acceptodds
Under review as a conference paper at ICLR 2027

MANIFEST: Manifold-Informed Learning of Feasibility-Enforcing Neural Solvers for Constrained Optimization

Abstract

Efficient solution of parametric and nonlinear optimization problems is important in several engineering applications. Neural network surrogates have emerged as a promising alternative to traditional optimization solvers in real-time applications where fast solutions are necessary. However, ensuring constraint satisfaction, especially with equality constraints, remains a key challenge in neural solvers. We address this by introducing MANIFEST, a Manifold-Informed Feasibility Enforcement Strategy, that directly incorporates the geometry of the equality-constraint manifold into neural surrogate training, thus guiding the learned neural map toward the feasible manifold before any correction is applied. Specifically, we design a training objective that integrates optimality, distance of the predictions from the feasible manifold, and first-order information about the local geometry and parametric variation of the feasible manifold, followed by a differentiable restoration step. In contrast to existing approaches that rely on post-prediction projection or restoration for constraint enforcement, MANIFEST trains the learned map itself to remain close to the feasible manifold, thereby avoiding substantial post-prediction corrections and associated numerical issues from projections in challenging optimization landscapes. We evaluate our framework on quadratically constrained quadratic programs (QCQPs), nonlinear programs (NLPs), and nonconvex nonlinear AC optimal power flow (AC-OPF). Across these problems, MANIFEST consistently outperforms state-of-the-art approaches, achieving near-optimal solution with negligible equality and inequality violations even without correction, with subsequent restoration further reducing violation residuals to numerical zero.

Primary area:
optimization
Keywords:
constrained optimization, neural surrogates, deep learning, optimal power flow

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3wJAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

MANIFEST is a neural framework for parametric constrained optimization. It trains a primal network using a Gauss–Newton-based geometric distance loss, a first-order loss on changes in equality residuals with respect to problem parameters, and a primal–dual objective; differentiable Gauss–Newton steps are also incorporated during training. The method is evaluated on parametric QCQPs, convex NLPs, and the IEEE 57-bus AC-OPF problem. Compared with DC3, FSNet, and NLPOpt-Net, MANIFEST achieves lower mean equality violations across the benchmarks. Its raw optimality gap is lower on the nonlinear NLP and AC-OPF tasks, but higher than those of FSNet and NLPOpt-Net on QCQP.

Critical strengths

The distance loss uses the equality-constraint Jacobian to measure the distance between a prediction and the feasible set, instead of penalizing the residual alone without accounting for constraint geometry. In its unregularized form, it is invariant to nonsingular reparameterizations of the equality constraints. The first-order loss penalizes changes in equality residuals as problem parameters vary, encouraging the solution map to track the feasible manifold across instances. With the primal–dual objective and differentiable Gauss–Newton updates, these terms allow MANIFEST to train from problem data without solver-generated solution labels.

Critical weaknesses
  1. The paper does not isolate the contributions of the geometric distance loss, the first-order parameter-variation loss, the primal–dual loss, or the differentiable Gauss–Newton steps. Nor does it provide a direct comparison between the geometric distance loss and a conventional equality-residual penalty.

  2. The QCQP results expose a trade-off between feasibility and optimality. MANIFEST achieves lower raw equality violations, but its raw relative optimality gap is substantially larger than those of FSNet and NLPOpt-Net. After correction, the baselines’ gaps decrease further, whereas MANIFEST’s remains unchanged. The paper notes that the baselines perform additional objective optimization during correction, so the corrected results are not directly comparable.

  3. Inference-time comparisons are not reported. Although MANIFEST can return the raw network prediction, it also permits Gauss–Newton restoration at inference to meet an equality-feasibility tolerance. Consequently, the cost of obtaining a feasible output remains unclear.

  4. The paper does not adequately clarify how MANIFEST relates to sensitivity-informed AC-OPF methods, particularly SI-DNN (Singh, Kekatos, and Giannakis, “Learning to Solve the AC-OPF Using Sensitivity-Informed Deep Neural Networks,” IEEE Transactions on Power Systems, 2022), which leverages solution sensitivities with respect to loads during training.

Questions for authors
  1. The authors should clarify the extent to which each training component contributes to feasibility and optimality, and in particular to compare the geometric distance loss with a standard equality-residual penalty under the same training setup.

  2. It would also be helpful to explain the apparent tension on QCQP, where MANIFEST achieves lower equality violation but a substantially larger raw optimality gap, and to provide a comparison that accounts for the additional objective optimization used when correcting the baselines.

  3. The paper should report wall-clock inference times for both the raw network prediction and the restored prediction, alongside the baselines under the same hardware and stopping criteria, and indicate how many Gauss–Newton iterations are typically required.

  4. Finally, the authors should clarify how MANIFEST’s local feasible-manifold analysis relates to the use of solution sensitivities in SI-DNN, what guarantees it offers beyond penalizing parameter-dependent residual changes, and what advantage follows from not requiring sensitivity labels.

3 comments

Thanks for your time and feedback.

Ablation studies:

[..] clarify the extent to which each training component contributes to feasibility and optimality, and [..] compare the geometric distance loss with a standard equality-residual penalty [..]

We have added ablation experiments in Appendix D on the AC-OPF case to demonstrate the contributions of the geometric loss components. Specifically, we examine

(i) the manifold-aware geometric distance loss LD\mathcal L_D by replacing it with a conventional equality-residual penalty ∣∣g(x,Fθ(x))∣∣22||g(x,F_{\theta}(x))||_2^2;

(ii) the first-order loss LFO\mathcal L_{FO} by removing this term and comparing against the full MANIFEST algorithm.

ConfigurationEq. MAEEq. MaxOpt. gap (%)
Conventional equality-residual1.17e-42.33e-20.01908
MANIFEST without LFO\mathcal L_{FO}7.42e-56.16e-30.007626
Full MANIFEST7.50e-64.50e-50.006600

The geometry-aware distance loss reduces equality MAE by approximately 36.8%, maximum equality violation by 73.5%, and optimality gap by 60% compared with the conventional residual loss. Including the first-order loss further substantially improves raw-prediction feasibility and optimality.

QCQP Experiment:

It would also be helpful to explain the apparent tension on QCQP, where MANIFEST achieves lower equality violation but a substantially larger raw optimality gap, and to provide a comparison that accounts for the additional objective optimization used when correcting the baselines.

The QCQP is a relatively simpler convex problem, where the additional objective optimization performed by competing methods during inference-time can effectively solve a substantial portion of the QCQP. Their improved post-correction optimality gaps therefore reflect additional optimization beyond the neural prediction, whereas MANIFEST's Gauss–Newton restoration targets only equality feasibility.

Importantly, both baselines' raw predictions and solution after restoration have substantially higher equality violations, so their smaller raw objective gaps do not necessarily indicate better solutions to the constrained problem. MANIFEST achieves over an order of magnitude lower raw equality violations while maintaining a small optimality gap of 0.52%, without further objective optimization.

Further, on the more challenging NLP and nonconvex AC-OPF benchmarks, MANIFEST achieves substantially better raw feasibility and optimality than the baselines (Table 1).

Inference times:

The paper should report wall-clock inference times [..] alongside the baselines

We have updated the manuscript with wall-clock inference comparisons (Table 1), including each method's native correction procedure.

MethodQCQP (ms)NLP (ms)AC-OPF (ms)
MANIFEST1.7061.68215.154
DC39.0998.71922.496
FSNet13.98514.689N/A
NLPOpt-Net199.699128.977764.932

MANIFEST achieves >5× faster total inference than DC3 on QCQP and NLP, and takes less than 0.016 s on the challenging AC-OPF benchmark, including restoration. The measured restoration uses only 3 Gauss–Newton iterations for QCQP/NLP and 11 for AC-OPF. This efficiency stems from shifting geometric feasibility enforcement to offline training, enabling the primal network to produce near-feasible raw predictions that require minimal inference-time correction, unlike competing methods that rely more heavily on post-prediction optimization.

SI-DNN Comparison:

[..] clarify how MANIFEST’s local feasible-manifold analysis relates to the use of solution sensitivities in SI-DNN, what guarantees it offers beyond penalizing parameter-dependent residual changes, and what advantage follows from not requiring sensitivity labels."

The key distinction is that SI-DNN is a supervised learning approach for AC-OPF that requires optimal solution labels and their sensitivities with respect to problem parameters, obtained by solving the underlying optimization problems and computing the corresponding sensitivities. In contrast, MANIFEST is entirely unsupervised, requiring neither optimal solutions nor sensitivity labels. Further, we provide a feasibility-oriented framework applicable to general nonlinear optimization problems beyond AC-OPF, without the computational burden of generating optimal solution and sensitivity labels.

