acceptodds
Under review as a conference paper at ICLR 2027

Decision-Focused Learning under Joint Uncertainty: Local Geometry and Complementarity Structure

Abstract

Decision-focused learning trains predictors for decisions in downstream optimisation problems. While objective uncertainty leaves the feasible set unchanged, constraint uncertainty can reshape the feasible set, render the represented decision infeasible for the true problem, and alter the active set governing local sensitivity. To study these effects, we focus on strongly convex quadratic programs and introduce by formalising a principled training loss function integrating objective-induced decision divergence, true-constraint violation, and active-set disagreement. We establish finite-sample bounds linking the loss to operational regret, true-violation frequency and magnitude, and dormant-bound prediction error. We then characterise the local learning information available through the represented optimiser, i.e., identify decision-sensitivity components alongside decision-null bound directions that remain observable through the primal–dual complementarity structure. We further show that is regionally convex quadratic in the prediction errors, with an affine output-space gradient and explicit active-set mismatch offsets. We finally demonstrate the effectiveness of on IEEE power-grid benchmarks, with particular emphasis on constraint shifts and their consequences for decision quality and true feasibility.

Primary area:
optimization
Keywords:
Decision-focused learning, Constrained convex optimisation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_riweOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies decision-focused learning for strongly convex quadratic programs in which both objective coefficients and constraint right-hand sides are predicted. It introduces sensitivity-informed geometric and nullspace-aware learning (SIGNAL), which combines an objective-induced Bregman divergence, a true-feasibility loss, and a signed primal–dual margin to supervise constraint activity. The analysis characterizes how prediction errors affect the represented optimizer, including constraint directions that are locally invisible to the decision, and derives finite-sample bounds on operational regret, true constraint violations, and dormant-bound prediction error. The method is evaluated on IEEE 14-bus and 57-bus power-grid benchmarks under constraint shifts and observation noise.

Critical strengths

The paper treats errors in objective prediction separately from errors in constraint prediction, since the latter can change the feasible set and make regret hard to interpret for infeasible decisions. Its signed primal–dual margin gives a learning signal for constraints that are inactive at the current solution but may become active after a shift. The analysis relates these signals to changes in the optimizer, feasibility, and constraint activity. The ablation compares three loss configurations: Dec-Only uses the decision-divergence loss, Dec-Feas adds the feasibility loss, and SIGNAL adds the signed-margin loss. The paper also compares these methods with other learning-based optimization methods on IEEE systems.

Critical weaknesses
  1. No comparison with direct constraint-parameter prediction. The ablation adds feasibility and regime losses in stages, but does not compare regime supervision with directly predicting the true constraint right-hand sides using normalized MSE. It is unclear whether the margin loss helps beyond fitting those values.

  2. The role of dual multipliers is unclear. The signed margin uses slack for inactive constraints and multipliers for active constraints. The experiments do not separate the multiplier signal from the slack signal or compare the full margin with activity-label supervision.

  3. The experiments do not show what the theory explains. The paper derives sensitivity relations, finite-sample bounds, and a condition for active-set recovery. The experiments report regret and constraint violations under shifts and noise, but do not show whether the bounds are informative, whether the recovery condition predicts active-set errors, or whether the sensitivity analysis explains the ablation results.

  4. The analysis assumes fixed problem matrices. It considers changes in the objective’s linear coefficients and the constraint right-hand sides, while keeping the Hessian and constraint left-hand-side matrices fixed. It therefore does not show how the method handles changes in objective curvature or constraint coefficients.

  5. The experiments do not test how performance changes with problem size. The two power-grid benchmarks use reduced problems with only two and six decision variables. The paper reports no results on larger problems, so it is unclear how solution quality and computational cost change as dimension grows.

Questions for authors
  1. How does SIGNAL compare with a scale-normalized MSE baseline that predicts all uncertain objective coefficients and constraint right-hand sides? Reporting results for uncertain inequality bounds separately would help isolate the effect of regime supervision.

  2. Which part of the signed-margin loss contributes to the results? Comparisons with slack-only, multiplier-only, and activity-label variants would help answer this.

  3. Can the theoretical bounds and active-set recovery condition help explain the experimental results? For example, are the bounds informative on the tested cases, and does the sensitivity analysis account for the ablation results?

