Riemannian Metric Matching: Methods & Applications

• Riemannian metric matching is a framework that defines or learns a smooth field of inner products so that induced geometric features (e.g., geodesic distances, volumes) agree with a target structure.
• It employs diverse methods—from pullback metrics in latent variable models to quotient and barrier-modified geometries in shape and SPD matrix analysis—to capture intrinsic, non-Euclidean properties.
• This matching framework enhances applications by aligning data representations with uncertainty, curvature, and symmetry, paving the way for improved optimization, interpolation, and manifold reconstruction.

Riemannian metric matching denotes a family of constructions in which a Riemannian metric is selected, estimated, or optimized so that the geometry it induces—local inner products, geodesic distances, interpolation paths, volume elements, transport costs, or tangent-space operators—agrees with a target structure. Across the literature, the term is not fully standardized. In some works it names a specific algorithmic framework, such as covariance fitting on the manifold of Hermitian positive definite matrices or denoising-based estimation of diffusion-geometric co-metrics; in others, the same principle appears through pullback metrics, quotient-shape geodesics, or optimization-oriented metric design (Cohen et al., 12 May 2026, Bamberger et al., 12 Jun 2026, Gruffaz et al., 7 Mar 2025).

1. Conceptual scope

A Riemannian metric on a manifold $M$ is a smooth field of inner products $g_p:T_pM\times T_pM\to\mathbb R$. It determines curve length, geodesic distance, exponential and logarithm maps, parallel transport, and volume. In the review formulation, Riemannian metric learning is expressed as

$$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$

so the learned object is not merely a tensor field but the full induced geometry (Gruffaz et al., 7 Mar 2025).

Within this broad formulation, several distinct matching problems recur. One may match a latent metric to observation-space deformation through a pullback; match a quotient-shape metric to reparameterization-invariant shape change; match model and sample covariance matrices on the HPD manifold; match a data-driven co-metric to diffusion geometry; or match the optimization metric to the second-order structure of a constrained objective. The review also notes that the manifold $\mathcal G_M$ of Riemannian metrics itself carries a natural metric,

$$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$

but practical learning methods rarely operate by directly minimizing a discrepancy between two full tensor fields (Gruffaz et al., 7 Mar 2025).

Setting	Matched object	Representative formulation
Generative latent models	Pullback or expected pullback geometry	$G=J^\top J$, $\mathbb E[J^\top J]$
Shape analysis	Quotient geodesics modulo reparameterization	$$\bar d(\overline{c_0},\overline{c_1})=\inf_{\varphi} d(c_0,c_1\circ\varphi)$$
HPD/SPD matrices	Covariance or geodesic discrepancy	$$\widehat p=\arg\min_{p\in\mathbb R_+^D}\mathcal D^2(R(p),\widehat R)$$
Diffusion geometry	Local co-metric / carré du champ	$$\Gamma_{\mathcal L}(f,h)=\frac12(f\mathcal L h+h\mathcal L f-\mathcal L(fh))$$
Optimization	Preconditioned metric	$g_p:T_pM\times T_pM\to\mathbb R$0

This variety explains a recurrent terminological subtlety. “Riemannian metric matching” may mean direct metric estimation, distance matching, geodesic matching, covariance fitting on a curved cone, or preconditioning by metric choice. The unifying theme is that Euclidean discrepancy is replaced by geometry induced by a nonconstant or non-Euclidean metric.

2. Pullback geometry and probabilistic manifold matching

A foundational instance appears in probabilistic generative dimensionality reduction. For a smooth generative map $g_p:T_pM\times T_pM\to\mathbb R$1, the natural latent-space metric is the pullback

$g_p:T_pM\times T_pM\to\mathbb R$2

with $g_p:T_pM\times T_pM\to\mathbb R$3 the Jacobian of $g_p:T_pM\times T_pM\to\mathbb R$4. Infinitesimal latent displacements are then measured so that their lengths match observation-space infinitesimal lengths under the map. In the probabilistic GP-LVM setting, the Jacobian is random, which induces a distribution over local metrics, and the expected metric satisfies

$g_p:T_pM\times T_pM\to\mathbb R$5

The covariance correction enlarges the metric in uncertain regions; accordingly, curve lengths increase there and geodesics avoid poorly supported parts of latent space (Tosi et al., 2014).

