acceptodds
Under review as a conference paper at ICLR 2027

Mind the Curvature: Correcting Structural Bias in Bures–Wasserstein Regression

Abstract

Tangent-space linearization makes scalar-on-distribution regression tractable for multivariate Gaussian inputs, but its predictions depend on the reference distribution. In Bures–Wasserstein space, nonzero curvature can turn reference mismatch into a nonlinear projection error that persists even when the input distributions are known exactly. We characterize the part of this error that remains after refitting the linear predictor and use it to motivate learning the reference from prediction residuals. A local risk expansion at general positive definite references quantifies the joint effect of reference displacement and data spread, with a signal-dependent commutator coefficient at the identity. The result distinguishes positive geometric bias from zero leading coefficients and exact predictive equivalence. An alternating procedure fits the regression and updates the reference covariance, while an empirical-risk bound links reference selection to the population target. Exact-covariance experiments verify the predicted coefficients and scaling quantitatively. Repeated prediction experiments assess validation-based selection and separate geometric from estimation error. Wearable, molecular, and sensing applications show task-dependent prediction gains with the same linear readout.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Reference misspecification, Riemannian curvature, Tangent space projection

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_7GyuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how reference choice affects linear regression on Gaussian distributions in Bures–Wasserstein geometry. It characterizes the prediction error that remains after refitting the regression coefficients and proposes learning the reference covariance by minimizing prediction residuals. Synthetic experiments verify the theoretical scaling, and 5 real-data applications evaluate improvements over fixed-reference regression and other linear baselines.

Critical strengths

A precise explanation of reference-induced error. The theory distinguishes changes that regression coefficients can absorb from genuine approximation error. It quantifies the latter and identifies cases where it vanishes, providing insight beyond the general observation that the geometry is curved.

Critical weaknesses
  1. The practical relevance of the main assumption remains unclear. The theory assumes that the response is linear at some suitable reference, and the synthetic experiments construct this scenario. The paper provides limited guidance for recognizing when this assumption is reasonable or assessing robustness when no such reference exists.

  2. The real-data evidence leaves uncertainty about the practical gains. Results use single outer splits without uncertainty intervals. Regularization is evaluated for raw covariance features but not fixed-reference tangent features, leaving open whether reference learning improves over a suitably regularized version of the same representation. All 600 runs in the additional simulations stopped at the iteration or backtracking limit. Testing sensitivity to initialization and optimization budgets would help establish whether the reported improvements are reproducible across a range of practical settings.

Questions for authors
  1. How can practitioners assess whether reference mismatch is a substantial source of prediction error? What evidence suggests that reference learning remains useful when no reference makes the response exactly linear?

  2. Given the reported stopping conditions, what evidence shows that the prediction gains are stable across different initial references and larger optimization budgets?

3 comments

Thank you for asking how a practitioner can use the theory when the generating reference is unknown, and whether simpler choices of regularization or stopping could explain the application gains.

1. Recognizing when reference adaptation helps. The tangent-linearity assumption gives a geometric interpretation to the excess risk. A practitioner can assess the predictive value of adaptation without knowing whether that assumption holds. Section 5 describes a paired comparison: fit both references on the same training data, select the update schedule within training, and compare squared errors on the same held-out observations. With the Gaussian inputs and the OLS readout held fixed, the paired loss difference measures the predictive value of learning the reference on that task. This gives practitioners a way to decide whether adaptation is worthwhile even when the generating reference is unknown (BioHL and KOA, Appendix G.7).

To examine misspecification directly, Appendix H.5 adds a quadratic term in the Gaussian mean with standard deviation one quarter that of the original noiseless signal. At fixed covariance every tangent-linear predictor is affine in the mean, so no Gaussian reference can represent this term exactly. Learning the reference nevertheless reduced mean test MSE in two independent sets of simulations. The quadratic dependence remains outside the model throughout this experiment. The gain shows that a better choice of coordinates can still improve the overall linear approximation without requiring an exactly tangent-linear response.

2. Practical gains, regularization and stopping. Your suggested ridge comparison isolates an important alternative explanation. We fitted ridge at the training BW barycenter, leaving the tangent features and ordinary covariates unchanged and selecting its penalty within training. Learned-reference OLS gave lower test MSE on all five application splits (Appendix G.4). Table 4 gives paired uncertainty intervals for the two small molecular test sets; on KOA the single-split interval crosses zero, while repeated grouped partitions favor the learned reference on average (Appendix G.7). Ridge changes the weight assigned to each tangent coordinate at the barycenter. Learning the reference changes the coordinates themselves, so a linear combination can express a different function of the same Gaussian input. The five matched comparisons show that this flexibility had predictive value even when the fixed-reference model could tune its regularization.