Our first-order loss (Eq. 13) directly penalizes the parameter-dependent variation of equality-constraint residuals, using Jacobian–vector products computed from the learned network and known constraints. Rather than learning the sensitivity of an optimal solution map, MANIFEST directly learns to preserve feasibility as problem parameters vary, based on the local geometry of the equality-feasible manifold.

Thank you for the additional experiments. In Weakness 1, I also asked about the contribution of the differentiable Gauss–Newton steps used during training. Table 3 does not test this component. If you have added a comparison without enabling the differentiable Gauss–Newton, keeping the other losses and training settings the same, could you also report the raw constraint violations and objective gaps? Also, does the conventional-residual baseline include the first-order loss? The reported 36.8% improvement comes from comparing the first two rows, so their first-order-loss settings need to be the same to attribute the difference to the geometric distance loss.

We thank the reviewer for the follow-up. We have conducted the additional differentiable Gauss–Newton ablation and updated Table 3 and the corresponding discussion in the revised manuscript to report the contributions of the geometric distance loss, first-order loss, and differentiable Gauss–Newton training.

1. Contribution of differentiable Gauss–Newton training. We evaluate MANIFEST with the differentiable Gauss–Newton steps disabled during training (K=0K=0), while retaining the first-order loss and all other training settings. The updated Table 3 reports the results.

MethodEq. MAEEq. MaxOpt. Gap (%)
Conventional equality residual (without FO term)1.17e-42.33e-21.91e-2
Geometric distance (without FO term)7.42e-56.16e-37.63e-3
MANIFEST without differentiable GN (K=0K=0, with FO term)1.63e-51.73e-46.52e-3
Full MANIFEST7.46e-64.47e-56.56e-3

Table 3: IEEE 57-bus AC-OPF raw-prediction ablations. All metrics are evaluated without inference-time restoration.

Including differentiable GN reduces raw equality MAE by 54.2% and maximum equality violation by 74.2%, while maintaining the same reported optimality gap. Importantly, these improvements are measured on raw neural predictions without inference-time restoration, demonstrating that differentiating through the Gauss–Newton iterations during training improves the learned solution map itself, rather than merely correcting its outputs afterward.

2. Geometric distance versus conventional residual. We confirm that both the conventional equality-residual baseline and the geometric-distance baseline exclude the first-order loss, i.e., ωFO=0.\omega_{\mathrm{FO}}=0.

Thus, the reported 36.8% reduction in equality MAE and 73.5% reduction in maximum equality violation are obtained under the same first-order-loss setting. These improvements therefore cannot be attributed to first-order regularization and demonstrate the benefit of the geometric-distance construction in the tested configurations. Furthermore, including the first-order term in full MANIFEST yields additional improvements in raw feasibility.

Together, these results demonstrate the complementary contributions of the geometric-distance loss, first-order regularization, and differentiable Gauss–Newton training to MANIFEST's raw-prediction feasibility. We have incorporated the expanded ablation table and these clarifications into the revised manuscript.

In summary, these experiments demonstrate the importance and critical role of each component in the MANIFEST framework.

We hope that this addresses your remaining concerns.

Reviewer_gzx1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a new framework for incorporating the geometry of the equality-constraint manifold into neural surrogate training by designing a new training objective with 3 components: geometric distance loss, first-order loss, and primal-dual training with differentiable restoration. The motivation of the paper is that the feasibility of outputs given by existing methods may be dominated by post-prediction correction mechanisms (e.g. iterative differentiable projection layers), as opposed to the learned neural surrogate’s proposed solutions. The method tries to minimize the impact of these correction mechanisms to reduce computation time and accumulation of numerical sensitivities. Results show competitive or improved optimality gaps and mean violations on a variety of problems and constraints.

