GENERATING SETS

KEITH CONRAD

1. Introduction
In Rn , every vector can be written as a (unique) linear combination of the standard basis
e1 , . . . , en . A notion weaker than a basis is a spanning set: a set of vectors in Rn is a
spanning set if its linear combinations fill up the whole space. The difference between a
spanning set and a basis is that a spanning set may contain more vectors than necessary
to span the space. For instance, {(1, 0), (3, 1), (2, 1)} is a spanning set for R2 but is not a
basis. Omitting a vector from this spanning set will give a basis of R2 . A basis is a minimal
spanning set. All bases of Rn have size n.
In a group, the analogue of a spanning set is called a generating set.
Definition 1.1. In a group G, a subset X ⊂ G is a generating set for G if every g ∈ G can
be written as a product of powers of elements taken from X:
(1.1) g = xa11 xa22 · · · xar r ,
where xi ∈ X and ai ∈ Z. We also say that X generates G and write G = hXi. If G has a
finite generating set, we say G is a finitely generated group.
Instead of using exponents in the definition of a generating set, we could write powers
as repeated copies of the same factor or its inverse (so g 3 h−2 = gggh−1 h−1 ). Therefore a
subset X of G is a generating set for G when every element of G is a product of elements
from X and inverses of elements from X. This is closer to what the idea of “generating set”
sounds like: through repeated use of the group operations (multiplication and inversion) we
can produce all elements of G from X.
Example 1.2. Every permutation in Sn is a product of cycles, so the cycles in Sn are a
generating set of Sn .
Example 1.3. The group Z/(m) × Z/(n) is generated by (1, 0) and (0, 1), since (a, b) =
a(1, 0) + b(0, 1)
Example 1.4. A group has a one-element generating set exactly when it is a cyclic group.
For instance, Z has the one-element generating sets {1} and {−1}.
Example 1.5. Dihedral groups have two generators: Dn = hr, si and every element is ri
or ri s. For a general group with two generators x and y, we usually can’t write elements in
the condensed form xm y n for some m and n in Z, e.g., yxyx2 is not x3 y 2 (or y 2 x3 ). We’ll
see an example in Section 4.
Example 1.6. The infinite nonabelian matrix group {( a0 1b ) : a = ±1, b ∈ Z} is finitely
generated. Since ( 10 1b ) = ( 10 11 )b and ( −1 b 1 b −1 0 1 1 b −1 0
0 1 ) = ( 0 1 )( 0 1 ) = ( 0 1 ) ( 0 1 ), this group has
−1 0 1 1
generating set {( 0 1 ), ( 0 1 )}.
1
2 KEITH CONRAD

Example 1.7. The group Q is not finitely generated: a finite set of rational numbers has a
common denominator, say N , and the subgroup of Q generated by these rational numbers
(their integral multiples and sums thereof) will only give rise to rational numbers with
denominators dividing N . Not all rationals have such denominators (try 1/(N + 1)), so Q
doesn’t have a finite set of generators as an additive group.
Example 1.8. A finitely generated group is at most countable, so an uncountable group
is not finitely generated.
In this handout, we look at generating sets of the groups Sn , An , SL2 (Z), GLn (R), and
SLn (R). The following table summarizes some the generating sets we will obtain for various
groups, and indicates where the proofs are found. At the end we discuss minimal generating
sets, which have some counterintuitive properties in nonabelian groups.

Group Generating Set Size Where

n(n−1)
Sn , n ≥ 2 (ij)’s 2 Theorem 2.1
(12), (13), . . . , (1n) n−1 Theorem 2.2
(12), (23), . . . , (n − 1 n) n−1 Theorem 2.3
(12), (12 . . . n) if n ≥ 3 2 Theorem 2.5
(12), (23 . . . n) if n ≥ 3 2 Corollary 2.6
(ab), (12 . . . n) if (b − a, n) = 1 2 Theorem 2.8
n(n−1)(n−2)
An , n ≥ 3 3-cycles 3 Lemma 3.1
(1ij)’s (n − 1)(n − 2) Theorem 3.2
(12i)’s n−2 Theorem 3.3
(i i + 1 i + 2)’s n−2 Theorem 3.4
(123), (12 . . . n) if n ≥ 4 odd 2 Theorem 3.5
(123),
 (23 . . . n) if n ≥ 4  even 2 Theorem 3.5
0 −1 1 1
SL2 (Z) , 2 Theorem 4.1
1 0 0 1
GLn (R), SLn (R) Elementary Matrices Infinite Theorem 5.2

2. Generators for Sn
The group Sn is generated by its cycles. The following theorem shows the 2-cycles (the
transpositions) are enough to generate Sn .
Theorem 2.1. For n ≥ 2, Sn is generated by its transpositions.
Proof. This is clear for n = 1 and 2. For n ≥ 3, we note (1) = (12)2 and every cycle of
length > 2 is a product of transpositions:
(i1 i2 . . . ik ) = (i1 i2 )(i2 i3 ) · · · (ik−1 ik ).
For example,
(13526) = (13)(35)(52)(26).
Since the cycles generate Sn , and products of transpositions give us all cycles, the transpo-
sitions generate Sn . 
Since transpositions have order 2 (though not every element of order 2 is a transposition,
e.g., (12)(34)), Theorem 2.1 tells us Sn is generated by elements of order 2.
 GENERATING SETS 3

The total number of transpositions in Sn is n2 = n(n−1)

2 , so Theorem 2.1 provides us
with a generating set of ≈ n2 /2 transpositions. The next theorem shows we can get a
generating set for Sn containing just n − 1 transpositions.
Theorem 2.2. For n ≥ 2, Sn is generated by the n − 1 transpositions
(12), (13), . . . , (1n).
Proof. The theorem is obvious for n = 2, so we take n ≥ 3.
By Theorem 2.1, it suffices to write each transposition in Sn as a product of the trans-
positions involving a 1. For a transposition (ij) where i and j are not 1, check that
(2.1) (ij) = (1i)(1j)(1i).

