Closed aarchiba closed 3 years ago
Fixing it seems to lead to discrepancies with tempo2 and with the reduced-rank versions, although not large discrepancies. Test failures, though.
The problem occurs specifically if you have a noise model but not EFAC/EQUAD.
In PINT's toa_covariance_matrix, if there is no noise model, you get a diagonal matrix with the square of each TOA's uncertainty on the diagonal. If there is a noise model, though, this term is not added, only the covariances coming from the noise components - so in particular if your noise model is just ECORRs the covariance matrix has zero eigenvalues and Cholesky decomposition fails. I do not think this can be correct.
https://github.com/nanograv/PINT/blob/9250b5a1d487acf866bdd5605ebd212d758bbb4e/src/pint/models/timing_model.py#L1106
yields:
Also:
yields