English:
Let . Since , is an orthogonal projection matrix of rank .
The objective is to minimize . is the projection of the column space of onto the -dimensional subspace spanned by , ensuring .
By the Eckart-Young-Mirsky Theorem, the matrix that minimizes subject to is .
Thus, we need to find such that .
Calculating for , we get:
(where represents columns to of ).
Therefore, the column space of must span the same space as the columns of .
The solution, using any orthogonal matrix (), is expressed as:
Solution: (The most basic solution is )
English:
, which is a symmetric matrix with eigenvalues .
The optimization problem is subject to .
By Ky Fan's Theorem (or the extended Rayleigh-Ritz theorem), the maximum of for an orthonormal matrix is the sum of the largest eigenvalues of the symmetric matrix (in this case, ).
This maximum is achieved when the column vectors of span the eigenspace (principal subspace) corresponding to the largest eigenvalues of .
The eigenvectors corresponding to the largest eigenvalues of are the first columns of , which is .
Therefore, must be an orthonormal matrix spanning the same space as .
Solution: (where is an arbitrary orthogonal matrix)
English:
Since the regular (invertible) matrix is written as , the matrices and must also be regular (invertible).
The objective function to be optimized is .
Let . Because and are invertible, holds for any matrix .
Therefore, the condition is completely equivalent to .
The problem can be rewritten as:
By the Eckart-Young-Mirsky Theorem, the solution to this optimization problem is .
Reverting to the original variable , the desired solution satisfies .
Since and are invertible, we multiply by on the left and on the right.
Solution: