Least Squares and perturbations

526 Views Asked by At

Consider

$Y = \theta X + \Delta$

where $\theta$ is the unknown matrix to be found, $Y, X$ are the data matrices of finite length $T$

$Y = [y_0 \ y_1 \ \dots \ y_T], \ X = [x_0 \ x_1 \ \dots \ x_T]$

and $\Delta$ is an unknown deterministic perturbation matrix, which has the same structure of $Y,X$ (can be Hankel or Toeplitz)

$\Delta = [\delta_0 \ \delta_1 \ \dots \ \delta_T]$.

When the perturbation is not present, the solution $\theta$ is found by Least Squares

$\theta = Y X^\dagger$

where $X^\dagger = X^\top (XX^\top)^{-1}$. When $\Delta$ is measurable, so it is known,

$\theta = (Y -\Delta)X^\dagger$.

The question is: how to correctly estimate $\theta$ when we don't have any information about $\Delta$ (I can only assume a structure and some bounds on the norm)?

Maybe total least squares (LS) can be useful in this way? Are there methods of Robust LS that can guarantee a correct estimate having finite samples of data?

1

There are 1 best solutions below

0
On

Notice that by construction $\Delta = Y -\theta X$, so if the basic elements of $\Delta$ are in $\mathbb{R}$ then the OLS estimator or MLE exists in $\mathbb{R}$. If you want a consistent OLS or MLE estimator, since the OLS or MLE is $\hat\theta=\theta_0+(X'X)^{-1}X'\Delta$, you must assume that plim $n^{-1}(X'\Delta)=0$ and that $\lim_{n\to \infty}n^{-1}X'X=Q$ where $Q$ is not singular.