Probabilistic group theory, combinatorics, and computing

Chapter 2
Estimation Problems and Randomised
Group Algorithms

´
Alice C. Niemeyer, Cheryl E. Praeger, and Akos Seress

2.1 Estimation and Randomization

2.1.1 Computation with Permutation Groups

In 1973, Charles Sims [89] proved the existence of the Lyons–Sims sporadic
simple group Ly by constructing its action as a group of permutations of a set of
cardinality 8,835,156 on a computer which could not even store and multiply the two
generators of Ly in this smallest degree permutation representation for the group!
The existence of this ﬁnite simple group, together with many of its properties, had
been predicted by Richard Lyons [60], but proof of existence was not established
until Sims’ construction. Leading up to this seminal achievement, Sims [88] had
developed concepts and computational methods that laid the foundation for his
general theory of permutation group computation.

A.C. Niemeyer ( )
Centre for the Mathematics of Symmetry and Computation, School of Mathematics and Statistics,
The University of Western Australia, 35 Stirling Highway, Crawley, WA 6009, Australia
e-mail: alice.niemeyer@uwa.edu.au
C.E. Praeger
King Abdulaziz University, Jeddah, Saudi Arabia
e-mail: cheryl.praeger@uwa.edu.au
´
A. Seress
The Ohio State University, Columbus, OH, USA
e-mail: akos@math.ohio-state.edu

A. Detinko et al. (eds.), Probabilistic Group Theory, Combinatorics, and Computing, 35
Lecture Notes in Mathematics 2070, DOI 10.1007/978-1-4471-4814-2 2,
© Springer-Verlag London 2013

36 A.C. Niemeyer et al.

Sims introduced the critical concept of a base of a permutation group G on a
finite set ˝, namely a sequence of points ˛1 ; : : : ; ˛b of ˝ such that only the identity
of G fixes all of them. For example, the dihedral group D2n D ha; bi acting on
f1; 2; : : : ; ng, where a D .1; 2; : : : ; n/ and b D .2; n/.3; n 1/ : : : , has a base
B D .1; 2/, since only the identity of D2n fixes both 1 and 2. Moreover the 2n
elements g 2 D2n produce 2n distinct image pairs .1g ; 2g / of the base B—for
example, a maps B to .2; 3/, b maps B to .1; n/.
Sims observed that elements of a permutation group G could always be
represented uniquely by the sequence of images of the points of a given base B.
He exploited this potentially compact representation of group elements, ingeniously
showing how to compute in G with these base images, via a so-called strong
generating set of G relative to B. Sims’ algorithm to construct a base and strong
generating set, called the Schreier–Sims algorithm, is of fundamental importance
for permutation group computation.
For groups possessing a small base, the Schreier–Sims algorithm is extremely
efficient, but for some groups every base has size close to the cardinality n D j˝j
of the point set. For such groups, the methods are not effective. Examples of such
large-base groups include the “giants”: the alternating group Alt.˝/ D An and the
symmetric group Sym.˝/ D Sn , which have minimum-sized bases .1; 2; : : : ; n 2/
and .1; 2; : : : ; n 1/ respectively.

2.1.2 Recognising the Permutation Group Giants

For computational purposes, a finite permutation group G on ˝ is given by a
(usually small) set X of generators. The group G consists of all products of arbitrary
length of elements from X . Since the Schreier–Sims algorithm is ineffective for
computation with the giants Alt.˝/ and Sym.˝/, it is important to determine in
advance (that is, before trying to find a base and strong generating set) whether or
not a given permutation group G D hX i is one of these giants. Thus the question
of identifying the giants Alt.˝/ and Sym.˝/, given only a generating set of
permutations, was a central issue in the development of general purpose group
theory computer systems.
Theoretically the problem of detecting these giants had engaged mathematicians
from the earliest studies of group theory. Since the seminal work of Camille Jordan
in the 1870s, it has been known that there are many kinds of permutations such that
the only transitive permutation groups containing them are the giants (we say that
G Ä Sym.˝/ is transitive if each pair of points of ˝ can be mapped one to the
other by an element of G). The most beautiful of these results that identifies a large
family of such elements is Jordan’s theorem below.
Let us call a permutation g 2 Sn a Jordan element if g contains a p-cycle, for
some prime p with n=2 < p < n 2. For example, g D .1; 2; 3; 4; 5/.6; 7/ 2 S9 is
a Jordan element (with n D 9; p D 5).

2 Estimation and Group Algorithms 37

Theorem 2.1. If a transitive permutation group G Ä Sn contains a Jordan element
then G is An or Sn .
Given a set of generators for G Ä Sym.˝/, it is easy to test whether G is
transitive. Hence, recognising the giants boils down to the question: how prevalent
are the Jordan elements in the giants? For a fixed prime p 2 .n=2; n 2/, the number
of elements in Sn containing a p-cycle is
!
n nŠ nŠ
.p 1/Š.n p/Š D .and in An /;
p p 2p

so the proportion of Jordan elements in An or Sn for thisP prime p is 1=p, and
1 c
therefore the proportion of Jordan elements in An or Sn is n=2<p<n 2 p log n
for some constant c. For n 100, c can be taken to be 1=5, which follows by
applying an inequality by Dusart [25, p. 414] to determine the number of primes p
with n=2 < p < n 2. So roughly c out of every log n independent, uniformly
distributed random elements from Sn or An will be Jordan. That is to say, we should
find a Jordan element with high probability by randomly selecting elements in
a giant.

2.1.3 Monte Carlo Algorithms

How do we turn the comments above into a justifiable algorithm? We want to make
some multiple of log n random selections from a transitive group G on n points
which we suspect may be Sn or An , but as yet we have no proof of this fact. We hope,
and expect, to find a Jordan element, thereby uncovering the secret and proving that
G really is a giant Sn or An .
Formally, we model this process as a Monte Carlo algorithm. The Monte Carlo
method was invented by Stanislaw Ulam in the 1940s; it was named after Monte
Carlo Casino in Monaco which Ulam’s uncle visited often (see the account in [62]).
The characteristics of a Monte Carlo algorithm are that it completes quickly, but
allows a small (user-controlled) probability of “error”, that is, of returning an
incorrect result. In our context, for a Monte Carlo algorithm, we begin with a
prescribed bound on the error probability " 2 .0; 1/. The algorithm typically makes
a number N D N."/ of random selections, depending on ", this number being
determined in advance to guarantee that the probability of an incorrect result is at
most ".
Here is a worked example of a Monte Carlo algorithm to recognise the giants Sn
and An among transitive permutation groups on n points.


Algorithm 1: JORDAN
Input: A transitive subgroup G D hx1 ; : : : ; xk i Ä Sn and a real number " 2 .0; 1/ (the error
probability bound);
Output: true or false;
# We hope the algorithm returns true if G is Sn or An – see the comments below;
for up to N D d.log " 1 /.log n/=ce random elements g from G do
if g is a Jordan element then
return true;
end
end
Return false;

Comments on the algorithm
1. The procedure completes after at most N repeats of the if statement, so it is an
algorithm! If it returns true then G D An or Sn by Jordan’s Theorem 2.1. On the
other hand, if the algorithm returns false then this may be incorrect, but only if
G does equal An or Sn , and we failed to find a Jordan element.
2. We have
Prob(we do not find a Jordan element, given that G D An or G D Sn ) Ä
ÁN
c
1 log n < ".
So Algorithm 1 is a Monte Carlo algorithm with error probability less than ".
This is a special kind of Monte Carlo algorithm: the result true is always correct,
and the possibility of an incorrect result is confined to the case where false is
returned. Such algorithms are called one-sided Monte Carlo algorithms.
3. This probability estimate assumes that the random selections made are indepen-
dent and uniformly distributed. There are algorithms available for producing “ap-
proximately random” elements from a group given by a generating set; see [3,18,
24]. We shall not discuss the theoretical details of these algorithms or their prac-
tical performance. Rather we assume in our discussion of randomised algorithms
that we are dealing with independent uniformly distributed random elements.
4. The design and discussion of this simple algorithm used concepts and results
from group theory to prove correctness, and from number theory to establish the
bound on the error probability. It is typical to gather and develop methods from
a variety of mathematical areas to achieve good algorithm design and analysis.
5. Algorithm 1 is essentially the algorithm used in GAP [37] and MAGMA [15]
for testing if a permutation group is a giant. It was first described by Parker and
Nikolai [79], preceding Sims’ work by a decade. The second author (Praeger)
recalls numerous discussions with John Cannon, over a number of years, about
the implementation of this algorithm in connection with his development of the
computer algebra system CAYLEY (a precursor to MAGMA). There was much
to learn about improving the practical performance of the algorithm to avoid its
becoming a bottle-neck for permutation group computation. A wider class of
“good elements” than the Jordan elements was used, based on generalisations of


Jordan’s Theorem (see [94, 13.10] and [83]), and better methods were developed
to produce “approximately random” elements.

2.1.4 What Kinds of Estimates and in What Groups?

Notice the role estimates played in Algorithm 1:
a lower bound for the proportion of Jordan elements gives an upper bound on
the error probability.
Does it matter if the estimate is far from the true value? We might, for different
reasons, propose one of two different answers to this question:
1. We might say “no”, because if there are more Jordan elements than our estimates
predict, then we simply find one more quickly and the algorithm confirms that
“G is a giant” more efficiently.
2. We might say “yes”, because if G is not a giant then we force the algorithm to
do needless work in testing too large a number of random elements so that the
algorithm runs more slowly than necessary on non-giants. Note that the algorithm
will never find a Jordan element in a non-giant by Theorem 2.1, so the full quota
of random elements will be tested before completion.
For general purpose algorithms such as Algorithm 1, which are used frequently
on arbitrary permutation groups, the quality of the estimates really does matter. We
should try to make estimates as good as possible, especially when they are used to
analyze a time-critical module of a randomised algorithm.
In the computer algebra systems GAP and MAGMA, new algorithms are under
development for computation with matrix groups and permutation groups. These
employ a tree-like data structure which allows a “divide and conquer” approach,
reducing to computations in normal subgroups and quotient groups. This approach
(see Sect. 2.4.1) reduces many computational problems to the case of finite simple
groups. Accordingly many of the topics chosen in this chapter are of relevance to
computing with finite simple groups.

2.1.5 What Group is That: Recognising Classical Groups
as Matrix Groups

As a more substantial example for group recognition, we describe an algorithm to
recognise a finite classical group in its natural representation. By this, we mean
that the algorithm will return the “name” of the group. We give a broad-brush
description of the classical recognition algorithm developed in [72] generalising the
Neumann–Praeger SL-recognition algorithm in [69].
The algorithm takes as input a subgroup G of a finite n-dimensional classical
group Class.n; q/ over a finite field Fq of order q, such as the general linear group
GL.n; q/ or a symplectic group Sp.n; q/, in its natural representation as a group of


matrices acting on the underlying vector space V .n; q/. The subgroup G is given
by a generating set of n n matrices over Fq . The algorithm seeks so-called ppd
elements in G which we describe as follows.
For an integer e > 1, a primitive prime divisor (ppd) of q e 1 is a prime r
dividing q e 1 such that r does not divide q i 1 for any i < e. It has been known
for a long time that primitive prime divisors exist unless q D 2, e D 6, or e D 2
and q C 1 is a power of 2; see [97]. Superficially, primitive prime divisors seem
interesting because the order of the classical group has the form
Y
j Class.n; q/j D q some power .q i ˙ 1/:
various i

We define a ppd-.qI e/ element g 2 Class.n; q/ as an element with order divisible
by a ppd of q e 1. The algorithm in [72] seeks two ppd elements, namely a
ppd-.qI e1/ and a ppd-.qI e2/ element for e1 6D e2 and e1 ; e2 > n=2, which
satisfy various additional minor conditions described in [72, Sects. 2 and 9]. We
call such a pair a good ppd matrix pair. Their importance lies in the following deep
theorem [72, Theorem 4.8], the proof of which relies heavily on the finite simple
group classification.
Theorem 2.2. If G Ä Class.n; q/ is irreducible on V .n; q/ and G contains a good
ppd matrix pair, then (essentially) G D Class.n; q/ or G is known explicitly.
Thus, provided that (a) we can test efficiently whether G is irreducible on
V .n; q/, (b) good ppd matrix pairs are sufficiently prevalent in Class.n; q/ and are
easily identifiable, and (c) the exceptions in Theorem 2.2 are easy to deal with, the
good ppd matrix pairs could play the role of the Jordan elements used to identify
the permutation group giants in Algorithm 1. We would then have an analogue of
Algorithm 1 for classical groups, underpinned by considerably deeper theory than
Jordan’s Theorem 2.1. It would look like this:

Algorithm 2: RECOGNISECLASSICAL
Input: An irreducible subgroup G D hX1 ; : : : ; Xk i Ä Class.n; q/ and a real number
" 2 .0; 1/ (the error probability bound).
Output: true or false
# If the output is true, we are certain that G D Class.n; q/;
# the output false may be incorrect;
for Many(depending on n; q; ") random elements g 2 G do
determine if g is a ppd element with the additional properties;
if a good ppd matrix pair is found then
if G is one of the exceptions then
return false
else
return true;
end
end
end
return false;


Comments on the algorithm
1. Note that if Algorithm 2 returns true then G really is Class.n; q/ by Theorem 2.2;
while if it returns false then the result may be incorrect (namely if G D
Class.n; q/ and we fail to ﬁnd the good ppd matrix pair).
2. The missing ingredient is our knowledge of the presence of good ppd matrix
pairs in Class.n; q/, and an estimate of their proportion. We need a positive lower
bound on their proportion in order to decide how Many random elements to test
to ensure an error probability of at most ". This is necessary to prove that we
have a one-sided Monte Carlo algorithm.
Estimating the proportion of ppd-.qI e/ elements in Class.n; q/: For the
details involved in dealing with the additional properties we refer the reader to [72].
For G D Class.n; q/ and e > n=2, let ppd.G; e/ be the proportion of ppd-.qI e/
elements in G. We give a few details for the general linear case.
Lemma 2.3. Let G D GL.n; q/ and let n
2 < e Ä n. Then 1
eC1 Ä ppd.G; e/ Ä 1 .
e

Proof. Let g 2 G be a ppd-.qI e/ element and let r be a ppd of q e 1 dividing
jgj. By considering a power of g of order r, we can show that g leaves invariant a
unique e-dimensional subspace U of V .n; q/, and acts irreducibly on U . Moreover
the induced element gjU is a ppd-.qI e/ element in GL.U /, and a straightforward
counting argument (see [72, Lemma 5.1]) shows that ppd.G; e/ D ppd.GL.U /; e/.
In other words, we may assume that n D e in the proof. With this assumption,
we have g irreducible on V .n; q/, and each such element lies in a Singer cycle
S D Zq e 1 of G. All Singer cycles are conjugate in G, and distinct Singer cycles
contain disjoint sets of irreducible elements. Moreover the number of Singer cycles
is jG W NG .S /j D jGj=.e.q e 1// (see [69, Lemma 2.1]). Hence ppd.G; e/ is equal
to .1=e/ (the proportion of such elements in the cyclic group S ).
This immediately gives ppd.G; e/ Ä 1=e. We need one more observation to
obtain the lower bound. Certainly each element of S of order not divisible by r
lies in the unique subgroup S0 of S of index r. Thus each element of S n S0 has
order divisible by r, and hence ppd.G; e/ .1=e/ .1 1=r/. Now e is the least
positive integer such that q e Á 1 .mod r/, and so q has order e modulo the prime r.
This implies that e divides r 1, and in particular r e C 1. Hence ppd.G; e/
.1=e/ e=.e C 1/ D 1=.e C 1/. t
u
A similar argument in [72, Theorem 5.7] shows that the bounds of Lemma 2.3
hold for the other classical groups for almost all values of e. Since each ppd element
corresponds to just one e-value (because e > n=2), we can ﬁnd a lower bound for
the proportion of ppd elements in G P adding the lower bounds for ppd.G; e/ over
by
all relevant e. For GL.n; q/, this is n=2<eÄn 1=e log 2 by Lemma 2.3. For the
other classical groups, the values of e occurring all have the same parity (odd for
unitary groups and even for symplectic and orthogonal groups), and for these groups
the proportion of ppd elements is roughly .log 2/=2 [72, Theorem 6.1].
These lower bounds (or rather, the equivalent ones we obtain in [72] after taking
into account the additional conditions on the ppd elements) allow us to decide


how many random selections to make in order to find a good ppd matrix pair
with probability at least 1 1=", and hence to determine the value for Many in
Algorithm 2.

