| Popis: |
Many data analysis problems take the form of a deconvolution, and one of the common difficulties is that the point-spread function (p.s.f.) is not completely specified. Indeed, it is often obtained from measurements which are very similar to the test data themselves. The correct probabilistic approach is to form the full joint posterior for the p.s.f. and the object, and then to obtain posteriors for features of interest in the object as marginals over this joint posterior, i.e., to integrate over uncertainties in the p.s.f. If the joint posterior is sharply peaked for some particular p.s.f., one may approximate this integral by choosing the p.s.f. at its modal value. If, however, the posterior is not sharply peaked, then this procedure may give misleading error estimates. Examples of this are shown, drawn from the analysis of drug absorption in man. |
