the time window is needed to estimate the data covariance matrix, one of
the "ingredients" for calculating the LCMV beamformer spatial filter
that will be applied to your sensor space data. Generally, the estimate
of this covariance matrix is better with more data samples. Thus,
spatial filters constructed on small snippets of your data will be less
reliable than spatial filters constructed on a longer time window.
Furthermore, if you use several time windows, i.e., several filters, I
suspect that this can potentially lead to discontinuities in your source
time series (if you intend to glue the output of your beamformers together).
Usually, to construct your data covariance matrix, you would use a time
window of interest, representing the activity you are interested in.
Thank you for your answer. I am no expert into the algorithms. To compute data covariance matrix, I guess the time window is one dimension that could be used as the data samples, but trial number could also play the role in the other dimension. By select a time window, a common spatial filter would be created that representing all the data within the time window. My question is: will this kind of spatial filter make the source output temporally smoothed or "affected"? (For it collapses the data sample in the time dimension) Especially when there are more than one sources in the time window of interest.
the data covariance matrix is directly involved in the computation of
the beamformer weights, and therefore has an impact on the output of the
beamformer. So, you may want to make sure that your covariance matrix is
representative of the signal you are interested in and also a good
estimate (involving enough data samples / a long enough time window).
The beamformer weights are then applied to your original time series,
thus, your source signal has the same temporal resolution as the
original signal (note that the beamformer is a spatial filter that
defines the contributions of your channels to a given source space
position).