2.1.6 What Group is That: Recognising Lie Type Groups
in Arbitrary Representations

Of course, we do not only encounter the classical groups in their natural represen-
tation. If G is a simple group of Lie type, given in any permutation or matrix group
representation, and the characteristic p of G is known, then we may proceed by an
extension of Algorithm 2. The procedure that we sketch was developed in [6].
Let e1 and e2 be the two highest ppd exponents, that is, integers e such that
G contains elements of order divisible by a primitive prime divisor of p e 1. It
was shown in [6] that for each pair of integers .e1 ; e2 /, there are at most seven
isomorphism types of Lie type groups of characteristic p with e1 ; e2 as the highest
ppd exponents in the group. Also, ppd elements with ppd exponents e1 and e2 are
frequent enough that we encounter them in a random sample of size polynomial in
the input length.
To distinguish between the possibilities for G with the same values e1 and e2 , we
consider the third highest ppd exponent in G and elements whose order is divisible
by a product of two ppd primes, corresponding to certain chosen ppd exponents. The
result is a polynomial-time Monte Carlo algorithm that names the isomorphism type
of G, with one exception: a polynomial-size random sample may not distinguish the
groups Sp.2m; p f / and O.2m C 1; p f /, for odd primes p. This last ambiguity was
handled by Altseimer and Borovik [1].

2.2 Proportions of Elements in Symmetric Groups

2.2.1 Notation

In this section we fix a set ˝ and consider the symmetric group Sym.˝/ on
˝. When ˝ D f1; : : : ; ng for some positive integer n we write Sn instead of
Sym.f1; : : : ; ng/. Elements of Sn are written in disjoint cycle notation. The number
of cycles of a given element g 2 Sn denotes the number of disjoint cycles g has on
f1; : : : ; ng including fixed points.

2.2.2 Historical Notes

The study of proportions of permutations has been of interest to mathematicians
for a long time. For example, in 1708 Monmort introduced and analyzed a game


of 13 cards which he called “jeu de Treize” (the game of thirteen) in his book on
the theory of games [64, pp. 54–64]. He later generalised the game to any number
of cards numbered from 1 to n [65, pp. 130–143]. In the game, a player has n
turns, each time announcing out loud the number of the turn and picking a card at
random from the deck of n cards without replacing it. The game is won if each time
the number of the card and the number announced are different. Leonhard Euler in
Solutio Quaestionis curiosae ex doctrina combinationum [34] describes the game as
follows: Data serie quotcunque litterarum a; b; c; d; e etc., quarum numerus sit n,
invenire quot modis earum ordo immutari possit, ut nullo in eo loco reperiatur, quem
initio occupaverat. This can be translated as Given an arbitrary series (sequence)
of letters a; b; c; d; e; : : :, let the number of which be n, find in how many ways their
order may be changed so that none reappears in the same place which it originally
occupied.1 In [33] Euler showed that the number of solutions is the integer closest
to nŠ=e. Earlier solutions had already been given; for example, Monmort presented
a solution by Nicolas Bernoulli [65, pp. 301–302]. De Moivre also mentions the
game already in the first edition of [23], and gives a solution in [23, Problem 35].
Today this problem is often called the hat-swapping problem: Suppose n men
each put a hat on a hat rack in a restaurant. When they leave they each choose a
random hat. What is the probability that no man chooses his own hat?
Nowadays we call a permutation in Sn which has no fixed points on f1; : : : ; ng a
derangement, and we would rephrase the game of thirteen, Euler’s question or the
hat-swapping problem as: How many derangements are there in Sn ?
In this section we will focus on certain other proportions of elements in Sn .
The proportions that we focus on arise either from algorithmic applications for
permutation groups or from applications to classical groups of Lie type (see
Sect. 2.3.2).

2.2.3 Orders of Permutations

The order of a permutation can easily be read off from its disjoint cycle notation;
namely it is the least common multiple of the cycle lengths. One of the oldest results
on the order of an element in a symmetric group is due to Landau, who determined
how large the order of an element in Sn can be asymptotically.
Theorem 2.4 (Landau [51]).

log maxg2Sn .ord.g//
lim p D 1:
n!1 n log.n/

1
Translation by Peter M. Neumann, The Queen’s College, University of Oxford.


Although the order of an element in Sn can be as large as the previous theorem
suggests, Erd˝ s and Tur´ n were able to prove, in the ﬁrst of a series of papers [27–
o a
32] on the subject of the statistics of permutations, that most elements have much
smaller order.
Theorem 2.5 (Erd˝ s and Tur´ n [27]). For "; ı > 0 there is a number N0 ."; ı/
o a
such that for all n N0 ."; ı/,

jfg 2 Sn j .1=2 "/ log2 .n/ Ä log.ord.g// Ä .1=2 C "/ log2 .n/gj
1 ı:
nŠ
Erd˝ s and Tur´ n proved many more insightful results on the order of elements
o a
in symmetric groups. For example, in [28] they investigated prime divisors of
the order of elements in symmetric groups. In [29] they described for any x the
limiting behaviour as n tends to inﬁnity of the proportion of elements g in Sn for
which log.ord.g// Ä 1 log2 .n/ C x log3=2 .n/. In [30] they considered, among other
2
problems, the number of different values that ord.g/ can have as g ranges over the
elements of Sn .
Goh and Schmutz [39] prove that the logarithm of the average order of a random
q R
p 1
permutation in Sn is c n= log.n/, where c D 2 2 0 log log 1 e t dt. This
e
constant is approximately 2:99.

2.2.4 Number of Cycles

Let a.n/ denote the average number of cycles of the elements in Sn . In a seminal
paper [40], Gonˇ arov examined various properties of random permutations. Among
c
many other results, he proved that the average number of cycles of a permutation in
Sn is close to log.n/.
Theorem 2.6 (Gonˇ arov [40]).
c

X1
n
a.n/ D D log.n/ C C o.1/
i D1
i

for n ! 1.
Plesken and Robertz [82] generalised these results to An and to wreath products
of groups with imprimitive action.

2.2.5 Generating Functions

One very powerful method of obtaining information about certain combinatorial
quantities is to employ generating functions.


Given a sequence .an /n2N of real numbers, the Ordinary Generating Function
for an is X
A.z/ WD an zn :
n 0

For example, an could be the number of certain elements in Sn .
A very intuitive way to view generating functions is given in the following quote
from Wilf’s aptly named book generatingfunctionology [96]: “A generating function
is a clothesline on which we hang up a sequence of numbers for display.” Here
we just highlight some of the ways in which generating functions can shed light
on some of our problems. To understand the power and beauty of the subject of
generating functions we refer the reader to both Wilf’s book [96] and a recent book
on analytic combinatorics by Flajolet and Sedgewick [35]. Both books also contain
various interesting results on proportions of permutations.
Several types of generating functions can be defined, and the type of generating
function chosen to attack a particular problem depends on the circumstances. In our
situation exponential generating functions are of particular interest. They are of the
form X an
A.z/ WD zn
n 0
nŠ

and ensure that the coefficients an of zn are manageable in situations where an
nŠ
is expected to grow almost as fast as nŠ. For example, if an is the number of
elements with a particular property in Sn ; then this number could grow rapidly and
using an ordinary generating function would quickly produce unwieldy coefficients.
However, dividing by the order of the group Sn ensures that the coefficients an =nŠ
are proportions of elements in Sn and thus all less than 1.
We study generating functions as elements of the ring of formal power series.
Analytic questions, convergence etc. do not concern us just yet. Generating func-
tions can be manipulated in various ways, and this theory is described in the above
mentioned books. Here we just state, as an example, how two generating functions
can be multiplied:
! ! !
X
1 X
1 X X
1 n
n
an z n
bn z D ak bn k zn :
nD0 nD0 nD0 kD0

The usefulness of taking a generating function approach in our situation can
be highlighted with the following example. A further example, that estimates the
proportion of regular semisimple elements in general linear groups, is given in
Sect. 2.3.6.

2.2.5.1 Example

Let b 1 be a fixed integer and let an denote the number of permutations in Sn all
of whose cycles have length at most b.


We would like to study the exponential generating function describing the
numbers an . So let X an
A.z/ WD zn :
n 0
nŠ

One very effective way of studying a generating function is to start from a
recursive equation for the coefficients an , and we employ this method here. Our
first task is to find a suitable recursion for an . Recall that we write permutations in
disjoint cycle notation. We are interested in finding an expression for the number an
of permutations in Sn all of whose cycles have length at most b in terms of am for
integers m smaller than n. We employ a combinatorial trick which has been used
e.g. in Beals et al. [9, Theorem 3.7].
We enumerate the permutations in Sn all of whose cycles have length at most
b according to the length d of the cycle containing the point 1. For a fixed d , we
have d 1 choices for the remaining points of the cycle of length d containing 1.
n
1
On these d points we can put any one of .d 1/Š different cycles and we have
an d choices for the permutation on the remaining n d points. Thus we obtain the
recursion
1 X
minfb;ng
an an d
D :
nŠ n .n d /Š
d D1

Note in particular that an D nŠ for n Ä b, which is in agreement with this recursion.
The recursion implies that
0 1
X an
1 X 1 minfb;ng an d
1 X
A.z/ WD zn D 1 C @ A zn
nŠ n .n d /Š
nD0 nD1 d D1

X X 1 an d
b 1 X X 1 an
b 1
D 1C zn D 1 C znCd :
n .n d /Š nD0
n C d nŠ
d D1 nDd d D1

A very useful trick when working with generating functions is to take the
derivative. This yields in our case

X X an
b 1 X
b X an
1 X
b
A0 .z/ D znCd 1
D zd 1
zn D zd 1
A.z/:
nŠ nŠ
d D1 nD0 d D1 nD0 d D1

Thus
A0 .z/ X b
D zd 1
A.z/
d D1

and so
X zd
b
log.A.z// D :
d
d D1


Therefore we see that our generating function is

X zd
b
A.z/ D exp. /:
d
d D1

While this has yielded a very succinct way of describing the number of elements
of interest, it does not as yet yield the desired upper and lower bounds for the
proportion of such elements. Thus we would like to know whether generating
functions can tell us about the limiting behaviour of the coefficients.
An elaborate theory of the asymptotic behaviour of the coefficients of the
generating functions exists. We mention here briefly a subject called “Saddlepoint
Analysis”. The theory is described in the above mentioned books (see also the
papers by Moser and Wyman [67, 68] and Bender [11]). We quote here one result
from Flajolet and Sedgewick’s book, which helps in the situation of our example.
The quoted result is based on a more general theorem by W.K. Hayman [43] (see
also Theorem VIII.4 of [35]). In line with the literature, we denote the coefficient
d
of zn in the generating function A.z/ by Œzn A.z/. The operator z d z is defined by
z d z W P .z/ 7! zP .z/.
d 0

P
Theorem 2.7 (see Corollary VIII.2 of [35]). Let P .z/ D n D1 aj zj have non-
j
negative coefficients and suppose gcd.fj j aj 6D 0g/ D 1. Let F .z/ D exp.P .z//.
Then
1 exp.P .r//
Œzn F .z/ p ;
2 rn
2
where r is defined by rP 0 .r/ D n and D z dz
d
P .r/.

2.2.5.2 Example of Saddlepoint Analysis
P d
Recall that A.z/ D exp. b D1 zd / is the exponential generating function for the
d
number of elements all of whose cycles have length at most b.
P d
Let P .z/ D b D1 zd . Then P .z/ is a polynomial in z with non-negative
d
coefficients and satisfies gcd.fd j coefficient of zd is nonzerog/ D 1. The first step
in applying Saddlepoint Analysis P to estimate rP
is determined by the equation n D
p
rP 0 .r/. We have n D rP 0 .r/ D r b D1 r d 1 D b D1 r d r b ; and so r Ä b n.
d d
r 2 Pb
The next step is to estimate , where D r dr P .r/ D r d D1 dr d 1 D
Pb Pb
d D1 dr Ä b d D1 r D bn.
d d
P d 1 Pb
Hence we have r Ä n1=b , Ä bn and P .r/ D b D1 rd d b d D1 r D b ;
d n

implying
1 exp.P .r// 1 e Án=b
Œzn A.z/ p p :
2 rn 2 bn n


2.2.6 Solutions to x m D 1 in Symmetric and Alternating
Groups

The number of solutions to an equation of the form x m D 1 for a fixed integer m
in symmetric and alternating groups of degree n has received quite a lot of attention
in the literature. More recently, interest in such equations has been rekindled due
to algorithmic applications. In particular, it has also been important for algorithmic
applications to find the asymptotic behaviour of the number of solutions of equations
of the form x m D 1 where m is allowed to grow with n.
We begin by outlining some of the results in the literature. For m fixed let

1
c.n; m/ D jfg 2 Sn j g m D 1gj:
nŠ
Let
X
1
Cm .z/ D c.n; m/zn
nD0

be the corresponding generating function.
Theorem 2.8 (Jacobsthal [47]). For a prime p we have

zp Á X
Œn=p
1
Cp .z/ D exp z C and c.n; m/ D :
p .n p/Š Šp
D1

The number nŠc.n; 2/ of solutions to the equation x 2 D 1 in symmetric groups of
degree n deserves particular attention, since it is also the sum of the degrees of the
irreducible representations of Sn . Chowla et al. [20] examined c.n; 2/ and showed
p
that n c.n; 2/ D c.n 1; 2/Cc.n 2; 2/. Thus they deduced that 1= n Ä c.n; 2/ Ä
p
1= n C n and found the dominant term of the asymptotic expansion for c.n; 2/.
1

Later, Chowla et al. [21] were able to generalise Jacobsthal’s expansion of Cp .x/
to Cm .x/ where m can be an arbitrary fixed integer, and they asked for an asymptotic
formula for c.n; m/.
Theorem 2.9 (Chowla et al. [21]).
0 1
X zd
Cm .z/ D exp @ A:
d
d jm

Moser and Wyman [66, 67] derived an asymptotic formula in terms of a contour
integral for c.n; 2/ and derived the first order term of the asymptotic value of
c.n; p/. Moreover, they were able to obtain corresponding results for alternating
groups.


