acceptodds
Under review as a conference paper at ICLR 2027

Optimal Transport Is Not Volume-Optimal for Conformal Prediction

10 views

Abstract

Transport-based methods construct multivariate prediction regions as pullbacks of quantile regions of a simple reference distribution: any invertible map that transports the reference measure to the conditional law of the response yields regions with exact conditional coverage. This validity, however, says nothing about the volume of the resulting regions. Recent work has favored optimal-transport (Brenier) maps, implicitly treating the induced regions as canonical. We show that this choice is provably suboptimal. For a standard Gaussian reference, we show that, among elliptical targets with finite second moments, exactly the Gaussian and spherical targets yield prediction regions whose volumes equal those of the corresponding highest-density regions (HDR) at every coverage level; for every other target in this class, the Brenier-induced region has strictly greater Lebesgue volume over a nonempty open interval of coverage levels. For the compactly supported references used in center-outward multivariate ranks — the uniform ball and the spherical uniform — the same volume gap holds for every non-spherical elliptical target, without exception. We then factor admissible transports into a Brenier map and a measure-preserving rearrangement, thereby separating coverage from volume. We parameterize this class using smooth measure-preserving flows generated by skew-symmetric matrix fields with at most \(2(d_y-1)\) nonzero entries and derive a stochastic volume objective. Experiments on synthetic and real data show that measure-preserving rearrangements yield smaller prediction regions.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
multidimensional conformal prediction, conformal prediction, optimal transport, generative modeling, prediction regions

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bB5ROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies transport-based multivariate prediction regions Tx(B(0,r1−α))T_x(\mathcal{B}(0,r_{1-\alpha})). Main negative result (Thm. 2/18): with a standard Gaussian reference, the Brenier (OT) map yields HDR-volume regions at every level only for Gaussian or spherical elliptical targets; for all other finite-second-moment elliptical targets the Brenier region is strictly larger on an open interval of levels. For uniform-ball and spherical-uniform references the gap holds for every non-spherical elliptical target (Thm. 17). Constructively, any transport is factored as Q∘SQ\circ S with SS measure-preserving; smooth μU\mu_U-preserving flows are characterized by skew-symmetric band matrix fields with at most 2(dy−1)2(d_y-1) entries (Prop. 5/Cor. 6), the HDR volume is shown attainable over measurable rearrangements (Prop. 8), and SS is trained with a log-domain surrogate; split conformal calibration on ∥S^−1(T^−1(y,x))∥2\|\hat S^{-1}(\hat T^{-1}(y,x))\|_2 gives marginal coverage. Experiments: Student-t sweep (Table 1), synthetic unit-Jacobian targets (Table 2), and five real multi-output datasets (Fig. 2, Table 4).

Critical strengths
  1. A clean, rigorous and somewhat surprising theoretical point: Brenier-induced quantile regions, widely treated as canonical, are provably volume-suboptimal except in exactly characterized cases (full dichotomy in Thm. 18, reference-independent rigidity lemma). Notably, Thm. 17 covers the uniform-ball reference used by center-outward-rank methods, so the result also gives a principled justification for the post-hoc volume corrections that recent OT-conformal methods add heuristically, e.g. the rank-space cutoff optimization in FACTOR (Gao & Han, 2026) and references therein. Assumptions and limits (second moment, "open interval" of levels) are stated honestly.
  2. A principled, post-hoc correction: coverage is fixed by QQ and volume controlled by SS; the sparse skew-symmetric parametrization gives exact measure preservation and a closed-form Jacobian det⁡DuSθ=exp⁡{(∥Sθ(u)∥2−∥u∥2)/2}\det D_uS_\theta=\exp\{(\|S_\theta(u)\|^2-\|u\|^2)/2\}, and it applies to any invertible transport (also demonstrated on RealNVP). Because SS only reshapes the latent set, it is in principle composable with monotone rank post-processing used elsewhere (e.g. subgroup synchronization of ranks).

