John, Thanks, I am glad that you can reproduce the problem. Based on the debug messages I would guess that your sph_neumann wraps your cyl_neumann, this is typically how its done. I tried calling cyl_neumann with nu = 8.5 and the same value of x (ref Arfkin). double ytest = boost::math::cyl_neumann(double(n)+0.5,a*kw); This works too. I'm not sure if my type casting matches what boost does. So, I can wrap my own and they should work. Based on what I expect of the bessel function I do not believe that there should be any ill-behaved points for non-zero values of x. When I got the overflow error I initially thought it was do to a recursive assignment as it was wrapped in a template function of my own and called from within a loop (not numerical instability). But when I could reproduce the problem outside the loop and the wrapper I figured I'd email the user group. Please let me know what you find. David ________________________________ From: John Maddock <boost.regex@virgin.net> To: boost-users@lists.boost.org Sent: Friday, March 23, 2012 6:10 AM Subject: Re: [Boost-users] spherical bessel functions
It doesn't matter if I define double x = a*kw and pass this into the interface, what seems to matter is the value of the index, when n = 8 it doesn't work. It works for 9 and 10 and all integers before 8. I can replace x with a, lambda or pi and it works for n = 8, so it seems to be this combination of numbers or something I'm doing wrong w/r to type casting variables etc.
OK I can reproduce here, and it's very specific to the calculation of Y[8.5](4*Pi), if you change nu at all, or add a tiny increment to the 4PI arg then it all works OK. The problem BTW is exquisite cancellation inside the algorithm, which leads me to suspect there's something "special" about this value. Unfortunately I don't have a fix at present (or understand the root cause yet), more investigation needed.... Thanks for the report, John. _______________________________________________ Boost-users mailing list Boost-users@lists.boost.org http://lists.boost.org/mailman/listinfo.cgi/boost-users