A Paley Wiener Theorem For The Inverse S PDF

PACIFIC JOURNAL OF MATHEMATICS
Pacific Journal of Mathematics

Volume 193 No. 1 March 2000 Pacific
Journal of
Mathematics
2000
Vol. 193, No. 1
Volume 193 No. 1 March 2000

http://www.pjmath.org
Founded in 1951 by
E. F. Beckenbach (1906–1982) F. Wolf (1904–1989)
EDITORS
V. S. Varadarajan (Managing Editor)
Department of Mathematics
University of California
Los Angeles, CA 90095-1555
pacific@math.ucla.edu
Vyjayanthi Chari Darren Long Sorin Popa
Department of Mathematics Department of Mathematics Department of Mathematics
University of California University of California University of California
Riverside, CA 92521-0135 Santa Barbara, CA 93106-3080 Los Angeles, CA 90095-1555
chari@math.ucr.edu long@math.ucsb.edu popa@math.ucla.edu
Robert Finn Jiang-Hua Lu Jie Qing
Stanford University The University of Hong Kong University of California
Stanford, CA 94305-2125 Pokfulam Rd., Hong Kong Santa Cruz, CA 95064
finn@math.stanford.edu jhlu@maths.hku.hk qing@cats.ucsc.edu
Kefeng Liu Sorin Popa Jonathan Rogawski
University of California University of California University of California
Los Angeles, CA 90095-1555 Los Angeles, CA 90095-1555 Los Angeles, CA 90095-1555
liu@math.ucla.edu popa@math.ucla.edu jonr@math.ucla.edu
PRODUCTION
pacific@math.berkeley.edu
Paulo Ney de Souza, Production Manager Silvio Levy, Senior Production Editor Nicholas Jackson, Production Editor
SUPPORTING INSTITUTIONS
ACADEMIA SINICA , TAIPEI UNIV. OF CALIF., SANTA BARBARA
CALIFORNIA INST. OF TECHNOLOGY UNIVERSIDAD DE LOS ANDES UNIV. OF CALIF., SANTA CRUZ
CHINESE UNIV. OF HONG KONG UNIV. OF ARIZONA UNIV. OF HAWAII
INST. DE MATEMÁTICA PURA E APLICADA UNIV. OF BRITISH COLUMBIA UNIV. OF MONTANA
KEIO UNIVERSITY UNIV. OF CALIFORNIA , BERKELEY UNIV. OF NEVADA , RENO
MATH . SCIENCES RESEARCH INSTITUTE UNIV. OF CALIFORNIA , DAVIS UNIV. OF OREGON
NEW MEXICO STATE UNIV. UNIV. OF CALIFORNIA , IRVINE UNIV. OF SOUTHERN CALIFORNIA
OREGON STATE UNIV. UNIV. OF CALIFORNIA , LOS ANGELES UNIV. OF UTAH
PEKING UNIVERSITY UNIV. OF CALIFORNIA , RIVERSIDE UNIV. OF WASHINGTON
STANFORD UNIVERSITY UNIV. OF CALIFORNIA , SAN DIEGO WASHINGTON STATE UNIVERSITY
These supporting institutions contribute to the cost of publication of this Journal, but they are not owners or publishers and have no respon-
sibility for its contents or policies.
See inside back cover or www.pjmath.org for submission instructions.
Regular subscription rate for 2006: $425.00 a year (10 issues). Special rate: $212.50 a year to individual members of supporting institutions.
Subscriptions, requests for back issues from the last three years and changes of subscribers address should be sent to Pacific Journal of
Mathematics, P.O. Box 4163, Berkeley, CA 94704-0163, U.S.A. Prior back issues are obtainable from Periodicals Service Company, 11
Main Street, Germantown, NY 12526-5635. The Pacific Journal of Mathematics is indexed by Mathematical Reviews, Zentralblatt MATH,
PASCAL CNRS Index, Referativnyi Zhurnal, Current Mathematical Publications and the Science Citation Index.
The Pacific Journal of Mathematics (ISSN 0030-8730) at the University of California, c/o Department of Mathematics, 969 Evans Hall,
Berkeley, CA 94720-3840 is published monthly except July and August. Periodical rate postage paid at Berkeley, CA 94704, and additional
mailing offices. POSTMASTER: send address changes to Pacific Journal of Mathematics, P.O. Box 4163, Berkeley, CA 94704-0163.
PUBLISHED BY PACIFIC JOURNAL OF MATHEMATICS
at the University of California, Berkeley 94720-3840
A NON-PROFIT CORPORATION
Typeset in LATEX
Copyright ©2006 by Pacific Journal of Mathematics
Vol. 193, No. 1, 2000
SIMPLE CONNECTIVITY OF THE MARKOV PARTITION

SPACE
L. Badoian and J.B. Wagoner
In Wagoner, 1987 the simplicial complex PA of Markov par-

titions was introduced as a tool for studying the group of au-
tomorphisms of a subshift of finite type (XA , σA ) built from
a zero-one transition matrix A. Triangles in PA led to the
matrix Triangle Identities in Wagoner, Pac. Journal, 1990
which have been used in Wagoner, 1990, 1990, 1990, 1992,
Kim, Roush & Wagoner, 1992, and the Williams Conjecture
counterexample paper Kim & Roush, to appear.
A key fact about PA is that it is contractible. See Wagoner,
1987. The purpose of this note is to correct the proof on pp.
99-100 in Wagoner, 1987 that PA is simply connected and
in the process to improve the bound in Proposition 2.13 of
Wagoner, 1987.
Proposition. A closed path in PA with L edges can be spanned by a

(possibly singular) triangulated 2-disc in PA having at most 8L2 + L trian-
gles.
The difficulty with the proof of (2.13) in [W1] occurs in the diagram of
Step 2 on p. 99, because it may not be the case that Vi−1 ∩ Vi+1 is a Markov
partition.
To correct this, it is better to change to a more straightforward notation
and let U −→ V rather than V − → U mean V < U < σA (V ) ∩ V . Recall
− −
−1
from [W1] that U −
→ V means U < V < U ∩ σA (U ). Then Definition 2.10
+
of [W1] becomes
U →V iff U−
→U ∩V −
→ V.
+ −
In particular, now U −
→ V implies that U → V in PA but with extra
−
information, whereas in [W1] the notation V −
→ U implied U → V , which
−
is somewhat contrary. Here are some properties of the arrows U −
→ V and
+
U−
→ V . See [W1].
−
1) If U −
→ V and U is a Markov partition, then V is a Markov partition.
+
If U −
→ V and V is a Markov partition, then U is a Markov partition.
−
1
2 L. BADOIAN AND J.B. WAGONER
2) If U → V and U and V are Markov partitions, then U ∩ V is a Markov

partition.
3) If U −→ V , W and U is a Markov partition, then U − → V ∩ W and
+ +
V ∩ W is a Markov partition. If U, V −
→ W and W is a Markov
−
partition, then U ∩ V −
→ W and U ∩ V is a Markov partition.
−
4) If U −
→X−
→ V and U , X, and V are Markov partitions, then U −
→
− + +
U ∩V −
→ V and U ∩ V is a Markov partition.
−
For completeness , we recall Definition 2.11 of [W1] giving the simplicial
structure on PA . Namely, an n-simplex of PA is an ordered (n + 1)-tuple
hV0 , V1 , . . . , Vn i such that Vi → Vj whenever i ≤ j.
The next step is to replace the diagram in Step 2 on p. 99 with the
following diagram
.. ..
. .
Vi+2 −→ Vi+1 ∩ Vi+2 ∩ Vi+3
+
&+ +%
↓ Vi+1 ∩ Vi+2 ↑
.− −-
Vi+1 ←− Vi ∩ Vi+1 ∩ Vi+2
−
-− −.
↑ Vi ∩ Vi+1 ↓
%+ +&
Vi −→ Vi−1 ∩ Vi ∩ Vi+1
+
&+ +%
↓ Vi−1 ∩ Vi ↑
.− −-
Vi−1 ←− Vi−2 ∩ Vi−1 ∩ Vi
−
-− −.
↑ Vi−2 ∩ Vi−1 ↓
%+ +&
Vi−2 −→ Vi−3 ∩ Vi−2 ∩ Vi−1
+
.. ..
. .
We can now deform a closed path of length L to a constant path as fol-

lows: Step 1 on p. 99 deforms the closed path of length L to an alternating
SIMPLE CONNECTIVITY 3
closed path of length 2L with vertices V0 , V1 , . . . , V2L . The number of trian-

gles in this deformation is at most L. Then the above diagram deforms the
alternating closed path of length 2L on the left to an alternating closed path
of length 2L on the right with vertices of the form Vi−1 ∩ Vi ∩ Vi+1 . Repeat-
ing the deformation in the diagram L-1 more times produces an alternating
closed path of length 2L with vertices of the form
Vi−L ∩ Vi−L+1 ∩ . . . ∩ Vi−1 ∩ Vi ∩ Vi+1 . . . ∩ Vi+L−1 ∩ Vi+L .
Thus all the vertices in this path are equal to
V0 ∩ V1 ∩ . . . ∩ V2L .
The total number of triangles in this deformation is at most 8L2 + L .
Remark. The argument in [W1] that Hn (PA ) = 0 for n ≥ 2 avoids the
Vi−1 ∩ Vi+1 type difficulty, because all intersections of Markov partitions en-
countered in the proof are Markov partitions as a consequence of properties
(1) through (4) above. There is a typographical change on p. 102, l.10:
Vps ∩ Vps should read Vps ∩ Vqs .
References
[KR] K.H. Kim and F.W. Roush, Williams’ Conjecture is false for irreducible subshifts,
1997, preprint from Alabama State University at Montgomery, to appear in Annals
of Mathematics.
[KRW] K.H. Kim, F.W. Roush and J.B. Wagoner, Automorphisms of the dimension group
and gyration numbers, Jour. Amer. Math. Soc., 5(1) (1992), 191-212.
[W1] J.B. Wagoner, Markov partitions and K2 , Pub. Math. IHES, 65 (1987), 91-129.
[W2] , Triangle identities and symmetries of a subshift of finite type, Pac. Jour.
Math., 144 (1990), 181-205.
[W3] , Eventual finite order generation for the kernel of the dimension group
representation, Trans. Amer. Math. Soc., 317(1) (1990), 331-350.
[W4] , Higher dimensional shift equivalence and strong shift equivalence are the
same over the integers, Proc. Amer. Math. Soc., 109(2) (1990), 527-536.
[W5] , Classification of subshifts of finite type revisited, Contemporary Mathe-
matics, 135 (1992), 423-444.
Received May 5, 1998. The authors were partially supported by NSF Grant # DMS
9322498.
Berkeley, CA 94720
E-mail address: badoian@math.berkeley.edu
Berkeley, CA 94720
4 L. BADOIAN AND J.B. WAGONER
E-mail address: wagoner@math.berkeley.edu

Vol. 193, No. 1, 2000
NILMANIFOLDS AND ASSOCIATED LIE ALGEBRAS

OVER THE INTEGERS
Bohumil Cenkl and Richard Porter
We consider the correspondence between nilmanifolds and

Lie algebras with rational basis, and we define spectral se-
quences converging to the respective cohomologies. The E2
terms of the spectral sequences are the cohomolgies of iso-
morphic graded Lie algebras.
Each nilmanifold gives rise to a Lie algebra with rational
basis. We give an example which illustrates that not all such
Lie algebras correspond to nilmanifolds. Given a Lie algebra
with rational basis we give a construction that produces a
nilmanifold with Lie algebra that is rationally equivalent to
the starting Lie algebra.
0. Introduction.
In Section 1 we consider the correspondence between groups and Lie alge-
bras over the reals and rationals. A nilmanifold is completely determined by
its fundamental group — a finitely-generated, torsion-free, nilpotent group.
In Section 2 we show how a nilmanifold determines a nilpotent Lie algebra
with a finite rational basis; we give an example to illustrate that not every
such Lie algebra comes from a nilmanifold. In Section 3 we define a spectral
sequence for a nilmanifold. We prove that spectral sequence converges to
the integer cohomology of the nilmanifold and has E2 term isomorphic to
the cohomolgy of a graded Lie algebra. In Section 4 we define the corre-
sponding spectral sequence for the Lie algebra associated to a nilmanifold.
The spectral sequence converges to the cohomology of the Lie algebra; the
E2 term is the cohomology of a graded Lie algebra, and if A is the smallest
subring of the rationals containing the structure constants of the Lie algebra
then with coefficients equal to A the E2 terms of the spectral sequences of
the nilmanifold and the Lie algebra are isomorphic. In Section 5 we give a
construction of a nilmanifold from a nilpotent Lie algebra with finite rational
basis; the Lie algebra associated to this nilmanifold is rationally equivalent
to the original Lie algebra.
5
6 BOHUMIL CENKL AND RICHARD PORTER
1. Nilmanifolds and corresponding Lie algebras over the reals

and the rationals.
A nilmanifold is the quotient of a simply-connected Lie group by a uniform
discrete subgroup. A uniform discrete subgroup can be identified with a
finitely-generated, torsion-free, nilpotent group. We denote such a group by
G. Let GR be a k-dimensional, connected, simply-connected, nilpotent Lie
group. Then GR can be identified with its real Lie algebra LR with bracket
[ , ]. With respect to any basis {d1 , . . . , dk } for LR , the group GR can be
identified with the set of elements ξ1 d1 + · · · + ξk dk , ξj ∈ R, in Rk . The Lie
algebra LR has real structure constants and the group structure ∗ on GR (as
well as on any subgroup) is given by the Campbell-Hausdorff formula:
Using the notation ad(x)(y) = [x, y], ad2 (x)(y) = ad(x)([x, y]) =
[x[x, y]], . . . we can write
(1) x ∗ y = z(x, y),
∞
X
z(x, y) = zn (x, y),
n=1
1 X 0 00
zn (x, y) = (z + zp,q ), where
n p+q=n p,q
0
X (−1)m+1 adp1 (x)adq1 (y) · · · adpm (x)(y)
zp,q = ,
p1 +···+pm =p
m p1 !q1 ! · · · pm !
q1 +···+qm−1 =q−1
pi +qi ≥1,pm ≥1
00
X (−1)m+1 adp1 (x)adq1 (y) · · · adpm−1 (y)(x)
zp,q = .
p1 +···+pm =p
m p1 !q1 ! · · · qm−1 !
q1 +···+qm−1 =q−1
pi +qi ≥1,pm ≥1
It is easy to find the first three homogeneous components of z. Namely,

z1 (x, y) = x + y,
1
z2 (x, y) = [x, y],
2
1 1
z3 (x, y) = [x, [x, y]] + [y, [y, x]].
12 12
It is clear that z(x, 0) = x, z(0, y) = y, and z(z(w, x), y) = z(w, z(x, y)).
Malcev [3] showed that a group GR contains a uniform discrete subgroup
G if and only if there is a basis for LR with rational structure constants. Let
{d1 , . . . , dk } be such a rational basis. Then the set of elements
ξ1 d1 + · · · + ξk dk , ξj ∈ Q,
NILMANIFOLDS AND ASSOCIATED LIE ALGEBRAS OVER THE INTEGERS 7
has the structure of a Lie algebra L = LQ . L as a subset of GR has group

structure given by the Campbell-Hausdorff formula. We denote this group
by GQ . Since L is a nilpotent Lie algebra, a rational basis for L can be chosen
so that {di , . . . , dk } is a basis for an ideal Li of L, with [Li , L] ⊆ Li+1 and
Lk+1 = 0. Such a basis is called a canonical basis for L. It follows from the
Jacobi identity that
(2) [Li , Lj ] ⊆ Li+j
for i, j ≥ 1. L
In particular, the bracket on L induces a graded Lie algebra
structure on i≥1 Li /Li+1 .
Let G be the group on the set of elements
(x1 d1 ) ∗ · · · ∗ (xk dk ), xj ∈ Q,
with group structure given by the map ∗. The elements
(xi di ) ∗ · · · ∗ (xk dk )
generate a normal subgroup G i of the group G.
G = G1 > G2 > · · · > Gk > 1
is a central series. The commutator (a, b) = a−1 b−1 ab for any a ∈ G and
b ∈ G i belongs to G i+1 , and the quotient groups G i /G i+1 are isomorphic to
Q.
Proposition 1 ([3]). The groups G and GQ are isomorphic. The isomor-
phism is given by the maps
ϕ ϕ
G GQ , (x1 d1 ) ∗ · · · ∗ (xk dk ) ξ1 d1 + · · · + ξk dk ,
ψ ψ
where
ξj = xj + ϕj (x1 , . . . , xj−1 ) and xj = ξj + ψj (ξ1 , . . . , ξj−1 )
are rational polynomials.
Proof. See [3].
The explicit formulas for ϕ and ψ follow from the Campbell-Hausdorff
formula.
Note that the maps ϕ and ψ give one-to-one maps of sets
ϕ
G i Li , i = 1, 2, . . . , k.
ψ
The group G, as a set, can be identified with Qk via the map e : G → Qk

given by
e : (x1 d1 ) ∗ · · · ∗ (xk dk ) −→ x = (x1 , . . . , xk ).
k
e induces on Q a group structure by the operation
x · y = e(e−1 x ∗ e−1 y).
It follows that
(3) x · y = ρ(x, y) = (ρ1 (x, y), . . . , ρk (x, y)),
where the ρ0j s are rational polynomials of the form
ρj (x, y) = xj + yj + τj (x1 , . . . , xj−1 , y1 , . . . , yj−1 )
= xj + yj + σj (x1 , . . . , xj−1 , y1 , · · · , yj−1 ) + · · · ,
and where σj = σj (x1 , . . . , yj−1 ) denotes the quadratic part of τj . Then Qk
can be given a Lie algebra structure with bracket [ , ] defined by setting
[x, y] = σ(x, y) − σ(y, x), where σ(x, y) = (σ1 , . . . , σk ).
This bracket on Qk
induces the original Lie algebra structure on L (= GQ )
via the maps e, ϕ, and ψ.
2. Finitely-generated torsion-free nilpotent groups.

Let G be a torsion-free, nilpotent group on k generators. A set of elements
{g1 , . . . , gk } of G is called a canonical basis for G if every element of G can
be expressed in the form
g1x1 · · · gkxk , xj ∈ Z,
so that the elements of the form gixi · · · gkxk generate a normal subgroup Gi
of G, for i = 1, 2, . . . , k, with the quotients Gi /Gi+1 being infinite cyclic.
Any such group is isomorphic to a uniform discrete subgroup of the real
Lie group GR . This isomorphism is given as the composition of maps
i ϕ inclusion
G −→ G −→ GQ −→ GR ,
where
i(g1x1 · · · gkxk ) = x1 g1 ∗ · · · ∗ xk gk ,
and
ϕ(x1 g1 ∗ · · · ∗ xk gk ) = ξ1 g1 + · · · + ξk gk .
As in Proposition 1, it follows that {g1 , . . . , gk } is a canonical basis for
the Lie algebra L. Note that while the x0j s are integers the ξj0 s are rational
numbers in general. Furthermore, recall that GQ = LQ = L and GR = LR as
sets. We identify G with its image i(G) in G. Thus we have:
(i) G is the subgroup of G consisting of the elements
(x1 g1 ) ∗ · · · ∗ (xk gk ), xj ∈ Z.
(ii) The groups Gi = G ∩ Gi of the elements
(xi gi ) ∗ · · · ∗ (xk gk ), xj ∈ Z
are normal subgroups of G.
L Gi /Gi+1 are infinite cyclic.

(iii) The quotient groups
(iv) The Z module i≥1 Gi /Gi+1 has a graded Lie algebra structure in-
duced by the commutators on G. This Lie algebra structure is induced
from that on i≥1 Li /Li+1 by the map ϕ ◦ i : G −→ L.
L
(v) The group G, as a set, can be identified with Zk ⊂ Qk via the map
i ◦ e. The group structure on Zk in (3) is induced by e. Therefore, the
rational polynomials ρj are Z-valued polynomials when restricted to
Zk × Zk . Each polynomial ρj (x, y) can be written as a sum (over Z)
of binomials

x1 x2 xr y1 y2 ys
··· ··· ,
a1 a2 ar b1 b2 bs
with au , bv nonnegative integers, r, s ≤ j, where not all a0u s and not
all b0v s are equal to zero and also as sum (over Q) of polynomials (see
[1]).
Thus, we can consider the group Zk with the group operation given by
(3). We denote by G the image of Zk in GQ under the one-to-one map
ch = ϕ ◦ e−1 : Zk −→ G ⊂ GQ .
Recall that, given a canonical basis {g1 , . . . , gk } for G, G k is the set of
elements ξ1 g1 + · · · + ξk gk , where ξ1 , . . . , ξk are certain rational numbers. G k
with the operation ∗ given by the Campbell-Hausdorff formula is a group.
It is important to note that the sequence of rational numbers ξ1 , . . . , ξk is
not an arbitrary sequence of rational numbers; it depends on the choice of
a canonical basis {g1 , . . . , gk } for G. Once a canonical basis is chosen then
ξj = xj + ϕj (x1 , . . . , xj−1 ), where the xu ’s are integers and the rational
polynomials ϕj are determined by the Lie algebra bracket.
We summarize our observations as follows:
Proposition 2. Let G be a torsion-free nilpotent group on k generators, and
let {g1 , . . . , gk } be a canonical basis for G. Then the group G is isomorphic
to the group Zk with the group operation defined by
(4) x · y = ch−1 (ch(x) ∗ ch(y)),
where x = (x1 , . . . , xk ), y = (y1 , ..., yk ) and where ∗ is given by the Campbell-
Hausdorff formula (1).
In order to illustrate the type of rational numbers ξ1 , . . . , ξk that can
occur, we look again at a classical example.
Example 1 (The Heisenberg group). Let G be the group of matrices
 
1 a1 a3 /k
0 1 a2  , where aj ∈ Z and k is a fixed integer > 0.
0 0 1
Then
      
 1 1 0 1 0 0 1 0 1/k 
g1 = 0 1 0 , g2 = 0 1 1 , g3 = 0 1 0 
0 0 1 0 0 1 0 0 1
 
is a canonical basis, and

 
1 x1 (x3 /k) + x1 x2
g1x1 g2x2 g3x3 = 0 1 x2 .
0 0 1
{g1 , g2 , g3 } is a canonical basis for L, [g1 , g2 ] = kg3 , and [gi , gj ] = 0 other-
wise. We have

k
ch(x) = x1 g1 + x2 g2 + x3 + x1 x2 g3 , x = (x1 , x2 , x3 );
2
therefore,
k
ξ1 = x1 , ξ2 = x2 , ξ3 = x3 + x1 x2 .
2
3
It can be checked that the group structure on Z is given by (4). Namely,
ch(x) ∗ ch(y) =(x1 + y1 )g1 + (x2 + y2 )g2

k
+ (x3 + y3 − ky1 x2 ) + (x1 + y1 )(x2 + y2 ) g3
2

=ϕ ((x1 + x2 )g1 ) ∗ ((y1 + y2 )g2 ) ∗ ((x3 + y3 − ky1 x2 )g3 )
=ϕ ◦ e−1 (x · y)
=ch(x · y).
As was mentioned above, a canonical basis for G determines a canonical
basis for the rational Lie algebra. However, the converse is not always true.
A canonical basis for L is not always a canonical basis for G. This fact is
demonstrated by the next example. We will see later that given a canonical
basis for a Lie algebra, there is always a rationally equivalent basis that is
induced by a canonical basis of the corresponding group.
Example 2. Let L be the Lie algebra over Z with generators g1 , . . . , g7 and
with product given by
[g1 , g2 ] = g4 − g5 ,
[g1 , g3 ] = g4 + g5 , [g2 , g3 ] = g5 ,
[g1 , g4 ] = g6 , [g2 , g4 ] = g6 , [g3 , g4 ] = g7 ,
[g1 , g5 ] = g6 , [g2 , g5 ] = g7 ,
[gi , gj ] = 0 otherwise.
L is a nilpotent finitely-generated Lie algebra with no torsion elements. The
set {g1 , . . . , g7 } forms a canonical basis for L with {gi , . . . , g7 } a basis for
the ideal Li . We want to show that {g1 , . . . , g7 } can not be a canonical
basis for a finitely-generated torsion-free nilpotent group whose associated

Lie algebra is L. The idea is the following. Assuming that {g1 , . . . , g7 } is
such a canonical basis we show that there are elements x, y ∈ Z7 such that
ch−1 (ch(x) ∗ ch(y)) fails to be a sequence of seven integers.
Let x = (x1 , . . . , x7 ) ∈ Z7 . Then ch(x) ∈ GQ ,
ch(x)
= ϕ(x1 g1 ∗ · · · ∗ x7 g7 )
= x1 g1 + x2 g2 + x3 g3
+ (x4 + 1/2(x1 x2 + x1 x3 ))g4 + (x5 + 1/2(−x1 x2 + x1 x3 + x2 x3 ))g5
+ (x6 + 1/12(−x1 x22 + 2x21 x3 + 2x1 x2 x3 + 6x1 x4 + 6x2 x4 + 6x2 x5 ))g6
+ (x7 + 1/12(x1 x22 − x1 x33 + x22 x3 − 2x1 x2 x3 + 6x3 x4 + 6x2 x5 ))g7
and there is a similar formula when x is replaced by y = (y1 , . . . , y7 ).

Now we assume that there is a sequence of integers z = (z1 , . . . , z7 ) ∈ Z7
such that ch(z) = ch(x) ∗ ch(y). A routine computation shows that
zi = xi + yi , i = 1, 2, 3,
z4 = x4 + y4 − x2 y1 − x3 y1 ,
z5 = x5 + y5 + x2 y1 − x3 y1 − x3 y2 ,
z6 = x6 + y6 − x4 y1 − x4 y2 − x5 y1
1
+ x2 y1 y2 + x3 y1 y2 + x3 y12 + x22 y1
2
z7 = x7 + y7 − x5 y2 − x4 y3 − x2 y1 y2 + x2 y1 y3 + x3 y1 y2
1 1 1
+ x3 y1 y3 + x2 x3 y1 + x23 y1 − x22 y1 x3 y22 .
2 2 2
From here it is immediate that for x and y such that x2 = 1, xj = 0, j 6= 2

and y1 = 1, yi = 0, i 6= 1 we get z6 = 1/2 and z7 = −1/2. Thus, the
cannonical basis {g1 , . . . , g7 } for L is not induced by a cannonical basis for
any group.
However, it is possible to find a nilpotent torsion-free Lie algebra L0 on

seven generators e1 , . . . , e7 over Z such that L and L0 are isomorphic over
Q (but not over Z) and that the generators e1 , . . . , e7 are induced by a
canonical basis for a torsion-free nilpotent group on 7 generators and that
L0 is the corresponding Lie algebra. Such a Lie algebra is the following:
Example 3. Let L0 be the Lie algebra over Z with generators e1 , . . . , e7

and products
[e1 , e2 ] = 2e4 − 2e5 ,
[e1 , e3 ] = 4e4 + 4e5 , [e2 , e3 ] = 2e5 ,
[e1 , e4 ] = 2e6 , [e2 , e4 ] = e6 , [e3 , e4 ] = 2e7 ,
[e1 , e5 ] = 2e6 , [e2 , e5 ] = e7 ,
[ei , ej ] = 0 otherwise.
In this case for x = (x1 , . . . , x7 ) and y = (y1 , . . . , y7 ) from Z7 , ch(x) and
ch(y) belong to G7 ⊂ GQ ,
ch(x)
= x1 e1 + x2 e2 + x3 e3
+ (x4 + 2x1 x3 + x1 x2 )e4 + (x5 + 2x1 x3 + x2 x3 − x1 x2 )e5
+ 1/12(12x6 + 12x1 x4 + 6x2 x4 + 12x1 x5 + 8x1 x2 x3 + 16x21 x3 − 2x1 x22 )e6
+ 1/12(12x7 + 12x3 x4 + 6x2 x5 − 8x1 x2 x3 − 8x1 x23 + 2x22 x3 + 2x1 x22 )e7 .
There are similar formulas for ch(y). The group structure on Z7 is given
by the formula z = ch−1 (ch(x) ∗ ch(y)),
z1 = x1 + y1 ,
z2 = x2 + y2 ,
z3 = x3 + y3 ,
z4 = x4 + y4 − 2x2 y1 − 4x3 y1
z5 = x5 + y5 + 2x2 y1 − 4x3 y1 − 2x3 y1
z6 = x6 + y6 − 2x4 y1 − x4 y2 − 2x5 y1 + 2x2 y1 y2 + 4x3 y1 y2 + 8x3 y12 + x22 y1
z7 = x7 + y7 − x5 y2 − x4 y3 − 2x2 y1 y2 + 4x2 y1 y3 + 4x3 y1 y2 + 8x3 y1 y3
+ 4x2 x3 y1 + 4x23 y1 − x22 y1 + x3 y22 .
Thus, the group G0 whose Lie algebra is L0 is the set of elements of the form
ex1 1 ∗ · · · ∗ ex7 7 with the product given by the above formulas.
Given a finitely-generated, torsion-free, nilpotent Lie Algebra over the in-
tegers, there is a family of Lie algebras over the integers that are equivalent
to it over the rationals. All such algebras can be provided with canonical
bases. However, as we have seen in Example 2, not all such bases are deter-
mined by a canonical basis for a finitely-generated, torsion-free, nilpotent
group. It seems natural to ask the following question: Suppose that we are
given a finitely-generated nilpotent Lie algebra over the integers. Can we
find a finitely-generated torsion-free nilpotent group and a canonical basis
for such a group so that its associated Lie algebra is a Lie algebra over the
integers that is rationally equivalent to the original one? It turns out that it
is possible. In the last part of this paper we describe a procedure for finding
such a group.
3. The spectral sequence for polynomial cochains.

