Log-Euclidean Riemannian Framework

• The Log-Euclidean Riemannian framework is a geometric method that maps symmetric positive definite matrices into a Euclidean space using the matrix logarithm.
• It simplifies intrinsic operations like geodesic distance and Fréchet means by converting nonlinear manifold tasks into linear Euclidean computations.
• The framework offers practical applications in areas such as gait analysis, image-set coding, and neural network classifiers while addressing challenges in correlation and operator-valued settings.

The Log-Euclidean Riemannian framework is a family of geometric constructions in which nonlinear matrix manifolds—most prominently the manifold of symmetric positive definite matrices and, in later extensions, the manifold of full-rank correlation matrices—are mapped by a logarithmic diffeomorphism into a Euclidean linear space, so that distances, geodesics, means, regression, and learning can be carried out in logarithmic coordinates and then transported back to the original manifold by the inverse map (Vemulapalli et al., 2015, Bisson et al., 17 Nov 2025). In the SPD setting, the canonical choice is the matrix logarithm, which converts manifold-valued objects such as covariance descriptors, diffusion operators, and connectivity matrices into symmetric matrices while preserving a rigorous Riemannian interpretation of the resulting computations (Chen et al., 2019).

1. Geometric definition

For $n\times n$ SPD matrices, the basic state space is

$$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$

or equivalently $$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$ in alternative notation (Vemulapalli et al., 2015, Utpala et al., 2022). Multiple papers emphasize that this set is not a Euclidean vector space: ordinary linear operations do not preserve positive definiteness, and Euclidean treatment ignores the intrinsic geometry of covariance-like data (Vemulapalli et al., 2015, Chen et al., 2019).

The classical Log-Euclidean construction uses the principal matrix logarithm

$\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$

with inverse $\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$, where $\mathcal{S}^n$ is the vector space of symmetric matrices (Vemulapalli et al., 2015). In eigencoordinates, if $X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$, then

$$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$

(Chen et al., 2019).

A central formulation treats the geometry as a pullback of the Euclidean metric through a diffeomorphism $\phi$. In the SPD case, $\phi=\log$ gives a metric for which the geodesic distance is

$$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$0

(Vemulapalli et al., 2015). In the broader theory of log-Euclidean Lie groups, if $$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$1 is globally diffeomorphic to a finite-dimensional vector space $$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$2, then

$$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$3

defines a commutative Lie group structure, and the pullback metric

$$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$4

is flat and bi-invariant (Bisson et al., 17 Nov 2025). Geodesics are straight lines in logarithmic coordinates,

$$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$5

and the distance is $$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$6 (Bisson et al., 17 Nov 2025). For SPD matrices, this yields the familiar group law

$$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$7

(Vemulapalli et al., 2015).

2. Core operations on SPD manifolds

The framework’s computational attractiveness follows from the fact that many intrinsic operations become ordinary Euclidean operations after log-mapping. The log-Euclidean distance is

$$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$8

and, with the vectorization

$$\mathcal{S}_{++}^n=\{P\in\mathbb{R}^{n\times n}\mid P=P^\top,\; x^\top P x>0\ \forall x\neq 0\},$$9

one obtains the isometric expression

$$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$0

(Chen et al., 2019, Utpala et al., 2022).

The Fréchet mean under the log-Euclidean metric has a closed form: $$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$1 (Utpala et al., 2022). This closed-form averaging is one of the recurrent reasons the framework is chosen over geometries that require iterative manifold optimization.

Several application pipelines use SPD covariance descriptors as the basic geometric object. In gait analysis, raw skeleton pose sequences are segmented into gait cycles, each cycle is arranged as $$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$2, and the sample covariance $$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$3 is used as the SPD descriptor; after eigendecomposition $$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$4, the log-map is written

$$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$5

and the final tangent-space fingerprint is

$$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$6

(Bůžek, 9 Dec 2025). The explicit transformation is

$$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$7

(Bůžek, 9 Dec 2025).

In image-set coding, covariance descriptors are likewise treated as points on an SPD manifold, then processed in the log domain to define kernels and centralized descriptors (Chen et al., 2019). In action recognition, an SPD covariance matrix $$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$8 is embedded by

$$\mathrm{SPD}(k)=\{X\in\mathbb{R}^{k\times k}\mid X^T=X,\; u^TXu>0\ \forall u\neq 0\}$$9

with off-diagonal entries scaled by $\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$0 so that Frobenius geometry is preserved after vectorization (Faraki et al., 2014).

3. Generalizations: correlation matrices, operator-valued settings, and unifying theories

A major extension replaces the SPD cone by the manifold of full-rank correlation matrices,

$\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$1

whose elements are symmetric positive definite matrices with unit diagonal (Bisson et al., 21 Jul 2025). Here ordinary Euclidean regression is especially problematic, because linear combinations of correlation matrices are generally not correlation matrices (Bisson et al., 21 Jul 2025). Two logarithmic diffeomorphisms are emphasized.

