- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies multivariate conformal prediction (CP) based on optimal transport (OT) and proposes learning a Mahalanobis ground cost over the compact set of determinant-one, eigenvalue-bounded positive-definite matrices M to minimize the Lebesgue volume of the resulting prediction region.
Empirically, they approximate with a GradNet$\Psi_\theta$ trained on a smoothed KDE of the estimation scores for both homogeneous and heterogeneous settings. Theoretical results establish finite-sample marginal coverage (Theorems 3.1, 4.1), and asymptotic conditional coverage for IMV-OTCP (Theorem 4.2). Experiments on synthetic 2D settings, five multivariate regression datasets (), and downstream robust optimization tasks evaluate the proposed methods against other benchmarks.
Critical strengths
-
Clear geometric characterization of volume and task-aware losses via the metric Brenier Jacobian (Section 3.2, Example 3.1, Appendix A.3). The paper cleanly formulates how changing the quadratic ground cost alters the OT pullback region . Example 3.1 (p. 4, proved in Appendix A.1) nicely illustrates that for an elliptical score law with dispersion and a fixed spherical reference , the volume-optimal quadratic ground cost is rather than or . Furthermore, writing the volume and downstream robust decision losses as expectations over the fixed reference ball weighted by the reverse-map Jacobian determinant (Eqs. 4–5) provides a unified framework that extends naturally to task-aware uncertainty sets (TA-OTCP, Appendix D.3–D.4) while preserving exact finite-sample marginal validity through data splitting or permutation-equivariant full-conformal assignment.
-
Thoroughness of the theoretical and experimental package. The manuscript is carefully written and complete in its technical execution: it includes polyhedral characterizations of both unweighted and localized Kantorovich prediction sets (Propositions B.1 and C.1), an anchor-amortized localized split-conformal extension (IMV-OTCP), a full-conformal localized extension (Appendix C), and evaluations across synthetic homogeneous, synthetic heterogeneous, five real multivariate regression datasets (up to ), and portfolio optimization tasks.
Critical weaknesses
-
It is not well justified why a separate linear matrix is needed when the reverse transport map is already approximated using a flexible neural network for (which is then discarded at calibration time in MV-OTCP, Eq. 7), nor why the linear transformation induced by cannot effectively be absorbed into the learned transport representation.
-
The matrix only globally rotates and rescales the score difference (which, as shown in Appendix F.4, p. 36, is mathematically equivalent to running standard Euclidean OT on linearly transformed scores ). It is unclear why this simple linear reweighting is preferable to basic score whitening (), especially since the localized ablation in Appendix D.2 (Figure 6) shows that learning reduces volume by only 1%–3% over fixing once local OT adaptation is present.
-
The methodological novelty is incremental relative to prior multivariate OT conformal inference papers (e.g., FOTCP of Ndiaye, 2025; SOTCP of Thurin et al., 2025 / Klein et al., 2025; Kondratyev et al., 2026), as the framework primarily replaces the squared Euclidean ground cost with the quadratic Mahalanobis cost , and the main real-data and heterogeneous benchmarks (Figure 4, Table 1) omit these OT-based baselines.
Questions for authors
-
In Example 3.1 and Figure 1, is suboptimal because transforms both and the spherical reference , while Appendix F.4 (p. 36) shows that discrete -assignment is equivalent to Euclidean OT on linearly transformed scores . If we simply standardize/whiten the scores via and run standard Euclidean FOTCP () with spherical , what are the resulting set volumes and symmetric differences in Figures 1, 2, 3, and 5? Any specific reason we are not pursuing this?
-
Why are OT-based baselines (FOTCP, SOTCP of Thurin et al. 2025 / Klein et al. 2025, Conditional Neural OT of Kondratyev et al. 2026, and IMV-OTCP with ) omitted from Figure 4 and Table 1? Can you report the performance of IMV-OTCP() and IMV-OTCP() across all the real datasets in Table 1 so we can see how much of the volume reduction actually comes from learning versus the localized GradNet OT pipeline?
-
By the Monge–Ampère identity (Eq. 10 and Eq. 24), the volume objective evaluates the inverse of a -dimensional Gaussian KDE on the fixed origin-centered ball .
- (a) How reliable is this KDE-based bilevel objective when (
rf1) or (scm1d) with , and why use a KDE-dependent Monge–Ampère loss + anchor networks rather than training a conditional convex potential flow / ICNN directly via maximum likelihood / the Knott–Smith criterion (e.g., Brenier, 1991; Huang et al., 2020; Kan et al., 2022) without any KDE? - (b) Since Eq. (24) shows that the volume integrand is , why fix the reference acceptance set to the origin-centered ball (which forces inclusion of low-density interior regions in multimodal or ring-like distributions) rather than optimizing the acceptance region/cutoffs in rank space using the Jacobian density or directly conformalizing the highest-density region (e.g., Izbicki et al., 2022; Camehl et al., 2024)?
- In MV-OTCP (Section 3.3), you fit a continuous Brenier map during metric learning, but then discard it and solve a discrete assignment problem (Eq. 7) at calibration/test time, whereas in IMV-OTCP and TA-OTCP you use split calibration with a forward GradNet. Why not use split calibration in MV-OTCP as well, and how is the volume of the union of polyhedral cells in Proposition B.1 computed numerically?
Minor suggestions and missing references (not affecting score): Cite Brenier (1991) when invoking Brenier's theorem in Lemma 3.1 and Appendix F.1; discuss conditional convex potential flows / multivariate quantile forecasters (Huang et al., 2020; Kan et al., 2022) and center-outward / superlevel-set quantile literature (Hallin et al., 2023; Camehl et al., 2024); report wall-clock training and inference times comparing SCP, MVCS, FOTCP, IMV-OTCP(), and IMV-OTCP; fix the typos on p. 14 line 752 ("for convinence") and p. 28 line 1511 ( should be ).
8 comments
W3
The methodological novelty is incremental relative to prior multivariate OT conformal inference papers (e.g., FOTCP of Ndiaye, 2025; SOTCP of Thurin et al., 2025 / Klein et al., 2025; Kondratyev et al., 2026), as the framework primarily replaces the squared Euclidean ground cost with the quadratic Mahalanobis cost , and the main real-data and heterogeneous benchmarks (Figure 4, Table 1) omit these OT-based baselines.
Response. Existing methods, including FOTCP (Ndiaye, 2025), SOTCP (Thurin et al., 2025), and conditional neural OT (Kondratyev et al., 2026), construct multivariate ranks under a prespecified ground cost. Our contribution is to optimize the ground metric for prediction-set volume or a downstream decision loss. The metric Brenier representation expresses the volume of as a Jacobian integral over a fixed reference ball. This lets us account for the dependence of the transport map on the metric through implicit differentiation. The quadratic cost also preserves the polyhedral representation of empirical assignment regions and the finite-sample conformal coverage guarantee.
We supplement the Figure 4 comparisons with FOTCP (Ndiaye, 2025), conditional neural OT (CNOT; Kondratyev et al., 2026), and Individualized Whiten Metric OTCP (IWM-OTCP), following the reviewer's suggestions in W2 and Q1. In the heterogeneous simulations, IWM-OTCP estimates using the same covariance network as SCP, trained by minimizing the Gaussian negative log-likelihood. It then fits a conditional Euclidean GradNet OT map to the whitened scores and calibrates the transport radius following the split OTCP procedure of Thurin et al. (2025). Each entry reports Volume / Worst-case coverage / , averaged over 100 paired repetitions.
| Method | -mixture | -clover | -Exp |
|---|---|---|---|
| SCP | 4.147 / 0.789 / 0.033 | 11.196 / 0.784 / 0.046 | 11.505 / 0.794 / 0.032 |
| MVCS | 4.121 / 0.770 / 0.045 | 16.288 / 0.741 / 0.071 | 11.121 / 0.743 / 0.068 |
| RLCP | 4.595 / 0.797 / 0.032 | 15.753 / 0.793 / 0.040 | 12.048 / 0.802 / 0.030 |
| CNOT | 4.188 / 0.715 / 0.083 | 16.943 / 0.748 / 0.070 | 13.295 / 0.723 / 0.078 |
| FOTCP | 4.415 / 0.704 / 0.087 | 18.011 / 0.743 / 0.070 | 15.636 / 0.709 / 0.088 |
| IWM-OTCP | 4.222 / 0.773 / 0.033 | 11.636 / 0.778 / 0.041 | 12.379 / 0.789 / 0.031 |
| IMV-OTCP | 4.031 / 0.783 / 0.026 | 8.871 / 0.795 / 0.023 | 7.263 / 0.793 / 0.028 |
Compared with FOTCP and IWM-OTCP, IMV-OTCP consistently achieves smaller prediction-set volumes and improves both coverage diagnostics across all three settings.
Moreover, we extend Table 1 to include IWM-OTCP, CNOT, FOTCP, and additional IMV-OTCP variants. Our IMV-OTCP framework is not tied to a particular OT estimation method. Following the reviewer's suggestion in Q3(a), we incorporate the KDE-free Energy Score approach of Kan et al. (2022), which learns a convex transport potential directly from samples. We evaluate IMV-OTCP using both GradNet-based and Energy Score-based OT estimation, denoted GradNet, learned and Energy Score, learned, respectively, along with their counterparts, GradNet, and Energy Score, . IWM-OTCP uses anchor-specific kNN covariance whitening followed by GradNet OT. Each entry reports Root volume / at 90% target coverage, averaged over 100 splits per dataset.
| Dataset () | GradNet, learned | GradNet, | Energy Score, learned | Energy Score, | IWM-OTCP | CNOT | FOTCP |
|---|---|---|---|---|---|---|---|
| Energy (2) | 1.434 / 0.092 | 1.467 / 0.090 | 1.358 / 0.088 | 1.393 / 0.089 | 1.647 / 0.096 | 1.463 / 0.094 | 1.971 / 0.098 |
| Bias (2) | 2.120 / 0.039 | 2.159 / 0.042 | 2.511 / 0.066 | 2.521 / 0.066 | 2.202 / 0.040 | 2.270 / 0.043 | 2.217 / 0.040 |
| Taxi (2) | 3.319 / 0.029 | 3.324 / 0.029 | 3.579 / 0.028 | 3.578 / 0.028 | 3.708 / 0.034 | 3.359 / 0.027 | 3.395 / 0.036 |
| rf1 (8) | 1.224 / 0.054 | 1.344 / 0.058 | 1.000 / 0.051 | 1.011 / 0.051 | 1.225 / 0.053 | 1.513 / 0.053 | 3.092 / 0.051 |
| scm1d (16) | 1.879 / 0.063 | 1.903 / 0.064 | 2.052 / 0.050 | 2.112 / 0.044 | 1.985 / 0.055 | 2.108 / 0.052 | 5.611 / 0.059 |
IMV-OTCP achieves the smallest root volume among the compared OT-based methods across all five datasets, using either GradNet or Energy Score fitting. In particular, the learned-metric variants generally outperform their counterparts and covariance-whitened IWM-OTCP, demonstrating the benefit of adapting the transport metric.
Q1
In Example 3.1 and Figure 1, is suboptimal because transforms both and the spherical reference , while Appendix F.4 (p. 36) shows that discrete -assignment is equivalent to Euclidean OT on linearly transformed scores . If we whiten the scores via and run Euclidean FOTCP () with spherical , what are the set volumes and symmetric differences in Figures 1, 2, 3 and 5? Why not pursue this approach?
Response. Following the reviewer's suggestion, we add Whiten Metric OTCP (WM-OTCP) as a baseline to Figures 2, 3, and 5 to directly assess whether covariance whitening can achieve comparable performance to learned metrics. Specifically, WM-OTCP applies Euclidean FOTCP to the whitened scores , where is the sample covariance of the estimation residuals. As shown in Appendix F.4, this is equivalent, in terms of empirical OT assignment, to MV-OTCP with a fixed Mahalanobis metric . The updated results are reported below. Since all methods maintain nominal marginal coverage, we report Volume / HDR symmetric difference, averaged over 100 repetitions. The smallest volumes are in bold, and the second-smallest are underlined.
| Setting | FOTCP | MVCS | WM-OTCP | MV-OTCP | |
|---|---|---|---|---|---|
| Five-star | 2 | 10.42/3.51 | 11.64/4.92 | 10.32/3.37 | /3.38 |
| Five-star | 4 | 10.97/3.92 | 11.75/4.97 | 10.55/3.43 | 10.55/3.45 |
| Five-star | 6 | 11.42/4.23 | 11.85/5.01 | 10.75/3.52 | /3.59 |
| Five-star | 8 | 11.87/4.55 | 11.96/4.98 | 10.94/3.58 | /3.62 |
| Five-star | 10 | 12.28/4.82 | 12.08/4.95 | 11.16/3.69 | /3.69 |
| Clover | 2 | 10.35/2.76 | 10.64/3.22 | 10.24/2.67 | /2.68 |
| Clover | 4 | 10.81/3.10 | 10.79/3.20 | 10.42/2.70 | /2.74 |
| Clover | 6 | 11.25/3.42 | 10.93/3.24 | 10.63/2.76 | /2.81 |
| Clover | 8 | 11.67/3.75 | 11.09/3.21 | 10.84/2.83 | /2.86 |
| Clover | 10 | 12.10/4.06 | /3.28 | 11.05/2.88 | 11.05/2.95 |
WM-OTCP and MV-OTCP achieve nearly identical volumes in these settings. This similarity is largely explained by the particular structure of the simulated distributions, rather than by any general optimality of covariance whitening. Before the linear stretch and the small isotropic noise, write a score in polar coordinates as . Here is the radius of the shape in that direction for . The score density is This has the same decreasing radial profile as the Gaussian case of Example 3.1, with a direction-dependent radius that gives the contours their lobes. The fivefold and fourfold rotational symmetries give the unstretched laws covariance . After the stretch and noise, the covariance is . Consequently, covariance whitening effectively removes the dominant global anisotropy, while the nonlinear OT map captures the remaining non-elliptical geometry. This explains why WM-OTCP performs comparably to MV-OTCP in these particular settings. For general score distributions, however, covariance whitening has no volume-optimality guarantee, motivating the direct optimization of the transport metric.