Here is a different generating set of n − 1 transpositions.
Theorem 2.3. For n ≥ 2, Sn is generated by the n − 1 transpositions
(1 2), (2 3), . . . , (n − 1 n).
Proof. By Theorem 2.1, it suffices to show each transposition (ab) in Sn is a product of
transpositions of the form (i i + 1) where i < n.
Since (ab) = (ba), without loss of generality a < b. We will argue by induction on b − a
that (ab) is a product of transpositions (i i + 1). This is obvious when b − a = 1, since
(ab) = (a a + 1) is one of the transpositions we want in the desired generating set. Now
suppose b − a = k > 1 and the theorem is settled for all transpositions moving a pair of
integers whose difference is less than k.
Consider the formula
(a b) = (a a + 1)(a + 1 b)(a a + 1).
The first and third transpositions on the right side lie in our desired generating set. The
middle transposition permutes a pair of integers with difference b − (a + 1) = k − 1 < k. By
induction, (a + 1 b) is a product of transpositions (i i + 1), so (a b) is as well. 
Now we are ready to cut down the size of a generating set for Sn to two.
Lemma 2.4. For a k-cycle (i1 i2 . . . ik ) in Sn and an arbitrary σ ∈ Sn ,
(2.2) σ(i1 i2 . . . ik )σ −1 = (σ(i1 )σ(i2 ) . . . σ(ik )).
Proof. To verify (2.2) in general, we will check both sides have the same effect on all integers
from 1 to n. (Permutations are equal when they have the same values at the same numbers.)
For an integer from 1 to n of the form σ(ir ), where 1 ≤ r < k, the left side of (2.2) has
the effect σ(ir ) 7→ ir 7→ ir+1 7→ σ(ir+1 ), and the right side also sends σ(ir ) to σ(ir+1 ). The
same reasoning shows both sides of (2.2) send σ(ik ) to σ(i1 ).
For an integer ` from 1 to n that is not among {σ(i1 ), . . . , σ(ik )}, it is not part of the
cycle on the right side of (2.2), so the right side fixes `. What does the left side of (2.2) do
to `? We want to compute
(σ(i1 i2 . . . ik )σ −1 )(`).
The permutation σ −1 sends ` to σ −1 (`). Since ` 6∈ {σ(i1 ), . . . , σ(ik )}, σ −1 (`) is not among
{i1 , . . . , ik }. Therefore the cycle (i1 i2 . . . ik ) fixes σ −1 (`), and then σ sends σ −1 (`) to `. Thus
the left side of (2.2) fixes `, just like the right side. 
4 KEITH CONRAD

Theorem 2.5. For n ≥ 2, Sn is generated by the transposition (12) and the n-cycle
(12 . . . n).
Proof. By Theorem 2.3, it suffices to show products of the permutations (12) and (12 . . . n)
yield all transpositions of the form (i i + 1). We may take n ≥ 3.
Set σ = (12 . . . n). Then
σ(12)σ −1 = (σ(1) σ(2)) = (23),
and more generally for k = 1, 2, . . . , n − 2,
σ k (12)σ −k = (σ k (1) σ k (2)) = (k + 1 k + 2).

Corollary 2.6. For n ≥ 3, Sn is generated by (12) and the (n − 1)-cycle (23 . . . n).
Proof. This is immediate from Theorem 2.5 since (12 . . . n) = (12)(23 . . . n). 
We’ve found a pair of generators for Sn consisting of a transposition and n-cycle, and
then a transposition and (n − 1)-cycle. These have orders 2 and n, and then 2 and n − 1.
How small can the orders of a pair of generators of Sn be?
Theorem 2.7. For n ≥ 3 except for n = 5, 6, 8, Sn is generated by an element of order 2
and an element of order 3.
We omit the proof of this theorem. It was first proved by G. A. Miller [4] in 1901, and
his proof relied on a choice of a prime between n and n/2, so it was not explicit in terms of
n. An explicit set of generators of order 2 and 3 was given by Dey and Wiegold [1] in 1971,
unaware of Miller’s earlier work. As an example, S9 is generated by
(14)(28)(59) and (123)(456)(789).
While Sn is generated by the particular transposition (12) and n-cycle (12 . . . n), it is
usually not true that Sn is generated by an arbitrary transposition and n-cycle. For instance,
(13) and (1234) do not generate S4 . The reason is that these two permutations preserve
congruences mod 2 (if two numbers in 1, 2, 3, 4 are both even or both odd, applying either
permutation to them returns values that are both even or both odd), so the subgroup they
generate in S4 has this property while S4 does not have this property.
Theorem 2.8. For 1 ≤ a < b ≤ n, the transposition (ab) and n-cycle (12 . . . n) generate
Sn if and only if (b − a, n) = 1.
Here the transposition (ab) is general, but the n-cycle is the standard one.
Proof. Let d = (b − a, n). We will show every g ∈ h(ab), (12 . . . n)i preserves mod d congru-
ences among {1, 2, . . . , n}:
(2.3) i ≡ j mod d =⇒ g(i) ≡ g(j) mod d.
It suffices to check this when g = (ab) and when g = (12 . . . n). For i 6= a or b, (ab)(i) = i.
Also (ab)(a) = b ≡ a mod d and (ab)(b) = a ≡ b mod d, so (ab)(i) ≡ i mod d for all i. Thus
(2.3) holds for g = (ab). As for g = (12 . . . n), we have (12 . . . n)(i) ≡ i + 1 mod n, so also
(12 . . . n)(i) ≡ i + 1 mod d since d | n. Therefore
i ≡ j mod d =⇒ i + 1 ≡ j + 1 mod d,
so (2.3) holds for g = (12 . . . n).
 GENERATING SETS 5

