 3x3 game without IIP. Example 4.2 with k = -2/3 is presented as a 3x3 game without IIP (payoff matrices (13)), and Section 4.2 states that 'for 3 x 3 games without IIP, degeneracy does not affect the convergency of FP.' Yet the same section and Figure 9 state that the CFP trajectory 'surrounds and approaches the set of NEs, but will never converge to any point in the set.' Approaching a continuum is not pointwise convergence under the paper's own definition. Thus Theorem 3.1 as stated is internally inconsistent with the example. If the theorem is meant only for nondegenerate games, the abstract, Theorem 3.1, and Section 4.2 must be revised; if it is meant as stated, the example must converge pointwise, which the text denies.","section":"Section 2.2 and Section 4.2, Example 4.2"},{"comment":"The proof of Theorem 3.2 is given only for 'the following typical case illustrated by Figure 4', with the statement that 'all the other cases are just similar.' Theorem 3.1 covers all 3x3 games without IIP, and Appendix D only lists possible cases without proving convergence for each one. Since the vector field and the Poincare map depend on the arrangement of best-response regions and indifferent lines, convergence in one configuration does not imply convergence in the others. The same issue appears in Theorem 3.3 ('the proof for case (2) is similar') and in Proposition 3.4, where cases (2) and (3) are deferred to Appendix F. This leaves the main theorem unproven even setting aside the degenerate counterexample.","section":"Section 3.3, proof of Theorem 3.2"},{"comment":"The reduction from PBRD to BRD is not justified. The projection mapping phi is non-injective: each point of the projected edge corresponds to a line segment in the original strategy simplex. The two-sentence proof says only that after PBRD converges, the BRD trajectory 'eventually belongs to the 4 cells around NE' and then invokes Lemma 2.2. It does not rule out persistent non-convergent motion along the fibers of phi within those cells. Because all subsequent contraction arguments are carried out in the projected plane, this step is load-bearing: if the reduction fails, the planar convergence results do not transfer to the original CFP dynamics.","section":"Section 3.2, Proposition 3.3(2)"},{"comment":"Proposition 3.4 asserts the existence of a saddle NE and a sink NE for games with multiple NEs. The proof of the sink part states: 'The existence of sink NE is a straightforward corollary of our main results Theorem 3.1, whose proof only relies on the existence of saddle NE.' But Theorem 3.1 is the statement being proved, and its proof uses Proposition 3.4. Even if the convergence argument uses only the saddle part, the proposition as stated invokes the main theorem in its own proof. The sink existence must be proved independently or removed from the proposition.","section":"Section 3.4, Proposition 3.4"}],"minor_comments":[{"comment":"The manuscript uses 'IP' and 'IIP' inconsistently, and contains the typo 'Internal Indiffernt Point'; these should be unified to 'IIP' and 'Indifferent Point'.","section":"Abstract and Section 1"},{"comment":"The definition of the edge of Delta_A writes 'x1 <= 0, x2 <= 0'; the inequalities should be 'xi >= 0' for a probability simplex edge.","section":"Section 2.1"},{"comment":"Proposition 3.3(1) states that phi(x,y) belongs to Z^B_j x Z^B_k if and only if (x,y) belongs to Z^B_j x Z^B_k; the second factor should be Z^A_k, not Z^B_k.","section":"Section 3.2, Proposition 3.3(1)"},{"comment":"There are numerous grammatical and typographical errors (e.g., 'convergency', 'have being receiving', 'the trajectory approach the continuum') that should be corrected before publication.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The geometric projection idea and the explicit examples are interesting, and a restricted theorem for nondegenerate games might be salvageable if the authors complete the case analysis, prove Proposition 3.3(2), and remove the circular use of Theorem 3.1 in Proposition 3.4. However, the main theorem as stated is contradicted by the paper's own Example 4.2, and the proof is not complete for all claimed cases. This is a load-bearing flaw rather than a presentation issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's the read. The headline: the main convergence theorem as stated is contradicted by the paper's own Example 4.2, and the proof has load-bearing gaps. Still, this is not a throwaway. The paper introduces a genuinely new geometric reduction and gives a convincing comparison with the quasi-supermodular class, so it is worth a referee's time—but not in its current form.