PaperlayerView exact revision on arXiv

Source-linked AI summary

Geometric Identification in Predict-Then-Optimize Learning

Jiaxiao Xu, Changhong Mou, Keji Liu, Dinghua Xu, Yeyu Zhang

arXiv:2609.33472v1cs.LGmath.OC

TL;DR

Decision-focused surrogate risk can preserve optimal downstream decisions without identifying the predictive cost report, even up to invisible directions. The paper characterizes the population SPO+ equality set and its geometric crossing conditions; under symmetry, these yield exact quotient-report identification, while contextual and application studies remain within stated population and evaluation limits.

  • Problem

    Decision quality alone cannot determine whether a predictive report is identified, and the conditions under which zero SPO+ excess risk identifies a quotient report are not characterized by decision consistency alone.

    Source: §1 · pp. 1–2

  • Method

    The paper characterizes the SPO+ population equality set through optimal-face crossing, then studies oracle separation, coercivity, symmetry boundaries, and conditional extensions.

    Source: §1 · p. 2

  • Results

    Under central symmetry, the mean class is the unique quotient Bayes report exactly when every nonzero effective displacement yields optimal-face departure with positive probability.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The identification target is the effective quotient report, not merely a downstream-optimal action; strict crossing alone does not identify the mean without symmetry.

    Source: §1 · pp. 1–2

    These are population results and do not guarantee finite-sample report recovery, neural-network convergence, or improved transfer regret; empirical checks do not verify the theorems’ universal quantifiers.

    Source: §8 · p. 10, §10 · p. 11

Abstract

from arXiv · show

Decision-focused surrogates can recover downstream decisions without identifying the quotient report. We characterize the equality set of the convex Smart Predict-then-Optimize surrogate (SPO+) population risk. Under central symmetry, the centered mean class is the unique Bayes minimizer exactly when every nonzero effective displacement makes the old optimizer leave the shifted optimal face with positive probability. This condition separates face crossing from selected-oracle disagreement and gives quantitative local coercivity. Without symmetry, strict crossing alone need not identify the mean; selection balance with reflected crossing restores quotient-report identification, and conditional versions extend the result to measurable predictors. These are population statements, without finite-sample report-recovery or generic transfer-regret guarantees. Closed-form mechanisms reproduce the analytic identities and rates. Portfolio, complete-matrix KuaiRec, and Energy/Storage studies measure predictive fidelity, shifted regret, and fitted-report geometry. A known data-generating process (DGP) companion retains their application geometries while isolating conditional-mean recovery and crossing, without testing the original observational assumptions.

1 Introduction

Decision quality does not determine a predictive report: different reports can induce equivalent downstream decisions. The paper characterizes when SPO+ population risk identifies the effective quotient report, and specifies symmetry, contextual, and empirical scope boundaries.

1 Introduction

  • Decision-equivalent reports need not remain equivalent after an objective, constraint, or feasible-set change, making transfer performance a testable hypothesis rather than an established benefit.The real-data studies evaluate pre-specified, application-specific transfer hypotheses.Source: §1 · p. 1
  • A one-dimensional example has SPO+ minimizers on [1, ∞), despite every point on that ray inducing the Bayes-optimal decision.The paper asks when SPO+ identifies the effective cost component rather than only a downstream-optimal action.Source: §1 · pp. 1–2
  • Without central symmetry, strict crossing alone may fail to identify the mean; reflection-balanced oracle selection with reflected crossing restores exact quotient-report identification.Approximate symmetry localizes the Bayes set under a coercivity condition.Source: §1 · p. 2
  • Conditional central symmetry and strict crossing yield uniqueness of the conditional-mean predictor after quotient projection, up to almost-sure equality.The result is a population Bayes statement and does not itself imply finite-sample report recovery, neural-network convergence, or improved shifted-task regret.Source: §1 · p. 2
  • Every nonzero effective displacement must make the old optimizer leave the entire shifted optimal face with positive probability for unique quotient-report identification under symmetry.This characterizes the complete population minimizer set after invisible directions are quotiented out.Source: §1 · p. 2
  • Selected-oracle disagreement can occur within a shared optimal face, while absolute continuity removes that ambiguity for each fixed displacement.Quantitative crossing gives local parameter coercivity, including quadratic growth under direction-covering polyhedral boundary mass and sharp one-dimensional exponents.Source: §1 · p. 2
  • Portfolio, complete-matrix KuaiRec, and Energy/Storage studies measure predictive fidelity, pre-specified-shift regret, and fitted-report crossing; observational outcomes do not establish population strict crossing.A known-DGP layer retains application feature rows and optimization geometry while setting the conditional cost law by construction; it tests conditional-mean recovery, crossing mass, and finite-family transfer regret, not a universal guarantee.Source: §1 · pp. 2–3

