{"id":"ec91b66c-7b5e-403e-925f-2a2f1be33c30","arxiv_id":"2607.24007","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Legendre-regularized policies—decisions defined as solutions of strongly convex regularized optimization problems—are feasible, smooth, surjective onto the relative interior, and universal-approximating, and beat decision-rule baselines in newsvendor and resource-allocation tests.","lead":"This paper trains decision policies that always respect hard constraints by making the policy's output the solution of a regularized optimization problem, giving smooth gradients for end-to-end training. Relevant for anyone building learned systems for inventory, resource allocation, or other constrained decisions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The practical claim rests on an untested computational premise: for a general convex S, evaluating and differentiating an admissible Legendre regularizer may be intractable, while all experiments exploit separability or very low-dimensional coupling.","rationale":"The theoretical core (Theorem 1, Proposition 2) is mathematically sound under Definition 1; the Fenchel–Legendre duality, the Lipschitz bounds, and the universal approximation construction all check out. The load-bearing risk is not a proof error but an unverified premise: that for feasible sets arising in practice one can instantiate an admissible Legendre regularizer and evaluate its conjugate Hessian cheaply enough for gradient-based training. The reader's weakest assumption already identifies this, and §4 concedes scalability is open. My stress-test reinforces this and proposes a concrete scaling experiment. The verdict remains CONDITIONAL, unchanged from the reader.","tokens_in":21020,"tokens_out":28386,"duration_ms":272748,"concrete_test":"Implement LRP-Log for a random bounded polytope S={w: Aw=b, Cw≥d} with n=100 variables and m=200 constraints, using the damped Newton iteration described in §A.4. Train on a synthetic contextual optimization problem for 50 steps and measure mean per-sample forward/backward pass time and numerical stability. Compare against an MLP followed by a differentiable Euclidean projection layer with analytic Jacobian under the same tuning protocol. If LRP-Log's per-pass time exceeds the projection baseline by an order of magnitude or fails to converge, the practical achievability claim for general constraints is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Legendre-regularized policies are simultaneously smooth, hard-feasible, expressive, and trainable—is conditional on Definition 1: one must produce a μ-strongly convex, essentially smooth φ with twice-differentiable conjugate, and then evaluate w_{F,φ}(z) and its Jacobian per forward pass. Theorem 1 and Proposition 2 both presuppose this. The examples in §2.3 cover polytopes via log/entropy barriers, but the experiments in §3 and §A.4 use only boxes and a two-group budget polytope, where coordinatewise closed forms or a 2-dimensional Newton solve exist. For a general polyhedral S with, say, n=100 and m=200 constraints, the log-barrier optimizer requires an interior-point/Newton solve on every forward pass and an additional linear solve for the vector–Jacobian product; the cost and numerical stability of doing this inside a training loop are not analyzed or demonstrated. The authors themselves list 'scalable algorithms for evaluating and differentiating Legendre-regularized optimization layers' as open (§4). Thus the claim that the policy class is broadly achievable in practice is not supported beyond the tested low-dimensional/separable cases; this is the load-bearing gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Legendre-regularized policies (LRPs) for contextual decision-making under hard feasibility constraints. A policy is defined as π_g(x)=w_{F,φ}(g(x)), where w_{F,φ}(z) is the minimizer of ⟨Fz,w⟩+φ(w) over the feasible set S and φ is an 'admissible Legendre regularizer.' The main theoretical results are: Theorem 1 gives single-valuedness, containment in the relative interior, an explicit Jacobian, Lipschitz continuity, and surjectivity; Proposition 1 establishes a converse necessity; Proposition 2 proves uniform universal approximation on compact contexts. The paper also identifies log-barrier, entropic, and Gaussian-perturbation regularizers as instances and reports experiments on a newsvendor problem and two resource-allocation problems with box or group-budget constraints.","tokens_in":21332,"tokens_out":22173,"duration_ms":226278,"significance":"The theoretical framework is elegant and largely correct. The affine-hull reduction and the duality expression w = w̄ + N∇ψ*(-Bz) make the policy analytically tractable, and the explicit Jacobian is a