acceptodds
Under review as a conference paper at ICLR 2027

Nested Inductive Bias Framework for Curvature-Aligned SPD Manifold Learning

Abstract

In Geometric Deep Learning, inductive biases serve two primary functions: enforcing manifold constraints and embedding relational priors. Currently, representation learning on SPD manifolds frequently relies on pullback Euclidean metrics to satisfy the former. While computationally efficient in avoiding domain boundary violations, these metrics induce a flat geometry that may fail to capture the intrinsic relational priors of datasets. While metrics such as the Poincaré metric are widely utilized to induce domain-aligned relational priors, generalizing them from standard vector representations to the SPD manifold has remained a challenge. To bridge this gap, we introduce a Nested Inductive Bias framework that utilizes a multistage diffeomorphic composition to formally pull back non-Euclidean target geometries onto the SPD manifold. Moreover, we identify that mapping unbounded Euclidean representations onto manifolds with finite injectivity radii induces a "Norm Mismatch," leading to degraded performance and diffeomorphic failure. We resolve this by introducing the Rational Conformal Metric (RCM), which preserves bijectivity as well as provides a strong regularization against outliers. Empirical evaluations on kinematic and signal processing benchmarks, together with synthetic experiments, demonstrate that deep manifold networks experience degradation in class separability due to curvature/norm mismatch between the metric and the intrinsic data distribution. Finally, our curvature-aligned Riemannian classifiers resolve the computational intractability of classical non-flat SPD metrics (like AIRM), achieving a 95% reduction in computational overhead.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Geometric Deep Learning, SPD Manifolds, Representation Learning, Riemannian Geometry, Diffeomorphism, Inductive Bias

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_yqRcOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces a Nested Inductive Bias framework that equips SPD manifolds with hyperbolic or spherical geometry by pulling back a target metric via a composition of logarithmic and exponential maps. It argues that manifold learning should account not only for SPD constraints but also for the latent relational structure of the data, and introduces a bounded Rational Conformal Metric (RCM) for outlier-robust representations.

Critical strengths

The nested metrics are constructed using a composition of diffeomorphisms and pullback metrics, relying on concepts from differential geometry rather than heuristics.

The paper explicitly distinguishes between enforcing manifold constraints and encoding relational structure, providing an interesting way to look at geometric inductive biases in SPD learning.

Critical weaknesses

The main construction is pullback metric obtained through a composition of established maps, and much of the theory follows from standard results on diffeomorphisms and pullbacks.

Limited comparison to existing metric-learning approaches on SPD manifolds. The paper does not convincingly establish that adaptive or learned SPD metrics cannot achieve similar effects without the nested construction.

The choice for the conformal factor in the Rational Conformal Metric could be better justified and compared to alternatives.

Questions for authors

The Rational Conformal Metric is an original part of the submission. Why is the specific conformal factor (1+\alpha|x|)^{-2} preferred over other possible factors? Is there a theoretical property that motivates this choice?

6 comments
AuthorsOpenReview ↗

Authors' Response to Reviewer yqRc - 1

We thank the reviewer for recognizing the geometric coherence of the nested construction and the originality of RCM. We appreciate the questions concerning novelty, adaptive SPD metrics, and the motivation for the conformal factor.

W1. Novelty Within Geometric Deep Learning

We agree that diffeomorphic composition and metric pullback are established differential-geometric principles. We do not claim novelty for these abstract operations. We situate our novelty within the subfield of Geometric Deep Learning and SPD manifold learning.

Our contribution for ICLR is to present a mathematically rigorous, computationally tractable SPD learning framework, without relying on heuristics.

The novel methodological contributions are:

(i) Curved SPD pullback metrics. To the best of our knowledge, we are the first to equip the SPD manifold with hyperbolic and spherical metrics via the nested inductive bias framework, imparting non-flat curvature and averting the high computational penalty of the classical AIRM metric.

(ii) Solves a> 5-year-old problem: It identifies the main cause (injectivity violation) and provides a principled solution (via RCM) to the problem of the instability of optimization on the projected hypersphere (mentioned by Skopek et al. [3] in 2020) using nested differential geometry, instead of relying on heuristics. We also provide beneficial properties of RCM in the rebuttal.

(iii) Intrinsic classification. Besides the Nested Bias framework, we also provide the derivation and implementation of the associated Riemannian classification procedure (Nested RMLR) specifically for the induced geometries.

(iv) Metric Alignment: We provide evidence for the alignment between the dataset's empirical ORC and the selected metric beyond merely satisfying the domain constraints. This concept is generalizable to non-SPD datasets as well.

W3 & Q1: Justification for the RCM Conformal Factor

The specific conformal factor used in the Rational Conformal Metric was deliberately derived to solve a fundamental tension in bounded representation learning: a valid bounding metric must simultaneously guarantee C∞C^\infty smoothness everywhere (including the origin) and preserve suitable training while bounding the codomain. We present both theoretical arguments and empirical results on the FPHA dataset, contrasting RCM against competitor bounded transformations.

