acceptodds
Under review as a conference paper at ICLR 2027

Universal Diffeomorphic Representation Learning on SPD Manifolds via the Spline-Pullback Metric

Abstract

Deep learning on Symmetric Positive Definite (SPD) manifolds has traditionally depended on predefined algebraic Riemannian metrics. Much like hand-engineered features in early machine learning, these fixed geometric structures constrain the expressive capacity and adaptability of the metrics. While recent works have attempted to learn parameterized geometries, they frequently breach foundational matrix function axioms via rank-dependent scaling or unbounded power operations. These violations result in spatial folding, a lack of global surjectivity, and gradient vanishing near spectral singularities. To overcome these limitations, we present the Spline-Pullback Metric (SPM), comprising both Spectral-SPM and Cholesky-SPM variants, marking a paradigm shift in SPD manifold learning from static metric choices to universal approximation. By modeling the global diffeomorphism with a monotonically constrained, rank-invariant B-spline, SPM functions as a dense universal approximator for strictly increasing diffeomorphisms. Our formulation theoretically generalizes existing pullback metrics while facilitating localized, non-linear modeling of the spectral domain. SPM guarantees a globally bijective pullback geometry, preventing gradient instabilities and rank-swapping discontinuities. We empirically demonstrate that SPM achieves superior performance results across five benchmark datasets utilizing Linear Probes, SPDNets, and deep Riemannian ResNet architectures.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Geometric Deep Learning, SPD Manifolds, Representation Learning, Riemannian Geometry, B-spline, Diffeomorphism, Stability

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_i6XeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes learnable extensions of the log-Euclidean and log-Cholesky metrics for SPD matrices. It replaces the fixed logarithmic map with an increasing spline transformation in log coordinates. Two variants termed Spectral-SPM and Cholesky-SPM are presented, where the transformation acts on eigenvalues and on positive Cholesky diagonal entries, respectively. The splines’ coefficients are learned jointly with the classification model. The paper presents theoretical results on invertibility, approximation capacity, and the induced pullback geometry. Experiments compare several metrics across five datasets and report runtime and memory use. Spectral-SPM demonstrates improved classification compared to several baselines, while Cholesky-SPM achieves only mixed results and generally lower computational cost.

Critical strengths
  • The idea of extending log-Euclidean and log-Cholesky metrics with learnable nonlinear transformations is new and useful.

  • The presented monotone spline constraints preserve invertibility while allowing the flexibility to expand or compress different input ranges.

  • The paper addresses approximation of both the scalar function and its first derivative. This is relevant to learning the metric and computing gradients.

  • The construction supports both spectral and Cholesky representations within one framework.

  • Experiments compare both commonly used Riemannian metrics and learnable metrics, where the proposed Spectral-SPM shows consistently superior results.

  • The paper reports runtime, memory use, and sensitivity to spline settings.

Critical weaknesses
  • Several empirical gains over the leading baseline are small. In addition, the Cholesky-SPM sometimes performs worse than log-Cholesky.

  • A simple learned affine map in log coordinates is missing as a baseline. This would help establish whether flexible nonlinear splines are necessary.

  • Theorem 1 allows degree-1 splines k≥1k \geq 1, which need not be C1C^1

Questions for authors
  • Do the empirical gains remain consistent across training seeds and compared with a learned affine map in log coordinates?

  • If I understand correctly, the second derivative can jump where the spline meets the linear tails. How do the claims about global curvature and parallel transport hold at these joins?

  • The derivative fθ′(x)=m/xf_\theta'(x)=m/x grows without bound near zero. How does the gradient stability guarantee hold in this case?

  • The links to the proofs in the theoretical results statements are unconventional. It would be clearer to refer to Appendix B for the proofs in the main text.

3 comments
AuthorsOpenReview ↗

Response to Reviewer i6Xe -1

We thank the reviewer for the positive assessment of the diffeomorphic construction, the approximation result, and the empirical comparisons. We appreciate the additional questions concerning affine baselines, regularity, and numerical stability.

W1: Empirical Gains

We agree that the improvement magnitude varies across datasets and architectures, and that C-SPM is not uniformly better than LCM. The two variants use different representations and inductive biases: S-SPM acts on the eigenvalue spectrum, whereas C-SPM acts on the positive diagonal entries of the Cholesky factor while retaining the lower-triangular components. Their relative performance therefore depends on the structure of the dataset. We will clarify that S-SPM provides the stronger empirical results in several settings, while C-SPM offers a computationally efficient alternative rather than claiming uniform superiority of both variants. Moreover, our proposed metric outperforms the baselines with statistical significance in the large scale AI-NLP and AFEW dataset results presented in the Appendix.

W2: A learned affine transformation in log coordinates

We agree that a learned affine map in log coordinates is a useful low-capacity reference. However, the relevant comparison must preserve the geometric requirements of the proposed metric. Related learnable metric formulations such as ALEM already explore restricted parameterized deformations. The central contribution of SPM is the ability to learn a more flexible nonlinear diffeomorphism while retaining the global bijectivity and differentiability required by the pullback construction. This mitigates the need for manual task-dependent metric selection, allowing the model to learn the required pullback Euclidean mapping.

The current submission includes a low-resolution spline ablation and a PLM comparison in the appendix. The piecewise linear map faces loss spikes at its C0C^0 kinks and diverges while SPM converges. However, adding your suggested affine baseline would strengthen the paper. We will add the learned affine comparison in the camera-ready version to quantify the benefit of nonlinear spline deformation over a low-capacity diffeomorphic transformation.

