Different from Arnoldi, here we suppose matrix A is an n x n Hermitian matrix and we plan to find the m most valueable eigenvalues. In the following, we will introduce Lanczos iterative algorithm.
Let matrix V be an n x m matrix, which can be written as , where are orthonormal Lanczos vectors. Let matrix T be a m x m tridiagonal real symmetric matrix with , which can be also written as
Note that can be expressed as AV = VT. Comparing the j-th column of both sides, we have
Furthermore, we have
In addition, we have
by multiplying at both sides of since , , and are orthogonal vectors.
However, in order to guarantee numerical stability, and should add additional terms. Therefore, they are expressed as follows:
since and are orthogonal vectors.
The algorithm is shown as follow, which is from wiki: