# Exponential (or other) families of distributions on manifolds.

The exponential family is a general parametrized class of probability distributions on $R^n$ that has many nice properties (ML estimation among them) and includes most of the "standard" distributions one encounters (Gaussian, multinomial, exponential, $\chi^2$ etc).

Are there similarly well-defined parametrized families of distributions for manifold-valued random variables ? Specifically, if you have a general Riemannian manifold ? Or asked another way, is there an equivalent notion of an "exponential family" for a Riemannian manifold ?

-
Why not? If you have a complete Riemannian manifold, it's straightforward to define a probability distribution on it that has a distinguished "center" and the distribution decays as, say, the exponential of the negative of the distance from the center. – Deane Yang Oct 28 '11 at 18:08
You just have to make sure that the "exponential function", however you define it, has finite integral. – Deane Yang Oct 28 '11 at 18:09
I guess my question isn't so much "can it be done" as much as "has it been done for any specific settings and what references do I need to look at". In particular, in the context of statistical estimation . – Suresh Venkat Oct 28 '11 at 18:24
Hi Joe, I'm specifically looking for random vars whose values are points on a manifold. @Deane, information geometry (which I'm familiar with) deals with how to represent families of distributions as manifolds (or submanifolds), which is different to building a probability distribution ON a manifold. – Suresh Venkat Oct 28 '11 at 22:03
I would guess that the answer is `not really'. As far as I know there is not a even a universally accepted definition of the 'normal distribution' on a Remanian Manifold. Probably the closest thing to the normal are those distributions that arise from generalisations of Brownian motion on manifolds. math.northwestern.edu/~ehsu/… – Robby McKilliam Oct 29 '11 at 0:32

## 3 Answers

The book Directional Statistics by Mardia and Jupp discusses concrete examples of distributions on:

1. Surface of the unit hypersphere
2. On Stiefel Manifolds
3. Maybe some others

EDIT In particular, have a look at §13.4.2 that discusses distributions on more general manifolds (e.g., on compact Riemannian manifolds). That section also provides several useful references.

I also recalled another source that might be useful to you:

Matrix Variate Distributions by Gupta and Nagar. In particular, see chapter 8.

-
Thanks. I'll check out both these references. The first one in particular has some elements that seem useful. – Suresh Venkat Oct 29 '11 at 20:58

I claim that the answer is yes. The key point that makes a distribution exponential is that there is a set of sufficient statistics, of the dimension of the manifold, that are additive.

On an $n$-dimensional manifold, chose $n$ functions $T_1$ through $T_n$ to $\mathbb R$. Given an $h(x)$ that is small when these functions are large, there is a unique $A(\eta)$ such that:

$\int h(x) e^{\sum_i \eta_i T_i - A(\eta)}=1$

That's the analogue of an exponential distribution.

Now, which $T_i$ and $h(x)$ should you pick? I don't know. It depends on the manifold.

-

There are several subtle aspects to these questions that I am sure have emerged for you since you first posted this question. The primary one is of course that probability spaces are assumed to have algebraic structures so that moments and other elements for statistical investigation of a random variable can be found -- or at least defined -- through direct analytic methods.

For example, the kth moment of a random variable with values in $(0,1)$ and probability distribution $f(x)$ is $\int_0^1 x^k f(x) dx.$ Note that you need to be able to multiply the values of the random variables to themselves and a real-valued function ($f$). Even multivariate statistics (which lacks multiplication) utilizes the vector-algebraic structure of $R^N$ to calculate moments etc.

All of this is lacking on a general manifold. Embedding the manifold into Euclidean space will not help, since you would typically get expected values, etc. living outside of the manifold, although it would be rather entertaining to tell someone that the expected location of the crash site of a satellite with decaying orbit is the center of the Earth.

Anyway, that's a long-winded introduction to point out that you should check out the following papers:
Pennec, X. "Probabilities and statistics on Riemannian manifolds: a geometric approach" http://hal.inria.fr/inria-00071490/

Pennec, X. "Probabilities and statistics on Riemannian manifolds: Basic tools for geometric measurements." IEEE Workshop on Nonlinear Signal and Image Processing. Vol. 4. 1999.

Bochner in his text ${\it Harmonic\ Analysis\ and\ the\ Theory\ of\ Probability}$ describes stable distributions for general probability spaces, which is picked up for hyperbolic spaces in Getoor, R. K. "Infinitely divisible probabilities on the hyperbolic plane." Pacific Journal of Mathematics 11.4 (1961): 1287-1308.

-