Neural Vector Bundle: Geometric Perspectives

• Neural vector bundles are mathematical frameworks that attach local vector spaces to nodes or manifold points, enabling structured transport via connections and heat operators.
• They enable processing signals using bundle diffusion, tangent-bundle filtering, and reproducing-kernel techniques, addressing challenges like over-smoothing and over-squashing.
• Applications include Bundle Neural Networks and Tangent Bundle Neural Networks, which demonstrate improved convergence and efficiency in denoising and graph benchmark tasks.

“Neural vector bundle” is not a standard formal term, but the phrase has acquired a coherent mathematical meaning across several research lines: neural computation is organized by vector-bundle geometry rather than by a single ambient feature space. In this usage, a model may attach a local vector space to each node or point, compare features through orthogonal transport maps or a connection, evolve signals by a bundle heat operator or Connection Laplacian, or interpret learned maps as elements of a vector-valued reproducing kernel Banach space. The phrase is used most directly for Bundle Neural Networks on graphs, Tangent Bundle Neural Networks on manifolds, and the vv-RKBS formulation of vector-valued networks and neural operators (Bamberger et al., 2024, Battiloro et al., 2022, Dummer et al., 30 Sep 2025).

1. Scope of the term

The literature supports at least four distinct but related meanings of a neural vector bundle. These meanings should not be conflated, because they live at different levels of structure: discrete transport on graphs, differential geometry on manifolds, reproducing-kernel function spaces, and classifying constructions in supergeometry (Bamberger et al., 2024, Battiloro et al., 2022, Dummer et al., 30 Sep 2025, Dumitrescu, 2010).

Interpretation	Base space	Defining mechanism
Graph bundle diffusion	Graph $G=(V,E)$	Local vector spaces, orthogonal transports, bundle heat diffusion
Tangent-bundle processing	Riemannian manifold $\mathcal M$	Vector fields, Connection Laplacian, tangent-bundle filters
vv-RKBS viewpoint	$X$, $\Omega$, or $Z$	Banach-valued outputs, reproducing kernel $K:X\times\Omega\to\twin(U,U^\diamond)$
Supergeometric transport	Manifold $M$ or supermanifold	Parallel transport along superpaths and connection data

A common misconception is to identify any vector-valued neural network with a neural vector bundle. The surveyed papers impose stronger structure. In BuNN, each node carries its own local coordinate system and features are transported through orthogonal maps before diffusion (Bamberger et al., 2024). In TNN, a signal is not merely $\mathbb R^d$-valued; it is a vector field $\mathbf F:\mathcal M\to\mathcal T\mathcal M$, so $\mathbf F(x)\in\mathcal T_x\mathcal M$ and processing must respect variation of tangent spaces across the manifold (Battiloro et al., 2022). In the vv-RKBS setting, the “bundle” aspect is interpretive rather than literal: the base is the input or conditioning domain, the fibers are Banach output spaces or function spaces, and the reproducing kernel supplies the geometry (Dummer et al., 30 Sep 2025).

2. Connection and transport as foundational bundle data

A foundational geometric model is supplied by the equivalence between $\mathcal M$0 parallel transport along superpaths and even connections on a $\mathcal M$1-graded vector bundle over an ordinary manifold $\mathcal M$2 (Dumitrescu, 2010). A connection on $\mathcal M$3 is a covariant derivative

$\mathcal M$4

and it is even when it preserves the grading. A superpath is a family

$\mathcal M$5

and parallel transport along superpaths is defined by horizontal sections satisfying

$\mathcal M$6

together with identity on constant superpaths, functoriality in the parameter $\mathcal M$7, gluing under concatenation, invariance under reparametrization by diffeomorphisms preserving the $\mathcal M$8-distribution, and compatibility with ordinary paths via the projection $\mathcal M$9.

The central theorem states a natural $X$0-$X$1 correspondence between $X$2 parallel transport on $X$3 over $X$4 and even connections on $X$5 over $X$6. The theorem is both an existence and uniqueness statement: every admissible $X$7 parallel transport comes from a unique even connection, and every even connection induces such a transport. A key step passes through the odd tangent bundle $X$8 and the projection $X$9, introducing odd-trivial connections on $\Omega$0. These are precisely the pullbacks of connections from $\Omega$1, characterized by

$\Omega$2

This framework is directly relevant when a neural vector bundle is understood through transport rules rather than solely through bundle charts. The result says that “moving vectors around” on a supermanifold is not extra structure beyond a connection: if the transport law satisfies the supergeometric axioms, it is exactly the holonomy or parallel transport of a unique connection (Dumitrescu, 2010). A plausible implication is that bundle-based neural models can be organized either by explicitly parameterizing a connection or by directly learning admissible transport laws, provided the two descriptions remain equivalent.

