Professional Documents
Culture Documents
Semigroups96 PDF
Semigroups96 PDF
SEMIGROUPS
Tero Harju
Department of Mathematics
University of Turku
FIN-20014 Turku, Finland
19961
1
Small corrections in 2010
Contents
1 Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1 Definitions and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Subsemigroups and Direct Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Homomorphisms and transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Partial orders and semilattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4 Combinatorial Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.1 The defect theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Ehrenfeucht’s conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3 Isbell’s Zig-Zag Theorem for Dominions . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5 Green’s Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.2 Green’s Lemma and Its Corollaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6 Inverse Semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.1 Regular Semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.2 Inverse Semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.3 Representations by Injective Partial Mappings . . . . . . . . . . . . . . . . . . . . . . 67
6.4 Congruences of Inverse semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Contents 3
Books on semigroups:
A.H. Clifford and G.B. Preston, The algebraic theory of semigroups, Vol. I & II, Math-
ematical Surveys of the Amer. Math. Soc. 7 (1961 & 1967). [A two-volume ‘handbook’
on semigroups from their algebraic point of view.]
P.M. Higgins, Techniques of semigroup theory, Oxford University Press, 1992 [Goes to
the advanced topics rather fast. Contains up-to-date proofs for free inverse semigroups,
topics on biordered sets, Isbell’s zig-zags, and some combinatorics of transformation
semigroups.]
J.M. Howie, An introduction to semigroup theory, Academic Press, 1976. [An easy
introduction. See the next entry.]
J.M. Howie, Fundamentals of semigroup theory, 1995. [Probably the best introduction.
Contains more advanced material such as Isbell’s zig-zags, and the free inverse semi-
groups].
G. Lallement, Semigroups and combinatorial applications, Wiley, 1979. [Concentrates
on combinatorial topics such as codes and free word semigroups, formal languages and
related topics.]
E.S. Lyapin, Semigroups, Translations of Math. Monographs, Vol 3, Amer. Math. Soc.,
1974. [Easy to read.]
M. Petrich, Introduction to semigroups, Merrill, 1973. [More advanced topics from the
structural point of view.]
M. Petrich, Lectures in semigroups, Wiley, 1977. [Bands, varieties, and lattices of vari-
eties.]
M. Petrich, Inverse semigroups, Wiley, 1984. [The bible on inverse semigroup.]
M. Petrich and N.R. Reilly, Completely regular semigroups, in preparation. [A large
book on one of the popular themes in semigroup theory.]
Related books:
J. Berstel and D. Perrin, Theory of Codes, Academic Press, 1985. [Formal languages,
codes, word semigroups.]
S. Burris and H.P. Sankappanavar, A course in universal algebra, Springer, 1981. [A
good introduction to general properties of algebras.]
P.M. Cohn, Universal algebra, Reidel, 1981. [A more comprehensive treatment of uni-
versal algebras.]
S. Eilenberg, Automata, Languages, and Machines, Vol. A, Academic Press, 1974. [One
of the basic books on automata.]
M. Lothaire, Combinatorics on Words, Addison-Wesley, 1983. [Combinatorial topic on
words and related areas.]
Contents 4
R.C. Lyndon and P. Schupp, Combinatorial group theory, Springer, 1977. [Contains
some theory of free groups among other topics.]
W. Magnus, A. Karrass and D. Solitar, Combinatorial group theory, Dover, 1976 [Con-
tains some theory of free groups among other topics.]
A. Salomaa and G. Rozenberg (eds), Handbook of formal languages, Vol. I, Springer,
1996. [A collection of survey articles by various authors on combinatorial topics on
words, and formal languages in general. Includes two proofs of Ehrenfeucht’s conjec-
ture; one of which is in these Lecture Notes. Includes also many other interesting results
of combinatorial nature.]
1
Basic Concepts
Semigroups
Let S be a set and σ : S ×S → S a binary operation that maps each ordered pair (x, y) of
S to an element σ(x, y) of S. The pair (S, σ) (or just S, if there is no fear of confusion)
is called a groupoid. The mapping σ is called a product of (S, σ). We shall mostly write
simply xy instead of σ(x, y). If we want to emphasize the place of the operation, then
we often write x · y. The element xy (= σ(x, y)) is the product of x and y in S.
We also use other notations for the product, when this is natural or otherwise conve-
nient. In particular, the following symbols may be used: ·, +, ⋆, ◦, ⊕, ⊗. If we do not
explicitly (or implicitly) mention the (notation for the) product, we choose the dot: x · y.
A groupoid S is a semigroup, if the operation σ is associative: for all x, y, z ∈ S,
x · (y · z) = (x · y) · z .
That is, σ(x, σ(y, z)) = σ(σ(x, y), z) using the unfriendly notation. This means that the
order in which the operation σ is carried out is irrelevant, and therefore we may write
x1 x2 . . . xn = x1 · (x2 · (· · · · xn )) . . . ) .
Of course, the order of the operands x1 , . . . , xn (where we can have repeated elements,
xi = xj ), is important.
For an element x ∈ S we let xn be the product of x with itself n times. So, x1 = x,
x = x · x, and xn+1 = x · xn for n ≥ 1.
2
∀x, y ∈ S : x · y = y · x .
Example 1.1. (1) (N, ·) is a semigroup for the usual multiplication of integers, n · m.
Also (N, +) is a semigroup, when + is the ordinary addition of integers. Define (N, ⋆)
by
n ⋆ m = max{n, m} .
1.1 Definitions and examples 6
(2) The previous case generalizes to other familiar sets of numbers: (Z, ·), (Z, +),
(Q, ·), (Q, +), (R, ·), (R, +), (C, ·), (C, +) of integers, rationals, reals and complex
numbers. In the multiplicative cases one may remove the zero element 0 from the do-
main, and still obtain a semigroup.
(3) Consider the upper triangular integer matrices
1 n
S= n≥1 .
0 1
Then S is a semigroup with the usual matrix product. All the semigroups in the above
examples have been commutative.
(4) Let TX be the set of all functions α : X → X. Then (TX , ◦) is a semigroup,
where ◦ is the composition of functions: (β ◦ α)(x) = β(α(x)) for all x ∈ X.
(5) Let S = {a, b, c} be a set of three elements, and define the product in the table.
Then S is a finite semigroup. In order to check this we should run · a b c
through all the triples: x · (y · z) = (x · y) · z. In this case, there a a b c
are some helpful restrictions in S. For instance, if z = c, then the b b a c
product is always c no matter what x and y are. ⊔
⊓ c c b c
∀y ∈ S : x · y = y .
∀y ∈ S : y · x = y .
Lemma 1.1. A semigroup S can have at most one identity. In fact, if S has a left identity
x and a right identity y, then x = y. In particular, the identity of a monoid is unique.
x · x−1 = 1 = x−1 · x .
Although semigroups and groups seem to be rather close to each other, semigroups lack
the basic symmetry properties of groups, and therefore the general theory of semigroups
differs quite a lot from the theory of groups. The theory of semigroups is more applicable
in theoretical computer science and also in some parts of combinatorial mathematics.
∀y ∈ S : x · y = x .
∀y ∈ S : y · x = x .
Proof. Exercise. ⊔
⊓
Example 1.2. (1) Consider the finite semigroup S as defined above in Example 1.1(5).
The element a is an identity of S, and hence S is a monoid. The element c is a right zero.
There are no left zeros, and hence S does not have a zero. Both a and c are idempotents.
Indeed, all left and right identities and zeros are idempotents. The element b is not an
idempotent, since b2 = a.
(2) The matrix semigroup from Example 1.1(3) has no idempotents let alone identi-
ties or zeros.
(3) Consider the set of all n-by-n-matrices with integer entries. This is a monoid
Zn×n . For instance,
1.1 Definitions and examples 8
a b
Z 2×2
= a, b, c, d ∈ Z .
c d
The identity matrix I is the identity of Z2×2 and the zero matrix 0 is its zero:
1 0 0 0
I= and 0 = .
0 1 0 0
−1 −1
This monoid has quite many idempotents. As an example, is an idempotent.
2 2
(4) Let GL(n, R) denote the monoid of all nonsingular n-by-n-matrices A with real
entries – so that det(A) 6= 0. Apart from the identity matrix GL(n, R) has no other
idempotents, because it is a group, called the generalized linear group over R. Also
the monoid SL(n, R) of all n-by-n-matrices A in Rn×n such that det(A) = 1 forms a
group. This is the special linear group over R.
(5) Let us define a product in a set S by
∀x, y ∈ S : x · y = x .
We obtain thus a semigroup, where each element x ∈ S is a right identity and at the
same time a left zero. Moreover, all the elements in this semigroup are idempotents,
S = ES . ⊔
⊓
Proof. Exercise. ⊔
⊓
Example 1.3. For any two subsets A, B ⊆ S of a semigroup S, define their product by
A · B = {a · b | a ∈ A, b ∈ B} .
The so obtained binary operation is associative, and hence the subsets of S form a semi-
group, the global semigroup of S. We denote this new semigroup by 2S . For a subset
A ⊆ S, we let, as usual, A2 = A · A. ⊔
⊓
Cancellation semigroups
A semigroup S is left cancellative, if
zx = zy =⇒ x = y ,
xz = yz =⇒ x = y .
and hence Z2×2 is not cancellative. However, if we take all matrices A with det(A) = 1,
then this semigroup is cancellative. ⊔
⊓
Subsemigroups
Let A 6= ∅ be a (nonempty) subset of a semigroup (S, ·). We say that (A, ·) is a sub-
semigroup of S, denoted by A ≤ S, if A is closed under the product of S:
∀x, y ∈ A : x · y ∈ A ,
that is,
A ≤ S ⇐⇒ A2 ⊆ A .
A subsemigroup uses the operation of its mother semigroup, and hence if A is a
subsemigroup of S, then certainly the operation of A is associative, and thus A is a
semigroup by its own right.
Example 1.5. Consider the additive semigroup S = (Q, +) of rationals. Now, (N, +) is
a subsemigroup of S, but (N, ·) is not, because its operation (multiplication) is not the
operation of S. ⊔
⊓
Proof. Write A = ∪∞ n n
n=1 X . It is easy to show that A ≤ S. Also, X ⊆ [X]S for all
n ≥ 1, since [X]S ≤ S, and hence the claim follows. ⊔
⊓
Example 1.6. (1) Let S be the finite semigroup from Example 1.1(5). Now, [a]S =
{a, a2 , a3 , . . . } = {a}, and hence {a} is a subsemigroup of S. Also, [c]S = {c}. On the
other hand, [b]S = {a, b} = [a, b]S , and [b, c]S = {a, b, c} = S. Here {b, c} generates
the full semigroup S.
(2) Consider the matrix semigroup S = Z2×2 , and let
0 1
M= .
−1 0
Now,
2 −1 0 3 0 −1 4 1 0
M = , M = , M = , M 5 = M,
0 −1 1 0 0 1
and hence [M ]S = {I, M, M 2 , M 3 } is a finite subsemigroup of S. In fact, [M ]S is a
subgroup of S, that is, a subsemigroup which is a group. Note that Z2×2 is not a group
itself. ⊔
⊓
Lemma 1.5. Let r be the index and p the period of an element x ∈ S. Then
Kx = {xr , xr+1 , . . . , xr+p−1 }
is a subgroup (i.e., a subsemigroup that is a group) of S.
Direct products
The direct product S × T of two semigroups S and T is defined by
It is easy to show that the so defined product is associative on the pairs (x, y) ∈ S × T ,
and hence the direct product is, indeed, a semigroup.
