nLab measurable locale

Measurable locales

Measurable locales

Idea

Measurable locales are certain locales, which may serve as a basis for measure theory much as general locales serve as a basis for topology (especially in constructive mathematics). Specifically, a measurable locale is equivalent to a compact strictly localizable measurable space (or dual to a commutative von Neumann algebra).

Ironically, in constructive mathematics, measurable locales are not locales (except for the empty space), on pain of excluded middle.

The concept is due to Dmitri Pavlov.

Definitions

At present, there is no purely order-theoretic definition of measurable locales. However, there are a few other ways of defining them.

In classical mathematics

First we give definitions appropriate for classical mathematics.

As a preliminary step, consider the category CompBoolAlg supComp Bool Alg_{sup} of complete boolean algebras and supremum-preserving homomorphisms of boolean algebras. This is a full subcategory of the category FrmFrm of frames and so its opposite category CompBoolAlg sup opComp Bool Alg_{sup}^{op} is a full subcategory of the category Loc of locales. The category MeasLocMeas Loc of measurable locales is yet further a full subcategory of CompBoolAlg sup opComp Bool Alg_{sup}^{op}, so our job is simply to specify which complete boolean algebras are the objects of this category. There are at least three equivalent ways of doing so.

  1. By one definition, a complete boolean algebra LL is a measurable locale if there is a (complete) compact strictly localizable measure space XX such that LL is isomorphic (as a boolean algebra) to the boolean algebra /𝒩\mathcal{M}/\mathcal{N} of measurable subsets of XX modulo null subsets. Note that the measure on XX is irrelevant except to specify the null subsets. In this way, MeasLocMeas Loc becomes equivalent to the category CSLEMSCSLEMS of (complete) compact strictly localizable enhanced measurable spaces. (Not every measure space has a complete boolean algebra as /𝒩\mathcal{M}/\mathcal{N}; those which do are called localizable, and an enhanced measurable space is a structure with both \mathcal{M} and 𝒩\mathcal{N} specified.)

  2. By another definition, a complete boolean algebra LL is a measurable locale if for every element x0x \neq 0 of LL there is a continuous valuation (see below) μ\mu on LL valued in [0,][0,\infty] such that μ(x)=1\mu(x) = 1. (It would be enough to demand that 0<μ(x)<0 \lt \mu(x) \lt \infty, since we can rescale the valuation.)

  3. By yet another definition, a complete boolean algebra LL is a measurable locale if there is a commutative W *W^*-algebra AA such that LL is isomorphic (as a boolean algebra) to the boolean algebra Proj(A)Proj(A) of projection operators in AA; (see idempotent operator for a construction of Proj(A)Proj(A)). In this way, MeasLocMeas Loc becomes dual to the category CommW *AlgComm W^* Alg of commutative W *W^*-algebras. For this definition, we need not require that LProj(A)L \cong Proj(A) be complete (or even a boolean algebra); this can be proved.

Whichever of these equivalent definitions is adopted, a measurable function between measurable locales is simply a continuous map between them as locales; these are the morphisms of MeasLocMeas Loc. In more elementary terms, a measurable function from LL to MM (thought of as measurable locales) is a function from MM to LL (thought of as Boolean algebras) that preserves all joins (and then it also preserves finite meets and so is a frame homomorphism). Or we can think of this function’s right adjoint, a function from LL to MM that preserves all meets.

In constructive mathematics

(This section is not due to Pavlov.)

For purposes of constructive mathematics, Toby Bartels suspects that it is appropriate to use the definition from W W^\star-algebras, so long as we allow norms to be lower real numbers. (If they are all located, then Proj(A)Proj(A) has decidable equality, which we don't want to require.) The other definitions seem more difficult to work with (primarily because they require completeness explicitly).

For this definition (as was remarked above), we need not require that LProj(A)L \cong Proj(A) be complete. Constructively, we cannot require this; Proj(A)Proj(A) need not be complete (although it is still a boolean algebra). Indeed, consider the point (see the examples), based on AA \coloneqq \mathbb{C}, which is not the complete Heyting algebra of all truth values but only the (possibly incomplete) boolean algebra {,}\{\bot, \top\} of classical truth values (corresponding to the self-adjoint idempotent complex numbers 00 and 11).

It seems that there is some notion of completeness that applies here; LL should in some sense be ‘measurably complete’. In the point, for example, the subsingleton {*|P}\{* \;|\; P\}, where PP is a truth value, is measurable iff PP is true or false, so the supremum of {|P}\{\top \;|\; P\} exists in LL under the same circumstances. Figuring this out could allow the definitions through measurable spaces or continuous valuations to work.

Simpson’s theory of sigma-locales, see below, is developed in classical framework, but could be a good starting point for a constructive theory, as it avoids the focus on complements. (But the main problem is completeness, not complements.)

Examples

The empty space, which is initial in MeasLocMeas Loc, is the terminal boolean algebra with one element.

The point, which is terminal in MeasLocMeas Loc, is the initial boolean algebra {,}\{\bot, \top\} of (classical) truth values.

The real line is the boolean algebra of Lebesgue-measurable sets of real numbers modulo the null sets. This is complete as a boolean algebra because … [argument needed] (and even constructively, it is still a boolean algebra).

Applying the classification of W *W^*-algebras, we find that (up to isomorphism), every measurable locale is a disjoint union (dually a direct product of boolean algebras) of {,} I\{\bot, \top\}^I for a set II of arbitrary cardinality. If II is finite, we get discrete spaces; if II is countable, we get the real line. Note that a single real line is already isomorphic to the union of countably infinitely many real lines. (Of course, we can't expect this classification theorem? to hold constructively.)

Borel measurable spaces

To do measure theory, it's important to know how to interpret the real line with the Borel sets (rather than Lebesgue sets) as a measurable space. This is actually a little tricky, because Lebesgue measure is not complete on the Borel sets, and passing to the Lebesgue sets gives a different measurable space. We might simply take the Borel sets as they are (so that only the empty set is null), but then this is not complete as a boolean algebra.

One might suspect that there is no Borel real line in MeasLocMeas Loc, which would cast serious doubts on that as a category for measure theory. But we can argue by abstract nonsense that it must exist, and more generally that any locale (including the locale of opens of any topological space) has a Borel measurable locale, as follows:

http://mathoverflow.net/questions/31603/why-do-probabilists-take-random-variables-to-be-borel-and-not-lebesgue-measurab/31724#31724representable functor theorem

In the case of the real line, the resulting measurable locale is actually rather complicated: something like an uncountable disjoint union of points and lines.

This process is a functor B:LocMeasLocB\colon Loc \to Meas Loc, which we can also extend to TopLocMeasLocTop \to Loc \to Meas Loc.

Lebesgue measurable spaces

Similarly, smooth manifolds give rise to (much simpler!) measurable locales by taking the Lebesgue sets (possible since we know which sets are null sets on a smooth manifold). This process is a functor L:SmoothMan submMeasLocL\colon SmoothMan_subm \to MeasLoc, where the morphisms in SmoothMan submSmoothMan_subm are the smooth submersions.

In the case of the real line, we get the usual line described under the basic examples above. In fact, any second-countable smooth manifold with positive dimension gives rise to this measurable locale; the really interesting thing about LL is what it does to the morphisms.

Measures

On any complete boolean algebra LL, given any abelian monoid RR equipped with a convergence structure (such as [0,][0,\infty]), a normal measure on LL valued in RR is a function μ:LR\mu\colon L \to R such that:

One could, of course, define a garden-variety measure by requiring continuity only when SS is the image of a decreasing sequence, but apparently normal measures are what we want for measurable locales. Normal measures are also known as continuous valuations.

Absolutely continuous measures on, say, Lebesgue space correspond to normal measures on the real line as a measurable locale. (Lebesgue measure is not normal on the σ\sigma-algebra of measurable sets, but it is normal on the boolean algebra of measurable sets modulo null sets, which is what we want.)

This should work in constructive mathematics, with [0,][0,\infty] (for a positive measure) being be the space of nonnegative lower reals.

This is related to σ\sigma-locales.

  • Alex Simpson, Measure, randomness and sublocales link

which is part of his work to develop synthetic probability theory.

References

The basic definition and some elementary properties are given in

This paper also establishes Gelfand-type duality for commutative von Neumann algebras?.

The theory of measurable locales originated on MathOverflow, in various questions and answers by Dmitri Pavlov. Here is an index:

  • Dmitri Pavlov, answer to Is there a category structure one can place on measure spaces so that category-theoretic products exist? (version: 2011-01-08)

Last revised on December 3, 2022 at 00:16:09. See the history of this page for a list of all contributions to it.