Notes on Lattice Theory

J. B. Nation
University of Hawaii
 Introduction

In the early 1890’s, Richard Dedekind was working on a revised and enlarged
edition of Dirichlet’s Vorlesungen über Zahlentheorie, and asked himself the following
question: Given three subgroups A, B, C of an abelian group G, how many different
subgroups can you get by taking intersections and sums, e.g., A + B, (A + B) ∩ C,
etc. The answer, as we shall see, is 28 (Chapter 7). In looking at this and related
questions, Dedekind was led to develop the basic theory of lattices, which he called
Dualgruppen. His two papers on the subject, Über Zerlegungen von Zahlen durch
ihre größten gemeinsamen Teiler (1897) and Über die von drei Moduln erzeugte
Dualgruppe (1900), are classics, remarkably modern in spirit, which have inspired
many later mathematicians.
“There is nothing new under the sun,” and so Dedekind found. Lattices, espe-
cially distributive lattices and Boolean algebras, arise naturally in logic, and thus
some of the elementary theory of lattices had been worked out earlier by Ernst
Schröder in his book Die Algebra der Logik. Nonetheless, it is the connection be-
tween modern algebra and lattice theory, which Dedekind recognized, that provided
the impetus for the development of lattice theory as a subject, and which remains
our primary interest.
Unfortunately, Dedekind was ahead of his time in making this connection, and
so nothing much happened in lattice theory for the next thirty years. Then, with
the development of universal algebra in the 1930’s by Garrett Birkhoff, Oystein Ore
and others, Dedekind’s work on lattices was rediscovered. From that time on, lattice
theory has been an active and growing subject, in terms of both its application to
algebra and its own intrinsic questions.
These notes are intended as the basis for a one-semester introduction to lattice
theory. Only a basic knowledge of modern algebra is presumed, and I have made no
attempt to be comprehensive on any aspect of lattice theory. Rather, the intention
is to provide a textbook covering what we lattice theorists would like to think every
mathematician should know about the subject, with some extra topics thrown in
for flavor, all done thoroughly enough to provide a basis for a second course for the
student who wants to go on in lattice theory or universal algebra.
It is a pleasure to acknowledge the contributions of students and colleagues to
these notes. I am particularly indebted to Michael Tischendorf, Alex Pogel and the
referee for their comments. Mahalo to you all.
Finally, I hope these notes will convey some of the beauty of lattice theory as I
learned it from two wonderful teachers, Bjarni Jónsson and Bob Dilworth.
 1. Ordered Sets

“And just how far would you like to go in?” he asked....

“Not too far but just far enough so’s we can say that we’ve been there,” said
the first chief.
“All right,” said Frank, “I’ll see what I can do.”
–Bob Dylan

In group theory, groups are defined algebraically as a model of permutations.

The Cayley representation theorem then shows that this model is “correct”: every
group is isomorphic to a group of permutations. In the same way, we want to define
a partial order to be an abstract model of set containment ⊆, and then we should
prove a representation theorem to show that this is what we have.
A partially ordered set, or more briefly just ordered set, is a system P = (P, ≤)
where P is a nonempty set and ≤ is a binary relation on P satisfying, for all
x, y, z ∈ P ,
(1) x ≤ x, (reflexivity)
(2) if x ≤ y and y ≤ x, then x = y, (antisymmetry)
(3) if x ≤ y and y ≤ z, then x ≤ z. (transitivity)
The most natural example of an ordered set is P(X), the collection of all subsets of
a set X, ordered by ⊆. Another familiar example is Sub G, all subgroups of a group
G, again ordered by set containment. You can think of lots of examples of this type.
Indeed, any nonempty collection Q of subsets of X, ordered by set containment,
forms an ordered set.
More generally, if P is an ordered set and Q ⊆ P , then the restriction of ≤ to Q
is a partial order, leading to a new ordered set Q.
The set < of real numbers with its natural order is an example of a rather special
type of partially ordered set, namely a totally ordered set, or chain. C is a chain
if for every x, y ∈ C, either x ≤ y or y ≤ x. At the opposite extreme we have
antichains, ordered sets in which ≤ coincides with the equality relation =.
We say that x is covered by y in P, written x ≺ y, if x < y and there is no z ∈ P
with x < z < y. It is clear that the covering relation determines the partial order
in a finite ordered set P. In fact, the order ≤ is the smallest reflexive, transitive
relation containing ≺. We can use this to define a Hasse diagram for a finite ordered
set P: the elements of P are represented by points in the plane, and a line is drawn
from a up to b precisely when a ≺ b. In fact this description is not precise, but it
1
 (b) antichain

(a) chain (c) P{x, y, z}

Note
(d) ∼
= (e)

(d) (e)
Figure 1.1

is close enough for government purposes. In particular, we can now generate lots of
examples of ordered sets using Hasse diagrams, as in Figure 1.1.
The natural maps associated with the category of ordered sets are the order
preserving maps, those satisfying the condition x ≤ y implies f (x) ≤ f (y). We
say that P is isomorphic to Q, written P ∼ = Q, if there is a map f : P → Q
which is one-to-one, onto, and both f and f −1 are order preserving, i.e., x ≤ y iff
f (x) ≤ f (y).
With that we can state the desired representation of any ordered set as a system
of sets ordered by containment.
Theorem 1.1. Let Q be an ordered set, and let φ : Q → P(Q) be defined by

φ(x) = {y ∈ Q : y ≤ x}.

Then Q is isomorphic to the range of φ ordered by ⊆.

Proof. If x ≤ y, then z ≤ x implies z ≤ y by transitivity, and hence φ(x) ⊆ φ(y).
Since x ∈ φ(x) by reflexivity, φ(x) ⊆ φ(y) implies x ≤ y. Thus x ≤ y iff φ(x) ⊆ φ(y).
That φ is one-to-one then follows by antisymmetry. 
A subset I of P is called an order ideal if x ≤ y ∈ I implies x ∈ I. The set
of all order ideals of P forms an ordered set O(P) under set inclusion. The map
2
φ of Theorem 1.1 embeds Q in O(Q). Note that we have the additional property
that the intersection of any collection of order ideals of P is again in an order ideal
(which may be empty).
Given an ordered set P = (P, ≤), we can form another ordered set P d = (P, ≤d ),
called the dual of P, with the order relation defined by x ≤d y iff y ≤ x. In the finite
case, the Hasse diagram of P d is obtained by simply turning the Hasse diagram of
P upside down (see Figure 1.2). Many concepts concerning ordered sets come in
dual pairs, where one version is obtained from the other by replacing “≤” by “≥”
throughout.

a b c

c a b
P Figure 1.2 Pd

For example, a subset F of P is called an order filter if x ≥ y ∈ F implies x ∈ F .

An order ideal of P is an order filter of P d , and vice versa.
The ordered set P has a maximum (or greatest) element if there exists x ∈ P such
that y ≤ x for all y ∈ P . An element x ∈ P is maximal if there is no element y ∈ P
with y > x. Clearly these concepts are different. Minimum and minimal elements
are defined dually.
The next lemma is simple but particularly important.
Lemma 1.2. The following are equivalent for an ordered set P.
(1) Every nonempty subset S ⊆ P contains an element minimal in S.
(2) P contains no infinite descending chain

a0 > a1 > a2 > . . .

(3) If
a0 ≥ a1 ≥ a2 ≥ . . .
in P, then there exists k such that an = ak for all n ≥ k.

Proof. The equivalence of (2) and (3) is clear, and likewise that (1) implies (2).
There is, however, a subtlety in the proof of (2) implies (1). Suppose P fails (1) and
that S ⊆ P has no minimal element. In order to find an infinite descending chain in
S, rather than just arbitrarily long finite chains, we must use the Axiom of Choice.
One way to do this is as follows.
3
 Let f be a choice function on the subsets of S, i.e., f assigns to each nonempty
subset T ⊆ S an element f (T ) ∈ T . Let a0 = f (S), and for each i ∈ ω define
ai+1 = f ({s ∈ S : s < ai }); the argument of f in this expression is nonempty
because S has no minimal element. The sequence so defined is an infinite descending
chain, and hence P fails (2). 
The conditions described by the preceding lemma are called the descending chain
condition (DCC). The dual notion is called the ascending chain condition (ACC).
These conditions should be familiar to you from ring theory (for ideals). The next
lemma just states that ordered sets satisfying the DCC are those for which the
principle of induction holds.
Lemma 1.3. Let P be an ordered set satisfying the DCC. If ϕ(x) is a statement
such that
(1) ϕ(x) holds for all minimal elements of P , and
(2) whenever ϕ(y) holds for all y < x, then ϕ(x) holds,
then ϕ(x) is true for every element of P .
Note that (1) is in fact a special case of (2). It is included in the statement of the
lemma because in practice minimal elements usually require a separate argument
(like the case n = 0 in ordinary induction).
The proof is immediate. The contrapositive of (2) states that the set F = {x ∈
P : ϕ(x) is false} has no minimal element. Since P satisfies the DCC, F must
therefore be empty.
We now turn our attention more specifically to the structure of ordered sets.
Define the width of an ordered set P by

w(P) = sup{|A| : A is an antichain in P}

where |A| denotes the cardinality of A.1 A second invariant is the chain covering
number c(P), defined to be the least cardinal γ such that P is the union of γ chains
in P. Because no chain can contain more than one element of a given
S antichain, we
must have |A| ≤ |I| whenever A is an antichain in P and P = i∈I Ci is a chain
covering. Therefore
w(P) ≤ c(P)

for any ordered set P. The following result, due to R. P. Dilworth [2], says in
particular that if P is finite, then w(P) = c(P).

1 Note that the width function w(P) does not distinguish, for example, between ordered sets

which contain arbitrarily large finite antichains and those which contain a countably infinite an-
tichain. For this reason, in ordered sets of infinite width it is sometimes useful to consider the
function µ(P), which is defined to be the least cardinal κ such that κ + 1 > |A| for every antichain
A of P. We will restrict our attention to w(P).
4
Theorem 1.4. If w(P) is finite, then w(P) = c(P).
Our discussion of the proof will take the scenic route. We begin with the case
when P is finite, using H. Tverberg’s nice proof [11].
Proof in the finite case. We need to show c(P) ≤ w(P), which is done by induction
on |P |. Let w(P) = k, and let C be a maximal chain in P. If P is a chain,
w(P) = c(P) = 1, so assume C 6= P. Because C can contain at most one element of
any maximal antichain, the width w(P −C) is either k or k−1, and both possibilities
can occur. If w(P − C) = k − 1, then P − C is the union of k − 1 chains, whence P
is a union of k chains.
So suppose w(P − C) = k, and let A = {a1 , . . . , ak } be a maximal antichain in
P − C. As |A| = k, it is also a maximal antichain in P. Set

L = {x ∈ P : x ≤ ai for some i},

U = {x ∈ P : x ≥ aj for some j}.

Since every element of P is comparable with some element of A, we have P = L ∪ U ,

while A = L ∩ U . Moreover, the maximality of C insures that the largest element of
C does not belong to L (remember A ⊆ P − C), so |L| < |P |. Dually, |U | < |P | also.
Hence L is a union of k chains, L = D1 ∪ · · · ∪ Dk , and similarly U = E1 ∪ · · · ∪ Ek as
a union of chains. By renumbering, if necessary, we may assume that ai ∈ Di ∩ Ei
for 1 ≤ i ≤ k, so that Ci = Di ∪ Ei is a chain. Thus

P = L ∪ U = C1 ∪ · · · ∪ Ck

is a union of k chains. 
So now we want to consider an infinite ordered set P of finite width k. Not
surprisingly, we will want to use one of the 210 equivalents of the Axiom of Choice!
(See H. Rubin and J. Rubin [9].) This requires some standard terminology.
Let P be an ordered set, and let S be a subset of P . We say that an element
x ∈ P is an upper bound for S if x ≥ s for all s ∈ S. An upper bound x need not
belong to S. We say that x is the least upper bound for S if x is an upper bound
for S and x ≤ y for every upper bound y of S. If the least upper bound of S exists,
then it is unique. Lower bound and greatest lower bound are defined dually.
Theorem 1.5. The following set theoretic axioms are equivalent.
(1) (Axiom of Choice) If X is a nonempty set, then there is a map φ :
P(X) → X such that φ(A) ∈ A for every nonempty A ⊆ X.
(2) (Zermelo well-ordering principle) Every nonempty set admits a well-
ordering (a total order satisfying the DCC ).
(3) (Hausdorff maximality principle) Every chain in an ordered set P can
be embedded in a maximal chain.
5
 (4) (Zorn’s Lemma) If every chain in an ordered set P has an upper bound in
P, then P contains a maximal element.
(5) If every chain in an ordered set P has a least upper bound in P, then P
contains a maximal element.

The proof of Theorem 1.5 is given in Appendix 2.

Our plan is to use Zorn’s Lemma to prove the compactness theorem (due to
K. Gödel [5]), and then the compactness theorem to prove the infinite case of Dil-
worth’s theorem. We need to first recall some of the basics of sentential logic.
Let S be a set, whose members will be called sentence symbols. Initially the
sentence symbols carry no intrinsic meaning; in applications they will correspond to
various mathematical statements.
We define well formed formulas (wff) on S by the following rules.
(1) Every sentence symbol is a wff.
(2) If α and β are wffs, then so are (¬α), (α ∧ β) and (α ∨ β).
(3) Only symbols generated by the first two rules are wffs.
The set of all wffs on S is denoted by S.2 Of course, we think of ¬, ∧ and ∨ as
corresponding to “not”, “and” and “or”, respectively.
A truth assignment on S is a map ν : S → {T, F }. Each truth assignment has a
natural extension ν : S → {T, F }. The map ν is defined recursively by the rules
(1) ν(¬ϕ) = T if and only if ν(ϕ) = F ,
(2) ν(ϕ ∧ ψ) = T if and only if ν(ϕ) = T and ν(ψ) = T ,
(3) ν(ϕ ∨ ψ) = T if and only if ν(ϕ) = T or ν(ψ) = T (including the case that
both are equal to T ).
A set Σ ⊆ S is satisfiable if there exists a truth assignment ν such that ν(φ) = T
for all φ ∈ Σ. Σ is finitely satisfiable if every finite subset Σ0 ⊆ Σ is satisfiable.
Note that these concepts refer only to the internal consistency of Σ; there is so far
no meaning attached to the sentence symbols themselves.
Theorem 1.6. (The compactness theorem) A set of wffs is satisfiable if and
only if it is finitely satisfiable.
Proof. Let S be a set of sentence symbols and S the corresponding set of wffs.
Assume that Σ ⊆ S is finitely satisfiable. Using Zorn’s Lemma, let ∆ be maximal
in P(S) such that
(1) Σ ⊆ ∆,
(2) ∆ is finitely satisfiable.
We claim that for all ϕ ∈ S, either ϕ ∈ ∆ or (¬ϕ) ∈ ∆ (but of course not both).
Otherwise, by the maximality of ∆, we could find a finite subset ∆0 ⊆ ∆ such
that ∆0 ∪ {ϕ} is not satisfiable, and a finite subset ∆1 ⊆ ∆ such that ∆1 ∪ {¬ϕ} is
2 Technically, S is just the absolutely free algebra generated by S with the operation symbols

given in (2).
6
not satisfiable. But ∆0 ∪ ∆1 is satisfiable, say by a truth assignment ν. If ν(ϕ) = T ,
this contradicts the choice of ∆0 , while ν(¬ϕ) = T contradicts the choice of ∆1 . So
the claim holds.
Now define a truth assignment µ as follows. For each sentence symbol p ∈ S,
define
µ(p) = T iff p ∈ ∆ .
Now we claim that for all ϕ ∈ S, µ(ϕ) = T iff ϕ ∈ ∆. This will yield µ(ϕ) = T for
all ϕ ∈ Σ, so that Σ is satisfiable.
To prove this last claim, let G = {ϕ ∈ S : µ(ϕ) = T iff ϕ ∈ ∆}. We have S ⊆ G,
and we need to show that G is closed under the operations ¬, ∧ and ∨, so that
G = S.
(1) Suppose ϕ = ¬β with β ∈ G. Then, using the first claim,

µ(ϕ) = T iff µ(β) = F

iff β∈
/∆
iff ¬β ∈ ∆
iff ϕ∈∆.

Hence ϕ = ¬β ∈ G.
(2) Suppose ϕ = α ∧ β with α, β ∈ G. Note that α ∧ β ∈ ∆ iff α ∈ ∆ and β ∈ ∆.
For if α ∧ β ∈ ∆, since {α ∧ β, ¬α} is not satisfiable we must have α ∈ ∆, and
similarly β ∈ ∆. Conversely, if α ∈ ∆ and β ∈ ∆, then since {α, β, ¬(α ∧ β)} is not
satisfiable, we have α ∧ β ∈ ∆. Thus

µ(ϕ) = T iff µ(α) = T and µ(β) = T

iff α ∈ ∆ and β ∈ ∆
iff (α ∧ β) ∈ ∆
iff ϕ ∈ ∆.

Hence ϕ = (α ∧ β) ∈ G.
(3) The case ϕ = α ∨ β is similar to (2). 
We return to considering an infinite ordered set P of width k. Let S = {cxi : x ∈
P, 1 ≤ i ≤ k}. We think of cxi as corresponding to the statement “x is in the i-th
chain.” Let Σ be all sentences of the form

(a) cx1 ∨ · · · ∨ cxk

for x ∈ P , and

(b) ¬(cxi ∧ cyi )

7
for all incomparable pairs x, y ∈ P and 1 ≤ i ≤ k. By the finite version of Dilworth’s
theorem, Σ is finitely satisfiable, so by the compactness theorem Σ is satisfiable, say
by ν. We obtain the desired representation by putting Ci = {x ∈ P : ν(cxi ) = T }.
The sentences (a) insure that C1 ∪ · · · ∪ Ck = P , and the sentences (b) say that each
Ci is a chain.
This completes the proof of Theorem 1.4.
A nice example due to M. Perles shows that Dilworth’s theorem is no longer true
when the width is allowed to be infinite [7]. Let κ be an infinite ordinal,3 and let P
be the direct product κ × κ, ordered pointwise. Then P has no infinite antichains,
so w(P) = ℵ0 , but c(P) = |κ|.
There is a nice discussion of the consequences and extensions of Dilworth’s The-
orem in Chapter 1 of [1]. Algorithmic aspects are discussed in Chapter 11 of [3],
while a nice alternate proof appears in F. Galvin [4].
It is clear that the collection of all partial orders on a set X, ordered by set
inclusion, is itself an ordered set PO(X). The least member of PO(X) is the
equality relation, corresponding to the antichain order. The maximal members of
PO(X) are the various total (chain) orders on X. Note that the intersection of a
collection of partial orders on X is again a partial order. The next theorem, due
to E. Szpilrajn, expresses an arbitrary partial ordering as an intersection of total
orders [10].
Theorem 1.7. Every partial ordering on a set X is the intersection of the total
orders on X containing it.
Szpilrajn’s theorem is an immediate consequence of the next lemma.
Lemma 1.8. Given an ordered set (P, ≤) and a 6≤ b, there exists an extension ≤∗
of ≤ such that (P, ≤∗ ) is a chain and b ≤∗ a.
Proof. Let a  b in P. Define

x≤y

x ≤0 y if or


x ≤ b and a ≤ y.

It is straightforward to check that this is a partial order with b ≤0 a.

If P is finite, repeated application of this construction yields a total order ≤∗
extending ≤0 , so that b ≤∗ a. For the infinite case, we can either use the compactness
theorem, or perhaps easier Zorn’s Lemma (the union of a chain of partial orders on
X is again one) to obtain a total order ≤∗ extending ≤0 . 
Define the dimension d(P) of an ordered set P to be the smallest cardinal κ
such that the order ≤ on P is the intersection of κ total orders. The next result
summarizes two basic facts about the dimension.
3 See Appendix 1.
8
Theorem 1.9. Let P be an ordered set. Then
(1) d(P) is the smallest cardinal γ such that P can be embedded into the direct
product of γ chains,
(2) d(P) ≤ c(P).

Proof. First suppose ≤ is the intersection of total orders ≤i (i ∈ I) on P . If we

Q let Ci
be the chain (P, ≤i ), then it is easy to see that the natural map ϕ : P → i∈I Ci ,
with (ϕ(x))i = x for all x ∈ P , satisfies x ≤ y iff ϕ(x) ≤ ϕ(y). Hence ϕ is an
embedding.
Q
Conversely, assume ϕ : P → i∈I Ci is an embedding of P into a direct product of
chains. We want to show that this leads to a representation of ≤ as the intersection
of |I| total orders. Define

x≤y

x Ri y if or


ϕ(x)i < ϕ(y)i .

You should check that Ri is a partial order extending ≤. By Lemma 1.8 each Ri
can be extended to a total order ≤i extending ≤. To see that ≤ is the intersection
of the ≤i ’s, suppose x  y. Since ϕ is an embedding, then ϕ(x)i  ϕ(y)i for some
i. Thus ϕ(x)i > ϕ(y)i , implying y Ri x and hence y ≤i x, or equivalently x i y (as
x 6= y), as desired.
Thus the order on P is the intersection of κ total orders if and only if P can be
embedded into the direct product of κ chains, yielding (1).
S
For (2), assume P = j∈J Cj with each Cj a chain. Then, for each j ∈ J, the
ordered set O(Cj ) of order ideals of Cj is also a chain. Define a map ϕ : P →
Q
j∈J O(Cj ) by (ϕ(x))j = {y ∈ Cj : y ≤ x}. (Note ∅ ∈ O(Cj ), and (ϕ(x))j = ∅
is certainly possible.) Then ϕ is clearly order-preserving. On the other hand, if
x  y in P and x ∈ Cj , then x ∈ (ϕ(x))j and x ∈ / (ϕ(y))j , so (ϕ(x))j * (ϕ(y))j and
ϕ(x)  ϕ(y). Thus P can be embedded into a direct product of |J| chains. Using
(1), this shows d(P ) ≤ c(P ). 

Now we have three invariants defined on ordered sets: w(P ), c(P ) and d(P ).
The exercises will provide you an opportunity to work with these in concrete cases.
We have shown that w(P ) ≤ c(P ) and d(P ) ≤ c(P ), but width and dimension are
independent. Indeed, if κ is an ordinal and κd its dual, then κ × κd has width |κ|
but dimension 2. It is a little harder to find examples of high dimension but low
width (necessarily infinite by Dilworth’s theorem), but it is possible (see [6] or [8]).
This concludes our brief introduction to ordered sets per se. We have covered
only the most classical results of what is now an active field of research, supporting
its own journal, Order.
9
 Exercises for Chapter 1
1. Draw the Hasse diagrams for all 4-element ordered sets (up to isomorphism).
2. Let N denote the positive integers. Show that the relation a | b (a divides b)
is a partial order on N . Draw the Hasse diagram for the ordered set of all divisors
of 60.
3. A partial map on a set X is a map σ : S → X where S = dom σ is a subset of
X. Define σ ≤ τ if dom σ ⊆ dom τ and τ (x) = σ(x) for all x ∈ dom σ. Show that
the collection of all partial maps on X is an ordered set.
4. (a) Give an example of a map f : P → Q which is one-to-one, onto and
order-preserving, but not an isomorphism.
(b) Show that the following are equivalent for ordered sets P and Q.
(i) P ∼= Q (as defined before Theorem 1.1).
(ii) There exists f : P  Q such that f (x) ≤ f (y) iff x ≤ y. ( means the map
is onto.)
(iii) There exist f : P → Q and g : Q → P , both order-preserving, with gf = idP
and f g = idQ .
5. Find all order ideals of the rational numbers Q with their usual order.
6. Prove that all chains in an ordered set P are finite if and only if P satisfies
both the ACC and DCC.
7. Find w(P), c(P) and d(P) for
(a) an antichain A with |A| = κ, where κ is a cardinal,
(b) Mκ , where κ is a cardinal, the ordered set diagrammed in Figure 1.3(a).
(c) an n-crown, the ordered set diagrammed in Figure 1.3(b).
(d) P(X) with X a finite set,
(e) P(X) with X infinite.

a1 a2 a3 an

c1 cκ

b1 b2 b3 bn

Figure 1.3

8. Embed Mn (2 ≤ n < ∞) into a direct product of two chains. Express the

order on Mn as the intersection of two totally ordered extensions.
9. Let P be a finite ordered set with at least ab + 1 elements. Prove that P
contains either an antichain with a + 1 elements, or a chain with b + 1 elements.
10
 10. Phillip Hall proved that if X is a finite set and S1 , . . . , Sn are subsets of X,
then there is a system of distinct representatives (SDR) a1 , . . . , anSwith aj ∈ Sj if
and only if for all 1 ≤ k ≤ n and distinct indices i1 , . . . , ik we have | 1≤j≤k Sij | ≥ k.
(a) Derive this result from Dilworth’s theorem.
(b) Prove Marshall Hall’s extended version: If Si (i ∈ I) are finite subsets of a
(possibly infinite) set X, then they have an SDR if and only if the condition
of P. Hall’s theorem holds for every n.
11. Let R be a binary relation on a set X which contains no cycle of the form
x0 R x1 R . . . R xn R x0 with xi 6= xi+1 . Show that the reflexive transitive closure
of R is a partial order.
12. A reflexive, transitive, binary relation is called a quasiorder.
(a) Let R be a quasiorder on a set X. Define x ≡ y if x R y and y R x. Prove
that ≡ is an equivalence relation, and that R induces a partial order on X/≡.
(b) Let P be an ordered set, and define a relation  on the subsets of P by
X  Y if for each x ∈ X there exists y ∈ Y with x ≤ y. Verify that  is a
quasiorder.
13. Let ω1 denote the first uncountable ordinal, and let P be the direct product
ω1 × ω1 . Prove that every antichain of P is finite, but c(P) = ℵ1 .

References
1. K. Bogart, R. Freese and J. Kung, Eds., The Dilworth Theorems, Birkhäuser, Boston, 1990.
2. R. P. Dilworth, A decomposition theorem for partially ordered sets, Annals of Math. 51 (1950),
161–166.
3. R. Freese, J. Ježek and J. B. Nation, Free Lattices, Mathematical Surveys and Monographs,
vol. 42, Amer. Math. Soc., Providence, R. I., 1995.
4. F. Galvin, A proof of Dilworth’s chain decomposition theorem, Amer. Math. Monthly 101
(1994), 352–353.
5. K. Gödel, Die Vollständigkeit der Axiome des logischen Funktionenkalküls, Monatsh. Math.
Phys. 37 (1930), 349–360.
6. J. B. Nation, D. Pickering and J. Schmerl, Dimension may exceed width, Order 5 (1988),
21–22.
7. M. A. Perles, On Dilworth’s theorem in the infinite case, Israel J. Math. 1 (1963), 108–109.
8. M. Pouzet, Généralisation d’une construction de Ben-Dushnik et E. W. Miller, Comptes
Rendus Acad. Sci. Paris 269 (1969), 877–879.
9. H. Rubin and J. Rubin, Equivalents of the Axiom of Choice II, Studies in Logic and the
Foundations of Mathematics, vol. 116, North-Holland, Amsterdam, 1985.
10. E. Szpilrajn, Sur l’extension de l’ordre partiel, Fund. Math. 16 (1930), 386–389.
11. H. Tverberg, On Dilworth’s decomposition theorem for partially ordered sets, J. Combinatorial
Theory 3 (1967), 305–306.

11
 2. Semilattices, Lattices and Complete Lattices

There’s nothing quite so fine

As an earful of Patsy Cline.
–Steve Goodman

The most important partially ordered sets come endowed with more structure
than that. For example, the significant feature about PO(X) for Theorem 1.7 is
not just its partial order, but that it is closed under intersection. In this chapter we
will meet several types of structures which arise naturally in algebra.
A semilattice is an algebra S = (S, ∗) satisfying, for all x, y, z ∈ S,
(1) x ∗ x = x,
(2) x ∗ y = y ∗ x,
(3) x ∗ (y ∗ z) = (x ∗ y) ∗ z.
In other words, a semilattice is an idempotent commutative semigroup. The symbol
∗ can be replaced by any binary operation symbol, and in fact we will most often
use one of ∨, ∧, + or ·, depending on the setting. The most natural example of
a semilattice is (P(X), ∩), or more generally any collection of subsets of X closed
under intersection. Thus the semilattice PO(X) of partial orders on X is naturally
contained in (P(X 2 ), ∩).
Theorem 2.1. In a semilattice S, define x ≤ y if and only if x ∗ y = x. Then
(S, ≤) is an ordered set in which every pair of elements has a greatest lower bound.
Conversely, given an ordered set P with that property, define x ∗ y = g.l.b. (x, y).
Then (P, ∗) is a semilattice.
Proof. Let (S, ∗) be a semilattice, and define ≤ as above. First we check that ≤ is
a partial order.
(1) x ∗ x = x implies x ≤ x.
(2) If x ≤ y and y ≤ x, then x = x ∗ y = y ∗ x = y.
(3) If x ≤ y ≤ z, then x ∗ z = (x ∗ y) ∗ z = x ∗ (y ∗ z) = x ∗ y = x, so x ≤ z.
Since (x ∗ y) ∗ x = x ∗ (x ∗ y) = (x ∗ x) ∗ y = x ∗ y) we have x ∗ y ≤ x; similarly
x ∗ y ≤ y. Thus x ∗ y is a lower bound for {x, y}. To see that it is the greatest lower
bound, suppose z ≤ x and z ≤ y. Then z ∗ (x ∗ y) = (z ∗ x) ∗ y = z ∗ y = z, so
z ≤ x ∗ y, as desired.
The proof of the converse is likewise a direct application of the definitions, and
is left to the reader. 
12
 A semilattice with the above ordering is usually called a meet semilattice, and as
a matter of convention ∧ or · is used for the operation symbol. In Figure 2.1, (a)
and (b) are meet semilattices, while (c) fails on several counts.

(a) (b) (c)

Figure 2.1

Sometimes it is more natural to use the dual order, setting x ≥ y iff x ∗ y = x.

In that case, S is referred to as a join semilattice, and the operation is denoted by
∨ or +.
A subsemilattice of S is a subset T ⊆ S which is closed under the operation ∗ of
S: if x, y ∈ T then x ∗ y ∈ T . Of course, that makes T a semilattice in its own right,
since the equations defining a semilattice still hold in (T, ∗).1
Similarly, a homomorphism between two semilattices is a map h : S → T with
the property that h(x ∗ y) = h(x) ∗ h(y). An isomorphism is a homomorphism that
is one-to-one and onto. It is worth noting that, because the operation is determined
by the order and vice versa, two semilattices are isomorphic if and only if they are
isomorphic as ordered sets.
The collection of all order ideals of a meet semilattice S forms a semilattice O(S)
under set intersection. The mapping from Theorem 1.1 gives us a set representation
for meet semilattices.
Theorem 2.2. Let S be a meet semilattice. Define φ : S → O(S) by

φ(x) = {y ∈ S : y ≤ x}.

Then S is isomorphic to (φ(S), ∩).

Proof. We already know that φ is an order embedding of S into O(S). Moreover,
φ(x ∧ y) = φ(x) ∧ φ(y) because x ∧ y is the greatest lower bound of x and y, so that
z ≤ x ∧ y if and only if z ≤ x and z ≤ y. 
A lattice is an algebra L = (L, ∧, ∨) satisfying, for all x, y, z ∈ S,
(1) x ∧ x = x and x ∨ x = x,
1 However, it is not enough that the elements of T form a semilattice under the ordering ≤. For

example, the sets {1, 2}, {1, 3} and ∅ do not form a subsemilattice of (P({1, 2, 3}), ∩).
13
 (2) x ∧ y = y ∧ x and x ∨ y = y ∨ x,
(3) x ∧ (y ∧ z) = (x ∧ y) ∧ z and x ∨ (y ∨ z) = (x ∨ y) ∨ z,
(4) x ∧ (x ∨ y) = x and x ∨ (x ∧ y) = x.
The first three pairs of axioms say that L is both a meet and join semilattice. The
fourth pair (called the absorption laws) say that both operations induce the same
order on L. The lattice operations are sometimes denoted by · and +; for the sake
of consistency we will stick with the ∧ and ∨ notation.
An example is the lattice (P(X), ∩, ∪) of all subsets of a set X, with the usual
set operations of intersection and union. This turns out not to be a very general
example, because subset lattices satisfy the distributive law

(D) A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C).

The corresponding lattice equation does not hold in all lattices: x ∧ (y ∨ z) =

(x ∧ y) ∨ (x ∧ z) fails, for example, in the two lattices in Figure 2.2. Hence we
cannot expect to prove a representation theorem which embeds an arbitrary lattice
in (P(X), ∩, ∪) for some set X (although we will prove such a result for distributive
lattices). A more general example would be the lattice Sub(G) of all subgroups of
a group G. Most of the remaining results in this section are designed to show how
lattices arise naturally in mathematics, and to point out additional properties which
some of these lattices have.

M3 N5
Figure 2.2

Theorem 2.3. In a lattice L, define x ≤ y if and only if x ∧ y = x. Then (L, ≤)

is an ordered set in which every pair of elements has a greatest lower bound and
a least upper bound. Conversely, given an ordered set P with that property, define
x ∧ y = g.l.b. (x, y) and x ∨ y = l.u.b. (x, y). Then (P, ∧, ∨) is a lattice.
The crucial observation in the proof is that, in a lattice, x ∧ y = x if and only
if x ∨ y = y by the absorption laws. The rest is a straightforward extension of
Theorem 2.1.
This time we leave it up to you to figure out the correct definitions of sublattice,
homomorphism and isomorphism for lattices. If a lattice has a least element, it is
14
denoted by 0; the greatest element, if it exists, is denoted by 1. Of special importance
are the quotient (or interval) sublattices:

a/b = {x ∈ L : b ≤ x ≤ a}
a/0 = {x ∈ L : x ≤ a}
1/a = {x ∈ L : a ≤ x}

The latter notations are used irrespective of whether L actually has a least element
0 or a greatest element 1.2
One further bit of notation will prove useful. For a subset A of an ordered set P,
let Au denote the set of all upper bounds of A, i.e.,

Au = {x ∈ P : x ≥ a for all a ∈ A}
\
= 1/a.
a∈A

Dually, A` is the set of all lower bounds of A,

A` = {x ∈ P : x ≤ a for all a ∈ A}
\
= a/0.
a∈A

Let us consider the question of when a subset A of an ordered set P has a least
upper bound. Clearly Au must be nonempty, and this will certainly be the case if P
has a greatest element. If moreover it happens that Au has a greatest lower bound
z in P, then in fact z ∈ Au , i.e., a ≤ z for all a ∈ A, because each a ∈ A is a lower
bound for Au . Therefore by definition z is the least
W upper bound of A. In this case
we say that the join of A exists, and write z = A (treating the join as a partially
defined operation). V
But if S is a finite meet semilattice with a greatest element, then Au exists for
every A ⊆ S. Thus we have the following result.
Theorem 2.4. Let S be a finite meet semilattice with greatest element 1. Then S
is a lattice with the join operation defined by
^ ^
x ∨ y = {x, y}u = (1/x ∩ 1/y).

This result not only yields an immediate supply of lattice examples, but it pro-
vides us with an efficient algorithm for deciding when a finite ordered set is a lattice:
2 There are several commonly used ways of denoting interval sublattices; the one we have adopted
is as good as any, but hardly universal. The most common alternative has a/b = [b, a], a/0 = (a]
and 1/a = [a). The notations ↓ a for a/0 and ↑ a for 1/a are also widely used.
15
if P has a greatest element and every pair of elements has a meet, then P is a lattice.
The dual version is of course equally useful.
Every finite subset of a lattice has a greatest lower bound and a least upper bound,
but these bounds need not exist for infinite subsets. Let us define a complete
V lattice
to be an ordered setWL in which every subset A has a greatest lower bound A and a
least upper bound A.3 Clearly every finite lattice is complete, and every complete
lattice is a lattice with 0 and 1 (but not conversely). Again P(X) is a natural (but
not very general) example of a complete lattice, and Sub(G) is a better one. The
rational numbers with their natural order form a lattice which is not complete.
Likewise, a complete meet semilattice is an ordered set S with a greatest element
and
V the property that every nonemptyV subset A of S has a greatest lower bound
A. By convention, we define ∅ = 1, the greatest element of S. The analogue of
Theorem 2.4 is as follows.
Theorem 2.5. If L is a complete meet semilattice, then L is a complete lattice with
the join operation defined by
_ ^ ^ \
A= Au = ( 1/a).
a∈A

Complete lattices abound in mathematics because of their connection with closure

systems. We will introduce three different ways of looking at these things, each with
certain advantages, and prove that they are equivalent.
A closure system on a set X is a collection C of subsets of X
T which is closed under
arbitrary intersections (including the empty intersection, so ∅ = X ∈ C). The sets
in C are called closed sets. By Theorem 2.5, the closed sets of a closure system form
a complete lattice. Various examples come to mind:
(i) closed subsets of a topological space,
(ii) subgroups of a group,
(iii) subspaces of a vector space,
(iv) order ideals of an ordered set,
(v) convex subsets of euclidean space <n .
You can probably think of other types of closure systems, and more will arise as we
go along.
A closure operator on a set X is a map Γ : P(X) → P(X) satisfying, for all
subsets A, B ⊆ X,
(1) A ⊆ Γ(A),
(2) A ⊆ B implies Γ(A) ⊆ Γ(B),
(3) Γ(Γ(A)) = Γ(A).
3 We could have defined complete lattices as a type of infinitary algebra satisfying some axioms,
but since these kinds of structures are not very familiar the above approach seems more natural.
Following standard usage, we only allow finitary operations
W in anValgebra (see Appendix 3). Thus a
complete lattice as such, with its arbitrary operations A and A, does not count as an algebra.
16
The closure operators associated with the closure systems above are as follows:
(i) X a topological space and Γ(A) the closure of A,
(ii) G a group and Sg(A) the subgroup generated by A,
(iii) V a vector space and Span(A) the set of all linear combinations of elements
of A,
(iv) P an ordered set and O(A) the order ideal generated by A,
(v) <n and H(A) the convex hull of A.
For a closure operator, a set D is called closed if Γ(D) = D, or equivalently (by
(3)), if D = Γ(A) for some A.
A set of closure rules on a set X is a collection Σ of properties ϕ(S) of subsets
of X, where each ϕ(S) has one of the forms

x∈S

or
Y ⊆ S =⇒ z ∈ S
with x, z ∈ X and Y ⊆ X. (Note that the first type of rule is a degenerate case
of the second, taking Y = ∅.) A subset D of X is said to be closed with respect to
these rules if ϕ(D) is true for each ϕ ∈ Σ. The closure rules corresponding to our
previous examples are:
(i) all rules Y ⊆ S =⇒ z ∈ S where z is an accumulation point of Y ,
(ii) the rule 1 ∈ S and all rules

x ∈ S =⇒ x−1 ∈ S
{x, y} ⊆ S =⇒ xy ∈ S

with x, y ∈ G,
(iii) 0 ∈ S and all rules {x, y} ⊆ S =⇒ ax + by ∈ S with a, b scalars,
(iv) for all pairs with x < y in P the rules y ∈ S =⇒ x ∈ S,
(v) for all x, y ∈ <n and 0 < t < 1, the rules {x, y} ⊆ S =⇒ tx + (1 − t)y ∈ S.
So the closure rules just list the properties that we check to determine if a set S is
closed or not.
The following theorem makes explicit the connection between these ideas.
Theorem 2.6. (1 ) If C is a closure system on a set X, then the map ΓC : P(X) →
P(X) defined by \
ΓC (A) = {D ∈ C : A ⊆ D}

is a closure operator. Moreover, ΓC (A) = A if and only if A ∈ C.

