Brownian Kernel Ladders

Overview

Brownian Kernel Ladders (BKLs) are a depth-indexed hierarchy of function spaces. Starting from linear functionals, each layer is an integral RKHS whose kernel is assembled from Brownian pullbacks of functions in the preceding layer. Depth is therefore encoded directly in the geometry of the function spaces, rather than through a fixed finite-dimensional network parameterization.

The framework is designed around three requirements that are rarely available simultaneously: an analytically explicit recursive construction, function classes that genuinely become richer with depth, and a statistical theory that records the cost of choosing a representation instead of treating that choice as free.

Core construction

The scalar Brownian kernel is

\[ k_{B}(s,t) = \frac{|s|+|t|-|s-t|}{2}, \qquad s,t\in\mathbb{R}. \]

The first layer is the RKHS of the linear kernel. Given a realized layer \(\mathcal H^{(\ell)}\), a probability measure \(\mu^{(\ell)}\) on a family of functions from that layer generates the next kernel through

\[ k^{(\ell)}(x,x') = \int k_B\!\left(u(x),u(x')\right) \,\mathrm d\mu^{(\ell)}(u), \]

and the next space is the RKHS associated with this kernel. Repeating the operation produces a ladder \(\mathcal H^{(1)},\ldots,\mathcal H^{(L)}\).

The nonnegative one-homogeneity of the Brownian kernel permits a kernel-preserving normalization of every layer measure onto the unit sphere of the preceding RKHS. Its square-root feature geometry also propagates explicit regularity and pointwise bounds through depth.

One framework, two mathematical levels

A central feature of the theory is the separation between the full representational family and the statistical model used for estimation.

Adaptive canonical envelope

All recursively admissible canonical ladder measures are allowed to vary. A function is assigned the infimum of its top-layer RKHS norm over every ladder that realizes it. This is the natural object for studying global geometry, regularity, expressivity, and variational structure.

Realized and controlled models

One ladder, or a finite dictionary of ladders, is fixed independently of the estimation sample. The corresponding union of top-layer RKHS balls is the statistical model. Selection among candidate ladders then has an explicit and measurable cost.

Quantifier matters. A supremum over representations outside a Gaussian expectation is not the same as unrestricted adaptive selection inside that expectation. The structural results apply to the full adaptive envelope; the near-parametric statistical guarantees apply to realized ladders and finite controlled dictionaries fixed before the estimation sample.

Main results

Structure of the adaptive envelope

Every depth-\(L\) function satisfies a \(2^{-(L-1)}\)-Hölder estimate and a corresponding pointwise bound, controlled by the infimal BKL complexity.
The adaptive envelope equipped with this complexity is a quasi-Banach space.
The hierarchy is nested. Under a geometric trace condition, every additional layer strictly enlarges the function class, and a separating function remains at positive uniform distance from each fixed-radius ball of the preceding level.
Regularized empirical-risk minimization over the adaptive envelope attains a solution for continuous losses that are uniformly bounded below.
For losses strictly convex in the prediction, population minimizers agree almost everywhere with respect to the input distribution; full support upgrades this to pointwise uniqueness.

Statistics of controlled ladder families

Let a dictionary of \(M\) canonical ladders be fixed independently of the estimation sample, and let the statistical class be the union of their radius-\(r\) top-layer RKHS balls. Its Gaussian complexity obeys

\[ \mathcal G_n \le \frac{r\left(1+\sqrt{2\log M}\right)}{\sqrt n} \left( \mathbb E\!\left[ \max_{1\le i\le n}\|X_i\|_2^2 \right] \right)^{2^{-L}}. \]

The estimate has the standard \(n^{-1/2}\) dependence for one realized ladder.
It contains no explicit ambient-dimension factor.
Selection among \(M\) candidate ladders costs the explicit factor \(1+\sqrt{2\log M}\).
Radius-constrained controlled estimators satisfy high-probability penalized oracle and excess-risk bounds.
Polynomial-size dictionaries retain a near-parametric \(\widetilde{O}(n^{-1/2})\) rate.

Interpretation

BKLs provide a two-level foundation for hierarchical kernel learning. The adaptive envelope describes what can be represented when all canonical ladder measures are allowed to vary. Realized ladders and controlled dictionaries describe what can be estimated with an accountable representation-selection cost.

This distinction prevents an overly broad conclusion: the \(n^{-1/2}\) complexity bound is not asserted for unrestricted sample-adaptive optimization over the entire envelope. It is proved for one ladder or a finite family chosen independently of the estimation sample. The full envelope instead supports the analytical and variational theory.

Under the geometric trace condition, depth is not merely a repeated reparameterization: it strictly increases representational power. At the same time, within controlled realized families, statistical complexity does not accumulate through a product of layerwise norm factors.

Connections

At depth two, the construction recovers the Brownian projection kernel used in Brownian kernel-enhanced random neural networks when the first layer is identified with Euclidean linear functionals.
BKLs share a recursive integral-RKHS architecture with neural Hilbert ladders, but use the same explicit Brownian pullback kernel at every layer.
The framework is related to deep RKBSs, chain RKBSs, Barron spaces, and recursive Banach-space models, while retaining an RKHS at each realized layer and an explicit separation between adaptive representation and controlled estimation.

Research directions

universality, approximation rates, and depth-dependent representation efficiency;
entropy conditions and complexity penalties for richer or infinite ladder families;
data-dependent ladder construction through independent pilot samples;
finite-dimensional approximations, sparse dictionaries, and scalable optimization;
links between hierarchical function-space geometry and practical representation learning.

Manuscript and citation

Brownian Kernel Ladders
Mahdi Mohammadigohari, Giuseppe Di Fatta, Giuseppe Nicosia, and Panos M. Pardalos.
arXiv:2606.15812

@article{mohammadigohari2026brownian, title = {Brownian Kernel Ladders}, author = {Mohammadigohari, Mahdi and Di Fatta, Giuseppe and Nicosia, Giuseppe and Pardalos, Panos M.}, year = {2026}, eprint = {2606.15812}, archivePrefix = {arXiv} }