Journal of Algebra 278 (2004) 294–313

www.elsevier.com/locate/jalgebra

Generalized Frobenius–Schur numbers

Daniel Bump a,∗ and David Ginzburg b
a Department of Mathematics, Stanford University, Stanford, CA 94035-2125, USA
b School of Mathematical Sciences, Sackler Faculty of Exact Sciences, Tel Aviv University,
Ramat-Aviv 69978, Israel
Received 4 August 2003
Available online 2 April 2004
Communicated by Jan Saxl

Let G be a finite group. Suppose that τ : G → G is an automorphism such that τ r = 1.

The norm map N : G → G is defined by
2 r−1
N(g) = g · τ g · τ g · · · · · τ g. (1)

Let χ be a character of G. We will study a class of character sums including

1
 
ετ (χ) = χ N(g) . (2)
|G|
g∈G

Theorem 9 gives a concrete interpretation of this sum (and its generalizations). Let (π, V )
be a representation of G affording χ . Let Eτ (π) be the vector space of r-linear forms
T : V × · · · × V → C satisfying
    r−1  
T π(g)v1 , π τ g v2 , . . . , π τ g vr = T (v1 , . . . , vr ).

Let cr : Eτ (π) → Eτ (π) be the linear transformation which replaces T by

(cr T )(v1 , v2 , . . . , vr ) = T (vr , v1 , . . . , vr−1 ).

We will show in Theorem 9 that ετ (χ) is the trace of cr . As a consequence (Theorem 10),
we see that ετ (χ) is an algebraic integer, in fact, an element of Z[e2πi/r ].

* Corresponding author.
E-mail address: bump@math.stanford.edu (D. Bump).

0021-8693/$ – see front matter  2004 Elsevier Inc. All rights reserved.
doi:10.1016/j.jalgebra.2004.02.012
 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 295

In many cases ετ (χ) ∈ Z, so it may be worth giving an example where it is not a rational
integer. Let

 
G = x, y, z | x 3 = y 3 = z3 = 1, zx = xz, zy = yz, xyx −1y −1 = z ,

a nonabelian group of order 27. Let τ : G → G be the outer automorphism of order r = 3

such that τ (x) = x, τ (y) = xy. Let χ be the character of degree 3 such that χ(z) = 3ρ,
where ρ is a primitive cube root of unity. We find that ετ (χ) = 1 + 2ρ.
Let us give an application. If g ∈ G let M(g) be the number of solutions to the equation
N(x) = g with x ∈ G. Philip Hall [14], generalizing a theorem of Frobenius, proved that
M(g) is divisible by the greatest common divisor of r and the order of the centralizer of g.
An interesting integer valued function on the group—is it perhaps a generalized character?
The answer is no, but we will now prove that M lies in the Z[e2πi/r ]-algebra generated
by the irreducible characters. Indeed, this is immediate from Theorem 10 and the identity

M(g) = ετ (χ)χ(g),

where the summation is over all irreducible characters χ of G. (The proof is a

straightforward adaptation of Proposition 1 below.) 
If r = 2, and τ is trivial, then ετ (χ) is (1/|G|) χ(g 2 ), and such sums were considered
by Frobenius and Schur [10]. If r = 2 and τ is not assumed to be trivial, these sums were
considered by Kawanaka and Matsuyama [18]. The work of Kawanaka and Matsuyama
has many interesting associations, for example to the notion of a generalized involution
model which we discuss.
We will consider a couple of different generalizations of (2) involving another
character η. In Theorem 9 the character η is a character of G, assumed τ -invariant. In
Theorem 12, the character η is not a character of G, but of its subgroup τ -fixed points.
These sums allow us to give a reinterpretation of a famous theorem of Shintani on the
lifting of characters for GL(n) over a finite field.
This paper touches on different works of Frobenius and Schur. An invaluable guide to
their work is Curtis [6].

1. Theorems of Frobenius–Schur and Kawanaka–Matsuyama

Let (π, V ) be an irreducible complex representation of a compact group G. We will

denote by χ = χπ the character of π . If π is self-contragradient, then there exists a
nondegenerate G-equivariant bilinear form B on V , unique up to scalar multiple and for
some constant ε = ±1 we must have

B(v, w) = εB(w, v).

296 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313

We will denote ε = ε1 (π) or ε1 (χ). If π is not self-contragradient, then we define

ε1 (π) = 0. Frobenius and Schur [10] proved

 
χ g 2 dg = ε1 (π).
G

These Frobenius–Schur numbers have a concrete interpretation. If ε1 (π) = 1, then

π(G) is conjugate to a subgroup of the orthogonal group O(n), where n = dim(V ). If
ε1 (π) = −1, then π(G) is conjugate to a subgroup of the symplectic group Sp(n), and n is
even. We therefore say that π is orthogonal or symplectic if ε1 (π) = 1 or −1, respectively.
An important generalization of this result was obtained by Kawanaka and Matsuyama
[18]. Let τ be an automorphism of G satisfying τ 2 = 1. Define ετ (π) as follows. If there
exists a nondegenerate bilinear form B on V such that
   
B π(g)v, π τ g w = B(v, w) (3)

then by the irreducibility of V , B is unique. The form B(w, v) has the same property, so
there exists a constant ε such that

B(v, w) = εB(w, v).

Moreover, B(v, w) = ε2 B(v, w) so ε = ±1. In this case we denote ε = ετ (π) = ετ (χ).

If no nondegenerate bilinear form exists satisfying (3), let ετ (π) = 0. Kawanaka and
Matsuyama proved that

 
ετ (π) = χ g · τ g dg. (4)

We are normalizing the Haar measure so the volume of G is one.

We will reformulate this result of Kawanaka and Matsuyama in preparation for a more
substantial generalization to multilinear forms. We begin by extending the definition of
ετ (g) to representations that are not necessarily irreducible. Let (π, V ) be an arbitrary
representation, not necessarily irreducible, and let χ be its character. Let Eτ (π) be the
space of bilinear forms B on V satisfying (3). Define an endomorphism c : Eτ (π) →
Eτ (π) by

cB(v, w) = B(w, v). (5)

Then we define ετ (π) to be the trace of c. Clearly this coincides with the previous definition
when π is irreducible.