Example 1.7. Let S = (N, +) and T = (N, ·). Then in the direct product S × T we have
(n, r) · (m, s) = (n + m, rs). ⊔
⊓
The direct product is a convenient way of combining two semigroup operations. The
new semigroup S × T inherits properties of both S and T .
The mappings π1 : S × T → S and π2 : S × T → T such that π1 (x, y) = x and
π2 (x, y) = y are called the projections of the direct product S × T .
In general S × T 6= T × S, as already verified in the previous example. However, the
direct product operation is associative on semigroups: S × (T × U ) = (S × T ) × U , and
hence we can define S1 ×S2 ×· · ·×Sn as the (finite) direct product of the semigroups Si .
Definition
Let (S, ·) and (P, ⋆) be two semigroups. A mapping α : S → P is a homomorphism, if
Example 1.8. (1) Let S = (N, +) and P = (N, ·), and define α(n) = 2n for all n ∈ N.
Now,
α(n + m) = 2n+m = 2n · 2m = α(n) · α(m) ,
and hence α : S → P is a homomorphism.
(2) Define S and P by the following tables:
· a b c d
⋆ e f g
a a b c d
e e f g
b b a d c
f f e g
c c d c d
g g g g
d d c d c
1.3 Homomorphisms and transformations 12
Proof. Exercise. ⊔
⊓
Proof. First of all the inverse mapping α−1 exists, because α is a bijection. Furthermore,
αα−1 = ι, and thus because α is a homomorphism,
and so α−1 (x) · α−1 (y) = α−1 (xy), which proves the claim. ⊔
⊓
Example 1.9. (1) Let S be an infinite monogenic semigroup, that is, S is generated by
one element S = [x] = {x, x2 , x3 , . . . }. Define α : S → (N+ , +) by α(xn ) = n.
Now, α is a homomorphism, since α(xn · xm ) = α(xn+m ) = n + m = α(xn ) +
α(xm ). Moreover, α is a bijection, and thus an isomorphism. Therefore, S ∼
= (N+ , +). In
conclusion, each infinite monogenic semigroup is isomorphic to the additive semigroup
of positive integers.
(2) If P ≤ S, then the inclusion mapping ι : P → S, ι(x) = x, is a homomorphism,
and it is injective (but need not be surjective). Therefore ι is an embedding (of P into S).
In particular, the identity function ι : S → S, ι(x) = x, is always an automorphism,
the trivial automorphism, of S. ⊔
⊓
1.3 Homomorphisms and transformations 14
Proof. Exercise. ⊔
⊓
One may check that there are no other elements in this semigroup, and hence S has seven
elements, S = {α, β, α2 , β 2 , αβ, βα, βα2 }. ⊔
⊓
Representations
A homomorphism ϕ : S → TX is a representation of the semigroup S. We say that ϕ
is a faithful representation, if it is an embedding, ϕ : S ֒→ TX .
The following theorem states that semigroups can be thought of as subsemigroups
of the full transformation semigroups, that is, for each semigroup S there exists a set X
such that S ∼ = P ≤ TX for a subsemigroup P of transformations.
1.3 Homomorphisms and transformations 15
Proof. Let X = S 1 , that is, add the identity 1 to S if S is not a monoid. Consider the
full transformation semigroup T = TS 1 . For each x ∈ S define a mapping
ρx : S 1 → S 1 , ρx (y) = xy (y ∈ S 1 ) .
Thus ρx ∈ T , and for all x, y ∈ S and for all z ∈ S 1 ,
ρxy (z) = (xy)z = ρx (yz) = ρx (ρy (z)) = ρx ρy (z) ,
and hence ρxy = ρx ρy . Consequently, the mapping
ϕ : S → T, ϕ(x) = ρx
is a homomorphism. For injectivity we observe that
ϕ(x) = ϕ(y) =⇒ ρx = ρy =⇒ ρx (1) = ρy (1) =⇒ x = y .
⊔
⊓
In the previous proof the special element 1 is needed (only) to ensure injectivity of
ϕ. It is needed, because there exists a semigroup S (without identity) that has two left
identities x 6= y. In such a case, xz = yz for all z ∈ S.
Hence, loosely speaking (by Theorem 1.4) the theory of semigroups can be taught of
as the theory of transformations. This situation is reflected in a more restricted manner
in the theory of groups, where we have the following fundamental theorem of represen-
tations.
Theorem 1.5 (Cayley). Each group can be represented faithfully as a (semi)group of
permutations (bijections X → X).
Proof. Exercise. ⊔
⊓
Example 1.11. Let S = {e, a, b, c} (see the table). · e a b c
By the above definitions we obtain the following faithful rep- e e a b c
resentation ϕ : S → T{e,a,b,c} for S. Notice that the mappings a a e c b
ρx are just the rows of the multiplication table on the right. b b b b b
c c c c c
e a bc eabc
ϕ(e) = , ϕ(a) = ,
e a bc a e cb
eabc e ab c
ϕ(b) = , ϕ(c) = .
b b bb c c cc
Equivalences
Let X be a nonempty set. A subset δ ⊆ X × X is a relation on X. We shall write both
(x, y) ∈ δ and xδy to denote that the (ordered) pair (x, y) is in relation δ. Let B(X) be
the set of all relations on X. This set forms a monoid under the operation of composition:
Semilattice of idempotents
The idempotents ES of a semigroup can be given a partial order (if ES 6= ∅) as follows.
Define for e, f ∈ ES :
e ≤ f ⇐⇒ ef = e = f e .
e = ef = ef h = eh and e = f e = hf e = he ,
Lemma 1.10. Let S be a semilattice. Then ef ∈ S is the greatest lower bound of the
elements e and f of S.
g(ef ) = (ge)f = gf = g ,
and hence g ≤ ef . Hence ef ∈ S is the greatest lower bound of the elements e and f in
the semilattice S. ⊔
⊓
2
General Structure Results
2.1 Quotients
Congruences
An equivalence relation ρ on a semigroup S is a left congruence, if
x1 x1 · x2 y1 · y2
x2
y1
y2
Proof. Suppose first that ρ is a congruence. If x1 ρy1 and x2 ρy2 , then, by the definition,
(x1 x2 )ρ(x1 y2 ) and (x1 y2 )ρ(y1 y2 ) and hence also (x1 x2 )ρ(y1 y2 ) by transitivity. In the
converse direction the claim is trivial. ⊔
⊓
2.1 Quotients 19
Syntactic congruences
We say that a congruence ρ saturates a subset X ⊆ S, if X is a union of congruence
classes of ρ.
Lemma 2.3. For all X ⊆ S the relation ΓX is a congruence, called the syntactic con-
gruence of X in S. It is the largest congruence that saturates X.
Assume xρy holds, and let u, v ∈ S 1 be any elements. By Lemma 2.3, also (ux)ρ(uy),
and, by the same lemma, (uxv)ρ(uyv). Hence (uxv)ρ = (uyv)ρ, from which it follows
that uxv ∈ X if and only if uyv ∈ X, since ρ saturates X. Thus xΓ y, and so ρ ⊆ Γ as
required. ⊔
⊓
2.1 Quotients 20
Quotients
Let ρ be a congruence of S, and let
S/ρ = {xρ | x ∈ S}
be the set of all congruence classes of ρ. We define a new semigroup, the quotient
semigroup (of S modulo ρ), on the domain S/ρ by
xρ · yρ = (xy)ρ .
This is a well defined binary operation by Lemma 2.1, and its associativity is easily
checked:
The quotient S/ρ is thus obtained from S by contracting each congruence class xρ
into a single element.
b =•yi ρ
y1
x2 z1
z2 a = xi ρ• • c = zi ρ
x1 z3
u1 u2 •
d = ui ρ
S S/ρ
Example 2.1. (1) Consider the finite semigroup S defined in the leftmost table. Here e
and f are idempotents. Indeed, e is an identity of S.
· e a f b
e e a f b · c d g
a a e b f c c d g
f f b f b d d c g
b b f b f g g g g
Let ρ be the relation, where apart from the relations xρx (x ∈ S), we have only
f ρb and bρf . Then ρ is a congruence of S, and its congruence classes (with a short
hand notations) are: c = {e}, d = {a} and g = {f, b}. The multiplication table for the
quotient semigroup S/ρ is given in the second table.
Also, the symmetric relation ρ′ (with ιS ⊆ ρ′ ) so that eρ′ a and f ρ′ b, is a congruence.
It has only two congruence classes, {e, a} and {f, b}, and hence the quotient S/ρ′ is a
semigroup having two elements.
2.2 Homomorphism theorem 21
The symmetric relation ρno (with ιS ⊆ ρno ) so that aρno b is not a congruence, be-
cause a · a = e and a · b = f in S, but eρno f does not hold. In this case, ρno does not
respect the product of S: aρno b holds, but (aa)ρno (ab) does not hold.
(2) Consider the semigroup S = (Z, +). If ρ is a congruence on S, then nρm implies
that (n + k)ρ(m + k) holds for all integers k, by the definition of a congruence. Suppose
that k is the smallest non-negative integer such that nρ(n + k) for some n ∈ Z. In
particular, (n − n)ρ(n + k − n), i.e., 0ρk. Denote by m the remainder of m when divided
by k: 0 ≤ m ≤ m and m ≡ m (mod k). By above mρm. Also the converse holds, and
thus the congruences of (Z, +) are exactly the ordinary number theoretic congruences,
ρ equals mod k (k ≥ 0). ⊔
⊓
We prove now that the congruences of a semigroup S are closed under intersections.
Proof. Denote ρ = ∩i∈I ρi for the congruences ρi (i ∈ I). Assume xρy and z ∈ S,
then also xρi y for all i ∈ I, and hence (zx)ρi (zy) and (xz)ρi (yz), which implies that
(zx)ρ(zy) and (xz)ρ(yz). Therefore ρ is a congruence of S.
For the other claim we observe first that δc is a congruence, since it is an intersection
of congruences. If ρ is any congruence such that δ ⊆ ρ, then ρ takes part in the intersec-
tion, and thus δc ⊆ ρ. The claim follows from this. ⊔
⊓
Natural homomorphisms
For a congruence ρ, define a mapping ρ♯ : S → S/ρ as follows
ρ♯ (x) = xρ .
One could also set ker(α) = αα−1 , where α−1 (y) = {x ∈ S | α(x) = y}, and αα−1
is thought as a product of relations. (Functions are read from right to left, and relations
from left to right; this causes some complications while writing compositions).
Proof. Indeed,
⊔
⊓
Corollary 2.1. Every congruence is a kernel of some homomorphism.
Example 2.2. Often homomorphisms are much easier to handle that congruences. Con-
sider S = (N, +). If α : S → P is a homomorphism, then
In particular, α(0) = 0. This implies that α depends only on 1. (Of course, this follows
already from the fact that 1 generates (N, +)). It is now easy to show that ker(α) equals
mod k for some k. ⊔
⊓
Homomorphism theorem
Theorem 2.2. Let α : S → P be a any homomorphism. There exists a unique embedding
β : S/ ker(α) ֒→ P such that the following diagram commutes.
α -P
S
ker(α)♯ β
R
S/ ker(α)
and so β(xρ) has a specific value (in P ), which is independent of the choice of the
representative of the congruence class xρ.
Secondly, β is a homomorphism, since
Proof. The proof is quite similar to the previous one. Here we observe that the mapping
β defined by β(xρ) = α(x) is well defined, since
since ρ ⊆ ker(α). ⊔
⊓
The homomorphism theorem, as well as the next isomorphism theorem, are standard
(universal) algebraic results, that is, they hold in all algebras (groups, rings, Boolean
algebras, and so on).
Below α(S) is the image semigroup of S in P . In particular, α(S) ≤ P , by
Lemma 1.1.
α(S) ∼
= S/ ker(α) .
2.3 Ideals and Rees congruences 24
Ideals
A nonempty subset I of a semigroup S is
• a left ideal, if SI ⊆ I;
• a right ideal, if IS ⊆ I;
• an ideal, if it is both a left and a right ideal.
Proof. Exercises. ⊔
⊓
Lemma 2.9. If I and J are left (right) ideals of a semigroup S with I ∩ J 6= ∅, then
I ∩ J is a left (right) ideal.
Proof. Let a ∈ I ∩ J and x ∈ S. Then sa ∈ I and sa ∈ J, since I and J are ideals, and
so sa ∈ I ∩ J. Hence (I ∩ J)S ⊆ (I ∩ J). ⊔
⊓
Rees congruences
Let I be an ideal of S. Define the Rees congruence (with respect to I) as follows:
ρI = I × I ∪ ι .
Hence xρI y holds if and only if either x, y ∈ I or x = y. This means that ρI contracts
the ideal I and leaves the rest of S as it was.
Proof. Exercise. ⊔
⊓
2.3 Ideals and Rees congruences 25
The quotient semigroup S/ρI will be denoted simply S/I, and it is called the Rees
quotient (of the ideal I). S/I has the elements I and {x} for x ∈ S \ I. In order to
simplify notation we identify the element {x} (= xρI ) with the element x ∈ S \ I. The
product of elements in S/I is as follows: x · y = xy for x, y ∈
/ I, and Ix = I = xI for
all x. Therefore I is a zero element of the semigroup S/I.
Minimal ideals
An ideal I of S is said to be minimal, if for all ideals J of S, J ⊆ I implies that J = I.
By the above
Every finite semigroup S does have a minimal ideal. Indeed, consider an ideal I,
which has the least number of elements. Such an ideal exists because S is finite and S is
its own ideal.
Example 2.3. The ideals of the semigroup (N, +) are exactly the sets n + N = {n + k |
k ≥ 0} (Exercise). Further, m + N ⊆ n + N if and only if m ≥ n. Therefore (N, +) has
no minimal ideals. ⊔
⊓
Simple semigroups
A semigroup S is said to be simple, if it has no ideals I 6= S.
Proof. Clearly, for any x ∈ S, SxS is an ideal of S, and hence if S is simple, then
S = SxS for all x ∈ S.
In converse, assume S = SxS for all x ∈ S. Let I is an ideal of S and let x ∈ I.
Hence S = SxS ⊆ I, and so I = S, from which it follows that S is simple. ⊔
⊓
Example 2.4. By the above, a semigroup S is simple if and only if for all s, r ∈ S the
equation xsy = r has a solution in S. Using this one can show that the semigroup of all
2-by-2-matrices
x 0
(x, y ∈ Q with x, y > 0)
y 1
is a simple semigroup. This is an exercise. ⊔
⊓
3
Free Semigroups and Presentations
Free semigroups
Let S be a semigroup. A subset X ⊆ S generates S freely, if S = [X]S and every
mapping α0 : X → P (where P is any semigroup) can be extended to a homomorphism
α : S → P such that α↾X = α0 . Here we say that α is a homomorphic extension of
the mapping α0 . If S is freely generated by some subset, then S is a free semigroup.
Example 3.2. (1) (N+ , +) is free, for X = {1} generates it freely: Let α0 : X → P be a
homomorphism, and define α : N+ → P by α(n) = α0 (1)n . Now, α↾X = α0 , and α is
a homomorphism: α(n + m) = α0 (1)n+m = α0 (1)n · α0 (1)m = α(n) · α(m).
(2) On the other hand, (N+ , ·) is not free. For, suppose X ⊆ N+ , choose P =
(N+ , +), and let α0 (n) = n for all n ∈ X. If α : (N+ , ·) → P is any homomorphism,
then α(n) = α(1 · n) = α(1) + α(n), and thus α(1) = 0 ∈ / P . So certainly no α can be
an extension of α0 .
(3) Let S = {a, a2 } be a monogenic semigroup with index 2 and period 1, that is,
a3 = a2 . If X generates S, then necessarily a ∈ X (since a is not a product of two
elements of S). Let α0 : S → P = (N+ , +) be such that α0 (a) = 1. If α : S → P is
an extension of α0 , then α(aa) = α(a) + α(a) = α0 (a) + α0 (a) = 2 and similarly
α(aaa) = 3. However, aa = aaa in S, and this is a contradiction. Thus S is not free.
⊔
⊓
3.1 Free semigroups 27
Theorem 3.1. For any alphabet A, A+ is a free semigroup, and it is freely generated
by A.
Proof. Clearly, A generates A+ , since all words are strings of letters from A. Let then
S be any semigroup, and α0 : A → S any mapping. We define α : A+ → S by
Proof. By the definition, each α0 has an extension. For the uniqueness suppose both
α : S → P and β : S → P are homomorphic extensions of α0 . Now, for all x ∈ S,
x = x1 x2 . . . xn for some xi ∈ X, since X generates S. Therefore,
and so α = β. ⊔
⊓
The next result states that every semigroup S is a homomorphic image of a word
semigroup, and therefore the word semigroups are basic semigroups wherefrom all other
semigroups can be derived.
Two sets X and Y have the same cardinality, denoted |X| = |Y |, if there is a
bijection α : X → Y . In this case the function α−1 : Y → X is also a bijection, and so
there is a 1-to-1 correspondence between the elements of X and Y . If X is finite, then
|X| = |Y | if and only if X and Y have exactly the same number of elements. However,
for infinite sets we cannot ‘count’ the number of elements. Indeed, |N| = |Q|, although
N ⊂ Q properly.
Theorem 3.3. For each semigroup S there exists an alphabet A and an epimorphism
ψ : A+ ։ S.
Proof. Let X be any generating set of S; you may even choose X = S. Let A be an
alphabet with |A| = |X|, and let ψ0 : A → X be a bijection. By definition of free-
ness, ψ0 has a homomorphic extension ψ : A+ → S. This extension is surjective, since
[ψ(X)]S = ψ([X]S ) = ψ(S), for X generates S. ⊔
⊓
S∼
= A+ / ker(ψ)
There are other free semigroups than the word semigroups, but all of them are iso-
morphic to word semigroups.
Corollary 3.3. If S and R are free semigroups generated by X and Y , respectively, such
that |X| = |Y |, then S ∼
= R.
Proof. We observe first that the claim holds for the word semigroups A+ , for which A
is the only minimal generating set. Let A be an alphabet such that |A| = |X| and let
α0 : X → A be a bijection.
Suppose X generates S freely, and that there is an x ∈ S, for which
x = x1 x2 . . . xn = y 1 y 2 . . . y m (xi , yi ∈ X) .
For the homomorphic extension α of α0 ,
α(x) = α0 (x1 )α0 (x2 ) . . . α0 (xn ) = α0 (y1 )α0 (y2 ) . . . α0 (ym )
in A+ . Since A+ satisfies the condition of the theorem, and α0 (xi ), α(yi ) are letters
for each i, we must have that α0 (xi ) = α0 (yi ) for all i = 1, 2, . . . , n (and m = n).
Moreover, α0 is injective, and so xi = yi for all i, and thus also S satisfies the claim.
Suppose then that S satisfies the uniqueness condition. Denote β0 = α−1 0 and let
β : A+ → S be the homomorphic extension of β0 . Now, β is surjective, since X gener-
ates S. It is also injective, because if β(u) = β(v) for some words u 6= v ∈ A+ , then
β(u) would have two different factorizations over X. Hence β is an isomorphism, and
the claim follows from this. ⊔
⊓
Example 3.3. (1) Let A = {a, b, c} be an alphabet. The words ab, bab, ba generate a
subsemigroup of A+ . This semigroup S = [ab, bab, ba] is not free, since the element
w = babab has two different factorizations in S: w = ba · bab = bab · ab.
(2) If A is as above, but S = [aab, ab, aa], then S is a free subsemigroup of A+ .
Indeed, if there were two different factorizations w = u1 u2 . . . un = v1 v2 . . . vm of a
word w ∈ S, then either u1 = v1 (and, by cancellation, there would be a shorter word
u2 . . . un = v2 . . . vm with two different factorizations) or u1 = aa and v1 = aab (or
symmetrically, u1 = aab and v1 = aa). But in this case u2 cannot be found, because it
should start with the letter b.
(3) Let then A = {a, b} be a binary alphabet, and consider
S = [abn | n ≥ 1] = [ab, ab2 , ab3 , . . . ] .
Clearly, Base(S) = S \ S 2 = {abn | n ≥ 1}, and S is freely generated by Base(S).
(4) For S = (R, ·), we have Base(S) = ∅. In particular, this semigroup cannot be
free. ⊔
⊓
3.1 Free semigroups 30
Free monoids
Observe that no monoid M can be a free semigroup, because 1 = 1·1. Hence the identity
1 cannot be among the free generator of M . It follows that 1 is a product of generators
of M , 1 = x1 x2 . . . xn , but now x = 1 · x gives two factorizations of an element x.
We say that M is a free monoid (freely generated by a subset X with 1 ∈ / X),
if X ∪ {1} is a generator set for M , and every mapping α0 : X → P (where P is a
monoid) extends to a monoid homomorphism α : M → P such that α↾X = α0 and
α(1M ) = 1P .
Again, if M is free and α0 : X → P a mapping to a monoid P , then its extension
α : M → P is unique.
We have immediately:
Corollary 3.5. The word monoid A∗ is a free monoid for all alphabets A.
Theorem 3.8. A monoid M is a free monoid if and only if M \{1M } is a free semigroup.
Proof. Suppose first that M is a free monoid, freely generated by X. Then M \ {1M }
is a subsemigroup of M , for otherwise, 1M = x1 x2 . . . xn for some xi ∈ X, and a
mapping α0 : X → A∗ with α0 (xi ) ∈ A does not extend. The rest of the claim follows
from the definition of a free semigroup.
The converse claim is an exercise. ⊔
⊓
The other results for free semigroups can also be modified for free monoids:
This is a monoid, called a special linear monoid, with the identity matrix I as its iden-
tity. Note that if det(M1 ) = 1 = det(M2 ), then also det(M1 M2 ) = 1.
3.2 Presentations of semigroups 31
with integer entries. Note that these matrices are not in SL(2, N).
Let M ∈ SL(2, N) be any matrix as above with M 6= I. Then
−1 n11 − n21 n12 − n22
A M=
n21 n22
−1 n11 n12
B M= .
n21 − n11 n22 − n12
Either one of these is in SL(2, N): A−1 M ∈ SL(2, N), if n11 ≥ n21 ; and B −1 M ∈
SL(2, N), if n11 < n21 .
Moreover, the trace of M , tr(M ) = n11 + n22 , is greater than the trace of A−1 M or
B −1 M whichever is in SL(2, N):
S∼
= A+ / ker(ψ) ,
S = ha1 , a2 , · · · | ui = vi (i ∈ I)i or S = hA | Ri ,
if ker(ψ) is the smallest congruence of A+ that contains the relation {(ui , vi ) | i ∈ I}.
In particular,
ψ(ui ) = ψ(vi ) for all ui = vi in R . (3.1)
The set R of relations is supposed to be symmetric, that is, if u = v is in R, then v = u
also holds.
