Continuous approximation methods for the regularization and smoothing of integral transforms (Q908677): Difference between revisions

Consider \(\int^{b}_{a}k(s,t)f(t)dt=g(s)\) where g(s) is not known explicitly but where n measured observations of g are given with a possible ``white-noise'' contamination. That is, the data is \((2)\quad g(s_ i)=\int^{b}_{a}k(s_ i,t)f(t)dt+\epsilon (s_ i),\) \(i=1,2,...,n\), where \(\epsilon (s_ i)\sim N(0,\sigma)\), \(i=1,2,...,n\), (i.e. \(\epsilon (s_ i)\) are independent errors with a normal distribution of mean zero and common variance \(\sigma^ 2\) (unknown). This paper deals with continuous approximation methods for obtaining a numerical solution f(t) in (1). The approximate solution is expressed in the linear form \(f^*=\sum a_ j\phi_ j\), where \(a_ j\) are parameters and \(\phi_ j\) are certain basis functions. The values of \(a_ j\) are determined by the minimization of a regularizing measure, which takes account of both the discrete \(I_ 2\) error in the integral transform and the continuous \(L_ 2\) norm of \(f^*\) or one of its derivatives. A generalized cross-validation technique, based on the work of \textit{G. Wahba} [SIAM J. Numer. Anal. 14, 651-667 (1977; Zbl 0402.65032)], is used for determining the smoothing parameter, and efficient algorithms are developed for three specific sets of basis functions \(\{\phi_ j\}\), including a novel algorithm when \(\{\phi_ j\}\) are chosen to be a set of eigenfunctions. Numerical examples are also given to compare the merits of the various algorithms. In the case where the function g(s) is not affected by noise, the established ``method of truncated solutions'' is adopted and an improved version of this method, based on B-splines, is described and then tested on numerical examples.