{"id":"9e2a77ba-87b2-4127-acc9-1107639f3fb9","arxiv_id":"2507.19965","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A single-trajectory model-free inverse LQR method is reformulated as a convex conic feasibility problem and solved by BSUM with an O(1/k) sublinear convergence rate claim.","lead":"This paper develops a method for recovering an expert's system dynamics and cost function from a single observed optimal trajectory, converting the non-convex estimation problem into a convex feasibility problem with an iterative solver. It claims the first explicit convergence-rate guarantee for model-free inverse optimal control, which could matter for safety-critical imitation learning and human-robot interaction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The dual in Theorem 3.8 is not the true Lagrangian dual: Eq. (18) requires U^T λ in range(Ω^TΩ), but Ω is rank-deficient by the scaling symmetry, so the O(1/k) result targets a pseudo-objective.","rationale":"The paper's goal is to give the first model-free IOC method with a provable O(1/k) convergence rate, and its route is: reformulate the non-convex problem as a convex conic feasibility problem, pass to the Lagrangian dual, solve the dual by BSUM, and cite Hong et al. for the rate. The most load-bearing step is therefore the correctness of the dual problem (19). That step fails: the pseudoinverse stationary-point formula is used without the necessary range condition, and the range condition is not vacuous because Ω has a null vector given by scaling the true parameters. This is an internal inconsistency, not a disagreement with community consensus. The concrete test above would settle it by exhibiting a feasible dual multiplier for which the true dual function is -∞ while the paper's formula is finite. The reader's weakest assumption—boundedness of the dual level set—is real and also stems from the singularity of Ω^TΩ, but it is downstream: even a bounded level set would not make Algorithm 1 a solver for the true dual if the dual derivation is invalid. The simulations show accurate trajectory reproduction, but they do not validate the theoretical dual and no code or derivative-estimation details are provided, so they cannot override the analytical failure. For these reasons the rejection stands, but on a different and more fundamental ground than the reader's stated weakest assumption.","tokens_in":17786,"tokens_out":11234,"duration_ms":158472,"concrete_test":"Using the paper's 3×3, m=2 example, assemble Ω from Eq. (15). (1) Verify numerical rank deficiency: form v = (vec(A^⊤P), vec(R), vec(Q), vec(P), vec(PA_K))^T from the ground-truth parameters and check Ωv = 0 to machine precision, so v ∈ Nul(Ω). (2) Set λ = (0,0,vec(I_n)) and compute the true value of L(ξ0 + t v; λ) for t→-∞; it diverges to -∞ while formula (19) returns the finite value (1/4)λ^THλ + λ^TW. This directly disproves Theorem 3.8. (3) Run Algorithm 1 from the stated initialization and compute Ωξ* at the returned ξ*; if the residual is nonzero, the algorithm has not solved primal problem (16). These checks are independent of any simulation details.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee rests on Lemma 3.7 and Theorem 3.8. The stationary condition in Lemma 3.7 is 2Ω^TΩξ + U^Tλ = 0. Since Ω^TΩ is positive semidefinite, this equation has a solution if and only if U^Tλ ∈ Ran(Ω^TΩ) = Nul(Ω)^⊥; otherwise L(ξ;λ) is unbounded below and the true dual function g(λ) is -∞. The paper substitutes the pseudoinverse ξ*(λ) = -(1/2)(Ω^TΩ)^†U^Tλ and obtains the finite quadratic (19), but it never verifies the range condition. The condition fails already because the constraints are scale-invariant: if (Z,Q,R,P,G) is feasible, so is c(Z,Q,R,P,G), so v = (vec(Z), vec(Q), vec(R), vec(P), vec(G))^T with ground-truth values satisfies Ωv = 0. For λ = (0,0,λ_P) with λ_P = I, U^Tλ·v = trace(P*) > 0, so U^Tλ ∉ Nul(Ω)^⊥. Along t v, ||Ω(tv)||² = 0 and L = t·trace(P*) - λ^TW, which tends to -∞ as t→-∞; hence g(λ) = -∞, whereas (19) gives a finite value. Therefore Theorem 3.8 is false as stated, Algorithm 1 minimizes a pseudo-dual, and ξ*(λ) from (18) need not be primal feasible. The level-set unboundedness identified by the reader is a symptom of this same rank deficiency, but boundedness alone cannot repair the argument.