j0 (and y0) in the range 2 <= x < (p/2)*log(2)

[Steve Kargl]

Anyone know where the approximations for j0 (and y0) come from?

msun/src/e_j0.c states

 *         for x in (2,inf)
 *              j0(x) = sqrt(2/(pi*x))*(p0(x)*cos(x0)-q0(x)*sin(x0))
 *         where x0 = x-pi/4. It is better to compute sin(x0),cos(x0)
 *         as follow:
 *              cos(x0) = cos(x)cos(pi/4)+sin(x)sin(pi/4)
 *                      = 1/sqrt(2) * (cos(x) + sin(x))
 *              sin(x0) = sin(x)cos(pi/4)-cos(x)sin(pi/4)
 *                      = 1/sqrt(2) * (sin(x) - cos(x))
 *         (To avoid cancellation, use
 *              sin(x) +- cos(x) = -cos(2x)/(sin(x) -+ cos(x))
 *          to compute the worse one.)

p0(x) and q0(x) are divergent asymptotic series.  If I extract
pzero() and qzero() from e_j0.c and compare the results against
summing truncated versions of p0(x) and q0(x), there are no obvious
connections.

Reading the documentation for the algorithms used in MPFR suggests
that x >= p/2*log(2), where p is precision of x, is required for use
of the large argument approximation for j0(x).  In double precision,
p = 53, so we have x >= 18.368...

Consider x=18.4 and sum up to N = 31 in the asymptotic series:

% ./pq 30 18.4
 p =  9.997932830701132e-01,  q = -6.781826311540553e-03  <-- series
pp =  9.997932830701132e-01, qq = -6.781826311540509e-03  <-- pzero,qzero
ulp(p, pp) = 0.000000e+00
ulp(q, qq) = 2.550000e+01

This is almost reasonable if 25.5 ULP is acceptable in q0(x).  Note
the series are computed in long double with 64 bits of precision.

Now, comparing x = 2 and summing N = 4 (best results).

% ./pq 4 2
 p =  9.894313812255859e-01,  q = -5.334472656250000e-02
pp =  9.862158212188928e-01, qq = -5.647769967932505e-02
ulp(p, pp) = 1.448159e+13
ulp(q, qq) = 2.257545e+14

For values of N > 4, the series start to diverge!  So, how does msun
use the large argument approximation for j0(x)?

-- 
Steve