Open ivan-pi opened 2 years ago
One more observation, in the Fortran interface we should add support for 1-based indexing.
Since CHOLMOD is a C library, it expects 0-based indexing. This should be handled as an (optional) argument in the call to the optimizer.
By default we'd assume Fortran indexing is used.
The biggest advantage of libdogleg is it supports problems with sparse Jacobians. This can result in huge savings for large least squares problems.
Generally speaking, we just need to recover the sparse storage buffers of the compressed-column sparse (CCS) matrix of type
cholmod_sparse
and fill them with our Jacobian values. In C, the unpacking is done as follows:Notably, the Jacobian matrix is referenced as
J^T
. The arrayJrowptr
is of sizeNmeas + 1
(recall the output vector is of sizeNmeas
). The arraysJcolidx
andJval
are of sizeJnnz
, which is the number of non-zero values in the sparse matrix.The number of non-zeros must be given explicitly to the libdogleg optimizer:
A full description of the CHOLMOD sparse storage format is given in the user guide: https://github.com/DrTimothyAldenDavis/SuiteSparse/blob/master/CHOLMOD/Doc/CHOLMOD_UserGuide.pdf