Theorem 1. If τ is an automorphism of the compact group G such that τ 2 = 1, and (π, V )

is a finite-dimensional representation, then (4) holds.
 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 297

Rather than deduce this from the result of Kawanaka and Matsuyama, we will instead
prove a more general result in Theorem 9 below.
If π is irreducible, Theorem 1 is the theorem of Kawanaka and Matsuyama. One might
expect that the general case follows from the special case in a simple way. Indeed the right
side of (4) is linear as a function of χ . So one might be tempted to assume that the space of
bilinear forms satisfying (3) decomposes into a direct sum of the corresponding spaces for
the irreducible constituents of χ and that the trace of c is therefore the sum of the traces
over these irreducible subspaces.
That this expectation is not true may be seen by an example. Let G be the cyclic group
of order 3, and let χ1 , χ2 be the two nontrivial characters of G, both of order 3. We take
τ = 1. Then ε1 (χi ) = 0, and there are no bilinear forms satisfying (3) for either χi . On
the other hand, let χ = χ1 + χ2 . Since the right side of (4) is linear as a function of χ ,
we must still have ε1 (χ) = 0. Now, however, there are two invariant bilinear forms. The
trace of c remains zero, but the space of bilinear forms does not decompose into subspaces
corresponding to the irreducible constituents.

2. Combinatorial interpretation of the “model”

If G is a finite group, we will denote by M a representation which contains every

irreducible representation of G exactly once. Let χM be its character. The values of χM
are algebraic integers, and since Gal(Q̄/Q) permutes the constituents of M, they are
rational integers. There is a strong tendency for these values to be nonnegative, so strong
that one might wonder whether this is always true.
Solomon [23] proved that the row sums of the character table are nonnegative. This
result is reproduced in Feit [8, p. 34], who remarks that Solomon and Thompson had
pointed out that the column sums of the character table (which are the values of χM )
can be negative. Still, we had to look fairly hard to find an example. The Mathieu group
M11 has the property that χM takes negative values for two conjugacy classes.
Given that the values of χM are, for many groups, nonnegative integers, it is natural to
ask whether χM has a combinatorial interpretation. Is, perhaps, χM (a) the cardinality of
some set naturally depending on M? We now prepare to prove Theorem 3, where such an
interpretation will be given, for certain groups.

Proposition 1. Let G be a finite group, and let a ∈ G. Then π χπ (a)ετ (π) is the number
of solutions g to the equation g · τ g = a. (The sum is over all irreducible representations π .)

Proof. Substituting the definition of ε1 (π) and interchanging the order of summation gives

 1
 τ 

χπ (a)χπ g · g .
g
|G| π

By Schur orthogonality, the expression in brackets is zero unless g · τ g is a conjugate of a,

in which it is the reciprocal of the cardinality of the conjugacy class of a. The result follows
by simple counting provided we know that the cardinality of the set S(a) of solutions to
298 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313

g · τ g = a is constant on conjugacy classes. If b = hah−1 , then g → hg τ h−1 is a bijection

S(a) → S(b), as required. 2

Proposition 2. A necessary and sufficient condition that ετ (π) = ±1 for all irreducible π
is that every g ∈ G is conjugate to τ g −1 .

Proof. The condition ετ (π) = ±1 amounts to the existence of a bilinear form B on the
space V of π such that B(π(g)v1 , v2 ) = B(v1 , π(τ g −1 )v2 ). Therefore π(g) and π(τ g −1 )
are adjoints with respect to B and have the same trace. Thus we see that all irreducible
characters take the same value on g and τ g −1 , so they are conjugate. 2

The following fact is well-known in the special case where τ = 1. The case where
z = 1 is in Howlett and Zworestine [15]. Motivated by examples of Prasad [21], Steinberg
[25] and Gow [12], and Proposition 3 below, we generalize this by using a central
element z (which must be of order two). If (π, V ) is an irreducible representation of G, let
ωπ : Z(G) → C× be the central character of π , so π(z)v = ωπ (z)v for all v ∈ V .

Theorem 2. Let G be a finite group, and let τ be an automorphism satisfying τ 2 = 1. Let

z ∈ Z(G) such that z2 = 1. The following are equivalent:

(i) Every irreducible representation of G satisfies ετ (π) = ωπ (z).

(ii) The dimension of M is equal to the number of g ∈ G such that g · τ g = z.

For example, if G = Sn these conditions are satisfied with τ = 1 and z = 1. Both

assertions have concrete interpretations: for (i), the irreducible representations of Sn can be
constructed over Q, a fortiori over R, and are therefore orthogonal. For (ii), the Robinson–
Schensted correspondence can be used to show that the number of g ∈ Sn such that g 2 = 1
is equal to the number of standard tableaux, which in turn is equal to the sum of the degrees
of the irreducible representations of G.

Proof. Using Proposition 1 and the identity χπ (az)ωπ (z) = χπ (a), (i) implies (ii).
On the other hand, assuming (ii), take a = z. Hypothesis (ii) means that





ωπ (z)ετ (π) χπ (1) = χπ (z)ετ (π) = χπ (1).
π π π

Since each |ωπ (z)ετ (π)|
1, and since the χπ (1) > 0, this implies that the numbers
ωπ (z)ετ (π) have absolute value 1, and the same complex argument. When π is the trivial
representation, ωπ (z)ετ (π) = 1, so this is true for all π . 2

Theorem 3. If the equivalent conditions of Theorem 2 are satisfied, then χM (a) is the
number of solutions to the equation g · τ g = az.

Thus if the equivalent conditions of Theorem 2 are satisfied, we have a suitable

combinatorial interpretation of χM .
 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 299

Proof. This is immediate from Proposition 1. 2

Let us recall results of Gow [13] and Klyachko [19]. Independently, Gow and Klyachko
proved two theorems, and proved the equivalence of them. Let G be GL(n, Fq ), and let
(π, V ) be an irreducible representation.

Theorem 4 (Gow, Klyachko). Let (π, V ) be an irreducible representation of GL(n, Fq ),

q odd. There exists a symmetric bilinear form B on V such that
   
B π(g)v, π t g −1 w = B(v, w). (6)

Theorem 5 (Gow, Klyachko). The sum of the dimensions of the irreducible representations
of G = GL(n, Fq ) (q odd) is equal to the number of symmetric matrices in G.

