Infinite products and zero-one laws in categorical probability

Any Cartesian monoidal category is obviously semicartesian. In the probabilistic setting that we are considering, this corresponds to a situation in which there is no randomness (in such a category, all morphisms are deterministic, cf. Definition 2.6). Hence it is somewhat uninteresting from our point of view.

\operatorname{\mathsf{FinStoch}} is the category of finite sets with stochastic matrices as morphisms. This means that for finite sets X and Y, a morphism f : X \to Y is a matrix (f_{xy})_{x \in X,y \in Y} of nonnegative real numbers such that \sum_y f_{xy} = 1 for every x. Stochastic matrices compose by matrix multiplication, which in this context is also known as the Chapman–Kolmogorov equation. We consider \operatorname{\mathsf{FinStoch}} as a symmetric monoidal category with respect to the cartesian product of sets on objects and the tensor product (Kronecker product) of stochastic matrices on morphisms; the monoidal structure isomorphisms are the obvious ones coming from the embedding \operatorname{\mathsf{FinSet}} \subseteq \operatorname{\mathsf{FinStoch}}.

Let \operatorname{\mathsf{CRing}}_+ be the category of commutative rings with maps which are merely additive and unit-preserving, considered as a symmetric monoidal category in the obvious way with respect to the tensor product of rings. Then since \operatorname{\mathsf{CRing}}_+ has the monoidal unit \mathbb{Z} as its initial object, we can conclude that \operatorname{\mathsf{CRing}}_+^{\mathrm{op}} is semicartesian monoidal.

\operatorname{\mathsf{Stoch}} is the category of measurable spaces (X,\Sigma_X) as objects and Markov kernels as morphisms. We sketch the definition here and refer to [markov_cats, Section 4] and references therein for the details. Given measurable spaces (X,\Sigma_X) and (Y,\Sigma_Y), a Markov kernel (X,\Sigma_X) \to (Y,\Sigma_Y) is a map

assigning to every x \in X a probability measure f(-|x) : \Sigma_Y \to [0,1] in such a way that for every S \in \Sigma_Y, the function x \mapsto f(S|x) is measurable. Intuitively, one can think of f as assigning to every x \in X a random element of Y; or one can think of f as a statistical model, specifying one probability measure f(-|x) for every parameter value x.

Composition of Markov kernels is again defined by a variant of the Chapman–Kolmogorov equation: for f : (X,\Sigma_X) \to (Y,\Sigma_Y) and g : (Y,\Sigma_Y) \to (Z,\Sigma_Z), the composite is given by, for all x \in X and T \in \Sigma_Z,

where the relevant measurability condition in order for this to define a Markov kernel follows by a standard application of the \pi-\lambda-theorem; associativity of composition uses Fubini’s theorem. In this way, we obtain a category \operatorname{\mathsf{Stoch}} which is, more or less by definition, isomorphic to the Kleisli category of the Giry monad [giry].

The symmetric monoidal structure on \operatorname{\mathsf{Stoch}} is defined in terms of the usual product of measurable spaces, given by the Cartesian product of underlying sets equipped with the product \sigma-algebra. The tensor product of morphisms f : (A,\Sigma_A) \to (X,\Sigma_X) and g : (B,\Sigma_B) \to (Y,\Sigma_Y) is given by the unique Markov kernel between the product measurable spaces which satisfies

The monoidal unit I is given by any one-element set with its unique \sigma-algebra. As a special case, morphisms I \to (X,\Sigma_X) can be identified with probability measures \Sigma_X \to [0,1]. sym-me-tric

There is a well-known subclass of particularly well-behaved measurable spaces, the standard Borel spaces, or equivalently Polish spaces equipped with their Borel \sigma-algebras. We write \operatorname{\mathsf{BorelStoch}} \subseteq\nobreak \operatorname{\mathsf{Stoch}} for the full subcategory of standard Borel spaces with Markov kernels. Since finite products of standard Borel spaces are again standard Borel, \operatorname{\mathsf{BorelStoch}} is again semicartesian symmetric monoidal. In summary, \operatorname{\mathsf{BorelStoch}} is the category of standard Borel spaces as objects with measurable Markov kernels as morphisms.

