Journal of Algebra V322 I11.2

Journal of Algebra
EDITOR-IN-CHIEF
Michel Broué
Journal of Algebra
Institut Henri Poincaré, 11, rue Pierre et Marie Curie,
F-75005 Paris, France
E-mail: jalgebra@ihp.jussieu.fr
EDITORIAL BOARD
Nicolás Andruskiewitsch Kazuhiko Kurano Aner Shalev

Universidad de Córdoba Meiji University Hebrew University
Facultad de Matemática, kurano@math.meiji.ac.jp Institute of Mathematics
Astronomía y Física shalev@math.huji.ac.il
andrus@famaf.unc.edu.ar Gus I. Lehrer
University of Sydney Ronald Solomon
Luchezar L. Avramov School of Mathematics and Statistics Ohio State University*
University of Nebraska g.lehrer@maths.usyd.edu.au solomon@math.ohio-state.edu
avramov@math.unl.edu
Martin Liebeck J.T. Stafford
Eva Bayer-Fluckiger Imperial College, London* Manchester University*
École Polytechnique Federale m.liebeck@ic.ac.uk Toby.Stafford@manchester.ac.uk
de Lausanne*
eva.bayer@epfl.ch Peter Littelmann Gernot Stroth
Universität zu Köln Universität Halle–Wittenberg
Steven Dale Cutkosky Mathematisches Institut stroth@mathematik.uni-halle.de
University of Missouri littelma@math.uni-koeln.de
cutkoskys@missouri.edu
Michel Van den Bergh
Dihua Jiang Laurent Moret-Bailly Limbergs Universitair Centrum
University of Minnesota Université de Rennes 1 Department WNI
School of Mathematics laurent.moret-bailly@univ-rennes1.fr vdbergh@luc.ac.be
dhjiang@math.umn.edu
Leonard L. Scott, Jr. Changchang Xi
Masaki Kashiwara University of Virginia* Beijing Normal University
University of Kyoto lls2l@virginia.edu School of Mathematical Sciences
masaki@kurims.kyoto-u.ac.jp xicc@bnu.edu.cn
Vera Serganova
E.I. Khukhro UC Berkeley Efim Zelmanov
Institute of Mathematics, SO RAN Department of Mathematics University of California, San Diego*
khukhro@yahoo.co.uk serganova@math.berkeley.edu ezelmano@euclid.ucsd.edu
EDITOR OF THE SECTION ON COMPUTATIONAL ALGEBRA

Gerhard Hiss
Lehrstuhl D für Mathematik
RWTH Aachen
E-mail: gerhard.hiss@math.rwth-aachen.de
SECTION ON COMPUTATIONAL ALGEBRA
Jon Carlson Meinolf Geck Gunter Malle

University of Georgia* Aberdeen University Universität Kaiserslautern
jfc@math.uga.edu King’s College Fachbereich Mathematik
Department of Mathematical Sciences malle@mathematik.uni-kl.de
John Cremona m.geck@maths.abdn.ac.uk
University of Warwick Eamonn O’Brien
Derek Holt The University of Auckland*
Mathematics Institute University of Warwick
j.e.cremona@warwick.ac.uk e.obrien@auckland.ac.nz
Mathematics Institute
dfh@maths.warwick.ac.uk Bruno Salvy
Patrick Dehornoy INRIA Rocquencourt
Université de Caen William M. Kantor
University of Oregon* Algorithms Project
Laboratoire de Mathématiques bruno.salvy@inria.fr
patrick.dehornoy@math.unicaen.fr kantor@darkwing.uoregon.edu
Reinhard Laubenbacher Jean-Yves Thibon
Harm Derksen Virginia Tech Université de Marne-la-Vallée
University of Michigan* Virginia Bioinformatics Institute Institut Gaspard Monge
hderksen@umich.edu reinhard@vbi.vt.edu jyt@univ-mlv.fr
FOUNDING EDITOR: Graham Higman, 1964–1984

EDITOR-IN-CHIEF: Walter Feit, 1985–2000
*Department of Mathematics
Journal of Algebra 322 (2009) 3797–3822
Contents lists available at ScienceDirect
Journal of Algebra
www.elsevier.com/locate/jalgebra
Irreducibility criteria for skew polynomials

Richard C. Churchill a,b , Yang Zhang c,∗
a
Department of Mathematics, Hunter College and Graduate Center, CUNY, New York, NY 10065, United States
b
University of Calgary, Canada
c
Department of Mathematics, University of Manitoba, Winnipeg, MB R3T 2N2 Canada
a r t i c l e i n f o a b s t r a c t
Article history: The paper presents two irreducibility criteria for the elements of
Received 28 August 2007 a large class of skew-polynomial rings. The proofs rely heavily on
Available online 29 September 2009 non-commutative valuations and extensions thereof. The results
Communicated by Bruno Salvy
apply, in particular, to ordinary linear differential operators and
Keywords:
linear difference operators having coefficients in not-necessarily-
Skew polynomials commutative fields with valuations.
Newton polygons © 2009 Elsevier Inc. All rights reserved.
Eisenstein criterion
1. Introduction
In recent decades irreducibility criteria and factorization algorithms for skew polynomials have at-
tracted increasing attention. Methods have now been developed involving Galois groups, p-curvature,
and eigenrings, etc. For an introduction to the literature see, e.g., Singer [25], Ulmer and Weil [26],
van der Put [27] and van Hoeij [30].
In this paper we are concerned with irreducibility. Specifically, we offer two irreducibility cri-
teria for the elements of a large class of skew-polynomial rings. The results apply, in particular,
to ordinary linear differential operators and linear difference operators having coefficients in not-
necessarily-commutative fields with valuations.
The first criterion (Theorem 8.1) generalizes a result of Dumas [10], which in van der Waerden [29,
p. 76] is described as a “far-reaching generalization” of the classical Eisenstein criterion. The second
(Theorem 9.7) extends Kovacic’s work on the Eisenstein criterion (see Kovacic [16]).
The proof of Theorem 8.1 is formulated in terms of Newton polygons. This is not absolutely nec-
essary, but it highlights the geometric ideas underlying Dumas’ original argument. The proofs of both
of our irreducibility criteria rely heavily on non-commutative valuations and extensions thereof.
* Corresponding author.
E-mail addresses: rchurchi@hunter.cuny.edu (R.C. Churchill), zhang39@cc.umanitoba.ca (Y. Zhang).
0021-8693/$ – see front matter © 2009 Elsevier Inc. All rights reserved.
doi:10.1016/j.jalgebra.2009.08.012
3798 R.C. Churchill, Y. Zhang / Journal of Algebra 322 (2009) 3797–3822
A skew-polynomial ring R [x; σ , δ] involves a not-necessarily-commutative ring R with unity, a ring

embedding σ : R → R, and a skew-derivation δ : R → R. (All the non-standard terms will be defined.)
In the literature it is common practice to assume δ σ = σ δ when dealing with such entities, but we
do not require this hypothesis.
Throughout the paper R denotes a ring, not necessarily commutative, with unity 1 = 1 R . All rings are
assumed to have unities, and ring homomorphisms are assumed to preserve unities. Unless specifically stated
to the contrary, we make no assumptions on the characteristic of R. We refer to R as a domain when R \ {0} is
assumed closed under multiplication (Cohn [6, §2.1, p. 8]), and as an integral domain when this is the case
and R is also assumed commutative.
By an R-algebra we mean a left and right R-module M admitting an associative R-bilinear mapping
(m1 , m2 ) ∈ M × M → m1m2 ∈ M which we refer to as “multiplication.” When confusion might otherwise
result we write m1 m2 as m1 · m2 . By a “unity” of this multiplication we mean a necessarily unique element
1 = 1 M ∈ M such that 1 · m = m · 1 = m for all m ∈ M.
F denotes the ring of functions f : R → R (wherein “multiplication” refers to composition). Elements
s ∈ R are identified with elements of F by means of left multiplication, i.e., by means of the correspondence
s → (ms : r ∈ R → sr ∈ R ). This is easily seen to endow F with the structure of an R-algebra, and that
structure is always assumed.
Z and Z+ always denote the usual ring of integers and the subset {1, 2, 3, . . .} of positive integers respec-
tively. Q, R and C always denote the rational, real and complex fields.
2. Background on skew polynomial rings
Let σ : R → R be an injective ring homomorphism. An additive group homomorphism δ : R → R is

a σ -derivation if
δ(rs) = σ (r )δ(s) + δ(r )s, r, s ∈ R . (2.1)
When σ is the identity automorphism 1 this is the standard definition of a derivation. When σ is
understood a σ -derivation is sometimes called a skew derivation.
When δ satisfies (2.1) we follow custom and use the notations δ(r ) and δr interchangeably, even
though δ is generally not R-linear. Note that δ 1 = δ(−1) = 0.
An R-algebra P with unity 1 = 1 P is a skew polynomial ring if the following three conditions are
satisfied.
SPR. 1. There is a distinguished element x ∈ P such that P is a free left R-module with basis (x j )∞
j =0
,
where x0 := 1.
n
By SPR. 1 any non-zero element p ∈ P has a unique representation of the form p = j =0 r j x
j
with r j ∈ R for j = 0, . . . , n and rn = 0. We refer to deg( p ) := n 0 as the degree of p, and to the

monomials r j x j with r j = 0 as the monomials of (or within) p.
The degree of the zero element is defined to be −∞, where −∞ ∈ / Z is assumed to satisfy
−∞ + n := n + −∞ = −∞ for any n ∈ Z ∪ {−∞}.
SPR. 2. The degree function is additive, i.e., p , q ∈ P one has
deg( pq) = deg( p ) + deg(q).
SPR. 3. There is an injective ring homomorphism σ : R → R and a σ -derivation δ : R → R such that
xr = σ (r )x + δ r for all r ∈ R . (2.2)

R.C. Churchill, Y. Zhang / Journal of Algebra 322 (2009) 3797–3822 3799
One sees from SPR. 1 that P = R [x] when both are regarded as R-modules, and for this reason the
custom is to replace the notation P with R [x]. The skew-polynomial ring is expressed accordingly as
R [x; σ , δ], and the symbol P is thereby freed for other duties. Nevertheless, we find it convenient to
retain the notation P in the presentation of examples.
The systematic study of skew polynomial rings was initiated by O. Ore in the early 1930s [21,22].
Our definition of such a ring is adapted from Ore [22, Chapter I, §1, p. 481].
Examples 2.1. Let R be a commutative integral domain.
(a) Polynomials. The usual polynomial ring P = R [x] in a single indeterminate x can easily be given
the structure of a skew polynomial ring: take σ = 1 (= id R ) and δ = 0 (i.e., the zero mapping
r → 0). When this structure is assumed one retains the notation R [x], i.e., one does not write
R [x; σ , δ].
(b) Differential operators. Let σ = id R and let δ : R → R be a derivation. Assuming the R-subalgebra
P := R [δ] ⊂ F generated by δ is free as an R-module with basis (δ j )∞ j =0
we see that P is a skew
polynomial algebra. We refer to such a skew polynomial algebra as an R-algebra of differential
operators. Here one writes R [δ] in place of R [x; 1, δ] (because x = δ ).
A sufficient hypothesis for freeness is that R contain the rational field Q as well as a primitive
n1, i.e., j an element r such that δr = 1. Indeed,
of if this module is not free we can write 0 =
j =0 r j δ ∈ F for some integer n > 0, where δ 0
:= 1 (:= id R ) and at least one r j = 0. Applying
the sum to 1 we find that r0 = 0, then successively to r , r 2 /2!, r 3 /3!, . . . , r n /n! that r1 = r2 = r3 =
· · · = rn = 0, and we have a contradiction.
We should point out that in the context of this particular example what we have called the
“degree” of an element is more commonly called the “order” of that element. For example, we
3
d z d d
would describe the linear operator L := dz 3 + z2 +1 dz + z ∈ C( z )[ dz ] has having degree 3, and
7
write deg( L ) = 3, whereas a classical analyst would more likely refer to 3 as the “order,” as in
describing L as a “third-order operator.” We have made our choice to handle polynomials and
operators simultaneously, and because switching to “order” would simply replace one problem
with another, i.e., an ambiguity in terminology would result when valuations are introduced.
(c) Difference operators. When σ : R → R is an injective ring homomorphism it is a simple matter to
check that δ := σ − 1 is a skew derivation. Assuming the R-subalgebra P := R [σ ] ⊂ F generated
by σ is free as an R-module with basis (σ j )∞ j =0
we see that P is a skew polynomial algebra. We
refer to such a skew polynomial algebra as an R-algebra of difference operators. In this context
one writes R [σ ] in place of R [x; σ , δ] = R [x; σ , σ − 1] (because x = σ ).
Note from σ = δ + 1 that when (σ j )∞ j =0
is a basis the collection (δ j )∞ j =0
will also span R [σ ].
A sufficient hypothesis for the second collection to be a basis is that R contain an element s1
such that δ s1 = 1 and, for j > 1, elements s j such that δ s j = s j −1 . As above, if the module is not
n
free on (δ j )∞
j =0
we can write 0 = j =0 r j δ for some integer n > 0, where δ := 1 (:= id R ) and at
j 0
least one r j = 0. Applying the sum to 1 we find that r0 = 0, then successively to s1 , s2 , s3 , . . . , sn

that r1 = r2 = r3 = · · · = rn = 0, and we have a contradiction.
Freeness on (σ j )∞ j =0
follows from freeness on (δ j )∞ j =0
(and conversely). If not we can write 0 =
n
j =0 r j σ j ∈ F for some n 0 and rn = 0. Replacing σ by δ + 1 then converts this expression to
n j ∞
j =0 r̂ j δ with r̂n = rn = 0. Since this contradicts freeness on (δ ) j =0 , the claim follows.
j
the form
(The argument is easily adapted to establish the converse.)
The following notational convention (from Jacobson [14, Chapter I, §1, pp. 2–3]) can be quite useful
for organizing calculations in a skew
polynomial ring R [x; σ , δ]. Let n and j be integers with n > 0
n
and 0 j n. Then there are j products λ1 λ2 · · · λn ∈ F in which λi ∈ {σ , δ} for 1 i n and σ
appears precisely j times. The sum of these products, which is again an element of F , is denoted S nj .
Finally, define S 00 to be the identity mapping 1, and S nj to be the zero mapping 0 when this function
has not otherwise been defined. One then has

S 00 = 1 (:= id), S 10 = δ, S 11 = σ ,
, (2.3)
S n,n−1 = σ n−1 δ + σ n−2 δ σ + · · · + δ σ n−1 and S nn = σ n
as well as the recursion relation
S nj = δ S n−1, j + σ S n−1, j −1 (2.4)
for all n 1 and all 1 j n − 1. The fundamental identity involving these entities is

i
xi b = ( S ik b)xk for all b ∈ R and all i 1, (2.5)
k =0
where S ik b := S ik (b), which has the immediate consequence

i
axi · bx j = a( S ik b)xk+ j , a , b ∈ R , i , j ∈ Z+ . (2.6)
k =0

One computes a product ( i ai x
i
) · ( j b j x j ) in R [x; σ , δ] by means of (2.6), i.e.,

i j i
i
ai x · j
b jx = ai x b j x = ai ( S ik b j )xk+ j . (2.7)
i j i, j i , j k =0
For additional background on skew polynomial rings, see, for example, Cohn [5,6] and Jacobson
[13,14].
3. Background on valuations and valuation rings
In this section we recall the definition of a non-commutative valuation and discuss those proper-
ties needed in our later work. For a more detailed discussion see, e.g., Cohn [6]. In the initial definition
we adopt the practice of that reference of using additive notation for expressing the binary operation
of a possibly non-commutative group.
Non-commutative valuation theory dates back to work of Artin, Noether et al. in the 1930s. Pre-
sentations in a more classical vein can be found in Schilling [24] and in Artin [2].
An ordered group is a (not necessarily commutative) group Γ with a total ordering α β preserved
by the group operation:
α β, α
β
⇒ α + α
β + β
for all α , α
, β, β
∈ Γ.
When dealing with an ordered group Γ it is customary to augment the structure to that of a (gener-
ally non-commutative) monoid by adding a symbol ∞ satisfying both ∞ > α and
α+∞=∞+α=∞+∞=∞
for all α ∈ Γ . We will follow this practice.

For the remainder of the section R is a ring (not necessarily commutative) and Γ is an ordered group
(again not necessarily commutative).
A valuation on R with values in Γ is a function ν : R → Γ ∪ {∞} satisfying the following properties

for all a, b ∈ R:
(V.1) ν assumes at least two distinct values;

(V.2) ν (a + b) min{ν (a), ν (b)}; and
(V.3) ν (ab) = ν (a) + ν (b).
When this is the case one refers to R as a valued ring, or as a valued field when R is a field, and to Γ
as the value group (of the valuation). Arguing by induction from (V.2) we see that

n

ν aj min ν (a j ) . (3.1)
j
j =1
When Γ is a cyclic subgroup of (R, +, 0) a valuation ν : R → Γ ∪ {∞} is said to be discrete.

Readers are assumed familiar with valuations in the context of commutative rings and commu-
tative value groups, and will therefore recognize that when R and Γ satisfy these conditions our
definition of a valuation reduces to the usual one. Our initial examples are the standard ones for the
commutative context. A more substantial example is given in Proposition 4.5.
Examples 3.1.
(a) The usual p-adic valuation ν p : Q → Z ∪ {∞}, for any prime p ∈ Z+ , is a discrete valuation in the
sense defined above.
(b) Let P1 denote the Riemann sphere and assume the usual identification of the field C( z) of rational
functions with the field M(P1 ) of meromorphic functions on P. Then for any point z0 ∈ P1 a
discrete valuation νz0 on M(P1 ) is defined by: νz0 ( f ) := ordz0 ( f ) if f = 0, i.e., νz0 ( f ) is the
exponent of the leading non-zero term of the Laurent expansion of f at z0 ; νz0 ( f ) := ∞ if f = 0.
Of course when (and only when) z0 = ∞ the Laurent expansion of f ∈ C( z) refers to the Laurent
expansion of f (1/t ) at t = 0. It is worth recalling that ord∞ ( f ) can be computed in a more direct
manner:
f = p /q = p ( z)/q( z) ∈ C( z) ⇒ ord∞ ( f ) = deg(q) − deg( p ). (i)
Proposition 3.2. Any valuation ν : R → Γ ∪ {∞} on R has the following properties:
(a) ν (0) = ∞. In particular, ν (0) > γ for all γ ∈ Γ (by (V.3)).

(b) ker ν := {a ∈ R | ν (a) = ∞} is a proper two-sided ideal of R (by (V.1)). In particular, ker ν = {0} if R is a
skew field or a simple ring. Moreover, R / ker ν is an integral domain.
(c) ν (1) = 0 and ν (a) = ν (−a).
(d) When a ∈ R is invertible one has ν (a−1 ) = −ν (a).
(e) If ν (a) = ν (b) then ν (a + b) = min{ν (a), ν (b)}.
The results are simple consequences of the definition of a valuation, e.g., see Cohn [6], §9.1.
It follows from Proposition 3.2(e) and induction on n that

n
ν aj = ν (ak ) if ν (ak ) < ν (ai ) for all i = k. (3.2)
j =1
Remark 3.3. It is worth pointing out explicitly that ker ν has not been defined in the usual manner,
i.e., it has not been defined as {a ∈ R | ν (a) = 0}. When ker ν = 0, as in Examples 3.1, ν is called a
proper valuation.
Let K be a skew field, i.e., a not necessarily commutative field. A subring V ⊂ K is:
• total if for any k ∈ K one either has k ∈ V or k−1 ∈ V (or both);

• invariant if for any k ∈ K \ {0} and any a ∈ V one has k−1 ak ∈ V ;
• a valuation ring if it is both total and invariant.
Of course when K is commutative the invariance condition is automatic, and one then defines a val-
uation ring using only the first condition. Indeed, in that context total subrings and valuation subrings
of K are identical concepts, and for that reason the “total” terminology is abandoned.
The connection between valuation rings and valuations is given by the initial assertion of the
following result (the proof is omitted).
Proposition 3.4. When K is a skew field and ν : K → Γ ∪ {∞} is a valuation the collection

V := a ∈ K : ν (a) 0
is a valuation ring of K . Moreover, the following assertions hold:
(a) Suppose a, b ∈ V . Then a is a left (and right ) multiple of b if and only if ν (a) ν (b).
(b) The group of units of V is U =: {a ∈ V | ν (a) = 0}.
(c) The set of non-units of V is m := {a ∈ V | ν (a) > 0}, and this set is the unique maximal left and right ideal
of V .
(d) For any a, b ∈ V the following assertions hold:
(i) ν (a) ν (b) ⇔ a = rb = br
for some r , r
∈ V .
(ii) ν (a) = ν (b) ⇔ a = rb = br
for some r , r
∈ U .
(iii) ν (a) > ν (b) ⇔ a = rb = br
for some r , r
∈ m.
4. Extending valuations to R [x; σ , δ]
Throughout the section R [x; σ , δ] denotes a skew polynomial ring and ν : R → Γ ∪ {∞} is a valuation.
The valuation ν is compatible with σ if

ν σ (a) = ν (a) for all a ∈ R , (4.1)
and is τ -compatible with δ , where τ ∈ Γ , if
ν (δa) ν (a) + τ for all a ∈ R . (4.2)
From (4.1) and induction (on i) we see that

ν σ i (a) = ν (a) for all (i , a) ∈ Z+ × R , (4.3)
and from (4.2) and induction that

ν δ i a ν (a) + i τ for all (i , a) ∈ Z+ × R . (4.4)
When both (4.1) and (4.2) hold we say that ν is τ -compatible with R [x; σ , δ].
The compatibility property is a common assumption for R [x; σ ], see, e.g., Cohn [6] and Mathiak
[18].
One can see the need for a ν -dependent τ in (4.2) by considering the rational function field
examples with ν = ordz0 for some z0 ∈ P1 in the complex case or ν = ordx0 for some x0 ∈ S1 in
the real case. We treat the complex case, but the results for the real case are identical. Choose any
non-constant r ∈ C( z).
First consider the case z0 ∈ C = P1 \ {∞} and suppose the Laurent expansion at z0 is
r = ak ( z − z0 )k + ak+1 ( z − z0 )k+1 + · · · , ak = 0. (4.5)
Then ordz0 (r ) = k and one easily verifies that

⎧
⎪
⎨ (a) ordz0 (r ) > 0 ⇒ ordz0 (r
) = ordz0 (r ) − 1,
(b) ordz0 (r ) = 0 ⇒ ordz0 (r
) 0 if r
= 0, and (4.6)
⎪
⎩
(c) ordz0 (r ) < 0 ⇒ ordz0 (r ) = ordz0 (r ) − 1.
For z0 = 0 and r = 1 + zn+1 with 1 n ∈ Z we have ordz0 (r ) = 0 and ordz0 (r

) = n > 0, showing that
strict inequality can occur in (b). In this case we see that (4.2) holds with τ = −1.
Now consider the case z0 = ∞ ∈ P1 and suppose the Laurent expansion at ∞ is
r = ak t k + ak+1 t k+1 + · · · , ak = 0, (4.7)
where, in the notation of Example 3.1(b), we have t = 1/ z. Then ord∞ (r ) = k, the chain-rule gives
d d
= −t 2 , (4.8)
dz dt
and we therefore have

⎧
⎪
⎨ (a) ord∞ (r ) > 0 ⇒ ord∞ (r
) = ord∞ (r ) + 1,
(b) ord∞ (r ) = 0 ⇒ ord∞ (r
) 0 if r
= 0, and (4.9)
⎪
⎩
(c) ord∞ (r ) < 0 ⇒ ord∞ (r
) = ord∞ (r ) + 1.
For n 1 and r = 1 + t n we have ord∞ (r ) = 0 and r

= −nt n+1 , hence ord∞ (r
) = n + 1, and we
conclude that strict inequality can occur in (b). In this case (4.2) holds with τ = 1.
As simple applications of these ideas we derive a few standard results familiar from (but seldom
proved in) elementary calculus.
Applications 4.1.
(a) There is no real rational function r ∈ R(x) such that r

= 1/x ∈ R(x), and there is no complex
rational function r ∈ C( z) such that r
= 1/ z ∈ C( z). The assertions are immediate from Eq. (4.6):
we have ord0 (1/x) = ord0 (1/ z) = −1, and those formulas show that ord0 (r
) = −1 for any r ∈
R(x) \ {0} and any r ∈ C( z) \ {0}.
(b) There is no real rational function r ∈ R(x) such that r
= 1/(x2 + 1) ∈ R(x) and there is no complex
rational function r ∈ C( z) such that r
= 1/( z2 + 1) ∈ C( z). Suppose r ∈ R(x) satisfies r
= 1/
(x2 + 1) or that r ∈ C( z) satisfies r
= 1/(z2 + 1). By writing x as z we may assume R(x) = R( z) ⊂
C( z), and thereby view r as an element of C( z) in both cases. Let s := i · ( z2 + 1)/(z2 − 1) ∈ C( z)
and set w := −i · (r ◦ s) ∈ C( z). Then one has w
= 1/ z, which contradicts the second assertion
in (a).
We now begin a series of technical results needed for extending the valuation to R [x; σ , δ].
Lemma 4.2. Suppose ν is τ -compatible with R [x; σ , δ] and S ik : R → R is as in the paragraph ending
with (2.7). Choose any b ∈ R. Then
ν ( S ik b) ν (b) + (i − k)τ for all 0 k i . (i)

Proof. For i = k = 0 we have ν ( S ik b) = ν ( S 00 b) = ν (b) + 0 · τ , and (i) therefore holds. For i > 0 and
0 k i recall that S ik is the sum ofk those monomials in σ and δ of total degree k in σ and
total degree i − k in δ , i.e., S ik = σ 1 δl1 · · · σ km δlm , where 0 k1 , . . . , km k, k1 + · · · + km = k,
0 l1 , . . . , lm i − k, and l1 + · · · + lm = i − k. As a consequence we have

ν ( S ik b) = ν σ k1 δl1 · · · σ km δlm (b)

min ν σ k1 δl1 · · · σ km δlm (b) .
Choosing exponents k
1 , . . . , km

, l
, . . . , l
which achieve this minimum we conclude that
1 m

ν ( S ik b) ν σ k1 δl1 σ k2 δl2 · · · σ km δlm (b)

= ν δl1 σ k2 δl2 · · · σ km δlm (b)

ν σ k2 δl2 · · · σ km δlm (b) + l
1 τ

= ν δl2 σ k3 · · · σ km δlm (b) + l
1 τ

ν σ k3 · · · σ km δlm (b) + l
2 τ + l
1 τ
..
.

ν (b) + lm + lm −1 + · · · + l 2 + l 1 τ
= ν (b) + (i − k)τ .
(Example: ν (δ 2 σ δ(b)) ν (σ δ(b)) + 2τ = ν (δ(b)) + 2τ ν (b) + τ + 2τ = ν (b) + 3τ = ν (b) + (4 − 1)τ .)

The result follows. 2
Corollary 4.3. For any a, b ∈ R one has
ν (aS ik b) ν (a) + ν (b) + (i − k)τ for all 0 k i .
Lemma 4.4. Suppose τ ∈ Γ and that ν is τ -compatible with R [x; σ , δ]. Choose non-zero a, b ∈ R, i , j ∈ Z+ ,
and write

i
axi · bx j = a( S ik b)xk+ j (i)
k =0
(as in (2.6)). Then

⎫
min ν a( S ik b) + (k + j )τ = ν (a) + ν (b) + (i + j )τ ⎪
⎬
k
= ν (ab) + (i + j )τ ⎪
. (ii)
⎭
= ν a( S ii b) + (i + j )τ
Proof. From Lemma 4.2 one has

ν a( S ik b) + (k + j )τ ν (ab) + (i + j )τ .
For the summation term in (i) corresponding to k = i we see from (4.3) that ν (a( S ii b)) = ν (aσ i (b)) =
ν (a) + ν (σ i (b)) = ν (a) + ν (b) = ν (ab), showing that equality is achieved when k = i. 2
Under the assumption that ν is τ -compatible with R [x; σ , δ] define ν̃ : R [x; σ , δ] → Γ ∪ {∞} by

l

k
ν̃ ck x := min ν (ck ) + kτ . (4.10)
k
k =0
Note from Lemma 4.4 and the summation conventions relating to ∞ that

ν̃ axi · bx j = ν̃ axi + ν̃ bx j . (4.11)
Proposition 4.5. When ν is τ -compatible with R [x; σ , δ] the function ν̃ defined in (4.10) is a valuation ex-
tending ν . Moreover, ν̃ is proper when ν has this property.
We include a proof for completeness. The proposition is undoubtedly well-known, but arguments
establishing the result are not easily located.
Proof. That ν̃ has values in Γ ∪ {∞} is obvious; all we need do is verify (V.1)–(V.3) of the definition
of a valuation. The cases involving the value ∞ are safely left to the reader.
(V.1): This is clear from the fact that ν̃ extends a function having this property.
n m
To verify (V.2) and (V.3) choose any p = i =0 ai x
i
,q = j =0 b j x
j
∈ R [x; σ , δ].
(V.2): We have (with the understanding that coefficients ak or bk with k greater than the degree of
the associated polynomial are zero)

k
ν̃ ( p + q) = ν̃ (ak + bk )x
k

= min ν (ak + bk ) + kτ
k

min min ν (ak ), ν (bk ) + kτ
k

= min min ν (ak ) + kτ , ν (bk ) + kτ
k

min min ν (ak ) + kτ , min ν (bk ) + kτ
k k

= min ν̃ ( p ), ν̃ (q) .

(V.3): From pq = i j ai x
i
b j x j (induction on) (V.2) and (4.11) we have

ν̃ ( pq) min ν̃ ai xi b j x j
ij

= min ν̃ ai xi + ν̃ b j x j
ij

= min ν̃ ai xi + min ν̃ b j x j
i j
= ν̃ ( p ) + ν̃ (q).
To establish (V.3) we must show that the inequality in the first line of the preceding computation
is actually an equality, i.e., that

ν̃ ( pq) = min ν (ai ) + i τ + ν (b j ) + j τ . (i)
ij
To this end imagine the product pq expressed in the form c 0 + c 1 x + · · · + cn+m xn+m ; we then have,
by definition,

ν̃ ( pq) = min ν (ck ) + kτ . (ii)
k
From (2.7) each ck is seen to be a sum of terms of the form
ai S i ,k− j b j , with k i + j ,
whereupon it follows from (3.1) that

ν (ck ) + kτ min ν (ai S i,k− j b j ) + kτ . (iii)
j ki + j
For any of the terms appearing within the bracket in (iii) we see from Corollary 4.3 that

ν (ai S i,k− j b j ) + kτ ν (ai ) + ν (b j ) + (i + j − k)τ + kτ
, (iv)
= ν (ai ) + i τ + ν (b j ) + j τ
and we therefore have

ν (ck ) + kτ min ν (ai ) + i τ + ν (b j ) + j τ : j k i + j for all 0 k n + m. (v)
i, j
Let i 0 and j 0 be the maximal choices for i and j which minimize ν (ai ) + i τ and ν (b j ) + j τ
respectively. We then see from (ii) and (v) that
ν̃ ( pq) ν (ai0 ) + i 0 τ + ν (b j0 ) + j 0 τ . (vi)
Set k0 := i 0 + j 0 . Recalling the final identity in (2.3) we then see that

⎫
ν (ai0 S i0 ,k0 − j0 b j0 ) + k0 τ = ν (ai0 S i0 i0 b j0 ) + k0 τ ⎪
⎪
⎪
= ν ai 0 σ i 0 b j 0 + k0 τ ⎬
i , (vii)
= ν (ai 0 ) + ν σ 0 b j 0 + k0 τ ⎪
⎪
⎪
⎭
= ν (ai 0 ) + ν (b j 0 ) + k0 τ
and we conclude that (iv) is in fact an equality for this particular choice of (k, i , j ). It is a strict
inequality for all other choices involving the same k (= k0 ). Otherwise there is a distinct pair (i , j )
satisfying both i + j k0 and ν (ai ) + i τ + ν (b j ) + j τ = ν (ai 0 ) + i 0 τ + ν (b j 0 ) + j 0 τ , whence ν (ai ) + i τ =
ν (ai0 ) + i 0 τ and ν (b j ) + j τ = ν (b j0 ) + j 0 τ . From the maximality conditions on i 0 and j 0 this forces
i i 0 and j j 0 , with at least one of these last two inequalities being strict. This gives i + j < k0 ,
thereby contradicting i + j k0 .
It now follows from (3.2) that
ν (ck0 ) + k0 τ = ν (ai0 ) + i 0 τ + ν (b j0 ) + j 0 τ ,
whereupon (i) becomes evident from (ii) and (v). 2
5. An elementary application to skew polynomials
Throughout this section R [x; σ , δ] denotes a skew polynomial ring, Γ is a (not necessarily commutative)
additive group, τ ∈ Γ , and ν : R → Γ ∪ {∞} is a proper valuation which is τ -compatible with R [x; σ , δ]. We
let ν̃ : R [x; σ , δ] → Γ ∪ {∞} denote the valuation extending ν given by Proposition 4.5.
n n n
When r ∈ R and n 1 we write r σ and r δ for σ n (r ) and δn (r ) respectively. We also write r δ as r (n) .
The following definition, wherein τ denotes a fixed element of Γ , is motivated by (4.6)(c). An
element r ∈ R \{0} has a pole of type τ at ν if for all n 0 one has both r (n) = 0 and ν (r (n) ) = ν (r )+ nτ .
The existence of such an r obviously implies δ = 0.
For example, any r ∈ C( z) having a pole at a point z0 ∈ C has a pole of type −1 at ordz0 . In
contrast, r = z2 ∈ C( z) has a pole at ∞, but it is not a pole of any type τ at ord∞ since ( z2 )(3) = 0.
Suppose γ1 , γ2 ∈ Γ . We say that γ1 is divisible by γ2 if there is a positive integer k such that
γ1 = k · γ2 (:= γ2+ γ2 + · · · + γ2 [k occurrences]), and to indicate this is the case we write γ1 /γ2 = k.
n
Suppose L = j =0 a j x ∈ R [x; σ , δ]. By a solution of the equation L y = 0 in R we mean an element
j
n n
r ∈ R such that an r δ + · · · + a1 r δ + a0 r = 0 if δ = 0, or such that an r σ + · · · + a1 r σ + a0 r = 0 if δ = 0.
n
Theorem 5.1. Let R be a domain. Suppose L = k=0 ak xk ∈ R [x; σ , δ] and L y = 0 admits a non-zero solution
r ∈ R having a pole of type τ ∈ Γ at ν . Then there must be indices 0 i < j n such that
ν (a j ) − ν (ai )
= −τ . (i)
j−i
In particular, when Γ is an additive subgroup of Z the indicated quotient must be an integer for at least one
such pair (i , j ).

k
Proof. By assumption there is an element r ∈ R \ {0} such that ak r δ = 0, wherein the prime indi-
cates summation over all k such that ak = 0, and we therefore have

k

∞ = ν (0) = ν ak r δ .
There can be no unique term in this sum having minimal value: otherwise the value of that sum
would belong to Γ by (3.2), contradicting ∞ ∈
/ Γ . As a consequence there must be indices i < j such
that
i j
ν ai r δ = ν a j r δ . (ii)
However, from the defining property of the pole-type hypothesis we have
k
ν ak r δ = ν (ak ) + ν (r ) + kτ
for all ak = 0, from which (ii) is easily seen to be equivalent to (i). 2
Corollary 5.2. Suppose τ = 0 and ν (a j ) = ν (an ) for all j such that a j = 0. Then L y = 0 has no solution in R
with a pole of type τ at ν .
ν ,τ
When τ ∈ Γ and ν is τ -compatible with R [x; σ , δ] we define the (ν , τ )-carrier Car L of a non-
n
zero skew polynomial L = j =0 a j x ∈ R [x; σ , δ] to be
j
ν ,τ
Car L := j , ν (a j ) + j τ : 0 j n and a j = 0 ⊂ Z × Γ. (5.1)
n
Corollary 5.3. Let L = j =0 a j x be a non-zero element of R [x; σ , δ] and suppose that for any two distinct
j
points (i , ν (ai ) + i τ ) and ( j , ν (a j ) + j τ ) of the (ν , τ )-carrier of L one has
ν (a j ) − ν (ai ) = (i − j )τ . (i)
Then L has no solution in R with a pole of type τ at ν .
When Γ is an additive subgroup of R condition (i) of the previous corollary has a simple geometric
interpretation.
n
Corollary 5.4. Suppose Γ is an additive subgroup of R and L = j =0 a j x
j
is a non-zero element of R [x; σ , δ].
Assume the slope
(ν (ai ) + i τ ) − (ν (a j ) + j τ )
i− j
of any line through two distinct points of the (ν , τ )-carrier of p is not zero, i.e., that none of these lines is
horizontal. Then L has no solution in R with a pole of type τ at ν .
d
Example 5.5. Let L ∈ C( z)[ dz ] be given by
d5 1 d2 1
L := − + .
dz5 z4 dz2 z3
We will use the preceding results to prove that the equation L y = 0 has no non-zero solution in C( z).
The (ord0 , −1)-carrier of L consists of the points (0, −3), (2, −6) and (5, −5), and none of the
three relevant line segments is horizontal. We conclude from Corollary 5.4 that the corresponding
equation has no solution in C( z) having a pole at 0.
For any non-zero z0 ∈ C the (ordz0 , −1)-carrier of L consists of the points (0, 0), (2, −2) and
(5, −5), and we conclude from a second appeal to Corollary 5.4 that L y = 0 has no solution in C( z)
having a pole at z0 .
In particular, if L y = 0 has a solution r ∈ C( z) \ {0} the only pole of r must be at ∞. It follows that
r must be a polynomial (see, for example, Conway [7], p. 111), and it is a simple matter to check that
L y = 0 has no non-zero polynomial solutions.
Proposition 5.6. Assume Γ is an additive subgroup of R. Let τ ∈ Z, and let C nw denote the closed second
quadrant of R2 . Then for any non-zero r , s ∈ R and any i , j ∈ Z+ the (ν , τ )-carrier of the product

i
rxi · sx j = r ( S ik s)xk+ j (i)
k =0
is contained in the convex set (i + j , ν (r ) + ν (s) + (i + j )τ ) + C nw and includes the vertex point (i + j , ν (r ) +
ν (s) + (i + j )τ ) of this set.
We refer to (i + j , ν (r ) + ν (s) + (i + j )τ ) as the corner point of the indicated convex set.
Proof. From Lemma 4.2 we have

ν r ( S ik s) = ν (r ) + ν ( S ik s)
ν (r ) + ν (s) + (i − k)τ . (ii)
The carrier point corresponding to r ( S ik s)xk+ j is therefore

k + j , ν r ( S ik s) + (k + j )τ = k + j , ν (r ) + ν ( S ik s) + (k + j )τ ,
and from (ii) we see that
ν (r ) + ν ( S ik s) + (k + j )τ ν (r ) + ν (s) + (i − k)τ + (k + j )τ
= ν (r ) + ν (s) + (i + j )τ .
This places all carrier points within the indicated convex set. That the vertex is a carrier point is seen
from the observation that the k = i term in (i) is r σ i (s)xi + j (recall (2.3)), and from compatibility that

ν r σ i (s) = ν (r ) + ν σ i (s) = ν (r ) + ν (s). 2
The geometric ideas in this section will be extended significantly in Section 7, where Newton
polygons are introduced.
6. Weights
One verifies easily that the equation of a non-vertical line Λ ⊂ R2 passing through two distinct
integer points can be written uniquely in the form
ax + by = c , a, b, c ∈ Z, b > 0, (6.1)
provided we normalize so as to achieve
(i) b = 1 if a = 0,
(ii) gcd(|a|, b) = 1 if a = 0.
Note that the slope is −a/b. With any such Λ we associate the Z-linear weight function
w Λ : (x, y ) ∈ Z × Z → ax + by ∈ Z.
The value of w Λ (x, y ) is called the Λ-weight of the integer point (x, y ). When Λ is understood we
simply refer to the weight of the point. Note that all integer points on Λ have weight c, and that all
integer points above (resp. below) Λ have weight greater (resp. less) than c.
Proposition 6.1. Suppose w Λ is as above and (d1 , e 1 ), (d2 , e 2 ) ∈ Z × Z are two points with the same weight
satisfying d1 < d2 . Let

d2 − d1 if a = 0,
m := (i)
gcd(d2 − d1 , |e 2 − e 1 |) if a = 0.
Then

(a) ma = e 2 − e 1 and
(ii)
(b) mb = d2 − d1 .
In particular,
b|(d2 − d1 ). (iii)
Proof. By hypothesis the two integer points lie on a line parallel to Λ.

One has a = 0 if and only if Λ is horizontal, in which case m = d2 − d1 , e 1 = e 2 and b = 1.
Conditions (ii) and (iii) are therefore satisfied.
For a = 0 we see from
−a e2 − e1
gcd |a|, b = 1 and = (iv)
b d2 − d1
that (iii) holds, say m̂b = d2 − d1 for some positive integer m̂. A second appeal to (iv) then gives (ii),
although with m replaced by m̂. The asserted condition m̂ = m = gcd(d2 − d1 , |e 2 − e 1 |) then follows
from the first equality in (iv), and the result is thereby established. 2
For the remainder of the section we fix τ ∈ Z and assume that R admits a proper valuation ν : R → Z∪{∞}
which is τ -compatible with R [x; σ , δ]. In addition we assume Λ ⊂ R2 is a non-vertical line passing through
two distinct integer points, and we let w Λ denote the corresponding weight function.
The Λ-weight w Λ (rx j ) of any non-zero monomial rx j ∈ R [x; σ , δ] is defined to be the Λ-weight of
the integer point ( j , ν (r ) + j τ ), i.e.,

w Λ rx j := w Λ j , ν (r ) + j τ = aj + b ν (r ) + j τ .
In view of (4.10), an equivalent definition would be

w Λ rx j := w Λ j , ν̃ rx j .
n
The Λ-weight w Λ ( p ) of any non-zero element p := j =0 a j x
j
∈ R [x; σ , δ] is defined by

w Λ ( p ) := min w Λ a j x j : a j = 0 .
j
The following proposition summarizes the fundamental additive properties of weights.
Proposition 6.2. Let rxn , sxm ∈ R [x; σ , δ] be non-zero monomials. Then the following assertions hold:
(1) When Λ has positive slope the non-zero monomial of minimal Λ-weight within the product rxn · sxm is
rsσ xn+m , and one has
n

w Λ rxn · sxm = w Λ rxn + w Λ sxm .
(2) When n = m one has (regardless of the sign of the slope of Λ)

w Λ rxn + sxn min w Λ rxn , w Λ sxn .
Proof. (1) Note that the positive slope assumption on Λ implies a < 0. Among all terms of the product
rxn · sxm the monomial rsσ xn+m assumes the minimal ν̃ -value (by Lemma 4.4), and therefore the
n
minimal Λ-weight (because a < 0). The integer points corresponding to all other non-zero monomials
of the product lie above Λ, and as a result these monomials must have greater Λ-weight. This gives
the initial assertion, and the displayed equality is seen from

n
n m k+m
w Λ rx · sx = wΛ r ( S nk s)x
k =0

= min w Λ r ( S nk s)xk+m
k
n
= w Λ rsσ xn+m
n
= a(n + m) + b ν rsσ + (n + m)τ
n
= an + am + b ν (r ) + ν sσ + nτ + mτ

= an + am + b ν (r ) + ν (s) + nτ + mτ

= an + b ν (r ) + nτ + am + b ν (s) + mτ

= w Λ rxn + w Λ sxm .
(2) We have

w Λ rxn + sxn = w Λ (r + s)xn

an + b min ν (r ), ν (s) + bnτ

= min an + b ν (r ) + nτ , an + b ν (s) + nτ

= min w Λ rxn , w Λ sxn . 2
Under the slope assumption of Proposition 6.2(1) one can give a much stronger result.
Proposition 6.3. When Λ has positive slope the Λ-weight function w Λ , when extended to a mapping of
R [x; σ , δ] into Z ∪ {∞} by sending 0 to ∞, is a proper valuation on R [x; σ , δ].
The proof is practically identical to that of Proposition 4.5, and is therefore omitted.
7. Newton polygons
Newton polygons, as originally conceived by Newton in 1676, were associated with an algorithm
for solving polynomial equations p (x, y ) = 0 locally for y as a function of x. Nowadays, however, one
finds references to such polygons in seemingly diverse areas of mathematics, and not all formulations
of the concepts are equivalent. We offer a definition sufficiently general to cover all the planar cases
we have encountered (which the reader should not take to mean “all the planar cases ever studied”),
but quickly restrict attention to Newton polygons associated with polynomials and skew polynomials
over rings with proper valuations (introduced in Remark 3.3).
We need a few definitions from geometry (as a general reference see Berger [3]): the intersection
C ⊂ R2 of a finite collection of closed half-planes in R2 is called a convex polyhedron, a line forming
the boundary of any one of these half-planes and intersecting C in more than one point is a support
line of C , the line segments, half-rays and/or lines obtained by intersecting these support lines with
C are the edges of C , and the points of intersection of two distinct edges are the vertices of C .
An example of a convex polyhedron is provided by the convex hull of any non-empty finite subset
A ⊂ R2 (see, e.g., Proposition 12.1.15 of Berger [3]).
Assume ∅ = A ⊂ Z2 ⊂ R2 and C ⊂ R2 is a convex polyhedron having the origin 0 as a vertex. The
Newton polygon of the data ( A , C ) is defined to be the convex hull of the subset A + C ⊂ R2 . The set A
is the carrier of the this Newton polygon, and C is the associated convex polyhedron. (For some authors
“Newton polygon” refers to the boundary of what we call the Newton polygon.)
Note that when A = { zi } and C are as in the previous paragraph we have

A+C = ( zi + C ). (7.1)
i
The
Newton polygon of the data ( A , C ) can therefore be defined as the convex hull of the subset
i ( z i + C ) ⊂ R . When A is finite with very few elements this characterization often provides the
2
easiest means of visualization.

Four common choices for the associated convex set C are:
• C 0 := {0}, where 0 denotes the origin of R2 , e.g., see §8, Chapter II, pp. 36–38 in Arnol’d [1] or
Chapter 7, §1, p. 295 in Cox [8].
• C + := {(0, r2 ) ∈ R2 : r2 0}, i.e., the non-negative x2 -axis in R2 , e.g., see Ianni [12] and
Murty [20]. The boundary of the convex hull of A + C + is sometimes called the lower convex
envelope of A, e.g., see Ianni [12].
• C ne := {(r1 , r2 ) ∈ R2 : r1 0, r2 0}, i.e., the closed first quadrant of R2 , e.g., see Chapter 8, §3,
particularly p. 380 in Bourbaki [4].
• C nw := {(r1 , r2 ) ∈ R2 : r1 0, r2 0}, i.e., the closed second quadrant of R2 , e.g., see Mal-
grange [17].
The role of the associated convex set C is most easily seen when A ⊂ Z2 is finite and the projec-
tion of A onto the first coordinate is injective, as is the case in Definition 7.1 and in every example
appearing between that definition and the end of the paper. For those C listed above one can then
show, e.g., by induction on the cardinality of A, that the Newton polygon is a convex polyhedron, and
as a consequence one can make reference to support lines, edges and vertices thereof. (The proof of
the existence of the polyhedral structure is elementary, and is therefore omitted.)
• When C = C 0 only compact edges are involved and the Newton polygon is compact.
• When C = C + two infinite vertical rays (the left and right sides of the convex hull of A + C + )
are also involved, and all remaining edges are on the “bottom” of the convex hull (hence “convex
lower envelope”).
• When C = C ne two infinite rays are involved, one vertical and one horizontal, and no edge has
strictly positive slope.
• When C = C nw two infinite rays are involved, one vertical and one horizontal, and no edge has
strictly negative slope.
A remark on terminology. To geometers a polygon is a convex polyhedron with non-empty interior

(see, e.g., Definition 12.1.1 in [3]). Readers are warned that what we have called a Newton polygon
need not be polygon in this sense. For example, when A = {(a1 , a2 )} ⊂ Z2 is a singleton and C = C 0
the associated Newton polygon is A, which is not what a geometer would call a polygon. Moreover,
our “polygons” can involve intersections of infinitely many closed half-spaces (e.g., see Chapter 9,
pp. 115ff in Murty [20]). Last but not least, polygons in our sense can have infinitely many edges;
something geometers would not permit.
Newton polygons have proven to be an ingeniously simple geometric device for exhibiting prop-
erties essential to the understanding of various types of polynomials and/or associated polynomial
equations. In our case they enable one to visually determine, almost instantaneously, which weight
function is appropriate for studying the irreducibility of a given skew polynomial. Our results can
be established without Newton polygons, but at the significant cost of a geometric formulation of
the ideas. Indeed, in Section 8 we follow §24 in van der Waerden [29], where the concept is never
mentioned.
Definition 7.1. Suppose R is an integral domain and ν : R → Z ∪ {∞} is a proper valuation which is
τ -compatible with R [x; σ , δ].
(a) The Newton polygon of a non-zero polynomial p (x) ∈ R [x] is the Newton polygon of the data
(Carνp , C + ), where the carrier Carνp of p (x) = nj=0 a j x j is defined as {( j , ν (a j )) ∈ Z × Z: a j = 0,
0 j n}.
(b) The Newton polygon of a non-zero skew polynomial L (x) ∈ R [x; σ , δ] is the Newton polygon of the
ν ,τ
data (Car L , C nw ).
In both cases the carrier is finite, and projection onto the first factor is injective; we conclude from
our earlier discussion that these Newton polygons are convex polyhedra.
Definition 7.1(a) is perhaps the most common definition of a “Newton polygon,” e.g., see Ianni [12],
Chapter 4, §3, p. 97 in Murty [20], or Chapter 9, p. 113 in Koblitz [15]. It will not be used in the sequel,
but is included so that readers may easily compare our approach to such polygons with others they
may encounter. Definition (b) will prove important in our work; it is adapted from the unpublished
notes Malgrange [17], where it is described as a modification of an idea of Ramis [23]. The same
notes inspired the treatment of linear differential operators by means of Newton polygons found in
Chapter 3, §3 in van der Put and Singer [28] (see the comments in the second paragraph of p. 86 of
that reference). Also see Della Dora et al. [9] and Grigoriev [11]. The choice of C nw in (b), which we
have noted guarantees that no edge has strictly negative slope, is needed to invoke Proposition 6.2(a).
8. Applications to skew polynomial factorizations
In this section R [x; σ , δ] denotes a skew-polynomial ring in which the underlying ring R is a domain. We
fix an integer τ ∈ Z and a proper valuation ν : R → Z ∪ {∞} which is τ -compatible with R [x; σ , δ]. Finally,
n
we fix an element L := j =0 a j x ∈ R [x; σ , δ], with an = 0, and we refer to n =: deg( L ) as the degree of L
j
ν ,τ
(as per the discussion in the final paragraph of Examples 2.1(b)). Car L ⊂ Z2 is the (ν , τ )-carrier of L,
ν ,τ
and P := P L ⊂ R2 is the Newton polygon of L. Readers need to keep in mind that P is a convex polyhedron.
Theorem 8.1. Let Λ ⊂ R2 be a support line of P with positive slope and let w Λ : (x, y ) ∈ Z × Z → ax + by
(with b > 0) be the corresponding weight function. Write

Λ∩P = i , ν (ai ) + i τ , j , ν (a j ) + j τ , where i < j . (i)
Then the quotient
j−i
m := (ii)
b
is a positive integer. Moreover, if L factors in R [x; σ , δ] as M N, the degrees of M and N must be at least m1 b
and m2 b respectively, where m1 and m2 are non-negative integers satisfying m1 + m2 = m.
Proof. That m ∈ Z+ is immediate from (ii) of Proposition 6.1.

Let c := w Λ (ai xi ) = w Λ (a j x j ) and note that from the support line assumption on Λ that

w Λ ah xh c whenever ah = 0, (iii)
i.e., c is the minimum of the Λ-weights of the monomials of L. Now observe that ai xi and a j x j are
the monomials of Λ-weight c of minimum and maximum orders respectively.
For the remainder of the proof we assume L factors as M N. Let akmin xkmin and akmax xkmax be the
(possibly non-distinct) monomials of M of least weight having minimal and maximal orders respec-
tively, and let almin xlmin and almax xlmax be the corresponding monomials of N. Then with two further
appeals to Proposition 6.1 we can find integers m1 , m2 ∈ N such that
bm1 = kmax − kmin , bm2 = lmax − lmin .
All we need do is verify that m1 and m2 have the required properties.

Set ck := w Λ (akmin xkmin ) = w Λ (akmax xkmax ) and cl := w Λ (almin xlmin ) = w Λ (almax xlmax ). From the min-
imal properties of c , ck and cl , the positive slope assumption and Proposition 6.2(a) we see that
c = ck + cl , and that terms of M and N of other (necessarily higher) Λ-weights cannot produce a
term of weight c. It is then immediate from the choices of kmin , lmin , kmax and lmax that
i = kmin + lmin and j = kmax + lmax ,
hence
bm = j − i = (kmax − kmin ) + (lmax − lmin ) = bm1 + bm2 ,
and m = m1 + m2 follows. To complete the proof simply note that
deg( M ) kmax kmax − kmin = bm1
and
deg( N ) lmax lmax − lmin = bm2 . 2
Corollary 8.2. Under the hypotheses of Theorem 8.1 the following assertions hold:
(a) at least one of M and N has degree at least b;

(b) if b = n − 1 then one of M and N must have the form ax + b; and
(c) when a0 = 0, and when w Λ (a0 ) = w Λ (an xn ) w Λ (a j x j ) for all 1 j < n such that a j = 0, the degrees
of both M and N must be divisible by b.
Proof. (a) Immediate from Theorem 8.1.

(b) By (a).
(c) In this case i = 0 and j = n, and from Theorem 8.1 we conclude that n = bm = bm1 + bm2
deg( M ) + deg( N ) = n. The equalities bm1 = deg( M ) and bm2 = deg( N ) follow. 2
Corollary 8.3. Assume the hypotheses and notations of Theorem 8.1 and, in addition, that the “constant term”
a0 of L is not zero. Then any of the following conditions guarantees the irreducibility of L over R [x; σ , δ]:
(a) b n;
(b) ν (an ) + nτ = ν (a0 ) + 1 and ν (a0 ) < ν (ak ) + kτ for all indices k satisfying ak = 0; and
(c) the integer m := ν (an ) + nτ − ν (a0 ) ∈ Z is positive, relatively prime to n, and ν (ak ) + kτ ν (a0 ) + m
holds for all indices k satisfying ak = 0.
Assertion (b) is one of our two candidates for an “Eisenstein irreducibility criterion” for skew
polynomials; the other is seen in Theorem 9.7.
Proof. (a): By Corollary 8.2(a).

(b) and (c): It suffices to prove (c) since (b) is the special case m = 1.
Let E denote the closed line segment [(0, ν (a0 )), (n, ν (a0 ) + k)] ⊂ R2 . Under the stated hypotheses
the Newton polygon of L is the union of all translates C nw + x with x ∈ E. The equation of the support
line Λ of the lower right edge E is −kx1 + nx2 = nν (a0 ), and from the hypothesis gcd(k, n) = 1 we
conclude that a = −k and b = n. The result now follows from (a). (The condition k > 0 ensures that Λ
has positive slope, as required in the hypothesis of Theorem 8.1.) 2
Example 8.4. In all the examples we take R = M(P1 ) C( z) and τ = −1.

3
d d d
(a) For 2 k ∈ Z and L = L k = 3 + zk dz + z4 dz 3 ∈ C( z )[ dz ] we have

4 d3
ν̃ z = ord0 z4 + 3 · (−1) = 4 − 3 = 1 = ord0 (a0 ) + 1 and
dz3
ord0 (a0 ) = 0 < ord0 (a1 ) + 1 · (−1) = k − 1.
d
It follows from Corollary 8.3(b) that L k is irreducible in C( z)[ dz ].
2
(b) The operator L = z13 + dz
d d
2 ∈ C( z )[ dz ] is irreducible: using ν = ord0 one computes that b = 2, and
Corollary 8.3(a) therefore applies.
2
d d
(c) The operator L = 1 + z dz 2 ∈ C( z )[ dz ] is irreducible: the argument of (b) also applies here.
2
d d
(d) We claim that for any positive odd integer m the operator L m := zm + dz 2 ∈ C( z )[ dz ] is irreducible.

(For m = 1 the operator corresponds to Airy’s equation y + zy = 0.) Indeed if the operator L
d
factors in C( z)[ dz ] then the form
d2 d 1
L ∞ := t 4 + 2t 3 + (i)
dt 2 dt tm
of L at ∞ factors in C(t ) when one switches from the local coordinate z near 0 to the local
coordinate t := 1/ z near ∞ ∈ P1 . However, for (i) we have
ord0 (a2 ) + 2 · (−1) = 4 + 2 · (−1) = 2
ord0 (a1 ) + 1 · (−1) = 3 + 1 · (−1) = 2, and
ord0 (a0 ) + 0 · (−1) = −m,
and the irreducibility of L ∞ is then seen from Corollary 8.3(c).
For additional applications of Newton polygons to linear differential operators see Chapter 3, §3,
pp. 86–98 in van der Put and Singer [28].
9. A second generalization of the Eisenstein criterion
In Kovacic [16] one finds an analogue of the Eisenstein irreducibility criterion for skew polynomials
in R [x; σ , δ] under the assumption σ δ = δ σ (for a brief description see van der Put and Singer [28,
Chapter 4, §2, p. 116]). Here we generalize Kovacic’s result; in particular, we drop the commutativity
assumption on the two operators.
Throughout this section K is a (commutative) field, Γ is an ordered group (written additively, although
not necessarily commutative), and ν : K → Γ ∪ {∞} is a proper valuation with associated valuation ring
R := {k ∈ K | ν (k) 0}. σ is an automorphism of K which restricts to an automorphism of R, and δ is σ -
derivation on K which restricts to a σ | R -derivation on R. When confusion cannot otherwise result we write
the restrictions σ | R and δ| R as σ and δ respectively. We view the skew-polynomial ring K [x; σ , δ] as an
extension of the skew-polynomial ring R [x; σ , δ]. We assume that for some τ ∈ Γ , ν is τ -compatible with
K [x; σ , δ]; as a consequence ν | R must be τ -compatible with R [x; σ , δ].
Recall that the τ -compatibility assumption involves two conditions:

ν σ (k) = ν (k) and ν δ(k) ν (k) + τ for all k ∈ K . (9.1)
Given the abundance of standing hypotheses it seems appropriate to begin with a somewhat de-
tailed example.
Example 9.1. Let A be an integral domain with quotient field K and let p ⊂ A be a prime ideal
such that the localization R := A p ⊂ K is a principal ideal domain. Then a proper surjective (discrete)
valuation ν : K → Z ∪ {∞} can be defined in the usual way: if π is a prime generating the unique
maximal ideal m ⊂ R then any non-zero a/b ∈ K can be written uniquely in the form a/b = π n u,
where n is an integer and u ∈ R is a unit, and we can therefore define ν (a/b) := n; ν (0) = ∞ (see,
e.g., Matsumura [19, Theorem 11.2, p. 79]). The definition is independent of the choice of π , R is the
associated (discrete) valuation ring, and m = {k ∈ K : ν (k) > 0}.
Note that we regard R as a subring of K . Moreover, we regard K as the quotient field of R.
Continuing with the example, assume σ : A → A is a ring automorphism, that δ : A → A is a σ -
derivation, and that p (as above) is both σ and δ -invariant. Then σ and δ are easily seen to have
unique extensions to both R and K which we again denote by σ and δ . With these conventions
in mind it now makes sense, for appropriate x, to view the skew-polynomial ring K [x; σ , δ] as an
extension of the skew-polynomial ring R [x; σ , δ].
We claim that
σ (π ) = π u , (i)
where u is a unit of R. Indeed, σ (π ) must also be a prime generator of m, and since this is the
unique prime ideal of R the two prime elements must be associates.
It follows from (i) that ν : K → Z ∪ {∞} is compatible with σ : K → K . To see this we again appeal
to the fact that any non-zero element k ∈ K can be uniquely written in the form k = π n u ∈ K with
n ∈ Z and u ∈ R a unit. We then have
n
ν σ (k) = ν σ π n u = ν σ (π ) σ (u )

= ν (π u 1 )n σ (u ) = ν π n un1 σ (u ) = n = ν π n u = ν (k),
where u 1 is also unit of R, and compatibility is thereby established.

Finally, we claim that ν : K → Z ∪ {∞} is 0-compatible with δ : K → K . To prove this we once
again write the typical non-zero k ∈ K in the now-familiar form π n u. We then have

δ(k) = δ π n u

= δ π n σ (u ) + π n δ(u )
= nπ n−1 δ(π )σ (u ) + π n δ(u )

= π n−1 nδ(π )σ (u ) + π δ(u ) . (ii)
Since δ(π ) ∈ m we know that δ(π ) is not a unit, and we can therefore write δ(π ) = π m u 1 with
m ∈ Z+ and u 1 ∈ R a unit. Substituting δ(π ) = π m u 1 into (ii) then gives

δ(k) = δ π n u = π n nπ m−1 u 1 σ (u ) + δ(u ) . (iii)
Note that nπ m−1 u 1 σ (u ) + δ(u ) ∈ R, and we therefore have ν (nπ m−1 u 1 σ (u ) + δ(u )) 0. It is then
immediate from (iii) (and the multiplicative property of ν ) that

ν δ(k) = ν π n nπ m−1 u 1 σ (u ) + δ(u )

= ν π n + ν nπ m−1 u 1 σ (u ) + δ(u )

ν πn
=n

= ν π nu
= ν (k),
and our final claim is established.
The ends the example; we now return to generalities.

Note from Proposition 4.5 that we can extend ν to a proper valuation ν̃ of K [x; σ , δ] by defining

ν̃ ki xi := min ν (ki ) + i τ . (9.2)
i
Definition 9.2. A non-zero skew polynomial p (x) ∈ K [x; σ , δ] is ν -primitive, or is primitive with respect
to ν , when ν̃ ( p (x)) = 0.
Example 9.3. Assume the context of Example 2.1(a) with R = Z and let ν = ν3 be the 3-adic valuation
on Z. Let ν̃3 be the extended valuation on Z[x] defined by (9.2).
(a) For the polynomial p (x) := 6x2 + 2x + 5 ∈ Z[x] we have ν̃3 ( p (x)) = 0, and the polynomial is
therefore ν3 -primitive. Of course p (x) is also primitive in the usual sense, i.e., the gcd of the
coefficients is 1.
(b) Let q(x) := 6x2 + 2x + 2 ∈ Z[x]. Then q(x) is ν3 -primitive but not primitive.
Remark 9.4. In Kovacic [16] a skew polynomial is called primitive if the ideal generated by the coeffi-
n
cients is the entire ring. That condition is implied by our definition: if ν̃ ( i =0 ai xi ) = 0 then ν (a j ) = 0
for some index j, and a j is then a unit (of the associated valuation ring) by Proposition 3.4(b). The
ideal of the associated valuation ring generated by all coefficients is therefore the entire ring.
It is well known that a straightforward generalization of the Gauss lemma fails in skew polynomial
rings (see, e.g., Kovacic [16]). Our reformulation of “primitive” circumvents that problem.
Lemma 9.5 (The Gauss valuation lemma). The product of two ν -primitive skew polynomials is again ν -
primitive.
Proof. From property (V.3) of a valuation we have ν̃ ( p (x)q(x)) = ν̃ ( p (x)) + ν̃ (q(x)) for any skew-
polynomials p (x) and q(x). It follows that ν̃ ( p (x)q(x)) = 0 when ν̃ ( p (x)) = ν̃ (q(x)) = 0. 2
Proposition 9.6. Suppose ν is surjective and 0-compatible with K [x; σ , δ], and f ∈ R [x; σ , δ] is reducible
in K [x; σ , δ], i.e., f = gh with g , h ∈ K [x; σ , δ]. Then there exist g 1 , h1 ∈ R [x; σ , δ] satisfying f = g 1 h1 ,
deg( g 1 ) = deg( g ) and deg(h1 ) = deg(h). Moreover, if f is ν -primitive the same is true of g 1 and h1 .
Proof. We will attempt a proof assuming τ -compatibility for arbitrary τ , and in the process reveal
our reasons for restricting to the case τ = 0.
First note that since ν is surjective one can find a non-zero k ∈ K be such that ν (k) = ν̃ ( g ). Let
g 1 = gk−1 , h1 = kh. Clearly, deg( g 1 ) = deg( g ), deg(h1 ) = deg(h), and f = g 1 h1 . Furthermore

(a) ν̃ ( g1 ) = ν̃ ( g ) + ν k−1 = ν̃ ( g ) − ν (k) = 0,
(i)
(b) ν̃ (h1 ) = ν (k) + ν̃ (h) = ν̃ ( g ) + ν̃ (h) = ν̃ ( f ).

Write g 1 = a i xi , a i ∈ K , h 1 = b j x j , b j ∈ K and f = cl xl , cl ∈ K . To prove g 1 , h1 ∈ R [x; σ , δ]
we have to show that ai , b j ∈ R for all i , j, i.e., that ν (ai ) 0 and ν (b j ) 0.
From (i)(a) we see that

0 = ν̃ ( g 1 ) = min ν (ai ) + i τ = ν (ai0 ) + i 0 τ for some index i 0 . (ii)
i
Case 1. Suppose τ > 0. Then ν (ai 0 ) = −i 0 τ < 0 implies ai 0 ∈ / R, and we would therefore have
g1 ∈
/ R [x; σ , δ]. This explains the need for restricting τ to non-positive values.
Now observe from (i)(b) that
ν̃ (h1 ) = ν (b j0 ) + j 0 τ = ν (c 0 ) + 0 τ = ν̃ ( f ) for some indices j 0 and 0 . (iii)
Case 2. Suppose τ < 0. Then (iii) is not sufficient to ensure h1 ∈ R [x; σ , δ]. Indeed, since j 0 and
0 may differ one cannot deduce ν (b j 0 ) 0 from ν (b j 0 ) + j 0 τ = ν (c 0 ) + 0 τ 0. We are thereby
reduced to the stated hypothesis τ = 0.
Assuming τ = 0, note from the first and second equalities of (ii) that ν̃ ( g 1 ) = mini {ν (ai )} = 0, and
we therefore have ν (ai ) 0 for all i. This gives ai ∈ R for all i, and g 1 ∈ R [x; σ , δ] follows.
Continuing under the assumption that τ = 0, note from the hypothesis f ∈ R [x; σ , δ] that ν (c ) 0
for all , and we therefore have ν̃ ( f ) = min {ν (c )} 0. Since ν̃ (h1 ) = ν̃ ( f ) (by (i)(b)), and since for
τ = 0 we have ν̃ (h1 ) = min j {ν (b j )}, it follows that ν (b j ) 0 for all j, hence that h1 ∈ R [x; σ , δ].
To complete the proof note from (i) that ν̃ ( g 1 ) and ν̃ (h1 ) must vanish if this is the case
for ν̃ ( f ). 2
We are now in a position to give the Eisenstein valuation criterion for skew polynomials.
Theorem 9.7 (The Eisenstein valuation criterion). Assume ν is both surjective and 0-compatible with
K [x; σ , δ], and set m := {r ∈ R | ν (r ) > 0}. Let f = cl xl + cl−1 xl−1 + · · · + c 0 ∈ R [x; σ , δ] be non-constant
and satisfy c 0 ∈
/ m, c i ∈ m for i = 1, 2, . . . , l, and cl ∈
/ m2 . Then f is irreducible in R [x; σ , δ].
Note from Proposition 3.4(c) that m is the unique maximal left-and-right ideal of the valuation
ring R.
Proof. We claim that m is invariant under δ , under σ , and under σ −1 . For δ this is immediate
from (4.2) (with τ = 0), and for σ from (4.3). As for σ −1 , choose any r ∈ m and set a := σ −1 (r ).
Then from σ (a) = r and (4.3) we see that ν (σ −1 (r )) = ν (a) = ν (σ (a)) = ν (r ), and the claim follows.
Suppose f is reducible min K [x; σ , δ] . nBy Proposition 9.6 we may then assume f = gh, where
g , h ∈ R [x; σ , δ]. Let g = i =0 ai xi , h = j =0 b j x with m > 0, n > 0 and am bn = 0. Then cl = cm+n =
j
am σ m (bn ) = 0. Since m is prime and c ∈ m \ m2 we have either am ∈ m and σ m (bn ) ∈/ m or am ∈/ m

and σ m (bn ) ∈ m.
Case 1. am ∈ m and σ m (bn ) ∈/ m.

m i
From Eq. (2.7) and the given hypotheses we have c 0 = i =0 ai bδ0 ∈ / m. It follows that there is a
(necessarily unique) index i 0 satisfying 0 i 0 m, ai 0 ∈
/ m, and ai 0 +1 , . . . , am ∈ m. Since n > 0 we
have n + i 0 > 0, and cn+i 0 ∈ m is then seen from the given hypotheses. On the other hand, by means
of a second appeal to Eq. (2.7) we can write

m
c n +i 0 = ai S ik (b j ). (i)
i =0 j +k=n+i 0
Since n + i 0 = j + k j + i n + i we have i i 0 , and the double sum therefore reduces to

m
c n +i 0 = ai S ik (b j )
i =i 0 j +k=n+i 0

m
= ai 0 S i 0 k (b j ) + ai S ik (b j ).
j +k=n+i 0 i =i 0 +1 j +k=n+i 0

Since j n, and since S ik = 0 unless 0 i k, we have j +k=n+i 0 S i 0 k (b j ) = S i 0 i 0 (bn ), which by the
final equality in (2.3) is equal to σ i0 (bn ), and we therefore have

m
cn+i 0 = ai 0 · σ i 0 (bn ) + ai S ik (b j ).
i =i 0 +1 j +k=n+i 0
Because all ai appearing in the final double summation are elements of m the same holds for the
value of that double sum, and ai 0 σ i 0 (bn ) ∈ m then follows from cn+i 0 ∈ m. Since ai 0 ∈ / m this forces
σ i0 (bn ) ∈ m, hence σ m (bn ) = σ m−i0 (σ i0 (bn )) ∈ m by σ -invariance, and we have a contradiction.
Case 2. am ∈
/ m and σ m (bn ) ∈ m.
In this instance the argument becomes more transparent by abandoning the S i j -notation.
By assumption there is a (necessarily unique) index j 0 satisfying 0 j 0 n, σ m (b j 0 ) ∈ / m, and
σ m (b j0 +1 ), . . . , σ m (bn ) ∈ m. The trick is to note, from the paragraph surrounding (2.3), that the ana-
logue of equality (i) of the previous case can be expressed as

m
i i −k
c m+ j 0 = ai σ k bδj .
k
i =0 j +k=m+ j 0
Since m + j 0 = j + k j + i j + m we have j j 0 , with equality holding throughout if and only if

j = j 0 , in which case k = i = m. Thus

m −1
m i i −k
c m+ j 0 = am σ k (b j ) + ai σ k bδj
k k
j +k=m+ j 0 i =0 j +k=m−1+ j 0 , j > j 0

m
= am σ m (b j 0 ) + am σ k (b j )
k
j +k=m+ j 0 +1, j > j 0

m −1
i i −k
+ ai σ k bδj .
k
i =0 j +k=m−1+ j 0 , j > j 0
From the invariance of m under δ, σ and σ −1 , and from the choice of j 0 , all terms σ k (b j ) and
i −k
σk
(bδj ) appearing within summands following the final equality sign are in m, whereas σ m (b j 0 ) ∈
/ m.
Moreover, since m > 0, cm+ j 0 ∈ m by hypothesis, hence am σ (b j 0 ) ∈ m. Because (in this case) am ∈
m
/ m,
the maximality of m forces σ m (b j 0 ) ∈ m, and this contradicts the choice of j 0 . 2
Corollary 9.8 (The classical Eisenstein criterion). Suppose f (x) = an xn +an−1 xn−1 + · · · + a0 ∈ Z[x] satisfies
the following condition: there is a prime p such that p |a0 , p |a j for j = 1, . . . , n, and p 2 |an . Then f is
irreducible in Z[x].
Proof. In the notation of Theorem 9.7 take K = Q, let ν be the p-adic valuation, let σ = idQ and let δ
be the trivial derivation. Then R = { ba ∈ Q | gcd(a, b) = 1, p b} and m = { ba ∈ R | gcd(a, b) = 1, p | a}.
The stated hypotheses imply a0 ∈ / m, an , . . . , a1 ∈ m, and an ∈
/ m2 , and from the theorem we conclude
that f (x) is irreducible in R [x]. Since R [x] contains Z[x], the result follows. 2
In the statement and proof of the following result we drop the standing hypotheses for the section.
In particular, we do not (initially) assume there is a valuation on the field K .
Corollary 9.9. (See Kovacic [16].) Let A be an integral domain with quotient field K and let p ⊂ A be a prime
ideal such that the localization A p ⊂ K is a principal ideal domain. Assume σ : A → A is a ring embedding,
that δ : A → A is a σ -derivation, and that p is both σ and δ -invariant. Let the extensions of σ and δ to both
l
A p and K again be denoted by σ and δ . Finally, suppose L = i =0 c i xi ∈ A [x; σ , δ] is such that c i ∈ p for
i = 1, . . . , l, c 0 ∈
/ p and cl ∈
/ p2 . Then L is irreducible over K .
Proof. Define a surjective valuation ν : K → Z ∪ {∞} as in Example 9.1, and note from that example
that ν is 0-compatible with K [x; σ , δ]. Let m ⊂ R := A p be as in that example, and note that p ⊂ m.
The hypotheses therefore ensure that one has c i ∈ m for i = 1, . . . , l, c 0 ∈
/ m and cl ∈
/ m2 , whereupon
irreducibility is immediate from Theorem 9.7. 2
Although we have assumed 0-compatibility in Theorem 9.7, the result also holds when ν is τ -
compatible with R [x; σ , δ] and τ 0. Simply observe that when τ 0 the τ -compatibility condition
in (9.1) implies the same condition with τ replaced by 0. It then follows from Proposition 4.5 that
when τ > 0 there are at least two ways to extend the valuation ν to K [x; σ , δ]:
1. τ -compatibility gives:

ν̃τ ki xi := min ν (ki ) + i τ . (9.7)
i
2. 0-compatibility gives:

ν̃0 ki xi := min ν (ki ) . (9.8)
i
In particular, when τ > 0 one can replace τ by 0 and then use Theorem 9.7 for verifying the irre-
ducibility criterion. From Eqs. (9.7) and (9.8) we see that the valuation extension of 0-compatibility is
simpler to deal with: it is easier to calculate the valuation ring and maximal ideal.
Examples 9.10. (Applications to differential operators over C( z).) Recall from Example 3.1(b) that
when z0 = ∞ the Laurent expansion of f ∈ C( z) refers to the Laurent expansion of f (1/t ) at t = 0.
A proper discrete valuation ν∞ on M(P1 ) is defined by ν∞ ( f ) := ord∞ ( f ), but actual values can be
computed in a more direct manner:
f = p /q = p ( z)/q( z) ∈ C( z) ⇒ ν∞ ( f ) = ord∞ ( f ) = deg(q) − deg( p ).
The corresponding valuation ring is given by

p
R ∞ := deg(q) deg( p ) , (i)
q
and the maximal ideal by

p
m := deg(q) > deg( p ) . (ii)
q
1
Note that m is generated by z
.
d
(a) Many differential operators not contained in R ∞ [ dz ] might still be amenable to analysis by means
of the ord∞ -valuation, often simply upon multiplication by an appropriate power of z. For exam-
ple, one has
!
d3 d d
L := z2 + 2z + z3 ∈
/ R∞ ,
dz3 dz dz
whereas
!
1 1 d3 2 d d
L= + + 1 ∈ R∞ .
z3 z dz3 z2 dz dz
1 2 1
For the operator z3
L we have c 0 := 1 ∈
/ m, c 1 := z2
∈ m, c 2 := z
∈ m, and c 2 ∈
/ m2 , whereupon
d
irreducibility over R ∞ [ dz ] is immediate from Theorem 9.7. By Proposition 9.6 the same result
follows over C( z)[ dz ] for z13 L, and it is
d
then an easy matter to see that L must be irreducible
d
over C( z)[ dz ].
(b) It is interesting to note that one can exploit the exclusion 1z ∈
/ m2 to construct an irreducible dif-
ferential operator of any prescribed degree. Specifically, for any positive integer n the differential
operator
1 dn dn−1 d
+ an−1 ( z) + · · · + a1 ( z ) +1
z dzn dzn−1 dz
must be irreducible over C( z) if one has an−1 ( z), . . . , a1 ( z) ∈ m, and one can choose such coeffi-
cients in many ways.
In both examples the key step for applying Theorem 9.7 is transforming in such a way that the
resulting leading coefficient does not belong to m.
Acknowledgments
The authors would like to thank the referees for their careful reading of the original manuscript,
corrections, suggestions for clarification, simplifications of several arguments, and additional refer-
ences.
References
[1] V.I. Arnol’d, Huygens and Barrow, Newton and Hooke, Birkhäuser-Verlag, Basel, 1990.
[2] E. Artin, Geometric Algebra, Interscience Publishers, Inc., 1957.
[3] M. Berger, Geometry, II, Springer-Verlag, Berlin, 1987.
[4] N. Bourbaki, Commutative Algebra, Chapters 1–7, Elem. Math. (Berlin), Springer-Verlag, Berlin, 1987.
[5] P.M. Cohn, Free Rings and Their Relations, Academic Press, London, 1985.
[6] P.M. Cohn, Skew Fields: Theory of General Division Rings, Encyclopedia Math. Appl., vol. 57, Cambridge Univ. Press, Cam-
bridge, 1995.
[7] J.B. Conway, Functions of One Complex Variable, second ed., Grad. Texts in Math., vol. 11, Springer-Verlag, 1978.
[8] D. Cox, J. Little, D. O’Shea, Using Algebraic Geometry, Grad. Texts in Math., vol. 185, Springer-Verlag, 1998.
[9] J. Della Dora, C. di Crescenzo, E. Tournier, An algorithm to obtain formal solutions of a linear homogeneous differential
equation at an irregular point, in: Lecture Notes in Comput. Sci., vol. 144, Springer-Verlag, 1982, pp. 273–280.
[10] G. Dumas, Sur quelques cas d’irreductibilite des polynomes á coefficients rationnels, J. Math. Pures Appl. (6) 2 (1906)
191–258.
[11] D. Grigoriev, Complexity of factoring and calculating the GCD of linear ordinary differential operators, J. Symbolic Com-
put. 7 (1990) 7–37.
[12] J.G. Ianni, Computing normalizations using Newton polygons, PhD thesis, Graduate Center, CUNY, 2001.
[13] N. Jacobson, The Theory of Rings, Amer. Math. Soc., New York, 1943.
[14] N. Jacobson, Finite-Dimensional Division Algebras over Fields, Springer-Verlag, 1996.
[15] N. Koblitz, p-Adic Numbers, p-Adic Analysis, and Zeta Functions, second ed., Grad. Texts in Math., vol. 58, Springer-Verlag,
New York, 1984.
[16] J. Kovacic, An Eisenstein criterion for noncommutative polynomials, Proc. Amer. Math. Soc. 34 (1) (1972) 25–29.
[17] B. Malgrange, Sur la réduction formelle des équations différentialles à singularités irrégulieres, Grenoble, preprint, 1979.
[18] K. Mathiak, Valuations of Skew Fields and Projective Hjelmslev Spaces, Lecture Notes in Math., vol. 1175, Springer-Verlag,
1986.
[19] H. Matsumura, Commutative Ring Theory, University Press, Cambridge, 1980.
[20] R.M. Murty, Introduction to p-Adic Analytic Number Theory, Stud. Adv. Math., vol. 27, AMS and International Press, Provi-
dence, RI, 2002.
[21] O. Ore, Formale Theorie der linearen Differentialgleichungen, J. Reine Angew. Math. 168 (1932) 233–252.
[22] O. Ore, Theory of non-commutative polynomials, Ann. of Math. 34 (22) (1933) 480–508.
[23] J.P. Ramis, Dévisage gevrey, Astérisque 59/60 (1978) 173–204.
[24] O. Schilling, The Theory of Valuations, Math. Surveys, vol. IV, Amer. Math. Soc., New York, 1950.
[25] M.F. Singer, Testing reducibility of linear differential operators: A group theoretic perspective, Appl. Algebra Engrg. Comm.
Comput. 7 (2) (1996) 77–104.
[26] F. Ulmer, J.A. Weil, Note on Kovacic’s algorithm, J. Symbolic Comput. 22 (1996) 179–200.
[27] M. van der Put, Reduction modulo p of differential equations, Indag. Math. (N.S.) 7 (3) (1996) 367–387.
[28] M. van der Put, M.F. Singer, Galois Theory of Linear Differential Equations, Springer-Verlag, Berlin, 2003.
[29] B.L. van der Waerden, Modern Algebra, vol. I, Ungar, New York, 1953.
[30] M. van Hoeij, Factorization of differential operators with rational function coefficients, J. Symbolic Comput. 24 (1997) 537–
561.
Journal of Algebra
Representation type of the blocks of category O S in types F 4

and G 2
Kenyon J. Platt 1
Department of Mathematics, Brigham Young University, Provo, UT 84602-6539, United States
Article history: In this paper a complete classification of the representation type

Received 17 September 2007 of the infinitesimal blocks of the parabolic category O S for the
Available online 15 May 2009 complex simple Lie algebras of types F 4 and G 2 is given.
Communicated by Jon Carlson
© 2009 Elsevier Inc. All rights reserved.
Keywords:
Category O
Representation type
Verma modules
Exceptional Lie algebras
1. Introduction
Indecomposable modules of a finite-dimensional algebra provide a complete description of all the

modules of the algebra. Consequently, classifying the indecomposable modules for a fixed finite-
dimensional algebra is a central theme in the representation theory of such algebras. One of the
first questions one can ask is, “How ‘classifiable’ are the indecomposable modules of a certain finite-
dimensional algebra?” A finite-dimensional algebra will fall into one of three classes depending on the
‘classifiability’ of its indecomposable modules. If there are only finitely many isomorphism classes of
indecomposable modules, then we say the algebra has finite representation type. If there are infinitely
many isomorphism classes of indecomposable modules, then we say the algebra has infinite represen-
tation type. If the algebra has infinite representation type, it can be further classified as having tame
representation type if, roughly speaking, these indecomposable modules can be parameterized in some
way, and wild representation type otherwise (see [Dro,CB]).
Cline, Parshall, and Scott [CPS, Thm. 3.6] proved that a highest weight category with finitely many
simple objects is equivalent to the category of finitely generated modules of some (finite-dimensional)
quasi-hereditary algebra. Projective modules of quasi-hereditary algebras admit filtrations by certain
E-mail address: platt@math.byu.edu.

1
Research partially supported by an NSF VIGRE fellowship at the University of Georgia.
doi:10.1016/j.jalgebra.2009.04.040
3824 K.J. Platt / Journal of Algebra 322 (2009) 3823–3838
Fig. 1. Dynkin diagrams for F 4 (left) and G 2 (right).
standard modules. Consequently, at times it is possible to directly deduce the structures of the pro-
jective modules. Using this information, one can express the algebra as a quiver with relations from
which one can potentially determine the representation type of the algebra.
If S is a subset of simple roots of a root system of a complex simple Lie algebra g relative to a fixed
maximal toral subalgebra, then category O S is a generalization of the ordinary Berstein–Gelfand–
Gelfand category O . These are highest weight categories having a decomposition into infinitesimal
blocks, with each block equivalent to the module category of some finite-dimensional quasi-hereditary
algebra. One is justified in talking about the representation type of an infinitesimal block of O S be-
cause of this underlying algebra.
μ
If μ is an integral weight, then the infinitesimal block O S contains all the simple modules with
highest weight linked to μ via the dot action of the Weyl group. The singular root system of μ has
simple roots J contained in the set of simple roots of the root system of g. Hence, the (singular)
integral infinitesimal blocks of category O S are determined by subsets J of the simple roots.
The classification of the representation type of the infinitesimal blocks of category O S began
with the independent work of Futorny, Nakano, and Pollack [FNP] and Brüstle, König, and Ma-
zorchuk [BKM] in classifying the representation type of the blocks of ordinary category O (where
S = ∅). Boe and Nakano [BN] later classified the representation type of all infinitesimal blocks of O S
with S ∩ J = ∅.
In this paper, the representation type of all integral infinitesimal blocks of category O S for any S
will be determined in the case when g is of type F 4 or G 2 . A computer was employed to determine
the representation type of the infinitesimal blocks. After tabulating the results, one can observe a
natural partition of the set of sixteen (resp., four) subroot systems of F 4 (resp., G 2 ) into nine (resp., 3)
equivalence classes. Furthermore, for both root systems, one can define a partial ordering on these
equivalence classes which encapsulates the classification of the representation type of the infinitesi-
mal blocks (see Theorems 5 and 6).
The paper is organized as follows. The necessary preliminaries and machinery are developed in
Section 2. Then in Section 4 some criteria will be developed for determining the representation type
of a given infinitesimal block. Section 5 is devoted to determining the representation type of all
infinitesimal blocks of category O S when g is of type F 4 or G 2 .
2. Preliminaries
2.1. Notation
Let g be a simple Lie algebra over the field C of complex numbers with root system Φ of type
F 4 or G 2 , determined by a maximal toral subalgebra h of g. If n ∈ {2, 4} is the rank of Φ , then let
= {α1 , . . . , αn } be a set of simple roots for Φ , ordered as given in the Dynkin diagrams for F 4 and
G 2 in Fig. 1. Let Φ + and Φ − be (respectively) the set of positive roots and the set of negative roots
with respect to .
For each α ∈ Φ , let gα be the α root space. If n+ (resp. n− ) is the sum of the positive (resp.
negative) root spaces, then we have the Cartan decomposition g = n− ⊕ h ⊕ n+ . Let b = h ⊕ n+ , the
Borel subalgebra defined by h and .
Denote the standard inner product on h∗ by (−, −). For each α ∈ Φ ⊆ h∗ , let sα denote the reflec-
tion corresponding to α . For a simple root αi ∈ , we will write the corresponding simple reflection
as si := sαi . Let W denote the Weyl group generated by the simple reflections. The length of w ∈ W
will be denoted l( w ). The relation w 1 w 2 denotes the Bruhat ordering on W . The Weyl group acts
on h∗ via the dot action:
w · μ = w (μ + ρ ) − ρ for all w ∈ W , μ ∈ h∗ ,
where ρ is the half sum of the positive roots.

K.J. Platt / Journal of Algebra 322 (2009) 3823–3838 3825
Write Z for the integers and Z0 for the non-negative integers. Denote the coroot of α ∈ Φ by
α̌ := (α2,αα ) . Let X = {μ ∈ h∗ | (μ, α̌ ) ∈ Z for all α ∈ Φ} denote the integral weight lattice, and let X + =
{μ ∈ X | (μ, α̌ ) ∈ Z0 for all α ∈ Φ + } denote the set of dominant integral weights.
2.2. The category O S
Fix S ⊆ , viewed where appropriate as a subset of {1, . . . , n} via the fixed ordering on simple
roots. Then S determines a standard parabolic subalgebra p S = m S ⊕ u+ S ⊇ b of g (see [RC]). The Lie
subalgebras m S and u+
S are called, respectively, the Levi factor and the nilradical. The subset

ΦS = Φ ∩ Zα
α∈ S
of Φ is the root system of m S with simple roots S and positive roots Φ S+ := Φ + ∩ Φ S . Denote the
Weyl group of Φ S by W S , viewed as a subgroup of W . Let w S denote the longest element of W S .
Let U (g) denote the universal enveloping algebra of g. We will work with U (g)-modules in the
parabolic category O S , defined as follows (see [RC]).
Definition 1. Let O S be the full subcategory of the category of U (g)-modules consisting of modules V
which satisfy the following conditions:
(i) V is a finitely generated U (g)-module.

(ii) As a U (m S )-module, V is the direct sum of finite-dimensional U (m S )-modules.
(iii) If v ∈ V , then dimC U (u+
S ) v < ∞.
Define X + ∗ +
S = {μ ∈ h | (μ, α̌ ) ∈ Z0 for all α ∈ Φ S }. The key objects in category O S are the
parabolic Verma modules, which are constructed as follows. Start with a finite-dimensional simple
m S -module F (μ) with highest weight μ ∈ X + +
S (F (μ) is finite-dimensional if and only if μ ∈ X S ).
+
Extend F (μ) to a p S -module by letting u S act by zero. The induced module
V (μ) = U (g) ⊗U (p S ) F (μ)
is a parabolic Verma module (or PVM for short). These are also called generalized Verma modules in the
literature.
V (μ) has the following properties. First, it is a highest weight module for g with highest weight μ.
Hence, it is a quotient of the ordinary Verma module M (μ) having highest weight μ. Also, V (μ) is
an object with finite length in category O S . Furthermore, V (μ) has a unique maximal submodule and
hence a unique simple quotient module, which we denote by L (μ); L (μ) is also the unique simple
quotient of M (μ). Each simple module in O S is isomorphic to some L (μ).
Every module in O S has a projective cover, and so there is a one-to-one correspondence between
the simple modules and the projective indecomposable modules in O S (see [RC]). Let P (μ) be the
projective cover of L (μ) in O S for each μ ∈ X + S .
Every projective module P in O S has a parabolic Verma composition series:
P = P 0 ⊇ P 1 ⊇ · · · ⊇ P r −1 ⊇ P r = 0
such that P i −1 / P i = V (νi ) for some νi ∈ X +

S (1 i r) (see [RC]). We have the following reciprocity
law:

P (μ) : V (ν ) = V (ν ) : L (μ) for all μ, ν ∈ X +
S .
2.3. Infinitesimal blocks of O S
Let Z be the center of U (g) and denote the set of algebra homomorphisms Z → C by Z . We say
that χ ∈ Z is the infinitesimal character of some nonzero V ∈ O S if zv = χ ( z) v for all z ∈ Z and all
χ
v ∈ V . For each χ ∈ Z , let O S be the full subcategory of O S consisting of modules V ∈ O S such that
for all z ∈ Z , each v ∈ V is annihilated by some power of z − χ ( z). We have the decomposition
χ
OS = OS
χ∈Z
of the category O S in which each module in O S decomposes into a direct sum of modules with each
χ χ
summand belonging to one of the subcategories O S . We call O S an infinitesimal block of category O S ,
corresponding to the infinitesimal character χ .
For each μ ∈ h∗ , the ordinary Verma module M (μ) (and any quotient thereof, such as V (μ) or
L (μ) if μ ∈ X +
S ) has an infinitesimal character which we will denote by χμ ∈ Z . Furthermore, if

χ ∈ Z is an infinitesimal character, then there exists μ ∈ h∗ such that χ = χμ . If χ = χμ , we can

μ χμ χ
write O S = O S = O(g, S , μ) for O S . The Harish-Chandra linkage principle yields
χμ = χν ⇐⇒ ν ∈ W · μ.
μ
Thus, V (ν ) (resp. L (ν ), P (ν )) is in O S if and only if ν = w S w · μ for some w ∈ W . Since PVM’s are
constructed from the finite-dimensional m S -modules with highest weights in X + S , the set of PVM’s in
O S is { V ( w S w · μ) | w S w · μ ∈ X +
μ
S } . Consequently, the PVM’s (as well as the simple modules and
projective indecomposable modules) in O S are parameterized by { w ∈ W | w S w · μ ∈ X +
μ
S }.
Assume from now on that μ is an integral weight and μ + ρ is antidominant; i.e., (μ + ρ , α̌ ) ∈ Z0
for all α ∈ ; (if it is not antidominant, we can replace it by a W -translate, so we are justified in
making this assumption). Let

Φμ = α ∈ Φ (μ + ρ , α̌ ) = 0 .
μ
If Φμ = ∅, then μ + ρ is a regular weight. If μ + ρ and ν + ρ are both regular weights, then O S is
equivalent to O νS by the Jantzen–Zuckerman translation principle. If μ + ρ is a regular weight, then
{w ∈ W | w S w · μ ∈ X +
S } is the set

S
W = w ∈ W l(sα w ) = l( w ) + 1 for all α ∈ S

= w ∈ W w −1 Φ S+ ⊆ Φ +
which is the set of smallest length representatives for the right cosets of W S in W .
Now, if μ ∈ h∗ is such that Φμ = ∅, then Φμ is a subroot system of Φ , and in this case μ + ρ is
a singular weight. Suppose α ∈ Φ + ∩ Φμ . Then (μ + ρ , α̌ ) = 0 and also we can write α = i =1 ai αi
n
for some ai ∈ Z 0 . Since μ + ρ is antidominant, (μ + ρ , α̌ i ) 0 for each 1 i n. Consequently,

n
0 = (μ + ρ , α ) = i =1 ai (μ + ρ , αi ) implies that (μ + ρ , αi ) = 0 for any i ∈ {1, . . . , n} such that ai = 0.
Set

J= α ∈ (μ + ρ , α ) = 0 .
Then Φμ is the root system Φ J which has simple roots J . Note that the Weyl group W J of Φ J is the
stabilizer of μ + ρ .
If μ + ρ is singular, then w S w · μ ∈ X +
S if and only if w W J ⊆ W . Since W J stabilizes μ + ρ , the
S
set

S
W J = w ∈ S W w < wsα ∈ S W for all α ∈ J
is the set of smallest length representatives for the left cosets w W J contained in S W . Consequently,
μ
the set S W J parameterizes the set of inequivalent irreducible modules in the infinitesimal block O S .
μ
That is, the set of simple modules in O S is the set { L ( w S w · μ) | w ∈ W } (see [BN]).
S J
As with S, we will frequently view the set J as a subset of {1, . . . , n}. We will use the notation
μ
O S = O(Φ, S , J ) = O(g, S , J )
when Φμ = Φ J .
3. Radical filtrations and the U α -algorithm
3.1. Radical filtrations and extensions
The radical of a g-module V , denoted rad V is the smallest submodule of V such that V / rad V
is semisimple. If V is a g-module, set rad0 V = V and for each i 1, set radi V = rad(radi −1 V ). We
thus have the radical filtration of V :
V = rad0 V ⊇ rad1 V ⊇ rad2 V ⊇ · · · .
If V is a finite length module (i.e., all chains of submodules in V have finite length), then for each
i 0, define radi V = radi V / radi +1 V , which is called the ith radical layer of V . Each PVM has a finite
radical filtration.
Let A be a finite-dimensional algebra over C. Then A is Morita equivalent to some basic alge-
bra Λ. Let L 1 , . . . , L r be a complete set of non-isomorphic simple Λ-modules with corresponding
projective covers P 1 , . . . , P r . The Ext1Λ -quiver Q (Λ) of Λ has its vertices in one-to-one correspon-
dence with the simple modules { L i }, and the number of arrows from vertex i to vertex j is equal to
dimC Ext1Λ ( L i , L j ) = dimC HomΛ ( P j , rad( P i ))/ HomΛ ( P j , rad2 ( P i )). From a theorem of Gabriel [Gab2],
the basic finite-dimensional algebra Λ is isomorphic to C Q (Λ)/ I for some ideal I of the path algebra
C Q (Λ); therefore, the category of A-modules is equivalent to the category of representations of some
path algebra of a quiver with relations.
Every extension between two irreducible modules in O S arises from an extension between them
in layers 0 and 1 of the radical filtration of some PVM. In fact, if Φμ = Φ J and x, w ∈ S W J satisfies
x < w, we have

dim Ext1O S L ( w S w · μ), L ( w S x · μ) = rad1 V ( w S w · μ) : L ( w S x · μ) .
Furthermore,
Ext1O S L ( w S w · μ), L ( w S x · μ) ∼
= Ext1O S L ( w S x · μ), L ( w S w · μ)
(see [BN, Sec. 2.3]). In particular, one has that if [rad1 V ( w S w · μ) : L ( w S x · μ)] =
0, then the Ext1 -
quiver for O S has an arrow from L ( w S w · μ) to L ( w S x · μ), and an arrow from L ( w S x · μ) to
L ( w S w · μ) .
μ
An infinitesimal block O S is semisimple if and only if there are no non-split extensions between
its simple modules. Because there are no self-extensions between simple modules in a highest weight
category, an infinitesimal block with only one PVM (and hence only one simple module) is necessarily
semisimple.
μ
Let S be a set of simple modules in O S corresponding to all the vertices in a single graph com-
μ μ
ponent of the Ext -quiver associated to O S . The full subcategory of O S consisting of those modules
1
μ μ
whose composition factors are contained in S is called a linkage class of O S . It is apparent that O S
μ μ
is semisimple if and only if rad1 V = 0 for all PVM’s V in O S if and only if each linkage class of O S
is composed of a single simple module.
3.2. The U α -algorithm
The U α algorithm is a tool used to compute radical filtrations of PVM’s in an infinitesimal block
μ
O S (see [Irv,Vog]).
First, let λ be a regular antidominant integral weight. Fix a simple reflection sα for some α ∈ .
If one composes the translation functors ‘onto’ and ‘out of’ the α -wall, one gets an exact covariant
functor θα on O S called translation through the α -wall. For w ∈ S W , θα L ( w S w · λ) = 0 unless w <
wsα ∈ S W ; in this case, θα L ( w S w · λ) has radical filtration layers
L ( w S w · λ),
U α L ( w S w · λ),
L ( w S w · λ),
where U α L ( w S w · λ) is a semisimple module defined as follows. Let

W = x ∈ S W x > xsα or xsα ∈
/ SW .
For x, w ∈ S W with x < w, let μ S (x, w ) be the coefficient of q(l( w )−l(x)−1)/2 in the relative Kazhdan–
Lusztig polynomial P xS, w (q), called the relative Kazhdan–Lusztig μ-function (see [CC]). In fact,
μ S (x, w ) = [rad1 V ( w S w · λ) : L ( w S x · λ)] (see [BN]). Now,

U α L ( w S w · λ) = L ( w S wsα · λ) ⊕ μ S (x, w ) L ( w S x · λ).
x∈W
One can start with V ( w S e · λ) = L ( w S e · λ) and use the fact that if w ∈ S W with w < wsα ∈ S W ,
then θα V ( w S w · λ) is a non-split extension of V ( w S wsα · λ) by V ( w S w · λ) to compute inductively
the composition factors of each V ( w S w · λ).
Using a ‘graded’ version of the U α -algorithm, one can compute not just the composition factors
but also the radical filtrations of the PVM’s (see [Bac,BGS,BN,Irv,Str]). Given a module M with filtration
{ M i }, define σ M to be the same module with filtration (σ M )i = M i −1 . Suppose w , wsα ∈ S W with
w < wsα and that the radical filtration of V ( w S w · λ) is known. Compute the radical filtration V =
rad0 V ⊇ rad1 V ⊇ rad2 V ⊇ · · · of V := V ( w S wsα · λ) as follows. First, the module θα V ( w S w · λ) has
the following filtration. For each i 0, let L ( w S y · λ) be a composition factor of radi V ( w S w · λ) with
y , ysα ∈ S W and y < ysα (so that θα L ( w S y · λ) = 0). If j = 0, 1, 2, then rad j θα L ( w S y · λ) occurs in
the (i + j )th layer of θα V ( w S w · λ). There is a short exact sequence
0 → σ V ( w S wsα · λ) → θα V ( w S w · λ) → V ( w S w · λ) → 0
of filtered modules. Hence, deleting the known radical filtration of V ( w S w · λ) from θα V ( w S w · λ)

leaves the radical filtration of V ( w S wsα · λ) (with all layers shifted up in index).
Now suppose that μ is any antidominant integral weight and let J ⊆ be the set of simple roots
on which μ + ρ is singular. If x, w ∈ S W J , then

radi V ( w S w · μ) : L ( w S x · μ) = radi V ( w S w · λ) : L ( w S x · λ) .
Consequently, the radical filtration of V ( w S w · μ) is obtained from that of V ( w S w · λ) by ignoring all

simple modules L ( w S y · λ) for y ∈
/ SW J.
4. Representation type of infinitesimal blocks of category O S
In this section, we compile some criteria to determine the representation type of a given infinites-
imal block of category O S . These criteria can be used whenever the structure of the PVM’s in the
μ
infinitesimal block O S is known; however, one criterion for wild representation type depends only
on knowing something about the Bruhat order on S W J (Proposition 4).
4.1. Triangular infinitesimal blocks
Suppose a linkage class of O (g, S , J ) has m simple modules, labeled L 1 , . . . , L m . If the correspond-
ing PVM’s V 1 , . . . , V m have radical filtration layers
V1 V2 V3 ··· V m −1 Vm
L1 L2 L3 ··· L m −1 Lm
L1 L2 ··· L m −2 L m −1
L1 ··· L m −3 L m −2 (4.1.1)
.. .. ..
. . .
L1 L2
L1
then we say that the linkage class of O (g, S , J ) is a triangular linkage class of length m. If O (g, S , J )
has only one linkage class and it is triangular of length m, then we say that O (g, S , J ) is a triangular
block of length m.
The following theorem classifies the representation type of all triangular infinitesimal blocks. For
its proof, see [FNP, Props. 5.3, 6.2, 7.1, 7.2]
Theorem 1. Suppose O (g, S , J ) is triangular of length m.
(i) If m = 1, then O (g, S , J ) is semisimple.

(ii) If m = 2 or m = 3, then O (g, S , J ) has finite representation type.
(iii) If m = 4, then O (g, S , J ) has tame representation type.
(iv) If m 5, then O (g, S , J ) has wild representation type.
4.2. Finite representation type
Suppose a linkage class of O (g, S , J ) has m simple modules. If these simple modules are labeled
L 1 , . . . , L m and if the corresponding PVM’s V 1 , . . . , V m have radical filtration layers
V1 V2 V3 ··· V m −1 Vm
L1 L2 L3 ··· L m −1 Lm (4.2.1)
L1 L2 ··· L m −2 L m −1
then we say that the linkage class of O (g, S , J ) is uniserial of length 2.

The following theorem says that there are very strict conditions placed on the structures of PVM’s
in a block having finite representation type.
Theorem 2. O (g, S , J ) has finite representation type if and only if the linkage classes of O (g, S , J ) are unise-
rial of length 2 or triangular of length 3.
For details, see [DoRe, Sec. 1] and [BN, Sec. 3.1].

Fig. 2. A kite.
Fig. 3. The ‘diamond’ Ext1 -quiver.
4.3. Wild representation type
4.3.1. Kite in the Ext1 -quiver

Let Λ be a finite-dimensional 2-nilpotent algebra. Gabriel’s Theorem [Gab1] asserts that the Ext1 -
quiver of Λ separates into a union of quivers whose underlying graphs are Dynkin diagrams if and
only if Λ has finite representation type. Furthermore, Dlab and Ringel [DlRi] proved that Λ has tame
representation type if and only if the Ext1 -quiver of Λ separates into a union of quivers whose
underlying graphs are Dynkin or extended Dynkin diagrams with at least one extended Dynkin di-
agram.
μ
Suppose the infinitesimal block O S is equivalent to the module category of a quasi-hereditary
algebra A. Consider the finite-dimensional 2-nilpotent algebra Λ = A / rad2 A. Since A Λ, if Λ has
wild representation type then so does A. Furthermore, Λ and A have the same Ext1 -quivers since
each extension between simple modules arises as an extension between layers 0 and 1 of the radical
μ
filtration of some PVM in O S .
1
Now suppose the Ext -quiver of Λ contains a ‘kite’ with any orientation on the arrows, such as
the kite shown in Fig. 2. Since the underlying graph is not a Dynkin diagram nor an extended Dynkin
diagram, Λ must have wild representation type. This proves the following proposition.
μ μ
Proposition 3. If the Ext1 -quiver associated to O S contains a kite, then O S has wild representation type.
4.3.2. Diamond infinitesimal blocks

The following argument is an adaptation of the argument in [FNP, Sec. 4.2] proving that
μ
O( A 1 × A 1 , ∅, ∅) has wild representation type. Suppose g is any simple Lie algebra and O S has ex-
actly four simple modules L 1 , L 2 , L 3 , L 4 and the corresponding PVM’s have radical filtration layers:
V1 V2 V3 V4
L4 (4.3.1)
L2 L3
L1 L2 L3
L1 L1
L1
By noting the simple modules in layers 0 and 1 of each PVM, we have the Ext1 -quiver for OS depicted
in Fig. 3. Using reciprocity, we can compute the radical filtration layers of the projective indecompos-
Fig. 4. Projective indecomposable modules for ‘diamond’.
Fig. 5. Subquiver of e Ae.
able modules [Irv]. These are shown in Fig. 4, where a line between simple modules represents an
extension between them.
μ
Let P = P 1 ⊕ P 2 ⊕ P 3 ⊕ P 4 and set A = EndOμ ( P )op , so O S is Morita equivalent to the category of
S
finitely generated A-modules. Consider the idempotent e = 1 L 1 + 1 L 4 . Now, A has wild representation
type whenever the algebra e Ae has wild representation type [Erd, I.4.7], and so localizing at e and
using the structure of P 1 and P 4 , we conclude that the quiver of e Ae has a subquiver shown in Fig. 5.
μ
[Erd, I.10.8(i)] implies that e Ae has wild representation type and therefore O S has wild representation
type.
One application of diamonds and kites is the following proposition due to [BN]. We say that four
distinct elements w , x1 , x2 , y ∈ W form a diamond if y < xi < w for i = 1, 2 and l( w ) = l( y ) + 2.
Proposition 4 (Boe–Nakano). If S W J
contains a diamond, then O (g, S , J ) has wild representation type.
This follows because a diamond in S W J gives rise to either a kite in the Ext1 -quiver or else a
linkage class of O (g, S , J ) which has exactly four simple modules with corresponding PVM’s with
structures as in (4.3.1).
5. Representation type of infinitesimal blocks in type F 4 and G 2
We determine the representation type of each infinitesimal block O (Φ, S , J ) when Φ is of type
F 4 or G 2 using the results in Sections 3 and 4. Most of the results were found using a computer.
This was done by first generating S W J to see if it is non-empty. If it contains at least four elements,
we generate the Hasse diagram of S W J induced by the Bruhat order on W , and look for diamonds.
If no diamonds are found or there are only two or three elements in S W J , we generate the radical
filtrations of the PVM’s in the infinitesimal block using the U α -algorithm and look for kites, uniserial
linkage classes, triangular linkage classes, or singleton linkage classes.
First, we adopt the following notation for the subroot systems of F 4 and G 2 . If T ⊆ contains only
long roots (i.e., T ⊆ {α1 , α2 } in F 4 or T = {α2 } in G 2 ), then Φ T is denoted A 1 or A 2 ; if T contains
only short roots (i.e., T ⊆ {α3 , α4 } in F 4 or T = {α1 } in G 2 ), then Φ T is denoted
A 1 or
A 2 . Since there
are two long simple roots and two short simple roots in type F 4 , if T = {α2 }, then we will denote Φ T
by A 1 and similarly if T = {α4 }, then we will denote Φ T by A 1 .
Table 1
Representation type of O (G 2 , Φ S , Φ J ).
ΦJ
G2 A1
A1 ∅
SS
G2 − − −
(1)
SS SS F
A1 − (2) (3) (6)
ΦS
SS SS F
A1 −
(3) (2) (6)
SS WT WT WD
∅
(1) (6) (6) (12)
The representation types of the infinitesimal blocks for types G 2 and F 4 are tabulated in Tables 1
and 2, respectively. The representation type of the infinitesimal block O (Φ, S , J ) is given on the row
labeled with the root system Φ S and the column labeled by the root system Φ J . SS means the block is
semisimple, F means it has finite representation type, and W means it has wild representation type;
there are no tame blocks for these algebras. A dash (−) means O (g, S , J ) = 0. The number below
each SS in the table represents the number of simple modules in the semisimple block; this is also
the number of linkage classes in the block. The number(s) below each F and W indicate the number
of simple modules in the corresponding block; pairs of numbers indicate that the infinitesimal block
splits into two linkage classes having the specified numbers of simple modules. For example, note
that O ( F 4 , A 2 ,
A 2 ) has two linkage classes, and they have, respectively, 20 and 12 simple modules
for a total of 32 simple modules in the infinitesimal block. If an infinitesimal block is not semisimple,
then the block does not split into more than two linkage classes in types F 4 and G 2 .
If Φ is of type F 4 or G 2 , all the infinitesimal blocks of finite representation type are composed of
linkage classes which are uniserial length two as in (4.2.1). The superscript above each W indicates
what condition was used to determine that the infinitesimal block has wild representation type. If
there is a diamond in the poset of S W J , then it is marked W D in the table. If the poset contains no
diamonds but there is a kite in the Ext1 -quiver, then it is marked W K in the table. If the infinitesimal
block is triangular of length at least five, then it is marked with W T . Notice that as one moves right
in a row or down in a column, one expects to eventually find diamonds in the poset of S W J .
Partition the subroot systems of G 2 as
[G 2 ] := {G 2 }, [ A 1 ] := { A 1 ,
A 1 }, [∅] := {∅}
and partition the subroot systems of F 4 as
[ F 4 ] := { F 4 }, [ B 3 ] := { B 3 }, [C 3 ] := {C 3 },

[ B 2 ] := A 2 × A 1 , B 2 ,
A2 × A1 , [ A 2 ] := { A 2 },

[
A 2 ] := { A 2 }, [ A1 × A 1 ] := A 1 × A1, A1 × A 1 , A 1 ×
A 1 ,

[ A 1 ] := A 1 , A 1 , A1,
A 1 , [∅] := {∅}.
These equivalence classes are enclosed by horizontal and vertical lines in Tables 1 and 2.
Using the data in Tables 1 and 2, one can make some observations. To facilitate what follows, for
the remainder of the paper the statement “representation type of O (Φ, S , J )” will mean one of the
four mutually exclusive conditions for the block: zero, semisimple, finite representation type (but not
semisimple), or wild representation type.
First, observe that the horizontal and vertical lines in Tables 1 and 2 divide the tables into rectan-
gles with entries in each rectangle having the same representation type. This proves the following.
Table 2
Representation type of O ( F 4 , Φ S , Φ J ).
∅
C3
B3
B2
F4
A2 ×
A2 × A1
A2
A2
A1 ×
A 1 ×
A1 ×
A1
A 1
A1
A
1
ΦJ
A 1
A 1
A1
A 1
ΦS
SS
F4 − − − − − − − − − − − − − − − (1)
SS F F F WK WK WK WK WD
B3 − − − − − − − (1) (2) (2) (2) (6) (6) (9) (9) (24)
SS F F F WK WK WK WK WD
C3 − − − − − − −
K.J. Platt / Journal of Algebra 322 (2009) 3823–3838

(1) (2) (2) (2) (9) (9) (6) (6) (24)
SS SS SS F F WK WK WK WD WD WD WD WD
A2 ×
A 1 − − − (3) (5) (4) (6) (6,6) (17) (17) (17) (36) (36) (44) (44) (96)
A2 × A1 − − − (5) (3) (4) (6,6) (6) (17) (17) (17) (44) (44) (36) (36) (96)
B2 − − − (4) (4) (9) (6,6) (6,6) (24) (24) (24) (60) (60) (60) (60) (144)
SS WT WT WT WD WD WD WD WD WD WD WD WD WD
A2 − − (1) (6) (6,6) (6,6) (12) (20,12) (36) (36) (36) (72) (72) (48,48) (96) (192)
SS WT WT WT WD WD WD WD WD WD WD WD WD WD
A2 − (1) − (6,6) (6) (6,6) (20,12) (12) (36) (36) (36) (96) (48,48) (72) (72) (192)
F F WK WK WK WD WD WD WD WD WD WD WD WD WD
A1 ×
A1 − (2) (2) (17) (17) (24) (36) (36) (61) (61) (61) (132) (132) (132) (132) (288)
A 1 ×
A 1 − (2) (2) (17) (17) (24) (36) (36) (61) (61) (61) (132) (132) (132) (132) (288)
A1 ×
A 1 − (2) (2) (17) (17) (24) (36) (36) (61) (61) (61) (132) (132) (132) (132) (288)
WK WK WD WD WD WD WD WD WD WD WD WD WD WD WD
A1 − (6) (9) (36) (44) (60) (72) (96) (132) (132) (132) (264) (264) (288) (288) (576)
A 1 − (6) (9) (36) (44) (60) (72) (96) (132) (132) (132) (264) (264) (288) (288) (576)
A1 − (9) (6) (44) (36) (60) (96) (72) (132) (132) (132) (288) (288) (264) (264) (576)
A 1 − (9) (6) (44) (36) (60) (96) (72) (132) (132) (132) (288) (288) (264) (264) (576)
SS WD WD WD WD WD WD WD WD WD WD WD WD WD WD WD
∅ (1) (24) (24) (96) (96) (144) (192) (192) (288) (288) (288) (576) (576) (576) (576) (1152)
3833
Fig. 6. Partial ordering on equivalence classes of subroot systems of G 2 (left) and F 4 (right).
Observation 1. If Φ is of type F 4 or G 2 , then O (Φ, S , J ) and O (Φ, S , J ) have the same representa-
tion type whenever Φ S and Φ S belong to the same equivalence class and Φ J and Φ J belong to the
same equivalence class.
We will define the representation type of O (Φ, [Φ S ], [Φ J ]) to be the representation type of

O(Φ, S , J ); this is well defined because of Observation 1.
Observation 2. If Φ is of type F 4 or G 2 , then O (Φ, [Φ S ], [Φ J ]) has the same representation type as

O(Φ, [Φ J ], [Φ S ]) except in the following cases: each of
O G 2 , [ A 1 ], [∅] , O F 4 , [ B 2 ], [ A 2 ] , O F 4 , [ B 2 ], [
A2]
has finite representation type, but
O G 2 , [∅], [ A 1 ] , O F 4 , [ A 2 ], [ B 2 ] , O F 4 , [
A 2 ], [ B 2 ]
have wild representation type. In particular, O (Φ, [Φ S ], [Φ J ]) is not zero (resp., semisimple) if and
only if O (Φ, [Φ J ], [Φ S ]) is not zero (resp., semisimple).
One observes this by noting that Tables 1 and 2 are symmetric across the main diagonal except in
the cases mentioned.
Now define a partial ordering on the equivalence classes of subroot systems of G 2 (resp., F 4 ) by
the left (resp., right) Hasse diagram in Fig. 6. Denote both of these partial orderings by ≺ (or by to
include the possibility of equality). We are now ready for the main theorem.
Theorem 5. Let Φ be of type F 4 or G 2 .
(i) Suppose S ⊆ .
(a) There exists an equivalence class [Φ S ] such that if [Φ S ] ≺ [Φ J ] then O (Φ, [Φ S ], [Φ J ]) = 0. Further-
more, if O (Φ, [Φ S ], [Φ J ]) = 0, then [Φ S ] ≺ [Φ J ] or [Φ J ] and [Φ S ] are incomparable. In particular,
O(Φ, [Φ S ], [Φ S ]) = 0.
Table 3
Associated equivalence classes for G 2 .
[ΦT ] [ΦT ] or [ΦT ] [Φ T ] [ΦT ]
[G 2 ] [∅] [∅] [∅]

[ A1 ] [ A1 ] [∅] [ A1 ]
[∅] [G 2 ] [G 2 ] [ A1 ]
(b) There exists an equivalence class [Φ S ] such that [Φ J ] ≺ [Φ S ] if and only if O (Φ, [Φ S ], [Φ J ]) has wild
representation type.
(c) If [Φ S ] ≺ [Φ S ], then O (Φ, [Φ S ], [Φ S ]) is semisimple and O (Φ, [Φ S ], [Φ S ]) has finite representation
type.
(d) If [Φ S ] ≺ [Φ S ], then for all J ⊆ , O (Φ, [Φ S ], [Φ J ]) is either zero or has wild representation type. In
particular O (Φ, [Φ S ], [Φ S ]) has wild representation type and O (Φ, [Φ S ], [Φ S ]) = 0.
(e) If [Φ S ] = [Φ S ], then O (Φ, [Φ S ], [Φ S ]) is semisimple or has finite representation type.
(f) If [Φ S ] is incomparable to [Φ S ], then [Φ S ] ∈ {[ A 2 ], [
A 2 ], [ A 1 ×
A 1 ]} in which case [Φ S ], [Φ S ] ∈
{[ B 3 ], [C 3 ]} and neither O(Φ, [Φ S ], [Φ S ]) nor O(Φ, [Φ S ], [Φ S ]) has wild representation type.
(ii) Suppose J ⊆ .
(a) There exists an equivalence class [Φ J ] such that if [Φ J ] ≺ [Φ S ] then O (Φ, [Φ S ], [Φ J ]) = 0. Further-
more, if O (Φ, [Φ S ], [Φ J ]) = 0, then [Φ J ] ≺ [Φ S ] or [Φ S ] and [Φ J ] are incomparable. In particular,
O(Φ, [Φ J ], [Φ J ]) = 0.
(b) There exists an equivalence class [Φ J ] such that [Φ S ] ≺ [Φ J ] if and only if O (Φ, [Φ S ], [Φ J ]) has wild
representation type.
(c) If [Φ J ] ≺ [Φ J ], then O (Φ, [Φ J ], [Φ J ]) is semisimple and O (Φ, [Φ J ], [Φ J ]) has finite representation
type.
(d) If [Φ J ] ≺ [Φ J ], then for all S ⊆ , O (Φ, [Φ S ], [Φ J ]) is either zero or has wild representation type. In
particular O (Φ, [Φ J ], [Φ J ]) has wild representation type and O (Φ, [Φ J ], [Φ J ]) = 0.
(e) If [Φ J ] = [Φ J ], then O (Φ, [Φ J ], [Φ J ]) is semisimple or has finite representation type.
(f) If [Φ J ] is incomparable to [Φ J ], then [Φ J ] = [ A 1 × A 1 ] and [Φ J ], [Φ J ] ∈ {[ B 3 ], [C 3 ]} and both
O(Φ, [Φ S ], [Φ J ] and O(Φ, [Φ S ], [Φ J ]) have finite representation type.
Proof. If T ⊆ , then the (not in general unique) equivalence classes [Φ T ], [Φ T ], [Φ T ], and [Φ T ]

corresponding to a given [Φ T ] are listed in Table 3 for G 2 and Table 4 for F 4 . The proof of the
theorem now follows by inspecting Tables 1 and 2. 2
The somewhat complicated description of the representation type of infinitesimal blocks for a Lie
algebra of type F 4 as given in Theorem 5 arises in part because for an arbitrary equivalence class [Φ T ],
the equivalence classes [Φ T ], [Φ T ], [Φ T ], and [Φ T ] are not in general unique. If one introduces three
more “virtual classes” that act as placeholders in the partial ordering, one gets the top-to-bottom
symmetric extended Hasse diagram given in Fig. 7. (These virtual classes come from Richardson orbits
which are not labeled by standard root systems.)
For each equivalence class [Φ T ] of a subroot system of F 4 , let [Φ T ]∗ be the (possibly virtual) class
obtained by reflecting [Φ T ] across the horizontal line of symmetry in the extended Hasse diagram

of F 4 . Now, given an equivalence class [Φ T ], define the (possibly virtual) class [Φ T ] by

[ΦT ]∗ if [ΦT ] =
[ B 2 ],

[Φ T]=
[B2] if [Φ T ] = [ B 2 ]
Table 4
Associated equivalence classes for F 4 .
[ΦT ] [ΦT ] or [ΦT ] [ΦT ] [ΦT ]

[F4] [∅] [∅] [∅]
[B3] [A2 ] [ A1 × A1 ] [ A1 × A1 ]
[C 3 ] [ A2 ] [ A1 × A1 ] [ A1 × A1 ]
[B2] [B2] [ A 2 ], [
A2 ] [B2]
[ A2 ] [C 3 ] [ B 3 ], [C 3 ] [B2]
[A2 ] [B3] [ B 3 ], [C 3 ] [B2]
[ A1 ×
A1 ] [ B 3 ], [C 3 ] [ B 3 ], [C 3 ] [ B 3 ], [C 3 ]
[ A1 ] [ B 3 ], [C 3 ] [F4] [F4]
[∅] [F4] [F4] [F4]
Fig. 7. Extended poset of equivalence classes in F 4 .
and let L(C ) ∈ {0, 1, . . . , 9} denote the layer index of the (possibly virtual) class C in the extended
Hasse diagram in Fig. 7.
Theorem 6. Let Φ be of type F 4 and let [Φ S ], [Φ J ] be equivalence classes of (true) subroot systems of F 4 .
(i) The following are equivalent.

(a) O ( F 4 , [Φ S ], [Φ J ]) is non-zero.

(b) [Φ J ] [Φ S ].

(c) [Φ S ] [Φ J ].
(ii) The following are equivalent.
(a) O ( F 4 , [Φ S ], [Φ J ]) is semisimple.

(b) [Φ J ] = [Φ S ].
(c) [Φ S ] = [Φ J ].
(iii) The following are equivalent.
(a) O ( F 4 , [Φ S ], [Φ J ]) has finite representation type.
(b) L([Φ J ]) = L([Φ S ]) − 1.
(c) [Φ J ] = [ A 2 ] and [Φ J ] = [ B 2 ]; or [Φ J ] = [ [ A 2 ], [
A 2 ] and [Φ S ] = [ B 2 ]; or [Φ J ] = A 2 ], [ B 2 ] and
L([Φ S ]) = L([Φ J ]) − 1.
(iv) The following are equivalent.
(a) O ( F 4 , [Φ S ], [Φ J ]) has wild representation type.
(b) L([Φ J ]) L([Φ S ]) − 2.
(c) [Φ J ] = [ A 2 ] or [Φ J ] = [
A 2 ] and L([Φ S ]) L([Φ [ A 2 ], [
J ]) − 3; or [Φ J ] = A 2 ] and L([Φ S ])
L([Φ J ]∗ ) − 2.
Proof. Note that Observation 2 implies that interchanging [Φ S ] and [Φ J ] does not affect whether the
block is 0 or semisimple. One can see in Table 2 that O ( F 4 , [Φ S ], [Φ J ]) is semisimple exactly when
[Φ S ] = [Φ J ] and zero exactly when [Φ J ] [Φ S ] to obtain (i) and (ii). For each [Φ S ], if [Φ J ] is a true
equivalence class of a subroot system which lies exactly one layer down from [Φ S ], then one verifies
in Table 2 that O ( F 4 , [Φ S ], [Φ J ]) has finite representation type and that this exhausts all of the in-
finitesimal blocks with finite type. Furthermore, according to Observation 2, the only cases for which
O( F 4 , [Φ S ], [Φ J ]) has finite representation type but O( F 4 , [Φ J ], [Φ S ]) does not have finite represen-
tation type are when [Φ S ] = [ B 2 ] and [Φ J ] = [ A 2 ] or [Φ J ] = [
A 2 ] in which case O ( F 4 , [Φ J ], [Φ S ]) has
wild representation type. This proves (iii) and since we exhausted all cases but that of wild represen-
tation type, it also proves (iv). 2
In principle, one can use a computer to determine the representation types of the blocks in types
E 6 , E 7 , and E 8 . However, their sizes make this a much more difficult task; therefore, a better approach
would be desirable in these cases. Preliminary calculations suggest that there is likely an analogue of
Theorem 6 for all the exceptional Lie algebras.
Acknowledgments
The author wishes to acknowledge the help and support of Brian Boe, his PhD advisor, as well as
many useful discussions with Bobbe Cooper and Daniel Nakano.
References
[Bac] E. Backelin, Koszul duality for parabolic and singular category O , Represent. Theory 3 (1999) 139–152.
[BGS] A. Beilinson, V. Ginsburg, W. Soergel, Koszul duality patterns in representation theory, J. Amer. Math. Soc. 9 (1996)
473–527.
[BN] B.D. Boe, D.K. Nakano, Representation type of the blocks of category O S , Adv. Math. 196 (2005) 193–256.
[BKM] Th. Brüstle, S. König, V. Mazorchuk, The coinvariant algebra and representation types of blocks of category O , Bull.
London Math. Soc. 33 (2001) 669–681.
[CB] W.W. Crawley-Boevey, On tame algebras and BOCS’s, Proc. London Math. Soc. 56 (1988) 451–483.
[CC] L.G. Casian, D.H. Collingwood, The Kazhdan–Lusztig conjecture for generalized Verma modules, Math. Z. 195 (1987)
581–600.
[CPS] E. Cline, B. Parshall, L. Scott, Finite dimensional algebras and highest weight categories, J. Reine Angew. Math. 391 (1988)
85–99.
[Dro] Y.A. Drozd, Tame and wild matrix problems, in: Representation Theory II, Springer Lecture Notes in Math. 832 (1980)
242–258.
[DlRi] V. Dlab, C.M. Ringel, Indecomposable representations of graphs and algebras, Mem. Amer. Math. Soc. 6 (1976) v+57.
[DoRe] S. Donkin, I. Reiten, On Schur algebras and related algebras V: Some quasi-hereditary algebras of finite type, J. Pure Appl.
Algebra 97 (1994) 117–134.
[Erd] K. Erdmann, Blocks of Tame Representation Type and Related Algebras, Lecture Notes in Math., vol. 1428, Springer,
Berlin, 1990.
[FNP] V. Futorny, D.K. Nakano, R.D. Pollack, Representation type of the blocks of category O , Q. J. Math. 52 (2001) 285–305.
[Gab1] P. Gabriel, Unzerlegbare Darstellungen I, Manuscripta Math. 6 (1972) 71–103.
[Gab2] P. Gabriel, Auslander–Reiten sequences and representation-finite algebras, in: Representation Theory I, Ottawa 1979, in:
Lecture Notes in Math., vol. 831, Springer-Verlag, Berlin/New York, 1980.
[Irv] R.S. Irving, A filtered category O S and applications, Mem. Amer. Math. Soc. 83 (1990) v+117.
[RC] A. Rocha-Caridi, Splitting criteria for g-modules induced from a parabolic and the Berňsteĭn–Gel’fand–Gel’fand resolution
of a finite dimensional, irreducible g-module, Trans. Amer. Math. Soc. 262 (1980) 335–366.
[Str] C. Stroppel, Category O : Gradings and translation functors, J. Algebra 268 (2003) 301–326.
[Vog] D.A. Vogan, Irreducible characters of semisimple Lie groups I, Duke Math. J. 46 (1979) 61–108.
Journal of Algebra
The computation of the logarithmic cohomology for plane

curves
Francisco-Jesús Castro-Jiménez a , Nobuki Takayama b,∗
a
Department of Mathematics, University of Sevilla, Sevilla, Spain
b
Department of Mathematics, Kobe University, Rokko 1-1, Kobe, Japan
Article history: We will give algorithms of computing bases of logarithmic coho-

Received 30 January 2008 mology groups for square-free polynomials in two variables.
Available online 26 August 2009 © 2009 Elsevier Inc. All rights reserved.
Communicated by Bruno Salvy
Keywords:
Algebraic geometry
Gröbner basis
Logarithmic cohomology groups
Curves
D-modules
1. Introduction
Let us denote by R = C[x] = C[x1 , . . . , xn ] the polynomial ring, by A n = Cx1 , . . . , xn , ∂1 , . . . , ∂n the

complex Weyl algebra of order n and by (Ω R• , d) (or simply (Ω • , d)) the complex of polynomial (or
regular) differential forms (i.e. the complex of differential forms with polynomial coefficients) where
d is the exterior derivative.
The elements of A n are called linear differential operators with polynomial coefficients. An element
P (x, ∂) in A n can be written as a finite sum P (x, ∂) = α aα (x)∂ α where α = (α1 , . . . , αn ) ∈ Nn ,
α α
aα (x) ∈ R and ∂ α = ∂1 1 · · · ∂n n . Here ∂i stands for the partial derivative ∂∂x .
i
For a non-zero polynomial f ∈ R we denote by R f the ring of rational functions

g
Rf = g ∈ R, m ∈ N
fm
E-mail addresses: castro@us.es (F.-J. Castro-Jiménez), takayama@math.kobe-u.ac.jp (N. Takayama).
doi:10.1016/j.jalgebra.2009.03.039
3840 F.-J. Castro-Jiménez, N. Takayama / Journal of Algebra 322 (2009) 3839–3851
and by (Ω •f , d) := ( R f ⊗ R Ω R• , d) the complex of rational differential forms with coefficients in R f

where d is the corresponding exterior derivative.
Let us denote by DerC ( R ) the free R-module of polynomial vector fields (or equivalently of
C-linear derivations of R). Following K. Saito [19] we will denote by Der R (− log f ) (or simply
Der(− log f ) if no confusion is possible) the R-module of logarithmic vector fields with respect to f ,
i.e.

n

Der R (− log f ) = δ = ai (x)∂i ∈ DerC ( R ) δ( f ) ∈ R · f .
i =1
Der R (− log f ) is canonically isomorphic to the R-module Syz R (∂1 ( f ), . . . , ∂n ( f ), f )

of syzygies among
(∂1 ( f ), . . . , ∂n ( f ), f ). This isomorphism associates the logarithmic vector field δ = i ai (x)∂i with the
δ( f )
syzygy (a1 (x), . . . , an (x), − f ).
If f is a non-zero constant, then Der(− log f ) = DerC ( R ). So we will assume from now that f is a
non-constant polynomial in R.
It is clear that
f DerC ( R ) ⊂ Der(− log f ) ⊂ DerC ( R )
and then Der(− log f ) has rank n as R-module. The R-module Der(− log f ) does not depend on the
polynomial f but only on the hypersurface D = V ( f ) := {a ∈ Cn | f (a) = 0} ⊂ Cn .
Assume f is reduced (i.e. f is square-free). According to K. Saito [19] a rational differential p-
p
form ω ∈ Ω f is said to be logarithmic with respect to f (or with respect to the hypersurface
p p +1
D = V ( f ) ⊂ Cn ) if both f ω and f dω are regular (i.e. f ω ∈ Ω R and f dω ∈ Ω R ). We denote by
p
Ω R (log f ) (or simply Ω p (log f )) the R-module of logarithmic differential p-forms with respect to f .
K. Saito [19, Corollary 1.6] proved that Der(− log f ) is a reflexive R-module whose dual is Ω 1 (log f ).
We denote by (Ω • (log f ), d) the complex
d d d
0 → Ω 0 (log f ) −
→ Ω 1 (log f ) − → Ω n (log f ) → 0
→ ··· −
which will be called the logarithmic de Rham complex and is also, for simple notation, denoted by
Ω • (log f ) if no confusion arises.
Algorithms of computing dimensions and bases of the de Rham cohomology groups H i (Ω •f ) are
given by T. Oaku and N. Takayama [15,17] and U. Walther [22]. Here, f is any non-zero polynomial
in n variables. The purpose of this paper is to give algorithms of computing dimensions and bases
of the logarithmic de Rham cohomology groups H i (Ω • (log f )) as C-vector spaces in the case of two
variables.
1.1. Logarithmic comparison theorem
The rings R and R f have natural structures of left A n -module where ∂i acts on a polynomial g
g
and on a rational function f m as the partial derivative with respect to xi .
The de Rham complex of a left A n -module M, denoted by DR( M ), is by definition the complex of
C-vector spaces ( M ⊗ R Ω R• , ∇ • ) where
p p +1
∇ p : M ⊗R ΩR → M ⊗R ΩR

is defined, for p 1, by ∇ p (m ⊗ ω) = ∇ 0 (m) ∧ ω + m ⊗ dω and ∇ 0 (m) = i ∂i (m) ⊗ dxi . Notice
that am ⊗ ω = m ⊗ aω for m ∈ M, ω ∈ Ω p and a ∈ R. The complexes Ω •f and DR( R f ) are naturally
isomorphic.
F.-J. Castro-Jiménez, N. Takayama / Journal of Algebra 322 (2009) 3839–3851 3841
For any non-zero f ∈ R, the inclusion i f is a natural morphism of complexes
i f : Ω • (log f ) → Ω •f .
We say (see [3]) that f satisfies the (global) logarithmic comparison theorem if the morphism i f is a
quasi-isomorphism (i.e. if i f induces an isomorphism H p (Ω • (log f )) → H p (Ω •f ) for any p).
If n = 2 and f is a quasi-homogeneous polynomial such that the origin is the only singular point of
the plane curve defined by f , then i f is a quasi-isomorphism ([3, Corollary 2.7] and [2, Theorem 1.3]).
1.2. The case n = 2. Bases for Der(− log f )
If n = 2, any finitely generated reflexive R-module is projective and then, by the Quillen–Suslin
theorem, this R-module is free. So, if n = 2, the R-module Der(− log f ) is free of rank 2. In this case,
we would like to compute a basis of Der(− log f ) by taking the polynomial f = f (x1 , x2 ) as input. By
using the isomorphism
Der(− log f ) Syz R ∂1 ( f ), ∂2 ( f ), f
and using Gröbner basis computation, a system of generators of Der(− log f ) can be calculated. Then
we can apply Quillen–Suslin algorithm (as presented for example in [9] and implemented in [7]) to
compute such a basis. Known Quillen–Suslin algorithms use Gröbner bases computation. Nevertheless,
in some cases, for a big family of polynomials f (x1 , x2 ) we will use an easier way to compute a basis
of Der(− log f ).
First of all, we can assume f to be a reduced polynomial since Der(− log f ) depends only on the
affine plane curve D = V ( f ) = {(a1 , a2 ) ∈ C2 | f (a1 , a2 ) = 0} ⊂ C2 .
Assume the plane curve D = V ( f ) is not smooth. The singular points of the plane curve D = V ( f )
(i.e. the affine algebraic set

Sing( D ) := V ( f , f 1 , f 2 ) = a = (a1 , a2 ) ∈ C2 f (a) = f 1 (a) = f 2 (a) = 0 ),
where f 1 = ∂1 ( f ), f 2 = ∂2 ( f ) – consists of a finite number of points (and it is not the empty set).
We will consider the affine plane C2 as a Zariski open subset of the projective plane P2 (C), the
affine point (a1 , a2 ) is mapped into the point with homogeneous coordinates (1 : a1 : a2 ). Coordinates
in P2 (C) will be denoted by (x0 : x1 : x2 ) and then the line at infinity is defined by x0 = 0.
Let us denote h = H ( f ), h1 = H ( f 1 ) and h2 = H ( f 2 ) where H (−) denotes homogenization with
respect to the variable x0 . Denote by S = C[x0 , x1 , x2 ] the polynomial ring graded by the degree of
the polynomials. If J = (h, h1 , h2 ) denotes the ideal in S generated by h, h1 , h2 then the quotient ring
S / J has Krull dimension 1. Let us denote by S + the irrelevant ideal in S, i.e. the ideal generated by
x0 , x1 , x2 . The following result is well known (see e.g. [10, Theorem 17.6, p. 136]).
Proposition 1.1. The graded ring S / J is Cohen–Macaulay if and only if S + is not an embedded prime associ-
ated with J .
If S / J is Cohen–Macaulay then the projective dimension of S / J is 2 and J satisfies the Hilbert–

Burch theorem [6], i.e. there exists an exact sequence
φ2 φ1
0 → S 2 −→ S 3 −→ J → 0,
where φ1 ( g 0 , g 1 , g 2 ) = g 0 h + g 1 h1 + g 2 h2 and φ2 is defined by a syzygy matrix of φ1 . In par-

ticular, since ker(φ1 ) = Syz S (h, h1 , h2 ) is a graded free S-module of rank 2 we can compute
{s(1) = (s10 , s11 , s12 ), s(2) = (s20 , s21 , s22 )} a minimal system of generators and this system is in fact a
basis of ker(φ1 ). By dehomogenization (i.e. by setting x0 = 1), we obtain a system {s(1) |x0 =1 , s(2) |x0 =1 }
of generators of Syz R ( f , f 1 , f 2 ) Der(− log f ) and since this R-module is free of rank 2, this last
system is in fact a basis.
If S / J is not Cohen–Macaulay we cannot apply, in general, the Hilbert–Burch theorem and the
previous procedure fails to compute a basis of Der(− log f ).
Example 1.2. (a) Consider the polynomial f = (x3 + y 4 + xy 3 )(x2 − y 2 ). With the notations as before
(and writing x1 = x, x2 = y, x0 = t) we can use Macaulay 2 to prove that the corresponding S / J is
Cohen–Macaulay and to compute a minimal system of generators of Syz S (h, h1 , h2 ) and then a basis
of Der(− log f ).
Macaulay 2, version 1.1 with packages: Classic, Core, Elimination,

IntegralClosure, LLLBases, Parsing,
PrimaryDecomposition, SchurRings, TangentCone
i1 : R=QQ[t,x,y];
i2 : f=(x^3+y^4+x*y^3)*(x^2-y^2);
i3 : f1=diff(x,f),f2=diff(y,f),h=homogenize(f,t),h1=homogenize(f1,t),h2=homogenize(f2,t);
i4 : Jf=ideal(h,h1,h2);
i5 : pdim coker gens Jf
o5 = 2
i6 : Syzf=kernel matrix({{h1,h2,h}});
i7 : mingens Syzf
o7 = {5} | 3x3+x2y-4xy2 -tx2+4txy+3x2y+4xy2-y3 |
{5} | 2x2y+xy2-3y3 tx2-txy+3ty2+2xy2+4y3 |
{6} | -15x2-5xy+18y2 5tx-18ty-15xy-23y2 |
3 2
o7 : Matrix R <--- R
Then a basis of Der(− log f ) is

1 4 2 1
x3 + x2 y − xy 2 ∂x + x2 y + xy 2 − y 3 ∂ y ,
3 3 3 3

2

−x + 4xy + 3x2 y + 4xy 2 − y 3 ∂x + x2 − xy + 3 y 2 + 2xy 2 + 4 y 3 ∂ y .
(b) Consider the polynomial g = (x3 + y 4 + xy 3 )(x2 + y 2 ). With the notations as before (and writing
x1 = x, x2 = y, x0 = t ) we can use Macaulay 2 to prove that the corresponding S / J is not Cohen–
Macaulay and the minimal number of generators of Syz S (h, h1 , h2 ) is 3. We can continue the last
Macaulay 2 session:
i8 : g=(x^3+y^4+x*y^3)*(x^2+y^2);
i9 : g1=diff(x,g),g2=diff(y,g),h=homogenize(g,t),h1=homogenize(g1,t),h2=homogenize(g2,t);
i10 : Jg=ideal(h,h1,h2);
i11 : pdim coker gens Jg
o11 = 3
i12 : Syzg=kernel matrix({{h1,h2,h}});
i13 : mingens Syzg
o13 =
{5} | 3tx2-15x3-12txy-20x2y-6xy2-5y3 3x4+4x3y+3x2y2+4xy3 3tx3-3tx2y+12x3y+12txy2+16x2y2+6xy3+4y4 |
{5} | 3tx2+3txy-10x2y-9ty2-15xy2-y3 2x3y+3x2y2+2xy3+3y4 -3txy2+8x2y2+9ty3+12xy3+2y4 |
{6} | -15tx+75x2+54ty+100xy+11y2 -15x3-20x2y-13xy2-18y3 -15tx2+15txy-60x2y-54ty2-80xy2-16y3 |
3 3
o13 : Matrix R <--- R
We will revisit this example in Example 4.1.