One should remember that the words w ∈ A+ are not elements of S but of the word
semigroup A+ , which is mapped onto S. We say that a word w ∈ A+ presents the
element ψ(w) of S. The same element can be presented by many different words. In
fact, if ψ(u) = ψ(v), then both u and v present the same element of S.
Let S = hA | Ri be a presentation. We show that S satisfies a relation u = v (that
is, ψ(u) = ψ(v)) if and only if there exists a finite sequence u = u1 , u2 , . . . , uk+1 = v
of words such that ui+1 is obtained from ui by substituting a factor ui by vi for some
ui = vi in R.
To be more precise, we say that a word v is directly derivable from the word u, if
ψ(u) = ψ(w1 u′ w2 ) = ψ(w1 )ψ(u′ )ψ(w2 ) = ψ(w1 )ψ(v ′ )ψ(w2 ) = ψ(w1 v ′ w2 ) = ψ(v) ,
u ≡ u1 = u2 = · · · = uk ≡ v .
In this presentation there are two generators and three relations. For instance, S satisfies
the relation baabbaa = bbaaaba, since u1 = baabbaa ≡ b · aab · baa = b · ba · baa ≡ u2
and u2 = bbabaa ≡ bba · ba · a = bba · aab · a ≡ bbaaaba. Also, aaab = aabb = abbb =
aaba = baa = bab in S and hence aaab = bab in S.
(2) A word semigroup does not need any relations: A+ = hA | ∅i.
(3) The presentation
Monoid presentations
All monoids have a semigroup presentation, but it is more convenient to use monoid
presentations for these in order to take advantage of the identity element:
M = ha1 , a2 , · · · | ui = vi (i ∈ I)i
for some k, m ≥ 0. The monoid M is a free commutative monoid, and it can be shown
that every commutative monoid, which is generated by two elements, is an epimorphic
image of M .
(2) The (monoid) presentation
ha, b | aba = 1i
defines a group. In fact, this group is isomorphic to (Z, +). Indeed, let M be a monoid
with the above presentation, i.e., M ∼ = A∗ /Rc , where A = {a, b} and R = {aba = 1},
∗
and let ψ : A ։ M be the corresponding epimorphism. Then M is generated by the
elements x = ψ(a) and y = ψ(b). Furthermore, ab = ab · aba ≡ aba · ba = ba,
and hence xy = yx. It follows that M is a commutative monoid. Consequently, every
element z ∈ M has the form z = ψ(an bm ) for some n, m ≥ 0. Also, a · ba = 1 and
ba · a = ab · a = 1 means that ba is a group inverse of a. Similarly, a2 is a group inverse
of b, b = (aa)−1 . This means that all elements of M have the form z = ψ(ak ) for k ∈ Z.
It is now easy to show that α : M → (Z, +) defined by α(a) = −1 and α(b) = 2 is an
isomorphism.
(3) Define two mappings α, β : N → N as follows:
(
0 if n = 0
α(n) = and β(n) = n + 1 (n ≥ 0) .
n − 1 if n ≥ 1
B = {β k αm | k, m ≥ 0} .
Furthermore, these elements are all different from each other: if γ = β k αm and δ =
β r αs , then
γ(n) = β k αm (n) = β k (n − m) = n + k − m ,
δ(n) = β r αs (n) = β r (n − s) = n + r − s .
B = ha, b | ab = 1i
is a (monoid) presentation of B.
The bicyclic monoid has many (semigroup) presentations, of which we mention
⊔
⊓
We shall use the fact that each semigroup is isomorphic to a subsemigroup of the full
transformation semigroup, and we actually prove
Proof. Define
Xn = {2n (2m − 1) − 1 | m = 1, 2, . . . } .
Now, Xn ∩ Xm = ∅ for all n 6= m, because k ∈ Xn if and only if n is the highest power
for which 2n divides k + 1. Also,
[
Xn = 2N + 1 ,
n≥1
and hence the sets Xn form a partition of the odd positive integers. Define
(
n−1 if n is even,
β1 (n) = 2n (n ≥ 1) , β2 (n) = −(k+1) −1
αk (2 (n + 1) + 2 ) if n ∈ Xk .
β22 β1k β2 β1 (n) = β22 β1k β2 (2n) = β22 β1k (2n − 1) = β22 (2k (2n − 1)) =
β2 (2k (2n − 1) − 1) = αk (2−(k+1) (2k (2n − 1) + 2−1 ) =
αk (2−1 (2n − 1 + 1) = αk (n) .
α(ai ) = abi (i = 1, 2, . . . , n)
Theorem 3.15. Each finitely generated word monoid A∗ can be embedded in a word
monoid B ∗ generated by two letters.
4
Combinatorial Topics
X + = [X]A∗
for the subsemigroup that X generates in the word monoid A∗ , that is, X + consists of
all catenations of the words in X. Also, the corresponding monoid is denoted by X ∗ ,
that is, X ∗ = X + ∪ {1}. Notice that if 1 ∈ X already, then X ∗ = X + .
If w ∈ A∗ is a word, then we write simply w∗ = {w}∗ .
Let u, v ∈ A+ be two words. Henceforth we write u = v (not u ≡ v) for the equality
in A∗ .
Then u is
• a factor of v, if v = w1 uw2 for some words w1 , w2 ∈ A∗ ;
• a prefix of v, if v = uw for some w ∈ A∗ ;
• a suffix of v, if v = wu for some w ∈ A∗ .
The length |w| of a word w ∈ A∗ is the number of the occurrences of the letters in
it: if w = a1 a2 . . . an with ai ∈ A, then |w| = n. The length of the empty word is zero.
Clearly, if w = uv, then |w| = |u| + |v|, and so the length function is a homomorphism
| | : A∗ → (N, +).
The next lemma is immediate.
We first prove another criterion for freeness of submonoids of a word monoid. For
a submonoid M of a word monoid A∗ we let M+ = M \ {1}. Here M+ is always
a subsemigroup of A+ , because the empty word cannot be a catenation of nonempty
words. The base of M , Base(M ) = M+ \ M+2 , is the base of the semigroup M+ . Recall
that the identity (the empty word) is automatically in a submonoid. The base of a free
monoid is called a code.
A submonoid M of a word monoid A∗ need not be free. As an example, consider the
monoid M = {a, ab, ba}∗ .
4.1 The defect theorem 38
Lemma 4.2. For all submonoids M of A∗ the base Base(M ) is the unique minimal
generating set of M (as a monoid), i.e., if a subset N generates M , then Base(M ) ⊆ N .
Proof. In order to show that Base(M ) generates M suppose to the contrary that there
exists a word w ∈ M that cannot be written as a product of words from Base(M ), and
let w is a shortest word such that w ∈ M+ but w ∈ / Base(M )+ . Since w ∈ / Base(M ),
there are two words u, v ∈ M+ such that w = uv. By the minimality of w, the shorter
words u and v are in Base(M )+ , but now also w ∈ Base(M )+ , which is a contradiction.
If a subset N ⊆ M generates M , then for all u ∈ Base(M ), u ∈ N ∗ (= M ), but u
is not a product of two or more words from M , and hence necessarily u ∈ N , too. ⊔
⊓
u, v, uw, wv ∈ M =⇒ w ∈ M .
and since M is freely generated by X, k+r = t+s and ui = vi for all i = 1, 2, . . . , k+r.
It follows that w = uk+1 . . . ut (since k ≤ t), and so w ∈ M as required.
Assume then that the condition is valid for M . Suppose that there exists a word,
which has two different factorizations over X,
u1 u2 . . . un = v1 v2 . . . vm (ui , vi ∈ X) .
u1 w ∈ M and u2 u3 . . . un = wv2 v3 . . . vm .
is a free submonoid. Clearly, it is the smallest free submonoid that contains the subset
X. The base of this intersection is denoted simply by Hull(X), and it is called the free
hull of X. In particular, X ∗ ⊆ Hull(X)∗ , since X ∗ is a monoid and X ⊆ Hull(X)∗ .
Theorem 4.3 (Defect theorem). Let X ⊆ A+ be a finite set of words. If X is not a code
(a base of a free monoid), then | Hull(X)| ≤ |X| − 1.
w = u1 u2 . . . un = v1 v2 . . . vm (ui , vi ∈ X) ,
Y = (Hull(X) \ {w})w∗ .
w1 w2 . . . wm = v1 v2 . . . vr for some wi , vi ∈ Y ,
Example 4.1. Let A = {a, b} and X = {ab, aba, bab, ba}. The submonoid X ∗ is not
free, since ab·aba = aba·ba. Consider any free submonoid M such that X ∗ ⊆ M . Now,
for u = ab, v = ba and w = a, u, v, uw, wv ∈ X ∗ ⊆ M , and hence by Theorem 4.1,
also w = a ∈ M . Similarly, b ∈ M , and hence M = A∗ , which means that Hull(X) =
{a, b}. Here |X| = 4 > 2 = | Hull(X)| as required by Theorem 4.3. ⊔
⊓
Some corollaries
A word w ∈ A+ is said to be primitive, if it is not a proper power of another word, that
is, w = uk implies that k = 1 and u = w.
Two words u, v ∈ A+ are conjugates (of each other), if there are words w1 , w2 ∈ A∗
such that u = w1 w2 and v = w2 w1 . Hence, if u is a conjugate of v, then u is obtained
from v by cyclically permuting the word v.
Corollary 4.2. Two words u, v ∈ A∗ commute, uv = vu, if and only if they are powers
of a common word.
Proof. If uv = vu, then X = {u, v} is not a code, and hence again | Hull(X)| < |X| =
2, and so u and v are powers of a common word. ⊔
⊓
Equality sets
Let α, β : A∗ → B ∗ be two homomorphisms between the word monoids A∗ and B ∗ ,
where we assume that the alphabets are finite. Their equality set is defined to be
4.2 Ehrenfeucht’s conjecture 41
Thus the equality set of α and β consists of all those words on which these homomor-
phisms agree.
A word w ∈ E(α, β) is also called a solution of (α, β). A solution w 6= 1 is a
nontrivial solution. Notice that we always have 1 ∈ E(α, β). The empty word is the
trivial solution of (α, β).
v, uv ∈ E(α, β) =⇒ u ∈ E(α, β) .
Proof. First of all the empty word is always in E = E(α, β), since α(1) = 1 = β(1).
We use a restricted version of Theorem 4.1 to show the rest of the claim. Suppose
v, uv ∈ E. Then α(uv) = α(u)α(v) = α(u)β(v) and α(uv) = β(uv) = β(u)β(v),
and hence α(u) = β(u), because A∗ is cancellative. This means that u ∈ E, too. Now,
if u, v, uw, wv ∈ E, then v, wv ∈ E and, by above, w ∈ W . Thus E is free. ⊔
⊓
The base Base(E(α, β)) is the set of all minimal solutions of (α, β). By above, it is
a code.
Example 4.2. (1) Consider the endomorphisms α, β : A∗ → A∗ , where A = {a, b}, and
α and β are defined by the images of the generators a and b:
Now, if w ∈ E = E(α, β) is nonempty, then it must begin with the letter a (because
α(b) and β(b) are ‘incomparable’). The next letter must be b, and, we obtain thus a
solution ab ∈ E: α(ab) = aba · ba = uivab · aba = β(ab). It is easy to show that
E = (ab)∗ = {(ab)n | n ≥ 0}.
