acceptodds
Under review as a conference paper at ICLR 2027

Minimum Volume Optimal Transport for Multivariate Conformal Prediction

10 views

Abstract

Optimal transport (OT) enables multivariate conformal prediction (CP) to construct flexible prediction sets that reflect the underlying complex distributional geometry while retaining distribution-free coverage guarantees. Existing OT-based methods typically rely on a fixed Euclidean transport cost, which can be suboptimal in terms of the volume of prediction sets. We propose Minimum-Volume OTCP, a framework that learns a Mahalanobis ground cost to directly minimize prediction-region volume over a class of positive-definite metrics. Using a metric Brenier representation, we derive a volume-based objective that depends explicitly on the Mahalanobis cost matrix and optimize it jointly with the transport map. We establish finite-sample marginal coverage of the output prediction sets and optimality of volume metrics. We further develop an individualized extension that learns covariate-dependent transport metrics, adapting the shape and orientation of prediction sets to heterogeneous conditional distributions. For this extension, we also show the asymptotic conditional coverage of our method. Simulations and real-data experiments demonstrate substantial reductions in prediction-set volume.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
conformal prediction, multivariate response, optimal transport, metric learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CV8LOpenReview ↗
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 cM(s,u)=(s−u)⊤M(s−u)c_M(s, u) = (s − u)^⊤M(s − u) 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 ∇gM(⋅)\nabla g_M(\cdot) 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 (d=2,8,16d = 2, 8, 16), and downstream robust optimization tasks evaluate the proposed methods against other benchmarks.

Critical strengths
  1. 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 cM(s,u)=(s−u)⊤M(s−u)c_M(s, u) = (s - u)^\top M(s - u) alters the OT pullback region TM−1(B1−α)T_M^{-1}(B_{1-\alpha}). Example 3.1 (p. 4, proved in Appendix A.1) nicely illustrates that for an elliptical score law with dispersion Σ\Sigma and a fixed spherical reference Q\mathcal{Q}, the volume-optimal quadratic ground cost is Mvol⋆∝Σ−1/2M_{\text{vol}}^\star \propto \Sigma^{-1/2} rather than M=IM = I or M∝Σ−1M \propto \Sigma^{-1}. Furthermore, writing the volume and downstream robust decision losses as expectations over the fixed reference ball B1−α\mathbb{B}_{1-\alpha} 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.

  2. 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 d=16d = 16), and portfolio optimization tasks.

Critical weaknesses
  1. It is not well justified why a separate linear matrix MM is needed when the reverse transport map GM(u)=M−1/2∇gM(M1/2u)G_M(u) = M^{-1/2}\nabla g_M(M^{1/2}u) is already approximated using a flexible neural network Ψθ\Psi_\theta for ∇gM(⋅)\nabla g_M(\cdot) (which is then discarded at calibration time in MV-OTCP, Eq. 7), nor why the linear transformation induced by MM cannot effectively be absorbed into the learned transport representation.

  2. The matrix MM only globally rotates and rescales the score difference s−us - u (which, as shown in Appendix F.4, p. 36, is mathematically equivalent to running standard Euclidean OT on linearly transformed scores Zi=MSiZ_i = M S_i). It is unclear why this simple linear reweighting is preferable to basic score whitening (S~=Σ^−1/2S\widetilde{S} = \widehat{\Sigma}^{-1/2} S), especially since the localized ablation in Appendix D.2 (Figure 6) shows that learning MxM_x reduces volume by only 1%–3% over fixing M=IM = I once local OT adaptation is present.

  3. 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 ∥s−u∥2\|s - u\|^2 with the quadratic Mahalanobis cost (s−u)⊤M(s−u)(s - u)^\top M (s - u), and the main real-data and heterogeneous benchmarks (Figure 4, Table 1) omit these OT-based baselines.

Questions for authors
  1. In Example 3.1 and Figure 1, M∝Σ−1M \propto \Sigma^{-1} is suboptimal because cM(s,u)=(s−u)⊤M(s−u)c_M(s, u) = (s - u)^\top M(s - u) transforms both ss and the spherical reference u∼Qu \sim \mathcal{Q}, while Appendix F.4 (p. 36) shows that discrete cMc_M-assignment is equivalent to Euclidean OT on linearly transformed scores Zi=MSiZ_i = M S_i. If we simply standardize/whiten the scores via s~=Σ^−1/2(s−m^)\tilde{s} = \widehat{\Sigma}^{-1/2}(s - \widehat{m}) and run standard Euclidean FOTCP (∥s~−u∥22\|\tilde{s} - u\|_2^2) with spherical u∼Qu \sim \mathcal{Q}, what are the resulting set volumes and symmetric differences in Figures 1, 2, 3, and 5? Any specific reason we are not pursuing this?

  2. 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 Mx=IM_x = I) omitted from Figure 4 and Table 1? Can you report the performance of IMV-OTCP(Mx=IM_x = I) and IMV-OTCP(Mx∝Σ^x−1/2M_x \propto \widehat{\Sigma}_x^{-1/2}) across all the real datasets in Table 1 so we can see how much of the volume reduction actually comes from learning MxM_x versus the localized GradNet OT pipeline?

  3. By the Monge–Ampère identity (Eq. 10 and Eq. 24), the volume objective J^(M,θ)\widehat{\mathcal{J}}(M, \theta) evaluates the inverse of a dd-dimensional Gaussian KDE f^\hat{f} on the fixed origin-centered ball B1−αB_{1-\alpha}.

  • (a) How reliable is this KDE-based bilevel objective when d=8d = 8 (rf1) or d=16d = 16 (scm1d) with n≤1000n \leq 1000, and why use a KDE-dependent Monge–Ampère loss + 2L2L 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 1/fP(GM(u))1 / f_{\mathcal{P}}(G_M(u)), why fix the reference acceptance set to the origin-centered ball B1−α\mathbb{B}_{1-\alpha} (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)?
  1. In MV-OTCP (Section 3.3), you fit a continuous Brenier map Ψθ^\Psi_{\widehat{\theta}} 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 Leb⁡{C^MV(x)}\operatorname{Leb}\{\widehat{\mathcal{C}}^{\text{MV}}(x)\} 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(M=IM = I), and IMV-OTCP; fix the typos on p. 14 line 752 ("for convinence") and p. 28 line 1511 (Y=g(X)+…Y = g(X) + \dots should be m(X)m(X)).

8 comments

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

It is not well justified why a separate linear matrix MM is needed when the reverse transport map GM(u)=M−1/2∇gM(M1/2u)G_M(u)=M^{-1/2}\nabla g_M(M^{1/2}u) is already approximated using a flexible neural network Ψθ\Psi_\theta for ∇gM(⋅)\nabla g_M(\cdot) (which is then discarded at calibration time in MV-OTCP, Eq. 7), nor why the linear transformation induced by MM cannot effectively be absorbed into the learned transport representation.

Response. The role of MM is to modify the OT cost, rather than merely increase the representational capacity of the transport map. For a fixed MM, Ψθ(M)\Psi_{\theta(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 MM 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)\Psi_{\theta(M)} acts as a nuisance component that provides a differentiable surrogate for optimizing MM. 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

The matrix MM only globally rotates and rescales the score difference s−us-u (which, as shown in Appendix F.4, p. 36, is mathematically equivalent to running standard Euclidean OT on linearly transformed scores Zi=MSiZ_i=MS_i). It is unclear why this simple linear reweighting is preferable to basic score whitening (S~=Σ^−1/2S\widetilde S=\widehat\Sigma^{-1/2}S), especially since the localized ablation in Appendix D.2 (Figure 6) shows that learning MxM_x reduces volume by only 1%–3% over fixing M=IM=I once local OT adaptation is present.

