Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

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}).

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 12:06 UTC pith:B54VWAM6

load-bearing objection 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. the 4 major comments →

arxiv 2601.04120 v2 pith:B54VWAM6 submitted 2026-01-07 math.OC cs.LG

A Single-Loop Bilevel Deep Learning Method for Optimal Control of Obstacle Problems

Yongcun Song , Shangzhi Zeng , Jin Zhang , Lvgang Zhang This is my paper
classification math.OC cs.LG MSC 68T0749M4165K1593-08
keywords optimal control of obstacle problemsbilevel optimizationdeep learningconstraint-embedding neural networkssingle-loop stochastic algorithmMoreau envelopeconvergence analysiselliptic variational inequalities
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

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.

Core claim

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

What carries the argument

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

Load-bearing premise

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.

What would settle it

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.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • 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.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

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.

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 (4)
  1. [§3.2.1 and Theorem 4.1] 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.
  2. [Assumptions A1–A2 and §3.1] 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.
  3. [§3.3, Stage 2] 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).
  4. [§5, Tables 5.1–5.2] 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.
minor comments (4)
  1. [§5, Figure 5.4 caption] Typo: 'Comparions' should be 'Comparisons'.
  2. [§6.3, text near Eq. (6.5)] Missing space in 'lossein'; also 'we set η=0.01' redefines the symbol η already used for step sizes, which can be confusing.
  3. [§2.1] The notation 'Gˆ ateaux' in the introduction has a corrupted accent; should be 'Gâteaux'.
  4. [§4.1, Assumption A1] The symbols θ := (θy, θu) are introduced after Assumption A1, but A1 already uses θy,θu jointly; the ordering of definitions could be improved.

Circularity Check

0 steps flagged

No circular reduction: Theorem 4.1 is a genuine stationarity-convergence result, and the numerical claims are benchmarked against independent solvers; the self-cited Moreau-envelope lemmas are independent mathematical support, not fitted inputs.

full rationale

The derivation chain is not circular. The central theorem, Theorem 4.1, is proved in Section 4 via explicit descent inequalities for the merit function φ_c; it concludes convergence to a stationary point of the penalized Moreau objective, and that conclusion is not obtained by fitting a parameter to the target outputs. The numerical results are validated against independent classical solvers and analytical solutions (active-set method [26], MPEC reference [28]), so the reported errors are not constructed from the training data. The Moreau-envelope machinery is imported from the authors' own prior work, [21] and [39], and this is the main self-citation burden; however, the cited results are parameter-free lemmas/theorems with stated assumptions (ρ-weak convexity, γ < 1/ρ, etc.) and do not already contain the paper's target result, so they qualify as independent support under the review rules. A real limitation, acknowledged by the paper itself, is that the reformulation (3.6) is equivalent to (3.5) only when the lower-level stationary set coincides with the minimizer set: Section 3.2.1 states this equivalence "holds, for instance, when e is convex with respect to θy", while Section 3.1 notes "the lower-level problem is, in general, nonconvex". This creates a semantic gap between convergence to a stationary point of φ_c and actually solving the bilevel problem (2.1), but it is a correctness/feasibility limitation, not a circularity, because Theorem 4.1 does not claim equivalence to (2.1) and the convergence result is not defined in terms of the original problem's solution.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper contributes an algorithm, not a new physical law; its free parameters are hand-tuned training hyperparameters rather than fitted constants. The real burden is conceptual: the proof assumes the NN lower-level objective is weakly convex/smooth and that stationarity of the lower level is equivalent to minimality, neither of which is demonstrated for the implemented nonconvex networks. No new physical entities are introduced.

