{"id":"b0a4e0ab-c9b5-44e7-b162-75ba03c82760","arxiv_id":"2601.04120","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Optimal control of obstacle problems is solved mesh-free by training constraint-embedding neural networks with S2-FOBA, a single-loop stochastic bilevel algorithm whose Moreau-penalized gradient converges sublinearly.","lead":"The paper introduces S2-FOBA, a single-loop stochastic algorithm that trains neural networks to solve optimal control problems constrained by obstacle equations without repeatedly solving expensive discretized subproblems. A smart generalist should read it because it promises a mesh-free route to nonsmooth bilevel control problems that classical finite-element methods handle only in low dimensions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Moreau reformulation (3.6) only encodes lower-level stationarity; with the nonconvex NN lower level, S2-FOBA can converge to a spurious stationary point that is not a lower-level minimizer, so Theorem 4.1 does not by itself support the claim that Algorithm 3.2 solves (2.1).","rationale":"The reader's weakest assumption identifies the same load-bearing issue: the Moreau-envelope reformulation equates the bilevel problem with a stationarity-based relaxation, and the NN parameterization of the lower level is explicitly nonconvex, so stationarity need not imply lower-level minimality. This is not an external disagreement with the authors; it is an internal limitation that the paper itself states in §3.2.1 and §3.3. The paper's convergence theorem is plausible as a statement about the penalized Moreau objective, and the numerical results and public code are genuine supporting evidence, but the theorem does not bridge from stationarity of φ_{c_k} to feasibility for the original bilevel problem (2.1). A minimal scalar example makes the failure mode concrete: the Moreau constraint is satisfied at a local maximum of the lower-level objective, and the penalized objective has a local minimum there, so S2-FOBA can converge to an infeasible point while Theorem 4.1 holds exactly as stated. Because the reader already judged the paper CONDITIONAL on this gap, my stress test does not move the verdict. The paper would need either to weaken its claims to convergence for the stationarity relaxation or to prove additional conditions under which the NN lower-level stationary set coincides with its minimizer set.","tokens_in":28020,"tokens_out":15184,"duration_ms":151358,"concrete_test":"Run the minimal instance with θu absent: lower level e(z)=(z²−1)², upper j(θ)=θ², γ=0.1<1/8, c_k=5k^{0.3}, η=α=10^{-3}, initialized at θ_y=0, z=0. Track θ_y; if it remains at 0 while the globally feasible lower-level minimizers are ±1, the penalized-Moreau limit is infeasible for (3.5), and Stage 2 would move the state to ±1, changing J from 0 to 1. This confirms the equivalence failure is not merely formal. An optional companion check on Example 1: after Stage 1, compute e(θ̂y,θ̃u)−min_z e(z,θ̃u) by multistart Adam and report the increase in J after Stage 2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is not the step-size analysis but the semantic content of the stationarity point. In §3.2.1 the authors state that (3.6) equals (3.5) only when the lower-level stationary set equals its minimizer set, “which holds, for instance, when e is convex in θy”; §3.1 states the implemented lower level is “in general, nonconvex”. For nonconvex e, equality in e(θy,θu) ≤ eγ(θy,θu) is equivalent to 0 ∈ ∂_{θy} e(θy,θu), i.e. stationarity, not global minimality. Theorem 4.1 only bounds E‖∇φ_{c_k}‖ and does not control the feasibility gap e−eγ, so a limit point can be a spurious stationary point (even a local maximum) of the lower-level objective. Stage 2 (§3.3) re-solves the lower level only for the final control; it cannot repair a control that Stage 1 chose against a non-optimal state. Concrete counterexample: e(z)=(z²−1)², j(θ)=θ², γ<1/8. At θ=0, ∇e=0, e(0)=eγ(0), and φ_c has a local minimum for every c, while the true bilevel solution is θ=±1. Thus the theorem's stationarity guarantee does not imply that Algorithm 3.2 solves (2.1).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a mesh-free bilevel deep learning method for optimal control of obstacle problems. The state and control are parameterized by constraint-embedding neural networks, and the resulting stochastic bilevel problem (3.5) is solved by a single-loop stochastic first-order algorithm (S2-FOBA) based on the Moreau-envelope reformulation (3.6) and a penalty strategy (3.10). The main theoretical result, Theorem 4.1, states that S2-FOBA drives the gradient of the penalized objective φ_{c_k} to zero at rate O(1/K^{(1-p)/2}) under assumptions A1–A3. Numerical experiments on benchmark obstacle-control problems, including control-constrained, irregular-domain, obstacle-control, and EVI examples, report relative L2 errors in the range 4e-3 to 3e-2 and favorable cost comparisons against an active-set method. A two-stage refinement (Algorithm 3.2) is added to improve lower-level feasibility of the final state.","tokens_in":28429,"tokens_out":3604,"duration_ms":36235,"significance":"If the convergence claim is