acceptodds
Under review as a conference paper at ICLR 2027

Spherical Harmonic Optimal Transport: Application to Climate Models Comparisons

Abstract

Optimal transport provides a powerful framework for comparing measures while respecting the geometry of their support, but comes with an expensive computational cost, hindering its potential application to real world use cases. On manifolds, convolutional algorithms based on the heat kernel have been proposed to alleviate this cost, but their theoretical properties remain largely unexplored. We establish that the heat kernel cost converges to the optimal transport cost as time vanishes in the balanced and unbalanced cases. In the specific case of the 2-sphere , we ensure that the associated Sinkhorn divergences retains the desirable geometric and analytic properties of classical optimal transport discrepancies. Moreover, we leverage the harmonic structure of the sphere to derive a fast Sinkhorn algorithm, requiring only memory and time per iteration, with fully dense GPU-friendly operations. We validate its computational efficiency on synthetic data, and discuss its potential use in the evaluation of global climate models, providing both spatial and seasonal insights into models performances.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
optimal transport

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_iVtBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies entropic balanced and unbalanced optimal transport on the 2-sphere, where the Gibbs kernel of the squared geodesic distance is replaced by the heat kernel at time ε/4, as in the convolutional OT framework of Solomon et al. (2015). The theoretical contributions are twofold. First, Theorem 3.1 shows that, on any compact connected Riemannian manifold, the entropic OT problem with the heat-kernel-induced cost C^ε=−εlog⁡Hε/4\hat C_\varepsilon = -\varepsilon \log H_{\varepsilon/4} converges to the unregularized OT problem with squared geodesic cost as ε → 0, in both the balanced and unbalanced settings. Second, Proposition 3.2 establishes that the associated Sinkhorn divergence on S² (and more generally on compact symmetric spaces) is symmetric, positive definite, convex in each argument, and metrizes weak convergence. Algorithmically, the authors observe that the heat kernel is radial and acts diagonally in the spherical harmonic basis. Each kernel application in the Sinkhorn iterations therefore reduces to a forward spherical harmonic transform, a multiplication by e−tℓ(ℓ+1)e^{-t\ell(\ell+1)}, and an inverse transform. The resulting method, SHOT, requires O(n) memory and O(n^{3/2}) time per iteration and runs on GPU. Validation consists of runtime benchmarks on random synthetic measures, using the Clenshaw–Curtis and HEALPix grids, against kernel-matrix Sinkhorn, a Cholesky-based convolutional solver, exact W₂, and two sliced methods (SSW and LSSOT). It also includes a comparison between heat and matrix convolutions in the appendix. Finally, an application compares the 2000–2014 seasonal precipitation climatologies of eight CMIP6 models with ERA5 on a HEALPix grid with Nside=64N_{\text{side}} = 64. The authors report model rankings under balanced and unbalanced divergences, regional bias indices derived from the gradient of the divergence, and an analysis of Arctic winter biases near the sea-ice edge.

Critical strengths
  1. Theorem 3.1 gives a rigorous justification for replacing the Gibbs kernel with the heat kernel in Sinkhorn iterations, a substitution that has been used heuristically since Solomon et al. (2015). The result covers the unbalanced case and does not follow directly from standard entropic OT convergence results, because the cost itself depends on ε. Together with Proposition 3.2, it also settles a real theoretical problem: the Gibbs kernel of the squared geodesic distance is not positive definite on the sphere, so the standard Sinkhorn divergence guarantees do not apply to it. The heat-kernel formulation restores them. This is the main reason I view the paper as a meaningful contribution.

  2. The algorithm is simple and practical. By applying the heat kernel as a diagonal filter in the spherical harmonic basis, SHOT avoids forming any kernel matrix, uses linear memory, and relies only on dense, parallel operations. Figure 1 shows that it scales to about 523,000 grid points, while kernel-matrix and Cholesky-based solvers run out of memory at moderate resolutions. Since the output is a differentiable divergence, the method could be useful well beyond the climate application presented here.

Critical weaknesses

The accuracy of the method at practical values of ε is not assessed. Theorem 3.1 is an asymptotic statement without a rate, and Appendix C.4 shows that the algorithm becomes unstable when ε is small relative to the spherical harmonic bandwidth, since the truncated heat kernel takes negative values. The only quantitative accuracy check (Appendix D.2) compares single convolutions after rescaling, not the resulting transport costs or divergences. As a result, it is unclear how closely the divergence computed at ε = 0.01 in the climate study reflects the squared-geodesic OT problem that motivates the work. This matters because the scalability argument depends on SHOT being a faithful substitute for standard Sinkhorn, and that has not been shown.

