\documentclass[leqno]{tac}
\usepackage{amsmath,amsfonts}
\author{George M. Bergman}
\thanks{%
Much of this work was done in 1986, when the author was
partially supported by NSF grant DMS~85-02330.
\protect\\\indent\indent %why are two \indent's needed?
arXiv:0711.0674.
}
\address{%
Department of Mathematics,\\
University of California\\
Berkeley, CA 94720-3840, USA}
\title{Colimits of representable algebra-valued functors}
\copyrightyear{2008}
\keywords{representable functor among varieties of algebras,
initial representable functor, colimit of representable functors,
final coalgebra, limit of coalgebras;
binar (set with one binary operation), semigroup, monoid, group,
ring, Boolean ring, Stone topological algebra}
\amsclass{Primary: 18A30, 18D35.
% (co)lims gp_objs&c
Secondary: 06E15, 08C05, 18C05, 20M50, 20N02
% Stone_sp cats_of_algs eq_cats semi&cats binar
}
\eaddress{gbergman@math.berkeley.edu}
% author macros BEGIN here
\hyphenation{
co-algeb-ra
co-algeb-ras
co-lim-it
co-lim-its
pre-coalgeb-ra
pre-coalgeb-ras
pseudo-co-algeb-ra
pseudo-co-algeb-ras
pseudo-co-pro-jection
pseudo-co-pro-jections
quasi-variety
sub-coalgeb-ra
zero-ary
}
\raggedbottom
\setlength{\mathsurround}{.167em}
\newcommand{\<}{\kern.0833em}
\newtheorem{question}[theorem]{Question}
\newcommand{\fb}{\mathbf}
\newcommand{\Rep}[2]{\fb{Rep}(#1,\protect\nolinebreak[2]#2)}
\newcommand{\coalg}[2]{\fb{Coalg}(#1,\protect\nolinebreak[2]#2)}
\newcommand{\pscoalg}[2]{\fb{Pseudocoalg}(#1,\protect\nolinebreak[2]#2)}
% next: instead of \mathbfdef{C}, to make $\!\C\!$-${\fb{Alg}}$$ look OK
\newcommand{\C}{\fb{C}\kern.05em}
\mathbfdef{D}
\mathbfdef[Bi]{Binar}
\mathbfdef[Se]{Semigp}
\mathbfdef[Gp]{Group}
\mathrmdef{ari}
\mathrmdef[cd]{card}
\newcommand{\cP}{\raisebox{.1em}{$\<\scriptscriptstyle\coprod$}\nolinebreak[2]}
\newcommand{\limit}{\varprojlim}
\newcommand{\colim}{\varinjlim}
\newcommand{\ba}{_\mathrm{base}}
\newcommand{\circsm}{\kern.25em{\scriptstyle\circ}\kern.25em\nolinebreak[3]}
% author macros END here
\begin{document}
\maketitle
\begin{abstract}
If $\C$ and $\D$ are varieties of algebras in the sense
of general algebra, then by a representable functor
$\C\to\D$ we understand a functor which, when
composed with the forgetful functor $\D\to\fb{Set},$ gives a
representable functor in the classical sense; Freyd showed that these
functors are determined by $\!\D\!$-coalgebra objects of $\C.$
Let $\Rep{\C}{\D}$ denote the category of all
such functors, a full subcategory of $\fb{Cat}(\C,\D),$
opposite to the category of $\!\D\!$-coalgebras in $\C.$
It is proved that $\Rep{\C}{\D}$ has small
colimits, and in certain situations, explicit constructions for
the representing coalgebras are obtained.
In particular, $\Rep{\C}{\D}$ always has an initial object.
This is shown to be ``trivial'' unless
$\C$ and $\D$ either both have {\em no} zeroary operations, or both
have {\em more than one} derived zeroary operation.
In those two cases, the functors in question may have
surprisingly opulent structures.
It is also shown that every set-valued representable functor
on $\C$ admits a universal morphism to a $\!\D\!$-valued
representable functor.
Several examples are worked out in detail, and areas for further
investigation are noted.
\end{abstract}
% - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
\section*{}
In \S\S\ref{S.intro}-\ref{S.preco>co} below we develop our general
results, and in \S\S\ref{S.BiBi}-\ref{S.Ring>Ab}, some examples.
(One example is also worked in \S\ref{S.Set}, to motivate
the ideas of \S\ref{S.precoalg}.)
R.~Par\'e has pointed out to me that my main result,
Theorem~\ref{T.final}, can be deduced from
\cite[Theorem~6.1.4, p.143, and following remark, and
{\em ibid.}\ Corollary~6.2.5, p.149]{M+P}.
However, as he observes, it is useful to have a direct proof.
% \pagebreak
\vspace{1em}
{\samepage\begin{center}{\bf I. GENERAL RESULTS.}\end{center}
\section{Conventions; algebras, coalgebras, and representable
functors.}\label{S.intro}
In} this note, morphisms in a category will be composed like
set-maps written to the left of their arguments.
The composition-symbol $\circsm$ will sometimes be introduced as an
aid to the eye, with no change in meaning.
So given morphisms $f:X\to Y$ and $g:Y\to Z,$ their composite
is $g\co}, where we develop methods
for the explicit construction of our colimit functors.)
If $\fb{A}$ is a category in which all families of
$<\lambda_\D$ objects have coproducts, an
{\em $\!\Omega_\D\!$-coalgebra} $R$ in $\fb{A}$ will mean
a pair $R=(|R\<|,(\alpha^R)_{\alpha\in\Omega_\C}),$ where $|R\<|$ is
an object of $\fb{A},$ and each $\alpha^R$ is a morphism
$|R\<|\to\coprod_{\ari(\alpha)}|R\<|;$
the $\alpha^R$ are the {\em co-operations} of the coalgebra.
For each $\alpha\in\Omega_\D$ and each object $A$ of $\fb{A},$
application of the hom-functor $\fb{A}(-,A)$ to the co-operation
$\alpha^R$ induces an operation
$\alpha_{\fb{A}(R,A)}: \fb{A}(|R\<|,A)^{\ari(\alpha)}\to
\fb{A}(|R\<|,A);$ these together
make $\fb{A}(|R\<|,A)$ an $\!\Omega_\D\!$-algebra,
which we will denote $\fb{A}(R,A).$
We thus get a functor
$\fb{A}(R,-):\fb{A}\to\mbox{$\Omega_\D\!$-$\!\fb{Alg}.$}$
Note that this is a nontrivial extension of standard notation:
If $A$ and $B$
are objects of $\fb{A},$ then $\fb{A}(A,B)$ denotes the hom-{\em set}\/;
if $R$ is an $\!\Omega_\D\!$-coalgebra in $\fb{A},$
then $\fb{A}(R,B)$ denotes an {\em $\!\Omega_\D\!$-algebra} with the
hom-set $\fb{A}(|R\<|,B)$ as underlying set.
We are also making the symbol $|\ |$ do double duty:
If $A$ is an object of the variety
$\C,$ then $|A\<|$ denotes its underlying set; if
$R$ is a coalgebra object in $\C,$ we denote by $|R\<|$
its underlying $\!\C\!$-algebra; thus
$||R\<||$ will be the underlying set of that underlying algebra.
In this situation, ``an element of $R\!$'',
``an element of $|R\<|\!$'' and ``an element of $||R\<||\!$'' will all
mean the same thing, the first two being shorthand for the third.
In symbols we will always write this $r\in||R\<||.$
We also extend the standard language under which the functor
$\fb{A}(|R\<|,-): \fb{A}\to\fb{Set}$ is said to
be {\em representable} with representing object $|R\<|,$ and call
$\fb{A}(R,-): \fb{A}\to\Omega_\D$\!-\!$\fb{Alg} $ a representable
(algebra-valued) functor, with representing coalgebra $R.$
A more minor notational point: Whenever we write something
like $\coprod_\kappa A,$ this will denote a $\!\kappa\!$-fold copower
of the object $A,$ with the index which ranges over $\kappa$ unnamed;
thus, a symbol such as $\coprod_\kappa A_\iota$ denotes the
$\!\kappa\!$-fold copower of a single object $A_\iota.$
On the few occasions (all in \S\ref{S.preco>co}) where we consider
coproducts other than copowers, we will show the variable index
explicitly in the subscript on the coproduct-symbol,
writing, for instance, $\coprod_{\iota\in\kappa}A_\iota$ to denote
(in contrast to the above)
a coproduct of a family of objects $A_\iota$ $(\iota\in\kappa).$
If $R=(|R\<|,(\alpha^R)_{\Omega_\C})$ is
an $\!\Omega_\D\!$-coalgebra in $\fb{A},$
and $s,t\in|T_{\Omega_\D}(\kappa)|,$ then the necessary
and sufficient condition for the algebras $\fb{A}(R,A)$ to
satisfy the identity $s=t$ for all objects $A$ of $\fb{A}$ is that the
$\!\kappa\!$-tuple of coprojection maps in the $\!\Omega_\D\!$-algebra
$\fb{A}(R,\,\coprod_{\kappa}|R\<|)$ satisfy that relation.
In this case, we shall say that the
$\!\Omega_\D\!$-coalgebra $R$ {\em cosatisfies} the identity $s=t.$
If $R$ cosatisfies all the identities in $\Phi_\D,$
we shall call $R$ a {\em $\!\D\!$-coalgebra} object of $\fb{A},$
and call the functor it represents
a representable $\!\D\!$-valued functor on $\fb{A}.$
Whenever a functor $F:\fb{A}\to\D$ has the property
that its composite with the underlying set functor $U_\D$
is representable in the classical sense, then the $\!\D\!$-algebra
structures on the values of $F$ in fact arise
in this way from a $\!\D\!$-coalgebra
structure on the representing $\!\fb{A}\!$-object.
If $\fb{A}$ not only has small coproducts but general small colimits,
the $\!\D\!$-valued functors that are representable
in this sense are precisely those that have left adjoints
\cite{Freyd}, \cite[Theorem~9.3.6]{245}, \cite[Theorem~8.14]{coalg}.
We shall write $\coalg{\fb{A}}{\D}$ for the category of
$\!\D\!$-coalgebras in $\fb{A},$ taking for the morphisms
$R\to S$ those morphisms $|R\<|\to|S|$ that make
commuting squares with the co-operations, and we shall
write $\Rep{\fb{A}}{\D}$ for the category of representable functors
$\fb{A}\to\D,$ a full subcategory of $\fb{Cat}(\fb{A},\D).$
Up to equivalence, $\Rep{\fb{A}}{\D}$ and
$\coalg{\fb{A}}{\D}$ are opposite categories.
The seminal paper on these concepts is \cite{Freyd} (though the
case where $\C$ is a variety of commutative rings
and $\D$ the variety of groups, or, occasionally, rings,
was already familiar to algebraic geometers, the representable
functors corresponding to the ``affine algebraic groups and rings'').
For a more recent exposition, see \cite[\S\S9.1-9.4]{245}.
In \cite{coalg} the structure of $\coalg{\C}{\D}$ is determined
for many particular varieties $\C$ and $\D.$
(The term ``coalgebra'' is sometimes used for a more elementary
concept: Given any functor $F: \fb{Set}\to\fb{Set},$
a set $A$ given with a morphism $A\to F(A)$ is called,
in that usage, an $\!F\!$-coalgebra.
The existence of final coalgebras in that sense has also been
studied \cite{PA+NM}, \cite{Barr}, and, as we shall see,
has a slight overlap with the concept studied in this note.
Still another use of ``coalgebra'', probably the earliest,
which also has relations to these two, and is basic to the theory of
Hopf algebras, is that of a module $M$ over a commutative ring,
given with a map $M\to M\otimes M;$ cf.\ \cite[pp.4 to end]{MS} and
\cite[\S\S29-32, \S43]{coalg}.
And in fact, as the referee has pointed out, results analogous to some
of those in this note were proved for coalgebras in that sense over
thirty years ago, by a similar approach~\cite{Barr_74}.)
The first step toward our results will be an easy observation.
\begin{lemma}\label{L.colim}
If $\fb{A}$ has small colimits, then so does $\coalg{\fb{A}}{\D};$
equivalently, $\Rep{\fb{A}}{\D}$ has small limits.
Moreover, the underlying $\!\fb{A}\!$-object of a colimit of
$\!\D\!$-coalgebras in $\fb{A}$ is the colimit of the underlying
$\!\fb{A}\!$-objects of these algebras.
Equivalently, the composite
of the limit of the corresponding representable functors with the
underlying set functor $U_\D$ is the limit of the
composites of the given functors with $U_\D,$ and may thus be evaluated
at the set level by taking limits of underlying sets.
\end{lemma}
\begin{proof}
Since $\D$ has small limits, the category of $\!\D\!$-valued functors
on any category also has small limits, which may be computed
object-wise; hence by properties of limits of
algebras, these commute with passing to underlying sets.
In particular, a functor $F$ from a small category $\fb{E}$ to
$\Rep{\fb{A}}{\D}$ will have a limit $L$ in $\D^\fb{A},$
and $U_\D\circsm L$ will
be the limit of the set-valued functors $U_\D\circsm F(E).$
The latter limit is represented by the colimit of the underlying
$\!\fb{A}\!$-objects of the coalgebras representing the $F(E);$
call this colimit object $|R\<|.$
As noted above, representability of $U_\D\circsm L: \fb{A}\to\fb{Set}$
by $|R\<|$ implies representability of $L:\fb{A}\to\D$
by a coalgebra $R$ with underlying $\!\fb{A}\!$-object $|R\<|.$
(One can get the same result starting at the coalgebra end,
using the general fact that ``colimits commute with colimits'' to
deduce that a colimit of underlying objects of a diagram of coalgebras
inherits co-operations from these, and verify that the resulting
coalgebra has the universal property of the desired colimit.)
\end{proof}
Knowing that $\Rep{\fb{A}}{\D}$ has small limits, to prove that it has
small {\em colimits} we need ``only'' prove that it satisfies the
appropriate solution-set condition (\cite[Theorem~V.6.1]{CW},
\cite[Theorem~7.10.1]{245}).
Easier said than done!
We shall get such a result in the case where $\fb{A}$ is a variety $\C.$
In the next section, we motivate the technique to be used (after
giving a couple of results showing cases to avoid when
thinking about examples), then preview the remainder of the paper.
\section{Background; trivial cases; motivation of the
proof; overview of the paper.}\label{S.motivate}
Let me first describe what led me to the questions answered below.
Let $\fb{Ring}^1$ denote the category of associative unital rings,
and for any field $K,$ let $\fb{Ring}^1_K$ denote the category of
associative unital $\!K\!$-algebras.
(When I write ``$\!K\!$-algebra'', ``algebra'' will be meant in
the ring-theoretic sense; otherwise it is
always meant in the general sense.)
In \cite[\S25]{coalg}, descriptions are obtained of all
representable functors from $\fb{Ring}^1_K$ into a number of
varieties of algebras (general sense!), including $\fb{Ring}^1.$
The result for the lastmentioned variety says
that for every representable functor
$F: \fb{Ring}^1_K\to\fb{Ring}^1,$ there exist two
{\em linearly compact} associative unital $\!K\!$-algebras $A$
and $B,$ such that $F$ is isomorphic to the functor
taking every object $S$ of $\fb{Ring}^1_K$ to the direct product
of completed tensor-product rings
$S\hat{\otimes}A\,\times\,S^{\mathrm{op}}\hat{\otimes}B.$
(One doesn't need to know precisely what these terms mean
to appreciate the point that is coming up.
For general background:
the category of linearly compact $\!K\!$-vector spaces is dual to the
category of all $\!K\!$-vector spaces, and a
linearly compact topology on an associative unital $\!K\!$-algebra
$A$ makes it an inverse limit of finite-dimensional $\!K\!$-algebras
\cite[\S24]{coalg}.
In this situation, the {\em completed} tensor product of a
$\!K\!$-algebra $S$ with $A$ is the inverse limit of the
tensor products of $S$ with those finite-dimensional $\!K\!$-algebras.)
Now the category of linearly compact associative unital $\!K\!$-algebras
has an initial object, the field $K$ given with the discrete topology,
and under the above characterization of representable
functors, the operation of completed tensor product with this
object describes the forgetful functor $\fb{Ring}^1_K\to\fb{Ring}^1.$
Hence the result cited shows
that $\Rep{\fb{Ring}^1_K}{\,\fb{Ring}^1}$ has an initial object, the
functor taking $S$ to $S\times S^{\mathrm{op}},$ regarded as a ring.
The coalgebra representing this functor has for underlying
object the free associative unital $\!K\!$-algebra
in two indeterminates, $K\langle x,y\rangle.$
(The same conclusion is true for $K$ any commutative ring, and,
in fact, in a still more general context \cite[\S28]{coalg}.
However, there is no analog in these contexts to the duality between
vector spaces and linearly compact vector spaces, and hence no
interpretation of representable functors in terms of completed
tensor products.
Indeed, the existence and description of the initial
object might not have been discovered
without the motivation of the case where $K$ is a field.)
The above result suggests that for general varieties $\C$ and
$\D,$ the category $\Rep{\C}{\D}$ might have an initial object,
which might have a non-obvious form.
It was this tantalizing hint that led to the present investigations.
Curiously, if, in the preceding example, the unitality condition is
dropped from the class of rings taken as the domain variety, {\em or}
the codomain variety, {\em or} both, then $\Rep{\C}{\D}$
still has an initial object, but not an ``interesting'' one
-- it is trivial, represented by the $\!0\!$- or
$\!1\!$-dimensional $\!K\!$-algebra depending on the case.
This sort of triviality occurs in some very general classes of
situations.
Let us now prove this, so that the reader who wishes to think about
the arguments of later sections in the light of examples of her or
his choosing will be able to consider
cases that have a chance of being nontrivial.
\begin{theorem}\label{T.7/9}
\textup{(i)}\ \ Suppose $\fb{A}$ is a variety of algebras
with no zeroary operations \textup{(}or more generally,
is a category with small coproducts such
that the initial object of $\fb{A}$ admits no
morphisms from non-initial objects into it\textup{)}.
Let $\D$ be
a variety of algebras having {\em at least one} zeroary operation.
Then for any representable functor $F:\fb{A}\to\D,$
the underlying $\!\fb{A}\!$-object of the coalgebra representing
$F$ is the initial object of $\fb{A};$ hence $F$ is the functor taking
every object to the $\!1\!$-element $\!\D\!$-algebra.
\textup{(}So, in particular, $\Rep{\fb{A}}{\D}$ has this trivial functor
as its initial object.\textup{)}\\[.5em]
%
\textup{(ii)}\ \ Suppose $\fb{A}$ is a variety of algebras
with a unique derived zeroary operation \textup{(}or more generally,
is a category with small coproducts whose initial object is also
a final object\textup{)}, and let $\D$ be any variety of algebras.
Then $\Rep{\fb{A}}{\D}$ has an initial object, namely the functor
taking all objects of $\fb{A}$ to the $\!1\!$-element algebra in $\D,$
represented by the initial-final object of $\fb{A}$ with the
unique $\!\D\!$-coalgebra structure that it admits.
\textup{(}However, in this case, $\Rep{\fb{A}}{\D}$ may have
nontrivial {\em non-initial} objects.\textup{)}\\[.5em]
%
\textup{(iii)}\ \ Suppose $\fb{A}$ is a variety of algebras with
more than one derived zeroary operation \textup{(}or more generally,
is a category with small coproducts and a final
object, such that the unique morphism
from the initial to the final object is an epimorphism but
not invertible\textup{),} and let $\D$ be any variety of algebras
having at most one derived zeroary operation.
Then $\Rep{\fb{A}}{\D}$ has an initial object $F.$
Namely --
\vspace{.5em}
\textup{(iii.a)}\ \ If $\D$ has no zeroary operations, $F$ is
represented by the final object of $\fb{A},$ which under the
above hypotheses has a unique $\!\D\!$-coalgebra
structure, and takes objects of $\fb{A}$ that admit morphisms
from the final object into them to the $\!1\!$-element
\textup{(}final\textup{)} $\!\D\!$-algebra, and objects which do not
admit such a morphism to the empty \textup{(}initial\textup{)}
$\!\D\!$-algebra.
\textup{(iii.b)}\ \ If $\D$ has exactly one derived zeroary operation,
then $F$ is represented by the initial object of $\fb{A},$ with
its unique $\!\D\!$-coalgebra
structure, and so takes all objects to the $\!1\!$-element algebra.