\n\nWhat is actually new: the class of 3x3 games without an internal indifferent point (IIP), and the projection mapping from the four-dimensional product simplex to the unit square. That mapping is a real idea, and the authors use it to turn best-response dynamics into a planar piecewise-linear system. The Poincaré map contraction for the typical cell path is explicit and checkable. The example showing the class is not contained in the quasi-supermodular games of Berger (2007) is solid. The higher-dimensional extension is speculative but interesting.\n\nNow the soft spots, in order of severity. First, Theorem 3.1 says every CFP approaches equilibrium in every 3x3 game without IIP. The paper's own Section 4.2, Example 4.2 at k = -2/3, is a 3x3 game without IIP, and the text says its trajectory 'surrounds and approaches the set of NEs, but will never converge to any point in the set.' Under the paper's own definition in Section 2.2, convergence means pointwise convergence to a strategy profile. So the theorem is internally inconsistent with the example. The authors hand-wave that degeneracy doesn't affect convergence, but that is exactly the case that breaks it. If the theorem is meant only for nondegenerate games, then the abstract, the theorem statement, and the degenerate section all need revision.\n\nSecond, the reduction step, Proposition 3.3, is not proven. The projection is non-injective: a point on the projected edge corresponds to a whole segment in the original simplex. The proof of part (2) is two sentences and invokes Lemma 2.2 after saying the trajectory 'eventually belongs to the 4 cells around NE.' It never rules out motion along the fibers of the projection. That is the bridge from the planar dynamics back to the original dynamics, and it is load-bearing.\n\nThird, the proof of Theorem 3.2 covers one typical case and says 'all the other cases are just similar.' That is not a proof for a theorem stated over all games without IIP. Appendix D enumerates cases but does not carry out the analysis for each.\n\nFourth, Proposition 3.4 contains an explicitly circular remark: the existence of a sink NE is declared 'a straightforward corollary of our main results Theorem 3.1,' and then Theorem 3.3 uses Proposition 3.4 to prove convergence. The authors claim it is not circular, but it is.\n\nThere is also a scope mismatch: the abstract advertises the fictitious play property (discrete-time), but the theorems concern CFP. The discrete-to-continuous gap is left open.\n\nWho this is for: researchers working on fictitious play and game dynamics. The projection idea and the non-overlap example are useful; the convergence claim is not supported as written. I would not cite the main theorem, but I would bring the paper to a reading group to discuss the technique and the failure mode.\n\nRecommendation: send to peer review, with the expectation of major revision or rejection. The authors need to fix the degenerate case, give a real proof of the reduction, and cover all claimed cases. As it stands, the central result is contradicted by their own example.","headline":"The paper's own Example 4.2 contradicts Theorem 3.1, but the projection method and the class comparison are genuinely useful and deserve referee attention.","tokens_in":23757,"tokens_out":3625,"would_cite":false,"duration_ms":32842,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A26","91A05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that continuous-time fictitious play converges to a Nash equilibrium from every starting point in every $3\\times3$ game whose indifference point lies outside both players' simplices, using a projection that reduces the…","keywords":["fictitious play","continuous-time fictitious play","Nash equilibrium","internal indifferent point","projection mapping","best-response dynamics","convergence","geometrical approach"],"falsifier":"Numerically integrate the continuous-time fictitious play equations (2) for any $3\\times3$ game whose indifferent point lies outside both simplices, starting from a point whose projection lands on the eight-cell loop; if the projection's return distance shrinks to zero but the original four-dimensional trajectory's distance to the unique Nash equilibrium does not (for instance, the trajectory keeps switching among all three actions), then Proposition 3.3(2) and Theorem 3.1 are false.","tokens_in":22647,"feed_emoji":"📐","tokens_out":12278,"duration_ms":106857,"temperature":0.7,"pith_summary":"The paper's central claim is that continuous-time fictitious play — the learning rule in which each player best responds to the opponent's empirical play — converges to a Nash equilibrium in every $3\\times3$ game whose indifferent point lies outside both players' strategy simplices. This defines a new class of games with the fictitious play property, distinct from the previously