acceptodds
Under review as a conference paper at ICLR 2027

Geometric Identification in Predict-Then-Optimize Learning

Abstract

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.

Primary area:
learning theory
Keywords:
decision-focused learning, SPO+, statistical identification, Fisher consistency, convex optimization, learning theory

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_ygPwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks when minimizing population SPO+ risk identifies the mean cost report, beyond recovering an optimal decision. It removes cost directions that cannot affect the optimization problem and studies identification in the remaining effective space. Under central symmetry, the key condition is strict crossing. Every nonzero effective displacement must have positive probability of moving the selected optimizer outside the shifted optimal face. The paper gives sufficient support conditions, distinguishes face crossing from changes in a tie-breaking oracle, and derives local risk-growth bounds.

Critical strengths
  1. The distinction between a correct decision and an identified report is worth making. Example 4.1 shows that an entire ray of reports can minimize the surrogate while inducing the same optimal decision. Quotienting out the invisible directions first makes the identification target precise, including for lower-dimensional or equality-constrained feasible sets.

  2. The treatment of ties is a strong part of the paper. A selected oracle can change its output while remaining within the same optimal face, so selected-output disagreement alone is not the appropriate criterion for identification.

  3. The quantitative analysis is good. Theorem 8.7 derives quadratic growth from boundary mass near normal-fan facets associated with a direction-covering set of edges spanning the effective space. The one-dimensional examples also show why strict crossing alone need not imply any polynomial growth rate. Together with the reflected-distribution result in Theorem 8.8, these results help clarify which assumptions are responsible for identification and quantitative risk growth.

Critical weaknesses
  1. The incremental novelty of the central identification result remains unclear. The paper acknowledges that the centered SPO+ excess-risk identity and uniqueness under stronger full-support conditions are already known. Given the nonnegative excess-risk representation, Theorem 5.2's equivalence between unique quotient-report identification and positive-probability crossing follows relatively directly after quotienting out invisible directions. Although the subsequent geometric and quantitative results are substantial, the paper does not sufficiently demonstrate important settings where the new characterization meaningfully advances existing SPO+ theory. Concrete examples where prior sufficient conditions fail but the new criterion establishes identification would strengthen the novelty claim.

  2. The application experiments provide limited evidence about the effective report. For exact top-k selection with fixed cardinality, adding the same value to every cost coordinate leaves the set of optimal selections unchanged. The large raw-RMSE differences in Table 4 may therefore partly be driven by an invisible component. The same issue would arise in Table 3 if the portfolio constraints fix total investment. The authors acknowledge the limitation of raw RMSE, but decomposing prediction residuals into effective-space and invisible components would make these comparisons substantially more informative. Projected residual error would still be a predictive diagnostic, rather than evidence of conditional-mean recovery.

  3. The experimental evaluation remains limited in scope, and several details needed to interpret or reproduce individual experiments are not clearly specified or mapped to the corresponding artifacts. For Table 7, the paper specifies the DGP and five seeds but not the training and held-out sample sizes, MLP architecture, optimizer, or training budget. The observational studies provide some protocol details but still lack dataset date ranges where applicable, retained sample counts, fuller preprocessing specifications, and complete feasible-set definitions. A consolidated experiment-by-experiment mapping to the relevant settings and files would improve reproducibility. The mean-recovery component of Appendix J measures Monte Carlo recovery of a sample mean and therefore does not test whether SPO+ identifies or learns that mean. Table 7 does train predictors, but only on a two-dimensional box, where the quotient coincides with the full ambient space. It therefore does not test report recovery in a genuinely nontrivial quotient setting.

Questions for authors
  1. For the KuaiRec top-k and Portfolio studies, can you report the corresponding effective and invisible subspaces and decompose the prediction residuals into these two components? For Portfolio, please also provide the complete feasible-set definition. This would clarify how much of the reported RMSE differences lies in directions that are actually relevant to the optimization problem.

  2. The novelty of Theorem 5.2 could be clarified. The paper notes that the centered SPO+ excess-risk identity and uniqueness under stronger full-support conditions are already known. Given that the excess risk can be written as �0��0� with �1��1�, the equivalence between unique quotient-report identification and positive-probability crossing appears relatively direct. Please clarify the substantive technical gain over prior SPO+ results, ideally with important settings where existing sufficient conditions fail but the new characterization yields a nontrivial conclusion.

  3. There remains a substantial gap between population identification and finite-sample recovery. The identification and coercivity results are population statements, and the approximate-ERM result still assumes a local uniform-deviation bound and localization of the learned report. Can these conditions be established under standard hypothesis-class assumptions to obtain an explicit finite-sample quotient-report recovery guarantee?

  4. The learned-predictor experiment does not test a nontrivial quotient space. Table 7 uses �2��2�, so �3��3� and �4��4�; quotient-report error therefore coincides with ordinary report error. Could the authors add a controlled learned-predictor experiment with �5��5�, where ambient recovery may fail while effective-space recovery succeeds? This would more directly validate the paper’s quotient-report formulation.

4 comments

Part 1/3 — W1+Q2 and W2+Q1 (residual decomposition)

We thank Reviewer ygPw for the detailed report and the concrete suggestions.

To W1+Q2

We agree with the contribution hierarchy. Liu and Grigas (2021) give the centered SPO+ excess-risk identity; Theorem 5.2 gives its exact equality condition after removing directions in L⊥L^\perp, with ties tested against the full shifted optimal face. Theorem 8.2 integrates a quantitative crossing lower bound, and the contextual statement applies the fixed-covariate result conditionally.

The equality condition covers cost laws without density or full support. For instance, on S=[−1/2,1/2]S=[-1/2,1/2], a symmetric law assigning probability 2−n−12^{-n-1} to each of C=1/nC=1/n and C=−1/nC=-1/n, for every n≥1n\ge1, crosses under every nonzero report displacement. Its unique Bayes report is zero. On the affine set S=(t,1−t):0≤t≤1S=\\{(t,1-t):0\le t\le1\\}, the cost C=Z(1,−1)C=Z(1,-1) with the same law for ZZ identifies the report modulo L⊥=span⁡(1,1)L^\perp=\operatorname{span}(1,1). These examples lie outside the earlier ambient sufficient conditions.

The main quantitative step is the direction-covering boundary-mass result. A positive projected-cost density near a finite family of normal-fan facets whose edge directions span LL yields RSPO+(cˉ+Δ)−RSPO+(cˉ)≥C0∣Δ∣2R_{\mathrm{SPO+}}(\bar c+\Delta)-R_{\mathrm{SPO+}}(\bar c)\ge C_0\\|\Delta\\|^2 locally for every Δ∈L\Delta\in L. In the four-strip example described in our response to wAPG, the density vanishes near the mean and has compact support, yet the bound holds with C0=1/16C_0=1/16 and r=1/10r=1/10. This controls effective report distance; Liu–Grigas calibration controls task regret under different conditions.

To W2+Q1

We decompose the Portfolio and KuaiRec residuals into effective and invisible components and state both feasible sets explicitly. For Portfolio with fixed total investment and for fixed-cardinality KuaiRec top-kk, decision differences lie in the subspace of vectors with zero coordinate sum. Write ri=c^(xi)−cir_i=\hat c(x_i)-c_i and

L=1⊥,L⊥=span⁡(1).L=\mathbf 1^\perp,\qquad L^\perp=\operatorname{span}(\mathbf 1).

The residuals decompose as

PL⊥ri=1⊤rid1,PLri=ri−PL⊥ri,P_{L^\perp}r_i=\frac{\mathbf1^\top r_i}{d}\mathbf1,\qquad P_Lr_i=r_i-P_{L^\perp}r_i,

and the vector RMSEs satisfy

RMSEambient2=RMSEL2+RMSEL⊥2.\mathrm{RMSE}_{\mathrm{ambient}}^2=\mathrm{RMSE}_{L}^2+\mathrm{RMSE}_{L^\perp}^2.

Portfolio uses d=48d=48 and KuaiRec uses d=2062d=2062. The table reports means and sample standard deviations over five training seeds. OLS and the two KuaiRec fits are deterministic (Table 4), so they have no seed variation.

ApplicationMethodAmbient RMSEEffective RMSEInvisible RMSE
Portfolio, value-weightedOLS0.1296±0.00000.1296\pm0.00000.0897±0.00000.0897\pm0.00000.0935±0.00000.0935\pm0.0000
Portfolio, value-weightedSPO+0.1374±0.00250.1374\pm0.00250.1070±0.00320.1070\pm0.00320.0861±0.00000.0861\pm0.0000
Portfolio, equal-weightedOLS0.1411±0.00000.1411\pm0.00000.0936±0.00000.0936\pm0.00000.1056±0.00000.1056\pm0.0000
Portfolio, equal-weightedSPO+0.2121±0.01620.2121\pm0.01620.1879±0.01810.1879\pm0.01810.0980±0.00000.0980\pm0.0000
KuaiRecRidge58.92158.92158.35158.3518.1778.177
KuaiRecLinear SPO+71.72871.72858.45958.45941.56341.563

On KuaiRec, the ambient RMSEs of Linear SPO+ and Ridge are 71.72871.728 and 58.92158.921, but the effective-space RMSEs are 58.45958.459 and 58.35158.351: the difference lies mainly in the invisible component. The reviewer's conjecture therefore holds for KuaiRec. It does not hold for Portfolio, where the gap is in the effective component (0.09360.0936 against 0.18790.1879 for equal-weighted returns) and the invisible components are similar. The SPO+ invisible component shows no seed variation because the SPO+ loss has no gradient in that direction, so its value is set by the initialization. The training objective is invariant to shifts in span⁡(1)\operatorname{span}(\mathbf1), so its gradient in that direction is zero; the fitted common-mode offset retains its initialized value. Projected residuals are held-out predictive diagnostics and do not by themselves establish conditional-mean recovery.

Part 2/3 — Continued from Part 1/3 (W2+Q1, then W3)

The complete Portfolio feasible set is

S_u=\left\\{w\in\mathbb R^{48}:\mathbf1^\top w=1,\quad 0\le w_i\le u,\quad 0.12\le\sum_{i\in G_g}w_i\le0.28\\ (g=1,\ldots,5)\right\\},

where the 48 industries are partitioned into five groups with indices 1–101\text{–}10, 11–2011\text{–}20, 21–3021\text{–}30, 31–4031\text{–}40, and 41–4841\text{–}48. Training uses u=0.10u=0.10, and the held-out evaluation changes the cap to u=0.12u=0.12. Under both constraint sets,

span⁡(Su−Su)=1⊥,(span⁡(Su−Su))⊥=span⁡(1).\operatorname{span}(S_u-S_u)=\mathbf1^\perp,\qquad \bigl(\operatorname{span}(S_u-S_u)\bigr)^\perp=\operatorname{span}(\mathbf1).

The KuaiRec action set at budget kk is

Sk=w∈0,12062:1⊤w=k,S_k=\\{w\in\\{0,1\\}^{2062}:\mathbf1^\top w=k\\},

with linear-optimization relaxation 0≤w≤1,1⊤w=k0\le w\le1,\\ \mathbf1^\top w=k, so again L=1⊥L=\mathbf1^\perp.

The learned-report results on these same application geometries are reported in our response to wAPG (W2). They use real feature rows and semisynthetic costs with known conditional means. In the flat-risk control, SPO+ achieves numerically zero excess risk and regret while its effective report RMSE remains large; MSE recovers the known mean much more closely.

To W3

We give the box experiment's sample sizes and training settings, the application preprocessing and feasible sets, and identify the experiment-to-artifact map for the added results. The two-dimensional box experiment of Table 7 is specified as follows. Let X∼Unif⁡([−1,1]2)X\sim\operatorname{Unif}([-1,1]^2), μ(X)=21+0.25X\mu(X)=2\mathbf1+0.25X, and

C=μ(X)+ε,εj∼iidUnif⁡[−a,a].C=\mu(X)+\varepsilon,\qquad \varepsilon_j\overset{\mathrm{iid}}{\sim}\operatorname{Unif}[-a,a].

The three regimes (SC-on, low mass, SC-off) use a=3a=3, 2.252.25, and 0.250.25. Each regime uses 1024 training samples, 4096 held-out samples, and 5 training seeds. The MLP has one hidden ReLU layer of width 16; the optimizer is Adam with learning rate 10−210^{-2}, batch size 64, and 60 epochs; the hybrid adds squared report error with weight 0.10.1. For this box, L=R2L=\mathbb R^2, so there are no invisible report directions.

The sample sizes, splits, and preprocessing of the application data are:

ApplicationAvailable samples and splitFeatures, target, preprocessing
Portfolio, each return definition computed separately14,119 usable dates each; train/validation/test 9,883/2,117/2,1199{,}883/2{,}117/2{,}119; test period 2018-01-24 to 2026-06-30355 features: five daily return lags per industry, 20-day rolling volatility, 252-day cumulative momentum, and 7 weekday plus 12 month indicators. Rolling features use lagged data; the warm-up period is dropped. Target: next-day cost ct=−rt+1c_t=-r_{t+1}.
KuaiRec1,411 users and 2,062 candidate videos; user split 987/211/213987/211/213, disjoint user IDs51 user features. Numeric missing values are imputed with training-user medians and standardized with training-user means and standard deviations; categorical vocabularies are fitted on training users only. Candidates are retained only when all held-out users have watch ratios. Target: negative watch ratio.
Energy/Storage787 complete 48-slot days; chronological split 550/118/119550/118/119; test period 2013-09-04 to 2013-12-319 features per half-hour slot: HolidayFlag, DayOfWeek, WeekOfYear, Day, Month, Year, PeriodOfDay, ForecastWindProduction, and CO2Intensity. Rows with missing targets or allowed features are dropped before retaining complete days. Target: SMPEP2.