  4. How would SIGNAL behave if the Hessian or constraint left-hand-side matrices were uncertain? If these cases are outside the paper’s scope, could the authors state this limitation and narrow the claims?

  5. How does performance change as the problem size grows? Results on larger QPs, with solution quality and computational cost reported against problem dimension, would help assess this.

4 comments

We sincerely thank the reviewer for their insightful comments and questions. Our revision incorporates a structural and empirical auxiliary-loss study (Appendices F, G, J8), empirical diagnostics (Appendix J6), a matrix-uncertainty scope discussion (Appendix H), training-cost measurements (J5), and a higher-dimensional EV-charging experiment (Appendix J9).

Response to Q1 and Q2.\textbf{Response to Q1 and Q2.} We thank the reviewer for pointing to the need for the adoption of normalised bound-prediction and various versions of complementarity-aware supervision within the same decision and feasibility framework. In the revised version, we have addressed this issue comprehensively by defining normalised-bound prediction, slack-only supervision, multiplier-only supervision, and sign-only supervision. In particular, in Appendices F and G, we analyse this distinction structurally for the four possible constraint-activity configurations. When a row is inactive in both true and represented problems, LregL_{\text{reg}} reduces exactly to slack-only supervision; if GiδzG_i\delta z is negligible, normalised-bound prediction provides an approximate parameter-space proxy. When a row is active in both problems, LregL_{\text{reg}} reduces exactly to multiplier-only supervision. In these stable one-sided regimes, the simpler continuous losses can therefore provide appropriate surrogates. The full signed margin is particularly distinctive when activity statuses differ. On a missed-active row, it combines the represented positive primal slack with the true dual multiplier; on a falsely active row, it combines the true positive primal slack with the represented dual multiplier. In both cases the two quantities enter with the same sign after taking the signed-margin difference. Sign-only supervision captures a different, coarser aspect of this structure. It assigns zero loss whenever the activity labels agree and a fixed loss of 1/k1/k whenever they disagree. Thus, it detects activity misclassification but discards continuous information about distance to activation, binding strength, and the severity of margin disagreement. Apart from this analysis, it is important to note that, for an inactive inequality, the primal slack si>0s_i>0 measures its distance to binding, equivalently the bound tightening required to reach activation. However, for a strictly active inequality, the dual multiplier μi>0\mu_i>0 quantifies its binding strength through the marginal objective-value improvement from bound relaxation; in the isolated-inequality setting, the relaxation required to release the constraint is di2μid_i^2\mu_i. Building on this observation, we show in Proposition 3 through a bound perturbation analysis that the signed primal-dual margin matters in all four constraint-activity configurations. Proposition 4 further establishes a deterministic certificate linking sufficiently small regime-supervision loss to exact active-set recovery on the supervised constraints. Finally, in Appendix J8, we compare the resulting variants (Dec-Feas-NMSE, Dec-Feas-Slack, Dec-Feas-Mult, and Dec-Feas-Sign) with SIGNAL, using the same predictor class, differentiable backend, componentwise clipping procedure, calibration protocol, and paired evaluation settings. This study evaluates five conditions: the OOD baseline, +2%+2\%/+4%+4\% relaxation of two training-exposed physical line ratings, and −2%-2\%/−4%-4\% tightening of two training-dormant physical line ratings. SIGNAL attains the lowest mean operational regret and violation magnitude in each condition. Dec-Feas-NMSE has a slightly lower mean violation frequency in the baseline and two relaxation conditions, but exhibits higher mean violation magnitude in all conditions. These metrics should be interpreted jointly, as lower violation frequency does not necessarily imply better overall performance. Dec-Feas-Slack regret is close to SIGNAL in several conditions. Nevertheless, the results suggest that combining slack and multiplier information in the full signed-margin formulation can be beneficial, as SIGNAL achieves lower violation frequency and magnitude than Dec-Feas-Slack across all evaluated conditions.