\operatorname{\mathsf{CRing}}_{+}^{^{\mathrm{op}}} is not causal. To see this, consider the ring \mathbb{Z}[t], where the following counterexample similar to [markov_cats, Example 11.33] can be formulated. Consider the additive unital maps given as the additive extensions of

On the basis monomials t^n, we have (fg)(t^n) = 1, and therefore

for all n,m \in \mathbb{N}, so that the hypothesis holds. If \operatorname{\mathsf{CRing}}_{+}^{^{\mathrm{op}}} were causal, then we would also have

for all n,m,\ell \in \mathbb{N}. But this clearly fails for n = \ell = 1 and m = 0.

The dual definition of infinite tensor products of algebraic structures as filtered colimits of finite tensor products is well-known in the literature, e.g. in the case of C*-algebras [blackadar, p. 315], although we do not know whether the preservation condition has been made explicit before. For example in the case where our category is \operatorname{\mathsf{CRing}}_+^{\mathrm{op}}, we recover the usual folklore definition of an infinite tensor product of rings \bigotimes_{i \in J} R_i in terms of formal sums of elementary tensors, where an elementary tensor is a family of elements (r_i)_{i \in J} such that all but finitely many are equal to the respective unit. The necessary preservation condition is easily seen to hold. Note that we do not yet consider the multiplication on \bigotimes_{i \in J} R_i, since Definition 3.1 is not yet concerned with the comonoid structures on the objects. We will get to this in the next section.

Intuitively, our definition matches up with the known algebraic ones under the categorical duality of algebra and geometry, where our Definition 3.1 is on the geometrical side of the duality.

We now consider infinite tensor products in \operatorname{\mathsf{Stoch}}, where we would like infinite tensor products to be given by the corresponding infinite products of measurable spaces in the usual sense. Since the Kolmogorov extension theorem does not hold for general (even merely countable) products of measurable spaces [AJ], this is not the case without further additional assumptions on the measurable spaces involved. Thus we do not know whether \operatorname{\mathsf{Stoch}} has infinite tensor products, although we suspect that it does not; but even in the unlikely case that it does, they are in general not the ones that one will care about for the purposes of probability theory.

However, the situation is much better for countable products in \operatorname{\mathsf{BorelStoch}}, for which the Kolmogorov extension theorem indeed holds [klenke, Theorem 14.35]. In other words, if \left( (X_i, \Sigma_i) \right)_{i \in \mathbb{N}} is a sequence of standard Borel spaces, then the cartesian product X_\mathbb{N} = \prod_{i \in \mathbb{N}} X_i, when equipped with the product \sigma-algebra \Sigma_\mathbb{N}, satisfies the universal property of an infinite tensor product with respect to maps out of I, since the marginalization maps implement a bijection between probability measures on X_{\mathbb{N}} = \prod_{i \in \mathbb{N}} X_i and compatible families of probability measures on the finite subproducts X_F = \prod_{i \in F} X_i, where the compatibility is with respect to marginalization to smaller subproducts specified by F' \subseteq F.

We now prove that this implies the universal property in general: if A and all the X_i are standard Borel spaces, then Markov kernels A \to X_J are in bijection with compatible families of Markov kernels A \to X_F. While we just saw that this is the case for A = I, we now show that it holds for arbitrary A. Thus suppose that

is a family of Markov kernels satisfying the compatibility condition \pi_{F',F} \circ g_{F'} = g_F for all finite F \subseteq F' \subseteq J. Then for every a \in A, the probability measures g_F(-|a) : \Sigma_F \to [0,1] are a compatible family to which the Kolmogorov extension theorem in the form of [klenke, Theorem 14.35] applies, and we obtain a unique probability measure g_J(-|a) : \Sigma_J \to [0,1] which has the original g_F(-|a) as its finite marginals. It remains to be shown that for every S \in \Sigma_J, the map a \mapsto g_J(S|a) is measurable. Since limits of pointwise convergent sequences of measurable real-valued functions are again measurable, the set of S for which this measurability holds is closed under countable disjoint union, and it is clearly closed under complements. We therefore have a \lambda-system. Since the map is measurable by assumption whenever S is a measurable cylinder set, and the cylinder sets form a \pi-system, the \pi-\lambda-theorem implies that a \mapsto g_J(S|a) is measurable for all S in the \sigma-algebra generated by the measurable cylinder sets, which is exactly the product \sigma-algebra \Sigma_J. Hence the existence part of the universal property indeed holds; and the uniqueness part is obvious by considering each a \in A separately and applying the Kolmogorov extension theorem.

