Three-dimensional covariance matrices are computed at each 5 latitude by 10 longitude location flora approximately 30 degrees S-60 degrees N at depths of 0m, 200m, 400m from temperature anomalies about the mean annual cycle for the 13 yeats from 1979-1991. Lags are chosen to focus on seasonal-to-intgrannual time scales resolved on gyre-to-basin space scales. These sample covariance maU'ices are fitted with a second-order auto-regressive (AR) model, allowing for the detection of horizontal wave propagation in the presence of dissipation. Over most of the ocean and at all depths, ENSO signals (i.e. with zero-crossing scales of 9-12 months) dominate the interannual variability, with eastward propagation found at the sea surface over most of the Pacific and Indian Oceans, and westward propagation found over most of the Pacific Ocean at 20(0 and 400m, the latter qualitatively consistent with Rossby wave propagation. The noise variance is approximately equal to the signal variance at all depths, while the ratio of zero-crossing scale-to-decay scale (i.e. 1.5-2.5) finds ENSO wave propagation critically dissipated almost everywhere. This effectively reduces the second-order AR model to a fast-order AR model, with decay scales and not zero-crossing scales defining decorrelation scales. Analyzing these covariance statistics within the framework of optimum interpolation allows sampling rates to be estimated for detecting sensonal-to-interannual variability to within prescribed error limits. Utilizing the fast-order AR covariance model, 1-2 observations per decorrelation scale in each dimension allows the interpolation error to be 0.8-0.6 of the signal standard deviation. A uniform set of decorrelation scales (e. g. 2.5 degrees latitude, 5 degrees longitude, 3 months) and 1.0 for the noise-to-signal ratio are chosen for detecting minimum gyre-scale biennial variability at all depths uniformly over the global domain. Therefore, sampling the global ocean at a rate of 2500-3000 observations per month yields maximum interpolation errors for biennial signals (i.e. : +/- 0.4 degrees C at the sea surface, +/- 0.5 degrees C at 200m, and +/- 0.2 degrees C at 400m) that are similar to those for the mean annual cycle. This sampling rate is adequate for detecting the space-time evolution of biennial signals (and, hence, ENSO signals) over the interior global ocean. But, it is inadequate for detecting year-to-year changes in upper ocean heat storage averaged over large portions of the global ocean as required by the WOCE Heat Budget Study.
