Number Theory

Paris 1992-93

R6dei-matrices and applications

Peter Stevenhagen

1. — Introduction
In this paper we describe an algebraic method to study the structure of
(parts of) class groups of abelian numberfields.The method goes back to the
Hungarian mathematician L. Redei, who used it to study the 2-primary part
of class groups of quadratic number fields in a series of papers [[18]-[24]]
that appeared between 1934 and 1953. The case of the /-primary part of the
class group of an arbitrary cyclic extension of prime degree / was studied
by Inaba [[12], 1940], who realized that one should look at the class group
as a module over the group ring. The matter was then taken up by Frohlich
[[6], 1954], who generalized Inaba's results by extending Redei's quadratic
method to the case of a cyclic field of prime power degree. In the seventies,
generalizations in the line of Inaba were given by G. Gras [[ 10]]. In all cases,
one studies /-primary parts of the class group of an abelian extension for
primes / that divide the degree.
Recently, completely different methods have been developed by Koly-
vagin and Rubin, showing that the structure of any /-primary part of the
class group of an abelian field of degree coprime to / can be described 'al-
gebraically'. For primes dividing the degree it is not yet clear whether the
approach works. The Kolyvagin-Rubin methods can be seen as refinements
of the analytic class number formula, and they are more general than the
Redei-Frohlich method as they work for most /. On the other hand, they
depend on the existence of infinite collections of auxiliary prime numbers,
so effective versions of the Cebotarev density theorem are needed to yield
deterministic algorithms. Because of this somewhat involved nature they
can only be used in practice for abelian fields of very small degree. More-
over, the method does not give any clue as to the average behaviour that is
to be expected when it is applied to a family of fields. For instance, it cannot
be used to compute the class number h+ of the maximal real subfield of the
n-th cyclotomic field for any n that is not very small. Also, it does not tell us

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 246 P. STEVENHAGEN

whether the fact that h^ is either 1 or very small, which can be shown to be
the case for all n < 200 having at least two prime divisors if one assumes
the generalized Riemann hypothesis [[31]], should be seen as a common
occurrence.
The Redei-Frohlich method is based only on class field theory and
therefore of a rather different nature. It can be used only to describe p-
primary parts of the class group when p does divide the degree of the
extension, which is exactly the case that had to be excluded before. We
will see that it gives in this case rise to density statements telling us how
many fields in a given infinite family will have some prescribed part in
their class group. An example : the class number hf3p is even for some
explicit collection of primes p of Dirichlet density 1/16 and divisible by 3 for
a collection of density 1/18.
An application of Redei's quadratic method that goes back to Redei
himself concerns criteria for the solvability of the negative Pell equation
x2 — Dy2 = — 1. This is a question that is closely related to the behaviour of
2-class fields, as was made clear by Scholz [[25]]. We will discuss it in the
two final sections of this paper.

2. — R€dei-matrices
In this section, K will denote a numberfieldthat is cyclic of prime degree
/ with Galois group G = Gal(if/Q). It is our intention to study the /-part
C of the class group of K. If / = 2, our convention will be that C is the 2-
part of the narrow class group of K. Correspondingly, we call an extension
unramified if all finite primes are unramified. The difference between the
narrow and the ordinary class group that may exist is of interest only when
Z = 2. It will be discussed in detail the following two sections.
As C is a finite abelian /-group with a natural G-action, it is a module
over the group ring over the /-adic integers Zi[G], The norm N = Y^geG9
annihilates the class group, so we can study C as a module over Zi[G]/N.
If (i denotes a primitive /-th root of unity, we have an isomorphism

Zi[G\/N-^

showing that A = Zi [G]/N is a discrete valuation ring whose maximal ideal

is generated by a — 1, with a a generator of G. The residue class field of A
is the finite field of/ elements F/. Every finite ^-module M is isomorphic to
a module of the form

with rii e Z>i for i = 1,2,..., s. Thus, we can specify the isomorphism

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 REDEI-MATRICES AND APPLICATIONS 247

class of the A-module M by giving the sequence of integers

rk = #{i : m > k} =

Note that r\{M) = s and that {rk(M)}k is a decreasing sequence with

rfc(M) = 0 for k sufficiently large.
The evaluation of r\{C) amounts to doing genus theory for the field K.
More precisely, class field theory associates to C an unramified extension
H of K, called the /-class field of Ky for which the Galois group Gal(H/K)
is canonically isomorphic to C. The quotient C/C<r~1 corresponds to an
unramified extension H\ of K that is known as the genus field of K. It is the
maximal unramified extension of K that is abelian over Q, and Gal(i7i/Q)
is isomorphic to the elementary abelian /-group G x C/C<7~1, If x denotes a
Dirichlet character generating the character group X of G that corresponds
to X, we can write x as a product x = 11!= 1 X;> where t is the number of
primes that ramifies in the extension K/Q and X; is a character of conductor
a power of some ramifying prime pi and of order /. The conductor of \% i s
equal to pi if and only if pi ^ /. The field H\ corresponds to the group of
Dirichlet characters

Xi = f[{Xi)-
i=l
It follows that Gal(Hi/K) has order ll~l, i.e. the (a - l)-rank n (C) is equal
to t — 1, where t is the number of ramifying primes in K/Q.
The subgroup CG = C[a — 1] of G-invariant ideal classes in C is known
as the subgroup of ambiguous ideal classes. As G is cyclic and C is finite,
the order of CG = H°(G, C) equals the order of Hl{G, C) = C/C^'1. It is
not difficult to check that CG is generated by the t classes [pi] of the ramified
primes pi of K. The order of CG is / t - 1 , so there is exactly one additional
relation between these classes that is independent of the obvious relations
[p{] = o.
The Redei-Frohlich theorem gives a description of r<i (C) by combining
the two descriptions of ri(C). Note that as abelian groups, we have :

IC/C*'1 x c-VC^" 1 ) 2 if / > 2,

and that the /-rank of C/Cl is equal to the sum Y^iZ^i r^- T n e theorem
is based on the observation that r2{C) can be obtained from an explicit
description of the natural map

: C[a - 1] —-> C/Ca-l

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 248 P. STEVENHAGEN

This is a homomorphism between elementary abelian /-groups, so it can be

viewed as a linear map between vector spaces over Fj. With this terminology,
the (a — l) 2 -rank of C is nothing but the F/-dimension of the kernel of </>.
This dimension can be given in terms of the rank of a certain matrix over
Fj, called the Redei matrix of K, as follows.

1. THEOREM (Redei-Frohlich). — Let K/Q be a cyclic extension of prime

degree I with group (a) and conductor / , and let x • (Z//Z)* —• Fj be a
generator of its character group, with values taken in the additive group Fj.
Letpi,p2, - - ,Pt be the primefactors off f andx = Y^\=i Xi the corresponding
decomposition ofx- Then the I-primary part C of the narrow class group of
K has (cr — l)2-rank
r2(C) =t- l-rankF/#,
where the entries a,ij e Fj of the Redei matrixR = (aijYij=1 are defined by :

if
aij =Xi(Pj) i^j]
t

Proof : as C[a — 1] is generated by the classes of the ramified primes

pi, we have a natural surjection p : Fj —> C[a — 1] that maps the j-th basis
vector ej to the class of pj. The group C/C*'1 is canonically isomorphic
to the subgroup Gal(Hi/K) of Gal(Jfi/Q) under the Artin map. We know
H\ explicitly from genus theory : it is the compositum of the cyclic fields
Q(Xt) of conductor a power of pi corresponding to the characters x%- Each
character Xi furnishes an isomorphism Gal(Q(xi)/Q) —* F;, and they can
be combined into an isomorphism 0*=1Xt • Gal(ffi/Q) ^ ^ Fj. The Redei
map R : F* —> F* is defined as the composed map

R : Fj -fU C[a - 1] - ^ C/C^1 ^U Gal(Hi/K) C Gal(JTi/Q) ^ Fj

of vector spaces over F/. As the kernel of p is of dimension 1, one has

(2) r 2 (C) = dimFz [ker (j>) = dimFi [ker R]-l=t-l- rankFi R,

as desired. The image of a basis vector ej is the Artin symbol of pj in

Gal(Hi/K). If i ^ j , the restriction of this symbol to Gal(Q(Xt)/Q) is the
Artin symbol of pj, and this is mapped to a,ij = x% (Pj) by x% • F o r the diagonal
entry an the Artin symbol and Xi (Pi) are not defined, but we can use the fact
that 0Xi maps Gal(Hi/K) to the hyperplane {(a.), e Fj : Y,\=iai = °}-
The desired identity J2l=i aij = ® follows immediately. •

