re: Limits

[Martin Escardo <m.escardo@cs.bham.ac.uk>]

Tobias Schroeder writes:
 > - Can the limit of a sequence of real numbers be expressed
 >   as a categorical limit (of course it can if the sequence is
 >   monotone, but what if it is not)?

I think I have an answer to this question (without cheating). It may
be well known or wrong (I haven't carefully checked the details, but I
believe that they are correct).

Given a metric space X with distance function d, construct a category,
also called X, as follows. The objects of the category X are the
points of the space X. An element of the hom-set X(x,y) is a triple
(r,x,y) with r a real number such that d(x,y)<=r. The composite of the
arrows r:x->y and s:y->z is the arrow s+r:x->z. This is well defined
by virtue of the triangle inequality d(x,z)<=d(x,y)+d(y,z). By virtue
of the condition d(x,x)=0, we have identities. Notice that all arrows
are mono.

Of course, because the category X is small and it is not a preordered
set, it doesn't have all limits. But some limits do exist.

CLAIM: Let x_n be a sequence of points of X, and, for each n, let the
arrow r_n:x_{n+1}->x_n be d(x_n,x_{n+1}). If the sum of r_k over k>=0
exists, then this omega^op-diagram has a categorical limit. The source
of the limiting cone is the metric limit l of the sequence. The
projection p_n:l->x_n is the sum of r_k over k>=n. If q_n:m->x_n is
another cone, then the mediating map u:l->m exists (and will be
automatically unique), because, by definition of cone and of our
category, q_n will have to be bigger than r_n, and then u=q_n-r_n does
the job.

Remarks. (1) For any given Cauchy sequence, one can construct an
equivalent Cauchy sequence for which the assumption in the second
sentence of the claim fails. Using classical logic, for any given
Cauchy sequence, one can construct an equivalent Cauchy sequence for
which the assumption holds.  

(2) In (some flavours of) constructive mathematics, the notion of a
Cauchy sequence "with fixed rate of convergence" is taken as
basic. This often is taken to mean that d(x_n,x_n+1)<=c^n for a fixed
c with 0<c<1. For such sequences, the assumption is satisfied. Recall
that a map f:X->Y is called non-expansive if d(fx,fx')<=d(x,x'). If
the natural numbers are metrized by d(m,n)=c^min(m,n) for m/=n, to get
a space N, then such a Cauchy sequence is just a non-expansive map
N->X. It converges if and only if the non-expansive map has a
non-expansive extension to N_{infty}, the metric completion of N
(which, topologically, is the one-point compactification of N). And
non-expansive maps are functors---see (3) below.

(3) Recall that a map f:X->Y is called lipschitz if there is a
constant c for which d(fx,fx')<=c.d(x,x').  A lipschitz map f:X->Y
gives rise to a functor f:X->Y defined by
f(r:x->x')=c.r:f(r)->f(x'). That is, the object part is given by the
map itself, and the arrow part is given by multiplication with the
lipschitz coefficient.

(4) We have taken the arrows r_n to be d(x_n,x_{n+1}). But actually
any choice of arrows does the job, provided the sum of r_k over k>=0
is finite.

(5) Other two conditions for the distance function of a metric space,
which were not used in the definition of the category X, are (i)
d(x,y)=0 implies x=y, and (ii) d(x,y)=d(y,x). By the first, our
category is skeletal. By the second, it is selfdual. Of course, people
have considered generalized metric spaces in which these are not
assumed to hold. See, for example, Lawvere's paper "Metric spaces,
generalized logic, and closed categories", in which he regards a
generalized metric space X as an enriched category with X(x,y)=d(x,y)
(so he has hom-numbers instead of hom-sets). Here we have hom-sets (of
numbers).

MHE