C. Gao and L. Han. FACTOR: Fairness-Aligned Conformal Transport for Optimal Regions. STAI-X 2026.

Critical weaknesses
  1. Theory/method gap and modest real-data gains. The suboptimality theory covers only elliptical targets, where the fix is a trivial affine whitening; nothing quantifies the gap or achievable gain for the non-elliptical targets the flow is designed for. On real data the rearrangement improves normalized log-volume by only ~0.01–0.05 per dimension (NOT: 4.53→4.50 Bio, 5.82→5.80 SCM20D), is neutral for NF on QM9 and worse on ONP (1.19→1.26). The claim that rearranged transports "dominate all baselines when dy≥4d_y\ge4" is largely inherited from the base NOT model (plain NOT already beats every baseline on SCM20D); CP-Split wins on both 2-D datasets.
  2. The surrogate and the parametrized class do not target the stated objective. The log-domain loss is only a lower bound on log-volume with a non-vanishing Jensen gap (Remark 10); flow images of a ball are diffeomorphic to a ball, so disconnected HDRs are unattainable (Remark 11); the oracle inequality (App. H) assumes a known TT and excludes approximation error. No guarantee links the trained SθS_\theta to the volume it is claimed to minimize.
  3. Missing comparison with simpler latent-set families and related volume-reduction work. The paper fixes the latent set to a ball and only reshapes it via SS, but it does not discuss or compare against cheaper alternatives that keep the Brenier map and change the latent set instead: (a) affine/whitened balls (sufficient for every target covered by Sec. 3), and (b) radial bands/annuli in rank space, which minimizes the same 1/f1/f-weighted volume estimator (Prop. 7's identity) over [rα,1,rα,2][r_{\alpha,1},r_{\alpha,2}]. The theory actually implies that radial-only corrections cannot reach the HDR for anisotropic targets (the obstruction is angular, Lemma 16), which would be a nice additional result to state and verify empirically.
Questions for authors
  1. What is the training/inference overhead of the ODE-based SS and S−1S^{-1} relative to the base NOT/NF model?
  2. Sec. 4.2 allows α\alpha as an input to SS (non-nested regions), but App. I.1 writes Sθ(⋅,x)S_\theta(\cdot,x) without α\alpha. Which was used, and how often does region crossing occur?
  3. How much of the gain over baselines is due to the rearrangement vs. the base model? Please add (i) an affine/whitening correction of QQ and (ii) a radial-band / cutoff-optimized baseline (keeping S=idS=\mathrm{id}) so that the value of reshaping, rather than rescaling, the latent set is isolated. Also explain why NF-R degrades on ONP.
  4. For Student-t in d=2d=2, can you provide a high-precision value of the Brenier–HDR gap to validate Table 1 beyond the k=1k=1 control?
  5. Is there a route to a measure-preserving correction for the uniform-ball reference (e.g. velocity fields tangent at the boundary), so the construction applies to center-outward ranks directly?
7 comments
AuthorsOpenReview ↗

Response to bB5R on the effect and cost of rearrangement

Thank you for the detailed review. We agree that the paper should distinguish the theoretical result about exact Brenier maps from the empirical gains attributable to rearranging a fitted map.

What changes when the base map is fixed? Comparing each backbone with its own rearrangement, the reported median normalized log-volume decreases for NOT on all five datasets and for RealNVP on three. RealNVP ties on QM9 at the reported precision (−0.05 for both) and worsens on ONP (1.19 to 1.26, with overlapping interquartile ranges). CP-Split has the smallest median volume on both 2D datasets, and plain NOT already beats the other families on SCM20D. There, the additional reduction from 5.82 to 5.80 corresponds to roughly 27% on the volume scale, calculated as 1−exp⁡(16(5.80−5.82))1-\exp(16(5.80-5.82)) from the rounded medians. For the non-elliptical 2D synthetic examples, where the HDR volumes are known, the median HDR-to-prediction volume ratios go from 0.43 to 0.96 (banana), 0.89 to 0.98 (sinusoidal), and 0.94 to 0.98 (star-shaped).

Surrogate, topology, and ONP. At a fixed level and under the assumptions of Lemma 9, the volume objective and its log surrogate identify the same optimal region over unrestricted measurable rearrangements. The Jensen gap can persist, so a decrease in the surrogate away from the optimum can leave volume unchanged or increase it. A smooth flow image of a ball cannot exactly represent a disconnected HDR, and Appendix H treats a known base map. An exact measure-preserving rearrangement leaves the fitted pushforward law unchanged. Errors in the base transport persist. After fitting and calibration, true conditional coverage can vary and the volume can increase, as it does on ONP. The current results do not distinguish the effects of base-map error, the surrogate, and the calibrated radius.

Simpler comparisons. With the true center and scatter of an elliptical target, unwhitening the radial map gives the HDR ellipsoids at every level. The composed map need not be Brenier for the original quadratic cost. The Elliptic comparator uses random forest residuals. A same-backbone whitening comparison would isolate the effect of whitening. In the discretized 2D Student-tt check, we compared all contiguous radial bands with the same reference mass (16,384 atoms). For every tested kk, the minimum-area band had zero inner radius, giving the original ball. Both same-backbone comparisons on real data remain open.

Overhead and nesting. With the base map frozen, the added training cost is fitting the flow, including forward ODE solves in the objective. The added scoring cost is applying Sθ−1S_\theta^{-1} to the latent point obtained from T−1T^{-1}. The closed-form determinant simplifies volume estimation. Comparable timings for the reported setup still need to be verified. The fitted Sθ(u,x)S_\theta(u,x) takes no coverage level as input. Levels are sampled during training to choose the latent balls. At prediction time all levels threshold the same score, so the regions are nested. We corrected the conflicting description in Section 4.2 and Appendix H. Appendix I already gave the level-independent model.

AuthorsOpenReview ↗

Follow-up to bB5R on the Student-t gap and uniform-ball reference

Student-tt reference check. The exact gap is zero for the spherical Student-tt control (k=1k=1) and for the anisotropic Gaussian controls. Both give zero at the precision of Table 1. For d=2d=2, ν=3\nu=3, and scatter diag(k,1/k)\mathrm{diag}(\sqrt{k},1/\sqrt{k}), the HDR area at mass cc is 3π[(1−c)−2/3−1]3\pi[(1-c)^{-2/3}-1]. We independently quantized the Gaussian reference with 65,536 equally weighted scrambled Sobol atoms and integrated the Student-tt density in the resulting mass-balanced Laguerre cells. At matched mass 0.90002, the NVF was approximately −0.0008, −0.0026, −0.0052, −0.0082, and −0.0111 for k=2,4,8,16,32k=2,4,8,16,32. The spherical control gave −0.00013 (exact value zero). At 16,384 atoms, the two scrambles gave −0.0116 and −0.0122 for k=32k=32. The check corroborates the scale and trend of the gap. Finite quantization limits the precision.