These papers were written before Kawanaka and Matsuyama [18], which clarifies the
equivalence of these two facts. The first theorem asserts that if τ g = t g −1 then ετ (π) = 1
for all irreducible representations. As Howlett and Zworestine noted, Theorem 2 shows the
equivalence in a transparent way.
Macdonald [20] removed the assumption that q is odd. Fulman and Guralnik [11] in a
remarkable paper give explicit formulas for χM at all conjugacy classes of G = GL(n, Fq ),
and striking interpretations of these results.
When n = 2, in addition to the involution τ , there is another involution θ of GL(2, Fq )
which can be used to illustrate the need for a central element z in Theorem 2. If g ∈
GL(2, Fq ), let θ g = det(g)−1 · g.

Proposition 3. If (π, V ) is an irreducible representation of GL(2, Fq ), then εθ (π) =

ωπ (−I ).

Proof. We note the identity


−1 0 −1
τ
g=η· g·η
θ
, η= .
1 0

By Theorem 4, there exists a symmetric bilinear form B on V satisfying (6). Define another
bilinear form T on V by T (v, w) = B(v, π(η)w). Then T (π(g)v, π(θ g)w) = T (v, w).
Since B is symmetric, and since τ η = η,
       
T (w, v) = B w, π(η)v = B π(η)v, w = B v, π η−1 w .

As η−1 = −η, this equals T (v, π(−I )w) = ωπ (−I )T (v, w) and the result follows. 2

As another example, let G = SL(2, Fq ). If q ≡ 1 mod 4, then every irreducible

representation is self-contragradient, that is ε1 (π) = ±1, for it is easily seen that every
element is conjugate to its inverse. This
 is easily checked for the semisimple
 conjugacy

classes, and for the unipotent ones, 10 11 is conjugate to its inverse by α0 α0−1 , where
300 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313

α 2 = −1. Moreover, it follows from Gow [12] Theorem 1 that ε1 (π) = ωπ (−I ). If q ≡ 3
mod 4, however, it is not true that every irreducible representation is self-contragradient.
It is still true that most conjugacy classes of SL(2, Fq ) are conjugate to their inverses, the
two unipotent classes being the only exceptions. And it is still true that most irreducible
representations are self-dual, in fact all but four are, two each of degrees (q + 1)/2 and
(q − 1)/2 being the only exceptions. And Gow’s result still asserts that when ε1 (π) = 0, it
equals ωπ (−I ). These four exceptions are a defect which can be eliminated by introducing
an involution. Let τ denote conjugation by

−1 0
ζ= ∈ GL(2, Fq ).
0 1

This has the advantage that τ g ∼ g −1 for all g, even the unipotent elements, so that
ετ (π) = ±1 for all π .

Proposition 4. If G = SL(2, Fq ) and τ is conjugation by ζ , then ετ (π) = 1 for all

irreducible representations (π, V ).

Proof. It may be computed that dim M = q(q + 1) for this group. On the other hand, the
solutions to g · τ g = I are the matrices
 
a b  2
a + bc = 1 ,
c a

and it is easy to count that there are q(q + 1) of these. The result follows by Theorem 2. 2

Propositions 3 and 4 both have generalizations to the symplectic group, due to

Vinroot [26]. Let
 
−In −In
J= , ζ=
In In

and for any field F let

 
Sp(2n, F ) = g ∈ GL(2n, F ) | gJ t g = J ,
 
GSp(2n, F ) = g ∈ GL(2n, F ) | gJ t g = µ(g)J ,

so that µ : GSp(2n, F ) → F × is a character. We have automorphisms θ and τ of

GSp(2n, Fq ) and Sp(2n, Fq ), respectively, defined by

θ
g = µ(g)−1 g, τ
g = ζgζ −1 .

Theorem 6 (Vinroot).

(i) If (π, V ) is an irreducible representation of GSp(2n, Fq ), then εθ (π) = ωπ (−I2n ).

 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 301

(ii) If (π, V ) is an irreducible representation of Sp(2n, Fq ), then ετ (π) = 1.

If n = 1, GSp(2n) = GL(2) and Sp(2n) = SL(2), so we have already proved this in

this case. If n > 1 and q ≡ 1 mod 4, Vinroot’s theorem may be deduced from results of
Gow [12]. If n > 1 and q ≡ 3 mod 4, however, Vinroot’s results improve Gow’s results.

3. Generalized involution models

The term “model” is used in two different ways, which we now explain. A model of a
(typically irreducible) representation π is an embedding of π in a multiplicity free induced
representation, typically induced from a one-dimensional representation of a subgroup
of G. The project of Bernstein, Gelfand and Gelfand [3] is to find several subgroups
H1 , . . . , Hn of G and characters ψi (typically one-dimensional) of Hi such that

∼
Hi (ψi ) = M.
IndG (7)
i

Of course this means that IndG Hi (ψi ) is multiplicity free, so we obtain a model for every
irreducible representation of G. Following Bernstein, Gelfand and Gelfand, such data are
sometimes called a model for G.
Let us assume that G satisfies the equivalent conditions of Theorem 2. Then we are
presented with a set of candidates for the groups Hi . Indeed, let G act on X = {x ∈ G |
x · τ x = z} by twisted conjugacy: g : x → g · x · τ g −1 . Let H1 , . . . , Hn be the stabilizers
of a set of orbit representatives for this action. Each orbit has [G : Hi ] elements, so

i [G : Hi ] = |X| = dim M. This numerical equality is a precondition for (7).
If τ = 1 and z = 1, then the Hi are precisely the centralizers of involutions, and such
a model is called an involution model. (For the purpose of this definition, the identity
element of G is considered an involution.) A good example is for Sn , where the model was
constructed by Inglis, Richardson and Saxl [16], and independently (details unpublished)
by Klyachko. Baddeley [1] shows that most but not all Weyl groups have involution
models.
If τ = 1, it is natural to call a model for G arising from a set of stabilizers for the action
of G on X by twisted conjugacy a generalized involution model. We will see that such a
model may not exist, even if the conditions of Theorem 2 are satisfied. In all the examples
that we will consider, z = 1.
If G = GL(n, Fq ), Klyachko gave a model for G in connection with Theorem 5.
A relationship can be seen between Klyachko’s model and the model of Inglis, Richardson
and Saxl. To explain this point, Klyachko’s models can be understood as follows. Let
2r < n. In the parabolic subgroup with Levi factor GL(2r) × GL(n − 2r), let Hr be the
group
 
h ∗ 
 h ∈ Sp(2r), k ∈ GL(n − 2r) .
0 k
302 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313