\textup{(}But in these situations, too, $\Rep{\fb{A}}{\D}$ may have
nontrivial non-initial objects.\textup{)}
\vspace{.5em}
Thus, assuming $\fb{A}$ a variety of algebras, the only situations
where $\Rep{\fb{A}}{\D}$ can have an initial object whose values are
not exclusively $\!0\!$- or $\!1\!$-element algebras are if
{\em neither} $\fb{A}$ nor $\D$ has zeroary operations, or if
{\em both} $\fb{A}$ and $\D$ have {\em more than} one derived zeroary
operation.
\end{theorem}
\begin{proof}
First, some general observations.
Recall that if a category $\fb{A}$ has an initial object $I,$ then
this is the colimit of the empty diagram, and in particular, is
the coproduct of the empty family of algebras.
It follows that any copower of $I$ is again $I.$
From this we can see that $I$ has a unique
$\!\D\!$-coalgebra structure for every variety $\D.$
More generally, suppose $\fb{A}$ has an initial object $I$
and that $J$ is an epimorph of $I.$
(For example, $\fb{A}$ might be the category of unital commutative
rings, so that $I$ is the ring
$\mathbb{Z}$ of integers, and $J$ might be a prime
field, $\mathbb{Z}/p\mathbb{Z}$ or $\mathbb{Q}.)$
Then a copower $\coprod_\kappa J$ can be identified with the
pushout of the system of maps from $I$
to a $\!\kappa\!$-tuple of copies of $J$ (because
$I$ is initial), so by the assumption that $I\to J$ is an epimorphism,
if $\kappa\neq 0$ that pushout will again be $J.$
(On the other hand, if $\kappa=0,$ then that pushout is $I.)$
The argument of the preceding paragraph now generalizes to show that
for every $\D$ without zeroary operations, such an
object $J$ also has a unique structure of $\!\D\!$-coalgebra.
With these observations in mind, we shall prove the various statements
of the theorem, in each case under the ``more general'' hypothesis.
(In each case, it is easy to see that {}that hypothesis holds
for the particular class of varieties with which the statement begins.)
In the situation of~(i), since $\D$ has a zeroary
operation, the representing object $R$ of $F$ must
have a zeroary co-operation, i.e., a morphism in $\fb{A}$ from
$|R\<|$ to the initial object.
But by assumption on $\fb{A},$
that can only happen if $|R\<|$ {\em is} the initial
object, giving the indicated conclusion.
In the situation of~(ii), we can form a $\!\D\!$-coalgebra in $\fb{A}$
by taking the initial-final object (often called a {\em zero} object)
$Z$ as underlying object, and noting as above that as the
initial object, $Z$ has a unique $\!\D\!$-coalgebra structure.
Because $Z$ is also a final object, the underlying $\!\fb{A}\!$-object
of every $\!\D\!$-coalgebra in $\fb{A}$ admits a unique map to $Z,$
and this clearly forms commuting squares with the co-operations.
Thus, the corresponding coalgebra is final in $\coalg{\fb{A}}{\D},$
and so determines an initial representable functor.
To verify the parenthetical assertion that representable functors other
than this one may also exist, let $\fb{A}=\Gp,$ and note that
the forgetful functor to $\Se$ is representable, as is
the identity functor of $\fb{A}.$
These two examples cover the cases where $\D$ has no zeroary
operations or a unique derived zeroary operation.
If $\D$ has more than one derived zeroary operation,
a representable functor from a category $\fb{A}$ of the indicated sort
must take values in the proper subvariety of $\D$ determined by the
identities saying that the values of all these operations are equal,
since an object of $\fb{A}$ clearly has a unique zeroary co-operation.
Although for some $\D$ (e.g., $\fb{Ring}^1)$ that subvariety is
trivial, for others it is not.
For instance, if $\D$ is the variety of groups with an
additional distinguished element, then the subvariety in
question consists of groups with the identity as that
element, and the functor from $\Gp$ to that subvariety
which leaves the group structure unchanged is not trivial.
In case~(iii.a), the assumption that the unique map from the initial to
the final object of $\fb{A}$ is an epimorphism implies (by the
observations of the second paragraph of this proof) that the
final object of $\fb{A}$ admits a unique $\!\D\!$-coalgebra structure
for every variety $\D$ with no zeroary operations.
It is easy to see that the resulting coalgebra is final in
$\coalg{\fb{A}}{\D},$ as asserted; the description
of the functor represented is also clear.
In the case~(iii.b), where $\D$ has a unique derived zeroary
operation, the conclusion actually requires no assumption
on $\fb{A}$ but that it have an initial object.
(Indeed, statements~(i) and~(ii) imply the same conclusion for
such $\D.)$
To get that conclusion, observe that the value of the unique zeroary
derived operation of $\D$ yields a unique one-element subalgebra in
every object of $\D$ (e.g., when $\D=\Gp$ or $\fb{Monoid},$
the subalgebra $\{e\}).$
Hence the functor represented by the initial object of $\fb{A}$
with its unique $\!\D\!$-coalgebra structure, taking every
object to the one-element $\!\D\!$-algebra, has a unique morphism to
every functor $\fb{A}\to\D,$ hence is initial in $\D^\fb{A},$ and
so, a fortiori, in~$\Rep{\fb{A}}{\D}.$
In these two situations, identity functors and forgetful functors again
show that not every representable functor need be trivial.
\end{proof}
The above theorem, in the case where $\fb{A}$ is
a variety $\C$ of algebras, is summarized in the chart below.
In that chart, $I$ means that the initial representable functor is
represented by the initial object, $F$ means that it is represented
by the final object,
$IF$ means it is represented by the initial-final object, an
exclamation point means that the functor so represented is
the only representable functor, and an exclamation point in
parenthesis means that, though the functor in question may not
be the only one, there is a strong restriction on representable
functors, namely that they take as values algebras in which all
derived zeroary operations are equal.
Stars mark the two cases in which nontrivial initial representable
functors can occur.
\begin{picture}(350,150)
\put(140,145){zeroary derived operations of $\D$}
\put(160,126){$0$}
\put(210,126){$1$}
\put(255,126){$>1$}
\multiput(140,120)(0,-35){4}{\line(1,0){150}}
\multiput(140,120)(50,0){4}{\line(0,-1){105}}
\put(121,098){$0$}
\put(160,098){\Large{$*$}}
\put(205,098){$I$!}
\put(256,098){$I$!}
\put(121,063){$1$}
\put(155,063){$IF$}
\put(205,063){$IF$}
\put(246,063){$IF$(!)}
\put(117,028){${>}\<1$}
\put(160,028){$F$}
\put(210,028){$I$}
\put(259,028){\Large{$*$}}
\put(065,105){zeroary}
\put(065,078){derived}
\put(050,051){operations}
\put(075,024){of $\C$}
\end{picture}
%
In contrast to the triviality of the seven cases covered by the above
theorem, the structure of the initial representable functor
in the two starred cases can be surprisingly rich; we shall
see this for a case belonging to the upper left-hand corner in
\S\ref{S.Set}, and for further examples of both cases
in \S\S\ref{S.BiBi}-\ref{S.Group}.
However, let us note a subcase of the upper left-hand
case where the functors one gets are again fairly degenerate.
\begin{proposition}\label{P.unary}
Suppose $\D$ is a polyunary variety; i.e.,
that $\ari(\alpha)=1$ for all $\alpha\in\Omega_\D.$
Then for any variety $\fb{A}$ of algebras \textup{(}or more
generally, for any category $\fb{A}$ having small colimits and a final
object\textup{)}, the category $\Rep{\fb{A}}{\D}$ has an initial object
$F,$ represented by the final object $T$ of $\fb{A},$ with its
unique $\!\D\!$-coalgebra structure, namely the structure in which
every primitive co-operation is the identity.
Thus, if $\fb{A}$ is a variety, then for every object $A$ of $\fb{A},$
the algebra $F(A)$ has for underlying set
the set of those $x\in|A\<|$ such that $\{x\}$ forms a
subalgebra of $A,$ and has the $\!\D\!$-algebra structure in
which every primitive operation of $\D$ acts as the identity.
\end{proposition}
\begin{proof}
Since $\fb{A}(T,T)$ is a trivial monoid, there is a
unique way to send the primitive unary operations of $\D$
to maps $T\to T,$ and this will make $T$ a $\!\D\!$-coalgebra.
Recalling that $T$ is final in $\fb{A}$ we see that for every
$\!\D\!$-coalgebra $R$ in $\fb{A},$ the unique morphism
$|R\<|\to T$ becomes a morphism of coalgebras, so the coalgebra
described is final in $\coalg{\fb{A}}{\D}.$
The description of the functor represented by this
coalgebra when $\fb{A}$ is a variety is immediate.
\end{proof}
From this point on, when the contrary is not stated,
the domain category of our functors will
be the variety $\C$ (assumed fixed in \S\ref{S.intro});
so when we use the terms ``representable functor'' and
``coalgebra'' without qualification, these will mean ``representable
functor $\C\to\D\!$'', and ``$\!\D\!$-coalgebra in $\C\!$''.
Let me now motivate the technique of proof of the existence of final
objects in $\coalg{\C}{\D},$ then
indicate how to extend that technique to other limits.
Our Lemma~\ref{L.colim} and Freyd's Initial Object Theorem will
establish the existence of an initial representable functor if we can
find a small set $W_\mathrm{funct}$ of representable functors such that
every object of $\Rep{\C}{\D}$ admits a morphism from some member of
$W_\mathrm{funct};$ equivalently, a small set $W_\mathrm{coalg}$ of
coalgebras such that every coalgebra admits a morphism
into a member of $W_\mathrm{coalg}.$
Let us call an object $R$ of $\coalg{\C}{\D}$ ``strongly quasifinal''
if every morphism of coalgebras with domain $R$
which is surjective on underlying algebras is an isomorphism.
(I will motivate this terminology in a moment.)
It will not be hard to show that every coalgebra admits a
morphism onto a strongly quasifinal coalgebra, so it will suffice to
show that up to isomorphism, there is only a small set of these.
To see how to get this smallness condition, note that no strongly
quasifinal coalgebra $R$ can be the codomain of
two distinct morphisms of coalgebras with a common domain $R';$
for if it were, then their coequalizer in $\coalg{\C}{\D}$ (which
by Lemma~\ref{L.colim} has for underlying $\!\C\!$-algebra the
coequalizer of the corresponding maps in $\C)$
would contradict the strong quasifinality condition.
(It is natural to call the property of admitting at most one
morphism from every object ``quasifinality'', hence our use of
``strong quasifinality'' for the above condition that implies it.)
Thus, if we can find a small set $V$ of coalgebras such
that every coalgebra is a union of subcoalgebras isomorphic to
members of $V$ (where we shall define a ``subcoalgebra'' of $R$
to mean a coalgebra which can be mapped into $R$ by a coalgebra
morphism that is one-to-one on underlying $\!\C\!$-algebras, and
also induces one-to-one maps of their copowers),
then the cardinalities of strongly quasifinal coalgebras $R$ will
be bounded by the sum of the cardinalities
of the members of $V$ (since each member of $V$ can be mapped
into $R$ in at most one way), giving the required smallness condition.
Naively, we would like to say that each coalgebra $R$ is the
union of the subcoalgebras ``generated'' by the elements $r\in|R\<|,$
bound the cardinalities of coalgebras that can be
``generated'' by single elements, and take the set of such
coalgebras as our $V.$
Unfortunately, there is not a well-defined concept of the
subcoalgebra generated by an element.
Nevertheless, we shall be able to build up, starting with any
element, or more generally, any subset $X$ of $||R\<||,$ a
subcoalgebra which contains $X,$ and whose cardinality can be
bounded in terms of that of $X.$
How?
We will begin by closing $X$ under the $\!\C\!$-algebra
operations of $|R\<|.$
We will then consider the image of the resulting subalgebra of $|R\<|$
under each {\em co-operation}
$\alpha^R: |R\<|\to\coprod_{\ari(\alpha)}|R\<|$ $(\alpha\in\Omega_\D).$
This image will be contained in the subalgebra of
$\coprod_{\ari(\alpha)}|R\<|$ generated by the images,
under the $\ari(\alpha)$ {\em coprojections}
$|R\<|\to\coprod_{\ari(\alpha)}|R\<|,$ of certain elements of $|R\<|.$
The set of these elements is not, in general, unique, but the number
of them that are needed can be bounded with the help of $\lambda_\C.$
Taking the subalgebra of $|R\<|$ that they generate, we then repeat this
process; after $\lambda_\C$ iterations, it will stabilize, giving
a subalgebra $|R\<|'\subseteq|R\<|$ such that each co-operation
$\alpha^R$ carries $|R\<|'$ into the subalgebra
of $\coprod_{\ari(\alpha)}|R\<|$ generated by the images of $|R\<|'$
under the coprojections.
This $|R\<|'$ still may not define a subcoalgebra of $R,$ because when
we use the inclusion $|R\<|'\subseteq|R\<|$ to induce homomorphisms
%
\begin{equation}\begin{minipage}[c]{35pc}\label{x.R'R}
$\coprod_{\ari(\alpha)}|R\<|'\ \to\ \coprod_{\ari(\alpha)}|R\<|
\qquad(\alpha\in\Omega_\D),$
\end{minipage}\end{equation}
%
these may not be one-to-one; hence the co-operations
$\alpha^R: |R\<|\to\coprod_{\ari(\alpha)}|R\<|,$
though they carry $|R\<|'$ into the image of~(\ref{x.R'R}),
may not lift to co-operations $|R\<|'\to\coprod_{\ari(\alpha)}|R\<|'.$
However, given $|R\<|',$ we can now find a larger subalgebra
$|R\<|'',$ whose cardinality we can again bound, such that the inclusion
$|R\<|''\subseteq|R\<|$ does induce one-to-one maps
$\coprod_\kappa|R\<|''\to\coprod_\kappa|R\<|$ for
all cardinals $\kappa.$
We then have to repeat the process of the preceding paragraph
so that the images of our new algebra under the co-operations
of $R$ are again contained in the subalgebras generated by
the images of the coprojections.
Applying these processes alternately (or better, applying one step
of each alternately, since nothing is gained by iterating one process
to completion before beginning the other), we get a {\em subcoalgebra}
of $R$ containing $X,$ whose cardinality we can bound.
Taking such a bound $\mu$ for the case $\cd(X)=1,$ any set of
representatives of the isomorphism classes of coalgebras of cardinality
$\leq\mu$ gives the $V$ needed to complete our proof.
In view of the messiness of the above construction, the bound on the
cardinality of the
final coalgebra that it leads to is rather large.
But this reflects the reality of the situation.
For instance, if $\C$ is $\fb{Set},$ and $\D$ the variety of sets given
with a single binary operation, we shall see in \S\ref{S.Set} that
the final object of $\coalg{\C}{\D}$ is the Cantor set, with
co-operation given by the natural bijection from that set
to the disjoint union of two copies of itself.
Since this coalgebra has cardinality $2^{\aleph_0},$ the functor
it represents takes a $\!2\!$-element set to a $\!\D\!$-algebra
of cardinality $2^{2^{\aleph_0}}.$
(Incidentally, because in this example, $\C=\fb{Set}$ and the variety
$\D$ is defined without the use of identities, this
final coalgebra is also the
final coalgebra in the sense of \cite{PA+NM} and \cite{Barr},
with respect to the functor taking
every set to the disjoint union of two copies of itself.)
How will we modify the above construction of final
coalgebras to get general small limits in $\coalg{\C}{\D}$?
Consider the task of finding a product of objects
$R_1$ and $R_2$ in this category, i.e., an object universal
among coalgebras $R$ with morphisms $f_1:R\to R_1$ and $f_2:R\to R_2.$
Such a pair of maps corresponds, at the algebra level, to a map
$f:|R\<|\to|R_1|\times|R_2|;$ let us call the latter algebra $S\ba.$
To express in terms of $f$ the fact that $f_1$ and $f_2$
are compatible with the co-operations
of $R_1$ and $R_2,$ let us, for each $\alpha\in\Omega_\D,$ define
$S_\alpha$ to be the $\!\C\!$-algebra
$\coprod_{\ari(\alpha)}|R_1|\times\coprod_{\ari(\alpha)}|R_2|.$
The $\ari(\alpha)$ coprojection maps
$|R_i|\to\coprod_{\ari(\alpha)}|R_i|$ $(i=1,2)$ induce
$\ari(\alpha)$ maps $S\ba\to S_\alpha;$ let us call these
``pseudocoprojections''; thus, each $S_\alpha$ is an object of
$\C$ with $\ari(\alpha)$ pseudocoprojection maps
of $S\ba$ into it, and an additional map of $S\ba$ into it,
induced by $\alpha^{R_1}$ and $\alpha^{R_2},$ which
we shall call the ``pseudo-co-operation'' $\alpha^S.$
We shall call systems $S$ of objects and morphisms of the
sort exemplified by this construction ``pseudocoalgebras''
(Definition~\ref{D.pseudo} below).
Then a coalgebra $R$ with morphisms into $R_1$ and $R_2$ can be regarded
as a coalgebra with a morphism into the above pseudocoalgebra $S.$
More generally, if we are given any small diagram of coalgebras,
then a cone from a coalgebra $R$ to that diagram is equivalent to a
morphism from $R$ to an appropriate pseudocoalgebra.
Thus, it will suffice to show that for every pseudocoalgebra $S,$
the category of coalgebras with morphisms to $S$ has a final object.
The construction sketched above for final coalgebras in fact
goes over with little change to this context.
After obtaining this existence result in the next two
sections, we will show that in many cases,
these colimits can be constructed more explicitly,
as inverse limits of what we shall call ``precoalgebras''.
(The example mentioned above where the final coalgebra
is the Cantor set will lead us to that approach.)
\section{Subcoalgebras of bounded cardinality.}\label{S.sub}
Recall that $\lambda_\C$ is a regular infinite
cardinal such that every primitive operation of $\alpha\in\Omega_\C$
has $\ari(\alpha)<\lambda_\C.$
A standard result is
\begin{lemma}[{\cite[Lemma~8.2.3]{245}}]\label{L.gen_by}
If $A$ is an algebra in $\C$ and $X$ a generating set for $A,$
then every element of $A$ is contained in a subalgebra generated
by $<\lambda_\C$ elements of $X.$\endproof
\end{lemma}
Let us fix a notation for algebras presented by
generators and relations.
For any set $X,$ let $F_\C(X)$ denote the free algebra on $X$ in $\C.$
If $Y$ is a subset of $|F_\C(X)|\times|F_\C(X)|,$
let $\langle X\mid Y\rangle_\C$ be the quotient of
$F_\C(X)$ by the congruence generated by $Y.$
If $A$ is a $\!\C\!$-algebra, a presentation of $A$
will mean an isomorphism with an algebra $\langle X\mid Y\rangle_\C.$
Every algebra $A$ has a canonical presentation, with $X=|A\<|,$ and
$Y$ consisting of all pairs $(\alpha_{F_\C(X)}(x),\alpha_A(x))$
with $\alpha\in\Omega_\C$ and $x\in|A\<|^{\ari(\alpha)}.$
If $X_0$ is a subset of $X,$ we shall often regard $F_\C(X_0)$
as a subalgebra of $F_\C(X).$
Here is the analog of the preceding lemma for relations.
\begin{corollary}\label{C.rels}
Let $A=\langle X\mid Y\rangle_\C,$ let $X_0$ be a subset
of $X,$ and let $p$ and $q$ be elements of $F_\C(X_0)$ which
fall together under the composite of natural
maps $F_\C(X_0)\hookrightarrow F_\C(X)\to A\<.$
Then there exist a set $X_1$ with $X_0\subseteq X_1\subseteq X,$
and a set $Y_1\subseteq Y\cap(|F_\C(X_1)|\times|F_\C(X_1)|),$ such that
the difference-set $X_1-X_0$ and the set $Y_1$ both
have cardinality $<\lambda_\C,$ and such that
$p$ and $q$ already fall together under the composite map
$F_\C(X_0)\hookrightarrow F_\C(X_1)\to
\langle X_1\mid Y_1\rangle_\C.$
\end{corollary}
\begin{proof}
By hypothesis, $(p,q)$ lies in the congruence on $F_\C(X)$
generated by $Y.$
A congruence on $F_\C(X)$ can be described as a subalgebra
of $F_\C(X)\times F_\C(X)$ which is also an equivalence relation
on $|F_\C(X)|,$ and the latter condition can
be expressed as saying that it contains all pairs $(r,r)$
$(r\in|F_\C(X)|),$ and is closed under both the unary
operation $(r,s)\mapsto(s,r)$ and the {\em partial}
binary operation carrying a pair of elements of
the form $((r,s),\ (s,t))$ to the element $(r,t).$
We can apply the preceding lemma to this situation, either using
the observation that the proof of that lemma works equally well
for structures with partial operations, or by noting that if we
extend the above partial operation to a total operation by
making it send $((r,s),\ (s',t))$ to $(r,s)$ if $s\neq s',$
then closure under that total operation is equivalent to closure
under the given partial operation.