interpreted as convergence to stationarity of the Moreau-penalized surrogate, the paper offers a useful single-loop, Hessian-free, stochastic algorithm for a hard nonsmooth bilevel problem class. The method is mesh-free, handles complex domains, and is supported by reproducible PyTorch code and comparisons with classical active-set and MPEC solvers. The authors also correctly identify and demonstrate the failure of naive objective-combination approaches. However, the paper's stronger claim that S2-FOBA solves the original obstacle-control problem (2.1) is not justified by the analysis as written: Theorem 4.1 controls only the gradient of φ_{c_k}, and the connection between φ_{c_k} and the original bilevel problem relies on an equivalence that the paper itself states requires lower-level convexity, which is not satisfied by the implemented nonconvex neural-network parametrization.","major_comments":[{"comment":"The load-bearing equivalence between (3.6) and (3.5) is stated by the authors to hold only when the stationary set of e(·, θu) coincides with its minimizer set, 'which holds, for instance, when e is convex in θy'. But §3.1 states that the lower-level problem is 'in general, nonconvex'. For nonconvex e, equality in the Moreau constraint e ≤ e_γ is equivalent to lower-level stationarity, not global minimality. Theorem 4.1 bounds E‖∇φ_{c_k}‖ and does not control the feasibility gap e − e_γ. Thus the paper does not establish that S2-FOBA produces a feasible lower-level minimizer for the original problem, even in the deterministic zero-variance limit. This gap should be addressed either by restricting the theoretical result to the stationarity-based surrogate, or by adding a mechanism that guarantees convergence to a true lower-level minimizer.","section":"§3.2.1 and Theorem 4.1"},{"comment":"Theorem 4.1 assumes ρ-weak convexity and L_e-smoothness of the NN lower-level objective e(θy,θu) in (3.5). For the implemented ResNet with Swish activations and the constraint-embedding maps (3.1)–(3.3), these properties are not verified. In particular, global weak convexity in θy is not generally inherited by deep residual networks. The paper should either prove or state sufficient architecture conditions under which A1–A2 hold, or explicitly frame the convergence analysis as conditional on these assumptions and discuss their plausibility for the tested networks.","section":"Assumptions A1–A2 and §3.1"},{"comment":"Stage 2 of Algorithm 3.2 re-solves the lower-level problem only for the final control parameter θ̃u, using θ̂y as initialization. This can only repair the state; it cannot repair a control that Stage 1 selected while the state parameter was at a spurious stationary point of the nonconvex lower-level objective. Consequently, the two-stage procedure does not restore feasibility of the coupled bilevel pair. The numerical examples may still be meaningful, but the paper should clarify that the method is not guaranteed to enforce bilevel feasibility, and should report any empirical check of lower-level optimality (e.g., comparison of e(θ̂y,θ̃u) with the value obtained by a full lower-level solve).","section":"§3.3, Stage 2"},{"comment":"The numerical results are reported for a fixed network architecture and fixed hyperparameters (γ, c_k, learning rates), and the convergence theorem's step-size conditions are not checked in the experiments. Since the theory is a central advertised contribution, the paper should at least discuss how the selected step-size and penalty schedules relate to the assumptions of Theorem 4.1, or state that the experiments are heuristic. This does not invalidate the numerical evidence, but it weakens the connection between theory and practice.","section":"§5, Tables 5.1–5.2"}],"minor_comments":[{"comment":"Typo: 'Comparions' should be 'Comparisons'.","section":"§5, Figure 5.4 caption"},{"comment":"Missing space in 'lossein'; also 'we set η=0.01' redefines the symbol η already used for step sizes, which can be confusing.","section":"§6.3, text near Eq. (6.5)"},{"comment":"The notation 'Gˆ ateaux' in the introduction has a corrupted accent; should be 'Gâteaux'.","section":"§2.1"},{"comment":"The symbols θ := (θy, θu) are introduced after Assumption A1, but A1 already uses θy,θu jointly; the ordering of definitions could be improved.","section":"§4.1, Assumption A1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable contribution to mesh-free bilevel training for obstacle control, with reproducible code and solid numerical evidence. The main concern is not the step-size analysis but the semantic gap between convergence to stationarity of the Moreau-penalized surrogate and solving the original bilevel problem under a nonconvex lower level. This is a central claim, so I recommend major revision rather than acceptance in the current form. If the authors can either reformulate the theoretical claim to match what is actually proved, or add verifiable conditions under which stationarity implies lower-level minimality, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take: the algorithm is coherent and the 2D numerics look decent, but the main theory is narrower than the abstract claims. Theorem 4.1 shows