2. Logarithmic A n -modules
An
Let us denote by M log f the quotient A n -module M log f = An Der(− log f )
. Moreover, we denote by
(− log f ) the set
Der

(− log f ) = δ + δ( f )
Der δ ∈ Der(− log f )
f
log f the quotient An -module

and by M
An
log f =
M .
(− log f )
A n Der
As quoted in Section 1.2, for n = 2 the R-module Der(− log f ) (and hence Ω 1 (log f )) is free of
rank 2. Moreover, by [19, 1.8] there exists an R-basis {δ1 , δ2 } of Der(− log f ) satisfying det( A ) = f
where
δi = ai1 ∂1 + ai2 ∂2 , i = 1, 2,
and A is the matrix (ai j ). Then the dual basis of {δ1 , δ2 } is {ω1 , ω2 } with
1 1
ω1 = (a22 dx1 − a21 dx2 ), ω2 = (−a12 dx1 + a11 dx2 ).
f f
The R-module Ω 2 (log f ) is free of rank 1 and ω1 ∧ ω2 is a basis of it. Moreover we have ω1 ∧ ω2 =
dx1 ∧dx2
f
.
Theorem 2.1. (See [1, Theorem 4.2.1].) Let f ∈ R = C[x, y ] be a non-zero reduced polynomial. There exists a
natural isomorphism in the corresponding derived category
Ω • (log f ) −
→ R Hom A 2 M log f , R ,
where the last complex is the solution complex of M log f with values in R.
Remark 2.2. This theorem is proved in [1] in the setting of analytic D -modules, using the notion of
V 0 -module and the logarithmic Spencer resolution of M log f . In our case, once a basis {δ1 , δ2 } is fixed
in Der(− log f ), the logarithmic Spencer resolution of M log f is nothing but
2 1 π
→ A 2 −→ A −
0→ A− → M log f → 0, (1)
where A stands for A 2 , the morphism π is the natural projection, the A-module morphism 1 is
defined by 1 ( P 1 , P 2 ) = P 1 δ1 + P 2 δ2 (for P i ∈ A) and 2 is defined by 2 ( Q ) = Q (−δ2 − b1 , δ1 − b2 )
for Q ∈ A where the polynomials b i are defined by the equality [δ1 , δ2 ] = δ1 δ2 − δ2 δ1 =
b1 δ1 + b2 δ2 .
Using the previous free resolution the solution complex R Hom A ( M log f , R ) is represented by the
complex of C-vector spaces
1∗ 2∗
0 → R −→ R 2 −→ R → 0,
where 1∗ ( g ) = (δ1 ( g ), δ2 ( g )) for g ∈ R and 2∗ (h1 , h2 ) = δ1 (h2 ) − δ2 (h1 ) − b1 h1 − b2 h2 for h i ∈ R.
The natural morphism of Theorem 2.1 is described as the morphism of complexes
d d
Ω 0 (log f ) = R −
→ Ω 1 (log f ) −
→ Ω 2 (log f )
η0 ↓ η1 ↓ η2 ↓
1∗ 2∗
R −→ R2 −→ R
where η0 = id, η1 (h1 ω1 + h2 ω2 ) = (h1 , h2 ) and η2 ( g ω1 ∧ ω2 ) = g for h1 , h2 , g ∈ R and where {ω1 , ω2 }

is the dual basis in Ω 1 (log f ) of the basis {δ1 , δ2 } in Der(− log f ). It is easy to check by com-
putation that this morphism η• of complexes of vector spaces is in fact an isomorphism of com-
plexes.
To each finitely generated left A n -module M we associate the complex of finitely generated right
A n -modules R Hom An ( M , A n ). To this one we associate the complex of finitely generated left A n -
modules Hom R (Ω Rn , R Hom An ( M , A n )) which is by definition the dual M ∗ of the left A n -module M
(see e.g. [11, Déf. 4.1.6]).
If M is holonomic (i.e. if the dimension of the characteristic variety is n) then it can be shown that
Ext iAn ( M , A n ) = 0 for i = n and then M ∗ is the left holonomic A n -module Hom R (Ω Rn , ExtnAn ( M , A n ))
An
(see e.g. [11, p. 41]). Assume ExtnAn ( M , A n ) = J
for some right ideal J ⊂ A n . Then Hom R (Ω Rn , A n / J )
An
is naturally isomorphic to the left A n -module JT
where J T is the left ideal J T = { P T | P ∈ J } and P T
is the formal adjoint of the operator P .
If N 1 , N 2 are finitely generated left A n -modules there exists a natural isomorphism of complexes
in the corresponding derived category (see e.g. [11, pp. 40–41])
R Hom An ( N 1 , N 2 ) → R Hom An R Hom An ( N 2 , A n ), R Hom An ( N 1 , A n )
and then a natural isomorphism
R Hom An ( N 1 , N 2 ) → R Hom An N 2∗ , N 1∗ .
In particular, if N 2 = R = C[x1 , . . . , xn ] then there exists a natural isomorphism from

R Hom An ( N 1 , R ) (i.e. the solution complex of N 1 ) to
R Hom An R ∗ , N 1∗ .
As the complex R Hom An ( R , A n ) is naturally isomorphic to Ω Rn we can identify R and R ∗ and then we
have a natural isomorphism
→ R Hom An R , N 1∗ −
R Hom An ( N 1 , R ) − → DR N 1∗ . (2)
The following theorem will be used later.
Theorem 2.3. (See [4, Theorem 3.1].) Let f ∈ C[x, y ] be a non-zero reduced polynomial. Then there exists a
log f .
natural isomorphism ( M log f )∗ M
The following corollary is an obvious consequence of Theorems 2.1 and 2.3 using also isomor-
phism (2).
log f )
Corollary 2.4. For any non-zero reduced polynomial f ∈ C[x, y ], the complexes Ω • (log f ) and DR( M
are naturally isomorphic in the derived category.
As a consequence of Corollary 2.4 and by [15,17,22], the cohomology of the complex Ω • (log f ) can
be computed from a given polynomial f , since a system of generators of the R-module Der (− log f )
can be computed by using the R-syzygies of (∂1 ( f ), ∂2 ( f ), f ) and we have algorithms of computing
DR( M ) for holonomic modules M. We note that M log f is shown to be holonomic [1, Corollary 4.2.2].
In order to compute bases of H i (Ω • (log f )), we give the explicit form
)
τ • : Ω • (log f ) → DR( M
of the quasi-isomorphism of complexes (regarded as objects in the abelian category) given in Corol-
lary 2.4. In general, this quasi-isomorphism is not an isomorphism between the chosen representatives
of the two complexes of vector spaces. This morphism is given as follows.
is defined by τ 0 ( g ) = g f where ( ) means the equivalence class modulo
τ 0 : R = Ω 0 (log f ) → M
(− log f ).
the ideal A 2 Der
⊗ R Ω 1 is defined by
τ 1 : Ω 1 (log f ) → M R

τ 1 ( g1 ω1 + g2 ω2 ) = g i ⊗ f ωi .
i
⊗ R Ω 2 is defined by τ 2 ( g ω1 ∧ ω2 ) = g ⊗ f ω1 ∧ ω2 .
τ 2 : Ω 2 (log f ) → M R
It is derived by a diagram chase of the double complex constructed from the Spencer resolution
and the Koszul resolution of R (see [21, Example 3.1] and [4]).
3. Algorithm
Let us summarize our algorithm of computing logarithmic cohomology groups in the two-
dimensional case. Most tensor products ⊗ in the sequel are over A 2 . If we omit the subscript A 2
for ⊗, it means that the tensor product is over A 2 .
Algorithm 3.1.
Input: a non-zero reduced polynomial f (x, y )
Output: dimensions and bases of H i (Ω • (log f )).
1. Compute a free basis {s = (s0 , s1 , s2 ), t = (t 0 , t 1 , t 2 )} of the syzygy module of f , f x , f y over the

polynomial ring C[x, y ]. This step can be performed by the following way.
(a) Compute the minimal syzygy of h( f ), h( f x ), h( f y ). Here, h( g ) is the homogenization of g. If
the number of generators is 2, then the dehomogenizations of these generators are s and t.
(b) If we fail on the first step, apply an algorithm for the Quillen–Suslin theorem to obtain s and
t (call the procedure Quillen–Suslin).
2. Define a left ideal in A 2 by
I = A 2 · {−s0 + s1 ∂x + s2 ∂ y , −t 0 + t 1 ∂x + t 2 ∂ y }. (3)
Compute the dimensions and bases of the de Rham cohomology groups for M = A 2 / I with the
algorithm in [15,17]. In other words, replace the A 2 -module C[x, y , 1/ f ] by A 2 / I of (3) in Algo-
rithm 1.2 in [15].
3. The bases of the previous step are given in A 2 /(∂x A 2 + ∂ y A 2 ) ⊗ M • where M • is a (1, 1, −1, −1)-
adapted free resolution of M. Here, M • = ( Ani , di ) is called a (1, 1, −1, −1)-adapted free resolu-
2
n n
tion when (gr A 2i , gr di ) is exact with a suitable degree shift where gr A 2i is the graded module
with the grading deg x = deg y = 1 and deg ∂x = deg ∂ y = −1 (see [17,18] as to details). Bases of
de Rham cohomology groups in Ω • ⊗ M q.is DR( M ) are determined by the transfer algorithm of
U. Walther [22, Theorem 2.5 (Transfer Theorem)]. Here, Ω • is the Koszul resolution of the right
A 2 -module A 2 /(∂x A 2 + ∂ y A 2 ).
4. Bases of cohomology groups in Ω • (log f ) is obtained by computing the preimage of τ i given in

). A procedure of computing the preimage is given at the end
Corollary 2.4; Ω • (log f ) qτ.is DR( M
of this section.
In the first step, we should firstly try to find the minimal syzygy. Because, usually it is faster than
applying implementations and algorithms for the Quillen–Suslin theorem.
The following example will illustrate how our algorithm works.
(− log f )
Example 3.2. We consider the case of f = xy (x − y ). Two canonical generators of I = A 2 Der
are
1 = 3 + x∂x + y ∂ y , 2 = −(2x − y ) + −x2 + xy ∂x .
The associated canonical logarithmic forms are
1 1
ω1 = x(x − y ) dy , ω2 = (− y dx + x dy ).
f f
Let us proceed on the step 2. We apply the procedure of computing the de Rham cohomology groups
[15,18] for A 2 / I . The maximal integral root of the b function for I = A 2 · {1 , 2 } with respect to the
weight (1, 1, −1, −1) is 1. The dehomogenization of the (1, 1, −1, −1)-minimal filtered free resolution
of A 2 / I , which is adapted, is
a −2 a −1
C •: A 2 [0] −−→ A 2 [1] ⊕ A 2 [0] −−→ A 2 [1], (4)
where
a−2 (c ) = c (−2 , 1 − 1) for c ∈ A 2 ,

1
a−1 (c , d) = (c , d) for (c , d) ∈ A 2 [1] ⊕ A 2 [0].
2
Put Ω(2) = A 2 /(∂x A 2 + ∂ y A 2 ) which is the right A 2 -module and is isomorphic to Ω R2 . Following
[15, Procedure 1.8], we truncate the complex Ω(2) ⊗ A 2 C • to the forms of (1, 1, −1, −1)-degree at
most 1 since the maximal integral root of the b-function is 1. The truncated complex is the following
complex of finite-dimensional vector spaces
ā −2 ā −1ā 0
C−−→ (C + Cx + C y ) ⊕ C −−→ (C + Cx + C y ) −→ 0. (5)
Here,
ā−2 (1) = (−2 , 1 − 1) mod ∂x A 2 + ∂ y A 2 = (0, 0),
ā−1 (a + bx + c y , d) = (a + bx + c y )1 + d2 mod ∂x A 2 + ∂ y A 2 = a.
Therefore, the cohomology groups are H 0 (Ω(2) ⊗ C • ) = Ker ā−2 = C, H 1 (Ω(2) ⊗ C • ) =

Ker ā−1 / Im ā−2 = (Cx + C y ) ⊕ C, H 2 (Ω(2) ⊗ C • ) = Ker ā0 / Im ā−1 = Cx + C y.
Finally, we perform the step 3 and 4. Put M = A 2 / I . In order to give bases of the cohomology
groups in M ⊗ R Ω i , we apply the transfer theorem (algorithm) of Uli Walther [22]. We consider the
R
double complex Ω • ⊗ C • constructed from Ω(2) ⊗ C • and Ω • ⊗ M̃. The transfer algorithm translates
cohomology classes in the complex Ω(2) ⊗ C • into those in the complex Ω • ⊗ M̃. This translation can
be performed by a diagram chase in the double complex with Gröbner basis computations. See 2.4
of [22].
Two cohomology classes 1 ⊗ x and 1 ⊗ y in H 2 (Ω(2) ⊗ C • ) are transfered to 1 ⊗ x and 1 ⊗ y in
H 2 (Ω • ⊗ M̃ ) respectively. It follows from the definition of τ 2 , xω1 ∧ ω2 and y ω1 ∧ ω2 is the basis of
H 2 (Ω • (log f )).
Let us compute transfers of bases of H 1 (Ω(2) ⊗ C • ). The cohomology class 1 ⊗ (x, 0) in
H 1 (Ω(2) ⊗ C • ) is transfered to (xy dx − x2 dy ) ⊗ 1 in H 1 (Ω • ⊗ M̃ ). Let us compute the preimage
by τ 1 . Solving c 1 f ω1 + c 2 f ω2 = xy dx − x2 dy (as cohomology class), we obtain c 1 = 0, c 2 = −x.
Therefore, 1 ⊗ (x, 0) stands for −xω2 . Analogously, 1 ⊗ ( y , 0) is transfered to − y 2 dx + xy dy and
stands for − y ω2 and 1 ⊗ (0, 1) is transfered to x( y − x) dy and stands for ω1 . In summary,
H 1 Ω • (log f ) = C(−x)ω2 + C(− y )ω2 + Cω1 .
) which is equal to
The base 1 ⊗ 1 in H 0 (Ω(2) ⊗ C • ) is transfered to xy (x − y ) ⊗ 1 ∈ H 0 (Ω • ⊗ M
τ 0 (1). Hence, H 0 (Ω • (log f )) = C · 1.
Before presenting implementations and larger examples, we explain about a procedure (step 4) to
find a preimage of τ i in general. The transfer algorithm gives an element in Ω i ⊗ A 2 M where Ω •
is the Koszul resolution of Ω(2) Ω R2 as the right A 2 -module. This element can be identified with
a differential form with coefficients in M. We need to find the preimage of it by τ i which lies in
Ω i (log f ). This can be performed by the method of undetermined coefficients.
Consider the case of τ 1 . Take an element c 1 ω1 + c 2 ω2 in Ω 1 (log f ) where c i ∈ R. We have seen in
Corollary 2.4 that

A2
1
τ (c1 ω1 + c2 ω2 ) = f ω̄1 ⊗ A 2 c̄1 + f ω̄2 ⊗ A 2 c̄2 ∈ ⊕ mod d M
⊗ A2 M . (6)
A2
1
0
Here, we identify ⊗ A 2 m1 with m1 ⊗ R dx and and when

⊗ A 2 m2 with m2 ⊗ R dy, mi ∈ M
0 ai
1
ωi = ai dx + bi dy, we denote bi
by ω̄i . We are given an element of the form m1 dx + m2 dy, mi ∈ M
as the output of the transfer algorithm. We regard mi as an element in A 2 in the sequel. We rewrite
f ωi as f ω1 = A dx + B dy and f ω2 = C dx + D dy. Assume I is generated by 1 and 2 . Then, it follows
j
from the definition of τ 1 and Corollary 2.4 that there exist c i ∈ R, di , e ∈ A 2 such that

2
j
Ac 1 + C c 2 = m1 + d1 j + ∂x e , (7)
j =1

2
j
Bc 1 + Dc 2 = m2 + d2 j + ∂ y e . (8)
j =1
j
To find the preimage c 1 ω1 + c 2 ω2 by τ 1 , we may solve (7) and (8). Fix a degree bound m for c i , di , e
and determine these elements by the method of unknown coefficients. The identities (7) and (8)
induce a system of linear equations over C for the coefficients. Increasing the degree bound and
solving the system, we will be able to obtain c 1 and c 2 in finite steps by virtue of the existence claim.
Consider the case of τ 2 . Since our basis in H 2 (Ω • ⊗ M̃ ) is given in terms of x and y and
f ω1 ∧ ω2 = dx ∧ dy, we need no computation to find the preimage by τ 2 .
Let us consider the case of τ 0 . Let m be an output of the transfer
2 algorithm. It lies in A 2 in general.
Finding the preimage g of τ 0 can be done by solving g f = m + j =1 d j j where g ∈ R and d j ∈ A 2 .
4. Implementation and examples
The second and third steps of Algorithm 3.1 can be performed with the help of the D-module pack-
age on Macaulay2 1.1 or later. Our code is merged to the D-module package [8] with the command
name logCohomology. Unfortunately, this implementation has not installed an efficient algorithm
of computing b-function by Noro [12] to get the truncated complex in [15,17]. Then, only relatively
small examples are feasible. Example 4.1 is computed by our Macaulay2 program. Example 4.2 is com-
puted by our implementation on kan/k0 (logc2.k, http://www.openxm.org) and Risa/Asir with
an implementation of [12] (the transfer algorithm has not been implemented yet for kan/k0). Our im-
plementation does not contain that for the Quillen–Suslin theorem. We utilize the implementation by
A. Fabianska on Maple [7] when the step 1(a) fails.
Example 4.1 (Continued from Example 1.2(b)). We will determine bases of H i (Ω • (log f )) where f =
(x3 + y 4 + xy 3 )(x2 + y 2 ). We firstly use Fabianska’s program for the Quillen–Suslin theorem to find
S S S
the 2 free generators of the syzygies of f , f x , f y . The two rows of the matrix S = S 11 S 12 S 13 are
21 22 23
generators where

115 43
S 11 = y − 5 /2 x − 6 y 3 − y2 + 9 y,
6 6

23
5
S 12 = − y + 1/2 x2 + y 3 + y 2 − 2 y x − y3 ,
6 6

1 2 2 1 4 3
S 13 = y + 1/2 x + −3 y + y x + y4 + y3 − y2,
3 2 3 2

46 24 22 12
S 21 = x2 + − y2 + y x+ y2 ,
15 25 75 5

46 3 4 2 2 8
S 22 = − x + y − y x2 − y 2 x,
75 25 25 15

4 12 4 2 2
S 23 = x3 − yx2 + y3 − y2 x − y3 .
75 25 25 75 5
S 12 S 13
1
The determinant of is f . We put ω1 = 1f ( S 23 dx − S 22 dy ) and ω2 = 1f (− S 13 dx + S 12 dy ).
√ S 22 S 23 3
( 3ωi agrees with the ωi in Theorem 2.4.)
We obtain the following
We apply the integration algorithm and the transfer algorithm for M.
• ))
result. (1) H (DR( M )) is spanned by 1 ⊗ f and then we have H (Ω (log f )) C · 1. (2) H 2 (DR( M
0 0
is spanned by 1 ⊗ a where a runs over
3 3 2 3 4
o9 = {{1}, {-x}, {y }, {-x*y }, {x*y }, {x y}, {y }}
(We have pasted the output of our Macaulay 2 program.) Then, we have
H 2 Ω • (log f ) C · 1 + C · (−x) + · · · + C · y 4 ω1 ∧ ω2 .
)) is spanned by 3 differential forms m1 dx + m2 dy where m1 , m2 are elements in A 2 , of

(3) H 1 (DR( M
which explicit expressions are a little lengthy. We solve the identities (7) and (8) to find c 1 and c 2 .
In other words, we need to compute preimages of m1 dx + m2 dy by τ 1 . As we explained, this can be
done by the method of undetermined coefficients degree by degree. We can find solutions when the
Table 1
p q Dimensions Timing in seconds

10 11 (8, 1, 1) 3.5
10 12 (9, 1, 1) 4.6
10 13 (10, 1, 1) 6.9
10 14 (11, 1, 1) 9.4
10 20 (17, 1, 1) 55.0
10 21 (18, 1, 1) 86.8
j
degree of c i , di , e with respect to x, y is 6 and that with respect to ∂x , ∂ y is 0. Here is a basis of the
3-dimensional vector space H 1 (Ω • (log f )) obtained by this method.
4
• − yxω1 − x2 ω2 ,
25

215 1101 367 2 43 2 367
• y− x− y ω1 + x − yx ω2 ,
28 280 56 35 350

11 28 13 14 4 112 2 56
• y− x− y3 − y2 + y ω1 + x2 + − y2 + y x+ y2 ω2 .
30 9 6 3 25 225 5 45
All programs and session logs to find this answer are obtainable from http://www.math.kobe-
u.ac.jp/OpenXM/Math/LogCohomology/readme.html. The logarithmic comparison theorem
does not hold for this example. In fact, the dimensions of the de Rham cohomology groups H i (Ω •f )
(i = 2, 1, 0) are 5, 3, 1 respectively.
Example 4.2. We apply a part of our algorithm to compute the dimensions of the cohomology groups
H i (Ω • (log f )) for f = x p + y q + xy q−1 . Here is Table 1 of p, q and the dimensions of H 2 , H 1 , H 0 and
timing data. The program is executed on a machine with 2G RAM and Pentium III (1G Hz).
The homogenization of f , f x , f y generates an ideal that is Cohen–Macaulay. These examples do

not need to call the subprocedure Quillen–Suslin. However, the logarithmic comparison theorem does
not hold for these examples (see [2]). Computation of de Rham cohomology groups is not feasible by
our implementation.
5. Another algorithm
In the previous sections, we have presented an algorithm of computing a basis of the logarithmic
cohomology groups for plane curves with respect to the canonical basis w 1 and w 2 . However, this
algorithm relies on algorithms for the Quillen–Suslin theorem and they are sometimes slow. We will
present another algorithm, which is independent of the Quillen–Suslin theorem, but it does not give
a basis with respect to the canonical basis.
Let f be a reduced polynomial in two variables. We denote by Ω kf the set of k-forms with co-
efficients in C[x, y , 1/ f ]. The k form ω ∈ Ω kf is called logarithmic k-form iff both of f ω and df ∧ ω
have polynomial coefficients. The space of logarithmic k-forms is denoted by Ω k (log f ). It is easy to
C[x, y ] dx∧dy
see that Ω 2 (log f ) = f
. Let us determine all the logarithmic 1-forms. Let ( p , q, r ) a triple of
polynomials such that
f y p − f x q + f r = 0 (syzygy equation). (9)
Note that (0, f , f x ), ( f , 0, − f y ), ( f x , f y , 0) are trivial solutions of the syzygy equation. For a solu-
p dx+q dy
tion ( p , q, r ) of the syzygy equation, ω= f
belongs to Ω 1 (log f ). Conversely, any logarithmic
1-form can be expressed in this way. In fact, the condition that df ∧ ω has a polynomial coefficient is
equivalent to that f y p − f x q is a multiple of f .
p dx+q dy dx∧dy
Put ω= f
. Let e (x, y ) be any polynomial. Then, d(e ω) = ( Le ) f
where
f y p − f xq
L = q ∂x − p ∂ y + q x − p y + .
f
We denote the Weyl algebra A 2 by D for simplicity in the sequel. Suppose that L i (i = 1, . . . , m),
stands for a set of generators
of the solution space of the syzygy equation, which is a C[x, y ]-module.
Then dΩ 1 (log f ) = y ] dx ∧ dy / f . Therefore, the computation of H 2 (Ω • (log f )) is nothing but
L i C[x,
m
the computation of C[x, y ]/ i =1 L i • C[x, y ] (see Oaku’s book [14] on the same question in the one
∗ ∗ ∗ }, which is a left D ideal. We denote by F the C-subvector
variable case). Put I = D · { L 1 , . . . , L m k
space of D ∩ C[x, y ] of which (1, 1, −1, −1)-order is less than or equal to k [20, pp. 14, 203]. Here,
(1, 1) is the weight for (x, y ) and (−1, −1) is that for (∂x , ∂ y ).
Algorithm 5.1. H 2 (Ω • (log f )).
Step 1. Find generators of the syzygy equation and obtain explicit expressions of L i .
Step 2. Compute a (1, 1, −1, −1)-Gröbner basis (standard basis) of I . We denote the elements of the
Gröbner basis by L i ∗ (renaming).
Step 3. Find the monic generator b(−∂x x − ∂ y y ) of in(1,1,−1,−1) ( I ) ∩ C[−∂x x − ∂ y y ].
Step 4. Let k0 be the maximal non-negative root of b(s) = 0. Then, return C-vector space basis {c i } of

F k0 / L i · F k0 −ord(1,1,−1,−1) ( L i ) .
i
{c i dx ∧ dy / f } is a basis of H 2 .
The steps 2–4 can also be done by computing D /( I ∗ + ∂x D + ∂ y D ) (0th integral module) where I ∗
is the formal adjoint of I . (As to details for the steps 2–4, see [16].)
(− log f ).
The left ideal generated by L ∗i is nothing but D · Der
Theorem 5.2. If dim V ( f , f x , f y ) 0, dim V ( f , f x ) 1, dim V ( f , f y ) 1, then Algorithm 5.1 is correct.
We note that when f is reduced, the assumption of the correctness holds.
Proof. Let I be the left ideal in D generated by L 1 , . . . , L m . We may assume that I contains f ∂x ,
f ∂ y and f y ∂x − f x ∂ y . Therefore, the characteristic variety of I is contained in V ( f (x, y )ξ, f (x, y )η,
f y (x, y )ξ − f x (x, y )η), of which dimension is less than or equal to 2 from the assumption. In
fact, assume (a, b) ∈ V ( f , f x , f y ). Then, ξ and η are free and then the dimension of the charac-
teristic variety is less than or equal to 2. Assume (a, b) ∈ V ( f , f x ) \ V ( f , f x , f y ). Then, we have
f (a, b) = 0, f x (a, b) = 0 and f y (a, b) = 0. Then, η is free and ξ = 0, then the dimension of the char-
acteristic variety is less than or equal to 2. The rest cases can be shown analogously. Therefore, D / I
is a holonomic D-module and hence a non-trivial b exists [20, Chapter 5, Theorem 5.1.2]. The rest
of the correctness proof is analogous with that of the 0th integration algorithm of D-modules [13],
[20, Chapter 5, Theorems 5.2.6 and 5.5.1]. 2
Example 5.3. For f = (x3 + y 4 + xy 3 )(x2 + y 2 ), we have dim H 2 (Ω • (log f )) = 7 with Algorithm 5.1.
The execution time is 1.9 s.
Let ω̃ , . . . , ω̃m be a set of generators of C[x, y ]-module Ω 1 (log f ). Define τ̃0 ( g ) = g f ,

m 1 dx∧dy
τ̃1 ( i=1 g i ω̃i ) = m i =1 ḡ i ⊗ f ω̃i , and τ̃2 ( g f
) = g ⊗ dx ∧ dy. Then, it is easy to show that τi = τ̃i
by expressing ω̃i in terms of the canonical basis ω1 , ω2 . Therefore, τ̃i gives the quasi-isomorphism of
Corollary 2.4. Hence, we have the following algorithm.
Algorithm 5.4.
Input: a non-zero reduced polynomial f .
Output: the dimension and bases of H i (Ω • (log f )), i = 0, 1.
Replace the free basis in step 1 of Algorithm 3.1 by a set of generators of the syzygy module and
perform the steps 2 and 3.
In step 4, use τ̃0 , τ̃1 to find preimages.
We close this paper with a remark for future study. Comparison theorems for logarithmic coho-
mology groups in the n variable case are studied under some conditions (see, e.g., [1,5] and their
references). It is an interesting problem to generalize our result to the n-variable case based on these
theorems and the algorithm given by Tsai and Walther [21].
Acknowledgments
The authors are grateful to M. Barakat and A. Fabianska for their help computing free bases of
syzygies by using her implementation for Quillen–Suslin’s theorem. The authors are also grateful to
L. Narváez-Macarro and the referees for their valuable comments and suggestions.
References
[1] F.J. Calderón-Moreno, Logarithmic differential operators and logarithmic de Rham complexes relative to a free divisor, Ann.
Sci. Ecole Norm. Sup. (4) 32 (1999) 701–714.
[2] F.J. Calderón-Moreno, D. Mond, L. Narváez-Macarro, F.J. Castro-Jiménez, Logarithmic cohomology of the complement of
a plane curve, Comment. Math. Helv. 77 (2002) 24–38.
[3] F.J. Castro-Jiménez, L. Narváez-Macarro, D. Mond, Cohomology of the complement of a free divisor, Trans. Amer. Math.
Soc. 348 (1996) 3037–3049.
[4] F.J. Castro-Jiménez, J.M. Ucha, Explicit comparison theorems for D -modules, J. Symbolic Comput. 32 (2001) 677–685.
[5] F.J. Castro-Jiménez, J.M. Ucha, Testing the logarithmic comparison theorem for free divisors, Experiment. Math. 13 (2004)
441–449.
[6] D. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry, Springer, New York, 1995.
[7] A. Fabianska, QuillenSuslin package, http://wwwb.math.rwth-aachen.de/QuillenSuslin/.
[8] A. Leykin, H. Tsai, D-module package for Macaulay2, 1999–2008, http://www.math.uiuc.edu/Macaulay2.
[9] A. Logar, B. Sturmfels, Algorithms for the Quillen–Suslin theorem, J. Algebra 145 (1992) 231–239.
[10] H. Matsumura, Commutative Ring Theory, Cambridge University Press, 1986.
[11] Z. Mebkhout, Le formalisme des six opérations de Grothendieck pour les D X -modules cohérents, Travaux en Cours, vol. 35,
Hermann, Paris, 1989.
[12] M. Noro, An efficient modular algorithm for computing the global B-function, in: A.M. Cohen, X.S. Gao, N. Takayama (Eds.),
Mathematical Software, Proceedings of the First International Congress of Mathematical Software, World Scientific, Beijing,
2002, pp. 147–157.
[13] T. Oaku, Algorithms for b-functions, restrictions, and algebraic local cohomology groups of D-modules, Adv. in Appl.
Math. 19 (1997) 61–105.
[14] T. Oaku, D-modules and Computational Mathematics, Asakura Publisher, 2002 (in Japanese).
[15] T. Oaku, N. Takayama, An algorithm for de Rham cohomology groups of the complement of an affine variety via D-module
computation, J. Pure Appl. Algebra 139 (1999) 201–233.
[16] T. Oaku, N. Takayama, H. Tsai, Polynomial and rational solutions of holonomic systems, J. Pure Appl. Algebra 164 (2001)
199–220.
[17] T. Oaku, N. Takayama, Algorithms for D-modules – restriction, tensor product, localization, and local cohomology groups,
J. Pure Appl. Algebra 156 (2001) 267–308.
[18] T. Oaku, N. Takayama, Minimal free resolutions of homogenized D-modules, J. Symbolic Comput. 32 (2001) 575–592.
[19] K. Saito, Theory of logarithmic differential forms and logarithmic vector fields, J. Fac. Sci. Tokyo Univ. Sec. IA 27 (1980)
265–291.
[20] M. Saito, B. Sturmfels, N. Takayama, Gröbner deformations of hypergeometric differential equations, in: Algorithms Comput.
Math., vol. 6, Springer-Verlag, Berlin, 2000, viii+254 pp.
[21] H. Tsai, U. Walther, Computing homomorphisms between holonomic D-modules, J. Symbolic Comput. 32 (2001) 597–617.
[22] U. Walther, Computing the cup product structure for complements of complex affine varieties, J. Pure Appl. Algebra 164
(2001) 247–273.
Journal of Algebra
A bound for orders in differential Nullstellensatz

Oleg Golubitsky a,1 , Marina Kondratieva b,2 , Alexey Ovchinnikov c,∗,3 ,
Agnes Szanto d,4
a
University of Western Ontario, Department of Computer Science, London, Ontario, Canada N6A 5B7
b
Moscow State University, Department of Mechanics and Mathematics, Leninskie gory, Moscow 119991, Russia
c
University of Illinois at Chicago, Department of Mathematics, Statistics, and Computer Science, 851 S. Morgan Street,
Chicago, IL 60607-7045, USA
d
North Carolina State University, Department of Mathematics, Raleigh, NC 27695-8205, USA
Article history: We give the first known bound for orders of differentiations
Received 3 March 2008 in differential Nullstellensatz for both partial and ordinary al-
Available online 27 June 2009 gebraic differential equations. This problem was previously ad-
Communicated by Reinhard Laubenbacher
dressed in [A. Seidenberg, An elimination theory for differential
algebra, Univ. of California Publ. in Math. III (2) (1956) 31–66]
Keywords:
Differential algebra but no complete solution was given. Our result is a complement
Characteristic sets to the corresponding result in algebraic geometry, which gives
Radical differential ideals a bound on degrees of polynomial coefficients in effective Null-
Differential Nullstellensatz stellensatz [G. Hermann, Die Frage der endlich vielen Schritte in
der Theorie der Polynomideale, Math. Ann. 95 (1) (1926) 736–788;
E.W. Mayr, A.W. Meyer, The complexity of the word problems for
commutative semigroups and polynomial ideals, Adv. Math. 46 (3)
(1982) 305–329; W.D. Brownawell, Bounds for the degrees in the
Nullstellensatz, Ann. of Math. 126 (3) (1987) 577–591; J. Kollár,
Sharp effective Nullstellensatz, J. Amer. Math. Soc. 1 (4) (1988)
963–975; L. Caniglia, A. Galligo, J. Heintz, Some new effectivity
bounds in computational geometry, in: Applied Algebra, Algebraic
Algorithms and Error-Correcting Codes, Rome, 1988, in: Lecture
Notes in Comput. Sci., vol. 357, Springer, Berlin, 1989, pp. 131–151;
N. Fitchas, A. Galligo, Nullstellensatz effectif et conjecture de Serre
(théorème de Quillen–Suslin) pour le calcul formel, Math. Nachr.
149 (1990) 231–253; T. Krick, L.M. Pardo, M. Sombra, Sharp es-
*
Corresponding author.
E-mail addresses: Oleg.Golubitsky@gmail.com (O. Golubitsky), kondratieva@sumail.ru (M. Kondratieva),
aiovchin@math.uic.edu (A. Ovchinnikov), aszanto@ncsu.edu (A. Szanto).
1
This author was partially supported by NSERC Grant PDF-301108-2004.
2
This author was partially supported by the Russian Foundation for Basic Research, project No. 08-01-90103-Mola .
3
This author was partially supported by NSF Grants CCF-0901175 and CCR-0096842.
4
This author was partially supported by NSF Grant CCR-0347506.
doi:10.1016/j.jalgebra.2009.05.032
O. Golubitsky et al. / Journal of Algebra 322 (2009) 3852–3877 3853
timates for the arithmetic Nullstellensatz, Duke Math. J. 109 (3)

(2001) 521–598; Z. Jelonek, On the effective Nullstellensatz, Invent.
Math. 162 (1) (2005) 1–17; T. Dubé, A combinatorial proof of the
effective Nullstellensatz, J. Symbolic Comput. 15 (3) (1993) 277–
296].
This paper is dedicated to the memory of Eugeny Pankratiev, who
was the advisor of the first three authors at Moscow State Univer-
sity.
1. Introduction
Given a system of algebraic partial differential equations F = 0, where F = f 1 , . . . , f k , and a dif-

ferential equation f = 0, one can effectively test if f is a differential algebraic consequence of F . In
this paper we develop a method that leads to an effective procedure which finds an algebraic expres-
sion of some power of f in terms of the elements of F and their derivatives (or shows that such
an expression does not exist). This procedure is called effective differential Nullstellensatz. A brute-force
algorithm solving this problem consists of two steps:
(1) find an upper bound h on the number of differentiations one needs to apply to F , and
(2) find an upper bound on the degrees of polynomial coefficients g i and a positive integer k
such that f k is a combination of the elements of F together with the derivatives up to the order h
and the coefficients g i . We solve the first problem in the paper. The second problem was addressed
and solved in [2] and further analyzed and improved in [3,4]. A purely algebraic solution was given
in [5]. A combinatorial approach via Hilbert polynomials was used in [10] (see also [11, Lecture XIII]).
Most of the references on the subject can be found in [6–9].
More precisely, our problem is as follows. We are given a finite set F of differential polynomials
such that a differential polynomial f belongs to the radical differential ideal generated by F in the
ring of differential polynomials. Knowing only the orders and degrees of the elements of F and the
order of f , we find a non-negative integer h such that f belongs to the radical of the algebraic ideal
generated by F and its derivatives up to the order h.
We give a complete solution to this problem using differential elimination. The problem is non-
trivial: the first attempt was made by Seidenberg [1], where it was conjectured that such a bound
would not be found. Here is where the main difficulty is coming from. In order to get the bound using
a differential elimination algorithm we need to estimate how many differentiation steps this algorithm
makes. Originally, termination proofs for such algorithms were based on the Ritt-Noetherianity of the
ring of differential polynomials, that is: every increasing chain of radical differential ideals terminates.
And this result does not say when the sequence terminates. We overcome this problem in our paper.
The article is organized as follows. We introduce basic notions of differential algebra in Section 2.
Then we formulate the main result, Theorem 1, in Section 3. In order to achieve this, we bound the
length of increasing sequences of radical differential ideals appearing in our differential elimination
in Section 5 (see Proposition 10). For that, in Section 4, we first bound the length of dicksonian
sequences of tuples of natural numbers with restricted growth of the maximal element in these
tuples (Lemma 8). Finally, we apply this to obtain the bound for the differential Nullstellensatz in
Theorem 15, from which Theorem 1 follows. We conclude by giving in Section 6 an alternative non-
constructive proof due to Michael Singer of existence of the bound, based on model theory.
There is some previous work on bounding orders in differential elimination algorithms. In the
ordinary case, we can bound the orders of derivatives of the output and all intermediate steps of
differential elimination [12], and this bound holds for any ranking. Also in the ordinary case, one
can give bounds for quantifier elimination [13] and for the orders and degrees of resolvents of prime
differential ideals of a certain type [14]. A related bound for involutive prolongation, based on the
analysis of stability of Spencer sequences, is obtained in [15].
3854 O. Golubitsky et al. / Journal of Algebra 322 (2009) 3852–3877
Note that, unlike the bounds for differential elimination mentioned above, the bound for the dif-
ferential Nullstellensatz proposed in this paper holds for the PDE case. Our bound is also based on
the analysis of differential elimination. But, due to the ranking-independent nature of the differential
Nullstellensatz, we could restrict our analysis to orderly rankings, which made it easier to treat not
only the ordinary case, but the PDE case as well.
2. Basic differential algebra
One can find recent tutorials on the constructive theory of differential ideals in [16–18]. One also
refers to [1,19–28] for differential elimination theory. A differential ring is a commutative ring with
unity endowed with a set of derivations Δ = {∂1 , . . . , ∂m }, which commute pairwise. The case of Δ =
{δ} is called ordinary. Construct the multiplicative monoid

km
Θ = ∂1k1 ∂2k2 · · · ∂m ki 0
of derivative operators. Let Y = { y 1 , . . . , yn } be a set whose elements are called differential indetermi-
nates. The elements of the set
Θ Y = {θ y | θ ∈ Θ, y ∈ Y }
are called derivatives. Derivative operators from Θ act on derivatives as θ1 (θ2 y i ) = (θ1 θ2 ) y i for all
θ1 , θ2 ∈ Θ and 1 i n.
The ring of differential polynomials in differential indeterminates Y over a differential field k is a
ring of commutative polynomials with coefficients in k in the infinite set of variables Θ Y . This ring
is denoted by
k{ y 1 , . . . , yn }.
We consider the case of char k = 0 only. An ideal I in k{ y 1 , . . . , yn } is called differential, if for all f ∈ I
and δ ∈ Δ, δ f ∈ I . Let F ⊂ k{ y 1 , . . . , yn } be a set of differential polynomials. For the algebraic ideal,
differential ideal, and radical differential ideal generated by F in k{ y 1 , . . . , yn }, we use notations ( F ),
[ F ], and { F }, respectively.
A ranking is a total order > on the set Θ Y satisfying the following conditions for all 1 = θ ∈ Θ
and u , v ∈ Θ Y :
(1) θ u > u ,
(2) u > v ⇒ θ u > θ v .
k k k
Let u be a derivative, that is, u = θ y j for θ = ∂1 1 δ22 · · · ∂mm ∈ Θ and 1 j n. The order of u is
defined as
ord u = ord θ = k1 + · · · + km .
If f is a differential polynomial, f ∈ / k, then ord f denotes the maximal order of derivatives appearing
effectively in f .
A ranking > is called orderly if ord u > ord v implies u > v for all derivatives u and v. Let a
ranking > be fixed. The derivative θ y j of the highest rank appearing in a differential polynomial
f ∈ k{ y 1 , . . . , yn } \ k is called the leader of f . We denote the leader by ld f or u f . Represent f as a
univariate polynomial in u f :
f = i f udf + a1 udf −1 + · · · + ad .
The monomial udf is called the rank of f and is denoted by rk f . Extend the ranking relation on
d d
derivatives to ranks: u 11 > u 22 if either u 1 > u 2 or u 1 = u 2 and d1 > d2 . The polynomial i f is called
the initial of f . Apply any derivation δ ∈ Δ to f :
∂f
δf = δ u f + δ i f udf + δa1 udf −1 + · · · + δad .
∂u f
The leader of δ f is δ u f and the initial of δ f is called the separant of f , denoted s f . If θ ∈ Θ \ {1},
then θ f is called a proper derivative of f . Note that the initial of any proper derivative of f is equal
to s f .
We say that a differential polynomial f is partially reduced w.r.t. g if no proper derivative of u g
appears in f . A differential polynomial f is algebraically reduced w.r.t. g if degu g f < degu g g. A dif-
ferential polynomial f is reduced w.r.t. a differential polynomial g if f is partially and algebraically
reduced w.r.t. g. Consider any subset A ⊂ k{ y 1 , . . . , yn } \ k. We say that A is autoreduced (respec-
tively, partially reduced, respectively, algebraically autoreduced) if each element of A is reduced
(respectively, partially reduced, respectively, algebraically reduced) w.r.t. all the others. A is called
weak d-triangular if no element of A belongs to k and the set of leaders of A is autoreduced. A weak
d-triangular set A is called d-triangular if A is partially autoreduced.
Every autoreduced set is finite [20, Chapter I, Section 9] (but an algebraically autoreduced set in a
ring of differential polynomials may be infinite). Every weak d-triangular set is finite, too [16, Propo-
sition 3.9]. For such sets we use capital calligraphic letters A, B , C , . . . and notation A = A 1 , . . . , A p
to specify the list of the elements of A arranged in order of increasing rank. We denote the sets of
initials and separants of elements of A by iA and sA , respectively. Let H A = iA ∪ sA . For a finite set
S of differential polynomials denote by S ∞ the multiplicative set containing 1 and generated by S.
Let I be an ideal in a commutative ring R. The saturated ideal I : S ∞ is defined as

a ∈ R ∃s ∈ S ∞ : sa ∈ I .
If I is a differential ideal then I : S ∞ is also a differential ideal (see [20]).

Let A = A 1 , . . . , A r and B = B 1 , . . . , B s be (algebraically) autoreduced (or weak d-triangular) sets.
We say that A has lower rank than B if
• there exists k min(r , s) such that rk A i = rk B i for 1 i < k, and rk A k < rk B k ,

• or if r > s and rk A i = rk B i for 1 i s.
We say that rk A = rk B if r = s and rk A i = rk B i for 1 i r. Let v be a derivative in k{ y 1 , . . . , yn }.

Denote by A v the set of the elements of A and their derivatives that have a leader ranking strictly
lower than v. A set A is called coherent if whenever A , B ∈ A are such that u A and u B have a
common derivative: v = ψ u A = φ u B , then
∞
s B ψ A − s A φ B ∈ (A v ) : H A .
3. Main result
For a finite set of differential polynomials F ⊂ k{ y 1 , . . . , yn } let D ( F ) be the maximal total degree
of a polynomial in F . For each i, 1 i n, let
h i ( F ) = ord y i ( F ), H ( F ) = max h i ( F ).
1i n
For h ∈ Z0 let F (h) denote the set of derivatives of the elements of F of order less than or equal
to h. The Ackermann function appearing in our main result is defined as follows [29, Section 2.5.5]:
A (0, n) = n + 1,
A (m + 1, 0) = A (m, 1),

A (m + 1, n + 1) = A m, A (m + 1, n) .
Theorem 1. Let F ⊂ k{ y 1 , . . . , yn } be a finite set, 0 = f ∈ { F } and let t ( F , f ) be the minimal non-negative

integer such that

f ∈ F (t ( F , f )) .
Then

t ( F , f ) A m + 8, max n, H ( F ∪ f ), D ( F ∪ f ) .
Proof. This result will be proved step-by-step in the following sections as described in the introduc-
tion and finally established in Theorem 15. 2
Remark 2. It is our own choice here to bound t ( F , f ) using solely the maximal orders and degrees of
F and f . One might come up with another bound using more information of F and f . But we em-
phasize that the bound on orders must depend on the degrees, number of differential indeterminates,
and number of basic differentiations as the following examples show.
Example 3. Let f = 1 and F = { y

− 1, yk } in k{ y }, the ordinary case. In order to express 1 in terms
of the elements of F , one has to differentiate yk k times.
In the linear and non-linear cases, consider the following examples showing that the bound must
depend on the number of variables and derivations.
Example 4. Let f = 1 and F = { y

1 , y 1 − y
2 , . . . , yn−1 − yn
, yn − a} in the ordinary differential ring
k{ y 1 , . . . , yn }, where a ∈ k is such that a(n) = 1. We have to differentiate the first n − 1 generators n
(n)
times to get yn into the corresponding algebraic ideal. Hence, t ( F , f ) = n.
Example 5. Let f = 1 and F = { y 21 , y 1 − y 22 , . . . , yn−1 − yn2 , 1 − yn

} ⊂ k{ y 1 , . . . , yn }, again in the ordi-
nary case. One can show that

F ⊂ y 1 , y 2 , . . . , yn , 1 − yn
= I 0 ,

F (1) ⊂ I 0 , y
1 , y
2 , . . . , yn
−1 , yn
= I 1 ,
(3 )
F (2) ⊂ I 1 , y
1 , . . . , yn
−2 , yn
−1 − 2, yn = I2,
(4 )
F (3) ⊂ I 2 , y
1 , . . . , y n −2 , y n −1 , y n = I3,

(4 ) (4 ) 4 (4 ) (5 )
F (4) ⊂ I 3 , y 1 , . . . , y n −2 − 22 , y n −1 , y n = I4,
2
..
.

−1
n 2n−k−1
(2n−1 ) (2n−1 ) 2k (2n−1 ) (2n−1 +1)
F ⊂ I 2n−1 −1 , y 1 − , y2 , . . . , yn = I 2n−1 ,
2k−1
k =1
..
.
(2n −1) (2n −1) (2n )
F (2 −1) ⊂ I 2n −2 , y 1
n
, y2 , . . . , yn .
/ ( F (2 −1) ). Thus, t ( F , f ) = 2n , because modulo 1 − yn

we have
n
Therefore, 1 ∈
(2n ) n (2n −1) n (2n −1)
= 2n yn2 −1 yn
≡ 2n yn2 −1 ≡ · · · ≡ 2n ! yn yn
n
yn2
≡ 2n ! yn
2 ≡ 1.
Example 6. If we replace F in the previous example by

G = u 2x1 , u x1 − u 2x2 , . . . , u xm−1 − u 2xm , 1 − u x2 ⊂ k{u }
m
with partial derivatives ∂x1 , . . . , ∂xm , we obtain an example which shows that the bound on orders
must depend on the number m of derivations. Again, the generators will have to be differentiated 2m
times to express 1.
4. Bounds on lengths of sequences
The results of this section will be further used in Section 5 to bound lengths of decreasing se-
quences of autoreduced sets appearing in the differential elimination algorithm that we use. In this
section the letters m and n will not mean the number of derivations and differential indeterminates,
respectively.
We begin by bounding the length of certain sequences of non-negative n-tuples. Call a sequence
t 1 , t 2 , . . . , tk
of n-tuples dicksonian if for all 1 i < j k there does not exist a non-negative n-tuple t such that
ti + t = t j .
For example, any lexicographically decreasing sequence is dicksonian. By Dickson’s lemma, every dick-
sonian sequence is finite. Our goal is to obtain an explicit upper bound for the length of a dicksonian
sequence, whose elements do not grow faster than a given function, in terms of this function, the first
element, and the size n of the tuples. Let

a11 , . . . , an1 , a21 , . . . , an2 , . . . , ak1 , . . . , akn
be a dicksonian sequence of n-tuples of non-negative integers such that
j j
max a1 , . . . , an f ( j ) (1)
for all j, 1 j k, where
f : Z0 → Z0
is a fixed function. We say that the growth of this sequence is bounded by the function f .
The following proposition closely resembles a particular case of our problem, namely that of
f ( i ) = m + i − 1.
However, in Proposition 7 the maximal coordinate must increase by 1 at each step, whereas in our
case it is allowed to decrease or remain the same. We will reduce the case of a dicksonian sequence
with the growth bounded by a function f from a certain large class of functions that “do not grow
too fast”, to the one treated in Proposition 7.
Proposition 7. (See [30, Proposition 1].) Let t 1 , t 2 , . . . , tk be a dicksonian sequence of n-tuples, such that the
maximal coordinate of t i equals m + i − 1, for all 1 i k. Then the maximal coordinate in the last tuple, tk ,
does not exceed
A (n, m − 1) − 1,
and there exists such a dicksonian sequence for which this bound is reached.
Note that in Proposition 7 we have: m is the maximal coordinate of t 1 and the length k of the
sequence is bounded by
A (n, m − 1) − m.
The general case (of any function f , not necessarily from our class) has also been studied in [31]
using a different approach. It is shown that the maximal possible length is primitive recursive in f
and recursive, but not primitive recursive (if f increases at least linearly), in n. Sequences yielding
the maximal possible length are constructed. Moreover, if f is linear, an explicit expression for the
maximal length is given in terms of a generalized Ackermann function. Our statement was motivated
by the need to obtain an explicit expression for the bound for a wider class of growth functions.
Let L f ,n denote the maximal length of a dicksonian sequence of n-tuples, whose growth is bounded
by f . For an increasing function f : Z0 → Z0 , let f −1 (x) be the least number k such that
f (k) x.
Lemma 8. Let f : Z0 → Z0 be an increasing function, and d ∈ Z0 be a number such that f (i + 1) −
f (i ) A (d, f (i ) − 1) for all i > 0. Then

L f ,n < f −1 A n + d, f (1) − 1 (2)
and the maximal entry of the last n-tuple does not exceed

A n + d, f (1) − 1 .
Proof. Consider a dicksonian sequence

a11 , . . . , an1 , a21 , . . . , an2 , . . . , ak1 , . . . , akn , (3)
whose growth is bounded by f . Construct from (3) a new sequence satisfying the conditions of Propo-
sition 7. Append to the first tuple d new coordinates, each equal to f (1), obtaining the following
(n + d)-tuple:

a11 , . . . , an1 , f (1), . . . , f (1) .
Then add f (2) − f (1) − 1 new (n + d)-tuples. The first n coordinates of these tuples are (a11 , . . . , an1 ).
The last d coordinates form a dicksonian sequence of d-tuples, starting with ( f (1), . . . , f (1)), with
the maximum coordinate growing exactly by 1 at each step. From Proposition 7 and condition

f (2) − f (1) A d, f (1) − 1 ,
such sequence exists. The last tuple will have the maximum coordinate equal to f (2) − 1. Next, add
the tuple

a21 , . . . , an2 , f (2), . . . , f (2) .
Since the growth of (3) is bounded by f , the maximal coordinate in this tuple equals f (2). Continue
by adding f (3) − f (2) − 1 new (n + d)-tuples, whose first n coordinates are (a21 , . . . , an2 ) and last d
coordinates form a dicksonian sequence growing by 1 at each step. Finally, when the tuple

ak1 , . . . , akn , f (k), . . . , f (k)
is reached, stop. We obtain a sequence of (n + d)-tuples in which the maximal coordinate grows by 1
at each step. We will show that this sequence is dicksonian. Suppose that it is not. Let t j , tl , j < l, be
two (n + d)-tuples from this sequence, for which there exists a tuple t such that
tl = t j + t .
Let t I , t II denote the first n coordinates and the last d coordinates of an (n + d)-tuple t, respectively.
Then we have
tlI = t Ij + t I and tlII = t IIj + t II .
If t j and tl have been added after the same tuple of the form

p i = a1i , . . . , ani , f (i ), . . . , f (i ) ,
or if t j coincides with such a tuple p i and tl has been added after p i , the equality
tlII = t IIj + t II
contradicts the fact that the last d coordinates of the tuples between p i and p i +1 , including p i and
excluding p i +1 , form a dicksonian sequence. If t j and tl have been added after different tuples p i and
p i
, the equality
tlI = t Ij + t I
contradicts the fact that sequence (3) is dicksonian. Therefore, our assumption was false and the
constructed sequence is dicksonian.
By Proposition 7, the maximum coordinate of its last element does not exceed
A (n + d, m − 1) − 1.
Since the maximum coordinate in the first element is f (1) and grows by 1 at each step, the number
of elements in the constructed sequence does not exceed

A n + d, f (1) − 1 − f (1).
On the other hand, the number of elements in the constructed sequence is:
f (2) − f (1) + f (3) − f (2) + · · · + f (k) − f (k − 1) + 1 = f (k) − f (1) + 1.
Therefore,

f (k) − f (1) + 1 A n + d, f (1) − 1 − f (1),
that is,

f (k) < A n + d, f (1) − 1 ,
and

k < f −1 A n + d, f (1) − 1 . 2
5. Differential elimination algorithm
Using the result of the previous section, we obtain an upper bound for the length of sequences
of d-triangular sets of decreasing rank produced by a differential elimination algorithm. The idea is
to put in correspondence with such a sequence a dicksonian sequence of tuples, whose growth is
bounded by a function derived from the algorithm.
We fix an orderly ranking. Algorithm 1 computes a characteristic decomposition of a radical dif-
ferential ideal given by a set of generators (correctness and termination are proven in Proposition 10).
It is designed in such a way that allows us to control:
• the orders and degrees of differential polynomials occurring in all intermediate steps together
with
• a bound on the number of iterations of this algorithm.
Note that the sets in this characteristic decomposition might not be autoreduced, but they are weak
d-triangular (as in [16]), that is, the set of leaders autoreduced. Also, in the algorithm:
• the procedure algrem computes an algebraic pseudo-remainder of a polynomial with respect to

an algebraic triangular set (a set is called triangular if the leaders of its elements are distinct).
• Algorithm MinimalTriangularSubset inputs a finite set of differential polynomials and outputs one
of its least rank triangular subsets.
Denote by Δ(C ) the set of “differential S-polynomials” of C defined in [16, Definition 4.2]. From now
on, m and n again denote the numbers of derivations and differential indeterminates, respectively.
Here is the main idea behind the algorithm. The proof of correctness can be found in Proposition 10:
(1) At each step of the algorithm, given ( F , C ), the set C¯ is formed from C by adding a new poly-
nomial f ∈ F and removing everything from C with leaders that are derivatives of ld f : these
elements form the set D. As a result, the set C¯ satisfies the property that the set of its leaders is
differentially autoreduced.
(2) We define the set G comprised of:
(a) the remaining elements of F ,
(b) the cross-derivatives (differential S-polynomials) of C¯,
(c) and the elements of D.
(3) We do the reductions in a way that allows us to control the orders of the remainders:
(a) first differentiate the elements of C¯ to the maximal necessary order b, obtaining a triangular
set B ,
(b) then autoreduce B algebraically,
(c) if the algebraic autoreduction changes the leaders of the polynomials in B , then B must be
inconsistent,
(d) if it does not change the leaders, we reduce elements of G algebraically with respect to alge-
braically autoreduced set B ,
(e) if all remainders are zeroes, we stop, otherwise we append the new remainders to the current
system F and continue to the next iteration of the while-loop.
Algorithm 1. RGBound( F 1 )
Input: a set F 1 ⊂ k{ y 1 , . . . , yn } with derivations {∂1 , . . .
, ∂m }
∞;
Output: A finite set T of triangular sets such that { F 1 } ⊂ C ∈ T {C } : H C
/ [C ] : H C∞ then C is coherent and weak d-triangular
if 1 ∈
otherwise 1 ∈ (C ) : H C ∞.
T := {∅}; U := {( F 1 , ∅)}
while U = ∅ do
Take and remove any ( F , C ) ∈ U
f := an element of F reduced w.r.t. C of the least rank
/ k then U := U ∪ {( F ∪ s f , C )} end if
if s f ∈
/ k then U := U ∪ {( F ∪ i f , C )} end if
if i f ∈
D := {C ∈ C | ld C = θ ld f for some θ ∈ Θ}
C¯ := (C \ D ) ∪ { f }
G := F ∪ Δ(C¯) ∪ D \ { f }
b := max g ∈G ord g
B := MinimalTriangularSubset({θ C | C ∈ C¯, ord θ C b})
B̄ := {algrem(h, B \ {h}) | h ∈ B}
if rk B̄ = rk B then T := T ∪ {B } else
R := {algrem( g , B ) | g ∈ G } \ {0}
if R = ∅ then U := U ∪ {( R ∪ F , C¯)} else T := T ∪ {C¯} end if
end if
end while
return T
We get the following bounds for the growth of the maximal degrees of the polynomials computed
at the ith iteration of Algorithm 1.
Proposition 9. Fix ( F i , Ci ) = ∅ ∈ U and let ( F i +1 , Ci +1 ) = ∅ be any of the elements obtained from ( F i , Ci )

after one iteration of the while-loop. We then have
(2H ( F i ∪mCi )+m)+1

D ( F i +1 ∪ Ci +1 ) 4D ( F i ∪ Ci ) .
Proof. Consider an iteration of the loop. In the first six lines of the loop the degrees do not change,
since adding an initial or a separant does not cause an increase in the degrees. So,
(1) the first place where the degrees may change is the computation of Δ(C¯), that is, computing
cross-derivatives. This at most doubles the degrees.
(2) After that, the only places where the degrees of polynomials may increase are calls to
algrem( g , B ) for g ∈ G or algrem(h, B \ {h}) for h ∈ B .
In both cases it is a sequence of at most |B | algebraic pseudodivisions. We will prove the bound for
the reduction of a fixed g ∈ G modulo B , and the other case follows similarly.
Assume that |B | = N, and let B = { B 1 , . . . , B N } be ordered so that
ld( B 1 ) > ld( B 2 ) > · · · > ld( B N ).
Let g (0) := g and

g (t ) := algrem g , { B 1 , . . . , B t }
for t > 0. Denote the maximal total degree of g (t ) ∪ B by δ(t ), t 0. Note that
1 δ(0) 2D ( F i ∪ Ci ).
Then g (t +1) is obtained by the pseudo-division of g (t ) with respect to the polynomial B t +1 . Thus,
g (t +1) = i B t +1 g (t ) − qB t +1 ,
where is a sufficiently large exponent specified below, q is the pseudo-quotient, and the degree of
g (t +1) in ld( B t +1 ) is smaller than the same degree of B t +1 . The exponent is bounded by

degld( B t +1 ) g (t ) − degld( B t +1 ) ( B t +1 ) + 1 δ(t ).
Therefore, if δ(t ) 2 the total degree of g (t +1) is bounded by
δ(t + 1) δ(t )δ(0) + δ(t ) + δ(0) 2δ(0)δ(t ),
because in this case we necessarily have δ(0) 2. If δ(t ) = 1 then B consists of linear polynomials
and, moreover, the polynomial g (t ) is linear as all its remainders are, that is, δ(t + 1) 1 . These two
cases imply that
N +1
δ( N ) 2δ(0) .
Using that

bi + m
δ(0) 2D ( F i ∪ Ci ), N , and b i 2H ( F i ∪ Ci ),
bi
we obtain the claim (the numbers b i correspond to the number b computed at the ith iteration of
the loop in Algorithm 1). 2
5.1. Differential bounds for splitting
Consider now the splitting part of differential elimination. Algorithm 1 removes an element ( F , C )
from U and within one iteration of the while-loop converts this element into zero, one, two, or three
elements. Moreover, if the set C does not change, the orders and degrees of the elements of F do not
increase after the conversion.
Let U be the set of all triples ( F , C , k), where ( F , C ) is a pair removed from the set U at the be-
ginning of the kth iteration of the while-loop (the iterations are numbered starting from 1). Introduce
a partial order on U : let
( F 1 , C1 , k 1 ) 0 ( F 2 , C2 , k 2 )
if and only if k1 < k2 , ( F 2 , C2 ) was added to U at iteration k1 and was not added to U at any iter-
ation k, k1 < k < k2 . Let the partial order on U be the transitive closure of the relation 0 . Each
triple u ∈ U , except u = ( F 1 , ∅, 1), has a unique direct predecessor u − 0 u and up to three direct
successors.
The relation satisfies the following first property: if
( F 1 , C1 , k 1 ) ( F 2 , C2 , k 2 ) (4)
and a polynomial f is reduced with respect to C2 , it is also reduced with respect to C1 . This is the
case because of the inclusion
Θ rk C2 ⊇ Θ rk C1 (5)
that holds by construction of C2 from C1 .

The second property of the relation is that, whenever (4) holds and
C 1 = C2 ,
for the elements g 1 and g 2 drawn respectively from F 1 and F 2 at iterations k1 and k2 , we have
deg g 1 > deg g 2 .
This property holds when ( F 1 , C1 , k1 ) is a direct predecessor of ( F 2 , C2 , k2 ), because in this case

C1 = C2 implies that g 2 is the initial or separant of g 1 ; by transitivity, it also holds in general for
any triples satisfying (4).
Consider Algorithm 1 with input F 1 and the corresponding partially ordered set U . Denote the
maximal length of a decreasing chain in U by L ( F 1 ).
Proposition 10. Algorithm 1 is correct and terminates. Moreover,

L ( F 1 ) log2 A m + 7, Q ( F 1 ) − 1 ,
where

Q ( F 1 ) = max 9, n, 29H ( F 1 ) , D ( F 1 ) . (6)
Proof. To demonstrate correctness we need to prove that for all A ∈ T , if 1 ∈ ∞ then A is

/ [ A] : H A
coherent, weak d-triangular, and
∞
{ F 1 } ⊆ { A} : H A .
Note that in the while loop of the algorithm there are two conditions used to add a set A to T .
The first is when rk A = rk Ā, where Ā is the set obtained by taking the algebraic remainders of the
elements of A by the rest of the elements. In this case we have 1 ∈ (A) : H A∞ by [12, Lemma 5].
The second case is when the while loop is entered with a pair ( F , C ) ∈ U and we add A := C¯ to T
if all elements of a set G, which contains F , reduce to zero by some derivatives of the elements of C¯.
Here C¯ is obtained from C by adding an element f to C and removing all elements from C with
leaders that are derivatives of the leader of f , so C¯ remains d-triangular if C is such. Since R = ∅,
the set Δ(C¯) is reducible to zero with respect to C¯. Therefore, C¯ is coherent by [16, Proposition 4.4].
To see the last claim, note that F 1 ⊂ F for all ( F , C ) ∈ U , so also F 1 ⊂ G, and, thus,
{ F 1 } ⊂ {C¯} : H C̄∞ = {A} : H A

∞
.
Termination of the algorithm and the bound for L ( F 1 ) will be proved as follows. To each element
( F , C , k) of U we associate an (m + 4)-tuple in such a way that any sequence of these tuples, corre-
sponding to a decreasing sequence of elements of U w.r.t. , is dicksonian. Note that by definition
L ( F 1 ) is the maximal length of such sequence. For the initial triple ( F 1 , ∅, 1), let the corresponding
(m + 4)-tuple be
(0, . . . , 0, n, 0).
For any other triple ( F , C , k) ∈ U , consider its direct predecessor ( F − , C− , k− ), that is, ( F − , C− , k− ) 0
( F , C , k). Let f be the element that was drawn from F − at the beginning of the k− th iteration. Let
i i d
rk f = ∂11 . . . ∂mm y j .
Let also g be the element drawn from F at the beginning of the kth iteration. If
F = F − ∪ {i f } or F = F − ∪ {s f },
this implies that g = i f or g = s f , respectively, and C = C− . Then the (m + 4)-tuple associated to

( F , C , k) is

a1 , . . . , am+3 , deg( g ) ,
where (a1 , . . . , am+4 ) is the (m + 4)-tuple associated to ( F − , C− , k− ). Otherwise (when f is added to

C− ), the (m + 4)-tuple associated to ( F , C , k) is

τ = i 1 , . . . , im , d, j , n − j , deg( g ) .
We need to show that, for any predecessor ( F̂ , Cˆ, k̂) of ( F , C , k), at least one coordinate of the tuple
τ̂ associated to ( F̂ , Cˆ, k̂) is greater than the corresponding coordinate of the tuple τ associated to
( F , C , k ).
If Cˆ = C , the statement holds because of the second property of the relation . Furthermore, if
Cˆ = ∅, the (m + 3)rd coordinate of τ̂ equals n, which is greater than n − j, the value of the (m + 3)rd
coordinate in τ (since j ∈ {1, . . . , n}). The non-trivial case of Cˆ = C and Cˆ = ∅ is considered in detail
below. Let

F̂ ∗ , Cˆ∗ , k̂∗
be the most recent predecessor of ( F̂ , Cˆ, k̂) with Cˆ∗ = Cˆ. Such predecessor exists, because Cˆ = ∅. Let
f̂ ∗ be the element drawn from F̂ ∗ at iteration k̂∗ . Because ( F̂ ∗ , Cˆ∗ , k̂∗ ) is the most recent predecessor
of ( F̂ , Cˆ, k̂) with Cˆ∗ = Cˆ, the element f̂ ∗ must have been added to Cˆ∗ at iteration k̂∗ to form C¯, from
which then the set A = Cˆ has been computed. Therefore,
rk f̂ ∗ ∈ rk Cˆ. (7)
Let
î d̂
rk f̂ ∗ = ∂11 . . . ∂mm y ĵ .
î
Then the tuple τ̂ associated to ( F̂ , Cˆ, k̂) is
τ̂ = (î 1 , . . . , îm , d̂, ĵ , n − ĵ , ρ̂ ),
where the last element ρ̂ is unimportant to us at the moment.

Similarly, let ( F ∗ , C ∗ , k∗ ) be the most recent predecessor of ( F , C , k) with C ∗ = C . Let f ∗ be the
element drawn from F ∗ at iteration k∗ . Let
i d
rk f ∗ = ∂11 . . . ∂mm y j .
i
Then the tuple τ associated to ( F , C , k) is
τ = (i 1 , . . . , im , d, j , n − j , ρ ),
where the last element ρ is again unimportant.

Recall that we need to show that at least one coordinate in τ̂ is greater than the corresponding
coordinate in τ . Observe first that if ĵ = j then either ĵ > j, or n − ĵ > n − j. Therefore, we may
assume that
ĵ = j .
By construction, f ∗ is reduced w.r.t. C ∗ . By the first property of the relation , it must be reduced
w.r.t. Cˆ as well. But then, given (7), we obtain that f ∗ is reduced w.r.t. f̂ ∗ . In particular, this im-
plies that rk f ∗ is reduced w.r.t. rk f̂ ∗ , which cannot be the case unless at least one of the following
inequalities holds:
î 1 > i 1 , . . . , îm > im , d̂ > d.
This concludes the proof of termination.

Note that if we remove n − j from the τ ’s, the sequence might not be dicksonian. Indeed, let
m = 1, F = { y 1 , y 2 } ⊂ k{ y 1 , y 2 , y 3 } with y 1 < y 2 < y 3 . We then would have τ1 = (0, 1, 1, 1) and
τ2 = (0, 1, 2, 1).
We now switch to proving the bound. Let H 1 = max( H ( F 1 ), m), H k+1 = 2H k , D 1 = D ( F 1 ), and
2H k +m
D k+1 = (4D k )( m )+1 .
By Proposition 9, the maximal coordinate of the (m + 4)-tuple τk does not exceed max( H k , D k , n). Let
√
3 uk (2+log2 uk )
u 1 = max n, 9, 29H ( F ) , D ( F ) , u k +1 = 2 .
Then the maximal coordinate of τk does not exceed uk for all k 1. Indeed, we will prove by induc-
tion that H k 19 log2 uk and D k uk . For k = 1 these inequalities hold by definition of u 1 . Assuming
that they hold for H k , D k , and uk , prove them for H k+1 , D k+1 , uk+1 . Since H k+1 = 2H k , we have
2 √
29H k+1 = 29·2H k = 29H k
3
uk2 = 22 log2 uk 2 uk log2 uk
< u k +1 ,
because uk 9. Next,

2H k + m
log2 D k+1 = + 1 (2 + log2 D k ) 22H k +m (2 + log2 uk )
m
3H k √
2 (2 + log2 uk ) 3
uk (2 + log2 uk ) = log2 uk+1 .
1
Here we used the fact that H k m, as well as the inequality H k 9
log2 uk proven above. Now
observe that for x 9, we have
√
3
x(2+log2 x)
2 2x+2 − 3 = A (3, x − 1).
Therefore, the sequence of (m + 4)-tuples τ0 , τ1 , τ2 , . . . satisfies the conditions of Lemma 8 with d = 3.

And, according to this lemma, the length of this sequence does not exceed

f −1 A m + 7, f (1) − 1 ,
where f (k) = uk 2k . We can now plug in f (1) = u 1 , and replace f −1 with log2 . 2
Corollary 11. The maximal orders and degrees of polynomials computed by Algorithm 1 do not exceed

A m + 7, Q ( F ) − 1 . (8)
Proof. Follows directly from Lemma 8 and Proposition 10. 2
5.2. Lifting the final bound from splitting
Note that for f ∈ k{ y 1 , . . . , yn } and F ⊂ k{ y 1 , . . . , yn } we have
1 ∈ {F } : f ⇔ f ∈ { F }.
Lemma 12. Let C 1 , . . . , C p = C ⊂ [ F ] be a coherent weak d-triangular set that reduces all elements of F to
zero, and let the ranking be orderly. Then for all f ∈ { F } we have

1 ∈ C (q) : H C∞ ∪ f ,
where q = max(ord f , b) − min g ∈C ord g with b = max g ∈C ord g.
Proof. We will first construct the sets

B = MinimalTriangularSubset {θ C | C ∈ C , ord θ C b}
and

B̄ = algrem h, B \ {h} h ∈ B .
If rk B = rk B̄ then
∞
1 ∈ (B ) : H B = (B) : H C∞
by [12, Lemma 5] and our lemma is proven.

Suppose that rk B = rk B̄ and take the sets C and H = H C as the input of [16, Algorithm 6.8]. We
get a differential regular system (A, H A ) as the output and A = A 1 , . . . , A p = ∅. This means that A
is a coherent d-triangular set such that
[C ] : H C∞ = [A] : H A
∞
.
Note that, since we have rk B̄ = rk B and the ranking is orderly, the elements of A can be constructed
as algrem(C , B0 ) for some C ∈ C and B0 ⊂ B .
∞ is radical and, therefore,
It follows from [16, Theorem 4.12] that the ideal [A] : H A
f ∈ { F } ⊂ {C } : H C∞ = {A} : H A
∞ ∞
= [A] : H A .
Let g be a partial pseudo-remainder of f with respect to A. Then, by the Rosenfeld lemma [16,
Theorem 4.8], we have
∞
g ∈ (A) : H A . (9)
Let
qi = b − ord Ci = b − ord Ai , 1 i p,
and
q0 = max(0, ord f − b).
∞ such that
Since the ranking is orderly, there exists h ∈ H A
(max(0,ord f −ord A 1 )) (max(0,ord f −ord A p ))

h · f − g ∈ A1 , . . . , Ap . (10)
Indeed, at each step of the partial pseudo-division the order of the resulting differential polynomial is
less than or equal to the one of the previous polynomial and ord f − ord A i represents the maximal
number of times one possibly needs to differentiate A i to perform one step of the partial pseudo-
reduction with respect to A i . For any i, 1 i p, we have
max(0, ord f − ord A i ) max(ord f , b) − ord A i
= b − ord A i + max(b, ord f ) − b = qi + q0 .
Hence, (9) and (10) imply that
(q1 +q0 ) (q p +q0 ) ∞

1 ∈ A1 , . . . , Ap : HA ∪f .
On the other hand,
(q1 +q0 ) (q p +q0 ) ∞

(q1 +q0 ) (q +q )
A1 , . . . , Ap : f ∪ HA ⊂ C1 , . . . , C p p 0 : f ∪ H C∞ .
Indeed, if 1 i p, by induction on r q i we shall prove that
(r ) (q ) (q )

Ai ⊂ C 1 1 , . . . , C p p : H C∞ . (11)
To prove (11) we note that for any A i ∈ A there exists the following representation

h Ai = βk θkj C j
{k, j | ord θkj ord C i −ord C j }
∞ . Then, after differentiating the above expression r times we obtain

with h ∈ H C
(r ) (ord C i −ord C 1 +r ) (ord C i −ord C p +r )

Ai ⊂ C1 ,..., Cp : H C∞ .
Note that
ord C i − ord C j + r ord C i − ord C j + b − ord C i = q j ,
and we have (11).

Now, by [12, Lemma 5] for the polynomials C i and A i , since rk C i = rk A i , we have
(q1 ) (q p )
s A i h i − sC i ∈ C 1 ,..., Cp (12)
and
(q1 ) (q p )
i A i h i − iC i ∈ C 1 ,..., Cp (13)
∞ . Let
for some h i ∈ H C
(q1 +q0 ) (q p +q0 ) ∞

a ∈ A1 , . . . , Ap : f ∪ HA .
This means that

(q1 +q0 ) (q p +q0 )
af h
∈ A 1 , . . . , Ap (14)
for
jp k kp
h
= s A1 · . . . · s A p i A1 · . . . · i A p .
j
1 1
Multiplying inclusions (12) and (13) by some powers of sC i and iC i and (14) by h i , we get
(q1 +q0 ) (q p +q0 )

af h
∈ C 1 ,..., Cp ,
∞ . Hence,
where h
∈ H C
(q1 +q0 ) (q p +q0 ) ∞

(q1 +q0 ) (q +q )
1 ∈ A1 , . . . , Ap : f ∪ HA ⊂ C1 , . . . , C p p 0 : f ∪ H C∞ .
Since
qi + q0 = b − ord C i + max(b, ord f ) − b max(b, ord f ) − min ord g = q,

g ∈C
the lemma is proven. 2
Lemma 13. Let F ⊂ k{ y 1 , . . . , yn } and f , f 1 , . . . , f k ∈ k{ y 1 , . . . , yn }. Suppose for some d ∈ Z1 we have

( f 1 · . . . · f k )d ∈ ( F ) : f . Then

(k+1) H +1 d)
θ1 f 1 · . . . · θk f k ∈ F (4 : f,
where θi ∈ Θ with ord θi H for all i, 1 i k.
Proof. It follows from the proof of [32, Lemma 1.7] that if ad ∈ ( F ) then

(∂i a)2d−1 ∈ F (d)
for any a ∈ k{ y 1 , . . . , yn } and ∂i ∈ Δ. Therefore,
2d−1 (d) ∞
(∂i f 1 ) · f 2 · . . . · f k + · · · + f 1 · . . . · f k−1 · (∂i f k ) ∈ F :f .
Multiplying by ((∂i f 1 ) · f 2 · . . . · f k )2d−1 , we obtain
2(2d−1)
(∂i f 1 ) · f 2 · . . . · f k ∈ F (d) : f ∞ .
To make the computation simpler, we have
4d
(∂i f 1 ) · f 2 · . . . · f k ∈ F (d) : f ∞ .
By induction, we conclude that
4 H d H H +1
(θ1 f 1 ) · f 2 · . . . · f k ∈ F (d(1+4+···+4 )) : f ∞ ⊂ F (4 d) : f ∞ .
Similarly, we obtain
42H d H +1 H +1 H
(θ1 f 1 ) · (θ2 f 2 ) · . . . · f k ∈ F (4 d+4 4 d) : f ∞
H +1 H
= F (4 d(1+4 )) : f ∞ .
Finally, by induction we get
4kH d (4 H +1 d(1+4 H +···+4(k−1) H )) ∞

(θ1 f 1 ) · (θ2 f 2 ) · . . . · (θk f k ) ∈ F :f
(4(k+1) H +1 d) ∞
⊂ F :f ,
which finishes the proof. 2
Let ( F , C ) ∈ U . We call a set A ∈ T a general component of ( F , C ) if
[A] ⊂ [ F , C ]. (15)
By the termination of the algorithm, the general component A of ( F , C ) always exists and
∞
[A] ⊂ { F , C } ⊂ { A } : H A .
In particular, there exists a general component of F 1 , that is, of ( F 1 , ∅).

Recall that t ( F , f ) is the minimal non-negative integer such that

f ∈ F (t ( F , f )) .
Lemma 14. For a finite subset F ⊂ k{ y 1 , . . . , yn } we have:
H ( F )·2 L ( F ) +1
t ( F , f ) ord f + H ( F ) · 2 L ( F ) + 4(n·2 +1)·t (G , f )+1
d, (16)
where
n·2 H ( F )·2L( F ) +m+ord f

d := max D ( f ), A m + 7, Q ( F ) − 1 , (17)
G ⊂ k{ y 1 , . . . , yn } is such that
L (G ) L ( F ) − 1 and H (G ) H ( F ) · 2 L ( F )
and Q ( F ) is defined in formula (6).
Proof. Let A be any general component computed by Algorithm 1 (see (15)) and p be the number of
elements of A. Note that
p n · 2 H (A)+m . (18)
We then have

p

p
∞
{ F } = [A] : H A ∩ { F , ii } ∩ { F , si }.
i =1 i =1
∞ , we have
If 1 ∈ (A) : H A

s1 · . . . · s p · i1 · . . . · i p ∈ (A).
Therefore,

f · s1 · . . . · s p · i1 · . . . · i p ∈ (A).
Consider now the case when A is as in Lemma 12 that gives us
j jp k kp
f · s11 · . . . · s p · i11 · . . . · i p ∈ A(ord f )
for some non-negative integers j 1 , . . . , j p and k1 , . . . , k p . Therefore,

f · s1 · . . . · s p · i1 · . . . · i p ∈ A(ord f ) .
Hence, by [5, Corollary 1.7], which gives an upper bound for degrees in the algebraic Nullstellensatz
when the degrees of the generating polynomials are not equal to two, after squaring all elements of
A(ord f ) of degree two, we obtain that

( f · s1 · . . . · s p · i1 · . . . · i p )d ∈ A(ord f ) , (19)
where d is defined by (17). Indeed,
n · 2 H (A)+m+ord f
bounds the number of algebraic indeterminates in A(ord f ) and f ,
H (A) H ( F ) · 2 L ( F )
(because at each step of Algorithm 1 the maximal order doubles at most), and by Corollary 11 we
have

max 4, D A(ord f ) = max 4, D (A) A m + 7, Q ( F ) − 1 .
We also have
1 ∈ { F , si } : f and 1 ∈ { F , ii } : f
for all i, 1 i p. Let G be F , ii or F , si with the maximal t (G , f ). Since in T there exists a general
component of G, we have
f k j = f j + h j,

(t (G , f )) (t (G , f ))
where f j ∈ ( F (t (G , f )) ), h j ∈ (ii ) or (si ), and k j is a natural number. Multiplying
the above expressions, we obtain that

kj
f = g + h, (20)
where g ∈ ( F (t (G , f )) ) and

h ∈ (θ1 s1 ) · . . . · (θ p s p ) · θ1
i1 · . . . · θ p
i p ord(θk ), ord θl
t (G , f ) .
Moreover, again since at each step of Algorithm 1 the maximal order doubles at most, we have
L( F )
A(q) ⊂ F (q+ H ( F )·2 )
for any q ∈ Z0 . By Lemma 13 and inclusion (19) we have
(2p +1)·t (G , f )+1 d)

(θ1 s1 ) · . . . · (θ p s p ) · θ1
i1 · . . . · θ p
i p ∈ A(ord f +4 : f,
where d is defined in (17). Hence,

(θ1 s1 ) · . . . · (θ p s p ) · θ1
i1 · . . . · θ p
i p

(2p +1)·t (G , f )+1 d)
∈ F (ord f + H ( F )·2 +4 :f
L( F )
(21)
for all θk and θk

with ord(θk ), ord(θk
) t (G , f ), 1 k p. Thus, from inequality (18) and inclu-
sion (21) it follows that

L ( F ) +4(2n·2 H (A) +1)·t (G , f )+1 d)

h∈ F (ord f + H ( F )·2 : f,
which finishes the proof because we have representation (20) and g ∈ ( F (t (G , f )) ). 2
Theorem 15. We have

t ( F , f ) A m + 8, max n, H ( F ∪ f ), D ( F ∪ f ) . (22)
Proof. We begin the proof by recalling Corollary 11, which states that at any stage of Algorithm 1
applied to F the orders and degrees of differential polynomials computed by this algorithm do not
exceed the bound

E := A m + 7, Q ( F ∪ f ) − 1 ,
where the function Q is defined in formula (6) and Q ( F ∪ f ) Q ( F ). In particular,
ord f E , H(F ) E,
and by Proposition 10
L ( F ) log2 E .
It follows directly from the definition of Q ( F ) that
n E, m E, 100 E .
Using these inequalities, we can bound the quantity d defined in (17) as
E 2 +2E EE
d E E ·2 EE ,
whence
E
EE
d · 4d 42d E E .
To simplify the latter formula, we note that
k... 2...
kk
22
,
p times pk times
for all natural numbers k and p, which can be easily derived by induction, using the inequality ab
ab , which holds for all natural numbers a, b 2. Thus, we get:
2...
d · 4d 42d
22 .
5E times
This allows us to obtain the following inequality from Lemma 14 for some G ⊂ k{ y 1 , . . . , yn } such
that L (G ) L ( F ) − 1 and H (G ) H ( F ) · 2 L ( F ) :
2... t (G , f )+1 2... t ( G , f )+1

t(F , f ) + 1 E + 1 + E2 + 2 2
22
3 ·
5E times 5E times
...
22
t (G , f )+1
2
.
6E times
Now we use the fact that

2...
A (4, k) = 22
−3
k+3 times
and simplify the above formula as
t ( F , f ) + 1 A (4, 6E )t (G , f )+1 ,
and noting that A (4, 6E ) A (4, A (5, E − 1)) = A (5, E ) yields
t ( F , f ) + 1 A (5, E )t (G , f )+1 .
Now recall that the set G, defined in the proof of Lemma 14, is the set obtained from F by
Algorithm 1 by adding to it an initial or separant. If now we take

E G := A m + 7, Q (G ) − 1
and let G 2 be the set obtained from G at the next iteration of Algorithm 1 by adding an initial or
separant, we can similarly write
t (G , f ) + 1 A (5, E G )t (G 2 , f )+1 .
We continue recursively writing similar inequalities for G 2 , G 3 , . . . , noting that the length of this chain
does not exceed the number of iterations in Algorithm 1, that is, L ( F ) log2 E. We note also that all
quantities E , E G , E G 2 , . . . arising in these inequalities can be uniformly bounded by

Ē := A m + 7, max 9, n, 29E , E = A m + 7, 29E A m + 7, A (4, E ) ,
since the orders and degrees are uniformly bounded by E. Thus, we have
W ...
t ( F , f ) + 1 W W
,
E times
where W = A (5, Ē ), which implies that
2...
t(F , f ) + 1 22
A (4, W E ) A 4, W 2 A 4, A (5, W − 1)
W E times

= A (5, W ) A 5, A (6, Ē − 1) = A (6, Ē )

= A 6, A m + 7, A (4, E ) A m + 6, A m + 7, A (4, E )

= A m + 7, A (4, E ) + 1

= A m + 7, A 4, A m + 7, Q ( F ∪ f ) − 1 + 1

A m + 7, A m + 6, A m + 7, Q ( F ∪ f ) − 1 + 1

= A m + 7, A m + 7, Q ( F ∪ f ) + 1 .
Note that
Q ( F ∪ f ) A (4, B ) A (m + 8, B − 1) − 1,
where B := max(n, H ( F ∪ f ), D ( F ∪ f )) 1, since n 1. Therefore,

A m + 7, A m + 7, Q ( F ∪ f ) + 1

A m + 7, A m + 7, A (m + 8, B − 1) − 1 + 1

A m + 7, A m + 7, A (m + 8, B − 1) − 1

= A m + 7, A (m + 8, B ) − 1
= A (m + 8, B ). 2
6. Model-theoretic proof of existence of the bound
The following argument was shown to the authors by Michael Singer. In this section we refer the
reader for ultrafilters and construction of ultraproducts to books in model theory, for instance [33,
34]. Let k̄ be the differential closure of k (see [35, Definition 3.2] and the references given there) and
q ∈ Z0 . We would like to emphasize that in the statement below we had to fix in advance the number
r of differential polynomials in F to be able to use ultraproducts. So, in Theorem 16 the variable r
is quantified before the bounding function β . However, the constructive bound that we obtained in
Theorem 15 does not have such a restriction, because it depends solely on the orders and degrees in
F and f , but not on the number of elements in F .
Theorem 16. For every r ∈ Z0 there exists a function β : Z30 → Z0 such that for any q ∈ Z0 and F ⊂
k{ y 1 , . . . , yn } with

| F | = r, max H ( F ), D ( F ) q, and 1 ∈ [F ]
we have

1 ∈ F (β(q,r ,n)) .
Proof. Assume that the statement is wrong, that is, there exist r , q ∈ Z0 such that for any α ∈ Z0
there exist p 1,α , . . . , p r ,α ∈ k{ y 1 , . . . , yn } with

max H ( p i j ), D ( p i j ) q
such that
1 ∈ [ p 1,α , . . . , p r ,α ]
and

r
α
( j)
1 = qi , j p i ,α (23)
i =1 j =0
for all q i , j ∈ k{ y 1 , . . . , yn } of order less than or equal to q + α . Again, it is essential here that r does

not depend on α . For a maximal differential ideal M in the differential ring i ∈Z0 k̄ denote the

differential ring ( i ∈Z0 k̄)/ M by K M . There is a natural differential ring homomorphism

k̄ { y 1 , . . . , yn } → K M { y 1 , . . . , yn } =: R .
i ∈Z0
We shall now make a special choice of the maximal differential ideal M. Let F be the filter consisting
of all cofinite subsets of Z0 . Then, there exists an ultrafilter U containing F . Since the field k̄ is
differentially closed, by Łoś’ theorem [33, Theorem 8.5.3] the ultraproduct

K := k̄/U
is a differentially closed field with the following property.

Let ā = (a0 , a1 , a2 , . . .) and b̄ = (b0 , b1 , b2 , . . .) ∈ K . Then, we have: if ā = b̄ then ai = b i for infinitely

many indices i. We now take M to be the kernel of the differential ring homomorphism

k̄ → K .
i ∈Z0
Let p̄ i be the image of ( p i ,1 , p i ,2 , p i ,3 , . . .) in R. This is defined correctly as all p i , j have orders and
degrees bounded. Assume that ( z1 , . . . , zn ) ∈ ( K M )n is a zero of p̄ i for all i. Then, for each i, 1 i r,
there exists V i ⊂ Z0 , V i ∈ U , such that
p i , j ( z1, j , . . . , zn, j ) = 0
for all j ∈ V i , where ( zt ,1 , zt ,2 , zt ,3 , . . .) is mapped to zt under the mentioned differential ring homo-
morphism for each t, 1 t n. Since V 1 ∩ · · · ∩ V r ∈ U and ∅ ∈ / U , there is an index
j ∈ V1 ∩ · · · ∩ Vr.
Therefore,
p 1, j ( z1, j , . . . , zn, j ) = · · · = p r , j ( z1, j , . . . , zn, j ) = 0.
Since k̄ is differentially closed, this contradicts to 1 ∈ [ p 1, j , . . . , p r , j ] in the differential ring

k̄{ y 1 , . . . , yn }. Therefore, since the field K M is differentially closed, we have 1 ∈ [ p¯1 , . . . , p¯r ]. Hence,
there exist γ ∈ Z0 and differential polynomials q¯i j ∈ K M { y 1 , . . . , yn } with ord q¯i j < γ + q so that
γ
r
1= q¯i j p̄ i ( j ) .
i =1 j =0
Again, due to our choice of M (that is, due to the fact that U is an ultrafilter), there exists α ∈ Z0
with α > γ such that
γ
r
( j)
1= qi j p i ,α ,
i =1 j =0
where qi , j ∈ k̄{ y 1 , . . . , yn } of order less than α + q. Since p i ,α ∈ k{ y 1 , . . . , yn } for all i, 1 i r, by

taking a basis of k̄ over k we may assume that in fact q i , j ∈ k{ y 1 , . . . , yn } for all i and j, 1 i r,
0 j γ . This contradicts to (23). Thus, our initial assumption was wrong. 2
7. Conclusions
We have obtained the first bound on orders for the differential Nullstellensatz. Surely, one can
improve the bound and find many applications of it. A general programme which is being realized
here is as follows. The differential elimination algorithms would be very useful for applications if
there were faster versions of them. Our work on bounding orders could lead to:
(1) understanding complexity estimates for the differential elimination,

(2) developing combined and separated differential and high performance algebraic algorithms.
One of the ideas is, instead of using the usual differential elimination, perform all differentiations at
the beginning of the process and then use only fast algebraic methods. We hope that our bounds will
contribute to this programme.
Acknowledgments
We thank Michael Singer for very helpful comments, support, and for the model theoretic proof
of the differential Nullstellensatz. We are grateful to Daniel Bertrand for encouraging us to solve the
problem. We appreciate the help of Erich Kaltofen, Teresa Krick, Alice Medvedev, and Eric Schost in
finding references to the previous work on the algebraic version of the effective Nullstellensatz, on
bounds for the lengths of monomial sequences, and on model theory. We are grateful to the referees
for important suggestions.
References
[1] A. Seidenberg, An elimination theory for differential algebra, Univ. of California Publ. in Math. III (2) (1956) 31–66.
[2] G. Hermann, Die Frage der endlich vielen Schritte in der Theorie der Polynomideale, Math. Ann. 95 (1) (1926) 736–788.
[3] E.W. Mayr, A.W. Meyer, The complexity of the word problems for commutative semigroups and polynomial ideals, Adv.
Math. 46 (3) (1982) 305–329.
[4] W.D. Brownawell, Bounds for the degrees in the Nullstellensatz, Ann. of Math. 126 (3) (1987) 577–591.
[5] J. Kollár, Sharp effective Nullstellensatz, J. Amer. Math. Soc. 1 (4) (1988) 963–975.
[6] L. Caniglia, A. Galligo, J. Heintz, Some new effectivity bounds in computational geometry, in: Applied Algebra, Algebraic
Algorithms and Error-Correcting Codes, Rome, 1988, in: Lecture Notes in Comput. Sci., vol. 357, Springer, Berlin, 1989,
pp. 131–151.
[7] N. Fitchas, A. Galligo, Nullstellensatz effectif et conjecture de Serre (théorème de Quillen–Suslin) pour le calcul formel,
Math. Nachr. 149 (1990) 231–253.
[8] T. Krick, L.M. Pardo, M. Sombra, Sharp estimates for the arithmetic Nullstellensatz, Duke Math. J. 109 (3) (2001) 521–598.
[9] Z. Jelonek, On the effective Nullstellensatz, Invent. Math. 162 (1) (2005) 1–17.
[10] T. Dubé, A combinatorial proof of the effective Nullstellensatz, J. Symbolic Comput. 15 (3) (1993) 277–296.
[11] C. Yap, Fundamental Problems of Algorithmic Algebra, Oxford University Press, New York, 2000.
[12] O. Golubitsky, M. Kondratieva, M. Moreno Maza, A. Ovchinnikov, A bound for Rosenfeld–Gröbner algorithm, J. Symbolic
Comput. 43 (8) (2008) 582–610.
[13] D. Grigoriev, Complexity of quantifier elimination in the theory of ordinary differential equations, in: Lecture Notes in
Comput. Sci., vol. 378, 1989, pp. 11–25.
[14] L. D’Alfonso, G. Jeronimo, P. Solernó, On the complexity of the resolvent representation of some prime differential ideals, J.
Complexity 22 (3) (2006) 396–430.
[15] W. Sweeney, The D-Neumann problem, Acta Math. 120 (1968) 223–277.
[16] E. Hubert, Notes on triangular sets and triangulation-decomposition algorithms II: Differential systems, in: Symbolic and
Numerical Scientific Computing 2001, 2003, pp. 40–87.
[17] W. Sit, The Ritt–Kolchin theory for differential polynomials, in: Differential Algebra and Related Topics, Proceedings of the
International Workshop (NJSU, 2–3 November 2000), 2002, pp. 1–70.
[18] F. Boulier, Triangularisation de systèmes de polynômes différentiels, Série IC2 (Information, Commande, Communication),
Hermès, 2000, never published (in French), http://hal.archives-ouvertes.fr/hal-00140006.
[19] J. Ritt, Differential Algebra, American Mathematical Society, New York, 1950.
[20] E. Kolchin, Differential Algebra and Algebraic Groups, Academic Press, New York, 1973.
[21] E. Hubert, Factorization-free decomposition algorithms in differential algebra, J. Symbolic Comput. 29 (4–5) (2000) 641–
662.
[22] E. Hubert, Improvements to a triangulation-decomposition algorithm for ordinary differential systems in higher degree
cases, in: Proceedings of ISSAC 2004, ACM Press, 2004, pp. 191–198.
[23] F. Boulier, D. Lazard, F. Ollivier, M. Petitot, Representation for the radical of a finitely generated differential ideal, in:
ISSAC’95: Proceedings of the 1995 International Symposium on Symbolic and Algebraic Computation, ACM Press, New
York, NY, USA, 1995, pp. 158–166.
[24] F. Boulier, D. Lazard, F. Ollivier, M. Petitot, Computing representations for radicals of finitely generated differential ideals,
Tech. Rep., Université Lille I, LIFL, 59655, Villeneuve d’Ascq, France, 1997.
[25] F. Boulier, F. Lemaire, Computing canonical representatives of regular differential ideals, in: ISSAC’00: Proceedings of the
2000 International Symposium on Symbolic and Algebraic Computation, ACM Press, New York, NY, USA, 2000, pp. 38–47.
[26] G. Carrà Ferro, A resultant theory for the systems of two ordinary algebraic differential equations, Appl. Algebra Engrg.
Comm. Comput. 8 (6) (1997) 539–560.
[27] M. Kondratieva, A. Levin, A. Mikhalev, E. Pankratiev, Differential and Difference Dimension Polynomials, Kluwer Academic
Publisher, 1999.
[28] D. Bouziane, A. Kandri Rodi, H. Maârouf, Unmixed-dimensional decomposition of a finitely generated perfect differential
ideal, J. Symbolic Comput. 31 (2001) 631–649.
[29] N. Cutland, Computability: An Introduction to Recursive Function Theory, Cambridge University Press, 1980.
[30] G. Moreno Socias, An Ackermannian polynomial ideal, in: Lecture Notes in Comput. Sci., vol. 539, Springer, Berlin/
Heidelberg, 1991, pp. 269–280.
[31] G. Moreno Socias, Length of polynomial ascending chains and primitive recursiveness, Math. Scand. 71 (2) (1992) 181–205.
[32] I. Kaplansky, An Introduction to Differential Algebra, Hermann, Paris, 1957.

[33] W. Hodges, A Shorter Model Theory, Cambridge University Press, Cambridge, 2000.
[34] D. Marker, Model Theory: An Introduction, Springer, New York, 2002.
[35] P.J. Cassidy, M.F. Singer, Galois theory of parametrized differential equations and linear differential algebraic group, in:
IRMA Lect. Math. Theor. Phys., vol. 9, 2007, pp. 113–157.
Journal of Algebra
The implicitization problem for φ : Pn (P1 )n+1

Nicolás Botbol a,b,∗,1
a
Departamento de Matemática, FCEN, Universidad de Buenos Aires, Argentina
b
Institut de Mathématiques de Jussieu, Université de P. et M. Curie, Paris VI, France
Article history: We develop in this paper methods for studying the implicitization
Received 4 March 2008 problem for a rational map φ : Pn (P1 )n+1 defining a hypersur-
Available online 25 March 2009 face in (P1 )n+1 , based on computing the determinant of a graded
strand of a Koszul complex. We show that the classical study of
Keywords:
Macaulay resultants and Koszul complexes coincides, in this case,
Elimination theory with the approach of approximation complexes and we study and
Rational map give a geometric interpretation for the acyclicity conditions.
Syzygy Under suitable hypotheses, these techniques enable us to obtain
Approximation complex the implicit equation, up to a power, and up to some extra factor.
Koszul complex We give algebraic and geometric conditions for determining when
Implicitization the computed equation defines the scheme theoretic image of φ ,
and, what are the extra varieties that appear. We also give some
applications to the problem of computing sparse discriminants.
1. Introduction
In this work we study the implicitization problem for a finite rational map φ : Pn (P1 )n+1 over
a field k, hence, its image is a hypersurface in (P1 )n+1 . Having a rational map φ : Pn (P1 )n+1
is equivalent to having n + 1 pairs of homogeneous polynomials f i , g i of the same degree di , for
i = 0, . . . , n, f i , g i with no common factors.
We show that the classical study of Macaulay resultants and Koszul complexes coincides with
the new approach introduced by L. Busé and J.-P. Jouanolou in [BJ03] and developed by them and
M. Chardin in [Cha06,BC05,Bus06,BCJ06], by means of approximation complexes, defined by J. Herzog,
A. Simis and W. Vasconcelos in [HSV82,HSV83].
*
Address for correspondence: Departamento de Matemática, FCEN, Universidad de Buenos Aires, Argentina.
E-mail address: nbotbol@dm.uba.ar.
1
The author was partially supported by: project ECOS-Sud A06E04, CONICET, UBACYT X042, PAV 120-03 (SECyT), and ANPCyT
PICT 20569, Argentina.
doi:10.1016/j.jalgebra.2009.03.006
N. Botbol / Journal of Algebra 322 (2009) 3878–3895 3879
The process consists of computing the implicit equation by means of the classical methods of elim-
ination theory adapted for this case. More precisely, we will consider the multigraded k-algebra B
that corresponds to the incidence variety associated to the given rational map φ . This algebra can be
presented as a quotient of the polynomial ring R in all the groups of variables, by some linear equa-
tions L 0 , . . . , L n . Consequently we propose as a resolution for B , the Koszul complex K•R ( L 0 , . . . , L n ),
denoted by K• , and we study and give a geometric interpretation of its acyclicity conditions.
In this case, we obtain the implicit equation (up to a power) by taking the determinant of a
suitable strand of a multigraded resolution, that is:

H deg(φ) = Res( L 0 , . . . , L n ) = det (K• )ν , for ν 0.
Later, we analyze the geometrical meaning of these results. We give algebraic and geometric condi-
tions for knowing when, the computed equation defines the scheme theoretic image of φ . And, when
it is not, we present a careful analysis of the extra varieties that appear.
Finally, we give some applications to the problem of computing sparse discriminants, or A-
discriminants (cf. [CD07]), by means of implicitization techniques, that were one of the reasons for
developing this technique.
The key point of our approach is that the hypotheses in our main Theorem 13 are generically sat-
isfied for rational parametrizations whose coordinates are rational functions of degree zero (defining
naturally a rational morphism to (P1 )n+1 ), while if we try to reduce this situation to the standard case
of a rational morphism to Pn by taking a common denominator, bad base points appear in general
and the results developed in [BJ03] do not apply.
2. Preliminaries on commutative algebra
2.1. The Koszul complex
We present here some basic tools of commutative algebra we will need for our purpose, starting
by a classical result due to Hurwitz (1913). He showed that in the generic case, when there are at
least as many variables as homogeneous polynomials, the Koszul complex is acyclic.
Let R be a ring, and P 0 , . . . , P n a sequence in R generating an ideal that we will denote by I .
Denote by K• the associated Koszul complex K• ( P 0 , . . . , P n ; R ):
+1
n i +1

i
0
∂n+1 ∂ i +1
R n+1 −−−→ · · · → R n+1 −−−→ R n+1 → · · · − R n+1 → 0,
∂1
K• : 0 → −→
i +1 i n+1
where the morphisms ∂i +1 : R n+1 → R , are defined in such way that the element
i +1 i +1
e k 1 ∧ · · · ∧ e k i +1 ∈ R n+1 is mapped to j =1 (− 1) j −1
P k j ek1 ∧ · · · ∧ e
k j ∧ · · · ∧ eki +1 .
If R is graded and every P i is homogeneous of degree di > 0, this complex inherits the grading. If
we introduce in K• the grading given by deg(ek1 ∧ · · · ∧ eki+1 ) = dk1 + · · · + dki+1 , the differentials are of
degree 0. It is a well-known fact that, P 0 , . . . , P n is a regular sequence of homogeneous polynomials
of positive degree if and only if K• is acyclic (see for instance [Bou07, Sec. 9, N.7, Cor. 2]). Naturally, if
R is graded or multigraded, and the P i are homogeneous with respect to the grading or multigrading,
this complex inherits this multigrading, and the same statement of acyclicity still holds. We will apply
this to the particular case where R = k[t 0 , . . . , tn ] ⊗k k[ X 0 , Y 0 , . . . , X n , Y n ], and k any field. Here, the
ring is naturally bigraded, and can also be seen as N × Nn+1 -graded, by considering one grading given
by t 0 , . . . , tn , and the Nn+1 -grading given by the n + 1 pairs X i , Y i . The polynomials P i will be hence
multihomogeneous of multidegree (di , 0, . . . , 1, . . . , 0), precisely P i ∈ k[t 0 , . . . , tn ]di ⊗k k[ X i , Y i ]1 , and
will be called L i because of their linearity in the second group of variables.
m A is a noetherian commutative ring, and R = A [ X 1 , . . . , Xm ]. Set I = ( P 0 , . . . , P n ), with
Assume
Pi = j =1 mi j X j , mi j ∈ A. A theorem due to L. Avramov [Avr81] gives necessary and sufficient con-
ditions for I to be a complete intersection in R in terms of the depth of certain ideals of minors of
M := (mi j )i , j ∈ Matm,n+1 ( A ).
3880 N. Botbol / Journal of Algebra 322 (2009) 3878–3895
Theorem 1 (L. Avramov). The ideal I is a complete intersection in R if and only if for all r ∈ {0, . . . , n},
codim A ( I r ) (n + 1) − r + 1. Here I r = I r ( M ) is the ideal of A generated by the r × r minors of M, for
0 r r0 := min{n + 1, m}. We define I 0 := A and I r = 0 for r > r0 .
For a proof we refer the reader to [Avr81, Prop. 1].

Later in this article, we will apply this result when A = k[t 0 , . . . , tn ], R = A [ X 0 , Y 0 , . . . , X n , Y n ], and
M is the 2(n + 1) × (n + 1) matrix
⎛ ⎞
− g0 . . . 0
⎜ f0 . . . 0⎟
⎜ ⎟
⎜ .. . . .. ⎟ ,
M =⎜ . ⎟ (1)
⎜ . . ⎟
⎝ 0 . . . − gn ⎠
0 ... fn
n
that defines a map of A-modules ψ : A n+1 → i =0 A [xi , y i ]1 ∼ = A 2(n+1) , and we see that the symmet-
∼
ric algebra Sym A (coker(ψ)) = A [X]/( P 0 , . . . , P n ), where X stands for the variables x0 , y 0 , . . . , xn , yn .
As A is a graded k-algebra, Sym A (coker(ψ)) is naturally multigraded and the graph of φ is an irre-
ducible component of MultiProj(Sym A (coker(ψ))) ⊂ Pn × (P1 )n+1 .
2.2. Approximation complexes
We give a brief outline of the construction of the approximation complexes of cycles, and show
that the complex Z• coincides in this particular case (under weak hypotheses) with a certain Koszul
complex.
Assume we are given a sequence f := ( f 0 , . . . , f n+1 ) of homogeneous elements of degree d over
the graded ring A = k[t 0 , . . . , tn ], generating an ideal I . Consider the two Koszul complexes K•A (f) and
K•R (f) = K•A (f) ⊗ A R. We have also, in A [x0 , . . . , xn+1 ], another relevant sequence to consider, let us
call it X := (x0 , . . . , xn+1 ), and we can consider also the corresponding Koszul complex K•R (X), which
is of course acyclic because of the regularity of the sequence X.
A straightforward computation permits to verify that their differential anticommutes, i.e. d f ◦ d X +
d X ◦ d f = 0, and this implies that d X induces a differential on the cycles, boundaries and homologies
of the complex K•R (f). We obtain in this way three complexes noted by Z• , B• and M• , called
respectively the approximation complexes of cycles, boundaries and homologies.
Remark 2. Recall that the homology modules of these complexes are, up to isomorphism, independent
of the choice of generators for I . (See for instance [Vas94, Cor. 3.2.7].)
A more explicit description of the Z -complex is the following:
d d d
Z• (f) : 0 → Z n+1 K•R (f) d(n + 1) (−n − 1) −−X→ · · · −−X→ Z 1 K•R (f) [d](−1) −−X→ A [X] → 0,
where Z i (K•R (f)) stands for the ith cycle of the complex K•R (f), and [−] and (−) are the shifts in the
two groups of variables t 0 , . . . , tn and x0 , . . . , xn+1 .
This complex has as objects the bigraded modules

Zi = Z i K• (f; A ) [di ] ⊗ A R (−i ), (2)
and in the future let Z i denote the module Z i (K• (f; A )).
In the case of a two-generated ideal, one has the following:
Proposition 3. With the notation above, if the sequence { f , g } is regular, then there exists a bigraded isomor-
phism of complexes

Z• ( f , g ) ∼
= K• L ; A [x, y ] ,
where L := y f − xg.
Proof. Given the sequence { f , g } the approximation complex is:
(x, y )
Z• ( f 0 , g 0 ) : 0 → Z 1 [d] ⊗ A A [x, y ](−1) −−−→ Z 0 ⊗ A A [x, y ] → 0.
As the sequence { f , g } is regular, H 1 (K•A (f)) = 0, hence Z 1 = (− g , f ) A ∼

= A by the isomorphism ψ
that maps a ∈ A to (− g .a, f .a) ∈ Z 1 , given by the left morphism of the acyclic Koszul complex K•A (f).
Tensoring with A [x, y ] we get an isomorphism of A-modules, Z1 ∼ = A [x, y ].
The commutativity of the following diagram shows that Z• ( f , g ) ∼ = K• ( L ; A [x, y ])
(x, y )
Z• : 0 Z 1 [d] ⊗ A A [x, y ](−1) Z 0 [d] ⊗ A A [x, y ] 0
ψ⊗ A 1 A [x, y ] =
L
K• : 0 A [x, y ][−d](−1) A [x, y ] 0,
where K• denotes K• ( L ; A [x, y ]), [−] denotes the degree shift for the grading on A and (−) the shift
in the variables x, y. 2
2.3. Elimination theory, the Macaulay resultant and the U-resultant
We recall here some basic properties of elimination theory, and also basic facts about resultants
that were introduced by F.S. Macaulay in 1902, a later formalized by J.-P. Jouanolou in “Le formalisme
du Résultant", cf. [Jou91]. This resultant corresponds to a generalization of the Sylvester resultant of
two homogeneous polynomials in two variables. Here, we present a brief outline in elimination theory
and its classical results.
Let Z be the ring of integers, t = t 0 , . . . , tn n + 1 indeterminates. Let d j be n + 1 non-negative
integers and let P j = |α |=d j X j ,α tα be n + 1 homogeneous polynomials of degree d j in the variables
t 0 , . . . , tn generating an ideal I .
Let us write B := Z[ X j ,α | j + 0, . . . , n, |α | = d j ], and R := B [t 0 , . . . , tn ]. Also let us call S the affine
spectrum of B, that is S := Spec( B ). With the assumption deg(t i ) = 1, we have that R / I is a Z-graded
B-algebra. So we can consider the projective S-scheme Z := Proj( R / I ) → PnS , the incidence variety:
i
Z = (X, t) ∈ Pn × S P j (X, t) = 0, ∀ j → PnS .
We denote by A the kernel of the following canonical map of rings A := ker( B = Γ ( S , O S ) →

j
Γ ( Z , O Z )) ∼ 0
= (Hm ( R / I ))0 , and set T := Spec( B /A) → Spec( B ) = S. We have the following commu-
tative diagram of schemes:
i
Z = Proj( R / I ) PnS = Pn ×Z S
p p (3)
j
T = Spec( B /A) S = Spec( B ).
Remark 4. The scheme S parametrizes sequences of polynomials ( P 0 , . . . , P n ) ⊂ Z[t 0 , . . . , tn ]. A closed

point x = V (m) of S belongs to T if and only if the sequence ( P 0 , . . . , P n ) of associated polynomials
has a common root in Pnk for some extension k of S /m.
Theorem 5 (J.-P. Jouanolou). With the notation above, the following statements are satisfied:
(i) A is a principal ideal in B, whose generator will be denoted by ResZ ( P 0 , . . . , P n ).

element Res( P 0 , . . . , P n ) of B is not a zero divisor, and for all j = 0, . . . , n is homogeneous of degree
(ii) The
i = j d i with respect to the variable X j .
(iii) Spec( B /( P 0 , . . . , P n )) is geometrically irreducible in S. Moreover, for any morphism of commutative
rings, Z[ X j ,α | |α | = d j ] → k, we define Resk ( ( P 0 ), . . . , ( P n )) := (ResZ ( P 0 , . . . , P n )) where is
extended to a morphism for B to k[t 0 , . . . , tn ], linearly in the t i ’s.
(iv) If k is a field, Resk ( ( P 0 ), . . . , ( P n )) = 0 if and only if ( P 0 ), . . . , ( P n ) have a common root (different
from zero) in an extension of k.
The original presentation of this result in these terms is in [Jou91, Prop. 2.3].
Remark 6. If the sequence { P 0 , . . . , P n } is regular, Res( P 0 , . . . , P n ) can be computed as the determi-

nant of K• ( P 0 , . . . , P n ; R )ν , for ν > η := (di − 1) (cf. [De84] or [Cha93b]).
3. The algebraic approach
We will establish here the relation between approximation complexes, tensor products of them,
and some Koszul complex we will present below.
Assume that the sequence { f i , g i } is regular for all i ∈ {0, . . . , n}. Then, as all n + 1 Koszul com-
plexes are acyclic, we have isomorphisms between A and the respective modules of cycles, as the
mentioned in the previous section, in Proposition 3.
Definition 7. Write K• the Koszul complex associated to the n + 1 polynomials L i := f i y i − g i xi .

Denote by Z• the complex
n obtained by tensoring the n + 1 approximation complexes Z• ( f i , g i )
over A, namely Z• := i =0 Z• ( f i , g i ).
Proposition 8. With the notation above, if the sequences { f i , g i } are regular for all i ∈ {0, . . . , n}, then there
exists an isomorphism of A-complexes Z• ∼ = K• .
Proof. It is enough to see that by definition of Z• and K• , and Proposition 3, we have

n
n

Z• := Z• ( f i , g i ) ∼
= K• L i ; A [ x i , y i ] ∼= K• 2
i =0 i =0
From Proposition 8 we deduce
Corollary 9. We can resolve the algebra
A [X]
B := Sym( f i , gi ) A [xi , y i ] ∼
= , (4)
( f i y i − g i xi )i =0,...,n
A
by means of the Z• complex, or, equivalently when the hypothesis of Proposition 8 are satisfied, by means of
the Koszul complex K• .
We have the following result.

Lemma 10. The Z• complex satisfies
(i) H 0 (Z• ) ∼ = H 0 (K• ) ∼= ( f i y i − gAi[xXi )]i=0,...,n , and

(ii) Z• is acyclic if and only if the Koszul complex K• ( L 0 , . . . , L n ; A [X]) is, that is, if and only if the sequence
L 0 , . . . , L n is regular.
Remark 11. We said that Avramov’s criterion, stated in Theorem 1, gives a necessary and sufficient
condition for the acyclicity of K• for being acyclic. That is, conditions on B for being a complete
intersection.
As mentioned in Section 2.3, the resultant of L 0 , . . . , L n

as polynomials in the variables t 0 , . . . , tn ,
can be computed as a MacRae invariant of ( R / I )ν , for ν > (di − 1), which is the determinant of a
suitable resolution of R / I in degree ν . Observe that these two complexes are naturally bigraded, one
grading corresponds to the X variables, and the other to the t variables. The acyclicity condition on
Z• is applied to the first group, and the notation ( R / I )ν stands for the grading on the second group.
Hence, for a fixed ν , (Z• )ν is a resolution of the k[X]-module of ( R / I )ν .
Consequently we get by this method a multi-homogeneous generator of the ideal A ⊂ k[X], that
is, we have the following implicitization result:
Corollary 12. If the sequence L 0 , . . . , L n is regular, then the complex
∂n+1 ∂2 ∂1
K• L 0 , . . . , Ln ; A [X] ν : 0 → (Zn+1 )ν −−→ · · · −→ (Z1 )ν −→ (Z0 )ν → 0

is acyclic for all ν > η := (di − 1).
Moreover, with the notation of Section 2.3, det((K• ( L 0 , . . . , L n ; A [X]))ν ) gives an implicit equation for the
closed image of p : Z = BiProj( R / I ) → S̃ = Proj( B ).
Proof. The result follows from Remark 6 and Theorem 5. For instance, by the universality of the re-
sultant (cf. Theorem 5 parts (iii) and (iv)) it is enough to see this for generic coefficients. By Remark 6
det((K• ( L 0 , . . . , L n ; A [X]))ν ) computes Res A ( L 0 , . . . , L n ) for ν > (di − 1). Again, by Theorem 5, this
resultant computes the divisor obtained as the projection of the incidence scheme according to dia-
gram (3) in the projective context:
i
Z = BiProj( R / I ) Pn = Pn ×Z S̃
S̃
p p
j
T = Proj( B /A) S̃ = Proj( B ).
The results follows from the fact that the image of this projection is the closed image of p : Z =
BiProj( R / I ) → S̃ = Proj( B ). 2
In the next section we will focus on the geometric interpretation of this fact, and in reinterpreting
this result in terms of the scheme theoretic image of φ .
4. The geometric approach
4.1. The implicit equation as a resultant
In this section we will focus on the geometrical aspects related to the acyclicity of the Koszul
complex K• , and the nature of the base locus of the rational map φ : Pn (P1 )n+1 . Recall φ is
defined by (t 0 : . . . : tn ) → ( f 0 : g 0 ) × · · · × ( f n : gn ) where f i , g i are homogeneous polynomials of

degree di . Write I (i ) := ( f i , g i ) the homogeneous ideal of A defined by f i , g i .
From Proposition 8, the regularity of the sequence { f i , g i } for all i implies that the complex K• co-
incides with the tensor product of the approximation complex of cycles, Z• (cf. Definition 7). Hence,
we will focus on the conditions on K• for being acyclic, and on its geometrical interpretation.
As we mentioned above, we want here to use a suitable complex for computing A :=
Res A ( L 0 , . . . , L n ). Also we will explain that under certain conditions on the codimension of some
ideals of minors of the matrix M defined in (1), we can assure that this resultant is exactly an
implicit equation of the scheme theoretic image of φ to a certain power, deg(φ).
Write X for the closed subscheme of Pn , given by the common zeroes of all 2(n + 1) polynomials,
and denote by W the base locus. That is:

(i ) (i )
X := Proj A / I and W := Proj A / I . (5)
i i
n n
Set-theoretically we have that X = i =0 V ( f i , g i ) and W = i =0 V ( f i , g i ) inside Pn . Clearly, we al-
ways have X ⊂ W .
Recall that given a matrix M as described in (1), we write I r = I r ( M ) for the ideal of A generated
by the r × r minors of M, for 0 r r0 := min{n + 1, m}, where I 0 := A and I r = 0 for r > r0 .
From Theorem 1 we have a condition for the acyclicity of the Koszul complex K• in terms of the
ideals I r .
For instance, as A is a Cohen Macaulay ring, codimension coincides with depth. In particular, when
n = 2, we have, from 1, that codimP2 ( V ( I 1 )) 3, this is X := V ( I 1 ) = V ( I (0) + I (1) + I (2) ) = ∅, and that
codimP2 ( V ( I 2 )) 2, which implies that V ( I 2 ) = { p 1 , . . . , p s } is a finite set. As V ( I 3 ) = W is the base
locus, codimP2 ( V ( I 3 )) 1 is satisfied when the base points of φ have codimension less than or equal
to 1. More generally:
Theorem 13. Let φ : Pn (P1 )n+1 be, as in the previous section, a finite rational map given by
(t 0 : . . . : tn ) → ( f 0 , g 0 ) × · · · × ( f n , gn ) where f i , g i are homogeneous polynomials of degree di , and for
all i = 0, . . . , n not both f i , g i are the zero polynomial. Let us denote by A the polynomial ring k[t 0 , . . . , tn ],
with L i the expression f i y i − g i xi , for i = 0, . . . , n, and with I (i ) the ideal ( f i , g i ) ⊂ A.
(i) The following statements are equivalent:

A [x0 , y 0 ,...,xn , yn ]
(a) The Koszul complex K• is a free resolution of B := ( L 0 ,..., Ln ) .
(b) codim A ( I r ) (n + 1) − r + 1 for all r = 1, . . . , n + 1.
(c) All the
following statements are true:
n
=1 V ( I ) = ∅.
(i) (i )
i
(ii) #( i < j ( V ( I (i ) · I ( j ) ))) < ∞.

(iii) dim( i < j <k ( V ( I
(i ) · I ( j ) · I (k) ))) 1.
..
.
(n) dim( i ( V ( I (0) · · · I (î ) · · · I (n) ))) n − 2.
(ii) If any (all) of the items before are satisfied, then:
(a) The (multi)homogeneous resultant Res A ,d0 ,...,dn ( L 0 , . . . , L n ) is not the zero polynomial in A.
(b) Denote by H the irreducible implicit equation of the scheme theoretic image H of φ . If for all
{i 0 , . . . , ik } ⊂ {0, . . . , n} we have that codim A ( I (i 0 ) + · · · + I (ik ) ) > k + 1, then, for ν > η =

i (d i − 1),

det (K• )ν = Res A ( L 0 , . . . , L n ) = H deg(φ) .
Proof. (i)(a) ⇔ (i)(b) follows from Avramov’s Theorem 1.

(i)(b) ⇔ (i)(c) note that each r × r-minor of M can be expressed as a product of r polynomials,
where for each column we choose either f or g. Then, the ideal of minors involving the columns
i 0 , . . . , i r −1 coincides with the ideal I (i 0 ) · · · I (ir −1 ) . As we assumed that for any i f i = 0 and g i = 0, the
condition dim( V ( I (0) · · · I (n) )) n − 1 is automatically satisfied.
(i)(a) ⇒ (ii)(a) this is a classical result first studied by J.-P. Jouanolou in [Jou95, §3.5], reviewed in
[GKZ94], and also used by L. Busé, M. Chardin and J.-P. Jouanolou, in their previous work in the area.
(ii)(b) Let us denote by

Z = (t , x): t = (t 0 , . . . , tn ), x = (x0 , y 0 , . . . , xn , yn ), L i (t , x) = 0, ∀i = 0, . . . , n .
As all the polynomials L i are multihomogeneous (and so homogeneous) in the variables xi , y i , and
homogeneous in the t i , then we can think Z as the incidence subvariety in Pn × (P1 )n+1 , and in
Pn × P2(n+1)−1 . We will return to this fact to reduce the problem of analysing the homogeneous
resultant in the space P2(n+1)−1 .
Write Z Ω for the open set defined by the points z ∈ Z , such that π1 ( z) ∈ Ω := Pn − W , where W
is the base locus, as in (5). The closed subscheme Z of Pn × (P1 )n+1 corresponds to the projective
scheme MultiProj(B ) and Z Ω is the open subset of Z that is isomorphic to the complement of the
base locus on Pn . For p in the base locus of φ , e.g. p ∈ V ( I (i ) ), there is a commutative diagram of
schemes as follows:
ZΩ Z Pn × (P1 )n+1
π2
π1−1 ( p ) π1
φ
Ω Pn (P1 )n+1 .
{ p} V ( I (i ) )
Observe that the closed subscheme Z is the Zariski closure of Z Ω , and that the closure of the
second projection of Z Ω coincides with the scheme theoretic image of φ , this is π2 ( Z Ω ) = H. Assume
that Res A ,d0 ,...,dn ( L 0 , . . . , L n ) = 0, hence, this equation defines a divisor [π2 ( Z )] in (P1 )n+1 .
Due to the multihomogenity of the resultant, the polynomial Res := Res A ,d0 ,...,dn ( L 0 , . . . , L n ) ∈ k[X]
is multihomogeneous, hence it defines a closed subscheme in (P1 )n+1 . By Theorem 5, this element is
the image via the specialization map : Z[coef( f i , g i ) : i = 0, . . . , n] → k, of the irreducible equation
ResZ,d0 ,...,dn ( L 0 , . . . , L n ) ∈ Z[coef( f i , g i ), ∀i ][X].
We claim that our hypotheses are the necessary ones to avoid extra factors. For this, set α :=
i 0 , . . . , ik , and write

k
(i j )
X α := Proj A / I , X j := X { j } for just one j and U α := X α − X j. (6)
j =0 j∈
/α
Observe that X α stands for the subset of W , containing X (cf. (5)), where the equations L i 0 , . . . , L ik
vanish identically. If U α is non-empty, consider p ∈ U α , then dim(π1−1 ( p )) = |α | = k + 1. As the
fibre is equidimensional, write Eα := π1−1 (U α ) ⊂ Pn × (P1 )n+1 for the fibre over U α , which defines a
(multi)projective bundle of rank |α |. Hence, we have

codim(Eα ) = n + 1 − (k + 1) + codimPn (U α ) .
The condition codim A ( I (i 0 ) + · · · + I (ik ) ) > k + 1, for all {i 0 , . . . , ik } ⊂ {1, . . . , n} implies that
codimPn (U α ) > k + 1. Hence
codim(Eα ) > n + 1 = codim( Z Ω ).
Observe that Z − Z U = Z Ω , where Z U := α Eα , and that codim( Z U ) > n + 1 = codim( Z Ω ) =

codim( Z Ω ).
As Spec( B ) is a complete intersection, in A2(n+1) it is unmixed and purely of codimension n + 1.
Thus, Z = ∅ is also purely of codimension n + 1. This and the fact that codim( Z U ) > n + 1 implies that
Z = ZΩ .
The graph Z Ω is irreducible, hence Z is irreducible as well, and its projection (the closure of the
image of φ ) is of codimension 1, and one has for ν > η :

det (K• )ν = divk[X] H 0 (K• )ν = divk[X] (Bν ) = lengthk[X]p (Bν )p [p].
p prime, codimk[X] (p)=1
as [det((K• )ν )] = divk[X] (Res) = e .[q] for some integer e and q := ( H ) ⊂ k[X], we have that
Also
p prime, codim(p)=1 lengthk[X]p ((Bν )p )[p] = e .[q] hence [det((K• )ν )] = lengthk[X]q ((Bν )q )[q]. Finally for
ν 0

lengthk[X]q (Bν )q = dimκ (q) Bν ⊗k[X]q κ (q) = deg(φ),
where κ (q) := k[X]q /qk[X]q . This shows that e = deg(φ) and completes the proof. 2
Remark 14. We showed that the scheme π2 ( Z ) is defined by the polynomial Res, while the closed
image of φ coincides with π2 ( Z Ω ), hence the polynomial H divides Res. Moreover, from the proof
above we conclude that H deg(φ) also divides Res. And if [Eα ] is an algebraic cycle of Pn × (P1 )n+1
of codimension n + 1, then [π2 (Eα )] is not a divisor in (P1 )n+1 , and consequently Res has no other
factor than H deg(φ) .
Remark 15. With the hypotheses of Theorem 13(ii)(b), denoting by degi the degree on the variables
xi , y i and by degtot the total one, we have:

(i) degi ( H ) deg(φ) = j =i d j ;

(ii) degtot ( H ) deg(φ) = i j =i d j .
In the rest of this section we will focus on the study of the difference between the closed schemes
Z Ω and Z , and its projection π2 ( Z − Z Ω ) on (P1 )n+1 . We will show how this can give extra factors
on the resultant.
4.2. Analysis of the extra factors
In this section our aim is to analyse in detail the nature of the extra factors that appear when the
codimension condition of Theorem 13(ii)(b) does not hold. We will see that these polynomials appear
due to the existence of big enough dimensional fibres over some of the components of the base locus.
In a few words, it happens that if the sum of the dimension of a component in the base locus plus
the dimension of its fibre is n, there is a “big dimensional” subscheme whose projection on (P1 )n+1
defines a hypersurface which gives rise to a factor in the computed resultant.
In order to understand this, we will first analyse some simple cases, namely, where this phe-
nomenon occurs over a finite set of points of the base locus; and later, we will deduce the general
implicitization result.
Example 16. Assume we are given a rational map φ : P2 P1 × P1 × P1 , where φ maps (u : v : w )
to ( f 0 : g 0 ) × ( f 1 : g 1 ) × ( f 2 : g 2 ), of degrees d, d and d respectively.
We may suppose that each of the pairs of polynomials { f 0 , g 0 }, { f 1 , g 1 } and { f 2 , g 2 } have no
common factors. Then, the condition codim A ( I (i ) ) 2 is automatically satisfied. Assume also that
X = ∅, this is, there are no common roots to all 6 polynomials.
We will show here that, if we do not ask for the “correct” codimension conditions, we could
be implicitizing some extra geometric objects. For instance, suppose that we take a simple point
p ∈ V ( I (0) + I (1) ) = ∅. Consequently L 0 (u , v , w , X) = L 1 (u , v , w , X) = 0 for all choices of X. Never-
theless, L 2 (u , v , w , X) = g 2 (u , v , w )x2 − f 2 (u , v , w ) y 2 = 0 imposes the nontrivial condition g 2 ( p )x2 −
f 2 ( p ) y 2 = 0 on (x2 , y 2 ), hence there is one point q = ( f 2 ( p ) : g 2 ( p )) ∈ P1 which is the solution of this
equation. We get π1−1 ( p ) = { p } × P1 × P1 × {q}. As we do not want the reader to focus on the precise
computation of this point q, we will usually write {∗} for the point {q} obtained as the solution of
the only nontrivial equation.
Suppose also that, for simplicity, V ( I (0) + I (1) ) = { p }, V ( I (0) + I (2) ) = ∅, and V ( I (1) + I (2) ) = ∅. This
says that if we compute π2 ( Z ), then we get

π2 ( Z ) = π2 π1−1 (Ω ∪ W ) = π2 π1−1 (Ω) ∪ π2 π1−1 ( W )

= π2 ( Z Ω ) ∪ π2 { p } × P1 × P1 × {∗}

= im(φ) ∪ P1 × P1 × {∗} ,

where W = Proj( A / i I (i ) ) the base locus of φ as in (5), and Ω = Pn − W its domain.
Hence, the Macaulay resultant contains some extra factor. Let us observe that if there is only one
extra hyperplane appearing (over a point p with multiplicity one), which corresponds to π2 (π1−1 ( p )),
then π1−1 ( p ) is a closed subscheme of Z , defined by the equation L 3 ( p ) = 0. Then, we will show that
Res(u , v , w ) ( L 0 , L 1 , L 2 ) = Hdeg(φ) · L 3 ( p ).
We will now generalize Theorem 13 in the spirit of the example above. For each i ∈ {0, . . . , n} take
X î := Proj( A / j =i I ( j ) ).
Proposition 17. Let φ : Pn (P1 )n+1 be a rational map that satisfies conditions (a)–(c) of Theorem 13.
Assume further that for all {i 0 , . . . , ik } ⊂ {0, . . . , n}, with k < n − 1, codim A ( I (i 0 ) + · · · + I (ik ) ) > k + 1. Then,
there exist non-negative integers μ p such that:

n
Res A ( L 0 , . . . , L n ) = H deg(φ) · L i ( p )μ p .
i =0 p ∈ X î
Proof. Denote by Γ := Z Ω the closure of the graph of φ , Z as before. From Remark 14, we can write
Res A ( L 0 , . . . , L n )
G := ,
H deg(φ)
the extra factor. It is clear that G defines a divisor in (P1 )n+1 with support on π2 ( Z − Γ ). From the
proof of Theorem 13, we have that Z and Γ coincide outside W × (P1 )n+1 . As Z is defined by linear
equations in the second group of variables, then Z − Γ is supported on a union of linear spaces
over the points of W , and so, its closure is supported on the union of the linear spaces (π1 )−1 ( p ) ∼ =
{ p } × ((P1 )n × {∗}), where {∗} is the point (x : y ) ∈ P1 such that L i ( p , x, y ) = 0 for suitable i. It follows
that π2 ((π1 )−1 ( p )) ⊂ V ( L i ) ⊂ (P1 )n+1 , and consequently

G= L i ( p )μ p ,
p∈ X
for some non-negative integers μp . 2
As we said at the beginning of this section, we are interested in understanding the extra factors.
We have shown above that these factors depend directly on the dimension of each component of the
base locus and on the dimension of its fibre; and that if their sum equals n this generates an extra
hypersurface, hence an extra factor. It is natural to ask, why this sum does not exceed n. The next
lemma gives an answer to this question.
Lemma 18. Let φ : Pn (P1 )n+1 , be a finite rational map as above. Under the hypothesis of Theorem 1,
k
codim A ( j =0 I (i j ) ) k + 1 for all k = 0, . . . , n and {i 0 , . . . , ik } ⊂ {0, . . . , n}.
Proof. To show this we will use Theorem 1. Let us denote by I the ideal I (i 0 ) + · · · + I (in ) , and for
k n
a fixed k write I = I + I , where I = j =0 I
(i j ) and I =
l=k+1 I
(il ) the sum over the comple-
mentary indices of {i 0 , . . . , ik }. As ( L 0 , . . . , L n ) is a complete intersection in R, also is ( L I 0 , . . . , L ik )

in A [xi 0 , y i 0 , . . . , xik , y ik ]. Applying Theorem 1 to the ideal ( L I 0 , . . . , L ik ), for r = 1 we have that
codim A ( I (i 0 ) + · · · + I (ik ) ) k + 1.
Observe that as I is generated by a subset of the set of generator of I then I is also a complete
intersection in R. Now, as it is generated by elements only depending on the variables xi j , y i j for
j = 0, . . . , k, we have that it is also a complete intersection in A [xi 0 , y i 0 , . . . , xik , y ik ]. 2
Remark 19. It is easy to see that the converse of Lemma 18 does not always hold (in fact almost
never). For example let us take for n = 1, f 0 = g 0 = f 1 = g 1 = 0.
We will now introduce some notation we will use to state and prove Theorem 22. The idea behind
these definitions is to split the base locus into irreducible subvarieties with the desired dimension, in
order to obtain a clear factorization of the extra component that appears.
k
Definition 20. For each α := (i 0 , . . . , ik ) set I α := j =0 I
(i j ) , and X := Proj( A / I α ) as defined in (6).
α
Denote by Θ the set of all α ⊂ {0, . . . , n} such that codim ( I α ) = |α | .
For each α := (i 0 , . . . , ik ) ∈ Θ , let I α = ( qi ∈Λα qi ) ∩ q , where Λα is the set of primary ideals of
codimension |α | containing I , and codim A (q ) > |α |. Write
X α ,i := Proj( A /qi ), with qi ∈ Λα , (7)
and let X αred ,i be the associated reduced variety. Denote by πα the projection onto the coordinates
(xi : y i ), for i ∈
/ α . Namely, let {ik+1 , . . . , in } := {1, . . . , n} − α , then πα is given by
n+1 n+1−|α |
πα : P1 → P1 ,
(x0 : y 0 ) × · · · × (xn : yn ) → (xik+1 : y ik+1 ) × · · · × (xin : y in ). (8)
Set φα : Pn Pα := πα ((P1 )n+1 ) defined as φα := πα ◦ φ . Hence, φα is explicitly defined as φα (t) =

( f ik+1 (t) : g ik+1 (t)) × · · · × ( f in (t) : g in (t)) as is shown in the diagram below
φ
Pn (P1 )n+1
φα
πα
Pα := πα ((P1 )n+1 ) ∼
= (P1 )n+1−|α | .

Denote by W α the base locus of φα and X := Proj( A / i I (i ) ). Since W := Proj( A / i I (i ) ) is the base
locus of φ as defined in (5), we have the inclusions X ⊂ W α ⊂ W .
Denote by Uα := Pn − W α , the open set where φα is well defined. Write Ωα := X α ∩ Uα and
Ωα ,i := X α ,i ∩ Uα . The open set Ωα ,i stands for the points of Pn which give rise to an extra factor.
This factor coincides with the restricted image of φ over Ωα ,i to a certain power.
If α is empty, we set πα = Id(P1 )n+1 , φα = φ , W α = W and Uα = Ωα = Ω .
Next, we show that this meticulous decomposition of the base locus has the properties we are
looking for. Namely, the fibre over each of the open sets Ωα ,i has dimension |α |, in such a way that
the sum of the dimensions of the base plus the dimension of the fiber is n.
Lemma 21. Let φ be as in Theorem 13. For each α ∈ Θ and each qi ∈ Λα , the following statements are satisfied:
(i) Ωα ,i is non-empty.
(ii) For all p ∈ Ωα ,i , dim(π1−1 ( p )) = |α |.
(iii) The restriction φα ,i of φ to Ωα ,i , defines a rational map
φα ,i : X α ,i Pα ∼
= (P1 )n+1−|α | . (9)
(iv) Z α ,i := π1−1 (Ωα ,i ) is a (multi)projective bundle Eα ,i of rank |α | over Ωα ,i .

Proof. Fix X α ,i ⊂ X α and write α := i 0 , . . . , ik . Note that Ωα ,i := X α ,i − j ∈/ α X { j } is an open subset

of X α ,i . If Ωα ,i = ∅, then X α ,i ⊂ j ∈/ α X { j } , and since it is irreducible, there exists j such that X α ,i ⊂
X { j } , hence X α ,i ⊂ X { j } ∩ X α = X α ∪{ j } . Setting α := α ∪ { j }, it follows that dim( X α ) dim( X α ,i ) =
n − |α | > n − |α |, which contradicts the hypothesis.
Let p ∈ Ωα ,i , π1−1 ( p ) = { p } × {qik+1 } × · · · × {qin } × (P1 )|α | , where the point q i j ∈ P1 is the only
solution to the nontrivial equation L i j ( p , xi j , y i j ) = y i j f i j ( p ) − xi j g i j ( p ) = 0. Then we deduce that
n
dim(π1−1 ( p )) = |α |, and that φα ,i : Ωα ,i → ∼
= (P1 )n+1−|α | given by p ∈ Ωα ,i → {qik+1 } ×
j =k+1 P(i j )
1
n+1
· · · × {qin } ∈ j =k+1 P(i j ) ,
1
is well defined.
The last statement follows immediately from the previous ones. 2
Note that X α ,i has dimension n − |α |, by the preceding lemma dim(π1−1 (Ωα ,i )) = n. Since
Res A ( L 0 , . . . , L n ) describes the codimension one part of π2 ( Z ), if dim(π2 (π1−1 (Ωα ,i ))) = n, then
Res A ( L 0 , . . . , L n ) is not irreducible. Denote by α ⊂ Λα the subset of primary ideals qi satisfying
dim(π2 (π1−1 ( X α ,i ))) = n. Observe that if |α | = n we are in the case of Proposition 17. The following
diagram illustrates this situation:
Z Pn × (P1 )n+1
π2
π1−1 ( p ) Eα , i π1
φ
{ p} Ωα ,i Pn (P1 )n+1
φα , i
πα
φα , i
X α ,i π2 (Eα ,i ) (P1 )n+1−|α | ,
where in this case π1−1 ( p ) = { p } × {qik+1 } × · · · × {qin } × (P1 )|α | is the |α |-dimensional fibre of Eα ,i
over p, for some q i j ∈ P1 .
We finally state our general result.
Theorem 22. Let φ : Pn (P1 )n+1 be a finite rational map, satisfying the hypotheses of Theorem 13 and
conditions (i)(a)–(i)(c) in that theorem, hence Res A ( L 0 , . . . , L n ) = 0. Denote by H the irreducible implicit
equation of the closure of the image of φ , and by H α ,i the irreducible implicit equation of the closure of the
image of φα ,i defined in (9) for each α ∈ Θ , and i such that qi ∈ α .
Then, there exist positive integers μα ,i such that:

Res A ( L 0 , . . . , L n ) = H deg(φ) · ( H α ,i )μα,i deg(φα,i ) .
α ∈Θ i :qi ∈α
Res ( L ,..., L )
Proof. Denote by Γ := Z Ω , the graph of φ , Z the incidence scheme, and G = A 0 n
the extra
H deg(φ)
factor.
As in Proposition 17, G defines a divisor in (P1 )n+1 with support on π2 ( Z − Γ ), and Z and Γ
coincide outside W × (P1 )n+1 .
By part (iii) of Lemma 21, for each α and each qi ∈ α ⊂ Λα , φα ,i defines a (multi)projective
bundle Eα ,i of rank |α | over Ωα ,i . By definition of α , π2 (Eα ,i ) is a closed subscheme of (P1 )n+1 of
codimension 1. Denoting by [Eα ,i ] = μα ,i .[Eαred
,i ] the class of Eα ,i as an algebraic cycle of codimension
n + 1 in Pn × (P1 )n+1 , we have (π2 )∗ [Eα ,i ] = μα ,i .(π2 )∗ [Eαred
,i ] = μα ,i . deg(φα ,i ).[pα ,i ], where pα ,i :=
( H α ,i ) .
As in Theorem 13, one has for ν > η :

det (K• )ν = divk[X] H 0 (K• )ν = divk[X] (Bν ) = lengthk[X]p (Bν )p [p].
p prime, codimk[X] (p)=1
We obtain that

det (K• )ν = lengthk[X]p (Bν )pα,i [pα ,i ] + lengthk[X]( H ) (Bν )( H ) ( H ) .
α ,i
α ∈Θ pα ,i
In the formula above, for each pα ,i we have

lengthk[X]p (Bν )pα,i = dimκ (pα,i ) Bν ⊗k[X]pα,i κ (pα ,i ) = μα ,i . deg(φα ,i ),
α ,i
where κ (pα ,i ) := k[X]pα,i /pα ,i k[X]pα,i . Consequently we get that for each α ∈ Θ , there is a factor of G,
denoted by H α ,i , that corresponds to the irreducible implicit equation of the scheme theoretic image
of φα ,i , raised to a certain power μα ,i . deg(φα ,i ). 2
Remark 23. It is important to remark that the set-theoretic approach does not tell us anything about
the scheme structure of the fibre (π1 )−1 ( W ). Namely the bijection
|α |
Xα × P1 × {∗} × · · · × {∗} −∼→ (π1 )−1 ( W )
α
is not necessarily a scheme isomorphism; for instance, Eα ,i need not be well-defined over all X α ,i , be-
cause the multiplicities of the components of X α ,i of Proj( A / I α ) are not necessarily preserved by π1−1 .
The fiber bundle Eα ,i is defined over the relative open set Ωα ,i of X α ,i ⊂ Proj( A / I α ), but the dimen-
sion of the fibre can increase on X α ,i − Ωα ,i .
Remark 24. Let us discuss briefly how the exponents μα ,i behave in the particular case where X α ,i
is a complete intersection subscheme of Pn . Take α ∈ Θ , defined as above, write α = {i 0 , . . . , ik }, and
fix pi ∈ Λα . As codim( I α ,i ) = |α |, we have dimPn ( X α ,i ) = n − |α | = n − (k + 1) =: m, and assume X α ,i
is irreducible.
Take G 0 , . . . , G k irreducible generators of I α ,i (recall X α ,i is a complete intersection), and G j := L i j
for j ∈ {k + 1, . . . , n}.
As G 0 , . . . , G k vanish over X α ,i , the element Res(G 0 , . . . , G n ) ∈ k[xik+1 , y ik+1 , . . . , xin , y in ] describes
exactly the irreducible implicit equation, H α ,i , of the scheme theoretic image, Hα ,i ⊂ P1i × · · · × P1in ,
k+1
of the restricted and correstricted map φα ,i .
Assume that L j lies in G := (G 0 , . . . , G n )μ j , and that μ j is maximum with this property, for all j.
Then, from the “Lemme de divisibilité général” by J.-P. Jouanolou (see [Jou91, Prop. 6.2.1]), we have
μj μα , i
Res(G 0 , . . . , G n ) = H α ,i . As this polynomial divides H α ,i ,
j
j μ j gives a lower bound for μα ,i .
Remark also that L i j always lies in the ideal G := (G 0 , . . . , G n )μ j for μi j = 1 by definition, when
j > k.
5. Examples
In this section we present several examples where we illustrate the theory developed in the pre-
vious sections. These computations were done in Maple 11, by means of the routines implemented in
the Multires developed by Galaad Team at INRIA, cf. [BM].
Example 25. In this example we show what happens when there is a point whose fibre has dimen-
sion 2.
> read"multires.mpl": with(linalg):

> f0 := u: g0 := v: f1 := u^2: g1 := v^2: f2 := v^2: g2 := w^2:
2 2 2 2
L0 := x0 v - y0 u, L1 := x1 v - y1 u, L2 := x2 w - y2 v
> M012w:=det(mresultant([L0,L1,L2],[u,v]));
2 4 2 2 2
M012w := -x2 w y1 (y1 x0 - x1 y0 ) y2
As mresultant gives a multiple of the desired Macaulay resultant, computing the greatest common
divisor over all permutations of L 0 , L 1 , L 2 we get that M w should be:
2 4 2 2 2
Mw := x2 w (y1 x0 - x1 y0 )
It remains to observe that x22 correspond to the equation L 2 ( p ) = 0, where p = (0 : 0 : 1) is the

point with 2-dimensional fibre, and use the fact that p has multiplicity 2.
Example 26. In this example we study the case of a two points (p 1 and p 2 ) base locus, where the
fibre above them has dimension 2. Take p 1 = (1 : 0 : 0) and p 2 = (0 : 0 : 1). The computation below
shows that in this case the methed yields a power of the irreducible implicit equation with some
extra factors. Those factors are powers of y 1 and x2 .
Why y 1 and x2 ? If we look at the linear equations L 0 , L 1 , L 2 evaluated in the two points, p 1 ,
p 2 , we see that L 0 ( p 1 ) = L 2 ( p 1 ) = 0 and L 1 ( p 1 ) = y 1 . If we do the same with p 2 we get L 0 ( p 2 ) =
L 1 ( p 2 ) = 0 and L 2 ( p 2 ) = x2 .
> f0 := u*w: g0 := v^2: f1 := u^2: g1 := v^2: f2 := v^2: g2 := w^2:

> L0:=x0*g0-y0*f0: L1:=x1*g1-y1*f1: L2:=x2*g2-y2*f2:
As we did before, we compute the greatest common divisor over all the permutation of L 0 , L 1 , L 2 .
The output M w, M v and Mu is:
2 8 3 2 2 2
Mw:= x2 w y1 (-y2 x1 y0 + x2 x0 y1)
8 3 4 2 2 2
Mv:= v y1 x2 (-y2 x1 y0 + x2 x0 y1)
2 8 4 2 2 2
Mu:= y1 u x2 (-y2 x1 y0 + x2 x0 y1)
The resultant can be obtained as the gcd of the three equations above.
Remark 27. In the spirit of Remark 24 take α = {0, 1}, that is, X α = V ( f 0 , g 0 , f 1 , g 1 ) = V (u , v ),
and consider G 0 = u, G 1 = v, and G 2 = L 2 = y 2 v 2 − x2 w 2 . By the multiplicatively of the resul-
tant (see for instance [Jou91, Sec. 5.7]) we have Res(G 0 , G 1 , G 2 ) = Res(u , v , −x2 w 2 ) = −x2 . Now, as
L 0 ∈ G := (G 0 , G 1 , G 2 ), but not in G 2 , and L 1 ∈ G 2 , but not in G 3 , we have μ0 = 1 and μ1 = 2.
Hence, Res(G 0 , G 1 , G 2 )2 = x22 divides Res( L 0 , L 1 , L 2 ), as the above computation showed. Moreover, in
this case, we see that Res(G 0 , G 1 , G 2 )2 = x22 coincides exactly with the extra factor.
Example 28. In this example we study the situation where the fibre along a 1-dimensional closed
subscheme of P2 is P1 , and is P2 above the points (1 : 0 : 0) and (0 : 0 : 1).
> f0 := u*v: g0 := u*w: f1 := u^2+v^2: g1 := v^2: f2 := v^2: g2 :=w^2:

> L0:=x0*g0-y0*f0; L1:=x1*g1-y1*f1; L2:=x2*g2-y2*f2;
8 2 3 2 2 2 2
Mw:= w x2 y1 y2 (x1 - y1) (-x0 y2 + y0 x2)
8 4 3 2 2 2 2
Mv:= v x2 y1 (x1 - y1) (-x0 y2 + y0 x2)
8 2 2 2 2
Mu:= u y1 (-x0 y2 + y0 x2)
This last equation corresponds to the situation on the open set u = 0. Is clear that the extra factor
y 1 is appearing because of the 2-dimensional fibre above p = (1 : 0 : 0), where L 0 ( p ) = L 2 ( p ) = 0, and
L 1 ( p ) = u 2 y 1 . Similarly x2 appears because of the 2-dimensional fibre above the point q = (0 : 0 : 1),
where L 0 (q) = L 1 (q) = 0, and L 2 (q) = w 2 x2 , as was shown in the second example. The other factor
that appears is (x1 − y 1 )2 , and it is due to the existence of a 1-dimensional closed subvariety V . In
this case, we see that the dimension of the fibre over a point p is 1, for any p in a suitable relative
open subset of V .
Precisely, we consider the fibre along the closed subvariety of P2 , V (u ) = {(u : v : w ): u = 0} and
we project from this closed subvariety of the incidence variety to the space P1 × P1 corresponding to
the second two copies (because the first one would be (0 : 0)). If we compute the implicit equation
by the method before, for the map
φ|(u =0) : (u = 0) ∼
= P1 P1 × P1 ,
( v : w ) → ( f 1 : g 1 ) × ( f 2 : g 2 ),
we get as output: M v := − v 4 (x1 − y1)2 and M w := x4 (x1 − y1)2 x22 . Observe that the x22 appearing
in the second equation is still the extra factor coming from the big fibre over the point q in the closed
set V (u ). We see here that the dimension of the fibre is not constant on V (u ), but on a relative open
set where it defines a fibre bundle E .
6. Applications to the computation of sparse discriminants
The computation of sparse discriminants is equivalent to the implicitization problem for a para-
metric variety, to which we can apply the techniques developed in the previous sections. In the
situation described in [CD07], a rational map f : Cn Cn given by homogeneous rational func-
tions of total degree zero is associated to an integer matrix B of full rank. This is done in such a way
that the corresponding implicit equation is a dehomogenization of a sparse discriminant of generic
polynomials with exponents in a Gale dual of B.
Suppose for instance that we take the matrix B below:
⎛ ⎞
1 0 0
⎜−2 1 0⎟
⎜ ⎟
B = ⎜ 1 −2 1 ⎟.
⎝ 0 1 −2
⎠
0 0 1
In this case, as the columns of B generate the affine relations of the lattice points {0, 1, 2, 3, 4}. The
closed image of the parametrization f is a dehomogenization of the classical discriminant of a generic
univariate polynomial of degree 4. Explicitly, from the matrix we get the linear forms l1 (u , v , w ) = u,
l2 (u , v , w ) = −2u + v, l3 (u , v , w ) = u − 2v + w, l4 (u , v , w ) = v − 2w, l5 (u , v , w ) = w (whose coeffi-
cients are read in the rows of B), and the polynomials f 0 = l1 · l3 , g 0 = l22 , f 1 = l2 · l4 , g 1 = l23 , f 2 = l3 · l5 ,
g 2 = l24 (the exponents of the linear forms are read from the columns of B). This construction gives
rise to the following rational map:
f : C3 C3 ,

u (u − 2v + w ) (−2u + v )( v − 2w ) (u − 2v + w ) w
(u , v , w ) → , , .
(−2u + v )2 (u − 2v + w )2 ( v − 2w )2
First, we see that we can get a map from P2C because of the homogeneity of the polynomials. Also,
taking common denominator, we can have a map to P3C , this is:
f : P2C P3C ,
(u : v : w ) → ( f 0 : f 1 : f 2 : f 3 ),
where f 0 = (−2u + v )2 (u − 2v + w )2 ( v − 2w )2 is the common denominator, f 1 = u (u − 2v + w )3 ×

( v − 2w )2 , f 2 = (−2u + v )3 ( v − 2w )3 and f 3 = (u − 2v + w ) w (−2u + v )2 (u − 2v + w )2 .
The problem with this way of projectivizing is that, in general, we cannot implement the theory
developed by L. Busé, M. Chardin, and J.-P. Jouanolou, because typically the base locus has awful
properties, as a consequence of taking common denominator.
As a possible way out, we propose in this work to consider the morphism of projective schemes
given by:
φ : P2 P1 × P1 × P1 ,
(u : v : w ) → ( f 0 : g 0 ) × ( f 1 : g 1 ) × ( f 2 : g 2 ),
where f 0 = u (u − 2v + w ), g 0 = (−2u + v )2 , f 1 = (−2u + v )( v − 2w ), g 1 = (u − 2v + w )2 , f 2 =

(u − 2v + w ) w g 2 = ( v − 2w )2 . For this particular example, we get that there are only two base
points giving rise to an extra factor, namely p = (1 : 2 : 3) and q = (3 : 2 : 1). Is easy to see that those
points give rise to two linear factors in the resultant.
First, we observe that this situation is better, because we are not adding common zeroes. Moreover,
if a point (u : v : w ) is a base point here, it also is in the two settings above.
Remember also that in the n = 2 case, the condition required on the Koszul complex associated to
this map for being acyclic is that the variety X , defined as the common zeroes of all the 6 polyno-
mials, be empty. In general, the conditions we should check are the ones imposed by the Avramov’s
theorem, as was shown in Theorem 13.
Note also that if we want to state this situation in the language of approximation complexes, we
need only to replace K• by Z• , because we can assume that { f i , g i } are regular sequences, due to
the fact that gcd( f i , g i ) = 1.
Remark 29. For a matrix like the B above, it is clear that the closed subvariety X is always empty, due
to the fact that all maximal minors of B are not zero, and the polynomials g i ’s involve independent
conditions. Then, the only common solution to l22 = l23 = l24 = 0 is (u , v , w ) = (0, 0, 0), and so X = ∅
in P2 . In this case, it is still better (from an algorithmic approach) to compute the discriminant of a
generic polynomial of degree 4 in a single variable and then dehomogenize, because, in our setting,
the number of variables is bigger than 1. But when the number of monomials of a sparse polynomial
in many variables is not big, this Gale dual approach for the computation of sparse discriminants
provides a good alternative.
We will give next an example where we show a more complicated case.
Example 30. Let C be the matrix given by
⎛ ⎞
1 −7 −6
⎜−1 4 3⎟
⎜ ⎟
C =⎜ 1 0 4 ⎟.
⎝ 0 1 −1
⎠
−1 2 0
As before, denoting by b i the ith row of C , we get the linear forms li (u , v , w ) = b i , (u , v , w ), asso-
ciated to the row vectors b i of B, where , stands for the inner product in C3 . Then we define the
homogeneous polynomials f 0 = l1 · l3 , g 0 = l2 · l5 , f 1 = l42 · l4 · l25 , g 1 = l71 , f 2 = l32 · l43 , f 2 = l61 · l4 . And we
obtain the following rational map:
φ : P2 P1 × P1 × P1 ,
(u : v : w ) → ( f 0 : g 0 ) × ( f 1 : g 1 ) × ( f 2 : g 2 ).
It is easy to see that in this case the variety X is not empty, for instance the point p = (1 : 1 : −1),
defined by l1 = l2 = 0 belongs to X .
As was shown by M.A. Cueto and A. Dickenstein in [CD07, Lemma 3.1 and Thm. 3.4], we can in-
terpret the discriminant computed from the matrix C in terms of the dehomogenized discriminant
associated to any matrix of the form C · M, where M is a square invertible matrix with integer coef-
ficients. That is, we are allowed to do operations on the columns of the matrix C , and still be able
to compute the desired discriminant in terms of the matrix obtained from C . In [CD07] they give an
explicit formula for this passage.
In this particular case, we can multiply C from the right by a determinant 1 matrix M, obtaining
⎛ ⎞ ⎛ ⎞
1 −7 −6 ⎛ ⎞ 1 0 0
⎜−1 4 3⎟ 1 12 −1 ⎜−1 −3 0⎟
⎜ ⎟ ⎜ ⎟
C ·M =⎜ 1 0 4 ⎟ · ⎝0 6 −1 ⎠ = ⎜ 1 −8 3 ⎟.
⎝ 0 1 −1
⎠ 0 5 1
⎝ 0 11 −2 ⎠
−1 2 0 −1 0 −1
Similar to what we have done before, we can see that the closed subvariety X associated to the ra-
tional map that we obtain from the matrix C · M is empty. Observe that #V ( I 2 ) is finite due to the fact
that l2 = l4 = 0 or l3 = l4 = 0 or l3 = l5 = 0 should hold. Moreover it is easy to verify that all maximal
minors are nonzero, and this condition implies that any of the previous conditions define a codimen-
sion 2 variety, this is, a finite one. A similar procedure works for seeing see that codim A ( I 3 ) 2.
Finally the first part of Theorem 13 implies that the Koszul complex K• is acyclic and so we can
compute the Macaulay resultant as its determinant.
Moreover, this property over the minors implies that codim A ( I (i 0 ) ) = 2 > k + 1 = 1 and that
codim A ( I (i 0 ) + I (i 1 ) ) = 3 > k + 1 = 2. So, the second part of Theorem 13 tells us that the determi-
nant of the Koszul complex K• in degree greater than (2 + 8 + 3) − 3 = 10 determines exactly the
implicit equation of the scheme theoretic image of φ . Observe that, as was shown in [CD07, Thm. 2.5],
for this map, we have that deg(φ) = 1.
We remark that the process implemented for triangulating the matrix C via M is not algorithmic
for the moment.
Acknowledgments
I would like to thank Laurent Busé, and my two advisors: Marc Chardin and Alicia Dickenstein,
for the very useful discussions, ideas and suggestions. Also to the Galaad group at INRIA, for their
hospitality. Finally I would like to thank the reviewer for his huge dedication and many corrections to
this article.
References
[Avr81] Luchezar L. Avramov, Complete intersections and symmetric algebras, J. Algebra 73 (1) (1981) 248–263.
[BC05] Laurent Busé, Marc Chardin, Implicitizing rational hypersurfaces using approximation complexes, J. Symbolic Com-
put. 40 (4–5) (2005) 1150–1168.
[BCJ06] Laurent Busé, Marc Chardin, Jean-Pierre Jouanolou, Torsion of the symmetric algebra and implicitization, arXiv:math.
AC/0610186, Proc. Amer. Math. Soc., in press.
[BJ03] Laurent Busé, Jean-Pierre Jouanolou, On the closed image of a rational map and the implicitization problem, J. Alge-
bra 265 (1) (2003) 312–357.
[BM] Laurent Busé, Bernard Mourrain, Multires package, http://www-sop.inria.fr/galaad/logiciels/multires/.
[Bou07] N. Bourbaki, Éléments de mathématique. Algèbre. Chapitre 10. Algèbre homologique, Springer-Verlag, Berlin, 2007,
Reprint of the 1980 original [Masson, Paris; MR0610795].
[Bus06] Laurent Busé, Elimination theory in codimension one and applications, INRIA research report 5918, p. 47, notes of
lectures given at the CIMPA–UNESCO–IRAN school in Zanjan, Iran, July 9–22 2005, 2006.
[Cha93b] Marc Chardin, The resultant via a Koszul complex, in: Computational Algebraic Geometry, Nice, 1992, in: Progr. Math.,
vol. 109, Birkhäuser Boston, Boston, MA, 1993, pp. 29–39.
[Cha06] Marc. Chardin, Implicitization using approximation complexes, in: Algebraic Geometry and Geometric Modeling, in:
Math. Vis., Springer, Berlin, 2006, pp. 23–35.
[CD07] María Angélica Cueto, Alicia Dickenstein, Some results on inhomogeneous discriminants, in: Proc. XVI Coloquio Lati-
noamericano de Álgebra Biblioteca de la Revista Matemática Iberoamericana, ISBN 978-84-611-7907-7, 2007.
[De84] Michel Demazure, Les Notes Informelles de Calcul Formel: Une définition constructive du resultant, Sciences et Tech-
nologies de l’Information et de la Communication à l’X, FRE CNRS 2341, 1984–1994.
[GKZ94] I.M. Gel’fand, M.M. Kapranov, A.V. Zelevinsky, Discriminants, Resultants, and Multidimensional Determinants, Math.
Theory Appl., Birkhäuser Boston, Boston, MA, 1994.
[HSV82] J. Herzog, A. Simis, W.V. Vasconcelos, Approximation complexes of blowing-up rings, J. Algebra 74 (2) (1982) 466–493.
[HSV83] J. Herzog, A. Simis, W.V. Vasconcelos, Approximation complexes of blowing-up rings. II, J. Algebra 82 (1) (1983) 53–83.
[Jou91] J.-P. Jouanolou, Le formalisme du résultant, Adv. Math. 90 (2) (1991) 117–263.
[Jou95] J.-P. Jouanolou, Aspects invariants de l’élimination, Adv. Math. 114 (1995) 1–174.
[Vas94] Wolmer V. Vasconcelos, Arithmetic of Blowup Algebras, London Math. Soc. Lecture Note Ser., vol. 195, Cambridge
University Press, Cambridge, 1994.
Journal of Algebra
Standard monomial bases and degenerations of SOm (C)

representations
Sangjib Kim
Department of Mathematics, Cornell University, Ithaca, NY 14853-4201, USA
Article history: We study explicit weight bases for the finite dimensional represen-
Received 2 May 2008 tations of SO(m). Using a flat degeneration technique, we describe
Available online 20 August 2009 the representation spaces via a lattice cone associated with the
Communicated by Harm Derksen
Gelfand–Tsetlin patterns for GL(m). Combinatorial descriptions of
Keywords:
the weight bases for SO(m) relative to those for GL(m) will be
Young tableaux given.
Gelfand–Tsetlin patterns © 2009 Elsevier Inc. All rights reserved.
Standard monomials
Toric degenerations
1. Introduction
For m = 2n or 2n + 1, let O (m) and SO(m) be respectively the orthogonal group and the special or-
thogonal group of rank n over the complex numbers C. The main object of this paper is the following
quotient algebra, which we will call the isotropic flag algebra,

C[ Mm,n ]U n /ISO(m) ∼
= σmD
r ( D )n
where M m,n is the space of m by n matrices, U n is the upper triangular unipotent subgroup of the
general linear group GL(n), σmD is the O (m) irreducible representation indexed by Young diagram D,
and the sum is over all Young diagrams with less than or equal to n rows. For the labeling system of
the irreducible representations, we shall follow [7, §19.5]. The ideal ISO(m) turns out to be generated
by quadratic relations derived from the symmetric bilinear form fixed for SO(m), which explains the
name of the quotient algebra.
E-mail address: sangjib@math.cornell.edu.
doi:10.1016/j.jalgebra.2009.08.004
S. Kim / Journal of Algebra 322 (2009) 3896–3911 3897
As an SO(m) module, the isotropic flag algebra contains every finite dimensional irreducible repre-
sentation exactly once. Our motivational object is the flag algebra which is the ring of regular functions
over SO(m)/U SO(m) where U SO(m) is the maximal unipotent subgroup of SO(m) [18,19]. The isotropic
flag algebra is isomorphic to the flag algebra as an SO(m) module, while its multiplicative structure is
slightly different when m is even.
Starting from a standard monomial theory for the U n -invariant ring C[ M m,n ]U n , we shall develop
a relative theory of SO(m) embedded in GL(m) by computing a Gröbner basis for the ideal ISO(m) . The
description of generators for ISO(m) will be derived from the fundamental representations of Sp (2m)
by embedding O (m) in Sp (2m). As a result, we obtain explicit weight bases of polynomial type for
the finite dimensional representations σmD . Then, from a characterization of the leading monomials
for the elements of ISO(m) and C[ M m,n ]U n , we show that the isotropic flag algebra is a deformation
of a subvariety of an affine toric variety encoded by an integral lattice cone associated with the
Gelfand–Tsetlin semigroup. The combinatorial description of our weight bases for σmD is compatible
with known tableaux models studied in [16,22].
Geometrically, our result can be understood in the context of a toric degeneration of a flag variety
for SO(m) by considering the isotropic flag algebra as a multi-graded homogeneous ring. The toric de-
generations of spherical varieties are now well known (e.g., [1,4,9]). In particular, the SAGBI–Gröbner
method describes degenerations of affine varieties in a very concrete way (e.g., [2,5,23]), and using
this method the toric degenerations of the flag varieties for GL(m) and Sp (2m) have been studied (cf.
[13–15,20]). Our approach can be understood as an application of this method to the case of SO(m).
2. Preliminaries
Let us fix our notations and explain the basic ingredients for our investigation.
2.1. U n -invariant ring C[ M m,n ]U n
Let GL(m) × GL(n) be acting on the space of complex m by n matrices

M m,n = ( zi , j ): 1 i m; 1 j n
by
− 1
( g 1 , g 2 ) Q = g 1t Q g 2−1
for g 1 ∈ GL(m), g 2 ∈ GL(n), and Q ∈ M m,n . Let us assume m n. Then, under the GL(m) × GL(n) action,
the coordinate ring C[ M m,n ] of M m,n has the following decomposition:

C[ Mm,n ] ∼
= ρmD ⊗ ρnD
r ( D )n
where ρkD is the irreducible representation of GL(k) labeled by Young diagram D, and the summation
is over all D with the number of rows r ( D ) less than or equal to n. This result is known as the
GL(m)–GL(n) duality (see, e.g., [10,12]). If U n is the subgroup of GL(n) consisting of upper triangular
matrices with 1’s on the diagonal, then by taking U n ∼ = 1 × U n invariants, we have
U n
C[ Mm,n ]U n ∼
= ρmD ⊗ ρnD
r ( D )n

∼
= ρmD .
r ( D )n
The second isomorphism is by the theorem of highest weight (see, e.g., [10, §5]).
3898 S. Kim / Journal of Algebra 322 (2009) 3896–3911
This representation decomposition turns out to be compatible with the grading structure of the
algebra. Since the diagonal subgroup A n of GL(n) normalizes U n , C[ M m,n ]U n is stabilized by the action
of A n . Note that A n acts on (ρnD )U n by the character
l l
diag(a1 , . . . , an ) → a11 · · · ann
given by Young diagram D = (l1 l2 · · · ln 0) ∈ Zn . Thus, ρm

D
ρmD ⊗ (ρnD )U n is the space of An -
eigenvectors of weight D in C[ M m,n ]U n and the C-algebra C[ M m,n ]U n is graded by the semigroup A n+
+ +
of dominant polynomial weights for GL(n), or equivalently the subsemigroup A n ⊂ A m of dominant
weights for GL(m). Therefore, we have
ρmD ρmE ⊆ ρmD + E (2.1)
where D = (l1 , . . . , ln ), E = (l
1 , . . . , ln
), and D + E = (l1 + l
1 , . . . , ln + ln
).
2.2. Isotropic flag algebra
Let m = 2n or 2n + 1 with n 2. We shall review some properties of the irreducible representations

of SO(m) and their realization inside the GL(m) representations. Our main references are [7, §19]
and [16].
D
Let us consider the GL(m) irreducible representation ρm labeled by Young diagram D as an SO(m)
module by restricting GL(m) to its subgroup SO(m). For a Young diagram D with the number of rows
D
r ( D ) less than n, ρm contains the irreducible representation σmD of SO(m) labeled by the same Young
diagram with multiplicity one:
ρmD = σmD ⊕ ImD (2.2)
where Im D
is its complement space. The main difference between m = 2n and m = 2n + 1 occurs
when r ( D ) = n. For m = 2n + 1, this decomposition works without any change. However, if m = 2n
D
then ρm contains a pair of associated irreducible representations of SO(2n) with highest weights
D = (l1 , l2 , . . . , ln ) ∈ Zn>0 and D − = (l1 , l2 , . . . , ln−1 , −ln ). In this case, we will let σ2n
+ D
denote the
O (2n) irreducible representation with highest weight D = D , i.e., +
D D D+ −
σ2n = σ2n ⊕ σ2n (2.3)
and continue to use the notation in (2.2).