Either way, we get the conclusion that if we extend the
$\!\C\!$-algebra structure of $F_\C(X)\times F_\C(X)$ to embrace the
two additional operations expressing symmetry and transitivity, then
our given element $(p,q)$ lies in the subalgebra of the
resulting structure generated by some subset
%
\begin{quote}
$Y_0\ \subseteq\ Y\cup\{(r,r)\mid r\in|F_\C(X)|\}$
\end{quote}
%
of cardinality $<\lambda_\C.$
The elements of $Y_0$ coming from $Y$ will form
a set $Y_1$ of cardinality $<\lambda_\C,$ and
by the preceding lemma, each component of each of these elements
will lie in the subalgebra of $F_\C(X)$
generated by some subset of $X$ of cardinality $<\lambda_\C,$
and the same will be true of each of the $<\lambda_\C$ elements $r$
occurring in pairs $(r,r)\in Y_0.$
As $\lambda_\C$ is a regular cardinal, the union of these
subsets will be a set
$X_1'\subseteq X$ of cardinality $<\lambda_\C$ such that $Y_0$
lies in $|F_\C(X_1')|\times |F_\C(X_1')|.$
By construction, $(p,q)$ lies in the subalgebra of this product
generated by $Y_0$ under our extended algebra structure.
Letting $X_1=X_0\cup X_1',$ we conclude that $p$ and $q$ fall together
in $\langle X_1\mid Y_1\rangle_\C,$ as claimed, and that $X_1$
and $Y_1$ satisfy the desired cardinality restrictions.
(Incidentally, this proof would not have worked if we had rendered the
elements $(r,r)$ by zeroary operations, rather than generating elements,
since $|F_\C(X_1)|\times |F_\C(X_1)|$ would not have been closed
in $|F_\C(X)|\times |F_\C(X)|$ under all these operations.)
\end{proof}
To go from the bounds on the cardinalities of the sets constructed
in the above lemma and corollary to bounds
on the cardinalities of the algebras they generate, we will want
\begin{definition}\label{D.exp<}
If $\kappa$ is a cardinal and $\lambda$ an infinite cardinal, then
$\kappa^{\lambda-}$ will denote the least cardinal $\mu\geq\kappa$
such that $\mu^\iota=\mu$ for all $\iota<\lambda.$
\end{definition}
To see that this makes sense, note that $\kappa^\lambda$ is
not changed on exponentiating by $\lambda,$ hence, a fortiori,
it is not changed on exponentiating by any positive $\iota<\lambda,$
so the class of cardinals $\mu$ with that property is nonempty; hence
it has a least member.
Immediate consequences of the above definition are
%
\begin{equation}\begin{minipage}[c]{35pc}\label{x.^*l-twice}
$(\kappa^{\lambda-})^{\lambda-}=\ \kappa^{\lambda-}.$
\end{minipage}\end{equation}
%
\begin{equation}\begin{minipage}[c]{35pc}\label{x.sup}
$\max(\kappa,\mu)^{\lambda-}\ =\ %
\max(\kappa^{\lambda-},\mu^{\lambda-}).$
\end{minipage}\end{equation}
%
Note also that for $\iota<\lambda$ and $\kappa>1$ we have
$\kappa^{\lambda-}\geq\kappa^\iota>\iota.$
Hence
%
\begin{equation}\begin{minipage}[c]{35pc}\label{x.geq*l}
If $\kappa>1,$ then $\kappa^{\lambda-}\geq\lambda.$
\end{minipage}\end{equation}
%
We also see
%
\begin{equation}\begin{minipage}[c]{35pc}\label{x.^aleph0}
For all $\kappa>1$ one has
$\kappa^{\aleph_0-}=\ \max(\kappa,\,\aleph_0).$
\end{minipage}\end{equation}
(Leo Harrington has pointed out to me that for $\lambda$ a
regular cardinal, which will always be the case below,
what I am calling $\kappa^{\lambda-}$ can be shown
equal to what set theorists call $\kappa^{<\lambda},$
namely $\sup_{\iota<\lambda}\kappa^\iota.$
However, we shall not need this fact.)
The following result is very likely known.
\begin{lemma}\label{L.gen_card}
If a $\!\C\!$-algebra $A$ is generated by a set $X,$ then
%
\begin{quote}
$\cd(|A\<|)\ \leq\ \max(\cd(X),\nolinebreak[2]\,
\cd(\Omega_\C),\,2)^{\lambda_\C-}.$
\end{quote}
%
\end{lemma}
\begin{proof}
We construct subsets $X_0\subseteq X_1\subseteq\dots
\subseteq X_{\lambda_\C}$ of $|A\<|$ as follows:
Take $X_0=X.$
For every successor ordinal $\iota\<{+}1$ let
$X_{\iota+1}$ consist of all elements of $X_\iota$ and all
elements of the form $\alpha_A(x)$ where $\alpha\in\Omega_\C$
and $x=(x_\gamma)_{\gamma\in\ari(\alpha)}\in X_\iota^{\ari(\alpha)}.$
For every limit ordinal $\iota$ let
$X_\iota=\bigcup_{\eta\in\iota} X_\eta.$
Since $\lambda_\C$ is a regular cardinal exceeding all the
cardinals $\ari(\alpha)$ $(\alpha\in\Omega_\C),$ we see that
$X_{\lambda_\C}$ is closed under the operations of $\C,$
so as it contains $X_0=X,$ it is all of $|A\<|.$
Let us now show by induction that for
all $\iota\leq\lambda_\C,$
%
\begin{equation}\begin{minipage}[c]{35pc}\label{x.leqmu}
$\cd(X_\iota)\ \leq\ \max(\cd(X),\,\cd(\Omega_\C),\,2)^{\lambda_\C-}.$
\end{minipage}\end{equation}
%
Clearly,~(\ref{x.leqmu}) holds for $\iota=0.$
Let us write $\mu$ for the right-hand side of~(\ref{x.leqmu}), which
is independent of $\iota,$ and by~(\ref{x.geq*l}) is $\geq\lambda_{\C}.$
At each successor ordinal $\iota\<{+}1,$ the number of elements
we adjoin as values of each operation $\alpha_A$ is at most
$\cd(X_\iota)^{\ari(\alpha)}\leq\mu^{\ari(\alpha)}\leq\mu,$
since $\ari(\alpha)<\lambda_\C.$
Hence, doing this for all $\cd(\Omega_\C)$ operations brings
in $\leq\mu\cdot\cd(\Omega_\C)\linebreak[0]\leq\mu$ elements.
Likewise, at a limit ordinal $\iota,$ we take the union of a family
of $\cd(\iota)\leq\linebreak[0]\lambda_\C\leq\mu$
sets of cardinality $\leq\mu,$
hence again get a set of cardinality $\leq\mu.$
Taking $\iota=\lambda_\C$ in~(\ref{x.leqmu}), we get
$\cd(|A\<|)\leq\mu,$ as required.
\end{proof}
For the step in our proof where we will enlarge an arbitrary
subalgebra of $|R\<|$ to a subcoalgebra of $R,$ we will first need
to define ``subcoalgebra''.
For this in turn we will need
\begin{definition}\label{D.pure}
A subalgebra $A$ of a $\!\C\!$-algebra $B$ will be called
{\em copower-pure} if for every cardinal $\kappa,$ the induced map
%
\begin{equation}\begin{minipage}[c]{35pc}\label{x.pure}
$\coprod_\kappa A\ \to\ \coprod_\kappa B$\ \ is one-to-one.
\end{minipage}\end{equation}
%
When this holds, we shall often identify
$\coprod_\kappa A$ with its image in~$\coprod_\kappa B\<.$
\end{definition}
An example of a subalgebra which is not copower-pure was noted
in \cite[discussion preceding Question~4.5]{embed}:
Let $\C$ be the variety of groups determined by the identities satisfied
in the infinite dihedral group, which include $x^2 y^2=y^2 x^2,$
but not $xy=yx,$ nor $x^n=1$ for any $n>0.$
Let $B$ be the infinite cyclic group $\langle x\rangle,$
which is free on one generator in $\C,$
and $A$ the subgroup $\langle x^2\rangle\subseteq B\<.$
Then $B\cP B$ is the free $\!\C\!$-algebra on two generators,
$x_0$ and $x_1,$ hence is noncommutative,
hence the same is true of $A\cP A.$
But the image of $A\cP A$ in $B\cP B$ is generated by $x_0^2$
and $x_1^2,$ which commute by the identity noted; so the
map $A\cP A\to B\cP B$ is not an embedding.
\begin{lemma}\label{L.pure<*l}
A subalgebra $A$ of a $\!\C\!$-algebra $B$ is copower-pure if and
only if~\textup{(\ref{x.pure})} holds for all $\kappa<\lambda_\C.$
\end{lemma}
\begin{proof}
``Only if'' is clear; for the converse,
assume~(\ref{x.pure}) holds whenever $\kappa<\lambda_\C.$
Suppose we are given $\kappa$ not necessarily $<\lambda_\C,$ and
distinct elements $p,q\in|\coprod_\kappa A\<|.$
Since $\coprod_\kappa A$ is generated by the images of $A$ under
the $\kappa$ coprojection maps, we see from Lemma~\ref{L.gen_by} that
$p$ and $q$ will lie in the subalgebra generated by the
copies of $A$ indexed by some subset $I\subseteq\kappa$
with $0<|I|<\lambda_\C.$
Now a set-theoretic retraction of $\kappa$ onto $I$
(a left inverse to the inclusion of $I$ in $\kappa)$
induces algebra retractions $\coprod_\kappa A\to\coprod_I A$ and
$\coprod_\kappa B\to\coprod_I B,$ making a commuting square
with the maps $\coprod_\kappa A\to\coprod_\kappa B$
and $\coprod_I A\to\coprod_I B.$
By choice of $I,$ the elements $p$ and $q$ lie in
the subalgebra $\coprod_I A\subseteq\coprod_\kappa A,$
and since $|I|<\lambda_\C,$ the map
$\coprod_I A\to\coprod_I B$ is one-to-one; so $p$ and $q$ have
distinct images in $\coprod_I B,$ and hence in $\coprod_\kappa B.$
\end{proof}
We are now ready for
\begin{definition}\label{D.subcoalg}
If $R$ and $R'$ are $\!\D\!$-coalgebras in $\C,$ then
we will call $R'$ a {\em subcoalgebra} of $R$ if $|R'|$ is
a copower-pure $\!\C\!$-subalgebra of $|R\<|,$ and for
each $\alpha\in\Omega_\D,$ the co-operation $\alpha^{R'}$
is the restriction to $|R'|\subseteq|R\<|$ of $\alpha^R.$
\end{definition}
Thus, for a subalgebra $A\subseteq |R\<|$ to yield a subcoalgebra
of $R,$ it must be copower-pure, and have the property that
each $\alpha^R$ carries $A$ into $\coprod_{\ari(\alpha)}A.$
(Remark: It would probably be more natural in the above definition
to require~(\ref{x.pure}) to hold only for $\kappa<\lambda_\D;$
and perhaps to remove that condition entirely when the arities of the
operations of $\D$ are all $\leq 1.$
But for simplicity, we will stick with the above definition.)
We shall now prove the existence of subcoalgebras satisfying
cardinality bounds, as sketched earlier.
(Note that the ``$\!\mu\!$'' of the next result is not necessarily the
same as the value so named in the proof of Lemma~\ref{L.gen_card},
since it also involves $\cd(\Omega_\D).)$
\begin{theorem}\label{T.subcoalg}
Let $R$ be a $\!\D\!$-coalgebra object of $\C,$ and $X$
a subset of $||R\<||.$
Then $R$ has a subcoalgebra $R'$ whose underlying set contains $X,$
and has cardinality at most
%
\begin{equation}\begin{minipage}[c]{35pc}\label{x.mu}
$\mu\ =\ %
\max(\cd(X),\,\cd(\Omega_\C),\,\cd(\Omega_\D),\,2)^{\lambda_\C-}.$
\end{minipage}\end{equation}
%
\end{theorem}
\begin{proof}
We shall construct a chain of subalgebras
$A_0\subseteq A_1\subseteq\dots\subseteq A_{\lambda_\C}$ of $|R\<|,$
and show that $A_{\lambda_\C}$ is the underlying $\!\C\!$-algebra of
a subcoalgebra $R'$ with the asserted properties.
We take for $A_0$ the subalgebra of $|R\<|$ generated by $X;$
by Lemma~\ref{L.gen_card} this has cardinality $\leq\mu.$
Assuming for some $\iota<\lambda_\C$ that $A_\iota$ has been
constructed, and has cardinality at most $\mu,$ we obtain
$A_{\iota+1}$ by adjoining $\leq\mu$ further elements chosen as follows.
First, for every $\kappa<\lambda_\C,$ and every pair of elements
$p,q\in|\coprod_\kappa A_\iota|$ which have equal image under the
natural map $\coprod_\kappa A_\iota\to\coprod_\kappa |R\<|,$
I claim we can adjoin to $A_\iota$ a set of
$<\lambda_\C$ elements whose presence causes the images of these
elements in the copower of the resulting algebra to fall together.
Indeed, this follows from Corollary~\ref{C.rels}, and the observation
that given a presentation $|R\<|=\langle X\mid Y\rangle_\C,$ the
copower $\coprod_\kappa |R\<|$ can be presented by taking the union of
$\kappa$ copies of $X,$ and for each of these, a copy of $Y.$
Note also that $\coprod_\kappa A_\iota$ is generated by
$\kappa$ copies of $A_\iota,$ which has cardinality $\leq\mu;$
hence by Lemma~\ref{L.gen_card} it itself has cardinality $\leq\mu,$
hence there are at most $\mu$ such pairs $p,q$ to deal with; so
for each $\kappa<\lambda_\C,$ we are adjoining at most $\mu$ elements.
The number of such cardinals $\kappa$ is $\leq\lambda_\C;$
so in handling all such pairs $p,q,$ for all $\kappa<\lambda_\C,$
we adjoin $\leq\mu$ new elements of $|R\<|$ to $A_\iota.$
In addition, for each operation symbol $\alpha\in\Omega_\D,$ and
each $p\in|A_\iota|,$ consider the image of $p$ under
the co-operation $\alpha^R: |R\<|\to\coprod_{\ari(\alpha)}|R\<|.$
By Lemma~\ref{L.gen_by}, this will lie in the subalgebra
of $\coprod_{\ari(\alpha)}|R\<|$ generated by a subset $X_{p,\alpha},$
having cardinality $<\lambda_\C,$ of the generating set for
that copower given by the union of the images of the coprojections.
So $X_{p,\alpha}$ is contained in the
union of the images, under those coprojections, of a
set $X'_{p,\alpha}$ of $<\lambda_\C$ elements of $|R\<|.$
For each $p\in|A_\iota|$ and $\alpha\in\Omega_\D,$
let us include such a set $X'_{p,\alpha}$
in the set of elements we are adjoining to $A_\iota.$
Letting $p$ run over the $\leq\mu$ elements of $A_\iota,$
and $\alpha$ over the elements of $\Omega_\D,$
we see that this process adjoins
$\leq\mu\cdot\cd(\Omega_\D)\cdot\lambda_\C\leq\mu$ new elements.
Let $A_{\iota+1}$ be the subalgebra of $|R\<|$ obtained by
adjoining to $A_\iota$ the two families of elements described in
this and the preceding paragraph.
On the other hand, if $\iota$ is a limit ordinal, we let $A_\iota$
be the $\!\C\!$-subalgebra of $|R\<|$ generated by
$\bigcup_{\eta\in\iota} |A_\eta|.$
That union, being a union of $\leq\lambda_\C$ sets of
cardinality $\leq\mu,$ will itself have
cardinality $\leq\mu,$ and it follows from Lemma~\ref{L.gen_card}
that $A_\iota$ will as well.
Now consider the subalgebra $A = A_{\lambda_\C}$ of $|R\<|.$
Since $\lambda_\C$ is by assumption a regular
cardinal, the union $\bigcup_{\eta\in\lambda_\C} |A_\eta|$
involved in the construction of $A$ is over a chain of cofinality
$\lambda_\C,$ which strictly majorizes the arities of
all operations of $\C;$ hence that union is closed
under those operations, so $|A\<|$ is that union.
I claim that $A$ is copower-pure in $|R\<|.$
Indeed, given any $\kappa<\lambda_\C,$ and elements $p,\ q$ of
$\coprod_\kappa A$ that fall together $\coprod_\kappa |R\<|,$
we can find $<\lambda_\C$ elements of $A$ such that $p$ and $q$ lie
in the subalgebra of $\coprod_\kappa A$ generated by images of
those elements under coprojection maps; and we can then find
some $\iota$ such that all those elements lie in $A_\iota.$
Thus we get $p',q'\in|\coprod_\kappa A_\iota|$ which map to
$p,q\in|\coprod_\kappa A\<|;$ so their images under the composite map
$\coprod_\kappa A_\iota\to \coprod_\kappa A \to\coprod_\kappa |R\<|$
fall together.
Hence, by the construction of $A_{\iota+1},$ the images of $p'$
and $q'$ fall together in $\coprod_\kappa A_{\iota+1},$ hence
they do so in $\coprod_\kappa A,$ i.e., $p=q,$ as required.
It remains to show that
each co-operation $\alpha^R$ $(\alpha\in\Omega_\D)$ carries
$A\subseteq|R\<|$ into the subalgebra
$\coprod_{\ari(\alpha)}A$ of $\coprod_{\ari(\alpha)}|R\<|.$
Every element $p\in|A\<|$ lies in some $A_\iota,$
and by construction, $A_{\iota+1}$ contains elements which
guarantee that $\alpha^R(p)$ lies in the subalgebra of
$\coprod_{\ari(\alpha)}|R\<|$ generated by the image of
$\coprod_{\ari(\alpha)}A_{\iota+1},$ hence, a fortiori,
in $\coprod_{\ari(\alpha)}A\<.$
\end{proof}
\begin{corollary}\label{C.subcoalg}
If $R$ is a $\!\D\!$-coalgebra object of $\C,$ then for every element
$p$ of $|R\<|$ there is a subcoalgebra $R'$ of $R$ whose underlying
$\!\C\!$-algebra contains $p,$ and has cardinality at most
$\max(\cd(\Omega_\C),\,\cd(\Omega_\D),\,2)^{\lambda_\C-}.$
In particular, if $\C$ and $\D$ each have at most countably
many operations, and all operations of $\C$ are finitary,
then every element of a $\!\D\!$-coalgebra object of $\C$ is
contained in a countable or finite subcoalgebra.\endproof
\end{corollary}
\section{Pseudocoalgebras, and the solution set
condition.}\label{S.pseudo}
We now come to the pseudocoalgebras of our sketched development.
\begin{definition}\label{D.pseudo}
By a {\em $\!\D\!$-pseudocoalgebra} in a category $\bf{A},$ we
shall mean a $\!4\!$-tuple
%
\begin{equation}\begin{minipage}[c]{35pc}\label{x.pseudo}
$S\ =\ (S\ba\<,\ (S_\alpha)_{\alpha\in\Omega_\D},\ %
(c^S_{\alpha,\<\iota})_{\alpha\in\Omega_\D,\,\<\iota\in\ari(\alpha)},\ %
(\alpha^S)_{\alpha\in\Omega_\D}),$
\end{minipage}\end{equation}
%
where $S\ba$ and the $S_\alpha$ are objects of $\fb{A}$
\textup{(}the ``base object'' and
the ``pseudocopower objects''\textup{)},
and for each $\alpha\in\Omega_\D,$ $\alpha^S$ \textup{(}the
$\!\alpha\!$-th ``pseudo-co-operation''\textup{)} and
the $c^S_{\alpha,\<\iota}$ \textup{(}the ``pseudocoprojections'',
one for each $\iota\in\ari(\alpha))$ are morphisms $S\ba\to S_\alpha.$
A morphism of $\!\D\!$-pseudocoalgebras $f:S\to S'$
will mean a family of morphisms
%
% below: switching to 34pc because 2-digit numbers reached; better fix?
\begin{equation}\begin{minipage}[c]{34pc}\label{x.psmorph}
$f\ba: S\ba\to S'\ba\<,$\quad and\quad
$f_\alpha: S_\alpha\to S'_\alpha$\quad
$(\alpha\in\Omega_\D)$
\end{minipage}\end{equation}
%
which make commuting squares with the
$c^S_{\alpha,\<\iota}$ and $c^{S'}_{\alpha,\<\iota},$
and with the $\alpha^S$ and $\alpha^{S'}.$
The category of $\!\D\!$-pseudocoalgebras in $\fb{A}$ will be
denoted $\pscoalg{\fb{A}}{\D}.$
\end{definition}
Note that the operation-set $\Omega_\D$ and arity-function $\ari_\D$
of $\D$ come into the definition of $\!\D\!$-pseudo\-coalgebra,
but the identities of $\D$ do not.
