10

I'd like to be able to say that a measure $\mu$ on a measurable space $X$ "is" a morphism $R \to X$, where $R$ is some incarnation of the real numbers in an appropriate category.

In other words, let $Mbl$ be the category of measurable spaces (an object is a set $X$ equipped with a $\sigma$-algebra; a morphism $f : X \to Y$ is a measurable function, i.e. the preimage of a measurable set is measurable). There is a functor $Meas : Mbl \to Set$, carrying $X$ to the set of measures on $X$.

Question: Is there some natural enlargement $\widetilde{Mbl}$ of the category $Mbl$ such that the functor $Meas : Mbl \to Set$ extends to a corepresentable functor $\widetilde{Mbl} \to Set$?

In other words, is there another category $\widetilde{Mbl}$, a fully faithful functor $\iota: Mbl \to \widetilde{Mbl}$, an object $R \in \widetilde{Mbl}$, and a natural isomorphism $Meas \cong Hom_{\widetilde{Mbl}}(R, \iota - )$?

(Of course, the answer is "trivially yes", but the point is that $\widetilde{Mbl}$ should be "natural" in some imprecise sense, ruling out the trivial answer.)

I'd also be happy for an answer where some slight modification of the category $Mbl$ is used. For example, perhaps we need to equip our measure spaces with an ideal of negligible sets or something...

Tim Campion
  • 60,951
  • 2
    I don't quite get what you're asking. A measure is (in particular) a monotone map from the set of measurable subsets to $[0, \infty]$. It is not required to preserve meets or joins in general, unlike the pullback map on measurable subsets, so it is difficult for me to imagine how measures are like pullback maps. Also, the functor that sends measurable spaces to the set of measurable sets is contravariant, not covariant. – Zhen Lin May 22 '23 at 04:13
  • 5
    If you restrict to finite discrete measurable spaces, i.e., finite sets with a real function on them, then the functor Meas is already pretty far from preserving finite products: there are many more measures on X⨯Y than measures on X and Y. So the functor ι would have to behave pretty wildly already on finite measurable spaces, to ensure that Meas can become representable. – Dmitri Pavlov May 22 '23 at 04:16
  • 4
    That being said, if you are willing to go to geometric objects like correspondences, then there is a fairly easy answer: Meas is corepresentable by a point once you embed Meas into the category of (enhanced) measurable spaces and correspondences between them, where a correspondence X→Y is just a (finite) measure on X⨯Y, and correspondences are composed like matrices. Not sure if this matches the intent of the question, though. – Dmitri Pavlov May 22 '23 at 04:18

2 Answers2

6

How about this? (But I will be enlarging $Mbl$ by a functor that is not full.)

Let's make a category whose objects can be called abstract $\sigma$-algebras. I'll call the category $\Sigma$.

A $\Sigma$-object consists of

(1) a set $F$,

(2) a symmetric binary relation on $F$ called "disjointness", and

(3) a rule "countable disjoint union" that assigns to each countable family $x:S\to F$ of pairwise disjoint elements an element $D(x)\in F$. (Here $S$ is a finite or countably infinite set, and "pairwise disjoint" means that $x(i)$ and $x(j)$ are disjoint if $i\neq j$.)

This is subject to some axioms. Call $x:S\to F$ and $x':S'\to F$ disjoint if $x(i)$ is always disjoint from $x'(i')$. The axioms are as follows:

(A1) If $E\in F$ is disjoint from $x(i)$ for every $i\in S$ then $E$ is disjoint from $D(x)$. Note that A1 implies that if $x:S\to F$ and $x':S'\to F$ are disjoint then $D(x)$ and $D(x')$ are disjoint.

(A2) For $x:\lbrace i_0\rbrace \to F$, $D(x)=x(i_0)$.

(A3) The disjoint union of disjoint unions is the disjoint union.

A $\Sigma$-morphism $T:F\to F'$ is a map of sets from $F$ to $F'$ that preserves disjointness and commutes with disjoint union in the sense that:

(M1) Whenever $E_1$ and $E_2$ are disjoint elements of $F$ then $T(E_1)$ and $T(E_2)$ are disjoint.

(M2) When $x:S\to F$ is a countable pairwise disjoint family (i.e. when $D(x)$ is defined) then $D(T\circ S))$ (which is defined, by (M1)) is equal to $T(D(S))$.

There is a functor $\iota:Mbl \to \Sigma^{op}$, taking each measurable space $X$ to the set $M_X$ of all measurable subsets of $X$ (with the usual notions of disjointness and countable disjoint union) and taking a measurable map $X\to Y$ to the map $M_Y\to M_X$ given by $E\mapsto f^{-1}(E)$.

The functor $\iota$ is not faithful, but it becomes faithful if we restrict attention to those measurable spaces in which points are determined by which measurable sets they belong to, which is not a big deal.

Here is the desired representing object. Let $P$ be the set $[0,+\infty]$, declare any two elements to be disjoint, and for a countable family $S\to P$ let $D(x)$ be the sum.

So a $\Sigma$-morphism $F\to P$ is simply a rule assigning an element of $[0,+\infty]$ to each element of $F$ and satisfying additivity for finite and countably infinite (abstract) disjoint unions. In particular a $\Sigma^{op}$-morphism $P\to\iota(X)$ is precisely a measure on the measurable space $X$.

Note that the structure on a set (abstract disjointness relation and abstract disjoint union) that makes it an abstract $\sigma$-algebra brings with it some things. There is an element $0=D(\emptyset)$ that is disjoint from all elements. There is a weak partial ordering: $E'\le E$ iff there exist $E''$ disjoint from $E'$ such that $E$ is the disjoint union of $E'$ and $E''$.

The functor $\iota$ is not full, but some of the new morphisms between measure spaces that appear are things that one might want to use sometimes. If $X$ is a measurable space and $A\subset X$ is a measurable subset, which we then consider as a measurable space in its own right, then there is a "wrong-way" map $\iota(X)\to \iota(A)$ that takes any measurable subset $E\subset A$ to $E\subset X$.

The object $\iota(\emptyset)$ is both initial and terminal.

  • Thanks, I like this a lot! I think a small tweak is needed, since addition of real numbers is not determined by the set of numbers being added (addition is not idempotent!). It seems we need to allow the input $S$ of $D$ to be a sequence, and then demand that it's invariant under permutations of its input. Alternatively, the input of $D$ is a multiset. If $x$ appears twice in $S$, I'm not sure whether we should demand that $x$ is disjoint from itself. – Tim Campion May 22 '23 at 12:12
  • Maybe with an eye toward allowing for signed measures as well, we might only have $D$ be defined on finite inputs, and then handle the $\sigma$-ness by saying something about countable directed sups / infs... – Tim Campion May 22 '23 at 12:15
  • Oops, good point! And I'm quite sure that $x$ should only be allowed to appear twice in $S$ if it is disjoint from itself. – Tom Goodwillie May 22 '23 at 12:21
  • But countable directed sups are not available. In an actual $\sigma$-algebra, the order relation determines both the disjointness relation and the disjoint union (both finite and countably infinite), but that's not the case in the object $P$. – Tom Goodwillie May 22 '23 at 12:25
  • I've done the required tweak now, perhaps correctly. – Tom Goodwillie May 22 '23 at 12:46
  • Another observation: an object in the image of $\iota$ has lots of extra properties, and most of that structure is preserved by $\Sigma^{op}$-morphisms between such objects. I mean finite and countably infinite unions, and relative complements, and (therefore) nonempty finite and countably infinite intersections (but not absolute complements). A morphism $\iota(X)\to \iota(Y)$ factors uniquely as one of those wrong=way maps associated to some $A\in M_X$ followed by a morphism $\iota(A)\to \iota(Y)$ that does take $Y$ to $A$ and therefore does preserve absolute complements. – Tom Goodwillie May 22 '23 at 16:24
  • Another observation: $\Sigma$ has limits, and $\iota:Mbl\to \Sigma^{op}$ preserves colimits, so i suppose there is a right adjoint $\Sigma^{op}\to Mbl$. – Tom Goodwillie May 22 '23 at 16:29
  • It occurs to me that the finitary version of these structures are literally just partial commutative monoids since the disjointness relation is nothing other than the domain of the binary sum operation. – Tim Campion May 23 '23 at 12:25
4

A very natural enlargement of the category of measurable spaces is $\mathsf{Stoch}$, the category of measurable spaces and Markov kernels, with composition given by the Chapman-Kolmogorov equation. This is equivalently the Kleisli category of the Giry monad, see Lawvere 1962 and Giry 1980. By construction, this has the property that the morphisms $1 \to X$ are exactly the probability measures on $X$.

I believe that this still works out just the same if you drop the normalization of probability and consider arbitrary measures instead (though I haven't worked out the details). An interesting intermediate choice could be that of s-finite measures and kernels.

As for why one might one way to consider categories with Markov kernels as morphisms, see this other answer.


As a technical detail, the functor $\mathsf{Mbl} \to \mathsf{Stoch}$ is not faithful, and because of this it perhaps doesn't qualify as an "enlargement" of $\mathsf{Mbl}$ in the sense of the OP. However, it is faithful on morphisms between (e.g.) standard Borel spaces, so restricting both categories to those objects should satisfy your desiderata.

Tobias Fritz
  • 5,785