Response. As shown in Appendix F.4, OT with cost cM(s,u)=(s−u)⊤M(s−u)c_M(s,u)=(s-u)^\top M(s-u) is equivalent to Euclidean OT on the transformed scores S~=MS\widetilde S=MS, with the reference held fixed. Example 3.1 establishes that M∝Σ−1/2M\propto\Sigma^{-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 XX-Exp setting for certain anchor counts, demonstrating that metric learning can provide additional benefits beyond localized OT adaptation.

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 ∣s−u∣2|s-u|^2 with the quadratic Mahalanobis cost (s−u)⊤M(s−u)(s-u)^\top M(s-u), 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 GM(B1−α)G_M(\mathbb{B}_{1-\alpha}) 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 Σ^x\widehat\Sigma_x 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 S~=Σ^x−1/2S\widetilde S=\widehat\Sigma_x^{-1/2}S and calibrates the transport radius following the split OTCP procedure of Thurin et al. (2025). Each entry reports Volume / Worst-case coverage / MSCE\sqrt{\mathrm{MSCE}}, averaged over 100 paired repetitions.

MethodXX-mixtureXX-cloverXX-Exp
SCP4.147 / 0.789 / 0.03311.196 / 0.784 / 0.04611.505 / 0.794 / 0.032
MVCS4.121 / 0.770 / 0.04516.288 / 0.741 / 0.07111.121 / 0.743 / 0.068
RLCP4.595 / 0.797 / 0.03215.753 / 0.793 / 0.04012.048 / 0.802 / 0.030
CNOT4.188 / 0.715 / 0.08316.943 / 0.748 / 0.07013.295 / 0.723 / 0.078
FOTCP4.415 / 0.704 / 0.08718.011 / 0.743 / 0.07015.636 / 0.709 / 0.088
IWM-OTCP4.222 / 0.773 / 0.03311.636 / 0.778 / 0.04112.379 / 0.789 / 0.031
IMV-OTCP4.031 / 0.783 / 0.0268.871 / 0.795 / 0.0237.263 / 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 Mx=IM_x=I counterparts, GradNet, Mx=IM_x=I and Energy Score, Mx=IM_x=I. IWM-OTCP uses anchor-specific kNN covariance whitening followed by GradNet OT. Each entry reports Root volume / MSCE\sqrt{\mathrm{MSCE}} at 90% target coverage, averaged over 100 splits per dataset.

Dataset (dd)GradNet, learnedGradNet, Mx=IM_x=IEnergy Score, learnedEnergy Score, Mx=IM_x=IIWM-OTCPCNOTFOTCP
Energy (2)1.434 / 0.0921.467 / 0.0901.358 / 0.0881.393 / 0.0891.647 / 0.0961.463 / 0.0941.971 / 0.098
Bias (2)2.120 / 0.0392.159 / 0.0422.511 / 0.0662.521 / 0.0662.202 / 0.0402.270 / 0.0432.217 / 0.040
Taxi (2)3.319 / 0.0293.324 / 0.0293.579 / 0.0283.578 / 0.0283.708 / 0.0343.359 / 0.0273.395 / 0.036
rf1 (8)1.224 / 0.0541.344 / 0.0581.000 / 0.0511.011 / 0.0511.225 / 0.0531.513 / 0.0533.092 / 0.051
scm1d (16)1.879 / 0.0631.903 / 0.0642.052 / 0.0502.112 / 0.0441.985 / 0.0552.108 / 0.0525.611 / 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 Mx=IM_x=I counterparts and covariance-whitened IWM-OTCP, demonstrating the benefit of adapting the transport metric.

Q1

In Example 3.1 and Figure 1, M∝Σ−1M\propto\Sigma^{-1} is suboptimal because cM(s,u)=(s−u)⊤M(s−u)c_M(s,u)=(s-u)^\top M(s-u) transforms both ss and the spherical reference u∼Qu\sim\mathcal Q, while Appendix F.4 (p. 36) shows that discrete cMc_M-assignment is equivalent to Euclidean OT on linearly transformed scores Zi=MSiZ_i=MS_i. If we whiten the scores via s~=Σ^−1/2(s−m^)\tilde s=\widehat\Sigma^{-1/2}(s-\widehat m) and run Euclidean FOTCP (∣s~−u∣22|\tilde s-u|_2^2) with spherical u∼Qu\sim\mathcal Q, 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 S~=Σ^−1/2S\widetilde S=\widehat\Sigma^{-1/2}S, where Σ^\widehat\Sigma 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 M∝Σ^−1/2M\propto\widehat\Sigma^{-1/2}. 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.

Settingρ\rhoFOTCPMVCSWM-OTCPMV-OTCP
Five-star210.42/3.5111.64/4.9210.32/3.3710.34‾\underline{10.34}/3.38
Five-star410.97/3.9211.75/4.9710.55/3.4310.55/3.45
Five-star611.42/4.2311.85/5.0110.75/3.5210.79‾\underline{10.79}/3.59
Five-star811.87/4.5511.96/4.9810.94/3.5810.99‾\underline{10.99}/3.62
Five-star1012.28/4.8212.08/4.9511.16/3.6911.18‾\underline{11.18}/3.69
Clover210.35/2.7610.64/3.2210.24/2.6710.26‾\underline{10.26}/2.68
Clover410.81/3.1010.79/3.2010.42/2.7010.44‾\underline{10.44}/2.74
Clover611.25/3.4210.93/3.2410.63/2.7610.66‾\underline{10.66}/2.81
Clover811.67/3.7511.09/3.2110.84/2.8310.87‾\underline{10.87}/2.86
Clover1012.10/4.0611.24‾\underline{11.24}/3.2811.05/2.8811.05/2.95

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 s=r(cos⁡θ,sin⁡θ)s=r(\cos\theta,\sin\theta). Here rj(θ)r_j(\theta) is the radius of the shape in that direction for j∈{Five-star,Clover}j\in\{\text{Five-star},\text{Clover}\}. The score density is fj(s) ∝exp⁡{−(rrj(θ))2/2}.f_j(s)\ \propto \exp\{-(\frac{r}{r_j(\theta)})^2/2\}. 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 σj2I\sigma_j^2 I. After the stretch AjA_j and noise, the covariance is σj2AjAj⊤+0.0152I\sigma_j^2 A_j A_j^\top+0.015^2 I. 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 Y=εY=\varepsilon. Let U=R(cos⁡Θ,sin⁡Θ)⊤U=R(\cos\Theta,\sin\Theta)^\top, where R∼Unif(0,1)R\sim\mathrm{Unif}(0,1) and Θ∼Unif(0,2π)\Theta\sim\mathrm{Unif}(0,2\pi) are independent. Define v(U)=(sign⁡(U1)/4,15sign⁡(U2)/4)⊤v(U)=(\operatorname{sign}(U_1)/4,\sqrt{15}\operatorname{sign}(U_2)/4)^\top and e1=(1,0)⊤e_1=(1,0)^\top, 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.

