acceptodds
Under review as a conference paper at ICLR 2027

DePICT: Decision-Preserving Interface for Constrained Downstream Tasks

Abstract

A constrained optimization problem may involve a parameter in its objective and in every active constraint, and yet its solution may be insensitive to small changes in that parameter. This raises a fundamental question: which inputs does a decision truly depend on? Building on this question we introduce DePICT, a procedure that constructs decision-preserving interfaces by ranking context directions with the optimizer's solution-sensitivity and aggregating them across an operating regime. We study this problem in a high-dimensional setting where primitive context parameterizes a constrained task and the downstream agent observes only a selected subset of context directions. For locally regular constrained programs, we derive a Karush–Kuhn–Tucker (KKT) based characterization of when a context direction is optimizer-relevant. Our analysis shows that appearing in the active optimization problem does not necessarily mean that a variable affects the final decision. Some context directions can change the KKT conditions while leaving the optimal solution unchanged because their effect is absorbed by the dual variables. DePICT is designed to remove exactly these directions. In a controlled diagnostis, it recovers the decision-relevant interface exactly and reduces linear-predictor regret from to . It also performs well on controlled KKT-regime and real solver-mediated tasks.

Primary area:
optimization
Keywords:
constrained optimization, decision-focused learning, feature selection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wwuTOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents DePICT, a sensitivity-based method for identifying the most relevant context variables that a downstream optimizer's decision actually depends on. The authors refer to subsets of context variables as an interface, and DePICT identifies a top-mm interface by ranking coordinates based on solution sensitivities computed through implicit differentiation of the KKT conditions (Theorem 1 and follow-on discussion). This relies on identifying the active set of constraints, and reducible coordinates can fall under two types: (1) the coordinate does not affect the active set (Layer 1), and (2) the coordinate affects the active set but does not change the optimal primal solution, as its effect can be absorbed into the dual variables via e.g. rescaling (Layer 2). The authors derive theoretical properties of irreducible directions (e.g. Theorem 7) and of their top-mm interface (Proposition 10). The method is benchmarked against baselines on knapsack, shortest path, and Warcraft problems, and a Fama-French (FF25) portfolio task, along with additional studies analyzing the stability and runtime cost.

Critical strengths

This work presents a sound theoretical framework for designing optimization problem interfaces as a subcomponent of learned decision-making systems. The analysis on the symbolic support vs. optimizer relevance is helpful and can be directly useful for a practitioner working with optimization embedded in learning pipelines through e.g. differentiable optimization layers. Example 4 in particular is a common scenario with learned constraints and provides a useful insight on the scale invariance of constraints with a particular linear structure.

Critical weaknesses

The main motivation for this method is somewhat weak in my opinion. The framework requires the model c↦y(c)c \mapsto y(c), but in order to design such a model, the designer already needs to know which constraints and corresponding context are important. In particular, if the active set of constraints is already known, then the designer can remove the inactive constraints from the model to get a smaller problem to solve. Yet a large benefit of using constrained optimization is that we don't necessarily know a priori which constraints will be binding at the solution, and solving the problem gives us that information. Moreover, small perturbations of the context may not affect the optimum, but large perturbations could still affect it. For example, in linear programs, changing a cost may not affect the solution until above a certain threshold, where the new solution then jumps to a different vertex. Thus using just local information might e.g. incorrectly decide to remove a sensor that was actually important. I think what is missing that would support the main motivation is a full application that demonstrates the actual cost savings of removing some portion of the context on some downstream task.

Edit for rebuttal: the authors' answers have addressed my concerns substantially, and the new experimental results provide a compelling real-world use case that demonstrates the benefits of the approach. I believe the work is strong and good fit for ICLR, and have raised my score from 3 →\to 4 accordingly.

Questions for authors

Can you provide an application that directly demonstrates a cost-savings vs. decision quality comparison for DePICT?

4 comments

We sincerely thank the reviewer for the positive assessment, and for noting that Example 4 reflects a common scenario with learned constraints.

Table R1. Full-input pipeline → DePICT pipeline, under the same decision-focused training (regret is the mean over 10 seeds).

Task / DFL methodInputsRegret ↓, LinearRegret ↓, MLPProjected monthly cost
Knapsack / QPTL200 → 604.029 → 1.7503.029 → 2.855$2.00 → $1.52 (−24%)
Shortest path / QPTL480 → 1603.046 → 1.9932.589 → 2.119$2.39 → $1.55 (−35%)
Shortest path / SPO+480 → 1601.795 → 0.7471.796 → 1.021$2.43 → $1.45 (−40%)

[W1] We strengthened the motivation in two ways. The introduction now cites work on the cost of sensing, communication and maintaining inputs (Sec. 1, lines 37–42). A new study shows cost savings together with decision quality (Sec. 3.6, lines 480–515; Table R1). Moreover, DePICT requires only the objective and the constraints of the problem (Sec. 1, lines 55–57; Eq. 1, lines 83–86). The model describes how each candidate input enters the problem, and DePICT finds which constraints and which context are important for moving the constrained optimum. DePICT does not assume that the active set is known. Removing inactive constraints is also not enough: a variable can appear in an active constraint and still be irrelevant (Example 4, Sec. 2.3, lines 186–202). Several existing selectors keep such variables (Sec. 3.3, lines 417–419). Our analysis is first-order, so we focus on local rather than global guarantees, and the paper states this as a limitation (Sec. 4, lines 527–528). The linear-program case is discussed in Sec. 2.3 (lines 229–233): a cost change does not move the solution while the active basis remains unchanged. DePICT combines the sensitivities from many representative contexts (Algorithm 1, lines 270–281), so an input that matters somewhere in the operating regime is kept (Sec. 2.1, lines 107–110). A stability analysis under perturbations is reported in Sec. 3.4. Reviewer BCyp, also emphasized missing experiments about cost saving and practical usage, hence we also added a study with a cost analysis of the full pipeline on two downstream tasks (Sec. 3.6, lines 480–515; Figure 6). Removing about two-thirds of the context lowers storage by 67–70% and the projected cost by 24–40%, and decisions are as good or better (Table R1). The tasks are controlled benchmarks, and the costs are projections under an assumed AWS deployment (App. D.1.19, lines 2666–2667).

[Q1] Sec. 3.6 (lines 480–515) and Figure 6 give the cost-savings vs. decision quality comparison for DePICT, and Table R1 shows the cost savings and the decision quality side by side. Full details are in Table 8 (App. D.1.19, line 2681).

Thank you again for your detailed review. As the discussion closes on 18 November, we have uploaded a revised PDF with the new cost vs. decision quality study (Sec. 3.6, Figure 6), added references on sensing and communication costs, the clarification that DePICT does not need the active set in advance, and the first-order scope stated as a limitation.

The other reviewers also recognized the value of the work. Reviewer BCyp noted that “the controlled diagnostics provide convincing examples” and that the paper “evaluates actual decision outcomes rather than prediction accuracy alone”, and after our response wrote “I do recognize that there would be an advantage of your method. I will raise my score.” Reviewer bamJ found “the question is clear and relevant” and that Example 4 explains “clearly why appearing in the active KKT system does not imply relevance.”

If our responses address your concerns, we would be grateful if you would consider adjusting your score.

Thank you and apologies for the late response. I appreciate your detailed answers, which have clarified the majority of my concerns regarding the motivation. Overall, I find the methodology sound and the new results are compelling and demonstrate the practical advantages of the method. I believe the work is strong and good fit for ICLR. I am willing to raise my score from 3 →\to 4.

We are delighted to know that the revisions and clarifications have adequately addressed your concerns. Your positive assessment meant a great deal to us. We appreciate the time and consideration you have devoted to reviewing our manuscript.

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

The authors propose DePICT, which constructs a fixed interface before predictor training by ranking context coordinates according to the sensitivity of the optimal solution. Using KKT sensitivity analysis, the paper characterizes optimizer relevance through two channels: changes in active constraint margins and changes in stationarity. It shows that variables appearing in the active KKT system may still be irrelevant to the optimal decision if their effects can be absorbed by the Lagrange multipliers. The paper further discusses the first-order consequences of removing decision-relevant variables and the stability of interface rankings.

Critical strengths
  • The question is clear and relevant: decision relevance versus predictive relevance, with the interface fixed before predictor training.
  • The two-channel decomposition (CA,GA)(C_{\mathcal{A}}, G_{\mathcal{A}}) and Example 4 explain clearly why appearing in the active KKT system does not imply relevance.
  • The authors state the first-order, fixed-active-set limitations and present Warcraft/FF25 as transfer tests rather than validations of the theory.
Critical weaknesses
  • Table 1's reported numbers are inconsistent considering \mboxF1/\mboxPrec.\mbox{F1}/\mbox{Prec.}.
  • Oracle support does not match the definition of irreducibility. Irreducibility is defined in Section 2.1. The oracle labels in App. D.1.3/D.1.4 are based on parameter dependence, which do not necessarily imply nonzero optimizer sensitivity.
  • Full-context information bypasses the knapsack interface (Eq. 42), so the complete decision pipeline is not depend on only the selected interface.
  • Regularity and numerical sensitivity are not fully established.
  • DePICT takes direct KKT/Jacobian computing, whereas the reported finite difference baselines perturb a task proxy.
  • Proposition 10(i), stated under Assumption 1 only, is false: under C2C^2 data, JJ is continuous but not necessarily locally Lipschitz.
  • Theorem 1 follows from classical KKT sensitivity analysis, Theorem 7 follows from chain rule, and the Jacobian is standard in differentiable optimization.
Questions for authors
  • Could the authors justify the local Lipschitz continuity of J(c)J(c) under Assumption 1? The implicit function theorem guarantees a continuously differentiable solution map under the stated assumptions, but this does not generally imply a locally Lipschitz Jacobian. Is an additional smoothness assumption required?

