Near-Application of Cauchy-Schwarz Inequality

120 Views Asked by At

I have the following situation:

I have two estimators of $\alpha$, both via maximum likelihood of the density:

$$ f(x,y\mid \alpha,\beta) = f(y \mid x,\alpha,\beta)f(x \mid \alpha) $$

One uses only $f(x \mid \alpha)$ and has variance $V_{x,\alpha \alpha}^{-1}$ via Fisher Information.

The other has variance $ (V_{x,\alpha \alpha} + V_{y,\alpha \alpha} - V_{y,\alpha \beta} V_{y,\beta \beta}^{-1} V_{y,\beta \alpha})^{-1}$ (once again via Fisher Information) by using the full likelihood.

Now I am supposed to show that the first estimator is less efficient:

$$ V_{x,\alpha \alpha}^{-1} - (V_{x,\alpha \alpha} + V_{y,\alpha \alpha} - V_{y,\alpha \beta} V_{y,\beta \beta}^{-1} V_{y,\beta \alpha})^{-1} \geq 0 $$

This is the same as:

$$ 1 \geq \frac{V_{y,\alpha \beta} V_{y,\beta \alpha}}{V_{y,\beta \beta} V_{y,\alpha \alpha}} $$

Which reminds of the Cauchy-Schwarz inequaliy, although I am not so sure how to apply this to the case at hand, since $V$'s are derivatives of the likelihood.

(If notation wasnt clear, $V_{w,\delta\nu}=\frac{\partial^2 f(w)}{\partial \delta \nu}$).

Can anybody provide me any insight? Maybe define some distance function using the derivatives and use Cauchy-Schwarz?

Thanks.