2 Related work

Prior work studies decision consistency, sufficient conditions for targets, and behavior near decision boundaries. This paper instead characterizes the complete zero-excess-risk equality set of population SPO+ after quotienting invisible directions.

Decision-focused learning with SPO+

  • Existing SPO+ analyses establish decision consistency or unique mean minimization under stronger conditions, but do not characterize the complete zero-excess-risk equality set.The paper’s contribution is the necessary-and-sufficient optimal-face strict-crossing condition under centered symmetry.Source: §2 · p. 3

What optimal decisions identify

  • Work on decision sufficiency, inverse optimization, and elicitable or calibrated surrogates targets information, proxies, observations, or regret rather than this paper’s Bayes equality set.Those targets ask different identification questions from whether zero SPO+ excess risk identifies one quotient report.Source: §2 · p. 3

Geometry near decision boundaries

  • Boundary-margin analyses study behavior near decision boundaries, whereas optimal-face crossing characterizes the global population-risk equality set and supports local coercivity.This distinction also separates disagreement by a selected oracle from departure from the full shifted optimal face.Source: §2 · p. 3

3 Problem Setup and Population Notions

The setup distinguishes downstream decision quality from identification of a predictive report. Because directions constant over the feasible set are invisible to the SPO+ risk, the relevant target is the report’s effective quotient component.

3 Problem Setup and Population Notions

  • The SPO task risk evaluates the fixed selected decision at a predicted report against the optimal value at the observed cost.The set-valued optimizer correspondence is paired with a deterministic measurable selection.Source: §3 · pp. 3–4
  • Invisible directions leave the risk unchanged, so identification concerns the quotient report [ˆc] or equivalently the effective projection P_Lˆc.For a full-dimensional feasible set, the quotient coincides with the ambient parameter space.Source: §3 · p. 4
  • Under central symmetry and integrability, the mean class is a Bayes minimizer of the convex SPO+ population risk.The fixed deterministic measurable oracle selection is chosen before population risks are formed.Source: §3 · p. 4
  • The paper separates Bayes existence, unique quotient-report identification, fixed-rule Fisher consistency, selection-independent decision consistency, and quantitative quotient-report identification.The quantitative notion is a local lower bound on excess risk for effective displacements, not task-regret calibration.Source: §3 · p. 4

4 Absolute Continuity Does Not Ensure Identification

Absolute continuity alone does not ensure that SPO+ identifies the quotient report. Identification requires cost mass in optimization-transition regions that are crossed by effective displacements.

4 Absolute Continuity Does Not Ensure Identification

  • For S = [−1/2, 1/2] and C ∼ Unif[1, 2] (mean 3/2), absolute continuity does not by itself make a decision-focused surrogate identify the quotient report, although every minimizer induces the same singleton decision as the mean.Source: §4 · p. 5
  • The risk is flat for reports ˆc ≥ 1 and has positive loss with positive probability for ˆc < 1.The same construction is given with a smooth compactly supported density in Appendix D.Source: §4 · p. 5
  • The obstruction is missing mass in optimization-transition regions, not lack of a density; strict crossing instead requires positive mass in every nonzero effective displacement’s crossing event.Intrinsic full support can suffice, whereas density confined to one decision region does not.Source: §4 · p. 5

5 Strict Crossing and Exact Equality-Set Geometry

Optimal-face strict crossing characterizes exactly when the SPO+ population minimizer set is one quotient-report class. Its criterion concerns departure from the shifted optimal face, not disagreement between selected oracle outputs.