concrete contribution for end-to-end training. The universality result is reassuring: smoothing does not come at the cost of expressiveness. The necessity result cleanly distinguishes Legendre regularization from projection-based or quadratic smoothing. The experiments are honest, including settings where the SAA benchmark wins, and the improvements over PADR/DD-SAA are meaningful in the tested regimes. The main caveat is that the computational premise—availability and efficient evaluation of an admissible regularizer—is only demonstrated for separable boxes and one low-dimensional coupled polytope.","major_comments":[{"comment":"The proof of Theorem 1 and the sentence 'Since φ has its effective domain contained in S' rely on a domain condition that is not stated in Definition 1. Definition 1 only says S°=(dom φ)°. For a closed convex S with nonempty relative interior, equality of relative interiors together with convexity does imply dom φ⊆S, but this implication is not shown; and if '°' is read as ordinary interior (as in the full-dimensional discussion), the condition is incompatible with lower-dimensional S. Please state explicitly that dom φ⊆S (or cl(dom φ)=S) and clarify that '°' denotes relative interior throughout. This is load-bearing because without such a condition the minimization in (2) is not restricted to S, so feasibility by construction is not guaranteed.","section":"§2.1, Definition 1; §2.2, Eq. (2)"},{"comment":"The practical claim that LRPs are broadly trainable for hard-constrained problems goes beyond the evidence. All experiments use either coordinatewise separable boxes, where the maps are closed-form, or one 20-dimensional coupled set with two group-budget inequalities, where the dual solve is two-dimensional. For a general polyhedral S with many coupled inequalities, evaluating w_{F,φ} requires a convex optimization of the same scale as the original decision problem, and differentiating it requires an additional KKT linear solve; no complexity or numerical-stability analysis is given, and §4 lists 'scalable algorithms' as an open direction. This does not invalidate Theorem 1 or Proposition 2, but it is a load-bearing gap for the paper's practical motivation. I recommend either adding a larger coupled experiment or explicitly narrowing the scope of the practical claim and stating the compu","section":"§3.3–3.4, §A.4, §4"}],"minor_comments":[{"comment":"The notation S° and (dom φ)° should be defined as relative interiors in one place; the current text alternates between 'relative interior' and 'interior' in the full-dimensional discussion.","section":"§2.1, Definition 1"},{"comment":"The claim that under the assumptions of Berthet et al. (2020) φ_ptb is strongly convex, Legendre-type, and has twice-differentiable conjugate would benefit from a precise reference (e.g., proposition number) and a short argument, since not every twice-differentiable expected maximum has a strongly convex conjugate.","section":"§2.3.3"},{"comment":"The projection step bπ_a(x) used for non-LRP baselines is mentioned but not described in full (e.g., the projection problem is assumed to be cheap); state whether this projection is computed exactly for all methods.","section":"§3.1"},{"comment":"Standard deviations are reported but no significance tests or paired comparisons. In several rows the LRP advantage over PADR is modest relative to spread; a brief statement on statistical significance would improve interpretation.","section":"Tables 1–3"},{"comment":"For the group-budget LRP-Log and LRP-Ent implementations, the derivation of the dual Newton iteration and the differentiation of the KKT system is sketched only verbally; a short pseudocode or derivation would help reproducibility.","section":"Appendix A.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's theoretical core is solid and the results are likely to be of interest to the math.OC/decision-focused learning community. My main concern is the mismatch between the breadth of the practical claim and the narrow set of feasible regions tested. If the authors tighten Definition 1 and either add a more demanding coupled experiment or temper the scope claims, I would support publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe punchline: the theory is solid and the paper deserves a serious referee, but the practical claim is wider than the evidence. The reader's CONDITIONAL verdict is about right; I'd push back a little on calling the mechanism 'mostly re-derived' — Prop 1 and Prop 2 are real additions.