(2 ) If Γ is a closure operator on a set X, let ΣΓ be the set of all rules

c∈S
17
where c ∈ Γ(∅), and all rules
Y ⊆ S =⇒ z ∈ S
with z ∈ Γ(Y ). Then a set D ⊆ X satisfies all the rules of ΣΓ if and only if
Γ(D) = D.
(3 ) If Σ is a set of closure rules on a set X, let CΣ be the collection of all subsets
of X which satisfy all the rules of Σ. Then CΣ is a closure system.
In other words, the collection of all closed sets of a closure operator forms a
complete lattice, and the property of being a closed set can be expressed in terms of
rules which are clearly preserved by set intersection. It is only a slight exaggeration
to say that all important lattices arise in this way. As a matter of notation, we will
also use CΓ to denote the lattice of Γ-closed sets, even though this particular variant
is skipped in the statement of the theorem.
Proof. Starting with a closure system C, define ΓC as above. Observe that ΓC (A) ∈ C
for any A ⊆ X, and Γ(D) = D for every D ∈ C. Therefore ΓC (ΓC (A)) = ΓC (A), and
the other axioms for a closure operator hold by elementary set theory.
Given a closure operator Γ, it is clear that Γ(D) ⊆ D iff D satisfies all the rules
of ΣΓ . Likewise, it is immediate because of the form of the rules that CΣ is always
a closure system. 
Note that if Γ is a closure operator on a set X, then the operations on CΓ are
given by
^ \
Di = Di
i∈I i∈I
_ [
Di = Γ( Di ).
i∈I i∈I

For example, in the lattice of closed subsets of a topological space, the join is the
closure of the union. In the lattice of subgroups of a group, the join of a collection
of subgroups is the subgroup generated by their union. The lattice of order ideals
is somewhat exceptional in this regard, because the union of a collection of order
ideals is already an order ideal.
One type of closure operator is especially important. If A = hA, F, Ci is an
algebra, then S ⊆ A is a subalgebra of A if c ∈ S for every constant c ∈ C, and
{s1 , . . . , sn } ⊆ S implies f (s1 , . . . , sn ) ∈ S for every basic operation f ∈ F . Of
course these are closure rules, so the intersection of any collection of subalgebras of
A is again one.4 For a subset B ⊆ A, define
\
Sg(B) = {S : S is a subalgebra of A and B ⊆ S}.
4 If A has no constants, then we have to worry about the empty set. We want to allow ∅ in the

subalgebra lattice in this case, but realize that it is an abuse of terminology to call it a subalgebra.
18
By Theorem 2.6, Sg is a closure operator, and Sg(B) is of course the subalgebra
generated by B. The corresponding lattice of closed sets is CSg = Sub A, the lattice
of subalgebras of A.
Galois connections provide another source of closure operators. These are rele-
gated to the exercises not because they are unimportant, but rather to encourage
you to grapple with how they work on your own.
For completeness, we include a representation theorem.

Theorem 2.7. If L is a complete lattice, define a closure operator ∆ on L by

_
∆(A) = {x ∈ L : x ≤ A}.

Then L is isomorphic to C∆ .

The isomorphism ϕ : L → C∆ is just given by ϕ(x) = x/0.

The representation of L as a closure system given by Theorem 2.7 can be greatly

improved upon in some circumstances. Here we will give a better representation for
lattices satisfying the ACC and DCC. In Chapter 3 we will do the same for another
class called algebraic lattices.
W
An element q of a lattice L is called join irreducible if q = F for a finite set F
implies q ∈ F , i.e., q is not the join of other elements. The set of all join irreducible
elementsW in L is denoted by J(L). Note that according to the definition 0 ∈ / J(L),
as 0 = ∅.5

Lemma 2.8. If a lattice L satisfies the DCC, then every element of L is a join of
finitely many join irreducible elements.

Proof. Suppose some element of L is not a join of join irreducible elements.W Let x
be a minimal such element. Then x is not itself join irreducible, so x = F for
W theWminimality of x, each f ∈ F
some finite set F of elements strictly below x. By
is the join of a finite set Gf ⊆ J(L). Then x = f ∈F Gf , a contradiction. 

If L also satisfies the ACC, then join irreducible elements can be identified as
those which cover a unique element, viz.,
_
q∗ = {x ∈ L : x < q}.

The representation of lattices satisfying both chain conditions (in particular, finite
lattices) as a closure system is quite straightforward.

5 This convention is not universal, as join irreducible is sometimes defined by q = r ∨ s implies

q = r or q = s, which is equivalent for nonzero elements.

19
Theorem 2.9. Let L be a lattice satisfying the ACC and DCC. Let Σ be the set of
all closure rules on J(L) of the form

F ⊆ S =⇒ q ∈ S
W
where q is join irreducible, F is a finite subset of J(L), and q ≤ F . (Include the
degenerate cases p ∈ S =⇒ q ∈ S for q ≤ p in J(L).) Then L is isomorphic to
the lattice CΣ of Σ-closed sets.
Proof. Define order preserving maps f : L → CΣ and g : CΣ → L by

f (x) = x/0 ∩ J(L)

_
g(S) = S.

Now gf (x) = x for all x ∈ L by Lemma 2.8.WOn the

W other hand, f g(S) = S for any
W we have S = F for some finite F ⊆ S, which
Σ-closed set, because by the ACC
puts every join irreducible q ≤ F in S by the closure rules. 
As an example of how we might apply these ideas, suppose we want to find
the subalgebra lattice of a finite algebra A. Now Sub A is finite, and every join
irreducible subalgebra is of the form Sg(a) for some a ∈ A (though not necessarily
conversely). Thus we may determine Sub A by first finding all the 1-generated
subalgebras Sg(a), and then computing the joins of sets of these.
Let us look at another type of closure operator. Of course, an ordered set need
not be complete. We say that a pair (L, φ) is a completion of the ordered set P if
L is a complete lattice and φ is an order embedding of P into L. A subset Q of a
complete lattice L is join dense if for every x ∈ L,
_
x = {q ∈ Q : q ≤ x}.

A completion (L, φ) is join dense if φ(P ) is join dense in L, i.e., for every x ∈ L,
_
x = {φ(p) : φ(p) ≤ x}.

It is not hard to see that every completion of P contains a join dense completion.
For, given
W a completion (L, φ) of P, let L0 be the set of allWelements of L of the
form {φ(p) : p ∈ A} for some subset A ⊆ P , including ∅ = 0. Then L0 is
a complete join subsemilattice of L, and hence a complete lattice. Moreover, L0
contains φ(p) for every p ∈ P , and (L0 , φ) is a join dense completion of P. Hence
we may reasonably restrict our attention to join dense completions.
Our first example of a join dense completion is the lattice of order ideals O(P).
Order ideals are the closed sets of the closure operator on P given by
[
O(A) = a/0,
a∈A
20
and the embedding φ is given by φ(p) = p/0. Note that the union of order ideals is
again an order ideal, so O(P) obeys the distributive law (D).
Another example is the MacNeille completion M(P), a.k.a. normal completion,
completion by cuts [2]. For subsets S, T ⊆ P recall that

S u = {x ∈ P : x ≥ s for all s ∈ S}
T ` = {y ∈ P : y ≤ t for all t ∈ T }.

The MacNeille completion is the lattice of closed sets of the closure operator on P
given by
M (A) = (Au )` ,
i.e., M (A) is the set of all lower bounds of all upper bounds of A. Note that M (A)
is an order ideal of P. Again the map φ(p) = p/0 embeds P into M(P).
Now every join dense completion V meets in P: if A ⊆ P and A
V preserves all existing
has a greatest lower bound b = A in P, then φ(b) = φ(A) (see Exercise 10). The
MacNeille completion has the nice property
W that it also preserves all existingW
joins in
P: if A has a least upper bound c = A in P, then φ(c) = c/0 = M (A) = φ(A).
In fact, every join dense completion corresponds to a closure operator on P .
Theorem 2.10. Let P be an ordered set. If Φ is a closure operator on P such that
Φ({p}) = p/0 for all p ∈ P , then (CΦ , φ) is a join dense completion of P, where
φ(p) = p/0. Conversely, if (L, φ) is a join dense completion of P, then the map Φ
defined by _
Φ(A) = {q ∈ P : φ(q) ≤ φ(a)}
a∈A

is a closure operator on P , Φ({p}) = p/0 for all p ∈ P , and CΦ ∼

= L.
Proof. For the first part, it is clear that (CΦ , φ) is a completion of P. It is a join
dense one because every closed set must be an order ideal, and thus for every C ∈ CΦ ,
_
C = {Φ({p}) : p ∈ C}
_
= {p/0 : p/0 ⊆ C}
_
= {φ(p) : φ(p) ≤ C}.

For the converse, it is clear that Φ defined W A ⊆ Φ(A), and A ⊆ B

W thusly satisfies
implies Φ(A) ⊆ Φ(B). But we also have q∈Φ(A) φ(q) = a∈A φ(a), so Φ(Φ(A)) =
Φ(A). W
To see that CΦ ∼= L, let f : CΦ → L by f (A) = a∈A φ(a), and let g : L → CΦ
by g(x) = {p ∈ P : φ(p) ≤ x}. Then both maps are order preserving, f g(x) = x for
x ∈ L by the definition of join density, and gf (A) = Φ(A) = A for A ∈ CΦ . Hence
both maps are isomorphisms. 
21
 Let K(P) be the collection of all closure operators on P such that Γ({p}) = p/0
for all p ∈ P . There is a natural order on K(P): Γ ≤ ∆ if Γ(A) ⊆ ∆(A) for all
A ⊆ P . Moreover, K(P) is closed under arbitrary meets, where by definition
^ \
( Γi )(A) = Γi (A).
i∈I i∈I

The least and greatest members of K(P) are the order ideal completion and the
MacNeille completion, respectively.
Theorem 2.11. K(P) is a complete lattice with least element O and greatest ele-
ment M .
Proof. The condition Γ({p}) = p/0 implies that O(A) ⊆ Γ(A) for all A ⊆ P , which
makes O the least element of K(P). On the other hand, for any Γ ∈ K(P), if b ≥ a
for all a ∈ A, then b/0 = Γ(b/0) ⊇ Γ(A). Thus
\
Γ(A) ⊆ (b/0) = (Au )` = M (A),
b∈Au

so M is its greatest element. 

The lattices K(P) have an interesting structure, which was investigated by the
author and Alex Pogel in [3].
We conclude this section with a classic theorem due to A. Tarski and Anne Davis
(Morel) [1], [4].
Theorem 2.12. A lattice L is complete if and only if every order preserving map
f : L → L has a fixed point.
Proof. One direction is easy. Given a complete lattice L and an order preserving
map f : W L → L, put A = {x ∈ L : f (x) ≥ x}. WNote A is nonempty W as 0 ∈ A.
Let a = A. Since a ≥ x for all x ∈ A, f (a) ≥ x∈A f (x) ≥ x∈A x = a. Thus
a ∈ A. But then a ≤ f (a) implies f (a) ≤ f 2 (a), so also f (a) ∈ A, whence f (a) ≤ a.
Therefore f (a) = a.
Conversely, let L be a lattice which is not a complete lattice.
Claim 1: Either L has no 1 or there exists a chain C ⊆ L which satisfies the
ACC and has no meet. For suppose L has a 1 and that every chain C in L satisfying
the ACC has a meet. We will show that every subset S ⊆ L has a join, which makes
L a complete lattice by the dual of Theorem 2.5.
Consider S u , the set of all upper bounds of S. Note S u 6= ∅ because 1 ∈ L. Let P
denote the collection of all chains C ⊆ S u satisfying the ACC, ordered by C1 ≤ C2
if C1 is a filter (dual ideal) of C2 . S
The order on P insures that if Ci (i ∈ I) is a chain of chains in P, then i∈I Ci ∈
P. Hence by Zorn’s Lemma, P contains a maximal element Cm . By hypothesis
22
V V W
Cm exists in VL, say Cm = a. In fact, a = S. For if s ∈ S, then s ≤ c for all
c ∈ Cm , so s ≤ Cm = a. Thus a ∈ S u , i.e., a is an upper bound for S. If a  t for
some t ∈ S u , then we would have a > a ∧ t ∈ S u , V
and theWchain Cm ∪ {a ∧ t} would
contradict the maximality of Cm . Therefore a = S u = S. This proves Claim 1;
Exercise 11 indicates why the argument is necessarily a bit involved.
If L has a 1, let C be a chain satisfying the ACC but having no meet; otherwise
take C = ∅. Dualizing the preceding argument, let Q be the set of all chains D ⊆ C `
satisfying the DCC, ordered by D1 ≤ D2 if D1 is an ideal of D2 . Now Q could be
empty, but only when C is not; if nonempty, Q has a maximal member Dm . Let
D = Dm if Q 6= ∅, and D = ∅ otherwise.
Claim 2: For all x ∈ L, either there exists c ∈ C with x  c, or there exists
d ∈ D with x  d. Supposing otherwise, let x ∈ L with x ≤ c for all c ∈ C and x ≥ d
for all d ∈ D.
V (The assumption x ∈ C ` means we are in the case Q 6= ∅.) Since
x ∈ C ` and C does not exist, there is a y ∈ C ` such that y  x. So x ∨ y > x ≥ d
for all d ∈ D, and the chain D ∪ {x ∨ y} contradicts the maximality of D = Dm in
Q.
Now define a map f : L → L as follows. For each x ∈ L, put

C(x) = {c ∈ C : x  c} ,
D(x) = {d ∈ D : x  d} .

We have shown that one of these two sets is nonempty for each x ∈ L. If C(x) 6= ∅,
let f (x) be its largest element (using the ACC); otherwise let f (x) be the least
element of D(x) (using the DCC). Now for any x ∈ L, either x  f (x) or x  f (x),
so f has no fixed point.
It remains to check that f is order preserving. Suppose x ≤ y. If C(x) 6= ∅ then
f (x) ∈ C and f (x)  y (else f (x) ≥ y ≥ x); hence C(y) 6= ∅ and f (y) ≥ f (x).
So assume C(x) = ∅, whence f (x) ∈ D. If perchance C(y) 6= ∅ then f (y) ∈ C, so
f (x) ≤ f (y). On the other hand, if C(y) = ∅ and f (y) ∈ D, then x  f (y) (else
y ≥ x ≥ f (y)), so again f (x) ≤ f (y). Therefore f is order preserving. 

Exercises for Chapter 2

1. Draw the Hasse diagrams for
(a) all 5 element (meet) semilattices,
(b) all 6 element lattices,
(c) the lattice of subspaces of the vector space <2 .
2. Prove that a lattice which has a 0 and satisfies the ACC is complete.
3. For the cyclic group Z4 , give explicitly the subgroup lattice, the closure oper-
ator Sg, and the closure rules for subgroups.
4. Define a closure operator F on <n by the rules {x, y} ⊆ S =⇒ tx+(1−t)y ∈ S
for all t ∈ <. Describe F (A). What is the geometric interpretation of F ?
23
 5. Prove that the following are equivalent for a subset Q of a complete lattice L.
W
(1) Q is join dense in L, i.e., x = {q ∈ Q : q ≤ x} for every x ∈ L.
(2) Every element of L is a join of elements in Q.
(3) If y < x in L, then there exists q ∈ Q with q ≤ x but q  y.
6. Find the completions O(P) and M(P) for the ordered sets in Figures 2.1 and
2.2.
7. Find the lattice K(P) of all join dense completions of the ordered sets in
Figures 2.1 and 2.2.
8. Show that the MacNeille operator satisfies M (A) = A iff A = B ` for some
B ⊆ P.
9. (a) Prove that if (L, φ)) is a join dense completion of the ordered set P, then
φ preserves all existing greatest lower bounds in P.
(b) Prove that the MacNeille completion preserves all existing least upper bounds
in P.
10. Prove that if φ is an order embedding of P into a complete lattice L, then φ
extends to an order embedding of M(P) into L.
11. Show that ω × ω1 has no cofinal chain. (A subset C ⊆ P is cofinal if for every
x ∈ P there exists c ∈ C with x ≤ c.)
12. Following Morgan Ward [5], we can generalize the notion of a closure operator
as follows. Let L be a complete lattice. (For the closure operators on a set X, L
will be P(X).) A closure operator on L is a function f : L → L which satisfies, for
all x, y ∈ L,
(i) x ≤ f (x),
(ii) x ≤ y implies f (x) ≤ f (y),
(iii) f (f (x)) = f (x).
(a) Prove that Cf = {x ∈ L : f (x) = x} is a complete meet subsemilattice of L.
(b) For any complete
V meet subsemilattice S of L, prove that the function fS
defined by fS (x) = {s ∈ S : s ≥ x} is a closure operator on L.
13. Let A and B be sets, and R ⊆ A × B a relation. For X ⊆ A and Y ⊆ B let

σ(X) = {b ∈ B : x R b for all x ∈ X}

π(Y ) = {a ∈ A : a R y for all y ∈ Y }.

Prove the following claims.

(a) X ⊆ πσ(X) and Y ⊆ σπ(Y ) for all X ⊆ A, Y ⊆ B.
(b) X ⊆ X 0 implies σ(X) ⊇ σ(X 0 ), and Y ⊆ Y 0 implies π(Y ) ⊇ π(Y 0 ).
(c) σ(X) = σπσ(X) and π(Y ) = πσπ(Y ) for all X ⊆ A, Y ⊆ B.
(d) πσ is a closure operator on A, and Cπσ = {π(Y ) : Y ⊆ B}. Likewise σπ is a
closure operator on B, and Cσπ = {σ(X) : X ⊆ A}.
(e) Cπσ is dually isomorphic to Cσπ .
24
The maps σ and π are said to establish a Galois connection between A and B. The
most familiar example is when A is a set, B a group acting on A, and a R b means
b fixes a. As another example, the MacNeille completion is Cπσ for the relation ≤
as a subset of P × P.

References
1. A. C. Davis, A characterization of complete lattices, Pacific J. Math. 5 (1955), 311–319.
2. H. M. MacNeille, Partially ordered sets, Trans. Amer. Math. Soc. 42 (1937), 90–96.
3. J. B. Nation and A. Pogel, The lattice of completions of an ordered set, Order 14 (1996), 1–7.
4. A. Tarski, A lattice-theoretical fixpoint theorem and its applications, Pacific J. Math. 5 (1955),
285–309.
5. M. Ward, The closure operators of a lattice, Annals of Math. 43 (1942), 191–196.

25
 3. Algebraic Lattices

The more I get, the more I want it seems ....

–King Oliver

In this section we want to focus our attention on the kind of closure operators
and lattices which are associated with modern algebra. A closure operator Γ on a
set X is said to be algebraic if for every B ⊆ X,
[
Γ(B) = {Γ(F ) : F is a finite subset of B}.

Equivalently, Γ is algebraic if the right hand side RHS of the above expression is
closed for every B ⊆ X, since B ⊆ RHS ⊆ Γ(B) holds for any closure operator.
A closure rule is said to be finitary if it is a rule of the form x ∈ S or the form
F ⊆ S =⇒ z ∈ S with F a finite set. Again the first form is a degenerate case of
the second, taking F = ∅. It is not hard to see that a closure operator is algebraic
if and only if it is determined by a set of finitary closure rules (see Theorem 3.1(1)).
Let us catalogue some important examples of algebraic closure operators.
(1) Let A be any algebra with only finitary operations – for example, a group,
ring, vector space, semilattice or lattice. The closure operator Sg on A such that
Sg(B) is the subalgebra of A generated by B is algebraic, because we have a ∈ Sg(B)
if and only if a can be expressed as a term a = t(b1 , . . . , bn ) for some finite subset
{b1 , . . . , bn } ⊆ B, in which case a ∈ Sg({b1 , . . . , bn }). The corresponding complete
lattice is of course the subalgebra lattice Sub A.
(2) Looking ahead a bit (to Chapter 5), the closure operator Cg on A × A such
that Cg(B) is the congruence on A generated by the set of pairs B is also algebraic.
The corresponding complete lattice is the congruence lattice Con A. For groups
this is isomorphic to the normal subgroup lattice; for rings, it is isomorphic to the
lattice of ideals.
(3) For ordered sets, the order ideal operator O is algebraic. In fact we have
[
O(B) = {O({b}) : b ∈ B}

for all B ⊆ P .
(4) Let S = (S; ∨) be a join semilattice with a least element 0. A subset J of S
is called an ideal if
(1) 0 ∈ J,
(2) x, y ∈ J implies x ∨ y ∈ J,
(3) z ≤ x ∈ J implies z ∈ J.
26
An ideal J is a principal ideal if J = x/0 for some x ∈ S. Since ideals are defined
by closure rules, the intersection of a set of ideals of S is again one.1 The closure
operator I on S such that I(B) is the ideal of S generated by B is given by
_
I(B) = {x ∈ S : x ≤ F for some finite F ⊆ B}.

Hence I is algebraic. The ideal lattice of a join semilattice is denoted by I(S).

(5) An ideal of a lattice is defined in the same way, since every lattice is in
particular a join semilattice. The ideal lattice of a lattice L is likewise denoted by
I(L). The dual of an ideal in a lattice is called a filter. (See Exercise 4.)
On the other hand, it is not hard to see that the closure operators associated with
the closed sets of a topological space are usually not algebraic, since the closure de-
pends on infinite sequences. The closure operator M associated with the MacNeille
completion is not in general algebraic, as is seen by considering the partially ordered
set P consisting of an infinite set X and all of its finite subsets, ordered by set con-
tainment. This ordered set is already a complete lattice, and hence its own MacNeille
completion. For any subset Y ⊆ X, let Yb = {S ∈ P : S ⊆ Y }. If Y is an infinite
S
proper subset of X, then M (Yb ) = X b ⊃ Yb = {M (F ) : F is a finite subset of Yb }.
We need to translate these ideas into the language ofWlattices. Let L be a complete
lattice. An element x ∈ L is W compact if whenever x ≤ A, then there exists a finite
subset F ⊆ A such that x ≤ F . The set of all compact elements of L is denoted by
Lc . An elementary argument shows that Lc is closed under finite joins and contains
0, so it is a join semilattice with a least element. However, Lc is usually not closed
under meets (see Figure 3.1(a), wherein x and y are compact but x ∧ y is not).
A lattice L is said to be algebraic
W (or compactly generated) if it is complete and Lc
is join dense in L, i.e., x = (x/0 ∩ Lc ) for every x ∈ L. Clearly every finite lattice
is algebraic. More generally, every element of a complete lattice L is compact, i.e.,
L = Lc if and only if L satisfies the ACC.
For an example of a complete lattice which is not algebraic, let K denote the
interval [0, 1] in the real numbers with the usual order. Then Kc = {0}, so K is
not algebraic. The non-algebraic lattice in Figure 3.1(b) is another good example
to keep in mind. (The element z is not compact, and hence in this case not a join
of compact elements.)
Theorem 3.1. (1 ) A closure operator Γ is algebraic if and only if Γ = ΓΣ for
some set Σ of finitary closure rules.
(2 ) Let Γ be an algebraic closure operator on a set X. Then CΓ is an algebraic
lattice whose compact elements are {Γ(F ) : F is a finite subset of X}.
Proof. If Γ is an algebraic closure operator on a set X, then a set S ⊆ X is closed if
and only if Γ(F ) ⊆ S for every finite subset F ⊆ S. Thus the collection of all rules
1 If S has no least element, then it is customary to allow the empty set as an ideal; however,

this convention is not universal.

27
 x y

(a) (b)
Figure 3.1

F ⊆ S =⇒ z ∈ S, with F a finite subset of X and z ∈ Γ(F ), determines closure

for Γ.2 Conversely, if Σ is a collection of finitary closure rules, then z ∈ ΓΣ (B) if
and only if z ∈ ΓΣ (F ) for some finite F ⊆ B, making ΓΣ algebraic.
For (2), let us first observe that for S
any closure
W operator Γ on X, and for any
collection of subsets Ai of X, we haveSΓ( Ai ) W= Γ(Ai ) where the join is computed
S
in
S the lattice
W C Γ . The inclusion
S Γ( A i )
W ⊇ Γ(A i )
W is immediate, while Ai ⊆
Γ(Ai ) ⊆ Γ(Ai ) implies Γ( Ai ) ⊆ Γ( Γ(Ai )) = Γ(Ai ).
Now assume that Γ is algebraic. Then, for all B ⊆ X,
[
Γ(B) = {Γ(F ) : F is a finite subset of B}
_
⊆ {Γ(F ) : F is a finite subset of B}
= Γ(B),

from which equality follows. Thus CΓ will be an algebraic lattice if we can show that
the closures of finite sets are compact. W
Let F be a finite subset of X. If Γ(F ) ≤ Ai in CΓ , then
_ [ [ [
F ⊆ Ai = Γ( Ai ) = {Γ(G) : G finite ⊆ Ai }.

Consequently each x ∈ F is in some Γ(Gx ), where S Gx is in turnWcontained in the

union of finitely many Ai ’s. Therefore Γ(F ) ⊆ Γ( x∈F Γ(Gx )) ⊆ j∈J Aj for some
finite subset J ⊆ I. We conclude that Γ(F ) is compact in CΓ .
W Conversely, let C be compact in CΓ . Since C is closed and Γ is algebraic, C =
{Γ(F ) : F finite ⊆ C}. Since C is compact, there exist finitely many finite subsets
2 In general there are also valid infinitary closure rules for Γ, but for algebraic closure operators

these are redundant.

28
of C, say F1 , . . . , Fn , such that C = Γ(F1 ) ∨ . . . ∨ Γ(Fn ) = Γ(F1 ∪ · · · ∪ Fn ). Thus
C is the closure of a finite set. 
Thus in a subalgebra lattice Sub A, the compact elements are the finitely gen-
erated subalgebras. In a congruence lattice Con A, the compact elements are the
finitely generated congruences.
It is not true that CΓ being algebraic implies that Γ is algebraic. For example,
let X be the disjoint union of a one element set {b} and an infinite set Y , and let
Γ be the closure operator on X such that Γ(A) = A if A is a proper subset of Y ,
Γ(Y ) = X and Γ(B) = X if b ∈ B.
The following theorem includes a representation of any algebraic lattice as the
lattice of closed sets of an algebraic closure operator.
Theorem 3.2. If S is a join semilattice with 0, then the ideal lattice I(S) is al-
gebraic. The compact elements of I(S) are the principal ideals x/0 with x ∈ S.
Conversely, if L is an algebraic lattice, then Lc is a join semilattice with 0, and
L∼= I(Lc ).
Proof. Let S be a join semilattice with 0. I is an algebraicWclosure operator, so I(S)
is an algebraic lattice. If F ⊆ S is finite, then I(F ) = ( F )/0, so compact ideals
are principal.
Now let L be an algebraic lattice. There are two natural
W maps: f : L → I(Lc )
by f (x) = x/0 ∩ Lc , and g : I(Lc ) → L by g(J) = J. Both W maps are clearly
order preserving, and they are mutually inverse:W f g(J) = ( J)/0 ∩ L c
= J by
the definition of compactness, and gf (x) = (x/0 ∩ Lc ) = x by the definition of
algebraic. Hence they are both isomorphisms, and L ∼ = I(Lc ). 
Let us digress for a moment into universal algebra. A classic result of Birkhoff
and Frink gives a concrete representation of algebraic closure operators [2].
Theorem 3.3. Let Γ be an algebraic closure operator on a set X. Then there is an
algebra A on the set X such that the subalgebras of A are precisely the closed sets
of Γ.
Corollary. Every algebraic lattice is isomorphic to the lattice of all subalgebras of
an algebra.
Proof. An algebra in general is described by A = hA; F, Ci where A is a set, F =
{fi : i ∈ I} a collection of operations on A (so fi : Ani → A), and C is a set of
constants in A. Appendix 3 reviews the basic definitions of universal algebra.
The carrier set for our algebra must of course be X. For each nonempty finite
set F ⊆ X and element x ∈ Γ(F ), we have an operation fF,x : X |F | → X given by

x if {a1 , . . . , an } = F
fF,x (a1 , . . . , an ) =
a1 otherwise.
29
Our constants are C = Γ(∅), the elements of the least closed set (which may be
empty).
Note that since Γ is algebraic, a set B ⊆ X is closed if and only if Γ(F ) ⊆ B for
every finite F ⊆ B. Using this, it is very easy to check that the subalgebras of A
are precisely the closed sets of CΓ . 
However, the algebra constructed in the proof of Theorem 3.3 will have |X|
operations when X is infinite. Having lots of operations is not necessarily a bad
thing: vector spaces are respectable algebras, and a vector space over a field F has
basic operations fr : V → V where fr (v) = rv for every r ∈ F . Nonetheless, we
like algebras to have few operations, like groups and lattices. A theorem due to Bill
Hanf tells us when we can get by with a small number of operations.3
Theorem 3.4. For any nontrivial algebraic lattice L the following conditions are
equivalent.
(1) Each compact element of L contains only countably many compact elements.
(2) There exists an algebra A with only countably many operations and constants
such that L is isomorphic to the subalgebra lattice of A.
(3) There exists an algebra B with one binary operation (and no constants) such
that L is isomorphic to the subalgebra lattice of B.

Proof. Of course (3) implies (2).

In general, if an algebra A has κ basic operations, λ constants and γ generators,
then it is a homomorphic image of the absolutely free algebra W (X) generated by
a set X with |X| = γ and the same κ operation symbols and λ constants. It is easy
to count that |W (X)| = max(γ, κ, λ, ℵ0 ), and |A| ≤ |W (X)|.
In particular then, if C is compact (i.e., finitely generated) in Sub A, and A has
only countably many basic operations and constants, then |C| ≤ ℵ0 . Therefore C has
only countably many finite subsets, and so there are only countably many finitely
generated subalgebras D contained in C. Thus (2) implies (1).
To show (1) implies (3), let L be a nontrivial algebraic lattice such that for each
x ∈ Lc , |x/0 ∩ Lc | ≤ ℵ0 . We will construct an algebra B whose universe is Lc − {0},
with one binary operation ∗, whose subalgebras are precisely the ideals of Lc with
0 removed. This makes Sub B ∼ = I(Lc ) ∼
= L, as desired.
For each c ∈ Lc − {0}, we make a sequence hci ii∈ω as follows. If 2 ≤ |c/0 ∩ Lc | =
n + 1 < ∞, arrange c/0 ∩ Lc − {0} into a cyclically repeating sequence: ci = cj iff
i ≡ j mod n. If c/0 ∩ Lc is infinite (and hence countable), arrange c/0 ∩ Lc − {0}
into a non-repeating sequence hci i. In both cases start the sequence with c0 = c.
Define the binary operation ∗ for c, d ∈ Lc − {0} by

c ∗ d = c ∨ d if c and d are incomparable,

c ∗ d = d ∗ c = ci+1 if d = ci ≤ c.
3 This result is unpublished but well known.
30
You can now check that ∗ is well defined, and that the algebra B = hLc ; ∗i has
exactly the sets of nonzero elements of ideals of Lc as subalgebras. 
The situation with respect to congruence lattices is considerably more compli-
cated. Nonetheless, the basic facts are the same: George Grätzer and E. T. Schmidt
proved that every algebraic lattice is isomorphic to the congruence lattice of some
algebra [6], and Bill Lampe showed that uncountably many operations may be re-
quired [5].
Ralph Freese and Walter Taylor modified Lampe’s original example to obtain a
very natural one. Let V be a vector space of countably infinite dimension over a
field F with |F | = κ > ℵ0 . Let L be the congruence lattice Con V, which for vector
spaces is isomorphic to the subspace lattice Sub V (since homomorphisms on vector
spaces are linear transformations, and any subspace of V is the kernel of a linear
transformation). The representation we have just given for L involves κ operations
fr (r ∈ F ). In fact, one can show that any algebra A with Con A ∼ = L must have
at least κ operations.
We now turn our attention to the structure of algebraic lattices. The lattice L is
said to be weakly atomic if whenever a > b in L, there exist elements u, v ∈ L such
that a ≥ u  v ≥ b.
Theorem 3.5. Every algebraic lattice is weakly atomic.
Proof. Let a > b in an algebraic lattice L. Then there is a compact element c ∈ Lc
with c ≤ a and c 6≤ b. Let P = {x ∈ a/b : c 6≤ x}. Note b ∈ P , and since c is
compact the join of a chain in P is again in P. Hence by Zorn’s Lemma, P contains
a maximal element v, and the element u = c ∨ v covers v. Thus b ≤ v ≺ u ≤ a. 
A lattice L is said to be upper continuous
W W if L is complete and, for every element
a ∈ L and every chain C in L, a ∧ C = c∈C a ∧ c.
Theorem 3.6. Every algebraic lattice is upper continuous.
W W
Proof. Let L W
be algebraic and C a chain in L. Of course c∈C (a ∧ c) ≤ W a ∧ C.
Let r = a ∧ C. For each d ∈ r/0 ∩ Lc , we have d ≤ a and d ≤ C. The
compactness
W of d implies
W d ≤ cd for some cd ∈ C, and hence d ≤ a ∧ cd . But then
r = (r/0 ∩ Lc ) ≤ c∈C a ∧ c, as desired. 
Two alternative forms of join continuity are often useful. An ordered set P is
said to be up-directed if for every x, y ∈ P there exists z ∈ P with x ≤ z and y ≤ z.
So, for example, any join semilattice is up-directed.
Theorem 3.7. For a complete lattice L, the following are equivalent.
(1) L is upper continuous. W W
(2) For every a ∈ L and up-directed set D ⊆ L, a ∧ D = d∈D a ∧ d.
(3) For every a ∈ L and S ⊆ L,
_ _ _
a∧ S = (a ∧ F ).
F finite ⊆S
31
Proof. It is straightforward that (3) implies (2) implies (1); we will show that (1)
implies (3) by induction on |S|. Property (3) is trivial if |S| is finite, so assume it is
infinite, and let λ be the least ordinal with |S| = |λ|. Arrange the elements of S into
a sequence hxξ : ξ < λi. PutW Sξ = {xν : ν < ξ}. Then |Sξ | < |S| for each ξ < λ, and
the elements of the form Sξ are a chain in L. Thus, using (1), we can calculate
_ __
a∧ S =a∧ Sξ
ξ<λ
_ _
= (a ∧ Sξ )
ξ<λ
_ _ _
= ( (a ∧ F ))
ξ<λ F finite ⊆Sξ
_ _
= (a ∧ F ),
F finite ⊆S

as desired. 
An element a ∈ L is called an atom if a  0, and a coatom if 1  a. Theorem 3.7
shows that every atom in an upper continuous lattice is compact. More generally,
if a/0 satisfies the ACC in an upper continuous lattice, then a is compact.
We know that every element x in an algebraic lattice can be expressed as the
join of x/0 ∩ Lc (by definition). It turns out to be at least as important to know
how x can be expressed as a meet of other elements. We say that an element q V in a
complete lattice L is completely meet irreducible if, for every subset S of L, q = S
implies q ∈ S. These are of course the elements which cannot be expressed as the
proper meet of other elements. Let M ∗ (L) denote theVset of all completely meet
irreducible elements of L. Note that 1 6∈ M ∗ (L) (since ∅ = 1 and 1 6∈ ∅).
Theorem 3.8. Let q ∈ L where L is a complete lattice. The following are equiva-
lent.
(1) Vq ∈ M ∗ (L).
(2) {x ∈ L : x > q} > q.
(3) There exists q ∗ ∈ L such that q ∗  q and for all x ∈ L, x > q implies x ≥ q ∗ .
V
The connection between (2) and (3) is of course q ∗ = {x ∈ L : x > q}. In a
finite lattice, q ∈ M ∗ (L) iff there is a unique element q ∗ covering q, but in general
we need the stronger property (3). V
A decomposition of an element a ∈ L is a representation a = Q where Q is
a set of completely meet irreducible elements of L. An element in an arbitrary
lattice may have any number of decompositions, including none. A theorem due
to Garrett Birkhoff says that every element in an algebraic lattice has at least one
decomposition [1].
32
Theorem 3.9. If L V is an algebraic lattice, then M ∗ (L) is meet dense in L. Thus
for every x ∈ L, x = (1/x ∩ M ∗ (L)).
V
Proof. Let m = (1/x ∩ M ∗ (L)), and suppose x < m. Then there exists a c ∈ Lc
with c ≤ m and c 6≤ x. Since c is compact, we can use Zorn’s Lemma to find
an element q which is maximal with respect to q ≥ x, q 6≥ c. For any y ∈ L,
y > q implies y ≥ q ∨ c, so q is completely meet irreducible with q ∗ = q ∨ c. Then
q ∈ 1/x ∩ M ∗ (L) implies q ≥ m ≥ c, a contradiction. Hence x = m. 
It is rare for an element in an algebraic lattice to have a unique decomposition. A
somewhat weaker V propertyVis for an element to have an irredundant decomposition,
meaning a = Q but a < (Q − {q}) for all q ∈ Q, where Q is a set of completely
meet irreducible elements. An element in an algebraic lattice need not have an
irredundant decomposition either. Let L be the lattice consisting of the empty
set and all cofinite subsets of an infinite set X, ordered by set inclusion. This
satisfies the ACC so it is algebraic. The completely meet irreducible elements of
L are its coatoms, the complements of one element subsets of X. The meet of
any infinite collection of coatoms is 0 (the empty set), but no such decomposition
is irredundant. Clearly also these are the only decompositions of 0, so 0 has no
irredundant decomposition.
A lattice is strongly atomic if a > b in L implies there exists u ∈ L such that a ≥
u  b. A beautiful result of Peter Crawley guarantees the existence of irredundant
decompositions in strongly atomic algebraic lattices [3].
Theorem 3.10. If an algebraic lattice L is strongly atomic, then every element of
L has an irredundant decomposition.
If L is also distributive, we obtain the uniqueness of irredundant decompositions.
Theorem 3.11. If L is a distributive, strongly atomic, algebraic lattice, then every
element of L has a unique irredundant decomposition.
The finite case of Theorem 3.11 is the dual of Theorem 8.6(c), which we will
prove later.
The theory of decompositions was studied extensively by Dilworth and Crawley,
and their book [4] contains most of the principal results.

Exercises for Chapter 3

1. Prove that anW upper continuous
W distributive lattice satisfies the infinite dis-
tributive law a ∧ ( i∈I bi ) = i∈I (a ∧ bi ).
2. Describe the complete sublattices of the real numbers < which are algebraic.
3. Show that the natural map from a lattice to its ideal lattice, ϕ : L → I(L) by
ϕ(x) = x/0, is a lattice embedding. Show that (I(L), ϕ) is a join dense completion
of L, and that it may differ from the MacNeille completion.
33
 4. Recall that a filter is a dual ideal. The filter lattice F(L) of a lattice L is
ordered by reverse set inclusion: F ≤ G iff F ⊇ G. Prove that L is naturally
embedded in F(L), and that F(L) is dually compactly generated.
5. Prove that every element of a complete lattice L is compact if and only if L
satisfies the ACC. (Cf. Exercise 2.2.) W
V 6. A subset S of a complete lattice L is a complete sublattice if A ∈ S and
A ∈ S for every nonempty subset A ⊆ S. Prove that a complete sublattice of an
algebraic lattice is algebraic.
7. (a) Represent the lattices M3 and N5 as Sub A for a finite algebra A.
(b) Show that M3 ∼ = Sub G for a (finite) group G, but that N5 cannot be so
represented.
8. A closure rule is unary if it is of the form x ∈ C =⇒ y ∈ C. Prove that if Σ is
a collection of unary closure rules, then unions of closed sets are closed, and hence
the lattice of closed sets CΣ is distributive. Conclude that the subalgebra lattice of
an algebra with only unary operations is distributive.
9. Let L be a complete lattice, J a join dense subset of L and M a meet dense
subset of L. Define maps σ : P(J) → P(M ) and π : P(M ) → P(J) by
σ(X) = X u ∩ M
π(Y ) = Y ` ∩ J.
By Exercise 2.13, with R the restriction of ≤ to J × M , πσ is a closure operator on
J and σπ is a closure operator on M . Prove that Cπσ ∼ = L and that Cσπ is dually
isomorphic to L.
10. A lattice is semimodular if a  a ∧ b implies a ∨ b  b. Prove that if every
element of a finite lattice L has a unique irredundant decomposition, then L is
semimodular. (Morgan Ward)V V
11. A decomposition a = Q is strongly irredundant if a < q ∗ ∧ (Q − {q})
for all q ∈ Q. Prove that every irredundant decomposition in a strongly atomic
semimodular lattice is strongly irredundant. (Keith Kearnes)
12. Let L be the lattice of ideals of the ring of integers Z. Find M ∗ (L) and all
decompositions of 0.

References
1. G. Birkhoff, Subdirect unions in universal algebra, Bull. Amer. Math. Soc. 50 (1944), 764–768.
2. G. Birkhoff and O. Frink, Representations of sets by lattices, Trans. Amer. Math. Soc. 64
(1948), 299–316.
3. P. Crawley, Decomposition theory for nonsemimodular lattices, Trans. Amer. Math. Soc. 99
(1961), 246–254.
4. P. Crawley and R. P. Dilworth, Algebraic Theory of Lattices, Prentice-Hall, Englewood Cliffs,
N. J., 1973.
5. R. Freese, W. A. Lampe and W. Taylor, Congruence lattices of algebras of fixed similarity
type, I, Pacific J. Math. 82 (1979), 59–68.
6. G. Grätzer and E. T. Schmidt, Characterizations of congruence lattices of abstract algebras,
Acta Sci. Math. (Szeged) 24 (1963), 34–59.
34
 4. Representation by Equivalence Relations

No taxation without representation!

So far we have no analogue for lattices of the Cayley theorem for groups, that
every group is isomorphic to a group of permutations. The corresponding represen-
tation theorem for lattices, that every lattice is isomorphic to a lattice of equivalence
relations, turns out to be considerably deeper. Its proof uses a recursive construction
technique which has become a standard tool of lattice theory and universal algebra.
An equivalence relation on a set X is a binary relation E satisfying, for all x, y, z ∈
X,
(1) x E x,
(2) x E y implies y E x,
(3) if x E y and y E z, then x E z.
We think of an equivalence relation as partitioning the set X into blocks of E-related
elements, called equivalence classes. Conversely, any partition of X into a disjoint
union of blocks induces an equivalence relation on X: x E y iff x and y are in the
same block. As usual with relations, we write x E y and (x, y) ∈ E interchangeably.
The most important equivalence relations are those induced by maps. If Y is
another set, and f : X → Y is any function, then

ker f = {(x, y) ∈ X 2 : f (x) = f (y)}

is an equivalence relation, called the kernel of f . If X and Y are algebras and

f : X → Y is a homomorphism, then ker f is a congruence relation.
Thinking of binary relations as subsets of X 2 , the axioms (1)–(3) for an equiv-
alence relation are finitary closure rules. Thus the collection of all equivalence
relations on X forms an algebraic lattice Eq X, with the order given by R ≤ S iff
(x, y) ∈ R =⇒ (x, y) ∈ S iff R ⊆ S in P(X 2 ). The greatest element of Eq X is
the universal relation X 2 , and its least element is the equality relation =. The
V meet
operation in Eq X is of course set intersection,W which means that (x, y) ∈ i∈I Ei
if and onlySif x Ei y for all i ∈ I.WThe join i∈I Ei is the transitive closure of the
set union i∈I Ei . Thus (x, y) ∈ Ei if and only if there exists a finite sequence of
elements xj and indices ij such that

x = x0 Ei1 x1 Ei2 x2 . . . xk−1 Eik xk = y.

35
The lattice Eq X has many nice properties: it is algebraic, strongly atomic, semi-
modular, relatively complemented and simple.1 The proofs of these facts are exer-
cises in this chapter and Chapter 11.
If R and S are relations on X, define the relative product R ◦ S to be the set of
all pairs (x, y) ∈ X 2 for which there exists a z ∈ X with x R z S y. If R and S are
equivalence relations, then because x R x we have S ⊆ R ◦ S; similarly R ⊆ R ◦ S.
Thus
R ◦ S ⊆ R ◦ S ◦ R ⊆ R ◦ S ◦ R ◦ S ⊆ ···
and it is not hard to see that R ∨ S is the union of this chain. It is possible, however,
that R ∨ S is in fact equal to some term in the chain; for example, this is always the
case when X is finite. Our proof will yield a representation in which this is always
the case, for any two equivalence relations which represent elements of the given
lattice.
To be precise, a representation of a lattice L is an ordered pair (X, F ) where X
is a set and F : L → Eq X is a lattice embedding. We say that the representation
is
(1) of type 1 if F (x) ∨ F (y) = F (x) ◦ F (y) for all x, y ∈ L,
(2) of type 2 if F (x) ∨ F (y) = F (x) ◦ F (y) ◦ F (x) for all x, y ∈ L,
(3) of type 3 if F (x) ∨ F (y) = F (x) ◦ F (y) ◦ F (x) ◦ F (y) for all x, y ∈ L.
P. M. Whitman [7] proved in 1946 that every lattice has a representation. In 1953
Bjarni Jónsson [4] found a simpler proof which gives a slightly stronger result.
Theorem 4.1. Every lattice has a type 3 representation.
Proof. Given a lattice L, we will use transfinite recursion to construct a type 3
representation of L.
A weak representation of L is a pair (U, F ) where U is a set and F : L → Eq U
is a one-to-one meet homomorphism. Let us order the weak representations of L by
(U, F ) v (V, G) if U ⊆ V and G(x) ∩ U 2 = F (x) for all x ∈ L.
We want to construct a (transfinite) sequence (Uξ , Fξ )ξ<λ of weak representations
of L, with (Uα , Fα ) v (UβS, Fβ ) whenever α ≤ β, whose limit (union) will be a lattice
embedding of L into Eq ξ<λ Uξ . We can begin our construction by letting (U0 , F0 )
be the weak representation with U0 = L and (y, z) ∈ F0 (x) iff y = z or y ∨ z ≤ x.
The crucial step is where we fix up the joins one at a time.
Sublemma 1. If (U, F ) is a weak representation of L and (p, q) ∈ F (x ∨ y), then
there exists (V, G) w (U, F ) with (p, q) ∈ G(x) ◦ G(y) ◦ G(x) ◦ G(y).
Proof of Sublemma 1. Form V by adding three new points to U , say V = U ∪˙ {r, s, t},
as in Figure 4.1. We want to make
p G(x) r G(y) s G(x) t G(y) q.
1 The terms relatively complemented and simple are defined in Chapter 10; we include them

here for the sake of completeness.

36
Accordingly, for z ∈ L we define G(z) to be the reflexive, symmetric relation on U
satisfying, for u, v ∈ U ,
(1) u G(z) v iff u F (z) v,
(2) u G(z) r iff z ≥ x and u F (z) p,
(3) u G(z) s iff z ≥ x ∨ y and u F (z) p,
(4) u G(z) t iff z ≥ y and u F (z) q,
(5) r G(z) s iff z ≥ y,
(6) s G(z) t iff z ≥ x,
(7) r G(z) t iff z ≥ x ∨ y.
You must check that each G(z) defined thusly really is an equivalence relation, i.e.,
that it is transitive. This is routine but a bit tedious to write down, so we leave
it to the reader. There are four cases, depending on whether or not z ≥ x and on
whether or not z ≥ y. Straightforward though it is, this verification would not work
if we had only added one or two new elements between p and q; see Theorems 4.5
and 4.6.

G(y) G(x)
r t
G(x) G(y)

p q
F (x ∨ y)

Figure 4.1

Now (1) says that G(z) ∩ U 2 = F (z). Since F is one-to-one, this implies G
is also. Note that for z, z 0 ∈ L we have z ∧ z 0 ≥ x iff z ≥ x and z 0 ≥ x, and
symmetrically for y. Using this with conditions (1)–(7), it is not hard to check that
G(z ∧ z 0 ) = G(z) ∩ G(z 0 ). Hence, G is a weak representation of L, and clearly
(U, F ) v (V, G). 
Sublemma 2. Let λ be a limit ordinal, and for ξ < λ let (Uξ , Fξ ) be weak S repre-
sentations of L such that α < β < λ implies (Uα , Fα ) v (Uβ , Fβ ). Let V = ξ<λ Uξ
S
and G(x) = ξ<λ Fξ (x) for all x ∈ L. Then (V, G) is a weak representation of L
with (Uξ , Fξ ) v (V, G) for each ξ < λ.
37
Proof. Let ξ < λ. Since Fα (x) = Fξ (x) ∩ Uα2 ⊆ Fξ (x) whenever α < ξ and Fξ (x) =
Fβ (x) ∩ Uξ2 whenever β ≥ ξ, for all x ∈ L we have
[

G(x) ∩ Uξ2 = Fγ (x) ∩ Uξ2
γ<λ
[
= (Fγ (x) ∩ Uξ2 )
γ<λ
= Fξ (x).

Thus (Uξ , Fξ ) v (V, G). Since F0 is one-to-one, this implies that G is also.
It remains to show that G is a meet homomorphism. Clearly G preserves order,
so for any x, y ∈ L we have G(x ∧ y) ⊆ G(x) ∩ G(y). On the other hand, if
(u, v) ∈ G(x) ∩ G(y), then there exists α < λ such that (u, v) ∈ Fα (x), and there
exists β < λ such that (u, v) ∈ Fβ (y). If γ is the larger of α and β, then (u, v) ∈
Fγ (x) ∩ Fγ (y) = Fγ (x ∧ y) ⊆ G(x ∧ y). Thus G(x) ∩ G(y) ⊆ G(x ∧ y). Combining
the two inclusions gives equality. 
Now we want to use these two sublemmas to construct a type 3 representation of
L, i.e., a weak representation which also satisfies G(x∨y) = G(x)◦G(y)◦G(x)◦G(y).
Start with an arbitrary weak representation (U0 , F0 ), and consider the set of all
quadruples (p, q, x, y) such that p, q ∈ U0 and x, y ∈ L and (p, q) ∈ F0 (x ∨ y).
Arrange these into a well ordered sequence (pξ , qξ , xξ , yξ ) for ξ < η. Applying the
sublemmas repeatedly, we can obtain a sequence of weak representations (Uξ , Fξ )
for ξ ≤ η such that
(1) if ξ < η, then (Uξ , Fξ ) v (Uξ+1 , Fξ+1 ) and (pξ , qξ ) ∈ Fξ+1 (xξ ) ◦ Fξ+1 (yξ ) ◦
Fξ+1 (xξ ) ◦ Fξ+1 (yξ ); S S
(2) if λ ≤ η is a limit ordinal, then Uλ = ξ<λ Uξ and Fλ (x) = ξ<λ Fξ (x) for
all x ∈ L.
Let V1 = Uη and G1 = Fη . If p, q ∈ U0 , and x, y ∈ L and p F0 (x ∨ y) q, then
(p, q, x, y) = (pξ , qξ , xξ , yξ ) for some ξ < η, so that (p, q) ∈ Fξ+1 (x) ◦ Fξ+1 (y) ◦
Fξ+1 (x) ◦ Fξ+1 (y). Consequently,

F0 (x ∨ y) ⊆ G1 (x) ◦ G1 (y) ◦ G1 (x) ◦ G1 (y).

Note (U0 , F0 ) v (V1 , G1 ).

Of course, along the way we have probably introduced lots of new failures of the
join property which need to be fixed up. So repeat this whole process ω times,
obtaining a sequence

(U0 , F0 ) = (V0 , G0 ) v (V1 , G1 ) v (V2 , G2 ) v · · ·

such that Gn (x∨y) ⊆ Gn+1 (x)◦Gn+1 (y)◦Gn+1 (x)◦Gn+1 (y) for all n ∈ w, x, y ∈ L.
38
 S S
Finally, let W = n∈w Vn and H(x) = n∈ω Gn (x) for all x ∈ L, and you get a
type 3 representation of L. 
Since the proof involves transfinite recursion, it produces a representation (X, F )
with X infinite, even when L is finite. For a long time one of the outstanding
questions of lattice theory was whether every finite lattice can be embedded into
the lattice of equivalence relations on a finite set. In 1980, Pavel Pudlák and Jı́ři
Tůma showed that the answer is yes [6]. The proof is quite difficult.
Theorem 4.2. Every finite lattice has a representation (Y, G) with Y finite.
One of the motivations for Whitman’s theorem was Garrett Birkhoff’s observa-
tion, made in the 1930’s, that a representation of a lattice L by equivalence relations
induces an embedding of L into the lattice of subgroups of a group. Given a rep-
resentation (X, F ) of L, let G be the group of all permutations on X which move
only finitely many elements, and let Sub G denote the lattice of subgroups of G.
Let h : L → Sub G by

h(a) = {π ∈ G : x F (a) π(x) for all x ∈ X}.

Then it is not too hard to check that h is an embedding.

Theorem 4.3. Every lattice can be embedded into the lattice of subgroups of a
group.
Not all lattices have representations of type 1 or 2, so it is natural to ask which
ones do. First we consider sublattices of Eq X with type 2 joins.
Lemma 4.4. Let L be a sublattice of Eq X with the property that R ∨ S = R ◦S ◦R
for all R, S ∈ L. Then L satisfies

(M ) x≥y implies x ∧ (y ∨ z) = y ∨ (x ∧ z).

The implication (M ) is known as the modular law.

Proof. Assume that L is a sublattice of Eq X with type 2 joins, and let A, B, C ∈ L
with A ≥ B. If p, q ∈ X and (p, q) ∈ A ∧ (B ∨ C), then

pAq
pBrCsBq

for some r, s ∈ X (see Figure 4.2). Since

rBpAqBs

and B ≤ A, we have (r, s) ∈ A ∧ C. It follows that (p, q) ∈ B ∨ (A ∧ C). Thus

A ∧ (B ∨ C) ≤ B ∨ (A ∧ C). The reverse inclusion is trivial, so we have equality. 
39
 C
r s

B B

p q
A
Figure 4.2

On the other hand, Jónsson gave a slight variation of the proof of Theorem
4.1 which shows that every modular lattice has a type 2 representation [4], [1].
Combining this with Lemma 4.4, we obtain the following.
Theorem 4.5. A lattice has a type 2 representation if and only if it is modular.
The modular law (M ) plays an important role in lattice theory, and we will see
it often. Note that (M ) fails in the pentagon N5 . It was invented in the 1890’s by
Richard Dedekind, who showed that the lattice of normal subgroups of a group is
modular. The modular law is equivalent to the equation,

(M 0 ) x ∧ ((x ∧ y) ∨ z) = (x ∧ y) ∨ (x ∧ z).

It is easily seen to be a special case of (and hence weaker than) the distributive law,

(D) x ∧ (y ∨ z) = (x ∧ y) ∨ (x ∧ z),

viz., (M ) says that (D) should hold for x ≥ y.

Note that the normal subgroup lattice of a group has a natural representation
(X, F ): take X = G and F (N ) = {(x, y) ∈ G2 : xy −1 ∈ N }. This representation is
in fact type 1 (Exercise 3), and Jónsson showed that lattices with a type 1 repre-
sentation, or equivalently sublattices of Eq X in which R ∨ S = R ◦ S, satisfy an
implication stronger than the modular law. A lattice is said to be Arguesian if it
satisfies

(A) (a0 ∨ b0 ) ∧ (a1 ∨ b1 ) ≤ a2 ∨ b2 implies c2 ≤ c0 ∨ c1

where
ci = (aj ∨ ak ) ∧ (bj ∨ bk )
for {i, j, k} = {0, 1, 2}. The Arguesian law is (less obviously) equivalent to a lattice
inclusion,

(A0 ) (a0 ∨ b0 ) ∧ (a1 ∨ b1 ) ∧ (a2 ∨ b2 ) ≤ a0 ∨ (b0 ∧ (c ∨ b1 ))

40
where
c = c2 ∧ (c0 ∨ c1 ).
These are two of several equivalent forms of this law, which is stronger than modular-
ity and weaker than distributivity. It is modelled after Desargues’ Law in projective
geometry.
Theorem 4.6. If L is a sublattice of Eq X with the property that R ∨ S = R ◦ S
for all R, S ∈ L, then L satisfies the Arguesian law.
Corollary. Every lattice which has a type 1 representation is Arguesian.
Proof. Let L be a sublattice of Eq X with type 1 joins. Assume (A0 ∨ B0 ) ∧ (A1 ∨
B1 ) ≤ A2 ∨ B2 , and suppose (p, q) ∈ C2 = (A0 ∨ A1 ) ∧ (B0 ∨ B1 ). Then there exist
r, s such that

p A0 r A1 q
p B0 s B1 q.

Since (r, s) ∈ (A0 ∨ B0 ) ∧ (A1 ∨ B1 ) ≤ A2 ∨ B2 , there exists t such that r A2 t B2 s.

Now you can check that

(p, t) ∈ (A0 ∨ A2 ) ∧ (B0 ∨ B2 ) = C1

(t, q) ∈ (A1 ∨ A2 ) ∧ (B1 ∨ B2 ) = C0

and hence (p, q) ∈ C0 ∨ C1 . Thus C2 ≤ C0 ∨ C1 , as desired. (This argument is

diagrammed in Figure 4.3.) 

A0 A1
A2

C1 t C0
p q

B2
B0 B1

s
Figure 4.3
41
 Mark Haiman has shown that the converse is false: there are Arguesian lattices
which do not have a type 1 representation [2], [3]. In fact, his proof shows that lat-
tices with a type 1 representation must satisfy equations which are strictly stronger
than the Arguesian law. It follows, in particular, that the lattice of normal sub-
groups of a group also satisfies these stronger equations. Interestingly, P. P. Pálfy
and Laszlo Szabó have shown that subgroup lattices of abelian groups satisfy an
equation which does not hold in all normal subgroup lattices [5].
The question remains: Does there exist a set of equations Σ such that a lattice
has a type 1 representation if and only if it satisfies all the equations of Σ? Haiman
proved that if such a Σ exists, it must contain infinitely many equations. In Chapter
7 we will see that a class of lattices is characterized by a set of equations if and only
if it is closed with respect to direct products, sublattices, and homomorphic images.
The class of lattices having a type 1 representation is easily seen to be closed under
sublattices and direct products, so the question is equivalent to: Is the class of all
lattices having a type 1 representation closed under homomorphic images?

Exercises for Chapter 4

1. Draw Eq X for |X| = 3, 4.
2. Find representations in Eq X for
(a) P(Y ), Y a set,
(b) N5 ,
(c) Mn , n < ∞.
3. Let G be a group. Let F : Sub G → Eq G be the standard representation by
cosets: F (H) = {(x, y) ∈ G2 : xy −1 ∈ H}.
(a) Verify that F (H) is indeed an equivalence relation.
(b) Verify that F is a lattice embedding.
(c) Show that F (H) ∨ F (K) = F (H) ◦ F (K) iff HK = KH (= H ∨ K).
(d) Conclude that the restriction of F to the normal subgroup lattice N (G) is a
type 1 representation.
4. Show that for R, S ∈ Eq X, R ∨ S = R ◦ S iff S ◦ R ⊆ R ◦ S iff R ◦ S = S ◦ R.
(For this reason, such equivalence relations are said to permute.)
5. Recall from Exercise 6 of Chapter 3 that a complete sublattice of an algebraic
lattice is algebraic.
(a) Let S be a join semilattice with 0. Assume that ϕ : S → Eq X is a join
homomorphism with the properties
(i) for each pair a, b ∈ X there exists σ(a, b) ∈ S such that (a, b) ∈ ϕ(s) iff
s ≥ σ(a, b), and
(ii) for each s ∈ S, there exists a pair (xs , ys ) such that (xs , ys ) ∈ ϕ(t) iff s ≤ t.
Show that ϕ induces a complete representation ϕ : I(S) → Eq X.
(b) Indicate how to modify the proof of Theorem 4.1 to obtain, for an arbitrary
42
join semilattice S with 0, a set X and a join homomorphism ϕ : S → Eq X
satisfying (i) and (ii).
(c) Conclude that a complete lattice L has a complete representation by equiva-
lence relations if and only if L is algebraic.
6. Prove that Eq X is a strongly atomic, semimodular, algebraic lattice.
7. Prove that a lattice with a type 1 representation satisfies the Arguesian inclu-
sion (A0 ).

References
1. P. Crawley and R. P. Dilworth, Algebraic Theory of Lattices, Prentice-Hall, Englewood Cliffs,
New Jersey, 1973.
2. M. Haiman, Arguesian lattices which are not linear, Bull. Amer. Math. Soc. 16 (1987), 121–
124.
3. M. Haiman, Arguesian lattices which are not type-1, Algebra Universalis 28 (1991), 128–137.
4. B. Jónsson, On the representation of lattices, Math. Scand. 1 (1953), 193–206.
5. P. P. Pálfy and Cs. Szabó, An identity for subgroup lattices of abelian groups, Algebra Univer-
salis 33 (1995), 191–195.
6. P. Pudlák and J. Tůma, Every finite lattice can be embedded in the lattice of all equivalences
over a finite set, Algebra Universalis 10 (1980), 74–95.
7. P. Whitman, Lattices, equivalence relations and subgroups, Bull. Amer. Math. Soc. 52 (1946),
507–522.

43
 5. Congruence Relations

“You’re young, Myrtle Mae. You’ve got a lot to learn, and I hope you never
learn it.”
– Vita in “Harvey”

You are doubtless familiar with the connection between homomorphisms and
normal subgroups of groups. In this chapter we will establish the corresponding ideas
for lattices (and other general algebras). Borrowing notation from group theory, if X
is a set and θ an equivalence relation on X, for x ∈ X let xθ denote the equivalence
class {y ∈ X : x θ y}, and let

X/θ = {xθ : x ∈ X}.

Thus the elements of X/θ are the equivalence classes of θ.

Recall that if L and K are lattices and h : L → K is a homomorphism, then the
kernel of h is the induced equivalence relation,

ker h = {(x, y) ∈ L2 : h(x) = h(y)}.

We can define lattice operations naturally on the equivalence classes of ker h, viz.,
if θ = ker h, then

xθ ∨ yθ = (x ∨ y)θ
(§)
xθ ∧ yθ = (x ∧ y)θ.

The homomorphism property shows that these operations are well defined, for if
(x, y) ∈ ker h and (r, s) ∈ ker h, then h(x ∨ r) = h(x) ∨ h(r) = h(y) ∨ h(s) = h(y ∨ s),
whence (x ∨ r, y ∨ s) ∈ ker h. Moreover, L/ ker h with these operations forms an
algebra L/ ker h isomorphic to the image h(L), which is a sublattice of K. Thus
L/ ker h is also a lattice.
Theorem 5.1. First Isomorphism Theorem. Let L and K be lattices, and let
h : L → K be a lattice homomorphism. Then L/ ker h with the operations defined by
(§) is a lattice L/ ker h, which is isomorphic to the image h(L) of L in K.
Let us define a congruence relation on a lattice L to be an equivalence relation θ
such that θ = ker h for some homomorphism h.1 We have see that, in addition to
1 This is not the standard definition, but we are about to show it is equivalent to it.
44
being equivalence relations, congruence relations must preserve the operations of L:
if θ is a congruence relation, then

(†) x θ y and r θ s implies x ∨ r θ y ∨ s,

and analogously for meets. Note that (†) is equivalent for an equivalence relation θ
to the apparently weaker, and easier to check, condition

(†0 ) x θ y implies x ∨ z θ y ∨ z.

For (†) implies (†0 ) because every equivalence relation is reflexive, while if θ has
the property (†0 ) and the hypotheses of (†) hold, then applying (†) twice yields
x ∨ r θ y ∨ r θ y ∨ s.
We want to show that, conversely, any equivalence relation satisfying (†0 ) and the
corresponding implication for meets is a congruence relation.
Theorem 5.2. Let L be a lattice, and let θ be an equivalence relation on L satisfying

x θ y implies x ∨ z θ y ∨ z,
(‡)
x θ y implies x ∧ z θ y ∧ z.

Define join and meet on L/θ by the formulas (§). Then L/θ = (L/θ, ∧, ∨) is a lattice,
and the map h : L → L/θ defined by h(x) = xθ is a surjective homomorphism with
ker h = θ.
Proof. The conditions (‡) ensure that the join and meet operations are well defined
on L/θ. By definition, we have

h(x ∨ y) = (x ∨ y)θ = xθ ∨ yθ = h(x) ∨ h(y)

and similarly for meets, so h is a homomorphism. The range of h is clearly L/θ.

It remains to show that L/θ satisfies the equations defining lattices. This follows
from the general principle that homomorphisms preserve the satisfaction of equa-
tions, i.e., if h : L  K is a surjective homomorphism and an equation p = q holds
in L, then it holds in K. (See Exercise 4.) For example, to check commutativity of
meets, let a, b ∈ K. Then there exist x, y ∈ L such that h(x) = a and h(y) = b.
Hence

a ∧ b = h(x) ∧ h(y) = h(x ∧ y)

= h(y ∧ x) = h(y) ∧ h(x) = b ∧ a.

Similar arguments allow us to verify the commutativity of joins, the idempotence

and associativity of both operations, and the absorption laws. Thus a homomorphic
45
image of a lattice is a lattice.2 As h : L → L/θ is a surjective homomorphism, we
conclude that L/θ is a lattice, which completes the proof. 
Thus congruence relations are precisely equivalence relations which satisfy (‡).
But the conditions of (‡) and the axioms for an equivalence relation are all finitary
closure rules on L2 . Hence, by Theorem 3.1, the set of congruence relations on a
lattice L forms an algebraic lattice Con L. The corresponding closure operator on
L2 is denoted by “con”. So for a set Q of ordered pairs, con Q is the congruence
relation generated by Q; for a single pair, Q = {(a, b)}, we write just con(a, b).
Moreover, the equivalence relation join (the transitive closure of the union) of a set
of congruence relations again satisfies (‡). For if θi (i ∈ I) are congruence relations
and x θi1 r1 θi2 r2 . . . θin y, then x ∨ z θi1 S
r1 ∨ z θi2 r2 ∨ z . . . θin y ∨ z, and likewise
for meets. Thus
W the transitive closure of i∈I θi is a congruence relation, and so it
is the join i∈I θi in Con L. Since the meet is also the same (set intersection) in
both lattices, Con L is a complete sublattice of Eq L.
Theorem 5.3. Con L is an algebraic lattice. A congruence relation θ is compact
in Con L if and only if it is finitely generated, i.e.,Wthere exist finitely many pairs
(a1 , b1 ), . . . , (ak , bk ) of elements of L such that θ = 1≤i≤k con(ai , bi ).

Note that the universal relation and the equality relation on L2 are both con-
gruence relations; they are the greatest and least elements of Con L, respectively.
Also, since x θ y if and only if x ∧ y θ x ∨ y, a congruence relation is determined by
the ordered pairs (a, b) with a < b which it contains.
A congruence relation θ is principal if θ = con(a, b) for some pair a, b ∈ L. The
principal congruence relations are join dense in Con L: for any congruence relation
θ, we have
_
θ = {con(a, b) : a θ b}.

It follows from the general theory of algebraic closure operators that principal
congruence W relations are compact, but this can be shown directly as follows: if
con(a, b) ≤ i∈I θi , then there exist elements c1 , . . . , cm and indices i0 , . . . , im such
that
a θi0 c1 θi1 c2 . . . θim b ,
W
whence (a, b) ∈ θi0 ∨ . . . ∨ θim and thus con(a, b) ≤ 0≤j≤m θij .
One of the most basic facts about congruences says that congruences of L/θ
correspond to congruences on L containing θ.
Theorem 5.4. Second Isomorphism Theorem. If θ ∈ Con L, then Con (L/θ)
is isomorphic to the interval 1/θ in Con L.

2 The corresponding statement is true for any equationally defined class of algebras, including

modular, Arguesian and distributive lattices.

46
Proof. A congruence relation on L/θ is an equivalence relation R on the θ-classes
of L such that
xθ R yθ implies xθ ∨ zθ R yθ ∨ zθ
and analogously for meets. Given R ∈ Con L/θ, define the corresponding relation
ρ on L by x ρ y iff xθ R yθ. Clearly ρ ∈ Eq L and θ ≤ ρ. Moreover, if x ρ y and
z ∈ L, then
(x ∨ z)θ = xθ ∨ zθ R yθ ∨ zθ = (y ∨ z)θ,
whence x ∨ z ρ y ∨ z, and similarly for meets. Hence ρ ∈ Con L, and we have
established an order preserving map f : Con L/θ → 1/θ.
Conversely, let σ ∈ 1/θ in Con L, and define a relation S on L/θ by xθ S yθ iff
x σ y. Since θ ≤ σ the relation S is well defined. If xθ S yθ and z ∈ L, then x σ y
implies x ∨ z σ y ∨ z, whence

xθ ∨ zθ = (x ∨ z)θ S (y ∨ z)θ = yθ ∨ zθ,

and likewise for meets. Thus S is a congruence relation on L/θ. This gives us an
order preserving map g : 1/θ → Con L/θ.
The definitions make f and g inverse maps, so they are in fact isomorphisms. 
It is interesting to interpret the Second Isomorphism Theorem in terms of homo-
morphisms. Essentially it corresponds to the fact that if h : L  K and f : L → M
are homomorphisms with h surjective, then there is a homomorphism g : K → M
with f = gh if and only if ker h ≤ ker f .
A lattice L is called simple if Con L is a two element chain, i.e., |L| > 1 and L has
no congruences except equality and the universal relation. For example, the lattice
Mn is simple whenever n ≥ 3. A lattice is subdirectly irreducible if it has a unique
minimum nonzero congruence relation, i.e., if 0 is completely meet irreducible in
Con L. So every simple lattice is subdirectly irreducible, and N5 is an example of
a subdirectly irreducible lattice which is not simple .
The following are immediate consequences of the Second Isomorphism Theorem.
Corollary. L/θ is simple if and only if 1  θ in Con L.
Corollary. L/θ is subdirectly irreducible if and only if θ is completely meet irre-
ducible in Con L.
Now we turn our attention to a decomposition of lattices which goes back to
R. Remak in 1930 (for groups) [7]. In what follows, it is important to remember
that the zero element of a congruence lattice is the equality relation.
V
Theorem 5.5.QIf 0 = i∈I θi in Con L, then L is isomorphic to a sublattice of the
direct product i∈I L/θi , and each of the natural homomorphisms πi : L → L/θi is
surjective.
47
 Q
Conversely, if L is isomorphic to a sublattice of a direct product i∈I Ki and each
∼
V the projection homomorphisms πi : L → Ki is surjective, then Ki = L/ ker πi and
of
i∈I ker πi = 0 in Con L.

Proof. ForQany collection θi (i ∈ I) in Con L, there is a natural homomorphism

π : L → L/θi with (π(x))i = xθi . Since two elements of V a direct product
V are
equal if and only if they agree in every component, ker π = θi . So if θi = 0,
then π is an embedding.
Q
Conversely,
V if π : L → Ki is an embedding, then ker π = 0, while as above
ker π = ker πi . Clearly, if πi (L) = Ki then Ki ∼
= L/ ker πi . 

A representation of L satisfying either of the equivalent conditions of Theorem 5.5

is called a subdirect decomposition, and the corresponding external construction is
called a subdirect product. For example, Figure 5.1 shows how a six element lattice
L can be written as a subdirect product of two copies of N5 .

1 h1, 1i

x hx, zi hz, xi

y hy, zi hz, yi

0 h0, 0i
N5 L
Figure 5.1

Next we should show that subdirectly irreducible lattices are indeed those which
have no proper subdirect decomposition.

Theorem 5.6. The following are equivalent for a lattice L.

(1) L is subdirectly irreducible, i.e., 0 is completely meet irreducible in Con L.
(2) There is a unique minimal nonzero congruence µ on L with the property that
θ ≥ µ for every nonzero θ ∈ Con L. Q
(3) If L is isomorphic to a sublattice of Ki , then some projection homomor-
phism πi is one-to-one.
(4) There exists a pair of elements a < b in L such that a θ b for every nonzero
congruence θ.
48
 The congruence µ of condition (2) is called the monolith of the subdirectly irre-
ducible lattice L, and the pair (a, b) of condition (4), which need not be unique, is
called a critical pair.
Proof. The equivalence of (1), (2) and (3) is a simple combination of Theorems 3.8
and 5.5. We get (2) implies (4) by taking a = x ∧ y and b = x ∨ y for any pair
of distinct elements with x µ y. On the other hand, if (4) holds we obtain (2) with
µ = con(a, b). 
Now we see the beauty of Birkhoff’s Theorem 3.9, that every element in an
algebraic lattice is a meet of completely meet irreducible elements. By applying this
to the zero element of Con L, we obtain the following fundamental result.
Theorem 5.7. Every lattice is a subdirect product of subdirectly irreducible lattices.
It should be clear that, with the appropriate modifications, Theorems 5.5 to 5.7
yield subdirect decompositions of groups, rings, semilattices, etc. into subdirectly
irreducible algebras of the same type. Keith Kearnes [5] has shown that there are
interesting varieties of algebras whose congruence lattices are strongly atomic. By
Theorem 3.10, these algebras have irredundant subdirect decompositions.
Subdirectly irreducible lattices play a particularly important role in the study of
varieties (Chapter 7).
So far we have just done universal algebra with lattices: with the appropriate
modifications, we can characterize congruence relations and show that Con A is an
algebraic lattice for any algebra A. (See Exercises 10 and 11.) However, the next
property is special to lattices (and related structures). It was first discovered by
N. Funayama and T. Nakayama [2] in the early 1940’s.
Theorem 5.8. If L is a lattice, then Con L is a distributive algebraic lattice.
Proof. In any lattice L, let

m(x, y, z) = (x ∧ y) ∨ (x ∧ z) ∨ (y ∧ z).

Then it is easy to see that m(x, y, z) is a majority polynomial, in that if any two
variables are equal then m(x, y, z) takes on that value:

m(x, x, z) = x
m(x, y, x) = x
m(x, z, z) = z.

Now let α, β, γ ∈ Con L. Clearly (α ∧ β) ∨ (α ∧ γ) ≤ α ∧ (β ∨ γ). To show the

reverse inclusion, let x, z ∈ α ∧ (β ∨ γ). Then x α z and there exist y1 , . . . , yk such
that
x = y0 β y1 γ y2 β . . . yk = z .
49
Let ti = m(x, yi , z) for 0 ≤ i ≤ k. Then

t0 = m(x, x, z) = x
tk = m(x, z, z) = z

and for all i,

ti = m(x, yi , z) α m(x, yi , x) = x ,
so ti α ti+1 by Exercise 4(b). If i is even, then

ti = m(x, yi , z) β m(x, yi+1 , z) = ti+1 ,

whence ti α ∧ β ti+1 . Similarly, if i is odd then ti α ∧ γ ti+1 . Thus

x = t0 α ∧ β t1 α ∧ γ t2 α ∧ β . . . tk = z

and we have shown that α ∧ (β ∨ γ) ≤ (α ∧ β) ∨ (α ∧ γ). As inclusion holds both

ways, we have equality. Therefore Con L is distributive. 
What else can you say about congruence lattices of lattices? Is every distribu-
tive algebraic lattice isomorphic to the congruence lattice of a lattice? This is the
$64,000 question. But while the general question remains open, it is known that a
distributive algebraic lattice D is isomorphic to the congruence lattice of a lattice if
(i) D ∼
= O(P) for some ordered set P (R. P. Dilworth, see [3]), or
(ii) the compact elements are a sublattice of D (E. T. Schmidt [8]), or
(iii) D has at most ℵ1 compact elements (A. Huhn [4]).
Thus a counterexample, if there is one, would have to be quite large and compli-
cated.3 In Chapter 10 we will prove (i), which includes the fact that every finite
distributive lattice is isomorphic to the congruence lattice of a (finite) lattice.
We need to understand the congruence operator con Q, where Q is a set of pairs,
a little better. A weaving polynomial on a lattice L is a member of the set W of
unary functions defined recursively by
(1) w(x) = x ∈ W ,
(2) if w(x) ∈ W and a ∈ L, then u(x) = w(x) ∧ a and v(x) = w(x) ∨ a are in W ,
(3) only these functions are in W .
Thus every weaving polynomial looks something like

w(x) = (. . . (((x ∧ s1 ) ∨ s2 ) ∧ s3 ) . . . ) ∨ sk

where si ∈ L for 1 ≤ i ≤ k. The following characterization is a modified version of

one found in Dilworth [1].
3 Itis not even known whether every distributive algebraic lattice is isomorphic to the congruence
lattice of an algebra with finitely many operations. For recent results on these problems, see
M. Tischendorf [9] and M. Ploščica, J. Tůma and F. Wehrung [6], [10].
50
 W
Theorem 5.9. Suppose ai < bi for i ∈ I. Then (x, y) ∈ i∈I con(ai , bi ) if and only
if there exist finitely many rj ∈ L, wj ∈ W , and ij ∈ I such that

x ∨ y = r0 ≥ r1 ≥ · · · ≥ rk = x ∧ y

with wj (bij ) = rj and wj (aij ) = rj+1 for 0 ≤ j < k.

Proof. Let R be the set of all pairs (x, y) satisfying the condition of the theorem. It
is clear that
i ) ∈ R for all i,
(1) (ai , bW
(2) R ⊆ i∈I con(ai , bi ).
Hence, if we can show that R is a congruence relation, it will follow that R =
W
i∈I con(ai , bi ).
Note that (x, y) ∈ R if and only if (x ∧ y, x ∨ y) ∈ R. It also helps to observe that
if x R y and x ≤ u ≤ v ≤ y, then u R v. To see this, replace the weaving polynomials
w(t) witnessing x R y by new polynomials w0 (t) = (w(t) ∨ u) ∧ v.
First we must show R ∈ Eq L. Reflexivity and symmetry are obvious, so let
x R y R z with
x ∨ y = r0 ≥ r1 ≥ · · · ≥ rk = x ∧ y
using polynomials wj ∈ W , and

y ∨ z = s0 ≥ s1 ≥ · · · ≥ sm = y ∧ z

via polynomials vj ∈ W , as in the statement of the theorem. Replacing wj (t) by

wj0 (t) = wj (t) ∨ y ∨ z, we obtain

x ∨ y ∨ z = r0 ∨ y ∨ z ≥ r1 ∨ y ∨ z ≥ · · · ≥ (x ∧ y) ∨ y ∨ z = y ∨ z.

Likewise, replacing wj (t) by wj00 (t) = wj (t) ∧ y ∧ z, we have

y ∧ z = (x ∨ y) ∧ y ∧ z ≥ r1 ∧ y ∧ z ≥ · · · ≥ rk ∧ y ∧ z = x ∧ y ∧ z.

Combining the two new sequences with the original one for y R z, we get a sequence
from x ∨ y ∨ z down to x ∧ y ∧ z. Hence x ∧ y ∧ z R x ∨ y ∨ z. By the observations
above, x ∧ z R x ∨ z and x R z, so R is transitive.
Now we must check (‡). Let x R y as before, and let z ∈ L. Replacing wj (t) by
uj (t) = wj (t) ∨ z, we obtain a sequence from x ∨ y ∨ z down to (x ∧ y) ∨ z. Thus
(x ∧ y) ∨ z R x ∨ y ∨ z, and since (x ∧ y) ∨ z ≤ (x ∨ z) ∧ (y ∨ z) ≤ x ∨ y ∨ z, this
implies x ∨ z R y ∨ z. The argument for meets is done similarly, and we conclude
that R is a congruence relation, as desired. 
The condition of Theorem 5.9 is a bit unwieldy, but not as bad to use as you
might think. Let us look at some consequences of the theorem.
51
 W
Corollary. If θi ∈ Con L for i ∈ I, then (x, y) ∈ i∈I θi if and only if there exist
finitely many rj ∈ L and ij ∈ I such that
x ∨ y = r 0 ≥ r1 ≥ · · · ≥ r k = x ∧ y
and rj θij rj+1 for 0 ≤ j < k.
At this point we need some basic facts about distributive algebraic lattices (like
Con L). W Recall that an element p of a complete lattice is completely join irreducible
W Q implies p = q for some q ∈ Q. An element p is completely join prime if
if p =
p ≤ Q implies p ≤ q for some q ∈ Q. Clearly every completely join prime element
is completely join irreducible, but in general completely join irreducible elements
need not be join prime.
Now every algebraic lattice has lots of completely meet irreducible elements (by
Theorem 3.9), but they may have no completely join irreducible elements. This hap-
pens, for example, in the lattice consisting of the empty set and all cofinite subsets of
an infinite set (which is distributive and algebraic). However, such completely join
irreducible elements as there are in a distributive algebraic lattice are completely
join prime!
Theorem 5.10. The following are equivalent for an element p in an algebraic dis-
tributive lattice.
(1) p is completely join prime.
(2) p is completely join irreducible.
(3) p is compact and (finitely) join irreducible.
Proof. Clearly (1) implies (2), and since every element in an algebraic lattice is a
join of compact elements, (2) implies (3). W
Let p be compact
W and finitely join irreducible, and assume p ≤ Q. As p is
compact,Wp ≤ WF for some finite subset F ⊆ Q. By distributivity, this implies
p = p ∧ ( F ) = q∈F p ∧ q. Since p is join irreducible, p = p ∧ q ≤ q for some q ∈ F .
Thus p is completely join prime. (Cf. Exercise 3.1) 
We will return to the theory of distributive lattices in Chapter 8, but let us now
apply what we know to Con L. As an immediate consequence of the Corollary to
Theorem 5.9 we have the following.
Theorem 5.11. If a ≺ b, then con(a, b) is completely join prime in Con L.
The converse is false, as there are infinite simple lattices with no covering rela-
tions. However, for finite lattices, or more generally principally chain finite lattices,
the converse does hold. A lattice is principally chain finite if every principal ideal
c/0 satisfies the ACC and DCC. This is a fairly natural finiteness condition which
includes many interesting infinite lattices, and many results for finite lattices can be
extended to principally chain finite lattices with a minimum of effort. Recall that if
x is a join irreducible element in such a lattice, then x∗ denotes the unique element
such that x  x∗ .
52
Theorem 5.12. Let L be a principally chain finite lattice. Then every congruence
relation on L is the join of completely join irreducible congruences. Moreover, ev-
ery completely join irreducible congruence is of the form con(x, x∗ ) for some join
irreducible element x of L.
Proof. Every congruence relation is a join of compact congruences, and every com-
pact congruence is a join of finitely many congruences con(a, b) with a > b. In a prin-
cipally chain finite lattice, every chain in a/b is finite by ExerciseW1.5, so there exists a
covering chain a = r0  r1  · · ·  rk = b. Clearly con(a, b) = 0≤j<k con(rj , rj+1 ),
and these latter are completely join prime by Theorem 5.11. Thus every congruence
relation on L is the join of completely join irreducible congruences con(r, s) with
r  s.
Now let a  b be any covering pair in L. By the DCC for a/0, there is an element
x which is minimal with respect to the properties x ≤ a and x  b. Since any
element strictly below x is below b, the element x is join irreducible and x∗ = x ∧ b.
It is also true that a = x ∨ b, since b < x ∨ b ≤ a, and it follows easily from these
two facts that con(a, b) = con(x, x∗ ). 
We will return to congruence lattices of principally chain finite lattices in Chap-
ter 10.

Exercises for Chapter 5

1. Find Con L for the lattices (a) Mn where n ≥ 3, (b) N5 , (c) the lattice L of
Figure 5.1, and the lattices in Figure 5.2.

(d) (e) (f )
Figure 5.2
W
2. An element p of a lattice L is join prime if for any finite subset F of L, p ≤ F
implies p ≤ f for some f ∈ F . Let P(L) denote the set of join prime elements of L,
and define
x∆y iff x/0 ∩ P(L) = y/0 ∩ P(L).
53
Prove that ∆ is a congruence relation on L.
3. Let X be any set. Define a binary relation on P(X) by A ≈ B iff the symmetric
difference (A−B)∪(B −A) is finite. Prove that ≈ is a congruence relation on P(X).
4. Lattice terms are defined in the proof of Theorem 6.1.
(a) Show that if p(x1 , . . . , xn ) is a lattice term and h : L → K is a homomor-
phism, then h(p(a1 , . . . , an )) = p(h(a1 ), . . . , h(an )) for all a1 , . . . , an ∈ L.
(b) Show that if p(x1 , . . . , xn ) is a lattice term and θ ∈ Con L and ai θ bi for
1 ≤ i ≤ n, then p(a1 , . . . , an ) θ p(b1 , . . . , bn ).
(c) Let p(x1 , . . . , xn ) and q(x1 , . . . , xn ) be lattice terms, and let h : L  K be a
surjective homomorphism. Prove that if p(a1 , . . . , an ) = q(a1 , . . . , an ) for all
a1 , . . . , an ∈ L, then p(c1 , . . . , cn ) = q(c1 , . . . , cn ) holds for all c1 , . . . , cn ∈ K.
5. Show that each element of a finite distributive lattice has a unique irredundant
decomposition. What does this say about subdirect decompositions of finite lattices?
6. Let θ ∈ Con L.
(a) Show that x  y implies xθ  yθ or xθ = yθ.
(b) Prove that if L is a finite semimodular lattice, then so is L/θ.
(c) Prove that a subdirect product of semimodular lattices is semimodular.
7. Prove that Con L1 × L2 ∼ = Con L1 × Con L2 . (Note that this is not true
for groups; see Exercise 9.)
8. Let L be a finitely generated lattice, and let θ be a congruence on L such that
L/θ is finite. Prove that θ is compact.
9. Find the congruence lattice of the abelian group Zp × Zp , where p is prime.
Find all finite abelian groups whose congruence lattice is distributive. (Recall that
the congruence lattice of an abelian group is isomorphic to its subgroup lattice.)
For Exercises 10 and 11 we refer to §3 (Universal Algebra) of Appendix 1.
10. Let A = hA; F, Ci be an algebra.
(a) Prove that if h : A → B is a homomorphism and θ = ker h, then for each
f ∈ F,

(U) xi θ yi for 1 ≤ i ≤ n implies f (x1 , . . . , xn ) θ f (y1 , . . . , yn ).

(b) Prove that (U) is equivalent to the apparently weaker condition that for all
f ∈ F and every i,

($) xi θ y implies f (x1 , . . . , xi , . . . xn ) θ f (x1 , . . . , y, . . . xn ).

(c) Show that if θ ∈ Eq A satisfies (U), then the natural map h : A → A/θ is a
homomorphism with θ = ker h.
Thus congruence relations, defined as homomorphism kernels, are precisely equiva-
lence relations satisfying (U).
54
 11. Accordingly, let Con A = {θ ∈ Eq A : θ satisfies (U)}.
(a) Prove that ConWA is aVcomplete sublattice of Eq A. (In particular, you
must show that and are the same in both lattices.)
(b) Show that Con A is an algebraic lattice.

References
1. R. P. Dilworth, The structure of relatively complemented lattices, Ann. of Math. 51 (1950),
348–359.
2. N. Funayama and T. Nakayama, On the distributivity of a lattice of lattice-congruences, Proc.
Imp. Acad. Tokyo 18 (1942), 553–554.
3. G. Grätzer and E. T. Schmidt, On congruence lattices of lattices, Acta Math. Acad. Sci.
Hungar. 13 (1962), 179–185.
4. A. Huhn, On the representation of distributive algebraic lattices II, Acta Sci. Math. (Szeged).
(1989).
5. K. A. Kearnes, Atomicity and Nilpotence, Canadian J. Math. 42 (1990), 1–18.
6. M. Ploščica, J. Tuma and F. Wehrung, Congruence lattices of free lattices in nondistributive
varieties, Coll. Math. (to appear).
7. R. Remak, Über die Darstellung der endlichen Gruppen als Untergruppen direkter Produckte,
Jour. für Math. 163 (1930), 1–44.
8. E. T. Schmidt, The ideal lattice of a distributive lattice with 0 is the congruence lattice of a
lattice, Acta Sci. Math. (Szeged). 43 (1981), 153–168.
9. M. Tischendorf, On the representation of distributive semilattices, Algebra Universalis 31
(1994), 446-455.
10. F. Wehrung, A uniform refinement property of certain congruence lattices, Proc. Amer. Math.
Soc. (to appear).

55
 6. Free Lattices

Freedom’s just another word for nothing left to lose ....

–Kris Kristofferson

If x, y and z are elements of a lattice, then x ∨ (y ∨ (x ∧ z)) = x ∨ y is always true,

while x ∨ y = z is usually not true. Is there an algorithm which, given two lattice
expressions p and q, determines whether p = q holds for every substitution of the
variables in every lattice? The answer is yes, and finding this algorithm (Corollary
to Theorem 6.2) is our original motivation for studying free lattices.
We say that a lattice L is generated by a set X ⊆ L if no proper sublattice of L
contains X. In terms of the subalgebra closure operator Sg introduced in Chapter 3,
this means Sg(X) = L.
A lattice F is freely generated by X if
(I) F is a lattice,
(II) X generates F,
(III) for every lattice L, every map h0 : X → L can be extended to a homomor-
phism h : F → L.
A free lattice is a lattice which is freely generated by one of its subsets.
Condition (I) is sort of redundant, but we include it because it is important
when constructing a free lattice to be sure that the algebra constructed is indeed
a lattice. In the presence of condition (II), there is only one way to define the
homomorphism h in condition (III): for example, if x, y, z ∈ X then we must have
h(x ∨ (y ∧ z)) = h0 (x) ∨ (h0 (y) ∧ h0 (z)). Condition (III) really says that this natural
extension is well defined. This in turn says that the only time two lattice terms in
the variables X are equal in F is when they are equal in every lattice.
Now the class of lattices is an equational class, i.e., it is the class of all algebras
with a fixed set of operation symbols (∨ and ∧) satisfying a given set of equations
(the idempotent, commutative, associative and absorption laws). Equational classes
are also known as varieties, and in Chapter 7 we will take a closer look at varieties
of lattices. A fundamental theorem of universal algebra, due to Garrett Birkhoff [1],
says that given any nontrivial1 equational class V and any set X, there is an algebra
in V freely generated by X. Thus the existence of free groups, free semilattices,
and in particular free lattices is guaranteed.2 Likewise, there are free distributive
1A variety T is trivial if it satisfies the equation x ≈ y, which means that every algebra in T
has exactly one element. This is of course the smallest variety of any type.
2 However, there is no free lunch.

56
lattices, free modular lattices, and free Arguesian lattices, since each of these laws
can be written as a lattice equation.
Theorem 6.1. For any nonempty set X, there exists a free lattice generated by X.
The proof uses three basic principles of universal algebra. These correspond for
lattices to Theorems 5.1, 5.4, and 5.5 respectively. However, the proofs of these
theorems involved nothing special to lattices except the operation symbols ∧ and ∨;
these can easily be changed to arbitrary operation symbols. Thus, with only minor
modification, the proof of this theorem can be adapted to show the existence of free
algebras in any nontrivial equational class of algebras.
Basic Principle 1. If h : A  B is a surjective homomorphism, then B ∼= A/ ker h.
Basic Principle 2. If f : A → B and g : A  C are homomorphism with g
surjective, and ker g ≤ ker f , then there exists h : C → B such that f = hg.

f
A B

g h
C
Figure 6.1
V
Basic Principle 3. If ψ = i∈I θi in Con A, then A/ψ is isomorphic to a
subalgebra of the direct product Πi∈I A/θi .
With these principles in hand, we proceed with the proof of Theorem 6.1.
Proof of Theorem 6.1. Given the set X, define the word algebra W (X) to be the set
of all formal expressions (strings of symbols) satisfying the following properties:
(1) X ⊆ W (X),
(2) if p, q ∈ W (X), then (p ∨ q) and (p ∧ q) are in W (X),
(3) only the expressions given by the first two rules are in W (X).
Thus W (X) is the absolutely free algebra with operation symbols ∨ and ∧ generated
by X. The elements of W (X), which are called terms, are all well-formed expressions
in the variables X and the operation symbols ∧ and ∨. Clearly W (X) is an algebra
generated by X, which is property (II) from the definition of a free lattice. Because
two terms are equal if and only if they are identical, W (X) has the mapping property
(III). On the other hand, it is definitely not a lattice. We need to identify those
pairs p, q ∈ W (X) which evaluate the same in every lattice, e.g., x and (x ∧ (x ∨ y)).
The point of the proof is that when this is done, properties (II) andV(III) still hold.
Let Λ = {θ ∈ Con W (X) : W (X)/θ is a lattice}, and let λ = Λ. We claim
that W (X)/λ is a lattice freely generated by {xλ : x ∈ X}.
57
 By Basic Principle 3, W (X)/λ is isomorphic to a subalgebra of a direct product
of lattices, so it is a lattice.3 Clearly W (X)/λ is generated by {xλ : x ∈ X}, and
because there exist nontrivial lattices (more than one element) for X to be mapped
to in different ways, x 6= y implies xλ 6= yλ for x, y ∈ X.
Now let L be a lattice and let f0 : X → L be any map. By the preceding
observation, the corresponding map h0 : X/λ → L defined by h0 (xλ) = f0 (x)
is well defined. Now f0 can be extended to a homomorphism f : W (X) → L,
whose range is some sublattice S of L. By Basic Principle 1, W (X)/ ker f ∼ = S so
ker f ∈ Λ, and hence ker f ≥ λ. If we use ε to denote the standard homomorphism
W (X)  W (X)/λ with ε(u) = uλ for all u ∈ W (X), then ker f ≥ ker ε = λ. Thus
by Basic Principle 2 there exists a homomorphism h : W (X)/λ → L with hε = f
(see Figure 6.2). This means h(uλ) = f (u) for all u ∈ W (X); in particular, h
extends h0 as required. 

f
W (X) L

ε h
W (X)/λ
Figure 6.2

It is easy to see, using the mapping property (III), that if F is a lattice freely
generated by X, G is a lattice freely generated by Y , and |X| = |Y |, then F ∼ = G.
Thus we can speak of the free lattice generated by X, which we will denote by
FL(X). If |X| = n, then we also denote this lattice by FL(n). The lattice FL(2)
has four elements, so there is not much to say about it. But FL(n) is infinite for
n ≥ 3, and we want to investigate its structure.
The advantage of the general construction we used is that it gives us the existence
of free algebras in any variety; the disadvantage is that it does not, indeed cannot,
tell us anything about the arithmetic of free lattices. For this we need a result due
to Thoralf Skolem [15] (reprinted in [16]), and independently, P. M. Whitman [18]
in 1941.4
3 This is where we use that lattices are equationally defined. For example, the class of integral

domains is not equationally defined, and the direct product of two or more integral domains is not
one.
4 The history here is rather interesting. Skolem, as part of his 1920 paper which proves the

Lowenheim-Skolem Theorem, solved the word problem not only for free lattices, but for finitely
presented lattices as well. But by the time the great awakening of lattice theory occurred in the
1930’s, his solution had been forgotten. Thus Whitman’s 1941 construction of free lattices became
the standard reference on the subject. It was not until 1992 that Stan Burris rediscovered Skolem’s
solution.
58
Theorem 6.2. Every free lattice FL(X) satisfies the following conditions, where
x, y ∈ X and p, q, p1 , p2 , q1 , q2 ∈ FL(X).
(1) x ≤ y iff x = y.
(2) x ≤ q1 ∨ q2 iff x ≤ q1 or x ≤ q2 .
(3) p1 ∧ p2 ≤ x iff p1 ≤ x or p2 ≤ x.
(4) p1 ∨ p2 ≤ q iff p1 ≤ q and p2 ≤ q.
(5) p ≤ q1 ∧ q2 iff p ≤ q1 and p ≤ q2 .
(6) p = p1 ∧ p2 ≤ q1 ∨ q2 = q iff p1 ≤ q or p2 ≤ q or p ≤ q1 or p ≤ q2 .
Finally, p = q iff p ≤ q and q ≤ p.
Condition (6) in Theorem 6.2 is known as Whitman’s condition, and it is usually
denoted by (W).
Proof of Theorem 6.2. Properties (4) and (5) hold in every lattice, by the definition
of least upper bound and greatest lower bound, respectively. Likewise, the “if ”
parts of the remaining conditions hold in every lattice.
We can take care of (1) and (2) simultaneously. Fixing x ∈ X, let
_
Gx = {w ∈ FL(X) : w ≥ x or w ≤ F for some finite F ⊆ X − {x}}.

Then X ⊆ Gx , and Gx is closed under W joins and meets, so Gx = FL(X). Thus every
w ∈ FL(X) is either above x or below F for some finite F ⊆ X − {x}. W Properties
(1) and (2) will follow if we can show that this “or” is exclusive: x  F for all
finite F ⊆ X − {x}. So let h0 : X → 2 (the two element chain) be defined by
h0 (x) = 1, and h0 (y) = 0 for y ∈ X − {x}. This map extends to a homomorphism W
h : FL(X) → W 2. For every finite F ⊆ X − {x} we have h(x) = 1  0 = h( F ),
whence x  F . V
Condition (3) is the dual of (2). Note that the proof shows x  G for all finite
G ⊆ X − {x}.
Whitman’s condition (6), or (W), can be proved using a slick construction due
to Alan Day [3]. This construction can be motivated by a simple example. In the
lattice of Figure 6.3(a), the elements a, b, c, d fail (W); in Figure 6.3(b) we have
“fixed” this failure by making a ∧ b  c ∨ d. Day’s method provides a formal way of
doing this for any (W)-failure.
Let I = u/v be an interval in a lattice L. We define a new lattice L[I] as follows.
The universe of L[I] is (L − I) ∪ (I × 2). Thus the elements of L[I] are of the form
x with x ∈/ I, and (y, i) with i ∈ {0, 1} and y ∈ I. The order on L[I] is defined by:

x ≤ y if x ≤L y
(x, i) ≤ y if x ≤L y
x ≤ (y, j) if x ≤L y
(x, i) ≤ (y, j) if x ≤L y and i ≤ j.
59
 (u, 1) a b
u a b (u, 0)
(v, 1)
c d v c d (v, 0)

(a) (b)
Figure 6.3

It is not hard to check the various cases to show that each pair of elements in L[I]
has a meet and join, so that L[I] is indeed a lattice.5 Moreover, the natural map
κ : L[I]  L with κ(x) = x and κ((y, i)) = y is a homomorphism. Figure 6.4 gives
another example of the doubling construction, where the doubled interval consists
of a single element {u}.
Now suppose a, b, c, d witness a failure of (W) in FL(X). Let u = c ∨ d, v = a ∧ b
and I = u/v. Let h0 : X → FL(X)[I] with h0 (x) = x if x ∈ / I, h0 (y) = (y, 0) if
y ∈ I, and extend this map to a homomorphism h. Now κh : FL(X) → FL(X) is
also a homomorphism, and since κh(x) = x for all x ∈ X, it is in fact the identity.
Therefore h(w) ∈ κ−1 (w) for all w ∈ FL(X). Since a, b, c, d ∈/ I, this means h(t) = t
for t ∈ {a, b, c, d}. Now v = a ∧ b ≤ c ∨ d = u in FL(X), so h(v) ≤ h(u). But we can
calculate

h(v) = h(a) ∧ h(b) = a ∧ b = (v, 1)  (u, 0) = c ∨ d = h(c) ∨ h(d) = h(u)

in FL(X)[I], a contradiction. Thus (W) holds in FL(X). 

Theorem 6.2 gives us a solution to the word problem for free lattices, i.e., an
algorithm for deciding whether two lattice terms p, q ∈ W (X) evaluate to the same
element in FL(X) (and hence in all lattices). Strictly speaking, we have an eval-
uation map ε : W (X) → FL(X) with ε(x) = x for all x ∈ X, and we want to
decide whether ε(p) = ε(q). Following tradition, however, we suppress the ε and
ask whether p = q in FL(X).

5 Thisconstruction yields a lattice if, instead of requiring that I be an interval, we only ask that
it be convex, i.e., if x, z ∈ I and x ≤ y ≤ z, then y ∈ I. This generalized construction has also
proved very useful, but we will not need it here.
60
 (u, 1)
u
(u, 0)

(a) (b)
Figure 6.4

Corollary. Let p, q ∈ W (X). To decide whether p ≤ q in FL(X), apply the condi-

tions of Theorem 6.2 recursively. To test whether p = q in FL(X), check both p ≤ q
and q ≤ p.

The algorithm works because it eventually reduces p ≤ q to a statement involving

the conjunction and disjunction of a number of inclusions of the form x ≤ y, each of
which holds if and only if x = y. Using the algorithm requires a little practice; you
should try showing that x∧(y ∨z)  (x∧y)∨(x∧z) in FL(X), which is equivalent to
the statement that not every lattice is distributive.6 To appreciate its significance,
you should know that it is not always possible to solve the word problem for free
algebras. For example, the word problem for a free modular lattice FM (X) is not
solvable if |X| ≥ 4 (see Chapter 7).
By isolating the properties which do not hold in every lattice, we can rephrase
Theorem 6.2 in the following useful form.
Theorem 6.3. A lattice F is freely generated by its subset X if and only if F is
generated by X, F satisfies (W ), and the following two conditions hold for each
x ∈ X:
W
(1) if x ≤ V G for some finite G ⊆ X, then x ∈ G;
(2) if x ≥ H for some finite H ⊆ X, then x ∈ H.

It is worthwhile to compare the roles of Eq X and FL(X): every lattice can be

embedded into a lattice of equivalence relations, while every lattice is a homomorphic
image of a free lattice.

6 The algorithm for the word problem, and other free lattice algorithms, can be efficiently
programmed; see Chapter XI of [6]. These programs have proved to be a useful tool in the
investigation of the structure of free lattices.
61
 Note that it follows from (W) that no element of FL(X) is properly both a meet
and a join, i.e., every element is either meet irreducible or join irreducible. Moreover,
the generators are the only elements which are both meet and join irreducible. It
follows that the generating set of FL(X) is unique. This is very different from the
situation say in free groups: the free group on {x, y} is also generated (freely) by
{x, xy}.
Each element w ∈ FL(X) corresponds to an equivalence class of terms in W (X).
Among the terms which evaluate to w, there may be several of minimal length (total
number of symbols), e.g., (x ∨ (y ∨ z)), ((y ∨ x) ∨ z), etc. Note that if a term p can
be obtained from a term q by applications of the associative and commutative laws
only, thenWp and q have the same length. This allows us to speak of the length of a
term t = ti without specifying the order or parenthesization of the joinands, and
likewise for meets. We want to show that a minimal length term for w is unique
up to associativity and commutativity. This is true for generators, so by duality it
suffices to consider the case when w is a join.
W
Lemma 6.4. Let t = ti in W (X), where each ti is either a generator or a meet.
Assume that ε(t) = w and ε(ti ) = wi under the evaluation map ε : W (X) → FL(X).
If t is a minimal length term representing w, then the following are true.
(1) Each ti is of minimal length.
(2) The wi ’s are pairwise incomparable.
V
(3) If ti is not a generator, so ti = j tij , then ε(tij ) = wij  w for all j.
V
Proof. Only
V (3) requires explanation. Suppose wi = wij in FL(X), corresponding
to ti = tij in W (X). Note that wi ≤ wij for all j. If for some j0 we also had
wij0 ≤ w, then _ _
w= wi ≤ wij0 ∨ wk ≤ w,
k6=i
W
whence w = wij0 ∨ k6=i wk . But then replacing ti by tij0 would yield a shorter term
representing w, a contradiction. 
If A and B are finite subsets of a lattice, we say that A refines B, written A  B,
if for each a ∈ A there exists b ∈ B with a ≤ b. We define dual refinement by C  D
if for each c ∈ C there exists d ∈ D with c ≥ d; note that because of the reversed
order of the quantification in the two statements, A  B is not the same as B  A.
The elementary properties of refinement can be set out as follows, with the proofs
left as an exercise.
Lemma 6.5. The refinement relation has the following properties.
W W
(1) A  B implies A ≤ B.
(2) The relation  is a quasiorder on the finite subsets of L.
(3) If A ⊆ B then A  B.
(4) If A is an antichain, A  B and B  A, then A ⊆ B.
62
 (5) If A and B are antichains with A  B and B  A, then A = B.
(6) If A  B and B  A, then A and B have the same set of maximal elements.
The preceding two lemmas are connected as follows.
W W
Lemma 6.6. Let w = 1≤i≤k wi = 1≤j≤m uj in FL(X). If each wi is either a
V
generator or a meet wi = j wij with wij  w for all j, then
{w1 , . . . , wm }  {u1 , . . . , un }.
W
Proof. For each i we have wi ≤ uj .VIf wi is a generator, this implies wi ≤ uk for
some k by TheoremV6.2(2). WIf wi = wij , we apply Whitman’s condition (W) to
the inclusion wi = wij ≤ uk = w. Since we are given that wij  w for all j, it
must be that wi ≤ uk for some k. Hence {w1 , . . . , wm }  {u1 , . . . , un }. 
W W
Now let t = ti and s = sj be two minimal length terms which W evaluate
W
to w in FL(X). Let ε(ti ) = wi and ε(sj ) = uj , so that w = wi = uj
in FL(X). By Lemma 6.4(1) each ti is a minimal length term for wi , and each
sj is a minimal length term for uj . By induction, these are unique up to as-
sociativity and commutativity. Hence we may assume that ti = sj whenever
wi = uj . By Lemma 6.4(2), the sets {w1 , . . . , wm } and {u1 , . . . , un } are antichains
in FL(X). By Lemma 6.4(3), the elements wi satisfy the hypothesis of Lemma 6.6,
so {w1 , . . . , wm }  {u1 , . . . , un }. Symmetrically, {u1 , . . . , un }  {w1 , . . . , wm }. Ap-
plying Lemma 6.5(5) yields {w1 , . . . , wm } = {u1 , . . . , un }, whence by our assumption
above {t1 , . . . , tm } = {s1 , . . . , sn }. Thus we obtain the desired uniqueness result.
Theorem 6.7. The minimal length term for w ∈ FL(X) is unique up to associa-
tivity and commutativity.
This minimal length term is called the canonical form of w. The canonical form
of a generator is just x. The proof of the theorem has shown that if w is a proper
join, then its canonical form is determined by the conditions of Lemma 6.4. If w is
a proper meet, then of course its canonical form must satisfy the dual conditions.
The proof of LemmaW6.4 gives us an algorithm for finding the canonical form of
a lattice term. Let t = ti in W (X), where each ti is either a generator or a meet,
and suppose that we have already put each ti into canonical form, which we can do
inductively. This will guarantee that condition (1) ofVLemma 6.4 holds when we are
done. For each ti which is not a generator, say ti = tij , check whether any tij ≤ t
in FL(X);
W if so, replace ti by tij . Continue this process until you have an expression
u = ui which satisfies condition (3). Finally, check whether uiW≤ uj in FL(X)
for any pair i 6= j; if so, delete ui . The resulting expression v = vi evaluates to
the same element as t in FL(X), and v satisfies (1), (2) and (3). Hence v is the
canonical Wform of t.
If w = wi canonically in FL(X), then the elements wi are called the canonical
joinands of w (dually, canonical meetands). It is important to note that these
elements satisfy the refinement property of Lemma 6.6.
63
 W
Corollary. If w is a proper join in FL(X) and w = U , then the set of canonical
joinands of w refines U .
This has an important structural consequence, observed by Bjarni Jónsson [9].
Theorem 6.8. Free lattices satisfy the following implications, for all u, v, a, b, c ∈
FL(X):
(SD∨ ) if u = a ∨ b = a ∨ c then u = a ∨ (b ∧ c),

(SD∧ ) if v = a ∧ b = a ∧ c then v = a ∧ (b ∨ c).

The implications (SD∨ ) and (SD∧ ) are known as the semidistributive laws.
Proof. We will prove that FL(X) satisfies (SD∨ ); then (SD∧ ) follows by duality.
We may assume that u is a proper join, for otherwise u is join irreducible and the
implication is trivial. So let u = u1 ∨ . . . ∨ un be the canonical join decomposition.
By the Corollary above, {u1 , . . . , un } refines both {a, b} and {a, c}. Any ui which is
not below a must be below both b and c, so in fact {u1 , . . . , un }  {a, b ∧ c}. Hence
_
u= ui ≤ a ∨ (b ∧ c) ≤ u ,

whence u = a ∨ (b ∧ c), as desired. 

Now let us recall some basic facts about free groups, so we can ask about their
analogs for free lattices. Every subgroup of a free group is free, and the countably
generated free group F G(ω) is isomorphic to a subgroup of F G(2). Every identity
which does not hold in all groups fails in some finite group.
Whitman used Theorem 6.3 and a clever construction to show that FL(ω) can
be embedded in FL(3). It is not known exactly which lattices are isomorphic to a
sublattice of a free lattice, but certainly they are not all free. The simplest result
(to state, not to prove) along these lines is due to J. B. Nation [11].
Theorem 6.9. A finite lattice can be embedded in a free lattice if and only if it
satisfies (W ), (SD∨ ) and (SD∧ ).
We can weaken the question somewhat and ask which ordered sets can be em-
bedded in free lattices. A characterization of sorts for these ordered sets was found
by Freese and Nation ([7] and [12]), but unfortunately it is not particularly enlight-
ening. We obtain a better picture of the structure of free lattices by considering the
following collection of results due to P. Crawley and R. A. Dean [2], B.. Jónsson [9],
and J. B. Nation and J. Schmerl [13], respectively.
Theorem 6.10. Every countable ordered set can be embedded in FL(3). On the
other hand, every chain in a free lattice is countable, so no uncountable chain can
be embedded in a free lattice. If P is an infinite ordered set which can be embedded
64
in a free lattice, then the dimension d(P) ≤ m, where m is the smallest cardinal such
that |P| ≤ 2m .
R. A. Dean showed that every equation which does not hold in all lattices fails
in some finite lattice [5] (see Exercise 7.5). It turns out (though this is not obvious)
that this is related to a beautiful structural result of Alan Day ([4], using [10]).
Theorem 6.11. If X is finite, then FL(X) is weakly atomic.
The book Free Lattices by Freese, Ježek and Nation [6] contains more information
about the surprisingly rich structure of free lattices.

Exercises for Chapter 6

1. Verify that if L is a lattice and I is an interval in L, then L[I] is a lattice.
2. Use the doubling construction to repair the (W)-failures in the lattices in
Figure 6.5. (Don’t forget to double elements which are both join and meet reducible.)
Then repeat the process until you either obtain a lattice satisfying (W), or else prove
that you never will get one in finitely many steps.

(a) (b)
Figure 6.5

3. (a) Show that x ∧ ((x ∧ y) ∨ z)  y ∨ (z ∧ (x ∨ y)) in FL(X).

(b) Find the canonical form of x ∧ ((x ∧ y) ∨ (x ∧ z)).
(c) Find the canonical form of (x ∧ ((x ∧ y) ∨ (x ∧ z) ∨ (y ∧ z))) ∨ (y ∧ z).
4. There are five small lattices which fail SD∨ , but have no proper sublattice
failing SD∨ . Find them.
5. Show that the following conditions are equivalent (to SD∨ ) in a finite lattice.
(a) u = a ∨ b = a ∨ c implies u = a ∨ (b ∧ c).
(b) For each m ∈ M (L) there is a unique j ∈ J(L) such that for all x ∈ L,
m∗ ∧ x 6≤ m iff x ≥ j.
65
 W
(c) For each a ∈ L, there
W is a set C ⊆ J(L) such that a = C, and for every
W B ⊆ L,
subset W a = B impliesWC  B.
(d) u = i ui = j vj implies u = i,j (ui ∧ vj ).
In a finite lattice satisfying these conditions, the elements of the set C given by part
(c) are called the canonical joinands of a.
6. An element p ∈ L is join prime if p ≤ x ∨ y implies p ≤ x or p ≤ y; meet
prime is defined dually. Let JP(L) denote the set of all join prime elements of L,
and let MP(L) denote the set of all meet prime elements of L. Let L be a finite
lattice satisfying SD∨ .
(a) Prove that the canonical joinands of 1 are join prime.
(b) Prove that the coatoms of L are meet prime.
(c) Show that for each q ∈ MP(L) there exists a unique element η(q) ∈ JP(L)
such that L is the disjoint union of q/0 and 1/η(q).
7. Prove Lemma 6.5.
8. Let A and B be lattices, and let X ⊆ A generate A. Prove that a map
h0 : X → B can be extended to a homomorphism h : A → B if and only if, for every
pair of lattice terms p and q, and all x1 , . . . , xn ∈ X,

p(x1 , . . . , xn ) = q(x1 , . . . , xn ) implies p(h0 (x1 ), . . . , h0 (xn )) = q(h0 (x1 ), . . . , h0 (xn )).

9. A complete lattice L has canonical decompositions if for eachVa ∈ L there exists

a set C ofVcompletely meet irreducible elements such that a = C irredundantly,
and a = B implies C  B. Prove that an upper continuous lattice has canon-
ical decompositions if and only if it is strongly atomic and satisfies SD∧ (Viktor
Gorbunov [8]).
For any ordered set P, a lattice F is said to be freely generated by P if F contains
a subset P such that
(1) P with the order it inherits from F is isomorphic to P,
(2) P generates F,
(3) for every lattice L, every order preserving map h0 : P → L can be extended
to a homomorphism h : F → L.
In much the same way as with free lattices, we can show that there is a unique (up
to isomorphism) lattice FL(P) generated by any ordered set P. Indeed, free lattices
FL(X) are just the case when P is an antichain.
10. (a) Find the lattice freely generated by {x, y, z} with x ≥ y.
(b) Find FL(P) for P = {x0 , x1 , x2 , z} with x0 ≤ x1 ≤ x2 .
The lattice freely generated by Q = {x0 , x1 , x2 , x3 , z} with x0 ≤ x1 ≤ x2 ≤ x3
is infinite, as is that generated by R = {x0 , x1 , y0 , y1 } with x0 ≤ x1 and y0 ≤ y1
(Yu. I. Sorkin [17], see [14]).
66
 11. A homomorphism h : L → K is lower bounded if for each a ∈ K, {x ∈ L :
h(x) ≥ a} is either empty or has a least element β(a). For example, if L satisfies
the DCC, then h is lower bounded. We regard β as a partial map from K to L. Let
h : L → K be a lower bounded homomorphism.
(a) Show that the domain of β is an ideal of K.
(b) Prove that β preserves finite joins.
(c) Show that if h is onto and L satisfies SD∨ , then so does K.

References
1. G. Birkhoff, On the structure of abstract algebras, Proc. Cambridge Phil. Soc. 31 (1935),
433–454.
2. P. Crawley and R. A. Dean, Free lattices with infinite operations, Trans. Amer. Math. Soc. 92
(1959), 35–47.
3. A. Day, A simple solution of the word problem for lattices, Canad. Math. Bull. 13 (1970),
253–254.
4. A. Day, Splitting lattices generate all lattices, Algebra Universalis 7 (1977), 163–170.
5. R. A. Dean, Component subsets of the free lattice on n generators, Proc. Amer. Math. Soc. 7
(1956), 220–226.
6. R. Freese, J. Ježek, and J. B. Nation, Free Lattices, Mathematical Surveys and Monographs,
vol. 42, Amer. Math. Soc., Providence, 1995.
7. R. Freese and J. B. Nation, Projective lattices, Pacific J. Math. 75 (1978), 93–106.
8. V. A. Gorbunov, Canonical decompositions in complete lattices, Algebra i Logika 17 (1978),
495–511,622. (Russian)
9. B. Jónsson, Sublattices of a free lattice, Canad. J. Math. 13 (1961), 256–264.
10. R. McKenzie, Equational bases and non-modular lattice varieties, Trans. Amer. Math. Soc.
174 (1972), 1–43.
11. J. B. Nation, Finite sublattices of a free lattice, Trans. Amer. Math. Soc. 269 (1982), 311–337.
12. J. B. Nation, On partially ordered sets embeddable in a free lattice, Algebra Universalis 18
(1984), 327–333.
13. J. B. Nation and J. H. Schmerl, The order dimension of relatively free lattices, Order 7 (1990),
97–99.
14. H. L. Rolf, The free lattice generated by a set of chains, Pacific J. Math. 8 (1958), 585–595.
15. T. Skolem, Logisch-kombinatorische Untersuchungen über die Erfüllbarkeit und Beweisbarkeit
mathematischen Sätze nebst einem Theoreme über dichte Mengen, Videnskapsselskapets skrifter
I, Matematisk-naturvidenskabelig klasse, Videnskabsakademiet i Kristiania 4 (1920), 1–36.
16. T. Skolem, Select Works in Logic, Scandinavian University Books, Oslo, 1970.
17. Yu. I. Sorkin, Free unions of lattices, Mat. Sbornik 30 (1952), 677–694. (Russian)
18. Ph. M. Whitman, Free lattices, Ann. of Math. (2) 42 (1941), 325–330.

67
 7. Varieties of Lattices

Variety is the spice of life.

A lattice equation is an expression p ≈ q where p and q are lattice terms. Our

intuitive notion of what it means for a lattice L to satisfy p ≈ q (that p(x1 , . . . , xn ) =
q(x1 , . . . , xn ) whenever elements of L are substituted for the variables) is captured
by the formal definition: L satisfies p ≈ q if h(p) = h(q) for every homomorphism
h : W (X) → L. We say that L satisfies a set Σ of equations if L satisfies every
equation in Σ. As long as we are dealing entirely with lattices, there is no loss of
generality in replacing p and q by the corresponding elements of FL(X), and in
practice it is often more simple and natural to do so (as in Theorem 7.2 below).
A variety (or equational class) of lattices is the class of all lattices satisfying some
set Σ of lattice equations. You are already familiar with several lattice varieties:
(1) the variety T of one-element lattices, satisfying x ≈ y (not very exciting);
(2) the variety D of distributive lattices, satisfying x ∧ (y ∨ z) ≈ (x ∧ y) ∨ (x ∧ z);
(3) the variety M of modular lattices, satisfying (x∨y)∧(x∨z) ≈ x∨(z∧(x∨y));
(4) the variety L of all lattices, satisfying x ≈ x.
If K is any class of lattices, we say that a lattice F is K-freely generated by its
subset X if
(1) F ∈ K,
(2) X generates F,
(3) for every lattice L ∈ K, every map h0 : X → L can be extended to a
homomorphism h : F → L.
A lattice is K-free if it is K-freely generated by one of its subsets, and relatively free
if it is K-free for some (unspecified) class K.
While these ideas floated around for some time before, it was Garrett Birkhoff [5]
who proved the basic theorem about varieties in the 1930’s.
Theorem 7.1. If K is a nonempty class of lattices, then the following are equiva-
lent.
(1) K is a variety.
(2) K is closed under the formation of homomorphic images, sublattices and
direct products.
(3) Either K = T (the variety of one-element lattices), or for every nonempty
set X there is a lattice FK (X) which is K-freely generated by X, and K is
closed under homomorphic images.
68
Proof. It is easy to see that varieties are closed under homomorphic images, sublat-
tices and direct products, so (1) implies (2).
The crucial step in the equivalence, the construction of relatively free lattices
FK (X), is a straightforward adaptation of the construction of FL(X). Let K be
a class which
T is closed under the formation of sublattices and direct products, and
let κ = {θ ∈ Con W (X) : W (X)/θ ∈ K}. Following the proof of Theorem 6.1,
we can show that W (X)/κ is a subdirect product of lattices in K, and that it is
K-freely generated by {xκ : x ∈ X}. Unless K = T, the classes xκ (x ∈ X) will be
distinct. Thus (2) implies (3).
Finally, suppose that K is a class of lattices which is closed under homomorphic
images and contains a K-freely generated lattice FK (X) for every nonempty set X.
For each nonempty X there is a homomorphism fX : W (X)  FK (X) which is
the identity on X. Fix the countably infinite set X0 = {x1 , x2 , x3 , . . . }, and let Σ
be the collection of all equations p ≈ q such that (p, q) ∈ ker fX0 . Thus p ≈ q is
in Σ if and only if p(x1 , . . . , xn ) = q(x1 , . . . , xn ) in the countably generated lattice
FK (X0 ) ∼= FK (ω).
Let VΣ be the variety of all lattices satisfying Σ; we want to show that K = VΣ .
We formulate the critical argument as a sublemma.
Sublemma. Let FK (Y ) be a relatively free lattice. Let p, q ∈ W (Y ) and let fY :
W (Y )  FK (Y ) with fY the identity on Y . Then K satisfies p ≈ q if and only if
fY (p) = fY (q).
Proof. If K satisfies p ≈ q, then fY (p) = fY (q) because FK (Y ) ∈ K. Conversely,
if fY (p) = fY (q), then by the mapping property (III) every lattice in K satisfies
p ≈ q.1 
Applying the Sublemma with Y = X0 , we conclude that K satisfies every equa-
tion of Σ, so K ⊆ VΣ .
Conversely, let L ∈ VΣ , and let X be a generating set for L. The identity map
on X extends to a surjective homomorphism h : W (X)  L, and we also have
the map fX : W (X) → FK (X). For any pair (p, q) ∈ ker fX , the Sublemma says
that K satisfies p ≈ q. Again by the Sublemma, there is a corresponding equation
in Σ (perhaps involving different variables). Since L ∈ VΣ this implies h(p) =
h(q). So ker fX ≤ ker h, and hence by the Second Isomorphism Theorem there is
a homomorphism g : FK (X) → L such that h = gfX . Thus L is a homomorphic
image of FK (X). Since K is closed under homomorphic images, this implies L ∈ K.
Hence VΣ ⊆ K, and equality follows. Therefore (3) implies (1). 
The three parts of Theorem 7.1 reflect three different ways of looking at varieties.
The first is to start with a set Σ of equations, and to consider the variety V (Σ) of all
1 However, if Y is finite and Y ⊆ Z, then FK (Y ) may satisfy equations not satisfied by FK (Z).
For example, for any lattice variety, FK (2) is distributive. The Sublemma only applies to equations
with at most |Y | variables.
69
lattices satisfying those equations. The given equations will in general imply other
equations, viz., all the relations holding in the relatively free lattices FV (Σ) (X). It
is important to notice that while the proof of Birkhoff’s theorem tells us abstractly
how to construct relatively free lattices, it does not tell us how to solve the word
problem for them. Consider the variety M of modular lattices. Richard Dedekind [6]
showed in the 1890’s that FM (3) has 28 elements; it is drawn in Figure 9.2. On the
other hand, Ralph Freese [9] proved in 1980 that the word problem for FM (5) is
unsolvable: there is no algorithm for determining whether p = q in FM (5). Christian
Herrmann [10] later showed that the word problem for FM (4) is also unsolvable. It
follows, by the way, that the variety of modular lattices is not generated by its finite
members:2 there is a lattice equation which holds in all finite modular lattices, but
not in all modular lattices.
Skipping to the third statement of Theorem 7.1, let V be a variety, and let κ be
the kernel of the natural homomorphism h : FL(X) → FV (X) with h(x) = x for all
x ∈ X. Then, of course, FV (X) ∼ = FL(X)/κ. We want to ask which congruences on
FL(X) arise in this way, i.e., for which θ ∈ Con FL(X) is FL(X)/θ relatively free?
To answer this, we need a couple of definitions.
An endomorphism of a lattice L is a homomorphism f : L → L. The set of
endomorphisms of L forms a semigroup End L under composition. It is worth
noting that an endomorphism of a lattice is determined by its action on a generating
set, since f (p(x1 , . . . , xn ) = p(f (x1 ), . . . , f (xn )) for any lattice term p. In particular,
an endomorphism f of FL(X) corresponds to a substitution xi 7→ f (xi ) of elements
for the generators.
A congruence relation θ is fully invariant if (x, y) ∈ θ implies (f (x), f (y)) ∈ θ for
every endomorphism f of L. The fully invariant congruences of L can be thought
of as the congruence relations of the algebra L∗ = (L, ∧, ∨, {f : f ∈ End L}). In
particular, they form an algebraic lattice, in fact a complete sublattice of Con L.
The answer to our question, in these terms, is again due to Garrett Birkhoff [4].
Theorem 7.2. FL(X)/θ is relatively freely generated by {xθ : x ∈ X} if and only
if θ is fully invariant.
Proof. Let V be a lattice variety and let h : FL(X) → FV (X) with h(x) = x for
all x ∈ X. Then h(p) = h(q) if and only if V satisfies p ≈ q (as in the Sublemma).
Hence, for any endomorphism f and elements p, q ∈ FL(X), if h(p) = h(q) then

hf (p) = h(f (p(x1 , . . . , xn ))) = h(p(f (x1 ), . . . , f (xn )))

= h(q(f (x1 ), . . . , f (xn )))
= h(f (q(x1 , . . . , xn ))) = hf (q)

2 If a variety V of algebras (1) has only finitely many operation symbols, (2) is finitely based,
and (3) is generated by its finite members, then the word problem for FV (X) is solvable. This
result is due to A. I. Malcev for groups; see T. Evans [7].
70
so that (f (p), f (q) ∈ ker h. Thus ker h is fully invariant.
Conversely, assume that θ is a fully invariant congruence on FL(X). If θ =
1Con FL(X) , then θ is fully invariant and FL(X)/θ is relatively free for the trivial
variety T. So without loss of generality, θ is not the universal relation. Let k :
FL(X) → FL(X)/θ be the canonical homomorphism with ker k = θ. Let V be the
variety determined by the set of equations Σ = {p ≈ q : (p, q) ∈ θ}. To show that
FL(X)/θ is V-freely generated by {xθ : x ∈ X}, we must verify that
(1) FL(X)/θ ∈ V, and
(2) if M ∈ V and h0 : X → M , then there is a homomorphism h : FL(X)/θ →
M such that h(xθ) = h0 (x), i.e., hk(x) = h0 (x) for all x ∈ X.
For (1), we must show that the lattice FL(X)/θ satisfies every equation of Σ, i.e.,
that if p(x1 , . . . , xn ) θ q(x1 , . . . , xn ) and w1 , . . . , wn are elements of FL(X), then
p(w1 , . . . , wn ) θ q(w1 , . . . , wn ). Since there is an endomorphism f of FL(X) with
f (xi ) = wi for all i, this follows from the fact that θ is fully invariant.
To prove (2), let g : FL(X) → M be the homomorphism such that g(x) = h0 (x)
for all x ∈ X. Since M is in V, g(p) = g(q) whenever p ≈ q is in Σ, and thus
θ = ker k ≤ ker g. By the Second Isomorphism Theorem, there is a homomorphism
h : FL(X)/θ → M such that hk = g, as desired. 
It follows that varieties of lattices are in one-to-one correspondence with fully
invariant congruences on FL(ω). The consequences of this fact can be summarized
as follows.
Theorem 7.3. The set of all lattice varieties ordered by containment forms a lattice
Λ dually isomorphic to the lattice of all fully invariant congruences of FL(ω). Thus
Λ is dually algebraic, and a variety V is dually compact in Λ if and only if V =
V (Σ) for some finite set of equations Σ.
Going back to statement (2) of Theorem 7.1, the third way of looking at varieties
is model theoretic: a variety is a class of lattices closed under the operators H
(homomorphic images), S (sublattices) and P (direct products). Now elementary
arguments show that, for any class K,

PS(K) ⊆ SP(K)
PH(K) ⊆ HP(K)
SH(K) ⊆ HS(K).

Thus the smallest variety containing a class K of lattices is HSP(K), the class of all
homomorphic images of sublattices of direct products of lattices in K. We refer to
HSP(K) as the variety generated by K. We can think of HSP as a closure operator,
but not an algebraic one: Λ is not upper continuous, so it cannot be algebraic (see
Exercise 5). The many advantages of this point of view will soon become apparent.
71
Lemma 7.4. Two lattice varieties are equal if and only if they contain the same
subdirectly irreducible lattices.
Proof. Recall from Theorem 5.6 that every lattice L is a subdirect product of subdi-
rectly irreducible lattices L/θ with θ completely meet irreducible in Con L. Suppose
V and K are varieties, and that the subdirectly irreducible lattices of V are all in
K. Then for any X the relatively free lattice FV (X), being a subdirect product of
subdirectly irreducible lattices FV (X)/θ in V, is a subdirect product of lattices in
K. Hence FV (X) ∈ K and V ⊆ K. The lemma follows by symmetry. 
This leads us directly to a crucial question: If K is a set of lattices, how can
we find the subdirectly irreducible lattices in HSP(K)? The answer, due to Bjarni
Jónsson, requires that we once again venture into the world of logic.
Let us recall that a filter (or dual ideal) of a lattice L with greatest element 1 is
a subset F of L such that
(1) 1 ∈ F ,
(2) x, y ∈ F implies x ∧ y ∈ F ,
(3) z ≥ x ∈ F implies z ∈ F .
For any x ∈ L, the set 1/x is called a principal filter. As an example of a nonprincipal
filter, in the lattice P(X) of all subsets of an infinite set X we have the filter F of all
complements of finite subsets of X. A maximal proper filter is called an ultrafilter.
We want to describe an important type of congruence relation on direct products.
Let Li (i ∈ I) be lattices, and let F be a filter on the lattice
Q of subsets P(I). We
define an equivalence relation ≡F on the direct product i∈I Li by

x ≡F y if {i ∈ I : xi = yi } ∈ F.

A routine check shows that ≡F is a congruence relation.

Lemma 7.5. (1) Let L be a lattice, F a filter on L, and a 6∈ F . Then there exists
a filter G on L maximal with respect to the properties F ⊆ G and a 6∈ G.
(2) A proper filter U on P(I) is an ultrafilter if and only if for every A ⊆ I,
either A ∈ U or I − A ∈ U .
(3) If U is an ultrafilter on P(I), then its complement P(I) − U is a maximal
proper ideal.
(4) If U is an ultrafilter and A1 ∪ · · · ∪ An ∈ U , then Ai ∈ U for some i.
(5) An ultrafilter U is nonprincipal if and only if it contains the filter of all
complements of finite subsets of I.
Proof. Part (1) is a straightforward Zorn’s Lemma argument. Moreover, it is clear
that a proper filter U is maximal if and only if for every A ∈
/ U there exists B ∈ U
such that A ∩ B = ∅, i.e., B ⊆ I − A. Thus U is an ultrafilter if and only if A ∈/U
implies I − A ∈ U , which is (2). DeMorgan’s Laws then yield (3), which in turn
implies (4). It follows from (4) that if an ultrafilter U on I contains a finite set,
72
then it contains a singleton {i0 }, and hence is principal with U = 1/{i0 } = {A ⊆ I :
i0 ∈ A}. Conversely, if U is a principal ultrafilter 1/S, then S must be a singleton.
Thus an ultrafilter is nonprincipal if and only if it contains no finite set, which by
(2) means that it contains the complement of every finite set. 
Corollary. If I is an infinite set, then there is a nonprincipal ultrafilter on P(I).
Proof. Apply Lemma 7.5(1) with L = P(I), F the filter of all complements of finite
subsets of I, and a = ∅. 
Q
If F is a filter on P(I), the quotient
Q lattice i∈I Li / ≡F is called a reduced
product. If U is an ultrafilter, then i∈I Li / ≡U is an ultraproduct. The interesting
case is when U is a nonprincipal ultrafilter. Good references on reduced products
and ultraproducts are [3] and [8].
Our next immediate goal is to investigate what properties are preserved by the
ultraproduct construction. In order to be precise, we begin with a slough of defini-
tions, reserving comment for later.
The elements of a first order language for lattices are
(1) a countable alphabet X = {x1 , x2 , x3 , . . . },
(2) equations p ≈ q with p, q ∈ W (X),
(3) logical connectives and, or, and ¬,
(4) quantifiers ∀xi and ∃xi for i = 1, 2, 3 . . . .
These symbols can be combined appropriately to form well formed formulas (wffs)
by the following rules.
(1) Every equation p ≈ q is a wff.
(2) If α and β are wffs, then so are (¬α), (α and β) and (α or β).
(3) If γ is a wff and i ∈ {1, 2, 3, . . . }, then ∀xi γ and ∃xi γ are wffs.
(4) Only expressions generated by the first three rules are wffs.
Now let L be a lattice, let h : W (X) → L be a homomorphism, and let ϕ be a
well formed formula. We say that the pair (L, h) models ϕ, written symbolically as
(L, h) |= ϕ, according to the following recursive definition.
(1) (L, h) |= p ≈ q if h(p) = h(q), i.e., if p(h(x1 ), . . . , h(xn )) = q(h(x1 ), . . . , h(xn )).
(2) (L, h) |= (¬α) if (L, h) does not model α (written (L, h) 6|= α).
(3) (L, h) |= (α and β) if (L, h) |= α and (L, h) |= β.
(4) (L, h) |= (α or β) if (L, h) |= α or (L, h) |= β (or both).
(5) (L, h) |= ∀xi γ if (L, g) |= γ for every g such that g|X−{xi } = h|X−{xi } .
(6) (L, h) |= ∃xi γ if (L, g) |= γ for some g such that g|X−{xi } = h|X−{xi } .
(For Y ⊆ X, g|Y denotes the restriction of g to Y .)
We say that L satisfies ϕ (or L models ϕ) if (L, h) models ϕ for every homomor-
phism h : W (X) → L.
We are particularly interested in well formed formulas ϕ for which all the variables
appearing in ϕ are quantified (by ∀ or ∃). The set Fϕ of variables that occur freely
73
in ϕ is defined recursively as follows.
(1) For an equation, Fp≈q is the set of all variables xi which actually appear in
p or q.
(2) F¬α = Fα .
(3) Fα and β = Fα ∪ Fβ .
(4) Fα or β = Fα ∪ Fβ .
(5) F∀xi α = Fα − {xi }.
(6) F∃xi α = Fα − {xi }.
A first order sentence is a well formed formula ϕ such that Fϕ is empty, i.e., no
variable occurs freely in ϕ. It is not hard to show inductively that, for a given
lattice L and any well formed formula ϕ, whether or not (L, h) |= ϕ depends only
on the values of h|Fϕ , i.e., if g|Fϕ = h|Fϕ , then (L, g) |= ϕ iff (L, h) |= ϕ. So if ϕ is
a sentence, then either L satisfies ϕ or L satisfies ¬ϕ.
Now some comments are in order. First of all, we did not include the predicate
p ≤ q because we can capture it with the equation p ∨ q ≈ q. Likewise, the logical
connective =⇒ is omitted because (α =⇒ β) is equivalent to (¬α) or β. On the
other hand, our language is redundant because or can be eliminated by the use of
DeMorgan’s law, and ∃xi ϕ is equivalent to ¬∀xi (¬ϕ).
Secondly, for any well formed formula ϕ, a lattice L satisfies ϕ if and only if it
satisfies the sentence ∀xi1 . . . ∀xik ϕ where the quantification runs over the variables
in Fϕ . Thus we can consistently speak of a lattice satisfying an equation or Whit-
man’s condition, for example, when what we really have in mind is the corresponding
universally quantified sentence.
Fortunately, our intuition about what sort of properties can be expressed as first
order sentences, and what it means for a lattice to satisfy a sentence ϕ, tends to be
pretty good, particularly after we have seen a lot of examples. With this in mind,
let us list some first order properties.
(1) L satisfies p ≈ q.
(2) L satisfies the semidistributive laws (SD∨ ) and (SD∧ ).
(3) L satisfies Whitman’s condition (W ).
(4) L has width 7.
(5) L has at most 7 elements.
(6) L has exactly 7 elements.
(7) L is isomorphic to M5 .
And, of course, we can do negations and finite conjunctions and disjunctions of these.
The sort of things which cannot be expressed by first order sentences includes the
following.
(1) L is finite.
(2) L satisfies the ACC.
(3) L has finite width.
(4) L is subdirectly irreducible.
74
 Now we are in a position to state for lattices the fundamental theorem about
ultraproducts, due to J. Los in 1955 [13].
Theorem 7.6. Let ϕ be a first order lattice Q sentence, Li (i ∈ I) lattices, and U
an ultrafilter on P(I). Then the ultraproduct i∈I Li / ≡U satisfies ϕ if and only if
{i ∈ I : Li satisfies ϕ} is in U .
Q
Corollary. If each Li satisfies ϕ, then so does the ultraproduct i∈I Li / ≡U .

Q Li (i ∈ I) and an ultrafilter U
Proof. Suppose we have a collection of lattices
on P(I). The elements of the ultraproduct Q i∈I Li /Q ≡U are equivalence classes
of elements of the direct product.
Q Let µ : Li → Li / ≡U be the canonical
homomorphism, and let πj : Li → Lj denote the projection map. We will prove
the following claim, which includes Theorem 7.6.
Q
Claim. Let h : W Q (X) → i∈I Li be a homomorphism, and let ϕ be a well formed
formula. Then ( Li / ≡U , µh) |= ϕ if and only if {i ∈ I : (Li , πi h) |= ϕ} ∈ U .
We proceed by induction on the complexity of ϕ. In view of the observations
above (e.g., DeMorgan’s Laws), it suffices to treat equations, and, ¬ and ∀. The
first three are quite straightforward.
Q
Note that for a, b ∈ Li we have µ(a) = µ(b) if and only if {i : πi (a) = πi (b)} ∈
U . Thus, for an equation p ≈ q, we have
Y
( Li / ≡U , µh) |= p ≈ q iff µh(p) = µh(q)
iff {i : πi h(p) = πi h(q)} ∈ U
iff {i : (Li , πi h) |= p ≈ q} ∈ U.

.
For a conjunction α and β, using A ∩ B ∈ U iff A ∈ U and B ∈ U , we have
Y Y Y
( Li / ≡U , µh) |= α and β iff ( Li / ≡U , µh) |= α and ( Li / ≡U , µh) |= β
iff {i : (Li , πi h) |= α} ∈ U and {i : (Li , πi h) |= β} ∈ U
iff {i : (Li , πi h) |= α and β} ∈ U.

.
For a negation ¬α, using the fact that A ∈ U iff I − A ∈
/ U , we have
Y Y
( Li / ≡U , µh) |= ¬α iff ( Li / ≡U , µh) 6|= α
iff {i : (Li , πi h) |= α} ∈
/U
iff {j : (Lj , πj h) 6|= α} ∈ U
iff {j : (Lj , πj h) |= ¬α} ∈ U.
75
.
Finally, we consider the case when ϕ has the Q form ∀xγ. First, assume A = {i :
(Li , πi h) |= ∀xγ} ∈ U , and let g : W (X) → Li be a homomorphism such that
µg|X−{x} = µh|X−{x} . This means that for each y ∈ X − {x}, the set By = {j :
πj g(y) = πj h(y)} ∈ T U . Since Fγ is a finite set and U is closed under intersection,
it follows that B = y∈Fγ −{x} By = {j : πj g(y) = πj h(y) for all y ∈ Fγ − {x}} ∈
U . Therefore A ∩ B = {i : (Li , πi h) |= ∀xγ and πi g|FγQ −{x} = πi h|Fγ −{x} } ∈ U .
Hence
Q {i : (Li , πi g) |= γ} ∈ U , and so by induction ( Li / ≡U , µg) |= γ. Thus
( Li / ≡U , µh) |= ∀xγ, as desired.
Conversely, suppose A = {i : (Li , πi h) |= ∀xγ} ∈ / U . Then the complement
I − A = {j : (Lj , πj h) 6|= ∀xγ} ∈ U . For each j ∈ I − A, there is a homomorphism
gj : W (X)Q→ Lj such that gj |X−{x} = πj h|X−{x} and (Lj , gj ) 6|= γ. Let g :
W (X) → Li be a homomorphism
Q gj for all j ∈ I − A. Then
such that πj g = Q
µg|X−{x} = µh|X−{x} but ( Li / ≡U , µg) 6|= γ. Thus ( Li / ≡U , µh) 6|= ∀xγ.
This completes the proof of Lemma 7.6. 

To our operators H, S and P let us add a fourth: Pu (K) is the class of all
ultraproducts of lattices from K. Finally we get to answer the question: Where do
subdirectly irreducibles come from?
Theorem 7.7. Jónsson’s Lemma. Let K be a class of lattices. If L is subdirectly
irreducible and L ∈ HSP(K), then L ∈ HSPu (K).
Proof.QNow L ∈ HSP(K) means that there are lattices Ki ∈ K (i ∈ I), a sublattice
S of i∈I Ki , and a surjective homomorphism h : S  L. If we also assume that
L is finitely subdirectly irreducible (this suffices), then ker h is meet irreducible in
Con S. Since Con S is distributive, this makes ker h meet prime. Q
For any J ⊆ I, let πJ be the kernel of the projection of S onto j∈J Kj . Thus
for a, b ∈ S we have a πJ b iff aj = bj for all j ∈ J. Note that H ⊇ J implies
πH ≤ πJ , and that πJ∪K = πJ ∧ πK .
Let H = {J ⊆ I : πJ ≤ ker h}. By the preceding observations,
(1) I ∈ H and ∅ ∈/ H,
(2) H is an order filter in P(I),
(3) J ∪ K ∈ H implies J ∈ H or K ∈ H.
However, H need not be a (lattice) filter. Let us therefore consider

Q = {F ⊆ P(I) : F is a filter on P(I) and F ⊆ H}.

By Zorn’s Lemma, Q contains a maximal member with respect to set inclusion, say
U . Let us show that U is an ultrafilter.
If not, then by Lemma 7.5(2) there exists A ⊆ I such that A and I − A are both
not in U . By the maximality of U , this means that there exists a subset X ∈ U
such that A ∩ X ∈ / H. Similarly, there is a Y ∈ U such that (I − A) ∩ Y ∈
/ H. Let
76
Z = X ∩ Y . Then Z ∈ U , and hence Z ∈ H. However, A ∩ Z ⊆ A ∩ X, whence
A∩Z ∈/ H by (2) above. Likewise (I − A) ∩ Z ∈
/ H. But

(A ∩ Z) ∪ ((I − A) ∩ Z) = Z ∈ H,

contradicting (3). Thus

Q U is an ultrafilter.
Now ≡U ∈ Con Ki , and its restriction
Q is a congruence on S. Moreover, S/ ≡U
is (isomorphic to) a sublattice of Ki / ≡U . If a, b are any pair of elements of S
such that a ≡U b, then J = {i : ai = bi } ∈ U . This implies J ∈ H and so πJ ≤ ker h,
whence h(a) = h(b). Thus the restriction of ≡U to S is below ker h, wherefore
L = h(S) is a homomorphic image of S/ ≡U . We conclude that L ∈ HSPu (K). 
The proof of Jónsson’s Lemma [11] uses the distributivity of Con L in a cru-
cial way, and its conclusion is not generally true for varieties of algebras which do
not have distributive congruence lattices. This means that varieties of lattices are
more well-behaved than varieties of other algebras, such as groups and rings. The
applications below will indicate some aspects of this.
Lemma 7.8. Let U be an ultrafilterQ on P(I) and J ∈ U . Then VQ= {B ⊆ J : B ∈
U } is an ultrafilter on P(J), and j∈J Lj / ≡V is isomorphic to i∈I Li / ≡U .
Proof. V is clearly a proper filter. Moreover, if A ⊆ J and A ∈ / V , then I − A ∈ U
and hence J − A = J ∩ (I − A) ∈ U . It follows by Lemma 7.5(2) that V is an
ultrafilter. Q Q
The projection ρJ : i∈I Li  j∈J Lj is a surjective homomorphism. As A ∩
Q
J ∈ UQif and only if A ∈ U , it induces a (well defined) isomorphism of i∈I Li / ≡U
onto j∈J Lj / ≡V . 
Theorem 7.9. Let K = {K1 , . . . , Kn } be a finite collection of finite lattices. If L
is a subdirectly irreducible lattice in the variety HSP(K), then L ∈ HS(Kj ) for some
j.
Proof. ByQJónsson’s Lemma, L is a homomorphic image of a sublattice of an ultra-
product i∈I Li / ≡U with each Li isomorphic to one of K1 , . . . , Kn . Let Aj = {i ∈
I : Li ∼
= Kj }. As A1 ∪ · · · ∪ An = I ∈ U , by Lemma 7.5(4) there is a j such that
Aj ∈ U . But then Lemma 7.8 says that there Q is an ultrafilter V on P(Aj ) such that
the original ultraproduct is isomorphic to k∈Aj Lk / ≡V , wherein each Lk = Kj . ∼
However, for any finite lattice K there is a first order sentence ϕK suchQthat a lattice
M satisfies ϕK if and only if M ∼ = K. Therefore, by Los’ Theorem, k∈Aj Lk / ≡V
is isomorphic to Kj . Hence L ∈ HS(Kj ), as claimed. 
Corollary. If V = HSP(K) where K is a finite collection of finite lattices, then V
contains only finitely many subvarieties.
Note that HSP({K1 , . . . , Kn }) = HSP(K1 ×· · ·×Kn ), so w.l.o.g. we can talk about
the variety generated by a single finite lattice. The author has recently shown that
77
the converse of the Corollary is false [17]: There is an infinite, subdirectly irreducible
lattice L such that HSP(L) has only finitely many subvarieties, each of which is
generated by a finite lattice.
Let us call a variety V finitely based if V = V (Σ) for some finite set of equations
Σ. These are just the varieties which are dually compact in the lattice Λ of lattice
varieties. Ralph McKenzie [14] proved the following nice result.
Theorem 7.10. The variety generated by a finite lattice is finitely based.
Kirby Baker [1] generalized this result by showing that if A is any finite algebra
in a variety V such that (i) V has only finitely many operation symbols, and (ii)
the congruence lattices of algebras in V are distributive, then HSP(A) is finitely
based. It is also true that the variety generated by a finite group is finitely based
(S. Oates and M. B. Powell [18]), and likewise the variety generated by a finite
ring (R. Kruse [12]). See R. McKenzie [15] for a common generalization of these
finite basis theorems. There are many natural examples of finite algebras which do
not generate a finitely based variety; see, e.g., G. McNulty [16].
We will return to the varieties generated by some particular finite lattices in the
next chapter.
If V is a lattice variety, let Vsi be the class of subdirectly irreducible lattices in
V. The next result is proved by a straightforward modification of the first part of
the proof of Theorem 7.9.
Theorem 7.11. If V and W are lattice varieties, then (V ∨ W)si = Vsi ∪ Wsi .
Corollary. Λ is distributive.
Theorem 7.11 does not extend to infinite joins (finite lattices generate the variety
of all lattices - see Exercise 5). We already knew the Corollary by Theorem 7.3,
because Λ is dually isomorphic to a sublattice of Con FL(ω), which is distributive,
but this provides an interesting way of looking at it.
In closing let us consider the lattice I(L) of ideals of L. An elementary argument
shows that the map x → x/0 embeds L into I(L). A classic theorem of Garrett
Birkhoff [4] says that I(L) satisfies every identity satisfied by L, i.e., I(L) ∈ HSP(L).
The following result of Kirby Baker and Alfred Hales [2] goes one better.
Theorem 7.12. For any lattice L, we have I(L) ∈ HSPu (L).
This is an ideal place to stop.

Exercises for Chapter 7

1. Show that fully invariant congruences form a complete sublattice of Con L.
2. Let L be a lattice and V a lattice variety. Show that there is a unique minimum
congruence ρV on L such that L/ρV ∈ V.
3. (a) Prove that if L is a subdirectly irreducible lattice, then HSP(L) is (finitely)
join irreducible in the lattice Λ of lattice varieties.
78
 (b) Prove that if a variety V is completely join irreducible in Λ, then V = HSP(K)
for some finitely generated, subdirectly irreducible lattice K. Q
4. Show that if F is a filter on P(I), then ≡F is a congruence relation on i∈I Li .
5. Prove that every lattice equation which does not hold in all lattices fails in some
finite lattice. (Let p 6= q in FL(X).
V Then there exist a finite join subsemilattice S of
FL(X) containing p, q and 0 = X, and a lattice homomorphism h : FL(X) → S,
such that h(p) = p and h(q) = q.)
The standard solution to Exercise 5 involves lattices which turn out to be lower
bounded (see Exercise 11 of Chapter 6). Hence they satisfy SD∨ , and any finite
collection of them generates a variety not containing M3 , while all together they
generate the variety of all lattices. On the other hand, the variety generated by M3
contains only the variety D of distributive lattices (generated by 2) and the trivial
variety T. It follows that the lattice Λ of lattice varieties is not join continuous.
6. Give a first order sentence characterizing each of the following properties of a
lattice L (i.e., L has the property iff L |= ϕ).
(a) L has a least element.
(b) L is atomic.
(c) L is strongly atomic.
(d) L is weakly atomic.
(e) L has no covering relations.
7. A lattice L has breadth n if L contains n elements whose join is irredundant,
but every join of n + 1 elements of L is redundant.
(a) Give a first order sentence characterizing lattices of breadth n (for a fixed
finite integer n ≥ 1).
(b) Show that the class of lattices of breadth ≤ n is not a variety.
(c) Show that a lattice L and its dual Ld have the same breadth.
8. Give a first order sentence ϕ such that a lattice L satisfies ϕ if and only if L
is isomorphic to the four element lattice 2 × 2.
9. Prove Theorem 7.11.
10. Prove that I(L) is distributive if and only if L is distributive. Similarly, show
that I(L) is modular if and only if L is modular.

References
1. K. Baker, Finite equational bases for finite algebras in a congruence-distributive equational
class, Advances in Math. 24 (1977), 207–243.
2. K. Baker and A. W. Hales, From a lattice to its ideal lattice, Algebra Universalis 4 (1974),
250–258.
3. J. L. Bell and A. B. Slomson, Models and Ultraproducts: an Introduction, North-Holland,
Amsterdam, 1971.
4. G. Birkhoff, On the lattice theory of ideals, Bull. Amer. Math. Soc. 40 (1934), 613–619.
5. G. Birkhoff, On the structure of abstract algebras, Proc. Cambridge Phil. Soc. 31 (1935),
433–454.
79
 6. R. Dedekind, Über die drei Moduln erzeugte Dualgruppe, Math. Annalen 53 (1900), 371–403.
7. T. Evans, Some connections between residual finiteness, finite embeddability and the word
problem, J. London. Math. Soc. (2) 2 (1969), 399–403.
8. T. Frayne, A. Morel and D. Scott, Reduced direct products, Fund. Math. 51 (1962), 195–228.
9. R. Freese, Free modular lattices, Trans. Amer. Math. Soc. 261 (1980), 81–91.
10. C. Herrmann, Uber die von vier Moduln erzeugte Dualgruppe, Abh. Braunschweig. Wiss. Ges.
33 (1982), 157–159.
11. B. Jónsson, Algebras whose congruence lattices are distributive, Math. Scand. 21 (1967), 110–
121.
12. R. Kruse, Identities satisfied by a finite ring, J. Algebra 26 (1973), 298–318.
13. J. Los, Quelques remarques, théorèmes et problèmes sur le classes définissables d’algèbras,
Mathematical interpretations of formal systems, North-Holland, Amsterdam, 1955.
14. R. McKenzie, Equational bases for lattice theories, Math. Scand. 27 (1970), 24–38.
15. R. McKenzie, Finite equational bases for congruence modular varieties, Algebra Universalis
24 (1987), 224-250.
16. G. McNulty, How to construct finite algebras which are not finitely based, Universal Algebra
and Lattice Theory (Charleston, S.C., 1984), Lecture Notes in Math., vol. 1149, Springer,
Berlin-New York, 1985, pp. 167–174.
17. J. B. Nation, A counterexample to the finite height conjecture, Order 13 (1996), 1–9.
18. S. Oates and M. B. Powell, Identical relations in finite groups, J. Algebra 1 (1964), 11–39.

80
 8. Distributive Lattices

Every dog must have his day.

In this chapter and the next we will look at the two most important lattice
varieties: distributive and modular lattices. Let us set the context for our study of
distributive lattices by considering varieties generated by a single finite lattice. A
variety V is said to be locally finite if every finitely generated lattice in V is finite.
Equivalently, V is locally finite if the relatively free lattice FV (n) is finite for every
integer n > 0.
Theorem 8.1. If L is a finite lattice and V = HSP(L), then
|FV (n)| ≤ |L||L| .
n

Hence HSP(L) is locally finite.

Proof. If K is any collection of lattices and V = HSP(K), then FV (X) ∼ = FL(X)/θ
where θ is the intersection of all homomorphism kernels ker f such that f : FL(X) →
L for some L ∈ K. (This is the technical way of saying that FL(X)/θ satisfies exactly
the equations which hold in every member of K.) When K consists of a single finite
lattice {L} and |X| = n, then there are |L|n distinct mappings of X into L, and
hence |L|n distinct homomorphisms fi : FL(X) → L (1 ≤ i ≤ |L| n 1
T ). The range of
∼
each fi is a sublattice of L. Hence FV (X) = FL(X)/θ with θ = ker fi means that
FV (X) is a subdirect
Q product of |L|n sublattices of L, and so a sublattice of the
direct product 1≤i≤|L|n L = L|L| , making its cardinality at most |L||L| .2 
n n

We should note that not every locally finite lattice variety is generated by a finite
lattice.
Now it is clear that there is a unique minimum nontrivial lattice variety, viz., the
one generated by the two element lattice 2, which is isomorphic to a sublattice of
any nontrivial lattice. We want to show that HSP(2) is the variety of all distributive
lattices.
Lemma 8.2. The following lattice equations are equivalent.
(1) x ∧ (y ∨ z) ≈ (x ∧ y) ∨ (x ∧ z)
(2) x ∨ (y ∧ z) ≈ (x ∨ y) ∧ (x ∨ z)
(3) (x ∨ y) ∧ (x ∨ z) ∧ (y ∨ z) ≈ (x ∧ y) ∨ (x ∧ z) ∨ (y ∧ z)
1 The kernels of distinct homomorphisms need not be distinct, of course, but that is okay.
n
is sometimes useful to view this argument constructively: FV (X) is the sublattice of L|L|
2 It

generated by the vectors x (x ∈ X) with xi = fi (x) for 1 ≤ i ≤ |L|n .