(2) Let then α, β : A∗ → A∗ for A = {a, b, c} be defined by
Remark 4.1. It can be shown that the problem whether E(α, β) contains a nonempty
word is algorithmically unsolvable, that is, there exists no algorithm which decides
whether or not E(α, β) = {1} on inputs α and β. This is know as the Post correspon-
dence problem. There are thus arbitrarily complex equality sets. ⊔
⊓
4.2 Ehrenfeucht’s conjecture 42
Test sets
We are going to prove the following result, which was conjectured by Ehrenfeucht at the
beginning of 1970s and proved by Albert and Lawrence and independently by Guba in
1985.
Theorem 4.4 (Ehrenfeucht’s conjecture). Let L ⊆ A∗ be any set of words. There exists
a finite subset T ⊆ L such that for all homomorphisms α, β : A∗ → B ∗
α↾T = β↾T ⇐⇒ α↾L = β↾L .
The subset T of L is called a test set of L, since in order to check whether any two
homomorphisms agree on the set L, it suffices to check whether they agree on the finite
subset T .
Theorem 4.4 may be restated in the following form:
Theorem 4.5. Let L ⊆ A∗ be any set of words. There exists a finite subset T ⊆ L such
that for all homomorphisms α, β : A∗ → B ∗
T ⊆ E(α, β) ⇐⇒ L ⊆ E(α, β) .
Systems of equations
Let X = {x1 , x2 , . . . , xn } be a special alphabet, the letters of which we call variables.
An equation over X is a pair of words (u, v) (for u, v ∈ X + ). We shall usually write
u = v instead of (u, v).
We say that (w1 , w2 , . . . , wn ) with wi ∈ A∗ is a solution of the equation u = v,
if the substitution xi 7→ wi in u and v results in the same word of A∗ . In other words,
a solution is a (monoid) homomorphism α : X ∗ → A∗ such that α(u) = α(v) (here
α(xi ) = wi ), that is, (u, v) ∈ ker(α).
Example 4.3. Let X = {x, y, z}. Then a solution of the equation xyx = zyxz is a
homomorphism α : X ∗ → A∗ (A is some alphabet) such that α(xyx) = α(zyxz), that
is, w1 w2 w1 = w3 w2 w1 w3 , where w1 = α(x), w2 = α(y) and w3 = α(z). Of course,
w1 = w2 = w3 = 1 is a solution. This is the trivial solution.
Let x 7→ w1 , y 7→ w2 , z 7→ w3 be any solution. By comparing the lengths we
obtain |w1 w2 w1 | = |w3 w2 w1 w3 |, that is, 2|w1 | + |w2 | = |w1 | + |w2 | + 2|w3 |, and so
|w1 | = 2|w3 |, which means that w1 = w3 w for a word w with |w| = |w3 | (since w1 and
w3 begin the same word). When we substitute w1 by w3 w, we obtain that w3 ww2 w3 w =
w3 w2 w3 ww3 , from which we get ww2 w3 w = w2 w3 ww3 by cancellation. Here w and
w3 end this word, and hence w = w3 , since |w| = |w3 |. Therefore, w3 w2 w3 w3 =
w2 w3 w3 w3 , and so w3 w2 = w2 w3 , which means that w2 and w3 are powers of the same
word, w2 = ur and w3 = us , and w1 = w3 w = w3 w3 = u2s . Further, if u is any word,
then w1 = u2s , w2 = ur , w3 = us is a solution for any r, s ≥ 0. We have thus found all
the solutions of the equation xyx = zyxz. ⊔
⊓
4.2 Ehrenfeucht’s conjecture 43
Example 4.4. Let X = {x1 , x2 , y1 , y2 } and consider the system of equations xi1 y1i =
xi2 y2i (i ≥ 1):
x1 y 1 = x2 y 2
x1 x1 y 1 y 1 = x2 x2 y 2 y 2 .
..
.
This system of equations has the trivial solution α(xi ) = 1 = α(yi ) (i = 1, 2), and
the solutions α(x1 ) = w1 = α(x2 ), α(y1 ) = w2 = α(y2 ) (in any alphabet); also
α(x1 ) = aa, α(x2 ) = a, α(y1 ) = a, α(y2 ) = aa is a solution. ⊔
⊓
Two systems E1 and E2 of equations are equivalent, if they have the same solutions.
The following was proved by Culik and Karhumäki:
Proof. For any alphabet B we let B = {a | a ∈ B} be a new alphabet, and write for
+
all words u = a1 a2 . . . ak ∈ B + , u = a1 a2 . . . ak ∈ B . Hence the function is an
∗
isomorphism B ∗ → B .
Let us first assume that Ehrenfeucht’s conjecture is true, and let E be a system of
equations in variables X = {x1 , x2 , . . . , xn }. Let X be a new set of variables defined as
+
above. To each equation u = v in E we attach a word uv ∈ X + · X . Let
L = {uv | u = v ∈ E} ⊆ (X ∪ X)∗ .
T = {u = v | uv ∈ T } ,
then α(u) = α(v) for all u = v in T, and α1 (uv) = α2 (vv) for all uv in T , and so
α1 (uv) = α2 (vv) for all uv in L, and α(u) = α(v) for all u = v in E. This proves that
E is equivalent to the finite subsystem T.
4.2 Ehrenfeucht’s conjecture 44
Assume then that each system of equations has an equivalent finite subsystem. Let
L ⊆ B ∗ be any set of words. Let B be a new alphabet as above. We form a system of
equations E from L as follows:
E = {u = u | u ∈ L} .
By assumption this system has an equivalent subsystem T. Let
T = {u | u = u ∈ T} .
Hence T is a finite subset of X. Let α, β : B ∗ → A∗ be homomorphisms such that
α(u) = β(u) for all u ∈ T . Let X = B ∪ B be our set of variables, and define a
homomorphism γ : X ∗ → A∗ : For all x ∈ B,
γ(x) = α(x) and γ(x) = β(x) .
Now, γ(u) = γ(u) for all u ∈ T , and hence γ is a solution to the finite system T. By
assumption, γ is then a solution to the whole system E, and, consequently, α(u) = β(u)
for all u ∈ L. This shows that T is a finite test set of L. ⊔
⊓
is an isomorphism.
Let X = {x1 , x2 , . . . , xk } be a set of word variables. We introduce for each xi four
new integer variables xi1 , xi2 , xi3 and xi4 , and denote
X̄ = {xij | 1 ≤ i ≤ k, 1 ≤ j ≤ 4} .
and let M(X̄) be the submonoid of the matrix monoid Z[X̄]2×2 generated by the matri-
ces X1 , X2 , . . . , Xk .
4.2 Ehrenfeucht’s conjecture 46
Proof. Consider two elements M = Xr1 Xr2 . . . Xrm and M ′ = Xs1 Xs2 . . . Xst of
M(X̄). The upper right corner entry M12 of M contains the monomial
for j = 1, 2, . . . , m. It is now easy to see that if these monomials exist in (M ′ )12 , then
t = m and Xrj = Xsj for j = 1, 2, . . . , m. Therefore, (M )12 = (M ′ )12 implies that
Xr1 = Xs1 , Xr2 = Xs2 , . . . , Xrm = Xsm . The claim follows from this. ⊔
⊓
⊔
⊓
is an isomorphism.
The following lemma is now immediate.
ϕ µ
? ?
α - SL(2, N)
M(X)
We shall now prove the main result of this section, from which Theorem 4.4.
Theorem 4.8. Each system of equations over a finite set of variables has an equivalent
finite subsystem.
Proof. Let X and X̄ be as above. Since, by Theorem 3.15, each monoid B ∗ can be
embedded into a free monoid generated by two elements, we can restrict our solutions
α : X ∗ → B ∗ to the case where B is binary, say B = {a, b}.
For a word w ∈ X ∗ we denote
Pw1 Pw2
ϕ(w) = ,
Pw3 Pw4
Now, α′ is a solution of the equations on the second line of the definition of P(u, v).
This is because α is into SL(2, N). Further, α′ extends in a unique way to a ring homo-
morphism α′ : Z[X̄] → Z, for which
′
α (Pw1 ) α′ (Pw2 )
αϕ(w) =
α′ (Pw3 ) α′ (Pw4 )
for all w ∈ X ∗ . Therefore by (4.4), α(u) = α(v) if and only if α′ is a solution of the
whole P(u, v).
Let now E = {ui = vi | i ∈ I} be a system of equations, and let P = ∪i∈I P(ui , vi ).
By Hilbert’s Basis Theorem, P has an equivalent finite subsystem P0 = ∪i∈I0 P(ui , vi ).
Consider the subsystem E0 = {ui = vi | i ∈ I0 } of E. By above,
α is a solution to E0 ⇐⇒ α′ is a solution to P0
⇐⇒ α′ is a solution to P ⇐⇒ α is a solution to E ,
where u′ and v ′ are obtained from u and v, resp., by substituting xa for a ∈ A. Therefore
each system of equations with constants A can be presented as a system of equations
over the variables X ∪ {xa | a ∈ A} together with a finite number of equations xa = a
containing a unique constant. Therefore Theorem 4.8 can be generalized:
4.2 Ehrenfeucht’s conjecture 48
Corollary 4.4. Each system of equation with constants is equivalent to a finite subsys-
tem.
Theorem 4.9. Each finite set of equations with constants is equivalent to a single equa-
tion with constants.
Note, however, that the single equation of Theorem 4.9 need not be among the origi-
nal ones.
Example 4.5. (1) Let F in(A∗ ) denote the monoid of all nonempty finite subsets of the
word monoid A∗ . It was shown by Lawrence that the monoid F in(A∗ ) does not satisfy
CP even when A is a binary alphabet. Indeed, in this case the system E of equations
x1 xi2 x1 = x1 xi3 x1 (i ≥ 1)
over three variables does not have an equivalent finite subsystem in F in(A∗ ).
(2) Recall that the bicyclic monoid B is a 2-generator and 1-relator semigroup with
a presentation ha, b | ab = 1i. The monoid B is isomorphic to the monoid generated by
the functions α, β : N → N:
xi1 xi2 x3 = x3 (i = 1, 2, . . . ) .
As is easily seen the homomorphism δj defined by δj (x1 ) = β, δj (x2 ) = α and
δj (x3 ) = γj , is a solution of xi1 xi2 x3 = x3 for all i ≤ j, but δj is not a solution of
xj+1 j+1
1 x2 x3 = x3 . Hence the system E does not have an equivalent finite subsystem,
and therefore the bicyclic monoid B does not satisfy CP. ⊔
⊓
Dominions
Let now S be any semigroup. An element x ∈ S is said to be dominated by a subset
U ⊆ S, if α(x) = β(x) for all homomorphisms α, β : S → P (to any semigroup P )
whenever α and β agree on U , α|U = β|U . The dominion of a subset U in a semigroup
S is the set of all elements of S that are dominated by U :
Zig-zag theorem
A rather neat condition for Dom(U, S) was given by Isbell in 1966 using zig-zags: Let
x ∈ S. A sequence
Lemma 4.7. If two elements x and y in S have the same spine u0 , u1 , . . . , u2m , then
x = y.
Proof. If
x = u0 y1 = x1 u1 y1 = · · · = xm u2m−1 ym = xm u2m ,
y = u0 v1 = z1 u1 v1 = · · · = zm u2m−1 vm = zm u2m ,
y = x1 u1 v1 = · · · = xm u2m−1 vm = xm u2m = x ,
We prove the following theorem only in one direction, the proof in the opposite direc-
tion requires either new methods (topological as in the original Isbell’s proof (1966), or
tensor products as in the proofs by Stenström (1971), Howie (1976) and Storrer (1976),
or semigroup diagrams and HNN extensions as in Jackson’s proof (1991)), or exten-
sive technics in combinatorics of presentations (as in Higgins’ proof (1991) based on
Jackson’s proof).