In the use we will make of pseudocoalgebras, the fact that those
identities are, by definition, cosatisfied by the
{\em $\!\D\!$-coalgebras} we map to them will be all that matters.
Let us make clear in what sense one can map coalgebras to
pseudocoalgebras.
\begin{definition}\label{D.psi}
If $R$ is a $\!\D\!$-coalgebra, or more generally,
an $\!\Omega_\D\!$-coalgebra, in $\fb{A},$ then we define
the associated $\!\D\!$-pseudocoalgebra $\psi(R)$ to have
%
\begin{quote}
$\psi(R)\ba\ =\ |R\<|,$\\[.17em]
$\psi(R)_\alpha\ =\ \coprod_{\ari(\alpha)}|R\<|,$\\[.17em]
$c^{\psi(R)}_{\alpha,\<\iota}=$ the $\!\iota\!$-th coprojection:
$|R\<|\to\coprod_{\ari(\alpha)}|R\<|,$\quad and\\[.17em]
$\alpha^{\psi(R)}\ =\ \alpha^R:\ |R\<|\to\coprod_{\ari(\alpha)}|R\<|\<.$
\end{quote}
Clearly, $\psi$ yields a full and faithful functor
$\coalg{\fb{A}}{\<\Omega_\D\mbox{-}\fb{Alg}}\to\pscoalg{\fb{A}}{\D},$
so when there is no danger of ambiguity, we shall
treat $\coalg{\fb{A}}{\<\Omega_\D\mbox{-}\fb{Alg}}$ and
its subcategory $\coalg{\fb{A}}{\D}$ as full subcategories of
$\pscoalg{\fb{A}}{\D};$ in particular
we shall speak of morphisms from coalgebras to pseudocoalgebras.
If $S$ is a $\!\D\!$-pseudocoalgebra,
then a $\!\D\!$-coalgebra given with a morphism
to $S,$ i.e., an object of the comma category
$(\coalg{\fb{A}}{\D}\downarrow S),$
will be called a $\!\D\!$-coalgebra {\em over}~$S.$
We shall say that a pseudocoalgebra $S$ ``is an
$\!\Omega_\D\!$-coalgebra'' if it is isomorphic to
$\psi(R)$ for some $\!\Omega_\D\!$-coalgebra $R,$ in
other words, if for every $\alpha,$ the object $S_\alpha$
is the copower $\coprod_{\ari(\alpha)}S\ba,$ with the
pseudocoprojections $c^S_{\alpha,\<\iota}$ as the coprojections.
We will say that $S$ is a $\!\D\!$-coalgebra if it is
an $\!\Omega_\D\!$-coalgebra $R$ which cosatisfies
the identities of~$\D.$
\end{definition}
Note that if $R$ is a coalgebra and $S$ a pseudocoalgebra, then
every morphism $f:R\to S$ is determined by
$f\ba:|R\<|\to S\ba,$ since once this is given, the components
$f_\alpha$ are uniquely determined by the property of
making commuting squares with the coprojections and pseudocoprojections,
via the universal property of $\coprod_{\ari(\alpha)}|R\<|.$
Of course, in general not every map $f\ba:|R\<|\to S\ba$
induces a morphism $f:R\to S,$ since the maps $f_\alpha$ so
determined by $f\ba$ may fail to satisfy
the remaining condition, that the squares they make with the
co-operations and pseudo-co-operations commute.
We now again restrict attention to the case where the
variety $\C$ plays the role of $\fb{A}.$
To the convention that ``coalgebra'', unmodified, means
``$\!\D\!$-coalgebra in $\!\C\!$''
we add the convention that ``pseudocoalgebra'', unmodified, means
``$\!\D\!$-pseudocoalgebra in $\!\C\!$''.
Note that these pseudocoalgebras are simply a kind of many-sorted
algebra, so there is no difficulty constructing limits of such objects.
We recall from Lemma~\ref{L.colim}
that $\coalg{\C}{\D}$ has small colimits,
given on underlying $\!\C\!$-algebras by the colimits of
the corresponding algebras in $\C.$
It follows that if we have a diagram of coalgebras over a fixed
pseudocoalgebra $S,$ its colimit coalgebra has an induced
morphism into $S,$ and so will also be the colimit of the given diagram
in the comma category $(\coalg{\C}{\D}\downarrow S).$
\begin{definition}\label{D.quasifinal}
A morphism of coalgebras will be called {\em surjective} if
it is surjective on underlying $\!\C\!$-objects.
If $f:R\to R'$ is a surjective morphism in
$(\coalg{\C}{\D}\downarrow S),$ for some pseudocoalgebra $S,$
then $R'$ \textup{(}given with the map $f$ from $R)$
will be called an {\em image} coalgebra of $R$ over $S.$
To avoid dealing with the non-small set of isomorphic copies of each
such image, we shall call an image coalgebra $R'$ of $R$ {\em standard}
if the map $||R\<||\to||R'||$ is the canonical map from a set to its
set of equivalence classes under an equivalence relation.
A coalgebra $R$ over $S$ will be called {\em strongly
quasifinal} over $S$ if the only surjective morphisms out of $R$ in
$(\coalg{\C}{\D}\downarrow S)$ are the isomorphisms.
\end{definition}
Given a coalgebra $R$ over a pseudocoalgebra $S,$
the category of all standard images of $R$ over $S$ will form a
partially ordered set, isomorphic to a sub-poset
of the lattice of congruences on the $\!\C\!$-algebra $|R\<|.$
We cannot expect that the set of congruences on $|R\<|$ such that
the $\!\D\!$-coalgebra structure of $R$ extends to the resulting
factor-algebra will be closed under intersections; but it will be
closed under arbitrary joins, since, as just noted, colimits of
coalgebras over $S$ correspond to
colimits of underlying $\!\C\!$-objects.
We deduce
\begin{lemma}\label{L.quasifinal}
Let $S$ be a pseudocoalgebra and $R$ a coalgebra over $S.$
Then the standard images of $R$ over $S$ form a
\textup{(}small\textup{)} complete lattice.
The greatest element of this lattice is, up to isomorphism, the
unique strongly quasifinal homomorphic image of $R$ over~$S.$\endproof
\end{lemma}
We can now show that, up to isomorphism, the strongly
quasifinal coalgebras over $S$ form a small set.
\begin{lemma}\label{L.no_two_iso}
Let $S$ be a pseudocoalgebra, and $R$ a strongly quasifinal
coalgebra over $S.$
Then distinct subcoalgebras of $R$ are nonisomorphic as
coalgebras over~$S.$
Hence in view of Corollary~\ref{C.subcoalg},
$\cd(||R\<||)$ is $\leq$ the sum of the cardinalities of
all \textup{(}up to isomorphism over $S)$ coalgebras over $S$
of cardinality at most
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.mu=}
$\max(\cd(\Omega_\C),\,\cd(\Omega_\D),\,2)^{\lambda_\C-}.$
\end{minipage}\end{equation}
%
This sum is at most
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.card}
$\max(\cd(|S\ba|),\lambda_\D)%
^{\max(\cd(\Omega_\C),\,\cd(\Omega_\D),\,2)^{\lambda_\C-}}.$
\end{minipage}\end{equation}
Hence, up to isomorphism, there is only a small set of coalgebras
$R$ strongly quasifinal over $S.$
\end{lemma}
\begin{proof}
If we had two distinct embeddings into $R$ over $S$ of some coalgebra
$R'$ over $S,$ then the coequalizer of the resulting diagram
$R'\stackrel{\longrightarrow}{\scriptstyle{\longrightarrow}}R$ would
be a proper image of $R$ over $S,$ contradicting the
strong quasifinality of $R.$
This gives the assertion of the first paragraph.
Hence if we break the subcoalgebras of $R$ of cardinality at
most~(\ref{x.mu=}) into their isomorphism classes over $S,$ no more
than one copy of each can occur, and by Corollary~\ref{C.subcoalg},
such subcoalgebras have union $R,$ giving the second assertion,
from which the final sentence of the lemma clearly follows.
To get the explicit bound~(\ref{x.card}), let us write $\mu$ for
the cardinal~(\ref{x.mu=}) and $\nu$ for~(\ref{x.card}).
We shall show that the number of structures of coalgebra over
$S$ on a set $X$ of cardinality $\leq\mu$ is at most $\nu.$
Since there are $\leq\mu<\nu$ cardinalities $\leq\mu,$
this will give $\leq\nu$ isomorphism classes of strongly quasifinal
coalgebras over $S$ altogether, and since each such coalgebra
has cardinality $\leq\mu<\nu,$ the sum of their cardinalities
will be $\leq\nu,$ as required.
To bound the number of structures of coalgebra over $S$ on $X,$
note that such a structure is determined by
several maps (subject to restrictions that we will not repeat
because they do not come into our calculations):
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.toS}
a map $X\to|S\ba|,$
\end{minipage}\end{equation}
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.fromprod}
for each $\alpha\in\Omega_\C,$ a map $X^{\ari(\alpha)}\to X,$
\end{minipage}\end{equation}
%
making $X$ a $\!\C\!$-algebra $A;$ and once this has been done,
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.tocoprod}
for each $\alpha\in\Omega_\D,$ a map $A\to\coprod_{\ari(\alpha)}A,$
\end{minipage}\end{equation}
%
giving the $\!\D\!$-coalgebra structure.
Given $X$ of cardinality $\leq\mu,$ the number of
possible % word inserted to avoid overfull box
choices for~(\ref{x.toS}) is bounded by $\cd(|S\ba|)^\mu.$
For each $\alpha\in\Omega_\C,$ the number of choices for the map
in~(\ref{x.fromprod}) is $\leq \mu^{\mu^{\ari(\alpha)}},$ but by
definition (see~(\ref{x.mu=})), $\mu$ is not increased by
exponentiation by
$\ari(\alpha)<\lambda_\C,$ so this bound is $\leq\mu^\mu.$
Letting $\alpha$ run over $\Omega_\C,$ we conclude that the
number of choices for~(\ref{x.fromprod}) is
$\leq(\mu^\mu)^{\cd(\Omega_\C)}=\mu^{\mu\,\cd(\Omega_\C)}.$
But again by definition, $\mu\geq\cd(\Omega_\C),$ so the product
$\mu\,\cd(\Omega_\C)$ simplifies to $\mu;$ hence the
number of choices for~(\ref{x.fromprod}) is $\leq\mu^\mu.$
Finally, for each $\alpha\in\Omega_\D,$ the copower
in~(\ref{x.tocoprod}) will be generated by an $\!\ari(\alpha)\!$-tuple
of copies of $X,$ hence by a set of cardinality
$\leq\ari(\alpha)\,\mu\leq\lambda_\D\,\mu,$
so by Lemma~\ref{L.gen_card} that copower has cardinality
$\leq\max(\lambda_\D\,\mu,\,\cd(\Omega_\C),\,2)^{\lambda_\C-}=
(\lambda_\D\,\mu)^{\lambda_\C-}$ (since
$\cd(\Omega_\C)$ and $2$ are majorized by $\mu).$
Hence for each $\alpha\in\Omega_\D$
the number of maps as in~(\ref{x.tocoprod}) is
$\leq((\lambda_\D\,\mu)^{\lambda_\C-})^\mu
\leq((\lambda_\D\,\mu)^\mu)^\mu=(\lambda_\D\,\mu)^\mu.$
Letting $\alpha$ run over $\Omega_\D,$ we get an additional
factor of $\cd(\Omega_\D)$ in the exponent, but this again is
absorbed by $\mu.$
Bringing together the choices made in~(\ref{x.toS}),~(\ref{x.fromprod})
and~(\ref{x.tocoprod}), we get the bound
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.total}
$\cd(|S\ba|)^\mu\ \mu^\mu\ (\lambda_\D^\mu\ \mu^\mu)$
\end{minipage}\end{equation}
%
on the number of possible structures.
Note also that for any infinite cardinal $\lambda$ and any
cardinal $\kappa>1,$ one knows that $\kappa^\lambda>\lambda,$
hence $\lambda^\lambda\leq(\kappa^\lambda)^\lambda=
\kappa^{\lambda\lambda}=\kappa^\lambda.$
Applying this with $\mu$ in the role
of $\lambda$ and $\lambda_\D$ in the role of $\kappa,$
we see that the $\mu^\mu$ terms can be dropped from~(\ref{x.total}).
Rewriting the product as a maximum, and putting in the
definition~(\ref{x.mu=}) of $\mu,$ we get the desired
bound~(\ref{x.card}).
\end{proof}
We deduce
\begin{theorem}\label{T.final}
Let $S$ be a pseudocoalgebra.
Then the category of coalgebras over $S$ has a final object.
The cardinality of the underlying set of this object is
at most~\textup{(\ref{x.card})}.
\end{theorem}
\begin{proof}
Since the category has small colimits, and every object has a
morphism into a strongly quasifinal object, and up to
isomorphism there is only a small set of such objects, the dual
statement to the Initial Object Theorem gives the required final object.
Since a final object is in particular strongly quasifinal, the
cardinality of its underlying set is bounded by~(\ref{x.card}).
\end{proof}
We can now show $\coalg{\C}{\D}$ complete.
Given a small category $\fb{E}$ and
a functor $F:\fb{E}\to\coalg{\C}{\D},$
each coalgebra $F(E)$ $(E\in\mathrm{Ob}(\fb{E}))$ can be
regarded as a pseudocoalgebra, $\psi(F(E)),$ and
this system of pseudocoalgebras has a limit, which can be
constructed objectwise.
Let
%
\begin{quote}
$S\ =\ \limit_\fb{E}\ \psi\,F.$
\end{quote}
If $R$ is a coalgebra, then a cone in $\coalg{\C}{\D}$ from
$R$ to the diagram of coalgebras $F$ is equivalent to
a morphism of pseudocoalgebras $R\to S.$
Hence the final object of $(\coalg{\C}{\D}\downarrow S)$
corresponds to a limit of $F.$
This gives
\begin{theorem}\label{T.lim}
$\coalg{\C}{\D}$ has small limits; equivalently,
$\Rep{\C}{\D}$ has small colimits.
Moreover, given a small category $\fb{E}$ and
a functor $F:\fb{E}\to\coalg{\C}{\D},$ we have
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.limcard}
$\cd(||\limit F||)\ \leq\ %
\max\<(\cd(\limit_\fb{E}||F(E)||),\,\lambda_\D)%
^{\max(\cd(\Omega_\C),\,\cd(\Omega_\D),\,2)^{\lambda_\C-}}.$
\end{minipage}
\end{equation}
\endproof
\end{theorem}
\vspace{.8em}
Since most of algebra is done with finitary operations, and
often with only finitely many of them, let us record what our result
says in that case.
\begin{corollary}\label{C.c}
If $\C$ and $\D$ each have only finitely many operations, and all
of these are finitary, then the final $\!\D\!$-coalgebra object
of $\C$ has underlying set of at most continuum cardinality.
In fact, this remains true if ``finitely many operations''
is generalized to ``at most countably many operations'', the assumption
of finite arity on the operations of $\D$ \textup{(}but not of
$\C)$ is generalized to that of arity less than the continuum,
and ``the final $\!\D\!$-coalgebra of $\C\!$''
is generalized to ``the limit of any
diagram of at most countably many $\!\D\!$-coalgebra objects of $\C,$
each of which has underlying set of at most continuum cardinality.''
\end{corollary}\noindent{\em Proof.}
In the situation of the generalized statement, the right-hand
side of~(\ref{x.limcard}) is bounded by
%
\begin{quote}
$\max((2^{\aleph_0})^{\aleph_0},\,
2^{\aleph_0})^{\max(\aleph_0,\,\aleph_0,\,2)^{\aleph_0-}}
=\ (2^{\aleph_0})^{\aleph_0^{\aleph_0-}}
=\ (2^{\aleph_0})^{\aleph_0}\ =\ 2^{\aleph_0}.$
\\[-3.2em]
\end{quote}
\endproof
\vspace{.8em}
%
We can apply Theorem~\ref{T.final} to other pseudocoalgebras $S$ than
those arising from limit diagrams.
One such application gives the next result;
we omit the cardinality estimate, which the reader can easily supply.
\begin{theorem}\label{T.cofree}
If $A$ is an object of $\C,$ then there is a universal
$\!\D\!$-coalgebra object $R$ of $\C$ with a $\!\C\!$-algebra
homomorphism $|R\<|\to A\<.$
In other words, the forgetful functor
$\coalg{\C}{\D}\to\C$ has a right adjoint; equivalently, the functor
$\Rep{\C}{\D}\to\Rep{\C}{\fb{Set}}$ given by composition with the
underlying set functor on $\D$ has a left adjoint.
More generally, suppose $\D'$ is a variety whose type $\Omega_{\D'}$
is a ``subtype'' of $\Omega_\D;$ i.e., such that the operation-symbols
of $\D'$ form a subset of the operation-symbols of $\D,$
and its arity function is the restriction of that of $\D.$
\textup{(}We make no assumption on the identities of $\D'.)$
Then for any $\!\D'\!$-coalgebra $R'$ of $\C$ there exists a universal
$\!\D\!$-coalgebra object $R$ of $\C$ with a $\!\C\!$-algebra
homomorphism $|R\<|\to|R'|$ that respects $\!\D'\!$-co-operations.
\end{theorem}
\begin{proof}
Let us establish the first assertion, then note how to adapt
the argument to the second.
Given $A,$ we define a pseudocoalgebra $S$ by letting
$S\ba=A,$ and, for each $\alpha\in\Omega_\D,$
taking $S_\alpha$ to be the final object of $\C,$ with the unique maps
from $S\ba$ to that final object serving (necessarily)
both as pseudocoprojections
$c^S_{\alpha,\<\iota}$ and as pseudo-co-operations $\alpha^S.$
Then a morphism from a $\!\D\!$-coalgebra object $R$ of $\C$
to $S$ is equivalent to a $\!\C\!$-algebra homomorphism $f:|R\<|\to A.$
Thus, the final coalgebra $R$ over $S$ of Theorem~\ref{T.final}
is a coalgebra with a universal map $|R\<|\to A\<.$
Given $\D'$ and $R'$ as in the second statement,
we form the $\!\D'\!$-pseudocoalgebra $\psi(R'),$
and extend this to a $\!\D\!$-pseudocoalgebra $S$ by letting
$S_\alpha$ be the final object of $\C$
for each $\alpha\in\Omega_\D-\Omega_{\D'}.$
The pseudocoprojections and pseudo-co-operations correspondinging
to these $\alpha$ are uniquely determined, and the $R$ given
by Theorem~\ref{T.final} again has the desired universal property.
\end{proof}
The construction of the first paragraph of the above
theorem can be thought of as giving a ``cofree coalgebra'' on
each object $A$ of $\C;$ equivalently a ``free'' representable
$\!\D\!$-valued functor $G$ on each representable set-valued
functor $E$ on $\C.$
The latter is not, of course, the composite of $E$ with the
free $\!\D\!$-algebra functor $F:\fb{Set}\to\D,$ which is
in general not representable.
For instance, for $\C=\Gp,$ $\D=\fb{Ab},$ and $E$ the
underlying set functor $U_\mathbf{Group},$
the representable functor $G:\Gp\to\fb{Ab}$
``free'' on $E$ is the trivial functor, since
there are no nontrivial representable functors $\Gp\to\fb{Ab},$
as follows from the description of all representable
functors $\Gp\to\Gp$ that will be recalled in \S\ref{S.Group} below.
In contrast, we will see in that section that for $\C=\D=\Gp$
or $\C=\D=\fb{Monoid},$ the free representable
$\!\D\!$-valued functor on any nontrivial representable set-valued
functor is nontrivial, and easy to describe.
A generalization of the first paragraph of Theorem~\ref{T.cofree}
that naturally suggests itself is that for every representable
functor $F:\D\to\fb{E}$ between varieties, the induced functor
$F\circsm-:\Rep{\C}{\D}\to\Rep{\C}{\fb{E}}$ should have a left adjoint.
This seems likely to be true: In the case where $F$ is
underlying-set-preserving, one can prove it from the present second
assertion of the theorem, by replacing $\D$ with an equivalent
variety, having the original operations of $\D,$ plus some additional
ones corresponding to those derived operations of $\D$ that yield the
operations of $\fb{E},$ so that $F$ becomes a forgetful functor.
However, this generalization of the first
assertion of the theorem, if true, would still not subsume the second
assertion, since the latter does not assume that the forgetful functor
$\D\to\Omega_{\D'}\!$-$\!\fb{Alg}$ has values in $\D'.$
The question of the ``right'' generalization of
Theorem~\ref{T.cofree} obviously merits investigation.
We shall, however, devote the rest of this note to exploring the
constructions described by our theorems as stated.
\section{Some examples with $\C=\fb{Set}.$}\label{S.Set}
The proofs in the preceding section were quite uninformative
as to the structures of the limit coalgebras obtained.
To investigate these, let us use the heuristic of
\cite[\S3.2]{245}: To construct an object with a right universal
property, consider an element $x$ of a (not necessarily universal)
object $A$ of the desired sort, ask what data one can describe
that such an element will determine, and what restrictions one can find
that these data must satisfy.
Then see whether the set of all possible values for data satisfying
those restrictions can be given a structure of object of the
indicated sort, induced by that of the assumed object $A.$
If so, the result should be the desired universal object.
We will here apply this approach to the construction of the final object
of $\coalg{\fb{Set}}{\Bi},$ where $\Bi$ is the category of sets with a
single binary operation $\beta,$ not subject to any identities.
Let $S$ be the final
$\!\Bi\!$-pseudocoalgebra in $\fb{Set},$ with $S\ba$ and
$S_\beta$ both $\!1\!$-element sets.
We want to find a right-universal coalgebra
over $S,$ so our heuristic says we should think about
a general $\!\Bi\!$-coalgebra $R$ in $\fb{Set}$ over $S,$
and an element $x\in|R\<|.$
(Since $\C$ is here $\fb{Set},$ what we would otherwise write
$||R\<||$ is simply $|R\<|.)$
Given $x,$ the first datum that we get from
it is its image in $S\ba\<;$ but this tells us nothing.
The structure of $\!\Bi\!$-coalgebra on $R$ also gives a map
$|R\<|\to|R\<|\cP|R\<|.$
Combining this with our ``useless'' map
$|R\<|\to S\ba,$ we get a map
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.01}
$|R\<|\ \to\ |R\<|\cP|R\<|\ \to\ S\ba\cP S\ba\<.$
\end{minipage}\end{equation}
%
Identifying the right-hand side with $\{0,1\},$ we thus see that
we can associate to every element of $|R\<|$ a member of the latter set.
Now that we have this map $|R\<|\to\{0,1\},$ we can
combine it in turn with the co-operation, getting a composite map
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.01^2}
$|R\<|\ \to\ |R\<|\cP|R\<|\ \to\ \{0,1\}\cP\{0,1\}\ \cong\ %
\{0,1\}\times\{0,1\}\ =\ \{0,1\}^2.$
\end{minipage}\end{equation}
%
Here, in the ``$\!\cong\!$'' step, we are using the
identification $X\cP X\cong\{0,1\}\times X$ for any set $X.$
Thus, the {\em first} coordinate of the image of an element
$x\in|R\<|$ under~(\ref{x.01^2}) specifies which copy of $|R\<|$
it is mapped into by $\beta:|R\<|\to|R\<|\cP|R\<|.$
This was already determined by the image of $x$ under~(\ref{x.01});
so the image of $x$ under~(\ref{x.01^2}) is gotten by bringing in a
second coordinate, specifying which ``side'' of the indicated copy
of $|R\<|$ the element $x$ maps to.
Repeating this process, we find at the next step that each
element of $|R\<|$ determines an element of $\{0,1\}^3$ extending its
image under~(\ref{x.01^2}); and so on.
Thus, after countably many steps, we get a map
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.01^w}
$|R\<|\ \to\ \{0,1\}^\omega.$
\end{minipage}\end{equation}
There is no evident way to get more information out of an
element of $|R\<|,$ since its image under
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.01^w+}
$|R\<|\ \to\ |R\<|\cP|R\<|\ \to\ %
\{0,1\}^\omega\cP\{0,1\}^\omega\ \cong\ \{0,1\}\times\{0,1\}^\omega$
\end{minipage}\end{equation}
%
is determined by its image in $\{0,1\}^\omega:$ an element
which maps to $(a_0,a_1,a_2,\dots)$ under~(\ref{x.01^w}) can be seen
to map to $(a_0,(a_1,a_2,\dots))$ under~(\ref{x.01^w+}).
This translates to say that~(\ref{x.01^w}) is a morphism
of coalgebras from $R$ to the coalgebra with underlying
$\!\fb{Set}\!$-object $\{0,1\}^\omega,$ and co-operation
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.01coop}
$(a_0,a_1,a_2,\dots)\ \mapsto\ (a_0,(a_1,a_2,\dots))\in
\{0,1\}\times\{0,1\}^\omega
\ \cong\ \{0,1\}^\omega\,\cP\,\{0,1\}^\omega.$
\end{minipage}\end{equation}
%
It is straightforward to verify that
the map~(\ref{x.01^w}) we have built up is the unique
morphism from $R$ to this coalgebra; so this coalgebra structure
on $\{0,1\}^\omega,$
the Cantor set, makes it the final object of $\coalg{\fb{Set}}{\Bi}.$
(This shows, by the way, that the cardinality bound of
Corollary~\ref{C.c} is best possible, even in the
case of the first paragraph thereof!
The proof of that result shows that $R$ is, however, a union of strongly
quasifinal subcoalgebras of cardinality $\leq 2^{\aleph_0-}=\aleph_0;$
the reader might find it interesting to identify these.)
The representable functor $\fb{Set}\to\Bi$
determined by the above final coalgebra carries every
set $A$ to the set of all $\!A\!$-valued functions (with no continuity
condition) on the Cantor set, furnished with
the binary operation that takes
two such functions $f$ and $g$ to the function $\beta(f,g)$ which can
be described -- using the standard geometric picture of the Cantor
set as a subset of the unit interval -- as gotten by drawing the graphs
of $f$ and $g$ on the real line, with a unit distance between them,
then ``compressing'' the abscissa by a factor of $1/3,$
so that the resulting function again has domain the Cantor set.
The heuristic cited above said that after finding
data that an element $x$ of a coalgebra would determine,
we should look for restrictions those data would have to satisfy.
Above, we found no such restrictions, because the
variety $\C$ had trivial algebra structure, the
variety $\Bi$ involved no identities, and
the pseudocoalgebra $S$ was trivial.
The above example would be enough to motivate the construction
of the next section; but having begun to look at representable
functors from $\fb{Set}$ to algebras with one binary operation,
let us see what happens when identities are imposed on that operation.
Because of the paucity of algebraic structure on $\fb{Set},$
many natural identities can be realized by representable functors
only in trivial ways.
For instance, for the commutative identity
%
\begin{quote}
$\beta(a,a')\ =\ \beta(a',a),$
\end{quote}
%
to be cosatisfied by $S$ means that if we map
any $x\in|S\<|$ by $\beta^S$ to $|S\<|\cP|S\<|,$ and then
apply to its image the map that interchanges the two copies of $|S\<|$
in that coproduct, the image is unchanged.
No element $x$ can have this property;
hence $|S\<|$ must be the empty set;
so the only representable functor from $\fb{Set}$ to algebras
with a commutative binary operation is the one taking
all sets to the $\!1\!$-element algebra.
An example of an identity of a less restrictive sort is
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.01*}
$\beta(a,\beta(a',a''))\ =\ \beta(a,\beta(a',a''')),$
\end{minipage}\end{equation}
%
saying that $\beta(-,\beta(-,-))$ is independent of its last argument.
Letting $\D$ be the subvariety of $\Bi$ determined by~(\ref{x.01*}),
we find, as before, that the data one can get from an element
$x$ of any object $S$ of $\coalg{\fb{Set}}{\D}$
is a string of $\!1\!$'s and $\!0\!$'s, and that the
co-operation on these is given by~(\ref{x.01coop}).
However, condition~(\ref{x.01*}) says that if the image of $x$
under $\beta^S$ lies in the second copy of $|S\<|,$
then on applying $\beta^S$ again to that copy, the new image cannot
also lie in the second copy of $|S\<|;$ for its image therein
would have to have equal images under
any pair of maps $a'',\ a'''$ from $|S\<|$ into any set.
This says that the image of $x\in|S\<|$ under~(\ref{x.01^2})
cannot be the string $11.$
Moreover,~(\ref{x.01*}) applies to occurrences of $\beta(-,\beta(-,-))$
within larger expressions; hence the image of $x$
under~(\ref{x.01^w}) cannot have two successive $\!1\!$'s anywhere.
The set $|R'|$ of sequences satisfying this condition forms
the underlying set of the final object $R'$ of $\coalg{\fb{Set}}{\D};$
it constitutes a closed subset of the Cantor set which is uncountable,
but has measure zero under the natural probability measure on that set.
(That this subset has measure zero follows from the fact that
the number of finite strings of length $n$ with no two
successive $\!1\!$'s is the Fibonacci number $f_{n+2},$ and
that $f_{n+2}/2^n\to 0$ as $n\to\infty.)$
Geometrically, one obtains $|R'|$ from the full Cantor set by deleting
the right-hand half of the right-hand half-Cantor-set, and doing the
same recursively to all the natural copies of the Cantor set within
itself.
(What I am calling a ``half-Cantor-set'' is the part of the Cantor set
within one of the non-deleted ``thirds'' in its geometric construction.)
Given any set $A,$ the value of our representable functor
$\fb{Set}\to\D$ at $A$ consists of all $\!A\!$-valued functions on
$|R'|.$
To compose two such functions $f$ and $g,$ one again draws
their graphs with unit space between them,
but this time, one strikes off the right-hand side of the
graph of $g;$ only then will compressing the abscissa give a
function $\beta(f,g)$ on $|R'|.$
It is not hard to see directly that this operation $\beta$
satisfies~(\ref{x.01*}).
Let us look next at the subvariety $\D$ of $\Bi$ determined
by the identity superficially similar to~(\ref{x.01*}),
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.0*3}
$\beta(a,\beta(a',a'''))\ =\ \beta(a,\beta(a'',a''')),$
\end{minipage}\end{equation}
%
saying that $\beta(-,\beta(-,-))$ is independent of its middle argument.
For an object $S$ of $\coalg{\fb{Set}}{\D},$ this
translates to say that the image of any $x\in|S\<|$ under~(\ref{x.01^w})
can contain no $1$ followed immediately by a $0.$
Thus, the only strings that can occur are those consisting
of a (possibly empty, possibly infinite) string of $\!0\!$'s
followed, if it is finite, by an infinite string of $\!1\!$'s.
This set is still infinite, but is now countable.
Writing $x_i$ $(i\in\omega)$ for the string consisting of
$i$ $\!0\!$'s followed by an infinite string of $\!1\!$'s, and
$x_{\omega}$ for the infinite string of $\!0\!$'s, we conclude that
the final object of $\coalg{\fb{Set}}{\D}$ has underlying set
$\{x_0,\->}
$\dots\to\{0,1\}^n\to\dots\to\{0,1\}^2\to\{0,1\}^1\to\{0,1\}^0\<.$
\end{minipage}\end{equation}
For none of the above sets were we given a map
$\{0,1\}^n\to\{0,1\}^n\cP\{0,1\}^n.$
Rather, for each of them we had a map
$\{0,1\}^n\to\{0,1\}^{n-1}\cP\{0,1\}^{n-1}.$
We shall abstract this sort of structure in the next definition.
I have so far avoided choosing a notation
for the coprojection maps from an object to a copower
of itself, though I have written the $\!\iota\!$-th
{\em pseudo\/}coprojection map of a pseudocoalgebra $S$ as
$c^S_{\alpha,\<\iota}:S\ba\to S_\alpha.$
Below, we will need to write down formulas relating
pseudocoprojections and genuine coprojections;
so given a cardinal $\kappa,$ an object $A$ of a category $\fb{A}$
having a $\!\kappa\!$-fold copower, and an $\iota\in\kappa,$ let us
write the $\!\iota\!$-th coprojection map of $A$ into that copower as
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.coproj}
$q^A_{\kappa,\<\iota}:\ A\ \to\ \coprod_\kappa A\<.$
\end{minipage}\end{equation}
In the next definition, I use Latin rather than Greek letters
for elements of the ordinal $\theta$ because
in our main application, $\theta$ will be $\omega\<{+}1,$
and most of what we do will concern its finite elements.
Also, on encountering expressions such as
$\coprod_{\ari(\alpha)}\,(P_j)\ba,$ the reader should
recall that, as indicated in \S\ref{S.intro}, this denotes the
$\!\ari(\alpha)\!$-fold copower of the single object $(P_j)\ba,$
not the coproduct of a family of different objects.
\begin{definition}\label{D.pre}
Let $\theta$ be an ordinal and $\fb{A}$ a
category with small coproducts.
By a $\!\theta\!$-indexed {\em precoalgebra} $P$ of type $\Omega_\D$ in
$\fb{A}$ we will mean an inverse system of
$\!\D\!$-pseudo\-coalgebras
$(P_k)_{k\in\theta}$ with connecting maps
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.pjk}
$p_{j,\1zeroary} for
particular varieties, it will be simplest if one only needs to study
cosatisfaction of a fixed family of defining identities.
\vspace{.5em}
Back to the task at hand.
Let $S$ be a $\!\D\!$-pseudocoalgebra in $\C.$
Then for any ordinal $\theta>0$ there is a natural recursive
construction of a $\!\theta\!$-indexed $\!\D\!$-precoalgebra $P$
having $P_0=S.$
Namely, suppose $00$
be an ordinal, and let $P$ be the $\!\theta\!$-indexed
$\!\D\!$-precoalgebra with $P_0=S$ constructed recursively
as described above.
Then $P$ is final among $\!\theta\!$-indexed
$\!\D\!$-precoalgebras $P'$ with $P'_0=S,$ or more
generally, given with a morphism $P'_0\to S.$
In particular, if $R$ is a genuine $\!\D\!$-coalgebra with a morphism
to $S,$ then the precoalgebra formed by taking $\theta$ copies of
$R$ \textup{(}regarded as pseudocoalgebras via the
functor $\psi),$ and connecting them by identity morphisms, admits
a unique map to $P$ extending the given map to $S=P_0.$\endproof
\end{proposition}
Let me sketch concretely how the morphism of the last
paragraph above is obtained.
Given $k>0,$ suppose that for all $jco}
Let us now see how to get coalgebras from the universal
precoalgebras $P$ constructed in~Proposition~\ref{P.P_is_final}.
A nice situation is if the construction of the $\!P_k\!$
eventually ``stabilizes'', in the sense that after some point, as one
adds new steps, the connecting morphisms
$p_{j,\$\cong\ \coprod_{\ari(\alpha)}\,
\limit_{k\in\omega}(P_k)\ba\ =\ %
\coprod_{\ari(\alpha)}(P_\omega)\ba\,.$
\end{tabbing}
\end{quote}
One finds that the pseudocoprojection maps
$(P_\omega)\ba\to (P_\omega)_\alpha\cong
\coprod_{\ari(\alpha)}(P_\omega)\ba$ are the genuine coprojection maps,
making $P_\omega$ a coalgebra.
As discussed in the lines following~(\ref{x.Pkbaselim}), the fact that
$\D$ is finitary and $(P_k)_{k\in\omega}$ cosatisfies the defining
identities of $\D$ as a precoalgebra implies that
$(P_k)_{k\in\omega+1}$ also cosatisfies them as a precoalgebra;
and one deduces from this that $P_\omega$
cosatisfies them as a coalgebra, i.e., is a $\!\D\!$-coalgebra.
The final assertion is easily seen from Proposition~\ref{P.P_is_final}.
\endproof
The next proposition establishes a sufficient condition
for~(\ref{x.limvscP}) (and another property we will
subsequently find useful) to hold in a variety $\C.$
Let me illustrate the technical-looking hypothesis by contrasting the
properties of coproducts in the two varieties $\Se$ and $\fb{Monoid}.$
Given any semigroup $A,$ consider the natural homomorphism from
$A\cP A$ to the coproduct $\{x\}\cP\{x\}$ of two copies
of the final (one-element) semigroup $\{x\}.$
The two generators of $\{x\}\cP\{x\}$ are
$q^{\{x\}}_{\<2,\<0}(x)$ and $q^{\{x\}}_{\<2,\<1}(x);$
and we see from the normal form for coproducts of
semigroups that if an element $u$ of $A\cP A$ maps to, say,
$q^{\{x\}}_{\<2,\<0}(x)\ q^{\{x\}}_{\<2,\<1}(x)\ %
q^{\{x\}}_{\<2,\<0}(x)$ in $\{x\}\cP\{x\},$ then it must have the form
$q^A_{\<2,\<0}(y_1)\ q^A_{\<2,\<1}(y_2)\ q^A_{\<2,\<0}(y_3)$
in $A\cP A,$ for unique $y_1,y_2,y_3\in|A\<|.$
On the other hand, in the variety $\fb{Monoid},$ the coproduct of
two copies of the final monoid $\{e\}$ is again $\{e\},$ so
every element of a monoid $A\cP A$ maps to the unique
element thereof, so knowing that an element $u\in|A\cP A\<|$
maps to $e$ does not yield a fixed monoid word that we can
say represents $u$ in terms of elements
$q^A_{\<2,\<0}(y_i)$ and $q^A_{\<2,\<1}(y_j).$
We now state for a general variety $\C$ the condition we have
shown that $\Se$ has, but $\fb{Monoid}$ does not have,
letting $\{x\}$ now denote the final $\!(1\!$-element) algebra of $\C:$
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.u=}
For every $u\in|\{x\}\cP\{x\}|$ there exists
a derived operation $s_u$ of $\C$ such that
\begin{quote}
$u\ =\ s_u(q^{\{x\}}_{\<2,\<0}(x),\,\dots,\,q^{\{x\}}_{\<2,\<0}(x),\ %
q^{\{x\}}_{\<2,\<1}(x),\,\dots,\,q^{\{x\}}_{\<2,\<1}(x)),$
\end{quote}
%
(say with $n_0$ arguments $q^{\{x\}}_{\<2,\<0}(x)$ and
$n_1$ arguments $q^{\{x\}}_{\<2,\<1}(x)),$ and
such that for every $\!\C\!$-algebra $A,$ and every element
$b\in|A\cP A\<|$ which is carried to $u$ by the map
$A\cP A\to\{x\}\cP\{x\}$ induced by the unique map $A\to\{x\},$ there
exist {\em unique} elements $b_1,\dots,b_{n_0+n_1}\in|A\<|$ such that
%
\begin{quote}
$b\ =\ s_u(q^{A}_{\<2,\<0}(b_1),\,\dots,\,q^{A}_{\<2,\<0}(b_{n_0}),\ %
q^{A}_{\<2,\<1}(b_{n_0+1}),\,\dots,\,q^{A}_{\<2,\<1}(b_{n_0+n_1})).$
\end{quote}
\end{minipage}\end{equation}
The next result shows that condition~(\ref{x.u=}) is useful,
but restrictive.
\begin{proposition}\label{P.red}
Let $\C$ be a finitary variety satisfying~\textup{(\ref{x.u=})}.
Then it satisfies the two conditions
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.arblimvscP}
For every functor $F:\fb{E}\to\C$
where $\fb{E}$ is a {\em connected} small category, and every
cardinal $\kappa,$ the comparison morphism
$\coprod_\kappa\,\limit_\fb{E}\,F(E)\to\,
\limit_\fb{E}\ \coprod_\kappa\0$ an ordinal,
and $P$ the final $\!\theta\!$-indexed $\!\D\!$-precoalgebra with
$P_0=S,$ constructed as in Proposition~\ref{P.P_is_final}.
Then\\[.5em]
%
\textup{(i)}\ \ If $\pi^S$ is one-to-one, then
$\pi^{P_k}$ is one-to-one for every $k\in\theta.$\\[.5em]
%
\textup{(ii)}\ \ If $\pi^S$ is surjective, and
$\D$ has no identities \textup{(}i.e., is the
variety of all $\!\Omega_\D\!$-algebras\textup{)},
then $\pi^{P_k}$ is surjective for all finite $k\in\theta.$\\[.5em]
%
\textup{(iii)}\ \ If $\pi^S$ is bijective \textup{(}e.g., if $S$ is the
trivial pseudocoalgebra\textup{)} and $\D$ has no identities,
then $\pi^{P_k}$ is bijective for every $k\in\theta.$
In this last situation, if we use the bijections $\pi^{P_k}$ to
identify each $(P_k)\ba$ with $\prod_{\alpha\in\Omega_\D} (P_k)_\alpha,$
then the homomorphisms $(p_{j,\0,$ the natural map
$(P_k)_\alpha\to\limit_{j->}
$\dots\ \to\ \{x_{00}\}\cP\{x_{01}\}\cP\{x_{10}\}\cP\{x_{11}\}\ \to\ %
\{x_0\}\cP\{x_1\}\ \to\ \{x\},$
\end{minipage}\end{equation}
%
where the morphisms act by
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.droplast}
$x_{a\,i}\ \mapsto\ x_a$ $(a\in\{0,1\}^j,\ i\in\{0,1\}).$
\end{minipage}\end{equation}
Back, now, to the case $\C=\Bi.$
I claim that in this case one can encode elements
of $(P_1)\ba=\{x_0\}\cP\{x_1\}$ by certain $\!\{0,1\}\!$-valued
functions on the Cantor set.