the iterates converge to a stationary point of the penalized Moreau objective φ_c. It does not show that this point is a lower-level minimizer for the control, which is what the original bilevel problem requires. The paper itself notes in §3.2.1 that the Moreau reformulation (3.6) is equivalent to the bilevel problem only when the lower-level stationary set equals the minimizer set, and it immediately says the implemented lower level is nonconvex. The stress-test counterexample is not artificial: with e(z)=(z²−1)² and a simple upper objective, θ=0 is a stationary point of e but not a minimizer, and φ_c has a local minimum there for every c. Stage 2 can fix the final state, but it cannot repair a control that was chosen against a non-optimal state. So the convergence result does not by itself support the \"solves (2.1)\" wording.\n\nWhat is genuinely here: S2-FOBA is a real single-loop stochastic method that avoids nested solves and does not need lower-level uniqueness. It adapts the authors' earlier Moreau-envelope machinery to a problem class that has not had this treatment. The experiments are honestly benchmarked against active-set and MPEC solvers, the code is public, and the paper shows clearly why the objective-combination baseline fails. The proof of the descent lemma is plausible under A1–A2. Those assumptions — joint weak convexity and smoothness of the network losses — are not verified for the ResNet/Swish architecture used, which is a real gap, though likely fixable with bounded-parameter arguments.\n\nSmaller issues: the high-dimensional claim is unsupported; every example is 2D. There are no repeated-seed statistics or error bars. And the rate in Theorem 4.1 is actually consistent once you read it as a bound on the norm rather than the squared norm — the worry about an inconsistency there seems like a misreading.\n\nBottom line: the paper contributes a workable training scheme and an interesting reformulation for mesh-free bilevel obstacle control, but the theoretical bridge to the original problem is incomplete. It deserves a serious referee who can press the authors on the equivalence gap and on numerical evidence in higher dimensions or with multiple seeds. I would send it to review, but I would not cite it in its current form.","headline":"A useful single-loop algorithm for bilevel obstacle control, but the convergence theorem only reaches a penalized stationarity point, not the true bilevel solution, and the paper leaves that gap open.","tokens_in":28920,"tokens_out":2999,"would_cite":false,"duration_ms":29932,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","49M41","65K15","93-08"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that optimal control of obstacle problems can be solved mesh-free by constraint-embedding neural networks, trained with a single-loop stochastic algorithm whose Moreau-penalized objective converges at O(K^{-(1-p)/2}).","keywords":["optimal control of obstacle problems","bilevel optimization","deep learning","constraint-embedding neural networks","single-loop stochastic algorithm","Moreau envelope","convergence analysis","elliptic variational inequalities"],"falsifier":"Construct a small lower-level network whose energy has two well-separated local minima for a fixed control parameter (e.g., a one-dimensional parameter with two energy wells), run S2-FOBA with the stated step sizes, and test whether the returned state parameter satisfies the lower-level optimality condition min_z e(z, θ_u) approximately. If the gradient of the penalized objective is small while e(θ_y, θ_u) stays above the true minimum, the convergence measure has decoupled from feasibility.","tokens_in":1643,"feed_emoji":"🎯","tokens_out":2793,"duration_ms":73499,"temperature":0.7,"pith_summary":"The paper tries to establish that a class of nonsmooth, bilevel optimal control problems—where the state is the solution of an obstacle problem—can be solved with deep neural networks without repeatedly solving discretized subproblems. Its central proposal is S2-FOBA, a single-loop stochastic first-order algorithm that trains state and control networks while explicitly preserving the bilevel structure through a Moreau-envelope reformulation. If correct, the method would give an accurate, constraint-satisfying state/control pair at a fraction of the cost of classical mesh-based active-set methods, and would extend naturally to complex geometries and to elliptic variational inequalities. The convergence theorem states that the gradient of the penalized Moreau objective vanishes at a sublinear rate under mild weak-convexity and smoothness assumptions, and the numerical experiments report relative L2 errors around 1e-2 to 1e-3.","feed_headline":"One training loop solves optimal control of obstacle problems","feed_subtitle":"Constraint-embedding networks plus a Moreau penalty reach stationary points without nested solves or lower-level uniqueness.","key_machinery":"The central mechanism is the Moreau-envelope-based reformulation: the bilevel constraint θ_y ∈ arg min e(θ_y, θ_u) is replaced by e(θ_y, θ_u) ≤ e_γ(θ_y, θ_u), where e_γ is the Moreau envelope e_γ(θ_y, θ_u) = min_z {e(z, θ_u) + (1/2γ)||z − θ_y||^2}. This