free parameters (5)
  • proximal parameter γ = 20 (Ex.1), 500 (Ex.2), 50 (Ex.3, Ex.4), 200000 (Ex.5)
    Hand-tuned per example in §5-§6; controls the Moreau proximal term (3.7). The numerical results depend on it; there is no automatic selection.
  • penalty parameter sequence c_k = 5k^0.3 (Ex.1,2,5), 5k^0.2 (Ex.3), (1/5)k^0.2 (Ex.4)
    Hand-set per example; Theorem 4.1 only requires c_k nondecreasing, but the practical growth rate is tuned to each benchmark.
  • learning rates α, β, η = 1e-3 with 0.8 decay per 1000 epochs (Ex.1-4); 2e-4 (Ex.5)
    Hand-tuned; the theorem requires α_0,β_0 sufficiently small relative to η_0, and the implemented values are chosen accordingly.
  • mini-batch size m = 512 for most examples
    Hand-selected; determines stochastic gradient variance and epoch cost in Algorithm 3.1.
  • NN architecture (depth/width/activation) = ResNet, 3 residual blocks, 16 neurons per layer, Swish
    Chosen by hand in §5; no ablation or capacity analysis links this choice to the convergence assumptions A1-A2.
axioms (5)
  • ad hoc to paper The NN lower-level objective e in (3.5) is ρ-weakly convex in (θy, θu) and L_e-smooth (A1-A2).
    Assumption (A1)-(A2) in §4.1. Standard Swish/ResNet parameterizations do not globally satisfy joint weak convexity/smoothness, so the convergence theorem may not apply to the implemented networks.
  • ad hoc to paper Stationary points of the lower-level NN problem coincide with its minimizers, making (3.6) equivalent to (3.5).
    Stated in §3.2.1 as a condition for equivalence; the lower-level problem is explicitly nonconvex, and no proof of this coincidence is given for the NN parameterization.
  • domain assumption Deep Ritz / universal approximation: minimizing the NN energy produces a state solving the obstacle problem, so the NN bilevel problem (3.5) approximates the continuous problem (2.1).
    Invoked via [17] in §3.1; no quantitative bound is given between the solution of (3.5) (or its stationary relaxation) and the solution of (2.1).
  • domain assumption Stochastic gradient oracles are unbiased with bounded variance (A3).
    Standard in stochastic optimization and satisfied by i.i.d. uniform sampling from the bounded domain; used throughout the proof of Theorem 4.1.
  • standard math Solution existence for (2.1) and standard Moreau-envelope properties from [39].
    Existence cited to [43,45] (§2.1); Moreau-envelope smoothness/Lipschitz lemmas taken from [39] and used as Lemma 4.2.

pith-pipeline@v1.3.0-alltime-deepseek · 27527 in / 19863 out tokens · 187877 ms · 2026-08-03T12:06:19.503169+00:00 · methodology

0 comments
read the original abstract

Optimal control of obstacle problems arises in a wide range of applications and is computationally challenging due to its nonsmoothness, nonlinearity, and bilevel structure. Classical numerical approaches rely on mesh-based discretization and typically require solving a sequence of costly subproblems. In this work, we propose a single-loop bilevel deep learning method, which is mesh-free, scalable to high-dimensional and complex domains, and avoids repeated solution of discretized subproblems. The method employs constraint-embedding neural networks to approximate the state and control and preserves the bilevel structure. To train the neural networks efficiently, we propose a Single-Loop Stochastic First-Order Bilevel Algorithm (S2-FOBA), which eliminates nested optimization and does not rely on restrictive lower-level uniqueness assumptions. We analyze the convergence behavior of S2-FOBA under mild assumptions. Numerical experiments on benchmark examples, including distributed and obstacle control problems with regular and irregular obstacles on complex domains, demonstrate that the proposed method achieves satisfactory accuracy while reducing computational cost compared to classical numerical methods.

Figures

Figures reproduced from arXiv: 2601.04120 by Jin Zhang, Lvgang Zhang, Shangzhi Zeng, Yongcun Song.

