
High-dimensional Online Calibration: Harmonic Weights and Harmonic Analysis
Maxwell Fishelson, speaking in the Institute for Advanced Study's Computer Science/Discrete Mathematics Seminar, presents new results on online calibration of multidimensional forecasts over convex sets. He focuses on forecasting over the hypercube, where his algorithm achieves epsilon-calibration in a number of rounds polynomial in the dimension, an exponential improvement over all prior bounds. The method itself is simple to state: on each round it predicts using a harmonically weighted distribution built from harmonically smoothed past outcomes, with the analysis resting on a property of the discrete Hilbert transform. The talk walks through the construction and the proof technique behind the dimension bound, aimed at an audience already familiar with online learning and forecasting theory. Recorded at Simonyi Hall, the seminar runs long enough to develop both the algorithm and the harmonic-analysis machinery that makes the improved bound possible.