The stopping flags you identified give no certificate of stationarity. We therefore tested the prediction consequence of the stopping rule directly. Across the exact-covariance conditions in Appendix H.6, validation-selected updates reduced pooled test MSE from both the perturbed start and the training barycenter at every budget examined. Extending the backtracking allowance for the perturbed-start runs that hit the limit had negligible effect on their selected test MSE. Validation could retain the starting point, and taking the last iterate at the largest budget gave higher pooled error than taking the last iterate at the smallest budget. Including zero updates makes the fixed BW barycenter an available choice whenever fitting starts there. A practitioner can therefore use validation to choose both whether to adapt and when to stop, while retaining an ordinary linear predictor for new distributions.

Thank you for your response, the additional experiments, and for addressing my questions. However, the response does not address the critical weakness concerning the practical relevance of the main assumptions and the scenarios in which they hold. Therefore, I will keep my score unchanged.

Thank you for asking when reference learning helps and how to assess its gains. Our previous reply reported the nonlinear-response experiment (Appendix H.5), the fixed-reference ridge comparison (G.4), molecular uncertainty and KOA refits (Table 4 and G.7), and initialization and budget sensitivity (H.6). This revision adds Propositions 2 and 3, referenced in Sections 3.3 and 3.5.

Weakness 1. Practical relevance of tangent linearity

1(a). When is the approximation reasonable? Proposition 2 (Appendix C.5) uses a first-order Taylor predictor in the tangent coordinates of a candidate reference. If the conditional mean has Hessian operator norm at most LL along the coordinate segments from the reference to the inputs, and the tangent coordinate vv has a finite fourth moment, the squared approximation error is at most (L2/4)E∥v∥4(L^2/4)\mathbb E\Vert v\Vert^4. The bound is small for a smooth response and inputs concentrated near the reference.

1(b). What happens when no reference is exactly linear? Write the conditional mean as g+ug+u, with gg tangent-linear at a candidate reference and ∥u∥L2≤δ\Vert u\Vert_{L^2}\leq\delta. Proposition 2 shows that the square root of the profiled excess risk differs from ∥(I−Πρ)g∥L2\Vert(I-\Pi_\rho)g\Vert_{L^2} by at most δ\delta, where Πρ\Pi_\rho projects onto the affine tangent feature space at ρ\rho. A correction that reduces this residual norm by more than 2δ2\delta strictly lowers the population risk after refitting.

In the nonlinear-response experiment (Appendix H.5), the centered quadratic term in the Gaussian mean is orthogonal to every affine tangent feature space for the latent inputs. It creates a positive approximation error at every reference and adds the same amount to each reference's profiled risk. In two sets of 20 paired replicates, learning the reference reduced mean test MSE by 16.7% and 18.5%. The 95% Monte Carlo intervals for fixed minus learned MSE were [0.132, 0.276] and [0.077, 0.212].

1(c). How can we decide whether to apply the correction? Our earlier reply proposed comparing fixed and learned references on the same held-out observations. Proposition 3 (Appendix H.6) considers independent i.i.d. validation of training-fitted checkpoints, including the zero-update fit. Under a conditional variance bound on the paired loss differences, with probability at least 1−α1-\alpha given training, the selected predictor's risk is within 2rv2r_v of the lowest checkpoint risk, where rvr_v is the validation error radius. On the same event, an observed paired validation MSE reduction greater than rvr_v implies lower risk than the zero-update fit.

Weakness 2. Evidence for practical gains

2(a). Split uncertainty and refitting. The 95% paired structure-group bootstrap intervals (Table 4 and Appendix G.7) compare fixed-reference OLS with learned-reference OLS, conditional on the fitted models and conformers. The MSE reductions were 0.0643 [0.0100, 0.1237] on BioHL and 0.0889 [−0.0541, 0.2507] on KOA. The KOA interval includes zero. To assess sensitivity to refitting, we also ran ten repetitions of grouped five-fold outer cross-validation on the same KOA molecules, with four-fold grouped selection within each outer training set. Mean out-of-fold MSE fell from 0.9510 to 0.8787 (7.61%), with a reduction in every repetition.

