I had a look at Knuth's The Art of Computer Programming, book 1. In chapter 1, section 1.2.2, exercise 25, he presents the following algorithm for calculating logarithm:

given $x\in[1,2)$, do the following:

L1: [Initialize] Set $y\leftarrow0$, $z\leftarrow x/2$, $k\leftarrow1$.

L2: [Test for end] If $x=1$, stop.

L3: [Compare] If $x-z<1$, go to L5.

L4: [Reduce values] Set $x\leftarrow x-z$, $z\leftarrow x/2^k$, $y\leftarrow y+\log_b(2^k/(2^k-1))$, and go to L2.

L5: [Shift] Set $z\leftarrow z/2$, $k\leftarrow k+1$, and go to L2.

(The algorithm needs an auxiliary table which stores $\log_b2$, $\log_b(4/3)$, and so on, to as many values as the precision of the computer.

Then Knuth concludes: this exercise is to explain why the above algorithm will terminate and why it computes an approximation of $y=\log_b x$.

Well, I see that it works, but I am not able to explain why...

  • $\begingroup$ Just to be clear, the result that you return is supposed to be the final value of $y$? $\endgroup$ – Aaron Sep 3 '11 at 17:39
  • $\begingroup$ Yes, $y$ is the result. I'll make it clear. $\endgroup$ – zar Sep 3 '11 at 19:06

The idea is that if we can express $x$ as a product $x=\prod a_i$, where the $1<a_{i+1}\leq a_{i}$, then $\log x = \sum \log a_i$, and if the $a_i$ decrease fast enough, then we can get a good approximation by taking partial sums, assuming that we have precomputed $\log a_i$.

Obviously, with such an algorithm, we very often will NOT terminate, and so we will have to stop when we are good enough. This is because if the $a_i$ are chosen from some countable set $B=\{b_i\}$, then only countably many numbers will have expressions as products of only finitely many $b_i$ (using repetitions). So step $L2$ needs to be modified to a condition like$|x-1|<\epsilon$, for some $\epsilon$ determining the accuracy.

So what does the algorithm do? At each stage, it starts with $x$, and finds the smallest $k$ such that $x-z=x(1-2^{-k})>1$, which is equivalent to $x>(1-2^{-k})^{-1}=\frac{2^k}{2^k-1}=a_i$. We then replace $x$ with $x/a_i$ and continue recursively, knowing that $\log x = \log a_i + \log x/a_i$.

In the end, we know that we will be off my a (multiplicative) factor of no more than $\log(1+\epsilon)\approx \epsilon$ (if $\epsilon$ is small, and we can use Taylor series to put explicit bounds on $\epsilon$ to make the error in this approximation as small as desired), and because of our assumption that $1\leq x <2$, this translates to an absolute error of at most $\log 2 \log (1+\epsilon) < 2\epsilon$.

  • $\begingroup$ There is a note, in Knuth's explanation, that says: "The algorithm involves an intentional computational error in step L5, as $z$ is shifted to the right one place, so that eventually $x$ will be reduced to 1 and the algorithm will terminate". So it seems that the algorithm is correct, and it will terminate without corrections. Am I missing something? $\endgroup$ – zar Sep 3 '11 at 22:23
  • 1
    $\begingroup$ @zar As far as I can see, the only way the algorithm will terminate as stated is if you assume that you are working with fixed precision. In this case, the difference between $x$ and $1$ will eventually be less than the level of precision. Even if you assume floating (where you only keep a certain number of significant digits), then because $x\approx 1$, you will still terminate (modulo problems with equality of floating numbers). However, if you assume arbitrary precision, the algorithm will usually not terminate as written. What assumptions does Knuth make about the mode of computation? $\endgroup$ – Aaron Sep 3 '11 at 22:34
  • $\begingroup$ Knuth just says "suppose we have a binary computer". But he notes that $z$ is shifted to the right one place, thene $x$ will eventually reduced to 1. So I think he uses fixed precision (working at low level, with bit shifting). $\endgroup$ – zar Sep 4 '11 at 8:49
  • $\begingroup$ @Aaron Sorry, Aaron, but I don't get the absolute value in $|x-1| < \epsilon$. Notice, in fact, that since $1 \leq x < 2$, $x-1$ will always be > 0. Moreover, where did you derive from that the multipicative factor left off in the computation is $\log(1+\epsilon)$, when you only know that $\log|x-1| < \log \epsilon$? $\endgroup$ – Riccardo Del Monte May 31 '14 at 17:20
  • $\begingroup$ @Aaron Also, I think you should review your upper bound for the error made by the algorithm: you say that it is $\log2\log(1+\epsilon)$, but first I don't see how you came up with this simply starting from the fact that $1 \leq x < 2$, and, most importantly, I don't think I would disagree with my computer which tells me, with tons of tests, that the relation $\log x-\log'x < \log(1+\epsilon)$ is valid, while yours is not. By the way, keep in mind that I'm not still completely sure that $\log(1+\epsilon)$ is the actual upper bound of the error; I only see that the error is less than such value $\endgroup$ – Riccardo Del Monte Jun 14 '14 at 19:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.