Figure 5.1
Figure 5.1. Figure 5.1: Training trajectories of Algorithm 3.2 for Example 1 Furthermore, the control constraint u ∈ Uad often presents challenges and typically requires additional techniques, such as active set detection. Within Algorithm 3.2, this issue can be effectively addressed by embedding u ∈ Uad via (3.3). To demonstrate the effect of this modification, we set Uad = {u ∈ L 2 (Ω) | 0 ≤ u(x) ≤ 0.7 a.e. in Ω}. The corresp… view at source ↗
Figure 5.2
Figure 5.2. Figure 5.2: Numerical results of Algorithm 3.2 for Example 1 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0 (a) Computed state ˆy 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.1 0.2 0.3 0.4 0.5 0.6 (b) Computed control ˆu with constraint [PITH_FULL_IMAGE:figures/full_fig_p016_5_2.png] view at source ↗
Figure 5.3
Figure 5.3. Figure 5.3: Numerical results of Algorithm 3.2 for Example 1 with control constraints 16 [PITH_FULL_IMAGE:figures/full_fig_p016_5_3.png] view at source ↗
Figure 5.4
Figure 5.4. Figure 5.4: Numerical results of Algorithm 2.1 for Example 1 [PITH_FULL_IMAGE:figures/full_fig_p017_5_4.png] view at source ↗
Figure 5.5
Figure 5.5. Figure 5.5: Numerical results of Algorithm 3.2 for Example 2 contrast, Algorithm 3.2 is mesh-free, which eliminates the need for costly mesh generation and allows for seamless handling of irregular geometries. Example 3. Let ρ1(ζ) = 2.25 + 0.21 sin(4ζ) + 0.18 cos(6ζ) + 0.135 cos(5ζ) and Ω =  (x1, x2) ∈ R 2 | x1 = r cos(ζ), x2 = r sin(ζ), 0 ≤ r < ρ1(ζ), 0 ≤ ζ < 2π [PITH_FULL_IMAGE:figures/full_fig_p018_5_5.png] view at source ↗
Figure 6.1
Figure 6.1. Figure 6.1: Numerical Results of Algorithm 3.2 for Example 3 where no control constraints are imposed on ψ. (6.1) min y∈H1 0 (Ω),ψ∈H1 0 (Ω) J(y, ψ) = 1 2 Z Ω |y − yd| 2dx + σ 2 Z Ω |∇ψ| 2dx, s.t. y = arg min y′∈Yad Z Ω  1 2 |∇y ′ | 2 − fy′  dx. We first approximate ψ by the NN (6.2) ψˆ(x; θψ) := m(x)N (x, θψ), where m ∈ C∞(Ω) is chosen such that m(x) = 0 ∀x ∈ ∂Ω and m(x) > 0 ∀x ∈ Ω. The state y is approximated by … view at source ↗
Figure 6.2
Figure 6.2. Figure 6.2: displays the computed state y, control ψ, and their corresponding reference solutions and pointwise errors. The final relative L 2 -errors for the state and the control, computed on a uniform grid with a mesh size of h = 1/200, are 4.25 × 10−2 and 2.99 × 10−2 , respectively. 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 2.0 1.5 1.0 0.5 0.0 0.5 2.0 1.5 1.0 0.5 0.0 0.5 (a) Reference state y ∗ 0.0 0.2 0.4… view at source ↗
Figure 6.3
Figure 6.3. Figure 6.3: Numerical results of Algorithm 3.2 for Example 5 REFERENCES [1] A. Alphonse, M. Hintermuller, A. Kister, C. H. Lun, and C. Sirotenko ¨ , A neural network approach to learning solutions of a class of elliptic variational inequalities, arXiv preprint arXiv:2411.18565, (2024). [2] V. Barbu, Optimal control of variational inequalities, Research Notes in Math., 100 (1984). [3] J. Barry-Straume, A. Sarshar, A.… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Constrained Neural Parameterization for Optimization in Function Spaces

    math.OC 2026-05 unverdicted novelty 6.0

    Neural parameterizations are constructed whose range lies in admissible sets for function-space optimization, converting constrained problems to unconstrained gradient-based ones.