known quasi-supermodular games. If the claim is true, it answers a long-standing question about which $3\\times3$ games are safe for fictitious play: non-convergence, as in the classic three-by-three counterexample, requires a fully mixed equilibrium in a specific sense. The proof's engine is a projection that collapses each player's strategy triangle to an edge, turning the four-dimensional best-response dynamics into a planar system, plus a redefinition of saddle and sink equilibria for the nonsmooth dynamics. The paper also shows the projection works for degenerate games and some $4\\times4$ games, where it can certify non-convergence.","feed_headline":"Convergence proven for 3x3 games without an internal indifferent point","feed_subtitle":"A new projection shows fictitious play always reaches a Nash equilibrium in this broad new game class.","key_machinery":"The central object is the projection mapping $\\varphi = \\varphi_A \\times \\varphi_B$, which sends each player's strategy simplex to an edge by drawing the line through a point and the game's indifferent point and recording its intersection with the chosen edge; when the indifferent point is absent, the projection runs parallel to the now-parallel indifference lines. On the resulting unit square, the image of the best-response dynamics is a planar piecewise-linear system (PBRD) with nine labeled cells, and the key identity is the return map $f$ on the entry edge, composed of eight fractional-linear maps that can be written down explicitly from similar triangles. The derivative of $f$ factors into ratios of cell widths and heights and is strictly less than 1, which is the contraction that rules out the endless eight-cell loop that would otherwise allow non-convergence. The paper also introduces geometric redefinitions of saddle and sink Nash equilibria tailored to the nonsmooth vector field, and proves these always exist when equilibria are multiple.","core_discovery":"On the paper's own terms, the discovery is Theorem 3.1: every continuous-time fictitious play process approaches a Nash equilibrium in every $3\\times3$ game without an internal indifferent point (IIP), the unique (possibly negative-probability) strategy profile at which all of a player's pure actions give equal payoff. Since that point lies outside both players' simplices, each player's strategy triangle contains at most two indifference lines, making the new projection mapping well-defined. The projected best-response dynamics live on a unit square divided into nine cells; the paper proves that any trajectory either enters a four- or six-cell path where a known $2\\times n$ convergence result forces convergence, or it follows an eight-cell loop controlled by a return map. The map's derivative factors as a product of ratios of cell widths and is strictly less than 1, so the return map is contractive and eventually pushes the trajectory into the convergent regime. For games with multiple equilibria, the paper defines saddle and sink equilibria geometrically, proves a saddle and a sink always coexist, and uses the saddle's separating lines to partition the square so that no trajectory can visit all three actions of both players forever.","pith_inferences":["Editorial inference: the derivative bound on the return map is quantitative, so the same proof should yield a rate at which trajectories on the eight-cell path approach the contractive regime; a natural extension is an explicit bound on the number of returns needed for $\\epsilon$-convergence in games without IIP.","Editorial inference: Proposition 3.3(2) is the step most worth probing, because $\\varphi$ identifies whole line segments with single points, so convergence on the projected square does not formally rule out motion along those fibers; simulating the original CFP for several initial conditions with the same projection would test whether the reduction hides any non-convergent behavior.","Editorial inference: the regime-shift examples suggest that the connectivity structure of the Nash equilibrium set can change abruptly along a payoff path while the dynamics remain tame; if pursued further, this connects to the difficulty of homotopy methods for equilibrium computation rather than to fictitious play itself."],"forward_implications":["The paper's Theorem 3.1 adds the entire class of $3\\times3$ games without an internal indifferent point to the known class of games with the fictitious play property, and Example 4.1 shows this class is not contained in the quasi-supermodular games previously identified.","Corollary 3.1 turns the long-run pattern of action profiles into a classification of the limit equilibrium: one action profile repeated means a pure Nash equilibrium, two switching profiles mean a