D = (1, . . . , 1) ∈ Zn then with respect to the standard action of GL(2n) and O (
Note that if 2n) on
C2n the space n C2n can be taken for both ρ2n D
and σ2nD
. Then, we may further decompose
n 2n
C
+ −
into D
σ2n ⊕ σ2n
D
by considering the +1 and −1 eigenspaces of the following map:
n n ∗ n
τ: C2n → C2n → C2n
where the first map is the isomorphism between C2n and (C2n )∗ given by the symmetric bilinear
form fixed for O (2n) and the second one is defined by the wedge product
n n 2n
C2n × C2n → C2n = C· e 1 ∧ · · · ∧ e 2n
See [10, §5.1.3] and [7, §19.2] for further details. For a combinatorial description of the individual
D+ D−
spaces σ2n and σ2n for a general D with r ( D ) = n, we refer to [16, §4].
The usual ordering can be given on A n+ , i.e., E is greater than D if E − D can be expressed by
a sum of positive roots. Then ImD
consists of the irreducible representations σmE of O (m) in ρm
D
where
E smaller than D. With this order and the grading structure (2.1), the following property can be easily
shown. See [7, §19.5]. A parallel result for the symplectic case is shown in [7, §17.3] and [14].
Proposition 2.1. For m = 2n or 2n + 1, with the notation given in (2.2) and (2.3), the sum

ISO(m) = ImD
r ( D )n
over all Young diagrams with r ( D ) n forms an ideal of C[ M m,n ]U n . Moreover, the quotient algebra
C[ Mm,n ]U n /ISO(m) , as an SO(m) module, contains every irreducible representation of SO(m) exactly once:

C[ Mm,n ]U n /ISO(m) ∼
= σmD . (2.4)
r ( D )n
We call the algebra C[ M m,n ]U n /ISO(m) the isotropic flag algebra for SO(m). The grading structure
(2.1) of C[ M m,n ]U n imposes the following grading structure on the isotropic flag algebra:

σmD + ISO(m) σmE + ISO(m) ⊆ σmD + E + ISO(m) , (2.5)
where D = (l1 , . . . , ln ), E = (l
1 , . . . , ln
), and D + E = (l1 + l
1 , . . . , ln + ln
).
2.3. Standard monomials for C[ M m,n ]U n
Let us briefly review a standard monomial theory for C[ M m,n ]U n with m n. A finite presentation
of C[ M m,n ]U n in terms of generators and relations is known: all the U n -invariant minors on M m,n
form a generating set and they satisfy the Plücker relations. More precisely, given Q = (t i , j ) ∈ M m,n
and I ∈ L (m, n) where

L (m, n) = [i 1 , i 2 , . . . , i p ]: 1 i 1 < · · · < i p m; 1 p n ,
we let δ I ( Q ) denote the minor of Q with the i 1 , i 2 , . . . , i p -th rows and the 1, 2, . . . , p-th columns:
⎡t ti1 2 · · · ti1 p ⎤
i1 1
⎢ ti2 1 ti2 2 · · · ti2 p ⎥
⎢ ⎥
δ I ( Q ) = det ⎢ . .. .. .. ⎥ .
⎣ .. . . . ⎦
ti p 1 ti p 2 · · · ti p p
Then the minors {δ I } indexed by L (m, n) generate C[ M m,n ]U n . We call I = [i 1 , i 2 , . . . , i p ] column

tableaux of length r ( I ) = p.
Definition 2.2. The tableau order on L (m, n) is a partial order defined as follows: for I =
[i 1 , i 2 , . . . , i p ] and J = [ j 1 , j 2 , . . . , jq ] in L (m, n), if p q and i r j r for all 1 r q, then I J .
c
l=1 δ I l ∈ C[ M m,n ] in {δ I : I ∈ L (m, n)} a standard monomial, if
Un
Let us call a monomial
I 1 , I 2 , . . . , I c form a multiple chain under the tableau order. The following result is known as a stan-
dard monomial theory. See, e.g., [3,6,9,11,17,20,21].
Theorem 2.3. The standard monomials form a C-vector space basis of C[ M m,n ]U n .
Here we note that every multiple chain I 1 I 2 · · · I c in L (m, n) corresponds to the semi-
standard Young tableau with columns I 1 , I 2 , . . . , I c , and every semistandard Young tableau can be
realized as a multiple chain in L (m, n) with respect to the tableau order. The weight bases for the
c ρm ⊂ C[ Mm,n ] can be Dalso deduced easily from the grading
D Un
individual irreducible representations
structure: a standard monomial l=1 δ I l is a weight basis for ρm if and only if the conjugate diagram
of (r ( I 1 ), r ( I 2 ), . . . , r ( I c )) is D, i.e., D = (r ( I 1 ), r ( I 2 ), . . . , r ( I c ))t . We refer to [14] for more details.
2.4. Leading monomial algebra
Next, we review some properties of the leading monomials of elements in C[ M m,n ]U n with respect
to the following monomial order.
Definition 2.4. (See [23].) Let >d be the total ordering on the coordinates za,b of M m,n given by za,i >d
zb, j if i < j, and za,i >d zb,i if a < b. The diagonal term order on C[ M m,n ] is the graded lexicographic
order on the set of monomials in C[ M m,n ] with respect to >d .
For a polynomial f ∈ C[ M m,n ], we let LM( f ) denote the leading monomial of f with respect to
the diagonal term order.
Let m = 5 and n = 4. For I 1 = [1, 2, 3, 5], I 2 = [1, 3, 4, 5], I 3 = [1, 3], I 4 = [3], the leading
Example 2.5.
4
monomial of l=1 δ I l is

4
4
LM δ Il = LM(δ I l )
l =1 l =1
= (z11 z22 z33 z54 )( z11 z32 z43 z54 )( z11 z32 )( z31 )
3 2 2
= z11 z31 z22 z32 z33 z43 z54
Note that it can be easily read from the corresponding tableau I 1 I 2 I 3 I 4 :
1 1 1 3
2 3 3
3 4
5 5
i.e., for each i and j, the degree of zi j is the number of i’s in the j-th row.
Definition 2.6. The leading monomial algebra in(C[ M m,n ]U n ) of C[ M m,n ]U n is the algebra generated by
{LM( f )} for all f ∈ C[ Mm,n ]U n .
Similarly, for X ⊂ C[ M m,n ]U n we let in( X ) denote the set of leading monomials of all elements in
X with respect to the diagonal term order. Then it can be easily shown that in(ISO(m) ) is an ideal
of in(C[ M m,n ]U n ). The leading monomials of C[ M m,n ]U n with respect to the diagonal term order have
the following nice properties.
Lemma 2.7. (See [14, §4.3].)
(i) For f ∈ C[ M m,n ]U n , its

leading monomial LM( f ) is equal to the leading monomial of a standard mono-
r
mial, i.e., LM( f ) = LM( l=1 δ I l ) for some I 1 · · · I r .
(ii) Distinct standard monomials of C[ M m,n ]U n have distinct leading monomials.
From the fact that our generators {δ I } form a SAGBI basis of C[ M m,n ]U n , i.e., {LM(δ I )} generates
in(C[ M m,n ]U n ), it can be shown that in(C[ M m,n ]U n ) is a flat degeneration of C[ M m,n ]U n . Moreover, if
m = n, the set of leading monomials can be identified with the set of Gelfand–Tsetlin patterns for
GL(m). Therefore, Spec(in(C[ M m,m ]U m )) can be understood as an affine toric variety encoded by the
Gelfand–Tsetlin patterns. See [9,14,15,20].
2.5. Gelfand–Tsetlin semigroup
We want to define Gelfand–Tsetlin patterns (cf. [8]) as order preserving maps defined on the fol-
lowing poset and impose a semigroup structure. The Gelfand–Tsetlin(GT) poset for GL(m) is
(i )
Γm = x j : 1 i m; 1 j i
(i +1) (i ) (i +1)
satisfying x j x j x j +1 for all i and j. We will draw it in a reversed triangular array:
(m) (m) (m)

x1 x2 ··· xm
(m−1) (m−1)
x1 ··· · · · xm −1
..
. ···
(2 ) (2 )
x1 x2
(1 )
x1
(i )
so that x j are decreasing from left to right along diagonals. Counting from bottom to top, we will
(k) (k) (k)
call x(k) = (x1 , x2 , . . . , xk ) the k-th row of Γm .
Definition 2.8. A GT pattern T of GL(m) is an order preserving map from Γm to the set of non-negative
integers:
T : Γm → Z0 ,
(k) (k)
and the k-th row of T is the image of x(k) , i.e., ( T (x1 ), . . . , T (xk )).
(k) (k) (k)

Note that the entries in the k-th row of T satisfy T (x1 ) T (x2 ) · · · T (xk ) for all k. If we
define the multiplication of two GT patterns as the addition of two order preserving maps, then the
set PGL(m) of all GT patterns of GL(m) forms a semigroup, and we will call it the GT semigroup. The
corresponding semigroup ring, called the GT semigroup ring, will be denoted by C[PGL(m) ].
There is a well-known bijection or conversion procedure between semistandard Young tableaux
and GT patterns compatible with successive branching rules of GL(k) down to GL(k − 1) for 2 k m
(see, e.g., [10, §8.1]). With the semigroup structure of GT patterns we have, only the correspondence
for column tableaux I ∈ L (m, m) is enough to define the bijection. Consider
I → T I (2.6)
where T I (Γm ) ⊂ {0, 1} such that the number of 1’s in the k-th row of T I is equal to the number of
entries in I which are less than or equal to k for 1 k m. Then the GT pattern corresponding to a
semistandard Young tableau or equivalently a multiple chain I 1 · · · I r in L (m, m) is the product of
the corresponding GT patterns [14, §3.4]:

r
I 1 · · · I r → T Il . (2.7)
l =1
Furthermore, as a consequence of Lemma 2.7, the bijection (2.7) provides the isomorphism be-
tween
r the leading monomial algebra and the GT semigroup ring for GL(m): for all standard monomials
l=1 δ I l of C[ M m,m ]
Um
,

in C[ M m,m ]U m → C[PGL(m) ],

r
r
LM δ Il → T Il . (2.8)
l =1 l =1
For l m, by restricting the isomorphism (2.8) to in(C[ M m,l ]U l ) ⊂ in(C[ M m,m ]U m ) we obtain the
following proposition.
Proposition 2.9. The leading monomial algebra in(C[ M m,l ]U l ) of C[ M m,l ]U l is isomorphic to the subring Am,l
(i )
of the GT semigroup ring C[PGL(m) ] consisting of GT patterns whose supports are in {x j ∈ Γm : 1 j l}.
For instance, the support of GT patterns appearing in A7,4 ∼

= in(C[ M 7,4 ]U 4 ) is
(7 ) (7 ) (7 ) (7 )
x1 x2 x3 x4
(6 ) (6 ) (6 ) (6 )
x1 x2 x3 x4
(5 ) (5 ) (5 ) (5 )
x1 x2 x3 x4
(4 ) (4 ) (4 ) (4 )
x1 x2 x3 x4
(3 ) (3 ) (3 )
x1 x2 x3
(2 ) (2 )
x1 x2
(1 )
x1
We omit its proof which is a mechanical calculation of the supports of GT patterns corresponding to
elements in L (m, l) under (2.6) and (2.7). A detailed proof for the case l = m is given in [14, §2.3].
2.6. Fundamental representations of Sp(2m)
The fundamental representations for the symplectic group Sp(2m) of rank m will be used to de-
scribe the generators of the ideal ISO(m) . We briefly review them and refer to [7,14] for further details.
Let us fix the skew-symmetric bilinear form , with respect to the elementary basis {εk } of C2m
as follows:
ε2i −1 , ε2i = − ε2i , ε2i −1 = 1 for 1 i m,

εa , εb = 0 otherwise.
p
The space C2m can be considered as the p-th fundamental
p 2m representation of GL(2m). For 2
p m, it contains the p-th fundamental representation ( C )prim of Sp(2m) exactly once:
p p p −2
C2m = C2m ⊕ω∧ C2m (2.9)
prim
m 2
where ω is the symplectic form i =1 ε2i −1 ∧ ε2i in C2m .
For I = [i 1 , i 2 , . . . , i p ] ∈ L (2m, m), let us set
p
ε I = εi 1 ∧ εi 2 ∧ · · · ∧ εi p ∈ C2m . (2.10)
Then the space spanned by {δ I ∈ C[ M 2m,m ]U m } where I ∈ L (2m, m) and r ( I ) = p is isomorphic to

p 2m
C under the correspondence between δ I and ε I . We have the following description of the lead-
ing monomials.
Proposition 2.10. (See [14, §5.2].) Let us fix the column tableau H 0 = [1, 3, 5, . . . , 2m − 1] ∈ L (2m, m).
p p −2
(i) For u r u ε I u ∈ C2m , if it is an element of ω ∧ ( C2m ), then the leading monomial of its corre-
sponding element u r u δ I u ∈ C[ M 2m,m ]U m is LM(δ I k ) for some I k H 0 .

(ii) Let ∈ C[ M 2m,m ]U m with r ( I u ) = p for all u be given. If LM(
u ru δ I u u ru δ I u ) = LM(δ I k ) for some
p −2 2m
I k H 0 , then u r u ε I u is in ω ∧ ( C ).
2.7. Symmetric bilinear form
For m = 2n or 2n + 1, we will use the following symmetric bilinear form ( , ). For the elementary
basis {er } of the space Cm ,
(1) if m = 2n, then (e s , et ) = 1 when (s, t ) = (2l − 1, 2l) or (2l, 2l − 1) for 1 l n and (e s , et ) = 0
otherwise;
(2) if m = 2n + 1, then (e s , et ) = 1 when (s, t ) = (2l − 1, 2l) or (2l, 2l − 1) or (2n + 1, 2n + 1) for
1 l n and (e s , et ) = 0 otherwise.
For a column tableau I ∈ L (m, n), we want to define its dual to be the column tableau under the
dual basis of {er } in (Cm )∗ with respect to the symmetric form ( , ). More precisely,
Definition 2.11. For I ∈ L (m, n), its dual I ∗ is the column tableau obtained from I by replacing 2i − 1
(resp. 2i) by 2i (resp. 2i − 1) for each i such that either 2i − 1 or 2i belongs to I for 1 i n.
3. SO(m) standard monomials
In this section, we describe a generating set of the ideal ISO(m) ⊂ C[ M m,n ]U n and study the leading
monomials of elements in ISO(m) by exploiting the embedding of O (m) in Sp(2m). As a result, we
obtain a standard monomial theory for SO(m). Throughout this and next sections, we assume m = 2n
or 2n + 1 with n 2.
3.1. O (m) and Sp(2m)
For a symmetric bilinear form ( , ) on Cm and a skew-symmetric bilinear form [ , ] on C2 , the space
C ∼
2m
= Cm ⊗ C2 has a natural skew-symmetric bilinear form , defined as follows:
v ⊗ w , v
⊗ w
= ( v , v
)[ w , w
]
for v , v
∈ Cm and w , w
∈ C2 . By letting e α ,β denote the elementary basis element e α ⊗ e β ∈ Cm ⊗C2 ,
we may realize the bilinear form , on C2m as follows: for 1 k n,
e 2k−1,1 , e 2k,2 = − e 2k,2 , e 2k−1,1 = 1;

e 2k,1 , e 2k−1,2 = − e 2k−1,2 , e 2k,1 = 1;
e i ,u , e j , v = 0 otherwise, (3.1)
with the additional condition e 2n+1,1 , e 2n+1,2 = − e 2n+1,2 , e 2n+1,1 = 1 for the case m = 2n + 1. We
can give an order on {e α ,β } by imposing the order of the elementary basis {εγ } for C2m under the
following bijection: for 1 i n,
e 2i −1,1 → ε4i −3 ;
e 2i ,2 → ε4i −2 ;
e 2i ,1 → ε4i −1 ;
e 2i −1,2 → ε4i , (3.2)
and we add e 2n+1,1 → ε4n+1 and e 2n+1,2 → ε4n+2 for the case m = 2n + 1. For example, if m = 5 then
we have
e 1,1 → ε1 e 1,2 → ε4 e 5,1 → ε9

e 2,1 → ε3 e 2,2 → ε2 e 5,2 → ε10
e 3,1 → ε5 e 3,2 → ε8
e 4,1 → ε7 e 4,2 → ε6
Note that for {e α ,2 : 1 α m}, we are using the dual basis by changing the order between e 2i −1,2
and e 2i ,2 for 1 i n. This description of the basis is compatible with the decomposition of C2m into
isotropic subspaces, i.e.,
Cm ⊕ (Cm )∗ ∼
= C2m ,
{e i ,1 } ∪ {e j ,2 } → {εγ }.
After identifying Cm and (Cm )∗ , we can extend this decomposition to the following injection
p q p +q
ι: Cm ⊗ Cm → C2m ι(e I ⊗ e J ) = ε I ∗ J (3.3)
where I = [i 1 , i 2 , . . . , i p ], J = [ j 1 , j 2 , . . . , j q ], and I ∗ J denotes the column tableau of length p + q

with entries corresponding to

(i 1 , 1), . . . , (i p , 1), ( j 1 , 2), . . . , ( jq , 2)
with respect to the bijection (3.2). See Example 3.3 below.

m ⊗2
2 2m (3.3), the symmetric form θ ∈ (C ) with respect to ( , )
We also note that under the embedding
maps to the symplectic form ω ∈ C with respect to , where

n
θ= (e 2i −1 ⊗ e 2i + e 2i ⊗ e 2i −1 ) + κ (e 2n+1 ⊗ e 2n+1 ),
i =1

m
ω= (ε2 j −1 ∧ ε2 j ). (3.4)
j =1
with κ = 0 if m = 2n; and κ = 1 if m = 2n + 1.
3.2. Generators of ISO(m)
For 1 l m/2, let ωl be the l-th fundamental weight

r of GL(m). Then, a copy of the irreducible
D
representation ρm of GL(m) with highest weight D = i =1 ωli can be obtained from the following
product:
l1 lr
Cm ⊗ · · · ⊗ Cm .
Recall that Im D
denotes the complement to the O (m) irreducible representation σmD in ρm D
. In
particular, for 1 d c n, let us consider the Young diagram (c , d)t with two columns of lengths c
t
and d. For 1 p < q r, we let
(l ,l )
Imp q denote the subspace of ImD spanned by

v
v
(l p ,lq )t
x1 ⊗ · · · ⊗ xup ⊗ ··· ⊗ xqu ⊗ · · · ⊗ xr : xup ⊗ xqu ∈ Im
u =1 u =1
li
where xi ∈ Cm for 1 i r. Since ImD can be obtained by intersecting ρmD with the space
m ⊗(| D |−2)
θ ⊗ (C ) generated by the symmetric form θ (see, for example, [7, §19]), the space Im D
is
l p ,lq )t l p m lq m
spanned by the subspaces
(
Im of quadratic relations coming from pairs C ⊗ C , i.e.,
(l p ,lq ) t
ImD =
Im .
1q< p r
Hence, we obtain the following result.
(c ,d)t
Lemma 3.1. The quadratic relations Im generate the ideal ISO(m) for 1 d c n.
We remark that these quadratic relations were studied in the language of tableaux in [16]. Our
(c ,d)t
next task is to characterize the elements of Im in terms of their leading monomials.
(c ,d)t
3.3. Leading monomials of Im
Note that the tensor space can be associated with the isotropic flag algebra via the following
Cartan product of GL(m) representations:
c d (c ,d) t
ζ: Cm ⊗ Cm → ρm ⊂ C[ Mm,n ]U n (3.5)
given by ζ (e I ⊗ e J ) = δ I δ J ∈ C[ M m,n ]U n . We also note that by Theorem 2.3, there is the section η of
ζ defined on the standard monomial basis {δ I δ J : I J }:
c d ζ t
Cm ⊗ Cm η ρm(c,d) ,
and consider the injection ϕ = ι ◦ η extending ι given in (3.3):
t c +d
ϕ : ρm(c,d) → C2m , ϕ (δ I δ J ) = ε I ∗ J . (3.6)
(c ,d)t (c ,d)t
The following proposition gives a characterization of the elements in Im ⊂ ρm in terms of their
leading monomials. We set H 0 = [1, 3, 5, . . . , 2m − 1] ∈ L (2m, m) as is given in Proposition 2.10.
t
Proposition 3.2. For a standard monomial δ I δ J in ρm(c,d) , if I ∗ J H 0 then LM(δ I δ J ) = LM( f ) for some
(c ,d)t (c ,d)t
f ∈ Im . Conversely, if f is an element in Im then LM( f ) = LM(δ I δ J ) for a standard monomial δ I δ J in
t
ρm(c,d) such that I ∗ J H 0 .
c (c ,d)t (c ,d)t
Proof. Recall that Cm ⊗ d Cm contains exactly one copy of ρm and that Im is the inter-
t
(c ,d) m ⊗(c +d−2) (c ,d)t (c ,d)t
section of ρm with θ ⊗ (C ) . Then, the correspondence (3.4) implies that Im ⊂ ρm
c +d−2 2m
is isomorphic to ϕ −1 (ω ∧ ( C )). Under the map ϕ defined in (3.6), a linear combination
of standard monomials δ I δ J maps to a linear combination of indecomposable wedge products ε I ∗ J .
(c ,d)t (c ,d)t
Therefore, the leading monomial of f ∈ Im ⊂ ρm can be characterized by the leading monomial
c+d−2 2m
of the element corresponding to ϕ ( f ) ∈ ω ∧ ( C ). The condition in the statements directly
follows from Proposition 2.10. 2
Example 3.3. For L (m, n), let m = 2n = 8. (i) Let us consider the standard monomial δ I δ J with I =
[1, 4, 5, 6] and J = [1, 4]. Then I ∗ J contains the elements corresponding to

(1, 1), (4, 1), (5, 1), (6, 1) ∪ (1, 2), (4, 2) ,
which is {1, 7, 9, 11} ∪ {4, 6} via (3.2). Therefore,
I ∗ J = [1, 4, 6, 7, 9, 11].
(4,2)t
Since I ∗ J H 0 , we have LM(δ I δ J ) = LM( f ) for any f ∈ Im .
(ii) On the other hand, if I = [1, 2, 5, 6] and J = [1, 3], then I ∗ J consists of {1, 3, 9, 11} ∪ {4, 8},
i.e.,
I ∗ J = [1, 3, 4, 8, 9, 11].
(4,2)t
Since the third element 4 is less than 5, I ∗ J H 0 , Therefore, LM(δ I δ J ) = LM( f ) for some f ∈ Im .
Here is a simple criterion for the leading monomial of a standard monomial to be divisible by
(c ,d)t
LM( g ) for some g ∈ Im .
r
Proposition 3.4. For I 1 · · · I r , let l=1 δ I l ∈ C[ Mm,n ]U n be given. Then its leading monomial is divisible
(c ,d)t
by LM( f ) for some f ∈ Im if and only if there is s such that I 1 ∗ I s H 0 for some 1 < s r.
r (c ,d)t
Proof. Let LM( l=1 δ I l ) be divisible by LM( f ) for f ∈ Im . Then by Lemma 2.7, we can assume
LM( f ) = LM(δ I u δ I s ) for some 1 u < s r. Moreover, by Proposition 3.2, I u ∗ I s H 0 , i.e., there is t
such that the t-th element of I u ∗ I s is less than 2t − 1. Since I 1 I u , the l-th entry in I 1 ∗ I s is less
than or equal to the l-th entry in I u ∗ I s for all l. Therefore, the t-th element of I 1 ∗ I s is also less than
2t − 1, i.e., I 1 ∗ I s H 0 . 2
3.4. SO(m) standard monomials
Let us define SO(m) standard monomials as follows.

r
l=1 δ I l of C[ M m,n
] r is SO(m) standard, if { I l : 1 l r } is a multiple
Un
Definition 3.5. An element
chain in L (m, n) and its leading monomial LM( l=1 δ I l ) with respect to the diagonal term order is not
divisible by leading monomials of elements in ISO(m) .
Our next task is to show the following standard monomial theory for the isotropic flag algebra:
Theorem 3.6. The SO(m) standard monomials in C[ M m ,nr]

Un
project to a vector space basis of
C[ Mm,n ] /ISO(m) . Moreover, the SO(m) standard monomials l=1 δ Il such that
Un
D = (l1 , l2 , . . . , lr )t
project to a basis of σmD ⊂ C[ M m,n ]U n /ISO(m) where lk = r ( I k ) for 1 k r.

r
We remark thatthe weight of l=1 δ I l can be read from the indexing tableau I 1 I 2 · · · I r ,
r
i.e., the weight of l=1 δ I l is ( w 1 , w 2 , . . . , w n ) where w k is equal to ν2k−1 − ν2k for all k and νi is the
number of i’s in the tableau for 1 i m. See [16,22] for more details.
Example 3.7. (i) If D = (1, 1, . . . , 1) ∈ Zk for k n, then every δ I with r ( I ) = k is SO(m) standard in
C[ Mm,n ]U n . Note that the space spanned by {δ I } with fixed length r ( I ) = k is isomorphic to the space
k
Cm which can be identified with σmD . See Theorem 5.1.7 in [10].
(ii) Let us consider the case of SO(5) with D = (2, 1). There are 40 standard monomials of the
form δ I δ J with r ( I ) = 2, r ( J ) = 1 and I , J ∈ L (5, 2). Among them only the followings 5 elements are
non-SO(5) standard:
{δ[1,2] δ[1] , δ[1,2] δ[2] , δ[1,3] δ[2] , δ[1,4] δ[2] , δ[1,5] δ[2] },
which agrees with the fact that the dimension of ρ5(2,1) is 40 and the dimension of σ5(2,1) is 35.
Now, let us consider a semistandard tableau Y D of shape D with entries from {1, 2, . . . , m}. We
let Y iD, j denote the entry in the i-th row and the j-th column of Y D , and α2k and β2k denote the
number of elements less than or equal to 2k in the first column and the second column of Y D
respectively. These notations and the following description of Proctor’s tableaux, which are tableaux
encoding weights of SO(m) representations, are from [16].
Lemma 3.8. (See [16,22].) The dimension of SO(m) representation σmD is equal to the number of semistandard
tableaux Y D of shape D satisfying the following conditions:
(i) α2k + β2k 2k for 1 k m;

(ii) if α2k + β2k = 2k for some k with Y αD ,1 = (2k − 1) and Y βD ,b = 2k for some b, then Y βD −1,b = 2k − 1.
2k 2k 2k
r
Recall that ρm D
= σmD ⊕ ImD from (2.2) and standard monomials l=1 δ Il whose indexing tableaux
I 1 · · · I r are of the shape D form a basis of ρm
D
(Theorem 2.3). Furthermore, by Lemma 2.7, the
D
leading monomials of these standard monomials form a basis for in(ρm ). Note that from the above
lemma, the dimension of Im is equal to the number of standard monomials in ρm
D D
that violate one of
the conditions in the above lemma. Now we characterize the leading monomials of Im D
.
r
Proposition 3.9. The dimension of Im
D
is equal to the number of standard monomials l=1 δ I l in ρmD whose
(l p ,lq )t
leading monomials are divisible by LM( f ) for some f ∈ Im with 1 p < q r.
Proof. From Proposition 3.2, it is enough to show that under the injection ϕ given in (3.6), the set
of tableaux Y D with two column tableaux I p I q violating any condition in Lemma 3.8 precisely
corresponds to the set of column tableaux I p ∗ I q such that I p ∗ I q H 0 where H 0 = [1, 3, 5, . . . ,
2m − 1]. Let us consider the semistandard Young tableau Y D with two columns I p I q such that
I p ∗ I q H 0 . Then there is k such that (i) (2k − 1)-th and (2k)-th entries in I p ∗ I q are (4k − 3) and
(4k − 2) respectively; or (ii) (2k)-th and (2k + 1)-th entries in I p ∗ I q are (4k − 1) and 4k respectively.
For the case (i), under (3.2), Y D satisfies either α2k + β2k > 2k or α2k + β2k = 2k with Y αD ,1 = 2k − 1
2k
and Y βD ,b = 2k. In the later case however, the second column does not contain (2k − 1). Therefore,
2k
Y D violates the second condition in Lemma 3.8. For the case (ii), we have α2k + β2k 2k + 1 in Y D .
Therefore Y D does not satisfy the first condition in Lemma 3.8. The other direction is also true by the
same argument. 2
(l p ,lq )t
Therefore, in(ImD
) is generated by in(Im ) for 1 p < q r, and we can conclude that our
generating set for ISO(m) is a Gröbner basis.
Corollary 3.10.
r
(i) For a standard monomial l=1 δ I l ∈ C[ Mm,n ]U n , its leading monomial is divisible by LM( f ) for some
(c ,d)t
f ∈ ISO(m) if and only if it is divisible by LM( g ) for some g ∈ Im for 1 d c n. Therefore,
in(ISO(m) ) is generated by
(c ,d)t
LM( f ): f ∈ Im , 1dc n .
r
(ii) A standard monomial l=1 δ I l ∈ C[ Mm,n ]U n is SO(m) standard if and only if it is not divisible by LM( f )
t
(c ,d)
for any f ∈ Im for 1 d c n.
Also, since as a vector space the dimension of σmD ∼ = ρmD /ImD is equal to the dimension of
in(ρD
m )/ in(Im
D
) (see, e.g., [2]), the dimension of σm
D
is equal to the number of SO(m) standard monomi-
r
als l=1 δ I l where D = (l1 , l2 , . . . , lr )t . The grading structure (2.5) of C[ M m,n ]U n /ISO(m) is compatible
with its representation decomposition (2.4). Hence, SO(m) standard monomials project to a basis of
C[ Mm,n ]U n /ISO(m) . This completes our proof for Theorem 3.6.
Finally, we have the following criterion for SO(m) standardness by using Proposition 3.4.
Corollary 3.11.
r
(i) The leading monomial of a standard monomial l=1 δ I l ∈ C[ M m,n ]U n is divisible by LM( f ) for some
f ∈ ISO(m) if and only
if I 1 ∗ I s H 0 for some 1 < s r.
r
(ii) A standard monomial l=1 δ I l ∈ C[ M m,n ]U n is SO(m) standard if and only if I 1 ∗ I s H 0 for all 1 < s r.
4. Flat degeneration and GT patterns
In this section, we study in(C[ M m,n ]U n )/in(ISO(m) ) as a flat degeneration of C[ M m,n ]U n /ISO(m) and
describe an integral lattice cone encoding its structure. Am,l is the subring of the GT semigroup ring
C[PGL(m) ] defined in Proposition 2.9.
Theorem 4.1. Spec(C[ M m,n ]U n /ISO(m) ) is a flat deformation of a subvariety of Spec(Am,n ).
Proof. Since in(C[ M m,n ]U n ) and in(ISO(m) ) are finitely generated, there is a positive weight vector α
that allows us to use the standard SAGBI–Gröbner deformation technique (see, e.g., [2,5]), i.e., there
exists a degree-filtration F α of C[ M m,n ]U n /ISO(m) given by α such that the associated graded ring of
the Rees algebra of C[ M m,n ]U n /ISO(m) with respect to F α is in(C[ M m,n ]U n )/in(ISO(m) ). This shows that
in(C[ M m,n ]U n )/in(ISO(m) ) is a flat degeneration of the isotropic flag algebra. 2
Recall that, by Proposition 2.9, in(C[ M m,n ]U n ) is isomorphic to a subring Am,n of the semigroup
(i )
ring C[PGL(m) ] generated by the GT patterns for GL(m) whose supports are in {x j ∈ Γm : 1 j n}.
Now, to obtain more explicit description of the integral lattice cone encoding the degeneration
of Spec(C[ M m,n ]U n /ISO(m) ), we want to identify the image of in(ISO(m) ) under the isomorphism
in(C[ M m,n ]U n ) Am,n .
Proposition 4.2. The quotient ring in(C[ M m,n ]U n )/in(ISO(m) ) is isomorphic to
Am,n /Θ
(i )
where Am,n is the semigroup ring over C generated by GT patterns with supports in {x j ∈ Γm : 1 j n}
and the ideal Θ is generated by T I + T J such that the sum of entries in the k-th row of T I + T J ∗ is greater than
k for some k.
Proof. From Proposition 2.9, we have in(C[ M m,n ]U n ) Am,n . Now we want to describe the ideal
in(ISO(m) ) in terms of GT patterns. From Corollary 3.10, it is enough to show that the condition on
I ∗ J in Proposition 3.2 is equivalent to the condition on T I + T J ∗ in the statement. Suppose I ∗ J H 0 ,
i.e., there is k such that the k-th element in I ∗ J is less than (2k − 1). Then it is equivalent that I ∗ J
contains more than (k − 1) elements which are less than (2k − 1) including, under the correspondence
(3.2), elements less than or equal to (k − 1) in I and in J ∗ respectively. This shows that the (k − 1)-th
row of T I + T J ∗ is greater than (k − 1). The converse is also true by the same argument. 2
Example 4.3. Let us consider the GT patterns corresponding to column tableaux given in Example 3.3.
(i) For I = [1, 4, 5, 6] and J = [1, 4], thus J ∗ = [2, 3]: the GT patterns T I and T J ∗ are, as SO(8) objects,
1 1 1 1 0 0 0 0 1 1 0 0 0 0 0 0
1 1 1 1 0 0 0 1 1 0 0 0 0 0
1 1 1 1 0 0 1 1 0 0 0 0
1 1 1 0 0 1 1 0 0 0
1 1 0 0 1 1 0 0
1 0 0 1 1 0
1 0 1 0
1 0
respectively, and therefore T I + T J ∗ is
2 2 1 1 0 0 0 0
2 2 1 1 0 0 0
2 2 1 1 0 0
2 2 1 0 0
2 2 0 0
2 1 0
2 0
1
Since the sum of entries in the k-th row is less than or equal to k for all k, the GT pattern T I + T J is
not in the ideal Θ . Note that we saw δ I δ J is SO(8) standard in Example 3.3.
(ii) On the other hand, for I = [1, 2, 5, 6] and J = [1, 3], thus J ∗ = [2, 4], the corresponding GT
patterns T I and T J ∗ are
1 1 1 1 0 0 0 0 1 1 0 0 0 0 0 0
1 1 1 1 0 0 0 1 1 0 0 0 0 0
1 1 1 1 0 0 1 1 0 0 0 0
1 1 1 0 0 1 1 0 0 0
1 1 0 0 1 1 0 0
1 1 0 1 0 0
1 1 1 0
1 0
respectively and T I + T J ∗ is
2 2 1 1 0 0 0 0
2 2 1 1 0 0 0
2 2 1 1 0 0
2 2 1 0 0
2 2 0 0
2 1 0
2 1
1
The sum of entries in the second row is 3. Therefore, T I + T J ∈ Θ . Note that we saw δ I δ J is not SO(8)
standard in Example 3.3.
Combined with Proposition 3.4, this shows that Spec(in(C[ M m,n ]U n )/in(ISO(m) )) is encoded by the
r (i )
semigroup of GT patterns l=1 T I l for GL(m) whose supports are in {x j ∈ Γm : 1 j n} and the
sum of entries in the k-th row of T I 1 + T I s∗ is less than or equal to k for all 1 k m and 1 < s r.
r
l=1 δ I l in ρm ⊂ C[ M m,n ] . From
D Un
Let us consider standard monomials Proposition 2.9, the iso-
r
morphism (2.8) provides the bijection between the leading monomials LM( l=1 δ I l ) of them and GT
patterns of GL(m) whose first rows are D. By describing GT patterns correspond to the complements
to in(Im D
), we obtain GT patterns correspond to SO(m) standard monomials of σmD ∼ = ρmD /ImD . In par-
ticular, we obtain the following corollary.
Corollary
r 4.4. The dimension of the SO(m) representation space σmD is equal to the number of GT patterns
l=1 T I l
of GL(m) such that their m-th rows are equal to D and that the sum of entries in the k-th row of
T I 1 + T I s∗ is less than or equal to k for 1 k m and 1 < s r.
Acknowledgments
The author expresses his sincere thanks to Roger Howe for many insightful conversations on this
project. The author is also very grateful to George Seligman and Eng-Chye Tan for their helpful com-
ments.
References
[1] V. Alexeev, M. Brion, Toric degenerations of spherical varieties, Selecta Math. (N.S.) 10 (4) (2004) 453–478.
[2] W. Bruns, A. Conca, Gröbner Bases, Initial Ideals and Initial Algebras. Homological Methods in Commutative Algebra, IPM
Proceedings, Teheran, 2004.
[3] W. Bruns, J. Herzog, Cohen–Macaulay Rings, Cambridge Stud. Adv. Math., vol. 39, Cambridge University Press, Cambridge,
1993.
[4] P. Caldero, Toric degenerations of Schubert varieties, Transform. Groups 7 (1) (2002) 51–60.
[5] A. Conca, J. Herzog, G. Valla, Sagbi bases with applications to blow-up algebras, J. Reine Angew. Math. 474 (1996) 113–138.
[6] C. De Concini, D. Eisenbud, C. Procesi, Young diagrams and determinantal varieties, Invent. Math. 56 (2) (1980) 129–165.
[7] W. Fulton, J. Harris, Representation Theory. A First Course, Grad. Texts in Math., vol. 129, Springer-Verlag, New York, 1991,
Readings in Mathematics.
[8] I.M. Gelfand, M.L. Tsetlin, Finite-dimensional representations of the group of unimodular matrices, Dokl. Akad. Nauk SSSR
(N.S.) 71 (1950) 825–828 (in Russian); English translation: I.M. Gelfand, in: S.G. Gindikin, V.W. Guillemin, A.A. Kirillov,
B. Kostant, S. Sternberg (Eds.), Collected Papers, vol. II, Springer-Verlag, Berlin, 1988, pp. 653–656.
[9] N. Gonciulea, V. Lakshmibai, Degenerations of flag and Schubert varieties to toric varieties, Transform. Groups 1 (3) (1996)
215–248.
[10] R. Goodman, N.R. Wallach, Representations and Invariants of the Classical Groups, Encyclopedia Math. Appl., vol. 68, Cam-
bridge University Press, Cambridge, 1998.
[11] W. Hodge, Some enumerative results in the theory of forms, Proc. Cambridge Philos. Soc. 39 (1943) 22–30.
[12] R. Howe, Perspectives on invariant theory: Schur duality, multiplicity-free actions and beyond, in: The Schur Lectures, Tel
Aviv, 1992, in: Israel Math. Conf. Proc., vol. 8, Bar-Ilan Univ., Ramat Gan, 1995, pp. 1–182.
[13] K. Kaveh, SAGBI Bases and degeneration of spherical varieties to toric varieties, Michigan Math. J. 53 (1) (2005) 109–121.
[14] S. Kim, Standard monomial theory for flag algebras of GL(n, C) and Sp(2n, C), J. Algebra 320 (2) (2008) 534–568.
[15] M. Kogan, E. Miller, Toric degeneration of Schubert varieties and Gelfand–Tsetlin polytopes, Adv. Math. 193 (1) (2005) 1–17.
[16] R.C. King, T.A. Welsh, Construction of orthogonal group modules using tableaux, Linear Multilinear Algebra 33 (3–4) (1993)
251–283.
[17] V. Lakshmibai, The development of standard monomial theory. II, in: A Tribute to C.S. Seshadri, Chennai, 2002, in: Trends
Math., Birkhäuser, Basel, 2003, pp. 283–309.
[18] G. Lancaster, J. Towber, Representation-functors and flag-algebras for the classical groups. I, J. Algebra 59 (1) (1979) 16–38.
[19] G. Lancaster, J. Towber, Representation-functors and flag-algebras for the classical groups. II, J. Algebra 94 (2) (1985) 265–
316.
[20] E. Miller, B. Sturmfels, Combinatorial Commutative Algebra, Grad. Texts in Math., vol. 227, Springer-Verlag, New York, 2005.
[21] C. Musili, The development of standard monomial theory. I, in: A Tribute to C.S. Seshadri, Chennai, 2002, in: Trends Math.,
Birkhäuser, Basel, 2003, pp. 385–420.
[22] R.A. Proctor, Young tableaux, Gelfand patterns, and branching rules for classical groups, J. Algebra 164 (2) (1994) 299–360.
[23] B. Sturmfels, Gröbner Bases and Convex Polytopes, Univ. Lecture Ser., vol. 8, Amer. Math. Soc., Providence, RI, 1996.
Journal of Algebra
New families of perfect nonlinear polynomial functions ✩

Zhengbang Zha a , Xueli Wang b,∗
a
School of Mathematical Sciences, Luoyang Normal University, Luoyang 471022, China
b
School of Mathematical Sciences, South China Normal University, Guangzhou 510631, China
Article history: A new class of almost perfect nonlinear (APN) polynomial functions
Received 23 May 2008 has been recently introduced. We give some generalizations of
Available online 23 May 2009 these functions and deduce new families of perfect nonlinear (PN)
Communicated by Gerhard Hiss
functions. We show that these PN functions are CCZ-inequivalent
Keywords:
to the known perfect nonlinear functions.
Almost perfect nonlinear functions © 2009 Elsevier Inc. All rights reserved.
Perfect nonlinear functions
Planar functions
Differential uniformity
1. Introduction
Let f (x) be a map GF ( pn ) → GF ( pn ). Let N (a, b) denote the number of solutions x ∈ GF ( pn ) of

f (x + a) − f (x) = b where a, b ∈ GF ( pn ), and let f = max{ N (a, b) | a, b ∈ GF ( pn ), a = 0}. Nyberg [17]
defined a map f to be differentially k-uniform if f = k. For applications in cryptography, one would
like to employ functions for which f is as small as possible. The differential 2-uniform function is
called an almost perfect nonlinear (APN) function, and we know that APN functions have the lowest
possible differential uniformity over GF (2n ). This is of interest in cryptography, since differential and
linear cryptanalysis exploit weaknesses in the uniformity of functions, which are used in many block
ciphers, such as DES. When k = 1, the differential 1-uniform function is called a perfect nonlinear
(PN) function. It is notable that differential 1-uniform functions have also been studied under the
name of planar functions, which are functions such that f (x + a) − f (x) is a permutation polynomial
for all a ∈ GF ( pn )∗ = GF ( pn ) \ {0}. In geometry, PN functions are known as planar functions. Planar
functions were introduced in Ref. [9] to describe projective planes with certain properties. In recent
✩
This work was partially supported by the National Natural Science Foundation (China) under Grant 10771078.
E-mail addresses: zzb322@yahoo.com.cn (Z. Zha), wangxuyuyan@yahoo.com.cn (X. Wang).
doi:10.1016/j.jalgebra.2009.04.042
Z. Zha, X. Wang / Journal of Algebra 322 (2009) 3912–3918 3913
papers [6,7], planar functions were used to describe new finite commutative semifields of odd order.
In prior studies [13,18], it was shown that a planar function yields either a skew Hadamard difference
set or a Paley type partial difference set depending on pn (mod 4). Additionally, in Refs. [11,12] planar
and APN functions are used to construct optimal constant-composition codes and signal sets.
Some polynomial PN functions and low differential uniformity functions have been recently discov-
ered (e.g., [3,8,13,14,16,19]), although it is very difficult to find more new inequivalent PN functions.
In this paper, we find that some APN functions defined over even characteristic fields can be general-
ized to fields with odd characteristics and they are proved to be PN functions. These new families of
PN functions are not EA-equivalent to the known ones. Our paper is organized as follows: in Section 2,
a brief overview of known results and some preliminary results (needed later in the paper) will be
presented. Also our Main Theorem will be presented in Section 2. In Section 3, some new families
of PN functions will be constructed, whose headspring comes from some known APN functions dis-
cussed in Section 2, and we show that these PN functions are always inequivalent to the known PN
monomials. In Section 4, we discuss the EA-equivalence of the new PN functions.
2. Preliminaries
Let p be a prime number. The p-weight of a nonnegative integer m is the sum of thedigits in
its p-adic representation, i.e., if m = i b i p i with 0 b i < p then the p-weight of m is i b i ∈ Z.
Recall that any function of GF ( pn ) can be represented by a polynomial over GF ( pn ) of degree less
than pn . Moreover, different polynomials define different functions. This allows us to identify the set
of functions of GF ( pn ) with the set of polynomials over GF ( pn ) with degree less than pn . The algebraic
degree of a polynomial over GF ( pn ) is the maximal p-weight of the exponents of its nonzero terms.
The GF ( p )-linear mappings L : GF ( pn ) → GF ( pn ) are represented by polynomials of algebraic de-
n−1 i
gree one with zero constant term, that is L (x) = i =0 c i x p , c i ∈ GF ( pn ). Such polynomials are called
linearized or p-polynomials. The sum of a linear function and a constant in GF ( pn ) is called an affine
function.
Two functions, F , G : GF ( pn ) → GF ( pn ) are called extended affine equivalent (EA-equivalent), if G =
A 1 ◦ F ◦ A 2 + A for some affine permutations A 1 , A 2 and affine function A. EA-equivalent nonconstant
functions have the same algebraic degree.
Two functions, F and G from GF ( pn ) to itself are called Carlet–Charpin–Zinoviev equivalent (CCZ-
equivalent) if the graphs of F and G are affine equivalent. CCZ-equivalent functions have the same
differential uniformity and the same extended Walsh spectrum [4].
In Ref. [5], it was shown that CCZ-equivalent functions have equal differential uniformity and that
EA-equivalence is a particular case of CCZ-equivalence. CCZ-equivalence is a coarser equivalence re-
lation for APN functions than EA-equivalence, but CCZ-equivalence and EA-equivalence are the same
for PN functions according to Ref. [15].
Let p be an odd prime number. The following are the currently known EA-inequivalent PN func-
tions:
(a) x2 in GF ( pn ) (folklore);
(b) x p +1 in GF ( pn ), k n/2 and n/(k, n) is odd [8,9];
k
(c) x10 + x6 − x2 in GF (3n ), n 5 is odd [8];

(d) x10 − x6 − x2 in GF (3n ), n 5 is odd [13];
2k+s
(e) x p +1 − u p −1 x p + p
s k k
in GF ( p 3k ), where (k, 3) = 1, k − s ≡ 0 (mod 3), s = k and k/(k, s) is odd,
and u is a primitive element of GF ( p 3k ) [19];
(f) x(3 +1)/2 in GF (3n ), k 3 is odd and (k, n) = 1 [8,14].
k
Note that the functions in (a)–(e) are of shape

n −1

ai , j x p + p ,
i j
ai , j ∈ GF pn .
i , j =0
3914 Z. Zha, X. Wang / Journal of Algebra 322 (2009) 3912–3918
The polynomials of this type are called Dembowski–Ostrom polynomials. Bierbrauer informed us that
the functions in (e) are also PN functions over GF ( p 4k ) where 4k/(4k, s) is odd, among other addi-
tional conditions.
In Ref. [1], a new family of APN functions were presented, given by the following:
k+s
Proposition 2.1. Let f be the function on GF (22k ) defined by f (x) = bx2 +1 + b2 x2 +2 + cx2 +1 +
s k k k
k−1 i +k
+2i where b , c
i =0 r i x2 ∈
/ GF (2k ), b is not a cube, and r i ∈ GF (2k ) for each i. Then f is APN over GF (22k ).
Dillon introduced quadratic hexanomials of the type F (x) = x( Ax2 + Bxq + C x2q ) + x2 ( Dxq + Ex2q ) +
Gx3q over GF (22m ) with q = 2m in Ref. [10]. If coefficients A , B , C , D , E , F , G satisfy some additional
conditions, F (x) can have differential uniformity of, at most, 4. In Ref. [2] Budaghyan and Carlet gave
various generalizations of this method and deduced the constructions of new infinite classes of APN
quadratic trinomials and hexanomials from GF (22m ) to GF (22m ) by the following propositions.
Proposition 2.2. Let m and i be any integers, q = 2m , n = 2m, gcd(i , m) = 1 and c , b ∈ GF (2n ) such that
cq+1 = 1, c ∈
/ {λ(2 +1)(q−1) , λ ∈ GF (2n )}, cbq + b = 0. Then the function F (x) = x2 +2 + bxq+1 + cxq(2 +2 )
i 2i i 2i i
n
is an APN function over GF (2 ).
Proposition 2.3. Let n be any even integer, q = 2n/2 , gcd(i , n/2) = 1, and c , s ∈ GF (2n ) such that s ∈
/ GF (q).
If the equation x2 +1 + cx2 + cq x + 1 = 0 has no solution x such that xq+1 = 1, and in particular if the
i i
polynomial X 2 +1 + c X 2 + cq X + 1 is irreducible over GF (2n ), then the function F (x) = x(x2 + xq + cx2 q ) +
i i i i
x2 (cq xq + sx2 q ) + x(2 +1)q is APN on GF (2n ).

i i i
We now present our main result:
Main Theorem. Let p be an odd prime, m, i are two positive integers with n = 2m, 0 < i < n, such that
n/ gcd(i , n) is an odd and q = pm . Let b, c ∈ GF ( pn ) such that cbq = b, cq+1 = 1, c e = 1, where e = (q + 1)/
gcd(q + 1, p i + 1). Then, the function F (x) = x p + p + bxq+1 + cxq( p + p ) defined over GF ( pn ) is a PN
2i i 2i i
function, which is CCZ-inequivalent to any of (a)–(f) listed above.
3. New PN functions over GF ( p n )
In this section, we will give some PN functions over GF ( pn ) with p an odd prime. First, we describe
a new family of PN trinomial functions over GF ( pn ), which is an analog of an odd characteristic field
of Proposition 2.2.
Proposition 3.1. In the case of the Main Theorem, the function F (x) is a PN function.
Proof. We need to count the number of solutions of F (x + a) − F (x) = b under the conditions defined
above for any a = 0, b in GF ( pn ). Equation F (x + a) − F (x) = b can be written as
+ pi
+ b(x + a)q+1 + c (x + a)q( p + pi ) + pi
− bxq+1 − cxq( p + pi )
2i 2i 2i 2i
(x + a) p − xp = b. (1)
Let = b − F (a). Then Eq. (1) becomes
i 2i 2i i i 2i 2i i
a p x p + a p x p + abxq + aq bx + caqp xqp + caqp xqp = . (2)
As Eq. (2) is affine, we just need to consider the case = 0. When = 0, replacing x by ua, from
Eq. (2) we get
+ pi
i i
+ baq+1 uq + u + caq( p + p ) uqp + uqp = 0,
2i 2i i 2i 2i
= ap up + up (3)
and
i i i i
c q = caqp +qp u qp + u qp + cbq aq+1 u q + u + c q+1 a p + p u p + u p = 0.
2i 2i 2i 2i
(4)
We get (cbq − b)aq+1 (u q + u ) = 0 by c q − = 0 and cq+1 = 1. Since a = 0 and cbq = b, u q = −u. We

replace u q by −u in Eq. (3) and get
2i 2i
a p + p − caq( p + p ) u p + u p = 0.
i 2i i i
(5)
If a p + p − caq( p + p ) = 0, then c = a(1−q)( p + p ) . This contradicts the initial assumption

i 2i i 2i i 2i
c (q+1)/ gcd(q+1, p +1) = 1. So, u + u p = 0. If n/ gcd(i , n) is odd, −1 is not a ( p i − 1)th power in GF ( pn ),

i i
then we get one solution u = 0 of Eq. (5) and F (x) is a PN function over GF ( pn ). 2
Next, we consider the EA- and CCZ-equivalence between the PN trinomial function in Proposi-
tion 3.1 and any of the PN monomials listed in (a)–(f).
Proposition 3.2. The PN function F (x) = x p + p + bxq+1 + cxq( p + p ) in Proposition 3.1 is EA-inequivalent
2i i 2i i
p r +1
to any Dembowski–Ostrom monomial g (x) = x n
over GF ( p ).
Proof. When r = 0, we assume that F (x) is EA-equivalent to x p +1 . From the definition of EA-
r
n−1 j n−1 j
equivalence, there exist linear permutations L 1 (x) = j =0 b j x p , L 2 (x) = j =0 c j x p and an affine
function L ∗ (x) such that L 1 (x p +1 ) = F ( L 2 (x)) + L ∗ (x), which can be written as
r
n −1
n −1
n −1

r p j p 2i p i j +2i
+ p l +i qp 2i qp i p j +2i +m + pl+i +m
b j x p +1 = c j cl x p +c cj cl x
j =0 j ,l=0 j ,l=0

n −1
j +m
+ pl
+ L ∗ (x).
q
+b c j cl x p (6)
j ,l=0
Since cq+1 = 1,
n −1
n −1

n −1
p m p r +1 p r p j q
j +m j +m
b j x p +1 + pl
+ cL ∗ (x)q − L ∗ (x)
q
c bj x − = cb − b c j cl x p (7)
j =0 j =0 j ,l=0
j +m
from c (6)q − (6). Since cbq = b, when m + j − l = ±r, there is no term of type x p + pl on the left
side of Eq. (7). Therefore,
pm pm
cl c j + c j +m cl−m = 0. (8)
pm
Since r = 0, assume l = m + j, such that 2cm+ j c j = 0 from Eq. (8). Assume c j = 0 for some j, such
that c j +m = 0. From Eq. (8) we obtain cl = 0 for m + j − l = ±r. Since c j = 0, then m = ±r. This is
not true for x p +1 , which is not a PN function in GF ( p 2m ). This means cl = 0 for any 0 l n − 1
m
and L 2 (x) = 0 is not a permutation. This is the contradiction. Thus, F (x) is not EA-equivalent to x p +1
r
when r = 0.
When r = 0, we assume that F (x) is EA-equivalent to x2 . Therefore, there exist linear permutations
n−1 n−1
c j x p and an affine function L ∗ (x) such that
j j
L 1 (x) = j =0 b j x p , L 2 (x) = j =0

n −1
n −1
2i i i p j
b j x p + p + bxq+1 + cxq( p + p ) c j cl x p + p + L ∗ (x).
2i j l
= (9)
j =0 j ,l=0
j j
We can see that there are no terms of form x2p on the left side of Eq. (9), so the coefficients of x2p
from the right side of Eq. (9) must equal zero. Then c 2j = 0 and c j = 0 for any 0 j n − 1. Obviously,
L 2 (x) = 0 is not a permutation. Therefore, we also see that F (x) is not EA-equivalent to x p +1 when
r
r = 0. The proof is completed. 2
Proposition 3.3. Let p be an odd prime. The PN function F (x) defined in Proposition 3.1 is EA-inequivalent to
any of the known PN functions listed in (a)–(f).
Proof. From Proposition 3.2, we know that the PN function F (x) is not EA-equivalent to either x p +1
r
3k +1
or x2 . Next we consider the other classes of PN functions. Since x 2 is not quadratic and it is not
affine, then the function EA-equivalent to it must not be quadratic. But the PN function F (x) of Propo-
3k +1
sition 3.1 is quadratic, so F (x) is not EA-equivalent to x 2 . And we note that the other PN functions
(c) and (d) only exist in GF (3n ) where n 5 is odd. But the PN function F (x) in Proposition 3.1 is
defined in GF ( p 2m ).
Now we consider the PN functions in case (e). We assume that the PN function F (x) in Proposi-
−k ∗ ∗ ∗ +s ∗
tion 3.1 is EA-equivalent to x p +1 − u p −1 x p + p in GF ( ptk ) where t = 3 or 4 and tk∗ / gcd(tk∗ , s)
s k k
n−1 j n−1 j
is odd. Therefore, there exist linear permutations L 1 (x) = j =0 b j x p , L 2 (x) = j =0 c j xp and an
affine function L ∗ (x) such that
n −1
n −1
n −1

s ∗−k ∗ ∗ +s p j p 2i p i j +2i
+ p l +i qp 2i qp i p j +2i +m + pl+i +m
b j x p +1 − u p −1 x p + p
k k
= c j cl x p +c cj cl x
j =0 j ,l=0 j ,l=0
n −1
j +m
+ pl
+ L ∗ (x)
q
+b c j cl x p (10)
j ,l=0
with 2m = tk∗ . From c (10) p − (10), we see

m

n −1 n −1

p m p s +1 k∗ p j+m s p j
∗ ∗ +s ∗ ∗ ∗ +s
−1 pk + p −k −k
b j x p +1 − u p −1 x p + p
k k
c bj x − up x −
j =0 j =0
n −1

j +m
+ pl
+ cL ∗ (x)q − L ∗ (x).
q
= cbq − b c j cl x p (11)
j ,l=0
ph pm+h ph pm+h
Since cbq = b, if j + m − l = ±s, ±(2k∗ − s), we get cl c j + c j +m cl−m = 0. When t = 3, s = 2k∗ from
the known conditions (k, 3) = 1 and k − s ≡ 0 (mod 3). Since 0 = ±s, ±(2k∗ − s), when j + m = l, we
pm
obtain c j c j +m = 0. If c j = 0 for some j, then c j +m = 0 and cl = 0 for any j + m − l = ±s, ±(2k∗ − s).
Since c j = 0, m = ±s or m = ±(2k∗ − s). These contradict the fact that tk∗ /(tk∗ , s) is odd. Therefore,
c j = 0 for any 0 j n − 1 and L 2 (x) is not an affine permutation. We have therefore proved that
the PN function F (x) in Proposition 3.1 is not EA-equivalent to case (e). The proof is completed. 2
Proof of Main Theorem. We have shown that the PN function F (x) in Main Theorem is EA-
inequivalent to any of the known PN functions listed in (a)–(f), and the CCZ-equivalence coincides
with the EA-equivalence for PN functions [15]. So the PN function F (x) is CCZ-inequivalent to any of
the known PN functions listed in (a)–(f), which completes the proof. 2
4. EA-equivalence of the new PN functions
We can give the analog of Proposition 2.1 as follows. The proof is just like the one of Proposi-
tion 3.1, which is omitted.
m +i m+h
Proposition 4.1. Let f (x) be the function defined on GF ( p 2m ) by f (x) = bx p +1 + b p x p + p + cx p + p +
i m m h
2m−1 j +k
r j x p + p where c ∈
/ GF ( pm ), b = λ p +1 for any λ ∈ GF ( p 2m ), 0 h 2m − 1 and r j ∈ GF ( pk ) for
j i
j =0
each j. Let 0 < i < 2m and gcd(i , 2m) = t. Then f (x) is a PN function when 2m/t is odd.
We can also show that the PN function in Proposition 4.1 is EA-equivalent to the PN function F (x)
in Main Theorem. In order to do this, we need the following lemma.
2m−1 l
Lemma 4.1. Let L 1 (x) be the linear function on GF ( p 2m ) defined by L 1 (x) = l=0 bl x p where bl ∈ GF ( p 2m ),
m m
p p
b0 = 1 − bm , bm + bm = 1 and bl+m = bl = −bl for 0 < l < m. Then L 1 (x) is a linear permutation.
Proof. For any x1 , a ∈ GF ( p 2m ), assume L 1 (x1 + a) = L 1 (x1 ), then we can get

2m −1
l
b l a p = 0. (12)
l =0
m
By subtracting Eq. (12) from the pm -power of Eq. (12) (we write simply as (12) p − (12)), we find
m m m m pm
that (b0 a + bm a p ) p − b0 a − bm a p = 0. Since b0 = 1 − bm , then (a − a p )(bm + bm − 1) = 0. Since
m m
p
bm + bm = 1, a = a p . From Eq. (12) we get b0 a + bm a = a = 0. Therefore, a = 0 and L 1 (x) is a
permutation. The proof is completed. 2
Similarly, we can obtain the following lemma.
i i +m j
Lemma 4.2. Let L 2 (x) be the linear function on GF ( p 2m ) defined by L 2 (x) = b1 x p + (1 − b1 )x p + b2 x p −
j +m pm pm
b2 x p , where b1 , b2 ∈ GF ( p 2m ), 1 − b1 = b1 , b2 + b2 0 and 0 i , j 2m − 1. Then L 2 (x) is a linear
=
permutation.
m +i
Proposition 4.2. Let b, c be as in Proposition 4.1, then the PN function g (x) = bx p +1 + b p x p + p +
i m m
m +i
+ p i is EA-equivalent to function f (x) = bx p i +1 m m +i
+ pm m +i
+ pi
2m−1 l+m
+ pl .
cx p + b p xp + cx p + l=0 rl x p
2m−1 l r i +r i +m
Proof. Since c ∈
/ GF ( pm ), we can find a linear function A (x) = l=0 bl x p , where bm = m ,
c p −c
rl+i +rl+i +m pm
b 0 = 1 − b m , bl = l l+m and bl+m = −bl for 0 < l < m. A direct computation shows that bl =
c p −c p
−bl = bl+m with l = 0. From Lemma 4.1, we know that A (x) is a linear permutation. We con-
2m−1 m +i m +i
sider the function A ( g (x)) = l=0 bl (bx p +1 + b p x p + p + cx p + p ) p . The coefficients of the
i m m i l
l+m
terms x p (1+ p i ) with l
= 0, m are bm+l bm+l + bl bm+l = 0. The coefficient of x1+ p is b0 b + bm b = b,
i
l +i
and the coefficient of x p (1+ p ) is b p . When l = 0, the coefficients of the terms x p (1+ p ) are
m i m m
l l+m
l+m c p (rl+i +rl+i +m ) cp (rl+i +rl+i +m )
= rl+i + rl+i +m , and the coefficient of x p (1+ p
l i m
bl c p + bl+m c p = l l+m − l l+m
) is
c p −c p c p −c p
m
b0 c + bm c p = c + r i + r i +m . Therefore, A ( g (x)) = f (x). From the definition of EA-equivalence, we see
that the function g (x) is EA-equivalence to f (x). 2
Remark 4.1. Similar to the proof of Proposition 4.2, we can also show, for the functions in Propo-
m +i m +i
sition 4.1, that the PN function f (x) = bx p +1 + b p x p + p + cx p + p is EA-equivalent to another
i m m i
m +i m+s
function h(x) = bx p +1 + b p x p + p + cx p + p with s = i , i + m. By Lemma 4.2 we have a linear per-
i m m s
m s −i s−i +m
cp c m c c
mutation A 1 (x) = m x − m xp + s −i s−i +m xp − s −i s−i +m xp . A direct computation
c p −c c p −c cp −c p cp −c p
pi p 2i + p i pm −1 pm ( p 2i + p i ) m
+1 ),
shows that A 1 ( f (x)) = h(x), and s = m − i, such that h(x ) = b(x +b x + bc x p
i
which shows that h(x p ) is equivalent to a PN functions F (x) in Proposition 3.1.
In Ref. [6], it is shown that the classification of finite presemifields of an odd order and the
one of PN Dembowski–Ostrom polynomials are equivalent. That is to say, in certain cases, the PN
Dembowski–Ostrom polynomials define isotopic presemifields if, and only if, they are EA-equivalent.
Then a new presemifield can be defined by the Main Theorem, which is not isotopic to any previously
known one.
Acknowledgments
The authors would like to thank the anonymous referees for their helpful suggestions that much
improved the organization and presentation of this paper.
References
[1] C. Bracken, E. Byrne, N. Markin, G. McGuire, New families of quadratic almost perfect nonlinear trinomials and multinomi-
als, Finite Fields Appl. 14 (2008) 703–714.
[2] L. Budaghyan, C. Carlet, Classes of quadratic APN trinomials and hexanomials and related structures, IEEE Trans. Inform.
Theory 54 (2008) 2354–2357.
[3] L. Budaghyan, C. Carlet, G. Leander, Two classes of quadratic APN binomials inequivalent to power functions, IEEE Trans.
Inform. Theory 54 (2008) 4218–4229.
[4] L. Budaghyan, C. Carlet, A. Pott, New classes of almost bent and almost perfect nonlinear polynomials, IEEE Trans. Inform.
Theory 52 (2006) 1141–1152.
[5] C. Carlet, P. Charpin, V. Zinoviev, Codes, bent functions and permutations suitable for DES-like cryptosystems, Des. Codes
Cryptogr. 15 (1998) 125–156.
[6] R.S. Coulter, M. Henderson, Commutative presemifields and semifields, Adv. Math. 217 (2008) 282–304.
[7] R.S. Coulter, M. Henderson, P. Kosick, Planar polynomials for commutative semifields with specified nuclei, Des. Codes
Cryptogr. 44 (2007) 275–286.
[8] R.S. Coulter, R.W. Matthews, Planar functions and planes of Lenz–Barlotti class II, Des. Codes Cryptogr. 10 (1997) 167–184.
[9] P. Dembowski, T. Ostrom, Planes of order n with collineation groups of order n2 , Math. Z. 103 (1968) 239–258.
[10] J.F. Dillon, APN polynomials and related codes, in: Polynomials over Finite Fields and Applications, Banff International
Research Station, November, 2006.
[11] C. Ding, J. Yin, Signal sets from functions with optimum nonlinearity, IEEE Trans. Commun. 55 (2007) 936–940.
[12] C. Ding, J. Yuan, A family of optimal constant-composition codes, IEEE Trans. Inform. Theory 51 (2005) 3668–3671.
[13] C. Ding, J. Yuan, A new family of skew Paley–Hadamard difference sets, J. Combin. Theory Ser. A 113 (2006) 1526–1535.
[14] T. Helleseth, D. Sandberg, Some power mappings with low differential uniformity, Appl. Algebra Engrg. Comm. Comput. 8
(1997) 363–370.
[15] G. Kyureghyan, A. Pott, Some remarks on planar mappings, in: Proceedings of WAIFI 2008, in: Lecture Notes in Comput.
Sci., vol. 5130, 2008, pp. 115–122.
[16] G.J. Ness, T. Helleseth, A new family of ternary almost perfect nonlinear mappings, IEEE Trans. Inform. Theory 53 (2007)
2581–2586.
[17] K. Nyberg, Differentially uniform mappings for cryptography, in: Advances in Cryptology – EUROCRYPT 93, in: Lecture
Notes in Comput. Sci., vol. 765, Springer-Verlag, New York, 1994, pp. 134–144.
[18] G. Weng, W. Qiu, Z. Wang, Q. Xiang, Pseudo-Paley graphs and skew Hadamard sets from presemifields, Des. Codes Cryp-
togr. 44 (2007) 49–62.
[19] Z. Zha, G. Kyureghyan, X. Wang, Perfect nonlinear binomials and their semifields, Finite Fields Appl. 15 (2009) 125–133.
Journal of Algebra
Vertices, sources and Green correspondents of the simple

modules for the large Mathieu groups
Susanne Danz ∗ , Burkhard Külshammer
Mathematisches Institut, Friedrich-Schiller-Universität Jena, 07737 Jena, Germany
Article history: We investigate the simple modules for the sporadic simple Mathieu
Received 27 May 2008 groups M 22 , M 23 and M 24 as well as those of the automorphism
Available online 24 April 2009 group, the covering groups and the bicyclic extensions of M 22 in
Communicated by Meinolf Geck
characteristics 2 and 3. We determine the vertices and sources as
Keywords:
well as the Green correspondents of these simple modules. We
Alperin’s Weight Conjecture also find two 3-blocks with elementary abelian defect groups of
Broué’s Abelian Defect Group Conjecture order 9 in these groups which are Morita equivalent to their Brauer
Covering group correspondents.
Green correspondent © 2009 Elsevier Inc. All rights reserved.
Mathieu group
Morita equivalence
Simple module
Source
Vertex
1. Introduction
More than 20 years ago, G.J.A. Schneider [38] determined the vertices, sources and Green corre-
spondents of the simple modules for the sporadic simple Mathieu groups M 11 and M 12 in character-
istic 2. Later H. Gollan [20] settled the characteristic 3 case, with a few exceptions. Gollan’s results
were reproved by S. Koshitani and K. Waki [29] who also determined the vertices and Green corre-
spondents of the modules left open in Gollan’s thesis. Both Schneider’s and Gollan’s results build on
computations with the computer algebra system CAYLEY, whereas Koshitani and Waki do without any
computer calculations.
The purpose of this article is to determine the vertices of the simple modules for the large Mathieu
groups M 22 , M 23 and M 24 over a field F of characteristics 2 and 3, respectively. For each of these
simple modules we also investigate its sources and its Green correspondent with respect to the nor-
E-mail addresses: susanned@minet.uni-jena.de (S. Danz), kuelshammer@uni-jena.de (B. Külshammer).
doi:10.1016/j.jalgebra.2009.03.035
3920 S. Danz, B. Külshammer / Journal of Algebra 322 (2009) 3919–3949
malizer of some vertex. Both the sources and the Green correspondents will be described in terms
of their Loewy series. All of the simple modules for the Mathieu groups M 22 , M 23 and M 24 in char-
acteristics 2 and 3 have precisely the defect groups of the respective blocks as vertices, with two
exceptions. The first of these exceptional cases occurs for M 24 in characteristic 2 where F M 24 has
only one block. The simple F M 24 -module D of dimension 1792 has the Sylow 2-subgroups of the
commutator subgroups of the maximal subgroups of M 24 with index 3795 as vertices. Moreover, each
vertex of D is a maximal subgroup of some Sylow 2-subgroup of M 24 . In characteristic 3 the principal
block of F M 24 contains a simple module of dimension 483 whose vertices are maximal subgroups of
Sylow 3-subgroups of M 24 and whose normalizers in M 24 are isomorphic to the automorphism group
of M 9 which is a split extension M 9 : S3 .
The Mathieu groups M 23 and M 24 have no non-trivial outer automorphisms, whereas the auto-
morphism group M 22 : 2 of M 22 is a split extension. Moreover, while, by [9], both M 23 and M 24 have
trivial Schur multipliers, M 22 has a Schur multiplier which is cyclic of order 12. This has been shown
by Mazet in [33]. In consequence, M 22 admits covering groups of order n| M 22 | = n · 27 · 32 · 5 · 7 · 11
for n ∈ {2, 3, 4, 6, 12}, and each of these is unique up to isomorphism. We will also focus on these
covering groups as well as on the automorphism group M 22 : 2 and the respective simple modules
over a field F of characteristics 2 and 3, respectively. In analogy to the case of the Mathieu groups,
we will determine the vertices of these simple modules, and give a description of their sources and
Green correspondents in terms of Loewy series. We will see below that, for our purposes, only the
covering groups 2. M 22 and 4. M 22 in characteristic 3, and the covering group 3. M 22 in characteristic 2
need to be considered explicitly. Once having determined vertices, sources and Green correspondents
for the simple modules in these cases, one obtains the respective results for the simple modules of
the remaining covering groups of M 22 as well.
Finally, we will also investigate the simple modules for the bicyclic extensions n. M 22 .2 of M 22
where n ∈ {2, 3, 4, 6, 12}. Here n. M 22 .2 is an extension of the automorphism group M 22 : 2 of M 22
by a cyclic group C = c of order n such that the outer automorphism of M 22 lifts to an outer
automorphism of n. M 22 mapping c to c −1 .
It turns out that all simple modules for the covering groups and the bicyclic extensions of M 22
considered throughout this article have precisely the defect groups of their blocks as vertices, and
in most cases this is a direct consequence of Knörr’s Theorem [27]. Two of the blocks investigated
here are particularly interesting, since all of their simple modules have trivial sources. We prove that
both blocks are Morita equivalent to their Brauer correspondents. This confirms Broué’s Abelian Defect
Group Conjecture for these blocks, in a strong form.
Suppose that G is an arbitrary finite group and that D is a simple F G-module with vertex Q
and trivial source. Then, by a theorem of Okuyama [36], the Green correspondent f ( D ) of D in
N G ( Q ) is again simple. Thus Q acts trivially on f ( D ), and f ( D ) can be viewed as a simple projective
F [ N G ( Q )/ Q ]-module. That is, the pair ( Q , [ f ( D )]) is a weight in Alperin’s sense [1]. In this way
every simple F G-module with trivial source defines a unique conjugacy class of weights, and non-
isomorphic simple F G-modules with trivial sources define different conjugacy classes of weights. Now
suppose that B is a block of F G such that every simple F G-module belonging to B has trivial source.
Then, in the way described above, we obtain an injection from the set of isomorphism classes of
simple F G-modules belonging to B into the set of conjugacy classes of F G-weights belonging to B.
According to Alperin’s Weight Conjecture [1], this injection should actually be a bijection. It would be
interesting to have a proof of this special case of Alperin’s Weight Conjecture.
If B has abelian defect group P , and still every simple F G-module belonging to B has trivial
source, then every simple F G-module belonging to B has vertex P , by Knörr’s Theorem [27], and
therefore the conjectured bijection of the previous paragraph should really be a one-to-one corre-
spondence between the set of isomorphism classes of simple F G-modules belonging to B and the set
of isomorphism classes of simple F N G ( P )-modules belonging to the Brauer correspondent b of B. In
view of the examples in this paper and other examples we are wondering whether B and b have to
be Morita equivalent in general.
We begin by introducing the notation used throughout and by summarizing some known facts
which will be needed later in this article. Afterwards, in Section 3, we will present our results on the
vertices, sources and Green correspondents of the simple modules under consideration. In Section 4
S. Danz, B. Külshammer / Journal of Algebra 322 (2009) 3919–3949 3921
we will then prove the existence of Morita equivalences between the blocks just mentioned above
and their respective Brauer correspondents.
Most of our results are due to calculations with the computer algebra system MAGMA [4]. In
this context we were faced with the task of constructing the simple modules for a finite group G
over a finite field. MAGMA provides a function AbsolutelyIrreducibleModulesBurnside
whose input are a finite permutation or matrix group G and a finite field F, and which constructs
the absolutely simple modules of G over appropriate extension fields of F. According to the MAGMA
instructions, the underlying algorithm is based on the Brauer–Burnside Theorem (cf. [5,8]). The lat-
ter asserts that, given any faithful F G-module V , each indecomposable projective F G-module occurs
as a direct summand of some tensor power of V . In particular, each simple F G-module then oc-
curs as composition factor of some tensor power of V . Most of the simple modules investigated here
have been constructed via the MAGMA function AbsolutelyIrreducibleModulesBurnside.
The function considers modules for splitting only up to some dimension d. The bound d can be de-
fined by the user, and is by default set to 2000. However, in order to get all simple modules of the
given group, the chosen bound d might be too small so that MAGMA then does not return all sim-
ple modules. Moreover, for our purposes it is often not necessary to construct all simple modules
of a group but only those in some particular block. Therefore, in some cases simple modules have
been constructed somewhat more interactively. Details are given at the beginning of Section 3. In
order to determine the Loewy series of the Green correspondents and the sources of the simple mod-
ules investigated, we repeatedly used the MAGMA function JacobsonRadical which computes the
radical of a given module over a group algebra.
The programs for carrying out the actual vertex computation had been developed by R. Zimmer-
mann in [43], and were later extended by R. Zimmermann and the authors in [16]. For details we refer
the reader to [15,16,43]. Moreover, the MAGMA source code of our algorithms as well as part of the
results of this note are available on-line at http://users.minet.uni-jena.de/~susanned/.
We have not yet developed actual code for computing the Green correspondent of a given inde-
composable F G-module M. In most cases we just restricted M to the normalizer of some vertex P ,
and determined an indecomposable direct sum decomposition of this restriction applying the built-
in MAGMA function IndecomposableSummands. Among these summands we then identified the
unique one with vertex P . In the case of the simple F M 24 -module of dimension 1243 in characteris-
tic 3 we proceeded slightly differently; details are given in Section 3.8.
The information on blocks and decomposition numbers for the groups investigated has been
taken from the GAP character table library [19] (see also http://www.math.rwth-aachen.de/
homes/MOC/decomposition/).
We would like to mention that sources and Green correspondents of simple modules for the Math-
ieu groups are also examined in [13] and [40] as has been pointed out by the referee.
2. Prerequisites
2.1. General notation
(1) In what follows, F will always denote a field of prime characteristic p, and all groups con-
sidered here are finite. Whenever A is an F -algebra then A is supposed to be finite-dimensional,
associative and unitary, and any A-module is understood to be a finitely generated left A-module.
Moreover, the endomorphism algebra End A ( V ) is supposed to be acting on V from the left as well.
By A ◦ we denote the opposite algebra of A. Given A-modules V and W such that W is isomorphic to
a direct summand of V , we write W | V .
(2) Suppose we are given groups G and H . For any F G-module V and any F H -module W , the
outer tensor product V ⊗ F W =: V W then becomes an F [G × H ]-module in the obvious way.
Moreover, there is a canonical F -algebra isomorphism F [G × H ] −→ F G ⊗ F F H . Identifying F [G × H ]
and F G ⊗ F F H via this isomorphism, each block of F [G × H ] may be written as B 1 ⊗ B 2 , for some
block B 1 of F G and some block B 2 of F H .
(3) When describing modules in terms of their Loewy series we will use the following notation:
given any F G-module V of Loewy length l ∈ N such that
Radi −1 ( V )/ Radi ( V ) = D i1 ⊕ · · · ⊕ D iri ,
for i = 1, . . . , l, appropriate r i ∈ N, and simple F G-modules D i1 , . . . , D iri , we write
⎡D ⎤
11 . . . D 1r1
⎢ D 21 . . . D 2r2 ⎥
V ∼⎢
⎣ .. ⎥.
⎦
.
D l1 . . . D lrl
If G is a p-group then we will only record the dimensions of the layers of the Loewy series of V ,
since these then determine the Loewy structure of V .
Suppose we are given any groups G and H , an F G-module V and an F H -module W with the
following properties:
• V and W have common Loewy length l ∈ N, and

• there is some bijection ϕ between the set of isomorphism classes of composition factors of V
and W such that Radi −1 ( V )/ Radi ( V ) ∼ = D i1 ⊕ · · · ⊕ D iri is equivalent to Radi −1 ( W )/ Radi ( W ) ∼
=
ϕ ( D i1 ) ⊕ · · · ⊕ ϕ ( D iri ), for i = 1, . . . , l and simple F G-modules D i1 , . . . , D iri .
Then we say that V and W have the same Loewy structure.

(4) For any integer n 1 we denote the symmetric and alternating group, respectively, of degree n
by Sn and An , respectively. Moreover, by D λ we understand the simple F Sn -module corresponding
to the p-regular partition λ of n. For details concerning the representation theory of the symmetric
groups we refer to [25].
2.2. Vertices, sources and the Green correspondence
Let G be a group, let V be an F G-module, and let H be a subgroup of G such that

V | IndGH (ResGH ( V )). Then V is said to be relatively H -projective. In the case that V is indecompos-
able, a subgroup P of G which is minimal subject to the condition that V is relatively P -projective
is called a vertex of V . The vertices of an indecomposable F G-module form a conjugacy class of
p-subgroups of G. Moreover, if P is a vertex of an indecomposable F G-module V then there is an
indecomposable F P -module W , unique up to isomorphism and conjugation with elements in N G ( P ),
such that V | IndGP ( W ). The module W is then called a source of V . For an introduction to the theory
of vertices, sources and Green correspondents of indecomposable F G-modules we refer the reader
to [35, Sections 4.3 and 4.4].
Suppose that D is a simple F G-module, and let P be a p-subgroup of G such that D is relatively
P -projective. As far as the question of determining the vertices of D is concerned, by Knörr’s The-
orem [27], we may suppose further that D belongs to a block with non-abelian defect groups. In
particular, D then cannot have cyclic vertices, by Erdmann’s Theorem [17]. Clearly, ResGP ( D ) has an
indecomposable direct summand whose vertices are also vertices of D and, as just mentioned, we
may ignore all those indecomposable direct summands of ResGP ( D ) which have cyclic vertices. In [16]
an algorithm is presented which detects and removes such summands from ResGP ( D ) without com-
puting an explicit indecomposable direct sum decomposition first. We have applied this algorithm
several times throughout our computations in order to make these less time and memory consuming.
In the following, we will write V = cyc when V is an F G-module all of whose indecomposable direct
summands have cyclic vertices. If V is projective then we write V = proj.
Consider, for instance, the simple F M 24 -module D (483)24 of dimension 483 in characteristic 3
M
which occurs in Section 3.8. Let P be a Sylow 3-subgroup of M 24 . Then Res P 24 ( D (483)24 ) =
U 1 ⊕ U 2 ⊕ cyc. Here U 1 and U 2 are indecomposable of dimension 3 each, and both have elemen-
tary abelian vertices of order 9. Restricting D := D (483)24 to P , splitting off all the indecomposable
direct summands with cyclic vertices, and determining the vertices of U 1 and U 2 required 11 MB of
main memory and took roughly 10 seconds. Calling the function VxStart(D,P) (cf. [43]) directly
required 2350 MB of main memory and took 15 hours.
For another example, let again be p = 3, and consider the simple F M 23 -module D := D (770)23 of
dimension 770 occurring in Section 3.7. Then D belongs to the principal block. If P ∈ Syl3 ( M 23 ) then
P is elementary abelian of order 9. By Knörr’s Theorem, D has thus vertex P . Moreover, Res P 23 ( D ) ∼
M
=
F ⊕ F ⊕ cyc. So D has trivial source. Restricting D to P , and splitting off all indecomposable direct
summands with cyclic vertices took 34 seconds and required 13 MB of main memory. When trying to
M
decompose Res P 23 ( D ) using the MAGMA function IndecomposableSummands, after 13 minutes
MAGMA ran out of memory (16 GB).
We mention that algorithms for splitting projective direct summands off a given module over some
p-group can also be found in [11,13,39].
Remark 2.1. Among the known criteria for testing a given F G-module V for relative projectivity with
respect to subgroups of G, especially Higman’s criterion which involves the relative trace map has
proved to be a valuable tool (cf. [35, Thm. 4.2.2]). Moreover, when H is a normal subgroup of G such
that |G : H | = p s , for some s ∈ N, also the result below is often helpful for testing an F G-module for
relative projectivity with respect to H . For this recall that the isomorphism classes of indecomposable
direct summands of an F G-module V bijectively correspond to the simple End F G ( V )-modules, via
Fitting correspondence. More precisely, consider a fixed decomposition
V ∼
= a 1 V 1 ⊕ · · · ⊕ ar V r ,
with mutually non-isomorphic indecomposable F G-modules V 1 , . . . , V r , and corresponding multiplic-

ities a1 , . . . , ar ∈ N. Furthermore, set E := End F G ( V ), and let
EE
∼
= b1 Ee 1 ⊕ · · · ⊕ bn Een
be a decomposition of the left E-module E, with pairwise orthogonal non-associate primitive idempo-
tents e 1 , . . . , en in E and multiplicities b1 , . . . , bn ∈ N. Then r = n, and after appropriate re-numbering,
we have V i ∼ = e i ( V ) = e i V and ai = bi , for i = 1, . . . , n. This follows from [35, Thm. 1.5.4].
For each f ∈ E, denote its image under the natural epimorphism E −→ E := E /J( E ) by f̄ . Then,
by [35, Thm. 1.4.5], { E ē 1 , . . . , E ēn } is a transversal for the isomorphism classes of simple E-modules,
and each of these occurs as composition factor of the E-module V . With this notation, we obtain:
Proposition 2.2. Let V be an F G-module such that
V ∼
= a1 V 1 ⊕ · · · ⊕ an V n ,
with pairwise non-isomorphic indecomposable F G-modules V 1 , . . . , V n and respective multiplicities

a1 , . . . , an ∈ N. As above, let { E ē 1 , . . . , E ēn } be the corresponding transversal for the isomorphism classes
of simple E-modules, and set di := [ V : E ē i ], for i = 1, . . . , n. Moreover, let E i := End E ( E ē i ), for i = 1, . . . , n.
Then ai · dim F ( E i ) = dim F ( E ē i ), and di · dim F ( E i ) = dim F ( V i ), for i = 1, . . . , n.
Proof. Let i ∈ {1, . . . , n}. In the notation of Remark 2.1, ai = b i which equals the multiplicity of the
simple E-module E ē i as composition factor of the E-module E E. By Wedderburn’s Theorem, this
multiplicity in turn equals dim E i ( E ē i ) = dim F ( E ē i )/ dim F ( E i ). Therefore, ai dim F ( E i ) = dim F ( E ē i ).
Furthermore, we have
◦
= (ē i E ē i )◦ ∼
E i = End E ( E ē i ) ∼ = (e i Ee i )/J(e i Ee i ) ,
so that the F -algebra e i Ee i possesses exactly one simple module which has F -dimension dim F ( E i ).
By [35, Thm. 1.8.11], the multiplicity di = [ V : E ē i ] equals the composition length of the e i Ee i -
module e i V . Consequently, di dim F ( E i ) = dim F (e i V ) = dim F ( V i ). This proves the proposition. 2
Corollary 2.3. In the notation of Proposition 2.2, assume that E ē i is absolutely simple, for some i ∈ {1, . . . , n}.
Then ai = dim F ( E ē i ), and di = dim F ( V i ). Moreover, the F G-module V i is then absolutely indecomposable. In
particular, if the E-module V has up to isomorphism precisely one composition factor and if this composition
factor is one-dimensional then V is absolutely indecomposable.
Proof. Consider any extension field F of F . Then
End F G ( F ⊗ F V ) ∼
= F ⊗ F End F G ( V ) ∼
= b1 ( F ⊗ F Ee 1 ) ⊕ · · · ⊕ bn ( F ⊗ F Een ),
by [35, Thm. 1.11.12]. Hence the F G-module V i is absolutely indecomposable if and only if the
E-module Ee i is absolutely indecomposable, or equivalently, if the E-module E ē i is absolutely simple.
The rest now follows from the previous proposition. 2
Corollary 2.4. Let H be a normal subgroup of G such that |G : H | = p s , for some s ∈ N. Moreover, let V be an
indecomposable F G-module such that the End F H ( V )-module ResGH ( V ) has an absolutely simple composition
factor occurring with multiplicity different from dim( V )/ p s . Then V is not relatively H -projective.
Proof. If V were relatively H -projective then we would have ResGH ( V ) ∼ = k( W 1 ⊕ · · · ⊕ W m ), for

mutually non-isomorphic and G-conjugate indecomposable F H -modules W 1 , . . . , W m . This follows
from [24, Thm. VII.9.3]. Moreover, V | IndGH ( W i ), for i = 1, . . . , m. By our hypothesis and Corollary 2.3,
W 1 , . . . , W m are absolutely indecomposable so that V ∼ = IndGH ( W i ) for i = 1, . . . , m, by Green’s Inde-
composability Theorem [21]. This yields the contradiction dim( V ) = p s dim( W i ) = dim( V ), and the
assertion follows. 2
Remark 2.5. (a) Corollary 2.4 is helpful for computational purposes, especially in the following situ-
ation: suppose that G is a p-group and that V is an indecomposable F G-module which we suspect
to have vertex G. Let { H 1 , . . . , H m } be a transversal for the conjugacy classes of maximal subgroups
of G. In order to verify that G is a vertex of V , it suffices to successively restrict V to each of
the maximal subgroups H i , compute the endomorphism algebras E i := End F H i (ResGH i ( V )), and search
for an absolutely simple composition factor of the F E i -module ResGH i ( V ) occurring with multiplic-
ity different from dim F ( V )/ p. This enables us to avoid the computation of the relative trace maps
TrGH i : End F H i (ResGH i ( V )) −→ End F G ( V ), for i = 1, . . . , m, which is required when applying Higman’s
criterion.
(b) Let G be arbitrary again. Suppose that F is any extension field of F , and let V be an indecom-
posable F G-module with vertex P . Then P is also a vertex of each indecomposable direct summand
of the F G-module F ⊗ F V . For a proof see for instance [18, L. II.4.14]. Each simple module D inves-
tigated in this paper has been constructed over a finite field Fq , for some p-power q, such that:
• D is absolutely simple,
• the sources and the Green correspondent of D are absolutely indecomposable,
• the composition factors of the Green correspondent of D are absolutely simple.
In order to check whether a given indecomposable Fq G-module is absolutely indecomposable, we

applied the criterion given by Corollary 2.3. Hence, the Fq G-module Fq ⊗Fq D is also simple and has
the same vertices as D. Moreover, if V and L are the Green correspondent and a source of D then
both Fq ⊗Fq V and Fq ⊗Fq L are indecomposable, and are the Green correspondent and a source of the
Fq G-module Fq ⊗Fq D. Also the Loewy structures of the modules in question do not change under this
field extension. Therefore, unless stated otherwise, for the remainder of this article we may assume
the field F to be algebraically closed.
In the sequel, we will have to deal with F G-modules which are obtained from modules for factor
groups of G via inflation. Therefore, recall that, given a normal subgroup N of G and an indecom-
posable F [G / N ]-module V , the inflation V := Inf NG ( V ) is an indecomposable F G-module. Moreover,
if P is a vertex of V then P N / N is a vertex of V . A proof for this can, for instance, be found
in [31, Prop. 2.1]. In particular, P ∩ N is then a Sylow p-subgroup of N, and P is a Sylow p-subgroup
of P N which implies N G / N ( P N / N ) = N G ( P ) N / N. In analogy, also the Green correspondents and the
sources of V and V are related as follows:
Proposition 2.6. Let N be a normal subgroup of G, and let V be an indecomposable F G-module such that
V = Inf NG ( V ), for an indecomposable F [G / N ]-module V . Suppose that V has vertex P so that N G / N ( P N / N ) =
N G ( P ) N / N. Let V be the Green correspondent of V with respect to (G , P , N G ( P )), and let V be the Green
correspondent of V with respect to (G / N , P N / N , N G ( P ) N / N ). Then V ∼
N ( P )N N ( P )N
= ResN G ( P ) (Inf N G
G
( V )). If L is
a P -source of V then there is a P N / N-source L of V such that L ∼
= Res PP N (Inf NP N ( L )). In particular, the Loewy
structures of V and V , as well as those of L and L coincide.
For a proof of Proposition 2.6 see [22].
Proposition 2.7. Let N be a normal subgroup of G, and let V be a simple F G-module with vertex Q N. Sup-
pose that | N G ( Q ) : N N ( Q )| = |G : N |, and denote the Green correspondence with respect to (G , Q , N G ( Q ))
by f 1 and the Green correspondence with respect to ( N , Q , N N ( Q )) by f 2 . Consider an indecomposable direct
sum decomposition ResGN ( V ) = V 1 ⊕ · · · ⊕ V n . Then Q is a vertex of V i for i = 1, . . . , n, and

Res N G ( Q ) f 1 ( V ) ∼
N (Q )
N
= f 2 ( V 1 ) ⊕ · · · ⊕ f 2 ( V n ).
Proof. We set H := N G ( Q ) and K := N N ( Q ). Then |G : H | = | N : K | so that the G-conjugates of Q are

precisely the N-conjugates of Q . By Clifford’s Theorem [35, Thm. 3.3.1], we have
ResGN ( V ) ∼
= k( W 1 ⊕ · · · ⊕ W m ),
for some k ∈ N, and simple F N-modules W 1 , . . . , W m which are pairwise conjugate in G. We may
assume that W 1 has vertex Q , and since the G-conjugates of Q are exactly the N-conjugates of Q ,
also W 2 , . . . , W m have vertex Q . Furthermore,
ResGH ( V ) = f 1 ( V ) ⊕ X 1 ⊕ · · · ⊕ X r ,
Res N
K ( W i ) = f 2 ( W i ) ⊕ Y i1 ⊕ · · · ⊕ Y isi , for i = 1, . . . , m.
Here, X 1 , . . . , X r and Y i1 , . . . , Y isi are indecomposable, and neither of these modules has vertex Q .
Now consider

m
k f 2 (W 1 ) ⊕ · · · ⊕ f 2 (W m ) ⊕ (Y i1 ⊕ · · · ⊕ Y isi )
i =1

∼
= ResGK ( V ) = Res HK f 1 ( V ) ⊕ Res HK ( X 1 ) ⊕ · · · ⊕ Res HK ( X r ),
and assume that k( f 2 ( W 1 ) ⊕ · · · ⊕ f 2 ( W m )) Res H K ( f 1 ( V )). Then there are some i ∈ {1, . . . , m} and
some j ∈ {1, . . . , r } such that f 2 ( W i ) | Res H K ( X j ) . In particular, Q is H -conjugate to a subgroup of
some vertex of X j , that is Q is itself a vertex of X j , a contradiction. Consequently, k( f 2 ( W 1 ) ⊕ · · · ⊕
f 2 ( W m )) | Res H H
K ( f 1 ( V )). Since f 1 ( V ) is relatively K -projective, Res K ( f 1 ( V )) is a direct sum of inde-
composable F K -modules which are pairwise conjugate in H . This follows from [24, Thm. VII. 9.3].
In particular, all indecomposable direct summands of Res H

K ( f 1 ( V )) have vertex Q , and this finally
yields

f 2(V 1) ⊕ · · · ⊕ f 2(V n) ∼
= k f 2 (W 1 ) ⊕ · · · ⊕ f 2 (W m ) ∼
= Res HK f 1 ( V ) . 2
Remark 2.8. Suppose that B is a block of F G with cyclic defect group C . The representation theory
of blocks with cyclic defect groups is well understood. In particular, such a block B is a Brauer tree
algebra. We refer to [23] for an introduction to the theory of Brauer trees. Given the Brauer tree of B
and the Green correspondent of one of the simple modules in B, one can also determine the Green
correspondents of the remaining simple modules in B. This is shown in [23, Sec. 4.4]. In fact, if B is
one of the blocks with cyclic defect groups investigated in this note then B occurs in characteristic 3,
and has a defect group C of order 3. In this special case, B contains either one or two simple F G-
modules, up to isomorphism. Moreover, by [23, Sections 6.4, 6.6 and 6.9], the Brauer tree of B is
either of shape
, (1)
where the node 3 is of type ◦ and the nodes 1 and 2 are of type ×, or of shape
, (2)
where the exceptional node 2 has type ◦, and the node 1 has type ×. In either case, [23, L. 4.4.12]
implies that all simple modules belonging to B have simple Green correspondents in N G (C ). Hence,
by [35, Thm. 4.7.8], these simple modules must have trivial sources.
We also mention a misprint in [23]. In the Brauer tree of the block B 8 of F [12. M 22 ] in character-
istic 3, the labelling of the nodes by × and ◦ should be interchanged. This block has a central defect
group of order 3; its Brauer tree is of shape (2), and also here the exceptional node has type ◦, and
the non-exceptional node has type ×.
2.3. Group extensions
In this subsection we briefly recall some facts concerning the covering groups of a finite group.
For more details and proofs of the results quoted here, we refer to [3, Sec. 33]. A covering group of
a group G is a group G containing a central subgroup H
H such that G /
G and H∼
= G. Here G

denotes the commutator subgroup of G. Provided that H has order 2, 3, etc., one also speaks of G as
a double cover, triple cover, etc., of G.
Now suppose that G is a perfect group, i.e. G equals its commutator subgroup G . Then there exists
a covering group G of G such that each covering group of G is isomorphic to a factor group of G.
Furthermore, G is determined up to isomorphism, and is called the Schur cover or Darstellungsgruppe
of G. Moreover, H Z (
G is itself a perfect group. If H∼
G /
G ) with = G then H is isomorphic to the
Schur multiplier of G.
Remark 2.9. As mentioned in the introduction, the Schur multiplier of M 22 is cyclic of order 12,
and M 22 therefore possesses covering groups of order 2| M 22 |, 3| M 22 |, 4| M 22 |, 6| M 22 | and 12| M 22 |. We
will use the Atlas notation, and write n. M 22 to denote the covering group of M 22 of order n| M 22 | =
n · 27 · 32 · 5 · 7 · 11, for n ∈ {2, 3, 4, 6, 12}.
Note that, as far as the determination of vertices of simple F [n. M 22 ]-modules is concerned,
we only need to consider the cases where n is coprime to p = char( F ). Namely, if n = p r q, for
some q, r ∈ N such that p q, then there is a cyclic subgroup Z Z (n. M 22 ) of order p r such that
n. M 22 / Z ∼
= q. M 22 . Moreover, by [35, Thm. 4.7.8], every simple F [n. M 22 ]-module D corresponds to
a simple F [q. M 22 ]-module D , via inflation. If P Z is a vertex of D then P / Z is a vertex of D ,
by [31, Prop. 2.1]. Furthermore, in consequence of Proposition 2.6 the Loewy structures of the Green
correspondents of D and D coincide. The same applies to the sources of D and D . Therefore, we
only need to investigate the simple F [3. M 22 ]-modules for p = 2, and the simple F [2. M 22 ]-modules
and the simple F [4. M 22 ]-modules for p = 3 explicitly.
Remark 2.10. There are five bicyclic extensions of M 22 , namely 2. M 22 .2, 3. M 22 : 2, 4. M 22 .2, 6. M 22 .2
and 12. M 22 .2. Let C 12 ∼
= C = c be the centre of 12. M 22 . The outer automorphism of M 22 lifts to an
outer automorphism of 12. M 22 which maps c to c −1 . In particular, 12. M 22 .2 has a centre of order 2.
If p = 2 then we only need to consider the simple modules for F [3. M 22 : 2]. For n ∈ {2, 4, 6, 12} any
simple F [n. M 22 .2]-module D is the inflation of a simple module D for F [ M 22 : 2] or F [3. M 22 : 2],
by [35, Thm. 4.7.8]. Moreover, in these cases Proposition 2.6 implies that the vertices as well as the
Loewy structures of the sources and the Green correspondents of D can be deduced from those of D .
Analogously, in the case where p = 3, it suffices to focus on the simple modules for F [2. M 22 .2]
and F [4. M 22 .2], respectively.
2.4. Equivalences between module categories
In Section 4 we will prove the existence of Morita equivalences between the faithful 3-block of
2. M 22 of defect 2 and its Brauer correspondent, and between the faithful 3-block of 2. M 22 .2 of de-
fect 2 and its Brauer correspondent. For this we now briefly recall some facts concerning equivalences
between module categories. For more details on this subject we refer to [2].
Given any F -algebra A, we denote the category of finitely generated left A-modules by A-mod.
Furthermore, A-stab denotes the stable category of A. That is, the objects in A-stab are the same
as the objects in A-mod. But the morphisms between objects V and W in A-stab are equivalence
classes of A-homomorphisms between V and W modulo A-homomorphisms which factor through
projective A-modules.
Definition 2.11. Consider F -algebras A and B.

(i) The algebras A and B are called Morita equivalent if and only if there exist an A-B-bimodule V
and a B- A-bimodule W such that V ⊗ B W ∼ = A as A- A-bimodules and W ⊗ A V ∼ = B as B-B-bimodules.
(ii) We say that there exists a stable equivalence of Morita type between A and B if and only if there
exist an A-B-bimodule V and a B- A-bimodule W such that the following conditions are satisfied:
• V is projective both as left A-module and right B-module.

• W is projective both as left B-module and right A-module.
• V ⊗B W ∼ = A ⊕ X as A- A-bimodules, where X is a projective A- A-bimodule.
• W ⊗A V ∼ = B ⊕ Y as B-B-bimodules, where Y is a projective B-B-bimodule.
Remark 2.12. By Morita’s Theorem, the F -algebras A and B are Morita equivalent if and only if the
module categories A-mod and B-mod are equivalent. If there exists a stable equivalence of Morita
type between A and B then the stable categories A-stab and B-stab are equivalent.
With this notation, the following result will be essential:
Theorem 2.13. (See Linckelmann [32, Thm. 2.1].) Let A and B be indecomposable non-simple symmetric
F -algebras, and let W be a B- A-bimodule which is projective both as left B-module and right A-module. Sup-
pose further that the functor W ⊗ A − induces a stable equivalence between A and B. Then the following
hold:
(i) If W is indecomposable and if D is a simple A-module, then W ⊗ A D is an indecomposable non-projective

B-module.
(ii) If, for every simple A-module D, also W ⊗ A D is a simple B-module then the functor W ⊗ A − induces
a Morita equivalence between A and B.
Remark 2.14. We consider subgroups K H G and an F G-module V . Then V K and V H will denote
the F -vector spaces of fixed points under the action of K and H , respectively, on V . Moreover, for
v ∈ V K , we define

Tr H
K ( v ) := g v.
g K ∈H /K
This yields an F -linear map Tr H K : V −→ V

K H
which does not depend on the choice of the transversal
for H / K , and is called relative trace.
Definition 2.15. Let P be a p-subgroup of G, and let V be an F G-module. Then we set

V ( P ) := V P Tr PQ V Q .
Q <P
Remark 2.16. (a) In this way we obtain a functor
F G-mod −→ F N G ( P )-mod, V −→ V ( P ),
which is called the Brauer functor or the Brauer construction with respect to P . In the following we list
some known facts concerning the Brauer functor. For details we refer to [6] and [41, §§11, 27].
(b) The Brauer functor is additive. That is, if V 1 and V 2 are F G-modules then the canonical map
V 1 ( P ) ⊕ V 2 ( P ) −→ ( V 1 ⊕ V 2 )( P ) is an isomorphism of F N G ( P )-modules. We identify V 1 ( P ) ⊕ V 2 ( P )
and ( V 1 ⊕ V 2 )( P ) in this way. Also, we have a canonical homomorphism V 1 ( P ) ⊗ F V 2 ( P ) −→
( V 1 ⊗ F V 2 )( P ).
(c) Suppose that A is a G-algebra over F . That is, A is both an F -algebra and an F G-module, and
for g ∈ G the map A −→ A , a −→ ga, is an F -algebra automorphism of A. Then the canonical map
A ( P ) ⊗ F A ( P ) −→ ( A ⊗ F A )( P ) −→ A ( P )
turns A ( P ) into an N G ( P )-algebra over F .

(d) Let A be a G-algebra over F , and let V be an A-module. Suppose further that V is an F G-
module and that the canonical map A ⊗ F V −→ V is an F G-homomorphism. Then the induced map
A ( P ) ⊗ F V ( P ) −→ ( A ⊗ F V )( P ) −→ V ( P )
gives an A ( P )-module structure on V ( P ) which is compatible with the action of N G ( P ).

(e) Suppose that V is a p-permutation F G-module. That is, for every p-subgroup P of G there is
an F -basis of V stabilized by P . Let further P and Q be p-subgroups of G such that Q P . Then
V (P ) ∼
= ( V ( Q ))( P ). Note that V ( Q ) is a p-permutation F N G ( Q )-module. In particular, V ( P ) = 0 then
also implies V ( Q ) = 0.
(f) Let V 1 and V 2 be F G-modules such that V 1 or V 2 is a p-permutation F G-module. Then the
canonical map V 1 ( P ) ⊗ F V 2 ( P ) −→ ( V 1 ⊗ F V 2 )( P ) is an isomorphism.
(g) The group algebra F G is both a G-algebra over F and a p-permutation F G-module. Further-
more, its Brauer construction ( F G )( P ) is canonically isomorphic to F C G ( P ), as an N G ( P )-algebra
over F . We identify ( F G )( P ) and F C G ( P ) in this way. If F G = F Ge 1 ⊕ · · · ⊕ F Ger is the block de-
composition of F G then, via this identification and additivity, we get
( F G )( P ) = ( F Ge 1 )( P ) ⊕ · · · ⊕ ( F Ger )( P ) = F C G ( P ) Br P (e 1 ) ⊕ · · · ⊕ F C G ( P ) Br P (er ),
where Br P is the usual Brauer homomorphism. Thus the Brauer construction of a block is either 0 or
a direct sum of blocks of F C G ( P ).
In the following, for each subgroup H of G, we denote the respective diagonal subgroup of G × G
by H := {(h, h) | h ∈ H }. Then the next result will play a crucial role in Section 4.
Theorem 2.17. (See Broué [7, Thm. 6.3].) Let G and H be groups with a common Sylow p-subgroup P such that
N G ( P )/C G ( P ) ∼
= N H ( P )/C H ( P ). Moreover, let e and f be central idempotents in F G and F H , respectively,
and set A := F Ge and B := F H f . We suppose that there are an A-B-bimodule V and a B- A-bimodule W
such that
G×H
(1) V | Ind P ( X ), for some permutation F [ P ]-module X .
(2) If Q is a non-trivial subgroup of P then F C G ( Q ) Br Q (e ) and F C H ( Q ) Br Q ( f ) are Morita equivalent via
the functors W ( Q ) ⊗ F C G ( Q ) Br Q (e) − and V ( Q ) ⊗ F C H ( Q ) Br Q ( f ) −.
Then the functors W ⊗ A − and V ⊗ B − induce a stable equivalence of Morita type between A and B.
3. Results
3.1. Constructing simple modules
We now present our results most of which have been obtained with computer assistance.
In the following, we consider M 22 < M 23 < M 24 < S24 and M 22 M 22 : 2 < S24 . For j ∈
{22, 23, 24}, by D (di ) j we understand the ith simple F M j -module of dimension d. We omit the in-
dex i whenever F M j has, up to isomorphism, exactly one simple module or two mutually dual simple
modules of dimension d. Analogously, the simple F [ M 22 : 2]-modules are denoted by D (di )22:2 , and
for n ∈ {2, 3, 4, 6, 12}, the simple F [n. M 22 ]-modules and the simple F [n. M 22 .2]-modules, respectively,
are denoted by D (di )n.22 and D (di )n.22.2 , respectively.
A permutation representation of 2. M 22 on 660 points and two permutation representations
of 4. M 22 on 4928 points are given in [42]. In the case of 4. M 22 we have worked with the first of
those two permutation representations. The simple modules for 4. M 22 in characteristic 3 investi-
gated here all appear to be composition factors of the corresponding permutation module over F9 .
The simple modules for 2. M 22 in characteristic 3 have been constructed via the MAGMA func-
tion AbsolutelyIrreducibleModulesBurnside. To obtain the simple modules for 2. M 22 .2
in characteristic 3, we proceeded as follows: we started with the 10-dimensional irreducible matrix
representation of 2. M 22 .2 over F3 available at [42]. The GAP function IsomorphismPermGroup
then enables us to construct a permutation representation on 9240 points of the respective matrix
group. This also yields a permutation representation of 2. M 22 2. M 22 .2 on 9240 points. The sim-
ple F9 [2. M 22 .2]-modules investigated in Section 3.6 could then be obtained as composition factors of
inductions of simple F9 [2. M 22 ]-modules to 2. M 22 .2.
In order to construct the simple modules for 3. M 22 in characteristic 2, we have taken the permuta-
tion representation for 3. M 22 on 693 points from [42]. Then each of the non-projective faithful simple
F4 [3. M 22 ]-modules occurs as composition factor of the corresponding permutation module over F4 .
The simple F9 M 24 -modules D (45)24 and D (990)24 of dimension 45 and 990, respectively, have
been obtained as composition factors of the induction of one of the 45-dimensional simple F9 M 23 -
modules. The remaining simple F9 M 24 -modules have been constructed as composition factors of
appropriate tensor powers of the simple F9 M 24 -module D (22)24 which is the unique non-trivial com-
position factor of the natural F9 M 24 -permutation module of dimension 24.
All remaining simple modules considered below have actually been constructed via the MAGMA
function AbsolutelyIrreducibleModulesBurnside. It turns out that all modules considered
throughout, that is all simple modules as well as their sources and Green correspondents can be
realized over F p 2 , and that they are then also absolutely indecomposable. Absolute indecomposability
has always been checked using the criterion given by Corollary 2.3.
3.2. Determining Brauer characters
For most of the simple modules occurring in this article we also determined their corresponding
Brauer characters. In order to do this we proceeded as follows: suppose that n 1, and let q := pn .
For our purposes n is at most 2. Then, following the MAGMA instructions, calling the function GF(q)
the finite field Fq is constructed as F p [x]/(C n ). Here C n denotes the Conway polynomial of degree n.
This is consistent with the Atlas, see [26, Sec. 3]. The image of x in F p [x]/(C n ) is denoted by Fq .1. If
n = 1 then F p .1 = 1.
With this notation, Fq .1 is a generator of the multiplicative group Fq× . Let further ζ :=
exp(2π i /(q − 1)). Then the lifting map
Z[ζ ] −→ Fq
used in the Atlas is that mapping ζ to Fq .1 (cf. [26, Sec. 2, 3]).

In [42], standard presentations of the groups examined here are given. If in addition representa-
tives for the p-regular conjugacy classes as words in the standard generators are given then one can
identify the Brauer character of a given simple module.
Example 3.1. Let p = 2, and consider the Mathieu group M 22 . Standard generators of M 22 are a and b
where a has order 2, b has order 4 and belongs to conjugacy class 4 A, and where both ab and ababb
have order 11. By [42], ababababbbabb belongs to conjugacy class 7 A, and ab belongs to conjugacy
class 11 A. The simple modules of dimension 10 occurring in Section 3.3 can be realized over F2 , the
ones of dimension 70 can be realized over F4 . Using the standard permutation representation of M 22
on 22 points from [42], and denoting Brauer characters as well as irrationalities in accordance with
the Atlas, we get:
Module Conj. class Modular character value

D (10)22 7A 0 = b7
D (70)22 11A F4 .12 = F4 .1 − 1 = −1 + b11
Hence, by [26], D (10)22 has Brauer character ϕ2 , and D (70)22 has Brauer character ϕ5 . This also
determines the labels of the Brauer characters of the simple modules D (10)∗22 , D (70)∗22 , D (34)22
and D (98)22 .
In characteristic 3, the double cover 2. M 22 has two simple modules of dimension 154. The val-
ues of the corresponding Brauer characters coincide, except on the conjugacy classes of elements of
order 8. There are two of the latter both of which are lying above the unique conjugacy class of
elements of order 8 of M 22 . Similarly, in characteristic 3, the covering group 4. M 22 has two dual
pairs of simple modules of dimension 56. Also here, in order to determine the labels of the Brauer
characters afforded by these modules, one needs to know their values on the four conjugacy classes
lying above the unique conjugacy class of elements of order 8 of M 22 . There does not seem to be
an obvious way to distinguish these conjugacy classes of elements of order 8. Moreover, in order to
distinguish between the faithful blocks B 5 and B 6 = B ∗5 of F [4. M 22 ] of defect 2, one has to make
a choice for a generator for the central subgroup of 4. M 22 of order 4. Therefore we do not determine
the precise Brauer character labels of the 154-dimensional simple F [2. M 22 ]-modules belonging to the
faithful block of defect 2, and we do not determine the precise Brauer character labels of the simple
F [4. M 22 ]-modules belonging to the faithful blocks B 5 and B 6 just mentioned. For similar reasons we
do not deduce the precise labelling of the Brauer characters of the simple F [2. M 22 .2]-modules in
characteristic 3 either.
In order to distinguish simple modules of equal dimension, in these three cases, we give elements
in the respective groups, written as words in the standard generators, on which the Brauer character
values of the modules in question differ. In addition, whenever we have simple modules of equal
dimension for which representing matrices are available at [42], we attach to each of these its “ID” as
given in [42].
Throughout we use the notation D ↔ ϕ ↔ a to indicate that a simple module D has Brauer char-
acter ϕ , and ID a in [42]. The Brauer character afforded by the dual module D ∗ is denoted by ϕ .
Which conjugacy classes we actually used in order to determine the Brauer characters of the simple
modules are listed at http://users.minet.uni-jena.de/~susanned/mathieu.html.
In view of Remark 2.5 and the considerations in Section 3.1, from now on we may again suppose
that F is algebraically closed. For the next three subsections let further p = 2.
3.3. M 22 in characteristic 2
We begin by investigating the Mathieu group M 22 and its extensions. For simplicity of notation,
we set M := M 22 , throughout this subsection.
Remark 3.2. (a) There is only one block of F M, i.e. the principal block of defect 7. Moreover, F M has
the following seven simple modules:
D (1)22 = F ↔ ϕ1 , D (10)22 ↔ ϕ2 ↔ a, D (10)∗22 ↔ ϕ3 ↔ b, D (34)22 ↔ ϕ4 ,

∗
D (70)22 ↔ ϕ5 ↔ a, D (70)22 ↔ ϕ6 ↔ b, D (98)22 ↔ ϕ7 .
(b) There is only one block of F [ M : 2], i.e. the principal one, containing the following six simple
modules:
D (1)22:2 = F ↔ ϕ1 , D (10)22:2 ↔ ϕ2 ↔ a, D (10)∗22:2 ↔ ϕ3 ↔ b,
D (34)22:2 ↔ ϕ4 , D (98)22:2 ↔ ϕ7 , D (140)22:2 ↔ ϕ5 .
M :2
Here, Ind M = D (140)22:2 ∼
( D (70)22 ) ∼ M :2
= Ind M M :2
( D (70)∗22 ), and ResM ( D (d)22:2 ) ∼
= D (d)22 , for d ∈ {1, 10,
34, 98}. In particular, D (140)22:2 is the only relatively M-projective simple F [ M : 2]-module.
With this notation, we have the following:
Proposition 3.3. All seven simple F M-modules have vertex Q ∈ Syl2 ( M ). Furthermore, the restriction of any
simple F M-module D to Q is also a source of D. For Q ∈ Syl2 ( M ) we have N M ( Q ) = Q , and thus the Green
correspondents of the simple F M-modules in Q are also sources. They have the following Loewy series:
Module D (1)22 D (10)22 D (10)∗22 D (34)22 D (70)22

Green 1 10 10 34 70
Layer dims. 1 1, 2, 1, 2, 1, 1, 1, 2 2, 3, 3, 5, 4, 1, 3, 4, 6, 7, 8, 9,
1, 1, 1, 1 2, 1, 1, 1 5, 4, 4, 2, 2 8, 7, 6, 5, 3, 2, 1
Module D (70)∗22 D (98)22

Green 70 98
Layer dims. 1, 3, 4, 6, 7, 8, 9, 2, 5, 6, 9, 10, 12,
8, 7, 6, 5, 3, 2, 1 12, 12, 10, 8, 6, 3, 2, 1
Proof. These results have been obtained by computer calculations. 2
Proposition 3.4. The simple F [ M : 2]-module D (140)22:2 has vertex Q ∈ Syl2 ( M ). Moreover, the restriction of
D (70)22 to Q is a source of D (140)22:2 . The remaining simple F [ M : 2]-modules have vertex P ∈ Syl2 ( M : 2),
and restrict indecomposably to P . Choosing Q P , we have N M :2 ( P ) = P = N M :2 ( Q ), and the Green corre-
spondents of the simple F [ M : 2]-modules in P have the following Loewy series:
Module D (1)22:2 D (10)22:2 D (10)∗22:2 D (34)22:2 D (98)22:2

Green 1 10 10 34 98
Layer dims. 1 1, 2, 1, 2, 1, 1, 1, 2, 2, 3, 3, 5, 4, 2, 5, 6, 9, 10, 12,
1, 1, 1, 1 2, 1, 1, 1 5, 4, 4, 2, 2 12, 12, 10, 8, 6, 3, 2, 1
Module D (140)22:2
Green 140
Layer dims. 1, 4, 7, 10, 13, 15, 17,
17, 15, 13, 11, 8, 5, 3, 1
Proof. As mentioned above, the only relatively M-projective simple F [ M : 2]-module is D (140)22:2 ,
M :2
and Res M ( D (140)22:2 ) ∼
= D (70)22 ⊕ D (70)∗22 . Thus, by the previous proposition, D (140)22:2 and
D (70)22 have common vertex Q and common source Res M Q ( D (70)22 ). Moreover, we have Q =
:2
N M ( Q ) N M :2 ( Q ) = P and | P : Q | = 2. Consequently, Res M
P ( D (140)22:2 ) is indecomposable and
therefore the Green correspondent of D (140)22:2 in P . Its Loewy series has been determined compu-
tationally. The remaining simple F [ M : 2]-modules restrict irreducibly to M and are thus not relatively
M-projective. Since |( M : 2) : M | = 2, Proposition 3.3 implies that all these modules have vertex P , and
restrict indecomposably to P . The Loewy series of their Green correspondents have been computed to
be as stated. 2
Remark 3.5. We observe that the Loewy series of the Green correspondents of the not relatively
M-projective simple F [ M : 2]-modules and those of the respective simple F M-modules coincide. How-
ever, we do not have a good explanation for this.
Remark 3.6. Consider the non-split central extension 3. M of M by a group of order 3. Besides the
principal block whose simple modules are obtained from the simple F M-modules via inflation, the
group algebra F [3. M ] has two faithful blocks B 2 and B 3 = B ∗2 of defect 7. The simple modules be-
longing to B 2 and B 3 , respectively, are:
B 2 : D (6)3.22 ↔ ϕ8 ↔ a, D (15)3.22 ↔ ϕ9 ↔ a, D (451 )3.22 ↔ ϕ11 ↔ b,
D (452 )3.22 ↔ ϕ10 ↔ a, D (84)3.22 ↔ ϕ12 ↔ a;

∗ ∗
B 3 : D (6)3.22 ↔ ϕ8 , D (15)3.22 ↔ ϕ9 , D (451 )∗3.22 ↔ ϕ11 ,
D (452 )∗3.22 ↔ ϕ10 , D (84)∗3.22 ↔ ϕ12 .
Moreover, F [3. M ] has two blocks of defect 0. In view of [35, Thm. 4.7.8] and Proposition 2.6, it
suffices to determine the vertices, sources and Green correspondents of the simple F [3. M ]-modules
belonging to the faithful blocks of positive defect.
Denoting a Sylow 2-subgroup of 3. M by P , we have N 3. M ( P ) ∼ = Z (3. M ) × P ∼
= C 3 × P . Hence
F [ N 3. M ( P )] possesses three blocks each of which has defect 7 and contains one simple module. All
of these simple modules are one-dimensional, and we denote them by 11 = F , 12 , 13 = 1∗2 where 12
belongs to the Brauer correspondent of B 2 , and 13 belongs to the Brauer correspondent of B 3 .
Proposition 3.7. All simple F [3. M ]-modules belonging to the blocks B 2 and B 3 , respectively, have vertex
P ∈ Syl2 (3. M ), and restrict indecomposably to P . Thus, with the above parametrization, 12 is the unique
composition factor of the Green correspondents of the simple modules belonging to B 2 , and 13 is the unique
composition factor of the Green correspondents of the simple modules belonging to B 3 . The Loewy series of
these Green correspondents are as follows:
Module D (6)3.22 D (6)∗3.22 D (15)3.22 D (15)∗3.22 D (451 )3.22 D (451 )∗3.22

Green 6 6 15 15 45 45
Layer dims. 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 1, 2, 2, 2, 2, 3, 4, 6, 5, 6, 1, 2, 3, 4, 5, 6,
1, 1, 1 1, 1, 1 2, 2, 2, 1, 1 2, 2, 2, 1, 1 5, 5, 3, 3, 2, 1 6, 6, 5, 4, 2, 1
Module D (452 )3.22 D (452 )∗3.22 D (84)3.22 D (84)∗3.22

Green 45 45 84 84
Layer dims. 1, 2, 3, 4, 5, 6, 2, 3, 4, 6, 5, 6, 2, 4, 6, 8, 9, 10, 10, 2, 4, 6, 8, 9, 10, 10,
6, 6, 5, 4, 2, 1 5, 5, 3, 3, 2, 1 10, 8, 6, 5, 3, 2, 1 10, 8, 6, 5, 3, 2, 1
Furthermore, for any simple F [3. M ]-module belonging to B 2 or B 3 , the Loewy series of its sources and those
of its Green correspondent coincide.
Proof. Consider the modules D (15)3.22 , D (451 )∗3.22 and D (452 )3.22 first. Since their dimensions are
not divisible by 2, they all have vertex P . Furthermore, we have determined the Loewy series of their
restrictions to N 3. M ( P ) with the computer. It turns out that each of these restricted modules has
a one-dimensional head, and, in particular, is thus indecomposable. Therefore the Green correspon-
dents of these three modules and their duals are precisely the indecomposable restrictions to N 3. M ( P ).
The concrete Loewy series have been determined computationally.
Finally we have computed that the simple modules D (6)3.22 and D (84)3.22 restrict indecomposably
to P , and that neither is relatively projective with respect to any maximal subgroup of P . Thus both
modules have vertex P , and their restrictions to P are also sources. The Loewy series of their Green
correspondents in N 3. M ( P ) have been determined by computer calculations to be as claimed.
The assertion concerning the Loewy series of the sources of the simple modules in question now
follows from [24, Thms. VII.7.21 and VII.9.15]. 2
Remark 3.8. (a) Let G := 3. M : 2. Then F G possesses three blocks: the principal block B 1 of defect 8,
the faithful block B 2 of defect 7 and the block B 3 of defect 0. The simple modules belonging to B 1
are obtained from the simple F [ M : 2]-modules via inflation. The simple modules belonging to the
block B 2 are the following:
D (12)3.22.2 ↔ ϕ8 , D (30)3.22.2 ↔ ϕ9 , D (90)3.22.2 ↔ ϕ11 ,
D (90)∗3.22.2 ↔ ϕ10 , D (168)3.22.2 ↔ ϕ12 .
Moreover, the outer automorphism of 3. M interchanges the faithful blocks of defect 7 of F [3. M ].
In consequence thereof, each of these faithful blocks of F [3. M ] is Morita equivalent to B 2 via the
respective induction functors. This is Fong’s first correspondence (cf. [30]). In particular, the simple
modules in B 2 are precisely the inductions of the faithful simple F [3. M ]-modules to G. We choose
notation such that Ind3G. M ( D (451 )3.22 ) ∼= D (90)3.22.2 ∼
= Ind3G. M ( D (452 )∗3.22 ).
(b) Let Q be a Sylow 2-subgroup of 3. M, that is a defect group of B 2 . Then N G ( Q ) is a split
extension N 3. M ( Q ) : 2 ∼= (C 3 × Q ) : 2. Moreover, N G ( Q )/ Q ∼ = S3 so that there are two simple
F N G ( Q )-modules: the trivial module and the inflation of the projective simple F S3 -module D (2,1)
of dimension 2; we denote the latter by 2. Here we obtain Morita equivalences between the block of
F [ N G ( Q )] containing 2 and each of the two non-principal blocks of F [ N 3. M ( Q )] both of which are
nilpotent blocks. Again the Morita equivalences are obtained via the induction functors, and we have
=2∼
Ind N G ( Q ) (12 ) ∼
N (Q ) N (Q )
3. M
= Ind N G ( Q ) (13 ) where 12 and 13 are as in Remark 3.6. The Green correspondence
3. M
with respect to (G , Q , N G ( Q )) will be denoted by f 2 .
Proposition 3.9. Let G := 3. M : 2. Let Q ∈ Syl2 (3. M ) so that Q is a defect group of B 2 . Then each sim-
ple F G-module belonging to B 2 has vertex Q , and the 2-dimensional simple F N G ( Q )-module is the unique
composition factor of its Green correspondent. The Loewy series of the Green correspondents and those of the
sources of the simple modules in B 2 are as follows:
Module D (12)3.22.2 D (30)3.22.2 D (90)3.22.2 D (90)∗3.22.2 D (168)3.22.2

Green 12 30 90 90 168
Layer 2, 2, 2, 2, 4, 4, 4, 4, 6, 8, 12, 10, 12, 2, 4, 6, 8, 10, 12, 4, 8, 12, 16, 18, 20, 20,
dims. 2, 2, 2 4, 4, 4, 2, 2 10, 10, 6, 6, 4, 2 12, 12, 10, 8, 4, 2 20, 16, 12, 10, 6, 4, 2
Sce. 6 15 45 45 84
Layer 1, 1, 1, 1, 2, 2, 2, 2, 3, 4, 6, 5, 6, 1, 2, 3, 4, 5, 6, 2, 4, 6, 8, 9, 10, 10,
dims. 1, 1, 1 2, 2, 2, 1, 1 5, 5, 3, 3, 2, 1 6, 6, 5, 4, 2, 1 10, 8, 6, 5, 3, 2, 1
Proof. We consider a simple F G-module V belonging to B 2 . Then there exists a simple F [3. M ]-
module U such that V ∼ = Ind3G. M (U ) and Res3G. M ( V ) ∼
= U ⊕ g U , for some g ∈ G \ 3. M. Both V and U
have vertex Q ∈ Syl2 (3. M ), by Proposition 3.7, and they also have common sources. Furthermore, we
have N 3. M ( Q ) N G ( Q ) =: N, and | N : N 3. M ( Q )| = 2. By Proposition 3.7, the Green correspondents in
N 3. M ( Q ) of U and g U are isomorphic to Res3N. M ( Q ) (U ) and Res3N. M ( Q ) ( g U ), respectively. From this,
3. M 3. M
together with Proposition 2.7, we now deduce that f 2 ( V ) = ResGN ( V ) and
g
Ind N 3. M ∼ G ∼ N 3. M
N 3. M ( Q ) Res N 3. M ( Q ) (U ) = Res N ( V ) = Ind N 3. M ( Q ) Res N 3. M ( Q ) U .
As explained in Remark 3.8, the induction functors yield Morita equivalences between each of the
two non-principal blocks of F [ N 3. M ( Q )] and the Brauer correspondent of B 2 . Thus, in particular, the
Green correspondent of U in N 3. M ( Q ) and f 2 ( V ) have the same Loewy structure, and the assertion
now follows from Proposition 3.7. 2
Remark 3.10. There are three blocks of F M 23 : the principal one and two dual blocks of defect 0. The
principal block has defect 7, and contains the following nine simple modules:
D (1)23 = F ↔ ϕ1 , D (11)23 ↔ ϕ3 ↔ b, D (11)∗23 ↔ ϕ2 ↔ a, D (44)23 ↔ ϕ4 ↔ a,

∗
D (44)23 ↔ ϕ5 ↔ b, D (120)23 ↔ ϕ6 , D (220)23 ↔ ϕ8 ↔ b, D (220)∗23 ↔ ϕ7 ↔ a,
D (252)23 ↔ ϕ9 .
Proposition 3.11. All simple F M 23 -modules belonging to the principal block have vertex P ∈ Syl2 ( M 23 ).
Moreover, D (120)23 has sources of dimension 56, and D (252)23 has sources of dimension 28. The remaining
modules in the principal block restrict indecomposably to P . We have N M 23 ( P ) = P , and the Green correspon-
dents in P of the simple F M 23 -modules in the principal block are thus also sources of these. Their Loewy series
are as follows:
Module D (1)23 D (11)23 D (11)∗23 D (44)23 D (44)∗23 D (120)23

Green 1 11 11 44 44 56
Layer dims. 1 2, 2, 1, 2, 1, 1, 1, 2, 3, 4, 4, 6, 5, 2, 3, 4, 5, 5, 2, 4, 5, 7, 8, 8,
1, 1, 1, 1 2, 2, 1, 1 6, 5, 5, 3, 2, 1 6, 6, 6, 3, 3, 1 7, 6, 4, 2, 2, 1
Module D (220)23 D (220)∗23 D (252)23

Green 220 220 28
Layer dims. 3, 8, 12, 18, 22, 26, 27, 5, 9, 13, 20, 23, 26, 3, 4, 4, 5,
27, 24, 20, 16, 9, 6, 2 28, 27, 22, 18, 14, 8, 5, 2 3, 3, 3, 2, 1
M
Proof. Consider the modules D (120)23 and D (252)23 first. Then Res P 23 ( D (120)23 ) = U ⊕ cyc, and
M
= V ⊕ W ⊕ cyc. Here U , V and W are indecomposable of dimension 56, 28 and 32,
Res P 23 ( D (252)23 )
respectively. As mentioned earlier, the indecomposable direct summands with cyclic vertices are eas-
ily detected. Both U and V have vertex P and are thus sources and Green correspondents of D (120)23
and D (252)23 , respectively. All other simple F M 23 -modules in the principal block restrict indecom-
posably to P , and are not relatively projective with respect to any maximal subgroup of P . The Loewy
series of the Green correspondents have been computed to be as stated. 2
Remark 3.12. (a) There is only one block of F M 24 , i.e. the principal one of defect 10. The 13 simple
F M 24 -modules are:
D (1)24 = F ↔ ϕ1 , D (11)24 ↔ ϕ2 ↔ a, D (11)∗24 ↔ ϕ3 ↔ b, D (44)24 ↔ ϕ5 ↔ b,

∗
D (44)24 ↔ ϕ4 ↔ a, D (120)24 ↔ ϕ6 , D (220)24 ↔ ϕ8 ↔ b, D (220)∗24 ↔ ϕ7 ↔ a,
D (252)24 ↔ ϕ9 , D (320)24 ↔ ϕ11 ↔ b, D (320)∗24 ↔ ϕ10 ↔ a, D (1242)24 ↔ ϕ12 ,
D (1792)24 ↔ ϕ13 .
(b) By [12], M 24 has three conjugacy classes of maximal subgroups with odd index in M 24 . These
are 24 : A8 of index 759, 26 : 3.S6 of index 1771, and 26 : ( L 3 (2) × S3 ) of index 3795.
Proposition 3.13. All simple F M 24 -modules, except D (1792)24 , have vertex P ∈ Syl2 ( M 24 ). Furthermore,
N M 24 ( P ) = P so that the Green correspondents of these modules are also sources. Their Loewy series are as
follows:
Module D (1)24 D (11)24 D (11)∗24 D (44)24 D (44)∗24

Green 1 11 11 44 44
Layer dims. 1 1, 1, 2, 2, 1, 1, 1, 2, 1, 2, 4, 4, 6, 6, 2, 2, 3, 4, 4, 6,
2, 1, 1, 1 2, 2, 1, 1 6, 5, 5, 3, 1, 1 6, 6, 4, 4, 2, 1
Module D (120)24 D (220)24 D (220)∗24 D (252)24

Green 120 220 220 252
Layer dims. 2, 4, 6, 9, 10, 14, 1, 3, 6, 10, 14, 20, 2, 4, 7, 11, 14, 18, 3, 5, 7, 13, 17, 22,
14, 14, 13, 12, 23, 26, 26, 25, 21, 22, 24, 24, 23, 21, 25, 29, 28, 27, 23,
9, 6, 4, 2, 1 18, 12, 8, 4, 2, 1 17, 14, 9, 6, 3, 1 20, 14, 9, 6, 3, 1
Module D (320)24 D (320)∗24 D (1242)24

Green 320 320 218
Layer dims. 1, 3, 6, 11, 17, 23, 29, 1, 3, 6, 11, 17, 23, 29, 3, 6, 9, 14, 17, 21,
34, 36, 36, 34, 29, 34, 36, 36, 34, 29, 23, 25, 23, 22, 18,
23, 17, 11, 6, 3, 1 23, 17, 11, 6, 3, 1 14, 10, 7, 3, 2, 1
The vertices of D (1792)24 have order 512, and are the M 24 -conjugates of the Sylow 2-subgroups
of the commutator subgroup of 26 : ( L 3 (2) × S3 ) M 24 . Furthermore, N M 24 ( Q ) has order 3072, and
N M 24 ( Q )/ Q ∼
= S3 . Denoting the trivial F [ N M 24 ( Q )]-module by 1, and the inflation of the two-dimensional
projective simple F S3 -module by 2, the Green correspondent of D (1792)24 has dimension 256 and the fol-
lowing Loewy series:
Layer 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Multiplicity of 1 0 1 4 8 11 11 8 5 5 8 11 11 8 4 1 0
Multiplicity of 2 1 2 2 3 4 6 10 13 13 10 6 3 2 2 2 1
Moreover, D (1792)24 has sources of dimension 128, and the dimensions of their Loewy layers are 1, 3, 5, 9,
13, 15, 18, 18, 15, 13, 9, 5, 3, 1.
Proof. All simple F M 24 -modules, except D (1792)24 and D (1242)24 , remain indecomposable when
restricted to P , and neither is relatively projective with respect to any maximal subgroup of P . More-
M
over, Res P 24 ( D (1242)24 ) = U ⊕ proj, where U is indecomposable of dimension 218 and has vertex P .
M
Finally, consider D (1792)24 . Our computations show that Res P 24 ( D (1792)24 ) = V ⊕ W ⊕ proj, where V
and W are indecomposable of dimension 256 and 512, respectively. It turns out that W has ver-
tices of order 128. Furthermore, there is a maximal subgroup Q of P such that Res PQ ( V ) = V 1 ⊕ V 2
=V ∼
and Ind PQ ( V 1 ) ∼ = Ind PQ ( V 2 ), for indecomposable F Q -modules V 1 and V 2 of dimension 128
both of which have vertex Q . Therefore also D (1792)24 has vertex Q , and V 1 and V 2 are sources
of D (1792)24 . Each maximal subgroup of M 24 of odd index in M 24 clearly contains an M 24 -conjugate
of Q . In fact, Q is conjugate to a Sylow 2-subgroup of (26 : ( L 3 (2)×S3 )) . The Loewy structures of the
Green correspondents and the sources of the simple F M 24 -modules have then also been determined
with the computer to be as claimed. 2
For the remainder of this article, let p = 3. We investigate the Mathieu group M 22 , its automor-
phism group M 22 : 2 and their extensions 2. M 22 , 4. M 22 , 2. M 22 .2, 4. M 22 .2 first. For this we again set
M := M 22 .
Remark 3.14. There are five blocks of F M: the principal block B 1 of defect 2 whose defect groups are
elementary abelian, the block B 2 of defect 1, and the blocks B 3 , B 4 = B ∗3 , and B 5 of defect 0. Thus,
in particular, all simple F M-modules have the defect groups of their blocks as vertices, by Knörr’s
Theorem [27]. For the sake of completeness, we will determine the sources and Green correspondents
of the simple modules belonging to the blocks of positive defect.
(a) The principal block B 1 contains the following simple modules:
D (1)22 = F ↔ ϕ1 , D (49)22 ↔ ϕ5 ↔ a, D (49)∗22 ↔ ϕ6 ↔ b, D (55)22 ↔ ϕ7 ,
D (231)22 ↔ ϕ10 .
Let P be a Sylow 3-subgroup of M. The normalizer N M ( P ) has order 23 · 32 = 72, and is isomorphic
to M 9 . Therefore F [ N M ( P )] has five simple modules: four of dimension 1, and one of dimension 2. We
denote these by 11 = F , 12 , 13 , 14 , 2. There are outer automorphisms ϕ and ψ of M 9 mapping 12 to 13
and 14 , respectively. Moreover, let f 1 be the Green correspondence with respect to ( M , P , N M ( P )).
(b) Now consider the block B 2 with defect group C ∼ = C 3 which contains the simple mod-
ules D (21)22 ↔ ϕ2 and D (210)22 ↔ ϕ9 . The normalizer N M (C ) has order 23 · 32 = 72; its Sylow
3-subgroups are elementary abelian of order 9, and its Sylow 2-subgroups are isomorphic to the dihe-
dral group D 8 of order 8. Furthermore, F [ N M (C )] possesses four simple modules: two of dimension 1,
and two of dimension 3. We denote them by 11 = F , 12 , 31 , 32 . In fact, C acts trivially on these simple
modules, by [35, Thm. 4.7.8]. Moreover, we have N M (C )/C ∼ = S4 . We choose notation such that 31 is
2
isomorphic to D (3,1) , and 32 is isomorphic to D (2,1 ) , when regarded as F S4 -modules. Considered
2
as F S4 -module, the one-dimensional module 12 is isomorphic to the alternating F S4 -module D (2 ) .
Denoting the Green correspondence with respect to ( M , C , N M (C )) by f 2 , we obtain the following:
Proposition 3.15. All simple F M-modules have the defect groups of their blocks as vertices. Furthermore, the
Loewy series of the Green correspondents and the sources of the simple modules in B 1 and B 2 are as follows:
Block B1 B2
Module D (1)22 D (49)22 D (49∗ )22 D (55)22 D (231)22 D (21)22 D (210)22

13 12 2
Green F 2 2 14 14 11 31 32
12 13 2

F F F
Source F FF FF F F F F
F F F
Proof. The assertion concerning the vertices is clear. The Loewy series of the Green correspondents of
the simple modules belonging to B 1 are given in the proof of [28, Prop. 4.2], and the Loewy series of
the sources of the simple modules in B 1 have been determined with the computer to be as claimed.
As mentioned in Remark 2.8, by [23, Sec. 6.4 and L. 4.4.12], the simple modules belonging to B 2 have
simple Green correspondents and thus trivial sources, by [35, Thm. 4.7.8]. We have explicitly checked
that 31 ∼
= f 2 ( D (21)22 ) which then implies f 2 ( D (210)22 ) ∼
= 32 . 2
Remark 3.16. Now, consider the automorphism group M : 2 of M. There are nine blocks of F [ M : 2].
The ones of positive defect are the principal block B 1 of defect 2, and the blocks B 2 and B 3 of defect 1.
(a) The principal block of F [ M : 2] contains the following simple modules:
D (11 )22:2 = F ↔ ϕ1,0 , D (12 )22:2 ↔ ϕ1,1 , D (551 )22:2 ↔ ϕ7,0 ↔ a, D (552 )22:2 ↔ ϕ7,1 ,
D (98)22:2 ↔ ϕ5 , D (2311 )22:2 ↔ ϕ10,0 ↔ a, D (2312 )22:2 ↔ ϕ10,1 .
Here, we have Ind M M :2

= D (98)22:2 ∼
( D (49)22 ) ∼ = Ind M M :2
( D (49)∗22 ) and ResMM :2
( D (98)22:2 ) ∼
= D (49)22 ⊕
∗
D (49)22 . The remaining simple F [ M : 2]-modules are the extensions of the simple F M-modules
F = D (1)22 , D (55)22 and D (231)22 , respectively. The Brauer characters ϕi ,0 are those whose values
are listed in the printed Atlas [26]. Let P be a Sylow 3-subgroup of M. By Proposition 3.15, all simple
F M-modules belonging to the principal block and thus also all simple F [ M : 2]-modules belonging to
the principal block have vertex P . Furthermore, given a simple F [ M : 2]-module D belonging to B 1
M :2
and a simple F M-module D such that D | Res M ( D ) then D and D have common sources.
Moreover, N M :2 ( P ) is isomorphic to a split extension N M ( P ) : 2 ∼ = M 9 : 2, and F [ M 9 : 2] has seven
simple modules: four of dimension 1, and three of dimension 2. We denote these by 11 = F , 12 , 13 ,
14 , 21 , 22 , 23 = 2∗2 . In the notation of Remark 3.14 and Proposition 3.15, the restrictions of 22 and 23
to M 9 are isomorphic to the 2-dimensional simple F M 9 -module, the restriction of 21 to M 9 splits into
the direct sum of the simple F M 9 -modules 12 and 13 . Furthermore, the restrictions of the F [ M 9 : 2]-
modules 12 and 14 to M 9 are isomorphic to the F M 9 -module 14 , and the restriction of 13 to M 9 is
trivial. We identify N M ( P ) with M 9 and N M :2 ( P ) with M 9 : 2. The Green correspondence with respect
to ( M : 2, P , N M ( P ) : 2) will be denoted by f 1 .
(b) The blocks of defect 1 contain the simple modules:
B 2 : D (211 )22:2 ↔ ϕ2,0 ↔ a, D (2101 )22:2 ↔ ϕ9,1 ,
B 3 : D (212 )22:2 ↔ ϕ2,1 , D (2102 )22:2 ↔ ϕ9,0 ↔ a,
where

Res M
M
:2
D (211 )22:2 ∼
= D (21)22 ∼
= ResM
M
:2
D (212 )22:2 ,

Res M
M
:2
D (2101 )22:2 ∼
= D (210)22 ∼
= ResM
M
:2
D (2102 )22:2 .
Again, the Brauer characters ϕi ,0 are those whose values occur in [26]. The simple modules in B 2
and B 3 have trivial sources and simple Green correspondents, by [23, Sec. 6.4 and L. 4.4.12],
and [35, Thm. 4.7.8]. If C M is a defect group of both B 2 and B 3 then N M :2 (C ) has order 144.
Moreover, N M :2 (C )/C ∼
= S4 × C 2 . In particular, the simple F [ N M :2 (C )]-modules are precisely the infla-
tions of the simple F [S4 × C 2 ]-modules.
Consider the natural epimorphism
= ∼
ν : S4 −→ S4 /A4 −
→ C2.
Then S4 × C 2 has two subgroups isomorphic to S4 , namely H 1 := {(x, 1) | x ∈ S4 } and H 2 :=

{(x, ν (x)) | x ∈ S4 }. Both of these have C 2 as a complement. In the following, we set N := N M :2 (C ), we
identify N /C with S4 × C 2 , and consider N /C as the inner direct product of H 1 and C 2 . Then there
are eight simple F [ N /C ]-modules, namely
2 2
F F, D (2 ) F , F sgn, D (2 ) sgn,
2 2
D (3,1) F , D (2,1 ) F , D (3,1) sgn, D (2,1 ) sgn.
Here sgn denotes the alternating F C 2 -module. Note that this labelling depends on the identification
of N /C with S4 × C 2 . If we choose an essentially different identification, that is if we replace H 1
by H 2 , then, for any simple F S4 -module D, the labelling of the module D F remains the same, but
2
the module D sgn is replaced by ( D ⊗ D (2 ) ) sgn.
With this notation we obtain:
Proposition 3.17. The simple F [ M : 2]-modules belonging to the blocks of positive defect have the defect
groups of their blocks as vertices. Their Green correspondents have the following Loewy series:
Block B1
Module D (11 )22:2 D (12 )22:2 D (551 )22:2 D (552 )22:2 D (98)22:2 D (2311 )22:2 D (2312 )22:2

21 22 23
Green F 13 12 14 22 23 11 14 12 13
21 23 22
Block B2 B3
Module D (211 )22:2 D (2101 )22:2 D (212 )22:2 D (2102 )22:2
2 2
Green InfCN ( D (3,1) F) InfCN ( D (2,1 ) sgn) InfCN ( D (3,1) sgn) InfCN ( D (2,1 ) F )
Proof. We have explicitly determined the isomorphism types of the Green correspondents of the
simple modules belonging to B 2 . These then also determine the isomorphism types of the Green
correspondents of the simple modules belonging to B 3 . We also obtain that f 1 ( D (11 )22:2 ) = F ,
f 1 ( D (12 )22:2 ) ∼
= 13 , f 1 ( D (551 )22:2 ) ∼
= 12 and f 1 ( D (552 )22:2 ) ∼
= 14 . From Propositions 3.15, 2.7
and [24, Thm. VII.7.21] we further deduce that

21 Y1 Y3
f 1 D (98)22:2 ∼ X1 X2 , f 1 D (2311 )22:2 ∼ Z1 Z2 , f 1 D (2312 )22:2 ∼ Z3 Z4 ,
21 Y2 Y4
where X 1 , X 2 ∈ {22 , 23 }, Y 1 , . . . , Y 4 ∈ {22 , 23 }, Z 1 , Z 3 ∈ {11 , 13 } and Z 2 , Z 4 ∈ {12 , 14 }. The actual iso-

morphism types of these simple modules can be read off from [28, proof of L. 4.5]. 2
Remark 3.18. The group algebra F [2. M ] possesses nine blocks. In view of [35, Thm. 4.7.8] and Proposi-
tion 2.6, it suffices to consider the simple F [2. M ]-modules belonging to the faithful blocks of F [2. M ].
These are the block B 3 of defect 2 whose defect groups are elementary abelian, and the block B 4 of
defect 1.
(a) The simple modules belonging to B 3 are
D (10)2.22 ↔ ϕ11 ↔ a, D (10)∗2.22 ↔ ϕ12 ↔ b, D (56)2.22 ↔ ϕ13 ,
D (154)2.22 , D (154)∗2.22 .
The Brauer character afforded by D (154)2.22 is either ϕ17 or ϕ18 , and representing matrices for
D (154)2.22 are available at [42]. Let A and B be standard generators of 2. M, let ζ := exp(2π i /8),
and consider the lifting map − : Z[ζ ] −→ F9 as in Section 3.2. Then g := A B A B A B B B A B B has or-
der 8. For the Brauer character ϕ afforded by D (154)2.22 , we have ϕ ( g ) = F9 .1 + 1 = −2i, and hence
ϕ ( g ) = 2i.
Let P ∈ Syl3 (2. M ), i.e. a defect group of B 3 . Then N 2. M ( P ) has order 144, and if Q ∈ Syl2 ( N 2. M ( P ))
then Q / Z (2. M ) is isomorphic to the quaternion group Q 8 of order 8. Furthermore, F [ N 2. M ( P )] has
ten simple modules which are obtained from the simple F [ N 2. M ( P )/ P ]-modules via inflation. These
will be denoted by F = 11 , 12 , . . . , 18 , 21 , 22 . Moreover, we may choose notation such that the simple
modules belonging to the Brauer correspondent b3 of B 3 are
15 , 16 = 1∗5 , 21 , 17 , 18 = 1∗7 .
(b) The block B 4 of defect 1 contains the simple modules D (120)2.22 ↔ ϕ14 and D (2101 )2.22 ↔ ϕ19 .
Let C ∼
= C 3 be a defect group of B 4 . Then N 2. M (C ) has order 144. Moreover, each simple F [ N 2. M (C )]-
module can be regarded as a simple F [ N 2. M (C )/C ]-module. Actually, N 2. M (C )/C ∼
= S4 × C 2 where C 2
denotes a cyclic group of order 2. In analogy to Remark 3.16 above, we now identify N 2. M (C )/C with
S4 × C 2 , and regard N 2. M (C )/C as the inner direct product of the subgroups H 1 := {(x, 1) | x ∈ S4 }
and C 2 . The eight simple F [ N 2. M (C )/C ]-modules are also denoted as in Remark 3.16.
For a suitable labelling we get:
Proposition 3.19. All simple F [2. M ]-modules belonging to B 3 and B 4 , respectively, have the defect groups of
their blocks as vertices. Moreover, all of these have trivial sources, and the following Green correspondents:
Block B3
Module D (10)2.22 D (10)∗2.22 D (56)2.22 D (154)2.22 D (154)∗2.22
Green 15 16 = 1∗5 21 17 18 = 1∗7
Block B4
Module D (120)2.22 D (2101 )2.22
N 2. M ( C ) N 2. M 22 (C ) 2
Green InfC 22 ( D (3,1) sgn) InfC ( D (2,1 ) sgn)
Proof. The assertion concerning the vertices is clear, by Knörr’s Theorem [27]. The sources and Green
correspondents have been determined computationally to be as claimed. 2
Remark 3.20. Next we turn to the group 4. M. Again it suffices to focus on the faithful blocks
of F [4. M ]. These are the blocks B 5 and B 6 = B ∗5 of defect 2 containing the following simple mod-
ules:
Block Module ID Brauer

B5 D (561 )4.22 a ϕ20 , ϕ21 , ϕ20 , ϕ21
D (562 )4.22 b ϕ20 , ϕ21 , ϕ20 , ϕ21
D (64)4.22 a ϕ22 , ϕ22
D (1601 )4.22 a ϕ25 , ϕ26
D (1602 )4.22 b ϕ26 , ϕ25
B6 D (561 )∗4.22 ϕ20 , ϕ21 , ϕ20 , ϕ21
D (562 )∗4.22 ϕ20 , ϕ21 , ϕ20 , ϕ21
D (64)∗4.22 ϕ22 , ϕ22
D (1601 )∗4.22 ϕ26 , ϕ25
D (1602 )∗4.22 ϕ25 , ϕ26
Column “Brauer” displays the possible Brauer characters of the simple modules in B 5 and B 6 ,
respectively. Again, we also record the ID’s of the modules appearing in [42]. Let A and B be standard
generators of 4. M as in [42], let ϕ be the Brauer character afforded by D (561 )4.22 , and let ψ be
the Brauer character afforded by D (562 )4.22 . Let further ζ := exp(2π i /8), and let − : Z[ζ ] −→ F9 be
the lifting map as in Section 3.2. Then g := A B A B A B B B A B B has order 8, ϕ ( g ) = 2F9 .1 = 2z8 , and
ψ( g ) = F9 .1 = −2z8 . The element z := ( A B A B A B A B A B 2 A B A B 2 A B 2 )63 ∈ Z (4. M 22 ) of order 4 acts
on the simple modules in B 5 via multiplication with F9 .12 , and on the simple modules in B 6 via
multiplication with F9 .16 .
Consider further P ∈ Syl3 (4. M ). Then P is elementary abelian of order 9 and a defect group of B 5
and B 6 . Its normalizer N 4. M ( P ) has order 288, and is isomorphic to a split extension P : (C 4 : C 8 ).
We denote the Green correspondence with respect to (4. M , P , N 4. M ( P )) by f . There are precisely 20
simple F [ N 4. M ( P )]-modules, namely the inflations of the simple F [C 4 : C 8 ]-modules. These will be
denoted by F = 11 , . . . , 116 , 21 , . . . , 24 .
In [34] J. Müller and M. Schaps have proved that the blocks B 5 and B 6 are derived equivalent to
their Brauer correspondents b5 and b6 = b∗5 , respectively, thereby proving Broué’s conjecture for B 5
and B 6 . We choose notation such that the simple N 4. M ( P )-modules belonging to the block b5 are
12 , 13 , 14 , 15 , 21 , and the simple N 4. M ( P )-modules belonging to the block b6 are 16 = 1∗2 , 17 = 1∗3 ,
18 = 1∗4 , 19 = 1∗5 , 22 = 2∗1 . For a suitable labelling of the simple F N 4. M ( P )-modules, we obtain the
following:
Proposition 3.21. The simple F [4. M ]-modules belonging to B 5 and B 6 , respectively, have vertex P . Moreover,
the Loewy series of their Green correspondents are as follows:
Block B5
Module D (561 )4.22 D (562 )4.22 D (64)4.22 D (1601 )4.22 D (1602 )4.22
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
12 21 13 21 14 21 15 21
12 14 15
⎢ 13 14 15 ⎥ ⎢ 13 21 ⎥ ⎢ 12 13 15 21 ⎥ ⎢ 12 13 14 21 ⎥
Green ⎣ ⎦ ⎣ ⎦ 21 21 ⎣ ⎦ ⎣ ⎦
21 21 12 14 15 12 13 15 21 12 13 14 21
13 14 15
12 21 14 21 15 21
Block B6
Module D (561 )∗4.22 D (562 )∗4.22 D (64)∗4.22 D (1601 )∗4.22 D (1602 )∗4.22
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1∗2 2∗1 1∗3 2∗1 1∗4 2∗1 1∗5 2∗1
1∗3 1∗4 1∗5
⎢ 1∗ 2∗ ⎥ ⎢ 1∗ 1∗ 1∗ 5⎥
∗ 1∗ 1∗ 2∗
⎢ 2 3 5 1⎥
1 ∗ 1∗ 1∗ 2∗
⎢ 2 3 4 1⎥
1
Green ⎣ 2 1
⎦ ⎣ 2 4
⎦ 2∗1 2∗1 ⎣ ∗ ∗ ∗ ∗⎦ ⎣ ∗ ∗ ∗ ∗⎦
1∗3 1∗4 1∗5 2∗1 2∗1 1 1 1 2 12 13 14 21
1∗2 1∗4 1∗5 2 3 5 1
2∗1 1∗3 1∗4 2∗1 1∗5 2∗1
If D is any of the simple modules belonging to B 5 or B 6 then the restriction of its Green correspondent f ( D )
N (P )
to P is also a source of D. Moreover, f ( D ) and Res P 4.M ( f ( D )) have the same Loewy lengths, and the di-
mensions of the respective Loewy layers coincide.
Proof. The assertion concerning the vertices is obvious. The dimensions of the Green correspondents
of the simple modules belonging to B 5 and B 6 , respectively, have already been determined in [34].
Moreover, the Green correspondents restrict indecomposably to P , by [34]. We have further calculated
the explicit Loewy structures of all Green correspondents with the computer, as stated. The assertion
regarding the sources follows from [24, Thm. VII.7.21]. 2
To close this subsection, we investigate the simple modules for the bicyclic extensions 2. M .2
and 4. M .2 of M.
Remark 3.22. (a) The group algebra F [2. M .2] has 13 blocks, and due to our previous considerations
it suffices to investigate only the faithful ones of positive defect explicitly. These are the block B 4 of
defect 2 and the blocks B 5 and B 6 of defect 1. In fact, the blocks B 5 and B 6 only differ by a linear
character, and are thus isomorphic.
(b) Consider the blocks B 5 and B 6 of defect 1 first. Here the block B 5 contains the simple modules
D (1201 )2.22.2 and D (2104 )2.22.2 , and the block B 6 contains D (1202 )2.22.2 and D (2103 )2.22.2 where

Res22.. M
M
.2
D (1201 )2.22.2 ∼
= D (120)2.22 ∼
= Res22.. M
M
.2
D (1202 )2.22.2 ,

Res22.. M
M
.2
D (2103 )2.22.2 ∼
= D (2101 )2.22 ∼
= Res22.. M
M
.2
D (2104 )2.22.2 .
Let c and d be standard generators of 2. M .2 as in [42]. Then g := (cdd)3 is an element of order 2

belonging to one of the conjugacy classes of 2. M .2 lying above the conjugacy class 2B of M : 2. Let
− : Z −→ F be the residue map. If ϕ denotes the Brauer character afforded by D (120 )
3 2 2.22.2 and if ψ
denotes the Brauer character afforded by D (2104 )2.22.2 then ϕ ( g ) = 2 = 8, and ψ( g ) = 1 = 28.
If C 2. M is a defect group of both B 5 and B 6 then N 2. M .2 (C ) ∼
= N 2. M (C ).2 has order 288, and
each simple F N 2. M (C )-module extends to two simple F N 2. M .2 (C )-modules. Thus there are 16 simple
F N 2. M .2 (C )-modules eight of which have dimension 1, and the remaining eight have dimension 3. We
may choose notation such that the Brauer correspondent b5 of B 5 contains the simple modules 33
and 34 , and the Brauer correspondent b6 of B 6 contains the simple modules 35 and 36 where
(C ) (2,12 )
Res N 2.M .(2C ) (33 ) ∼ sgn ∼
N (C ) N N (C )
2. M
= InfC 2.M D = ResN 22..MM .(2C ) (35 )
and
(C ) (3,1)
Res N 2.M .(2C ) (34 ) ∼ sgn ∼
N (C ) N N (C )
2. M
= InfC 2.M D = Res N 22..MM .(2C ) (36 ),
in the notation of Remark 3.16(b).

