Abstract
It is determined how long a time series must be to estimate covariances and moments up to fourth order with a specified statistical significance. For a given averaging time T there is a systematic difference between the true flux or moment and the ensemble average of the time means of the same quantities. This difference, referred to here as the systematic error, is a decreasing function of T tending to zero for T→∞. The variance of the time mean of the flux or moment, the so-called error variance, represents the random scatter of individual realizations, which, when T is much larger than the integral time scale T of the time series, is also a decreasing function of T. This makes it possible to assess the minimum value of T necessary to obtain systematic and random errors smaller than specified values. Assuming that the time series are either Gaussian processes with exponential correlation functions or a skewed process derived from a Gaussian, we obtain expressions for the systematic and random errors. These expressions show that the systematic error and the error variance in the limit of large T are both inversely proportional to T, which means that the random error, that is, the square root of the error variance, will in this limit be larger than the systematic error. It is demonstrated theoretically, as well as experimentally with aircraft data from the convective boundary layer over the ocean and over land, that the assumption that the time series are Gaussian leads to underestimation of the random errors, while derived processes with a more realistic skewness and kurtosis give better estimates. For fluxes, the systematic and random errors are estimated when the time series are sampled instantaneously, but the samples separated in time by an amount Δ. It is found that the random error variance and the systematic error increase by less than 8% over continuously sampled data if Δ is no larger than the integral scale obtained from the flux time series and the cospectrum, respectively.
