Work

Metastability, variation and learnability for subadditive and ergodic averages

Year: 2026
Type: preprint
Abstract: For integrable subadditive processes on probability spaces, we obtain explicit rates of metastability for convergence of f_n/n in L^1. For nonnegative data of mean at most B, the rate has the form h̃^(⌈8B/ε⌉)(1), where h̃ is constructed explicitly from a tail cutoff and a Hilbert-space mean ergodic bound. Signed data require uniform control of the negative tails, and finite cyclic examples show that this datum is necessary. For r ≥ 2, nonnegative processes satisfy sharp r-variation bounds under an L(log L)^(1/r) singleton moment. The rectangular estimate holds on every increasing chain for r > 2, and also for r=2 in dimension two. For signed processes, a critical negative logarithmic moment gives strong variation for r > 2 and a sharp logarithmically weighted quadratic estimate at r=2. We answer Gabor's 2019 subsequence question negatively by an increasing tempered Følner sequence. Positive results on Følner subsequences give the sharp 1/N pointwise upcrossing tail. Examples answer Neri–Powell's 2024 questions about marginal tightness and learnability. We also determine sharp learnability orders for Birkhoff, stationary Markov and ratio averages. For bounded Birkhoff inputs, pointwise and uniform counts have respective orders (B/ε)^2 log(1/λ) and (B/ε)^2/λ. On the counterexample Følner sequence both counts have the latter order after every thinning.
Language: English
Open Access status: green
OpenAlex ID: W7236556118
Cites:
Cited by:
Related to: 10