Let f 5 be the Green correspondence with respect to (2. M .2, C , N 2. M .2 (C )). In consequence of
Propositions 3.19 and 2.7, the simple modules belonging to B 5 and B 6 , respectively, have simple
Green correspondents and trivial sources. More precisely, we have:
f 5 ( D (1201 )2.22.2 ) ∼
= 34 , f 5 ( D (1202 )2.22.2 ) ∼
= 36 , f 5 ( D (2103 )2.22.2 ) ∼
= 35 ,
f 5 ( D (2104 )2.22.2 ) ∼
= 33 .
(c) The faithful block B 4 contains the simple modules
D (101 )2.22.2 , D (102 )2.22.2 , D (101 )∗2.22.2 , D (102 )∗2.22.2 ,
D (561 )2.22.2 , D (562 )2.22.2 , D (308)2.22.2 ,

where notation is chosen such that

Ind22.. M
M
.2
D (10)2.22 ∼
= D (101 )2.22.2 ⊕ D (102 )2.22.2 ,

Ind22.. M
M
.2
D (10)∗2.22 ∼
= D (101 )∗2.22.2 ⊕ D (102 )∗2.22.2 ,

Ind22.. M
M
.2
D (56)2.22 ∼
= D (561 )2.22.2 ⊕ D (562 )2.22.2 ,

Ind22.. M
M
.2
D (154)2.22 ∼
= D (308)2.22.2 ∼
= Ind22.. M
M
.2
D (154)∗2.22 .
In particular, all these simple modules have trivial sources, by Proposition 3.19, and thus simple
Green correspondents, by Okuyama’s Theorem [36]. Let P 2. M be a Sylow 3-subgroup of 2. M .2,
i.e. a defect group of B 4 . Then N 2. M .2 ( P ) ∼
= N 2. M ( P ).2, and there are 14 simple F [ N 2. M .2 ( P )]-modules
seven of which belong to the Brauer correspondent b4 of B 4 . Among these are four of dimension 1
and three of dimension 2. Denoting the simple F N 2. M ( P )-modules as in Remark 3.18, we may label
the simple F N 2. M .2 ( P )-modules belonging to b4 as 15 , 16 , 17 = 1∗6 , 18 = 1∗5 , 23 , 25 , 26 such that:
2. M .2 Brauer N 2. M . 2 ( P ) 2. M N 2. M ( P )
D (101 )2.22.2 ϕ11,0 , ϕ11,1 15 D (10)2.22 15
D (102 )2.22.2 ϕ11,0 , ϕ11,1 16 D (10)2.22 15
D (101 )∗2.22.2 ϕ12,0 , ϕ12,1 1∗5 D (10)∗2.22 1∗5
D (102 )∗2.22.2 ϕ12,0 , ϕ12,1 1∗6 D (10)∗2.22 1∗5
D (561 )2.22.2 ϕ13,0 , ϕ13,1 25 D (56)2.22 21
D (562 )2.22.2 ϕ13,0 , ϕ13,1 26 D (56)2.22 21
D (308)2.22.2 ϕ17 23 D (154)2.22 ⊕ D (154)∗2.22 17 ⊕ 18
The entries of this table should be read as follows: the first column displays the simple modules
in B 4 , the fourth column their restrictions to 2. M, and the third column their Green correspondents
in N 2. M .2 ( P ). The last column contains the restrictions of these Green correspondents to N 2. M ( P ).
Column “Brauer” lists for each simple module in B 4 the corresponding possible Brauer characters.
Again the ϕi ,0 denote the Brauer characters whose values are printed in [26]. Let c and d be standard
generators of 2. M .2 as in [42], let ζ := exp(2π i /8), and let − : Z[ζ ] −→ F9 be the lifting map as in
Section 3.2. Then g := ccdcdcdcddcdd and h := g 3 are non-conjugate elements of order 14. Moreover,
x := cdcdd has order 10. The modular character values of the above simple modules on these elements
are as follows:
Module g h x
D (101 )2.22.2 F9 .1 + 2 = b7 ∗ ∗ 2F9 .1 = b7 0
D (102 )2.22.2 F9 .13 = −b7 ∗ ∗ F9 .1 = −b7 0
D (101 )∗2.22.2 b7 b7 ∗ ∗ 0
D (101 )∗2.22.2 −b7 −b7 ∗ ∗ 0
D (561 )2.22.2 0 0 −(F9 .1 + 1) = −r5
D (562 )2.22.2 0 0 F9 .1 + 1 = r5
Remark 3.23. Finally consider 4. M .2. There are 16 blocks of F [4. M .2], and in view of our previous
results we only need to investigate the faithful block B 7 of defect 2 containing five simple modules.
These are precisely the inductions of the simple F [4. M ]-modules belonging to the faithful blocks B 5
and B 6 of F [4. M ]. Here we actually have Morita equivalences between B 7 and each of the blocks B 5
and B 6 of F [4. M ], induced by the respective induction functors. Again this is Fong’s first correspon-
dence, see [30]. Similarly, via the induction functors, we also have Morita equivalences between the
Brauer correspondent b7 of B 7 and each of the Brauer correspondents b5 and b6 of the blocks B 5
and B 6 , respectively. Consequently, any simple F [4. M .2]-module in B 7 and the corresponding simple
F [4. M ]-modules belonging to B 5 and B 6 , respectively, have common vertices and common sources.
Furthermore, the Loewy structures of their Green correspondents coincide.
Remark 3.24. There are seven blocks of F M 23 , namely the principal block B 1 of defect 2 whose defect
groups are elementary abelian, the block B 4 of defect 1, and the blocks B 2 , B 3 = B ∗2 , B 5 , B 6 = B ∗5
and B 7 of defect 0. Thus also all simple F M 23 -modules have the defect groups of their blocks as
vertices, by Knörr’s Theorem [27].
(a) The simple modules belonging to the principal block B 1 are:
D (1)23 = F ↔ ϕ1 , D (22)23 ↔ ϕ2 , D (104)23 ↔ ϕ6 ↔ b, D (104)∗23 ↔ ϕ5 ↔ a,
D (253)23 ↔ ϕ8 , D (770)23 ↔ ϕ10 , D (770)∗23 ↔ ϕ9 ↔ a.
The normalizer N M 23 ( P ) of a Sylow 3-subgroup P of M 23 has order 24 · 32 = 144. Actually,

N M 23 ( P ) ∼
= M 9 : 2, and we denote the seven simple F [ M 9 : 2]-modules as in Remark 3.16. The Green
correspondence with respect to ( M 23 , P , N M 23 ( P )) is denoted by f 1 .
(b) The block B 4 contains precisely one simple F M 23 -module, namely D (231)23 ↔ ϕ7 . Let C ∼ = C3
be a defect group of B 4 . The normalizer N M 23 (C ) has order 5 · 32 · 23 = 360, and it is isomorphic to
a split extension (A5 × C 3 ) : 2. Furthermore, N M 23 (C )/C ∼ = S5 , and F [ N M 23 (C )] has five simple mod-
ules. These are the inflations of the simple F S5 -modules. Precisely one of them is projective, namely
2
the inflation of the simple F S5 -module D (3,1 ) of dimension 6. Denoting the Green correspondence
with respect to ( M 23 , C , N M 23 (C )) by f 4 , we then have:
Proposition 3.25. All simple F M 23 -modules have the defect groups of their blocks as vertices. The Loewy series
of the sources and Green correspondents of the simple modules in B 1 and B 4 are as follows:
Block B1
Module D (1)23 D (22)23 D (104)23 D (104)∗23 D (253)23 D (770)23 D (770)∗23

14 21
Green F 13 23 22 12 22 23
21 14

F F F
Source F F F F F F F F F
F F F
Block B4
Module D (231)23
N M 23 (C ) 2
Green InfC ( D (3,1 ) )
Source F
Proof. Again the assertion about the vertices is obvious. The simple module D (22)23 is a composition
factor, and hence a direct summand, of the natural permutation F M 23 -module on 23 points. Therefore,
D (22)23 has trivial source. Furthermore, Res P 23 ( D (104)23 ) = V ⊕ cyc, Res P 23 ( D (253)23 ) ∼
M M
= F ⊕ cyc and
Res P 23 ( D (770)23 ) ∼
M
= F ⊕ F ⊕ cyc. Here V is indecomposable of dimension 5 and therefore a source
of D (104)23 . By Okuyama’s Theorem [36], D (22)23 , D (770)23 , D (770)∗23 and D (253)23 have simple
Green correspondents. Their actual isomorphism types, the Loewy series of the Green correspondents
of the remaining simple modules in B 1 and the sources of D (104)23 and D (104)∗23 have been deter-
mined computationally.
Finally, consider D (231)23 . By [23, Sec. 6.6 and L. 4.4.12], it has a simple Green correspondent and
thus, by [35, Thm. 4.7.8], a trivial source. In view of Remark 3.24, when regarded as F S5 -module,
2
f 4 ( D (231)23 ) has to be the unique projective simple F S5 -module D (3,1 ) . This proves the proposi-
tion. 2
Remark 3.26. There are six blocks of F M 24 . These are the principal block B 1 of defect 3 whose defect
groups are extraspecial of exponent 3, the blocks B 2 , B 3 = B ∗2 , B 4 , B 5 of defect 1, and B 6 of defect 0.
The simple modules belonging to the blocks B 2 , . . . , B 6 clearly have the respective defect groups as
their vertices.
(a) The simple F M 24 -modules belonging to B 1 are:
D (1)24 = F ↔ ϕ1 , D (22)24 ↔ ϕ2 , D (231)24 ↔ ϕ5 , D (483)24 ↔ ϕ7 ,
D (770)24 ↔ ϕ9 ↔ b, D (770)∗24 ↔ ϕ8 ↔ a, D (1243)24 ↔ ϕ13 .
The normalizer N M 24 ( P ) of some Sylow 3-subgroup P of M 24 has order 216, and N M 24 ( P )/ P is

isomorphic to the dihedral group D 8 of order 8. Consequently, the simple F [ N M 24 ( P )]-modules are
precisely the inflations of the simple F D 8 -modules, and are denoted by F = 11 , 12 , 13 , 14 , 2. The mod-
ules 12 and 14 are interchanged by an outer automorphism of N M 24 ( P ). The Green correspondence
with respect to ( M 24 , P , N M 24 ( P )) will be denoted by f 1 .
(b) By [37], M 24 has two conjugacy classes of subgroups of order 3. If C 3,1 and C 3,2 are represen-
tatives for these, then N M 24 (C 3,1 ) ∼ = L 2 (7) × S3 and N M 24 (C 3,2 ) is isomorphic to an extension 3.S6 .
We will from now on identify N M 24 (C 3,1 ) with L 2 (7) × S3 and N M 24 (C 3,2 ) with 3.S6 via these
isomorphisms. As subgroups of S24 the group C 3,1 is generated by a product of eight disjoint
3-cycles, and C 3,2 is generated by a product of six disjoint 3-cycles. The defect groups of B 2 , B 3
and B 5 are conjugate to C 3,1 , and the defect groups of B 4 are conjugate to C 3,2 . By [26], F L 2 (7)
has five simple modules: F = 11 , 31 , 32 = 3∗1 , 61 , 71 . Hence F [ L 2 (7) × S3 ] has ten simple modules:
F = 11 F , F sgn, 31 F , 32 F , 31 sgn, 32 sgn, 61 F , 61 sgn, 71 F , 71 sgn. Here, sgn de-
notes the alternating F S3 -module D (2,1) . Note that L 2 (7)×S3 possesses exactly one normal subgroup
isomorphic to S3 , namely H := {(1, x) | x ∈ S3 }. For i = 2, 3, 5 we denote the Brauer correspondent
of B i by b i .
The group algebra F [3.S6 ] has seven simple modules. Namely, by [35, Thm. 4.7.8], each simple
F [3.S6 ]-module is also a simple F S6 -module, and there are seven of those. There are two projective
2 2
simple F S6 -modules: D (4,2) and D (2 ,1 ) = D (4,2) ⊗ sgn which have dimension 9. We denote the
Brauer correspondent of B 5 by b5 . For i = 2, 3, 4, 5 and a suitable labelling, the blocks B i and b i ,
respectively, contain the following simple modules:
B2 B3 B5 B4
D (45)24 ↔ ϕ3 ↔ a D (45)∗
24 ↔ ϕ4 ↔ b D (1035)24 ↔ ϕ12 D (252)24 ↔ ϕ6
D (990)24 ↔ ϕ10 D (990)∗24 ↔ ϕ11 ↔ b D (2277)24 ↔ ϕ14 D (5544)24 ↔ ϕ15
b2 b3 b5 b4
NM ( C 3, 2 )
31 F 32 F 61 F InfC 3,224 ( D (4,2) )
N M ( C 3, 2 ) 2 2
31 sgn 32 sgn 61 sgn InfC 3,224 ( D (2 ,1 ) )
With the above notation, the following holds:
Proposition 3.27. Apart from the module D (483)24 , all simple F M 24 -modules belonging to the principal block
have vertex P ∈ Syl3 ( M 24 ). The vertices of D (483)24 are the M 24 -conjugates of a subgroup of M 24 of order 9
whose normalizer in M 24 is isomorphic to the automorphism group of M 9 which is a split extension M 9 : S3 .

Furthermore, the Loewy series of the sources are as follows:
Module D (1)24 D (22)24 D (231)24 D (483)24 D (770)24 D (770)∗24 D (1243)24

Sce. 1 4 3 1 5 5 19
Layer dims. 1 1, 2, 1 1, 1, 1 1 1, 2, 2 2, 2, 1 2, 3, 4, 2, 4, 2, 2
Proof. The modules D (1)24 , D (22)24 , D (770)24 , D (770)∗24 and D (1243)24 obviously have vertex P ,
M
since their dimensions are not divisible by 3. Moreover, Res P 24 ( D (231)24 ) = V 1 ⊕ V 2 ⊕ cyc, where
V 1 and V 2 are indecomposable of dimension 3 with vertex P and therefore sources of D (231)24 .
M
For D (483)24 we have Res P 24 ( D (483)24 ) = U 1 ⊕ U 2 ⊕ cyc, where U 1 and U 2 are indecomposable of
dimension 3 and have vertex Q of order 9. Actually, N M 24 ( Q ) ∼
= Aut( M 9 ) ∼
= M 9 : S3 which determines
N M 24 ( Q ) up to M 24 -conjugacy, by [37]. The Loewy series of the sources of the non-trivial simple
F M 24 -modules belonging to B 1 have been obtained by computer calculations. 2
Remark 3.28. Let Q be a vertex of D (483)24 . By the above proposition, its normalizer N M 24 ( Q ) is
isomorphic to M 9 : S3 . The factor group N M 24 ( Q )/ Q is isomorphic to GL(2, 3) which is one of the
double covers of S4 . So each simple F [ N M 24 ( Q )]-module can be regarded as simple F [GL(2, 3)]-
module. Furthermore, F [GL(2, 3)] has six simple modules two of which are projective. The latter
2
are precisely the inflations of the projective simple F S4 -modules D (3,1) and D (2,1 ) . We denote the
2
simple F N M 24 ( Q )-module obtained from D (3,1) via inflation as 31 , and the one obtained from D (2,1 )
via inflation by 32 . For the following, let further f be the Green correspondence with respect to
( M 24 , Q , N M 24 ( Q )).
Proposition 3.29. With the notation of Remarks 3.26 and 3.28, the Green correspondents of the simple F M 24 -
modules belonging to the principal block have the following Loewy series:
Module D (1)24 D (22)24 D (231)24 D (770)24 D (770)∗24 D (1243)24 D (483)

⎡ ⎤
2
⎢ 11 12 13 ⎥
⎢ ⎥
12 2 14 13 14 ⎢ 22 ⎥
⎢ ⎥
Green F 2 11 13 2 2 ⎢ 13 14 ⎥ 31
⎢ ⎥
12 2 13 14 14 ⎢ 22 ⎥
⎣ 1 1 ⎦
1 2
2
Proof. Since D (483)24 has trivial source, its Green correspondent f ( D (483)24 ) is simple, by
Okuyama’s Theorem [36]. Furthermore, f ( D (483)24 ) can also be regarded as a simple projective
F [ N M 24 ( Q )/ Q ]-module. By the previous remark, f ( D (483)24 ) has to be the inflation of one of the
projective simple F S4 -modules. We have checked that it is actually isomorphic to that of D (3,1) .
The remaining Green correspondents and their Loewy series have been determined with the com-
puter. In order to determine f 1 ( D (1243)24 ), let W be a source of D (1243)24 , and set N := N M 24 ( P ). By
Proposition 3.27, dim( W ) = 19. Furthermore, our computations show that Ind NP ( W ) = U 1 ⊕ · · · ⊕ U 6
where U 1 , . . . , U 6 are pairwise non-isomorphic indecomposable modules such that dim(U i ) = 19
and dim(U j ) = 38, for i ∈ {1, 2, 3, 4} and j ∈ {5, 6}. Precisely one of these summands is isomor-
M
phic to f 1 ( D (1243)24 ). Now Res P 24 ( D (1243)24 ) = W ⊕ 2W 1 ⊕ 2W 2 ⊕ proj where W 1 and W 2
are non-isomorphic indecomposable modules of dimension 9. Our computations also show that
M
Res N 24 ( D (1243)24 ) has a submodule U isomorphic to one of the 19-dimensional modules U 1 , . . . , U 4
such that U has vertex P , and

Res NP (U ) ⊕ Res NP Res N 24 D (1243)24 /U ∼
M
= W ⊕ 2W 1 ⊕ 2W 2 ⊕ proj.
M
Hence Res NP (U ) is a direct summand of Res P 24 ( D (1243)24 ), by [10, Thm. 1.5.8]. Since U is relatively
M
P -projective, [14, Thm. 19.2] implies that U is a direct summand of Res N 24 ( D (1243)24 ). Consequently,
f 1 ( D (1243)24 ) ∼
= U , and its Loewy series has been computed to be as stated. 2
Proposition 3.30. The simple F M 24 -modules belonging to the blocks B 2 , B 3 , B 4 and B 5 have the defect groups
of their blocks as vertices. Moreover, all these modules have trivial sources, and the following Green correspon-
dents:
Block B2 B3 B5
Module D (45)24 D (990)24 D (45)∗24 D (990)∗24 D (1035)24 D (2277)24
Green 31 sgn 31 F 32 sgn 32 F 61 F 61 sgn
Block B4
Module D (252)24 D (5544)24
N M ( C 3, 2 ) NM ( C 3, 2 ) 2
, 12 ) )
Green InfC 3,224 ( D (4,2) ) InfC 3,224 ( D (2
Proof. The assertion about the vertices is obvious, and, by [23, Sec. 6.9 and L. 4.4.12], we also know
that all of these simple modules have simple Green correspondents and therefore trivial sources,
by [35, Thm. 4.7.8]. Let f 2 be the Green correspondence with respect to ( M 24 , C 3,1 , N M 24 (C 3,1 )), and
let f 4 be the Green correspondence with respect to ( M 24 , C 3,2 , N M 24 (C 3,2 )). We have constructed
f 2 ( D (45)24 ), f 2 ( D (1035)24 ) and f 4 ( D (252)24 ). The restriction of f 2 ( D (45)24 ) to the normal subgroup
H∼ = S3 of N M 24 (C 3,1 ) decomposes into the direct sum of three copies of the alternating module,
and the restriction of f 2 ( D (1035)24 ) to H decomposes into the direct sum of six copies of the trivial
module. This determines the Green correspondents of the simple modules in B 2 , B 3 and B 5 . Moreover,
we have computed that f 4 ( D (252)24 ) is isomorphic to D (4,2) when regarded as F S6 -module. This
then also yields the claimed assertion on the Green correspondents of the simple modules in B 4 . 2
4. Morita equivalent blocks
In the following, let p = 3, M := M 22 , G = 2. M and P ∈ Syl3 (G ). As we have seen in Proposi-

tion 3.19, all simple F G-modules belonging to the faithful block B := B 3 of F G with defect group P
have vertex P , trivial sources and simple Green correspondents. The aim of this section is to show
that B is Morita equivalent to its Brauer correspondent b in N G ( P ) =: H .
The block B is an indecomposable F [G × G ]-module with vertex P = {( g , g ) | g ∈ P } and trivial
sources, by [35, Thm. 5.10.8]. In particular, G × H and H × G both contain N G ×G ( P ). We write
V := ResGG × G
×H ( B ) = V 0 ⊕ V 1 and W := ResGH× G
×G ( B ) = W 0 ⊕ W 1 ,
with submodules V 0 , V 1 , W 0 , W 1 where V 0 and W 0 are the Green correspondents of B in G × H and

H × G, respectively. Thus V 0 is an indecomposable F [G × H ]-module with vertex P belonging to
the block B ⊗ b, and W 0 is an indecomposable F [ H × G ]-module with vertex P belonging to the
block b ⊗ B. We will show that the functor W 0 ⊗ B − gives a Morita equivalence between B and b.
Remark 4.1. (a) We recall that P ∼ = C 3 × C 3 , | H | = 144 and C G ( P ) = C H ( P ) ∼

= C 3 × C 6 . Moreover, if
Q is a subgroup of P of order 3 then C G ( Q ) ∼ = A4 × C 6 where A4 denotes the alternating group of
degree 4, and C H ( Q ) = C H ( P ) ∼
= C3 × C6.
(b) We denote the block idempotents of B and b by e and f , respectively. Then f is the unique
non-principal block idempotent of F C G ( P ), by (a), and f = z − 1 where Z (G ) = z. Thus e f = e.
By making use of Theorem 2.13, we now show:

Proposition 4.2. The functors V 0 ⊗b − and W 0 ⊗ B − induce a stable equivalence of Morita type between B
and b.
Proof. Since V 0 has vertex P and trivial source, it suffices to show that condition (2) of The-
orem 2.17 is satisfied, with V 0 and W 0 in place of V and W . We show first that the func-
tors V 0 ( P ) ⊗ F C H ( P ) Br P ( f ) − and W 0 ( P ) ⊗ F C G ( P ) Br P (e) − induce Morita equivalences between
F C G ( P ) Br P (e ) and F C H ( P ) Br P ( f ). In fact, by Remark 4.1, we have C G ( P ) = C H ( P ) and Br P (e ) =
Br P ( f ) = f = z − 1. Similarly,
V 0 ( P ) ⊕ V 1 ( P ) = V ( P ) = B ( P ) = F C G ( P ) Br P (e ) = F C G ( P ) f
is an indecomposable F C G ( P ) f -F C G ( P ) f -bimodule. Since V 0 ( P ) = 0, this implies V 1 ( P ) = 0. In

the same way, W 0 ( P ) = F C G ( P ) f is an indecomposable F C G ( P ) f -F C G ( P ) f -bimodule. Trivially the
functor F C G ( P ) f ⊗ F C G ( P ) f − induces a Morita equivalence between F C G ( P ) f and F C G ( P ) f .
Now let Q be a subgroup of P of order 3. It remains to show that the functors
V 0 ( Q ) ⊗ F C H ( Q ) Br Q ( f ) − and W 0 ( Q ) ⊗ F C G ( Q ) Br Q (e) −
induce Morita equivalences between F C G ( Q ) Br Q (e ) and F C H ( Q ) Br Q ( f ). By Remark 4.1, we have

C H (Q ) ∼= C 3 × C 6 and Br Q ( f ) = f = z − 1. Also, we have C G ( Q ) ∼ = A4 × C 6 . Note that Br Q (e ) f =
Br Q (e ) Br Q ( f ) = Br Q (e f ) = Br Q (e ). Since Br Q (e ) is a sum of block idempotents with defect group P ,
this means that Br Q (e ) = e 1 f where e 1 is the principal block idempotent of F A4 . Since F A4 e 1 ∼ = F C3,
this implies that F C G ( Q ) Br Q (e ) ∼ = F [C 3 × C 3 ] ∼
= F C H ( Q ) Br Q ( f ). Arguing as above, we also obtain
V 0 ( Q ) = F C G ( Q ) Br Q (e ) ∼
= F C H ( Q ) Br Q ( f )
and V 1 ( Q ) = 0. Hence the functors V 0 ( Q ) ⊗ F C H ( Q ) Br Q ( f ) − and W 0 ( Q ) ⊗ F C G ( Q ) Br Q (e) − trivially

induce Morita equivalences between the blocks F C G ( Q ) Br Q (e ) and F C H ( Q ) Br Q ( f ). 2
We now prove the main result of this section:
Theorem 4.3. The functor W 0 ⊗ B − induces a Morita equivalence between B and b.
Proof. By Proposition 4.2 and Theorem 2.13(ii), it suffices to show that W 0 ⊗ B D is a simple b-module,
for every simple B-module D. But
W 0 ⊗ B D | W ⊗ B D = ResGH× G G
×G ( B ) ⊗ B D = Res H ( D ),
and W 0 ⊗ B D is an indecomposable non-projective F H -module, by Proposition 4.2 and Theo-

rem 2.13(i). Since computer calculations show that ResGH ( D ) has a unique non-projective indecompos-
able direct summand, W 0 ⊗ B D is the Green correspondent of D in any case, and Proposition 3.19(i)
shows that W 0 ⊗ B D is in fact a simple b-module. This proves the theorem. 2
To close, we now consider the bicyclic extension

G := 2. M .2 of M and the faithful block B 4 of F G
which has defect group P . As we have seen in Remark 3.22, also all simple modules belonging to B 4
have vertex P , trivial sources and simple Green correspondents. In analogy to the previous theorem,
we will show that B 4 is Morita equivalent to its Brauer correspondent b4 in N
G ( P ) =: H . For this we
set
G ×

V := Res G
(B4) = V 1 and
V0 ⊕ W := ResG ×G ( B 4 ) =
W0 ⊕
W 1,
G ×
H H ×G
where V 0 and G×
W 0 denote the Green correspondents of B 4 in H ×
H and G, respectively. With
this we now obtain:
Theorem 4.4. The functor

W 0 ⊗ B 4 − induces a Morita equivalence between the blocks B 4 and b4 .
Proof. The proof is similar to that of Theorem 4.3, so we omit the details here. Notice that
∼
H ( P ) = C
C G ( P ) = C G ( P ) = C3 × C6,
and if Q is any subgroup of P of order 3 then
∼ ∼
G ( Q ) = S4 × C 6
C and H ( Q ) = S3 × C 6 .
C
Arguing as in the proof of Proposition 4.2 and applying Theorem 2.17, we deduce that the functors

V 0 ⊗b4 − and W 0 ⊗ B 4 − induce a stable equivalence of Morita type between the blocks B 4 and b4 .

Moreover, our computations show that, given any simple B 4 -module D, the F
H -module ResG
( D ) has
H
a unique non-projective indecomposable direct summand, namely the simple Green correspondent
of D in
H . Since, by Theorem 2.13(i),

W 0 ⊗ B 4 D | ResG
H
(D)
is indecomposable and non-projective, W 0 ⊗ B 4 D is the Green correspondent of D and thus a simple

b4 -module. The assertion of the theorem now follows from Theorem 2.13(ii). 2
Acknowledgments
The authors are grateful to the Mathematical Sciences Research Institute in Berkeley for its hos-
pitality and financial support during their stay in the framework of the program ‘Representation
Theory of Finite Groups and Related Topics’. The first author’s research was also supported by the
Deutsche Forschungsgemeinschaft through grant # DA 1115/1-1. Moreover, both authors wish to thank
R. Kessar, S. Koshitani, M. Linckelmann, K. Lux and J. Müller for helpful discussions on the topic of this
note. They also gratefully acknowledge the comments of referees on earlier versions of the manuscript.
References
[1] J.L. Alperin, Weights for finite groups, in: The Arcata Conference on Representations of Finite Groups, Arcata, CA, 1986, in:
Proc. Sympos. Pure Math., vol. 47, Part 1, Amer. Math. Soc., Providence, RI, 1987, pp. 369–379.
[2] F.W. Anderson, K.R. Fuller, Rings and Categories of Modules, Grad. Texts in Math., vol. 13, Springer-Verlag, New York, 1992.
[3] M. Aschbacher, Finite Group Theory, second ed., Cambridge Stud. Adv. Math., vol. 10, Cambridge Univ. Press, Cambridge,
2000.
[4] W. Bosma, J. Cannon, C. Playoust, The Magma algebra system. I. The user language, J. Symbolic Comput. 24 (3–4) (1997)
235–265.
[5] R. Brauer, A note on theorems of Burnside and Blichfeldt, Proc. Amer. Math. Soc. 15 (1964) 31–34.
[6] M. Broué, On Scott modules and p-permutation modules: An approach through the Brauer morphism, Proc. Amer. Math.
Soc. 93 (1985) 401–408.
[7] M. Broué, Equivalences of blocks of group algebras, in: Finite-Dimensional Algebras and Related Topics, Ottawa, ON, 1992,
in: NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., vol. 424, Kluwer Acad. Publ., Dordrecht, 1994, pp. 1–26.
[8] R.M. Bryant, L.G. Kovács, Tensor products of representations of finite groups, Bull. London Math. Soc. 4 (1972) 133–135.
[9] N. Burgoyne, P. Fong, Multipliers of the Mathieu groups, Nagoya Math. J. 27 (1966) 733–745.
[10] J.F. Carlson, L. Townsley, L. Valero-Elizondo, M. Zhang, Cohomology Rings of Finite Groups, Algebr. Appl., vol. 3, Kluwer
Acad. Publ., Dordrecht, 2003.
[11] J.F. Carlson, Support varieties for modules, in: Discovering Mathematics with Magma, in: Algorithms Comput. Math., vol. 19,
Springer, Berlin, 2006, pp. 187–204.
[12] J. Conway, R. Curtis, S. Norton, R. Parker, R. Wilson, Atlas of Finite Groups, Clarendon Press, Oxford, 1985.
[13] D.A. Craven, Algebraic modules for finite simple groups, PhD thesis, Oxford University, 2007.
[14] C.W. Curtis, I. Reiner, Methods of Representation Theory, vol. I, With Applications to Finite Groups and Orders, Pure Appl.
Math., A Wiley–Interscience Publication, John Wiley & Sons, Inc., New York, 1981.
[15] S. Danz, Theoretische und algorithmische Methoden zur Berechnung von Vertizes irreduzibler Moduln symmetrischer
Gruppen, PhD thesis, University of Jena, 2007, http://www.minet.uni-jena.de/algebra/preprints/preprints.html.
[16] S. Danz, B. Külshammer, R. Zimmermann, On vertices of simple modules for symmetric groups of small degrees, J. Alge-
bra 320 (2008) 680–707.
[17] K. Erdmann, Blocks and simple modules with cyclic vertex, Bull. London Math. Soc. 9 (1977) 216–218.
[18] W. Feit, The Representation Theory of Finite Groups, North-Holland Math. Library, vol. 25, North-Holland Publishing Co.,
Amsterdam, New York, 1982.
[19] The GAP group, GAP-4 – Groups, algorithms and programming, version 4.4.9, 2006.
[20] H. Gollan, Die 3-modularen Darstellungen der Mathieu-Gruppen M 11 und M 12 , Diploma thesis, Essen, 1985.
[21] J.A. Green, On the indecomposable representations of a finite group, Math. Z. 70 (1959) 430–445.
[22] M.E. Harris, A note on the Green invariants in finite group modular representative theory, Results Math. 51 (2008) 249–259.
[23] G. Hiss, K. Lux, Brauer Trees of Sporadic Groups, Oxford Science Publications, Clarendon Press, Oxford University Press,
Oxford, 1989.
[24] B. Huppert, N. Blackburn, Finite Groups II, Grundlehren Math. Wiss., vol. 243, Springer-Verlag, Berlin, New York, 1982.
[25] G.D. James, A. Kerber, The Representation Theory of the Symmetric Group, Encyclopedia Math. Appl., vol. 16, Addison–
Wesley, Reading, 1981.
[26] C. Jansen, K. Lux, R. Parker, R. Wilson, An Atlas of Brauer Characters, Clarendon Press, Oxford, 1995.
[27] R. Knörr, On the vertices of irreducible modules, Ann. of Math. (2) 110 (1979) 487–499.
[28] S. Koshitani, N. Kunugi, Broué’s conjecture holds for principal 3-blocks with elementary abelian defect group of order 9,
J. Algebra 248 (2002) 575–604.
[29] S. Koshitani, K. Waki, The Green correspondents of the Mathieu group M 12 in characteristic 3, Comm. Algebra 27 (1) (1999)
37–66.
[30] B. Külshammer, On p-blocks of p-solvable groups, Comm. Algebra 9 (1981) 1763–1785.
[31] B. Külshammer, Some indecomposable modules and their vertices, J. Pure Appl. Algebra 86 (1993) 65–73.
[32] M. Linckelmann, Stable equivalences of Morita type for self-injective algebras and p-groups, Math. Z. 223 (1996) 87–100.
[33] P. Mazet, Sur le multiplicateur de Schur du groupe de Mathieu M 22 , C. R. Acad. Sci. Paris Sér. A 289 (1979) 659–661.
[34] J. Müller, M. Schaps, The Broué conjecture for the faithful 3-blocks of 4. M 22 , J. Algebra 319 (2008) 3588–3602.
[35] H. Nagao, Y. Tsushima, Representations of Finite Groups, Academic Press, San Diego, 1989.
[36] T. Okuyama, Module correspondence in finite groups, Hokkaido Math. J. 10 (1981) 299–318.
[37] G. Pfeiffer, The subgroups of M 24 , 1995; http://schmidt.nuigalway.ie/~goetz/pub/m24sub.html.
[38] G.J.A. Schneider, The vertices of the simple modules of M 12 over a field of characteristic 2, J. Algebra 83 (1983) 189–200.
[39] G.J.A. Schneider, Computing with endomorphism rings of modular representations, J. Symbolic Comput. 9 (1990) 607–636.
[40] M. Szőke, Examining Green correspondents of weight modules, PhD thesis, RWTH Aachen University, 1998.
[41] J. Thévenaz, G-Algebras and Modular Representation Theory, Oxford Math. Monogr., Oxford Science Publications, Clarendon
Press, Oxford University Press, New York, 1995.
[42] R. Wilson, et al., Atlas of finite group representations, Version 3, http://brauer.maths.qmul.ac.uk/Atlas/v3/.
[43] R. Zimmermann, Vertizes einfacher Moduln der Symmetrischen Gruppen, PhD thesis, University of Jena, 2004, http://www.
minet.uni-jena.de/algebra/preprints/preprints.html.
Journal of Algebra
Constructing modular separating invariants ✩

Müfit Sezer
Department of Mathematics, Bilkent University, Ankara 06800, Turkey
Article history: We consider a finite dimensional modular representation V of

Received 2 December 2008 a cyclic group of prime order p. We show that two points in V
Available online 25 July 2009 that are in different orbits can be separated by a homogeneous
invariant polynomial that has degree one or p and that involves
Keywords:
variables from at most two summands in the dual representation.
Separating invariants Simultaneously, we describe an explicit construction for a separat-
Modular groups ing set consisting of polynomials with these properties.
Introduction
Let V denote a finite dimensional representation of a group G over a field F . The induced ac-
tion on the dual space V ∗ extends to the symmetric algebra F [ V ] := S ( V ∗ ) of polynomial functions
on V . More precisely, the action of σ ∈ G on f ∈ F [ V ] is given by (σ f )( v ) = f (σ −1 v ) for v ∈ V .
The subalgebra in F [ V ] of polynomials that are left fixed under the action of the group is denoted
by F [ V ]G . Any invariant polynomial f ∈ F [ V ]G is constant on the G-orbits in V . A subset A ⊆ F [ V ]G
is said to be separating (for V ) if for any pairs of vectors v , w ∈ V , we have: If f ( v ) = f ( w ) for all
f ∈ A, then f ( v ) = f ( w ) for all f ∈ F [ V ]G . If G is finite, this is equivalent to saying that whenever
v , w ∈ V are in different G-orbits, there exists f ∈ A such that f ( v ) = f ( w ). Although the concept of
separating invariants dates back to the origins of the invariant theory there has been a recent interest
in the topic initiated by Derksen and Kemper [2] who pointed out that one can get nice separating
subalgebras as opposed to the full invariant ring which is often complicated in terms of constructive
and ring theoretical considerations. For instance, there always exists a finite separating set [2, 2.3.15]
and the Noether bound (for finite groups) holds with no restriction on the characteristic of the field
[2, 3.9.14]. Since then, separating invariants have been studied by several people and further evidence
for their well behavior has been revealed, see [3,5–7,9,13,14]. We direct the reader to [2, 2.3.2, 3.9.4]
and [12] for more background and motivation on the subject.
✩
Research supported by a grant from Tübitak: 109T384.
E-mail address: sezer@fen.bilkent.edu.tr.
doi:10.1016/j.jalgebra.2009.07.011
4100 M. Sezer / Journal of Algebra 322 (2009) 4099–4104
In this paper we study separating invariants for representations of a cyclic group Z / p of prime
order p, over a field F of characteristic p. Invariants of cyclic p-groups over characteristic p are dif-
ficult to describe. Although exact degree bounds for the algebra generators for the invariant rings of
all representations of Z / p are known [10], explicit generating sets are available only for handful of
cases. Actually, generating sets for the two and the three dimensional indecomposable representa-
tions of Z / p were given by Dickson [4] as early as the beginning of the twentieth century. It turns
out that these invariant rings are generated by two and four elements respectively. But things get
complicated very quickly. The only other indecomposable representations where a generating set for
the corresponding ring of invariants are known are the four and the five dimensional representations
which were computed by Shank through difficult computations, see [15]. His methods were later used
to work out some decomposable cases. A generating set that applies to all representations of Z / p is
described by Hughes and Kemper [11]. Their set consists of norms (orbit products) of some variables,
transfers (orbit sums) and invariants up to a certain quite optimal degree. The reason to include in-
variants up to some degree is that norms and transfers can be employed to decompose invariants
only after some degree and there are invariants in small degrees that are not expressible using norms
and transfers. In [13] it is shown that this uncertainty in small degrees is not an issue for separating
purposes: The sum of relative transfers with respect to maximal subgroups together with the norms
of certain variables is separating for any representation of any p-group. But unfortunately this sep-
arating set is infinite dimensional as a vector space. In this paper we restrict ourselves to Z / p and
show that a separating set for an indecomposable representation V n of dimension n can be obtained
by adding to any separating set for the indecomposable subrepresentation V n−1 some explicitly de-
scribed transfers and the norm of the terminal variable of V n∗ , see Theorem 3. The set which we add
to a separating set for V n−1 consists of polynomials of degree p. Inductively this yields a separating
set of polynomials of degree one or p for an indecomposable representation V n , see Remark 4. But
the size of the separating set for V n obtained from Theorem 3 is not optimal, see again the discussion
in Remark 4.
Next we consider decomposable representations. A major result concerning decomposable repre-
sentations is that the polarization of separating invariants yields a separating set over any character-
istic, see Draisma et al. [6] which does not hold for generating invariants. A result by Domokos [5]
states that for the direct sum of any number copies of a representation V there exists a separating
set of polynomials each of which involve variables from at most 2n summands in V ∗ , where n is the
dimension of V . If the group is reductive 2n can be replaced by n + 1. We obtain a sharpening of
this result for Z / p as follows. Let W be a Z / p representation over characteristic p. We show that
the separating invariants for a particular proper subrepresentation of W union separating invariants
for the indecomposable summands in W together with an explicitly constructed set of transfers form
a separating set for W , see Theorem 6. These transfers involve variables from two summands only
and are of degree p. Hence we obtain by induction that for any representation W of Z / p there is a
separating set consisting of degree one and degree p polynomials that involve variables from at most
two summands in W ∗ .
Modular separating invariants
Let p > 0 be a prime number. For the rest of the paper G denotes the cyclic group of order p,
and F denotes a field of characteristic p. We fix a generator σ of G. It is well known that there
are exactly p indecomposable representations V 1 , V 2 , . . . , V p of G up to isomorphism where σ acts
on V n for 1 n p by a Jordan block of dimension n with ones on the diagonal. Let e 1 , e 2 , . . . , en
be the Jordan block basis for V n with σ (e i ) = e i + e i −1 for 2 i n and σ (e 1 ) = e 1 . We identify
each e i with the column vector with 1 on the i-th coordinate and zero elsewhere. Let x1 , x2 , . . . , xn
denote the corresponding elements in the dual space V n∗ . Since V n∗ is indecomposable it is isomorphic
to V n . In fact, x1 , x2 , . . . , xn forms a Jordan block basis for V n∗ in the reverse order. We may assume
that σ (xi ) = xi + xi +1 for 1 i n − 1 and σ (xn ) = xn . We have F [ V n ] = F [x1 , x2 , . . . , xn ]. Pick a
column vector (c 1 , c 2 , . . . , cn )t in V n , where c i ∈ F for 1 i n. There is a G-equivariant surjection
V n → V n−1 given by (c 1 , c 2 , . . . , cn )t → (c 2 , c 3 . . . , cn )t . We use the convention that V 0 is the zero
representation. Dual to this surjection, the subspace in V n∗ generated by x2 , x3 , . . . , xn is closed under
M. Sezer / Journal of Algebra 322 (2009) 4099–4104 4101
the G-action and is isomorphic to V n∗−1 . Hence F [ V n−1 ] = F [x2 , x3 , . . . , xn ] sits as a subalgebra in

F [ V n ]. For f ∈ F [ V n ], the norm of f , denoted by N ( f ), is defined by l
0l p −1 σ ( f ). Moreover

define Tr = 0l p −1 σ l , which we call the transfer map. Note that both N ( f ) and Tr( f ) are invariant
polynomials. For a positive integer k, let J k denote the ideal in F [ V n ] generated by xk , xk+1 , . . . , xn if
1 k n and let J k denote the zero ideal if k > n. We need the following well-known fact.

Lemma 1. Let a be a positive integer. Then 0l p −1 l
a
≡ −1 mod p if p − 1 divides a and 0l p −1 l
a
≡0
mod p, otherwise.
Proof. We direct the reader to [1, 9.4] for a proof. 2
p −1
Lemma 2. Let 2 i n − 1 be an integer. Then there exist f 1 , f 2 ∈ F [x2 , x3 , . . . , xn ] such that Tr(x1 xi )=
p −1
f 1 x1 + f 2 . Moreover, f 1 ≡ −xi +1 mod J i +2 .
Proof. Since the vector space generated by x2 , x3 , . . . , xn is closed under the G-action and σ (x1 ) =
p −1
x1 + x2 , it follows that Tr(x1 xi ) as a polynomial in x1 (with coefficients in F [ V n−1 ]) is of degree at
most one. Therefore the first assertion of the lemma follows.
For 0 l p − 1 we have
p −1
l
p −1 l l
σ x1 xi = x1 + lx2 + x3 + · · · xi + lxi +1 + x i +2 + · · · .
2 2
Let a, b be non-negative integers with a + b = p − 1. Then the coefficient of x1 xai xbi+1 in σ l (x1 xip−1 )
p −1 b p −1 p −1 b
is b
l . Therefore the coefficient of in x1 xai xbi+1
is 0l p −1 b
Tr(x1 xi )
l . By the previous
lemma this number is zero unless b = p − 1 and is −1 if b = p − 1. This completes the proof. 2
Let S ⊆ F [ V n−1 ]G be a separating set of invariants for V n−1 . Our next result describes a finite set
of invariant polynomials in F [ V n ]G such that, when added to S, one gets a separating set for V n .
p −1
Theorem 3. Let S ⊆ F [ V n−1 ]G be a separating set for V n−1 . Then S together with N (x1 ), Tr(x1 xi ) for
2 i n − 1 is a separating set for V n .
Proof. Let v 1 = (c 1 , c 2 , . . . , cn )t and v 2 = (d1 , d2 , . . . , dn )t be two column vectors in V n in different

G-orbits, where c i , di ∈ F for 1 i n. If (c 2 , c 3 , . . . , cn )t and (d2 , d3 , . . . , dn )t are in different orbits
in V n−1 , then there exists a polynomial in S that separates these points because S ⊆ F [ V n−1 ]G is
separating. Therefore this polynomial separates v 1 and v 2 as well. Hence by replacing v 2 with a
suitable element in its orbit we may assume that c i = di for 2 i n. Note that with this assumption
we must have c 1 = d1 . First assume that there exists index 3 i n such that c i = di = 0. Let j denote
p −1
the largest integer n such that c j = 0. We show that Tr(x1 x j −1 ) separates v 1 and v 2 as follows. By
p −1
the previous lemma we can write Tr(x1 x j −1 ) = f 1 x1 + f 2 such that f 1 , f 2 ∈ F [x2 , x3 , . . . , xn ] with f 1 ≡
p −1
−x j mod J j +1 . Since c i = di for 2 i n, we have f 2 ( v 1 ) = f 2 ( v 2 ) and f 1 ( v 1 ) = f 1 ( v 2 ). Moreover
p −1
f 1 ( v 1 ) = −c j because c i = 0 for j < i, so f 1 ( v 1 ) (and hence f 1 ( v 2 )) is non-zero. It follows that
f 1 x1 + f 2 separates v 1 and v 2 because the first coordinates of these vectors are different. We now
assume that c i = di = 0 for 3 i n. We show that in this case N (x1 ) separates v 1 and v 2 . Note that
N (x1 )( v 1 ) = 0l p −1 (c 1 + lc 2 ). We define a polynomial Q (x) = 0l p −1 (x + lc 2 ) ∈ F [x]. Notice that
N (x1 )( v 1 ) = Q (c 1 ) and that Q (c 1 ) = Q (c 1 + c 2 ) = Q (c 1 + 2c 2 ) = · · · = Q (c 1 + ( p − 1)c 2 ). Since Q (x) is
a polynomial of degree p, it follows that c 1 , c 1 + c 2 , c 1 + 2c 2 , . . . ,
c 1 + ( p − 1)c 2 are the only solutions
to Q (x) = Q (c 1 ). Therefore if N (x1 )( v 2 ) = 0l p −1 (d1 + ld2 ) = 0l p −1 (d1 + lc 2 ) = Q (d1 ) is equal
to N (x1 )( v 1 ) = Q (c 1 ), then d1 must be equal to c 1 + lc 2 for some 0 l p − 1. Equivalently we must
have σ l ( v 1 ) = v 2 . This is a contradiction because then v 1 and v 2 are in the same G-orbit. 2
Remark 4. The invariants of the two dimensional indecomposable representation V 2 is a regular ring
generated by the fixed variable of V 2∗ and the norm of the terminal variable of V 2∗ , see [4]. The set
in Theorem 3 which we add to a separating set for V n−1 consists of n − 1 polynomials of degree p.
Hence, inductively this yields a separating set of n(n − 1)/2 + 1 polynomials of degree one or p for
an indecomposable representation V n . We note that there is always a separating set of size 2n + 2
for any representation of dimension n of any group. This fact was forwarded to us with a sketch of a
proof during the refereeing process of [13] and it also appears in [8]. However, the proof is not con-
structive as opposed to Theorem 3. We will see that it is possible to obtain separating sets consisting
of polynomials of degree one or p for decomposable representations as well, see Corollary 7.
We have mentioned in the introduction that the computations [15] for the invariants of V 4 and V 5
are difficult and the generating sets are more complicated compared to the two and the three di-
mensional representations. On the other hand our result yields a much more simpler separating
subalgebra. Consider F [x2 , x3 , x4 ] = F [ V 3 ] ⊆ F [x1 , x2 , x3 , x4 ] = F [ V 4 ]. Then F [x2 , x3 , x4 ]G is generated
p −1
by x4 , x23 − 2x2 x4 − x3 x4 , Tr(x3 x4 ), N (x2 ), see [4]. Since these four polynomials form a separating set
p −1 p −1
in F [x2 , x3 , x4 ]G , by the previous theorem, this set together with N (x1 ), Tr(x1 x2 ), Tr(x1 x3 ) is a
separating set in F [x1 , x2 , x3 , x4 ]G . It is instructive to compare this separating set with the generating
set given in [15].
m
We now consider decomposable representations of G. Let W = i =1 W i , where W i is an inde-
composable G representation of dimension q i p, i.e., W i = V qi . Let e i ,1 , e i ,2 , . . . , e i ,qi denote the
standard basis vectors for W i , where e i , j is the column vector of dimension q i with one at the j-th
coordinate and zero elsewhere. As before, we assume that these vectors form a Jordan block basis for
W i with σ (e i , j ) = e i , j −1 for 2 j qi and σ (e i ,1 ) = e i ,1 . Let xi ,1 , xi ,2 , . . . , xi ,qi denote the correspond-
ing elements in the dual W i∗ . Define T i = V qi −1 and recall that there is a G-equivariant surjection
πi : W
i m→ T i , given by (c i ,1 , c i ,2 , . . . , c i ,qi ) → (c i ,2 , c i ,3 , . . . , c i ,qi ) , where c i , j ∈ F for 1 j qi . Define
t t
T = i =1 T i . We identify W as the vector space of m-tuples ( w 1 , w 2 , . . . , w m ) with w i ∈ W i and T

as the vector space of m-tuples (t 1 , t 2 , . . . , tm ) with t i ∈ T i . Then we have a G-equivariant surjection
π : W → T given by ( w 1 , w 2 , . . . , w m ) → (π1 ( w 1 ), π2 ( w 2 ), . . . , πm ( w m )). Dual to this surjection, the
subspace in W ∗ generated by xi , j for 1 i m, 2 j q i is isomorphic to T ∗ . Therefore we get the
inclusion
F [ T ] = F [xi , j | 1 i m, 2 j qi ] ⊆ F [ W ] = F [xi , j | 1 i m, 1 j qi ].
We prove a result along the same lines of Lemma 2. Let k be a positive integer and for 1 j m, let
J j ,k denote the ideal in F [ W j ] generated by x j ,k , x j ,k+1 , . . . , x j ,q j . Set J j ,k = 0 if k > q j .
Lemma 5. Let i , j , k be integers satisfying 1 i , j m, i = j and 1 k q j − 1. Then there exist f 1 ∈ F [ W j ]

p −1 p −1
and f 2 ∈ F [ T i ] ⊗ F [ W j ] such that Tr(xi ,1 x j ,k ) = f 1 xi ,1 + f 2 . Moreover, f 1 ≡ −x j ,k+1 mod J j ,k+2 .
Proof. The proof essentially carries over from Lemma 2. For 0 l p − 1, we have
p −1
l
p −1 l l
σ xi ,1 x j ,k = xi ,1 + lxi ,2 + xi ,3 + · · · x j ,k + lx j ,k+1 + x j ,k+2 + · · · .
2 2
Note that no monomial in the above expansion is divisible by x2i ,1 . Therefore as a polynomial in xi ,1 ,
p −1
the transfer Tr(xi ,1 x j ,k ) is of degree at most one and moreover if a monomial m that appears in
p −1
Tr(xi ,1 x j ,k ) is divisible by xi ,1 , then m/xi ,1 is in F [ W j ]. Finally, for non-negative integers a and b
p −1 b
with a + b = p − 1 the coefficient of xi ,1 xaj,k xbj,k+1 in σ l (x1 xip−1 ) is b
l . Therefore the coeffi-
p −1 p −1 b
cient of xi ,1 xaj,k xbj,k+1 in Tr(xi ,1 x j ,k ) is 0l p −1 b
l . Hence the final statement follows as in
Lemma 2. 2
M. Sezer / Journal of Algebra 322 (2009) 4099–4104 4103
p −1
For 1 i , j m with i = j and 1 k q j − 1 define H ki, j = Tr(xi ,1 x j ,k ). Set

H = H ki, j
1 i , j m, i = j , 1 k q j − 1 .
We show that the union polynomials in H and the separating sets for T , W i for 1 i m gives a
separating set for W .
Theorem 6. Assume the notation and the convention of the previous paragraphs. For 1 1 m, let S i ⊆
F [ W i ]G denote a separating set for W i and S ⊆ F [ T ]G denote a separating set for T . Then the union of the
polynomials in S, S 1 , S 2 , . . . , S m , H is a separating set for W .
Proof. Let v 1 = (c 1 , . . . , cm ) and v 2 = (d1 , . . . , dm ) be two vectors in W in different G-orbits, where

c i , di ∈ W i for 1 i m. Say c i = (c i ,1 , c i ,2 , . . . , c i ,qi )t and di = (di ,1 , di ,2 , . . . , di ,qi )t , where c i , j , di , j ∈ F
for 1 i m and 1 j q i . If π ( v 1 ) and π ( v 2 ) in T were in different G-orbits, then there exists a
polynomial in S that separates π ( v 1 ) and π ( v 2 ) because S is a separating set for T . This polynomial
separates v 1 and v 2 as well. Therefore we may assume that π ( v 1 ) and π ( v 2 ) are in the same G-orbit.
Hence by replacing v 2 with a suitable vector in its orbit we may assume that π ( v 1 ) = π ( v 2 ), that is
c i , j = di , j for 1 i m and 2 j q i . Also if c i and di are in different G-orbits for some 1 i m,
then there exists a polynomial in S i that separates c i and di because S i is a separating set for W i .
Then v 1 and v 2 is separated by this polynomial as well. Therefore we may assume that c i and di are
in the same G-orbit for 1 i m.
First assume that there exists 1 r m and 3 k qr such that cr ,k = dr ,k = 0. By replacing k
with a larger integer if necessary, we may assume that cr ,k = dr ,k = 0 for k < k qr . Since cr , dr ∈ W r
are in the same G-orbit, there exists an integer 0 l p − 1 such that σ l (cr ) = dr . Since cr ,k = 0 for
k > k, the (k − 1)-st coordinate of σ l (cr ) is equal to cr ,k−1 + lcr ,k . Therefore the equality of the vectors
σ l (cr ) and dr gives cr ,k−1 + lcr ,k = dr ,k−1 . But since we have cr ,k−1 = dr ,k−1 , it follows that l = 0, that is
cr = dr . On the other hand since v 1 = v 2 , there exists 1 b m, b = r such that cb = db . Equivalently
p −1
cb,1 = db,1 . We show that H bk− 1
,r = Tr(xb,1 xr ,k−1 ) separates v 1 and v 2 . By the previous lemma we can
p −1 p −1
write Tr(xb,1 xr ,k−1 ) = f 1 xb,1 + f 2 with f 2 ∈ F [ T b ] ⊗ F [ W r ], f 1 ∈ F [ W r ] with f 1 ≡ −xr ,k mod J r ,k+1 .
Since cb,k = db,k for k > 1, cr = dr and f 2 ∈ F [ T b ] ⊗ F [ W r ], it follows that f 2 ( v 1 ) = f 2 ( v 2 ). We also
p −1 p −1
have f 1 ( v 1 ) = −cr ,k because cr ,k = 0 for k > k. Similarly f 1 ( v 2 ) = −dr ,k . Since cr ,k = dr ,k = 0, it
follows that f 1 ( v 1 ) = f 1 ( v 2 ) = 0. Hence H bk− 1
,r separates v 1 and v 2 because cb,1 = db,1 .
Next we consider the case c i , j = di , j = 0 for all 1 i m and 3 j q i . We look into two
subcases. First assume that there exists 1 r m such that cr ,2 = dr ,2 = 0. Since cr and dr are in
the same G-orbit there exists an integer 0 l p − 1 such that σ l (cr ) = dr . Moreover, since c i , j = 0
for 1 i m and 3 j q i , all coordinates of σ l (c i ) and c i are the same except the first one for
1 i m. That is π ( v 1 ) = π (σ l ( v 1 )). Therefore by replacing v 1 with σ l ( v 1 ), we may assume that
cr ,1 = dr ,1 as well. On the other hand since v 1 = v 2 , there exists 1 b m, b = r such that cb = db .
Now we have reduced to the situation considered in the previous paragraph: There exists 1 r m
such that cr = dr and 1 b m such that cb,1 = db,1 . Since cr ,k = dr ,k = 0 for 2 < k qr , the
p −1
argument in the previous paragraph shows that H b1,r = Tr(xb,1 xr ,1 ) separates v 1 and v 2 . Finally if
c i , j = di , j = 0 for all 1 i m and 2 j q i , then each c i and di is a fixed point in W i . Hence if
c i and di are in the same G-orbit, then c i = di . Since this is true for all 1 i m, it follows that
v1 = v2. 2
Note that the dimensions of indecomposable summands in T are one less than the dimensions
of the indecomposable summands in W . Meanwhile, the polynomials in S i may be chosen to be of
degree one and p for all 1 i m by Remark 4 and the polynomials in H are of degree p and
involve variables from two summands. Therefore by induction on the maximum dimension of an
indecomposable summand in a representation one easily gets the following.
Corollary 7. Let W be a Z / p representation over characteristic p. Then there exists a separating set for W
consisting of polynomials each of which has degree one or p and involves variables from at most two summands
in W ∗ .
References
[1] H.E.A. Campbell, I.P. Hughes, R.J. Shank, D.L. Wehlau, Bases for rings of coinvariants, Transform. Groups 1 (4) (1996) 307–
336.
[2] Harm Derksen, Gregor Kemper, Computational invariant theory, in: Invariant Theory and Algebraic Transformation
Groups, I, in: Encyclopaedia Math. Sci., vol. 130, Springer-Verlag, Berlin, 2002.
[3] Harm Derksen, Gregor Kemper, Computing invariants of algebraic groups in arbitrary characteristic, Adv. Math. 217 (5)
(2008) 2089–2129.
[4] Leonard Eugene Dickson, On Invariants and the Theory of Numbers, Reprinted by Dover Publications Inc., New York, 1966.
[5] M. Domokos, Typical separating invariants, Transform. Groups 12 (1) (2007) 49–63.
[6] Jan Draisma, Gregor Kemper, David Wehlau, Polarization of separating invariants, Canad. J. Math. 60 (3) (2008) 556–571.
[7] E. Dufresne, J. Elmer, M. Kohls, The Cohen–Macaulay property of separating invariants of finite groups, preprint,
arXiv:0904.1069, 2009.
[8] Emilie Dufresne, Separating invariants, PhD thesis, Queen’s University, Kingston, Ontario, 2008.
[9] Emilie Dufresne, Separating invariants and finite reflection groups, Adv. Math. 221 (6) (2009) 1979–1989.
[10] P. Fleischmann, M. Sezer, R.J. Shank, C.F. Woodcock, The Noether numbers for cyclic groups of prime order, Adv.
Math. 207 (1) (2006) 149–155.
[11] Ian Hughes, Gregor Kemper, Symmetric powers of modular representations, Hilbert series and degree bounds, Comm.
Algebra 28 (4) (2000) 2059–2088.
[12] G. Kemper, Separating invariants, J. Symbolic Comput. 44 (2009) 1212–1222.
[13] M.D. Neusel, M. Sezer, Separating invariants for modular p-groups and groups acting diagonally, preprint, available at
http://www.fen.bilkent.edu.tr/~sezer/, 2008.
[14] Mufit Sezer, Lexsegment and Gotzmann ideals associated with the diagonal action of Z / p, preprint, available at http://
www.fen.bilkent.edu.tr/~sezer/, 2008.
[15] R. James Shank, S.A.G.B.I. bases for rings of formal modular seminvariants [semi-invariants]., Comment. Math. Helv. 73 (4)
(1998) 548–565.
Journal of Algebra
Totally chiral maps and hypermaps of small genus

Antonio Breda D’Azevedo a,∗,1 , Gareth A. Jones b
a
Departamento de Matematica, Universidade de Aveiro, 3810-193 Aveiro, Portugal
b
School of Mathematics, University of Southampton, Southampton SO17 1BJ, United Kingdom
Article history: An orientably regular hypermap is totally chiral if it and its mirror
Received 31 July 2008 image have no non-trivial common quotients. We classify the
Available online 17 March 2009 totally chiral hypermaps of genus up to 1001, and prove that the
Communicated by Eamonn O’Brien
least genus of any totally chiral hypermap is 211, attained by
Keywords:
twelve orientably regular hypermaps with monodromy group A 7
Map and type (3, 4, 4) (up to triality). The least genus of any totally
Hypermap chiral map is 631, attained by a chiral pair of orientably regular
Totally chiral map maps of type {11, 4}, together with their duals; their monodromy
Totally chiral hypermap group is the Mathieu group M 11 . This is also the least genus of
Chirality group any totally chiral hypermap with non-simple monodromy group, in
2-generated groups this case the perfect triple covering 3 . A 7 of A 7 . The least genus of
any totally chiral map with non-simple monodromy group is 1457,
attained by 48 maps with monodromy group isomorphic to the
central extension 2 . Sz(8).
1. Introduction
Maps on compact surfaces have for many years been a common topic of interest among those
working on finite and discrete groups, on Riemann surfaces, and in areas as diverse as combina-
torics and hyperbolic geometry. Their generalisations to hypermaps (surface embeddings of hyper-
graphs), introduced by Cori [Cor] in 1975, have recently attracted considerable attention in view of
Grothendieck’s theory of dessins d’enfants [Gro], in which Belyı̆’s theorem [Bel] characterises the Rie-
mann surfaces defined (as projective algebraic curves) over the field of algebraic numbers as those
obtained in a canonical way from hypermaps. This means that these combinatorial objects play a ma-
*
E-mail addresses: breda@mat.ua.pt (A. Breda D’Azevedo), g.a.jones@maths.soton.ac.uk (G.A. Jones).
1
Supported in part by UI&D Matemática e aplicações of University of Aveiro, through Program POCTI of FCT co-financed by
the European Community fund FEDER.
doi:10.1016/j.jalgebra.2009.02.017
3972 A. Breda D’Azevedo, G.A. Jones / Journal of Algebra 322 (2009) 3971–3996
jor role in relating the Galois theory of algebraic number fields to the Teichmüller theory of Riemann
surfaces (see [JSin], for instance).
Chirality is the phenomenon which occurs when an oriented object, such as a map, hypermap
or Riemann surface, is not isomorphic to its mirror image under any orientation-preserving trans-
formations. This concept is of great importance in science: for instance, a molecule and its mirror
image may have very different chemical properties. It is also important in mathematics: for instance,
the well-known Poincaré dodecahedral space and Weber–Seifert space both occur as chiral pairs of
oriented 3-manifolds [Mon, §§3.7, 3.8, 3.13], [WS].
Within the category of oriented hypermaps, the most symmetric objects are the orientably regular
(or rotary) hypermaps, those for which the orientation-preserving automorphism group acts transi-
tively on darts. Such a hypermap H is said to be regular (or reflexible) if it has an orientation-reversing
automorphism, so that H is isomorphic to its mirror image H; otherwise it is chiral. Early classifi-
cations of maps and hypermaps showed chirality to be a rather unusual phenomenon for low genus.
For instance, there are no chiral hypermaps on the sphere, and on the torus they form three easily
described infinite families, whose types are permutations of (2, 3, 6), (2, 4, 4) and (3, 3, 3). For each
genus g > 1 there are (up to isomorphism) only finitely many orientably regular hypermaps, and
when g is small relatively few of these are chiral. For instance, there are no chiral maps of genus g
for 2 g 6, the first examples as g increases being the chiral pair of Edmonds maps of genus 7
and type {7, 7}, with monodromy group (and hence automorphism group) isomorphic to the affine
group AGL1 (8), together with a chiral pair of maps of type {6, 9} and their duals, with a monodromy
group of order 54. Among the proper hypermaps (those of type ( p , q, r ) with p , q, r > 2) there are
no chiral pairs of genus 2, and (up to triality) just one of genus 3, having type (3, 3, 7) and a mon-
odromy group of order 21; similarly, there is just one pair each for g = 4, 6 and 7 and there are
none for g = 5. A computer classification by Conder [Con07] up to genus 101 has shown that chiral-
ity becomes more common as g increases, though regularity still predominates in this range. He has
shown that there are, up to isomorphism and duality, 3378 regular orientable maps of genus g for
2 g 101, but only 594 chiral pairs. Similarly for proper hypermaps in this range there are (up to
isomorphism and triality) 14647 regular orientable hypermaps and 2191 chiral pairs. At present it is
still not clear what the asymptotic behaviour is as g → ∞.
In [BJNŠ], Nedela, Škoviera and the present authors introduced numerical and algebraic measures
of chirality, called the chirality index and the chirality group, which indicate how much a hypermap
differs from its mirror image. Here we consider the most extreme form of chirality, called total chi-
rality, in which these two hypermaps are as unlike as possible, in the sense of having no non-trivial
common quotients; this is a combinatorial analogue of two integers being mutually coprime. As noted
in [BJNŠ], although there are infinitely many totally chiral hypermaps, those of small genus seem to
be very rare. Indeed, the least genus of such a hypermap described there is 481, having type (7, 3, 7)
with monodromy group isomorphic to the alternating group A 7 (see Theorem 13 of [BJNŠ] and the
remark following Theorem 14). Our aim here is to classify the totally chiral hypermaps of genus up
to 1001 (see Theorem B and the tables in Appendix A), and in particular to describe the totally chi-
ral hypermaps and maps of least genus. These genera turn out to be surprisingly large, namely 211
and 631 respectively (see Theorems A and C, and Examples 1, 2 and 5). The monodromy groups of
the hypermaps and maps achieving these lower bounds are respectively isomorphic to A 7 and to the
Mathieu group M 11 . These groups are both simple, and in Theorem D we show that the least genus
of any totally chiral hypermap with a composite monodromy group is also 631; in this case the mon-
odromy group is the perfect triple covering group 3 . A 7 of A 7 , and the hypermaps are unbranched
triple coverings of those of genus 211 mentioned above.
Our method is based on the observation that the monodromy group G of a totally chiral hypermap
must be a perfect covering group of a non-abelian finite simple group S which is not of type L 2 (q).
Such groups S are comparatively rare: for instance, there are only ten of order less than 105 . The
restriction g 1001, together with the Hurwitz bound, implies that |G | 84000, so that determining
the relevant covering groups G is not difficult, using the Schur multiplier and the representation
theory of S to deal with central and non-central extensions. Finding the chiral hypermaps associated
with G is equivalent to classifying, up to automorphisms, the generating pairs for G which are not
simultaneously inverted by any automorphism; in specific cases this can be done by character theory,
A. Breda D’Azevedo, G.A. Jones / Journal of Algebra 322 (2009) 3971–3996 3973
using the character tables and other information in the ATLAS of Finite Groups [CCNPW], but in order to
deal with the numerous possibilities which need to be eliminated we have also made extensive use of
GAP [GAP]. If G is simple then total chirality is equivalent to chirality, and in other cases the groups G
are sufficiently close to being simple that it is not difficult to determine which of the associated chiral
hypermaps are totally chiral.
At the end of the paper we describe some constructions of infinite families of totally chiral hy-
permaps. By the nature of some constructions (and one of them given in [BJNŠ]) it is not properly a
surprise that there are infinitely many totally chiral hypermaps. However, what is unexpected is that
every totally chiral hypermap (and map) has infinitely many totally chiral hypermaps as unbranched
coverings (Corollary 8).
2. Hypermaps
Throughout this paper, all hypermaps are assumed to be finite and oriented, and all automor-
phisms are assumed to be orientation-preserving. A hypermap H of type (l, m, n) can be regarded as
a transitive permutation representation of the triangle group

= (l, m, n) = X , Y , Z X l = Y m = Z n = X Y Z = 1
on a set of darts: the hypervertices, hyperedges and hyperfaces correspond to the cycles of X , Y
and Z , with incidence given by non-empty intersection; equivalently H corresponds to a conjugacy
class of subgroups K of finite index in , the stabilisers of darts, called hypermap subgroups. The
monodromy group Mon H of H is the group of permutations of the darts induced by , and the
automorphism group Aut H is its centraliser in the symmetric group.
We say that H is orientably regular (or rotary) if Aut H acts transitively on the darts, or equivalently
K is a torsion-free normal subgroup of , in which case Mon H ∼ = Aut H ∼= / K . Thus the orientably
regular hypermaps with a given monodromy group G correspond to the orbits of Aut G on generating
triples for G of type (l, m, n) (that is, triples (x, y , z) of elements of order l, m, n that generate G and
satisfy xyz = 1), or equivalently on pairs of generators x and y of G such that x, y and xy have
orders l, m and n. Such a hypermap is denoted by (G ; x, y ), and it is isomorphic to (G ; x , y ) if
and only if there is an isomorphism G → G taking x to x and y to y . The darts of (G ; x, y ) can
then be identified with the elements of G, permuted regularly, and the hypervertices, hyperedges and
hyperfaces with the cosets of the cyclic subgroups generated by x, y and z; incidence corresponds
to non-empty intersection of cosets, and local orientation is determined by the successive powers of
these generators. Maps are hypermaps of type (l, 2, n).
For the rest of this paper we will let H denote an orientably regular hypermap of type (l, m, n)
with monodromy group G. If N G (l, m, n) is the number of the above generating triples in G, then
since Aut G acts freely on them the number of such hypermaps is N G (l, m, n)/|Aut G |. They all have
genus

|G | 1 1 1
g =1+ 1− − − . (2.1)
2 l m n
Quotients H of H correspond to subgroups K of containing K , and the degree d of the covering

H → H is equal to the index | K : K | of K in K . If H and H have Euler characteristics χ and χ ,
and genera g and g , then
χ = dχ − b dχ ,
where b is the total order of branching, so
b
g = 1 + d( g − 1) + 1 + d( g − 1), (2.2)
2
with equality in either case if and only if the covering is unbranched.

Machì’s operations [Mac] rename the hypervertices, hyperedges and hyperfaces of a hypermap,
thus permuting l, m and n, without changing the surface or the monodromy group; the hypermaps
obtained from H in this way are its associates. For instance, replacing (x, y , z) with (x, z, y z ) gives a
hypermap H(12) of type (l, n, m) in which the hyperedges and hyperfaces of H have been transposed.
The relationship between H and its associates is sometimes known as triality.
The mirror image H of H corresponds to the image K of K under the automorphism of in-
verting X and Y . We say that H is regular if H ∼ = H, that is, K = K , or equivalently G has an
automorphism inverting x and y; otherwise H is chiral. The reflection operation H → H commutes
with Machì’s operations, and together they generate a group Γ of twelve hypermap operations, iso-
morphic to S 3 × C 2 . Under the action of Γ , each orientably regular hypermap lies in an orbit of length
dividing 12, all sharing the same genus and monodromy group. In classifying hypermaps with specific
properties, we will generally pick one representative from each orbit of Γ , usually choosing the type
so that l m n. In particular, if l, m and n are all distinct, then regular and chiral hypermaps of type
(l, m, n) lie in orbits of length 6 and 12 respectively.
The chirality index | K K : K | of H, introduced in [BJNŠ], is a measure of the chirality of H, equal
to 1 if and only if H is regular. The most extreme form of chirality occurs when the co-chirality index
| : K K | of H is 1, or equivalently K K = , so that H and H have no non-trivial common quotients;
in this case we say that H is totally chiral. Note that K K is a normal subgroup of containing K , so
if G is simple then H must be either regular or totally chiral (see [BJNŠ, Lemma 11]).
We finish this section with two examples, which we will later show to be the totally chiral hyper-
maps of least genus.
Example 1. The permutations x = (1, 2, 3)(4, 5, 6) and y = (1, 2)(3, 5, 6, 7) clearly generate a transitive
subgroup of the alternating group A 7 . Since x2 y = (1, 5, 4, 7, 3), and no transitive proper subgroup
of A 7 has order divisible by 5, it follows that x and y generate A 7 . We will denote the resulting ori-
entably regular hypermap ( A 7 ; x, y ) by A7 [1] , since it is our first example with monodromy group A 7 .
Since z := (xy )−1 = (2, 3, 7, 5)(4, 6) it has type (3, 4, 4); since | A 7 | = 7!/2 = 2520, Eq. (2.1) implies
that it has genus 211. Now A 7 has automorphism group S 7 , acting on A 7 by conjugation, and it
is easily seen that no element of S 7 conjugates both x and y to their inverses, so A7 [1] is chiral.
Since A 7 is simple, it follows that A7 [1] is totally chiral, as are all members of its Γ -orbit. In fact
A7 [1] ∼
(12)
= A7 [1] since g = (1, 6, 3, 5, 2, 4) conjugates the pair (x, y ) to (x, z), so this orbit contains
six totally chiral hypermaps of genus 211, forming a chiral pair of each of the types (3, 4, 4), (4, 3, 4)
and (4, 4, 3).
Example 2. Similar arguments show that the permutations x = (1, 2, 3)(4, 5, 6) and y = (1, 5)(2, 4, 6, 7)
determine a totally chiral hypermap A7 [2] of type (3, 4, 4) and genus 211 with monodromy group A 7 .
Again we have A7 [2] ∼
(12)
= A7 [2] , so this Γ -orbit contains six hypermaps of genus 211. These are
distinct from those in Example 1 since there is no element of S 7 which conjugates the generators x
and y to those in Example 1 or their inverses.
In Lemma 5 and Corollary 6 we will give a more general method for proving chirality and non-
isomorphism of orientably regular hypermaps.
3. Counting triples and hypermaps
We will show that the hypermaps A7 [1] , A7 [2] , A7 [1] and A7 [2] are the only orientably regular
hypermaps of type (3, 4, 4) with monodromy group A 7 . For this, and in many other places in this
paper, we need to count generating triples of a given type in a finite group. If the group is not too
large, one can easily do this by using a program such as GAP. It is also possible to do this by hand,
and thus to check the computer calculations, by using character tables. Specifically, if X , Y and Z
are conjugacy classes in a finite group G, then the number of solutions of xyz = 1 in G, with x ∈ X ,
y ∈ Y and z ∈ Z , is given by the formula
|X | . |Y | . |Z | χ (x)χ ( y )χ (z)
, (3.1)
|G | χ
χ (1)
where χ ranges over the irreducible complex characters of G (see [Ser, §7.2] for this and other similar
results). Such a triple generates G if and only if it is not contained in any maximal subgroup of G,
so knowledge of the maximal subgroups allows one to count the generating triples obtained from the
classes X , Y and Z . By summing over all triples of classes of elements of orders l, m and n we then
obtain the total number N G (l, m, n) of generating triples of type (l, m, n). As an example, we have the
following result. (In the proof, and throughout this paper, we use ATLAS notation [CCNPW] for finite
groups, together with their conjugacy classes and their characters; in particular, L n (q) denotes the
projective special linear group PSLn (q) = SLn (q)/{± I } over the field of q elements.)
Proposition 1. The group G = A 7 has 20160 generating triples of type (3, 4, 4).
Proof. The elements of order 3 in A 7 form two conjugacy classes, namely the 70 elements of class 3A
(in ATLAS notation) with cycle-structure 314 in the natural representation of A 7 , and the 280 elements
of class 3B with cycle-structure 32 1. There is a single class 4A of 630 elements of order 4, with cycle-
structure 421. It is easy to see that if x and y are in classes 3A and 4A, and the subgroup they generate
is transitive, then xy cannot have order 4; thus there are no generating triples of type (3, 4, 4) with x
in 3A, so we may assume that x is in 3B. Applying formula (3.1) with x in 3B and y and z in 4A, we
find from the character table of A 7 in [CCNPW] that the resulting number of triples is

280 × 6302 (−1) × 12
1+ = 42840 = 17|G |,
2520 35
with only the characters χ1 and χ9 , of degrees 1 and 35, making non-zero contributions to the sum.
It is easy to see that if x, y
is intransitive it must have two orbits, of lengths 1 and 6 or 3 and 4.
In the first case, x and y lie in exactly one of the seven conjugates of A 6 in A 7 ; now the character
table of A 6 shows that it has 1080 triples of type (3, 4, 4) with x having cycle-structure 32, so we
7
obtain 7560 = 3|G | such triples in A 7 . In the second case x and y lie in exactly one of the 3 = 35
conjugates of the even subgroup of S 3 × S 4 in A 7 ; by inspection, this group contains 6 × 24 = 144
such triples, so we obtain 5040 = 2|G | triples of this type. The only proper subgroups of G of order
divisible by 3, 4 and 7 are the two conjugacy classes of 15 subgroups isomorphic to L 2 (7), so if x
and y generate a transitive proper subgroup it must be one of these. The character table of L 2 (7)
shows that it contains 336 such triples, so the number of them in A 7 is 30 × 336 = 10080 = 4|G |.
This leaves (17 − 3 − 2 − 4)|G | = 8|G | = 20160 triples generating G. 2
Corollary 2. The hypermaps A7 [1] , A7 [1] , A7 [2] and A7 [2] described in Examples 1 and 2 are the only ori-
entably regular hypermaps of type (3, 4, 4) with monodromy group A 7 .
Proof. Since A 7 has automorphism group S 7 of order 7!, it follows from Proposition 1 and the re-
marks in Section 2 that there are 20160/7! = 4 such hypermaps, so they must be the four described
in Examples 1 and 2. 2
Clearly the corresponding result also applies to the hypermaps of type (4, 3, 4) and (4, 4, 3) de-
scribed in Examples 1 and 2.
4. Numerical bounds
In order to prove our main theorems, we need an argument which restricts the triples that we
have to consider. There are no chiral hypermaps on the sphere, and those on the torus are not totally
chiral since their monodromy groups are not perfect (see Lemma 4(b)). Thus a totally chiral hypermap
must have genus g > 1, so its type (l, m, n) must be a ‘hyperbolic’ triple, that is,
1 1 1
+ + < 1.
l m n
We can then write (2.1) as
|G |
g =1+ , (4.1)
λ
where
−1
1 1 1
λ := 2 1 − − −
l m n
depends only on (l, m, n). If g 211 (the genus in Examples 1 and 2) then (4.1) gives
|G |
λ , (4.2)
210
and this, together with the following lemma, can be used to show that various triples cannot corre-
spond to totally chiral hypermaps of smaller genus.
Lemma 3.
(a) The only hyperbolic types (l, m, n) with λ 936/35 = 26.742 . . . are the following (up to permutations
of l, m and n):
(2, 3, n) for n = 7, 8, 9 and 10, and (2, 4, 5).
(b) The only hyperbolic types (l, m, n) with 936/35 = 26.742 . . . > λ 20.16 are the following (up to per-
mutations of l, m and n):
(2, 3, n) for n = 11, 12, 13 and 14, and (2, 4, 6) and (3, 3, 4).
(c) The only hyperbolic types (l, m, n) with λ 12 and l, m, n 7 are the following (up to permutations
of l, m and n):
(2, 3, 7), (2, 4, 5), (2, 4, 6), (2, 4, 7), (2, 5, 5), (2, 5, 6), (2, 5, 7), (2, 6, 6), (3, 3, 4),
(3, 3, 5), (3, 3, 6) and (3, 4, 4).
The proof of this is a straightforward but tedious exercise in elementary arithmetic, and is there-
fore omitted. The rather strange bounds on λ in parts (a) and (b) are required in the proofs of
Theorems A and B. The relevant types are listed in Table 1: on the left are those in part (a) and
(after a gap) part (b), and on the right those in part (c), in each case in decreasing order of λ. The
first column gives the type, the second column gives the number of types formed by applying triality,
and the third column gives the corresponding value of λ.
A more extensive table, including the values of λ for all the arithmetic triangle groups, can be
found in [BJ]; this omits (2, 3, 13) and (2, 5, 7) since the associated triangle groups are not arithmetic
groups.
Table 1
Hypermap types and values of λ appearing in Lemma 3.
Type No. λ (a) Type No. λ (c)

(2, 3, 7) 6 84 (2, 3, 7) 6 84
(2, 3, 8) 6 48 (2, 4, 5) 6 40
(2, 4, 5) 6 40 (2, 4, 6) 6 24
(2, 3, 9) 6 36 (3, 3, 4) 3 24
(2, 3, 10) 6 30 (2, 5, 5) 3 20
(2, 4, 7) 6 56/3 = 18.6 . . .
(b) (2, 5, 6) 6 15
(2, 3, 11) 6 132/5 = 26.4 . . . (3, 3, 5) 3 15
(2, 3, 12) 6 24 (2, 5, 7) 6 140/11 = 12.7 . . .
(2, 4, 6) 6 24 (2, 6, 6) 3 12
(3, 3, 4) 3 24 (3, 3, 6) 3 12
(2, 3, 13) 6 156/7 = 22.2 . . . (3, 4, 4) 3 12
(2, 3, 14) 6 21
5. The totally chiral hypermaps of least genus
Before we prove our first main result we establish the following consequences of totally chirality.
Lemma 4. Let H be a totally chiral hypermap.
(a) Any non-trivial orientably regular quotient of H is also totally chiral.

(b) The monodromy group G of H is a perfect group.
(c) G does not have any simple group L 2 (q) as a quotient group.
Proof. (a) If H is totally chiral, then so is any orientably regular quotient of H, since its hypermap
subgroup of contains that corresponding to H.
(b) An abelian group has an automorphism inverting all its elements, so it cannot be the mon-
odromy group of a totally chiral hypermap. It therefore follows from (a) that G has no non-trivial
abelian quotients, so G is perfect.
(c) As shown in [BJNŠ, Proposition 8, Corollary 10], any pair of generators of L 2 (q) is inverted by
an automorphism of that group, so (c) also follows immediately from (a). 2
The converse of Lemma 4(a) is false, since an orientably regular covering of a totally chiral hy-
permap need not be totally chiral (see Example 8 in Section 11). We will prove a partial converse in
Proposition 6.
It follows from Lemma 4 that G is a covering of a non-abelian simple group S L 2 (q). For small
orders, there are very few such simple groups, for instance only four of order less than 20160. This
forms the basis of the proof of our first main result:
Theorem A. The least genus of any totally chiral hypermap is 211, attained only by the twelve hypermaps
formed from A7 [1] and A7 [2] by taking associates and mirror images. They all have monodromy group A 7 and
type (3, 4, 4), (4, 3, 4) or (4, 4, 3).
Proof. Let G be the monodromy group of a totally chiral hypermap of least genus g. By Examples 1
and 2 we have g 211, so |G | 84( g − 1) 17640 by the Hurwitz bound. By Lemma 4(b) G is
perfect, and moreover G must be simple, for otherwise Lemma 4(a) implies that a proper epimorphic
image of G would be the monodromy group of a totally chiral hypermap H , and by (2.2) this would
have genus g < g. We therefore consider the non-abelian simple groups of order at most 17640 as
possible monodromy groups of totally chiral hypermaps. According to the ATLAS the simple groups
within this range are:
(a) L 2 (q) for various odd q 31 and even q 16,

(b) A 7 of order 2520 = 23 . 32 . 5 . 7,
(c) L 3 (3) of order 5616 = 24 . 33 . 13,
(d) U 3 (3) of order 6048 = 25 . 33 . 7,
(e) M 11 of order 7920 = 24 . 32 . 5 . 11.
We can immediately eliminate case (a) by Lemma 4(c). We will eliminate cases (c), (d) and (e) by
using the inequality (4.2) and Lemma 3(a).
First let G = L 3 (3), so |G |/210 = 5616/210 = 936/35 = 26.742 . . . . By Lemma 3(a) the only types
satisfying λ |G |/210 are (2, 3, n) for n = 7, 8, 9 and 10, and (2, 4, 5). However (2, 3, 7), (2, 3, 9),
(2, 3, 10) and (2, 4, 5) are impossible since G has no elements of order 5, 7 or 9, and a calculation
similar to the proof of Proposition 1, using GAP or the character table of L 3 (3), shows that G has no
generating triples of type (2, 3, 8).
If G is the unitary group U 3 (3) then |G |/210 = 28.2, so by Lemma 3(a) and Table 1 the same five
types must be considered. In this case (2, 3, 10) and (2, 4, 5) are impossible since G has no elements
of order 5, and we find that there are no generating triples of the other three types.
If G is the Mathieu group M 11 then |G |/210 = 37.714 . . . , so the only types to consider are (2, 3, 7),
(2, 3, 8) and (2, 4, 5). The first is impossible since G has no elements of order 7, and there are no
generating triples of the other two types.
Thus G = A 7 , so |G |/210 = 12. Since A 7 has no elements of order greater than 7 it follows that the
only types we need to consider are the twelve listed in Lemma 3(c). The methods used earlier show
that there are no generating triples of types (2, 3, 7), (2, 4, 5), (2, 4, 6), (2, 5, 5), (2, 5, 6), (2, 6, 6) or
(3, 3, 4) in A 7 . There are 10080 each of types (2, 4, 7) and (2, 5, 7), giving two hypermaps in each
case, of genus g = 136 and 199; by considering generating pairs one sees that these are both regular.
There are 5040 each of types (3, 3, 5) and (3, 3, 6), giving one hypermap in each case, of genus 169
and 211; by their uniqueness these are necessarily regular. Finally there are 20160 generating triples
of type (3, 4, 4), giving the four hypermaps of genus 211 described in Examples 1 and 2. Applying
triality gives the other eight hypermaps of genus 211 described there. 2
Remark. One could alternatively regard the smallest totally chiral hypermaps as those having the
smallest monodromy group. It is clear from the proof of Theorem A that A 7 is the smallest mon-
odromy group which can occur, so in this sense the smallest totally chiral hypermaps are those with
monodromy group A 7 . These include all those of genus 211 appearing in Theorem A, but also other
totally chiral hypermaps of higher genus such as those of genus 481 mentioned in Section 1; Table 2
lists all those of genus g 1001.
6. Testing hypermaps for isomorphism and reflexibility
The method of proof of Theorem A can be extended to classify all the totally chiral hypermaps of
comparatively small genus. We shall do this for genus g 1001, though in principle it is possible to
go further.
Before proceeding with the classification, we describe a simple method for determining iso-
morphism or regularity of hypermaps, which we will use repeatedly. Given an orientably regular
hypermap H with monodromy group G, any transitive permutation representation ρ : G → S k ,
with a point stabiliser H G, gives rise to a quotient hypermap H∗ = H/ H with k darts. This
hypermap is orientable, but not orientably regular unless H is normal in G; it has monodromy
group G ∗ = ρ (G ) ∼=
G / ker(ρ ). We call H∗ a faithful quotient of H if ρ is faithful, or equivalently
∗ ∼
g ∈G H = ker(ρ ) in G, so that G = G. In these circumstances, H is uniquely
g
H has trivial core
determined by H∗ as its minimal orientably regular covering, in the sense that any orientably regular
covering of H∗ is also a covering of H.
The objective is to find a faithful quotient H∗ of H which adequately reflects certain properties
of H, so that it can be used as a simpler substitute for H in specific calculations (see Lemma 5 and
Corollary 6, for instance). For convenience it is desirable (though not always possible) that H∗ should
be spherical, that is, have genus 0, and that the degree k = |G : H | should be small. For simple
groups G, the ATLAS gives useful information about maximal subgroups and hence transitive rep-
resentations of comparatively small degrees, all of them necessarily faithful.
If k is sufficiently small, it can be useful to represent H∗ visually by its Walsh map [Wal]. This con-
struction (which can be applied to any hypermap) produces a bipartite map on the same surface, with
black and white vertices, edges and faces corresponding to the hypervertices, hyperedges, darts and
hyperfaces of H∗ , and with the rotation of edges around each black or white vertex determined by
the cyclic order of darts in the corresponding hypervertex or hyperedge. Isomorphisms of hypermaps
correspond to colour-preserving isomorphisms of their Walsh maps, and it often easy to see from a
diagram whether or not these exist.
Lemma 5. Let H and H be orientably regular hypermaps with monodromy groups isomorphic to G, and let
H be a subgroup of G with trivial core.
(a) If H/ H ∼= H / H then H ∼
= H .
(b) If H = H and every subgroup of G equivalent under Aut G to H is conjugate in G to H , then H/ H ∼
∼
=
H / H .
Proof. (a) The hypermap subgroups K and K of corresponding to H and H are the cores in
of those corresponding to H/ H and H / H . If these quotients are isomorphic then their hypermap
subgroups are conjugate and hence have the same core, so K = K and hence H ∼ = H .

(b) If H and H are isomorphic then they correspond to the same (normal) hypermap subgroup K
of , with / K ∼ = G. Let L and L be the subgroups of , containing K , which correspond to H
under the isomorphisms of Mon H and Mon H with / K . Then L / K and L / K are equivalent under
automorphisms of / K , so by the hypothesis they are conjugate in / K and hence L and L are
conjugate in ; but these are the hypermap subgroups corresponding to H/ H and H / H , so these
quotients are isomorphic. 2
Corollary 6. Let H be an orientably regular hypermap with monodromy group G, and let H be a subgroup
of G with trivial core.
(a) If H/ H ∼= H/ H then H is regular.

(b) If H is regular and every subgroup of G equivalent under Aut G to H is conjugate to H , then H/ H ∼
= H/ H .
Proof. This follows immediately from Lemma 5, with H = H. 2
When testing hypermaps for isomorphism or regularity, Lemma 5 and Corollary 6 can be useful in
replacing large hypermaps with much smaller ones, as shown by the following example.
Example 3. The hypermap H = A7 [1] in Example 1 has monodromy group G = A 7 ; the subgroup
H = A 6 , a point stabiliser in the natural representation of A 7 , has trivial core, and its conjugates form
the unique conjugacy class of subgroups of index 7 in A 7 , so it satisfies the hypotheses of Lemma 5
and Corollary 6. The hypermap A7 [1] has | A 7 | = 2520 darts, whereas A7 [1] / A 6 has only 7; it is easy to
use the permutations x and y to draw the Walsh map for A7 [1] / A 6 ; Fig. 1(a) shows that this map is
not isomorphic to its mirror image, so A7 [1] is chiral by Corollary 6(b). Similarly Fig. 1(b) shows that
the hypermap A7 [2] in Example 2 is chiral. Since these two maps cannot be transformed into each
other by reflection and/or transposition of colours, it follows from Lemma 5(b) that the hypermaps in
Examples 1 and 2 form different orbits of Γ .
Parts (b) of Lemma 5 and Corollary 6 can fail if there is a second conjugacy class of subgroups of G
which are equivalent to H under outer automorphisms of G. The following example is instructive.
Example 4. The group G = L 3 (3) has 33696 generating triples of type (3, 3, 4). Since Aut G = PGL3 (3)
has order 11232, these correspond to three orientably regular hypermaps of genus 235 with mon-
odromy group L 3 (3). Two of these form a chiral pair H1 = L3 (3)[1] and H2 = H1 ∼ = H1 (01) , which are
Fig. 1. The quotient hypermaps A7 [1] / A 6 and A7 [2] / A 6 .
Fig. 2. The quotient hypermaps L3 (3)[1] / H and H3 / H .
totally chiral since L 3 (3) is simple, whereas the third is regular and self-dual. We can represent L 3 (3)
faithfully and transitively on the 13 points of the projective plane over the field F 3 , with point sta-
biliser H ∼
= AGL2 (3). The faithful quotient H1 / H , shown in Fig. 2(a), corresponds to the generating pair
x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11), y = (1, 13, 8)(2, 3, 5)(4, 12, 6)(7, 9, 11), and H1 / H , the mirror
image of Fig. 2(a), corresponds to their inverses. These two quotient hypermaps form a chiral pair,
but we cannot use Corollary 6(b) to deduce the same for their coverings H1 and H1 since the outer
automorphism of L 3 (3) (induced by the point-line duality of the projective plane) transposes the
point stabilisers with a second conjugacy class of subgroups, the stabilisers of lines. Instead we can
argue that H1 is chiral because w (x, y ) = (x−1 y )2 ( yxy )2 (xy −1 )3 x(xy )2 [x, y ] is the identity permuta-
tion while w (x−1 , y −1 ) = (1, 13, 2, 6, 9, 12, 11, 10, 8, 4, 3, 7, 5) has order 13, where [x, y ] = x−1 y −1 xy
is the commutator of x and y, so H1 and H1 correspond to distinct normal subgroups of . By
applying triality we obtain a Γ -orbit of six totally chiral hypermaps of genus 235 with monodromy
group L 3 (3), forming three chiral pairs of types (3, 3, 4), (4, 3, 3) and (3, 4, 3).
In the case of the third hypermap H3 , we have the same x as for H1 = L3 (3)[1] but now y =
(1, 2, 3)(5, 6, 7)(8, 9, 10)(11, 12, 13). The quotient hypermap H3 / H , shown in Fig. 2(b), is again chiral,
since a reflection transposes the vertex-colours, but in this case the hypermap itself is regular. As in
the case of L3 (3)[1] , Corollary 6(b) does not apply to H . In both cases, if we want a subgroup which
satisfies the condition in Corollary 6(b), we could take H to be the stabiliser of an incident point-line
pair: this has index 52 in G, so it gives quotient hypermaps with four times as many darts.
7. Classifying totally chiral hypermaps of low genus
Before using the techniques developed in earlier sections to prove our main theorem, we first need
to explain some central extensions which appear in its statement and proof; see [CCNPW, §4.1] or
[Hup, §V.23] for further details. A group G is a central extension of a group S by a group N if G has a
central subgroup N such that G / N ∼ = S; this extension is proper if N is contained in the commutator
subgroup G of G. The groups N appearing in proper central extensions of S are all quotients of a
group Mult( S ), the Schur multiplier of S. If S is perfect then there is a unique proper central extension
G = Ŝ of S by Mult( S ); this is the universal covering group of S, and any proper central extension of S
is a quotient of Ŝ by some subgroup of Mult( S ).
Each alternating group S = A n (n 5) has a unique proper central extension G by N = C 2 , denoted
by 2 . A n . One can construct this extension by embedding A n , as the group of even permutation
matrices, in the Lie group SO(n); this has simply connected covering group Spin(n), a central extension
of SO(n) by its fundamental group π1 SO(n) ∼ = C 2 [Che, §XII], and 2 . An is the inverse image of An in
Spin(n). Schur [Sch] showed that if n = 6, 7 then Mult( A n ) ∼
= C 2 , so that 2 . An is the universal covering
group Â n of A n . However, if n = 6 or 7 then Mult( A n ) ∼ = C 6 , so there are proper central extensions
d . A n of A n by C d for d = 2, 3 and 6. The triple covering group 3 . A 7 can be constructed from the
well-known Hoffman–Singleton graph [HS], [Cam, §3.6], which has the simple unitary group U 3 (5) as
a subgroup of index 2 in its automorphism group. The stabiliser in U 3 (5) of one of the 50 vertices
is isomorphic to A 7 , and its inverse image under the natural epimorphism SU 3 (5) → U 3 (5), taking
matrices to projective transformations, is the required group 3 . A 7 , with centre N ∼ = C 3 consisting of
the scalar matrices α I where α 3 = 1 in the Galois field F 25 .
Theorem B. The totally chiral hypermaps of genus g 1001 are those listed in Tables 2, 3, 4, 5, 6, 7 and 8
(see Appendix A), together with those formed from them by taking mirror images and associates. In each case
the monodromy group is a simple group A 7 , L 3 (3), U 3 (3), M 11 or Sz(8), or the proper central extension d . A 7
of A 7 by C d for d = 2 or 3.
Proof. In order to extend our method of proof of Theorem A up to genus 1001, we need to consider
monodromy groups G of order at most 84 × 1000 = 84000, and we need to modify inequality (4.2)
to give λ |G |/1000.
By Lemma 4(b) the monodromy group G of a totally chiral hypermap H is perfect, so if N is any
maximal normal subgroup of G then G / N is a non-abelian simple group S; by Lemma 4(a) S is the
monodromy group of a totally chiral hypermap H = H/ N, which has genus g 1001 by (2.2), and by
Lemma 4(c) we have S L 2 (q). In addition to the simple groups A 7 , L 3 (3), U 3 (3) and M 11 considered
in the proof of Theorem A, the only other non-abelian simple groups S L 2 (q) with | S | 84000 are:
A8 ∼
= L 4 (2) and L 3 (4), both of order 20160 = 26 . 32 . 5 . 7,
the symplectic group S 4 (3) ∼
= U 4 (2) of order 25920 = 26 . 34 . 5,
the Suzuki group Sz(8) of order 29120 = 26 . 5 . 7 . 13, and
the unitary group U 3 (4) of order 62400 = 26 . 3 . 52 . 13.
We will first show that of these five extra groups, only S = Sz(8) can occur. It is easiest to deal
with the largest groups first, since they correspond to the fewest triples.
If S = U 3 (4) then λ | S |/1000 = 62.4, and Lemma 3 and Table 1 show that the only type satis-
fying this condition is (2, 3, 7). Since U 3 (4) has no elements of order 7, it follows that neither this
group S nor any of its coverings G can appear as the monodromy group of a totally chiral hypermap
of genus g 1001.
Similarly, if S = S 4 (3), with elements of orders 1, 2, 3, 4, 5, 6, 9 and 12, it follows from Lemma 3
and Table 1 that the only possible types are (2, 3, 9) and (2, 4, 5). If S = L 3 (4), with elements of orders
1, 2, 3, 4, 5 and 7, only (2, 3, 7), (2, 4, 5) and (3, 3, 4) are possible. In the case of S = A 8 , which has
elements of orders 1, 2, 3, 4, 5, 6, 7 and 15, we obtain just these three types, together with (2, 4, 6).
However, GAP shows that none of these three groups S has a generating triple of any of these listed
types, so they can be eliminated.
This leaves the simple groups S = A 7 , L 3 (3), U 3 (3), M 11 and Sz(8), together with their covering
groups G, as possible monodromy groups of totally chiral hypermaps of genus at most 1001. Now
the quotient hypermap H is totally chiral, so by Theorem A it has genus g 211; it then follows
from (2.2) that H has genus g 210| N | + 1. Since g 1001 this implies that | N | 4. Now groups of
order at most 4 have solvable automorphism groups, whereas G is perfect, so N is in the centre of G,
and is isomorphic to a quotient of the Schur multiplier Mult( S ) of S. By [CCNPW] the groups S =
L 3 (3), U 3 (3) and M 11 all have trivial Schur multipliers, so if S is one of these then G = S. Since
Mult( A 7 ) ∼= C 6 and Mult(Sz(8)) ∼ = C 2 × C 2 , the only other possibilities are that S = A 7 and N ∼ = C1, C2
or C 3 , or that S = Sz(8) and N ∼ = C 1 , C 2 or C 2 × C 2 .
If G = Sz(8) then λ 29.12. Since Suzuki groups have no elements of order 3, Lemma 3 and
Table 1 show that the only possible type is (2, 4, 5). As shown in [JS], there are four totally chiral
hypermaps of this type with monodromy group Sz(8), forming two chiral pairs of genus 729 (see
Table 6). Any covering of Sz(8) with | N | 2 would give genus g 1457 by (2.2), and this is outside
our range.
If S = M 11 then λ 7.92. We have not listed all the triples satisfying this inequality, but since M 11
has elements of orders 1, 2, 3, 4, 5, 6, 8 and 11, the only ones we need consider are (2, 3, 11), (2, 4, 6),
(2, 4, 8), (2, 4, 11), (2, 5, 5), (2, 5, 6), (2, 5, 8), (2, 5, 11), (2, 6, 6), (2, 6, 8), (2, 6, 11), (2, 8, 8), (3, 3, 4),
(3, 3, 5), (3, 3, 6), (3, 3, 8), (3, 3, 11), (3, 4, 4), (3, 4, 5), (3, 4, 6) and (4, 4, 4). For the first three types,
together with (2, 5, 5), (2, 5, 6), (2, 6, 6), (3, 3, 4), (3, 3, 5), (3, 3, 6), (3, 3, 11), (3, 4, 4) and (4, 4, 4),
GAP shows that there are no generating triples. For the remaining types the corresponding hypermaps,
as found by GAP, are listed in Table 5.
Similarly the group G = U 3 (3) has elements of orders 1, 2, 3, 4, 6, 7, 8 and 12, with λ 6.048, so
we obtain only the types (2, 4, 6), (2, 4, 7), (2, 4, 8), (2, 4, 12), (2, 6, 6), (2, 6, 7), (2, 6, 8), (2, 6, 12),
(2, 7, 7), (2, 7, 8), (2, 7, 12), (2, 8, 8), (2, 8, 12), (3, 3, 4), (3, 3, 6), (3, 3, 7), (3, 3, 8), (3, 3, 12), (3, 4, 4),
(3, 4, 6), (3, 4, 7), (3, 4, 8) and (4, 4, 4). The first six of these, together with (2, 6, 12), (2, 8, 8), (3, 3, 4),
(3, 4, 4) and (4, 4, 4), correspond to no generating triples. The types (2, 6, 7), (2, 6, 8), (2, 7, 7) and
(2, 8, 12) are each realised by 12096 = |Aut G | generating triples, giving one regular hypermap in each
case, while the types (2, 7, 8) and (2, 7, 12) are realised by 24192 generating triples, in each case
giving two regular hypermaps. The hypermaps corresponding to the remaining types are listed in
Table 4.
The elements of G = L 3 (3) have order 1, 2, 3, 4, 6, 8 or 13. In this case λ 5.616, so the only pos-
sible types are (2, 3, 13), (2, 4, 6), (2, 4, 8), (2, 4, 13), (2, 6, 6), (2, 6, 8), (2, 6, 13), (2, 8, 8), (2, 8, 13),
(2, 13, 13), (3, 3, 4), (3, 3, 6), (3, 3, 8), (3, 3, 13), (3, 4, 4), (3, 4, 6), (3, 4, 8), (3, 4, 13), (3, 6, 6), (4, 4, 4)
and (4, 4, 6). The types (2, 4, 6), (2, 4, 8), (2, 6, 6) and (2, 6, 8) correspond to no generating triples,
while (2, 3, 13), (2, 4, 13) and (2, 6, 13) are each realised by 22464 = 2|Aut G | generating triples, giv-
ing rise to two regular hypermaps in each case. Type (2, 8, 8) is realised by 11232 triples, giving one
regular hypermap, type (2, 8, 13) is realised by 44928 triples, giving four regular hypermaps, while
type (2, 13, 13) is realised by 89856 triples, giving eight regular hypermaps. The hypermaps corre-
sponding to the remaining types are listed in Table 3.
If G = A 7 , with elements of orders 1, 2, 3, 4, 5, 6 and 7 and with λ 2.52, the possible types
are (2, 6, 6), (2, 6, 7), (2, 7, 7), (3, 3, 6), (3, 3, 7), (3, 4, 4), (3, 4, 5), (3, 4, 6), (3, 4, 7), (3, 5, 5), (3, 5, 6),
(3, 5, 7), (3, 6, 6), (3, 6, 7), (3, 7, 7), (4, 4, 4), (4, 4, 5), (4, 4, 6), (4, 4, 7), (4, 5, 5), (4, 5, 6), (4, 5, 7),
(4, 6, 6), (4, 6, 7), (4, 7, 7), (5, 5, 5), (5, 5, 6), (5, 5, 7), (5, 6, 6), (5, 6, 7), (5, 7, 7), (6, 6, 6), (6, 6, 7),
(6, 7, 7) and (7, 7, 7). There are no generating triples of type (2, 6, 6). For type (2, 6, 7) there are
10080 = 2|Aut G | generating triples, giving rise to two regular hypermaps, while for type (2, 7, 7)
there are 15120 triples giving three regular maps, for type (3, 3, 6) there are 5040 triples giving one
regular hypermap, and for type (3, 3, 7) there are 10080 triples giving two regular hypermaps. The
hypermaps corresponding to the remaining types are listed in Table 2.
There is a unique perfect double covering G = 2 . A 7 of A 7 , with elements of orders 1, 2, 3, 4,
5, 6, 7, 8, 10, 12 and 14. Using λ 5.04, we find totally chiral hypermaps of genus g 1001 for
the following types: there is one totally chiral pair and one regular hypermap of type (3, 5, 5) and
genus 673; one totally chiral pair and two regular hypermaps of type (3, 5, 7) and genus 817; two
totally chiral pairs and one regular hypermap of type (3, 5, 8) and genus 862; two totally chiral pairs
and three regular hypermaps of type (3, 7, 7) and genus 961; finally two totally chiral pairs and one
regular hypermap of type (3, 5, 14) and genus 997. (There are also a few non-totally chiral hypermaps,
with chirality index 2, but only for genera g > 1001: the least such genus is 1216, corresponding to
two chiral pairs of type (4, 7, 8).) The totally chiral hypermaps are listed in Table 7.
Similarly, there is a unique perfect triple covering G = 3 . A 7 , with elements of orders 1, 2, 3, 4,
5, 6, 7, 12, 15 and 21. In this case we find chiral hypermaps of genus g 1001 for the following
types: there is one non-totally chiral pair and one regular hypermap of type (3, 3, 5) and genus 505;
four non-totally chiral pairs and one regular hypermap of type (3, 3, 6) and genus 631; two totally
chiral pairs of type (3, 4, 4) and genus 631; two totally chiral pairs and one regular hypermap of type
(3, 4, 5) and genus 820; finally three totally chiral pairs, one non-totally chiral pair and one regular
hypermap of type (3, 4, 6) and genus 946. Here the non-totally chiral hypermaps all have chirality
index 3. The chiral hypermaps are listed in Table 8. 2
In dealing with 2 . A 7 , by starting with the presentation

4 2
A 7 = a, b a−3 ab2 = b5 = b2 a−1 ba = 1
in [CRKMW] we found a convenient presentation
3 2
2 . A 7 = x, y x14 = y 10 = xy −1 = y 5 , x = 1, (xy )4 = y 3 x−1 yxy −1 = y 5 .
This group has a unique involution, namely the central involution y 5 (= x7 ), so a subgroup H has
trivial core if and only if it has odd order. The largest such subgroups have order 21 and index 240,
and they form a single conjugacy class represented by H = xy 5 , yx−1 y −1 xyxy 2 x−1
, so this subgroup
can be used in applications of Lemma 5 and Corollary 6. The fact that Aut(2 . A 7 ) ∼ = S 7 allows us to
count the hypermaps of various types with monodromy group 2 . A 7 .
8. The totally chiral maps of least genus
Maps are simply hypermaps of type (l, 2, n) (that is, of map type {n, l} in the notation of [CM,
Chapter 8]), so they are duals of hypermaps of type (2, l, n). The classification in Theorem B and its
associated tables can therefore be used to determine the totally chiral maps of genus g 1001, and
in particular it tells us those of least genus, described in more detail in the following example.
Example 5. Table 5 shows that M 11 is the monodromy group of an orientably regular hypermap
M11 [1] of type (2, 4, 11), and hence of an orientably regular map of type {11, 4}. One can use char-
acter theory to show that there are exactly two such maps. The group M 11 has a single conjugacy
class of 990 elements of order 4, a single conjugacy class of 165 elements of order 2, and two con-
jugacy classes each of 720 elements of order 11. By formula (3.1) it follows from the character table
in [CCNPW] that the number of triples of type (4, 2, 11) in this group is

990 × 165 × 720 2 × 2 × −1 1 × −3 × 1
2× 1+ + = 2| M 11 |,
7920 10 45
with only the characters χ1 , χ2 and χ9 of degrees 1, 10 and 45 contributing to the sum. No maximal
subgroup of M 11 contains elements of orders 4 and 11, so each such triple generates M 11 . Since M 11
has no outer automorphisms it follows that there are two orientably regular maps of type {11, 4} with
monodromy group M 11 . By (2.1) these maps have genus 631. They form a chiral pair M and M since
the elements of order 11 in M 11 are not inverted by any automorphism. Alternatively, one can use the
natural representation of degree 11 of M 11 and note that the faithful quotient M11 [1] / M 10 of M11 [1]
by the point stabiliser H = M 10 is chiral, as shown in Fig. 3 with x = (2, 10)(4, 11)(5, 7)(8, 9) and
y = (1, 4, 3, 8)(2, 5, 6, 9); now Corollary 6(b) applies to H since all automorphisms of M 11 are inner,
so we have a chiral pair. Since M 11 is simple these two maps are totally chiral, as are their associates,
which include a chiral pair of maps of type {4, 11}. Table 5 shows all the totally chiral hypermaps of
genus at most 1001 with monodromy group M 11 , including M11 [1] of type (2, 4, 11), a dual of one of
the chiral pair. The table also shows that there are totally chiral maps of genus 694, 826, 829, 961 and
991 with monodromy group M 11 .
To summarise, we have proved:

Fig. 3. The faithful quotient M11 [1] / M 10 .
Theorem C. The least genus of any totally chiral map is 631, attained by four maps with monodromy group
isomorphic to the Mathieu group M 11 . They consist of a chiral pair of type {11, 4} and their duals of type {4, 11}.
It may seem strange that M 11 appears in Theorem C, and not a smaller perfect group such as d . A 7
(d = 1, 2, 3), L 3 (3) or U 3 (3). The reason is simple: there are no chiral maps with these monodromy
groups. For instance, there are very few orientably regular maps with monodromy group A 7 : there
are just two regular maps of type {6, 7} and their duals, of genus 241, and one self-dual regular map
of type {7, 7} and genus 271. In the case of 2 . A 7 there are no orientably regular maps at all, only
hypermaps, since this group has only one involution, lying in the centre. All the orientably regular
hypermaps with monodromy group L 3 (3) are regular, the least genus being 253 attained by two
regular maps of type {3, 13} and their duals. The same applies to U 3 (3), the least genus in this case
being 577 attained by a regular map of type {6, 7} and its dual, and also to 3 . A 7 , where the least
genus is 721, attained by two regular maps of type {6, 7} and their duals.
9. The totally chiral composite hypermaps (and maps) of least genus
Let us define an orientably regular hypermap H to be simple if its monodromy group G is a

simple group, and composite otherwise. By Theorem B, the totally chiral composite hypermaps of
genus g 1001 are those in Tables 7 and 8 corresponding to the central extensions G = 2 . A 7 and
3 . A 7 of A 7 , and by inspection the least genus of such a hypermap is 631. More specifically, we have:
Theorem D. The least genus of any totally chiral composite hypermap is 631, attained by twelve hyper-
maps formed from 3 . A7 [1] and 3 . A7 [2] by taking mirror images and associates. They all have monodromy
group 3 . A 7 and type (3, 4, 4), (4, 3, 4) or (4, 4, 3).
Remarks. 1. The quotients of these hypermaps by the centre N ∼ = C 3 of 3 . A 7 are the twelve hyper-
maps of genus 211 in the Γ -orbits of A7 [1] and A7 [2] , appearing in Examples 1 and 2 and Theorem A.
2. Fig. 4(a) shows the faithful quotient of 3 . A7 [1] by a subgroup H ∼
= S 5 of index 63 in 3 . A 7 . This
Walsh map is on a torus, and it is easy to see a translation which is a fixed-point-free automorphism
of order 3 of the map, induced by the centre N ∼ = C 3 of the automorphism group 3 . A 7 . The quotient of
this map by N (Fig. 4(b)) is the faithful quotient, again on a torus, of the orientably regular hypermap
A7 [1] = 3 . A7 [1] / N, with monodromy group A 7 , by H ; the torus in Fig. 4(a) is tessellated by three
copies of a square fundamental region for N.
3. Although 631 is the least genus of any totally chiral hypermap with monodromy group 3 . A 7 ,
there are chiral hypermaps of smaller genus with this monodromy group. The least such genus is 505,
attained by a chiral pair H, H and a regular hypermap, all of type (3, 3, 5), together with their
associates. The quotient hypermap H/ N ∼ = H/ N of genus 169 does not contradict the minimality
of the genus 211 in Table 2, since it is regular; thus H has chirality index 3 and is not totally chiral.
Among the orientably regular hypermaps with monodromy group 3 . A 7 , the genus 631 is also attained
by four chiral pairs and a regular hypermap, all of type (3, 3, 6), together with their associates. Again,
the chiral hypermaps have chirality index 3, so they are not totally chiral.
Fig. 4. (a) The faithful quotient of the totally chiral composite hypermap 3 . A7 [1] by H . (b) The faithful quotient A7 [1] / H . Both
are on tori.
4. It is strange that this genus 631 is also the least genus of any totally chiral map (see Section 8),
although there is no apparent direct connection between the two sets of hypermaps attaining these
lower bounds, between their monodromy groups M 11 and 3 . A 7 , between the corresponding triangle
groups (2, 4, 11) and (3, 4, 4), or between the Riemann surfaces and algebraic curves associated
with these hypermaps in Grothendieck’s theory of dessins d’enfants.
5. It may also seem strange that it is 3 . A 7 , and not the smaller group 2 . A 7 , which is the mon-
odromy group of the totally chiral composite hypermaps of least genus. The reason is that 2 . A 7 has
no generating triples with large values of λ, corresponding to hypermaps of low genus: there are none
of type (2, m, n) for any m and n (since the only involution is in the centre), none of type (3, 3, n)
for any n < 10, and none of type (3, 4, n) or (4, 4, n) for any n. The least genera of orientably regular
and chiral hypermaps with monodromy group 2 . A 7 are respectively 589, attained by a regular hyper-
map of type (3, 3, 10), and 673, attained by a totally chiral pair and a regular hypermap, all of type
(3, 5, 5).
Although it takes us outside our self-imposed range g 1001, we conclude this section by com-
puting the least genus of any totally chiral composite map.
Theorem E. The least genus of any totally chiral composite map is 1457, attained by 48 maps with monodromy
group 2 . Sz(8). They consist of 12 chiral pairs of type {4, 5}, together with their duals of type {5, 4}.
Proof. The ATLAS gives a presentation
7 7
x, y x2 = y 4 = (xy )5 = xy 2 = xyxy −1 xy 2 = 1
for the double covering G = 2 . S of the Suzuki group S = Sz(8), showing that G is the monodromy
group of an orientably regular hypermap H of type (2, 4, 5) and genus 1457. This is a double covering
of an orientably regular hypermap H of that type and of genus 729. Applying a triality operation
gives a corresponding pair of maps M and M of type {4, 5} and of these genera, with monodromy
groups G and S. The chirality group of M is a normal subgroup of G, so it is either trivial, the central
subgroup of order 2, or G; the first two cases can be eliminated since they would imply that M is
chiral, whereas no element of order 4 in S is inverted by any automorphism [Suz]. Thus M is totally
chiral (and H is one of the four totally chiral hypermaps of type (2, 4, 5) and genus 729 appearing in
the proof of Theorem B and in Table 8).
Now suppose that M is a totally chiral composite map of genus g 1457. Being composite, it
must be a d-sheeted covering, for some d 2, of a simple hypermap M∗ , which is totally chiral by
Lemma 4(a). By Theorem D, M∗ has genus g ∗ 631, so g = d( g ∗ − 1)+ 1 630d + 1, giving d = 2 and
hence g ∗ 729. Now Theorem B and the associated Tables 2–6 for the simple groups A 7 , L 3 (3), U 3 (3),
M 11 and Sz(8) show that the only possibilities are g ∗ = 631, 694 or 729, corresponding to hypermaps
of type (2, 4, 11), (2, 5, 8) or (2, 4, 5), with monodromy groups M 11 , M 11 or Sz(8) respectively. Since
d = 2 the d-sheeted covering is central, whereas M 11 has trivial Schur multiplier, so the first two cases
cannot arise here. It follows that g ∗ = 729, g = 1457 and the monodromy group of M is the unique
proper double covering G of S.
A calculation with GAP shows that there are 698880 = 24| S | generating triples of type (2, 4, 5)
in G. Since Aut G ∼ = G / Z (G ) ∼
= S it follows that we obtain 24 totally chiral maps M of type {4, 5},
forming 12 chiral pairs, together with their 24 duals. 2
Remark. The Schur multiplier of the group S = Sz(8) is isomorphic to C 2 × C 2 , so the universal
covering group Ŝ = 22 . S has three central subgroups of order 2, the quotient by each giving a double
covering 2 . S of S. The outer automorphism group C 3 of S, induced by the Galois group of the field F 8 ,
lifts to a group of outer automorphisms of Ŝ, permuting these three central subgroups transitively,
so the corresponding covering groups 2 . S are isomorphic to each other, and each can be taken to
be G.
Each epimorphism from = (2, 4, 5) to S factors through Ŝ, so each normal subgroup N
of with quotient S contains a normal subgroup M of with / M ∼ = Ŝ and N / M ∼= C2 × C2.
It follows that there are three normal subgroups K of with M < K < N, corresponding to
three totally chiral maps M of type {4, 5}, genus 1457 and monodromy group / K ∼ = G, which
are double coverings of the map M of genus 729 corresponding to N. However, these are not
the only double coverings of M with these properties: the group has a derived group of
index 2, so each of these groups K satisfies N /( K ∩ ) ∼ = C 2 × C 2 , giving three subgroups of
index 2 in N containing K ∩ . All three are normal in , since their images are central in
/( K ∩ ) ∼ = S × C 2 , one is K , with / K ∼
= G × C 2 . One of them is N ∩ , with /( N ∩ ) ∼ = G,
∗ ∗ ∼
and the third (let us call it K ) also has quotient group / K = G. The six groups K and K ∗ cor-
respond to six double coverings M of M with monodromy group G, so the four totally chiral
maps M of type {4, 5} and genus 729 enumerated earlier lift to 24 totally chiral maps M of type
{4, 5} and genus 1457. The same applies to their duals, thus accounting for the 48 maps in Theo-
rem E.
10. Infinite families
So far we have concentrated on individual examples or finite families of totally chiral hypermaps,
but it is possible to give uniform constructions of infinite families.
Example 6. It is shown in [JS] that each of the Suzuki groups Sz(q) (q = 2e , odd e 3) is the mon-
odromy group of a family of orientably regular maps of type {4, 5} and genus
|Sz(q)| q2 (q2 + 1)(q − 1)

1+ =1+ .
40 40
These maps and their duals are chiral, since the elements of order 4 in Sz(q) are not inverted by any
automorphism of the group; since Sz(q) is simple they are totally chiral. There is a similar construction
of totally chiral maps of type {3, 7} based on the Ree groups Re(3e ), which are simple for all odd
e 3 [Jon].
Example 7. Theorem 13 of [BJNŠ] describes a totally chiral hypermap with monodromy group A n for
each n 7. If n is odd, one can take the generators x = (1 2 . . . n) and y = (1 2 4), giving a hypermap
1
of type (n, 3, n) and genus 1 + 6
(n − 1)!(n − 3). If n is even, one can take x = (1 2 . . . n − 1) and
y = (1 n)(2 3), in which case the type is (n − 1, 2, n − 1) and the genus is 1 + 14 (n − 2)!n(n − 5). (Values
n < 7 are excluded by Lemma 4, since A n is solvable for n 4, while A 5 ∼ = L 2 (4) and A 6 ∼= L 2 (9).)
One can also produce infinite families of totally chiral hypermaps as coverings of a single hyper-
map.
Proposition 7. Let H and H̃ be orientably regular hypermaps corresponding to normal subgroups K and L
of , where K L and K / L is a solvable group of order coprime to | : K |. If H is totally chiral then so is H̃.
Proof. Suppose that H is totally chiral and H̃ is not, so K K = but M := LL < . Recall that K is
the image of K under the automorphism of inverting the generators X and Y .
We first show that K M = , or equivalently K L = . Since = K K we have / K L ∼ = K /( K ∩ K L ),
which is an epimorphic image of K / L ∼ = K / L and is therefore solvable. On the other hand, / K L is
also an epimorphic image of the perfect group / K , so it is trivial. Thus = K L = K M.
It follows that / M = K M / M ∼ = K / K ∩ M with K ∩ M L, so / M is a quotient of K / L and is
therefore solvable, of order coprime to |G |. It follows that has a proper normal subgroup N M
such that / N is an elementary abelian p-group P for some prime p not dividing | : K |. We
have K N = , so Q := K / K ∩ N ∼ = / N ∼
= P . Since K is isomorphic to the fundamental group π1 H
of H it follows that Q is a quotient of the first homology group H 1 (H; F p ) ∼ = K / K K p over the
field F p , where K is the commutator subgroup of K and K p is the subgroup generated by its
pth powers. Now H 1 (H; F p ) is an F p G-module where G = / K , with the action of G as Aut H
on H 1 (H; F p ) identified with its induced action by conjugation on K / K K p . Since p is coprime
to |G |, Maschke’s theorem implies that H 1 (H; F p ) is a direct sum of irreducible submodules. We
have / K ∩ N = K N / K ∩ N = ( K / K ∩ N ) × ( N / K ∩ N ) = Q × ( N / K ∩ N ), so G, acting as N / K ∩ N,
acts trivially on Q . Thus the principal F p G-module has non-zero multiplicity as a direct summand
of H 1 (H; F p ). However, it follows from a theorem of Sah [Sah, Theorem 3.5] that if a finite group G
has an orientation-preserving action on a compact orientable surface S, and F is a field of character-
istic not dividing |G |, then the dimension of the subspace of H 1 ( S ; F ) fixed by G is twice the genus
of S /G. When S is the underlying surface of an orientably regular hypermap with automorphism
group G this genus is 0, so the principal module has zero multiplicity. This contradiction proves the
result. 2
(Sah’s theorem is stated and proved in [Sah] for finite groups acting conformally on compact Rie-
mann surfaces. In the above proof, S inherits a Riemann surface structure, preserved by G, from that
of the simply connected Riemann surface — in this case the hyperbolic plane — on which the triangle
group acts conformally.)
Corollary 8. Each totally chiral hypermap has infinitely many totally chiral covering hypermaps of the same
type. (In particular, any totally chiral map has infinitely many totally chiral covering maps of the same type.)
Proof. Let H be a totally chiral hypermap of genus g, corresponding to a subgroup K of as in

Proposition 7. Then K is a surface group of genus g 1 (in fact g 211 by Theorem A), so there
are infinitely many characteristic subgroups L of K such that K / L is a solvable group of finite order
coprime to | : K |: for instance we could take L = K K n , so that K / L ∼
2g
= C n , for any integer n coprime
to | : K |. Each such L is normal in , so the corresponding hypermap H̃ is orientably regular, and
of the same type as H; by Proposition 7 it is totally chiral. 2
The existence of totally chiral hypermaps with monodromy groups 2 . A 7 or 3 . A 7 (see Tables 7
and 8) shows that the conclusion of Proposition 7 can sometimes be valid even if | K : L | and
| : K | are not coprime. However, in this case there are many counterexamples, such as the fol-
lowing:
Example 8. Table 2 shows that there is a totally chiral hypermap H of type (4, 4, 4) and genus 316
with monodromy group G = A 7 . Let K be the corresponding hypermap subgroup of = (4, 4, 4),
and let L = K ∩ , so K / L ∼
= / ∼= C 4 × C 4 . Then LL = , so the hypermap H̃ corresponding to L
is not totally chiral. The point is that Sah’s theorem does not apply as in the proof of Proposition 7
when |G | is divisible by p (= 2 here).
Appendix A
Here we present the tables which accompany Theorem B, listing all the totally chiral hypermaps
of genus g 1001. There is one table for each of the possible monodromy groups G = A 7 , L 3 (3),
U 3 (3), M 11 , Sz(8), 2 . A 7 and 3 . A 7 . In each table we describe one representative of each Γ -orbit,
where Γ ∼ = S 3 × C 2 is the group of hypermap operations generated by triality and reflection. We have
chosen each representative to have type (l, m, n) with l m n, so the Γ -orbits containing maps are
those with l = 2. In each case we give the genus of the hypermaps and the number of them in the
orbit.
Table 2
The chiral hypermaps with monodromy group A 7 and g 1001.
Type g Trip. Refl. Chir. H
(3, 4, 4) 211 20160 0 4 A7 [1] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 2)(3, 5, 6, 7))
A7 [2] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 5)(2, 4, 6, 7))
(3, 4, 5) 274 25200 1 4 A7 [3] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 3, 5, 7)(4, 6))
A7 [4] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 4, 6, 5)(2, 7))
(3, 4, 6) 316 15120 1 2 A7 [5] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 7, 3, 6)(2, 4))
(4, 4, 4) 316 60480 0 12 A7 [6] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 3)(2, 5, 6, 7))
A7 [7] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 5)(2, 4, 7, 3))
A7 [8] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 4)(2, 3, 5, 7))
A7 [9] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 6, 4, 2)(5, 7))
A7 [10] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 6, 7, 5)(3, 4))
A7 [11] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 7)(3, 4, 6, 5))
(3, 5, 5) 337 15120 1 2 A7 [12] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 7, 6, 5, 4))
(3, 4, 7) 346 40320 0 8 A7 [13] = ( A 7 ; (1, 2, 3), (2, 7, 6, 4)(3, 5))

A7 [14] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 4, 5, 6)(3, 7))
A7 [15] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 7, 6, 3)(2, 4))
A7 [16] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 5)(2, 7, 3, 4))
(4, 4, 5) 379 55440 3 8 A7 [17] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 4, 5, 6)(3, 7))
A7 [18] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 4)(3, 6, 5, 7))
A7 [19] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 4)(2, 5, 6, 7))
A7 [20] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 4, 7, 5)(2, 3))
(3, 5, 7) 409 45360 3 6 A7 [21] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 6, 4, 7, 5))

A7 [22] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 5, 3, 7, 4))
A7 [23] = ( A 7 ; (1, 2, 3)(4, 5, 6), (3, 7, 6, 4, 5))
(4, 4, 6) 421 30240 2 4 A7 [24] = ( A 7 ; (1, 2, 3, 4)(5, 6), (2, 4, 7, 6)(3, 5))
A7 [25] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 7, 4, 6)(2, 5))
(4, 5, 5) 442 20160 0 4 A7 [26] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 7, 6, 5, 2))

A7 [27] = ( A 7 ; (1, 2, 3, 4)(5, 6), (2, 7, 6, 4, 3))
(3, 6, 7) 451 20160 2 2 A7 [28] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 4, 5)(2, 6)(3, 7))
Table 2 (continued)
Type g Trip. Refl. Chir. H

(4, 4, 7) 451 110880 6 16 A7 [29] = ( A 7 ; (1, 2, 3, 4)(5, 6), (2, 7, 6, 4)(3, 5))
A7 [30] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 6, 2, 5)(4, 7))
A7 [31] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 7, 6, 4)(2, 5))
A7 [32] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 7, 6, 3)(2, 4))
A7 [33] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 3, 4, 5)(2, 7))
A7 [34] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 2, 4, 7)(3, 6))
A7 [35] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 2, 6, 3)(5, 7))
A7 [36] = ( A 7 ; (1, 2, 3, 4)(5, 6), (2, 5)(3, 7, 4, 6))
(3, 7, 7) 481 65520 7 6 A7 [37] = ( A 7 ; (1, 2, 3), (1, 5, 2, 7, 4, 6, 3))
A7 [38] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 6, 2, 5, 3, 7, 4))
A7 [39] = ( A 7 ; (1, 2, 3)(4, 5, 6), (1, 2, 5, 6, 3, 4, 7))
(4, 5, 6) 484 20160 2 2 A7 [40] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 3, 7, 2, 6))
(4, 5, 7) 514 90720 0 18 A7 [41] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 6, 4, 7, 5))
A7 [42] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 5, 7, 4, 6))
A7 [43] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 2, 4, 6, 7))
A7 [44] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 3, 4, 5, 7))
A7 [45] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 5, 3, 7, 4))
A7 [46] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 7, 5, 3, 6))
A7 [47] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 5, 2, 4, 7))
A7 [48] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 2, 6, 3, 7))
A7 [49] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 2, 3, 5, 7))
(5, 5, 6) 547 20160 2 2 A7 [50] = ( A 7 ; (1, 2, 3, 4, 5), (1, 7, 5, 3, 6))
(4, 6, 7) 556 40320 4 4 A7 [51] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 5)(2, 3, 6)(4, 7))
A7 [52] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 7, 6)(2, 4)(3, 5))
(5, 5, 7) 577 60480 4 8 A7 [53] = ( A 7 ; (1, 2, 3, 4, 5), (1, 6, 4, 7, 5))
A7 [54] = ( A 7 ; (1, 2, 3, 4, 5), (1, 2, 4, 6, 7))
A7 [55] = ( A 7 ; (1, 2, 3, 4, 5), (1, 2, 5, 6, 7))
A7 [56] = ( A 7 ; (1, 2, 3, 4, 5), (1, 3, 6, 5, 7))
(4, 7, 7) 586 100800 0 20 A7 [57] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 6, 4, 2, 7, 5, 3))
A7 [58] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 5, 2, 6, 3, 7, 4))
A7 [59] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 3, 5, 7, 2, 4, 6))
A7 [60] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 5, 2, 7, 4, 6, 3))
A7 [61] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 6, 2, 3, 4, 7, 5))
A7 [62] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 3, 4, 5, 7, 2, 6))
A7 [63] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 6, 7, 2, 5, 3, 4))
A7 [64] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 6, 3, 4, 2, 7, 5))
A7 [65] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 5, 3, 7, 4, 2, 6))
A7 [66] = ( A 7 ; (1, 2, 3, 4)(5, 6), (1, 5, 3, 6, 2, 7, 4))
(5, 6, 7) 619 35280 3 4 A7 [67] = ( A 7 ; (1, 2, 3, 4, 5), (1, 4)(2, 5, 6)(3, 7))
A7 [68] = ( A 7 ; (1, 2, 3, 4, 5), (1, 4)(2, 6)(3, 5, 7))
(5, 7, 7) 649 120960 12 12 A7 [69] = ( A 7 ; (1, 2, 3, 4, 5), (1, 4, 5, 6, 2, 7, 3))
A7 [70] = ( A 7 ; (1, 2, 3, 4, 5), (1, 3, 4, 7, 2, 5, 6))
A7 [71] = ( A 7 ; (1, 2, 3, 4, 5), (1, 4, 6, 2, 7, 3, 5))
A7 [72] = ( A 7 ; (1, 2, 3, 4, 5), (1, 4, 6, 7, 2, 3, 5))
A7 [73] = ( A 7 ; (1, 2, 3, 4, 5), (1, 2, 5, 6, 3, 4, 7))
A7 [74] = ( A 7 ; (1, 2, 3, 4, 5), (1, 6, 7, 2, 5, 3, 4))
(6, 7, 7) 691 35280 5 2 A7 [75] = ( A 7 ; (1, 2, 3)(4, 5)(6, 7), (1, 5, 2, 7, 4, 6, 3))
(7, 7, 7) 721 115920 15 8 A7 [76] = ( A 7 ; (1, 2, 3, 4, 5, 6, 7), (1, 4, 5, 6, 7, 2, 3))
A7 [77] = ( A 7 ; (1, 2, 3, 4, 5, 6, 7), (1, 5, 7, 4, 6, 2, 3))
A7 [78] = ( A 7 ; (1, 2, 3, 4, 5, 6, 7), (1, 6, 7, 4, 2, 5, 3))
A7 [79] = ( A 7 ; (1, 2, 3, 4, 5, 6, 7), (1, 6, 7, 2, 5, 3, 4))
Table 3
The chiral hypermaps with monodromy group L 3 (3) and g 1001.
Type g Trip. Refl. Chir. H (x, y )

(3, 3, 4) 235 33696 1 2 L3 (3)[1] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 13, 8)(2, 3, 5)(4, 12, 6)(7, 9, 11)
(3, 3, 6) 469 44928 2 2 L3 (3)[2] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 9, 5)(2, 4, 6)(7, 10, 12)(8, 13, 11)
(3, 4, 4) 469 44928 0 4 L3 (3)[3] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 8)(2, 9, 7, 5)(4, 11)(6, 13, 12, 10)
L3 (3)[4] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 2, 5, 12)(3, 4)(6, 11, 8, 7)(10, 13)
(3, 3, 8) 586 112320 4 6 L3 (3)[5] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 10, 7)(2, 9, 12)(3, 5, 8)(6, 11, 13)
L3 (3)[6] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 8, 11)(2, 6, 13)(3, 12, 7)(5, 10, 9)
L3 (3)[7] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 3, 8)(2, 11, 5)(4, 9, 7)(6, 12, 10)
(3, 4, 6) 703 44928 2 2 L3 (3)[8] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 8)(2, 10, 5, 3)(4, 11)(6, 12, 9, 7)
(4, 4, 4) 703 33696 1 2 L3 (3)[9] x = (2, 11)(3, 12, 9, 7)(4, 10, 8, 6)(5, 13)
y = (1, 8, 12, 6)(2, 4, 5, 3)(7, 9)(10, 11)
(3, 3, 13) 721 112320 8 2 L3 (3)[10] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 5, 4)(3, 7, 8)(6, 11, 9)(10, 13, 12)
(3, 4, 8) 820 112320 4 6 L3 (3)[11] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 12, 4, 3)(2, 10, 5, 7)(6, 8)(9, 13)
L3 (3)[12] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 9)(2, 11, 12, 8)(3, 5)(6, 10, 7, 13)
L3 (3)[13] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 2, 4, 3)(5, 6, 12, 13)(7, 11)(9, 10)
(3, 6, 6) 937 67392 2 4 L3 (3)[14] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 2, 8, 9, 11, 12)(3, 13, 10)(6, 7)
L3 (3)[15] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 12, 13)(2, 4, 3, 5, 8, 6)(9, 10)
(4, 4, 6) 937 67392 4 2 L3 (3)[16] x = (2, 11)(3, 12, 9, 7)(4, 10, 8, 6)(5, 13)
y = (1, 12, 2, 5)(3, 7, 4, 11)(6, 8)(9, 13)
(3, 4, 13) 955 157248 8 6 L3 (3)[17] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 12)(2, 6, 3, 9)(4, 10, 11, 5)(7, 13)
L3 (3)[18] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 12, 2, 5)(3, 7, 4, 11)(6, 8)(9, 13)
L3 (3)[19] x = (2, 6, 10)(3, 7, 4)(5, 13, 8)(9, 12, 11)
y = (1, 4, 5, 2)(3, 12)(6, 11)(7, 13, 10, 8)
Table 4
The chiral hypermaps with monodromy group U 3 (3) and g 1001.