Uniform-ball reference. A sufficient construction is a smooth skew-symmetric field AA supported strictly inside the ball. Its row-wise divergence vi=∑j∂jAijv_i=\sum_j\partial_j A_{ij} is divergence-free by skew symmetry and vanishes near the boundary. Its flow remains inside the ball and preserves uniform volume. The spherical-uniform reference with a different radial density requires a weighted construction. Our experiments use the Gaussian reference.

AuthorsOpenReview ↗

Response to bB5R (1/3): two statements added in the revision

Thank you for the detailed review. This follow-up addresses requests left open in our earlier reply. Two statements prompted by your review are proved in the revision, in a new Appendix J; the numbers of all earlier statements are unchanged.

W3(b). Radial corrections cannot reach the HDR (Corollary 35). Work in the setting of Lemma 16 (dy≥2d_y\ge 2, elliptical target with a continuous, strictly decreasing generator): Gaussian reference and a non-Gaussian, non-spherical target with finite second moment, or the uniform-ball or spherical-uniform reference and a non-spherical target with finite first moment. Then f∘Qf\circ Q is not almost everywhere a radial function. Consequently, no family of radial latent sets (balls, annuli, unions of shells) has the HDRs as images, even for a dense set of levels, and there is a nonempty open interval of levels on which every radial latent set of the given mass has an image strictly larger than the HDR. The proof needs no regularity of QQ. If the preimages of the HDR ellipsoids were radial for a dense set of levels, then the Mahalanobis radius would satisfy ρΣ(Q(u))=β(∥u∥)\rho_\Sigma(Q(u))=\beta(\|u\|) almost everywhere for a Borel function β\beta. Monotonicity of a convex gradient and convexity of the ellipsoids force β\beta to be nondecreasing up to a null set of radii, hence equal almost everywhere to the radial map aa of Lemma 16. Every Brenier region would then be an HDR, which that lemma excludes.

Among radial sets of a given reference mass, the smallest image is obtained by thresholding M(r)=E[1/f(Q(U))∣∥U∥=r]M(r)=\mathbb{E}[1/f(Q(U))\mid \|U\|=r], with partial selection at a tied threshold if needed. The ball is the best radial set at every level if and only if MM coincides almost everywhere with a nondecreasing function. We do not prove that MM is monotone for elliptical targets. In our two-dimensional Student-tt check (16,384 atoms, level 0.9), the best contiguous band was the ball for every tested kk, so a band correction would have gained nothing there. Radial sets do suffice exactly when f∘Qf\circ Q is radial, for example for a spherical target with a ring-shaped density. In that example, an annular HDR can be attained by a single band at levels where the corresponding radial preimage is an annulus.

W1. Beyond elliptical targets (Proposition 28). For any target with density f>0f>0 and any almost-everywhere invertible transport TT from a radial reference, write gT=f∘Tg_T=f\circ T and let gT♯g_T^{\sharp} be its radially nonincreasing rearrangement under μU\mu_U. Then, with Vmin⁡(r)V_{\min}(r) the minimum volume at the same mass,

Vol(T[B(0,r)])−Vmin⁡(r)=∫B(0,r)(1gT(u)−1gT♯(∥u∥)) dμU(u)≥0.\mathrm{Vol}\bigl(T[B(0,r)]\bigr)-V_{\min}(r)=\int_{B(0,r)}\Bigl(\frac{1}{g_T(u)}-\frac{1}{g_T^{\sharp}(\Vert u\Vert)}\Bigr)\thinspace d\mu_U(u)\quad\ge\quad 0.

The gap vanishes at every level if and only if gTg_T is, almost everywhere, a nonincreasing function of ∥u∥\|u\|; for a C1C^1 diffeomorphism and a reference density γ(∥u∥)\gamma(\|u\|), if and only if ∣det⁡DT(u)∣=J(∥u∥)|\det DT(u)|=J(\|u\|) with γ/J\gamma/J nonincreasing (Theorem 3 of Kondratyev et al., ICLR 2026, gives sufficiency for Brenier maps under strict decrease). This is a criterion and a computable quantity, not a closed-form gap for a family of targets, and it is elementary. The right-hand side can be estimated from the fitted base density and reference draws. Under the fitted law, it bounds what any rearrangement can gain. For the three non-elliptical synthetic targets of Table 2, where the HDR volume is known, the HDR-to-region volume ratio goes from 0.43 to 0.96 (banana), 0.89 to 0.98 (sinusoidal) and 0.94 to 0.98 (star-shaped) in dimension 2.

Reviewer_kEzVOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper focuses on conformal prediction: Given a response distribution mu_{Y|X=x} and threshold alpha, find a set C_{1-alpha}(x) such that the probability that Y is contained in C_{1-alpha}(x) is 1-alpha. The set C_{1-alpha}(x) can be parametrized as the image of a ball under a transport map T_x between a reference distribution mu_U (say Gaussian) and the response distribution mu_{Y|X=x}, where the ball has probability 1-alpha under the reference distribution. The main question addressed in the paper is which map T_x has property that the volume of C_{1-alpha}(x) is the smallest. The first main result of this paper is that the Brenier map of optimal transport does not minimize the volume. The paper then formulates the above problem by using the standard polar decomposition of a transport map T_x as composition between the Brenier map and a measure-preserving map S_x. Using this result the paper writes down the optimization problem of minimizing the volume of C_{1-alpha}(x) over the maps S_x, and provides surrogate objectives that numerically are more stable. To ensure that the maps S_x are measure-preserving they are parametrized as flow maps using vector fields which guarantee the desired property. Lastly, the paper provides a few numerical experiments to show that their method yields conformal prediction sets with lower volume while maintaining adequate coverage.