Energy/Storage uses the fixed instance SchedulingInstances/load3/day01.txt with 30-minute slots, 5 machines, and 20 tasks. Writing ximsx_{ims} for assigning task ii to machine mm starting at slot ss, the feasibility constraints are

∑m,sxims=1for each task i,\sum_{m,s}x_{ims}=1 \quad\text{for each task }i,

∑i,s:,s≤t<s+dirixims≤Km(m=1,…,5;t=0,…,47),\sum_{i,s:\\,s\le t < s+d_i}r_i x_{ims}\le K_m \quad (m=1,\ldots,5;\\ t=0,\ldots,47),

with xims=0x_{ims}=0 outside the task time windows and xims∈0,1x_{ims}\in\\{0,1\\}. The linear-programming evaluation uses the relaxation xims∈[0,1]x_{ims}\in[0,1]. The objective coefficients are the task power times 0.50.5 hours times the sum of the prices over the occupied slots. The parameters di,ri,Kmd_i,r_i,K_m, the time windows, and the powers are taken from the instance file.

We agree that the mean-recovery computation in Appendix J is a Monte Carlo estimate for a given conditional mean and trains no predictors; it is therefore not evidence of SPO+ learned-report recovery. The trained-report experiments above and under Q4 address this gap. An experiment-to-artifact map in the experimental appendix lists the settings, scripts, and machine-readable results for each added diagnostic, including the corrected Portfolio runs, known-DGP grids, and simplex learned-report experiment.

Part 3/3 — Continued from Part 2/3 (Q3 and Q4)

To Q3

Yes. The explicit high-probability bound in our response to Z9bk (Q3) applies to a finite, bounded, LL-valued predictor class containing μ(x)=E[PLC∣X=x]\mu(x)=\mathbb E[P_LC\mid X=x], with bounded projected costs and uniform conditional quadratic growth. It gives O(n−1/4)O(n^{-1/4}) effective report error for exact ERM. The empirical minimizer need not be assumed to lie inside the local growth radius: convexity extends the local lower bound along rays to every report reachable by the bounded class. A uniform deviation bound for a bounded class can replace the finite-class Hoeffding term.

The bounded-cost and realizability assumptions can also be relaxed. With sub-exponential centered losses and approximation error aF=inf⁡f∈FR(f)−R(μ)a_{\mathcal F}=\inf_{f\in\mathcal F}\mathcal R(f)-\mathcal R(\mu), the corresponding bound has the form

(EX∣f^n(X)−μ(X)∣2)1/2≤[aF+2ϵn,δ+ηnC∗,tail]1/2,\left(\mathbb E_X\\|\hat f_n(X)-\mu(X)\\|^2\right)^{1/2}\le\left[\frac{a_{\mathcal F}+2\epsilon_{n,\delta}+\eta_n}{C_{\ast,\mathrm{tail}}}\right]^{1/2},

where ϵn,δ\epsilon_{n,\delta} is the sub-exponential uniform deviation term and C∗,tailC_{\ast,\mathrm{tail}} is the growth constant extended to the bounded report range. The common conditional growth assumption remains essential.

We checked that assumption in a two-context finite-class experiment with S=[−1,1]S=[-1,1], a=0.1a=0.1, and C∣X=x∼ρxUnif⁡[−a,a]+(1−ρx)(δ−a+δa)/2C\mid X=x\sim\rho_x\operatorname{Unif}[-a,a]+(1-\rho_x)(\delta_{-a}+\delta_a)/2. For ∣Δ∣≤a/2|\Delta|\le a/2, the conditional excess risk equals 2ρxΔ2/a2\rho_x\Delta^2/a. With ρ0=0.5\rho_0=0.5, 250 exact ERM runs, and n=524,288n=524{,}288, the 95th-percentile report error divided by aa rises from 00 at ρ1=0.5\rho_1=0.5, to 0.05300.0530 at ρ1=0.05\rho_1=0.05, 0.28280.2828 at ρ1=0.01\rho_1=0.01, and 0.35360.3536 when ρ1=0\rho_1=0. A flat context therefore prevents uniform report recovery in this controlled setting.

To Q4

Yes. We trained predictors on a three-action simplex with a two-dimensional effective space and a one-dimensional invisible space. We train MLP reports on the three-action simplex

S=w∈R3:w≥0,1⊤w=1,L=1⊥,L⊥=span⁡(1).S=\\{w\in\mathbb R^3:w\ge0,\\ \mathbf1^\top w=1\\},\qquad L=\mathbf1^\perp,\quad L^\perp=\operatorname{span}(\mathbf1).

The effective space is two-dimensional and the invisible space is one-dimensional. Features are X∼Unif⁡([−1,1]2)X\sim\operatorname{Unif}([-1,1]^2), with known conditional mean

μ(X)=(0,1,2)⊤+1.251+Q(0.15X),\mu(X)=(0,1,2)^\top+1.25\mathbf1+Q(0.15X),

where the columns of QQ form an orthogonal basis of LL. Under SC-on, centered Gaussian noise of scale 1.51.5 is added in LL; under SC-off, costs are deterministic given XX. The training set has 1024 samples and the held-out set has 8192. The model is a two-layer ReLU network of width 32, with batch size 128 and 60 epochs; the Adam learning rate is 0.0030.003 for the effective network and 0.010.01 for the common-mode intercept. The table reports means and sample standard deviations over five seeds.

RegimeMethodAmbient RMSEEffective RMSEInvisible RMSESPO+ excessRegret
SC-onMLP-MSE0.267±0.1940.267\pm0.1940.153±0.0320.153\pm0.0320.178±0.2380.178\pm0.2380.0128±0.00500.0128\pm0.00500.533±0.0020.533\pm0.002
SC-onMLP-SPO+3.903±1.3663.903\pm1.3660.198±0.0540.198\pm0.0543.897±1.3693.897\pm1.3690.0221±0.02000.0221\pm0.02000.540±0.0150.540\pm0.015
SC-offMLP-MSE0.181±0.2350.181\pm0.2350.0087±0.00040.0087\pm0.00040.178±0.2380.178\pm0.2380.0000±0.00000.0000\pm0.00000.0000±0.00000.0000\pm0.0000
SC-offMLP-SPO+3.944±1.3583.944\pm1.3580.574±0.1380.574\pm0.1383.897±1.3693.897\pm1.3690.0000±0.00000.0000\pm0.00000.0000±0.00000.0000\pm0.0000

Under SC-on, MLP-SPO+ has ambient RMSE 3.9033.903, of which the invisible component is 3.8973.897 and the effective component is 0.1980.198. The SPO+ objective has no gradient in the common-mode component, so that component retains its initialization. The five common-mode intercepts start at −1,−0.5,0,0.5,1-1,-0.5,0,0.5,1, while the true common-mode value is 2.252.25. Their invisible errors are 3,∣b−2.25∣\sqrt3\\,|b-2.25|; across the five seeds these have mean 3.8973.897 and sample standard deviation 1.3691.369. This is the case the reviewer asks for: ambient recovery fails, while the effective component is recovered to within 0.1980.198, comparable to 0.1530.153 for MLP-MSE. Under SC-off, the SPO+ effective report error exceeds the MSE value, but the test excess and regret are zero.

