MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Suppose $f(t)$ is a continuous-valued, zero-mean stochastic signal with Gaussian autocorrelation (with variance $\sigma^2$). Suppose I then pass this signal through a step function, producing a new $\pm$1 signal $g(t)$ that has value +1 wherever $f(t) \ge 0$, and -1 wherever $f(t) < 0$. This signal is similar to a random telegraph signal, though perhaps with different statistics. My question is: does the autocorrelation of $g(t)$ have a simple form? If it were a random telegraph signal where the switching was described by a Poisson process, then I take it the autocorrelation would decay exponentially with the lag. But the switches between the two states (+1 and -1) don't seem to be independent here. Any guidance or references would be helpful.

More generally, suppose f is composed with a "soft" step function, such as the error function $\mathrm{erf}(u)$. For instance, let $g(t)=\mathrm{erf}(\alpha f(t))$. For large $\alpha$, this approaches the case above. In general, what is the autocorrelation of $g$?

share|cite|improve this question
Why do you make the statement or the claim that the switches are not independent here? Obviously, switches between $+1$ alternate with $-1$, so in one sense of sequence alone with disregard to the time between the sign changes there is a predictable component: $+1$ follows $-1$ follows $+1$... However, the time interval between the sign changes is still a random variable, isn't it? Also, the $\LaTeX$ command you want for $\pm$ is \pm, not \plusminus – sleepless in beantown Jan 5 '11 at 6:28
I could be wrong about the independence (and thanks for thinking about this). I was thinking that the switches are not independent in the sense of a Poisson process. Thank you, Didier, for the answer below. Is there are reference you could point me to? Also, is it possible to generalize this in the case where f is composed with not the sgn function, but the function $h(u) = \erf(\alpha u)$ (i.e., $g(t) = h(f(t))$). For large \alpha, this approaches the case above. Is there a simple form for the autocorrelation of g in general? Thank you. – James Hsieh Jan 5 '11 at 13:32
Hi Didier -- thank you very much for your reply. I've edited my original question to include the more general one. – James Hsieh Jan 5 '11 at 17:14
up vote 4 down vote accepted

Assume $(f(t))_t$ is Gaussian and centered with variance $\sigma^2$ and autocorrelation $E[f(t)f(s)]=c(t,s)$ and let $g(t)=\mathrm{sgn}(f(t))$. Then $(g(t))_t$ is $\pm1$ and centered with autocorrelation $$ E[g(t)g(s)]=\frac{2}{\pi}\arcsin\left(\frac{c(t,s)}{\sigma^2}\right).$$

Now I answer the OP's more general question.

The computation in the sign function case is based on the fact that $(f(t),f(s))$ is a Gaussian vector distributed like $(N,aN+bN')$ with $a=c(t,s)/\sigma^2$ and $b=\sqrt{1-a^2}$, for two independent centered Gaussian $N$ and $N'$ with variance $\sigma^2$. One computes $$ E[g(t)g(s)]=2P[N>0,N'\ge -uN] $$ as a two-dimensional Gaussian integral for the suitable value $u=a/b$ and this yields the result.

Likewise, if $g(t)=h(f(t))$ for a given odd function $h$, $E[g(t)g(s)]=E[h(N)h(aN+bN')]$ and it remains to compute this two-dimensional Gaussian integral.

If $h(u)=\mathrm{erf}(\alpha u)$ as suggested by the OP, (I believe that) one gets $$ E[g(t)g(s)]=\frac{2}{\pi}\arcsin\left(\frac{\alpha^2c(t,s)}{1+\alpha^2\sigma^2}\right). $$

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.