acceptodds
Under review as a conference paper at ICLR 2027

A Riemannian Framework for Normalized SPD Matrix Learning

14 views

Abstract

Symmetric positive definite matrices are widely used to represent covariance and second-order statistics in vision, neuroscience, and time-series learning. Existing methods often rely on the Log-Euclidean metric, which applies the scalar logarithm to eigenvalues and maps SPD matrices to a Euclidean space. However, many practical SPD representations are normalized or regularized so that their spectra lie in a bounded interval, typically . In this setting, the logarithm is only one possible coordinate choice, and alternative spectral geometries remain largely unexplored.We propose a unified framework for Normalized SPD learning based on diffeomorphic spectral coordinates. Given any scalar diffeomorphism , we lift it to SPD matrices by applying it to eigenvalues, obtaining a global coordinate map from bounded SPD matrices to symmetric matrices. This construction induces a pullback Euclidean geometry with explicit distances, geodesics, and Riemannian operators. We instantiate the framework with logit, algebraic square-root, rational, tangent, and Gumbel coordinates, which differ in symmetry, boundary behavior, and numerical conditioning.The proposed coordinates can be plugged into existing SPD architectures without changing their overall design. We evaluate them in SPD residual networks, diffeomorphic generative modeling, covariance pooling for visual recognition, and large-matrix approximation settings. Experiments across action recognition, EEG/fMRI analysis, fine-grained recognition, and ImageNet-scale covariance models show that spectral coordinate choice is an important design dimension for bounded SPD learning.

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

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9oSLOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The authors introduce a family of Deformed-Euclidean Metrics (DEMs) meant to be applied to normalized SPD (NSPD) matrices, where standard approaches such as the log-Euclidean metric fail to cover the whole real line and leave the upper boundary unhandled. By introducing a scalar diffeomorphism ff that maps the bounded NSPD eigenvalues to the entire real line, all presented DEMs have global geometry (flat and geodesically complete), bi-invariance (distances remains unchanged after group operations), and orthogonal and scale invariance (distances are robust to rotations and scaling). New geometric operators and eigendecomposition-free approximations are presented to make the new map computationally feasible and useful. Experiments on synthetic data, visual recognition, and neuroimaging show strong downstream utility and distributional fidelity against SPD-based diffeomorphism baselines.

Critical strengths
  • Strong Methodological Novelty: The authors present a non-trivial but well-founded and intuitive approach to address limitations in diffeomorphisms in the SPD cone applied to NSPD matrices, which are commonly used in neuroimaging and other applications. This enables an improved manifold for downstream machine learning tasks with convenient implementation properties.

  • Paper Clarity and Experiments: The method and experimental design are clear, with enough details presented for reproducibility. Results show reasonable gains against baselines, empirically showing the advantages of using the presented NSPD metrics across a wide range of popular datasets with diverse applications.

Critical weaknesses
  • Unclear Motivation in Writing: While the authors do a good job at explaining the method, explicitly mentioning the limitations of using LEM or other metrics on NSPD matrices (failure to cover the whole real line and unhandled upper bound) would be helpful for the reader to understand the motivation behind the DEM family.

  • Inconsistency of Best-Performing DEM: Results can appear inconsistent across datasets and metrics. While the five presented DEMs are all of valid utility, the best performance within these appears to be arbitrary, limiting the applicability of the method. More details on the leading DEM between protocols would be useful.

Questions for authors
  1. Could you add discussion on selecting an appropriate DEM? If I were to implement this myself on a new dataset, there is no clear guideline on which DEM I should use.
Reviewer_hq4hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the Riemannian geometry of SPD matrices. The motivation for this work is as follows. In many applications that use SPD matrices, the SPD matrices are normalized, but later processing still uses geometry defined on the full SPD cone. This motivates the authors to study the space of normalized SPD matrices, which is defined as the restricted space of SPD matrices with eigenvalues between zero and one. They construct a family of Riemannian geometries on this space by pulling back Euclidean geometry through several smooth spectral maps that restrict the eigenvalues of SPD matrices to be between zero and one. These geometries have explicit formulas for distances, geodesics, and means. The paper also develops fast approximations and evaluates the geometries induced by these maps on visual recognition, Riemannian classification, and brain connectivity generation.