2(b). Ridge on fixed tangent features. The fixed-reference baseline fits ridge to the BW tangent features at the training barycenter with the same ordinary covariates and an unpenalized intercept. We scaled the features on training data and chose the penalty from a 15-value grid by leave-one-out error (Appendix G.4 and Section 4.4). Learned-reference OLS lowered test MSE relative to this baseline by 9.88%, 20.47%, 42.01%, 10.67%, and 18.32% on the NHANES, ESOL, BioHL, KOA, and Gas splits, respectively.

2(c). Stopping, initialization, and update budgets. The original runs ended at the iteration cap in 574 of 600 cases and at the backtracking limit in 26 (Appendix H.2). The sensitivity analysis uses matched samples for two initial references across 12 simulation settings with 50 replicates each (Appendix H.6). At caps of 60, 120, and 240 updates, validation-selected pooled test MSE fell by 4.29%, 5.02%, and 5.45% relative to the frozen perturbed reference and by 2.04%, 2.37%, and 2.56% relative to the frozen training barycenter. At 240 updates, six of 12 condition means improved from the barycenter and validation selected zero updates in 314 of 600 runs. Using the final iterate increased pooled test MSE at both starts as the cap rose from 60 to 240. Doubling the backtracking allowance from 40 to 80 for the 27 affected runs in the perturbed-start comparison changed the selected test MSE by at most 2.05×10−82.05\times10^{-8}.

Reviewer_U2DwOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies how the choice of reference affects scalar regression on Gaussian distributional covariates. Assuming that the conditional mean is linear in tangent features at a generating reference, it characterizes the prediction error that remains at another reference after refitting the intercept and coefficients. The main result gives a local leading term t2r4CSt^2r^4C_S, where tt controls reference displacement and rr controls data spread. The coefficient depends on the response direction and the covariate distribution. The authors use this analysis to motivate learning the reference covariance through profiled least squares, and evaluate the method through exact-covariance calculations, repeated simulations, and five applications.

Critical strengths
  1. The most useful result is the calculation of prediction bias after the regression coefficients have been refitted. Theorem 2 and Corollary 1 show why noncommuting directions alone do not guarantee a prediction penalty: the effect also depends on the signal and on what the affine predictor can absorb. The general-reference expansion and the distinction between a zero leading coefficient and exact equivalence make this a useful account of when reference choice matters. Its contribution beyond earlier supervised reference learning, such as Mialon et al. (2021), is this explicit characterization within the Gaussian tangent model.

  2. Table 1 provides a direct check of the theory by comparing predicted coefficients with numerical risks using exact covariances. The commuting and signal-direction controls in Appendix H.1 strengthen this evidence. Remark 1 also makes Table 2 interpretable as a comparison of identity, barycenter, and learned references: the barycenter improves on the identity in four of five tasks, and the learned reference improves on both in all five. This supports the practical value of reference choice, although these application comparisons do not isolate the contribution of the local curvature term.

Critical weaknesses
  1. The magnitude and consistency of the application gains need a more careful assessment. Table 2 reports one outer split per task without uncertainty estimates; BioHL and KOA have only 30 and 54 test molecules, respectively (Table 3). The paper itself shows sensitivity in BioHL: validation MSE reductions against matched fixed BW range from 13.03% to 56.92%, with the largest reductions occurring in conditions where fixed BW has the highest error (Table 8C). All four conditions improve on fixed BW, but under seed 10023 the optimized model has validation RMSE 0.5005 versus 0.4888 for Mean + SD OLS (G.4). Thus, the advantage over simpler representations is not consistent across these conditions. Table 8A also reports validation/test MSE reductions of 3.58%/10.49% for KOA and 6.00%/18.24% for Gas. These differences do not by themselves establish instability, but they make paired uncertainty estimates important for interpreting the size of the practical benefit.

  2. Figure 2 cannot be reconstructed from the manuscript's description. Section 4.2 and Appendix E.2 specify the reference mismatch, sample sizes, and replication protocol, but do not fully specify the covariate generator, response coefficients, or noise distribution and level. Appendix E.1 says that the renderer restores this figure from a stored PDF asset. This leaves a reproducibility gap for the experiment that includes estimated Gaussian inputs. The separately specified exact-feature simulations in Appendix H.2 do not resolve that gap.

Questions for authors
  1. Could you use the paired predictions described in G.5 to report confidence intervals for the BioHL and KOA test MSE differences, accounting for any shared molecular groups? These would help assess the uncertainty in the reported comparisons.

  2. NHANES and Gas end near the numerical SPD floor (E.3, G.3). What are the learned covariances' generalized eigenvalues relative to their training barycenters? At these references, how closely do the iterative features used during optimization agree with the spectral features used for prediction? These diagnostics would clarify the geometric and numerical regime of the application results.

  3. Could you provide the complete data-generating model for Figure 2, including the distribution of Gaussian means and covariances, the response coefficients, and the noise model? Is code available to regenerate the plotted results from that model?

3 comments

Thank you for your careful reading of the risk expansion and for identifying where the application results and Figure 2 needed stronger evidence.

1. How consistent are the application gains? Your observation about the BioHL conformers clarifies what the comparison can establish. Learning the reference improved validation MSE over its matched fixed BW reference in all four conformer conditions; the ranking across different representations varied (Appendix G.4). To quantify uncertainty on the original molecular test sets, we resampled molecular structure groups jointly across methods. The paired interval for the test MSE reduction over fixed BW lies above zero on BioHL and includes zero on KOA (Table 4). On the same KOA molecules, repeated grouped cross-validation gave lower out-of-fold MSE in every round, by 7.61% on average (Appendix G.7). Because the two fits use the same molecular inputs, feature dimension and OLS readout, this paired comparison directly assesses the value of learning the reference. On KOA, that benefit recurs as different molecules are held out.

2. Reconstructing Figure 2. You identified a gap specific to the estimated-input experiment. Appendix E.2 now gives its Gaussian input generator, response coefficients and noise scaling, within-distribution sampling and covariance estimator, split, initial reference, update budgets and error bars. We regenerated all 24 conditions from that model with 20 replicates per condition. The supplementary code README gives the instructions for reproducing the figure from simulated observations. The design varies the number of independent distributions separately from the number of observations within each one. With moderate or large reference mismatch and at least 100 distributions, learning the reference narrowed the test-error gap to the generating-reference oracle, even though all fits used the same estimated Gaussian inputs.

3. Sensitivity to the positive-definiteness constraint. At the default floor, the learned covariance contracted sharply along one direction on each task. With the OLS coefficients held fixed, iterative and spectral feature calculations gave nearly identical test predictions (Appendix E.3). After raising the eigenvalue floor from 10−610^{-6} to 10−310^{-3} and refitting, learned BW retained test MSE reductions of 9.08% on NHANES and 17.01% on Gas relative to fixed BW. These checks show that the reference can adapt selectively across covariance directions and improve prediction on both tasks even when the most extreme contraction is excluded.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer U2Dw,

As the discussion period is drawing to a close, please let us know whether our responses and additional experiments have resolved your concerns and helped you reconsider your assessment of our work.

Thank you again for your time and thoughtful consideration.

Thank you for the clarifications and additional analyses. These substantially address my concerns. I maintain my positive score, while noting the remaining uncertainty in KOA’s original test-set improvement.

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

This paper studies scalar-on-distribution in Bures-Wasserstein space, where nontrivial curvature can lead to irreducible error within a Gaussian tangent-feature family with linear readout. The authors theoretically characterize the geometric structural bias through the profiled prediction risk and quantify how mismatch with reference distribution causes the error through a local risk expansion. Subsequently, they proposed an algorithm that repeatedly fits OLS at the current reference and corrects the reference covariance in the suitable direction. Experimental results on synthetic and real-data tasks confirm the theoretical predictions and lower test error after optimizing the reference.

Critical strengths
  • The problem is well motivated: scalar-on-distribution that associates covariates with empirical distribution is a highly active and relatively underexplored area. This paper aims to investigate how curved BW geometry beyond one dimension can become problematic and create additional errors. The extension to higher dimensions is always important to further our understanding,

  • Clear empirical results: the experiments sections are comprehensively reported; tables and plots confirm the theoretical validity and improvement from the algorithm cleanly.

Critical weaknesses
  • Theoretical novelty is limited: The main theoretical argument in this paper is a natural generalization from classical wisdom in linear algebra. In particular, Theorem 1 mirrors common phenomena about projection: the orthogonal component causes certain loss of information and is not recoverable by linear refitting, as the feature is no longer observed anyways. The derivation mainly relies on the classical identity:

    bρ=Πρbρ+(I−Πρ)bρb_\rho = \Pi_\rho b_\rho+ (I-\Pi_\rho)b_\rho

This is properly modified and rephrased in the BW setting, where nontrivial curvature affects the projection. Theorem 2 is more technically sound and well motivated, but its central message is still a toy example where linear refit cannot absorb the residual.

