Kernel methods join the Deep Ritz method with new error guarantees

Hendrik Kleikamp and Tizian Wenzel introduce a way to combine kernel methods with the Deep Ritz method for approximating weak solutions of elliptic partial differential equations. The peer-reviewed, open-access paper was published on 22 October 2026 in a Springer Nature journal covering computational science and engineering, and it appears as an early-release version that is citable and carries a permanent DOI, subject to further editorial edits before the final version of record.

The core idea is to use kernel approximations as ansatz functions inside the Deep Ritz method. In this setting, the authors approximate weak solutions of elliptic partial differential equations while enforcing boundary conditions weakly through Nitsche’s method, a well-known technique in numerical analysis. Rather than relying on neural networks as the function class, the approach draws on kernel interpolation, which brings a different set of theoretical tools to bear on the problem.

A central contribution of the paper is a set of a priori error estimates, proven in different norms. The authors derive these estimates by combining standard results for weak solutions of elliptic equations with established convergence results for kernel methods. According to the abstract, the availability of such error estimates is what renders the method useful for practical purposes where theoretical guarantees are important, since practitioners can know in advance how the approximation error is expected to behave.

Beyond the theory, the paper describes the procedure in detail and provides practical hints and implementation details. The authors then evaluate the performance of the proposed approach through numerical examples, which they report also serve to verify the theoretical findings. In a high-dimensional test case, they apply two-layered kernels, a configuration that the authors note goes beyond a purely linear method.

One of the paper’s most distinctive claims concerns comparison. To the best of the authors’ knowledge, it contains the first extensive numerical studies and discussions comparing four approaches side by side: kernel interpolation, the kernel Galerkin method, the Deep Ritz method with kernels, and the classical Deep Ritz method with neural networks. Through these experiments, the authors numerically verify advantages and disadvantages of the particular methods and show under which assumptions each individual approach is beneficial.

Related open-access articles listed by the publisher include work on solving approximation tasks with greedy deep kernel methods, the Deep Ritz method for parametric p-Dirichlet problems, and adaptive importance sampling for Deep Ritz, indicating sustained interest in machine-learning-based solvers for partial differential equations.

By replacing them with kernels while retaining weak enforcement of boundary conditions via Nitsche’s method, the new approach aims to keep the machinery of the method intact while adding convergence guarantees of the kind kernel methods are known for.

The paper is currently available as an accepted-manuscript share, which the publisher notes is peer-reviewed, citable, and carries a permanent DOI, with the final version of record to replace it automatically. Readers can consult the full open-access article, including the numerical studies comparing the four methods, through the journal’s website.

Subject of Research: Kernel methods as ansatz functions in the Deep Ritz method for elliptic partial differential equations, with a priori error estimates and numerical comparisons.

Article Title: Kernel methods in the deep Ritz framework: theory and practice

Article References: Kleikamp, H., & Wenzel, T. (2026). Kernel methods in the deep Ritz framework: theory and practice. Computational Science and Engineering. https://doi.org/10.1007/s44207-026-00019-6

Image Credits: AI Generated

DOI: 10.1007/s44207-026-00019-6

Keywords: Deep Ritz method, kernel methods, elliptic partial differential equations, Nitsche's method, a priori error estimates, weak solutions, kernel Galerkin method, neural networks, numerical analysis, machine learning, high-dimensional PDEs, open access