81
Thus each of these equations determines the variety D of all distributive lattices.
Proof. If (1) holds in a lattice L, then for any x, y, z ∈ L we have

(x ∨ y) ∧ (x ∨ z) = [(x ∨ y) ∧ x] ∨ [(x ∨ y) ∧ z]
= x ∨ (x ∧ z) ∨ (y ∧ z)
= x ∨ (y ∧ z)

whence (2) holds. Thus (1) implies (2), and dually (2) implies (1).
Similarly, applying (1) to the left hand side of (3) yields the right hand side, so
(1) implies (3). Conversely, assume that (3) holds in a lattice L. For x ≥ y, equation
(3) reduces to x ∧ (y ∨ z) = y ∨ (x ∧ z), which is the modular law, so L must be
modular. Now for arbitrary x, y, z in L, meet x with both sides of (3) and then use
modularity to obtain

x ∧ (y ∨ z) = x ∧ [(x ∧ y) ∨ (x ∧ z) ∨ (y ∧ z)]
= (x ∧ y) ∨ (x ∧ z) ∨ (x ∧ y ∧ z)
= (x ∧ y) ∨ (x ∧ z)

since x ≥ (x ∧ y) ∨ (x ∧ z). Thus (3) implies (1). (Note that since (3) is self-dual,
the second argument actually makes the first one redundant.) 
In the first Corollary of the next chapter, we will see that a lattice is distributive
if and only if it contains neither N5 nor M3 as a sublattice. But before that, let us
look at the wonderful representation theory of distributive lattices. A few moments
reflection on the kernel of a homomorphism h : L  2 should yield the following
conclusions.3
Lemma 8.3. Let L be a lattice and h : L  2 = {0, 1} a surjective homomorphism.
Then h−1 (0) is an ideal of L, h−1 (1) is a filter, and L is the disjoint union of h−1 (0)
and h−1 (1).
Conversely, if I is an ideal of L and F a filter such that L = I ∪F ˙ (disjoint
union), then the map h : L → 2 given by

0 if x ∈ I,
h(x) =
1 if x ∈ F.

is a surjective homomorphism.
This raises the question: When is the complement L − I of an ideal a filter? The
answer is easy. A proper ideal I of a lattice L is said to be prime if x ∧ y ∈ I implies
3 This is one point where we really don’t want to assume that L has a 0 and 1. So in this
chapter, an ideal of a lattice means a nonempty subset I such that x ∨ y ∈ I whenever x, y ∈ I,
and z ∈ I whenever z ≤ x ∈ I. A filter is defined dually.
82
x ∈ I or y ∈ I. Dually, a proper filter F is prime if x ∨ y ∈ F implies x ∈ F or
y ∈ F . It is straightforward that the complement of an ideal I is a filter iff I is a
prime ideal iff L − I is a prime filter.
This simple observation allows us to work with prime ideals or prime filters (in-
terchangeably), rather than ideal/filter pairs, and we shall do so.

Theorem 8.4. Let D be a distributive lattice, and let a  b in D. Then there exists
a prime filter F with a ∈ F and b ∈
/ F.

Proof. Now 1/a is a filter of D containing a and not b, so by Zorn’s Lemma there
is a maximal such filter (with respect to set containment), say M . For any x ∈
/ M,
the filter generated by x and M must contain b, whence b ≥ x ∧ m for some m ∈ M .
Suppose x, y ∈ / M , with say b ≥ x ∧ m and b ≥ y ∧ n where m, n ∈ M . Then by
distributivity

b ≥ (x ∧ m) ∨ (y ∧ n) = (x ∨ y) ∧ (x ∨ n) ∧ (m ∨ y) ∧ (m ∨ n).

The last three terms are in M , so we must have x ∨ y ∈

/ M . Thus M is a prime
filter. 

Now let D be any distributive lattice, and let TD = {ϕ ∈ Con D : D/ϕ ∼ = 2}.
TheoremT8.4 says that if a 6= b in D, then there exists ϕ ∈ TD with (a, b) ∈
/ ϕ,
whence TD = 0 in Con D, i.e., D is a subdirect product of two element lattices.

Corollary. The two element lattice 2 is the only subdirectly irreducible distributive
lattice. Hence D = HSP(2).

Corollary. D is locally finite.

Another consequence of Theorem 8.4 is that every distributive lattice can be

embedded into a lattice of subsets, with set union and intersection as the lattice
operations.

Theorem 8.5. Let D be a distributive lattice, and let S be the set of all prime filters
of D. Then the map φ : D → P(S) by

φ(x) = {F ∈ S : x ∈ F }

is a lattice embedding.

For finite distributive lattices, this representation takes on a particularly nice

form. Recall that an element p ∈ L is said to be join prime if it is nonzero and
p ≤ x ∨ y implies p ≤ x or p ≤ y. In a finite lattice, prime filters are necessarily of
the form 1/p where p is a join prime element.
83
Theorem 8.6. Let D be a finite distributive lattice, and let J(D) denote the ordered
set of all nonzero join irreducible elements of D. Then the following are true.
(1) Every element of J(D) is join prime.
(2) D is isomorphic to the lattice of order ideals O(J(D)). W
(3) Every element a ∈ D has a unique irredundant join decomposition a = A
with A ⊆ J(D).

Proof. In a distributive lattice, every join irreducible element is join prime, because
p ≤ x ∨ y is the same as p = p ∧ (x ∨ y) = (p ∧ x) ∨ (p ∧ y).
For any finite lattice, the map φ : L → O(J(L)) given by φ(x) = x/0 ∩ J(L)
is order preserving (in fact, meet preserving) and one-to-one. To establish the
isomorphism of (2), we need to know that for a distributive lattice it is onto. If D is
distributive
W and I is an order ideal Wof J(D), then for p ∈ J(D) we have by (1) that
p ≤ I iff p ∈ I, and hence I = φ( I).
The join decomposition of (3) is then obtained by taking A to be the set of
maximal elements of a/0 ∩ J(D). 
It is clear that the same proof works if D is an algebraic distributive lattice
whose compact elements satisfy the DCC. In Theorem 10.6 we will characterize
those distributive lattices isomorphic to O(P) for some ordered set P.
As an application, we can give a neat description of the free distributive lattice
FD (n) for any finite n, which we already know to be a finite distributive lattice. Let
X = {x1 , . . . , xn }. Now it is not hard to see that any element
W in a free distributive
lattice can be written as a join of meets of generators, w = wi with wi = xi1 ∧. . .∧
xik . Another easy argument shows that the meet V of a nonempty
V proper subset of
the generators is join prime in FD (X); note that ∅ = 1 and X = 0 do not count.
(See Exercise 3). Thus the set of join irreducible elements of FD (X) is isomorphic
to the (dual of, but it is self-dual) ordered set of nonempty, proper subsets of X, and
the free distributive lattice is isomorphic to the lattice of order ideals of that. As
an example, FD (3) and its ordered set of join irreducibles are shown in Figure 8.1.
Dedekind [6] showed that |FD (3)| = 18 and |FD (4)| = 166. Several other small
values are known exactly, and the rest can be obtained in principle, but they grow
quickly (see Quackenbush [10]). While there exist more accurate expressions, the
simplest estimate is an asymptotic formula due to D. J. Kleitman:
 
n
log2 |FD (n)| ∼ .
bn/2c

The representation by sets of Theorem 8.5 does not preserve infinite joins and
meets. The corresponding characterization of complete distributive lattices which
have a complete representation as a lattice of subsets is derived from work of Alfred
Tarski and S. Papert [9], and was surely known to both of them. W An element p of
a complete lattice L is said to be completely join prime if p ≤ X implies p ≤ x
84
 x y z

x y z

x∧y x∧z y∧z

x∧y x∧z y∧z

J(FD (3)) FD (3)

Figure 8.1

for some x ∈ X. It is not necessary to assume that D is distributive in the next

theorem, though of course it will turn out to be so.
Theorem 8.7. Let D be a complete lattice. There exists a complete lattice embed-
ding φ : D → P(X) for some set X if and only if x 6≤ y in D implies there exists a
completely join prime element p with p ≤ x and p 6≤ y.
Thus, for example, the interval [0, 1] in the real numbers is a complete distributive
lattice which cannot be represented as a complete lattice of subsets of some set.
In a lattice with 0 and 1, the pair of elements a and b are said to be complements
if a ∧ b = 0 and a ∨ b = 1. A lattice is complemented if every element has at
least one complement. For example, the lattice of subspaces of a vector space is a
complemented modular lattice. In general, an element can have many complements,
but it is not hard to see that each element in a distributive lattice can have at most
one complement.
A Boolean algebra is a complemented distributive lattice. Of course, the lattice
P(X) of subsets of a set is a Boolean algebra. On the other hand, it is easy to
see that O(P) is complemented if and only if P is an antichain, in which case
O(P) = P(P). Thus every finite Boolean algebra is isomorphic to the lattice P(A)
of subsets of its atoms.
For a very different example, the finite and cofinite subsets of an infinite set form
a Boolean algebra.
85
 If we regard Boolean algebras as algebras B = hB, ∧, ∨, 0, 1, c i, then they form
a variety, and hence there is a free Boolean algebra FBA(X) generated by a set X.
If X is finite, say X = {x1 , . . . , xn }, then FBA(X) has 2n atoms, viz., all meets
z1 ∧ . . . ∧ zn where each zi is either xi or xci . Thus in this case FBA(X) ∼ = P(A)
where |A| = 2n . On the other hand, if X is infinite then FBA(X) has no atoms;
if |X| = ℵ0 , then FBA(X) is the unique (up to isomorphism) countable atomless
Boolean algebra!
Another natural example is the Boolean algebra of all clopen (closed and open)
subsets of a topological space. In fact, by adding a topology to the representation
of Theorem 8.5, we obtain the celebrated Stone representation theorem for Boolean
algebras [13]. Recall that a topological space is totally disconnected if for every pair
of distinct points x, y there is a clopen set V with x ∈ V and y ∈ / V.
Theorem 8.8. Every Boolean algebra is isomorphic to the Boolean algebra of clopen
subsets of a compact totally disconnected (Hausdorff ) space.
Proof. Let B be a distributive lattice. (We will add the other properties to make B
a Boolean algebra as we go along.) Let Fp be the set of all prime filters of B, and
for x ∈ B let
Vx = {F ∈ Fp : x ∈ F }.

The sets Vx will form a basis for the Stone topology on Fp .

With only trivial changes, the argument for Theorem 8.4 yields the following
stronger version.
Sublemma A. Let B be a distributive lattice, G a filter on B and x ∈
/ G. Then
there exists a prime filter F ∈ Fp such that G ⊆ F and x ∈
/ F.
Next we establish the basic properties of the sets Vx , all of which are easy to
prove.
(1) Vx ⊆ Vy iff x ≤ y.
(2) Vx ∩ Vy = Vx∧y .
(3) Vx ∪ Vy = Vx∨y .
(4) If B has a least element 0, then V0 = ∅. Thus Vx ∩ Vy = ∅ iff x ∧ y = 0.
(5) If B has a greatest element 1, then V1 = Fp . Thus Vx ∪ Vy = Fp iff x ∨ y = 1.
Property (3) is where we use the primality of the filters in the sets Vx .SIn particular,
the family of sets Vx is closed under finite intersections, and of course x∈B Vx = Fp ,
so we can legitimately take {Vx : x ∈ B} as a basis for a topology on Fp .
Now we would like to show that if B has a largest element 1, then Fp is a compact
space. It suffices to consider covers by basic open sets, so this follows from the next
Sublemma.
S
Sublemma B. If B has a greatest W element 1 and Sx∈S Vx = Fp , then there exists
a finite subset T ⊆ S such that T = 1, and hence x∈T Vx = Fp .
86
 W
Proof. Set I0 = { T : T ⊆ S, T finite}. If 1 ∈
/ I0 , then I0 generates an ideal I of
B with 1 ∈
/ I. By the dual of Sublemma A, there exists a prime ideal SH containing
I and not 1. Its complement B − H is a prime filter K. Then K ∈ / x∈S Vx , else
z ∈ K for some z ∈ S, whilst z ∈ I0 ⊆ B − K. This contradicts our hypothesis, so
we must have 1 ∈ I0 , as claimed. 
The argument thus far has only required that B be a distributive lattice with 1.
For the last two steps, we need B to be Boolean. Let xc denote the complement of
x in B.
First, note that by properties (4) and (5) above, Vx ∩ Vxc = ∅ and Vx ∪ Vxc = Fp .
Thus each set VxS(x ∈ B) is clopen. On the other hand, let W be a clopen set. As
it is open, W = x∈S Vx for some set S ⊆ B. But W S is also a closed subset of the
compact space Fp , and hence compact. Thus W = x∈T Vx = VW T for some finite
T ⊆ S. Therefore W is a clopen subset of Fp if and only if W = Vx for some x ∈ B.
It remains to show that Fp is totally disconnected (which makes it Hausdorff).
Let F and G be distinct prime filters on B, with say F 6⊆ G. Let x ∈ F − G. Then
F ∈ Vx and G ∈ / Vx , so that Vx is a clopen set containing F and not G. 
There are similar topological representation theorems for arbitrary distributive
lattices, the most useful being that due to Hilary Priestley in terms of ordered
topological spaces. A good introduction is in Davey and Priestley [5].
In 1904 Huntington [8] conjectured that every uniquely complemented lattice
must be distributive (and hence a Boolean algebra). It turns out that if we assume
almost any additional finiteness condition on a uniquely complemented lattice, then
it must be distributive. As an example, we have the following theorem of Garrett
Birkhoff and Morgan Ward [4].
Theorem 8.9. Every complete, atomic, uniquely complemented lattice is isomor-
phic to the Boolean algebra of all subsets of its atoms.
Other finiteness restrictions which insure that a uniquely complemented lattice
will be distributive include weak atomicity (Bandelt and Padmanabhan [3]) and
upper continuity (Bandelt [2] and Saliı̆ [11], [12] independently). A monograph
written by Saliı̆ [14] gives an excellent survey of results of this type.
Nonetheless, Huntington’s conjecture is very far from true. In 1945, R. P. Dil-
worth [7] proved that every lattice can be embedded in a uniquely complemented
lattice. (For a strengthened version, see Adams and Sichler [1]).

Exercises for Chapter 8

1. Show that a lattice L is distributive if and only if x ∨ (y ∧ z) ≥ (x ∨ y) ∧ z for
all x, y, z ∈ L. (J. Bowden)
2. (a) Prove that every maximal ideal of a distributive lattice is prime.
(b) Show that a distributive lattice D with 0 and 1 is complemented if and only
if every prime ideal of D is maximal.
87
 3. These are the details of the construction of the free distributive lattice given
in the text. Let X be a finite set.
(a) Let δ denote the kernel of the natural homomorphism from FL(X)  FD (X)
with x 7→ x. Thus u δ v iff u(x1 , . . . , xn ) = v(x1 , . . . , xn ) in all distributive
lattices. Prove that for every w ∈ FL(X) there exists w0 which is a join of
meets of generators such that w δ w0 . (Show that the set of all such elements
w is a sublattice of FL(X) containing the generators.)
(b) Let L be any lattice generated V W X, and let ∅ ⊂ Y ⊂ X. Show that
by a set
for all w ∈ L, V either w
W ≥ Y or w ≤ (X − Y ).
(c) Show that Y  (X V − Y ) in F
W D (X) by exhibiting a homomorphism
h : FD (X) → 2 with h( Y )  h( (X − Y )).
(d) Generalize these results to the case when X is a finite ordered set (as in the
next exercise).
4. Find the free distributive lattice generated by
(a) {x0 , x1 , y0 , y1 } with x0 < x1 and y0 < y1 ,
(b) {x0 , x1 , x2 , y} with x0 < x1 < x2 .
5. Let P = Q∪R ˙ be the disjoint union of two ordered sets, so that q and r are
incomparable whenever q ∈ Q, r ∈ R. Show that O(P) ∼ = O(Q) × O(R).
6. Let D be a distributive lattice with 0 and 1, and let x and y be complements
in D. Prove that D ∼ = 1/x × 1/y. (Dually, D ∼ = x/0 × y/0; in fact, 1/x ∼ = y/0 and
∼ ∼
1/y = x/0. This explains why Con L1 × L2 = Con L1 × Con L2 (Exercise 5.6).)
7. Show that the following are true in a finite distributive lattice D.
W
(a) For each join irreducible element x of D, let κ(x) = {y ∈ D : y  x}. Then
κ(x) is meet irreducible and κ(x)  x.
(b) For each x ∈ J(D), D = 1/x ∪˙ κ(x)/0.
(c) The map κ : J(D) → M (D) is an order isomorphism.
8. A join semilattice with 0 is distributive if x ≤ y ∨ z implies there exist y 0 ≤ y
and z 0 ≤ z such that x = y 0 ∨ z 0 . Prove that an algebraic lattice is distributive if
and only if its compact elements form a distributive semilattice.
9. Prove Theorem 8.7.
10. Prove Papert’s characterization of lattices of closed sets of a topological space
[9]: Let D be a complete distributive lattice. There is a topological space T and an
isomorphism φ mapping D onto the lattice of closed subsets of T , preserving finite
joins and infinite meets, if and only if x 6≤ y in D implies there exists a (finitely)
join prime element p with p ≤ x and p 6≤ y.

References
1. M. E. Adams and J. Sichler, Lattices with unique complementation, Pacific J. Math. 92 (1981),
1–13.
2. H. J. Bandelt, Complemented continuous lattices, Arch. Math. (Basel) 36 (1981), 474–475.
3. H. J. Bandelt and R. Padmanabhan, A note on lattices with unique comparable complements,
Abh. Math. Sem. Univ. Hamburg 48 (1979), 112–113.
88
 4. G. Birkhoff and M. Ward, A characterization of Boolean algebras, Ann. of Math. 40 (1939),
609–610.
5. B. Davey and H. Priestley, Introduction to Lattices and Order, Cambridge University Press,
Cambridge, 1990.
6. R. Dedekind, Über die drei Moduln erzeugte Dualgruppe, Math. Annalen 53 (1900), 371–403.
7. R. P. Dilworth, Lattices with unique complements, Trans. Amer. Math. Soc. 57 (1945), 123–
154.
8. E. V. Huntington, Sets of independent postulates for the algebra of logic, Trans. Amer. Math.
Soc. 5 (1904), 288–309.
9. S. Papert, Which distributive lattices are lattices of closed sets?, Proc. Cambridge Phil. Soc.
55 (1959), 172–176.
10. R. Quackenbush, Dedekind’s problem, Order 2 (1986), 415–417.
11. V. N. Saliı̆, A compactly generated lattice with unique complements is distributive, Mat. Za-
metki 12 (1972), 617–620. (Russian)
12. V. N. Saliı̆, A continuous uniquely complemented lattice is distributive, Fifth All-Union Conf.
Math. Logic, Abstracts of Reports, Inst. Mat. Sibirsk. Otdel. Akad. Nauk SSSR, Novosibirsk,
1979, p. 134. (Russian)
13. M. H. Stone, The theory of representations of Boolean Algebras, Trans. Amer. Math. Soc. 40
(1936), 37–111.
14. V. N. Saliı̆, Lattices with unique complements, Translations of the Amer. Math. Soc., vol. 69,
Amer. Math. Soc., Providence, R. I., 1988.

89
 9. Modular Lattices

To dance beneath the diamond sky with one hand waving free ...
–Bob Dylan

The modular law was invented by Dedekind to reflect a crucial property of the
lattice of subgroups of an abelian group, or more generally the lattice of normal
subgroups of a group. In this chapter on modular lattices you will see the lattice
theoretic versions of some familiar theorems from group theory. This will lead us
naturally to consider semimodular lattices.
Likewise, the lattice of submodules of a module over a ring is modular. Thus our
results on modular lattices apply to the lattice of ideals of a ring, or the lattice of
subspaces of a vector space. These applications make modular lattices particularly
important.
The smallest nonmodular lattice is N5 , which is called the pentagon. Dedekind’s
characterization of modular lattices is simple [3].

Theorem 9.1. A lattice is modular if and only if it does not contain the pentagon
as a sublattice.

Proof. Clearly, a modular lattice cannot contain N5 as a sublattice. Conversely,

suppose L is a nonmodular lattice. Then there exist x > y and z in L such that
x ∧ (y ∨ z) > y ∨ (x ∧ z). Now the lattice freely generated by x, y, z with x ≥ y is
shown in Figure 9.1; you should verify that it is correct. The elements x ∧ (y ∨ z),
y ∨ (x ∧ z), z, x ∧ z and y ∨ z form a pentagon there, and likewise in L. Since the
pentagon is subdirectly irreducible and x ∧ (y ∨ z)/y ∨ (x ∧ z) is the critical quotient,
these five elements are distinct. 

Birkhoff [1] showed that there is a similar characterization of distributive lattices

within the class of modular lattices. The diamond is M3 , which is the smallest
nondistributive modular lattice.

Theorem 9.2. A modular lattice is distributive if and only if it does not contain
the diamond as a sublattice.

Proof. Again clearly, a distributive lattice cannot have a sublattice isomorphic to

M3 . Conversely, let L be a nondistributive modular lattice. Then, by Lemma 8.2,
there exist x, y, z in L such that (x ∨ y) ∧ (x ∨ z) ∧ (y ∨ z) > (x ∧ y) ∨ (x ∧ z) ∨ (y ∧ z).
Now the free modular lattice FM (3) is diagrammed in Figure 9.2; again you should
90
 x∨z

x y∨z

x ∧ (y ∨ z)
z
y ∨ (x ∧ z)

y x∧z

y∧z
Figure 9.1: FL(2 + 1)

verify that it is correct.1 The interval between the two elements above is a diamond
in FM (3), and the corresponding elements will form a diamond in L.
The details go as follows. The middle elements of our diamond should be

[x ∧ (y ∨ z)] ∨ (y ∧ z) = [x ∨ (y ∧ z)] ∧ (y ∨ z)
[y ∧ (x ∨ z)] ∨ (x ∧ z) = [y ∨ (x ∧ z)] ∧ (x ∨ z)
[z ∧ (x ∨ y)] ∨ (x ∧ y) = [z ∨ (x ∧ y)] ∧ (x ∨ y)

where in each case the equality follows from modularity. The join of the first pair
of elements is (using the first expressions)

[x ∧ (y ∨ z)] ∨ (y ∧ z) ∨ [y ∧ (x ∨ z)] ∨ (x ∨ z) = [x ∧ (y ∨ z)] ∨ [y ∧ (x ∨ z)]

= [(x ∧ (y ∨ z)) ∨ y] ∧ (x ∨ z)
= (x ∨ y) ∧ (x ∨ z) ∧ (y ∨ z).

Symmetrically, the other pairs of elements also join to (x ∨ y) ∧ (x ∨ z) ∧ (y ∨ z).

Since the second expression for each element is dual to the first, each pair of these
three elements meets to (x ∧ y) ∨ (x ∧ z) ∨ (y ∧ z). Because the diamond is simple, the
five elements will be distinct, and hence form a sublattice isomorphic to M3 . 

1 Recall from Chapter 7, though, that F (n) is infinite and has an unsolvable word problem
M
for n ≥ 4.
91
 x y z

Figure 9.2: FM (3)

Corollary. A lattice is distributive if and only if it has neither N5 nor M3 as a

sublattice.
The preceding two results tell us something more about the bottom of the lattice
Λ of lattice varieties. We already know that the trivial variety T is uniquely covered
by D = HSP(2), which is in turn covered by HSP(N5 ) and HSP(M3 ). By the
Corollary, these are the only two varieties covering D.
Much more is known about the bottom of Λ. Both HSP(N5 ) and HSP(M3 ) are
covered by their join HSP{N5 , M3 } = HSP(N5 × M3 ). George Grätzer and Bjarni
Jónsson ([6], [7]) showed that HSP(M3 ) has two additional covers, and Jónsson and
Ivan Rival [8] proved that HSP(N5 ) has exactly fifteen other covers, each generated
by a finite subdirectly irreducible lattice. You are encouraged to try and find these
covers. Because of Jónsson’s Lemma, it is never hard to tell if HSP(K) covers
HSP(L) when K and L are finite lattices; the hard part is determining whether
your list of covers is complete. Since a variety generated by a finite lattice can have
infinitely many covering varieties, or a covering variety generated by an infinite
subdirectly irreducible lattice, this can only be done near the bottom of Λ.
Now we return to modular lattices. For any two elements a, b in a lattice L there
92
are natural maps µa : (a ∨ b)/b → a/(a ∧ b) and νb : a/(a ∧ b) → (a ∨ b)/b given by

µa (x) = x ∧ a
νb (x) = x ∨ b.

Dedekind showed that these maps play a special role in the structure of modular
lattices.
Theorem 9.3. If a and b are elements of a modular lattice L, then µa and νb are
mutually inverse isomorphisms, whence (a ∨ b)/b ∼
= a/(a ∧ b).
Proof. Clearly, µa and νb are order preserving. They are mutually inverse maps by
modularity: for if x ∈ (a ∨ b)/b, then

νb µa (x) = b ∨ (a ∧ x) = (b ∨ a) ∧ x = x

and, dually, µa νb (y) = y for all y ∈ a/(a ∧ b). 

Corollary. In a modular lattice, a  a ∧ b if and only if a ∨ b  b.
For groups we actually have somewhat more. The First Isomorphism Theorem
says that if A and B are subgroups of a group G, and B is normal in A ∨ B, then
the quotient groups A/A ∧ B and A ∨ B/B are isomorphic.
A lattice L is said to be semimodular (or upper semimodular) if a  a ∧ b implies
a ∨ b  b in L. Equivalently, L is semimodular if u  v implies u ∨ x  v ∨ x, where
a  b means a covers or equals b. The dual property is called lower semimodular.
Traditionally, semimodular by itself always refers to upper semimodularity. Clearly
the Corollary shows that modular lattices are both upper and lower semimodular.
A strongly atomic, algebraic lattice which is both upper and lower semimodular is
modular. (See Theorem 3.7 of [2]; you are asked to prove the finite dimensional
version of this in Exercise 3.)
Our next result is a version of the Jordan-Hölder Theorem for semimodular lat-
tices, first proved for modular lattices by Dedekind in 1897.
Theorem 9.4. Let L be a semimodular lattice and let a < b in L. If there is a
finite maximal chain from a to b, then every chain from a to b is finite, and all the
maximal ones have the same length.
Proof. We are given that there is a finite maximal chain in b/a, say

a = a0 ≺ a1 ≺ · · · ≺ an = b.

If n = 1, i.e., a ≺ b, then the theorem is trivially true. So we may assume inductively

that it holds for any interval containing a maximal chain of length less than n.
Let C be another maximal chain in b/a. If, perchance, c ≥ a1 for all c ∈ C − {a},
then C − {a} is a maximal chain in b/a1 . In that case, C − {a} has length n − 1 by
induction, and so C has length n.
93
 Thus we may assume that there is an element d ∈ C − {a} such that d  a1 .
Moreover, since b/a1 has finite length, we can choose d such that d ∨ a1 is minimal,
i.e., e ∨ a1 ≥ d ∨ a1 for all e ∈ C − {a}. We can show that d  a as follows. Suppose
not. Then d > e > a for some e ∈ L; since C is a maximal chain containing a and
d, we can choose e ∈ C. Now a1  a = d ∧ a1 = e ∧ a1 . Hence by semimodularity
d ∨ a1  d and e ∨ a1  e. But the choice of d implies e ∨ a1 ≥ d ∨ a1  d > e,
contradicting the second covering relation. Therefore d  a.
Now we are quickly done. As a1 and d both cover a, their join a1 ∨ d covers both
of them. Since a1 ∨ d  a1 , every maximal chain in b/(a1 ∨ d) has length n − 2.
Then every chain in b/d has length n − 1, and C has length n, as desired. 
Now let L be a semimodular lattice in which every principal ideal x/0 has a finite
maximal chain. Then we can define a dimension function δ on L by letting δ(x) be
the length of a maximal chain from 0 to x:

δ(x) = n if 0 = c0 ≺ c1 ≺ · · · ≺ cn = x.

By Theorem 9.4, δ is well defined. For semimodular lattices the properties of the
dimension function can be summarized as follows.
Theorem 9.5. If L is a semimodular lattice and every principal ideal has only finite
maximal chains, then the dimension function on L has the following properties.
(1) δ(0) = 0,
(2) x > y implies δ(x) > δ(y),
(3) x  y implies δ(x) = δ(y) + 1,
(4) δ(x ∨ y) + δ(x ∧ y) ≤ δ(x) + δ(y).
Conversely, if L is a lattice which admits an integer valued function δ satisfying (1)–
(4), then L is semimodular and principal ideals have only finite maximal chains.
Proof. Given a semimodular lattice L in which principal ideals have only finite
maximal chains, properties (1) and (2) are obvious, while (3) is a consequence of
Theorem 9.4. The only (not very) hard part is to establish the inequality (4). Let x
and y be elements of L, and consider the join map νx : y/(x ∧ y) → (x ∨ y)/x defined
by νx (z) = z ∨ x. Recall that, by semimodularity, u  v implies u ∨ x  v ∨ x. Hence
νx takes maximal chains in y/(x ∧ y) to maximal chains in (x ∨ y)/x. So the length
of (x ∨ y)/x is at most that of y/(x ∧ y), i.e.,

δ(x ∨ y) − δ(x) ≤ δ(y) − δ(x ∧ y)

which establishes the desired inequality.

Conversely, suppose L is a lattice which admits a function δ satisfying (1)–(4).
Note that, by (2), δ(x) ≥ δ(z) + 2 whenever x > y > z; hence δ(x) = δ(z) + 1 implies
x  z.
94
 To establish semimodularity, assume a  a ∧ b in L. By (3) we have δ(a) =
δ(a ∧ b) + 1, and so by (4)

δ(a ∨ b) + δ(a ∧ b) ≤ δ(a) + δ(b)

= δ(a ∧ b) + 1 + δ(b)

whence δ(a ∨ b) ≤ δ(b) + 1. As a ∨ b > b, in fact δ(a ∨ b) = δ(b) + 1 and a ∨ b  b,

as desired.
For any a ∈ L, if a = ak > ak−1 > · · · > a0 is any chain in a/0, then δ(aj ) >
δ(aj−1 so k ≤ δ(a). Thus every chain in a/0 has length at most δ(a). 
For modular lattices, the map µx is an isomorphism, so we obtain instead equality.
It also turns out that we can dispense with the third condition, though this is not
very important.
Theorem 9.6. If L is a modular lattice and every principal ideal has only finite
maximal chains, then
(1) δ(0) = 0,
(2) x > y implies δ(x) > δ(y),
(3) δ(x ∨ y) + δ(x ∧ y) = δ(x) + δ(y).
Conversely, if L is a lattice which admits an integer-valued function δ satisfying
(1)–(3), then L is modular and principal ideals have only finite maximal chains.
At this point, it is perhaps useful to have some examples of semimodular lat-
tices. The lattice of equivalence relations Eq X is semimodular, but nonmodular
for |X| ≥ 4. The lattice in Figure 9.3 is semimodular, but not modular. We will see
more semimodular lattices as we go along, arising from group theory (subnormal
subgroups) in this chapter and from geometry in Chapter 11.

Figure 9.3

For our applications to group theory, we need a supplement to Theorem 9.4.

This in turn requires a definition. We say that a quotient a/b transposes up to
95
c/d if a ∨ d = c and a ∧ d = b. We then say that c/d transposes down to a/b.
We then define projectivity to be the smallest equivalence relation on the set of all
quotients of a lattice L which contains all transposed pairs hx/(x ∧ y), (x ∨ y)/yi.
Thus a/b is projective to c/d if and only if there exists a sequence of quotients
a/b = a0 /b0 , a1 /b1 , . . . , an /bn = c/d such that ai /bi and ai+1 /bi+1 are transposes
(up or down).
The strengthened version of Theorem 9.4, again due to Dedekind for modular
lattices, goes thusly.
Theorem 9.7. Let C and D be two maximal chains in a finite length semimodular
lattice, say

0 = c0 ≺ c1 ≺ · · · ≺ cn = 1
0 = d0 ≺ d1 ≺ · · · ≺ dn = 1.

Then there is a permutation π of the set {1, . . . , n} such that ci /ci−1 is projective
to dπ(i) /dπ(i)−1 for all i.
Proof. Again we use induction on n. We may assume c1 6= d1 , for otherwise the
result follows by induction. Then c1 /0 transposes up to (c1 ∨ d1 )/d1 , and d1 /0
transposes up to (c1 ∨ d1 )/c1 .
Let c1 ∨ d1 = e2 ≺ e3 ≺ · · · ≺ en = 1 be a maximal chain in 1/(c1 ∨ d1 ). By
induction, there is a permutation σ of {2, . . . , n} such that ci /ci−1 is projective to
eσ(i) /eσ(i)−1 if σ(i) 6= 2, and ci /ci−1 is projective to e2 /c1 = (c1 ∨ d1 )/c1 if σ(i) = 2.
Similarly, there is a permutation τ of {2, . . . , n} such that dj /dj−1 is projective
to eτ (j) /eτ (j)−1 if τ (j) 6= 2, and dj /dj−1 is projective to e2 /d1 = (c1 ∨ d1 )/d1 if
τ (j) = 2. Now just check that the permutation π of {1, . . . , n} given by
 −1
 τ σ(k) if k > 1 and σ(k) 6= 2

π(k) = 1 if k > 1 and σ(k) = 2

 −1
τ (2) if k = 1

has the property that ck /ck−1 is projective to dπ(k) /dπ(k)−1 .

This argument is illustrated in Figure 9.4. 
Theorems 9.4 and 9.7 are important in group theory. A chief series of a group G
is a maximal chain in the lattice of normal subgroups N (G). Since N (G) is modular,
our theorems apply.
Corollary. If a group G has a finite chief series of length k,

{1} = N0 < N1 < · · · < Nk = G

then every chief series of G has length k. Moreover, if

{1} = H0 < H1 < · · · < Hk = G

96
 E

C e2 D

c1 d1

Figure 9.4

is another chief series of G, then there is a permutation π of {1, . . . , k} such that

Hi /Hi−1 ∼
= Nπ(i) /Nπ(i)−1 for all i.
A subgroup H is subnormal in a group G, written H // G, if there is a chain in
Sub G,
H = H0 / H1 / . . . / Hk = G
with each Hi−1 normal in Hi (but not necessarily in G for i < k). Herman Wielandt
proved that the subnormal subgroups of a finite group form a lattice [10].
Theorem 9.8. If G is a finite group, then the subnormal subgroups of G form a
lower semimodular sublattice SN (G) of Sub G.
Proof. Let H and K be subnormal in G, with say

H = H0 / H1 / . . . / Hm = G
K = K0 / K1 / . . . / Kn = G.

Then H ∩ Ki / H ∩ Ki+1 , and so we have the series

H ∩ K / H ∩ K1 / H ∩ K2 / . . . H ∩ G = H / H1 / . . . / G.

Thus H ∩ K // G. Note that this argument shows that if H, K // G and K ≤ H,

then K // H.
The proof that SN (G) is closed under joins is a bit trickier. Let H, K / / G
as before. Without loss of generality, H and K are incomparable. By induction,
we may assume that |G| is minimal and that the result holds for larger subnormal
subgroups of G, i.e.,
(1) the join of subnormal subgroups is again subnormal in any group G 0 with
|G 0 | < |G|,
(2) if H < L // G, then L ∨ K // G; likewise, if K < M // G, then H ∨ M // G.
97
If there is a subnormal proper subgroup S of G which contains both H and K, then
H and K are subnormal subgroups of S (by the observation above). In that case,
H ∨ K // S by the first assumption, whence H ∨ K // G. Thus we may assume that
(3) no subnormal proper subgroup of G contains both H and K.
Combining this with assumption (2) yields
(4) H1 ∨ K = G = H ∨ K1 .
Finally, if both H and K are normal in G, then so is H ∨ K. Thus we may assume
(by symmetry) that
(5) H is not normal in G, and hence H < H1 ≤ Hm−1 < G.
Now G is generated by the set union H1 ∪ K (assumption (4)), so we must have
x−1 Hx 6= H for some x ∈ H1 ∪ K. But H / H1 , so k−1 Hk 6= H for some k ∈ K.
However, k−1 Hk is a subnormal subgroup of Hm−1 , because
k−1 Hk / k−1 H1 k / . . . / k−1 Hm−1 k = Hm−1
as Hm−1 /G. Thus, by assumption (1), H ∨k−1 Hk is a subnormal subgroup of Hm−1 ,
and hence of G. But H < H ∨ k−1 Hk ≤ H ∨ K, so (H ∨ k−1 Hk) ∨ K = H ∨ K.
Therefore H ∨ K is subnormal in G by assumption (2).
Finally, if H ∨ K  H in SN (G), then H / H ∨ K (using the observation after
the first argument), and (H ∨ K)/H is simple. By the First Isomorphism Theorem,
K/(H ∧K) is likewise simple, so K  H ∧K. Thus SN (G) is lower semimodular. 
A maximal chain in SN (G) is called a composition series for G. As SN (G) is
lower semimodular, the duals of Theorems 9.4 and 9.7 yield the following important
structural theorem for groups.
Corollary. If a finite group G has a composition series of length n,
{1} = H0 / H1 / . . . / Hn = G
then every composition series of G has length n. Moreover, if
{1} = K0 / K1 / . . . / Kn = G
is another composition series for G, then there is a permutation π of {1, . . . , n} such
that Ki /Ki−1 ∼
= Hπ(i) /Hπ(i)−1 for all i.
V
A finite decomposition of an element a ∈ L is an expression a = Q where
Q is a finite set of meet irreducible elements. If L satisfies the ACC, then every
element has a finite decomposition. We have seen that every element of a finite
distributive lattice has a unique irredundant decomposition. In a finite dimensional
modular lattice, an element can have many different finite decompositions, but the
number of elements in any irredundant decomposition is always the same. This
is a consequence of the following replacement property (known as the Kurosh-Ore
Theorem).
98
Theorem 9.9. If a is an element of a modular lattice and

a = q 1 ∧ . . . ∧ q m = r1 ∧ . . . ∧ r n

are two irredundant decompositions of a, then m = n and for each qi there is an rj

such that ^
a = rj ∧ qk
k6=i

is an irredundant decomposition.
V V
Proof. Let a = Q = R be two irredundant finite V decompositions (dropping
the subscripts temporarily). Fix q ∈ Q, and let q = (Q − {q}). By modularity,
q ∨ q/q ∼
= q/q ∧ q = q/a. Since q is meet V irreducible
V in L, this implies that a is meet
irreducible in q/a. However, a = q ∧ R = r∈R (q ∧ r) takes place in q/a, so we
must have a = q ∧ r for some r ∈ V R.
Next we observe
V that a = r ∧ (Q − {q}) is irredundant. For if not, we would
have a = r ∧ S irredundantly for some proper subset V S ⊂V Q − {q}. Reapplying
the first argument to V the two decompositions a = r ∧ S = Q with the element
r, we obtain a = q 0 ∧ S for some q 0 ∈ Q, contrary to the irredundance of Q.
It remains to show that |Q| = |R|. Let Q = V {q1 , . . . , qm } say.VBy the first part,
there is an element r1 ∈ R such that a = r1 ∧ (Q − {q1 }) = R irredundantly.
Applying the argument to these two V decompositions and V q2 , there is an element
r2 ∈ R such that a = r1 ∧ r2 ∧ (Q − {q1 , q2 }) = V R. Moreover, r1 and r2
are distinct, for otherwise we would
V have a = r 1 ∧ (Q − {q1 , q2 }), contradicting
the irredundance of a = r1 ∧ (Q − {q1 }). Continuing, we can replace q3 by an
element r3 of R, distinct from r1 and r2 , and so forth. After m steps, we obtain
a = r1 ∧ · · · ∧ rm , whence R = {r1 , . . . , rm }. Thus |Q| = |R|. 
With a bit of effort, this can be improved to a simultaneous exchange theorem.
V V
Theorem 9.10. If a is an element of a modular lattice and a = Q = R are
two irredundant finite decompositions of a, then for each q ∈ Q there is an r ∈ R
such that ^ ^
a = r ∧ (Q − {q}) = q ∧ (R − {r}).