Theorem 2.10 (Moser and Wyman [66, 67]). For a prime p > 2,

1 1 n Án.1 1
p/
1
p
c.n; p/ p en :
nŠ p e

Herrera [44] gives the following recursive formula for the number nŠb.n; m/ of
elements in Sn of order m:
X .n 1/Š X
nŠb.n; m/ D b.n s; t/; where gcd.t; s/ D m:
s
.n s/Š t

Other authors (e.g. Chernoff [19], Mineev and Pavlov [63], and Pavlov [80])
studied the number of elements in Sn or An satisfying an equation of the form
x m D a for some element a 2 Sn .
In 1986 Volynets [92], Wilf [95] and Pavlov [81] independently determined the
limiting behaviour of c.n; m/ for fixed m, and n tending to infinity. The following
theorem is Wilf’s formulation of the result.
Theorem 2.11. Let m be a fixed positive integer. Define ".n; m/ D 0 if m is odd
and ".n; m/ D 1=.2m2n/ if m is even; and let
0 1
1 @ 1 X
D 1=m 1 C nd=n C ".n; m/A :
n nm
d jm;d <m

Then for n ! 1 we have
n X 1
c.n; m/ p expf g:
2 mn d d
d jm

The above result has been generalised in the literature in various directions and
we shall mention some of these.

2.2.6.1 Families of m

Ben-Ezra [10] generalised these formulae as follows. Let ˘ be a set of primes and
let ˘ 0 denote the set of all primes not in ˘ . Further, let C˘ .z/ denote the generating
function for the proportion c.n; ˘ / of all elements whose order only involves
primes in ˘ , and let C˘ 0 .z/ denote the generating function for the proportion
c.n; ˘ 0 / of all elements whose order only involves primes in ˘ 0 . For a finite set
Q
B of integers, define jjBjj D 1 if B D ¿ and jjBjj D b2B b otherwise. Then


Theorem 2.12 (Ben-Ezra [10]).
Q . 1/jBjC1
1. C˘ .z/ D BÂ˘ 0 .1 zjjBjj / jjBjj .
jBj<1
Q . 1/jBjC1
2. C˘ 0 .z/ D BÂ˘ .1 zjjBjj / jjBjj .
jBj<1

2.2.6.2 Growing m

The ﬁrst author to consider an equation of the form x m D 1 in symmetric groups of
degree n in which m is not assumed to be ﬁxed was Warlimont [93], who considers
the case m D n. In particular, he shows that
Theorem 2.13 (Warlimont [93]).
Â Ã
1 2c 1 2c 1
C 2 Ä c.n; n/ Ä C 2 C O ;
n n n n n3 o.1/

where c D 0 if n is odd and c D 1 if n is even.
In 1990 Erd˝ s and Szalay [26] considered the case where m lies in the range
o
log.n/=.2 log log.n// Ä m Ä n.1=4/ " , and derived an asymptotic formula for
c.n; m/.
Volynets [92] proved the following result via the Saddlepoint method.
Theorem 2.14 (Volynets [92]). For primes p, and for positive integers n such that
n and p tend to 1 and p=n ! 0,

1 n Án.1 X .n1=p /mCkp
1
1=p/
c.n; p/ D p 1=2 .1 C o.1//;
nŠ e .m C kp/Š
kD0

where m D n pŒn=p. In particular, if n1=p =p 2 ! 0 then

1 n Án.1 1=p/ 1=p
c.n; p/ D p 1=2 e n .1 C o.1//;
nŠ e

while if n1=p =p ! 0 then

1 n Án.1 1=p/ nm=p
c.n; p/ D p 1=2 .1 C o.1//:
nŠ e mŠ
Finally A.V. Kolchin [49] proved the following theorem using the method of
generalised schemes of allocation (see [50, Chap. 5]).


Theorem 2.15 (Kolchin [49]). For d , n positive integers such that d log log.n/=
log.n/ ! 0 and for ı D 0 if d is odd and ı D 1=.2d / when d is even, the following
holds:
8 9
1 nn.1 1=d / 1 <X nj=d =
c.n; d / D p exp ı .1 C o.1//:
nŠ en d : j ;
j jd

Another generalisation of the above to the case where the cycle lengths are
elements of particular sets can be found in [90]. Finally we would like to refer
the interested reader to V.F. Kolchin’s book on random graphs [50], which contains
many references and notes to the above mentioned, and other, results on random
permutations.

2.2.7 The Munchausen Method (Bootstrapping)
¨

The previous results highlight how difficult it is to obtain the overall limiting
behaviour for c.n; m/ when m Ä `n for some constant ` and m is allowed to grow
with n. However, for our algorithmic applications (see Sect. 2.2.8 below), we require
good upper bounds for c.n; m/ in the case where m D r.n k/ for r 2 f1; 2; 3g
and k Ä 6.
To obtain bounds for c.n; m/ in cases where n 1 Ä m Ä `n for some constant `,
we return to more basic methods and highlight some of the ideas in a proof of the
limiting behaviour of c.n; m/ in such cases.
A popular folk tale tells the story of how Baron M¨ nchausen found himself stuck
u
in a swamp while riding his horse. He then managed to save himself and his horse
by pulling himself out of the swamp by his own ponytail.
We employ a similar strategy to obtain good estimates for our required propor-
tions. We begin by deriving a first crude estimate and then using this to refine our
estimates. This method (also called bootstrapping) was employed in [9] and later in
[73].
The overall estimate for c.n; m/ is obtained in two steps. The first step yields a
very crude estimate. This in turn is employed in a second step to yield a more refined
estimate. 8
<2 for 360 < m
Define .m/ WD 2:5 for 60 < m Ä 360
:
3:345 for m Ä 60.
A first crude estimate for c.n; m/ is given in the following theorem.
Theorem 2.16. Let m; n 2 N with m n 1. Then

1 .m/m
c.n; m/ Ä C :
n n2


Proof-Idea for Crude Estimate

The proof of our first crude estimate relies on a simple idea. It divides the problem
of estimating c.n; m/ into several smaller problems by considering the following
proportions in Sn (see [9]) according to how many cycles the numbers 1, 2 and 3 lie
in. Define proportions
1. c .1/ .n; m/ of those g 2 Sn which have 1; 2; 3 in the same g-cycle.
2. c .2/ .n; m/ of those g 2 Sn which have 1; 2; 3 in two g-cycles.
3. c .3/ .n; m/ of those g 2 Sn which have 1; 2; 3 in three g-cycles.
Then it is clear that

c.n; m/ D c .1/ .n; m/ C c .2/ .n; m/ C c .3/ .n; m/:

For each i with i 2 f1; 2; 3g, we can now hope to use the extra knowledge about
the elements that contribute to the proportion c .i / .n; m/ to obtain a first estimate for
this proportion.
For example, we show how we can obtain an estimate for c .1/ .n; m/. Elements
g 2 Sn contributing to this proportion must contain a cycle C of length d with the
following properties:
1. d j m and 3 Ä d .
2. The cycle C of length d contains 1,2,3.
3. The remaining cycles of g all have lengths dividing m.
Now we can obtain an expression for c .1/ .n; m/ by considering all allowable cycle
lengths d and counting the number of cycles C on d points that contain the points
1, 2 and 3 and ensuring that the remaining n d points all have lengths dividing m.
As C has to contain 1, 2 and 3, we have n 3 points left to choose the remaining
d 3 points of C ; and having chosen a set of d points (which contains the points 1,
2 and 3), we have .d 1/Š ways of arranging them into different cycles. The number
of permutations on the remaining n d points all of whose cycle lengths divide m
is c.n d; m/.n d /Š. Hence
!
1 X n 3
c .n; m/ D
.1/
.d 1/Šc.n d; m/.n d /Š
nŠ d 3
d jm;d 3

.n 3/Š X
D .d 1/.d 2/c.n d; m/:
nŠ
d jm;3Äd Än

As we are currently only interested in obtaining a first crude estimate, we apply a
very rough upper bound on c.n d; m/, by replacing it with the constant 1. We
therefore find


.n 3/Š X
c .1/ .n; m/ Ä .d 1/.d 2/
nŠ
d jm;3Äd Än

.n 3/Š X m m
Ä . 1/. 2/
nŠ t t
m=nÄt Äm=3
Z !
m=3
.n 3/Š m2
Ä .n 1/.n 2/ C dt
nŠ m=n t2
.n 3/Š
Ä f.n 1/.n 2/ C mn 3mg
nŠ
1 m
< C 2:
n n

We can employ similar estimates to obtain crude upper bounds for c .2/ .n; m/
and c .3/ .n; m/, which we omit here. Having obtained a first crude estimate, we now
insert this estimate when trying to get a better estimate for c.n; m/.

2.2.7.1 The Pull

Enumerating g by the g-cycle of length d on 1 and recalling that n 1 Ä m yields

1 X
c.n; m/ D c.n d; m/
n d jm
1Äd Än

1 1 X
Ä C c.n d; m/:
m n d jm
1Äd Äm=2

For example, in the case where m D n or m D n 1, inserting the crude estimate
for c.n d; m/ in the equations above we find that

1 1 X Â 1 .m/m
Ã
c.n; m/ Ä C C
m n d jm
n d .n d /2
1Äd Äm=2

1 d.m/.2 C 4 .m//
Ä C ;
m n2

where d.m/ denotes the number of positive integer divisors of m. The above results
allow us to prove the following strong corollaries.
Corollary 2.17. Let n 19. Let f 2 fn 3; n 2g be odd. Then
1. The conditional probability that a random element g has an n-cycle given that it
satisfies g n D 1 is at least 1=2.


2. The conditional probability that a random element g has an f -cycle given that
it satisfies g 2f D 1 and jg f j D 2 is at least 1=3.
Finally, we highlight some of the results proved in [75] estimating c.n; m/, where
m D rn for a fixed value of r. The proof of this theorem relies on ideas similar to
those outlined above, combined with an idea of Warlimont’s [93] dividing cycles of
permutations into large and small cycles.
Theorem 2.18. For positive integers r; n with r fixed and n sufficiently large,
Â Ã
1 a.r/ 1
c.n; rn/ D C 2 CO 5
n n n2 o.1/

P
where a.r/ D i;j .1 C i Cj /, 1 Ä i; j Ä r 2 , ij D r 2 and r C i; r C j divide rn.
2r
2
Moreover, the conditional probability that an element g Á Sn is an n-cycle, given
a.r/ 1
that its order divides rn, is at least 1 n O n3=2 o.1/ .

2.2.8 Algorithmic Applications of Proportions
in Symmetric Groups

Warlimont’s result is very useful for algorithmic purposes. It tells us that most
permutations g satisfying the equation g n D 1 are n-cycles. Moreover, it also
identifies the cycle structure of the second most abundant set of permutations
satisfying the equation g n D 1; namely permutations which consist of two cycles of
length n=2, and these only occur when n is even. This translates into the algorithm
below to find an n-cycle. Note that the algorithm works in any permutation or matrix
group representation of Sn , where we may not easily recognise the cycle structure
of an element in the natural representation. Such algorithms are called black box
group algorithms; for a formal definition, see Sect. 2.4.2.
Suppose we are given a group G and we believe G might be isomorphic to Sn
under a putative, yet unknown, isomorphism W G ! Sn . We find an element
g 2 G which would map to an n-cycle under with high probability by Algorithm 3
below.

Algorithm 3: FINDNCYCLE
Input: G a group, n 19 an integer, 0 < " < 1 real;
Output: g or fail;
# If the output is g, then g n D 1;
for up to n log." 1 / random elements g 2 G do
if g n D 1 then
return g;
end
end
Return fail;


The algorithm takes as input a real " such that 0 < " < 1, and this input is
used to control the probability of failure. We require that the probability that G is
isomorphic to Sn and the algorithm returns fail to be at most ". Note that on each
random selection, the probability of finding an n-cycle is 1=n. Hence the probability
of failing to find an n-cycle in N."/ random selections is .1 1=n/N."/ and we have
.1 1=n/N."/ < " when N."/ log." 1 /=. log.1 1=n//. In particular, this is
1
the case when N."/ n log." /.
Thus the above algorithm returns with probability at least 1 " an element g 2 G
satisfying g n D 1. Therefore, if G Š Sn then with probability at least 1=2 this
element is an n-cycle, by the above corollary.
Niemeyer and Praeger [74] generalise Warlimont’s result and consider the case
where m n, namely rn Ä m < .r C 1/n for fixed positive integers r.
Algorithm 3 is part of a procedure which decides whether a black box group G is
isomorphic to the full symmetric group Sn for a given natural number n. The full al-
gorithm is described in [9]. First, we have to describe a presentation for the group Sn .
Theorem 2.19 (Coxeter and Moser, 1957).

hr; s j r n D s 2 D .rs/.n 1/
D Œs; r j 2 D 1 for 2 Ä j Ä n=2i

is a presentation for Sn . Moreover, if some group G has generators r; s satisfying
this presentation and r 2 ¤ 1 then G is isomorphic to Sn .
Definition 2.20. The transposition y matches the n-cycle x if y moves two adjacent
points in x.
Lemma 2.21. For n 5, an n-cycle and a matching transposition satisfy the
presentation in Theorem 2.19.
Now we are ready to sketch the algorithm BBRECOGNISESN of [9].

Algorithm 4: BBRECOGNISESN
Input: G D hXi a black box group, n 5;
Output: true and a map W G ! Sn , or fail;
repeat
1. find r 2 G with r n D 1.
# is .r/ an n-cycle?
2. find h 2 G with h2m D 1 where m 2 fn 2; n 3g odd.
# is .hm / a transposition?
3. find a random conjugate s of hm with Œs; s g 6D 1.
# does .s/ interchange two points of .r/?
until repeated too often;
if r or s not found then return fail;
else
define by
• .r/ D .1; : : : ; n/ and
• .s/ D .1; 2/.
Return true and W G ! Sn ;
end


We test whether hr; si Š Sn via the presentation described in Theorem 2.19.
Theorem 2.22. Given a black box group G isomorphic to Sn ; the probability that
the algorithm BBRECOGNISESN.G; n; "/ returns fail is at most ". The cost of the
algorithm is
O..n C n log.n/ / log." 1 //;
where is the cost of ﬁnding a random element in a black box group and the cost
of a black box group operation.

