Advancements in Fractional Neural Operators with Adaptive Hybrid Kernels in Multiscale Sobolev Spaces

Romulo Damaselin Chaves dos Santos, Jorge Henrique de Oliveira Sales · 2025

Read Advancements in Fractional Neural Operators with Adaptive Hybrid Kernels in Multiscale Sobolev Spaces by Romulo Damaselin Chaves dos Santos, Jorge Henrique de Oliveira Sales (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-11680.txt.

Introduction

Romulo Damaselin Chaves dos Santos and Jorge Henrique de Oliveira Sales set out new math in a 2025 paper on neural operators, models that learn to solve hard equations. Their work adds kernels that adapt and a way to weigh noise across scales.

The authors prove four main results. One splits bias from spread for a class of spread-out operators. One shows the models stay stable under heavy-tailed noise. One beats the growth of cost as the number of inputs grows. One gives a faster search method drawn from quantum ideas. As a test case, they apply the tools to plasma swirls in fusion work and report a large gain over the best prior models.

Catalog record

CREATORRomulo Damaselin Chaves dos Santos, Jorge Henrique de Oliveira Sales
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-11680

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Advancements in Fractional Neural Operators with Adaptive Hybrid Kernels in Multiscale Sobolev Spaces Romulo Damaselin Chaves dos Santos, Jorge Henrique de Oliveira Sales Abstract This paper introduces significant advancements in fractional neural operators (FNOs) through the integration of adaptive hybrid kernels and stochastic multiscale analysis. We address several open problems in the existing literature by establishing four foundational theorems. First, we achieve exact bias-variance separation for Riemann-Liouville operators using fractional Prokhorov metrics, providing a robust framework for handling non-local dependencies in differential equations. Second, we demonstrate Lévy-regularized Hadamard stability with sharp convergence rates in Besov-Morrey spaces, enhancing FNO stability under heavy-tailed noise processes. Third, we overcome the curse of dimensionality in multivariate settings by achieving tensorized RL-Caputo convergence in RdRd with anisotropic Hölder regularity. Finally, we develop a quantum-inspired fractional gradient descent algorithm that significantly improves convergence rates in practical applications. Our proofs employ advanced techniques such as multiphase homogenization, Malliavin-Skorokhod calculus, and nonlocal divergence theorems, ensuring mathematical rigor and robustness. The practical implications of our work are demonstrated through applications in fusion plasma turbulence, where our methods yield a 70% improvement over state-of-the-art FNOs. This enhancement is particularly notable in modeling the multifractal δBδB fields in ITER tokamak data, where the anisotropic penalty corresponds to enstrophy cascade rates in Hasegawa-Wakatani models, showcasing its versatility and potential for broader application. Keywords: Fractional Neural Operators (FNOs). Adaptive Hybrid Kernels. Stochastic Multiscale Analysis. Lévy-Regularized Stability. 1 Introduction Fractional neural operators (FNOs) have emerged as a transformative approach for addressing non-local differential equations in scientific machine learning. Building upon the pioneering work of [ 3 ] , who established convergence rates for Caputo-activated networks, and [ 1 ] , who introduced the Voronovskaya-Damaselin theorem for symmetrized operators, the field has witnessed rapid theoretical advancements. However, several critical limitations persist in three key areas. The seminal work by [ 2 ] on Riemann-Liouville (RL) derivatives revealed an intrinsic approximation bias \mathcal{O}(n^{-\alpha}) when handling constant functions. Conversely, [ 4 ] demonstrated the superior stability properties of Hadamard operators in log-scale spaces. Recent efforts by [ 1 ] attempted to hybridize these approaches through static weighting schemes but failed to address spatial heterogeneity in fractional smoothness \alpha(x) , heavy-tailed noise resilience in infinite-dimensional settings, and the curse of dimensionality for multivariate operators. Our work addresses these limitations through four fundamental innovations. First, we introduce adaptive hybrid kernels, which are dynamically weighted RL-Caputo operators with \theta(x) -dependent blending (Theorem 4.1 ), effectively eliminating the constant-bias problem identified by [ 2 ] . Second, we establish Lévy-regularized Hadamard stability, providing the first convergence guarantees under \alpha -stable noise processes (Sec. 3 ), thereby extending [ 3 ] ’s Gaussian stability results. Third, we develop tensorized multivariate operators that achieve dimension-independent rates through anisotropic Hölder adaptation (Eq. 3.1 ), overcoming the \mathcal{O}(n^{-k/d}) bottleneck observed in [ 1 ] . The urgent need for such advances is underscored by one emerging application. In fusion energy, ITER tokamak data reveals multifractal \delta B fields that require \alpha(x) -adaptive operators for confinement modeling. Here, the anisotropic penalty \int|\nabla\alpha|^{2}/\alpha^{5/2}dx corresponds to enstrophy cascade rates in Hasegawa-Wakatani models. Our kernel \Phi_{H}^{L} effectively matches empirical jump intensities with \text{Var}[\Phi_{H}^{L}]\sim\mathcal{O}(t^{2/\alpha-1}) , showcasing its robustness and potential for broader applications. 2 Mathematical Background In this section, we establish the mathematical foundation necessary for understanding the subsequent theorems and their proofs. We begin by reviewing key concepts in fractional calculus, functional analysis, and stochastic processes. 2.1 Fractional Calculus Fractional calculus extends the classical calculus to non-integer orders, allowing for the differentiation and integration of functions to any real or complex order. The Riemann-Liouville (RL) and Caputo fractional derivatives are fundamental to this field. Riemann-Liouville Fractional Derivative: For a function f\in L^{1}([a,b]) , the RL fractional derivative of order \alpha is defined as: D_{RL}^{\alpha}f(x)=\frac{1}{\Gamma(1-\alpha)}\frac{d}{dx}\int_{a}^{x}\frac{f(t)}{(x-t)^{\alpha}}\,dt, (2.1) where \Gamma is the Gamma function. Caputo Fractional Derivative: The Caputo fractional derivative of order \alpha is defined as: D_{*}^{\alpha}f(x)=\frac{1}{\Gamma(1-\alpha)}\int_{a}^{x}\frac{f^{\prime}(t)}{(x-t)^{\alpha}}\,dt. (2.2) 2.2 Functional Analysis Functional analysis provides the tools necessary to study spaces of functions and operators acting on these spaces. Key concepts include normed spaces, Banach spaces, and Sobolev spaces. Sobolev Spaces: For \Omega\subset\mathbb{R}^{n} and s\in\mathbb{R} , the Sobolev space W^{s,p}(\Omega) consists of functions f such that: \|f\|_{W^{s,p}(\Omega)}=\left(\int_{\Omega}|(I-\Delta)^{s/2}f(x)|^{p}\,dx\right)^{1/p}<\infty, (2.3) where (I-\Delta)^{s/2} is the Bessel potential operator. 2.3 Stochastic Processes Stochastic processes, particularly Lévy processes, are essential for modeling systems with random fluctuations. Lévy Process: A Lévy process L_{t} is a stochastic process with independent and stationary increments. It is characterized by a triplet (a,\sigma^{2},\nu) , where \nu is the Lévy measure. 2.4 Multiphase Homogenization Multiphase homogenization provides the foundation for resolving scale-separated phenomena in fractional operators. By decomposing the domain \Omega into microscale ( \epsilon -periodic) and macroscale components, we derive effective hybrid kernels \Phi_{H}^{L} that adaptively weight Riemann-Liouville (RL) and Caputo derivatives (Eq. 3.1 ). This approach rigorously handles heterogeneous fractional orders \alpha(x) through variational principles in negative Hölder spaces, ensuring consistency across scales. The homogenized operator D_{AH}^{\alpha(x)} eliminates the constant-bias problem by dynamically balancing nonlocal interactions (Theorem 4.1 ). 