This week’s topic is one of several suggested again by Mr Wu, blogger and Singaporean mathematics tutor. He’d suggested several topics, overlapping in their subject matter, and I was challenged to pick one.
The reputation of mathematics has two aspects: difficulty and truth. Put “difficulty” to the side. “Truth” seems inarguable. We expect mathematics to produce sound, deductive arguments for everything. And that is an ideal. But we often want to know things we can’t do, or can’t do exactly. We can handle that often. If we can show that a number we want must be within some error range of a number we can calculate, we have a “numerical solution”. If we can show that a number we want must be within every error range of a number we can calculate, we have an “analytic solution”.
There are many things we’d like to calculate and can’t exactly. Many of them are integrals, which seem like they should be easy. We can represent any integral as finding the area, or volume, of a shape. The trick is that there’s only a few shapes with volumes we can find exact formulas for. You may remember the area of a triangle or a parallelogram. You have no idea what the area of a regular nonagon is. The trick we rely on is to approximate the shape we want with shapes we know formulas for. This usually gives us a numerical solution.
If you’re any bit devious you’ve had the impulse to think of a shape that can’t be broken up like that. There are such things, and a good swath of mathematics in the late 19th and early 20th centuries was arguments about how to handle them. I don’t mean to discuss them here. I’m more interested in the practical problems of breaking complicated shapes up into simpler ones and adding them all together.
One catch, an obvious one, is that if the shape is complicated you need a lot of simpler shapes added together to get a decent approximation. Less obvious is that you need way more shapes to do a three-dimensional volume well than you need for a two-dimensional area. That’s important because you need even way-er more to do a four-dimensional hypervolume. And more and more and more for a five-dimensional hypervolume. And so on.
That matters because many of the integrals we’d like to work out represent things like the energy of a large number of gas particles. Each of those particles carries six dimensions with it. Three dimensions describe its position and three dimensions describe its momentum. Worse, each particle has its own set of six dimensions. The position of particle 1 tells you nothing about the position of particle 2. So you end up needing ridiculously, impossibly many shapes to get even a rough approximation.
With no alternative, then, we try wisdom instead. We train ourselves to think of deductive reasoning as the only path to certainty. By the rules of deductive logic it is. But there are other unshakeable truths. One of them is randomness.
We can show — by deductive logic, so we trust the conclusion — that the purely random is predictable. Not in the way that lets us say how a ball will bounce off the floor. In the way that we can describe the shape of a great number of grains of sand dropped slowly on the floor.
The trick is one we might get if we were bad at darts. If we toss darts at a dartboard, badly, some will land on the board and some on the wall behind. How many hit the dartboard, compared to the total number we throw? If we’re as likely to hit every spot of the wall, then the fraction that hit the dartboard, times the area of the wall, should be about the area of the dartboard.
So we can do something equivalent to this dart-throwing to find the volumes of these complicated, hyper-dimensional shapes. It’s a kind of numerical integration. It isn’t particularly sensitive to how complicated the shape is, though. It takes more work to find the volume of a shape with more dimensions, yes. But it takes less more-work than the breaking-up-into-known-shapes method does. There are wide swaths of mathematics and mathematical physics where this is the best way to calculate the integral.
This bit that I’ve described is called “Monte Carlo integration”. The “integration” part of the name because that’s what we started out doing. To call it “Monte Carlo” implies either the method was first developed there or the person naming it was thinking of the famous casinos. The case is the latter. Monte Carlo methods as we know them come from Stanislaw Ulam, mathematical physicist working on atomic weapon design. While ill, he got to playing the game of Canfield solitaire, about which I know nothing except that Stanislaw Ulam was playing it in 1946 while ill. He wondered what the chance was that a given game was winnable. The most practical approach was sampling: set a computer to play a great many games and see what fractions of them were won. (The method comes from Ulam and John von Neumann. The name itself comes from their colleague Nicholas Metropolis.)
There are many Monte Carlo methods, with integration being only one very useful one. They hold in common that they’re build on randomness. We try calculations — often simple ones — many times over with many different possible values. And the regularity, the predictability, of randomness serves us. The results come together to an average that is close to the thing we do want to know.
I hope to return in a week with a fresh A-to-Z essay. This week’s essay, and all the essays for the Little Mathematics A-to-Z, should be at this link. And all of this year’s essays, and all A-to-Z essays from past years, should be at this link. And if you’d like to shape the next several essays, please let me know of some topics worth writing about! Thank you for reading.