We consider the representation 1 ⊗ Γn−2r , where Γn−2r is the Gelfand–Graev represen-

tation of GL(n − 2r). Inducing to GL(n, Fq ) and summing over r gives M. The key ob-
servation is that Weyl group of Hr is W (Cr ) × Sn−2r , where W (Cr ) is the Weyl group
of Sp(2r) realized as a subgroup of S2r , and these groups are precisely the centralizers of
involutions in Sn used in the involution model.
Inglis and Saxl [17], being aware of this connection, and gave a variant of Klyachko’s
model in which this relationship between Klyachko’s GL(n) model and the models for the
symmetric group is made clarified. But despite this relationship with the involution model
for Sn , Klyachko’s model is itself not a generalized involution model.
We show that GL(2, Fq ) does not have a generalized involution model with τ g = tg −1 .
In this case the subgroups H1 and H2 are the orthogonal groups of the two equivalence
classes of binary quadratic forms. Their orders are |H1 | = 2(q − 1) and |H2 | = 2(q + 1).

Theorem 7. In this setting, there do not exist characters ψi of the Hi such that (7) is
satisfied.

Proof. Let W be the cyclic subgroup of G = GL(2, Fq ) generated by −I . Thus W has

order two and is contained in the center of G. Then M = M+ ⊕ M− , where M+ is the
sum of irreducible representations on which W acts trivially, and M− is the sum of the
irreducible representation on which it acts nontrivially. We have
  1    1
dim M+ = q 2 + 1 (q − 1), dim M− = (q − 1)2 (q + 1).
2 2
Assume that ψ1 and ψ2 are characters of H1 and H2 such that (7) is satisfied. If
ψi (−I ) = 1, then every irreducible character in the induced module ψiG is in M+ , and if
ψi (−1) = −1, then every irreducible character in ψiG is in M− . We must have ψ1G = M−
and ψ2G = M+ or ψ1G = M+ and ψ2G = M− . However the indices of H1 and H2 in G
are
1   1
[G : H1 ] = q q 2 − 1 , [G : H2 ] = q(q − 1)2 .
2 2
It is therefore impossible. 2

On the other hand, a variant of this setup does produce a model for PGL(2, Fq ), where
q is odd. The following construction is similar to one proposed by Soto-Andrade [24].
However Soto-Andrade makes exceptions for the one-dimensional representations, and
we do not. In PGL(2, Fq ), let H1 be the normalizer of the diagonal torus, let H2 be the
normalizer of an anisotropic torus H2◦ , and let H3 = G. We take ψ1 to be the trivial
character of H1 , ψ2 to be the quadratic character of H2 whose kernel is H2◦ , and let ψ3
be the unique nontrivial one-dimensional character of H3 = G.

Theorem 8. For G = PGL(2, Fq ), we have M = ψ1G ⊕ ψ2G ⊕ ψ3G . The groups Hi are the
centralizers of the three conjugacy classes of involutions, so this model is an involution
model.
 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 303


Proof. We have i [G : Hi ] = q 2 + 1 = dim(M), so it is enough to show that ψ1G ⊕ ψ2G ⊕
ψ3G is multiplicity-free. The restriction of ψ3 to H1 and H2 differs from ψ1 and ψ2 so does
not occur in ψ1G ⊕ ψ2G . It is sufficient to show that ψ1G and ψ2G are multiplicity-free and
disjoint.

Lemma 1 (Gelfand). Let G be a group and let ι : G → G be an anticommutative involution.

Let H be a subgroup of G and ψ a character of H such that H and ψ are stabilized by ι.
Assume that every double coset H gH is ι-invariant, and has a representative g such that
ι g = h gh where ψ(h )ψ(h ) = 1. Then ψ G is multiplicity-free.
1 2 1 2

Proof. By the geometric form of Mackey’s theorem the endomorphism ring of ψ G

consists of the convolution ring of functions ∆ on G satisfying ∆(hgh
) = ψ(h)∆(g)ψ(h
)
for h, h
∈ H (see Bump [4, Proposition 4.1.2]). Defining ι ∆(g) = ∆(ι g), since ι is
anticommutative, ι (∆1 ∗ ∆2 ) = ι ∆2 ∗ ι ∆1 . The assumption that every double coset is
ι-invariant means that ι ∆ = ∆, since with g as in the statement of the lemma, we have
∆(ι g) = ∆(g). Therefore EndG (ψ G ) is commutative. It follows that ψ G is multiplicity
free. 2

Returning to the proof of the theorem, we may apply the lemma with H = H1 , ψ = ψ1
and ι the transpose involution. We see that ψ1G is multiplicity free.
We may take H2◦ to be
 
a b  2
 a + βb2 = 0 ,
−βb a

with −β a fixed nonsquare in F×

q . We may then apply the Gelfand Lemma with
 
β −1 0 β 0
ι
g= t
g .
0 1 0 1

We may take the coset representatives for H2 \G/H2 to be diagonal. We see that ψ2G is
multiplicity free.
Finally we must show that ψ1G and ψ2G are disjoint. By Mackey’s theorem, it is
sufficient to show that for every double coset H1 gH2 there exists γ ∈ H1 such that
g −1 γ g ∈ H2 and ψ1 (γ ) = ψ2 (g −1 γ g). If this is true for one coset representative g it is
true for every representative. We may take the representative g to be a unipotent matrix
g = 10 x1 . Now choose γ = m0 10 ∈ H1 , where m = β/(1 + x 2 β). Then ψ1 (γ ) = 1 while
ψ2 (g −1 γ g) = −1. 2

The alternating groups An give interesting examples and counterexamples for general-
ized involution models. The involution τ will be trivial for some n, and for others, it will
be conjugation by the transposition (12) ∈ Sn − An .

Proposition 5. In An (n > 2) every element is conjugate to its inverse if n = 5, 6, 10 or

14. If n = 3, 4, 7, 8 or 12, then every element g is conjugate to (12)g(12). These are the
304 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313

only cases where there exists an involution τ of An such that g is conjugate to τ g −1 for all
g ∈ An .