The 10-seed scaling results, reported under wAPG (W2), show effective RMSE falling from 0.5950.595 to 0.1180.118 under crossing; without crossing, it stays near 0.510.51 despite zero excess risk and regret.

Thank you for the detailed clarifications. I appreciate the additional analyses and experiments. However, the finite-sample guarantee still relies on additional assumptions, and the new experiment remains relatively limited in scope. More importantly, my central concern about the incremental theoretical novelty remains. Although the new example helps illustrate the characterization, the equivalence in Theorem 5.2 follows relatively directly from the previously established nonnegative excess-risk identity.

Overall, I maintain my original assessment, although I consider the paper borderline. That said, I would not oppose acceptance if the other reviewers reach a consensus in favor of the paper.

Reviewer_Z9bkOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies which geometric conditions guarantee that every minimizer of the population SPO+ risk (a convex surrogate for the SPO risk) coincides with the mean cost vector, up to directions that do not affect the optimization problem.

The paper shows that, for centrally symmetric cost distributions, recovery holds exactly when every nonzero shift in a relevant direction makes the original selected decision strictly suboptimal on a set with positive probability mass. It proves that full support in the relevant space is a sufficient condition. It also gives lower bounds on excess SPO+ risk under stronger crossing conditions. Conversely, a polynomial lower bound on excess risk requires a corresponding lower bound on crossing probability. Under additional density assumptions near decision boundaries, the authors establish local quadratic growth of excess SPO+ risk for polytopes. Finally, the authors relax the symmetry assumption and show when recovery still holds. They also give bounds on the minimizers under approximate symmetry. The appendix also extends the identification results to the contextual setting.

Synthetic examples illustrate the recovery conditions and the bounds. Experiments on portfolio optimization, recommendation, and energy scheduling compare cost prediction error and decision regret.

Critical strengths
  1. The research question is interesting. The paper gives a necessary and sufficient condition for identifying the mean under symmetry, up to directions that do not affect the optimization problem. It also handles ties, where the selected decision can change while the original decision remains optimal.

  2. The investigation is broad. For polytopes, Theorem 8.7 gives local quadratic growth using density bounds near decision boundaries that cover the relevant directions. This gives a local bound on the distance to the mean in the relevant directions, in terms of excess SPO+ risk. Theorem 8.8 gives conditions for identification without central symmetry. The authors also investigate how approximate symmetry affects the minimizers and prove identification in the contextual setting.

Critical weaknesses
  1. The presentation is a major weakness of the paper. The introduction uses terms such as “quotient report” and “shifted optimal face” and discusses GΔG_{\Delta} before giving enough context and definitions. The SPO+ loss is defined before the SPO loss, although it is its convex surrogate. Some discussions focus heavily on proof steps, as in the detailed proof sketch following Theorem 6.1, while the importance of the results and their relationship to previous work need more explanation. Several short sections, especially Sections 4–7, could be combined to make the progression clearer. Some proof references are also vague; for example, Theorem 6.1 refers to “the appendix” without specifying which one. These issues make the contribution unnecessarily difficult to assess. Additionally, the empirical findings need more explanation in the main text. Section 9 gives only a brief account of the results and includes no empirical figures. The reader must go to the appendix to understand much of this part of the paper. Including a representative figure and explaining what the findings demonstrate and how they relate to the identification results, would make the empirical contribution easier to assess in the main text.

  2. Although the authors acknowledge the earlier SPO+ results, the comparison could be more explicit for both Theorems 5.2 and 8.7. Remark 5.3 explains the new characterization in Theorem 5.2, but it would help to clarify which additional cases it covers compared with earlier sufficient conditions for uniqueness. For Theorem 8.7, a direct comparison with the assumptions and bounds in Risk Bounds and Calibration for a Smart Predict-then-Optimize Method (Liu and Grigas, 2021) would clarify what the new conditions add.

Questions for authors
  1. Remark 5.3 explains how Theorem 5.2 relates to the known excess-risk identity. Could you clarify which additional cases the characterization covers compared with earlier sufficient conditions for uniqueness?

  2. For Theorem 8.7, could you compare its assumptions and quadratic bound directly with those in Risk Bounds and Calibration for a Smart Predict-then-Optimize Method (Liu and Grigas, 2021)? Which additional cases are covered by the new geometric conditions?

  3. Possible extension: Could Corollary E.1 yield an explicit finite-sample bound, holding with high probability, on the distance between an empirical SPO+ minimizer and the mean, up to irrelevant directions? What additional assumptions would be needed, including to ensure that the empirical minimizer lies in the region where the local growth bound holds?

2 comments

We thank Reviewer Z9bk. We address the presentation points first, then the two theory comparisons and the finite-sample question.

W1: presentation.

The Introduction now explains why decisions can leave reports undetermined before naming the quotient (lines 029–043); Section 2 defines the quotient after defining SPO and its surrogate SPO+ (lines 103–116). Section 5 defines the crossing gap and event before Theorem 5.2 (lines 212–249). The main text now moves from problem setup to related work, the continuity counterexample, and identification (lines 092, 162, 194, 212); geometry, oracle disagreement, and quantitative growth are subsections of Section 5 (lines 260, 312, 347). Theorem 5.4 has a shorter proof sketch followed by its geometric interpretation (lines 294–303). Proof pointers name Appendices A, C, and E (lines 249, 298, 362, 410–411). Section 6 now includes Portfolio, KuaiRec, and Energy/Storage figures (lines 448–449, 504–506, 520–521); the Portfolio interpretation is at lines 471–473.

W2 and Q1: what the identification result adds.

We agree that the centered excess-risk identity is known. Theorem 5.2 states its equality condition: under central symmetry and integrability, the complete Bayes-report set is cˉ+L⊥\bar c+L^\perp, where L=span⁡(S−S)L=\operatorname{span}(S-S), exactly when every nonzero Δ∈L\Delta\in L has positive probability of moving the selected optimizer off the full optimal face of C+2ΔC+2\Delta. This covers ties and lower-dimensional feasible sets without requiring an ambient density. For example, on S=[−1/2,1/2]S=[-1/2,1/2], the symmetric discrete law Pr⁡(C=±1/n)=2−n−1\Pr(C=\pm1/n)=2^{-n-1}, n≥1n\ge1, crosses for every nonzero shift although it has no density or full support. With S=(t,1−t):0≤t≤1S=\\{(t,1-t):0\le t\le1\\} and C=Z(1,−1)C=Z(1,-1) for the same law ZZ, the Bayes set is L⊥=span⁡(1,1)L^\perp=\operatorname{span}(1,1). These examples give identification outside the earlier ambient sufficient conditions.

W2 and Q2: the quantitative comparison.

Liu and Grigas (2021) calibrate task regret against SPO+ excess risk under ambient Gaussian domination and full-dimensionality. Our boundary-mass result instead controls effective report distance. A density floor on neighborhoods of a finite direction-covering family of normal-fan facets gives

RSPO+(cˉ+Δ)−RSPO+(cˉ)≥C0∣Δ∣2,Δ∈L,∣Δ∣≤r.R_{\mathrm{SPO+}}(\bar c+\Delta)-R_{\mathrm{SPO+}}(\bar c) \ge C_0\\|\Delta\\|^2,\qquad \Delta\in L,\quad \\|\Delta\\|\le r.