Critical strengths

None

Critical weaknesses

The first result of this paper is not at all surprising. The Brenier map is optimal for quadratic cost and there is no reason to think that it is optimal for minimizing the volume of the conformal prediction set.

The paper lacks a theoretical characterization of the optimal S_x, that is, what theoretical properties does it have? (For example, the Brenier map is always the gradient of a convex function, a fact which is highly important.) On the other hand, the algorithm proposed in the paper is not novel, and the numerical results are not very convincing, showing small improvement while adding computational cost (of the map S_x.

Overall, the paper is will benefit from being re-written for clarity.

Questions for authors

The paper never discusses why we should we care about the conformal prediction set to have the smallest volume, since, by construction, it always has probability 1-alpha. The justification should come from downward tasks that practitioners would care about. Can the authors clarify this?

3 comments
AuthorsOpenReview ↗

Response to kEzV on the contribution, optimum, and downstream use

We disagree with the review's assessment on three substantive points.

Proposition 8 already characterizes every volume-optimal rearrangement. The review says the paper lacks a theoretical characterization of the optimal SxS_x. Proposition 8 in the submitted main text proves attainment and gives a necessary and sufficient condition. At a fixed coverage level, under the proposition's assumptions, an almost-everywhere invertible measure-preserving SxS_x in the polar factorization Tx=Qx∘SxT_x=Q_x\circ S_x minimizes volume exactly when it sends the reference ball onto Qx−1(Hx)Q_x^{-1}(H_x) up to reference-null sets, where HxH_x is the highest-density region at that level. Lemma 12 shows that the resulting region is unique up to null sets. Multiple rearrangements can attain it. Brenier's convex-gradient representation is an example of a different kind of characterization. Our geometric criterion identifies the full set of optimizers in the measurable class. Exact attainment within the smooth flow class is a separate question, and the paper discusses its topological limitations.

Fixing coverage is what makes the volume comparison meaningful. Coverage determines how often the set should contain the response. Volume measures how much of the response space the set leaves plausible at that reliability. Two regions can each contain 90% of the probability mass and differ substantially in volume. The submitted introduction states this distinction and Figure 1 illustrates it with two transports to the same target. The request for a downstream example is reasonable. The revised text explains how a tighter uncertainty set can reduce conservatism in robust optimization when it excludes scenarios that drive the worst-case cost. For estimated transports, split conformal calibration guarantees finite-sample marginal coverage. We assess conditional coverage empirically.

The theoretical result is a dichotomy, and the empirical gains depend on the dataset. We agree that quadratic transport optimality alone gives no reason to expect minimum-volume regions. The new result determines exactly when the Brenier map is volume-optimal throughout the family of levels for the elliptical targets under consideration. With a standard Gaussian reference, Gaussian and spherical targets attain the HDR volume at every level. The remaining elliptical targets have a strict gap on an open interval. For both compact radial references studied, the gap occurs for every non-spherical elliptical target covered by the theorem.

Polar factorization and measure-preserving flows are classical. The methodological contribution is to optimize prediction-region volume within the orbit of a fixed base transport, using a sparse Gaussian-preserving flow and a tractable objective. Under Lemma 9's conditions, the log-domain objective and the volume have the same minimizers over the unrestricted class at a fixed level. Training and inverting the flow adds computation. In the reported experiments, median normalized log-volume falls in eight of ten comparisons against the same base model. RealNVP ties on QM9 at the reported precision and worsens on ONP. On SCM20D, the NOT change from 5.82 to 5.80 in 16 dimensions corresponds to roughly 27% less volume on the scale implied by the rounded medians. In the controlled 2D banana example, the median HDR-to-prediction volume ratio rises from 0.43 to 0.96. These results show where the extra computation buys a meaningful reduction and where the method still needs improvement.

AuthorsOpenReview ↗

Response to kEzV (1/2): structure of the optimal map

Thank you for the review. This follow-up addresses the structure of a single map across coverage levels and clarifies how we summarized the volume gains.

1. Structure of the optimal map. You asked for a structural property of the optimal SxS_x, comparable to "the Brenier map is the gradient of a convex function". Proposition 8 completely characterizes the optimizers at a fixed level through the image of one ball. Your question asks for a structural description of a single map across all levels. The revision adds a statement about the map, valid for all levels at once (Propositions 28, 29 and 31, in a new Appendix J; the numbers of all earlier statements are unchanged). It is elementary: it follows from the change of variables and the bathtub principle. Let μU\mu_U have a positive radial density γ(∥u∥)\gamma(\|u\|), let GG be the distribution function of ∥U∥\|U\|, and let TT be an almost-everywhere invertible transport from μU\mu_U to a target with density f>0f>0.

