We claim the following innovations
We propose a generalized realization form for rational functions in n-variables (for any "n"), which are described in the Lagrange basis;
We show that the n-dimensional Loewner matrix can be written as the solution of a series of cascaded Sylvester equations;
We demonstrate that the required variables to be determined, i.e. the barycentric coefficients, can be computed using a sequence of small-scale 1-dimensional Loewner matrices instead of the large-scale (NxN) n-dimensional one, therefore drastically reducing the both computational effort and memory needs, and improving accuracy;
We show that this decomposition achieves variables decoupling; thus connecting the Loewner framework for rational interpolation of multivariate functions and the Kolmogorov Superposition Theorem (KST), restricted to rational functions. The result is the formulation of KST for the special case of rational functions;
Connections with KAN neural nets follows (detailed in future work).
This page provides supplementary materials to the paper entitled "On the Loewner framework, the Kolmogorov superposition theorem, and the curse of dimensionality", by A.C. Antoulas, I-V. Gosea and C. Poussot-Vassal (available as arXiv).
We provide additional (regularly updated) numerical benches and results, grounded on the above contribution. Next, we treat problems of three different nature:
Approximation from an available function (set #1):
the multi-variables function H(s1,s2,...sn) - either rational or non-rational / complex or real-valued - is available.
Approximation from an available tensor data (set #2):
the (large-scale) n-D tensor data is available. In this configuration, the multi-variables function H(s1,s2,...sn) is unknown, but its evaluation at some points has been computed, and is provided. The n-D tensor denotes the evaluation of H at some points, thus leading to a n-dimensional matrix.
Nonlinear eigenvalue approximation (set #3):
The objective, based on one of the above ingredients, is to construct a rational model that matches / approximates the data.
The proposed data-sets consist of a collection of examples, either constructed by us or proposed in the literature (other references, MOR Wiki...). For each case, we provide a ZIP file which contains data in Matlab format ".mat" files. More specifically ("X" denotes the name or the number of the experiment)
"tab_X.mat": contains the n-D tensor. More specifically, it contains
tab: the n-D tensor
p_c: the column interpolation points
p_r: the row interpolation points
"model_X.mat": contains the statistics and model, all gathered in the structure "stat". More specificalkly, it contains the following fields
H: the handle function
n: the number of variables
ord: the order of each variables
p_c: the interpolation points
p_r: the (row) interpolation points
w: the response of the original system at
c_full: the Lagrangian coefficients (obtained with the full n-D Loewner matrix null-space computation)
c_rec: the Lagrangian coefficients (obtained with the recursive 1-D Loewner matrix null-space computation)
flop_full: FLOPs (obtained with the full n-D Loewner matrix null-space computation)
flop_rec: FLOPs (obtained with the recursive 1-D Loewner matrix null-space computation)
time_full: CPU time (obtained with the full n-D Loewner matrix null-space computation)
time_rec: CPU time (obtained with the recursive 1-D Loewner matrix null-space computation)
Here we consider first a collection of functions H(s1,s2,...,sn), that are rational / polynomial and accessible. We aim at demonstrating the scalability with respect to the number of variables "n".
>> Set #1 Data (Matlab)
>> Our score sheet (PDF)
Additionally, we consider possible an other function set (possibly non-rational), obtained from literature
>> Set #2 Data (Matlab)
>> Our score sheet (PDF)
Here we consider examples where the tensor data is has been computed and we want to construct a rational parametric model. Most of the examples come from the MOR Wiki page. In most of these cases, the parametric realization is known but its parametric orders dependency is unknown.
>> Data (Matlab)
>> Our score sheet (PDF)
Here we consider a collection of data related to the non-linear eigenvalue problem.
>> Data (Matlab)
>> Our score sheet (PDF)