Pullback Riemannian Metric

• Pullback Riemannian metrics are defined by transferring a target manifold’s metric to a domain via a smooth mapping.
• They enable applications in machine learning, robotics, and geometric control with explicit computational formulas for geodesics, distances, and stability.
• Key insights include accurate geodesic computation, curvature transfer, and practical real-time stability in non-Euclidean frameworks.

A pullback Riemannian metric is a fundamental construction in differential geometry wherein a smooth map between manifolds is used to transfer the geometric structure of a Riemannian metric from a target (or ambient) manifold to the domain manifold. This enables a wide class of applications, including geometry-aware machine learning, statistical modeling on constrained spaces, generative modeling on matrix manifolds, geometric control in robotics, and harmonic map theory. The construction underpins numerous modern algorithmic advances where non-Euclidean geometry must be imposed or leveraged for analysis, modeling, or inference.

1. Definition and Construction of Pullback Riemannian Metrics

Given a smooth map $f: (M, g_M) \rightarrow (N, g_N)$ between differentiable manifolds, with $g_N$ a Riemannian metric on $N$, the pullback metric $f^* g_N$ is the symmetric (0,2)-tensor field on $M$ defined by

$$(f^* g_N)_x(u,v) = g_N|_{f(x)} \left( Df_x[u], Df_x[v] \right), \quad u,v \in T_x M.$$

In local coordinates, the metric tensor is given by

$$(f^*g_N)_{ij}(x) = \frac{\partial f^\alpha}{\partial x^i} \frac{\partial f^\beta}{\partial x^j} (g_N)_{\alpha \beta}(f(x)),$$

with $Df_x$ the Jacobian matrix of $f$ at $x$.

If $f$ is a diffeomorphism onto its image and $g_N$ is positive-definite, then $f^* g_N$ is a Riemannian metric on $M$. If $N$ is Euclidean and $g_N$ is the canonical metric, then

$$g_x(u,v) = \langle Df_x[u], Df_x[v] \rangle_{\mathbb{R}^n}.$$

This definition underlies most applications, including manifold-valued generative modeling (Collas et al., 20 May 2025), geometry-aware statistical learning (Augenstein et al., 2024, Diepeveen, 2024), and geometric deep learning (Diepeveen et al., 2024).

2. Geometric Properties and Computational Formulas

Under a pullback metric, basic geometric operations on the domain $M$ reduce to calculations on the range $N$ via $f$:

• Geodesics: If $N$ is Euclidean and $f$ is a diffeomorphism, geodesics on $(M, f^*g_N)$ take the form

$$\gamma(t) = f^{-1}\left((1-t)f(x_0) + t f(x_1)\right).$$

• Distance: The Riemannian distance between $x_0, x_1 \in M$ satisfies

$d_{f^*g_N}(x_0, x_1) = \|f(x_0) - f(x_1)\|_2.$

• Exponential and logarithm maps: Exponential maps are computed as

$\exp^M_x(v) = f^{-1}(\exp^N_{f(x)}(Df_x(v))),$

where $\exp^N$ is the exponential map in $N$ (Collas et al., 20 May 2025, Diepeveen, 2024).

• Volume element: The Riemannian volume form is given by $\sqrt{\det (Df_x^T Df_x)}$ in the Euclidean pullback case (Tennenholtz et al., 2021).

In more general scenarios (e.g., mappings between non-Euclidean or stochastic manifolds), the pullback metric incorporates both the local geometry via the target metric and the local deformation via the map's Jacobian (Augenstein et al., 2024, Rozo et al., 7 Mar 2025).

3. Applications in Machine Learning and Statistical Modeling

3.1. Generative Models and Manifold Learning

Pullback Riemannian metrics are central in latent variable models where encoders or decoders parameterize a generative mapping $f: Z \rightarrow X$ between a low-dimensional latent space $Z$ and a data manifold $X$. The induced metric on $Z$ captures geometric properties of the data distribution and allows for:

• Geometry-aware distances for interpolation and path-finding (Diepeveen, 2024, Tennenholtz et al., 2021).
• Uncertainty quantification via volume or curvature of the induced metric tensor (Tennenholtz et al., 2021, Augenstein et al., 2024).
• Riemannian autoencoders with intrinsic dimension estimation (Diepeveen et al., 2024).

3.2. Matrix Manifolds and Information Geometry

For structured data such as symmetric positive-definite (SPD) matrices or correlation matrices, diffeomorphisms (e.g., matrix logarithm, Cholesky decomposition) are used to transfer the Euclidean metric to these manifolds:

• Log-Euclidean pullback: The diffeomorphism $$\phi_{\mathrm{SPD}}(\Sigma) = \mathrm{veclt}(\log \Sigma)$$ yields a pullback metric with closed-form expressions for geodesics, norms, and flows on SPD matrices (Collas et al., 20 May 2025).
• Correlation manifold: The normalized Cholesky transform induces an efficient pullback metric for correlation matrices (Collas et al., 20 May 2025).
• Hilbert pullback metric for Gaussians: The mixture-geodesic and Hilbert metric are realized as pullbacks via embedding into a higher-dimensional SPD cone, providing computationally tractable surrogates for the Fisher–Rao geometry (Nielsen, 2023).

