Markov Processes for Category Theorists

Posted on by Owen Lynch


Recall the definition of an affine space over a ring A: an affine space is a space X equipped with the ability to take “affine combinations” of points in X, i.e. a_1 x_1 + \ldots a_n x_n where a_1 + \ldots + a_n = 1.

We can similarly define an “affine space” over a rig (ring without negation). An “affine space” over the rig \mathbb{R}_{\geq 0} is precisely a convex space. Using this definition makes it obvious how to define “convex morphisms”: a convex morphism is one that preserves affine combinations over \mathbb{R}_{\geq 0}. Note that this is different from the typical definition of a “convex function”, which has

f(\lambda x + (1-\lambda)y) \leq \lambda f(x) + (1-\lambda)f(y)

However, this definition of “convex morphism” has

f(\lambda x + (1-\lambda)y) = \lambda f(x) + (1-\lambda)f(y)

Soon we will pick a different name for “convex morphism” that fixes this difficulty.

The most important convex spaces are the n-simplices:

\Delta^n = \{x \in \mathbb{R}_{\geq 0}^{n+1} \mid x_0 + \ldots + x_n = 1 \}

These are important for many reasons, one of which is that an element of \Delta^n represents a probability mass function on n+1 elements. More generally, the set of (probability) measures on any space X with sigma-algebra \Omega of What has previously been called a stochastic map is precisely a convex map from a probability simplex to itself.

I propose that we should generalize the concept of stochastic map to cases where to domain and codomain are not a probability simplex. Namely, I propose that we define a stochastic map to be what we previously called a “convex morphism”.

We then can consider the category of convex spaces and stochastic maps. Call this category \mathbf{Cvx}.

This website supports webmentions, a standard for collating reactions across many platforms. If you have a blog that supports sending webmentions and you put a link to this post, your blog post will show up as a response here. You can also respond via twitter or respond via mastodon (on your preferred mastodon server); through the magic of all tweets or toots with links to this post will show up below (subject to moderation).

Site proudly generated by Hakyll with stylistic inspiration from Tufte CSS