Accurate Large-sample Uncertainty Quantification using Stochastic Gradient Markov Chain Monte Carlo

In three linesNew arXiv paper proposing discrete-time approximations to SG(L)D with and without momentum, enabling accurate predictions of stationary covariance and integrated autocorrelation time. Non-asymptotic error bounds for practical tuning and uncertainty quantification, validated on misspecified models and large batch sizes.

Summary generated by Claude — human-verified