For d > 1, the group Sn does not preserve mod d congruences: pick i 6≡ j mod d and
consider the transposition (ij). So if h(ab), (12 . . . n)i = Sn then we must have d = 1.
Now we prove the converse direction: if (b − a, n) = 1 then h(ab), (12 . . . n)i = Sn . Let
σ = (12 . . . n), so σ i (a) ≡ a + i mod n. Therefore σ b−a (a) ≡ b mod n, and both sides of the
congruence are between 1 and n, so σ b−a (a) = b. Since (b − a, n) = 1, hσi = hσ b−a i and
σ b−a is an n-cycle sending a to b, so σ b−a = (ab . . . ) where the dots are other numbers in
the range {1, 2 . . . , n}. Then
h(ab), σi = h(ab), σ b−a i = h(ab), (ab . . . )i.
A suitable relabeling of the numbers 1, 2 . . . , n (that is, making an overall conjugation on Sn )
turns (ab) into (12) and (ab . . . ) into (12 . . . n), so h(ab), σi is conjugate to h(12), (12 . . . n)i,
which is Sn by Theorem 2.5. Therefore h(ab), σi = Sn . 
Corollary 2.9. For each transposition τ = (ab) and n-cycle σ in Sn , hτ, σi = Sn if and
only if (k, n) = 1, where σ k (a) = b.
Proof. Exercise. 
Corollary 2.10. For a prime number p, Sp is generated by an arbitrary transposition and
p-cycle.
Proof. An arbitrary p-cycle in Sp can be written as (12 . . . p) by relabeling the objects being
permuted (that means by applying an overall conjugation on Sp ), so to show an arbitrary
transposition and p-cycle generate Sp it suffices to show each transposition and the specific
p-cycle (12 . . . p) generate Sp . For a transposition (ab) where 1 ≤ a < b ≤ p, we have
(b − a, p) = 1, so h(ab), (12 . . . p)i = Sp by Theorem 2.8. 

3. Generators for An
We now look for generating sets of An , where n ≥ 3. (For n = 1 and 2, An is trivial.)
Our development will parallel in large part what we did for Sn . In particular, we will see
that An can be generated by two permutations.
Within Sn , every element of An is a product of transpositions, but the transpositions
themselves do not lie in An . The smallest cycles in An (excluding the trivial 1-cycle) are
the 3-cycles. Do these generate An ? Yes.
Lemma 3.1. For n ≥ 3, each element of An is a product of 3-cycles. Therefore the 3-cycles
generate An .
Proof. The identity (1) is (123)(132), which is a product of 3-cycles. Now pick a non-identity
element of An , say σ. Write it as a product of transpositions in Sn :
σ = τ1 τ2 · · · τr .
The left side has sign 1 and the right side has sign (−1)r , so r is even. Therefore we can
collect the products on the right into successive transpositions τi τi+1 , where i = 1, 3, . . .
is odd. We will now show every product of two transpositions in Sn is a product of two
3-cycles, so σ is a product of 3-cycles.
Case 1: τi and τi+1 are equal. Then τi τi+1 = (1) = (123)(132), so we can replace τi τi+1
with a product of two 3-cycles.
Case 2: τi and τi+1 have exactly one element in common. Let the common element be a,
so we can write τi = (ab) and τi+1 = (ac), where b 6= c. Then
τi τi+1 = (ab)(ac) = (acb) = (abc)(abc),
6 KEITH CONRAD

so we can replace τi τi+1 with a product of two 3-cycles.