W3 Spline Degree We thank the reviewer for catching this typographic oversight. We shall correct the degree to k>=2 for C1C^1. This doesn't affect our experiments, as we used k=3.

Q1: Do gains remain consistent across seeds and compared with a learned affine map?

The five-fold results on HDM05 and Radar provide evidence across different data partitions, and Appendix G shows consistent learned spline shapes across the five Radar folds. Moreover, Table 6 shows relatively stable SPM performance against spline order and grid size changes. However, we agree that FPHA uses a predefined train/test split and does not currently report a multi-seed standard deviation. We aim to add multi-seed evaluation and the learned affine baseline in the camera-ready version.

Q2 Second Derivative

We thank the reviewer for identifying the distinction between the regularity of the active spline region and that of the finite linear extrapolation tails. This distinction is explicitly addressed in Appendix B.1 of our submission.

In our proof of Theorem 1, we state, "Theoretically, by extending the grid boundaries t0 → −∞ and tm → ∞, the linear tails vanish, and C2C^2 smoothness holds globally on R>0. Higher-order differentiability can easily be obtained by increasing k."

For the cubic B-spline used in our experiments, the mapping is C2C^2 throughout the active knot interval. The linear extrapolation tails are introduced solely as a computational fallback for extreme spectral outliers, matching the spline in value and first derivative and thus preserving global C1C^1 regularity. On any prescribed compact spectral domain, the knot boundaries can be chosen sufficiently wide that the entire domain lies within the C2C^2 spline region.

Furthermore, none of the evaluated inputs in all of our reported experiments required linear-tail extrapolation: the spectral values remained within the prescribed spline domain. Consequently, the finite transition joins were not encountered during the reported experiments, and the evaluated transformations operated entirely within the C2C^2 region, precluding concerns regarding curvature calculation and parallel transport.

We will make this distinction clearer in the manuscript.

AuthorsOpenReview ↗

Response to Reviewer i6Xe -2

Q3: The derivative grows without bound near zero. How does the gradient-stability guarantee hold?

We thank the reviewer for identifying this distinction. Although theoretically the derivative can grow near 0. Since our dataset uses SPD matrices, exact 0 eigenvalues do not appear.

Furthermore, we avert arbitrarily large gradients as x approaches 0 by our numerical safeguards.

Our implementation addresses the numerical issues associated with small and repeated eigenvalues through a two-tier positive diagonal perturbation strategy, described in Appendix H.2. These small perturbations not only prevent multiple eigenvalues from being exactly the same, but also bound how small an eigenvalue can become.

Thus, the eigenvalues used by the numerical spectral transformation are prevented from approaching zero arbitrarily closely. Weyl's eigenvalue perturbation inequality additionally guarantees that the displacement of each eigenvalue is bounded by the operator norm of the input perturbation, 2×10−82\times10^{-8}. The non-similar eigenvalues ensure a non-singular denominator in the Dalecki-Krein off-diagonal elements, preventing gradient instabilities.

Q4: The proof references should point clearly to Appendix B

We will standardize the theorem and proposition proof references in the main text to direct readers to the corresponding subsections of Appendix B.

We thank the reviewer for motivating these clarifications. Kindly let us know if you have any remaining concerns/ suggestions, and we respectfully ask that you consider these clarifications in your final assessment and rating.

I thank the authors for the detailed responses. They address some of my concerns. However, considering the comments raised by the other reviewers, I still consider this paper to be borderline. It has merit and nice ideas, but also many aspects need more work.

Reviewer_gA5rOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Spline Pullback Metric (SPM) for learning Riemannian metrics in deep networks on SPD manifolds. This paper points out that traditional metrics such as LEM, LCM, and AIRM lack adaptive capabilities, while learnable metrics such as ALEM and PCM have limitations in terms of global differential homeomorphism properties. Therefore, this paper transforms metric learning into a learning problem of monotonic scalar differential homeomorphism, and constructs two implementations, S-SPM and C-SPM, based on B-spline. The author conducted experiments on three datasets, HDM05, FPHA, and Radar, using three architectures, Linear Probe, SPD-MLR, and RResNet. The results indicate that SPM achieved the highest classification accuracy with at least one variant in all nine combinations of datasets and architectures.

Critical strengths

1.This paper rephrases metric learning on SPD manifolds as learning a scalar differential homeomorphism, using the Spline function to implement adaptive metrics that vary with the task, and solving the gradient instability and rank jump problems of previously learnable metrics. 2.This paper rigorously proves that the existing method ALEM is poorly defined at repeated feature values due to the non uniqueness of the feature space, and PCM lacks global surjections. In addition, this paper also uses the general approximation theorem (Theorem 2) to illustrate that LEM, LCM, and PCM are all special cases of this class. 3.The experimental design of this paper is very comprehensive and effective. Three standard datasets (HDM05, FPHA, Radar), three architectures (Linear Probe, SPDNetMLR, Deep RResNet), and seven metrics were evaluated under a unified backbone protocol, and extended to two large-scale scenarios: AI-NLPand AFEW.

