Longer, more informal write-ups of the research on this site: the findings that don't fit neatly into a project page.
The Bayes filter of a partially observed robot is a contraction mapping, and the mechanism is counterintuitive, so this post measures it: motion noise erases wrong priors, observations provably cannot, and the two live in different metrics.
Read postWe fitted eight models, from SIR and Lotka-Volterra to ARIMA, Neural ODEs, and PINNs, to fifty years of US presidential vote shares. The models that failed were as informative as the ones that fit, and the fitted numbers sketch an unflattering portrait of polarization.
Read postThe sum tree inside Ape-X's replay buffer, ported to C and OpenACC: why the GPU-friendly batch update does asymptotically more work on purpose, the race condition that produced silently wrong probabilities, and honest crossover measurements.
Read postWhat breaks in the ordinary Q-learning fixed point when an adversary picks half of every transition, and why the fix turns the greedy max into a small linear program over mixed strategies.
Read postTheory and measurements for tree search when the tree is infinite and the utility function is approximate: why depth rents out evaluator error at gamma^d, why the winner's curse fights back, and why A* tolerates optimism for free while noisy search cannot.
Read postA single parity theorem from 1980s origami math (Maekawa's theorem) turns out to be exactly the condition a printer toolpath needs.
Read post