{"id":"89b0abb6-eb75-4703-a38e-93fee7c9a651","arxiv_id":"2608.04386","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"For stochastic linear-quadratic differential games, the authors construct an alpha-potential function, bound the approximation parameter by model coefficients and control radius, and reduce minimization to a finite-dimensional LQ control problem.","lead":"This paper shows that a class of N-player stochastic linear-quadratic differential games can be approached through alpha-potential functions, which give an explicit bound on how far a minimizer is from a Nash equilibrium and reduce the search to a standard finite-dimensional control problem. A network example from prior work is re-derived more directly, matching the earlier feedback control.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quadratic representation of the α-potential in §4.2 has an incorrect S_t matrix: the off-diagonal block (y_l, u_j) is written as S_j^l − S_j^j instead of S_l^j − S_j^j, so the derived Riccati feedback does not minimize the stated Φ for general cross terms.","rationale":"The reader's weakest assumption concerned Riccati solvability and the dependence of α on the unspecified radius L; those are genuine limitations but they are conditional assumptions explicitly stated in the paper. My stress-test found a more fundamental algebraic issue: the S_t matrix in §4.2 does not match the potential function Φ defined in (4.3). If this is correct, the reduction to a finite-dimensional LQ problem is wrong for general S_i, and the main theorem's conclusion does not follow from the stated Riccati equations. The network application sets S_i = 0, so it is not affected, which explains why the inconsistency was not noticed in the example. The concern is concrete and testable: a two-player scalar check immediately distinguishes the paper's S_t from the S_t required by (4.3)–(4.4). I recommend CONDITIONAL: the framework may be salvageable by correcting S_t and re-deriving the Riccati system, but as written the general claim is not supported.","tokens_in":29476,"tokens_out":23593,"duration_ms":200221,"concrete_test":"Take N=2 with scalar states and controls, no drift or diffusion, R_i = 1, Q_i = 0, and set S_1^1 = S_2^2 = 1, S_1^2 = 0.5, S_2^1 = 0.3. Compute the quadratic coefficient of Y_1 u_2 and Y_2 u_1 directly from (4.3)–(4.4): the correct coefficients are −0.25 and −0.35, respectively. The paper's displayed S_t produces −0.35 and −0.25, i.e., the two are swapped. If this mismatch reproduces, the LQ representation in §4.2 is incorrect for general cross terms, and the Riccati-based minimizer must be re-derived with the corrected S_t.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central reduction in Section 4.2 claims that the potential function Φ in (4.3) is represented by the quadratic form x^T Q_t x + 2x^T S_t u + u^T R_t u with the displayed matrices. I checked the S-related terms in (4.3)–(4.4). For each player i, the S-contribution to Φ is (1/2) y_i^T S_i u from the first term and u_i^T (S_i^i)^T (x − (1/2)ȳ) from the second. Expanding and collecting the coefficient of y_l^T u_j gives (1/2) S_l^j − (1/2) S_j^j, so the matrix S_t must have block (l,j) equal to (1/4)(S_l^j − S_j^j). The displayed S_t instead has, for row l and column j, (1/4)(S_j^l − S_j^j). This is not a notational difference: S_j^l is n×k_l while S_j^j is n×k_j, so the subtraction is dimensionally inconsistent when k_l ≠ k_j, and even when k_l = k_j the value differs unless S_l^j = S_j^l. A concrete asymmetric example reproduces the mismatch. Since Θ_t, H_t, and the Riccati system (4.13)–(4.15) are built from this S_t, the feedback û_s = −H_s^{-1}(Θ_s X̂_s + ϑ_s) is not the minimizer of the actual Φ unless S_i terms vanish (as in the network application). Thus the general claim that minimization of the α-potential reduces to the stated finite-dimensional LQ problem is not established as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops an α-potential game approach for N-player stochastic linear-quadratic differential games. In the closed-loop setting with multiplicative noise, it derives probabilistic and PDE representations for first- and second-order linear derivatives of players' costs and proves their equivalence (Theorem 3.1). In the open-loop setting, it constructs an α-potential function from the linear-derivative formula, obtains an explicit bound on α in terms of model coefficients and an H2-control radius (Proposition 4.2), and reduces minimization of the potential to an extended-state LQ control problem whose Riccati system yields a candidate open-loop α-Nash equilibrium (Section 4.2, Theorem 4.1). The paper then applies the method to a network LQ game from [6], showing that the feedback obtained from the reduced Riccati system coincides with the feedback from the conditional McKean-Vlasov formulation (Proposition 5.1, Corollary 5.1).","tokens_in":29875,"tokens_out":21131,"duration_ms":203185,"significance":"If the results hold as stated, the paper gives a useful bridge between the abstract α-potential framework and explicit stochastic LQ computations. The