This construction yields an uncertainty-aware form of metric matching. The deterministic pullback term aligns latent and observation geometry, while the additive $g_p:T_pM\times T_pM\to\mathbb R$6 term aligns the metric with epistemic uncertainty. The same work defines geodesics by minimizing

$g_p:T_pM\times T_pM\to\mathbb R$7

and uses them for interpolation in latent space. For handwritten digits, rotating objects, and motion capture, geodesic interpolation follows the learned manifold support more closely than Euclidean straight lines, yielding more plausible intermediate reconstructions (Tosi et al., 2014).

A more recent generative formulation replaces Euclidean conditional paths in flow matching by approximate geodesics of a data-induced ambient metric. Metric Flow Matching defines a smooth SPD-valued field $g_p:T_pM\times T_pM\to\mathbb R$8 on $g_p:T_pM\times T_pM\to\mathbb R$9, learns interpolants $$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$0 by minimizing the metric kinetic energy

$$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$1

and then matches vector fields under the corresponding Riemannian norm,

$$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$2

Here the matched quantity is the conditional tangent velocity along approximate geodesics of the learned metric, rather than along Euclidean straight segments (Kapuśniak et al., 2024).

Taken together, these works establish a recurrent pattern: metric matching in generative models frequently means choosing geometry so that latent or interpolating trajectories respect the data manifold, uncertainty, and observation-space semantics, rather than an a priori Euclidean chart.

3. Shape spaces, quotient geometry, and registration

In shape analysis, metric matching is often inseparable from quotient geometry. The space of parametrized curves or surfaces carries nuisance symmetries, chiefly reparameterization, and the relevant metric must descend to the quotient. For manifold-valued open curves $$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$3, a reparameterization-invariant metric $$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$4 satisfies

$$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$5

which induces a quotient metric on shape space $$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$6. The resulting distance is

$$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$7

Optimal matching is then geodesic computation in the quotient, realized through horizontal lifts in the total space (Brigant et al., 2024). A related unifying account treats intrinsic Sobolev metrics on immersions, outer metrics induced by ambient diffeomorphisms, and hybrid combinations through the common language of Riemannian submersions and optimal control (Bauer et al., 2018).

For elastic metrics on curves, an important computational development is the replacement of explicit optimization over the reparameterization group by a varifold-based endpoint relaxation. The relaxed matching problem

$$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$8

keeps the intrinsic Riemannian deformation model while avoiding direct discretization of $$\operatorname*{argmin}_{g\in \mathcal G_\theta\subset \mathcal G_M} L(\mathcal F(g),X), \qquad \mathcal F(g)=(d_g,\operatorname{Exp}_g,T_g,\mathrm{vol}_g),$$9. This supports both open and closed curves, includes first-order elastic metrics and second-order $\mathcal G_M$0-type metrics, and can also quotient out Euclidean motions or scaling (Bauer et al., 2018).

For parametrized regular surfaces in $\mathcal G_M$1, an inner Sobolev $\mathcal G_M$2-type metric is defined directly on deformation vector fields along the surface,

$\mathcal G_M$3

yielding a geodesic shooting framework on the manifold of surfaces itself rather than on an ambient diffeomorphism group (Bauer et al., 2011). This is a particularly direct form of metric matching: deformation cost is measured by a geometry attached to the evolving surface.

A finite-dimensional analogue appears for planar triangular meshes with fixed connectivity. There the admissible configurations form a smooth manifold, and a complete metric

$\mathcal G_M$4

is built from barrier terms involving reciprocal triangle heights, reciprocal boundary vertex-edge distances, and a coercive quadratic term. The resulting geodesics preserve mesh validity, prevent triangle collapse, and keep cell aspect ratios bounded away from zero along arbitrarily long geodesic curves (Herzog et al., 2020).

These constructions share a common principle: the matched geometry is not the raw ambient Euclidean geometry of coordinates, but a quotient, intrinsic, or barrier-modified geometry tailored to the admissible notion of shape change.

