The perturbation timescale-dependences of the CO2 flux and the random flux sampling error are evaluated from eddy-covariance tower observations in the mid-day convective boundary layer over mid-latitude conifer forests. The perturbation timescale is the timescale used in the standard Reynolds decomposition to define mean and perturbations quantities. The random error due to inadequate sampling of the turbulence is estimated using two different approaches (traditional and daily-differencing). A fixed record length of 3.6 h (dyadic timescale) is used for all results, where the record length is the timescale over which the products of perturbations are averaged (flux averaging timescale). Long multiple-hour records are required to evaluate the sampling errors.
When high temporal resolution flux estimates are of interest (e.g., sub-daily timescales), including incremental contributions to the flux from transport on timescales longer than 10 min cannot be justified based on the magnitude of the incremental increase in the random sampling error. That is, the additional flux obtained by increasing the perturbation timescale beyond 10 min is dominated by random sampling error. This result is supported by both the traditional and daily-differencing approaches. For a perturbation timescale of 30 min, the relative random error (random error divided by the flux) is 38% at the taller tower and 27% at the shorter tower, and increases with increasing perturbation timescale. The cost associated with reducing the random error by using the shorter 10 min perturbation timescale, compared to the standard practice of 30 min, is an increase in the systematic flux error from 3% to 7% (averaged over three sites). Such error, while systematic, may be small in comparison to other sources of uncertainty. The choice of the perturbation timescale, and the trade-offs between reducing systematic or random errors, may depend on the intended application of the flux data. When only longer term flux estimates (e.g., monthly or annual averages) are of interest the random sampling error tends to cancel because of the larger number of samples, and the perturbation timescale can be increased to further reduce the systematic flux error.