The first statement, about An when n is conjugate to its inverse is also proved in
Berggren [2].

Proof. For any g ∈ An , g is conjugate to its inverse in Sn . If the centralizer CSn (g) is not
contained in An then it follows that g ∼ g −1 in An , too. It is easy to see that if CSn (g) ⊆ An
then the cycle type of g is a partition of n into distinct odd integers. If we write n as a sum
of k distinct odd integers, then g is conjugate to (12)g −1 (12) if (n − k)/2 is even, and g
is conjugate to g −1 if (n − k)/2 is odd. The partitions of n into distinct odd integers are
presented in Table 1.
If distinct partitions with (n − k)/2 of both parities can be found, clearly no τ exists.
This is true for odd n  9 and even n  16. The result is now evident. 2

Now let us investigate which of these have a generalized involution model. Let τ be the
identity An → An if n = 5, 6, 10 or 14, and τ is conjugation by (12) if n = 3, 4, 7, 8 or
12. Let X be the set of solutions to g τ g = 1. If τ = 1, this is the set of involutions (i.e.,
elements satisfying g 2 = 1) in An . On the other hand, if τ is conjugation by (12), then
g ∈ X if and only if g(12) = h is an involution in Sn − An . The map g → g(12) also helps
us to understand the action of G = An on X. The action on X is twisted conjugacy, but the
corresponding action on involutions in Sn − An is ordinary conjugacy. Thus the stabilizers
Hi in (7) will be the centralizers of a set of representatives of the conjugacy classes of
involutions in An when τ is trivial; and when τ is not trivial, they will be the centralizers
in An of the An -conjugacy classes of involutions in Sn − An .
A3 has a generalized involution model: there is one conjugacy class of involutions
in S3 − A3 , and the stabilizer is the group {1}. Inducing its unique irreducible character
gives M.

Table 1
Partitions into odd and unequal parts
n Partition k (n − k)/2
3 3 1 1
4 3+1 2 1
5 5 1 2
6 5+1 2 2
7 7 1 3
8 7 + 1 or 5 + 3 2 3
9 9 or 5 + 3 + 1 1 or 3 4 or 3
10 9 + 1 or 7 + 3 2 4
11 11 or 7 + 3 + 1 1 or 3 5 or 4
12 11 + 1 or 9 + 3 or 7 + 5 2 5
13 13 or 9 + 3 + 1 or 7 + 5 + 1 1 or 3 6 or 5
14 13 + 1 or 11 + 3 or 9 + 5 2 6
15 15 or 11 + 3 + 1 or 9 + 5 + 1 or 7 + 5 + 3 1 or 3 7 or 6
16 15 + 1 or 7 + 5 + 3 + 1 etc. 2 or 4 7 or 6
 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 305

A4 also has a generalized involution model with τ nontrivial. There is one A4 conjugacy
class of involutions in S4 − A4 , represented by (12). The centralizer in A4 is the cyclic
group H = (12)(34) of order 2. Inducing the trivial character of H gives M.
A5 has a generalized involution model with τ = 1. There are two conjugacy classes
of involutions in A5 , representatives being 1 and (12)(34). The centralizers are H1 = A5
and the abelian subgroup H2 of order 4 generated by (12)(34) and (13)(24). Inducing
any nontrivial character ψ2 of H2 gives a 15-dimensional representation of A5 containing
every nontrivial irreducible representation precisely once, so we take ψ1 to be the trivial
representation of H1 .
A6 has a generalized involution model with τ = 1. There are two conjugacy classes
of involutions in A6 , representatives being 1 and (12)(34). The centralizers are H1 = A6
and the dihedral group H2 of order 8 with generators u = (1324)(56) and t = (12)(56),
satisfying tut −1 = u−1 , t 2 = u4 = 1. The group H2 has four one-dimensional characters,
two of which take value −1 on u. Choosing ψ2 to be either of these, and ψ1 to be the trivial
character gives a generalized involution model.
A7 has a generalized involution model with τ = 1. There are two A7 -conjugacy classes
of involutions in S7 − A7 . The first has centralizer H1 = CA7 ((67)). We may construct
an isomorphism S5 → H1 by mapping σ ∈ S5 (acting on {1, 2, 3, 4, 5}) to itself if σ
is even, and to σ (67) if σ is odd. We take ψ1 to be the trivial character of H1 . The
second centralizer is H2 = CA7 ((12)(34)(56)). It is a group of order 24 which may be
identified with S4 . To construct an isomorphism H2 → S4 note that H2 has four 3-Sylow
subgroups, (135)(246) , (136)(245) , (145)(236) and (146)(235) . Acting on these
by conjugation gives a faithful permutation representation and an isomorphism H2 → S4 .
We take ψ2 to be the nontrivial one-dimensional character of H2 . Labeling the irreducible
characters as in the Atlas of Finite Groups [5], we find that ψ1G = χ1 + χ2 + χ5 , while
ψ2G = χ3 + χ4 + χ6 + χ7 + χ8 + χ9 , and so we have a generalized involution model in this
case, too.
A8 has no generalized involution model with τ conjugation by (12). Indeed, there
are two conjugacy classes of involutions in S8 − A8 , one of which is (12)(34)(56). Its
centralizer is a group of order 48 which admits four characters of degree one. No matter
which of these is chosen, the induced representation of A8 is not multiplicity free, so no
model can be constructed using this subgroup.
We do not know whether generalized involution models exist in the remaining case of
A10 , A12 and A14 .

4. Higher order automorphisms

Let r be a positive integer, and let τ be an automorphism of the compact group G

such that τ r = 1. As before, twisted conjugacy is the action of G on itself given by
g : x → g · x · τ g −1 . A clue to the meaning of twisted conjugacy is given by the following
result.
306 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313

Proposition 6. Suppose that G is finite, and τ is an automorphism of G. The number

of twisted conjugacy classes of G with respect to τ equals the number of τ -invariant
irreducible representations of G.

This is implied by Proposition 1.2 of Digne [7], which is stated without proof; it was
most likely known to Shintani. We give a simple counting argument.

