# Article

**Keywords:**

sparse direct solver; hierarchical condensation; finite element method; sparse matrices; algorithm

**Summary:**

We present a method that in certain sense stores the inverse of the stiffness matrix in $O(N\log N)$ memory places, where $N$ is the number of degrees of freedom and hence the matrix size. The setup of this storage format requires $O(N^{3/2})$ arithmetic operations. However, once the setup is done, the multiplication of the inverse matrix and a vector can be performed with $O(N\log N)$ operations. This approach applies to the first order finite element discretization of linear elliptic and parabolic problems in triangular domains, but it can be generalized to higher-order elements, variety of problems, and general domains. The method is based on a special hierarchical enumeration of vertices and on a hierarchical elimination of suitable degrees of freedom. Therefore, we call it hierarchical condensation of degrees of freedom.