Thu, 19 Jul 2007
More about fixed points and attractors
I picked a few example functions, some of which worked and some of which didn't.
One glaring omission from the article was that I forgot to mention the so-called "Babylonian method" for calculating square roots. The Babylonian method for calculating √n is simply to iterate the function x → ½(x + n/x). (This is a special case of the Newton-Raphson method for finding the zeroes of a function. In this case the function whose zeroes are being found is is x → x2 - n.) The Babylonian method converges quickly for almost all initial values of x. As I was writing the article, at 3 AM, I had the nagging feeling that I was leaving out an important example function, and then later on realized what it was. Oops.
But there's a happy outcome, which is that the Babylonian method points the way to a nice general extension of this general technique. Suppose you've found a function f that has your target value, say √2, as a fixed point, but you find that iterating f doesn't work for some reason. For example, one of the functions I considered in the article was x → 2/x. No matter what initial value you start with (other than √2 and -√2) iterating the function gets you nowhere; the values just hop back and forth between x and 2/x forever.
But as I said in the original article, functions that have √2 as a fixed point are easy to find. Suppose we have such a function, f, which is badly-behaved because the fixed point repels, or because of the hopping-back-and-forth problem. Then we can perturb the function by trying instead x → ½(x + f(x)), which has the same fixed points, but which might be better-behaved. (More generally, x → (ax + bf(x)) / (a + b) has the same fixed points as f for any nonzero a and b, but in this article we'll leave a = b = 1.) Applying this transformation to the function x → 2/x gives us the Babylonian method.
I tried applying this transform to the other example I used in the original article, which was x → x2 + x - 2. This has √2 as a fixed point, but the √2 is a repelling fixed point. √2 ± ε → √2 ± (1 + 2√2)ε, so the error gets bigger instead of smaller. I hoped that perturbing this function might improve its behavior, and at first it seemed that it didn't. The transformed version is x → ½(x + x2 + x - 2) = x2/2 + x - 1. That comes to pretty much the same thing. It takes √2 ± ε → √2 + (1 + √2)ε, which has the same problem. So that didn't work; oh well.
But actually things had improved a bit. The original function also has -√2 as a fixed point, and again it's one that repels from both sides, because -√2 ± ε → -√2 ± (1 - 2√2)ε, and |1 - 2√2| > 1. But the transformed function, unlike the original, has -√2 as an attractor, since it takes -√2 ± ε → -√2 ± (1 - √2)ε and |1 - √2| < 1.
So the perturbed function works for calculating √2, in a slightly backwards way; you pick a value close to -√2 and iterate the function, and the iterated values get increasingly close to -√2. Or you can get rid of the minus signs entirely by transforming the function again, and considering -f(-x) instead of f(x). This turns x2/2 + x - 1 into -x2/2 + x + 1. The fixed points change places, so now √2 is the attractor, and -√2 is the repeller, since √2 ± ε → √2 ± (1 - √2)ε. Starting with x = 1, we get:
I had meant to write about Möbius transformations, but that will have to wait until next week, I think.
[ Addendum 20201018: Another followup article, but I never did get around to discussing the Möbius transformations. ]