Hacker News Reader

Comment by srean | Hacker News Reader

Care to elaborate on the martingale part, very curious.

Martingales with barrier reflections are a fantastic way to model all kinds of parameters that are randomly distributed over time, but stay roughly the same value in the short haul. They're also very well studied, so you can make strong guarantees about their expected values and distributions, unlike naive random walks that stay within a given range.

We published a paper about such an application to modelling sensor error here: https://doi.org/10.1177/1932296817711297

The martingale comes up under the "simulator" section, starting in the paragraph "In our previous simulation study".

Thanks a bunch, looking forward to reading the paper. I am especially intrigued by the claim one can give stronger guarantees than say a simple random walk.

By "naive" random walk, I'm referring to how one would _implement_ a random walk that's restricted to a given range. The most straightforward way to do it is to "clip" any values that would exit the given range by simply coercing them to the bound, eg:

if next_value > upper_bound: next_value = upper_bound

if next_value < lower_bound: next_value = lower_bound

This works fine, but serves to concentrate probability mass near the boundaries, so it's no longer uniformly distributed. By reflecting across boundaries rather than coercing, you're effectively flattening that concentration. If you coerce, you also reduce the expected value, which may or may not be a desirable property.

Reply on news.ycombinator.com