Critical strengths
  • The motivation is clear and relevant. SPD matrices are often normalized, so studying geometry on the resulting restricted domain is useful.
  • The idea is simple and the derivations are clear and elegant. One general construction gives several Riemannian geometries on NSPD.
  • The proposed geometries have clear guarantees and explicit formulas for distances, geodesics, and means.
  • The presentation is well organized.
  • The paper addresses computation through fast approximations, operation counts, and runtime tests.
  • The experiments cover several applications and show consistent gains in visual recognition.
Critical weaknesses
  • The presented empirical improvements are not consistent across applications.
  • The best spectral map changes across datasets, tasks, and evaluation measures. The paper gives no clear rule for choosing a map.
  • The experiments do not include explicit comparisons with some commonly-used Riemmanian metrics on the SPD manifold. Particularly, the affine-invariant metric (which is arguably the most common) or log-determinant metric (which has shown SOTA performance in recent papers).
  • The geometric guarantees apply to the full NSPD domain. Normalized inputs satisfy an extra constraint, such as unit trace or unit Frobenius norm. The proposed geodesics and means stay in NSPD but need not preserve this constraint.
  • Most theoretical guarantees follow from the standard pullback construction. The theoretical novelty is therefore limited.
  • The fast approximations are tested empirically, but they do not preserve the exact maps’ geometric guarantees.
Questions for authors
  • How should a user choose the spectral map? Is it possible to come up with a rule based on validation data?
  • How do the methods compare with affine-invariant and log-determinant metrics?
  • Does leaving the unit-trace or unit-norm subset affect the results?
  • Some maps seem to improve generation fidelity but reduce downstream performance. Is there an explanation for these results?
  • Note the extra characters in the title
Reviewer_bWC3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a family of Deformed-Euclidean Metrics (DEMs) on the normalized SPD (NSPD) domain by pulling back the Euclidean metric through scalar spectral diffeomorphisms. It introduces five specific deformations (i.e., Logit, IAS, Rational, Cauchy, and Gumbel) and derives closed-form expressions for the corresponding distances, geodesics, exponential and logarithmic maps, parallel transport, and weighted Fréchet means. It also employs polynomial (i.e., Taylor, Chebyshev, Legendre, and Laguerre) and rational (Padé) approximations to enable forward computation without eigendecomposition. The experiments cover fine-grained recognition, ImageNet-1K, classification using Riemannian ResNet with BatchNorm, and fMRI connectome generation using the DiffeoCFM pipeline.

Critical strengths
  1. The paper evaluates the proposed method not only on fine-grained visual recognition and ImageNet-1K, but also on Riemannian neural network classification and fMRI functional connectivity matrix generation. The results across different tasks provide a preliminary assessment of the applicability of the proposed framework, rather than demonstrating its performance on a single dataset alone.

  2. The authors investigate several approximation schemes, including Chebyshev, Legendre, Taylor, and Padé approximations, and report approximation errors and runtime measurements. By enabling forward transformations without explicit eigendecomposition, the proposed approach has potential practical engineering value, particularly for large-scale matrix computations.

Critical weaknesses

#Novelty

  1. NSPD:The paper states:“In covariance pooling and iterative whitening, trace or Frobenius normalization places the spectrum in the convergence range of fast matrix-function algorithms (Li et al., 2018; Huang et al., 2019; Song et al., 2022b; Rahman et al., 2026). ” However, this statement does not fully reflect the normalization procedures used in the cited works. For iSQRT-COV (Li et al., 2018), although normalization is performed before the Newton–Schulz (NS) iterations, the scaling factor is subsequently restored to recover the appropriate magnitude of the matrix function, as shown in the implementation: y = YZY * torch.sqrt(normA).view(batchSize, 1, 1).expand_as(x) (see https://github.com/jiangtaoxie/fast-MPN-COV/blob/master/src/representation/MPNCOV.py ). Moreover, Equation (13) in (Song et al. 2022b https://arxiv.org/pdf/2201.08663) also shows that a compensation factor is applied after normalization. Furthermore, Rahman et al. (2026) employ normalization by the mean eigenvalue, as specified in Equation (2), rather than simply using the trace or Frobenius normalization described in the paper. Since Rahman et al. (2026) use mean-eigenvalue normalization, the eigenvalues of the normalized matrices are not necessarily confined to the interval (0,1). Together with the compensation mechanisms used in iSQRT-COV (Li et al., 2018) and Song et al. (2022b), this raises concerns about the paper’s characterization of the normalization procedures in prior work. Among the four non-EEG methods discussed, three do not necessarily produce matrices whose spectra lie strictly within (0,1).

