6
$\begingroup$

The orthogonal Procrustes problem seeks a matrix $M$ that minimizes $||AM-B||_F$ subject to $M^TM=I$, where $M$ is $d\times d$ and both $A$ and $B$ are $n\times d$. Geometrically, $M$ rotates a set of $n$ vectors (the rows of $A$) so that they best align with another set of $n$ vectors (the rows of $B$) in a least-squares sense.

It has been shown that there is an elegant closed-form solution to this problem, namely $M=UV^T$ where $A^TB=U\Sigma V^T$.

I am interested in solving a similar problem and wondering if there is an equally elegant closed-form solution. Given vector sets $A$ and $B$ as before, I would like to find $M$ that makes the vectors in $A$ as "orthogonal" as possible to corresponding vectors in $B$. That is, I want $a_i^TMb_i$ to be as close to zero as possible for $i\in[1,n]$, where $a_i$ and $b_i$ represent particular rows of $A$ and $B$:

$M^*=\underset{M}{\operatorname{argmin}}\sum_i (a_i^TMb_i)^2$ subject to $M^TM=I$.

Is there a name for this formulation? Anyone know of a solution that can be easily expressed (and computed) as a function of $A$ and $B$, or a reason that it cannot?

$\endgroup$

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.