[1]
(3, 3, 6) 505 48384 0 4 U3 (3) x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 16, 23)(2, 25, 20)(3, 18, 22)(4, 11, 24)(5, 15, 9)(6, 27, 8)(7, 13, 10)
(14, 19, 21)(17, 28, 26)
U3 (3)[2] x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 21, 13)(2, 20, 28)(3, 15, 10)(4, 7, 5)(8, 17, 26)(9, 18, 22)(11, 19, 23)
(12, 25, 16)(14, 27, 24)
Table 4 (continued)

[3]
(3, 3, 7) 577 120960 2 8 U3 (3) x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 4, 2)(3, 8, 20)(5, 10, 11)(6, 14, 25)(7, 26, 22)(9, 23, 28)(12, 24, 16)
(13, 17, 19)(15, 21, 18)
U3 (3)[4] x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 22, 11)(2, 15, 19)(3, 13, 7)(4, 8, 23)(5, 20, 18)(6, 27, 12)(9, 10, 17)
(14, 24, 16)(21, 26, 28)
U3 (3)[5] x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 8, 4)(2, 20, 16)(3, 15, 24)(5, 17, 11)(6, 18, 9)(7, 19, 13)(10, 14, 27)
(12, 25, 28)(21, 23, 26)
U3 (3)[6] x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 21, 20)(2, 14, 28)(3, 19, 22)(4, 13, 5)(6, 11, 9)(7, 27, 24)(8, 10, 15)
(12, 23, 26)(16, 17, 18)
(3, 3, 8) 631 120960 4 6 U3 (3)[7] x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 14, 16)(2, 6, 17)(3, 20, 13)(4, 11, 12)(5, 23, 7)(8, 24, 26)(9, 15, 25)
(10, 21, 18)(19, 27, 28)
U3 (3)[8] x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 21, 11)(3, 9, 25)(4, 23, 18)(5, 14, 20)(6, 12, 8)(7, 10, 22)(13, 26, 15)
(16, 24, 27)(17, 19, 28)
U3 (3)[9] x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 5, 13)(2, 7, 12)(3, 17, 28)(4, 25, 20)(6, 14, 15)(8, 26, 18)(9, 23, 19)
(10, 22, 24)(11, 27, 16)
[10]
(3, 3, 12) 757 72576 4 2 U3 (3) x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 20, 24)(2, 5, 13)(3, 19, 25)(4, 28, 14)(6, 16, 10)(7, 21, 27)(8, 23, 11)
(9, 12, 17)(15, 18, 26)
[11]
(3, 4, 7) 829 84672 5 2 U3 (3) x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 7, 26, 11)(2, 14)(3, 12)(4, 18, 13, 22)(5, 21, 8, 19)(6, 17, 9, 23)
(10, 15, 24, 27)(16, 20, 28, 25)
[12]
(3, 4, 8) 883 72576 2 4 U3 (3) x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 11, 26, 7)(2, 14)(3, 12)(4, 22, 13, 18)(5, 19, 8, 21)(6, 23, 9, 17)
(10, 27, 24, 15)(16, 25, 28, 20)
U3 (3)[13] x = (2, 17, 25)(3, 5, 7)(4, 21, 14)(6, 15, 26)(8, 18, 13)(9, 20, 28)(10, 12, 19)
(11, 24, 16)(22, 23, 27)
y = (1, 13, 27, 14)(2, 24)(3, 25, 19, 22)(4, 21)(5, 20, 28, 7)(6, 11, 26, 18)
(8, 10, 17, 12)(9, 16, 15, 23)
Table 5
The chiral hypermaps with monodromy group M 11 and g 1001.

[1]
(2, 4, 11) 631 15840 0 2 M11 x = (2, 10)(4, 11)(5, 7)(8, 9)
y = (1, 4, 3, 8)(2, 5, 6, 9)
(2, 5, 8) 694 47520 0 6 M11 [2] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 3, 11, 5, 2)(4, 8, 9, 6, 7)
M11 [3] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 3, 6, 9, 11)(2, 5, 4, 8, 7)
M11 [4] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 5, 4, 7, 10)(2, 9, 6, 11, 8)
[5]
(2, 6, 8) 826 47520 0 6 M11 x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 5, 9)(2, 6, 7, 10, 4, 8)(3, 11)
M11 [6] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 6, 5, 10, 3, 4)(2, 8, 7)(9, 11)
M11 [7] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 7, 8, 5, 2, 3)(4, 9)(6, 10, 11)
(3, 3, 8) 826 31680 0 4 M11 [8] x = (3, 10, 6)(4, 8, 9)(5, 7, 11)
y = (1, 9, 4)(2, 6, 10)(3, 8, 5)
M11 [9] x = (3, 10, 6)(4, 8, 9)(5, 7, 11)
y = (1, 2, 8)(3, 10, 9)(4, 11, 7)
(2, 5, 11) 829 15840 0 2 M11 [10] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 3, 11, 7, 10)(2, 8, 5, 9, 6)
[11]
(3, 4, 5) 859 39600 1 4 M11 x = (3, 10, 6)(4, 8, 9)(5, 7, 11)
y = (1, 11, 7, 6)(2, 4, 3, 10)
M11 [12] x = (3, 10, 6)(4, 8, 9)(5, 7, 11)
y = (1, 9, 11, 2)(5, 10, 8, 7)
[13]
(2, 6, 11) 961 47520 0 6 M11 x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 11, 10, 2, 9, 7)(3, 4)(5, 8, 6)
M11 [14] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 3, 8, 7, 5, 2)(4, 11)(6, 9, 10)
M11 [15] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 8, 3)(2, 11)(4, 6, 5, 7, 9, 10)
(2, 8, 8) 991 31680 0 4 M11 [16] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 4, 11, 10, 5, 6, 2, 3)(8, 9)
M11 [17] x = (3, 7)(4, 8)(5, 10)(6, 11)
y = (1, 4, 8, 9, 10, 2, 5, 6)(7, 11)
(3, 4, 6) 991 39600 1 4 M11 [18] x = (3, 10, 6)(4, 8, 9)(5, 7, 11)
y = (1, 4, 3, 8)(2, 5, 6, 9)
M11 [19] x = (3, 10, 6)(4, 8, 9)(5, 7, 11)
y = (1, 3, 10, 11)(2, 8, 5, 4)
The central extension 2 . A 7 has a faithful permutation representation of degree 240 on the cosets
of H = xy 5 , yx−1 y −1 xyxy 2 x−1
given by
x = (1, 2)(3, 11, 39, 9, 38, 118, 30, 7, 29, 21, 5, 20, 49, 12)
(4, 16, 66, 195, 98, 25, 6, 8, 34, 127, 225, 143, 42, 10)
(13, 53, 171, 51, 79, 130, 96, 31, 121, 200, 120, 134, 72, 54)
Table 6
The chiral hypermaps with monodromy group Sz(8) and g 1001.

[1]
(2, 4, 5) 729 349440 0 4 S z(8) x = (1, 2)(3, 4)(5, 7)(6, 9)(8, 12)(10, 13)(11, 15)(14, 19)(16, 21)(17, 23)
(18, 25)(20, 28)(22, 31)(24, 33)(26, 35)(27, 32)(29, 37)(30, 39)
(34, 43)(36, 46)(38, 48)(41, 51)(42, 44)(45, 55)(47, 50)(49, 58)
(52, 60)(53, 61)(54, 59)(56, 62)(57, 63)(64, 65)
y = (1, 3, 5, 8)(4, 6, 10, 14)(7, 11, 16, 22)(9, 12, 17, 24)(13, 18, 26, 36)
(15, 20, 29, 38)(19, 27, 31, 28)(21, 30, 40, 50)(23, 32, 41, 52)
(25, 34, 44, 54)(33, 42, 53, 43)(35, 45, 56, 63)(37, 47, 51, 46)
(39, 49, 59, 60)(48, 57, 55, 58)(61, 64, 62, 65)
[2]
S z(8) x = (2, 45)(3, 27)(4, 49)(5, 24)(6, 10)(7, 38)(8, 25)(9, 51)(11, 39)
(12, 63)(13, 60)(14, 55)(15, 44)(16, 48)(17, 19)(18, 53)(20, 26)
(21, 61)(22, 57)(23, 35)(28, 40)(29, 52)(30, 50)(31, 34)(32, 62)
(33, 46)(36, 54)(37, 47)(41, 58)(42, 56)(43, 65)(59, 64)
y = (1, 12, 10, 27)(2, 37, 33, 3)(4, 21, 43, 59)(5, 60, 14, 35)(6, 52, 62, 41)
(7, 46, 32, 64)(8, 18, 15, 29)(9, 24, 48, 25)(11, 30, 53, 40)
(13, 26, 47, 19)(16, 61, 57, 42)(17, 55, 39, 50)(20, 34, 36, 58)
(23, 45, 54, 65)(28, 56, 49, 44)(31, 51, 63, 38)
(14, 56, 131, 157, 108, 27, 107, 32, 88, 73, 153, 45, 44, 57)
(15, 61, 41, 140, 93, 202, 124, 33, 52, 23, 90, 142, 186, 62)
(17, 70, 167, 84, 219, 109, 129, 35, 75, 170, 137, 174, 150, 71)
(18, 26, 103, 149, 101, 230, 132, 36, 43, 146, 106, 67, 169, 74)
(19, 77, 159, 47, 158, 114, 55, 37, 133, 172, 116, 227, 97, 78)
(22, 86, 65, 69, 144, 216, 91, 40, 138, 60, 128, 100, 190, 87)
(24, 94, 155, 240, 135, 237, 113, 28, 112, 236, 213, 81, 207, 95)
(46, 85, 119, 203, 154, 175, 83, 115, 68, 50, 168, 204, 197, 99)
(48, 163, 235, 145, 224, 126, 238, 117, 233, 231, 102, 192, 64, 164)
(58, 177, 212, 176, 208, 226, 220, 122, 193, 206, 182, 160, 229, 178)
(59, 180, 147, 221, 211, 80, 210, 123, 201, 104, 189, 205, 76, 181)
(63, 185, 141, 188, 139, 173, 156, 125, 194, 92, 198, 89, 165, 161)
(82, 151, 218, 187, 166, 152, 191, 136, 110, 217, 228, 162, 111, 215)
(105, 209, 239, 214, 222, 183, 179, 148, 232, 184, 223, 196, 199, 234),
y = (1, 3, 13, 35, 8, 2, 7, 31, 17, 4)(5, 22, 88, 87, 41, 9, 40, 56, 91, 23)
(6, 26, 104, 150, 44, 10, 43, 147, 109, 27)(11, 45, 154, 236, 115, 29, 108, 204, 155, 46)
(12, 34, 101, 230, 119, 30, 16, 67, 169, 50)(14, 58, 179, 176, 123, 32, 122, 234, 182, 59)
(15, 63, 189, 149, 126, 33, 125, 221, 106, 64)(18, 75, 171, 141, 132, 36, 70, 200, 92, 74)
(19, 79, 167, 84, 21, 37, 134, 170, 137, 39)(20, 81, 148, 172, 52, 38, 135, 105, 159, 61)
(24, 96, 68, 198, 114, 28, 54, 85, 188, 97)(25, 99, 117, 173, 144, 42, 83, 48, 165, 100)
(47, 160, 55, 90, 156, 116, 208, 78, 140, 161)(49, 162, 180, 239, 233, 118, 166, 201, 184, 163)
(51, 95, 226, 220, 86, 120, 113, 229, 178, 138)(53, 72, 183, 128, 93, 121, 130, 199, 69, 142)
(57, 111, 224, 127, 197, 107, 152, 192, 66, 175)(60, 177, 157, 219, 194, 65, 193, 153, 174, 185)
(62, 187, 203, 71, 202, 124, 228, 168, 129, 186)(73, 181, 158, 110, 215, 131, 210, 227, 151, 191)
(76, 206, 237, 133, 98, 80, 212, 207, 77, 143)(82, 139, 214, 211, 190, 136, 89, 223, 205, 216)
(94, 225, 209, 146, 222, 112, 195, 232, 103, 196)(102, 218, 164, 213, 235, 145, 217, 238, 240, 231).
Table 7 is based on these generators.
Table 7
The totally chiral hypermaps with automorphism group 2 . A 7 and g 1001.
Type g #Trip. Refl. Chir. H (a, b) H/ C 2

(3, 5, 5) 673 15120 1 2 2A7 [1] a = yxy −1 x3 yx−2
(totally chiral) b = y −1 x2 yx3 A7 [12]
(3, 5, 7) 817 20160 2 2 2A7 [2] a = yxy −1 x3 yx−2
(totally chiral) b = xyx3 y −1 x A7 [23]
(3, 5, 8) 862 25200 1 4 2A7 [3] a = yxy −1 x3 yx−2
(12)
(totally chiral) b = y −2 A7 [4]
2A7 [4] a = yxy −1 x3 yx−2
(12)
(totally chiral) b = x−2 y −1 x−3 A7 [3]
(3, 7, 7) 961 35280 3 4 2A7 [5] a = yxy −1 x3 yx−2
(totally chiral) b = yxy A7 [39]
[6]
2A7 a= yxy −1 x3 yx−2
(totally chiral) b = yx−2 y −1 A7 [38]
[7]
(3, 5, 14) 997 25200 1 4 2A7 a= yxy −1 x3 yx−2
(totally chiral) b = x2 yx3 A7 [21]
[8]
2A7 a= yxy −1 x3 yx−2
(totally chiral) b = x2 y −1 x−7 A7 [22]
For 3 . A 7 the following permutation representation of degree 63 satisfy Lemma 5 and Corollary 6,
and Table 8 that follows is based on these generators.
x = (1, 30, 36)(2, 29, 46)(3, 27, 47)(4, 31, 44)(5, 34, 53)(7, 18, 19)(8, 32, 43)(9, 23, 48)(10, 22, 49)
(11, 24, 42)(12, 25, 50)(14, 28, 17)(15, 33, 51)(16, 26, 52)(20, 35, 21)(37, 62, 40)(38, 59, 58)
(39, 55, 60)(45, 61, 63),
y = (1, 42, 21, 15, 45)(2, 34, 3, 23, 55)(4, 11, 40, 33, 18)(5, 36, 48, 63, 46)(6, 43, 56, 12, 38)
(7, 44, 10, 37, 28)(8, 13, 24, 51, 54)(9, 19, 29, 53, 31)(14, 41, 32, 57, 49)(16, 17, 39, 47, 22)
(25, 26, 59, 60, 27)(30, 50, 35, 58, 61).

Table 8
The totally chiral hypermaps with automorphism group 3 . A 7 and g 1001.
Type g #Trip. Refl. Chir. H (a, b)

[1]
(3, 4, 4) 631 20160 0 4 3 . A7 a = yx−1 y −2 xyxyx
(H/C 3 = A7 [1] ) b = y −1 x−1 y 2 xyx−1 y −1
(totally chiral)
3 . A7 [2] a = yx−1 y −2 xyxyx
(H/C 3 = A7 [2] ) b = y −1 xy 2 xyxy 2 x−1 yx
(totally chiral)
(3, 4, 5) 820 25200 1 4 3 . A7 [3] a = yx−1 y −2 xyxyx

(H/C 3 = A7 [4] ) b = yxy −1 xy −1 x−1 yxy −2 x−1 y 2 xy −1
(totally chiral)
3 . A7 [4] a = yx−1 y −2 xyxyx
(H/C 3 = A7 [3] ) b = y −1 xy −1 x−1 y −2 xy −2
(totally chiral)
(3, 4, 6) 946 45360 1 8 3 . A7 [5] a = yx−1 y −2 xyxyx

(H/C 3 = A7 [5] ) b = xy −1 x−1 y −1 x−1 y −2 x−1 y
(totally chiral)
3 . A7 [6] a = yx−1 y −2 xyxyx
(H/C 3 = A7 [5] ) b = y −1 x−1 y 2 x−1 yx−1 y 2 xyx−1 y −1 xy −1 x−1
(totally chiral)
3 . A7 [7] a = yx−1 y −2 xyxyx
(H/C 3 = A7 [5] ) b = y −1 xy 2 x2 y −3 x−1 y −1 x−1
(totally chiral)
References
[BJ] M. Belolipetsky, G.A. Jones, A bound for the number of automorphisms of an arithmetic Riemann surface, Math. Proc.
Cambridge Philos. Soc. 138 (2005) 289–299. MR 2005k: 30077.
[Bel] G.V. Belyı̆, Galois extensions of a maximal cyclotomic field, Izv. Akad. Nauk SSSR Ser. Mat. 43 (1979) 267–276, 479
(in Russian).
[BJNŠ] A. Breda D’Azevedo, G. Jones, R. Nedela, M. Škoviera, Chirality groups of maps and hypermaps, J. Algebraic Combin.,
in press, doi:10.1007/s10801-008-0138-z.
[Cam] P.J. Cameron, Permutation Groups, London Math. Soc. Stud. Texts, vol. 45, Cambridge Univ. Press, Cambridge, 1999.
MR 2001c: 20008.
[CRKMW] C.M. Campbell, E.F. Robertson, T. Kawamata, I. Miyamoto, P.D. Williams, Deficiency zero presentations for certain
perfect groups, Proc. Roy. Soc. Edinburgh Sect. A 103 (1986) 63–71. MR 87k: 20060.
[Che] C. Chevalley, Theory of Lie Groups, I, Princeton Univ. Press, Princeton, 1946. MR 7, 412c.
[Con07] M.D.E. Conder, http://www.math.auckland.ac.nz/~conder/.
[CCNPW] J.H. Conway, R.T. Curtis, S.P. Norton, R.A. Parker, R.A. Wilson, ATLAS of Finite Groups, Clarendon Press, Oxford, 1985.
MR 88g: 20025.
[Cor] R. Cori, Un code pour les graphes planaires et ses applications, Astérisque 27 (1975).
[CM] H.S.M. Coxeter, W.O.J. Moser, Generators and Relations for Discrete Groups, third ed., Springer-Verlag, Berlin, 1972.
MR 50, 2313.
[GAP] The GAP Group, GAP—Groups, Algorithms, and Programming, Version 4.4.10, http://www.gap-system.org, 2007.
[Gro] A. Grothendieck, Esquisse d’un programme, in: Geometric Galois Actions, 1, in: London Math. Soc. Lecture Note Ser.,
vol. 242, Cambridge Univ. Press, Cambridge, 1997, pp. 5–48. English translation: Sketch of a programme, pp. 243–283.
[HS] A.J. Hoffman, R.R. Singleton, On Moore graphs of diameter two and three, IBM J. Res. Develop. 4 (1960) 497–504.
MR 25, 3857.
[Hup] B. Huppert, Endliche Gruppen I, Springer-Verlag, 1967. MR 37, 302.
[Jon] G.A. Jones, Ree groups and Riemann surfaces, J. Algebra 165 (1994) 41–62. MR 95d: 20030.
[JS] G.A. Jones, S.A. Silver, Suzuki groups and surfaces, J. London Math. Soc. (2) 48 (1993) 117–125. MR 94i: 30037.
[JSin] G.A. Jones, D. Singerman, Belyı̆ functions, hypermaps and Galois groups, Bull. London Math. Soc. 28 (1996) 561–590.
[Mac] A. Machì, On the complexity of a hypermap, Discrete Math. 42 (1982) 221–226. MR 84a: 05024.
[Mon] J.M. Montesinos, Classical Tessellations and Three-Manifolds, Springer-Verlag, Berlin, 1987.
[Sah] C.-H. Sah, Groups related to compact Riemann surfaces, Acta Math. 123 (1969) 13–42. MR 40, 4447.
[Sch] I. Schur, Über die Darstellungen der symmetrischen und alternierenden Gruppen durch gebrochene lineare Substitu-
tionen, J. Math. 139 (1911) 155–250.
[Ser] J.-P. Serre, Topics in Galois Theory, Jones and Bartlett, Boston, 1992. MR 94d: 12006.
[Suz] M. Suzuki, On a class of doubly transitive groups, Ann. of Math. (2) 75 (1962) 105–145.
[Wal] T.R.S. Walsh, Hypermaps versus bipartite maps, J. Combin. Theory Ser. B 18 (1975) 155–163. MR 50, 12778.
[WS] C. Weber, H. Seifert, Die beiden Dodekaederäume, Math. Z. 37 (1933) 237–253.
Journal of Algebra
Some algorithms for semi-invariants of quivers

D.A. Shmelkin
117437, Ostrovitianova, 9-4-187, Moscow, Russia
Article history: We present some theorems and algorithms for calculating perpen-
Received 22 August 2008 dicular categories and locally semi-simple decompositions. We im-
Available online 28 May 2009 plemented a computer program TETIVA based on these algorithms
and we offer this program for everybody’s use.
Keywords:
Generic decomposition
Perpendicular categories
Luna stratification
1. Introduction
Let Q be a finite quiver with the set Q 0 of vertices and Q 1 of arrows; for an arrow ϕ ∈ Q 1 denote
by t ϕ and hϕ its tail and its head, respectively. Let k be an algebraically closed field of characteristic 0.
Q
For a dimension vector α ∈ Z+ 0 consider a tuple (kαi , i ∈ Q 0 ) of standard vector spaces over k of
the corresponding dimensions and the set R ( Q , α ) = ϕ ∈ Q 1 Hom(kαt ϕ , kαhϕ ) of the representations

of Q of dimension α . The group GL(α ) = i ∈ Q 0 GL(αi ) acts naturally on R ( Q , α ) and the orbits of
this action are the isomorphism classes of representations.
Q
Recall that β ∈ Z+ 0 is called a root if R ( Q , β) contains an indecomposable representation and
a Schur root if generic element in R ( Q , β) is indecomposable, and in this case the automorphism
group of generic representation is by [4] the kernel k∗ of the action GL(α ) : R ( Q , α ) consist-
ing of scalar operators. By the Krull–Schmidt theorem for any representation V ∈ R ( Q , α ) there
is a decomposition V = R 1 ⊕ R 2 ⊕ · · · ⊕ R t into a sum of indecomposable summands and it is
unique up to permutations and isomorphisms of summands. In particular, V yields a decomposition
α = dim R 1 + · · · + dim R t into the sum of roots. This decomposition is an important invariant of V
and in some cases it allows to recover the group Aut( V ) or even the isomorphism class of V . It is
shown in [4] that among all the decompositions of α into the sum of roots there is a generic element
such that there is an open dense subset in R ( Q , α ) consisting of representations with such a decom-
position; Kac called this canonical decomposition. We however follow another tradition and prefer the
E-mail address: mitia@mccme.ru.
doi:10.1016/j.jalgebra.2009.05.019
3998 D.A. Shmelkin / Journal of Algebra 322 (2009) 3997–4010
name generic for the same object. By [4] each root in the generic decomposition is a Schur root, and
moreover a decomposition α = β1 + · · · + βt is generic if and only if for generic B i ∈ R ( Q , βi ) holds:
Aut( B i ) = k∗ and Ext( B i , B j ) = 0 for i , j = 1, . . . , t, i = j. Kac addressed in [4] the problem of finding
a combinatorial algorithm for the generic decomposition. In [2] Derksen and Weyman proposed an
elegant and fast algorithm for this.
Denote by SL(α ) the commutator subgroup of GL(α ), SL(α ) = i ∈ Q 0 SL(αi ) ⊆ GL(α ). A natural task
of the Invariant Theory in the quiver setup is to study the GL(α )-semi-invariant functions on R ( Q , α ),
which are also SL(α )-invariant. To be precise,the character group of GL(α ) is isomorphic to Z Q 0 such
χ
that χ ∈ Z Q 0 gives rise to the character χ = a∈ Q 0 , αa >0 deta a and we have
SL(α ) (GL(α ))
k R ( Q , α) = k R ( Q , α) χ , (1)
χ ∈Z Q 0
(GL(α ))
where k[ R ( Q , α )]χ = { f ∈ k[ R ( Q , α )] | g f = χ ( g ) f , ∀ g ∈ GL(α )} is the module of the semi-
.
invariants of the given weight χ . Note also that k[ R ( Q , α )](χGL(α )) = {0} implies χ (α ) = a∈ Q 0 χa αa =
0. Denote by πSL(α ) : R ( Q , α ) → R ( Q , α )//SL(α ) = Spec k[ R ( Q , α )]SL(α ) the categorical quotient.
Recall that Ringel introduced in [8] the Euler bilinear form

α , β
= αa βa − αt ϕ βhϕ , (2)
a∈ Q 0 ϕ∈ Q 1
such that the Tits form can be written as q Q (α ) = α , α

, and proved the formula:
dim Ext(U , V ) = dim Hom(U , V ) − dim U , dim V

. (3)
Schofield introduced in [9] a correspondence between representations and semi-invariants. Namely,

for any representation W there is a determinantal semi-invariant c W such that c W ( V ) = 0 if and only
if dim V , dim W
= 0 and Hom( V , W ) = 0, hence also Ext( V , W ) = 0 by Ringel formula (3). More-
over, the weight of c W is equal − ·, dim W
. Besides, the representations V such that c W ( V ) = 0
constitute an Abelian subcategory closed under homomorphisms, extensions, direct sum and sum-
mands and this subcategory is denoted by ⊥ W . Similarly, the subcategory W ⊥ consists of those V
such that Hom( W , V ) = 0 and Ext( W , V ) = 0.
Let β be a prehomogeneous dimension vector, that is R ( Q , β) contains a dense orbit GL(β) V . In
this case Schofield proved in [9] that the category V ⊥ is isomorphic to the category of representations
of a quiver Σ without oriented cycles and moreover, if W 1 , . . . , W l are the simple objects of V ⊥ ,
then the semi-invariants c W 1 , . . . , c W l are algebraically independent generators of k[ R ( Q , β)]SL(β) . We
revisit Schofield’s proof and find that, together with the Derksen–Weyman algorithm it yields an
algorithm for the dimensions of the simple objects in W ⊥ (see Algorithm 3.6).
Assume from now on that Q has no oriented cycles. In [11] we introduced a class of representa-
tions helping to study the semi-invariants of quiver from the geometrical point of view:
Theorem 1.1. (See [11].) Let V = m1 S 1 + · · · + mt S t ∈ R ( Q , α ) be a decomposition into indecomposable

summands. The following properties of V are equivalent:
(i) the SL(α )-orbit of V is closed in R ( Q , α );

(GL(α ))
(ii) the GL(α )-orbit of V is closed in R ( Q , α ) f , f ∈ k[ R ( Q , α )]χ ;
⊥
(iii) S 1 , . . . , S t are simple objects in W for a representation W .
We called the representations satisfying the equivalent properties of the above theorem locally
semi-simple. By (iii) these representations have the property:
dim Hom( S i , S j ) = δi j . (4)

D.A. Shmelkin / Journal of Algebra 322 (2009) 3997–4010 3999
This property yields an equality Aut( V ) ∼ = GL(m1 ) × · · · × GL(mt ) so the decomposition corre-
sponding to V completely defines the embedding Aut( V ) ⊆ GL(α ). Hence, such a decomposition (we
call them locally semi-simple) yields a conjugate class ( H ) of subgroups in GL(α ) and a stratum
( R ( Q , α )//SL(α ))( H ) consisting of all ξ ∈ R ( Q , α )//SL(α ) such that the automorphism group of the
−1
locally semi-simple representations in πSL (α ) (ξ ) belong to ( H ). In [11] we proved that this is a strat-
ification of R ( Q , α )//SL(α ) by locally closed smooth subsets, and moreover, this is a refinement of
Luna’s stratification of R ( Q , α )//SL(α ) in the sense of [7]. So we call the set of the locally semi-
simple decompositions of α the GL(α )-stratification of R ( Q , α )//SL(α ). Of particular interest are the
open stratum and the corresponding decomposition that we call generic locally semi-simple.
In [3] Derksen and Weyman introduced the notion of quiver Schur sequence and related this notion
with the geometry of the cone generated by the weights of semi-invariants. We prove that the decom-
positions corresponding to the quiver Schur sequences are locally semi-simple (Theorem 2.6). Next,
we get in Theorem 3.9 a complete description of the Luna GL(β)-stratification for prehomogeneous β
in terms of perpendicular categories, which is an invariant-theoretic counterpart of [3, Theorem 5.1]
for this case. Therefore, applying Algorithm 3.6 for perpendicular categories, we can compute all lo-
cally semi-simple decompositions of β and in particular, the generic one (Corollary 3.10). Finally, we
propose Algorithm 4.7 for the generic locally semi-simple decomposition for arbitrary dimension vec-
tor α .
Besides proving theorems and algorithms we implemented a computer program for doing all these
types of calculations, namely, allowing to calculate generic and generic locally semi-simple decompo-
sitions for arbitrary dimension vectors and perpendicular categories for a prehomogeneous vector.
This program is called TETIVA and is available at [12].
2. Schur sequences and locally semi-simple representations
In this section we relate locally semi-simple decompositions of dimension vector to various other
decompositions and start with the σ -stable ones.
Let α be a dimension vector and σ ∈ Z Q 0 be a weight such that σ (α ) = 0. Recall that King in-
troduced in [5] the notion of (semi-)stability of representations of dimension α . Assume that generic
(GL(α ))
representation of dimension α is σ -semi-stable, or, equivalently, k[ R ( Q , α )]nσ = {0} for some
n ∈ Z+ . Then each σ -semi-stable V ∈ R ( Q , α ) has a filtration in the subcategory of σ -semi-stable
representations with the σ -stable factors, that is, Jordan–Hölder factors. So V yields a decomposition
of α into the linear combination of the dimensions of σ -stable representations, and for V generic we
get the so-called σ -stable decomposition of α :
α = m1 α1 + · · · + ms αs . (5)
Note that for Q being a tame quiver and σ the defect the σ -stable representations are the regular
ones and the σ -stable decomposition is Ringel’s canonical one, see [8]. By [11, Proposition 10, Theo-
rem 11], we get:
Proposition 2.1. The σ -stable decomposition is locally semi-simple.
Definition 2.2. For dimension vectors α , β denote by hom(α , β) and ext(α , β) the dimensions of
Hom( A , B ) and Ext( A , B ) for generic A ∈ R ( Q , α ), B ∈ R ( Q , β), respectively. Further, write α ⊥ β if
hom(α , β) = 0 = ext(α , β).
Definition 2.3. A sequence α1 , . . . , αs of dimension vectors is called (right) perpendicular if it consists

of Schur roots and for 1 i < j s holds αi ⊥ α j .
Now we introduce an important notion from [3]:

Definition 2.4. A perpendicular sequence α1 , . . . , αs is called a Schur sequence if

. (GL(αi ))
αi ◦ α j = dim k R ( Q , αi ) − ·,α j
= 1, 1 i < j s. (6)
This sequence is called quiver Schur if additionally α j , αi

0 for i < j.
The notion of quiver Schur sequence can be interpreted in terms of local quiver. The idea of
the latter goes back to [6], where it was applied for semi-simple representations and we used
it in [11] for locally semi-simple ones, too. Moreover, the definition works for any representation
V = m1 S 1 + · · · + mt S t satisfying condition (4) and we define Σ V to be the quiver with t vertices
corresponding to the summands S 1 , . . . , S t and δi j − dim S i , dim S j
arrows from i to j. Note that,
thanks to the condition (4) and the Ringel formula (3), δi j − dim S i , dim S j
= dim Ext( S i , S j ) 0. On
the other hand, the definition of Σ V only depends on the sequence α = (dim S 1 , . . . , dim S t ), not on
the multiplicities, and even not on the summands themselves, provided the homomorphism spaces
are trivial. So it is possible and even more correct to denote the quiver Σα . This local quiver plays
a crucial role in [6] and [11] because for locally semi-simple V the slice representation at V is (by
formula (9) in [11]):

Aut( V ), Ext( V , V ) ∼
= GL(γ ), R (Σα , γ ) , (7)
where γ = (m1 , . . . , mt ) is a dimension vector for Σα . Therefore Luna’s étale slice theorem relates the
local equivariant structure of a neighborhood of V in R ( Q , α ) with that of 0 in R (Σα , γ ).
Proposition 2.5. Let α = (α1 , . . . , αt ) be a sequence of Schur roots.
1. If α is a quiver Schur sequence then hom(αi , α j ) = 0 for any i = j and the quiver Σα has no oriented
cycles except loops.
2. If hom(αi , α j ) = 0 for any i = j and Σα has no oriented cycles except loops, then after a reordering we
have αi ⊥ α j for i < j. If moreover αi is imaginary for at most one i, then after a reordering α becomes
a quiver Schur sequence.
Proof. 1. For i < j, αi ⊥ α j implies hom(αi , α j ) = 0 and ext(αi , α j ) = 0. The latter together
with [10, Theorem 4.1] implies that either hom(α j , αi ) = 0 or ext(α j , αi ) = 0. By Ringel formula,
hom(α j , αi ) − ext(α j , αi ) = α j , αi
0, hence hom(α j , αi ) = 0 as well. Since all non-loop arrows
of Σα go from j to i with j > i, this quiver does not contain non-loop oriented cycles.
2. Each oriented graph having no oriented cycles admits an order such that all arrows go from
bigger to smaller indices; forget the loops of Σα and fix such an order. Since hom(αi , α j ) = 0 for any
1 i = j t and ext(αi , α j ) = 0 for any i < j in our order, we have αi ⊥ α j . If moreover, at least one
from αi , α j is real then αi ◦ α j = 1 by [3, Lemma 4.2]. 2
Therefore the quiver Schur sequences are very close to the sequences with trivial mutual ho-
momorphisms and without oriented cycles. Of course, not each locally semi-simple representation
meets the latter condition: for instance take a tame quiver as Q and the sum of the simple non-
homogeneous regular representations over an orbit of Coxeter functor; then this representation is
locally semi-simple by [11, Proposition 20] but the local quiver is an oriented cycle by [11, Proposi-
tion 21].
Theorem 2.6. For a quiver Schur sequence α = (α1 , . . . , αt ) and a tuple (m1 , . . . , mt ) the decomposition
β = m1 α1 + · · · + mt αt is locally semi-simple.
Proof. By Theorem 1.1 the fact that the decomposition is locally semi-simple does not depend on
the multiplicities so we may assume m1 = m2 = · · · = mt = 1. Then we apply Theorem 5.1 from [3].
Denote by Σ( Q , β) the set of weights of the semi-invariants in k[ R ( Q , α )]SL(α ) . Theorem 5.1 asserts
that the cone R+ Σ( Q , β) has a face F = R+ Σ( Q , β) ∩ {σ ∈ R Q 0 | σ (α1 ) = · · · = σ (αt ) = 0}. Moreover,
the theorem guarantees that for σ from the relative interior of F , β = α1 + · · · + αt is the σ -stable
decomposition of β . Now the assertion follows from Proposition 2.1. 2
3. Prehomogeneous dimension vectors
Recall that a dimension vector β is called prehomogeneous if R ( Q , β) contains a dense GL(β)-orbit.

By [4] this is equivalent to the generic decomposition of β containing only real Schur roots. For this
particular case we are able to calculate the Luna stratification completely.
Proposition 3.1. If β is prehomogeneous and α is a sequence of Schur roots such that β = m1 α1 + · · · + mt αt

then this decomposition is locally semi-simple if and only if α is a quiver Schur sequence up to order.
Proof. By Theorem 2.6 and Proposition 2.5 we only need to prove that if the decomposition is locally
semi-simple then Σα does not contain oriented cycles (in particular, the absence of loops means that
all summands are real roots). Indeed, take a locally semi-simple representation W = m1 S 1 + · · · + mt S t
corresponding to the decomposition. Then the orbit GL(β) W is closed in an open affine neighbor-
hood R 0 ⊆ R ( Q , β), R 0 W and by formula (7) the slice representation at W is isomorphic to
(GL(γ ), R (Σα , γ )). The étale slice theorem yields an étale morphism of R (Σα , γ )//GL(γ ) to R 0 //GL(β).
Since GL(β) has an open orbit in R ( Q , β), it is contained in R 0 and so R 0 //GL(β) is a point. Conse-
quently, R (Σα , γ )//GL(γ ) is a point, hence, Σα does not have oriented cycles. 2
The way we compute the locally semi-simple decompositions of prehomogeneous dimension vec-
tors is based on Schofield’s theorem:
Theorem 3.2. (See [9, Theorem 2.5].) If β is prehomogeneous and W = m1 S 1 + · · · + mt S t is the decompo-
sition into indecomposable summands of a representation from the dense orbit, then W ⊥ is isomorphic to the
category of representations of a quiver with n − t vertices and without oriented cycles, where n = | Q 0 |. The
same is true for ⊥ W .
We want to generalize this theorem and to give an algorithm for calculation of the perpendicu-
lar category based on the original proof. The above theorem can be reformulated as follows: there
are n − t representations R 1 , . . . , R n−t , which are all non-isomorphic simple objects in W ⊥ . By Theo-
rem 1.1 R = R 1 + · · · + R n−t is locally semi-simple and the above theorem additionally asserts that the
local quiver Σ of R does not have oriented cycles. A particular case of the theorem is when β is a real
Schur root, and actually the proof in [9] deduces the general case from this particular one, where the
proof is based on the notion of projective and injective representations that we recall following [9].
For i , j ∈ Q 0 denote by [i , j ] the k-vector space on the basis of oriented paths from i to j in Q .
For a ∈ Q 0 consider the representation P a such that P a (i ) = [a, i ], i ∈ Q 0 , and for any arrow ϕ ∈ Q 1
the map P a (ϕ ) takes a path T from a to t ϕ to the concatenation T ϕ , which is a path a to hϕ .
Dually, consider the representation I a such that I a (i ) = [i , a]∗ , i ∈ Q 0 , and I a (ϕ ) is the dual map to
the natural one from [hϕ , a] to [t ϕ , a]. These representations have nice properties with respect to the
homomorphisms and extensions: for any representation V of Q hold

∗
Hom( P a , V ) ∼
= V (a), Hom( V , I a ) ∼
= V (a) , Ext( P a , V ) = 0, Ext( V , I a ) = 0. (8)
Theorem 3.3. Let β be a real Schur root for a quiver Q with n vertices and without oriented cycles and let W
be a generic representation of dimension β . Then:
1. If β = dim P a for some a ∈ Q 0 , then the simple objects of the category W ⊥ are all the simple representa-
tions of Q but S a . Otherwise, the dimensions of n − 1 projective objects of W ⊥ are the indecomposable
summands of the generic decomposition for dimension vectors dim P a − β, dim P a

β , where a runs
over Q 0 .
2. If β = dim I a for some a ∈ Q 0 , then the simple objects of the category ⊥ W are all the simple representa-
tions of Q but S a . Otherwise, the dimensions of n − 1 injective objects of ⊥ W are the indecomposable
summands of the generic decomposition for dimension vectors dim I a − dim I a , β
β , where a runs
over Q 0 .
Proof. We prove assertion 1, the proof for 2 being similar. First of all, for β = dim P a the asser-
tion follows from formulae (8). In the opposite case, we apply the proof of Theorems 2.3 and 3.1
from [9]. From the proof of Theorem 2.3 we learn that all indecomposable projective objects in W ⊥
can be
obtained as the indecomposable summands of a generic extension of sW by Λ, where
Λ = a∈ Q 0 P a , s = dim Ext( W , Λ). For each individual P = P a Theorem 3.1 considers a generic exact
sequence 0 → P → P ∼ → sW → 0, where s = dim Ext( W , P ) and states that P ∼ is projective in W ⊥ .
Then mutual extensions of the indecomposable summands of P ∼ are trivial in W ⊥ by (8), hence,
trivial in Mod( Q ), because W ⊥ is closed under extensions. Therefore, P ∼ is generic in its dimension
and the dimensions of the indecomposable summands of P ∼ are the summands of the generic de-
composition of dim P ∼ = dim P + dim Ext( W , P )β = dim P − β, dim P
β , the latter equality following
from Hom( W , P ) = 0. 2
The above theorem yields a fast algorithm as follows:
Algorithm 3.4. Right perpendicular category of a Schur root

input: a quiver Q with n vertices and without oriented cycles,
a real Schur root γ
output: n − 1 dimensions of the simple objects in W ⊥ such that
GL(γ ) W is dense in R ( Q , γ ).
1. Calculate the dimensions ρ1 , . . . , ρn of indecomposable projectives.
If γ = ρ j , then return ε1 , . . . ε ĵ . . . , εn .
2. Loop on i = 1, . . . , n
Calculate the generic decomposition of ρi − γ , ρi
γ
Add each summand to an array provided it is not yet there
After step 2 we must have distinct summands β1 , . . . , βn−1 in the array
3. Color the entries 1, . . . , n − 1 white, the number of black entries b = 0
Loop while b < n − 1
Loop on j = 1, . . . , n − 1
If jth entry is white and for each other white entry k, βk , β j
= 0, then
remember this entry j, which is going to be black, break the loop
Set next simple dimension α j = β j
Loop on k = 1, . . . , n − 1
If kth entry is black then α j = α j − βk , β j
αk
b = b + 1.
4. return α1 , . . . , αn−1 .
Proof. After Theorem 3.3 only step 3 of the algorithm needs to be explained. In that step we do the
inverse to step 1, i.e., we recover the dimensions of the simple objects from those of indecompos-
able projectives. The idea of the step is that, though we do not know the quiver Σ of the simple
objects, we know that the Euler form on Σ is the same that inherited from Q . But by formulae (8)
ext(βi , β j ) = 0, hence again by those formulae, βi , β j
with respect to the Euler form of Σ is equal
to the number of paths from j to i in Σ . So the first vertex becoming black is a sink of Σ , hence,
the corresponding projective is simple. Furthermore, each next vertex becoming black is the sink of Σ
with removed black vertices, hence the corresponding projective is the simple plus the sum over black
vertices of the already obtained simple dimensions multiplied by the number of paths to there. This
completes the proof. 2
Remark 3.1. Dualizing Algorithm 3.4 as in Theorem 3.3, we get one for the left perpendicular category.
Now we generalize Theorem 3.2 as follows:
Theorem 3.5. If W = m1 S 1 + · · · + mt S t , where S 1 , . . . , S t are Schur indecomposable summands and

(dim S 1 , . . . , dim S t ) is a perpendicular sequence of real Schur roots, then there is a sequence α =
(α1 , . . . , αn−t ), n = | Q 0 | of real Schur roots such that the corresponding indecomposable representations
are all simple objects in W ⊥ and Σα has no oriented cycles. The same is true for ⊥ W .
Proof. We just present an algorithm for calculation of α based on Algorithm 3.4:
Algorithm 3.6.
a perpendicular sequence (dim S 1 , . . . , dim S t ) of real Schur roots
output: dimensions α1 , . . . , αn−t of the simple objects in W ⊥
Loop on i = 1, . . . , t
We have dimensions α1 , . . . , αn+1−i of the simple objects of the current
category. For i = 1 the category is the whole of Mod( Q ).
Calculate the quiver Σi of the current category by means of Euler form
Expand dim S i as a linear combination of α1 , . . . , αn+1−i : the coefficients
yield a dimension vector γi for Σi
Find the dimensions α1 , . . . , αn−i of the simple objects in the right
perpendicular category to γi for Σi
calculate new α1 , . . . , αn−i as the linear combinations of old
α1 , . . . , αn+1−i with coefficients from α1 , . . . , αn−i
return α1 , . . . , αn−t .
The key idea of this clear algorithm is that we obtain the perpendicular category to W as a se-
quence of subcategories, where the next is obtained as the perpendicular category to a Schur root.
The algorithm for the left perpendicular category is similar, we only go from S t to S 1 . 2
The above theorem makes possible to introduce an operation ⊥ mapping a real (i.e., consisting
of real roots) perpendicular sequence α of t roots to real perpendicular sequences α ⊥ and ⊥ α of
n − t roots being the sequence of dimensions of simple objects in the right and left perpendicular
categories, respectively. The first natural application of Algorithm 3.6 is given by the following inter-
pretation of the result from [9]:
Theorem 3.7. Let β be a prehomogeneous dimension vector and β = (β1 , . . . , βt ) be a perpendicular sequence
of the summands for the generic decomposition of β . For each member γi ∈ β ⊥ pick a generic T i ∈ R ( Q , γi ).
Then the determinantal semi-invariants c γi = c T i , i = 1, . . . , n − t, constitute an algebraically independent
system of generators for k[ R ( Q , β)]SL(β) .
For any real perpendicular sequence α the sequence α ⊥ is a real quiver Schur sequence, so if α is
not a quiver Schur sequence, then ⊥ (α ⊥ ) is different from α (cf. Corollary 3.10). Otherwise we have:
Proposition 3.8. If α is a real quiver Schur sequence, then α =⊥ (α ⊥ ).
Proof. Denote α ⊥ by β = (β1 , . . . , βn−t ) and ⊥ β by γ = (γ1 , . . . , γt ). Each αi ∈ α has the property
αi ⊥ β j for each β j ∈ β , hence, αi decomposes as αi = ρi1 γ1 + · · · + ρit γt with non-negative integers
ρi1 , . . . , ρit . By [11, Proposition 13] the sequence ρ = (ρ1 , . . . , ρt ) is a real quiver Schur sequence for
the quiver Σγ . Since this quiver has t vertices, we have ⊥ ρ is empty and therefore, there are no semi-
invariants non-vanishing on the representation O corresponding to the decomposition ρ1 + · · · + ρt .
However, by Theorem 2.6 O is locally semi-simple, hence O is the zero point in R (Σγ , dim O ). In
other words, ρ1 , . . . , ρt are the dimensions of simple representations, so α = γ up to order. 2
Recall that the set of all locally semi-simple decompositions of a dimension vector β is the same
as the Luna GL(β)-stratification of R ( Q , β)//SL(β). For β prehomogeneous we are able to describe this
stratification completely:
Theorem 3.9. Let β be a prehomogeneous dimension vector and β = (β1 , . . . , βt ) be a perpendicular sequence
of the summands for the generic decomposition of β .
1. There is a bijection of the Luna GL(β)-stratification of R ( Q , β)//SL(β) with the set of subsequences in
β ⊥ = (γ1 , . . . , γn−t )
γ ⊆ β ⊥ → β = m1 ρ1 + · · · + ms ρs , (ρ1 , . . . , ρs ) =⊥ γ , m1 , . . . , ms ∈ Z+ .
2. The stratum corresponding to γ is {ξ ∈ R ( Q , β)//SL(β) | c γi (ξ ) = 0 ⇔ γi ∈ γ }.

3. This bijection preserves the order on the sets: if γ1 ⊆ γ2 , then the stratum corresponding to γ1 is contained
in the closure of that for γ2 .
Before proving the theorem we state an important
Corollary 3.10. The generic locally semi-simple decomposition of β is β = m1 β1 + · · · + mt βt , where
β =⊥ β ⊥ .
Proof of Theorem 3.9. 1. First of all we show that the map is well defined, i.e., there is a unique
decomposition of β in ⊥ γ and this decomposition is locally semi-simple. By definition of β ⊥ we have
for each βi ∈ β and each γ j ∈ γ , βi ⊥ γ j . Then by definition of ⊥ γ each βi is a Z+ -linear combination
of elements of ⊥ γ , hence, β is. This decomposition is locally semi-simple by Theorem 1.1 and is
unique, because ⊥ γ is clearly linear independent. To prove that our correspondence is a bijection
we only need to check that each locally semi-simple decomposition β = p 1 ρ1 + · · · + p s ρs can be
obtained this way. First of all by Proposition 3.1 we may assume that ρ = (ρ1 , . . . , ρs ) is a real quiver
Schur sequence. In particular, there is a unique representation V ⊆ R ( Q , β) corresponding to this
decomposition, up to isomorphism. Then the sequence ρ ⊥ is a subsequence in β ⊥ . Indeed, each semi-
invariant that does not vanish on V does not vanish generically on R ( Q , β) so each element of ρ⊥
is presented as a Z+ -linear combination over β ⊥ . Then ρ ⊥ can be identified with the subsequence
{i ∈ {1, . . . , n − t } | c ( V ) = 0}. So we may set γ = ρ ⊥ and recover ρ as ⊥ γ by Proposition 3.8.
γi
2. The locally semi-simple representations over the stratum corresponding to γ constitute one or-
bit GL(β) V , where V corresponds to the decomposition β = m1 ρ1 + · · · + ms ρs , ρ =⊥ γ . By definition
of ⊥ γ , c γi ( V ) = 0 for γi ∈ γ . For any j = 1, . . . , n − t it follows from the properties of determinantal
semi-invariants that c γ j ( V ) = 0 implies ρi , γ j
= 0 for each i = 1, . . . , s. Since γ1 , . . . , γn−t are linear
independent, the number of semi-invariants c γ j non-vanishing on V is less than or equal to the codi-
mension of the common kernel of the corresponding forms ·, γ j
on Q Q 0 . Since ρ1 , . . . , ρs are linear
independent and belong to that kernel, its codimension cannot be more than n − s, so c γ j ( V ) = 0 for
γ j ∈/ γ . Assertion 3 clearly follows from 2. 2
Remark 3.2. This assertion is closely related with [3, Theorem 5.1] because quiver Schur sequences are
in bijection with the Luna GL(β)-strata by Proposition 3.1 and the subsequences in β ⊥ are in bijection
with the faces of the cone R+ Σ( Q , β), which is simplicial in this case.
4. Generic locally semi-simple decomposition
There are cases, where the only locally semi-simple representation in R ( Q , α ) is the trivial semi-
simple one. We have a criterion for this, as follows:
Proposition 4.1. Let α = m1 γ1 + · · · + mt γt be the generic decomposition with γi = γ j for i = j. Then

k[ R ( Q , α )]SL(α ) = k if and only if | Q 0 | = t.
Proof. Assume that k[ R ( Q , α )]SL(α ) = k. By [1] this algebra is generated by determinantal semi-
invariants, in particular, there is W ∈ R ( Q , δ) such that 0 ≡ c W ∈ k[ R ( Q , α )]SL(α ) , and moreover,
c W does not vanish on the summands of generic representation, in particular, γi , δ
= 0, i = 1, . . . , t.
On the other hand, by Remark 4.6 and Corollary 4.12 from [3], γ1 , . . . , γt are linear independent; since
these are contained in a proper Q-vector subspace of Q Q 0 , we conclude t < | Q 0 |.
Conversely, assume k[ R ( Q , α )]SL(α ) = k. Since SL(α ) has no non-trivial characters and R ( Q , α ) is
a vector space, this assumption is equivalent to SL(α ) acting with a dense orbit on R ( Q , α ), hence,
α being a prehomogeneous dimension vector. But (γ1 , . . . , γt )⊥ must be empty because there are no
non-trivial semi-invariants. Then by Theorem 3.2, t = | Q 0 |. 2
Let α = m1 β1 + · · · + mt βt be a locally semi-simple decomposition. For a root β consider the

β -piece of the decomposition, i.e., the sum of all summands being multiples of β . If β is real, then
R ( Q , β) contains a unique isomorphism class of indecomposable representations so the β -piece is
just mβ with m arbitrary. If β is imaginary, and the β -piece is of the form mi 1 β + · · · + mik β , we say
that β occurs k times with multiplicities mi 1 , . . . , mik .
Definition 4.2. We call a locally semi-simple decomposition almost loopless if the multiplicity of each
imaginary Schur root β is 1 and if q Q (β) < 0, then this root occurs one time.
Theorem 4.3. A locally semi-simple decomposition α = p 1 δ1 + · · · + p s δs is generic if and only if it is almost

loopless, the local quiver Σδ has no oriented cycles except for the loops, and the number of different roots in δ
is equal to that for the generic decomposition of α .
Proof. By [11, Theorem 19] the multiplicities of imaginary roots in generic locally semi-simple de-
composition are equal to 1. Moreover, if q Q (β) < 0, then by [10, Theorem 3.7] nβ is again a Schur
root, so the generic locally semi-simple decomposition is almost loopless. By formula (7) the slice
representation corresponding to the decomposition is (GL(ρ ), R (Σδ , ρ )) and the condition that the
decomposition is almost loopless implies that the loops of Σδ exist only at vertices with dimension 1.
By [11, Corollary 5] the decomposition is generic locally semi-simple if and only if k[ R (Σδ , ρ )]SL(ρ )
is generated by the GL(ρ )-invariant submodule in R (Σδ , ρ ), which is the sum of scalar endomor-
phisms corresponding to the loops of Σδ . Removing the loops of Σδ we define a quiver Σ such
that the GL(ρ )-modules R (Σδ , ρ ) and R (Σ, ρ ) only differ by an invariant factor, hence, the generic
and the generic locally semi-simple decompositions of ρ with respect to Σ are the same as for Σδ .
Thus the decomposition is generic locally semi-simple if and only if k[ R (Σ, ρ )]SL(ρ ) = k. If Σ has
oriented cycles, then the latter is false, in the opposite case by Proposition 4.1 the decomposition
is generic locally semi-simple if and only if the number of different roots in the generic decompo-
sition for R (Σ, ρ ) (or, equivalently, for R (Σδ , ρ )) is equal to s. By [11, Proposition 14], the generic
decomposition of R ( Q , α ) is the result of substituting δ1 , . . . , δs instead the standard basis into the
generic decomposition of R (Σδ , ρ ), so we are done for the case, where δ1 , . . . , δs are distinct. Note
that, since the decomposition is almost loopless by assumption, equal summands δi 1 = · · · = δik = δ
are isotropy roots, q Q (δ) = 0, with mi p = ρi p = 1 for p = 1, . . . , k. So the above property of the generic
decomposition for R (Σδ , ρ ) implies that it has summands ρ 1 , . . . , ρ k such that ρipq = δ pq . Besides, the
decomposition α = p 1 δ1 + · · · + p s δs is symmetric by i 1 , . . . , ik , hence so is the set ρ 1 , . . . , ρ k . Thus
the summands ρ 1 , . . . , ρ k yield equal dimension vectors for Q and this completes the proof. 2
Remark 4.1. It follows from the last part of the proof that if some (isotropic) root δ occurs k times
in the generic locally semi-simple decomposition, then some (isotropic) root occurs k times in the
generic decomposition.
In what follows we will intensively use the following well-known fact.
Proposition 4.4. An exact sequence of homomorphisms 0 → U → V → W → 0 for representations of Q

yields for any representation X exact sequences:
0 → Hom( W , X ) → Hom( V , X ) → Hom(U , X )
→ Ext( W , X ) → Ext( V , X ) → Ext(U , X ) → 0, (9)
0 → Hom( X , U ) → Hom( X , V ) → Hom( X , W )
→ Ext( X , U ) → Ext( X , V ) → Ext( X , W ) → 0. (10)
Our idea of algorithm for generic locally semi-simple decomposition for a dimension vector α
is similar to that for generic decomposition from [2]. That algorithm works with perpendicular se-
quences and glues and permutes two items each time there is a non-trivial extension between them.
We proceed as follows: starting from the generic decomposition, we transform it into the generic
locally semi-simple one by slightly splitting summands by each other on the fact of non-trivial ho-
momorphism. More precisely, we do it only when at least one of the items is imaginary, for the
homomorphisms between real summands we apply Corollary 3.10. The most simple step of this pro-
cedure is: having Schur roots α , β with α ⊥ β, ext(β, α ) = 0 but hom(β, α ) = 0, we “factorize” the
imaginary root by the real one to get an imaginary Schur root with trivial homomorphism spaces
with the real root:
Proposition 4.5. Let α and β be Schur roots such that ext(α , β) = 0 = ext(β, α ).
1. If both α and β are imaginary, then hom(α , β) = 0 = hom(β, α ).

2. In any case either hom(α , β) = 0 or hom(β, α ) = 0.
3. Assume that hom(α , β) = 0, dim hom(β, α ) = p > 0.
A: If α is imaginary, then for generic A ∈ R ( Q , α ), B ∈ R ( Q , β) there is an exact sequence of ho-
momorphisms: 0 → p B → A → C → 0, C is Schurian, q Q (dim C ) = q Q (α ), Hom(C , B ) = 0,
Hom( B , C ) = 0, Ext( B , C ) = 0.
B: If β is imaginary, then for generic A ∈ R ( Q , α ), B ∈ R ( Q , β) there is an exact sequence of ho-
momorphisms: 0 → C → B → p A → 0, C is Schurian, q Q (dim C ) = q Q (β), Hom( A , C ) = 0,
Hom(C , A ) = 0, Ext(C , A ) = 0.
Moreover, in both cases A, B, if γ ⊥ α and γ ⊥ β , then γ ⊥ dim C ; if α ⊥ γ and β ⊥ γ , then dim C ⊥ γ .
Proof. Assertions 1 and 2 follow from Theorems 4.1 and 2.4 of [10], respectively. We prove 3A the
proof for 3B being similar. Consider the decomposition ρ = α + β ; the conditions ext(α , β) = 0 =
ext(β, α ) imply that this is a generic one. Then by Theorem 4.3 the generic locally semi-simple de-
composition of ρ has two summands, ρ = m1 γ + m2 δ and moreover, by Remark 4.1, if γ is imaginary,
then m1 = 1 and the same for δ . We want to find this decomposition and its local quiver Σ .
Assume that both summands γ , δ are imaginary. Then, either Σ does not have arrows different
from loops, hence ρ = γ + δ is the generic decomposition, which is impossible because it is different
from ρ = α + β . Or Σ does have arrows, but does not have oriented cycles by Theorem 4.3. Then,
by [2, Corollary 12] ρ = γ + δ is itself a Schur root, a contradiction with the generic decomposition
ρ = α + β.
Assume that both γ , δ are real. Then the generic locally semi-simple representations in
R ( Q , ρ ) constitute one GL(ρ )-orbit, hence, GL(ρ ) acts with a dense orbit on R ( Q , ρ )//SL(ρ ) and
k( R ( Q , ρ ))GL(ρ ) = k( R ( Q , ρ )//SL(ρ ))GL(ρ ) = k. So ρ is prehomogeneous in contradiction with the

generic decomposition ρ = α + β .
So the decomposition has the form ρ = aγ + δ , where δ is an imaginary Schur root and γ is real.
As above, Σ has r > 0 arrows of the same orientation between two vertices and besides 1 − q Q (δ)
loops at the vertex corresponding to δ . Consider the generic decomposition of (a, 1) on Σ , it must
have a form (a, 1) = (a − q)(1, 0) + (q, 1), q > 0. One can easily calculate the values of the Euler
form: (q, 1), (1, 0)
Σ and (1, 0), (q, 1)
Σ are equal q and q − r up to transposition. By [11, Propo-
sitions 14, 15] these values are equal to α , β
Q and β, α
Q , which are 0 and p. So we have,
q = r = p, and β = (1, 0) = γ , α = ( p , 1) = p γ + δ . Finally, the generic locally semi-simple decom-
position of ρ is ρ = ( p + 1)β + α − p β . In particular, α − p β is a Schur root, hom(β, α − p β) = 0,
and β, α − p β
= β, α
− pq Q (β) = 0, so β ⊥ α − p β . Hence, by [10, Theorem 3.3] generic rep-
resentation of dimension α has a subrepresentation of dimension p β , so isomorphic to p B be-
cause generic. Thus we have the claimed exact sequence. Write: q Q (α − p β) = α − p β, α − p β
=
q Q (α ) + p 2 − p ( α , β
+ β, α
) = q Q (α ).
The rest of the assertion will be deduced from formulae (9), (10) for U = p B , V = A , W = C . Since
Ext( V , U ) = 0 and Hom( V , U ) = 0, (10) with X = V yields Hom( V , W ) ∼ = Hom( V , V ) = k. Then (9)
with X = W yields Hom( W , W ) = k, so C is Schurian. Next, Hom( V , U ) = 0 and (9) with X = U yield
Hom( W , U ) = 0, hence, Hom(C , B ) = 0. Finally, apply (10) with X = U and note that Ext(U , V ) = 0
by assumption, Ext(U , U ) = 0 because β is real, Hom(U , U ) ∼ = Hom(U , V ) ∼
2
= k p ; then Ext(U , W ) and
Hom(U , W ) vanish. That dim C is perpendicular to any dimension vector γ , which is perpendicular
to α and β follows directly from formulae (9), (10) with X being generic representation of dimen-
sion γ ; with such a choice of X four members in the exact sequence vanish, hence all six vanish. 2
Definition 4.6. Assume that α , β are the subsequent members of a decomposition. If α , β meet the
conditions of Proposition 4.5.3A (resp. Proposition 4.5.3B), we call the replacement of α , β by β ,
α − p β (resp. β − p α , α ) pushing α to the right (resp. pushing β to the left). This operation also includes
the obvious recalculation of multiplicities on the decomposition. We also may apply both terms to
the transposition of the members α , β (even when both are real) such that α ⊥ β and β ⊥ α .
Now we present our algorithm for the generic locally semi-simple decomposition of a dimension
vector α with a given generic decomposition α = m1 γ1 + · · · + mt γt such that γ is a Schur sequence.
Recall that the decomposition of this sort is the result of the algorithm from [2].
Algorithm 4.7.
a Schur sequence (γ1 , . . . , γt ) of the summands
for the generic decomposition, multiplicities (m1 , . . . , mt )
output: a quiver Schur sequence (α1 , . . . , αt ) of the summands
for the generic locally semi-simple decomposition,
multiplicities ( p 1 , . . . , pt )
>> First stage:
for i = 2, . . . , t
Set j = i
while γ j is imaginary and γ j −1 is real
push γ j to the left and j = j − 1
>> Result of the first stage: first s t members of γ are imaginary,
>> last t − s are real.
>> Second stage:
Replace the subsequence γ = (γs+1 , . . . , γt ) by ⊥ γ ⊥
>> Result of the second stage: hom(γi , γ j ) = 0 for i , j > s
>> Third stage:
while there is 1 i < j t with γ j , γi
> 0
guarantee that the segment [i , j ] is minimal with such a property
transfer γ j to position i + 1
push γi to the right
>> Result of the third stage: the sequence γ is now a quiver Schur sequence.
return the current γ1 , . . . , γt and (m1 , . . . , mt ).
Now we are going to prove the algorithm.
Proposition 4.8. During the first stage of the algorithm the sequence γ remains perpendicular and each time
when γ j is imaginary and γq is real ext(γ j , γq ) = 0.
Proof. By Proposition 4.5 the sequence γ remains perpendicular after pushing γ j to the left pro-
vided γ was perpendicular before it. We need to prove additionally that C in the exact sequence
from Proposition 4.5.3B has the property Ext(C , D ) = 0 for D being generic representation of dimen-
sion γq , q < j, provided γq is real. When we start pushing to the left this γ j the property holds,
because real roots do not change within the first stage. On each push of γ j we have Ext( B , D ) = 0 by
induction, so applying (9), we get Ext(C , D ) = 0. 2
Proposition 4.9. After the first stage we have γi ⊥ γ j for 1 i = j s.
Proof. Assume that i < j. Then γi ⊥ γ j follows from the previous proposition and hom(γ j , γi ) = 0
follows from [10, Theorem 4.1], because both γi and γ j are imaginary. Assume that ext(γ j , γi ) > 0.
Then we can apply the algorithm for generic decomposition from [2] to the perpendicular sequence γ .
From that algorithm follows that in such a situation we can glue together γi and γ j and get a perpen-
dicular sequence with t − 1 members and, continuing the algorithm, we get the generic decomposition
with less than t members. Contradiction. 2
Proposition 4.10. After the second stage of the algorithm the sequence γ remains orthogonal and
hom(γi , γ j ) > 0 implies i > s, j s.
Proof. Before the second stage γ consists of two segments, (γ1 , . . . , γs ) and γ = (γs+1 , . . . , γt ). Both
segments are orthogonal sequences with trivial Ext spaces, the first segment by the previous propo-
sition and γ because it is a subsequence in the original γ . Consider two dimension vectors ρ1 , ρ2
being the linear combinations of the two segments with the multiplicities. The fact that γ is orthog-
onal is equivalent to ρ1 ⊥ ρ2 , that is, generic representation R 1 of dimension ρ1 is perpendicular
to that in dimension ρ2 . The second stage consists in replacing the generic decomposition for ρ2
by the generic locally semi-simple one. Therefore γ remains orthogonal because R 1 is perpendicular
to generic locally semi-simple representation of dimension ρ2 otherwise the determinantal semi-
invariant defined by R 1 vanishes on R ( Q , ρ2 ). The fact about Hom-spaces follows from the previous
proposition and the feature of generic locally semi-simple decompositions. 2
The third stage of the algorithm is based on the following
Lemma 4.11. If α , β , γ is a perpendicular sequence of Schur roots such that α is imaginary and
hom(γ , α ) > 0, then ext(γ , β) = 0.
Proof. Assume ext(γ , β) > 0. Pick generic representations U ∈ R ( Q , α ), V ∈ R ( Q , β), W ∈ R ( Q , γ )

p
and consider a non-split exact sequence 0 → V → X −→ W → 0. Then X is Schurian (see the
proof of [3, Corollary 12]) and U ⊥ V , U ⊥ W together with (10) imply U ⊥ X . By [10, Theo-
rem 4.1], hom(γ , α ) > 0 implies ext(γ , α ) = 0, so we may apply Proposition 4.5 and, since α is
imaginary, we get γ is real and we are in the case 4.5.3A. In particular, any non-trivial homomor-
phism h ∈ Hom( W , U ) is not surjective. Then the composition hp ∈ Hom( X , U ) is also non-trivial,
because p is surjective, and is not surjective, because h is not. The kernel of hp contains the im-
age of V so hp is neither injective nor surjective. This contradicts to [10, Lemma 2.3], because X is
Schurian and Ext(U , X ) = 0. 2
Proposition 4.12. During the third stage of the algorithm assume γ j , γi

> 0:
1. If [i , j ] is minimal with the property and γi is imaginary, then we have γ j , γk

= 0 for i < k < j.
2. γ j is real and γi is imaginary.
3. The third stage of the algorithm finishes after finitely many steps.
Proof. First of all, for any i < j holds ext(γi , γ j ) = 0 and by [10, Theorem 4.1] either ext(γ j , γi ) = 0
or hom(γ j , γi ) = 0, and the latter is the case if both γi and γ j are imaginary. So either γ j , γi
< 0
and in this case ext(γ j , γi ) > 0 and hom(γ j , γi ) = 0, or γ j , γi
> 0 and in this case ext(γ j , γi ) = 0
and hom(γ j , γi ) > 0, or else γ j , γi
= 0 and in this case ext(γ j , γi ) = 0 = hom(γ j , γi ).
1. Applying Lemma 4.11 to the sequence γi , γk , γ j , i < k < j, we get ext(γ j , γk ) = 0. So we have
γ j , γk
0 and γ j , γk
> 0 contradicts to the minimality of [i , j ].
2. In the third stage we do not change the real roots, only permute them, hence, exactly one of γi
and γ j is imaginary. We prove that γ j is real and γi is imaginary applying induction. At the beginning
each imaginary is to the left of each real. Then we apply the step of the third type to a minimal
segment [i , j ] and have γi imaginary by induction. Then by 1 γ j is perpendicular from both sides
to γk , i < k < j , so the property remains true for all pairs containing γ j . As for the imaginary root
being the result of pushing γi to the right, it is given by Proposition 4.5.3A for α = γi , β = γ j and
we have the exact sequence 0 → p B → A → C → 0 from Proposition 4.5.3A. Taking X to be a generic
representation of dimension γl , l < i , we have Hom( A , X ) = 0, because γi , γl
0 by induction.
Then by (9), Hom(C , X ) = 0, so hom(γi − p γ j , γl ) = 0, hence, γi − p γ j , γl
0 and the property
remains true.
3. Each step of the stage decreases the total dimension of an imaginary member of γ so we cannot
do it infinitely many times. 2
Theorem 4.13. Algorithm 4.7 yields the generic locally semi-simple decomposition.
Proof. The above propositions convinced us that after finitely many steps we obtain a decomposition
with γ being a perpendicular sequence of Schur roots such that hom(γi , γ j ) = 0 if i = j. We also
claim that this is a Schur sequence, that is, γi ◦ γ j = 1 for i < j. Indeed this was true for the starting
sequence and by [3, Lemma 4.2] this needs to be checked only for γi and γ j being imaginary. We now
show that the property is preserved by any pushing of the imaginary root. So assume that γi and γ j
are imaginary γk is real, i < k and we replace γ j by γ j − p γk . Then the vector space of the semi-
invariants on R ( Q , γi ) with the weight − ·, γ j − p γk
is embedded to that of the weight − ·, γ j
by
multiplying with a semi-invariant of the weight − ·, p γk
. So γi ◦ γ j − p γk > 1 would imply γi ◦ γ j > 1.
A similar argument works for pushing γi so we proved that the output of the algorithm is a quiver
Schur sequence. Hence, by Theorem 2.6 the resulting decomposition is locally semi-simple and by
Theorem 4.3 this is the generic locally semi-simple decomposition. 2
References
[1] H. Derksen, J. Weyman, Semi-invariants of quivers and saturation for Littlewood–Richardson coefficients, J. Amer. Math.
Soc. 13 (2000) 467–479.
[2] H. Derksen, J. Weyman, On the canonical decomposition of quiver representations, Compos. Math. 133 (2002) 245–265.
[3] H. Derksen, J. Weyman, The combinatorics of quiver representations, preprint, arXiv:math.RT/0608288.
[4] V. Kac, Infinite root systems, representations of graphs, and invariant theory, II, J. Algebra 78 (1982) 141–162.
[5] A.D. King, Moduli of representations of finite dimensional algebras, Quart. J. Math. Oxford Ser. (2) 45 (1994) 515–530.
[6] L. Le Bruyn, C. Procesi, Semisimple representations of quivers, Trans. Amer. Math. Soc. 317 (1990) 585–598.
[7] D. Luna, Slices étales, Bull. Soc. Math. France (1973) 81–105.
[8] C.M. Ringel, Rational invariants of the tame quivers, Invent. Math. 58 (1980) 217–239.
[9] A. Schofield, Semi-invariants of quivers, J. London Math. Soc. 43 (1991) 383–395.

[10] A. Schofield, General representations of quivers, Proc. London Math. Soc. (3) 65 (1992) 46–64.
[11] D.A. Shmelkin, Locally semi-simple representations of quivers, Transform. Groups 12 (2007) 153–173.
[12] D.A. Shmelkin, TETIVA a computer program available at http://www.mccme.ru/~mitia.
Journal of Algebra
Classification of semifields of order 64

I.F. Rúa a,∗,1 , Elías F. Combarro b,2 , J. Ranilla b,2
a
Departamento de Matemáticas, Universidad de Oviedo, Spain
b
Artificial Ingelligence Center, University of Oviedo, Spain
Article history: A finite semifield D is a finite nonassociative ring with identity

Received 25 August 2008 such that the set D ∗ = D \ {0} is closed under the product. In this
Available online 2 April 2009 paper we obtain a computer-assisted description of all semifields
Communicated by William M. Kantor
of order 64, which completes the classification of finite semifields
Keywords:
of order at most 125.
Finite semifield © 2009 Elsevier Inc. All rights reserved.
Finite division ring
Projective planes
1. Introduction
A finite semifield (or finite division ring) D is a finite nonassociative ring with identity such that
the set D ∗ = D \ {0} is closed under the product, i.e., it is a loop [20,7]. Finite semifields have been
traditionally considered in the context of finite geometries since they coordinatize projective semifield
planes [11]. Recent applications to coding theory [6,17,10], combinatorics and graph theory [21], have
broadened the potential interest in these rings.
Because of their diversity, the obtaining of general theoretical algebraic results seems to be a rather
difficult (and challenging) task. On the other hand, because of their finiteness, computational methods
can be naturally considered in the study of these objects. So, the classification of finite semifields of
a given order is a rather natural problem to use computations. For instance, computers were used in
the classification of semifields or order 32 [25,20] and 81 [8].
In this paper we present a classification of semifields with 64 elements up to isotopy. Because of
the complexity of the problem, the algorithms used in the papers mentioned above cannot be used
directly to solve the case of 64 elements. Our techniques combine two known methods with the help
E-mail addresses: rua@uniovi.es (I.F. Rúa), elias@aic.uniovi.es (E.F. Combarro), ranilla@aic.uniovi.es (J. Ranilla).
1
Partially supported by MEC-MTM-2007-67884-C04-01 and IB08-147.
2
Partially supported by MEC-TIN-2007-61273 and MEC-TIN-2007-29664-E.
doi:10.1016/j.jalgebra.2009.02.020
4012 I.F. Rúa et al. / Journal of Algebra 322 (2009) 4011–4029
of important observations which allow us to reduce the computational effort. The resulting classifica-
tion is somewhat surprising in that only one tenth of the planes found were known beforehand.
The structure of the paper is as follows. In Section 2, basic properties of finite semifields are
reviewed. Section 3 describes known constructions of semifields of order 64. In Section 4, we present
the method we used to classify all 64-element semifields. Finally, in Section 5, a complete description
of the semifields is given, together with several of their properties.
2. Preliminaries
In this section we collect definitions and facts on finite semifields. Proofs can be found, for in-
stance, in [20,7].
A finite nonassociative ring D is called presemifield, if the set of nonzero elements D ∗ is closed
under the product. If D has an identity element, then it is called (finite) semifield. If D is a finite
semifield, then D ∗ is a multiplicative loop. That is, there exists an element e ∈ D ∗ (the identity of D)
such that ex = xe = x, for all x ∈ D and, for all a, b ∈ D ∗ , the equation ax = b (respectively xa = b) has
a unique solution.
Apart from finite fields (which are obviously finite semifields), proper finite semifields were first
considered by L.E. Dickson [9] and were deeply studied by A.A. Albert [1–4]. The term finite semifield
was introduced in 1965 by D.E. Knuth [20]. These rings play an important role in the study of certain
projective planes, called semifield planes [11,20].
The characteristic of a finite presemifield D is a prime number p, and D is a finite-dimensional al-
gebra over GF (q) (q = p c ) of dimension d, for some c , d ∈ N, so that | D | = qd . If D is a finite semifield,
then GF (q) can be chosen to be its associative–commutative center Z ( D ). Other relevant subsets of a
finite semifield are the left, right, and middle nuclei (N l , N r , N m ), and the nucleus N [7].
Isomorphism of presemifields is defined as usual for algebras, and the classification of finite semi-
fields up to isomorphism can be naturally considered. Because of the connections to finite geometries,
we must also consider the following notion. If D 1 , D 2 are two presemifields over the same prime
field GF ( p ), then an isotopy between D 1 and D 2 is a triple ( F , G , H ) of bijective linear maps D 1 → D 2
over GF ( p ) such that
H (ab) = F (a)G (b), ∀a, b ∈ D 1 .
Clearly, any isomorphism between two presemifields is an isotopy, but the converse is not nec-
essarily true. Any presemifield is isotopic to a finite semifield [20, Theorem 4.5.4]. From a presemi-
field D, a projective plane P ( D ) can be constructed. We refer to [11,20] for the details. Theorem 6
in [3] shows that isotopy of finite semifields is the algebraic translation of the isomorphism between
the corresponding projective planes. Two finite semifields D 1 , D 2 are isotopic if, and only if, the
projective planes P ( D 1 ), P ( D 2 ) are isomorphic.
The set of isotopies from a finite semifield D to itself is a group under composition, called the
autotopism group, and denoted At( D ). This group acts on the fundamental triangle of the plane P ( D ),
that is, it leaves invariant each of the three lines L x = {(1, x, 0) | x ∈ D } ∪ {(0, 1, 0)}, L y = {(1, 0, y ) |
y ∈ D } ∪ {(0, 0, 1)}, L ∞ = {(0, 1, z) | z ∈ D } ∪ {(0, 0, 1)}.
Given a finite semifield D, it is possible to construct the set D of all its isotopic but nonnecessarily
isomorphic finite semifields. It is a subset of the set of principal isotopes of D [14,20]. A principal
isotope of D is a finite semifield D ( y ,z) (where y , z ∈ D ∗ ) such that ( D ( y ,z) , +) = ( D , +) and multipli-
cation is given by the rule
a · b = R−1 −1
z (a) L y (b ), ∀a, b ∈ D ,
where R z , L y : D → D are the maps R z (a) = az, L y (a) = ya, for all a ∈ D. Moreover, there is a re-
lation between the order of At( D ) and the orders of the automorphism groups of the elements
I.F. Rúa et al. / Journal of Algebra 322 (2009) 4011–4029 4013
in D [20, Theorem 3.3.4]. If D is a finite semifield, and D is the set of all nonisomorphic semifields
isotopic to D, then
2 1
| D | − 1 = At( D )
|Aut( E )|
E ∈D
where Aut( E ) denotes the automorphism group of a finite semifield E. The sum of the right term
will be called the Semifield/Automorphism (S/A) sum [20, Theorem 3.3.4], and it provides the number
of nonisomorphic semifields generating the same plane, and the order of their automorphism groups.
So, for instance, the S/A sum 28 1
+ 72 means that the plane is coordinatized by 35 nonisomorphic
semifields, 28 of them have trivial automorphism group, 7 have an automorphism group of order 2.
If B = [x1 , . . . , xd ] is a GF (q)-basis of a presemifield D, then there exists a unique set of constants
A D ,B = { A i 1 i 2 i 3 }di ,i ,i =1 ⊆ GF (q) such that
1 2 3

d
xi 1 xi 2 = A i 1 i 2 i 3 xi 3 , ∀i 1 , i 2 ∈ {1, . . . , d}.
i 3 =1
This set of constants is known as cubical array or 3-cube corresponding to D with respect to the
basis B , and it completely determines the multiplication in D.
A remarkable fact is that permutation of the indexes of a 3-cube preserves the absence of nonzero
divisors. Namely, if D is a presemifield, and σ ∈ S 3 (the symmetric group on the set {1, 2, 3}), then
the set
AσD ,B = { A i σ (1) i σ (2) i σ (3) }di1 ,i 2 ,i 3 =1 ⊆ GF (q)
is the 3-cube of a GF (q)-algebra D σB without zero divisors [20, Theorem 4.3.1]. Notice that, in gen-
eral, different choices of bases B , B lead to nonisomorphic presemifields D σB , D σB . However, these
presemifields are always isotopic [20, Theorems 4.4.2 and 4.2.3].
By [20, Theorem 5.2.1], up to six projective planes can be constructed from a given finite semi-
field D using the transformations of the group S 3 . Actually, S 3 acts on the set of semifield planes
of a given order. So, the classification of finite semifields can be reduced to the classification of the
corresponding projective planes up to the action of the group S 3 . In this setting, we will consider
a plane as new if no known3 finite semifield coordinatizes a plane in its S 3 -class.
We shall use a graphical representation to distinguish between the different cases. The vertices of
an hexagon will depict the six different planes obtained from a given finite semifield (cf. [20, Theo-
rem 5.2.1]):
3
We have considered as known semifields those appearing in the, up to our knowledge, last survey on the topic, [18], together
with the planes coordinatized by nonprimitive semifields [12].
A dotted line between two planes shows that the corresponding finite semifields are isotopic.
If the line is continuous, then the corresponding planes can be coordinatized by a commutative or
symplectic semifield [17,16]. So, for instance, the following two pictures represent a plane P ( D ) which
is isotopic to P ( D σ ), for all σ ∈ S 3 . In the second case the plane can be coordinatized by commutative
and symplectic semifields.
The construction of finite semifields of a given order can be rephrased as a matrix problem [8],
[12, Proposition 3]. We state this proposition in the particular case of semifields of order 64.
Proposition 1. There exists a finite semifield D of 64 elements if, and only if, there exists a set of 6 matrices
B Σ = { A 1 , . . . , A 6 } ⊆ GL(6, 2) such that:
1. A 1 is the identity matrix.

6
i =1 λi A i ∈ GL(6, 2), for all non-zero tuples (λ1 , . . . , λ6 ) ∈ Z2 .
6
2.
3. The first column of the matrix A i is the column vector with a 1 in the ith position, and 0 everywhere else.
In such a case, the set { B i jk }6i , j ,k=1 , where B i jk = ( A j )ik , is the 3-cube corresponding to D with respect to the
standard basis of Z62 . In [8], the set B Σ , and its linear span Σ are called standard basis and semifield spread
set (SSS), respectively.
The following convention will be used through this paper to represent a semifield of order 64.
Let B Σ = ( A 1 , A 2 , A 3 , A 4 , A 5 , A 6 ) be one of its standard bases. Recall that the first column of A i has
a one in the ith position and zeroes elsewhere. If the remaining columns of A i are
⎛ ⎞
a29 a23 a17 a11 a5
⎜ a28 a22 a16 a10 a4 ⎟
⎜ ⎟
⎜ a27 a21 a15 a9 a3 ⎟
⎜ ⎟
⎜ a26 a20 a14 a8 a2 ⎟
⎝a a19 a13 a7 a1
⎠
25
a24 a18 a12 a6 a0
29 j
then we will encode A i as the number j =0 a j 2
. For a concrete representation of the semifield one
6
can identify the semifield with Z62 , and the multiplication with x ∗ y = i =1 xi A i y.
Remark 1. Because of [3, Lemma 5], the characteristic polynomial of any matrix in Σ \ {06 , I 6 } has
no linear factors. We will say that a monic polynomial is admissible if it has no linear factors. On the
other hand, if a semifield of order 64 is primitive [26] then it has a standard basis such that A 2 is
a companion matrix whose characteristic polynomial is primitive (i.e. its multiplicative order is 63)
[12, Proposition 2, Corollary 1].
Finally, the following result relates the SSS of isotopic finite semifields (cf. [8, Section 2]).
Proposition 2. If Σ is the SSS of a semifield D of order 64, and Σ is the SSS of an isotope of D, then there
exists a matrix Q ∈ GL(6, 2), and S ∈ Σ , such that Σ = Q −1 Σ S −1 Q .
3. The known semifield planes of order 64
In this section we give a classification of the known semifield planes of order 64. We consid-
ered the examples of [18], and the 36 nonprimitive finite semifields of [12], and explored which
planes can be coordinatized by these constructions. This yielded to 35 semifield planes, divided
into 13 S 3 -classes, that we list below. A semifield representative is given for each plane, together
with the order of its automorphism group. Of these planes only two can be coordinatized by com-
mutative semifields. The total number of nonisomorphic commutative semifields coordinatizing these
planes is 14. In the next section we shall see that no other commutative semifield of order 64 ex-
ists.
I (Desarguesian plane)
Finite field GF (64) (6 automorphisms).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274593, 67639409, 33954937, 25632381, 566730623).
II (Twisted field plane)

Twisted field (1 automorphism) with parameters:
• j ∈ GF (64) such that j 6 + j + 1 = 0;

• α ∈ Aut(GF (64)) such that α (x) = x4 , for all x ∈ GF (64);
2
• β ∈ Aut(GF (64)) such that β(x) = x4 , for all x ∈ GF (64).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274593, 225354480, 673682562, 25632381, 199628676).
III
Knuth’s semifield of type 2 (1 automorphism) with parameters:
• f ∈ GF (8) such that f 3 + f + 1 = 0;

• g ∈ GF (8) such that g + 1 = 0;
• σ ∈ Aut(GF (8)) such that σ (x) = x2 , for all x ∈ GF (8).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274596, 27112887, 35119969, 253266042, 1070246993).
IV
Knuth’s semifield of type 5 (1 automorphism) with parameters:
• f ∈ GF (8) such that f 3 + f + 1 = 0;

• g ∈ GF (8) such that g + 1 = 0;
• σ ∈ Aut(GF (8)) such that σ (x) = x2 , for all x ∈ GF (8).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274593, 189853287, 236639294, 212321269, 624416899).
V (II Huang and Johnson plane [13])

Huang and Johnson sporadic semifield of type II (3 automorphisms).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274623, 1022013944, 102205750, 429859362, 652592216).
VI (III Huang and Johnson plane)

Huang and Johnson sporadic semifield of type III (3 automorphisms).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274605, 1022014833, 374827988, 557069354, 336124018).

VII (IV Huang and Johnson plane)

Huang and Johnson sporadic semifield of type IV (1 automorphism).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274605, 427572234, 1072787891, 401402255, 192290736).
VIII (VI Huang and Johnson plane)

Huang and Johnson sporadic semifield of type VI (2 automorphisms).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274593, 189853287, 580915984, 793113293, 782199145).
IX (VII Huang and Johnson plane)

Huang and Johnson sporadic semifield of type VII (1 automorphism).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274605, 67640187, 851743451, 194887306, 617256025).
X (VIII Huang and Johnson plane)

Huang and Johnson sporadic semifield of type VIII (1 automorphism).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274593, 189853287, 1000703633, 930902659, 782199145).
XI
(Plane coordinatized by a commutative semifield,
S 3 -equivalent to a Kantor–Williams symplectic presemifield plane)
Commutative semifield (6 automorphisms) with tuple of matrices
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274594, 70580276, 37685996, 25345988, 584237329).
XII (Two-sided nonprimitive plane)

H (semifield # 1 in [12, p. 1423]), the unique nonprimitive semifield of order 64 [12] (6 automor-
phisms).
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (146808934, 811798971, 308657185, 563815286, 374228233).
XIII (One-sided nonprimitive plane)

Semifield # 2 in [12, p. 1423], one-sided nonprimitive semifield (1 automorphism) with tuple of
matrices
( A 2 , A 3 , A 4 , A 5 , A 6 ) = (135274600, 518296613, 253216863, 778190320, 47879003).
For each of these 13 semifield representatives, we computed the order of their center and nuclei
Z N = ( Z , N , Nl , N m , N r ), the list of all principal isotopes, and the order of their isomorphism groups.
Some information on the autotopism group was computed r as well as the length of the orbits in the
fundamental triangle ( L x , L ∞ , L y ), given in the form i =1 [b i ], if ai cycles of length b i (i = 1, . . . , r)
a i
exist. Specifically, the computation of the autotopism group of a semifield D was achieved by a direct
computation of tuples ( F , G , H ) of bijective linear maps such that H (ab) = F (a)G (b), for all a, b ∈ D.
The coordinate matrices of these maps were later processed by the software Magma [5] to obtain
information on the autotopism group. All these data are collected in Table 1.
Remark 2. 1. Let us notice that the Huang and Johnson plane of type V [13] is S 3 -equivalent to the
plane VII above. Namely, it is the plane VII(23) .
2. All these planes can be coordinatized by a finite semifield containing a primitive element (even
the planes XII and XIII).
Table 1
Known semifields of order 64 and their properties.
Plane S 3 -class |At| (L x , L∞ , L y ) S/A sum ZN
2[1] + 1[63]
I 23814 2[1] + 1[63] 1
6
(64, 64, 64, 64, 64)
2[1] + 1[63]
2[1] + 1[63]
567
II 2[1] + 1[63] 7
(4, 4, 4, 4, 4)
Z63 Z9 1
2[1] + 1[63]
2[1] + 2[7] + 1[49]

49
III 2[1] + 2[7] + 1[49] 81
(2, 2, 2, 2, 2)
Z27 1
2[1] + 2[7] + 1[49]
2[1] + 1[63]
441
IV 2[1] + 1[63] 9
(2, 2, 8, 4, 8)
Z27 Z9 1
2[1] + 2[7] + 1[49]
3[1] + 1[2] + 2[3] + 9[6]

42
V 2[1] + 1[7] + 1[14] + 1[42] 92
+ 2
+ 4
+ 1
(2, 2, 8, 2, 2)
Z2 × (Z7 Z3 ) 1 2 3 6
2[1] + 1[7] + 1[14] + 1[42]
3[1] + 1[2] + 2[3] + 9[6]

42
VI 2[1] + 1[7] + 1[14] + 1[42] 92
+ 2
+ 4
+ 1
(2, 2, 8, 2, 2)
Z2 × (Z7 Z3 ) 1 2 3 6
2[1] + 1[7] + 1[14] + 1[42]
9[1] + 28[2]
14
VII 2[1] + 1[7] + 4[14] 280
1
+ 7
2
(2, 2, 8, 2, 2)
Cyclic
2[1] + 1[7] + 4[14]
2[1] + 4[3] + 1[6] + 3[9] + 1[18]

126
VIII 2[1] + 1[21] + 1[42] 28
+ 7
(2, 2, 8, 4, 4)
Z7 × (Z23 Z2 ) 1 2
2[1] + 1[21] + 1[42]
3[1] + 1[2] + 6[3] + 7[6]

42
IX 2[1] + 1[21] + 1[42] 91
+ 7
(2, 2, 8, 2, 2)
Z7 × S 3 1 2
2[1] + 1[21] + 1[42]
5[1] + 6[2] + 4[3] + 6[6]

42
X 2[1] + 1[21] + 1[42] 91
+ 7
(2, 2, 8, 2, 2)
Z7 × S 3 1 2
2[1] + 1[21] + 1[42]
2[1] + 1[3] + 1[6] + 2[9] + 2[18]

18
XI 2[1] + 1[3] + 1[6] + 2[9] + 2[18] 211
+ 16
+ 4
+ 1
(2, 2, 2, 4, 2)
Z3 × S 3 1 2 3 6
3[1] + 1[2] + 2[3] + 9[6]
3[1] + 1[2] + 6[3] + 7[6]

6
XII 3[1] + 1[2] + 6[3] + 7[6] 636
1
+ 48
2
+ 4
3
+ 1
6
(2, 2, 2, 2, 2)
Symmetric
3[1] + 1[2] + 6[3] + 7[6]
17[1] + 16[3]
3
XIII 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
Since the same plane can be coordinatized by several constructions, we include Table 2, where
every plane is followed by a list of those (pre)semifield constructions that coordinatize it:
• FF: Finite field;

• TF: Twisted field;
• K: Knuth’s semifield of types 1 to 5;
Table 2
Known 64-element semifield planes and their coordinatizing constructions.
Plane\Permutation 1 (1 2) (1 3) (1 2 3) (2 3) (1 3 2)
I FF – – – – –
II TF – – – – –
III K2 K2 – – – –
IV K5 – K3 / JJ8 – K4 –
V HJII NONE – NONE – –
VI HJIII NONE – NONE – –
VII HJIV NONE NONE NONE HJV NONE
VIII HJVI / JJ4 NONE – NONE – –
IX HJVII NONE – NONE – –
X HJVIII NONE – NONE – –
XI KWC – KWS – NONE –
XII NP2 – – – – –
XIII NP1 NP1 – NONE – –
• JJ: Jha–Johnson semifield constructed over GF (4) or GF (8) [15];

• HJ: Huang–Johnson sporadic semifields of types II–VII;
• KW: Kantor–Williams symplectic presemifield or a S 3 -equivalent commutative presemifield;
• NP: Nonprimitive semifield (1- or 2-sided).
4. Search algorithm
In this section we describe an algorithm to generate all semifields of order 64. As we have previ-
ously noticed (Proposition 1), any semifield of this order can be described by a standard basis (a tuple
of 6 matrices satisfying certain conditions). So, the output of the algorithm will be tuples of matrices
which correspond to finite semifields. Not all possible tuples satisfying the conditions of Proposition 1
will be listed. It is only necessary to obtain representatives of all S 3 -classes of equivalence.
Our method is based on the algorithm proposed in [12]. This algorithm proved to be effi-
cient in the computation of all nonprimitive semifields of order 64, and it can be adapted to our
case. It first fixes matrices A 1 and A 2 , the first equal to the identity matrix, the second one cho-
sen among a small amount of matrices (see [12]). Then, it computes 15 lists L (λ3 ,λ4 ,λ5 ,λ6 ) (for all
(λ3 , λ4 , λ5 , λ6 ) ∈ Z42 \ {(0, 0, 0, 0)}) containing matrices B of GL(6, 2) such that the first column of B
is the vector (0, 0, λ3 , λ4 , λ5 , λ6 )t , and such that the matrices B + A2, B + A1 and B + A2 + A1 are
elements of GL(6, 2). If B Σ = ( A 1 , A 2 , A 3 , A 4 , A 5 , A 6 ) is a standard basis of a semifield of order 64,
6
then any nonzero linear combination of the form i =3 λi A i must be contained in the corresponding
list L (λ3 ,λ4 ,λ5 ,λ6 ) . This provides a fast test to check if a tuple of matrices is a standard basis or not.
The algorithm of [12] uses these lists to produce consistent tuples of matrices ( A 3 , A 4 , A 5 , A 6 ).
The main feature of this procedure is that, for fixed matrix A 3 , the lists of matrices A 4 , A 5 and A 6 are
sieved with the help of the other lists. Once all the tuples are generated, another algorithm is used to
classify (up to isomorphism) the corresponding finite semifields.
This method lead to some of the results of [12], even though the computational effort was remark-
able big (if run in a single machine, 10 months of computing time). However, it cannot be directly
applied to our case. The main obstruction is the size of the lists L (λ3 ,λ4 ,λ5 ,λ6 ) . The lists created in [12],
because of some extra conditions,4 contained approximately 2 million matrices each. In the new situ-
ation, the lists contain more than 7 million matrices each, since no restriction in the characteristic
polynomial of the matrices can be imposed. If applied directly, we would require approximately
466 months to complete the task. This means that, in a certain sense, the classification problem
of semifields of order 64 is almost 50 times more difficult than the primitivity problem. Clearly, the
drawback of the method is that all tuples of matrices must be computed first, and classification is
only achieved after search.
4
Namely, the characteristic polynomial of any matrix in those lists is required to be one of seven known polynomi-
als [12, Corollary 2, Lemma 3].
On the other hand, the algorithm which was used in [8] to classify semifields of order 81 combines
search and classification. It works as follows. Standard bases of semifields of order 81 are obtained
in 4 steps. In the ith step a list of nonequivalent partial standard bases of size i is computed [8]. This
means that a list of tuples ( A 1 , . . . , A i ) of i matrices is produced with the following two conditions.
First, any tuple in the list can be potentially extended to a standard basis (so the name partial).
Second, none of them can be obtained from any other by means of a transformation of the form of
Proposition 2 (so the name nonequivalent). This ensures that the resulting standard bases correspond
to nonisotopic finite semifields. This method lead to the classification of semifields of order 81 in a few
days on a PC [8].
This algorithm cannot be used directly in our setting either. First, there is a significant increase in
the search space since the matrices have size 6 × 6 instead of 4 × 4. Besides, we run into problems
because the number of nonequivalent partial standard bases of size 4 is very large.
We combine these two methods in an effective way. Namely, we use first the method of [8] to pro-
duce nonequivalent partial standard bases of size 3. We introduce a small variation in that algorithm,
since S 3 -classes are enough for the classification of finite semifields. So, we consider nonequivalent
partial standard bases up to S 3 -equivalence (notice that the matrices of a partial standard basis of
a semifield D, when transposed, are a partial standard basis of D (1,3) ). Then, the algorithm of [12] is
used to complete these partial standard bases to tuples of 6 matrices. Finally, the semifields obtained
from those standard bases are classified up to S 3 -equivalence.
Main algorithm: Classification algorithm (up to S 3 -equivalence) of semifields of order 64
• Input: None
• Output: A complete set of representatives of S 3 -equivalence classes of semifields of order 64
(semifields are described by standard bases)
• Procedure:
S := ∅ // Set of representatives
P := PartialStandardBasesOfSize3() // Generation of nonequivalent partial
standard bases of size 3
for A in P do
if L ( A 2 ) has not been previously computed then
L ( A 2 ) := Lists( A 2 ) // Generation of 15 lists of compatible matrices
end
S := S ∪ Complete( A 2 , A 3 , L ( A 2 )) // Complete the partial standard basis
end
return Classify(S) // Classification up to S3 -equivalence
Algorithm 1: PartialStandardBasesOfSize3
• Input: None
• Output: A set of nonequivalent partial bases (of size 3) of semifields of order 64
• Procedure:
T := ∅ // Set of nonequivalent partial standard bases
C := {Companion matrices with primitive characteristic polynomial}
H := {Matrices whose first column is (001000)t , with admissible characteristic polynomial}
for A 2 in C , A 3 in H do
if any nonzero linear combination of { I 6 , A 2 , A 3 } is invertible then T := T ∪ {( I 6 , A 2 , A 3 )}
end // Generation of all partial standard bases of size 3, with A2 in C
end
for A in T do // Removal of equivalent partial standard bases to A
for Σ in { I 6 , A 2 , A 3 , I 6 , A t2 , A t3 } do
for S in Σ do // Generation of equivalent partial standard bases to A

for E ∈ Σ , Q ∈ GL(6, 2) such that Q −1 E S −1 Q is in C do // New matrix A2
for B ∈ Σ \ S , E such that the first column of Q −1 B S −1 Q is
(001000)t do // New matrix A3
if ( I 6 , Q −1 E S −1 Q , Q −1 B S −1 Q ) = A then
Remove ( I 6 , Q −1 E S −1 Q , Q −1 B S −1 Q ) from T
end
end
end
end
end
end
return T
Algorithm 2: Lists
• Input: A companion matrix A 2

• Output: Fifteen lists, containing those matrices which can be used to complete ( I 6 , A 2 ) to a SSS
of a semifield of order 64
• Procedure:
for i = 1, . . . , 15 do
λ6 λ5 λ4 λ3 := Binary representation of i
L (λ3 ,λ4 ,λ5 ,λ6 ) := ∅ // Matrices in a completion of ( I , A 2 ) with first
column (00λ3 λ4 λ5 λ6 )t
for B in GL(6, 2) with first column equal to (00λ3 λ4 λ5 λ6 )t do
if Characteristic polynomials of B and B + A 2 are admissible then
L (λ3 ,λ4 ,λ5 ,λ6 ) := L (λ3 ,λ4 ,λ5 ,λ6 ) ∪ { B }
end
end
end
return [ L (0,0,0,1) , . . . , L (1,1,1,1) ]
Algorithm 3: Complete
• Input: A companion matrix A 2 , a compatible matrix A 3 , and L ( A 2 ) (fifteen lists of matrices

indexed by the binary representation of the natural numbers in the range 1 to 15)
• Output: All possible standard bases of a semifield of order 64 extending the partial standard
basis ( I 6 , A 2 , A 3 )
• Procedure:
T := ∅ // Set of partial standard bases
Create a list LL (0,1,0,0) of matrices A 4 ∈ L (0,1,0,0) such that A 3 + A 4 ∈ L (1,1,0,0)
for A 4 in LL (0,1,0,0) do
Create a list LLL (0,0,1,0) of matrices A 5 ∈ LL (0,0,1,0) such that
A 4 + A 5 ∈ L (0,1,1,0) and A 3 + A 4 + A 5 ∈ L (1,1,1,0)
Create a list LLL (0,0,0,1) of matrices A 6 ∈ LL (0,0,0,1) such that
A 4 + A 6 ∈ L (0,1,0,1) and A 3 + A 4 + A 6 ∈ L (1,1,0,1)
for A 5 ∈ LLL (0,0,1,0) and A 6 ∈ LLL (0,0,0,1) do
if A 5 + A 6 ∈ L (0,0,1,1) and A 3 + A 5 + A 6 ∈ L (1,0,1,1) and

A 4 + A 5 + A 6 ∈ L (0,1,1,1) and A 3 + A 4 + A 5 + A 6 ∈ L (1,1,1,1)
then T := T ∪ {( I 6 , A 2 , A 3 , A 4 , A 5 , A 6 )}
end
end
end
return T
Algorithm 4: Classify
• Input: A collection S of standard bases of semifields of order 64

• Output: Representatives of the S 3 -equivalence classes of the semifields induced by the standard
bases in the collection
• Procedure:
W := ∅ // Set of representatives
C := ∅ // Set of all standard bases of semifields S3 -equivalent to a
finite semifield with standard basis in W
for A in S do // Standard basis of a semifield D
if A ∈
/ C then
W := W ∪ { A }
for σ ∈ S 3 do
for each principal isotope J of D σ do
C := C ∪ {all standard bases of J}
end
end
end
end
return W
The combination of methods which lead to the previous algorithm produced satisfactory results
since there was a 99% reduction in the number of cases to be explored. Originally, six companion
matrices A 2 were considered, corresponding to the six primitive polynomials of degree 6 over Z2 :
x6 + x + 1, x6 + x5 + x2 + x + 1, x6 + x5 + x4 + x + 1,
x6 + x5 + 1, x6 + x5 + x3 + x2 + 1, x6 + x4 + x3 + x + 1.
That is, we restricted our search to primitive semifields (notice that any nonprimitive semifield
of order 64 is S 3 -equivalent to a primitive semifield, since planes XII and XIII can be coordina-
tized by a primitive semifield). From these 6 matrices, a total amount of 46252032 partial stan-
dard basis of size 3 were produced by the algorithm PartialStandardBases. But, after removal of
equivalent ones, just 399866 nonequivalent partial standard bases of size 3 remained. Of these,
377675 had matrix A 2 = C (x6 + x + 1), and the rest had matrix A 2 = C (x6 + x5 + x3 + x2 + 1).
Let us remark that all partial standard bases with A 2 = C ( p (x)), and p (x) one of the poly-
nomials x6 + x5 + x2 + x + 1, x6 + x5 + x4 + x + 1, x6 + x5 + 1, are equivalent to a partial
standard basis of the first kind, while all partial standard bases with A 2 = C (x6 + x4 + x3 +
x + 1) are equivalent to a partial standard basis of the second kind. For instance, any par-
tial standard basis of the form ( I 6 , C (x6 + x + 1), A 3 ) is transformed into a partial standard ba-
sis of the form ( I 6 , C (x6 + x5 + x2 + x + 1), A 3 ) simply by choosing E = I 6 + C (x6 + x + 1)
and
Table 3
Number of 64-element finite semifields.
Number of classes S 3 -action Isotopy Isomorphism

Previously known 13 35 8560
Actual number 80 332 376971
Table 4
Standard bases of new semifields of order 64 ( A 1 = I 6 ).
# A2 A3 A4 A5 A6
XIV 135274594 833399958 260289148 1031543734 289062724
XV 135274608 929017139 43661225 236902583 58939658
XVI 135274599 780494300 790514579 230544263 113930782
XVII 135274613 129115169 42444851 21901924 898304298
XVIII 135274600 61518615 173580196 594359470 490618435
XIX 135274596 127101463 659059070 333784654 916173704
XX 135274594 71778966 584275037 294214292 47830349
XXI 135274623 210604913 369031692 869572955 513934562
XXII 135274614 298719213 371084575 198518457 1064230576
XXIII 135274593 861665485 782365624 171288969 901459391
XXIV 135274599 329921844 727137562 76464013 227638817
XXV 135274611 502051974 921041233 799525002 1031595865
XXVI 135274608 245456463 960054086 892794578 298620733
XXVII 135274605 685174186 978351539 30452336 770639372
XXVIII 135274605 287116498 229601764 116769706 659794001
XXIX 135274600 391353891 935896110 611263392 219544639
XXX 135274605 253357853 818841952 214536771 376932474
XXXI 135274605 927823728 1043775209 180241271 355149199
XXXII 135274594 830091251 496454571 294214292 937736232
XXXIII 135274594 625566993 820897994 1041833019 527914637
XXXIV 135274605 1015682606 420972778 331973660 602075080
XXXV 135274605 760830577 942032486 331689540 1059713288
XXXVI 135274600 1013759534 967035803 19745382 54171530
XXXVII 135274618 702966560 581719755 551123260 248773288
XXXVIII 135274594 615209500 653723442 597803524 64203530
XXXIX 135274594 766078631 695326175 915348146 1069733783
XL 135274623 151599276 350755192 628485436 451306380
XLI 135274603 977768416 817386359 868142796 1019542329
XLII 135274618 815130072 525371889 729166901 111689247
XLIII 135274606 106019670 834133431 978786631 810064469
XLIV 135274605 127050072 1067943835 187195280 46720452
XLV 135274611 1029317650 520088048 219035419 1059375133
XLVI 135274593 331766335 841368844 1036592040 609803946
XLVII 135274608 153781051 728397374 250468680 37621084
XLVIII 135274613 194046320 574609178 123846514 234580732
XLIX 135274606 769446985 220872112 247312724 317866821
L 135274620 658083501 394974963 168700243 63130518
LI 135274623 834590458 173628128 932232776 129367971
LII 135274623 836224690 207465030 645936619 243736264
LIII 135274613 799452523 265350121 339974318 626807932
LIV 135274617 1021351255 508837061 331738527 774541306
LV 135274603 685722191 420150503 99266464 259409008
LVI 135274594 256099125 620378737 948416852 976951193
LVII 135274596 709090705 611750851 500630203 252344113
LVIII 135274599 296969012 813505114 902870605 253337313
LIX 135274620 1001102903 554659815 887933802 199592399
LX 135274617 1033514144 124586730 857117276 1057874229
LXI 135274596 24792404 923093719 241684897 311785609
LXII 135274600 335867443 64403431 539538757 1048273875
LXIII 135274599 379947884 468255421 166163471 737134591
LXIV 135274608 675671366 173741043 1064349970 330692042
LXV 135274599 773727160 109764209 24488523 96297681
Table 4 (continued)
# A2 A3 A4 A5 A6
LXVI 135274603 116650580 995749484 402467592 539691559
LXVII 135274617 17661426 345529328 1017107050 175067510
LXVIII 135274618 213725930 722775913 632723229 390126804
LXIX 135274614 298218413 614428189 791447010 595936461
LXX 135274617 885643496 663907199 176085274 942663142
LXXI 135274620 508419752 1024350995 67020677 581629534
LXXII 135274606 559395716 262131527 86612471 850591393
LXXIII 135274593 808819530 46309136 231311151 1016044841
LXXIV 135274608 382084651 723363276 222713684 40992056
LXXV 135274605 844321142 295097610 835339006 805165097
LXXVI 135274617 342415979 844418594 360800935 877880127
LXXVII 135274596 61984053 93250846 172201558 898164556
LXXVIII 135274613 844649372 514653314 956814412 890109555
LXXIX 135274608 580417165 384216079 701045724 922685158
LXXX 135274614 599051760 670195531 755889110 1021850782
⎛ ⎞
1 0 0 0 0 0
⎜0 0 0 0 0 1⎟
⎜ ⎟
⎜0 0 0 0 1 0⎟
Q =⎜ ⎟
⎜0 0 0 1 1 0⎟
⎝0 0 1 0 1 0⎠
0 1 1 1 1 1
(B can be chosen freely) in the running of the algorithm PartialStandardBasesOfSize3.

The implemented algorithm in language C required 30 days on a 12 2.5 GHz CPU linux cluster
(1 computer year on a single PC). The output, before classification, consisted in 95877 standard bases.
Classification of these matrices up to S 3 -equivalence lead to the results of the next section.
Let us remark that the extraordinary feature of our algorithm is that it allowed us to solve the
classification problem in approximately the same time it was needed in [12] to solve the primitivity
problem. As noticed above, this problem is, in a certain sense, 50 times simpler.
5. The semifield planes of order 64: A classification
The classification of 64-element semifields that we present in this section, completes the clas-
sification of finite semifields of order 125 or less [19,25,20,8,22,23]. Let us compare the number of
S 3 -equivalence classes, semifield planes, and coordinatizing finite semifields which were found, with
those previously known (Table 3).
As we can see approximately one tenth of the semifield planes were previously known. The stan-
dard bases of the coordinatizing semifields, from S 3 -classes XIV to LXXX are collected in Table 4 ( A 1 is
always the identity matrix). A semifield representative with maximal number of automorphisms was
chosen for each S 3 -class.
We processed these semifield representatives to obtain the order of their center and nuclei, the
list of all principal isotopes, and the order of their isomorphism groups. Also, the length of the orbits
in the fundamental triangle and some information on the autotopism group was computed. All this
data is collected in Table 5.
6. Concluding remarks
In this paper we present a computer-assisted classification of finite semifields of order 64. Sur-
prisingly, the classification resulted in many new examples of semifields since only one tenth of the
planes found were known beforehand, and it completes the classification of finite semifields up to
order 125. This is a new step towards the classification of semifield planes of order 256 or less sug-
gested in [18]. Our classification algorithm combines two known methods with the help of important
observations. It proved to be quite suitable to the problem considered. We hope this classification
Table 5
Properties of new semifields of order 64.
2[1] + 1[63]
63
XIV 3[1] + 2[3] + 2[7] + 2[21] 63
(2, 2, 2, 2, 4)
Z7 Z9 1
2[1] + 1[63]
2[1] + 3[7] + 2[21]

21
XV 2[1] + 3[7] + 2[21] 186
+ 9
(2, 2, 2, 2, 2)
Z7 Z3 1 3
2[1] + 3[7] + 2[21]
2[1] + 3[7] + 2[21]

21
XVI 2[1] + 3[7] + 2[21] 186
+ 9
(2, 2, 2, 2, 2)
Z7 Z3 1 3
2[1] + 3[7] + 2[21]
2[1] + 1[3] + 4[15]

15
XVII 2[1] + 1[3] + 4[15] 264
1
+ 3
5
(2, 2, 2, 4, 2)
Cyclic
5[1] + 12[5]
2[1] + 1[3] + 3[5] + 3[15]

15
XVIII 2[1] + 1[3] + 4[15] 264
1
+ 3
5
(2, 2, 2, 2, 2)
Cyclic
2[1] + 1[3] + 4[15]
2[1] + 1[7] + 4[14]

14
XIX 2[1] + 1[7] + 4[14] 280
1
+ 7
2
(2, 2, 2, 2, 2)
Cyclic
2[1] + 1[7] + 4[14]
2[1] + 6[3] + 5[9]

9
XX 2[1] + 6[3] + 5[9] 440
+ 3
(2, 2, 2, 2, 2)
Z23 1 3
2[1] + 6[3] + 5[9]
2[1] + 6[3] + 5[9]

9
XXI 2[1] + 6[3] + 5[9] 440
+ 3
(2, 2, 2, 2, 2)
Z23 1 3
2[1] + 6[3] + 5[9]
2[1] + 7[9]
9
XXII 2[1] + 7[9] 441
1
(2, 2, 2, 4, 2)
Cyclic
5[1] + 20[3]
5[1] + 20[3]
9
XXIII 2[1] + 7[9] 441
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 7[9]
2[1] + 21[3]
9
XXIV 2[1] + 21[3] 441
(2, 2, 4, 2, 4)
Z23 1
2[1] + 6[3] + 5[9]
2[1] + 7[9]
9
XXV 5[1] + 20[3] 441
1
(2, 2, 2, 2, 4)
Cyclic
2[1] + 7[9]
2[1] + 21[3]
9
XXVI 2[1] + 21[3] 441
(2, 2, 4, 2, 4)
Z23 1
2[1] + 6[3] + 5[9]
2[1] + 21[3]
9
XXVII 2[1] + 21[3] 441
(2, 2, 4, 2, 4)
Z23 1
2[1] + 6[3] + 5[9]
2[1] + 6[3] + 5[9]

9
XXVIII 2[1] + 6[3] + 5[9] 440
+ 3
(2, 2, 2, 2, 2)
Z23 1 3
2[1] + 6[3] + 5[9]
Table 5 (continued)
2[1] + 6[3] + 5[9]

9
XXIX 2[1] + 6[3] + 5[9] 440
+ 3
(2, 2, 2, 2, 2)
Z23 1 3
2[1] + 6[3] + 5[9]
2[1] + 6[3] + 5[9]

9
XXX 2[1] + 6[3] + 5[9] 440
+ 3
(2, 2, 2, 2, 2)
Z23 1 3
2[1] + 6[3] + 5[9]
2[1] + 6[3] + 5[9]

9
XXXI 2[1] + 6[3] + 5[9] 440
+ 3
(2, 2, 2, 2, 2)
Z23 1 3
2[1] + 6[3] + 5[9]
2[1] + 6[3] + 5[9]

9
XXXII 2[1] + 6[3] + 5[9] 440
+ 3
(2, 2, 2, 2, 2)
Z23 1 3
2[1] + 6[3] + 5[9]
2[1] + 9[7]
7
XXXIII 9[1] + 8[7] 567
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 9[7]
2[1] + 7[3] + 7[6]

6
XXXIV 2[1] + 7[3] + 7[6] 637
1
+ 49
2
(2, 2, 2, 2, 2)
Symmetric
5[1] + 6[2] + 4[3] + 6[6]
3[1] + 1[2] + 2[3] + 9[6]

6
XXXV 3[1] + 1[2] + 2[3] + 9[6] 652
1
+ 16
2
+ 4
3
+ 1
6
(2, 2, 2, 2, 2)
Cyclic
3[1] + 1[2] + 2[3] + 9[6]
3[1] + 1[2] + 2[3] + 9[6]

6
XXXVI 3[1] + 1[2] + 2[3] + 9[6] 652
1
+ 16
2
+ 4
3
+ 1
6
(2, 2, 2, 2, 2)
Cyclic
3[1] + 1[2] + 2[3] + 9[6]
2[1] + 7[3] + 7[6]

6
XXXVII 2[1] + 7[3] + 7[6] 637
1
+ 49
2
(2, 2, 2, 2, 2)
Symmetric
5[1] + 6[2] + 4[3] + 6[6]
2[1] + 7[3] + 7[6]

6
XXXVIII 2[1] + 7[3] + 7[6] 637
1
+ 49
2
(2, 2, 2, 2, 2)
Symmetric
5[1] + 6[2] + 4[3] + 6[6]
5[1] + 12[5]
5
XXXIX 5[1] + 12[5] 792
1
+ 9
5
(2, 2, 2, 2, 2)
Cyclic
5[1] + 12[5]
2[1] + 21[3]
3
XL 2[1] + 21[3] 1323
1
(2, 2, 2, 4, 2)
Cyclic
65[1]
65[1]
3
XLI 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
XLII 65[1] 1323
1
(2, 2, 2, 2, 4)
Cyclic
2[1] + 21[3]
(continued on next page)

Table 5 (continued)
65[1]
3
XLIII 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
XLIV 2[1] + 21[3] 1323
1
(2, 2, 2, 4, 2)
Cyclic
65[1]
65[1]
3
XLV 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
XLVI 17[1] + 16[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
XLVII 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
17[1] + 16[3]
2[1] + 21[3]
3
XLVIII 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
17[1] + 16[3]
2[1] + 21[3]
3
XLIX 17[1] + 16[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
L 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
17[1] + 16[3]
17[1] + 16[3]
3
LI 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
17[1] + 16[3]
3
LII 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
65[1]
3
LIII 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
65[1]
3
LIV 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
LV 65[1] 1323
1
(2, 2, 2, 2, 4)
Cyclic
2[1] + 21[3]
17[1] + 16[3]
3
LVI 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
17[1] + 16[3]
3
LVII 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
Table 5 (continued)
65[1]
3
LVIII 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
17[1] + 16[3]
3
LIX 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
LX 17[1] + 16[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
17[1] + 16[3]
3
LXI 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
LXII 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
17[1] + 16[3]
65[1]
3
LXIII 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
LXIV 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
17[1] + 16[3]
17[1] + 16[3]
3
LXV 2[1] + 21[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
LXVI 17[1] + 16[3] 1323
1
(2, 2, 2, 2, 2)
Cyclic
2[1] + 21[3]
65[1]
3
LXVII 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
65[1]
3
LXVIII 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
65[1]
3
LXIX 2[1] + 21[3] 1323
1
(2, 2, 4, 2, 2)
Cyclic
2[1] + 21[3]
2[1] + 21[3]
3
LXX 2[1] + 21[3] 1323
1
(2, 2, 2, 4, 2)
Cyclic
65[1]
9[1] + 28[2]
2
LXXI 9[1] + 28[2] 1960
1
+ 49
2
(2, 2, 2, 2, 2)
Cyclic
9[1] + 28[2]
9[1] + 28[2]
2
LXXII 9[1] + 28[2] 1960
1
+ 49
2
(2, 2, 2, 2, 2)
Cyclic
9[1] + 28[2]
(continued on next page)

Table 5 (continued)
9[1] + 28[2]
2
LXXIII 9[1] + 28[2] 1960
1
+ 49
2
(2, 2, 2, 2, 2)
Cyclic
9[1] + 28[2]
9[1] + 28[2]
2
LXXIV 9[1] + 28[2] 1960
1
+ 49
2
(2, 2, 2, 2, 2)
Cyclic
9[1] + 28[2]
65[1]
1
LXXV 65[1] 3969
1
(2, 2, 2, 2, 2)
Cyclic
65[1]
65[1]
1
LXXVI 65[1] 3969
1
(2, 2, 2, 2, 2)
Cyclic
65[1]
65[1]
1
LXXVII 65[1] 3969
1
(2, 2, 2, 2, 2)
Cyclic
65[1]
65[1]
1
LXXVIII 65[1] 3969
1
(2, 2, 2, 2, 2)
Cyclic
65[1]
65[1]
1
LXXIX 65[1] 3969
1
(2, 2, 2, 2, 2)
Cyclic
65[1]
65[1]
1
LXXX 65[1] 3969
1
(2, 2, 2, 2, 2)
Cyclic
65[1]
will be helpful in the discovery of new semifield constructions, a problem which has not been ad-
dressed in this paper. In this direction, using the Oyama representation of matrices [24], and following
a method similar to the one mentioned in [8], we managed to obtain different algebraic descrip-
tions for the semifield representatives of Section 5. This description can be found at the website
http://www.aic.uniovi.es/pir/semifields64.
Acknowledgments
We thank the referees and editor for providing constructive comments and help in improving the
contents of this paper.
References
[1] A.A. Albert, On nonassociative division algebras, Trans. Amer. Math. Soc. 72 (1952) 296–309.
[2] A.A. Albert, Finite noncommutative division algebras, Proc. Amer. Math. Soc. 9 (1958) 928–932.
[3] A.A. Albert, Finite division algebras and finite planes, Proc. Sympos. Appl. Math. 10 (1960) 53–70.
[4] A.A. Albert, Generalized twisted fields, Pacific J. Math. 11 (1961) 1–8.
235–265.
[6] A.R. Calderbank, P.J. Cameron, W.M. Kantor, J.J. Seidel, Z4 -Kerdock codes, orthogonal spreads, and extremal Euclidean line-
sets, Proc. London Math. Soc. (3) 75 (1997) 436–480.
[7] M. Cordero, G.P. Wene, A survey of finite semifields, Discrete Math. 208/209 (1999) 125–137.
[8] U. Dempwolff, Semifield planes of order 81, J. Geom. 89 (2008) 1–16.
[9] L.E. Dickson, Linear algebras in which division is always uniquely possible, Trans. Amer. Math. Soc. 7 (1906) 370–390.
[10] S. González, C. Martínez, I.F. Rúa, Symplectic spread based generalized Kerdock codes, Des. Codes Cryptogr. 42 (2) (2007)
213–226.
[11] M. Hall Jr., The Theory of Groups, Macmillan, 1959.

[12] I.R. Hentzel, I.F. Rúa, Primitivity of finite semifields with 64 and 81 elements, Internat. J. Algebra Comput. 17 (7) (2007)
1411–1429.
[13] H. Huang, N.L. Johnson, 8 semifield planes or order 82 , Discrete Math. 80 (1990) 69–79.
[14] D.R. Hughes, F.C. Piper, Projective Planes, Grad. Texts in Math., vol. 6, Springer-Verlag, New York, Berlin, 1973.
[15] V. Jha, N.L. Johnson, An analog of the Albert–Knuth theorem on the orders of finite semifields, and a complete solution to
Cofman’s subplane problem, Algebras Groups Geom. 6 (1989) 1–35.
[16] W.M. Kantor, Commutative semifields and symplectic spreads, J. Algebra 270 (2003) 96–114.
[17] W.M. Kantor, M.E. Williams, Symplectic semifield planes and Z4 -linear codes, Trans. Amer. Math. Soc. 356 (2004) 895–938.
[18] W.M. Kantor, Finite semifields, in: Finite Geometries, Groups, and Computation, Proc. of Conf. at Pingree Park, CO Sept.
2005, de Gruyter, Berlin, New York, 2006.
[19] E. Kleinfeld, Techniques for enumerating Veblen–Wedderburn systems, J. ACM 7 (1960) 330–337.
[20] D.E. Knuth, Finite semifields and projective planes, J. Algebra 2 (1965) 182–217.
[21] J.P. May, D. Saunders, Z. Wan, Efficient matrix rank computation with applications to the study of strongly regular graphs,
in: Proceedings of ISSAC 2007, ACM, New York, 2007, pp. 277–284.
[22] G. Menichetti, Algebre Tridimensionali su un campo di Galois, Ann. Mat. Pura Appl. 97 (1973) 293–302.
[23] G. Menichetti, On a Kaplansky conjecture concerning three-dimensional division algebras over a finite field, J. Algebra 47
(1977) 400–410.
[24] T. Oyama, On quasifields, Osaka J. Math. 22 (1985) 35–54.
[25] R.J. Walker, Determination of division algebras with 32 elements, Proc. Sympos. Appl. Math. 75 (1962) 83–85.
[26] G.P. Wene, On the multiplicative structure of finite division rings, Aequationes Math. 41 (1991) 222–233.
Journal of Algebra
A subalgebra of 0-Hecke algebra

Xuhua He 1
Department of Mathematics, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong
Article history: Let ( W , I ) be a finite Coxeter group. In the case where W

Received 16 September 2008 is a Weyl group, Berenstein and Kazhdan in [A. Berenstein, D.
Available online 17 April 2009 Kazhdan, Geometric and unipotent crystals. II. From unipotent
Communicated by Meinolf Geck
bicrystals to crystal bases, in: Quantum Groups, in: Contemp.
Keywords:
Math., vol. 433, Amer. Math. Soc., Providence, RI, 2007, pp. 13–88]
Coxeter groups constructed a monoid structure on the set of all subsets of I using
0-Hecke algebra unipotent χ -linear bicrystals. In this paper, we will generalize
this result to all types of finite Coxeter groups (including non-
crystallographic types). Our approach is more elementary, based on
some combinatorics of Coxeter groups. Moreover, we will calculate
this monoid structure explicitly for each type.
1.1. Let W be a Coxeter group generated by the simply reflections si (for i ∈ I ). Let H be the
Iwahori–Hecke algebra associated to W with parameter q = 0, i.e., H is a Q-algebra generated by T si
for si ∈ I with relations T s2i = − T si and the braid relations. The algebra H is called 0-Hecke algebra. It
was introduced by Norton in [No]. Representations of H were later studied in the work of Carter [Ca],
Hivert, Novelli and Thibon [HNT] and etc. More recently, Stembridge [St] used the 0-Hecke algebra to
obtain a new proof for the Möbius function of the Bruhat order of W .
1.2. Set T s i = − T si . For w ∈ W , we define T w

= T · · · T , where w = s · · · s is a reduced
si si i1 ik
1 k
expression of w. Tit’s theorem implies that T w is well defined. Moreover, we have a binary operation
∗ : W × W → W such that T x T y = T x ∗ y for any x, y ∈ W . It is easy to see that ( W , ∗) is a monoid

with unit element 1.
Now we state our main theorem.
E-mail address: maxhhe@ust.hk.

1
The author is partially supported by (USA) NSF grant DMS 0700589 and (HK) RGC grant DAG08/09.SC03.
doi:10.1016/j.jalgebra.2009.04.003
X. He / Journal of Algebra 322 (2009) 4030–4039 4031
J
Theorem 1. Let W be a finite Coxeter group. For any subset J ⊂ I , let w 0 be the maximal element in the
J
subgroup generated by s j (for j ∈ J ). Then { w 0 w 0I ; J ⊂ I } is a commutative submonoid of ( W , ∗). In other
words, there exists a commutative monoid structure I on the set of subsets of I , such that
T J T J = T J I J 2 .
w 0 1 w 0I w 0 2 w 0I w 01 w 0I
Remark. In the case where W is a Weyl group, this result was discovered by Berenstein and Kazhdan
in [BK, Proposition 2.30]. Their approach was based on unipotent χ -linear bicrystals. The proof below
is more elementary. It is based on some combinatorial properties of Coxeter groups. In the end, we
will calculate the operator I explicitly for each type.
Below we introduce a binary operation : W × W → W and study some properties about the
operations ∗ and . In the case where W is a finite Weyl group, these properties were proved in [HL]
using geometry of flag varieties.
1.3. We denote by l the length function and the Bruhat order on W .

By [He, Lemma 1.4(1)], for x, y ∈ W , the subset {u y ; u x} of W contains a unique minimal
element which we denote by x y. Moreover, x y = u y for some u x and l(x y ) = l( y ) − l(u ).
We also have that if si x > x, then (si x) y = min{si (x y ), x y }.
There is a similar description for x ∗ y.
Lemma 1. Let x, y ∈ W . Then the subset {uv ; u x, v y } contains a unique maximal element, which equals
x ∗ y. Moreover, x ∗ y = u y = xv for some u x and v y and l(x ∗ y ) = l(u ) + l( y ) = l(x) + l( v ).
Remark. A slightly weaker version was proved in [He, Lemma 1.4 (2)]. The proof here is similar
to [He].
By definition, si ∗ w = max{ w , si w }. Now for a, b ∈ W and i ∈ I with si a > a, we have that
T ( si a)∗b = T s i a T b = T s i T a T b = T s i T a ∗b .
Hence (si a) ∗ b = max{a ∗ b, si (a ∗ b)} if si a > a.

We only prove that {uv ; u x, v y } contains a unique maximal element which equals x ∗ y and
x ∗ y = u y for some u x with l(x ∗ y ) = l(u ) + l( y ).
We argue by induction on l(x). For l(x) = 0, the statement is clear. Assume that l(x) > 0 and that
the statement holds for all x with l(x ) < l(x). Then there exists i ∈ I such that si x < x. By induction
hypothesis, the subset {uv ; u si x, v y } contains a unique maximal element (si x) ∗ y and there
exists u 1 si x such that (si x) ∗ y = u 1 y and l(u 1 y ) = l(u 1 ) + l( y ).
Set z = x ∗ y. Then z = max{u 1 y , si u 1 y } and zy −1 = u 1 or si u 1 . In either case, we have that
zy −1 x. If zy −1 = u 1 , then we already know that l( z) = l( zy −1 ) + l( y ). If zy −1 = si u 1 , then
si u 1 y > u 1 y and
l(si u 1 y ) = l(u 1 y ) + 1 = l(u 1 ) + l( y ) + 1 l(si u 1 ) + l( y ) l(si u 1 y ).
Thus we must have that l(si u 1 ) = l(u 1 ) + 1 and l( z) = l( zy −1 ) + l( y ).

Now for any u x and v y. By [Lu, Corollary 2.5], u si x or si u si x. By the definition of
(si x) ∗ y, we have that uv or si uv is less than or equal to (si x) ∗ y = u 1 y z. By [Lu, Corollary 2.5],
uv z. Therefore z is the unique maximal element in the subset {uv ; u x, v y }. 2
Corollary 1. Let x x and y y. Then x ∗ y x ∗ y.

4032 X. He / Journal of Algebra 322 (2009) 4030–4039
Lemma 2. Let x x and y y, then x y x y.
By definition, x y x y . Now we prove that x y x y by induction on l(x).

For l(x) = 0, x y = y y = x y. Now assume that l(x) > 0 and that x1 y x1 y for any x1
with l(x1 ) < l(x). Then there exists i ∈ I such that si x < x. By induction hypothesis, (si x) y (si x) y.
By [Lu, Corollary 2.5], x y = min{(si x) y , si ((si x) y )} (si x) y , si ((si x) y ). Hence x y
min{(si x) y , si ((si x) y )} = x y. The statement is proved. 2
Lemma 3. The action ( W , ∗) × W → W , (x, y ) → x y is a left action of the monoid ( W , ∗).
By definition, 1 x = x for any x ∈ W .

Let x, y , z ∈ W . Then there exists u x and v y such that y z = v z and x ( v z) = uv z. By
definition uv x ∗ y. So (x ∗ y ) z x ( y z).
On the other hand, there exists w x such that x ∗ y = w y and l( w y ) = l( w ) + l( y ). Then there
exists w w y such that ( w y ) z = w z. Since l( w y ) = l( w )+ l( y ), we may write w as w = w 1 y 1 for
some w 1 w and y 1 y. Thus (x ∗ y ) z = w 1 ( y 1 z). Notice that w 1 x and y 1 y and y 1 z y z.
By the previous lemma, x ( y z) w 1 ( y 1 z) = (x ∗ y ) z. The lemma is proved. 2
Lemma 4. Assume that W is finite. Then (x y ) w 0I = x ∗ ( y w 0I ).
By definition, x y is the unique minimal element in {u y ; u x}. Hence (x y ) w 0I is the unique

maximal element in {u y w 0 ; u x}. By Lemma 1, this unique maximal element is x ∗ ( y w 0I ). 2
J J J J J
Lemma 5. Let J 1 , J 2 ⊂ I . If ( w 0 1 w 0I ) ∗ ( w 0 2 w 0I ) = w 0 3 w 0I for some J 3 ⊂ I , then ( w 0 2 w 0I ) ∗ ( w 0 1 w 0I ) =
J J J J J J J
w 0 3 w 0I . In other words, if ( w 0 1 w 0I ) w 02 = w 03 for some J 3 ⊂ I , then ( w 0 2 w 0I ) w 01 = w 03 .
This was proved in [BK, Proposition 2.30 (c)] by applying the anti-automorphism w → w 0I w −1 w 0I
on W .
Lemma 6. If J 2 = K K with sk sk = sk sk for any k ∈ K and k ∈ K . Then for any J 1 ⊂ I , we have that
J J J J
w 0 1 w 0I w 0 2 = w 0 1 w 0I w 0K w 0 1 w 0I w 0K .
J J
We fix a reduced expression w 0 1 w 0I = si 1 si 2 · · · sin for i 1 , i 2 , . . . , in ∈ I . Assume that v w 0 1 w 0I
J J J J J
with ( w 0 1 w 0 ) w 0 2 = v w 0 2 . Then v w 0 2 . Hence v ∈ W J 2 . Let v = sil · · · sil be a reduced subex-
1 k
pression. Then l1 , . . . , lk ∈ J 2 = K ∪ K . By assumption on K and K , v
= v 1 v 2 for v 1 ∈ W K and
J J J
v 2 ∈ W K . So v w 0 2 = v 1 v 2 w 0K w 0K = ( v 1 w 0K )( v 2 w 0K ) (( w 0 1 w 0I ) w 0K )(( w 0 1 w 0I ) w 0K ).
J J J
On the other hand, assume that v 1 , v 2 w 0 1 w 0I with ( w 0 1 w 0I ) w 0K = v 1 w 0K and ( w 0 1 w 0I ) w 0K =

v 2 w 0K . Then v 1 ∈ W K and v 2 ∈ W K and there exist 1 t 1 < t 2 < · · · < t u n and 1 t 1 < t 2 < · · · <
t u n such that v 1 = sit · · · situ is a reduced subexpression of v 1 and v 2 = sit · · · sit is a reduced
1 1 u
subexpression of v 2 .
Since K ∩ K = ∅, {t 1 , . . . , t u } and {t 1 , . . . , t u } are disjoint subsets of {1, . . . , n}. Let v be the
element that corresponds to the subexpression {t 1 , . . . , t u } {t 1 , . . . , t u }. Then it is easy to see
J J J
that v = v 1 v 2 . Hence (( w 0 1 w 0I ) w 0K )(( w 0 1 w 0I ) w 0K ) = ( v 1 w 0K )( v 2 w 0K ) = v 1 v 2 w 0K w 0K = v w 0 1
J J
( w 0 1 w 0I ) w 02 . 2
Lemma 7. Let J 1 , J 2 ⊂ I . Then for any J 1 ⊂ J 1 , we have that
J J J J J J
w 0 1 w 0I w 0 2 = w 0 1 w 0 1 w 0 1 w 0I w 0 2 .
J J J J J J J
Notice that w 0 1 w 0I = ( w 0 1 w 0 1 )( w 0 2 w 0I ) = ( w 0 1 w 0 1 ) ∗ ( w 0 1 w 0I ). The lemma follows from
Lemma 3. 2
Below is the key lemma.
Lemma 8. Assume that W is a irreducible finite Coxeter group and i , i ∈ I are end points of the Coxeter graph
of W . Let J 1 = I − {i } and J 2 = I − {i }. Then ( w 0 1 w 0I ) w 0 2 = w 0 3 for some J 3 ⊂ J 1 ∩ J 2 .
J J J
We will prove this lemma in subsection 1.5. The proof is based on a case-by-case checking. We
will also use the result to give an explicit description of the operator I for each type.
Before proving the lemma, we will show that the key lemma implies the main theorem.
1.4. Proof of Theorem 1
By Lemma 5, if T J T J = T J , then T J T J = T J . Using Lemma 4, we may

w 0 1 w 0I w 0 2 w 0I w 0 3 w 0I w 0 2 w 0I w 0 1 w 0I w 0 3 w 0I
reformulate the main theorem as follows:
J J J
For any J 1 , J 2 ⊂ I , we have that ( w 0 1 w 0I ) w 0 2 = w 0 3 for some J 3 ⊂ I .
We argue by induction on the cardinality of I . By Lemma 6 and Lemma 7, it suffices to prove the
case where W is irreducible and J 1 , J 2 are connected in the Coxeter graph.
It is easy to see that ( w 0 w 0I ) w 0I = ( w 0I w 0I ) w 0 = w 0 and ( w ∅0 w 0I ) w 0 = ( w 0 w 0I ) w ∅0 = 1 =
J J J J J
w ∅0 . Now assume that J 1 , J 2 are proper connected subgraph in the Coxeter graph. Then there exists
end points i , i ∈ I such that i ∈ / J 2 . Set J 1 = I − {i } and J 2 = I − {i }. Then J 1 ⊂ J 1 and
/ J 1 and i ∈

J 2 ⊂ J 2.
J J J J J J J J J
By Lemma 7, ( w 0 2 w 0I ) w 0 1 = ( w 0 2 w 0 2 ) (( w 0 2 w 0I ) w 0 1 ). By Lemma 8, ( w 0 2 w 0I ) w 0 1 = w 0 3
J
for some J 3 ⊂ J 1 ∩ J 2 . By induction hypothesis on W J , we have that ( w 0 2 w 0 2 ) w 0 3 = w 0 4 for some
J J J
2
J
J 4 ⊂ J 1 ∩ J 2 . By Lemma 5, ( w 0 1 w 0I ) w 0 2 = w 0 4 .
J J
J J J J J J J J J
Again by Lemma 7, ( w 0 1 w 0I ) w 0 2 = ( w 0 1 w 0 1 ) (( w 0 1 w 0I ) w 0 2 ) = ( w 0 1 w 0 1 ) w 0 4 . By induction
J J J J
hypothesis on W J , we have that
1
( w 01 w 01 ) w 0 4 = w 0 5 for some J 5 ⊂ J 1 ∩ J 2 .
1.5. Proof of Lemma 8
We use the same labeling of Coxeter graph as in [Bo].

For 1 a, b n, set

sa sa−1 · · · sb , if a b,
s[a,b] =
1, otherwise.
Type A n
I −{1} I I −{n} I
We have that w 0 w 0 = s− 1
[n,1] and w 0 w 0 = s[n,1] . Hence
I −{1} I −{1} I −{1} I −{1} I −{1,2}

w0 w 0I w 0 = s− 1
[n,1] w 0 = s− 1
[n,2] w 0 = w0 ,
I −{1} I −{n} I −{n} I −{n} I −{1,n}
w0 I
w0 w0 = s− 1
[n,1] w 0 = s− 1
[n−1,1] w 0 = w0 ,
I −{n} I
I −{n} I −{n} I −{n} I −{n−1,n}
w0 w0 w0 = s[n,1] w 0 = s[n−1,1] w 0 = w0 .
Type B n
I −{1} I I −{n} I
[n−1,1] s[n,1] and w 0 w 0 = sn s− 1 −1
[n,n−1] · · · s[n,1] . Hence
I −{1} 1 I −{1} I −{1} I −{1} I −{1,2}

w0 = s−
w 0I w 0
[n−1,1] s[n,1] w 0 = s− 1
[n−1,2] s[n,2] w 0 = w0 ,
I −{1} I I −{n} 1 I −{n} I −{n} I −{n−1,n}
w0 w0 w0 = s−
[n−1,1] s[n,1] w 0 = s− 1
[n−1,1] s[n,1] w 0 = s− 1
[n−1,1] w 0
I −{1,n−1,n}
= w0 ,
I −{n} I I −{n} −1 I −{n} I −{n}
w0 w0 w0 = sn s− 1
[n,n−1] · · · s[n,1] w 0 = sn−1 s− 1 −1
[n−1,n−2] · · · s[n−1,1] w 0 = 1.
Type D n
Set

1, if 2 n;
=
0, if 2 | n.
We have that
I −{1}
w0 w 0I = s− 1
[n−2,1] s[n,1] ,
I −{n−1}
w0 w 0I = sn−1 (sn−2 sn ) · · · s− 1
[n−2,2] sn− s− 1
[n−2,1] sn−1+ ,
I −{n}
w0 w 0I = sn (sn−2 sn−1 ) · · · s− 1
[n−2,2] sn−1+ s− 1
[n−2,1] sn− .
Hence
I −{1} 1 I −{1} I −{1} I −{1} I −{1,2}

w0 = s−
w 0I w 0[n−2,1] s[n,1] w 0 = s− 1
[n−2,2] s[n,2] w 0 = w0 ,
I −{1} I I −{n−1} I −{n−1} I −{n−1} I −{1,n−1,n}
w0 w0 w0 = s− 1
[n−2,1] s[n,1] w 0 = s− 1
[n−2,1] sn s[n−2,1] w 0 = w0 ,
I −{n−1} I I −{n} 1 −1 I −{n}
w0 w0 w0 = sn−1 (sn−2 sn ) · · · s−
[n−2,2] sn− s[n−2,1] sn−1+ w 0
I −{n}
= sn−1 sn−2 · · · s− 1 −1
[n−2+ ,2] s[n−1− ,1] w 0

s s · · · sn−2 , if 2 n;
= 1 3
s2 s4 · · · sn−2 , if 2 | n,
I −{n} I I −{n} 1 −1 I −{n}
w0 w0 w0 = sn (sn−2 sn−1 ) · · · s−[n−2,2] sn−1+ s[n−2,1] sn− w 0
I −{n}
= s− 1 −1 −1
[n−1,n−2] · · · s[n−1− ,2] s[n−2+ ,1] w 0

s s · · · sn−1 , if 2 n;
= 2 4
s1 s3 · · · sn−1 , if 2 | n.
Applying the automorphism σ : W → W which exchanges sn−1 and sn , we also have that
I −{1} I −{n} I −{1,n−1,n}
(w0 w 0I ) w 0 = w0 and

I −{n−1} I −{n−1} (s2 s4 · · · sn−3 )sn , if 2 n;
w0 w 0I w 0 =
(s1 s3 · · · sn−3 )sn , if 2 | n.
For type E, set x = s4 s3 s5 s4 s2 .