For a radial bounding mapping f(r)f(r) applied to a vector xx with norm r=∥x∥2r = \|x\|_2, the vector mapping is F(x)=xf(r)rF(x) = x \frac{f(r)}{r}. We compared our RCM formulation against alternative mathematical bounds:

PropertyL2L_2 NormSimple RationalRational SqrtExponentialTanhProposed RCM
f(r)f(r)11r1+r\frac{r}{1+r}r1+r2\frac{r}{\sqrt{1+r^2}}1−e−r1-e^{-r}tanh⁡(r)\tanh(r)r1+α(r2+ϵ2−ϵ)\frac{r}{1+\alpha(\sqrt{r^2+\epsilon^2}-\epsilon)}
C∞C^\infty Smooth EverywhereNoNoYesNoYesYes (ϵ>0\epsilon > 0)
BijectiveNoYesYesYesYesYes
Radial Gradient Decay (∂f∂r\frac{\partial f}{\partial r})00O(r−2)O(r^{-2})O(r−3)O(r^{-3})O(e−r)O(e^{-r})O(e−2r)O(e^{-2r})O(r−2)O(r^{-2})
C∞C^\infty Origin and O(r−2)O(r^{-2}) FlowNoNoNoNoNoYes

Note: RCM parameter conditions α>0\alpha>0, ε>0\varepsilon>0, and αε<1\alpha\varepsilon<1.

Theoretical Motivation for the Specific Factor:

  1. Standard bounds like Tanh and Exponential shrink the gradient exponentially O(e−r)O(e^{-r}). For high-variance datasets, this induces catastrophic gradient vanishing quickly, stalling training, as we show below. RCM exhibits polynomial O(r−2)O(r^{-2}) radial derivative decay, which avoids the exponentially decaying gradient alternatives considered here.
  2. L2L_2 and the simple rational bounds are not C∞C^\infty at the origin, making them unsuitable for our diffeomorphic framework, as curvature calculations require higher-order differentiability.
  3. The parameter α\alpha in RCM controls the limiting radius of the radial transformation. This provides a geometrically interpretable way to control the bounding envelope, which might be learnably optimized in future work.

To empirically validate this specific conformal factor, we ablated RCM against these popular alternatives on the FPHA dataset (Linear Probe, multiple seeds):

Radial Bounding MappingMean Accuracy ±\pm Std
Proposed RCM (Ours - O(r−2)O(r^{-2}))84.52% ±\pm 0.51%
Rational Sqrt (O(r−3)O(r^{-3}))9.16% ±\pm 1.08%
Exponential (O(e−r)O(e^{-r}))2.78% ±\pm 0.00%
Tanh (O(e−2r)O(e^{-2r}))2.78% ±\pm 0.00%
(Unconstrained NSM baseline)13.28% ±\pm 2.26%

While the remaining mappings suffer from rapid gradient vanishing (even the Rational Sqrt mapping's cubic (O(r−3)O(r^{-3})) decay proves too aggressive here), the specific rational formulation in RCM provides both differentiability and proper training.

AuthorsOpenReview ↗

Authors' Response to Reviewer yqRc - 2

W2. Adaptive and Learned SPD Metrics

We agree that learned SPD metrics are important empirical alternatives. The reviewer correctly identifies that adaptive metric-learning approaches exist for SPD manifolds (even as recently as 2026). However, we situate our novelty within the current literature by noting that these adaptive base metrics generally pull back the flat Euclidean metric.

The primary motivation for these pullback Euclidean metrics becoming so popular in recent SPD literature is to overcome the computational burden of the classical non-positively curved AIRM metric, but this comes at the cost of sacrificing curvature. Learnable metrics (such as ALEM[1], SSPM, and CSPM[2]) optimize the weights of the mapping from the SPD manifold to flat Euclidean space to improve class separation. They do not imbue the manifold with non-zero curvature.

Instead of forsaking these Euclidean advances, we designed our framework in a modular fashion that can easily integrate these pullback Euclidean metrics. Since the first diffeomorphism in our nested framework is the Log-Euclidean transformation, we can replace it with a learnable metric's Log transformation while still extending them to non-Euclidean geometries. Notably, these metrics should themselves be valid C∞C^\infty Riemannian metrics to substitute LEM.

To test this, we evaluated LEM, SSPM, ALEM, and CSPM as standalone Euclidean RMLR classifiers on FPHA (Linear Probe), and then integrated their representations into the three nested target configurations.

Base MetricStandalone (Flat Euclidean RMLR)Nested in NSM (Unconstrained)Nested in NSM-RCM (Proposed)Nested in NHM (Hyperbolic)
LEM83.13%10.78%83.30%85.57%
SSPM85.22%20.87%85.22%87.30%
ALEM84.00%7.83%87.13%89.22%
CSPM58.96%6.78%47.65%62.43%

These observations follow.

  1. While adaptive metrics remain competitive (such as standalone SSPM at 85.22%) and improve upon the standard LEM, they fall short of the performance achieved by simply placing the LEM into our Hyperbolic framework. This provides evidence that even a learnable flat metric may not fully substitute for aligning the metric's curvature with the dataset's intrinsic curvature.

  2. When we apply our nested framework on top of the adaptive metrics, especially the FPHA's curvature-aligned NHM, performance increases (ALEM-NHM achieves the highest 89.22%). Pairwise McNemar's tests with the Holm-Bonferroni correction confirm this complementary performance gain is statistically significant(p<0.05p<0.05). Both Nested ALEM_NSM_RCM and Nested ALEM_NHM significantly outperform standalone ALEM. Moreover, both Nested ALEM_NSM_RCM and Nested ALEM_NHM significantly outperform their standard LEM-based counterparts. This demonstrates the complementary nature of these frameworks.