3. Tangent-bundle neural networks on manifolds

On a compact smooth Riemannian manifold $\Omega$3, the natural bundle-valued signal is a vector field

$\Omega$4

with $\Omega$5. The geometric operator governing such signals is the Connection Laplacian $\Omega$6, defined as the trace of the second covariant derivative induced by the Levi-Civita connection; it is self-adjoint, elliptic, and negative semidefinite (Battiloro et al., 2022). The associated heat equation

$\Omega$7

has solution $\Omega$8, and this yields the tangent-bundle filter

$\Omega$9

A Tangent Bundle Neural Network stacks such filters with a differential-preserving nonlinearity. For layer $Z$0, the $Z$1-th output channel is

$Z$2

The architecture is continuous: the signal lives in the tangent bundle, the filters are spectral multipliers of the Connection Laplacian, and the nonlinearity is constrained to preserve bundle structure.

The discrete realization proceeds by sampling $Z$3 into a point cloud, constructing a geometric graph, estimating local tangent bases by local PCA, and using SVD to project transport estimates to orthogonal matrices $Z$4. This produces an $Z$5-bundle in the form of a cellular sheaf and a normalized Sheaf Laplacian $Z$6. After time discretization, the discretized space-time TNN becomes

$Z$7

which the paper identifies as a principled variant of Sheaf Neural Networks (Battiloro et al., 2022).

The convergence theorem is structural: under bandlimitedness, non-amplifying and Lipschitz-continuous filters in frequency, differential-preserving nonlinearities, and bandlimited sampled sheaf signals, the discrete architecture converges in probability to the underlying continuous TNN. The numerical illustration is denoising of a tangent vector field on the unit $Z$8-sphere; DD-TNN consistently achieves lower MSE than a one-layer Manifold Neural Network baseline across $Z$9, $K:X\times\Omega\to\twin(U,U^\diamond)$0, and noise levels $K:X\times\Omega\to\twin(U,U^\diamond)$1, $K:X\times\Omega\to\twin(U,U^\diamond)$2, and $K:X\times\Omega\to\twin(U,U^\diamond)$3 (Battiloro et al., 2022).

4. Bundle Neural Networks on graphs

Bundle Neural Networks recast graph message propagation as diffusion on a flat vector bundle rather than ordinary neighbor aggregation (Bamberger et al., 2024). For an undirected graph $K:X\times\Omega\to\twin(U,U^\diamond)$4, each node $K:X\times\Omega\to\twin(U,U^\diamond)$5 carries a local vector space $K:X\times\Omega\to\twin(U,U^\diamond)$6, typically $K:X\times\Omega\to\twin(U,U^\diamond)$7, and each edge induces an orthogonal transport map

$K:X\times\Omega\to\twin(U,U^\diamond)$8

in the flat case. Flatness means path-independence and factorization through per-node orthogonal frames $K:X\times\Omega\to\twin(U,U^\diamond)$9. The bundle Laplacian is

$M$0

and the bundle Dirichlet energy is

$M$1

Smoothness therefore means agreement after parallel transport, not literal equality in a global coordinate system.

The BuNN layer is derived from the heat equation on the vector bundle,

$M$2

with solution

$M$3

For flat bundles,

$M$4

A BuNN layer has four steps:

1. Compute bundle maps:

$M$5

2. Bundle-aware local update:

$M$6

3. Diffuse over the bundle:

$M$7

4. Apply nonlinearity:

$M$8

BuNNs are a special case of Sheaf Neural Networks, but the restriction to flat vector bundles yields a closed-form heat kernel and nodewise orthogonal frames, rather than general edgewise sheaf maps. The paper argues that this improves tractability and changes the communication regime from iterative local message passing to global diffusion (Bamberger et al., 2024).

The theoretical claims are threefold. First, BuNN mitigates over-smoothing because the infinite-time limit is not forced to be constant across nodes; aligned outputs satisfy $M$9, while the nodewise outputs $\mathbb R^d$0 can remain distinct. Second, BuNN mitigates over-squashing because for a linear BuNN layer

$\mathbb R^d$1

and $\mathbb R^d$2 for all nodes in a connected graph and any $\mathbb R^d$3, so every node can influence every other node in one layer. Third, with injective positional encodings, $\mathbb R^d$4-layer BuNNs with encoder and decoder have compact uniform approximation over connected graph families; the width bound is $\mathbb R^d$5 on any finite subfamily $\mathbb R^d$6.

The empirical evidence follows the same structure. On synthetic barbell and clique tasks, BuNN nearly solves both, with MSE around $\mathbb R^d$7 on barbell and $\mathbb R^d$8 on clique. On the heterophily suite from Platonov et al.—roman-empire, amazon-ratings, minesweeper, tolokers, and questions—BuNN achieves the best score on all five tasks, with a relative improvement of about $\mathbb R^d$9 on average. On the Long Range Graph Benchmark, BuNN-Hop achieves $\mathbf F:\mathcal M\to\mathcal T\mathcal M$0 average precision on Peptides-func, reported as a new state of the art (Bamberger et al., 2024).