\n\nWhat's actually new: the necessity converse (Prop 1) and the universal approximation preservation (Prop 2) are not in the cited Fenchel–Young or perturbed-optimizer literature. The clean representation w = w̄ + N∇ψ*(−Bz) with explicit Jacobian is a nice packaging of classical Legendre duality, but the paper earns credit for the precise conditions and the bijectivity/surjectivity onto the relative interior. The proof sketch checks out; the Lipschitz bound and the inverse map are fine. The paper is also honest about the open scalability problem in §4, which is more than many papers do.\n\nSoft spots, in proportion: the stress-test note is right that Definition 1's existence is not guaranteed for general convex S. All experiments use boxes or a two-group budget polytope where closed forms or a 2-D Newton solve exist. For a general polytope with dozens of constraints, the per-forward-pass solve is an interior-point method, and the authors don't analyze cost or stability. That is a real gap, and they admit it. The experiments also lack the projection-based baseline that would isolate whether Legendre smoothing is what drives the gains; the linear-resource setting where SAA ties or wins weakens the headline. No significance tests, no released code/data. None of this undermines the theorems; it limits the empirical claims.\n\nWho should read it: anyone building differentiable optimization layers or decision-focused policies; the theory will be useful even if large-scale implementation is incomplete.\n\nRecommendation: send it to serious peer review. A competent referee can check the proofs quickly, and the novelty of Prop 1/Prop 2 deserves a shot. I'd ask the authors for a projection baseline, significance testing on the margins, and at least one non-separable polytope experiment before acceptance.","headline":"Solid theory with a narrower practical reach than claimed; worth peer review, with experiments needing a projection baseline and significance testing.","tokens_in":21966,"tokens_out":1748,"would_cite":true,"duration_ms":18523,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","90C15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A regularized optimization wrapper turns hard-constrained decision policies into smooth, differentiable maps without losing expressive power.","keywords":["decision-focused learning","contextual optimization","Legendre-regularized policies","hard constraints","differentiable optimization layers","universal approximation","regularized optimizers","prescriptive analytics"],"falsifier":"Look for any convex feasible set S and non-Legendre regularizer whose solution map is single-valued, everywhere differentiable, Lipschitz, and bijective onto the relative interior; Proposition 1 says none exists. The natural suspect, a quadratic penalty on a box (Euclidean projection), fails at boundary saturation and differentiability, so a valid counterexample would immediately refute the necessity result.","tokens_in":20788,"feed_emoji":"🧮","tokens_out":6520,"duration_ms":68266,"temperature":0.7,"pith_summary":"Decision-focused learning faces a three-way tension: policies must be expressive, must respect hard feasibility constraints, and must be smooth enough for gradient training. This paper proposes Legendre-regularized policies, where a learned latent vector is fed into a regularized optimization problem over the original feasible region and the solution is the decision. The paper establishes that, when the regularizer is Legendre-type (strongly convex, essentially smooth, with twice-differentiable conjugate), the resulting solution map is single-valued, maps onto the relative interior of the feasible set, has an explicit Jacobian, is Lipschitz continuous, and can be made arbitrarily smooth. It also proves that these Legendre conditions are necessary for those properties, and that composing the map with a universal approximator yields uniform approximation of any continuous feasible policy on compact contexts. If correct, this resolves the trilemma and unifies explicit regularizers like log-barrier and entropy with perturbation-based differentiable optimizers; experiments on newsvendor and resource allocation report lower decision costs than the benchmarks.","feed_headline":"Regularized optimizer maps make constrained policies smooth and trainable","feed_subtitle":"The solution map is single-valued, covers the feasible interior, and is Lipschitz — so no projection step is needed.","key_machinery":"The central object is the admissible Legendre regularizer φ (Definition 1): a proper lower-semicontinuous convex function whose effective domain has relative interior equal to S°, and whose reduction ψ(y) = φ(w̄ + Ny) on the affine hull is μ-strongly convex, essentially smooth, and has twice-differentiable convex conjugate. The workhorse identity is w_{F,φ}(z) = w̄ + N∇ψ*(−Bz), derived from the first-order condition Bz + ∇ψ(y) = 0; it converts the optimization into the