ρ\rhoFOTCPWM-OTCPMV-OTCPVolume reduction vs. WM
1.510.234/5.44613.155/8.4327.771/3.01140.9%
2.011.508/6.74813.155/8.4337.668/2.89141.7%
2.512.248/7.50413.156/8.4337.980/3.20339.3%
3.012.735/7.99913.155/8.4328.506/3.72835.3%
3.513.111/8.38613.155/8.4318.757/3.98433.4%
4.013.430/8.71313.154/8.4309.361/4.59128.8%

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.

Reviewer_Ec8aOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Recent work on multivariate conformal prediction uses optimal transport (OT) to rank vector-valued nonconformity scores: scores are mapped to a reference distribution (e.g., uniform on the unit ball), a quantile ball is calibrated there, and the prediction region is its pullback. Existing methods fix the squared Euclidean cost, and the resulting regions need not have minimal volume.

The paper proposes to learn the ground cost instead. It uses a Mahalanobis cost cM(s,u)=(s−u)⊤M(s−u)c_M(s,u) = (s-u)^\top M (s-u) with M≻0M \succ 0, det⁡M=1\det M = 1 and bounded eigenvalues. A metric Brenier lemma shows that the cMc_M-optimal map is a Euclidean Brenier map after the change of variables s↦M1/2ss \mapsto M^{1/2} s. This expresses the region's volume as an expectation, over the fixed reference ball, of the Hessian determinant of the transport potential, which is then minimized over MM. In practice, the map is fitted with GradNetOT and the profiled objective is differentiated through the inner OT fit by implicit differentiation. The framework extends to task-aware decision losses, with volume as a special case. Two methods result: MV-OTCP (one global metric, full-conformal discrete assignment) and IMV-OTCP (covariate-dependent local metrics and maps, split calibration).

Theoretically, the paper establishes finite-sample marginal coverage for both methods, consistency of the learned metric under a high-level transport-consistency assumption, asymptotic agreement of the MV-OTCP region with the oracle region, and asymptotic conditional coverage for IMV-OTCP. For elliptical scores, the volume-optimal metric is shown to be M∝Σ−1/2M \propto \Sigma^{-1/2}, which recovers the minimum-volume density level set.

Experiments include 2-D simulations and five regression benchmarks with response dimension 2 to 16, against SCP, MVCS, FOTCP and RLCP. All methods attain nominal marginal coverage; the proposed methods produce the smallest regions in most settings (all real datasets except rf1), with gains over Euclidean-cost OTCP increasing with anisotropy.

Critical strengths
  1. The paper addresses an important and actively discussed issue: the volume efficiency of OT-based conformal prediction regions. It identifies the fixed squared Euclidean cost as a hidden design choice and shows, analytically (for elliptical scores the volume-optimal metric is M∝Σ−1/2M \propto \Sigma^{-1/2}, not M=IM = I) and experimentally, that it is volume-suboptimal. The proposed fix is elegant: the metric Brenier representation turns region volume into an explicit, differentiable objective in MM (a Hessian determinant over a fixed reference ball), while coverage is untouched because the metric is learned on separate data.

  2. The paper is well written and communicates the method, theory and experiments clearly. Finite-sample marginal coverage holds for both variants, and consistency and oracle-recovery results are provided. Empirically, the methods keep nominal coverage while giving the smallest regions in most settings (4 of 5 real datasets), and the gain over Euclidean-cost OTCP grows with anisotropy, which is exactly the regime the theory predicts. This matters because it indicates the improvement comes from the proposed mechanism rather than from tuning.

Critical weaknesses
  1. Learning MM requires a bilevel procedure: a neural OT map (GradNetOT) fitted through a KDE-based Monge–Ampère loss, implicit differentiation through that fit, and, for IMV-OTCP, a second network with inverse and cycle-consistency losses refitted at each anchor. More importantly, the paper's own Example 3.1 suggests a closed-form plug-in, M^∝Σ^−1/2\hat M \propto \hat\Sigma^{-1/2}, but OT-based CP with this whitening metric is not evaluated as a baseline in the experiments (it appears only in the illustration of Figure 1), and runtimes are not reported. It is unclear whether the gains justify the added complexity, or whether most of them would be obtained by simply whitening the scores before Euclidean OT.

  2. The prediction regions themselves can be non-elliptical, since they are pullbacks under a nonlinear OT map. However, the metric MM only rescales and rotates the score space globally. For distributions whose geometry is curved or varies across the score space (e.g., banana-shaped scores), no single MM aligns with the local orientation, and the benefit over the Euclidean cost should largely vanish.

Questions for authors

I thank authors for their work and find contribution very interesting. My questions are mostly methodological:

  1. For elliptical score distributions, ellipsoidal conformal methods (Messoudi et al., 2022) and conformalized Gaussian scoring (Braun et al., 2025) already recover the HDR set, and their localized versions adapt the shape and orientation of the region to xx. The experiments include only the global Mahalanobis score (SCP) and MVCS. Could the authors compare with localized ellipsoidal CP (e.g., nearest-neighbour covariance estimates as in Messoudi et al., 2022) and with Braun et al. (2025), particularly in the heterogeneous settings? This would isolate what the nonlinear transport adds beyond an adaptive ellipsoid.

  2. The learned cost appears closely related to Euclidean OT applied after whitening. If the scores are whitened, s~=Σ^−1/2(s−m^)\tilde s = \hat\Sigma^{-1/2}(s - \hat m), and a Euclidean OT map from s~\tilde s to QQ is fitted, the pulled-back region already coincides with the HDR ellipsoid in the elliptical setting of Example 3.1. Could the authors compare MV-OTCP with this plug-in procedure, and IMV-OTCP with its localized analogue in which m^(x)\hat m(x) and Σ^(x)\hat\Sigma(x) are estimated from nearest neighbours. If whitening recovers most of the volume reduction, the case for the bilevel optimization is considerably weaker.

  3. The method already relies on a kernel density estimate of the transformed scores to fit GradNetOT. Given this, density-based conformal sets that directly target HDR level sets, such as KDE level-set prediction sets HPD-split (Izbicki et al., 2022) is a natural baseline. How does the proposed method compare with them in volume and conditional coverage?

3 comments

We thank the reviewer for the positive assessment of our work and the constructive suggestions. In response, we have expanded our experiments to include comparisons with covariance whitening, adaptive ellipsoidal conformal methods, and density-level-set approaches, providing a more comprehensive evaluation of the benefits of our proposed framework. Below, we address each concern in detail and provide clarifications and additional numerical results.

W1

Learning MM requires a bilevel procedure: a neural OT map (GradNetOT) fitted through a KDE-based Monge–Ampère loss, implicit differentiation through that fit, and, for IMV-OTCP, a second network with inverse and cycle-consistency losses refitted at each anchor. More importantly, the paper's own Example 3.1 suggests a closed-form plug-in, M^∝Σ^−1/2\widehat M\propto\widehat\Sigma^{-1/2}, but OT-based CP with this whitening metric is not evaluated as a baseline in the experiments (it appears only in the illustration of Figure 1), and runtimes are not reported. It is unclear whether the gains justify the added complexity, or whether most of them would be obtained by simply whitening the scores before Euclidean OT.