  • In Eq. (42), the optimizer still uses w(c)w(c) and B(c)B(c) computed from the full context. How does this agree with the proposed decision-preserving interface and the information restriction considered in Theorem 7?

4 comments

We sincerely thank the reviewer for the careful and precise review, and for recognizing the clarity of our question and the two-channel decomposition.

[W1] We have updated the support metrics and the corresponding descriptions in the table to reflect budget-optimal rather than exact full-support recovery.

[W2] Oracle support and irreducibility are different by definition, and the paper states how they relate: Corollary 3 shows that support only upper-bounds irreducibility (Sec. 2.3, lines 166–168). An input is irreducible when changing it moves the optimal decision (Sec. 2.1, lines 103–107). Oracle labels mark which inputs enter the task parameters. The paper calls them construction-level labels (App. D.1.6, line 2051). Irreducibility itself is tested in the Layer-2 diagnostic, where the labels are defined by the KKT structure (App. D.1.16, lines 2475–2483). There, DePICT keeps every irreducible input and no nuisance input (Sec. 3.3, lines 416–417; Table 4).

[W3] The experiments follow the standard setting of Mandi et al. (2020): the selected interface is used to predict the item values, and the solver receives the true weights and capacity. We state this explicitly (App. D.1.5, lines 2022–2028). Shortest path and Warcraft are not affected: their constraints do not depend on the context, so the complete decision pipeline depends only on the selected interface.

[W4] Regularity is an explicit assumption of the theory (Assumption 1, Sec. 2.1, lines 91–94; App. C.1, lines 876–882). The paper states that its guarantees hold only under this assumption (Sec. 4, lines 524–526). Numerical sensitivity is handled in two ways. A threshold removes numerical noise from the scores (App. B, lines 805–806). The stability study shows that the selected interface stays almost the same under perturbations, with overlap above 0.979 (Sec. 3.4, lines 459–460).

[W5] DePICT uses the KKT system. The FD baseline is a standard perturbation method that does not use it (App. D.0.5, lines 1586–1591); it represents selectors that have no access to the solver's structure. Finite differences of the solution itself would estimate the same Jacobian as DePICT, but they need one extra solve per input, and FD cost grows from 37 s to 373 s as the inputs grow (Sec. 3.5, lines 471–472).

[W6] We updated the Assumption 1, which now requires C3C^3 data (Sec. 2.1, lines 91–94), and the proof of Lemma 2 is also updated (App. C.3.1, lines 1188–1194). Proposition 10(i) (line 294) holds under the revised Assumption 1. No experimental result changes, because all maps in our experiments are C∞C^\infty.

[W7] KKT sensitivity analysis, the chain rule and the Jacobian are classical tools, and the paper says so (App. A, lines 727–737). Our contributions (Sec. 1, lines 68–78) are what we do with them.

[Q1] We answer Q1 in W6.

[Q2] The experiments follow the standard setting of Mandi et al. (2020), now stated explicitly (App. D.1.5, lines 2022–2028). The interface governs the predicted item values, and w(c)w(c) and B(c)B(c) are known parameters of the solver. Theorem 7 extends to this case: the decision cannot respond to any input that is missing from both the interface and the known parameters (Sec. 2.4, lines 261–264). In Sec. 3.6 (lines 485–509), ww and BB are also predicted from the interface, so the information restriction of Theorem 7 holds exactly.

Thank you again for your detailed review. As the discussion closes on 18 November, we have uploaded a revised PDF with the strengthened smoothness assumption (C3C^3) and the corrected proof of Lemma 2, the explicit knapsack protocol, the new remark extending Theorem 7 to solvers with known parameters, the new experiment in which the decision depends only on the selected interface (Sec. 3.6), and the updated support metrics in Table 1.