  3. The importance of RCM is also apparent. For the unconstrained NSM, even when equipped with learnable base metrics such as ALEM or SSPM, it fails due to a norm-and-curvature mismatch. The RCM intervention successfully "saves" its performance significantly, demonstrating its necessity regardless of the base metric used.

While the results show empirical complementarity, they do not guarantee improvement for every base metric replacement over LEM, contingent on the metric's inherent stability and valid diffeomorphism properties.

We hope these additions properly address your concerns, and we respectfully ask you to consider updating your assessment of our submission in light of this new evidence.

References: [1] Chen, Z., Song, Y., Xu, T., Huang, Z., Wu, X. J., & Sebe, N. (2024). Adaptive Log-Euclidean metrics for SPD matrix learning. IEEE Transactions on Image Processing, 33, 5194-5205.

[2] Das, T., Dutta, S., Neogy, S., & Singh, K. K. (2026). Beyond Rigid Geometries: The Spline-Pullback Metric for Universal Diffeomorphic SPD Representation Learning. arXiv preprint arXiv:2605.04406.

[3] Skopek, O., Ganea, O.-E., & Bécigneul, G. (2020). Mixed-curvature variational autoencoders. In International Conference on Learning Representations. https://openreview.net/forum?id=S1g6xeSKDS

Dear Reviewer yqRc,

As the discussion period comes to an end, we wish to thank you for your thoughtful review. We hope that our provided clarifications and new experimental results have addressed all your raised concerns. Please let us know if any final clarifications would be helpful before the window closes.

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

This paper proposes a nested pullback framework for introducing hyperbolic and spherical geometries into SPD learning. It also introduces the Rational Conformal Metric (RCM), which provides a smooth bounded intermediate mapping to address the finite injectivity domain of the spherical exponential map. Experiments evaluate the framework on action recognition and radar classification.

Critical strengths
  1. The distinction between SPD domain constraints and relational geometric priors is clear and well motivated.
  2. RCM provides an explicit treatment of spherical mapping constraints, substantially improving performance over unconstrained spherical classification.
  3. The framework achieves promising improvements on the kinematic benchmarks, with reported runtime comparable to LEM and substantially lower than native AIRM RMLR.
Critical weaknesses

My main concern is the applicability of a single global geometric prior. Each NHM/NSM instantiation adopts a constant-curvature target space, which may not suit heterogeneous structures across classes or local regions. Noise, missing observations, or changes in acquisition conditions may also create a mismatch between the data and the selected geometry. The current robustness experiments mainly use two-dimensional synthetic data, leaving it unclear whether the method retains its advantages when real downstream data deteriorate.

Questions for authors
  1. How are the geometry type and curvature selected, and how sensitive is performance to a mismatched prior?
  2. Could the authors evaluate increasing noise or missing observations on a real downstream task and compare performance degradation against LEM?
4 comments
AuthorsOpenReview ↗

Authors' Response to Reviewer Ynpt - 1

We thank the reviewer for their positive assessment and for recognizing the distinction between SPD domain constraints and relational geometric priors.

We distinguish geometric mismatch, arising from an unsuitable target geometry, from acquisition mismatch, arising from noise or deteriorating observations. We address both below.

W1 & Q1. Geometry Selection and Sensitivity

The framework allows different target geometries, principally hyperbolic (NHM) and spherical (NSM). In our current framework, the target geometry (Hyperbolic vs. Spherical) is selected prior to training by estimating the discrete Ollivier-Ricci Curvature (ORC) of the dataset. As demonstrated in the manuscript, model performance is highly sensitive to a mismatched prior. This is expected. Imposing a Spherical geometry (NSM) onto the negative/hierarchical dataset leads to a norm and curvature mismatch, causing a significant performance drop. Aligning the prior (NHM) with the dataset ORC yields the highest accuracy since the validated datasets show a negative curvature. This is the core thesis of our framework: induced geometry should be selected to align with the dataset's inherent empirical curvature to achieve best classification performance.

Although sensitivity to geometric mismatch is an expected phenomenon that can be averted by selecting the correct metric using our proposed analysis, sensitivity to noise should be measured separately, as we do below.

W2 & Q2. Real-World Noise Robustness

To evaluate robustness beyond two-dimensional synthetic examples, we conducted an additional corruption experiment on the real-world FPHA dataset.

You rightly point out that changing acquisition conditions may alter the data. If a dataset is inherently noisy, the initial ORC calculation captures this, and the appropriate prior is selected. However, to evaluate robustness against progressive downstream deterioration after a metric is selected, we conducted a rigorous noise degradation experiment on the real-world FPHA dataset, instead of relying on synthetic experiments.

We injected symmetric Gaussian noise of different levels directly into the intrinsic manifold structure of the covariance matrices. We evaluated increasing noise conditions (σ∈[0.1,0.4]\sigma \in [0.1, 0.4]) and retrained the networks to assess robustness against the Log-Euclidean Metric (LEM).

Noise Level (σ\sigma)LEM (Flat Baseline)NHM (Hyperbolic)NSM-RCM (Proposed)NSM (Unconstr.)
0.184.17%84.70%83.48%10.26%
0.282.78%85.04%82.96%7.30%
0.379.83%84.87%82.78%9.39%
0.474.61%82.43%78.96%6.96%

The results show:

  1. The flat geometry of LEM is highly sensitive to downstream noise, resulting in a nearly 10% drop in accuracy. The proposed NHM, on the other hand, proves remarkably robust to real-world structural noise, maintaining 82.43% accuracy even under high noise. While unconstrained NSM fails due to norm and metric mismatch as highlighted in the manuscript, our RCM intervention "saves" NSM significantly. Hence, NSM-RCM remains more noise-robust than LEM in this corruption-and-retraining protocol. .

