Pradyumna S Upadrashta wrote:
> Hi,
>
> Does anyone know of any (simple) papers that discuss the
relationship
> between nonlinearity and long-autocorrelation times in time-series?
> Many times I have run across the suggestion that "long
> autocorrelation times" suggest nonlinear behavior, but aside from
the
> clearly non-stationarity of such processes, what else can be said
> about long autocorrelation times?
>
> P
>

Ref: Jones DA (1976) Nonlinear autoregressive processes. Proc Roy Soc
Lond A v360 71-95

This paper contains an example of a nonlinear process (1-step
dependent) where the autocorrelations are found exactly and shown not
to match those of a linear AR(1) process. This might be thought to
relate to your question.

However, non-linearity and length of autocorrelation are not directly
related. It is well-known that you can construct a linear process with
essentially any (valid) autocorrelation function you like, so
non-linearity isn't necessary for long acf's.

There may be an argument about "simple structures". If you demand that
a model should have a simple structure (however you might judge this),
then maybe you do need to resort to a simple non-linear structure to
get a long acf. But "simple structure" shouldn't be confused with a
structure you are familiar with, as in ARMA models, where FARMA are
known for their ability to give long acfs.

I don't understand "Many times I have run across the suggestion that
"long
 autocorrelation times" suggest nonlinear behavior, but aside from the
 clearly non-stationarity of such processes, what else can be said
 about long autocorrelation times?" ...

  ....long acfs are not necessarily non-stationary
  ....nonlinear processes are not necessarily non-stationary

David Jones



.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to