Type E 6
We have that
I −{1}
w0 w 0I = s1 s− 1 −1 −1
[6,3] x s[6,1] ,
I −{2}
w0 w 0I = x−1 s− 1
[6,1] s[5,1] x,
I −{6}
w0 w 0I = s[6,1] s− 1 −1
[4,6] s[3,5] s2 s4 s3 s1 .
Hence
I −{1} I −{1} I −{1} 1 −1 −1 I −{1} {2,4,5}

w0 w 0I w 0 = s1 s− 1 −1 −1
[6,3] x s[6,1] w 0 = s−
[6,3] x s[6,2] w 0 = w0 ,
I −{1} I −{2} I −{2} −1 I −{2}

w0 w 0I w 0 = s1 s− 1 −1 −1
[6,3] x s[6,1] w 0 = s1 s− 1
[6,3] s4 s3 s5 s4 s1 s[6,3] w 0
I −{1,2} I −{1,2}
= s1 s− 1
[6,3] s4 s3 s5 s4 w 0 = s− 1
[6,3] s4 s3 s5 s4 w 0 = s4 s6 ,
I −{1} I −{6} I −{6} I −{6} {3,4,5}

w0 w 0I w 0 = s1 s− 1 −1 −1
[6,3] x s[6,1] w 0 = s1 s− 1 −1 −1
[5,3] x s[5,1] w 0 = w0 ,
I −{2} I −{2} I −{2} I −{2}

w0 w 0I w 0 = x−1 s− 1
[6,1] s[5,1] x w 0 = x−1 s1 s− 1
[6,3] s[5,3] s1 s4 s3 s5 s4 w 0
= x−1 (s3 s5 ) = 1.
I −{6} I −{6} {2,3,4}

Applying the nontrivial diagram automorphism, we also have that ( w 0 w 0I ) w 0 = w0
I −{6} I I −{2}
and (w0 w0) w0 = s1 s4 .
Type E 7
We have that
I −{1}
w0 w 0I = s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1 ,
I −{2}
w0 w 0I = s2 s4 s3 s1 s[5,2] s− 1
[6,4] s[5,1] xs[7,1] xs[6,3] s[7,4] s2 ,
I −{7}
w0 w 0I = s[7,1] xs[6,3] s1 s[7,2] s− 1
[7,4] .
Hence
I −{1} I −{1} I −{1}

w0 w 0I w 0 = s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1 w 0
I −{1}
= (s3 s4 s2 s[5,3] ) s[6,2] s− 1
[6,4] s[7,2] xs[6,3] w 0
{2,4,5,7}
= (s3 s4 s2 s[5,3] ) w 0 = s2 s5 s7 ,
I −{1} I −{2} I −{2}

w0 w 0I w 0 = s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1 w 0
I −{2}
= s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] (s[7,1] xs[6,3] s1 ) w 0

= s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] (s1 s4 s[6,3] s1 ) = 1,
I −{1} I −{7} I −{7}

w0 w 0I w 0 = s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1 w 0
I −{7}
= s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] (s[6,1] xs[6,3] s1 ) w 0
I −{6,7}
= s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] w 0
I −{6,7} {3,4,5}
= s1 s3 s4 s2 s[5,3] s1 s[5,2] s− 1
[5,4] w 0 = w0 ,
I −{2} I −{2} I −{2}

w0 w 0I w 0 = s2 s4 s3 s1 s[5,2] s− 1
[6,4] s[5,1] xs[7,1] xs[6,3] s[7,4] s2 w 0
I −{2}
= s2 s4 s3 s1 s[5,2] s− 1
[6,4] s[5,1] x (s[7,1] xs[6,3] s[7,4] ) w 0

= s2 s4 s3 s1 s[5,2] s− 1
[6,4] s[5,1] x (s1 s4 s6 ) = 1,
I −{2} I −{7} I −{7}

w0 w 0I w 0 = s2 s4 s3 s1 s[5,2] s− 1
[6,4] s[5,1] xs[7,1] xs[6,3] s[7,4] s2 w 0
I −{7}
= s2 s4 s3 s1 s[5,2] s− 1
[6,4] (s[5,1] xs[6,1] xs[6,3] s[6,4] s2 ) w 0
{1,3,4,6}
= s2 s4 s3 s1 s[5,2] s− 1
[6,4] w 0 = 1,
I −{7} I −{7} I −{7} I −{7}

w0 w 0I w 0 = s[7,1] xs[6,3] s1 s[7,2] s− 1
[7,4] w 0 = s[6,1] xs[6,3] s1 s[6,2] s− 1
[6,4] w 0
{2,3,4,5}
= w0 .
Type E 8
We have that
I −{1}
w0 w 0I = s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1 −1
[6,4] s[7,1] xs[6,3] s1 s[8,1] xs[6,3] s1 s[7,2] s[6,4] s[8,1] xs[6,3] s1 ,
I −{2}
w0 w 0I = s2 s4 s3 s1 s[5,2] s− 1 −1
[6,4] s[5,1] xs[7,1] xs[6,3] s[7,4] s2 s[8,1] xs[6,3] s1 s[7,2] s[7,4] s[8,1] xs[6,3] s[7,4] s2 ,
I −{8}
w0 w 0I = s[8,1] xs[6,3] s1 s[7,2] s− 1 −1
[7,4] s[8,1] xs[6,3] s1 s[7,2] s[8,4] .
Hence
I −{1} I −{1}
w0 w 0I w 0 = s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1
I −{1}
s[8,1] xs[6,3] s1 s[7,2] s− 1
[6,4] s[8,1] xs[6,3] s1 w 0

= s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1 (s3 s5 s7 )
= 1,
I −{1} I −{2}
w0 w 0I w 0 = s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1 s[8,1] xs[6,3] s1
I −{2}
s[7,2] s− 1
[6,4] s[8,1] xs[6,3] s1 w 0

= s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1 s[8,1] xs[6,3] s1 (s1 s3 s[6,3] s1 )
= 1,
I −{1} I −{8}
w0 w 0I w 0 = s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1
I −{8}
s[8,1] xs[6,3] s1 s[7,2] s− 1
[6,4] s[8,1] xs[6,3] s1 w 0
{2,3,4,5,6}
= s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s1 w 0
{2,3,4,5,6}
= s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] (s[7,1] xs[6,3] s1 ) w 0
{3,4,5}
= s1 s3 s4 s2 s[5,3] s1 s[6,2] s− 1
[6,4] w 0 = 1,
I −{2} I −{2} −1
w0 w 0I w 0 = s2 s4 s3 s1 s[5,2] s− 1
[6,4] s[5,1] xs[7,1] xs[6,3] s[7,4] s2 s[8,1] xs[6,3] s1 s[7,2] s[7,4]
I −{2}
(s[8,1] xs[6,3] s[7,4] s2 ) w 0
−1
= s2 s4 s3 s1 s[5,2] s− 1
[6,4] s[5,1] xs[7,1] xs[6,3] s[7,4] s2 s[8,1] xs[6,3] s1 s[7,2] s[7,4]
(s1 s4 s6 s[7,3] s1 ) = 1,
I −{2} I −{8}
w0 w 0I w 0 = s2 s4 s3 s1 s[5,2] s− 1
[6,4] s[5,1] xs[7,1]
I −{8}
xs[6,3] s[6,4] s2 s[7,1] xs[6,3] s1 s[6,2] s− 1
[6,4] s[7,1] xs[6,3] s[7,4] s2 w 0
{1,5,6}
= s2 s4 s3 s1 s[5,2] s− 1
[6,4] s[5,1] xs[7,1] w 0 = 1,
I −{8} I −{8} I −{8} {2,3,4,5}

w0 w 0I w 0 = s[6,1] xs[6,3] s1 s[6,2] s− 1 −1
[6,4] s[7,1] xs[6,3] s1 s[7,2] s[7,4] w 0 = w0 .
Type F 4
I −{1} I I −{4} I
[4,1] s2 s3 s2 s1 s2 s3 s2 s[4,1] and w 0 w 0 = s[4,1] s3 s2 s3 s4 s3 s2 s3 s− 1
[4,1] . Hence
I −{1} 1 I −{1} I −{1}

w0 = s−
w 0I w 0
[4,1] s2 s3 s2 s1 s2 s3 s2 s[4,1] w 0 = 1,
I −{1} I I −{4} 1 I −{4}
w0 w0 w0 = s−
[4,1] s2 s3 s2 s1 s2 s3 s2 s[4,1] w 0 = 1.
I −{4} I −{4}
Applying the nontrivial diagram automorphism, we also have that ( w 0 w 0I ) w 0 = 1.
Type H 3
I −{1} I I −{3} I
We have that w 0 w 0 = s1 s2 s1 s2 s3 s2 s1 s2 s1 s3 s2 s1 and w 0 w 0 = s3 s2 s1 s2 s1 s3 s2 s1 s2 s3 . Hence it
I −{1} I −{1} I −{1} I −{3} I −{3} I −{3}
is easy to see that ( w 0 w 0I ) w 0 = (w0 w 0I ) w 0 = (w0 w 0I ) w 0 = 1.
Type H 4
I −{1} I I −{4} I
We have that w 0 w 0 = s1 s2 s1 s2 s3 s2 s1 s2 s1 s3 s2 s1 (s[4,1] s2 s1 s[3,1] s2 s3 )4 s3 s2 and w 0 w0 =
I −{1} I −{1} I −{1} I −{4}
(s[4,1] s2 s1 s[3,1] s2 s3 )4 s4 . Hence it is easy to see that ( w 0 w 0I ) w 0 = (w0 w 0I ) w 0 =
I −{4} I I −{4}
(w0 w0) w0 = 1.
Type I m
I −{1} I I −{1} I −{1} I I −{2} I −{2} I I −{2}
It is easy to see that ( w 0 w0) w0 = (w0 w0) w0 = (w0 w0) w0 = 1.
1.6. Now we will calculate the operator I explicitly for each type. This is based on the previous
subsection, the equalities J 1 I J 2 = J 2 I J 1 (see Lemma 5), J 1 I J 2 = J 1 J 1 ( J 1 I J 2 ) for any J 1 ⊂ J 1
(see Lemma 7) and the inequality J 1 I J 2 ⊂ J 1 I J 2 for J 1 ⊂ J 1 and J 2 ⊂ J 2 (see Lemma 2). We just
list below the cases where J 1 and J 2 are proper connected subgraph of the Coxeter graph of W .
For the case where J 1 and J 2 are not necessarily connected and J 1 , . . . , J l (resp. J 1 , . . . , J l ) are
the connected components of J 1 (resp. J 2 ), we have that J 1 I J 2 = 1i l,1i l ( J i I J i ) (see
Lemma 6).
Type A n : Let J 1 = {a, a + 1, . . . , n − b}, J 2 = {a , a + 1, . . . , n − b }. Then
J 1 I J 2 = {a + a − 1, a + a , . . . , n − b − b }.
Type B n : Let J 1 = {a, a + 1, . . . , n − b}, J 2 = {a , a + 1, . . . , n − b }. Then

{a + a − 1, a + a , . . . , n}, if b = b = 0;
J 1 I J 2 = {a + a − 1, a + a , . . . , n − b − 1}, if b = 0, b 1;

∅, if b, b 1.
Type D n : If I − J 1 = n − 1 or n and I − J 2 = n − 1 or n, then J 1 I J 2 was already calculated in the

previous subsection.
If {n − 1, n} J 1 and {n − 1, n} J 2 and J 1 or J 2 contains at most n − 2 elements, then J 1 I J 2 = ∅.
Otherwise, we may assume that {n − 1, n} ⊂ J 1 , i.e. J 1 = {a, a + 1, . . . , n} for some a. If {n −
1, n} ⊂ J 2 , i.e. J 2 = {a , a + 1, . . . , n} for some a , then J 1 I J 2 = {a + a − 1, a + a , . . . n}.
If {n − 1, n} J 2 , we may assume without loss of generality that n ∈ / J 2 , i.e. J 2 = {a , a + 1, . . . ,
n − b } for some a , b with b 1. Then J 1 I J 2 = {a + a − 1, a + a , . . . , n − b − 1}.

Type E 6 : If 2 ∈
/ J 1 and 2 ∈/ J 2 , then J 1 I J 2 = ∅.
If 2 ∈ J 1 and 2 ∈/ J 2 , then

{4, 6}, if J 1 = I − {1}, J 2 = I − {2},
J 1 I J 2 = {1, 4}, if J 1 = I − {6}, J 2 = I − {2},
∅, otherwise.
If 2 ∈ J 1 ∩ J 2 , then
⎧ {2, 4, 5}, if J 1 = J 2 = I − {1},

⎪
⎪
⎪ {2, 3, 4},
⎪ if J 1 = J 2 = I − {6},
⎪
⎨ {3, 4, 5},
if { J 1 , J 2 } = I − {1}, J 2 = I − {6} ,
J 1 I J 2 =
⎪
⎪ {4}, if { J 1 , J 2 } = I − {1}, I − {1, 6} ,
⎪
⎪
⎪ {4},
⎩ if { J 1 , J 2 } = I − {6}, I − {1, 6} ,
∅, otherwise.
Type E 7 : For proper subsets J 1 and J 2 , we have that
⎧ {2, 5, 7}, if J 1 = J 2 = I − {1},

⎪
⎪
⎪
⎪ {2, 3, 4, 5}, if J 1 = J 2 = I − {7},
⎪
⎨ {3, 4, 5},
if { J 1 , J 2 } = I − {1}, I − {7} ,
J 1 I J 2 =
⎪
⎪ {4}, if { J 1 , J 2 } = I − {7}, I − {1, 7} ,
⎪
⎪
⎪ {
⎩ 4}, if { J 1 , J 2 } = I − {7}, I − {6, 7} ,
∅, otherwise.
Type E 8 : For proper subsets J 1 and J 2 , we have that

{2, 3, 4, 5}, if J 1 = J 2 = I − {8},
J 1 I J 2 =
∅, otherwise.
Type F , H and I : For proper subsets J 1 , J 2 of I , we always have that J 1 I J 2 = ∅.
Acknowledgments
During my visit in University of Oregon, Berenstein told me the result [BK, Proposition 2.30] that
was discovered in his joint work with Kazhdan and asked me if I can find a combinatorial explanation.
It is my pleasure to thank him. Furthermore, I thank the referee for his/her valuable comments and
suggestions, especially for the suggestion of using the language of 0-Hecke algebras and for pointing
out some references on this topics. The computations for exceptional groups were done with the aid
of CHEVIE package [CH] of [GAP].
References
[BK] A. Berenstein, D. Kazhdan, Geometric and unipotent crystals. II. From unipotent bicrystals to crystal bases, in: Quantum
Groups, in: Contemp. Math., vol. 433, Amer. Math. Soc., Providence, RI, 2007, pp. 13–88.
[Bo] N. Bourbaki, Groupes et algèbres de Lie, Ch. 4,5,6, Hermann, Paris, 1968.
[Ca] R.W. Carter, Representation theory of the 0-Hecke algebra, J. Algebra 104 (1) (1986) 89–103.
[CH] M. Geck, G. Hiss, F. Lübeck, G. Malle, G. Pfeiffer, CHEVIE – A system for computing and processing generic character tables
for finite groups of Lie type, Weyl groups and Hecke algebras, Appl. Algebra Engrg. Comm. Comput. 7 (1996) 175–210.
[GAP] Martin Schönert, et al., GAP – Groups, Algorithms, and Programming – version 3 release 4 patchlevel 4. Lehrstuhl D für
Mathematik, Rheinisch Westfälische Technische Hochschule, Aachen, Germany, 1997.
[He] X. He, Minimal length elements in some double cosets of Coxeter groups, Adv. Math. 215 (2007) 469–503.
[HL] X. He, J.-H. Lu, On intersections of G-stable pieces and B − × B-orbits in the wonderful compactification, in preparation.
[HNT] F. Hivert, J.-C. Novelli, J.-Y. Thibon, Yang–Baxter bases of 0-Hecke algebras and representation theory of 0-Ariki–Koike–
Shoji algebras, Adv. Math. 205 (2) (2006) 504–548.
[Lu] G. Lusztig, Hecke Algebras with Unequal Parameters, CRM Monogr. Ser., vol. 18, Amer. Math. Soc., 2003.
[No] P.N. Norton, 0-Hecke algebras, J. Austral. Math. Soc. Ser. A 27 (3) (1979) 337–357.
[St] J. Stembridge, A short derivation of the Möbius function for the Bruhat order, J. Algebraic Combin. 25 (2) (2007) 141–148.
Journal of Algebra
Generic extensions and generic polynomials for semidirect

products
Sebastian Jambor
Lehrstuhl B für Mathematik, RWTH Aachen University, Templergraben 64, D-52062 Aachen, Germany
Article history: This paper presents a generalization of a theorem of Saltman

Received 7 October 2008 on the existence of generic extensions with group A G over
Available online 18 June 2009 an infinite field K , where A is abelian, using less restrictive
Communicated by Gunter Malle
requirements on A and G. The method is constructive, thereby
Keywords:
allowing the explicit construction of generic polynomials for those
Constructive Galois theory groups, and it gives new bounds on the generic dimension.
Kummer theory Generic polynomials for several small groups are constructed.
Generic polynomials © 2009 Elsevier Inc. All rights reserved.
Semidirect products
1. Introduction
Inverse Galois theory is concerned with the question of whether a given finite group G is realizable
as Galois group over some field K (cf. [8]). Once this question is settled, one can go a step further and
ask for a description of all Galois extensions of K with group G. This is done using parametric polyno-
mials, i.e. polynomials f (x1 , . . . , xk , X ) with coefficients in some rational function field K (x1 , . . . , xk ),
such that the splitting field of f over K (x1 , . . . , xk ) has Galois group G and that every G-extension of
K is the splitting field of f (a1 , . . . , ak , X ) for a specialization of f with certain a1 , . . . , ak ∈ K . Usually,
one also requires that these polynomials describe all Galois extensions with group G, where the fixed
field is an arbitrary extension field of K ; in this case, f is called generic. For an excellent reference
for generic polynomials see [3].
In [10], the concept of generic extensions for a group G is introduced, and Ledet [7] showed that
over infinite ground fields K , the existence of a generic G-extension is equivalent to the existence
of a generic G-polynomial. Kemper [4] then showed that every generic polynomial is in fact descent
generic, i.e. every subgroup of G is the splitting field of some specialization.
In this paper, we will prove the following theorem:
E-mail address: sebastian@momo.math.rwth-aachen.de.
doi:10.1016/j.jalgebra.2009.05.040
S. Jambor / Journal of Algebra 322 (2009) 4040–4052 4041
Theorem 1. Let A be a finite abelian group and K an infinite field. Let G be a finite group acting on A by
automorphisms, such that for every prime p the order of the image of G in Aut( A p ) is coprime to p, where
A p denotes the p-Sylow subgroup of A. If there exist generic extensions for G and A over K then there exists a
generic extension for A G over K .
This is a generalization of Saltman’s Theorem [10, Theorem 3.5], where the same result is proved
under the condition that | A | and |G | are coprime. Saltman proves this as an easy corollary of his
theorem which states that a generic A G-extension exists, provided that generic extensions for A
and G exist. While this conclusion is quite elegant and in particular shows the existence of generic
A G-polynomials for certain A and G, it is not trivial to extract those polynomials from the generic
extensions. This extraction has been carried out in [3, Section 5.5] for dihedral groups of prime power
degree (i.e. G = C2 ), but the number of parameters of the generic polynomials thus constructed is
not optimal. The authors of [3] deem it already too involved to construct generic polynomials for
Frobenius groups of prime degree, which is naturally the next family of semidirect products to study
after dihedral groups.
The approach of this paper uses Kummer theory as outlined in Section 2. The proof of Theorem 1
is constructive, and the extraction of a generic A G-polynomial is straightforward. In particular, the
construction of generic polynomials for dihedral groups or Frobenius groups is now an easy task.
In Section 4 we will construct generic polynomials for most of the groups of order 24.
2. Kummer theory
We prove two easy results in Kummer theory, which will be our motivation for the construction
of the generic A G-extensions. In this whole section, let n ∈ N, let K be a field with characteristic
coprime to n, and let L / K be a finite Galois extension such that L contains a primitive nth root
of unity ζn . By the classical Kummer theory, the abelian extensions E / L with exp(Aut E ( L ))|n are in
bijection to the
√ subgroups U L ∗ / L ∗n , where the bijection maps each such subgroup U to the field
n
extension L ( U )/ L. (Cf. [6, VI, §8]. Here, exp( A ) denotes
√ the exponent of the group A, i.e. the least
n
common multiple of orders of elements in A, and L ( U ) is the field extension of L where all nth
roots of elements a ∈ L ∗ with aL ∗n ∈ U are adjoined.)√
n
The first result is a criterion to decide whether L ( U )/ K is Galois. Note that the action of Aut K ( L )
on L ∗ induces an action on L ∗ / L ∗n .
√
Proposition 2. Let U L ∗ / L ∗n . The extension L ( U )/ K is a Galois extension if and only if U is invariant
n
under the action of the Galois group of L / K .
√
Proof. Set G := Aut K ( L ). Let first U be invariant under G. By definition, L ( U ) is the splitting field
n
of PU = {t n − a | a ∈ L ∗ and aL ∗n ∈ U } over L, and this set is invariant under G, i.e. for t n − a ∈ PU and
g ∈ G we have t n − g (a) ∈ PU . Consider the set

n
∗ ∗n
Q := t − b a ∈ L and aL ∈ U ⊆ K [t ],
b∈Ga
√
n
where Ga is the G-orbit of a. Then L ( U ) is the splitting field of Q over L, and it contains the
splitting field E of
√ Q over K , which is therefore Galois over K , as the polynomials in Q are separable.
To prove that L ( U ) = E, we have to show that L ⊆ E; but for a ∈ L ∗ the polynomial t n − an is an
n
element of PU√and hence a divisor of an element of Q, therefore a ∈ E.

n
Now let √L ( U )/ K be Galois, and consider the set Q as above. Then every polynomial in Q has
a zero in L ( U ) and hence splits completely. In particular, the polynomial t n − g (a) splits for every
n
g ∈ G and every aL ∗n , and by the bijection between subgroups of L ∗ / L ∗n and abelian extensions of L
this implies g (a) L ∗n ∈ U . 2
4042 S. Jambor / Journal of Algebra 322 (2009) 4040–4052
√
n
If we know that an extension L ( U )/ K is Galois, we can try to determine the isomorphism type
of the Galois√group. This can be done if the Galois group is known to be a semidirect product of
N := Aut L ( L ( U )) by G := Aut K ( L ). To describe the action of G on N we use the following non-
n
degenerate pairing (cf. [6, VI, §8]):
h(α )
·,· : N × U → ζn : h, xL ∗n
→ for any root α of t n − x.
α
Proposition√ 3. Let U L ∗ / L ∗n be a finite √

subgroup invariant under G := Aut K ( L ). Assume that N :=
Aut L ( L ( U )) has a complement in Aut K ( L ( U )) (which is then isomorphic to G). Let g ∈ G and h ∈ N;
n n
then ghg −1 is determined by

ghg −1 , xL ∗n = g h, g −1 (x) L ∗n for all xL ∗n ∈ U .
√
Proof. Let xL ∗n ∈ U and let α ∈ L ( n U ) be a root of t n − x. Then

ghg −1 (α ) hg −1 (α )
ghg −1 , xL ∗n = =g = g h, g −1 (x) L ∗n .
α g −1 (α )
Since the pairing is non-degenerate, this determines ghg −1 uniquely. 2
Corollary 4. Assume that K contains a primitive nth root of unity. Let U L ∗ / L ∗n be a subgroup isomorphic
to (Z/nZ)k for some k ∈ N, and assume √ that U is invariant under the action of G :=√ Aut K ( L ). Furthermore,
assume that the Galois group of L ( U )/ K is a semidirect product of N := Aut L ( L ( U )) by G. Then N is
n n
isomorphic to U ∗ (the dual of U ) as (Z/nZ)G-module.
The whole construction of the generic extension for A G is motivated by this corollary.
3. Generic extension
3.1. Definitions and notations
Let G be a finite group and K a field.
Definition 5. (See [10, Definition 1.1].) A Galois extension S / R of commutative K -algebras with
group G is called a generic G-extension over K , if
1. R is of the form K [t, 1/t ] for some number d of indeterminates t = (t 1 , . . . , td ) and an element
0 = t ∈ K [t], and
2. whenever K is an extension field of K and L/K is a Galois algebra with group G, there is a
K -algebra homomorphism ϕ : R → K, such that S ⊗ϕ K/K and L/K are isomorphic as Galois
extensions. The map ϕ is called a specialization.
Since a lot of arguments involve extending the scalars of K -algebras to K (ζn ), we adopt the fol-
lowing notation.
Notation. Let n ∈ N and K a field of characteristic coprime to n. Whenever X is a K -algebra, we set

X n := K (ζn ) ⊗ K X , where ζn is a primitive nth root of unity.
3.2. Construction of generic extensions
We have the following analogue to Corollary 4 for Galois algebras.
Lemma 6. Let n ∈ N, and let K be a field of characteristic coprime to n. Let A = (Z/nZ)k for some k ∈ N and
let G be a finite group acting faithfully on A. Define C := Aut K ( K n ). Let T / K be a Galois algebra with group
A G and S := T A .
Then there exist α1 , . . . , αk ∈ T n∗ with αin ∈ S n∗ such that T n = S n [α1 , . . . , αk ]. The group
α1 S n∗ , . . . , αk S n∗ T n∗ / S n∗ is isomorphic to A ∗ as (Z/nZ)G-modules, and for any α S n∗ ∈ α1 S n∗ , . . . , αk S n∗
of order n, the group α S n∗ is isomorphic to ζn as (Z/nZ)C -module.
Proof. Let A i be the ith component of (Z/nZ)k , and let A i be the canonical complement (i = 1, . . . , k).
Then T A i / S and hence T n i / S n is Galois with group A i ∼
A
= Cn . Since S n is a direct sum of isomor-
phic fields, by Hilbert 90 for Galois algebras there exists αi ∈ T n i with T n i = S n [αi ] and αin ∈ S n∗ .
A A
Furthermore, A i acts on S n [αi ] by multiplying αi with roots of unity. Since T n is generated by the
A
subalgebras T n i , we have T n = S n [α1 , . . . , αk ].
For the last statement consider the non-degenerate pairing
σ (α )
·,· : A × α1 S n∗ , . . . , αk S n∗ → ζn : σ , α S n∗
→ .
α
We have

γ γ σ γ −1 (α ) σ γ −1 (α )
σ , α S n∗ = =γ − 1
= γ σ , γ −1 (α ) S n∗ ,
α γ (α )
for any γ ∈ C × G. Since ζn ⊆ K n = S nG we see g σ , α S n∗ = σ , g −1 (α ) S n∗ for all g ∈ G, which

proves that α1 S n∗ , . . . , αk S n∗ is isomorphic to A ∗ as (Z/nZ)G-modules. Now let κ ∈ C and e ∈ N
with κ (ζn ) = ζne . Then we have σ , κ (α ) S n∗ = κ σ , κ (α ) S n∗ = σ , α S n∗ e , and since the pairing is bi-
multiplicative and non-degenerate, we have κ (α ) S n∗ = α e S n∗ , which concludes the proof. 2
Our next goal is to describe the action of G on T n more precisely: we know the images of the αi
only up to elements in S n∗ ; we will therefore choose new generators β1 , . . . , βk of T n where we can
describe the action explicitly. To do this, we make the following assumption: we let n = q for a prime
power q and assume that A = (Z/qZ)k is cyclic as (Z/qZ)G-module. For any submodule M (Z/qZ)G
isomorphic to A ∗ we then have M A ∗ = A ∗ . More precisely, for any x ∈ A ∗ with (Z/qZ)x = A ∗ we have
Mx = A ∗ . Fix such an M and let (m1 , . . . , mk ) be a basis of M such that m1 is a cyclic generator of M;
then m1 · x is again a cyclic generator of A ∗ .
The passage from T q∗ to T q∗ / S q∗ corresponds to the passage from Z to Z/qZ. The idea is therefore
to choose preimages of Z/qZ in Z for M and transfer the resulting relations onto T q∗ : Let Λ : G →
(Z/qZ)k×k : g
→ Λ g be the representation
induced by M with respect to the basis (m1 , . . . , mk ). Fur-
thermore, let Γi ∈ Z/qZ with mi = g ∈G Γi g for all g ∈ G and i ∈ {1, . . . , k}. Choose λ g ∈ Zk×k and
g g
γi ∈ Z such that Λ = λ mod q and Γi = γig mod q for all g ∈ G and i ∈ {1, . . . , k}.
g g g g
Then we have on the one hand
−1 h
gmi = γih gh = γi g h mod q,
h∈G h∈G
on the other hand

k
g

k
g

gmi = λ ji m j = λ ji γ jh h mod q,
j =1 j =1 h∈G
−1 h k
thus γi g = γ jh λ gji + t (i , g , h)q for some t (i , g , h) ∈ Z.
j =1
Finally, let (σ1 , . . . , σk ) ∈ A k be the dual basis of (m1 , . . . , mk ) and C := Aut K ( K q ).
Lemma 7. With the assumptions and notations above, let T / K be a Galois algebra with group A G and set
S := T A . Choose a generator α S q∗ of α1 S q∗ , . . . , αk S q∗ as (Z/qZ)G-module and set a := α q ∈ S q∗ . For κ ∈ C
h
with κ (ζq ) = ζqe let y κ ∈ S q∗ with κ (α ) = α e y κ . Then there exist roots βi ∈ T q of t q − h∈G h(aγi ) such that
T q = S q [β1 , . . . , βk ], and C , G, and A act on T q by
γ h
k
λ
g
g δ
κ (βi ) = βie h yκ i , g (βi ) = β j ji zi , σ j (βi ) = ζq i j βi ,
h∈G j =1
g
where κ ∈ C , g ∈ G, zi = h∈G h(at (i , g ,h) ), and δi j is the Kronecker delta.
h
Proof. Let βi := h∈G h(α γi ) for i = 1, . . . , k; then βi S q∗ = mi (α S q∗ ), i.e.

α1 S q∗ , . . . , αk S q∗ = M α S q∗ = β1 S q∗ , . . . , βk S q∗
and hence T q = S q [β1 , . . . , βk ].

h
Since βi is a root of t q − h∈G h(aγi ), its image under g must be a root of
g −1 h k

g

γ h λ ji q
tq − h a γi = tq − h a j · h at (i , g ,h) .
h∈G j =1 h∈G h∈G
k g
λ ji z g for = 0, . . . , q − 1, so for every g
These roots are ζq j =1 (β j )
i
∈ G and every i ∈ {1, . . . , k} there
k
g g
exists i such that g (βi ) = ζq i λ ji g
g
j =1 (β j ) z i .
Consider the (Z/qZ)G-module (Z/qZ)k ⊕ Z/qZ with basis (ξ1 , . . . , ξk+1 ), where G acts by g ξi =
k g g
j =1 λ ji ξ j + i ξk+1 for i = 1, . . . , k, and g ξk+1 = ξk+1 (i.e. ξi corresponds to βi for i = 1, . . . , k, and
ξk+1 corresponds to ζq ). Then ((Z/qZ)k ⊕ Z/qZ)/ξk+1 ∼ = A ∗ , i.e. there exist μ1 , . . . , μk ∈ Z such
μ
that ξ1 + μ1 ξk+1 , . . . , ξk + μk ξk+1 is a submodule isomorphic to A ∗ . Setting βi := ζq i βi yields the
h
desired result for the action of G. For the action of C just note that κ (βi ) = h∈G h(κ (α )γi ) =
h
(βi )e h∈G h(( y κ )γi ). 2
The last lemma gives us the formulae to construct Galois extensions with group A G, given a
G-extension and a Cq -extension.
Lemma 8. Let R ⊆ S ⊆ U be K -algebras, such that S /R is a G-extension and U /S is a Cq -extension. Assume

that Uq = Sq [θ] with θ ∈ Uq such that v := θ q ∈ Sq ; for κ ∈ C with κ (ζq ) = ζqe let y κ ∈ Sq with κ (θ) = θ e y κ .
q h
Set Tq := Sq [θ1 , . . . , θk ], where θi = h∈G h( v γi ); define the action of C , G, and A on Tq by
γ h
k
λ
g
g
κ (θi ) = θie h yκ i , g (θi ) = θ j ji zi , σ j (θi ) = ζ δi j θi ,
h∈G j =1
g
where κ ∈ C , g ∈ G, zi = h∈G h( v t (i , g ,h) ), and δi j is the Kronecker delta.
Then Tq /R is a C × ( A G )-extension.
k i
Proof. Let s := ∈(Z0 )k s θ be an arbitrary element in Tq , where s ∈ Sq and θ :=

i =1 θi , for
k g
all ∈ (Z0 )k . Define g (s) := ∈(Z0 )k g (s ) g (θ) , where g (θ) := i =1 g (θi ). By the definition of zi
q
we have g (θi )q = g (θi ), so g (s) is well defined. We show that this defines an action of G on Tq : let
k g h gh
g , h ∈ G and i ∈ {1, . . . , k}. Since Λ is a representation, we have j =1 λ j λ ji = λ i + s(i , , g , h )q for
q h
some s(i , , g , h) ∈ Z. We have to prove g (h(θi )) = ( gh)(θi ). Set b i := θi = h∈G h( v γi ). On the one
hand we have
k
k λ g j λhji g λh k
λ s(i , , g ,h) g λhji h
gh k k
g h(θi ) = θ z j ji g zhi = θ i b zj g zi ,
j =1 =1 =1 =1 j =1
k λ i gh
gh
on the other hand we have ( gh)(θ j ) = =1 θ zi , hence we have to show

k
s(i , j , g ,h) g λ ji h
h
gh
bj zj g zi = zi .
j =1
This amounts to show

k
α
γ j s(i , j , g , h) + t ( j , g , α )λhji + t i , h, g −1 α = t (i , gh, α ) (1)
j =1
for all α ∈ G. But

k −1 α k

k
j =1 γ jg λhji − =1 γ
α λ gh
i

k
t ( j, g , α )λhji = − γ jα s(i , j , g , h)
q
j =1 j =1
and
−1 α k −1 α k −1 α
−1
γi( gh) − j =1 γ jg λhji j =1 ( jγ α λ gh
ji − γ j
g
λhji )
t i , h, g α = = t (i , gh, α ) + ,
q q
which proves (1). We show that the actions of A and G on Tq result in an A G-extension Tq /Rq .
k
For i , j ∈ {1, . . . , k} and g ∈ G choose μ ji ∈ Z such that g −1 σi g =
g g
j =1 μ ji σ j ; since A acts trivially
k
on Sq , it is enough to prove that σi g and g j =1 μ gji σ j act in the same way on θ for ∈ {1, . . . , k}.
We have on the one hand
g
k
λ
g
g
σi g (θ ) = ζ λi θ j j z ,
j =1
on the other hand

k
g g g
k
λ
g
g
g μ σ j (θ ) = g ζ μ i θ = ζ μ i
ji
θ j j z .
j =1 j =1
g g
Since (σ1 , . . . , σk ) is the dual basis of (m1 , . . . , mk ), we have λ i ≡ μi mod q.
Finally, it is easy to verify that our definition gives an action of C on Tq and that the actions of C
and A G commute, which finishes the proof of the lemma. 2
Remark. Whenever there exists a generic Cq -extension over K , we can choose one of the form
U /S such that Uq = Sq [θ] for some θ ∈ Uq with θ q ∈ Sq , as in the last lemma. The proof is al-
most identical to the proof that a generic extension can be chosen to have a normal basis, using the
equivalence of the existence of generic G-extensions over K and the retract-rationality of the exten-
sion K (t 1 , . . . , tn )/ K (t 1 , . . . , tn )G , where G acts faithfully and transitively on {t 1 , . . . , tn } (cf. [3, Remark,
p. 100] and [10, Corollary 5.4]).
Note that the generic Cq -extensions constructed by Saltman [10] already are of this form.
Now letA be any finite abelian group. For every prime p let A p be the p-Sylow subgroup of A.
Then A ∼= p prime A p and Aut( A ) ∼ = p prime Aut( A p ), so if a group G is acting on A, this action
induces actions on the groups A p . We can get a finer decomposition of A: For each prime p, the
group A p is a Z p G-module, where Z p G is the group ring of G over the ring of p-adic integers Z p .
We assume p G; then Z p G is a direct sum of full matrix rings of unramified extensions of Z p (cf. [2,
Satz 11.1], [1, Proposition 2.2.28]), i.e.

Zp G ∼
n(i )×n(i )
= Ri ,
i =1
where n(i ) ∈ N, and R i is an unramified extension of Z p , for i = 1, . . . , . By multiplying A p with the

corresponding central idempotents, it suffices to analyze finite R n×n -modules, where R is an unram-
ified extension of Z p and n ∈ N. Using Morita equivalence and the Fundamental Theorem of finitely
generated modules over PIDs we conclude

m
ki
Ap ∼
= Z/ p ei Z ,
i =1
for some m, ni , ki ∈ N, where each (Z/ p e i Z)ki is an irreducible Z p G-module.

We are now able to prove Theorem 1.
Proof of Theorem 1. Let S /R be a generic G-extension over K . We can assume that S /R has a
normal basis (cf. [3, p. 105]), and thus S N /R has a normal basis for any normal subgroup N P G.
We assume first that A is of the form A ∼ = (Z/qZ)k for some prime power q and some k ∈ N, such
that A is cyclic as (Z/qZ)G-module. Let N P G be the kernel of the action of G on A. Let V /U be a
generic Cq -extension over K ; by the previous remark we can assume that Vq = Uq [θ] for some θ ∈ Vq
with θ q ∈ Uq .
By the definition of generic extensions, R is of the form R = K [r1 , . . . , rm , 1/r ] and U is of
the form U = K [u 1 , . . . , un , 1/u ]. Let u = f (u 1 , . . . , un ) for some polynomial f . Let (s1 , . . . , s )
y = ( y 11 , . . . , yn ) over S and set ρ :=

S /R; choose n · indeterminates
N
be a free basis of

f ( j =1 y 1 j s j , . . . j =1 ynj s j ) ∈ S N [y]. Then ρ :=
g ∈G / N g (ρ ) ∈ R [ ] = K [r, , 1/r ], where r =
y y

(r1 , . . . , rm ). Set R:= R[y, 1/ρ ] and S := S [y, 1/ρ ], and define a homomorphism
ϕ : U → (S )N : u i
→ j=1 y i j s j ; we denote the extension of ϕ to Uq → (S )qN again by ϕ . Set
V := V ⊗ϕ (S )N ; then V /S is a Cq -extension, and Vq = (S )qN [ θ ], where θ q = ϕ (θ q ). Set v := ϕ (θ q ),
and define θ1 , . . . , θk as in Lemma 8 to get a C ×( A G / N )-extension (S )q [θ1 , . . . , θk ]/ R . Then Tq :=
N
Sq [θ1 , . . . , θk ]/ R is a C × ( A G )-extension, where N acts trivially on the θi . Set T := Sq [θ1 , . . . , θk ]C .
We claim that T /R is a generic A G-extension.
Let T / K be an A G-extension (by Jensen et al. [3, Proposition 1.1.5] it suffices to consider
A G-extensions of K instead of extension fields L ⊇ K ); set S := T A . There exists a specialization
ψ : R → K such that S ⊗ψ K / R ⊗ψ K ∼ = S / K , and ψ(s1 ), . . . , ψ(s ) is a K -basis of S N .
Choose α ∈ ( T qN )∗ as in Lemma 7 and let L := S qN [α ]C . Then L / S N is a Cq -extension, so there
exists a specialization χ : U → S N such that V ⊗χ S N / S N ∼ = L / S N . There exist zi j ∈ K with χ (u i ) =
∼
j =1 z i j ψ(s j ), and extending the map ψ to R by y i j
→ z i j we get an isomorphism Tq ⊗ψ K / K =
T q / K which maps θi to βi , i.e. an isomorphism of C × ( A G )-extensions. Restricting to the fixed
algebras, we get T ⊗ψ K / K ∼ = T / K as A G-extensions.

Now let A be arbitrary; we have A ∼ = p prime A p as G-module. Let p be a prime and N P G
the kernel of the action of G on A p . Then there exist n1 , . . . , nm ∈ N and k1 , . . . , km ∈ N such that
m
Ap ∼ = i =1 (Z/ pni Z)ki as G / N-module and such that (Z/ pni Z)ki is cyclic as (Z/ pni Z)G / N-module.
The process above can be carried out for each of those cyclic submodules to give a generic A G-
extension. 2
3.3. Generic polynomials and generic dimension
The generic dimension of a group G over a field K , denoted by gd K G, is the minimal num-
ber of parameters in a generic G-polynomial over K , or ∞ if no generic polynomial exists (cf. [3,
Section 8.5]). The following bounds can be derived from Saltman’s results about generic extensions
for semidirect products A G: If G is a finite group acting on the finite abelian group A by au-
tomorphisms with kernel N P G, and if K is an infinite field and |G | and | A | are coprime, then
gd K ( A G ) gd K (G ) + [G : N ] gd K ( A ) (cf. [3, Proposition 8.5.6]). Using the construction in Theorem 1,
these bounds can be considerably improved:
Let A be a finite abelian group and G a group acting on A, such that for every prime p the image
of G in Aut( A p ) has order coprime to p. Let p p be the exponent of A p and N p P G the kernel of
p
the action on A p ; then there exist k1 , . . . , k p such that A p ∼
(Z/ p i Z)ki as Z p (G / N p )-modules.
= i =1
For every i, let γ ( p , i ) denote the minimal number of generators of (Z/ p i Z)ki as (Z/ p i Z)(G / N p )-
modules.
Corollary 9. Let A be a finite abelian group and G a group acting on A, such that for every prime p the image
of G in Aut( A p ) has order coprime to p. For every prime p, define N p , p , and γ ( p , i ) as above. Then
p

gd K ( A G ) gd K (G ) + [G : N p ] γ ( p , i ) gd K (C p i ).
p prime i =1
Saltman gives an explicit construction of generic Cq -extensions for prime powers q with 8 q (cf.
[10, Proposition 2.6]), and Jensen, Ledet, and Yui use these extensions to construct generic polynomials
in ϕ (q)/2 parameters for odd q (cf. [3, Proposition 5.3.4]). This allows us to construct generic A G-
polynomials over Q:
Corollary 10. Let A be a finite abelian group with 8 exp( A ). Let G be a group acting on A, such that for every
prime p the image of G in Aut( A p ) has order coprime to p. For every prime p, define N p , p , and γ ( p , i ) as
above. Let R = K [r1 , . . . , rm , 1/r ] and let S /R be a generic G-extension with a normal basis. Then a generic
( A G )-polynomial over Q with
p

2
ϕ( pi )
m + [G : N 2 ] γ (2, i )ϕ 2i + [G : N p ] γ ( p, i)
2
i =1 p prime i =1
p 3
parameters can be effectively constructed.
Proof. It suffices to consider the case where A is of the form A ∼ = (Z/qZ)k for some prime power q
and some k ∈ N, such that A is cyclic as (Z/qZ)G-module. In the general case we can take a product
of the generic polynomials. Let N P G be the kernel of the action of G on A. Let T / K be a Galois
algebra with group A G. Choose α and βi in T qN as in Lemma 7. Then ( S qN [β1 ])C / S N is Galois with
group Cq . Following the argument in Jensen et al. [3, p. 103] we see that there exists j ∈ {1, . . . , q − 1}
with ( j , q) = 1 and a specialization ϕ : T → K such that θ1 maps to β1j and TrTq /T (θ1 ) maps to a
primitive element of ( S qN [β1 ])C / S N . Using the pairing in Lemma 6, we see that the Galois closure
of ( S qN [β1 ])C / K is T / K , since β1 S q∗ = m1 α S q∗ , and we chose m1 as cyclic generator. Thus the prod-
uct of the minimal polynomial of TrTq /T (θ1 ) and a (suitable) generic polynomial for G is a generic
polynomial for A G. 2
Remark. The generic polynomials can be simplified in special cases.
1. Assume that G acts faithfully on the cyclic (Z/qZ)G-module A ∼

= (Z/qZ)k , where q is a prime
j
power and k ∈ N. By replacing α by α j and β1 by β1 in the situation above we can assume that
j = 1, i.e. Tr(β1 ) is a primitive element of ( S q [β1 ])C / S, where S := T A . We claim that the Galois
closure of K [Tr(β1 )]/ K is T , i.e. the minimal polynomial of TrTq /T (θ1 ) is generic (i.e. there is no
need for the additional generic polynomial for G in the proof above):
Let A A be the orthogonal complement of β1 S q∗ under the pairing in Lemma 6, then
S q [β1 ] = T qA and hence S [Tr(β1 )] = T A . Thus K [Tr(β1 )] = T B for some A B A G, and since

q|[ K [Tr(β1 )] : K ] we
have B ∩ A = A. The Galois closure of T is T , where B is the core of B
B B
in A G, i.e. B = x∈ A G B x ; we show that B is trivial. Since β1 S q∗ = m1 α S q∗ and we chose m1

as cyclic generator, we see that the normal closure of β1 S q∗ in A ∗ G is A ∗ . Since the pairing
respects the G-action, we get that the normal core of A is trivial, hence B intersects A trivially.
The group A B splits over A, so B is a subgroup of G. But G acts faithfully on A, i.e. B = 1.
2. Now assume A = Z/qZ for an odd prime power q = pn and G acts faithfully on A. In (1) above
we saw that we can get an irreducible generic polynomial of degree |G | · q; now, we make some
further reductions which will give a generic polynomial of degree q.
Since G is isomorphic to a subgroup of Aut(Cq ), it is cyclic of order , generated by an ele-
q −1 −2
ment g ∈ G. We can choose the γih such that θ1 = v k g ( v k ) · · · g −1 ( v ) =: Ψ ( v ) in Lemma 8
for some k ∈ N. Furthermore, we can choose V /U as the generic Cq -extension constructed by
Saltman: Let d = ϕ (q), let e ∈ N be of order pd modulo pq and choose a generator κ of
d−1
AutQ (Q(ζq )). Set x := u 1 + u 2 (ζq + 1/ζq ) + · · · + ud/2 (ζq + 1/ζq )d/2 + (ζq − 1/ζq ) and u := i =0 κ (x).
d−1 d −2
Now let U := Q[u 1 , . . . , ud/2 , 1/u ] and Vq := Uq [θ] with θ q = xe ) · · · κ d−1 (x) =: Φ(x).
κ (xe
κ
Then Vq / U is a generic Cq -extension (cf. [3, Section 5.3]). Next, let S / R be a generic
G-extension with free basis (s1 , . . . , s ), and replace each u i by j y i j s j , so from x we get
−1 d−1 i j
X := ( j y 1 j s j ) + · · · + ( j yd/2, j s j )(ζq + 1/ζq )d/2 + (ζq − 1/ζq ). Set ρ := i =0 j =0 g (κ ( X )),
κ
R := R[y, 1/ρ ], and S := S [y, 1/ρ ], and let Tq := Sq [θ1 ], where θ1 = Ψ (Φ( X )). Then Tq / R
q
is a generic Cq G-extension. Let s ∈ S be a generator for a normal basis of S /R . Then the
minimal polynomial of TrTq /T G (θ1 s) is generic:
The argument is analogous to the proof of Jensen et al. [3, Proposition 5.3.4]: The extension
Tq /R is generated by {ζ i θ1 g m (s) | 0 i < ϕ (q), 0 j < q, 0 m < }, so T /R is generated by
j
their traces. We only have to consider the cases i = 0 and (q, j ) = 1, since the other traces are
either conjugate to one of those or lie in a subextension. If T / K is a Cq G-extension and S :=
q j
T G , then T q = S q [β1 ] with β1 = Ψ (Φ(b)) for some b ∈ S, and some element Tr T q / T G (θ1 g m (s)) =
Tr T q / T G ( g −m (θ1 )s) specializes to a primitive element. We have g −m (Ψ (Φ(b)) j ) = Ψ (Φ( g −m (b j ))),
j
so by sending X to g −m (b j ) we get the desired result.
Remark. The theory developed here gives an interpretation for the element M τ (b) in [10, Theo-
rem 2.3] or Φ(b) in [3, Section 5.3], which is used to characterize cyclic extensions of prime power
degree q: it is an image of an element in (Z/qZ)C which generates a submodule isomorphic to the
(Z/qZ)C -module ζq (where C := Aut K ( K (ζq )), and K is the base field).
4. Examples
Example 11 (S3 = D6 = C3 C2 ). A generic C2 -extension is given by R = K [r , 1/r ] and S := R [α ]

with α 2 = r, where C2 acts by changing the sign, and s := 1 − α generates a free basis. We have
X := y 1 + y 2 α + ζ − ζ 2 and θ13 = Ψ (Φ( X )) = X 4 κ ( X 2 ) g ( X 2 )κ ( g ( X )). We replace θ1 by θ1 / X to remove
the fourth power, and get the trace t of sθ as
θ 2 g (s) θ 2s
t = θs + + + θ g (s).
g (x1 )κ (x1 ) g (x1 )κ (x1 )
The generator σ of C3 simply acts on the summands by multiplication with roots of unity, and we
can calculate the minimal polynomial of t as

μ := X 3 − 12 A 2 + 12 y 21 X + 16( A − 6) A 2 + 12 y 21 ,
where A := r y 22 − y 21 + 3.
The easy example of the generic S 3 -polynomial allows us to construct generic polynomials for
groups of order 24.
Example 12 (Groups of order 24). There are 15 groups of order 24, and it is known for all of them
whether generic polynomials over Q exist (cf. [3, Exercise 7.3]). We are now able to actually compute
generic polynomials for those groups.
For C3 × C8 ∼ = C24 and C3 C8 there are no generic polynomials over Q. For SL(2, 3), Rikuna [9]
proved that the invariant field of a four-dimensional representation is purely transcendental, so a
generic polynomial can be constructed, e.g. using the methods of Kemper and Mattig [5].
If the group is a direct product (i.e. D8 × C3 , Q8 × C3 , C4 × C6 , V4 × C6 , S3 × C4 , S3 × V4 ∼
= D12 × C2 ,
(C3 C4 ) × C2 , and C2 C3 ∼
= A4 × C2 ), a generic polynomial can be constructed by taking a product
of generic polynomials for each factor. Generic polynomials for S4 are well known, so we are left to
deal with the groups C3 Q8 , G 1 := C3 D8 , where the kernel of the action is C4 , and G 2 := C3 D8 ,
where the kernel of the action is V4 .
We start with C3 Q8 . Let
4 A+B+C 2 (1 − r1 r2 r3 )3
F (r1 , r2 , r3 , X ) := X 2 − 1 − 2(1 − r1 r2 r3 )2 X2 − 1 −8 X2 − 1
A BC A BC
A 2 + B 2 + C 2 − 2 A B − 2 AC − 2BC
+ (1 − r1 r2 r3 )4 ,
A2 B 2 C 2
−1/2
where A := 1 + r12 + r12 r22 , B := 1 + r22 + r22 r32 , and C := 1 + r32 + r12 r32 , and set μ1 := r44 F (r1 , r2 , r3 , r4 X ).
Then μ1 is a generic Q8 -polynomial [3, Theorem 6.1.12], and a quadratic subextension is parametrized
by the polynomial X 2 − (1 + r12 + r12 r22 )(1 + r22 + r22 r32 ). By Example 11,

μ2 := X 3 − 12 A 2 + 12 y 21 X + 16( A − 6) A 2 + 12 y 21
with A := (1 + r12 + r12 r22 )(1 + r22 + r22 r32 ) y 22 − y 21 + 3 is a ‘generic polynomial’ for the S3 -subextension.
Thus μ1 μ2 is a generic C3 Q8 -polynomial.
Now we consider the semidirect products C3 D8 . Let
μ1 := X 4 − 2r1 r2 X 2 + r12 r2 (r2 − 1) ∈ Q(r1 , r2 , X ).

√
Then μ1 is a generic D8 -polynomial;
√ furthermore, if F /Q(r1 , r2 ) is the splitting field, then F /Q( r2 )
is a V4 -extension, and F /Q( r2 − 1 ) is a C4 -extension (cf. [3, Theorem 2.2.7 and Corollary 2.2.8]). Set

μ2 := X 3 − 12 A 2 + 12 y 21 X + 16( A − 6) A 2 + 12 y 21 ,

μ3 := X 3 − 12 B 2 + 12 y 21 X + 16( B − 6) B 2 + 12 y 21 ,
where A := r2 y 22 − y 21 + 3 and B := (r2 − 1) y 22 − y 21 + 3, then μ1 μ2 is generic for G 2 -extensions, and

μ1 μ3 is generic for G 1 -extensions.
As mentioned in the introduction, it is now quite simple to describe an algorithm to compute

generic polynomials over Q for Frobenius groups C p C , where C acts faithfully on C p and 8 .
However, the actual computation of the polynomials is practically infeasible, as there is no computer
algebra system known to the author with an efficient method for computations in the rational func-
tion field Q( z1 , . . . , zk ) and algebraic extensions thereof.
Instead, we will use the theory developed here to construct single polynomials having a prescribed
semidirect product as Galois group.
Example 13 (Polynomials with prescribed Galois groups). 1. First, we set G := C4 and construct polyno-
mials with Galois group Fk5 C4 for several k ∈ N. Let L /Q := Q(ζ5 )/Q and let g be the generator of
the Galois group C4 which maps ζ5 to ζ52 .
The subgroup ζ5 L ∗5 L ∗ / L ∗5 is invariant under C4 , so any root of t 5 − ζ5 generates a field ex-
tension E /Q with Galois group C20 . In fact, E = Q(ζ25 ).
The F5 C4 -submodule of L ∗ / L ∗5 generated by (1 + ζ5 ) L ∗5 is (1 + ζ5 ) L ∗5 , (1 + ζ52 ) L ∗5 , and g acts
0 2
with the matrix . Thus the Galois group of
11

3

t 5 − g i (1 + ζ5 ) = t 20 − 3t 15 + 4t 10 − 2t 5 + 1
i =0
1 0
is F25 C4 , where g acts on F25 via , by Proposition 3.
03
Similarly, the F5 C4 -submodule of L ∗ / L ∗5 generated by (1 − ζ5 ) L ∗5 is

(1 − ζ5 ) L ∗5 , 1 − ζ52 L ∗5 , 1 − ζ54 L ∗5 ,
0 0 3
and g acts with the matrix 10 1 , so we know the Galois group of t 20 − 5t 15 + 10t 10 − 10t 5 + 5,
012
1 0 0
namely the semidirect product F35 C4 , where g acts by 02 0 .
003
Last, consider the element x := (1 + ζ5 − ζ52 ); then xL ∗5 generates a submodule isomorphic to

the regular module. The element 1 − g 2 ∈ F5 C4 is a cyclic generator of the two-dimensional self-
dual faithful F5 C4 -module, so any root of t 5 − x/ g 2 (x) generates a field extension E /Q whose Galois
1 0
closure has Galois group F25 C4 , where g acts on F25 by ; the minimal polynomial of the roots
04
1 15
of t 5 − x/ g 2 (x) over Q is t 20 + 11 t − 19
11
1 5
t 10 + 11 t + 1.
2. As a second example, we construct a polynomial with Galois group F43 D8 , where D8 := a, b |
a4 , b2 , (ab)2 is the dihedral group of order 8, and the action of D8 on M := F43 is defined by
⎛ ⎞ ⎛ ⎞
1 0 0 0 2 0 0 0
⎜0 2 0 0⎟ ⎜0 1 0 0⎟
a
→ ⎝ ⎠, b
→ ⎝ ⎠.
0 0 0 1 0 0 0 1
0 0 2 0 0 0 1 0
√
Let S := Q( 4√−2, i ) be√the splitting field of √ t 4 + 2 over
√ Q; its Galois group is generated by the ele-
ments α = ( −2
→ i −2, i
→ i ) and β = ( 4 −2
→ 4 −2, i
→ −i ), and it is isomorphic
4 4
√ to D8 .
Now let ζ3 be a primitive third root of unity and consider the element x := 4 −2 + i + ζ3 ∈ S (ζ3 )∗ .
Then xS (ζ3 )∗3 generates an eight-dimensional submodule of the F3 D8 -module S (ζ3 )∗ / S (ζ3 )∗3 . The
element 1 + 2ba ∈ F3 D8 generates a submodule of F3 D8 isomorphic to M, thus y := xβ(α (x2 )) S (ζ3 )∗3
generates a submodule of S (ζ3 )∗ / S (ζ3 )∗3 isomorphic to M.
Set z := y 2 κ ( y ), where κ is a generator of AutQ (Q(ζ3 )) and let θ be a root of t 3 − z. Then
S (Tr S (ζ3 ,θ)/ S (ζ3 ,θ)κ (θ))/ S is a C3 -extension (cf. [3, Section 5.3]). The Galois closure U /Q(ζ3 ) of
S (ζ3 , θ)/Q(ζ3 ) has Galois group M D8 , since M is self-dual, thus the Galois closure T /Q of
S (Tr S (ζ3 ,θ)/ S (ζ3 ,θ)κ (θ))/Q has Galois group M D8 . Since we know the action of the Galois group
on U /Q, we can compute the minimal polynomial of Tr S (ζ3 ,θ)/ S (ζ3 ,θ)κ (θ) as
t 24 − 144t 22 + 472t 21 + 7524t 20 − 30456t 19 − 266608t 18 − 981864t 17 + 30277458t 16
+ 9496600t 15 − 1093991688t 14 − 1140063288t 13 + 30808510272t 12

+ 31632046632t 11 − 495311379648t 10 − 865959612792t 9 + 5149493226585t 8
+ 14478424454376t 7 − 28713293762728t 6 − 144781282966176t 5
− 41870309411988t 4 + 619972209753552t 3 + 1309616138896848t 2
+ 1104816334207968t + 352192366019556.
Acknowledgments
I would like to express my gratitude to Professor Wilhelm Plesken for many invaluable discussions
and helpful comments. Furthermore, I would like to thank the anonymous referee for pointing out
additional references (especially concerning the existence of generic SL(2, 3)-polynomials over Q) and
a wrong citation in the proof of the theorem.
References
[1] D.F. Holt, W. Plesken, Perfect Groups, Oxford Math. Monogr., The Clarendon Press, Oxford University Press, New York, 1989,
with an appendix by W. Hanrath, Oxford Science Publications.
[2] H. Jacobinski, Maximalordnungen und erbliche Ordnungen, Vorlesungen aus dem Fachbereich Mathematik der Universität
Essen (Lecture Notes in Mathematics at the University of Essen), vol. 6, Universität Essen Fachbereich Mathematik, Essen,
1981.
[3] C.U. Jensen, A. Ledet, N. Yui, Generic Polynomials: Constructive Aspects of the Inverse Galois Problem, Math. Sci. Res. Inst.
Publ., vol. 45, Cambridge University Press, Cambridge, 2002.
[4] G. Kemper, Generic polynomials are descent-generic, Manuscripta Math. 105 (1) (2001) 139–141.
[5] G. Kemper, E. Mattig, Generic polynomials with few parameters, in: Algorithmic Methods in Galois Theory, J. Symbolic
Comput. 30 (6) (2000) 843–857.
[6] S. Lang, Algebra, 3rd edition, Grad. Texts in Math., vol. 211, Springer-Verlag, New York, 2002.
[7] A. Ledet, Generic extensions and generic polynomials, in: Algorithmic Methods in Galois Theory, J. Symbolic Comput. 30 (6)
(2000) 867–872.
[8] G. Malle, B.H. Matzat, Inverse Galois theory, Springer Monogr. Math., Springer-Verlag, Berlin, 1999.
[9] Y. Rikuna, The existence of a generic polynomial for SL(2, 3) over Q , preprint, 2004.
[10] D.J. Saltman, Generic Galois extensions and problems in field theory, Adv. Math. 43 (3) (1982) 250–283.
Journal of Algebra
Dendriform equations
Kurusch Ebrahimi-Fard a,∗ , Dominique Manchon b
a
Laboratoire MIA, Université de Haute Alsace, 4 rue des Frères Lumière, 68093 Mulhouse, France
b
Université Blaise Pascal, C.N.R.S.-UMR 6620, 63177 Aubière, France
Article history: We investigate solutions for a particular class of linear equations

Received 15 October 2008 in dendriform algebras. Motivations as well as several applications
Available online 10 July 2009 are provided. The latter follow naturally from the intimate link
Communicated by Jean-Yves Thibon
between dendriform algebras and Rota–Baxter operators, e.g. the
Keywords:
Riemann integral map or Jackson’s q-integral.
Linear differential equation © 2009 Elsevier Inc. All rights reserved.
Linear integral equation
Lie bracket flow
Riccati equation
Magnus expansion
Fer expansion
Dendriform algebra
Pre-Lie algebra
Hopf algebra
Rota–Baxter algebra
Planar rooted trees
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4054
2. Dendriform power sums expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4056
3. Power sums expansions in unital dendriform algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4058
3.1. Equation of degree (1, 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4058
4. The pre-Lie Magnus expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4063
5. An equation of degree (2, 0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4064
5.1. Matrix dendriform algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4065
5.2. Transformation into a dendriform system of degree (1, 0) . . . . . . . . . . . . . . . . . . . . . . . . . 4065
*
E-mail addresses: kurusch.ebrahimi-fard@uha.fr (K. Ebrahimi-Fard), manchon@math.univ-bpclermont.fr (D. Manchon).
URLs: http://www.th.physik.uni-bonn.de/th/People/fard/ (K. Ebrahimi-Fard), http://math.univ-bpclermont.fr/~manchon/
(D. Manchon).
doi:10.1016/j.jalgebra.2009.06.002
4054 K. Ebrahimi-Fard, D. Manchon / Journal of Algebra 322 (2009) 4053–4079
6. Higher-order equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4067

7. Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4068
7.1. Rota–Baxter algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4069
7.2. A link with generalized initial value problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4070
7.3. A link with the Riccati differential equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4071
7.4. Vogel’s identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4072
7.5. Rooted trees and the coefficients in the pre-Lie Magnus expansion . . . . . . . . . . . . . . . . . . 4073
8. A concluding remark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4077
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4078
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4078
1. Introduction
This paper sets out to develop a systematic study of a particular class of linear equations in Lo-
day’s dendriform algebras [Lod01,Ag04]. The guiding principle follows from the intimate connection
of such algebras to associative Rota–Baxter algebras [Bax60,Rot69,Ag00,Eb02], where analog equations
naturally appear in the context of applications ranging from numerical analysis to renormalization in
perturbative quantum field theory. The reader interested in more details may want to consult some
of the following references [At63,EGP07,EMP07,EMP07b].
In the setting of dendriform algebras we obtain the recursive formula for the logarithm of the
solutions of the two linear dendriform equations Y = 1 + λa ≺ Y and Z = 1 − λ Z a, a ∈ D, in D [[λ]],
where ( D , ≺, ) is a unital dendriform algebra. This way, our work provides a refined approach to
the classical Magnus expansion [Mag54], well-known in the field of numerical analysis in the context
of approximations of ordinary first-order linear differential and integral equations, e.g. see [Ise84,
Ise02,BCOR98,BCOR09]. In our approach this expansion involves the pre-Lie and associative products
naturally associated with the underlying dendriform structure and displays their interesting interplay.
Let us emphasize that the language of dendriform and pre-Lie algebras unveils a hitherto unobserved
new structure in these expansions, related to operadic aspects in the context of free pre-Lie algebra
[ChaLiv01,Cha08]. Moreover, the two linear equations above may be interpreted as first-order cases
with respect to the parameter λ. We will introduce equations of arbitrary order in λ and show by
transforming the higher-order equations into dendriform matrix equations that its solutions follow
from the first-order ones.
Putting the two dendriform products together into a single equation:
X = a + λ X b + λc ≺ X , (1)
we find a natural link to Lie bracket flow equations. However, for general dendriform algebras the Lie
bracket is replaced by a pre-Lie product. We will show how the solutions Y = Y (c ) and Z = Z (b) of
the first two recursions above naturally lead to the solution of the general equation (1):

X = Y ∗ Y −1 a ≺ Z −1 ∗ Z . (2)
Here, the product ∗ =≺ + is the associative product in the dendriform algebra D. In fact, we ob-
serve that in the special case where −b = c, the solution X can be represented in terms of two
non-commuting group actions.
Eventually our approach brings together the works of Magnus [Mag54], Fer [F58] and Baxter
[Bax60] on linear initial value problems and the corresponding integral equations. A simple exam-
ple might help to elucidate this point of view. Details will be provided in the sequel. Let k be a field
of characteristic zero and F an—associative—k-algebra, say, of operator-valued functions on the real
line, e.g. smooth n × n matrix-valued functions. Recall that the solution to the initial value problem:
K. Ebrahimi-Fard, D. Manchon / Journal of Algebra 322 (2009) 4053–4079 4055

Ẏ (t ) = Y (t ), A (t ) = (Y A )(t ) − ( AY )(t ), t > 0, Y (0) = Y 0 ,
for A ∈ F , can be written in terms of the solution of the initial value problem U̇ (t ) = A (t )U (t ),
U (0) = 1, i.e. Y (t ) = − Ad U (t ) (Y 0 ) =: −U (t )Y 0 U −1 (t ). Following Magnus [Mag54], see also [IMNZ00],
we can write U (t ) = exp(Ω( A )(t )), where Ω( A )(0) = 0 and:
adΩ
Ω̇( A )(t ) = ( A )(t ),
exp(adΩ ) − 1
which implies:
Y (t ) = − Adexp(Ω( A )(t )) (Y 0 ).
Here, we may remark that Magnus’ exponential solution gives rise to the following nontrivial identity:

T exp I ( A )(t ) := 1 + I ( A )(t ) + I A I ( A ) (t ) + I A I A I ( A ) (t ) + · · · = exp Ω( A )(t ) ,
t
I ( X )(t ) = 0 X (s) ds denotes the Riemann integral on F . The expression on the left-hand side of the
first equality is known as the time-ordered exponential, i.e. the T -operation denotes the time order-
ing operator [OR00]. We may rewrite the Lie bracket flow equation for A ∈ F , as follows:

Ẏ (t ) = I (Ẏ )(t ) + Y 0 , A (t ) = Y 0 , A (t ) + I (Ẏ )(t ), A (t ) = A 0 (t ) + I (Ẏ )(t ), A (t ) . (3)
The crucial observation is that recursion (3) actually takes place inside a pre-Lie algebra. Indeed, the
integration by parts rule implies that the product defined in terms of the Riemann integral and the
ordinary Lie bracket, i.e. X Z := [ I ( X ), Z ] satisfies the—left—pre-Lie identity, leading to the pre-Lie
flow equation:

Ẏ (t ) = A 0 (t ) + (Ẏ A )(t ), Ẏ (0) = A 0 (0) = Y 0 , A (0) . (4)
One may interpret this as a linear initial value problem on a pre-Lie algebra. When expanding the Lie
bracket, [ I ( X ), Z ] = I ( X ) Z − Z I ( X ), the newly introduced pre-Lie product, X Z = [ I ( X ), Z ] on F , can
be written as a linear combination of two binary compositions, X Z := I ( X ) Z , and Z ≺ X := Z I ( X ),
which we call left- and right-dendriform products, respectively:
Ẏ (t ) = A 0 (t ) + (Ẏ A )(t ) − ( A ≺ Ẏ )(t ), Ẏ (0) = A 0 (0). (5)
The rules these left- and right-dendriform compositions have to satisfy so as to provide a pre-Lie
product are called dendriform axioms [Lod01]. In the next section we will present them in detail.
Iserles in [Ise01] further generalized the above Lie bracket flow by looking at the initial value problem
Ẏ = Y A − B Y , t > 0, Y (0) = Y 0 which can be written:

Ẏ (t ) = Y A (t ) − B Y (t ) = I (Ẏ )(t ) + Y 0 A (t ) − B (t ) I (Ẏ )(t ) + Y 0

= C (t ) + I (Ẏ ) A (t ) − B I (Ẏ ) (t ), (6)
with C (t ) := Y 0 A (t ) − B (t )Y 0 . Using the dendriform products, ≺ and , this writes elegantly as fol-
lows:
Ẏ = C + Ẏ A − B ≺ Ẏ , Ẏ (0) = C (0) = Y 0 A (0) − B (0)Y 0 . (7)

As we will see in the next section, the surprising fact is, that the same dendriform rules also give
rise to an associative product, and both together, the associative and the pre-Lie product may be used
to find an exponential solution for Iserles’ problem (7), and hence also for the above recursions (4),
respectively (5), though in the more general context of dendriform algebras.
The link to Baxter’s work [Bax60] follows naturally from the fact that we solve the above recursions
using purely dendriform algebra, which allows us to replace the Riemann integral map I by more
general integral type operators, e.g. Rota–Baxter operators of arbitrary weight [Bax60,Rot69,Eb02],
including Jackson’s q-integral as well as Riemann sums.
Finally, we will present several applications at the end of the article, including the noncommutative
generalization of a Spitzer type identity first presented in a commutative context by Vogel in [V63],
coming from probability theory. A link to Riccati’s differential equation is outlined, which we hope to
deepen in the near future. We will use planar rooted trees to encode the pre-Lie Magnus expansion
and also comment on an observation made in an earlier publication [EM09], indicating a reduced
number of terms in the pre-Lie Magnus and Fer expansions due to the pre-Lie relation.
The paper is organized as follows. In Section 2 we recall the definition of a dendriform algebra and
introduce the associated left and right pre-Lie products. The augmentation D of a dendriform algebra
D by a unit 1 is also recalled [Ron00,Cha02,Ron02]. In Section 3 we introduce our most general
equation in D [[λ]]: it is said to be of degree (m, n) as it involves monomials of degree m + 1 (resp.
n + 1) with respect to the right (resp. left) dendriform product (resp. ≺). We prove in detail that (2)
gives a complete solution of the case (m, n) = (1, 1), i.e. Eq. (1) for a, b, c ∈ D, in terms of solutions
of the cases of degree (1, 0) and (0, 1). The particular (and simpler) case, where a = 1 and b, c ∈ D is
also considered. Section 4 is devoted to the pre-Lie Magnus expansion, recalling the results of [EM09].
In Section 5 we introduce matrix dendriform algebras, which allow us to solve the equations of degree
(2, 0) and (0, 2). The procedure is generalized to equations of degree (m, 0) and (0, n) in the following
section. Finally, Section 7 is devoted to applications in the context of Rota–Baxter algebras, which
furnish a large collection of dendriform algebras. We touch initial value problems in matrix algebras
[Ise01], an identity discovered by W. Vogel [V63], and show how the Riccati differential equation
is linked with the degree (2, 0) dendriform equation. Eventually, we introduce a noncommutative
Butcher series like formula for the pre-Lie Magnus expansion and indicate how the pre-Lie structure
may be used to reduce the number of terms in it.
2. Dendriform power sums expansions
Loday’s dendriform algebra may be seen at the same time as an associative as well as a pre-
Lie algebra. Indeed, it was found that in any dendriform algebra both products can be written as
particular linear combinations of the dendriform left and right binary compositions. Main examples
of dendriform algebras are provided by the shuffle and quasi-shuffle algebra as well as associative
Rota–Baxter algebras. The link between dendriform algebras and the latter is quite natural as shuffle
type products and their noncommutative generalizations appear in the construction of free associative
Rota–Baxter algebras [EG05].
We briefly introduce the setting of dendriform algebra. Let k be a field of characteristic zero. Recall
that a dendriform algebra [Lod01] over the field k is a k-vector space A endowed with two bilinear
operations, denoted ≺ and and called right and left products, respectively, subject to the three
axioms below:
(a ≺ b) ≺ c = a ≺ (b ≺ c + b c ), (8)
(a b) ≺ c = a (b ≺ c ), (9)
a (b c ) = (a ≺ b + a b) c . (10)
One readily verifies that these relations yield associativity for the product:
a ∗ b := a ≺ b + a b. (11)
However, at the same time the dendriform relations imply that the bilinear products and defined
by:
a b := a b − b ≺ a, a b := a ≺ b − b a, (12)
are left pre-Lie and right pre-Lie, respectively, which means that we have:
(a b) c − a (b c ) = (b a) c − b (a c ), (13)
(a b) c − a (b c ) = (a c ) b − a (c b). (14)
Observe that a b = −b a. The associative operation ∗ and the pre-Lie operations , all define
the same Lie bracket:
[[a, b]] := a ∗ b − b ∗ a = a b − b a = a b − b a.
Observe that:
a ∗ b + b a = a b + b a.
We stress here that in the commutative case (commutative dendriform algebras are also named Zinbiel
algebras [Lod95,Lod01]), the left and right operations are further required to identify, so that a b =
b ≺ a. In this case both pre-Lie products vanish. A natural example of a commutative dendriform
algebra is given by the shuffle algebra in terms of half-shuffles [Sch58]. Any associative algebra A
equipped with a linear integral-like map I : A → A satisfying the integration by parts rule also gives
a dendriform algebra, when a ≺ b := aI (b) and a b := I (a)b. Note that Loday and Ronco introduced
in [LR04] the notion of tridendriform algebra, that is, a k-vector space T equipped with three binary
operations, <, > and • satisfying seven relations. One can show that any tridendriform algebra gives
a dendriform algebra [Eb02].
Let A = A ⊕ k.1 be our dendriform algebra augmented by a unit 1, such that:
a ≺ 1 := a =: 1 a, 1 ≺ a := 0 =: a 1, (15)
implying a ∗ 1 = 1 ∗ a = a. Note that 1 ∗ 1 = 1, but that 1 ≺ 1 and 1 1 are not defined, see [Ron00,
Cha02] for more details.
We recursively define the following set of so-called left- respectively right-dendriform iterated
elements of degree n ∈ N in A [[λ]] for fixed x1 , . . . , xn ∈ A:
(n)

w (x1 , . . . , xn ) := · · · (x1 x2 ) x3 · · · xn ,
(n)

w≺ (x1 , . . . , xn ) := x1 ≺ x2 ≺ · · · (xn−1 ≺ xn ) · · · .
(0) (0)
Defining w ≺ (x1 , . . . , xn ) := 1 =: w (x1 , . . . , xn ), we may write these left- respectively right-
dendriform iterated elements more compactly:
(n)
(n−1)

w (x1 , . . . , xn ) := w (x1 , . . . , xn−1 ) xn ,
(n)
(n−1)

w≺ (x1 , . . . , xn ) := x1 ≺ w ≺ (x2 , . . . , xn ) .
(n) (n) (n) (n)

In case that x1 = · · · = xn = x we simply write w ≺ (x, . . . , x) = w ≺ ({x}n ) = w ≺ (x) and w (x, . . . ,
(n) (n)
x) = w ({x}n ) = w (x).
Let us recall from Chapoton and Ronco [Ron00], [Ron02, Proposition 2.6.3], [Cha02, Section 3.4]
that, on the free dendriform algebra on one generator a, augmented by a unit element, there is a
(n) (n)
Hopf algebra structure with respect to the associative product (11). The elements w := w (a)
generate a cocommutative graded connected Hopf subalgebra ( H , ∗) with coproduct:
(n)
(n) (n) (n−m)
w = w ⊗ 1 + 1 ⊗ w + w (m) ⊗ w , (16)
0<m<n
(n) (n) (n)

and antipode S ( w ) = (−1)n w ≺ . In fact the w generate a free associative subalgebra of the free
dendriform algebra on a for the associative product [LR98, Theorem 3.8], so that one can use the
(n)
previous formula for the coproduct action on w as a definition of the Hopf algebra structure on H .
As an important consequence, it follows that H is isomorphic, as a Hopf algebra, to the Hopf algebra
of noncommutative symmetric functions [GKLLRT95]. Let us finally introduce the exponential and
logarithm map in terms of the associative product (11). For any x ∈ λ A [[λ]]:

exp∗ (x) := x∗n /n! resp. log∗ (1 + x) := − (−1)n x∗n /n.
n0 n >0
3. Power sums expansions in unital dendriform algebras
Let ( A , ≺, ) be a dendriform algebra, A its augmentation by a unit 1. Let us introduce the fol-
lowing equation of degree (m, n) in A [[λ]]:

m
(q+1)

n
( p +1 )
X =a+ λq ω ( X , bq1 , . . . , bqq ) + λ p ω≺ (c p1 , . . . , c pp , X ), (17)
q =1 p =1
with m, n ∈ N+ and a, b i j , 1 j i m, c i j , 1 j i n in A. Here, the degree deg( X ) ∈ N2+ of X

is defined as the maximal degree of the left- respectively right-dendriform iterated elements in X
minus one. In this work we give a detailed account on the solutions of the cases (m, 0), (0, n) and
(1, 1), leaving the general case for further studies.
3.1. Equation of degree (1, 1)
We first study the case where deg( X ) = (1, 1) in (17). Its solution follows from the solutions of
the cases (1, 0) and (0, 1). Later, we will see how the more general cases (m, 0) and (0, n), m and n
bigger than one reduce to the cases (1, 0) and (0, 1) by embedding the dendriform algebra ( A , ≺, )
into the matrix dendriform algebra M N ( A ) of size N = N ≺ and N = N , respectively, with:
(m − 1)m (n − 1)n
N := 1 + and N ≺ := 1 + .
2 2
Relation (17) for m = n = 1 simplifies to the following equation:
X = a + λ X b + λc ≺ X , (18)
in A [[λ]] for fixed elements a, b, c ∈ A.

First, observe that for c = −b ∈ A, relation (18) reduces to a degree one recursion involving the
dendriform left pre-Lie product (12):
X = a + λ X b, (19)
in A [[λ]] for fixed elements a, b ∈ A. For this pre-Lie type recursion the special case a = 1 leads to the
simple solution X = 1, since 1 x = x 1 = 0.
Proposition 1. Let ( A , ≺, ) be a unital dendriform algebra. Let Y = Y (c ) and Z = Z (b), c , b ∈ A be solutions

of the recursions:
Y = 1 + λc ≺ Y resp. Z = 1 + λ Z b, (20)
of degree deg(Y ) = (0, 1) respectively deg( Z ) = (1, 0), in A [[λ]]. One verifies that the product Y ∗ Z gives a
solution to Eq. (18) for the particular case a = 1 ∈ A.
Proof. Indeed, a simple calculation shows that:
Y ∗ Z = (1 + λc ≺ Y ) ∗ (1 + Z λb)
= 1 + λc ≺ Y + Z λb + (λc ≺ Y ) ∗ ( Z λb)
= 1 + (λc ≺ Y ) ≺ (1 + Z λb) + (1 + λc ≺ Y ) ( Z λb)
= 1 + (λc ≺ Y ) ≺ Z + Y ( Z λb)
= 1 + λc ≺ (Y ∗ Z ) + (Y ∗ Z ) λb. 2
It is obvious that formal solutions to (20) are given by:

Y= λn w ≺
(n)
(c ) resp. Z = λn w
(n)
(b). (21)
n0 n0
Notice that, due to the definition of the Hopf algebra structure on H , these two series behave as
group-like elements with respect to the coproduct , see (16), (up to the extension
of the scalars from
k to k[λ] and the natural extension of the Hopf algebra structure on H = n0 H n to its completion

Ĥ = n0 H n with respect to the grading). Hence, we remark here that with respect to Eqs. (20) the
solutions Ỹ = Ỹ (c ) and Z̃ = Z̃ (b) of the equations:
Ỹ = 1 − λỸ c and Z̃ = 1 − λb ≺ Z̃ (22)
satisfy:
Ỹ ∗ Y = 1 = Y ∗ Ỹ resp. Z̃ ∗ Z = 1 = Z ∗ Z̃ .
Indeed we have, using (21) and the antipode:

Ỹ = (−1)n λn w
(n)
(c )
n0

=S λn w ≺
(n)
(c ) = S (Y ) = Y −1 ,
n0
and similarly for Z . Hence, we may write Ỹ = Y −1 , and Z̃ = Z −1 , respectively.

We now slightly generalize equations (20) by allowing them to start with a non-unital element:
E = a + λb ≺ E resp. F = c + λ F d, (23)
in A [[λ]] for fixed elements a, b, c , d ∈ A.

Recall the recursively defined set of left- respectively right-dendriform iterated elements of degree
n ∈ N in A [[λ]], n > 0:
(n+1)
(n)

w≺ { y }n , x := y ≺ w ≺ { y }n−1 , x ,
(n+1)
(n−1)

w x, { y }n := w x, { y }n−1 y,
in A and for fixed x, y ∈ A. In A [[λ]] we find immediately the formal solutions for Eqs. (23):

E =a+ λn−1 w ≺
(n)
{b}n−1 , a and F = c + λn−1 w
(n)
c , {d}n−1 .
n2 n2
Our first result is the following:
Lemma 2. Let A be a unital dendriform algebra. Let Y = Y (b) and Z = Z (d), b, d ∈ A, be solutions in A [[λ]]
to Eqs. (20), respectively. For a, c ∈ A Eqs. (23) are solved by:

E = Y ∗ Y −1 a and F = c ≺ Z −1 ∗ Z (24)
in A [[λ]], respectively.
Proof. Let us verify this for the first equation. The other case works analogously:

Y ∗ Y −1 a = (1 + λb ≺ Y ) ∗ Y −1 a

= Y −1 a + (λb ≺ Y ) ∗ Y −1 a

= Y −1 a + (λb ≺ Y ) ∗ Y −1 a + λb ≺ Y ∗ Y −1 a

= Y −1 a + (Y − 1) ∗ Y −1 a + λb ≺ Y ∗ Y −1 a

= a + λb ≺ Y ∗ Y −1 a . 2
An important remark concerning the restriction a, c ∈ A is in order. Eqs. (23) reduce to the ones
in (20) for a = c = 1. However, for the solutions in (24) we must be careful with respect to the
particular properties of the dendriform unit, i.e. we have to avoid 1 ≺ 1 respectively 1 1, see (15).
Indeed, for a ∈ A one may rewrite (24):

E = Y ∗ Y −1 a = Y ∗ 1 a − λ Y −1 b a = Y ∗ a − λ Y ∗ Y −1 b a .
Written in this form we may now put a = 1, keeping in mind that b = 1. Hence, we find E = Y ,
i.e. that E, with b instead of c, is a solution of the first relation in (20) as expected. From this we
conclude our next result, i.e. the solution to Eq. (18).
Theorem 3. Let A be a unital dendriform algebra. Let Y = Y (c ) and Z = Z (b), c , b ∈ A, be solutions in A [[λ]]
to the equations Y = 1 + λc ≺ Y , Z = 1 + λ Z b, of degrees (0, 1), (1, 0), respectively. Then the equation:
X = a + λ X b + λc ≺ X ,
of degree deg( X ) = (1, 1) and a ∈ A is solved by:

X = Y ∗ Y −1 a ≺ Z −1 ∗ Z . (25)
Proof. Let us verify this in detail. Recall Eqs. (22):

Y ∗ Y −1 a ≺ Z −1 ∗ Z

= (1 + λc ≺ Y ) ∗ Y −1 a ≺ Z −1 ∗ Z

= Y −1 a ≺ Z −1 ∗ Z + (λc ≺ Y ) ≺ Y −1 a ≺ Z −1 ∗ Z

+ (λc ≺ Y ) Y −1 a ≺ Z −1 ∗ Z

= Y −1 a ≺ Z −1 ∗ Z + λc ≺ Y ∗ Y −1 a ≺ Z −1 ∗ Z

+ (λc ≺ Y ) ∗ Y −1 a ≺ Z −1 ∗ Z

= Y −1 a ≺ Z −1 ∗ Z + λc ≺ Y ∗ Y −1 a ≺ Z −1 Z

+ λc ≺ Y ∗ Y −1 a ≺ Z −1 ≺ Z + (λc ≺ Y ) ∗ Y −1 a ≺ Z −1 ∗ Z

= Y −1 a ≺ Z −1 ∗ Z + λc ≺ Y ∗ Y −1 a ≺ Z −1 Z

+ λc ≺ Y ∗ Y −1 a ≺ Z −1 ∗ Z + (Y − 1) ∗ Y −1 a ≺ Z −1 ∗ Z

= λc ≺ Y ∗ Y −1 a ≺ Z −1 Z + λc ≺ Y ∗ Y −1 a ≺ Z −1 ∗ Z + a ≺ Z −1 ∗ Z

= a + λc ≺ Y ∗ Y −1 a ≺ Z −1 ∗ Z + a ≺ Z −1 + (λc ≺ Y ) ≺ Y −1 a ≺ Z −1 Z

= a + λc ≺ Y ∗ Y −1 a ≺ Z −1 ∗ Z + a ≺ Z −1 + (λc ≺ Y ) ≺ Y −1 a ≺ Z −1 (λ Z b)

= a + λc ≺ Y ∗ Y −1 a ≺ Z −1 ∗ Z

+ λ a ≺ Z −1 ∗ Z + (λc ≺ Y ) ≺ Y −1 a ≺ Z −1 ∗ Z b

= a + λc ≺ Y ∗ Y −1 a ≺ Z −1 ∗ Z + λ a ≺ Z −1 ∗ Z + Y ≺ Y −1 a ≺ Z −1 ∗ Z b

= a + λc ≺ Y ∗ Y −1 a ≺ Z −1 ∗ Z + λ a ≺ Z −1 ∗ Z + Y ∗ Y −1 a ≺ Z −1 ∗ Z b

− λ Y ∗ Y −1 a ≺ Z −1 ∗ Z b

= a + λc ≺ Y ∗ Y −1 a ≺ Z −1 ∗ Z + λ Y ∗ Y −1 a ≺ Z −1 ∗ Z b
= a + λc ≺ X + λ X b. 2
Corollary 4. Let A be a unital dendriform algebra and a, b ∈ A. Let Z = Z (b) be a solution to Z = 1 + λ Z b.

The solution to the “left” pre-Lie equation in A [[λ]]:
X = a + λX b (26)
is given by:

X = Z −1 ∗ Z a ≺ Z −1 ∗ Z . (27)
Recall that the solution Z (b) of the equation of degree (1, 0), Z = 1 + λ Z b is a group-like
element with respect to the coproduct , see (16). We may therefore write the above solution (27)
as:

X = Ad∗Z −1 Z a ≺ Z −1 ,
where as usual Ad∗F (G ) := F ∗ G ∗ F −1 . Due to the second dendriform axiom (9) we may interpret the
term Z a ≺ Z −1 as a second group action, which we denote by:
Θ Z (a) := Z a ≺ Z −1 . (28)
Lemma 5. Let A be a unital dendriform algebra and a, b ∈ A. Let A , B be two solutions of the equation of
degree (1, 0).
(1) Θ A ◦ Θ B (x) = Θ A ∗ B (x);

(2) Θ A ◦ Θ A −1 (x) = Θ A −1 ◦ Θ A (x) = x.
Proof. The second property follows from the first. Let us verify the first identity:

Θ A ◦ Θ B (x) = A B x ≺ B −1 ≺ A −1
= ( A ∗ B ) x ≺ ( A ∗ B )−1 = Θ A ∗ B (x).
This follows immediately from the dendriform axioms. 2
We remark that the two group actions in general do not commute, i.e. Ad∗A ◦ Θ B = Θ B ◦ Ad∗A .
The corresponding infinitesimal actions write as follows: let Y be the solution of the equation Y =
1 + λa ≺ Y of degree (0, 1) in A [[λ]].
(1) ΘY (x) = Y x ≺ Y −1 ∼ x + λa x + O (λ2 );

(2) Ad∗Y (x) = Y ∗ x ∗ Y −1 ∼ x + λ[[a, x]] + O (λ2 ).
Recall that the Dynkin operator is the linear endomorphism of the tensor algebra T ( X ) over an
alphabet X = {x1 , . . . , xn , . . .} into itself the action of which on words y 1 . . . yn , y i ∈ X is given by the
left-to-right iteration of the associated Lie bracket:

D ( y 1 , . . . , yn ) = · · · [ y 1 , y 2 ], y 3 · · · , yn ,
where [x, y ] := xy − yx. The Dynkin operator is a quasi-idempotent: its action on a homogeneous
element of degree n satisfies D 2 = nD. The associated projector D /n sends T n ( X ), the component
of degree n of the tensor algebra, to the component of degree n of the free Lie algebra over X . The
tensor algebra is a graded connected cocommutative Hopf algebra, and it is natural to extend the def-
inition of D to any such Hopf algebra as the convolution product of the antipode S with the grading
operator N: D := S N [PR02,EGP06,EGP07,EMP07]. This applies in particular in the dendriform con-
text to the Hopf algebra H introduced above. We will write D n for D ◦ pn , where pn is the canonical
projection from T ( X ) (resp. H ) to T n ( X ) (resp. H n ).
Lemma 6. (See [EMP07].) For any integer n 1 and for any x ∈ A we have:
(n)

D w (x) = · · · (x x) x · · · x. (29)
Let us come back to Eq. (18) respectively Corollary 4. Then, using D (1) = 0, we see immediately
that the solution Z = Z (a) to the equation Z = 1 + λ Z a is mapped to the solution:

X := λ−1 D ( Z ) = Ad∗Z −1 ◦ Θ Z (a) = Z −1 ∗ Z a ≺ Z −1 ∗ Z
of the corresponding pre-Lie equation (19) for a = b = −c ∈ A:
X = a + λ X a.
4. The pre-Lie Magnus expansion
In this section we recall results from [EM09] where we obtained a recursive formula for the loga-
rithm of the solutions of the two dendriform equations X = 1 + λa ≺ X and Y = 1 − λY a in A [[λ]],
a ∈ A, of degree (0, 1), (1, 0), respectively.
Let us introduce the following operators in ( A , ≺, ), where a is any element of A:
L ≺ [a](b) := a ≺ b, L [a](b) := a b, R ≺ [a](b) := b ≺ a, R [a](b) := b a,
L [a](b) := a b, L [a](b) := a b, R [a](b) := b a, R [a](b) := b a.
In the following theorem we find a recursive formula for the logarithm of the solutions of the
equations in (20). We will call it the pre-Lie Magnus expansion for reasons which will become clear in
Remark 8.
Theorem 7 (Pre-Lie Magnus expansion). (See [EM09].) Let Ω := Ω (λa), a ∈ A, be the element of λ A [[λ]]
such that Y = exp∗ (Ω ) and Y −1 = exp∗ (−Ω ), where Y and Y −1 are the solutions of the two equations
Y = 1 +λa ≺ Y and Y −1 = 1 −λY −1 a of degree (0, 1), (1, 0), respectively. This element obeys the following
recursive equation:
R [Ω ] B m (m)
Ω (λa) = (λa) = (−1)m R [Ω ](λa), (30)
1 − exp(− R [Ω ]) m!
m0
or alternatively:
L [Ω ] B m (m)
Ω (λa) = (λa) = L [Ω ](λa), (31)
exp( L [Ω ]) − 1 m!
m0
where the B l ’s are the Bernoulli numbers.
Returning to Theorem 3 we see that the solution of Eq. (18) writes:

X = exp∗ Ω (c ) ∗ exp∗ −Ω (c ) a ≺ exp∗ Ω (−b) ∗ exp∗ −Ω (−b)
and the solution to the pre-Lie recursion, i.e. Eq. (18) for c = −b:
X = Ad∗exp∗ (−Ω (−b)) ◦ Θexp∗ (Ω (−b)) (a).

Remark 8 (Classical Magnus expansion). Let F be any algebra of operator-valued functions

x on the real
line, e.g. smooth n × n matrix-valued functions, closed under integrals I (U )(x) := 0 U (s) ds. Then,
F is a dendriform algebra for the operations:
x x
A ≺ B (x) := A (x) · B (t ) dt , A B (x) := A (t ) dt · B (x)
0 0
with A , B ∈ F . This is a particular example of a dendriform structure arising from a Rota–Baxter

algebra structure—of weight zero [Ag00]. We refer to Section 7 for further details on Rota–Baxter
algebras and their connections to dendriform algebras. Here, let us simply mention that the Rota–
Baxter operator on F giving rise to the dendriform structure is the Riemann integral map: I ( A )(x) :=
x
0
A (t ) dt. In this particular example we see that the pre-Lie Magnus expansion (30) with the pre-Lie
product L [ A ]( B ) = ( A B )(t ) = [ I ( A ), B ](t ):
B m (m)
Ω ( A )(s) = Ω̇(s) = ad A (s) , (32)
m! Ω(s)
m0
coincides with Magnus’ logarithm [Mag54], Ω(t ) = log(Y (t )), of the solution of the initial value prob-
lem Ẏ (t ) = A (t )Y (t ), Y (0) = Y 0 = 1, such that:
t
Ω(t )n
Y (t ) = exp Ω̇(s) ds = , Ω(0) = 0.
n!
0 n0
For more details on Magnus’ work see also [Wil67,MP70,St87,KO89,GKLLRT95,Z96,IN99,CaIs06]. By

now we hope the above used prime notation, Ω , in Eqs. (30) and (31) of Theorem 7 has become
self-evident. Notice that if we suppose the algebra F to be unital, the unit (which we denote by 1)
has nothing to do with the added unit 1 of the underlying dendriform algebra. In fact, we extend the
Rota–Baxter algebra structure to F by setting:
R (1) := 1, R̃ (1) := −1 and 1.x = x.1 = 0 for any x ∈ F . (33)
This is consistent with the axioms (15) which in particular yield 1 x = R (1)x and x ≺ 1 = −x R̃ (1).
Later, in Section 7 we will provide more details in the context of applications.
Remark 9. We recently learned that the classical Magnus expansion (32) appeared in the general
pre-Lie context, corresponding to the Riemann integral, as early as 1980 in [AG80]. The dendriform
context is needed to interpret it as a logarithm with respect to the corresponding associative product.
A pre-Lie version of Fer’s infinite product expansion of this logarithm [F58] is also available, see
[EM09].
5. An equation of degree (2, 0)
Let D be a dendriform algebra and D its augmentation by a unit 1, see (15). We are interested in
solving the following equation of degree (2, 0) in D [[λ]]:
X = d + λ X c + λ2 ( X b ) a , (34)
for a, b, c ∈ D and d ∈ D. We introduce matrix dendriform algebras, in which Eq. (34) will be trans-
formed into a system of order (1, 0) equations.
5.1. Matrix dendriform algebras
Let ( D , ≺, ) be a dendriform algebra. Recall from [Lod01] that the space Mn ( D ) of square n × n-
matrices with entries in D is a dendriform algebra, with operations defined by:

n
n
(a ≺ b)i j := aik ≺ bkj , (a b)i j := aik bkj .
i =1 i =1
We can indeed verify the first dendriform axiom:

n
n
(a ≺ b) ≺ c ij
= (a ≺ b)ik ≺ ckj = (ail ≺ blk ) ≺ ckj
k =1 k,l=1

n
n
= ail ≺ (blk ∗ ckj ) = ail ≺ (b ∗ c )lj
k,l=1 l =1

= a ≺ (b ∗ c ) i j ,
hence (a ≺ b) ≺ c = a ≺ (b ∗ c ) in Mn ( D ), and the two other axioms can be checked similarly. If

D := D ⊕ k.1 (resp. Mn ( D ) := Mn ( D ) ⊕ k.1n ) is the dendriform algebra D (resp. Mn ( D )) augmented
with a unit, then we can identify the unit 1n with the diagonal matrix with dendriform units 1’s on
the diagonal and 0 elsewhere.
5.2. Transformation into a dendriform system of degree (1, 0)
First, we will show how to transform equation (34) for d = 1 into a “first-order” equation in
M2 ( D )[[λ]]. Observe the size of the matrices, i.e. 2 = 1 + (2 − 1)2/2.
Lemma 10. Let Z ∈ M2 ( D )[[λ]] be the solution of the “matrix dendriform” equation of degree (1, 0):
Z = 12 + λ Z M . (35)
Then for any v = ( v 1 , v 2 ) ∈ k2 the solution in ( D [[λ]])2 of the equation:
Y = v12 + λY M , (36)
with Y M := (Y 1 M 11 + Y 2 M 21 , Y 1 M 12 + Y 2 M 22 ) is given by:
v Z := ( v 1 Z 11 + v 2 Z 21 , v 1 Z 12 + v 2 Z 22 ).
Proof. We immediately check:
v Z = v12 + λ v ( Z M )
= v12 + λ( v Z ) M . 2
Proposition 11. Let X be the solution of Eq. (34) in D [[λ]] with d = 1. Then the vector Y := ( X , λ X b) in
( D [[λ]])2 is the solution of Eq. (36) with v = (1, 0) ∈ k2 and:

c b
M= .
a 0
Proof. One easily checks that:

c b
v12 + λY M = (1, 0) + λ( X , λ X b)
a 0

= (1, 0) + λ X c + λ( X b) a, X b
= (1, 0) + ( X − 1, λ X b)
= Y. 2
Using Theorem 7, i.e. the pre-Lie Magnus expansion, we find immediately the solution to (35):
c b
Corollary 12. Let Ω = log∗ Z ∈ M2 ( D [[λ]]), where Z is the solution of Eq. (35) with M = , given by the
a0
pre-Lie Magnus expansion:
1 1 1
Ω (λ M ) = λ M − λ2 M M + λ3 ( M M ) M + λ3 M (M M ) + · · · .
2 4 12
Then the solution X of Eq. (34), d = 1, is such that the line vector ( X , λ X b) is the first line of the matrix
Z = exp∗ (Ω ), i.e.:
X = (1, 0) Z (1, 0)t .
The following corollaries are readily verified.
Corollary 13. Let X be the solution of Eq. (34) in D [[λ]]. Then the vector Y := ( X , λ X b) in ( D [[λ]])2 is the
solution of equation:
Y = v (d ∗ 12 ) + λY M , (37)
with v = (1, 0) ∈ k2 and M as in Proposition 11.
Here d ∗ 12 is the diagonal matrix in M2 ( D ) with d on its diagonal and 0 else. Using Lemma 2 we
immediately get:
Corollary 14. Let Ω = log∗ Z ∈ M2 ( D [[λ]]), where Z is the solution of Eq. (35) with M as in Proposition 11,
given by the pre-Lie Magnus expansion:
1 1 1
Ω (λ M ) = λ M − λ2 M M + λ3 ( M M ) M + λ3 M (M M ) · · · .
2 4 12
The line vector ( X , λ X b) is the first line of the matrix U := (d12 ≺ exp∗ (−Ω )) ∗ exp∗ (Ω ).
The equation of degree (0, 2) in D [[λ]]:
X̂ = 1 + λ c ≺ X + λ2 a ≺ (b ≺ X̂ ) (38)
can be treated similarly:
Proposition 15. Let X̂ be the solution of Eq. (38) in D [[λ]]. Then the column vector Ŷ := ( X̂ , λb ≺ X̂ )t in
( D [[λ]])2 is a solution of equation:
Ŷ = 12 v + λ M t ≺ Ŷ , (39)
with v = (1, 0)t ∈ k2 and:

c a
Mt = .
b 0
Proof. The solution of Eq. (38) in D [[λ]] is exactly the solution of Eq. (34) in D [[λ]] where D is the
dendriform algebra ( D , , ), with x y := y x and x y := y ≺ x. 2
6. Higher-order equations
The general degree (m, 0) equation writes:

m
(q+1)
X = a00 + λq ω ( X , aq1 , . . . , aqq ). (40)
q =1
The cases m = 3 and m = 4 can be worked out by hand and will give a hint for the general case.
Indeed, the length 4 line vector:

Y 3 := X , λ X a21 , λ X a31 , λ2 ( X a31 ) a32
is the solution of the equation:
⎛ ⎞
a11 a21 a31 0
⎜ a22 0 0 0 ⎟
Y 3 = (a00 , 0, 0, 0) + λY 3 M 3 , with M3 = ⎝ ⎠.
0 0 0 a32
a33 0 0 0
The length 7 line vector:

Y 4 := X , λ X a21 , λ X a31 , λ2 ( X a31 ) a32 , λ X a41 ,

λ2 ( X a41 ) a42 , λ3 ( X a41 ) a42 a43
is the solution of the equation:

⎛ ⎞
a11 a21 a31 0 a41 0 0
⎜ a22 0 0 0 0 0 0 ⎟
⎜ ⎟
⎜ 0 0 0 a32 0 0 0 ⎟
⎜ ⎟
Y 4 = (a00 , 0, 0, 0, 0, 0, 0) + λY 4 M 4 , with M 4 = ⎜ a33 0 0 0 0 0 0 ⎟.
⎜ ⎟
⎜ 0 0 0 0 0 a42 0 ⎟
⎝ 0 0 0 0 0 0 a43
⎠
a44 0 0 0 0 0 0
We recognize the matrix M 3 as the upper left 4 × 4-submatrix of M 4 . The solution of the general
equation (40) will be the first coefficient of a vector Y m of length 1 + m(m − 1)/2, whose coefficients
j +1
(discarding the first one) are given by λ j ω ( X , aq1 , . . . , aqj ), 1 j < q m. They are displayed ac-
cording to the lexicographic order of the pairs (q, j ) indexing them. The vector Y m is the solution of
the equation
Y m = (a00 , 0, . . . , 0) + λY m M m ,

m(m−1)
2
times
where the square matrix M m of size 1 + m(m − 1)/2 is recursively defined by:
⎛ ⎛ ⎞ ⎞
am1 0 0 ··· 0
⎜ ⎜ 0 0 0 ··· 0⎟ ⎟
⎜ M m −1 ⎜ . .. .. .. ⎟ ⎟
⎜ ⎝ .. . . ··· .⎠ ⎟
⎜ ⎟

⎜⎜ 0 0 0 ··· 0 ⎟
⎟
M m −1 A1 ⎜ ⎞⎟
Mm = = ⎜⎛ 0 0 ··· 0
⎞ ⎛
0 am2 0 ··· 0 ⎟. (41)
A2 A3 ⎜ ⎟
⎜⎜ 0 0 ··· 0⎟ ⎜0 0 am3 ··· 0 ⎟⎟
⎜⎜ .⎟ ⎜. ⎟⎟
⎜⎜ .. .. ⎟ ⎜. .. .. .. .. ⎟⎟
⎜⎜ . . · · · .. ⎟ ⎜. . . . . ⎟⎟
⎝⎝ ⎠ ⎝ ⎠⎠
0 0 ··· 0 0 0 0 · · · am(m−1)
amm 0 ··· 0 0 0 0 ··· 0
We proceed then as in Section 5.2: according to Lemma 2. The line vector Y m is the first line of the
matrix of size N := 1 + m(m − 1)/2 defined by ((a00 ∗ 1 N ) ≺ Z ∗−1 ) ∗ Z , where Z is the solution of the
order (1, 0) matrix dendriform equation:
Z = 1N + Z Mm . (42)
In the simpler case when a00 = 1 instead of being an element of D, the line vector Y m is just the first
line of the solution Z of Eq. (42).
The solution of general equation of order (0, n) can then be derived immediately, by the same
trick as in the proof of Proposition 15. Higher-order equations of type (17) of order (m, n) remain to
be solved.
7. Applications
Recall from the introduction the link between recursions in dendriform algebras and simple initial
value problems, say, for instance in algebras of matrix valued functions closed under the Riemann
integral map. In fact, the link here is provided by the relation between associative Rota–Baxter and
dendriform algebras.
7.1. Rota–Baxter algebras
Recall [Bax60,Rot69,Eb02] that an associative Rota–Baxter algebra (over a field k) is an associative

k-algebra A endowed with a k-linear map R : A → A subject to the following relation:

R (a) R (b) = R R (a)b + aR (b) + θ ab , (43)
where θ ∈ k. The map R is called a Rota–Baxter operator of weight θ . The map R̃ := −θ id − R also is a
weight θ Rota–Baxter map. Both the image of R and R̃ form subalgebras in A. Associative Rota–Baxter
algebras arise in many mathematical contexts, e.g. in integral and finite differences calculus, but also
in perturbative renormalization in quantum field theory [EGM06].
A few examples are in order. First, recall the classical integration by parts rule showing that the or-
dinary Riemann integral is a weight zero Rota–Baxter map. Moreover, on a suitable class of functions,
we define the following Riemann summation operators:
[
x/θ ] [x
/θ ]−1
R θ ( f )(x) := θ f (nθ) and R θ ( f )(x) := θ f (nθ), (44)
n =1 n =1
which satisfy the weight −θ and the weight θ Rota–Baxter relation, respectively. The summation
operator:

S ( f )(x) := θ f (x + θ n) (45)
n1
also satisfies relation (43) on a suitable class of functions. Another useful example is given in term of
the q-dilatation operator, σq f (x) := f (qx). In fact, one shows that the q-summation operator:

S q ( f )(x) := σqn f (x) = f qn x (46)
n0 n0
satisfies relation (43) on a suitable class of functions, with θ = −1. In fact, one may think of (43) as a
generalized integration by parts identity. Indeed, the reader will have no difficulty in checking duality
of (43) with the ‘skewderivation’ rule [CEFG06]:
∂( f g ) = ∂( f ) g + f ∂( g ) + θ∂( f )∂( g ),
i.e. R ∂ x = x + c and ∂ Rx = x, x ∈ A and c ∈ ker ∂ . For instance, in the case of the q-summation opera-
tor (46), we find ∂q f (x) := (id − σq ) f (x) = f (x) − f (qx), satisfying:
∂q ( f g ) = ∂q ( f ) g + σq ( f )∂q g
= ∂q ( f ) g + f ∂q ( g ) − ∂q ( f )∂q ( g ).
From this one can prove the Rota–Baxter property for the q-summation operator S q (weight θ = −1),
which is equivalent to the so called q-integration by parts rule:

Sq( f )Sq(g) = Sq Sq( f )g + Sq σq ( f ) S q ( g ) .
We call a k-algebra A with both, a Rota–Baxter map R and its corresponding skewderivation ∂ a Rota–
Baxter pair ( A , R , ∂). Although, we should underline that, in general, Rota–Baxter algebras do not
come with such a skewderivation, e.g. let A be a K-algebra which decomposes directly into subalge-
bras A 1 and A 2 , then the projection to A 1 , R : A → A, R (a1 , a2 ) = a1 , is an idempotent Rota–Baxter
operator, i.e. of weight θ = −1, see [CEFG06,EM09] for more details.
Proposition 16. (See [Eb02].) Any associative Rota–Baxter algebra gives a tridendriform algebra, ( T R , <, >,
•θ ), in the sense that the Rota–Baxter structure yields three binary operations:
a < b := aR (b), a > b := R (a)b, a •θ b := θ ab,
satisfying the tridendriform algebra axioms [LR04].
The associated associative product ∗θ is given by:
a ∗θ b := aR (b) + R (a)b + θ ab. (47)
It is sometimes called the “double Rota–Baxter product”, and verifies:
R (a ∗θ b) = R (a) R (b), R̃ (a ∗θ b) = − R̃ (a) R̃ (b) (48)
which is just a reformulation of the Rota–Baxter relation (43). In fact, one easily shows that for any
Rota–Baxter algebra A, the vector space A equipped with the new product (47) is again a Rota–Baxter
algebra, denoted A R , though in general non-unital.
From the general link between dendriform and tridendriform algebras it simply follows:
Corollary 17. Any associative Rota–Baxter algebra gives rise to a dendriform algebra structure, ( D R , ≺, ),
given by:
a ≺ b := aR (b) + θ ab = −a R̃ (b), a b := R (a)b. (49)
The dendriform pre-Lie products, (13) and (14), can be written explicitly:

a b = R (a), b − θ ba, a b = a, R (b) + θ ab.
Recall [EM09] that we extend the dendriform structure ( D R , ≺, ) on the Rota–Baxter algebra
( A , R ) to D R by setting R (1) := 1, R̃ (1) := −1, and 1.x = x.1 = 0 for any x ∈ D R . This is consistent
with the axioms (15) which in particular yield 1 x = R (1)x and x ≺ 1 = −x R̃ (1).
7.2. A link with generalized initial value problems
Let us go back to Remark 8. Assume for the moment that we work in a noncommutative unital
function algebra F , say, suitable n × n-matrix valued functions, with a Rota–Baxter pair ( R , ∂). Let us
consider the initial value problems (IVPs)
Ẏ := ∂ Y = Y B − C Y , Y (0) = Y 0 (50)
with B , C ∈ F . Iserles in [Ise01] looked at this equation in the classical setting of Remark 8, where its
straightforward solution is given in terms of the classical Magnus expansion (32):

Y = − exp Ω(C ) Y 0 exp −Ω( B ) . (51)
Here U := exp(Ω( B )) and V := exp(−Ω(C )) solve the IVPs, U̇ = −U B, U (0) = 1 and V̇ = C V ,

V (0) = 1, respectively.
Coming back to Eq. (50) in the context of a Rota–Baxter pair ( R , ∂) of weight zero, we may trans-
form it into the recursion for Ẏ :
Ẏ = A + R (Ẏ ) B − C R (Ẏ ), A = Y 0 B − C Y 0,
which now becomes a dendriform equation of degree (1, 1) for the element X := Ẏ :
X = A + X B + C ≺ X.
One verifies by using the relation between ∂ and R as well as (48), that R maps the solution X ,
see (25), of this equation to (51), i.e. R ( X ) = Y .
However, in the light of the fact that Rota–Baxter algebras in general may not come with a corre-
sponding skewderivation, it is crucial to realize that X is simply an element in A R , the Rota–Baxter
algebra with product (47). Indeed, we would like to emphasize that our findings in the context of
dendriform algebras hold through for any associative Rota–Baxter algebra ( A , R ).
Hence, starting with a Rota–Baxter algebra ( A , R ), Theorem 7 gives us the solution Ŷ =
exp( R (Ω (a))) to the equation Ŷ = 1 + λ R (a Ŷ ). We may interpret this as the “integral-like” equa-
tion corresponding to the IVP, Ẏ = aY , Y 0 = 1. Though, we only used (48) as well as (15), and
interpret Ẏ purely as an element in A R . Theorem 3 then implies for b = 0 the solution to recursion
Ê = a + λ R (c Ê ), given by Ê = exp( R (Ω (λc ))) R (exp( R (Ω (λc )))a), analogously for c = 0. From a den-
driform algebraic point of view the classical IVP (50) should be replaced by an equation in terms of
the pre-Lie product:
X = a + X b,
a, b ∈ A, where its solution (27) again is an element in A R :
X = Ad∗exp∗ (−Ω (−b)) ◦ Θexp∗ (Ω (−b)) (a).
The formal limit of θ → 0 reduces to the Lie bracket flow like equation. It is important to remark
that, as there are two natural Rota–Baxter maps on A, i.e. R and R̃, we can map X ∈ A R either via R
to Y := R ( X ) ∈ A or via R̃ to Ỹ := R̃ ( X ) ∈ A.
7.3. A link with the Riccati differential equation
The Riccati differential equation writes:
ẏ = p + qy + r y 2 , (52)
where y , p , q and r are differentiable functions on the real line. We suppose for convenience that r
never vanishes. Looking at the transformation:
ẇ
y=− (53)
rw
one easily checks that the solutions of (52) are also solutions of the following linear homogeneous
second-order differential equation:
r ẅ − (ṙ + qr ) ẇ + pr 2 w = 0. (54)
Now consider Eq. (34). Recall Remark 8, in this particular situation, we find a dendriform algebra
structure on the set F of twice differentiable functions on R, which we denote by D, with dendriform
left and right products:
f ≺ g := f · R ( g ), f g := R ( f ) · g .
Here R = I is the Riemann integral operator on F . Recall from [EMP07b,EM09] that the operator R
(which is a Rota–Baxter map of weight zero) extends to D by putting R (1) = 1, where 1 is the usual
unit of the algebra of functions, i.e. the function equal to 1 everywhere. The operator R extends
naturally to D [[λ]] by λ-linearity. Starting from a solution X of (34) and putting y := R ( X ) we have:

y = 1 + λ R R ( X )c + λ2 R R R ( X )a b ,
and hence:
ẏ = λ yc + λ2 R ( ya)b.
We suppose at this point that b is a nowhere vanishing function on R. Dividing out by b and differ-
entiating once again we obtain:
1 ḃ d c c
ÿ − ẏ = λ y + λ ẏ + λ2 ya,
b b2 dt b b
hence:

d c
b ÿ − (ḃ + λbc ) ẏ − λ + λ2 a b 2 y = 0.
dt b
d q
We recognize a formal version of Eq. (54), with p = −( dt b
+ λ2 a), q = λc and r = b.
7.4. Vogel’s identity
We present here a generalization of an operator identity and its solution, found by Vogel in [V63]
in the context of probability theory, to noncommutative associative Rota–Baxter algebras.
Let ( A , R ) be an associative Rota–Baxter algebra of weight θ ∈ K, a, b, c ∈ A. The following equation
defined in A [[λ]] appeared in [V63]:
X = a + λ R ( X )b + λc R̃ ( X ). (55)
For A being commutative Vogel found the solution, X = B − C , where C = C (b, c ) and B = B (b, c ) are
defined in A [[λ]] as follows:

C := exp R α (λ) R a exp R θ −1 log(1 + λθ c ) + R̃ θ −1 log(1 + λθ b)
, (56)
−1 −1
B := exp − R̃ α (λ) R̃ a exp R θ log(1 + λθ c ) + R̃ θ log(1 + λθ b) , (57)
and θ α (λ) := log(1 + λθ b) − log(1 + λθ c ).1 It is obvious that Vogel’s identity simply translates into the
dendriform identity (18):
1
We should remark that, following our convention, Vogel uses a commutative Rota–Baxter algebra of weight θ = −1.
X = a + λ X b − λc ≺ X ,
with its general solution X = Y ∗ (Y −1 a ≺ Z −1 ) ∗ Z , Y = Y (−c ) and Z = Z (b) = Y (−b)−1 , i.e.:

X = exp∗ Ω (−λc ) ∗ exp∗ −Ω (−λc ) a ≺ exp∗ Ω (−λb) ∗ exp∗ −Ω (−λb) .
Rewriting it into the Rota–Baxter representation gives the solution to Vogel’s identity in a noncommu-
tative Rota–Baxter algebra. It is an interesting exercise to retrieve Vogel’s solution from the dendriform
one in the case of a commutative Rota–Baxter algebra. Hence, first observe that in this case:

a b = a b − b ≺ a = R (a)b − −b R̃ (a) = R (a)b − θ ba − bR (a) = R (a), b − θ ba = −θ ba.
Such that:
B n (n) Bn n log(1 − λθ a)
Ω (λa) = L [Ω ](λa) = (−θ)n λ Ω (λa) a = − .
n! n! θ
n0 n0
By recalling our convention (33) and using commutativity as well as the well-known identity for the
product (47), θ a ∗ b = R̃ (a) R̃ (b) − R (a) R (b), we find:

log(1 + λθ c ) log(1 + λθ b) log(1 + λθ c )
X = exp R̃ exp − R̃ R̃ a exp R
θ θ θ

log(1 + λθ b) log(1 + λθ c ) log(1 + λθ b)
+ R̃ − exp − R exp R
θ θ θ

log(1 + λθ c ) log(1 + λθ b)
× R a exp R + R̃ .
θ θ
7.5. Rooted trees and the coefficients in the pre-Lie Magnus expansion
The final section is devoted to a closer analysis of the pre-Lie Magnus expansion. Iserles and
Nørsett [IN99] used planar rooted binary trees to disentangle the combinatorial structure underlying
the classical Magnus expansion, see also [IMNZ00,Ise02]. Here we propose the use of planar rooted
trees to represent the pre-Lie Magnus expansion. This, of course, incorporates the classical case but,
is more in the line of—noncommutative—Butcher series.
However, let us first use planar rooted binary trees. Recall Ω (λa) = n>0 λn Ωn (a):
B n (m)
Ω (λa) = L [Ω ](λa).
n!
n0
This writes out as:
1 1 1 1
Ω (λa) = λa − λ2 a a + λ3 (a a) a + λ3 a (a a) − λ4 (a a) a a
2 4 12 8
1
− λ4 a (a a) a + a (a a) a + (a a) (a a) + · · · .
24
At fifth order we have ten terms. Using planar binary rooted trees to encode the pre-Lie product:
∼ a a,
we find at fourth order:

1 1
− − + + .
8 24
In [EM09] we observed that from order four upwards the left pre-Lie relation (13) implies a reduction
in the number of terms. Indeed, at order four we observe a reduction from the four terms above to
the following two terms:
1 1
− −
6 12
thanks to the left pre-Lie relation:
− = − .
To go beyond this order we propose another approach in the spirit of noncommutative Butcher se-
ries. In fact, one can write the pre-Lie Magnus expansion like a classical Butcher series using ordinary
planar rooted trees. By definition a planar rooted tree t is made out of vertices and nonintersect-
ing oriented edges such that all but one vertex have exactly one incoming line. For a tree t we
denote by V (t ) and E (t ) the sets of vertices and edges, respectively. By f ( v ), v ∈ V (t ) we denote
the number of outgoing edges, i.e. the fertility of the vertex v of the rooted tree t. The degree of
a tree deg(t ) is given by its number of vertices. We denote the graded vector space of planar rooted
(m)
trees by T pl := m>0 T pl . Let t = B + (t 1 , . . . , tn ), where B + denotes the grafting operation, join-
ing the nroots via n new edges to a new vertex, which becomes the root of the new tree t and
n
deg(t ) = i =1 deg(t i ) + 1. Now, we define the linear map α : T pl → k:
Bn B f (v )
n
α (t ) := α (t i ) = .
n! f ( v )!
i =1 v ∈ V (t )
We see immediately that this function maps a rooted tree t containing vertices of fertility f ( v ) =
2n + 1, n > 0, v ∈ V (t ) to zero.
Lemma 18. ker(α ) = {t ∈ T pl | ∃ v ∈ V (t ): f ( v ) = 2n + 1, n > 0}.
In the following we denote by T ple1 the graded vector subspace of planar rooted trees exclud-
ing trees t with vertices of fertility f ( v ) = 2n + 1, n > 0, v ∈ V (t ). We introduce the notation
(n)
r (a1 , . . . , an ) := a1 (a2 (a3 · · · (an−1 an )) · · ·). Next, we introduce the tree functional F . Defined
by F [ ](a) = a and for t = B + (t 1 · · · tn ), n 1:
(n+1)
F [t ](a) := r F [t 1 ](a), . . . , F [tn ](a), a .
Theorem 19. The pre-Lie Magnus expansion can be written:

Ω (a) = α (t ) F [t ](a). (58)
t ∈T ple1
Proof. The fact that the sum only includes planar rooted trees with even vertices and vertices of
fertility one follows from Lemma 18. We just have to prove that the RHS of (58) verifies Eq. (31)
(with λ set to 1). Indeed, if we denote by Ω this right-hand side we have:

Ω = α (t ) F [t ](a)
t ∈T ple1
Bm (m+1)
= α (t 1 ) · · · α (tm )r F [t 1 ](a), . . . , F [tn ](a), a
m!
m0 t 1 ,...,tm ∈T e1
pl
B m (m+1)
= r (Ω , . . . , Ω , a),
m!
m0 m times
hence Ω = Ω . 2

Remark 20. Let T ( z) = k0 T k zk be the Poincaré series of T ple1 . Using the methods of [De04] we can
prove that T is the solution of the following functional equation:
1
T ( z) − z2 T ( z)3 = ,
1−z
which in turn yields the recursive definition of the coefficients T n :

Tn = 1 + T p Tq Tr .
p ,q,r 0
p +q+r =n−2
This sequence matches with sequence [A049130] in Sloane’s encyclopedia [S]. On can see this by
noticing that the series W = zT satisfies the functional equation2 :
z
W ( z) − W ( z)3 = ,
1−z
z( z−1)( z+1)
hence W is the reversion of the series z → z3 − z+1
. The first terms of the sequence are:
1, 1, 2, 4, 10, 26, 73, 211, 630, 1918, . . . .
This formula provides a simple and efficient way to expand Magnus’—pre-Lie—recursion. Let us
check a few examples:
Ω (a) = α ( ) F [ ](a) + α ( ) F [ ](a) + α ( ) F [ ](a) + α ( ) F [ ](a) + · · ·

(2 ) (2 ) B 2 (2 )
= a + B 1 r F [ ](a), a + B 21 r F [ ](a), a + r F [ ](a), F [ ](a), a + · · · .
2!
2
We greatly thank Loïc Foissy for having indicated this point to us.
Recall that T pl denotes the vector space of planar rooted trees. Hence, at order four we find:
Ω (a) = · · · + α ( ) F [ ](a) + α ( )F [ ](a) + α ( )F [ ](a) + α ( )F [ ](a) + · · ·

(2 ) B2 (2 ) B2 (2 )
= · · · + B 31 r F [ ](a), a + B 1 r F [ ](a), F [ ](a), a + B 1 r F [ ](a), F [ ](a), a
2! 2!
B 2 (2 )
+ B 1 r F [ ](a), a + · · · .
2!
Recall that B 3 = 0. Due to the recursive nature of the pre-Lie Magnus expansion at order five the
following set of 8 trees appear:

B + ( ), B + ( ), B + ( , , , ), B + ( , ), B + ( , ), B + ( , ), B + ( , ), B + ( , ) .
Of course, the first two terms have different coefficients. Other trees do not appear due to the fact
that odd Bernoulli numbers are zero, i.e. B 3 = 0. Hence, for the order five contribution we find:
1 1
Ω5 (a) = − B 1 (a a) a a a − B 1 a (a a) a a
6 12
B4 B2

+ a a a (a a) + 2 a (a a) (a a) +
4! 2!2!
B 22 B2
+ a a (a a) a + B 21 (a a) (a a) a
2!2! 2!
B2 2 B2 2
+ B (a a) a (a a) + B a (a a) a a .
2! 1 2! 1
Using the left pre-Lie relation we may write:
B2 B2 B2 2
B 21 (a a) a (a a) = B 21 a
(a a) a a − B a (a a) a a
2! 2! 2! 1
B 2 2
+ B 1 (a a) a a a
2!
leading to seven terms:
B 2 5 B2 1
Ω5 (a) = − B 1 (a a) a a a − B 1 a (a a) a a
2! 2 2! 2
B4 B 22
+ a a a (a a) + a (a a) (a a) +
4! 2!2!
B 22 B2
+ a a (a a) a + B 21 (a a) (a a) a
2!2! 2!

+ B 2 B 21 a (a a) a a .
It is an open question whether one can further reduce the order five term using the left pre-Lie
identity. Moreover, it would be interesting to find a recursion for Ω (λa) which already incorporates
the pre-Lie identity.
8. A concluding remark
The notion of mould has been introduced by J. Ecalle [Ec81,Ec92] as a powerful tool in the analysis
of singularities (theory of resurgence). A mould is a function with a variable number of variables:
more precisely a rational mould will be a collection ( f n )n1 of rational functions with rational coeffi-
cients in n variables un . The graded vector space thus obtained:

Mould = Mouldn
n1
is an anticyclic non-symmetric operad [Cha07], with partial compositions ◦i : Mould p × Mouldq →

Mould p +q−1 given by:
( f ◦i g )(u 1 , . . . , u p +q−1 )
:= (u i + · · · + u i +q−1 ) f (u 1 , . . . , u i −1 , u i + · · · + u i +q−1 , u i +q , . . . , u p +q−1 ) g (u i , . . . , u i +q−1 ).
The rational function 1 = 1/u 1 ∈ Mould1 is the unit, i.e. 1 ◦1 f = f and f ◦i 1 = f for any f ∈ Mouldn
and for any i ∈ {1, . . . , n}. The anticyclic structure is given by the push τ : Mouldn → Mouldn defined
by:

τ ( f )(u 1 , . . . , un ) = f −(u 1 + · · · + un ), u 1 , . . . , un−1 .
One can easily check that τ (1) = −1 and that τ is of order n + 1 on Mouldn . Now consider the
space H of “continuous formal power series” with coefficients in some associative complete topolog-
ical algebra A, i.e. formal integrals:
+∞
h(t ) := h u t u −1 du ,
1
where h u ∈ A for any u ∈ R+ . A rational mould f = ( f n )n1 uniquely gives rise to a nonlinear opera-
tor F f on H given by:

F f [h](t ) := 1 + ··· f n (u 1 , . . . , un )h u 1 · · · h un t u 1 +···+un du 1 · · · dun .
n1
We will denote by I the (linear) operator corresponding to the unit mould 1 (which has only one
component in one variable which is 11 = 1 = 1/u 1 ). It is nothing but the Riemann integral operator
given by:
+∞ t
1 u1
I [h](t ) = F 1 [h](t ) = hu1 t du 1 = h(s) ds.
u1
1 0
We identify Mould with its image by f → F f . It is shown in [Cha07, Sections 3, 4] and in [CHNT08,
Example 3.15] that Mould is a dendriform algebra with:

( F G )[h] = I F [h]Ġ [h] , ( F ≺ G )[h] = I Ḟ [h]G [h] , (59)
and that the free dendriform algebra with one generator D embeds naturally into Mould. We have:
˙ G )[h] = ( Ḟ Ġ )[h] and ( F ≺

(F ˙ G )[h] = ( Ḟ Ġ )[h] (60)
where we have set:

( F G )[h] = I F [h] G [h], ( F G )[h] = F [h] I G [h] . (61)
As a consequence, if we consider a solution of some dendriform equation in D [[λ]], the derivative of its
image in Mould (more precisely in the nonlinear operators on H) satisfies the corresponding differ-
ential equation along the lines of Section 7.3. In other words, if the solution of a classical differential
equation can be expressed in a formula in terms of dendriform operations, then the derivative of the
solution of the same differential equation in (the image of) D [[λ]] is given by the same formula.
Acknowledgments
We would like to thank A. Iserles, A. Lundervold, H. Munthe-Kaas and A. Zanna for helpful dis-
cussions and useful remarks. We also thank the referee for the concluding remark in relation with
references [Cha07] and [CHNT08].
References
[AG80] A. Agrachev, R. Gamkrelidze, Chronological algebras and nonstationary vector fields, J. Sov. Math. 17 (1) (1981)
1650–1675.
[Ag00] M. Aguiar, Prepoisson algebras, Lett. Math. Phys. 54 (4) (2000) 263–277.
[Ag04] M. Aguiar, Infinitesimal bialgebras, pre-Lie and dendriform algebras, in: Hopf Algebras, in: Lect. Notes Pure Appl.
Math., vol. 237, Dekker, New York, 2004, pp. 1–33.
[At63] F.V. Atkinson, Some aspects of Baxter’s functional equation, J. Math. Anal. Appl. 7 (1963) 1–30.
[Bax60] G. Baxter, An analytic problem whose solution follows from a simple algebraic identity, Pacific J. Math. 10 (1960)
731–742.
[BCOR09] S. Blanes, F. Casas, J.A. Oteo, J. Ros, Magnus expansion: Mathematical study and physical applications, Phys. Rep. 470
(2009) 151–238.
[BCOR98] S. Blanes, F. Casas, J.A. Oteo, J. Ros, Magnus and Fer expansions for matrix differential equations: The convergence
problem, J. Phys. A 31 (1998) 259–268.
[CEFG06] J. Cariñena, K. Ebrahimi-Fard, H. Figueroa, J.M. Gracia-Bondía, Hopf algebras in dynamical systems theory, Int. J.
Geom. Methods Mod. Phys. 4 (4) (2007) 577–646.
[CaIs06] F. Casas, A. Iserles, Explicit Magnus expansion for nonlinear equations, J. Phys. A 39 (2006) 5445–5461.
[ChaLiv01] F. Chapoton, M. Livernet, Pre-Lie algebras and the rooted trees operad, Int. Math. Res. Not. 8 (2001) 395–408.
[Cha02] F. Chapoton, Un théorème de Cartier-Milnor-Moore-Quillen pour les algèbres dendriformes et les algèbres braces,
J. Pure Appl. Algebra 168 (2002) 1–18.
[Cha07] F. Chapoton, The anticyclic operad of moulds, Int. Math. Res. Not. 20 (2007).
[Cha08] F. Chapoton, Operads and algebraic combinatorics on trees, Sem. Lothar. Combin. 58 (2008), Article B58c.
[CHNT08] F. Chapoton, F. Hivert, J.-C. Novelli, J.-Y. Thibon, An operational calculus for the Mould operad, Int. Math. Res. Not. 9
(2008).
[De04] E. Deutsch, Ordered trees with prescribed root degrees, node degrees, and branch lengths, Discrete Math. 282
(2004) 89–94.
[Eb02] K. Ebrahimi-Fard, Loday-type algebras and the Rota–Baxter relation, Lett. Math. Phys. 61 (2002) 139–147.
[EG05] K. Ebrahimi-Fard, L. Guo, Free Rota–Baxter algebras and rooted trees, J. Algebra Appl. 7 (2) (2008) 1–28, arXiv:math/
0510266.
[EM09] K. Ebrahimi-Fard, D. Manchon, A Magnus- and Fer-type formula in dendriform algebras, Found. Comput. Math. 9 (3)
(2009) 295–316, arXiv:0707.0607.
[EGM06] K. Ebrahimi-Fard, L. Guo, D. Manchon, Birkhoff type decompositions and the Baker–Campbell–Hausdorff recursion,
Comm. Math. Phys. 267 (2006) 821–845.
[EGP06] K. Ebrahimi-Fard, J.M. Gracia-Bondía, F. Patras, A Lie theoretic approach to renormalization, Comm. Math. Phys. 276
(2007) 519–549, arXiv:hep-th/0609035.
[EGP07] K. Ebrahimi-Fard, J.M. Gracia-Bondía, F. Patras, Rota–Baxter algebras and new combinatorial identities, Lett. Math.
Phys. 81 (2007) 61–75, arXiv:math.CO/0701031.
[EMP07] K. Ebrahimi-Fard, D. Manchon, F. Patras, A Bohnenblust–Spitzer identity for noncommutative Rota–Baxter alge-
bras solves Bogoliubov’s counterterm recursion, J. Noncommut. Geom. 3 (2) (2009) 181–222, arXiv:0705.1265v1
[math.CO].
[EMP07b] K. Ebrahimi-Fard, D. Manchon, F. Patras, New identities in dendriform algebras, J. Algebra 320 (2008) 708–727,
arXiv:0705.2636v1 [math.CO].
[Ec81] J. Ecalle, Les fonctions résurgentes, Tomes I, II, et III, Publ. Math. d’Orsay.
[Ec92] J. Ecalle, Singularités non abordables par la géométrie, Ann. Inst. Fourier (Grenoble) 42 (1–2) (1992) 73–164.
[F58] F. Fer, Résolution de l’equation matricielle U̇ = pU par produit infini d’exponentielles matricielles, Bull. Classe des
Sci. Acad. Roy. Belg. 44 (1958) 818–829.
[GKLLRT95] I.M. Gelfand, D. Krob, A. Lascoux, B. Leclerc, V. Retakh, J.-Y. Thibon, Noncommutative symmetric functions, Adv.
Math. 112 (1995) 218–348, arXiv:hep-th/9407124.
[Ise84] A. Iserles, Solving linear differential equations by exponentials of iterated commutators, Numer. Math. 45 (1984)
183–199.
[Ise01] A. Iserles, A Magnus expansion for the equation Y = AY − Y B, J. Comput. Math. 19 (2001) 15–26.
[Ise02] A. Iserles, Expansions that grow on trees, Notices Amer. Math. Soc. 49 (2002) 430–440.
[IMNZ00] A. Iserles, H.Z. Munthe-Kaas, S.P. Nørsett, A. Zanna, Lie-group methods, Acta Numer. 9 (2000) 215–365.
[IN99] A. Iserles, S.P. Nørsett, On the solution of linear differential equations in Lie groups, in: C.J. Budd, A. Iserles (Eds.),
Geometric Integration: Numerical Solution of Differential Equations on Manifolds, in: Philos. Trans. R. Soc. A,
vol. 357, London Mathematical Society, 1999, pp. 983–1020.
[KO89] S. Klarsfeld, J.A. Oteo, Recursive generation of higher-order terms in the Magnus expansion, Phys. Rev. A 39 (1989)
3270–3273.
[Lod95] J.-L. Loday, Cup-product for Leibniz cohomology and dual Leibniz algebras, Math. Scand. 77 (2) (1995) 189–196.
[Lod01] J.-L. Loday, Dialgebras, in: Lecture Notes in Math., vol. 1763, Springer, Berlin, 2001, pp. 7–66.
[LR98] J.-L. Loday, M. Ronco, Hopf algebra of the planar binary trees, Adv. Math. 139 (1998) 293–309.
[LR04] J.-L. Loday, M. Ronco, Trialgebras and families of polytopes, in: Homotopy Theory: Relations with Algebraic Geom-
etry, Group Cohomology, and Algebraic K-theory, in: Contemp. Math., vol. 346, Amer. Math. Soc., Providence, RI,
2004, pp. 369–398, arXiv:math.AT/0205043.
[Mag54] W. Magnus, On the exponential solution of differential equations for a linear operator, Comm. Pure Appl. Math. 7
(1954) 649–673.
[MP70] B. Mielnik, J. Plebański, Combinatorial approach to Baker–Campbell–Hausdorff exponents, Ann. Inst. H. Poincaré A
XII (1970) 215–254.
[OR00] J.A. Oteo, J. Ros, From time-ordered products to Magnus expansion, J. Math. Phys. 41 (2000) 3268–3277.
[PR02] F. Patras, C. Reutenauer, On Dynkin and Klyachko idempotents in graded bialgebras, Adv. in Appl. Math. 28 (2002)
560–579.
[Ron00] M. Ronco, Primitive elements in a free dendriform algebra, in: Contemp. Math., vol. 207, 2000, pp. 245–263.
[Ron02] M. Ronco, Eulerian idempotents and Milnor–Moore theorem for certain non-commutative Hopf algebras, J. Alge-
bra 254 (2002) 152–172.
[Rot69] G.-C. Rota, Baxter algebras and combinatorial identities, I, II, Bull. Amer. Math. Soc. 75 (1969) 325–329; Bull. Amer.
Math. Soc. 75 (1969) 330–334.
[S] N.J.A. Sloane, The on-line encyclopedia of integer sequences, http://www.research.att.com/~njas/sequences/.
[Sch58] M.P. Schützenberger, Sur une propriété combinatoire des algèbres de Lie libres pouvant être utilisée dans un prob-
lème de mathématiques appliquées, Séminaire Dubreil–Jacotin Pisot (Algèbre et théorie des nombres), Paris, Année
1958/59.
[St87] R.S. Strichartz, The Campbell–Baker–Hausdorff–Dynkin formula and solutions of differential equations, J. Funct.
Anal. 72 (1987) 320–345.
[V63] W. Vogel, Die kombinatorische Lösung einer Operator-Gleichung, Z. Wahrscheinlichkeitstheorie 2 (1963) 122–134.
[Wil67] R.M. Wilcox, Exponential operators and parameter differentiation in quantum physics, J. Math. Phys. 8 (1967) 962–
982.
[Z96] A. Zanna, The method of iterated commutators for ordinary differential equations on Lie groups, Technical Report
1996/NA12, Department of Applied Mathematics and Theoretical Physics, University of Cambridge.
Journal of Algebra
Enumeration of nilpotent loops via cohomology

Daniel Daly, Petr Vojtěchovský ∗
Department of Mathematics, University of Denver, 2360 S. Gaylord St., Denver, CO 80208, USA
Article history: The isomorphism problem for centrally nilpotent loops can be
Received 26 October 2008 tackled by methods of cohomology. We develop tools based on
Available online 27 May 2009 cohomology and linear algebra that either lend themselves to
Communicated by Patrick Dehornoy
direct count of the isomorphism classes (notably in the case
Keywords:
of nilpotent loops of order 2q, q a prime), or lead to efficient
Nilpotent loop classification computer programs. This allows us to enumerate all
Classification of nilpotent loops nilpotent loops of order less than 24.
Loop cohomology © 2009 Published by Elsevier Inc.
Group cohomology
Central extension
Latin square
1. Introduction
A nonempty set Q equipped with a binary operation · is a loop if it possesses a neutral element 1
satisfying 1 · x = x · 1 = x for every x ∈ Q , and if for every x ∈ Q the mappings Q → Q , y → x · y and
Q → Q , y → y · x are bijections of Q . From now on we will abbreviate x · y as xy.
Note that multiplication tables of finite loops are precisely normalized Latin squares, and that
groups are precisely associative loops.
The center Z ( Q ) of a loop Q consists of all elements x ∈ Q such that
xy = yx, (xy ) z = x( yz), ( yx) z = y (xz), ( yz)x = y ( zx)
for every y, z ∈ Q . Normal subloops are kernels of loop homomorphisms. The center Z ( Q ) is a normal
subloop of Q . The upper central series Z 0 ( Q ) Z 1 ( Q ) · · · is defined by
E-mail addresses: ddaly@math.du.edu (D. Daly), petr@math.du.edu (P. Vojtěchovský).
0021-8693/$ – see front matter © 2009 Published by Elsevier Inc.

doi:10.1016/j.jalgebra.2009.03.042
D. Daly, P. Vojtěchovský / Journal of Algebra 322 (2009) 4080–4098 4081

Z 0 ( Q ) = 1, Q / Z i +1 ( Q ) = Z Q / Z i ( Q ) .
If there is n 0 such that Z n−1 ( Q ) < Z n ( Q ) = Q , we say that Q is (centrally) nilpotent of class n.
The goal of this paper is to initiate the classification of small nilpotent loops up to isomorphism,
where by small we mean either that the order | Q | of Q is a small integer, or that the prime factor-
ization of | Q | involves few primes.
Here is a summary of the paper, with A = ( A , +) a finite abelian group and F = ( F , ·) a finite loop
throughout.
Section 2: Central extensions of A by F are in one-to-one correspondence with (normalized) cocy-
cles θ : F × F → A. Let Q( F , A , θ) be the central extension of A by F via θ . If θ − μ is a coboundary
then Q( F , A , θ) ∼
= Q( F , A , μ), that is, the two loops are isomorphic.
Section 3: The group Aut( F , A ) = Aut( F ) × Aut( A ) acts on the cocycles by

(α , β) : θ → (α ,β) θ, (α ,β)
θ : (x, y ) → βθ α −1 x, α −1 y .
For every (α , β) ∈ Aut( F , A ) we have Q( F , A , θ) ∼

= Q( F , A , (α ,β) θ).
Fix a cocycle θ , and let us write θ ∼ μ if there is (α , β) ∈ Aut( F , A ) such that (α ,β) θ − μ is a
coboundary. If θ ∼ μ, we have Q( F , A , θ) ∼ = Q( F , A , μ). If the converse is true for every μ, we say
that θ is separable. We describe several situations in which all cocycles are separable.
Section 4: If all cocycles are separable, the isomorphism problem for central extensions reduces to
the study of the equivalence classes of ∼.
For (α , β) ∈ Aut( F , A ), let

Inv(α , β) = θ; θ − (α ,β) θ is a coboundary ,
and for H ⊆ Aut( F , A ), let

Inv( H ) = Inv(α , β).
(α ,β)∈ H
Then Inv( H ) is a subgroup of cocycles, and Inv( H ) = Inv( H ), where H is the subgroup of Aut( F , A )
generated by H .
For H Aut( F , A ), let

Inv∗ ( H ) = Inv( H ) \ Inv( K ).
H < K Aut( F , A )
When θ ∈ Inv∗ ( H ), the ∼-equivalence class [θ]∼ of θ is a union of precisely [Aut( F , A ) : H ] cosets of
coboundaries. It is not necessarily true that [θ]∼ is contained in Inv∗ ( H ), however, it is contained in

Invc∗ ( H ) = Inv∗ ( K ),
K
where the union is taken over all subgroups K of Aut( F , A ) conjugate to H . Moreover, | Invc∗ ( H )| =
| Inv∗ ( H )| · [Aut( F , A ) : N Aut( F , A ) ( H )], where N G ( H ) is the normalizer of H in G.
Hence, if every cocycle is separable, we can enumerate all central extensions of A by F up to
isomorphism as soon as we know | Inv∗ ( H )| for every H Aut( F , A ), cf. Theorem 4.5.
Section 5: For H , K Aut( F , A ), we have Inv( H ) ∩ Inv( K ) = Inv( H ∪ K ). Hence | Inv∗ ( K )| can be
deduced from the cardinalities of the subgroups Inv( H ) via the principle of inclusion and exclusion
based on the subgroup lattice of Aut( F , A ).
In turn, to find | Inv( H )|, it suffices to determine the cardinalities of Inv(α , β) for every (α , β) ∈ H ,
and the way these subgroups intersect. When A is a prime field, the action θ → (α ,β) θ can be seen
4082 D. Daly, P. Vojtěchovský / Journal of Algebra 322 (2009) 4080–4098
as a matrix operator on the vector space of cocycles, and its preimage of coboundaries is Inv(α , β). It
is therefore not difficult to find Inv(α , β) by means of (computer) linear algebra even for rather large
prime fields A and loops F .
Section 6: When A = Z p , F = Zq and p = q are primes, the dimension of Inv(α , β) can be found
without the assistance of a computer, cf. Theorem 6.5.
Section 7: Since every cocycle is separable when p = 2 and q is odd, Theorems 4.5 and 6.5 give
a formula for the number of nilpotent loops of order 2q, up to isomorphism, cf. Theorem 7.1. The
asymptotic growth of the number of nilpotent loops of order 2q is determined in Theorem 7.3.
Section 8: Every central subloop contains A = Z p for some prime p. Not every choice of A and
F results in separable cocycles, but we can work around this problem when A and F are small
by excluding the subset W ( F , A ) = {θ; Z (Q( F , A , θ)) > A }, because all remaining cocycles will be
separable. When W ( F , A ) is small, the isomorphism problem for {Q( F , A , θ); θ ∈ W ( F , A )} can be
tackled by a direct isomorphism check, using the GAP package LOOPS.
Section 9: This allows us to enumerate all nilpotent loops of order n less than 24 up to isomor-
phism, cf. Table 2. The computational difficulties are nontrivial, notably for n = 16 and n = 20. We
accompany Table 2 by a short narrative describing the difficulties and how they were overcome.
There are 2, 623, 755 nilpotent loops F of order 12, which is why the case n = 24 is out of reach
of the methods developed here.
Section 10: In order not to distract from the exposition, we have collected references to related
work and ideas at the end of the paper.
2. Central extensions, cocycles and coboundaries
We say that a loop Q is a central extension of A by F if A Z ( Q ) and Q / A ∼

= F.
A mapping θ : F × F → A is a normalized cocycle (or cocycle) if it satisfies
θ(1, x) = θ(x, 1) = 0 for every x ∈ F . (2.1)
For a cocycle θ : F × F → A, define Q( F , A , θ) on F × A by

(x, a)( y , b) = xy , a + b + θ(x, y ) . (2.2)
The following characterization of central loop extensions is well known, and is in complete analogy
with the associative case:
Theorem 2.1. The loop Q is a central extension of A by F if and only if there is a cocycle θ : F × F → A such
that Q ∼
= Q( F , A , θ).
The cocycles F × F → A form an abelian group C( F , A ) with respect to addition
(θ + μ)(x, y ) = θ(x, y ) + μ(x, y ).
When A is a field, C( F , A ) is a vector space over A with scalar multiplication
(c θ)(x, y ) = c · θ(x, y ).
Let

Map0 ( F , A ) = τ : F → A ; τ (1) = 0 ,
Hom( F , A ) = {τ : F → A ; τ is a homomorphism of loops},
and observe:
Lemma 2.2. The mapping : Map0 ( F , A ) → C( F , A ), τ →

τ defined by

τ (x, y ) = τ (xy ) − τ (x) − τ ( y )
is a homomorphism of groups with kernel Hom( F , A ).
( F , A) ∼
The image B( F , A ) = C = Map0 ( F , A )/ Hom( F , A ) is a subgroup (subspace) of C( F , A ), and its
elements are referred to as coboundaries.
When A is a field, the vector space Map0 ( F , A ) has basis {τc ; c ∈ F \ {1}}, where

1, if x = c ,
τc : F → A , τc (x) = (2.3)
0, otherwise.
Hence the vector space B( F , A ) is generated by {

τc ; c ∈ F \ {1}}. Observe that for x, y ∈ F \ {1} we
have
⎧
⎪ 1, if xy = c ,
⎪
⎨ −1, if x = c or y = c but not x = y ,

τc (x, y ) = (2.4)
⎪
⎪ −2, if x = y = c ,
⎩
0, otherwise.
Coboundaries play a prominent role in classifications due to this simple observation:
Lemma 2.3. Let

τ ∈ B( F , A ). Then f : Q( F , A , θ) → Q( F , A , θ +
τ ) defined by

f (x, a) = x, a + τ (x)
is an isomorphism of loops.
The converse of Lemma 2.3 does not hold, making the classification of loops up to isomorphism
nontrivial even in highly structured subvarieties, such as groups. Nevertheless it is clear that it suffices
to consider cocycles modulo coboundaries, and we therefore define the (second) cohomology H( F , A ) =
C( F , A )/B( F , A ).
3. The action of the automorphism groups and separability
Let Aut( F , A ) = Aut( F ) × Aut( A ). The group Aut( F , A ) acts on C( F , A ) via

θ → (α ,β) θ, (α ,β)
θ : (x, y ) → βθ α −1 x, α −1 y .
Indeed, we have (αγ ,βδ) θ = (α ,β) ((γ ,δ) θ), and (α ,β) (θ + μ) = (α ,β) θ + (α ,β) μ. Since
(α ,β)
τ = β
τ α −1 ,
the action of Aut( F , A ) on C( F , A ) induces an action on B( F , A ) and on H( F , A ). Moreover:
Lemma 3.1. Let (α , β) ∈ Aut( F , A ). Then f : Q( F , A , θ) → Q( F , A , (α ,β) θ) defined by
f (x, a) = (α x, β a)
is an isomorphism of loops.
Proof. Let · be the multiplication in Q( F , A , θ) and ∗ the multiplication in Q( F , A , (α ,β) θ). Then

f (x, a) · ( y , b) = f xy , a + b + θ(x, y ) = α (xy ), β a + b + θ(x, y )

= α (x)α ( y ), β(a) + β(b) + βθ α −1 α x, α −1 α y

= α (x)α ( y ), β(a) + β(b) + (α ,β) θ(α x, α y )
= (α x, β a) ∗ (α y , β b) = f (x, a) ∗ f ( y , b). 2
As in Section 1, write θ ∼ μ if there is (α , β) ∈ Aut( F , A ) such that (α ,β) θ − μ ∈ B( F , A ). Then ∼ is

an equivalence relation on C( F , A ), and the equivalence class of θ is
(α ,β)
[θ]∼ = θ + B( F , A ) .
(α ,β)∈Aut( F , A )
By Lemmas 2.3 and 3.1, if θ ∼ μ then Q( F , A , θ) ∼ = Q( F , A , μ). We say that θ is separable if the
converse is also true, that is, if Q( F , A , θ) ∼
= Q( F , A , μ) if and only if θ ∼ μ.
We remark that there exists an inseparable cocycle already in C(Z6 , Z2 ). In the rest of this section
we describe situations that guarantee separability.
Proposition 3.2. Let Q = Q( F , A , θ). If Aut( Q ) acts transitively on { K Z ( Q ); K ∼

= A, Q / K ∼
= F } then
θ is separable.
Proof. Let Q = Q( F , A , θ), and let f : Q → Q( F , A , μ) be an isomorphism. Let K = f −1 (1 × A ).