5. Function-space geometry: vv-RKBS as a bundle-like neural formalism

A different line of work shifts from geometric transport to function-space structure. A vector-valued reproducing kernel Banach space is a Banach space $\mathbf F:\mathcal M\to\mathcal T\mathcal M$1 of functions $\mathbf F:\mathcal M\to\mathcal T\mathcal M$2, where $\mathbf F:\mathcal M\to\mathcal T\mathcal M$3 is a Banach space, such that point evaluations are bounded: $\mathbf F:\mathcal M\to\mathcal T\mathcal M$4 Equivalently,

$\mathbf F:\mathcal M\to\mathcal T\mathcal M$5

with

$\mathbf F:\mathcal M\to\mathcal T\mathcal M$6

The paper develops a general kernel notion

$\mathbf F:\mathcal M\to\mathcal T\mathcal M$7

where $\mathbf F:\mathcal M\to\mathcal T\mathcal M$8 denotes twin operators associated with the dual pair $\mathbf F:\mathcal M\to\mathcal T\mathcal M$9 (Dummer et al., 30 Sep 2025).

This formulation is deliberately more general than vv-RKHS: it allows asymmetric domains $\mathbf F(x)\in\mathcal T_x\mathcal M$0, infinite-dimensional outputs, and avoids assumptions of reflexivity, separability, and symmetry. The reproducing property is written through duality pairings rather than an inner product: $\mathbf F(x)\in\mathcal T_x\mathcal M$1

The neural specialization uses integral and neural vv-RKBSs. For shallow $\mathbf F(x)\in\mathcal T_x\mathcal M$2-valued networks, the feature function is

$\mathbf F(x)\in\mathcal T_x\mathcal M$3

and the representer theorem yields a sparse atomic measure

$\mathbf F(x)\in\mathcal T_x\mathcal M$4

hence

$\mathbf F(x)\in\mathcal T_x\mathcal M$5

or, in matrix notation,

$\mathbf F(x)\in\mathcal T_x\mathcal M$6

The theorem guarantees at most $\mathbf F(x)\in\mathcal T_x\mathcal M$7 atoms, not necessarily $\mathbf F(x)\in\mathcal T_x\mathcal M$8.

The same machinery covers neural operators. DeepONet is recovered as

$\mathbf F(x)\in\mathcal T_x\mathcal M$9

while the joint representer theorem for hypernetworks and function-space hypernetworks yields

$\mathcal M$00

The paper states explicitly that “Neural Vector Bundle” is not a formal term defined there, but it is a good conceptual reading: the base is the input or conditioning domain, the fibers are the output Banach spaces or function spaces, a network is a section or function valued in those fibers, and the kernel acts like a bundle morphism transporting information between fibers (Dummer et al., 30 Sep 2025).

6. Universal bundles and classifying constructions

A further extension comes from supergeometry, where vector bundles are organized by classifying spaces rather than by diffusion operators or function-space norms. The paper on $\mathcal M$01-grassmannians introduces $\mathcal M$02-domains equipped with an odd involution

$\mathcal M$03

and constructs the real $\mathcal M$04-grassmannian $\mathcal M$05 by gluing $\mathcal M$06-domains of dimension $\mathcal M$07 (Afshari et al., 2018). Over each chart, the canonical super vector bundle is

$\mathcal M$08

and the local pieces glue to a globally defined bundle

$\mathcal M$09

The universal property is expressed via a Gauss supermap. For a finite-type super vector bundle $\mathcal M$10, a finite trivializing cover and a partition of unity yield

$\mathcal M$11

represented by a Gauss supermatrix $\mathcal M$12. This induces a morphism

$\mathcal M$13

and the main theorem states

$\mathcal M$14

The paper also gives a homotopy classification theorem: after stabilization, the induced morphisms of Gauss supermaps are homotopic.

Although this is not a neural architecture, it supplies a universal-bundle interpretation of the term. This suggests a fourth sense of neural vector bundle: not a specific learning layer, but a classifying or parameterizing object for bundle-valued models. In that broader reading, the surveyed literature separates three principal questions. One asks how transport determines geometry, answered by the equivalence of superpath transport and connection (Dumitrescu, 2010). Another asks how signals and features evolve on bundles, answered by TNNs and BuNNs through Connection Laplacians and bundle heat diffusion (Battiloro et al., 2022, Bamberger et al., 2024). A third asks what function spaces underlie vector-valued neural maps, answered by vv-RKBS theory and its representer theorems (Dummer et al., 30 Sep 2025).