Circular distributions

[ghost]

May math knowledge is quite limited. But in the past, I more than once had to deal with circular data. I usually use a Von Mises distribution to approximate my data. Having support in Boost.Math would have been very helpful back then.

[NAThompson]

Not an unreasonable request. Looks like all the building blocks are already in Boost.Math. Got a number of other things to do at the moment, but I'll try to get this added as time permits.

[ghost]

Is there anything I can do to assist you?

[NAThompson]

@pcjmuenster : If you want to submit a PR I could help with the details. I'd probably just cp normal.hpp von_mises.hpp and just make the necessary adjustments to make it a Von Mises distribution.

[NAThompson]

Looking at the Von Mises distribution wiki, the mean, variance, mode, pdf and entropy are straightforward. The only one that would require significant testing is the CDF.

Also a decision would have to be made about the support(), since [-pi,pi] is just as good as (say) [0, 2pi].

In fact this is a good first task for anyone who wants to learn how to contribute to the library.

[ghost]

@NAThompson Thanks, I will try to look into it this weekend.

[ghost]

Notes in case of my unforeseen death: For large values of kappa, the naive computation of the PDF fails since both exp(kappa) and bessel_i0(kappa) become quite large. For instance with kappa=100, the values exceed the range of single precision floating point numbers. However the maximum of such a distribution is still a fairly small number (cf. Wolfram Alpha). In a first attempt, I tried to approximate the logarithm of bessel_i0(kappa) instead and exponentiate the difference. I used the Hankel Expansions for this. Boost currently uses different sets of polynomials to approximate the Bessel functions but doesn't yet include functions to directly retrieve their logarithms. For large kappa these could be trivially adapted from the polynomial approximations. Suppose we use a similar a approach to calculate the variance. The quotient of bessel_i1 and bessel_i0 converges fast to 1. This could lead to a loss of precision when we substract the result from 1. I have yet to figure out how severe that is. (A simple test showed that under the assumption we could compute the logarithms of the Bessel functions exactly – which we can not – the error is 0.05% from the true variance.)

[NAThompson]

@pcjmuenster : Could direct numerical quadrature work in the case where the series converges slowly? We have quite a few quadrature routines available, though some require expensive node-weight precomputation.

Note that the general goal is 1ULP evaluation in arbitrary precision; see here; you can create the plots using my garbage software here.

[ghost]

I have worked on the CDF yesterday and came across the following issue: The literature gives an algorithm that is also used in all other libraries I could find. The algorithm assumes the graph is centered around the mean. Thus the support would be [mu-pi, mu+pi]. I'd like to know if this kind of instance-dependent – in contrast to class-dependent – beahvior is allowed according to the contract of support. The alternative is to shift the tails around to fit them into the "standard" interval. That way the CDF would not be symmetric anymore. This also affects the implementation of quantile. I can't see any reasonable definition for "the lowest X%".

[NAThompson]

@pcjmuenster : My opinion, such as it is, is that as long as it is documented, it's fine to have the support be [mu-pi, mu+pi].

@jzmaddock, @pabristow : What's your opinion?

[pabristow]

Glad you are yet to meet an unforeseen premature demise ;-)

[mu-pi, mu+pi] looks fine by me - it only vital to do what it says on the tin.

I don't think we have been very precise what 'support' is anyway?

[NAThompson]

Yeah, for periodic functions, support() is a bit ambiguous. Maybe it should be the whole real line?

[ghost]

Independent of the return value of support, a decision has to be made where cdf shall return 0 and 1, this also impacts quantile and median. The latter would (kind of) reasonably be equal to mean like the "normal" normal distribution. Otherwise, somebody has to find a formula for skewness and kurtosis cause I am totally out of my field then … range, i. e. the domain of pdf, can naturally be the set of real numbers. But support … that would be weird. OTOH: being able to calculate the PDF but not the CDF for angles outside of [mu - pi, mu + pi] is also counter-intuitive. In the end, I think it would make even more sense if one had to ask directly for an interval of angles instead of substracting to CDFs. Then some magical formula would provide the area of the circular segment between them (counter-clockwise).

[NAThompson]

@pcjmuenster : I think this will require writing very careful error messages that self-document the requirements of the function.

I might start with the definition support() := [µ-π, µ+π). Then cdf(µ-π) := 0, and cdf(µ+π) := 1. Don't allow pdf to be evaluated anywhere, throw when x \not \in [µ-π, µ+π].

Also, should µ be constrained?

[ghost]

I tried every trick I know (not many) to increase the precision. But even then most functions don't work for distributions with kappa > 10. Maybe somebody else is able to complete the work.

[pabristow]

@jzmaddock is a man who knows a very lot of precision tricks.

https://en.wikipedia.org/wiki/Von_Mises_distribution says

"as {\displaystyle \kappa }\kappa increases, the distribution approaches a normal distribution in x with mean μ and variance 1/{\displaystyle \kappa }\kappa "

So would normal do at some (computable) cross-over point?

Or be close enough approximation that some 'adjustment' could be made?

[NAThompson]

@pcjmuenster : Open a PR and we can see what you're doing.