Proof. (1) Assume first that there exists a zig-zag (4.5) of x, and let α, β : S → P be
two homomorphisms that agree on the subsemigroup U : α|U = β|U . In particular,
5.1 Definitions
xLy ⇐⇒ S 1 x = S 1 y
xRy ⇐⇒ xS 1 = yS 1
xJy ⇐⇒ S 1 xS 1 = S 1 yS 1 .
Here S 1 x, xS 1 and S 1 xS 1 are the principal left ideal, the principal right ideal and the
principal ideal generated by x ∈ S. By the definitions,
As an exercise we have
Therefore, αRβ implies that α(X) = β(X). On the other hand, if α(X) = β(X), then
define γ ∈ TX as a function, for which
Then, in the above notations, βγ(x) = β(x) = α(x), and hence α ∈ βTX1 . Similarly,
β ∈ αTX1 , and hence, αRβ ⇐⇒ α(X) = β(X). ⊔
⊓
Proof. Suppose that (x, y) ∈ L ◦ R. This means that there exists an element z ∈ S such
that xLz and zRy. Therefore there are elements s, s′ , r, r ′ ∈ S such that
x = sz, z = s′ x, z = yr, y = zr ′ .
t = sz · r ′ = xr ′ ,
x = sz = syr = szr ′ r = tr ,
t = s · zr ′ = sy,
y = zr ′ = s′ xr ′ = s′ szr ′ = s′ t ,
which implies that yLt. Since xRt and tLy, also (x, y) ∈ R ◦ L. We have shown that
L ◦ R ⊆ R ◦ L. The inclusion in the other direction is proved in the same way. ⊔
⊓
D- and H-relations
The previous result is important because it implies that the product
D =L◦R
is an equivalence relation.
Lemma 5.2. D is the smallest equivalence relation that contains both L and R.
Proof. Clearly, L ⊆ D and R ⊆ D, since xLx and xRx hold for all x ∈ S 1 .
In order to show that D is an equivalence relation one needs to show that it is transi-
tive: if xDz and zDy, then also xDy. This is an exercise.
In order to show that D is the smallest equivalence relation containing L ∪ R one
assumes an equivalence relation C with L ∪ R ⊆ C, and shows that D ⊆ C. This is an
exercise. ⊔
⊓
5.1 Definitions 54
J
L R
Let
H = L ∩ R.
Hence H is an equivalence relation, and it is the largest equivalence relation of S that is
contained in both L and R.
The inclusion diagram of the Green’s relations is given in Fig. 5.1. Here one needs
to observe that D ⊆ J by Lemma 5.2.
We denote by Dx and Hx the corresponding equivalence classes that contain the
element x ∈ S. Clearly, for all x ∈ S
H x = L x ∩ Rx .
On D-classes
Lemma 5.3. For each semigroup S we have
xDy ⇐⇒ Lx ∩ Ry 6= ∅ ⇐⇒ Ly ∩ Rx 6= ∅ .
Moreover, [ [
Dx = Ly = Ry .
y∈Dx y∈Dx
Rz Ly ∩ Rz
Example 5.3. Let X = {1, 2, . . . , 7}, and consider the subsemigroup S of the full trans-
formation semigroup TX generated by α and β as defined in the table.
1 2 3 4 5 6 7
α 2 2 5 6 5 6 5
β 3 4 3 4 7 4 7
This semigroup consists of six elements α, β, αβ, βα, αβα and βαβ.
By computing we get βαβα = βα, and hence βαRβαβ. Also, αβRαβα holds, as
do βαLαβα and αβLβαβ.
αβ αβα
βαβ βα
We conclude (after some calculations) that the eggbox for the class Dβα looks like
in the figure. ⊔
⊓
5.2 Green’s Lemma and Its Corollaries 56
Green’s lemma(s)
Let S be a semigroup. For each s ∈ S we define (as before) the mapping ρs : S 1 → S 1
by
∀z ∈ S 1 : ρs (z) = sz .
The usefulness of the Green’s relations are due to the following version of Green’s
lemma.
Lemma 5.4. Let S be a semigroup, xLy, and let s, s′ ∈ S 1 be such that sx = y and
s′ y = x. Then
1. ρs : Rx → Ry is a bijection, and ρs′ : Ry → Rx is a bijection,
2. ρs′ = ρ−1
s is the inverse function of ρs restricted to Rx ,
3. ρs fixes the L-classes, that is, zLρs (z) for all z ∈ Rx ,
4. ρs preserves H-classes, that is, for all u, v ∈ Rx : uHv ⇐⇒ ρs (u)Hρs (v) .
Rx x z
Ry y ρs (z)
Lx
Proof. First we have to prove that ρs maps Rx into Ry . For this, let z ∈ Rx , that is,
zS 1 = xS 1 . We have now
szS 1 = sxS 1 = yS 1 ,
which shows that ρs (z) = sz ∈ Ry as required. A symmetric argument shows that ρs′
maps Ry into Rx .
If z ∈ Rx , then zRx and therefore there are elements u, u′ ∈ S 1 such that z = xu
and x = zu′ . Now,
s′ sz = s′ s · xu = s′ · sx · u = s′ yu = xu = z ,
and hence
∀z ∈ Rx : s′ sz = z . (5.1)
5.2 Green’s Lemma and Its Corollaries 57
Hence
ρs′ ρs (z) = ρs′ (sz) = s′ sz = z ,
and so ρs′ ρs is the identity mapping of Rx . Similarly, ρs ρs′ is the idenitity mapping of
Ry , from which we conclude Claims (1) and (2).
For Case (3) assume z ∈ Rx . By (5.1), the elements z and ρs (z) (= sz) are in the
same L-class.
For Case (4) we notice that uHv if and only if uLv and uRv; and so
by Case (3) and since R is a left congruence (and ρs (u) = su, ρs (v) = sv). On the other
hand, if ρs (u)Hρs (v), that is, suHsv, then uHv, since, by equation (5.1), u = s′ su and
v = s′ sv (and since R is a left congruence and ρs′ fixes the L-classes). ⊔
⊓
Lemma 5.5. Let S be a semigroup, yRz, and let r, r ′ ∈ S 1 be such that yr = z and
zr ′ = y. Then
1. λr : Ly → Lz is a bijection, and λr′ : Lz → Ly is a bijection,
2. λr′ = λ−1
r is the inverse function of λr restricted to Ly , and
3. λr preserves the R-classes, that is, wRλr (w) for all w ∈ Ly .
λr
?
- x u
ρs′ ρs
y z
6
λr ′
Proof. If xLe, then x ∈ Le , that is, x = se for some s ∈ S 1 . Therefore, since e = ee,
x = se · e = xe. The other case is similar. ⊔
⊓
Corollary 5.2. The H-classes inside a D-class have the same cardinality, that is, if xDy,
then there exists a bijection between Hx and Hy .
Proof. Indeed, if x, z ∈ S are in the same D-class such that xLy, yRz with
sx = y, s′ y = x, yr = z, zr ′ = y ,
Rx x xy
Ry e y
Lx Ly
Because zLx, there exists a u ∈ S 1 with z = ux. We have thus obtained xux =
xz = x, and thus zz = uxux = ux = z, which means that z ∈ ES .
In the other direction, suppose there exists an idempotent e ∈ Ry ∩ Lx . Now, by
Lemma 5.6, ey = y and xe = x. From eRy we obtain that xeRxy, and thus xRxy.
From eLx we obtain that eyLxy, and thus yLxy. So xy ∈ Rx ∩ Ly as required.
Finally, if Ry ∩ Lx contains an idempotent, it must be unique by Corollary 5.1, since
Ry ∩ Lx is an H-class. ⊔
⊓
Proof. That Case (1) implies Case (2) is clear, since we can choose x = e = y for the
idempotent e ∈ H.
Assume Case (2). Then Case (1) follows by Theorem 5.2, since now Rx ∩ Ly =
H = Ry ∩ Lx . Therefore H contains an idempotent e, and hence by the converse claim
of Theorem 5.2, we know that H is a subsemigroup of S. Further, H is a monoid with
identity e by Lemma 5.6. We apply then Lemma 5.7 for e = f . This shows that H is a
group. Hence Case (2) implies Case (3).
Case (3) implies Case (1), since the identity element of the group H is an idempotent
of S. ⊔
⊓
Proof. The inclusion D ⊆ J is valid for all semigroups. Thus we have to show that
J ⊆ D.
Suppose that xJy. Therefore there exist u, v, r, s ∈ S 1 such that
Now,
∀i ≥ 0 : x = (ur)i x(sv)i and y = (ru)i y(vs)i .
By Lemma 5.8, there exists a nonnegative n such that (ur)n ∈ ES . We may suppose
here that ur 6= 1, for otherwise already xRy and thus xDy, since R ⊆ D.
Denote z = rx, for short. We have then
Example 6.1. (1) All groups are regular: x = xx−1 x, where x−1 is the group inverse of
x. A regular semigroup is a group if and only if it has exactly one idempotent (Exercise).
(2) If e ∈ ES is an idempotent, then e = eee, and hence all idempotents of a semi-
group are regular elements of S.
(3) The full transformation semigroup TX is regular (Exercise). ⊔
⊓
We obtain immediately
Proof. Assume first that x is regular, and let x = xyx. Then e = xy ∈ ES and f = yx
satisfy xRe and xLf .
Suppose then that xRe. Now, x = ex by Lemma 5.6, and there exists a y ∈ S 1 such
that e = xy. Therefore x = ex = xyx, and x is regular. The case for L is similar. ⊔
⊓
Rx = Re x e = xy
Rf f = yx y
Lx = Lf Le
Regular D-classes
We say that a class Dx is regular, if it contains only regular elements.
Proof. Suppose that x is regular. Hence, by Lemma 6.2, the class Dx contains an idem-
potent e, that is, Dx = De .
Let then z ∈ Dx . Now, zDe, and thus there exists u ∈ Dx with eRu and uLz. So
there are elements r, s, s′ ∈ S such that
z = s′ u = s′ eu = s′ esz = s′ ursz = z · rs · z ,
and so z is regular. ⊔
⊓
Proof. When x ∈ D, then x = xyx for some y, and hence xLyx and xRxy, where xy
and yx are idempotents. ⊔
⊓
Note that an inverse element of x, if such an element exists, need not be unique.
Proof. If x ∈ S is regular, then for some y ∈ S, x = xyx. Now, yxy = yxy · x · yxy,
and so yxy is also regular. Also, x = x · yxy · x and consequently yxy is an inverse
element of x. ⊔
⊓
Lallement’s lemma
Lemma 6.6. Let S be a regular semigroup, and let α : S ։ P be an epimorphism onto
a semigroup P . If e ∈ EP , then α−1 (e) ∩ ES 6= ∅, that is, there exists an idempotent
f ∈ ES such that α(f ) = e.
As a consequence we have
xρ ∈ ES/ρ =⇒ ∃e ∈ ES : xρ = eρ .
As an exercise
Example
A semigroup S is called an inverse semigroup, if each x ∈ S has a unique inverse
element x−1 :
x = xx−1 x and x−1 = x−1 xx−1 .
6.2 Inverse Semigroups 64
Example 6.2. (1) If S is a group, then it is an inverse semigroup, and x−1 is the group
inverse of the element x.