  • The inverse-score regions T[B(0,r)]T[B(0,r)] defined in Appendix J have minimum volume at every level if and only if f(T(u))f(T(u)) is, almost everywhere, a nonincreasing function of ∥u∥\|u\|. For a C1C^1 diffeomorphism this means ∣det⁡DT(u)∣=J(∥u∥)|\det DT(u)|=J(\|u\|) with γ/J\gamma/J nonincreasing. A smooth Brenier map has a symmetric positive semidefinite Jacobian. In the diffeomorphic setting above, volume optimality across levels is characterized by the radial determinant and the monotonicity of the density ratio. Theorem 3 of Kondratyev et al. (ICLR 2026) gives sufficiency for Brenier maps under strict decrease. Proposition 29 gives a necessary and sufficient condition for the stated smooth transport class.
  • For T=Q∘ST=Q\circ S, the rearrangement SS is optimal at every level if and only if (f∘Q)∘S(f\circ Q)\circ S is the radially nonincreasing rearrangement of f∘Qf\circ Q under μU\mu_U. If f(Y)f(Y) has no atoms, optimal rearrangements are unique up to μU\mu_U-preserving maps that preserve ∥u∥\|u\|.
  • The loss we train, with the level sampled uniformly, equals, up to an additive constant, E[−log⁡G(∥U∥)⋅(−log⁡f(Q(S(U))))]\mathbb{E}\bigl[-\log G(\|U\|)\cdot(-\log f(Q(S(U))))\bigr]. At fixed xx, if log⁡f(Y)∈Lp\log f(Y)\in L^p for some p>1p>1, then over almost-everywhere invertible measure-preserving maps its minimizers are exactly the rearrangements of the previous item, whenever one exists.

Limits: an invertible optimal SS need not exist (for example in dimension one), there is no canonical choice among optimal maps, and if the target density has no plateaus and the composed transport is a diffeomorphism, each HDR must agree up to null sets with the corresponding diffeomorphic image of a ball. In no case do we have a guarantee that the trained flow reaches the optimum.

2. "Not surprising." We agree that quadratic-cost optimality gives no reason to expect minimum volume. It is also known that Monge–Kantorovich quantile maps are not affine equivariant and that the center-outward map of a non-spherical elliptical law is not the Mahalanobis map (Chernozhukov et al., 2017; Hallin, La Vecchia and Liu, 2022). The paper proves a statement about regions and volumes. Among elliptical targets with a continuous, strictly decreasing generator and finite second moment, the Gaussian-reference Brenier regions are HDRs at every level exactly for Gaussian and spherical targets, and are strictly larger on an open interval of levels otherwise. This does not follow from the statement about maps: composing an HDR-generating map with a map that preserves spheres leaves every region unchanged, and excluding this case is the content of Lemma 16. Outside the elliptical class, the criterion is the first item above.

AuthorsOpenReview ↗

Response to kEzV (2/2): motivation, novelty, size and cost of the gains

3. Why volume. At fixed coverage, the size of the set is the standard efficiency criterion for prediction sets (Lei, Robins and Wasserman, 2013; Sadinle, Lei and Wasserman, 2019; Izbicki et al., 2022), and the HDR is the minimum-volume set. The submitted introduction and Figure 1 already distinguish coverage and volume. The revised manuscript states the efficiency criterion explicitly and adds a downstream illustration. The sentence on robust optimization is an illustration: a smaller set does not by itself guarantee a lower worst-case cost. The paper does not evaluate a downstream task.

4. Novelty of the algorithm. Polar factorization and flows that preserve a reference measure are classical, and the skew-symmetric form of such fields is standard. What the paper adds is a sparse band parametrization for the Gaussian reference, with at most 2(dy−1)2(d_y-1) entries and a closed-form Jacobian determinant, and its use to reduce the volume of the regions of a fixed base transport. When an all-level optimum exists and the fitted conditional density has no plateaus, its regions equal fitted HDRs at the same fitted mass. Those HDRs can also be obtained by thresholding the fitted density without training a rearrangement.

5. Size and cost of the gains. The roughly 27% figure for NOT on SCM20D is a conversion of rounded medians. Rounding alone allows reductions of roughly 15% to 38%. The interquartile ranges overlap, and the table provides separate medians rather than paired differences. Against the same base model, over ten (dataset, backbone) pairs, there are six decreases with separated interquartile ranges, two decreases within overlapping ranges (NOT on QM9 and SCM20D), one tie (RealNVP on QM9) and one increase (RealNVP on ONP). The metric is per dimension. For NOT on ONP, the change from 1.04 [1.03,1.04]1.04\,[1.03,1.04] to 1.00 [0.99,1.00]1.00\,[0.99,1.00] in dimension 8 is about 4% per axis and about 27% in volume (21% to 33% given rounding). By the same conversion, the increase for RealNVP on ONP is about 75% (62% to 90% given rounding, with overlapping interquartile ranges). On the two-dimensional datasets the gains for NOT are about 6% and 10% in volume. On the two-dimensional banana target, the HDR-to-region volume ratio goes from 0.43 to 0.96.

Fitting and inverting the flow adds computation. We do not yet have verified timings under the protocol of Appendix I. The preliminary measurements use a different training protocol.

6. Clarity. We did not rewrite the paper. The revision makes three clarifications. The list of contributions separates the fixed-level statement about sets (Proposition 8) from the criterion valid at all levels (Proposition 28). Section 4 states once that SS does not depend on the level and what the level-averaged loss targets (Proposition 31). Section 5 states that Table 1 reports the gap of the fitted map.

Reviewer_XS3NOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper demonstrates that Brenier maps can produce volume-suboptimal multivariate prediction regions and proposes a measure-preserving rearrangement framework to reduce prediction-region volume while maintaining conformal coverage guarantees.

Critical strengths
  1. Conformal prediction using transport maps for multivariate outcomes is an active research area. The paper addresses an important and interesting problem of constructing efficient prediction regions.