Response. As shown in Appendix F.4, OT with ground cost cM(s,u)=(s−u)⊤M(s−u)c_M(s,u)=(s-u)^\top M(s-u) is equivalent to Euclidean OT on the transformed scores S~=MS\widetilde S=MS, with the reference distribution held fixed. Example 3.1 establishes that M∝Σ−1/2M\propto\Sigma^{-1/2} yields the minimum-volume HDR for radially decreasing elliptical score distributions, provided the metric is admissible. However, for general score distributions, covariance whitening has no volume-optimality guarantee.

To investigate this distinction, as suggested by the reviewer, we introduce Whiten Metric OTCP (WM-OTCP), which applies FOTCP (Ndiaye, 2025) to S~=Σ^−1/2S\widetilde S=\widehat\Sigma^{-1/2}S, where Σ^\widehat\Sigma is the sample covariance of the estimation residuals. We compare it with standard FOTCP and MV-OTCP, which directly learns MM by minimizing prediction-set volume. All three methods use the same empirical OT assignment and conformal calibration procedure.

To demonstrate the benefits of metric learning, we consider the following setting. For simplicity, we set the regression mean to zero and let Y=εY=\varepsilon. Generate U=R(cos⁡Θ,sin⁡Θ)⊤U=R(\cos\Theta,\sin\Theta)^\top, where R∼Unif(0,1)R\sim\mathrm{Unif}(0,1) and Θ∼Unif(0,2π)\Theta\sim\mathrm{Unif}(0,2\pi) are independent. Define v(U)=(sign⁡(U1),15sign⁡(U2))⊤/4v(U)=(\operatorname{sign}(U_1),\sqrt{15}\operatorname{sign}(U_2))^\top/4 and e1=(1,0)⊤e_1=(1,0)^\top, and set $\varepsilon=\sqrt{2}\operatorname{diag}(\rho^{-1/2},\rho^{1/2}) \left[ U+24,\operatorname{sign}(U_1)\mathbf 1{|U_1|>0.94}e_1 +2,\mathbf 1{v(U)^\top U>0.9}v(U) \right].$ The tail perturbations leave the central 90% mass unchanged before stretching but substantially affect the covariance, creating a discrepancy between covariance geometry and the target HDR. The table reports Volume / HDR symmetric difference over 100 repetitions. All methods maintain nominal marginal coverage. The smallest volume is marked in bold.

ρ\rhoFOTCPWM-OTCPMV-OTCPVolume reduction vs. WM
1.510.234 / 5.44613.155 / 8.4327.771 / 3.01140.9%
2.011.508 / 6.74813.155 / 8.4337.668 / 2.89141.7%
2.512.248 / 7.50413.156 / 8.4337.980 / 3.20339.3%
3.012.735 / 7.99913.155 / 8.4328.506 / 3.72835.3%
3.513.111 / 8.38613.155 / 8.4318.757 / 3.98433.4%
4.013.430 / 8.71313.154 / 8.4309.361 / 4.59128.8%

Across the values of ρ\rho considered above, MV-OTCP reduces volume by 28.8%–41.7% relative to WM-OTCP and achieves smaller HDR symmetric differences. These results demonstrate the effectiveness and importance of metric learning for more general distributions beyond the setting of Example 3.1.

In the single-thread benchmark at d=2d=2 and nest=1000n_{\rm est}=1000, mean fitting and calibration times are 3.15 s for MV-OTCP and 0.412 s for FOTCP. Although metric learning introduces additional computational overhead, the runtime remains manageable, reflecting a reasonable trade-off between prediction efficiency and computational cost.

For the individualized setting, we have additionally included Individualized Whiten Metric OTCP (IWM-OTCP) as a whitening-based baseline, with the corresponding results reported in Q2–Q3.

W2

The prediction regions themselves can be non-elliptical, since they are pullbacks under a nonlinear OT map. However, the metric MM only rescales and rotates the score space globally. For distributions whose geometry is curved or varies across the score space (e.g., banana-shaped scores), no single MM aligns with the local orientation, and the benefit over the Euclidean cost should largely vanish.

Response. We agree that a global Mahalanobis metric cannot directly align spatially varying local orientations. However, this does not imply that metric learning becomes ineffective for curved score distributions. The metric MM enters the OT cost and changes the optimal coupling, going beyond a global linear transformation of the prediction region. Specifically, the resulting region is TM−1(B1−α)T_M^{-1}(\mathbb B_{1-\alpha}), where the nonlinear transport map can capture curved and nonconvex geometries even with a fixed MM.

Nevertheless, we acknowledge that when the distributional geometry is dominated by spatially varying local orientations, with limited benefit from adjusting the global coupling, the additional gains from learning MM may be small. Extending our framework to spatially adaptive metrics that can capture such local geometric variations is an interesting direction for future work.

Q1

For elliptical score distributions, ellipsoidal conformal methods (Messoudi et al., 2022) and conformalized Gaussian scoring (Braun et al., 2025) already recover the HDR set, and their localized versions adapt the shape and orientation of the region to xx. The experiments include only the global Mahalanobis score (SCP) and MVCS. Could the authors compare with localized ellipsoidal CP (e.g., nearest-neighbour covariance estimates as in Messoudi et al., 2022) and with Braun et al. (2025), particularly in the heterogeneous settings? This would isolate what the nonlinear transport adds beyond an adaptive ellipsoid.

Response. We clarify that both SCP and MVCS in our original heterogeneous simulations already adapt their prediction regions to the covariates. Specifically, SCP uses a covariance network that outputs the Cholesky factor LxL_x of Σx=LxLx⊤\Sigma_x=L_xL_x^\top, trained by minimizing the Gaussian negative log-likelihood. It then conformalizes the standardized residual norm ∥Σ^x−1/2S∥2\|\widehat\Sigma_x^{-1/2}S\|_2, consistent with the approach of Braun et al. (2025). MVCS (Braun et al., 2026) similarly learns a covariate-dependent matrix for its transformed LpL_p score. Thus, our original experiments already include adaptive ellipsoidal and norm-based baselines.

The simulation results in Q2 show that IMV-OTCP improves both volume and MSCE\sqrt{\mathrm{MSCE}} over this adaptive SCP baseline in all three settings. Our SCP comparison uses a fitted covariance network. A separate nearest-neighbour ellipsoid would assess the effect of the covariance estimator in the ellipsoidal baseline.

Q2

The learned cost appears closely related to Euclidean OT applied after whitening. If scores are whitened, s~=Σ^−1/2(s−m^)\tilde s=\widehat\Sigma^{-1/2}(s-\widehat m), and a Euclidean OT map from s~\tilde s to QQ is fitted, the pulled-back region coincides with the HDR ellipsoid in the setting of Example 3.1. Could the authors compare MV-OTCP with this plug-in procedure, and IMV-OTCP with its localized analogue in which m^(x)\widehat m(x) and Σ^(x)\widehat\Sigma(x) are estimated from nearest neighbours. If whitening recovers most of the volume reduction, the case for the bilevel optimization is weaker.

Response. Please see W1 for the homogeneous comparison with WM-OTCP. In the heterogeneous simulations, as suggested by reviewer, we introduce Individualized Whiten Metric OTCP (IWM-OTCP), which first whitens the scores using a covariate-dependent covariance estimate, S~=Σ^x−1/2S\widetilde S=\widehat\Sigma_x^{-1/2}S. It then fits a conditional Euclidean OT map to the whitened scores and calibrates the transport radius following the split OTCP procedure of Thurin et al. (2025). We additionally include conditional neural OT (CNOT; Kondratyev et al., 2026) and FOTCP (Ndiaye, 2025).

