The day in conversation for 2026-11-06

discussion

Transcript

The discussion: Ada: That O(T 1/2R) rate for the Neyman Regret seems quite robust; are we talking about that convergence holding strictly under the standard regularity conditions, or does that require some specific structural assumptions on the underlying probability space? Emmy: It's stated as minimax optimal precisely because those conditions cover the necessary setup, but I wonder if that optimality hinges more on the choice of weight function within the Augmented Inverse Probability Weighting estimators rather than just the general adaptive design framework. Kolya: From my perspective, I’m more curious about how that rate compares to what we see in simpler bandits; does this specific augmentation technique actually provide a noticeable improvement over standard FTRL when R is small? Ada: If we look back at last week’s work on sequential testing, the complexity often explodes when you move into these weighted settings, so seeing a clean power law dependency on R suggests the design manages that explosion effectively. Emmy: But linking that to the other papers today—especially the one on Eisenstein series—it feels like we jump from highly controlled estimation bounds into deep spectral analysis of zeros, which is a completely different kind of control structure. Kolya: Exactly; when we look at those odd weight Eisenstein series, determining the distribution of their zeros within fundamental domains implies a level of density that seems far removed from the finite-time regret bounds we discuss in adaptive experimental design. Ada: Perhaps the connection lies in how both fields use spectral properties

More episodes

← Home