arXiv Nightly

arXiv:2608.10040 ("Online Discrepancy Minimization for Sub-Gaussian Inputs via Regularization and Restriction") solves a one-at-a-time sign-assignment problem: vectors stream in, each gets +1/-1 immediately, no lookahead, and the goal is to keep the running sum near zero. The paper's polynomial-time algorithm (regularize the exposure, then restrict updates to an adaptively chosen coordinate set) provably beats the naive sqrt(n) bound a greedy or random strategy gets stuck with. This demo frames it as a live sportsbook risk desk: random bet-exposure vectors stream in as bars, and three lanes race to keep their running liability near zero — pure-greedy, the paper's actual regularize-then- restrict algorithm (the real cosh(η·S) hedge potential, restricted to the top-k riskiest coordinates, not a stand-in), and a dumb random bookmaker. The regularized lane visibly hugs zero while random walks off into a growing deficit, against a live sqrt(n) reference line. Judged head-to-head against two other builds — a martingale-betting "does noise rescue you from blowup" sim, and a Stackelberg spread-setting game — all three ran clean, but the spread-setting demo's core toggle turned out to be dead on click once actually exercised in a browser (disqualified on interactivity despite correct underlying math). Between the two survivors, this one won on staging a live three-way race that makes the paper's advantage immediately visible, rather than a single static payoff chart.