Response to Q3.\textbf{Response to Q3.} We thank the reviewer for their question regarding empirical diagnostics of the theoretical results. In the revised version, we addressed this issue. In particular, to directly connect our theoretical analysis with empirical behaviour, we have added empirical diagnostics in Appendix J6, covering the IEEE 57-bus experiment. First, the active-set certificate is verified across all evaluated instances, with zero active-set mismatches among 16,87416,874 certified ID and 5,8065,806 certified OOD instances; the deterministic downstream bounds associated with the finite-sample mean loss are also satisfied, albeit conservatively, consistent with their role as analytical controls. Second, the tangent-normal sensitivity decomposition is confirmed to numerical precision, showing that the relative contributions of objective and active-constraint errors differ markedly between ID and OOD conditions: objective-induced tangent motion accounts for 88.1%88.1\% of eligible ID decision loss, whereas active-constraint RHS errors account for 67.2%67.2\% of eligible OOD decision loss through normal motion. Third, we directly verify the dormant-bound blind spot: 88.2%88.2\% of ID and 65.7%65.7\% of OOD physical-rating coordinates are locally decision-null, with vanishing decision and feasibility gradients, while regime supervision provides nonzero gradients on all evaluated such coordinates. Together, these findings establish an empirical connection between the proposed geometric analysis, constraint-activity supervision, and the distinct learning signals available under joint objective-constraint uncertainty.

Response to Q4.\textbf{Response to Q4.} We thank the reviewer for stating their concern about the scope of the proposed framework. In the revised version, we have addressed this issue by providing a discussion about the scope beyond strongly convex QPs. we clarify that true-feasibility shaping extends naturally by evaluating predicted decisions against true constraints, while signed-margin supervision and its active-set certificate follow from primal-dual complementarity, provided corresponding constraints, compatible positive scales, and well-defined KKT-pair selections are available. In contrast, the fixed objective-curvature metric, canonical normalisation, and exact regional affine sensitivity structure rely on the strongly convex QP formulation with fixed problem matrices. We then discuss the extension challenges for different constrained optimisation problems. In particular, we highlight that, for a locally regular active set, perturbations in HH and in the active-constraint matrix JJ modify the KKT differential to \begin{align} \begin{bmatrix} H&J^\top \newline J&0 \end{bmatrix} \begin{bmatrix} dz \newline dy \end{bmatrix} =\begin{bmatrix} -(dH)z-dc-(dJ)^\top y\newline dr-(dJ)z \end{bmatrix}. \end{align} The terms (dH)z(dH)z, (dJ)⊤y(dJ)^\top y, and (dJ)z(dJ)z are absent from the fixed-geometry analysis. If the Hessian varies, positive definiteness must be maintained for the strongly convex solution setting and the true-objective divergence uses the true sample-specific curvature. If constraint coefficients vary, active manifold orientations and normalisations change; compatible correspondence and scaling are needed for true/represented margin comparisons. Although local implicit differentiation remains possible under a nonsingular KKT system, the exact fixed-region affine map, fixed HH-orthogonality statements across parameter changes, and regionally affine gradients require further analysis.

Response to Q5.\textbf{Response to Q5.} We thank the reviewer for stating their concern about training-cost measurements and transferability of the proposed framework. In the revised version, we have addressed these issues by providing a training time comparison of the compared methods and adding a higher-dimensional EV-charging experiment. In particular, in Appendix J5, we report training wall-clock time on the IEEE 57-bus benchmark across training sample sizes. For example, at Ntrain=1500N_{\mathrm{train}}=1500, SIGNAL takes 9.252±0.3069.252\pm0.306 seconds, versus 7.701±0.2137.701\pm0.213 seconds for Dec-Feas and 13.116±0.44913.116\pm0.449 seconds for dQP-DFL. Thus, the regime supervision incurs additional training time over Dec-Feas, while SIGNAL is faster than dQP-DFL, as the main comparator capable of handling joint uncertainty by default. Furthermore, in Appendix J9, we provide a network-constrained EV-charging experiment. It schedules 1212 controlled chargers over eight periods, producing 9696 decision variables, 1,0241,024 polyhedral electrical-network inequalities, and 1212 energy-limit inequalities. We compare SIGNAL against the two principal comparators applicable directly to joint objective-constraint uncertainty: MSE-PtO and dQP-DFL. Across the reported conditions, SIGNAL attains lower mean operational regret and violation measures than MSE-PtO and dQP-DFL. This is evidence of transfer to a higher-dimensional problem with a different constraint structure.

