2
$\begingroup$

I have a time-series observation dataset that has been distorted. I want to recover the best approximation of the original signal as possible. Disclaimer:: I know only the basics of linear algebra, so please bear with me.

I have a good model of the distortion, represented by a square matrix. In theory, if I can find an inverse of the transformation, I can recover the original signal. However, the model of the distortion as a matrix is ill-conditioned (i.e. almost singular).

Is it possible to take this matrix model, and generate an invertible matrix that is an approximation of the original matrix?

$\endgroup$
1
  • $\begingroup$ The most robust way I know of is the SVD. $\endgroup$
    – MBaz
    Jun 24, 2016 at 17:25

2 Answers 2

1
$\begingroup$

The usual way to approach inverting a matrix that is rank deficient is to use a generalized inverse.

This means if your matrix equation is: $$ y = \mathbf{A}x $$ where $\mathbf{A}$ is an $n \times m$ matrix, then you can use, for example, the Moore-Penrose pseudo-inverse, $\mathbf{A}^{+}$: $$ \mathbf{A}^{+} = (\mathbf{A}^*\mathbf{A})^{-1} \mathbf{A}^* $$ if a left inverse is required or $$ \mathbf{A}^{+} = \mathbf{A}^*(\mathbf{A}\mathbf{A}^*)^{-1} $$ for the right inverse.

If $n=m$ then either be used.

However, as per the link above, there are many different possibilities for selecting an inverse in the rank-deficient / non-square case because the system of equations is underdetermined.

As per @MBaz's comment, you can calculate this using the singular value decomposition of $\mathbf{A}$.

$\endgroup$
1
  • $\begingroup$ Yes, I calculated the Moore-Penrose pseudo-inverse with Mathematica. However, the results were unsatisfactory, i.e. the pseudo inverse was not very good at recovering the original signal. $\endgroup$
    – Justin G
    Jun 27, 2016 at 21:02
0
$\begingroup$

Simple Answer

A simple way to solve this is by using diagonal loading (see this answer for a related example). If your square matrix is $R$, then instead of inverting $R$, invert $R + \sigma I$, where $\sigma$ is a small value.

Why This Helps

An ill-conditioned matrix (as you probably know) has some near-zero eigenvalues. The condition number of the matrix is the ratio of its largest and smallest eigenvalues, so you can see why you would run into problems with near-zero eigenvalues. Diagonal loading increases all the eigenvalues slightly, moving the smallest ones away from zero and improving the condition number of the matrix.

$\endgroup$
2
  • $\begingroup$ The problem with this method is that it introduces a basis-dependent bias which is hard to reason about. I wouldn't recommend it. $\endgroup$
    – Jazzmaniac
    Feb 29 at 15:09
  • $\begingroup$ It's not perfect, but if computation time is part of the equation, it is less demanding than SVD, for example. It is therefore used in some practical cases. $\endgroup$
    – Gillespie
    Feb 29 at 16:46

Your Answer

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

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