Neural Algorithmic Solvers for High-Dimensional Linear Systems

Fuente: Zenodo
Guardado en:
Detalles Bibliográficos
Autor principal: SÉRGIO DE ANDRADE, PAULO
Formato: Recurso digital
Publicado: Zenodo 2025
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866901789275586560
author SÉRGIO DE ANDRADE, PAULO
author_facet SÉRGIO DE ANDRADE, PAULO
contents High-dimensional linear systems are ubiquitous in scientific computing, engineering, and data science, arising from discretization of partial differential equations, statistical models, and large-scale optimization problems. Solving these systems efficiently and accurately presents significant computational challenges due due to the curse of dimensionality, memory constraints, and the inherent complexity of large-scale matrix operations. Traditional direct methods often become intractable, while iterative solvers, though more memory-efficient, can suffer from slow convergence or sensitivity to matrix conditioning. This paper explores the emerging paradigm of neural algorithmic solvers as a novel approach to tackle high-dimensional linear systems. We investigate how deep learning architectures, particularly those inspired by recurrent processes and graph neural networks, can learn to approximate or accelerate the solution process. The core idea is to leverage the approximation capabilities of neural networks to emulate, improve, or discover iterative solution algorithms that are robust, efficient, and scalable to high dimensions. We present a conceptual framework for designing such solvers, discussing potential architectures, training methodologies, and evaluation metrics. The envisioned benefits include reduced computational complexity, enhanced adaptability to various matrix properties, and the potential for superior generalization compared to traditional methods. Furthermore, we delve into the challenges associated with interpretability, theoretical guarantees, and the generation of appropriate training data for these data-driven solvers.
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_17687854
institution Zenodo
language
publishDate 2025
publisher Zenodo
record_format zenodo
spellingShingle Neural Algorithmic Solvers for High-Dimensional Linear Systems
SÉRGIO DE ANDRADE, PAULO
High-dimensional linear systems are ubiquitous in scientific computing, engineering, and data science, arising from discretization of partial differential equations, statistical models, and large-scale optimization problems. Solving these systems efficiently and accurately presents significant computational challenges due due to the curse of dimensionality, memory constraints, and the inherent complexity of large-scale matrix operations. Traditional direct methods often become intractable, while iterative solvers, though more memory-efficient, can suffer from slow convergence or sensitivity to matrix conditioning. This paper explores the emerging paradigm of neural algorithmic solvers as a novel approach to tackle high-dimensional linear systems. We investigate how deep learning architectures, particularly those inspired by recurrent processes and graph neural networks, can learn to approximate or accelerate the solution process. The core idea is to leverage the approximation capabilities of neural networks to emulate, improve, or discover iterative solution algorithms that are robust, efficient, and scalable to high dimensions. We present a conceptual framework for designing such solvers, discussing potential architectures, training methodologies, and evaluation metrics. The envisioned benefits include reduced computational complexity, enhanced adaptability to various matrix properties, and the potential for superior generalization compared to traditional methods. Furthermore, we delve into the challenges associated with interpretability, theoretical guarantees, and the generation of appropriate training data for these data-driven solvers.
title Neural Algorithmic Solvers for High-Dimensional Linear Systems
url https://doi.org/10.5281/zenodo.17687854