The other reviewers also recognized the value of the work. Reviewer wwuT wrote that the paper "presents a sound theoretical framework for designing optimization problem interfaces" and that the analysis "is helpful and can be directly useful for a practitioner". Reviewer BCyp noted that "the controlled diagnostics provide convincing examples", and after our response wrote "I do recognize that there would be an advantage of your method. I will raise my score."

If our responses address your concerns, we would be grateful if you would consider adjusting your score.

Thank you for the detailed responses and revisions. Some concerns remain. In particular, the main theoretical ingredients are based on classical KKT sensitivity analysis and the chain rule, so the methodological novelty remains moderate.

Thank you for the reply. We addressed all of your Questions and Weakness with our extensive experiments in our response, could you be more specific which of them have not been addressed??

We agree that KKT sensitivity analysis and the chain rule are classical tools (App. A, lines 727–737). The contribution of the paper is what we ask and what we find with them. Below we state each part and what is new in it.

1. A new problem. We ask which inputs a constrained decision needs, before any data is collected or any model is trained (Sec. 2.1, lines 103–107). To our knowledge, this question has not been studied for constrained optimization.

2. An exact test (Theorem 1, lines 126–133). Classical theory gives a formula for the sensitivity J. We give a necessary and sufficient condition for when an input has no effect on the decision.

3. A new finding (Sec. 2.3; Example 4, lines 186–202). An input can appear in the active constraints and still have no effect on the decision. We also show when this happens and when it cannot (Corollary 3, lines 166–168). Several existing selection methods keep 50–70% of such inputs (Sec. 3.3, lines 417–419).

4. A new consequence (Corollary 8, lines 255–259). If a needed input is removed, the resulting error cannot be fixed by more data or better training.

5. A stability guarantee (Proposition 10, line 294). We give a radius within which the selected inputs do not change.

6. A practical method (Algorithm 1, lines 270–281). DePICT needs no labels and no trained model. Its runtime stays near 5 seconds when the number of inputs grows from 200 to 2,000 (Sec. 3.5).

7. Evidence. On four benchmarks, DePICT matches or beats seven selection methods. With decision-focused training, it has the lowest regret in all six settings while using one-third of the inputs (Sec. 3.6, Figure 6).

Reviewer_BCypOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DePICT selects input variables using the sensitivity of the downstream optimal decision. Its goal is to preserve decision-relevant information while reducing the input interface, rather than to replace decision-focused training. The paper characterizes locally irrelevant directions through KKT analysis and evaluates the selected inputs using prediction-trained models followed by downstream optimization. Controlled diagnostics and real-data experiments assess support recovery and decision quality.

Critical strengths
  1. The distinction between variables that appear in an optimization problem and variables that actually affect its solution is useful. The controlled diagnostics provide convincing examples of this distinction and connect the removal of irrelevant directions to improved downstream regret.
  2. The feature-selection comparisons use matched downstream predictors and optimization pipelines, and the paper evaluates actual decision outcomes rather than prediction accuracy alone. Experiments with different predictor families and repeated runs provide useful evidence for input selection within the evaluated pipelines.
Critical weaknesses
  1. The practical need for reducing the input interface is not sufficiently established in the evaluated settings. Sensing, communication, and processing costs motivate the paper, but the experiments mainly compare selectors under a prescribed feature budget. They do not show that input acquisition or processing is a material bottleneck in these tasks, or quantify an end-to-end cost reduction that justifies discarding information. Selection runtime alone does not establish this benefit. Consequently, the experiments demonstrate how to select under a budget more clearly than why the proposed reduction is practically needed.
  2. The additional value relative to decision-focused training remains unresolved. The downstream predictors are trained with prediction loss, so the results do not establish whether the selected inputs improve the performance–cost tradeoff over full-input DFL or DFL with an explicit feature budget. I recognize that DePICT is intended to complement rather than replace DFL. Nevertheless, this comparison matters to its practical motivation: if using the full input is inexpensive, decision-focused training is a natural alternative that may already reduce the influence of irrelevant information. The current experiments do not establish when explicit structural selection provides a meaningful additional benefit.

=> The authors somewhat addressed my main concern. While I still am not certain that this method is super useful, but I do see some value of this approach. I am raising my score.

Questions for authors
  1. In the evaluated applications, which input-related cost is sufficiently significant to justify reducing the interface? Can the existing evidence connect feature reduction to a meaningful end-to-end cost saving while preserving decision quality, rather than only to a smaller input count or faster selection?
  2. Does DePICT provide an additional benefit when the downstream predictor is trained for decision quality? A matched comparison of full-input DFL, DePICT followed by DFL, and a decision-focused selection method under the same feature budget would clarify the practical role of the structural criterion.
3 comments

We sincerely thank the reviewer for the thoughtful comments.