The table reports Volume / Worst-case coverage / MSCE\sqrt{\mathrm{MSCE}} over 100 repetitions.

MethodXX-mixtureXX-cloverXX-Exp
SCP4.147/0.789/0.03311.196/0.784/0.04611.505/0.794/0.032
MVCS4.121/0.770/0.04516.288/0.741/0.07111.121/0.743/0.068
RLCP4.595/0.797/0.03215.753/0.793/0.04012.048/0.802/0.030
CNOT4.188/0.715/0.08316.943/0.748/0.07013.295/0.723/0.078
FOTCP4.415/0.704/0.08718.011/0.743/0.07015.636/0.709/0.088
IWM-OTCP4.222/0.773/0.03311.636/0.778/0.04112.379/0.789/0.031
IMV-OTCP4.031/0.783/0.0268.871/0.795/0.0237.263/0.793/0.028

IMV-OTCP achieves the smallest volume and the lowest MSCE\sqrt{\mathrm{MSCE}} in all three settings. Compared with IWM-OTCP, it also improves worst-case coverage in all three settings. The corresponding real-data results are reported in Q3.

Q3

The method already relies on a kernel density estimate of the transformed scores to fit GradNetOT. Given this, density-based conformal sets that directly target HDR level sets, such as KDE level-set prediction sets HPD-split (Izbicki et al., 2022) is a natural baseline. How does the proposed method compare with them in volume and conditional coverage?

Response. We clarify that our framework is not restricted to GradNetOT and can accommodate alternative transport estimators. To demonstrate this flexibility, we evaluate two implementations of IMV-OTCP, using GradNetOT and the KDE-free Energy Score approach of Kan et al. (2022) for transport estimation, labeled GradNet, learned and Energy Score, learned, respectively. The latter learns a convex transport potential directly from samples without requiring density estimation. We also include HPD-split (Izbicki et al., 2022), which estimates conditional density rankings to approximate highest-density regions (HDRs). As HPD-split targets population minimum-volume regions over a different prediction-set family, the comparison is informative but not strictly like-for-like. For the real-data IWM-OTCP baseline, we follow the reviewer's suggestion and construct local whitening metrics using kNN-based residual covariance estimates at the nearest anchor.

The table reports Root volume / MSCE\sqrt{\mathrm{MSCE}} at 90% target coverage with 100 random splits. The smallest volumes are in bold, and the second-smallest are underlined.

Dataset (dd)GradNet, learnedEnergy Score, learnedMVCSIWM-OTCPHPD-splitCNOTFOTCP
Energy (2)1.434 / 0.0921.358‾\underline{1.358} / 0.0881.337 / 0.1011.647 / 0.0961.388 / 0.0931.463 / 0.0941.971 / 0.098
Bias (2)2.120‾\underline{2.120} / 0.0392.511 / 0.0662.177 / 0.0422.202 / 0.0402.105 / 0.0392.270 / 0.0432.217 / 0.040
Taxi (2)3.319‾\underline{3.319} / 0.0293.579 / 0.0283.492 / 0.0293.708 / 0.0343.236 / 0.0263.359 / 0.0273.395 / 0.036
rf1 (8)1.224‾\underline{1.224} / 0.0541.000 / 0.0511.286 / 0.0611.225 / 0.0531.336 / 0.0601.513 / 0.0533.092 / 0.051
scm1d (16)1.879 / 0.0632.052 / 0.0502.076 / 0.0711.985 / 0.0551.921‾\underline{1.921} / 0.0652.108 / 0.0525.611 / 0.059

The proposed IMV-OTCP framework achieves the smallest root volume among the OT-based methods on all five datasets, and outperforms MVCS, which also targets prediction-set volume, on four of the five datasets. Notably, IMV-OTCP also achieves smaller volumes than HPD-split on Energy, rf1, and scm1d, despite the latter directly targeting the conditional HDR. Overall, these results demonstrate the competitive volume efficiency of our OT-based framework, particularly against existing OT methods and MVCS. They also show that KDE-free transport estimation can improve performance in higher-dimensional settings, further supporting the flexibility of our metric-learning framework.

Reference

[1] Ndiaye, E. (2025). Beyond Uncertainty Sets: Leveraging Optimal Transport to Extend Conformal Predictive Distribution to Multivariate Settings. arXiv preprint arXiv:2511.15146.

[2] Kondratyev, V., Fishkov, A., Hegazy, M., Kotelevskii, N., Flamary, R., Panov, M., & Moulines, E. (2026, April). Neural optimal transport meets multivariate conformal prediction. In International Conference on Learning Representations (Vol. 2026, pp. 77879-77917).

[3] Thurin, G., Nadjahi, K., & Boyer, C. (2025). Optimal transport-based conformal prediction. arXiv preprint arXiv:2501.18991.

[4] Kan, K., Aubet, F. X., Januschowski, T., Park, Y., Benidis, K., Ruthotto, L., & Gasthaus, J. (2022, May). Multivariate quantile function forecaster. In International Conference on Artificial Intelligence and Statistics (pp. 10603-10621). PMLR.

[5] Izbicki, R., Shimizu, G., & Stern, R. B. (2022). Cd-split and hpd-split: Efficient conformal regions in high dimensions. Journal of Machine Learning Research, 23(87), 1-32.

[6] Braun, S., Berta, E., Jordan, M. I., & Bach, F. (2025). Multivariate conformal prediction via conformalized gaussian scoring. arXiv e-prints, arXiv-2507.

[7] Braun, S., Aolaritei, L., Jordan, M. I., & Bach, F. (2026). Minimum volume conformal sets for multivariate regression. Journal of the American Statistical Association, (just-accepted), 1-27.

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

The paper studies optimal-transport-based multivariate conformal prediction (OTCP). Existing OTCP methods rank residual scores with the squared-Euclidean transport cost. The authors argue that this cost can produce needlessly large prediction regions. They propose MV-OTCP, which replaces the cost with a Mahalanobis cost c_M(s,u) = (s-u)^T M (s-u), where M is positive-definite with det(M) = 1 and bounded eigenvalues. M is chosen to minimize the Lebesgue volume of the resulting OT prediction region.

  • Volume objective (Section 3.2). A "metric Brenier" reduction (Lemma 3.1) writes the volume as an expectation over the reference law of a Hessian determinant (Eq. 4). Example 3.1 shows that for elliptical scores the volume-optimal metric is proportional to Sigma^{-1/2}.
  • Estimation (Section 3.3, Appendix A.2). Empirically, the reference-to-score map is a GradNet fitted with a Monge–Ampère residual loss on a KDE (Eq. 10). M is updated with implicit differentiation through this inner fit (Algorithm 1). The learned M is then plugged into the full-conformal assignment of FOTCP (Eq. 7).
  • Individualized variant (Section 4, Appendix A.4). IMV-OTCP learns covariate-dependent metrics and maps at kernel-weighted anchors, then uses split-conformal calibration of the transported radius.
  • Task-aware variant (Section 3.2, Appendix A.3). TA-OTCP replaces volume with an integrated downstream decision loss.
  • Theory. The paper proves finite-sample marginal coverage (Theorems 3.1, 4.1, C.1, C.2). It gives a uniform deviation bound and consistency of the learned metric under high-level assumptions (Theorem 3.2, Assumption 3.1; Theorem B.1). It shows the learned region converges to the oracle region (Theorem 3.3) and gives asymptotic conditional coverage of IMV-OTCP (Theorem 4.2, Assumption 4.1).
  • Experiments. Bivariate homogeneous and heterogeneous simulations compare against SCP, MVCS, FOTCP and RLCP (Figures 2–4, with aspect-ratio and anchor ablations in Appendix D). Five real regression datasets compare IMV-OTCP against SCP, MVCS and RLCP (Table 1). Task-aware experiments include a portfolio study (Appendix D.3–D.4).