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies model-free inverse optimal control for continuous-time LQR systems, where both the dynamics (A,B) and cost weights (Q,R) are unknown. It proposes to identify the optimal feedback gain K* from a single expert trajectory, then reformulates the joint estimation problem as the convex feasibility problem (14) and the equivalent least-squares conic program (16). The authors derive a Lagrangian dual (19), solve it with a block successive upper-bound minimization algorithm, and claim a sublinear O(1/k) convergence rate. Simulation comparisons with Pontryagin Differentiable Programming are reported as showing faster and more accurate reconstruction.","tokens_in":18156,"tokens_out":7705,"duration_ms":94594,"significance":"If the central claims were correct, the paper would make a substantial contribution: a convex reformulation of a nonconvex inverse LQR problem, a closed-form block-coordinate solver, and the first explicit non-asymptotic convergence-rate guarantee for model-free IOC. The paper contains several genuinely useful pieces: Lemma 3.3 and Lemma 3.4 give a clean trajectory-equivalence argument, and the SDP feasibility formulation in (14) is a reasonable convex relaxation of the original problem. However, the load-bearing dual derivation in Lemma 3.7 and Theorem 3.8 is invalid for the stated problem, and the convergence-rate bound in Theorem 3.13 rests on unverified assumptions. Because these flaws directly undermine the algorithm's claimed guarantees, the paper cannot be accepted in its present form.","major_comments":[{"comment":"The derivation of the dual (19) is invalid because the paper replaces the stationary condition 2Ω^TΩξ + U^Tλ = 0 by the pseudoinverse formula (18) without checking that U^Tλ ∈ Ran(Ω^TΩ). If this range condition fails, the Lagrangian L(ξ;λ) is unbounded below and the true dual function is −∞, not the finite quadratic in (19). The condition fails in the present problem: the equality constraints in (14) are scale-invariant, so for any feasible v = (vec(Z), vec(R), vec(Q), vec(P), vec(G))^T built from ground-truth parameters, Ωv = 0; taking λ = (0, I, 0) with the identity multiplier for P gives U^Tλ · v = trace(P*) > 0, so U^Tλ ∉ Nul(Ω)^⊥ = Ran(Ω^TΩ), and along ξ = tv with t < 0, L(ξ;λ) → −∞. Thus Theorem 3.8 is false as stated, and Algorithm 1 in Section 3.2 minimizes a pseudo-objective rather than the actual dual of (16).","section":"§3.1, Lemma 3.7 and Theorem 3.8"},{"comment":"The stated O(1/k) bound depends on R = sup_{λ∈F, λ*∈Γ*_+} ∥λ−λ*∥, but the paper never proves that the level set F = {λ : Jdual(λ) ≤ Jdual(λ0)} is bounded. Since H is only positive semidefinite and the PSD cones are unbounded, directions inside the feasible cone along which the quadratic part vanishes can make F unbounded or the constant σ = 1/(9∥H∥R²) degenerate; no argument rules this out. The proof merely says the problem satisfies all assumptions of Theorem 2 in [26], but boundedness and related regularity hypotheses are not verified. Consequently the advertised convergence-rate guarantee is not established.","section":"§3.2, Theorem 3.13"},{"comment":"Even if the dual sequence converged, the algorithm's output ξ*(λ^{k+1}) = −(1/2)(Ω^TΩ)^†U^Tλ^{k+1} is not shown to be primal feasible. The paper proves only a dual objective gap (under the contested Theorem 3.13); it does not show that ∥Ωξ*(λ^{k+1})∥ → 0 or that the cone constraints are satisfied, and the pseudoinverse can return a point with Ωξ ≠ 0 when U^Tλ is not in Ran(Ω^TΩ). The stopping criterion ∥λ^{k+1}−λ^k∥ < ε is a dual-iterate gap and has no proven relation to reconstruction accuracy. Hence the claim that Algorithm 1 converges to a point in the equivalence set Φ* is unsupported.","section":"§3.2, Algorithm 1 and Eq. (18)"},{"comment":"Strong duality is asserted via Slater's condition, but strict feasibility of (16) is not proved. A strictly feasible point would require Ωξ = 0 with Q ≻ 0, P ≻ εI, and R ≻ εI; the true parameters only guarantee Q ≽ 0, and Q may be singular, so no argument shows that a positive definite Q can satisfy the equality constraints. Without Slater's condition, the KKT-based recovery of a primal solution from λ* via (18) lacks a rigorous justification. This is an additional unverified assumption in the chain leading to the convergence-rate claim.","section":"Remark 3.9 and problem (16)"}],"minor_comments":[{"comment":"The notation e^AK t should be written as exp((Â − B̂K̂)t) or e^{(Â−B̂K̂)t}; as typeset it is ambiguous between e^A K t and e^{AK}t.","section":"Equation (8)"},{"comment":"The characteristic polynomial p(λ) = λ^n + a_{n−1}λ^{n−1} + ... + a_0 is followed by equations (9)–(10) that use coefficients a_1, ..., a_n inconsistently; the induction step should be re-indexed to match a_0, ..., a_{n−1}.","section":"Lemma 3.3 