[math-fun] Floating point matrix product anomalies

From: Bill Gosper <billgosper@gmail.com> Subject: [math-fun] Floating point matrix product anomalies

--Wow, that is strange. Perhaps related is this: 1. Take N consecutive rightward steps of length 1+noise (where noise is a tiny-variance mean=0 random variable). You will reach N+-sqrt(N)*NoiseAmplitude. 2. Instead take a step, double it, add noise, and keep on doing that for lg(N) doublings. You will reach N+-N*Constant*NoiseAmplitude. 3. If in (1) the noise each step is negatively correlated to the noise the preceding step, then we do even better. -- Warren D. Smith http://RangeVoting.org <-- add your endorsement (by clicking "endorse" as 1st step)