For S=[−1,1]2S=[-1,1]^2, the four-strip density p(c)=121A∪B(c)p(c)=\tfrac12\mathbf1_{A\cup B}(c), with A=∣c1∣≤14,,1≤∣c2∣≤2A=\\{|c_1|\le\tfrac14,\\,1\le|c_2|\le2\\} and BB its coordinate swap, vanishes near the mean and outside its compact support. The earlier full-support and Gaussian-domination conditions fail, while the boundary-mass argument yields C0=1/16C_0=1/16 for r=1/10r=1/10. The two bounds control different quantities; the complete assumptions and derivation are in our response to wAPG.

Q3: finite-sample recovery.

Yes, under a finite LL-valued predictor class F\mathcal F containing μ(x)=E[PLC∣X=x]\mu(x)=\mathbb E[P_LC\mid X=x], bounded projected costs and reports, and conditional quadratic growth with common constants C0,r>0C_0,r>0 on one full-measure set. Put H=DS(Bc+2Bf)H=D_S(B_c+2B_f) and C∗=C0min⁡1,r/(Bc+Bf)C_{\ast}=C_0\min\\{1,r/(B_c+B_f)\\}. For an ηn\eta_n-approximate empirical minimizer, with probability at least 1−δ1-\delta,

(EX∣f^n(X)−μ(X)∣2)1/2≤[2Hlog⁡(2∣F∣/δ)/(2n)+ηnC∗]1/2.\left(\mathbb E_X\\|\hat f_n(X)-\mu(X)\\|^2\right)^{1/2} \le \left[\frac{2H\sqrt{\log(2|\mathcal F|/\delta)/(2n)}+\eta_n}{C_{\ast}}\right]^{1/2}.

Take C∗=C0C_{\ast}=C_0 when Bc+Bf=0B_c+B_f=0. Exact ERM gives an O(n−1/4)O(n^{-1/4}) effective-report bound for fixed F\mathcal F. A separate localization assumption is unnecessary: bounded predictions and convexity extend local growth along rays to every report the class can reach. This guarantee still requires realizability and uniform conditional growth.

After reading the response to my comments and to other reviewers, I have changed my score from 2 - weak reject to 3 - weak accept.

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

This paper studies when the SPO+ surrogate identifies a predictive cost report, beyond recovering a downstream-optimal decision. For a compact convex feasible set (S), the authors formulate identification on the effective space, accounting for report directions that are invisible to the optimization problem. Under central symmetry of the cost distribution, the main result characterizes unique quotient-report identification through optimal-face strict crossing: every nonzero effective displacement must cause the original selected optimizer to leave the shifted optimal face with positive probability.

Critical strengths

The paper clearly explains why predicting a cost vector that induces a good decision does not necessarily recover the underlying predictive report. The quotient formulation is particularly appropriate: directions in (L⊥L^\perp) cannot affect objective differences over (S), so ambient-space identification would be an unsuitable target for lower-dimensional feasible sets. Sections 3 and 6 and Proposition C.1 make this distinction mathematically explicit. This provides a useful framework for discussing what decision-focused training can identify, especially in problems with equality constraints.

Critical weaknesses
  1. The headline characterization provides a relatively modest advance over the established SPO+ risk identity E[GΔ(C)]\mathbb{E}[G_\Delta(C)] and the fact that a nonnegative random variable has positive expectation exactly when it is positive with positive probability. Liu and Grigas (2021, Theorem 2.1) already state the same centered optimization-gap identity and establish unique mean minimization under stronger assumptions. The quotient formulation, weaker distributional assumptions, and explicit handling of ties are worthwhile refinements, and the authors appropriately acknowledge the known identity. Nevertheless, several prominent results remain direct consequences of this representation: Theorem 8.2 integrates the lower bound in the definition of quantitative crossing, while the contextual identification result follows by conditioning. The direction-covering boundary-mass theorem therefore needs to carry much of the substantive novelty.

  2. The experimental results are not promising. The paper states these limitations explicitly. However, they substantially restrict what this experimental layer contributes. Table 7 does evaluate learned quotient reports, but only in the controlled two-dimensional box example. The application-geometry study uses just eight, eight, and three feature rows for Portfolio, KuaiRec, and Energy/Storage, respectively, and does not extend that learned-report evidence to these settings.

Questions for authors

See Weaknesses

2 comments

Part 1/2 — W1 and W2 (learned-predictor study)

We thank Reviewer wAPG for laying out the dependencies among the theoretical results and the specific gaps in experimental coverage.

To W1

We agree with the proposed contribution hierarchy. Liu and Grigas (2021) give the centered SPO+ excess-risk identity. Theorem 5.2 characterizes its equality set after quotienting invisible directions; Theorem 8.2 integrates the quantitative crossing condition, and the contextual result applies the fixed-covariate characterization conditionally. The main new quantitative step is the direction-covering boundary-mass result.

Let SS be a compact polytope with effective space L=span⁡(S−S)L=\operatorname{span}(S-S). Suppose PLCP_LC has a density on LL. Take a finite set of edges whose directions span LL, and assume the density is bounded below by m>0m>0 on two-sided neighborhoods of positive-measure patches of their normal-fan facets. Under central symmetry and integrability, there are C0,r>0C_0,r>0 such that

RSPO+(cˉ+Δ)−RSPO+(cˉ)≥C0∣Δ∣2,Δ∈L,∣Δ∣≤r.R_{\mathrm{SPO+}}(\bar c+\Delta)-R_{\mathrm{SPO+}}(\bar c) \ge C_0\\|\Delta\\|^2,\qquad \Delta\in L,\quad \\|\Delta\\|\le r.

Every effective displacement crosses at least one controlled facet; both the crossing mass and gap scale with its length, giving the quadratic bound uniformly over directions. Liu–Grigas calibration instead relates task regret to SPO+ excess risk under ambient Gaussian domination and full-dimensionality. The two results control different quantities.

Here is a distribution where the earlier sufficient conditions fail. Take S=[−1,1]2S=[-1,1]^2 and

p(c)=121A∪B(c),A=∣c1∣≤14,1≤∣c2∣≤2,B=∣c2∣≤14,1≤∣c1∣≤2.p(c)=\tfrac12\mathbf1_{A\cup B}(c),\quad A=\\{|c_1|\le\tfrac14,\\ 1\le|c_2|\le2\\},\quad B=\\{|c_2|\le\tfrac14,\\ 1\le|c_1|\le2\\}.

The four compact strips leave an open neighborhood of the mean with zero density; at the zero mean, every action is optimal. Hence neither unique optimality at the mean nor ambient Gaussian domination holds. A horizontal and a vertical edge cover all effective directions; their facet neighborhoods have density floor 1/21/2. The bound above becomes

RSPO+(Δ)−RSPO+(0)≥116∣Δ∣2,∣Δ∣≤110.R_{\mathrm{SPO+}}(\Delta)-R_{\mathrm{SPO+}}(0) \ge \tfrac1{16}\\|\Delta\\|^2,\qquad \\|\Delta\\|\le\tfrac1{10}.