  2. Pairwise McNemar's tests (with Holm-Bonferroni correction) confirm that both NHM and NSM-RCM outperform LEM under high noise with statistical significance (σ=0.4\sigma = 0.4, p<0.01p < 0.01). This answers your question: our proposed framework is significantly more robust to downstream noise and signal deterioration than standard Euclidean baselines.

The fixed-radius synthetic experiment reported to Reviewer TarY provides complementary evidence that the relative performance of spherical and hyperbolic classifiers changes with corruption and that RCM substantially improves the spherical classifier in the high-noise regime.

AuthorsOpenReview ↗

Authors' Response to Reviewer Ynpt - 2

W3. Heterogeneous Structures and Learnable Curvature

We agree that modeling heterogeneous structures perfectly would require a manifold with learnable, locally variable curvature.

The present framework prioritizes explicit geometric operations and computational efficiency. More flexible variable-curvature constructions are possible in principle, but need not retain the closed-form geometric operators required for efficient deep learning.

However, to evaluate adaptability while preserving closed-form computational efficiency, we make the Global Curvature learnable. We extended our RMLR formulation to make the global curvature parameter cc learnable via gradient descent, while retaining the sign of the selected geometric family.

We retrained the FPHA Linear Probe model with the learnable curvature formulation:

Metric TopologyInitial CurvatureLearned CurvatureTest Accuracy
NHM (Hyperbolic)-1.0000-0.801785.57%
NSM-RCM (Spherical)+1.0000+1.022782.09%
NSM (Unconstrained)+1.0000+1.002414.78%

The learned parameters demonstrate that the curvature scalar can be optimized jointly with the classifier. Even in this adaptive regime, the necessity of the RCM intervention to "save" NSM is apparent.

Extending this framework to Product Manifolds (such as S×H×E\mathbb{S} \times \mathbb{H} \times \mathbb{E}) for further representation of heterogeneous local structures remains a promising direction for future work.

We thank the reviewer for highlighting these important distinctions, which we shall address in the camera-ready version. Kindly let us know if you have any remaining concerns, and we respectfully ask that you consider this new evidence in your final assessment and rating.

Thank you for your reply. Most of my concerns have been resolved, so I will keep the score.

AuthorsOpenReview ↗

Thank You!

We are glad that most of your concerns have been resolved. We thank you for your review and constructive feedback throughout the review process.

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

This paper proposes a Nested Inductive Bias framework for representation learning on symmetric positive definite (SPD) manifolds. The framework combines Log-Euclidean mapping with diffeomorphic pullbacks of non-Euclidean target geometries, allowing hyperbolic and spherical relational priors to be incorporated into SPD representations. The authors further introduce the Rational Conformal Metric (RCM), which provides a bounded radial transformation intended to resolve injectivity-domain violations in spherical mappings and improve robustness to outliers. The proposed Nested Hyperbolic Metric (NHM) and RCM-stabilized Nested Spherical Metric (NSM) are evaluated on FPHA, HDM05, and Radar classification benchmarks using linear probes, SPDNet, and Riemannian ResNet architectures. Additional synthetic experiments examine curvature alignment and outlier regularization. The reported results show promising improvements in classification accuracy, particularly on kinematic datasets, together with substantially lower computational overhead than native AIRM classification.

Critical strengths
  1. A principled and coherent geometric framework. The paper addresses a meaningful distinction between preserving SPD manifold constraints and incorporating relational geometric priors. The proposed nested diffeomorphic construction provides a mathematically motivated mechanism for transferring non-Euclidean geometries to SPD representations. The treatment of spherical injectivity-domain limitations through RCM is particularly interesting because it connects a practical numerical problem to an explicit geometric construction. This is an important strength supporting the paper's methodological contribution.

  2. Encouraging empirical improvements with practical computational advantages. The results demonstrate substantial gains on the kinematic benchmarks. In particular, NHM improves RResNet accuracy on HDM05 from 55.14% with LEM to 69.83%, and on FPHA from 76.52% to 86.43%. The RCM intervention also restores much of the performance lost by unconstrained spherical mapping. Furthermore, the reported HDM05 runtime decreases from approximately 751 seconds per epoch for native AIRM RMLR to approximately 38 seconds for the proposed nested metrics. These results provide meaningful evidence of practical utility rather than relying exclusively on theoretical motivation.

Critical weaknesses
  1. The evidence for the general curvature-alignment hypothesis is not yet sufficiently strong. The paper argues that matching model curvature to intrinsic data geometry improves representation learning. However, all three real-world datasets exhibit negative estimated Ollivier-Ricci curvature, so the experiments mainly support the effectiveness of hyperbolic representations on these datasets rather than the broader curvature-alignment principle. The reported correlation of approximately 0.993 is based on only three datasets and should therefore be interpreted cautiously. The synthetic experiment provides useful complementary evidence, but changing the radial embedding scale also changes the relative strength of the underlying geometric signal and ambient noise. Consequently, it is difficult to attribute the observed performance differences exclusively to curvature alignment. These limitations matter because curvature alignment is a central explanatory claim of the paper, beyond the demonstrated classification improvements.