Case 3: τi and τi+1 have no elements in common. This means τi and τi+1 are disjoint, so
we can write τi = (ab) and τi+1 = (cd) where a, b, c, d are distinct (so n ≥ 4). Then
τi τi+1 = (ab)(cd) = (ab)(bc)(bc)(cd) = (bca)(cdb) = (abc)(bcd),
so we can replace τi τi+1 with a product of two 3-cycles. 
Let’s reduce the number of 3-cycles needed.
Theorem 3.2. For n ≥ 3, the group An is generated by 3-cycles of the form (1ij).
Proof. For each 3-cycle (abc) not containing a 1, we have (abc) = (1ab)(1bc). Now use
Lemma 3.1. 
Theorem 3.2 is an analogue for alternating groups (on at least 3 letters) of Theorem 2.2
for symmetric groups (on at least 2 letters): each theorem says there is a generating set of
cycles that contain 1 and have the smallest length consistent with the parity allowed in the
group. The 3-cycles containing a 1 leave two undetermined terms, so our generating set for
An in Theorem 3.2 is quite a bit larger than the generating set for Sn in Theorem 2.2. Here
is a sharper analogue of Theorem 2.2 for An , where we only allow one undetermined entry
in the 3-cycles.
Theorem 3.3. For n ≥ 3, the group An is generated by 3-cycles of the form (12i).
Proof. For n = 3, the only such 3-cycle is (123), and we know A3 = {(1), (123), (132)} is
generated by (123). We now take n ≥ 4.
Since (12i)−1 = (1i2), each 3-cycle in An containing 1 and 2 is generated by the 3-cycles
of the form (12i). For a 3-cycle containing 1 but not 2, say (1ij), check
(1ij) = (12j)(12j)(12i)(12j).
By Theorem 3.2, we’re done. 
We can describe Theorem 3.3 in a “coordinate-free” manner as follows: An is generated
by the 3-cycles moving a common pair of terms.
Here is an An -analogue of Theorem 2.3.
Theorem 3.4. For n ≥ 3, the consecutive 3-cycles (i i + 1 i + 2), with 1 ≤ i ≤ n − 2,
generate An .
Proof. This is true for n = 3 since A3 = {(1), (123), (132)} is cyclic with generator (123).
We know by Theorem 3.3 that A4 is generated by (123) and (124). Since
(3.1) (124) = (123)(123)(234)(123),
we see (123) and (234) also generate A4 .
Now take n ≥ 5. Theorem 3.3 says a generating set for An is the set of 3-cycles (12i).
We argue by induction on i that these particular 3-cycles can be produced from products
of consecutive 3-cycles. This is obvious for i = 3, and is shown in (3.1) for i = 4. For i ≥ 5,
assume (12j) is a product of consecutive 3-cycles for 3 ≤ j < i. Then the equation
(1 2 i) = (1 2 i − 2)(1 2 i − 1)(i − 2 i − 1 i)(1 2 i − 2)(1 2 i − 1)
and the inductive hypothesis show (12i) is a product of consecutive 3-cycles. 
Here is an analogue of Theorem 2.5.
 GENERATING SETS 7

Theorem 3.5. For n ≥ 4, An is generated by two elements:

(
(12 . . . n), if n is odd,
(123) and
(23 . . . n), if n is even.
The theorem is true in a redundant sense when n = 3.
Proof. It suffices, by Theorem 3.4, to obtain all (i i+1 i+2) from the indicated permutations.
First suppose n is odd. Let σ = (12 . . . n), so σ ∈ An . Then, for 1 ≤ k ≤ n − 3.
σ k (123)σ −k = (σ k (1)σ k (2)σ k (3)) = (k + 1 k + 2 k + 3).
Now suppose n is even. Let σ = (23 . . . n), so σ ∈ An . Then, for 1 ≤ k ≤ n − 3,
σ k (123)σ −k = (σ k (1)σ k (2)σ k (3)) = (1 k + 2 k + 3).
This does not give us the consecutive 3-cycles right away, but we can obtain them from
what we now have, since
(k k + 1 k + 2) = (1 k k + 1)(1 k + 1 k + 2).

There is an analogue of Theorem 2.7 for alternating groups, with a slightly different set
of exceptions.
Theorem 3.6. For n ≥ 3 except for n = 6, 7, 8, An is generated by an element of order 2
and an element of order 3.
We omit the proof. As an example, A9 is generated by
(14)(29)(37)(56) and (123)(456)(789).

4. SL2 (Z)
The group SL2 (Z) consists of 2 × 2 integer matrices with determinant 1 (under multi-
plication). For instance, ( 26 7
11 3 ) is in SL2 (Z). There are two important matrices in this
group:    
0 −1 1 1
S= , T = .
1 0 0 1
Since S 2 = ( −1 0 4 10 k 1k
0 −1 ) and S = ( 0 1 ), the matrix S has order 4. Since T = ( 0 1 ) for all
k ∈ Z, so T has infinite order.
Theorem 4.1. The group SL2 (Z) is generated by S and T .
Proof. As the proof will reveal, this theorem is essentially the Euclidean algorithm in dis-
guise. A worked example follows the proof.
First we check how S and a power of T change the entries in a matrix. Verify that
       
a b −c −d k a b a + ck b + dk
S = and T = .
c d a b c d c d
Thus, up to a sign change, multiplying by S on the left interchanges the rows. Multiplying
by a power of T on the left adds a multiple of the second row to the first row and does not
change the second row. Given a matrix ( ac db ) in SL2 (Z), we can carry out the Euclidean
algorithm on a and c by using left multiplication by S and powers of T . We use the power
of T to carry out the division (if a = cq + r, use k = −q) and use S to interchange the
8 KEITH CONRAD

roles of a and c to guarantee that the larger of the two numbers (in absolute value) is in
the upper-left corner. Multiplication by S will cause a sign change in the upper row, but
this has no serious effect on the algorithm.
Since ad − bc = 1, a and c are relatively prime, so the last step of Euclid’s algorithm
will have a remainder of 1. This means, after suitable multiplication by
S’s and
T ’s, we
will have transformed the matrix ( ac db ) into one with first column ±1 0 or ±1 0
. Left-
multiplying by S interchanges the rows up to a sign, so we can suppose the first column
is ±1 1x