We are grateful to the reviewer for their constructive feedback. We hope that the additional theoretical analysis, experimental evidence, and clarifications in the revised manuscript satisfactorily address their concerns and strengthen their assessment of our work.

Thanks for the response. The new baseline and the margin ablations do address my main concern about what the proposed supervision adds. Compared with slack-only, the full margin consistently reduces the share of test instances with a constraint violation, though the gains are modest and regret remains close in several settings. The larger EV experiment also broadens the evaluation of SIGNAL. What still limits the evidence is that the added value of the multipliers is directly supported by only one power-grid benchmark. Even so, the new evidence addresses my main concerns, and I am raising my score.

Reviewer_erXEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies decision-focused learning with jointly uncertain objective coefficients and constraint right-hand sides in strongly convex quadratic programs with fixed problem matrices. It proposes SIGNAL, a training loss combining Hessian-weighted decision discrepancy, normalized true-constraint violation penalties, and supervision of signed primal–dual margins constructed from constraint slacks and multipliers. The margin loss provides direct learning signals for inactive predicted bounds that are locally invisible to decision-based gradients.

The theoretical analysis characterizes local decision geometry, dormant-bound sensitivity, regional loss structure, and deterministic empirical bounds linking the loss to downstream metrics. Experiments on two IEEE-based dispatch benchmarks report improvements in operational regret and violation metrics under observation noise, increased load, and varying training-set sizes.

Critical strengths

The paper provides a useful account of learning signals under joint uncertainty. It clearly explains why inactive predicted bounds can receive no direct gradient from decision and feasibility losses, despite becoming consequential under a changed operating condition. Connecting this blind spot to primal–dual information is informative and motivates auxiliary supervision.

The reported empirical improvements are consistent. SIGNAL improves violation frequency and magnitude alongside operational regret across the tested high-load conditions. The repeated-run evaluation and nested ablations support the usefulness of adding feasibility shaping and regime supervision within the proposed framework.

Critical weaknesses

The distinctive contribution relative to constraint-activity supervision and existing optimization-layer differentiation is insufficiently established. The main algorithmic contribution appears to be a new training loss rather than a new optimization or gradient-computation mechanism. Much of the analysis specializes classical KKT sensitivity and standard inequalities to this loss. Although supervising a continuous slack–multiplier margin differs from predicting binary activity labels, the paper does not sufficiently position this design against existing active-set prediction or constraint-state supervision approaches. The theoretical results explain how the proposed loss provides additional gradients, but do not establish why this particular representation is preferable to simpler supervision.

The validation does not isolate the central design or establish its transferability beyond fixed-geometry QPs. Comparing Dec-Feas with SIGNAL adds direct supervision in previously decision-null bound directions, but does not distinguish this benefit from the benefit of the specific primal–dual encoding. A normalized bound-MSE or slack-supervision replacement using the same optimization-layer differentiation would be more informative than pure MSE-PtO or sequential baselines. Furthermore, the exact affine sensitivity and regional quadratic structure rely strongly on fixed linear constraints and a fixed positive-definite Hessian. The authors acknowledge this limitation, but the paper does not establish which parts of the approach survive nonlinear constraints or more general optimization problems. This limits the breadth of the contribution, particularly given the small reduced decision dimensions in the experiments.

Questions for authors

Novelty and mechanism clarification: How does SIGNAL differ from existing active-set prediction or constraint-activity supervision methods, beyond replacing activity labels with continuous slack–multiplier margins? Which benefit is attributable to the supervision representation, and which follows from established differentiation through an optimization layer?

Assessment-changing comparison: How does SIGNAL compare with Dec-Feas augmented by normalized bound MSE or slack-only supervision, using the same predictor, differentiation backend, and calibration protocol? These comparisons would help determine whether the multiplier component and signed-margin construction are responsible for the gains.

Scope clarification: Section 8 identifies broader optimization problems as future work. Without requiring a new experimental domain, can the authors clarify which aspects of the loss and gradient construction extend to smooth nonlinear constraints? In particular, which results would need modification when sensitivity depends on the Lagrangian Hessian and changing constraint Jacobians? Please distinguish extensions to nonlinear convex programs from general nonconvex problems.

3 comments

We sincerely thank the reviewer for their insightful comments and questions. Our revision incorporates a discussion about the related work on active-set prediction and constraint-activity learning (Appendix E), a clarification regarding the distinction with optimisation-layer differentiation in Section 2, a structural and empirical auxiliary-loss study (Appendices F, G, J8), and a discussion about the scope of our contributions (Appendix H).

Response to Q1.\textbf{Response to Q1.} We thank the reviewer for highlighting the need to clarify our contributions relative to differentiation layers, active-set prediction, and constraint-activity learning. In the revised version, we have addressed this issue by improving and extending the related work discussion. In particular, in Section 2, we clarify that differentiable optimisation determines how gradients propagate through the optimiser; however, our analysis determines which learning information those gradients reveal under joint objective-constraint uncertainty and accordingly motivates regime supervision that captures information beyond local decision sensitivity. In Appendix E, we cite five additional references on active-set prediction and constraint-activity learning, and point out that, in contrast to methods that predict active-set labels for solver replacement or constraint screening, SIGNAL predicts objective and constraint parameters while retaining the represented optimisation problem, whose solution implicitly determines a predicted active set, without requiring a separate active-set classifier. Moreover, SIGNAL learns constraint-activity information through continuous signed primal-dual margin supervision, which captures both activity disagreements and differences in slack or multiplier magnitudes.

Response to Q2.\textbf{Response to Q2.} We thank the reviewer for pointing to the need for the adoption of normalised bound-prediction and slack-only supervision within the same decision and feasibility framework. In the revised version, we have addressed this issue comprehensively by defining normalised-bound prediction, slack-only supervision, multiplier-only supervision, and sign-only supervision. In particular, in Appendices F and G, we analyse this distinction structurally for the four possible constraint-activity configurations. When a row is inactive in both true and represented problems, LregL_{\text{reg}} reduces exactly to slack-only supervision; if GiδzG_i\delta z is negligible, normalised-bound prediction provides an approximate parameter-space proxy. When a row is active in both problems, LregL_{\text{reg}} reduces exactly to multiplier-only supervision. In these stable one-sided regimes, the simpler continuous losses can therefore provide appropriate surrogates. The full signed margin is particularly distinctive when activity statuses differ. On a missed-active row, it combines the represented positive primal slack with the true dual multiplier; on a falsely active row, it combines the true positive primal slack with the represented dual multiplier. In both cases the two quantities enter with the same sign after taking the signed-margin difference. Sign-only supervision captures a different, coarser aspect of this structure. It assigns zero loss whenever the activity labels agree and a fixed loss of 1/k1/k whenever they disagree. Thus, it detects activity misclassification but discards continuous information about distance to activation, binding strength, and the severity of margin disagreement. Apart from this analysis, it is important to note that, for an inactive inequality, the primal slack si>0s_i>0 measures its distance to binding, equivalently the bound tightening required to reach activation. However, for a strictly active inequality, the dual multiplier μi>0\mu_i>0 quantifies its binding strength through the marginal objective-value improvement from bound relaxation; in the isolated-inequality setting, the relaxation required to release the constraint is di2μid_i^2\mu_i. Building on this observation, we show in Proposition 3 through a bound perturbation analysis that the signed primal-dual margin matters in all four constraint-activity configurations. Proposition 4 further establishes a deterministic certificate linking sufficiently small regime-supervision loss to exact active-set recovery on the supervised constraints. Finally, in Appendix J8, we compare the resulting variants (Dec-Feas-NMSE, Dec-Feas-Slack, Dec-Feas-Mult, and Dec-Feas-Sign) with SIGNAL, using the same predictor class, differentiable backend, componentwise clipping procedure, calibration protocol, and paired evaluation settings. This study evaluates five conditions: the OOD baseline, +2%+2\%/+4%+4\% relaxation of two training-exposed physical line ratings, and −2%-2\%/−4%-4\% tightening of two training-dormant physical line ratings. SIGNAL attains the lowest mean operational regret and violation magnitude in each condition. Dec-Feas-NMSE has a slightly lower mean violation frequency in the baseline and two relaxation conditions, but exhibits higher mean violation magnitude in all conditions. These metrics should be interpreted jointly, as lower violation frequency does not necessarily imply better overall performance. Dec-Feas-Slack regret is close to SIGNAL in several conditions. Nevertheless, the results suggest that combining slack and multiplier information in the full signed-margin formulation can be beneficial, as SIGNAL achieves lower violation frequency and magnitude than Dec-Feas-Slack across all evaluated conditions.

