I am trying to find a unitary tramsformation, $M$, that minimizes $\Vert MA-B \Vert_F^2$ where $A$ and $B$ are $N\times L,\;L\ge N$.

I know how to solve it without the unitary constraint. I thought using Lagrange multipliers with the constraint $\Vert M^HM-I \Vert_F^2 = tr\left\{ \left( M^HM-I \right)^H \left( M^HM-I \right)\right\} = 0$ but the it is quite difficult to solve. Is there any simpler way?


  • $\begingroup$ Why wouldn't the constraint $M^HM=I$ be appropriate? (I may not grasp the context of the question fully -- it looks to me like you are talking about unitary matrices $M$.) $\endgroup$ Feb 25 '14 at 16:29
  • $\begingroup$ @JeffSnider I am indeed talking about a unitary matrix $M$. However, $M^HM=I$ imposes $N^2$ constraints to my understanding. $\endgroup$
    – ThP
    Feb 25 '14 at 20:10
  • $\begingroup$ Since $M^HM$ is Hermitian and the non-diagonals are constrained to be zero it imposes only $N(N+1)/2$ constraints. There may be a special way to ensure $M$ is unitary other than that constraint, but I don't know what it would be. $\endgroup$ Feb 25 '14 at 20:36
  • 2
    $\begingroup$ This is the Orthogonal Procrustes Problem ( en.wikipedia.org/wiki/Orthogonal_Procrustes_problem ) $\endgroup$ Aug 8 '16 at 10:32

Since $M$ is unitary you can write $$\Vert MA-B\Vert_F^2=\Vert MA\Vert_F^2-2\langle MA,B\rangle_F+\Vert B\Vert_F^2\\ = \Vert A\Vert_F^2-2\mathrm{Re}\langle M,BA^*\rangle_F+\Vert B\Vert_F^2.$$ Therefore, your optimization reduces to maximizing $\mathrm{Re}\langle M,BA^*\rangle_F$. Using H$\ddot{\text{o}}$lder inequality and SVD of $BA^*=USV^*$, it's straightforward to show that the maximum is the Schatten 1-norm of $BA^*$ which is attained at $M=UV^*$.

  • $\begingroup$ Can you explain how to show that the maximum of $\mathrm{Re}\langle M,BA^*\rangle_F$ is obtained at $M=UV^*$? $\endgroup$ Feb 26 '14 at 9:17
  • 1
    $\begingroup$ First, use the fact that $\Vert M\Vert = 1$ and H$\ddot{\text{o}}$lder inequality to show that $\mathrm{Re}\langle M,BA^*\rangle_F$ is upper-bounded by the Schatten 1-norm of $BA^*$. Then, simply verify that this bound is attained at $M=UV^*$. $\endgroup$
    – S.B.
    Feb 26 '14 at 10:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.