Proof. There are two permutation actions of τ which we may consider: the action on
the conjugacy classes of G, and the action on the irreducible representations of G. Both
permutation actions are realized on the center of C[G], the first by taking the basis of
the center comprised of the conjugacy class sums, the second by taking the basis of
central idempotents parametrized by the irreducible representations. Thus these actions are
equivalent, and the number of τ invariant representations equals the number of τ -invariant
conjugacy classes.
We must therefore show that the number of τ -invariant conjugacy classes equals the
number of twisted conjugacy classes. We will count the number N of solutions (x, g) to
the equation g −1 xg = τ x in two different ways.
First, N equals

 
 
# g ∈ G | g −1 xg = τ x = CG (x)
x ∼τx x ∼τx

because given x which is conjugate to τ x, we may count the number of g conjugating x

to τ x by fixing one such g, then noting that any other such g must differ from that one by
an element of the centralizer |CG (x)|. Because the number of elements of the conjugacy
class of x is |G|/|CG (x)|, we see that N is |G| times the number of τ -invariant conjugacy
classes.
On the other hand, N equals

 
# x ∈ G | xg τ x −1 = g .
g∈G

The index in G of #{x ∈ G | xg τ x −1 = g} is the order of the twisted conjugacy class of g,

so this is |G| times the number of twisted conjugacy classes. 2

Returning to the more general case of a compact group, let (π, V ) be a complex
representation of G with character η. We do not assume that π is irreducible. Let
(σ, W ) be another representation of G. We assume that σ is irreducible, and moreover
that τ σ ∼
= σ . This implies that there is a linear transformation J : W → W such that
J ◦ σ (g) = σ (τ g) ◦ J .
Since we are assuming σ is irreducible, Schur’s Lemma implies that J r acts by a
scalar on W . We may therefore choose a normalization of J so that J r is the identity
transformation of W .
Let Eτ (π, σ ) be the vector space of r-linear maps T : V × · · · × V → W which satisfy
    r−1  
T π(g)v1 , π τ g v2 , . . . , π τ g vr = σ (g)T (v1 , . . . , vr ). (8)
 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 307

We define a linear transformation cr : Eτ (π, σ ) → Eτ (π, σ ) by

(cr T )(v1 , . . . , vr ) = J T (vr , v1 , . . . , vr−1 ). (9)

Let ετ (π, σ ) be the trace of cr on Eτ (π, σ ).

We will denote by N : G → G the map

τ2 τ r−1
N(g) = g · τ g · g · ··· · g.

Theorem 9. We have

   
ετ (π, σ ) = η N(g) tr J ◦ σ (g) dg. (10)
G

Proof. We will make use of the group G  which is the semidirect product of G by a
cyclic group t with generator t of order t, acting on G by conjugation according to

the automorphism τ : tgt −1 = τ g for g ∈ G. Let U = V ⊗ · · · ⊗ V . We will give it a G
module structure which was noted by Shintani [22, Lemma 1-4]. Specifically, let G act by
the representation
τ   r−1 
Π(g) = π(g) ⊗ π g ⊗ ··· ⊗ π τ g ,

 by letting
which we extend to G

Π(t)(v1 ⊗ · · · ⊗ vr ) = v2 ⊗ · · · ⊗ vr ⊗ v1 .

The content of Shintani’s Lemma 1-4 is that if g ∈ G then

 
tr Π(tg) = η N(g) . (11)

Let us verify this. This follows immediately once we check that if A1 , . . . , Ar are
endomorphisms of V , then the trace of the endomorphism

v1 ⊗ · · · ⊗ vr → A2 v2 ⊗ · · · ⊗ Ar vr ⊗ A1 v1 (12)

is tr(A1 A2 · · · Ar ). If x1 , . . . , xn is a basis of V , and if Ai xj = k aikj xk , then (12) takes
xi1 ⊗ · · · ⊗ xir to


a2j2 i2 a3j3 i3 · · · arjr ir a1j1 i1 xj2 ⊗ · · · ⊗ xjr ⊗ xj1 .
j1, ...,jr

The trace is the sum of the contributions with j2 = i1 , j3 = i2 , etc., that is,


a2i1 i2 a3i2 i3 · · · arir−1 ir a1ir i1 = tr(A2 A3 · · · A1 ) = tr(A1 A2 · · · Ar ).
i1 ,...,ir
308 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313

This proves (11).

We may also extend the representation σ of G on W to G  by letting σ (t) = J .
Now let T be an r-linear form on V satisfying (8). By the universal property of the
tensor product, T induces a map U → W which is G-equivariant. Thus we may identify
the space Eτ (π, σ ) of such linear maps with HomG (U, W ). We give HomG (U, W ) a

G-module structure by the representation λ(ĝ)f = σ (ĝ) ◦ f ◦ Π(ĝ)−1 . It is clear from
the definitions that we have a commutative diagram:

Eτ (π, σ ) ∼
= HomG (U, W )
cr λ(t )

Eτ (π, σ ) ∼
= HomG (U, W )

so ετ (π, σ ) = tr λ(t) and this is what we must compute.

 → C× be the linear character which is
Let ζ be a primitive rth root of unity. Let ρ : G
trivial on G, and which maps t → ζ . We can decompose


r−1
 
HomG (U, W ) =  U, W ⊗ ρ ,
HomG k

k=0

−k f .
 (U, W ⊗ ρ ) is the space of f ∈ HomG (U, W ) for which λ(t)f = ζ
and HomG k

Therefore

r−1
 
ετ (π, σ ) = tr λ(t) = ζ −k dim HomG
 U, W ⊗ ρ .
k

i=0

 (U, W ⊗ρ ) is the inner product of the characters, this equals

Since the dimension of HomG k

r−1 
     k
ζk tr Π ĝ tr σ ĝ ρ ĝ dĝ.
i=0 
G

 Since we
Of course k ζ −k ρ k is r times the characteristic function of the coset tG in G.
are normalizing both groups to have Haar volume one, the restriction of the normalized
 to G is 1 dg. Thus
Haar measure d ĝ on G r

ετ (π, σ ) = tr Π(tg) tr σ (tg) dg
G

and the theorem is proved. 2

Theorem 10. The character sum (10) is an algebraic integer, in fact an element of
Z[e2πi/r ].
 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 309

Proof. The linear endomorphism cr of Eτ (π, σ ) has order r, so its eigenvalues are rth
roots of unity. Therefore its trace is in Z[e2πi/r ]. 2

5. A formula of Frobenius

The purpose of this section is to briefly indicate a connection between Theorem 9 and
Frobenius–Schur duality, and the representation
 theory of the symmetric group.
Let V = Cn . The vector space r V has commuting actions of U (n) and Sr , a left
action of U (n):

g · (v1 ⊗ · · · ⊗ vr ) = gv1 ⊗ · · · ⊗ gvr , g ∈ U (n),

and a right action of Sr :

(v1 ⊗ · · · ⊗ vr ) · σ = vσ (1) ⊗ · · · ⊗ vσ (r) , r ∈ Sr .