proof"},{"comment":"The runtime comparison (0.3 s vs. 98.4 s) is reported without hardware/software details, implementation effort, or multiple runs; as a single trial it is anecdotal. Reporting distributions over the Monte Carlo runs would better support the claimed order-of-magnitude improvement.","section":"Section 4.1 and Table 1"}],"recommendation":"reject","confidential_remarks":"The dual-function flaw in Theorem 3.8 is fundamental and affects the paper's main novelty. The authors could potentially repair the approach by adding constraints that enforce U^Tλ ∈ Ran(Ω^TΩ) or by analyzing a different dual, but the current manuscript does not contain such a repair, and the theorem as stated is false. This is a load-bearing error that cannot be fixed by local editing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The convexification is the real contribution here, and it's a good one. Rewriting the ARE using Z = A^T P and turning the joint estimation problem into an SDP feasibility problem with linear equality constraints is a clean idea. The identification of K* from the trajectory and then asking for (A,B,Q,R,P) that are consistent with it is the right way to set up model-free IOC. The block-coordinate solver is natural, and the simulations show the method works on the tested examples. If the theoretical claims were correct, the O(1/k) rate would indeed be a first.\n\nBut the theory doesn't hold. The Lagrangian dual derivation in Lemma 3.7 and Theorem 3.8 is wrong. The stationary condition 2Ω^TΩξ + U^Tλ = 0 only has a solution when U^Tλ is in the range of Ω^TΩ. The paper substitutes the pseudoinverse without checking this, and the condition fails. Because the constraints are scale-invariant, Ω has a nontrivial nullspace along the scaling direction. For a multiplier that pairs with that direction, the true dual function is -∞, while formula (19) gives a finite value. So the \"dual\" being solved by Algorithm 1 is not the Lagrangian dual of (16), and the primal recovery via (18) is not justified. The O(1/k) bound in Theorem 3.13 is a corollary of a theorem in [26] applied to this pseudo-objective, and it also assumes a bounded level set that is never verified. The rate constant R may well be infinite.\n\nOther issues are smaller but real: the abstract calls the problem SOCP when it is actually SDP; Slater's condition is asserted, not proved; and the simulations provide no code or derivative-estimation details, so the empirical comparison with PDP is not reproducible. The trajectory reconstruction is enforced in-sample by the data matrices, so the low MSE is partly by construction.\n\nI'd still like to see this line of work continue. The convexification may be salvageable by adding a normalization constraint to break the scale symmetry, or by working with the true dual restricted to the range condition. But as written, the central guarantee is not established, and the paper overstates what is proven.\n\nIf I were the editor, I'd send it to review — the idea is good enough that a careful referee report could help the authors fix the dual derivation. But I would not accept the current version. I wouldn't cite it in my own work until the theorem is repaired.","headline":"Clever convex reformulation of model-free IOC, but the O(1/k) convergence claim rests on a false Lagrangian dual derivation and the main theorem does not hold as stated.","tokens_in":18700,"tokens_out":6140,"would_cite":false,"duration_ms":68527,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N45","90C25","90C22","93B30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Model-free inverse optimal control can be reformulated as a convex program, with a provable O(1/k) convergence rate for the block-coordinate solver.","keywords":["inverse optimal control","linear quadratic regulator","model-free","convex optimization","conic programming","block coordinate descent","convergence rate","system identification"],"falsifier":"Compute the Hessian $H$ of the dual objective and look for a nonzero direction $d$ inside the semidefinite cone with $H d = 0$ and $W^\\top d = 0$; such a direction makes the objective constant along a ray, so the level-set radius $R$ is infinite, $\\sigma = 1/(9\\|H\\|R^2)$ vanishes, and the theorem's $\\mathcal{O}(1/k)$ inequality is vacuous. Run the proposed updates on an instance with a rank-deficient $Q$ on the data-supported subspace and check whether the dual gap obeys the claimed bound with the stated constants.","tokens_in":17507,"feed_emoji":"🎯","tokens_out":10651,"duration_ms":113565,"temperature":0.7,"pith_summary":"Model-free