September 1992, 26 pp.

The effects of sampling frequency on the first- and second-moment statistics of selected ECMWF model variables are investigated in a simulation of "perpetual July" with a diurnal cycle included and with surface and atmospheric fields saved at hourly intervals. The shortest characteristic time scales (as determined by the e-folding time of lagged autocorrelation functions) are those of ground heat fluxes and temperatures, precipitation and run-off, convective processes, cloud poperties, and atmospheric vertical motion, while the longest time scales are exhibited by soil temperature and moisture, surface pressure, and atmospheric specific humidity, temparature and wind. The time scales of surface heat and momentum fluxes and of convective processes are substantially shorter over land than over the oceans.

An appropriate sampling frequency for each model variable is obtained by comparing the estimates of first- and second-moment climate statistics determined at intervals ranging from 2 to 24 hours with the "best" estimates obtained from hourly sampling. Relatively accurate estimation of first- and second-moment climate statistics (10 percent errors in means, 20 percent errors in variances) can be achieved by sampling a model variable at intervals that usually are longer than the bandwidth of its time series, but that often are shorter than its characteristic time scale.

For the surface variables, sampling at intervals that are non-integral divisors of a 24-hour day yields relatively more accurate time-mean statistics because of a reduction in errors accociated with aliasing of the diurnal cycle and higher-frequency harmonics. The superior estimates of first-moment statistics are accompanied by inferior estimates of the variance of the daily means due to the presence of sytematic biases, but these probably can be avoided by defining a different measure of low-frequency variability. Estimates of the intradiurnal variance of accumulated precipitation and surface run-off also are strongly impacted by the length of the storage interval. In light of these results, some alternative strategies for storage of the EMWF model variables are recommended.(pdf file)

UCRL-MI-123395