Critical weaknesses

1.The innovation and shortcomings of this paper lie in the fact that although a learnable metric is constructed using spline functions, its essence is simply to select a universal one-dimensional monotonic function approximator and superimpose it on existing LEM (spectral coordinates) and LCM (Cholesky coordinates), without specifically incorporating spline and SPD manifolds in curvature modeling, intrinsic distance, or geodesic structures. 2.The complexity of the measurement design in this paper is high. As the measurement design in this article is based on LEM and LCM, although ALEM and PCM ensure mathematical properties compared to previous works, the complexity is too high. In the field of Riemannian manifold deep learning, I believe that if the feature structure extraction is good, the design of subsequent metrics should be simplified to reduce the burden of network learning. 3.There is a problem with the expression of "Universal" in the title of this paper. The cross-sectional curvature of the global pullback metric is always zero, SPM can only be used universally within the class of flat scalar pullback metrics, and bending metrics such as AIRM and Bures Wasserstein are not any globally homeomorphic pullback metrics. The "consumers" of Corollary 1 and the abstract "generalizes existing pullback metrics" must be limited. 4.The statistical rigor of the main experiment in this article is insufficient. The author conducted McNemar+FDR on two large-scale experiments, AI-NLP and AFEW; But it was not done on the three core main datasets (Tables 1-3).

Questions for authors

1.Monotonic spline transformations (with strictly monotonic control point increments and log domain parameterization) have been maturely implemented in normalizing flows, such as Neural Spline Flows and B-Spline Flows. SPM is designed based on LCM and LEM metrics, so what is the truly innovative aspect? 2.How many learnable parameters, operator steps, and training burden have been added to the design of SPM? Given sufficient feature structure extraction, what is the approximate additional complexity brought by splines compared to directly using fixed LEM/LCM. 3.Has the author attempted a more simplified metric design, such as fixed logarithmic deformation, few nodes or low order splines, and piecewise monotonic deformation? Such a lightweight design may be more suitable for adaptive metric learning. 4.Is there a scalar f that allows SPM to accurately reproduce bending metrics such as AIRM and Bures Wasserstein when the cross-sectional curvature of the global pullback metric is always zero? 5.Can paired McNemar tests and FDR corrections consistent with AI-NLP/AFEW be performed on the HDM05, FPHA, and Radar core datasets (Tables 1-3)?

2 comments
AuthorsOpenReview ↗

Response to Reviewer gA5r-1

We thank the reviewer for the detailed assessment and for recognizing the theoretical analysis of existing learnable metrics, the broad experimental protocol, and the connection to established SPD metrics. We respond to the concerns below.

W1: The use of monotone splines

We agree that monotone B-splines and their use in invertible transformations are established tools; we do not claim to introduce the spline parameterization itself. We situate our novelty within the Geometric Deep Learning subfield of SPD metric learning and our spline formulation and its use as a learnable Riemannian metric construction on SPD manifolds.

Recent work on SPD metric learning, including Adaptive Log Euclidean Metric (ALEM, 2024) and Power Cholesky Metric (ICLR, 2026), demonstrates that learning pullback or parameterized mappings on SPD manifolds is an active research direction. However, these approaches have distinct limitations in their expressivity, global diffeomorphic properties and behavior at spectral degeneracies, which we demonstrated in the paper. SPM provides a different formulation that subsumes these hand-crafted formulations and instantiates this construction in both spectral and Cholesky coordinates.

The novelty therefore lies in the SPD metric-learning formulation, the diffeomorphism guarantees, the spectral and Cholesky instantiations, the linear tail approximation fallbacks, and the resulting geometric properties, not in the existence of monotone splines as a general function-approximation tool. Hence, we believe in the literature where different task-specific hand-crafted pullback Euclidean metric formulae are being constructed for incremental performance gains on specific tasks, SPM, with its scalar approximation properties, provides a meaningful advance by allowing the model to learn the required mapping formula without the need to retrain the model with multiple separate pullback Euclidean metrics, etc.

We will clarify the related-work discussion to distinguish this contribution from spline-based normalizing flows and other uses of splines.

W2: The measurement design is complex, and subsequent metric learning should be simplified if feature structure extraction is already effective. We would like to clarify the complementary purposes of our two evaluation settings.

The Linear Probe experiments deliberately restrict feature extraction and use a linear classifier after the metric transformation. This design isolates the contribution of the learned metric, minimizing the influence of a powerful downstream feature extractor. It therefore allows us to evaluate whether SPM itself improves the representation of SPD inputs.

In contrast, the Riemannian ResNet experiments evaluate whether the metric construction remains beneficial when combined with a more expressive feature-learning architecture. The observed gains in these experiments indicate that the benefit of SPM is not restricted to the linear-probe setting.

Regarding computational complexity, the default SPM configuration used across all experiments uniformly in the manuscript introduces only 13 trainable scalar parameters for the shared spline transformation, independently of the SPD matrix dimension. The same spline is applied across all eigenvalues in S-SPM or all Cholesky diagonal entries in C-SPM, rather than learning separate transformations for individual matrix components.

The existing runtime comparisons also demonstrate that this additional flexibility introduces only little practical overhead. For example, on the FPHA Linear Probe experiment, S-SPM requires 1.32 s/epoch, compared with 1.30 s/epoch for the fixed LE/LC baselines.

Thus, SPM introduces a small parameter overhead and a modest measured computational cost while providing an adaptive transformation that fixed LEM and LCM do not offer. Compared to an MLP, which often requires deep layers for universal approximation guarantees and may not enforce bijectivity, SPM attains universal scalar spectral and Cholesky pullback Euclidean SPD metric approximation, while maintaining high parameter efficiency. We will make these distinctions explicit in the camera-ready version.

AuthorsOpenReview ↗

Response to Reviewer gA5r-2

W3: The universal approximation claim

We respectfully clarify that the submission does not claim that SPM approximates arbitrary Riemannian metrics or arbitrary curved geometries. The approximation domain is explicitly restricted in Theorem 2 and Corollary 1: SPM is a universal approximator of scalar diffeomorphisms over compact spectral intervals, and consequently subsumes the corresponding scalar-diffeomorphism-based spectral and Cholesky pullback Euclidean metrics. Theorem 5 establishes the globally flat curvature of the resulting pullback geometry.