2.2.9 Restrictions on Cycle Lengths

An extensive amount of literature exists on the topic of random permutations whose
cycle lengths lie in a given set L or lie in a particular arithmetic progression. Early
work includes that of Touchard [91], Gonˇ arov [40] and Gruder [42].
c
Let L be a set of natural numbers. Let dL .n/ denote the proportion of elements
in Sn all of whose cycle lengths lie in L and let dL .n; k/ denote the proportion of
elements in Sn with exactly k cycles all of whose lengths lie in L . A generating
function for dL .n/ can be found in [91]. This proportion has been studied by many
authors; we just mention brieﬂy some of Gruder’s results.
Theorem 2.23 (Gruder [42]).

1 X 1
dL .n; k/ D :
kŠ x1 xk
.x1 ;:::;xk /2L k
x1 C Cxk Dn

P za P1
Put H.z/ D a2L a and let D.z/ D nD0 dL .n/zn .
1. D.z/ D exp.H.z//. P Pn
2. D.z/x D exp.xH.z// D 1
nD0 kD0 dL .n; k/x
k n
z .
Bolker and Gleason [13] obtain an explicit asymptotic formula for dL .n/ when
L is an arithmetic progression.
Let pa .n/ denote the proportion of elements in Sn all of whose cycle lengths are
at least a for some a 2.
1. limn!1 1
pa .n/ D exp.1 C 1
2ÁC C 1
a 1 /.
Pn
2. log lima!1 limn!1 1
pa .n/
D ; where D limn!1 1
i D1 i log.n/ is the
Euler constant.
V.F. Kolchin summarises many of the asymptotic results known about this case
in his book [50]. We refer the interested reader to [50] and references therein.


Finally, we mention one particular proportion that has been of considerable
interest in various applications. For positive integers b, let p:b .n/ denote the
proportion of elements in Sn with no cycle of length divisible by b. This proportion
was first studied for primes b in [28], where Erd˝ s and Tur´ n give an explicit
o a
formula for it. This formula immediately generalises to arbitrary positive integers b.
For a prime b, Erd˝ s and Tur´ n also give the limiting distribution of p:b .n/. Many
o a
other authors have also considered this proportion; for example [12], [14, Sect. 2],
[38]. Here we quote a result from [8, Theorem 2.3(b)].
Theorem 2.26. Let n b. Then
Â Ã1=b Â Ã1=b
b .1 1
/ b .1 C n /
2
n
1
Ä p:b .n/ Ä 1
:
n .1 b
/ n .1 b /

Ben-Ezra [10] obtained a similar result for b D 2. A formula for the proportion
of elements in An with no cycle of length divisible by b is also given in [8]. Mar´ ti
o
[61] generalises this, and gives a formula for the proportion of elements of order not
divisible by b in arbitrary permutation groups.
The above estimates have proved to be very useful in deriving proportions of
certain elements in finite classical groups of Lie type. Suppose G is a finite classical
group of Lie type given in natural dimension n with n 2. Using the method
outlined in Sect. 2.3.4, [58] shows that the proportion of elements in G for which
some power is an involution with a large 1-eigenspace of dimension d with n=3 Ä
d Ä 2n=3 is at least c= log.n/ for some constant c.

2.3 Estimation Techniques in Lie Type Groups

We start with a seemingly simple result about permutation groups, discuss the deep
Lie-theoretic analysis underpinning it, and indicate how this approach has led to a
powerful estimation technique for Lie type groups.

2.3.1 p-Singular Elements in Permutation Groups

The following beautiful and surprising result of Isaacs et al. [46] was published in
1995.
Theorem 2.27 (Isaacs, Kantor and Spaltenstein [46]). Let G Ä Sn and let p be a
prime dividing jGj. Then there is at least 1 chance in n that a uniformly distributed
random permutation in G has order a multiple of p.
This result is about any permutation group—not necessarily primitive, nor even
transitive. It is best possible for such a general result, since if n D p then in the


affine group AGL.1; p/ there are exactly p 1 elements of order divisible by p out
of a total of p.p 1/ elements in the group.
The only known proof of Theorem 2.27 requires the finite simple group
classification. The proof strategy is first to make an elementary reduction to the
case where G is a nonabelian simple group. Then the simple groups are dealt with.
There are no difficulties with the alternating groups An and the sporadic simple
groups. This leaves the finite simple groups of Lie type to be considered, and this
is where the authors of [46] “wave a magic wand” with a sophisticated argument
from the theory of Lie type groups. We (Niemeyer and Praeger) were at first baffled
by this proof, as well as fascinated by what it achieved, so set about trying to
understand it. Along the way there was help from Klaus Lux and Frank L¨ beck. u
With Frank L¨ beck we made our first full-blown application of the theory in [58] to
u
estimate the proportion of a certain family of even ordered elements in classical
groups. We discovered that this beautiful theory had been introduced by Gus
Lehrer [54,55] to count various element classes and representation theoretic objects
associated with Lie type groups. Recently Arjeh Cohen and Scott Murray [22]
also used this approach to develop algorithms for computing with finite Lie
algebras.
Our objective became: to formalise the ideas into a framework for estimating
proportions of a wide class of subsets of finite Lie type groups. The framework was
first set out in [58] and in general in [76]. We describe it in the next subsection.

2.3.2 Quokka Subsets of Finite Groups

For a finite group G and a prime p dividing jGj, each group element g can be
written uniquely as a commuting product g D us D su, where u is a p-element and
s is a p 0 -element (that is, ord.u/ is a power of p while ord.s/ is coprime to p). This
is called the Jordan p-decomposition of g.
To find this decomposition write ord.g/ D p a b where p − b and a 0. Then
since p a and b are coprime, there are integers r; t such that rp a C tb D 1. It is
a
straightforward to check that the elements u D gt b and s D g rp have the required
properties, and that u; s are independent of the choices for r; t. This decompo-
sition is critical for defining the kinds of subsets amenable to this approach for
estimation.
Definition 2.28. Let G be a finite group and p a prime dividing jGj. A non-empty
subset Q of G is called a quokka set, or a p-quokka set if we wish to emphasise the
prime p, if the following two properties hold:
(a) Q is closed under conjugation by elements of G.
(b) For g 2 G with Jordan p-decomposition g D us D su, g 2 Q if and only if
s 2 Q.


A natural place to find p-quokka sets is in finite Lie type groups in characteristic
p; for example, in G D GL.n; q/ with q a power of p. Here, in a Jordan
p-decomposition g D us D su, the element u is unipotent and s is semisimple. The
elements u; s are called the unipotent part and the semisimple part of g, respectively.
Some of the subsets already discussed in this chapter turn out to be quokka sets. We
give an example.
Example 2.29. Let G D GL.n; q/ or SL.n; q/, with q a power of p, let e be an
integer such that e > n=2, and suppose that q e 1 has a primitive prime divisor.
Then the subset Q of ppd-.n; qI e/ elements of G is a p-quokka set. To see this, note
that Q is closed under conjugation since conjugate elements have the same order.
Also, for a Jordan p-decomposition g D us D su, a ppd r of q e 1 divides ord.g/
if and only if r divides ord.s/.

2.3.3 Estimation Theory for Quokka Sets

The standard reference for the concepts discussed below is Roger Carter’s book [17],
and an account of the required theory is given in [76].
The groups: We start with a connected reductive algebraic group G defined over
the algebraic closure Fq of the finite field Fq of order q, where q is a power
of a prime q0 . A Frobenius morphism F W G ! G defines a finite group of
Lie type G F D fg 2 G j F .g/ D gg as its fixed point subgroup. We use the
following example to illustrate the concepts as they arise. For the algebraic group
q
G D SL.n; Fq / and Frobenius morphism F W .aij / 7! .aij /, the finite group of Lie
type is G D SL.n; q/, since the fixed field of the map a 7! aq is Fq .
F

Maximal tori: A torus in an algebraic group is a subgroup T that is isomorphic
to a direct product of a finite number of copies of the multiplicative group of Fq . In
particular, T is abelian. A torus T is F -stable if F .T / D T , and T is a maximal
torus if T is closed and not properly contained in another torus. All F -stable
maximal tori in G are conjugate. In our example G D SL.n; Fq /, the subgroup T0
of diagonal matrices in G is a maximal torus that is isomorphic to a direct product
of n 1 copies of .Fq / .
The Weyl group: Choose an F -stable maximal torus T0 in G. The Weyl group W
is defined as the quotient NG .T0 /=T0 . Since F -stable maximal tori are conjugate,
the group W is independent of the choice of T0 . In our example G D SL.n; Fq /,
with T0 the subgroup of diagonal matrices, NG .T0 / is the subgroup of monomial
matrices in G, and W D NG .T0 /=T0 is isomorphic to the group of n n permutation
matrices, so W Š Sn .
F -conjugacy: Elements v; w 2 W are said to be F -conjugate if there is an
element x 2 W such that v D x 1 wF .x/. Notice that we abuse notation a little
in this definition, since x 2 W is a coset x D x0 T0 and by F .x/ we mean F .x0 /T0


(which is well deﬁned since T0 is F -stable). In our example G D SL.n; Fq /,
F -conjugation is ordinary conjugation (since each x 2 W has a representative
monomial matrix with entries 0 or ˙1, and hence x is ﬁxed by F ).
A crucial correspondence and the Quokka Theorem: For an F -stable maximal
torus T of G, the intersection T F D T G F D fg 2 T j F .g/ D gg is
called a maximal torus of G F ; although all F -stable maximal tori of G are
G-conjugate, there are usually several G F -conjugacy classes of F -stable maximal
tori T F , and the structure of the T F is governed by the Weyl group. There is a
1–1 correspondence between G F -conjugacy classes of F -stable maximal tori and
F -conjugacy classes of the Weyl group. This is a crucial ingredient in proving the
main theorem below. Let C be the set of F -conjugacy classes in W , and for C 2 C ,
let TC denote a representative F -stable maximal torus of G F corresponding to C .
F

Theorem 2.30. Let G; F; T0 ; W and C be as above, and let Q G F be a quokka
set. Then
jQj X jC j jT F Qj
D C
:
jG F j jW j jTC j
F
C 2C

Bounds on proportions: Essentially Theorem 2.30 allows us to separate an
estimation problem within a Lie type group G F into two simpler problems, one
within the Weyl group and the other within various maximal tori. The expression for
jQj
jG F j
in Theorem 2.30 as an exact summation can lead to usable bounds. Suppose
that CO is a union of F -conjugacy classes and that `Q is a positive constant such
F
`Q for all C 2 CO. Then Theorem 2.30 implies that
jTC Qj jQj jC O
that F
jTC j jG F j
`Q jW jj .
F
Ä uQ for all C 2 CO, then
jTC Qj jQj jC O
Similarly, if uQ is such that F
jTC j jG F j
Ä uQ jW jj .

A worked example: Let G D SL.n; Fq / and let Q be the quokka set of ppd-
.n; qI e/ elements of G, for some e 2 .n=2; n/—see Example 2.29. We use this
“quokka theory” to re-derive Lemma 2.3. The Weyl group is W Š Sn , and each
F
maximal torus TC containing an element of Q is of the form

TC D Zq e
F
1 other cyclic factors: (2.1)

As we discussed in the last paragraph of the proof of Lemma 2.3, for each such torus,
jT F Qj
C 1
the proportion jT F j lies between 1 eC1 and 1. The F -conjugacy class C in W
C
corresponding to such a torus consists of certain elements of W D Sn containing an
F
e-cycle, and all classes C with this property correspond to tori TC as in (2.29). Let
CO be the subset of W of all elements containing an e-cycle. Then jCOj=jW j D 1=e,
and as we discussed above, jQj=jG F j lies between .1 eC1 / 1 D eC1 and 1 .
1
e
1
e

r-abundant elements: The original impetus to study the work of Isaacs et al. [46]
so closely came from efforts of Niemeyer and Praeger to understand whether, for a
prime r, the lower bound given in [46] for the proportion of r-singular elements in


finite classical groups was close to the true proportion. (An r-singular element is one
with order a multiple of r.) Niemeyer conducted a computer experiment on general
linear groups G D GL.n; p a /, for various dimensions n and primes p and r, where
r divides jGj and r ¤ p, to discover the kinds of r-singular elements in G which
appeared frequently on repeated independent random selections from G. It turned
out that a good proportion of the r-singular elements that we found left invariant,
and acted irreducibly on, a subspace of dimension greater than n=2. Moreover, their
frequency seemed to be roughly proportional to 1=e, where e is the smallest positive
integer such that r divides p ae 1. We decided to call these elements r-abundant. It
seemed at first that the r-abundant elements alone occurred with frequency greater
than the lower bound predicted in [46]. However, it was pointed out to us by
Klaus Lux that, hidden in the proofs in [46] was a lower bound on the proportion
of r-singular elements of the form c=e for some constant c, with e as above. If
e > n=2 then these r-singular elements are the ppd-.n; p a I e/ elements used in
the classical recognition algorithm in [72], and in general r-abundant elements
are as easily recognisable as ppd elements from properties of their characteristic
polynomials: namely, there is an irreducible factor f .x/ of degree greater than n=2
and a multiple of e, such that x has order a multiple of r modulo f .x/ in the
polynomial ring Fpa Œx. A detailed study of r-abundant elements was carried out
by Niemeyer and Praeger with Tomasz Popiel [71] to prove that the experimentally
observed proportion of r-singular elements in general linear groups is correct, and to
find and prove analogues for other finite classical groups. The r-abundant elements
form a quokka set, and their proportion was determined [71, Theorem 1.1] using
Theorem 2.30. For the general linear group GL.n; p a /, the proportion is
Â Ã
1 ln.2/
1 t 1 .r C 1/
r e

with an error term of the form c=n for some constant c, where r t is the largest power
of r dividing p ae 1. It would be interesting to know if r-abundant elements could
be useful algorithmically to identify classical groups. To aid our understanding of
such elements, Sabina Pannek is undertaking a Ph.D. project to find which maximal
subgroups of finite classical groups contain elements with an irreducible invariant
subspace of the natural module of more than half the dimension.