The reported results support three claims: finite-sample marginal validity, smaller volume than FOTCP in the bivariate homogeneous simulations, and smaller root volume than the reported baselines on four of five real datasets. The claimed consistency and conditional-coverage properties rest on high-level assumptions that are not verified for the implemented estimator.

Critical strengths
  1. A natural, well-motivated idea with a clean characterization. Treating the OT ground cost as a learnable object optimized for set volume (Section 3.2) fits well with recent OT-based conformal methods. Lemma 3.1 and the change-of-variables formula (Eq. 4, and its equivalent form in Eq. 24, J(M) = E_Q[1{U in B} / f_P(G_M(U))]) turn the implicit dependence on M into a computable objective. Example 3.1 and Figure 1 give a concrete case where the volume-optimal metric is known in closed form (Sigma^{-1/2}) and differs from both the identity and Sigma^{-1}. This is a useful insight for practitioners using OTCP.
  2. Coverage guarantees are kept and are correctly argued. The metric and maps are fitted on data independent of calibration. The marginal-coverage proofs (Appendix F.2, F.5, F.9, F.10) are standard exchangeability arguments and appear correct, including the weighted, randomized-center full-conformal localized construction in Appendix C. The polyhedral representation (Proposition B.1, Proposition C.1) correctly extends the FOTCP cell structure to Mahalanobis costs.
  3. Controlled synthetic evidence that the learned metric helps in the homogeneous setting. In the aspect-ratio ablations (Figure 3, Figure 5), only the anisotropy of the residual law varies while its area is held fixed (Appendix D.1). The paired improvement of MV-OTCP over FOTCP grows with the aspect ratio. This directly isolates the claimed mechanism. Coverage stays near nominal throughout.
Critical weaknesses
  1. The real-data evidence does not isolate the contribution of metric learning, and important baselines are missing.
    • Table 1 compares IMV-OTCP only with SCP, MVCS and RLCP. It omits the OT baselines the method builds on (FOTCP, the split OTCP of Thurin et al. 2025 / Klein et al. 2025, conditional neural OT of Kondratyev et al. 2026) and the IMV-OTCP(M = I) ablation. It also omits other flexible multivariate CP methods the paper itself cites (e.g., Feldman et al. 2023, Fang et al. 2025, Izbicki et al. 2022, the scores benchmarked in Dheur et al. 2025).
    • The real-data configuration uses only four outer metric updates with learning rate 0.02 (Appendix E.4), and the initialization of B is not stated. M may therefore stay close to its starting value, and the volume gains in Table 1 may come mainly from localized OT ranking rather than from the learned cost.
    • The only localized ablation (Figure 6, Appendix D.2) is on bivariate simulations. There the improvement over M = I is a few percent on X-clover and can be negative on X-Exp for some anchor counts.
    • IMV-OTCP's anchor count and bandwidth multiplier are tuned per dataset (Appendix E.4), while MVCS uses a single configuration. The very large MVCS standard error on scm1d (4.266, s.e. 0.616) suggests that this baseline is not well tuned.
    • MVCS is better on rf1.
    • Taken together, the headline claim that learning the ground metric substantially reduces volume on real data is not yet established.
  2. The theory is mostly conditional on high-level assumptions that do not cover the implemented estimator.
    • Theorem 3.2 adds a standard covering/Hoeffding bound over the d(d+1)/2-dimensional metric class to the term E_N^obj. Assumption 3.1(ii) simply posits E_N^obj = o_P(1), i.e., uniform-in-M consistency of a GradNet fitted with a KDE and a stabilized determinant. This is the hard part, and it is not verified.
    • A related concern is that the Monge–Ampère residual loss (Eq. 10, Eq. 16) is enforced only pointwise at reference draws, against a KDE with full support. Without the second boundary condition (the pushforward constraint), a convex solution of this residual equation need not be the Brenier map. Identifiability of g_M by the inner fit is therefore unclear.
    • The analysis also treats M_hat as an exact global minimizer of a non-convex bilevel problem. In practice it is computed with 4–20 damped implicit-gradient steps.
    • For IMV-OTCP, Theorem 4.2 assumes R_N(x,Y) converges to R*(x,Y) (Assumption 4.1). The implemented method uses a fixed number of nearest-anchor fits with a non-vanishing bandwidth (Remark A.2, Appendix E.3–E.4), so this assumption cannot hold as stated. No conditions on the number of anchors or the bandwidth are given.
    • The oracle conditional coverage in Theorem 4.2 holds for any fixed metric, because T_{M,x}#P_x = Q for every M. The theorem therefore says nothing specific about the learned metric.
    • The abstract's claim of "optimality of volume metrics" overstates what is shown, which is consistency toward the minimizer within the Mahalanobis class under a separation assumption (Assumption 3.1(iii)).
  3. Key computational and evaluation details are missing.
    • The paper does not explain how the Lebesgue volume of the resulting regions is computed. These regions are non-convex: unions of polyhedral cells for MV-OTCP, and inverse images of a ball under a learned map for IMV-OTCP. This matters most for d = 8 and d = 16 (rf1, scm1d), where the precision of a Monte Carlo or grid estimate directly affects Table 1.
    • No runtime or memory comparison is reported. The cost of per-anchor bilevel fits with Hessian–vector/conjugate-gradient implicit differentiation, and of the per-trial-y assignment of full-conformal MV-OTCP, is unknown.
    • MV-OTCP itself, the paper's main method, is evaluated only on bivariate simulations.
    • The theory assumes i.i.d. reference draws (Theorem 3.3, Step 4), whereas the experiments use deterministic radial-shell grids (Appendix E.2–E.4). This mismatch is minor but unacknowledged.
Questions for authors
  1. Can you add FOTCP or split OTCP (Thurin et al. 2025; Klein et al. 2025), conditional neural OT (Kondratyev et al. 2026), and IMV-OTCP with M = I to Table 1, with the same anchors, bandwidths and GradNet budget? It would also help to report how far the learned M moves from its initialization, for example via its condition number or ||M_hat - I||_F. A clear gain over the M = I variant on real data would substantially strengthen the main empirical claim.
  2. How is the volume of the non-convex prediction regions computed, especially in d = 8 and d = 16? What is the estimation error of the volume estimator, and is the same estimator used for all methods?
  3. Does the pointwise Monge–Ampère residual loss (Eq. 10) identify the Brenier map g_M without an explicit pushforward or boundary constraint? Can you give primitive conditions, or empirical diagnostics such as the pushforward fit of the learned map, under which Assumption 3.1(ii) plausibly holds uniformly in M? The validity of the consistency results depends on this.
  4. Under what joint rates of the number of anchors L and the localization bandwidth h would Assumption 4.1 hold for the anchor-amortized IMV-OTCP actually used in the experiments? Without such conditions, Theorem 4.2 does not apply to the reported method.
  5. What are the wall-clock costs of MV-OTCP, IMV-OTCP and TA-OTCP relative to FOTCP, MVCS and RLCP, and how do they scale with d and n? This determines the practical relevance of the method.
6 comments

Thank you for your detailed review. We appreciate your positive assessment of the motivation and geometric characterization of our method, the finite-sample coverage guarantees, and the synthetic evidence supporting metric learning. We also thank you for highlighting important concerns regarding the incremental contribution of metric learning, the theoretical assumptions, and computational efficiency. In particular, we agree that the benefits of learning MxM_x should be distinguished from those of the overall localized OT pipeline. Below, we address these concerns in detail and provide additional experimental results and theoretical clarifications.