4. SPD and HPD matrices as metric-matching domains

A second major regime concerns matrix manifolds, especially symmetric or Hermitian positive definite matrices. In array processing, Riemannian metric matching is formulated as covariance fitting on the HPD manifold. Given a sample covariance $\mathcal G_M$5 and a model covariance

$\mathcal G_M$6

the generic problem is

$\mathcal G_M$7

The SERCOM method replaces Euclidean covariance discrepancies by the Jensen–Bregman LogDet divergence

$\mathcal G_M$8

arguing that HPD covariance matrices should be matched in a geometry-aware manner rather than by Frobenius subtraction. The method is explicitly described as a Riemannian-geometry-aware divergence surrogate rather than exact affine-invariant geodesic matching (Cohen et al., 12 May 2026).

For supervised comparison of SPD-valued data, log-Euclidean metric learning provides a different form of matching. An SPD matrix is mapped to the log domain, vectorized, and endowed with a learned Mahalanobis form $\mathcal G_M$9, inducing the geodesic distance

$$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$0

This reduces Riemannian metric learning for SPD matrices to metric learning in the matrix-logarithm domain while preserving a log-Euclidean Riemannian interpretation (Vemulapalli et al., 2015).

A third line constructs a new geometry on $$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$1 through Cholesky factors. The Log-Cholesky metric is induced by a Lie group structure on lower triangular matrices with positive diagonal, yielding explicit formulas for geodesics, logarithm and exponential maps, Fréchet means, and parallel transport. Its distance reduces to

$$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$2

and the determinant of the Log-Cholesky average is the geometric mean of the input determinants, which eliminates swelling in the sense defined in the paper (Lin, 2019).

Across these works, the matched object varies: model covariances to empirical covariances, task geometry to labeled SPD descriptors, or interpolation geometry to determinant-preserving matrix statistics. What unifies them is the replacement of flat Euclidean comparison by geometry native to the positive definite cone.

5. Data-driven geometry from distributions, operators, and graphs

Another strand learns geometry directly from sampled data. One formulation chooses a metric $$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$3 from a family $$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$4 by maximizing normalized inverse volume at the observed points,

$$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$5

equivalently as maximum likelihood under the density

$$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$6

On the multinomial simplex, the construction uses pullbacks of the Fisher information metric under a parametric family $$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$7, leading to a learned geodesic distance that resembles TFIDF reweighting but is derived from Riemannian geometry (Lebanon, 2012).

A more recent formulation uses diffusion geometry and learns the carré du champ operator rather than a tensor field in coordinates. For a diffusion generator $$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$8,

$$\mathbf G_g(\mathfrak h,\mathfrak k) = \int_M \operatorname{tr}\!\left(g_x^{-1}\mathfrak h_x g_x^{-1}\mathfrak k_x\right)\,d\mathrm{vol}_g(x),$$9

and on a Riemannian manifold this satisfies

$G=J^\top J$0

The key observation is that, after Gaussian corruption $G=J^\top J$1, the kernel CDC can be written as a conditional expectation, leading to the training loss

$G=J^\top J$2

This is the paper’s explicit use of the phrase “Riemannian metric matching”: a denoising probabilistic framework for learning local co-metric information with sample-wise training and amortized inference (Bamberger et al., 12 Jun 2026).

There are also exact or near-exact reconstruction results in discrete settings. On a fixed Euclidean polyhedral surface with cotangent Laplacian, the discrete Laplace–Beltrami operator determines the discrete metric uniquely up to global scale. The proof constructs a convex energy $G=J^\top J$3 with gradient equal to the edge-weight vector, showing that cotangent weights determine edge lengths modulo scaling (Gu et al., 2010). In a random geometric graph model whose edge probabilities depend monotonically on intrinsic Riemannian distance, there is an $G=J^\top J$4-time algorithm that reconstructs all pairwise geodesic distances between sampled latent points with uniform additive error

$G=J^\top J$5

under the paper’s regularity assumptions (Huang et al., 7 Nov 2025).

These approaches suggest a broad interpretation: metric matching can target local volume, diffusion operators, Laplace-type observables, or graph connectivity, provided those objects encode the desired geometry.

6. Metric choice as preconditioning and finite-dimensional approximation

In optimization, metric matching often means selecting a Riemannian metric that reflects the second-order structure of the objective. For equality-constrained problems and quotient manifolds, Riemannian preconditioning proposes to derive the metric from the Hessian of the Lagrangian. A representative construction is

$G=J^\top J$6

or a regularized variant using separate cost-related and constraint-related terms. The central claim is that on quotient manifolds the Lagrangian is degenerate in vertical directions but captures the relevant second-order information on the horizontal space, so choosing the metric from this structure yields a principled preconditioner (Mishra et al., 2014).