Reference graph

Works this paper leans on

59 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Alphonse, M

    A. Alphonse, M. Hinterm \"u ller, A. Kister, C. H. Lun, and C. Sirotenko , A neural network approach to learning solutions of a class of elliptic variational inequalities , arXiv preprint arXiv:2411.18565, (2024)

    Pith/arXiv arXiv 2024

  2. [2]

    Barbu , Optimal control of variational inequalities , Research Notes in Math., 100 (1984)

    V. Barbu , Optimal control of variational inequalities , Research Notes in Math., 100 (1984)

    1984

  3. [3]

    Barry-Straume, A

    J. Barry-Straume, A. Sarshar, A. A. Popov, and A. Sandu , Physics-informed neural networks for PDE-constrained optimization and control , Communications on Applied Mathematics and Computation, (2025), pp. 1--24

    2025

  4. [4]

    Bergounioux , Use of augmented Lagrangian methods for the optimal control of obstacle problems , Journal of Optimization Theory and Applications, 95 (1997), pp

    M. Bergounioux , Use of augmented Lagrangian methods for the optimal control of obstacle problems , Journal of Optimization Theory and Applications, 95 (1997), pp. 101--126

    1997

  5. [5]

    Bergounioux and S

    M. Bergounioux and S. Lenhart , Optimal control of bilateral obstacle problems , SIAM Journal on Control and Optimization, 43 (2004), pp. 240--255

    2004

  6. [6]

    Bourgat and G

    J. Bourgat and G. Duvaut , Numerical analysis of flow with or without wake past a symmetric two-dimensional profile without incidence , International Journal for Numerical Methods in Engineering, 11 (1977), pp. 975--993

    1977

  7. [7]

    u ller, and C. L \

    C. Brett, C. M. Elliott, M. Hinterm \"u ller, and C. L \"o bhard , Mesh adaptivity in optimal control of elliptic variational inequalities with point-tracking of the state , Interfaces and Free Boundaries, 17 (2015), pp. 21--53

    2015

  8. [8]

    Brezis and G

    H. Brezis and G. Stampacchia , The hodograph method in fluid-dynamics in the light of variational inequalities , Archive for Rational Mechanics and Analysis, 61 (1976), pp. 1--18

    1976

  9. [9]

    Y. Cao, C. C. So, Y. Dai, S. P. Yung, and J.-M. Wang , Adversarial physics-informed neural networks with hard constraints for optimal control of PDEs , Journal of Computational Physics, 541 (2025), p. 114307

    2025

  10. [10]

    T. Chen, Y. Sun, and W. Yin , Closing the gap: Tighter analysis of alternating stochastic gradient methods for bilevel problems , Advances in Neural Information Processing Systems, 34 (2021), pp. 25294--25307

    2021

  11. [11]

    Cheng, X

    X. Cheng, X. Shen, X. Wang, and K. Liang , A deep neural network-based method for solving obstacle problems , Nonlinear Analysis: Real World Applications, 72 (2023), p. 103864

    2023

  12. [12]

    P. L. Combettes and J.-C. Pesquet , Deep neural network structures solving variational inequalities , Set-Valued and Variational Analysis, 28 (2020), pp. 491--518

    2020

  13. [13]

    C. W. Cryer , The method of Christopherson for solving free boundary problems for infinite journal bearings by means of finite differences , Mathematics of Computation, 25 (1971), pp. 435--443

    1971

  14. [14]

    Cybenko , Approximation by superpositions of a sigmoidal function , Mathematics of Control, Signals and Systems, 2 (1989), pp

    G. Cybenko , Approximation by superpositions of a sigmoidal function , Mathematics of Control, Signals and Systems, 2 (1989), pp. 303--314

    1989

  15. [15]

    Darehmiraki , A deep learning approach for the obstacle problem , in Proceedings of Academia-Industry Consortium for Data Science: AICDS 2020, Springer, 2022, pp