2.3.4 Strong Involutions in Classical Groups

In [53], Leedham-Green and O’Brien introduced a new Las Vegas algorithm to
find standard generators for a finite simple n-dimensional classical group H in odd
characteristic in its natural action. (Recall that a randomised algorithm is called
Las Vegas if the output, if it exists, is always correct; the algorithm may report
failure with a small probability.) The algorithm of [53] proceeds by constructing
recursively various centralisers of involutions (elements of order 2), the details of


Table 2.1 The classical S X n
groups for Theorem 2.31 and
SL.` C 1; q/ GL.` C 1; q/ `C1
Corollary 2.32
SU.` C 1; q/ GU.` C 1; q/ `C1
Sp.2`; q/ GSp.2`; q/ 2`
SO.2` C 1; q/ GO.2` C 1; q/ 2` C 1
SO˙ .2`; q/ GO˙ .2`; q/0 2`

which are discussed further in Sect. 2.4.3. The issue we address here is how to find
an appropriate involution. Leedham-Green and O’Brien wished to work with an
involution whose centraliser would be essentially a product of two smaller classical
groups, each of roughly half the dimension. They called such involutions “strong”:
an involution is strong if its fixed point subspace has dimension in Œn=3; 2n=3/,
or equivalently if its 1-eigenspace has dimension in .n=3; 2n=3. Let I denote
the subset of strong involutions in H . Leedham-Green and O’Brien constructed
elements of I by making independent, uniformly distributed random selections from
H to find an element of even order which powered up to a strong involution. We call
such elements preinvolutions. To estimate how readily a preinvolution can be found
by random selection, we need to estimate the size of the set

P.H; I / D fh 2 H j ord.h/ is even, hord.h/=2 2 I g: (2.2)

Leedham-Green and O’Brien estimated that it would require O.n C n4 log n C
4
n log q/ elementary field operations (that is, additions, multiplications or inver-
sions) to compute a strong involution in H , where is an upper bound on
the number of elementary field operations required to produce an independent,
uniformly distributed random element of H ; see [53, Theorem 8.27]. Underpinning
this complexity estimate was their estimate that the proportion of preinvolutions in
H was at least c=n, for a constant c.
Niemeyer and Praeger, with Frank L¨ beck, used the approach described in
u
Sect. 2.3.3 to obtain an improved estimate for this proportion [58, Theorem 1.1].
They considered any n-dimensional classical group H satisfying S Ä H Ä X ,
where S , X , n are as in one of the lines of Table 2.1 with q odd. Here
GO˙ .2`; q/0 denotes the connected general orthogonal group—the index 2 sub-
group of GO˙ .2`; q/ that does not interchange the two SO˙ .2`; q/-classes of
maximal isotropic subspaces.
Theorem 2.31. Let H satisfy S Ä H Ä X , with S , X , n as in one of the lines of
Table 2.1, with q odd and ` 2, and let I H be the set of strong involutions.
Then
jP.H; I /j 1
:
jH j 5000 log2 .`/
The weak constant of 1=5000 arises from the fact that the estimation only
considered one class of elements that power up to a strong involution, and from the
fact that it determined one constant that is valid uniformly for all classical groups.


A more detailed analysis taking into account a wider family of preinvolutions would
yield a larger value for the constant.
L¨ beck, Niemeyer and Praeger also obtained similar lower bounds for projective
u
groups: note that, for Z0 Ä Z.X /, since the subset I of involutions in Theorem 2.31
contains no central elements, the set I WD IZ0 =Z0 is a subset of involutions in the
projective group H WD HZ0 =Z0 .
Corollary 2.32. With the above notation, jP.H ; I /j=jH j 1=.5000 log2 `/.
Using this new lower bound reduces the complexity of computing a strong
involution in [53] to O.log.n/ C n4 log n C n4 log q/; that is, replacing the first
summand n by log.n/ . It seems to be typical that whenever “quokka theory” is
applicable, it produces superior estimates to more intuitive geometric methods.
In Sect. 2.4.3, the algorithm in [53] will be discussed further. Here we just
mention that the proof of [58, Theorem 1.1] could have been given for a more
general class of involutions called “balanced involutions”. For constants ˛; ˇ such
that 0 < ˛ < 1=2 < ˇ < 1, an .˛; ˇ/-balanced involution in an n-dimensional
classical group H is one with fixed point subspace having dimension in Œ˛n; ˇn/.
The resulting lower bound on the proportion of .˛; ˇ/-balanced involutions in H
would be c= log2 .n/, for a constant c depending only on ˛ and ˇ.

2.3.5 More Comments on Strong Involutions

Before leaving this topic we make some comments about the proof of Theorem 2.31.
First, it is not difficult to see that P.H; I / is a quokka set: it is non-empty since I ¤
;; it is conjugacy closed since I is a union of H -conjugacy classes; and finally, since
q is odd, if g D us D su is the Jordan p-decomposition then gord.g/=2 D s ord.s/=2 ,
and hence g 2 P.H; I / if and only if s 2 P.H; I /.
To obtain the lower bound in Theorem 2.31 we used Theorem 2.30. A special
subset C0 of F -conjugacy classes of W was examined, for which it was possible
both to estimate w0 WD j [C 2C0 C j=jW j and to find a good positive lower bound on
jTC P.H; I /j for each C 2 C0 . To give an understanding of this subset of W ,
F

while avoiding the technicalities associated with small dimensions and the other
types of classical groups, we confine our attention to H D GL.n; q/ with n 7.
Here C0 is a set of conjugacy classes in W D Sn . We choose a particular positive
integer a as follows, and take W0 WD [C 2C0 C to consist of all permutations with a
single cycle of length 2a k 2 .n=3; 2n=3, for some integer k, and no other cycle of
length divisible by 2a . For a0 D log2 ln 2 C log2 log2 n, we take a to be the integer
in the interval Œa0 1=2; a0 C 1=2/. We note for later use that, since n 7, we have
a 1 and .13=4/ 2a Ä n.
First we show that jTC P.H; I /j=jTC j
F F
1=2, for C 2 C0 with a cycle of
a F
length 2 k as above. Each torus TC in the H -conjugacy class of tori corresponding
a
to C is of the form Z A, where Z is cyclic of order q 2 k 1 leaving invariant
a subspace U of dimension 2a k and acting as a Singer cycle on U , and for each
x 2 A, the 2-part of ord.x/ (that is, the highest power of 2 dividing ord.x/) is


a
strictly less than the 2-part of q 2 k 1. Now half of the elements z 2 Z are such
a
that the 2-part of ord.z/ is equal to the 2-part of q 2 k 1, and for each such z, and
any x 2 A, the element zx has even order, and .zx/jzxj=2 is the unique involution z0
in Z. The element z0 acts as I on the subspace U and has ﬁxed point subspace
of dimension n 2a k 2 Œn=3; 2n=3/; that is to say, z0 is a strong involution and
zx 2 P.H; I /. Thus jTC P.H; I /j=jTC j 1=2.
F F

Theorem 2.30 now implies that

jP.H; I /j 1 jW0 j
;
jH j 2 jW j

so it remains to estimate the size of W0 . A straightforward counting argument yields

jW0 j X p:2a .n 2a k/
D ; (2.3)
jW j 2a k
k

where the sum is over integers k such that n=3 < 2a k Ä 2n=3, and p:2a .n 2a k/
is the proportion of elements in Sn 2a k with no cycle of length divisible by 2a . By
Lemma 4.2(a) of [58], which is based on Theorem 2.26,
1 1=2a 1 1=2a
p:2a .n 2a k/ > .n 2a k/ > n :
4 4
a
Thus each summand in (2.3) is at least 3=.8n1C1=2 / since 2a k Ä 2n=3. The number
of summands in (2.3) is at least .2n=3 n=3/=2a 1 D n=.3 2a / 1, which is at
least n=.39 2a / (since .13=4/ 2a Ä n). Hence

jP.H; I /j 1 jW0 j 1 n 3 1 1
D ;
jH j 2 jW j 2 39 2a 8n1C1=2a 208 2a n1=2a

which is greater than 208 3 log .n/ D 624 log .n/ . This proves Theorem 2.31 for H D
1 1 1
2 2
GL.n; q/.
The family W0 of elements of the Weyl group W gives a far better lower bound
than bounds obtained by geometric arguments. However we have not considered all
conjugacy classes in W , and indeed it seems that, for this problem, application of
“quokka theory” does not yield an upper bound. It is reasonable to ask how good
the lower bound of Theorem 2.31 is. To attempt to answer this question, we quote a
few sentences from [58, p. 3399].
We did some numerical experiments for small q 2 f3; 5; 9; 13g and groups from the
theorem up to dimension 1000. We computed many pseudo-random elements and checked
if they powered up to an involution with a ﬁxed point space of dimension in the right
range. The proportion of these elements is not a monotonic function in the dimension,
but the trend was that the proportion was about 25% for small dimensions and went
down to about 15% in dimension 1000 (independently of the type of the group and q).
Further, statistical tests on the data from the groups H we sampled strongly indicates that
P .H; I /=jH j D O.1= log.`//. This seems to suggest at least that we cannot expect that
there is a lower bound independent of the rank of the group.


2.3.6 Regular Semisimple Elements and Generating Functions

Let H be an n-dimensional classical group in odd characteristic, as in one of the
lines of Table 2.1. The methods described in Sect. 2.3.4 show how to find a strong
involution efficiently, or more generally, how to find an .˛; ˇ/-balanced involution z.
The problem of constructing the centraliser CH .z/ of such an involution will be
discussed in Sect. 2.4. In this section we explore an estimation problem connected
with part of the construction. An essential component in finding CH .z/ is to take
random conjugates zg to find a “nice product” y WD zzg , where “nice” means “close
to regular semisimple”. This procedure is discussed in the seminal paper [78] by
Christopher Parker and Rob Wilson. They estimate that O.n/ random products will
produce a nice product with high probability. The approach taken by Praeger and
Seress [86], and described in this section, shows that only O.log n/ random products
are required.
Written in an appropriate basis, the product y D zzg of an involution z and a
random conjugate zg of z has the following form, where y0 has no ˙1-eigenvectors:
0 1
Ir 0 0
y WD zzg D @ 0 y0 0 A :
0 0 Is

Typically, the dimension r is close to 2m n, where m is the maximum of the
dimensions of the ˙1-eigenspaces of z, and s is close to 0. The question arises:
what kind of matrix do we expect for y0 “typically”? Let us restrict attention to
the simplest case where H D GL.n; q/ with q odd. By considering the results
of computer experiments on various .˛; ˇ/-balanced involutions and their random
conjugates for various n and odd q, we discovered that often y0 is “regular
semisimple”. For the following discussion, let us assume that y D y0 .
An element y of GL.n; q/ is called semisimple if is diagonalisable over some
extension field of Fq (see [17, p. 11]), and this is equivalent to its minimal
polynomial my .t/ being multiplicity free. Also y is called regular if its centraliser in
the corresponding general linear group over the algebraic closure of Fq has minimal
possible dimension, namely n (see [17, p. 29]). It turns out that an element y of a
general linear group is regular if and only if my .t/ D cy .t/, where cy .t/ denotes the
characteristic polynomial of y. These two conditions for elements of finite classical
groups are discussed and compared in [69, Note 8.1]. The regular semisimple
elements are those which are both regular and semisimple. In fact, for elements y of
H D GL.n; q/, y is regular semisimple if and only if the characteristic polynomial
cy .t/ for its action on V .n; q/ satisfies

cy .t/ D a product of pairwise distinct irreducible polynomials.

Looking into the analysis of this situation in the paper [78], it is clear that Parker and
Wilson recognised that regular semisimple elements y occur frequently. Moreover,


the proportion of regular semisimple elements in the full n-dimensional matrix
algebra was estimated by Neumann and Praeger [70]. The main result of [86] (cf.
Theorem 2.34) is a strengthening of the estimates in [70, 78].
The characteristic polynomial cy .t/ has two special properties: firstly, when
y D y0 the element y has no ˙1-eigenvectors, so cy .t/ is not divisible by t ˙ 1.
Secondly, since y z D z 1 .zzg /z D zg z D y 1 , the characteristic polynomials of
y and y 1 are equal. Now cy 1 .t/ D cy .t/ is the conjugate polynomial of cy .t/
where, for an arbitrary polynomial f .t/ with f .0/ ¤ 0, its conjugate polynomial is
f .t/ WD f .0/ 1 t degf f .t 1 /. Thus cy .t/ D cy .t/ is self-conjugate. We have seen
that conjugation by z inverts y, and similarly conjugation by zg inverts y. Inverting
a regular semisimple matrix pins down the conjugacy class of the involution z, as
shown in [86, Lemma 3.1]. For n even and q odd, let C Â GL.n; q/ denote the the
conjugacy class of involutions with fixed point space of dimension n=2.
Lemma 2.33. Let z; y 2 GL.n; q/ with q odd, such that y is regular semisimple
with characteristic polynomial cy .t/ coprime to t 2 1, and z is an involution
inverting y. Then n is even, z 2 C , and zy is also an involution which inverts y.
By Lemma 2.33, we have a bijection .z0 ; z/ 7! .y; z/ between the sets
ˇ
ˇ y WD zz0 regular semisimple
X D .z; z0 / 2 C C ˇ
ˇ with cy .t/ coprime to t 2 1

and 8 ˇ 9
< ˇ y; z 2 GL.n; q/; z2 D 1; y z D y 1
=
ˇ
Y D .y; z/ ˇ y regular semisimple, and :
: ˇ ;
ˇ c .t/ coprime to t 2 1
y

The set X is relevant for algorithmic purposes, while the set Y is more amenable
to estimation techniques. For the algorithm, we are given (that is to say, we have
already found) the involution z 2 C , and we want to know the proportion of z0 2 C
such that .z; z0 / 2 X . This is

jfz0 2 C j .z; z0 / 2 X gj jX j jY j j GL.n; q/j jY j
D D D
jC j jC j 2 jC j 2 jC j 2 j GL.n; q/j
4
and the first factor on the right of the equality, namely j GL.n;q/j D j GL.n=2;q/j , lies
jC j2 j GL.n;q/j
between .1 q 1 /7 and .1 q 1 /2 . Thus the essential problem is to estimate

jY j
ss.n; q/ WD :
j GL.n; q/j

Parker and Wilson [78] give a heuristic that estimates this quantity as being at least
c=n if we require in addition that y has odd order. Our approach gives a surprisingly
precise answer; see [86, Theorem 1.2]. Since n is even we consider ss.2d; q/.