convex-analytic inversion ∇ψ* = (∇ψ)^{-1}. The Jacobian −N∇²ψ*(−Bz)B is the Hessian of the conjugate, which explains why smoothness of the conjugate transfers to the policy and why the Lipschitz constant is inherited from the","core_discovery":"At the center is the optimizer map w_{F,φ}(z) = argmin_{w∈S} {⟨Fz,w⟩ + φ(w)}. The paper claims that under Definition 1 (admissible Legendre regularizer) this map is a single-valued function from R^k onto the relative interior S°, with the explicit form w = w̄ + N ∇ψ*(−Bz) and Jacobian −N ∇²ψ*(−Bz)B, and it is Lipschitz with constant bounded by ||F||/μ. The same Legendre assumptions are shown necessary: any regularized optimizer that is uniquely defined, bijective onto the interior, Lipschitz, and everywhere differentiable forces the regularizer's reduction to be strongly convex, essentially smooth, and have a twice-differentiable conjugate. The closing theoretical step is a universal approxi","pith_inferences":["Editorial extension: the necessity result suggests a design checklist for any smooth-optimizer layer: if the layer must be single-valued, everywhere differentiable, bijective onto the feasible interior, and Lipschitz, its regularizer must be of Legendre type; quadratic penalties fail the essential-smoothness test.","The bijection between latent space and the relative interior points to a regularizer-induced coordinate system on the feasible region, which could be exploited for uncertainty quantification or calibrated decision sets.","A testable direction the paper leaves open is scaling: for general polyhedral feasible sets with no closed-form layer, a Newton-iteration layer with implicit differentiation would need to be benchmarked against the closed-form box and polytope cases to see whether the theoretical guarantees survive finite iterations.","Because the approximation theorem only needs a dense latent model, any dense function class (not only neural networks) could replace the predictor and retain the same uniform approximation guarantee."],"forward_implications":["Because smoothness lives in the decision map rather than a surrogate loss, Legendre-regularized policies can be trained on any subdifferentiable downstream loss, including nonlinear and piecewise-linear costs.","Policies are feasible by construction at every training step; no projection or repair layer is needed, and gradients exist everywhere rather than only almost everywhere.","Universal approximation transfers from the latent model to the policy class, so hard feasibility and smoothness do not cap expressive power.","Log-barrier, entropic, and perturbation-based smooth optimizers are instances of one construction, giving a unified view of explicit and implicit smoothing.","With smoother regularizers the policy becomes C^{k-1}, which enables sensitivity analysis and higher-order training methods, not just first-order backpropagation."],"fun_headline_variants":["Hard constraints, smooth gradients: Legendre-regularized policies","Single-valued optimizer maps make feasible policies trainable","Smooth, feasible policies via Legendre-regularized optimizers","Explicit Jacobians give hard-constrained policies smooth learning"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on being able to construct an admissible Legendre regularizer — a μ-strongly convex, essentially smooth function with twice-differentiable conjugate — for the specific feasible set, and to evaluate its solution map and Jacobian cheaply; the paper demonstrates such regularizers for boxes and grouped-budget polytopes but gives no general recipe for arbitrary convex feasible sets.","fun_headline_variants_meta":{"raw":{"variants":["Hard constraints, smooth gradients: Legendre-regularized policies","Single-valued optimizer maps make feasible policies trainable","Smooth, feasible policies via Legendre-regularized optimizers","Explicit Jacobians give hard-constrained policies smooth learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1289,"prompt_tokens":746,"completion_tokens":543,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":475}},"tokens_in":490,"tokens_out":543,"duration_ms":6400,"temperature":1.0,"reasoning_tokens":475,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:19:23.636819+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Look for any convex feasible set S and non-Legendre regularizer whose solution map is single-valued, everywhere differentiable, Lipschitz, and bijective onto the relative interior; Proposition 1 says none exists. The natural suspect, a quadratic penalty on a box (Euclidean projection), fails at boundary saturation and differentiability, so a valid counterexample would immediately refute the necessity result.","supporting_citations":[],"review_version":1}