Critical strengths
  1. The framework is conceptually coherent, with individual components that have clear geometric interpretations.
  2. The empirical results across problem classes are strong and show good performance on hard problems with nonconvex constraints, including an AC-OPF example.
Critical weaknesses
  1. The work is motivated by the desire to eliminate/reduce the computational cost of projection methods. However, it never quantifies that impact, either on training or inference time.
  2. The designed objective has several computationally nontrivial components, including a Jacobian-based geometric loss, first-order derivatives, a dual network, and a differentiable Gauss-Newton correction. However, at this stage, the paper does not provide sufficient computational comparisons or ablations to show how much each component contributes to the final performance.
Questions for authors
  1. Could the authors provide ablation experiments on including or not including each component/loss in the framework? E.g., is a primal/dual NN necessary? How does each component contribute to raw-prediction feasibility and optimality?
  2. Could the authors quantify the impact on inference time against the baselines, and include a comparison with classical optimization algorithm baselines? Because one of the main problems the paper addresses is reducing computation time (by reducing the need for post-prediction corrections), a broader suite of baselines and additional metrics on computational efficiency would help.
  3. Could the authors discuss the scalability of the framework?
3 comments

We thank the reviewer for their time and critical feedback.

Computation Time:

Could the authors quantify the impact on inference time against the baselines, and include a comparison with classical optimization algorithm baselines? Because one of the main problems the paper addresses is reducing computation time (by reducing the need for post-prediction corrections), a broader suite of baselines and additional metrics on computational efficiency would help.

Thank you for this question. We have updated the manuscript with wall-clock training and inference comparisons (Table 1 and Appendix D). The following table reports total inference time, including each method's native correction procedure.

MethodQCQP (ms)NLP (ms)AC-OPF (ms)
MANIFEST1.7061.68215.154
DC39.0998.71922.496
FSNet13.98514.689N/A
NLPOpt-Net199.699128.977764.932

MANIFEST achieves over 5× faster total inference than DC3 on QCQP and NLP, and takes <0.016 s for the challenging AC-OPF case, including restoration. MANIFEST also achieves faster training in the QCQP and NLP cases (Appendix D). While MANIFEST incurs greater offline training cost in AC-OPF than DC3, we note that DC3 produces a suboptimal solution with significant equality violations that persist despite correction in this setting.

Regarding classical optimization baselines, our comparisons focus on state-of-the-art learning-based optimization methods solving the same constrained optimization problems, enabling direct comparisons under common formulations. Classical optimization solvers are used as references for evaluating solution quality, and not as the primary computational baselines.

Ablation Studies:

The designed objective has several computationally nontrivial components [..] However, at this stage, the paper does not provide sufficient computational comparisons or ablations to show how much each component contributes to the final performance."

Could the authors provide ablation experiments on including or not including each component/loss in the framework? E.g., is a primal/dual NN necessary? How does each component contribute to raw-prediction feasibility and optimality?"

We have added ablation experiments in Appendix D (Table 3) to evaluate the contributions of the geometric distance and first-order loss terms on the AC-OPF case.

ConfigurationEq. MAEEq. MaxOpt. gap (%)
Conventional residual loss1.17e-42.33e-20.01908
MANIFEST without first-order loss7.42e-56.16e-30.007626
Full MANIFEST7.50e-64.50e-50.006600

These results demonstrate the benefits of the proposed geometric losses, with full MANIFEST achieving substantially lower raw equality violations and improved optimality.

Regarding the remaining components, primal–dual methods are well established in constrained optimization theory and are also employed in learning-based methods such as NLPOpt-Net. They provide a principled framework for jointly promoting optimality and constraint satisfaction through adaptive multipliers.

The differentiable Gauss–Newton restoration is also theoretically supported: Theorem 1(ii) establishes local quadratic convergence of equality restoration, faster than the linear rate of gradient-based feasibility seeking under the stated conditions, while Proposition 10 connects the geometric distance loss to a sufficient initialization condition for this convergence. We additionally quantify the computational costs of these components in Appendix D (Table 5).

Scalability:

Could the authors discuss the scalability of the framework?

MANIFEST is designed for general parametric constrained optimization without problem-specific completion or solver-generated training labels. The principal scalability advantage is the low online computational cost: once trained, raw inference requires only a forward pass through the primal network, independent of the geometric losses, dual network, and Jacobian-based linear solves used during training. This makes the approach particularly suitable for repeated optimization across many parameter instances.