W1

The real-data evidence does not isolate the contribution of metric learning, and important baselines are missing.

W1(1)

Table 1 compares IMV-OTCP only with SCP, MVCS and RLCP. It omits the OT baselines the method builds on (FOTCP, the split OTCP of Thurin et al. 2025 / Klein et al. 2025, conditional neural OT of Kondratyev et al. 2026) and the IMV-OTCP(M=IM=I) ablation. It also omits other flexible multivariate CP methods the paper itself cites (e.g., Feldman et al. 2023, Fang et al. 2025, Izbicki et al. 2022, the scores benchmarked in Dheur et al. 2025).

Response. We expand Table 1 to include FOTCP (Ndiaye, 2025), conditional neural OT (CNOT; Kondratyev et al., 2026), and Individualized Whiten Metric OTCP (IWM-OTCP). IWM-OTCP corresponds to IMV-OTCP with a fixed whitening metric Mx∝Σ^x−1/2M_x\propto\widehat\Sigma_x^{-1/2}, where Σ^x\widehat\Sigma_x is estimated from the kNN residual covariance at the nearest anchor. We also evaluate two implementations of IMV-OTCP, using GradNet and the KDE-free Energy Score approach of Kan et al. (2022) for OT estimation, denoted GradNet, learned and Energy Score, learned, respectively. To isolate the contribution of metric learning, we include their corresponding fixed-identity (Mx=IM_x=I) variants under the same OT estimation procedures.

The table reports Root volume / MSCE\sqrt{\mathrm{MSCE}} at 90% target coverage, together with the relative Frobenius displacement ΔM=∥M^−M(0)∥F/∥M(0)∥F\Delta_M=\|\widehat M-M^{(0)}\|_F/\|M^{(0)}\|_F, where M(0)∝Σ^−1/2M^{(0)}\propto\widehat\Sigma^{-1/2} and Σ^\widehat\Sigma denotes the global sample covariance matrix. We use 100 seed-paired splits per dataset. The fixed-identity and IWM-OTCP controls are matched to the GradNet, learned variant in anchors, localization bandwidths, score-map losses, and map-fitting budgets. The smallest volumes are in bold, and the second-smallest are underlined.

Dataset (dd)GradNet, learnedGradNet, Mx=IM_x=IEnergy Score, learnedEnergy Score, Mx=IM_x=IIWM-OTCPMVCSCNOTFOTCPΔM\Delta_M (Energy Score)
Energy (2)1.434/0.0921.467/0.0901.358‾\underline{1.358}/0.0881.393/0.0891.647/0.0961.337/0.1011.463/0.0941.971/0.0987.9%
Bias (2)2.120/0.0392.159‾\underline{2.159}/0.0422.511/0.0662.521/0.0662.202/0.0402.177/0.0422.270/0.0432.217/0.0409.5%
Taxi (2)3.319/0.0293.324‾\underline{3.324}/0.0293.579/0.0283.578/0.0283.708/0.0343.492/0.0293.359/0.0273.395/0.03614.8%
rf1 (8)1.224/0.0541.344/0.0581.000/0.0511.011‾\underline{1.011}/0.0511.225/0.0531.286/0.0611.513/0.0533.092/0.05113.5%
scm1d (16)1.879/0.0631.903‾\underline{1.903}/0.0642.052/0.0502.112/0.0441.985/0.0552.076/0.0712.108/0.0525.611/0.05913.9%

Our IMV-OTCP framework achieves the smallest root volume on four of the five datasets and remains competitive on Energy. In particular, learned GradNet consistently outperforms both its fixed-identity counterpart and covariance-whitened IWM-OTCP across all five datasets. Under matched fitting configurations, these comparisons demonstrate that optimizing the ground metric can improve prediction-set efficiency beyond localized OT fitting alone.

To address the reviewer's suggestion regarding flexible multivariate CP methods, we additionally include HPD-split (Izbicki et al., 2022), which estimates conditional density rankings to construct high-density prediction regions. Unlike our approach, HPD-split directly targets the conditional highest-density region (HDR), which is the population minimum-volume set at a given conditional coverage level. Thus, the two methods optimize over different prediction-set families. Notably, our IMV-OTCP framework achieves smaller root volumes than HPD-split on three of the five datasets: Energy (1.358 vs. 1.388), rf1 (1.000 vs. 1.336), and scm1d (1.879 vs. 1.921). HPD-split performs better on Bias (2.105 vs. 2.120) and Taxi (3.236 vs. 3.319). These results demonstrate that our OT-based framework remains competitive with methods explicitly targeting the conditional HDR, while allowing flexible transport estimation without requiring conditional density estimation.

W1(2)

The real-data configuration uses only four outer metric updates with learning rate 0.02 (Appendix E.4), and the initialization of BB is not stated. MM may therefore stay close to its starting value, and the volume gains in Table 1 may come mainly from localized OT ranking rather than from the learned cost.

Response. We initialize the metric at M(0)∝Σ^−1/2M^{(0)}\propto\widehat\Sigma^{-1/2} and set B(0)=log⁡M(0)B^{(0)}=\log M^{(0)}. We will clarify this initialization and the optimization settings in Appendix E.4. As reported in W1(1), the learned Energy Score metrics exhibit relative Frobenius displacements of 7.9%–14.8% from initialization, confirming that the optimization updates the metrics.

More importantly, the matched comparisons in W1(1) directly address whether the improvements arise primarily from localized OT fitting. Both the learned-metric and fixed-metric variants use the same localized OT pipeline and fitting budgets, yet learned GradNet achieves smaller root volumes than both Mx=IM_x=I and IWM-OTCP across all five datasets. The Energy Score ablations also show improvements on four of the five datasets. These results provide empirical evidence that optimizing the ground metric contributes additional volume efficiency beyond localized OT ranking alone.

W1(3)

The only localized ablation (Figure 6, Appendix D.2) is on bivariate simulations. There the improvement over M=IM=I is a few percent on X-clover and can be negative on X-Exp for some anchor counts.

Response. The negative gains in Figure 6 observed in some X-Exp configurations occur primarily when relatively few anchors are used, where local metric and transport estimation can be less stable. As the anchor count increases, the performance becomes more stable, supporting the importance of sufficient localization resolution for effective metric learning. The real-data comparisons in W1(1) extend the metric ablation to response dimensions 2, 8, and 16.