We shall formally regard the Cantor set as
$\{0,1\}^\omega,$ but for convenience in description, I will often
use language corresponding to the ``middle thirds in the unit
interval'' Cantor set, referring, for instance,
to the ``left'' and ``right'' halves of the set.
(I hope this will be more helpful than confusing.)
Here is the encoding:
The elements $x_0$ and $x_1$ will be represented by the constant
functions $0$ and $1$ respectively.
If two elements $s$ and $t$ are represented by functions $f_s$ and
$f_t$ respectively, let us represent $\beta(s,t)$ by the
function whose graph on the ``left half'' of the Cantor set is
a compressed copy of the graph of $f_s,$ and on the ``right half'',
a compressed copy of the graph of $f_t.$
In terms of strings of $\!0\!$'s and $\!1\!$'s (which we will write
without parentheses or commas), this says that
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.beta}
$f_{\beta(s,\->}),~(\ref{x.droplast})
are seen to correspond to the operation of sending
$\!\{0,1\}^{j+1}\!$-valued functions to $\!\{0,1\}^j\!$-valued
functions by dropping the final $0$ or $1$ in the output-symbols.
It is not hard to deduce that on passing to the inverse limit, we get a
bijection between elements of
$(P_\omega)\ba=\limit_{k\in\omega} (P_k)\ba$ and
functions $\{0,1\}^\omega\to\{0,1\}^\omega$ continuous with respect
to the standard topology on both copies of $\{0,1\}^\omega.$
The $\!\Bi\!$-operation on this set is again described
by~(\ref{x.beta}).
Turning to co-operations, we have noted that by
Proposition~\ref{P.dropbase}(iii), for each $j\in\omega,$ the
pseudo-co-operation
$\beta^{P_{j+1}}: (P_{j+1})\ba\to (P_{j+1})_\beta=(P_j)\ba\cP(P_j)\ba$
is an isomorphism, taking each generator of the
form $x_{0a}$ to $q^{(P_j)\ba}_{2,\<0}(x_a),$ and each
generator of the form $x_{1a}$ to $q^{(P_j)\ba}_{2,\<1}(x_a).$
Identifying $(P_{j+1})_\beta$ with $(P_{j+1})\ba$ via this
map, we see that the image of the first pseudocoprojection,
$c^{P_{j+1}}_{\beta,0}:(P_{j+1})\ba\to (P_{j+1})_\beta,$ is
the subalgebra generated by the elements of the form $x_{0a}$
$(a\in\{0,1\}^j),$ and that of the second pseudocoprojection,
$c^{P_{j+1}}_{\beta,1},$
is the subalgebra generated by the elements $x_{1a}.$
Passing to the inverse limit, and using the ``middle thirds''
image of the Cantor set, we see that $q^{P_\omega}_{2,\<0}$ acts
by {\em vertically} compressing the graph of an element of $P_\omega$ so
as to get a function taking values in the lower half of the
Cantor set, while $q^{P_\omega}_{\<2,1}$ similarly compresses it to
a function with values in the upper half.
Regarding $P_\omega$ as a coalgebra $R,$ the co-operation
$\beta^R: |R\<|\to|R\<|\cP|R\<|$ thus works by ``breaking up''
the graph of a general element $y$ of $R$ into segments
in which the value is in the lower half-Cantor-set and segments
in which it is in the upper half-Cantor-set (each segment having for
domain a ``sub-Cantor-set'' consisting of all elements beginning
with some specified string of $\!0\!$-s and $\!1\!$-s), and using
this decomposition to write $y$ as a $\!\beta_{|R\<|}\!$-word in the
corresponding elements of $q^{P_\omega}_{\<2,\<0}(|R\<|)$ and
$q^{P_\omega}_{\<2,1}(|R\<|).$
This completes the description of the universal object $R$!
(That description was sketched, without using the language of
coalgebras, as \cite[last two exercises of \S8.3]{245}.)
\vspace{.5em}
Though the set $||R\<||$ has a more complicated form than
the underlying set of the initial object of $\coalg{\fb{Set}}{\Bi}$
found in \S\ref{S.Set}, note that its cardinality is still that
of the continuum; for a continuous map $\{0,1\}^\omega\to\{0,1\}^\omega$
is determined by its values at a countable dense subset of its domain,
so there are at most $(2^{\aleph_0})^{\aleph_0}=2^{\aleph_0}$ such maps.
A curious property of this example is that not only does the
co-operation $\beta^R$ give an isomorphism of $|R\<|$ with
the coproduct in $\Bi$ of two copies of itself,
as required by Proposition~\ref{P.dropbase}; the operation
$\beta_{|R\<|}$ likewise gives a bijection of $||R\<||$ with
the product in $\fb{Set}$ of two copies of itself.
I do not know whether this is an instance of some general result.
The set $||R\<||$ of continuous maps from the Cantor set to itself
also has a natural monoid structure, given by composition of maps.
I likewise do not know whether this has any interpretation as
``additional structure'' on the coalgebra $R$ we have constructed.
\section{The final object of $\coalg{\Bi}{\Se}.$}\label{S.BiSe}
When we turn to co-$\!\Se\!$ objects of $\Bi,$ we cannot apply
Proposition~\ref{P.dropbase}(iii) directly, since the variety
$\Se$ is defined using a nonempty set of identities; but we
can take the above description of the final $\!\Bi\!$-precoalgebra
and $\!\Bi\!$-coalgebra,
and use Proposition~\ref{P.imagecoalg}(iv) and~(ii)
to obtain the corresponding
final $\!\Se\!$-precoalgebra and coalgebra as substructures thereof.
Let us write $P'$ for the precoalgebra we called $P$ in the preceding
section, and let $P$ here
denote the subprecoalgebra thereof determined by
the condition that $\beta$ cosatisfy the associative identity.
The first instance of that identity arises in $P'_2.$
In constructing the two derived pseudo-co-operation maps that we must
equalize, we in each case begin
by mapping $(P'_2\<)\ba$ into $(P'_1\<)\ba\cP(P'_1\<)\ba$ by
$\beta^{P'_2},$ and then map this, in different ways, into
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.lmr}
$(P'_0\<)\ba\,\cP\,(P'_0\<)\ba\,\cP\,(P'_0\<)\ba\<.$
\end{minipage}\end{equation}
%
Namely, to get one derived pseudo-co-operation,
we map our first $(P'_1\<)\ba$
by $\beta^{P'_1}$ into $(P'_0\<)\ba\cP(P'_0\<)\ba,$
which we identify with the subalgebra of~(\ref{x.lmr})
generated by the first two copies of~$(P'_0\<)\ba,$
while mapping the second $(P'_1\<)\ba$ by
$(p_{0,1})\ba$ into $(P'_0\<)\ba$ which we identify with the third copy.
To get the other derived pseudo-co-operation,
we apply $(p_{0,1})\ba$ to our
first $(P'_1\<)\ba,$ and $\beta^{P'_1}$ to the second,
identifying its codomain with the subalgebra of~(\ref{x.lmr})
generated by the second and third copies.
Denoting the images of the unique element $x$ of $(P'_0\<)\ba$
under the three coprojections into~(\ref{x.lmr})
by $x_\lambda,\,x_\mu,\,x_\rho$ (mnemonic for the {\em left,
middle} and {\em right} terms in the associative law), we find that
our two derived pseudo-co-operations act by
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.lmr_r}
$x_{00}\ \mapsto\ x_\lambda,\qquad
x_{01}\ \mapsto\ x_\mu,\qquad
x_{10}\ \mapsto\ x_\rho,\qquad
x_{11}\ \mapsto\ x_\rho,$
\end{minipage}\end{equation}
%
and
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.lmr_l}
$x_{00}\ \mapsto\ x_\lambda,\qquad
x_{01}\ \mapsto\ x_\lambda,\qquad
x_{10}\ \mapsto\ x_\mu,\qquad
x_{11}\ \mapsto\ x_\rho.$
\end{minipage}\end{equation}
%
respectively.
Note that the only generators of $(P'_2\<)\ba$ on which these
two maps agree are $x_{00}$ and $x_{11}.$
Now the object of $\Bi$ freely generated by any set $X$ of
idempotent elements has a normal form, consisting of all
expressions in these generators with no subexpression which can
be simplified by the idempotence relations, i.e., no
$\beta(x,x)$ with $x\in X.$
Comparing normal forms in the domains and codomains of the
above two maps, it is not hard to
verify that their equalizer will be the subcoalgebra
of $(P'_2\<)\ba$ generated by $x_{00}$ and $x_{11}.$
(Perhaps the easiest way to see this is to regard members of
$P_2',$ as in the preceding section, as $\!\{00,\<01,10,11\}\!$-valued
functions on the Cantor set, and~(\ref{x.lmr_r}),~(\ref{x.lmr_l})
as determining maps from these to
$\!\{\lambda,\<\mu,\rho\}\!$-valued functions on that set.)
Hence $(P_2)\ba$ is the subalgebra $\langle x_{00},\,x_{11}\rangle$
of $(P'_2\<)\ba\<.$
Instances of coassociativity at higher levels similarly give
the condition that for any generator $x_{i_0\dots i_k}$
occurring in the expression for any element of $(P_k)\ba,$
successive indices $i_j,\,i_{j+1}$ must be equal, so
$(P'_j\<)\ba$ is just $\langle x_{0\dots 0},\,x_{1\dots 1}\rangle.$
The limit object is therefore
$\langle x_{0^\infty}, x_{1^\infty}\rangle;$
its co-operation takes $x_{0^\infty}$
to its own image under the first coprojection, and
$x_{1^\infty}$ to its own image under the second coprojection.
The corresponding initial object of $\Rep{\Bi}{\Se}$
can be described as taking each binar, $A,$ to the set of pairs
$\{(a,b)\in|A\<|\times|A\<|\mid a^2{\<=\<}a,\, b^2{\<=\<}b\},$
with operation $(a,b)\cdot(c,d)=(a,d).$
(Like the construction of the initial semigroup-valued
representable functor on $\fb{Set},$ this is a
``rectangular band'' construction.)
\vspace{.5em}
What if we modify the task of the preceding section in the opposite way,
and seek to describe the final object of $\coalg{\Se}{\Bi},$
equivalently, the initial representable functor $\Se\to\Bi$?
Since our new ``$\!\C\!$'', $\Se,$ like $\Bi,$ satisfies~(\ref{x.u=}),
while ``$\!\D\!$'' is again $\Bi,$ we can again say
that the representing object will have for underlying algebra the
inverse limit of the chain of idempotent-generated objects (in this
case, semigroups) written as in~(\ref{x.x->->}) and~(\ref{x.droplast}).
In the preceding section, we used
the ``rigidity'' of the normal form in $\Bi$ to identify elements
of that inverse limit with certain functions on the Cantor set.
No such model is evident when the base-category is $\Se.$
An element of $(P_1)\ba$ can be represented by a
finite alternating string of $\!0\!$'s and $\!1\!$'s;
an element of $(P_2)\ba$ mapping to such an element
is obtained by replacing each of the $\!0\!$'s with a
finite alternating string of $\!00\!$'s and $\!01\!$'s,
and each of the $\!1\!$'s with a finite alternating string
of $\!10\!$'s and $\!11\!$'s; and so on.
But I don't see any ``geometric'' description of the inverse
limit of these semigroups.
Nevertheless, the description of
each $P_k$ $(k\in\omega)$ as a semigroup
freely generated by certain idempotents will prove
useful in the next section, where we will construct the subprecoalgebra
of this $P$ that yields the final object of $\coalg{\Se}{\Se}.$
\section{The final object of $\coalg{\Se}{\Se}.$}\label{S.SeSe}
In our construction of the final object of $\coalg{\Bi}{\Se}$ above,
the pseudocoalgebras $P_j$ stabilized quickly: $P_1$ already
made all the distinctions that were going to be made among elements
of sets $(P_j)\ba,$ and once these propagated up to the pseudocopower
object at the next stage, we could have verified that $P_2$ was a
genuine coalgebra, and was our desired final object.
In constructing the final object of $\coalg{\Se}{\Se},$
our structure will also stabilize early:
All the distinctions among elements of that coalgebra will be present
in $(P_2)\ba\<;$ when we reach $(P_3)\ba$ the process of eliminating
elements at which the associative identity is not
cosatisfied will stabilize, and at the next step, the pseudocopower
object $(P_4)_\beta$ will catch up, making $P_4$ the desired coalgebra.
Let us, in this section, write $P'$ for the $\!\omega\!$-indexed
$\!\Bi\!$-precoalgebra in $\Se$ referred to at the end of the last
section, built up universally on the trivial pseudocoalgebra $P'_0;$
thus the inverse system of semigroups $(P'_k\<)\ba$
is given by~(\ref{x.x->->}) and~(\ref{x.droplast}),
with ``$\!\cP\!$'' interpreted as coproduct of semigroups.
By Proposition~\ref{P.imagecoalg}(iv), the precoalgebra we want is the
largest subprecoalgebra of $P'$ cosatisfying the associative identity.
As before, the first instance of that identity occurs in $P'_2.$
The elements of $(P'_2)\ba$ at which that identity is cosatisfied
are those at which the two homomorphisms into the semigroup freely
generated by three idempotent elements $x_\lambda, x_\mu, x_\rho$ given
by~(\ref{x.lmr_r}) and~(\ref{x.lmr_l}) agree.
However, because semigroups have a less rigid structure than
binars, the equalizer of those two homomorphisms is larger than
the subsemigroup generated by the two elements of
$\{x_{00},\1,$ then its expression includes
occurrences of both $x_{00}$ and $x_{11}.$
\end{lemma}
\begin{proof}
(i): If $r$ began with $x_{01},$ then its image
under~(\ref{x.lmr_l}) would begin with $x_\lambda,$
while its image under~(\ref{x.lmr_r}) would
begin with $x_\mu,$ a contradiction.
The same reasoning applies with ``end'' in place of ``begin'', and
the analogous computation excludes elements that begin or end
with $x_{10}.$
To get the remaining assertions, let us
first note that when either~(\ref{x.lmr_r})
or~(\ref{x.lmr_m}) is applied to $r,$ only the generator
$x_{00}$ is mapped to $x_\lambda;$ hence for any $i,$
the $\!i\!$-th occurrence of $x_{00}$ in $r$ (if such exists)
yields the $\!i\!$-th occurrence of $x_\lambda$ in the image element;
hence the initial string of $r$ up to (respectively, through) the
$\!i\!$-th occurrence of $x_{00}$ is mapped
by {\em both}~(\ref{x.lmr_r}) and~(\ref{x.lmr_m}) to
the initial string of the image of $r$ up to (respectively, through)
the $\!i\!$-th occurrence of $x_\lambda;$ and the same holds for
the terminal strings following (respectively beginning with)
the $\!i\!$-th occurrences of $x_{00}$ and $x_\lambda.$
These conclusions are not true of the images
of these substrings under~(\ref{x.lmr_l}); rather,
if we write ``(\ref{x.lmr_l}) and~(\ref{x.lmr_m})''
in place of ``(\ref{x.lmr_r}) and~(\ref{x.lmr_m})'',
we get the analogous results with $x_{11}$ and $x_\rho$
in place of $x_{00}$ and $x_\lambda.$
To see (ii) now, simply observe that if the $\!i\!$-th occurrence
of $x_{00}$ in $r$ were immediately preceded by
$x_{10},$ then in the image of $r$ under~(\ref{x.lmr_r}), the $\!i\!$-th
occurrence of $x_\lambda$ would be immediately preceded
by $x_\rho,$ while in the image under~(\ref{x.lmr_m}) it
would be immediately preceded by $x_\mu,$ a contradiction.
The statement for $x_{10}$ immediately following $x_{00},$ and
the corresponding statements for $x_{11}$ and $x_{01},$
are seen in the same way.
To see (iii), suppose that the $\!i\!$-th and
$\!i\<{+}1\!$st occurrences of $x_{00}$ have no $x_{11}$ between them.
Since they cannot be adjacent (our expression for $r$ being in normal
form), they have a nonempty string of $\!x_{01}\!$'s and $\!x_{10}\!$'s
between them.
Hence in the image of $r$ under~(\ref{x.lmr_m}), the $\!i\!$-th and
$\!i\<{+}1\!$st occurrences of $x_\lambda$ have
precisely an $x_\mu$ between them;
but the only way we can get this in the image under~(\ref{x.lmr_r})
is if they have only a single $x_{01}$ between them.
Again, the second statement is proved in the same way.
To get (iv), suppose the $\!i\!$-th occurrence of $x_{00}$ in $r$ is
immediately followed by the $\!j\!$-th occurrence of $x_{11},$
and we factor $r$ at this point, writing $r=s\,x_{00}\,x_{11}\,t.$
Applying~(\ref{x.lmr_m}), we see that the
$\!i\!$-th occurrence of $x_\lambda$ in the image element is
immediately followed by the $\!j\!$-th occurrence of $x_\rho,$
giving a factorization of that image as $u=v\,x_\lambda\,x_\rho\,w.$
Here $v\,x_\lambda,$ the image of $s\,x_{00}$ under~(\ref{x.lmr_m}),
is both the initial string of $u$ through the
$\!i\!$-th occurrence of $x_\lambda,$ and the initial
string up to (but not including) the $\!j\!$-th occurrence of $x_\rho.$
By our earlier observations, the former characterization also makes
this element the image of $s\,x_{00}$ under~(\ref{x.lmr_r}), while
the latter makes it the image of $s\,x_{00}$ under~(\ref{x.lmr_l}).
Hence those three images of $s\,x_{00}$ are equal;
so $s\,x_{00}\in|R\<|.$
The statement that $x_{11}\,t\in|R\<|$ is obtained in the same way,
as are the corresponding results when $x_{00}$ follows $x_{11}.$
To show~(v), suppose first that $r$ does not contain $x_{11}.$
Then~(i) and~(iii) imply that has the form $x_{00}(x_{01}x_{00})^i$
for some $i\geq 0.$
The images of this element under~(\ref{x.lmr_r})
and~(\ref{x.lmr_l}) are $x_\lambda(x_\mu x_\lambda)^i$
and $x_\lambda$ respectively, so $i=0,$ showing
that $r$ has length~$1.$
We get the same conclusion if $r$ does not contain $x_{00},$
completing the proof of~(v).
\end{proof}
Let us now use the above tools
to dig our way into the structure of an element $r\in|R\<|.$
By~(i) above, such an element must begin with $x_{00}$ or $x_{11}.$
Assume the former without loss of generality.
That may be all of $r,$ for it is easy to see that $x_{00}\in|R\<|.$
If it is not all of $r,$ then by~(ii), the following factor must be
$x_{11}$ or $x_{01}.$
In the former case, writing $r=x_{00}\,x_{11}\,t,$ we know by~(iv)
that each of the factors $x_{00}$ and $x_{11}\,t,$ lies in $|R\<|,$
and we are reduced to studying elements with shorter expressions.
So suppose the second factor is $x_{01}.$
This may be followed either by another $x_{00}$ or by an $x_{10}.$
In the former case, this new $x_{00}$
must again be followed by $x_{01}.$
(It can't be terminal by~(v), and it can't be followed by $x_{11},$
because if it were, then by~(iv) the product of the terms up to that
point, $x_{00}\,x_{01}\,x_{00},$ would belong to $|R\<|,$ again
contradicting~(v).)
Repeating these considerations, we see that $r$ will begin with some
string $(x_{00}\,x_{01})^{i+1}$ $(i\in\omega),$ followed by an $x_{10}.$
This $x_{10}$ cannot terminate $r$ by~(i), so it must be followed
by $x_{11}$ or $x_{01}.$
If we get $x_{01},$ then the next term can only be another $x_{10};$
the apparent alternative $x_{00}$ is ruled out by the fact
that such an $x_{00}$ and the preceding $x_{00}$ would have
$x_{01}\,x_{10}\,x_{01}$ between them, contradicting~(iii).
