Lanczos Algorithm

The Lanczos method is an iterative algorithm that reduces a symmetric matrix $A$ of size $N \times N$ to a tridiagonal matrix $T$ of size $m \times m$, where $m \ll N$. The eigenvalues of $T$ approximate the extremal eigenvalues of $A$, and the corresponding eigenvectors can be reconstructed.

Key Steps of the Lanczos Method

1. Initialization:

   • Choose a random starting vector $v_1$ with unit norm.
   • Set $\beta_0 = 0$ and $v_0 = 0$.

2. Iteration: For $j = 1, 2, \dots, m$:

   • Compute $w = A v_j - \beta_{j-1} v_{j-1}$.
   • Compute $\alpha_j = v_j^\top w$.
   • Compute $w = w - \alpha_j v_j$.
   • Compute $\beta_j = \|w\|$.
   • If $\beta_j = 0$, stop; otherwise, set $v_{j+1} = w / \beta_j$.

3. Tridiagonal Matrix: After $m$ iterations, the matrix $T$ is constructed as:

   $$T = \begin{pmatrix} \alpha_1 & \beta_1 & 0 & \dots & 0 \\\ \beta_1 & \alpha_2 & \beta_2 & \dots & 0 \\\ 0 & \beta_2 & \alpha_3 & \dots & 0 \\\ \vdots & \vdots & \vdots & \ddots & \beta_{m-1} \\\ 0 & 0 & 0 & \beta_{m-1} & \alpha_m \end{pmatrix}$$

4. Diagonalization of $T$:

   • Diagonalize $T$ using standard dense matrix techniques (e.g., QR algorithm).
   • The eigenvalues of $T$ approximate the extremal eigenvalues of $A$.
   • The corresponding eigenvectors of $A$ can be reconstructed from the Lanczos vectors $v_j$.

Advantages of the Lanczos Method

1. Efficiency:

   • Only matrix-vector products are required, making it suitable for sparse matrices.
   • Memory usage is $O(N \cdot m)$ instead of $O(N^2)$.

2. Scalability:

   • Works well for very large matrices where dense methods are infeasible.

3. Focus on Extremal Eigenvalues:

   • The Lanczos method is particularly effective at finding the largest or smallest eigenvalues and their eigenvectors.

Challenges and Considerations

1. Loss of Orthogonality:

   • In finite-precision arithmetic, the Lanczos vectors $v_j$ can lose orthogonality, leading to spurious eigenvalues.
   • Remedies include reorthogonalization or using more advanced variants like the Implicitly Restarted Lanczos Method.

2. Choice of $m$:

   • The number of iterations $m$ must be chosen carefully to balance accuracy and computational cost.

3. Convergence:

   • Convergence to the extremal eigenvalues is typically fast, but interior eigenvalues may require many iterations.