Theorem 2.34. For a fixed odd prime power q, the limit of ss.2d; q/ as d ! 1
exists and
ss.1; q/ WD lim ss.2d; q/ D .1 q 1 /2 :
d !1
p
Moreover jss.2d; q/ ss.1; q/j D o.q0 d / for any q0 such that 1 < q0 < q.
Corollary 2.35. There exists c > 0 with the property that for any z 2 C the
proportion of z0 2 C such that .z; z0 / 2 X is bounded below by c.
We use generating functions discussed in Sect. 2.2.5 to study the quantities
ss.2d; q/. We define

X
1
S.u/ D ss.2d; q/ud where ss.0; q/ D 1:
d D0

Since y is regular semisimple, cy .t/ is multiplicity-free, and since y is inverted by
the involution z, we have a factorisation
! 0 s 1
Y r Y
cy .t/ D fi .t/ @ gj .t/gj .t/A (2.4)
i D1 j D1

where each fi D fi has even degree, and each gj ¤ gj , with the fi ; gj ; gj
pairwise distinct monic irreducibles. We use this decomposition to find in [86,
Lemma 3.2] that the number of pairs .y 0 ; z/ 2 Y such that y 0 has characteristic
polynomial cy .t/ is equal to

j GL.2d; q/j
Qr 1
Á Q Á:
deg fi s deg gj
i D1 .q 1/ j D1 .q 1/
2

Summing over all possible cy .t/ gives an expression for ss.2d; q/j GL.2d; q/j.
Comparing the expression we obtain for ss.2d; q/ by this process with the
coefficient of ud in the infinite product
! Â Ã
Y 1
u 2 deg f Y udeg g
1C 1
1C ;
q 2 deg f 1 q deg g 1
f Df ; irred: fg;g g; g¤g ; irred:

we see that the two expressions are the same. Hence S.u/ is equal to this infinite
product. The contribution to the infinite product from each irreducible polynomial
f or conjugate pair fg; g g of non-self-conjugate polynomials depends only on the
degrees of the polynomials. Thus

YÂ um
ÃN .qI2m/ YÂ um
ÃM .qIm/
S.u/ D 1C 1C (2.5)
m 1
qm 1 m 1
qm 1


where the exponents are

N .qI m/ D # monic irreducible self-conjugate polynomials over
Fq of degree m:
M .qI m/ D # (unordered) conjugate pairs of monic irreducible
non-self-conjugate polynomials over Fq of degree m:

It turned out that a somewhat similar infinite product arose when Praeger was
studying separable matrices in finite unitary groups with Jason Fulman and Peter
Neumann in [36]. A similar analysis to that given in [36] for these matrices yielded:
1. S.u/ is analytic for juj < 1 with a simple pole at u D 1.
p
2. S.u/ D .1 u/ 1 H.u/, with H.u/ analytic for juj < q.
Completing the analysis we found the asymptotic behaviour of the ss.2d; q/, as
in Theorem 2.34.

2.4 Computing Centralisers of Involutions

The results in the previous section play a significant role in the analysis of algo-
rithms to compute centralisers of involutions. In general the problem of computing
centralisers is of great importance in theoretical computer science and in group
theory. In computer science, the main interest stems from the connection with the
graph isomorphism problem.
Problem 2.36. (ISO) Given: graphs 1 .V1 ; E1 / and 2 .V2 ; E2 /.
Find: an edge-preserving bijection between V1 and V2 , or prove that no such
bijection exists.
ISO is polynomial-time reducible to the following computational problems with
permutation groups.
Problem 2.37. (STAB) Given: a permutation group G Ä Sym.˝/ and a subset
Â ˝.
Find: the set stabiliser StabG . / D fg 2 G j g D g.
Problem 2.38. (INT) Given: permutation groups G; H Ä Sym.˝/.
Find: the intersection G H .
Problem 2.39. (CENT) Given: permutation groups G; H Ä Sym.˝/.
Find: the centraliser CG .H / D fg 2 G j hg D h for all h 2 H g.
Problems 2.37–2.39 are in the same class of the complexity hierarchy, which
means that they can be reduced to each other in time polynomial in the input
length [59].


The reduction of ISO is easiest to STAB or INT. First, we notice that 1 .V1 ; E1 /
and 2 .V2 ; E2 / are isomorphic if and only if 1 [ 2 (disjoint copies of 1 and 2 )
has an automorphism that exchanges V1 and V2 . Therefore, it is enough to compute
automorphism groups of graphs. Given a graph .V; E/, define ˝ as the set of
unordered pairs in V . Then E corresponds to a subset Â ˝, and Sym.V / acts as
a group G on ˝. We can compute Aut. / as Aut. / D StabG . / or Aut. / D
G .Sym. / Sym.˝ n //.
Although, using backtrack methods (see e.g. [87, Chap. 9]), ISO and CENT
are usually easy to solve in practice, no polynomial-time solution is known for
Problems 2.36–2.39. Special cases with polynomial-time solutions are of great
theoretical and practical interest.
In group theory, the most important case of centraliser computations is to
construct centralisers of involutions. On the theoretical side, a major tool in the
study and classification of finite simple groups is the investigation of their involution
centralisers [41]. On the computational side, in the last decade involution centraliser
computations became prevalent [1, 7, 45, 53, 56, 78]. In the next subsections, we
describe some applications of centraliser computations; Bray’s algorithm [16] for
computing centralisers of involutions; and efforts to analyze Bray’s algorithm.

2.4.1 Applications of Centralisers of Involutions Computations

A recent active area of computational group theory is the so-called matrix group
recognition project. Let V be a finite dimensional vector space over a finite
field Fq . Given G D hS i Ä GL.V /, the goal is to compute quantitative and
structural information about G such as the order, a composition series, and important
characteristic subgroups like the largest solvable normal subgroup of G.
There are two main approaches to matrix group recognition. The geometric
approach, initiated by Neumann and Praeger [69] and currently led by Leedham-
Green and O’Brien [52,77], is based on Aschbacher’s classification of matrix groups
[2]. Aschbacher defines nine categories of matrix groups G. In seven of these
categories, there is a natural normal subgroup N C G that can be used to divide
the recognition problem into two smaller subproblems on N and G=N . Based on
that result, the geometric approach tries to find a homomorphism ' W G ! H into
an appropriate permutation or matrix group H , and recursively recognise Im.'/
and Ker.'/. In contrast, the black-box group approach of Babai and Beals [4] aims
for the abstract group theoretic structure of G. Babai and Beals define a series of
characteristic subgroups, present in all finite groups, and initiate a program that
tries to compute a composition series going through these characteristic subgroups.
Both approaches eventually lead to simple (or quasisimple) matrix groups, where
further divide-and-conquer is impossible. For such groups, a major issue is the
solution of the constructive membership problem.


2.4.2 Constructive Membership in Lie Type Groups

Definition 2.40. A black-box group G is a group whose elements are encoded by
bit strings (strings consisting of 0s and 1s) of uniform length. Moreover, there are
oracles for the following tasks. Given strings representing g; h 2 G, we can compute
a string representing gh; a string for g 1 ; and we can decide whether g D 1.
A black-box algorithm is an algorithm that, given G by a set of generators, uses
only the black-box oracles.
The definition of black-box groups covers the “concrete” representations of
groups as permutation groups or matrix groups defined over finite fields. Note that if
G is a black-box group and N is a recognisable normal subgroup (i.e., given a string
representing some g 2 G, we can decide whether g 2 N ), then G=N is also a black-
box group. This observation plays a crucial role in recursive algorithms, allowing
us to work in factor groups. Also note that we require only that N is recognisable,
but N is not necessarily constructed (i.e., we may not have a generating set for N
in hand). Examples of recognisable normal subgroups that may be hard to construct
are the centre and the largest soluble normal subgroup of G. Black-box groups were
introduced by Babai and Szemer´ di [5]. For an introduction to the basic black-box
e
group algorithms, see [87, Chap. 2].
A black-box group algorithm does not use specific features of the group
representation, nor particulars of how group operations are performed. For example,
we lose all information stored implicitly in the cycle structure of a permutation, or
in the characteristic polynomial of a matrix. In practice, and also in some theoretical
considerations, we often allow oracles for some other operations; an example is an
oracle to compute element orders.
The very reasonable and justified question arises: why do we handicap ourselves
with black-box group algorithms? One answer is that in certain situations, we cannot
do more than the black-box operations. For example, to generate random elements
in a matrix group, so far every algorithm takes repeated products and inverses of
the given generators, and after a while declares the last element constructed as a
random element of the input group [3, 18, 24]. Bray’s algorithm (see Sect. 2.4.4)
for computing centralisers of involutions is another example of a black-box group
algorithm, with a possible enhancement using element order oracles. Another, more
unusual answer is that elements of a permutation group can be described as unique
words in a strong generating set (SGS), constructed in a canonical way. The group
operations are performed using the images of elements of the base associated with
the SGS. For the important class of small-base groups, these group operations
are much faster than permutation multiplication, but the algorithms using this
representation are strictly black-box. For details, we refer to [87, Chap. 5.4].
Next, we define the notion of a straight-line program (SLP). Expressing elements
of a group G in a given set of generators may result in words of length proportional
to jGj; intuitively, SLPs are shortcuts, to reach group elements faster from a set of
generators. By [5], every g 2 G can be reached from any set of generators by an
SLP of length at most .1 C log jGj/2 .


Definition 2.41. Given G D hS i and g 2 G, a straight-line program (SLP)
reaching g from S is a sequence of expressions W D .w1 ; : : : ; wm / such that, for
i D 1; 2; : : : ; m,
1. wi is a symbol for some s 2 S ; or
2. wi D .wj ; wk / for some j; k < i ; or
3. wi D .wj ; 1/ for some j < i .
We define the evaluation of W the natural way: eval.wj ; wk / D eval.wj /eval.wk /
and eval.wj ; 1/ D eval.wj / 1 ; and require that eval.wm / D g.
Finally, we are ready to define the constructive membership problem.
Definition 2.42. A constructive membership algorithm for a group G is a black-
box group algorithm that, given the black-box group G D hS i and g 2 G,
constructs an SLP reaching g from S .
The main result of this subsection is the following theorem by Holmes et al. [45].
Theorem 2.43 ([45]). Let G be a black-box group equipped with an order oracle.
There is a black-box Monte Carlo algorithm which reduces the constructive
membership problem for G to three instances of the same problem for centralisers
of involutions of G.
Proof. Let G D hS i and g 2 G. An algorithm constructing an SLP reaching g from
S consists of the following steps.
1. Find h 2 G with ord.gh/ D 2`. Define z WD .gh/` .
2. Find an involution x 2 G with ord.xz/ D 2m. Define y WD .xz/m .
3. Construct X D CG .x/.
4. Solve the constructive membership problem for y 2 X .
5. Construct Y D CG .y/.
6. Solve the constructive membership problem for z 2 Y .
7. Construct Z D CG .z/.
8. Solve the constructive membership problem for gh 2 Z.
9. Compute and return an SLP for g.
To prove the correctness of the algorithm, observe that z, constructed in Step 1, is
an involution centralising gh. In Step 2, y is in the centre of the dihedral group
hx; zi, so x is an involution centralising y and y is an involution centralising z.
Hence Steps 3, 5 and 7 compute centralisers of involutions, and the constructive
membership problems in Steps 4, 6 and 8 indeed try to reach elements of G that
are in the appropriate subgroups. Finally, note that the construction of x provides
an SLP reaching x from S and, consequently, we have SLPs reaching y, then z,
then gh from S . Also, in Step 1, we construct an SLP reaching h from S . Hence, in
Step 9, we can construct an SLP reaching g from S . t
u
Remark 2.44. We note that the hypothesis of Theorem 2.43 that G has an order
oracle can be relaxed. The only places in the algorithm where the order oracle is
used are in Steps 1 and 2. For example, at the construction of z in Step 1, we


can proceed the following way. Instead of computing `, we can raise gh to an
appropriate multiple of the odd part of jGj. To find such a multiple (without knowing
jGj), it is enough to know a superset of primes occurring in jGj or, in the case of
a matrix group G Ä GL.n; q/, we can work with the set of pseudoprimes: these
are the largest divisors of the numbers q e 1 for e Ä n, that are relatively prime
to q j 1 for all j < e. The pseudoprimes can be computed in polynomial time
(polynomial in terms of n and log q). For details, see [4]. The use of the order oracle
in Step 2 can be avoided in exactly the same way.
In [45], Holmes et al. show that if G is a simple group of Lie type then the
algorithm described in Theorem 2.43, not counting the time requirement of Steps 4,
6 and 8, runs in polynomial time. However, we cannot apply the theorem recursively
to the groups in these steps, because they are not simple. Therefore, we need a
recursive scheme involving all groups, not only the simple ones. Such a scheme is
designed by Babai et al. in [7]; Theorem 2.43 is a crucial ingredient in the following
result.
Theorem 2.45 ([7]). There is a randomised polynomial-time algorithm, employing
certain number-theoretical oracles, which, given a matrix group G Ä GL.n; q/ of
odd characteristic, solves the constructive membership problem in G.
The required number-theoretical oracles are the factorisation of integers of the
form q e 1, for 1 Ä e Ä n, and the solution of the discrete logarithm problem:
given a; b 2 Fq e , decide whether a 2 hbi; and, if the answer is affirmative, then
find an integer x such that a D b x . In polynomial-time algorithms for matrix
groups, it is customary to assume the use of these number theory oracles as they
are already needed in finding a composition series and the order of a 1 1 matrix
group over Fq . We note that Theorem 2.45 extends to matrix groups defined over
fields of characteristic 2, with some restrictions on the composition factors of G. It
is expected that these restrictions will be removed in the near future, as constructive
membership algorithms in all simple groups are in the offing.

2.4.3 Constructive Recognition of Lie Type Groups

Membership testing is an important first step in exploring a permutation or matrix
group G; however, for studying the structure of G and constructing important
subgroups, it is beneficial to identify the composition factors of G with standard
copies of these factor groups. For alternating and classical groups, the standard copy
is the natural permutation and matrix representation, respectively. For exceptional
groups, the definition of a standard copy is not so clear-cut: we may choose the
smallest-dimensional matrix representation, or a Bruhat decomposition, or any other
representation we may be able to control. Here we only give a formal definition for
classical groups, taken from [48].


Definition 2.46. Constructive recognition of a black-box group G D hS i isomor-
phic to a simple classical group defined on some vector space over a field of given
characteristic p is an algorithm that verifies that there is, indeed, an isomorphism,
and finds the following:
(i) The field size q D p e ; as well as the type and the dimension d of G.
(ii) A new set S generating G; a vector space Fd , and a monomorphism W G !
q
PSL.d; q/; specified by the image of S ; such that G acts projectively on Fd
q
as a classical group defined on Fd .
q

Moreover; the data structures underlying (ii) yield deterministic algorithms for each
of the following:
(iii) Given g 2 G; find g and a straight-line program from S to g.
(iv) Given h 2 PGL.d; q/; decide whether or not h 2 G I and; if it is; find h 1
and a straight-line program from S to h 1 .
(v) Find a form on Fd involved in the definition of G as a classical group; if G 6Š
q
PSL.d; q/.
Although Definition 2.46 is formulated in the general context of black-box
groups, of course it can be applied to any given permutation or matrix representation
of G. The simplest but most important case is when G is already given in its
natural representation, and the only task is to find “nice” generators S such that
each element of G can be reached easily from S . For classical groups of odd
characteristic, this task has been accomplished by Leedham-Green and O’Brien by
a highly efficient algorithm [53]. A rough outline of their procedure is given in
Algorithm 5.

Algorithm 5: CONSTRUCTIVERECOGNITION
Input: G D hSi Ä GL.V / Š GL.n; q/, q odd, G is a classical group in its natural
representation;
Output: A data structure for constructive recognition of G;
.1/ repeat
y WD random element of G;
until ord.y/ is even and x WD y ord.y/=2 has ˙1-eigenspaces E1 ; E 1 with
dim.Ei / 2 .n=3; 2n=3/;
.2/ Construct H D CG .x/;
.3/ Recursively solve constructive recognition for the restriction of H to its action on E1
and E 1 ;
.4/ Use the result of Step .3/ to obtain nice generators and data structure for constructive
recognition of G;

The following simple lemma from [84] implies that Step .3/ is indeed a recursive
call.


