Pullback Geometry Fundamentals

Updated 7 May 2026

Pullback geometry is the study of transferring geometric structures, such as metrics and connections, between spaces via smooth maps.
It underpins applications in Riemannian geometry, fiber bundle theory, and categorical constructions by formalizing how intrinsic properties are inherited.
In data science, pullback methods drive isometric learning and generative modeling, enabling accurate manifold representation and improved interpolation.

Pullback geometry is the study of geometric structures induced on one mathematical space via a smooth mapping from another space equipped with a given geometric or analytic structure. The construction of a “pullback” transfers or lifts metrics, connections, measures, or other structures along a differentiable, algebraic, or categorical map, allowing the domain space to inherit geometric data from the codomain. This principle underlies core techniques in Riemannian geometry, analysis, category theory, data science, and geometric representation learning, providing a unifying language for modeling manifold-valued data, constructing generative models, extending currents, and systematizing categorical fiber-product constructions.

1. Pullback of Riemannian Metrics and General Definition

Given a smooth map $\phi : Z \to X$ between differentiable manifolds where $X$ is equipped with a Riemannian metric $g_X$ , the pullback metric $g_Z = \phi^* g_X$ is defined by

$g_Z(z)(v, w) = g_X(\phi(z))[ d\phi_z(v), d\phi_z(w) ]$

for all $v, w \in T_z Z$ . This construction equips $Z$ with a Riemannian metric such that for any tangent vectors at $z$ , their inner product is measured by first mapping them forward via $d\phi_z$ and then by applying the metric tensor on $X$ at $X$ 0. The pullback operation extends to more general geometric structures, including connections and forms, retaining properties such as compatibility with differentials and commutativity with pullback for composition of maps (where defined) (Kruiff et al., 2024).

This formalism generalizes to non-Riemannian settings, such as the pullback of metrics on metric measure spaces for continuous or weakly differentiable maps, where the pullback is typically defined on differentiable structures such as cotangent modules or Sobolev classes, using Jacobians and differentials in weaker senses (Ikonen et al., 2021).

2. Isometric Learning, Neural Parameterizations, and Flow Matching

A core application of pullback geometry in modern machine learning consists in learning an isometric diffeomorphism $X$ 1 that aligns a latent or ambient space $X$ 2 with a data manifold $X$ 3 (potentially embedded nonlinearly or with complex intrinsic geometry). The isometry condition requires

$X$ 4

or, at the level of metrics,

$X$ 5

for all $X$ 6. Practically, isometric learning is enforced with a composite loss function:

Graph-matching (global isometry): Penalizes deviation of geodesic distances between pairs of points under $X$ 7 from target distances measured on $X$ 8:

$X$ 9

Stability (control of Jacobian): Regularizes local variations in the metric by enforcing near-isometry of the Jacobian.

Parameterizations via Neural ODEs allow highly expressive, invertible diffeomorphisms $g_X$ 0 to be learned by integrating vector fields $g_X$ 1, with the inverse map available via time-reversal (Kruiff et al., 2024).

Given such maps, manifold-aware generative models become computationally tractable, as flow-matching objectives and closed-form geodesic interpolation can be performed efficiently in the (designed) latent space.

3. Pullback Metrics in Data-Driven Geometry and Representation Learning

Pullback geometry provides a principled approach to extracting and manipulating the intrinsic geometry of data manifolds or feature representations:

In persistent homology encodings, the pullback metric induced on the input data manifold via a PH encoding $g_X$ 2 is

$g_X$ 3

highlighting directions where PH is maximally sensitive, with importance ranked by the singular values of $g_X$ 4. The low rank of the pullback metric reflects information bottlenecks in the encoding and provides a model-agnostic quantification of sensitivity and feature relevance (Liang et al., 2023).

In probabilistic generative modeling for latent variable models (especially in hyperbolic or curved latent spaces), the pullback metric accounts for both the nonlinear mapping and the ambient geometry. For a nonlinear latent-to-data mapping $g_X$ 5 and ambient metric $g_X$ 6, the pullback

$g_X$ 7

ensures that metric computations, Riemannian distances, and geodesics in the latent space respect the true data manifold density, yielding improved interpolation and uncertainty properties, particularly for hierarchical or non-Euclidean latent representations (Augenstein et al., 2024).

In normalizing flows, autoencoders, and score-based models, pullback metrics provide the backbone for defining geodesics, designing dimension-reducing encoders, and minimizing volume or isometry distortion, often with closed-form filterings for Gaussian or structured densities (Diepeveen et al., 2024).

