1: \begin{abstract} 

2:     A novel orthogonalization-free method together with two specific

3:     algorithms are proposed to solve extreme eigenvalue problems. On

4:     top of gradient-based algorithms, the proposed algorithms

5:     modify the multi-column gradient such that earlier columns are

6:     decoupled from later ones. Global convergence to eigenvectors

7:     instead of eigenspace is guaranteed almost surely. Locally,

8:     algorithms converge linearly with convergence rate depending on

9:     eigengaps. Momentum acceleration, exact linesearch, and column

10:     locking are incorporated to further accelerate both algorithms and

11:     reduce their computational costs. We demonstrate the efficiency

12:     of both algorithms on several random matrices with different

13:     spectrum distribution and matrices from computational chemistry.

14: \end{abstract}

15: