On Tue, Mar 9, 2021 at 2:23 PM Dirk Eddelbuettel <e...@debian.org> wrote: > > > On 9 March 2021 at 14:26, Graham Inggs wrote: > | Control: tags -1 - fixed-upstream + upstream > | Control: notforwarded -1 > | > | > | >From TMB upstream [1]: > | > | > Digging a bit deeper I found that after calling > | > > | > M_cholmod_factorize_p(A, mm, (int*)NULL, 0 /*fsize*/, f, &c) > | > > | > the cholmod_common struct contains bogus, e.g. c.status=14903956 and > c.dtype=15216988. > | > | > | [1] https://github.com/kaskr/adcomp/issues/340#issuecomment-790606775 > > Appreciate the hard-core digging Graham! > > And said call is made by glmmTMB in error, or is the call done by Matrix as > part of exposing CHOLMOD?
Indeed, that's the question ... If glmmTMB calls Matrix-exported C functions and something goes wrong does not necessarily point the finger to Matrix, even though it did not happen for previous versions of Matrix. In this case, it could be that glmmTMB relied on an undocumented Matrix behavior which I had decided to change... though I must say, I would still be surprised.. I'd rather expect something more complicated where at least two factors are leading to the problem you see.. Martin -- Martin Maechler, Seminar fuer Statistik, ETH Zurich