Repeating this argument, we get a (possibly empty)
string of alternating $\!x_{10}\!$'s and $\!x_{01}\!$'s,
finally followed by $x_{10}\,x_{11};$ i.e., $r$ begins
%
\begin{quote}
$(x_{00}\,x_{01})^{i+1}(x_{10}\,x_{01})^j(x_{10}\,x_{11})$
\quad$(i,j\in\omega).$
\end{quote}
%
(Note that if $j=0,$ the ``$\!(x_{10}\,x_{01})^j\!$'' in this
expression is not a semigroup element, but merely
means that nothing is inserted between the
$(x_{00}\,x_{01})^{i+1}$ and the $(x_{10}\,x_{11}).$
This interpretation of possibly zero exponents
will be in effect throughout this section.)
If the above $x_{11}$ is not the final term of $r,$ then it must be
followed either by $x_{00}$ (in which case we can again reduce to the
study of two shorter elements of $|R\<|$ using (iv)), or by
another $x_{10}.$
Let us collect as large a power of $x_{10}\,x_{11}$ as we can, and
note what the next two factors, if any, can be; we conclude that
if the element we have been considering cannot be factored
at an $x_{00}$-$x_{11}$ interface, then it must have the form
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.start_p}
$r\ =\ (x_{00}\,x_{01})^{i+1}(x_{10}\,x_{01})^j
(x_{10}\,x_{11})^{k+1}\,s,$ where $i,j,k\in\omega,$ and $s$ is either
the empty string, or begins with $x_{10}\,x_{01}.$
\end{minipage}\end{equation}
Let us now apply (\ref{x.lmr_r}) and (\ref{x.lmr_l})
to this description~(\ref{x.start_p}) of $r.$
The image of~(\ref{x.start_p}) under~(\ref{x.lmr_r}) begins
with $(x_\lambda\,x_\mu)^{i+1}(x_\rho\,x_\mu)^j\,x_\rho$
(where, in evaluating the image of $(x_{10}\,x_{11})^{k+1},$ we
have used the idempotence of $x_\rho),$ and the last clause
of~(\ref{x.start_p}) shows that if this is not the whole of
that image, it is followed by another $x_\mu.$
On the other hand, the image of~(\ref{x.start_p})
under~(\ref{x.lmr_l}) begins
$x_\lambda\,(x_\mu\,x_\lambda)^j(x_\mu\,x_\rho)^{k+1}$
(where in evaluating
the first part, we use the idempotence of $x_\lambda),$
possibly followed by a string beginning $x_\mu\,x_\lambda.$
Equating the powers of $x_\lambda\,x_\mu$ with which these
images begin, we see that $j=i.$
If we then compare the number of terms $x_\rho$ occurring
before each image either ends or shows another $x_\lambda,$ we see that
this is at least $j\<{+}1$ in the first case and exactly $k\<{+}1$
in the second, so $k\geq j=i;$
so the term $(x_{10}\,x_{11})^{k+1}$ of~(\ref{x.start_p})
has a left factor $(x_{10}\,x_{11})^{i+1}.$
We conclude that $r$ begins with a factor to which we give the name
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.p_i}
$p_i\ =\ (x_{00}\,x_{01})^{i+1}(x_{10}\,x_{01})^i
(x_{10}\,x_{11})^{i+1}$\quad $(i\in\omega).$
\end{minipage}\end{equation}
%
We remark that if we write the final factor $x_{01}$ of the
initial string $(x_{00}\,x_{01})^{i+1}$ in~(\ref{x.p_i}) twice
(as we may since all our generators are idempotent),
and likewise the initial factor $x_{10}$ of the
terminal string $(x_{10}\,x_{11})^{i+1},$ and combine these
with the middle string, then the expression assumes
the non-reduced, but somewhat more elegant form
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.p_i_alt}
$(x_{00}\,x_{01})^{i+1}(x_{01}\,x_{10})^{i+1}(x_{10}\,x_{11})^{i+1}.$
\end{minipage}\end{equation}
%
It is straightforward to verify that the element $p_i$ described
by~(\ref{x.p_i}), equivalently,~(\ref{x.p_i_alt}),
lies in $|R\<|:$ its image
under each of (\ref{x.lmr_r}), (\ref{x.lmr_l}) and (\ref{x.lmr_m}) is
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.p_i_sig}
$(x_\lambda\,x_\mu)^{i+1}\,(x_\mu\,x_\rho)^{i+1}$
\end{minipage}\end{equation}
%
(where again, for elegance, I am using a non-reduced
expression, with the middle occurrence of $x_\mu$ repeated).
What about the remaining factor of $r,$ if any?
If it is nonempty,
let us write $r=p_i\,s$ and apply (\ref{x.lmr_r}), (\ref{x.lmr_l})
and (\ref{x.lmr_m}), calling the images of $s$ under these
three maps $s_1,\ s_2$ and $s_3,$ respectively.
Then the fact that $p_i\,s\in|R\<|$ tells us that
%
\begin{quote}
$(x_\lambda\,x_\mu)^{i+1}\,(x_\mu\,x_\rho)^{i+1}\ s_1\ =\ %
(x_\lambda\,x_\mu)^{i+1}\,(x_\mu\,x_\rho)^{i+1}\ s_2\ =\ %
(x_\lambda\,x_\mu)^{i+1}\,(x_\mu\,x_\rho)^{i+1}\ s_3.$
\end{quote}
This does not imply that $s_1=s_2=s_3,$ since the idempotent
$x_\rho$ to the left of these three elements may mask distinctions
between them.
But it clearly implies that $x_\rho\,s_1=x_\rho\,s_2=x_\rho\,s_3,$
hence that $x_{11}\,s\in|R\<|.$
Thus, if we write $r$ as $p_i\cdot x_{11}\,s,$ using the fact that
$p_i$ ends with the idempotent element $x_{11},$ then the factors
$p_i$ and $x_{11}\,s$ lie in $|R\<|,$ and we can apply the same
considerations to the second of these, which is shorter than $r.$
We assumed above that $r$ began with $x_{00}.$
If instead it begins with $x_{11},$ then the nontrivial case
is when this is followed by something other than $x_{00},$ and in
that case, by symmetry (interchanging subscripts $0$ and $1,$
subscripts $\lambda$ and $\rho,$ and applications of
(\ref{x.lmr_r}) and (\ref{x.lmr_l}) throughout the preceding
argument), we get an initial string which we name
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.q_i}
$q_i\ =\ (x_{11}\,x_{10})^{i+1}(x_{01}\,x_{10})^i
(x_{01}\,x_{00})^{i+1}$\quad $(i\in\omega),$
\end{minipage}\end{equation}
%
which likewise belongs to $|R\<|,$ and, if it is not the whole of $r,$
can be split off so as to leave a shorter factor in $|R\<|.$
We conclude that $|R\<|$ is generated by the elements
$x_{00},\ x_{11},\ p_i$ and $q_i$ $(i\in\omega).$
It is easy to check the relations these satisfy
in $(P'_2)\ba,$ and deduce
\begin{lemma}\label{L.genrel}
The semigroup $|R\<|$ has the presentation
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.genrel}
$\langle\P}
$\coprod_{\ari(\beta)}\,(f_2)\ba\circsm\beta^S\ =
\ \beta^R\circsm(f_2)\ba\<,$
\end{minipage}\end{equation}
%
showing that $(f_2)\ba$ is a morphism of coalgebras $S\to R.$
To show that this is the only morphism of coalgebras $S\to R,$
note first that the inclusion of the subsemigroup
$|R\<|$ in $(P_2)\ba$ induces a morphism $g$ from the coalgebra
$R$ to the final $\!3\!$-indexed precobinar $(P'_k\<)_{k\in 3},$
and that this morphism separates elements of $|R\<|.$
Hence if we had two distinct
morphisms $S\to R,$ then composing these with $g,$
we would get distinct morphisms $S\to(P'_k\<)_{k\in 3},$
contradicting the universal property
of that $\!3\!$-indexed precoalgebra
(Proposition~\ref{P.P_is_final}).
Thus
\begin{theorem}\label{T.finalcosemi}
The cosemigroup $R$ with underlying semigroup~\textup{(\ref{x.genrel})}
and co-opera\-tion~\textup{(\ref{x.beta^R})}
is the final object of $\coalg{\Se}{\Se}.$\endproof
\end{theorem}
What does the functor represented by this coalgebra $R,$ i.e.,
the initial object $E$ of $\Rep{\Se}{\Se},$ look like?
Writing $a,$ $b,$ $c_i,$ $d_i$ for the images of $x_{00},$ $x_{11},$
$p_i$ and $q_i$ respectively under a semigroup homomorphism on $|R\<|,$
it can be described as taking a semigroup
$A$ to the semigroup with underlying set
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.efab}
$|E(A)|\ =\ \{\,(a;\,b;\,c_0,c_1,\dots;\,d_0,d_1,\dots\,)\in
|A\<|^{1+1+\omega+\omega}\ \mid\\[.17em]
\hspace*{7em}a^2=a,\ \ b^2=b,\ \ a\0.$
There are no representable functors that I was previously aware of
whose morphisms from the initial representable functor
involve those coordinates; but one can easily design such functors.
Our initial functor $E$ itself is one, of course.
One can also cut it down, throwing away all the relations
in~(\ref{x.efab}), and all but three of the coordinates,
getting the construction sending $A$ to the set of all
$\!3\!$-tuples $(a,b,c)$ of elements of $|A\<|,$ with the
multiplication $(a,\,b,\,c)(a',\,b',\,c')=(a,\,b',\,c\,(a'b)^i c'),$
for arbitrary fixed~$i.$
It is easy to verify that this is associative, and that the unique
morphism from $E$ to this functor uses the coordinates $a,$ $b,$
and $c_i.$
More generally, if one takes any semigroup word $w$ in $m+n$
variables, one can define a representable functor $F$ taking every
semigroup $A$ to the set of all $\!m+n+1\!$-tuples
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.m+n+1}
$(a_1,\ \dots,\ a_m;\ b_1,\ \dots,\ b_n;\ c)$
\end{minipage}\end{equation}
%
of elements of $A,$ with the operation
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.w-mult}
$(a_1,\dots,a_m;\ b_1,\dots,b_n;\ c)
\ (a'_1,\dots,a'_m;\ b'_1,\dots,b'_n;\ c')\\[.17em]
\hspace*{1em}
=\ (a_1,\,\dots,\,a_m;\ b'_1,\,\dots,\,b'_n;\ c
\ w(a'_1,\,\dots,\,a'_m,\,b_1,\,\dots,\,b_n\,)\ c'),$
\end{minipage}\end{equation}
%
which it is easy to verify is associative.
To describe the map from $E$ to this functor, break the
word $w$ into alternating blocks of occurrences of
the ``$\!a\!$'' variables (the first $m)$ and
the ``$\!b\!$'' variables (the last $n),$ and count these blocks,
starting from the first block of ``$\!a\!$'' variables
(ignoring any ``$\!b\!$'' variables that may precede them), and
ending with the last block of ``$\!b\!$'' variables
(ignoring any ``$\!a\!$'' variables that may follow that block).
Say these constitute $i$ blocks of ``$\!a\!$'' variables alternating
with $i$ block ``$\!b\!$'' variables, for some $i\in\omega.$
(This $i$ may be $0,$ namely, if $w$ consists only of a
word in the ``$\!b\!$'' variables followed by a word in the ``$\!a\!$''
variables, with one of those words possibly empty.)
Then the unique morphism $E\to F$ assigns to all $a_r$ the value
$a,$ to all $b_r$ the value $b,$ and to $c$ the value $c_i.$
We remark that~(\ref{x.w-mult}) looks more natural if we rewrite
our $\!m\<{+}\1$ derived zeroary
operations.}\label{S.>1zeroary}
We saw in Theorem~\ref{T.7/9} (last paragraph) that the only
situations where the initial object of $\Rep{\C}{\D}$ can have
nontrivial nonempty algebras among its values
are when $\C$ and $\D$ either both have no zeroary operations,
or both have more than one derived zeroary operation.
So far, we have looked at cases of the former sort.
A very elementary example of a variety with more than one derived
zeroary operation is the one
determined by a single primitive zeroary operation $\alpha_0,$
a single primitive unary operation $\alpha_1,$
and no identities; i.e., the variety of sets with a distinguished
element, and an endomap not assumed to fix that element.
(The derived zeroary operations are $\alpha_0,
\ \alpha_1(\alpha_0),\ \dots,\ \alpha_1^i(\alpha_0),\ \dots\ .)$
Taking any variety $\C,$ and the above variety as $\D,$
let $P$ be the final $\!\D\!$-precoalgebra in $\C$ having
for $P_0$ the trivial pseudocoalgebra.
By Proposition~\ref{P.dropbase}(iii),
$(P_1)\ba$ can be identified with the direct product
of the $\!0\!$-fold and $\!1\!$-fold copowers of $(P_0)\ba\<.$
The $\!1\!$-fold copower is just $(P_0)\ba,$ the trivial algebra;
hence that direct product can be identified with the $\!0\!$-fold
copower of $(P_0)\ba;$ i.e., the initial object of $\C.$
Let us denote this by $I_\C;$ we know that
it will be nonempty and nontrivial precisely if $\C$ has more
than one derived zeroary operation.
We find that passage to each subsequent level of $P$
brings in one more copy of $I_\C,$ so that $(P_k)\ba=I_\C^k.$
Taking the limit, we get
$(P_\omega)\ba=I_\C^\omega\<,$ where the pseudo-co-operation
$\alpha_0^{P_\omega}$ projects $I_\C^\omega$
to its component indexed by $0,$
while $\alpha_1^{P_\omega}$ is the ``left shift'' map.
Propositions~\ref{P.compar_iso} and~\ref{P.red} tell us that
if $\C$ is a finitary variety satisfying~\textup{(\ref{x.u=})},
then $P_\omega$ will be a coalgebra,
the final object of $\coalg{\C}{\D}.$
But in fact, those hypotheses on $\C$ are not needed here.
Their purpose was to guarantee that $\!\ari(\alpha)\!$-fold
copowers commuted with $\!\omega\!$-indexed inverse limits
for all $\alpha\in\Omega_\D,$ but this is automatic
when all elements of $\Omega_\D$ are zeroary or unary.
(The $\!1\!$-fold copower functor is the identity and
commutes with everything; the $\!0\!$-fold copower is the
constant functor giving the initial algebra, and constant functors
commute with limits over connected categories.)
Indeed, either by following the above precoalgebra approach,
or by directly checking the universal property, it is not hard to
verify the following general statement.
\begin{lemma}\label{L.01ary}
In a category $\fb{X}$ with an initial object, let such an
object be denoted~$I_\fb{X}.$
Suppose $\fb{A}$ is a category having small products and
coproducts, and $\D$ a variety with only zeroary and unary
primitive operations \textup{(}with or without identities\textup{)}.
Then $\coalg{\fb{A}}{\D}$ has a final object $R,$ with
underlying $\!\fb{A}\!$-object $|R\<|=I_\fb{A}^{|I_\D|},$ where the
\textup{(}primitive or arbitrary\textup{)}
zeroary $\!\D\!$-co-operations of $R$ are given by the projections to
the components indexed by the values in $|I_\D|$ of the corresponding
zeroary operations of $I_\D,$ and where the
\textup{(}primitive or arbitrary\textup{)}
unary $\!\D\!$-co-operations of $R$ are determined
\textup{(}contravariantly\textup{)} by the action of the corresponding
unary operations of $I_\D$ on the index-set $|I_\D|.$\endproof
\end{lemma}
For concreteness, let us return to the case where $\D$
is the variety defined by one zeroary operation,
one unary operation, and no identities.
Let $\C$ be an arbitrary variety with more than one derived zeroary
operation, and let $R$ be the final $\!\D\!$-coalgebra of $\C,$
which by the above observations
has underlying $\!\C\!$-algebra $I_\C^\omega.$
If $I_\C$ is finite or countable, then $R$ has the
cardinality of the continuum; and for many choices of $\C$ -- e.g.,
$\D$ itself, or the variety of sets with two distinguished elements,
or the variety of vector-spaces with a distinguished vector,
over a countable or finite field -- we find that the corresponding
representable functor $\C(R,-)$ takes all nontrivial finite or
countable objects of $\C$ to
objects of $\D$ having cardinality~$2^{\aleph_0^{\aleph_0}}.$
As a curious exception, suppose $\C$ is the
category of abelian groups with one distinguished element, so that
$I_\C$ is $\mathbb{Z}$ with distinguished element $1.$
Then $|R\<|$ is the group $\mathbb{Z}^\omega$
of all sequences of integers, with the
constant sequence $(1,1,\dots)$ as distinguished element.
It is known that the only group homomorphisms
$\mathbb{Z}^\omega\to\mathbb{Z}$ are the finite linear combinations
of the projection maps \cite[Proposition~94.1]{Fuchs}.
Hence, if we apply our functor $\C(R,-)$ to $\mathbb{Z},$ with any
$n\in|\mathbb{Z}|$ made the distinguished element, we get a
countable $\!\D\!$-object, the set of those
finite linear combinations of projection maps
$\mathbb{Z}^\omega\to\mathbb{Z}$ whose coefficients sum to $n$
(with distinguished element given by $n$ times the projection onto the
$\!0\!$-component, and endomap given by the {\em right} shift operator
on those strings of coefficients).
On the other hand, if we apply $\C(R,-)$ to
$\mathbb{Z}/p\mathbb{Z}$ or $\mathbb{Q}$
(with any distinguished element), we find that the resulting object
again has cardinality $2^{\aleph_0^{\aleph_0}}.$
\vspace{.5em}
A very different class of cases where we can get a nice handle on
the final object of a category $\coalg{\C}{\D}$ is when $\D$ is an
arbitrary variety, and $\C$ the variety $\fb{Boole}$ of Boolean rings
(commutative associative unital
rings satisfying the identities $2=0$ and $x^2=x).$
The variety $\fb{Boole}$ is dual to the category $\fb{Stone}$ of
Stone spaces, i.e., totally disconnected compact Hausdorff spaces,
via the functor taking every Boolean ring $A$ to its prime spectrum,
equivalently, to the set $\fb{Boole}(A,2)$ of its homomorphisms into
the $\!2\!$-element Boolean ring, under the function topology;
the inverse functor takes each Stone space $X$ to
the set $\fb{Stone}(X,2)$ of continuous $\!\{0,1\}\!$-valued
functions on $X,$ made a Boolean ring under pointwise operations,
equivalently, to the set of clopen subsets of $X,$ made a
Boolean ring in the usual way \cite[Introduction and \S II.4]{PJ.Sto}.
Hence, {\em $\!\D\!$-coalgebras} in $\fb{Boole}$ correspond
to {\em $\!\D\!$-algebra} objects in $\fb{Stone}.$
These can be described as $\!\D\!$-algebras given with totally
disconnected compact Hausdorff topologies on their underlying sets,
with respect to which the $\!\D\!$-operations are continuous.
The representable functor $\fb{Boole}\to \D$
corresponding to such a topological $\!\D\!$-algebra $X$ can
be described as taking each Boolean ring $B$ to the
$\!\D\!$-algebra of continuous $\!X\!$-valued functions on the
Stone space of $B.$
(In particular, we can recover the underlying $\!\D\!$-algebra of
$X$ from the functor by evaluating the latter on the Boolean ring $2,$
corresponding to the $\!1\!$-point Stone space.)
Any finite set with the discrete topology is a Stone space, so any
finite $\!\D\!$-algebra gives a $\!\D\!$-algebra object of $\fb{Stone}.$
For some varieties $\D,$ such as those of associative (unital
or nonunital) rings, semigroups,
monoids, groups, and distributive lattices (and, automatically,
all subvarieties of these), it is known
\cite[VI.2.6-9]{PJ.Sto}, \cite{250B_stone} that
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.profin}
The topological $\!\D\!$-algebras with Stone topologies are precisely
the inverse limits of systems of finite $\!\D\!$-algebras, with the
topologies induced by the discrete topologies on those finite algebras.
\end{minipage}\end{equation}
When this holds, every such Stone algebra is, in particular,
the inverse limit of all its finite homomorphic images;
and given an arbitrary $\!\D\!$-algebra $A,$
the category of Stone $\!\D\!$-algebras furnished with homomorphisms
of $A$ into them will have as initial object the inverse limit
of all finite homomorphic images of $A.$
Now note that the category $(A\downarrow\D)$ of
(non-topologized) $\!\D\!$-algebras with
homomorphisms of $A$ into them can be regarded as a variety $\D',$
by taking any presentation $\langle X\mid Y\rangle_\D$ for $A$
in $\D,$ and letting $\D'$ have, in addition to the operations and
identities of $\D,$ an $\!X\!$-tuple of additional zeroary
operations, subject to the relations $Y.$
So the above inverse limit of finite homomorphic images of $A$
can be regarded as the initial Stone topological $\!\D'\!$-algebra.