Thus the local facet condition gives report growth in a concrete setting outside the older sufficient conditions.

To W2

We added a learned-predictor study on a nontrivial quotient, expanded each application-geometry grid to 24 feature rows, and trained reports on the known-mean Portfolio and KuaiRec geometries. These experiments test the two regimes in Theorem 5.2: effective error decreases with training size when crossing holds in the controlled simplex, while a flat-risk control permits large report error alongside near-zero excess and regret. Here “SC-on” uses centered cost noise and “SC-off” is deterministic conditional on the features.

In the two-dimensional box experiment (Table 7), L=R2L=\mathbb R^2, so there are no invisible report directions. To examine learned reports under a nontrivial quotient, we use the three-action simplex

S=w∈R3:w≥0,1⊤w=1,L=1⊥,L⊥=span⁡(1).S=\\{w\in\mathbb R^3:w\ge0,\\ \mathbf 1^\top w=1\\},\qquad L=\mathbf 1^\perp,\quad L^\perp=\operatorname{span}(\mathbf 1).

The effective space is two-dimensional and the invisible space is one-dimensional. Under SC-on, MLP-SPO+ attains effective report RMSE 0.198±0.0540.198\pm0.054, invisible-component RMSE 3.897±1.3693.897\pm1.369, and regret 0.540±0.0150.540\pm0.015. For reference, MLP-MSE attains effective RMSE 0.153±0.0320.153\pm0.032 and regret 0.533±0.0020.533\pm0.002. SPO+ has no gradient in the common-mode component: that component is unconstrained by its objective and does not change decisions. SPO+ thus recovers the effective component to an accuracy comparable to MSE and leaves the invisible component untrained, which is the behavior the quotient formulation describes.

A 10-seed scaling diagnostic trains the same MLP-SPO+ at n∈128,512,2048,8192n\in\\{128,512,2048,8192\\}. Under SC-on, effective RMSE decreases from 0.5950.595 to 0.1180.118, and held-out excess risk decreases from 0.13030.1303 to 0.00590.0059. Under SC-off, held-out excess and regret are zero at all four sample sizes, and effective RMSE stays near 0.510.51. This is the predicted dichotomy: the effective error decreases with nn only when crossing holds, whereas zero excess and regret under SC-off do not constrain the report.

Part 2/2 — Continued from Part 1/2 (W2)

Known-DGP geometry grid.

In response to the concern about the 8, 8, and 3 feature rows, we reran the diagnostic with 24 fixed test feature rows per application geometry and 8 conditional cost draws per row. It evaluates three effective directions, both signs, and five displacement magnitudes, giving 30 direction–sign–scale cells. SC-on uses five independent cost-noise seeds; SC-off uses one reference cost. This enlarged grid replaces the earlier small-grid crossing counts; it does not replace Appendix J's separate Monte Carlo mean-recovery diagnostic. The grid sizes and tested displacements differ, so the old and new counts should not be compared cell by cell. The counts of cells with an observed crossing are:

Application geometrySC-onSC-off
Portfolio, value-weighted27/3027/303/303/30
Portfolio, equal-weighted28/3028/301/301/30
KuaiRec14/3014/300/300/30
Energy/Storage29/3029/305/305/30

These counts describe crossing in the tested cells under the specified conditional distributions. They do not establish the population condition, which requires positive crossing probability for every nonzero effective displacement. The observational datasets do not provide the conditional means needed to check central symmetry. The deterministic control is locally flat for most, but not all, tested cells.

Trained reports on the known-mean geometries.

For the Portfolio (both return definitions) and KuaiRec, we train MLPs on the application feature rows and feasible sets. Costs are drawn from the known conditional mean plus generator noise, and the reports are evaluated on held-out feature rows.

SC-off control.

The table reports means and sample standard deviations over five seeds; effective RMSE is listed as MSE / SPO+.

ApplicationTraining rowsEffective RMSE (MSE / SPO+)SPO+ excessSPO+ regret
Portfolio, value-weighted20480.0052±0.0008/3.288±1.0910.0052\pm0.0008\\ /\\ 3.288\pm1.091<3×10−8< 3\times10^{-8}<5×10−10< 5\times10^{-10}
Portfolio, equal-weighted20480.0086±0.0018/2.607±0.7780.0086\pm0.0018\\ /\\ 2.607\pm0.778<3×10−8< 3\times10^{-8}<5×10−10< 5\times10^{-10}
KuaiRec9871.113±0.108/50.177±1.6351.113\pm0.108\\ /\\ 50.177\pm1.635<3×10−8< 3\times10^{-8}<5×10−10< 5\times10^{-10}

Across the two Portfolio return definitions and KuaiRec, the SC-off runs give SPO+ numerically zero excess and regret, although its effective report RMSE exceeds the MSE error. This illustrates the distinction in Theorem 5.2: when crossing fails, the decision may be optimal even though the report is not identified. The bounds in the last two columns cover the absolute values of all five seed-level SPO+ results in each setting.

Reviewer_LgKhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how the geometry of the feasible set affects the identification of the cost vector through SPO+, beyond achieving optimal downstream decisions, which may improve the reusability of SPO+ predictions. To this end, the authors define identification through an equivalence class constructed based on the orthogonal complement of the feasible set's difference span, and then characterize the conditions under which SPO+ achieves identification.

Critical strengths

The paper characterizes how the geometric structure of the feasible set affects the identification properties of SPO+, providing theoretical insights into when the predicted cost vector can be reused in other downstream optimization problems.

Critical weaknesses

My main concern is the practical motivation for this work, particularly why it is important to understand the identification of the conditional mean through SPO surrogates. The distinction between end-to-end decision optimality and prediction optimality is definitely interesting, and I totally agree that "the distinction becomes important when the prediction is reused." However, based on my understanding, the predicted cost vector in SPO serves more as an intermediate variable for achieving end-to-end decision optimality, and the "prediction" is intentionally tailored to the downstream optimization problem, potentially sacrificing predictive accuracy in exchange for better decision performance. Thus, reusing SPO predictions may not be the most natural choice given the original design of the method, and I find the motivation for studying their identification properties less compelling unless such reuse of SPO predictions is common in practice.

If we want something reusable and less dependent on the specific downstream optimization problem, one simple idea is to train a separate model by minimizing an MSE loss, where we no longer need to understand the identification of an end-to-end method. Also, when it comes to reusing predictions, the identification property studied in this work depends on the geometric structure of the feasible set. If S changes drastically, I am not sure whether the SPO predictions would remain reusable, even if they satisfy the identification property established in this paper.

Questions for authors
  1. As my main concern, could the authors elaborate more on why it is important to understand the identification of the conditional mean through SPO surrogates?

  2. I wonder what the results will be like if the objective is non-linear. I think SPO structure can be easily extended into non-linear objectives, then how about the geometric identification?

