More laws

More laws

[Tom Leinster]

Thanks very much to the many people who provided helpful, expert replies.

To recap, I asked (among other things): given an algebra A for some
theory, what can be said about the algebras obeying all the equational
laws that A obeys?

The question was posed in the context of finitary algebraic theories, and
in that context, there's a straightforward description of such algebras:
they are exactly the quotients of subalgebras of (possibly infinite)
powers of A.  As explained by Peter Selinger and George Janelidze, this
comes from Birkhoff's Theorem.

I wanted to understand this particular situation - finitary algebraic
theories and their equational laws - and I do understand it better than I
did before.  However, what I ultimately want to understand is something
slightly different and more general, which I'll now describe.

This will take a while, so I'll start with the punchline: we get some very
simple universal characterizations of some quite sophisticated objects.
For example, we'll get a characterization of the Stone spaces among all
topological spaces, and a construction of the space of Borel probability
measures on a compact space.

Here goes.  Take an algebraic theory and write F for the free algebra
functor.  Any equational law w = w' in a set X of variables generates a
congruence ~ on FX (identify w and w'), hence a quotient map

e: FX ---> (FX)/~.

An algebra A obeys this law iff e is orthogonal to A, i.e. every map FX
---> A factors uniquely through e.

All such maps e encoding laws are regular epi (and more besides).
However, let's consider *all* maps whose domain is a free object.  I also
want to consider all adjunctions, not just those arising from an algebraic
theory - although they will provide important examples.

So the general set-up is this.  Fix an adjunction U: D ---> C, F: C ---> D
(F left adjoint to U).  A *law* is an arrow e: FX ---> E where X is in C
and E is in D.  An object A of D *obeys* the law e if e is orthogonal to
A.  An object B of D is *A-complete* if B obeys every law that A obeys.
The *A-completion* B[A] of B is the free A-complete object on B, assuming
it exists.

Note that "law" now means something more general than it did before; we're
no longer just interested in *equational* laws.  For instance, if the
adjunction is the usual one between D = Group and C = Set,
torsion-freeness is defined by laws, whereas it's not defined by
equational laws (is it?).  Indeed, a group is torsion-free iff it obeys
the law 1 ---> Z/pZ for every prime p.

Examples (without proofs):

1. Take C = D = Set and the identity adjunction.  Write n for an n-element
set.  The 0-complete objects are 0 and 1.  The only 1-complete object is
1.  If |A| > 1 then every set is A-complete, since A obeys "no laws" (i.e.
only those laws that are isomorphisms).

(If we were just using equational laws then 0 would be 1-complete too.)

2. Take the usual adjunction between D = abelian groups and C = sets.  Let
k be either the field of rational numbers or the field of p elements for
some prime p; these fields have the property that an abelian group can be
a k-vector space in at most one way.  Then an abelian group is k-complete
iff it is a k-vector space.

3. Take D to be a v-semilattice and C to be the terminal category.  Let a
be in D.  Then an element b of D is a-complete iff b >= a, so b[a] = a v
b.
This might suggest that in general, B[A] should be functorial in A (as
well as in B), but it's not.  You can use example (1) to show this.

4. Take the discrete-space/point-set adjunction between D = Top and C =
Set.  Write 2 for the discrete 2-point space.  Then a space is 2-complete
iff it is a Stone space, i.e. compact Hausdorff totally disconnected.  So
if B is any space, B[2] is the reflection of B into Stone spaces.

It's striking that the adjunction between Top and Set, together with the
simple space 2, give rise to the notion of Stone space.  Poetically, "the
theory of the 2-point space is the theory of Stone spaces".

All of this remains true if 2 is replaced by any other Stone space with >
1 element.  (But it's most dramatic if you use 2.)

5. This is the most substantial example, and it's what started me off on
all this.  It's from work of Matthias Schroeder and Alex Simpson: see
Simpson's talk "Probabilistic Observations and Valuations" at

http://homepages.inf.ed.ac.uk/als/Talks/

Let D be the category of topological spaces equipped with a binary
operation and C = Top, with the obvious adjunction between them.  Let I =
[0, 1] with topology generated by all the subintervals (x, 1], and the
midpoint (mean) operation.  Schroeder and Simpson prove, among other
things, that if X is a compact Hausdorff space then (FX)[I] is the space
of regular Borel probability measures on X, equipped with the weak
topology (and the obvious midpoint operation).  Using the universal
property of (FX)[I], this also gives the definition of integration on X.


I'd be interested to know what other people make of this.  Perhaps, for
instance, someone knows a good way to describe which properties can be
defined by "laws" in the sense above, and perhaps someone can shed some
light on the non-functoriality noted in (3).

Thanks for reading,
Tom