For example, suppose that $\D=\Gp$ and
$\langle X\mid Y\rangle_\D$ is the infinite cyclic group (where
we can take $X$ a singleton and $Y$ empty), or that
$\D=\fb{Ring}^1$ and $\langle X\mid Y\rangle_\D$ is
the initial ring $(X$ and $Y$ both empty).
In these cases, the initial Stone $\!\D'\!$-algebra will be the
inverse limit of the finite groups or rings $\mathbb{Z}/n\mathbb{Z}.$
(This inverse limit, called the ``profinite completion
of $\mathbb{Z}\!$'', is the direct product over all primes $p$
of the groups or rings of $\!p\!$-adic integers.)
In this example, our Stone algebra, and hence the value of our initial
representable functor at the Boolean ring $2,$
merely has the cardinality of
the continuum, and the final coalgebra to which it corresponds is
merely countable, in contrast to some of our earlier examples.
This is an instance of a general phenomenon.
Before formulating it, let us note that if $\D$
satisfies~(\ref{x.profin}), then that property will be inherited
by $\D'=(A\downarrow\D)$ for any $\!\D\!$-algebra $A,$ and that
if $A$ is finitely generated, then if $\D$ is describable in terms
of finitely many primitive operations, all of finite arity, these
properties will be inherited by $\D'.$
Hence it suffices to prove the next result for such a variety $\D,$
understanding that it will then be applicable to varieties
$\D'=(A\downarrow\D)$ when $A$ is a finitely generated $\!\D\!$-algebra.
\begin{lemma}\label{L.countable}
Suppose $\D$ is a finitary variety with only finitely many primitive
operations, which satisfies\textup{~(\ref{x.profin})}.
Then the final object of $\coalg{\fb{Boole}}{\D}$ is at most countable;
equivalently, the initial Stone $\!\D\!$-algebra has separable
topology.
\end{lemma}
\begin{proof}
From the fact that $\D$ is finitary with finitely many primitive
operations, one sees that up to isomorphism, there are at
most countably many finite $\!\D\!$-algebras.
Hence the initial $\!\D\!$-algebra
has at most countably many finite homomorphic images,
so the initial Stone $\!\D\!$-algebra is the inverse limit
of an at most countable inverse system of finite algebras.
Every continuous $\!\{0,1\}\!$-valued function on that inverse
limit is induced by such a function on one of these algebras; so the
Boolean ring of all such functions is at most countable.
\end{proof}
Sometimes our final coalgebras are even smaller than the
above result requires.
If one takes for $\D$ the variety of groups, and
lets $\langle X\mid Y\rangle_\D$ be an infinite group with no finite
homomorphic images, such as
%
\begin{equation}\begin{minipage}[c]{34pc}\label{x.wxyz}
$\langle\,w,x,y,z\ \mid\ wxw^{-1}=x^2,\ xyx^{-1}=y^2,\ %
yzy^{-1}=z^2,\ zwz^{-1}=w^2\,\rangle,$
\end{minipage}\end{equation}
%
(or, going outside the context of Lemma~\ref{L.countable} to
a non-finitely-generated example, but still using the fact
that $\Gp$ satisfies~(\ref{x.profin}), the
additive group of rational numbers), then the initial
Stone $\!\D'\!$-algebra is trivial, making the final object of
$\coalg{\fb{Boole}}{\D'}$ the $\!2\!$-element (initial) Boolean ring.
In this case, the derived zeroary operations of $\D'$ cannot be
distinguished by representable functors on $\fb{Boole},$ and we are
effectively reduced to case~(iii) of Theorem~\ref{T.7/9}.
In the other direction, taking varieties $\D$ not
satisfying~(\ref{x.profin}), one again
gets examples where the final object
of $\coalg{\fb{Boole}}{\D}$ has continuum cardinality.
For instance, if we let $\D$ be, as at the beginning of this section,
the variety of sets with a single zeroary and a single binary
operation, then by the discussion there, the final object of
$\coalg{\fb{Boole}}{\D}$ will be $2^\omega,$ the countable direct
power of the initial object $2=\{0,1\}$ of $\fb{Boole}.$
The corresponding Stone $\!\D\!$-algebra can be described as the
Stone-\v{C}ech compactification of $\omega,$ with the element
$0\in\omega$ as the value of the zeroary operation, and the
endomap induced by $n\mapsto n+1$ as the unary operation.
Some more familiar varieties not
satisfying~(\ref{x.profin}) are the variety
of Lie algebras over a finite field (this may be deduced from
\cite[Example~25.49, p.126]{coalg}) and the variety
of all lattices (\cite[second example on p.10]{250B_stone}).
Incidentally, one's first impression on looking at the statement of
Lemma~\ref{L.countable} might be that it should surely be possible
to weaken the hypothesis ``finitely many operations''
to ``at most countably many''; but this is not so.
If $\D$ is the variety of sets with an $\!\omega\!$-tuple of
zeroary operations, then~(\ref{x.profin}) is inherited from
the variety $\fb{Set},$ but by Lemma~\ref{L.01ary}, the final
object of $\coalg{\fb{Boole}}{\D}$ again has for underlying algebra
the uncountable Boolean algebra $2^\omega.$
Infinite compact Hausdorff spaces are ``typically''
uncountable, but not always.
Can the initial Stone object of a variety $\D$ be countably infinite?
I found this a hard one to answer; but here is an example.
Let $\D$ be the variety of lattices (or upper semilattices,
or lower semilattices) with one additional zeroary operation
$\alpha_0,$ and one additional unary operation $\alpha_1,$ subject
to the identities saying that $\alpha_1$ is {\em increasing,}
i.e., $\alpha_1(x)\wedge x=x$ and/or $\alpha_1(x)\vee x=\alpha_1(x)$
(depending on which operations are assumed present).
I leave the determination of the initial Stone topological
object of $\D$ to the reader who would like a challenging exercise.
(Note that one cannot assume that $\D$ satisfies~(\ref{x.profin}).)
If $\D$ is a variety whose initial object is finite, then even
without~(\ref{x.profin}) one can see that this finite object,
regarded as a finite Stone space, will be the initial Stone object
of $\D,$ and so will determine the initial representable functor
$\fb{Boole}\to\D.$
We find, in particular, that the initial representable
functor $\fb{Boole}\to\fb{Boole}$ is the identity.
\vspace{.5em}
There are other dualities between varieties of algebras and
structures with compact topologies; cf.\ \cite{RA+Kap}, \cite{DC+BD}.
It would be worth seeing whether these also lead to useful
results about coalgebras, equivalently, representable functors.
It would also be worth examining whether some of
the properties of varieties
$\D$ that allow one to prove~(\ref{x.profin}) might also allow
one to prove results about $\!\D\!$-coalgebras in general varieties.
In both~\S\ref{S.BiSe} and~\S\ref{S.SeSe} above, the
associative identity trimmed uncountable cobinars $P'_\omega$
down to countable cosemigroups $R=P_\omega.$
Is this a case of a more general phenomenon?
Cf.\ also \cite[\S32, first paragraph]{coalg}, where a
statement formally resembling~(\ref{x.profin})
is translated into a dual statement about ``coalgebras'',
though in one of the other senses of that word.
\vspace{.5em}
Yet another example of an initial representable functor between
varieties with more than one derived zeroary operation is the one
that motivated this investigation, as noted in \S\ref{S.motivate}:
the initial object of $\Rep{\fb{Ring}^1}{\,\fb{Ring}^1},$
given by $S\mapsto S\times S^{\mathrm{op}}$
\cite[Corollary~25.22]{coalg}.
(That result is generalized a bit in \cite[Exercise~31.12(vi)]{coalg}.)
\section{Universal constructions in $\coalg{\Gp}{\Gp}$ and\\
$\coalg{\fb{Monoid}}{\fb{Monoid}}.$}\label{S.Group}
Since the varieties of groups and of monoids have unique derived zeroary
operations (giving the identity element), Theorem~\ref{T.7/9}
tells us that the initial representable functor from either of
them to any variety, or from any variety to either of them, is trivial;
equivalently, that the corresponding final coalgebras have
initial algebras for underlying algebra.
But other limits of coalgebras
involving these varieties need not be trivial.
Let us briefly examine {\em products} in $\coalg{\Gp}{\,\Gp}.$
Kan \cite{DK.Mon} determined the structure of all comonoid
objects in $\Gp.$
These all extend uniquely to cogroup structures, so we shall describe
his result as determining the cogroups in $\Gp.$
The functors these
represent are simply the direct powers of the identity functor.
If we take an arbitrary cogroup $R$ in $\Gp,$ the precise
statement is that its underlying group $|R\<|$ is
free on the set consisting of the nonidentity elements
$x$ satisfying $\beta^R(x)=x^{(0)} x^{(1)}$ (where
$\beta\in\Omega_\Gp$ again denotes the primitive binary operation).
Now a morphism $R\to S$ of cogroups must take elements of $R$
satisfying $\beta^R(x)=x^{(0)} x^{(1)}$ to elements of $S$ with the same
property; so it must take each member of the canonical free generating
set for $|R\<|$ either to a member of the corresponding
generating set for $|S|,$ {\em or} to the identity element, $e.$
If we write $\mathrm{Id}_\Gp$ for the identity functor of
$\Gp,$ and $\mathrm{Id}_\Gp^n$ for the $\!n\!$-fold
direct power of this functor (the functor taking each group $G$
to $G^n$ -- not, of course, the $\!n\!$-fold composite
of $\mathrm{Id}_\Gp),$ then this is equivalent to saying that a morphism
$\mathrm{Id}_\Gp^m\to \mathrm{Id}_\Gp^n$ is determined
by specifying, as the value to be assigned at each of the $n$
coordinates of the codomain groups, either a specified
one of the $m$ coordinates of the domain groups, or the value $e.$
So, though one could phrase Kan's
result as saying that the isomorphism classes of cogroups
in $\Gp$ correspond to the isomorphism classes of sets, the
category $\coalg{\Gp}{\,\Gp}$
is actually equivalent, not to $\fb{Set},$ but
to the category $\fb{Set}^\mathrm{pt}$ of pointed sets,
by the functor taking each coalgebra $R$ to the set
$\{x\mid\beta^R(x)=x^{(0)} x^{(1)}\},$ with $e$ as basepoint.
The product of two pointed sets $(X,e_X)$ and $(Y,e_Y)$ is
$(X\times Y,\ (e_X,e_Y)).$
Thus, the product of the cogroups in $\Gp$ represented by
the free groups on $m$ and on $n$ generators will be free on
$(m+1)(n+1)-1$ generators (with $+1$ to count the basepoint $e$
in the description of the relevant pointed sets, and $-1$ to
discount the basepoint in the product set).
In terms of representable functors, the coproduct of
$\mathrm{Id}_\Gp^m$ and $\mathrm{Id}_\Gp^n$ in
$\Rep{\Gp}{\,\Gp}$ is
therefore $\mathrm{Id}_\Gp^{mn+m+n}.$
To describe the universal maps
$\mathrm{Id}_\Gp^m\to \mathrm{Id}_\Gp^{mn+m+n}$ and
$\mathrm{Id}_\Gp^n\to \mathrm{Id}_\Gp^{mn+m+n},$
note that for {\em any} such pair of maps,
each coordinate of $\mathrm{Id}_\Gp^{mn+m+n}$ must select either
some coordinate of $\mathrm{Id}_\Gp^m$ or the identity, and
either some coordinate of $\mathrm{Id}_\Gp^n$ or the identity.
The universal pair is such that
the $mn+m+n$ coordinates together cover all possible choices exactly
once, except for the simultaneous choice of the identity for
both coordinates.
It is interesting to observe that for $S$ and $S'$ objects of
$\coalg{\Gp}{\,\Gp}$ and $S\times S'$ their product
in this category, the pair of projection maps $S\times S'\to S$ and
$S\times S'\to S'$ does not in general separate elements of
the underlying set $||S\times S'\<||$ of the product coalgebra.
For example, let us write the underlying free group of the coalgebra
representing $\mathrm{Id}_\Gp$ as $|S\<|=\langle x\rangle,$
and the underlying free group of the product coalgebra $S\times S$ as
$|S\times S\<|= \langle g_{x,\Ab}
Let $\fb{Ring}^1$ denote the variety of associative rings with $1,$ and
$\fb{Ab}$ the variety of abelian groups, which we will write additively.
In \cite[Theorem~13.15]{coalg} it is shown, inter alia, that
$\coalg{\fb{Ring}^1}{\fb{Ab}}$ is equivalent to $\fb{Ab}.$
Precisely, if we denote by $U$ the forgetful functor taking every
ring to its underlying abelian group, then
every representable functor $\fb{Ring}^1\to\fb{Ab}$
can be obtained by following $U$ with a representable functor
$\fb{Ab}(A,-):\fb{Ab}\to\fb{Ab},$ for some abelian group $A.$
(As is well-known, every abelian group is the representing
object of a unique representable functor
$\fb{Ab}\to\fb{Ab},$ the group structure being given
by elementwise operations on group homomorphisms.)
The object of $\coalg{\fb{Ring}^1}{\fb{Ab}}$ representing
this composite is the tensor ring
$\mathbb{Z}\langle A\<\rangle,$ i.e., the result of applying to
$A$ the left adjoint to $U,$ with a
co-$\!\fb{Ab}\!$-structure that the reader can easily write down; and
these constructions yield a functorial equivalence.
Limits in $\fb{Ab}$ are easy to describe, and yield descriptions
of the corresponding limits in $\coalg{\fb{Ring}^1}{\fb{Ab}}.$
Note that the tensor ring construction does not respect
one-one-ness of maps.
For instance, let $A$ be the
abelian group $\langle a,b\mid pb=0\rangle$ for any prime $p$
(the direct sum of $\mathbb{Z}$ and $\mathbb{Z}/p\mathbb{Z}),$ let $B$
be the subgroup of $A$ generated by $pa$ and $b,$ and let $f:B\to A$ be
the inclusion map.
Then in $\mathbb{Z}\langle B\<\rangle$
the element $(pa)b$ is easily shown
to be nonzero, but its image in $\mathbb{Z}\langle A\<\rangle$
can be written $p(ab)=a(pb)=a0=0.$
Consequently, statements about $\fb{Ab}$ that refer to one-one-ness
may not go over to $\coalg{\fb{Ring}^1}{\fb{Ab}}.$
For instance, in $\fb{Ab},$ every one-to-one map is an equalizer (e.g.,
of its canonical map onto its cokernel group, and the zero map
thereto); in particular, this is true of the inclusion $A\subseteq B$
of the preceding paragraph; hence the induced map
in the equivalent category $\coalg{\fb{Ring}^1}{\fb{Ab}}$
is also an equalizer; showing that equalizers in that category
need not be one-to-one on underlying sets.
% - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
\begin{thebibliography}{00}
\bibitem{PA+NM} Peter Aczel and Nax Mendler,
{\em A final coalgebra theorem}, in
{\em Category theory and computer science} (Manchester, 1989), 357--365,
Lecture Notes in Computer Science, 389, Springer, 1989.
~MR~{\bf 91f}:18001.
\bibitem{RA+Kap} Richard F. Arens and Irving Kaplansky,
{\em Topological representations of algebras,}
Trans. A. M. S. {\bf 63} (1948) 457-481.
~MR~{\bf 10},\ p.\,7.
\bibitem{Barr_74} Michael Barr,
{\em Coalgebras over a commutative ring,}
J. Algebra {\bf 32} (1974) 600--610.
MR~{\bf 53}\#5707.
\bibitem{Barr} Michael Barr,
{\em Terminal coalgebras in well-founded set theory,}
Theoretical Computer Science {\bf 114} (1993) 299--315.
~MR~{\bf 94m}:18006a.
Additions and corrections, {\em ibid.}
{\bf 124} (1994) 189--192.
~MR~{\bf 94m}:18006b.
\bibitem{tower} George M. Bergman,
{\em Some category-theoretic ideas in algebra \textup{(}a too-brief
tour of algebraic structure, monads, and the adjoint tower}),
Proceedings of the 1974 International Congress of Mathematicians
(Vancouver) v.\,I, pp.\,285-296, Canadian Mathematical Congress, 1975.
~MR~{\bf 58}\#22222.
\bibitem{245} George M. Bergman,
{\em An Invitation to General Algebra and Universal Constructions,}
pub. Henry Helson, Berkeley, CA, 1998.
Readable online at
http://math.berkeley.\linebreak[0]edu/{$\!\sim$}gbergman/245\<.
~MR~{\bf 99h}:18001.
\bibitem{embed} George M. Bergman,
{\em Some results on embeddings of algebras, after de Bruijn and
McKenzie,} Indag. Math. {\bf 18} (2007) 349--403.
arXiv:0704.0275.
MR2373687. %(no review yet as of 15/7/08)
http://\linebreak[0]www.sciencedirect.com/science/journal/00193577\,.
\bibitem{250B_stone} George M. Bergman,
{\em Infinite Galois theory, Stone spaces, and profinite groups,}
course handout, 1997, 12\,pp.,
/{$\sim$}gbergman/grad.hndts%
/\linebreak[0]infGal$\!+\!$profin.ps\<.
\bibitem{coalg} George M. Bergman and Adam O. Hausknecht,
{\em Cogroups and Co-rings in Categories of Associative Rings,}
A. M. S. Mathematical Surveys and Monographs,
v.45, ix${+}$388\,pp., 1996.
~MR~{\bf 97k}:16001.
\bibitem{GB.Lat} Garrett Birkhoff,
{\em Lattice Theory,}
3rd edition, A. M. S. Colloquium Publications, v.\,XXV, 1967.
~MR~{\bf 37}\#2638.
\bibitem{DC+BD} David M. Clark and Brian A. Davey,
{\em Natural Dualities for the Working Algebraist,}
Cambridge University Press, 1998.
~MR~{\bf 2000d}:18001.
\bibitem{C+P}
A. H. Clifford and G. B. Preston,
{\em The algebraic theory of semigroups, v.\,I.,}
Mathematical Surveys No.~7, A. M. S., 1961.
~MR~{\bf 24}\#A2627.
\bibitem{PC.UA} P. M. Cohn,
{\em Universal Algebra,}
2nd ed., Reidel, 1981.
~MR~{\bf 82j}:08001.
\bibitem{Freyd} Peter Freyd, % QA1.C59
{\em Algebra-valued functors in general and tensor products
in particular,}
Colloquium Mathematicum (Wroc{\l}aw) {\bf 14} (1966) 89--106.
~MR~{\bf 33}\#4116.
\bibitem{Fuchs} Fuchs, L\'aszl\'o,
{\em Infinite abelian groups. Vol. II,}
Pure and Applied Mathematics. Vol. 36-II. Academic Press, 1973.
~MR~{\bf 50}\#2362.
\bibitem{PJ.Sto} Peter T. Johnstone,
{\em Stone Spaces,}
Cambridge Studies In Advanced Math., vol.\,3,
Cambridge University Press, 1982.
~MR~{\bf 85f}:54002.
\bibitem{DK.Mon} Daniel M. Kan,
{\em On monoids and their dual,}
Bolet\'in de la Sociedad Matem\'atica Mexicana
(2) {\bf 3} (1958) 52-61.
~MR~{\bf 22}\#1900.
\bibitem{CW} Saunders Mac~Lane,
{\em Categories for the Working Mathematician,}
Springer Graduate Texts in Mathematics, v.5,~1971.
~MR~{\bf 50}\#7275.
\bibitem{M+P} Michael Makkai and Robert Par\'e,
{\em Accessible Categories: the foundations of categorical model
theory,} Contemporary Mathematics, {\bf 104}, 1989, ISBN 0-8218-5111-X.
~MR~{\bf 91a}:03118.
\bibitem{MS} Moss E. Sweedler,
{\em Hopf Algebras,}
Math. Lecture Note Series, Benjamin, N.Y., 1969.
~MR~{\bf 40}\#5705.
\bibitem{TW.Rep} D. O. Tall and G. C. Wraith,
{\em Representable functors and operations on rings,}
Proc. London Math. Soc. (3) {\bf 20} (1970) 619--643. %QA1 L6
~MR~{\bf 42}\#258.
\bibitem{GW.aar} Gavin C. Wraith,
{\em Algebraic Theories,}
Lecture Note Series, vol.\,22,
Aarhus Universitet, Matematisk Institut,
Lectures Autumn 1969, revised version of notes, Feb.~1975.
(MR\,{\bf 41}\#6943 reviews the original version;
the revision has far fewer errors.
But note re p.49, line 6 from bottom: it is not true that ``...~every
monad on $A^\mathrm{b}$ arises in this way''.)
\end{thebibliography}
\end{document}