4 comments

Part 1/3 — Motivation and Q1

We thank Reviewer LgKh for following up on the motivation question. We address the main concern and the two questions in turn.

To the main concern (motivation) and Q1

SPO+ is trained to optimize downstream decision quality. The predictor first outputs a cost vector c^\hat c, which the downstream optimization problem maps to a decision ww. Prior work has primarily evaluated the resulting decisions. We ask what constraints SPO+ training places on this intermediate cost report.

This question arises even with the feasible set held fixed. Different reports can induce the same action, and the SPO+ loss is exactly invariant to shifts along certain report directions. Low decision regret alone therefore does not tell us how much cost information the model has recovered. Our identification condition states when the population risk uniquely determines the effective report after directions invisible to SS are removed. When this condition holds under conditional central symmetry, the identified report equals the conditional mean.

This distinction matters when evaluating or using a trained report. Regret evaluates performance when the goal is a good decision under the current feasible set. If the cost predictions are to be interpreted, compared across models, or reused after constraints change, one needs to know which report directions SPO+ determines and which it leaves unconstrained. Along directions invisible to SS, the loss is invariant, so the fitted component can depend on initialization; an auxiliary predictive loss can constrain it. MSE directly targets the conditional mean. Our result gives conditions under which an SPO+-trained model's output also has a conditional-mean interpretation.

Formally, let L=span⁡(S−S)L=\operatorname{span}(S-S). For any δ∈L⊥\delta\in L^\perp, δ⊤w\delta^\top w is constant over w∈Sw\in S, so the corresponding terms cancel in the SPO+ loss:

ℓSPO+S(c^+δ,c)=ℓSPO+S(c^,c),δ∈L⊥.\ell_{\mathrm{SPO+}}^{S}(\hat c+\delta,c) = \ell_{\mathrm{SPO+}}^{S}(\hat c,c), \qquad \delta\in L^\perp.

This invariance holds for every observation. Prior work (Elmachtoub and Grigas, 2022) shows that, under continuity and full support of the cost distribution, central symmetry about the conditional mean, and a unique optimizer at that mean, the conditional mean uniquely minimizes the SPO+ population risk. Theorem 5.2 states that, under central symmetry, the complete SPO+ Bayes-report set is cˉ+L⊥\bar c+L^\perp if and only if every nonzero effective displacement has positive-probability optimal-face crossing (lines 237–249). Under conditional central symmetry,

arg⁡min⁡c^RSPO+(c^)=cˉ+L⊥⟺Pr⁡w∗(C)∉W∗(C+2Δ)>0∀Δ∈L∖0.\arg\min_{\hat c}R_{\mathrm{SPO+}}(\hat c) = \bar c+L^\perp \quad\Longleftrightarrow\quad \Pr\\{w^*(C)\notin W^*(C+2\Delta)\\} > 0 \quad \forall\Delta\in L\setminus\\{0\\}.

The held-out Portfolio comparison separates report accuracy from decision regret after the position cap changes from u=0.10u=0.10 to u=0.12u=0.12. Report RMSE is the root mean squared Euclidean error against realized costs; regret compares each chosen portfolio with the hindsight optimum under the same cap.

Return definitionReport RMSE (OLS)Report RMSE (SPO+)Regret at u=0.12u=0.12 (OLS)Regret at u=0.12u=0.12 (SPO+)
Value-weighted0.1296±0.00000.1296\pm0.00000.1374±0.00250.1374\pm0.00250.01488±0.000000.01488\pm0.000000.01492±0.000080.01492\pm0.00008
Equal-weighted0.1411±0.00000.1411\pm0.00000.2121±0.01620.2121\pm0.01620.01542±0.000000.01542\pm0.000000.01546±0.000110.01546\pm0.00011

Entries are means ±\pm sample standard deviations over five training seeds, using the same held-out dates within each return definition; OLS is a deterministic fit, so its standard deviation is 0.

These are new runs, not a relabeling of Table 3. The legacy parser read the value-weighted and equal-weighted blocks of the source CSV together, duplicated dates, and mixed return definitions in lagged features and next-day targets. Table 3 reported OLS/SPO+ RMSEs of 0.1324/0.2036 and shifted regrets of 0.015078/0.015160 on those processed rows. The original protocol did not identify a unique weighting convention, so we cannot assign Table 3 to either clean definition. We rebuilt the two definitions separately and reran the five-seed comparison shown here. The accompanying manuscript PDF still contains the legacy Table 3 and figure; the clean rerun values in this comment supersede those legacy numbers for the Portfolio comparison.

OLS has the lower report RMSE under both return definitions. The regret difference is 0.000040.00004 in each definition, smaller than the SPO+ seed-level standard deviation in these runs. The comparison illustrates the reviewer's point that MSE is the direct objective for mean prediction; the identification result concerns population risk.

Part 2/3 — Continued from Part 1/3 (Motivation and Q1)

What happens when SS changes (cross-set reuse)

The reviewer asks whether SPO+ predictions remain reusable when SS changes drastically. The quotient formulation gives a direct answer. Suppose the source SPO+ risk identifies the conditional mean up to Lsrc⊥L_{\mathrm{src}}^\perp, and the source and target use the same cost coordinates. If Lnew⊆LsrcL_{\mathrm{new}}\subseteq L_{\mathrm{src}}, then every source Bayes report c^=cˉ+n\hat c=\bar c+n, n∈Lsrc⊥n\in L_{\mathrm{src}}^\perp, satisfies

n∈Lnew⊥,arg⁡min⁡w∈Snewc^⊤w=arg⁡min⁡w∈Snewcˉ⊤w.n\in L_{\mathrm{new}}^\perp, \qquad \arg\min_{w\in S_{\mathrm{new}}}\hat c^\top w = \arg\min_{w\in S_{\mathrm{new}}}\bar c^\top w.

The source report therefore preserves the full target optimizer set under this condition. In words, reuse is safe whenever the new feasible set introduces no decision direction that the source set could not see, for example when bounds or caps change within the same affine hull, as in the Portfolio cap change.

The guarantee can fail when the target set introduces a direction that the source set cannot see. Let Ssrc=(t,0):−1≤t≤1S_{\mathrm{src}}=\\{(t,0):-1\le t\le1\\} and C=(Z,1+V)C=(Z,1+V), where Z,VZ,V are independent centered Gaussian variables with positive variances. Every (0,b)(0,b) is a source Bayes report. On Snew=conv⁡(−1,0),(0,0),(1,1)S_{\mathrm{new}}=\operatorname{conv}\\{(-1,0),(0,0),(1,1)\\}, the mean (0,1)(0,1) minimizes on the edge between (−1,0)(-1,0) and (0,0)(0,0), while the source Bayes report (0,−1)(0,-1) uniquely selects the strictly suboptimal action (1,1)(1,1).

The Portfolio results evaluate the prespecified cap change from u=0.10u=0.10 to u=0.12u=0.12. The quadratic-objective experiment below tests transfer in both directions. These experiments measure performance under the tested shifts; we make no claim about how often predictions are reused in practice.