W1(4)

IMV-OTCP's anchor count and bandwidth multiplier are tuned per dataset (Appendix E.4), while MVCS uses a single configuration. The very large MVCS standard error on scm1d (4.266, s.e. 0.616) suggests that this baseline is not well tuned.

Response. We have retuned MVCS on scm1d, reducing its root volume from 4.266 to 2.076 (s.e. 0.087). The updated result is reported in W1(1).

W1(5)

MVCS is better on rf1.

Response. In the updated rf1 comparison in W1(1), the root volume is 1.286 for MVCS and 1.000 for learned Energy Score IMV-OTCP. The learned GradNet variant also has a smaller root volume than MVCS.

W1(6)

Taken together, the headline claim that learning the ground metric substantially reduces volume on real data is not yet established.

Response. Please refer to our responses to W1(1)–W1(5), where we provide additional real-data comparisons, metric-learning ablations, and experimental evidence supporting the contribution of learned ground metrics to prediction-set efficiency.

W2

The theory is mostly conditional on high-level assumptions that do not cover the implemented estimator.

W2(1)

Theorem 3.2 adds a standard covering/Hoeffding bound over the d(d+1)/2d(d+1)/2-dimensional metric class to the term ENobj\mathcal E_N^{\mathrm{obj}}. Assumption 3.1(ii) simply posits ENobj=oP(1)\mathcal{E}_N^{\mathrm{obj}}=o_P(1), i.e., uniform-in-MM consistency of a GradNet fitted with a KDE and a stabilized determinant. This is the hard part, and it is not verified.

Response. We define the volume objective using the exact reverse OT map GMG_M for each candidate ground metric MM. The same objective can be evaluated with different fitted maps G^M\widehat G_M, including those obtained with GradNet or Energy Score. Assumption 3.1(i) gives the regularity needed for uniform concentration of the exact-map average (Appendix F.3). Part (ii) controls the error from using the fitted Jacobian and its stabilizer on the same reference draws.

When the exact and fitted reverse-map Jacobians and 1/fQ1/f_Q are uniformly bounded on B1−αB_{1-\alpha}, the determinant is Lipschitz on the relevant bounded set. The error in Assumption 3.1(ii) then satisfies $\mathcal E_N^{\mathrm{obj}}\le C\delta_N+C\sup_{M\in\mathcal M_0}\frac1N\sum_{i:U_i\in\mathbb B_{1-\alpha}}\lVert\nabla\widehat G_M(U_i)-\nabla G_M(U_i)\rVert_F. $ Thus δN→0\delta_N\to0 and convergence of G^M\widehat G_M to GMG_M in C1(B1−α)C^1({B}_{1-\alpha}), uniformly over M∈M0M\in\mathcal M_0, suffice. Theorem 1.1 of González-Sanz and Sheng (2024) and Theorem 1.5 of Caja-López et al. (2026) establish differentiability of the exact Brenier potential in C2,βC^{2,\beta} along smooth paths of positive densities on fixed, uniformly convex domains. A density plug-in OT estimator can therefore achieve the stated convergence by integrating the linearized bounds along a path between the estimated and true densities. This requires the bounds to be uniform over MM and both the density error in the required Hölder norm and the numerical solver error in C1C^1 to vanish uniformly over MM.The conditions above give a sufficient route for a density plug-in estimator.

W2(2)

A related concern is that the Monge–Ampère residual loss (Eq. 10, Eq. 16) is enforced only pointwise at reference draws, against a KDE with full support. Without the second boundary condition (the pushforward constraint), a convex solution of this residual equation need not be the Brenier map. Identifiability of gMg_M by the inner fit is therefore unclear.

Response. We clarify that, at the population level, the Monge–Ampère equation can identify the Brenier map under suitable regularity conditions, without separately imposing the second boundary condition. For a fixed MM, let Pˉ\bar P and Qˉ\bar Q denote the transformed score and reference distributions, with densities pˉ\bar p and qˉ\bar q. Suppose gg is convex, ∇g\nabla g is continuously differentiable and injective on the reference domain ΩQˉ\Omega_{\bar Q}, and pˉ(∇g(u))det⁡∇2g(u)=qˉ(u)\bar p(\nabla g(u))\det\nabla^2g(u)=\bar q(u) holds almost everywhere on ΩQˉ\Omega_{\bar Q}. By the change-of-variables formula,Pˉ(∇g(ΩQˉ))=∫ΩQˉqˉ(u) du=1\bar P\big(\nabla g(\Omega_{\bar Q})\big)=\int_{\Omega_{\bar Q}}\bar q(u)\,du=1. Hence (\nabla g)_{\\#}\bar Q=\bar P. Since ∇g\nabla g is the gradient of a convex function, Brenier's uniqueness theorem (Brenier, 1991) identifies it as the quadratic optimal transport map. Thus, the pushforward condition follows from the population equation under these assumptions.

We acknowledge that minimizing the empirical Monge–Ampère residual at finitely many reference points does not by itself establish these conditions. Assumption 3.1(ii) provides a high-level condition on the consistency of the resulting plug-in objective. We will clarify this distinction and the corresponding scope of our theoretical results.

W2(3)

The analysis also treats M^\widehat M as an exact global minimizer of a non-convex bilevel problem. In practice it is computed with 4–20 damped implicit-gradient steps.

Response. The returned metric can have a positive empirical objective gap εN\varepsilon_N after finitely many updates. This is the optimization error in a nonconvex empirical minimization problem. Theorems 3.2 and B.1 study the statistical error of empirical minimizers. If ΔN\Delta_N is the uniform deviation of the empirical objective from its population counterpart, the returned metric has population excess objective at most 2ΔN+εN2\Delta_N+\varepsilon_N. The same consistency argument under separation applies when εN=oP(1)\varepsilon_N=o_P(1).

Comments on the paper

1 comment

Changes to the paper

We have uploaded a revised PDF of the manuscript, with modifications highlighted in blue. The main changes are summarized below:

  • We qualify the abstract's volume-optimality claim as applying within the specified metric class. Appendix B.2 shows how the population Monge–Ampère equation identifies the Brenier map under explicit regularity conditions. Appendix G.7.1 adds a radius error bound and sufficient conditions for Assumption 4.1 under a growing-anchor design.
  • We link sample-whitened WM-OTCP to the assignment identity in Eq. (29) and compare it with MV-OTCP and FOTCP across six aspect ratios (Appendix E.1, Figure 6) and in a four-dimensional homogeneous setting (Appendix E.2, Table 2).
  • We update the real-data results in Table 1, including the retuned MVCS result on scm1d. Table 3 (Appendix E.4) adds FOTCP, CNOT, and locally whitened IWM-OTCP and compares learned and identity metrics for the GradNet and KDE-free Energy Score implementations of IMV-OTCP.
  • Appendix F.4 describes the additional OT implementations and how we compute calibrated set volumes, including importance sampling for IMV-OTCP. It also specifies the initial metric for GradNet-based IMV-OTCP as the inverse square root of the local residual covariance, normalized to determinant one.
  • Table 4 (Appendix E.7) reports single-thread fitting and calibration times for nine methods, including both IMV-OTCP implementations, across response dimensions and estimation sample sizes.
  • We expand the related work in Appendix A and cite Brenier (1991) in Section 3.2 and Appendix G.1. We also correct “For convenience” in Appendix B.3 and the regression mean notation to m(X)m(X) in Appendix F.3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.