  • Algorithm 1 is computationally heavy: Given the first weakness, it would not be surprising to see that in synthetic experiments, reference mismatch increases test error, and reference correction using residual information is a natural strategy for optimization. My main concern is that Algorithm 1 might be extremely slow for higher dimensions; by the end of Section 3, it requires taking costly derivative and matrix roots at O(nd3)O(nd^3) and refitting OLS at O(np2)O(np^2). However, as p=d+d(d+1)/2=O(d2)p = d + d(d+1)/2 = O(d^2) by Equation 2; this algorithm can grow quartically in the dimension dd.

Currently, all of the experiments only include the low-dimensional regime (d=4−6d = 4-6), but scalability of this algorithm is unclear and can be very limited. Also certain datasets (like BioHL and KOA) have small test sets (30 & 54), and the results can be subject to statistical uncertainty.

Since the reference correction algorithm appears to depend heavily on a low dimension, and the theoretical analysis does not provide a new perspective, I am leaning towards rejection for now.

Questions for authors
  • In the introduction, the authors mentioned distribution representations like kernel mean embeddings and sliced Wasserstein features. Would it be possible to incorporate them in the experiments as well?
  • Consider some rewriting of the theory section (2, 3); for instance, I would suggest adding some definitions to avoid defining things in chunks and aggregating certain results in 2 and 3,5 into new theorems.
  • If time permits, can you include experiments where the data dimension is moderate (d = 20-50)? It would be useful to see how this algorithm’s compute scales with data dimension and how resistant it is to the curse of dimensionality?
5 comments

Thank you for distinguishing the classical projection argument from the geometric claim. Your review raises two questions that bear directly on the contribution: what BW geometry tells us after the regression has been refitted, and what it costs to use that information.

1. What the projection identity leaves open. We agree that the orthogonal decomposition in Theorem 1 is classical. We disagree, however, with the interpretation that the residual arises because a feature is no longer observed. At either reference the full transport coordinates recover both the mean and the covariance of the input Gaussian (Appendix G.6). In Euclidean regression, moving the origin translates the predictors, which an intercept absorbs exactly. In BW regression, changing the reference can induce a nonlinear change of coordinates. The input is still fully observed, yet a response that is linear in the first coordinates can have positive approximation error in the second. Figure 1 illustrates the underlying change of coordinates. To make the distinction explicit, Section 2 now defines the prediction space and its projection together after Eq. (2).

Theorem 2 quantifies this effect at a general positive definite generating covariance. Under its local model, the BW transport equation gives a t2r4t^2r^4 term in excess risk after optimal refitting, its coefficient, and a uniform remainder. The two-dimensional example evaluates the general coefficient in closed form. Holding the inputs and the reference displacement fixed while rotating the response changes the coefficient from 16/4516/45 to zero (Appendix H.1). The projection identity identifies the residual; the geometric calculation predicts when a leading penalty appears and how large it is.

2. Computational cost and empirical scope. Your quartic scaling calculation is correct. With p=d+d(d+1)/2p=d+d(d+1)/2 and n≥pn\ge p, a dense OLS refit costs O(np2)=O(nd4)O(np^2)=O(nd^4), in addition to the matrix roots. Appendix H.4, Table 14, measures the accuracy–runtime trade-off at a fixed update budget and gives ECoG timings. In the signal-aligned, large-sample simulation, reference updates lowered test error at each dimension examined. For the small molecular test sets you identified, Table 4 also gives paired uncertainty intervals, and Appendix G.7 evaluates KOA across repeated grouped partitions. The reference search and its repeated OLS fits are paid for once during training. For each new Gaussian, prediction requires a single transport map to evaluate the fitted linear model, with cost independent of both the training sample size and the number of reference updates.

3. Alternative distribution representations. Table 9 compares KME, SW and BW with matched feature dimensions and the same OLS readout. Learned BW had the lowest test MSE on ESOL and BioHL, SW on NHANES, and KME on KOA and Gas. BW gives OLS direct access to the joint affine transport from the reference to each input, so a response linear in that deformation is exactly representable at the appropriate reference. Holding the Gaussian inputs and OLS model fixed, learning the reference lowered test MSE relative to the training barycenter on all five application splits (Table 2). On ESOL, for example, all three representations use 14 regression coordinates: test MSE is 1.476 for learned BW, 1.744 for KME and 1.877 for SW. The BW coordinates describe translation and symmetric deformation directly, whereas the matched KME and SW fits use principal components of Fourier features and projected moments, respectively.