4. Pullback Geometry in Manifold Theory and Bundle Constructions

Beyond data-driven contexts, pullback geometry is central in classical manifold theory:

The pullback of Riemannian metrics by submersions or immersions is fundamental in the study of submanifolds and fiber bundles. For vector bundles $g_X$ 8 classified via maps $g_X$ 9, the pullback of the universal connection from the Grassmannian imparts rich geometric properties to $g_Z = \phi^* g_X$ 0—its curvature, fatness, and stability are determined by geometric features of $g_Z = \phi^* g_X$ 1 (such as its Wirtinger angle and immersion properties). For example, a “fat” pullback connection arises iff $g_Z = \phi^* g_X$ 2 is an immersion with bounded Wirtinger angle, and parallelism of the pullback connection is characterized via vanishing of shape operators of $g_Z = \phi^* g_X$ 3 (Tapp, 2014).
For principal bundles and Riemannian submersions with totally geodesic fibers, the non-negativity of curvature of the pullback total space imposes strict constraints—namely, the level sets of the pullback map must form totally geodesic submanifolds of the base. This “A-flat” foliation condition provides obstructions to non-negative curvature in constructions such as exotic spheres and Rigas bundles (Durán et al., 2013).

5. Pullback Factorization and Categorical Pullbacks

In categorical settings, pullbacks unify intersection and fiber product constructions:

In graph of groups and groupoid categories, categorical pullbacks (fibered products) of morphisms encapsulate subgroup intersections at the level of fundamental groups. These are realized constructively as $g_Z = \phi^* g_X$ 4-products or fibered box products, with existence and universal properties depending on the pointing and acylindricity conditions of the ambient graph (Delgado et al., 6 Aug 2025).
In the abstract theory of tangent categories, the precise characterization of morphisms along which pullbacks exist and are preserved by tangent functors is given by “tangent display maps.” These coincide with submersions in the classical manifold setting and ensure that constructions such as differential bundles, connections, slice categories, and restriction structures can be performed functorially and stably (Cruttwell et al., 28 Feb 2025).

The ability to realize the full strength of pullback constructions in these categorical frameworks is linked to structural properties such as existence, uniqueness, and preservation under functorial images.

6. Pullback of Currents, Forms, and Orbifold Bundles

For analytic and algebro-geometric currents:

The pullback of currents (generalized differential forms) via holomorphic maps between complex manifolds extends the classical form pullback to a much broader class of “PS-currents,” enabling the transfer of both cycle and singular structures. This is implemented via Gysin-type constructions—using graph embeddings, normal bundles, and Chern classes—to define pullback operations compatible with cohomology and intersection theory, including for meromorphic maps and singular spaces (Kalm, 2020).
In orbifold and parabolic bundle theory, the pullback geometry is governed by the base orbifold structure, ramification, and stack-theoretic maps. Stability of vector bundles under pullback is preserved if and only if the morphism is genuinely ramified (i.e., inertia groups map surjectively), ensuring that degrees, slopes, and filtration weights are transferred correctly without degenerating stability properties (Das et al., 2021).

7. Applications in Data Manifold Learning and Manifold-Constrained Generation

Pullback geometry underpins advanced generative and representation learning frameworks:

Pullback Flow Matching (PFM): Efficient generative modeling on data manifolds is achieved by learning an invertible diffeomorphism $g_Z = \phi^* g_X$ 5 between a latent space and the data manifold, enforcing (approximate) isometry, and training flow models on the latent that pull back to well-behaved geodesic transport in data space. This yields significant advances in interpolation, sample novelty, and property-guided generation, as demonstrated for synthetic manifolds and protein design (Kruiff et al., 2024).
Structured generative models on matrix manifolds: For structured spaces such as SPD or correlation matrices, global analytical diffeomorphisms (log-Euclidean, normalized Cholesky) enable the transfer of Euclidean generative modeling techniques onto matrix manifolds by replacing all geometric operations with their pullback via the chosen diffeomorphism. This enables efficient training and sampling, preserving manifold constraints and yielding state-of-the-art empirical results in neuroimaging (Collas et al., 20 May 2025).
Data-driven Riemannian geometry: Score-based pullback metrics and learned normalizing flows enable direct extraction of the data manifold’s geometry, allowing accurate geodesic computation, robust barycenter estimation, and automatic discovery of intrinsic dimension. The resulting manifold-aware autoencoders realize provable error bounds and faithful low-dimensional chartings (Diepeveen et al., 2024, Diepeveen, 2024).
Probabilistic pullback geometry in hyperbolic latent manifolds: In Gaussian process LVMs with hyperbolic latent spaces, the pullback metric accounts for both the ambient curvature and mapping distortion, ensuring that interpolation and uncertainty quantification reliably follow the data support and avoid high-uncertainty gaps, which is essential for robust generative tasks in hierarchical domains (Augenstein et al., 2024).

Pullback geometry constitutes a unifying principle enabling rigorous geometric structures to be transferred, inherited, or reconstructed across mappings, with deep implications in differential geometry, algebraic topology, category theory, data science, and modern representation learning. It provides the foundational mechanisms by which intrinsic data geometry, curvature, and manifold structure are respected and exploited in high-dimensional modeling, with broad impact on generative modeling, inference, and the analysis of structured data.