It is important to choose the ``on-line" averaging time 
, i.e.,
the averaging applied to the data before they are written to an initial
storage medium, carefully. The data will be averaged relative to a
particular phase tracking center and this on-line averaging cannot be
undone. Any excess averaging done off-line, after the data have been
archived, can be undone later, though it is obviously inefficient to do
this.
The effects of finite averaging time were outlined in
Lectures 3 and 10, and are also discussed in detail by Bridle
& Schwab (1989), whose results I simply quote here.
As is increased, phase winding of a point source at radius
from the phase center both smears
and attenuates the synthesized response to such a
source. The effect is worst on a given baseline when Earth rotation
moves the source perpendicular to the fringes associated with that
baseline, but is zero when the feature moves parallel to the fringes.
The magnitude of the effect therefore depends on baseline orientation,
hour-angle and declination. For an array observing a point source
near the north celestial pole, the average reduction in
amplitude 
can be estimated in terms
of the angular distance from the phase tracking center as:
where
for uniform square u-v
coverage,
for uniform circular u-v coverage and
for
circular, Gaussian tapered, u-v coverage, and
is in seconds.
The choice of balances several issues. First, you
must decide what amplitude reduction 
due to time-average
smearing will be acceptable at 
. Then estimate the
corresponding from Equation 2 using
appropriate values of 
and 
. If sensitivity or hardware
considerations force you to choose an IF bandwidth that is wider than
the 
given by Equation 1, estimate
so that the effects of time-average smearing at 
are
slightly less than those of chromatic aberration. (See
VLA Averaging Time
for a VLA-based example.)
Ideally, you would observe with set to the value
estimated from Equation 2 unless this exceeds
the expected coherence time 
for the atmospheric phase
fluctuations on the longest baselines, or unless it is too long to let
you edit the data satisfactorily. In either of those cases, you would
observe with a shorter . You may later average the edited,
calibrated data to the value estimated from
Equation 2, to minimize the data volume and
the image processing time.