 
If (θ, Mθ ) is a representation of Sr then Vθ = ( r V ) C[Sr ] Mθ is a representation of
U (n). 
It can be checked that the vector space of endomorphisms of r V which commute with
the linear transformations of Sr is spanned by the linear transformations of U (n). It follows
that Vθ , if nonzero is irreducible, and (using the fact that the Mθ are self contragradient)


r 
V∼
= Vθ ⊗ Mθ , (13)
θ

where θ runs through the irreducible representations of Sr such that Vθ is nonzero. (This
will be all irreducible representations of Sr if n  r.) Let ηθ : U (n) → C be the character
of Vθ , and let χθ be the character of θ . This fact is known as Frobenius–Schur duality.
Now let λ  be a partition of r. Thus λ = (λ1 , λ2 , . . . , λl ) where λ1  λ2  · · · are positive
integers and λi = r. Let cλ be a representation of the conjugacy class of Sr of cycle
type λ. For example if λ = (3, 2, 2) then we can take cλ = (123)(45)(67). A fundamental
formula, known to Frobenius, asserts that if g ∈ U (n), then
  

tr g λi = χθ (cλ )ηθ (g). (14)
θ

See Frobenius [9, (1.) of Section 4]. Frobenius’ result is stated in terms of symmetric
polynomials, but it can be translated into a statement about characters of irreducible
representations of unitary groups. His m is our n, and the quotient


[λ1 , . . . , λm ]x λ1 · · · x λm ∆(x1 , . . . , xm ),

where the summation is over permutations of a fixed (λ1 , . . . , λm ), is a Schur polynomial.

Its values on the eigenvalues of g ∈ U (m) is the character of an irreducible representation
of U (m), corresponding to our ηθ .
310 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313

We will use Theorem 9 to prove (14) in the special case where λ = (r), so cλ is a
r-cycle. The general case (14) can be deduced from the special case using considerations
about induction from Sλ1 × Sλ2 × · · · to Sr , but we will omit this discussion. It may be
instructive to ask whether (14) suggests generalizations of Theorem 9, though we have not
pursued this question. We show therefore that
 

tr g r = χθ (cr )ηθ (g),
θ

where cr is an r-cycle. Indeed, those characters of irreducible representations of U (n)

which are homogeneous polynomials of degree r are precisely the ηθ . Since tr(g r ) is a
class function on U (n) and a homogeneous polynomial of degree r, there are coefficients
Aθ such that
 
tr g r = Aθ ηθ (g),

and by orthogonality

 
Aθ = tr g r ηθ (g) dg.

By Theorem 9, this is the trace of cr on the space of r-linear forms

V × · · · × V → Vθ

which are U (n)-equivariant, or equivalently on HomU (n) ( k V , Vθ ) which by (13) is
isomorphic to Mθ as an Sk -module. Therefore Aθ = χθ (cr ). (We have used the fact that
Aθ is real.)

6. Base change for GL(n, Fq )

In this section we will consider some character sums which are similar to the ones
in the previous section. We recall some results of Shintani [22] on base change for
GL(n). Shintani found, first for representations of GL(n), and later for automorphic
representations associated with holomorphic modular forms, character identities which led
to the modern theory of base change for automorphic representations, by Langlands and
by Arthur and Clozel.
We will give an apparently new formulation of Shintani’s base change correspondence
in terms of certain character sums. These sums resemble those of Theorem 9, but differ in
that η is now not a character of G, but of its τ -fixed points.
Let G = GL(n, Fq r ), and let τ : G → G be the Frobenius automorphism, corresponding
to the Galois automorphism x → x q of Fq . The fixed subgroup τ is H = GL(n, Fq ). Let
N : G → G be the “norm map” defined by (1). We let G act on itself by twisted conjugacy:
g : x → g · x · τ g −1 . The orbits are naturally called twisted conjugacy classes. Since
 
N g · x · τ g −1 = g · N(x) · g −1 ,
 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 311

N takes twisted conjugacy classes to ordinary conjugacy classes. It is not necessarily

true that N(x) ∈ H . However N(x) is conjugate in G to elements of a unique conjugacy
class in H , so N induces a bijection between the twisted conjugacy classes of G and the
conjugacy classes of H .
If η is a class function on H , we extend it to a class function on G by:

η(h) if g is conjugate in G to h ∈ H,
η(g) = (15)
0 if g is not conjugate to an element of H.

The second case is only for definiteness, since η will appear principally in composition
with N : G → G and N(g) is always conjugate to an element of H .
Suppose that χ is an irreducible character of G which is τ -invariant, that is, such that
τ χ = χ . If (π, V ) is an irreducible G-module affording χ , then τ π ∼ π , so there exists an
=
intertwining map I : V → V such that
 
I ◦ π(g) = π τ g ◦ I. (16)

This identity only determines I up to a scalar, but Shintani’s correspondence will specify
I uniquely.

Theorem 11 (Shintani). There is a bijection between the τ -invariant irreducible characters

χ of G and the irreducible characters η of H . In this correspondence, we have
   
η N(g) = tr I ◦ π(g) , (17)

where η is extended to a class function on G by (15), and I : V → V is a linear

transformation satisfying (16), where (π, V ) is a representation affording the character χ .
The map I satisfies I r = ±1, and I r = 1 if r is even.

If χ and η are related as in Shintani’s theorem, then we say that χ is the base change
of η.

Theorem 12. Let ξ and η be irreducible characters of G and H , respectively. Then

1
   η(1)/ξ(1) if ξ is the base change of η,

ξ(g)η N(g) =
|G| 0 otherwise.
g∈G

We have extended η to a class function on G by (15).

