choice of linear solver in ceres

[amberwood31]

Hi,

Thanks for sharing the code!
I'm curious about the choice of solvers in ceres, which is "SPARSE_SCHUR" for linear solver with the trust region strategy being DOGLEG.
As shown in the functionMarginalizationError::updateErrorComputation(), the reduced camera matrix is enforced to be positive semidefinite by setting small eigenvalues to zero, which means the Jacobian and Hessian could be rank-deficient.
To the best of my understanding, cholmod ("SPARSE_SCHUR is supposed to be based on cholmod from ceres documentation) is used for full-rank matrix.

So, what strategy is being used here to handle the rank-deficiency? Is it handled inside the DOGLEG method?

This might seems more like a ceres related question, but I'm interested in what specific properties in the okvis formulation might have contributed to the choice of solvers.

Any insight is highly appreciated!