We agree that this scope should be made explicit wherever the term "universal" appears. In particular, SPM is not claimed to reproduce AIRM or Bures-Wasserstein geometries with nonzero sectional curvature. Its universality is within the specified family of scalar-diffeomorphism-induced, globally flat pullback metrics. The flatness is also a deliberate computational property of the construction: it enables closed-form operations in the pullback space, rather than requiring iterative procedures associated with general. We will ensure the scope of our claims is clear in the camera-ready version.

W4: Statistical rigor The current submission reports paired McNemar tests with continuity correction, Benjamini–Hochberg FDR adjustment, and bootstrap confidence intervals for the extended AI-NLP and AFEW evaluations.

We agree that equivalent statistical analyses on HDM05, FPHA, and Radar would strengthen the main experimental results. Due to limited time in the rebuttal window, we aim to incorporate those results in the camera-ready version.

Q1 Monotonic spline transformations We have highlighted our distinction above, and we will make it explicit in the manuscript.

Q2: How many learnable parameters, operator steps, and additional computational burden does SPM introduce compared with fixed LEM/LCM?

As stated, the current implementation uses a grid size of 10 and spline order 3. The number of trainable spline parameters is 13. These 13 parameters are shared across all eigenvalues in S-SPM and all Cholesky diagonal entries in C-SPM. The parameter count therefore does not scale with the SPD matrix dimension or the number of samples. The ablation results show that the additional flexibility is obtained with a modest practical computational cost.

Q3: Have simpler metric designs We have evaluated several simpler alternatives and lower-capacity configurations.

First, the fixed LEM and LCM transformations serve as non-learnable baselines, allowing us to compare adaptive spline-based transformations against established fixed coordinate mappings.

Second, Appendix E, Table 6 reports a nine-configuration FPHA Linear Probe ablation over grid sizes 5, 10, and 20 and spline orders 2, 3, and 4. The corresponding configurations range from 7 to 24 trainable parameters, with the default configuration using 13. Performance remains reasonably consistent across these configurations, indicating that SPM does not rely on a single narrowly selected spline resolution or order.

Third, Appendix J compares SPM with a monotone piecewise-linear mapping. Although the piecewise-linear mapping preserves monotonicity, its C0C^0 kinks lead to highly unstable optimization and active divergence in the eigenvalue optimization experiment, whereas SPM's C2C^2 interior regularity enables smooth convergence.

Together, these comparisons evaluate fixed transformations, lower- and higher-capacity splines, and a simpler monotone piecewise-linear alternative. We agree that lightweight design is crucial for adaptive metric learning, and SPM was designed keeping this explicitly in mind.

Q4: AIRM and Bures Wasserstein The purpose of SPM is not to reproduce all possible SPD geometries, but to provide a learnable and theoretically characterized family of flat pullback metrics, motivated by their popularity in the literature and favorable computational tractability over classical curved metrics. Hence, as established in Theorem 5, the global pullback metrics induced by the SPM diffeomorphism are flat by design.

However, extending SPM to curved geometries, by potentially replacing the fixed conformal factor of a conformal metric with a learnable spline, remains an exciting future direction.

Q5 McNemar tests: As stated above, we aim to incorporate the additional test results in the camera-ready version.

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

This paper provide a new aspect for Symmetric Positive Definite (SPD) manifolds, it criqutes fixed method(e.g. LEM, LCM, AIRM) has limited expressive capacity, and recent works has different weakness in each method. Therefore this paper propose Spline-Pullback Metric (SPM), also two varient(Spectral-SPM, Cholesky-SPM). And authors conduct experiment on HDM05, FPHA, Radar with Linear Probe, SPDNet, RResNet and compared with 6 metrics, and shows SPM have the SOTA performence.

Critical strengths
  • Authors conduct richness experiment results and analysis.
  • This paper proposal a simple, well-motivated core idea(fixed vs learnable).
  • Corollart 1 gives a clear connection between SPM and established metrics such as LEM and LCM, because Corollart 1 shows these metrics are special cases of SPM.
Critical weaknesses
  • Missing experiment to show the improvement is come from SPM's powerful expressibility, rather than learnable. Because ablation did not include the case of low expressibility method's performence.
  • There has no direct experiment results support training stability claim, the paper claims that SPM has stable backpropagation, but this claim is supported only indirectly.
  • Minor improvement in HDM05 RResNet and Radar RResNet.
  • limited task diversity(mainly on classification accuracy) and dataset scale.
Questions for authors
  • Could ablation report one more column for control coefficients (e.g. a SPM variant with only 2-3 control coefficients cic_i)?
  • Could authors add experiment related to stabillity claims? such as gradient-norm curves, NaN/failure rates per metric, how often each safeguard is triggered, and results with the safeguards disabled.
  • Could author eval multi random seed result?
3 comments
AuthorsOpenReview ↗

Response to Reviewer BhnP - 1

We thank the reviewer for recognizing the motivation and breadth of the experiments, and for the constructive suggestions on isolating expressivity and directly measuring optimization stability.

W1: SPM's expressive capacity and learnability

We would like to clarify that the FPHA hyperparameter ablation in Appendix E does vary both grid size and spline order. In our implementation, changing the grid size changes the number of spline control points, and therefore the available deformation resolution and parameterization capacity. The nine configurations in Table 6 consequently provide an existing low- versus higher-resolution spline ablation.

