Variation Brownian Kernel Ladders

Overview

Variation Brownian Kernel Ladders (VBKLs) are a path-atomic function-space model for hierarchical learning. The construction first builds a nonlinear recursive dictionary of compositional Brownian paths. Linear superposition is introduced only after the target depth has been reached, through a finite signed measure over the completed dictionary.

This order separates three objects that are often combined: the geometry of one recursive atom, the total variation needed to combine atoms, and the finite architecture used to approximate or learn the resulting function. The placement of superposition is therefore part of the mathematical definition, not merely an implementation choice.

Path-atomic construction

The scalar Brownian kernel is

\[ k_B(s,t) = \frac{|s|+|t|-|s-t|}{2}. \]

Its RKHS \(\mathcal H_B\) is the anchored Cameron–Martin space of absolutely continuous functions satisfying \(g(0)=0\) and \(g'\in L^2\), with squared norm \(\|g\|_{\mathcal H_B}^2=\int |g'(t)|^2\,\mathrm dt\).

Starting from an admissible set \(\Omega\) of first-layer directions, the recursive dictionaries are

\[ \mathcal U_1 = \left\{x\mapsto \omega^{\top}x:\omega\in\Omega\right\}, \qquad \mathcal U_\ell = \left\{g\circ u: u\in\mathcal U_{\ell-1}, \ g\in\mathcal H_B, \ \|g\|_{\mathcal H_B}\le 1 \right\}. \]

Thus a depth-\(L\) atom follows one first-layer projection through exactly \(L-1\) normalized Brownian profiles. Each dictionary is also exactly a union of unit balls of Brownian pullback RKHSs.

The depth-\(L\) variation space and its intrinsic complexity are

\[ \mathcal V^{(L)} = \left\{ F=\int_{\mathcal U_L}u\,\mathrm d\mu(u): \mu\text{ is a finite signed measure} \right\}, \] \[ \widehat{\mathcal C}^{(L)}_{\mathrm{var}}(F) = \inf\left\{ \|\mu\|_{\mathrm{TV}}: F=\int_{\mathcal U_L}u\,\mathrm d\mu(u) \right\}. \]

Nonlinear recursion constructs the atoms. Only the final signed measure introduces linear variation. No convexification or measure superposition is inserted at the intermediate layers.

Main results

Geometry of the full VBKL space

The variation complexity is well defined and nondegenerate, and it controls both pointwise magnitude and \(2^{-(L-1)}\)-Hölder regularity.
Under compactness of the input domain and first-layer direction set, the recursive dictionary is compact and a minimum-total-variation representation is attained.
When the input measure has full support, the Hölder representative is unique and gives a continuous embedding into the corresponding Hölder space.
Consecutive VBKL spaces are nested. Under the stated local nondegeneracy condition, with the relevant trace lying in the support of the input measure, the inclusion is strict.

Statistical guarantees for explicit finite architectures

The paper also defines associated finite lower-support architectures with piecewise-linear Brownian profiles and normalized mixing. For their radius-\(R\) outer variation classes, empirical and expected Rademacher bounds are obtained through Brownian quadratic chaos, signed threshold traces, and VC entropy. The estimates separate:

the outer total-variation radius;
the range scale induced by recursive Brownian geometry;
the parameter complexity of the finite lower-support architecture;
the sample size.

Corresponding high-probability generalization guarantees are proved for these finite-architecture Brownian variation classes.

Constructive approximation of the full space

Every element of the full VBKL space admits a two-stage approximation. First, the outer signed measure is replaced by at most \(M\) recursive atoms. Second, only the selected atoms' outer Brownian profiles are replaced by piecewise-linear interpolants of resolution \(m\). The lower-level supports remain unchanged.

\[ \|F-F_{M,m}\|_{L^2(\nu)} \le \widehat{\mathcal C}^{(L)}_{\mathrm{var}}(F) \left( R_L M^{-1/2} + \sqrt{\frac{A_0}{2}}\,m^{-1/2} \right). \]

The two terms measure different resources: outer-measure discretization and outer-profile interpolation.
Choosing \(M=m=N\) gives a balanced \(O(N^{-1/2})\) approximation rate under the stated fixed quantities.
The uniform Brownian-profile interpolation constant \(\sqrt{A/2}\) and the \(m^{-1/2}\) rate are sharp.
Evaluation uses at most \(2M\) active outer-profile hat-basis contributions, independently of the profile resolution \(m\).

Scope of the guarantees

Full path-atomic VBKL space

The regularity, compactness, attainment, strict-depth, and constructive approximation results concern the full infinite-dimensional variation space generated by signed measures over the recursive path dictionary.

Associated finite architectures

The Rademacher and generalization bounds concern explicitly parameterized finite Brownian variation classes. Their lower supports may contain intermediate linear mixing.

The distinction is essential. The general mixed finite architecture is not identified with the full path-atomic VBKL ball. Its path-only specialization is a subclass of the full space, but no such inclusion is asserted for the general mixed architecture. Finite-parameter statistical bounds are therefore not attributed to the unrestricted infinite-dimensional variation ball.

Relation to Brownian Kernel Ladders

Brownian Kernel Ladders and VBKLs use the same Brownian function-generation mechanism but organize representation in different ways. A BKL layer averages Brownian pullback kernels over a probability measure and produces a new integral RKHS. VBKL instead follows individual compositional supports to build a nonlinear path dictionary, then applies a signed-measure variation hull only at the outermost level.

The two frameworks are therefore complementary: BKL emphasizes recursive Hilbert-space geometry and the separation between adaptive envelopes and controlled realized ladders, while VBKL emphasizes the geometry and finite approximation of path-atomic variation spaces.

Theory-directed computational evidence

The experiments are organized around the mechanisms in the theory. They separately test signed-measure discretization, Brownian-profile interpolation, balanced refinement, finite-model optimization, and controlled statistical learning. A normalized tent construction attains the sharp worst-case interpolation bound.

The supervised studies indicate favorable behavior in limited-data and parameter-efficiency regimes. They are presented as controlled evidence for the proposed mechanisms, not as a claim of universal predictive dominance or global optimization convergence.

Manuscript and citation

Variation Brownian Kernel Ladders
Mahdi Mohammadigohari.
arXiv:2608.13882

@article{mohammadigohari2026variation, title = {Variation Brownian Kernel Ladders}, author = {Mohammadigohari, Mahdi}, year = {2026}, eprint = {2608.13882}, archivePrefix = {arXiv} }