strengths are concrete: Theorem 3.1 is proved by a detailed Itô calculation; Lemma 4.2 supplies explicit constants for the variational estimates; Proposition 5.1 gives a clean algebraic proof of equivalence with the recalled formulation; and Example 4.1 shows a case where the new bound is sharper than a previous BSDE-based estimate. The main caveat is that the general LQ reduction in Section 4.2 rests on a quadratic representation of the potential whose S_t matrix is mis-indexed; until that is corrected, the general α-Nash statement is not established. The Section 5 application, which has S_i = 0, is not affected.","major_comments":[{"comment":"Expanding the integrand F in (4.4) shows that the coefficient of y_l^T u_j is (1/2)(S_l^j - S_j^j): the term (1/2) y_l^T S_l u gives (1/2) S_l^j, and the term u_j^T (S_j^j)^T (x - (1/2) sum_i y_i) gives -(1/2) S_j^j. Since the quadratic representation is written as 2 x^T S_t u, the block (l,j) of S_t for l,j >= 2, l != j should therefore be (1/4)(S_l^j - S_j^j), with analogous corrections in the first column. The displayed S_t instead places (1/4)(S_j^l - S_j^j) in that block. This is not a notational variant: S_j^l has dimension n x k_l and cannot occupy an n x k_j block when k_l != k_j, and even when k_l = k_j the two expressions differ unless S_l^j = S_j^l. Since Theta_s, H_s, the Riccati system (4.13)-(4.15), and the feedback u_hat_s = -H_s^{-1}(Theta_s X_hat_s + vartheta_s) are all built from this S_t, the proof that u_hat minimizes the actual potential Phi is not valid for general cross terms S_i. The Section 5 application has S_i = 0 and is unaffected, but the general claim in Theorem 4.1 and the abstract requires correcting S_t and re-deriving the subsequent LQ formulas.","section":"Section 4.2 (displayed S_t before Eq. (4.13))"}],"minor_comments":[{"comment":"The abstract should qualify the alpha-Nash equilibrium claim with the hypotheses of Theorem 4.1 (Riccati solvability, HJB regularity, and admissibility of the feedback control).","section":"Abstract"},{"comment":"The symbol V is used both for the value function defined in (4.10) and for the verification candidate; please use distinct symbols, for example V and tilde V, to avoid confusion in the statement and proof.","section":"Theorem 4.1"},{"comment":"The displayed matrices Q_t, S_t, R_t would be easier to verify if they carried equation numbers; in particular, the reader needs to compare them directly with (4.4).","section":"Section 4.2"},{"comment":"The comparison with the BSDE-based estimate in [5] is terse: the value L_y^b = kappa/N and the resulting O(N^{-1}) bound are asserted without derivation. Since this example is used to advertise the sharpness of Proposition 4.2, please expand the calculation.","section":"Example 4.1"},{"comment":"Section 3 assumes a one-dimensional Brownian motion while Section 4 uses d_W dimensions; please state explicitly that the results of Section 3 extend componentwise to multidimensional noise.","section":"Sections 3 and 4"},{"comment":"There are several typos and notation slips, for example 'Itˆ o' for Ito in the proof of Theorem 3.1, 'a d W-dimensional' in Section 4, and 'F s' in Section 5.1; a careful proofreading pass is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript is within scope for math.OC. The main technical defect is the mis-indexed S_t matrix in Section 4.2; it is local and fixable, and the rest of the paper (Theorem 3.1, Proposition 4.2, and Section 5) appears sound. I recommend major revision rather than rejection. The reference list is heavily drawn from the same research group, which is natural for this line of work and does not appear to create circularity; however, the authors should double-check that the imported Theorem 2.1 is stated with the same hypotheses as in [6]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the Hao-Shi paper on α-potential games for N-player LQ differential games. The paper has a real contribution in Section 3 and 4.1, but the central reduction in Section 4.2 contains a load-bearing algebraic error that invalidates the main theorem as stated.\n\nWhat is genuinely new: the closed-loop linear derivative representations for multiplicative noise (Theorem 3.1) are carefully done. The explicit α bound in Proposition 4.2, with the blockwise constants, is a solid piece of work, and the comparison with the BSDE bound in Example 4.1 shows a real improvement. The network example in Section 5 is a nice consistency check: the reduced Riccati system is shown to match the conditional McKean-Vlasov characterization, and that part checks out.\n\nThe problem is the quadratic representation in Section 4.2. The stress-test note is correct. The S_t matrix has the indices swapped. From (4.3)-(4.4), the coefficient of y_l^T u_j is 1/2 S_l^j - 1/2 S_j^j, so the block should be 1/4(S_l^j - S_j^j). The displayed S_t has 1/4(S_j^l - S_j^j). That's dimensionally inconsistent when k_l ≠ k_j, and wrong even when the dimensions match. Since Θ_t and the Riccati system (4.13)-(4.15) are built from this S_t, the feedback u_hat does not minimize the actual Φ for games with nonzero S_i. So Theorem 4.1's verification argument does not apply to the real potential function. The network example survives only because S_i = 0 there.