Custom Query (1030 matches)

The sqrt, log and atan functions are failing over significant portions of their domain (particularly for rational or complex rational arguments), even when it is entirely reasonable that a good result should be returned. Some typical examples are:

(sqrt (expt 10 47)) => 3.1622778E+23 ;CCL errors
(sqrt (/ (expt 10 47) 3)) => 1.8257418E+23 ;CCL errors
(sqrt (complex (expt 10 46) (expt 10 47))) => #C(2.3505187E+23 2.12719E+23) ;CCL errors
(sqrt (complex most-positive-short-float most-positive-short-float)) => #C(2.0267142E+19 8.394926E+18) ;CCL errors

(log (expt 10 -66)) => -151.97063 ;CCL errors
(log (- (expt 10 66))) => #C(151.97063 3.1415927) ;CCL returns wrong type (double, not single)
(log (complex (expt 10 65) (expt 10 66))) => #C(151.9756 1.4711276) ;CCL errors
(log (complex (expt 10 -65) (expt 10 -66))) => #C(-149.66307 0.09966865) ;CCL errors
(log 8.0d0 2) => 3.0D0 ;CCL returns wildly inaccurate result
(log #C(0.0 1.0) 0) => #C(0.0 0.0) ;CCL returns wrong type (integer)

(atan (expt 10 46) (expt 10 47)) => 0.09966865 ;CCL errors
(atan (expt 10 -46) (expt 10 -47)) => 1.4711276 ;CCL returns zero

I'm attaching a file that demonstrates how these could all be fixed by making changes to CCL's sqrt, log, log-e and atan functions (but the changes may not be implemented in the most efficient way). Making these fixes would also have a knock-on effect and correct dependent infelicities in both the arg and phase functions (which aren't noted above).

In ccl 1.6 (32-bit) we have

1 > (atanh #c(-2d0 -1d-20))
#C(-0.5493061443340549D0 -1.5707963267948966D0)

but

1 > (atanh #c(-2d0 -0d0))
#C(-0.5493061443340549D0 1.5707963267948966D0)

The spec says that for on the negative real axis less than -1, atanh is continuous with Quadrant III. Since (float-sign -0d0) -> -1d0, both answers should have the same signs and the imaginary part should be negative.

testing attachments