{"id":"43b9480e-154b-4e89-afe8-49d5d15fa9e8","arxiv_id":"2412.18802","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper characterizes feedback Stackelberg equilibria for forward linear-quadratic stochastic games with affine constraints under four assumptions (H1-H4).","lead":"This paper derives feedback Stackelberg equilibria for linear-quadratic stochastic differential games in which the leader faces affine constraints, using Riccati equations, FBSDEs, and Lagrangian duality. It provides a dual problem and KKT condition, with two examples that appear to contain algebraic errors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The examples do not verify H3: the claimed solution of the stochastic Riccati equation (19) violates its terminal condition, so the main theorem's hypotheses have no demonstrated instance.","rationale":"The reader's weakest-assumption identification is exactly where the paper is most exposed: H3 is an assumption, not a theorem, and the only concrete evidence offered for it is wrong. My independent check of the terminal condition in (19) and the residual in (8) confirms that the Section 5 examples are not valid instantiations of the hypotheses. This does not make the conditional statement false, but it does mean the abstract's claim that the examples 'support our main results' is unsupported. Because the reader already returned CONDITIONAL, no verdict change is needed; the concern lands exactly on the same load-bearing assumption and reinforces the need for revision rather than altering the level of confidence.","tokens_in":169,"tokens_out":10410,"duration_ms":106399,"concrete_test":"For the data of Section 5, enforce the terminal condition phi2(1) = diag(-1,0) and solve (19) backward, first with the ansatz ephi2 = 0 so that the Riccati block becomes a deterministic 2x2 matrix ODE; then check whether the resulting solution is bounded on [0,1] and whether substituting it into (33) gives a control satisfying the constraint with equality and strong duality (26). If the ODE has no bounded solution, H3 fails in the paper's only example; if it does, recompute the example's lambda* and the equilibrium formulas and compare them with the paper's values.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing object in Theorem 4.5 is the pair (phi2, ephi2) solving the nonstandard stochastic Riccati equation (19); H3 merely assumes this solution exists. From it are built the leader feedback (21), the KKT system (32), and the equilibrium (33). In Section 5 the alleged solution is phi2 = diag(-1, 1 - e^{-t}), ephi2 = 0. This cannot satisfy (19): the terminal condition stated in (19) is phi2(1) = diag(G2,0) = diag(-1,0), while the proposed matrix has bottom-right entry 1 - e^{-1} at t = 1, not 0. The companion claim phi1 = (1,0) also fails (8): with A=C=1, D1=-2, S1=1/16, the bracket in (8) is -2+1+1+1-1/16 = 15/16, not 0. Thus the examples give no valid instance satisfying H1/H3, and the paper gives no sufficient condition under which H3 holds. The conditional theorem may still be true, but its advertised non-degenerate illustrations do not support it, and the existence of any model satisfying the key hypothesis remains open.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a non-zero-sum linear-quadratic stochastic Stackelberg differential game with affine constraints on the leader's admissible strategies, where the constraints depend on the state, the follower's best response, and the leader's control. It derives the follower's feedback strategy from a stochastic Riccati equation (SRE), reformulates the leader's problem through a fully coupled FBSDE and a nonstandard SRE, and constructs a Lagrangian dual. Under assumptions (H1)-(H4) it proves strong duality, gives a KKT condition, and states a feedback Stackelberg equilibrium in Theorem 4.5. Two examples with indefinite coefficients are presented as illustrations.","tokens_in":21364,"tokens_out":11289,"duration_ms":97955,"significance":"The paper's architecture—constraint reformulation via FBSDEs, dual problem, KKT conditions, and a positivity condition for uniqueness—is a reasonable extension of known techniques and, if the hypotheses can be guaranteed, would provide a useful recipe for constrained leader-follower LQ games. It is a strength that the main result is stated as a conditional theorem with explicit hypotheses rather than as a formal derivation from weaker assumptions. However, the paper's only nontrivial instances verifying the key hypothesis (H3) are the Section 5 examples, and these examples contain checkable algebraic errors; moreover, H3 itself is assumed with no sufficient condition. Thus the current manuscript does not demonstrate that its framework applies to any non-degenerate