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 REDEI-MATRICES AND APPLICATIONS 249

The Redei matrix R is by definition a singular (t x t) -matrix since the sum

of its rows is zero. It is said to have maximal rank if the rank equals t — 1.
Obviously, the rank is maximal if and only if T2{C) = 0. We will meet
this condition in the next section when investigating the solvability of the
negative Pell equation.
The field H2 corresponding to the quotient C/C^"~1)2 is the central /-class
field of Ky i.e. the largest unramified extension E of Hi that is normal over
<Q> and for which the group extension
0 —> Gal(£/#i) —> Gal(£/Q) —> Gal(ffi/Q) —> 0
is a central extension. In Frohlich's terminology [[711, the central class field
Hi is a field of class two : its Galois group ft = Gal(#2/Q) is not in
general abelian but its lower central series has length at most two. This
is equivalent to saying that the commutator subgroup [ft, ft] is contained in
the center of ft, or that [ft, [ft, ft]] = 0. The Redei-Frohlich method enables
us to obtain Gal(H2/K) from very simple rational data. More precisely, we
can determine T{ = ri(C) for i = 1,2 in terms of the prime factors of the
discriminant, and this leads to

„ f (Z/2Z) ri " r2 x (Z/4Z) r2 if / = 2;

" \(Z//Z)ri+ra if/ > 2.
The first isomorphism is an isomorphism of modules over the ring A =
Zi[G\/N, the second is an isomorphism of abelian groups.

3. — Applications
As a first application, we will obtain divisibility results for the real
cyclotomic class numbers /i+ of the type discussed in the introduction.
Recall that h£ is the class number of the maximal real subfield Fn =
Q(Cn + Cn1) of the cyclotomic field of conductor n.

3. LEMMA. — Suppose that I > 2 and that the l-class group C of the field
K in theorem 1 has (a — I)2-rankr2{C) = r. Then lr divides h+, with n the
conductor ofK.
Proof: as K is real of conductor n, it is contained in Fn. The genus field
Hi of K is equal to H C\Fn, so H2Fn/Fn is an unramified abelian extension
of Fn of degree [H2 : Hi] = lr. This degree divides /i+ by class field theory,
so we are done. •
This lemma provides us with an easy method of constructing infinitely many
n for which /i+ is divisible by an arbitrarily high power of a prime number

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 250 P. STEVENHAGEN

/. One simply takes those n for which Fn contains a subfield K of degree /

over Q for which the Redei-matrix is of rank much smaller than t — 1. By
taking it equal to the zero matrix, the following result is obtained.

4. THEOREM. — Letn be divisible by t distinct primes congruent to 1 mod /

such that each of these primes is an l-th power modulo ail others. Then Z t-1
divides h^. •
For fixed tt there are infinitely many pairwise coprime n satisfying the
hypothesis of the theorem. This follows easily from Dirichlet's theorem on
primes in arithmetic progressions. If t — 1 primes congruent to 1 mod / have
been chosen such that each is an l-th power modulo the others, the t-th
prime that makes the hypothesis of the theorem hold true can be chosen
from an infinite collection of Dirichlet density [(/ - l)/ 2 ^" 1 )]" 1 . This follows
from the Cebotarev density theorem, as the condition on the t-th prime is
that for each of the t — 1 previous primes p, it splits completely in the field Ep
that is obtained by adjoining an l-th root of unity (i and tfp to the subfield
of degree / in the p-th cyclotomicfield.The fields Ep are all of degree (/ — I)/ 2
over Q, and they are linearly disjoint over Q((z). As a very special case, we
obtain the claim made in the introduction that hf3p is divisible by 3 for a
set of primes p of Dirichlet density 1/18.
For odd /, results similar to those in the preceding theorem have been
proved by Cornell and Rosen [[3], 1984] using cohomological methods that
go back to Furuta [[9]]. For t = 2 and t = 3 their results are identical to
those following from lemma 3, for large t their method is better. Neither
method gives any result for the prime conductor case t = 1.
For / = 2, the lemma and the arguments given above have to be adapted
for several reasons. First of all one has to take care of the ramification of real
primes. This leads one to consider only those quadratic fields K that have
real genus fields, i.e. real quadratic fields for which all odd prime divisors of
the discriminant are congruent to 1 modulo 4. One only obtains a divisibility
result 2r~1|/i+, which is weaker than lemma 3. In order to find 2r|/i+ one
needs to show that H2 is real. This can sometimes be done using a method
of Scholz [[25]]. Secondly, one has to adapt the density computation given
above as the fields Ep have smaller degree, a statement equivalent to the
quadratic reciprocity law. Rather than working out all details here, we give a
characteristic example. It is stronger than the cohomological result in [[3]].