5 Strict Crossing and Exact Equality-Set Geometry

  • The equality event is set-valued: the old decision remains in the shifted optimal face when the crossing gap is zero.The centered excess-risk argument connects positive expected gap to positive probability of strict crossing.Source: §5 · p. 6
  • The minimizer set is the mean class plus invisible directions exactly when every nonzero effective displacement produces optimal-face departure with positive probability.Theorem 5.2 gives the if-and-only-if equality-set characterization.Source: §5 · p. 6
  • Selected-oracle disagreement is not the criterion, because oracle outputs can differ while remaining within a shared optimal face.Theorem 5.2 identifies the exact equality case after invisible directions are quotiented out; prior centered representations and stronger-condition uniqueness are acknowledged as known [Liu and Grigas, 2021].Source: §5 · p. 6

6 Geometry of Strict Crossing

Strict crossing is established through geometry intrinsic to the effective space, while quotient-report identification disregards directions that do not change optimization. Intrinsic full support guarantees crossing.

6 Geometry of Strict Crossing

  • The geometry is intrinsic to L: translating the feasible set preserves its optimizers, and the restricted support function describes its set-valued optimal faces.For polytopes, minimizers are constant on relative interiors of minimization normal cones.Source: §6 · p. 6
  • Full support of the projected cost on the effective space L is sufficient for optimal-face strict crossing.The proof first constructs a relatively open crossing neighborhood, then uses full support to give it positive probability.Source: §6 · pp. 6–7
  • Figure 2 links optimal-face departure to exact quotient identification and indicates that side conditions and quantitative refinement accompany the result.Source: Figure 2 · p. 7
  • Only the effective component is identifiable: directions orthogonal to the feasible-set differences leave SPO+ risk unchanged for every cost distribution.The target is the quotient report [c̄] = c̄ + L⊥, equivalently the projection onto L; ambient full dimensionality is unnecessary.Source: §6 · p. 8

7 Oracle Disagreement and Optimal-Face Crossing

Optimal-face crossing and selected-oracle disagreement are distinct events: disagreement can occur without crossing, while absolute continuity removes this ambiguity for each fixed displacement.

7 Oracle Disagreement and Optimal-Face Crossing

  • Selected-oracle disagreement always implies crossing, but the reverse can fail when the shifted optimum is a non-singleton face.Source: §7 · p. 8
  • Both atomic and atom-free examples have disagreement probability 1/2 and crossing probability 0.In these examples the oracle moves tangentially within a common optimal face, leaving the crossing gap zero; atom-freeness alone is insufficient.Source: §7 · p. 8
  • Under absolute continuity of the projected cost law, selected-output disagreement and departure from the shifted singleton optimal face coincide almost surely for each fixed displacement.The result is displacement-wise, not a simultaneous claim over all displacements.Source: §7 · p. 8

8 Quantitative Crossing and Excess-Risk Growth

Quantitative crossing links local excess-risk growth to crossing mass and gap, with a polyhedral boundary-mass condition yielding quadratic growth. Asymmetric laws require additional selection balance for exact mean identification.

8 Quantitative Crossing and Excess-Risk Growth

  • Quantitative strict crossing and local polynomial coercivity are two views of the same phenomenon, with QSC exponents α + β yielding growth exponent p = α + β.Conversely, a growth exponent p yields QSC(p, p − 1); qualitative strict crossing alone supplies no modulus.Source: §8 · p. 9
  • Direction-covering boundary mass on a finite family of polytope edges yields a local quadratic excess-risk lower bound in every effective displacement direction.The family’s edge directions must span L, and the density condition is intrinsic to L, so ambient cost density is not required.Source: §8 · pp. 9–10
  • The growth bound is a population inverse quotient-report identification modulus, not a task-regret transfer guarantee.It does not automatically establish finite-sample report recovery, neural-network optimization convergence, or improved shifted-task regret.Source: §8 · p. 10
  • A positive boundary density gives a quadratic one-dimensional rate, while vanishing boundary density gives a flatter polynomial rate; strict crossing alone need not give a polynomial modulus.Source: §8 · p. 10
  • For asymmetric laws, strict crossing alone need not identify the mean; selection balance plus reflected crossing restores exact quotient-report identification.The balance condition equates average deployed oracles under observed and reflected costs; it is not generic and is not directly observable from ordinary observational data.Source: §8 · pp. 10–11

9 Real-Data Diagnostics

The real-data studies report distinct patterns across portfolio, KuaiRec, and Energy/Storage, while observational results do not test the theorem’s assumptions.