The results remain reasonably consistent across grid sizes 5, 10 and 20 and spline orders 2, 3 and 4, indicating that the observed performance is not dependent on one narrowly selected configuration. In addition, Appendix J compares SPM with a simpler monotone piecewise-linear map. The PLM fails in the eigenvalue optimization experiment because its (C^0) kinks produce unstable step-gradients, while SPM's (C^2) interior regularity enables smooth convergence. Together, these experiments support the importance of the smooth, constrained parameterization rather than learnability alone.

W2: Training stability We would like to distinguish the mathematical properties of the map from the computational safeguards used in finite-precision training. The monotonic parameterization guarantees a globally bijective scalar map with a strictly positive derivative. Separately, the manuscript describes a computational safety protocol for handling repeated or nearly repeated eigenvalues and numerical decomposition failures. We do not intend the safeguard protocol to be interpreted as a theorem guaranteeing stability of the entire deep network's optimization.

The submission already contains several relevant empirical observations: (i) the nine-configuration FPHA ablation in Table 6 shows consistent accuracy and runtime; (ii) Appendix G visualizes the learned S-SPM transformations across five Radar folds, showing consistent nonlinear shapes; and (iii) Appendix J directly compares optimization under a (C^0) PLM and the (C^2)-interior SPM, where the PLM diverges while SPM converges in fewer than 20 epochs.

We agree that gradient-norm trajectories, safeguard activation/failure counts and controlled safeguard ablations would provide more direct evidence. We will add these diagnostics in the camera-ready version and will report the numerical safeguards separately from the continuous mathematical guarantees.

W3: The improvements on HDM05 RResNet and Radar RResNet

We agree that the gains are not uniform in magnitude across all dataset--architecture combinations. The contribution is not that every SPM variant produces a large improvement in every setting. Rather, the results show that the learnable pullback construction can provide benefits across different SPD representations and architectures, with the strongest gains appearing in particular settings. We also respectfully highlight the statistically significant gains on the large-scale AI-NLP and AFEW datasets. We will moderate any wording that could imply uniformly large gains and discuss the complementary roles of S-SPM and C-SPM more carefully in the camera-ready version.

W4: Task diversity and dataset scale

We respectfully wish to clarify that, in addition to the three core SPD benchmarks, the submission also includes two extended large-scale evaluations: DAIR AI-NLP with 20,000 covariance matrices and AFEW with (400x400) covariance matrices. The sources of both of these datasets (NLP and facial expressions) remain distinct from the 3 SPD benchmarks. These experiments broaden the evaluation beyond the core action-recognition and radar settings and include paired statistical testing with FDR adjustment and bootstrap confidence intervals. We agree that additional task types and independent distance-based evaluations would strengthen the work further, and we aim to include these as future camera-ready extensions where feasible.

Q1 Control coefficients

We have already provided a control-point ablation. Howerver, we agree that an explicitly minimal 2--3-control-point configuration column would make the capacity trend clearer and we aim to add it in the camera-ready version.

AuthorsOpenReview ↗

Response to Reviewer BhnP - 2

Q2: Gradient norms, NaN/failure rates, safeguard triggers and safeguard-disabled results

The existing paper provides a computational safeguard protocol and empirical evidence from the FPHA hyperparameter ablation, Radar fold-wise spline visualizations and the PLM-versus-SPM eigenvalue optimization experiment. The latter directly illustrates the optimization consequences of a nonsmooth monotone map. We also tracked the percentage of violations of our linear-tail spline safeguard across all training epochs across all experiments and found them to be exactly 0.0%

Nevertheless, we agree that systematic diagnostic of gradient norms, safeguard activation frequencies, and non-safeguard results will strengthen the paper. We will add gradient-norm trajectories, numerical failure counts and a controlled safeguard ablation in the camera-ready version.

Q3:Random seeds

HDM05 and Radar are evaluated using five-fold cross-validation, and Appendix G shows consistent learned spline shapes across the five Radar folds. FPHA, however, uses a predefined official train/test split, and the reported result does not include a standard deviation over independent training seeds. We agree that cross-validation variability and random-seed variability measure different sources of variation. We aim to add multi-seed evaluation for representative experiments in the camera-ready version.

Thank you for the response. While some points have been clarified, but my main concerns (minimal-capacity ablation, direct stability evidence, and multi-seed results) remain unaddressed. I therefore keep my score unchanged. But I am open to adjusting it after discussion with the AC and the other reviewers.

Reviewer_q1mjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

the paper argues that instead of manually choosing a fixed Riemannian metric for symmetric positive definite (SPD) matrices, we should learn the geometry directly from data, while guaranteeing that the resulting transformation remains mathematically valid

Critical strengths

Introduction of a learnable SPD metric that preserves desirable geometric properties by construction.The paper addresses an important problem in geometric deep learning: learning expressive, data-adaptive Riemannian metrics on SPD manifolds while preserving mathematical validity.

Critical weaknesses

Metric Learning for Clifford Group Equivariant Neural Networks, is relevant prior work for the motivation behind the Spline-Pullback Metric (SPM) paper --> could be added to the references (Learnable symmetric metric matrix)

The paper repeatedly describes SPM as a universal framework for Riemannian metric learning. However, Theorem 2 establishes universal approximation only for strictly increasing C1C^1 scalar diffeomorphisms on compact intervals --> does not imply universal approximation of arbitrary Riemannian metrics on the SPD manifold

Questions for authors

The paper argues that asymmetric eigenvalue perturbations bound the Lipschitz constant of the spectral backward pass. --> Can the authors provide a rigorous bound on the backward operator norm in terms of the spline parameters and quantify how much their spectral perturbations bias the computed gradients?

How are the global curvature and Levi-Civita connection claims justified when the spline transformation is only C1C^1 at its extrapolation boundaries? (theorem 1)

2 comments
AuthorsOpenReview ↗

Response to Reviewer q1mj

We thank the reviewer for recognizing the importance of preserving mathematical validity in learnable SPD metrics and for the constructive questions regarding prior work, the spectral backward pass and boundary regularity.

W1 Related Works: We thank the reviewer for pointing out this relevant work. We will add Metric Learning for Clifford Group Equivariant Neural Networks to the related-work discussion and compare its equivariant representation approach with our strictly monotone scalar Spline diffeomorphism approach.

W2: The universality claim We agree with the mathematical distinction and respectfully clarify that the theorem is already stated for a restricted function class. Theorem 2 establishes density in the space of strictly increasing scalar diffeomorphisms over compact spectral intervals. Corollary 1 applies this result to the corresponding scalar-diffeomorphism-based pullback Euclidean metric families. Theorem 5 establishes that these pullback metrics are globally flat. Thus, we do not claim approximation of arbitrary Riemannian metrics. However, we agree that some wordings may have sounded more general. We will make this scope more explicit in the final wording.

Q1 We thank the reviewer for this important question. The backward operator is characterized by the Daleckii-Krein divided-difference matrix. Because the eigenvector matrix is orthogonal, the operator norm of the backward pass is bounded by the maximum absolute value among the elements in this divided-difference matrix.

For our chosen parameterization (which applies a B-spline to the matrix logarithm), the scalar derivative is simply the derivative of the spline divided by the input value. The B-spline formulation guarantees that the spline's derivative is bounded globally by a constant, determined entirely by its degree, control-point increments, and knot spacings. Let us refer to this maximum spline derivative bound as M.

Our numerical implementation further employs a two-tier perturbation strategy (detailed in Appendix H.2). First, a positive diagonal perturbation is added to the input SPD matrix before eigendecomposition. Since this perturbation is positive definite, it preserves positive definiteness and ensures a strict positive lower bound on the perturbed eigenvalues. In our implementation, this creates a minimum eigenvalue floor of 10^-8. Second, if nearly repeated eigenvalues persist, an asymmetric eigenvalue perturbation enforces a minimum separation between them.

Combining this positive eigenvalue floor with the globally bounded spline derivative, and applying the mean value theorem to the divided differences, we obtain our final numerical bound: the norm of the perturbed backward operator is strictly bounded by the maximum spline derivative (M) divided by the minimum eigenvalue floor (10^-8).

Thus, the positive eigenvalue floor controls the potentially unbounded scalar derivative near zero, while the asymmetric perturbation provides additional numerical protection against eigenvalue degeneracy.

Regarding perturbation-induced gradient bias, Weyl's inequality bounds the displacement of each eigenvalue due to the first-tier input perturbation by 2 x 10^-8. However, this alone does not quantify the complete matrix-gradient error, which also depends on the variation of the spline derivative and the eigenvectors. We will clarify this distinction in the revised manuscript rather than overstate the existing perturbation guarantee.

Q2 We thank the reviewer for identifying the distinction between the regularity of the active spline region and that of the finite linear extrapolation tails. This distinction is explicitly addressed in Appendix B.1 of our submission.

In our proof of Theorem 1, we state, "Theoretically, by extending the grid boundaries t0 → −∞ and tm → ∞, the linear tails vanish, and C2C^2 smoothness holds globally on R>0. Higher-order differentiability can easily be obtained by increasing k."

For the cubic B-spline used in our experiments, the mapping is C2C^2 throughout the active knot interval. The linear extrapolation tails are introduced solely as a computational fallback for extreme spectral outliers, matching the spline in value and first derivative and thus preserving global C1C^1 regularity. On any prescribed compact spectral domain, the knot boundaries can be chosen sufficiently wide that the entire domain lies within the C2C^2 spline region.

Furthermore, none of the evaluated inputs in all of our reported experiments required linear-tail extrapolation: the spectral values remained within the prescribed spline domain. Consequently, the finite transition joins were not encountered during the reported experiments, and the evaluated transformations operated entirely within the C2C^2 region, precluding concerns regarding curvature calculation and parallel transport.

We will make this distinction clearer in the manuscript.

Reviewer_q1mjOpenReview ↗

Comment

Thank you for the detailed response. I share the other reviewers' concerns regarding the missing controlled baselines and direct empirical evidence of numerical stability. Several relevant experiments have been proposed for the camera-ready version but are not yet available for assessment, so I will keep my current score.

Reviewer_QE9COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a learnable Riemannian metric framework for SPD manifolds (spline-pullback metric). Instead of using predefined algebraic Riemannian metrics, the method utilizes monotonically constrained B-splines to parameterize a globally invertible transformation of SPD matrices. The framework has two variants - 1) Spetral-SPM (applying the spline to eigenvalues) and 2) Cholesky-SPM(applying the spline to the diagnoal elements of Cholesky factors).

The primary contributions is that it uses B-splines to construct learnable pullback Riemannian metrics while preserving mathematical properties such as global diffeomorphism. The authors' experiments show that SPM achieves better classification performance compared to several existing metrics across benchmark datasets.

Critical strengths
  1. This work is novel as an assembly. To my knowledge, it might be the first work that uses monotonically constrained B-splines to construct learnable pullback metrics on SPD manifolds. The method's validity is also apparently shown in its experiments comparing SPM to several existing Riemannian metrics. The improvement of the method, especially S-SPM, is consistent across the reported classification benchmarks.

  2. This work has a clear and consistent use of relevant mathematical knowledge. It's relatively well-written. The theoretical results regarding global diffeomorphism, universal approximation, and the resulting Riemannian geometric properties are presented in a logical order.

Critical weaknesses
  1. The proposed method heavily relies on supervised classification to learn the Riemannian metric, but the learned metric itself is not independently evaluated. The B-spline transformation is jointly trained with a downstream classifier, which means the learning objective is essentially to find a nonlinear transformation of SPD matrices that improves classification accuracy. Two concerns follow this fact: 1) the learned metric is not necessarily capturing any meaningful underlying geometric structure of the SPD manifold, but rather learning a task-specific feature transformation that makes the classes more easily separable. This thus reduces the significance of the paper's contribution, since it is unclear how this differs from ordinary supervised feature learning; 2) all the experiments evaluate the performance through downstream classification without directly evaluating the learned distances between SPD matrices. For instance, nearest-neighbor classification, clustering, or retrieval using the learned geodesic distances would actually provide more convincing evidences that the metric itself is meaningful. It is also unclear whether the learned metric can be transferred to other tasks without retraining the spline. Based on the above two points, the lack of independent evaluation of the learned metric is a major weakness of this work.

  2. The performance improvement might simply be coming from introducing additional nonlinear parameters rather than the proposed Riemannian geometry. In Section 6, the authors compare SPM with several existing metrics, most of which are either fixed analytical transformations or have more restricted parameterizations. However the proposed B-spline is a learnable nonlinear function, which naturally has higher expressivity compared to fixed transformations such as Log-Euclidean. Therefore, the improvement in classification accuracy might simply comes from the additional learnable parameters. A natural baseline would be applying the Log-Euclidean metric followed by a small MLP that has approximately the same number of parameters as SPM - if such a baseline can achieve comparable performance, then the claimed advantage of the proposed geometric formulation would become much less convincing. The current experiments don't really distinguish whether the improvement comes from learning a Riemannian metric or simply adding a nonlinear feature transformation.

  3. The universal approximation claim is potentially overstated. In Theorem 2, the authors prove that the proposed B-spline can approximate strictly increasing scalar diffeomorphisms on compact spectral intervals. However this is only a universal approximation result for a restricted family of 1D functions rather than arbitrary Riemannian metrics on SPD manifolds. Also, Theorem 5 states that the proposed pullback metrics have globally zero sectional curvature, which means they cannot represent general Riemannian geometries with nonzero curvature such as AIRM. In another word, the expressivity of the proposed method is fundamentally limited to a family of globally flat pullback metrics. Although the theoretical result is valid within the specified function class, I think the paper's emphasis on "universal" Riemannian metric learning is somewhat misleading, since it doesn't really cover general Riemannian geometries.

Questions for authors
  1. Can you evaluate the learned metric independently of the classifier (such as nearest-neighbor classification, clustering, or retrieval) using the learned geodesic distances?
  2. Can the learned B-spline metric generalize to a different downstream task without retraining? What happens if the spline is frozen and a new classifier is trained?
  3. Can you compare SPM against a Log-Euclidean transformation followed by a parameter-matched MLP? How much of the improvement comes from the additional nonlinear expressivity rather than the Riemannian geometric construction?
  4. Regarding Theorem 2, since SPM only approximates scalar diffeomorphisms and induces globally flat metrics, in what sense is the proposed method a universal approximator of Riemannian metrics?
  5. What exactly is the loss function used to train the B-spline? Is it only the downstream classification loss or are there additional geometric objectives or regularization terms?
3 comments
AuthorsOpenReview ↗

Response to Reviewer QE9C - 1

We thank the reviewer for recognizing the novelty of using monotonically constrained B-splines to construct learnable pullback metrics on SPD manifolds, and for the positive assessment of the mathematical development and consistent S-SPM results. We address the concerns below:

W1.1: Joint training with a classifier We agree that the learned geometry is task-adaptive: the spline parameters are optimized jointly with the downstream objective. Our objective is to learn a mathematically valid, task-adaptive pullback metric that would benefit the classification task at hand while maintaining diffeomorphism guarantees, not to claim that the learned metric necessarily recovers a geometry independent of the task.

Importantly, the manuscript uses the Linear Probe setting in which the uncompressed SPD matrices are mapped using each metric and are then passed only to a single linear classifier, without a deep feature extractor that could compensate for a weak metric. The favorable results achieved by SPM over other metrics provide evidence that the learned metric transformation itself contributes to class separability, rather than the gains arising solely from a highly expressive downstream backbone.

We acknowledge that a linear probe is still a supervised evaluation and is not a substitute for direct distance-based evaluation. We are conducting the experiment and will add nearest-neighbor classification/retrieval using the learned pullback distance in the camera-ready version. We will also clarify the distinction between task-adaptive metric learning and task-independent geometric recovery.

W1.2: Transfer to a different downstream task

We agree that frozen-spline transfer is a useful test of whether the learned deformation captures reusable structure beyond the training objective. The current work focuses on end-to-end task-adaptive metric learning, and does not claim task-independent pretraining or guaranteed cross-task transfer.

The existing Radar analysis in Appendix G provides related evidence: the learned S-SPM transformations are visualized across five independent cross-validation folds, where they converge to consistent nonlinear shapes despite changes in the data split. This supports stability of the learned spectral deformation across folds, but we do not present it as a cross-task transfer experiment.