    M. Darehmiraki , A deep learning approach for the obstacle problem , in Proceedings of Academia-Industry Consortium for Data Science: AICDS 2020, Springer, 2022, pp. 179--188

    2020

  16. [16]

    De Los Reyes , On the optimal control of some nonsmooth distributed parameter systems arising in mechanics , GAMM-Mitteilungen, 40 (2018), pp

    J.-C. De Los Reyes , On the optimal control of some nonsmooth distributed parameter systems arising in mechanics , GAMM-Mitteilungen, 40 (2018), pp. 268--286

    2018

  17. [17]

    W. E and B. Yu , The deep Ritz method: A deep learning-based numerical algorithm for solving variational problems , Communications in Mathematics and Statistics, 6 (2018), pp. 1--12

    2018

  18. [18]

    El Bahja, J

    H. El Bahja, J. C. Hauffen, P. Jung, B. Bah, and I. Karambal , A physics-informed neural network framework for modeling obstacle-related equations , Nonlinear Dynamics, 113 (2025), pp. 1--12

    2025

  19. [19]

    Friedman , Variational Principles and Free-Boundary Problems , Dover Books on Mathematics, Dover Publications, 2010

    A. Friedman , Variational Principles and Free-Boundary Problems , Dover Books on Mathematics, Dover Publications, 2010

    2010

  20. [20]

    u ller, R. H. Hoppe, and C. L \

    A. Gaevskaya, M. Hinterm \"u ller, R. H. Hoppe, and C. L \"o bhard , Adaptive finite elements for optimally controlled elliptic variational inequalities of obstacle type , in Optimization with PDE Constraints: ESF networking program'OPTPDE', Springer, 2014, pp. 95--150

    2014

  21. [21]

    L. L. Gao, J. J. Ye, H. Yin, S. Zeng, and J. Zhang , Moreau envelope based difference-of-weakly-convex reformulation and algorithm for bilevel programs , arXiv preprint arXiv:2306.16761, (2023)

    Pith/arXiv arXiv 2023

  22. [22]

    Y. Gao, Y. Song, Z. Tan, H. Yue, and S. Zeng , Prox-PINNs : A deep learning algorithmic framework for elliptic variational inequalities , arXiv preprint arXiv:2505.14430, (2025)

    arXiv 2025

  23. [23]

    Glowinski , Lectures on Numerical Methods for Non-Linear Variational Problems , Springer Science & Business Media, 2008

    R. Glowinski , Lectures on Numerical Methods for Non-Linear Variational Problems , Springer Science & Business Media, 2008

    2008

  24. [24]

    Glowinski , Variational Methods for the Numerical Solution of Nonlinear Elliptic Problems , SIAM, 2015

    R. Glowinski , Variational Methods for the Numerical Solution of Nonlinear Elliptic Problems , SIAM, 2015

    2015

  25. [25]

    Z. Hao, C. Ying, H. Su, J. Zhu, J. Song, and Z. Cheng , Bi-level physics-informed neural networks for PDE constrained optimization using Broyden's hypergradients , The 11th International Conference on Learning Representations, (2023)

    2023

  26. [26]

    M. Hinterm \"u ller , An active-set equality constrained Newton solver with feasibility restoration for inverse coefficient problems in elliptic variational inequalities , Inverse Problems, 24 (2008), p. 034017

    2008

  27. [27]

    u ller, R. H. Hoppe, and C. L \

    M. Hinterm \"u ller, R. H. Hoppe, and C. L \"o bhard , Dual-weighted goal-oriented adaptive finite elements for optimal control of elliptic variational inequalities , ESAIM: Control, Optimisation and Calculus of Variations, 20 (2014), pp. 524--546

    2014

  28. [28]

    Hinterm \"u ller and I

    M. Hinterm \"u ller and I. Kopacka , Mathematical programs with complementarity constraints in function space: C-and strong stationarity and a path-following algorithm , SIAM Journal on Optimization, 20 (2009), pp. 868--902

