Adaptive Log-Euclidean Metrics

• Adaptive Log-Euclidean Metrics (ALEMs) are Riemannian metrics defined on SPD matrices that incorporate tunable parameters to adapt the geometry to specific data characteristics.
• They extend the classical Log-Euclidean metric by enabling learnable log bases or Mahalanobis formulations, thereby improving discrimination in tasks like metric learning and deep neural networks.
• Empirical results in applications such as face matching, clustering, and EEG classification demonstrate significant accuracy improvements over fixed metric approaches.

Adaptive Log-Euclidean Metrics (ALEMs) are a class of Riemannian metrics and associated distances defined on the manifold of symmetric positive definite (SPD) matrices. They generalize the standard Log-Euclidean metric by introducing tunable or learnable parameters into the definition of the matrix logarithm or its associated pullback metric, thereby allowing the geometry to adapt to data-driven or architectural requirements. Adaptive Log-Euclidean Metrics have been developed in several mathematically and algorithmically distinct frameworks, but share the principle that the classical Log-Euclidean metric (LEM) can be generalized to a parametric or learned family, retaining mathematical structure while enabling improved discrimination and adaptability in applications such as metric learning and deep learning on SPD-valued data.

1. Mathematical Foundations of Log-Euclidean Metrics

The manifold $\mathrm{SPD}_n$ of real $n\times n$ symmetric positive definite matrices naturally carries several Riemannian metrics. The Log-Euclidean metric (LEM), as introduced by Arsigny et al., is defined by the pullback via the principal matrix logarithm: $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$ LEM equips $\mathrm{SPD}_n$ with a flat Riemannian metric in which geodesic distance is

$$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$

and the corresponding Fréchet mean, exponential and logarithm maps, and parallel transport all admit closed-form expressions. This metric possesses affine and similarity invariance and makes $\mathrm{SPD}_n$ an abelian Lie group under the operation $X\odot Y = \exp (\log X + \log Y)$ (Chen et al., 2023).

2. Motivations and Principles of Adaptivity

Despite the mathematical simplicity and efficiency of fixed Riemannian metrics such as LEM or the affine-invariant metric (AIM), fixed metric choices may be suboptimal for learning tasks:

• Deep SPD-valued neural networks frequently employ a hard-wired metric, even as feature statistics vary across layers and tasks.
• In classification or clustering, the directions of maximal intra-class variance and inter-class separation are typically dataset- or task-specific.
• The standard LEM (i.e., natural log base $e$ applied uniformly to all eigenvalues) treats all spectral directions equally, which may underutilize discriminative information encoded in the geometry of the data distribution.

Adaptive Log-Euclidean Metrics address these issues by introducing learnable or tunable parameters (either in the “logarithm base” or via a Mahalanobis structure in the log domain), enabling the metric to conform to the requirements of the task and the observed data (Vemulapalli et al., 2015, Chen et al., 2023, Yger et al., 2015).

3. Variants and Parameterizations of ALEMs

3.1. Pullback-based ALEMs with Learnable Log Bases

A major recent advance is the introduction of per-eigenvalue base parameters. Let $\alpha = (a_1, \ldots, a_n)$ with $a_i>0,\, a_i \neq 1$. The adaptive matrix logarithm is defined as

$n\times n$0

where $n\times n$1 and $n\times n$2 is the eigendecomposition of $n\times n$3.

The Adaptive Log-Euclidean Metric (ALEM) associated to parameters $n\times n$4 is

$n\times n$5

This parameterization retains all key algebraic properties of LEM (flatness, Lie-group bi-invariance, closed-form mean, similarity-invariance), but allows the geometry to be modulated by learning $n\times n$6 during end-to-end training. Three equivalent parameterizations (RELU, MUL, DIV) are used in practice (e.g. $n\times n$7, $n\times n$8, or $n\times n$9), and gradients are computed via spectral matrix-backpropagation (Chen et al., 2023).

3.2. Mahalanobis-type Adaptive Log-Euclidean Metrics

Alternatively, one can treat the log-embedded representation $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$0 ($$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$1) and learn a positive definite Mahalanobis matrix $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$2: $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$3 $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$4 is typically learned via Information-Theoretic Metric Learning (ITML), which seeks an $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$5 close to a prior $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$6 (in LogDet divergence), subject to supervised (must-link/cannot-link) constraints. The advantage is that $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$7 adapts the metric to intra-class and inter-class structure and yields a data-driven Riemannian geodesic (Vemulapalli et al., 2015).

3.3. Congruence-centered (Reference-based) ALEMs

Another approach parameterizes the log-Euclidean metric via a congruence transformation about a reference $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$8: $$\mathrm{mln}_e(S) = U \ln(\Sigma) U^\top,\ \text{where}\ S=U\Sigma U^\top.$$9 Here, $\mathrm{SPD}_n$0 is learned by maximizing a supervised kernel-target alignment criterion, typically via Riemannian gradient methods in the SDP cone (Yger et al., 2015).