converts the constraint into a smooth penalty ψ_c = j + c(e − e_γ), avoiding the nonsmooth value function. The constraint-embedding network architectures (3.1)–(3.4) bake the obstacle and control constraints directly into the network output, so no penalty or active-set detection is needed for feasibility. The auxiliary sequence z_k approximates the proximal center z*_γ and is updated inside the same loop, which eliminates the inner minimizati","core_discovery":"The paper's central claim is that the bilevel optimal control problem (2.1) can be approximated by a stochastic bilevel problem over neural-network parameters (3.5), and that S2-FOBA solves that problem without nested optimization. The Moreau-envelope reformulation replaces the lower-level constraint with e(θ_y, θ_u) ≤ e_γ(θ_y, θ_u), where e_γ is a continuously differentiable Moreau envelope whose proximal center is tracked by an auxiliary variable z_k. The algorithm minimizes the penalized objective φ_c(θ) = j(θ)/c + (e(θ) − e_γ(θ)) by alternating stochastic gradient steps for z, θ_y, and θ_u in one loop. Theorem 4.1 states that under ρ-weak convexity and Lipschitz smoothness, with step siz","pith_inferences":["Inference: Theorem 4.1 bounds the gradient of the penalized Moreau objective, not the gap to the original bilevel optimum. If the lower-level network has spurious stationary points, the stated rate could hold while the control remains suboptimal; the practical value then depends on how often Stage 2's refinement is needed.","Inference: Stage 2 repairs only the state, never the control. If Stage 1 returns a poor control, the refinement cannot fix it, so end-to-end accuracy relies on Stage 1 having already found a good control.","Inference: The mesh-free nature suggests a natural extension to three-dimensional or higher-dimensional domains, where classical active-set methods would become expensive and where the paper's complex-domain experiments hint that the method's error would remain roughly mesh-independent.","Inference: The manual choices of step sizes and penalty schedules (e.g., c_k = 5k^0.3) leave room for automated tuning; testing a wider range of growth exponents would separate the method's robustness from its sensitivity to these hyperparameters."],"forward_implications":["With trained networks, evaluating the solution at a new mesh is only a forward pass, whereas classical active-set methods require re-meshing and re-solving; the paper's Table 5.2 shows order-of-magnitude time savings at moderate resolutions.","The bilevel training avoids the failure mode of weighted single-level objectives such as (2.4), which cannot produce a state that minimizes the lower-level energy for the computed control.","The two-stage refinement restores lower-level feasibility of the state for the Stage-1 control, yielding relative L2 errors around 1e-2 to 1e-3 in the benchmarks.","Because S2-FOBA needs only first-order stochastic gradients and does not assume lower-level uniqueness, it implements as a single loop and scales to larger networks and domains than nested-solve approaches.","The same algorithm extends to obstacle control and to optimal control of elliptic variational inequalities by replacing the lower-level loss with a projection fixed-point residual."],"fun_headline_variants":["Single-loop deep learning for obstacle control","Mesh-free optimal control of obstacles in one loop","No nested solves: deep bilevel for obstacle problems","S2-FOBA: one-loop bilevel training for obstacle control","Moreau-envelope trick speeds up obstacle control learning"],"cache_read_input_tokens":30080,"weakest_assumption_plain":"The argument depends on the assumption that a point where the lower-level network's objective has zero gradient is actually a global minimizer of that objective; the paper establishes this when the lower level is convex in the state parameters, but the implemented neural network is explicitly nonconvex.","fun_headline_variants_meta":{"raw":{"variants":["Single-loop deep learning for obstacle control","Mesh-free optimal control of obstacles in one loop","No nested solves: deep bilevel for obstacle problems","S2-FOBA: one-loop bilevel training for obstacle control","Moreau-envelope trick speeds up obstacle control learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000147,"raw_usage":{"total_tokens":1021,"prompt_tokens":743,"completion_tokens":278,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":202}},"tokens_in":487,"tokens_out":278,"duration_ms":3298,"temperature":1.0,"reasoning_tokens":202,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T12:06:19.503169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a small lower-level network whose energy has two well-separated local minima for a fixed control parameter (e.g., a one-dimensional parameter with two energy wells), run S2-FOBA with the stated step sizes, and test whether the returned state parameter satisfies the lower-level optimality condition min_z e(z, θ_u) approximately. If the gradient of the penalized objective is small while e(θ_y, θ_u) stays above the true minimum, the convergence measure has decoupled from feasibility.","supporting_citations":[],"review_version":1}