    2009

  29. [29]

    Hinterm \"u ller and I

    M. Hinterm \"u ller and I. Kopacka , A smooth penalty approach and a nonlinear multigrid algorithm for elliptic MPECs , Computational Optimization and Applications, 50 (2011), pp. 111--145

    2011

  30. [30]

    u ller, C. L \

    M. Hinterm \"u ller, C. L \"o bhard, and M. Tber , An ^1 -penalty scheme for the optimal control of elliptic variational inequalities , in Numerical Analysis and Optimization: NAO-III, Muscat, Oman, January 2014, Springer, 2015, pp. 151--190

    2014

  31. [31]

    Hinterm \"u ller and T

    M. Hinterm \"u ller and T. Surowiec , A bundle-free implicit programming approach for a class of elliptic MPECs in function space , Mathematical Programming, 160 (2016), pp. 271--305

    2016

  32. [32]

    Hornik , Approximation capabilities of multilayer feedforward networks , Neural Networks, 4 (1991), pp

    K. Hornik , Approximation capabilities of multilayer feedforward networks , Neural Networks, 4 (1991), pp. 251--257

    1991

  33. [33]

    Hornik, M

    K. Hornik, M. Stinchcombe, and H. White , Multilayer feedforward networks are universal approximators , Neural Networks, 2 (1989), pp. 359--366

    1989

  34. [34]

    Ito and K

    K. Ito and K. Kunisch , Semi-smooth Newton methods for variational inequalitiesof the first kind , ESAIM: Mathematical Modelling and Numerical Analysis, 37 (2003), pp. 41--62

    2003

  35. [35]

    Ito and K

    K. Ito and K. Kunisch , Optimal control of obstacle problems by H^1 -obstacles , Applied Mathematics and Optimization, 56 (2007), pp. 1--17

    2007

  36. [36]

    Jaillet, D

    P. Jaillet, D. Lamberton, and B. Lapeyre , Variational inequalities and the pricing of American options , Acta Applicandae Mathematica, 21 (1990), pp. 263--289

    1990

  37. [37]

    Kinderlehrer and G

    D. Kinderlehrer and G. Stampacchia , An Introduction to Variational Inequalities and Their Applications , SIAM, 2000

    2000

  38. [38]

    M.-C. Lai, Y. Song, X. Yuan, H. Yue, and T. Zeng , The hard-constraint PINNs for interface optimal control problems , SIAM Journal on Scientific Computing, 47 (2025), pp. C601--C629

    2025

  39. [39]

    R. Liu, Z. Liu, W. Yao, S. Zeng, and J. Zhang , Moreau envelope for nonconvex bi-level optimization: A single-loop and Hessian -free solution strategy , in International Conference on Machine Learning, PMLR, 2024, pp. 31566--31596

    2024

  40. [40]

    L. Lu, R. Pestourie, W. Yao, Z. Wang, F. Verdugo, and S. G. Johnson , Physics-informed neural networks with hard constraints for inverse design , SIAM Journal on Scientific Computing, 43 (2021), pp. B1105--B1132

    2021

  41. [41]

    D. Luo, T. O’Leary-Roseberry, P. Chen, and O. Ghattas , Efficient PDE-constrained optimization under high-dimensional uncertainty using derivative-informed neural operators , SIAM Journal on Scientific Computing, 47 (2025), pp. C899--C931

    2025

  42. [42]

    Meyer, A

    C. Meyer, A. Rademacher, and W. Wollner , Adaptive optimal control of the obstacle problem , SIAM Journal on Scientific Computing, 37 (2015), pp. A918--A945

    2015

  43. [43]

    Meyer and O

    C. Meyer and O. Thoma , A priori finite element error analysis for optimal control of the obstacle problem , SIAM Journal on Numerical Analysis, 51 (2013), pp. 605--628

    2013

  44. [44]

