2
$\begingroup$

I am considering the model :

$$ \mathbf{x} = \mathbf{A}\mathbf{s} $$

where $\mathbf A \in \mathcal{M}_{n,p}(\mathbb{R})$ and $\mathbf s \in \mathbb{R}^{p}$ such that the entries of $\mathbf s$ are i.i.d. with $\text{Laplace}(1/2)$ distribution. In other words:

$$ p(\mathbf{s}) = \prod \limits_{i=1}^{p} \frac{1}{2} \, \exp\left( -\frac{1}{2}\vert s_{i} \vert \right) = \frac{1}{2^{p}} \, \exp \left( -\frac{1}{2} \Vert \mathbf{s} \Vert_{1} \right). $$

I am wondering whether this is identifiable. Because if I replace $\mathbf{A}$ with $-\mathbf{A}$ and $\mathbf{s}$ with $-\mathbf{s}$, I will obtain the same observations $\mathbf{x}$ and the components of $-\mathbf{s}$ are still i.i.d. with $\text{Laplace}(1/2)$ distribution. Am I right? How can I make this model identifiable? Shall I choose another distribution instead of the Laplace distribution?

$\endgroup$
2
$\begingroup$

No, it's not identifiable, in the sense that you can choose some matrix $\boldsymbol J$ which is diagonal with every nonzero being $\pm 1$ and do:

$\boldsymbol x = \boldsymbol A \boldsymbol J \boldsymbol J \boldsymbol s$

So that $\boldsymbol A'=\boldsymbol A \boldsymbol J$ and $\boldsymbol s'=\boldsymbol J \boldsymbol s$ but $p(\boldsymbol s)=p(\boldsymbol s')$.

So yes you are right.

One option is to sample a $\boldsymbol J$ at each iteration (guessing this is frequentist though). Another is to constrain $\boldsymbol s$ or $\boldsymbol A$ to be positive on the understanding that any appropriate (column, entry) pair could be flipped in sign.

$\endgroup$
0

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.