On Linear Versus Nonlinear Approximation in the Average Case Setting
We compare the average errors of n-term linear and nonlinear approximations assuming that the coefficients in an orthogonal expansion of the approximated element are scaled i.i.d. random variables. We show that generally the n-term nonlinear approximation can be even exponentially better than the n-term linear approximation. On the other hand, if the scaling parameters decay no faster than polynomially then the average errors of nonlinear approximations do not converge to zero faster than those of linear approximations, as n → +∞. The main motivation and application is the approximation of Gaussian processes. In this particular case, the nonlinear approximation is, roughly, no more than n times better than its linear counterpart.
- 10.Kuo, H.-H.: Gaussian Measures in Banach Spaces. Lecture Notes in Mathematics, vol. 463. Springer, New York (1975)Google Scholar
- 12.Rauhut, H.: Compressive sensing and structured random matrices. In: Fornasier, M. (ed.) Theoretical Foundations and Numerical Methods for Sparse Recovery, vol. 9, pp. 1–92. Walter de Gruyter, Berlin (2012)Google Scholar