To Q2

Yes. The population identification argument extends to continuous function-valued objectives on a compact decision space. We give the assumptions, excess-risk identity, crossing criterion, and a quadratic example here; the full proof will accompany the nonlinear extension.

Setting. Fix X=xX=x, let W\mathcal W be a compact metric decision space, and let C∈C(W)C\in C(\mathcal W) be the conditional random objective function, with μ=E[C∣X=x]\mu=\mathbb E[C\mid X=x]. Let the report class be A=r+V\mathcal A=r+\mathcal V, where r∈C(W)r\in C(\mathcal W) is a fixed reference function and V⊆C(W)\mathcal V\subseteq C(\mathcal W) is a closed linear subspace. Assume that CC is Bochner integrable, belongs to A\mathcal A almost surely, and has mean μ∈A\mu\in\mathcal A. Its law is centrally symmetric about μ\mu as a C(W)C(\mathcal W)-valued random element, so C=d2μ−CC\overset{d}{=}2\mu-C. Let ω(C)∈arg⁡min⁡w∈WC(w)\omega(C)\in\arg\min_{w\in\mathcal W}C(w) be a fixed measurable optimizer selection.

The function-valued SPO+ loss is

ℓSPO+W(f,C)=sup⁡w∈WC(w)−2f(w)+2f(ω(C))−min⁡w∈WC(w).\ell_{\mathrm{SPO+}}^{\mathcal W}(f,C) = \sup_{w\in\mathcal W}\\{C(w)-2f(w)\\}+2f(\omega(C))-\min_{w\in\mathcal W}C(w).

For any Δ∈V\Delta\in\mathcal V, the population excess risk satisfies

R(μ+Δ)−R(μ)=E![C(ω(C))+2Δ(ω(C))−min⁡w∈WC(w)+2Δ(w)].\mathfrak R(\mu+\Delta)-\mathfrak R(\mu) = \mathbb E\\!\left[C(\omega(C))+2\Delta(\omega(C))-\min_{w\in\mathcal W}\\{C(w)+2\Delta(w)\\}\right].

Let

N=Δ∈V:Δ is constant on W.\mathcal N = \\{\Delta\in\mathcal V:\Delta\text{ is constant on }\mathcal W\\}.

The Bayes report set is exactly μ+N\mu+\mathcal N if and only if every nonconstant displacement Δ∈V∖N\Delta\in\mathcal V\setminus\mathcal N has positive probability of moving the selected optimizer outside the optimizer set of the shifted objective:

\omega(C)\notin \arg\min_{w\in\mathcal W}\\{C(w)+2\Delta(w)\\} \right\\} > 0.$$ The proof uses joint symmetry in $C(\mathcal W)$ and the nonnegative shifted-optimum gap, as in the linear case. We will include the full argument in the nonlinear-extension appendix. For linear costs, $\mathcal N$ corresponds to $L^\perp$. When $\mathcal V=C(\mathcal W)$, full support of the random objective in the supremum-norm topology, together with joint central symmetry, is sufficient for strict crossing.

Part 3/3 — Continued from Part 2/3 (Cross-set reuse and Q2)

As a check, a finite-action experiment uses the nonlinear objective

Fθ(w)=θ1w+θ2w2,ϕ(w)=(w,w2),Θ∼N((0,0.5),0.82I2).F_\theta(w)=\theta_1w+\theta_2w^2,\qquad\phi(w)=(w,w^2),\qquad\Theta\sim\mathcal N\bigl((0,0.5),0.8^2I_2\bigr).

Let W0=−1,1\mathcal W_0=\\{-1,1\\} and W1=−1,0,1\mathcal W_1=\\{-1,0,1\\}. In feature coordinates,

L0=span⁡ϕ(1)−ϕ(−1)=span⁡(1,0),L1=R2.L_0=\operatorname{span}\\{\phi(1)-\phi(-1)\\}=\operatorname{span}\\{(1,0)\\},\qquad L_1=\mathbb R^2.

The quadratic coefficient is invisible on W0\mathcal W_0 and visible on W1\mathcal W_1. We use n=1024n=1024 training samples, 30 independent training replicates, and the same 50,000 held-out coefficient draws for each replicate. Empirical SPO+ risk is minimized exactly by linear programming. Regret is measured against the best action for each draw, so it is positive even for the true mean. Coefficient RMSE here is ∣θ^−μ∣2/2\\|\hat\theta-\mu\\|_2/\sqrt2, averaged over the two coefficients; the Portfolio tables use the unnormalized vector-norm RMSE.

Training method and source setTarget setCoefficient RMSEDecision regret
MSE sample meanW1\mathcal W_10.0233±0.01330.0233\pm0.01330.4419±0.00000.4419\pm0.0000
SPO+ trained on W1\mathcal W_1W1\mathcal W_10.0278±0.01230.0278\pm0.01230.4419±0.00000.4419\pm0.0000
SPO+ trained on W0\mathcal W_0W1\mathcal W_10.3542±0.00070.3542\pm0.00070.9460±0.00060.9460\pm0.0006
MSE sample meanW0\mathcal W_00.0233±0.01330.0233\pm0.01330.6390±0.00060.6390\pm0.0006
SPO+ trained on W1\mathcal W_1W0\mathcal W_00.0278±0.01230.0278\pm0.01230.6389±0.00060.6389\pm0.0006

Entries are means ±\pm sample standard deviations over 30 training replicates. For SPO+ trained on W0\mathcal W_0, every value of θ^2\hat\theta_2 minimizes the empirical risk; we report the representative θ^2=0\hat\theta_2=0.

The report trained on W1\mathcal W_1 transfers to W0\mathcal W_0 with regret 0.63890.6389, close to the MSE value 0.63900.6390. In the reverse direction, the endpoint-trained report has regret 0.94600.9460 on W1\mathcal W_1, compared with 0.44190.4419 for the MSE fit. This regret depends on the arbitrary choice of θ^2\hat\theta_2, which is the point of the example: a coefficient that the source set cannot see is not determined by SPO+ training, so reuse on a set that can see it is unreliable.

A separate population-risk diagnostic uses W=[−1,1]\mathcal W=[-1,1] and shifts the quadratic report coefficient by 0.50.5, giving Δ(w)=0.5w2\Delta(w)=0.5w^2. This displacement varies over [−1,1][-1,1], but is constant on −1,1\\{-1,1\\}, where w2≡1w^2\equiv1. We estimate the expectation from 100,000 antithetically paired Gaussian draws and compute each one-dimensional quadratic extremum exactly. On [−1,1][-1,1], the excess risk is 0.192680.19268 (Monte Carlo SE 0.000490.00049) and the crossing probability is 0.922240.92224 (SE 0.000570.00057). Both are zero on the endpoint set.

The function-valued result is a population identification criterion. The quantitative-growth and finite-sample results in the paper and in this response are proved for linear polyhedral problems, and we do not claim them for nonlinear objectives.

Dear Reviewer

We have only one day left for author-reviewer discussion. If the rebuttal clarifies your concerns, can you please acknowledge and increase the score.

Thanks

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.