Estimation of sparse polynomial approximation error to continuous function

Abstract

The sparse polynomial approximation of continuous functions has emerged as a prominent area of interest in function approximation theory in recent years. A key challenge within this domain is the accurate estimation of approximation errors. This paper focuses on continuous functions, characterizing their sampled values as a combination of the values of their best approximation polynomials within a finite-dimensional polynomial space and the associated remainder terms. Consequently, the sampled values of a function can be interpreted as noisy samples of the values of its best approximation polynomial, with the noise equivalent to the remainder term's values at those points. By selecting a uniformly bounded orthonormal polynomial system as the basis for this finite-dimensional space, it becomes feasible to formulate noise constraint inequalities and l1-minimization problems or their weighted l1-minimization variants. This paper provides estimations for the approximation error of the sparse polynomial derived from the l1-minimization method, characterizing the error in terms of the quasi-norm of the sampled function or its best uniform approximation polynomial, the sparsity, and the best approximation error. The analysis reveals that if the sampled function is a sparse polynomial from a finite-dimensional space, it can be reconstructed exactly. Moreover, it is observed that the smoother the sampled function, the fewer degrees of the sparse polynomial are required to attain a given approximation accuracy. The paper also extends this analysis to estimate the L2-norm approximation error for the sparse polynomial obtained via the weighted l1-minimization method, noting that in this context, the orthonormal polynomial system does not need to be uniformly bounded for the conclusions to hold.