In this part we start with a finitely-generated, torsion-free, nilpotent group
G with a canonical basis. The induced Lie algebra is a Lie algebra over the
rationals together with the induced canonical basis. In [1] we proved that the
integral cohomology of G is isomorphic to the cohomology of the differential
algebra P (G) of rational polynomial maps from Zk to Z, where k is the
number of generators of the group G. The differential on P (G) is induced
by the group structure on Z. In this section we use the structure of the group
G to define a filtration on P (G), and study the resulting spectral sequence.
Starting with the sequence of normal subgroups of G, G = G1 > G2 >
· · · > Gk > 1, denote by i1 the largest positive integer such that G1 /Gi1 +1
is a torsion-free, abelian group and by i2 the largest positive integer such
that Gi1 +1 /Gi2 +1 is a torsion-free, abelian group. Inductively, let 0 = i0 <
i1 < i2 < · · · < iN −1 < iN = k be the largest positive integers such that
Gij−1 +1 /Gij +1 , j = 1, 2, . . . , N − 1, are torsion-free, abelian groups. Then
G = G1 > G2 > · · · > GN +1 = 1, Gj+1 = Gij +1 , j = 0, 1, 2, . . . , N − 1
is the shortest central series with torsion-free quotients. Then ϕ maps this
sequence of normal subgroups to a sequence of ideals of the Lie algebra L.
L = L1 ⊃ L2 ⊃ · · · ⊃ LN +1 = 0, Lj+1 = Lij +1 ,
ϕ (Gj ) ⊂ Lj , j = 1, 2, . . . , N.
From (4) it follows that the commutators induce a Lie algebra structure on
the Z-module gr G = ⊕j≥1 Gj /Gj+1 . The Lie bracket [ , ] on L induces a
Lie algebra structure on the Q-module gr L = ⊕j≥1 Lj /Lj+1 . In addition the
graded Lie algebras gr G ⊗ Q and gr L are isomorphic. The isomorphism is
induced by ϕ.
For the remainder of this paper, we assume that the canonical basis
{g1 , . . . , gk } for G is such that {g1 , . . . , gi1 } projects to a basis for G1 /G2 ;
{gi1 +1 , . . . , gi2 } projects to a basis for G2 /G3 ; and in general
{gij−1 +1 , . . . , gij } projects to a basis for Gj /Gj+1 .
To each basis element gj of g we assign a weight as follows. If gj projects
to a basis element for Gr /Gr+1 , then its weight is |gj | = r. In particular,
|gir−1 +1 | = · · · = |gir | = r for r = 1, 2, . . . , and for every rational number q
we set
|qgj | = |gj |.
Furthermore,
|gr + gs | = max{|gr |, |gs |}.
Since a canonical basis for G induces a canonical basis for L, we set the
weight of an element gj of L to be the weight of gj considered as an element
of G.
From (2) it follows that for any two elements of the canonical basis gr ∈
La , gs ∈ Lb the Lie bracket is
X r,s
[gr , gs ] = qj gj ,
j≥r+s
where qjr,s ∈ Q and gj ∈ La+b . Thus,

[gr , gs ] ≥ |gr | + |gs |.
The above summation goes over all j such that
|gj | ≥ |gr | + |gs |.
Next we assign a weight to each binomial map Zk → Z according to the
rules:
|xj | = |gj |,

xj
a = a|xj |, a = 0, 1, 2, . . . ,

xj xj
p
a = a
for any integer p 6= 0,
|u + v| = max{|u|, |v|},
|uv| = |u| + |v|.
Analogously, we assign a weight to each monomial map Zk → Z as follows:
|xj | = |gj |,
|xaj | = a|xj |,
|qxj | = |xj | for any rational number q 6= 0,
|u + v| = max{|u|, |v|},
|uv| = |u| + |v|.
Observe that each binomial of weight w is a sum of monomials of weight
≤ w and that each monomial of weight w is a sum of binomials of weight
≤ w.
Lemma 1. Let the group structure on G be induced by the ·-product (3).
Then each ρj (x, y) is a sum (over Z) of binomials. ρj (x, y) can be also
expressed as a sum (over Q) of monomials of weight ≤ |gj |, j = 1, 2, . . . .
Proof. Because the group structure on Zk is given by restricting the ∗-
product from G to e−1 (Zk ), it follows that the statement is proved if we
verify that
|ρj (x, y)| ≤ |gj |, j = 1, 2, . . . , k
for every x, y ∈ Zk .
Equivalently, it suffices to show that the above inequality is satisfied in
the product formula on G,
((x1 g1 ) ∗ · · · ∗ (xk gk )) ∗ ((y1 g1 ) ∗ · · · ∗ (yk gk )) = (ρ1 (x, y)g1 ∗ · · · ∗ ρk (x, y)gk ).

ϕ
But the maps ϕ, ψ; G L are such that
ψ
ψ(ξ1 g1 + · · · + ξk gk )
= (ξ1 g1 ) ∗ · · · ∗ (ξj + ψj (ξ1 , . . . , ξj−1 ))gj ∗ · · · ∗ (ξk + ψk (ξ1 , . . . , ξk−1 ))gk ,
ϕ(x1 g1 ∗ · · · ∗ xk gk )
= x1 g1 + · · · + (xj + gj (x1 , . . . , xj−1 ))gj + · · · + (xk + gk (x1 , . . . , xk−1 ))gk ,
where
|ξj + ψj (ξ1 , . . . , ξj−1 )| ≤ |ξj | = |gj |
and
|xj + ϕj (x1 , . . . , xj−1 )| ≤ |xj | = |gj |.
It follows that the statement is proved if we show that the product in L,

given by (1)
ξ ∗ η = (ξ1 g1 + · · · + ξk gk ) ∗ (η1 g1 + · · · + ηk gk )
= (β1 (ξ, η)g1 + · · · + βk (ξ, η)gk )
holds with βj (ξ, η) = βj (ξ1 , . . . , ξj−1 , η1 , . . . , ηj−1 ) and |βj (ξ, η)| ≤ |gj | for
j = 1, 2, . . . , k. But
1
βj (ξ, η) = ξj + ηj + [ξ, η]j + · · · .
2
Hence, it is enough to verify that |[ξ, η]j | ≤ |gj | if |ξj | ≤ |gj | and |ηj | ≤ |gj |
for all j = 1, 2, . . . .
Since
X r,s
[gr , gs ] = aj gj , ar,s
j ∈ Q,
j≥r+s
where the sum is over all gj0 s such that |gj | ≥ |gr | + |gs | and
k
ξr ηs ar,s
X X
[ξ, η] = j gj ,
r,s=1 j≥r+s
the result follows from the identities

r,s
X
|[ξ, η]j | =
ξr ηs aj
1≤r,s≤k
r+s≤j
≤ max {|ξr ηs ar,s
j |}
1≤r,s≤k
r+s≤j
= max{|ξr ηs |}
= max{|ξr | + |ηs |}
≤ max{|gr | + |gs |} ≤ |gj |.

Recall from [1] that an element of P n (G), n ≥ 1, is given as a linear
combination, with integral coefficients, of the binomials
Y xij
,
aij
1≤i≤k
1≤j≤n
where the a0ij s are nonnegative integers. P 0 (G) = Z. The differential d on

P (G), as a derivation, is completely determined by its values on P 1 (G). On
a typical binomial
x1 x2 xk
u= ···
a1 a2 ak
1
in P (G)

y1 yk z1 zk ρ1 ρk
(5) du = ··· + ··· − ··· ,
a1 ak a1 ak a1 ak
where ρj = yj + zj + τj , or more explicitly
ρj (y1 , . . . , yj , z1 , . . . , zj ) = yj + zj + τj (y1 , . . . , yj−1 , z1 , . . . , zj−1 ),
where τj is a polynomial of degree ≥ 2. To each polynomial cochain u in
P (G) we assign a nonnegative integer kuk called its norm, according to the
following rules:
kqgi k = kgi k for any rational number q,

xi
a = akxi k for i = 1, 2, . . . , k,

xi xi
p
a = a for any integer p,

kuvk = kuk + kvk,

ku + vk = max{kuk, kvk} for any u ∈ P i (G), v ∈ P j (G).
Lemma 2. The polynomial functions ρj and τj have their norms bounded

from above. Namely,
kρj (x1 , . . . , xj , y1 , . . . , yj )k ≤ |gj | − 1,
kτj (x1 , . . . , xj−1 , y1 , . . . , yj−1 )k ≤ |gj | − 2.
Proof. ρj (x1 , . . . , xj , y1 , . . . , yj ) = xj + yj + τj (x1 , . . . , xj−1 , y1 , . . . , yj−1 ),
where τj is a sum of binomials

xi1 xir yj1 yjs
β= ··· ··· ,
ai1 air bj1 bjs
r ≥ 1, s ≥ 1, with integer coefficients. The norm of such a binomial β is
X X
kβk = au kxu k + bv kyv k
X X
= au (|xu | − 1) + bv (|yv | − 1)
X X
≤ au |xu | + bv |yv | − 2
= |β| − 2
≤ |ρj | − 2
≤ |gj | − 2
according to Lemma 1. Therefore, kτj k ≤ |gj | − 2. Since kρj k =
max{kxj k, kβk} over all binomials β in ρj , it follows that kρj k = kxj k.
Hence, kρj k = |gj | − 1.
Using the norm k k, we define a filtration on P (G) by setting

F i = F i P (G) = {u ∈ P (G)| kuk ≤ i}, i = 0, 1, 2, . . . ,
0 1 2 i−1 i
F ⊂ F ⊂ F ⊂ ···F ⊂ F ⊂ · · · ⊂ P (G).
This filtration has the following properties:
Lemma 3. (i) Under the cup product on chains
F i × F j −→ F i+j .
(ii) The differential d on P (G) preserves the filtration; i.e.,
dF i ⊂ F i .
Proof. (i) follows immediately from the definition.
(ii) follows once we verify that for every binomial u ∈ F i , du ∈ F i .
Let
x1 xk
u= ··· .
a1 ak
Then du is given by formula (5) above. Note that

X k
y1 yk

a1 · · · = aj (|gj | − 1) ≤ i
ak
j=1
and
X k
z1 zk
a1 · · · ak = aj (|gj | − 1) ≤ i.

j=1
And according to Lemma 2,
X k k
ρ1 ρk X
a1 · · · ak = aj kρj k ≤ aj (|gj | − 1),

j=1 j=1
which is ≤ i, by the assumption. Therefore, kduk ≤ i. Hence, du ∈ F i .
3.1. E0 and E1 -terms of the spectral sequence. Let {Er , dr } be the

spectral sequence corresponding to the filtration {F i }. In particular, we
have
E0 = F i /F i−1 .
Let p be the projection F i → E0 . The next step is to compute E1 . Since

d0 = pd and because each element of F i is an integral linear combination
of binomials, we look more closely at the last term in formula (5) for du,
where u ∈ F i .

ρ1 ρk y 1 + z1 + τ 1 y k + zk + τ k
··· = ···
a1 ak a1 ak
 
k  aXj −1 
Y y j + zj τj y j + zj τj
+ + .
 aj aj r aj − r 
j=1 r=1
Observe that
kτj k ≤ |gj | − 2
by Lemma 2. Therefore,

τj
aj = aj kτj k < aj (|gj | − 1)

and

aj −1
X y j + zj τj

r=1 r aj − r

y j + zj τj
= max
1≤r≤aj −1 r aj − r
= max {rkyj + zj k + (aj − r)kτj k}
1≤r≤aj −1
= max {rkyj k + (aj − r)kτj k, rkzj k + (aj − r)kτj k}
1≤r≤aj −1
= max {rkyj k + (aj − r)kτj k}
1≤r≤aj −1
since kyj k = kzj k. Because kyj k = |gj | − 1 and kτj k < |gj | − 1, the last term
is strictly smaller than
max {(r + (aj − r))(|gj | − 1)} = aj (|gj | − 1).
1≤r≤aj −1
ρ1 ρk

Therefore, the binomial a1 ··· ak has the form
Y k
ρ1 ρk
··· = {Xj + Tj },
a1 ak
j=1
where
kXj k = aj (|gj | − 1), kTj k < aj (|gj | − 1).
But
k
Y k
Y X
(Xj + Tj ) = Xj + Ti1 · · · Tir Xj1 · · · Xjk−r ,
j=1 j=1
where the sum is over all sequences i1 , . . . , ir of elements from 1, 2, . . . , k
of length r ≥ 1 and where j1 , . . . , jk−r is the complementary sequence.
Observe that
X
Ti1 · · · Tir Xj1 · · · Xjk−r

= max kTi1 · · · Xjk−r k
(all sequences)
< max {ai1 (|gi1 | − 1) + · · · + air (|gir | − 1)
(all sequences)
+ aj1 (|gj1 | − 1) + · · · + ajk−r (|gjk−r | − 1)}
which is strictly smaller than i by the assumption that u ∈ F i . Therefore,
 
k k k 
Y yj Y zj Y y j + zj
d0 u = p + −
 aj aj aj 
j=1 j=1 j=1
xi

or d0 u = pd1 u, where d1 u = 0 for u = when 1 ≤ i ≤ k and
1
j −1
k aY
Y yj zj
d1 u = − ,
r aj − r
j=1 r=1
when aj ≥ 2. But this is the differential on P 1 (A), when A is the Z-module

G1 /G2 ⊕ · · · ⊕ Gj /Gj+1 ⊕ · · · considered as an abelian group. Using the
results of [1], we obtain:
Theorem 1. The E1 -term V of the spectral sequence {Er , dr } is isomorphic
to the exterior algebra Z (x1 , . . . , xk ) on one-dimensional generators.
3.2. E2 -term of the spectral sequence. Since the terms of the spectral
sequence
V are algebras, it is sufficient to find the value of d1 on the generators
xj of Z (x1 , . . . , xk ). If xj , kxj k = i, is a representative of xj then d1 xj
will be determined by the projection of dxj into the submodule of P (G) of
elements whose norm is i − 1.
From the definition of d on P (G) it follows that
dxj = −τj (y1 , . . . , yj−1 , z1 , . . . , zj−1 ),
where
τj (y1 , . . . , yj−1 , z1 , . . . , zj−1 )

X y1 yj−1 z1 zj−1
= c(a1 , . . . , aj−1 , b1 , . . . , bj−1 ) ··· ··· ,
a1 aj−1 b1 bj−1
where the coefficients c(· · · ) are integers and at least two of the nonnegative
integers a1 , . . . , bj−1 are nonzero. Note that
X j−1
y1 zj−1
a1 · · · bj−1 = (at kyt k + bt kzt k)

t=1
j−1
X
= (at + bt )(|gt | − 1)
t=1
j−1
X j−1
X
= (at + bt )|gt | − (at + bt ).
t=1 t=1
We are looking for the sequences of nonnegative integers a1 , . . . , aj−1 ,
b1 , . . . , bj−1 such that

y1 yj−1 z1 zj−1
a1 · · · aj−1 ···
= i − 1 = |gj | − 2.
b1 bj−1
From the above identity, it follows that such sequences satisfy
j−1
X j−1
X
(at + bt )|gt | − (at + bt ) = |gt | − 2.
t=1 t=1
But from Lemma 1, it follows that

j−1 j−1
y1 zj−1 X X

a1 · · · = (a |y
t t | + b |z
t t |) = (at + bt )|gt | ≤ |gj |.
bj−1
t=1 t=1
Therefore,
j−1
X j−1
X
(at + bt )|gt | = (at + bt ) + |gj | − 2 ≤ |gj |.
t−1 t=1
Hence, we must have

j−1
X
(at + bt ) ≤ 2.
t=1
Pj−1
But if t=1 (at + bt ) ≤ 1, then there cannot be two nonzero integers among
a1 , . . . , bj−1 . Hence,
j−1
X
(at + bt ) = 2.
t=1
But then exactly one of the at ’s and exactly one of the bt ’s must be equal to
1 and all the other integers must be equal to 0. Therefore, d1 xj is determined
by the terms in τj of the form
j−1
cr,s cr,s
X
j xr ys , j ∈ Z,
r,s=1
where
kxr ys k = kxr k + kys k = |gj | − 2.
Theorem 2. The E2 -term of the spectral sequence {Er , dr } is isomorphic
to the integral cohomology of the graded Lie algebra
M
grG = Gj /Gj+1 .
j≥1
4. The spectral sequence for the Lie algebra cochains.

Let G be a finitely-generated, torsion-free, nilpotent group with canonical
basis {g1 , . . . , gk }. We denote by A the smallest subring of the rationals
containing the structure constants of the corresponding Lie algebra L. Let
{ω1 , . . . , ωk } be the A-dual of the canonical basis {g1 , . . . , gk } of L. The
algebra of A-valued cochains on L is isomorphic to the exterior algebra
^ ^
(L∗ ) = {ω1 , . . . , ωk },
A
L∗ = HomA (L, A). Recall that there is a pairing

r
^ r
^
h, i: (L∗ ) ⊗ (L) −→ A,
∗ ∗ ∗
Vr
Vr that for every v = v1 ∧ · · · ∧ vr ∈
such (L∗ ) and any v = v1 ∧ · · · ∧ vr ∈
(L)
hv ∗ , vi = det(hvi∗ , vj i).
In particular, for every f ∈ r (L∗ ) and g ∈ s (L∗ ), f ∧ g ∈ r+s (L∗ ) and

V V V
hf ∧ g, v1 ∧ · · · ∧ vr+s i
X
= (sgnπ)hf, vπ(1) ∧ · · · ∧ vπ(r) ihg, vπ(r+1) ∧ · · · ∧ vπ(r+s) i.
r,s shuffle
V2
For every ω = a1 ω1 + · · · + ak ωk , aj ∈ A, the differential d : L∗ −→ L∗
takes the form
dω = a1 ω1 + · · · + ar ωk
and extends to the map d :V r (L∗ ) −→ r+1 (L∗ ) as a derivation.
V V
Let k k be the norm on A (ω1 , . . . , ωk ) defined by the formulas
kωj k = |gj | − 1, kqωj k = kωj k for q ∈ A,

∗ ∗ ∗ ∗
ku ∧ v k = ku k + kv k,
ku∗ + v ∗ k = max{ku∗ k + kv ∗ k}.
Using this norm, we define an ascending filtration on (L∗ ) by setting

V
^ n ^ o
F i = F i (L∗ ) = v ∗ ∈ (L∗ ) | kv ∗ k ≤ i .
Then
^
F 0 ⊂ F 1 ⊂ · · · ⊂ F i ⊂ F i+1 ⊂ · · · ⊂ (L∗ ).
(L∗ ) lowers the filtration.

V
Lemma 4. The differential d on
Proof. Since d is a derivation it suffices to verify that if kωj k = i, then

kdωj k < i. But
kdωj k = max{kωr k + kωs k}

= max{kgr k + kgs k},
where the max is taken over all pairs kgr k + kgs k ≤ kgj k − 1 < kgj k = i.
Thus, dF i ⊂ F i−1 , i ≥ 1.
4.1. E0 , E1 , and E2 terms of the spectral sequence for Lie algebra

cochains. We denote by {Er , dr } the spectral sequence associated with the
above filtration. In particular, let
E0i = F i /F i−1 ,
and let p : F i −→ F i /F i−1 be the projection. Set ω ∗j = pωj∗ . Then d0 on
the element ω j with kωj k = i is determined by those elements in dωj which
have norm exactly equal to i. But kωj k ≤ i − 1. Hence, d0 ≡ 0.
V
Lemma 5. E1 is isomorphic to the exterior algebra A (ω 1 , . . . , ω k ) on gen-
erators of dimension one.
The differential d on (L∗ ) is dual to the Lie bracket in the sense that
V
hdωj , gr ∧ gs i = −hωj , [gr , gs ]i

X r,s
=− qi hωj , gi i
X u,v
=− qj hωu ∧ ωv , gr ∧ gs i.
u<v
Note that the first summation is always zero if |gj | < |gr | + |gs | because
|gr | + |gs | ≤ |gi | and hωj , gi i = 0 whenever |gj | < |gi |. Hence,
X r,s
dωj = − qj ωr ∧ ωs
r<s
with |gr | + |gs | ≤ |gj | or |ωr | +V|ωs | ≤ |ωj |, when |ωi | = kωi k + 1.
The differential d1 on ω j ∈ A (ω 1 , . . . , ω k ) is determined by the terms in
ωj of the form
X r,s
qj ωr ∧ ωs ,
r<s
where kgr k + kgs k = kgj k − 1 or equivalently |gr | + |gs | = |gj |. This implies:
Theorem 3. The E2 -term of the spectral sequence {Er , dr } is isomorphic to
the cohomology of the graded Lie algebra
M
gr L = (Lj /Lj+1 ⊗ A) .
j≥1
Suppose that the group G is two-stage nilpotent, and let G = G1 > G2 >
1 be the shortest central series with torsion free quotients. Let {g1 , . . . , gk }
be a canonical basis such that {g1 , . . . , gi } projects to a basis for G1 /G2 and
{gi+1 , . . . , gk } projects to a basis for G2 . Then {g1 , . . . , gk } determines a
canonical basis for the induced Lie algebra L. We use the same symbols for
the canonical bases of the group and its associated Lie algebra. Then there
is a sequence of ideals of L, L = L1 ⊃ L2 ⊃ 0 and the subset {g1 , . . . , gi }
of the canonical basis for L projects to a basis for L1 /L2 and the subset
{gi+1 , . . . , gk } projects to a basis for L2 .
Note that for the elements of the canonical basis for G and L, the norm
is
kga k = 0 for a = 1, . . . , i; kgb k = 1 for b = i + 1, . . . , k,
and there are no elements of norm higher than 1. Note that the differential
dxj on xj in P 1 (G) can be written in the form
dxj = δ0 (xj ) + δ1 (xj ) + · · · + δr (xj ) + · · · ,
where δr (xj ) is the element of the submodule of P (G) whose norm is kxj k−r.
The differential dr in the spectral sequence Er for the group G is determined
by δr (xj ). Thus, in the case of a two-stage nilpotent group dr = 0 for r ≥ 2.
A similar argument shows that the differentials dr for r ≥ 2 for the Lie
algebra spectral sequence Er are also zero. Therefore, the spectral sequence
{Er , dr } converges to
H ∗ (G; Z); i.e., E2 ∼
= E∞ ∼= grH ∗ (G; Z),
where grH ∗ (G; Z) is the graded module associated with the filtration {F i }
of P (G). If the Lie algebra L is a Lie algebra over the integers, then A = Z
and we can also conclude that the spectral sequence {Er , dr } converges to
H ∗ (L; Z); i.e.,
E2 ∼
= E∞ ∼ = grH ∗ (L; Z).
Therefore, from Theorem 2 and Theorem 3 we obtain:
Corollary 1. Let G be a two-stage, nilpotent, finitely-generated, torsion-
free group, and let L be its Lie algebra such that L is a Lie algebra over the
integers. Then there is an isomorphism of graded modules
grH ∗ (G; Z) ∼
= grH ∗ (L; Z).
If we do not assume that G is a two-stage nilpotent group and that the
associated Lie algebra L is over the integers, then we get the following:
Corollary 2. Let A be the smallest subring of rationals containing the struc-
ture constants of the Lie algebra L. Then there is an isomorphism of A-
modules
E2 ⊗ A ∼= E2 ∼
= H ∗ (grL; A).
The conclusion of Corollary 1 is valid even for certain (perhaps for all)
three-stage, nilpotent, finitely-generated, torsion-free groups. However, the
argument based only on the length of lower central series used in the proof of
the Corollary 1 is not sufficiently strong to prove the statement. An explicit
computation of the E2 -term and of the differential d2 is needed even in the
example below.
Example 3 (continued). Let G0 and L0 be the group and its associated
Lie algebra as stated above. Recall that {g1 , . . . , g7 } is a canonical basis for
G0 and also for L0 (as we keep the same notation for both). The shortest
lower central series with torsion-free quotients is G = G1 > G2 > G3 > 1.

Then {g1 , g2 , g3 } projects to a basis of G1 /G2 , {g4 , g5 } projects to a basis of
G2 /G3 , and {g6 , g7 } projects to a basis of G3 . The Lie algebra L0 is filtered
by the ideals L0 = L01 ⊃ L02 ⊃ L03 ⊃ 0 with appropriate quotients. The
weights are kg1 k = kg2 k = kg3 k = 0, kg4 k = kg5 k = 1, kg6 k = kg7 k = 2.
Since the comparison of Theorems 2 and 3 implies that the E2 -term of the
spectral sequence for the group G0 and the E2 -term of the spectral sequence
for the Lie algebra L0 are isomorphic Z-modules, it suffices to compute only
one of them. We choose to compute the E2 -term for L0 .
Let {ω1 , . . . , ω7 } be the Z-dual basis of the canonical basis {g1 , . . . , g7 }
of L0 . Then the exterior algebra
∗
^ ^ ^
= L0 = HomZ (L0 , Z) = (ω1 , . . . , ω7 )
is filtered and graded. The elements ω1 , . . . , ω7 have dimension one and

filtration kω1 k = kω2 k = kω3 k = 0, kωV
4 k = kω5 k = 1, kωV
6 k = kω7 k = 2. The
increasing filtration on the Z-module q of elements of of dimension q is
defined as the sequence of modules
q
^ q
^
Fp = {ω ∈ | kωk ≤ p}.
Vq Vq
Setting F−p = Fp , we get a descending filtration
q
^ q
^ q
^ q
^
F0 ⊂ F−1 ⊂ · · · ⊂ F−p+1 ⊂ ··· ⊂ .
Then
q−p q−p
E0−p,q
^ ^
= F−p /F−p+1
d0 = 0. Thus, E1−p,q =
with V ∼ E −p,q and d1 is determined by the full differential
0
d on . From the explicit formulas below we can see that kdωj k = kωj k − 1.
Since the differential d is dual to the Lie bracket, we get
dω1 = dω2 = dω3 ,

dω4 = 2ω2 ω1 + 4ω3 ω1 ,
dω5 = −2ω2 ω1 + 4ω3 ω1 + 2ω3 ω2 ,
dω6 = 2ω4 ω1 + 2ω5 ω1 + ω4 ω2 ,
dω7 = ω5 ω2 + 2ω4 ω3 .
Then an explicit computation shows that
E20,q = 0, 4 ≤ q,
E2−1,q = 0, q = 2, 5 ≤ q,
E2−2,q = 0, q = 3, 7 ≤ q,
E2−3,q = 0, q ≤ 5, 9 ≤ q,
E2−4,q = 0, q ≤ 7, 10 ≤ q,
E2−5,q = 0, q ≤ 9, 12 ≤ q,
E2−6,q = 0, q ≤ 10
E2−p,q = 0, p ≥ 7.
d2 is determined by that part of the differential d that lowers the filtration

by two. But in our case d lowers the filtration only by one. Thus d2 = 0,
and E∞ ∼ = E2 ∼ = grH ∗ (L0 ; Z). But, that is not the case for the differential d
on P (G0 ). That differential does have a component that lowers the degree
by two. However, since there is an isomorphism E2 ∼ = E2 of the second terms
of the spectral sequences for the group G0 and for the Lie algebra L0 , d2 = 0
for dimensional reasons in all cases except possibly for d2 : E2−2,4 −→ E20,3
because E2−2,4 ∼ = E2
−2,4 ∼
= Z[α], where α is represented by the element ω5 ω4 −
4ω6 ω1 − 3ω6 ω2 − 2ω6 ω3 − 4ω7 ω1 − 2ω7 ω2 , and E20,3 ∼ = E2 ∼
0,3
= Z2 [β], where
β is represented by ω3 ω2 ω1 . In order to compute the differential d2 we need
an explicit form of the isomorphism E2 ∼ = E2 . This isomorphism is induced
by the map sending ωj to the polynomial xj in P (G0 ) and the products
ωj1 ωj2 · · · ωjs to the monomials xj1 xj2 · · · xjs where j1 > j2 > · · · > js . The
technique of [1] is then used to show that d2 = 0. Thus, E3 ∼ = E2 and d3 = 0
for dimensional reasons. In fact all the higher differentials are also zero for
dimensional reasons. Then we get E∞ ∼ = grH ∗ (G0 ; Z) and the comparison
of Theorems 2 and 3 gives grH ∗ (G0 ; Z) ∼ = grH ∗ (L0 ; Z).
5. Construction of a group from a Lie algebra.

In his section we describe a method for generating examples of groups start-
ing with Lie algebras. More precisely, let L be a nilpotent torsion-free Lie
algebra over Z on a finite set of generators {g1 , . . . , gk }. Our aim is to find a
torsion-free nilpotent group G on k generators whose associated Lie algebra
L(G) is a Lie algebra over the integers. Then a canonical basis for G induces
a canonical basis for L(G) and the group structure on G is determined via
the group structure on Zk by the formula x · y = ch−1 (ch(x) ∗ ch(y)). In
some instances L(G) = L (see Example 1 and Example 3). Example 2 shows
that it is not always the case. We will show below that there is always a
group G such that L(G) is rationally equivalent to L. However, G is not

uniquely determined by L.
Starting with a Lie algebra L with a canonical basis {g1 , . . . , gk }, we first
check to see whether the basis is induced by a canonical basis of some group
G.
Assume that G is the set of elements g1x1 . . . , gkxk , x = (x1 , . . . , xk ) ∈ Zk .
Then compute ch(x) and ch(y) for x, y ∈ Zk . Note that ch(x) is completely
determined by the Lie algebra via the Campbell-Hausdorff formula (1), and
hence, so is the product ch(x) ∗ ch(y). Since ch : Zk −→ Gk is a one-to-one
map, it is enough to check whether or not ch(x) ∗ ch(y) has the form ch(z)
for some z ∈ Zk and for any choice of x and y. If the answer is yes, then G
is the desired group and {g1 , . . . , gk } is a canonical basis for G induced by
the given basis {g1 , . . . , gk } for L. This was the case in Example 3.
However, if there are elements x, y ∈ Zk such that ch(x)∗ch(y) is not equal
to ch(z) for some z ∈ Zk , then we proceed as follows: Let ξ = ξ1 g1 +· · ·+ξk gk
and η = η1 g1 + · · · + ηk gk be any two elements of LQ = the Lie algebra L
over Z, considered as a Lie algebra over Q; i.e., ξj , ηj are rational numbers.
Then by (1)
ξ ∗ η = β1 (ξ, η)g1 + · · · + βk (ξ, η)gk ;
βj (ξ, η) = βj (ξ1 , . . . , ξj−1 , η1 , . . . , ηj−1 ).
Let Nj be the smallest positive integer such that
1
βj (ξ, η) = Bj (ξ, η); Bj (ξ, η) = Bj (ξ1 , . . . , ξj−1 , η1 , . . . , ηj−1 ),
Nj
where Bj (ξ, η) is a polynomial with integer coefficients. {g1 , . . . , gk } is as-
sumed to be a canonical basis for L. Therefore, {gj , . . . , gk } generates an
ideal of L, where L = L1 ⊃ L2 ⊃ · · · is the shortest sequence of such ideals
with abelian quotients. Then {g1 , . . . , gk } can be grouped into subsets such
that {gij−1 +1 , . . . , gij } projects to a basis for Lj /Lj+1 , j = 1, 2, . . . . Then
we define the norm of each such element gs , ij−1 ≤ s ≤ ij , to be kgs k = s−1.
We assume that the norm satisfies the identities: kpgj k = kgj k for any inte-
ger p, 1 ≤ j ≤ k, and ku + vk = max{kuk, kvk} for any u, v ∈ L.
Furthermore, we set kξj k = kgj k, 1 ≤ j ≤ k and kξr ξs k = kξr k + kξs k, for
any polynomial variables. Therefore, kBj+1 (ξ, η)k = αi1 +1 kξi1 +1 k + · · · +
αj kξj k (note that kξ1 k = · · · = kξi1 k = 0). Denote the sum of the coefficients
by
εj = αi1 +1 + · · · + αj .
Now we define the subset D of the Q-vector space LQ to be the set of
elements
m1 m2 mk
g1 + g2 + · · · + gk ,
n1 n2 nk
where m1 , . . . , mk are arbitrary integers and

ε
nj+1 = nj j Nj+1 , j = 1, 2, . . . .
n1 can be any nonzero integer. We choose n1 = 1. It follows that the

operation ∗ induces on D the structure of a group. D is a uniform discrete
subgroup of the real Lie group GR and the quotient is a nilmanifold. The
above construction of D is similar to that given by Malcev [3]. For different
choices of a canonical basis for D we get different but rationally equivalent
Lie algebras.
Now we make a choice of a canonical basis for D that the associated
Lie algebra is a Lie algebra over the integers. Starting with the original
canonical basis {g1 , . . . , gk } for L we define
1
ej = gj , j = 1, 2, . . . , k.
nj
Note that for every r and s, [gr , gs ] = qur,s gu , where the summation is over
P
all u such that kgr k + kgs k < kgu k. Therefore, u > max{r, s, }. Hence,
1 1 X r,s
[er , es ] = [gr , gs ] = qu gu
nr ns nr ns
nu X r,s gu nu X r,s
= qu = qu eu ,
nr ns nu nr ns
where the qur,s are integers and nu

nr ns is also an integer since u > max{r, s}.
Therefore, the new Lie algebgra Lb is a Lie algebra over the integers that is
rationally equivalent to L. Then the group G is defined as the set of elements
{ex1 1 , . . . , exk k } with the operation (ex1 1 · · · exk k )·(ey11 · · · eykk ) = ez11 · · · ezkk given
by the formula z = ch−1 (ch(x) ∗ ch(y)).
Example 2 (continued). The new Lie algebra L, b in this case, has canoni-
cal basis {e1 = g1 , e2 = g2 , e3 = g3 , e4 = 2 g4 , e5 = 12 g5 , e6 = 12
1 1 1
g6 , e7 = 12 g7 }
and bracket
[e1 , e2 ] = 2e4 − 2e5
[e1 , e3 ] = 2e4 + 2e5 [e2 , e3 ] = 2e5
[e1 , e4 ] = 6e6 [e2 , e4 ] = 6e6 [e3 , e4 ] = 6e7
[e1 , e5 ] = 6e6 [e2 , e5 ] = 6e7
[ei , ej ] = 0 otherwise.
The group G is the set of elements {ex1 1 , . . . , ex7 7 } with the product induced
by the formula ch−1 (ch(x) ∗ ch(y)).
References
[1] B. Cenkl and R. Porter, Polynomial cochains on nilmanifolds, preprint.
[2] P. Hall, Nilpotent groups, Queens Mary College, Mathematics Notes, 1969.
[3] A. Malcev, On a class of homogeneous spaces, Izv. Akad. Nauk SSSR Ser. Mat., 13
(1949), 9-32 (Russian); English transl. in Math. USSR-Izv., 39 (1949).
Received June 11, 1998.
Northeastern University
Boston, MA 02115-5096
E-mail address: cenkl@neu.edu
Northeastern University
Boston, MA 02115-5096
E-mail address: rdp@neu.edu
Vol. 193, No. 1, 2000
RIGIDITY OF COMPACT MINIMAL SUBMANIFOLDS IN

