Errors for large correlation lengths
For large correlation lengths, discrepancies can occur (e.g. stripes) in the random field. This can be avoided by increasing the embedding factor. However, to the user this is not obvious. (Only for larger verbosity settings, the number of large negative values of the covariance matrix is given).
The minimal solution would be to throw an exception if the large negative values occur, in order to give a clear feedback that the embedding factor should be adjusted.
From a user perspective, the optimal solution would be an option, where the embedding factor is chosen automatically. (To limit the computational effort when increasing the embedding factor, individual factors for each dimension might be useful as well)