The proof of this, and much more on the general theory of decompositions in
lattices, can be found in Crawley and Dilworth [2]; see also Dilworth [5].
Now Theorems 9.9 and 9.10 are exactly what we want in a finite dimensional mod-
ular lattice. However, in algebraic modular lattices, finite decompositions into meet
irreducible elements need not coincide with the (possibly infinite) decomposition
into completely meet irreducible elements given by Birkhoff’s Theorem. Consider,
for example, the chain C = (ω + 1)d , the dual of ω + 1. This satisfies the ACC, and
hence is algebraic. The least element of C is meet irreducible, but not completely
99
meet irreducible. In the direct product C n , the least element has a finite decom-
position into n meet irreducible elements, but every decomposition into completely
meet irreducibles is necessarily infinite.
Fortunately, the proof of Theorem 9.9 adapts nicely to give us a version suitable
for algebraic modular lattices.
V
Theorem 9.11. Let a be an element of a modular lattice. If a = Q is a finite, V
irredundant decomposition into completely meet irreducible elements, and a = R
is another decomposition into meet irreducible
V elements, then there exists a finite
subset R0 ⊆ R with |R0 | = |Q| such that a = R0 irredundantly.
The application of Theorem 9.11 to subdirect products is immediate.
Corollary. Let A be an algebra such that Con A is a modular lattice. If A has
a finite subdirect decomposition into subdirectly irreducible algebras, then every ir-
redundant subdirect decomposition of A into subdirectly irreducible algebras has the
same number of factors.
A more important application is to the theory of direct decompositions of con-
gruence modular algebras. (The corresponding congruences form a complemented
sublattice of Con A.) This subject is treated thoroughly in McKenzie, McNulty
and Taylor [9].
Let us close this section by mentioning a nice combinatorial result about finite
modular lattices, due to R. P. Dilworth [4].
Theorem 9.12. In a finite modular lattice L, let Jk (L) be the set of elements which
cover exactly k elements, and let Mk (L) be the set of elements which are covered by
exactly k elements. Then |Jk (L)| = |Mk (L)| for any integer k ≥ 0.
In particular, the number of join irreducible elements in a finite modular lattice
is equal to the number of meet irreducible elements.
We will return to modular lattices in Chapter 12.

Exercises for Chapter 9

1. (a) Prove that a lattice L is distributive if and only if it has the property that
a ∨ c = b ∨ c and a ∧ c = b ∧ c imply a = b.
(b) Show that L is modular if and only if, whenever a ≥ b and c ∈ L, a ∨ c = b ∨ c
and a ∧ c = b ∧ c imply a = b.
2. Show that every finite dimensional distributive lattice is finite.
3. Prove that if a finite dimensional lattice is both upper and lower semimodular,
then it is modular.
4. Prove that the following conditions are equivalent for a strongly atomic, alge-
braic lattice.
(i) L is semimodular: a  a ∧ b implies a ∨ b  b.
(ii) If a and b both cover a ∧ b, then a ∨ b covers both a and b.
100
 (iii) If b and c are incomparable and b ∧ c < a < c, then there exists x such that
b ∧ c < x ≤ b and a = c ∧ (a ∨ x).
5. (a) Find infinitely many simple modular lattices of width 4.
(b) Prove that the variety generated by all lattices of width ≤ 4 contains subdi-
rectly irreducible lattices of width ≤ 4 only.
6. Prove that every arguesian lattice is modular.
7. Let L be a lattice, and suppose there exist an ideal I and a filter F of L such
that L = I ∪ F and I ∩ F 6= ∅.
(a) Show that L is distributive if and only if both I and F are distributive.
(b) Show that L is modular if and only if both I and F are modular.
(R. P. Dilworth)
8. Show that modular lattices satisfy the equation
x ∧ (y ∨ (z ∧ (x ∨ t))) = x ∧ (z ∨ (y ∧ (x ∨ t))).
9. Let C and D be two chains in a modular lattice L. Prove that C ∪ D generates
a distributive sublattice of L. (Bjarni Jónsson)
10. Let a and b be two elements in a modular lattice L such that a ∧ b = 0. Prove
that the sublattice generated by a/0 ∪ b/0 is isomorphic to the direct product
a/0 × b/0.
11. Prove Theorem
Q 9.11. (Mimic the proof of Theorem 9.9.)
12. Let A = i∈ω Z2 be the direct product of countably many copies V of theVtwo
element group. Describe two decompositions of 0 in Sub A, say 0 = Q = R,
such that |Q| = ℵ0 and |R| = 2ℵ0 .

References
1. G. Birkhoff, Lattice Theory, First edition, Colloquium Publications, vol. 25, Amer. Math. Soc.,
Providence, R. I., 1940.
2. P. Crawley and R. P. Dilworth, Algebraic Theory of Lattices, Prentice-Hall, Englewood Cliffs,
N. J., 1973.
3. R. Dedekind, Über Zerlegungen von Zahlen durch ihre grössten gemeinsamen Teiler, Fest-
schrift der Herzogl. technische Hochschule zur Naturforscher-Versammlung, Braunschweig (1897),
Reprinted in “Gesammelte mathematische Werke”, Vol. 2, pp. 103–148, Chelsea, New York,
1968.
4. R. P. Dilworth, Proof of a conjecture on finite modular lattices, Ann. of Math. 60 (1954),
359–364.
5. R. P. Dilworth, Structure and Decomposition Theory, Proceedings of Symposia on Pure Math-
ematics: Lattice Theory, vol. II, Amer. Math. Soc., Providence, R. I., 1961.
6. G. Grätzer, Equational classes of lattices, Duke Jour. Math. 33 (1966), 613–622.
7. B. Jónsson, Equational classes of lattices, Math. Scand. 22 (1968), 187–196.
8. B. Jónsson and I. Rival, Lattice varieties covering the smallest non-modular variety, Pacific J.
Math. 82 (1979), 463–478.
9. R. McKenzie, G. McNulty and W. Taylor, Algebras, Lattices, Varieties, vol. I, Wadsworth and
Brooks-Cole, Belmont, CA, 1987.
10. H. Wielandt, Eine Verallgemeinerung der invarianten Untergruppen, Math. Zeit. 45 (1939),
209–244.
101
 10. Finite Lattices and their Congruence Lattices

If memories are all I sing

I’d rather drive a truck.
–Ricky Nelson

In this chapter we want to study the structure of finite lattices, and how it is
reflected in their congruence lattices. There are different ways of looking at lattices,
each with its own advantages. For the purpose of studying congruences, it is useful
to represent a finite lattice as the lattice of closed sets of a closure operator on its
set of join irreducible elements. This is an efficient way to encode the structure, and
will serve us well.1
The approach to congruences taken in this chapter is not the traditional one. It
evolved from techniques developed over a period of time by Ralph McKenzie, Bjarni
Jónsson, Alan Day, Ralph Freese and J. B. Nation for dealing with various specific
questions (see [1], [4], [6], [7], [8], [9]). Gradually, the general usefulness of these
methods dawned on us.
In the simplest case, recall that a finite distributive lattice L is isomorphic to
the lattice of order ideals O(J(L)), where J(L) is the ordered set of nonzero join
irreducible elements of L. This reflects the fact that join irreducible elements in a
distributive lattice are join prime. In a nondistributive lattice, we seek a modification
which will keep track of the ways in which one join irreducible is below the join of
others. In order to do this, we must first develop some terminology.
Rather than just considering finite lattices, we can include with modest addi-
tional effort a larger class of lattices satisfying a strong finiteness condition. Recall
that a lattice L is principally chain finite if no principal ideal of L contains an infi-
nite chain (equivalently, every principal ideal x/0 satisfies the ACC and DCC). In
Theorem 11.1, we will see where this class arises naturally in an important setting.2
Recall that if X, Y ⊆ L, we say that X refines Y (written X  Y ) if for each
x ∈ X there exists y ∈ Y with x ≤ y. It is easy to see that the relation  is a
quasiorder (reflexive and transitive), but not in general antisymmetric. Note X ⊆ Y
implies X  Y .
If q ∈ J(L) is completely join irreducible, let q∗ denote the unique element of
L with q  q∗ . Note that if L is principally chain finite, then q∗ exists for each
q ∈ J(L).
1 For an alternate approach, see Appendix. 3
2 Many of the results in this chapter can be generalized to arbitrary lattices. However, these
generalizations have not yet proved to be very useful unless one assumes at least the DCC.
101
 W
W expression of a ∈ L is a finite set B such that a = B. A join expression
A join
a = B is minimal ifWit is irredundant and B cannot be properly refined, i.e.,
B ⊆ J(L) and a > b∗ ∨W (B − {b}) for each bW∈ B. An equivalent way to write this
technically is that a = B minimally if a = C and C  W B implies B ⊆ C.
A join coverW of p ∈ L is a finite set A such that p ≤ A. A join cover A of p
is minimal if A isWirredundant and A cannot be properly refined to another join
cover of p, i.e., p ≤ B and B  A implies A ⊆ B.
We define a binary relation D on J(L) as follows: p D q if there exists x ∈ L
such that p ≤ q ∨ x but p 6≤ q∗ ∨ x. This relation will play an important role in our
analysis of the congruences of a principally chain finite lattice.3
The following lemma summarizes some properties of principally chain finite lat-
tices and the relation D.
Lemma 10.1. Let L be a principally chain finite lattice.
(1) If b 6≤ a in L, then there exists p ∈ J(L) with p ≤ b and p 6≤ a.
(2) Every join expression in L refines to a minimal join expression, and every
join cover refines to a minimal join cover.
(3) For p, q ∈ J(L) we have p D q if and only if q ∈ A for some minimal join
cover A of p.

Proof. (1) Since b 6≤ a and b/0 satisfies the DCC, the set {x ∈ b/0 : x 6≤ a} has at
least
W one minimal element p. Because y < p implies y ≤ a for any y ∈ L, we have
{y ∈ L : y < p} ≤ p ∧ a < p, and hence p ∈ J(L) with p∗ = p ∧ a. W
(2) Suppose L contains an element s with a join representation s = F which
does not refine to a minimal one. Since the DCC holds in s/0, there W is an element
t ≤ s minimal with respect to having a join representation t = A which fails to
refine to a minimal one.
W Clearly t is join reducible, and there is a proper, irredundant
join expression t = B with B  A.
Let B = {b1 , . . . , bk }. Using the DCC on b1 /0, we can find c1 ≤ b1 such that
t = c1 ∨b2 ∨. . .∨bk , but c1 cannot be replaced by any lower element: t > u∨b2 ∨. . . bk
whenever u < c1 . Now apply the same argument to b2 and {c1 , b2 , . . . , bk }. After k
such steps we obtain a join cover C which refines B and is minimal pointwise: no
element can be replaced by a (single) lower element.
The elements of C may not be join irreducible, but each element of C is strictly
below t, and hence has a minimal join expression. ChooseSa minimal join expression
Ec for each c ∈ C. It is not hard to check that E = c∈C Ec is a minimal join
expression for t, and E  C  B  A, which contradicts the choice W of t and B.
Now let u ∈ L and W let A be a join cover of u, i.e., u ≤ A. We can find
B ⊆ A such that u ≤ B irredundantly. As above, refine B to a pointwise minimal

3 Note that D is reflexive, i.e., p D p for all p ∈ J(L). The relation D, defined similarly
except that it requires p 6= q, is also important, and D stands for “D or equal to.” For describing
congruences, it makes more sense to use D rather than D.
102
join cover
S C. Now we know that minimal join expressions exist, so we may define
E = c∈C Ec exactly as before. Then E will be a minimal join cover of u, and again
E  C  B  A.
(3) Assume p D q, and let x ∈ L be such that p ≤ q ∨ x but p 6≤ q∗ ∨ x. By (2),
we can find a minimal join cover A of p with A  {q, x}. Since p 6≤ q∗ ∨ x, we must
have q ∈ A.
Conversely, if A is a minimal Wjoin cover of p, and q ∈ A, then we fulfill the
definition of p D q by setting x = (A − {q}). 

Now we want to define a closure operator on the join irreducible elements of a

principally chain finite lattice. This closure operator should encode the structure of
L in the same way the order ideal operator O does for a finite distributive lattice.
For S ⊆ J(L), let
_
Γ(S) = {p ∈ J(L) : p ≤ F for some finite F ⊆ S}.

It is easy to check that Γ is an algebraic closure operator. The compact W (i.e.,

finitely generated) Γ-closed sets are of the form Γ(F ) = {p ∈ J(L) : p ≤ F } for
some finite subset F of J(L). In general, we would expect these to be only a join
subsemilattice of the lattice CΓ of closed sets; however, for a principally chain finite
lattice L the compact closed sets actually formW an ideal (and hence a sublattice) of
CWΓ . For if S ⊆ Γ(F ) with F finite, then S ⊆ F/0, whichWsatisfies
W the ACC. Hence
{ G : G ⊆ S and G is finite} has a largest element. So S = G for some finite
G ⊆ S, from which it follows that Γ(S) = Γ(G), and Γ(S) is compact. In particular,
if L has a largest element 1, then every closed set will be compact.
With that preliminary observation out of the way, we proceed with our general-
ization of the order ideal representation for finite distributive lattices.

Theorem 10.2. If L is a principally chain finite lattice, then the map φ with φ(x) =
{p ∈ J(L) : p ≤ x} is an isomorphism of L onto the lattice of compact Γ-closed
subsets of J(L).
W
Proof. Note that if x = A is a minimal join expression, then φ(x) = Γ(A), so φ(x)
is indeed a compact Γ-closed set. The map φ is clearly order preserving, andWit is
one-to-one by part (1) of Lemma 10.1. Finally, φ is onto because Γ(F ) = φ( F )
for each finite F ⊆ J(L). 

To use this result, we need a structural characterization of Γ-closed sets.

Theorem 10.3. Let L be a principally chain finite lattice. A subset C of J(L) is

Γ-closed if and only if
(1) C is an order ideal of J(L), and
(2) if A is a minimal join cover of p ∈ J(L) and A ⊆ C, then p ∈ C.
103
Proof. It is easy to see that Γ-closed sets have these properties. Conversely, let
C ⊆WJ(L) satisfy (1) and (2). We want to show Γ(C) ⊆ C. If p ∈ Γ(C), then
p ≤ F for some finite subset F ⊆ C. By Lemma 10.1(2), there is a minimal join
cover A of p refining F ; since C is an order ideal, A ⊆ C. But then the second
closure property gives that p ∈ C, as desired. 
In words, Theorem 10.3 says that for principally chain finite lattices, Γ is deter-
mined by the order on J(L) and the minimal join covers of elements of J(L). Hence,
by Theorem 10.2, L is determined by the same factors. Now we would like to see
how much of this information we can extract from Con L. The answer is, “not
much.” We will see that from Con L we can find J(L) modulo a certain equivalence
relation. We can determine nothing of the order on J(L), nor can we recover the
minimal join covers, but we can recover the D relation (up to the equivalence). This
turns out to be enough to characterize the congruence lattices of principally chain
finite lattices.
Now for a group G, the map τ : Con G → N (G) given by τ (θ) = {x ∈ G : x θ 1}
is a lattice isomorphism. The next two theorems and corollary establish a similar
correspondence for principally chain finite lattices.
Theorem 10.4. Let L be a principally chain finite lattice. Let σ map Con L to
the lattice of subsets P(J(L)) by
σ(θ) = {p ∈ J(L) : p θ p∗ }.
Then σ is a one-to-one complete lattice homomorphism.
Proof. Clearly σ is order preserving: θ ≤ ψ implies σ(θ) ⊆ σ(ψ).
To see that σ is one-to-one, assume θ 6≤ ψ. Then there exists a pair of elements
a, b ∈ L with a < b and (a, b) ∈ θ − ψ. Since (a, b) 6∈ ψ, we also have (x, b) 6∈ ψ
for any element x with x ≤ a. Let p ≤ b be minimal with respect to the property
p ψ x implies x 6≤ a. We claim that p is join irreducible. W If y1W , . . . , yn < p,Wthen for
each i there exists an xi such that yi ψ xi ≤ a. Hence yi ψ xi ≤ a, so yi < p.
Now p = p ∧ b θ p ∧ a ≤ p∗ , implying p θ p∗ , i.e., p ∈ σ(θ). But (p, p∗ ) 6∈ ψ because
p∗ ψ x ≤ a for some x; thus V p 6∈ σ(ψ).
T Therefore σ(θ) 6⊆ σ(ψ).
It is easy to see that σ( θi ) = σ(θS i ) for any collection
W of congruences θi (i ∈ I).
Since
W σ is order
S preserving, we have σ(θi ) ⊆ σ( θi ), and it remains to show that
σ( θi ) ⊆ σ(θ W ).
i
If (p, p∗ ) ∈ θi , then there exists a connecting sequence
p = x0 θi1 x1 θi2 x2 . . . xk−1 θik xk = p∗ .
Let yj = (xj ∨p∗ )∧p. Then y0 = p, yk = p∗ , and p∗ ≤ yj ≤ p implies yj ∈ {p∗ , p} for
each j. Moreover, we have yj−1 θij yj for j ≥ 1. There
S must exist a j with yj−1 = p
and yj = p∗ , whence p θij p∗ and p ∈ σ(θij ) ⊆ σ(θi ). We conclude that σ also
preserves arbitrary joins. 
Next we need to identify the range of σ.
104
Theorem 10.5. Let L be a principally chain finite lattice, and let S ⊆ J(L). Then
S = σ(θ) for some θ ∈ Con L if and only if p D q ∈ S implies p ∈ S.
Proof. Let S = σ(θ). If q ∈ S and p D q, then q θ q∗ , and for some x ∈ L we have
p ≤ q ∨ x but p 6≤ q∗ ∨ x. Thus
p = p ∧ (q ∨ x) θ p ∧ (q∗ ∨ x) < p.
Hence p θ p∗ and p ∈ σ(θ) = S.
Conversely, assume we are given S ⊆ J(L) satisfying the condition of the theorem.
Then we must produce a congruence relation θ such that σ(θ) = S. Let T =
J(L) − S, and note that T has the property that q ∈ T whenever p D q and p ∈ T .
Define
x θ y if x/0 ∩ T = y/0 ∩ T.
The motivation for this definition is outlined in the exercises: θ is the kernel of
the standard homomorphism from L onto the join subsemilattice of L generated by
T ∪ {0}.
Three things should be clear: θ is an equivalence relation; x θ y implies x∧z θ y∧z;
and for p ∈ J(L), p θ p∗ if and only if p 6∈ T , i.e., p ∈ S. (The last statement will
imply that σ(θ) = S.) It remains to show that θ respects joins.
Assume x θ y, and let z ∈ L. We want to show (x∨ z)/0∩ T ⊆ (y ∨ z)/0∩ T , so let
p ∈ T and p ≤ x ∨ z. Then there exists a minimal join cover Q of p with Q  {x, z}.
If q ∈ Q and q ≤ z, then of course q ≤ y ∨ z. Otherwise q ≤ x, and since p ∈ T
and p D q (by Lemma 10.1(3)), W we have q ∈ T . Thus q ∈ x/0 ∩ T = y/0 ∩ T , so
q ≤ y ≤ y∨z. It follows that p ≤ Q ≤ y∨z. This shows (x∨z)/0∩T ⊆ (y∨z)/0∩T ;
by symmetry, they are equal. Hence x ∨ z θ y ∨ z. 
In order to interpret the consequences of these two theorems, let E denote the
transitive closure of D on J(L). Then E is a quasiorder (reflexive and transitive),
and so it induces an equivalence relation ≡ on J(L), modulo which E is a partial
order, viz., p ≡ q if and only if p E q and q E p. If we let QL denote the partially
ordered set (J(L)/ ≡, E), then Theorem 10.5 translates as follows.
Corollary. If L is a principally chain finite lattice, then Con L ∼= O(QL ).
Because the D relation is easy to determine, it is not hard to find QL for a finite
lattice L. Hence this result provides a reasonably efficient algorithm for determining
the congruence lattice of a finite lattice. Hopefully, the exercises will convince you
of this. As an application, we have the following observation.
Corollary. A principally chain finite lattice L is subdirectly irreducible if and only
if QL has a least element.
Now let us turn our attention to the problem of representing a given distributive
algebraic lattice D as the congruence lattice of a lattice.4 Not every distributive
4 Recall from Chapter 5 that it is an open problem whether every distributive algebraic lattice

is isomorphic to the congruence lattice of a lattice.

105
algebraic lattice is isomorphic to O(P) for an ordered set P. Indeed, those which
are have a nice characterization.

Lemma 10.6. The following are equivalent for a distributive algebraic lattice D.
(1) D is isomorphic to the lattice of order ideals of an ordered set.
(2) Every element of D is a join of completely join prime elements.
(3) Every compact element of D is a join of (finitely many) join irreducible
compact elements.

Proof. An order ideal I is compact in O(P) if and only if it is finitely generated,

i.e., I = p1 /0 ∪ · · · ∪ pk /0 for some p1 , . . . , pk ∈ P . Moreover, each pi /0 is join
irreducible in O(P). Thus O(P) has the property (3).
Note that if D is a distributive algebraic lattice and W p is a join irreducible
W 0 compact
element, then p is completely join prime. For if p ≤ U , then p ≤ U for some
finite subset U 0 ⊆ U ; as join irreducible elements are join prime in a distributive
lattice, this implies p ≤ u for some u ∈ U 0 . On the other hand, a completely join
prime element is clearly compact and join irreducible, so these elements coincide.
If every compact element is a join of join irreducible compact elements, then so is
every element of D, whence (3) implies (2).
Now assume that the completely join prime elements of D are join dense, and
let P denote the set of completely join prime elements with the order they inherit
from D. Then it is straightforward to show that the map φ : D → O(P) given by
φ(x) = x/0 ∩ P is an isomorphism. 

Now it is not hard to find lattices where these conditions fail. Nonetheless,
distributive algebraic lattices with the properties of Lemma 10.6 are a nice class
(including all finite distributive lattices), and it behooves us to try to represent each
of them as Con L for some principally chain finite lattice L. We need to begin by
seeing how QL can be recovered from Con L.

Theorem 10.7. Let L be a principally chain finite lattice. A congruence relation

θ is join irreducible and compact in Con L if and only if θ = con(p, p∗ ) for some
p ∈ J. Moreover, for p, q ∈ J, we have con(q, q∗ ) ≤ con(p, p∗ ) iff q E p.

Proof. We want to use the representation Con L ∼ = O(QSL ). Note that if Q is a

partially ordered set and I is an order ideal of Q, then I = x∈I x/0, and, of course,
set union is the join operation in O(Q). Hence join irreducible compact ideals are
exactly those of the form x/0 for some x ∈ Q.
Applying these remarks to our situation, using the isomorphism, join irreducible
compact congruences are precisely those with σ(θ) = {q ∈ J(L) : q E p} for some
p ∈ J(L). Recalling that p ∈ σ(θ) if and only if p θ p∗ , and con(p, p∗ ) is the least
congruence with p θ p∗ , the conclusions of the theorem follow. 

106
Theorem 10.8. Let D be a distributive algebraic lattice which is isomorphic to
O(P) for some ordered set P. Then there is a principally chain finite lattice L such
that D ∼
= Con L.
Proof. We must construct L with QL ∼ = P. In view of Theorem 10.3 we should try
to describe L as the lattice of finitely generated closed sets of a closure operator on
an ordered set J. Let P 0 and P 1 be two unordered copies of the base set P of P,
disjoint except on the maximal elements of P. Thus J = P 0 ∪ P 1 is an antichain,
and p0 = p1 if and only if p is maximal in P. Define a subset C of J to be closed if
{pj , q k } ⊆ C implies pi ∈ C whenever p < q in P and {i, j} = {0, 1}. Our lattice L
will consist of all finite closed subsets of J, ordered by set inclusion.
It should be clear that we have made the elements of J atoms of L and

p i ≤ pj ∨ q k

whenever p < q in P. Thus pi D q k iff p ≤ q. (This is where you want only one copy
of each maximal element). It remains to check that L is indeed a principally chain
finite lattice with QL ∼ = P, as desired. The crucial observation is that the closure of
a finite set is finite. We will leave this verification to the reader. 
Theorem 10.8 is due to R. P. Dilworth in the 1940’s, but his proof was never
published. The construction given is from George Grätzer and E. T. Schmidt [5].
We close this section with a new look at a pair of classic results. A lattice is said
to be relatively complemented if a < x < b implies there exists y such that x ∧ y = a
and x ∨ y = b.5
Theorem 10.9. If L is a principally chain finite lattice which is either modular
or relatively complemented, then the relation D is symmetric on J(L), and hence
Con L is a Boolean algebra.
Proof. First assume L is modular, and let p D q with p ≤ q ∨ x but p 6≤ q∗ ∨ x.
Using modularity, we have

(q ∧ (p ∨ x)) ∨ x = (q ∨ x) ∧ (p ∨ x) ≥ p,

so q ≤ p ∨ x. On the other hand, if q ≤ p∗ ∨ x, we would have

p = p ∧ (q ∨ x) ≤ p ∧ (p∗ ∨ x) = p∗ ∨ (x ∧ p) = p∗ ,

a contradiction. Hence q 6≤ p∗ ∨ x, and q D p.

Now assume L is relatively complemented and p D q as above. Observe that
a join irreducible element in a relatively complemented lattice must be an atom.
5 Thus a relatively complemented lattice with 0 and 1 is complemented, but otherwise it need

not be.
107
Hence p∗ = q∗ = 0, and given x such that p ≤ q ∨ x, p 6≤ x, we want to find y
such that q ≤ p ∨ y, q 6≤ y. Using the ACC in q ∨ x/0, let y be maximal such that
x ≤ y < q ∨ x and p 6≤ y. If y < p ∨ y < q ∨ x, then the relative complement z of
p ∨ y in q ∨ x/y satisfies z > y and z 6≥ p, contrary to the maximality of y. Hence
p ∨ y = q ∨ x, i.e., q ≤ p ∨ y. Thus q D p.
Finally, if D is symmetric, then QL is an antichain, and thus O(QL ) is isomorphic
to the Boolean algebra P(QL ). 
A lattice is simple if |L| > 1 and L has no proper nontrivial congruence relations,
i.e., Con L ∼ = 2. Theorem 10.9 says that a subdirectly irreducible, modular or
relatively complemented, principally chain finite lattice must be simple.
In the relatively complemented case we P get even more. Let Li (i ∈ I) be a collec-
tion of lattices with 0. The direct sum Li is the sublattice of the direct product
consisting of all elements which are only finitely non-zero. Combining Theorems
10.2 and 10.9, we obtain relatively easily a fine result of Dilworth [2].
Theorem 10.10. A relatively complemented principally chain finite lattice is a
direct sum of simple (relatively complemented principally chain finite) lattices.
Proof. Let L be a relatively complemented principally chain finite lattice. Then
every element of L is a finite join of join irreducible elements, every join irreducible
element is an atom, and the D relation is symmetric, i.e., p D q implies p ≡ q. We
Ṡ
can write J(L) as a disjoint union of ≡-classes, J(L) = i∈I Ai . Let
_
Li = {x ∈ L : x = F for some finite F ⊆ Ai }.

P to show that the Li ’s are ideals (and hence sublattices) of L, and that
We want
L∼= i∈I Li . W
The crucial technical detail is this: if p ∈ J(L), F ⊆ J(L) is finite, and p ≤ F ,
then p ≡ f for some f ∈ F . For F can be refined to a minimal join cover G of p,
and since join irreducible elements are atoms, we must have G ⊆ F . But p D g (and
hence p ≡ g) for each g ∈ G.
NowW we can show that each Li is W an ideal of L. Suppose y ≤ x ∈ Li . Then
x = F for some F ⊆ Ai , and y = H for some H ⊆ J(L). By the preceding
observation, H ⊆ Ai , and thus
P y ∈ Li . W
Define a map φ : L → i∈I Li by φ(x) = (xi )i∈I , where xi = (x/0 ∩ Ai ).
There are several things to check: that φ(x) is only finitely nonzero, that φ is one-
to-one and onto, and that it preserves meets and joins. None is very hard, so we
will only do the last one, and leave the rest to the reader.
We want to show that φ preserves joins, i.e., that (x ∨ y)i = xi ∨ yi . It suffices
to show that if p ∈ J(L) and p ≤ (x ∨ y)i , then p ≤ xi ∨ yi . Since Li is an ideal,
we have p ∈ Ai . Furthermore, since p ≤ x ∨ y, there is a minimal join cover F of p
refining {x, y}. For each f ∈ F , weWhave f ≤ x or f ≤ y, and p D f implies f ∈ Ai ;
hence f ≤ xi or f ≤ yi . Thus p ≤ F ≤ xi ∨ yi . 
108
 Exercises for Chapter 10
1. Do Exercise 1 of Chapter 5 using the methods of this chapter.
2. Use the construction from the proof of Theorem 10.8 to represent the distribu-
tive lattices in Figure 10.1 as congruence lattices of lattices.

(a) (b) (c)

Figure 10.1
W
3. Let a = B be a join expression in a lattice L. Prove that the following two
properties (used to define minimality) really are equivalent.
W
(a) B ⊆W J(L) and a > b∗ ∨ (B − {b}) for each b ∈ B.
(b) a = C and C  B implies B ⊆ C.
4. Let P be an ordered set satisfying the DCC, and let Q be the set of finite
antichains of P, ordered by . Show that Q satisfies the DCC. (This argument is
rather tricky, but it is the proper explanation of Lemma 10.1(2).)
5. Let p be a join irreducible element in a principally chain finite lattice. Show
that p is join prime if and only if p D q implies p = q.
6. Let L be a principally chain finite lattice, and p ∈ J(L). Prove that there is a
congruence ψp on L such that, for all θ ∈ Con L, (p, p∗ ) 6∈ θ if and only if θ ≤ ψp .
(More generally, the following is true: Given a lattice L and a filter F of L, there
is a unique congruence ψF maximal with respect to the property that (x, f ) ∈ θ
implies x ∈ F for all x ∈ L and f ∈ F .)
7. Prove that a distributive lattice is isomorphic to O(P) for some ordered set P
if and only if it is algebraic and dually algebraic. (This extends Lemma 10.6.)
8. Let L be a principally chain finite lattice, and let T ⊆ J(L) have the property
that p D q and p ∈ T implies q ∈ T .
(a) Show Wthat the join subsemilattice S of L generated by T ∪ {0}, i.e., the set
of all F where F is a finite subset of T ∪ {0}, is a lattice. (S need not a
be sublattice of L, because the meet operation is different.)
109
 W
(b) Prove that the map f : L  S given by f (x) = (x/0 ∩ T ) is a lattice
homomorphism.
(c) Show that the kernel of f is the congruence relation θ in the proof of Theorem
10.5.
9. Prove that if L is a finite lattice, then L can be embedded into a finite lattice
K such that Con L ∼ = Con K and every element of K is a join of atoms. (Michael
Tischendorf)
10. Express the lattice of all finite subsets of a set X as a direct sum of two-
element lattices.
11. Show that if A is a torsion abelian group, then the compact subgroups of A
form a principally chain finite lattice (Khalib Benabdallah).

The main arguments in this chapter originated in a slightly different setting,

geared towards application to lattice varieties [7], the structure of finitely generated
free lattices [4], or finitely presented lattices [3]. The last three exercises give the
version of these results which has proved most useful for these types of applications,
with an example.
A lattice homomorphism f : L → K is lower bounded if for every a ∈ K, the set
{x ∈ L : f (x) ≥ a} is either empty or has a least element, which is denoted β(a). If
f is onto, this is equivalent to saying that each congruence class of ker f has a least
element. For example, if L satisfies the DCC, then every homomorphism f : L → K
will be lower bounded. The dual condition is called upper bounded. These notions
were introduced by Ralph McKenzie in [7].
12. Let L be a lattice with 0, K a finite lattice, and f : L  K a lower bounded,
surjective homomorphism. Let T = {β(p) : p ∈ J(K)}. Show that:
(a) T ⊆ J(L);
(b) K is isomorphic to the join subsemilattice S of L generated by T ∪ {0};
(c) for each t ∈ T , every join cover of t in L refines to a join cover of t contained
in T .
13. Conversely, let L be a lattice with 0, and let T be a finite subset of J(L)
satisfying condition (c) of Exercise 12. Let S denote the join subsemilattice
W of L
generated by T ∪ {0}. Prove that the map f : L  S given by f (x) = (x/0 ∩ T )
is a lower bounded lattice homomorphism with βf (t) = t for all t ∈ T .
14. Let f be the (essentially unique) homomorphism from F L(3) onto N5 . Show
that f is lower bounded. (By duality, f is also upper bounded.)

References
1. A. Day, Characterizations of finite lattices that are bounded-homomorphic images or sublat-
tices of free lattices, Canad. J. Math 31 (1979), 69–78.
2. R. P. Dilworth, The structure of relatively complemented lattices, Ann. of Math. 51 (1950),
348–359.
110
3. R. Freese, Finitely presented lattices: canonical forms and the covering relation, Trans. Amer.
Math. Soc. 312 (1989), 841–860.
4. R. Freese and J. B. Nation, Covers in free lattices, Trans. Amer. Math. Soc. 288 (1985), 1–42.
5. G. Grätzer and E. T. Schmidt, On congruence lattices of lattices, Acta Math. Acad. Sci.
Hungar. 13 (1962), 179–185.
6. B. Jónsson and J. B. Nation, A report on sublattices of a free lattice, Coll. Math. Soc. János
Bolyai 17 (1977), 233–257.
7. R. McKenzie, Equational bases and non-modular lattice varieties, Trans. Amer. Math. Soc.
174 (1972), 1–43.
8. J. B. Nation, Lattice varieties covering V (L1 ), Algebra Universalis 23 (1986), 132–166.
9. J. B. Nation, An approach to lattice varieties of finite height, Algebra Universalis 27 (1990),
521–543.

111
 11. Geometric Lattices

Many’s the time I’ve been mistaken

And many times confused . . . .
–Paul Simon

Now let us consider how we might use lattices to describe elementary geometry.
There are two basic aspects of geometry: incidence, involving such statements as
“the point p lies on the line l,” and measurement, involving such concepts as angles
and length. We will restrict our attention to incidence, which is most naturally
stated in terms of lattices.
What properties should a geometry have? Without being too formal, surely we
would want to include the following.
(1) The elements of a geometry (points, lines, planes, etc.) are subsets of a given
set P of points.
(2) The set P of all points is an element of the geometry, and the intersection
of any collection of elements is again one.
(3) There is a dimension function on the elements of the geometry, satisfying
some sort of reasonable conditions.
If we order the elements of a geometry by set inclusion, then we obtain a lattice in
which the atoms correspond to points of the geometry, every element is a join of
atoms, and there is a well-behaved dimension function defined. With a little more
care we can show that “well-behaved” means “semimodular” (recall Theorem 9.6).
On the other hand, there is no harm if we allow some elements to have infinite
dimension.
Accordingly, we define a geometric lattice to be an algebraic semimodular lattice
in which every element is a join of atoms. As we have already described, the points,
lines, planes, etc. (and the empty set) of a finite dimensional Euclidean geometry
(<n ) form a geometric lattice. Other examples are the lattice of all subspaces of
a vector space, and the lattice Eq X of equivalence relations on a set X. More
examples are included in the exercises.1
We should note here that the geometric dimension of an element is generally
one less than the lattice dimension δ: points are elements with δ(p) = 1, lines are
1 The basic properties of geometric lattices were developed by Garrett Birkhoff in the 1930’s [2].

Similar ideas were pursued by K. Menger, F. Alt and O. Schreiber at about the same time [10].
Traditionally, geometric lattices were required to be finite dimensional, meaning δ(1) = n < ∞.
The last two examples show that this restriction is artificial.
112
elements with δ(l) = 2, and so forth.
A lattice is said to be atomistic if every element is a join of atoms.
Theorem 11.1. The following are equivalent.
(1) L is a geometric lattice.
(2) L is an upper continuous, atomistic, semimodular lattice.
(3) L is isomorphic to the lattice of ideals of an atomistic, semimodular, prin-
cipally chain finite lattice.

Proof. Every algebraic lattice is upper continuous, so (1) implies (2).