A SPHERE
Sharief Deshmukh
In this paper we study n-dimensional compact minimal

submanifolds in S n+p with scalar curvature S satisfying the
pinching condition S > n(n − 2). We show that for p ≤ 2
these submanifolds are totally geodesic (cf. Theorem 3.2 and
Corollary 3.1). However, for codimension p ≥ 2, we prove the
result under an additional restrictions on the curvature ten-
sor corresponding to the normal connection (cf. Theorem 3.1
and Corollary 4.1). We also show that the scalar curvature S
of a non-totally geodesic n-dimensional non-negatively curved
minimal submanifold in S n+p with flat normal connection sat-
isfies n(n − p − 1) ≤ S ≤ n(n − 2) (cf. Theorem 4.1). Since
for a compact hypersurface M of S n+1 the normal connection
is flat, we use the above estimate for a scalar curvature S of
a non-negatively curved minimal hypersurface M in S n+1 to
infer that either M is totally geodesicor else it is isometric
q q
m n−m
to the hypersurface S m n
× S n−m n
. As a conse-
quence this result, we conclude that the only non-negatively
curved compact minimal hypersurfaces in S n+1 which are dif-
feomorphic to S n is totally geodesic sphere.
1. Introduction.
Let M be an n-dimensional compact minimal submanifold in a unit sphere
S n+p . One of the interesting questions in the geometry of the minimal
submanifolds of S n+p is to obtain conditions under which they are totally
geodesic. These conditions generally involve the pinching of the sectional
curvatures, Ricci curvatures, or the scalar curvature (or equivalently the
square of the length of the second fundamental form). In [3] for n ≥ 4, it
is proved that if the Ricci curvatures of a minimal submanifold in S n+p are
greater than n − 2, then M is totally geodesic, where n = dim M . For n = 3
also the above result holds and is proved in [6]. As a natural generalization of
this result one may expect that if the scalar curvature S of an n-dimensional
compact minimal submanifold of S n+p satisfies S > n(n − 2), then M is
totally geodesic. This will improve all the existing results involving pinching
of the scalar curvature for the minimal submanifolds of S n+p , and specially
31
32 SHARIEF DESHMUKH
that in [2]. However, it is not known whether this result holds even in
dimension 3. Recently in [6] a slightly courser (than expected above) result
is proved in dimension 3, namely there it is shown that for a 3-dimensional
compact minimal submanifold M of S 3+p , the condition S > 4 implies that
M is totally geodesic.
In this paper we study n-dimensional compact minimal submanifolds in
S n+p with scalar curvature satisfying the pinching condition S > n(n − 2).
We show that for p ≤ 2 these submanifolds are totally geodesic (cf. Theorem
3.2 and Corollary 3.1). However, for codimension p ≥ 2, we have to put addi-
tional restrictions on the curvature tensor corresponding to the normal con-
nection to get the result (cf. Theorem 3.1 and Corollary 4.1). We also show
that the scalar curvature S of a non-totally geodesic n-dimensional non-
negatively curved minimal submanifold in S n+p with flat normal connection
satisfies n(n − p − 1) ≤ S ≤ n(n − 2) (cf. Theorem 4.1). Since for a compact
hypersurface M of S n+1 the normal connection is flat, we use the above
estimate for the scalar curvature S of a non-negatively curved minimal hy-
persurface M in S n+1 to infer (cf. Corollary 4.2), that either M is totally
q ge-
m
pm n−m n−m
odesic or else it is isometric to the hypersurface S n × S n
considered in [2].
As a consequence of Corollary 4.2, we conclude that the only non-negativ-

ely curved compact minimal hypersurface in S n+1 which is diffeomorphic to
S n is totally geodesic sphere, giving a result in the direction of a question
asked by Yau in [7] (cf. p. 692, Problem 99).
2. Preliminaries.
Let M be an n-dimensional compact minimal submanifold of the unit sphere
S n+p . We denote by g the Riemannian metric on S n+p as well as that
induced on M , and by ∇¯ and ∇ the Riemannian connections S n+p on M
and respectively. Then we have
(2.1) ¯ X Y = ∇X Y + h(X, Y ),
∇ ¯ X N = −AN X + ∇⊥ Y,
∇ X
X, Y ∈ χ(M ), N ∈ Γ(ν), where h is the second fundamental form, χ(M ) is

the Lie algebra of smooth vector fields on M , Γ(ν) is the space of smooth
sections of the normal bundle ν of M , ∇⊥ is the normal connection and
AN is the Weingarten map corresponding to the normal N ∈ Γ(ν) which
satisfies
(2.2) g(AN X, Y ) = g(h(X, Y ), N ), X, Y ∈ χ(M ), N ∈ Γ(ν).

MINIMAL SUBMANIFOLDS 33
For the submanifold M the equations of Gauss, Codazzi and Ricci are re-
spectively
(2.3) R(X, Y )Z = g(Y, Z)X − g(X, Z)Y + Ah(Y,Z) X − Ah(X,Z) Y
(2.4) (∇h) (X, Y, Z) = (∇h)(Y, Z, X)
(2.5) R⊥ (X, Y, N1 , N2 ) = g ([ANI , AN2 ](X), Y )
for X, Y, Z ∈ χ(M ), N1 , N2 ∈ Γ(ν), where R, R⊥ are the curvature tensors

corresponding to the connections ∇ and ∇⊥ respectively and the covariant
derivative (∇h)(X, Y, Z) is given by
(∇h)(X, Y, Z) = ∇⊥
X h(Y, Z) − h(∇X Y, Z) − h(Y, ∇X Z).
The second covariant derivative (∇2 h)(X, Y, Z, W ) is defined as
(∇2 h)(X, Y, Z, W ) = ∇⊥
X (∇h)(Y, Z, W ) − (∇h)(∇X Y, Z, W )
− (∇h)(Y, ∇X Z, W ) − (∇h)(Y, Z, ∇X W )
and we have the Ricci identity
(2.6) (∇2 h)(X, Y, Z, W ) − (∇2 h)(Y, X, Z, W )
= R⊥ (X, Y )h(Z, W ) − h(R(X, Y )Z, W ) − h(Z, R(X, Y )W ).
For a local orthonormal frame {e1 , . . . , en } on M , as M is a minimal

submanifold we have
X X X
(2.7) h(ei , ei ) = 0, (∇h)(X, ei , ei ) = 0, (∇2 h)(X, Y, ei , ei ) = 0.
i i i
Let Ric be the Ricci tensor of M . Then the Ricci operator Q is a sym-
metric operator defined by
Ric (X, Y ) = g(Q(X), Y ), X, Y ∈ χ(M ).
The Gauss Equation (2.3) gives the following expression for the Ricci oper-
ator Q of the minimal submanifold M
X
(2.8) Q(X) = (n − 1)X − Ah(ei ,X) ei ,
i
P{e1 , . . . , en } is a local orthonormal frame on M . The scalar curvature

where
S = i Ric (ei , ei ) of the minimal submanifold is given by
(2.9) S = n(n − 1) − khk2 ,
where khk2 = 2
P
ij kh(ei , ej )k is the square of the length of the second
fundamental form.
34 SHARIEF DESHMUKH
Lemma 2.1. Let M be an n-dimensional compact minimal submanifold of

the unit sphere S n+p . Then
Z Xh
(2.10) k∇hk2 + R⊥ (ek , ei ; h(ej , ek ), h(ei , ej ))
M ijk
i X
− R(ek , ei ; ej , Ah(ei ,ej ) ek ) + Ric ei , Ah(ei ,ej ) ej dv = 0.
ij
Proof. Define f : M → R by f = 12 khk2 . Then the Laplacian of f given by

X
∆f = [ek ek (f ) − ∇ek ek (f )] ,
k
can be shown to satisfy
X X
∆f = g (∇2 h)(ek , ek , ei , ej ), h(ei , ej ) + k(∇h)(ei , ej , ek )k2 .
ijk ijk
Now, using the Ricci identity (2.6) and Equations (2.4), (2.7) in the above
equation and integrating the resulting equation we get the integral formula
(2.10).
Next we define kRk2 the square of the length of the curvature tensor field
of M by
X
kRk2 = kR(ei , ej )ek k2 ,
ijk
where {e1 , . . . , en } is a local orthonormal frame on M . Then using Equa-

tions (2.3) and (2.7) we immediately get:
(2.11) kRk2 = 2n(n − 1) − 4 khk2 + 2 kAh k2
X
−2 g Ah(ej ,ek ) ei , Ah(ei ,ek ) ej
ijk
P 2
where kAh k2 = ijk Ah(ei ,ej ) ek and {e1 , . . . , en } is a local orthonormal

frame on M .
X 1
(2.12) R ek , ei ; ej , Ah(ei ,ej ) ek = kRk2 − S,
2
ijk
where S is the scalar curvature of M and {e1 , . . . , en } is a local orthonormal

frame on M .
Proof. Equation (2.3) implies

Ah(ei ,ej ) ek = R(ek , ei )ej + Ah(ek ,ej ) ei − δij ek + δkj ei .
Taking inner product with Ah(ei ,ej ) ek in the above equation and noticing
that
X X
R ek , ei ; ej , Ah(ei ,ej ) ek = − R ek , ei ; ej , Ah(ek ,ej ) ei
ijk ijk
we get (2.12).
3. Minimal submanifolds with pinched scalar curvature.

Let M be an n-dimensional compact minimal submanifold of the unit sphere
S n+p . We choose a local orthonormal frames {e1 , . . . , en } on M and {N1 , ...,
Np } that of normals and define the function K ⊥ : M → R by
Xh i2
K⊥ = R⊥ (ei , ej , Nα , Nβ
ijαβ
and call it the normal curvature of the minimal submanifold. We also define
a function ϕ : M → R by
X
ϕ=2 kAα k2 kAβ k2 ,
α<β
where Aα = ANα is the Weingarten map in the direction of the normal

vector Nα .
Theorem 3.1. Let M be an n-dimensional compact minimal submanifold
of the unit sphere S n+p . If the scalar curvature S and the normal curvature
K ⊥ of M satisfy S > n(n − 2) and K ⊥ ≤ ϕ respectively, then M is totally
geodesic.
Proof. Let {e1 , . . . , en } be a local orthonormal frame on M . Then Equations
(2.3) and (2.8) imply
(3.1) Ah(ej ,ek ) ei = R(ei , ek )ej − δkj ei + δij ek + Ah(ei ,ej ) ek .
X
(3.2) Ah(ej ,ek ) ek = (n − 1)ej − Q(ej ).
k
Taking inner product in (3.1) with Ah(ei ,ej ) ek , we arrive at

X
(3.3) g Ah(ei ,ej ) ek , Ah(ej ,ek ) ei
ijk
X
= kAh k2 + R ei , ek ; ej , Ah(ei ,ej ) ek − khk2 .
ijk
36 SHARIEF DESHMUKH
Similarly Equation (3.2) gives

X
(3.4) g Ah(ej ,ek ) ek , Ah(ei ,ej ) ei = (n − 1) khk2
ijk
X
− Ric ej , Ah(ei ,ej ) ei .
ij
Now, using Equations (3.3) and (3.4) in (2.5) we arrive at

X
(3.5) R⊥ (ek , ei , h(ei , ej ), h(ej , ek )) = R ei , ek ; ej , Ah(ei ,ej ) ek
ijk
X
+ Ric ej , Ah(ei ,ej ) ei
ij
+ kAh k2 − n khk2 .
For a local orthonormal frame {N1 , . . . , Np } of normals we have
X X
Ric ej , Ah(ei ,ej ) ei = g(Aα ei , ej )Ric(ej , Aα ei )
ij αij
 
X X
(3.6) = g(Aα ei , ej )Ric  g(Aα ei , ej )ej , Aα ei 
αi j
X
= Ric(Aα ei , Aα ei ).
αi
Now, using (2.8) in (3.6) we arrive at
X X X
(3.7) Ric ej , Ah(ei ,ej ) ei = (n − 1) kAα ei k2 − kh(ej , Aα ei )k2
ij αi αij
X
= (n − 1) khk2 − kAα Aβ k2 .
αβ
Also we observe that

X X
(3.8) g Ah(ej ,ek ) ei , Ah(ei ,ek ) ej = g(Aα ei , Aβ ej )g(Aα ej , Aβ ei )
ijk ijαβ
and
X X
(3.9) kAα Aβ − Aβ Aα k2 = g ((Aα Aβ − Aβ Aα )(ei ), ej )2
αβ ijαβ
X
= −2 g(Aα ei , Aβ ej )g(Aα ej , Aβ ei )
ijαβ
X
+2 kAα Aβ k2
αβ
where we have used kAα Aβ k2 = kAβ Aα k2 which follows from the fact that
Aα and Aβ are symmetric. We use Equations (3.8) and (3.9) in (2.11) to
arrive at
kRk2 = 2n(n − 1) − 4 khk2 + 2 kAh k2

X X
+ kAα Aβ − Aβ Aα k2 − 2 kAα Aβ k2
αβ αβ
and consequently Lemma 2.3 gives

X
2 R ek , ei ; ej , Ah(ei ,ej ) ek = 2n(n − 1) − 4 khk2 + 2 kAh k2
ijk
X X
(3.10) − 2S + kAα Aβ − Aβ Aα k2 −2 kAα Aβ k2 .
αβ αβ
The integral formula (2.10) in view of (3.5) takes the form

Z  X
(3.11) k∇hk2 + kAh k2 − n khk2 − 2 R ek , ei ; ej , Ah(ei ,ej ) ek
M ijk

X 
+2 Ric ej , Ah(ei ,ej ) ei dv = 0.

ij
Thus using (3.7), (3.10) and (2.9) in (3.11), we conclude

 
Z  X 
(3.12) k∇hk2 + n khk2 − kAh k2 − kAα Aβ − Aβ Aα k2 dv = 0.
M αβ

Note that in view of Equation (2.5) we have
Xh i2 X
(3.13) K⊥ = R⊥ (ei , ej , Nα , Nβ ) = g ([Aα , Aβ ](ei ), ej )2
ijαβ ijαβ
 
X X
=  g ((Aα Aβ − Aβ Aα )(ei ), ej )2 
αβ ij
X
= kAα Aβ − Aβ Aα k2 .
αβ
38 SHARIEF DESHMUKH
Also we have
X 2 X
kAh k2 = h(ei ,ej ) k =
A e g (Aα ei , ej )2 kAα ek k2

ijk ijkα
X X
= kAα k g (Aα ei , ej )2 =
2
kAα k4
ijα α
!2
X X
= kAα k2 −2 kAα k2 kAβ k2 ,
α α<β
which gives
X
(3.14) kAh k2 = khk4 − 2 kAα k2 kAβ k2 .
α<β
Using Equations (3.13) and (3.14) in (3.12), we arrive at

Z n o
(3.15) k∇hk2 + n − khk2 khk2 + ϕ − K ⊥ dv = 0.
M
Now using S > n(n − 2) with (2.4), we conclude that khk2 < n and hence
the integral formula (3.15) implies that M is totally geodesic.

of the unit sphere S n+2 . If the scalar curvature S of M satisfies S > n(n−2),
then M is totally geodesic.
Proof. In codimension 2 the integral formula (3.12) together with (3.14)

gives
Z n o
k∇hk2 + n − khk2 khk2 + 2 kA1 k2 kA2 k2 −kA1 A2 − A2 A1 k2 dv = 0,
M
where {N1 , N2 } is a local orthonormal frame of normals on M and Aα =

ANα , α = 1, 2. Then using Lemma 1 in ([2], p. 64), and the hypothesis
S > n(n − 2), we get from above integral formula that khk2 = 0, that is M
is totally geodesic.
Remark 3.1. For Veronese surface M in S 4 , one can show that the normal
curvature K ⊥ and the function ϕ appearing in Theorem 3.1 satisfy K ⊥ =
ϕ = 98 . Indeed we can choose a local orthonormal frame {e1 , e2 } on M and
a local orthonormal frame {N1 , N2 } of normals such that the Weingarten
maps AN1 , AN2 take the form (cf. [2])

0 1 1 0
A1 = , A2 = .
1 0 0 −1
Remark 3.2. For a compact minimal hypersurface M in S n+1 we have

K ⊥ = 0 and ϕ = 0 and consequently the integral formula (3.15) takes the
form
Z n o
k∇Ak2 + n − kAk2 kAk2 dv = 0,
M
where A is the Weingarten map of the hypersurface. Thus we have:
Corollary 3.1. Let M be an n-dimensional compact minimal hypersurface

of the unit sphere S n+1 . If the scalar curvature S of M satisfies S ≥ n(n−2),
then M is either totally
q geodesic or else is isometric to the hypersurface
m
pm n−m n−m
S n ×S n .
Proof. If M is not totally geodesic, then the above integral formula gives
that the Weingarten map A is parallel and kAk2 = n, consequently
q as in [2],
m
pm n−m n−m
M is isometric to the hypersurface S n ×S n of S n+1 .
4. Minimal submanifolds with flat normal connection.

In this section we study compact minimal submanifolds of the unit sphere
S n+p which have flat normal connection, that is R⊥ = 0. First, we prove the
following theorem which estimates the scalar curvature of such submanifolds
which are non-totally geodesic and non-negatively curved.
Theorem 4.1. Let M be an n-dimensional compact non-negatively curved

minimal submanifold of the unit sphere S n+p with flat normal connection.
Then either M is totally geodesic or else the scalar curvature S of M satisfies
n(n − p − 1) ≤ S ≤ n(n − 2).
Proof. Since the normal connection is flat we have R⊥ = 0 and that all the
Weingarten maps AN , N ∈ Γ(ν), can be diagonalized with respect to the
same local orthonormal frame {e1 , . . . , en } (cf. [1], p. 127). Choose a local
orthonormal frame {N1 , . . . , Np } of normals such that Aα (ei ) = λαi ei , where
40 SHARIEF DESHMUKH
λαi are the smooth functions and α = 1, . . . , p. Then we have

X X
Ric ej , Ah(ei ,ej ) ej − R ek , ei ; ej , Ah(ei ,ej ) ek
ij ijk
 
X X X
=  g (Aα ei , ej ) Ric (ei , Aα ej ) − g (Aα ei , ej ) R (ek , ei ; ej , Aα ek )
α ij ijk
 
X X X
=  Ric (Aα ej , Aα ej ) − R (ek , Aα ej ; ej , Aα ek )
α j jk
 
X X X
=  R (ek , Aα ej ; Aα ej , ek ) − R (ek , Aα ej ; ej , Aα ek )
α jk jk
X
(λαj )2 Kkj − λαj λαk Kkj

=
αjk
where Kkj = R(ek , ej ; ej , ek ) is the sectional curvature of the plane section

spanned by {ek , ej }. Consequently we have
X X
(4.1) Ric ei , Ah(ei ,ej ) ej − R ek , ei ; ej , Ah(ei ,ej ) ek
ij ijk
 
1 X X
=  2(λαj )2 Kkj − 2 λαj λαk Kkj 
2
αjk αjk
 
1 X α 2 X X
= (λj ) Kkj + (λαk )2 Kkj − 2 λαj λαk Kkj 
2
αjk αjk αjk
 
1 X
=  (λαj − λαk )2 Kkj  ≥ 0.
2
αjk
Then the integral formula (2.10) with R⊥ = 0 and (4.1) gives
(4.2) k∇hk = 0,
X X
Ric ei , Ah(ei ,ej ) ej − R ek , ei ; ej , Ah(ei ,ej ) ek = 0.
ij ijk
Now Equation (3.5) with R⊥ = 0 and the second equation in (4.2) yields
(4.3) kAh k2 − n khk2 = 0.

Also, we have
!2
X X X
(4.4) kAh k2 = kAα k4 = kAα k2 −2 kAα k2 kAβ k2
α α α<β
X
= khk4 − 2 kAα k2 kAβ k2 .
α<β
Combining (4.3) and (4.4), we arrive at

(4.5) khk2 khk2 − n ≥ 0.
P
Also from (4.3), we have α kAα k4 − n kAα k2 = 0, or equivalently
X n 2 n2 p
(4.6) kAα k2 − = .
α
2 4
Now using Schwarz inequality, we get
" #2
X 2 n 2 1 X 2 n 1 np 2
(4.7) kAα k − ≥ kAα k − = khk2 − .
α
2 p α 2 p 2
Combining (4.6) and (4.7), we conclude that

(4.8) khk2 khk2 − np ≤ 0.
This proves that either M is totally geodesic or khk2 ≤ np. If M is not

totally geodesic, then from khk2 ≤ np together with (4.5) we get n ≤ khk2 ≤
np, which by Equation (2.9) is equivalent to n(n − p − 1) ≤ S ≤ n(n − 2).
Next we prove the following:
of the unit sphere S n+p with flat normal connection. If the scalar curvature S
of M satisfies S ≥ n(n − 2), then M is qeither totally
geodesic or is isometric
m
pm n−m n−m
to the hypersurface S n ×S n .
Proof. Since R⊥ = 0, Equation (3.5) gives

X X
(4.9) Ric ei , Ah(ei ,ej ) ej − R ek , ei ; ej , Ah(ei ,ej ) ek
ij ijk
= n khk2 − kAh k2 .
Thus the integral formula (2.10) in view of (4.9) and R⊥ = 0, takes the form
Z n o
(4.10) k∇hk2 + n khk2 − kAh k2 dv = 0.
M
42 SHARIEF DESHMUKH
For a local orthonormal frame {N1 , . . . , Np } of normals, using the expression

for kAh k2 as given in (4.4) into (4.10) we arrive at
 
Z  X 
(4.11) k∇hk2 + n − khk2 khk2 + 2 kAα k2 kAβ k2 dv = 0.
M α<β

SinceS ≥ n(n− 2), that is khk2 ≤ n, above integral gives k∇hk = 0,

khk2 n − khk2 = 0, α<β kAα k2 kAβ k2 = 0. The last two equations give
P
either M is totally geodesic or else khk2 = n and Aβ = 0, β = 2, . . . , p. In

the second case, weqhave asin [2] that M is isometric to the hypersurface
m
pm n−m n−m
S n ×S n of S n+1 .
Now as applications of these theorems, first we have:
Corollary 4.1. Let M be an n-dimensional compact minimal submanifold
of the unit sphere S n+p with flat normal connection. If the scalar curvature
S of M satisfies S > n(n − 2), then M is totally geodesic.
Corollary 4.2. Let M be a non-negatively curved compact minimal hyper-
surface of S n+1 . Then M is either
q totally
geodesic or is isometric to the
m
pm n−m n−m
hypersurface S n ×S n .
Proof. Since M is hypersurface, the normal connection is flat, by Theo-

rem 4.1, we have either M is totally geodesic or the scalar curvature is
given by S = n(n − 2). If M is non-totally geodesic, we have khk2 = n
and consequently by
qthe result
in [2], M is isometric to the hypersurface
m
pm n−m n−m
S n ×S n .
Remark 4.1. In [2], it is asked that the values of khk2 should determine
the compact hypersurfaces up to rigid motion in the ambient sphere S n+1
(cf. p. 75). The above Corollary shows that for non-negatively curved
hypersurfaces in S n+1 , khk2 takes only two values 0 and n and the corre-
sponding hypersurfaces are known. Thus the Corollary 4.2 can be thought
of as a result in the direction of this question.
We also have the following consequence of Corollary 4.2:
Corollary 4.3. Let M be a non-negatively curved compact minimal hyper-
surface of S n+1 . If M is diffeomorphic to S n , then M is totally geodesic.
Remark 4.2. One of the important questions concerning the hypersurfaces
of S n+1 is to show that “the only embedded minimal hypersurface in S n+1
which is diffeomorphic to S n is the totally geodesic sphere” (cf. [7], p. 692,
Problem-99). Corollary 4.3 can be considered as a result in this direction
for the class of non-negatively curved minimal hypersurfaces in S n+1 .
Finally, we prove:
of the unit sphere S n+p with flat normal connection. If M is of constant
sectional curvature c ≥ 0, then either M is flat (c = 0) and p ≥ (n − 1), or
totally geodesic.
Proof. Let {e1 , . . . , en } be a local orthonormal frame on M . Then for the
minimal submanifold M of constant sectional curvature c, Equation (2.3)
can be restated as
(c − 1) (δjk ei − δik ej ) = Ah(ej ,ek ) ei − Ah(ei ,ek ) ej .
Taking inner product with Ah(ej ,ek ) ei and summing the resulting equation
we arrive at
X
(4.12) (1 − c) khk2 = kAh k2 − g Ah(ei ,ek ) ej , Ah(ej ,ek ) ei .
ijk
Now as the normal connection is flat, R⊥ = 0 and with the help of

Equation (2.5) we compute
X X
g Ah(ei ,ek ) ej , Ah(ej ,ek ) ei = g Ah(ej ,ek ) ej , Ah(ei ,ek ) ei
ijk ijk
 
X X
= g Ah(ej ,ek ) ej , Ah(ei ,ek ) ei 
ik j
X
= g (n − 1)ek − Q(ek ), Ah(ei ,ek ) ei
ik
X
= (n − 1) khk2 −

Ric ek , , Ah(ei ,ek ) ei .
ik
However for M the Ricci curvature satisfies Ric (X, Y ) = (n − 1)cg(X, Y )

and consequently,
X
g Ah(ei ,ek ) ej , Ah(ej ,ek ) ei = (n − 1)(1 − c) khk2 .
ijk
Using this equation in (4.12), we get

(4.13) kAh k2 = n(1 − c) khk2 .
Note that Equation (4.3) holds for non-negatively curved compact minimal
submanifolds of S n+p with flat normal connection. Thus combining (4.3)
and (4.13), we conclude that c khk2 = 0 and this proves that either c = 0
or M is totally geodesic. However if c = 0, then S = 0 and consequently by
Thorem 4.1, n(n − p − 1) ≤ 0 (as M is non-totally geodesic), which implies
p ≥ n − 1.
44 SHARIEF DESHMUKH
Corollary 4.4. There is no compact flat minimal hypersurface in S 4 .
References
[1] B.Y. Chen, Total Mean Curvature and Submanifolds of Finite Type, World Scientific,
1983.
[2] S.S. Chern, M. DoCarmo and S. Kobayashi, Minimal submanifolds of a sphere with
second fundamental form of constant length, Functional Analysis and Related Topics,
Springer Verlag, (1970), 59-75.
[3] N. Ejiri, Compact minimal submanifolds of a sphere with positive Ricci curvature, J.
Math. Soc. Japan, 31 (1979), 251-256.
[4] K. Moshimo, Minimal immersions of 3-dimensional sphere in to spheres, Osaka J.
Math., 21 (1984), 721-732.
[5] J. Simon, Minimal varieties in Riemannian manifolds, Ann. of Math., 88(2) (1968),
62-105.
[6] Y.B. Shen, Curvature pinching for three-dimensional minimal submanifolds in a sphere,
Proc. Amer. Math. Soc., 115(3) (1992), 791-795.
[7] S.T. Yau, Problem section, Seminar on Differential Geometry, Princeton Univ. Press,
1982.
Received July 15, 1998.
King Saud University

Riyadh-11451
Saudi Arabia
E-mail address: shariefd@ksu.edu.sa
Vol. 193, No. 1, 2000
THE COBORDISM OF ORIENTED THREE DIMENSIONAL

ORBIFOLDS
K.S. Druschel
In this paper we construct a decomposition for oriented orb-

ifold cobordism and apply it to establish that every oriented
three dimensional orbifold bounds an oriented four orbifold.
Introduction.
In a previous paper [2], using orbifold Pontrjagin numbers, we had estab-
lished that rationally every odd dimensional oriented orbifold bounds and,
also, if an orbifold bounds, then some multiple of it bounds with no increase
in the set of local groups. The proof involves systematically making the
orbifold rationally cobordant to one with a smaller set of local groups. Thus
to study torsion in orbifold cobordism one must gain control of the way
one introduces new local groups when one tries to make the singularities
associated with the original orbifold bound. In Section one we collect some
definitions associated with orbifolds and their cobordism. In Section two
we consider various orbifold cobordism groups with restrictions on the set
of local groups and fit these into a commutative diagram with cobordism
groups of certain orbibundles over manifolds. This is much in the spirit of
[1]. In Section three we apply this to show that every two dimensional and
every three dimensional orbifold bounds.
1. Orbifold Cobordism.
In this section we briefly review some material associated with orbifolds and
cobordism.
An orbifold chart on a topological space X is defined by a triple (U, G, π)
where U is an open subset of Rn , G is a a finite group acting effectively and
differentiably on U , and π : U → X factors as hpU where pU is the natural
orbit map and h is open and injective. Two charts (Ui , Gi , πi ), i = 1, 2, are
compatible if for x ∈ π1 (U1 ) ∩ π2 (U2 ) and x̄ a lift of x in U1 there is an open
set U ⊂ U1 about x̄ and a diffeomorphism f : U → U 0 onto an open subset
U 0 ⊂ U2 such that π2 f = π1 . Such a map is called an overlap map. An n
dimensional orbifold Q is defined by a Hausdorff underlying space, denoted
|Q|, together with a maximal atlas of compatible orbifold charts on |Q|.
45
46 K.S. DRUSCHEL
An n dimensional orbifold with boundary is defined as above except that

U is an open subset of Hn = {(x1 , x2 , . . . , xn ) ∈ Rn : xn ≥ 0}. If W
is an orbifold with boundary, the orbifold boundary of W , denoted δo W ,
has underlying space {w ∈ |W | : w = πα (x), x ∈ δHn } where πα is the
map for some chart for W . An orbifold Q is orientable if it has an atlas
{(Uα , Gα , πα ), α ∈ Λ} such that the Uα ’s can all be oriented consistently
with the G0α s and the overlap maps orientation preserving. It is oriented
once one chooses a consistent orientation of the charts.
Two compact, oriented n dimensional orbifolds Q1 , Q2 , are orbifold cobor-
dant`if there is an oriented orbifold with boundary W , such that δo W =
Q1 −Q2 with the appropriate orientations. As with manifold cobordism
(Ωn ) one has the group of cobordism classes of oriented n dimensional orb-
ifolds, which we denote by Ωn,orb .
For x ∈ |Q| the local group at x is the isotropy group Gx̂ for a chart
(U, G, π) with x̂ ∈ U a lift of x. Using the differential, Gx may be considered
as a subgroup of Gl(n, R) and is then well defined up to conjugacy. We
denote this conjugacy class by (Gx ) and the set of conjugacy classes of local
groups for Q by GQ . If Q has a Riemannian metric, which is possible if |Q|
is compact, then we can view Gx as a subgroup of O(n), and then, if Q is
oriented, as a subgroup of SO(n). For a subgroup K of Gl(n, R) let j be
the dimension of the subspace of Rn on which K acts trivially. The degree
of K (deg K) is n − j. For (H) ∈ GQ the H singular set of Q , denoted QH ,
is defined to be the set {x ∈ Q|(H) < (Gx )}.
We will be particularly interested in two examples of (F, G) orbibundles
[2, 1.8]. For an orbifold Q with atlas {(Uα , Gα , πα ), α ∈ Λ} an l dimensional
orbivector bundle Ψ over Q has total space (T ot Ψ) an orbifold E which
locally looks like Uα ×Gα Rl where the action of Gα and the overlap maps
for E are linear on the second factor. If these are also orthogonal, then we
can associate with Ψ the sphere and disc bundles SΨ and DΨ. If they are
orientation preserving then we say Ψ is oriented.
To study the H singular set in an orbifold Q we look at the normal bundle
νH [2, p. 310], an (Rl /H, NGl(l,R) (H)/H) bundle over an orbifold QH . Here
NK (G) denotes the normalizer of the group G in K, l is the degree of H, and
QH is the best orbifold approximation to QH [2, 2.9]. When H is maximal
in GQ QH = QH and T ot νH is diffeomorphic to a neighborhood of QH in Q.
When |Q| is compact we can replace Gl(l, R) with O(l) and have associated
sphere and disc bundles, SνH and DνH .
2. A Decomposition for Oriented Orbifold Cobordism.