  2. The independent contribution of RCM and the interpretation of its numerical advantages need clearer justification. RCM is a promising construction, but the outlier robustness results do not fully distinguish its specific benefits from those of generic smooth, bounded, direction-preserving radial transformations. For example, in the directional heavy-tailed experiment, L2 normalization already achieves 100% core accuracy and 91.8% total accuracy, compared with 92.3% total accuracy for RCM. This suggests that further controlled comparisons would be useful for isolating the advantage of the proposed metric. In addition, the claim regarding asymptotic gradient stability should be formulated more precisely, since the radial derivative of the bounded transformation approaches zero for sufficiently large input norms. The spherical classifier also retains a potential cut-locus limitation for class-anchor logarithmic maps, although the authors report no violations in their experiments. Clarifying these distinctions would strengthen confidence in the claimed theoretical and practical advantages.

Questions for authors
  1. Evidence for curvature alignment (clarification).How sensitive are the estimated dataset curvatures and the corresponding metric-performance comparisons to the k-nearest-neighbor graph construction and curvature estimation settings? Given that all three real-world datasets exhibit negative estimated curvature, how strongly do the authors believe the current experiments support the general curvature-alignment hypothesis beyond the effectiveness of hyperbolic representations?

  2. Independent contribution of RCM (clarification and experimental attribution). Can the authors clarify what benefits are specific to the proposed RCM formulation compared with other smooth, bounded, direction-preserving radial mappings? In particular, which existing experiments isolate its geometric advantages from the general effects of feature compression and outlier magnitude limitation?

  3. Numerical interpretation and evaluation protocol (clarification).How should the asymptotic gradient-stability claim be interpreted given the behavior of the radial derivative at large input norms? Could the authors also clarify the scope of the spherical classifier's cut-locus guarantee, the distinction between native and tangent-space AIRM baselines, and the uncertainty of the FPHA results? These clarifications would help assess the robustness of the main conclusions.

3 comments
AuthorsOpenReview ↗

Authors' Response to Reviewer TarY - 1

We thank the reviewer for their positive assessment of our geometric framework, RCM construction, and computational improvements. We address your the concerns regarding curvature alignment, RCM's independent contribution, and numerical guarantees below.

Weaknesses:

W1. Curvature-Alignment Hypothesis:

We agree that more evidence is required for the curvature-alignment hypothesis, as all three datasets are negatively curved and the provided correlation should indeed be interpreted cautiously.

To provide more evidence, we comprehensively expanded the scope of our SPD matrix synthetic experiment (not the outlier regularization one). We explicitly decoupled the radial embedding from the added ambient noise. We kept the signal radius fixed at 1.5 and changed only the ambient noise levels (σ∈[0.1,0.3]\sigma \in [0.1, 0.3]): this avoided confounding factors and allowed us to judge two things at once: the robustness of metrics to added noise (which also addresses a point raised by Reviewer Ynpt) and the impact of curvature on the classification performance of the different classifiers.

Noise Std (σ\sigma)Mean ORCNHMNSMNSM-RCM
0.100+0.05291.7%93.1%91.0%
0.200-0.04988.5%79.2%86.0%
0.300 (High)-0.0584.4%42.3%82.5%

These results empirically demonstrate both central claims of our framework: curvature alignment and RCM's necessity.

  1. Bidirectional Curvature Alignment (Q1): At low noise level, the data retains a spherical structure due to its spherical generation, where both NSM significantly outperforms the NHM (Holm-corrected McNemar’s test p<0.001p < 0.001). Moderate norms avert norm mismatch, yielding similar NSM/NSM-RCM accuracies here. However, as high noise is injected and the geometry shifts negative, NSM collapses. NHM aligns with this negative curvature, significantly outperforming NSM (p<10−10p < 10^{-10}), while retaining a high accuracy.

  2. RCM's Requirement: At σ=0.300\sigma=0.300, NHM significantly outperforms NSM-RCM, which in turn significantly outperforms unconstrained NSM (p<10−10p < 10^{-10}). This demonstrates two things: RCM significantly restores NSM's performance in this noisy setting by preventing norm mismatch, while NHM still maintains the highest overall performance due to best metric alignment.

W2. RCM's Independent Contribution:

Comparing RCM with generic bounded radial mappings is necessary. RCM's distinctive contribution is its smooth, invertible construction, whose image matches the spherical exponential map's injectivity domain. We present theoretical and empirical evidence for RCM on the real FPHA dataset, instead of synthetic experiments.

1. Theoretical Arguments:

L2L_2 normalization lacks origin differentiability, making it unsuitable for our diffeomorphic framework where curvature calculations require higher-order smoothness. A valid bounding metric must guarantee C∞C^\infty smoothness everywhere and proper training. RCM uniquely achieves this balance.