(2) Consider Z × Z as a drawing board so that you can draw a unit line up u, down
d, right r and left ℓ. A word w ∈ A∗ = {u, d, r, ℓ}∗ produces a figure on Z × Z, when
you start from (0, 0) and follow the instructions given by the word w.
· ·
· • · •
◦ ◦ ·
w = uurd v = ruℓdur
We identify the words w1 and w2 , and denote w1 ⊲⊳ w2 , if they draw the same figure
with the same endpoint. This means that
Theorem 6.4. Let S be an inverse semigroup. Then the idempotents ES form a sub-
semigroup of S. Moreover, ES is a semilattice, that is, the idempotents of an inverse
semigroup commute.
Proof. Let e, f ∈ ES and consider the (unique) inverse element x = (ef )−1 of ef .
Now, (
ef · xe · ef
ef = ef · x · ef =
ef · f x · ef
and
xe · ef · xe = xef x · e = xe ,
f x · ef · f x = f · xef x = f x .
6.2 Inverse Semigroups 65
x2 = xe · f x = x · ef · x = x ,
Corollary 6.1. Assume S = [X]S . If each generator x ∈ X has a unique inverse ele-
ment, then S is an inverse semigroup:
for all xi ∈ X.
Proof. Assume x, y ∈ S have unique inverse elements x−1 and y −1 , respectively. Then
A characterization
Theorem 6.5. Let S be a semigroup. The following are equivalent:
1. S is an inverse semigroup.
2. S is regular and its idempotents commute.
3. Each L-class and R-class contains an idempotent.
e = xf = xf f = ef = f e = yee = ye = f .
As exercises we have
∀x ∈ S : x−1 ES x ⊆ ES .
e ≤ f ⇐⇒ ef = e = f e .
x ≤ y ⇐⇒ ∃e ∈ ES : x = ey .
Indeed, here ≤ is
• reflexive, since x = (xx−1 ) · x, where xx−1 ∈ ES ;
• antisymmetric, since if x = ey and y = f x, then x = ey = eey = ex, and so
x = ey = ef x = f ex = f x = y;
• transitive, since if x = ey and y = f z, then also x = ey = ef z, where ef ∈ ES .
If you restrict ≤ onto ES you get the above partial order of idempotents. Indeed, if
e ≤ f , then there exists g ∈ ES such that e = gf , and here e = gf f = ef = f e as
required.
As an exercise we state
x ≤ y ⇐⇒ ∃e ∈ ES : x = ye ⇐⇒ xx−1 = yx−1 ⇐⇒ x = xy −1 x
⇐⇒ xx−1 = xy −1 ⇐⇒ x−1 x = y −1 x ⇐⇒ x−1 x = x−1 y ⇐⇒ x = xx−1 y .
3. ∀x ∈ S ∃!x′ ∈ S : xx′ ∈ ES .
4. S is an inverse semigroup that satisfies: x = xyx =⇒ y = yxy.
Proof. Case (1) implies Case (2) by choosing x′ = x−1 , the group inverse of x.
Case (2) implies Case (3): In one direction, the claim is clear, since if x = xx′ x, then
xx ∈ ES . Conversely assume that x = xx′ x for a unique x′ (in which case xx′ ∈ ES ),
′
and xy ∈ ES . Now,
xyx = xy · xy · x = xyx · x′ · xyx ,
and
xyx = xyx · y · xyx ,
wherefrom we obtain that y = x′ , when (2) is applied to xyx.
Case (3) implies Case (4), since if x = xyx, then xy ∈ ES , and from xy = xyxy ∈
ES we get y = yxy by the uniqueness assumption, and here y = x′ = x−1 .
Case (4) implies Case (1), since
Partial mappings
Let X 6= ∅ be a set. A partial mapping α : X → X is a function from a subset Y =
dom(α) of X onto ran(α) = α(Y ) ⊆ X. A partial mapping α : X → X is undefined
on all x ∈
/ dom(α).
Example 6.3. Let X = {1, 2, . . . , 5} and let dom(α) = {2, 4, 5} with α(2) = 1, α(4) =
5 and α(5) = 1. We can represent α as follows
245
α= .
151
We observe that
that is, dom(α−1 ) = ran(α) and ran(α−1 ) = dom(α). Moreover, α−1 (α(x)) = x for
all x ∈ dom(α). Clearly, α−1 ∈ IX . Furthermore,
Now, αα−1 α = α and α−1 αα−1 = α−1 , and so α−1 is an inverse element of α, and IX
is a regular semigroup.
By Theorem 6.4, we need to show that the idempotents of IX commute. For this we
prove that
ε ∈ EIX ⇐⇒ ε = ιY for some Y ⊆ X .
Indeed, suppose ε is an idempotent of IX . Then by (6.2),
and so by injectivity,
which implies that dom(ε) ⊆ ran(ε). Similarly, ran(ε2 ) = ran(ε) implies that ran(ε) ⊆
dom(ε), and so ran(ε) = dom(ε). Further, ε2 (x) = ε(x) for all x ∈ dom(ε), and thus,
by injectivity, ε(x) = x for all x ∈ dom(ε).
Finally, the idempotents commute, since for all Y, Z ⊆ X,
ιZ ιY = ιY ∩Z = ιY ιZ ,
First of all
x−1 S = x−1 xS, (6.3)
because x−1 xS ⊆ x−1 S, and x−1 S = x−1 xx−1 S ⊆ x−1 xS. In particular, since
x−1 x ∈ ES ,
y ∈ dom(τx ) = x−1 S ⇐⇒ y = x−1 xy . (6.4)
To show that τx ∈ IX we need to show that it is injective. For this let y, z ∈ x−1 S be
such that τx (y) = τx (z). Thus xy = xz, and so by (6.4), y = x−1 xy = x−1 xz = z.
We need also:
dom(τyx ) = dom(τy τx ) . (6.5)
To prove this we notice that
that is
z ∈ dom(τy τx ) ⇐⇒ z = x−1 xz and xz = y −1 yxz ,
by (6.4). Therefore if z ∈ dom(τy τx ), then
ϕ(x) = τx (x ∈ S) .
Heritage of images
Lemma 6.8. Let S be an inverse semigroup and α : S → P a homomorphism. Then
α(S) is an inverse subsemigroup of P .
Proof. We have for all x ∈ S, α(x) = α(xx−1 x) = α(x) · α(x−1 ) · α(x), and so α(S)
is a regular subsemigroup of P . Again, it sufficies to show that the idempotents of α(S)
commute.
Let g, h ∈ Eα(S) . By Lemma 6.6, there exist e, f ∈ ES so that g = α(e) and
h = α(f ). Now, gf = α(ef ) = α(f e) = f g, from which the claim follows. ⊔
⊓
In particular,
Therefore,
xρy ⇐⇒ x−1 ρy −1 .
∀x ∈ S : α(x−1 ) = α(x)−1 .
Proof. First of all, if α(x) = e = α(y), then α(xy) = e2 = e, and so xy ∈ α−1 (e).
This means that α−1 (e) is a sunsemigroup of S.
The claim follows when we show that x ∈ α−1 (e) implies x−1 ∈ α−1 (e). For this
we just notice that if α(x) = e, then α(x−1 ) = α(x)−1 = e−1 = e. ⊔
⊓
For ideals we have the following rather strong closure property (which will be an
exercise). In general, if I is an ideal of a semigroup S, then S is called an ideal extension
of I by S/I, where S/I is the Rees quotient.
Theorem 6.12. Let S be an inverse semigroup. Then for all congruences ρ and δ,
ρ ⊆ δ ⇐⇒ ∀e ∈ ES : eρ ⊆ eδ .
Proof. In one direction the claim is trivial. Suppose then that eρ ⊆ eδ for all e ∈ ES .
Now,
and thus
xρy =⇒ xδyx−1 y . (6.6)
On the other hand,
and, by combining this with (6.6), we obtain that xδy holds. This shows that ρ ⊆ δ. ⊔
⊓
ρ = δ ⇐⇒ ∀e ∈ ES : eρ = eδ
Theorem 6.13. Let S be an inverse semigroup, and let ρ and δ be its congruences. Then
α(x) ∈ EP ⇐⇒ β(x) ∈ ET ,
α(e) = α(f ) ⇐⇒ β(e) = β(f ) .
Proof. In one direction the claim is again trivial. Suppose then that ker(ρ) = ker(δ) and
tr(ρ) = tr(δ). If e ∈ ES , then
eρx =⇒ ∃f ∈ ES : f δx =⇒ f f −1 = f δxx−1 ,
eρx =⇒ ee−1 = eρxx−1 =⇒ eδxx−1 ,
and thus eδf and f δx hold, from which we obtain eδx, that is, eρ ⊆ eδ. The case for
eδ ⊆ eρ is similar, and hence we can conclude that eρ = eδ, and finally ρ = δ by
Corollary 6.5. ⊔
⊓
Proof. Indeed, by commuting idempotents, x−1 ex · x−1 ex = x−1 xx−1 eex = x−1 ex.
⊔
⊓
Theorem 6.14. For a congruence ρ of an inverse semigroup S, ρmin is the smallest con-
gruence whose trace equals tr(ρ).
6.4 Congruences of Inverse semigroups 73
Proof. To show that ρmin is an equivalence relation demands some calculations, which
we leave as an exercise.
We show that ρmin is a congruence. For this suppose xρmin y and let e ∈ ES be such
that xe = ye. Let z ∈ S. Now, f = z −1 ez ∈ ES , and
xz · f = xzz −1 · ez = xe · zz −1 z = ye · zz −1 z = yzz −1 · ez = yz · f ,
and therefore xzρmin yz. On the other hand, xe = ye implies that ex−1 = ey −1 by
taking inverses of both sides, and using this we obtain
which shows that zxρmin zy, since ex−1 z −1 zx is an idempotent. Thus ρmin is a congru-
ence.
Next we show that tr(ρmin ) = tr(ρ). For f, g ∈ ES ,
Theorem 6.15. Let S be an inverse semigroup and ρ its congruence. Then ρmax is the
largest congruence of S whose trace equals tr(ρ).
Proof. First of all we show that ρmax is a congruence. That it is an equivalence relation
is an easy exercise. Let then xρmax y and let z ∈ S. We have
Since z −1 ez ∈ ES , we have
xρmax y =⇒ x−1 · z −1 ez · xρy −1 · z −1 ez · y ,
and thus zxρmax zy holds if xρmax y holds. Similarly, for xzρmax yz. We conclude that
ρmax is a congruence.
Next we show that tr(ρmax ) = tr(ρ). For this let f, g ∈ ES . Then
f ρmax g =⇒ f −1 f f ρg −1 f g =⇒ f ρf g ,
and in the same way, f ρmax g =⇒ gρf g, from which we get that f ρg, and so
tr(ρmax ) ⊆ tr(ρ). In the other direction, for all e ∈ ES , if f ρg, then f −1 ef ρg −1 eg
and so we deduce that f ρmax g, that is, the traces are the same.
Finally ρmax is the largest of such congruences. For this assume δ isa congruence
such that tr(δ) = tr(ρ). Now, for all e ∈ ES ,
xδy =⇒ x−1 eaδy −1 ey =⇒ x−1 eaρy −1 ey =⇒ xρmax y ,
since x−1 ex, y −1 ey ∈ ES ; and so δ ⊆ ρmax as required. ⊔
⊓
The above theorem states that ρmax identifies as many elements of S as possible
with the restriction that it does not identify any idempotents unless ρ does so. Certainly,
ρ ⊆ ρmax , and so the quotient S/ρmax is an epimorphic image of S/ρ.
Group congruences
We say that a congruence ρ of a semigroup S is a group congruence, if S/ρ is a group.
The following lemma holds already for regular semigroups.
Lemma 6.13. An inverse semigroup is a group if and only if it has a unique idempotent.
For a congruence ρ of an inverse semigroup, we have by Theorem 6.2 that
xρ ∈ ES/ρ =⇒ ∃e ∈ ES : eρ = xρ ,
and hence
Theorem 6.16. A congruence ρ of an inverse semigroup S is a group congruence if and
only if tr(ρ) = ES × ES .
Proof. If tr(ρ) = ES × ES , then S/ρ has exactly one idempotent by Theorem 6.2. In
the other direction the claim is equally clear. ⊔
⊓
Remark 6.1. If S is not an inverse semigroup, it need not have the smallest group con-
gruence. As an example consider (N+ , +). The group congruences of this semigroup are
exactly ρn = {(p, q) | p ≡ p( mod n)}. ⊔
⊓
xσS y ⇐⇒ ∃e ∈ ES : xe = ye .
As an exercise we have
∀e, f ∈ ES : eρf =⇒ e = f .
By Theorem 6.15, for each inverse semigroup S there exists the greatest idempotent
separating congruence, which will be denoted by µS .
6.5 Free Inverse Semigroups 76
xµS y ⇐⇒ ∃e ∈ ES : x−1 ex = y −1 ey .
Proof. This is clear, since from Theorem 6.15, since tr(µS ) = ιE , and so the claim
follows. ⊔
⊓
Generalized notion
Free inverse semigroups (and monoids) are defined in a similar way as free semigroups
except now the homomorphic extensions of mappings must be homomorphisms of in-
verse semigroups.
It is somewhat more convenient to consider a generalized (but equivalent) notion of
freeness. This is as follows.
Let L be a class of semigroups. Let F ∈ L be a semigroup in this class, X a
nonempty set, and ϕ : X → F an injective mapping into F such that ϕ(X) generates
F . We say that the pair (F, ϕ) is a free L-semigroup on X, if for any P ∈ L and any
mapping α0 : X → P there exists a homomorphism α : F → P such that the following
diagram commutes, that is, α0 = αϕ.
ϕ
X -F
α0 α
R
P
In this case, we say also that ϕ(X) generates F freely. The homomorphism α is
called a ϕ-extension of the mapping α0 . The mapping ϕ is often omitted from the nota-
tion, and then we talk about free L-semigroups F (on a set X).
Note that here we do not assume that X is a subset of F . Nor do we claim that such
a L-free semigroup exists for a given class L.
Note also that we require that ϕ is a bijection from X onto a generating set ϕ(X) of
F . In literature one usually omits this requirement, and derives it from a more general
condition in certain classes (like ‘varieties of semigroups’).
The following is immediate.
6.5 Free Inverse Semigroups 77
Example 6.4. (1) If S is an inverse semigroup, then the mapping δ(x) = x−1 , where x−1
is the inverse element of x in S, is an involution. This is the involution we are mainly
interested in.
(2) If G is the special linear group of n-by-n-matrices (with entries in Z and determi-
nant equal to 1), then the mapping δ(A) = AT , which maps a matrix A to its transpose,
is an involution of G. ⊔
⊓
that is, if α is a homomorphism that respects (or is compatible with) the involutions.
If in the above both S and P are inverse semigroups and their involutions are the
inverse functions, x 7→ x−1 , then the latter condition becomes:
∀x ∈ S : α(x−1 ) = α(x)−1 .
Let X and X ′ be two nonempty sets of the same cardinality such that X ∩ X ′ = ∅,
and let ψ : X → X ′ be a bijection. Write Y = X∪X ′ , and define the mapping w 7→ w−1
of words w ∈ Y + by (
−1 ψ(x) if x ∈ X ,
x =
ψ (x) if x ∈ X ′ .
−1
and
6.5 Free Inverse Semigroups 78
w = x1 x2 . . . xn =⇒ w−1 = x−1 −1 −1
n xn−1 . . . x1 .
Now, (Y + ,−1 ) is a semigroup with involution, and the following theorem states that it
is a free semigroup with involution. Note that here Y + itself is just the free semigroup
on the set Y of letters – it is not an inverse semigroup.
Later we shall usually write
X ′ = X −1 = {x−1 | x ∈ X} and Y = X ∪ X −1 .
Theorem 6.21. (Y + ,−1 ) is a free semigroup with involution, that is, if (P, δ) is a semi-
group with involution and α0 : X → P is any mapping, then there exists a (unique)
homomorphism α : (Y + ,−1 ) → (P, δ) such that α0 (x) = α(x) and α0 (x−1 ) = α(x)δ
for all x ∈ X.
By the second condition, α is a homomorphism, and by the first condition it respects the
involutions:
⊔
⊓
and arbitrary direct products. It was shown by Birkhoff some 60 years ago that if a
class V of algebras forms a variety, then it has the free algebras on every nonempty set
X. Birkhoff also gave a characterization of the varieties by identities that are special
relations on words.
We prove the existence of a free inverse semigroup by taking the approach of Vagner.
Let (Y + ,−1 ) be the free semigroup with involution as constructed in the above. De-
fine a relation ρX on (Y + ,−1 ) as follows,
Further, let ρcX be the congruence generated by ρX , that is, ρcX is the smallest congruence
that contains the relation ρX . Let us denote by
ZX = (Y + ,−1 )/ρcX
uρcX = [u] ,
so that ZX = {[u] | u ∈ Y + } and the product [u][v] = [uv] satisfies the following
properties (of ρX ):
First of all each [uu−1 ] is an idempotent, since [uu−1 ] = [u][u]−1 and ZX is regular.
Suppose then that [w] is an idempotent of ZX . Then [u−1 u][uu−1 ] = [uu−1 ][u−1 u] by
the definition of ρX , and consequently,
[w] = [w][w] = [ww−1 w][ww−1 w] = [w][w−1 w][ww−1 ][w] = [w][ww−1 ][w−1 w][w]
= [w]2 [w−1 ]2 [w]2 = [w][w−1 ]2 [w] = [ww−1 ][w−1 w] = [(ww−1 )(ww−1 )−1 ] ,
X −1 = {x−1 | x ∈ X}
Example 6.5. For X = {a, b}, we have the following reduction of the word w =
ab−1 abaa−1 b−1 ba−1 b,
where the last word is reduced, and presents the same word w in GX . ⊔
⊓
Lemma 6.16. For each word w ∈ Y ∗ there exists a unique reduced word r(w) such that
w = r(w) in GX . Therefore each congruence class [u] contains a unique reduced word.
Proof. Assume there are two different reduced words u and v in the same class, and
let u = w1 , w2 , . . . , wk = v bePa sequence of words, where wi+1 ∼ wi using once
a rule (I) or (D). Denote N = ki=1 |wi | (the sum of the lengths of the words in this
sequence). We can assume that the sequence w1 , w2 , . . . , wk is chosen such that N is
minimum. Since u and v are both reduced (and different), we must have |u| < |w2 | and
|wk−1 | > |v|, that is, rule (I) is used in w1 ∼ w2 and (D) is used in wk−1 ∼ wk . It
follows that the sequence |w1 |, |w2 |, . . . , |wk | has a maximal value (a turning point): for
some i, |wi−1 | < |wi | > |wi+1 |. This means that wi is obtained from wi−1 by adding
a factor xx−1 , and wi+1 by deleting a factor yy −1 . These factors cannot coincide, for
otherwise wi−1 = wi+1 contradicts the minimality of N . If the factors overlap with
6.6 Concrete Free Inverse Semigroups 82
each other, then x−1 = y and wi contains a factor xx−1 x or x−1 xx−1 . In this case,
again wi−1 = wi+1 . Hence the factors xx−1 and yy −1 should be disjoint. However, now
we get a smaller value of N by replacing wi by the word where yy −1 is removed from
wi−1 and replacing wi+1 by the word where the factor xx−1 is then added at its place.
⊔
⊓
FX = {w ∈ Y ∗ | w = r(w)}
u · v = r(uv)
for all u, v ∈ FX .
The relations of GX are just the group axioms, and therefore
where xi ∈ X and εi ∈ {−1, +1} (so that x+1 = x), and xi 6= xi+1 . Extend α0 to FX
by setting
α(w) = α0 (x1 )ε1 α0 (x2 )ε2 . . . α0 (xk )εk .
By the uniqueness in (6.9), this is a well defined mapping, and it is rather immediate that
α is a group homomorphism. ⊔
⊓
there exists a word u ∈ Y ∗ such that e = [uu−1 ]. If here the word u is not reduced, say
u = v1 xx−1 v2 , then
uu−1 = v1 xx−1 v2 · v2−1 xx−1 v1−1 = v1 · v1−1 v1 xx−1 v2 · v2−1 xx−1 v1−1
= v1 xx−1 v1−1 v1 v2 v2−1 v1−1
= (v1 x)(v1 x)−1 (v1 v2 )(v1 v2 )−1 ,
and therefore e = [(v1 x)(v1 x)−1 ][(v1 v2 )(v1 v2 )−1 ], where the factors v1 x and v1 v2 are
shorter than u and they more reduced. From this observation we obtain
6.6 Concrete Free Inverse Semigroups 83
Note that the mapping (6.11) is not injective: for e = xyy −1 x−1 = xyy −1 x−1 ·xx−1 ,
we have both {xy} 7→ e and {xy, x} 7→ e. To overcome the noninjectivity we do the
following. Let A = {u1 , u2 , . . . , un } 7→ e, so that (6.10) holds. If here ui = v1 v2 , then
u1 u−1 −1 −1 −1 −1 −1 −1
1 . . . ui ui . . . un un = u1 u1 . . . v1 v2 v2 v1 . . . un un
= u1 u−1 −1 −1 −1 −1
1 . . . v1 v2 v2 v1 · v1 v1 . . . un un
= u1 u−1 −1 −1
1 . . . un un · v1 v1
and so also A ∪ {v1 } 7→ e. This means that we can safely add all the prefixes of the
words ui into A without changing its value e. Denote
A = {v | v a prefix of a u ∈ A}
the set of all words that are prefixes of words from A, including the empty word 1 and
the words u ∈ A, themselves. By the above, A 7→ e, and this mapping is injective.
In conclusion, let
K = {A | A = A}
be the set of all finite nonempty subsets of reduced words that are closed under taking
prefixes,
A ∈ K if and only if w = uv ∈ A =⇒ u ∈ A .
1 ) given by A 7→ e is a bijection.
Lemma 6.19. The mapping K → EM (where M = ZX
and
Ae ∪ Af = Aef .
Therefore the idempoyents of M can be identified with the closed sets A of reduced
words, and the product of the idempotents becomes defined as the union of such sets.
6.6 Concrete Free Inverse Semigroups 84
uρcX 7→ (w,
b r(w))
is a bijection. In fact, one more proof shows that ψ is an isomorphism.
6.6 Concrete Free Inverse Semigroups 85
On Reynolds construction
Reynolds (1986) proved that we get the free inverse semigroups also from the partial
mappings as follows.
Recall that IFX consists of the injective partial mappings of the free group FX into
itself. Let ψ : X → IFX be defined by
On Munn’s construction
Munn presented the free inverse semigroups using (oriented word) trees. His variant is
probably the most intuitive, but it is also (probably) less useful than Scheiblich’s con-
struction.