inverse optimal control asks for the unknown system matrices and unknown LQR cost weights that explain one observed optimal trajectory; this paper claims that this non-convex joint estimation problem is exactly equivalent to a convex conic feasibility problem. The key is a variable change that turns the bilinear algebraic Riccati equation into linear equalities, so the remaining constraints are positive-semidefinite cones. The paper solves the Lagrange dual with block successive upper-bound minimization, proves that every block update has a closed form, and shows the dual objective gap converges at $\\mathcal{O}(1/k)$. If the proof stands, this is the first model-free IOC method with an explicit, non-asymptotic convergence-rate guarantee, and the simulations show it reproducing the expert trajectory with mean-squared error around $10^{-6}$ in well under a second.","feed_headline":"Model-free inverse LQR becomes a convex problem with O(1/k) rate","feed_subtitle":"A block-coordinate dual solver provably closes the objective gap with closed-form updates and no line search.","key_machinery":"The load-bearing device is the variable change $Z=\\hat{A}^\\top \\hat{P}$ and $G=\\hat{P}\\hat{A}_K$, which turns the bilinear Riccati equation into linear equalities and makes the joint estimation a convex feasibility problem in the stacked variable $\\xi$. The solver works on the Lagrange dual, where the KKT relation $\\xi^*(\\lambda)=-\\tfrac12(\\Omega^\\top\\Omega)^\\dagger U^\\top\\lambda$ produces the quadratic-conic dual objective, and each of the three block updates is a closed-form resolvent/projection step. This structure is what allows the proof of the $\\mathcal{O}(1/k)$ convergence rate.","core_discovery":"The paper's central claim is that jointly recovering $(A,B,Q,R)$ from a single optimal LQR trajectory, usually regarded as a non-convex inverse problem, can be recast without loss as a convex problem. Concretely, Theorem 3.5 proves that optimal solutions of the original parameter optimization are exactly the feasible points of a system of linear equalities plus semidefinite cone constraints. Vectorizing those equalities and minimizing the residual gives a least-squares problem over PSD cones whose Lagrangian dual is a convex quadratic-conic objective; Theorem 3.13 proves that the block-coordinate dual updates reduce the objective gap at rate $\\mathcal{O}(1/k)$. The paper further claims this is the first explicit convergence-rate guarantee for model-free IOC and reports an order-of-magnitude speedup over a differentiable-programming benchmark in reconstruction accuracy and robustness.","pith_inferences":["The paper leaves implicit that the real identifiability bottleneck is the identification of the feedback gain $K^*$ from the trajectory; if the data fail the persistent-excitation condition, no convex reformulation can recover a meaningful cost because $K^*$ itself is not unique.","A natural testable extension would be to apply the same variable-change and dual block-coordinate machinery to discrete-time LQR or output-feedback problems, where the cones and block updates change but the convexification strategy may carry over.","Because convergence is measured on the dual objective gap rather than on parameter error, practitioners should validate recovered models by trajectory reproduction error, as the simulations do, rather than by closeness of $(A,B,Q,R)$ to an unknown ground truth.","A downstream safety implication the paper does not develop: since positive scaling of $(Q,R)$ does not change the optimal policy, the magnitude of the recovered cost is not identifiable, so any guarantee stated in terms of $Q$ or $R$ must be invariant to that scaling."],"forward_implications":["If the reformulation is valid, model-free IOC no longer needs non-convex optimization: any data-compatible $(A,B,Q,R)$ can be found by convex feasibility, so local minima are not an issue.","The $\\mathcal{O}(1/k)$ bound gives an explicit iteration budget, so a safety-critical deployment can precompute how many iterations are needed for a required tolerance on the dual objective gap.","Closed-form block updates make each iteration cheap and step-size free; in the paper's experiment the method converges in 19 iterations and 0.3 seconds, versus 2000 iterations for the differentiable-programming baseline.","Because all zero-objective solutions form an equivalence class, the algorithm's guarantee is convergence to an equivalent LQR explanation of the expert's behavior, not to a unique ground-truth parameter