Tensor-based empirical interpolation method and its application in model reduction

TL;DR

This paper introduces a tensor-based empirical interpolation method for matrix-valued functions that reduces computational effort while maintaining accuracy, extending existing methods like EIM and DEIM.

Contribution

It develops a tensor-based interpolation approach that avoids vectorization, providing theoretical insights and demonstrating efficiency in model reduction tasks.

Findings

The method generates interpolation points on a rectangular grid.

It is equivalent to applying DEIM in each tensor direction.

The proposed method is faster with minor accuracy loss.

Abstract

In general, matrix or tensor-valued functions are approximated using the method developed for vector-valued functions by transforming the matrix-valued function into vector form. This paper proposes a tensor-based interpolation method to approximate a matrix-valued function without transforming it into the vector form. The tensor-based technique has the advantage of reducing offline and online computation without sacrificing much accuracy. The proposed method is an extension of the empirical interpolation method (EIM) for tensor bases. This paper presents a necessary theoretical framework to understand the method's functioning and limitations. Our mathematical analysis establishes a key characteristic of the proposed method: it consistently generates interpolation points in the form of a rectangular grid. This observation underscores a fundamental limitation that applies to any matrix-based approach relying on widely used techniques like EIM or DEIM method. It has also been theoretically shown that the proposed method is equivalent to the DEIM method applied in each direction due to the rectangular grid structure of the interpolation points. The application of the proposed method is shown in the model reduction of the semi-linear matrix differential equation. We have compared the approximation result of our proposed method with the DEIM method used to approximate a vector-valued function. The comparison result shows that the proposed method takes less time, albeit with a minor compromise with accuracy.