Response to Q3.\textbf{Response to Q3.} We thank the reviewer for stating their concern about the scope of the proposed framework. In the revised version, we have addressed this issue by providing a discussion about the scope beyond strongly convex QPs. More specifically, in Appendix H, we clarify that true-feasibility shaping extends naturally by evaluating predicted decisions against true constraints, while signed-margin supervision and its active-set certificate follow from primal-dual complementarity, provided corresponding constraints, compatible positive scales, and well-defined KKT-pair selections are available. In contrast, the fixed objective-curvature metric, canonical normalisation, and exact regional affine sensitivity structure rely on the strongly convex QP formulation with fixed problem matrices. We then discuss the extension challenges for different constrained optimisation problems. In particular, for LPs, the objective-induced Bregman divergence vanishes identically, requiring an alternative decision loss or strongly convex regularisation to retain meaningful decision supervision. Moreover, potential nonuniqueness of primal solutions and dual multipliers necessitates consistent selection rules or appropriate regularisation. In smooth nonlinear convex programmes, local KKT sensitivities are computed from \begin{align} \begin{bmatrix} W & J^\top \newline J&0 \end{bmatrix} \begin{bmatrix} dz \newline dy\end{bmatrix} =-\begin{bmatrix} \partial_\eta\nabla_z\mathcal L \newline \partial_\eta c_{\mathcal A} \end{bmatrix}d\eta, \end{align} where W=∇zz2LW=\nabla^2_{zz}\mathcal L and J=∇zcAJ=\nabla_z c_{\mathcal A}, under local regularity ensuring a differentiable KKT branch. Both WW and JJ now vary with the solution and parameters. Consequently, fixed tangent/normal geometry, exact affine finite-displacement sensitivities, regional convex-quadratic loss, and affine output gradients do not transfer unchanged. For a smooth strongly convex objective, the true-objective Bregman divergence remains nonnegative and controls decision displacement, but nonlinear slacks require appropriate local scaling rather than the QP's exact HH-metric distance. Finally, for general nonconvex programmes (not mentioned explicitly), complementarity can still provide a signed margin at a chosen KKT pair, but KKT points need not be global optima; the selected solution branch can be set-valued, unstable, or discontinuous; and the objective Bregman divergence need not be nonnegative. Thus, we claim neither global regret guarantees nor the present geometric identities in that setting.

We are grateful to the reviewer for their constructive feedback. We hope that the additional theoretical analysis, experimental evidence, and clarifications in the revised manuscript satisfactorily address their concerns and strengthen their assessment of our work.

Reviewer_tdzMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SIGNAL studies decision-focused learning with joint uncertainty in objective coefficients and constraint bounds. For strongly convex quadratic programs with fixed objective curvature and constraint matrices, it combines a decision loss, a true-feasibility loss, and slack–dual supervision intended to reach constraints that receive no local decision gradient. The paper develops a geometric analysis and evaluates the method on two reduced power-grid QP benchmarks.

