1
$\begingroup$

I would like to basically understand what an autoregressive model is used for (so I don't really attempt maths in the answers).

I just started a signal processing course and the model was introduced.

This is what I understood but I don't think I understood it well :

We have $n-1$ samples of a signal and we want to guess what the $n$'th will be.

With Young Walker, we have an approximation of the $n$'th sample that is : $\tilde{x}[n]=\sum a_k x[n-k]$

But $\tilde{x}[n]$ is just an approximation of $x[n]$.

So we have : $x[n]=\tilde{x}[n]+\epsilon$ where $\epsilon$ is the "error" I make when I use $\tilde{x}[n]$.

The autoregressive models assumes that $\epsilon$ is a white noise with average zero (and this makes sense because there is no reason in our model to have more often $\tilde{x}[n]>x[n]$ than $\tilde{x}[n]<x[n]$ ).

Did I understand the physical meaning of the model well ?

Thanks !!

$\endgroup$
1
$\begingroup$

Your understanding is right. Just a few points...

  • I think you meant Yule-Walker equations.
  • It is better to distinguish $N$ from $n$. In an AR($N$) model we consider last $N$ terms to approximate the next term.
  • An important interpretation that you missed was the fact that when we consider $$\epsilon=x[n]-\tilde{x}[n]=x[n]-\sum_ka_kx[n-k],$$ or in a more standard way (representing the process by variable $y$ instead of $x$ $$y[n]=\sum_ka_ky[n-k]+\epsilon[n],$$ it can be seen as an input-output relationship of an LTI filter, $\epsilon[n]$ being input and $y[n]$ the output. More specifically,

AR model can be seen as the output of an all-pole IIR filter which is excited by a white noise input.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you for your answer. So I assumed that you mainly agreed with what I said you just put further details. But I didn't understand why we can see it as an input-output relation between $y[n]$ and $\epsilon[n]$ only. Indeed we also need all the $y[n-k]$ to compute the ouput. So should't it be a filter with $P+1$ input ($1$ for $\epsilon$ and $P$ for all the $y[n-k]$ needed). $\endgroup$ – StarBucK Nov 7 '16 at 12:09
  • $\begingroup$ Consider an IIR filter with input signal $x$ and output signal $y$. We just input one sample at each time, and get one sample of output ($x[n]$ and $y[n]$). But the only input is $x[n]$. Although technically the filter considers the outputs from some previous times in the calculation of the next sample, they are not inputs of the filter. $\endgroup$ – msm Nov 7 '16 at 17:18
  • $\begingroup$ Does it make sense? $\endgroup$ – msm Nov 8 '16 at 12:42
  • $\begingroup$ Hmm sorry but I don't get it :/. I am not familiar with IIR filter at all (I just saw it on wikipedia now). If at "time" n I need y[n], then I need into my computation $ \epsilon [n]$ AND $y[n-k]$ : I didn't understand why you said that the previous time output are not input of the filter $\endgroup$ – StarBucK Nov 8 '16 at 19:47
  • $\begingroup$ You don't have another way to explain it ? :s $\endgroup$ – StarBucK Nov 11 '16 at 11:37

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.