nLab convex mixture

Idea

Convex mixtures are to convex combinations as integrals are to sums.

In other words, convex mixtures are given by integrating a function with respect to a measure whose total normalization is one (hence, a probability measure).

The concept is used in analysis and geometry as a generalization of midpoints, as well as in probability theory to take expectation values of random variables.

In terms of category theory, it is tightly related to the Giry monad and other probability monads.

Constructions

For real numbers

Let $X$ be a measurable space. We can view a function $f \colon X\to\mathbb{R}$ as an indexed collection of real numbers. If $X$ is finite, one can take the average, or midpoint:

$$\frac{f(x_1)+\dots+f(x_n)}{n}$$

More generally, given a probability distribution $p$ on $X$, one can take the weighted average, or convex combination,

$$\sum_{x\in X} f(x)\,p(x) .$$

The resulting number is a convex combination of the numbers $f(x)$.

Convex mixtures generalize this to the infinite case: if the function $f$ is measurable, given a probability measure $p$ on $X$, one can take, if it exists, the integral

$$\int_X f(x) \,p(d x) \,.$$

One can view the resulting value as a mixture, analogous to a convex combination, of the points $f(x)$. (More precisely, of the points in the support of the pushforward measure $f_*p$ on $\mathbb{R}$.)

In the language of probability theory, we are taking the expectation value of the random variable $f$ on the probability space $(X,p)$.

A very similar construction can be given by replacing $\mathbb{R}$ with a generic Banach space, using Bochner integrals?.

For probability measures

Given measurable spaces $X$ and $Y$, we can consider the space of probability measures $P Y$ on $Y$, and equip it with its canonical sigma-algebra (see Giry monad). This way, a measurable function $f:X\to P Y$ is equivalently a Markov kernel $k_f:X\to Y$. Given a probability measure $p$ on $X$, the resulting convex mixture is, equivalently, the composition of the Markov kernel with the measure $p$ (seen as a kernel from the one-point space): for every measurable subset $B$ of $Y$,

$$\int_X f(x)(B)\,p(d x) \;=\; \int_X k_f(B|x)\,p(d x) .$$

In other words, mixtures of measures are equivalently compositions in the category of Markov kernels.

In particular, taking $f$ to be the identity (i.e. nonparametrically), a mixture of probability measures on $Y$ is given by integrating a probability measure over probability measures over $Y$, $\pi\in P P Y$:

$$B \;\mapsto\; \int_{PY} q(B) \, \pi(d q) .$$

This is equivalently the multiplication of the Giry monad and of most probability monads (more below).

In terms of monad algebras

The Giry monad is the most general monad of probability measures on the category of measurable spaces. Algebras of the Giry monad, therefore, can be interpreted exactly as spaces where one can form arbitrary convex mixtures. This is analogous to how convex spaces, where one can form arbitrary convex combinations, are the algebras of the distribution monad.

In particular, mixtures of probability measures can be seen as instances of the multiplication of the Giry monad, or more generally, as compositions of Markov kernels. Indeed, Markov kernels are the Kleisli morphisms of the Giry monad, and hence their composition is defined in terms of the monad multiplication.

Similar notions of mixture can be given using other probability monads. Note that, in order to have a convex mixture, one needs a form of normalization.

Examples

• In probability theory, the expected value of an integrable random variable is a convex mixture.
• In physics, convex mixtures give the center of mass? of non-discrete bodies.
• The multiplication of the Giry monad, and of most probability monads, consists of taking a convex mixture of probability measure.
• The ergodic decomposition theorem says that in some cases, every invariant measure is a convex mixture of ergodic ones.

See also

• Lebesgue integral, Bochner integral?, expectation value
• convex space, convex combination, mean
• probability monad, Giry monad, distribution monad
• ergodic decomposition theorem, de Finetti theorem

category: probability, analysis, functional analysis