  2. The paper is theoretically sound, and the proposed method is well supported by rigorous theoretical analysis.

Critical weaknesses
  1. The experimental evaluation could be extended to demonstrate the broader applicability and advantages of the proposed method.

  2. Some experimental results require further discussion and analysis, particularly cases where the proposed rearrangement does not improve prediction-region volume or conditional coverage.

Questions for authors
  1. The effectiveness of the proposed rearrangement may depend on the pretrained transport map. In particular, the training objective

L(S)=E[log⁡∣det⁡DT(S(U))∣+12∥S(U)∥22]\mathcal{L}(S)=\mathbb{E}\left[\log|\det DT(S(U))|+\frac{1}{2}\|S(U)\|_2^2\right]

is determined by the estimated transport TT. Since the rearrangement preserves the reference distribution, it cannot correct errors in the target distribution learned by TT. Consequently, when the original transport poorly approximates the true target distribution, minimizing this objective may not lead to smaller prediction regions under the true distribution, even after conformal calibration. Could the authors discuss this limitation? Or if there is a controlled experiments to investigate the robustness of the proposed rearrangement to varying levels of accuracy in the pretrained transport, it will be helpful for readers.

  1. Although the proposed rearrangement generally improves prediction-region volume, Figure 2 shows that it does not consistently improve volume or maintain the same ECR across all datasets. The authors briefly acknowledge the cases where volume does not improve, but a more detailed discussion of these results would strengthen the paper. In particular, could the authors explain why rearrangement sometimes increases prediction-region volume or worsens ECR, despite the theoretical motivation? It would also be useful to investigate whether these outcomes are related to the quality of the pretrained transport map or the discrepancy between the surrogate objective and actual volume.

  2. Since the proposed rearrangement framework is applicable to general invertible transport maps, it would be helpful to extend the experiments to additional transport architectures beyond Neural Optimal Transport (NOT) and RealNVP (NF, as used in CONTRA). I believe the authors used NOT and NF (i.e., CONTRA) since they are commonly used or already published results. Other architectures, such as continuous normalizing flows (CNFs), could also be considered. My understanding is that the theoretical properties of the proposed rearrangement do not depend on the specific architecture of the base transport, provided the required assumptions are satisfied. Therefore, evaluating additional transport architectures could further demonstrate the general applicability of the proposed method and provide insights into how its effectiveness depends on the choice and quality of the base transport.

11 comments

We sincerely thank the reviewer for the careful reading of our paper and for the constructive and encouraging feedback. We are glad that the reviewer finds the problem important and the theoretical analysis rigorous and sound. Below we answer each question and describe the revisions we are making.

AuthorsOpenReview ↗

Q1: Dependence on the quality of the pretrained transport.

We thank the reviewer for raising this point. We would like to emphasize that the main contribution of the paper is, to the best of our knowledge, the first result showing that Brenier-induced prediction regions, which recent work implicitly treats as canonical, are provably volume-suboptimal. The gap also holds for the uniform-ball and spherical-uniform references used for center-outward ranks (Theorem 17). The factorization T=˜Q˜∘˜S˜T\~=\~Q\~\circ\~S separates two roles: the volume is controlled by the rearrangement SS, while the coverage, including conditional coverage as measured by ECR and WSC, is determined by QQ, whose estimation is outside the scope of our work. Hence, we do not aim to provide a better estimate of the transport map itself, or to show how to correct the inexactness in the badly estimated tranportation map QQ; rather, we show that even a perfectly estimated OT map is suboptimal for conformal prediction. With the rearrangement step, our goal is, given a reasonably well-estimated transport, to select the representative within its measure-preserving orbit that has the smallest prediction region.

The reviewer is right that SS preserves μU\mu_U and therefore cannot correct errors in the target law learned by T^\hat{T}. If the base transport is poorly estimated, the level sets of the estimated density are not aligned with those of the true density, and the rearrangement may fail to recover the true HDR sets. Moreover, since neither the transport nor the rearrangement is exact in practice, the conformal calibration step may substantially inflate the resulting regions. Hence, no volume guarantee can be given once estimation errors are taken into account. Marginal coverage, however, remains guaranteed by conformal calibration regardless of the quality of T^\hat{T} and S^\hat{S}. We will add this discussion to the Limitations paragraph.

AuthorsOpenReview ↗

Q2: Cases in Figure 2 where volume or ECR does not improve.

We thank the reviewer for this remark. We agree that these cases deserve a more detailed discussion, and we will expand it in the revision.

Volume. The rearrangement reduces the volume of its own base transport in nine of the ten (dataset, base model) pairs. The exception is RealNVP on onp, where the two results lie within each other's interquartile ranges. This behaviour is in line with our theory: the Brenier map already produces the HDR family exactly for Gaussian and spherical targets (Theorem 18). When a conditional response law is close to one of these cases, the base regions are already close to optimal and little can be gained. The controlled experiments support this intuition: the gain is largest for the strongly curved, non-convex banana target and smallest for targets whose level sets are closer to ellipsoidal (Table 3). We will add this explanation to Section 5. We will also revise Figure 2 to show directly the (non-normalized) volume of each base transport next to its rearrangement, which makes the effect of the rearrangement easier to read.

ECR. The rearrangement is not designed to improve conditional coverage, and we do not expect it to. Since SS preserves μU\mu_U, the conditional law described by the model is unchanged, and so is the conditional coverage \emph{under the model}. In practice, the ODE discretization of SS adds a small amount of numerical error. Accordingly, we observe small changes in ECR in both directions, within the variability across seeds. The relevant conclusion is that the rearrangement reduces volume without degrading conditional coverage, which is what Figure 2 is meant to show. We will state this explicitly.

