Ln(x) it X MP b IX 1 b j 1 IX ( j 1) j 1 Error The error in xMP is b xMP b X 11b t 2 2 Expand log x log xMP 1 xMP / xMP log xMP xMP / xMP In the case that x is close to 1, this is the dominant error. Write log xMP as t 1 log xMP b LogX 11 b i LogX i i2 The ratio of the first term left our of the series to the first is Rat b it 2 b / 2 In the case that x is not close to 1, this is the dominant error, for x very close to 1 the error in (1-x) dominates. The calculation speeds up as (1-x) 0, so it costs little to calculate to the relative accuracy, but users need to bear in mind that if (1-x) < b--it-1, that there is no accuracy in the stated log. The lack of accuracy is also true for double precision logarithms. The problem can be avoided by always passing 1-x in the argument. mpkroot2.for mplog2.for. 2k’th root of (x-1)+1 This starts by finding 1 c x 1 1 (1) Note that for k=0, c = x-1 and the expansion in (4) gives the answer. log x 2k log 1 c (2) 2 k The left side of 1 is (1+c)2 1 + 2c (1+2c)21+22c … (1+2k-1c) 1+2kc So that c (x-1)/2k (3) This is found much more accurately in KthRoot2.doc - mpln.for. c i 1 i log 1 c i (4) Estimation of M and k If the sum in (4) is truncated at i=M, the error is Series c i M 1 i i c M 1 M 1 c i c M 1 1 c M 1 i 0 (5) For c << 1, the c in the denominator can be dropped so that, the value of log(x) is Ln x 2 M k c i 1 i c M 1 M 1 i 2 k (6) For small c, the natural log is approximately 2kc with relative truncation error c M 1 M rel ,T (7) The absolute error in the multi precision value of the natural log is - MPDetail.doc 1 b IC 1it (8) 2 IC(1)=log(c)/14-1 The first term in this expansion is approximately c so the relative error is ||/c, so that equating this to the truncation error gives M 1 c 1 ln( c ) / 14*ln(2) it b 2 M 1 (9) Or log 2 log(c) / 14*log(2) it log b M 1 log c log M 1 Note that for our b = 16384 = 214, this becomes log 2 log(c) / 14*log(2) it log b M 1 log c log M 1 log c(1 M 1) it log b log 2 / M 1 (10) (11) M it log b log 2 log( M 1) / log c To a reasonable approximation, the last two terms on the last line of (11) can be dropped and 2k c = x, so that M it log b / k log 2 log x For x such that k log2 >> log x, this becomes M 14it / k The number of multiplications required to find log(x) are approximately1 7 k + M or ops M 98it / M This is a minimum with M 10 it (12) k 1.4 it The total number of multilications is 20 (it)1/2 The 7 is approximately the number of iterations required to solve (1+c) M = 1+x where M=2k by Newton’s method. 1 M and k are found semi-iteratively. Assume that M is given by the first line in (12). Then use the approximation in (3), c=(x-1)/2k, to find k from the last line of (11) M it log b log 2 log( M 1) / log x 1 / 2 k log x 1 / 2k it log b log 2 log( M 1) / M k log 2 log( x 1) it log b log 2 log( M 1) / M k it log b / log 2 1 log( M 1) / log 2 / M log x 1 / log 2 k 14it / M 1/ log 2 log( M 1) / log 2 log( x 1) / log 2 Ignoring the value of x, it is shown below (12) that the log series should contain M = 10 it terms. This implies that c M Mb it 1 x 1 2 k M 1 Mb M 1 log( x 1) k log 2 log M log b Then