5. THEOREM. — Let p and q be primes congruent to 1 moduio 4 that

are mutual quadratic residues, and suppose that the fourth power residue
symbols (^) and (^) are equal Then hp~q is even. •
It is not difficult to see that we obtain the claim in the introduction for
q = 13. For n = pq = 5 • 29 = 145 it follows that hf45 is even. This is one

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 REDEI-MATRICES AND APPLICATIONS 251

of the two values n < 200 with n not a prime power for which /i+ > 1. In
fact, one can use Odlyzko's discriminant minorations to show [[31]] that
h+45 = 2.
There are no results of a similar algebraic nature in the prime conductor
case t = 1, and we cannot produce infinite families of primes p for which
h+ is even. See [[28]] for a more complete discussion.
A second application of the technique of Redei matrices arises in the study
of the solvability in integers of the negative Pell equation x2 — Dy2 = — 1,
where D > 1 is a squarefree integer. With CD a fundamental unit in Q(y/D)
and N the norm to Q one has

x1 - Dy2 = - 1 is solvable in integers <=> NCD = — 1.

Indeed, if the equation is solvable there are units of norm —1, so the
fundamental unit cannot have norm -h 1. Conversely, if Nen = — 1 it may be
that en is not in Z[y/D]t but as its cube e3D always is we still get an integral
solution to the equation. As it is more natural to work with discriminants
than radicands, we will further take D to be a quadratic discriminant
and say that the negative Pell equation is solvable for D if the equation
x2 — Dy2 = —4 has integral solutions. If the equation is solvable for JD,
then D is positive and —1 is a quadratic residue modulo every prime divisor
of Dy so D must be in the set V of real quadratic discriminants that are
not divisible by any prime congruent to 3 mod 4. A question that has been
studied by many people but that is still completely open is the following.

6. PROBLEM. — LetV(-l) cV be the set of real quadratic discriminants

for which the negative Pell equation is solvable. Decide whether the limit

exists and if so, determine it

This is a very hard problem, and to my knowledge is is not even known
whether the liminf and the limsup of this expression are in the open interval
(0,1).
The relation between the solvability of the negative Pell equation and the
previous section is given by the following immediate consequence of class
field theory.

7. LEMMA. — The negative Pell equation is solvable for a quadratic

discriminant D if and only if the narrow 2-Hilbert class field ofQ(y/D) is
real

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 252 P. STEVENHAGEN

Proof : both statements are equivalent to the fact that Q(y/D) is a real
field for which the narrow Hilbert class field coincides with the ordinary
Hilbert class field. •
Let H be the narrow 2-Hilbert class field of K = Q(VD). This is the
situation of the preceding section, with / equal to 2. From the lemma, we
see that the negative Pell equation is solvable for D if and only Hk is real
for all k > 1. The condition that the genus field Hi is real is equivalent to
the requirement that D is in V, since H\ is obtained from K by adjoining
a square root of (—l)^" 1 )/ 2 for each odd prime divisor p of D. If H = Hi
this condition is also sufficient for solvability of the negative Pell equation.
8. LEMMA. — The negative PeU. equation is solvable for D e V if the Redei
matrix ofQ(VD) has maximal rank.
Proof : the condition implies an equality Hi = H<i, so H = Hi is real
and we are done by the previous lemma. •
If D e V has t distinct prime divisors, the corresponding Redei matrix R is
by the quadratic reciprocity law a symmetric (t x t)-matrix over F 2 whose
rows and columns add up to zero. Let Rr be the (t — 1) x (t — l)-minor
obtained by leaving out the last row and column from R. If D ranges over
the subset Vt of V consisting of those discriminants that have exactly t
distinct prime divisors, it is intuitively clear that the corresponding Redei
minor R'D behaves like a random symmetric (t — 1) x (t — 1)-matrix over
F 2 , i.e. that
#{DeVt:D<X and R'D = S}
oo #{DeVt:D< X}
exists and does not depend on the choice of the symmetric matrix S. The
statement is a reformulation of the fact that the vector consisting of the Q)
Legendre symbols ( £i ) of an element D = pip2 ... pt is randomly distributed
as a function on Vt. The details of a correct proof are not trivial. Redei's
original proof [[22]] proceeds by induction on t, and so does the proof of the
rediscovery of the result in [[5]]. There is also an easy way out by adapting
the notion of density [[17]].
Once one knows that a discriminant D e Vt gives rise to a random
symmetric matrix, one can determine how likely it is that such a matrix is
non-singular. We give a slightly more general result for future reference.
9. PROPOSITION. — Let n > 1 be an integer and q a prime power Then
there are
(t1) l<k<n
fcodd

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 REDEI-MATRICES AND APPLICATIONS 253

symmetric (n x n)-matrices over the field of q elements ¥q that are non-

singular. The number of matrices of arbitrary rank r e {0,1,... n} equals
[™] Ar(q), where [™] denotes the number of r -dimensional subspaces of a
vector space of dimension n over ¥q.
Proof : the result for q — 2 occurs in rather cumbersome terminology
and with a lengthy proof in [[22]]. A completely elementary proof by induc-
tion on n can be found in [[13]]. In order to see that the statement given
there is identical to ours one needs the explicit value

The first half of the proposition immediately implies the second half, as
symmetric matrices correspond bijectively to symmetric bilinear forms and
giving a symmetric bilinear form of rank r on V = F™ is equivalent to giving
a subspace W C V of dimension n — r and a non-degenerate symmetric
bilinear form of the factor space V/W. This remark also shows that the
numbers An (q) can be computed inductively from the relation

r=0

so it suffices to check that the given expression satisfies this relation. An

elegant way of doing this is given in [[511. •
We will only use the preceding proposition for q = 2 and n = t — 1, so we
write At-i(2) = 2®at with
[t/2]

Set Vt(-l) =Vtn V(-l). We now know that for fixed t, we have a lower
density for Vt{— 1) in Vtf since the two preceding lemmas imply

The numerical value a^ = .4194224... is already in [[22]]. The density

result has been reproved in [[17]], [[11]] and [[5]]. The formulation given by
these authors is different from Redei's, as they interpret the Redei matrix
as an incidence matrix of a graph on t points.

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 254 P. STEVENHAGEN

The equations V = Ut>i©t and V(-l) = \Jt>\Dt{-\) make it very

plausible that the lower density of V(-l) in V is not smaller than a^.
However, I do not know how to prove this. The problem is that each Vt is
a subset of zero density in Vy and it seems non-trivial to prove a density
result for V from a density result for each of the subsets Vt.
The preceding argument can be further refined in order to obtain a still
higher value of the lower density of Vt(-l) in Vt. In particular, we will push
the limit value for t —> oo over the value 1/2 that has been suggested as a
possible value [[17]]. However, we need to pass to the next higher level, i.e.
the field H3, in order to do this.

4. — Higher levels
In principle, the Redei-Frohlich method for determining T2{C) can be
extended to determine inductively all values rk(C). Having defined the first
Redei map
R = Rt : Ff — > C [ a - l ] - ^ C/C^'1 >-* ¥j,
one can repeat the procedure and consider the higher Redei maps

