Stochastic Parameterization: The Importance of Nonlocality and Memory

Abstract

Stochastic parameterizations deployed in models of the Earth system frequently invoke locality assumptions such as Markovianity or spatial locality. This work highlights the impact of such assumptions on predictive performance. Both in terms of short-term forecasting and the representation of long-term statistics, we find locality assumptions to be detrimental in idealized experiments. We show, however, that judicious choice of Markovian parameterization can mitigate errors due to assuming Markovianity. We propose a simple modification to standard Markovian parameterizations, which yields significant improvements in predictive skill while reducing computational cost. We further note a divergence between configurations of a parameterization which perform best in short-term prediction and those which best represent time-invariant statistics.