  2. This raises a fundamental question about the motivation for defining the NSPD domain: Is the proposed NSPD formulation sufficiently justified, given that the normalization schemes used in the cited literature do not necessarily constrain the eigenvalues to (0,1)? More importantly, the Introduction does not clearly explain what specific problems NSPD solves relative to the conventional SPD domain, nor does it demonstrate why existing SPD geometries are inadequate for normalized SPD representations.

  3. DEM: Compared with the pullback-Euclidean framework (Chen et al. 2024b), DEM appears to differ in three main aspects: (1) the domain is specifically fixed to NSPD; (2) the representation maps are restricted to scalar spectral diffeomorphisms; and (3) the framework is instantiated using five specific scalar deformations—Logit, IAS, Rational, Cauchy, and Gumbel—together with corresponding polynomial and rational approximations. However, these differences appear to be primarily domain-specific design choices and implementation extensions, rather than fundamentally new geometric principles. Consequently, the methodological novelty of DEM appears limited unless the authors can demonstrate substantial theoretical advances or clear practical advantages over existing pullback-Euclidean metrics, particularly LEM and ALEM.

#Fairness of Experimental Comparisons

  1. Fine-grained dataset fine-tuning experiments (Appendix D.1). The experimental details regarding pretraining and subsequent fine-tuning on fine-grained datasets are insufficiently described. In particular, iSQRT-COV also reports results on Cars and Aircraft, but its performance is substantially lower than that reported in this paper. Without further details, it is difficult to determine the reasons for this performance gap, raising concerns about the fairness of the comparison.

  2. Experimental settings, computational complexity, and numerical precision on ImageNet-1K. The experimental settings on ImageNet-1K differ from those used in iSQRT-COV, particularly in terms of the learning rate schedule and the number of training epochs. Consequently, it is unclear whether the reported improvements are attributable to the proposed method itself or to the improved training strategy. If possible, the authors are encouraged to report the results of iSQRT-COV under the same training strategy adopted in this paper to enable a fairer comparison. Furthermore, it would be helpful to report the computational complexity of the methods evaluated on ImageNet-1K, including the number of parameters, FLOPs, and throughput. The paper states that FP64 precision is used, but does not provide a comparison of GPU memory consumption. The authors are encouraged to report the memory requirements of the proposed method relative to the baselines. If possible, results using FP32 or FP16 precision under the best-performing configuration would also be valuable for assessing the method's practical efficiency.

  3. Robustness of the proposed method. In Table 3, the Riemannian network GEM performs particularly poorly on FPHA, while achieving reasonably good results on Cora. In addition, Table 4 shows that, for REM and CEM on the ANDI dataset, the (\beta) values are approximately 0.17 and the (H) values are approximately 0.23. These results raise concerns about the robustness and generalizability of the proposed method across different datasets and settings. Further analysis would help clarify the reasons for these performance variations and establish the robustness of the proposed approach.

  4. Comparison of the ALEM. ALEM (Chen et al. 2024b) can learn the optimal deformation, whereas the proposed method uses a fixed deformation. One would therefore expect the proposed method to offer compensating advantages in terms of computational efficiency or numerical stability. However, the paper does not include ALEM as a baseline, making it difficult to assess the relative merits and trade-offs of the two approaches.

Questions for authors
  1. Figures 4 and 5: The use of images to present the results on fine-grained datasets and ImageNet-1K makes the results less intuitive to interpret and compare. The authors are encouraged to consider using tables or more informative visualizations to improve clarity and facilitate quantitative comparisons.

  2. Possible formatting issues in the appendix: Some text appears to contain garbled characters or formatting errors. For example, an issue can be found at the end of the line on page 13, line 701. The authors should carefully proofread the appendix and correct any such issues.

  3. Title typo: There appears to be an extra “s” at the end of the title. The authors should check and correct this typographical error.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.