By our assumption, there is g ∈ Aut( Q ) such that g (1 × A ) = K . Then f g : Q → Q( F , A , μ) is an
isomorphism mapping 1 × A onto itself. We can therefore assume without loss of generality that
already f has this property.
Denote by · the multiplication in Q and by ∗ the multiplication in Q( F , A , μ). Define β : A → A
by (1, β(a)) = f (1, a). Then

1, β(a + b) = f (1, a + b) = f (1, a) · (1, b) = f (1, a) ∗ f (1, b) = (1, β a) ∗ (1, β b)
= (1, β a + β b),
which means that β ∈ Aut( A ).

Define τ : F → A and α : F → F by f (x, 0) = (α x, τ x). Since f (1, 0) = (1, 0), we have τ ∈
Map0 ( F , A ). Moreover, calculating modulo A in both loops, we have

α (xy ), 0 ≡ f (xy , 0) ≡ f (x, 0) · ( y , 0) ≡ f (x, 0) ∗ f ( y , 0) ≡ (α x, 0) ∗ (α y , 0) ≡ α (x)α ( y ), 0 ,
and α ∈ Aut( F ) follows.

The isomorphism f satisfies

f (x, a) = f (1, a) · (x, 0) = f (1, a) ∗ f (x, 0) = (1, β a) ∗ (α x, τ x) = (α x, β a + τ x).
If is therefore the composition of the isomorphism (x, a) → (x, a + β −1 τ x) of Lemma 2.3 (with β −1 τ
in place of τ ) and of the isomorphism (x, a) → (α x, β a) of Lemma 3.1. This means that μ = (α ,β) (θ +
β
−1 τ ), so μ ∈ (α ,β) θ + B( F , A ), θ ∼ μ. 2
We now investigate separability in abelian groups. The next two results can be proved in many
ways from the Fundamental Theorem of Finitely Generated Abelian Groups, which we use without
warning.
Lemma 3.3. Let p be a prime, and let
A = Z p e 1 × · · · × Z p en (3.1)
be an abelian p-group, where e 1 · · · en . Let x ∈ A be an element of order p. Then there exists a unique
integer e j such that: there is a complemented cyclic subgroup B A satisfying x ∈ B and | B | = p e j . Moreover,
A /x ∼
= Z p f 1 × · · · × Z p fn ,
where f i = e i for every i = j, and f j = e j − 1.
Proof. Every element x ∈ A of order p is of the form

x = x1 p e1 −1 , . . . , xn p en −1 ,
where xi ∈ {0, . . . , p − 1} for every 1 i n, and where xi = 0 for some 1 i n. Let j be the least
integer such that x j = 0. Consider the element
x
y= = 0, . . . , 0, x j , x j +1 p e j+1 −e j , . . . , xn p en −e j .
p e j −1
Then B = y contains x, | B | = p e j , and
C = Z p e 1 × · · · × Z p e j −1 × 0 × Z p e j +1 × · · · × Z p e n
is a complement of B in A (that is, B ∩ C = 0 and B ∪ C = A). 2
Proposition 3.4. Let A be a finite abelian group. For a prime p dividing | A | and for a finite abelian group F of
order | A |/ p, let

X ( p , F ) = x ∈ A ; |x| = p and A /x ∼
=F .
Then the sets X ( p , F ) that are nonempty are precisely the orbits of the action of Aut( A ) on A.
Proof. For a prime p, let A p be the p-primary component of A. Then A = A p 1 × · · · × A pm , for

some distinct primes p 1 , . . . , pm , and Aut( A ) = Aut( A p 1 ) × · · · × Aut( A pn ). (For a detailed proof, see
[8, Lemma 2.1].) We can therefore assume that A = A p is a p-group.
It is obvious that every orbit of Aut( A ) is contained in one of the sets X ( p , F ). It therefore suffices
to prove that if x, y ∈ X ( p , F ) then there is ϕ ∈ Aut( A ) such that ϕ (x) = y.
Let A be as in (3.1). If A is cyclic of order p e1 then A /x ∼ = Z pe1 −1 , and we can assume that
x = ap e1 −1 , y = bp e1 −1 , where 1 a, b p − 1. The automorphism of A determined by 1 → b/a
(modulo p) then maps a to b and hence x to y.
Assume that n > 1. Let B x , B y be the complemented cyclic subgroups B obtained by Lemma 3.3
for x, y, respectively. Then | B x | = | B y | since A /x ∼
= A / y , and hence the integer e j determined by
Lemma 3.3 is the same for x and y. We can in fact assume that already j is the same. Furthermore,
we can assume that the isomorphism from
A /x ∼
= Z pe1 × · · · × Z pe j−1 × B x /x × Z pe j+1 × · · · × Z pen
to
A / y ∼
= Z pe1 × · · · × Z pe j−1 × B y / y × Z pe j+1 × · · · × Z pen
is componentwise, and maps B x /x to B y / y . We can then extend B x /x → B y / y to an isomor-

phism B x → B y while sending x to y by the case n = 1, and hence obtain the desired automorphism
of A. 2
Corollary 3.5. Let Q = Q( F , A , θ) be an abelian group, A = Z p , p a prime. Then θ is separable.
Proof. Combine Propositions 3.2 and 3.4. 2
Finally, we show that all cocycles are separable in “small” situations.
Lemma 3.6. There is no loop Q with [ Q : Z ( Q )] = 2.
Proof. Assume, for a contradiction, that | Q / Z ( Q )| = 2, and let a ∈ Q \ Z ( Q ). Then every element of
Q can be written as ai z, where i ∈ {0, 1} and z ∈ Z ( Q ). For every i, j, k ∈ {0, 1} and z1 , z2 , z3 ∈ Z ( Q )
we have ai z1 · (a j z2 · ak z3 ) = ai (a j ak ) · z1 z2 z3 , and similarly, (ai z1 · a j z2 ) · ak z3 = (ai a j )ak · z1 z2 z3 . The
two expressions are equal if any of i, j, k vanishes. So it remains to discuss the case i = j = k = 1.
But then a(aa) = (aa)a, because a2 ∈ Z ( Q ). Hence Q is a group. It is well known that if Q is a group
and Q / Z ( Q ) is cyclic then Q = Z ( Q ), a contradiction. 2
Lemma 3.6 cannot be improved: for every odd prime p there is a nonassociative loop Q such that
| Q / Z ( Q )| = p, cf. Theorem 7.1.
Lemma 3.7. Let Q = Q( F , A , θ), A = Z p , p a prime. Assume further that one of the following conditions is
satisfied:
(i) | Q | = p,
(ii) | Q | = pq, where q is a prime,
(iii) [ Q : Z ( Q )] 2,
(iv) | Q | < 12.
Then θ is separable.
Proof. When (i) or (iii) hold then Q is an abelian group by Lemma 3.6, and so θ is separable by
Corollary 3.5.
Assume that (ii) holds. If Z ( Q ) > A then Z ( Q ) = Q and we are done by Corollary 3.5. Else
Z ( Q ) = A and θ is separable by Proposition 3.2, for trivial reasons.
To finish (iv), it remains to discuss the case | Q | = 8. If Z ( Q ) = A, θ is separable by Proposition 3.2.
If Z ( Q ) > A then Z ( Q ) = Q by Lemma 3.6, and we are done by Corollary 3.5. 2
4. The invariant subspaces
For (α , β) ∈ Aut( F , A ), let

Inv(α , β) = θ ∈ C( F , A ); θ − (α ,β) θ ∈ B( F , A ) . (4.1)
For ∅ = H ⊆ Aut( F , A ), let

Inv( H ) = Inv(α , β). (4.2)
(α ,β)∈ H
Lemma 4.1. Let ∅ = H ⊆ Aut( F , A ). Then Inv( H ) = Inv( H ).

Proof. Assume that θ ∈ Inv(α , β) ∩ Inv(γ , δ). Then θ − (α ,β) θ ∈ B( F , A ) and θ − (γ ,δ) θ ∈ B( F , A ). The
second equation is equivalent to (α ,β) θ − (α ,β) ((γ ,δ) θ) ∈ B( F , A ). Adding this to the first equation yields
θ − (α ,β) ((γ ,δ) θ) = θ − (αγ ,βδ) θ ∈ B( F , A ). 2
Corollary 4.2. Let H , K Aut( F , A ). Then Inv( H ) ∩ Inv( K ) = Inv( H ∪ K ).
For α , γ ∈ Aut( F ) and β , δ ∈ Aut( A ), let γ α = γ αγ −1 , δ β = δβδ −1 .
Lemma 4.3. Let (α , β), (γ , δ) ∈ Aut( F , A ). Then θ ∈ Inv(α , β) if and only if (γ ,δ) θ ∈ Inv(γ α , δ β).
Proof. The following conditions are equivalent:
(γ ,δ)

θ ∈ Inv γ α , δ β ,
(γ ,δ) γ δ
θ − ( α , β) (γ ,δ) θ ∈ B( F , A ),
(γ ,δ)
θ − (γ α ,δβ) θ ∈ B( F , A ),
(γ ,δ)

θ − (α ,β) θ ∈ B( F , A ),
θ − (α ,β) θ ∈ B( F , A ),
θ ∈ Inv(α , β). 2
For H Aut( F , A ), let

Inv∗ ( H ) = θ ∈ C( F , A ); θ ∈ Inv(α , β) if and only if (α , β) ∈ H ,

Invc∗ ( H ) = Inv∗ (α ,β) H .
(α ,β)∈Aut( F , A )
As we are going to see, the cardinality of the equivalence class [θ]∼ can be easily calculated for
θ ∈ Inv∗ ( H ), provided θ is separable.
If G is a group and H G, let N G ( H ) = {a ∈ G ; a H = H } be the normalizer of H in G.
Lemma 4.4. Let H G = Aut( F , A ). Then | Invc∗ ( H )| = | Inv∗ ( H )| · [G : N G ( H )].
Proof. Since a H = b H if and only if a−1 b ∈ N G ( H ), there are precisely [G : N G ( H )] subgroups K of G

conjugate to H .
Assume that K = H are conjugate, K = (α ,β) H . The mapping f : C( F , A ) → C( F , A ), θ → (α ,β) θ
is a bijection. By Lemma 4.3, f (Inv( H )) = Inv( K ), and f (Inv∗ ( H )) = Inv∗ ( K ), proving | Inv∗ ( H )| =
| Inv∗ ( K )|. Since K = H , we have Inv∗ ( H ) ∩ Inv∗ ( K ) = ∅ by definition. 2
For a group G, denote by Subc (G ) a set of subgroups of G such that for every H G there is
precisely one K ∈ Subc (G ) such that K is conjugate to H .
Theorem 4.5. Let F be a loop and A an abelian group. Assume that θ is separable for every θ ∈ C( F , A ). Let
G = Aut( F , A ). Then there are
| Invc∗ ( H )| | Inv∗ ( H )|
= (4.3)
|B( F , A )| · [G : H ] |B( F , A )| · [ N G ( H ) : H ]
H ∈Subc (G ) H ∈Subc (G )
central extensions of A by F , up to isomorphism.

Proof. By Lemma 4.1,

C( F , A ) = Inv∗ ( H ) = Invc∗ ( H ),
H G H ∈Subc (G )
where the unions are disjoint. Let θ ∈ Invc∗ ( H ), for some H G. Let X = {μ ∈ C( F , A ); Q( F , A , μ) ∼
=
Q( F , A , θ)}. Since θ is separable, we have

X = [θ]∼ = (α ,β)
θ + B( F , A ) ⊆ Inv∗ ( H ), c
(α ,β)∈G
where the first equality follows by separability of θ , and the inclusion from Lemma 4.3.
Let K be the unique conjugate of H such that θ ∈ Inv∗ ( K ). We have (α ,β) θ − (γ ,δ) θ ∈ B( F , A ) if and
−1 −1
only if θ − (α γ ,β δ) θ ∈ B( F , A ), which holds if and only if θ ∈ Inv(α −1 γ , β −1 δ). Since θ ∈ Inv∗ ( K ),
we see that (α ,β) θ − (γ ,δ) θ ∈ B( F , A ) holds if and only if (α −1 γ , β −1 δ) ∈ K , or (α , β) K = (γ , δ) K .
Hence [θ]∼ is a union of [G : K ] = [G : H ] cosets of B( F , A ). We have established the first sum of (4.3).
The second sum then follows from Lemma 4.4. 2
5. Calculating the subspaces Inv(α , β) by computer
Assume throughout this section that A = Z p , where p is a prime. Then C( F , A ), B( F , A ) and

H( F , A ) are vector spaces over GF( p ).
For (α , β) ∈ Aut( F , A ) let R = R (α , β), S = S (α , β) be the linear operators C( F , A ) → C( F , A ) de-
fined by
R (α , β)θ = (α ,β) θ,
S (α , β)θ = θ − (α ,β) θ.
Hence R (α , β) is invertible, and S (α , β) = I − R (α , β), where I : C( F , A ) → C( F , A ) is the identity

operator.
As β ∈ Aut(Z p ) is a scalar multiplication by β(1), let us identify β with β(1). Then R (α , β) is a
matrix operator with rows and columns labeled by pairs of nonidentity elements of F , where the only
nonzero coefficient in row (x, y ) is −β in column (α −1 x, α −1 y ).
By definition of Inv(α , β) and S (α , β), we have

Inv(α , β) = θ ∈ C( F , A ); S (α , β)θ ∈ B( F , A ) = S (α , β)−1 B( F , A ).
In order to calculate Inv(α , β), we can proceed as follows:
• calculate the subspace B( F , A ) as the span of {

τc ; 1 = c ∈ F },
• calculate the kernel Ker S (α , β) and image Im S (α , β) as usual,
• find a basis B of the subspace B( F , A ) ∩ Im S (α , β),
• for b ∈ B , find a particular solution θb to the system S (α , β)θb = b,
• then Inv(α , β) = Ker S (α , β) ⊕ θb ; b ∈ B .
In particular, with S = S (α , β), we have

dim Inv(α , β) = dim Ker S + dim Im S ∩ B( F , A )

= dim Ker S + dim Im S + dim B( F , A ) − dim Im S + B( F , A )
2
= | F | − 1 + dim B( F , A ) − dim Im S + B( F , A ) .
Using a computer, it is therefore not difficult to find Inv(α , β) and its dimension even for rather
large loops A = Z p and F . See Section 9 for more details.
Remark 5.1. If it is preferable to operate modulo coboundaries, note that S (α , β)(θ + τ ) = S (α , β)θ +
τ −
β τ α −1 , and view S (α , β) as a linear operator S (α , β) : H( F , A ) → H( F , A ) defined by

S (α , β) θ + B( F , A ) = θ − (α ,β) θ + B( F , A ).
Then Inv(α , β)/B( F , A ) = Ker S (α , β).
6. The subspaces Inv(α , β) for A = Z p , F = Z q
If H K Aut( F , A ), we have Inv( K ) Inv( H ). Hence the subgroups Inv( H ) will be incident in
accordance with the upside down subgroup lattice of Aut( F , A ), except that some edges in the lattice
can collapse, i.e., it can happen that Inv( H ) = Inv( K ) although H < K :
Example 6.1. Let A = Z2 , F = Z2 × Z2 . Then Aut( F ) ∼

= S 3 . Let H be the subgroup of Aut( F ) generated
by a 3-cycle. Then it turns out that Inv( H ) = Inv(Aut( F )).
Such a collapse has no impact on the formula (4.3) of Theorem 4.5, since only subgroups H with
Inv∗ ( H ) = ∅ contribute to it.
We proceed to determine dim Inv(α , β).
In addition to the operators R (α , β) and S (α , β) on C( F , A ), define T (α , β) by
T (α , β)θ = θ + R (α , β)θ + · · · + R (α , β)k−1 θ,
where k = |α |.
Lemma 6.2. Let R, S, T be operators on a finite-dimensional vector space V such that R k = I , S = I − R,

T = I + R + · · · + R k−1 . Then Im T Ker S and Im S Ker T . If Im T = Ker S then Ker T = Im S.
Proof. We have T S = ( I + R + · · · + R k−1 )( I − R ) = I − R k = 0 and S T = ( I − R )( I + R + · · · + R k−1 ) = 0,

which shows Im T Ker S, Im S Ker T .
Assume that Im T = Ker S. By the Fundamental Homomorphism Theorem,
dim Im T + dim Ker T = dim V = dim Im S + dim Ker S = dim Im S + dim Im T ,
so dim Ker T = dim Im S. Since Im S Ker T , we conclude that Im S = Ker T . 2
Lemma 6.3. Let p, q be primes, A = Z p , F = Zq , α ∈ Aut( F ), β ∈ Aut( A ).
(i) If |β| does not divide |α | then S (α , β) is invertible.

(ii) If |β| divides |α | then Ker S (α , β) = Im T (α , β) and dim Ker S (α , β) = (q − 1)2 /|α |.
Proof. Let F ∗ = F \ {0}, k = |α |. The automorphism α acts on F ∗ × F ∗ via (x, y )α = (α −1 x, α −1 y ).

Every α -orbit has size k. Let t = (q − 1)2 /k, and let O1 , . . . , Ot be all the distinct α -orbits on F ∗ × F ∗ .
Let R = R (α , β), S = S (α , β), T = T (α , β). Throughout the proof, let θ ∈ Ker S, i.e.,

θ(x, y ) = βθ α −1 x, α −1 y (6.1)
for every x, y ∈ F ∗ . For every 1 i t, let (xi , y i ) ∈ Oi . Define θi ∈ C( F , A ) by

θ(xi , y i ), if (x, y ) = (xi , y i ),
θi (x, y ) =
0, otherwise.
Then

t
t
θ= T θi = T θi ,
i =1 i =1
thus Ker S Im T .
The condition (6.1) implies θ(x, y ) = β k θ(x, y ) for every x, y ∈ F ∗ . If |β| does not divide |α |, we
have β k = 1, and therefore θ = 0, proving Ker S = 0.
Assume that |β| divides |α |. Then R k = I , and Im T Ker S by Lemma 6.2. Thus Im T = Ker S and
Ker T = Im S. Since θ is determined by the values θ(xi , y i ), for 1 i t, and since these values can
be arbitrary, we see that dim Ker S = t. 2
Lemma 6.4. Let p, q be distinct primes, A = Z p , F = Zq , α ∈ Aut( F ), β ∈ Aut( A ), and assume that |β|
divides |α |. Then

1
dim Ker T (α , β) ∩ B( F , A ) = (q − 1) 1 − .
|α |
τc ; c ∈ F ∗ } is linearly independent thanks to p = q. Let

Proof. The set {

τ= λc
τc ,
c∈ F ∗
for some λc ∈ A. An inspection of (2.4) reveals that
τc = β
R (α , β) ταc . (6.2)
Thus
τ belongs to Ker T (α , β) ∩ B( F , A ) if and only if

λc
τc + β ταc + · · · + β k−1
λc λc
τα (k−1) c = 0. (6.3)
c c c
The coefficient of τc in (6.3) is λc + βλα −1 c + · · · + β k−1 λα −(k−1) c , so the system (6.3) can be rewritten
in terms of the coefficients λc as
λc + βλα −1 c + · · · + β k−1 λα −(k−1) c = 0, for c ∈ F ∗ . (6.4)
For any 1 i k, the equation for c is a scalar multiple of the equation for α i c. On the other
hand, each equation involves scalars c from only one orbit of α . Hence (6.4) reduces to a system
of (q − 1)/|α | linearly independent equations in variables λc , c ∈ F ∗ . It follows that the subspace of
homogeneous solutions has dimension (q − 1)(1 − 1/|α |). 2
Theorem 6.5. Let p = q be primes, A = Z p , F = Zq , α ∈ Aut( F ), β ∈ Aut( A ). Then
Inv(α , β) = Ker S (α , β) + B( F , A ).
Moreover,

q − 1, if |β| does not divide |α |,
dim Inv(α , β) =
(q − 1) + (q − 1)(q − 2)/|α |, otherwise.
Thus

0, if |β| does not divide |α |,
dim Inv(α , β)/B( F , A ) =
(q − 1)(q − 2)/|α |, otherwise.
Proof. Let R = R (α , β), S = S (α , β), T = T (α , β) and B = B( F , A ). Assume that |β| does not di-
vide |α |. Then S is invertible by Lemma 6.3, so Inv(α , β) = S −1 B = B = Ker S + B, and we have
dim Inv(α , β) = dim B = q − 1 thanks to p = q.
Now assume that |β| divides |α |. By Lemmas 6.2, 6.3 and 6.4, we have Im T = Ker S, dim(Im S ∩
B ) = (q − 1)(1 − 1/|α |), and dim Ker S = (q − 1)2 /|α |, so
dim Inv(α , β) = dim Ker S + dim(Im S ∩ B )

= (q − 1)2 /|α | + (q − 1) 1 − 1/|α |
= (q − 1) + (q − 1)(q − 2)/|α |.
It remains to show that Inv(α , β) = Ker S + B.

Let k = |α |. The coboundaries { τc ; c ∈ F ∗ } are linearly independent thanks to p = q.For 1 i
m = (q − 1)/k, let c i be a representative of the coset c i α in F ∗ , and assume that
m
i =1 c i α =
∗
F . By (6.2), the set { R
τci ; 0
k − 2, 1 i m} is linearly independent, and so is its S-image
{ R

τci − R
+1
τci ; 0
k − 2, 1 i m} ⊆ B. This shows that dim( S ( B ) ∩ B ) (q − 1)(1 − 1/k). On
the other hand, dim(Im S ∩ B ) = (q − 1)(1 − 1/k) by Lemma 6.4. Thus Im S ∩ B = S ( B ) ∩ B = S ( B ). But
this means that Inv(α , β) = S −1 B is equal to Ker S + B. 2
7. Nilpotent loops of order 2q, q a prime
For n 1, let N (n) be the number of nilpotent loops of order n up to isomorphism. In this section
we find a formula for N (2q), where q is a prime, and describe the asymptotic behavior of N (2q) as
q → ∞.
Loops of order 4 are associative, and, up to isomorphism, there are 2 nilpotent groups of order 4,
namely Z4 and Z2 × Z2 .
Theorem 7.1. Let q be an odd prime. For a positive integer d, let
Pred(d) = {d ; 1 d < d, d/d is a prime}
be the set of all maximal proper divisors of d. Then the number of nilpotent loops of order 2q up to isomorphism
is

1
N (2q) = 2(q−2)d + (−1)| D | · 2(q−2) gcd D . (7.1)
d
d divides q−1 ∅= D ⊆Pred(d)
Proof. By Lemma 3.6, the only central extension of Zq by Z2 is the cyclic group Z2q . Since this
group can also be obtained as a central extension of Z2 by Zq , we can set A = Z2 , F = Zq . Then by
Lemma 3.7, every θ ∈ C( F , A ) is separable, so Theorem 4.5 applies.
Table 1
The number N (2q) of nilpotent loops of order 2q, q a prime, up to isomorphism.
2q N (2q)
4 2
6 3
10 1,044
14 178, 962, 784
22 123, 794, 003, 928, 541, 545, 927, 226, 368
26 453, 709, 822, 561, 251, 284, 623, 981, 727, 533, 724, 162, 048
34 110, 427, 941, 548, 649, 020, 598, 956, 093, 796, 432, 407, 322, 294, 493, 291, 283, 427, 083, 203, 517, 192, 617, 984
We have Aut( F , A ) = Aut( F ) = α ∼

= Zq−1 . The subgroup structure of Aut( F ) is therefore transpar-
ent: for every divisor d of q − 1 there is a unique subgroup H d = α d of order (q − 1)/d, and if d,
d are two divisors of q − 1 then H d ∪ H d = H gcd(d,d ) .
By Theorem 6.5,

dim Inv( H d ) = dim Inv αd = (q − 1) + (q − 1)(q − 2)/ (q − 1)/d = (q − 1) + (q − 2)d,
so dim(Inv( H d )/B( F , A )) = (q − 2)d.

Note that H d is a maximal subgroup of H d if and only if d ∈ Pred(d). For ∅ =
D ⊆ Pred(d), we
have H d ; d ∈ D = H gcd D , and so

Inv( H d ) = Inv( H gcd D )
d ∈ D
by Corollary 4.2. Then
∗
Inv ( H d ) = Inv( H d ) + (−1)| D | · Inv( H gcd D )
∅= D ⊆Pred(d)
by the principle of inclusion and exclusion.

As Aut( F ) is abelian, [ N Aut( F ) ( H d ) : H d ] = [Aut( F ) : H d ] = d. The formula (7.1) then follows by The-
orem 4.5. 2
Example 7.2. To illustrate (7.1), let us determine N (14) = N (2 · 7). The divisors of q − 1 = 6 are 6, 3,
2, 1. Hence

N (14) = 25·6 − 25·3 − 25·2 + 25·1 /6 + 25·3 − 25·1 /3 + 25·2 − 25·1 /2 + 25·1 /1
= 178, 962, 784. (7.2)
Table 1 lists the number of nilpotent loops of order 2q up to isomorphism for small primes q. (It is
by no means difficult to evaluate (7.1) for larger primes, say up to q 100, but the decimal expansion
of N (2q) becomes too long to display neatly in Table 1.)
Here is the asymptotic growth of N (2q):
Theorem 7.3. Let q be an odd prime. Then the number of nilpotent loops of order 2q up to isomorphism is
approximately 2(q−2)(q−1) /(q − 1). More precisely,
q−1
lim N (2q) · = 1.
q prime, q→∞ 2(q−2)(q−1)
Proof. We prove the assertion by a simple estimate. To illustrate the main idea, note that (7.2) can be
rewritten as
230 /6 + 215 (1/3 − 1/6) + 210 (1/2 − 1/6) + 25 (1 − 1/2 − 1/3 + 1/6).
Thus, upon rewriting (7.1) in a similar fashion, there will be no more than q − 1 summands, each of
the form

2(q−2)d (1/d1 ± 1/d2 ± · · · ± 1/dm ). (7.3)
A reciprocal 1/d appears in (7.3) if and only if there is a divisor d of q − 1 and D ⊆ Pred(d) such
that gcd D = d . Now, for every divisor d of q − 1 there is at most one subset D ⊆ Pred(d) such
that gcd D = d (because if D = {e 1 , . . . , en }, d/e i = p i is a prime, then gcd D = d/( p 1 · · · pn ) uniquely
determines D). Hence the number of reciprocals in (7.3) cannot exceed q − 1. Finally, the largest
proper divisor of q − 1 is (q − 1)/2. Altogether,
2(q−2)(q−1) 2(q−2)(q−1)
− (q − 1)2(q−2)(q−1)/2 (q − 1) N (2q) + (q − 1)2(q−2)(q−1)/2 (q − 1),
q−1 q−1
thus
(q − 1)3 (q − 1) (q − 1)3
1 − (q−2)(q−1)/2 N (2q) · (q−2)(q−1) 1 + (q−2)(q−1)/2 ,
2 2 2
and the result follows by the Squeeze Theorem. 2
8. Inseparable cocycles
Let A = Z p , F be as usual. The easiest (but slow) way to deal with inseparable cocycles θ ∈ C( F , A )
is to treat separately the subset

W ( F , A ) = θ ∈ C( F , A ); Z Q( F , A , θ) > A ⊆ C( F , A ).
We will refer to elements of W ( F , A ) informally as large center cocycles. Note that the adjective “large”
is relative to A. The subset W ( F , A ) can be determined computationally as follows.
Let Q = Q( F , A , θ). The element (x, a) belongs to Z ( Q ) if and only if {(x, b); b ∈ A } ⊆ Z ( Q ),
which happens if and only if x ∈ Z ( F ) and θ satisfies
θ(x, y ) = θ( y , x),
θ(x, y ) + θ(xy , z) = θ( y , z) + θ(x, yz),
θ( y , x) + θ( yx, z) = θ(x, z) + θ( y , xz),
θ( y , z) + θ( yz, x) = θ( z, x) + θ( y , zx)
for all y, z ∈ F . The first condition ensures that (x, a) commutes with all elements of Q , and the last
three conditions ensure that (x, a) associates with all elements of Q , no matter in which position
(x, a) happens to be in the associative law. (Note that the last condition is a consequence of the first
three.)
Hence for every 1 = x ∈ Z ( F ) we can solve the above linear equations and obtain the subspace
W x ( F , A ) C( F , A ) such that θ ∈ W x ( F , A ) if and only if (x, A ) ⊆ Z (Q( F , A , θ)). Then

W ( F , A) = W x ( F , A ),
1=x∈ Z ( F )
and this subset can be determined by the principle of inclusion and exclusions on the subspaces
W x ( F , A ), 1 = x ∈ Z ( F ).
Importantly, every cocycle θ ∈ C( F , A ) \ W ( F , A ) is separable, since then Q( F , A , θ) possesses a
unique central subloop of the cardinality | A |, namely A.
When A, F are small, we can complete the isomorphism problem by first constructing the loops
Q( F , A , θ) for all θ ∈ W ( F , A )/B( F , A ) and then sorting them up to isomorphism by standard algo-
rithms of loop theory. Since these algorithms are slow, dealing with large center cocycles is the main
obstacle in pushing the enumeration of nilpotent loops past order n = 23.
9. Enumeration of nilpotent loops of order less than 24
The results are summarized in Table 2. A typical line of Table 2 can be read as follows: “#Q ” is
the number of nilpotent loops (up to isomorphism) of order n that are central extensions of the cyclic
group A = Z p by the nilpotent loop F of order n/ p. If only the order of F is given, F is any of the
nilpotent loops of order n/ p. If no information about A and F is given, any pair ( A , F ) with A = Z p ,
F nilpotent of order n/ p can be used. Finally, “#Q , Z ( Q ) > A” is the number of nilpotent loops with
center larger than A. Since this makes sense only when A is specified, we omit “#Q , Z ( Q ) > A” in
the other cases.
By Lemma 3.7, we can apply the formula (4.3) safely until we reach order n = 12.
For every prime p there is a unique nilpotent loop of order p up to isomorphism, namely the
cyclic group Z p .
The number of nilpotent loops of order 2q, q a prime, is determined by Theorem 7.1. Note, how-
ever, that the theorem does not produce the loops. Since we need all nilpotent loops of order 6 and 10
explicitly in order to compute the number of nilpotent loops of order 12, 18 and 20, we must obtain
the nilpotent loops of order 6, 10 by other means (a direct isomorphism check on H( F , A ) will do).
In accordance with Theorem 7.1, there are 3 nilpotent loops of order 6. Beside the cyclic group of
order 6, the other two loops are
L 6,2 1 2 3 4 5 6 L 6,3 1 2 3 4 5 6
1 1 2 3 4 5 6 1 1 2 3 4 5 6
2 2 1 4 3 6 5 2 2 1 4 3 6 5
3 3 4 5 6 1 2, 3 3 4 5 6 1 2.
4 4 3 6 5 2 1 4 4 3 6 5 2 1
5 5 6 2 1 3 4 5 5 6 1 2 4 3
6 6 5 1 2 4 3. 6 6 5 2 1 3 4
9.1. n = 8
Case A = Z2 , F = Z4 . We have Aut( A ) = 1, Aut( F ) = α ∼ = Z2 , and dim Hom( F , A ) = 1,

dim B( F , A ) = 2, dim C( F , A ) = 9. Computer yields dim Inv(α ) = 7. Hence (4.3) shows that there are
27 29 − 27
+ = 80
22 22 · 2
central extensions of Z2 by Z4 , up to isomorphism.

Table 2
The number of nilpotent loops up to isomorphism.
n A F #Q , Z ( Q ) > A #Q
4 Z2 Z2 2 2
6 Z2 Z3 1 3
8 Z2 Z4 2 80
8 Z2 Z2 × Z 2 2 60
8 Z2 4 3 139
9 Z3 Z3 2 10
10 Z2 Z5 1 1044
12 Z2 Z6 6 1,049,560
12 Z2 L 6,2 4 1,048,576
12 Z2 L 6,3 4 525,312
12 Z2 6 11 2,623,485
12 Z3 Z4 1 196
12 Z3 Z2 × Z 2 1 76
12 Z3 4 2 272
12 2,623,755
14 Z2 Z7 1 178,962,784
15 Z3 Z5 1 66,626
15 Z5 Z3 1 5
15 66,630
16 Z2 8 9,284 466,409,543,467,341
18 Z2 9 34 157,625,987,549,892,128
18 Z3 Z6 10 2,615,147,350
18 Z3 L 6,2 14 5,230,176,602
18 Z3 L 6,3 10 2,615,147,350
18 Z3 6 34 10,460,471,302
18 157,625,998,010,363,396
20 Z2 10 2,798,987 4,836,883,870,081,433,134,082,379
20 Z5 Z4 1 1985
20 Z5 Z2 × Z 2 1 685
20 Z5 4 2 2670
20 4,836,883,870,081,433,134,085,047
21 Z3 Z7 1 17,157,596,742,628
21 Z7 Z3 1 6
21 17,157,596,742,633
22 Z2 Z11 1 123,794,003,928,541,545,927,226,368
Case A = Z2 , F = Z2 × Z2 . We have Aut( A ) = 1, Aut( F ) = σ , ρ ∼ = S 3 , where |σ | = 2, |ρ | = 3.

Furthermore, dim Hom( F , A ) = 2, dim B( F , A ) = 1 and dim C( F , A ) = 9. The three subspaces Inv(σ ),
Inv(σρ ), Inv(σρ 2 ) have dimension 6, and any two of them intersect precisely in Inv(ρ ) (see Exam-
ple 6.1), which has dimension 3. By (4.3), there are
23 26 − 23 29 − 3 · 26 + 2 · 23
+3· + = 60
2 2·3 2·6
central extensions of Z2 by Z2 × Z2 .
In order to pinpoint the number of nilpotent loops of order 8, we must determine which loops
are obtained both as central extensions of Z2 by Z4 and of Z2 by Z2 × Z2 . First of all, Z2 × Z4 is
such a loop. Assume that Q is another such loop. Then | Z ( Q )| > 2 and hence Q is an abelian group
by Lemma 3.6. Now, Q = Z8 since every factor Z 8 /x by an involution is isomorphic to Z4 . Finally,
Q = Z2 × Z2 × Z2 since every factor by an involution is of exponent 2. We conclude that there are

80 + 60 − 1 = 139 nilpotent loops of order 8.
9.2. n = 9
We have A = Z3 , F = Z3 , dim Hom( F , A ) = 1, dim B( F , A ) = 1 and dim C( F , A ) = 4. Also, Aut( A ) =

β ∼
= Z2 , Aut( F ) = α ∼
= Z2 .
By computer, Inv(β) = B( F , A ) has dimension 1, dim Inv(α ) = 2, dim Inv(α β) = 3, and Inv(α β) ∩
Inv(α ) = Inv(β). Then (4.3) gives
3 33 − 3 32 − 3 34 − 33 − 32 + 3
+ + + = 10
3 3·2 3·2 3·4
nilpotent loops of order 9.
9.3. n = 12
For the first time we have to worry about separability, and hence we have to calculate the sub-
sets W ( F , A ).
Case A = Z2 , F = Z6 . Let Aut( F ) = α ∼ = Z2 . The subset W ( F , A ) is in fact a subspace: Let x ∈ F
be the unique involution and y ∈ F an element of order 3. If θ ∈ W y ( F , A ) then Z (Q( F , A , θ)) =
Q( F , A , θ) by Lemma 3.6. Thus W y ( F , A ) ⊆ W x ( F , A ) = W ( F , A ).
Computer calculation yields dim W ( F , A ) = 7, dim B( F , A ) = 4, dim Inv(α ) = 15, and
dim( W ( F , A ) ∩ Inv(α )) = 6. Thus there are
225 − 27 − (215 − 26 ) 215 − 26

+ = 1,049,594
24 · 2 24
loops Q with | Q | = 12 and Q / Z ( Q ) = Z6 . Among the 27 /24 = 8 loops constructed from the large
center cocycles, 6 are nonisomorphic.
Case A = Z2 , F = L 6,2 . By computer, Aut( F ) = 1, dim B( F , A ) = 5, dim W ( F , A ) = 7. Thus there are
225 − 27
= 1,048,572
25
loops Q with | Q | = 12 and Q / Z ( Q ) = F . The 27 /25 = 4 loops corresponding to cocycles in W ( F , A )

are pairwise nonisomorphic.
Case A = Z2 , F = L 6,3 . Then computer gives Aut( F ) = α ∼
= Z2 , dim B( F , A ) = 5, dim W ( F , A ) = 7,
dim Inv(α ) = 16, and W ( F , A ) Inv(α ). Thus there are
225 − 216 216 − 27

+ = 525,308
25 ·2 25
nilpotent loops Q with | Q | = 12 and Q / Z ( Q ) = F . The 27 /25 = 4 loops corresponding to large center
cocycles are pairwise nonisomorphic.
Among the 6 + 4 + 4 loops with | Z ( Q )| > 2 found so far, 11 are nonisomorphic.
Case A = Z3 , | F | = 4. If Z ( Q ) > A then [ Q : Z ( Q )] 2, so all cocycles in C( F , A ) are separable by
Lemma 3.7. The details are in Table 2.
If a nilpotent loop of order 12 is a central extension of both Z2 and of Z3 , it is an abelian group
by Lemma 3.6, and hence it is isomorphic to Z2 × Z2 × Z3 or to Z4 × Z3 . We have counted these two
loops twice and must take this into account.
9.4. n = 15
Either A = Z3 , F = Z5 or A = Z5 , F = Z3 . In both cases, all cocycles are separable by Lemma 3.7.

Most subspaces Inv( H ) can be determined by Theorem 6.5. The two cases overlap only in Z3 × Z5 .
9.5. n = 16
This is a more difficult case due to the 139 nilpotent loops F 1 , . . . , F 139 of order 8.
Cases A = Z2 , F = F i . We calculate the subsets W i = W ( F i , A ), and treat separable cocycles out-
side W i as usual. (In one of the cases, the automorphism group Aut( F , A ) = Aut( F ) is the simple
group of order 168, the largest automorphism group we had to deal with in the entire search.) We
filter the large center loops up to isomorphism.
We now need to filter the union of the 139 sets of large center loops up to isomorphism. This can
be done efficiently as follows: Let Q = Q( F , A , θ) where θ ∈ W i . For every central involution x of Q ,
calculate Q /x and determine its isomorphism type. If Q /x is isomorphic to some F j with j < i,
we have already seen Q and can discard it.
9.6. n = 18
See Table 2.
9.7. n = 20
This is the computationally most difficult case, due to the 1,044 nilpotent loops of order 10. See
Section 10 for more. The efficient filtering of large center loops is crucial here. On the other hand,
1,008 out of the 1,044 nilpotent loops of order 10 have trivial automorphism groups.
9.8. n = 21
This case is analogous to n = 15.
10. Related ideas and concluding remarks
For an introduction to loop theory see Bruck [1] or Pflugfelder [14].

The study of (central) extensions of groups by means of cocycles goes back to Schreier [16]. The
abstract cohomology theory for groups was initiated by Eilenberg and MacLane in [2–4], and it has
grown into a vast subject.
Eilenberg and MacLane were also the first to investigate cohomology of loops. In [5], they imposed
conditions on loop cocycles that mimic those of group cocycles, and calculated some cohomology
groups. A more natural theory (by many measures) of loop cohomology has been developed in [9]
by Johnson and Leedham-Green. As in this paper, their third cohomology group vanishes, since they
impose no conditions on the (normalized) loop 2-cocycles.
We are not aware of any work on the classification of nilpotent loops per se. In the recent pa-
per [10], McKay, Mynert and Myrvold enumerated all loops of order n 10 up to isomorphism. We
believe that all results in Section 4–Section 9 are new.
This being said, the central notion of separable cocycles must have surely been noticed before,
but since it is of limited utility in group theory (where much stronger structural results are available
to attack the isomorphism problem of central extensions), it has not been investigated in the more
general setting of loops. The experienced reader will recognize the mappings S and T of Section 6 as
the consecutive differentials in a free resolution of a cyclic group, cf. [6, Ch. 2].
The computational tools developed here are applicable to finitely based varieties of loops, and can
therefore be used to classify nilpotent loops of small orders in such varieties. One merely has to start
with the appropriate space of cocycles (determined by a system of linear equations, just as in the
group case). The first author intends to undertake this classification for loops of Bol–Moufang type,
cf. [15]. The classification of all Moufang loops of order n 64 and n = 81 can be found in [12]. The
classification of Bol loops has been started in [11]. The LOOPS [13] package contains libraries of small
loops in certain varieties, including Bol and Moufang loops.
All calculations in this paper have been carried out in the GAP [7] package LOOPS. We wrote two
mostly independent codes, and the calculations have been done at least twice. The enumeration of
nilpotent loops of order 20 took more than 90 percent of the total calculation time, about 2 days on a
single-processor Unix machine. Both codes and the multiplication tables of all nilpotent loops of order
n 10 can be downloaded at the second author’s web site http://www.math.du.edu/˜petr.
References
[1] R.H. Bruck, A Survey of Binary Systems, third printing, corrected, Ergeb. Math. Grenzgeb., vol. 20, Springer-Verlag, 1971.
[2] S. Eilenberg, S. MacLane, Group extensions and homology, Ann. of Math. (2) 43 (4) (Oct. 1942) 757–831.
[3] S. Eilenberg, S. MacLane, Cohomology theory in abstract groups I, Ann. of Math. (2) 48 (1) (Jan. 1947) 51–78.
[4] S. Eilenberg, S. MacLane, Cohomology theory in abstract groups II, Ann. of Math. (2) 48 (1) (Apr. 1947) 326–341.
[5] S. Eilenberg, S. MacLane, Algebraic cohomology groups and loops, Duke Math. J. 14 (1947) 435–463.
[6] L. Evens, The Cohomology of Groups, Oxford Math. Monogr., The Clarendon Press, 1991.
[7] The GAP Group, GAP — Groups, algorithms, and programming, version 4.4, Aachen, St Andrews, visit, http://www-
gap.dcs.st-and.ac.uk/~gap, 2006.
[8] C.J. Hillar, D.L. Rhea, Automorphisms of finite abelian groups, Amer. Math. Monthly 114 (10) (2007) 917–923.
[9] K.W. Johnson, C.R. Leedham-Green, Loop cohomology, Czechoslovak Math. J. 40 (115) (2) (1990) 182–194.
[10] B. McKay, A. Meynert, W. Myrvold, Small Latin squares, quasigroups, and loops, J. Combin. Des. 15 (2) (2007) 98–119.
[11] G.E. Moorhouse, Bol loops of small order, technical report, http://www.uwyo.edu/moorhouse/pub/bol/.
[12] G.P. Nagy, P. Vojtěchovský, The Moufang loops of order 64 and 81, J. Symbolic Comput. 42 (9) (2007) 871–883.
[13] G.P. Nagy, P. Vojtěchovský, LOOPS: Computing with quasigroups and loops in GAP, download at http://www.math.du.
edu/loops.
[14] H.O. Pflugfelder, Quasigroups and Loops: Introduction, Sigma Ser. Pure Math., vol. 7, Heldermann Verlag, Berlin, 1990.
[15] J.D. Phillips, P. Vojtěchovský, The varieties of loops of Bol–Moufang type, Algebra Universalis 54 (3) (2005) 259–271.
[16] O. Schreier, Über die Erweiterung von Gruppen I, Monatsh. Math. Phys. 34 (1) (1926) 165–180.
Journal of Algebra
Constructing arithmetic subgroups of unipotent groups

Willem A. de Graaf a,∗ , Andrea Pavan b,1
a
Dipartimento di Matematica, University of Trento, Via Sommarive 14, 38050 Povo (Trento), Italy
b
Dipartimento di Matematica Pura ed Applicata, University of Padova, Via Trieste 63, 35121 Padova, Italy
Article history: Let G be a unipotent algebraic subgroup of some GLm (C) defined
Received 3 July 2008 over Q. We describe an algorithm for finding a finite set of
Available online 8 April 2009 generators of the subgroup G (Z) = G ∩ GLm (Z). This is based on
Communicated by Gunter Malle
a new proof of the result (in more general form due to Borel and
Keywords:
Harish-Chandra) that such a finite generating set exists.
Algebraic groups © 2009 Elsevier Inc. All rights reserved.
Arithmetic groups
Lie algebras
Algorithms
1. Introduction
Let G be an algebraic subgroup of GLm (C) defined over Q, where m 1. Then for a subring R of C
we set
G ( R ) = G ∩ GLm ( R ).
The group G (Z) and any other subgroup Γ of G (Q) commensurable with it are called arithmetic
subgroups of G.
Arithmetic groups occur in many contexts. Examples are: the automorphism group of a finitely
generated nilpotent group (see [10], Chapter 6), the group of units of the ring of integers of a number
field, and the group of units of the group algebra ZG , where G is a finite group. A celebrated theorem
of Borel and Harish-Chandra [3] says that arithmetic groups are finitely generated. In this paper we
consider the problem of computing a finite set of generators of an arithmetic subgroup of a unipotent
group.
This problem was also treated in the paper [9] by Grunewald and Segal, where a general algorithm
for all arithmetic groups was outlined. However, their declared aim was to show that such a computa-
*
E-mail addresses: degraaf@science.unitn.it (W.A. de Graaf), pan@math.unipd.it (A. Pavan).
1
The author would like to thank the University of Trento for its hospitality during 2008, when the work on the paper was
carried out.
doi:10.1016/j.jalgebra.2009.03.027
W.A. de Graaf, A. Pavan / Journal of Algebra 322 (2009) 3950–3970 3951
tion is, at least in principle, feasible, and no attempt was made to make the algorithms efficient. And
unfortunately their algorithm appears to be unsuited for practical computation. Indeed, one step (i.e.,
Algorithm 4.2.2) of the algorithm for the unipotent case requires the enumeration of a set of size at
2
least (2(m!(1 + Δ)m + 1))m , where Δ is a non-negative integer. This is impractical, even for small m.
In this paper we describe a practical algorithm for finding a finite set of generators of G (Z), in
case G is unipotent, that is to say, all its elements are unipotent matrices. As a byproduct this yields
an independent proof of the Borel–Harish-Chandra theorem in this case. Also, we can show that the
groups G (Z) are T-groups of Hirsch length equal to dim G. In order to show that the algorithm can
be used to compute practical examples, we have implemented it in the language of the computer
algebra system GAP4 (cf. [8]). However, we do remark that our algorithm, in the worst case, has a
time complexity that is exponential in m.
We now sketch the main idea of the algorithm. Let V be the vector space on which G acts natu-
rally. Let
0 = V0 < V1 < · · · < Vn = V
be a flag of V with respect to the action of G (this means that for v ∈ V i and g ∈ G we have g v ≡
v mod V i −1 ). Then we can form the G-module V = V n−1 ⊕ VV . In informal terms the matrix of a
1
g ∈ G acting on V is formed from the matrix of its action on V by taking the block in the upper
left part of the matrix, and the block in the bottom right part of the matrix, and constructing the
block matrix consisting of these two blocks. Now let Q be the image of G in GL( V ); then we can
recursively compute generators of Q (Z). The recursion works because the Q -flag in V has smaller
length. Let π : G → Q be the projection. In Section 2 we describe π (G (Z)) (Proposition 2.7). We show
how to find generators of π (G (Z)), and their preimages in G (Z). Let N (Z) ⊂ G (Z) denote the kernel
of π . We find a finite set of generators of N (Z), and joined to the elements of G (Z) found earlier this
solves the problem.
These ideas are detailed in Section 2. In Section 3 we illustrate them with a simple example. The
constructions of Section 2 do not immediately yield an implementable algorithm. In order to obtain
that we need some technical preparation. In Section 4 we describe some results that allow us to work
with the Lie algebra of G rather than with G itself. Section 5 contains some material on T-groups. In
Section 6 we describe some algorithms for lattices that we need. Then in Section 7 we give a detailed
description of the main algorithm, and prove its correctness. The last section describes some practical
experiences with our implementation of this algorithm in GAP4.
2. The derived representation
The goal of this section is to introduce some notation, and to prove the results that underpin the
main algorithm.
Let V be a finite-dimensional vector space over Q, and L a full-dimensional lattice of V . We do
not prove the next lemma here; it will follow from Lemma 6.3.
Lemma 2.1. Let U and W be two subspaces of V with U ⊆ W . Then there exist subspaces U
and W
of V
such that
W
⊆ U
and equalities
U ⊕ U
= V = W ⊕ W
and
(L ∩ U ) + (L ∩ U
) = L = (L ∩ W ) + (L ∩ W
)
hold.
3952 W.A. de Graaf, A. Pavan / Journal of Algebra 322 (2009) 3950–3970
In the notations of the lemma above, we say that W

⊆ U
is a system of L-complements for
U ⊆ W.
Now let
0 = V0 < V1 < · · · < Vn = V
be a chain of subspaces of V with n 1. Then we consider the vector space
V
V = V n −1 ⊕ .
V1
We call it the derived vector space. Also, we have the full-dimensional lattice
L + V1
L = ( L ∩ V n −1 ) +
V1
of V , which we call the derived lattice, and the chain of subspaces
0 = V 0 < V 1 < · · · < V n−1 = V
of V where
V i +1
V i = V i ⊕
V1
which we call the derived chain. Note that its length is n − 1, which is strictly less than the length of
the chain of V .
ξ
Now let W n−1 ⊆ W 1 be a system of L-complements to V 1 ⊆ V n−1 , and let us denote by V −→ V 1
the projection of V onto V 1 along W 1 . By Hom( W n−1 , V 1 ) we denote the space of all linear maps
W n−1 → V 1 . Then we consider the map
: End( V ) −→ Hom( W n−1 , V 1 ), ϕ → ξ ◦ ϕ|W n−1 . (1)
We refer to it as the error map induced by the system W n−1 ⊆ W 1 . Further we define

Γ = γ ∈ Hom( W n−1 , V 1 ) γ ( L ∩ W n−1 ) ⊆ L ∩ V 1 . (2)
Since L ∩ W n−1 is full-dimensional in W n−1 , Γ is a lattice of Hom( W n−1 , V 1 ), and it is full-

dimensional since L ∩ V 1 is full-dimensional in V 1 . We refer to it as the lattice induced by the system
W n−1 ⊆ W 1 .
Now let G be a unipotent algebraic group defined over Q acting faithfully on V , and suppose that
0 = V0 < V1 < · · · < Vn = V
is a flag of V with respect to G, i.e., for all v ∈ V i we have g v ≡ v mod V i −1 for all g ∈ G. We consider
the subgroup

G L = g ∈ G (Q) g L = L
of G (Q). We want to find a finite set of generators of this group.

Since V 1 and V n−1 are G-stable subspaces of V , G acts on both V n−1 and VV , hence on their direct
1
sum, that is to say, on the derived vector space. We refer to the action of G on V as the derived
action. Further, we denote by N its kernel, which is of course a unipotent algebraic group over Q
acting faithfully on V , by Q its image, which is a unipotent algebraic group over Q acting faithfully
on V , and by π the projection of G onto Q . The following lemma is well known; it follows directly
from the commutativity of diagram (3) in Section 4.
Lemma 2.2. The projection π maps G (Q) surjectively onto Q (Q).
Also we set

N L = g ∈ N (Q) g L = L
and

Q L = q ∈ Q (Q) qL = L .
Of course the derived chain is a flag for V with respect to the action of Q . Since G acts faithfully
on V we can regard elements in G (Q) as automorphisms of V . The same consideration applies to
elements of N (Q). So we can apply the map to the elements of these groups.
Proposition 2.3. For every g ∈ G (Q) and h ∈ N (Q) we have
( g · h) = ( g ) + (h),
where is as in (1).
V
Proof. Let v ∈ W n−1 . Since h is an automorphism of V acting as the identity on V1
,
h( v ) − v ∈ V 1 .
Further, g is an automorphism of V acting as the identity on V 1 , hence

g h( v ) − v = h( v ) − v .
Since W n−1 ⊆ W 1 , we have ξ( v ) = 0. Hence applying ξ to both sides of the previous identity and
using linearity we obtain
ξ ◦ g ◦ h( v ) = ξ ◦ g ( v ) + ξ ◦ h( v )
hence the thesis. 2
Of course, N (Q) acts on G (Q) by multiplication on the right. Once we endow Hom( W n−1 , V 1 ) with
the obvious group structure given by addition, the previous proposition implies that the restriction of
to N (Q) is a group morphism. Hence N (Q) acts on Hom( W n−1 , V 1 ) by
Hom( W n−1 , V 1 ) × N (Q) → Hom( W n−1 , V 1 ), (x, h) → x + (h).
With these observations, we can restate the previous proposition saying that the restriction of to
G (Q) is a morphism of N (Q)-sets.
Now let us denote by W the image of N (Q) under . Then Proposition 2.3 gives us:
Hom( W n−1 , V 1 )
Corollary 2.4. There exists a unique map ˆ : Q (Q) −→ W
such that the diagram

G (Q) Hom( W n−1 , V 1 )
π
ˆ Hom( W n−1 , V 1 )
Q (Q) W
is commutative.
Proof. By Lemma 2.2, the map π : G (Q) → Q (Q) is surjective. Also, if g , g

∈ G (Q) are such that
π ( g ) = π ( g
) then g −1 g
∈ N (Q), hence due to Proposition 2.3 we obtain

( g
) = gg −1 g
= ( g ) + g −1 g
thus
(g) + W = (g
) + W .
These two facts show that the function
Hom( W n−1 , V 1 )
ˆ : Q (Q) −→ , q → ( g ) + W
W
where g is any element of G (Q) such that π ( g ) = q, is well defined and, of course, it makes the
diagram above commutative. If ˆ
is another such a function, then
ˆ ◦ π = ˆ
◦ π
hence, by surjectivity of π it follows that ˆ

= ˆ . 2
Also we set

G L = g ∈ G (Q) g L = L .
In other words, an element g ∈ G (Q) lies in G L if and only if π ( g ) lies in Q L . Of course, G L contains
both G L and N (Q).
Lemma 2.5. Let g ∈ G L . Then g ∈ G L if and only if ( g ) ∈ Γ .
Proof. If g ∈ G L then g ( L ∩ W n−1 ) ⊆ L, hence

ξ ◦ g ( L ∩ W n−1 ) ⊆ ξ( L ) = ξ ( L ∩ V 1 ) ⊕ ( L ∩ W 1 ) = L ∩ V 1
hence ( g ) ∈ Γ .
Now let g ∈ G L such that ( g ) ∈ Γ . Then g is an automorphism of V fixing both L + V 1 and
L ∩ V n−1 , and such that ξ ◦ g sends L ∩ W n−1 in L ∩ V 1 . In particular, since the preimage of L ∩ V 1
under ξ is W 1 + ( L ∩ V 1 ), we have that g ( L ∩ W n−1 ) ⊆ W 1 + ( L ∩ V 1 ). Further, since L ∩ W n−1 ⊆ L + V 1
and g fixes L + V 1 , we have that g ( L ∩ W n−1 ) ⊆ L + V 1 . Hence

g ( L ∩ W n −1 ) ⊆ ( L + V 1 ) ∩ W 1 + ( L ∩ V 1 ) .
Since L = ( L ∩ V 1 ) + ( L ∩ W 1 ), applying Dedekind’s modular law we obtain equality

L = (L + V 1 ) ∩ W 1 + (L ∩ V 1 )
hence g ( L ∩ W n−1 ) ⊆ L. Since L = ( L ∩ V n−1 ) + ( L ∩ W n−1 ) and g fixes L ∩ V n−1 , this shows that
g ( L ) ⊆ L. Now let l ∈ L. Since g fixes V 1 + L, there exist l1 ∈ V 1 and l2 ∈ L such that g (l1 ) + g (l2 ) = l.
Since g ( L ) ⊆ L, we have that g (l2 ) ∈ L, hence in particular that g (l1 ) = l − g (l2 ) ∈ L. Since g fixes V 1 ,
we also have that g (l1 ) ∈ V 1 , hence g (l1 ) ∈ L ∩ V 1 . Since V 1 ⊆ V n−1 , also g (l1 ) ∈ L ∩ V n−1 holds. Since
g fixes L ∩ V n−1 , we obtain that l1 ∈ L ∩ V n−1 , hence l1 + l2 ∈ L, hence L ⊆ g ( L ). So g ∈ G L . 2
Proposition 2.6. The map given by the chain
Hom( W n−1 , V 1 )
G L −→ Hom( W n−1 , V 1 ) −→
Γ
is a group morphism with kernel G L .
Proof. Let f , g ∈ G L . Then they are automorphisms of V acting as the identity on V 1 , fixing V n−1
and acting as the identity on V V . Further, they fix L ∩ V n−1 and L + V 1 . Now let l ∈ L. Of course,
n −1
g (l) − l ∈ V n−1 ∩ ( L + V 1 ). Since V 1 ⊆ V n−1 , applying Dedekind’s modular law we obtain V n−1 ∩
( L + V 1 ) = V 1 + ( L ∩ V n−1 ), which shows that

f g (l) − l − g (l) − l ∈ L ∩ V n−1
hence
f ◦ g (l) − f (l) − g (l) ∈ L .
Since ξ( L ) = L ∩ V 1 , we finally obtain
ξ ◦ f ◦ g (l) − ξ ◦ f (l) − ξ ◦ g (l) ∈ L ∩ V 1
which shows that the map is a group morphism. By Lemma 2.5, its kernel is G L . 2
Proposition 2.7. Let ˆ be as in Corollary 2.4. The map Ψ given by the chain
ˆ Hom( W n−1 , V 1 ) Hom( W n−1 , V 1 )

Q L −→ −→
W W +Γ
is a group morphism. Its kernel is equal to the image of G L under π .
Proof. Since π : G (Q) → Q (Q) is surjective and g ∈ G (Q) is in G L if and only if π ( g ) ∈ Q L , we

obtain a surjective map π : G L → Q L . By commutativity of the diagram in Corollary 2.4, we also
obtain the commutative diagram
Hom( W n−1 , V 1 )
G L Hom( W n−1 , V 1 ) Γ
Ψ Hom( W n−1 , V 1 )
Q L Γ +W
By Proposition 2.6, the top row is a group morphism, hence also the bottom row is. Again by Propo-
sition 2.6, G L is the kernel of the top row. Hence, since the diagram above is commutative, the image
of G L under π lies in the kernel of Ψ . Now let q be in the kernel of Ψ , and let g ∈ G L be a preimage
of q under π . By commutativity of the diagram above, we have
(g) ∈ Γ + W .
Now let w ∈ W such that ( g ) + w ∈ Γ . Since W is the image of N (Q) under , there exists h ∈ N (Q)
such that (h) = w. Hence by Proposition 2.3 we have
( g · h) = ( g ) + (h) ∈ Γ
thus by Lemma 2.5 we obtain that g · h ∈ G L . Of course, π ( g · h) = q. 2
3. An example
Let us consider
⎧⎛ ⎞ ⎫
⎪
⎪ 1 0 a b ⎪
⎪
⎨⎜ 1 2⎟ ⎬
0 1 c c
G= ⎜
⎝0 0 1
2 ⎟ ∈ GL4 (C) such that a, b, c ∈ C .
⎠
⎪
⎪ c ⎪
⎪
⎩ ⎭
0 0 0 1
It is easy to check that G is an algebraic subgroup of GL4 (C) defined over Q. Since it is contained
in the set of upper-unitriangular matrices of GL4 (C), it is unipotent. In this example it is rather
straightforward to find a set of generators of G (Z) directly. However, in this section we illustrate the
results of the previous section by showing how they help us finding a finite set of generators for G (Z).
G acts faithfully on Q4 by matrix–vector multiplication; also, Z4 is a full-dimensional lattice of Q4 .
Thus we can consider the subgroup G Z4 of G, and it is easily seen that
G (Z) = G Z4 .
The chain of subspaces
0 = V 0 < e 1 , e 2 = V 1 < e 1 , e 2 , e 3 = V 2 < V 3 = Q4
where e 1 , e 2 , e 3 and e 4 are the standard basis of Q4 , is a flag of V with respect to the action of G. The
derived vector space has basis given by (e 1 , 0), (e 2 , 0), (e 3 , 0), (0, e 3 + V 1 ) and (0, e 4 + V 1 ); through
this basis we can identify it with Q5 , and under this identification the derived lattice corresponds
to Z5 . The kernel of the derived action is
⎧⎛ ⎞ ⎫
⎪
⎪ 1 0 0 b ⎪
⎪
⎨ ⎜0 1 0 0⎟ ⎬
N= ⎝⎜ ⎟ ∈ GL (C) such that b ∈ C .
⎪
⎪ 0 0 1 0⎠ 4
⎪
⎪
⎩ ⎭
0 0 0 1
In particular,
⎧⎛ ⎞ ⎫
⎪
⎪ 1 0 0 b ⎪
⎪
⎨⎜ ⎟ ⎬
0 1 0 0
N Z4 = ⎜⎝
⎟ ∈ GL4 (C) such that b ∈ Z
⎠
⎪
⎪ 0 0 1 0 ⎪
⎪
⎩ ⎭
0 0 0 1
and it is straightforward to check that it is an infinite cyclic group with generator
⎛ ⎞
1 0 0 1
⎜0 1 0 0⎟
⎜
n=⎝ ⎟.
0 0 1 0⎠
0 0 0 1
The subgroup
⎧⎛ ⎞ ⎫
⎪
⎪ 1 0 a 0 0 ⎪
⎪
⎪
⎪ 0 0⎟
⎪
⎪
⎨⎜⎜0 1 c ⎟ ⎬
Q = ⎜⎜0 0 1 0 0⎟⎟ ∈ GL 5 (C) such that a , c ∈ C
⎪
⎪⎝ 0 ⎪
⎪
⎪
⎪ 0 0 1 c⎠ ⎪
⎪
⎩ ⎭
0 0 0 0 1
of GL5 (C) is the image of the derived action, the projection of G onto Q being
⎛ ⎞
⎛ ⎞ 1 0 a 0 0
1 0 a b ⎜
⎜0 1 c 1 2⎟ ⎜0 1 c 0 0⎟⎟
⎜ c ⎟ ⎜0 0 1
π :G → Q, 2 →
⎜ 0 0⎟⎟.
⎝0 0 1 c ⎠ ⎝0 0 0 1 c⎠
0 0 0 1
0 0 0 0 1
Now Q Z5 is a torsion free abelian group of rank 2 with basis given by
⎛ ⎞ ⎛ ⎞
1 0 1 0 0 1 0 0 0 0
⎜0 1 0 0 0⎟ ⎜0 1 1 0 0⎟
⎜ ⎟ ⎜ ⎟
q1 = ⎜
⎜0 0 1 0 0⎟⎟, q2 = ⎜
⎜0 0 1 0 0⎟⎟.
⎝0 0 0 1 0⎠ ⎝0 0 0 1 1⎠
0 0 0 0 1 0 0 0 0 1
A system of Z4 -complements for V 1 ⊆ V 2 is given by W 2 ⊆ W 1 where
W 2 = e 4 , W 1 = e 3 , e 4 .
Using the basis e 4 for W 2 and the basis e 1 , e 2 for V 1 we can identify Hom( W 2 , V 1 ) with M2×1 (Q);
under this identification, the induced lattice Γ corresponds to M2×1 (Z). Further, using the standard
basis of Q4 we can identify End(Q4 ) with M4×4 (Q). In this way, the error map is
⎛ ⎞
a1,1 a1,2 a1,3 a1,4
⎜ a2,1 a2,2 a2,3 a2,4 ⎟
: M4×4 (Q) → M2×1 (Q), ⎜ ⎟ → a1,4 .
⎝ a3,1 a3,2 a3,3 a3,4 ⎠ a2,4
a4,1 a4,2 a4,3 a4,4
1
The image of the rational points of N under is the subspace of M2×1 (Q) generated by . So
1 0
in the notation of Section 2 we have W = 0
and Γ = M2×1 (Z). Furthermore, the map Ψ from
M2×1 (Q)
Proposition 2.7 goes from the rational points of Q to W +Γ
.
Now we need two matrices g 1 and g 2 in G (Q) whose images under π are q1 and q2 , respectively.
Their existence is guaranteed by the surjectivity of π . For example, we can take
⎛ ⎞ ⎛ ⎞
1 0 1 0 1 0 0 0
⎜0 1 0 0⎟ ⎜0 1 1 ⎟ 1
⎜
g1 = ⎝ ⎟, g2 = ⎜ ⎟. 2
0 0 1 0⎠ ⎝ 0 0 1 1⎠
0 0 0 1 0 0 0 1
This shows in particular that g 1 and g 2 are in G Z5 . Also, using commutativity of the diagram in
Corollary 2.4, we have that

0
Ψ (q1 ) = 0 + W + Γ, Ψ (q2 ) = 1 + W + Γ.
2
Now the kernel of Ψ is generated by q1 and q22 . Therefore, by Proposition 2.7, π (G Z4 ) is generated
by q1 and q22 . Their preimages, g 1 and g 22 are only guaranteed to lie in G Z5 ; however, here we see
that they are already in G Z4 . Now the kernel of π restricted to G Z4 is N Z4 . So g 1 N Z4 and g 22 N Z4
G Z4
generate . We conclude that n, g 1 , g 22 generate G Z4 .
N Z4
Of course we could have made a different choice for the preimages of q1 and q2 . For example, we
could have taken
⎛ 1 ⎞
1 0 1 2
⎜ ⎟
g 1
= ⎜
⎝0
0 1 0 0⎟
0 1 0⎠
0 0 0 1
instead of g 1 . Then g 1
is in G Z5 but it is not in G Z4 . However, since q1 is in the kernel of Ψ , we get
ˆ (q1 ) ∈ W + Γ and by the commutativity of the diagram in Corollary 2.4, ( g1
) ∈ W + Γ . This can of
course also be checked directly as
1
( g1
) = 2
.
0
Now we note that
⎛ 1 ⎞
1 0 0 2
⎜0 1 0 0 ⎟ 1
n =⎜
⎟
⎝ 0 0 1 0 ⎠ ∈ N (Q) with (n ) =
2
.
0
0 0 0 1
The existence of such an n

is guaranteed by the fact that W is the image of the rational points of N
under . Since N is in the kernel of π we have π ( g 1
· (n
)−1 ) = π ( g 1
) = q1 . So we can work with
g 1
(n
)−1 as preimage of q1 . This choice of preimage works for us since by Proposition 2.3 we have
that

−1

1 1
0
g1 · (n ) = g 1 − (n ) = 2
− 2
= ∈ Γ.
0 0 0
Hence g 1
(n
)−1 ∈ G Z4 and we can apply the same considerations as above to prove that n, g 1
(n
)−1
and g 22 are a generating set for G Z4 . As a matter of coincidence, we note that
g 1
· (n
)−1 = g 1 .
Finally we note that n commutes with both g 1 and g 22 , hence the chain of subgroups

1 < n < n, g 1 < n, g 1 , g 22 = G Z4
is a central series for G Z4 with infinite cyclic factors.
4. The Lie algebra connection
Let G ⊂ GLm (C) be a unipotent algebraic group. As illustrated in Section 3, the results in Section 2
in principle yield an algorithm for finding a finite set of generators of G (Z). However, to make it work
efficiently in practice we rather work with the Lie algebra of G than with G itself. In this section we
describe the main results that we need for that.
First we review some standard facts on the Lie algebra of an algebraic group; for more details we
refer to [2,5,12,13].
As customary we denote the Lie algebras of the algebraic groups G , H , . . . by g, h, . . . . Let G be
a unipotent algebraic group defined over Q, acting on a vector space V . Then g also acts on V .
Furthermore, G is connected, and hence a subspace U ⊂ V is G-stable if and only if it is g-stable. If
this is the case then we get a G-action and a g-action on U , and those are compatible, in the sense
that the corresponding g-representation is the differential of the G-representation. Similarly we get
compatible G- and g-actions on quotients and direct sums of modules.
An important role in our algorithm is played by the exponential mapping. For a nilpotent x ∈
glm (C) we set

n −1
xi
exp(x) = ,
i!
i =0
and for a unipotent u ∈ GLm (C),
n −1
( u − 1) i
log(u ) = (−1)i −1 .
i
i =1
Since G is defined over Q we have that g ⊂ glm (C) has a basis such that all elements have coefficients
in Q. The Q-span of such a basis is denoted gQ . Then it is well known that the maps exp : gQ → G (Q)
and log : G (Q) → gQ are mutually inverse. In particular, when we work with the Lie algebra gQ we
keep control over the elements of G (Q) by these mappings.
Now let W be another finite-dimensional vector space over Q, H a unipotent algebraic subgroup
of GL( W ), and ϕ : G → H a morphism of algebraic groups. Then we have the diagram
log
G (Q) g
exp
ϕ dϕ (3)
log
H (Q) h
exp
and it turns out that it is commutative, i.e., exp(dϕ (x)) = ϕ (exp(x)) for all x ∈ g (cf. [5], Chapter V,
§4, Proposition 15).
Now we return to the setting of Section 2. A sequence of subspaces
0 = V0 < V1 < · · · < Vn = V
is a flag for the action of G if and only if it is a flag for the action of g. (The latter means that
g · V i ⊂ V i −1 for i > 0.) In particular, g acts on the derived vector space V , and the corresponding
representation of g is the differential of the representation of G on V . In particular this means that n,
which is the Lie algebra of N, is the kernel of dπ and q = dπ (g), where q is the Lie algebra of Q .
Proposition 4.1. n is central in g.
Proof. Let x ∈ g and y ∈ n. Then x is an endomorphism of V such that x. V ⊆ V n−1 and x. V 1 = 0,

and y is an endomorphism of V such that y . V ⊆ V 1 and y . V n−1 = 0. Now let v ∈ V . Then y . v ∈ V 1 ,
hence x. y . v = 0. Also, y . v ∈ V 1 , hence x. y . v = 0. Thus [x, y ]. v = 0. Since g acts faithfully on V ,
[x, y ] = 0. 2
Now let L be a full-dimensional lattice of V , W n−1 ⊆ W 1 a system of L-complements to V 1 ⊆

V n−1 , and let us denote by and by Γ the induced error map and the induced lattice, respectively.
Since g acts faithfully on V , we can regard its elements as endomorphisms of V . The same consider-
ation applies to n. So we can consider the restriction of the map to g and n.
Proposition 4.2. The restriction of the induced map to n is injective.
Proof. Let x ∈ n such that (x) = 0. Since x( V ) ⊆ V 1 , for every v ∈ V we have
x( v ) = ξ ◦ x( v ).
Since (x) = 0, ξ ◦ x( W n−1 ) = 0. Hence x( W n−1 ) = 0. Since x( V n−1 ) = 0, we finally have x( V ) = 0. 2
Further, we define

n L = x ∈ n (x) ∈ Γ .
By the previous proposition, it is a full-dimensional lattice of n. Since n acts faithfully on V and V

admits a flag with respect to this action, then n, regarded as a Lie subalgebra of gl( V ), consists of
nilpotent endomorphisms. Hence we can consider the diagram
exp
n GL( V )

Hom( W n−1 , V 1 )
Proposition 4.3. The diagram above is commutative.
Proof. Let us denote by id V the identity endomorphism of V . Since W n−1 ⊆ W 1 , (id V ) = 0. Now let
x ∈ n. Then x( V ) ⊆ V 1 and x( V n−1 ) = 0, hence x2 ( V ) = 0, and
exp x = id V + x.
Thus
(exp x) = (id V + x) = (id V ) + (x) = (x). 2

5. T-groups
The groups G (Z) that we are after are finitely-generated nilpotent and torsion free. Such groups
are called T-groups in the literature (cf. [10]). In this section we review some facts that we need on
T-groups.
It is known that any T-group admits a (proper normal) central series with infinite cyclic factors;
the other way round, every group admitting such a series is clearly a T-group. Now let G be a group,
and let g 1 , . . . , gn be an (ordered) set of elements of G. Then we can consider the chain of subgroups
G 1 G 2 · · · G n G n +1 = 1
of G where for every i = 1, . . . , n,
G i = g i , . . . , gn .
We call it the chain associated to g 1 , . . . , gn . Further, we say that g 1 , . . . , gn is a T-sequence for G if

the associated chain is a proper central series for G with infinite cyclic factors. It will be convenient
to extend this terminology saying that the empty set is a T-sequence for the trivial group. Every
T-group G has a T-sequence, and the length of a T-sequence is an invariant of the group, called the
Hirsch-length of G. Furthermore, if g 1 , . . . , gn is a T-sequence for G then every element g ∈ G can be
e e
written g = g 11 · · · gnn , where e i ∈ Z.
Now let A be an abelian group, and let a1 , . . . , an ∈ A. Then we consider

L = (e 1 , . . . , en ) ∈ Zn a11 · · · ann = 1
e e
which is of course a subgroup of Zn . We call it the relation lattice of a1 , . . . , an in A. For every

e = (e 1 , . . . , en ) ∈ L, e = 0, we can define its height as the minimum j = 1, . . . , n such that e j = 0, and
its leading coefficient as the integer e j . Now let e (1) , . . . , e (m) be a basis of L, where for j = 1, . . . , m,
( j) ( j)
e ( j ) = e 1 , . . . , en .
We say that the basis is in Hermite normal form if the matrix
⎛ ⎞
(1 ) (1 )
e1 · · · en
⎜ . .. ⎟
⎜ . ⎟
⎝ . . ⎠ ∈ Mm×n (Z)
(m)
e1 · · · en(m)
is. This means that there exists a sequence of integers
1 i 1 < · · · < im n
such that
( j)
ei =0
for all j = 1, . . . , m and all 1 i < i j , and that
(k) ( j)
0 ei < ei j
j
for every 1 k < j m. It is known that L admits a unique basis in Hermite normal form. We note
that the height of any non-zero element of L is one of the integers i 1 , . . . , im . Further, if its height is
( j)
i j for some j = 1, . . . , m, then its leading coefficient is a (non-zero) multiple of e i .
j
The next lemma is a somewhat stronger version of a result of Eick (cf. [6], Lemma 3.19).
Lemma 5.1. Let G be a T-group, A an abelian group, and let ϕ : G → A be a morphism of groups. Further, let
g 1 , . . . , gn be a T-sequence for G, and let e (1) , . . . , e (m) be the basis in Hermite normal form of the relation
lattice of ϕ ( g 1 ), . . . , ϕ ( gn ) in A. For j = 1, . . . , m set
( j) ( j)
e e
k j = g 11 · · · gnn .
Then k1 , . . . , km is a T-sequence for the kernel of ϕ .
Proof. Let us denote by
K 1 K 2 · · · K m K m +1 = 1
the chain of subgroups of G associated to k1 , . . . , km . We want to show that it is a (proper normal)

central series for ker ϕ with infinite cyclic factors. Since the basis e (1) , . . . , e (m) is in Hermite normal
form, we have the sequence
1 i 1 < i 2 < · · · < im n
defined as above. It is convenient to set i 0 = 0 and im+1 = n + 1. Also, let us denote by
G = G 1 > G 2 > · · · > G n > G n +1 = 1
the proper central sequence with infinite cyclic factors for G associated to g 1 , . . . , gn . Then it is
enough to prove that for every j = 1, . . . , m + 1 we have
ker ϕ ∩ G i j−1 +1 = · · · = ker ϕ ∩ G i j −1 = ker ϕ ∩ G i j = K j .
Indeed, suppose that the previous equalities hold. Then K 1 = ker ϕ ∩ G i 0 +1 = ker ϕ ∩ G 1 = ker ϕ ∩ G =
ker ϕ . Also, for every j = 1, . . . , m + 1, G i j P G, hence K j = ker ϕ ∩ G i j P ker ϕ ∩ G = ker ϕ . This shows
G
that the chain is a normal series for ker ϕ . Further, for every j = 1, . . . , m, the map ker ϕ → G i j +1
ker ϕ G
has kernel ker ϕ ∩ G i j +1 = K j +1 . Hence it factors through a group monomorphism K j +1
→ G i j +1
. The
Kj K j G i j +1 (ker ϕ ∩G i j )G i j +1 (ker ϕ G i j +1 )∩G i j Gi j
image of K j +1
under it is G i j +1
= G i j +1
= G i j +1
G i j +1
, and the image of k j K j +1
( j)
ei
j
is g i G i j +1 . This shows that the series is central and with infinite cyclic factors.
j
So we have to prove the previous equalities. It is clear that for every j = 1, . . . , m + 1,
ker ϕ ∩ G i j−1 +1 ⊇ · · · ⊇ ker ϕ ∩ G i j −1 ⊇ ker ϕ ∩ G i j ⊇ K j
and it remains to prove the reverse inclusions. We proceed by induction on j. Let us consider the
base case j = m + 1. Then K j is trivial, and all we have to show is that for every l = im + 1, . . . , n + 1,
ker ϕ ∩ G l is trivial, too. Again, we proceed by induction on l. In the base case l = n + 1, it is obviously
true. Now let l = im + 1, . . . , n and suppose that ker ϕ ∩ G l+1 is trivial. Let g ∈ ker ϕ ∩ G l . Then g = gle h
for some e ∈ Z and h ∈ G l+1 , and
ϕ ( gl )e + ϕ (h) = 0.
Since h ∈ gl+1 , . . . , gn , then ϕ (h) ∈ ϕ ( gl+1 ), . . . , ϕ ( gn ). Thus if e = 0, then there would exist an ele-
ment in the relation lattice with height l, which is impossible. Hence e = 0, hence g = h ∈ ker ϕ ∩ G l+1 ,
hence g = 1 by the inductive hypothesis. This concludes the case j = m + 1. Now let j = 1, . . . , m, and
suppose that
ker ϕ ∩ G i j +1 = · · · = ker ϕ ∩ G i j+1 −1 = ker ϕ ∩ G i j+1 = K j +1 .
In this case we have to show that for every l = i j −1 + 1, . . . , i j ,
ker ϕ ∩ G l = K j
and again we proceed by induction on l. Let us just consider the base case l = i j , the inductive step
being similar to the one in the case j = m + 1. Let g ∈ ker ϕ ∩ G i j . Then g = g ie h for some e ∈ Z
j
and some h ∈ G i j +1 . If e = 0 then g ∈ G i j +1 and we conclude by inductive hypothesis that g ∈ K j +1 .
Now let us suppose e = 0. Arguing as before, the relation lattice contains an element of height i j and
( j) f
leading coefficient e. Thus e i divides e. Let us denote by f the quotient. Then gG i j +1 = ki G i j +1 ,
j
−f
hence by inductive hypothesis gk j ∈ ker ϕ ∩ G i j +1 = K j +1 , hence finally g ∈ K j . 2
6. Some algorithms for lattices
In this section we describe some algorithms that solve several problems related to lattices. We
mainly work with matrices whose rows span lattices or subspaces in Qn . We say that a matrix is
integral if it has integer entries.
A basic algorithm that we use is the Smith normal form: given an m × n integral matrix A this
algorithm finds an m × n integral matrix S, and integral unimodular square matrices P and Q with:
1. S is in Smith normal form (this means that there is an r such that di = S (i , i ) is positive for
1 i r, S has no other non-zero entries, and di divides di +1 for 1 i < r),
2. S = P A Q .
For details on this algorithm we refer to [11], §8.3. One property that we note is the following (cf.
[11], Chapter 8, Corollary 3.4).
Lemma 6.1. Let q1 , . . . , qn denote the rows of Q −1 . They form a basis of Zn , and di is the smallest non-negative
integer such that di qi lies in the span of the rows of A for 1 i r.
We also need an algorithm that appears to be well known: in the computer algebra system Magma
[4] it is implemented under the name of “saturation.” However, we have not been able to find a
reference for it in the literature. For this reason we sketch a solution here. Let A be an m × n-matrix
with integer entries. Let V ⊂ Qn be the Q-space spanned by the rows of A. The problem is to find a
Z-basis for the lattice Zn ∩ V . Without loss of generality we assume that the rows of A are linearly
independent. The key observation is the following. Let B be an m × n integral matrix whose rows
span V . Then its rows span Zn ∩ V if and only if the Smith normal form of B has diagonal entries
that are all equal to 1. This follows from Lemma 6.1. This implies the correctness of the following
algorithm.
Algorithm 1 (Saturation).
Input: an m × n integral matrix A with linearly independent rows.
Output: an m × n integral matrix B whose rows span Zn ∩ V , where V ⊂ Qn is the Q-space spanned
by the rows of A.
(1) Let S, P , Q be the output of the Smith normal form algorithm with input A.
(2) Let S
be the matrix obtained from S by setting the diagonal entries equal to 1.
(3) Return B = P −1 S
Q −1 .
Algorithm 2 (Intersection of lattice and subspace).

Input: an n × n integral matrix A whose rows span the full-dimensional lattice L in Qn , and an m × n
matrix B whose rows span an m-dimensional Q-subspace W of Qn .
Output: an n × n integral matrix whose rows span L, and whose first m rows span the lattice W ∩ L.
n
(1) Let e 1 , . . . , en and b1 , . . . , bm denote the rows of A and B, respectively. Write b i = j =1 βi j e j , and
let B
= (βi j ); if necessary multiply the rows of B
by integers in order to get integral entries.
(2) Let C be the output of Algorithm 1 with input B
.
(3) Let S, P , Q be the output of the Smith normal form algorithm with input C .
(4) Return Q −1 A.
Lemma 6.2. Algorithm 2 is correct.
Proof. The idea is to use the given basis of L as a basis of Qn . Let ψ : Qn → Qn be the corresponding
isomorphism. So if v ∈ Qn then ψ( v ) is the vector that contains the coefficients of v with respect to
the basis of L. So after the first step the rows of B
form a basis of ψ( W ). Of course ψ( L ) = Zn ⊂ Qn .
So the rows of C form a basis of ψ( W ) ∩ ψ( L ) = ψ( W ∩ L ). Furthermore, the Smith normal form
S of C has diagonal entries equal to 1. Therefore the rows of Q −1 form a basis of Zn and the first m
rows form a Z-basis of ψ( W ∩ L ). Note that for v ∈ Qn we have ψ −1 ( v ) = ψ( v ) A. Therefore the rows
of Q −1 A form a basis of L, and the first m are a Z-basis of W ∩ L. 2
Algorithm 3 (L-complements).
Input: an n × n integral matrix A whose rows span a full-dimensional lattice L in V = Qn ; and bases
of subspaces V 1 ⊆ V n−1 ⊂ V .
Output: an n × n integral matrix C with the following properties:
– The rows of C span L.

– The first s rows of C span L ∩ V 1 (s = dim V 1 ).
– The first t rows of C span L ∩ V n−1 (t = dim V n−1 ).
(1) Execute Algorithm 2 with input A and a matrix whose rows span V n−1 . Let w 1 , . . . , w n denote
the rows of the output.
t
(2) Let v 1 , . . . , v s be the given basis of V 1 , and write v i = j =1 αi j w j . Let A = (αi j ).

(3) Execute Algorithm 2 with input the t × t-identity matrix, and A . Let B denote the output.
(4) Let C
be the product of B and the t × n matrix whose rows are w 1 , . . . , w t . Let C be the matrix
obtained from C
by appending w t +1 , . . . , w n .
Lemma 6.3. Algorithm 3 is correct. Let u 1 , . . . , un denote the rows of its output matrix B. Let W 1 , W n−1 ⊂ Qn
be the subspaces spanned by u s+1 , . . . , un and ut +1 , . . . , un , respectively. Then W 1 ⊂ W n−1 are a system of
L-complements to V 1 ⊂ V n−1 .
Proof. After the first step w 1 , . . . , w n span L and w 1 , . . . , w t span L ∩ V n−1 . Next we work in V n−1
using the basis w 1 , . . . , w t . We rewrite the basis elements of V 1 with respect to this basis. We note
that multiplying a v ∈ Qt with the matrix with rows w 1 , . . . , w t is the reverse transformation. So the
rows of C
span L ∩ V n−1 and the first s rows of C
span L ∩ V 1 . Therefore the output is correct.
The last statement follows directly from the definition of L-complements. 2
Remark 6.4. The output of Algorithm 3 has one more useful property. Let W 1 , W n−1 be as in Lem-
ma 6.3. The bases of the spaces W 1 and W n−1 that are produced by this algorithm are bases of
L ∩ W 1 and L ∩ W n−1 , respectively.
Algorithm 4 (Integral relations).

Input: An m × n-matrix A with rational coefficients.
Output: an m × n integral matrix whose rows are a basis of the lattice

m
m n
Λ = (e 1 , . . . , em ) ∈ Z e i ai ∈ Z ,
i =1
where a1 , . . . , am are the rows of A.
(1) Let M be the matrix obtained by appending the n × n-identity matrix at the bottom of A.
(2) Let v 1 , . . . , v m be a basis of the space { v ∈ Qm+n | v M = 0}. If necessary multiply each v i by an
integer to ensure that it has integral coefficients.
(3) Let B be the output of the saturation algorithm (Algorithm 1) applied to the matrix with the v i
as rows.
(4) Output the rows of B with the last n coefficients deleted.
Lemma 6.5. Algorithm 4 is correct.
Proof. Note that the matrix M has rank n; therefore in step (2) we find m linearly independent basis
vectors. Now set

Λ
= e = (e 1 , . . . , em+n ) ∈ Zm+n eM = 0 .
Then (e 1 , . . . , em+n ) → (e 1 , . . . , em ) is a bijection Λ → Λ

. Now after step (3) B is a basis of Λ
. We
conclude that the output is a basis of Λ. 2
7. The main algorithm
Now we return to our initial problem. Let G ⊂ GLm (C) be a unipotent algebraic group defined
over Q. Set V = Qm and let L be a full-dimensional lattice in V . The problem is to compute a finite
set of generators of the group G L = { g ∈ G (Q) | g ( L ) = L }.
We assume that G is defined as subset of GLm (C) by polynomial equations that have coefficients
in Q. Then as a first step we find the Lie algebra g of G as follows. First we compute a set S of
generators for the radical of the ideal generated by the polynomials that define G (cf. [1]). Then we
obtain the Lie algebra by differentiating the elements of S .
The second step will be to compute a flag 0 = V 0 < V 1 < · · · < V n = V of V for the action of g.
This is done by straightforward linear algebra: V 1 is the space killed by all elements of g; V 2 / V 1 is
the subspace of V / V 1 that is killed by all elements of g, and so on.
Now we have the input for our main algorithm which we now state. We use the notation of
Sections 2, 4.
Algorithm 5 (Main algorithm).

Input: a non-zero finite-dimensional vector space V over Q, a full-dimensional lattice L of V , a Lie
subalgebra g ⊂ gl( V ) that is the Lie algebra of a unipotent algebraic group G, and a flag
0 = V0 < V1 < · · · < Vn = V
of V with respect to the action of g.

Output: a T-sequence for G L .
(1) If n = 1 then return the empty set, else go to step (2).

(2) Compute the derived vector space V , the derived lattice L , the derived flag 0 = V 0 < V 1 <
· · · < V n−1 = V .
(3) Compute the kernel n and the image q of the derived action of g on V , together with the
projection dπ : g → q.
(4) Apply the algorithm recursively to the vector space V , the lattice L , the Lie algebra q and the
derived flag. Denote by q1 , . . . , qk the result.
(5) Compute a system W n−1 ⊆ W 1 of L-complements to V 1 ⊆ V n−1 (Algorithm 3), the induced lat-
tice Γ and the induced error map : End( V ) → Hom( W n−1 , V 1 ).
(6) Compute a basis x1 , . . . , xl of n L and set ni = exp(xi ) for 1 i l.
(7) For 1 i k compute a preimage x of log(q i ) under dπ and set gqi = exp(x).
(8) Compute the image W
of n under .
(9) Compute a basis W = { w 1 , . . . , w k } in Hermite normal form of the relation lattice of the ele-
ments ( gqi ) + Γ + W
in
Hom( W n−1 , V 1 )
Γ +W
for i = 1, . . . , k.
(10) For each w i in W do the following
(i ) (i )
(a) Write w i = (e 1 , . . . , ek ) and set
(i ) (i )
e e
g w i = gq11 · · · gqkk .
(b) Compute v w i ∈ W
and γ w i ∈ Γ such that v w i + γ w i = ( g w i ).
(c) Compute the preimage n w i of − v w i under : n → Hom( W n−1 , V 1 ).
(d) Compute g i = g w i · exp(n w i ).
(11) Return g 1 , . . . , gk , n1 , . . . , nl .
We start by commenting on the computability of various steps.
5. We note that Algorithm 3 produces bases of L ∩ W n−1 and L ∩ V 1 . We use these bases to represent
an element of Hom( W n−1 , V 1 ) as an s × t-matrix (where s = dim( V 1 ), t = dim( W n−1 )). Then a
Z-basis of Γ is the set of elementary s × t-matrices, which have one coefficient equal to 1, and
all other coefficients equal to 0. Computing (a) for an a ∈ End( V ) is standard linear algebra.
Indeed, for a ∈ End( V ) and w ∈ W n−1 write aw = v 1 + w 1 , where v 1 ∈ V 1 and w 1 ∈ W 1 . Then
(a) w = v 1 .
6. Here we first compute a basis of the space (n) ⊂ Hom( W n−1 , V 1 ). Using Algorithm 2 we find
a Z-basis of Γ ∩ (n). The inverse images of the basis elements under are then a basis of n L
(Proposition 4.2).
8. In step (6) we already obtained a basis of W
, and a basis γ1 , . . . , γst of Γ such that γ1 , . . . , γm
are a basis of W
∩ Γ . Let N = st− m and for γ ∈ Hom( W n−1 , V 1 ) set ψ(γ ) = (cm+1 , . . . , c st ),
where the c i are defined by γ = i c i γi . Then ψ : Hom( W n−1 , V 1 ) → Q N is a linear map and
γ ∈ Γ + W
if and only if ψ(γ ) ∈ ZN . Set u i = ( gqi ) for 1 i k. We want to compute a
Z-basis of the lattice

k

Λ = (e 1 , . . . , ek ) ∈ Zk ei ui ∈ W
+ Γ .
i =1

i e i ψ(u i ) ∈ Z . So we get a basis of Λ by applying
N
Now (e 1 , . . . , ek ) lies in Λ if and only if
Algorithm 4 with input the matrix with rows ψ(u i ). Then we compute the Hermite normal form
(cf. [11], §8.1) of the basis of Λ obtained.
The other steps are straightforward. We now prove the correctness of the algorithm.
Theorem 7.1 (Correctness of the main algorithm). Let V be a non-zero finite-dimensional vector space over Q,
L a full-dimensional lattice of V , and G a unipotent algebraic subgroup of GL( V ). Further, let g ⊂ gl( V ) be the
Lie algebra of G. Then Algorithm 5, with input V , L and g and a flag
0 = V0 < V1 < · · · < Vn = V
of V with respect to the action of g, returns a T-sequence for the subgroup G L of GL( V ).
Proof. As a first thing we notice that, since g is the Lie algebra of G, as seen in Section 4 the given
flag is also a flag of V with respect to the action of G. Hence the hypothesis of Section 2 is satisfied,
and we can consider all the constructions described there. As seen in Section 4, n ⊂ gl( V ) is the Lie
algebra of N and q ⊂ gl( V ) is the Lie algebra of Q . Now we proceed by induction on the length n of
the flag.
If n = 1, then every vector of V is G-fixed, hence G is the trivial subgroup of GL( V ) and G L is the
trivial subgroup of GL( V ), hence the empty set is a T-sequence for G L .
Now suppose that n 2. By the inductive hypothesis, q1 , . . . , qk is a T-sequence for the subgroup
Q L of GL( V ). By results in Section 4, the exponential map gives a bijection from n to N (Q) and,
since by Proposition 4.1 the Lie algebra n is abelian, it is also a group morphism. Further, using
Proposition 4.3 and Lemma 2.5, it is easily seen that the image of n L under the exponential map
is N L . Hence n1 , . . . , ns is a T-sequence for N L , regardless of their order. As seen in Section 4, we
have exp(dπ (x)) = π (exp(x)) for all x ∈ g. So for i = 1, . . . , k we get
π ( gqi ) = q i . (4)
In particular, gqi ∈ G L . By commutativity of the diagram in Corollary 2.4,
Ψ (qi ) = ( gqi ) + Γ + W
and, due to Proposition 4.3, W

is equal to the image W of N (Q) under . Hence W is a basis in
Hom( W n−1 , V 1 )
Hermite normal form for the relation lattice of Ψ (q i ), i = 1, . . . , k, in Γ +W . For 1 i k set
(i ) (i )
e e
h i = q11 · · · qkk ,
(i ) (i )
where w i = (e 1 , . . . , ek ) is as in the algorithm. By Lemma 5.1, the ordered set h1 , . . . , hk is a T-
sequence for the kernel of Ψ . Owing to (4) we have that g w i is an element of G L satisfying
π ( g w i ) = hi
for 1 i k. Again by commutativity of the diagram in Corollary 2.4, we have that ( g w i ) ∈ Γ + W =

Γ + W
(hence step (b) makes sense). Since exp(n w i ) ∈ N (Q), we have g i ∈ G L and
π ( gi ) = hi .
Further, using Propositions 2.3 and 4.3, we obtain that
( g i ) = ( g w i ) + (n w i ) = v w i − v w i + γ w i ∈ Γ.
Hence by Lemma 2.5 we have that g i ∈ G L . Thus by Proposition 2.7 the ordered set g 1 N L , . . . , gk N L is
a T-sequence for G
NL
L
. Using Proposition 4.1 we get that N is central in G, hence N L is central in G L .
Therefore we finally obtain that g 1 , . . . , gk , n1 , . . . , nl is a T-sequence for G L . 2
Corollary 7.2. Let the notation be as in Theorem 7.1. Let H denote the T-group generated by the output of
Algorithm 5. Then the Hirsch length of H is equal to the dimension of G. Moreover, the Lie algebra of the
radicable hull of H is isomorphic over Q to g.
Proof. We use the notation of the proof of Theorem 7.1. By induction, q1 , . . . , qk is a T-sequence
for Q L . This group has dimension equal to dim G − dim N = dim G − l. Therefore the T-sequence
output by the algorithm is of length equal to dim G.
Let h1 , . . . , hr denote the T-sequence output by the algorithm. From [10], Chapter 6, we get that
the Lie algebra of the radicable hull of H is isomorphic to the Lie algebra spanned by log(h i ). But the
latter one is g. 2
Remark 7.3. The algorithm can be slightly modified in order to compute even a finite presentation
of G L . To show how this can be done, we need to introduce some further notation.
Let G be a T-group and g 1 , . . . , gn a T-sequence for G . We say that a word w in g 1 , . . . , gn is
e e
normal if it is of the form g 11 · · · gnn for some e i ∈ Z. If this is the case, the depth of w is the
minimum i such that e i is non-zero. It is an easy induction to prove that every element g of G can
be written in a unique way as a normal word w in g 1 , . . . , gn . Also, g ∈ Gi = g i , . . . , gn if and only
if w has depth at least i. Since g 1 , . . . , gn is a T-sequence, for every 1 i < j n we have that
[ g i , g j ] ∈ G j +1 . Hence there exist a unique normal word w [ gi , g j ] (of depth at least j + 1) such that
[ g i , g j ] = w [ gi , g j ] .
It is well known that

g 1 , . . . , gn [ g i , g j ] = w [ gi , g j ] for 1 i < j n
is a finite presentation for G. We call it the standard presentation for G with respect to g 1 , . . . , gn .
Recall that the algorithm as it stands computes a T-sequence g 1 , . . . , gk , n1 , . . . , nl for G L . Now we
want to sketch how it can be modified in order to compute the standard presentation for G L with
respect to such a T-sequence. Of course it will be enough to show how to compute the normal words
of the forms w [ gi , g j ] , w [ gi ,n j ] and w [ni ,n j ] (for any suitable choice of the indexes i and j). The proof
of Theorem 7.1 shows that N L is central in G L . Hence it follows at once that
w [ gi ,n j ] = w [ni ,n j ] = 1.
So the only hard part is to compute the words of the form w [ gi , g j ] . To this end we can suppose
by inductive hypothesis that step (4) of the algorithm, along with a T-sequence q1 , . . . , qk for Q L ,
produces also the standard presentation

q1 , . . . , qk [qi , q j ] = w [qi ,q j ] for 1 i < j k
of Q L with respect to q1 , . . . , qk . Recall that, using the notations of the proof of Theorem 7.1,
h1 , . . . , hk is a T-sequence for the kernel of Ψ . Furthermore, the proof even shows that we can ef-
fectively write its elements as normal words in q1 , . . . , qk (of increasing depth). There are algorithms
known for computing a standard presentation of a subgroup of a T-group (cf. [11]; an implementa-
tion of the algorithms for this purpose is available in the GAP4 package “polycyclic,” [7]). So we can
compute the standard presentation

h1 , . . . , hk [h i , h j ] = w [hi ,h j ] for 1 i < j k
for the kernel of Ψ with respect to h1 , . . . , hk . Since every w [hi ,h j ] is a normal word in h1 , . . . , hk , we
can evaluate it in g 1 , . . . , gk (that it to say, substituting every h i with g i ). In this way we obtain a
normal word u [ gi , g j ] in g 1 , . . . , gn . Since π : G L → K sends g i in hi and has kernel N L , we obtain that
n = [ g i , g j ]u − 1
[g ,g ] ∈ N L .
i j
Since n1 , . . . , ns is a T-sequence for N L , we can write n as a normal word v in n1 , . . . , ns . This can be

done effectively, being equivalent to write an element of a lattice in terms of a (ordered) basis. Since
u [ gi , g j ] v is a normal word in g 1 , . . . , gk , n1 , . . . , nl , it now follows easily that
w [ gi , g j ] = u [ gi , g j ] v .
8. Practical performance
It is rather straightforward to see that the complexity of Algorithm 5 is exponential in the length
of the flag of V . Indeed, if the flag has maximal length, or in other words, dim V i = i, then dim V =
2n − 2. So in the worst case the dimension of the ambient vector space is roughly doubled in each
step of the recursion.
It turns out that the dimension of the spaces Hom( W n−1 , V 1 ) increases even faster (in the worst
case by about a factor of 4 each step of the recursion). For this reason we avoid working with the
entire space Hom( W n−1 , V 1 ). Instead we consider the associative algebra with one A ⊂ End( V ) gen-
erated by the elements of the Lie algebra g. Let U = ( A ) ⊂ Hom( W n−1 , V 1 ). Then (G ) ⊂ U , so we
can work with the space U instead of Hom( W n−1 , V 1 ). In fact we choose to work with a potentially
somewhat bigger space, namely the subspace of Hom( W n−1 , V 1 ) consisting of the matrices that have
non-zero entries only in those positions for which there are elements in U that have non-zero en-
tries in those positions. This space has the advantage that its intersection with the lattice Γ is easily
computed. Furthermore, in practice it is only slightly bigger than U .
We have implemented the algorithm in the language of GAP4 [8]. We use three series of Lie
algebras in gln (Q) to generate test inputs to the algorithm. The terms of the first series are denoted gn ,
which is spanned by
x i = e 1 , i +1 for 1 i n − 1,
n −1

xn = e j , j +1
j =2
(here e i , j is the n × n-matrix with a 1 on position (i , j ) and zeros elsewhere). The only non-zero
commutators are [xi , xn ] = xi +1 for 1 i n − 2. So gn is of dimension n and of nilpotency class
n − 1.
The second series of Lie algebras is denoted hn , which is spanned by

n −1
y1 = ie i ,i +1 ,
i =1
n−k

yk = e i ,i +k for 2 k n − 1.
i =1
Here the only non-zero commutators are [ y 1 , yk ] = −kyk+1 for 2 k n − 2. So hn is of dimension

n − 1 and of nilpotency class n − 2. In fact, as abstract Lie algebras, hn ∼
= gn−1 . We note that both Lie
algebras have a flag of maximal length.
Table 1
Time (in seconds) for the main algorithm with input gn , hn and un .
n time gn time hn time un

6 0 .7 0 .4 1.8
7 3 3 11
8 24 16 95
9 204 133 963
The terms of the third series, denoted un , are spanned by all matrices e i , j for 1 i < j n. In
other words, un is the full upper triangular Lie algebra. Also this Lie algebra has a flag of maximal
length.
In Table 1 we list the running times2 of the algorithm with input the Lie algebras gn , hn and un ,
for n = 6, 7, 8, 9. In all cases for the lattice L we have taken Zn ⊂ Qn .
On Table 1 we make the following comments:
• We see that the algorithm is efficient enough to tackle nontrivial examples.

• However, the running times do confirm the analysis above that the complexity of the algorithm
is exponential in the length of the flag.
∼ gn−1 , but
• Also we see that the length of the flag has a great bearing on the running time, as hn =
the algorithm needs markedly longer for hn than for gn−1 .
• Of course, for groups of higher dimension, but equal flag length, the algorithm also has to work
harder, as is shown by comparing gn and hn on the one hand, and un on the other hand.
• Finally we remark that it turned out that for n 10 our program was not able to complete the
computation within 1 GB of memory. So for higher flag lengths a more memory efficient imple-
mentation would be needed, for example using sparse matrices (that are not currently available
in GAP).
References
[1] Thomas Becker, Volker Weispfenning, Gröbner Bases, A Computational Approach to Commutative Algebra, Grad. Texts in
Math., vol. 141, Springer-Verlag, New York, 1993 (in cooperation with Heinz Kredel).
[2] Armand Borel, Linear Algebraic Groups, second ed., Grad. Texts in Math., vol. 126, Springer-Verlag, New York, 1991.
[3] Armand Borel, Harish-Chandra, Arithmetic subgroups of algebraic groups, Ann. of Math. (2) 75 (1962) 485–535.
[4] Wieb Bosma, John Cannon, Catherine Playoust, The Magma algebra system, I. The user language, in: Computational Algebra
and Number Theory, London, 1993, J. Symbolic Comput. 24 (3–4) (1997) 235–265.
[5] Claude Chevalley, Théorie des groupes de Lie. Tome III. Théorèmes généraux sur les algèbres de Lie, Actualités Sci. Ind.,
vol. 1226, Hermann & Cie, Paris, 1955.
[6] Bettina Eick, Algorithms for polycyclic groups, Habilitation Thesis, Technische Universität Braunschweig, 2001.
[7] Bettina Eick, Werner Nickel, Polycyclic, a GAP package, http://www.gap-system.org/Packages/polycyclic.html, 2004
(accepted).
[8] The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.4.10, http://www.gap-system.org, 2007.
[9] Fritz Grunewald, Daniel Segal, Some general algorithms, I. Arithmetic groups, Ann. of Math. (2) 112 (3) (1980) 531–583.
[10] Daniel Segal, Polycyclic Groups, Cambridge Tracts in Math., vol. 82, Cambridge University Press, Cambridge, 1983.
[11] Charles C. Sims, Computation with Finitely Presented Groups, Encyclopedia Math. Appl., vol. 48, Cambridge University Press,
Cambridge, 1994.
[12] Patrice Tauvel, Rupert W.T. Yu, Lie Algebras and Algebraic Groups, Springer Monogr. Math., Springer-Verlag, Berlin, 2005.
[13] William C. Waterhouse, Introduction to Affine Group Schemes, Grad. Texts in Math., vol. 66, Springer-Verlag, New York,
1979.
2
The computations were done on a 2 GHz processor with 1 GB of memory for GAP.
Journal of Algebra
Hyperoctahedral Chen calculus for effective Hamiltonians

Christian Brouder a , Frédéric Patras b,∗
a
Institut de Minéralogie et de Physique des Milieux Condensés, CNRS UMR7590, Universités Paris 6 et 7, IPGP, 140 rue de Lourmel,
75015 Paris, France
b
Laboratoire J.-A. Dieudonné, CNRS UMR 6621, Université de Nice, Parc Valrose, 06108 Nice Cedex 02, France
Article history: We tackle the problem of unraveling the algebraic structure of

Received 11 December 2008 computations of effective Hamiltonians. This is an important
Available online 31 July 2009 subject in view of applications to chemistry, solid state physics
Communicated by Jean-Yves Thibon
or quantum field theory. We show, among other things, that
Keywords:
the correct framework for these computations is provided by
Noncommutative symmetric functions the hyperoctahedral group algebras. We define several structures
Descent algebras on these algebras and give various applications. For example,
Hyperoctahedral groups we show that the adiabatic evolution operator (in the time-
dependent interaction representation of an effective Hamiltonian)
can be written naturally as a Picard-type series and has a natural
exponential expansion.
Introduction
We start with a short overview of the classical theory of Chen calculus, that is, iterated in-
tegral computations. The subject is classical but is rarely presented from the suitable theoretical
perspective—that is, emphasizing the role of the convolution product on the direct sum of the sym-
metric groups group algebras. We give therefore a brief account of the theory that takes into account
this point of view—this will be useful later in the article. Then, we recall the construction of effective
Hamiltonians in the time-dependent interaction representation.
Section 3 is devoted to the investigation of the structure of the hyperoctahedral group algebras.
Although we are really interested in the applications of these objects to the study of effective Hamil-
tonians, and although the definitions we introduce are motivated by the behavior of the iterated
integrals showing up in this setting, we postpone the description of the way the two theories in-
teract to a later stage of the article. Roughly stated, we show that the descent algebra approach to
E-mail address: patras@unice.fr (F. Patras).
doi:10.1016/j.jalgebra.2009.07.017
4106 C. Brouder, F. Patras / Journal of Algebra 322 (2009) 4105–4120
Lie calculus, as emphasized in Reutenauer’s [23] can be lifted to the hyperoctahedral setting. This
extends previous works [14,1,5,2,19] on the subject and shows that these results (focusing largely on
Solomon’s algebras of hyperoctahedral groups and other wreath product group algebras) are naturally
connected to the study of physical systems through the properties of their Hamiltonians and of the
corresponding differential equations, very much as the classical theory of free Lie algebras relates nat-
urally to the study of differential equations and topological groups. Notice however that the statistics
we introduce here on hyperoctahedral groups seems to be new—and is different from the statis-
tics naturally associated to the noncommutative representation theoretic approach to hyperoctahedral
groups, as it appears in these works.
Section 4 studies the effective adiabatic evolution operator and shows that it can be expanded
as a generalized Picard series by means of the statistics introduced on hyperoctahedral groups.1 As
a corollary, we derive in the last section an exponential expansion for the evolution operator. Such
expansions are particularly useful in view of numerical computations, since they usually lead to ap-
proximating series converging much faster than the ones obtained from the Picard series.
1. The algebra of iterated integrals
Let us recall the basis of Chen’s iterated integrals calculus, starting with a first order linear differ-
ential equation (with, say, operator or matrix coefficients):
A (t ) = H (t ) A (t ), A (0) = 1.
The solution can be expanded as the Picard series:
t t t1
A (t ) = 1 + H (x) dx + H (t 1 ) H (t 2 ) dt 1 dt 2 + · · · + H (t 1 ) · · · H (tn ) + · · ·
0 0 0 nt
where nt := {0 tn · · · t 1 t } and where the measure dt 1 . . . dtn is implicit. Notice, for further
use, that when we allow for a general initial condition A (x) = 1, with possibly x = −∞, the integra-
[x,t ]
tion simplex nt should be replaced by n := {x tn · · · t 1 t }.
Solving for A (t ) = exp(Ω(t )) (see [3,16]), and more generally any computation with A (t ), requires
the computation of products of iterated integrals of the form:

σ := H (t σ (1) ) · · · H (t σ (n) ), σ = σ (1), . . . , σ (n) ∈ S n ,
nt
where S n stands for the symmetric group of order n. Notice that we represent an element σ in S n by
the sequence (σ (1), . . . , σ (
n)).
In general, for any μ = n σ ∈ S n μσ · σ ∈ S := n Q[ S n ], the direct
sum of the group algebras of
the symmetric groups S n over the rationals, we will write μ for σ ∈ S n μσ · σ . This allows, for
n
example, to write A (t ) as I , where I := n (1, . . . , n) is the formal sum of the identity elements in
the symmetric group algebras.
The formula for the product of σ with β is a variant of Chen’s formula for the product of two
iterated integrals of functions or of differential forms (a proof of the formula will be given in Section 3
in a more general framework; the formula also holds for arbitrary integration bounds, that is when
nt is replaced by n[x,t ] ):
σ · β = σ ∗ β,
1
Picard series are often referred to as Dyson or Dyson–Chen series in the literature, especially in contemporary physics, but
we prefer to stick to the most classical terminology.
C. Brouder, F. Patras / Journal of Algebra 322 (2009) 4105–4120 4107
where σ ∗ β is the convolution product of the two permutations, that is, for σ ∈ S n , β ∈ S m :
2
σ ∗ β is the sum of the n+nm permutations γ ∈ S n+m with st(γ (1), . . . , γ (n)) = (σ (1), . . . , σ (n))
and st(γ (n + 1), . . . , γ (n + m)) = (β(1), . . . , β(m)). Here, st stands for the standardization map, the
action of which on sequences is obtained by replacing (i 1 , . . . , in ), i j ∈ N∗ by the (necessarily unique)
permutation σ ∈ S n , such that σ ( p ) < σ (q) for p < q if and only if i p i q . In words, each number i j
is replaced by the position of i j in the increasing ordering of i 1 , . . . , in . If we take the example of
(5, 8, 2), the position of 5, 8 and 2 in the ordering 2 < 5 < 8 is 2, 3 and 1. Thus, st(5, 8, 2) = (2, 3, 1).
For instance,
(2, 3, 1) ∗ (1) = (2, 3, 1, 4) + (2, 4, 1, 3) + (3, 4, 1, 2) + (3, 4, 2, 1),

(1, 2) ∗ (2, 1) = (1, 2, 4, 3) + (1, 3, 4, 2) + (1, 4, 3, 2) + (2, 3, 4, 1) + (2, 4, 3, 1) + (3, 4, 2, 1).
The convolution product relates to the shuffle product [23] as it appears in Chen’s work [7] and
in the parametrization of the product of iterated integrals of functions or differential forms. Indeed,
for σ , β as above:
−1
σ −1 ∗ β −1 = σ β[n]
where β[n] stands for the sequence (β(1) + n, . . . , β(m) + n). Recall, for completeness sake, that the
shuffle product A B of two words (or sequences) A = a A = aa2 · · · ak , B = bB = bb2 · · · bl is defined
recursively by
A B = a( A B ) + b( A B)
of ∗ follows immediately from the definition, the

we refer to [25] and [23] for details. Associativity
unit is 1 ∈ S0 = Q, and the graduation on S = n Q[ S n ] is compatible with ∗, so that:
Lemma 1.1. The convolution product provides S with the structure of a graded connected associative (but
noncommutative) unital algebra.
For completeness, recall that connected means simply that S0 = Q. In fact, S carries the richer
structure of a Hopf algebra, and as such is referred to as the Malvenuto–Reutenauer Hopf algebra [13].
From the point of view of the theory of noncommutative symmetric functions, the elements of S
should be understood as free quasisymmetric functions [8]. This definition of the convolution product
on S allows, for example, to express simply the coefficients of the continuous Baker–Campbell–
Hausdorff formula (compare with the original solution [16]):

Ω(t ) = log( I ) .
Here log( I ) identifies, in S, with the formal sum of Solomon’s (also called Eulerian or canonical)
idempotents [26]. We refer to [20,23,21,22,9] for an explanation and a Hopf algebraic approach to
these idempotents and, more generally, for a Hopf algebraic approach to Lie computations. We will
return later with more details to Solomon’s idempotent but mention only, for the time being, that one
of the main purposes of the present article is to extend these ideas to the more general framework
required by the study of effective Hamiltonians.
2
This is one possible definition of the convolution product, there are several equivalent ones that can be obtained using the
various natural set automorphisms of the symmetric groups (such as inversion or conjugacy by the element of maximal length).
They result into various (but essentially equivalent) associative algebra structures on the direct sum of the symmetric groups
group algebras, see e.g. [13].
2. Iterated integrals in time-dependent perturbation theory
The ultimate aim of quantum physics is the knowledge of the eigenvalues and eigenstates of the
Hamiltonian H describing a physical system. In most cases, the eigenvalue problem cannot be solved
exactly, but the eigenvalues and eigenstates of a simpler and closely related Hamiltonian H 0 are
known, at least numerically. Then, H can be rewritten H = H 0 + V , where V is referred to as the
perturbation term. For example, in molecular physics, H 0 is the Hamiltonian describing the interac-
tions of N e electrons with N n nuclei, whereas H describes the interaction of the N e electrons with
themselves and with the N n nuclei. In that case, the perturbation V describes the electron–electron
interaction, this simple approach paving the way to most of the numerical methods in the field.
In other terms, perturbation theory provides a systematic way to calculate an eigenstate of H from
an eigenstate of H 0 . In the time-dependent perturbation theory, we first define a time-dependent
Hamiltonian H (t ) = H 0 + e− |t | V . When is small, this means physically that the interaction is very
slowly switched on from t = −∞ where H (−∞) = H 0 to t = 0 where H (0) = H . The basic idea is
that, if is small enough, then an eigenstate of H 0 can be transformed into an eigenstate of H by the
time-dependent perturbation e− |t | V . For example, the ground state |Φ0 (the eigenstate associated to
the lowest eigenvalue E 0 of H 0 ) should hopefully be transformed into the ground state of H . We also
assume for the time being that the ground state is nondegenerate, that is that the eigenspace associ-
ated to the highest eigenvalue E 0 is one-dimensional. We write from now on |Φi for the eigenvectors
of H 0 and assume that the eigenvalues E i are ordered increasingly (E i E i +1 ).
To implement this picture of perturbation theory, the time-dependent Schrödinger equation
i ∂|Ψ S (t )/∂ t = H (t )|Ψ S (t ) should be solved. However, the solutions |Ψ S (t ) of this equation vary
like e−i E 0 t |Φ0 for large negative t (where |Φ0 satisfies H 0 |Φ0 = E 0 |Φ0 ) and have no limit for
t → −∞. To compensate for this time variation, one looks instead for |Ψ (t ) = ei H 0 t |Ψ S (t ) that sat-
isfies i ∂|Ψ (t )/∂ t = H I (t )|Ψ (t ), with H I = ei H 0 t V e−i H 0 t e− |t | . Now H I (−∞) = 0, and |Ψ (−∞)
makes sense. Using H I , we can start consistently from the ground state |Φ0 of H 0 and solve the
time-dependent Schrödinger equation with the boundary condition |Ψ (−∞) = |Φ0 . When no eigen-
value crossing takes place, |Φ0 should be transformed into the ground state |Ψ (0) of H .
At first sight, solving the time-dependent Schrödinger equation for |Ψ (t ) does not look simpler
than solving the time-independent Schrödinger equation with the Hamiltonian H . However, if V is
small enough, fairly accurate approximations of the true eigenstates can be obtained from the first
terms of the perturbative expansion of |Ψ (t ). In general, instead of calculating directly |Ψ (t ), it is
convenient to define the unitary operator U (t ) as the solution of i ∂ U (t )/∂ t = H I (t )U (t ), with the
boundary condition U (−∞) = 1. Thus, |Ψ (t ) = U (t )|Φ0 . The connection with iterated integrals
appears when solving iteratively the equation for U (t ):
t t t1
U (t ) = 1 + (−i ) dt 1 H I (t 1 ) + (−i )2 dt 1 dt 2 H I (t 1 ) H I (t 2 ) + · · · .
−∞ −∞ −∞
A straightforward calculation [11] gives us
∞
|Φin Φin | V |Φin−1 · · · Φi 2 | V |Φi 1 Φi 1 | V |Φ0
U (0)|Φ0 = |Φ0 + , (1)
( E 0 − E in + ni ) · · · ( E 0 − E i 2 + 2i )( E 0 − E i 1 + i )
n=1 i 1 ...in

where we use the completeness relation 1 = i |Φi Φi |. However, it immediately appears that this
expression has no limit for → 0 because the sum over all i p contains the terms i p = 0, for which
the denominator has a factor pi . In 1951, Gell-Mann and Low [10] conjectured that
U (0)|Φ0
|ΨGL = lim
→0 Φ0 |U (0)|Φ0
exists and is an eigenstate of H . The fact that the Gell-Mann and Low formula indeed removes all the
divergences of Eq. (1) was proved much later by Nenciu and Rasche [18].
The above scheme works nicely when the ground state of H 0 is nondegenerate. When it is degen-
erate, that is when the eigenspace E 0 associated to the lowest eigenvalue of H 0 has dimension > 1,
the problem is more subtle, see [17,6,15]. To understand the algebraic phenomena underlying time-
dependent perturbation theory for degenerate systems is actually the purpose of the present article.
Let us write P for the projection on this eigenspace. The natural extension of the Gell-Mann and
Low formula then reads as a definition of a “Gell-Mann and Low” operator acting on the degenerate
eigenspace E 0 :
−1
U GL := lim U (0) P P U (0) P , (2)
→0
or, U GL = lim →0 U GL ( ), U GL ( ) := U (0) P ( P U (0) P )−1 . It can be shown that the operator P U (0) P
is invertible within the image of P if no vector in the image of U (0) P is annihilated by P [4]. We will
assume that this property is satisfied by the systems we consider. The operator U GL was first proposed
by Morita in 1963 [17]. It shows up e.g. in the time-dependent interaction representation of the
effective Hamiltonian H eff := lim →0 P H U (0) P [ P U (0) P ]−1 classically used to solve the eigenvalue
problem.
This operator U GL is the one we will be interested in here, postponing to further work the analysis
of concrete applications to the study of degenerate systems. In other terms, we will investigate and
unravel the fine algebraic structure of the iterative expansions of U (t ) and U GL ( ).
3. Wreath product convolution algebras
Let us explain further our motivation. In the previous section, we observed that the study of ef-
fective Hamiltonians leads to the study of Picard-type expansions involving the operators H I (t ) and
P H I (t ) or, equivalently, A (t ) := −i (1 − P ) H I (t ) and B (t ) := i P H I (t ). Expanding these expressions
will lead to the study of iterated integrals involving the two operators A (t ) and B (t ) such as, say:
t A (t 2 ) B (t 3 ) A (t 1 ). The idea underlying the forthcoming algebraic constructions is to encode such
3
an expression by a signed permutation and to lift computations with iterated integrals to an abstract
algebraic setting: in the previous example, the signed permutation would be (2, 3̄, 1) (see below for
precise definitions).
In more abstract (but equivalent) terms, iterated integrals on two operators are conveniently en-
coded by elements of the hyperoctahedral groups, whereas iterated integrals on k operators would be
conveniently encoded by more general colored permutations or elements of the wreath product of S n
with the cyclic group of order k. Recall the definition of the hyperoctahedral group B n of order n. The
hyperoctahedral group is the group defined either as the wreath product of the symmetric group of
order n with the cyclic group of order 2, or, in a more concrete way, as the group of “signed per-
mutations” the elements of which are written as sequences of integers i ∈ N∗ and of integers with
an upper bar ī , i ∈ N∗ , so that, when the bars are erased, one recovers the expression of a permuta-
tion. The composition rule is the usual one for permutations, together with the sign rule for bars: for
example, if σ̄ ∈ B 3 = (2, 3̄, 1) and β̄ = (3̄, 1, 2̄), then:
β̄ ◦ σ̄ (2) = β̄(3̄) = 2̄¯ = 2,

β̄ ◦ σ̄ (3) = β̄(1) = 3̄.

By analogy with S, we equip B := n B n with the structure of a graded connected (associative
but noncommutative) algebra with a unit. This algebra structure agrees with the ones introduced
in [1,19,2] (possibly up to an isomorphism: for example, the relationship between the product we
consider and the shifted shuffle product of [19] reflects the relationship between the convolution and
(shifted) shuffle product on S that we recalled in Section 1 of the present article).
The standardization st of a signed sequence w̄ (i.e. a sequence of integers and of integers marked
with an upper bar) is defined analogously to the classical standardization, except for the fact that
upper bars are left unchanged (or, equivalently, have to be reintroduced at their initial positions after
the standardization of the sequence w has been performed, where we write w for w̄ where the upper
bars have been erased). For example, st(2̄, 7, 1̄, 2) = (2̄, 4, 1̄, 3).
Definition 3.1. Let σ̄ , β̄ belong to B n , resp. B m . Their convolution product is defined by:

σ̄ ∗ β̄ := τ̄
τ̄
n+m
where τ̄ runs over the n
elements of B n+m with st(τ̄ (1), . . . , τ̄ (n)) = σ̄ , st(τ̄ (n + 1), . . . ,
τ̄ (n + m)) = β̄ .
For instance,
(2̄, 3, 1) ∗ (1̄) = (2̄, 3, 1, 4̄) + (2̄, 4, 1, 3̄) + (3̄, 4, 1, 2̄) + (3̄, 4, 2, 1̄),
(1, 2̄) ∗ (2, 1̄) = (1, 2̄, 4, 3̄) + (1, 3̄, 4, 2̄) + (1, 4̄, 3, 2̄) + (2, 3̄, 4, 1̄) + (2, 4̄, 3, 1̄) + (3, 4̄, 2, 1̄).
Notice that this definition is dictated, for us, by iterated integrals computations, similarly to the clas-
sical one-Hamiltonian case dealt with in Section 1. Indeed, let A (t ), B (t ) be two time-dependent
operators. For σ̄ ∈ B n , let us write σ̄ for the iterated integrals obtained by the usual process,
with the extra prescription that upper indices (empty set or bar) in σ̄ indicate that the operator
used
at the corresponding level of the integral is A or B, so that e.g. σ̄ = (3̄, 1, 2̄)
is associated to:
t B (t 3 ) A (t 1 ) B (t 2 ). For an arbitrary γ̄ =
3
n σ̄ ∈ B n aσ̄ · σ̄ ∈ B, we write γ̄ for n σ̄ ∈ B n aσ̄ · σ̄ .
Proposition 3.2. The product of two iterated integrals σ̄ × β̄ is given by:
σ̄ × β̄ = σ̄ ∗ β̄.
This formula is a noncommutative variant of the classical Chen formulas for the product of iterated
integrals of differential forms [7]. It includes as a particular case the formula for the product of two
iterated integrals depending on a single time-dependent Hamiltonian given in Section 1 of the article.
We detail the proof for the sake of completeness, and since the formula is crucial for our purposes.
For a permutation σ̄ we denote by σ the same permutation without bars (e.g. if σ̄ = (2̄, 3, 1̄), then
σ = (2, 3, 1)) and we define X (t σ (i) ) = A (t σ (i) ) if σ̄ (i ) has no bar and X (t σ (i) ) = B (t σ (i) ) if σ̄ (i ) has
a bar. Therefore,
t
t n −1 t tn
+m−1
σ̄ × β̄ = dt 1 . . . dtn X (t σ (1) ) · · · X (t σ (n) ) dtn+1 . . . dtn+m X (tn+β(1) ) · · · X (tn+β(m) ).

0 0 0 0
By Fubini’s theorem, this can be rewritten as the integral of X (t σ (1) ) · · · X (tn+β(m) ) over the domain
n+m
nt × mt
. The idea is now to rewrite this domain as a sum of n
domains isomorphic to nt +m .
For instance, the product of the domain 0 tn · · · t 1 t with the domain 0 tn+1 t is the sum
of the n + 1 domains obtained by inserting tn+1 between 0 and t 1 , then between t 1 and t 2 , up to
between tn and t. More generally the product of nt by m t
is the sum of all the domains obtained
by “mixing” the two conditions 0 tn · · · t 1 t and 0 tn+m · · · tn+1 t, i.e. by ordering the
n + m variables t i so that these conditions are satisfied. If ρ (i ) is the position of variable t i in one
of these orderings (where the variables are ordered from the largest to the smallest), the conditions
imply that ρ (1) < · · · < ρ (n) and ρ (n + 1) < · · · < ρ (n + m). For example, if 0 t 2 t 1 t and
0 t 4 t 3 t, for the domain 0 t 4 t 2 t 1 t 3 t, t 3 is in the first place (i.e. largest), t 1 in

the second, t 2 in the third and t 4 in the fourth (smallest), and the permutation is ρ = (2, 3, 1, 4). In
general, we get

nt × m
t
= (t τ (1) , . . . , t τ (n+m) ) 0 tn+m · · · t 1 t ,
τ
where τ runs over the permutations in S n+m such that τ (1) < · · · < τ (n) and τ (n + 1) < · · · <
τ (n + m). Equivalently, τ runs over the permutations such that: st(τ (1), . . . , τ (n)) = (1, . . . , n) and
st(τ (n + 1), . . . , τ (n + m)) = (1, . . . , m). Now,

X (xσ (1) ) · · · X (xσ (n) ) X (xn+β(1) ) · · · X (xn+β(m) )
{(x1 =t τ (1) ,...,xn+m =t τ (n+m) )|0tn+m ···t 1 t }

= X (t τ (σ (1)) ) · · · X (t τ (σ (n)) ) X (t τ (n+β(1)) ) · · · X (t τ (n+β(m)) )
{0tn+m ···t 1 t }
so that finally, taking into account the bars of the permutations (that is the fact that X is A or B,
depending only on its position
in the sequence X (t τ (σ (1)) ) · · · X (t τ (σ (n)) ) X (t τ (n+β(1)) ) · · · X (t τ (n+β(m)) )),
we obtain σ̄ × β̄ = γ̄ γ̄ , with st(γ̄ (1) · · · γ̄ (n)) = (σ̄ (1), . . . , σ̄ (n)) and st(γ̄ (n + 1) · · · γ̄ (n +
m)) = (β̄(1), . . . , β̄(n)). This concludes the proof.
Remark 3.3. The same proof leads to a general noncommutative Chen formula. Let A 1 , . . . , A n and
B 1 , . . . , B m be noncommutative (e.g. matrix-valued) functions and let
t
t n −1
Aα = dt 1 . . . dtn A 1 (t α (1) ) · · · A n (t α (n) ),

0 0
t tm−1
Bβ = dt 1 . . . dtm B 1 (t β(1) ) · · · B m (t β(m) ),

0 0
t tn
+m−1
( A B )σ = dt 1 . . . dtn+m A 1 (t σ (1) ) · · · A n (t σ (n) ) B 1 (t σ (n+1) ) · · · B m (t σ (n+m) ),

0 0
then A α × B β = ( A B )α ∗β .
Proposition 3.4. The convolution product provides B with the structure of an associative (but noncommuta-
tive) algebra with a unit.
The proposition can be checked directly from the combinatorial definition of the convolution prod-
uct, we refer to the original proofs [1,19]. In our setting, it also follows from the associativity of the
product of iterated integrals. Notice that the general noncommutative Chen formula would relate
similarly to the algebraic structures on colored permutations and wreath product group algebras in-
troduced in [19].
4. Progressions and regressions
Modern noncommutative representation theory originates largely in the work of Solomon [27].
From this point of view, it is natural to partition hyperoctahedral groups into “descent classes”, simi-
larly to the partition of symmetric groups into descent classes (such a partition is often referred to as
a statistics on S n ).
Recall that a permutation σ ∈ S n has a descent in position i < n if and only if σ (i ) > σ (i + 1).
The descent set Desc(σ ) of σ is the setof all i < n such that σ has a descent in position i. The
S n = I ⊂[n−1] {σ , Desc(σ ) = I }. The descent algebra D is the linear
partition into descent classes read:
span of Solomon’s elements D nS := σ ∈ S n , Desc(σ )⊆ S σ , where S ⊆ [n − 1] and n ∈ N∗ (with the con-
vention D 0∅ = 1). It is provided with a free associative algebra structure by the convolution product ∗
on S ⊃ D , see [23, Chap. 9]. This algebra has various natural generating families as a free associative
algebra—for instance, the family of the D n∅ . It is therefore also isomorphic to the algebra of non-
commutative symmetric functions Sym, from which it follows that the structure theorems for these
functions can be carried back to the descent algebra—a point of view introduced and developed in [9]
and a subsequent series of articles starting with [12].
The corresponding descent statistics on B n is obtained by considering the total order n̄ < n − 1 <
· · · < 1̄ < 1 < · · · < n. A signed permutation σ̄ ∈ B n has a descent in position i < n if and only if σ̄ (i ) >
σ̄ (i + 1) [14, Def. 3.2]. Descent classes are defined accordingly. The problem with this noncommutative
representation theoretical statistics and with the corresponding algebraic structures is that they do
not fit the needs of iterated integral computations for effective Hamiltonians, as we shall see in the
forthcoming sections. Neither do the generalized descent algebras of [19, Sect. 5]. Notice that this is
not the case when symmetric groups are considered: the statistics of descent classes fits the needs
of noncommutative representation theory as well as the needs of Lie theoretical computations, as
emphasized in [23,9].
For this reason, we introduce another statistics on B n . It seems to be new, and has surprisingly
nice properties, in that it allows to generalize very naturally many algebraic properties of symmetric
groups descent classes.
We say that an element ᾱ = (α (1), . . . , α (n)) ∈ B n has a progression in position i if either:
1. |α (i )| < |α (i + 1)| and α (i + 1) ∈ N∗ ;

2. |α (i )| > |α (i + 1)| and α (i + 1) ∈ N̄∗ .
Else, we say that α has a regression in position i. Here, the operation | | is the operation of forget-
ting the bars, so that e.g. |6̄| = 6. The terminology is motivated by the quantum physical idea that
particles (associated to unmarked integers) propagate forward in time, whereas holes (associated to
marked integers in our framework) propagate backward. We refer the reader to Goldstone diagrams
expansions [11] of the Gell-Mann Low eigenstate |ΨGL for further insights into the physical motiva-
tions. Further details on these topics are contained in the following sections of this article, but we do
not develop here fully the physical implications of our approach, the focus being on their mathemat-
ical background.
We write Reg(α ) for the set of regressions of α . For example: Reg(4, 3̄, 5̄, 6, 2̄, 1) = {2, 5} since the
sequence (4, 3̄, 5̄, 6, 2̄, 1) has only two regressions, in positions2 and 5. For an arbitrary subset S of
[n − 1], we mimic now the descent statistics and write R nS := σ ∈ B n , Reg(σ )= S σ . It is also convenient

to introduce the elements T nS := σ ∈ B n , Reg(σ )⊆ S σ = U ⊆ S R nU .
Lemma 4.1. The elements R nS (resp. T nS ), S ⊆ [n − 1], form a family of linearly independent elements in the
group algebra Q[ B n ].
The first assertion follows from the very definition of the R nS , since it is easily checked that
{σ̄ ∈ B n , Reg(σ̄ ) = S } = ∅ for any S ⊆ [n − 1]. The second case follows from the Möbius inversion
formula:

R nS = (−1)| S |−|U | T nS ,
U ⊆S
where | S | stands for the number of elements in S.

Lemma 4.2. We have, for S ⊆ [n − 1], U ⊆ [m − 1]:
T nS ∗ T Um = T S ∪{n}∪(U +n) ,
where U + n = {u + n, u ∈ U }.

Indeed, by definition, for σ̄ ∈ B n , β̄ ∈ B m , with Reg(σ̄ ) = X ⊆ S, Reg(β̄) = Y ⊆ U , σ̄ ∗ β̄ = τ̄ τ̄ ,
where τ̄ runs over the elements of B n+m with st(τ̄ (1), . . . , τ̄ (n)) = σ̄ and st(τ̄ (n + 1), . . . , τ̄ (n +
m)) = β̄ . In particular, for any such τ̄ and by definition of the standardization process:
Reg(τ̄ ) ⊆ X ∪ {n} ∪ (Y + n).
Conversely, any τ̄ ∈ B n+m appears in the expansion of st(τ̄ (1), . . . , τ̄ (n)) ∗ st(τ̄ (n + 1), . . . , τ̄ (n + m))
by the very definition of ∗ and does not appear in the expansion of any other product σ̄ ∗ β̄ with
Reg(σ̄ ) = Reg(st(τ̄ (1), . . . , τ̄ (n))), Reg(β̄) = Reg(st(τ̄ (n + 1), . . . , τ̄ (n + m))), from which the lemma fol-
lows.
Corollary 4.3. For S, U as above, with the notation of the previous sections:

+m
T nS × T Um = T nS ∪{n}∪(U +n)
so that:
n n n +···+n
T ∅1 × · · · × T ∅k = T {n1 ,...,n +···+
k
nk−1 } .
1 1
Theorem 4.1. The linear span R of the elements T nS (equivalently, of the R nS ), n ∈ N, S ⊆ [n − 1], is closed
under the convolution product in B. This algebra, referred to from now on as the (hyperoctahedral) Regression
algebra, is isomorphic to the descent algebra D and to the algebra of noncommutative symmetric functions
Sym.
The second part of the theorem follows from the product rule in D , that reads:
n+m
D nS ∗ D m
U = D S ∪{n}∪(U +n) .
The proof for this last identity can be obtained similarly to the one in Lemma 4.2—see also [23].
Now we study in more detail the elements R n∅ that will play an important role in the following.
The lowest order R n∅ are
R 1∅ = (1) + (1̄),
R 2∅ = (1, 2) + (1̄, 2) + (2, 1̄) + (2̄, 1̄),
R 3∅ = (1, 2, 3) + (1̄, 2, 3) + (1, 3, 2̄) + (1̄, 3, 2̄) + (2, 1̄, 3) + (2̄, 1̄, 3) + (2, 3, 1̄) + (2̄, 3, 1̄)
+ (3, 1̄, 2) + (3̄, 1̄, 2) + (3, 2̄, 1̄) + (3̄, 2̄, 1̄).
We first observe that, if σ̄ ∈ B n is a term of R n∅ (and therefore has no regression), then the barred
integers of σ̄ are entirely determined by the permutation σ = (|σ̄ (1)|, . . . , |σ̄ (n)|), except for σ̄ (1).
Indeed, by definition of a progression, σ̄ (i + 1) ∈ N∗ if σ (i ) < σ (i + 1) and σ̄ (i + 1) ∈ N̄∗ if σ (i ) >
σ (i + 1). In other words, σ̄ (i + 1) ∈ N̄∗ iff σ has a descent at i. The integer σ̄ (1) is not determined
by σ and can be barred or not. Therefore, the number of terms of R n∅ is 2 · n!.
5. A Picard-type hyperoctahedral expansion
When it comes to expand ΨGL or U GL , as introduced in Section 2, the classical strategy introduced
by Goldstone (at least for nondegenerate states, that is for ΨGL [11]) consists in appealing to the
hole/particle duality of quantum physics. Goldstone’s theory was generalized to degenerate states by
Michels and Suttorp [15], but this part of the theory has remained largely in infancy and relies on
shaky mathematical grounds. The purpose of this section is to show that hyperoctahedral groups
provide a convenient way to derive and study such expansions, so as to build the foundations of
a group-theoretic approach to the perturbative computation of the ground states of physical systems,
with a particular view toward the degenerate case.
To sum up, we want to compute U GL ( ) = U (0) P ( P U (0) P )−1 . Let us write H i for −i H I and
A (t ) := (1 − P ) H i (t ), B (t ) := − P H i (t ) (notice the −1 sign in the definition of B). From the Picard
expansion, we have
0 0 t1
U (0) = 1 + H i (x) dx + H i (t 1 ) H i (t 2 ) dt 1 dt 2 + · · · + H i (t 1 ) · · · H i (tn ) + · · · .
−∞ −∞ −∞ n [−∞,0]
In this section (following a suggestion by the referee whose remarks helped us to simplify notably the
presentation of the following computations—we take the opportunity to thank him or her warmly),
we introduce a new encoding of iterated integrals involving A, B and H i .
The notation is best explained through an example:
0 t1 t2 t3 t4

A (t 3 ) H i (t 1 ) B (t 2 ) B (t 5 ) H i (t 4 ) dt 1 dt 2 dt 3 dt 4 dt 5 =: [2̂, 3̄, 1, 5̂, 4̄].
−∞ −∞ −∞ −∞ −∞
Concretely, in an arbitrary iterated integral involving A, B and H i , we look recursively at the po-
sitions i 1 , . . . , in of t 1 , . . . , tn in the integrand and decorate i j with a bar (resp. a hat, resp. no
decoration) if the corresponding operator is B (resp. H i , resp. A). In our example, t 1 (resp. t 2 . . .)
is in position 2 (resp. 3 . . .) in the product A (t 3 ) H i (t 1 ) B (t 2 ) B (t 5 ) H i (t 4 ) and appears as a param-
eter for H i (resp. B), so that we map 1 to 2̂, 2 to 3̄, and so on. The so-obtained sequence of
decorated integers is written inside brackets to avoid confusion with our previous notation. No-
tice that, if σ̄ ∈ B n , σ̄ = [σ̄ −1 (1), . . . , σ̄ −1 (n)] (where we use the definition of Section 3 for σ̄ ).
We will also use some self-explaining multilinear extensions of this notation, so that, for example:
[3, (1̄, 2̂) + (2, 1̂)] = [3, 1̄, 2̂] + [3, 2, 1̂], and so on. Notice in particular that the identity H i = A − B
translates formally into k̂ = k − k̄ so that, for example, [1̂, 2̄] = [1, 2̄] − [1̄, 2̄].
Theorem 5.1. The effective adiabatic evolution operator U GL has the hyperoctahedral Picard-type expansion:

U GL = lim P + (1 − P ) R n∅ P.
→0
n∈N∗
Indeed, let us expand [1̂, . . . , n̂] = [−∞,0] H i (t 1 ) · · · H i (tn ) with the A and B operators. In order to
n
do so, we introduce still another notation: for σ̄ ∈ B k , k < n, we set:

σ̄ ; n − k = X (t σ (1) ) · · · X (t σ (k) ) H i (tk+1 ) · · · H i (tn )
[−∞,t σ (k) ]
k[−∞,0] ×n−k
[−∞,0] [−∞,t σ (k) ]

where k × n−k is a shortcut for:

(t 1 , . . . , tn ) −∞ tk · · · t 1 0, −∞ tn · · · tk+1 t σ (k) ;
where σ stands, as usual, for the image of σ̄ in S k (obtained by forgetting the decorations), and
where X (t σ (i ) ) = A (t σ (i ) ) if σ (i ) = σ̄ (i ) and B (t σ (i ) ) else. For example,

(21̄3); 2 = A (t 2 ) B (t 1 ) A (t 3 ) H i (t 4 ) H i (t 5 ),
[−∞,t 3 ]
[−∞,
3
0]
×2

(23̄1); 2 = A (t 2 ) B (t 3 ) A (t 1 ) H i (t 4 ) H i (t 5 ).
−∞t 3 t 2 t 1 0, −∞t 5 t 4 t 1
The more general symbols X ; n − k are defined, as usual, by extending linearly these conventions to
arbitrary elements X ∈ Q[ B k ], k < n.
Lemma 5.1. The symbol σ̄ ; n − k can be expanded as:

σ̄ ; n − k = σ̄ −1 (1), . . . , σ̄ −1 (i ), σ̄ −1 (i + 1), . . . , σ̄ −1 (k) (k
+ 1, . . . , n̂) ,
where σ̄ ∈ B k and i := σ (k).
Indeed:

(t 1 , . . . , tn ) −∞ tk · · · t 1 0, −∞ tn · · · tk+1 t σ (k)

= (t 1 , . . . , tn ) −∞ tk · · · t σ (k) , −∞ tn · · · tk+1 t σ (k) , t σ (k) · · · t 1 0

= (t 1 , t 2 , . . . , tn ) −∞ t σ (k) · · · t 1 0,

(t σ (k)+1 , . . . , tn ) ∈ (u 1 , . . . , uk−σ (k) ) −∞ u σ (k) · · · u 1 t σ (k)

× ( v 1 , . . . , v n−k ) −∞ v n−k · · · v 1 t σ (k) ,
where × stands for the Cartesian product. Since the Cartesian product of simplices is reflected in the
shuffle product (see e.g. the proof of Proposition 3.2), the lemma follows.
We then have
[1̂, . . . , n̂] = [1, 2̂, . . . , n̂] − [1̄, 2̂, . . . , n̂] = −[1̄, 2̂, . . . , n̂] + [1, 2, 3̂, . . . , n̂] − [1, 2̄, 3̂, . . . , n̂]
= P [1̂, . . . , n̂] + [1, 2, 3̂, . . . , n̂] − [1, 2̄, 3̂, . . . , n̂]
= P [1̂, . . . , n̂] + [1, 2, 3̂, . . . , n̂] − [1, 2̄, 3̂, . . . , n̂]

+ − 2̄, (1) (3̂, . . . , n̂) + 2̄, (1) (3̂, . . . , n̂) .
Now, from the recursive definition of the shuffle product:

[1, 2̄, 3̂, . . . , n̂] + 2̄, (1) (3̂, . . . , n̂) = (1) (2̄, 3̂, . . . , n̂)
0
= A (t ) dt B (t 1 ) H i (t 2 ) · · · H i (tn−1 )
−∞ n[−∞,0]
−1

= −(1 − P ) R 1∅ P [1̂, . . . , n
− 1].
Here, we have used that P is a projection, so that (1 − P ) B (t ) = −(1 − P ) P H i (t ) = 0, to rewrite
0 0

A (t ) dt = (1 − P ) A (t ) + B (t ) dt = (1 − P ) R 1∅ .
−∞ −∞
We get finally, since (1, 2); n − 2 = [1, 2, 3̂, . . . , n̂] and (according to Lemma 5.1) (2, 1̄); n − 2 =
[2̄, (1) (3̂, . . . , n̂)]:

[1̂, . . . , n̂] = P [1̂, . . . , n̂] + (1 − P ) R 1∅ P [1̂, . . . , n
− 1] + (1 − P ) R 2∅ ; n − 2 ,
where the last identity follows, once again, from (1 − P ) B (t ) = 0 (we won’t comment any more on
this rewriting trick from now on).
The proof of the theorem can be obtained along these principles by recursion. Let us indeed as-
sume for a while that:

R k∅ ; n − k = R k∅ P [1̂, . . . , n
− k] + R k∅+1 ; n − k − 1 .
Then we get, by induction:

[1̂, . . . , n̂] = P [1̂, . . . , n̂] + (1 − P ) R 1∅ P [1̂, . . . , n
− 1]
2
+ (1 − P ) R ∅ P [1̂, . . . , n − 2] + · · · + (1 − P ) R n∅−1 P [1̂] + (1 − P ) R n∅ .

Since U (0) = n [1̂, . . . , n̂], this implies
∞

U (0) = P U (0) + (1 − P ) R n∅ P U (0) − 1 + 1 ,
n =1
or, since P 2 = P :
∞

n
U (0) P = P + (1 − P ) R ∅ P P U (0) P ,
n =1
and the theorem follows.