For larger problem dimensions, the geometric computations and optional Gauss–Newton restoration incur additional costs associated with Jacobian evaluations and linear solves. However, the first-order loss uses Jacobian–vector products, avoiding construction of the full parameter-residual Jacobian. On the challenging IEEE 57-bus AC-OPF benchmark, MANIFEST achieves <0.016s for the complete inference, demonstrating efficient online computation even for nonlinear, nonconvex constraints. These properties provide a promising basis for scaling to larger applications.

Thank you to the authors for addressing these concerns and comments! After reading this response and the discussion between the authors and other reviewers, I maintain my score of 2.

We thank the reviewer for taking the time to examine our responses and the broader discussion. We appreciate the acknowledgment that we have addressed the raised concerns.

We have provided substantial additional experimental and theoretical evidence, including component ablations, detailed training and end-to-end inference comparisons, and clarification of our convergence guarantees. These additions directly respond to the weaknesses identified in the original review.

Given these substantive revisions, we would appreciate clarification on which specific technical concerns remain unresolved and continue to motivate the current score. We respectfully encourage the reviewer to reconsider the assessment in light of the additional evidence, particularly the demonstrated improvements in raw feasibility and computational efficiency over existing methods.

Reviewer_bH4NOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MANIFEST, an unsupervised framework for learning solutions to parametric constrained optimization problems. It combines a Jacobian-based geometric distance loss and a penalty on constraint-residual derivatives with primal–dual training and differentiable Gauss–Newton restoration. Experiments on QCQPs, nonlinear programs, and IEEE 57-bus AC-OPF evaluate solution quality and constraint violations before and after correction, showing improvements particularly in the equality feasibility of raw predictions.

Critical strengths

The paper addresses the important challenge of learning near-feasible solutions to constrained optimization problems without optimal-solution labels. Separately reporting raw and restored predictions is informative, and the substantial reductions in raw equality violations on the NLP and AC-OPF benchmarks are promising.

Critical weaknesses
  1. The proposed method still requires Jacobian evaluations and linear solves during training and inference-time restoration. However, the paper provides neither a sufficiently detailed computational cost comparison nor training and inference runtimes, leaving its practical efficiency advantages unclear.

  2. The theoretical analysis mainly characterizes local geometry and restoration steps. It does not establish convergence of the primal–dual network training procedure or provide feasibility guarantees or generalization error bounds for unseen parameter instances.

  3. The experiments cover only fixed-size synthetic problems and a single IEEE 57-bus system. Without varying the number of variables, constraints, or buses, the results do not establish scalability to larger optimization problems.

  4. Although inequality violations are penalized in Eq. (15), the manifold construction and Gauss–Newton restoration target only equality constraints. As acknowledged in Appendix A, restoration need not preserve inequality feasibility. Table 1 still reports a maximum AC-OPF inequality violation of 3.7×10−63.7\times10^{-6} after restoration. Claims that all violation residuals reach numerical zero should therefore be qualified, and the adopted inequality-feasibility tolerance should be stated.

Questions for authors
  1. How does MANIFEST compare with the baselines in training time and total inference time, including restoration, under the same feasibility tolerances? Please clarify the costs of Jacobian evaluations and linear solves.

  2. How is inequality feasibility maintained during equality restoration in AC-OPF? What inequality-feasibility tolerance is used, and what fraction of test instances satisfy all constraints after restoration?

5 comments

Thanks for your time and feedback. We address your questions below.

Computation time:

How does MANIFEST compare with the baselines in training time and total inference time, including restoration [..]? Please clarify the costs of Jacobian evaluations and linear solves.

First, we clarify that MANIFEST does not require Jacobian evaluations or linear solves for raw inference. These operations are confined to training and optional Gauss–Newton restoration; raw inference uses only the primal network and produces near-feasible predictions without the dual network or geometric loss computations (Sec. 3.2).

We have also updated the manuscript with wall-clock training (Appendix D) and inference times (Table 1, summarized below).

MethodQCQP (ms)NLP (ms)AC-OPF (ms)
MANIFEST1.7061.68215.154
DC39.0998.71922.496
FSNet13.98514.689N/A
NLPOpt-Net199.699128.977764.932

MANIFEST achieves over 5× faster total inference than DC3 on QCQP and NLP, and takes <0.016s for the challenging AC-OPF case, including restoration. MANIFEST also achieves faster training in the QCQP and NLP cases (Appendix D). While MANIFEST incurs greater offline training cost in AC-OPF than DC3, we note that DC3 produces a sub-optimal solution with significant equality violations that persist despite correction in this setting. We additionally report Jacobian-related training costs in Appendix D (Table 5), with Gauss–Newton restoration, manifold-distance evaluation, and first-order Jacobian–vector products requiring 2.411 s, 0.207 s, and 3.230 s per epoch, respectively.