3.4. One-parameter Families: Alpha-Procrustes Metrics

The $\mathrm{SPD}_n$1-Procrustes family provides a smooth interpolation between the Log-Euclidean metric ($\mathrm{SPD}_n$2) and the Wasserstein/Bures metric ($\mathrm{SPD}_n$3). The distance is

$\mathrm{SPD}_n$4

As $\mathrm{SPD}_n$5, $\mathrm{SPD}_n$6, i.e., the Log-Euclidean metric. This formulation extends to infinite-dimensional Hilbert spaces and RKHS covariance operators (Quang, 2019).

4. Geometric and Algebraic Properties

All ALEMs based on pullback constructions or $\mathrm{SPD}_n$7-parametrizations inherit structural properties from the Log-Euclidean framework:

• Bi-invariance: The metric is invariant under congruence transformations.
• Flatness: Sectional curvature is zero for LEM and remains controlled for $\mathrm{SPD}_n$8 in a neighborhood of zero; the Wasserstein metric introduces non-negative curvature.
• Closed-form mean (Fréchet mean): For weights $\mathrm{SPD}_n$9,

$$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$0

where $$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$1 is the inverse of $$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$2.

• Similarity invariance: $$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$3 is invariant under congruence by orthogonal transformations and global scaling.
• Riemannian/Euclidean consistency and EMI: The affine-invariant metric $$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$4 is always greater than or equal to the parameterized log-Euclidean distance, with equality for geodesics through the parameter point (Yger et al., 2015).

5. Optimization, Learning Algorithms, and Computational Considerations

5.1. Learning Parameterized Metrics

• ITML in log-domain: For Mahalanobis-type ALEMs, learning proceeds by solving a LogDet-regularized constrained optimization to satisfy similarity/dissimilarity constraints (Vemulapalli et al., 2015).
• Kernel alignment and gradient methods: For reference-centered parameterizations, kernel-target alignment is optimized over $$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$5 using geodesic (Riemannian) gradient ascent, with each update following the exponential map in $$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$6 (Yger et al., 2015).
• End-to-end differentiation: For per-eigenvalue ALEMs (pullback-learning), gradients are propagated with respect to eigenvalue log bases or their equivalent parameterizations during standard deep network training. Riemannian SGD can be employed for updating parameters on SPD manifolds.

5.2. Computational Aspects

• Matrix logarithms: Each application requires an $$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$7 eigen-decomposition for $$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$8 matrices.
• Mahalanobis learning: Each ITML iteration is $$d_{\mathrm{LEM}}(X,Y) = \| \mathrm{mln}_e(X) - \mathrm{mln}_e(Y) \|_F,$$9 ($\mathrm{SPD}_n$0).
• Riemannian gradient methods: Per-iteration cost is $\mathrm{SPD}_n$1.
• Parameter storage: For per-eigenvalue ALEMs, only an $\mathrm{SPD}_n$2-vector (or $\mathrm{SPD}_n$3 diagonal matrix) of bases is required.
• No explicit global asymptotic complexity or convergence rates are provided in the literature; per-iteration costs are dominated by matrix functions and Gram matrix computations.

6. Applications and Empirical Results

ALEMs have been validated on a variety of classification, clustering, and deep learning tasks involving SPD descriptors:

• Face matching (LFW): Adaptive log-Euclidean (ITML in log-domain) yields 69.4% accuracy vs. best fixed distance 61.6%, and outperforms log-Frobenius and affine-invariant metrics by ~9% (Vemulapalli et al., 2015).
• Semi-supervised clustering (ETH80): Adaptive log-Euclidean metric with ITML achieves 73.8% accuracy versus 55.7% for fixed metrics, a gain of ~18% (Vemulapalli et al., 2015).
• EEG and texture classification: Learned reference-point ALEMs improve mean accuracy on EEG data from 70% to 72% and on textures from 78.5% to 80.2% (Yger et al., 2015).
• Deep SPD networks: Integration of learned log base ALEMs as adaptive LogEig layers in SPDNet yields improvements of up to 2.4 percentage points over fixed-log architectures, and is consistently beneficial across skeleton-action, hand-gesture, and deep-feature covariance datasets (Chen et al., 2023).
• Ablations: Fixed, non-learned log bases (e.g., log-base-2) fail to yield consistent improvements; learned parameterizations alone yield robust performance gains and adaptation to data geometry.

7. Connections, Extensions, and Theoretical Perspectives

ALEMs unify a spectrum of Riemannian geometries, interpolating via the Alpha-Procrustes family between Log-Euclidean and Wasserstein frameworks, and accommodating both finite- and infinite-dimensional settings including kernelized (RKHS) covariance operators (Quang, 2019). Geometric structure (e.g., curvature) is controlled by adaptive parameters, and the learned metrics remain true geodesic distances. In the context of kernel learning, ALEMs induce tunable Riemannian kernels applicable to complex supervised learning objectives (Yger et al., 2015).

Key future directions include mini-batch and stochastic extensions to improve scalability, joint learning of kernel mixtures, low-rank metric parameterizations for dimensionality reduction, and re-derivation of Riemannian network primitives under adaptive metrics. These efforts aim to further exploit the ability of ALEMs to tailor data geometry for improved learning efficacy while retaining the essential mathematical tractability and structure of the Log-Euclidean paradigm.