### abstract

- Many major oceanographic internal wave observational programs of the last 4 decades are reanalyzed in order to characterize variability of the deep ocean internal wavefield. The observations are discussed in the context of the universal spectral model proposed by C. J. R. Garrett and W. H. Munk. The Garrett and Munk model is a good description of wintertime conditions at Site D on the continental rise north of the Gulf Stream. Elsewhere and at other times, significant deviations in terms of amplitude, separability of the 2-D vertical wavenumber-frequency spectrum, and departure from the model's functional form are reported. Specifically, the Garrett and Munk model overestimates annual average frequency domain spectral levels both at Site D and in general. The bias at Site D is associated with the Garrett and Munk model being a fit to wintertime data from Site D and the presence of an annual cycle in high-frequency energy in the western subtropical North Atlantic having a maximum in winter. The wave spectrum is generally nonseparable, with near-inertial waves typically having greater bandwidth (occupying smaller vertical scales) than continuum frequency waves. Separability is a better approximation for more energetic states, such as wintertime conditions at Site D. Subtle geographic differences from the high-frequency and high vertical wavenumber power laws of the Garrett and Munk spectrum are apparent. Such deviations tend to covary: whiter frequency spectra are partnered with redder vertical wavenumber spectra. We review a general theoretical framework of statistical radiative balance equations and interpret the observed variability in terms of the interplay between generation, propagation, and nonlinearity. First, nonlinearity is a fundamental organizing principle in this work. The observed power laws lie close to the induced diffusion stationary states of the resonant kinetic equation describing the lowest-order nonlinear transfers. Second, eddy variability and by implication wave mean interactions are also an organizing principle. Observations from regions of low eddy variability tend to be outliers in terms of their parametric spectral representation; other data tend to cluster in two regions of parameter space. More tentatively, the seasonal cycle of high-frequency energy is in phase with the near-inertial seasonal cycle in regions of significant eddy variability. In regions of low eddy variability, the seasonal cycle in high-frequency energy lags that of near-inertial energy. The induced diffusion stationary states are approximate analytic solutions to the resonant kinetic equation, and the Garrett and Munk spectrum represents one such analytic solution. We present numerical solutions of the resonant kinetic equation, however, that are inconsistent with the Garrett and Munk model representing a stationary state, either alone or in combination with other physical mechanisms. We believe this to be the case for other regional characterizations as well. We argue that nonstationarity of the numerical solutions is related to local transfers in horizontal wavenumber, whereas the analytic induced diffusion stationary states consider only nonlocal transfers in vertical wavenumber. Consequences for understanding the pathways by which energy is transferred from sources to sinks are considered. Further progress likely requires self-consistent solutions to a broadened kinetic equation.