Since the resulting comparison morphism Remark 3.3 is an isomorphism by construction—both sides are given by the corresponding products of measurable spaces—it follows that the required preservation condition of infinite tensor products holds as well.

In conclusion, \operatorname{\mathsf{BorelStoch}} indeed has countable tensor products. It is plausible that there is another subcategory of \operatorname{\mathsf{Stoch}} strictly larger than \operatorname{\mathsf{BorelStoch}} which has all infinite tensor products; for example, one can try to construct such a subcategory by imposing compactness or perfectness conditions [fremlin4, §451] on the Markov kernels, since under such assumption one again has a version of the Kolmogorov extension theorem [fremlin4, Corollary 454G]. However, we have so far not been able to find a definition of “compact Markov kernel” or “perfect Markov kernel” which would guarantee closure under composition, due to problems with perfect measures not being stable under mixtures [ramachandran]. Hence the problem of finding a subcategory of \operatorname{\mathsf{Stoch}} with all infinite tensor products remains open.

Infinite tensor products never exist in \operatorname{\mathsf{FinStoch}}, in the following sense: if the family (X_i)_{i \in J} is such that no X_i is empty and infinitely many of them contain at least two elements, then \bigotimes_i X_i does not exist. One way to see this is to use the fact that the hom-sets of \operatorname{\mathsf{FinStoch}} are convex subsets of the vector space of matrices, and that composition distributes over these convex combinations, making \operatorname{\mathsf{FinStoch}} into a category enriched in convex sets.compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a () command inside the argument of command inside the argument of command inside the argument of command inside the argument of command inside the argument of command inside the argument of command inside the argument of command inside the argument of See also golem.ph.utexas.edu/category/2010/09/what\_is\_this\_category\_enriched.html for discussion of this enrichment. Since these hom-sets are finite-dimensional, for every Y and Z there is n \in \mathbb{N} such that among every n morphisms Y \to Z, one of them can be written as a convex combination of the others. Now suppose that the product \bigotimes_i X_i existed. Then by choosing varying elements of each X_j, we could construct uncountably many morphisms 1 \to \bigotimes_i X_i whose marginalizations 1 \to X_F are all deterministic. Per the above, one of these hypothetical morphisms can be written as a convex combination of finitely many others. By choosing F \subseteq J suitably, we can achieve that the finite marginalizations 1 \to X_F of the finitely many morphisms involved in this convex combination are all distinct. Since these morphisms 1 \to X_F are deterministic by construction, and no deterministic morphism in \operatorname{\mathsf{FinStoch}} can be written as a convex combination of other deterministic morphisms, we have arrived at a contradiction. Thus \bigotimes_i X_i does not exist.

Suppose that \mathsf{C} is a partially ordered set, considered as a category with X \geq Y if and only if the morphism X \to Y exists.compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a compositionalityarticleYou must not put a () command inside the argument of command inside the argument of command inside the argument of command inside the argument of command inside the argument of command inside the argument of command inside the argument of command inside the argument of This direction of the ordering is opposite to the usual convention, but necessary for the following to work without reversing direction. Then the monoidal structure makes \mathsf{C} into a preordered commutative monoid, which we write additively. Here, the functoriality of the monoidal structure amounts to the assumption that addition is monotone,

Furthermore, the monoidal structure is semicartesian if and only if X \geq 0 for every X. One can then apply Lemma 3.10 to show that \mathsf{C} has infinite tensor products if and only if \mathsf{C} is a finitary complete monoid in the sense of Goldstern and Karner [karner, Section 2.2], and these infinite tensor products then correspond to infinite sums \sum_{i \in J} X_i in the monoid.

Let \operatorname{\mathsf{SetMulti}} be the category with objects ordinary sets, morphisms f: X \to Y given by multivalued functions, which are maps f: X \to \mathcal{P}(Y)\setminus \emptyset, with composition of morphisms f : X \to Y and g : Y \to Z given by

\operatorname{\mathsf{SetMulti}} could also be described as the Kleisli category of the nonempty powerset monad, which sends a set X to \mathcal{P}(X) \setminus \emptyset. In particular, \operatorname{\mathsf{FinSetMulti}} is closely related to \operatorname{\mathsf{Rel}}, the category of sets and relations, since the latter can equivalently be regarded as the Kleisli category of the powerset monad (where the empty set is allowed). Intuitively, \operatorname{\mathsf{SetMulti}} is the closest relative of \operatorname{\mathsf{Rel}} to which our formalism can be applied. See also [markov_cats, Example 2.6] for a treatment of the full subcategory \operatorname{\mathsf{FinSetMulti}} on finite sets.

The Cartesian product of sets equips \operatorname{\mathsf{SetMulti}} with a symmetric monoidal structure, if for f : A \to X and g: B \to Y we set

The monoidal unit is any one-element set I. With the obvious structure isomorphisms inherited from \mathsf{Set}, this turns \operatorname{\mathsf{SetMulti}} into a semicartesian symmetric monoidal category, and we can give it a Markov category structure by setting \mathsf{copy}_{X}(x) := \{(x,x)\} \subset X \times X. Then the deterministic morphisms are precisely those sending each point into a singleton set, resulting in an isomorphism of categories \operatorname{\mathsf{SetMulti}}_\mathrm{det} \cong \mathsf{Set}.

Of course, \mathsf{Set} has all infinite products. But perhaps surprisingly, these products are not Kolmogorov products, or even infinite tensor products in \operatorname{\mathsf{SetMulti}}. To see this, consider for example the family (X_n)_{n\in\mathbb{N}} with X_n = \{0,1\} for all n. A map I \to \prod_{n\in \mathbb{N}} \{0,1\} is just a nonempty subset of this infinite product. Given such a subset U, the finite marginals I \to \prod_{n \in F} \{0,1\} for finite F \subset \mathbb{N} are the images of U under the projection map. But these images do not determine U. For instance, consider the subset U of those sequences containing at least one 1. Its image in every \prod_{n \in F} \{0,1\} is the whole set, since any finite sequence of 0s and 1s can be extended to one containing a 1, but U itself is a proper subset. In particular, U has the same compatible family of finite marginalizations as e.g. the set of those sequences containing at least one 0, and hence the uniqueness part of the universal property fails.

It is clear that this counterexample is quite generic and works for any infinite family of nonempty sets, infinitely many of which contain at least two elements, as in Example 3.7. Thus \operatorname{\mathsf{SetMulti}} is very far from having Kolmogorov products. A good way to remedy this deficiency is to consider topological spaces instead of sets: as we will see in ?, the Kleisli category of the lower Vietoris monad on \operatorname{\mathsf{Top}} does have Kolmogorov products, and it can be thought of as the topological counterpart of \operatorname{\mathsf{SetMulti}}.

Let A = \{*\} and X = Y = Z = [0,1] with the usual topology. Let f: A \to HX send * to X, and define

Using the definition of the hyperspace topology, it is straightforward to see that these maps are continuous. To check the hypothesis of the statement of causality, we must verify that the resulting two maps A \to Y \times Z agree; these are just closed subsets of [0,1]^2. From the definition of the composition, we see that any point (y,z) with y \in g(x) for some x and z \in h_1(y) must lie in the set corresponding to h_1, and similarly for h_2. Hence the set corresponding to h_1 contains all points, while the set corresponding to h_2 contains as a subset all points (y,z) with y\neq 1/2, as well as (1/2,0). But the closure of \{(y,z) \mid y \neq 1/2\} in [0,1]^2 is the whole space, so that both sets are indeed equal to the whole space.

Hence the hypothesis is satisfied. But one can verify that the point (x,y,z) = (1/2,1/2,1) is in the subset of X \times Y \times Z corresponding to the h_1 side of (2.7), but not so for the h_2 side.