Throughout this we assume that every orbifold is compact, smooth, and
oriented unless stated otherwise.
From [2, 2.14] and [2, 1.15] we conclude the following.
THE COBORDISM OF ORIENTED THREE DIMENSIONAL ORBIFOLDS 47
Lemma 2.1. If an orbifold Q bounds then some multiple s of Q bounds an

orbifold W such that GQ = GW , i.e. sQ bounds with no increase in local
groups.
Now suppose (H) is maximal in GQ . Then QH is a manifold and the
normal bundle to QH in Q is an ordinary (Rn /H, NO(n) (H)/H) bundle. [2,
p. 305]. (Here n is the degree of H.) Also when sQ = δorb W with GQ = GW ,
we have the normal bundle to W H in W is an ordinary (Rn /H, NO(n) (H)/H)
bundle over the manifold W H and an extension of the normal bundle to sQ.
Hence one wants to first consider the following.
Definition 2.2. Suppose H < SO(n) has degree n. We let Bk (H) be the
set of cobordism classes of (Rn /H, NO(n) (H)/H) bundles over k-dimensional
manifolds where we require that the total space of the bundle is oriented.
By using the classifying maps for such bundles we have that Bk (H) is iso-
morphic to Ωk,t (BNO(n) (H)/H) [2, 2.3] (or Ωk (BNO(n) (H)/H) if NO(n) (H)
< SO(n)). Here Ωk (X) denotes the group of bordism classes of maps from
oriented k dimensional manifolds into the space X. If there is some canon-
ical orbibundle Γ over X then Ωk,t (X) consists of bordism classes of maps
f : M k → X such that f ∗ Γ ⊕ T M is oriented. Ωk,t (BNO(n) (H)/H) is ra-
tionally isomorphic to the twisted homology H∗,t (BNO(n) (H)/H, Ω∗ ). [2,
2.4].
Once we are asking whether an orbifold Q bounds exactly and not just
rationally, we find that we can no longer restrict the local groups of the
bounded orbifold W to be those of Q. In other words if Q = δorb W it
may be necessary that GW is strictly larger than GQ . This is born out by
the following illustration. This illustration is due to Quach Ngoc Du, as
communicated to me by Mike Davis, as is a generalization to prove that an
oriented two dimensional orbifold bounds.
Figure 2.1.
48 K.S. DRUSCHEL
W is topologically a three disc. As an orbifold it has singular set a half

open line segment labelled Cn , the cyclic group of order n, a point labelled by
the dihedral group Dn , and an open line segment labelled C2 . The orbifold
boundary of W is a teardrop orbifold with singular set a point labelled Cn .
For H = Cn , W H is a folded closed line segment with its fold point labelled
C2 .
Thus we see that W H need not be a manifold and that Bk (H) is too re-
strictive. However, in general, if QH bounds an orbifold T and the (Rn /H,
NO(n) (H)/H) bundle over QH extends to an (Rn /H, NO(n) (H)/H) orbi-
bundle Ψ over T , then Q is orbifold cobordant to Q − (int (DνH )) ∪SνH SΨ.
(See the proof of 2.8.) For the resulting orbifold one has local groups those
of Q with H thrown out, and then those of SΨ added in. One wants to gain
some control of which local groups are added in by SΨ so we consider the
following.
Let F denote a set of conjugacy classes of finite subgroups of SO(m). F
is closed if whenever (K) ∈ F then (L) ∈ F for every (L) ∈ GRm /K . Let H
be a finite linear group of degree l and F̂ ⊃ F indicate any two closed sets
of classes of finite subgroups of SO(l + k + 1) and SO(l + k), respectively,
with (H) ∈ F.
Definition 2.3. Suppose (H) is maximal in F. Define Bk (H, F, F̂) to be
the equivalence classes of pairs (M, ν) where ν is an (Rl /H, NO(l) (H)/H)
bundle over a k dimensional manifold M with the total space of ν oriented.
The equivalence relation is defined by (M1 , ν1 ) ∼ (M2 , ν2 ) if there is an
(Rl /H, NO(l) (H)/H) orbibundle Ψ over an orbifold W so that
i) T ot Ψ is oriented
ii) GT ot`
Ψ ⊂ F̂, GSΨ ⊂ F − (H)
iii) M1 M2 = δorb W (not necessarily orientably)
iv) Ψ|Mi = νi and T ot νi has the inherited orientation.
Note that if G ∈ GT ot Γ for Γ an (Rl /H, NO(l) (H)/H) orbibundle over an
orbifold Qn with T ot Γ oriented then G < NSO(n+l) (H) and hence
G < NO(l) (H) ⊕ O(n). Also, tracing through the definition of an (Rl /H,
NO(l) (H)/H) orbibundle one sees that ker ρ2|G = H. Here ρi : O(l) ⊕
O(n) → O(s), i = 1, 2 is projection onto the first, respectively second factor
with s = l, respectively n.
For Ψ as in the above definition, the second part of ii) implies that if
G ∈ GT ot Ψ and (K) ∈ GRl+k+1 /G with deg ρ1 (K) < l (so K shows up as
a local group for SΨ) then (K) ∈ F − (H). Thus we have the following
definition.
Definition 2.4. Let F be as above and H < G < NO(l) (H) ⊕ O(k + 1) with
ker ρ2|G = H . Then G satisfies condition SF ,H if for every (K) ∈ GRl+k+1 /G
with deg ρ1 (K) < l, (K) ∈ F − (H).
Now let F ⊂ F̂ be a pair of closed sets of conjugacy classes of finite

subgroups of SO(n), respectively SO(n + 1).
Definition 2.5. Let Ωn,orb (F, F̂) = {Qn : GQ ⊂ F}/ ∼, where Q1 ∼ Q2 if

`
there is an orbifold W with GW ⊂ F̂ and Q1 −Q2 the oriented orbifold
boundary of W .
Let Fall,m be the set of all conjugacy classes of finite subgroups of SO(m)
and Ωn,orb (F) = Ωn,orb (F, Fall,n+1 ).
Proposition 2.6.
i) Ωn,orb = Ωn,orb (Fall,n ).

ii) Since GQ is finite for a compact orbifold Q, Ωn,orb (F, F̂) = ∪Ωn,orb (Fα ,
F̂) where the union is over all closed finite subsets Fα of F.
iii) If F1 ⊂ F2 ⊂ F̂ then Ωn,orb (F1 , F̂) ⊂ Ωn,orb (F2 , F̂).
iv) If F ⊂ F̂1 ⊂ F̂2 then the natural map τ : Ωn,orb (F, F̂1 ) → Ωn,orb (F, F̂2 )
is onto.
From Lemma 2.1 we conclude:
Proposition 2.7. ker τ ⊂ torsion (Ωn,orb (F, F̂1 )).
The various cobordism groups, Bk (H, F, F̂), of Rn /H bundles, and the

orbifold cobordism groups with various restrictions on the local groups fit
together according to the following commutative diagram.
Here g1 , g2 , g3 are inclusion maps, hi [Q] = [(QH , νH )] and ki [(M, ν)] =
[Sν], and [(M, ν)] or [Q] denotes the appropriate cobordism class.
Also we must assume that H is maximal in F and if L ∈ F̂ with H < L
then NL (H) has condition SF ,H as in Definition 2.4. The latter ensures
that h2 is well defined. Indeed, suppose [Qi ] ∈ Ωn,orb (F, F̂), i = 1, 2, with
[Q1 ] = [Q2 ] via the orbifold with boundary W . If the normal bundle of
WH in W is denoted by Ψ, then the local groups of T ot Ψ are of the form
NGx (H), x ∈ WH , and condition SF ,H guarantees that [(QH H
1 , νH (Q1 ))] =
[(QH H
2 , νH (Q2 ))] in Bn−deg H (H, F, F̂) via (WH , Ψ).
50 K.S. DRUSCHEL
Ωn,orb (F − (H), F) Ωn,orb (F − (H), F̂) Ωn,orb (F − (H))

g1 g2 g3
? ? ?
Ωn,orb (F, F) Ωn,orb (F, F̂) Ωn,orb (F)
h1 h2
? ?
Bn−deg H (H) Bn−deg H (H, F, F̂)
k1 k2
?
Ωn−1,orb (G(S n−1 /H) , F − (H))
The following lemma allows us to tell when an orbifold Q with local groups
in F is cobordant to one with local groups in F − (H) by examining the
normal bundle to QH . The commutivity of the diagram allows more control
of the local groups in the cobordism.
Lemma 2.8. The two columns in the above diagram are exact.
Proof. To prove ker hi ⊂ im gi , note that if [(QH , νH )] = 0 in Bn−deg H (H,
F, F̂) via (W, Ψ), then the local groups of SΨ are in F − (H). Thus Q
is cobordant to (Q − DνH ) ∪SνH SΨ, whose local groups are in F − (H),
since those of Q − DνH and SΨ are. This cobordism is via (Q × I) ∪(DνH ,1)
DΨ, which has local groups in F̂, since both Q and T otΨ do. The reverse
inclusion of image and kernal is true vacuously. We have im hi ⊂ ker ki since
SνH bounds Q − DνH . If for [(M, Ψ)] ∈ Bn−deg H (H, F, F̂), ki [(M, Ψ)] = 0,
then SΨ = δorb W with GW ⊂ F − (H). Hence (M, Ψ) ∼ (QH , νH ) where
Q = W ∪SΨ DΨ.
3. Applications.
We illustrate Lemma 2.8 with a proof of the following. Consideration of
Figure 2.1 leads to an alternative proof of this proposition.
Proposition 3.1. Ω2,orb = 0.
Proof. If F is a finite set of conjugacy classes of finite subgroups of SO(2) it
consists of cyclic subgroups, say Cn1 , . . . Cnk . Suppose C2 ∈ F. If H = Cni ,
ni 6= 2, and F̂ = F ∪ Dni then Dni satisfies condition SF ,H . We note
that in general B0 (H, F, F̂) is generated by the trivial Rl /H bundle over

a point. Here this bundle bounds the bundle Ψ with total space I ×Z2
(R2 /Cni ) where Z2 folds I and is reflection across the x axis for R2 and
GSΨ = {(C2 ), (e)}. Thus B0 (H, F, F̂) = 0 and gi , i = 2, 3, are isomorphisms
giving that Ω2,orb (F) ∼ = Ω2,orb (F − (Cni )). Repeating this argument for
F − (Cni ) we get Ω2,orb (F) ∼ = Ω2,orb ((C2 ), (e)). From Proposition 2.6 ii)
we have Ω2,orb ∼ = Ω2,orb ((C2 ), (e)). Every two dimensional orbifold Q with
points x1 , . . . , xk labelled C2 is cobordant to ` another orbifold Q1 with an
2
even number of points labelled C2 via (Q×I k(I ×Z2 (R /Cni )))/ ∼ where
each [(0, R2 /Cni )] is glued to a chart neighborhood of an xi in Q ` × 1. Q1 is
cobordant to a manifold with k more handles than Q via Q1 × I k(I ×Z2
(R2 /Cni ))/ ∼ where the ends of the I × (R2 /Cni )’s are glued to chart
neighborhoods of the singular points in Q1 × 1. Thus Ω2,orb ((C2 ), (e)) ∼ =
Ω2 = 0.
Theorem 3.2. Every compact, oriented, three dimensional orbifold bounds
an oriented, compact orbifold.
Proof. The set of conjugacy classes in O(3) of all finite subgroups of SO(3)
consists of two infinite families coming from the cyclic and dihedral groups,
and three other classes from the tetrahedral group, T , the octahedral group,
O, and the icosahedral group, I.
Let F be any set of conjugacy classes that contains I and the dihedral
groups Ds , s = 2, 3, 5, and F̂ = F ∪ (I) where I = hI, −id4 i. Here −id4
refers to the 4 × 4 diagonal matrix with -1’s down the diagonal. Then
(F, F̂) satisfies the conditions for Lemma 2.8 with H = I. We see that
B0 (I, F, F̂) = 0 as the R3 /I bundle over a point bounds the R3 /I bundle
Ψ over [−1, 1]/Z2 where T ot Ψ = [−1, 1] ×h−id4 i R3 /I and SΨ has maximal
local groups Ds , s = 2, 3, 5, and thus local groups in F −(H). (See Figure 3.1
below for a picture of the singular set of SΨ, where the singular set of S 3 /I
on the left is folded by −id4 to form that of SΨ.) From Lemma 2.8 we thus
have g2 is an isomorphism and then so is g3 : Ω3,orb (F − (I)) → Ω3,orb (F).
In particular Ω3,orb ∼= Ω3,orb (Fall,3 − (I)).
Replacing I with O in the above paragraph and letting s = 2, 3, 4 and
F = Fall,3 − (I), yields Ω3,orb (Fall,3 − (I)) ∼ = Ω3,orb (Fall,3 − (I) − (O)).
Applying this method again with H = T and s = 2, 3, we obtain Ω3,orb ∼ =
Ω3,orb (F1 ) where F1 consists of the classes of dihedral and cyclic groups.
We cannot use the above method to get rid of Dn . Also when H = Dn
and (H) ∈ F ⊂ F1 the map k1 : B0 (H) → Ω2,orb (GS 2 /H , F − (H)) is nonzero
as S 2 /Dn cannot bound with local groups in F − (H). (Indeed the point
labelled Cn in S 2 /Dn can only bound a folded line segment labelled Cn and
such a line segment can only possibly fit into Dn ,I,O, or T , none of which
are in F − (H).) Hence, by Lemma 2.8, h1 : Ω3,orb (F, F) → B0 (H) is the
zero map. In other words, for Q with local groups in F the normal bundle
52 K.S. DRUSCHEL
to QH bounds in B0 (H), meaning there are even number of points labelled

(H). Thus gi , i = 1, 2, 3, are isomorphisms. Then using Proposition 2.6 ii)
we have Ω3,orb (F1 ) ∼
= Ω3,orb (F2 ) where F2 consists of the cyclic groups in
SO(3).
Now B1 (Ck ) is generated by R2 /Ck bundles over S 1 whose total space is
oriented. These bundles are trivial (since π1 (B(NSO(2) (Ck )/Ck )) = π1 (BS 1 )
= 0) and so bound D2 × R2 /Ck . This gives us that B1 (Ck ) = 0 and hence
g1 and g3 are isomorphisms. Thus Ω3,orb ∼ = Ω3,orb (F2 ) ∼
= Ω3 , which is zero
[3].
Singular set for S 3 /I. Singular set for S 3 /hI, −id4 i.
Figure 3.1.
We now illustrate the proof of Theorem 3.2 by showing how a three orb-
ifold Q with singular set as in Figure 3.2 is cobordant to a three dimensional
manifold. Here unlabelled arcs have singular set C2 . Figure 3.3 gives the
singular set of a four dimensional orbifold with boundary which makes Q
cobordant to an orbifold Q1 with no local groups I, as in the first para-
graph of the proof. In Figure 3.3 the singular set around the point labelled
hI, −id4 i consists of four planes labelled by Ck , k = 3, 5, 2, 2, meeting in
lines labelled by Dk , k = 3, 5, 2, and I. Figure 3.4 shows the singular set
of Q1 and another orbifold Q2 which is cobordant to Q1 and has dihedral
and cyclic local groups. Note that there is an even number of points in Q2
labelled by each dihedral group so, as in the third paragraph of the proof, we
can systematically make Q2 cobordant to an orbifold with fewer dihedral lo-
cal groups. Figure 3.5 shows the singular sets of Q3 and Q4 , both cobordant
to Q2 . In Q3 the dihedral groups D3 have been removed, and in Q4 the only
local dihedral groups are D2 . Q4 is cobordant to an orbifold Q5 with singular
set two circles labelled C2 , two labelled C3 , and one labelled C`
5 . This is via
(Q4 × I) ((D2 /D2 ) × I)/ ∼ where, since DνD2 (Q4 ) ∼ = D2 /D2 D2 /D2 , we
`
identify DνD2 (Q4 ) with the ends of (D2 /D2 )×I). One uses similar orbifolds
with boundary for the cobordisms between Q2 and Q3 , and then between Q3
and Q4 . Finally, one makes Q5 cobordant to a three manifold via the orb-
ifold (Q5 × I) ∪Y X where X consists of a disjoint union of D2 × (D2 /Ck )’s,
k = 3, 3, 2, 2, 5, and Y is diffeomorphic to a disjoint union of S 1 × (D2 /Ck )’s,
k = 3, 3, 2, 2, 5.
Figure 3.2. Singular set for Q.
Figure 3.3. Singular set for (Q × I) ∪(DνI ,1) DΨ.

54 K.S. DRUSCHEL
Singular set for Q1 . Singular set for Q2 .
Figure 3.4.
Singular set for Q3 . Singular set for Q4 .
Figure 3.5.
References
[1] P.E. Conner and E.E. Floyd, Differentiable Periodic Maps, Springer, Berlin, 1964.
[2] K.S. Druschel, Oriented Orbifold Cobordism, Pacific J. Math., 164(2) (1994), 299-319.
[3] V.A. Rohlin, A three-dimensional manifold is the boundary of a four-dimensional one,
Dokl. Akad. Nauk SSSR, 81 (1951), 355-357.
Received September 7, 1996 and revised June 29, 1998.
St. Louis University

St. Louis, MO 63103
E-mail address: druschelks@slu.edu
Vol. 193, No. 1, 2000
TWO-SIDED BRAID GROUPS AND ASYMPTOTIC

INCLUSIONS
Juliana Erlijman
We show that the two-sided inclusions of braid subfactors

of type A coincide with the asymptotic inclusions for the one-
sided pairs. We also include an elementary method for ob-
taining the higher relative commutants of the one-sided pairs,
which were already computed by H. Wenzl.
Introduction.
In [W-1] and [W-2] Wenzl constructed new examples of subfactors of the
hyperfinite II1 factor. He considered unitary representations ρ of the infinite
braid group B∞ , whose restrictions to finite braid groups Bn generate finite
dimensional C∗ -algebras, with additional properties such as the existence of
a positive Markov trace on the quotients. His pairs of subfactors are given
by the von Neumann algebras generated by {gi : i ∈ N, i > m} and by
{gi : i ∈ N0 } in the trace representation, where gi := (πtr ◦ ρ)(σi ), σi are
the braid generators for i ∈ N0 , and m ∈ N0 is arbitrary.
These special unitary braid representations can be obtained in connection
with the representation theory of the classical Lie algebras. The ones corre-
sponding to the Lie type A factor through the Hecke algebra H∞ (q) for q a
root of unity, [W-1]. The ones corresponding to the the types B,C,D factor
through the Birman-Murakami-Wenzl algebra C∞ (r, q) for special values of
r and q, [W-2]. Wenzl subfactors are a generalisation of Jones subfactors
which arise in connection with the sl(2) case, [J].
In [E] the two-sided versions of the Wenzl subfactors for classical Lie types
were considered; for the Jones subfactors this had been done by M. Choda
in [Ch]. The two-sided subfactors are defined by extending the unitary
braid representations ρ to the infinite two-sided braid group with generators
σi , with i ∈ Z. Thus, the two-sided pairs are generated by {gi : i ∈
Z\{0, . . . , m}} and by {gi : i ∈ Z}.
It turns out that if we set m = 0 — where m is as in the first paragraph
— then the asymptotic inclusions (in the sense of Ocneanu, [O]) for Wenzl’s
one-sided pairs associated with the Lie types A,B,C,D coincide with the
corresponding two-sided versions. This was easily proved in [E] for the
B,C,D types. In this paper we shall prove this fact for the remaining type A
57
58 JULIANA ERLIJMAN
case. It has also been shown by S. Goto in [G] independently by a different

method.
The method we employ here is fairly elementary. Roughly, we show that
the iterations of the Jones’ basic construction for the one-sided pair at the
finite dimensional level are obtained by adding generators “to the left”, and
then by reducing by a special projection. In this way we can relate the
Jones’ basic iterations of the one-sided pair with the two-sided pair. For
example, for large n ∈ N, the first basic construction for
hg1 , . . . , gn i ⊂ hg0 , . . . , gn i
is isomorphic to the algebra
(∗) phg−k+1 , . . . , gn ip = he, phg0 , . . . , gn ipi.
Here, the parameter k corresponds to sl(k), and p and e are special projec-
tions in hg−k+1 , . . . , g−2 i and in hg−k , . . . , g−2 i, respectively. The projection
e behaves like the Jones’ basic projection. When q = 1, e is precisely the
antisymmetrizer ak of the symmetric group algebra CSk , acting on V ⊗k by
permuting the tensor factors, where V is a k-dimensional vector space; and
p coincides with the antisymmetrizer ak−1 , which projects V ⊗k−1 onto the
dual representation of V , regarded as an Sl(k)-module. The subfactors for
this case, q = 1, were studied by A. Wassermann, [Wa].
At each iteration, we do a similar procedure as in (∗) for the first basic
construction, with variations depending on its parity, and one obtains, in
the limit, a reduced two-sided pair. Later, by applying some Bisch-McDuff
results, we show that this reduced two-sided pair is in fact conjugate to the
two-sided pair.
This paper is organised as follows: In the first section we include some
results by Wenzl on his type A subfactors and definitions. In the second
section we prove some technical lemmas and establish the relation between
the Jones’ iterations of the one-sided finite dimensional approximants and
the two-sided approximants. In the last section, first part, we show that we
can apply Bisch-McDuff results that imply that the pairs in consideration are
stable, that is, that they split a copy of the hyperfinite II1 factor. Later on,
we express the asymptotic inclusion as some “infinitely reduced” version of
the two-sided pair, and using inductive arguments and stableness we prove
our main result. As a remark, the “infinitely reduced” two-sided pair is
conjugate to the two-sided pair in a more general setting: In fact, in our
proof we only assume the properties of more general braid-type subfactors
as constructed in [E]. We also include in Section 3.2 an expression for
the higher relative commutants of Wenzl’s (finite depth) type A subfactors,
which have already been computed by him, [W-4], and later also computed
by D. Evans and Y. Kawahigashi in [EK].
TWO-SIDED BRAID GROUPS 59
Acknowledgements. I am very grateful to Hans Wenzl, who taught me

what the higher relative commutants for his type A subfactors are, and by
several discussions. I am also grateful to Dietmar Bisch for informing me
about his McDuff-type splitting results for pairs, and to Fred Goodman. I
would also like to thank the Fields Institute and George Elliott for their
hospitality and support.
1. Preliminaries.
Let us recall that the finite dimensional Hecke algebra Hn (q) is the free
complex algebra with generators 1, T0 , . . . , Tn−2 , and relations, depending
on a parameter q ∈ C,
(B1 ) Ti+1 Ti Ti+1 = Ti Ti+1 Ti , for i = 0, . . . , n − 3,
(B2 ) Ti Tj = Tj Ti , for |i − j| ≥ 2,
(H) Ti2 = (q − 1)Ti + q, for i = 0, . . . , n − 2.
It can be shown by induction that theseS complex algebras have dimension
n!, independent of q. Set H∞ (q) = Hn (q). The representations consid-
ered for defining the braid subfactors are interesting in the case that the
2πi
parameter q is a root of unity, q 6= 1. So, we shall fix q = e± l , with l ≥ 3.
We shall briefly summarise the parametrisations of the semisimple quotients
of Hn (q), with q as above, which are associated with sl(k) for 1 < k < l (see
[W-1]).
For k ∈ N, and k < l, a (k, l) Young diagram λ of size n is a k-tuple
λ = (λ1 , . . . , λk ) with λ1 ≥ λ2 ≥ . . . ≥ λk ≥ 0, λ1 − λk ≤ l − k, and
Pk (k,l)
i=1 λi = n. We denote the set of (k, l) diagrams of size n by Λn . We
can also regard a (k, l) diagram λ = (λ1 , . . . , λk ) of size n as k ordered rows
of boxes with λi boxes in the ith row. A (k, l) tableau t of shape λ of size n
is a standard tableau, such that for each j ≤ n, the subdiagram occupied by
(k,l)
the numbers {1, . . . , j} is an element of Λj . A (k, l) standard tableau t
of shape λ can also be regarded as an increasing sequence of (k, l) diagrams
(k,l)
∅ = λ0 ⊂ [1] = λ1 ⊂ λ2 ⊂ . . . ⊂ λn = λ. Denote by Tλ the set of (k, l)
tableaux of shape λ.
(k,l)
For each diagram λ ∈ Λn let Vλ be a vector space with basis {vt }
(k,l)
labelled by Tλ . Wenzl defined in [W-1] an irreducible representation
(k,l)
πλ of Hn (q) on Vλ given by
(k,l) {(1 − q d+1 )(1 − q d−1 )}1/2

πλ (Ti )~vt = bd (q)~vt + ~vsi (t) ,
1 − qd
where d = d(t, i) = c(i+1)−c(i)−r(i+1)+r(i), with c(j) and r(j) denoting
the column and the row of the box containing j respectively, and where
d (1−q)
bd (q) = q(1−q d ) , and si (t) is the tableau obtained from t by interchanging
60 JULIANA ERLIJMAN
the numbers i and i + 1 (if the resulting tableau is not standard then it
appears with zero coefficient).
(k,l)
Different diagrams in Λn give inequivalent representations. Consider
Hn−1 (q) as the subalgebra of Hn (q) generated by 1, T0 , . . . , Tn−3 . The re-
(k,l) (k,l)
striction rule is given by πλ |Hn−1 (q) ∼ 0
L
= λ0 <λ πλ0 , where λ < λ means
(k,l)
that λ can be obtained by adding one box to λ0 ∈ Λn−1 . We can then
(k,l) (k,l) L (k,l)
define a representation πn of Hn (q) by πn (x) = λ∈Λ(k,l) πλ (x), for
n
(k,l)
x ∈ Hn (q). Its restriction to Hn−1 (q) is equivalent to πn−1 . Finally, we have
(k,l)
a well defined representation π (k,l) of H∞ (q) given by π (k,l) (x) = πn (x),
if x ∈ Hn (q). The representation π (k,l) is locally finite dimensional, i.e., the
algebras (k,l) (H (q)) are finite dimensional C∗ -algebras. We have
∼ L Bn := π n
Bn = λ∈Λ(k,l) Maλ (C), that is, the equivalence classes of minimal idempo-
n
(k,l)
tents of Bn are labelled by Λn . The representation π (k,l) is also unitary
(k,l)
(that is, gi := πn (Ti ) is a unitary element of Bn = hg0 , . . . , gn−2 i, for
i = 0, . . . , n − 2, and every n ∈ N), and has the following properties:
(i) Any element x ∈ Bn+1 can be written as a sum of elements agn−1 b + c
with a, b, c ∈ Bn .
(ii) The ascending sequence of finite dimensional C∗ -algebras (Bn ) is pe-
riodic with period k, in the sense of Wenzl,
S [W-1].
(iii) The unique positive faithful trace tr on Bn has the Markov property:
tr (gn−1 x) = η tr (x),
for all x ∈ Bn , for all n ∈ N, where η ∈ C is fixed. Given condition
(i), the Markov condition implies the multiplicativity property for the
trace:
(iii)0 If x and y are in ∗-subalgebras generated by disjoint subsets of gener-
ators gi , then
tr (xy) = tr (x) tr (y).
(iv) Existence of a projection p ∈ Bk , where k is the periodicity, with the
contraction property: p ∈ Bk has the contraction property if for all
n ∈ N,
pBn+k p ∼
= pBk+1,n+k+1 ∼ = Bk+1,n+k+1 ,
where Bs,t is the algebra generated by {gs , . . . , gt−2 }.
Property (iv) is equivalent to a special propertyL in connection with the
structure coefficients for the multiplication in n K0 (Bn ). More precisely,
given a locally finite dimensional representation of the braid group
S L B∞ =
B
n n , one has an associative, commutative, graded product on n K0 (Bn )
defined as follows (see [GW] for more details). For projections x ∈ Bn
and y ∈ Bm define x ⊗ y = x(ρ ◦ shiftn )(y) ∈ Bn+m , where shiftn :
CBm → CBn+m is determined by σi 7→ σi+n . Then, [x] ⊗ [y] = [x ⊗ y]
L
defines the multiplication on n K0 (Bn ). Denote the structure constants
ν
of this multiplication by cλµ . That is, if pλ and pµ are minimal projections
(k,l) (k,l)
classes labelled by λ ∈ Λn and µ ∈ Λm respectively, then [pλ ] ⊗
in the P
[pµ ] = ν∈Λ(k,l) cνλµ [pν ]. The existence of a projection p with the contraction
n+m
property is equivalent to (v) or (vi) below:
(k,l) (k,l)
(v) For all n ∈ N0 there exists an injective map j : Λn → Λn+k L that pre-
serves the structure coefficients for the multiplication in n K0 (Bn ),
ν j(ν) j(ν) (k,l) (k,l)
that is, such that cλµ = cj(λ)µ = cλj(µ) for all λ ∈ Λn , µ ∈ Λm ,
(k,l) (k,l)
ν ∈ Λn+m , and also such that cλj(µ) = 0 if ∈/ j(Λn+k ).
(vi) There exists a projection p ∈ Bk such that for every minimal projection
pλ ∈ Bn , and for all n ∈ N, the projection p ⊗ pλ remains minimal in
Bn+k . Moreover, if λ 6= λ0 then p ⊗ pλ and p ⊗ pλ0 are not equivalent.
For a proof of these see [W-3], [E]. An implication of these last conditions
is that the injective map j preserves the coefficients of the inclusion matrices
for the pairs Bn ⊂ Bn+1 , Gn = (gλµ )λ∈Λ(k,l) ,µ∈Λ(k,l) , that is, gλµ = gj(λ)j(µ) ,
n n+1
since gλµ = cµ[1]λ . It is also easy to show that the periodicity condition on
(k,l) (k,l)
(Bn ) forces the injective map j : Λn → Λn+k to be a bijection for large
n (see [E]). It was shown in [GW] that an idempotent p[1k ] ∈ Bk in the
equivalence class given by the diagram [1k ] (a column of height k boxes)
(k,l)
has the contraction property: The map j maps an element µ ∈ Λn to
(k,l)
a diagram j(µ) ∈ Λn+k obtained by adding to µ a column of k boxes to
the left. Let us remark that the minimal projections p[1s ] ∈ Bs correspond
to “q-antisymmetrizers”: If q = 1, p[1s ] is just the antisymmetrizer of CSs
acting on V ⊗s – where V is an s-dimensional vector space – by permuting
tensor factors.
We include here a well-known lemma related to the structure coefficients
that will be needed in the next section:
Lemma 1.
j(η) (k,l) (k,l)
(a) cση[1] = c[1k−1 ]σ , for η ∈ Λn and σ ∈ Λn+1 , where the integer co-
efficients cµνρ are theLstructure coefficients for the associative graded
multiplication ⊗ on n K0 (Bn ).
j(η)
(b) c[1k−1 ]σ 6= 0 ⇐⇒ η < σ.
j(η)
Proof. (a) The coefficients cση[1] and c[1k−1 ]σ coincide with the classical Little-
wood-Richardson coefficients by the Young-Pieri Rules (see [GW]). So we
only need to apply the classical Littlewood-Richardson rule, which says that
the (classical) cνλµ coefficient is the number of ways the Young diagram λ
can be expanded to the Young diagram ν by strict µ-expansion (check for
62 JULIANA ERLIJMAN
these rules e.g. in [FH]). By this rule,

(
1 if σ > η
(1.1) cση[1] =
0 otherwise.
If cση[1] = 1 then σ > η, i.e., σ− {one box} = η, so that j(σ− {one box}) =
j(η), where σ − {one box} means a diagram obtained by taking away one
box of σ. Thus, one can see that j(η) is a strict σ-expansion of [1k−1 ], and
j(η) j(η)
it follows that c[1k−1 ]σ = 1. Conversely, if c[1k−1 ]σ > 0 then j(η) can be
obtained by adding the boxes of σ to [1k−1 ]. Since j(η) has at least one
column with k boxes, then j(η) is formed by adding one box of σ to [1k−1 ]
to obtain [1k ], and then “attaching” the resulting σ − {one box} to the
“right” of [1k ]. Then, η = σ − {one box}, so that cση[1] = 1. It follows that
j(η)
cση[1] = c[1k−1 ]σ .
(b) It is an obvious consequence of (a) and (1.1).
The so-called one-sided and two-sided inclusions of hyperfinite II1 factors
corresponding to the representation π (k,l) of H∞ (q) are defined as follows:
For r ∈ N set
Ar := hg1 , . . . , gr−2 i, Br := hg0 , . . . , gr−2 i.
The pair Ar ⊂ Br is conjugate to Br−1 ⊂ Br . The inclusion of ascending
sequences (Ar ) ⊂ (Br ) is periodic, and the Wenzl one-sided pair is given
by the inclusion
S of hyperfinite
S II1 factors A ⊂ B, with A and B the weak
closures of r Ar and r Br respectively, in the GNS representation with
respect to the unique trace (see [W-1]). For the two-sided inclusion (see
[E]), consider the ascending sequences of finite dimensional C ∗ -algebras
C̃r := hg−r+1 , . . . , g−1 , g1 , . . . , gr−1 i,
D̃r := hg−r+1 , . . . , g−1 , g0 , g1 , . . . , gr−1 i,
after extending our representation to the two-sided infinite braid group. We
have the identifications
C̃r ∼
= Br ⊗ Br , D̃r ∼
= B2r ,
via the trace preserving automorphism g−i gj 7→ gr−i ⊗gj , and gi 7→ gr+i . The
pair C̃r ⊂ D̃r is conjugate to Br ⊗ Br ⊂ B2r , and the inclusion here
L depends
on the structure coefficients for the graded multiplication on n K0 (Bn ),
as explained in properties (iv)-(vi). The inclusion of ascending sequences
(C̃r ) ⊂ (D̃r ) is periodic as well, and the two-sided pair is given by the
inclusion
S of hyperfinite
S II1 factors C ⊂ D with C and D the weak closures
of r C̃r and r D̃r respectively, in the GNS representation with respect to
the unique trace.
In the rest of the paper we shall assume a basic knowledge on Wenzl’s

Hecke algebra article, [W-1].
2. Some technical lemmas and the basic iterations for Ar ⊂ Br .