Rk. ker ^ _ > C[a - 1 ] n c ^ - 1 ) " 1 - ^

Just as in the case k = 1, we obtain the (a - l) fc+1 -rank from this map by

rk+i(C) = dim Fi [ker Rk] - 1 = rk(C) -

which is the analogue of (2). Despite the close analogy, a serious compli-
cation arises for these higher levels. For k = 1, we were able to embed
C(^-!) /C(a~^ in a canonical way in a vector space of dimension t over
F/. This was due to the fact that we could describe the genus field Hi very
explicitly in terms of Dirichlet characters. The fields Hk for k > 2 are no
longer abelian over Q, and no general method is known to describe them
explicitly. This is a serious drawback that accounts for the fact that there
is no generalisation of theorem 1 to higher levels that is of a comparable
simplicity. For the same reason, we do not have general density results for
these levels that resemble those in the preceding section.
Only in the special case where k = I = 2, there is a more explicit version
of the theory that goes back to Redei [[22]] and was further developed by
Frohlich [[71]. We can formulate it in modern terms as follows.
Let D be a quadratic discriminant, and D = 11*=i &% i t s factorization
into prime power discriminants. The set V of discriminantal divisors of
D is defined as the set of divisors d of D of the form d = n L i ^V "^^
ei e {0,1}. This is in a natural way a vector space of dimension t over F2
with a canonical basis consisting of the divisors d{. The natural surjection

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 REDEI-MATRICES AND APPLICATIONS 255

V —> C[2] maps di to the class of the ramified prime d{ of K that divides
di. As the genus field H\ of K = Q(\/D) is generated over Q by the square
roots \fd{, Kummer theory tells us that the Galois group Gal(ifi/Q) can
be seen as the dual space V* = Hom(V, F2) of V. The kernel of the Redei
map R\ : V —> V* consists of divisors d e V for which the associated Artin
symbol <J0 e Gal(H/K) is the identity on Hi. The kernel of the dual map
R* : V = V** -> V* consists of those d € V for which \fd is left invariant
by the Artin symbols of all ideals that have order 2 in the class group. Note
that D itself is always in this kernel.
A decomposition D = Di • J92 with Di e ker Rl is called a decomposition
of the second kind. These decompositions are characterized by the fact Di
is a square modulo all prime divisors of D 2 and vice versa. The prime 2
needs special attention here. Given a decomposition D = Di • D2 that
is of the second kind, Redei explicitly constructs a quadratic extension of
Q(>/S7, y/lh) that is cyclic of degree 4 and unramified over K. This is
possible since the equation x1 — Diy2 — D2Z2 = 0 has non-zero rational
solutions by Legendre's theorem and the assumption on the decomposition.
For a primitive integral solution (#, y, z) with well-chosen 2-adic behavior,
the extension E that is generated over K by a square root 7 ^ of x 4- yy/Di
has the desired properties. The extension E/K depends on the choice of
the solution, but the quadratic extension EHi = Hi (JD1 ) of Hi does not.
Every element a e Gal(H2/Hi) is determined by its action on the elements
7£)x for Di e keriJJ, so we can view this Galois group as a subspace of
Hom(keri?J,F2) = (keriJJ)*. With these identifications, we can describe
the second Redei map :

R2 : ker#! —• C[2] n C2 —> C 2 /C 4 - ^ Gal(ff2/-ffi) C (ker/JJ)*

explicitly as an F2 -linear map between vector spaces of dimension 1+r 2 (C).

The 8-rank r3(C) of the narrow class group of K is given by the formula

which is non-negative as i?2 is always singular.

As soon as one chooses a basis for ker/^i and for the space kerfi*
of decompositions of the second kind, R2 is given by a matrix whose
entries describe the action of the Artin symbols <JV coming from d e ker Ri
on explicit elements 7^ with d! e kerit^. Note the equality Ri = R\
in case D is in the set V of discriminants that are of interest for the
negative Pell equation. The entries of i?2 are quadratic symbols of quadratic
irrationals and can be computed rather easily. Redei's paper [[22]] has
numerous identities that express these 'new number theoretic symbols',
as he calls them, in rational terms, and Frohlich [[8]] does the same in a

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 256 P. STEVENHAGEN

more systematic way. However, these expressions are usually given in terms
of the chosen solution of x2 — D\y2 — D2z2 = 0, and this makes it difficult
to obtain density results in terms of the prime divisors of D. Special cases
have been dealt with by Morton [[14H16]], and density statements for the
behavior of C/C8 have been proved by the author [[26]] in the case that
t — 1 prime divisors of the discriminant are fixed and the last one varies.
For t = 2 this yields results that had been known for some time.
In the previous section, we showed that the negative Pell equation is solvable
for all D i n a subset of density at of Vt since these D have r2(C) = 0.
Following an idea that goes back to Redei [[21]] and Scholz [[25]], we can
use the 8-rank theory to enlarge this set even further by looking at those D
that have r2{C) = 1. We will indicate briefly how this is done.
The density (3t of the set of D e Vt having r2{C) = 1 follows easily
from proposition 9. One has (3t = <*t if t is even and (3t = {I — 21~t)at if t
is odd. Note that lim^oo fit = lim^oo at = aoo- For D as above, there is
exactly one non-trivial decomposition D = D\D2y and one can show that
the higher Redei matrix R2 equals

R2 =

which has to be interpreted in the obvious way as a matrix over F2. As the
biquadratic residue symbols have value ±1, there are 4 possible values for
this matrix, and they each occur for a set of D that has density j(3t in Vt.
If ( ^ ) 4 and ( ^ ) 4 are both equal to 1, the matrix R2 is the zero matrix and
H2 is strictly smaller than the 2-Hilbert class field H. In all other cases,
its rank is one and H = H2. We can determine whether H2 is real by a
generalization of the argument used in proving theorem 5. One has :

It follows that H = H2 is totally complex and the Pell equation is not

solvable if the biquadratic residue symbols are not equal, which happens
for a collection of discriminants of density /3t/2 in Vt. If both symbols equal
— 1, the Pell equation is solvable, and this happens for a set of density
/?t/4. Taking together the two collections of D for which the Pell equation is
solvable, we conclude that for each fixed t, the set Vt(—1) has lower density
ott + jfit and upper density l — ^-pt inside Vt. For increasing t these values
rapidly converge to :

f-aoo = .52428...
1 - 5 •<*«> = .79029...

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 REDEI-MATRICES AND APPLICATIONS 257

It remains a challenging problem to deduce any non-trivial density result

forD(-l) in V,
Hardly anything is known about the distribution of the ranks rk{C) when
k > 4. Some numerical data are available in the quadratic case [[2],
[27]], mainly for cyclic C. The best known example is probably that of the
quadratic field QCv^-p), with p a prime congruent to 1 mod 4. In this case
the discriminant D = — 4p has two prime divisors and C is a non-trivial
cyclic 2-group. It follows from theorem 1 that the order of C is divisible by
4 exactly when p = mod8, and the 8-rank results quoted above imply that
the order is divisible by 8 if and only if p splits completely in Q(Cs, v7! 4-z)>
which is a non-abelian field of degree 8 over Q. All numerical evidence
suggests strongly that the order of C is divisible by 16 for a set of primes of
density 1/16, but the existing techniques do not even suffice to show that
this happens infinitely often. The question is closely related to the 2-adic
behavior of the fundamental unit ep in the field Q(y/p), see [[27]].

Note added in proof

It is now conjectured that the limit value in problem 6 exists and equals
1 - tfoo = .5805775582 ..., with a^ as in section 3, see [[29]]. The heuristics
can be extended to the case of quadratic orders [[30]]. They have been
confirmed by extensive computer calculations [[11].
Manuscrit regu le 7 septembre 1994

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 258 P. STEVENHAGEN

References
[1] W. BOSMA, P. STEVENHAGEN. — Density computations for real quadratic
units, preprint (1994).
[2] H. COHN, J.C. LAGARIAS. — On the existence of fields governing the 2-
invariants of the class group of Q(y/dp) as p varies, Math. Comp. 41,
711-730 (1983).
[3] G. CORNELL, M.I. ROSEN. — The £-rank of the real class group ofcyclotOTnic
fields, Compositio Math. 53, 133-141 (1984).
[5] J. E. CREMONA, R.W.K. ODONI. — Some density results for negative Pell
equations; an application of graph theory, J. London Math. Soc. (2) 39,
16-28 (1989).
[6] A. FROHLICH. — The generalization of a theorem ofL. RedeVs, Quart. J.
Math. Oxford (2) 5, 130-140 (1954).
[7] A. FROHLICH. — On fields of class two, Proc. Lond. Math. Soc. (3) 4,
235-256 (1954).
[8] A. FROHLICH. — A prime decomposition symbol for certain non Abelian
number fields, Acta Sci. Math. 21, 229-246 (1960).
[9] Y. FURUTA. — On class field towers and the rank of ideal class groups,
Nagoya Math. J. 48, 147-157 (1972).
[10] G. GRAS. — Sur les l-classes d*ideaux dans les extensions cycliques
relatives de degre premier I, Ann. Inst. Fourier, Grenoble 23,3, 1-48
(1973).
[11] J. HURRELBRINK. — On the norm of the fundamental unit, preprint,
Louisiana State University (1990).
[12] E. INABA. — Uber die Struktur der l-Klassengruppe zyklischer Zahlkorper
vom Primzahlgrad I, J. Fac. Sci. Imp. Univ. Tokyo, section I, vol. IV 2,
61-115(1940).
[13] J. MACWILLIAMS. — Orthogonal matrices over finite fields, Amer. Math.
Monthly 76, 152-164(1969).
[ 14] P. MORTON. — Density results for the 2-classgroups of imaginary quadra-
tic fields, J. reine angew. Math. 332, 156-187 (1982).
[15] P. MORTON. — Density results for the 2-classgroups and fundamental
units of real quadratic fields, Studia Scientiarum Math. Hungarica 17,
21-43(1982).
[16] P. MORTON. — The quadratic number fields with cyclic 2-class groups,
Pac. J. Math. 108, 165-175 (1983).