Lemma 2.47. Let G, x, E1 , E 1 be as in Algorithm 5, with G classical but not
linear. Then V D E1 ? E 1 , and both E1 and E 1 are nondegenerate (and of even
dimension if G is symplectic).
Proof. For u 2 E1 and w 2 E 1 , we have .u; w/ D .u; w/x D .u; w/ and hence
.u; w/ D 0. Thus E1 Â E ?1 . Since the bilinear form is nondegenerate, dim.E1 / D
n dim.E 1 / D dim.E ?1 / and hence E1 D E ?1 . Therefore, E 1 E ?1 D 0 so
E 1 , and similarly also E1 , are nondegenerate. In particular, E1 and E 1 both have
even dimension if G is symplectic. t
u
Since, for i 2 f1; 1g, x acts as a scalar matrix on Ei , Lemma 2.47 implies that
the restriction of H to Ei is a classical group of the same type as G and Step .3/ is
indeed a recursive call. Note that the requirement dim.Ei / 2 .n=3; 2n=3/ ensures
that CG .x/ can be split into two parts of roughly equal size, thereby ensuring that
the depth of the recursion is logarithmic in n.
To analyze Algorithm 5, for the ﬁrst two steps we have to estimate (i) the
proportion of elements y as in Step (1); and (ii) give a running time estimate for the
construction of involution centralisers. Task (i) has been accomplished in Sect. 2.3.4.
In the next two subsections, we describe and analyze an algorithm for computing
involution centralisers.

2.4.4 Computation of an Element Centralising an Involution

In this subsection we describe an algorithm by Bray [16] that constructs an element
in the centraliser of a given involution.

Algorithm 6: CENTRALISINGELEMENT
Input: G D hSi and an involution x 2 G;
Output: An element of CG .x/;
.1/ g WD random element of G;
.2/ y WD x x g and m WD ord.y/;
.3/ if m is even then
return .g/ WD y m=2
else
return .g/ WD y .mC1/=2 g 1
end

We note that the order computation in Step (2) may be avoided, using a superset
of primes occurring in G, or pseudoprimes (see Remark 2.44).
Lemma 2.48. The output of Algorithm 6 is correct: no matter which g 2 G is
chosen in Step .1/, we have .g/ 2 CG .x/.


Proof. For any g 2 G, the group D WD hx; x g i is dihedral, of order 2m. If m is
even then .g/ 2 Z.D/; in particular, .g/ centralises x 2 D.
If m is odd then, using that x 2 D 1, we obtain
mC1
m 1 mC1
xy D .xg 1 xg/ x.xg 1 xg/ D xg :
2
2 2

mC1
Comparison of the leftmost and rightmost terms gives .g/ D y 2 g 1
2 CG .x/.
t
u
We say that g 2 G is of even type if y D xx g has even order, and g 2 G is of
odd type if y D xx g has odd order. Note that for any c 2 CG .x/, x cg D x g , so
xx g D xx cg and consequently g and cg have the same type. Moreover, .xx g /c D
xx gc so xx g and xx gc are conjugate, have the same order, and g and gc have the
same type. Combining the last two observations, we obtain that in a double coset
CG .x/ g CG .x/, all elements have the same type.
Lemma 2.49. (i) If g is chosen from the uniform distribution on the set of odd type
elements of G then .g/ is a uniformly distributed random element of CG .x/.
(ii) If g is chosen from the uniform distribution on the set of even type elements
of G and .g/ is in the conjugacy class C of involutions in CG .x/ then .g/ is
a uniformly distributed random element of C .
mC1
Proof. (i) Suppose that g is of odd type. For c 2 CG .x/, we have y 2 .cg/ 1 D
mC1
y 2 g 1 c 1 and so .cg/ D .g/c 1 . Hence, as cg runs through the coset
mC1
CG .x/ g, y 2 g 1 c 1 runs through CG .x/. This implies that if g runs through
the elements of G of odd type then each element of CG .x/ occurs as .g/
exactly the same number of times.
(ii) Suppose now that g is of even type. Then .g/ D .xx g /m=2 is an involution; let
C denote its conjugacy class in CG .x/. As gc runs through the coset g CG .x/,
.gc/ D .xx gc /m=2 D ..xx g /m=2 /c covers each element of C the same number
of times. Hence each element of a ﬁxed conjugacy class C of involutions in
CG .x/ has the same chance to occur as .g/ for some g of even type. t
u

2.4.5 Computation of the Full Centraliser

In order to compute a set X of generators of CG .x/ for a given group G and
involution x 2 G, we may construct a sequence .g1 ; : : : ; gm / of random elements
in G and take X WD f .gi / j 1 Ä i Ä mg. By Lemma 2.48, we always have
hX i Ä CG .x/, but when can we stop? How large should m be so that, with high
probability, X generates the entire group CG .x/?
By Lemma 2.49, random elements gi of odd type are highly desirable, since then
.gi / is a uniformly distributed random element of CG .x/. Such a random element


.gi / 2 CG .x/, added to an already constructed proper subgroup H < CG .x/,
increases H with probability 1 1=jCG .x/ W H j 1=2, so if we know an
upper bound ` for the length of subgroup chains in CG .x/ then we may estimate
how many elements gi of odd type we need to encounter. For polynomial-time
computations, the trivial bound ` Ä log2 jGj suffices, but sometimes we have much
better estimates for the number of required random generators. In particular, in the
especially important case when G is a simple group of Lie type defined over a field
of odd characteristic, the structure of involution centralisers is known. Consequently,
for any involution x 2 G, the number of uniformly distributed random elements
needed to generate CG .x/ with probability greater than 1 " can be bounded by a
function of ", independent of G and x [57]. Therefore, the following seminal result
of Parker and Wilson [78] has great importance in the analysis of many matrix group
algorithms.
Theorem 2.50 ([78]). There exists a positive constant c such that:
(i) If G is a simple exceptional group of Lie type defined over a field of odd order,
and x is any involution in G, then the probability that a uniformly distributed
random element g 2 G is of odd type is bounded below by c.
(ii) If G is a simple classical group defined over a field of odd order, with natural
module of dimension n, and x is any involution in G, then the probability that a
uniformly distributed random element g 2 G is of odd type is bounded below by
c=n. Moreover, the order of magnitude 1=n for a lower bound is best possible.
Parker and Wilson [78, p. 886] give an indication of how big the constants can
be: “The constants c that can be obtained from our proofs are of the order of 1=1000,
but we have made no attempt to calculate them explicitly, as we conjecture that the
best possible constants are nearer 1=4.”
The basic idea of the proof of Theorem 2.50 is to identify a set of dihedral
subgroups D of twice odd order in G, each D containing the given involution x.
If the random conjugate x g falls into one of these subgroups D then xx g has odd
order and g is of odd type. In order to avoid double counting, we also require that
generators of the maximal cyclic normal subgroup of D be regular semisimple in a
suitable subgroup H Ä G. (Here H depends on D but H is also of Lie type. We
require the generators of D to be regular semisimple as elements of this Lie type
group, as defined in Sect. 2.3.6.)
While Theorem 2.50 is sufficient to prove polynomial running time of centraliser
of involution computations in Lie type simple groups, the scarcity of elements of
odd type raises the the following questions. Is there an algorithm that uses the
lower quality random elements .gi / 2 CG .x/, obtained from gi of even type,
to generate CG .x/? Can the asymptotic running time of this algorithm be faster than
the construction of CG .x/ using the uniformly distributed .gi / obtained from gi of
odd type? To formulate this problem precisely, we need some definitions.
We consider finite classical groups H of dimension n over a finite field Fq of
odd order q. We denote by H the generalized Fitting subgroup of H (for example
H D SL.n; q/ if H D GL.n; q/). Let ˛; ˇ be real numbers such that 0 < ˛ <


1=2 < ˇ < 1, and let x 2 H be of order 2. Recall that x is called an .˛; ˇ/-
balanced involution in H if the subspace E1 .x/ of fixed points of x in the underlying
vector space has dimension r where ˛n Ä r < ˇn. For a given sequence X D
.C1 ; : : : ; Cm / of conjugacy classes of .˛; ˇ/-balanced involutions in H , a c-tuple
.g1 ; : : : ; gm / is a class-random sequence from X if gi is a uniformly distributed
random element of Ci for each i D 1; : : : ; m, and the gi are mutually independent.
Given a classical group G Ä GL.n; q/ and an involution x 2 G, the centraliser
CG .x/ modulo x is the direct product of two classical groups H .1/ and H . 1/ , acting
on E1 .x/ and E 1 .x/, respectively. If g 2 G is of even type then .g/ acts as
an involution g .J / on EJ , for J 2 f1; 1g, and if .g1 ; : : : ; gm / is a sequence of
uniformly distributed random elements of even type in G then Lemma 2.49 implies
.J / .J /
that .g1 ; : : : ; gm / is a class-random sequence from some conjugacy classes of
.J / .J /
involutions X .J / D .C1 ; : : : ; Cm /.
With an application in Algorithm 5 in mind, we propose the following problems.
We use the notation and definitions of the previous paragraphs.
Problem 2.51. Given a classical group G Ä GL.n; q/ and a .1=3; 2=3/-balanced
involution x in G, estimate the probability p that for a uniformly distributed g 2 G
of even type, g .J / is an .˛; ˇ/-balanced involution in H .J / , for both J 2 f1; 1g.
Here ˛; ˇ are constants, chosen appropriately.
Problem 2.52. Let G Ä GL.n; q/ be a classical group and let X D .C1 ; : : : ; Cm /
be a sequence of conjugacy classes of .˛; ˇ/-balanced involutions in G. Estimate
the minimum value of m such that, with high probability, a class-random sequence
from X generates a subgroup of G containing G .
If the product .1=p/m, for the probability p from Problem 2.51 and the minimum
value m from Problem 2.52, satisfies .1=p/m D o.n/ then the elements .g/
obtained from even type g generate CG .x/ asymptotically faster than the elements
.g/ obtained from odd type g.
Problem 2.52 has been solved for all classical groups.
Theorem 2.53 ([84]). Let ˛; ˇ be real numbers such that 0 < ˛ < 1=2 < ˇ < 1.
Then there exist integers m D m.˛; ˇ/ and n.˛; ˇ/ such that, for G, n, q as above,
with q odd, if n > n.˛; ˇ/ and X D .C1 ; : : : ; Cm / is a given sequence of conjugacy
classes of .˛; ˇ/-balanced involutions in G, then a class-random sequence from X
generates a subgroup containing G with probability at least 1 q n .
The basic idea of the proof of Theorem 2.53 is standard: if a class-random
sequence .g1 ; : : : ; gm / does not generate G then all gi belong to some maximal
subgroup M < G, with M not containing G . Since gi is uniformly distributed in
its conjugacy class, we have to estimate the ratios jM Ci j=jCi j for all maximal
subgroups M . Maximal subgroups are characterised by Aschbacher’s theorem [2];
it turns out that the most difficult case is when M is reducible (has a proper invariant
subspace).
Much less is known about Problem 2.51. At present, a solution is known only in
the case when G D SL.n; q/.


Theorem 2.54 ([85]). There exist c and n0 such that if n > n0 , SL.n; q/ Ä
G Ä GL.n; q/, x is a .1=3; 2=3/-balanced involution of G, and g 2 G is a
uniformly distributed random element among the elements of G of even type, then
with probability at least c= log n, g.1/ and g . 1/ are .1=6; 2=3/-balanced involutions
on the eigenspaces E1 .x/ and E 1 .x/ respectively.
The proof of Theorem 2.54 uses a significant enhancement of the generating
function method described in Sect. 2.3.6, and also some ideas from [58].

Acknowledgements This chapter forms part of our Australian Research Council Discovery
Project DP110101153. Praeger and Seress are supported by an Australian Research Council
Federation Fellowship and Professorial Fellowship, respectively. Niemeyer thanks the Lehrstuhl D
f¨ r Mathematik at RWTH Aachen for their hospitality, and acknowledges a DFG grant in SPP1489.
u
All three of us warmly thank the de Br´ n Centre for Computational Algebra at National University
u
of Ireland, Galway, for their hospitality during the Workshop on Groups, Combinatorics and
Computing in April 2011, where we presented the short lecture course that led to the development
of this chapter. We are very grateful to Peter M. Neumann for many thoughtful comments and
advice, and his translation of Euler’s words in Sect. 2.2.2.

References

1. C. Altseimer, A.V. Borovik, Probabilistic Recognition of Orthogonal and Symplectic Groups,
in Groups and Computation, III, vol. 8, Columbus, OH, 1999 (Ohio State University Mathe-
matical Research Institute Publications/de Gruyter, Berlin, 2001), pp. 1–20
2. M. Aschbacher, On the maximal subgroups of the finite classical groups. Invent. Math. 76(3),
469–514 (1984)
3. L. Babai, Local Expansion of Vertex-Transitive Graphs and Random Generation in Finite
Groups, in 23rd ACM Symposium on Theory of Computing (ACM, New York, 1991),
pp. 164–174
4. L. Babai, R. Beals, A Polynomial-Time Theory of Black Box Groups. I, in Groups St. Andrews
1997 in Bath, I. London Mathematical Society Lecture Note Series, vol. 260 (Cambridge
University Press, Cambridge, 1999), pp. 30–64
5. L. Babai, E. Szemer´ di, On the Complexity of Matrix Group Problems I, in 25th Annual
e
Symposium on Foundations of Computer Science (IEEE Computer Society Press, Los
Alamitos, 1984), pp. 229–240
´
6. L. Babai, W.M. Kantor, P.P. P´ lfy, A. Seress, Black-box recognition of finite simple groups of
a
Lie type by statistics of element orders. J. Group Theor. 5(4), 383–401 (2002)
´
7. L. Babai, R. Beals, A. Seress, Polynomial-Time Theory of Matrix Groups, in 41st ACM
Symposium on Theory of Computing, Bethesda, MD, 2009 (ACM, New York, 2009),
pp. 55–64
´
8. R. Beals, C.R. Leedham-Green, A.C. Niemeyer, C.E. Praeger, A. Seress, Permutations with
restricted cycle structure and an algorithmic application. Combin. Probab. Comput. 11(5), 447–
464 (2002)
´
9. R. Beals, C.R. Leedham-Green, A.C. Niemeyer, C.E. Praeger, A. Seress, A black-box group
algorithm for recognizing finite symmetric and alternating groups. I. Trans. Am. Math. Soc.
355(5), 2097–2113 (2003)
10. D.E.-C. Ben-Ezra, Counting elements in the symmetric group, Int. J. Algebra Comput. 19(3),
305–313 (2009)
11. E.A. Bender, Asymptotic methods in enumeration. SIAM Rev. 16, 485–515 (1974)


