Hi all, Unfortunately, I must report about a bug. The bug appeared inside oem.m, if O.sxnorm was true.
The O.sxnorm option was introduced to handle cases where the diagonal elements in Sx differ with many orders of magnitude, which can cause numerical problems when OEM has to invert matrices. In practice this is done by an internal rescaling, and a renormalisation is needed for the output. However, I was a bit too quick for the averaging kernel, thought that the terms cancelled each other, missed the ratio was between different elements of x (ie. x_i/x_j, not x_i/x_i). In Qpack2, the sxnorm option is activated automatically, by considering values in Sx. Hence, you could have been using the option without an active selection. That is, your old Qpack2 averaging kernels could be incorrect. The safest option is to recalculate your data. Or at least, compare one case with newest atmlab, with your old inversion. I had done the same mistake in qp2_rel2vmr but as that is a new function, I hope that bug don't affect anyone. These bugs were tracked down thanks to input from Ole-Martin. Bye, Patrick _______________________________________________ qpack mailing list [email protected] https://www.sat.ltu.se/mailman/listinfo/qpack