In this section we shall first prove some rather technical lemmas. These
will be used later in this section to show how the iterations of the basic
construction for Ar ⊂ Br are related to the two-sided pair. We use the same
notation as in the preliminaries, for i, j ∈ Z, i < j − 1,
Bi,j := hgi , . . . , gj−2 i,
so that Br = B0,r . The algebra Bi,j = shifti (Bj−i ) is just the shifting “to
the right (or left) by |i| ” of the algebra Bj−i . If k is the periodicity constant,
n ∈ N, and if (
−ik + 1 if n = 2i − 1
k(n) := ,
−ik if n = 2i
define the projections
(
p(n) := shiftk(n) (p[1k−1 ] ) ∈ shiftk(n) (Bk−1 )
.
q (n) := shiftk(n) (p[1k ] ) ∈ shiftk(n) (Bk )
By this we mean that p(n) and q (n) are the q-antisymmetrizers p[1k−1 ] ∈ Bk−1
and p[1k ] ∈ Bk , but with the appropriate shifting of the generators’ indexes
depending on the parity of n. Set also
i
Y
(i)
P := p(2s−1) = p(1) p(3) . . . p(2i−1) .
s=1
Below we list some properties which will be used frequently in the next
section. The reference for these is basically [W-1].
Remarks.
(1) For every n ∈ N, p(n) ∈ shiftk(n) (Bk−1 ) and q (n) ∈ shiftk(n) (Bk ) are
minimal. ([W-1, p. 368].)
(2) For i ∈ N, by the braid relations, the p(2i−1) ’s are mutually commuting,
and the p(2i) ’s are also mutually commuting. In particular, P (i) is a
projection.
(3) For every n ∈ N, q (n) ≤ p(n) . Furthermore, if we regard p(n) in
shiftk(n) (Bk ), then p(n) splits orthogonally as p(n) = q (n) ⊕ q̃ (n) , where
q̃ (n) ∈ shiftk(n) (Bk ) is also minimal. ([W-1, p. 368].)
(4) q (2i) ≤ p(2i−1) : Since shift1 (p[1k−1 ] ) ∼ p[1k−1 ] in Bk , p[1k ] is a rank one
minimal central projection in Bk , and p[1k ] ≤ p[1k−1 ] by (3), then p[1k ] ≤
shift1 (p[1k−1 ] ). Thus, q (2i) = shift−ik (p[1k ] ) ≤ shift−ik (shift1 (p[1k−1 ] )) =
p(2i−1) .
64 JULIANA ERLIJMAN
(5) By the braid relations, [q (n) , Bs,t ] = 0 for any n ∈ N, and s, t ∈ Z such
that −(i − 1)k + 1 ≤ s ≤ t − 2, or such that −(i − 1)k ≤ s ≤ t − 2
(depending on parity of n). In particular, q (2i−1) and q (2i) commute
with p(2j−1) for j = 1, . . . , i − 1.
(6) q (n) ∈ shiftk(n) (Bk ) remains minimal in shiftk(n) (Bk+1 ): This can be
shown by using the contraction property of p[1k ] , or by using the prop-
erties for p[1k ] stated in [W-1, p. 368].
(7) Let A ⊂ B be finite dimensional C∗ algebras acting on a Hilbert space
H with a faithful trace on B, and e be a projection on H. Then, if
(∗)ebe = eA (b)e, where A is the trace preserving conditional expec-
tation onto A, and (∗∗)eA ∼ = A, then hB, ei ∼ = Q ⊕ K, where Q is
isomorphic to the basic construction for A ⊂ B, and K is isomorphic
to a subalgebra of B. [W-1, Theorem 1.1, (i)].
Next in Lemma 2 we prove some relations involving the projections q (1) ,

q (2) ,
and trace preserving conditional expectations. This will enable us to
show that in fact the projections q (n) behave as the Jones’ basic projections.
Lemma 2. If N ⊂ M are von Neumann algebras and p ∈ M , set Np to

be the von Neumann algebra generated by {pxp : x ∈ N }. So, for r ∈ N
consider the inclusions Ar p(1) ⊂ Br p(1) ⊂ B−k+1,r p(1) . Let E1 : Br p(1) →
Ar p(1) and E2 : B−k+1,r p(1) → Br p(1) denote the conditional expectations
with respect to tr onto Ar p(1) and Br p(1) , respectively. We have:
(a) q (1) commutes with Ar p(1) .
(b) q (2) commutes with Br p(1) .
(c) q (1) bq (1) = E1 (b)q (1) , for every b ∈ Br p(1) .
(d) q (2) q (1) q (2) = αq (2) , with α ∈ C such that E2 (p(1) q (1) p(1) ) = αp(1) .
(e) hq (1) , Br p(1) i contains a subalgebra, Q1 , isomorphic to hAr p(1) , e1 i.
Proof. Parts (a) and (b) are true by remarks (3), (4) and (5).
(c) Any element x ∈ Br can be written as a linear combination of elements
of the form ug0 v + w, with u, v, w ∈ Ar (see [W-1]). Since p(1) commutes
with Br , then any element in Br p(1) can be written as a linear combination
of elements of the form ug0 v + w, with u, v, w ∈ Ar p(1) . By the bimodule
property of the conditional expectation (E1 (uzv) = uE1 (z)v if u, v ∈ Ar p(1) ),
if u, v, w ∈ Ar p(1) then E1 (ug0 v + w) = αuv + w, where α = tr(g0 ) (the
last equality follows from multiplicativity property of tr; see (iii)0 in the
preliminaries). As q (1) commutes with Ar and q (1) p(1) = q (1) = p(1) q (1) , the
claim follows as soon as we show q (1) g0 q (1) = αq (1) . Since q (1) ∈ B−k+1,2 is
minimal then q (1) g0 q (1) = βq (1) , for some β ∈ C. By the Markov property
β = tr(g0 ) = α.
(d) q (2) remains minimal in B−k,1 and q (2) p(1) = q (2) . (Remarks (4),(6).)
tr (q (2) x)
In particular, if x ∈ B−k+1,1 , q (2) p(1) xp(1) q (2) = αx q (2) , with αx := tr (q (2) )
.
tr (p(1) x)
Also, for y ∈ Br , tr (p(1) xp(1) yp(1) ) = βx tr (p(1) yp(1) ), where βx := tr (p(1) )
,
by multiplicativity of tr, so that E2 (p(1) xp(1) ) = βx p(1) . We shall show that
βx = αx for every x ∈ B−k+1,r . For x ∈ B−k+1,0 , p(1) xp(1) = βx p(1) , by
minimality of p(1) in B−k+1,0 . Then, αx q (2) = βx q (2) , so that αx = βx . For
x = g−1 , αx = βx by multiplicativity of tr. Finally, because αx = βx for
x ∈ B−k+1,0 and for x = g−1 , then the equality holds for every x ∈ B−k+1,1 ,
since any element in B−k+1,1 is a linear combination of elements of the form
tr (q (2) ug−1 v)
ug−1 v + w, with u, v, w ∈ B−k+1,0 . (For x = ug−1 v, αx = tr (q (2) )
=
tr (vq (2) ug−1 ) tr (vq (2) u)tr (g−1 )
tr (q (2) )
= = αuv tr (g−1 ) = βuv tr (g−1 ) = βx .) It now
tr (q (2) )
follows that for x ∈ B−k+1,1 , q (2) p(1) xp(1) q (2) = E2 (p(1) xp(1) )q (2) . In par-

ticular, (d) follows if we take x = q (1) .

(e) The projection q (1) and the algebra Br p(1) verify the conditions of
Remark (7) (conditions (a) and (c) of this lemma). Therefore, hq (1) , Br p(1) i ∼ =
he1 , Br p(1) i⊕K1r , where K1r is isomorphic to a subalgebra of Br p(1) , and where
e1 is the Jones’ basic projection for Ar p(1) ⊂ Br p(1) .
The lemmas above and the following results have the following purpose:
We want to obtain all the iterations of the Jones’ basic construction for
the pair Ar ⊂ Br , in a way that they can be related to the two-sided
construction. The odd iterations will be obtained by adding k − 1 “reduced”
generators on the “left” of the previous iteration, and then by reducing by
the q-antisymmetrizer p(n) for the appropriate n. The even iterations will
be obtained by just adding one reduced generator on the left of the previous
iteration. We shall now show this for the first two iterations, and afterwards
for the general case, by induction.
Proposition 3. For large r ∈ N, consider the inclusion Ar p(1) ⊂ Br p(1) as
before:
(i) The 1st basic construction is isomorphic to B−k+1,r p(1) = hq (1) , Br p(1) i.
(ii) The 2nd basic construction is isomorphic to B−k,r p(1) = hq (2) ,
B−k+1,r p(1) i.
Proof. (i) By Lemma 2 (e), the algebra hq (1) , Br p(1) i contains a subalgebra,
Q1 , isomorphic to the basic construction for Ar p(1) ⊂ Br p(1) . Consider the
inclusions:
t
Br p(1) ⊂G Q1 , and Br p(1) ⊂L B−k+1,r p(1) ,
where G is the inclusion matrix for Ar ⊂ Br . Our goal is to show that
L = Gt , for then we shall have dim Q1 = dim B−k+1,r p(1) , and so the desired
66 JULIANA ERLIJMAN
equality Q1 = hq (1) , Br p(1) i = B−k+1,r p(1) . A minimal idempotent in Br p(1) is

given by p(1) pλ p(1) = p(1) pλ , where pλ is a minimal idempotent in Br in the
(k,l)
central summand labelled by the diagram λ ∈ Λr . We can write the class
of p pλ in B−k+1,r p(1) as follows, using Lemma 1 (a): [p(1) pλ ] = [p[1k−1 ] ⊗
(1)
(k,l) (k,l)
pλ ] = ν∈Λ(k,l) cλ[1]ν [pj(ν) ], where we use that the map j : Λn → Λn+k
P
r−1
as in the preliminaries is a bijection for large n. Therefore, the inclusion
(k,l) (k,l)
matrix L is given by Lλν = cλ[1]ν for λ ∈ Λr , ν ∈ Λr−1 , which coincides
with the matrix Gt .
(ii) Using (i) and Lemma 2 (d), we shall show that hq (2) , B−k+1,r p(1) i
contains a subalgebra, Q2 , isomorphic to the basic construction for Br p(1) ⊂
B−k+1,r p(1) . By Remark (7), Lemma 2 (b), and Remark (5) we just need
to show that q (2) bq (2) = E2 (b)q (2) for all b ∈ B−k+1,r p(1) , where E2 is the
unique trace preserving conditional expectation onto Br p(1) . The equation
above is true for b ∈ Br p(1) , and by Lemma 2 (d) it also holds for b = q (1) .
Since B−k+1,r p(1) = hq (1) , Br p(1) i by part (i), then the equation holds for all
b ∈ B−k+1,r p(1) by the bimodule property of E2 , and because any element of
hq (1) , Br p(1) i is a linear combination of elements of the form xq (1) y + z, with
x, y, z ∈ Br p(1) . Consider the inclusions:
B−k+1,r p(1) ⊂G Q2 and B−k+1,r p(1) ⊂H B−k,r p(1) ,
where G is again the inclusion matrix for Ar ⊂ Br . Since Q2 ⊆ B−k,r p(1) ,

it is enough to show that G = H in order to prove that Q2 = B−k,r p(1) .
(k,l) (k,l)
By Lemma 1 and the fact that the map j : Λn → Λn+k is a bijection
for large n, it is easy to showPthat the equivalent class of p(1) in B−k+1,r
is given by [p[1k−1 ] ⊗ 1r ] = λ∈Λ
(k,l) αλ [pj(λ) ], with αλ 6= 0 for every
r−1
(k,l)
λ ∈ Λr−1 . Therefore, the classes of minimal idempotents in B−k+1,r are
(k,l)
given by [pj(λ) ] with λ ∈ Λr−1 . The class of pj(λ) in B−k,r p(1) is [pj(λ) ] =
P δ
P µ
(k,l) c [p ] = (k,l) c
δ∈Λr+k
j(λ)[1] δ µ∈Λ r λ[1] [pj(µ) ], so that the inclusion matrix H
(k,l) (k,l)
is given by Hλµ = cµλ[1] for λ ∈ Λr−1 and µ ∈ Λr . So H coincides with
G.
Lemma 4. Let 1 ∈ A ⊂ B be finite dimensional C ∗ algebras acting on a

Hilbert space H with a faithful trace tr on B. Let p ∈ A be a full projection
(i.e., the central support of p is 1). Then, the algebra hBp , peA i is isomorphic
to the basic construction for Ap ⊂ Bp , where eA is the Jones projection for
A ⊂ B.
Proof. The algebras Ap , Bp and peA satisfy the conditions of Remark (7).
Therefore, hBp , peA i ∼
= hBp , eAp i ⊕ K, where K is isomorphic to a subalge-
bra of Bp . Because p is full in A, then the inclusion matrix for Ap ⊂ Bp
1 ~
coincides with that for A ⊂ B, G. Thus, tr(p) tA = ~tAp ≥ GGt~thBp ,peA i =
1
GGt~thB,e i = 1 ~tA , so that an equality holds, and by faithfulness of
tr(p) A tr(p)
tr, K must be zero.
We shall obtain now all the iterations of the basic construction, by doing
the same procedure as in the last proposition, by induction. Define
B r (2i − 1) := B−ki+1,r
B r (2i) := B−ki,r .
Corollary 5. For r ∈ N large, the nth basic construction for (Ar ) ([ n+1 ]) ⊂
P 2
(Br ) ([ n+1 ]) is isomorphic to
P 2
D E
B r (n) ([ n+1 ]) = q (n) , B r (n − 1) ([ n+1 ]) .
P 2 P 2
Proof. The statement is true for the cases n = 1, 2 by Proposition 3. We

shall proceed now by induction in n ∈ N. Suppose that the statement is
true for m ≤ n. We want to show that then it is also true for m = n + 1.
By inductive hypothesis, the n − 1st basic construction for (Ar )P ([ n2 ]) ⊂
(Br )P ([ n2 ]) is isomorphic to B r (n − 1)P ([ n2 ]) , and the nth basic construction
for (Ar ) ([ n+1 ]) ⊂ (Br ) ([ n+1 ]) is isomorphic to B r (n) ([ n+1 ]) . Thus, the
P 2 P 2 P 2
n + 1st basic construction for (Ar ) ([ n+2 ]) ⊂ (Br ) ([ n+2 ]) is isomorphic to the
P 2 P 2
first basic construction for for B r (n − 1) ([ n+2 ]) ⊂ B r (n) ([ n+2 ]) . (Here we
P 2 P 2
n+2 n
also use that P ([ 2 ]) = P ([ 2 ]) p, with p = p(n) or p = p(n+1) according to
parity.)
(a) If n is odd, write n + 1 = 2i. The inclusion B r (2i − 2)P (i) ⊂
B r (2i − 1)P (i) is conjugate, via shift(i−1)k , to (B r+(i−1)k (0)p(1) )p̃ ⊂
Qi−1 (2s−1)
(B r+(i−1)k (1)p(1) )p̃ , with p̃ = shift(i−1)k ( s=1 p ). By Proposition 3
(ii) and Lemma 4, the first basic construction for the latter is given by
(B r+(i−1)k (2)p(1) )p̃ = hq (2) , B r+(i−1)k (1)p(1) ip̃ . Then, by applying the in-
verse of shift(i−1)k , shift −(i−1)k , the basic construction for B r (2i − 2)P (i) ⊂
B r (2i − 1)P (i) is isomorphic to B r (2i)P (i) = hq (2i) , B r (2i − 1)P (i) i.
(b) If n is even, write n + 1 = 2i − 1. As in (a), apply shift(i−1)k to
obtain that B r (2i − 3)P (i) ⊂ B r (2i − 2)P (i) is conjugate to (Ar+(i−1) )p(1) p̃ ⊂
(Br+(i−1) )p(1) p̃ , with basic construction isomorphic to (B r+(i−1)k (1)p(1) )p̃ =
hq (1) , (Br+(i−1)k )p(1) ip̃ . Thus, the basic construction for B r (2i − 3)P (i) ⊂
B r (2i − 2)P (i) is isomorphic to B r (2i − 1)P (i) = hq (2i−1) , B r (2i − 2)P (i) i.
68 JULIANA ERLIJMAN
Section 3.
3.1. Denote by B (n) the Jones’ nth basic construction for the one-sided
pair A ⊂ B, (see [J]). The asymptotic inclusion for A ⊂ B is the inclusion
A∨(A0 ∩B (∞) ) ⊂ B (∞) , where B (∞) := ( n B (n) )00 . We shall proceed first by
S
showing that both the two-sided pair C ⊂ D and the asymptotic inclusion
for the one-sided pair A ⊂ B satisfy the conditions of some splitting results
due to D. McDuff and D. Bisch (see [McD], [B]).
Lemma 6.
(a) The asymptotic inclusion A ∨ (A0 ∩ B (∞) ) ⊂ B (∞) of the one-sided pair
A ⊂ B has two non-trivial non-commuting central sequences.
(b) The two-sided pair C ⊂ D has two non-trivial non-commuting central
sequences.
Proof. Let us recall first that a central sequence (xn ) for a II1 factor M is
a bounded sequence in M with the property that k[xn , x]k2 → 0, for all
x ∈ M , where k . k2 is the norm defined on M via its trace. A central
sequence (xn ) ∈ M is said to be trivial if kxn − λn .1k2 → 0, where λn ∈ C
for all n ∈ N. A central sequence for a pair of II1 factors N ⊂ M is a central
sequence (xn ) for M which is contained in N .
(a) The candidates are the sequences (eB (n) )n∈N and (eB (n+1) )n∈N , where
eB (n) is the projection corresponding to the nth Jones’ basic construction for
A ⊂ B. It is obvious that these sequences lie in A ∨ (A0 ∩ B (∞) ). To show
that they are non-trivial, consider a sequence (λn ) of complex numbers. For
all n ∈ N,
1
keB (n) − λn .1k22 = τ |1 − λn |2 + (1 − τ )|λn |2 ≥ min{τ, (1 − τ )} > 0,
2
because tr (eB (n) ) = τ ∈ (0, 1) for all n ∈ N. To show that (eB (n) )n and
(eB (n+1) )n are non-commuting, write
k[eB (n) , eB (n+1) ]k22 = −2tr (ηeB (n) eB (n+1) ) + 2tr (eB (n) eB (n+1) )
= 2η(1 − η) ≡ c > 0,
since eB (n+1) eB (n) eB (n+1) = ηeB (n+1) , with 0 < η = tr(eB (n) ) for all n ∈ N.
Finally, to show that (eB (n) )n and (eB (n+1) )n are central with respect to B (∞)
fix an element x ∈ B (∞) . There exists a sequence (xn ) with xn ∈ B (n) which
converges to x in the k . k2 norm. Given > 0, choose some j0 ∈ N with
kx − xj0 k2 < . Then, [eB (n) , xj0 ] = 0 for n > j0 + 1, so that
k[eB (n) , x]k2 ≤ 2keB (n) k.kx − xj0 k2 + keB (n) xj0 − xj0 eB (n) k2 < 2.
(b) For the two-sided pair C ⊂ D consider the sequences (en )n∈N and
(en+1 )n∈N , where en is the image (under the braid representation consid-
ered) of the spectral projection corresponding to the eigenvalue −1 of the
generator Tn of the Hecke algebra. In other words, for q 6= −1, en = q−g q+1
n
(see [W-1]). These sequences belong to C (since none of its terms is e0 ),

and they are non-trivial: As in (a), for a complex sequence (λn ),
1
ken − λn .1k22 = d|1 − λn |2 + (1 − d)|λn |2 ≥ min{d, (1 − d)} > 0,
2
for all n ∈ N, since d := tr(en ) ∈ (0, 1) for all n ∈ N. For showing that
they are non-commuting we use one of the defining relations for the Hecke
algebra generators, namely, en en+1 en − αen = en+1 en en+1 − αen+1 , for 0 6=
q
α = (1+q) 2 . Hence,
k[en , en+1 ]k22 = −2tr (en en+1 en en+1 − αen en+1 + αen en+1 ) + 2tr (en en+1 )
= 2αtr (en ) − 2αtr (en )2 ≡ κ > 0,
where we also use the fact that tr (en en+1 ) = tr(en )2 , which is constant for
all n ∈ N because of the Markov property of tr. Finally, we need to show
that these are central sequences with respect to D. This follows from a
similar procedure as in (a). Let x be an element in D. Then, there exists
a sequence (xn ) with xn ∈ he−n , . . . , en i, which converges to x in the k . k2
norm. For > 0, take j0 ∈ N such that kx − xj k2 < . Then, [en , xj0 ] = 0
for n > j0 + 1, because xj0 ∈ he−j0 , . . . , ej0 i and [ei , ej ] = 0 for |i − j| > 1.
Therefore,
k[en , x]k2 ≤ 2ken k kx − xj0 k2 + ken xj0 − xj0 en k2 < 2.

Now we are able to use the Bisch-McDuff splitting results mentioned
above: In [McD], McDuff gave necessary and sufficient conditions for a
k . k2 -separable II1 factor M to be isomorphic to M ⊗R, ¯ where R is the
separable II1 factor. In [B, Theorem 3.1], based on the work by McDuff,
Bisch gave necessary and sufficient conditions for a pair N ⊂ M of separable
¯ ⊂ M ⊗R
II1 factors to be stable, that is, for N ⊂ M to be conjugate to N ⊗R ¯
¯
(i.e., existence of an isomorphism Φ : M → M ⊗R with Φ(N ) = N ⊗R). ¯
The equivalent condition that we shall use is the following, namely, that
the algebra M 0 ∩ N ω is non-commutative, where ω is a free ultrafilter in
N. In particular, it is enough to show that there exist two non-trivial non-
commutative central sequences for M which are contained in N . Therefore,
as a result of Lemma 6:
Corollary 7. Both the asymptotic pair A ∨ (A0 ∩ B (∞) ) ⊂ B (∞) for the
one-sided pair A ⊂ B and the two-sided pair C ⊂ D are stable.
Remark. Let R⊗∞ be the II1 factor given by the inductive limit of R⊗n
with the canonical embeddings and trace. It can easily be shown that it is
hyperfinite, so that it is isomorphic to R. For a non-zero projection p ∈ R, let
70 JULIANA ERLIJMAN
Rp⊗∞ be the II1 factor given by the inductive limit of the factors Rp⊗n , with
the embeddings Rp⊗n ∼ = Rp⊗n ⊗ p ⊂ Rp⊗n+1 , together with an appropriate
renormalisation of the canonical trace at each step. It can also be easily
shown that Rp⊗∞ is hyperfinite, and so isomorphic to R. Therefore, if a pair
N ⊂ M splits a copy of R, then it also splits a copy of R⊗n , R⊗∞ , or Rp⊗∞ .
3.2. In Section 2 we obtained all the iterations of the Jones’ basic con-
structions for Ar ⊂ Br in a way which relates them to the two-sided pair
C ⊂ D. Corollary 5 says that for large r, and for every s, the 2s + 1st basic
construction for the pair (Ar )P (s) ⊂ (Br )P (s) is isomorphic to B r (2s − 1)P (s) .
Then, B (∞) can be identified with the II1 factor defined by the inductive
limit of (Ds )P (s) , with
Ds := hg−sk+1 , g−sk , . . . i00 ,
and the embeddings ψs : (Ds )P (s) → (Ds+1 )P (s+1) given by (Ds )P (s) = ∼
(Ds )P (s+1) ⊂ (Ds+1 )P (s+1) (note that [p (s+1) , Ds ] = 0 for all s ∈ N), to-
gether with an appropriate renormalisation of the trace tr at each step; that
is, define tr(s) on each (Ds )P (s) by tr(s) := tr(p 1k−1 )s tr. With the same iden-
[1 ]
tification as above, the subalgebra of B (2s−1) given by A ∨ (A0 ∩ B (2s−1) ) is

isomorphic to (Cs )P (s) , where
Cs := hg−sk+1 , . . . , g−1 , ĝ0 , g1 , . . . i00 :
For showing this, one should prove that
hg−sk+1 , . . . , g−1 iP (s) = hg1 , . . . i00P (s) ∩ hg−sk+1 , . . . i00P (s) .
One inclusion,
hg−sk+1 , . . . , g−1 iP (s) ⊂ hg1 , . . . i00P (s) ∩ hg−sk+1 , . . . i00P (s) ,
is obvious. For the other inclusion, one needs to apply an estimate, [W-1,
(2s−1)
Theorem 1.6] — dim A0 ∩ B (2s−1) ≤ dim p(A0r ∩ Br ), for any non-zero
projection p ∈ Ar and sufficiently large r — as it was used in [W-1, Theorem
3.7, (b)], where p ∈ Ar is chosen to be a projection with the contraction
property. So, as a corollary we re-obtain:
Wenzl’s Theorem. The nth -higher relative commutant A0 ∩ B (n) of the
pair A ⊂ B is isomorphic to B 1 (n) ([ n+1 ]) .
P 2
Proof. It is a corollary of Corollary 5, and the above identification. (We

include at the end of this article a rule of how to draw the inclusion diagrams
for A0 ∩B (n) ⊂ A0 ∩B (n+1) and a description of the principal graphs, together
with two examples.)
We have just seen that the asymptotic pair A ∨ (A0 ∩ B (∞) ) ⊂ B (∞) for
the one-sided inclusion A ⊂ B is conjugate to the von Neumann algebra
inductive limit of (Cs )P (s) ⊂ (Ds )P (s) — which we shall denote by Cp∞ ⊂
Dp∞ for convenience — with the embeddings and the renormalisation of the
trace at each step as above. We need to show that Cp∞ ⊂ Dp∞ is conjugate
to the two-sided pair C ⊂ D. In other words, we need to “eliminate” the
projections P (s) . That is the reason in the next results we need consider
larger von Neumann algebras, of the form (Cs )P (s) ⊗ R⊗s .
Lemma 8. Consider s large enough so that P (s) is full in B 1 (2s − 1) =
B−sk+1,1 . Then, there exist a projection P̃ (s) ∈ Cs with P̃ (s) ≤ P (s) , and a
unitary us ∈ Cs ⊗ R⊗s such that:
(a) ubs (P̃ (s) ⊗ 1R⊗s ) := us (P̃ (s) ⊗ 1R⊗s )u∗s = 1C ⊗ qs , with qs a projection
in R⊗s .
(b) P̃ (s) is “maximal” with the property described in (i). That is, if there is
0 0
a projection P (s) ∈ Cs with P (s) ≤ P (s) , and a unitary u0s ∈ Cs ⊗ R⊗s
0 0
with u b0s (P (s) ⊗ 1R⊗s ) = 1C ⊗ qs0 for some qs0 ∈ R⊗s , then P (s) . P̃ (s) .
Proof. (a) Let Λ be the finite set of (k, l) Young diagrams that label the
minimal central projections z̃λ of B 1 (2s − 1), and zλ = z̃λ ⊗ 1 the minimal
central projections in Cs = hg−sk+1 , . . . , g−1 , g1 , . . . i00 . For λ ∈ Λ let qλ be a
(s)
projection in R⊗s with tr(qλ ) = tr(P zλ )
tr(zλ ) . Since in a II1 factor projections
are unitarily equivalent if and only if their traces coincide, there is a uni-
tary ũλ,s in the factor (Cs ⊗ R⊗s )zλ ⊗1R⊗s such that u b̃λ,s (P (s) zλ ⊗ 1R⊗s ) =
ũλ,s (P (s) zλ ⊗ 1R⊗s )ũ∗λ,s = zλ ⊗ qλ . Take qs ∈ R⊗s with tr(qs ) = min{tr(qλ )},
λ∈Λ
so that zλ ⊗ qλ & zλ ⊗ qs in (Cs ⊗ R⊗s )zλ ⊗1R⊗s for all λ. It follows that there
is a projection P̃ (s) ∈ Cs ⊗ R⊗s and a unitary uλ,s in (Cs ⊗ R⊗s )zλ ⊗1R⊗s with
P̃ (s) zλ ≤ P (s) zλ and u
bλ,s (P̃ (s) zλ ⊗ 1R⊗s ) = zλ ⊗ qs . Thus, if us := λ uλ,s ,
P
then us ∈ Cs ⊗ R⊗s is a unitary and u bs (P̃ (s) ⊗ 1R⊗s ) = 1C ⊗ qs .