For product manifolds, the same idea is made systematic by designing a block-diagonal operator $G=J^\top J$7 that approximates the diagonal blocks of the Euclidean-metric Riemannian Hessian, and then defining

$G=J^\top J$8

Exact block-diagonal preconditioning, left and right preconditioning, and Gauss–Newton type preconditioning are all presented as concrete ways to match the metric to the local curvature of the cost while keeping application and inversion tractable. Canonical correlation analysis, truncated singular value decomposition, and tensor ring completion are given as examples where this metric design accelerates Riemannian methods (Gao et al., 2023).

A more structural finite-dimensional approximation problem appears in the theory of Riemannian Bergman metrics. For a closed manifold $G=J^\top J$9, the first $\mathbb E[J^\top J]$0 Laplace eigenspaces define an embedding into Euclidean space whose pullback metric lies in a finite-dimensional Bergman family. The paper constructs a map $\mathbb E[J^\top J]$1 such that

$\mathbb E[J^\top J]$2

and shows that Bergman metrics are $\mathbb E[J^\top J]$3-dense in the space of all Riemannian metrics after the appropriate spectral normalization (Potash, 2013). This yields finite-dimensional symmetric-space approximations to metric space itself, so metric matching may be carried out through finite-dimensional spectral representatives rather than directly in the infinite-dimensional manifold of smooth metrics.

In all these cases, the operative notion of matching is algorithmic or representational: the metric is tailored to curvature, quotient structure, or a finite-dimensional encoding of an otherwise infinite-dimensional geometric object.

7. Terminological issues, limitations, and recurring distinctions

Several distinctions recur across the literature. First, direct tensor-field matching remains comparatively uncommon. The review explicitly notes that most practical methods do not optimize a discrepancy between two full metrics in $\mathbb E[J^\top J]$4; instead they match induced distances, geodesics, transport, or local operators (Gruffaz et al., 7 Mar 2025). This suggests that “Riemannian metric matching” usually refers to induced-geometry matching rather than pointwise tensor regression.

Second, surrogate geometry is common. SERCOM, for example, operates with JBLD, which is described as a geometry-preserving surrogate tied to affine-invariant geometry rather than the AIRM geodesic distance itself (Cohen et al., 12 May 2026). Likewise, relaxed shape matching with varifold endpoint penalties does not directly optimize over reparameterizations, but replaces hard quotient constraints by an invariant terminal discrepancy (Bauer et al., 2018).

Third, some constructions are explicitly approximate. In probabilistic latent-variable models, the full object is a random Riemannian manifold, but practical algorithms typically use only the expected metric $\mathbb E[J^\top J]$5. The source text warns that this may be misleading in regimes where the metric distribution is not sharply concentrated (Tosi et al., 2014). In diffusion geometry, the learned object may be a local CDC matrix or tangent projector rather than a globally integrated geodesic metric (Bamberger et al., 12 Jun 2026). In random geometric graphs, the output is pairwise geodesic distance recovery on sampled points, not a smooth tensor field on the underlying manifold (Huang et al., 7 Nov 2025).

Fourth, many exact identifiability results come with stringent structural assumptions. The cotangent Laplace–Beltrami operator determines the metric only on a fixed triangulation and only up to scale (Gu et al., 2010). Mesh-based complete metrics presuppose fixed connectivity and known vertex correspondence (Herzog et al., 2020). HPD covariance matching assumes known model structure and, in the main formulation, known noise variance (Cohen et al., 12 May 2026). Such conditions are not incidental; they delimit the regime in which a given form of metric matching is well posed.

A plausible synthesis is therefore that Riemannian metric matching is best understood as a family of geometry-aware estimation, fitting, and optimization problems rather than a single method. What is matched may be a pullback metric to observation-space change, a quotient metric to shape equivalence, a covariance model to an HPD sample, a co-metric to diffusion geometry, a discrete metric to operator data, or an optimization metric to Hessian structure. The concept is unified by one principle: the relevant comparison object is no longer Euclidean distance alone, but the Riemannian geometry induced by the metric.