12. E.A. Bertram, B. Gordon, Counting special permutations. Eur. J. Comb. 10(3), 221–226 (1989)
13. E.D. Bolker, A.M. Gleason, Counting permutations. J. Comb. Theor. Ser. A 29(2), 236–242
(1980)
14. M. B´ na, A. McLennan, D. White, Permutations with roots. Random Struct. Algorithm 17(2),
o
157–167 (2000)
15. W. Bosma, J. Cannon, C. Playoust, The Magma algebra system. I. The user language.
J. Symbolic Comput. 24, 235–265 (1997)
16. J.N. Bray, An improved method for generating the centralizer of an involution. Arch. Math.
(Basel) 74(4), 241–245 (2000)
17. R.W. Carter, Finite Groups of Lie Type (Wiley Classics Library, Wiley, Chichester, 1993),
Conjugacy classes and complex characters, Reprint of the 1985 original, A Wiley-Interscience
Publication
18. F. Celler, C.R. Leedham-Green, S.H. Murray, A.C. Niemeyer, E.A. O’Brien, Generating
random elements of a finite group. Comm. Algebra 23(13), 4931–4948 (1995)
19. W.W. Chernoff, Solutions to x r D ˛ in the alternating group. Ars Combin. 29(C), 226–227
(1990) (Twelfth British Combinatorial Conference, Norwich, 1989)
20. S. Chowla, I.N. Herstein, W.K. Moore, On recursions connected with symmetric groups. I. Can.
J. Math. 3, 328–334 (1951)
21. S. Chowla, I.N. Herstein, W.R. Scott, The solutions of x d D 1 in symmetric groups. Norske
Vid. Selsk. Forh. Trondheim 25, 29–31 (1952/1953)
22. A.M. Cohen, S.H. Murray, An algorithm for Lang’s Theorem. J. Algebra 322(3), 675–702
(2009)
23. A. de Moivre, The Doctrine of Chances: Or, A Method of Calculating the Probability of Events
in Play, 2nd edn. (H. Woodfall, London, 1738)
24. J.D. Dixon, Generating random elements in finite groups. Electron. J. Comb. 15(1), Research
Paper 94 (2008)
25. P. Dusart, The kth prime is greater than k.ln k C ln ln k 1/ for k 2. Math. Comp. 68(225),
411–415 (2009)
26. P. Erd˝ s, M. Szalay, On Some Problems of the Statistical Theory of Partitions, in Number
o
Theory, vol. I, Budapest, 1987. Colloq. Math. Soc. J´ nos Bolyai, vol. 51 (North-Holland,
a
Amsterdam, 1990), pp. 93–110
27. P. Erd˝ s, P. Tur´ n, On some problems of a statistical group-theory. I. Z. Wahrscheinlichkeits-
o a
theorie und Verw. Gebiete 4, 175–186 (1965)
28. P. Erd˝ s, P. Tur´ n, On some problems of a statistical group-theory. II. Acta Math. Acad. Sci.
o a
Hung. 18, 151–163 (1967)
29. P. Erd˝ s, P. Tur´ n, On some problems of a statistical group-theory. III. Acta Math. Acad. Sci.
o a
Hung. 18, 309–320 (1967)
30. P. Erd˝ s, P. Tur´ n, On some problems of a statistical group-theory. IV. Acta Math. Acad. Sci.
o a
Hung. 19, 413–435 (1968)
31. P. Erd˝ s, P. Tur´ n, On some problems of a statistical group theory. VI. J. Indian Math. Soc.
o a
34(3–4), 175–192 (1970/1971)
32. P. Erd˝ s, P. Tur´ n, On some problems of a statistical group theory. V. Period. Math. Hung.
o a
1(1), 5–13 (1971)
33. L. Euler, Calcul de la probabilit´ dans le jeu de rencontre. M´ moires de l’Academie des
e e
Sciences de Berlin, 7 (1751) 1753, pp. 255–270. Reprinted in Opera Omnia: Series 1, vol.
7, pp. 11–25. Available through The Euler Archive at www.EulerArchive.org.
34. L. Euler, Solutio Quaestionis curiosae ex doctrina combinationum. M´ moires de l’Acad´ mie
e e
des Sciences de St.-Petersbourg, 3:57–64, 1811. Reprinted in Opera Omnia: Series 1, vol. 7,
pp. 435–440. Available through The Euler Archive at www.EulerArchive.org.
35. P. Flajolet, R. Sedgewick, Analytic Combinatorics (Cambridge University Press, Cambridge,
2009)
36. J. Fulman, P.M. Neumann, C.E. Praeger, A generating function approach to the enumeration of
matrices in classical groups over finite fields. Mem. Am. Math. Soc. 176(830), vi+90 (2005)


37. The GAP Group, GAP — Groups, Algorithms, and Programming, Version 4.5.2(beta), 2011,
http://www.gap-system.org/
38. S.P. Glasby, Using recurrence relations to count certain elements in symmetric groups. Eur.
J. Comb. 22(4), 497–501 (2001)
39. W.M.Y. Goh, E. Schmutz, The expected order of a random permutation. Bull. Lond. Math.
Soc. 23(1), 34–42 (1991)
40. V. Gonˇ arov, On the field of combinatory analysis. Am. Math. Soc. Transl. 19(2), 1–46 (1962)
c
41. D. Gorenstein, R. Lyons, R. Solomon, The Classification of the Finite Simple Groups.
Mathematical Surveys and Monographs, vol. 40 (American Mathematical Society, Providence,
1994)
42. O. Gruder, Zur Theorie der Zerlegung von Permutationen in Zyklen. Ark. Mat. 2(5), 385–414
(1953)
43. W.K. Hayman, A generalisation of Stirling’s formula. J. Reine Angew. Math. 196, 67–95
(1956)
44. R.B. Herrera, The number of elements of given period in finite symmetric groups. Am. Math.
Mon. 64, 488–490 (1957)
45. P.E. Holmes, S.A. Linton, E.A. O’Brien, A.J.E. Ryba, R.A. Wilson, Constructive membership
in black-box groups. J. Group Theor. 11(6), 747–763 (2008)
46. I.M. Isaacs, W.M. Kantor, N. Spaltenstein, On the probability that a group element is
p-singular. J. Algebra 176(1), 139–181 (1995)
47. E. Jacobsthal, Sur le nombre d’´ l´ ments du groupe sym´ trique Sn dont l’ordre est un nombre
ee e
premier. Norske Vid. Selsk. Forh. Trondheim 21(12), 49–51 (1949)
´
48. W.M. Kantor, A. Seress, Black box classical groups. Mem. Am. Math. Soc. 149(708), viii+168
(2001)
49. A.V. Kolchin, Equations that contain an unknown permutation. Diskret. Mat. 6(1), 100–115
(1994)
50. V.F. Kolchin, Random Graphs. Encyclopedia of Mathematics and Its Applications, vol. 53
(Cambridge University Press, Cambridge, 1999)
51. E. Landau, Handbuch der Lehre von der Verteilung der Primzahlen. 2 B¨ nde, 2nd edn.
a
(Chelsea Publishing Co., New York, 1953), With an appendix by Paul T. Bateman
52. C.R. Leedham-Green, The Computational Matrix Group Project, in Groups and Computation,
III, vol. 8, Columbus, OH, 1999 (Ohio State University Mathematical Research Institute
Publications/de Gruyter, Berlin, 2001), pp. 229–247
53. C.R. Leedham-Green, E.A. O’Brien, Constructive recognition of classical groups in odd
characteristic. J. Algebra 322(3), 833–881 (2009)
54. G.I. Lehrer, Rational tori, semisimple orbits and the topology of hyperplane complements.
Comment. Math. Helv. 67(2), 226–251 (1992)
55. G.I. Lehrer, The cohomology of the regular semisimple variety. J. Algebra 199(2), 666–689
(1998)
56. M.W. Liebeck, E.A. O’Brien, Finding the characteristic of a group of Lie type. J. Lond. Math.
Soc. (2) 75(3), 741–754 (2007)
57. M.W. Liebeck, A. Shalev, The probability of generating a finite simple group. Geom. Dedicata
56(1), 103–113 (1995)
58. F. L¨ beck, A.C. Niemeyer, C.E. Praeger, Finding involutions in finite Lie type groups of odd
u
characteristic. J. Algebra 321(11), 3397–3417 (2009)
59. E.M. Luks, Permutation Groups and Polynomial-Time Computation, in Groups and compu-
tation, New Brunswick, NJ, 1991. DIMACS Series in Discrete Mathematics and Theoretical
Computer Science, vol. 11 (American Mathematical Society, Providence, 1993), pp. 139–175
60. R. Lyons, Evidence for a new finite simple group. J. Algebra 20, 540–569 (1972)
61. A. Mar´ ti, Symmetric functions, generalized blocks, and permutations with restricted cycle
o
structure. Eur. J. Comb. 28(3), 942–963 (2007)
62. N. Metropolis, The beginnings of the Monte Carlo method. Los Alamos Sci. 15 (Special Issue),
125–130 (1987)


63. M.P. Mineev, A.I. Pavlov, The number of permutations of a special form. Mat. Sb. (N.S.)
99(141)(3), 468–476, 480 (1976)
64. P.R. de Monmort, Essay d’analyse sur les jeux de hazard (J. Quillau, Paris, 1708)
65. P.R. de Monmort, Essay d’analyse sur les jeux de hazard, 2nd edn. (J. Quillau, Paris, 1713)
66. L. Moser, M. Wyman, On solutions of x d D 1 in symmetric groups. Can. J. Math. 7, 159–168
(1955)
67. L. Moser, M. Wyman, Asymptotic expansions. Can. J. Math. 8, 225–233 (1956)
68. L. Moser, M. Wyman, Asymptotic expansions. II. Can. J. Math. 9, 194–209 (1957)
69. P.M. Neumann, C.E. Praeger, A recognition algorithm for special linear groups. Proc. Lond.
Math. Soc. (3) 65 (3), 555–603 (1992)
70. P.M. Neumann, C.E. Praeger, Cyclic matrices over finite fields. J. Lond. Math. Soc. (2) 52,
263–284 (1995)
71. A.C. Niemeyer, T. Popiel, C.E. Praeger, Abundant p-singular elements in finite classical
groups, preprint (2012) http://arxiv.org/abs/1205.1454v2
72. A.C. Niemeyer, C.E. Praeger, A recognition algorithm for classical groups over finite fields.
Proc. Lond. Math. Soc. (3) 77 (1), 117–169 (1998)
73. A.C. Niemeyer, C.E. Praeger, On the frequency of permutations containing a long cycle.
J. Algebra 300(1), 289–304 (2006)
74. A.C. Niemeyer, C.E. Praeger, On permutations of order dividing a given integer. J. Algebr.
Comb. 26(1), 125–142 (2007)
75. A.C. Niemeyer, C.E. Praeger, On the proportion of permutations of order a multiple of the
degree. J. Lond. Math. Soc. (2) 76(3), 622–632 (2007)
76. A.C. Niemeyer, C.E. Praeger, Estimating proportions of elements in finite groups of Lie type.
J. Algebra 324(1), 122–145 (2010)
77. E.A. O’Brien, Algorithms for Matrix Groups, in Groups St. Andrews 2009 in Bath, vol. 2.
London Mathematical Society Lecture Note Series, vol. 388 (Cambridge University Press,
Cambridge, 2011), pp. 297–323
78. C.W. Parker, R.A. Wilson, Recognising simplicity of black-box groups by constructing
involutions and their centralisers. J. Algebra 324(5), 885–915 (2010)
79. E.T. Parker, P.J. Nikolai, A search for analogues of the Mathieu groups. Math. Tables Aids
Comput. 12, 38–43 (1958)
80. A.I. Pavlov, An equation in a symmetric semigroup. Trudy Mat. Inst. Steklov. 177, 114–121,
208 (1986); Proc. Steklov Inst. Math. 1988(4), 121–129, Probabilistic problems of discrete
mathematics
81. A.I. Pavlov, On permutations with cycle lengths from a fixed set. Theor. Probab. Appl. 31,
618–619 (1986)
82. W. Plesken, D. Robertz, The average number of cycles. Arch. Math. (Basel) 93(5), 445–449
(2009)
83. C.E. Praeger, On elements of prime order in primitive permutation groups. J. Algebra 60(1),
126–157 (1979)
´
84. C.E. Praeger, A. Seress, Probabilistic generation of finite classical groups in odd characteristic
by involutions. J. Group Theor. 14(4), 521–545 (2011)
´
85. C.E. Praeger, A. Seress, Balanced involutions in the centralisers of involutions in finite general
linear groups of odd characteristic (in preparation)
´
86. C.E. Praeger, A. Seress, Regular semisimple elements and involutions in finite general linear
groups of odd characteristic. Proc. Am. Math. Soc. 140, 3003–3015 (2012)
´
87. A. Seress, Permutation Group Algorithms. Cambridge Tracts in Mathematics, vol. 152
(Cambridge University Press, Cambridge, 2003)
88. C.C. Sims, Computational Methods in the Study of Permutation Groups, in Computational
Problems in Abstract Algebra, Proceedings of the Conference, Oxford, 1967 (Pergamon,
Oxford, 1970), pp. 169–183
89. C.C. Sims, The Existence and Uniqueness of Lyons’ Group, in Finite groups ’72, Proceedings
of the Gainesville Conference, University of Florida, Gainesville, FL, 1972. North–Holland
Mathematical Studies, vol. 7 (North-Holland, Amsterdam, 1973), pp. 138–141


90. A.N. Timashev, Random permutations with cycle lengths in a given ﬁnite set. Diskret. Mat.
20(1), 25–37 (2008)
91. J. Touchard, Sur les cycles des substitutions. Acta Math. 70(1), 243–297 (1939)
92. L.M. Volynets, The number of solutions of the equation x s D e in a symmetric group. Mat.
Zametki 40(2), 155–160, 286 (1986)
¨
93. R. Warlimont, Uber die Anzahl der L¨ sungen von x n D 1 in der symmetrischen Gruppe Sn .
o
Arch. Math. (Basel) 30 (6), 591–594 (1978)
94. H. Wielandt, Finite Permutation Groups, Translated from the German by R. Bercov (Academic,
New York, 1964)
95. H.S. Wilf, The asymptotics of e P .z/ and the number of elements of each order in Sn . Bull. Am.
Math. Soc. (N.S.) 15 (2), 228–232 (1986)
96. H.S. Wilf, Generatingfunctionology, 2nd edn. (Academic, Boston, 1994)
97. K. Zsigmondy, Zur Theorie der Potenzreste. Monatsh. f¨ r Math. U. Phys. 3, 265–284 (1892)
u

http://www.springer.com/978-1-4471-4813-5

Probabilistic group theory, combinatorics, and computing

More Related Content

What's hot (19)

Similar to Probabilistic group theory, combinatorics, and computing (20)

More from Springer (20)

Probabilistic group theory, combinatorics, and computing