set."],"supporting_citations":[{"why":"Establishes the classic connection between optimal linear control laws and the structure of the cost functional, which the paper's reformulation builds on.","marker":"[6]"},{"why":"Supplies the LMI/optimality characterization used to express compatibility between the identified gain and the cost weights as convex constraints.","marker":"[7]"},{"why":"Provides the differentiable-programming baseline against which the simulations compare convergence speed, accuracy, and robustness.","marker":"[22]"},{"why":"Supplies the block-coordinate-descent convergence theorem whose assumptions the paper verifies to obtain the O(1/k) rate.","marker":"[26]"},{"why":"Documents non-uniqueness of inverse optimal control solutions, motivating the paper's equivalence-class notion of convergence.","marker":"[28]"},{"why":"Supports the persistent-excitation argument that lets the feedback gain be identified from the trajectory by least squares.","marker":"[29]"}],"fun_headline_variants":["IOC: non-convex joint estimation becomes convex with O(1/k)","First convex reformulation for model-free IOC with rate guarantee","Inverse LQR without dynamics: convex and provably fast","Model-free IOC gets convex form and O(1/k) convergence proof","Provably convergent convex IOC: joint A,B,Q,R recovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The $\\mathcal{O}(1/k)$ proof assumes the set of dual variables no worse than the starting point stays within a bounded region; the paper does not prove this, and with unbounded semidefinite cones the rate constant can become infinite, making the bound vacuous.","fun_headline_variants_meta":{"raw":{"variants":["IOC: non-convex joint estimation becomes convex with O(1/k)","First convex reformulation for model-free IOC with rate guarantee","Inverse LQR without dynamics: convex and provably fast","Model-free IOC gets convex form and O(1/k) convergence proof","Provably convergent convex IOC: joint A,B,Q,R recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000313,"raw_usage":{"total_tokens":1784,"prompt_tokens":956,"completion_tokens":828,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":738}},"tokens_in":572,"tokens_out":828,"duration_ms":7860,"temperature":1.0,"reasoning_tokens":738,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:53:32.870542+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the Hessian $H$ of the dual objective and look for a nonzero direction $d$ inside the semidefinite cone with $H d = 0$ and $W^\\top d = 0$; such a direction makes the objective constant along a ray, so the level-set radius $R$ is infinite, $\\sigma = 1/(9\\|H\\|R^2)$ vanishes, and the theorem's $\\mathcal{O}(1/k)$ inequality is vacuous. Run the proposed updates on an instance with a rank-deficient $Q$ on the data-supported subspace and check whether the dual gap obeys the claimed bound with the stated constants.","supporting_citations":[{"cited_title":"When is a linear control system optimal? Journal of Basic Engineering , 86(1):51–60, 1964","cited_arxiv_id":null,"evidence_quote":"Establishes the classic connection between optimal linear control laws and the structure of the cost functional, which the paper's reformulation builds on."},{"cited_title":"Linear matrix inequalities in system and control theory","cited_arxiv_id":null,"evidence_quote":"Supplies the LMI/optimality characterization used to express compatibility between the identified gain and the cost weights as convex constraints."},{"cited_title":"Pontryagin differentiable program- ming: An end-to-end learning and control framework","cited_arxiv_id":null,"evidence_quote":"Provides the differentiable-programming baseline against which the simulations compare convergence speed, accuracy, and robustness."},{"cited_title":"Iteration complexity analysis of block coordinate descent methods","cited_arxiv_id":null,"evidence_quote":"Supplies the block-coordinate-descent convergence theorem whose assumptions the paper verifies to obtain the O(1/k) rate."},{"cited_title":"Nonuniqueness and convergence to equivalent solutions in observer-based inverse reinforcement learning.Automatica, 171:111977, 2025","cited_arxiv_id":null,"evidence_quote":"Documents non-uniqueness of inverse optimal control solutions, motivating the paper's equivalence-class notion of convergence."},{"cited_title":"System identification—a survey.Automatica, 7(2):123–162, 1971","cited_arxiv_id":null,"evidence_quote":"Supports the persistent-excitation argument that lets the feedback gain be identified from the trajectory by least squares."}],"review_version":1}