As the name suggests, circular data is data that is measured on a circle. One example of circular data is direction (e.g. how many degrees from True North); another is time. (Note: Time is interesting because we can think of it as linear (Jan 2019 -> … -> Dec 2019 -> Jan 2020) or circular (Jan -> … -> Dec -> Jan).)
A common probability distribution we see with circular data is the von Mises distribution (named after Richard von Mises, an early 20th-century Austrian mathematician). It is a close approximation to the wrapped normal distribution which is the normal distribution “wrapped” around a unit circle. The von Mises distribution is simpler and more tractable, and hence is more commonly used.
The von Mises distribution has two parameters: , which is a measure of location, and , which is a measure of concentration. For any , the probability density function (PDF) is given by
where is the modified Bessel function of order 0. Below is a picture of the PDF for different values of (taken from Wikipedia). You can see that it looks very much like the normal distribution. As increases, the distribution approaches the distribution. For small , the distribution is very close to uniform, with corresponding exactly to the uniform distribution over the interval.
Below is a really nice visualization of circular data overlaid with the estimated von Mises distribution, taken from this blog post:
The mean, median and mode of this distribution is . The variance is equal to , where is the modified Bessel function of order .
References:
- Wikipedia. von Mises distribution.
- Zeileis, A. Circular regression trees and forests.