0 0
(b) Suppose there exist P (s) ∈ Cs with P (s) ≤ P (s) and a unitary u0s ∈
0
Cs ⊗ R⊗s such that u b0s (P (s) ⊗ 1R⊗s ) = 1C ⊗ qs0 for some qs0 ∈ R⊗s . Then for
all λ ∈ Λ
0
tr(zλ )tr(qs0 ) = tr(P (s) zλ ) ≤ tr(P (s) zλ ) = tr(zλ )tr(qλ ),
0
so that tr(qs0 ) ≤ tr(qλ ). Thus, tr(qs0 ) ≤ min tr(qλ ) = tr(qs ), and so P (s) .
λ∈Λ
P̃ (s) .
Lemma 9. Fix s large. There is a projection P̃ (2s) ∈ C2s with P̃ (s) ⊗ P̃ (s) ≤
P̃ (2s) ≤ P (2s) , a projection q2s ∈ R⊗2s with q2s ≥ qs ⊗ qs (P̃ (s) and qs as in
the lemma above), and a unitary u2s in C2s ⊗ R⊗2s such that:
(a) As in Lemma 8, (a), for 2s.
72 JULIANA ERLIJMAN
(b) As in Lemma 8, (b), for 2s.

(c) The following is a commuting diagram:
u
bs
(Cs ⊗ R⊗s )(P̃ (s) ⊗1 ⊗s ) −→ (Cs ⊗ R⊗s )(1C ⊗qs )
 R 
 
ιes 
y
 es
y (A)
s ⊗ us
u\
(C2s ⊗ R⊗2s )(P̃ (s) ⊗P̃ (s) ⊗1 ⊗2s ) −→ (C2s ⊗ R⊗2s )(1C ⊗qs ⊗qs )
 R 
 
ιes 
e
y
 e
y es (B)
u
b2s
(C2s ⊗ R⊗2s )(P̃ (2s) ⊗1 −→ (C2s ⊗ R⊗2s )(1C ⊗q2s ) ,
R⊗2s )
with the embeddings



 ιes = reduction by the projection P̃ (s) ⊗ P̃ (s) ⊗ 1R⊗2s ,

e = reduction
s by the projection 1C ⊗ qs ⊗ qs ,

 ιes = reduction
e by the projection P̃ (2s) ⊗ 1R⊗2s ,

es = reduction by the projection 1C ⊗ q2s .
e
(Note that the maps above are embeddings since P̃ (s) ⊗ P̃ (s) ⊗1R⊗2s and
1C ⊗qs ⊗qs commute with P̃ (s) ⊗1R⊗s and with 1C ⊗qs respectively, and
also P̃ (2s) ⊗ 1R⊗2s ≥ P̃ (s) ⊗ P̃ (s) ⊗ 1R⊗2s , and 1C ⊗ q2s ≥ 1C ⊗ qs ⊗ qs ).
Proof. Note that we identify B 1 (2s − 1) ⊗ B 1 (2s − 1) with a subalgebra of
B 1 (4s − 1) according to the multiplication rules for the structure coefficients
that describe the embeddings Br ⊗ Br ⊂ B2r , as in the preliminaries. By
Lemma 8 we can find P] (2s) ≤ P (2s) , with P
] (2s) ⊗1 ⊗2s ∈ C ⊗R⊗2s maximal
R 2s
with respect to the property of being unitarily equivalent to a projection of
the form 1C ⊗ q̃2s , for some q̃2s ∈ R⊗2s .
We can also define the unitary ‘us ⊗ us ’ ∈ C2s ⊗ R⊗2s since, by the
construction of us ∈ Cs ⊗ R⊗s in the proof of Lemma 8, one has that in fact
us ∈ B 1 (2s − 1) ⊗ R⊗s (so that us ⊗ us ∈ B 1 (2s − 1) ⊗ B 1 (2s − 1) ⊗ R⊗2s ⊂
C2s ⊗ R⊗2s ). For P̃ (s) as in Lemma 8, P̃ (s) ⊗ P̃ (s) ≤ P (s) ⊗ P (s) = P (2s) ,
and P (s) ⊗ P (s) ⊗ 1R⊗2s is unitarily equivalent to 1C ⊗ qs ⊗ qs via us ⊗ us ,
so that P̃ (s) ⊗ P̃ (s) . P](2s) by maximality of P ](2s) . Since this last relation
holds in each factor summand of C2s ⊗ R⊗2s , then for every minimal central
projection zλ ∈ C2s there is a projection P̃ (2s) zλ such that P̃ (2s) zλ ∼ P] (2s) z
λ
(s) (s)
and (P̃ ⊗ P̃ )zλ ≤ P̃ (2s) zλ ≤ P (2s) (s) (s)
zλ . Thus, P̃ ⊗ P̃ ≤ P̃ (2s) ≤ P (2s) ,
and P̃ (2s) ⊗ 1R⊗2s is maximal with respect to being unitarily equivalent to
a projection of the form 1C ⊗ q2s ∈ C2s ⊗ R⊗2s , where we can assume
that q2s ≥ qs ⊗ qs . Since their traces agree in each factor summand, the
projections (P̃ (2s) − P̃ (s) ⊗ P̃ (s) ) ⊗ 1R⊗2s and 1C ⊗ (q2s − qs ⊗ qs ) are unitarily
equivalent via some unitary w2s ∈ C2s ⊗ R⊗2s .
Note that the maps ιe es and e es are clearly embeddings, and that the fact
the morphisms ιes and es are injective follows from the faithfulness and mul-
tiplicativity of the trace (property (iii)0 in the preliminaries).
To find a unitary u2s ∈ C2s ⊗ R⊗2s that makes the diagram commute,
take the partial isometry
v2s := w2s (P̃ (2s) − P̃ (s) ⊗ P̃ (s) ) ⊗ 1R⊗2s + (us ⊗ us )(P̃ (s) ⊗ P̃ (s) ⊗ 1R⊗2s ),

and a partial isometry v2s 0 with initial projection given by (1 − P̃ (2s) ) ⊗

C
1R⊗2s and final projection given by 1C ⊗ (1R⊗2s − q2s ). (The latter can be
defined since the trace of both projections agree in each factor summand of
C2s ⊗ R⊗2s .) One can see easily that u2s := v2s + v2s 0 is a unitary and that
u
b2s (P̃ (2s) ⊗ 1R⊗2s ) = 1C ⊗ q2s . Finally, by the definitions of the embeddings
and the unitaries, the diagrams (A) and (B) commute.
Lemma 10. Fix s large. The pairs of von Neumann algebra inductive limits
n n
lim (Cs2n )P̃ (s2n ) ⊗ R⊗s2 ⊂ lim (Ds2n )P̃ (s2n ) ⊗ R⊗s2 ,
−→ −→
n n
and
n n
lim Cs2n ⊗ (R⊗s2 )qs2n ⊂ lim Ds2n ⊗ (R⊗s2 )qs2n
−→ −→
n n
are conjugate with the embeddings as above, and the trace normalisation at
each step as described at the beginning of this section.
Proof. By Lemma 8 and Lemma 9, for s large and fixed, we can inductively
n n n
define sequences of projections (P̃ (s2 ) )n , with P̃ (s2 ) ∈ Cs2n ⊗ R⊗s2 and
n n n
P̃ (s2 ) ≤ P (s2 ) , and (qs2n )n with qs2n ∈ R⊗s2 , and also a sequence of
n
unitaries (us2n )n with us2n ∈ Cs2n ⊗ R⊗s2 , such that the diagrams below
commute for all n:
74 JULIANA ERLIJMAN
n u
bs2n n
(Cs2n )P̃ (s2n ) ⊗ R⊗s2 −→ Cs2n ⊗ (R⊗s2 )qs2n
& .
n u
bs2n n
(Ds2n )P̃ (s2n ) ⊗ R⊗s2 −→ (Ds2n )P̃ (s2n ) ⊗ (R⊗s2 )qs2n
   
   
ιs2 ιs2n s2n  s2n
 n 
y y 
y y
n+1 n+1
(Ds2n+1 )P̃ (s2n+1 ) ⊗ R⊗s2 −→ Ds2n+1 ⊗ (R⊗s2 )qs2n+1
u
bs2n+1
% -
n+1 n+1
(Cs2n+1 )P̃ (s2n+1 ) ⊗ R⊗s2 −→ Cs2n+1 ⊗ (R⊗s2 )qs2n+1 .
u
bs2n+1

Proposition 11. The inclusion of inductive limits obtained via “reducing”
n
by the subprojection sequence (P̃ (s2 ) )n≥1 ,
n n
lim (Cs2n )P̃ (s2n ) ⊗ R⊗s2 ⊂ lim (Ds2n )P̃ (s2n ) ⊗ R⊗s2 ,
−→ −→
n n
is conjugate to that obtained via reducing by the sequence of projections

n
(P (s2 ) )n≥1 ,
n n
lim (Cs2n )P (s2n ) ⊗ R⊗s2 ⊂ lim (Ds2n )P (s2n ) ⊗ R⊗s2 .
−→ −→
n n
For proving this we shall first need the following lemma:

Lemma 12.
(n)
tr(P (n) zλ )
lim (n)
= 1,
n→∞ tr(P (n) ) tr(zλ )
(k,l) (n)
for any Young diagram λ ∈ Λnk , where zλ is the minimal central idem-
(n+1) (n+1)
potent of B 1 (2n − 1) corresponding to λ, and where zλ is zj(λ) , with
(k,l)
j(λ) ∈ Λ(n+1)k obtained by adding to to the left of λ a column of height k.
Proof. Let G be the inclusion matrix for B 1 (2n − 1) ⊂ B 1 (2n + 1). Consider
n > n0 large, and note that G does not depend on n by periodicity. Let G̃
be the one for B 1 (2n − 1)P (n) ⊂ B 1 (2n + 1)P (n+1) , where the embedding is
as before: For x ∈ B 1 (2n − 1), P (n) xP (n) 7→ P (n+1) xP (n+1) . The trace is
renormalised at B 1 (2n − 1)P (n) as before: tr(n) = tr(P1(n) ) tr, where tr is the
usual trace.
Because n is large, P (n) is full in B 1 (2n − 1), and G̃ does not depend
on n (the graph is periodic) and is “like” G, with the difference that the
multiplicity of some edges of G may be larger than the corresponding ones
for G̃. By [W-1], both G̃ and G are normal. In particular, the Perron-
Frobenius eigenvectors for G and Gt (resp. for G̃ and G̃t ) coincide. Also, by
fullness of P (n) , the trace weight vectors ~t (n) and ~t̃ (n) for B 1 (2n − 1) and
for B 1 (2n − 1)P (n) are multiples of each other; moreover, ~t̃ (n) = tr(P
~t (n)
(n) ) . Let
~a (n) and ~ã (n) be the dimension vectors for B 1 (2n−1) and for B 1 (2n−1) (n) , P
respectively. By P-F theory (see [W-1]), if β (resp. β̃) is the P-F eigenvalue
for G (resp. G̃), then
t n !n
G G̃ t
lim ~a (n0 ) (n )
= c ~t 0 and lim ~ã (n0 ) = c̃ ~t̃ (n0 ) ,
n→∞ β n→∞ β̃
where c, c̃ ∈ C. We can compute these constants:
n
Gt

~t (n+n0 )~a (n+n0 ) ~t (n+n0 ) β n
X X
(n0 )
1 = tr(1) = λ λ = λ ~a
β λ
λ λ
2
~t (n0 ) c ~t (n0 )
X
−→ = c ~t (n0 ) ,

n→∞ λ λ
λ
so that c = k~t (n0 ) k−2 . Similarly, c̃ = tr(P (n0 ) )2 k~t (n0 ) k−2 . Therefore, if
(k,l)
λ ∈ Λ(n+n0 )k ,
~t̃ (n+n0 )~ã (n0 ) 2

c̃ ~t̃
(n+n0 )
tr(P (n+n0 ) zλ ) λ
(n+n0 )
λ
lim = lim = = 1.
n→∞ tr(P (n+n0 ) ) tr(z (n+n0 ) ) n→∞ ~ (n+n0 ) (n+n0 ) (n0 ) 2
λ tλ ~a c ~t λ

Proof of Proposition 11. By definition, if n0 is large and n > n0 , we have

(n)
tr(P (n) zλ ) (n) (k,l)
tr(P̃ (n) ) = (n) , for some zλ ∈ Λnk . Since we can identify all the fi-
tr(zλ )
(k,l) tr(P̃ (n) )
nite sets Λnk via the bijection j, it follows by Lemma 12 that lim tr(P (n) ) =
n→∞
n
1. Consider the canonical embeddings given by φn : (Ds2n )P (s2 ) n ⊗ R⊗s2 →
n n n n n
(Ds2n )P̃ (s2n ) ⊗ R⊗s2 , with φn (P (s2 ) xP (s2 ) ⊗ r) = P̃ (s2 ) xP̃ (s2 ) ⊗ r for
76 JULIANA ERLIJMAN
n
x ∈ Ds2n , r ∈ R⊗s2 . Then
n n
φ = (φn ) : lim (Ds2n )P (s2n ) ⊗ R⊗s2 → lim (Ds2n )P̃ (s2n ) ⊗ R⊗s2
−→ −→
n n
is an isomorphism of the inductive limits which conjugates our pairs: φ is

tr(P̃ (n) ) (s2n ) ⊗ 1) ) = (P̃ (s2n ) ⊗
clearly surjective. Since lim tr(P (n) ) = 1, then φ((P n
n→∞
1)n , and injectivity is a consequence of this fact.
Theorem 13. The asymptotic inclusion for the one-sided inclusion of sub-
factors A ⊂ B, A ∨ (A0 ∩ B (∞) ) ⊂ B (∞) , is conjugate to the two-sided
inclusion C ⊂ D.
Proof. By the identification shown at the beginning of Section 3.2, before
Lemma 8, we have A ∨ (A0 ∩ B (∞) ) ⊂ B (∞) ∼ = Cp∞ ⊂ Dp∞ . The statement
follows from this observation, Corollary 7, Lemma 10 and Proposition 11.

Remark. If C ⊂ D is a two-sided braid inclusion of hyperfinite II1 factors

defined as in [E] (i.e., via unitary braid representations which are locally
finite dimensional and have properties as (i)-(vi) in the preliminaries), then
C ⊂ D is conjugate to Cp∞ ⊂ Dp∞ , where the latter is an “infinitely re-
duced” version of C ⊂ D by a descending sequence of projections (pn ), with
pn ∈ Cn and tr(pn ) = cn for some fixed c ∈ C. This is just the observation
that for proving Theorem 13 we only used these assumptions.
3.3. We include here the description for the inclusions A0 ∩ B (n) ⊂ A0 ∩
B (n+1) mentioned in the proof of Wenzl’s theorem. The embeddings are
given by the ψn ’s as at the beginning of 3.2. To draw the inclusion diagrams
for A0 ∩B (n) ⊂ A0 ∩B (n+1) in general, for any 1 < k < l, l ≥ 3, one alternates
according to the parity of n in the following way:
If n is odd, then A0 ∩ B (n) ⊂ A0 ∩ B (n+1) is given by simply considering
the diagram for B( n+1 )k ⊂ B( n+1 )k+1 , and restricting it to the simple ideals
2 2
of B( n+1 )k which correspond to classes of minimal idempotents pλ , with
2
n+1
pλ ≤ p[1k−1 ] ⊗ 2 .
If n is even, then A0 ∩ B (n) ⊂ A0 ∩ B (n+1) is given by a subgraph of
the graph for B( n2 )k+1 ⊂ B( n2 +1)k , which is obtained by restricting it to the
simple ideals of B( n2 )k+1 which correspond to classes of minimal idempotents
⊗n
pλ , with pλ ≤ p[1k−1 ] 2 , and to the simple ideals of of B( n2 +1)k which
n
correspond to classes of minimal idempotents pµ , with pµ ≤ p[1k−1 ] ⊗ 2 +1 .
The principal graphs in general can be described the following simple way:
Fix n0 large. For the (k, l) case the principal graph is given by the inclusion
diagram for Bn0 k ⊂ Bn0 k+1 , where the even vertices are labelled by (k, l)
diagrams of n0 k boxes, and the odd vertices by (k, l) diagrams of n0 k + 1
boxes. The ∗-vertex is given by the k × n0 rectangle. Let us also observe

that the dual principal graph is the same, which follows by applying similar
arguments.
For example, the Bratteli diagrams and principal graphs in the cases k = 3
and l = 6, 7 are the following:
78 JULIANA ERLIJMAN
References
[B] D. Bisch, On the existence of central sequences in subfactors, Trans. AMS, 321
(1990), 117-128.
[Ch] M. Choda, Index for factors generated by Jones’ two sided sequence of projections,
Pacific J. Math., 139 (1989), 1-16.
[E] J. Erlijman, New braid subfactors from braid group representations, Trans. AMS,
350 (1998), 185-211.
[EK] D. Evans and Y. Kawahigashi, Orbifold subfactors from Hecke algebras, Comm.
Math. Phys., 165 (1994), 445-484.
[FH] W. Fulton and J. Harris, Representation Theory. A First Course, Springer-Verlag,
Graduate Texts in Mathematics, 129 (1991).
[GW] F. Goodman and H. Wenzl, Littlewood-Richardson coefficients for the Hecke algebra
at roots of unity, Adv. Math., 82 (1990), 24-45.
[G] S. Goto, Quantum double construction for subfactors arising from periodic com-
muting squares, preprint, 1996.
[J] V. Jones, Index for subfactors, Invent. Math., 72 (1983), 1-25.
[McD] D. McDuff, Central sequences and the hyperfinite II1 factor, Proc. London Math.
Soc., 21 (1970), 443-341.
[O] A. Ocneanu, Chirality of Operator Algebras, ‘Subfactors’, Taniguchi Symposium
on Operator Algebras, World Scient., (1994), 39-63.
[Wa] A. Wassermann, Coactions and Young-Baxter equations for ergodic actions and
subfactors, London Math. Society, Lect. Note Series, 136 (1988), 203-236.
[W-1] H. Wenzl, Hecke algebras of type An and subfactors, Invent. Math., 92 (1988),
349-383.
[W-2] , Quantum groups and subfactors of Lie type B, C, and D, Comm. Math.
Phys., 133 (1990), 383-433.
[W-3] , Braids and invariants of 3-manifolds, Invent. Math., 114 (1993), 235-275.
[W-4] , Lecture at the Mittag-Leffler Institute, 1988, personal communication.
Received April 1, 1998 and revised November 10, 1998.
University of Regina
Regina, Sask. S4S 0A2
Canada
E-mail address: erlijman@math.uregina.ca
Vol. 193, No. 1, 2000
CHARACTERIZATION OF HARDY SPACES BY

SINGULAR INTEGRALS AND ‘DIVERGENCE-FREE’
WAVELETS
J.E. Gilbert, J.A. Hogan, and J.D. Lakey
In memory of A. Uchiyama
The Hardy space Hρ1r (Rn ) consists of all divergence free

r-form distributions f whose non-tangential maximal func-
tions are in L1 (Rn ). We say that a system of singular integrals
characterizes Hρ1r (Rn ) if this space consists precisely of those
divergence-free r-form distributions f whose images under the
singular integral operators are integrable. When the opera-
tors are determined by Fourier multipliers, necessary and suf-
ficient conditions are prescribed on the multipliers in order
that the system characterize Hρ1r (Rn ). The condition is anal-
ogous to the Janson–Uchiyama condition for the scalar-valued
case and the characterization follows the lines of Uchiyama’s
constructive decomposition of BMO. In particular, it is shown
how to build divergence-free r-form wavelets which play the
same role that the R. Fefferman–Chang elementary decompo-
sition played in Uchiyama’s work.
1. Introduction.
In this paper we continue to develop real-variable theory for Hardy spaces
Hρpr (Rn ) of divergence-free, or, more precisely, ‘co-closed’ r-forms. Though
these spaces arise as the ‘real parts’ of boundary values of r + 1-form so-
lutions of the Hodge-deRham (d, d∗ )-system in the upper half-space Rn+1 + ,
the most natural definition for them is a strictly real variable one using the
non-tangential maximal function. A characterization by singular integrals
then seeks an understanding of the relationship between the two possible
definitions. The characterization given here is based on Uchiyama’s con-
structive decomposition arguments, with a view towards broadening the
scope of those arguments to other geometric settings.
Let H be a finite-dimensional real Hilbert space with inner product h · , · i
and norm k · kH . The real variable Hardy space H p (Rn , H), 0 < p < ∞,
79
80 J.E. GILBERT, J.A. HOGAN, AND J.D. LAKEY
consists of those H-valued distributions for which

Z 1/p
(1.1) kf kH p = |N (f )(x)|p dx
Rn
is finite where
(1.2) N (f )(x) = sup ku(y, t)kH
|x−y|<t
is the non-tangential maximal function of the Poisson extension u(x, t) of

f . As is well-known, H p (Rn , H) = Lp (Rn , H) when 1 < p < ∞ and the
notation H p , Lp will be used interchangably for such values of p. The usual
Hardy H p (Rn )-spaces correspond to the scalar case H = R. For the most
part in this paper H will be the rth -exterior power Λr (Rn ) of Rn , and we
shall use the term divergence free r-form rather loosely to mean any function
or distribution f : Rn → Λr (Rn ) such that
n
X ∂f
(1.3) d∗ f = µ∗j =0
∂xj
j=1
where µ∗jdenotes interior multiplication by the element ej in the standard
basis for Rn (cf. Section 3 for more details on forms and their properties). By
a distribution we shall always mean one which is restricted at infinity in the
sense of Stein so that harmonic extensions and the action of singular integrals
on such distributions are well-defined ([16], p. 123). The divergence-free
condition then singles out a closed subspace of H 1 (Rn , Λr (Rn )).
Definition 1.4. The Hardy space Hρ1r (Rn ) consists of all divergence free
r-form distributions f whose non-tangential maximal function N (f ) is in
L1 (Rn ).
Alternatively, one can follow the Stein and Weiss formulation of Euclidean
H p -theory in the 1960’s (e.g., [17], [18]), beginning with boundary values
of a space of r + 1-form solutions of the Hodge-deRham (d, d∗ )-system on
the upper half-space Rn+1
+ . Due to relations among their components these
boundary values form a closed subspace of L1 (Rn , Λr+1 (Rn+1 )). Indeed,
Xn
(1.5) f −→ e0 ∧ f + ej ∧ (Rj f ) = (I + R)f
j=1
defines a 1-1 mapping from Hρ1r (Rn ) onto this space of boundary values,
turning Hρ1r (Rn ) into a space of ‘real parts’ of boundary values ([7], [8]). A
divergence-free distribution f then belongs to Hρ1r (Rn ) if and only if (I +
R)f belongs to L1 (Rn , Λr+1 (Rn+1 )). This definition emphasizes the relation
between over-determinedness of the (d, d∗ )-system and the ‘divergence-free’
or so-called ‘gauge’ condition on these ‘real parts’; implicit also is the role
of sub-harmonicity in arriving at the L1 -space of boundary values.
CHARACTERIZATION OF HARDY SPACES 81
A major advance in understanding H 1 -spaces without regard to harmonic-

ity occured when C. Fefferman and Stein characterized H 1 (Rn ) in terms of
the grand maximal function ([5]). But could anything comparable be said
about singular integrals? The need to understand the relationship between
these two approaches – one requiring the non-degeneracy of various maxi-
mal functions, the other requiring non-degeneracy of a system of singular
integrals – led C. Fefferman to raise the question of describing the families of
singular integral operators that characterize H 1 ([4]). In the present context
singular integral will mean a standard Calderón-Zygmund integral operator
Z D E
f (x) −→ fˆ(ξ), m (ξ) e2πix.ξ dξ
Rn
determined by its ‘multiplier’ m = m (ξ), a Λr (Cn )-valued C ∞ -function

which is homogeneous of degree 0. The multiplier can thus be identified
with a function m = m (ω) on the unit sphere Σn−1 and we shall refer to
such an operator, or systems of such operators, as a singular integral oper-
ator having ‘smooth symbol m’; in the case of a system K = (K1 , . . . , KM )
the symbol will be an operator m = m (ω) : Λr (Cn ) → CM . The operator
will be real if and only if m(−ω) = m(ω), ω ∈ Σn−1 . Fefferman’s question
now becomes:
Problem 1.6. Prescribe necessary and sufficient conditions for a system

of singular integrals K = (K1 , . . . , KM ) to characterize Hρ1r (Rn ) in the sense
that Hρ1r (Rn ) consists precisely of those divergence-free r-form distributions
f for which Kf ∈ L1 (Rn , CM ).
One might also hope that (1.6) provides further insight into the role of
over-determinedness in (1.5) without regard to subharmonicity. On one
hand, we can consider Hρ1r as a prototype of Hardy spaces of distributions
satisfying some added geometric constraint – the divergence-free condition
in this case – and our approach to problem (1.6) reflects this point of view.
On the other hand, the precise geometric structure of the present case allows
for a particularly simple solution as in the proof of Theorem A.
To arrive at a solution to (1.6) we interpret the divergence free condition
on the Fourier transform side. The crucial idea is to use (1.3) to provide
an orthogonal decomposition Λr (Cn ) = Nr (ω) ⊕ Tr (ω) of Λr (Cn ) at each
point ω of the unit sphere Σn−1 in Rn , splitting Λr (Cn ) into its respective
‘normal’ and ‘tangential’ components
Nr (ω) = a ∈ Λr (Cn ) : ω ∧ a = 0 , Tr (ω) = Nr (ω)⊥

(1.7)
at ω. This splitting corresponds to the Hodge Decomposition of H 1 (Rn , Λr ).

Theorem A. A system of singular integrals K characterizes Hρ1r (Rn ) if

and only if for each ω ∈ Σn−1 its symbol m(ω) has the property

n−1
rank(m(ω) ⊕ m(−ω)) = 2
r
as a mapping (a, b) → m(ω)a + m(−ω)b from Tr (ω) ⊕ Tr (−ω) into CM .
The algebraic condition in Theorem A thus requires the restriction of

m(ω) ⊕ m(−ω) to Tr (ω) ⊕ Tr (−ω) be 1-1 since

n−1 n−1
(1.8) dimC Nr (ω) = , dimC Tr (ω) = .
r−1 r
It might be 1-1 on a larger subspace of Λr (Cn ) ⊕ Λr (Cn ), of course, but it is

only the behaviour on Tr (ω)⊕Tr (−ω) that is of importance for characterizing
Hρ1r (Rn ). For 0-forms, T0 (ω) = C and so in the case of complex-valued
functions Theorem A reduces to the well-known result of Uchiyama and
Janson ([10], [19]). In the more general form-valued case Theorem A follows
from the corresponding characterization of H 1 (Rn , Λr (Cn )). To achieve this
we augment the system K obtaining an auxiliary system K̃ that characterizes
H 1 (Rn , Λr (Cn )) and whose restriction to Hρ1r (Rn ) agrees with K.
This simple proof does not capture the full import of Uchiyama’s solution
of problem (1.6), however, because he gave at the same time a construc-
tive version of the C. Fefferman-Stein decomposition of BMO(Rn ). Indeed,
Uchiyama showed that:
(i) the maximal rank condition on m(ω) provides an ‘adaptive-pseudo-
invertibility’ criterion on K, so that
(ii) for each g ∈ BMO(Rn ) one can construct inductively a function h ∈
L∞ (Rn , CM ) with K∗ h = g.
The construction is carried out by applying an appropriate pseudoinverse
to each term in the R. Fefferman-Chang elementary particle expansion of
g. In fact, Uchiyama’s argument is valid in any vector-valued BMO-type
space so long as the functions possess an elementary particle expansion and
the norm is determined by a Carleson condition on the coefficients in the
expansion. More precisely, let B be a closed subspace of BMO(Rn , H) where
H is a finite-dimensional Hilbert space. Suppose furthermore,
(1.9)(a) for each dyadic cube Q there is a function ψQ ∈ B with

Z
supp(ψQ ) ⊆ mQ, ψQ (x) dx = 0, |Dα ψQ (x)| ≤ const. |Q|−1/2−|α|/n ,
Rn
uniformly in Q for all 0 ≤ |α| ≤ 1,

P
(1.9)(b) each g in B has an expansion g ∼ Q cQ ψQ , converging in the sense
of distributions, so that
 1/2
1 X
g −→ sup  |cQ0 |2 
Q |Q|
Q0 ⊆Q
provides an equivalent norm on B, and
(1.9)(c) conversely, to each sequence {cQ } satisfying the

P Carleson condition
of (1.9)(b) there corresponds a function g ∼ Q cQ ψQ .
The restriction to B of any standard Calderón-Zygmund operator with

symbol m(ω) : H → CM will automatically map B into BMO(Rn , CM ), and
its adjoint will map BMO(Rn , CM ) into BMO(Rn , H). But this adjoint need
not have range in B, of course, and the abstract formulation of Uchiyama’s
result has to reflect this fact. For convenience, it is stated here only for real
systems.
Constructive Decomposition Theorem. Let B be a subspace of

BMO(Rn , H) satisfying (1.9) and let K be a real standard singular inte-
gral system such that to every ν ∈ ΣM −1 there corresponds a system Lν for
which
K∗ ◦ Lν = I, ν · Lν = 0
as operators from B into BMO(Rn , H) and BMO(Rn , CM ), respectively.
Then for each g in B one can construct an h in L∞ (Rn , CM ) and f in
BMO(Rn , H) so that K∗ h + f = g and khk∞ ≤ const. kgkB ; furthermore,
if B can be identified with the dual space of a subspace of H 1 (Rn , H), then
hf, G i = 0 for all G in the pre-dual of B.
The detailed proof in the scalar case given by [1] carries over to the vector-
valued case by using techniques that Uchiyama showed us after he read an
earlier draft of this paper [20]. Both the support and C 1 conditions on the
ψQ in (1.9), as well as the hypothesis that K is real, can be weakened at the
expense of more technical book keeping. The term f arises in the induc-
tive construction because at each stage a nonlinear splitting in L∞ (Rn , CM )
is performed on terms of the form Lν (ψQ ) that takes these terms outside
Lν (B). When B = BMO(Rn , H), then f is necessarily zero.
In the present setting, let B = BMOρr denote the dual space of Hρ1r
(cf. (4.4)). The Hodge decomposition ensures that Hρ1r is a complemented
subspace of H 1 (Rn , Λr ), and so in turn B = BMOρr is a complemented
subspace of BMO(Rn , Λr ). In particular, B can be identified with the dual
of a subspace of H 1 (Rn , Λr ).
Theorem B. If K satisfies the rank condition of Theorem A, then for each

g ∈ BMO
ρr one can construct an h ∈ L∞ (Rn , CM ) with khk∞ ≤ CkgkBMOρr
∗
and (K H 1 ) h = g.

ρr
Since KH 1 = K ◦ (R∗ R) (cf. (3.7)), the Constructive Decomposition the-

ρr
orem can be applied to K 1 . The term f in that theorem is identified
Hρr
with zero as an element of the dual of Hρ1r . Theorem B is thus a corol-
lary of Uchiyama’s Constructive Decomposition theorem once one produces
the family ψQ and verifies the existence of pseudoinverses Lν satisfying the
pseudoinverse condition in the Constructive Decomposition theorem. The
ψQ that we construct will form a bi-orthogonal system of wavelets – refine-
ments of the R. Fefferman-Chang ‘elementary particles’ which, in addition
to satisfying (1.9), are divergence-free. (cf. Theorem C and Theorem 4.6).
The properties of these wavelets are more convenient than crucial. For ex-
ample, the images of standard wavelets under R∗ R would provide sufficient
tools for the constructive decomposition, though they would not have com-
pact support. But the wavelets that we construct should be of independent
use; moreover, their tensor product structure also facilitates the proof of
necessity of the rank condition in Theorem A.
After some preliminary sections, Theorem A is established in Section 5.
The existence of operators Lν in the case B = BMOρr can be established
along lines similar to those used in proving the sufficiency of the rank con-
dition in Theorem A, that is, by augmenting K to K̃, finding corresponding
L̃ν , then restricting to Hρ11 . However, in keeping with the point of view of
the abstract Uchiyama theorem, we will give a more complicated proof of
the pseudoinverse condition – one that can be followed for decomposition of
spaces B where the ‘augmentation trick’ could fail. The pseudoinverse cri-
terion is established in Section 6. Theorem B will then follow immediately
from the abstract Uchiyama theorem for the case B = BMOρr . To keep
technicalities to a minimum we only verify Theorem A and the pseudoin-
verse criterion for real systems K, although our arguments can be extended
to prove Theorems A and B for complex K. There are very important ex-
amples of K that do not preserve real-valued functions, see Section 2. In
Section 7 the wavelets are constructed to complete the proof of Theorem B.
2. One-form case, examples.

The significance of the rank condition in Theorem A is most easily illus-
trated in the case of 1-forms f = (f1 , . . . , fn ), i.e., Rn -valued functions and
distributions, where ‘divergence-free’ has its traditional meaning
∂f1 ∂fn
(2.1) d∗ f = + ··· + = 0.
∂x1 ∂xn
The decomposition Cn = N1 (ω) ⊕ T1 (ω) of Cn corresponds to the familiar

splitting of any 1-form into its ‘divergence’ and ‘curl-free’ parts. It also lives
up to the geometric terminology used:
N1 (ω) = a ∈ Cn : a ∧ ω = 0 = C ω

(2.2)(i)
can be identified with the complexification of the one-dimensional subspace
of Rn passing through ω, while the subspace
T1 (ω) = a ∈ Cn : a · ω = 0 = N1 (ω)⊥

(2.2)(ii)
orthogonal to N1 (ω) can be identified with the complexification of the tan-
gent plane a ∈ Rn : a · ω = 0 to Σn−1 at ω; in other words, N1 (ω) is the

complexification of the normal line to this tangent plane.
2.3. Canonical Example. The symbol of the mapping (1.5) is given by

m(ω) : a −→ e0 ∧ a − i ω ∧ a (a ∈ Cn )
where {e0 , . . . , en } is the extension of the standard basis of Rn to a basis
for Rn+1 . Thus
m(ω)a + m(−ω)b = e0 ∧ (a + b) − i ω ∧ (a − b) = 0 (a, b ∈ Cn )
if and only if a + b = 0 and a − b ∈ N1 (ω), i.e., ker(m(ω) ⊕ m(−ω)) =
{(a, −a) : a ∈ N1 (ω) }. Hence the restriction of m(ω) ⊕ m(−ω) to T1 (ω) ⊕
T1 (−ω) has maximal rank, so I + R characterizes Hρ11 (Rn ).
2.4. Partition of Unity Example. Using spherical polar coordinates

ω = (sin θ ω0 , cos θ), ω0 ∈ Σn−2 , 0 ≤ θ ≤ π,
for Σn−1 , we can use Theorem A to show that each f in Hρ11 (Rn ) is the sum
of 2n − 1 divergence-free 1-forms each of which can be extended analytically
to some tube domain. This is in complete analogy to Carleson’s result for
H 1 (Rn ), and, just as in Uchiyama’s proof of that result, a partition of unity
φ1 (ω) + · · · + φ2n−1 (ω) = 1 (ω ∈ Σn−1 )
is required in which each φj is a real function having support in a portion
of Σn−1 lying in some half-space (cf. [19], Corollary 3). The restriction on
the support ensures that for each ω there exist j, k such that
φj (ω) 6= 0, φj (−ω) = 0; φk (ω) = 0, φk (−ω) 6= 0.
Consequently, if
m(ω) = (φ1 (ω), . . . , φ2n−1 (ω)),
then m(ω) ⊕ m(−ω) is 1-1 from T1 (ω) ⊕ T1 (−ω) into C2n−1 , and so the
corresponding K having m = m(ω) as symbol characterizes Hρ11 (Rn ). The
connection of this analytic extension of Hρ11 (Rn ) functions with the func-
tional calculus developed by McIntosh remains to be investigated ([12]).
The construction of the φj proceeds by induction on n. For the unit circle

Σ1 there exist C ∞ -functions forming a partition of unity φ1 + φ2 + φ3 ≡ 1
where each function has support in some arc of length less than π. In
passing to Σ2 , the previous three functions provide a partition of unity
with respect to the longitudinal variable ω0 , but some care must be taken
with the latitudinal variable θ because of the ambiguity of spherical polar
coordinates at the North and South Poles. Let φ+ = φ+ (θ) be an even C ∞
bump-function on R such that
φ+ (θ) = 0, |θ| ≥ θ0 , φ+ (θ) = 1, |θ| ≤ θ1
for some 0 < θ1 < θ0 < π/2, and then define φ− by φ− (θ) = φ+ (π − θ).
Finally, let φ = φ(θ) be a C ∞ -function having support in (0, π) so that
φ+ + φ + φ− is a partition of unity on [0, π]. By construction,
φ+ (θ) + {φ1 (ω0 ) + φ2 (ω0 ) + φ3 (ω0 )}φ(θ) + φ− (θ)
will then provide an unambiguously defined partition of unity by 5 functions
on Σ2 each of which has support in a portion of Σ2 lying in some half-space.
It is clear how to generalize this to any Σn−1 .
2.5. Basis Example. The previous example can be refined to express

1 n
P2n−1 Pn−1
each f ∈ Hρ1 (R ) as a sum f = j=1 k=1 fjk of basis functions each
of which admits an analytic extension to some tube domain. We need a
cross-section κ : Σn−1 \ {−1} → SOn . With the notation of (2.4) set
I − (1 − cos θ) ω0 ⊗ ω0 − sin θ ω0t
 
κ(ω) =  ,
sin θ ω0 cos θ
regarding ω0 as a row matrix and its transpose ω0t as a column matrix. Under
matrix multiplication on the right, 1.κ(ω) = ω; in addition, κ(ω) maps the
tangent space T1 (1) ∼
= Rn−1 at the North Pole to the tangent space T1 (ω) at
ω. The image of the {e1 , . . . , en−1 } for T1 (1) will thus be a basis for T1 (ω),
and we can use this together with the previous partition of unity to define
a new symbol
 
φ1 (ω)κ(ω)e1 . . . φ2n−2 (ω)κ(ω)e1 φ2n−1 (ω)κ(−ω)e1
 . ... . . 
m(ω) =  ,
 . ... . . 
φ1 (ω)κ(ω)en−1 . . . φ2n−2 (ω)κ(ω)en−1 φ2n−1 (ω)κ(−ω)en−1
where it is assumed that φ2n−1 = 1 in a neighborhood of the South Pole to
ensure that κ(−ω) is well-defined. For the same reasons as before, m(ω) ⊕
m(−ω) is 1-1 from T1 (ω) ⊕ T1 (−ω) into C(n−1)×(2n−1) , and so again the
corresponding K characterizes Hρ11 (Rn ). Notice that m(ω) ⊕ m(−ω) will not
be 1-1 on all of Cn ⊕ Cn . This example should also be contrasted with the

wavelet representation of f given later in (4.6).
2.50 . Normal Space Example. Let Hρ̃11 denote the complement of Hρ11 ,
that is, the image of RR∗ , inside of H 1 (Rn , Rn ). Just as above, we can
construct a symbol

n(ω) = φ1 (ω)κ(ω)en . . . φ2n−2 (ω)κ(ω)en φ2n−1 (ω)κ(−ω)en .
Then n(ω) ⊕ n(−ω) will be 1-1 on N1 (ω) ⊕ N1 (−ω) into C1×(2n−1) , and so
the same principles imply that the corresponding system N will characterize
Hρ̃11 (Rn ). This example can readily be generalized to a system Nr charac-
terizing Hρ̃1r . Crucial to the proof of Theorem A will be the fact that for
any f ∈ Hρ1r , one has Nr f = 0.
2.6. Failure of Projection Examples. As an operator on Cn the symbol

of the operator R∗ R has the property

X
m(ω)a, b = ωj ωk aj bk ,
j,k
so m(ω)a + m(−ω)(−a) = 0 for any a ∈ Cn . Thus the Hodge projection

R∗ R does not characterize Hρ11 (Rn ). Similarly, neither of I + R∗ R, I + RR∗
characterizes Hρ11 (Rn ). These last two examples are analogous to the first
‘nontrivial’ scalar-valued example {I, R · R} which failed to characterize
H 1 (Rn ) (cf. [6]).
3. Hodge decomposition, r-forms.

To facilitate reading of the paper it may be helpful to recall some known
though not always familiar algebraic ideas ([9], pp. 49-65). The exterior
algebra
X n
n
(3.1) Λ∗ (R ) = ⊕ Λr (Rn ), Λ0 (Rn ) = R, Λ1 (Rn ) = Rn , . . .
r=0
associated with Rn is the algebra of linear combinations of wedge products
v1 ∧ · · · ∧ vr of vectors in Rn . Wedge products
(3.2) eα = ei1 ∧. . .∧eir ; α ∈ Pr = {(i1 , . . . , ir ) : 1 ≤ i1 < · · · < ir ≤ n}
of the standard basis {e1 , . . . , en } for Rn provide a basis for Λr (Rn ); by
convention, e∅ = 1 in the case r = 0 when α is then the empty set ∅
and Λ0 (Rn ) = R. The standard basis for Rn also extends to a standard
basis {e0 , e1 , . . . , en } for Rn+1 . There is a canonical inner product h·, ·i on
Λr (Rn ) with respect to which it is a real Hilbert space of dimension nr
having {eα : α ∈ Pr } as an orthonormal basis; Λ∗ (Rn ) is then a sum of
Hilbert spaces, hence a real Hilbert space itself. To simplify notation we

shall often write Λr , Λ∗ instead of Λr (Rn ), Λ∗ (Rn ). When Fourier transforms
arise the corresponding complex exterior algebra Λ∗ (Cn ) along with its inner
product structure will be needed, but in such cases full notation will always
be retained to avoid any confusion.
The inner product structure on Λ∗ permits the introduction of Banach
spaces of Λ∗ -valued functions. For instance, the Lebesgue space of all square-
integrable r-forms
X
(3.3)(i) f : Rn −→ Λr , f (x) = fα (x)eα , (0 ≤ r ≤ n)
α∈Pr
on Rn will be denoted by L2 (Rn , Λ 2 n

r ); similarly, L (R , Λ∗ ) will denote the
space of all square integrable forms without restriction on r. In both cases
the inner product is given by
Z

(3.3)(ii) f,g = f (y), g(y) dy
Rn
with respect to the inner product on Λ∗ . Many operators on forms exploit
the algebra structure on Λ∗ : The wedge product a → ej ∧ a defines the
exterior multiplication operator µj : Λ∗ → Λ∗ mapping Λr to Λr+1 , 0 ≤ r <
n; its Hilbert space adjoint is the interior multiplication operator µ∗j : Λ∗ →
Λ∗ mapping Λr+1 to Λr , 0 ≤ r < n. Direct computation shows that
(3.4) µj µk + µk µj = 0; µ∗j µ∗k + µ∗k µ∗j = 0; µ∗j µk + µk µ∗j = δjk I.
The exterior and interior Riesz transform operators
n
X n
X
(3.5) R= µk Rk ; R∗ = − µ∗k Rk ,
k=1 k=1
thus define bounded operators on L2 (Rn , Λ ∗) such that

(3.6) Z Z
f (x), R∗ g(x) dx = f, R∗ g .

Rf, g = Rf (x), g(x) dx =
Rn Rn
In view of (3.4),
(3.7) RR = 0, R∗ R∗ = 0; RR∗ + R∗ R = I
since k Rk2 = −I. An immediate consequence of this is a crucial splitting
P
of L2 (Rn , Λr ).
Theorem 3.8 ([7] Hodge decomposition). The operators RR∗ and R∗ R
define orthogonal projections from L2 (Rn , Λr ) onto
{f ∈ L2 (Rn , Λr ) : Rf = 0} and {f ∈ L2 (Rn , Λr ) : R∗ f = 0}
respectively for each r = 0, 1, . . . , n.
The relation of this to the standard decomposition in terms of the Hodge-

deRham (d, d∗ )-system on Rn is clear from the fact that
R = (−∆)−1/2 ◦ d; R∗ = (−∆)−1/2 ◦ d∗ .
In particular, we shall make frequent use of the fact that the massless con-
dition R∗ f = 0 and the divergence-free condition d∗ f = 0 are equivalent in
the sense of distributions modulo polynomials ([8]). For future reference it
will also be important to note that all of these properties will persist in any
space of forms on which Riesz transforms are well-defined and bounded.
4. Wavelet analogues.
Wavelets give a precise analogue of the R. Fefferman-Chang elementary par-
ticle decomposition that played a fundamental role in the constructive de-
composition of BMO as described by Uchiyama and Christ-Geller. Better
yet, divergence-free wavelets are now available thanks to independent work
of Lemarié and Battle in the r = 1 case. Our construction of divergence-
free r-form wavelets in Section 7 is suggested, in turn, by a splitting of the
index set Pr in (3.2) analogous with the splitting (1.7) of Λr (Cn ). Here we
will describe the wavelets, their role in norming Hρ1r and its dual, and their
convenience in describing some Calderón-Zygmund operators.
Denote by E ∗ the family of (2n − 1) binary n-tuples = (1 , . . . , n )
in {0, 1}n \ (0, . . . , 0). E ∗ is sufficient to parametrize the mother wavelets
needed in the real-valued case, i.e., when r = 0 ([3], p. 317; [13], p. 84).
Indeed, there exist real-valued scaling functions φ = φ(x) with associated
compactly supported mother wavelets {ψ : ∈ E ∗ } such that the family of
usual translates and dilates
(4.1) ψ,Q (x) = 2nj/2 ψ (2j x − k) (Q = Qjk , ∈ E ∗ )
of all these wavelets provide an orthonormal basis for L2 (Rn ). The index
Q = Qjk ∈ Qj ⊂ Q denotes the dyadic cube of sidelength l(Q) = 2−j with
vertex at k/2j . Any such basis extends componentwise to the vector-valued
case.
Let {eα }α∈Pr denote the standard basis for Λr . Then {ψ ⊗eα : ∈ E ∗ , α ∈
Pr } is a family of (card E ∗ )(card Pr ) mother wavelets whose translates and
dilates provide an orthonormal basis of compactly supported functions for
L2 (Rn , Λr ). Because R∗ R is a singular operator, the divergence free im-
ages R∗ R(ψ ⊗ eα ) need not be compactly supported elements of Hρ2r (Rn ),
however. In fact, for the r = 1 case Lemarié has shown that there does
not exist an orthonormal basis of compactly supported wavelets for Hρ21 (Rn ),
though it is possible to build bi-orthogonal wavelets where the reconstruct-
ing wavelets are divergence-free, but no such constraint is imposed on the
analyzing wavelets ([11]). Thus the price one pays for local properties in
the case r > 0 is a loss of orthogonality, but biorthogonality is more than

sufficient for the present purpose.
For each ∈ E ∗ set
(4.2) i = min{i : i = 1 } ( ∈ E ∗ ).
The index set Pr of Λr (Cn ) can be split into two parts
(4.3)(i) Tr = {α : i ∈
/ α }, Nr = {α : i ∈ α }
analogous to the splitting (1.7) of Λr (Cn ). Note that
(4.3)(ii) dimC Nr (ξ) = card Nr , dimC Tr (ξ) = card Tr ,
independent of ∈ E ∗ . In Section 7 we shall prove the following result.
Theorem C. For each r, 1 ≤ r < n, and each integer m > 0 there is
a family of compactly supported r-forms {(ψ ,α , η ,α ) : ∈ E ∗ , α ∈ Tr }
having continuous derivatives as well as vanishing moments up to order m;
furthermore,
(i) ψ ,α is divergence-free and {(ψ ,α , η ,α ) : ∈ E ∗ , α ∈ Tr } is bi-orthogo-
nal in the sense that
0 ,α0
D E
,α
d∗ ψ ,α = 0, ψQ , ηQ 0 = δ0 δαα0 δQQ0 ,
hold, while
(ii) for 1 ≤ p < ∞, the mapping
XXD ,α
E
,α
f −→ R∗ Rf (x) = f, ηQ ψQ (x)
,Q α∈Tr
is a projection from H p (Rn , Λr ) onto Hρpr (Rn ).
These wavelets furnish convenient equivalent norms on Hρ1r (Rn ) and its
dual space BMOρr (Rn ). The latter is the quotient space of BMO(Rn , Λr )
having norm defined by induction on r:
( Z 1 )
1 2
2
(4.4) kgkBMO = sup inf |g(x) − Rh(x)| dx ,
Q h∈BMOρr−1 |Q| Q
the supremum being taken over all cubes in Rn having sides parallel to the
coordinate axes (cf., [8]). In the case of 0-forms it reduces to the usual
BM O-space of real-valued functions for which
( Z 1/2 )
1
(4.5) sup inf |h(x) − c|2 dx
Q c∈R |Q| Q
is finite.
Theorem 4.6 ([8]).

(i) Each f ∈ Hρ1r (Rn ) has a wavelet expansion
,α ,α
XX
f (x) = hf, ηQ iψQ
,Q α∈Tr
whose square function norm

 1/2
Z X X 2 2
,α ,α
kS(f )kL1 = hf, ηQ i ψQ (x)  dx

Rn ,Q α∈Tr
is equivalent to the non-tangential maximal norm (1.1) on Hρ1r (Rn ).

(ii) Each g ∈ BMOρr (Rn ) has a weak∗ -convergent expansion
X X ,α ,α
g(x) = cQ ψQ
,Q α∈Tr
for which the Carleson condition

 1/2
1 X ,α 2
sup  cQ0  < ∞
Q |Q| 0
Q ⊆Q∈Q
defines a norm on BMOρr (Rn ) equivalent to (4.4). Conversely, any

such sequence {c,α n
Q } gives rise to such an element g ∈ BMOρr (R ).
This result follows from tent space arguments based on the fact that
Theorem C provides a discrete resolution of the operator R∗ R. In fact,
similar arguments establish boundedness results for a large family of discrete
sum operators
D E
,α
Ψ,α
XX
(4.7) D : f −→ |Q|−1/2 f, ηQ Q (x)
,Q α∈Tr
where each Ψ,α

Q CM -valued
is a function which is assumed to have a vanish-
ing moment as well as satisfy ‘standard’ size and smoothness conditions, cf.,
[7]. Indeed, this observation plays a crucial role in the constructive decom-
position of BMOρr , where Ψ,α ,α
Q is the image of ψQ under a pseudoinverse
LνQ of K as described in Section 6.
The wavelet expansions in (4.6) are particularly useful in computing re-
striction results. Let 1 = (0, . . . , 0, 1) be the North Pole in Σn−1 and set
Z Z ∞
P1 : f −→ P1 f (t) = f (y + t1) dy = e2πist fˆ(s1) ds (t ∈ R).
Rn−1 −∞
Formally, this is given on the transform side as the restriction to the line
through the North Pole. Basic to the proof of necessity of the rank condition
in Theorem A is the identification of the image of Hρ1r (Rn ) under P1 .
Theorem 4.8. The restriction P1 is bounded from Hρ1r (Rn ) onto

H 1 (R, Λr (Rn−1 )).
By the closed graph theorem therefore,
kP1 f kH 1 = inf kgkH 1 : g ∈ Hρ1r (Rn ) : P1 g = P1 f .

(4.9)
The corresponding result for L1 (Rn , RM ) is well-known (cf. [15]). A proof

of (4.8) appears at the end of Section 7.
5. Proof of Theorem A.
In what follows we shall assume that K is real. Proof of necessity of the
maximal rank condition in Theorem A exploits some of the same ideas as
Janson’s proof in the scalar-valued case by reducing the setting to any one-
dimensional subspace of Rn which after rotation we can always assume is
R1. The restriction of K to this line defines a singular integral
Z
(5.1) K1 : h −→ e2πist m(s1)ĥ(s) ds = m(1)(h − ih̃) + m(−1)(h + ih̃),
R
where h̃ is the Hilbert transform of h. Furthermore, K1 commutes with P1

in the sense that K1 ◦ P1 = P1 ◦ K.
Proof of necessity. Now suppose that K characterizes Hρ1r (Rn ). By the
closed graph theorem, therefore,
(5.2) A0 kf kH 1 ≤ kK1 f kL1 + · · · + kKM f kL1 (f ∈ Hρ1r (Rn ))
for some constant A0 > 0. Using (4.8) we shall prove that
(5.3) khkH 1 ≤ const. kK1 hkL1
then holds for all h ∈ H 1 (R, Λr (Rn−1 )). Since K, and hence K1 is real, (5.3)
will continue to hold for all H 1 -functions h : R → Λr (Cn−1 ) as well, enabling
us to apply it also to h − ih̃, h + ih̃. From this the necessity of the maximal
rank condition follows quickly.
First we establish (5.3). By (4.9) the inequality
(5.4) A kP1 f kH 1 ≤ inf {kK1 gkL1 + · · · + kKM gkL1 } ,
g
holds after restriction of (5.2) to R1, the infimum being taken over g ∈
Hρ1r (Rn ) with P1 g = P1 f . To proceed further we need a slightly stronger
way of describing the norm of the Kj g ([15], p. 115). Observe first that
convolution
Z
(τ ∗ f )(x) = τ (v)f (x − v) dv (f ∈ L1 (Rn , RM ))
Rn−1
by any real-valued function τ in L1 (Rn−1 ) can be defined componentwise;

furthermore, τ ∗ f belongs to L1 (Rn , RM ) and
Z Z
P1 : τ ∗ f −→ τ (v)f (y + t1 − v) dv dy = τ̂ (0)(P1 f )(t).
Rn−1 Rn−1
In particular, P1 (τ ∗ f ) = P1 f whenever τ̂ (0) = 1.
Lemma 5.5. For each f = (f1 , . . . , fM ) in L1 (Rn , RM )
Z Z Z
kP1 f (t)kRM dt = inf |(τ ∗ f1 )(x)|dx + · · · + |(τ ∗ fM )(x)|dx ,
R τ Rn Rn
the infimum being taken over all τ ∈ L1 (Rn−1 ), τ̂ (0) = 1.

Proof of Lemma 5.5. Obviously
Z Z Z
kP1 f (t)kRM dt ≤ |(τ ∗ f1 )(x)|dx + · · · + |(τ ∗ fM )(x)|dx.
R Rn Rn
Conversely, suppose the fj are real-valued continuous functions all of which
have support in a compact subset, say K, of Rn ; in particular, kfj kL∞ ≤ N ,
for some N > 0. Then to each > 0 there corresponds a function τ ∈
L1 (Rn−1 ), τ̂ (0) = 1, such that
Z Z Z
|(τ ∗ f1 )(x)|dx + · · · + |(τ ∗ fM )(x)|dx ≤ kP1 f (t)kRM dt + ,
Rn Rn R
([15], pp. 113, 115). This proves the theorem since such functions are dense
in L1 (Rn , RM ) as K and N are allowed to vary.
Proof of necessity of the maximal rank condition can now be completed.
Since K commutes with translation, hence with convolution by τ , (5.4) to-
gether with (5.5) ensures that
Z
kP1 f kH 1 ≤ inf kτ ∗ (Kf )kL1 = kK1 f (t)kRM dt,
τ R
establishing (5.3). This inequality will be applied to a copy of H 1 (R) in
H 1 (R, Λr (Cn−1 )). For any a, b ∈ Λr (Cn−1 ),
h −→ hab = (a − b)h − i(a + b)h̃ = a(h − ih̃) − b(h + ih̃)
embeds H 1 (R) in H 1 (R, Λr (Cn−1 )) so that
khab kH 1 = k(a − b)h − i(a + b)h̃kH 1 ∼
= (kak + kbk) khkH 1 .
Now apply K1 to such functions:
K1 hab = (m(1)a − m(−1)b)h − i(m(1)a + m(−1)b)h̃.
But if m(1)a + m(−1)b = 0, then
(kak + kbk) khkH 1 ≤ const. (km(1)a − m(−1)bkCM ) khkL1
can occur for all f ∈ H 1 (R) only if a = b = 0. Thus m(1) ⊕ m(−1) must
have maximal rank on Tr (1) ⊕ Tr (−1). After rotation the proof is complete.
This proves that if K characterizes Hρ1r then the rank condition holds.
Proof of sufficiency. We begin with a Λr (Cn )-valued version of Uchiyama’s
characterization of H 1 .
Lemma 5.6 ([20]). Let K̃ be a standard singular integral operator hav-
ing symbol m̃(ω) mapping Λr (Cn )-valued functions to CM -valued functions.
Then K̃ characterizes H 1 (Rn , Λr (Cn )) if and only if for each ω ∈ Σn−1 ,

n
rank(m̃(ω) ⊕ m̃(−ω)) = 2 .
r
Now suppose that the symbol of K satisfies rank(m(ω)⊕m(−ω)) = 2 n−1

r
as a mapping (a, b) → m(ω)a + m(−ω)b from Tr (ω) ⊕ Tr (−ω) into CM . Let
Nr be the system of singular integrals in Example (2.5 0
) with
symbol n,
m(ω)
and consider the augmented system K̃ having symbol . Clearly m̃
n(ω)
satisfies rank(m̃(ω) ⊕ m̃(−ω)) = 2 nr as a mapping on CM ⊕ CM . By (5.6),

K̃ then characterizes H 1 (Rn , Λr (CM )). Since any f ∈ Hρ1r lies in the kernel
of Nr it follows that kf kHρ1r ∼ kK̃f kL1 = kKf kL1 . Therefore K characterizes
Hρ1r . This completes the proof of Theorem A.
6. Pseudo-inverses.
Pseudoinverses play a fundamental role in Uchiyama’s constructive decom-
position. In our setting these are systems of real singular integrals Lν =
(Lν1 , . . . , LνM ),
Z D E
(6.1) Lνj f (x) = e2πix.ξ fˆ(ξ), θνj (ξ) dξ
Rn
indexed by elements ν = (ν1 , . . . , νM ) in the unit sphere ΣM −1 of RM and

satisfying standard estimates independent of ν; here θν1 , . . . , θνM are Λr (C)-
valued functions on Σn−1 extended to homogeneous functions of degree zero
on Rn \ {0}. We shall follow the same path, always mindful of the fact that
the mapping m(ω)⊕m(−ω) need not be 1-1 everywhere on Λr (Cn )⊕Λr (Cn ).
Theorem D. If the mapping m(ω) ⊕ m(−ω) associated with the symbol of
K has the property that for each ω ∈ Σn−1

n−1
rank(m(ω) ⊕ m(−ω)) = 2
r
as a mapping from Tr (ω) ⊕ Tr (−ω) into CM , then there exists a system of
real singular integrals Lν = (Lν1 , . . . , LνM ), ν ∈ ΣM −1 , depending smoothly
on ν and such that

X
νj Lνj = 0, K∗ ◦ Lν = I
j
on Hρ1r (Rn ) for each ν.

Smooth dependence i.e., C ∞ -dependence, ensures that the standard esti-
mates for Lν are uniform in ν.
Clearly the system Lν acts as a ‘partial’ inverse to K in the sense that
it inverts K on Hρ1r (Rn ), but not necessarily on all of H 1 (Rn , Λr ), just as
every mapping A between finite dimensional spaces admits a partial inverse
- the Moore-Penrose pseudo-inverse A+ of A, for instance ([14]). Because
A+ inverts A on the subspace orthogonal to the kernel of A, the maximal
rank condition on K will allow us to define Lν via its symbol by applying
the Moore-Penrose pseudo-inverse construction at each point ω of Σn−1 . As
we have to ensure that everything can be done smoothly both in ω and ν,
it may be helpful to recall some basic ideas underlying the construction of
A+ from A. First Penrose shows that equations
(6.2)(i) XX ∗ A∗ = X, XAA∗ = A∗
have a unique solution X which he takes as the ‘generalized inverse’ A+ of
A; of course, if A has an inverse A−1 in the usual sense, then A−1 obviously
satisfies (6.2)(i), so A+ = A−1 when the latter is well-defined. He then
shows that BA∗ is a solution of (6.2)(i) provided
(6.2)(ii) BA∗ AA∗ = A∗
and obtains one such B using the Cayley-Hamilton theorem. It is this last
construction that we need to show can be done smoothly, though it should be
observed that Penrose himself remarked that it could be done continuously
([14], p. 408).
Theorem 6.3. Let Aα be a family of m × n complex matrices varying
smoothly with α on some open set A. Then the Moore-Penrose inverse
A+α of Aα varies smoothly with α on any open subset of A on which the rank
of Aα remains constant.
Proof. Let
fα (λ) = det A∗α Aα − λI = a0 (α) + a1 (α)λ + · · · + an (α)λn

be the characteristic polynomial of A∗α Aα . The coefficients of fα will be

smooth on A and A∗ A will be a solution
of the polynomial equation fα (λ) =
∗

0. On the other hand, since rank Aα = rank Aα Aα ,
fα (λ) = λn−r (an−r (α) + · · · + an (α)λr ) , (r = rank Aα ).
Thus
a0 (α) = a1 (α) = · · · = an−r−1 (α) = 0, an−r (α) 6= 0
on any open subset of A on which Aα has constant rank r, while

∗ n−r ∗ r
(6.4) (Aα Aα ) an−r (α)I + · · · + an (α)(Aα Aα ) = 0
on this same subset. Now define Bα by

1
an−r+1 (α)I + · · · + an (α)(A∗α Aα )r−1 .