The first is the off-log map,

$\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$2

where $\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$3 is the vector space of symmetric hollow matrices (Bisson et al., 21 Jul 2025). Its inverse is

$\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$4

with $\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$5 the unique diagonal correction ensuring that the output lies in $\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$6 (Bisson et al., 21 Jul 2025). The second is the log-scaling map,

$\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$7

which lands in $\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$8, the space of symmetric matrices with zero row sum, and has inverse

$\log:\mathcal{S}_{++}^n\to \mathcal{S}^n,$9

(Bisson et al., 21 Jul 2025). In both cases geodesics are straight lines in transformed coordinates.

The 2025 theory of log-Euclidean Lie groups extends this picture beyond isolated matrix examples. It proves an isomorphic isometry theorem: any two log-Euclidean Lie groups of the same dimension, equipped with Frobenius-induced log-Euclidean metrics, are isometrically isomorphic as Riemannian manifolds (Bisson et al., 17 Nov 2025). In particular, $\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$0, $\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$1, and $\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$2 have matching dimensions, and explicit isometries are constructed between $\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$3 with the standard log-Euclidean metric and $\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$4 with off-log or log-scaling metrics (Bisson et al., 17 Nov 2025).

A separate unification appears in the Alpha Procrustes family of distances. For SPD matrices $\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$5, the $\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$6-Procrustes distance has a closed form and converges, as $\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$7, to

$\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$8

(Quang, 2019). In that formulation, the Log-Euclidean metric and the Bures–Wasserstein metric arise as distinguished limits or special cases of one Riemannian family (Quang, 2019). The same paper also extends the logarithmic construction to positive definite unitized Hilbert-Schmidt operators, yielding the Log-Hilbert-Schmidt metric

$\exp:\mathcal{S}^n\to \mathcal{S}_{++}^n$9

(Quang, 2019).

4. Statistical and machine-learning procedures

Because log-Euclidean geometries are Euclidean after transformation, they support closed-form or standard algorithms that still admit intrinsic interpretation. In metric learning, one first computes

$\mathcal{S}^n$0

then learns a Mahalanobis matrix $\mathcal{S}^n$1 in the vector space of log-matrices by Information-Theoretic Metric Learning, producing the learned distance

$\mathcal{S}^n$2

(Vemulapalli et al., 2015). The key claim of that formulation is that this is still a genuine geodesic distance induced by a Riemannian metric on $\mathcal{S}^n$3 (Vemulapalli et al., 2015).

For nonparametric regression on SPD-valued responses, the framework appears as a Euclidean Pullback Metric. With $\mathcal{S}^n$4, the intrinsic local polynomial regression estimator has the closed form

$\mathcal{S}^n$5

which, in the local constant case, reduces to

$\mathcal{S}^n$6

(Reyes et al., 2023). The same paper proves a local-linear asymptotic bias formula and states consistency for the EPM setting (Reyes et al., 2023).

In SPD neural networks, pullback Euclidean metrics are used to define intrinsic classifiers. Under $\mathcal{S}^n$7, the Riemannian multinomial logistic regression model becomes

$\mathcal{S}^n$8

(Chen et al., 2023). That paper gives an intrinsic explanation of the common LogEig classifier and proves equivalence between the LEM-based SPD MLR and a LogEig MLR under the standard LEM and aligned optimization choices (Chen et al., 2023).

A further development replaces the fixed logarithm by a learnable generalized logarithm,

$\mathcal{S}^n$9

with induced adaptive distance

$X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$0

(Chen et al., 2023). The resulting Adaptive Log-Euclidean Metrics retain pullback geometry, Lie-group structure, closed-form Fréchet means, and explicit Riemannian operators while introducing learnable parameters into the geometry itself (Chen et al., 2023).

In privacy-preserving statistics, the same flattening principle supports the tangent Gaussian mechanism for the differentially private Fréchet mean. The private release is defined by sampling in log-coordinates and mapping back: $X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$1 after Gaussian perturbation of $X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$2 (Utpala et al., 2022). The paper derives a sensitivity bound

$X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$3

for data in a geodesic ball and reports utility scaling $X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$4, in contrast to $X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$5 for the earlier Riemannian Laplace mechanism (Utpala et al., 2022).

5. Representative application domains

In gait kinematics, the framework is used to compare Euclidean and manifold-based representations of locomotor variability across slow, medium, and fast conditions (Bůžek, 9 Dec 2025). Euclidean descriptors were defined directly on raw coordinates, including

$X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$6

whereas manifold-based descriptors used log-transformed SPD covariance representations, including $X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$7, $X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$8, and $X=U\,\mathrm{Diag}(e_1,\dots,e_n)\,U^T$9 (Bůžek, 9 Dec 2025). The reported empirical contrast is specific: Euclidean metrics showed a strictly monotonic pattern, $$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$0, while the Riemannian metrics showed an inverted-U pattern, with variance increasing from slow to medium and then decreasing slightly at fast speed (Bůžek, 9 Dec 2025). The paper interprets this as stabilization at high speed and possible adherence to geodesic, energy-efficient trajectories (Bůžek, 9 Dec 2025).

