acceptodds
Under review as a conference paper at ICLR 2027

Manifold Constrained Conformal Prediction For Spatial Distributions

Abstract

We introduce a conformal prediction method for spatial distributions observed through finite samples of events, such as tropical cyclone genesis and earthquake locations. Our approach scores empirical measures using spherical sliced Wasserstein distance, then constrains the resulting prediction set to lie near the training data manifold. We derive a finite sample coverage lower bound for future empirical measures using a data adaptive manifold tolerance and then extend this to the underlying distribution by upper bounding the for sampling error. Because the resulting set is not analytically tractable, we introduce a modified flow based sampling procedure that represents candidate distributions by particle clouds and produces an ensemble over spatial distributions. Numerical experiments assess empirical measure coverage on synthetic data and compare energy distance and manifold distance on synthetic processes, tropical cyclone genesis, and earthquake occurrences.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal, Wasserstein, Spatial, Uncertainty Quantification, Events

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LmYuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper quantifies uncertainty for spatial distributions represented by event locations over an entire period. Events are encoded as empirical probability measures on the sphere, and a conformal prediction set is constructed using spherical sliced Wasserstein (SSW) distance. An empirical manifold constraint based on proximity to training empirical measures is introduced to restrict spatial configurations that deviate from historical patterns. The paper bounds the additional miscoverage induced by this constraint and derives a coverage bound for the underlying distribution after enlarging the set. A constrained velocity field generates candidate distributions. Simulations and cyclone and earthquake case studies illustrate trade-offs between manifold discrepancy (MD) and energy distance (ED).

Critical strengths

The prediction target is an entire period's spatial distribution rather than a single location or scalar, making the problem relevant to spatial statistics and geoscience.

The theory explicitly distinguishes empirical measures from underlying distributions and quantifies the coverage cost of the additional constraint.

The authors acknowledge limitations involving temporal splitting, the sampling-error radius, and the interpretation of the generated ensemble. The experiments also show that the constraint is not uniformly beneficial across datasets.

Critical weaknesses
  1. The relationship between the theoretical constraint and the implementation as described is unclear. Equation (8) measures proximity to a single complete training empirical measure, whereas Section 3.3 describes pointwise matching to pooled reference support. These distances are generally different, leaving the connection between the numerical outputs and the theoretically defined set unresolved. Moreover, the constraint is operationally a union of SSW balls around training measures; no smooth-manifold structure is assumed or established. The terminology should distinguish empirical proximity from geometric manifold estimation.

  2. The reported real-data coverages of 11/13 and 15/15 concern the ambient SSW ball and do not establish coverage of its intersection with the empirical manifold constraint. The stated lower bound, 1−α−1/(n0+1)1-\alpha-1/(n_0+1), applies when ε≥ε^\varepsilon\ge\widehat{\varepsilon}, where ε^\widehat{\varepsilon} is the training leave-one-out maximum defined in Equation (10). Smaller tolerances are not covered by this guarantee, and the bound can be substantially below the nominal level when the training sample is small.

  3. The practical benefits and scientific interpretation require stronger evidence. The constraint mainly improves MD, which is aligned with its optimization objective, whereas ED does not improve consistently. Proximity to historical patterns alone does not establish physical consistency. Although the authors acknowledge that set coverage does not imply calibration of arbitrary ensemble summaries, the practical calibration of ensemble-derived quantities remains unestablished. Temporal dependence and distribution shift further limit the applicability of the coverage guarantee.

Questions for authors
  1. Does the velocity field use the distance in Equation (8), or the SSW distance between the candidate cloud and the locally matched reference cloud described in Section 3.3? If the latter, how is it related to the theoretical constraint, and what fraction of generated outputs satisfies the constraint when evaluated using Equation (8)?

  2. Can the authors separately report coverage of the ambient SSW ball, the empirical manifold constraint, and their intersection, and explain how the additional coverage loss is accounted for within the target miscoverage budget?

  3. Are coverage and predictive scores stable under rolling calibration using only historical data and evaluation on consecutive future periods? How are interannual dependence and long-term distribution shift addressed?

  4. Can the authors provide geographic or scientific evaluation metrics independent of the empirical manifold constraint, together with numerical constraint residuals, failure rates, and step-size sensitivity?

  5. Beyond proximity to historical spatial patterns, what evidence supports the claimed geophysical coherence? How does the constraint behave for genuine but rare or shifted spatial configurations?

