Pahan Dewasurendra

Extrapolation Is an Assumption: Sharp Limits for Pass@k Forecasting

Pooled Exploration Closes the Distributed Linear-Bandit Gap

Multiscale Reward Hedging from Correct Demonstrations

The first horizon-free guarantee for continuous classes is given, to hedge in one shared vote over tolerant optimality tests at every accuracy scale, and an exact MovieLens audit runs in 1.7 CPU seconds across ten users and improves mean latent gap over both a demonstrated-rating policy and a proper online baseline.

The Quadratic Cost of Replicable Distribution Estimation

It is shown that cells with low average distortion cover constant measure, that each such cell has exponentially small measure, and that Gaussian-profile isoperimetry forces total boundary Ω( p k ) .

Self-Bounding Regret Matching+ in Potential Games and Product-Simplex Optimization

An exact one-step conservation law for RM+.

Interaction Is Unnecessary for Order-Optimal One-Bit Mean Estimation

A fully non-adaptive protocol whose query list is fixed before any bit is observed and whose sample complexity matches the adaptive one-bit minimax rate in every moment regime.

Two Dimensions Govern Agnostic Multiclass Transductive Learning

A new without-replacement multiplicative-weights lemma preserves the fast $d_{DS}/n$ term, and consequently, agnostic multiclass PAC and transductive learning obey the same two-dimension law up to logarithmic factors.

Sharp Root Anti-Concentration for Online Algorithm Configuration

Anytime Last Iterates in Fixed Dimension

Convex Networks Remain Hard to Certify: Dimension-Accuracy Barriers for Lipschitz Constants

The lifted-selector reduction has an inverse-polynomial radial gap, proved through a quantitative theorem for rational cyclic zonogons, and the results apply to Euclidean zonotope radius and positive-semidefinite binary quadratic maximization parameterized by rank.

Dirichlet Follow-the-Leader Closes the Gap in Simultaneous Multiclass U-Calibration

A single, horizon-free algorithm that satisfies the optimal regret rate for every bounded proper loss and also adapt to every smooth proper loss, covering nondifferentiable losses and changes of the active simplex face.

Optimal-Dimension U-Calibration by Bayesian Bootstrap

It is shown that the classical Bayesian bootstrap closes this gap in U-calibration, which asks one online probability fore-caster to have low regret for every bounded proper loss, including losses unknown when the forecasts are made.