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

 REDEI-MATRICES AND APPLICATIONS 259

[17] R. V. PERLIS. — On the density of fields with N(e) = - 1 , preprint,

Louisiana State University (1990).
[18] L. REDEI, H. REICHARDT. — Die Anzahl der durch 4 teilbaren Invarianten
der Klassengruppe ein.es beliebigen quadraUschen Zahlkorpers, J. reine
angew. Math. 170, 69-74 (1934).
[19] L. REDEI. — Arithmetischer Beweis des Satzes uber die Anzahl der durch
vier teilbaren Invarianten der absoluten Klassengruppe im quadratLschen
Zahlkorper, J. reine angew. Math. 171, 55-60 (1935).
[20] L. REDEI. — Uber die Grundeinheit und die durch 8 teilbaren Inva-
rianten der absoluten Klassengruppe im quadratischen Zahlkorper; J.
reine angew. Math. 171, 131-148 (1935).
[21] L. REDEI. — Uber einige Mittelwertfragen im quadratischen Zahlkorper,
J. reine angew. Math. 174, 131-148 (1936).
[22] L. REDEI. — Ein neues zahlentheorettsches Symbol mitAnwendungen auf
die Theorie der quadratischen Zahlkorper, J. reine angew. Math. 180, 1-
43 (1939).
[23] L. REDEI. — Bedingtes Artinsches Symbol mit Anwendungen in der
Klassenkorperfheorie, Acta Math. Acad. Sci. Hung. 4, 1-29 (1953).
[24] L. REDEI. — Die 2-Ringklassengruppe des quadratischen Zahlkorpers und
die Theorie derPellschen Gleichung, Acta Math. Acad. Sci. Hung. 4, 3 1 -
87 (1953).
[25] A. SCHOLZ. — Uber die Losbarkeit der Gleichung t1 — Du2 = — 4, Math.
Zeitschrift 39
[26] P. STEVENHAGEN. — Class groups and governing fields, Publ. Math. Fac.
Sci. Besangon, annee 1989/90, 1-94 (1990).
[27] P. STEVENHAGEN. — On the 2-power divisibility of certain quadratic class
numbers, J. Number Theory 43 (1), 1-19 (1993).
[28] P. STEVENHAGEN. — Class number parity for the p-th cyclotomic field,
Math. Comp. 63 no. 208 (to appear, 1994).
[29] P. STEVENHAGEN. — The number of real quadratic fields having units of
negative norm, Exp. Math. 2 (2), 121-136 (1993).
[30] P. STEVENHAGEN. — Frobenius distributions for real quadratic orders, J.
Theorie des Nombres Bordeaux (to appear, 1995).
[31] F.VAN DER LINDEN. — Class number computations of real abelian number
fields, Math. Comp. 39, 693-707 (1982).

Peter Stevenhagen
Faculteit Wiskunde en Informatica
Plantage Muidergracht 24
1018 TV Amsterdam, Netherlands
e-mail: psh0fwi.uva.nl

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press

https://doi.org/10.1017/CBO9780511661990.015 Published online by Cambridge University Press