  6. Apart from the change in prediction target and spherical geometry, what technical developments go beyond existing OT-based conformal methods (e.g., Thurin et al., 2025; Klein et al., 2025) and the nearest-neighbor conformal argument of Feldman et al. (2025)? Can the authors distinguish these developments from the combination of established components?

3 comments
AuthorsOpenReview ↗

Response to Critical Weaknesses and Questions 1,2,3 from Reviewer LmYu

Critical Weakness 1)

Thank you for helping us identify this inconsistency in our notation. The correct definition should have been the one in Section 3.3. The definition in Eqn. ~8 was a holdover from a previous iteration of the paper. We have updated our definition and proposition 3.1 in response to this fix. Because question was raised by all reviewers, we have included a more detailed response in our global response at the top. We have updated the definition in Eqn ~8 and updated Proposition 3.1 to use this definition. Proposition 3.1 only changes in the 1/(n+1)1/(n + 1) term, which is now 1/(ntrain+1)1/(n_{\rm{train}} + 1) instead of 1/(ncal+1)1/(n_{\rm{cal}} + 1).

“Moreover, the constraint is operationally a union of SSW balls around training measures; no smooth-manifold structure is assumed or established. The terminology should distinguish empirical proximity from geometric manifold estimation”

Thank you for identifying that the terminology can imply a stronger geometric assumption than the method requires. We do not estimate a smooth differential manifold, and the coverage argument does not require differentiability, curvature, or a fixed intrinsic dimension. The statistical object is empirical proximity to a family/support of historically plausible distributions. We will make this explicit and use terms such as ''empirical support proximity'' or ''empirical plausibility tube'' where greater precision is needed.

Critical Weakness 2)

Yes, you are correct in that the theory holds when ϵ≥ϵ^\epsilon \geq \hat\epsilon and not when ≤\leq, however, we choose ϵ^=ϵ\hat\epsilon = \epsilon. This is a conservative finite sample lower bound rather than an expected empirical coverage.

Critical Weakness 3)

"Temporal dependence and distribution shift further limit the applicability of the coverage guarantee"

Yes, we shared the same concerns in the real data, however, our temporal diagnostics in the appendix (Table 8 and 9) show no strong evidence of substantial trend, serial dependence over time, or shifts between the calibration and test conformity scores. We interpret these analyses as empirical checks that we can not reject the assumption that the scores are exchangeable.

Question 1)

The constrained velocity field uses the manifold penalty described in 3.3, we have updated the notation that will replace Eq. 8. in the official comment.

"what fraction of generated outputs satisfies the constraint when evaluated using Equation (8)"

We replaced Eq. 8. since it was a hold over from a previous version. For the correct Eq. 8. 20 out of 21 of the tropical cyclone genesis CVF samples satisfied the manifold tube constraint, while all 26 earthquake samples satisfied the manifold constraint.

Question 2)

Yes, we will provide the ambient ball coverage, the manifold coverage, and the intersection of the two in the updated manuscript as in the official comment. The manifold coverages for the real data examples had 13/13 coverage for tropical cyclone genesis and 15/15 for earthquakes meaning their intersection coverage is 11/13 and 14/15 respectively. The earlier 15/15 ambient coverage for earthquakes was updated after rechecking the ambient coverage. For the simulations Figure 2 shows that if we set ϵ too low then we incur a loss of coverage. However, if we set it to the calibrated number or higher we do not lose any statistically discernible coverage.

Question 3)

This is a great idea. Although our statistical checks (Tables 8 and 9) did not show evidence of serial correlation or distribution shift, it stands to reason that tropical cyclones and earthquakes, which may depend on slowly evolving processes, may not be exchangeable year to year. We have updated our results to include an additional adaptive conformal method (Xu and Xie, 2021, Conformal Prediction Interval for Dynamic Time-Series, ICML) to construct a rolling prediction set. The results show that, at α=0.10\alpha=0.10, ambient SSW coverage on tropical cyclones is 12/13=0.923.12/13=0.923. When we include the manifold constraint, the coverage is unchanged. The same 12/13=0.923.12/13=0.923. tropical cyclones were covered. For earthquakes at the same α\alpha coverage was 14/15=0.93314/15=0.933 using the adaptive method. Again, all 15/1515/15 held out yearly distributions satisfy the calibrated localized manifold condition, so manifold restricted coverage is still 14/15=0.933.14/15=0.933.

AuthorsOpenReview ↗

Response to Questions 4, 5, 6 from Reviewer LmYu

Question 4)

This is a great point and something the other reviewers have asked about. For each of the real data analysis examples we have included additional geographic validation based on independent data. We have included the full table of comparisons in the global response, which shows that the proposed CVF attains the independently derived geographic boundaries more frequently than VF.

We also agree with you in which diagnostics related to the flow should be reported. We will report the fraction of generated clouds satisfying each and both constraints in the final manuscript. For the earthquake example, only one sample generated was found to be outside of manifold tube. For tropical cyclone genesis locations, all samples were contained within the manifold tube.

Question 5)

"Beyond proximity to historical spatial patterns, what evidence supports the claimed geophysical coherence?"

For these two applications, geophysical coherence has a directly observable spatial component. Tropical cyclone genesis is geographically constrained by known genesis climatology, while earthquake occurrences are structurally tied to fault systems. We believe the additional tests provided in our official comment directly evaluate these aspects of geophysical coherence.

Question 6)

We apologize that we did not make the distinction between our work and prior work sufficiently clear. Our contributions beyond these existing approaches are: 1) constructing prediction sets that jointly combine forecast uncertainty with proximity to an empirical data manifold, and 2) developing a particle-flow representation of these sets, which would otherwise be intractable sets over spaces of measures.

Thurin et al. and Klein et al. use optimal transport to construct ranks from multivariate nonconformity scores and form prediction sets for multivariate data. In our setting, SSW distance measures discrepancy between a candidate distribution and the forecast distribution, producing a scalar nonconformity score. These approaches use OT to rank multivariate scores for multivariate responses, whereas we use an OT metric to define a scalar score between distribution responses.

We were not aware of Feldman et al. at the time of writing and have now added this work to the discussion. Their nearest neighbor construction is also distinct from our manifold restriction. We define

\in \arg\min_{\nu:\operatorname{supp}(\nu)\subseteq A} W_2(\mu,\nu),$$ followed by $$d_{\mathrm{pool}}(\mu;A) = SSW(\mu,\nu_A^*(\mu)),$$ which is minimized when the individual events comprising the candidate distribution are close to historically observed locations. Feldman et al. instead use $$d_F(\mu)=\min_j W_2(\mu,\widehat P_j),$$ which favors candidate distributions resembling a single observed training distribution. We initially considered this latter construction, but found through the particle flow that it produced underdispersed samples. Moreover, our prediction sets are centered around the forecast distribution and then restricted by empirical spatial support, whereas Feldman et al. construct balls around training observations. These therefore represent different notions of uncertainty. A further technical contribution is the representation of the resulting manifold-intersected prediction sets. We extend the flow-based representation of Harris (2026) to the intersection of a conformal ball and an empirical support restriction. This requires finding point clouds that simultaneously satisfy the chosen conformal level and the tolerance for proximity to pooled training locations. We derive a joint particle velocity enforcing both conditions. Repeating this construction across conformal levels produces a predictive ensemble of spatial distributions. This goes beyond geometric characterization or membership testing by providing sampled forecast scenarios that can be directly evaluated using distributional scores. The closest general framework to ours is Lugosi and Matabuena (2025), which we had not cited originally and have now included following reviewer QFR5. Their work develops conformal prediction for general metric spaces, and our ambient construction using spherical sliced Wasserstein distance can therefore be viewed as an instance of metric conformal prediction. Their framework already accommodates distribution-valued responses, which we now clarify in Section 3.1. Our additional contribution is to restrict this ambient prediction set using proximity to the pooled spatial support of the training data. Their kNN procedure selects observations with nearby covariates to estimate a local prediction radius. Our nearest neighbor operation instead matches individual candidate event locations to historical event locations, imposing an additional constraint on the spatial distributions admitted by the prediction set. We then provide the particle-flow construction needed to represent this intersection in practice.

Thanks for your explanations and the additional results.

Reviewer_QFR5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces a conformal prediction method for spatial distributions by relying on geometric inductive biases and optimal transport. The problem is relevant, and the method combines existing tools in a potentially interesting way. However, the technical novelty of the work is not clear, and there are several issues with the presentation and discussion of the method. Empirical results are not entirely convincing.