\n\nThis is a big flaw, because the reduction to a finite-dimensional LQ problem is the paper's headline result. It's fixable in principle—replace S_j^l with S_l^j and redo the algebra—but as written the general claim is unproven. The abstract also drops the conditional hypotheses of Theorem 4.1, which is a lesser but real issue.\n\nI'd still send this to a referee: the framework is useful, and the error is located. But the referee should be told to check the S_t matrix carefully. My recommendation is major revision, not acceptance.","headline":"A useful framework with a load-bearing index error in the central S_t matrix; Section 4.2 needs a major fix before the main result can stand.","tokens_in":30386,"tokens_out":8331,"would_cite":false,"duration_ms":70861,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A23","49N10","93E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"One function yields approximate Nash equilibria in stochastic LQ games","keywords":["stochastic differential games","linear-quadratic games","alpha-potential games","approximate Nash equilibrium","linear derivatives","Riccati equations","multiplicative noise","open-loop equilibrium"],"falsifier":"Compute, in a two-player version of the network LQ model, the actual best-response gain a player can obtain by deviating from the potential-minimizing control, and compare it with $L^2\\max_i\\sum_{j\\ne i}\\Lambda_{ij}$; if the gain exceeds the bound for some admissible $L$, Proposition 4.2 is wrong. Alternatively, exhibit coefficients satisfying the standing assumptions for which the Riccati system has no solution on $[t,T]$ with $H_s$ uniformly positive definite, which would make Theorem 4.1 inapplicable.","tokens_in":29226,"feed_emoji":"🎯","tokens_out":11027,"duration_ms":92516,"temperature":0.7,"pith_summary":"This paper is trying to establish that a large class of $N$-player stochastic linear-quadratic differential games—where the state evolves linearly, costs are quadratic, and noise multiplies both state and controls—can be encoded by a single $\\alpha$-potential function. In the open-loop setting, the potential is built from the players' linear cost derivatives, and the approximation error $\\alpha$ is bounded explicitly by the model coefficients and the squared radius $L^2$ of the admissible control set. Minimizing this potential function is reduced to a standard finite-dimensional stochastic LQ control problem, whose Riccati solution produces a feedback control that is an open-loop $\\alpha$-Nash equilibrium. The paper also shows that in a known network LQ game, this construction reproduces the existing feedback equilibrium through a direct Riccati system rather than a conditional-law formulation. That gives a practical route from a hard multi-player equilibrium search to solving one classical control problem.","feed_headline":"One function yields approximate Nash equilibria in stochastic LQ games","feed_subtitle":"The approximation error is explicit, and the equilibrium comes from solving a Riccati equation.","key_machinery":"The central object is the linear derivative $\\frac{\\delta J_i}{\\delta u_i}$ of a cost functional with respect to player $i$'s strategy, which measures how a unilateral perturbation of that player's control changes her cost. The $\\alpha$-potential function $\\Phi$ is defined as the path integral of these derivatives along the line from zero to $u$, and the load-bearing identity $X^{ru}=X^u-(1-r)\\bar Y^u$ expresses the state under the scaled control $ru$ in terms of the original state and the summed variational process, allowing $\\Phi$ to be written as the cost of an augmented LQ state $(X,Y_1,\\dots,Y_N)$. The associated Riccati system with the positivity condition $H_s\\ge\\delta I$ then turns potential minimization into a classical finite-dimensional stochastic LQ problem.","core_discovery":"Define the open-loop $\\alpha$-potential function by integrating each player's first-order linear derivative along the line segment from zero control to $u$: $\\Phi(u) = \\int_0^1 \\sum_i \\frac{\\delta J_i}{\\delta u_i}(ru; u_i)\\,dr$. The paper proves that, when all controls are $H^2$-bounded by $L$, $\\Phi$ is an $\\alpha$-potential function with $\\alpha \\le L^2 \\max_i \\sum_{j\\ne i}\\Lambda_{ij}$, where each $\\Lambda_{ij}$ is an explicit blockwise combination of coefficient norms and variational-process bounds. Using the identity $X^{ru}=X^u-(1-r)\\bar Y^u$, the minimization of $\\Phi$ is rewritten as a finite-dimensional LQ problem in the augmented state $(X,Y_1,\\dots,Y_N)$, so the classical verification theorem applies: if the associated Riccati system admits a solution with $H_s\\ge\\delta I$ and the feedback $\\hat u_s=-H_s^{-1}(\\Theta_s\\hat X_s+\\vartheta_s)$ is admissible, then $\\hat u$ minimizes $\\Phi$ and is an open-loop $\\alpha$-Nash equilibrium. Before this, the paper establishes equivalence between probabilistic and PDE representations of the first- and second-order