For real x, the magnitude of x must not be so large that most of the computer word contains the integer part of x. Likewise, x must not be too near an integer multiple of π, although x close to the origin causes no accuracy loss. Finally, x must not be so close to the origin that COT(X) ≈ 1/x overflows.
For complex arguments, let z = x + iy. If |sin z|2 is very small, that is, if x is very close to a multiple of π and if |y| is small, then cot z is nearly singular and a fatal error condition is reported. If |sin z|2 is somewhat larger but still small, then the result will be less accurate than half precision. When |2x| is so large that sin 2x cannot be evaluated accurately to even zero precision, the following situation results. If |y| < 3/2, then CCOT cannot be evaluated accurately to be better than one significant figure. If 3/2 ≤|y| < -1/2 ln ε/2, where ε = AMACH(4) is the machine precision, then CCOT can be evaluated by ignoring the real part of the argument; however, the answer will be less accurate than half precision. Finally, |z| must not be so small that cot z ≈ 1/z overflows.
3 2 Result of CCOT(Z) is accurate to less than one-half precision because the real part of Z is too near a multiple of π when the imaginary part of Z is zero, or because the absolute value of the real part is very large and the absolute value of the imaginary part is small.
1. Referencing COT(X) is NOT the same as computing 1.0/TAN(X) because the error conditions are quite different. For example, when X is near π /2, TAN(X) cannot be evaluated accurately and an error message must be issued. However, COT(X) can be evaluated accurately in the sense of absolute error.
Visual Numerics, Inc.