3.3. Probabilistic Latent Variable Models

In GP-LVMs and related probabilistic encoders, where mappings $f: Z \rightarrow X$ are stochastic, the expected pullback metric at a point is evaluated as

$$\mathbb{E}[G(z)] = \mathbb{E}[J(z)]^T G_X(f(z)) \mathbb{E}[J(z)] + \mathrm{Tr}[G_X(f(z)) \Sigma_J],$$

with $J(z)$ the (random) Jacobian and $\Sigma_J$ its covariance. This allows for data-aware metrics and geodesics that avoid low-data or high-uncertainty regions (Augenstein et al., 2024, Rozo et al., 7 Mar 2025).

4. Pullback Metrics in Geometric Analysis and Harmonic Maps

The pullback metric is the first fundamental form associated with a differentiable map between Riemannian manifolds. In the context of harmonic maps $f: (M, g) \to (N, h)$, the pullback metric $g^* = f^* h$ encodes the metric distortion induced by $f$:

• The energy density is $e(f) = \frac{1}{2} \operatorname{tr}_g f^* h$.
• The divergence $\delta(g^*)$ and the tension field $\tau(f) = \operatorname{tr}_g Ddf$ obey the identity $\delta(g^*) + 2 d e(f) = -h(\tau(f), df(-))$, with harmonicity (minimization of energy) characterized by the vanishing of the tension field (Stepanov et al., 10 Jul 2025).
• Classical orthogonal splittings of the pullback metric (Berger–Ebin, York decompositions) elucidate its trace (isotropic scaling), exact (Lie derivative), and TT (transverse traceless) components, providing fine-grained control over the geometric deformation encoded by $f$ (Stepanov et al., 10 Jul 2025).

5. Robotics and Geometric Control via Pullback Bundle Metrics

In robotic motion planning and control, pullback metrics are used to transfer task-space Riemannian structure onto configuration spaces:

• Let $Q$ be the robot's configuration manifold, $T$ a task-space, and $\phi: Q \to T$ a task map. If $G(y)$ is a metric on $T$, the pullback metric on $Q$ is

$g(q) = D\phi(q)^T G(\phi(q)) D\phi(q).$

• This enables modular design and fusion of multiple geometric task objectives (attractors, constraints, damping) into a single globally consistent metric structure, supporting least-squares fusion and real-time control (Bylard et al., 2021).
• Position-dependent pullbacks guarantee compatibility and geometric well-posedness for both robot configuration spaces and general manifolds (Bylard et al., 2021).

6. Best Practices and Stability in Pullback Geometry

The expressiveness and stability of pullback metrics depend on careful design of the base diffeomorphism or generative map:

• Mapping data manifolds into totally geodesic submanifolds of the target ensures that interpolation (geodesics) and barycentre computations remain well-posed and that the induced geometry reflects the intrinsic structure of the data (Diepeveen, 2024).
• Local isometry—ensured by controlling the Lipschitz constants of both $\phi$ and $\phi^{-1}$—maximizes stability in geometric computations and prevents artificially induced curvature or metric collapse (Diepeveen, 2024).
• Pulling back positive curvature (e.g., from spheres) can cause instability in estimation and interpolation, while hyperbolic or locally flat pullback settings promote robustness (Diepeveen, 2024).
• For deep models, invertible normalizing flows with isometry regularization and subspace constraints yield stable pullback geometries, recover intrinsic dimension, and provide global charts (Diepeveen et al., 2024, Diepeveen, 2024).

7. Algorithmic and Numerical Aspects

Implementation of pullback metrics in practical algorithms involves:

• Computation of the Jacobian $Df_x$ (or higher-order derivatives for stochastic models) via automatic differentiation.
• Geodesic, log, and exp map computation via closed forms in special settings or via discrete energy minimization and numerical ODE solvers (Collas et al., 20 May 2025, Augenstein et al., 2024).
• Efficient linear algebraic approximations (e.g., extreme eigenvalue solvers for Hilbert pullback distance) and scalable techniques for high-dimensional data (Nielsen, 2023).
• Empirical regularization (isometry, subspace) and careful design of diffeomorphic mappings in deep learning pipelines (Diepeveen et al., 2024, Diepeveen, 2024).

---

The pullback Riemannian metric is a unifying and computationally tractable mechanism for transferring and adapting geometric structure, enabling advanced data analysis, machine learning on manifolds, information geometry, geometric control, and the study of harmonic maps. Recent developments navigate both theoretical (e.g. stability, intrinsic dimension, analytic criteria for harmonicity) and practical (e.g., fast sampling, real-time computation, deep-learning parametrizations) domains (Collas et al., 20 May 2025, Augenstein et al., 2024, Diepeveen, 2024, Diepeveen et al., 2024, Nielsen, 2023, Stepanov et al., 10 Jul 2025, Tennenholtz et al., 2021, Rozo et al., 7 Mar 2025, Bylard et al., 2021).