HilbNets: Banach Enrichment & Geometric Learning

• HilbNets are frameworks that integrate Banach-enriched symmetric monoidal multicategories with Hilbert space methods and bounded multilinear maps.
• They employ a functorial spectral theorem to extend self-adjoint operator calculus, ensuring universality and convergence from continuum to discrete settings.
• HilbNets support geometric deep learning using Hilbert bundles and sheaf Laplacians, enabling efficient, transferably validated networks for manifold-valued data.

HilbNets encompass two independent, but thematically related, frameworks in contemporary mathematics and machine learning: (1) a Banach-enriched symmetric monoidal multicategory formalism for Hilbert spaces and bounded multilinear maps (Chang, 17 Nov 2025), and (2) a geometric deep learning architecture for infinite-dimensional bundle-valued signals based on convolution with a connection Laplacian (Tandon et al., 7 May 2026). Both constructions place Hilbert space methods, operator theory, and categorical or geometric semantics at their core. This entry systematically surveys the formal constructions and results in both lines of work, emphasizing analytic, categorical, and learning-theoretic properties as well as empirical validations.

1. Banach-Enriched Symmetric Monoidal Multicategory Structure

HilbNets (denoted $\mathbf{HilbMult}$) are defined as a symmetric monoidal multicategory enriched over Banach spaces (Chang, 17 Nov 2025). The objects are complex Hilbert spaces $H, K, \ldots$. For any finite list of Hilbert spaces $(H_1,\ldots,H_n)$ and a target object $K$, the multimorphisms $\mathrm{Hom}(H_1,\ldots,H_n;K)$ are the spaces of bounded $n$-linear maps:

$$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$

each equipped with the operator norm

$$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$

making every hom-set a Banach space. Composition is defined by substitution of multilinear maps:

$$(S \circ (T_1,\ldots,T_m))(x_{1,1},\ldots,x_{m,n_m}) = S\bigl(T_1(x_{1,1},\ldots),\,\ldots,\,T_m(\ldots,x_{m,n_m})\bigr).$$

This operation is contractive ($\|S\circ(T_j)\|\leq\|S\|\prod_j\|T_j\|$), identities satisfy $H, K, \ldots$0, and associativity/unitality hold up to multicategory coherence isomorphisms.

The symmetric monoidal structure is given by the completed Hilbert-space tensor product $H, K, \ldots$1 (unit object is $H, K, \ldots$2), with $H, K, \ldots$3 acting on morphisms as

$H, K, \ldots$4

where the operator norm is multiplicative, $H, K, \ldots$5. Structural isomorphisms—associators, unitors, and symmetries—are given by isometries.

The Banach enrichment is certified by norm completeness (Lemma 3.1), and composition/tensoring are multilinear contractions, making $H, K, \ldots$6 a Banach-enriched symmetric monoidal multicategory.

2. Functorial Spectral Theorem and Universality

The functorial spectral theorem provides a categorical lift of the self-adjoint operator functional calculus. Given a bounded self-adjoint operator $H, K, \ldots$7, and a compatible family of $H, K, \ldots$8-linear vertex operators $H, K, \ldots$9 (satisfying spectral locality and compatibility with polynomial calculus), there exists a unique Banach-enriched symmetric monoidal multifunctor

$(H_1,\ldots,H_n)$0

such that, for polynomials $(H_1,\ldots,H_n)$1,

$(H_1,\ldots,H_n)$2

Functoriality is ensured by the compatibility conditions. When $(H_1,\ldots,H_n)$3 represents pointwise multiplication in the spectral representation $(H_1,\ldots,H_n)$4, the multifunctor recovers the standard continuous functional calculus $(H_1,\ldots,H_n)$5 and extends to all continuous functions via uniform limits.

The universality (Proposition 6.2) asserts that $(H_1,\ldots,H_n)$6 is the universal Banach-enriched symmetric monoidal multicategory supporting self-adjoint operator calculus: for any such multicategory $(H_1,\ldots,H_n)$7, every compatible family of data $(H_1,\ldots,H_n)$8 factors uniquely through a multifunctor from $(H_1,\ldots,H_n)$9.

3. Geometric Deep Learning with HilbNets

A distinct framework introduces HilbNets as spectral neural networks for infinite-dimensional data (Tandon et al., 7 May 2026). Here, a Hilbert bundle $K$0 (with fibers $K$1 separable real Hilbert spaces) models signals $K$2, enabling learning on data such as time series or distributions parameterized by a manifold.

A metric connection $K$3 on $K$4 induces a self-adjoint connection Laplacian

$K$5

which in local frames recovers $K$6. For any bounded Borel $K$7, the spectral calculus $K$8 defines a convolution operator on sections: $K$9.

This generalizes spectral graph convolutional networks, with the spectral profile $\mathrm{Hom}(H_1,\ldots,H_n;K)$0 parameterized as a learnable filter bank.

4. Sampling, Discretization, and Consistency

Implementation relies on a two-stage discretization. First, the manifold is sampled at $\mathrm{Hom}(H_1,\ldots,H_n;K)$1 points $\mathrm{Hom}(H_1,\ldots,H_n;K)$2, constructing a geometric graph $\mathrm{Hom}(H_1,\ldots,H_n;K)$3 with kernel weights

$\mathrm{Hom}(H_1,\ldots,H_n;K)$4

A Hilbert cellular sheaf $\mathrm{Hom}(H_1,\ldots,H_n;K)$5 assigns fiber spaces to nodes and geodesic midpoint fibers to edges, with restriction maps involving parallel transport induced by $\mathrm{Hom}(H_1,\ldots,H_n;K)$6. The sheaf Laplacian $\mathrm{Hom}(H_1,\ldots,H_n;K)$7 generalizes vector-valued Laplacians.

Signal discretization is performed by truncating to a $\mathrm{Hom}(H_1,\ldots,H_n;K)$8-dimensional orthonormal basis in each fiber, yielding a network sheaf $\mathrm{Hom}(H_1,\ldots,H_n;K)$9 and signal vectors in $n$0. The block matrix Laplacian acts as

$n$1

where $n$2 are finite-dimensional parallel transports.

The resulting implementable $n$3-HilbNet architecture propagates signals by applying polynomial spectral filters and nonlinearities layerwise:

$n$4

with $n$5 trainable weights.

5. Theoretical Guarantees: Convergence and Transferability

Several theorems establish analytic consistency of the discrete sheaf formulation with the underlying continuum theory (Tandon et al., 7 May 2026).

• Laplacian Convergence: Under suitable scaling of $n$6 and regularity, as $n$7, the rescaled sampled Laplacian $n$8 converges in probability, both pointwise and in $n$9 norm, to $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$0.
• Finite-Rank Approximation: There exists a deterministic sequence $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$1 such that the projected discrete Laplacian $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$2 converges in $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$3-mean to $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$4.
• Architectural Consistency: For any fixed filter bank and Lipschitz nonlinearity, the output of the discrete $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$5-HilbNet converges in mean-square $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$6 to the continuum HilbNet as $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$7.
• Transferability: For independent graph samples $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$8, the outputs of corresponding $$\mathrm{Hom}(H_1,\ldots,H_n; K) = \mathcal B(H_1 \times \cdots \times H_n, K),$$9-HilbNets become indistinguishable in $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$0 as $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$1, with explicit sample-independent error bounds.

6. Parametric Flexibility and Empirical Evaluation

The sheaf-theoretic HilbNet admits three parameterizations of the transport operators $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$2 on sheaf edges:

• Frozen identity: $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$3 (recovers standard GCN),
• Free Householder: $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$4 is a product of Householder reflections, fully expressing any $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$5 operator,
• Circulant: $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$6 is an orthogonal circulant, parameter-efficient for time-series.

Parameter counts per edge vary: identity $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$7, Householder $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$8, circulant $$\|f\| = \sup_{\|x_i\| \leq 1} \|f(x_1,\ldots,x_n)\|_K,$$9.

Experimental results include two major settings:

• Synthetic Transport Recovery: On $$(S \circ (T_1,\ldots,T_m))(x_{1,1},\ldots,x_{m,n_m}) = S\bigl(T_1(x_{1,1},\ldots),\,\ldots,\,T_m(\ldots,x_{m,n_m})\bigr).$$0 with Otto–Wasserstein geometry, the free Householder model achieves near-perfect recovery of ground-truth parallel transport, with MSE $$(S \circ (T_1,\ldots,T_m))(x_{1,1},\ldots,x_{m,n_m}) = S\bigl(T_1(x_{1,1},\ldots),\,\ldots,\,T_m(\ldots,x_{m,n_m})\bigr).$$1, exceeding the best attainable by circulant or identity transports.
• Traffic Forecasting: On METR-LA and PEMS-BAY datasets, HilbNets (with learned transports) consistently achieve lower MAE, RMSE, and MAPE than fiber-only MLP and baseline GCNs. Free $$(S \circ (T_1,\ldots,T_m))(x_{1,1},\ldots,x_{m,n_m}) = S\bigl(T_1(x_{1,1},\ldots),\,\ldots,\,T_m(\ldots,x_{m,n_m})\bigr).$$2 models attain the best performance, while circulant parametrization is nearly comparable yet significantly more parameter-efficient. Both HilbNets outperform FC-LSTM and approach STAEformer while using far fewer parameters (Tandon et al., 7 May 2026).

7. Significance and Connections

HilbNets as developed in both category-theoretic (Chang, 17 Nov 2025) and geometric learning (Tandon et al., 7 May 2026) contexts provide a rigorous analytic and compositional foundation for studying networks of Hilbert spaces and operators, with categorical semantics that canonically encompass operator algebra, quantum processes, and modern geometric learning architectures. The Banach-enriched multicategory formalism supplies universality, coherence, and a functorial spectral calculus for self-adjoint operators. The Hilbert bundle sheaf framework extends the scope of graph and manifold learning, allowing transferability and convergence guarantees for bundle-valued, possibly infinite-dimensional data. Together, these frameworks enable canonical links between functional analysis, category theory, operator representation, and data-centric learning methods.