1: \begin{abstract}

2:   We consider the problem of minimizing a function over the manifold of orthogonal matrices.

3:   %

4:   The majority of algorithms for this problem compute a direction in the tangent space, and then use a retraction to move in that direction while staying on the manifold.

5:   %

6:   Unfortunately, the numerical computation of retractions on the orthogonal manifold always involves some expensive linear algebra operation, such as matrix inversion, exponential or square-root.

7:   %

8:   These operations quickly become expensive as the dimension of the matrices grows.

9:   %

10:   To bypass this limitation, we propose the landing algorithm which does not use retractions.

11:   %

12:   The algorithm is not constrained to stay on the manifold but its evolution is driven by a potential energy which progressively attracts it towards the manifold.

13:   %

14:   One iteration of the landing algorithm only involves matrix multiplications, which makes it cheap compared to its retraction counterparts.

15:   %

16:   We provide an analysis of the convergence of the algorithm, and demonstrate its promises on large-scale and deep learning problems, where it is faster and less prone to numerical errors than retraction-based methods.  

17: \end{abstract}

18: