Tuesday, November 20, 2012

Orthogonal Series and Wrapped Circular Distribution

This is just a quick follow-up on the previous posting.

recommended reading: Mardia and Jupp, section 3.5.7 on wrapped distributions http://www.amazon.com/Directional-Statistics-Kanti-V-Mardia/dp/0471953334

To construct a wrapped distributions on a circle, we can take a distribution that is defined on the real line, like the normal, cauchy, t or stable distribution, and wrap it around the circle. Essentially it's just taking the support modulo (2 pi) and adding the overlapping densities. For some distributions the wrapped density has a nice closed form expression, for example the wrapped cauchy distribution that is also available in scipy.stats.
For other distributions, the density is given as infinite sum, that however converges in many cases very fast.
Mardia and Jupp show how to construct the series representation of the wrapped distribution from the characteristic function of the original, not wrapped distribution.
The basic idea is that for circular wrapped distributions the characteristic function is only evaluated at the integers, and we can construct the Fourier expansion of the wrapped density directly from the real and imaginary parts of the characteristic function. (In contrast, for densities on the real line we need a continuous inverse Fourier transform that involves integration.)

To see that it works, I did a "proof by plotting"

For the wrapped Cauchy distribution, I can use scipy.stats.wrapcauchy.pdf as check. For both wrapped Cauchy and wrapped normal distributions, I also coded directly the series from Mardia and Jupp's book (pdf-series1). I also draw a large number (10000) of random numbers to be able to compare to the histogram. The generic construction from only the characteristic function is pdf-series2-chf in the plots. I used 10 terms in the series representation.
The plots are a bit "boring" because all 2 resp. 3 lines for the density coincide up to a few decimals

Here's the wrapped Cauchy:


And here's the wrapped normal distribution:

No comments:

Post a Comment