ABSTRACT
A plausible approach would involve selection of a. distritmt~ion P whosc ccmditiona,l distributions are mil~inlally distant from the distrihtions giver1 by A arid B using some convenient measure of' distitnce between distributions. Crit>c:ri:t (13.3.10) ancl (13.3.11) arc nctually of this forrn. An attmctive ;~lternative is to use the Kullback-Leibler ir~format,ion (,o rneaswc distaice between clistrihtions. From this viewpoint Arnold and Gokhale (1998a) suggest choosing P to rrlinirriize the following ol~jective fimction
The distribution l'" which irrii~iruizes (13.3.13) can be ohtainecl by arl :q~plication of the 1)arrocl-1-rat cliff (1 972) iterat,ivc scaling algorithm.
