2
$\begingroup$

According to the wikipedia(Link), autocorrelation has two definition.

In statistics, the definition of the autocorrelation between times $s$ and $t$ is like the following: $$\displaystyle R(s,t) = \frac{\mathbb{E}[(X_t-\mu_t)(X_s-\mu_s]}{\sigma_t\sigma_s}$$


However, in signal processing, the above definition is often used without the normalizaiton.

According to digital communication written by Bernard Sklar, $$R_x(\tau)=\int_{-\infty}^{\infty}x(t)x(t+\tau)du$$

$$R_X(\tau)=\mathbb{E}\{X(t)X(t+\tau)\}$$

Where \begin{array}{a} \tau & \mbox{ is the difference between } t \mbox{ and }s.\\ x(t) & \mbox{ is real-valued energy signal.}\\ X(t) & \mbox{ is a random process. }\end{array}

$\endgroup$
  • 1
    $\begingroup$ sometimes all we care about are relative values of the autocorrelation. normalizing it simply means that it's relative to $R_x(0)$. $\endgroup$ – robert bristow-johnson Jan 22 '16 at 5:51
1
$\begingroup$

Autocorrelation (sometimes also termed aucovariance) is often used in the context of stationnary time signals. As the $\sigma_t$ and $\sigma_s$ are "equal". Since the autocorrelation is symmetric and peaks at $\tau=0$, $R_x(\tau)$ (as noted by @robert bristow-johnson) is often regarded as the natural reference. In a lot of applications, one mostly look at its scale-variant properties: as a measure on how signal chunks look similar, or modeling its decay, or by fitting a linear model through it.

Sometimes, to distinguish the normalized version, it is called an autocorrelation coefficient function.

When dealing with non-stationnary signals, it would be unnatural to stationnarize it artificially. For instance, the Wigner-Ville distribution is based on the Fourier transform of the unnormalized non-stationary autocorrelation function.

| improve this answer | |
$\endgroup$
0
$\begingroup$

Usually such operation are made as "filtering" technique to improve signal to noise ratio. a threshold than is set to derive probability of detection given a certain false alarm rate. These two probabilities are not impacted by the scaling of the two signals - thus the normalization is nor required and the processing associated with it is saved.

| improve this answer | |
$\endgroup$

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.