Radial bounding mappings f(r)f(r) applied to vector vv with norm r=∥v∥2r = \|v\|_2:

PropertyL2L_2 NormSimple RationalRational SqrtExponentialTanhProposed RCM
f(r)f(r)11r1+r\frac{r}{1+r}r1+r2\frac{r}{\sqrt{1+r^2}}1−e−r1-e^{-r}tanh⁡(r)\tanh(r)r1+α(r2+ϵ2−ϵ)\frac{r}{1+\alpha(\sqrt{r^2+\epsilon^2}-\epsilon)}
C∞C^\infty Smooth EverywhereNoNoYesNoYesYes (ϵ>0\epsilon > 0)
BijectiveNoYesYesYesYesYes
Radial Gradient Decay (∂f∂r\frac{\partial f}{\partial r})00O(r−2)O(r^{-2})O(r−3)O(r^{-3})O(e−r)O(e^{-r})O(e−2r)O(e^{-2r})O(r−2)O(r^{-2})
Smooth Origin and O(r−2)O(r^{-2}) FlowNoNoNoNoNoYes

Note: RCM parameter conditions α>0\alpha>0, ε>0\varepsilon>0, and αε<1\alpha\varepsilon<1.

2. Empirical Validation:

To empirically validate this theoretical advantage, we ablated RCM against these alternative bounded mappings on the high-variance FPHA dataset (Linear Probe, multiple seeds):

Radial Bounding MappingMean Accuracy ±\pm Std
Proposed RCM (Ours - O(r−2)O(r^{-2}))84.52% ±\pm 0.51%
Rational Sqrt (O(r−3)O(r^{-3}))9.16% ±\pm 1.08%
Exponential (O(e−r)O(e^{-r}))2.78% ±\pm 0.00%
Tanh (O(e−2r)O(e^{-2r}))2.78% ±\pm 0.00%
(Unconstrained NSM baseline)13.28% ±\pm 2.26%

Tanh and Exponential mappings shrink gradients exponentially, inducing catastrophic vanishing. The C∞C^\infty smooth Rational Sqrt mapping decays cubically (O(r−3)O(r^{-3})), which is also too aggressive for high-variance data. The RCM’s O(r−2)O(r^{-2}) polynomial gradient flow and smoothness prove beneficial and unique compared to the presented bounds. Pairwise McNemar’s tests (with Holm-Bonferroni correction) confirm that RCM's superiority over all compared bounding maps is highly statistically significant (p<1.12×10−86p < 1.12 \times 10^{-86} for all pairwise comparisons).

AuthorsOpenReview ↗

Authors' Response to Reviewer TarY - 2

W3. Asymptotic Gradient Stability:

The reviewer is correct: for a monotonic transformation to map an unbounded domain (Rn\mathbb{R}^n) into a bounded radial envelope, its radial derivative must asymptotically approach zero as r→∞r \to \infty. This is a necessary and desirable mathematical property for bounding outliers, not a defect.

However, in deep learning optimization, the critical distinction lies in the rate of decay as inputs reach large, finite magnitudes. Standard bounding mappings, such as Tanh or Exponential, exhibit exponential gradient decay. For high-variance datasets (with large norms), this causes premature gradient vanishing, stalling the network before it can learn (empirically demonstrated above). The Rational Conformal Metric (RCM) achieves the required topological bound using a sustainable polynomial decay rate (O(r−2)O(r^{-2})). We will revise the manuscript to formulate this claim more precisely.

W4. Clarification on the Cut-Locus:

We already acknowledge in the manuscript that cut-locus violation remains a theoretical issue of applying the established RMLR framework to spherical spaces, if the optimizer were to push an anchor to the opposite antipode. However, cross-entropy optimization empirically averts it, as the loss function strongly penalizes antipodal divergence. We also provided another solution in the manuscript to impose a stronger radial bound α=4c/π\alpha=4\sqrt{c}/\pi. In both cases, in our original experiments, we achieved exactly 0.0% cut-locus violations. An enhanced version of the RMLR framework that theoretically prevents the cut locus without reducing the bound radius remains a promising direction for future work.

To further verify this robustness, we tracked the Riemannian distance between all data points and their corresponding class anchors across every epoch of our new multi-seed FPHA ablations as well as the synthetic experiments. Across all runs, the cut-locus violation rate remained exactly 0.00%. We will update the Appendix to explicitly acknowledge this distinction between the theoretical limit and the robust empirical reality driven by cross-entropy dynamics.

Questions:

Q1: Sensitivity of ORC & Curvature Alignment

To address concerns regarding the stability of the Ollivier-Ricci Curvature (ORC) estimation, we conducted a comprehensive sensitivity analysis across graph hyperparameters: neighborhood size (k∈{5,6,7,8,9,10}k \in \{5, 6, 7, 8, 9, 10\}), random walk idleness (α∈{0,0.25,0.5}\alpha \in \{0, 0.25, 0.5\}), and edge weighting. In all cases, the curvature of all 3 datasets remained negative, demonstrating the stability of the curvature calculation. As shown in the representative subset below evaluated at k=5k=5, while the exact magnitude scales with graph idleness, the sign remains negative, and the relative geometric hierarchy reported in our original manuscript remains stable: HDM05 shows the most negative curvature, followed by FPHA and Radar.

Datasetk=5,α=0.0k=5, \alpha=0.0α=0.25\alpha=0.25α=0.50\alpha=0.50
Radar-0.209-0.104-0.0439
FPHA-0.230-0.113-0.0886
HDM05-0.233-0.128-0.1021

Beyond hyperbolic representations, the synthetic noise experiments (detailed in W1 above) provide statistically significant evidence of bidirectional metric alignment (NSM-RCM outperforming NHM in spherical settings and vice versa in hyperbolic settings).

Q2: Independent Contribution of RCM

As detailed theoretically and empirically in Section W2 above, RCM's superiority over generic smooth competitor metrics lies uniquely in its ability to combine C∞C^\infty origin smoothness with a sustainable O(r−2)O(r^{-2}) gradient flow, avoiding the exponential gradient death that stalls other bounded radial mappings. This isolates its geometric advantage beyond general feature compression.

AuthorsOpenReview ↗

Authors' Response to Reviewer TarY - 3

Q3: Numerical Interpretation, Cut-Locus, and Evaluation Protocols

We have provided clarifications regarding asymptotic gradient stability (Section W3) and the cut-locus guarantee (Section W4) and shall include them in the camera-ready version.

Regarding the distinction between native and tangent-space AIRM baselines: both are established in the literature. In the native case, we use the native AIRM RMLR, which directly draws hyperplanes on the SPD manifold using the AIRM metric, making it computationally intractable for higher-dimensional data. Tangent Space AIRM (TS-AIRM) is a common way to implement the classifier that uses the AIRM metric but draws hyperplanes in the metric's tangent space using a Euclidean network, which greatly reduces complexity. We implemented TS-AIRM for HDM05 strictly due to its extremely high computational burden. This highlights the limitation of the native AIRM metric's utility in real-world high-dimensional settings, as demonstrated in our manuscript.

We shall explicitly clarify this distinction and expand the descriptions of both classifiers in the final text.

Uncertainty of FPHA results: As requested, we conducted multi-seed evaluations for our rebuttal ablations on FPHA. The empirical variances across the metrics are low. This shows that the point estimates reported in our manuscript are stable. We aim to include the variance across all FPHA metric experiments in the camera-ready version, as rerunning all architectures with multiple seeds is computationally difficult during the rebuttal period.

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

Comments on the paper

3 comments

We thank the reviewers and the Area Chair for the discussion. The central question is whether choosing a target geometry adds anything once one already has a good SPD representation. The learned-metric comparison answers this empirically, and the spherical construction explains precisely what can be guaranteed before training.

1. A learned SPD representation and a curved target make distinct contributions. Let ψ\psi be a smooth invertible SPD coordinate map. The metric ψ∗gEucl\psi^*g_{\mathrm{Eucl}} has zero sectional curvature for every such ψ\psi. On the domains established in Theorems 2 and 5, our pullbacks have sectional curvature −c-c or +c+c. In dimension at least two, invariance of sectional curvature rules out even a local isometry between these flat and curved metrics. Whether this geometric distinction improves classification can then be tested by comparing each learned base with its own standalone classifier.

In the FPHA linear-probe ablation, switching from flat classification to NHM raises accuracy for every tested base family. LEM goes from 83.13% to 85.57%, SSPM from 85.22% to 87.30%, ALEM from 84.00% to 89.22%, and CSPM from 58.96% to 62.43%. The comparison also works in the other direction. Keeping NHM as the target and changing the base from LEM to ALEM raises accuracy from 85.57% to 89.22%. Thus the best result combines a learned base with a curved target. The spherical results delimit this conclusion particularly well. NSM-RCM improves ALEM from 84.00% to 87.13%, ties standalone SSPM at 85.22%, and lowers CSPM from 58.96% to 47.65%. The choice of target continues to matter after the base family has been chosen.

2. RCM resolves an exact loss of distinguishability, and its rational tail has a calculable effect. Consider St=diag(et,1,…,1)S_t=\mathrm{diag}(e^t,1,\ldots,1). Its log-Euclidean vector is te1t e_1. Under the unrestricted round-sphere exponential in Eq. (1), t=0t=0 and t=π/ct=\pi/\sqrt c both map to the stereographic origin. Every downstream classifier therefore gives identical scores to these two distinct SPD inputs. This obstruction exists for every fixed c>0c>0, independently of the dataset or optimizer. With α=2c/π\alpha=2\sqrt c/\pi, the RCM radius instead increases strictly from zero to the limiting radius R=1/αR=1/\alpha. Its image is exactly the open injectivity ball. The resulting spherical representation distinguishes every pair of finite SPD inputs (Theorem 5).

The gradient question also admits a direct calculation. Write f(r)=r/[1+α(r2+ϵ2−ϵ)]f(r)=r/[1+\alpha(\sqrt{r^2+\epsilon^2}-\epsilon)] and A=(1−αϵ)/α2>0A=(1-\alpha\epsilon)/\alpha^2>0 under the paper's parameter condition. Then R−f(r)∼A/rR-f(r)\sim A/r and f′(r)∼A/r2f'(r)\sim A/r^2. The reviewer is correct that the isolated RCM radial derivative vanishes. For the composed map, Eq. (1) gives the stereographic output radius h(r)=tan⁡(c f(r))/ch(r)=\tan(\sqrt c\,f(r))/\sqrt c. Using cot⁡u∼1/u\cot u\sim1/u and h′(r)=sec⁡2(c f(r))f′(r)h'(r)=\sec^2(\sqrt c\,f(r))f'(r) yields lim⁡r→∞h(r)r=lim⁡r→∞h′(r)=1cA.\lim_{r\to\infty}\frac{h(r)}r=\lim_{r\to\infty}h'(r)=\frac1{cA}. Thus both tangential and radial derivatives from log-Euclidean to stereographic coordinates approach a finite positive constant. This gives a specific reason for the rational tail. The limit describes the coordinate map in exact arithmetic. Class-anchor logarithms have a separate antipodal singularity. Constraining both anchors and inputs to the same open hemisphere excludes antipodal pairs by the triangle inequality.