2.5 Malliavin-Skorokhod Calculus To stabilize solutions under Lévy-driven noise, we employ Malliavin-Skorokhod calculus for infinite-dimensional jump processes. This framework quantifies the sensitivity of stochastic fractional integrals to Lévy measures \nu(ds)\sim s^{-1-\gamma}ds , enabling the regularization of Hadamard kernels (Eq. 3.5 ). Specifically, the Skorokhod integral bounds the variance of \Phi_{H}^{L} under \alpha -stable perturbations, while Malliavin derivatives control anisotropic Hölder singularities in \mathbb{R}^{d} (Theorem 4.4 ). 2.6 Nonlocal Divergence Theorems Nonlocal divergence theorems generalize Gauss-Green identities to fractional Sobolev spaces W^{\alpha,p}(\Omega) . For tensorized kernels, these theorems enable dimension-independent convergence by reconciling anisotropic smoothness with the nonlocal flux conditions: \int_{\mathbb{R}^{d}}\nabla^{\alpha}\cdot\mathbf{F}\,dx=\oint_{\partial\Omega}\mathbf{F}\cdot\mathbf{n}^{\alpha}\,dS, (2.4) where \mathbf{n}^{\alpha} is the fractional co-normal vector. This underpins the \mathcal{O}(n^{-2+\alpha}) -optimality of our quantum-inspired gradient descent algorithm. 2.7 ITER Tokamak Data & Enstrophy Cascade In fusion plasma turbulence, ITER tokamak measurements reveal multifractal \delta B fields governed by Hasegawa-Wakatani models. The anisotropic penalty term \int|\nabla\alpha|^{2}/\alpha^{5/2}dx in Eq. 4.1 corresponds to enstrophy cascade rates \epsilon\sim\nu^{3/2}k_{\perp}^{2} , where k_{\perp} is the perpendicular wavenumber. Our adaptive kernels \Phi_{H}^{L} resolve this by aligning the fractional order \alpha(x) with the local turbulent intensity, matching empirical cascade spectra within 5% error (Sec. 5 ). These tools collectively address the open challenges in FNOs, bridging rigorous analysis with critical applications in energy and finance. 3 Mathematical Foundations 3.1 Adaptive Hybrid Derivatives in Negative Hölder Spaces We introduce adaptive hybrid derivatives, which combine the strengths of Riemann-Liouville (RL) and Caputo fractional derivatives to address spatial heterogeneity in fractional smoothness. Let \alpha\in C^{1,\delta}(\Omega) with \delta>1/2 and 0<\alpha_{0}\leq\alpha(x)\leq\alpha_{1}<1 . The adaptive hybrid derivative of a function f is defined as: D_{AH}^{\alpha(x)}f:=\theta(x)D_{RL}^{\alpha(x)}f+(1-\theta(x))D_{*}^{\alpha(x)}f, (3.1) where the RL and Caputo components are: \displaystyle D_{RL}^{\alpha(x)}f(x) \displaystyle:=\frac{1}{\Gamma(1-\alpha(x))}\frac{d}{dx}\int_{x-\epsilon}^{x}\frac{f(y)}{(x-y)^{\alpha(x)}}dy\quad(\epsilon>0), (3.2) \displaystyle D_{*}^{\alpha(x)}f(x) \displaystyle:=\frac{1}{\Gamma(1-\alpha(x))}\int_{x}^{\infty}\frac{f^{\prime}(y)}{(y-x)^{\alpha(x)}}dy. (3.3) The weighting function \theta(x) balances the contributions of the RL and Caputo derivatives: \theta(x)=\frac{\|D_{RL}^{\alpha(x)}f\|_{\mathcal{M}}}{\|D_{RL}^{\alpha(x)}f\|_{\mathcal{M}}+\|D_{*}^{\alpha(x)}f\|_{L^{1}}}, (3.4) where \|g\|_{\mathcal{M}} denotes the norm in the space of Radon measures. This adaptive approach captures both local and non-local behaviors of f , making it robust for heterogeneous fractional smoothness. 3.2 Lévy-Regularized Hadamard Kernels Lévy processes generalize Brownian motion by allowing jumps and heavy tails. We use a Lévy subordinator L_{t} with triplet (0,\sigma^{2},\nu) , where \nu(ds)\sim s^{-1-\gamma}ds , to construct regularized Hadamard kernels. The regularized Hadamard kernel is: \Phi_{H}^{L}(z):=\mathbb{E}_{L_{t}}\left[\int_{0}^{\infty}(\ln(1+|z|+s))^{-\alpha}\nu(ds)\right], (3.5) where \mathbb{E}_{L_{t}} denotes the expectation with respect

Read the complete text