The New York Times carried an obituary for Dr Arianna Wright Rosenbluth. She died in December from Covid-19 and the United States’s mass-murderous handling of Covid-19. And she’s a person I am not sure I knew anything substantial about. I had known her name, but not anything more. This is a chance to correct that a bit.
Rosenbluth was a PhD in physics (and an Olympics-qualified fencer). Her postdoctoral work was with the Atomic Energy Commission, bringing her to a position at Los Alamos National Laboratory in the early 1950s. And a moment in computer science that touches very many people’s work, my own included. This is in what we call Metropolis-Hastings Markov Chain Monte Carlo.
Monte Carlo methods are numerical techniques that rely on randomness. The name references the casinos. Markov Chain refers to techniques that create a sequence of things. Each thing exists in some set of possibilities. If we’re talking about Markov Chain Monte Carlo this is usually an enormous set of possibilities, too many to deal with by hand, except for little tutorial problems. The trick is that what the next item in the sequence is depends on what the current item is, and nothing more. This may sound implausible — when does anything in the real world not depend on its history? — but the technique works well regardless. Metropolis-Hastings is a way of finding states that meet some condition well. Usually this is a maximum, or minimum, of some interesting property. The Metropolis-Hastings rule has the chance of going to an improved state, one with more of whatever the property we like, be 1, a certainty. The chance of going to a worsened state, with less of the property, be not zero. The worse the new state is, the less likely it is, but it’s never zero. The result is a sequence of states which, most of the time, improve whatever it is you’re looking for. It sometimes tries out some worse fits, in the hopes that this leads us to a better fit, for the same reason sometimes you have to go downhill to reach a larger hill. The technique works quite well at finding approximately-optimum states when it’s hard to find the best state, but it’s easy to judge which of two states is better. Also when you can have a computer do a lot of calculations, because it needs a lot of calculations.
So here we come to Rosenbluth. She and her then-husband, according to an interview he gave in 2003, were the primary workers behind the 1953 paper that set out the technique. And, particularly, she wrote the MANIAC computer program which ran the algorithm. It’s important work and an uncounted number of mathematicians, physicists, chemists, biologists, economists, and other planners have followed. She would go on to study statistical mechanics problems, in particular simulations of molecules. It’s still a rich field of study.