In dynamic functional connectivity, time-varying Pearson correlation matrices from sliding windows are smoothed by mapping each $$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$1 to Euclidean coordinates $$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$2, fitting a degree-6 polynomial on 10 evenly spaced samples, and pulling the fitted curve back by $$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$3 (Bisson et al., 21 Jul 2025). The key reported outcome is that Euclidean regression leaves the SPD cone, SPD log-Euclidean regression still requires rescaling, and that rescaling can alter correlations by up to $$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$4, whereas the proposed pullback framework stays inside $$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$5 and avoids that distortion (Bisson et al., 21 Jul 2025).

In image-set coding, the framework underlies the construction of CovDs-S, a kernel-matrix descriptor built from Gaussian embeddings of sub-image sets into the SPD manifold, mean centralization in the log domain, and mixtures of Log-Euclidean arc-cosine kernels,

$$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$6

(Chen et al., 2019). The final descriptor

$$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$7

uses kernel alignment to learn the mixing weights (Chen et al., 2019).

In human action recognition, the Log-Euclidean Bag of Words framework embeds covariance matrices of spatio-temporal HOF features by $$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$8, then applies $$\log(X)=U\,\mathrm{Diag}(\log(e_1),\dots,\log(e_n))\,U^T, \qquad \exp(X)=U\,\mathrm{Diag}(\exp(e_1),\dots,\exp(e_n))\,U^T$$9-means, hard assignment, spatio-temporal pyramids, or sparse coding in the embedded Euclidean space (Faraki et al., 2014). The paper reports performance improvements on KTH, Olympic Sports, and ADL, while also noting the complexity trade-off between HA, STP, and SC encodings (Faraki et al., 2014).

In dataset comparison without alignment, the log-Euclidean signature framework constructs SPD diffusion operators that approximate the heat kernel or Laplace-Beltrami operator, then compares them by a lower bound on the log-Euclidean metric: $\phi$0 (Shnitzer et al., 2022). The truncated descriptor

$\phi$1

enables comparison across datasets with different sizes, dimensions, and modalities (Shnitzer et al., 2022).

6. Boundaries, misconceptions, and theoretical caveats

A recurrent misconception is that every use of the matrix logarithm is intrinsically identical. Several papers distinguish between genuinely pullback-based intrinsic methods and extrinsic approximations that merely operate on transformed coordinates (Chen et al., 2023, Reyes et al., 2023). In the action-recognition literature, the Log-Euclidean BoW model is explicitly described as an extrinsic approach because it depends on a Euclidean embedding of the manifold (Faraki et al., 2014). By contrast, pullback formulations for classification or regression explicitly define the Riemannian metric on the original manifold and then use the transformed space only as an isometric coordinate system (Chen et al., 2023, Reyes et al., 2023).

A second caveat concerns correlation matrices. The 2025 theory proves that for any $\phi$2, there is no choice of log-Euclidean metrics on $\phi$3 and on $\phi$4 for which the inclusion

$\phi$5

is an isometric embedding (Bisson et al., 17 Nov 2025). This rules out the common simplification that correlation matrices can always be treated as an isometric submanifold of the SPD cone under some convenient logarithmic metric (Bisson et al., 17 Nov 2025).

A third boundary is that not every alignment-free or spectrally derived quantity is itself a metric. The log-Euclidean signature distance is explicitly stated not to satisfy identity of indiscernibles in general; isospectral but non-isometric manifolds can have zero LES distance (Shnitzer et al., 2022). Likewise, smooth visual appearance in transformed PCA plots is not itself evidence of intrinsic smoothness: in dynamic functional connectivity, the authors note that visual regularity in transformed trajectories can partly be an embedding artifact (Bisson et al., 21 Jul 2025).

Method-specific limitations also remain. The Riemannian multinomial logistic regression framework is developed for Pullback Euclidean Metrics rather than arbitrary Riemannian metrics, and extension to AIM is identified as future work (Chen et al., 2023). In local polynomial regression on SPD manifolds, the strongest closed-form results apply to Euclidean Pullback Metrics such as Log-Euclidean and Log-Cholesky, not to arbitrary geometries (Reyes et al., 2023). In high-dimensional SPD learning, adaptive or alternative pullback metrics may offer better time-error trade-offs than the standard logarithm, especially when matrix-log computations dominate runtime (Chen et al., 2023, Reyes et al., 2023).

Taken together, these caveats delimit the framework rather than weaken it. The common structural fact is that log-Euclidean methods are exact when the geometry is genuinely induced by a logarithmic diffeomorphism and when downstream procedures respect that pullback structure. Where additional constraints arise—unit diagonal, quotient structure, nonmetric spectral summaries, or metrics outside the pullback class—the framework either requires a modified logarithmic chart or ceases to provide the full set of intrinsic guarantees established in the classical SPD case (Bisson et al., 17 Nov 2025, Bisson et al., 21 Jul 2025).