Critical strengths
  1. The paper gives a clear account of why currently inactive constraints can receive no decision gradient despite becoming important under changing operating conditions. The distinction between decision error, true infeasibility, and constraint-regime information provides a coherent motivation for the loss design. The authors explicitly state the assumptions of the QP analysis and disclose the reduced benchmark construction and use of shifted development data. These limitations should be evaluated as restrictions on the contribution's scope rather than undisclosed flaws.
  2. The experiments include nested loss ablations and a joint-learning differentiable-QP baseline that already penalizes true constraint violations, alongside sequential adaptations of PEAR and other methods. The improvements therefore cannot be dismissed simply as a comparison against methods that never learn constraints or consider feasibility. Repeated runs and variations in observation noise and training data provide useful evidence within the chosen setting.
Critical weaknesses
  1. The distinctive benefit of slack–dual supervision remains insufficiently isolated. The comparison with Dec-Feas is relevant because both methods jointly train the objective and constraint predictors, but adding the proposed term also adds direct supervision of constraint information. For inactive constraints, that signal is closely related to normalized bound-prediction error. A simple constraint-prediction auxiliary loss within the same decision-and-feasibility framework is therefore an important missing comparator. The sequential PEAR comparison cannot resolve this question, since it also changes whether the constraint predictor is frozen.
  2. The submitted implementation introduces a further difference in the key ablation: SIGNAL clips gradients separately for each loss component before summing them, whereas Dec-Feas clips the gradient of the combined loss. These updates need not coincide when clipping is active. The empirical importance of this difference is unknown, but the current comparison does not isolate the auxiliary loss alone. A matched gradient-handling protocol is needed to support a stronger attribution of the improvement to the proposed loss.
  3. The applicability and empirical coverage are narrow. The established formulation relies on strong convexity and fixed objective curvature and constraint matrices, with uncertainty restricted to objective coefficients and constraint bounds. Both benchmarks are reduced power-grid QPs with a similar construction in which dormant constraints become relevant under higher load. The larger case does retain a substantial constraint system, and the authors acknowledge the restrictions, but noise and sample-size sweeps do not establish transfer across materially different joint-uncertainty problems. Evidence from another problem structure, even within the stated QP class, would strengthen the practical case.
Questions for authors
  1. Does slack–dual supervision improve on a normalized constraint-prediction auxiliary loss added to the same decision and feasibility losses? In which constraint regimes should the dual information provide a benefit beyond directly supervising the bounds?
  2. Do the gains over Dec-Feas persist when both methods use the same gradient-clipping protocol and comparable tuning? Please clarify the role of componentwise clipping in the reported ablation.
  3. Which parts of SIGNAL could extend to LPs or broader constrained optimization problems, and which fundamentally require a positive-definite quadratic objective and fixed constraint geometry? In particular, how would the decision divergence, normalization, and treatment of nonunique solutions change? Evidence on a structurally different problem within the current assumptions would also help distinguish a broadly useful mechanism from one validated mainly by the chosen activation-shift construction.
3 comments

We sincerely thank the reviewer for their insightful comments and questions. Our revision incorporates a structural and empirical auxiliary-loss study (Appendices F, G, J8), a clipping-controlled ablation (Appendix J7), a higher-dimensional EV-charging experiment (Appendix J9), and a discussion about the scope of our contributions (Appendix H).

Response to Q1.\textbf{Response to Q1.} We thank the reviewer for suggesting that a normalised bound-prediction loss within the same decision and feasibility framework is an appropriate control. In the revised version, we have addressed this issue comprehensively by defining normalised-bound prediction, slack-only supervision, multiplier-only supervision, and sign-only supervision. In particular, in Appendices F and G, we analyse this distinction structurally for the four possible constraint-activity configurations. When a row is inactive in both true and represented problems, LregL_{\text{reg}} reduces exactly to slack-only supervision; if GiδzG_i\delta z is negligible, normalised-bound prediction provides an approximate parameter-space proxy. When a row is active in both problems, LregL_{\text{reg}} reduces exactly to multiplier-only supervision. In these stable one-sided regimes, the simpler continuous losses can therefore provide appropriate surrogates. The full signed margin is particularly distinctive when activity statuses differ. On a missed-active row, it combines the represented positive primal slack with the true dual multiplier; on a falsely active row, it combines the true positive primal slack with the represented dual multiplier. In both cases the two quantities enter with the same sign after taking the signed-margin difference. Sign-only supervision captures a different, coarser aspect of this structure. It assigns zero loss whenever the activity labels agree and a fixed loss of 1/k1/k whenever they disagree. Thus, it detects activity misclassification but discards continuous information about distance to activation, binding strength, and the severity of margin disagreement. Apart from this analysis, it is important to note that, for an inactive inequality, the primal slack si>0s_i>0 measures its distance to binding, equivalently the bound tightening required to reach activation. However, for a strictly active inequality, the dual multiplier μi>0\mu_i>0 quantifies its binding strength through the marginal objective-value improvement from bound relaxation; in the isolated-inequality setting, the relaxation required to release the constraint is di2μid_i^2\mu_i. Building on this observation, we show in Proposition 3 through a bound perturbation analysis that the signed primal-dual margin matters in all four constraint-activity configurations. Proposition 4 further establishes a deterministic certificate linking sufficiently small regime-supervision loss to exact active-set recovery on the supervised constraints. Finally, in Appendix J8, we compare the resulting variants (Dec-Feas-NMSE, Dec-Feas-Slack, Dec-Feas-Mult, and Dec-Feas-Sign) with SIGNAL, using the same predictor class, differentiable backend, componentwise clipping procedure, calibration protocol, and paired evaluation settings. This study evaluates five conditions: the OOD baseline, +2%+2\%/+4%+4\% relaxation of two training-exposed physical line ratings, and −2%-2\%/−4%-4\% tightening of two training-dormant physical line ratings. SIGNAL attains the lowest mean operational regret and violation magnitude in each condition. Dec-Feas-NMSE has a slightly lower mean violation frequency in the baseline and two relaxation conditions, but exhibits higher mean violation magnitude in all conditions. These metrics should be interpreted jointly, as lower violation frequency does not necessarily imply better overall performance. Dec-Feas-Slack regret is close to SIGNAL in several conditions. Nevertheless, the results suggest that combining slack and multiplier information in the full signed-margin formulation can be beneficial, as SIGNAL achieves lower violation frequency and magnitude than Dec-Feas-Slack across all evaluated conditions.

