In a recent technical report [1], we have shown that our
algorithm converges in the case where the
working set size is equal to 2 and without shrinking, for any kernel
that verifies Mercer's conditions. To do so, we have used a theorem proved
by Keerthi *et al* [6]. Note however that more recently, Lin [9]
has shown the convergence of our algorithm for any value
of the working set size (but again without shrinking), under the
following hypothesis:

where

Journal of Machine Learning Research