model. The contribution would be significant after the examples are corrected and H3 is supported.","major_comments":[{"comment":"The claimed solution (φ1,eφ1)=(1,0) does not solve SRE (8). For the data of (34)-(35) one has A=C=1, B1=1/2, E1=4, D1=-2, G1=1, hence S1=1/16 and the drift bracket in (8) equals D1+φ1A+A⊤φ1+C⊤eφ1+eφ1C+C⊤φ1C−φ1S1φ1 = -2+1+1+0+0+1−1/16 = 15/16 ≠0. Therefore dφ1 = −15/16 dt, not zero, and the feedback formula (10) in the examples is not established. This is not a typographical slip in the final formula only; the same φ1 is used in the construction of ψ1 and in the leader's SRE data.","section":"Section 5, Eq. (8)"},{"comment":"The proposed φ2 = diag(-1,1−e^{-t}) violates the terminal condition of (19): φ2(1) must equal diag(G2,0)=diag(-1,0), but its (2,2)-entry at t=1 is 1−e^{-1}≠0. Since (19) is the nonstandard SRE whose solvability is exactly assumption (H3), and since (21), (32), and (33) all depend on that solution, neither Example 5.1 nor Example 5.2 verifies the hypotheses of Theorem 4.5. The printed Stackelberg equilibrium formulas in these examples therefore do not follow from the stated assumptions.","section":"Section 5, Eq. (19)"},{"comment":"The paper assumes existence and uniqueness of a solution to the stochastic Riccati equation (19) without providing any sufficient condition. The matrix F in (17) is indefinite (with −S1 on the off-diagonal) and the terminal data are singular, so this assumption is not covered by the standard Riccati theory invoked for (8). Because the leader's feedback strategy (21), the KKT system (32), and the equilibrium (33) all require φ2, the class of models for which the main theorem is applicable remains unspecified. A proof of H3 under checkable conditions, or at least corrected examples with a verifiable solution, is needed.","section":"Section 4, Assumption (H3)"}],"minor_comments":[{"comment":"The text says that eJ(λ*,u*_2(λ*)) is 'strictly concave' in a and reaches its minimum at a=0. With λ*=2a/(e^{-2}-1), the expression eJ=-1-2aλ*-(1-e^{-2})λ*^2/2 equals -1+2a^2/(1-e^{-2}), which is strictly convex in a, not concave; the stated 'minimum at a=0' is consistent with convexity, so the wording should be corrected.","section":"Section 5, Example 5.1"},{"comment":"With λ* = max(2a/(e^{-2}-1),0), the value function equals -1 for a≥0 and is strictly convex in a for a<0; the minimum is attained for all a≥0, not 'when a≤0' as stated. The sentence about strengthening constraints should be revised accordingly.","section":"Section 5, Example 5.2"},{"comment":"The admissible space in the infimum is written as L^2_F([s,T],R^n), but u2 is R^m-valued; it should be L^2_F([s,T],R^m).","section":"Problem 4.1"}],"recommendation":"major_revision","confidential_remarks":"The algebraic checks in Major Comments 1 and 2 are direct and unambiguous; before resubmission the authors should either supply corrected solutions of (8) and (19) for their examples or replace the examples. Because H3 is simply assumed and the paper acknowledges in the Conclusion that the positive-definiteness of S0 needs further research, the advertised contribution is currently weaker than the claims in the Abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a reasonable conditional-theory paper whose advertised examples are wrong. The problem—forward LQ Stackelberg games with affine constraints on the players' strategies—is new, and the FBSDE rewriting that makes the affine constraints depend only on the leader's strategy is a genuine technical device. The Lagrangian-duality/KKT structure is the standard one for this literature, but here it is applied to an indefinite setting where it is not routine, and Theorem 4.5 is a defensible conditional statement under (H1)-(H4).\n\nThe soft spot is Section 5, and it is not minor. The claimed solution (φ1, eφ1) = (1, 0) does not satisfy SRE (8): with A=C=1, D1=-2, S1=1/16, the bracket is 15/16, not 0. The claimed (φ2, eφ2) = (diag(-1, 1-e^{-t}), 0) fails the terminal condition of (19), which forces φ2(1)=diag(-1, 0). So H3 is not verified, and the two 'non-degenerate examples' illustrate nothing except the need for revision. Since H3 is a bare existence/uniqueness assumption on a nonstandard stochastic Riccati equation with indefinite F, the paper currently has no demonstrated instance satisfying its own hypotheses.\n\nSmaller gripes: Example 5.2 is called an equality constraint but is written with ≤; the separation proof of Theorem 4.3 is a sketch; and the paper would be stronger with any sufficient condition for H3. The citation pattern is fair, including the authors' prior work, which is the right motivation.