Eqution P'(x)=0 is  also bad because P'(x) may has multiple roots. It should start from d(N-1)P/dx=0 which is linear and go up applying binary search on each stage. If root multiple then on previous stage it was simple and is founded. As result all candidates for roots will be found with hight precision. The problem is to make choice which of them are roots of  P(x)=0. Here we must use existence theoren of Sturm . Simplest variant is P(a)*P(b)<0=> x in [a,b]. I used with success long aritmetic and now on WA16.

AC finally.
In Java only using my BigFraction baseg on BigInt
I implemented famous Stourm's theorem which garantees
absolutely correct considering multiple roots
without any eps and so on.
But in question what degree of found multiple root
I needed in unknown constant h which garantes that if
d/dtP(t)!=0 and P(t)==0 than in (t-h,t+h) d/dtP(t)!=0.
I taken h=1e-12. In other words I think that if
P(t)==0 and in (t-h,t+h) exists t1 such that d/dtP(t1)==0
then t==t1.

h=1e-6 works also.

Edited by author 08.12.2007 19:43