Computing Empirical Orthogonal Functions (EOFs) can take a lot of computer time. First the covariance matrix is estimated, this takes N_x²N_t time, with N_x the number of spatial points and N_t the number of time steps. Next the eigenvalues are computed, taking N_x³ time.

The computation can therefore be sped up by computing the EOFs on a coarser grid than the orginal data, reducing N_x. If one averages over two grid boxes in longitude and two in latitude, N_x is reduced by a factor 4, and the EOF are computed between 16 and 64 times faster.

If an EOF computation takes too long, please kill it (using the link provided on the next page) and retry with larger numbers in these fields.

There are more efficient algorithms to compute he EOFs, but as far as I know these do not work when there is missing data. I plan to implement a faster method for fields without missing data in the future.

The covariance between two grid points is only considered valid when percentage of the time series both have valid data. Enter a smaller number to get more valid data in the EOFs, but the quality of these data will be lower. A higher number gives fewer but higher-quality data points. At very low values the EOF procedure will fail, as the estimate of the covariance matrix is no longer positive-definite.