The Frequency Reduced-Basis method: Reduced order models for time-dependent problems using the Laplace transform

Ricardo Reyes · 2025

Read The Frequency Reduced-Basis method: Reduced order models for time-dependent problems using the Laplace transform by Ricardo Reyes (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-2502-19120.txt.

Introduction

Ricardo Reyes speeds up time-dependent simulations by detouring through the Laplace transform. Solving an evolving equation at full size, step after step, is costly. His method first transforms the problem into a steady one that depends on a complex parameter. A set of sample solves at chosen parameter values then forms the raw material.

A compression step distills those samples into a small basis of representative shapes. The time problem is then solved inside that tiny space with any standard stepping scheme. Hardy space theory, through a classic theorem, justifies the swap. It implies the compressed basis is accurate to an exponential degree. Tests on several evolution equations measure the speed and error in practice. The work offers model reduction a new, frequency-first route.

Catalog record

CREATORRicardo Reyes
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE81 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-19120

Browse the full library · About the Tower of Babel

Full text

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

The Frequency Reduced-Basis method: Reduced order models for time-dependent problems using the Laplace transform Ricardo Reyes Abstract In this paper, we propose the Frequency Reduced-Basis method: a reduced-basis method to solve time-dependent partial differential equations based on the Laplace transform. Unlike traditional approaches, we begin by applying the Laplace transform to the evolution problem, yielding a time-independent boundary value problem that depends on the complex Laplace parameter. First, in an offline stage, we appropriately sample the Laplace parameter and solve the collection of problems using the finite element method. Next, we apply a Proper Orthogonal Decomposition (POD) to this collection of solutions in order to obtain a reduced basis that is of dimension much smaller than that of the original solution space. This reduced basis, in turn, is then used to solve the evolution problem using any suitable time-stepping method. A key insight to justify the formulation of the method relies on Hardy spaces of analytic functions. By applying the Paley-Wiener theorem we can then define an isometry between the solution of the time-dependent problem and its Laplace transform. As a consequence of this result, one may conclude that computing a POD with samples taken in the Laplace domain produces an exponentially accurate reduced basis for the time-dependent problem. Numerical experiments characterizing the performance of the method, in terms of accuracy and speed-up, are included for a variety of relevant time-evolution equations. 1 Introduction Solving multiple-query or forecasting problems often requires the solution of multiple parametric partial differential equations quickly and reliably. Unfortunately, numerical approximations are often either computationally expensive or insufficiently accurate. At this point, the development of surrogate models that are fast and reliable becomes necessary. Reduced basis methods are suitable surrogate models, where the original high-dimensional problem is replaced by a low-rank approximation. These methods consist of two stages: an offline part, where we construct a reduced basis from a collection of ‘high-fidelity’ solutions at selected sample points, and an online part, where we solve the low-dimensional problem defined as the projection of the original equation onto the subspace spanned by this reduced basis. Reduced bases for time-dependent problems are usually constructed using a standard POD [ 1 , 2 , 3 ] , which allows for a detailed data sampling. In problems where the solution can evolve without a clear pattern or at least without one that we can easily predict, we may need to sample data at a high frequency and over a long time window to be able to construct an appropriate reduced basis. If we also consider any other additional parametrization (for example physical or geometrical parameters), we may face issues sampling the data for large problems as they can become prohibitively expensive and time consuming. For general parametrised problems, a substantial body of work applies a Greedy sampling technique [ 4 , 5 , 6 ] , which decreases the computational burden of sampling. Although this is ideal for most parametric problems, the evolution behaviour of a time-dependent problem does not allow us to follow a random sampling in most cases. Therefore, when we deal with parametrised time-dependent problems, it is common to find techniques that have a mix of both methods: a Greedy approach to select the sample set in the parameter space, and either an impulse response [ 7 , 8 , 9 ] , or a regular interval sampling [ 10 , 11 , 12 , 13 , 3 ] in the time-domain. Although any of these approaches can give accurate results with a good enough sampling, we still face the same problem, a big computational cost for large problems. Thus, for real-world applications with computational and time constraints it is not always possible to construct an affordable and accurate reduced basis by following a traditional reduced-basis method. 1 1 1 Schleuß et al. [ 2 ] take an original approach to basis generation of the time-dependent problem, where rather than sampling over the global solution, the method captures a general behaviour of the problem by sampling randomly drawn starting points with random initial conditions in parallel. 2 2 2 It is worth noting the work in space-time approximations where the method aims to construct a reduced basis to project the entire space-time system, rather than solving a low-dimensional system for each time step [ 14 , 15 ] An alternative approach to building a reduced model consists of finding a surrogate for the transfer function in the frequency domain [ 16 , 17 , 18 , 3 ] . In this approach, mostly used in signal processing and control problems, the problem is written as an input-output linear time-invariant system, with a clearly defined transfer function. Then by using the Laplace transform we can convert the problem from the time domain to the frequency domain, find a surrogate for the transfer function and bring it back to the time domain using and inverse transform. (In this case it would be necessary to review some inverse Laplace transform methods [ 19 ] .) The major issue with this approach is how difficult and costly it can be to compute the numerical approximation of the inverse Laplace transform, which may be problem dependent and sometimes unattainable in practice. The Frequency Reduced-Basis method combines key ideas from frequency-domain surrogate modelling [ 16 , 20 ] and standard POD . While inspired by classical transfer-function approximations, our approach departs from them in that we avoid computing an inverse Laplace transform. Instead, we exploit the isometric structure of Hardy spaces to construct a reduced basis in the frequency domain that is then used to solve the original time-dependent problem. A summary of the method and its main advantages is presented in section 1.1 . We intend to explain why this method works using some elements of harmonic analysis. We start by defining the approximation spaces for each problem and the equivalence between their norms: a L^{2} Hilbert space for the time-domain problem and a H^{2} Hardy space for the frequency-domain problem. Then we map the frequency-domain problem, set in the complex half-plane, onto the unit circle. And lastly, we justify why we only use the real part of the solution as data input using the properties of analytic functions. We validate the method using three numerical examples: a wave equation and a heat equation to evaluate convergence, efficiency and performance, and a advection problem to test stability. 1.1 Method Summary In this work, we introduce the Frequency Reduced-Basis (FRB) method: a model reduction technique for time-dependent partial differential equations. The method uses the Laplace transform and the POD to build a model reduction framework that operates in the frequency domain during the offline stage and in the time domain during the online stage. Offline stage (frequency domain) We apply the Laplace transform to the time-dependent PDE, which yields a parametric elliptic boundary value problem in the complex Laplace domain. We then evaluate the solution at a selected set of complex frequencies and collect the corresponding snapshots. Finally, we construct a reduced basis using a Singular Value Decomposition (SVD) of these frequency-domain solutions. Online stage (time domain) We project the original time-dependent PDE onto the reduced subspace, following the standard reduced-basis procedure. Finally, we

Read the complete text