I agree to everything Ingo states below (sorry for my mistake in explaining the number of SVs in the course of the vigorous handwaving explanation I advanced) All I meant was that there is no universal best algorithm, as Ingo also states below. Balaji -----Original Message----- From: Ingo Steinwart <[log in to unmask]> To: [log in to unmask] Date: Thu, 17 Feb 2005 14:51:31 -0700 Subject: Re: svm and curse of dimensionality Hi all, here are just a very few comments: to answers 1,2: the fraction of the training samples that are support vectors tends to TWICE the minimum achievable error rate for an optimal classifier. And this only holds if you use e.g. a Gaussian kernel and simply sum up the slacks. Otherwise you should, in general, expect even more support vectors. to answer 3.c: It is well known that there exists no best classifier, and consequently, neither SVMs nor any other method can be a best classifier. See the book of Devroye, Gyoerfi and Lugosi. ingo