Surrogate objective. The surrogate is a lower bound on the log-volume, and its Jensen gap does not vanish (Remark 10). However, it shares its population minimizer with the volume (Lemma 9). In our experiments it gave slightly smaller volumes than the original objective and trained more stably. We, therefore, do not believe that the surrogate is the source of the cases discussed above.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of the revision

We thank the three reviewers. This comment summarizes changes in the revised PDF.

New statements. Propositions 28, 29 and 31: a transport gives minimum-volume regions at every level if and only if the target density, read in latent coordinates, is a nonincreasing function of the latent norm; smooth form in terms of the Jacobian determinant (Theorem 3 of Kondratyev et al., ICLR 2026, gives sufficiency for Brenier maps under strict decrease); gap identity, characterization of all-level optimal rearrangements up to radius-preserving maps when there are no density plateaus, and minimizers of the level-averaged loss when the class contains an all-level optimum. Corollary 35, with Lemma 34: for the targets of Theorems 2 and 17, no family of radial latent sets recovers the HDR family. They are in a new last appendix (Appendix J), so that the numbers of the earlier statements are unchanged.

Corrections. Section 5 now states that six of the eight within-backbone decreases have separated interquartile ranges, and reports the changes in worst-slab coverage. Table 1 is described as the gap of the fitted map; reference values for d=2d=2 and a Euclidean-ball comparator are added. Volume rankings against baselines whose volume is estimated in a bounding box are removed for dy≥8d_y\ge 8. The claims on the parametrization of measure-preserving fields are narrowed to the sparse band representation.

Not done. A flow-matching backbone; same-backbone whitening and radial-set comparisons on real data; a controlled study of the effect of base-map error; the cause of the increase on ONP with RealNVP; an independent validation of Table 1 for d≥4d\ge 4; timings under the protocol of Appendix I.

Scope of the claims. The main contribution is the theory of Section 3. Proposition 28 is an elementary criterion that complements it. The rearrangement is a trained map whose ball images are meant to approximate the HDRs of the fitted transport. In the experiments the median volume decreases in eight of ten within-backbone comparisons and increases in one.

AuthorsOpenReview ↗

Real-data experiments: current results and their scope (2/2): coverage, scope, current PDF

(Continuation of part 1/2.)

4. Coverage. Calibration guarantees marginal coverage, not conditional coverage. Between a base model and its rearrangement, the values of the excess coverage risk (ECR, an estimate of the deviation from conditional coverage) posted on 15 November change by at most 0.003: the ECR increases in six pairs (for example NF on Bio, 0.054 → 0.057), decreases in four (for example NF on SCM20D, 0.076 → 0.073) and is unchanged in five. For NOT these changes are within one standard deviation; no standard deviation is posted for NF and FM. The comments of 15 November do not report worst-slab coverage. For NOT, the changes that we reported on 13 November still apply (.90 to .88 on SCM20D and .90 to .89 on ONP); for the re-tuned NF runs and for FM we have not reported it. We have therefore removed from our comments of 15 November the statements that the volume reduction comes "without a loss in conditional coverage", "at no cost in marginal or conditional coverage" or "at no cost in coverage", and we have limited "unchanged within one standard deviation" to NOT. This concerns conditional coverage only: marginal coverage is guaranteed by calibration, and the values posted on 15 November for NOT are within 0.002 of 0.90. What we wrote on 13 November stands: for fitted transports, the statement that the rearrangement reduces the volume without degrading conditional coverage is not supported.

5. Near-Gaussian targets. On 15 November we attributed the absence of a gain on QM9 to targets that are close to Gaussian, and wrote that the gains of NOT on SCM20D and QM9, which are within one standard deviation, are consistent with such targets. This is a conjecture that we have not tested, and our comments now say so. Theorem 18 concerns the exact Brenier map: its regions are HDRs at every level when the target is exactly Gaussian or spherical. It gives no bound for targets that are only close to Gaussian, for a fitted map such as NOT, or for RealNVP and flow-matching maps. On 13 November we withdrew this explanation for the two cases that did not improve at that time (RealNVP on QM9 and on ONP) and wrote that we had not identified the cause of the tie on QM9. The RealNVP runs on QM9 are unchanged, so the withdrawal and the statement on the tie both still hold. The manuscript does not use this explanation.

6. Scope. The statements of 15 November that the rearrangement "never increases" the volume and is "a backbone-agnostic, plug-in improvement" described the current runs; we have reworded them. They are not guarantees: the result for NF on ONP changed with the hyperparameters of the base model, and, as we wrote on 13 November, nothing guarantees that the trained flow reduces the volume. On the computational cost of the rearrangement, our reply of 13 November is unchanged: fitting and inverting the flow adds computation, and we do not yet have verified timings under the protocol of Appendix I.

7. Current PDF. Current runs: Table 2 (NOT), Figure 2 (NOT, NF, FM) and Figure 5 (all methods). Earlier runs: the NF and NF-R entries for SCM20D and ONP in Tables 5 to 8 of Appendix I; these tables have no FM rows. Appendix I states both points, and the captions of Tables 5 to 8 and of Figures 2 and 5 state that these NF entries come from earlier runs. The PDF does not describe the FM model or the re-tuned NF hyperparameters; its reproducibility statement says so. Of the tables announced on 15 November, the PDF contains Table 2 only: the normalized log-volume and the ECR of NF and FM are shown as box plots in Figure 2, and the table of percentages is not in the PDF. Section 5 no longer contains the counts described in our summary of 13 November (eight of ten, six with separated interquartile ranges); the changes in worst-slab coverage are reported in Appendix I.