    Mignot , Contr \^o le dans les in \'e quations variationelles elliptiques , Journal of Functional Analysis, 22 (1976), pp

    F. Mignot , Contr \^o le dans les in \'e quations variationelles elliptiques , Journal of Functional Analysis, 22 (1976), pp. 130--185

    1976

  45. [45]

    Mignot and J

    F. Mignot and J. P. Puel , Optimal control in some variational inequalities , SIAM Journal on Control and Optimization, 22 (1984), pp. 466--476

    1984

  46. [46]

    Mowlavi and S

    S. Mowlavi and S. Nabi , Optimal control of PDEs using physics-informed neural networks , Journal of Computational Physics, 473 (2023), p. 111731

    2023

  47. [47]

    Nemirovski, A

    A. Nemirovski, A. Juditsky, G. Lan, and A. Shapiro , Robust stochastic approximation approach to stochastic programming , SIAM Journal on Optimization, 19 (2009), pp. 1574--1609

    2009

  48. [48]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis , Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , Journal of Computational Physics, 378 (2019), pp. 686--707

    2019

  49. [49]

    Schiela and D

    A. Schiela and D. Wachsmuth , Convergence analysis of smoothing methods for optimal control of stationary variational inequalities with control constraints , ESAIM: Mathematical Modelling and Numerical Analysis, 47 (2013), pp. 771--787

    2013

  50. [50]

    Sirignano and K

    J. Sirignano and K. Spiliopoulos , DGM : A deep learning algorithm for solving partial differential equations , Journal of Computational Physics, 375 (2018), pp. 1339--1364

    2018

  51. [51]

    Y. Song, X. Yuan, and H. Yue , Accelerated primal-dual methods with enlarged step sizes and operator learning for nonsmooth optimal control problems , arXiv preprint arXiv:2307.00296, (2023)

    Pith/arXiv arXiv 2023

  52. [52]

    Y. Song, X. Yuan, and H. Yue , The ADMM-PINNs algorithmic framework for nonsmooth PDE-constrained optimization: A deep learning approach , SIAM Journal on Scientific Computing, 46 (2024), pp. C659--C687

    2024

  53. [53]

    Y. Song, X. Yuan, H. Yue, and T. Zeng , An operator learning approach to nonsmooth optimal control of nonlinear PDEs , arXiv preprint arXiv:2409.14417, (2024)

    Pith/arXiv arXiv 2024

  54. [54]

    T. M. Surowiec , Numerical optimization methods for the optimal control of elliptic variational inequalities , in Frontiers in PDE-Constrained Optimization, Springer, 2018, pp. 123--170

    2018

  55. [55]

    Wachsmuth , Strong stationarity for optimal control of the obstacle problem with control constraints , SIAM Journal on Optimization, 24 (2014), pp

    G. Wachsmuth , Strong stationarity for optimal control of the obstacle problem with control constraints , SIAM Journal on Optimization, 24 (2014), pp. 1914--1932

    2014

  56. [56]

    Wachsmuth , Towards M-stationarity for optimal control of the obstacle problem with control constraints , SIAM Journal on Control and Optimization, 54 (2016), pp

    G. Wachsmuth , Towards M-stationarity for optimal control of the obstacle problem with control constraints , SIAM Journal on Control and Optimization, 54 (2016), pp. 964--986

    2016

  57. [57]

    S. Wang, M. A. Bhouri, and P. Perdikaris , Fast PDE-constrained optimization via self-supervised operator learning , arXiv preprint arXiv:2110.13297, (2021)

    Pith/arXiv arXiv 2021

  58. [58]

    X. E. Zhao, W. Hao, and B. Hu , Two neural-network-based methods for solving elliptic obstacle problems , Chaos, Solitons & Fractals, 161 (2022), p. 112313

    2022

  59. [59]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTION or pop #1 'skip if FUNCTION new.block.checka empty 'skip 'new.block if FUNCTION field.or.null duplicate empty pop "" 'skip ...