0 . Each matrix of the form ( 0y ) in SL 2 (Z) must have y = 1 (the determinant is 1),
1 x x −1 x
and then it is ( 0 1 ) = T . A matrix ( 0 y ) in SL2 (Z) must have y = −1, so the matrix is
( −1 x
0 −1 ) = −T
−x = ( −1 0 )T x . Since ( −1 0 ) = S 2 , we can finally unwind and express our
0 −1 0 −1
original matrix in terms of S’s and T ’s. 
Example 4.2. Take A = ( 26 7
11 3 ). Since 26 = 11 · 2 + 4, we want to subtract 11 · 2 from 26:
 
−2 4 1
T A= .
11 3
Now we want to switch the roles of 4 and 11. Multiply by S:
 
−2 −11 −3
ST A = .
4 1
Divide −11 by 4: −11 = 4 · (−3) + 1, so we want to add 4 · 3 to −11. Multiply by T 3 :
 
3 −2 1 0
T ST A = .
4 1
Once again, multiply by S to switch the entries of the first column (up to sign):
 
3 −2 −4 −1
ST ST A = .
1 0
Our final division is: −4 = 1(−4) + 0. We want to add 4 to −4, so multiply by T 4 :
 
4 3 −2 0 −1
T ST ST A = = S.
1 0
Thus, left-multiplying by the inverses of all the S’s and T ’s on the left side, we obtain
A = T 2 S −1 T −3 S −1 T −4 S.
Since S has order 4, we can write S −1 as S 3 if we wish to use a positive exponent on S.
However, a similar idea does not apply to the negative powers of T .
In this example, we wrote ( 26 7 a b b a
11 3 ) in terms of S and T , but not in the form S T or T S .
The next theorem shows that is impossible.
Theorem 4.3. Each matrix of the form S a T b or T b S a has 0 in one of its entries.
Proof. Since S has order 4, each matrix S a T b or T b S a has the form S a ( 10 1b ) and ( 10 1b )S a
where 0 ≤ a ≤ 3. Since S 0 = I2 and S 2 = −I2 , this gives us 6 matrix products:
     
b 1 b b 0 −1 2 b b 2 −1 −b
T = , ST = , S T =T S = ,
0 1 1 b 0 −1
     
3 b 0 1 b b −1 b 3 −b 1
S T = , T S= , T S = .
−1 −b 1 0 −1 0
In all cases, the product has 0 in one of its entries. 
 GENERATING SETS 9

Since SL2 (Z) = hS, T i, we can also say SL2 (Z) = hS, ST i. This is interesting because
S has order 4 and ST = ( 01 −1 1 ) has order 6. In other words, SL2 (Z) is an infinite group
that can be generated by two elements of finite order. Moreover, in the quotient group1
PSL2 (Z) := SL2 (Z)/{±I2 } = hS, ST i, S has order 2 and ST has order 3. This quotient
group is “universal” for the property of being generated by elements of order dividing 2
and 3: if G = hx, yi is a group generated by elements x and y where x2 = e and y 3 = e,
there is a unique group homomorphism PSL2 (Z) → G sending S to x and ST to y, so G
is isomorphic to a quotient group of PSL2 (Z). For example, An and Sn are generated by
elements of order 2 and 3 when n ≥ 9, so these groups are quotient groups of PSL2 (Z).

5. GLn (R) and SLn (R)

In linear algebra, matrices get into row echelon form by elementary row operations:
• Multiply a row by a nonzero scalar.
• Add a multiple of one row to another.
• Exchange two rows.
Elementary row operations on an m × n matrix can be expressed using left multiplication
by an m × m matrix having a particularly simple form: just one entry might not be 0 or 1.
We will illustrate this on a 2 × 3 matrix ( ad eb fc ).
(1) Multiplying a row by λ is left multiplication by ( λ0 10 ) or ( 10 λ0 ), for instance
    
λa λb λc λ 0 a b c
= .
d e f 0 1 d e f
(2) Adding λ times one row to another row is left multiplication by ( 10 λ1 ) or ( λ1 01 ), for
instance
    
a + λd b + λe c + λf 1 λ a b c
= .
d e f 0 1 d e f
(3) Exchanging rows is left multiplication by ( 01 10 ):
    
d e f 0 1 a b c
= .
a b c 1 0 d e f
An elementary matrix is one that corresponds to an elementary row operation. They
come in three flavors:
• Di (λ) has 1’s on the main diagonal except for a λ 6= 0, 1 in the ith slot along the
main diagonal.
• In + λEij with i 6= j is a matrix with 1’s on the main diagonal and a nonzero λ in
position (i, j).
• Rij with i 6= j has 1’s in positions (i, j) and (j, i) and also in every diagonal position
except the ith and jth, and 0’s everywhere else. Note Rij = Rji . (This is a matrix
analogue of the transposition (ij).)
The inverse of such a matrix is of the same type: Di (λ)−1 = Di (1/λ), (In + λEij )−1 =
−1
In − λEij , and Rij = Rij .
Theorem 5.1. The group GLn (R) is generated by elementary matrices.
1The group {±I } is the center of SL (Z).
2 2
10 KEITH CONRAD