On the multi-seed FPHA radial-map ablation, RCM achieved 84.52% ± 0.51%, rational square-root 9.16% ± 1.08%, exponential and tanh each 2.78% ± 0.00%, and unconstrained NSM 13.28% ± 2.26%. Bounded maps therefore behaved differently under the tested protocol. The derivative calculation motivates the design independently of attributing those accuracy gaps to a single cause.

3. The empirical conclusion survives a narrower reading of curvature alignment. Across all nine dataset–architecture comparisons in Tables 1–3, NHM exceeds LEM and NSM-RCM exceeds unconstrained NSM in reported accuracy. At fixed synthetic signal radius, the ranking also reverses. At noise 0.100, NSM reaches 93.1%, NHM 91.7%, and NSM-RCM 91.0%. At noise 0.300, the respective values are 42.3%, 84.4%, and 82.5%. On real FPHA data, with retraining at each corruption level, the decline from noise 0.1 to 0.4 is 2.27 percentage points for NHM, 4.52 for NSM-RCM, and 9.56 for LEM. The three real-data mean graph curvatures remained negative across the tested graph settings. We interpret their correlation of approximately 0.993 with the mean accuracy gap as descriptive. Selecting a geometry for a new or locally heterogeneous dataset remains an open model-selection question.

Finally, on HDM05, NHM took 38.29 seconds per epoch, NSM-RCM 38.30, LEM 38.24, and native AIRM RMLR 750.70 (Table 4, NVIDIA T4, batch size 32). The HDM05 accuracy table uses tangent-space AIRM.

We therefore request the reviewers to weigh together the explicit spherical-domain guarantee, the gains over learned flat baselines, and the measured cost of obtaining those gains.

An exact classification example with the SPD features held fixed

We close with an example that answers a question raised in the discussion. Can changing the target geometry alter what the classifier in Eq. (8) can represent, even when it receives exactly the same SPD features? On the following class-balanced problem, the optimal flat probe attains exactly 75% accuracy, while a single hyperbolic decision boundary attains 100%. Both values follow analytically.

Consider the three commuting matrices St=diag⁡(e,et)S_t=\operatorname{diag}(e,e^t) for t∈{−2,0,2}t\in\{-2,0,2\}. Assign positive label to S0S_0 with probability 1/21/2 and negative label to each of S−2,S2S_{-2},S_2 with probability 1/41/4. In the two diagonal log-Euclidean coordinates, the common feature map gives vt=(1,t)v_t=(1,t).

Flat target. For binary Euclidean RMLR, the difference between the two class logits is affine in the input, even when the classes have separate anchors and normals. Write this difference as g(v)g(v). Since v0=(v−2+v2)/2v_0=(v_{-2}+v_2)/2, we have g(v0)=12g(v−2)+12g(v2).g(v_0)=\tfrac12 g(v_{-2})+\tfrac12 g(v_2). An affine probe therefore cannot assign the positive label to the midpoint and the negative label to both endpoints. It must incur error of at least 1/41/4. The rule that predicts positive when t>−1t>-1 attains this bound, so 75% is the exact optimum. Any affine reparameterization of the common features preserves the midpoint identity.

Hyperbolic target. Take curvature −1-1. In the paper’s normalization, the origin exponential followed by conversion to Klein coordinates gives K(v)=tanh⁡(2∥v∥) v/∥v∥K(v)=\tanh(2\|v\|)\,v/\|v\|. Examine just the first coordinate:

K_1(v_{-2})=K_1(v_2)=\frac{\tanh(2\sqrt5)}{\sqrt5}<\frac1{\sqrt5}<\tfrac12.$$ The hyperbolic hyperplane $K_1=1/2$ therefore separates all three inputs strictly. This separator is realized by Eq. (8) itself. To verify this, write $K_1=X_1/X_0$ in unit hyperboloid coordinates. Choose a shared class anchor $P$ on $X_1=X_0/2$ and opposite tangent normals $\pm A$, where $A=(1/2,1,0,\ldots)$. Since $\langle P,A\rangle_L=0$, the positive-class logit is $$\langle\log_P X,A\rangle_L =\frac{d}{\sinh d}\,(X_1-X_0/2),\qquad d=d(P,X).$$ The factor is positive, and the other logit has the opposite sign. Hence the paper’s binary hyperbolic head attains 100% accuracy on the same features. **The gap persists under perturbation.** Add an independent perturbation $U$ to the log coordinates, drawn from the same distribution for all three inputs. Conditional on any $U=u$, the translated positive point remains the midpoint of the translated negative pair. The flat upper bound of 75% therefore survives averaging over $U$. The hyperbolic inequalities and the attaining flat threshold above are all strict at the three inputs. By continuity, both retain their predictions for every perturbation in a sufficiently small ball. Thus the exact 75% versus 100% gap also holds for a distribution with a smooth density in the full log-coordinate space. This example isolates what choosing the target contributes. The matrices, feature map, dimension, and labels are held fixed. Changing the target geometry allows the classifier to realize a labeling that every flat linear probe on those features must miss. The argument is independent of curvature estimation, optimization, and sampling variability. The within-base FPHA comparisons address the complementary empirical question of whether this additional modeling choice helps on observed data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.