Proof. Let χ be the irreducible character of G which is the base change of η. By

Shintani’s theorem this means that η(N(g)) = tr(I ◦ π) where (π, V ) is a representation
with character χ and I : V → V is as in (17).
312 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313

Suppose first that ξ = χ . Let · , · be a G-invariant inner product on V . Then it is also


G-invariant. Let v1 , . . . , vd 
be an orthonormal basis of V . The trace of any linear transfor-
mation T : V → V equals i T vi , vi . Since η(N(g)) = tr(I ◦ π(g)) = tr(π(g) ◦ I ), we
have
1
  1

  
χ(g)η N(g) = π(g)vi , vi π(g)I vj , vj .
|G| |G|
g∈G g∈G i,j

Interchanging the order of summation and using Schur orthogonality for matrix coefficients
this equals

1

vi , I vj vi , vj .
χ(1)
i,j

We have vi , vj = δij (Kronecker δ), so we obtain

1
1
vi , I vi = tr(I ).
χ(1) χ(1)
i

But using (17) with g = 1, tr(I ) = η(1), so this is just η(1)/χ(1).

If ξ and χ are different characters, we proceed similarly, but we have

1


 
π (g)vi , vi π(g)I vj , vj ,
|G|
g∈G i,j

where (π
, V
) is an irreducible representation affording ξ , and this is zero by Schur
orthogonality. 2

There is considerable literature generalizing Shintani’s results to other algebraic groups

over finite fields, and our Theorem 12 should extend to these. Another example is as
follows. Let H be a finite group (nonabelian if the example is not to be too trivial).
Let G = H × · · · × H (r copies); let τ be the automorphism cyclicly permuting the
components. An analog of Theorem 12 is true for this example, also.

Acknowledgments

We thank Persi Diaconis, Ryan Vinroot, editor Jan Saxl and the referee for helpful
comments. This work was supported in part by NSF grant DMS-9970841.

References

[1] R. Baddeley, Models and involution models for wreath products and certain Weyl groups, J. London Math.
Soc. (2) 44 (1991) 55–74.
 D. Bump, D. Ginzburg / Journal of Algebra 278 (2004) 294–313 313

[2] J.L. Berggren, Finite groups in which every element is conjugate to its inverse, Pacific J. Math. 28 (1969)
289–293.
[3] J. Bernstein, I. Gelfand, S. Gelfand, Models of representations of compact Lie groups, Funct. Anal. Appl. 9
(1976) 322–324.
[4] D. Bump, Automorphic Forms and Representations, Cambridge University Press, 1997.
[5] J. Conway, R. Curtis, S. Norton, R. Parker, R. Wilson, Atlas of Finite Groups, 1985.
[6] C. Curtis, Pioneers of Representation Theory: Frobenius, Burnside, Schur and Brauer, Amer. Math. Soc.,
1999.
[7] F. Digne, Shintani descent and L functions on Deligne–Lusztig varieties, in: P. Fong (Ed.), The Arcata
Conference on Representations of Finite Groups, in: Proc. Sympos. Pure Math., part 1, vol. 47, Amer. Math.
Soc., 1986, pp. 61–68.
[8] W. Feit, Characters of Finite Groups, Benjamin, New York–Amsterdam, 1967.
[9] G. Frobenius, Über die charakterisischen Einheiten der symmetrischen Gruppe, Sitz.ber. Akad. Wiss. Berlin
(1903) 504–537.
[10] G. Frobenius, I. Schur, Über die rellen Darstellungen der endlichen Gruppen, Sitz.ber. Akad. Wiss. Berlin
(1906) 186–208.
[11] J. Fulman, R. Guralnik, Conjugacy class properties of the extension of gl(n, Fq ), J. Algebra (2004).
[12] R. Gow, Real representations of the finite orthogonal and symplectic groups of odd characteristic,
J. Algebra 96 (1985) 249–274.
[13] R. Gow, Properties of the characters of the finite general linear group related to the transpose-inverse
involution, Proc. London Math. Soc. (3) 47 (1993) 493–506.
[14] P. Hall, On a theorem of Frobenius, Proc. London Math. Soc. (2) 40 (1935) 468–501.
[15] R. Howlett, C. Zworestine, On Klyachko’s model for the representations of finite general linear groups, in:
Representations and Quantizations (Shanghai, 1998), 2000, pp. 229–245.
[16] N. Inglis, R. Richardson, J. Saxl, An explicit model for the complex representations of Sn , Arch. Math.
(Basel) 54 (1990) 258–259.
[17] N.F.J. Inglis, J. Saxl, An explicit model for the complex representations of the finite general linear groups,
Arch. Math. (Basel) 57 (5) (1991) 424–431.
[18] N. Kawanaka, H. Matsuyama, A twisted version of the Frobenius–Schur indicator and multiplicity-free
permutation representation, Hokkaido Math. J. 19 (1990) 495–508.
[19] A. Klyachko, Models for complex representations of the groups GL(n, q), Mat. Sb. (N.S.) 120 (162) (1983)
371–386.
[20] I.G. Macdonald, Symmetric Functions and Hall Polynomials, second ed., in: Oxford Math. Monogr.,
Clarendon Press, Oxford University Press, New York, 1995, with contributions by A. Zelevinsky, Oxford
Science Publications.
[21] D. Prasad, On the self-dual representations of finite groups of Lie type, J. Algebra (1998).
[22] T. Shintani, Two remarks on the irreducible characters of finite general linear groups, J. Math. Soc. Japan 28
(1976) 396–414.
[23] L. Solomon, On the sum of the elements in the character table of a finite group, Proc. Amer. Math. Soc. 12
(1961) 962–963.
[24] J. Soto-Andrade, Geometrical Gelfand models, tensor quotients, and Weyl representations, in: P. Fong (Ed.),
The Arcata Conference on Representations of Finite Groups, in: Proc. Sympos. Pure Math., part 1, vol. 47,
Amer. Math. Soc., 1986, pp. 305–316.
[25] R. Steinberg, Lectures on Chevalley Groups, Yale University Press, 1968.
[26] R. Vinroot, PhD thesis, Stanford University, 2004, in preparation.