We also reiterate that the goal is not to eliminate Jacobian computations altogether, but to learn near-feasible raw predictions that reduce inference-time correction costs.

Theoretical analysis:

[..] does not establish convergence of the primal–dual network training procedure or provide feasibility guarantees or generalization error bounds [..]

Thanks for raising this point. We have added Section 3.3, "Convergence Guarantees," (revised manuscript), establishing convergence results for both the outer primal–dual training procedure and the inner feasibility-restoration iterations. for the outer primal–dual training and inner restoration loops. Specifically, Theorem 1(i) establishes linear convergence of the primal and dual optimality gaps under standard two-sided Polyak–Łojasiewicz (PL) and smoothness conditions, without requiring convexity or concavity. Further, Theorem 1(ii) establishes quadratic convergence of the Gauss–Newton feasibility-restoration iterations, compared with the linear convergence rate of gradient-based feasibility seeking approaches under the PL condition (e.g., FSNet (Nguyen & Donti, 2025)). Under the stated regularity conditions, this improves the theoretical iteration complexity for reaching equality-feasibility tolerance ϵ\epsilon from O(log⁡(1/ϵ))O(\log(1/\epsilon)) to O(log⁡log⁡(1/ϵ))O(\log\log(1/\epsilon)). Empirically, all experiments in Section 4 are tested on unseen parameter instances (held-out test set), which addresses generalization.

Scalability:

[..] experiments cover only fixed-size synthetic problems [..] Without varying the number of variables, constraints, or buses, the results do not establish scalability [..]

First, varying the number of variables, constraints, or buses defines a different optimization problem, not another parameter instance. Second, our experiments use standard benchmark problem sizes comparable to those in existing works, enabling direct comparisons. The IEEE 57-bus AC-OPF is already a challenging NP-hard nonconvex optimization problem, as evidenced by the substantially degraded feasibility and optimality performance of the competing methods (Table 1). Finally, our primary focus is not scalability with problem dimension, but improving raw-prediction feasibility and optimality while reducing inference-time correction costs.

Inequality feasibility:

How is inequality feasibility maintained during equality restoration in AC-OPF? What inequality-feasibility tolerance is used, and what fraction of test instances satisfy all constraints after restoration?

As explicit in Eq. (15), inequality feasibility is promoted during training through the augmented Lagrangian, while Gauss–Newton restoration targets equality constraints. The reported AC-OPF inequality violations are already very small, with a mean of 1.7×10−81.7\times10^{-8} after restoration (Table 1). Thus, our claim of numerical-zero residuals refers specifically to equality constraints after restoration, not exact inequality feasibility; we have clarified this in the revision.

Finally, a uniform feasibility tolerance across AC-OPF constraints is not necessarily meaningful due to their different physical scales. We therefore report the mean and maximum violations in Table 1 rather than relying on an arbitrary binary feasibility rate.

  1. Runtime. Thank you for adding the timings. They show a speedup at the tested sizes; whether it persists on larger instances remains unknown.
  2. Convergence. The outer-loop result assumes two-sided PL and uses full gradients. The paper neither establishes PL for the network objective nor proves convergence of the actual minibatch training.
  3. DC3 comparison. The QCQP and NLP have fixed equality Jacobians, so DC3 can precompute its affine completion. Was this enabled in the timed baseline? Moreover, both synthetic benchmarks use a fixed equality matrix. Could the authors also test a setting with A=A(x)A=A(x) to assess whether the inference speedup persists when a single factorization cannot be reused across instances?

Thanks for engaging with our rebuttal and acknowledging our experimental performance. We address your remaining questions below.

DC3 comparison. QCQP and NLP have fixed equality Jacobians, so DC3 can precompute its affine completion. Was this enabled [..]? [..] both synthetic benchmarks use a fixed equality matrix. Could the authors also test a setting with (A=A(x)) [..]?