We will add a frozen-spline experiment in the camera-ready version: the spline learned on one task will be frozen while a new downstream classifier is trained, and its performance will be compared with a jointly trained SPM model. We will also add an independent distance-based evaluation as described above.

W2: A parameter-matched Log-Euclidean-MLP and expressivity

We agree that the ablated baselines are more rigid than SPM, but this is a central motivation of SPM: we wish to present a more expressive SPD pullback Euclidean metric that would reduce reliance on handcrafted rigid metrics popular in recent literature. We also appreciate this suggestion of the MLP experiment. We would like to clarify that the central distinction is not only the forward expressivity of the transformation. SPM is designed to parameterize a globally bijective, differentiable map with a differentiable inverse (a diffeomorphism), so that it induces a valid pullback geometry and supports the associated Riemannian operations. A generic MLP, even when it has a comparable number of parameters and can represent a flexible forward transformation, does not by itself guarantee injectivity, surjectivity, a nonsingular Jacobian everywhere, or a differentiable inverse required by a Riemannian metric.

The monotonic B-spline parameterization provides these properties by construction through positive control-point increments. Its smoothness can also be increased systematically by increasing the spline degree/order. In particular, cubic splines have (C^2) regularity in the interior, and higher-degree splines provide higher finite-order differentiability. Thus, the comparison is not simply between two nonlinear function approximators: it is between a generic nonlinear transformation and a constrained diffeomorphic transformation designed to define a valid pullback metric.

The submission already includes a relevant controlled experiment in the Appendix J comparing SPM with a monotonically constrained piecewise-linear map (PLM), which is a simpler low-complexity monotone transformation. In an eigenvalue optimization task targeting 6.0, the PLM encountered a gradient explosion at its (C^0) kink, with a loss spike, and actively diverged after 700 epochs, In contrast, SPM's (C^2) produced smooth gradients and converged to 6.13 with final loss 0.208 in fewer than 20 epochs.

We will make this distinction clearer in the rebuttal and include a parameter-matched MLP comparison as an additional camera-ready control, while emphasizing that a generic unconstrained MLP is not itself a diffeomorphic metric baseline.

AuthorsOpenReview ↗

Response to Reviewer QE9C - 2

W3: The universal approximation claim

We respectfully clarify that the submission does not claim that SPM approximates arbitrary Riemannian metrics or arbitrary curved geometries. The approximation domain is explicitly restricted in Theorem 2 and Corollary 1: SPM is a universal approximator of scalar diffeomorphisms over compact spectral intervals, and consequently subsumes the corresponding scalar-diffeomorphism-based spectral and Cholesky pullback Euclidean metrics. Theorem 5 establishes the globally flat curvature of the resulting pullback geometry.

We agree that this scope should be made explicit wherever the term "universal" appears. In particular, SPM is not claimed to reproduce AIRM or Bures-Wasserstein geometries with nonzero sectional curvature. Its universality is within the specified family of scalar-diffeomorphism-induced, globally flat pullback metrics. The flatness is also a deliberate computational property of the construction: it enables closed-form operations in the pullback space, rather than requiring iterative procedures associated with general. We will ensure the scope of our claims is clear in the camera-ready version.

Q1: Independent Evaluation

As mentioned above, this ablation will surely strengthen the paper. We are conducting the experiments and will incorporate them in the camera-ready version. In the current version, lightweight linear probes partially isolate the metric's impact from heavy downstream processing.

Q2: Can the learned spline transfer to a different task when frozen?

As mentioned above, we will add a frozen-spline/new-classifier experiment in the camera-ready version to evaluate this question directly. Additionally, another observation that might be beneficial for this question is that, for the other datasets as well, a spline profile similar to the non-linear spline profile reported for the Radar dataset in the manuscript was observed. This might provide complementary evidence of the generalizability of the spline's learning across different tasks; however, the explicit experiment will demonstrate this definitively.

Q3: Non-linear expressivity

Non -linear learnability is a key motivation behind SPM over rigid existing metrics, though the key distinction is that SPM's nonlinear map is constrained to be a global diffeomorphism, rather than being an arbitrary nonlinear feature transformation. We agree that a parameter-matched MLP is a useful additional reference for forward expressivity and will include it in the camera-ready version.

Q4: SPM's universality

SPM situates its novelty and universality in the popular and active subfield of pullback Euclidean metrics, all of which aim to induce flat geometry due to their computational benefits. However, future work can readily extend the SPM formulation to curved geometric pullback as well. A simple formulation would be to make the conformal factor of a conformal metric SPM-learnable. As said above, we will make these distinctions clearer in the camera-ready version.

Q5: Loss function

The spline parameters are optimized jointly with the downstream model using the downstream classification objective. There is no additional geodesic-distance supervision or separate geometric regularization term in the reported experiments. The monotonicity constraint is imposed architecturally through the cumulative positive-increment parameterization of the control coefficients, rather than through a penalty term. We will state this explicitly.

However, as we mentioned in the conclusion of our manuscript, using SPM opens the promising future direction of Riemannian Metric Regularization (RMR) in which adding a loss penalty specific to the spline parameters could uncover interesting interactions between controlling metric expressivity and downstream performance. We leave that as future work.

We thank the reviewer for motivating these clarifications and suggesting additional experiments. Kindly let us know if you have any remaining concerns/ suggestions, and we respectfully ask that you consider these clarifications in your final assessment and rating.

Reviewer_QE9COpenReview ↗

Response to authors

I appreciate the authors for their rebuttal. However, since there are no new experiment results that resolve my concerns and a lot of the changes are postponed to either future works or camera-ready version, I maintain my current score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.