9 Real-Data Diagnostics

  • Observational data cannot test Theorem 5.2 because they do not reveal conditional means, central symmetry, or population strict crossing.The known-cost-law companion retains application features and geometries but does not test the original observational assumptions.Source: §9 · p. 11
  • The studies use a 48-industry portfolio cap raised from 0.10 to 0.12, user-disjoint complete KuaiRec matrices, and Energy/Storage prices with an intercept and tilts.KuaiRec tests k ∈ {5, 10, 15}; Energy/Storage uses tilts s ∈ {0, 0.5, 1}.Source: §9 · p. 11
  • Portfolio shifted regrets are close despite up to sevenfold report-RMSE differences.Fitted crossing is descriptive; Appendix J shows improved mean recovery and crossing on tested geometries, not population strict crossing or universal transfer.Source: §9 · p. 11
  • Linear SPO+ leads formal methods on KuaiRec at all tested budgets, while Energy SPO+ leads only at the largest tilt.Appendix J’s known-DGP results do not establish population strict crossing or universal transfer.Source: §9 · p. 11

10 Discussion, limitations, and conclusion

The discussion distinguishes symmetry-based mean identification from asymmetric selection balance and confines the findings to population geometry and tested empirical checks.

10 Discussion, limitations, and conclusion

  • Under symmetry, E[C] is the Bayes target; without symmetry, selection balance with reflected crossing identifies the quotient report, while strict crossing alone may fail.Approximate symmetry localizes the Bayes set, but does not make strict crossing alone sufficient.Source: §10 · p. 11
  • The results do not imply finite-sample recovery, neural convergence, or transfer regret, and lower-dimensional feasible sets identify only a quotient.Empirical checks do not verify the theorems’ universal quantifiers.Source: §10 · p. 11

Reproducibility Statement

The source bundle includes materials supporting proofs, analytic illustrations, and reproduction of the experiments.

Reproducibility Statement

AI Use Statement

This appendix supports the paper’s population identification results with proofs, geometric qualifications, counterexamples, and quantitative arguments, and records how generative AI was used. It relates SPO+ risk growth to optimizer crossing and conditional-mean identification modulo effective-report equivalence under stated conditions, without claiming finite-sample recovery or shifted-task regret guarantees.

AI Use Statement

  • The authors used generative AI for research planning, literature synthesis, mathematical and proof assistance, experimental review and interpretation, and prose revision, but not for data generation, data cleaning or reformatting, or qualitative analysis.They manually checked mathematical material, verified software changes and numerical outputs with reported tests, and take responsibility for the final work.Source: AI Use Statement · p. 12

A Proofs for Centering and Strict Crossing

  • Under central symmetry, the centered mean is a global minimizer, and strict crossing is equivalent to its unique minimization on the quotient by the invisible subspace.If strict crossing fails for a nonzero effective displacement, that displacement gives a second quotient minimizer.Source: A Proofs for Centering and Strict… · p. 14

B Optimal-Set Containment for General Compact Convex Sets

  • For the set-valued SPO risk, a singleton optimizer at c* contained in the mean-optimal face suffices for c* to minimize risk.The theorem uses compactness and a worst-decision evaluation of predicted optimal faces; this risk should not be conflated with fixed-selection risk on ties, and arbitrary face-interior points need not be exposable.Source: B Optimal-Set Containment for General… · p. 15

C Geometry of Tie Loci and Full Support

  • Directions orthogonal to span(S − S) leave the SPO+ risk unchanged, so ambient-space uniqueness is impossible when that invisible subspace is nontrivial.Under central symmetry, strict crossing instead characterizes uniqueness on the mean’s affine space of effective directions.Source: C Geometry of Tie Loci and Full Support · p. 16
  • For polytopes with an absolutely continuous projected cost, crossing occurs when the shifted cost leaves the formerly optimal vertex’s closed normal cone; remaining on a shared boundary does not count.Absolute continuity makes the source optimizer unique almost surely, while the closed-cone test preserves the set-valued condition at the shifted cost.Source: C Geometry of Tie Loci and Full Support · pp. 16–17