For (2) implies (3), we first note
W that the atoms of an upper continuous lattice
are compact.WFor if Wa  0Wand a  F for every W finite F ⊆ U , then by Theorem 3.7
we have a ∧ U = (a ∧ F ) = 0, whence a  U . Thus in a lattice L satisfying
condition (2), the compact elements are precisely the elements which are the join of
finitely many atoms, in other words (using semimodularity) the finite dimensional
elements. Let K denote the ideal of all finite dimensional elements of L. Then K is
a semimodular principally chain finite sublattice of L, and it is not hard to see that
the map φ : L → I(K) by φ(x) = x/0 ∩ K is an isomorphism.
Finally, we need to show that if K is a semimodular principally chain finite lattice
with every element the join of atoms, then I(K) is a geometric lattice. Clearly I(K)
is algebraic, and every ideal is the join of the elements, and hence the atoms, it
contains. It remains to show that I(K) is semimodular.
Suppose I  I ∩ J in I(K). Fix an atom a ∈ I − J. Then I = (I ∩ J) ∨ a/0,
and hence I ∨ J = a/0 ∨ J. Let x be any element in (I ∨ J) − J. Since x ∈ I ∨ J,
there exists j ∈ J such that x ≤ a ∨ j. Because K is semimodular, a ∨ j  j. On the
other hand, every element of K is a join of finitely many atoms, so x 6∈ J implies
there exists an atom b ≤ x with b 6∈ J. Now b ≤ a ∨ j and b 6≤ j, so b ∨ j = a ∨ j,
whence a ≤ b ∨ j. Thus b/0 ∨ J = I ∨ J; a fortiori it follows that x/0 ∨ J = I ∨ J.
As this holds for every x ∈ (I ∨ J) − J, we have I ∨ J  J, as desired. 
At the heart of the preceding proof is the following little argument: if L is
semimodular, a and b are atoms of L, t ∈ L, and b ≤ a ∨ t but b 6≤ t, then a ≤ b ∨ t.
It is useful to interpret this property in terms of closure operators.
A closure operator Γ has the exchange property if y ∈ Γ(B ∪ {x}) and y 6∈ Γ(B)
implies x ∈ Γ(B ∪ {y}). Examples of algebraic closure operators with the exchange
property include the span of a set of vectors in a vector space, the geometric closure
of a set of points in Euclidean space, and the convex closure of a set of points in
Euclidean space. More generally, we have the following representation theorem for
geometric lattices, due to Saunders Mac Lane [9].
Theorem 11.2. A lattice L is geometric if and only if L is isomorphic to the lattice
of closed sets of an algebraic closure operator with the exchange property.
Proof. Given a geometric lattice L, we can define a closure operator Γ on the set A
113
of atoms of L by
Γ(X) = {a ∈ A : a ≤ ∨X}.
Since the atoms are compact, this is an algebraic closure operator. By the little
argument above, Γ has the exchange property. Because every element is a join of
atoms, the map φ : L → CΓ given by φ(x) = {a ∈ A : a ≤ x} is an isomorphism.
Now assume we have an algebraic closure operator Γ with the exchange property.
Then CΓ is an algebraic lattice. The exchange property insures that the closure of
a singleton, Γ(x), is an atom of CΓ : if y W
∈ Γ(x), then x ∈ Γ(y), so Γ(x) = Γ(y).
Clearly, for every closed set we have B = b∈B Γ(B). It remains to show that CΓ is
semimodular.
Let B and C be closed sets with B  B ∩ C. Then B = Γ({x} ∪ (B ∩ C)) for any
x ∈ B − (B ∩ C). Suppose C < D ≤ B ∨ C = Γ(B ∪ C), and let y be any element
in D − C. Fix any element x ∈ B − (B ∩ C). Then y ∈ Γ(C ∪ {x}) = B ∨ C, and
y 6∈ Γ(C) = C. Hence x ∈ Γ(C ∪ {y}), and B ≤ Γ(C ∪ {y}) ≤ D. Thus D = B ∨ C,
and we conclude that CΓ is semimodular. 
Now we turn our attention to the structure of geometric lattices.
Theorem 11.3. Every geometric lattice is relatively complemented.
Proof. Let a < x < b in a geometric lattice. By upper continuity and Zorn’s Lemma,
there exists an element y maximal with respect to the properties a ≤ y ≤ b and
x ∧ y = a. Suppose x ∨ y < b. Then there is an atom p with p ≤ b and p 6≤ x ∨ y.
By the maximality of y we have x ∧ (y ∨ p) > a; hence there is an atom q with
q ≤ x ∧ (y ∨ p) and q 6≤ a. Now q ≤ y ∨ p but q 6≤ y, so by our usual argument
p ≤ q ∨ y ≤ x ∨ y, a contradiction. Thus x ∨ y = b, and y is a relative complement
of x in b/a. 
Let L be a geometric lattice, and let K be the ideal of compact elements of L. By
Theorem 10.10, K is a direct sum of simple lattices, and by Theorem 11.1, L ∼ = I(K).
So what we need now is a relation between the ideal lattice of a direct sum and the
direct product of the corresponding ideal lattices.
P Q
Lemma 11.4. For any collection of lattices Li (i ∈ I), we have I( Li ) ∼ = I(Li ).
P
Proof. If we identify Li with the set of all vectors in P Li which Q are zero except
in the i-th place, then there is a natural map φ : I( Li ) → I(Li ) given by
φ(J) = hJi ii∈I , where Ji = {x ∈ Li : x ∈ J}. It will be a relatively straightforward
argument to show that this is an isomorphism. Clearly Ji ∈ I(Li ), and the map φ
is order preserving. P
Assume J, K ∈ I( Li ) with J 6≤ K, and let x ∈ J − K. There exists an i0 such
that xi0 6∈ K, and hence Ji0 6≤ Ki0 , whence φ(J) 6≤ φ(K). Thus φ(J) ≤ φ(K) if and
only if J ≤ K, and φ is one-to-one. Q P
It remains to show that φ is P onto. Given hTi ii∈I ∈ I(Li ), let J = {x ∈ Li :
xi ∈ Ti for all i}. Then J ∈ I( Li ), and it is not hard to see that Ji = Ti for all
i, and hence φ(J) = hTi ii∈I , as desired. 
114
 So we are left with the task of describing the lattice of ideals of a simple semi-
modular principally chain finite lattice in which every element is a join of atoms. If
L = I(K) where K is such a lattice, then L is subdirectly irreducible: the unique
minimal congruence µ is generated by collapsing all the finite dimensional elements
to zero. So if K is finite dimensional (whence L ∼ = K), then L is simple, and it may
be otherwise, as is the case with Eq X. On the other hand, if K is modular and
infinite dimensional, then µ will identify only those pairs (a, b) such that a∨b/a∧b is
finite dimensional, and so L will not be simple. Summarizing, we have the following
result.
Theorem 11.5. Every geometric lattice is a direct product of subdirectly irreducible
geometric lattices. Every finite dimensional geometric lattice is a direct product of
simple geometric lattices.
The finite dimensional case of Theorem 11.5 should be credited to Dilworth [3],
and the extension is due to J. Hashimoto [6]. The best version of Hashimoto’s
theorem states that a complete, weakly atomic, relatively complemented lattice is a
direct product of subdirectly irreducible lattices. A nice variation, due to L. Libkin [8],
is that every atomistic algebraic lattice is a direct product of directly indecomposable
(atomistic algebraic) lattices.
Before going on to modular geometric lattices, we should mention one of the most
intriguing problems in combinatorial lattice theory. Let L be a finite geometric
lattice, and let
wk = |{x ∈ L : δ(x) = k}|.
The unimodal conjecture states that there is always an integer m such that

1 = w0 ≤ w1 ≤ . . . wm−1 ≤ wm ≥ wm+1 ≥ . . . wn−1 ≥ wn = 1.

This is true if L is modular, and also for L = Eq X with X finite ([5] and [7]). It
is known that w1 ≤ wk always holds for for 1 ≤ k < n ([1] and [4]). But a general
resolution of the conjecture still seems to be a long way off.

Exercises for Chapter 11

1. Let L be a finite geometric lattice, and let F be a nonempty order filter on L
(i.e., x ≥ f ∈ F implies x ∈ F ). Show that the lattice L0 obtained by identifying all
the elements of F (a join semilattice congruence) is geometric.
2. Draw the following geometric lattices and their corresponding geometries:
(a) Eq 4,
(b) Sub (Z2 )3 , the lattice of subspaces of a 3-dimensional vector space over Z2 .
3. Show that each of the following is an algebraic closure operator on <n with
the exchange property, and interpret them geometrically.
Pk
(a) Span(A) = { i=1 λi ai : k ≥ 1, ai ∈ A ∪ {0}}
115
 Pk Pk
(b) Γ(A) = { i=1 λi ai : k ≥ 1, ai ∈ A, λi = 1}
Pk Pi=1
k
(c) ∆(A) = { i=1 λi ai : k ≥ 1, ai ∈ A, i=1 λi = 1, λi ≥ 0}
4. Let G be a simple graph (no loops or multiple edges), and let X be the set of
all edges of G. Define S ⊆ X to be closed if whenever S contains all but one edge
of a cycle, then it contains the entire cycle. Verify that the corresponding closure
operator E is an algebraic closure operator with the exchange property. The lattice
of E-closed subsets is called the edge lattice of G. Find the edge lattices of the
graphs in Figure 11.1.

(a) (b) (c)

Figure 11.1

5. Show that the lattice for plane Euclidean geometry (<2 ) is not modular. (Hint:
Use two parallel lines and a point on one of them.)
6. (a) Let P and L be nonempty sets, which we will think of as “points” and
“lines” respectively. Suppose we are given an arbitrary incidence relation ∈ on P ×L.
Then we can make P ∪ L ∪ {0, 1} into a partially ordered set K in the obvious way,
interpreting p ∈ l as p ≤ l. When is K a lattice? atomistic? semimodular? modular?
subdirectly irreducible?
(b) Compare these results with Hilbert’s axioms for a plane geometry.
(i) There exists at least one line.
(ii) On each line there exist at least two points.
(iii) Not all points are on the same line.
(iv) There is one and only one line passing through two given distinct points.
7. Let L be a geometric lattice,
W and let A denote the set of atoms of L. A subset
S ⊆ A is independent if p 6≤ (S − {p}) for all p ∈ S. A subset B ⊆ A is a basis for
L if B is independent and ∨B = 1.
(a) Prove that L has a basis.
(b) Prove that if B and C are bases for L, then |B| = |C|.
(c) Show that the sublattice generated by an independent set S is isomorphic
to the lattice of all finite subsets of S.
8. A lattice is atomic if for every x > 0 there exists a ∈ L with x ≥ a  0. Prove
that every element of a complete, relatively complemented, atomic lattice is a join
of atoms.
116
 9. Let I be an infinite set, and let X = {pi : i ∈ I}∪{q
˙ i : i ∈ I}. Define a subset
S of X to be closed if S = X or, for all i, at most one of pi , qi is in S. Let L be the
lattice of all closed subsets of X.
(a) Prove that L is a relatively complemented algebraic lattice with every ele-
ment the join of atoms.
(b) Show that the compact elements of L do not form an ideal.
(This example shows that the semimodularity hypothesis of Theorem 11.1 cannot
be omitted.)
10. Prove that Eq X is relatively complemented and simple.

References
1. J. G. Basterfield and L. M. Kelly, A characterization of sets of n points which determine n
hyperplanes, Proc. Camb. Phil. Soc. 64 (1968), 585–588.
2. G. Birkhoff, Abstract linear dependence and lattices, Amer. J. Math. 57 (1935), 800–804.
3. R. P. Dilworth, A decomposition theorem for partially ordered sets, Ann. of Math. 51 (1950),
161–166.
4. C. Greene, A rank inequality for finite geometric lattices, J. Comb. Theory 9 (1970), 357–364.
5. L. H. Harper, The morphology of partially ordered sets, J. Combin. Theory Ser. A 17 (1974),
44–58.
6. J. Hashimoto, Direct, subdirect decompositions and congruence relations, Osaka J. Math. 9
(1957), 87–112.
7. W. N. Hsieh and D. J. Kleitman, Normalized matching in direct products of partial orders,
Stud. Appl. Math. 52 (1973), 258–289.
8. L. Libkin, Direct decompositions of atomistic algebraic lattices, Algebra Universalis 33 (1995),
127–135.
9. S. Mac Lane, A lattice formulation for transcendence degrees and p-bases, Duke Math. J. 4
(1938), 455–468.
10. K. Menger, F. Alt, and O. Schreiber, New foundations of projective and affine geometry, Ann.
of Math. 37 (1936), 456–482.

117
 12. Complemented Modular Lattices

Lookout, the saints are comin’ through

And it’s all over now, Baby Blue.
–Bob Dylan

Traditionally, complemented modular lattices are the alii 1 of lattices, and sub-
directly irreducible geometric modular lattices are the alii nui 2 . In this chapter we
will see why.
Lemma 12.1. Every complemented modular lattice is relatively complemented.
Proof. Let c be a complement of x in a modular lattice, and let a ≥ x ≥ b. Consider
the element z = a ∧ (b ∨ c), and note that z = b ∨ (a ∧ c) by modularity. Then
x ∧ z = x ∧ (b ∨ c) = b ∨ (x ∧ c) = b ∨ 0 = b, and dually x ∨ z = a. Thus z is a relative
complement of x in a/b. 
Theorem 12.2. Every algebraic complemented modular lattice is geometric.
Proof. Let L be an algebraic complemented modular lattice. We need to show that
the atoms of L are join dense, i.e., a > b implies that there is an atom p with
p ≤ a and p  b. But we know that L is weakly atomic, so there exist elements
c, d such that a ≥ c  d ≥ b. Let p be a relative complement of d in c/0. Then
c/d = (p ∨ d)/d ∼= p/(p ∧ d) = p/0, whence p is an atom. Also p ≤ c ≤ a, while
p  d implies p  b. 
The next result, though not needed in the sequel, is quite nice. It is due to Bjarni
Jónsson [2], extending O. Frink [1].
Theorem 12.3. Every complemented modular lattice can be embedded in a geomet-
ric modular lattice.
Sketch of Proof. Let L be a complemented modular lattice. Let F(L) denote the
lattice of filters of L ordered by reverse set inclusion, i.e., F ≤ G iff F ⊇ G. Since L
has a 0, every filter of L is contained in a maximal (w.r.t. set inclusion) filter; hence
F(L) is atomic (every nonzero element contains an atom). However, F(L) is dually
algebraic, so we form the ideal lattice I(F(L)), which is algebraic and is still atomic
(but Wit may not be atomistic). Let A denote the set of atoms of I(F(L)), and let
A = A, so that A is the ideal of F(L) generated by the maximal filters. Then the
1 Hawaiian chiefs.
2 Hawaiian big chiefs.
118
sublattice K = A/0 of I(F(L)) is modular, algebraic, and atomistic, whence it is
geometric. W
It remains to show that L is embedded into K by the map h(x) = {F ∈ A : x ∈
F }. This is a nontrivial (but doable) exercise. 
Since every geometric modular lattice is a direct product of subdirectly irreducible
ones, the following lemma comes in handy.
Lemma 12.4. A geometric modular lattice is subdirectly irreducible if and only if
for each pair of distinct atoms p, q there is a third atom r such that p, q and r
generate a sublattice isomorphic to M3 .
Proof. First, suppose that the condition of the lemma holds in a geometric modular
lattice L. Since a geometric lattice is relatively complemented, every nontrivial
congruence relation on L collapses some atom to 0. The condition implies that
if one atom collapses to 0, then they all do. Thus the congruence µ generated
by collapsing all the atoms to 0 is the unique minimum nontrivial congruence of
L, making it subdirectly irreducible. (In fact, µ collapses every finite dimensional
quotient a/b of L; unless the lattice is finite dimensional, this is not the universal
congruence. See exercise 3.)
Conversely, assume that L is subdirectly irreducible. Then the compact elements
of L form a simple, principally chain finite lattice; otherwise, as in the proof of
Theorem 11.5, Lc would be a proper direct sum and L would be a direct product.
Thus |QLc | = 1, and for any two atoms p, q of L there is a sequence of atoms with
p = p0 D p1 D . . . D pn = q.
Therefore, by induction, it will suffice to prove the following two claims.
(i) If p, q and r are atoms of L with p D q D r, then p D r.
(ii) If p and q are distinct atoms with p D q, then there is an atom s such that
p, q and s generate a diamond.
Let us prove (ii) first. If p and q are distinct and p D q, then by definition there is
an element x ∈ L such that p ≤ q ∨ x but p  q∗ ∨ x = x. Note that, by modularity,
q ∨ x  x and hence p ∨ x = q ∨ x. Set s = x ∧ (p ∨ q). Since p ∨ q has dimension
2 and p  x, we have s  0. Now p  x implies p ∧ s ≤ p ∧ x = 0. Similarly
q  x and that implies q ∧ s = 0, while p 6= q gives p ∧ q = 0. On the other hand,
p ∨ s = (p ∨ x) ∧ (p ∨ q) = p ∨ q, and similarly q ∨ s = p ∨ q. Thus p, q and s generate
an M3 .
To prove (i), assume p D q D r. Without loss of generality, these are distinct,
and hence by (ii) there exist atoms x and y such that p, q, x generate a diamond
and q, r, y likewise. Set z = (p ∨ r) ∧ (x ∨ y). Then
r ∨ z = (p ∨ r) ∧ (r ∨ x ∨ y)
= (p ∨ r) ∧ (x ∨ q ∨ y)
= (p ∨ r) ∧ (p ∨ q ∨ y) ≥ p.
119
Now there are two possibilities. If p  z, then p D r via z, and we are done. So
assume p ≤ z. Then p ≤ x ∨ y and x ∨ y has dimension 2, so x ∨ y = p ∨ x = q ∨ x.
But then q ≤ x ∨ y, whence x ∨ y = q ∨ y also, i.e., the tops of the two diamonds
coincide. In particular p, q and r join pairwise to x ∨ y, so that again p D r. 
It turns out that dimension plays an important role in subdirectly irreducible
geometric lattices. We define the dimension of a geometric lattice L to be the
length of a maximal chain in L. Thus δ(L) = δ(1) if δ(1) = n < ∞; more generally,
δ(L) = |B| where B is a basis for L (see exercise 11.7).
Of course 2 is the only geometric lattice with δ(L) = 1. Geometric lattices with
δ(L) = 2 are isomorphic to Mκ for some cardinal κ, and these are simple whenever
κ > 2.
Subdirectly irreducible geometric modular lattices with δ(L) > 2 correspond to
projective geometries of geometric dimension ≥ 2. In particular, for δ(L) = 3 they
correspond to projective planes. Projective planes come in two types: arguesian and
nonarguesian. The nonarguesian projective planes are sort of strange: we can con-
struct lots of examples of them, but there is no really good representation theorem
for them.
On the other hand, a theorem of classical projective geometry translates as fol-
lows.
Theorem 12.5. Every subdirectly irreducible geometric modular lattice with δ(L) ≥
4 is arguesian.
Now we are ready for the best representation theorem of them all, due to Birkhoff
and Frink (but based on older ideas from projective geometry). Recall that the lat-
tice of subspaces of any vector space is a subdirectly irreducible geometric arguesian
lattice.
Theorem 12.6. Let L be a subdirectly irreducible geometric arguesian lattice with
δ(L) = κ ≥ 3. Then there is a division ring D such that L is isomorphic to the
lattice of all subspaces of a κ-dimensional vector space over D.
A later version of these notes will include a proof of Theorem 12.6, but not this
one.

That’s all, folks!

Exercises for Chapter 12

1. For what values of n is Mn ∼ = Sub V for some vector space V ?
2. The following steps carry you through the proof of Theorem 12.3.
(a) Show that if L is a complete, upper continuous, modular lattice and 1 is a
join of atoms, then L is geometric (i.e., algebraic and atomistic).
(b) Find a finite semimodular lattice in which the atoms join to 1, but not every
element is a join of atoms.
120
 (c) For the map h : L → K given in the text, h(x) is an ideal of F(L), and hence
a set of filters. Show that F ∈ h(x) iff F is the intersection of finitely many
maximal (w.r.t. set inclusion) filters and x ∈ F . Note that this implies that
h preserves order.
(d) Use (c) to show that h(x ∧ y) = h(x) ∧ h(y).
(e) Show that, in order to prove that h preserves joins, it suffices to prove that
h(x ∨ y) ≤ h(x) ∨ h(y) whenever x ∧ y = 0.
(f) Prove that if x ∧ y = 0 and F is a maximal filter with F ∈ h(x ∨ y), then
there exist G ∈ h(x) and H ∈ h(y) such that F ≤ G ∨ H, i.e., F ⊇ G ∩ H.
Thus h preserves joins (by (e)).
3. Define a relation ξ on a modular lattice L by ha, bi ∈ ξ iff (a ∨ b)/(a ∧ b) is
finite dimensional. Show that ξ ∈ Con L. Give examples to show that ξ can be 0,
1 or neither in Con L.

References
1. O. Frink, Complemented modular lattices and projective spaces of infinite dimension, Trans.
Amer. Math. Soc. 60 (1946), 452–467.
2. B. Jónsson, Modular lattices and Desargues’ theorem, Math. Scand. 2 (1954), 295–314.
3. J. von Neumann, Continuous Geometries, I. Halperin, ed., Princeton University Press, Prince-
ton, 1960.

121
 Appendix 1: Cardinals, Ordinals and Universal Algebra

In these notes we are assuming you have a working knowledge of cardinals and
ordinals. Just in case, this appendix will give an informal summary of the most
basic part of this theory. We also include an introduction to the terminology of
universal algebra.

1. Ordinals
Let C be a well ordered set, i.e., a chain satisfying the descending chain condition
(DCC). A segment of C is a proper ideal of C, which (because of the DCC) is
necessarily of the form {c ∈ C : c < d} for some d ∈ C.
Lemma. Let C and D be well ordered sets. Then
(1) C is not isomorphic to any segment of itself.
(2) Either C ∼= D, or C is isomorphic to a segment of D, or D is isomorphic
to a segment of C.
We say that two well ordered sets have the same type if C ∼ = D. An ordinal is
an order type of well ordered sets. These are usually denoted by lower case Greek
letters: α, β, γ, etc. For example, ω denotes the order type of the natural numbers,
which is the smallest infinite ordinal. We can order ordinals by setting α ≤ β if
α∼ = β or α is isomorphic to a segment of β. There are too many ordinals in the
class of all ordinals to call this an ordered set without getting into set theoretic
paradoxes, but we can say that locally it behaves like one big well ordered set.
Theorem. Let β be an ordinal, and let B be the set of all ordinals α with α < β,
ordered by ≤. Then B ∼
= β.
For example, ω is isomorphic to the collection of all finite ordinals.
Recall that the Zermelo well ordering principle (which is equivalent to the Axiom
of Choice) says that every set can be well ordered. Another way of putting this is
that every set can be indexed by ordinals,
X = {xα : α < β}
for some β. Transfinite induction is a method of proof which involves indexing a set
by ordinals, and then applying induction on the indices. This makes sense because
the indices satisfy the DCC.
In doing transfinite induction, it is important to distinguish two types of ordinals.
β is a successor ordinal if {α : α < β} has a largest element. Otherwise, β is called
a limit ordinal. For example, every finite ordinal is a successor ordinal, and ω is a
limit ordinal.
122
 2. Cardinals
We say that two sets X and Y have the same cardinality, written |X| = |Y |,
if there exists a one-to-one onto map f : X → Y . It is easy to see that “having
the same cardinality” is an equivalence relation on the class of all sets, and the
equivalence classes of this relation are called cardinal numbers. We will use lower
case german letters such as m, n and p to denote unidentified cardinal numbers.
We order cardinal numbers as follows. Let X and Y be sets with |X| = m and
|Y | = n. Put m ≤ n if there exists a one-to-one map f : X  Y (equivalently, if
there exists an onto map g : Y  X). The Cantor-Bernstein theorem says that this
relation is anti-symmetric: if m ≤ n ≤ m, then m = n, which is the hard part of
showing that it is a partial order.

Theorem. Let m be any cardinal. Then there is a least ordinal α with |α| = m.

Theorem. Any set of cardinal numbers is well ordered.1

Now let |X| = m and |Y | = n with X and Y disjoint. We introduce operations

on cardinals (which agree with the standard operations in the finite case) as follows.

m + n = |X ∪ Y |
m · n = |X × Y |
mn = |{f : Y → X}|

It should be clear how to extend + and · to arbitrary sums and products.

The basic arithmetic of infinite cardinals is fairly simple.

Theorem. Let m and n be infinite cardinals. Then

(1) m + n = m · n = max {m, n},
(2) 2m > m.

The finer points of the arithmetic can get complicated, but that will not bother
us here. The following facts are used frequently.

Theorem. Let X be an infinite set, P(X) the lattice of subsets of X, and Pf (X)
the lattice of finite subsets of X. Then |P(X)| = 2|X| and |Pf (X)| = |X|.

A fine little book [2] by Irving Kaplansky, Set Theory and Metric Spaces, is easy
reading and contains the proofs of the theorems above. The book Introduction
to Modern Set Theory by Judith Roitman [4] is recommended for a slightly more
advanced introduction.

1 Again, there are too many cardinals to talk about the “set of all cardinals”.
123
 3. Universal Algebra
Once you have seen enough different kinds of algebras: vector spaces, groups,
rings, semigroups, lattices, even semilattices, you should be driven to abstraction.
The proper abstraction in this case is the general notion of an “algebra”. Univer-
sal algebra is the study of the properties which different types of algebras have in
common. Historically, lattice theory and universal algebra developed together, more
like Siamese twins than cousins. In these notes we do not assume you know much
universal algebra, but where appropriate we do use its terminology.
An operation on a set A is just a function f : An → A for some n ∈ ω. An algebra
is a system A = hA; Fi where A is a nonempty set and F is a set of operations on
A. Note that we allow infinitely many operations, but each has only finitely many
arguments. For example, lattices have two binary operations, ∧ and ∨. We use
different fonts to distinguish between an algebra and the set of its elements, e.g., A
and A.
Many algebras have distinguished elements, or constants. For example, groups
have a unit element e, rings have both 0 and 1. Technically, these constants are
nullary operations (with no arguments), and are included in the set F of operations.
However, in these notes we sometimes revert to a more old-fashioned notation and
write them separately, as A = hA; F, Ci, where F is the set of operations with at
least one argument and C is the set of constants. There is no requirement that
constants with different names, e.g., 0 and 1, be distinct.
A subalgebra of A is a subset S of A which is closed under the operations, i.e.,
if s1 , . . . , sn ∈ S and f ∈ F, then f (s1 , . . . , sn ) ∈ S. This means in particular that
all the constants of A are contained in S. If A has no constants, then we allow the
empty set as a subalgebra (even though it is not properly an algebra). Thus the
empty set is a sublattice of a lattice, but not a subgroup of a group. A nonempty
subalgebra S of A can of course be regarded as an algebra S of the same type as A.
If A and B are algebras with the same operation symbols (including constants),
then a homomorphism from A to B is a mapping h : A → B which preserves the
operations, i.e., h(f (a1 , . . . , an )) = f (h(a1 ), . . . , h(an )) for all a1 , . . . , an ∈ A and
f ∈ F. This includes that h(c) = c for all c ∈ C.
A homomorphism which is one-to-one is called an embedding, and sometimes
written h : A  B or h : A ≤ B. A homomorphism which is both one-to-one and
onto is called an isomorphism, denoted h : A ∼ = B.
These notions directly generalize notions which should be perfectly familiar to
you for say groups or rings. Note that we have given only terminology, but no
results. The basic theorems of universal algebra are included in the text, either
in full generality, or for lattices in a form which is easy to generalize. For deeper
results in universal algebra, there are several nice textbooks available, including A
Course in Universal Algebra by S. Burris and H. P. Sankappanavar [1], and Algebras,
Lattices, Varieties by R. McKenzie, G. McNulty and W. Taylor [3].
124
 References
1. S. Burris and H. P. Sankappanavar, A Course in Universal Algebra, Springer-Verlag, New
York, 1980.
2. I. Kaplansky, Set Theory and Metric Spaces, Allyn and Bacon, Boston, 1972.
3. R. McKenzie, G. McNulty and W. Taylor, Algebras, Lattices, Varieties, vol. I, Wadsworth and
Brooks-Cole, Belmont, CA, 1987.
4. J. Roitman, Introduction to Modern Set Theory, Wiley, New York, 1990.

125
 Appendix 2: The Axiom of Choice

In this appendix we want to prove Theorem 1.5.

Theorem 1.5. The following set theoretic axioms are equivalent.

(1) (Axiom of Choice) If X is a nonempty set, then there is a map φ :
P(X) → X such that φ(A) ∈ A for every nonempty A ⊆ X.
(2) (Zermelo well-ordering principle) Every nonempty set admits a well-
ordering (a total order satisfying the DCC ).
(3) (Hausdorff maximality principle) Every chain in an ordered set P can
be embedded in a maximal chain.
(4) (Zorn’s Lemma) If every chain in an ordered set P has an upper bound in
P, then P contains a maximal element.
(5) If every chain in an ordered set P has a least upper bound in P, then P
contains a maximal element.

Let us start by proving the equivalence of (1), (2) and (4).

(4) =⇒ (2): Given a nonempty set X, let Q be the collection of all pairs (Y, R)
such that Y ⊆ X and R is a well ordering of Y , i.e., R ⊆ Y × Y is a total order
satisfying the DCC. Order Q by (Y, R) v (Z, S) if Y ⊆ Z and R ⊆ S. In order
to apply Zorn’s
S Lemma, check that if {(Yα , Rα ) : α ∈ A} is a chain in Q, then
S
(Y , R) = ( Yα , Rα ) ∈ Q and (Yα , Rα ) v (Y , R) for every α ∈ A, and so (Y , R)
is an upper bound for {(Yα , Rα ) : α ∈ A}. Thus Q contains a maximal element
(U, T ). Moreover, we must have U = X. For otherwise we could choose an element
z ∈ X −U , and then the pair (U 0 , T 0 ) with U 0 = U ∪{z} and T 0 = T ∪{(u, z) : u ∈ U }
would satisfy (U, T ) @ (U 0 , T 0 ), a contradiction. Therefore T is a well ordering of
U = X, as desired.
(2) =⇒ (1): Given a well ordering ≤ of X, we can define a choice function φ
on the nonempty subsets of X by letting φ(A) be the least element of A under the
ordering ≤.
(1) =⇒ (4): For a subset S of an ordered set P, let S u denote the set of all
upper bounds of S, i.e., S u = {x ∈ P : x ≥ s for all s ∈ S}.
Let P be an ordered set in which every chain has an upper bound. By the Axiom
of Choice there is a function φ on the subsets of P such that φ(S) ∈ S for every
nonempty S ⊆ P . We use the choice function φ to construct a function which
assigns a strict upper bound to every subset of P which has one as follows: if S ⊆ P
and S u − S = {x ∈ P : x > s for all s ∈ S} is nonempty, define γ(S) = φ(S u − S).
126
 Fix an element x0 ∈ P . Let B be the collection of all subsets B ⊆ P satisfying
the following properties.
(1) B is a chain.
(2) x0 ∈ B.
(3) x0 ≤ y for all y ∈ B.
(4) If A is a nonempty order ideal of B and z ∈ B ∩ (Au − A), then γ(A) ∈
B ∩ z/0.
The last condition says that if A is a proper ideal of B, then γ(A) is in B, and
moreover it is the least element of B strictly above every member of A.
Note that B is nonempty, since {x0 } ∈ B.
Next, we claim that if B and C are both in B, then either B is an order ideal of
C or C is an order ideal of B. Suppose not, and let A = {t ∈ B ∩ C : t/0 ∩ B =
t/0 ∩ C}. Thus A is the largest common ideal of B and C; it contains x0 , and by
assumption is a proper ideal of both B and C. Let b ∈ B − A and c ∈ C − A. Now
B is a chain and A is an ideal of B, so b ∈ / A implies b > a for all a ∈ A, whence
b ∈ B ∩(Au −A). Likewise c ∈ C ∩(Au −A). Hence by (4), γ(A) ∈ B ∩C. Moreover,
since b was arbitrary in B − A, again by (4) we have γ(A) ≤ b for all b ∈ B − A,
and similarly γ(A) ≤ c for all c ∈ C − A. Therefore

γ(A)/0 ∩ B = A ∪ {γ(A)} = γ(A)/0 ∩ C

whence γ(A) ∈ A, contrary to the definition of γ.

It follows, that if B and C are in B, b ∈ B and c ∈ C, and b ≤ c, then b ∈ C.
Also, you can easily check that if B ∈ B and B u − B is nonempty, then B ∪
{γ(B)} ∈ B. S
Now let U = B∈B B. We claim that U ∈ B. It is a chain because for any two
elements b, c ∈ U there exist B, C ∈ B with b ∈ B and c ∈ C; one of B and C is
an ideal of the other, so both are contained in the larger set and hence comparable.
Conditions (2) and (3) are immediate. If a nonempty ideal A of U has a strict upper
bound z ∈ U , then z ∈ C for some C ∈ B. By the observation above, A is an ideal
of C, and hence the conclusion of (4) holds.
Now U is a chain in P, and hence by hypothesis U has an upper bound x. On
the other hand, U u − U must be empty, for otherwise U ∪ {γ(U )} ∈ B, whence
γ(U ) ∈ U , a contradiction. Therefore x ∈ U and x is maximal in P. In particular,
P has a maximal element, as desired.
Now we prove the equivalence of (3), (4) and (5).
(4) =⇒ (5): This is obvious, since the hypothesis of (5) is stronger.
(5) =⇒ (3): Given an ordered set P, let Q be the set of allSchains in P, ordered
by set containment. If {Cα : α ∈ A} is a chain in Q, then Cα is a chain in P
which is the least upper bound of {Cα : α ∈ A}. Thus Q satisfies the hypothesis of
(5), and hence it contains a maximal element C, which is a maximal chain in P.
127
 (3) =⇒ (4): Let P be an ordered set such that every chain in P has an upper
bound in P . By (3), there is a maximal chain C in P. If b is an upper bound for
C, then in fact b ∈ C (by maximality), and b is a maximal element of P.
(There are many variations of the proof of Theorem 1.5, but it can always be
arranged so that there is only one hard step, and the rest easy. The above version
seems fairly natural.)

128
 Appendix 3: Formal Concept Analysis

Exercise 13 of Chapter 2 is to show that a binary relation R ⊆ A × B induces a

pair of closure operators, described as follows. For X ⊆ A, let

σ(X) = {b ∈ B : x R b for all x ∈ X}.

Similarly, for Y ⊆ B, let

π(Y ) = {a ∈ A : a R y for all y ∈ Y }.

Then the composition πσ : P(A) → P(A) is a closure operator on A, given by

πσ(X) = {a ∈ A : a R b whenever x R b for all x ∈ X}.

Likewise, σπ is a closure operator on B, and for Y ⊆ B,

σπ(Y ) = {b ∈ B : a R b whenever a R y for all y ∈ Y }.

In this situation, the lattice of closed sets Cπσ ⊆ P(A) is dually isomorphic to
Cσπ ⊆ P(B), and we say that R establishes a Galois connection between the πσ-
closed subsets of A and the σπ-closed subsets of B.
Of course, Cπσ is a complete lattice. Moreover, every complete lattice can be
represented via a Galois connection.
Theorem. Let L be a complete lattice, A a join dense subset of L and B a meet
dense subset of L. Define R ⊆ A × B by a R b if and only if a ≤ b. Then, with σ
and π defined as above, L ∼
= Cπσ (and L is dually isomorphic to Cσπ ).
In particular, for an arbitrary complete lattice, we can always take A = B = L. If
L is algebraic, a more natural choice is A = Lc and B = M ∗ (L) (compact elements
and completely meet irreducibles). If L is finite, the most natural choice is A = J(L)
and B = M (L). Again the proof of this theorem is elementary.
Formal Concept Analysis is a method developed by Rudolf Wille and his col-
leagues in Darmstadt (Germany), whereby the philosophical Galois connection be-
tween objects and their properties is used to provide a systematic analysis of cer-
tain very general situations. Abstractly, it goes like this. Let G be a set of “ob-
jects” (Gegenstände) and M a set of relevant “attributes” (Merkmale). The relation
I ⊆ G × M consists of all those pairs hg, mi such that g has the property m. A
concept is a pair hX, Y i with X ⊆ G, Y ⊆ M , X = π(Y ) and Y = σ(X). Thus
129
hX, Y i is a concept if X is the set of all elements with the properties of Y , and
Y is exactly the set of properties shared by the elements of X. It follows (as in
exercise 12, Chapter 2) that X ∈ Cπσ and Y ∈ Cσπ . Thus if we order concepts by
hX, Y i ≤ hU, V i iff X ⊆ U (which is equivalent to Y ⊇ V ), then we obtain a lattice
B(G, M, I) isomorphic to Cπσ .
A small example will illustrate how this works. The rows of Table A1 correspond
to seven fine musicians, and the columns to eight possible attributes (chosen by a
musically trained sociologist). An × in the table indicates that the musician has
that attribute.1 The corresponding concept lattice is given in Figure A2, where the
musicians are abbreviated by lower case letters and their attributes by capitals.

Instrument Classical Jazz Country Black White Male Female

J. S. Bach × × × ×
Rachmaninoff × × × ×
King Oliver × × × ×
W. Marsalis × × × × ×
B. Holiday × × ×
Emmylou H. × × ×
Chet Atkins × × × × ×

Table A1.

I=M J

B W

Cl o Co F

b=r m a h e

Figure A2
1 To avoid confusion, androgynous rock stars were not included.
130
 Formal concept analysis has been applied to hundreds of real situations outside
of mathematics (e.g., law, medicine, psychology), and has proved to be a useful tool
for understanding the relation between the concepts involved. Typically, these ap-
plications involve large numbers of objects and attributes, and computer programs
have been developed to navigate through the concept lattice. A good brief intro-
duction to concept analysis may be found in Wille [2] or [3], and the whole business
is explained thoroughly in Ganter and Wille [1].
Likewise, the representation of a finite lattice as the concept lattice induced by
the order relation between join and meet irreducible elements (i.e., ≤ restricted to
J(L) × M (L)) provides and effective and tractable encoding of its structure. As an
example of the method, let us show how one can extract the ordered set QL such
that Con L ∼ = O(Q(L)) from the table.
Given a finite lattice L, for g ∈ J(L) and m ∈ M (L), define
g%m if g  m but g ≤ m∗ , i.e., g ≤ n for all n > m,
m&g if m  g but m ≥ g∗ , i.e., m ≥ h for all h < g,
glm if g % m and m & g.
Note that these relations can easily be added to the table of J(L) × M (L).
These relations connect with the relation D of Chapter 10 as follows.
Lemma. Let L be a finite lattice and g, h ∈ J(L). Then g D h if and only if there
exists m ∈ M (L) such that g % m & h.
Proof. If g D h, then there exists x ∈ L such that g ≤ h ∨ x but g  h∗ ∨ x. Let m
be maximal such that m ≥ h∗ ∨ x but m  g. Then m ∈ M (L), g ≤ m∗ , m ≥ h∗
but m  h. Thus g % m & h.
Conversely, suppose g % m & h. Then g ≤ m∗ ≤ h ∨ m while g  m = h∗ ∨ m.
Therefore g D h. 
As an example, the table for the lattice in Figure A2 is given in Table A3. This is
a reduction of the original Table A1: J(L) is a subset of the original set of objects,
and likewise M (L) is contained in the original attributes. Arrows indicating the
relations %, & and l have been added. The Lemma allows us to calculate D
quickly, and we find that |QL | = 1, whence L is simple.

I=M Cl J Co B W F
b=r × × l l & × l
o × l × × % %
m × × × & × l l
h l & × & × l ×
e l & l × l × ×
a × l × × l × l
131
 Table A3.

References
1. B. Ganter and R. Wille, Formale Begriffsanalyse: Mathematische Grundlagen, Springer-
Verlag, Berlin-Heidelberg, 1996.
2. R. Wille, Restructuring lattice theory: an approach based on hierarchies of concepts, Ordered
Sets, I. Rival, ed., Reidel, Dordrecht-Boston, 1982, pp. 445–470.
3. R. Wille, Concept lattices and conceptual knowledge systems, Computers and Mathematics
with Applications 23 (1992), 493–515.

132