The climate application does not establish that the proposed metric is more informative than existing alternatives, and SHOT is only used as a post-hoc evaluation tool. The study includes no baselines (RMSE, pattern correlation, balanced OT metrics, sliced spherical OT, or the spherical convolutional Wasserstein distance of Garrett et al. (2024)). It also provides no external evidence that the unbalanced ranking, which differs substantially from the balanced one, is the more faithful assessment. Several conclusions, such as links to sea-ice parameterization complexity and the need for changes to convection schemes, are not supported by the analysis. Although the authors motivate SHOT as a differentiable, geometry-aware objective, it is never used for learning, and there is no test of whether it predicts out-of-sample skill. Since the climate study is the main evidence of practical value, these gaps substantially weaken the empirical contribution.

Questions for authors

Clarifications:

  • At ε = 0.01, how close is SHOT’s Sinkhorn divergence to the exact W₂ cost and to standard Sinkhorn with the geodesic Gibbs kernel? A comparison on grids where all methods fit in memory, across a range of ε, would address my main concern. Do the model rankings change if the exact or Gibbs-kernel divergence is used instead?
  • Can a convergence rate be given for Theorem 3.1, or at least an empirical characterization of the gap as a function of ε and the bandwidth L?
  • How do the model rankings compare with RMSE, pattern correlation, balanced OT, and Garrett et al. (2024)? Is there any external evidence, for example from established assessments of these CMIP6 models, that the unbalanced ranking is more reliable?
  • Are the values in Figure 3 normalized in the same way as those in Figure 13? From Figure 13, the Arctic DJF−JJA difference appears to be around 0.3 to 0.4 rather than 0.06 to 0.09. Also, why are the regions described as non-overlapping when the ITCZ lies inside the Tropics?
  • On the HEALPix grid used in the climate study, SHOT is slower than several baselines. What is its runtime if the data are regridded to the Clenshaw–Curtis grid, and how many Sinkhorn iterations does each method require?

Suggestions:

  • The paper motivates SHOT as a differentiable objective, but its usefulness for building better predictive models is not demonstrated. One direct test would be to use it as a training loss for a spherical forecasting model, such as an SFNO trained on ERA5 precipitation. The model could be evaluated on held-out years with independent metrics (power spectra, Fractions Skill Score, extreme-precipitation statistics) and compared with MSE and spectral losses. A less expensive test, using the CMIP6 data already collected, would be a perfect-model experiment. Each model would be treated in turn as pseudo-truth, the others weighted by their historical divergence, and the weighted projections checked against the pseudo-truth’s future period, comparing SHOT-based weights with RMSE-based weights. Either experiment would show whether the metric captures aspects of model quality that matter for prediction.
  • Please show a plain difference map alongside the OT gradient map, to support the claim that the gradient localizes displacement errors that a difference map cannot.
2 comments

We thank the reviewer for all the comments that hopefully helped us improve the clarity of our paper and the scope of the experimental results.

Climate models comparison.

Comparison with other methods (C1&C3)

We added this comparison in Section 4.2.2 of the revised version. We rank the 8 models with SHOT and with the absolute global-mean bias, RMSE, pattern correlation, RMSE after smoothing both fields with the same heat kernel as SHOT and the original spherical sliced Wasserstein distance (SSW). The balanced version of SHOT agrees with SSW (Spearman correlation 0.98 in DJF, 0.62 in JJA) and only weakly with RMSE and pattern correlation (0.48 in DJF, −0.17 in JJA) which is consistent with the double penalty of pointwise scores on displaced features. The unbalanced divergence moves between two known regimes as τ\tau varies: for large τ\tau its ranking matches the ranking by global-mean bias and for small τ\tau that of the smoothed RMSE. Large rank changes between the balanced and unbalanced divergences therefore reflect global-mean biases.

Garrett et al. (2024) compare, at each location, the distribution of daily values over time with one-dimensional OT. Their metric needs daily time series and cannot be computed on the seasonal climatologies we use. It also answers a different question (the distribution of values at each point, rather than the spatial pattern), so we see it as complementary.

We have no external evidence that the unbalanced ranking is more reliable, and we do not claim it: ERA5 is itself an imperfect reference, and the balanced and unbalanced divergences answer different questions that are selected by τ\tau. What we can show is what each metric responds to. In a new experiment (Appendix E.4), we add errors of known type and size to ERA5: a shift of the tropical rain band, a global or tropical scaling of the intensity, and a blur. When the band is shifted, the balanced divergence grows as the square of the shift, while RMSE grows more and more slowly once the shifted band stops overlapping its original position (double penalty). Balanced SHOT, SSW, LSSOT and pattern correlation are exactly blind to a global scaling, and the unbalanced divergence puts displacement errors on its transport part and intensity.

Comparison with "classical" Sinkhorn.

Rate of convergence (C2).

Yes, in the balanced case. We added it to the revised version (Remark 3.3, with the proof in Appendix A.2). For probability measures P,QP,Q on S2\mathbb S^2 and 0<ε<10<\varepsilon<1 we have ∣OTC^ε,ε(P,Q)−W22(P,Q)∣≤32 εlog⁡(1/ε)+O(ε).|\mathrm{OT}_{\widehat C_\varepsilon,\varepsilon}(P,Q)-W_2^2(P,Q)|\le\frac32\,\varepsilon\log(1/\varepsilon)+\mathcal O(\varepsilon). The proof combines two results from the literature: the sharp two-sided estimates of the spherical heat kernel of Nowak, Sjögren and Szarek (2019) and the convergence rate of entropic optimal transport for Lipschitz costs on compact manifolds of Carlier, Pegon and Tamanini (2023).

This result gives an order in ε\varepsilon, not a numerical bound: the constants of the heat kernel estimates are not explicit and as Theorem 3.1, it concerns the exact heat kernel and not its degree-LL truncation. For this reason, the accuracy at the values of ε\varepsilon and LL used in practice is measured empirically, in the comparison described below.

Small scale comparison (C1).

We added this comparison in Appendix D.3 of the revised version (Table 2): we compared SHOT against a dense log-domain Sinkhorn solver of Eq. (2) with the Gibbs kernel e−d2/εe^{-d^2/\varepsilon} on the same grid, on 10 pairs of random measures (Clenshaw–Curtis L=96L=96 and HEALPix Nside=32N_{\rm side}=32), balanced and unbalanced (τ=100\tau=100), for ε∈{0.01,0.1,1,10}\varepsilon\in\{0.01,0.1,1,10\}. For ε≤0.1\varepsilon\le0.1, which covers the settings of our experiments, the two solvers agree closely: the relative differences of the divergence values and of their gradients are below 1%. The difference grows for larger ε\varepsilon because the heat kernel Hε/4\mathcal H_{\varepsilon/4} moves away from the Gibbs kernel, not because of the solver. Table 2 also reports the gap of both divergences to the exact (unregularized) W22W_2^2: it is the same for SHOT and for the dense solver, so it comes from the entropic regularization and not from SHOT.

Runtime after regridding (C5).

We took the 16 divergences of the climate study (each of the 8 models against ERA5, in DJF and JJA) and computed each one twice with SHOT: once on the HEALPix grid (Nside=32N_{\rm side}=32, 12,288 points), and once after moving both fields to the Clenshaw–Curtis grid with about the same number of points (L=79L=79, 12,403 points) by using an area-weighted average over the HEALPix pixels overlapping each Clenshaw–Curtis cell, rescaled in a way to keep the total mass. Settings: ε=0.05\varepsilon=0.05, τ=100\tau=100 (unbalanced) or τ=∞\tau=\infty (balanced) and with threshold 10−910^{-9} checked every 10 iterations for the convergence and at most 1000 iterations in total. Times are for one transport problem (model vs ERA5) on a Tesla V100 GPU with medians over the 16 pairs.

griditerationstime (s)time / iteration (ms)relative difference due to regridding ∣SCC−SHP∣/SHP\mid S_{\rm CC}-S_{\rm HP} \mid /S_{\rm HP}
balancedHEALPix5965.49.1–
CC (regridded)5961.01.60.6%
unbalancedHEALPix4814.28.7–
CC (regridded)4760.81.80.2%

SHOT needs almost the same number of iterations on both grids, so the gain comes entirely from the cost of one iteration: as discussed in Section 4.1, this reflects the maturity of the two spherical harmonic transform libraries: torch-harmonics, used on the Clenshaw–Curtis grid, provides optimized CUDA kernels while s2fft, used on HEALPix, does not yet. Regridding takes roughly 1 ms and only slightly changes the divergences values.

Differentiable objective (S1).

We agree that the paper in the current state does not showcase SHOT as a training objective and it is left for future work (Sections 4.2 and 5). Both experiments the reviewer suggest are what we could do next, but each is a study of its own: training a spherical neural operator on ERA5 precipitation with SHOT and comparing it with MSE and spectral losses on held-out years requires a full training and evaluation pipeline, and the perfect-model test requires future-scenario runs for all 8 models. Neither of those experiments could be done within the discussion period. As a first step, in the resubmitted version of the paper, Appendix D.4 shows SHOT used as a differentiable loss: interpolations between measures on S2\mathbb S^2 obtained only by gradient descent through the SHOT divergence, which move mass along great circles, whereas the L2L^2 interpolation fades one configuration into the other. This simple example stands as an empirical validation of the gradients SHOT produced, and to some extent validates it can be used in a learning pipeline.

Gradient maps vs. difference maps (S2).

We added this comparison in the revised version (Section 4.2.3 and Appendix E): Figure 12 shows the plain difference map next to the SHOT gradient maps, and Table 5 gives, for each (model, season, τ\tau) the correlation between the gradient and the difference map, raw or smoothed with the heat kernel.

Reviewer_ft5KOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This work studies the problem of computing optimal transport maps on compact, connected Riemannian manifolds. The authors first formalize the connection between heat kernel entropic regularized unbalanced OT and exact unbalanced OT on the manifold and that Sinkhorn diverges are well defined at the heat kernel approximation. Then, using the spherical harmonic transform, the introduce a Sinkhorn algorithm for computing the entropic UOT map on the sphere at an O(n^{3/2}) cost, which is substantially faster than the naive O(n^2) cost. Experiments show that this approach scales to much larger grid sizes than most baseline approaches. Finally, the authors use their approach to detect distributional biases in climate model generated precipitation fields against ERA5 reanalysis data.

Critical strengths
  1. This paper is exceptionally well written and easy to follow. The problem is well motivated, and the proposed method is rigorously developed through a series of theoretical results culminating in a fast, scalable algorithm. The proofs are quite technical, but straightforward enough to follow and appear to be correct.

  2. This paper also included a well-designed application using their method to validate climate models against historical observations. This is a classic important problem in climate science and precipitation is notoriously hard to assess precisely due to the spatial fidelity issue the authors identify in 4.2.2. The deep dive on Arctic precipitation biases was particularly interesting and connects to known model biases.

Critical weaknesses
  1. Section 4 does not directly establish that SHOT can detect climate model biases that other methods could not. While it is entirely possible that this method does, the experiments do not directly compare biases / ranks/ norms against any baselines methods. For instance, though RMSE is wrong for this application (despite its wide usage here), might it show similar precip gradient attribution maps?
Questions for authors
  1. How do the rankings / detected biases of SHOT compare to baseline methods such as RMSE or other OT methods such as Garrett 2024 or Hyun 2022?

  2. Would it be possible to present a condensed algorithm in Section 3.2? The full algorithm in the appendix is too much, but a small condensed version might be useful for summarizing the actual method.

  3. It would be nice to show the Sinkhorn algorithm, or at least its iterates, since its fundamental to the approach and the paragraph on Convolutional Optimal Transport directly references its iterates and states that they require computing a convolution.

2 comments

We thank the reviewer for the very positive comments.

Comparison with other methods.

We added this comparison in Section 4.2.2 of the revised version. We rank the 8 models with SHOT and with the absolute global-mean bias, RMSE, pattern correlation, RMSE after smoothing both fields with the same heat kernel as SHOT and the original spherical sliced Wasserstein distance (SSW). The balanced version of SHOT agrees with SSW (Spearman correlation 0.98 in DJF, 0.62 in JJA) and only weakly with RMSE and pattern correlation (0.48 in DJF, −0.17 in JJA) which is consistent with the double penalty of pointwise scores on displaced features. The unbalanced divergence moves between two known regimes as τ\tau varies: for large τ\tau its ranking matches the ranking by global-mean bias and for small τ\tau that of the smoothed RMSE. Large rank changes between the balanced and unbalanced divergences therefore reflect global-mean biases.