New experiment. For simplicity, we set the regression mean to zero, so that . Let , where and are independent. Define and , and generate $\varepsilon=\sqrt{2}\operatorname{diag}(\rho^{-1/2},\rho^{1/2}) \left[ U+24,\operatorname{sign}(U_1)\mathbf 1{|U_1|>0.94}e_1 +2,\mathbf 1{v(U)^\top U>0.9}v(U) \right].$ The table reports the Volume / HDR symmetric difference over 100 repetitions for each setting. All methods maintain nominal marginal coverage, with mean coverage between 0.9003 and 0.9013.
| FOTCP | WM-OTCP | MV-OTCP | Volume reduction vs. WM | |
|---|---|---|---|---|
| 1.5 | 10.234/5.446 | 13.155/8.432 | 7.771/3.011 | 40.9% |
| 2.0 | 11.508/6.748 | 13.155/8.433 | 7.668/2.891 | 41.7% |
| 2.5 | 12.248/7.504 | 13.156/8.433 | 7.980/3.203 | 39.3% |
| 3.0 | 12.735/7.999 | 13.155/8.432 | 8.506/3.728 | 35.3% |
| 3.5 | 13.111/8.386 | 13.155/8.431 | 8.757/3.984 | 33.4% |
| 4.0 | 13.430/8.713 | 13.154/8.430 | 9.361/4.591 | 28.8% |
MV-OTCP consistently achieves the smallest volume and HDR symmetric difference across all six aspect ratios, reducing volume by 28.8%–41.7% relative to WM-OTCP. By directly optimizing prediction-set volume, MV-OTCP achieves substantially better efficiency, demonstrating the practical value of learning the transport metric beyond covariance whitening.
Thank you for your thoughtful review and positive assessment of our geometric characterization of prediction-set volume and task-aware losses. We appreciate your constructive comments regarding the novelty of metric learning, its relationship to existing OT methods, and the practical benefits of learning the ground cost. We would like to emphasize that our main contribution is to explicitly optimize the transport geometry for prediction-set volume or downstream decision loss, rather than simply replacing the Euclidean cost with a Mahalanobis cost. Below, we address each concern in detail and provide additional theoretical clarifications and experimental evidence.
W1
Response. The role of M is to modify the OT cost, rather than merely increase the representational capacity of the transport map. For a fixed M, Ψθ(M) approximates the corresponding Brenier map in the transformed coordinates. Although a sufficiently flexible neural network can represent the maps associated with different metrics, optimizing a network under the fixed Euclidean OT cost still targets the unique Euclidean Brenier map. In contrast, learning M changes the OT objective and hence the optimal coupling, which cannot generally be achieved through network reparameterization alone.
In MV-OTCP, the neural transport map Ψθ(M) acts as a nuisance component that provides a differentiable surrogate for optimizing M. Once the metric is learned, the final full-conformal prediction regions are constructed directly from the empirical OT assignment under the learned cost, without requiring the fitted neural map at calibration time. These prediction regions admit a representation as unions of polyhedral cells.
W2
Response. As shown in Appendix F.4, OT with cost cM(s,u)=(s−u)⊤M(s−u) is equivalent to Euclidean OT on the transformed scores S=MS, with the reference held fixed. Example 3.1 establishes that M∝Σ−1/2 minimizes prediction-set volume for radially decreasing elliptical score distributions, provided the metric is admissible. For general score distributions, however, covariance whitening is not guaranteed to be volume-optimal. We have therefore added whitening-based OT baselines, with the corresponding results reported in W3 and Q1.
The gains from metric learning depend on the underlying distribution and the number of anchors. While improvements are modest in some configurations, Figure 6 shows volume reductions of approximately 10% on X-Exp setting for certain anchor counts, demonstrating that metric learning can provide additional benefits beyond localized OT adaptation.