So, let us check that the formula for R k∅ ; n − k holds. Let us consider an arbitrary element σ̄ ∈ B k
with Reg(σ̄ ) = ∅. Then, with the notation of Lemma 5.1:

σ̄ ; n − k = σ̄ −1 (1), . . . , σ̄ −1 (i ), σ̄ −1 (i + 1), . . . , σ̄ −1 (k) (k + 1, k
+ 2, . . . , n̂)

− σ̄ −1 (1), . . . , σ̄ −1 (i ), σ̄ −1 (i + 1), . . . , σ̄ −1 (k) (k + 1, k
+ 2, . . . , n̂) .
Let us denote the first term by T 1 and the second by T 2 , so that σ̄ ; n − k = T 1 − T 2 . To calculate T 1 ,
let us use another (equivalent, the equivalence follows from the recursive definition of the shuffle
product and is left to the reader) recursive definition of the shuffle product, namely:

k
a1 a2 · · · ak bb2 · · · bl = a1 · · · ai b(ai +1 · · · ak b2 · · · bl ).
i =0
We get

k

T1 = σ̄ −1 (1), . . . , σ̄ −1 ( j ), k + 1, σ̄ −1 ( j + 1), . . . , σ̄ −1 (k) (k
+ 2, . . . , n̂)
j =i

k
= β̄ j ; n − k − 1
j =i
where β̄ j := (σ̄ −1 (1), . . . , σ̄ −1 ( j ), k + 1, σ̄ −1 ( j + 1), . . . , σ̄ −1 (k))−1 . We notice then that, for l < k,
β j (l) < β j (l + 1) ⇐⇒ σ (l) < σ (l + 1).
In particular, β̄ j has no regression in position less than k. Now, β j (k + 1) = β̄ j (k + 1) = j β j (k) = i,

which implies that β̄ j has no regression in position k. Finally, β̄ j has no regression.
Let us enumerate the number of (necessarily distinct) signed permutations β̄ j obtained in that
way. There are 2(k − 1)! elements σ̄ of R k∅ with a given value j of σ (k), where j runs from 1 to k. For
a given σ̄ with σ (k) = j, T 1 provides k − j + 1 elements of R k∅+1 . Thus, the expansion of T 1 provides
(k + 1)! different elements of R k∅+1 when σ̄ runs over R k∅ .
The term T 2 can be computed similarly. Using the same recursive formula for the shuffle product
as above, we get

k

T2 = σ̄ −1 (1), . . . , σ̄ −1 ( j ), k + 1, σ̄ −1 ( j + 1), . . . , σ̄ −1 (k) (k
+ 2, . . . , n̂)
j =i

= σ̄ −1 (1), . . . , σ̄ −1 (k) (k + 1, . . . , n̄)

− σ̄ −1 (1), . . . , σ̄ −1 ( j ), k + 1, σ̄ −1 ( j + 1), . . . , σ̄ −1 (k) (k
+ 2, . . . , n̂) .
j <i
In the first term in this expansion of T 2 , we recognize −σ̄ P [1̂, . . . , n

− k], so that these terms sum
up to − R k∅ P [1̂, . . . , n
− k] when σ̄ runs over elements in B k without regressions. In the second term
(the sum over j < i), the same reasoning as for T 1 shows that each term of the expansion is of the
form β̄ j ; n − k − 1, where β̄ j has no regression. Again, when σ̄ runs over elements in B k without
regressions, this provides (k + 1)! such elements in the expansion. These terms are pairwise distinct
and pairwise distinct from the elements showing up in the expansion of T 1 , from which we conclude

T 1 − T 2 = R k∅ P [1̂, . . . , n
− k] + R k∅+1 ; n − k + 1 .
The theorem follows.
6. A Magnus expansion for the evolution operator
In the classical case, that is when the solution X (t ) of a first order linear differential equation
is obtained from its Picard series expansion, the resulting approximating series converges relatively
slowly to the solution. This problem—let us call it the Magnus problem—is solved by reorganizing
the series expansion, often by looking for an exponential expansion X (t ) = exp Ω(t ) of the solution,
known as its Magnus expansion. Many numerical techniques have been developed along this idea that
go much beyond the formal-algebraic problem of deriving a formal expression for Ω(t ). However,
deriving such an expression is a decisive step towards the understanding of the behavior of Ω(t ).
This problem was solved, in the classical case, by Bialynicki-Birula, Mielnik and Plebański [3,16] who
obtained a formula for Ω(t ) in terms of Solomon’s elements D nS .
The purpose of the present section is to solve the Magnus problem for the analysis of solutions
in time-dependent perturbation theory. This provides the general term of time-dependent coupled-
cluster theory [24]. Our previous results pave the way toward the solution of the problem. Namely, as
it appears from Theorem 5.1, the natural object to look at is not so much the effective Hamiltonian
H = lim P 0 H U GL ( )
→0
or the effective adiabatic evolution operator U GL , than the Picard-type series

Pic := R n∅ .
n∈N
Notice that we define Pic as the sum of the R n∅ over all the integers (and not over N∗ ) in order to
have the identity operator 1 = H R 0 as the first term of the series. Of course, we have
∅

n n
U GL ( ) = P + (1 − P ) R ∅ P = P + (1 − P ) R ∅ P = P + (1 − P ) Pic P .
n∈N∗ n∈N
In other terms, we are interested in the expansion:
U GL ( ) = P + (1 − P ) exp(Ω ) P ,
where

n
n
Ω = log R∅ = log R∅ .
n∈N n∈N
n n n +···+n
Since R ∅1 ∗ · · · ∗ R ∅k = R {n11 ,...,n1 +···+
k
nk−1 } , a first expression of Ω R = log n∈N R n∅ follows:
(−1)| S | n
ΩR = R ,
|S| + 1 S
n∈N∗ S ⊆[n−1]
where one can recognize the hyperoctahedral analogue of Solomon’s Eulerian idempotent [23, Chap. 3,
Lem. 3.14]:
(−1)| S | n
soln = D .
|S| + 1 S
S ⊆[n−1]
The analogy is not merely formal and follows from the isomorphism of Theorem 4.1 together with the
existence of a logarithmic expansion of soln , which is actually best understood from a Hopf algebraic
point of view, see [20,23,21,22]:

n n
sol = log D∅ .
n∈N∗ n∈N
As a corollary of Theorem 4.1, we also get the expansion of Ω R in the canonical basis of
n∈N∗ Q[ B n ]:
Proposition 6.1. We have

(−1)| S | n − 1 −1 n
ΩR = TS
n |S|
n∈N∗ S ⊆[n−1]

(−1)| S | n − 1 −1
= σ̄ .
n |S|
n∈N∗ S ⊆[n−1] σ̄ ∈ B n , Reg(σ )= S
The proposition follows from the analogous expansion for soln [23], together with the algebra
isomorphism Theorem 4.1:

(−1)| S | n − 1 −1
soln = σ.
n |S|
n∈N∗ S ⊆[n−1] σ ∈ S n , Desc(σ )= S
Corollary 6.2. The hyperoctahedral Magnus expansion of the effective Hamiltonian H reads, when truncated
at the third order:

1
H = lim P H I P + (1 − P ) exp H (1) + H (1̄) + [ H (12) + H (1̄2) + H (21̄) + H (2̄1̄) − H (12̄)
→0 2
1
− H (1̄2̄) − H (21) − H (2̄1) ] + [ H (123) + H (1̄23) + H (132̄) + H (1̄32̄) + H (21̄3) + H (2̄1̄3)
3
+ H (231̄) + H (2̄31̄) + H (32̄1̄) + H (3̄2̄1̄) + H (31̄2) + H (3̄1̄2) + H (321) + H (3̄21) + H (23̄1)
+ H (2̄3̄1) + H (12̄3̄) + H (1̄2̄3̄) + H (13̄2) + H (1̄3̄2) + H (213̄) + H (2̄13̄) + H (312̄) + H (3̄12̄) ]
1
− [ H (132) + H (1̄32) + H (231) + H (2̄31) + H (213) + H (2̄13) + H (312) + H (3̄12)
6
+ H (13̄2̄) + H (1̄3̄2̄) + H (12̄3) + H (1̄2̄3) + H (21̄3̄) + H (2̄1̄3̄) + H (31̄2̄) + H (3̄1̄2̄) + H (23̄1̄)

+ H (2̄3̄1̄) + H (32̄1) + H (3̄2̄1) + H (123̄) + H (1̄23̄) + H (321̄) + H (3̄21̄) ] P .
Acknowledgments
We thank Kurusch Ebrahimi-Fard, Jean-Yves Thibon, and the anonymous referee. Their comments
and suggestions resulted into several improvements with respect to a preliminary version of this
article.
References
[1] M. Aguiar, N. Bergeron, K. Nyman, The peak algebra and the descent algebra of type B and D, Trans. Amer. Math. Soc. 356
(2004) 2781–2824.
[2] M. Aguiar, J.-C. Novelli, J.-Y. Thibon, Unital versions of the higher order peak algebras, arXiv:0810.4634v1.
[3] I. Bialynicki-Birula, B. Mielnik, J. Plebański, Explicit solution of the continuous Baker–Campbell–Hausdorff problem, Ann.
Physics 51 (1969) 187–200.
[4] C. Bloch, Sur la théorie des perturbations des états liés, Nuclear Phys. 6 (1958) 329–347.
[5] C. Bonnafé, C. Hohlweg, Generalized descent algebra and construction of irreducible characters of hyperoctahedral groups,
Ann. Inst. Fourier (Grenoble) 56 (2006) 131–181.
[6] L.N. Bulaevskii, Sov. Phys. JETP 24 (1967) 154.
[7] K.T. Chen, Iterated integrals of differential forms and loop space homology, Ann. of Math. 97 (1973) 217–246.
[8] G. Duchamp, F. Hivert, J.-Y. Thibon, Noncommutative symmetric functions VI: Free quasi-symmetric functions and related
algebras, Internat. J. Algebra Comput. 12 (2002) 671–717.
[9] I.M. Gelfand, D. Krob, A. Lascoux, B. Leclerc, V.S. Retakh, J.-Y. Thibon, Noncommutative symmetric functions, Adv. Math. 112
(1995) 218–348.
[10] M. Gell-Mann, F. Low, Bound states in quantum field theory, Phys. Rev. 84 (1951) 350–354.
[11] J. Goldstone, Derivation of the Brueckner many-body theory. Lectures on the many-body problem, Proc. R. Soc. London A
239 (1957) 267–279.
[12] D. Krob, B. Leclerc, J.-Y. Thibon, Noncommutative symmetric functions II: Transformations of alphabets, Internat. J. Algebra
Comput. 7 (1997) 181–264.
[13] C. Malvenuto, Ch. Reutenauer, Duality between quasi-symmetric functions and Solomon’s descent algebra, J. Algebra 177
(1995) 967–982.
[14] R. Mantaci, C. Reutenauer, A generalization of Solomon’s algebra for hyperoctahedral groups and other wreath products,
Comm. Algebra 23 (1995) 27–56.
[15] M.A.J. Michels, L.G. Suttorp, Diagrammatic analysis of adiabatic and time-independent perturbation theory for degenerate
energy levels, Phys. A 58 (1978) 385–388.
[16] B. Mielnik, J. Plebański, Combinatorial approach to Baker–Campbell–Hausdorff exponents, Ann. Inst. H. Poincaré Sect. A XII
(1970) 215–254.
[17] T. Morita, Progr. Theoret. Phys. 29 (1963) 351.
[18] G. Nenciu, G. Rasche, Adiabatic theorem and Gell-Mann–Low formula, Helv. Phys. Acta 62 (1989) 372–388.
[19] J.-C. Novelli, J.-Y. Thibon, Free quasi-symmetric functions and descent algebras for wreath products, and noncommutative
multi-symmetric functions, arXiv:0806.3682v1.
[20] F. Patras, Homothéties simpliciales, PhD thesis, Univ. Paris 7, January 1992.
[21] F. Patras, La décomposition en poids des algèbres de Hopf, Ann. Inst. Fourier (Grenoble) 43 (1993) 1067–1087.
[22] F. Patras, L’algèbre des descentes d’une bigèbre graduée, J. Algebra 170 (1994) 547–566.
[23] C. Reutenauer, Free Lie Algebras, Oxford University Press, Oxford, 1993.
[24] K. Schönhammer, O. Gunnarsson, Time-dependent approach to the calculation of spectra functions, Phys. Rev. B 18 (1978)
6606–6614.
[25] M.P. Schützenberger, Sur une propriété combinatoire des algèbres de Lie libres pouvant être utilisée dans un problème de
mathématiques appliquées, Séminaire Dubreil–Jacotin Pisot (Algèbre et théorie des nombres), Paris, Année 1958/59.
[26] L. Solomon, On the Poincaré–Birkhoff–Witt theorem, J. Combin. Theory 4 (1968) 363–375.
[27] L. Solomon, A Mackey formula in the group algebra of a finite Coxeter group, J. Algebra 41 (1976) 255–268.
Journal of Algebra
Polynomial relations among principal minors

of a 4 × 4-matrix
Shaowei Lin, Bernd Sturmfels ∗
Department of Mathematics, University of California, Berkeley, CA 94720, USA
Article history: The image of the principal minor map for n × n-matrices is
Received 19 December 2008 shown to be closed. In the 19th century, Nanson and Muir
Available online 11 August 2009 studied the implicitization problem of finding all relations among
principal minors when n = 4. We complete their partial results
Keywords:
by constructing explicit polynomials of degree 12 that scheme-
Principal minors theoretically define this affine variety and also its projective closure
Hyperdeterminant in P15 . The latter is the main component in the singular locus of
Polynomial ideal the 2 × 2 × 2 × 2-hyperdeterminant.
1. Introduction
Principal minors of square matrices appear in numerous applications. A basic question is the
Principal Minors Assignment Problem [3] which asks for necessary and sufficient conditions for a col-
lection of 2n numbers to arise as the principal minors of an n × n-matrix. When the matrix is
symmetric, this question was recently answered by Oeding [11] who extended work of Holtz and
Sturmfels [4] to show that the principal minors of a symmetric matrix are characterized by certain
hyperdeterminantal equations of degree four.
This question is harder for general matrices than it is for symmetric matrices. For example, con-
sider our Theorem 1 which says the image of the principal minor map is closed. The same statement
is trivially true for symmetric matrices, but the proof for non-symmetric matrices is quite subtle.
We denote the principal minors of a complex n × n-matrix A by A I where I ⊆ [n] = {1, 2, . . . , n}.
Here, A I is the minor of A whose rows and columns are indexed by I , including the 0 × 0-minor
A ∅ = 1. Together they form a vector A ∗ of length 2n . We are interested in an algebraic characterization
n 2 n
of all vectors in C2 which can be written in this form. The map φa : Cn → C2 , A → A ∗ is called the
affine principal minor map for n × n-matrices.
E-mail addresses: shaowei@math.berkeley.edu (S. Lin), bernd@math.berkeley.edu (B. Sturmfels).
doi:10.1016/j.jalgebra.2009.06.026
4122 S. Lin, B. Sturmfels / Journal of Algebra 322 (2009) 4121–4131
n
Theorem 1. The image of the affine principal minor map is closed in C2 .
This result says that Im φa is a complex algebraic variety. The dimension of this variety is n2 −
n + 1. This number is an upper bound because the principal minors remain unchanged under the
transformation A → D A D −1 for diagonal matrices D, and it is not hard to see that this upper bound
is attained [14]. What we are interested in here is the prime ideal of polynomials that vanish on the
irreducible variety Im φa . We determine this prime ideal in the first non-trivial case n = 4. Here, we
ignore the trivial relation A ∅ = 1.
Theorem 2. When n=4, the prime ideal of the 13-dimensional variety Im φa is minimally generated by 65
polynomials of degree 12 in the unknowns A I .
This theorem completes the line of research started by MacMahon, Muir and Nanson [8–10] in the
late 19th century. Our proof of Theorem 2 takes advantage of their classical results, and it will be
presented in Section 3.
Algebraic geometers would consider it more natural to study the projective version of our problem.
We define the projective principal minor map as
2 2 n
φ : Cn × Cn → C2 , ( A , B ) → det( A I B [n]\ I ) I ⊆[n] . (1)
Here we take two unknown n × n-matrices A and B to form an n × 2n-matrix ( A , B ), and for each
I ⊆ [n], we evaluate the n × n-minor with column indices I on A and column indices [n]\ I on B.
The image of φ is a closed affine cone in C2 , to be regarded as a projective variety in P2 −1 =
n n
P(C2 ⊗ · · · ⊗ C2 ).
What makes the projective version more interesting than the affine version is that Im φ is invariant
under the natural action of the group G = GL2 (C)n . This was observed by J.M. Landsberg (cf. [4,11]).
In Section 4 we shall prove
Theorem 3. When n = 4, the projective variety Im φ is cut out scheme-theoretically by 718 linearly inde-
pendent homogeneous polynomials of degree 12. This space of polynomials is the direct sum of 14 irreducible
G-modules.
The polynomials described in Theorems 2 and 3 are available online at
http://math.berkeley.edu/~shaowei/minors.html.
A geometric interpretation of our projective variety is given in Section 5. We shall see that Im φ is
the main component in the singular locus of the 2 × 2 × 2 × 2-hyperdeterminant. This is based on
work of Weyman and Zelevinsky [15], and it relates our current study to the computations in [5]. We
begin in Section 2 by rewriting principal minors in terms of cycle-sums. The resulting combinatorial
structures are key to our proof of Theorem 1.
After posting the first version of this paper on the arXiv, Eric Rains informed us that some of our
findings overlap with results in Section 4 of his article [1] with Alexei Borodin. The relationship to
their work, which we had been unaware of, will be discussed at the end of this paper, in Remark 14.
2. Cycle-sums and closure
In this section, we define cycle-sums and determine their relationship to the principal minors. We
then prove that a certain ring generated by monomials called cycles is integral over the ring generated
by the principal minors. This integrality result will be our main tool for proving the closure theorem.
Let X = (xi j ) be an n × n-matrix of indeterminates and C[ X ] the polynomial ring generated by
these indeterminates. Let P I ∈ C[ X ] denote the principal minor of X indexed by I ⊆ [n], including
S. Lin, B. Sturmfels / Journal of Algebra 322 (2009) 4121–4131 4123
P ∅ = 1. Together these minors form a vector P ∗ of length 2n . Thus A ∗ = P ∗ ( A ) if A is a complex

n × n-matrix. Now, given a permutation π ∈ Sn of [n], define the monomial c π = i =π (i ) xi π (i ) ∈ C[ X ]
where the product is taken over the support of π . We call c π a k-cycle if π is a cycle of length k. For
I ⊆ [n], | I | 2, define the cycle-sum C I = π ∈C I c π where C I is the set of all cycles with support I .
Also, let C ∅ = 1 and C i = xii , i ∈ [n]. Together they form a vector C ∗ of length 2n . The cycles, cycle-
sums and principal minors generate subrings C[c ∗ ], C[C ∗ ] and C[ P ∗ ] of C[ X ]. The next result shows
that C[C ∗ ] = C[ P ∗ ].
Proposition 4. The principal minors and cycle-sums satisfy the following relations: for any subset I ⊆ [n] of
cardinality d 1, we have

PI = (−1)k+d C I 1 · · · C I k , (2)
I = I 1
···
I k

CI = (−1)k+d (k − 1)! P I 1 · · · P I k , (3)
I = I 1
···
I k
where the sums are taken over all set partitions I 1

· · ·
I k of I .
Proof. The first equation is Leibniz’s formula for the determinant. The second equation is derived
from the first formula by applying the Möbius inversion formula [12, §3.7] for the lattice of all set
partitions of [n]. 2
n n n
Let ψ : C2 → C2 be the polynomial map given by (3). We say that u ∗ ∈ C2 is realizable as
principal minors if u ∗ = P ∗ ( A ) for some complex matrix A. Similarly, u ∗ is realizable as cycle-sums if
u ∗ = C ∗ ( A ) for some A.
n
Corollary 5. A vector u ∗ ∈ C2 is realizable as principal minors if and only if its image ψ(u ∗ ) is realizable as
cycle-sums.
A monomial in C[ X ] can be represented by a directed multigraph on n vertices as follows: label

the vertices 1, . . . , n and for each xkij appearing in the monomial, draw k directed edges from vertex i
to vertex j. Cycles c (i 1 ...ik ) correspond to cycle graphs i 1 → · · · → ik → i 1 which we write as (i 1 . . . ik ).
We are interested in studying when a product of cycles can be written as a product of smaller ones.
This is equivalent to decomposing a union of cycles into smaller cycles. For instance, the relation
c (123) c (132) = c (12) c (23) c (13) says that the union of these two 3-cycles can be decomposed into three
2-cycles.
Lemma 6. Let π1 , π2 , . . . , πm be m 2 distinct cycles of length k 3 with the same support. Then, the
product c π1 c π2 · · · c πm can be expressed as a product of strictly smaller cycles.
Proof. We may assume that all the cycles have support [k]. Note that it suffices to prove our
lemma for m = 2, 3. The following is our key claim: given an l-cycle c (1si 3 ...il ) , l k, s = 2, not equal
to c (1s(s+1)...k) , the product c (1...k) c (1si 3 ...il ) can be expressed as a product of cycles of length smaller
than k. Indeed, suppose that no such expression exists. Let the graphs of c (1...k) and c (1si 3 ...il ) be G1
and G2 respectively. Color the edges of G1 red and G2 blue. Then G = G1 ∪ G2 contains the cycle
C1 = (1s(s + 1) . . . k) where the first edge 1 → s is blue while every other edge is red. Since s = 2,
C1 has fewer than k vertices. The following algorithm decomposes G \C1 into cycles:
1. Initialize i = 1 and v 1 = s.
2. Begin with vertex v i and take the directed blue path until a vertex v i +1 from the set
{1, 2, . . . , v i − 1} is encountered.
3. Take the red path from v i +1 to v i . Call the resulting cycle Ci +1 .

4. If v i +1 = 1, we are done. Otherwise, increase i by 1 and go to step 2.
Since no decomposition into smaller cycles exists, one of the cycles Ci has k vertices. In particular,
Ci contains the vertex 1, so by the above construction, v i = 1. Let P be the blue path in Ci from v i −1
to v i = 1. Since s cannot lie in the interior of the red path Ci − P from 1 to v i −1 , it must lie on P .
The blue edge into s emanates from 1. However, 1 is the last vertex of P , so s must be the first
vertex of P , i.e. v i −1 = s. This shows that P is a path from s to 1 with vertex set {s, s + 1, . . . , k, 1}.
Its union with the blue edge 1 → s gives a blue cycle contained in G2 , so G2 equals this cycle. Since
G2 is not the cycle (1s(s + 1) . . . k), it contains an edge α → β with s α < β k and β = α + 1. The
same argument with α and β relabeled as 1 and s now shows that the vertex set of G2 in the old
labeling is {β, β + 1, . . . , k, 1, . . . , s, . . . , α }. This is a contradiction, which proves the key claim.
We return to the lemma. For m = 2 we simply use the key claim. Suppose m = 3. Let G1 , G2
and G3 be the three cycles. The m = 2 case tells us that G1 ∪ G2 can be decomposed into smaller
cycles C1 , . . . , Cr . The trick now is to take the union of some Ci with G3 and apply the key claim. If
Ci has at most |Ci | − 2 directed edges in common with G3 , we are done. Indeed, we can label the
vertices so that G3 = (12 . . . k) and Ci = (1si 3 . . . il ) with s = 2. Also, Ci = (1s(s + 1) . . . k), otherwise it
has |Ci | − 1 edges in common with G3 . Hence, the key claim applies. We are left with the case where
each Ci has |Ci | − 1 edges in common with G3 . Assume further that C1 , . . . , Cr are those constructed
by the algorithm in the key claim. It is then not difficult to deduce that either G1 = G3 or G2 = G3 .
This contradicts the assumption that the three graphs are distinct. 2
Proposition 7. The algebra C[c ∗ ] is integral over its subalgebra C[ P ∗ ].
Proof. Let R k = C[ P ∗ , {c π }|π |k ] ⊂ C[ X ] be the subring generated by the principal minors and cy-
cles of length at most k. Note that R n = C[c ∗ ] and R 2 = C[ P ∗ ] since c (i j ) = P i P j − P i j for all distinct
i , j ∈ [n]. Thus, it suffices to show that R k is integral over R k−1 for all 3 k n. In particular, we
need to show that each k-cycle c π is the root of a monic polynomial in R k−1 [ z] where z is an inde-
terminate.
We claim that the monic polynomial p ( z) = π ∈C I ( z − c π ) is in R k−1 [ z] for all I ⊆ [n]. Indeed,
the coefficient of z N −d , 1 d N = |C I |, in p ( z) is

αd = (−1)d c π1 c π2 · · · c πd .
{π1 ,...,πd }⊆C I
Observe that α1 = −C [k] which lies in C[ P ∗ ] ⊆ R k−1 by Proposition 4. For d > 1, we apply Lemma 6
which implies that each monomial in αd can be expressed as a product of smaller cycles. This shows
that αd ∈ R k−1 . 2
Corollary 8. If { A k }k>0 is a sequence of complex n × n-matrices whose principal minors are bounded, then the
cycles c π ( A k ) are also bounded.
Proof. Proposition 7 implies that c π satisfies a monic polynomial with coefficients in C[ P ∗ ]. Since
the principal minors are bounded, these coefficients are also bounded, so the same is true for c π . 2
Proof of Theorem 1. Suppose { A k }k>0 is a sequence of complex n × n-matrices whose principal minors
n
tend to u ∗ ∈ C2 . Since the cycle values c π ( A k ) are bounded, we can pass to a subsequence and
assume that the sequence of values for each cycle converges to a complex number v π . Lemma 9
below states that the image of the cycle map is closed. Hence there exists an n × n-matrix A such that
c π ( A ) = v π for all cycles. The limit minor u I is expressed in terms of the v π using the formula (2).
We conclude that the principal minors of the matrix A satisfy P I ( A ) = u I for all I . 2
The following lemma concludes the proof of Theorem 1 and this section.
2
Lemma 9. Let M be the number of cycles and consider the map γ : Cn → C M whose coordinates are the
cycle monomials c π in C[ X ]. Then the image of the monomial map γ is a closed subset of C M . (So, it is a toric
variety.)
Proof. The general question of when the image of a given monomial map between affine spaces is
closed was studied and answered independently in [2] and in [6]. We can apply the characterizations
given in either of these papers to show that the image of our map γ is closed. The key observation
is that the cycle monomials generate the ring of invariants of the (C∗ )n -action on C[ X ] given by
X → D · X · D −1 where D is an invertible diagonal matrix. Equivalently, the exponent vectors of
2
the monomials c π are the minimal generators of a subsemigroup of Nn that is the solution set of a
2
system of linear equations on n unknowns. The geometric meaning of this key observation is that the
monomial map γ represents the quotient map Cn → Cn /(C∗ )n in the sense of geometric invariant
2 2
theory. Now, the results on images of monomial maps in [2, §3] and [6] ensure that Im γ is closed. 2
3. Sixty-five affine relations
We seek to identify generators for the prime ideal In of polynomials in C[ A ∗ ] that vanish on
the image Im φa of the affine principal minor map. Here the 2n coordinates A I of the vector A ∗ are
n
regarded as indeterminates. For n 3, every vector u ∗ ∈ C2 , u ∅ = 1, is realizable as the principal
minors of some n × n-matrix, so In = 0. In this section, we determine In for the case n = 4.
Finding relations among the principal minors of a 4 × 4-matrix is a classical problem posed by
MacMahon in 1894 and partially solved by Nanson in 1897 [8–10]. The relations were discovered by
means of “devertebrated minors” and trigonometry. Here, we write the Nanson relations in terms of
the cycle-sums. They are the maximal 4 × 4-minors of the following 5 × 4-matrix:
⎛ ⎞
C 123 C 14 C 124 C 13 C 134 C 12 2C 234 C 12 C 13 C 14 + C 134 C 124 C 123
⎜ C 124 C 23 C 123 C 24 C 234 C 21 2C 134 C 21 C 23 C 24 + C 234 C 124 C 123 ⎟
⎜ ⎟
⎜ ⎟
⎜ C 134 C 32 C 234 C 31 C 123 C 34 2C 124 C 31 C 32 C 34 + C 234 C 134 C 123 ⎟ . (4)
⎜ ⎟
⎝ C 234 C 41 C 134 C 42 C 124 C 43 2C 123 C 41 C 42 C 43 + C 234 C 134 C 124 ⎠
1 1 1 C 1234
Each of the cycle-sums in this matrix can be rewritten as a polynomial in the principal minors P I
using the relations (3). An explicit example is
C 123 = 2 A 1 A 2 A 3 − A 12 A 3 − A 13 A 2 − A 23 A 1 + A 123 .
The maximal minors of (4) give us five polynomials in the ideal I4 . Each can be expanded either in
terms of cycle-sums or in terms of principal minors.
Muir [9] and Nanson [10] left open the question of whether additional polynomials are needed
to generate the ideal I4 , even up to radical. We applied computer algebra methods to answer this
question. In the course of our experimental investigations, we discovered the 65 affine relations that
generate the ideal. They are the generators of the ideal quotient (K : g ) where K is the ideal generated
by the five 4 × 4-minors above, and g is the principal 3 × 3-minor corresponding to the first three
rows and columns of (4). Thus the main stepping stone in the proof of Theorem 2 is the identity
I4 = (K : g ). (5)
Before proving this, we present a census of the 65 ideal generators, and we explain why all 65 poly-
nomials are needed and to what extent they are uniquely characterized by the equality of ideals
Table 1
Multidegrees of the 65 minimal generators of I4 .
No. Cycle-sums Principal minors No. Cycle-sums Principal minors

Size Degree Size Multidegree Size Degree Size Multidegree
1 32 8 5234 [4, 5, 5, 5] 34 163 10 5234 [4, 5, 5, 7]
2 32 8 5234 [5, 4, 5, 5] 35 163 10 5234 [4, 5, 7, 5]
3 32 8 5234 [5, 5, 4, 5] 36 163 10 5234 [4, 7, 5, 5]
4 32 8 5234 [5, 5, 5, 4] 37 163 10 5234 [5, 4, 5, 7]
5 42 9 4912 [4, 4, 6, 6] 38 163 10 5234 [5, 4, 7, 5]
6 42 9 4912 [4, 6, 4, 6] 39 163 10 5234 [5, 5, 4, 7]
7 42 9 4912 [4, 6, 6, 4] 40 163 10 5234 [5, 5, 7, 4]
8 42 9 4912 [6, 4, 4, 6] 41 163 10 5234 [5, 7, 4, 5]
9 42 9 4912 [6, 4, 6, 4] 42 163 10 5234 [5, 7, 5, 4]
10 42 9 4912 [6, 6, 4, 4] 43 163 10 5234 [7, 4, 5, 5]
44 163 10 5234 [7, 5, 4, 5]
11 80 9 5126 [4, 5, 5, 6] 45 163 10 5234 [7, 5, 5, 4]
12 80 9 5126 [4, 5, 6, 5]
13 80 9 5126 [4, 6, 5, 5] 46 254 10 5558 [4, 5, 6, 6]
14 80 9 5126 [5, 4, 5, 6] 47 254 10 5558 [4, 6, 5, 6]
15 80 9 5126 [5, 4, 6, 5] 48 254 10 5558 [4, 6, 6, 5]
16 80 9 5126 [5, 5, 4, 6] 49 214 10 6716 [5, 4, 6, 6]
17 80 9 5126 [5, 5, 6, 4] 50 214 10 6716 [5, 6, 4, 6]
18 80 9 5126 [5, 6, 4, 5] 51 214 10 6716 [5, 6, 6, 4]
19 80 9 5126 [5, 6, 5, 4] 52 254 10 5558 [6, 4, 5, 6]
20 80 9 5126 [6, 4, 5, 5] 53 254 10 5558 [6, 4, 6, 5]
21 80 9 5126 [6, 5, 4, 5] 54 254 10 5558 [6, 5, 4, 6]
22 80 9 5126 [6, 5, 5, 4] 55 254 10 5558 [6, 5, 6, 4]
56 254 10 5558 [6, 6, 4, 5]
23 116 9 5656 [5, 5, 5, 5] 57 254 10 5558 [6, 6, 5, 4]
24 116 9 5656 [5, 5, 5, 5]
25 116 9 5656 [5, 5, 5, 5] 58 354 10 5993 [5, 5, 5, 6]
59 354 10 5993 [5, 5, 6, 5]
26 91 10 6088 [3, 6, 6, 6] 60 354 10 5993 [5, 6, 5, 5]
27 91 10 6088 [6, 3, 6, 6] 61 364 10 8224 [6, 5, 5, 5]
28 91 10 6088 [6, 6, 3, 6]
29 91 10 6088 [6, 6, 6, 3] 62 685 11 5915 [4, 6, 6, 6]
63 685 11 5915 [6, 4, 6, 6]
30 834 11 5779 [5, 5, 5, 7] 64 685 11 5915 [6, 6, 4, 6]
31 834 11 5779 [5, 5, 7, 5] 65 685 11 5915 [6, 6, 6, 4]
32 834 11 5779 [5, 7, 5, 5]
33 834 11 5779 [7, 5, 5, 5]
in (5). The polynomial ring C[ A ∗ ] has 15 indeterminates that are indexed by non-empty subsets of
{1, 2, 3, 4}. It has the following natural multigrading by the group Z4 :
deg( A 1 ) = [1, 0, 0, 0], deg( A 2 ) = [0, 1, 0, 0], ..., deg( A 4 ) = [0, 0, 0, 1],
deg( A 12 ) = [1, 1, 0, 0], ..., deg( A 234 ) = [0, 1, 1, 1], deg( A 1234 ) = [1, 1, 1, 1].
This Z4 -grading is a positive grading, i.e., each graded component is a finite-dimensional C-vector
space. Both the ideal K and the prime ideal I4 are homogeneous in this Z4 -grading. This means that
the minimal generators of both ideals can be chosen to be Z4 -homogeneous, and their number is
unique.
Our computation revealed that this number of generators is 65. Moreover, we found that the
generators lie in 63 distinct Z4 -graded components. The component in degree [5, 5, 5, 5] happens
to be three-dimensional. We chose a C-basis for this 3-dimensional space of polynomials. All 62
other components are one-dimensional, and these give rise to generators with integer coefficients
and content 1 that are unique up to sign. The complete census of all 65 generators is presented in
Table 1. For each generator we list its multidegree and its number of monomials (“size”) in the two
expansions, namely, in terms of cycle-sums C I and in terms of principal minors A I .
The first four rows of Table 1 refer to the four maximal minors of the matrix (4) which involve the
last row. The expansion of any of these four minors in terms of cycle-sums has 32 monomials and is
of total degree 8. However, the expansion of that polynomial in terms of principal minors A I is much
larger: it has 5234 monomials.
Proof of Theorem 2. We compute the ideal (K : g ) and find that it has the 65 minimal generators
above. We check that each of the five generators of K vanishes on Im φa but g does not vanish
on Im φa . This implies (K : g ) ⊆ I4 . To prove the reverse inclusion we argue as follows. Computation
of a Gröbner basis in terms of cycle-sums reveals that (K : g ) is an ideal of codimension 2, and we
know that this is also the codimension of the prime ideal I4 . Therefore I4 is a minimal associated
prime of (K : g ). To complete the proof, it therefore suffices to show that (K : g ) is a prime ideal. We
do this using the following lemma:
Lemma 10. Let J ⊂ k[x1 , x2 , . . . , xn ] be an ideal containing a polynomial f = gx1 + h, with g , h not involv-
ing x1 and g a non-zero divisor modulo J . Then, J is prime if and only if the elimination ideal J ∩ k[x2 , . . . , xn ]
is prime.
Lemma 10 is due to M. Stillman and appears in [13, Proposition 4.4(b)]. We apply this lemma to
our ideal J = (K : g ) in the polynomial ring C[ A ∗ ], with x1 = A 1234 as the special variable, and we
take the special polynomial f to be the 4 × 4-minor of the matrix (4) formed by deleting the fourth
row.
We have f = g A 1234 + h where g , h are polynomials that do not involve A 1234 . A computation
verifies that ( J : g ) = J , so g is not a zero-divisor modulo J . It remains to show that the elimination
ideal J ∩ C[ A ∗ \ A 1234 ] is prime. Now, since J has codimension two, this elimination ideal is principal.
Indeed, its generator is the 4 × 4-minor of (4) given by the first four rows. This polynomial has Z4 -
degree [6, 6, 6, 6]. We check using computer algebra that it is absolutely irreducible, and conclude
that J is prime. 2
4. A pinch of representation theory
In this section we prove Theorem 3, and we explicitly determine the 14 polynomials of degree 12
that serve as highest weight vectors for the relevant irreducible G-modules. We begin by describing
the general setting for n 4.
Let Vn ⊂ P2 −1 be the image of the projective principal minor map φ , and let Jn ∈ C[ A ∗ ] be
n
the homogeneous prime ideal of polynomials in 2n unknowns A I that vanish on Vn . Clearly, Jn is

invariant under the action of Sn on C[ A ∗ ] which comes from permuting the rows and columns of
the n × n-matrix. Let GL2 (C) denote the group of invertible complex 2 × 2-matrices, and consider the
vector space V = V 1 ⊗ V 2 ⊗ · · · ⊗ V n where each V i C2 with basis e 0i , e 1i . We identify a basis vector
e 1j ⊗ e 2j ⊗ · · · ⊗ enj of the tensor product V with the unknown A I where i ∈ I if and only if j i = 1,
1 2 n
for all i.
n
The natural action of the n-fold product G = GL2 (C) × · · · × GL2 (C) on V C2 extends to its
coordinate ring C[ A ∗ ]. This action commutes with the map φ in (1). Here G acts on the parameter
space of n × 2n-matrices ( A , B ) by having its ith factor GL2 (C) act on the n × 2-matrix formed by
the ith columns of A and B. This argument, due to J.M. Landsberg, shows that the prime ideal Jn is
invariant under G. See also [4, §6] and [11, Theorem 1.1].
Corollary 11. The set Vn is closed in P2 −1 , i.e. it is a projective variety.

n
Proof. For any index set I ⊆ [n], there exists a group element g ∈ G which takes A ∅ to A I . Thus, every
affine piece U I = {u ∗ ∈ Vn : u I = 0} of the constructible set Vn = Im φ is isomorphic to U ∅ = Im φa .
The latter image is closed by Theorem 1. Therefore, Vn is an irreducible projective variety. 2
The action of the Lie group G gives rise to an action of the Lie algebra g = gl2 (C) × · · · × gl2 (C) on
the polynomial ring C[ A ∗ ] by differential operators. Here gl2 (C) denotes the ring of complex 2 × 2-
matrices. Indeed, the vector

w x
0, . . . , 0, , 0, . . . , 0 ∈ g,
y z
whose entries are zero matrices of format 2 × 2 except in the ith coordinate, acts on the polynomial
ring C[ A ∗ ] as the linear differential operator
∂ ∂ ∂ ∂

w AI + x A I ∪{i } + y AI + z A I ∪{i } .
∂ AI ∂ AI ∂ A I ∪{i } ∂ A I ∪{i }
i∈
/I
Here the sum is over all I ⊆ [n] not containing i. We extend this action to all of g by linearity. If all
the coordinate matrices of an element h ∈ g are strictly upper triangular, we say that h is a raising
operator. Similarly, h is a lowering operator if all the matrices are strictly lower triangular.
We now focus on the case n = 4. Let I h be the ideal generated by the homogenizations of the
65 affine generators in Theorem 2 with respect to A ∅ . This is a subideal of the homogeneous prime
ideal J = J4 we are interested in. Our strategy is to identify a suitable intermediate ideal between
Ih ⊂ J .
Proof of Theorem 3. Let K = G I h be the ideal generated by the image of I h under the group G. Then
K is a G-invariant ideal of C[ A ∗ ] that is contained in the prime ideal J . Since G acts transitively on
the affine charts U I , and since I h coincides with the unknown ideal J on the chart U ∅ , we conclude
that the ideal K defines our projective variety V4 scheme-theoretically.
By definition, the ideal K is generated by its degree-12 component K12 . To prove Theorem 3 we
need to show that K12 is a G-module of C-dimension 718, and that it decomposes into 14 irreducible
G-modules. As a G-module, the graded component K12 is generated by the homogenizations of the
65 polynomials in Table 1. Representation theory as in [7,11] tells us that the unique highest weight
vectors of the irreducible G-modules contained in K12 can be found by applying raising operators
h ∈ g to these 65 generators.
Indeed, consider the 1st, 5th and 26th polynomials in Table 1. When written in terms of cycle-
sums, these polynomials have 32, 42 and 91 terms respectively. Let D, E and F be their homoge-
nizations with respect to A ∅ . By applying the raising and lowering operators in the Lie algebra g, one
checks that all 65 generators lie in the G S4 -orbit of D, E and F . Thus,
K12 = M D ⊕ M E ⊕ M F ,
where M D , M E and M F are the G-modules spanned by the G S4 -orbits of the polynomials D, E
and F respectively. Furthermore, as in [7], we write

S i jkl = S (12−i ,i ) C2 ⊗ S (12− j , j ) C2 ⊗ S (12−k,k) C2 ⊗ S (12−l,l) C2
for the tensor product of Schur powers of C2 . The dimension of S i jkl equals (13 − 2i )(13 − 2 j ) ×
(13 − 2k)(13 − 2l). Our three G-modules decompose as
M D S 4555 ⊕ S 5455 ⊕ S 5545 ⊕ S 5554 ,
M E S 4466 ⊕ S 4646 ⊕ S 4664 ⊕ S 6446 ⊕ S 6464 ⊕ S 6644 ,
M F S 3666 ⊕ S 6366 ⊕ S 6636 ⊕ S 6663 .

The above three vector spaces have dimensions 540, 150 and 28 respectively. This shows that
dim(K12 ) = 718, and the proof of Theorem 3 is complete. 2
Many questions about relations among principal minors remain open at this point, even for n = 4.
The most basic question is whether the ideal K is prime, that is, whether K = J holds (cf. Re-
mark 14). Next, it would be desirable to find a nice determinantal representation for the polynomials
E and F , in analogy to D being the homogenization of the determinant of the last four rows in (4).
We know little about the prime ideal In for n 5. It contains various natural images of the ideal I4 ,
but we do not know whether these generate. The most optimistic conjecture would state that In is
generated by the GL2 (C)n -orbit of the polynomials D, E and F . The work of Oeding [11] gives hope
that at least the set-theoretic version might be within reach:
Conjecture 12. The variety Vn ⊂ P2 −1 is cut out by equations of degree 12.

n
5. Singularities of the hyperdeterminant
Our object of study is the projective variety V4 which is parametrized by the principal minors
of a generic 4 × 4-matrix. We have seen that V4 is a variety of codimension two in the projective
space P15 . That ambient space is the projectivization of the vector space C2 ⊗ C2 ⊗ C2 ⊗ C2 of
2 × 2 × 2 × 2-tables a = (ai jkl ). This section offers a geometric characterization of the variety V4 .
The articles [4,11] show that the variety parametrized by the principal minors of a symmetric
matrix is closely related to the 2 × 2 × 2-hyperdeterminant. It is thus quite natural for us to ask
whether such a relationship also exists in the non-symmetric case. We shall argue that this is indeed
the case.
The hyperdeterminant of format 2 × 2 × 2 × 2 is a homogeneous polynomial of degree 24 in 16
unknowns having 2,894,276 terms [5]. Its expansion into cycle-sums was found to have 13,819 terms.
The hypersurface ∇ of this hyperdeterminant consists of all tables a ∈ P15 such that the hypersurface

1
1
1
1
1 1 1 1
Ha = (x, y , z, w ) ∈ P × P × P × P : ai jkl xi y j zk w l = 0
i =0 j =0 k =0 l =0
has a singular point. Weyman and Zelevinsky [15] showed that the hyperdeterminantal hypersurface
∇ ⊂ P15 is singular in codimension one. More precisely, by [15, Theorem 0.5(5)], the singular lo-
cus ∇sing of ∇ is the union in P15 of eight irreducible projective varieties, each having dimension 13:

∇sing = ∇node (∅) ∪ ∇node {i , j } ∪ ∇cusp . (6)
1i < j 4
Here, the node component ∇node (∅) is the closure of the set of tables a such that Ha has two singular
points (x, y , z, w ) and (x , y , z , w ) with x = x , y = y , z = z and w = w . The extraneous component
∇node ({1, 2}) is the closure of the set of tables a such that Ha has two singular points (x, y , z, w )
and (x, y , z , w ), and similarly for the other five extraneous components. Finally, the cusp compo-
nent ∇cusp parametrizes all tables a for which Ha has a triple point. The connection to our study is
given by the following result:
Theorem 13. The node component in the singular locus of the 2 × 2 × 2 × 2-hyperdeterminant coincides with
the variety parametrized by the principal minors of a generic 4 × 4-matrix. In symbols, we have V4 = ∇node (∅).
Proof. We now dehomogenize by setting x0 = y 0 = z0 = w 0 = 1, x1 = x, y 1 = y, z1 = z and w 1 = w.

Let (c i j ) be a generic complex 4 × 4-matrix and consider the ideal in C[x, y , z, w ] generated by the
3 × 3-minors of
⎛ ⎞
c 11 + x c 12 c 13 c 14
⎜ c c 22 + y ⎟
⎜ 21 c 23 c 24 ⎟
⎜ ⎟. (7)
⎝ c 31 c 32 c 33 + z c 34 ⎠
c 41 c 42 c 43 c 44 + w
We claim that the variety of this ideal consists of two distinct points in C4 . We prove this by a compu-
tation. Regarding the c i j as unknowns, we compute a Gröbner basis for the ideal of 3 × 3-minors with
respect to the lexicographic term order x > y > z > w over the base field K = Q(c 11 , c 12 , . . . , c 44 ). The
output shows that the ideal is radical and the initial ideal equals x, y , z, w 2 .
The determinant of (7) is the affine multilinear form

1
1
1
1
F (x, y , z, w ) = ai jkl · xi y j zk w l ,
i =0 j =0 k =0 l =0
whose coefficients are the principal minors of the 4 × 4-matrix (c i j ). The claim established in the
previous paragraph implies that the system of equations
∂F ∂F ∂F ∂F
F= = = = =0
∂x ∂y ∂z ∂w
has two distinct solutions over the algebraic closure of K = Q(c 11 , . . . , c 44 ). These two solutions cor-
respond to two distinct singular points of the hypersurface Ha in P1 × P1 × P1 × P1 . From this we
conclude that the table a = (ai jkl ) of principal minors of (c i j ) lies in the node component ∇node (∅).
Our argument establishes the inclusion V4 ⊆ ∇node (∅). Both sides are irreducible subvarieties
of P15 , and in fact, they share the same dimension, namely 13. This means they must be equal. 2
Our results in Section 4 give an explicit description of the equations that define the first compo-
nent in the decomposition (6). This raises the problem of identifying the equations of the other seven
components.
Remark 14. After posting the first version of this paper on the arXiv, we learned that some of
the results in this paper have already been addressed in [1, §4]. Specifically, Landsberg’s observation
that Vn is G-invariant coincides with [1, Theorem 4.2], and our Theorem 13 coincides with [1, Theo-
rem 4.6]. Coincidentally, without reference to dimensions, we proved V4 ⊆ ∇node (∅) directly while [1,
Theorem 4.6] gives the other inclusion. Moreover, at the very end of the paper [1], it is stated that
“. . . the variety has degree 28, with ideal generated by a whopping 718 degree 12 polynomials”. This appears
to prove our conjecture (at the end of Section 4) that the ideal K is actually prime. We verified that
the ideal K has degree 28, but we did not yet succeed in verifying that K is saturated with respect
to the irrelevant maximal ideal. We tried to do this computation by specializing the 16 unknowns
to linear in fewer unknowns but this leads to an ideal which is not prime. It thus appears that the
ideal K is not Cohen–Macaulay.
Acknowledgments
Shaowei Lin was supported by graduate fellowship from A*STAR (Agency for Science, Technology
and Research, Singapore). Bernd Sturmfels was supported in part by the U.S. National Science Foun-
dation (DMS-0456960 and DMS-0757236). We thank Luke Oeding for conversations and software that
helped us in Section 4, we thank Harald Helfgott for bringing the classical papers [9,10,13] to our
attention, and we thank Eric Rains for pointing us to his earlier results in [1, §4], discussed above.
References
[1] A. Borodin, E. Rains, Eynard–Mehta theorem, Schur process, and their Pfaffian analogs, J. Stat. Phys. 121 (2005) 291–317.
[2] D. Geiger, C. Meek, B. Sturmfels, On the toric algebra of graphical models, Ann. Statist. 34 (2006) 1463–1492.
[3] O. Holtz, H. Schneider, Open problems on GKK τ -matrices, Linear Algebra Appl. 345 (2002) 263–267.
[4] O. Holtz, B. Sturmfels, Hyperdeterminantal relations among symmetric principal minors, J. Algebra 316 (2007) 634–648.
[5] P. Huggins, B. Sturmfels, J. Yu, D. Yuster, The hyperdeterminant and triangulations of the 4-cube, Math. Comp. 77 (2008)
1653–1679.
[6] A. Katsabekis, A. Thoma, Toric sets and orbits on toric varieties, J. Pure Appl. Algebra 181 (2003) 75–83.
[7] J.M. Landsberg, L. Manivel, On the ideals of secant varieties of Segre varieties, Found. Comput. Math. (2004) 397–422.
[8] P.A. MacMahon, A certain class of generating functions in the theory of numbers, Philos. Trans. Roy. Soc. London 185 (1894)
111–160.
[9] T. Muir, The relations between the coaxial minors of a determinant of the fourth order, Trans. Roy. Soc. Edinburgh 39
(1898) 323–339.
[10] E.J. Nanson, On the relations between the coaxial minors of a determinant, Philos. Magazine (5) 44 (1897) 362–367.
[11] L. Oeding, Set-theoretic defining equations of the variety of principal minors of symmetric matrices, arXiv:0809.4236.
[12] R. Stanley, Enumerative Combinatorics 1, Cambridge Univ. Press, 1997.
[13] M. Stillman, Tools for computing primary decompositions and applications to ideals associated to Bayesian networks, in:
A. Dickenstein, I. Emiris (Eds.), Solving Polynomial Equations. Foundations, Algorithms and Applications, Springer-Verlag,
Heidelberg, 2005, pp. 203–239.
[14] E.B. Stouffer, On the independence of principal minors of determinants, Trans. Amer. Math. Soc. 26 (1924) 356–368.
[15] J. Weyman, A. Zelevinsky, Singularities of hyperdeterminants, Ann. Inst. Fourier 46 (1996) 591–644.
Journal of Algebra
Binary Hermitian forms over a cyclotomic field

Dan Yasaki 1
Department of Mathematics and Statistics, 146 Petty Building, University of North Carolina at Greensboro, Greensboro, NC 27402-6170,
United States
Article history: Let ζ be a primitive fifth root of unity and let F be the cyclotomic
Received 21 January 2009 field F = Q(ζ ). Let O ⊂ F be the ring of integers. We compute the
Available online 8 July 2009 Voronoï polyhedron of binary Hermitian forms over F and classify
Communicated by John Cremona
GL2 (O )-conjugacy classes of perfect forms. The combinatorial data
Keywords:
of this polyhedron can be used to compute the cohomology of the
Voronoï polyhedron arithmetic group GL2 (O ) and Hecke eigenforms.
Hermitian forms © 2009 Elsevier Inc. All rights reserved.
Perfect forms
1. Introduction
Let F /Q be a number field, and let O denote its ring of integers. There exists an algorithm to
compute all of the GLn (O )-equivalence classes of perfect n-ary quadratic forms over F once an initial
perfect form is found [14,16]. This is investigated in the totally real number field case in [13,17,19].
We remark that a different notion of perfection has been investigated in [3,6,20]. In this paper, we
consider the cyclotomic field Q(ζ5 ).
Although Hermitian forms over number fields are of interest in their own right, our main motiva-
tion for this computation comes from an investigation of a Taniyama–Shimura type correspondence
over F , relating Hecke eigenforms over F with integer eigenvalues and elliptic curves over F . Due to
the work of Wiles et al. such a correspondence is known to exist for F = Q [9]. It is an open problem
for [ F : Q] > 1. This has been investigated for F an imaginary quadratic field in [5,7,8,18,21,22] and
for F a real quadratic field in [10,11]. In an ongoing project joint with P. Gunnells, F. Hajir, and D. Ra-
makrishnan, we are investigating the complex quartic field F = Q(ζ5 ). To this end, we first compute
the Voronoï polyhedron associated to binary Hermitian forms over Q(ζ5 ). The Voronoï polyhedron
provides a combinatorial structure in which to perform the Hecke eigenvalue computations.
E-mail address: d_yasaki@uncg.edu.

1
Partially supported by UNCG New Faculty Grant.
doi:10.1016/j.jalgebra.2009.06.009
D. Yasaki / Journal of Algebra 322 (2009) 4132–4142 4133
The results of the computations in this paper could also be used to compute the elliptic points
of GL2 (O ) using techniques of [23]. Since the forms and their minimal vectors are given explic-
itly here, one could also extract invariants such as an additive analogue of the Hermite constant
for Q(ζ5 ) [3,20].
The paper is organized as follows. In Section 2, we set notation. In Sections 3–5 we recall the
Voronoï polyhedron and its relation to Hermitian forms over Q(ζ5 ). The Voronoï polyhedron for Q(ζ5 )
is computed in Section 6.
2. Notation
First we set notation and recall and collect a few basic facts from algebraic number theory that
will be used later.
2.1. Field
Let ζ = ζ5 = e 2π i /5 be a primitive fifth root of unity, and let F be the cyclotomic field F = Q(ζ ).
Let O ⊂ F denote √ the ring of integers, and let ¯· denote complex conjugation. Let
√ k ⊂ F denote the real
subfield k = Q( 5 ), and let Ok denote its ring of integers. Then u 5 = (1 + 5 )/2 is a fundamental
unit for k. Let Ok+ denote the totally positive elements of O K .
Let ι = (ι1 , ι2 ) denote the (non-complex conjugate) embeddings
ι: F → C × C
√ √ √
given by sending 5 to ( 5, − 5 ), or equivalently given by sending ζ to (ζ, ζ 3 ). Denote the non-
trivial embedding by · . Specifically, for α ∈ F , let (α , α ) denote ι(α ).
2.2. Binary Hermitian forms over F
Definition 2.1. A binary Hermitian form over F is a map φ : F 2 → k of the form
φ(x, y ) = axx̄ + bx ȳ + b̄ x̄ y + c y ȳ ,
where a, c ∈ k and b ∈ F such that φ and φ , viewed as forms on F ⊗Q R are positive definite.
Note that φ̂ = φ + φ takes values in Q. Indeed, φ̂ is precisely the composition Trk/Q ◦ φ , and by
choosing a Q-basis for F , φ̂ can be viewed as a quadratic form over Q. In particular, it follows that
φ̂(O2 ) is discrete in Q.
Using φ̂ , we can define minimal vectors and perfection. Specifically, the minimum of φ is
m(φ) = inf φ̂( v ).

v ∈O 2 \{0}
A vector v ∈ O 2 is minimal vector for φ if φ( v ) = m(φ). The set of minimal vectors for φ is de-
noted M (φ). A Hermitian form over F is perfect if it is uniquely determined by M (φ) and m(φ).
3. Self-adjoint homogeneous cone
3.1. Symmetric space
Let G be the Q-group Res F /Q (GL2 ) and let G = G(R) the corresponding group of real points.
Let K ⊂ G be a maximal compact subgroup, and let A G be the identity component of the maxi-
mal Q-split torus in the center of G. Then the symmetric space associated to G is a 7-dimensional
space X = G / K A G .
4134 D. Yasaki / Journal of Algebra 322 (2009) 4132–4142
3.2. Cone of Hermitian forms over C
Every Hermitian form over C can be represented by a Hermitian matrix. Let C be the cone of
positive definite binary complex Hermitian forms, viewed as a subset of V , the R-vector space of
2 × 2 complex Hermitian matrices. Let ·∗ denote complex conjugate transpose. Then the usual action
of GL2 (C) on C is given by
( g · φ)( v ) = φ( g ∗ v ), where g ∈ GL2 (C) and φ ∈ C . (1)
3.3. Cone of Hermitian forms over F
Let V be the Weil restriction Res F /Q H 2 of the variety H 2 of 2 × 2 complex Hermitian matrices
defined over F . Then V := V(R)
V × V and

a b
V(Q)
H 2 ( F ) = : a, c ∈ k and b ∈ F .
b̄ c
Let C be the cone C = C × C ⊂ V . Since C is the space of positive-definite Hermitian forms on C2 ,

we can use ι to view C as the space of forms on F 2 . Specifically, for φ = (φ1 , φ2 ) ∈ C and v ∈ F 2 , we
define

φ( v ) = φ1 ι1 ( v ) + φ2 ι2 ( v ) .
Definition 3.1. Let φ be a Hermitian form over F . A 2 × 2 matrix A with coefficients in F is associated
to φ if φ = ι( A ).
Since the map ι is injective, this matrix is unique and will be denoted A φ . For v ∈ F 2 , then φ( v )
is just the trace. Specifically, if φ ∈ V(Q) with associated matrix A φ , then
φ( v ) = Trk/Q ( v ∗ A φ v ).
3.4. C as a symmetric space
The embedding ι gives an isomorphism
G → GL2 (C) × GL2 (C). (2)
Under this identification, ι( A G ) = {(r I , r I ) | r > 0}, where I is the 2 × 2 identity matrix.
Combining (1) and (2), we get an action of G on C . Let φ0 denote the binary Hermitian form
represented by I . Then the stabilizer in Γ of ι(φ0 ) is a maximal compact subgroup K . The group A G
acts on C by positive real homotheties, and we have
X = C /R>0
X C × X C × R,
where X C is the symmetric space for SL2 (C).
3.5. C as a self-adjoint homogeneous cone
Let C¯ denote the closure of C in V . Each vector w ∈ C2 gives a rank 1 Hermitian form w w ∗ (here
w is viewed as a column vector). Combined with ι, we get a map q : F 2 → C¯ given by
q( v ) = ( v v ∗ , v v ∗ ).
The cone C is a self-adjoint homogeneous cone [1]. In particular, C is endowed with a scalar product,
and the interpretation of C as a space of forms over F is reflected in the scalar product. Specifically,
suppose φ ∈ V(Q) and v ∈ O 2 . Then φ, q( v ) = φ( v ).
4. Voronoï polyhedron
In this section we recall properties of the Voronoï polyhedron. This section follows [15], and more
details can be found there.
4.1. Rational structure of V
Fix Λ ⊂ V to be the lattice generated by q( v ) for v ∈ O 2 , and let Λ = Λ \ {0}. We will refer to
points of C ∩ V(Q) as Hermitian forms over F .
Let R ( v ) be the ray R>0 · q( v ) ⊂ C¯. The set of rational boundary components C1 of C is the set of
rays of the form R ( v ), v ∈ F 2 [1].
Definition 4.1. The Voronoï polyhedron Π is the closed convex hull in C¯ of the points C1 ∩ Λ .
4.2. Voronoï decomposition
By construction GL2 (O ) acts on Π . By taking the cones on the faces of Π , one obtains a
Γ -admissible decomposition of C for Γ = GL2 (O) [1]. Since the action of GL2 (O) commutes with
the homotheties, this decomposition descends to a GL2 (O )-equivariant tessellation of X by ideal
7-dimensional polytopes.
5. Primitivity and minimal vectors
There is another notion of minimal vectors that is explored in [1,14]. In this section, we show that
our notion of minimality agrees with theirs in this case by first defining primitive points for the case
of interest and examining both notions of minimal vectors. This should follow from general results
of [16].
The statements in this section are valid only because the class number of k is 1. The results should
extend more generally to a CM field F with real maximal subfield k of class number 1.
5.1. Primitive points
There are several notions of primitivity that we will need.

α
Definition 5.1. A vector v = β ∈ O 2 is primitive if the ideal (α , β) = O .
Proposition 5.2. Let u , v ∈ O 2 be primitive vectors. Then q(u ) = q( v ) if and only if u = τ v for some torsion
unit τ ∈ O .
Proof. One direction is immediate. Specifically, if τ ∈ O is a torsion unit, then τ τ̄ = τ τ̄ = 1, and

thus q(τ v ) = q( v ) follows from the definition of Λ.
α ξα
For the converse, if v = β then u = ηβ for some ξ, η ∈ F . Since q(u ) = q( v ),
ξ ξ̄ = ηη̄ = 1 and ξ η̄ = ξ̄ η = 1.
It follows that u = ξ v for some ξ ∈ F that satisfies ξ ξ̄ = 1. Write ξ as ξ = λ/μ for some λ, μ ∈ O
with (λ, μ) = O . Since ξ α ∈ O , it follows that μ | α . Similarly, since ξβ ∈ O , it follows that μ | β .
Since v is primitive, μ is a unit, and hence ξ ∈ O . Then ξ must be a torsion unit because ξ ξ̄ = 1. 2
a c
Definition 5.3. A form φ ∈ Λ ∩ C¯ is primitive form if there exists a matrix A = c̄ b
with gcd(a, b) = 1
and ι( A ) = φ .
Note that if φ = ι( A ) is primitive, then c ∈ O and a, b ∈ Ok+ .
Lemma 5.4. Let a, b ∈ Ok× be totally positive with gcd(a, b) = 1. Then ab ∈ N F /k ( F ) if and only if a ∈ N F /k ( F )
and b ∈ N F /k ( F ).
Proof. If a and b are norms, then their product is clearly a norm. For the other direction, suppose
m = ab = N F /k (ξ ) for some ξ ∈ F . Since gcd(a, b) = 1, it suffices

to consider the case where m is
square-free in k. Factor (ξ ) as a product of prime ideals (ξ ) = π in F . Since m is square-free, we
must have that each π lies over 5 or a rational prime p such that p ≡ 1 mod 5. It follows that (ξ )
has a factorization in k
(ξ ) = ρ ℘1 ℘2 · · · ℘n ,
√
where the ℘i are prime ideals in k and ρ is either trivial or generated by 5. Each of these factors
has a generator which is a norm, and so it follows that (a) and (b) have generators which are norms.
Specifically, e 1 a and e 2 b are norms for some choice of units e 1 , e 2 ∈ Ok∗ . The fundamental unit u 5 of k
is positive in one embedding into R and negative in the other. In particular since a, b, e 1 a, and e 2 b
are totally positive, e 1 and e 2 must be even powers of u 5 . It follows that a and b are themselves
norms. 2
Proposition 5.5. Let φ ∈ Λ ∩ C¯ be a primitive rank 1 form. Then there exists a primitive vector v ∈ O 2 such
that q( v ) = φ .
a c
Proof. Since φ is primitive, there exists a matrix A = c̄ b with gcd(a, b) = 1 and ι( A ) = φ . Since φ
is rank 1, c c̄ = N F /k (c ) = ab. By Lemma 5.4, there exist α , β ∈ F with N F /k (α ) = a and N F /k (β) = b.
Since gcd(a, b) = 1, it follows that v = (α , β)t is a primitive vector with q( v ) = φ as desired. 2
Remark. Since Λ is a lattice, there is another notion of primitivity for φ ∈ Λ. Specifically, φ is primitive
in Λ if φ can be extended to a basis of the lattice. Note that Proposition 5.5 shows that primitive
rank 1 forms correspond to primitive vectors in O 2 and primitive vectors in O 2 give rise to primitive
rank 1 forms. However, while primitive rank 1 forms are primitive in Λ, there are many vectors
primitive in Λ that are not primitive as forms.
Even when we restrict ourselves to considering vectors that are primitive in Λ and rank 1, if
we relax the condition of gcd(a, b) = 1, we can no longer guarantee that this vector comes from
a primitive vector in O 2 . For example, 4 + u 5 is totally positive, but there does not exist α ∈ F such
4+ u 5 0
that α ᾱ = 4 + u 5 . Thus the point ι(
0 0
) is not in q(O2 ). However, we do have the following
result, which basically says that although there are rank 1 rational forms which are not in the image
of O 2 , they are contained in Ok · q(O 2 ).
Proposition 5.6. Let φ ∈ Λ ∩ C1 be a rank 1 form. Then there exists α ∈ Ok+ and a primitive vector v ∈ O 2
such that α · q( v ) = φ .
a c
Proof. Write φ as φ = ι( A ), where A = c̄ b
for some a, b ∈ Ok+ and c ∈ O . If φ is primitive, the result
follows from Proposition 5.5 with α = 1. If φ is not primitive, choose α ∈ Ok+ such that α = gcd(a, b).
Since φ is rank 1, it follows that c c̄ = ab. In particular α 2 | c c̄. Since α ∈ Ok , we have that α | c
and α | c̄. Thus we have φ = ι(α A 0 ), where A 0 = α −1 A. Since A 0 corresponds to a primitive rank 1
form, Proposition 5.5 implies that there exists a primitive vector v ∈ O 2 such that q( v ) = ι( A 0 ). The
result follows. 2
Proposition 5.7. Let b ∈ Ok+ . Then there exists an α ∈ O \ {0} and t ∈ Ok+ ∪ {0} such that
b = α ᾱ + t .
Proof. The square of the fundamental unit η = u 25 ∈ Ok is totally positive. Since η = u 5 ū 5 , multi-
plication by η acts on Ok+ and preserves N F /k (O ). In particular it suffices to show the result for
a fundamental domain for the action of η on Ok+ . Once can take the positive cone spanned by 1
and η2 . It is clear that every point in the cone has the form η2 + t for some t ∈ Ok+ except for 1
and 2. The condition is trivially satisfied for 1 and 2. 2
5.2. Minimal vectors
There is another notion of minimal vectors and perfect forms described in [14]. Specifically one
can define
m̂(φ) = inf φ, ψ
ψ∈Λ ∩C1
and

M̂ (φ) = ψ ∈ Λ ∩ C1 φ, ψ = m̂(φ) .
Notice that if ψ ∈ M̂ (φ), then ψ is primitive in Λ. It is clear that m̂(φ) m(φ). If m̂(φ) = m(φ), then

q( v ) v ∈ M (φ) ⊆ M̂ (φ).
Proposition 5.8. Let φ ∈ C . Then
m̂(φ) = min m(b · φ).

b∈Ok+
Proof. For ψ ∈ Λ ∩ C1 , there exists b ∈ Ok+ and v ∈ O 2 such that ψ = b · q( v ) by Proposition 5.6. It
follows that

φ, ψ = φ, b · q( v ) = Trk/Q ( A φ bv v ∗ ) = Trk/Q ( v ∗ b A φ v ).
Then
m̂(φ) = inf φ, ψ = inf Trk/Q ( v ∗ b A φ v ) = inf m(b · φ). 2

ψ∈Λ ∩C1 v ∈O 2 b∈Ok+
b∈Ok+
Proposition 5.9. Let φ ∈ C . Then m(φ) = m̂(φ).
Proof. By Proposition 5.8,
m̂(φ) = inf m(b · φ).

b∈Ok+
In particular m̂(φ) m(φ), and to prove the result it suffices to show that
m(φ) m(b · φ) for every b ∈ Ok+ . (3)

By Proposition 5.7, there exists an α ∈ O and t ∈ Ok+ such that b = α ᾱ + t. We compute

Trk/Q ( v ∗ b A φ v ) = Trk/Q v ∗ (α ᾱ + t ) A φ v

= Trk/Q (α v )∗ A φ (α v ) + Trk/Q (t v ∗ A φ v ).
We have that Trk/Q ((α v )∗ A φ (α v )) m(φ). Furthermore Trk/Q (t v ∗ A φ v ) > 0 since t and v ∗ A φ v are
both totally positive. The result follows. 2
6. Voronoï cones for F
In this section, we describe the GL2 (O )-conjugacy classes of Voronoï cones. We implement the
method described in [14].
6.1. Perfect forms
We find one perfect form φ , with associated matrix

1 ζ3 + ζ2 + 3 ζ3 − ζ2 + ζ − 1
Aφ = .
5 −2ζ 3 − ζ − 2 ζ3 + ζ2 + 3
This is done using Magma [4] as follows.
(i) Fix a list L ⊂ O 2 of vectors large enough so that the conditions

φ( v ) = 1 v ∈ L
has a unique solution.

(ii) Ensure that φ is positive definite.
(iii) Check that L ⊆ M (φ).
Step (i) is accomplished by including more vectors into the list L until the linear system has a unique
solution. Steps (ii) and (iii) are accomplished by picking a Z-basis for O 2 and expressing φ as
a quadratic form on Z8 .
The perfect form φ has 240 minimal vectors. It is clear that if v ∈ M (φ) then τ v ∈ M (φ) for
any torsion unit τ ∈ O . There are 24 minimal vectors (modulo torsion units). Let ω denote the unit
ω = ζ + ζ 2 . Then the form φ has (modulo torsion units) minimal vectors
3
−ζ + 1 −ζ + 1 1 1 1 1 1
, , , , , , ,
ζ3 + 1 1 −ω −ζ 2 0 ζ3 −ζ 2 + 1
−1 −1
1 1 1 1 1 ω ω
, , , , , , ,
1 ζ3 + 1 ζ +1 ζ3 + ζ + 1 −ζ 4 ζ4 ζ4 − 1
−1
ω ω −1 ω −1 ω ω ω
, , , , , ,
−1 −ζ 3 − 1 −ζ 3 − ζ 2 − 1 ω+1 −ζ 3 0

ω ω 0 0
2 , , , . (4)
ζ ω 1 ω
By Proposition 5.2, these give rise to 24 distinct points in C¯. These 24 points are vertices of a top-
dimensional Voronoï cone. We show in Section 6.2 that this is the only top-dimensional Voronoï cone
modulo GL2 (O ).
Proposition 6.1. There is 1 GL2 (O )-conjugacy class of 8-dimensional cones. The corresponding perfect form φ
has (modulo torsion units) 24 minimal vectors.
6.2. Top cones
The perfect forms correspond to 8-dimensional Voronoï cones. Specifically, each top cone corre-
sponds to a facet F of the Voronoï polyhedron. There is a unique point φ F ∈ C ∩ V(Q) [14] such
that
(i) F = {x ∈ Π | x, φ F = 1}, and

(ii) for all x ∈ Π \ F , we have x, φ F > 1.
Let Z F be the minimal vertices of φ F , the finite set of points x ∈ Λ ∩ C1 such that x, φ F = 1. Then F
is the convex hull of Z F . The form φ F is a perfect form, and

q( v ): v ∈ M (φ F ) = Z F .
For γ ∈ GL2 (O ), let Θ γ = (γ ∗ )−1 . The action of GL2 (O ) on perfect forms, minimal vectors, and
minimal vertices are related by the following.
Proposition 6.2. Let F be a top cone and let φ F be the corresponding perfect form with minimal vectors M (φ F )
and minimal vertices Z F . For γ ∈ GL2 (O ),
(i) Θ γ · φ = φγ · F , and
(ii) M (Θ γ · φ) = γ · M (φ).
In particular, two perfect forms are GL2 (O )-conjugate if and only if their minimal vectors or minimal vertices
are GL2 (O )-conjugate.
Thus to classify the perfect forms modulo GL2 (O ), one can instead classify the top-dimensional
Voronoï cones. The top-dimensional cone corresponding to φ , denoted Cφ is has faces given by the
convex hull of {q( v ) | v ∈ M (φ)}. The program Polymake [12] is used to compute the convex hull.
There are 118 codimension 1 faces, corresponding to 118 neighboring top-dimensional Voronoï cones
and 118 perfect forms. There are 14 faces with 12 vertices, 80 faces with 9 vertices, and 24 faces
with 7 vertices. Using Magma, the stabilizer S φ of φ is computed. The group S φ has order 600 and
Magma type 600, 54.
In order to cut down the number of computations that need to be made, we first classify the faces
of Cφ modulo S φ , the stabilizer of M (φ). Indeed, let ψ be a perfect form such that Cψ meets Cφ in
a codimension 1 face F . If γ ∈ S φ , then γ · ψ is the perfect form corresponding to top cone γ · Cψ ,
which meets Cφ in a codimension 1 face γ · F .
It is clear that S φ conjugate faces must have the same number of vertices. One computes that
there are 3 orbits of faces with 12 vertices. One orbit S φ · F 1 consists of 12 faces, and the other
2 orbits consist of 1 element each, denoted F 2 and F 3 . One shows that F 2 is GL2 (O )-conjugate to F 3
by computing a matrix that sends F 2 to F 3 . There are 4 orbits of faces with 9 vertices. The orbits
S φ · F 4 , S φ · F 5 , S φ · F 6 , and S φ · F 7 consist of 20 faces each. There are 2 orbits of faces with 7 vertices.
The orbits S φ · F 8 and S φ · F 9 consist of 12 faces each.
To show that there is only 1 GL2 (O )-class of perfect form, it suffices to show that the perfect
forms φi associated to the top cones that neighbor Cφ at face F i are each GL2 (O )-conjugate to φ . To
this end, we use the following lemma.
Lemma 6.3. Let F , F be two faces of Cφ with associated perfect forms ψ , ψ . If γ ∈ GL2 (O )\ S φ and γ · F = F ,
then both ψ and ψ are GL2 (O )-conjugate to φ .
Proof. Since Cφ is a top-dimensional cone, and F and F are codimension 1 faces of Cφ , we must
/ S φ , it follows that γ · Cφ = Cψ . Thus Θ γ · φ = ψ .
have that γ · Cφ = Cφ or γ · Cφ = Cψ . Since γ ∈
Repeating the argument using γ −1
shows that Θ γ −1 · φ = ψ . 2
Table 1
GL2 (O )-conjugacy classes of Voronoï cones.
Type # of v Cone rank Stabilizer

A 24 8 600, 54
B1 9 7 30, 4
B2 9 7 30, 4
B3 12 7 100, 14
B4 12 7 600, 54
B5 7 7 50, 5
C1 6 6 30, 4
C2 6 6 10, 2
C3 7 6 10, 2
C4 6 6 60, 11
C5 6 6 50, 5
C6 6 6 20, 2
C7 7 6 10, 2
C8 6 6 20, 2
C9 6 6 30, 4
C 10 6 6 50, 5
D1 5 5 10, 2
D2 5 5 10, 2
D3 5 5 20, 5
D4 5 5 10, 2
D5 5 5 10, 2
D6 5 5 10, 2
D7 5 5 10, 2
D8 5 5 10, 2
D9 5 5 20, 5
D 10 5 5 100, 14
D 11 6 5 20, 5
E1 4 4 20, 5
E2 4 4 20, 2
E3 4 4 20, 2
E4 4 4 10, 2
E5 4 4 10, 2
E6 4 4 20, 5
E7 4 4 20, 5
E8 4 4 10, 2
E9 4 4 200, 31
F1 3 3 60, 11
F2 3 3 20, 5
F3 3 3 100, 16
F4 3 3 20, 5
G1 2 2 200, 31
G2 2 2 Γ∞
6.3. Lower cones
In an analogous way, one classifies the lower-dimensional faces.
Theorem 6.4. There is exactly 1 GL2 (O )-class of 8-cones, 5 GL2 (O )-classes of 7-cones, 10 classes of 6-cones,
11 classes of 5-cones, 9 classes of 4-cones, 4 classes of 3-cones, and 2 classes of 2-cones.
Table 1 gives the GL2 (O )-classes of Voronoï cones along with the number of Π vertices, the rank
of the cone, and the stabilizer in GL2 (O ) of the cone. The Magma small group type of each stabilizer
is given for the finite groups. In particular, the first component is the order of the group. Let U ∗
denote the subgroup of upper triangular matrices in GL2 (O ) such that the top left entry is a torsion
unit.
6.4. Classification of forms
Note that Theorem 6.4 gives a classification of binary Hermitian forms over F based on the
configuration of the minimal vectors. Indeed, by duality, the vertices of the cones that arise above
correspond to GL2 (O )-classes of configurations of minimal vectors for forms over F [2]. For example,
there are 4 distinct classes of cones with 3 vertices. Hence there are 4 distinct GL2 (O )-types of binary
Hermitian forms over F with exactly 3 minimal vectors. One can distinguish the types by studying
the configuration of minimal vectors.
Let [·] denote the subset of the minimal vectors of the perfect form in the order given in (4). For
example, [5, 23] = {e 1 , e 2 }. Since F has class number one, every primitive vector in O 2 is GL2 (O )-
conjugate to e 1 . Combined with Theorem 6.4, one computes the following.
Theorem 6.5. Let φ be a binary Hermitian form over F . Then M (φ) is GL2 (O )-conjugate to exactly one of the
following
[5], [5, 20], [5, 23], [5, 8, 23], [5, 22, 23], [5, 20, 23], [5, 10, 23], [5, 8, 22, 23],
[4, 5, 8, 23], [5, 8, 10, 23], [5, 8, 18, 23], [5, 18, 20, 23], [5, 15, 17, 23], [5, 19, 20, 23],
[5, 18, 19, 23], [5, 20, 23, 24], [4, 5, 8, 18, 23], [5, 8, 10, 12, 23], [5, 8, 20, 22, 23],
[4, 5, 8, 9, 23], [5, 8, 18, 19, 23], [5, 8, 12, 20, 23], [5, 8, 18, 22, 23], [5, 18, 19, 20, 23],
[5, 20, 22–24], [5, 9, 15, 17, 23], [5, 8, 18, 19, 20, 22], [5, 8, 10, 12, 21, 23],
[5, 8, 12, 20, 21, 23], [5, 8, 18, 19, 20, 22, 23], [5, 8, 20, 22–24], [5, 8, 9, 15, 17, 23],
[5, 8, 10, 22–24], [3–5, 8, 15, 18, 23], [4, 5, 8, 18, 22, 23], [4, 5, 8, 9, 15, 23],
[5, 9, 15–17, 23], [5, 8, 10, 12, 20–24], [3, 5, 8, 18–20, 22–24],
[3–5, 7, 8, 10, 13, 15, 18, 22–24], [1, 4, 5, 8, 9, 11, 13–17, 23], [5, 6, 9, 15–17, 23], [1–24]
Acknowledgments
I thank P. Gunnells for many helpful tips, tricks and explanations. I would also like to thank F. Hajir
for helpful conversations. I also thank the reviewer for many helpful comments.
References
[1] A. Ash, Deformation retracts with lowest possible dimension of arithmetic quotients of self-adjoint homogeneous cones,
Math. Ann. 225 (1) (1977) 69–76.
[2] A. Ash, M. McConnell, Cohomology at infinity and the well-rounded retract for general linear groups, Duke Math. J. 90 (3)
(1997) 549–576.
[3] R. Baeza, R. Coulangeon, M.I. Icaza, M. O’Ryan, Hermite’s constant for quadratic number fields, Experiment. Math. 10 (4)
(2001) 543–551.
[4] W. Bosma, J. Cannon, C. Playoust, The Magma algebra system. I. The user language, in: Computational Algebra and Number
Theory, London, 1993, J. Symbolic Comput. 24 (3–4) (1997) 235–265.
[5] J. Bygott, Modular forms and modular symbols over imaginary quadratic fields, PhD thesis, Exeter University, 1998.
[6] R. Coulangeon, Voronoï theory over algebraic number fields, in: Réseaux euclidiens, designs sphériques et formes modu-
laires, in: Monogr. Enseign. Math., vol. 37, Enseignement Math., Geneva, 2001, pp. 147–162.
[7] J.E. Cremona, E. Whitley, Periods of cusp forms and elliptic curves over imaginary quadratic fields, Math. Comp. 62 (205)
(1994) 407–429.
[8] J.E. Cremona, Periods of cusp forms and elliptic curves over imaginary quadratic fields, in: Elliptic Curves and Related
Topics, in: CRM Proc. Lecture Notes, vol. 4, Amer. Math. Soc., Providence, RI, 1994, pp. 29–44.
[9] H. Darmon, A proof of the full Shimura–Taniyama–Weil conjecture is announced, Notices Amer. Math. Soc. 46 (11) (1999)
1397–1401.
[10] L. Dembélé, An algorithm for modular elliptic curves over real quadratic
√ fields, Experiment. Math. 17 (4) (2008) 427–438.
[11] L. Dembélé, Explicit computations of Hilbert modular forms on Q( 5 ), Experiment. Math. 14 (4) (2005) 457–466.
[12] E. Gawrilow, M. Joswig, Polymake: An approach to modular software design in computational geometry, in: Proceedings of
the 17th Annual Symposium on Computational Geometry, Medford, MA, 2001, June 3–5, ACM, 2001, pp. 222–231.
[13] P.E. Gunnells, D. Yasaki, Perfect forms over totally real number fields, submitted for publication.
[14] P.E. Gunnells, Modular symbols for Q-rank one groups and Voronoï reduction, J. Number Theory 75 (2) (1999) 198–219.
[15] P.E. Gunnells, D. Yasaki, Hecke operators and Hilbert modular forms, in: Algorithmic Number Theory, in: Lecture Notes in
Comput. Sci., vol. 5011, Springer, Berlin, 2008, pp. 387–401.
[16] M. Koecher, Beiträge zu einer Reduktionstheorie in Positivitätsbereichen. I, Math. Ann. 141 (1960) 384–432.
[17] A. Leibak, An explicit construction of initial perfect quadratic forms over some families of totally real number fields, in:
Algorithmic Number Theory, in: Lecture Notes in Comput. Sci., vol. 5011, Springer, Berlin, 2008, pp. 240–252.
[18] M. Lingham, Modular forms and elliptic curves over imaginary quadratic fields, PhD thesis, University of Nottingham, 2005.
[19] H.E. Ong, Perfect quadratic forms over real-quadratic number fields, Geom. Dedicata 20 (1) (1986) 51–77.
[20] M.E. Pohst, M. Wagner, On the computation of Hermite–Humbert constants for real quadratic number fields, J. Théor.
Nombres Bordeaux 17 (3) (2005) 905–920.
[21] M.H. Şengün, The nonexistence of certain representations of the absolute Galois group of quadratic fields, Proc. Amer. Math.
Soc. 137 (1) (2009) 27–35.
[22] E. Whitley, Modular symbols and elliptic curves over imaginary quadratic number fields, PhD thesis, Exeter University,
1990.
[23] D. Yasaki, Elliptic points of the Picard modular group, Monatsh. Math. 156 (2009) 391–396.
Journal of Algebra
Wythoff polytopes and low-dimensional homology

of Mathieu groups
Mathieu Dutour Sikirić a,∗,1 , Graham Ellis b
a
Rudjer Bosković Institute, Bijenicka 54, 10000 Zagreb, Croatia
b
Mathematics Department, National University of Ireland, Galway, Ireland
Article history: We describe two methods for computing the low-dimensional

Received 5 February 2009 integral homology of the Mathieu simple groups and use them to
Available online 9 October 2009 make computations such as H 5 ( M 23 , Z) = Z7 and H 3 ( M 24 , Z) =
Communicated by Derek Holt
Z12 . One method works via Sylow subgroups. The other method
Keywords:
uses a Wythoff polytope and perturbation techniques to produce
Group homology an explicit free Z M n -resolution. Both methods apply in principle
Resolution to arbitrary finite groups.
Polytope © 2009 Elsevier Inc. All rights reserved.
Wythoff construction
1. Introduction
We describe two methods for computing the integral homology for the Mathieu simple groups
presented on Table 1. The first homology H 1 (G , Z) is trivial for any simple group and so is omitted
from the table (see [3] for an exposition of relevant facts on group homology). The second homology
of Mathieu groups is well-known [16]. A computer method for the second homology of a permutation
group was illustrated on the Mathieu groups M 21 and M 22 in [15]. The mod p cohomology H ∗ (G , F p )
is now known for all Mathieu groups except M 24 [21,1,2,17]. With the help of the Bockstein spectral
sequence it is, in principle, possible to obtain integral homology from mod p cohomology (p ranging
over the prime divisors of the group order), though the details can be difficult. For example, the
calculation of H n ( M 23 , Z) was obtained in this way for 1 n 6 by Milgram [17] and provided the
first example of a non-trivial finite group with trivial integral homology in dimensions 3. It seems
that the mod p cohomology of M 24 is not known for all primes p (see [14] for the case p = 3) and
so we can assign the status of a new theorem to the following result.
*
E-mail addresses: mdsikir@irb.hr (M. Dutour Sikirić), graham.ellis@nuigalway.ie (G. Ellis).
1
First author has been supported by Marie Curie fellowship MTKD-CT-2006-042685 and by the Croatian Ministry of Science,
Education and Sport under contract 098-0982705-2707.
doi:10.1016/j.jalgebra.2009.09.031
4144 M.Dutour Sikirić, G. Ellis / Journal of Algebra 322 (2009) 4143–4150
Table 1
Low-dimensional homology of Mathieu groups with 0 a 53 and 0 b 1.
G H 2 (G , Z) H 3 (G , Z) H 4 (G , Z) H 5 (G , Z)
M 11 0 Z8 0 Z2
M 12 Z2 Z6 ⊕ Z8 Z3 (Z2 )3
M 21 Z4 ⊕ Z12 Z5 0 (Z2 )4 ⊕ Z4 ⊕ Z7
M 22 Z12 0 0 Z2 ⊕ Z2 ⊕ Z7
M 23 0 0 0 Z7
M 24 0 Z12 0 (Z2 )a ⊕ (Z4 )b ⊕ Z7
Table 2
Values of n expressed in term of k 1 such that H n ( M m , Z)( p ) = Z p .
H n ( M m , Z)( p ) = Z p m = 11 m = 12 m = 21 m = 22 m = 23 m = 24
n= n= n= n= n= n=
p=5 8k − 1 8k − 1 4k − 1 8k − 1 8k − 1 8k − 1
p=7 – – 6k − 1 6k − 1 6k − 1 6k − 1
p = 11 10k − 1 10k − 1 – 10k − 1 10k − 1 20k − 1
p = 23 – – – – 22k − 1 22k − 1
Theorem 1. H 3 ( M 24 , Z) = Z12 and H 4 ( M 24 , Z) = 0.
This result (and other table entries) can be obtained from the hap homological algebra pack-
age [10] for the gap computational algebra system [12] using (variants of) the following command.
gap> GroupHomology(MathieuGroup(24),3);
gap> [4, 3]
The algorithm underlying this command is explained in Section 2. The current implementation is
unable to determine the integers a, b in Table 1 though it does establish the ranges 0 a 53,
0 b 1.
Abelian invariants of a (co)homology group are the easiest cohomological information to access.
More difficult information would be, for example, explicit cocycles G n → A corresponding to coho-
mology classes in H n (G , A ). Explicit cocycles are constructed in hap using the induced chain map
B ∗G → R ∗G from the bar resolution B ∗G to an explicit small free ZG-resolution R ∗G of Z. In Sections 3–5
we explain how the Wythoff polytope construction can be used to produce such a resolution R ∗G . This
resolution provides an alternative computation of H 3 ( M 24 , Z).
In Section 6 we determine the p-part H n ( M m , Z)( p ) of the integral homology of the Mathieu
groups for n 1 and primes p 5. For p ∈ {5, 7, 11, 23} the p-part is either trivial or Z p ; it is trivial
for all other primes p 5. Table 2 lists the values of n for which the p-part is non-trivial.
Although the paper focuses on Mathieu groups, the techniques are applicable in principle to arbi-
trary finite groups. In some cases the Wythoff polytopal method is a significantly faster method for
computing the homology groups.
2. Algorithm underlying the HAP function
Given a group G, a free ZG-resolution of the trivial module Z is an exact sequence
0 ← Z ← R 0G ← R 1G ← · · · ← R kG ← · · ·
of free ZG-modules R iG . A previous paper [9] describes an algorithm for computing free ZG-
resolutions for finite G. This has now been implemented as part of the hap package. It takes as
input a finite group G and a positive integer n. It returns:
M.Dutour Sikirić, G. Ellis / Journal of Algebra 322 (2009) 4143–4150 4145
• The rank of the kth module R kG in a free ZG-resolution R ∗G (0 k n).

• The image of the ith free ZG-generator of R kG under the boundary homomorphism dk : R kG → R kG−1
(1 k n).
• The image of the ith free Z-generator of R kG under a contracting homotopy hk : R kG → R kG+1 (0
k n − 1).
The contracting homotopies hk satisfy, by definition, hk dk+1 + dk+2 hk+1 = 1 and need to be specified
on a set of free Abelian group generators of R k since they are not G-equivariant. The homotopy can
be used to make constructive the following frequent element of choice.
For x ∈ ker(dk : R kG → R kG−1 ) choose an element x̃ ∈ R kG+1 such that dk+1 (x̃) = x.
One sets x̃ = hk (x). In particular, for any group homomorphism φ : G → G , the homotopy allows one

to define an induced φ -equivariant chain map φ∗ : R ∗G → R ∗G .
The algorithm in [9] can only handle fairly small groups. For example, the hap implementation
takes 20 seconds on a 2.66 GHz Intel PC with 2 G of memory to compute eight terms of a free ZG-
resolution R ∗G for the symmetric group G = S 5 ; the ZG-rank of R 8G is 115. However, for any group G
there is a surjection
H n (Syl p , Z) → H n (G , Z)( p )
from the homology of a Sylow p-subgroup Syl p = Syl p (G ) onto the p-part of the homology of G. For
a Sylow p-subgroup P there is a description of the kernel of the surjection H n ( P , Z ) → H n (G , Z )( p )
due to Cartan and Eilenberg [4]. It is generated by elements
φ K (a) − φxK x−1 (a),
where x ranges over the double coset representatives of P in G, K = P ∩ xP x−1 , the homomorphisms
φ K , φx−1 K x : H n ( K , Z) → H n ( P , Z) are induced by the inclusion K → P , k → k and the conjugated
inclusion K → P , k → x−1 kx, and a ranges over the generators of H n ( K , Z). Thus, the homology of a
large finite group G can be computed from free resolutions (with specified contracting homotopy) for
each of its Sylow subgroups. Our implementation of the algorithm in [9] can be used to produce six
terms of free Z(Syl p )-resolutions for all Sylow subgroups Syl p of all Mathieu groups except M 24 . The
Sylow subgroup Syl2 ( M 24 ) has order 1024 and requires a specific application of a general technique.
To explain the technique suppose that G is a group, possibly infinite, for which we have some
ZG-resolution of Z
C ∗ : · · · → C n → C n−1 → · · · → C 0 → Z,
but that C ∗ is not free. Suppose that for each m we have a free ZG-resolution of the module C m
D m∗ : → D m,n → D m,n−1 → · · · → D m,0 → C m .
Theorem 2. (See [20].) There is a free ZG-resolution R ∗G → Z with

R nG = D p ,q .
p +q=n
The proof of this theorem of C.T.C. Wall can be made constructive by using contracting homotopies
on the resolutions D m∗ . Furthermore, a contracting homotopy on R ∗G can be constructed by a formula
involving contracting homotopies on the D m∗ and on C ∗ . Details are given in [11].
Table 3
Low-dimensional homology of Sylow subgroups of M 24 for p = 2, 3.
p H 1 (Syl p , Z) H 2 (Syl p , Z) H 3 (Syl p , Z) H 4 (Syl p , Z) H 5 (Syl p , Z)

2 (Z2 )4 (Z2 )8 (Z2 )11 ⊕ (Z4 )6 (Z2 )32 d (Z2 )52 ⊕ Z4
3 (Z3 )2 (Z3 )2 (Z3 )4 (Z3 )3 (Z3 )4 ⊕ Z9
Suppose now that N is a normal subgroup of G and that C ∗ is a free Z(G / N )-resolution. Then,
regarding C ∗ as a ZG-resolution, each free ZG-generator of C m is stabilized by N. Any free Z N-
resolution of Z can be used to construct a free ZG-resolution D m∗ of C m . Thus, using Theorem 2,
we can construct a free ZG-resolution R ∗G from a free Z N-resolution R ∗N and free Z(G / N )-resolution
G /N
R∗ . The constructed resolution is often referred to as a twisted tensor product and denoted by R ∗G =
R∗ ⊗
N ˜ R ∗G / N .
This twisted tensor product has been implemented in hap and can be used to provide free reso-
lutions for the Sylow subgroup Syl p ( M 24 ). Since | M 24 | = 210 · 33 · 5 · 7 · 11 · 23 the non-cyclic Sylow
subgroups occur only for p = 2, 3. Their low-dimensional integral homology can be computed using
hap and is given in Table 3.
In degrees n = 5 the current version of hap fails to determine the image of H n (Syl2 , Z) in
H n ( M 24 , Z). It succeeds in constructing the image as a finitely presented group but fails to deter-
mine the group from this presentation. This failure should be resolved in a future release of hap.
The remainder of the paper is aimed at constructing small free resolutions for large groups such
as M 24 .
3. Orbit polytopes
Suppose that a finite group G acts linearly on Rn . For a vector v ∈ Rn we consider the convex hull

P = P (G , v ) = Conv v g : g ∈ G
of the orbit of v under the action of G. The polytope P has a natural cell structure with respect to
which we can consider the cellular chain complex C ∗ ( P ). The action of G on Rn induces an action
of G on C ∗ ( P ) and we can view C ∗ ( P ) as a chain complex of ZG-modules. Since P is contractible we
have H i (C ∗ ( P )) = 0 for all i 1 and H 0 (C ∗ ( P )) = Z. Furthermore, if the polytope is of dimension m
then H 0 (C ∗ ( P )) ∼
=Z∼ = C m ( P ). So there is a homomorphism C 0 ( P ) → C m−1 ( P ) which can be used to
splice together infinitely many copies of C ∗ ( P ) to form an infinite ZG-resolution
· · · → C 1 → C 0 → C m −1 → · · · → C 2 → C 1 → C 0 → Z
of the trivial ZG-module Z. In principle one can use Theorem 2 to convert C ∗ to a free ZG-resolution.
Precise details are given in [11]. To put this idea into practice one requires:
1. The face lattice of the orbit polytope P (G , v ).

2. For each orbit of cell e in P (G , v ), the subgroup Stab(G , e ) G of elements that stabilize e glob-
ally.
Stab(G ,e )
3. A free Z Stab(G , e )-resolution R ∗ for each stabilizer Stab(G , e ).
Stab(G ,e )
Assuming that the stabilizer groups Stab(G , e ) are reasonably small, resolutions R ∗ are readily
obtained from hap’s implementation of the algorithm in [9]. Thus, to convert C ∗ to a free ZG-
resolution, we must focus on requirements (1) and (2).
One could use computational geometry software such as polymake [13] to determine the combina-
torial structure of P (G , v ) for small groups G. For instance, any permutation group G S n acts on Rn
by π (x1 , . . . , xn ) = (xπ −1 (1) , . . . , xπ −1 (n) ) for π ∈ G. In particular, the Mathieu group M 10 of order 720,
generated by π1 = (1, 9, 6, 7, 5)(2, 10, 3, 8, 4) and π2 = (1, 10, 7, 8)(2, 9, 4, 6), acts on R10 . For the
Fig. 1. Two Wythoff polytopes constructed from D 3 .
vector v = (1, 2, 3, 4, 5, 6, 7, 8, 9, 10) the polytope P ( M 10 , v ) is 9-dimensional with 720 vertices each
of degree 632. The polytope thus has 227 520 edges.
4. Orbit polytopes of finite reflection groups
Let W be a finite reflection group generated by a simple system of Euclidean reflections S =

{s1 , . . . , sn }. For each reflection s ∈ S let H s denote the corresponding reflecting hyperplane and Δ
the fundamental simplex for S. The Coxeter–Dynkin reduced diagram is the graph on S with two
reflections adjacent if they do not commute. Fix a subset ∅ V ⊆ S. The type T = t ( v ) ⊂ S of a point
v ∈ Δ is the set of s ∈ S such that v ∈ / H s . Choose a point v of type V . Let P ( W ; V , v ) denote the
n-dimensional polytope formed by the convex hull of the orbit of v under the action of W .
As an example, consider the 3-dimensional reflection group W = B3 generated by reflections
s1 , s2 , s3 where (s1 s2 )3 = 1, (s1 s3 )2 = 1 and (s2 s3 )4 = 1. For V = {s1 , s2 , s3 } and vector v ∈ R3 in gen-
eral position but close to the mirrors H s1 and H s3 the polytope P ( W ; V , v ) is pictured in Fig. 1(a).
For V = {s2 , s3 } and v ∈ H s1 the polytope P ( W ; V , v ) is pictured in Fig. 1(b).
Proposition 3. The combinatorial type of P ( W ; V , v ) is independent of the choice of v.
Proof. For V = S the polytope obtained is the well-known permutahedron, whose face-lattice is in-
dependent of v. The stabilizer of a face of P ( W ; V , v ) is a parabolic subgroup and this establishes
an isomorphism between the face lattice of P ( W ; V , v ) and the lattice of parabolic subgroups of W .
Furthermore, the 1-skeleton of P ( W ; S , v ) is the Cayley graph Cay( W , S ) of W with respect to the
generating set S. Observe that in Cay( W , S ) the length of any edge labelled by generator s ∈ S de-
creases as v is moved towards the hyperplane H s , and that the edge ceases to exist when v moves
into H s . Denote by Cay V ( W , S ) the obtained reduced Cayley graph. The group W acts transitively on
its vertex set. The set of vectors v of type t ( v ) = S has measure one in the set of all vectors. Thus
any face F of P ( W ; V , v ) is obtained as the limit of some face G of the permutahedron and the set
of vertices contained in F is the limit of the vertices contained in G. Thus the graph Cay V ( W , S )
determines the face-lattice and this proves the required result. 2
We are going to describe explicitly the face lattice P W , V of the polytope P ( W ; V , v ) for any v of
type V . Our description is of course equivalent to the classic one given in [5,6,22] and reproduces the
one of [19]. Since we assume that W is finite, the Coxeter–Dynkin reduced diagram of W is a tree
and given any two vertices u and v of it we denote by [u , v ] the unique path from u to v.
For two subsets U , U ∈ S we say that U blocks U (from V ) if for all u ∈ U and v ∈ V there is a
u ∈ U , such that u ∈ [u , v ]. This defines a binary relation on subsets of S, which we will denote by

U U . We also write U ∼ U if U U and U U , and we write U < U if U U and U U .

It is easy to see that is reflexive and transitive, which implies that ∼ is an equivalence relation.
Let [U ] denote the equivalence class containing U . It can be shown that if U ∼ U then U ∩ U ∼
U ∼ U ∪ U . This yields that every equivalence class X contains a unique smallest (under inclusion)
subset m( X ) and unique largest subset M ( X ). The subsets m( X ) will be called the essential subsets
of S (with respect to V ). Let E ( V ) be the set of all essential subsets of S. Clearly, the above relation <
is a partial order on E ( V ). Also, V ∈ E ( V ) and V is the smallest element of E ( V ) with respect to <.
The faces F of P ( W ; V , v ) are indexed by their isobarycenters g ( F ). The stabilizer of F is the
stabilizer of g ( F ), that is the parabolic subgroup of W generated by S − t ( g ( F )). The type of such
an isobarycenter is an essential subset of S and all essential subsets are realized as isobarycenters of
faces. The rank of an essential subset is the dimension of the corresponding face. Given two faces F , F
of P ( W ; V , v ), F ⊂ F if and only if we have the type inequality t ( g ( F )) < t ( g ( F )) and { g ( F ), g ( F )}
is contained in at least one image g (Δ) with g ∈ W of the fundamental simplex Δ.
We can use the above formalism to obtain the combinatorial structure of the orbit poly-
tope P ( M 24 , v ) where the Mathieu group acts on R24 by permuting basis vectors, and v =
(1, 2, 3, 4, 5, 0, . . . , 0) ∈ R24 . Since M 24 is a 5-transitive permutation group we have
P ( M 24 , v ) = P ( S 24 , v ).
The symmetric group S 24 is a finite reflection group with simple generating system S = {si = (i , i + 1):
1 i 23}. The vector v lies in those mirrors H si for 6 i 23. So P ( M 24 , v ) = P ( S 24 , V , v ) for
V = { s 1 , . . . , s 5 }.
Our proof of Proposition 3 implies that the polytope P ( M 24 , v ) has | S 24 /si : 6 i 23| =
5 100 480 vertices. The essential subsets of rank 1 defining edges are V − {sk } for 1 k 4 and
( V − {s5 }) ∪ {s6 }. So, the number of edges is

S 24 /s5 , si : 7 i 23 + S 24 /sk , si : 6 i 23 = 58 655 520.
1k4
Each vertex of the polytope has the same degree d say. Thus the number of edges is d × 5 100 480/2 =
58 655 520 from which d = 23. Since P ( M 24 , v ) is of dimension 23, this shows that it is simple.
Each vertex of P ( M 24 , v ) has stabilizer group Stab( M 24 , v ) = M 24 ∩ si : 6 i 23 ∼
= (C 2 × C 2 ×
C 2 × C 2 ) : C 3 of order 48. Under M 24 , for 1 k 4, there is only one orbit of edges of type V − {sk };
they have stabilizer Stab( M 24 , v ) : C 2 of order 96. Under M 24 there are two orbits of edges of type
( V − {s5 }) ∪ {s6 }, one with stabilizer S 3 , the other with stabilizer a 2-group of order 32.
The formalism of essential subsets is a useful tool to determine the face lattice of P ( W ; V , v ) for
a Coxeter group W and provides ready access to the lattice for homology computations. The equality
between the polytopes P ( M 24 , v ) and P ( S 24 , v ) was essential for being able to apply this formalism
and thus get a reasonably simple description of the face lattice.
For an arbitrary vector v and group G we cannot expect to have a simple combinatorial descrip-
tion of the face lattice of P (G , v ) and we need to use specific computational techniques. If G is large,
then we cannot expect to be able to store the vertex set of P (G , v ). Fortunately, by the group ac-
tion, the full face lattice is encoded in the set S ( v ) of vertices adjacent to v. This set S ( v ) can be
computed iteratively by using the Poincaré polyhedron theorem (see [18,7] for some example of such
computations). Once the list of neighbours is known the face-lattice follows easily.
After one has obtained the low-dimensional faces of P ( M 24 , v ) and their stabilizer groups, we can
use Theorem 2 to compute the initial terms of a free Z M 24 -resolution of Z.
5. Wythoff construction for polytopes
The Wythoff construction can also be defined for partially ordered sets. A flag in a poset is an
arbitrary completely ordered subset. We say that a connected poset K is a d-dimensional complex (or,
simply, a d-complex) if every maximal flag in K has size d + 1. In a d-complex K every element x can
be uniquely assigned a number dim(x) ∈ {0, . . . , d}, called the dimension of x, in such a way, that the
minimal elements of K have dimension zero and dim( y ) = dim(x) + 1 whenever x < y and there is
no z with x < z < y. The elements of a complex K are called faces, or k-faces if the dimension of the
face needs to be specified. Furthermore, 0-faces are called vertices and d-faces (maximal faces) are
called facets. If x < y and dim(x) = k, we will say that x is a k-face of y.
Table 4
Rank of resolutions of M 22 , M 23 , M 24 obtained from the Wythoff construction.
G P V Free rank of resolution in degrees 0, 1, 2, . . .

M 22 α21 {0, 1, 2} 1, 7, 33, 113, 301, 694
M 23 α22 {0, 1, 2, 3, 4} 2, 20, 116, 451, 1334, 3279
M 24 α23 {0, 1, 2, 3, 4} 1, 9, 50, 204, 649
For a flag f ⊂ K define its type as the set t ( f ) = {dim( F ): F ∈ f }. Clearly, t ( f ) is a subset of
S = {0, . . . , d} and, conversely, every subset of S is the type of some flag. Let Ω be the set of all non-
empty subsets of S and fix an arbitrary V ∈ Ω . For two subsets U , U ∈ Ω we say that U blocks U
(from V ) if for all u ∈ U and v ∈ V there is a u ∈ U and u u v or v u u. With this notion
of blocking we can define the notion of essential subset of S and the inequality < in the same way
as for Coxeter groups.
The construction of P (K; V ) mimics the one of P ( W ; D , v ) above for Coxeter groups. The Wythoff
complex P (K; V ) consists of all flags F such that t ( F ) is essential. For two such flags F and F , we
have F < F whenever t ( F ) < t ( F ) and F is compatible with F , that is, F ∪ F is a flag. It can be
shown that P (K, V ) is again a d-complex.
The face lattice K( P ) of a (d + 1)-dimensional polytope P is a d-complex, which is a CW-complex
topologically equivalent to a sphere. It is proved in [19] that the topological type of P (K; V ) is the
same as the one of K. This version of the Wythoff construction when applied to a regular polytope
gives a face lattice which is isomorphic to the one obtained by applying the Wythoff construction
to the corresponding Coxeter group. The complex P (K( P ), {0}) is equal to K( P ) and P (K( P ), {d}) is
the complex of the polytope dual to P . In general P (K( P ), V ) is not a polytope since the notion of
convexity is not well preserved by the Wythoff construction without any regularity assumption.
The topological invariance means that if a group G acts on a polytope P then we can apply the
orbit polytope construction to P (K( P ), V ) for a chosen V in order to compute H i (G , Z).
In the case of M 24 , we take as polytope the 23-dimensional simplex α23 and we build the Wythoff
polytope P (α23 ; {0, 1, 2, 3, 4}). In Table 4 we give the results obtained for the larger Mathieu groups.
The method applies to any finite group acting on n points by using the simplex αn−1 . We do not
need G to act transitively. All programs are available from [8].
6. Homology at p = 5, 7, 11, 23
Suppose that a group G has Sylow p-subgroup P = C p of prime order. The Cartan–Eilenberg double
coset formula implies that the surjection
πn : H n ( P , Z) → H n (G , Z)( p)
has kernel generated by the elements
H n (φ g )(a) − a
for g ∈ N G ( P ), a ∈ H n ( P , Z) and φ g : P → P , p → gpg −1 . Here N G ( P ) is the normalizer of P in G.

Using the isomorphism H n−1 ( P , Z) ∼ = H n ( P , Z) and the cohomology ring structure H ∗ ( P , Z) ∼=
Z p [x2 ], we see that a group homomorphism φ : P → P , p → pm induces a homology homomorphism
k
H 2k−1 (φ) : H 2k−1 ( P , Z) → H 2k−1 ( P , Z), a → am .
For p ∈ {5, 7, 11, 23} the Mathieu groups have Sylow p-subgroups which are either trivial or of
prime order. One can use gap to determine their normalizers. It is thus a routine exercise to determine
the p-part of the integral homology of the Mathieu groups, the results of which are given in the
Introduction.
References
[1] A. Adem, J. Maginnis, R.J. Milgram, The geometry and cohomology of the Mathieu group M 12 , J. Algebra 139 (1991) 90–133.
[2] A. Adem, R.J. Milgram, The cohomology of the Mathieu group M 22 , Topology 34 (1995) 389–410.
[3] K.S. Brown, Cohomology of Groups, Springer-Verlag, 1994.
[4] H. Cartan, S. Eilenberg, Homological Algebra, Princeton University Press, 1956.
[5] H.S.M. Coxeter, Wythoff’s construction for uniform polytopes, Proc. London Math. Soc. 38 (1935) 327–339;
Reprinted in H.S.M. Coxeter, Twelve Geometrical Essays, Southern Illinois University Press, Carbondale, 1968, pp. 40–53.
[6] H.S.M. Coxeter, Regular Polytopes, Dover Publications, 1973.
[7] M. Deraux, Deforming the R-Fuchsian (4, 4, 4)-triangle group into a lattice, Topology 45 (2006) 989–1020.
[8] M. Dutour Sikirić, Polyhedral, http://www.liga.ens.fr/~dutour/polyhedral, 2008.
[9] G. Ellis, Computing group resolutions, J. Symbolic Comput. 38 (2004) 1077–1118.
[10] G. Ellis, HAP – Homological algebra programming, Version 1.8, a package for the GAP computational algebra system,
http://www.gap-system.org/Packages/hap.html, 2007.
[11] G. Ellis, J. Harris, E. Sköldberg, Polytopal resolutions for finite groups, J. Reine Angew. Math. 598 (2006) 131–137.
[12] The GAP Group, GAP – Groups, algorithms, and programming, Version 4.4.9, http://www.gap-system.org, 2006.
[13] E. Gawrilow, M. Joswig, Polymake: A framework for analyzing convex polytopes, in: Gil Kalai, Günter M. Ziegler (Eds.),
Polytopes — Combinatorics and Computation, Birkhäuser, 2000, pp. 43–74.
[14] D.J. Green, The 3-local cohomology of the Mathieu group M 24 , Glasg. Math. J. 38 (1996) 69–75.
[15] D.F. Holt, The calculation of the Schur multiplier of a permutation group, in: Computational Group Theory, Durham, 1982,
Academic Press, 1984, pp. 307–319.
[16] P. Mazet, Sur les multiplicateurs de Schur de Mathieu, J. Algebra 77 (1982) 552–576.
[17] R.J. Milgram, The cohomology of the Mathieu group M 23 , J. Group Theory 3 (2000) 7–26.
[18] R. Riley, Application of a computer implementation of Poincare’s theorem on fundamental polyhedra, Math. Comp. 40
(1983) 607–632.
[19] R. Scharlau, Geometrical realizations of shadow geometries, Proc. London Math. Soc. 61 (1990) 615–656.
[20] C.T.C. Wall, Resolutions of extensions of groups, Math. Proc. Cambridge Philos. Soc. 57 (1961) 251–255.
[21] P.J. Webb, A local method in group cohomology, Comment. Math. Helv. 62 (1987) 135–167.
[22] W.A. Wythoff, A relation between the polytopes of the C 600 -family, Proc. Sect. Sci. 20 (1918) 966–970, Koninklijke
Akademie van Wetenschappen te Amsterdam.
Journal of Algebra
Deciding finiteness of matrix groups in positive

characteristic ✩
A.S. Detinko a , D.L. Flannery a,∗ , E.A. O’Brien b
a
School of Mathematics, Statistics and Applied Mathematics, National University of Ireland, Galway, Ireland
b
Department of Mathematics, University of Auckland, Auckland, New Zealand
Article history: We present a new algorithm to decide finiteness of matrix

Received 29 June 2009 groups defined over a field of positive characteristic. Together
Available online 1 October 2009 with previous work for groups in zero characteristic, this provides
the first complete solution of the finiteness problem for finitely
Keywords:
generated matrix groups over a field. We also give an algorithm to
Matrix group compute the order of a finite matrix group over a function field of
Function field positive characteristic by constructing an isomorphic copy of the
Finiteness problem group over a finite field. Our implementations of these algorithms
Order problem are publicly available in Magma.
Algorithm © 2009 Elsevier Inc. All rights reserved.
1. Introduction
Deciding finiteness is a fundamental problem for any class of potentially infinite groups. For matrix
groups over a field of zero characteristic, the algorithms of [1,6] provide a solution of this problem,
and their implementations perform satisfactorily for reasonably large input (cf. [6, Section 4]). De-
ciding finiteness over a purely transcendental extension F of a finite field was considered by several
authors [3,9,10]. The approach taken in [10] relies on the fact that a subgroup G of GL(n, F) is fi-
nite if and only if, for every finite subfield Fq of F, the enveloping algebra G Fq is finite. Since
the Fq -dimension of G Fq may depend exponentially on n (see [10, Theorem 3.3]), this leads to
exponential-time algorithms. The polynomial-time algorithms of [3,9] involve significant computing
over function fields, and so we expect that they are practical only for small input. We know of no
implementations of the algorithms of [3,9,10].
✩
Supported in part by Science Foundation Ireland, grants 07/MI/007 and 08/RFP/MTH1331, and by the Marsden Fund of New
Zealand, grant UOA 721.
*
E-mail address: dane.flannery@nuigalway.ie (D.L. Flannery).
doi:10.1016/j.jalgebra.2009.08.020
4152 A.S. Detinko et al. / Journal of Algebra 322 (2009) 4151–4160
A uniform approach to deciding finiteness of matrix groups over an infinite field via congruence
homomorphisms was proposed in [5, Section 4.3], and applied to nilpotent groups. We implemented
this approach, for rational nilpotent groups, in the computer algebra systems Magma [2] and GAP (see
the ‘Nilmat’ package [4]). Its performance is usually much better than existing procedures in GAP and
Magma.
The idea of using congruence homomorphisms to decide finiteness of matrix groups was further
developed in [6], for groups over a function field of zero characteristic. In this paper we extend the
ideas of [6] to positive characteristic. As in that earlier paper, our main method is the application
of congruence homomorphisms to enable a comparison of dimensions of certain enveloping alge-
bras. However, the finiteness problem in positive characteristic is more complicated: a finite subgroup
of GL(n, F) need not be completely reducible, and it can be unboundedly large. The opposite holds in
characteristic zero.
Despite these difficulties, we obtain a substantial improvement upon the algorithms of [3,9,10].
We avoid their most inefficient step; namely, computing a basis of the enveloping algebra of the
input group over a function field (see Sections 2 and 3). As in [6], much of the computation takes
place in the coefficient field – which is finite here. Although the number of (function and finite)
field operations of our finiteness testing algorithm is polynomial in certain parameters of the input,
our primary goal was to develop a practical algorithm. We have implemented it in Magma [2] and
demonstrate that it performs well for a range of input.
We also give an algorithm to compute the order of a finite matrix group G over a function field of
positive characteristic, relying on the same strategy used to decide finiteness. This algorithm finds an
isomorphic copy of G over a finite field, which can be used to derive additional information about G.
In Section 4 we present a simplified finiteness test for nilpotent groups. Finally, in Section 5 we report
on the performance of our Magma implementation of the algorithms.
By elementary structure theory of finitely generated field extensions, any finitely generated matrix
group G is defined over a finite extension of a function field. As explained below, we can construct an
isomorphism of G onto a group defined over the function field, in larger degree. Thus the results of
this paper together with [1,6] effectively allow us to decide finiteness of a finitely generated matrix
group over any field (cf. also [6, Section 3.2.2]).
2. Preliminaries and background
Let F be a field of characteristic p > 0, and let G = S , where S = { S 1 , . . . , S r } ⊆ GL(n, F). We

may assume that F is a finite extension of a function field E = Fq ( X 1 , . . . , X m ), where the X i are
algebraically independent indeterminates, and Fq is the finite field of size q. Replacement of elements
of F by matrices over E according to the multiplication action of F on an E-basis of F defines an
isomorphism of G into GL(nl, E), where l = |F : E|. So without loss of generality, from now on F =
Fq ( X 1 , . . . , Xm ), m 1, and q is a power of the prime p.
In fact G is contained in GL(n, R ) for a finitely generated integral domain R ⊆ F. We can take
R = 1f Fq [ X 1 , . . . , X m ], where f = f ( X 1 , . . . , X m ) is a common multiple of the denominators of the
non-zero entries of the S i and S −1
i , 1 i r. We say that α = (α1 , . . . , αm ) is admissible (or
S -admissible) if f (α ) = 0. Here the αi are in the algebraic closure Fq of Fq ; note that Fq need not
contain αi such that α is admissible. For an admissible α , let ν denote the positive integer such
that Fq (α ) := Fq (α1 , . . . , αm ) = Fqν . Let ϕα be the ring homomorphism R → Fqν whose kernel is
generated by the monomials X i − αi , 1 i m. If necessary, we extend ϕα to a homomorphism

R = 1f Fqμ [ X 1 , . . . , X m ] → Fqμν for μ 1 in the obvious way. With a slight abuse of notation, the
induced congruence homomorphisms on GL(n, R ) and on the full matrix algebra Mat(n, R ) will also
be denoted ϕα . Evaluation of ϕα on a subset M of Mat(n, R ) is simply substitution of αi for X i in
the entries of the elements of M, 1 i m. We denote ϕα (M) as M(α ).
Lemma 2.1. If G is finite then the kernel of ϕα on G is a p-group.

A.S. Detinko et al. / Journal of Algebra 322 (2009) 4151–4160 4153
Proof. This holds for m = 1 by [5, Proposition 3.2 and Example 3.6]. The result for m > 1 follows
readily: the kernel of a composite of congruence homomorphisms, all of whose kernels are p-groups,
is a p-group. 2
Corollary 2.2. If G is finite and completely reducible, then ϕα is an isomorphism from G onto ϕα (G ) for every
admissible α .
Let L/K be a field extension, and suppose that T is a finite subset of GL(n, L) such that the
enveloping algebra T K is finite-dimensional as a K-vector space. We now describe a standard pro-
cedure that constructs a basis of T K consisting of elements from the monoid generated by T . (Since
we use the procedure to compute an enveloping algebra basis only over a finite field, we assume that
L is finite in the description.)
BasisEnvAlgebra(T , K)
Input: T ⊆ GL(n, L), L a finite field, and K a subfield of L.
Output: a basis of the enveloping algebra H K , where H = T .
(I) A := { I n }.
(II) While there exist A ∈ A and T ∈ T such that AT ∈
/ spanK (A) do A := A ∪ { AT }.
(III) Return A.
We now set up a convention. Suppose that S (α ) is duplicate-free. For A (α ) ∈ Mat(n, Fqν ) that
is a word in the elements of S (α ), we canonically define a pre-image A of A (α ) in GL(n, F): if
A (α ) = S i 1 (α ) · · · S it (α ) then A = S i 1 · · · S it .
Lemma 2.3. Matrices B 1 , . . . , B l in Mat(n, F) are Fq -linearly independent if and only if they are Fqμ -linearly
independent.
l
Proof. The non-trivial Fqμ -linear dependence i =1 ai B i = 0n between the B i yields a system of equa-
tions with coefficients in F. Since (a1 , . . . , al ) is a solution of this system, ai ∈ F ∩ Fqμ = Fq for all i.
Thus, if the B i are Fq -linearly independent then they must be Fqμ -linearly independent. The other
direction is obvious. 2
Corollary 2.4. If G is finite then dimFq G Fq = dimFqμ G Fqμ .
Proof. By Lemma 2.3, dimFq G Fq dimFqμ G Fqμ . Conversely, G Fqμ has a basis consisting of ele-
ments of G; that basis is therefore an Fq -linearly independent subset of G Fq . Hence dimFqμ G Fqμ
dimFq G Fq . 2
We write
F for Fqμ ( X 1 , . . . , Xm ).
Lemma 2.5. If G is finite then the kernel of ϕα on G Fqμ is contained in the radical of G Fqμ and the radical
of G
F.
Proof. The proofs of Proposition 3.2 and Corollary 3.3 in [3] carry over. 2
Lemma 2.6. If G is completely reducible, then G is finite if and only if ϕα : G Fqμ → G (α )Fqμ is an isomor-
phism, for any S -admissible α and μ 1.
Proof. If G is finite then G is completely reducible over the extension field F of F (see e.g. [8, 1.8,
p. 12]), so the radical of G
F is zero. Lemma 2.5 now implies that ker ϕ α on G Fqμ is trivial. 2
Note that Lemma 2.6 implies Corollary 2.2.
Lemma 2.7. The algebras G Fqμ and G (α )Fqμ are isomorphic if and only if
dimFqμ G Fqμ = dimFqμ G (α )Fqμ .
Proof. A basis of G Fqμ maps under ϕα to a spanning set of G (α )Fqμ , which is a basis if and only
if the Fqμ -dimensions of these two algebras are equal. 2
Corollary 2.8. If G is completely reducible, then G is finite if and only if, for every S -admissible α ,
dimFqμ G Fqμ = dimFqμ G (α )Fqμ = dimFq G Fq = dimFq G (α )Fq .
Proof. This follows from Corollary 2.4, Lemma 2.6 and Lemma 2.7. 2
Lemma 2.9. If A 1 (α ), . . . , A d (α ) are Fqμ -linearly independent, then A 1 , . . . , A d are Fqμ -linearly indepen-
dent.
Proof. Clear, since ϕα is Fqμ -linear. 2
Now we state an algorithm to decide whether an enveloping algebra G Fqμ and its congruence
image G (α )Fqμ are isomorphic, for admissible α and μ 1. This uses the same approach as the
algorithm IsFiniteMatGroupFuncNF of [6].
IsIsomorphismEnvAlgebras(S , α , μ)
Input: a finite subset S = { S 1 , . . . , S r } of GL(n, F), an S -admissible α , a positive integer μ.
Output: ‘true’ if ϕα acts on G Fqμ as an isomorphism, where G = S ; ‘false’ otherwise.
(I) If S (α ) has duplicates then return ‘false’.

(II) Construct A(α ) = { A 1 (α ), . . . , A d (α )} := BasisEnvAlgebra(S (α ), Fqμ ).
Let A be the set of canonical pre-images { A 1 , . . . , A d }.
(III) For A i (α ) ∈ A(α ) and S j (α ) ∈ S (α )
d
find ak ∈ Fqμ such that A i (α ) S j (α ) = k=1 ak A k (α ).
d
If A i S j = k=1 ak A k , then return ‘false’.
(IV) Return ‘true’.
If IsIsomorphismEnvAlgebras(S , α , μ) returns ‘true’ then G is finite, and the set A found

in step (II) is a basis of S Fqμ = G Fqμ . (For A is a spanning set by step (III), and it is linearly
independent by Lemma 2.9.) Observe that we obtain this basis after a calculation over a finite field,
rather than over the function field F.
By Lemma 2.6, the following algorithm decides finiteness of a completely reducible subgroup of
GL(n, F).
IsFiniteCRMatGroupFuncFF(S )
Input: a finite subset S of GL(n, F) such that G = S is completely reducible.
Output: ‘true’ if G is finite; ‘false’ otherwise.
(I) Find an S -admissible α .

(II) Return IsIsomorphismEnvAlgebras(S , α , ν ).
Corollary 2.8 implies that we can also decide finiteness of a completely reducible group G by
testing whether ϕα acts as an isomorphism on G Fqμ , for any given μ 1. However dimFqμ G (α )Fqμ
might be larger than dimFqν G (α )Fqν , which is bounded above by n2 .
Now suppose that G is a (finitely generated, not necessarily completely reducible) subgroup
of GL(n, F), and we know α such that ϕα is an isomorphism on G Fqν if G is finite. We may
now decide finiteness of G just as in IsFiniteCRMatGroupFuncFF – namely, by applying
IsIsomorphismEnvAlgebras. Unfortunately, such α need not exist. On the other hand, there
always exist α such that ϕα is an isomorphism on G Fq if G is finite. We consider these issues again
at the end of Section 3.
3. Deciding finiteness and computing orders in positive characteristic
We now present a general algorithm to decide finiteness of a finitely generated subgroup G of

GL(n, F). The approach is similar to the finiteness testing algorithm of [3], but avoids its most compli-
cated step: computing a basis of G F over F. We also outline a simple method to compute the order
of a finite subgroup of GL(n, F).
We continue with established notation. That is, α is an S -admissible m-tuple of elements from Fq
such that S (α ) is duplicate-free, and A(α ) = { A 1 (α ), . . . , A d (α )} is a basis of G (α )Fq (α ) com-
puted via BasisEnvAlgebra, with canonical pre-image A = { A 1 , . . . , A d }. For i and j such that
d d
A i (α ) S j (α ) = k=1 ak A k (α ), where ak ∈ Fqν = Fq (α ), define D = A i S j − k=1 ak A k . We assume that
p does not divide ν . For a ∈ Fqν , denote the trace of a over Fq by tr(a):
tr(a) = a + σ (a) + · · · + σ ν −1 (a), Gal(Fqν /Fq ) = σ .
d
Observe that D := ν A i S j − k=1 tr(ak ) A k is in G F .
Lemma 3.1. Let D and D be as defined above. If G is finite and D = 0n , then D is a non-zero element of the
radical of G F .
d d
Proof. If D = 0n then A i S j = where bk = ν1 tr(ak ) ∈ Fq . In fact A i (α ) S j (α ) = k=1 bk A k (α )
k=1 bk A k
d
implies that bk = ak for all k. But this contradicts D = A i S j − k=1 ak A k = 0n . Hence D is non-zero.
We verify that D ∈ as in the proof of [3, Corollary 3.5]. 2
Lemma 3.2. The nullspace of the radical of G F is a non-zero G-module.
Proof. For all g ∈ G and u in the nullspace U of , we have gu = u = 0, since is an ideal

of G F . Thus GU ⊆ U as required. 2
So if G is finite and D = 0n , then the nullspace of D contains a non-trivial G-module. We can find
such a module using the following procedure.
ModuleViaNullspace(S , E )
Input: a finite subset S of GL(n, F), and E ∈ Mat(n, F).
Output: a G-module U in the nullspace of E, for G = S .
(I) U := Nullspace( E ).
(II) While there exists S i ∈ S such that U ∩ S i U = U do U := U ∩ S i U .
(III) Return U .
Since each pass through the while loop reduces the dimension of U , ModuleViaNullspace
terminates in at most n iterations. If E is a non-zero element of (for example, if G is finite and
E = D for D = 0n ), then the output is a proper non-zero G-submodule of the underlying space V .
Now we present our main algorithm for deciding finiteness. We use the following notation. Let U
be a G-submodule of V and extend a basis of U to a basis of V . Write G with respect to the latter
basis in block triangular form; then ρU denotes the projection homomorphism from G onto the block
diagonal group, whose kernel is the unitriangular subgroup that fixes U and V /U elementwise.
IsFiniteMatGroupFuncFF(S )
Input: a finite subset S of GL(n, F).
Output: ‘true’ if G = S is finite; ‘false’ otherwise.
(I) Find an S -admissible α such that p does not divide ν = |Fq (α )/Fq |.
If S i (α ) = S j (α ) for distinct S i , S j ∈ S , then set E = S i − S j and go to (IV).
(II) A(α ) := BasisEnvAlgebra(S (α ), Fqν ) = { A 1 (α ), . . . , A d (α )}.
Let A be the canonical pre-image { A 1 , . . . , A d } of A(α ).
d
(III) If there exist A i ∈ A and S j ∈ S such that A i S j = k=1 ak A k , where ak ∈ Fqν and A i (α ) S (α ) =
d d
k=1 ak A k ( α ), then set E = ν A i S j − k=1 tr(ak ) A k ;
else return ‘true’.
(IV) U 1 := ModuleViaNullspace(S , E ).
If U 1 = {0} then return ‘false’;
else let ρ = ρU 1 , U 2 = V /U 1 ,
for k = 1, 2 do
A := {ρ ( A 1 )|U k , . . . , ρ ( A d )|U k }, S := {ρ ( S 1 )|U k , . . . , ρ ( S r )|U k }, go to (III).
At any stage of IsFiniteMatGroupFuncFF, we test finiteness of constituents G |U of G in block

triangular form. In looping back to step (III) from step (IV), the dimension of the G-module U strictly
reduces. Thus, eventually the algorithm finds either that all constituents are finite, or that one of them
is infinite. In the former case G has a finite homomorphic image whose kernel is a finitely generated
unipotent subgroup of GL(n, F), and so is also finite; in the latter case G is infinite.
The maximum number of iterations of IsFiniteMatGroupFuncFF is 2n, and its main com-
ponent BasisEnvAlgebra has cost O (rn8 ) finite field operations. The principal difference be-
tween IsFiniteMatGroupFuncFF and the simpler alternative IsFiniteCRMatGroupFuncFF
for completely reducible input is that the former calls ModuleViaNullspace. The operations car-
ried out over the function field are matrix addition, matrix multiplication, and nullspace and intersec-
tion of subspaces. All use O (nk ) field operations where k 3. For just one indeterminate, admissible
α always exist in Fqd+1 where d is the largest degree of denominators in entries of the matrices
in S ; a similar estimate holds for m > 1. In practice, admissible alpha may be found by repeatedly
evaluating the denominator polynomial f ( X 1 , . . . , X m ) for X i chosen in finite extensions of the prime
subfield, until a non-root is obtained.
We turn now to the problem of computing the order of a finite subgroup of GL(n, F). Below we
give a simple procedure to solve this problem, based on the next lemma.
Lemma 3.3. Let M be a finite subset of Mat(n, F). There are infinitely many admissible α = (α1 , . . . , αm ),
αi ∈ Fq , such that |M| = |M(α )|. If m = 1 then |M| = |M(α )| for all but finitely many admissible α .
Proof. Let M = { M 1 , . . . , M k }. For each pair i , j, where i < j, choose a position in which M i and
M j have different entries, and let di j be the difference of the entries. Denote by h the product

1i < j k d i j of all these differences. If h (α ) = 0 then |M| = |M(α )|. Since there are infinitely many
admissible α such that h(α ) = 0, and only finitely many admissible α such that h(α ) = 0 if m = 1,
the result follows. 2
Corollary 3.4. Let G GL(n, F) be finite. There are infinitely many admissible α such that |G | = |G (α )| and
|G Fq | = |G (α )Fq |. If m = 1 then |G | = |G (α )| and |G Fq | = |G (α )Fq | for all but finitely many admissi-
ble α .
Remark 3.5. It is not true that if G is finite then there are infinitely many admissible α such that
|G Fqν | = |G (α )Fqν |. Indeed dimFqν G (α )Fqν may be less than dimFq G (α )Fq for every admissi-

1 1 1 X
ble α . For example, consider the subgroup G of GL(2, F2 ( X )) generated by 01
and
01
. For all
α ∈ F2 we have dimF2 (α ) G F2 (α ) = 3, whereas dimF2 (α ) G (α )F2 (α ) = 2.
By Corollary 3.4, if G is finite and m = 1, then there is a positive integer δ such that ϕα is an iso-
morphism on G Fq whenever α ∈ Fq \ Fqδ . As such δ may be impracticably large, our implementation
of the following algorithm uses the intrinsic random selection function in Magma.
SizeFiniteMatGroupFuncFF(S )
Input: S ⊆ GL(n, F) such that G = S is finite.
Output: |G |.
(m)
(I) Select an S -admissible α ∈ Fq .
(II) If IsIsomorphismEnvAlgebras(S , α , 1) = ‘true’ then return |G (α )|;
(m) (m)
else replace Fq by Fq \ {α } and go to (I).
We end this section with some comments on SizeFiniteMatGroupFuncFF. Recall that

dimFq G Fq may depend exponentially on n. However, sometimes we can replace (S , α , 1) by (S , α , ν )
in step (II) above, thereby bringing the relevant dimension back to no more than n2 . For instance, this
is valid if G is cyclic or completely reducible. However, in general we cannot make this modification
(cf. Remark 3.5).
Notice that SizeFiniteMatGroupFuncFF constructs an isomorphic copy of G GL(n, F) de-
fined over a finite field. We can use this copy and machinery for matrix groups over finite fields to
answer other questions about G.
4. Deciding finiteness of nilpotent matrix groups
In this section we develop a specialized algorithm to decide finiteness of nilpotent subgroups

of GL(n, F). We remove the limitation of [5, Section 4.3] that the ground field is perfect. Our algorithm
represents an improvement of the positive characteristic finiteness testing algorithm in [5], including
a more efficient transfer to the completely reducible case. One important application is to decide
whether a single element g (equivalently, subgroup g ) of GL(n, F) has finite order.
For the rest of this section, G GL(n, F) is nilpotent. We let g s and g u denote respectively the
diagonalizable and unipotent parts of g ∈ GL(n, F). That is, g s and g u are the unique matrices such
that g s ∈ GL(n, F) is diagonalizable, g u ∈ GL(n, F) is unipotent, and g = g s g u = g u g s .
Lemma 4.1. If g ∈ GL(n, F) has finite order then g s and g u are both in g .
Proof. Cf. [11, Corollary 1, p. 135]. 2
Define G s = ( S 1 )s , . . . , ( S r )s and G u = ( S 1 )u , . . . , ( S r )u . The next result follows from part of [11,

Proposition 3, pp. 136–137] (which does not require that the ground field be perfect).
Lemma 4.2.
(i) The maps defined by g → g s and g → g u for g ∈ G are homomorphisms; thus G s = { g s | g ∈ G } and
G u = { g u | g ∈ G }.
(ii) G G s × G u .
Lemma 4.3. G is finite if and only if G s is finite.

Proof. By Lemma 4.2(i), G s is finite if G is finite. As a finitely generated periodic matrix group, G u is
finite. Hence if G s is finite then G is finite by Lemma 4.2(ii). 2
Let γ be the positive integer such that p γ −1 < n p γ . By [12, p. 192], p γ is the maximum order
γ pγ γ γ
of a unipotent element of GL(n, F). Define S p = { S i | 1 i r } and G p = S p .
Lemma 4.4.
γ
(i) G is finite if and only if G p is finite.
γ
(ii) If G is finite then G p = G s is completely reducible.
γ γ
Proof. (i) Certainly G p G is finite if G is finite. Suppose that G p is finite. Then each S i has finite
pγ pγ pγ
order, so ( S i )s has order coprime to p. Thus ( S i )s ∈ ( S i )s . Since ( S i )s ∈ S i by Lemma 4.1, we
γ
have G s G p , and so G s is finite. Lemma 4.3 now completes the proof of this item.
γ γ
(ii) If G is finite then G s G p . Further, G p G s since each generator of the nilpotent group
pγ
G G has trivial unipotent part (by the choice of γ ). 2
Lemma 4.4 establishes correctness of the following algorithm to decide finiteness of nilpotent sub-
groups of GL(n, F).
IsFiniteNilpotentMatGroupFuncFF(S )
Input: a finite subset S of GL(n, F) such that G = S is nilpotent.
Output: ‘true’ if G is finite; ‘false’ otherwise.
γ pγ
(I) S p := { S i | 1 i r }.
γ
(II) Return IsFiniteCRMatGroupFuncFF(S p ).
For nilpotent input, IsFiniteNilpotentMatGroupFuncFF is superior to IsFiniteMat-

GroupFuncFF, because it immediately reduces to the completely reducible case.
IsFiniteNilpotentMatGroupFuncFF may be further refined. Rather than computing a basis
γ
of an enveloping algebra in step (II), it suffices to test whether ϕα has trivial kernel on G p . A practi-
cal method to do this is given at the end of [5, Section 4.2]. Likewise, computing orders can be made
more efficient for nilpotent input. A specialized method to compute the order of a nilpotent subgroup
of GL(n, q) is implemented in Nilmat [4], and may be used in step (II) of SizeFiniteMatGroup-
FuncFF.
5. Implementation and performance
Implementations of our algorithms are publicly available in Magma. In this section we report on
their performance and dependence on the main input parameters: the degree n, the number of gen-
erators r, and size q of the coefficient field. We also investigated how runtimes vary with the degrees,
coefficients and number of summands of polynomials appearing in matrix entries.
The experiments reported in Table 1 were undertaken on a 3.0 GHz machine with 4 GB RAM
running Magma V2.15-10.
As tests, we chose groups with extremal properties, that pass through all stages of each algorithm.
The column ‘Runtime.1’ in Table 1 lists the CPU time in seconds of IsFiniteMatGroupFuncFF for
input G i j . The column ‘Runtime.2’ lists the time for IsFiniteNilpotentMatGroupFuncFF when
G i j is nilpotent. Note that the G i1 are finite and the G i2 are infinite for 1 i 4.
Polynomials in the matrix entries of G 1 j , G 2 j have degrees up to 1000, and many summands
with large coefficients. The G 1 j are absolutely irreducible: G 11 is a conjugate of GL(40, 57 ) in
GL(40, F57 ( X )), whereas G 12 is generated by G 11 and infinite order matrices in SL(40, F57 ( X )). Testing
each group necessitates computing an algebra basis of maximal size 402 = 1600 in Mat(40, 57 ). The
Table 1
Group n r q Runtime.1 Runtime.2

G 11 40 2 57 1646 –
G 12 40 10 57 1124 –
G 21 54 20 294 806 –
G 22 54 23 294 474 –
G 31 36 520 78 2506 113
G 32 36 522 78 252 20
G 41 100 1 312 423 16
G 42 100 1 312 8 4
Table 2
Group n r q Order Runtime

H1 20 3 17 20!280 33
H2 40 24 310 522 76 56
H3 24 16 72 34 54 73 233
H4 40 1 510 53 230
performance of both IsFiniteMatGroupFuncFF and IsFiniteCRMatGroupFuncFF is essen-

tially identical for this input.
The G 2 j have non-trivial unipotent normal subgroups, and so are not completely reducible. The
group G 21 is the Kronecker product of a conjugate of GL(6, 294 ) in GL(6, F294 ( X )) with a 10-generator
unipotent subgroup of GL(9, F294 ( X )). The group G 22 is generated by G 21 and infinite order matrices
of the form g ⊗ I 9 , where g is an upper triangular element of SL(6, F294 ( X )).
The G 3 j are nilpotent and not completely reducible. The group G 31 is the Kronecker product
of a 3-dimensional unipotent group with a 12-dimensional completely reducible nilpotent group
over F78 ( X ). Specifically, the latter group is a conjugate of a 2 × 2 block diagonal group, whose blocks
are a Sylow 3-subgroup and a Sylow 5-subgroup of SL(6, 78 ). The group G 32 is generated by G 31 and
infinite order diagonal matrices of the form g ⊗ I 18 , where g ∈ SL(2, F78 ( X )).
The G 4 j are cyclic. The group G 41 is generated by h1 ⊗ h, where h, h1 ∈ GL(10, F312 ( X )), h is
unipotent, and h1 is a conjugate of a randomly chosen 3 -element of GL(10, 312 ). Also G 42 = h2 ⊗ h
where h2 is a lower triangular element of SL(10, F312 ( X )). Comparison of the last two columns of
Table 1 for G 3 j and G 4 j demonstrates the superiority of IsFiniteNilpotentMatGroupFuncFF
for nilpotent input.
The performance of SizeFiniteMatGroupFuncFF depends on the algorithm used to deter-
mine the order of a matrix group defined over a finite field. Magma uses the (random) Schreier–Sims
algorithm [7, Section 7.8]. In Table 2 we report on using SizeFiniteMatGroupFuncFF to com-
pute the orders of the following groups over a univariate function field: H 1 is a conjugate of the full
monomial subgroup of GL(20, 17), H 2 and H 3 are nilpotent groups constructed in the same manner
as G 31 (H 2 but not H 3 is completely reducible), and H 4 is cyclic unipotent.
References
[1] L. Babai, R. Beals, D.N. Rockmore, Deciding finiteness of matrix groups in deterministic polynomial time, in: Proc. of Inter-
national Symposium on Symbolic and Algebraic Computation ISSAC ’93, ACM Press, 1993, pp. 117–126.
235–265.
[3] A.S. Detinko, On deciding finiteness for matrix groups over fields of positive characteristic, LMS J. Comput. Math. 4 (2001)
64–72 (electronic).
[4] A.S. Detinko, B. Eick, D.L. Flannery, Nilmat — Computing with nilpotent matrix groups. A refereed GAP 4 package, see http://
www.gap-system.org/Packages/nilmat.html, 2007.
[5] A.S. Detinko, D.L. Flannery, Algorithms for computing with nilpotent matrix groups over infinite domains, J. Symbolic
Comput. 43 (2008) 8–26.
[6] A.S. Detinko, D.L. Flannery, On deciding finiteness of matrix groups, J. Symbolic Comput. 44 (2009) 1037–1043.
[7] Derek F. Holt, Bettina Eick, Eamonn A. O’Brien, Handbook of Computational Group Theory, Chapman & Hall/CRC, London,
2005.
[8] B. Huppert, N. Blackburn, Finite Groups II, Springer, 1982.

[9] G. Ivanyos, Deciding finiteness for matrix semigroups over function fields over finite fields, Israel J. Math. 124 (2001)
185–188.
[10] D.N. Rockmore, K.-S. Tan, R. Beals, Deciding finiteness for matrix groups over function fields, Israel J. Math. 109 (1999)
93–116.
[11] D. Segal, Polycyclic Groups, Cambridge University Press, Cambridge, 1983.
[12] D.A. Suprunenko, Matrix Groups, Transl. Math. Monogr., vol. 45, Amer. Math. Soc., Providence, RI, 1976.
Journal of Algebra
Corrigendum
Corrigendum to “Irreducible (2, 3, 7)-subgroups of PGLn (F),

n 7” [J. Algebra 300 (1) (2006) 339–362]
M. Clara Tamburini a , M.A. Vsemirnov b,∗
a
Dipartimento di Matematica e Fisica, Università Cattolica del Sacro Cuore, Via dei Musei 41, 25121 Brescia, Italy
b
Steklov Institute of Mathematics, 27 Fontanka, Saint Petersburg 191023, Russia
a r t i c l e i n f o
Article history:
Received 6 February 2009
Available online 7 May 2009
1. Due to a misquotation of the main result from [3] the classification of conjugacy types of ir-
reducible (2, 3, 7)-generated subgroups of PSL5 (F) was stated incorrectly on pp. 343 and 352. For
a correction one has to redefine n5 . Namely, for p = 0, 5, 7, let n5 be the order of p modulo 5 times
the order of p 2 modulo 7, and let n5 = 4 if p = 7.
For p = 0, one can show that the matrices x, y on p. 352 are actually defined over a proper subfield
of Q(δ), where δ is a primitive 35th root of 1, namely, over the subfield generated by + −1 =
δ 5 + δ −5 and η = δ 7 .
For the convenience of the reader we also reproduce the correct list of conjugacy types of (2, 3, 7)-
groups depending on the residue of p modulo 35:
• H 5 H 2 , p = 2, 3;
• H 5 PSL5 ( p ), p ≡ 1, 6 (mod 35);
• H 5 PSL5 ( p 3 ), p ≡ 11, 16, 26, 31 (mod 35);
• H 5 PSU5 ( p 2 ), p ≡ 29, 34 (mod 35);
• H 5 PSU5 ( p 4 ), p ≡ 8, 13, 22, 27 (mod 35) or p = 7;
• H 5 PSU5 ( p 6 ), p ≡ 4, 9, 19, 24 (mod 35);
• H 5 PSU5 ( p 12 ), p ≡ 2, 3, 12, 17, 18, 23, 32, 33 (mod 35);
• p = 0, H 5 is defined over Q(δ 5 + δ −5 , δ 7 ), where δ is a primitive 35th root of 1, and fixes a non-
degenerate hermitian form.
DOI of original article: 10.1016/j.jalgebra.2006.02.030.

E-mail addresses: c.tamburini@dmf.unicatt.it (M.C. Tamburini), vsemir@pdmi.ras.ru (M.A. Vsemirnov).
doi:10.1016/j.jalgebra.2009.03.038
4162 M.C. Tamburini, M.A. Vsemirnov / Journal of Algebra 322 (2009) 4161–4162
For the unitary case we use the notation of [1], i.e., SUn (q2 ) SLn (q2 ) and preserves a hermitian
form associated with the automorphism α → α q of GF(q2 ).
2. The arguments used in the analysis of the members of class C5 in Theorems 6 and 8 should be
more accurate. Instead of considering y (respectively z) one also has to consider their scalar multiples.
Namely, in the proof of Theorem 6 the claim that the projective image of x, y is not contained
in a maximal subgroup from the class C5 follows from the fact that any scalar multiple of y of
determinant 1 has an invariant factor of the form t ± β i , i ∈ {1, 4, 7}. In a similar way, in Theorem 8
we use the fact that any scalar multiple of z of determinant 1 has an invariant factor of the form
t − γ i , i ∈ {1, 8, 15, 22, 29, 36, 43}.
3. In the proof of Theorem 6 we stated erroneously that the group 23 : PSL3 (2) is Hurwitz. In fact,
it is not even (2, 3)-generated. Thus, the analysis of the subgroups of M 22 can give the only Hurwitz
subgroup PSL3 (2) PSL2 (7), which has been already excluded by considering Schur multipliers. An-
other way to avoid the analysis of M 22 and its subgroups is to notice that this case may arise only for
p = 2 and M 22 PSU6 (4), while we have n6 = 6 for p = 2 and hence PSU6 (4) ∈ C5 .
Acknowledgments
We are grateful to the anonymous referee of our article [2] who pointed out the mistakes 1 and 3
above made in the previous paper.
References
[1] R.W. Carter, Simple Groups of Lie Type, Pure Appl. Math., vol. 28, John Wiley & Sons, 1972.
[2] M.C. Tamburini, M.A. Vsemirnov, Irreducible (2, 3, 7)-subgroups of PGLn (F), n 7, II, J. Algebra 321 (8) (2009) 2119–2138.
[3] M.C. Tamburini, A. Zalesski, Classical groups in dimension 5 which are Hurwitz, in: C.Y. Ho, P. Sin, P.H. Tiep, A. Turull (Eds.),
Proceedings of the Gainesville Conference on Finite Groups, de Gruyter, 2004, pp. 363–371.

Journal of Algebra V322 I11.2

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Journal of Algebra V322 I11.2

Transféré par

Droits d'auteur :

Formats disponibles

Journal of Algebra

Nicolás Andruskiewitsch Kazuhiko Kurano Aner Shalev

EDITOR OF THE SECTION ON COMPUTATIONAL ALGEBRA

SECTION ON COMPUTATIONAL ALGEBRA

Jon Carlson Meinolf Geck Gunter Malle

FOUNDING EDITOR: Graham Higman, 1964–1984

Contents lists available at ScienceDirect

Irreducibility criteria for skew polynomials

A skew-polynomial ring R [x; σ , δ] involves a not-necessarily-commutative ring R with unity, a ring

2. Background on skew polynomial rings

Let σ : R → R be an injective ring homomorphism. An additive group homomorphism δ : R → R is

δ(rs) = σ (r )δ(s) + δ(r )s, r, s ∈ R . (2.1)

with r j ∈ R for j = 0, . . . , n and rn = 0. We refer to deg( p ) := n 0 as the degree of p, and to the

SPR. 2. The degree function is additive, i.e., p , q ∈ P one has

deg( pq) = deg( p ) + deg(q).

SPR. 3. There is an injective ring homomorphism σ : R → R and a σ -derivation δ : R → R such that

xr = σ (r )x + δ r for all r ∈ R . (2.2)

Examples 2.1. Let R be a commutative integral domain.

least one r j = 0. Applying the sum to 1 we ﬁnd that r0 = 0, then successively to s1 , s2 , s3 , . . . , sn

as well as the recursion relation

S nj = δ S n−1, j + σ S n−1, j −1 (2.4)

where S ik b := S ik (b), which has the immediate consequence

3. Background on valuations and valuation rings

for all α ∈ Γ . We will follow this practice.

A valuation on R with values in Γ is a function ν : R → Γ ∪ {∞} satisfying the following properties

(V.1) ν assumes at least two distinct values;

When Γ is a cyclic subgroup of (R, +, 0) a valuation ν : R → Γ ∪ {∞} is said to be discrete.

f = p /q = p ( z)/q( z) ∈ C( z) ⇒ ord∞ ( f ) = deg(q) − deg( p ). (i)

Proposition 3.2. Any valuation ν : R → Γ ∪ {∞} on R has the following properties:

(a) ν (0) = ∞. In particular, ν (0) > γ for all γ ∈ Γ (by (V.3)).

• total if for any k ∈ K one either has k ∈ V or k−1 ∈ V (or both);

4. Extending valuations to R [x; σ , δ]

and is τ -compatible with δ , where τ ∈ Γ , if

ν (δa) ν (a) + τ for all a ∈ R . (4.2)

From (4.1) and induction (on i) we see that

and from (4.2) and induction that

r = ak ( z − z0 )k + ak+1 ( z − z0 )k+1 + · · · , ak = 0. (4.5)

Then ordz0 (r ) = k and one easily veriﬁes that

(c) ordz0 (r ) < 0 ⇒ ordz0 (r ) = ordz0 (r ) − 1.

For z0 = 0 and r = 1 + zn+1 with 1 n ∈ Z we have ordz0 (r ) = 0 and ordz0 (r

r = ak t k + ak+1 t k+1 + · · · , ak = 0, (4.7)

and we therefore have

For n 1 and r = 1 + t n we have ord∞ (r ) = 0 and r

(a) There is no real rational function r ∈ R(x) such that r

ν ( S ik b) ν (b) + (i − k)τ for all 0 k i . (i)

(Example: ν (δ 2 σ δ(b)) ν (σ δ(b)) + 2τ = ν (δ(b)) + 2τ ν (b) + τ + 2τ = ν (b) + 3τ = ν (b) + (4 − 1)τ .)

Corollary 4.3. For any a, b ∈ R one has

ν (aS ik b) ν (a) + ν (b) + (i − k)τ for all 0 k i .

(as in (2.6)). Then

Proof. From Lemma 4.2 one has

From (2.7) each ck is seen to be a sum of terms of the form

whereupon it follows from (3.1) that

and we therefore have

ν̃ ( pq) ν (ai0 ) + i 0 τ + ν (b j0 ) + j 0 τ . (vi)

Set k0 := i 0 + j 0 . Recalling the ﬁnal identity in (2.3) we then see that

5. An elementary application to skew polynomials

However, from the deﬁning property of the pole-type hypothesis we have

for all ak = 0, from which (ii) is easily seen to be equivalent to (i). 2

points (i , ν (ai ) + i τ ) and ( j , ν (a j ) + j τ ) of the (ν , τ )-carrier of L one has

Then L has no solution in R with a pole of type τ at ν .

We refer to (i + j , ν (r ) + ν (s) + (i + j )τ ) as the corner point of the indicated convex set.

Proof. From Lemma 4.2 we have

The carrier point corresponding to r ( S ik s)xk+ j is therefore

and from (ii) we see that

with r j ∈ R for j = 0, . . . , n and rn = 0. We refer to deg( p ) := n 0 as the degree of p, and to the

least one r j = 0. Applying the sum to 1 we ﬁnd that r0 = 0, then successively to s1 , s2 , s3 , . . . , sn

r = ak ( z − z0 )k + ak+1 ( z − z0 )k+1 + · · · , ak = 0. (4.5)

r = ak t k + ak+1 t k+1 + · · · , ak = 0, (4.7)

for all ak = 0, from which (ii) is easily seen to be equivalent to (i). 2

Example 8.4. In all the examples we take R = M(P1 ) C( z) and τ = −1.

am σ m (bn ) = 0. Since m is prime and c ∈ m \ m2 we have either am ∈ m and σ m (bn ) ∈/ m or am ∈/ m

χ ∈ Z is an inﬁnitesimal character, then there exists μ ∈ h∗ such that χ = χμ . If χ = χμ , we can