\n\nWho this is for: specialists in stochastic control and dynamic games. They would get real value from the constraint-rewriting method and the dual formulation, and they would want to know the main theorem once the examples are repaired or replaced. I would send it to a referee, with a note that the examples are algebraically wrong and the existence of a valid instance satisfying H1-H4 is open. After a substantive revision, the paper is likely publishable.","headline":"The problem and framework are a genuine extension of the constrained LQ-SSDG literature, but the paper's own examples fail to verify the key solvability assumption H3, so the submission is not ready as is.","tokens_in":21872,"tokens_out":4266,"would_cite":false,"duration_ms":57731,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N70","91A15","91A65","93B52"],"pacs":[],"model":"deepseek-v4-flash","headline":"Constrained leader-follower stochastic games get explicit feedback equilibria","keywords":["linear-quadratic games","Stackelberg differential games","affine constraints","feedback Stackelberg equilibrium","KKT conditions","stochastic Riccati equations","forward-backward SDEs","Lagrangian duality"],"falsifier":"In Example 5.1, evaluate the proposed $(\\phi_2,\\tilde\\phi_2)$ at $t=1$ and compare with the terminal condition $\\phi_2(1)=\\operatorname{diag}(G_2,0_{n\\times n})$; if the entries do not match, the example does not actually verify (H3).","tokens_in":20906,"feed_emoji":"🎯","tokens_out":7327,"duration_ms":58789,"temperature":0.7,"pith_summary":"The paper sets out to extend the linear-quadratic stochastic Stackelberg differential game to the case where the leader's admissible strategies are defined by affine equality and inequality constraints coupling the state, the follower's best response, and the leader's own control. It claims that, under four assumptions (H1)-(H4), a feedback Stackelberg equilibrium exists and can be written explicitly as affine functions of an augmented state and a Lagrange multiplier. The central technical step rewrites the affine constraints as constraints on the leader's strategy alone, using a new FBSDE, which makes the Slater condition easy to check. If the main theorem is right, constrained leader-follower stochastic control problems become solvable by a concrete KKT recipe rather than by separate case analyses.","feed_headline":"KKT recipe solves constrained stochastic leader-follower games","feed_subtitle":"A dual problem, Slater condition, and Riccati equations turn affine constraints into a solvable KKT system.","key_machinery":"The proof chain uses three linked devices. First, the follower's optimal strategy is put in feedback form $u_1^* = -E_1^{-1}B_1^\\top(\\phi_1 X^* + \\psi_1)$ through the stochastic Riccati equation (8) and the BSDE (9). Second, the leader's relaxed problem is rewritten as the fully coupled FBSDE (17), and a four-step scheme reduces solvability to the nonstandard stochastic Riccati equation (19) for $(\\phi_2,\\tilde\\phi_2)$, whose unique solution is assumed in (H3). Third, Lagrangian duality replaces the constrained leader problem by the dual problem (4.2), and the KKT system (32) ties the dual multiplier to the feedback gains, yielding the equilibrium pair (33).","core_discovery":"The paper's central claim is Theorem 4.5: under (H1)-(H4), the pair $(\\hat u_1,\\hat u_2)$ defined by (33) is a feedback Stackelberg equilibrium of Problem 2.1. The optimal strategies are affine in the augmented state $\\hat Z$ and the dual multiplier $\\lambda^*$, and $\\lambda^*$ is characterized by the KKT conditions (32), with $\\rho_i(\\lambda^*)-a_i$ giving the slack of the $i$-th affine constraint. When the matrix $S_0$ in (23) is strictly positive definite, $\\lambda^*$ is the unique maximizer of the dual problem (Problem 4.2).","pith_inferences":["A natural follow-up is to test the Slater-condition verification on mean-field or jump-diffusion Stackelberg models, where the same FBSDE trick may bypass the usual verification bottleneck.","The positive-definiteness of $S_0$ could be checked numerically before solving the game; if it fails, the dual maximizer may be nonunique, and the equilibrium selection would need extra criteria.","If correct, the affine-constraint machinery could be combined with the approximation in Example 1.1 to handle quadratic and risk constraints in Stackelberg games.","The new FBSDE rewriting of the constraints appears to be a transferable tool, likely useful in other hierarchical control problems where the Slater condition is the main obstruction."],"forward_implications":["Constrained LQ Stackelberg games are reduced to solving a stochastic Riccati equation and a KKT system, so standard numerical solvers for Riccati equations and quadratic programs can be applied directly.","Under (H1)-(H4) with $S_0\\in\\mathbb{S}^l_{++}$, the dual problem has a unique solution, so the KKT system (32) has a unique $\\lambda^*$ and the feedback equilibrium is uniquely identified.","The affine constraints can be checked through the explicit formula for $\\rho(u_2^*(\\lambda))$ in Proposition 4.1, giving a computable membership test for the leader's admissible set.","The method generalizes the affine-constraint approach previously developed for single-player stochastic LQ control, pointing to a uniform treatment of constraints in hierarchical stochastic control."],"supporting_citations":[{"why":"Supplies existence and uniqueness of solutions to the stochastic Riccati equation (8) used for the follower's feedback strategy.","marker":"[22]"},{"why":"Provides solvability of the BSDEs and FBSDEs used for $\\psi_1,\\psi_2$, and the augmented state $Z$.","marker":"[29]"},{"why":"Supplies the four-step scheme that transforms the coupled FBSDE (17) into the Riccati equation (19).","marker":"[13]"},{"why":"Establishes the feedback representation for linear-quadratic leader-follower games that the present construction extends to constraints.","marker":"[28]"},{"why":"Introduces the affine-constraint approximation for stochastic LQ control that motivates the constraint class studied here.","marker":"[7]"},{"why":"Provides the perturbation analysis background used for the KKT condition and strong duality in Theorem 4.4.","marker":"[2]"}],"fun_headline_variants":["KKT conditions crack constrained leader-follower games","Dual problem unlocks constrained stochastic Stackelberg","Affine constraints solved in stochastic games via KKT","Feedback equilibrium via KKT for constrained Stackelberg","Riccati and duality tame affine constraints in games"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Assumption (H3), posed in Section 4 before Theorem 4.2, states that the nonstandard stochastic Riccati equation (19) for $(\\phi_2,\\tilde\\phi_2)$ has a unique solution; all feedback formulas for the leader, the KKT system, and the final equilibrium depend on this solution, and the paper assumes it rather than proving a sufficient condition.","fun_headline_variants_meta":{"raw":{"variants":["KKT conditions crack constrained leader-follower games","Dual problem unlocks constrained stochastic Stackelberg","Affine constraints solved in stochastic games via KKT","Feedback equilibrium via KKT for constrained Stackelberg","Riccati and duality tame affine constraints in games"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1561,"prompt_tokens":841,"completion_tokens":720,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":646}},"tokens_in":457,"tokens_out":720,"duration_ms":24423,"temperature":1.0,"reasoning_tokens":646,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:29:04.200599+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In Example 5.1, evaluate the proposed $(\\phi_2,\\tilde\\phi_2)$ at $t=1$ and compare with the terminal condition $\\phi_2(1)=\\operatorname{diag}(G_2,0_{n\\times n})$; if the entries do not match, the example does not actually verify (H3).","supporting_citations":[{"cited_title":"Indefinite stochastic linear-quadratic optimal control problems with random coefficients: Closed-loop representation of open-loop optimal controls","cited_arxiv_id":null,"evidence_quote":"Supplies existence and uniqueness of solutions to the stochastic Riccati equation (8) used for the follower's feedback strategy."},{"cited_title":"Stochastic optimal control–A concise introduction","cited_arxiv_id":null,"evidence_quote":"Provides solvability of the BSDEs and FBSDEs used for $\\psi_1,\\psi_2$, and the augmented state $Z$."},{"cited_title":"Forward-backward Stochastic Differential Equations and their Applications","cited_arxiv_id":null,"evidence_quote":"Supplies the four-step scheme that transforms the coupled FBSDE (17) into the Riccati equation (19)."},{"cited_title":"A leader-follower stochastic linear quadratic differential game.SIAM Journal on Control and Optimization , 41(4):1015–1041, 2002","cited_arxiv_id":null,"evidence_quote":"Establishes the feedback representation for linear-quadratic leader-follower games that the present construction extends to constraints."},{"cited_title":"Stochastic linear-quadratic control problems with affine constraints","cited_arxiv_id":null,"evidence_quote":"Introduces the affine-constraint approximation for stochastic LQ control that motivates the constraint class studied here."},{"cited_title":"Fr´ ed´ eric Bonnans and Alexander Shapiro.Perturbation Analysis of Optimization Problems","cited_arxiv_id":null,"evidence_quote":"Provides the perturbation analysis background used for the KKT condition and strong duality in Theorem 4.4."}],"review_version":1}