D Counterexample Details

  • Counterexamples show that compact-support and smooth symmetric costs can have risk plateaus, while a singular atom-free law can permit selected-output changes without the old decision leaving the shifted optimal face.The singular example is supported on a tie slice, so atom-freeness alone does not rule out such behavior.Source: D Counterexample Details · pp. 17–18
  • First moments alone do not yield a uniformly bounded SPO+ loss, and strict crossing alone does not guarantee polynomial local excess-risk growth.Finite-sample arguments using a uniform loss bound need an additional bounded-support or tail assumption; an example satisfying strict crossing has excess smaller than every polynomial near zero.Source: D Counterexample Details · pp. 18–19

E Quantitative Proofs and Sharpness

  • The quantitative proofs derive a lower bound on crossing probability from a local excess-risk bound and use that event’s probability and gap to establish a quantitative risk bound.The appendix also labels one-dimensional sharp exponents, but the supplied passages do not state their values.Source: E Quantitative Proofs and Sharpness · pp. 19–20
  • A local lower bound on population risk growth gives conditional approximate-ERM recovery on the effective slice, but is not a standalone finite-sample rate theorem.The recovery statement is deterministic when local uniform deviation and local membership hold.Source: E Quantitative Proofs and Sharpness · pp. 20–21
  • In one dimension, boundary density proportional to x^ν yields a sharp risk-growth exponent of 2 + ν, through crossing probability and gap scaling with the displacement.With density bounded above and below by multiples of x^ν near the decision boundary, only costs in (−2t, 0) cross it; the crossing gap is 2B(c+2t).Source: Figure 4 · p. 20, E Quantitative Proofs and Sharpness · p. 21

F Contextual Prediction and Further Extensions

  • Strict crossing alone can fail to identify the mean, whereas selection balance with reflection and full support restores exact quotient-report identification.The counterexample has a full-support mean-zero cost law satisfying strict crossing for which zero is not a SPO+ population-risk minimizer.Source: F Contextual Prediction and Further… · p. 21
  • Under the contextual theorem’s conditions, the conditional mean uniquely minimizes risk among eligible measurable predictors modulo equality of effective reports; this does not guarantee finite-sample recovery or shifted-task regret.The stated diagnostics do not extend the population Bayes result to finite-sample, neural-optimization, or shifted-task regret guarantees.Source: F Contextual Prediction and Further… · p. 22

G Proofs for Symmetry Relaxation and Contextual Prediction

This section extends crossing-based quotient identification to approximate symmetry, contextual predictors, and Fenchel transition losses, while clarifying the limits of real-data diagnostics. Deterministic mechanism experiments illustrate how crossing, oracle disagreement, boundary density, and fitted-report geometry relate to surrogate risk and identification, without establishing general identification or transfer guarantees.

G Proofs for Symmetry Relaxation and Contextual Prediction

  • For p > 1, if the symmetrized SPO+ risk increase from µ is at least C∥∆∥^p for 0 < ∥∆∥ ≤ r and 4BSη < Cr^(p−1), a global minimizer exists and every quotient minimizer is localized; exact uniqueness is not established.Source: G Proofs for Symmetry Relaxation and… · p. 22
  • Conditional-risk minimizers are exactly reports with the same projected report as µ(x), and strict conditional excess yields uniqueness modulo almost-sure equality of projected reports.The contextual quantitative corollary applies a conditional lower bound at each effective displacement and integrates it using Tonelli’s theorem.Source: G Proofs for Symmetry Relaxation and… · pp. 22–23
  • With selection balance and reflected crossing, excess SPO+ risk is strictly positive for every nonzero effective displacement, yielding quotient minimizers µ + N; failure of crossing defeats quotient uniqueness.The random excess term vanishes exactly when the reflected old optimizer remains in the shifted optimal face.Source: G Proofs for Symmetry Relaxation and… · p. 23

H Real-Data Protocol and Evidence Scope

  • The known-DGP experiment preserves application geometries while replacing costs with registered laws of known conditional mean; its finite shift tests do not establish observational strict crossing or universal transfer.The realized run was a mechanism and figure run, not the preregistered high-budget confirmation; finite grids cannot establish the all-displacement population condition.Source: H Real-Data Protocol and Evidence Scope · p. 24, J Experiments with Known… · p. 33

I.1 Portfolio

  • Portfolio’s pre-specified cap shift changes regret similarly across methods despite report RMSE differing by up to roughly sevenfold.The λ = 1 penalty endpoint had similar shifted regret to λ = 0.01 but RMSE (3.14 ± 2.74) × 10^6 and unreliable crossing slopes; this was numerically unstable, not evidence of a geometric failure boundary.Source: §I.1 · pp. 25–26, Table 3 · p. 25, Figure 5 · p. 26

I.2 Complete-matrix KuaiRec recommendation

  • Linear SPO+ has the lowest formal-method mean regret at every KuaiRec top-k budget, while the hybrid is close and validation-selected PG is less competitive.The exact complete-matrix evaluation supports additive top-k regret, but is not logged-bandit or off-policy evaluation.Source: §I.2 · p. 27, Figure 8 · p. 28

I.3 Energy/Storage

  • SPO+ improves with the tested Energy training budget, but the largest report penalty and tested PG settings are numerically unstable, and the grid does not establish a universally optimal penalty.At s = 1, LP regrets were 251.86, 48.29, and 42.43 thousand for T = 10, 40, and 80; the best tested PG cell still gave 759.87 thousand.Source: §I.3 · p. 31, Figure 13 · p. 31

J Experiments with Known Data-Generating Processes on Real Application Geometries

  • Known-DGP SC-ON runs had positive crossing in 16/36 Portfolio, 5/36 KuaiRec, and 10/20 Energy/Storage direction-scale rows, versus 2/36, 0/36, and 5/20 under SC-OFF.Conditional-mean curves improved with repeated outcomes under SC-ON, while deterministic SC-OFF recovery reached numerical zero; finite-grid results do not establish population strict crossing.Source: J Experiments with Known… · p. 33, Figure 16 · p. 35

L Details for the Deterministic Illustrations

  • Analytic constructions illustrate the C1 plateau, boundary-density slopes 2 + ν, and oracle disagreement without crossing.These illustrations are analytic constructions, not empirical evidence.Source: Figure 18 · p. 36, N Mechanism Experiments · p. 38

M Fenchel Transition Extension (Supporting)

  • Under central symmetry and the stated integrability assumptions, the Fenchel transition identity yields unique identification exactly when every nonzero displacement has positive selected Bregman divergence with positive probability.The extension is supporting, not a general decision-focused surrogate identification theorem; SPO+ recovery requires the aligned support-function selection, and no generic task-consistency statement follows.Source: M Fenchel Transition Extension… · p. 37
  • Under absolute continuity, selected-oracle disagreement and optimal-face crossing agree almost surely for each fixed displacement.The result uses that the translated tie locus has Lebesgue measure zero.Source: M Fenchel Transition Extension… · p. 38

N Mechanism Experiments

  • Boundary-density log-log slopes are exactly 2 + ν, while strict crossing under a flat C∞ density yields no polynomial risk modulus.Analytic constructions also illustrate the C1 plateau; these are deterministic evaluations, not empirical evidence.Source: Figure 18 · p. 36, N Mechanism Experiments · p. 38, Figure 19 · p. 40
  • At ε = 0, selected-oracle disagreement is 1/2 while optimal-face departure is 0; for ε > 0, the probabilities coincide.The atomic-tie and Cantor examples use increasing noise, and the gap-based events are insensitive to the tested tolerances.Source: Figure 18 · p. 36, N Mechanism Experiments · p. 38, Figure 20 · p. 40
  • For the box, crossing mass and conditional gap scale as t and excess risk as t2; removing a boundary patch creates a dead zone, and the mechanism also holds on a non-symmetric triangle.Lowering the density floor preserves the slopes while shrinking constants.Source: N Mechanism Experiments · p. 38, Figure 21 · p. 40
  • In the controlled SC-off experiment, held-out crossing mass is 0 for displacement (0.5, 0), while all four methods have zero observed regret but differing quotient errors and report variances.The diagnostics are finite-sample illustrations of designed regimes, not estimates of real-data assumptions or neural-convergence results.Source: N Mechanism Experiments · pp. 38–39, Table 7 · p. 39

Related papers

This paper cites

Loading 2609.33472v1…