For later discussion, let us first introduce the operator defined by
The adjoint of with respect to the standard inner products in and is the operator
The main step at each iteration of our algorithms is the computation of the search direction from the symmetrized Newton equation (with respect to an invertible matrix P which is usually chosen as a function of the current iterate X,Z) given below.
where and is the centering parameter. Here is the symmetrization operator defined by
and and are the linear operators
where denotes the linear operator defined by
Assuming that , we compute the search direction via a Schur complement equation as follows (the reader is referred to [2] and [8] for details). First compute from the Schur complement equation
where
Then compute and from the equations
If , solving (9) by a direct method is overwhelmingly expensive; in this case, the user should consider using an implementation that solves (9) by an iterative method such as CG or QMR. In our package, we assume that and (9) is solved by a direct method such as LU or Cholesky decomposition.