Math precision requirements of C and C++ standard

12
votes

Do the C and C++ standards require the math operations in math.h on floating points (i.e. sqrt, exp, log, sin, ...) to return numerically best solution?

For a given (exact and valid) input there can obviously in general not be an exact floating point output from these functions. But is the output required to be the representable value nearest to the mathematically exact one?

If not, are there any requirements on precision whatsoever (possibly platform-specific / in other standards ?), so that I am able to make worst-case estimates of calculation errors in my code? What are typical limits on numerical errors of modern implementations?

1
c++cfloating-pointmath.h
Basically no. IEEE-754 doesn't even specify most of these. There are some limits on the number of representable digits but there was a time when wonky proprietary floating-point implementations proliferated, and as a result the language leaves most everything about the floating-point results unspecified. That is the C spirit after all – doynax
@doynax: +1 An implementation needn't even follow IEEE 754 specification for floats/doubles. – legends2k
@doynax: What do you mean IEEE 754 does not specify most of these? IEEE 754-2008 table 9.1 recommends exp, log, sin, and more. – Eric Postpischil
@EricPostpischil: It would appear I am out-of-date, having only read IEEE 754-1985. From a cursory reading of the 2008 revision it seems that the accuracy of these functions is left unspecified aside from a few identities and the curious requirement to signal inexact results if-and-only-if they are inexact. This seems needlessly expensive to handle. Surely no-one would expect exact results from the transcendental functions anyway, with the possible exception of the exponential function? – doynax
@doynax: Section 9.2, where the table appears, says the functions should be correctly rounded. That means the error must be the minimum possible given the rounding mode and the format; in round-to-nearest mode, the nearest representable value must be returned (breaking ties with the usual rule). I agree, these results should not generally be expected except for those that have been demonstrated to be feasible (as by the CRlibm project. See my comment with MSalter’s answer; this was changed post-committee. – Eric Postpischil

1 Answers

4
votes

No, and for good reason. In general, you'd need an infinite precision (and infinite time) to determine the exact mathematical result. Now most of the times you need only a few extra iterations to determine sufficient bits for rounding, but this number of extra bits depend on the exact result (simply put: when the result is close to .5 ULP). Even determining the extra number of iterations required is highly non-trivial. As a result, requiring exact results is far, far slower than a pragmatic approach.