I discovered some strange behaviour of the PolyLog[] Function in Mathematica which seems to me like a bug in the function implementation.
It looks like PolyLog is breaking down for big arguments and non-integer orders
In[1]:= -PolyLog[5/2, -E^500] // N
Out[1]:= 2.691151407447062*10^8626 + 9.87331945684138*10^8623 I
which is for sure incorrect. For such big arguments the PolyLog function has nice asymptotic approximations (which are e.g. used for complete Fermi-Dirac Integrals which are related to PolyLog's) through $F_n(x)=-\mathrm{Li}_{n+1}(-e^x)$.
This can be demonstrated by using even orders for the PolyLog[] with similar big arguments
In[2]:= -PolyLog[4, -E^500] // N
Out[2]:= 2.60437*10^9
if compared to the asymptotic approximation $\frac{u^n}{\Gamma(n+1)}$ of $-\mathrm{Li}_n(-e^u)$
In[3]:= u^4/Gamma[5] /. u->500 //N
Out[3]:= 2.60417*10^9
the results look o.k.
The expected result of above given argument E^500 and order 5/2 should be
In[4]:= u^(5/2)/Gamma[5/2 + 1] /. u -> 500 // N
Out[4]:= 1.68209*10^6
I assume that the function algorithm somehow gets troubles with internal accuracy. Any ideas to "tell" Mathematica to increase the accuracy for the numerical calculation?
N[Sin[E^500]], I would be more comfortable withN[Sin[E^500], 20], after increasing$MaxExtraPrecisionto 400 or so. – Mark McClure Mar 29 '13 at 23:32