Critical strengths
  • Interesting approach that tackles a relevant problem by combining existing tools.
Critical weaknesses

Both the presentation and the discussion have several issues.

  • The technical novelty of the proposed approach is not clear, and comparison with existing methods is limited.

    • Discussion of related work is mostly limited to the introduction.
    • The method proposed in 3.1 shares many similarities with existing methods. What's the relationship between the proposed method and approaches to conformal prediction in metric spaces such as [1]?
    • It is not clear what the technical contribution of the proposed approach is beyond the constraints introduced in Section 3.2, which seems to be a modest contribution. There should be a clear discussion of what aspects of the work are technically novel and what constitutes an adaptation of existing methods. The novel contributions are not clearly spelled out, not even in the introduction.
    • Some of the references seem misrepresented, e.g., Tibshirani et al. 2019 and Barber et al. 2023 do not seem the best references to discuss "spatially adaptive conformal methods".
    • The paper is not self-contained. Section 3.3 heavily relies on the method presented in Harris 2026, and it is difficult for a reader to fully grasp and understand the proposed approach from Section 3.3 alone.
  • The limitations of the approach discussed in 3.3 are not fully discussed and clearly acknowledged. The paper comments on how the imposed constraints could be incompatible but does not fully acknowledge that, to my understanding, the samples obtained by such sampling might not even belong to the set in Eq. 9. This should be checked at least in the empirical analysis. Also, the approach described in lines 250 to 253 is not really clear.

  • Please improve the discussion on the impact of the i.i.d. assumption made in some of the theoretical results in relation to the real-world data.

Similarly, the empirical analysis is not entirely convincing:

  • Tables 2 and 3 in the main text do not report standard errors, which are instead reported for the same results in the appendix. The appendix shows that the errors are often significantly larger than the difference between the reported metrics. Please comment on this and explain how the errors were obtained. It is not clear whether the results are meaningful.

  • In the experiment on synthetic data, which is the benchmark where a good amount of data is available, the performance of CVF is barely different w.r.t. VF and has high deviations. Moreover, VF clearly outperforms CVF on the Chicago benchmark, which is reported only in the appendix.

  • Coverage results with and without the manifold constraint should be reported and better discussed for the experiments on real-world data for all the baselines.

Minor issues

  • In the empirical results, the meaning of the Pred–Pred and True–Pred rows has to be inferred, with no clear definition in the text.
  • The paper calls "concurrent work" papers from 2 years ago, which is questionable at best.

[1] Gábor Lugosi, Marcos Matabuena, “Conformal and kNN Predictive Uncertainty Quantification Algorithms in Metric Spaces”, arxiv 2025

Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Critical Weaknesses Addressed by Reviewer QFR5

"Similarly, the empirical analysis is not entirely convincing"

"The appendix shows that the errors are often significantly larger than the difference between the reported metrics."

We apologize for this mistake. After careful code review, we realize we only reported standard deviation for the real data instead of standard error. We will divide these by their respective test sizes. We will replace these by

SE⁡^(mˉ)=sn,\widehat{\operatorname{SE}}(\bar m) = \frac{s}{\sqrt{n}},

using the standard error consistently throughout the manuscript and appendix.

"In the experiment on synthetic data, which is the benchmark where a good amount of data is available, the performance of CVF is barely different w.r.t. VF and has high deviations"

Yes, we agree that the simulations do not show a uniform advantage of CVF over VF. This is consistent with the intended role of the manifold constraint. CVF is not designed to improve upon VF when the empirical support contains little additional structure. In such settings, CVF can incur a modest cost from imposing an unnecessary restriction. The Chicago example is an important instance of the latter. We will present Chicago as a limiting case illustrating when the additional constraint is unneces- sary rather than as a setting in which CVF should improve VF.

To more directly evaluate the setting in which the manifold constraint is intended to improve upon VF, we have additionally included two simulations that contain recurring spatial support structure and regions of the sphere in which events should not occur. These are intended was targeted evaluations of the mechanism introduced by CVF rather than replacements of the existing simulations. These two simulations also more closely resemble the spatial structure present in our tropical cyclone and earthquake datasets, but with substantially more samples.

Table. Results for the shifting mode curve simulation with 50 calibration and 50 test clouds of 50 points. Values in parentheses are standard errors computed across the 50 test clouds.

MethodED (×102)(\times 10^2)Pred--PredPred--TrueMD (×102)(\times 10^2)In region (%)
M1: Minimum Area10.59 (0.764)0.0572 (0.0000)0.1657 (0.0052)1.823 (0.017)10.04 (0.64)
M2: Conformal Minimum Area11.08 (0.796)0.0510 (0.0000)0.1651 (0.0054)1.917 (0.019)10.84 (0.60)
M3: Diffusion4.03 (0.679)0.1230 (0.0000)0.1659 (0.0048)1.972 (0.071)56.00 (1.04)
M4: Flow2.35 (0.539)0.1349 (0.0000)0.1634 (0.0041)4.509 (0.090)30.52 (0.95)
M5: VAE7.63 (0.735)0.0789 (0.0000)0.1618 (0.0051)2.401 (0.026)32.32 (0.86)
VF: Velocity Field1.25 (0.607)0.1318 (0.0000)0.1563 (0.0045)3.076 (0.229)48.56 (4.12)
CVF: Constrained Velocity Field1.31 (0.629)0.1307 (0.0000)0.1561 (0.0046)2.405 (0.196)67.08 (3.32)

CVF increases the proportion of generated points in valid regions from 48.5648.56% for VF to 67.0867.08% while also decreasing the manifold distance from 3.076 to 2.405 while maintaining similar ED.

Table. Results for the independent fault support simulation with 50 calibration and 50 test clouds of 80 points. Values in parentheses are standard errors computed across the 50 test clouds.

MethodED (×102)(\times 10^2)Pred--PredPred--TrueMD (×102)(\times 10^2)In region (%)
M1: Minimum Area3.99 (0.771)0.0345 (0.0000)0.1012 (0.0052)1.118 (0.016)48.14 (0.78)
M2: Conformal Minimum Area4.17 (0.785)0.0338 (0.0000)0.1017 (0.0053)1.175 (0.017)47.70 (0.73)
M3: Diffusion1.06 (0.563)0.0808 (0.0000)0.1097 (0.0042)2.069 (0.080)34.34 (1.03)
M4: Flow1.33 (0.434)0.0867 (0.0000)0.1140 (0.0035)3.312 (0.092)26.28 (0.93)
M5: VAE1.85 (0.629)0.0692 (0.0000)0.1078 (0.0045)1.791 (0.028)30.90 (0.60)
VF: Velocity Field0.19 (0.489)0.1005 (0.0000)0.1152 (0.0039)1.564 (0.117)51.44 (4.34)
CVF: Constrained Velocity Field0.16 (0.489)0.1001 (0.0000)0.1149 (0.0039)1.167 (0.090)57.84 (4.05)

CVF increases the proportion of points in the valid region from 51.4451.44% to 57.8457.84% for the fault support simulation all while decreasing the manifold distance from 1.564 to 1.167 again with similar ED. Both experiments illustrate the intended distinction that the manifold constraint provides additional information when recurring empirical support structure is not fully captured by the distributional score.

"Coverage results with and without the manifold constraint should be reported and better discussed for the experiments on real-world data for all the baselines."

Yes, multiple reviewers shared the same sentiment. We included an official comment describing where we will include the ambient ball results, manifold containment, and intersection of the two regions coverages.

AuthorsOpenReview ↗

Responses to Critical Weaknesses and Minor Weaknesses Addressed by Reviewer QFR5

Minor Issues

"the meaning of the Pred–Pred and True–Pred rows has to be inferred"

"The paper calls concurrent work papers from 2 years ago, which is questionable at best"

Thank you for the suggestions of presentation and wording. We will replace "concurrent work" with "recent related work" where appropriate and define Pred--Pred and True--Pred explicitly at first use in the table subtext.

Critical Weakness on 3.3 and lines 250-253 clarity

Yes, we agree and apologize for the clarity of 3.3. As other reviewers have noticed and we posted an official comment regarding the issue, as the used manifold penalty was the one described in 250-253, which is described in the official comment.

"the samples obtained by such sampling might not even belong to the set in Eq. 9"

Yes, reviewer LmYu shared a similar sentiment. We will include the tropical cyclone genesis and earthquake CVF sample coverage along with the CVF sample coverages for the real data in the appendix of the updated manuscipt.

Self Containment using Harris 2026

Yes, you are correct in which Section~3.3 assumes too much familiarity with Harris (2026). We will first state the single score boundary dynamics used as the computational starting point and then derive our multivariate approach and extension of the method. The statistical prediction set will be defined before the numerical sampler so that the two roles are clearly separated.

IID Assumption for Real Data

We agree that this should be discussed more explicitly. The conformal guarantee requires exchangeability rather than strict temporal independence. We therefore will connect the temporal tests in the appendix to the real data discussion and clarify that they provide evidence against temporal dependence, but do not necessarily prove exchangeability of the data.

Incorrect Citation

Thank you for identifying this citation error. Tibshirani et al. (2019) and Barber et al. (2023) should not have been described as spatially adaptive conformal methods. We will retain them only in the appropriate covariate shift or nonexchangeability discussion and replace the spatial conformal citations with work that directly studies spatial prediction.

AuthorsOpenReview ↗

Response to Reviewer QFR5 Regarding Novelty of Method

"The technical novelty of the proposed approach is not clear, and comparison with existing methods is limited"

"What's the relationship between the proposed method and approaches to conformal prediction in metric spaces"

"not clear what the technical contribution of the proposed approach is beyond the constraints introduced in Section 3.2"

We apologize that we did not make the distinction between our work and prior work sufficiently clear in the appropriate part of the manuscript. Our contributions beyond existing approaches are: 1) constructing prediction sets that jointly combine forecast uncertainty with proximity to an empirical data manifold, and 2) developing a particle-flow representation of these sets, which would otherwise be intractable sets over spaces of measures.

Thurin et al. and Klein et al. suggested by reviewer LmYu, use optimal transport to construct ranks from multivariate nonconformity scores and form prediction sets for multivariate data. In our setting, SSW distance measures discrepancy between a candidate distribution and the forecast distribution, producing a scalar nonconformity score. These approaches use OT to rank multivariate scores for multivariate responses, whereas we use an OT metric to define a scalar score between distribution responses.

We have added Feldman et al. to the discussion. Their nearest neighbor construction is also distinct from our manifold restriction. We define

\in \arg\min_{\nu:\operatorname{supp}(\nu)\subseteq A} W_2(\mu,\nu),$$ followed by $$d_{\mathrm{pool}}(\mu;A) = SSW(\mu,\nu_A^*(\mu)),$$ which is minimized when the individual events comprising the candidate distribution are close to historically observed locations. Feldman et al. instead use $$d_F(\mu)=\min_j W_2(\mu,\widehat P_j),$$ which favors candidate distributions resembling a single observed training distribution. We initially considered this latter construction, but found through the particle flow that it produced underdispersed samples. In addition, our prediction sets are centered around the forecast distribution and then restricted by empirical spatial support, whereas Feldman et al. construct balls around training observations. These therefore represent different notions of uncertainty. A further technical contribution is the representation of the resulting manifold intersected prediction sets. Rather than simply applying the scalar flow of Harris (2026), we derive a multi-constraint extension in which the dynamics are governed jointly by the conformal score and the manifold discrepancy. This yields a multidimensional constraint system and a corresponding joint particle velocity that simultaneously drives candidate point clouds toward the chosen conformal level while enforcing proximity to the pooled empirical support. Repeating this construction across conformal levels produces a predictive ensemble of spatial distributions. Thus, our novelty is not only the definition of the intersected prediction set, but also a new constrained velocity field construction for sampling from sets defined by multiple distributional constraints. By explicitly introducing the scalar construction of Harris (2026) in section 3.2 as you suggested before deriving our multi-constraint extension, we believe the revised manuscript makes the additional velocity field contribution substantially clearer. We agree that Lugosi and Matabuena (2025), would be one of the most similar works and should be cited in section 3.1. Their work develops conformal prediction for general metric spaces, and our ambient construction using spherical sliced Wasserstein distance can therefore be viewed as an instance of metric conformal prediction. Their framework already accommodates distribution-valued responses, which we now clarify in Section 3.1. Our additional contribution is to restrict this ambient prediction set using proximity to the pooled spatial support of the training data. Their kNN procedure selects observations with nearby covariates to estimate a local prediction radius. Our nearest neighbor operation instead matches individual candidate event locations to historical event locations, imposing an additional constraint on the spatial distributions admitted by the prediction set. We then provide the aforementioned particle-flow construction needed to represent this intersection in practice.
Reviewer_ZqJkOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduced a conformal prediction method for spatial distributions observed through finite samples of events, such as tropical cyclone genesis and earthquake locations. The underlying idea scores empirical measures using the spherical sliced Wasserstein distance, then constrains the resulting set of predictions to lie near the training data manifold. The paper also derives a finite-sample coverage lower bound for future empirical measures using a data-adaptive manifold tolerance. Finally, it introduces a modified flow-based sampling procedure that represents candidate distributions by particle clouds and produces an ensemble over spatial distributions.

Critical strengths
  1. The idea of using the entire spatial response is well motivated and captures the uncertainty that scalar summaries or event-wise geographic areas might miss.
  2. The theory provides a nice result on the coverage lost through the training-based restriction.
  3. The experiments, such as on cyclones and earthquake data, suggest improvement in geographic concentration, whilst retaining the distributional scores.
Critical weaknesses
  1. It appears that the discrepancy used in implementation is not the same as the one analysed here. Eq. 8 seems to define manifold discrepancy as the minimum SSW distance to the entire training empirical measure. However, looking at Section 3.3. it describes constructing a reference cloud by matching each candidate point to its nearest point in the pooled training support. Aren't these operations different? The second one, in particular, can combine points from multiple training observations and therefore need not enforce proximity to any one observed distribution. Also, Section 4.2 states that the reported MD is this local numerical discrepancy. Could the authors clarify/explain the relationship between the two quantities and confirm whether generated samples satisfy the actual constraint in Eq. 9?
  2. Can the authors comment on whether the geographic plausibility can be evaluated independently of the discrepancy optimised by CVF?
  3. Can the authors comment on whether the reported coverage distinguishes ambient-ball coverage from coverage after applying the manifold restriction?
Questions for authors

Please see the weaknesses above.

1 comment
AuthorsOpenReview ↗

Response to Critical Weaknesses Addressed by Reviewer ZqJk

Thank you again for raising these concerns regarding the methodology.

1) Inconsistency in the manifold constraint definition

Yes, we agree and apologize for the inconsistency between 3.3 and Eq. 8. All three reviewers have found this inconsistency, thus we decided to make an official comment updating our resolution to this.

“The second one, in particular, can combine points from multiple training observations and therefore need not enforce proximity to any one observed distribution.”

We agree with this characterization. Our goal with the manifold constraint is to enforce proximity of the candidate locations to the empirically observed spatial manifold, rather than proximity to any single observed distribution. The ambient SSW constraint already requires the candidate distribution to remain within the prediction region around the forecast. The pooled support constraint then adds complementary historical structure by restricting the candidate towards locations represented in the observed data. Due to the pooled support containing multiple observations, the construction preserves variability along the empirical manifold rather than tying it to one reference cloud. This methodology, does however, change underlying theory. We have updated the definition in Eqn ~8 and updated Proposition 3.1 to use this definition. Proposition 3.1 only changes in the 1/(n+1)1/(n + 1) term, which is now 1/(ntrain+1)1/(n_{\rm{train}} + 1) instead of 1/(ncal+1)1/(n_{\rm{cal}} + 1).

2) Geographic Plausibility

Yes! We added independent geographic validation for both real data applications in an official comment. For earthquakes, we provide proximity to active fault traces from the GEM Global Active Faults Database. For tropical cyclone genesis, we construct a non convex geographic support from the independent historical GDEX dataset and evaluate how often generated and observed genesis locations fall into theses regions. Both corroborate that CVF has successfully increased proximity to geographic structure relative to VF and other methods.

3) Ambient Ball Coverage

Yes, we agree that coverage of both the manifold constraint and intersection should be reported. All 13 held out test observations for tropical cyclone genesis and all 15 held out test observations for earthquake genesis fall within the manifold tube, therefore the intersection of the regions remain what the ambient SSW ball coverages are which are 11/1311/13 for tropical cyclone genesis and 14/1514/15 for earthquakes. The earthquake value was corrected after rechecking the ambient coverage, which had been incorrectly stated as 15/1515/15 in the manuscript. Thus ambient coverage before and after applying the constraint to the real data remains unchanged. We will include this information in the manuscript, as described in our official author comment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Manifold Constraint inconsistency, Geographic Plausibility, and Ambient Coverage

