Take the 2-minute tour ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Hi all,

Previously I asked a question that currently has no satisfactory answer Least sum squares given constraints on subcomponents

It comes from an engineering problem. I was thinking to formulate it differently and hope that someone becomes interested and/or know how to solve it.

By formulating differently, I will have the following system:

$\mathbf{D} \mathbf{R} \mathbf{\theta} = \mathbf{0}_{2N \times 1}$

D is a $2N \times 6N$ block diagonal matrix that contains unknowns $\mathbf{D} = diag(\mathbf{x}^T - a_1 \mathbf{z}^T, \mathbf{y}^T - b_1 \mathbf{z}^T, \dots, \mathbf{x}^T - a_N \mathbf{z}^T, \mathbf{x}^T - b_N \mathbf{z}^T)$ where $\mathbf{x}, \mathbf{y}, \mathbf{z}$ are 3x1 orthogonal unit vectors that we need to find (ie., $\mathbf{x}^T \mathbf{x} = 1, \mathbf{x}^T \mathbf{y} = 0, \dots$). $a_i, b_i$ are known parameters.

R is a $6N \times M$ matrix that contains only numerical entries (measured and computed). $\mathbf{\theta}$ is a vector of M other unknown parameters.

By doing so, I isolated unknowns in two separate matrices (vectors). However, it is still not trivial.

I tried to solve this, but there is no obvious way. One way is that I tried to find $\mathbf{x}, \mathbf{y}, \mathbf{z}$ such that $\mathbf{D} \mathbf{R}$ is rank-deficient. I'm not sure if this can be solved, either exactly or in least-squared, in closed-form or numerically. Any idea, discussion is appreciated.

edit: I was not clear. If I set determinant of any MxM sub-matrix of $\mathbf{D} \mathbf{R}$ to 0 and express it as a function of elements in $\mathbf{x}, \mathbf{y}, \mathbf{z}$, then I have a number of homogeneous polynomials. That's why the title comes.

edit2: M is much smaller than 2N.

share|improve this question

1 Answer 1

One suggests that we can have a good prior for $\theta$ (but not for $\mathbf{D}$). Then, he proposed that we fix $\theta$, find $\mathbf{R}$ for least square $|| \mathbf{D} \mathbf{R} \theta||$ given the constraints above; then fix $\mathbf{R}$ and find $\theta$ for least square $|| \mathbf{D} \mathbf{R} \theta|| $ given $||\theta|| = 1$. Then keep repeating (fix $\theta$ then $\mathbf{R}$) until they converges.

The latter least square is easy, and the former, I believe there is a closed form or numerical solution.

But my question is if such numerical scheme works? Is there any proof for it? What is the name of this method? I could not recall this numerical scheme.

Please enlighten.

share|improve this answer
    
I don't understand why you can vary $R$. Doesn't that represent some measurement which is fixed? From your question I inferred that $D$ and $\theta$ contain all the unknowns you want to solve for. Assuming you meant $D$ rather than $R$, then the iterative procedure you described is simply partial optimization; for instance to minimize $f(x,y)$, we first fix $x$ and minimize $y$ and then minimize $x$ for the fixed $y$. This certainly doesn't converge to the global minimum in general, but will if the graph of $f$ is strictly convex. –  John Jiang Mar 14 '12 at 0:55

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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