The importance of nonlinearities in the governing equations for the dynamical behavior of the atmosphere on long timescales is a repeatedly debated issue in the atmospheric science community. We present evidence for nonlinear signatures and regime-like behavior in the geopotential height field of a very long general circulation model (GCM) integration. When projected onto the phase-space spanned by the leading empirical orthogonal functions (EOF) the probability density function (PDF) exhibits regions of enhanced probability, indicating the existence of preferred atmospheric circulation patterns. More strikingly, in some subspaces the projected mean phase-space tendencies form a distinct double-swirl pattern, pointing directly to nonlinear dynamical behavior.

By fitting a nonlinear stochastic model with multiplicative noise in the highly truncated phase space, we find that the nonlinear signatures in the tendencies are sufficient to produce the observed regions of enhanced probability in the PDF. We address the fundamental question, for which timescales is a stochastic approach valid. On the timescales for which the stochastic model captures the temporal characteristics of the GCM, the associated Fokker-Planck Equation can determine the mean and spread of ensemble predictions, depending on the initial-state distribution.