1. Inconsistency in the manifold constraint definition

We would like to thank the reviewers for identifying this large inconsistency in our presentation, and we apologize for the confusion this has caused. Our simulations and real data experiments all use the Section 3.3 definition of the manifold constraint, which was not fully defined. The correct definition of the manifold penalty is

\arg\min_{\nu:\operatorname{supp}(\nu)\subseteq A} W_2(\mu,\nu),$$ followed by $$d_{\rm pool}(\mu;A)=SSW(\mu,\nu_A^*(\mu)).$$ Where $A$ is the pooled calibration reference support. The definition in Eqn. 8 was a hold over from a previous iteration of our coverage lower bound and should have been replaced with the above definition. We have updated the definition in Eqn. 8 and updated Proposition 3.1 to use this definition. Proposition 3.1 only changes in the $1/(n+ 1)$ term, which is now $1/(n_{\mathrm{train}} + 1)$ instead of $1/(n_{\mathrm{cal}} + 1)$. # 2. Independent geographic plausibility We have now added external geographic validations for both real data applications in Appendix G.6 and G.7. For tropical cyclones, we construct geographic support regions by drawing nonconvex hulls around the historical GDEX genesis locations (https://gdex.ucar.edu/datasets/d824001/). We validate the plausibility of these regions using our dataset, IBTrACS (https://www.ncei.noaa.gov/products/international-best-track-archive). As shown in Table 1, CVF places $82.86\%$ of generated points within the geographic support, compared with $76.16\%$ for VF. This is close to the independent IBTrACS tropical cyclone genesis observations, for which $94.96\%$ of observed locations lie within the same support regions. **Table 1. Tropical cyclone geographic support based on historical GDEX genesis locations.** "Points in support" is the pooled percentage of genesis locations within the fixed support regions. Standard errors are provided in parentheses $(s/\sqrt{n})$. | Method | Points in support (%) | Mean per cloud (%) | SE (%) | |---|---:|---:|---:| | IBTrACS observed genesis clouds | 94.96 | 94.95 | 0.95 | | M1: Minimum Area | 82.21 | 82.21 | 1.32 | | M2: Conformal Minimum Area | 81.54 | 81.54 | 1.30 | | M3: Diffusion | 63.94 | 63.94 | 1.53 | | M4: Flow | 45.10 | 45.10 | 1.85 | | M5: VAE | 62.98 | 62.98 | 1.55 | | VF: Velocity Field | 76.16 | 76.35 | 2.54 | | **CVF: Constrained Velocity Field** | **82.86** | **82.97** | 1.61 | For earthquakes, we use active fault traces from the GEM Global Active Faults Database (https://github.com/GEMScienceTools/gem-global-active-faults). Each observed or generated earthquake location is assigned its pointwise distance to the nearest trace in the fixed relevant fault system. As shown in Table 2, CVF improves upon VF in mean distance to a mapped fault, decreasing from $83.79$ km to $74.34$ km, median distance, decreasing from $70.85$ km to $57.85$ km, and the fraction of generated locations within $5$ km of a mapped fault, increasing from $1.66\%$ to $4.35\%$. **Table 2. Pointwise distance of observed and generated earthquake locations to the relevant GEM active fault system.** Lower mean and median distances and a larger fraction within 5 km indicate greater geographic consistency with the mapped fault structure. | Method | Mean distance (km) | Median distance (km) | Within 5 km (%) | |---|---:|---:|---:| | NOAA NCEI observed earthquakes | 7.40 | 3.03 | 56.00 | | M1: Minimum Area | 60.68 | 60.24 | 4.27 | | M2: Conformal Minimum Area | 64.29 | 62.41 | 4.04 | | M3: Diffusion | 285.71 | 208.17 | 0.45 | | M4: Flow | 91.96 | 86.20 | 1.19 | | M5: VAE | 150.53 | 117.34 | 1.16 | | VF: Velocity Field | 83.79 | 70.85 | 1.66 | | **CVF: Constrained Velocity Field** | **74.34** | **57.85** | **4.35** | # 3. Ambient versus manifold-restricted coverage We have revised Sections 4.3 and 4.4 to indicate that coverage on the real data does not change if we include or exclude the manifold constraint, because all members of both the tropical cyclone genesis and earthquake datasets fall within the manifold tube. However, as shown in Table 1, including the manifold constraint can improve geographic plausibility.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.