Table R1. Decision regret, mean ± SD over 10 matched seeds (lower is better).

TaskDFL methodPredictorFull + DFLDePICT + DFLBudgeted DFL
KnapsackQPTLLinear4.029 ± 0.2381.750 ± 0.2642.138 ± 0.283
KnapsackQPTLMLP3.029 ± 0.7532.855 ± 0.7173.826 ± 0.545
Shortest pathQPTLLinear3.046 ± 0.1641.993 ± 0.1413.073 ± 0.154
Shortest pathQPTLMLP2.589 ± 0.2312.119 ± 0.2313.081 ± 0.281
Shortest pathSPO+Linear1.795 ± 0.0840.747 ± 0.0691.597 ± 0.282
Shortest pathSPO+MLP1.796 ± 0.0731.021 ± 0.0462.178 ± 0.162

Table R2. Full-input pipeline → DePICT pipeline (linear predictor).

Task / DFL methodInputsStorage per sampleRegret ↓Projected monthly cost
Knapsack / QPTL200 → 60−70%4.029 → 1.750$2.00 → $1.52 (−24%)
Shortest path / QPTL480 → 160−67%3.046 → 1.993$2.39 → $1.55 (−35%)
Shortest path / SPO+480 → 160−67%1.795 → 0.747$2.43 → $1.45 (−40%)

[W1] We now establish the need for reducing the input interface in two ways: with evidence from the literature (Sec. 1, lines 37–42) and with a new cost experiment (Sec. 3.6, lines 480–515; Figure 6). Each cost is now supported. Sensing: Joshi & Boyd (2008), Krause et al. (2008) and Janisch et al. (2019). Communication: Shao et al. (2021), cited in the same lines. Processing: we now measure it directly (Table R2; Figure 6, row 2). We now quantify the cost reduction. Storage falls by 67–70% and the projected operating cost falls by 24–40% (Table R2). The discarded information was not needed: regret is lower after the reduction (Table R2). Input acquisition is not measured, as it is outside the scope of this paper, but it is supported by the cited work. The practical need is shown using storage, training time and projected operating cost (Tables 6–8, App. D.1.19, lines 2659–2681). The new experiment shows why the reduction is needed: with one-third of the inputs, decisions are better and the cost is lower (Table R2).

[W2] The additional value relative to DFL is now resolved. Sec. 3.6 (lines 480–515) trains every pipeline with decision-focused learning, and DePICT adds value in every setting (Table R1). The new experiment uses decision-focused training (QPTL and SPO+), not prediction loss. DePICT + DFL has the lowest regret in all settings, with 30–33% of the inputs (Table R1) and a lower cost (Table R2). The results confirm that DePICT is complementary to DFL: DePICT + DFL performs better than DFL alone (Table R1), and the paper also states that the two are complementary (Sec. 4, lines 524–525). Decision-focused training on the full input does not reduce the influence of irrelevant information enough: Full + DFL has higher regret than DePICT + DFL in every setting (Table R1). Tables R1 and R2 now establish that structural selection provides additional benefit.

[Q1] Reviewer wwuT, also emphasized missing experiments about cost saving and practical usage. The new experiments justify input-related cost: storage falls by 67–70%, and the projected operating cost falls by 24–40% (Table R2). Table R2 also shows the cost saving and the decision quality together: the cost falls and regret improves in every row. The costs are projections under an assumed AWS deployment (App. D.1.19).

[Q2] DePICT provides additional benefit as DePICT + DFL has the lowest regret (Table R1; Figure 6). We ran exactly this comparison: full-input DFL, DePICT followed by DFL, and a decision-focused selection method (Budgeted DFL), with the same feature budget, the same training and 10 matched seeds (Sec. 3.6, lines 480–510). DePICT followed by DFL is the best of the three (Table R1; paired confidence intervals in Table 5, App. D.1.18, line 2633). The role of the structural criterion is clear: decision-focused training decides how to use the inputs, and DePICT decides which inputs are needed.

We added the new material is in Sec. 1 (lines 37–42), Sec. 3.6 (lines 480–515), Figure 6, and App. D.1.18–D.1.19 to support our claim. The content of table R1, R2 are included in the final version.

Thanks. My main concern was W1, and I do recognize that there would be an advantage of your method. I will raise my score.

Thank you for in detail analysis of our response and for raising your score. Your comments on the practical need for the method also mentioned by reviewer wwuT led to the new study in Sec. 3.6, and the paper is stronger for it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.