Runtime. [..] They show a speedup at the tested sizes; whether it persists on larger instances remains unknown.

a) DC3 comparison. For QCQP and NLP, we evaluate DC3 using the native implementation and prescribed correction procedure, with the factorization reuse enabled. Thus, our comparison already accounts for the computational advantage arising from the fixed equality matrix. Even with this optimization, MANIFEST achieves substantially lower end-to-end inference times: 1.706 ms versus 9.099 ms for DC3 on QCQP, and 1.682 ms versus 8.719 ms on NLP, including MANIFEST's Gauss-Newton restoration and linear solves.

b) Evaluation with varying Jacobians. We already include a nonlinear benchmark with varying equality Jacobians, in the more challenging IEEE 57-bus AC-OPF case. From the AC power-balance constraints in Eqs. (45)–(46) of Appendix C, the equality Jacobian with respect to y=(Pg,Qg,Vr,Vi)y=(P_g,Q_g,V_r,V_i) is

\begin{bmatrix} C & 0 & -J_{P,V_r} & -J_{P,V_i} \\\\ 0 & C & -J_{Q,V_r} & -J_{Q,V_i} \end{bmatrix},$$ where $C$ is the generator-to-bus incidence matrix and with $Y_{\mathrm{bus}}=G+jB$, and $$J_{P,V_r} = \operatorname{diag}(GV_r-BV_i) + \operatorname{diag}(V_r)G + \operatorname{diag}(V_i)B,$$ $$J_{P,V_i} = \operatorname{diag}(BV_r+GV_i) - \operatorname{diag}(V_r)B + \operatorname{diag}(V_i)G,$$ $$J_{Q,V_r} = -\operatorname{diag}(BV_r+GV_i) - \operatorname{diag}(V_r)B + \operatorname{diag}(V_i)G,$$ $$J_{Q,V_i} = \operatorname{diag}(GV_r-BV_i) - \operatorname{diag}(V_i)B - \operatorname{diag}(V_r)G.$$ Unlike the fixed equality matrices in the synthetic benchmarks, the Jacobian depends explicitly on the predicted voltages $V_r,V_i$ and changes across problem instances, so neither the numerical Jacobian nor a single factorization of its associated linear system can be precomputed and reused. **Importantly, our inference advantage persists in this varying-Jacobian setting,** with end-to-end inference time of **15.154 ms vs 22.496 ms for DC3**. c) **Larger instances.** Our experiments span three established benchmark classes, including 100-variable QCQP and NLP and nonlinear, nonconvex AC-OPF with 128 variables and 115 equality constraints, consistent with the experimental scope of SoTA works. Importantly, **these are already challenging problems where existing methods struggle to achieve feasibility and optimality**; e.g., on AC-OPF, FSNet fails to produce finite solutions, DC3 exhibits significant equality violations, and NLPOpt-Net suffers from severe feasibility and optimality degradation. In contrast, MANIFEST achieves near-optimal solutions with substantially improved feasibility and lower end-to-end inference time. These findings demonstrate our effectiveness beyond synthetic problems, highlighting the practical merits alongside the rigorous theoretical guarantees established in the paper. > **Convergence.** [..] assumes two-sided PL and uses full gradients. The paper neither establishes PL for the network objective nor proves convergence of the actual minibatch training. a) **Convergence of minibatch training.** While we provide the general convergence results in Theorem 1, the convergence of minibatch training follows from this result; we provide a full proof in Corollary 2 of the revised manuscript. b) **PL conditions.** We emphasize that unconditional global convergence guarantees for arbitrary nonlinear, nonconvex optimization is mathematically impossible in general. Even classical optimization requires structural assumptions for global guarantees, e.g., linear programming is convex, and general nonconvex QP is already NP-hard. Thus, convergence analysis for these problems necessarily relies on additional regularity conditions. The PL condition is a well-established and widely adopted framework in nonconvex optimization, providing convergence guarantees without requiring convexity. Our theoretical analysis follows this established framework, consistent with SoTA methods such as FSNet. Notably, FSNet assumes global PL regularity, whereas our analysis requires only local Jacobian regularity, which is a less restrictive assumption. For primal-dual network training, we explicitly state the PL and smoothness assumptions for convergence and additionally establish a stochastic-gradient bound accounting for gradient variance and dual-response errors. Thus, we provide rigorous convergence guarantees under clearly specified conditions, consistent with established theoretical practice in this area.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.