This is quick list of some possible fixes for the next revision of quantreg
Version:  June 11 2004

1.  refine the choice of eps parameter in rq.fit.fn  (this has been a
perennial problem.  As several examples show, it is sometimes possible
to get into trouble with the current 1e-6 tolerance with rq.fit.fn
returning NaN's while a  more relaxed tolerance say 1e-4 produces a
good answer, verifiable from "br".

2.  Introduce percentile methods for the bootstrap methods.

3.  Make the anova.rq function somewhat more flexible, allowing e.g.
to test equality of one slope coef across several quantiles.