saddle, and four switching profiles mean a sink.","Because CFP and BRD follow the same trajectory shapes and differ only in time scaling, the contraction of the return map on the projected square transfers to the original dynamics, giving convergence from every initial condition in the no-IIP class.","The projection method also applies outside the convergent setting: it reproduces a known $4\\times4$ non-convergent example and reveals regime shifts in one-parameter families of games, where a continuum of Nash equilibria can emerge and the dynamics may approach the continuum without converging to any point of it."],"supporting_citations":[{"why":"Supplies Lemma 2.2, the convergence result for games where one player has two pure strategies, which the reduction uses once the trajectory is confined to the cells around an equilibrium.","marker":"[12]"},{"why":"Defines the quasi-supermodular game classes with the fictitious play property that the paper compares against and proves are distinct from the no-IIP class.","marker":"[14]"},{"why":"Supplies the definition of best-response regions and the linear equivalence relation used in the comparison section.","marker":"[33]"},{"why":"Provides the classic $3\\times3$ example where fictitious play does not converge, the baseline the new convergent class must be contrasted with.","marker":"[15]"},{"why":"Supplies a known $4\\times4$ example whose non-convergence the paper re-derives with its projection method in the higher-dimensional extension.","marker":"[16]"}],"fun_headline_variants":["Fictitious play convergence proven for 3x3 games without IIP","New proof: FP converges in 3x3 games without internal indifferent point","Geometric proof shows FP reaches Nash in broad 3x3 game class","3x3 games without IIP: FP now guaranteed to converge","Projection mapping settles convergence for a new class of 3x3 games"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that convergence of the projected planar dynamics forces convergence of the original four-dimensional dynamics, even though the projection identifies many different original states with a single point on the plane.","fun_headline_variants_meta":{"raw":{"variants":["Fictitious play convergence proven for 3x3 games without IIP","New proof: FP converges in 3x3 games without internal indifferent point","Geometric proof shows FP reaches Nash in broad 3x3 game class","3x3 games without IIP: FP now guaranteed to converge","Projection mapping settles convergence for a new class of 3x3 games"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000423,"raw_usage":{"total_tokens":2165,"prompt_tokens":934,"completion_tokens":1231,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":1132}},"tokens_in":550,"tokens_out":1231,"duration_ms":9134,"temperature":1.0,"reasoning_tokens":1132,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:50:37.983484+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Numerically integrate the continuous-time fictitious play equations (2) for any $3\\times3$ game whose indifferent point lies outside both simplices, starting from a point whose projection lands on the eight-cell loop; if the projection's return distance shrinks to zero but the original four-dimensional trajectory's distance to the unique Nash equilibrium does not (for instance, the trajectory keeps switching among all three actions), then Proposition 3.3(2) and Theorem 3.1 are false.","supporting_citations":[{"cited_title":"Fictitious play in 2×n games","cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 2.2, the convergence result for games where one player has two pure strategies, which the reduction uses once the trajectory is confined to the cells around an equilibrium."},{"cited_title":"Two more classes of games with the continuous-time fictitious play property.Games and Economic Behavior, 60(2):247–261, 2007","cited_arxiv_id":null,"evidence_quote":"Defines the quasi-supermodular game classes with the fictitious play property that the paper compares against and proves are distinct from the no-IIP class."},{"cited_title":"Payoff performance of fictitious play","cited_arxiv_id":null,"evidence_quote":"Supplies the definition of best-response regions and the linear equivalence relation used in the comparison section."},{"cited_title":"Some topics in two-person games","cited_arxiv_id":null,"evidence_quote":"Provides the classic $3\\times3$ example where fictitious play does not converge, the baseline the new convergent class must be contrasted with."},{"cited_title":"Shapley Polygons in 4 x 4 Games","cited_arxiv_id":null,"evidence_quote":"Supplies a known $4\\times4$ example whose non-convergence the paper re-derives with its projection method in the higher-dimensional extension."}],"review_version":1}