Why is cross-correlation not defined in a normalized sense?

47 Views Asked by At

When correlation is defined in systems and signals, as well as in the statistical sense, it is defined as a normalized measure with respect to the Cauchy-Schwarz inequality.

$\space$

In systems and signals, correlation is defined as

$$\rho = \frac{ \langle x(t), \space y(t) \rangle}{\| x(t)\|\|y(t)\|} = \frac{1}{\sqrt{E_xE_y}}\int_{-\infty}^{\infty}x(t)y^*(t)dt$$

Here, the inner product is normalized by dividing it by the square root of the energies, or the norms, of the two signals in question. The Cauchy-Schwarz inequality for this case is stated as

$$ |\langle x(t), \space y(t)\rangle| \leq \| x(t)\|\|y(t)\|$$

$\space$

In statistics, correlation is defined as

$$\rho_{xy} = \frac{Cov(X,Y)}{\sigma_x\sigma_y}$$

Here, the covariance is normalized by dividing it by the square root of the variances, or the standard deviations, of the two random variables in question. The Cauchy-Schwarz inequality for this case is stated as

$$|Cov(X,Y)| \leq \sigma_x\sigma_y$$

$\space$

So, it seems as if there is a universal trend for defining correlation in a normalized sense. That is until we get to cross-correlation. For some reason cross-correlation is defined as

$$ \psi(\tau) = \langle x(t), \space y(t-\tau) \rangle = \int_{-\infty}^{\infty} x(t)y^*(t-\tau)dt $$

Here, the cross-correlation is simply the time-shifted inner product. No normalization has been applied in this definition.

$\space$

So, why do we define correlation in a normalized sense across the board, but when it comes to its' extended time delayed version, cross-correlation, we "forget" to normalize it all of a sudden?