Garrett et al. (2024) compare, at each location, the distribution of daily values over time with one-dimensional OT. Their metric needs daily time series and cannot be computed on the seasonal climatologies we use. It also answers a different question (the distribution of values at each point, rather than the spatial pattern), so we see it as complementary.

We have no external evidence that the unbalanced ranking is more reliable, and we do not claim it: ERA5 is itself an imperfect reference, and the balanced and unbalanced divergences answer different questions that are selected by τ\tau. What we can show is what each metric responds to. In a new experiment (Appendix E.4), we add errors of known type and size to ERA5: a shift of the tropical rain band, a global or tropical scaling of the intensity, and a blur. When the band is shifted, the balanced divergence grows as the square of the shift, while RMSE grows more and more slowly once the shifted band stops overlapping its original position (double penalty). Balanced SHOT, SSW, LSSOT and pattern correlation are exactly blind to a global scaling, and the unbalanced divergence puts displacement errors on its transport part and intensity.

Presentation of Sinkhorn Algorithm (Q2&Q3). We agree that it could improve the reading. Following this suggestion, we now describe a concise version of Sinkhorn's algorithm in Section 2 (Background). We then present SHOT in Section 3.2, by referring to the steps of Algorithm 1 that it replaces: in the two updates, the kernel operator becomes the heat convolution F−1(h⊙F[⋅])\mathcal F^{-1}(h\odot\mathcal F[\cdot]), hℓ=e−εℓ(ℓ+1)/4h_\ell=e^{-\varepsilon\ell(\ell+1)/4}, and in the returned value, integrals are computed by the grid quadrature. The full version describing SHOT will stay in the appendix. We submitted a new version of the paper for the rebuttal according to those changes.

Thank you for including the additional analysis comparing against the baseline ranking methods. I think its quite interesting how the SHOT rankings vary between the smoothed RMSE and the global mean bias. Of course, SHOT provides a lot more than these approaches, but its good to see some connection with more traditional metrics. The added Sinkhorn algorithm and the explanation in 3.2 are also very appreciated. While I understood the method before, the revised explanation is more self contained so its very easy to see which parts are being modified without having to pull up a separate entropic OT paper for reference.

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

The paper presents Spherical Harmonic Optimal Transport (SHOT), a method for computing heat-kernel-based Sinkhorn divergences between measures on the sphere, including measures with unequal total mass. It establishes convergence of the heat-kernel-regularized transport costs to squared-geodesic optimal transport in both balanced and unbalanced settings, and verifies properties of the associated continuous Sinkhorn divergences. Algorithmically, SHOT uses spherical harmonic transforms to accelerate heat-kernel applications, reporting O(n^{3/2}) time per iteration and O(n) memory. Validation includes synthetic runtime benchmarks on Clenshaw–Curtis and HEALPix grids, supplementary approximation and parameter-sensitivity experiments, and comparisons of eight CMIP6 precipitation models against ERA5. The climate application illustrates model rankings and seasonal differences through divergence scores and gradient-based regional diagnostics.

Critical strengths
  1. SHOT exploits spherical harmonic transforms to apply the heat kernel without explicitly storing a dense pairwise kernel matrix. The benchmarks demonstrate scalability to large grids, and the precipitation application provides a concrete use case involving unequal total mass. This practical capability is the strongest argument for acceptance, provided the numerical accuracy is adequately established.
  2. The paper establishes a small-regularization limit for balanced and unbalanced transport and verifies conditions supporting the continuous heat-kernel Sinkhorn divergence. The unbalanced analysis is particularly relevant to comparing unnormalized physical fields and offers a potentially meaningful contribution beyond computational implementation.
Critical weaknesses
  1. The algorithmic novelty is insufficiently distinguished from directly relevant prior work. Berman, “The Sinkhorn algorithm, parabolic optimal transport and geometric Monge–Ampère equations” (Numerische Mathematik, 2020), §6.3.2 and Theorem 6.2, already considers band-limited spherical heat kernels, forward/inverse spherical Fourier transforms, and Sinkhorn iterations with O(N^{3/2}) complexity, with O(N\log^2N) discussed for equiangular grids. This work is neither cited nor compared against. Consequently, the central computational construction and complexity bound do not establish novelty by themselves. The authors should delineate the additional theoretical and algorithmic contributions and demonstrate the practical advantages of their implementation.
  2. The validation does not sufficiently connect the continuous theory to the implemented solver. The guarantees concern the exact heat kernel, whereas the implementation uses finite spectral truncation and numerical quadrature; the paper acknowledges that truncation can produce negative convolution outputs and invalidate Sinkhorn updates. Runtime comparisons and rescaled single-convolution errors do not establish the accuracy of the final transport values or gradients. Furthermore, Algorithm 1’s unbalanced scaling updates require clarification to reconcile them with the product-reference entropy in Eq. (2). Small-scale comparisons against a direct solver of the stated objective, together with accuracy–runtime and stability evaluations, are necessary to establish that the reported efficiency applies to a correct and reliable approximation.
Questions for authors
  1. How does SHOT’s computational construction differ from Berman (2020)? Which contributions extend beyond this prior work, and which concern implementation or the unbalanced setting?
  2. Can you clarify whether Algorithm 1’s unbalanced updates and returned objective are consistent with Eq. (2)? In particular, how is the product-reference entropy incorporated into the updates, and does the implementation use the same formulation?
  3. What evidence establishes the accuracy of the final divergence values and gradients at the experimental settings? Can you provide small-scale comparisons against a direct solver of the same objective and clarify how nonpositive convolution outputs are detected and handled? These checks would materially affect my confidence in the reported efficiency and downstream conclusions.
1 comment

We thank the reviewer for the very precise and constructive comments.

Novelty (Q1).

We thank the reviewer for highlighting this reference, which we had clearly missed in our literature review. We will cite and discuss it in Sections 1 and 3. We agree that §6.3.2 and Theorem 6.2 contain the computational construction: band-limited heat kernels on S2\mathbb S^2, forward and inverse spherical harmonic transforms, O(N3/2)\mathcal O(N^{3/2}) per iteration and O(Nlog⁡2N)\mathcal O(N\log^2N) on equiangular grids. We will credit it to Berman and no longer present it or its complexity as new. Yet, we believe our contributions go beyond Berman's paragraph and delineation of the idea. What remains of our contribution beyond it is:

  • Theory. Theorem 6.2 of (Berman, 2020) concerns balanced Sinkhorn, its potentials and the transport cost. Our new results are the unbalanced case of Theorem 3.1 (τ<∞\tau<\infty), where plans have free marginals and mass (mass bound and block approximation proven in Lemmas A.1–A.2), and Prop. 3.2: on S2\mathbb S^2 the debiased divergence S^ετ\hat S^\tau_\varepsilon is positive definite, convex and metrizes weak convergence (both in the balanced and unbalanced case) which justifies its use as a metric and its gradients.
  • Implementation. Theorem 6.2 is a complexity result on quasi-Monte Carlo point sets and to our reading no implementation is reported, nor any practical implementation details are discussed. SHOT brings a differentiable GPU implementation for Clenshaw–Curtis and HEALPix grids.
  • Application. We discuss and evaluate such an approach and the unbalanced variants on the evaluation of climate models, which is fully new and original.

Implementation and Validation.

Algorithm 1 vs Eq. (2) (Q2).

You are completely right and we thank the reviewer for this question. As explained in the general answer, our updates used the uniform measure as reference in the entropy instead of the product measure P⊗QP\otimes Q of Eq. (2), and the implementation followed Algorithm 1. Since KL(π ∣ σ⊗σ)=KL(π ∣ P⊗Q)+∫log⁡p dπ1+∫log⁡q dπ2+cst,\mathrm{KL}(\pi\,|\,\sigma\otimes\sigma)=\mathrm{KL}(\pi\,|\,P\otimes Q)+\textstyle\int\log p\,d\pi^1+\int\log q\,d\pi^2+\mathrm{cst}, both problems have the same optimal plan when the marginals are fixed, and the extra terms cancel in the Sinkhorn divergence: balanced results are unchanged. In the unbalanced case the marginals are free and the problems differ. The corrected algorithms (Algorithm 1 in Section 2 and Algorithm 2 in Appendix C) and our implementation solve Eq. (2). With κ=τ/(τ+ε)\kappa=\tau/(\tau+\varepsilon), the originally submitted update u←(p⊘HeatConv(v))κ\mathbf u\leftarrow\big(\mathbf p\oslash\mathrm{HeatConv}(\mathbf v)\big)^{\kappa} becomes u←(1⊘HeatConv(q⊙v))κ,v←(1⊘HeatConv(p⊙u))κ:\mathbf u\leftarrow\big(\mathbf 1\oslash\mathrm{HeatConv}(\mathbf q\odot\mathbf v)\big)^{\kappa},\qquad\mathbf v\leftarrow\big(\mathbf 1\oslash\mathrm{HeatConv}(\mathbf p\odot\mathbf u)\big)^{\kappa}: the product-reference entropy enters through q\mathbf q (resp. p\mathbf p) inside the convolution, instead of p\mathbf p (resp. q\mathbf q) in the numerator. The returned value is the full dual of Eq. (2), ∫τ(1−e−φ/τ) dP+∫τ(1−e−ψ/τ) dQ−ε (m(π)−m(P) m(Q)),\textstyle\int\tau\big(1-e^{-\varphi/\tau}\big)\,dP+\int\tau\big(1-e^{-\psi/\tau}\big)\,dQ-\varepsilon\,\big(m(\pi)-m(P)\,m(Q)\big), whose last term, which was omitted in the originally submitted Algorithm 1, depends on the mass of the plan and vanishes only when τ=∞\tau=\infty.

Negative convolutions (Q3).

The exact heat operator is strictly positive but as pointed out in the paper its degree-LL truncation which is evaluated in floating point can return values ≤0\le 0 when the output is very small. We do not clamp: as noted in App. C.4 clamping can bias the potentials without preventing failure of the Sinkhorn iterations. In the log-domain update a value ≤0\le 0 makes the potential ±∞\pm\infty or NaN. After every iteration our algorithm checks that both potentials are finite and if it is not the case it stops and returns the run as failed.

Small scale comparison (Q3).

We added this comparison in Appendix D.3 of the revised version (Table 2): we compared SHOT against a dense log-domain Sinkhorn solver of Eq. (2) with the Gibbs kernel e−d2/εe^{-d^2/\varepsilon} on the same grid, on 10 pairs of random measures (Clenshaw–Curtis L=96L=96 and HEALPix Nside=32N_{\rm side}=32), balanced and unbalanced (τ=100\tau=100), for ε∈{0.01,0.1,1,10}\varepsilon\in\{0.01,0.1,1,10\}. For ε≤0.1\varepsilon\le0.1, which covers the settings of our experiments, the two solvers agree closely: the relative differences of the divergence values and of their gradients are below 1%. The difference grows for larger ε\varepsilon because the heat kernel Hε/4\mathcal H_{\varepsilon/4} moves away from the Gibbs kernel, not because of the solver. Table 2 also reports the gap of both divergences to the exact (unregularized) W22W_2^2: it is the same for SHOT and for the dense solver, and the difference can mainly be explained by effect of regularization on the transport problem.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comment to the attention of all reviewers

We thank the reviewers for their careful reading of the paper and for pointing out missing technical details or different aspects of the evaluation of the method that we overlooked.

Thanks to the question of reviewer TSgm on Algorithm 1’s unbalanced scaling updates, we acknowledge an inconsistency regarding the product measure reference in the KL term of Eq. (2). In our implementation, we used the uniform measure as a reference (as can be found in most balanced entropy-regularized OT implementations), whereas a correct implementation for the unbalanced case should use the product measure (as described in [Chizat et al., 2018; Séjourné et al., 2019]). We have now corrected this point in our implementation and in the description of the algorithm given in the appendix, and it did not change significantly the performances reported in the paper. Especially, note that this change doesn't impact the running time experiment given in section 4.1. Together with remarks of reviewers ft5K and iVtB that asked for a more quantitative evaluation of the method with respect to other baseline scoring methods, we decided, along with the corrected implementation and results, to rework parts of the experimental Section on climate models and focus more deeply on quantitative explanations of the produced rankings. As a byproduct and to maintain the page limit, we removed the paragraph on sea-ice parameterisation and convection schemes. All figures and tables in Sec. 4.2 have been regenerated, and we added the baselines and analyses described below in the direct responses to reviewers.

The new uploaded version of the paper carries those modifications (highlighted in red), and we hope that in its current form the paper addresses all the reviewers' reserves about our method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.