Dear Reviewer bRgQ,

As the discussion period is drawing to a close, please let us know whether our responses and additional experiments have resolved your concerns and helped you reconsider your assessment of our work.

Thank you again for your time and thoughtful consideration.

Thank you again for your constructive review. Building on our earlier response, we explain how the BW prediction-risk analysis, supervised reference learning and empirical evidence address your concerns and questions.

Weakness 1: Theoretical contribution and the two-dimensional example.

Our contribution within the transport-representation literature [1,2] is a quantitative characterization of BW prediction error after optimal intercept and coefficient refitting. At either reference, the full Gaussian transport coordinates recover the input mean and covariance (G.6). Reference change can induce a nonlinear coordinate transformation and alter the functions expressible by a linear predictor. Thus the analysis quantifies the task-dependent risk arising from a reference-dependent prediction class with the distributional information retained.

Theorem 1 connects this prediction class to profiled risk. Theorem 2 supplies the BW-specific calculation in general dimension at a general positive definite generating covariance: a local t2r4t^2r^4 excess-risk term, its explicit signal-dependent coefficient and a uniform remainder. Corollary 1 gives the commutator expression at the identity. The two-dimensional example evaluates the general coefficient in closed form; Table 1 and H.1 verify the coefficients and joint scaling. This is the quantitative link between BW geometry and prediction after refitting that motivates learning the reference from the response.

Weakness 2 / Question 3: Computational cost and moderate dimensions.

For p=d+d(d+1)/2p=d+d(d+1)/2 and n≥pn\ge p, dense OLS costs O(np2)=O(nd4)O(np^2)=O(nd^4), alongside O(nd3)O(nd^3) matrix-root/derivative work. H.4, Table 14, directly measures the accuracy–runtime trade-off. With 10,000 training distributions, three updates at d=20,35,40,50d=20,35,40,50 reduce mean test MSE by 43.47%, 48.73%, 49.53% and 38.68% in five paired signal-aligned simulations. Corresponding median fitting-plus-prediction times are 26.37, 45.86, 48.87 and 84.26 seconds over three serial Apple M4 runs after input preparation. The 48-dimensional ECoG runtime study reports 18.85 seconds for three-update fitting and prediction.

Reference search is paid during training. Each new Gaussian requires one transport map and the fitted linear readout, with prediction cost independent of the training sample size and number of reference updates. This retains an explicit linear predictor while adapting its coordinates to the task.

Weakness 2: Uncertainty in the molecular results.

Table 4 reports paired 95% percentile intervals for test-MSE differences on BioHL and KOA by jointly resampling molecular structure groups across methods (10,000 draws; 30 and 54 test groups). G.7 further evaluates split variation through ten repetitions of grouped five-fold outer CV with four-fold inner schedule selection on the same 270 KOA molecules. Mean out-of-fold MSE falls from 0.9510 to 0.8787, a 7.61% reduction relative to fixed BW OLS, with improvement in all ten repetition means. These analyses complement the original test splits with paired uncertainty and repeated evaluation of the reference-selection procedure.

Question 1: KME and sliced Wasserstein comparisons.

G.6, Table 9, compares KME, SW and BW using matched feature dimensions and the same OLS readout. Learned BW achieves the lowest test MSE on ESOL and BioHL. On ESOL, all three use 14 distribution coordinates: MSE is 1.476 for learned BW, 1.744 for KME and 1.877 for SW. BW features describe translation and joint symmetric deformation directly. Holding the Gaussian inputs and OLS readout fixed, reference learning lowers test MSE on all five application splits by 9.91%–46.39% relative to fixed BW (Table 2), isolating the contribution of reference adaptation.

Question 2: Theory presentation.

As reported on earlier, Section 2 now defines the prediction space and its projection together after Eq. (2). Sections 3.3–3.4 present the profiled-risk identity and general BW expansion as theorems, followed by reference optimization in Section 3.5. This makes the definitions, statistical characterization, geometric calculation and resulting method explicit in sequence.

Thank you again for your time and constructive feedback. Please feel free to raise any further questions here; we would be happy to clarify. If our responses and supporting evidence address your concerns, we would be grateful if you would consider revising your score upward.

References

[1] Wang, W., et al. (2013). A Linear Optimal Transportation Framework for Quantifying and Visualizing Variations in Sets of Images. IJCV, 101(2), 254–269. DOI.

[2] Mialon, G., et al. (2021). A Trainable Optimal Transport Embedding for Feature Aggregation and its Relationship to Attention. ICLR. Paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.