linear derivatives for the closed-loop LQ game with multiplicative noise, which supplies the derivative formulas the potential construction uses.","pith_inferences":["The paper leaves implicit that the same augmented-state reduction could be applied to closed-loop policy classes if the potential's minimizer is allowed to depend on the variational states, since the obstacle is only the restricted admissible class.","A testable extension is to let the control radius $L$ shrink with $N$; in the network example the bound is $O(1/N^2)$, suggesting that large sparse games can have near-exact potential structure even when the exact potential condition fails.","Because the $\\alpha$ bound is stated blockwise in terms of coefficient differences, games with nearly symmetric cross-player costs should admit near-zero $\\alpha$; quantifying that near-symmetry is a natural next step.","The equivalence between probabilistic and PDE derivative representations suggests that derivative-based learning algorithms for stochastic LQ games could use either representation to estimate the potential function, a direction not pursued here."],"forward_implications":["A single Riccati-equation computation replaces the search for an open-loop Nash equilibrium in this class of stochastic LQ games.","The approximation error grows at most like $L^2$ times a coefficient-dependent constant, so shrinking the admissible control radius tightens the $\\alpha$-Nash guarantee.","When the mixed second-order linear derivatives are symmetric, the same construction yields an exact potential game with $\\alpha=0$.","The probabilistic and PDE derivative representations agree, so derivatives can be computed either by simulation or by solving ODE systems, whichever is more convenient.","In the network LQ game, the equilibrium is obtained without introducing an auxiliary random variable, since the reduced finite-dimensional LQ problem reproduces the same feedback policy."],"supporting_citations":[{"why":"Supplies the theorem that an $\\alpha$-potential minimizer is an $\\alpha$-Nash equilibrium and the definition of the $\\alpha$-potential constructed from linear derivatives.","marker":"[6, 7]"},{"why":"Introduces the linear-derivative calculus and gives the closed-loop LQ representation that Section 3 extends to multiplicative noise.","marker":"[9]"},{"why":"Gives the BSDE-based $\\alpha$-estimate that Example 4.1 compares against to show the blockwise bound is sharper.","marker":"[5]"},{"why":"Provides the classical verification theorem used to solve the reduced finite-dimensional LQ control problem.","marker":"[10]"}],"fun_headline_variants":["Alpha-potential functions crack N-player stochastic LQ games","Explicit alpha bound yields Nash equilibria in stochastic games","Stochastic LQ games: one potential function ties it together","From derivatives to equilibria: alpha-potential approach for N players","Potential function reduces stochastic LQ to finite-dimensional control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result is conditional on the Riccati equations admitting a well-behaved solution with $H_s\\ge\\delta I$ and on the feedback policy they generate being an allowed square-integrable control; the paper assumes both rather than proving them for general coefficient choices.","fun_headline_variants_meta":{"raw":{"variants":["Alpha-potential functions crack N-player stochastic LQ games","Explicit alpha bound yields Nash equilibria in stochastic games","Stochastic LQ games: one potential function ties it together","From derivatives to equilibria: alpha-potential approach for N players","Potential function reduces stochastic LQ to finite-dimensional control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000633,"raw_usage":{"total_tokens":2985,"prompt_tokens":1073,"completion_tokens":1912,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":1839}},"tokens_in":689,"tokens_out":1912,"duration_ms":13280,"temperature":1.0,"reasoning_tokens":1839,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:33:41.027149+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, in a two-player version of the network LQ model, the actual best-response gain a player can obtain by deviating from the potential-minimizing control, and compare it with $L^2\\max_i\\sum_{j\\ne i}\\Lambda_{ij}$; if the gain exceeds the bound for some admissible $L$, Proposition 4.2 is wrong. Alternatively, exhibit coefficients satisfying the standing assumptions for which the Riccati system has no solution on $[t,T]$ with $H_s$ uniformly positive definite, which would make Theorem 4.1 inapplicable.","supporting_citations":[{"cited_title":"Guo and Y","cited_arxiv_id":null,"evidence_quote":"Introduces the linear-derivative calculus and gives the closed-loop LQ representation that Section 3 extends to multiplicative noise."},{"cited_title":"Pham,Continuous-Time Stochastic Control and Optimization with Financial Applica- tions, Springer, 2009","cited_arxiv_id":null,"evidence_quote":"Provides the classical verification theorem used to solve the reduced finite-dimensional LQ control problem."}],"review_version":1}