Proof. Pick A ∈ GLn (R). Its reduced row echelon form is In , the only n × n invertible
reduced row echelon matrix. Elementary row operations correspond to left multiplication
by an elementary matrix, so there are elementary matrices E1 , E2 , . . . , Ek such that
E1 E2 . . . Ek A = In .
Therefore A = Ek−1 · · · E2−1 E1−1 . The inverse of an elementary matrix is an elementary
matrix, so A is a product of elementary matrices. 
Is the subgroup SLn (R) generated by the elementary matrices In + λEij ? They have
determinant 1 and the other elementary matrices do not: det(Di (λ)) = λ and det(Rij ) =
−1. The proof of Theorem 5.1 does not work in SLn (R) since row reduction can involve
row exchanges or scaling a row to make an entry 1, and these correspond to the two types
of elementary matrices not in SLn (R). Nevertheless, it turns out that the guess is correct:
Theorem 5.2. For n ≥ 2, the matrices In + λEij with i 6= j and λ ∈ F × generate SLn (R).
Proof. This involves tedious computations. First we treat the case n = 2. Pick ( ac db ) in
SL2 (R). To show it is a product of matrices ( 10 1∗ ) or ( 1∗ 01 ), first suppose b 6= 0. Then
     
a b 1 0 1 b 1 0
= .
c d (d − 1)/b 1 0 1 (a − 1)/b 1
If c 6= 0 then
     
a b 1 (a − 1)/c 1 0 1 (d − 1)/c
= .
c d 0 1 c 1 0 1
If b = 0 and c = 0 then the matrix is ( a0 1/a
0
), and
      
a 0 1 0 1 1 1 0 1 −1/a
= .
0 1/a (1 − a)/a 1 0 1 a−1 1 0 1
To handle n > 2, we induct on n. Pick A ∈ SLn (R). We will show multiplying A on the
left or right by matrices In + λEij leads to a block matrix ( 01 A00 ). Since this is in SLn (R)
we have det A0 = 1, so A0 ∈ SLn−1 (R). By induction A0 is a product of elementary matrices
In−1 + λEij , so ( 01 A00 ) is a product of block matrices ( 01 In−1 +λE
0
ij
) = In + λEi+1 j+1 . Thus
(product of some In + λEij )A(product of some In + λEij ) = product of some In + λEij ,
and we can solve for A to see that it is a product of matrices In + λEij .
Multiplying A by In + λEij on the left or right is an elementary row or column operation:
a11 ··· a1n
 
.. .. ..

 . . .


(In + λEij )A =  ai1 + λaj1 · · · ain + λajn  ,
 
 .. .. .. 
 . . . 
an1 ··· ann
where the ith row equals ith row of A + λ(jth row of A, and
 
a11 · · · a1j + λa1i · · · a1n
 .. .. .. .. ..  ,
A(In + λEij ) =  . . . . . 
an1 · · · anj + λani · · · ann
where the jth column equals the jth column of A + λ(ith column of A).
 GENERATING SETS 11

Looking along the first column of A, some entry is not 0 (since det A = 1). If ak1 6= 0
where k > 1 then
!
1 ···
 
1 − a11
(5.1) In + E1k A = .. . . .
ak1 . .
If a21 , . . . , an1 are all 0, then a11 6= 0 and
 
  a11 · · ·
1
In + E21 A =  1 ... .
 
a11 .. ..
. .
Then by (5.1) with k = 2,
!
1 ···
 
1
(In + (1 − a11 )E12 ) In + E21 A = .. . . ,
a11 . .
Once we have a matrix with upper left entry 1, multiplying it on the left by In + λEi1
for i 6= 1 will add λ to the (i, 1)-entry, so with a suitable λ we can make the (i, 1)-entry of
the matrix 0. Thus multiplication on the left by suitable matrices of the form In + λEij
produces a block matrix ( 01 B∗ ) whose first column is all 0’s except for the upper left entry,
which is 1. Multiplying this matrix on the right by In + λE1j for j 6= 1 adds λ to the
(1, j)-entry without changing a column other than the jth column. With a suitable choice
of λ we can make the (1, j)-entry equal to 0, and carrying this out for j = 2, . . . , n leads to
a block matrix ( 01 A00 ), which is what we need to conclude the proof by induction. 

6. Minimal Generating Sets

Since a basis of Rn is the same thing as a minimal spanning set, perhaps the analogue
of a basis of Rn in the context of groups is a minimal generating set: a generating set that
is no longer a generating set when some element is removed. In contrast to bases for Rn ,
different minimal generating sets of a nonabelian group may not have the same size.
Example 6.1. Two minimal generating sets for S4 are {(12), (23), (34)} and {(12), (1234)}.
That these sets each generate S4 is a special case of Theorems 2.3 and 2.5.
Why is the first generating set minimal? Any two transpositions in it have a common
fixed point or commute. These properties are preserved under multiplication, and S4 has
neither property, so two permutations in the first set do not generate S4 . In the second set,
removing a permutation leaves just one, which does not generate S4 since S4 is not cyclic.
Example 6.2. The generating sets in Theorems 2.2 and 2.3 are both minimal. Why?
Well, removing a transposition (1i) from the generating set in Theorem 2.2 will leave a
set of transpositions with a common fixed point i, and thus they can’t generate all of Sn .
To explain why the generating set in Theorem 2.3 is minimal is a bit more subtle. It we
take out (12) or (n − 1 n), then at most we will be able to generate permutations having a
common fixed point (1 or n), and thus we do not get all of Sn . If we take out (i i + 1) for
some i from 2 to n − 1, then we are left with
(12), (23), . . . , (i − 1 i), (i + 1 i + 2), . . . , (n − 1 n).
Convince yourself that these transpositions can never produce the transposition that swaps
1 and n. The “bridge” linking 1 to n through the generating set has been broken when
(i i + 1) is taken out.
12 KEITH CONRAD

Theorem 6.3. A set of transpositions that generates Sn must have at least n − 1 transpo-
sitions, so a set of n − 1 transpositions in Sn that generates Sn is a minimal generating set
of Sn .

Proof. Let τ1 , . . . , τm be distinct transpositions in Sn , and assume they generate Sn . We

want to show m ≥ n − 1. We will use graph theory.
Create a graph whose vertices are {1, 2, . . . , n} and make an edge between i and j if
(ij) is one of the transpositions τ1 , . . . , τm . So our graph has n vertices and m edges. For
two vertices i and j in the graph, to say there is an element of the group hτ1 , . . . , τm i that
sends i to j is equivalent to saying there is a path in our graph starting at i and ending at
j. Therefore when hτ1 , . . . , τm i = Sn , this graph is connected. A connected graph with n
vertices must have at least n − 1 edges, so m ≥ n − 1. 

Remark 6.4. Your prior experience with bases in linear algebra might suggest that 2-
element generating sets for Sn , being as small as possible (for n ≥ 3), should be the most
useful generating sets. In fact, the generating set of transpositions (i i + 1) in Theorem 2.3
turns out to be the most important one.

Because minimal generating sets in a non-abelian group need not have a common size,
there is not a theory of “bases” for finitely generated groups that permits the same degree
of elegance as bases for finite-dimensional vector spaces in linear algebra. In fact, finitely
generated groups can fail to satisfy natural analogues of theorems about finite-dimensional
vector spaces:
• a subgroup of a finitely generated group need not be finitely generated, (!)
• even if a subgroup of a finitely generated group is finitely generated, it might require
more generators than the original group. (!)
For a finite group G, let d(G) denote the smallest possible size of a generating set for G.
(With analogies to linear algebra in mind, we might naively think of d(G) as a “dimension,”
hence the notation.) For instance, d(G) = 1 precisely when G is a cyclic group. Theorems
2.5 and 3.5 tell us d(Sn ) and d(An ) equal 2 when these groups are non-abelian (that is,
except for very small n). Consider the following questions about d(G) for a general finite
group G, motivated by an analogy with bases in Rn :
• Do all minimal generating sets for G have d(G) terms?
• Does every generating set for G contain a generating set of d(G) elements?
• If H ⊂ G is a subgroup, is d(H) ≤ d(G)?
The answer to these questions, in the context of all finite groups, is no. Indeed, the
symmetric groups Sn for n ≥ 4 answer the first two questions in the negative, using the
minimal generating set (12), (13), . . . , (1n) from Theorem 2.2. The symmetric groups also
provide examples that negate the third question, as follows. Let H be the subgroup of Sn
generated by the transpositions (12), (34), . . . , (2j − 1 2j), . . . . While d(Sn ) = 2, it can be
shown that d(H) = [n/2]. In particular, d(H) > d(Sn ) for n ≥ 6.
Despite this bad news in general, if we restrict our attention to finite groups G of prime-
power order then the answers to the first two questions above are yes: all minimal generating
sets of G have the same size and every generating set of G contains a generating set with
that minimal size. The main result leading to this is called the Burnside Basis Theorem,
which involves the Frattini subgroup.
 GENERATING SETS 13

Definition 6.5. For a nontrivial finite group G, its Frattini subgroup Φ(G) is the intersection
of the maximal (⇒ proper) subgroups of G:
\
Φ(G) = H
max.subgp.

If G is trivial, define Φ(G) to be trivial too.

Always Φ(G) C G, since conjugates of maximal subgroups are also maximal subgroups.
When G is a finite p-group, its maximal subgroups are its subgroups of index p and they
are all normal subgroups. For a maximal subgroup M , G/M has order p, so [G, G] ⊂ M
from G/M being abelian. Thus [G, G] ⊂ Φ(G), so G/Φ(G) is an abelian p-group. Since p-th
powers in each G/M are trivial, we have {g p : g ∈ G} ⊂ Φ(G), so all nontrivial elements of
G/Φ(G) have order p. Thus G/Φ(G) ∼ = (Z/(p))r for some r ≥ 1. We can think of r as a
dimension: G/Φ(G) is a vector space over Z/(p) (scale by exponentiation: (a mod p)g = g a )
and r = dimZ/(p) (G/Φ(G)).
Example 6.6. If G = hgi is nontrivial cyclic of p-power order then it has only one maximal
proper subgroup, hg p i, so Φ(G) = hg p i and G/Φ(G) = hgi/hg p i is cyclic of order p.
Example 6.7. The quaternion group Q8 has maximal subgroups hii, hji, hki, which inter-
sect in {±1}, so Φ(Q8 ) = {±1} and Q8 /Φ(Q8 ) = hi, ji ∼
= (Z/(2))2 .
Example 6.8. If   
 1 a b 
G=  0 1 c  : a, b, c ∈ Z/(p)
0 0 1
 
for a prime p then   
 1 0 b 
Φ(G) =  0 1 0  : b ∈ Z/(p)
0 0 1
 

and G/Φ(G) ∼
= (Z/(p))2 .
Theorem 6.9 (Burnside). Let G be a finite p-group. For each subset {g1 , . . . , gk } of G,
hg1 , . . . , gk i = G ⇐⇒ hg 1 , . . . , g k i = G/Φ(G).
A proof of Theorem 6.9 can be found in specialized texts on group theory. The theorem
reduces many questions about generating sets of a finite p-group G to questions about
spanning sets of a finite-dimensional vector space. In particular, all minimal generating sets
of G have common size equal to the size of a basis of G/Φ(G) as a vector space over Z/(p)
and every generating set for G must contain a generating set of d(G) elements because every
spanning set of a vector space of dimension d has at least d elements.
The implication H ⊂ G ⇒ d(H) ≤ d(G) has counterexamples among finite p-groups.
(There wouldn’t be counterexamples if we always had H ∩Φ(G) = Φ(H), but such a formula
isn’t even true if H is a nontrivial subgroup of a cyclic p-group.2) Before describing a family
of counterexamples, we describe a situation where the implication is valid for p-groups.
Let Gn be the subgroup of a finite group G generated by its nth powers (this may
include elements that are not nth powers, just as the commutator subgroup is generated
by commutators but its elements may not all be commutators). Then Gn C G, and a finite
2 If G = hgi, set H = Φ(G) = hg p i. Then Φ(H) = hg p2 i, so H ∩ Φ(G) = Φ(G) 6= Φ(H) if |G| > p.
14 KEITH CONRAD

p-group is called powerful if G/Gp is abelian for p 6= 2 or G/G4 is abelian for p = 2. All
abelian p-groups are powerful, and powerful p-groups share some features of abelian groups,
e.g., if G is a powerful p-group then Gp = {g p : g ∈ G} [2, Prop. 2.6]. It turns out that for
powerful p-groups G, H ⊂ G ⇒ d(H) ≤ d(G) [2, Theorem 2.9].
Here is a construction of a finite p-group G where d(H) > d(G) for a subgroup H.
Example 6.10. For a prime p, let A be the p × p matrix over Z/(p) that permutes the
standard basis cyclically: A(e1 ) = e2 , A(e2 ) = e3 , . . . , A(ep ) = e1 and A is extended to
(Z/(p))p linearly:  
0 0 0 ... 0 1
 1 0 0 ... 0 0 
 
 0 1 0 ... 0 0 
A =  .. .. .. . . .. .
 
 . . .
 . . 

 0 0 0 ... 0 0 
0 0 0 ... 1 0
This matrix has order p. Set V := (Z/(p))p and G := V o hAi (a semi-direct product), with
powers of A acting on V by matrix-vector multiplication. The group law is
(v, Ai )(w, Aj ) = (v + Ai w, Ai+j ).
Then |G| = pp+1 .
Concretely, we can think of G as certain affine transformations of V : (v, Ai ) ∈ G is the
mapping V → V where x 7→ Ai x + v. For v ∈ V , the product vA in G is the mapping
x 7→ Ax + v while the product Av in G is the mapping x 7→ A(x + v) = Ax + Av. Therefore
A and v don’t commute in G if Av 6= v, so A does not commute in G with members of the
standard basis of V .
Since powers of A send e1 to the entire standard basis of V and repeated addition sends
each ei to its Z/(p)-multiples, from A and e1 we can get all of V and thus A and e1 generate
the group G. Neither A nor e1 generates G by itself (G is not abelian), so d(G) = 2. The
subgroup H := V of G is a p-dimensional vector space over Z/(p), so d(H) = p. Thus when
p 6= 2 we have constructed a p-group G and a subgroup H where d(H) > d(G).
An example where d(G) > d(H) for 2-groups can also be constructed, but this is omitted
here.

References
[1] I. M. S. Dey and J. Wiegold, “Generators for alternating and symmetric groups,” J. Australian Math.
Soc. 12 (1971), 63–68.
[2] J. D. Dixon, M. P. F. du Sautoy, A. Mann, D. Segal, Analytic Pro-p Groups, 2nd ed., Cambridge Univ.
Press, Cambridge, 1999.
[3] A. Lubotzky, A. Mann, “Powerful p-groups I. Finite Groups,” J. Algebra 105 (1987), 484–505.
[4] G. A. Miller, “On the groups generated by 2 operators,” Bull. Amer. Math. Soc. 7 (1901), 14–32.