Numbering: the PDF has a new Table 2 and a new Figure 2. In the reviews and in our comments of 11 and 13 November, "Table 2" (synthetic targets) is now Table 3, "Table 4" (normalized log-volume) is now Table 5, and "Figure 2" (all methods) is now Figure 5, which now includes FM and the re-tuned NF runs; the tables of ECR, worst-slab coverage and marginal coverage are now Tables 6 to 8.

The abstract of the PDF and of the forum page has been reworded: it now states the generator and moment conditions of Theorems 2 and 17 and says that, for exact transports, the rearrangement can change the volume of the regions but not their coverage. The theorems are unchanged.

Later revision of 15 November: two paragraphs are added to the main text, which now ends on page 10. One, in Section 4.2, states results of Appendix J; the other, in Section 5 ("Reading the results"), states the points above on volume, on the dependence on the base model, on conditional coverage and on cost. No numbered statement, equation, table or figure is added or renumbered.

AuthorsOpenReview ↗

Real-data experiments: current results and their scope (1/2): runs, volume, percentages

This comment (two parts) concerns the real-data experiments only; the theoretical results and the synthetic experiments are not affected. Our comments of 11 and 13 November and those of 15 November describe partly different runs, and some sentences of 15 November said more than the results show. We have edited our two comments of 15 November accordingly; their earlier versions remain in the revision history. NOT, NF (RealNVP) and FM (flow matching) denote the base models, "-R" the rearranged model, and d the response dimension. Table and figure numbers are those of the current PDF (see item 7 in part 2/2).

1. Current runs. Following the reviewers' remarks on the results on SCM20D and ONP (in particular the increase for NF on ONP), we re-tuned the hyperparameters of the NF base model. We also added FM: on 13 November we had withdrawn our promise of a flow-matching base model and listed it as not done; these runs have now been made. The NOT results, and the NF results on Bio, Blog and QM9, are unchanged. Normalized log-volume (log-volume divided by d), NF → NF-R:

  • earlier runs (medians): SCM20D 6.76 → 6.53; ONP 1.19 → 1.26;
  • current runs (values posted on 15 November): SCM20D ≈5.92 → ≈5.90; ONP ≈1.05 → ≈1.00.

The re-tuned base model has a smaller volume on both datasets, which is consistent with the earlier model being insufficiently trained, as we suggested on 15 November. We have not tested whether this explains the earlier increase on ONP. Our statements of 11 and 13 November about RealNVP on these two datasets describe the earlier runs: the increase on ONP (about 75% in volume), the decrease on SCM20D counted among the six decreases with separated interquartile ranges, the count of eight decreases in ten comparisons, the statement that the median ECR is unchanged at two decimals except for RealNVP on Bio, and the worst-slab coverage values.

2. Volume. At two decimals, the values of the normalized log-volume posted on 15 November decrease in 13 of the 15 (base model, dataset) pairs and are equal in two (NF and FM on QM9). The medians of Figure 2, rounded to two decimals, give the same count, which is the one in the abstract; on QM9 the medians of NF and NF-R, and of FM and FM-R, differ by less than 0.01. Our comments of 11 and 13 November give medians, whereas Table 2 and the NOT columns of 15 November give means with standard deviations (hence 5.82 → 5.80 before and 5.82 → 5.81 now for NOT on SCM20D). For NOT, Table 2 gives the total log-volume (not divided by d): the mean decreases on all five datasets, by 0.07 on Bio, 0.10 on Blog and 0.33 on ONP, against standard deviations of 0.02, 0.07 and 0.07 to 0.08, and by 0.03 on QM9 and 0.22 on SCM20D, against standard deviations of 0.04 to 0.05 and 0.66 to 0.76. As on 13 November, these are separate summaries of the base model and of its rearrangement; we still do not report paired differences. The NF and FM values, marked "≈", are within 0.015 of the medians of Figure 2 and are posted without a measure of dispersion; Figure 2 shows box plots.

3. Percentages. The percentages of 15 November are 1 − exp(−d·δ), where δ is the decrease of the normalized log-volume; for NOT, d·δ is the decrease of the mean total log-volume of Table 2. This changes the scale, not the precision.

  • SCM20D (d = 16): NOT "19%", from a decrease of the mean of 0.22 (20% with the rounded means of Table 2), smaller than one standard deviation; NF "27%", from δ ≈ 0.02 (15% to 38% for δ between 0.01 and 0.03); FM "80%", from δ ≈ 0.10 (76% to 83% for δ between 0.09 and 0.11).
  • ONP (d = 8): NOT "29%", from a decrease of the mean of 0.33 (28% with the rounded means of Table 2), about four standard deviations; NF "33%", from δ ≈ 0.05 (27% to 38% for δ between 0.04 and 0.06); FM "21%", from δ ≈ 0.03 (15% to 27% for δ between 0.02 and 0.04).

For NOT, the value of about 27% that we gave on 13 November for each of these two datasets came from rounded medians of the same runs. For NF on SCM20D we have removed the word "substantial"; the manuscript describes this reduction as limited.

(Continued in part 2/2: coverage, near-Gaussian targets, scope, current PDF.)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.