Response to Q2.\textbf{Response to Q2.} We thank the reviewer for suggesting that clipping effects need to be assessed carefully. In the revised version, we have addressed this issue by introducing Dec-Feas (CW-Clip), which shares SIGNAL's componentwise clipping procedure while retaining the Dec-Feas (Default) training loss. In particular, in Appendix J7, we compare three methods: Dec-Feas (CW-Clip), Dec-Feas (Default), and SIGNAL. These methods share the same predictor class, differentiable backend, calibration protocol, and paired evaluation settings. The results show that the mean operational regret is 2.15±0.052.15\pm0.05 for Dec-Feas (CW-Clip), 2.36±0.132.36\pm0.13 for Dec-Feas (Default), and 0.95±0.060.95\pm0.06 for SIGNAL. Componentwise clipping improves Dec-Feas by 8.9%8.9\% in regret, confirming that it matters; however, SIGNAL still reduces regret by 55.8%55.8\% relative to the matched-clipping Dec-Feas (CW-Clip) comparator. The corresponding reductions in violation frequency and magnitude are 25.4%25.4\% and 57.5%57.5\%. Thus, clipping alone cannot explain the gain.

Response to Q3.\textbf{Response to Q3.} We thank the reviewer for stating their concern about transferability of the proposed framework. In the revised version, we have addressed this issue by adding a higher-dimensional EV-charging experiment and providing a discussion about the scope beyond strongly convex QPs. In particular, in Appendix J9, we provide a network-constrained EV-charging experiment. It schedules 1212 controlled chargers over eight periods, producing 9696 decision variables, 1,0241,024 polyhedral electrical-network inequalities, and 1212 energy-limit inequalities. We compare SIGNAL against the two principal comparators applicable directly to joint objective-constraint uncertainty: MSE-PtO and dQP-DFL. Across the reported conditions, SIGNAL attains lower mean operational regret and violation measures than MSE-PtO and dQP-DFL. This is evidence of transfer to a higher-dimensional problem with a different constraint structure. Furthermore, in Appendix H, we separate the underlying principles from QP-specific properties. More specifically, we clarify that true-feasibility shaping extends naturally by evaluating predicted decisions against true constraints, while signed-margin supervision and its active-set certificate follow from primal-dual complementarity, provided corresponding constraints, compatible positive scales, and well-defined KKT-pair selections are available. In contrast, the fixed objective-curvature metric, canonical normalisation, and exact regional affine sensitivity structure rely on the strongly convex QP formulation with fixed problem matrices. We then discuss the extension challenges for different constrained optimisation problems. In particular, for LPs, we highlight that the objective-induced Bregman divergence vanishes identically, requiring an strongly convex regularisation, and that potential nonuniqueness of primal solutions and dual multipliers necessitates consistent selection rules or appropriate regularisation.

We are grateful to the reviewer for their constructive feedback. We hope that the additional theoretical analysis, experimental evidence, and clarifications in the revised manuscript satisfactorily address their concerns and strengthen their assessment of our work.

Thank you for detailed responses. As I mentioned in my review, I liked the suggestion from the authors about inactive constraints and gradient information. I already gave a good score, so I will maintain mine.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.