Bα = −
an−r (α)
Property (6.4) ensures that
Bα (A∗α Aα )n−r+1 = (A∗α Aα )n−r ,
which after succesive cancellation by Aα and A∗α on the right reduces to
Bα A∗α Aα A∗α = A∗α
([14], p. 406). This completes the proof.
We can now begin the proof of Theorem D. For each ν in ΣM −1 of RM
and ω in Σn−1 define µν : C ⊕ Λr (Cn ) → CM by
(6.7) µν (ω) : (λ, a) −→ µν (ω)(λ, a) = λν + m(ω)πT (ω)a
where πT (ω) is the orthogonal projection from Λr (Cn ) onto Tr (ω). The first
result brings out very clearly how the algebraic condition on the symbol of K
and the fact that K is real permit use of the Moore-Penrose pseudo-inverse
µ+
ν (ω) of µν (ω).
Theorem 6.8. If the symbol of K satisfies the maximal rank condition of

Theorem A, then

n−1
ker (µν (ω)) = {(0, b) : b ∈ Nr (ω) }, rank (µν (ω)) = 1 +
r
for each (ν, ω) ∈ ΣM −1 ×Σn−1 . In particular, µ+
ν (ω) is smooth as a function
on ΣM −1 × Σn−1 and µ+ ν (ω)µν (ω) = I on C ⊕ T r (ω).
Proof. Obviously,
ker (µν (ω)) ⊇ {(0, b) : b ∈ Nr (ω) }.
The reverse inclusion will follow from the restrictions on the symbol of K.
Suppose (λ, a) lies in the kernel of µν (ω) and let a = a0 + a1 be the decom-
position of an element of Λr (Cn ) into its respective normal and tangential
components a0 , a1 at ω. Then
µν (ω)(λ, a) = λν + m(ω)a1 = 0,
and so
m(ω)a1 = 0 (λ = 0), m(ω)(a1 /λ) = −ν (λ 6= 0).
But the maximal rank condition ensures that m(ω) is 1-1 on Tr (ω); conse-
quently, a1 = 0 if λ = 0. If λ 6= 0 on the other hand, set b = a1 /λ. Then
b ∈ Tr (−ω) and, because K is real,
m(ω)(b) = m(−ω)(b) ∈ ΣM −1 ;
in this case, m(ω)b + m(−ω)(−b) = 0, contradicting the maximal rank
condition. Hence ker (µν (ω)) ⊆ {(0, b) : b ∈ Nr (ω) }. The remainder of
the
n−1
theorem now follows immediately from (6.3) since dimC (Tr (ω)) = r .
In view of (6.8), therefore, there exist
γ = (γ1 , . . . , γM ), θ(ν, ω) = (θν1 (ω), . . . , θνM (ω))
with γ ∈ CM and θ = θ(ν, ω) a C ∞ -function ΣM −1 × Σn−1 → CM ⊗ Λr (Cn )
such that
 
X
(6.9) µ+
ν (ω) : β = (β1 , . . . , βM ) −→
hβ, γi , βj θνj (ω) .
j
Thus on C ⊕ Tr (ω),

µ+
ν (ω)µν (ω) : (λ, a) −→ λ hν, γi + hm(ω)a, γi ,
X X
λ νj θνj (ω) + ha, mj (ω) i θνj (ω) = (λ, a).
j j
But this can happen only if

X X
(6.10) νj θνj (ω) ∈ Nr (ω), ha, mj (ω) i θνj (ω) = a
j j
for all a ∈ Tr (ω). These properties of (θν1 (ω), . . . , θνM (ω)) ensure that the
singular integrals defined by (6.1) have the properties required of the system
Lν = (Lν1 , . . . , LνM ) in Theorem D, completing the proof of that theorem.
7. Divergence-free r-form wavelets.

In this section we present the construction of the r-form wavelets described
in Theorem C (cf. [11]). The first step will be the proof of the following
result.
Lemma 7.1. For each r, 1 ≤ r ≤ n, there is a bi-orthogonal family
{(γ ,α , η ,α ) : ∈ E ∗ , α ∈ Pr } of compactly supported r-form wavelets so
that each f in L2 (Rn , Λr ) admits a wavelet expansion
,α ,α
X X
(7.2)(i) f (x) = f, ηQ γQ (x)

α∈Pr ,Q
which commutes with interior differentiation in the sense that

 
∂  X ,α ,α
µ∗k

f, ηQ γQ (x)
∂xk
Q∈Qm
,α\k ,α\k
X
(7.2)(ii) = µ∗k Dk f, ηQ γQ (x) (k ∈ α)

Q∈Qm
holds for each , α, k and m where α \ k denotes the (r − 1)-tuple obtained

by omitting k from the r-tuple α.
Note that the sum in (7.2)(i) is taken over the set Q of all dyadic cubes,
while the sums in (7.2)(ii) are taken over the set Qm of dyadic cubes having
side-length l(Q) = 2−m . The divergence-free wavelets ψ,α will arise as
suitable linear combinations of the γ,α - this will be the second step in the
construction.
Let φ = φ(s), s ∈ R, be a compactly supported real-valued scaling func-
tion on R and ψ = ψ(s) the associated wavelet. The usual translates and
dilates of ψ produce an orthonormal basis for L2 (R), while tensor prod-
ucts of {φ, ψ} generate wavelets on Rn ([13], p. 79): For each n-tuple
= (1 , . . . , n ) in E ∗ and x = (x1 , . . . , xn ) in Rn set
n
Y
(7.3) ψ (x) = (j φ(xj ) + (1 − j )ψ(xj )) .
j=1
Then translates and dilates of the family {ψ : ∈ E ∗ } provide an orthonor-

mal basis for L2 (Rn ).
The fundamental idea of Lemarié was to use multi-resolution analysis for
the bi-orthogonal case ([2], p. 128) to derive bi-orthogonal pairs of scaling
functions {φ+ , φ− } and wavelets {ψ+ , ψ− } from the initial choice of (φ, ψ)
by exploiting Viete’s formula
∞
sin πσ Y 1
(7.4) = cos πσ (σ ∈ R),
πσ 2k
k=1
in conjunction with the difference operators
(7.5) ∆+ : f (s) −→ f (s + 1) − f (s), ∆− : f (s) −→ f (s) − f (s − 1)
for functions on R. Property (7.4) ensures that

−πiσ sin πσ
πσ
−πiσ
(7.6) φ̂+ (σ) = e φ̂(σ), φ̂− (σ) = e φ̂(σ)
πσ sin πσ
define new scaling functions {φ+ , φ− } which will have compact support
provided φ is sufficiently smooth, while translates and dilates of the corre-
sponding pair of compactly supported wavelets {(ψ+ , ψ− ) },

2i πiσ
(7.7) ψ̂+ (σ) = ψ̂(σ), ψ̂− (σ) = ψ̂(σ)
πσ 2
provide a bi-orthogonal basis for L2 (R). On the other hand, since
πσ πσ
(Df )ˆ(σ) = e−πiξ (∆+ f )ˆ(σ) = eπiξ (∆− f )ˆ(σ),
sin πσ sin πσ
property (7.6) ensures that
(7.8) Dφ+ = ∆− φ, Dφ = ∆+ φ− , Dψ+ = −4ψ, Dψ = 4ψ− .
Finally, translates and dilates of the family {(ψ,+ , ψ,− ) : ∈ E ∗ } of tensor
products
n
Y
(7.9) ψ,+ (x) = (j ψ+ (xj ) + (1 − j )φ+ (xj )) ,
j=1
Yn
ψ,− (x) = (j ψ− (xj ) + (1 − j )φ− (xj ))
j=1
generate a bi-orthogonal wavelet basis for L2 (Rn ).

The bi-orthogonal wavelets in (7.1), and hence those needed for Theorem
B, will be obtained by the same tensor product construction as before, but in
an important sense to be made precise shortly, the construction will produce
families intermediate to those in (7.3) and (7.9).
Proof of Lemma 7.1. For each ∈ E ∗ and α ∈ Pr set

Y Y
γ,α (x) = (j ψ+ (xj ) + (1 − j )φ+ (xj )) (j ψ(xj ) + (1 − j )φ(xj ))
j∈α j ∈α
/
and
Y Y
η,α (x) = (j ψ− (xj ) + (1 − j )φ− (xj )) (j ψ(xj ) + (1 − j )φ(xj )) .
j∈α j ∈α
/
Notice that γ,α = η,α = ψ when α is the empty set, i.e., r = 0, whereas
γ,α = ψ,+ and η,α = ψ,− when α is the n-tuple (1, 2, . . . , n), i.e., when
r = n. Since each of the latter two families generates a bi-orthogonal basis
for L2 (Rn ) as varies over E ∗ , so the family {(γ,α , η,α ) : ∈ E ∗ } generates
a bi-orthogonal basis for L2 (Rn ) for each fixed α ∈ Pr . To convert these real-
valued functions into r-forms set
(7.10) γ ,α (x) = γ,α (x)eα , η ,α (x) = η,α (x)eα
with {eα : α ∈ Pr } the orthonormal basis for Λr of wedge products (3.2). By

construction therefore, the translates and dilates of {(γ ,α , η ,α ) : ∈ E ∗ , α ∈
Pr } provide a bi-orthogonal basis for L2 (Rn , Λr ), establishing (7.2)(i).
On the other hand, because of (7.8),
∂γ ,α
(7.11)(i) µ∗k = hµ∗k (eα ), eα\k i (1 − k )∆− (k) − 4k γ ,α\k

(k ∈ α)
∂xk
while
∂η ,α\k
= hµk (eα\k ), eα i (1−k )∆+ (k)+4k η ,α

(7.11)(ii) µk (k ∈ α)
∂xk
where
∆+ (k) : f (x) −→ f (x + ek ) − f (x), ∆− (k) : f (x) −→ f (x) − f (x − ek )
are the difference operators ∆± applied now to the xk variable of a function
on Rn . Property (7.2)(ii) is now easily established. When k = 0, for
instance, summation by parts ensures that
 
∂ X D ,α
E
,α 
X D ,α\k
E
,α\k
µ∗k  f, ηQ γQ =− f, µk Dk ηQ γQ ,
∂xk
Q∈Qm Q∈Qm
from which (7.2)(ii) follows after integrating by parts; whereas when k = 1,

∂
,α ,α
D
,α\k
E
,α\k
µ∗k f, ηQ iγQ = − f, µk Dk ηQ γQ
∂xk
and again (7.2)(ii) follows after integrating by parts. This establishes (7.1).

Now we come to the final step in the proof of Theorem C, replacing of
{γ ,α : ∈ E ∗ , α ∈ Pr } by a family {ψ ,α : ∈ E ∗ , α ∈ Tr } of divergence-free
wavelets so that
,α ,α
X X
(7.12) f −→ f, ηQ ψQ (x)
,Q α∈Tr
is the Hodge projection. Recall that Pr = Tr ∪ Nr where

Tr = {α ∈ Pr : i ∈
/ α }, Nr = {β ∈ Pr : i ∈ β }, i = min{k : k = 1 },
and set
(
1 α ∈ Tr ,
(7.13) ψ ,α
(x) = − hµ∗i (ea ), eα id∗ γ ,a
4 a = α ∪ i .
Clearly each ψ ,α is a divergence-free r-form; in addition, at the risk of
forcing ψ ,α , η ,α to have large compact support, we can assume that both
functions have continuous derivatives and vanishing moments up to some
prescribed order - this can be achieved by choosing the original φ to be a
compactly supported scaling function having sufficiently many derivatives
and scaling filter with sufficiently many zeroes ([13], p. 93). On the other
hand, by (7.11) and the choice a = α ∪ i ,
1 X ∂γ ,a
ψ ,α = γ ,α (x) − hµ∗i (ea ), eα i µ∗k
4 ∂xk
k∈α
X
= γ ,α + hµ∗i (ea ), eα i hµ∗k (ea ), ea\k iγ ,a\k
k∈α
since i = 1. Since a \ k ∈ Nr when k ∈ α, {(ψ ,α , η ,α ) : ∈ E ∗ , α ∈ Tr } is

a bi-orthogonal family in the sense that

,α 0 ,α0
ψQ , ηQ0 = δ0 δαα0 δQQ0 (α, α0 ∈ Tr ).
This establishes part (i) of Theorem C, leaving only the proof of the projec-
tion property (ii). Now certainly (7.12) maps L2 (Rn , Λr ) into Hρ2r (Rn ) as
,α
each ψQ (x) is massless. To complete the proof, therefore, we have to show
that each f in Hρ2r (Rn ) can be represented as
,α ,α
X X
(7.14)(i) f (x) = f, ηQ ψQ (x).

,Q α∈Tr
But every f has a wavelet expansion

,β ,β
X X
(7.14)(ii) f (x) = f, ηQ γQ (x),

β∈Pr ,Q
so it is enough to show that the expansions in (7.14) coincide on Hρ2r (Rn ).

Let us assume for the moment that the following technical result has been
established.
Lemma 7.15. For each in E ∗ the identity
 
∂  X X ,β ,β
X X ,α ,α 
µ∗i

f, ηQ γQ − f, ηQ ψQ
∂xi
β∈Pr Q∈Qm α∈Tr Q∈Qm
,δ ,δ
X X
d∗ f, ηQ γQ

=
δ∈Tr−1 Q∈Qm
holds for all f in L2 (Rn , Λr ) and all levels of resolution m.

The expression F in parentheses on the left hand side of (7.15) has the
P P ,δ ,δ ,α
form F = δ∈Nr Q∈Qm cQ γQ since, by the definition of ψQ , the Tr
terms cancel. But δ ∈ Nr implies i ∈ δ so that µ∗i ∂x∂i F = 0 implies F = 0.

The expansions in (7.14) thus coincide on L2 -massless r-forms completing
the proof of Theorem C.
Proof of Lemma 7.15. In view of the definition of Tr and Nr ,

 
∂ X X
,β ,β 
µ∗i

 f, ηQ γQ
∂xi
β∈Pr Q∈Qm
,δ ,δ
X X
= µ∗i Di f, ηQ γQ .
δ∈Tr−1 Q∈Qm
On the other hand, for any α ∈ Tr ,

,α ,α
X
f, ηQ ψQ
Q∈Qm
1 X
,α ∗ ,a
= − hµ∗i (ea ), eα i f, ηQ d γQ
4
Q∈Qm
 
X
,α ,α 1 X X
,α\k ,a\k 
= f, ηQ γQ ± hµ∗i (ea ), eα i  µ∗k Dk f , ηQ γQ
4
Q∈Qm k∈α Q∈Qm
where ‘±’ depends on whether k < i . Consequently,

   
∂ X
,α ,α 
X X ,α\k ,α\k
µ∗i µ∗k Dk f, ηQ

 f, ηQ ψQ =−  γQ  .
∂xi
Q∈Qm k∈α Q∈Qm
The identity now follows, completing the proof.

The usefulness of linking wavelet constructions with the geometry of Eu-
clidean space becomes evident in the proof of Theorem 4.8, the only step
remaining to complete the proof of Theorem A and all ancillary results.
Proof of Theorem 4.8. We prove first that
Z
P1 : f −→ P1 f (t) = f (y + t1) dy
Rn−1
maps f ∈ Hρ1r (Rn ) into H 1 (R, Λr (Rn−1 )). Now each f ∈ Hρ1r (Rn ) has a
wavelet expansion
2mn a,α
X X
,α m
f (x) = k,m ψ (2 x − k) (k ∈ Zn , m ∈ Z)
∈E ∗ ,α∈Tr k,m
where Z
a,α f (y), η ,α (2m y − k) dy

k,m =
Rn
and
 1/2
Z
kf kH 1 ∼  2mn a,α 2 ψ ,α (2m x − k) 2 
X
= k,m Λr
dx.
Rn k,m
But by construction, P1 γ ,α = 0 unless = (0, . . . , 0, 1), while

P1 γ ,α (t) = ψ(t)eα ( = (0, . . . , 0, 1)).

(7.16)
Thus Z
,α
P1 : ψ −→ ψ(t)eα + ( . ) dy
Rn−1
where the integrand consists of partial derivatives of compactly supported
functions. Consequently, (7.16) remains true with γ ,α replaced by ψ ,α (cf.
(7.13)), and so P1 f has a wavelet expansion
XX
(7.17) P1 f (t) = 2m bαl,m ψ(2m t − l)eα ( = (0, . . . , 0, 1))
α∈Tr l,m
with coefficients
a,α
X
bαl,m = K,l,m
K∈Zn−1
and the same restriction on . A simple application of Hölder’s inequality
now shows that
1/2
Z X
2 b ψ(2m t − l)2
m α
l,m dt ≤ const. kf kH 1
R l,m
for each α ∈ Tr . Thus P1 f is a function in H 1 (R, Λr (Rn−1 )) since the wedge

products {eα : α ∈ Tr } form an orthonormal basis for Λr (Rn−1 ); in addition
kP1 f kH 1 ≤ const. kf kH 1 (f ∈ Hρ1r ).
To prove that P1 maps f ∈ Hρ1r (Rn ) onto H 1 (R, Λr (Rn−1 )) we use the
atomic decomposition for H 1 -spaces and show first that P1 maps H 1 (Rn )
onto H 1 (R). It is enough to show that each atom aI having support in
the interval I = [c, d], say, has an atomic pre-image AI 0 . For each x =
(x1 , . . . , xn ) ∈ Rn set
(
1 aI (xn ), c ≤ xj ≤ d,
AI 0 (x) = n−1
|I| 0, otherwise.
Clearly this defines an atom on Rn having support in the cube I 0 = {x ∈ Rn :
c ≤ xj ≤ d } and such that P1 (AI 0 ) = aI . Now fix f ∈ H 1 (R, Λr (Rn−1 )); it
will have a pre-image g ∈ H 1 (Rn , Tr (1)) since Tr (1) = Λr (Rn−1 ). But then
F = (R∗ R)g is a function in Hρ1r (Rn ) such that
P1 F = P1 (R∗ R g) = P1 g = f,
establishing the onto property. This, finally, completes the proof.
Acknowledgement. The authors would like to both referees for several

helpful suggestions.
References
[1] M. Christ and D. Geller, Singular integral characterizations of Hardy spaces on ho-
mogeneous groups, Duke Math. J., 51 (1984), 547-598.
[2] A. Cohen, Biorthogonal wavelets, in ‘Wavelets: A Tutorial in Theory and Applica-
tions,’ C.K. Chui ed., pp. 123-152, Academic Press, San Diego, 1992.
[3] I. Daubechies, Ten Lectures on Wavelets, CBMS-NSF Regional Conference Series,
SIAM, Philadelphia, 1992.
[4] C. Fefferman, Harmonic analysis and H p -spaces, in ‘Studies in Harmonic Analysis,’
J. M. Ash, ed., pp. 38-75, MAA, 1976.
[5] C. Fefferman and E. Stein, H p spaces of several variables, Acta Math., 129 (1972),
137-193.
[6] A. Gandulfo, J. Garcia-Cuerva and M. Taibleson, Conjugate system characterization
of H 1 : counter-examples for the Euclidean plane and local fields, Bull. Amer. Math.
Soc., 82 (1976), 83-85.
[7] J.E. Gilbert, J.A. Hogan and J.D. Lakey, Frame decompositions of form-valued Hardy
spaces, in ‘Clifford Algebras in Analysis and Related Topics,’ CRC Press, Boca Raton,
(1995), 239-259.
[8] , Atomic decomposition of divergence-free Hardy spaces, Math. Moravica, Spe-
cial volume, Proc. IWAA (1997), 33–52.
[9] J.E. Gilbert and M.A.M. Murray, Clifford algebras and Dirac operators in harmonic
analysis, Cambridge University Press, Cambridge, 1991.
[10] S. Janson, Characterization of H 1 by singular integral transforms on martingales and
Rn , Math. Scand., 41 (1977), 140-152.
[11] P.G. Lemarié-Rieusset, Ondelettes vecteurs á divergence nulle, C.R. Acad. Sci. Paris,
313 (1991), 213-216.
[12] A. McIntosh, Clifford algebras, Fourier theory, Singular integrals, and Harmonic func-
tions in Lipschitz domains, in ‘Clifford Algebras in Analysis and Related Topics,’ CRC
Press, Boca Raton, (1995), 33-87.
[13] Y. Meyer, Ondelettes et operateurs, I, II, III, Hermann ed., Paris, 1990.
[14] R. Penrose, A generalized inverse for matrices, Proc. Camb. Phil. Soc., 51 (1955),
406-413.
[15] H. Reiter, Classical Harmonic Analysis and Locally Compact Groups, Oxford Univ.
Press, Oxford, 1968.
[16] E.M. Stein, Harmonic Analysis, Real variable methods, orthogonality, and oscillatory
integrals, Princeton University Press, Princeton, N.J., 1993.
[17] E.M. Stein and G. Weiss, Introduction to Fourier analysis on Euclidean Spaces,
Princeton University Press, Princeton, N.J., 1970.
[18] , Generalization of the Cauchy-Riemann equations and representations of the
rotation group, Amer. J. Math., 90 (1968), 163-196.
[19] A. Uchiyama, A constructive proof of the Fefferman-Stein decomposition of BMO
(Rn ), Acta Math., 148 (1982), 215-241.
[20] , Private communication.
Received December 15, 1996 and revised February 23, 1999. The first and the third
authors were supported in part by NSF grant DMS 93-07655. The second author was
supported by Australian Research Council.
The University of Texas at Austin

Austin TX 78712-1082
E-mail address: gilbert@math.utexas.edu
Macquarie University
NSW 2109
Australia
E-mail address: jeffh@mpce.mq.edu.au
New Mexico State University

Las Cruces, NM 88003-8001
E-mail address: jlakey@nmsu.edu
Vol. 193, No. 1, 2000
TOTAL DETERMINATION OF MATERIAL PARAMETERS

FROM ELECTROMAGNETIC BOUNDARY INFORMATION
Mark S. Joshi and Stephen R. McDowall
In this paper we complete the proof that the material pa-

rameters can be obtained for a chiral electromagnetic body
from the boundary admittance map. We prove that from the
admittance map, the parameters are uniquely determined to
infinite order at the boundary. This removes the assumption
of such knowledge in the result of the second author regarding
interior determination for chiral media.
Introduction.
In this paper we complete the proof that the material parameters can be
obtained for a chiral electromagnetic body from the boundary admittance
map. We achieve this by improving the boundary determination result of
[7]. This enables the removal of the assumption in the interior determination
result of [8] in which it is assumed that the material parameters are known
to infinite order at the boundary.
The behavior of electromagnetic fields in a body is described by Maxwell’s
equations. The electric displacement and the magnetic induction of the
body are related to the fields by the constituent equations which are defined
in terms of a number of material parameters. The parameters typically
considered are the conductivity, the electric permittivity and the magnetic
permeability. A fourth, often neglected, characteristic of an electromagnetic
body is its chirality. Chirality is an asymmetry in the molecular structure:
A molecule is chiral if it cannot be superimposed onto its mirror image, and
the presence of chirality results in a rotation of electromagnetic fields (see
[5]).
In [11], Somersalo et. al. presented the boundary admittance map for
time-harmonic fields at a fixed frequency for non-chiral bodies, and posed
the inverse problem of whether or not the material parameters of a body
could be determined from knowledge of this boundary map. In [9] it was
shown that this is in fact the case assuming knowledge of the parameters
near the boundary. In [8], the second author showed that the admittance
map is well defined also in the case of a chiral body, and that knowledge
of this map determines the material parameters of the body, including the
107
108 MARK S. JOSHI AND STEPHEN R. MCDOWALL
chirality, throughout the body, assuming that the parameters are known to
agree to infinite order at the boundary.
It is therefore pertinent to ask whether or not the admittance map in each
case determines the material parameters at the boundary of the body. In
[7] the second author showed that the admittance map does determine the
parameters and their first normal derivatives and it was conjectured that
the same should be true for all higher order derivatives. It was infeasible,
however, to carry out the computations arrived at in the proof in [7] to prove
determination of the higher order derivatives. In the present paper we prove
that the admittance map does indeed determine the parameters to infinite
order at the boundary in both the chiral and non-chiral cases, thus removing
the aforementioned assumptions in the interior determination results. We
work exclusively with time-harmonic fields at fixed frequency. In order to
pose the problem precisely we shall need the following function spaces: If
Ω is a smoothly bounded open set in R3 , H s (Ω)k consists of k-dimensional
vector fields whose components are in the usual L2 -based Sobolev space H s .
Let Div denote the surface divergence on the boundary of Ω, and ν(x) be
the outward unit normal vector at x ∈ ∂Ω, and define the following space
of tangential fields:
1 n 1 1
o
2
T HDiv (∂Ω) = F ∈ H 2 (∂Ω)3 | ν · F = 0, and DivF ∈ H 2 (∂Ω) .
1
If F ∈ T HDiv
2
(∂Ω), let (E, H) ∈ D0 (Ω)3 × D0 (Ω)3 be the unique solution to
Maxwell’s equations with boundary condition ν ∧ E|∂Ω = F . For Maxwell’s
equations, see Section 1.1 in the non-chiral case, and Section 2.1 in the chiral
case. Solvability of the boundary value problem is given in [11] and [8] for
non-chiral and chiral bodies respectively. The boundary admittance map
1 1
2
Λ : T HDiv (∂Ω) → T HDiv
2
(∂Ω) is then defined by
(0.1) ΛF = ν ∧ H|∂Ω .
We use Λ to denote the non-chiral admittance map, and Π to denote the
chiral admittance map.
For a non-chiral body, ε is a complex parameter with real part the electric
permittivity and imaginary part 1/ω times the conductivity of the body,
and µ is the real-valued magnetic permeability of the body. We assume that
|ε| ≥ ε0 > 0 and µ ≥ µ0 > 0. Suppose that we have two electromagnetic
bodies (Ω; ε, µ) and (Ω; ε0 , µ0 ) with the same boundary ∂Ω. Let Λ and Λ0
be the associated admittance maps. By the statement Λ = Λ0 we mean
1
the following: For every F ∈ T HDiv
2
(∂Ω), if (E, H) and (E 0 , H 0 ) are the
solutions to the boundary value problem above for the bodies (Ω; ε, µ) and
(Ω; ε0 µ0 ) respectively, then
ΛF = ν ∧ H|∂Ω = ν ∧ H 0 |∂Ω = Λ0 F.
TOTAL DETERMINATION OF MATERIAL PARAMETERS 109
We give the analogous meaning to Π = Π0 .

We may now state the results. In Section 1 we prove:
Theorem A. Let (Ω; ε, µ) and (Ω; ε0 , µ0 ) be bounded electromagnetic bodies
in R3 with electromagnetic parameters (ε, µ), (ε0 , µ0 ) ∈ C ∞ (R3 ), and with
smooth boundary ∂Ω. Let Λ and Λ0 be the associated admittance maps, and
assume that Λ = Λ0 . Then on ∂Ω
ε = ε0 , and µ = µ0
and the same is true of all derivatives at the boundary.
In Section 2 we handle the case of a chiral body. The method of proof is
essentially the same as in the non-chiral case, however the underlying ideas
become somewhat obscured by the more complicated setting, and for this
reason we have presented the two proofs separately, the first illustrating the
ideas more accessibly. The reason for the difference between the proofs is
that the non-chiral Maxwell’s equations decouple in a way that the chiral
equations do not. The parameters describing the electromagnetic properties
of a chiral body used here are (as in [8]) a change of variables from the
ones in the Born-Federov formulation (see [5]). The parameters ε and µ
satisfy the same conditions as for the non-chiral case above, and β describes
the chirality, and is a purely imaginary smooth function. We assume that
εµ+β 2 6= 0 which amounts to assuming that the fields never become parallel.
The result of Section 2 is:
Theorem B. Let (Ω; ε, µ, β) and (Ω; ε0 , µ0 , β 0 ) be two bounded electromag-
netic bodies in R3 with chirality described by β and β 0 respectively. Suppose
that (ε, µ, β), (ε0 , µ0 , β 0 ) are in C ∞ (R3 ) and ∂Ω is smooth. If the associated
admittance maps are equal, Π = Π0 , then on ∂Ω
ε = ε0 , µ = µ0 , and β = β0
and the same is true of all derivatives at the boundary.
It was shown in [7], although not stated explicitly, that the admittance
maps are pseudo-differential operators of order one. Our approach is an
inductive one. We will assume that the Taylor series of the parameters
are known to agree to some order at the boundary, and then compute the
principal symbol of the difference of the two associated admittance maps.
We show that this principal symbol then determines the next term in the
Taylor series of the parameters. This approach has already been exploited
in various inverse scattering problems, [1], [2], [3] and [4]. Computing the
principal symbol turns out to be computationally much simpler than trying
to compute the lower order terms of the symbol of a single admittance map
as we are able to work invariantly and solely with principal symbols. This is
facilitated by the study of families of classical pseudo-differential operators
Pt ∈ ΨDOcl m parameterized by normal distance from the boundary. These
have the property that the total symbol is of lower order when restricted to
the boundary. We are then able to define an invariant vector of principal
dl
symbols for this family by taking the principal symbols of l!1 dt Pt |t=0 . We
study and utilize the properties of this vector under composition with normal
differentiation at the boundary and with pseudo-differential operators.
In [8] it was shown that for a chiral body the admittance map uniquely
determines the material parameters throughout the body under the assump-
tion that they are known to infinite order at the boundary. We thus have
the following corollary:
Corollary C. With (Ω; ε, µ, β) and (Ω; ε0 , µ0 , β 0 ) as in Theorem B, if Π =
Π0 then
ε = ε0 , µ = µ0 , and β = β0
throughout Ω.
We thus conclude that, in principle, the material parameters of a chiral
electromagnetic body are recoverable from information obtained only at the
boundary of the body, namely knowledge of the boundary admittance map.
1. The Non-Chiral Case.

1.1. The equations. We assume that we are working in a compact convex
body, Ω, in R3 and that we have picked geodesic normal coordinates local
to a point on the boundary so that ∂Ω is x3 = 0 and the Euclidean metric
becomes
X
dx23 + gij (x)dxi dxj .
i,j<3
For the moment, we make no constraints on gij but shall do so later.

As we have a metric, there is a natural pairing and therefore an iso-
morphism between one-forms and vector fields - we shall work exclusively
with one-forms. For time-harmonic fields at fixed frequency ω, Maxwell’s
equations then take the form
∗dH = −iωεE
(1.1)
∗dE = iωµH
and
δ(εE) = 0, δ(µH) = 0.
Here ∗ is the Hodge star operator, d is exterior differentiation and δ = ∗d∗
is its adjoint. We have

1
∗d ∗ dE = ω 2 εE
µ
and thus

1
µ∗d ∧ ∗dE + ∗d ∗ dE = ω 2 µεE.
µ
Or
− ∗ d(log µ) ∧ ∗dE + ∗d ∗ dE = ω 2 µεE.
Now on one-forms on a three-manifold, ∆ = − ∗ d ∗ d + d ∗ d∗ so
−∆E + d ∗ d ∗ E − ∗(d log µ ∧ ∗dE) − ω 2 µεE = 0.
We also know
δ(E) = −E · d log ε.
(Here we are taking the inner product on one-forms induced by the metric
- this is an invariant statement and holds in the flat coordinates.) So
−∆E − d(E · d log ε) − ∗(d log µ ∧ ∗dE) − ω 2 µεE = 0.
This is of the form
h 0
i
ME = Dx23 I − ∆x3 I − M (x, Dx0 ) − iP (x)Dx3 − R(x) E = 0,
where as in [7] M consists of all terms involving first order differentiation in
x1 and x2 , P is the coefficient matrix of ∂/∂x3 , R consists of all zero order
0
terms and ∆x3 is the Laplacian in x1 , x2 for the value of x3 .
1.2. Factorization.
Proposition 1.1. There exists B(x, Dx0 ) ∈ ΨDO1 with principal symbol
−|ξ 0 |x depending smoothly on x3 such that
M = (Dx3 I − iP (x) − iB)(Dx3 I + iB)
up to smoothing and B is unique up to smoothing.
Proof. As in [7]
∆0 + i[Dx3 , B] + M + P B + B 2 + R = 0.
The principal symbol of B is therefore ±|ξ 0 |x ; we take −|ξ 0 |x . Now suppose
we have chosen Bj ∈ ΨDO1−j such that
GN = B0 + B1 + · · · + BN
satisfies
∆0 + i[Dx3 , GN ] + M + P GN + G2N + R = EN ∈ ΨDO1−N .
Then we must have
2 −N
i[Dx3 , BN +1 ] + P BN +1 + BN +1 + GN BN +1 + BN +1 GN − EN ∈ ΨDO .
The principal symbol of this is just
−2|ξ 0 |x σ−N (BN +1 ) − σ1−N (EN ).
So picking the symbol of BN +1 appropriately this is zero away from ξ 0 = 0

and we can construct all the BN ; we let B be the asymptotic sum. By
construction, B is unique modulo a smoothing operator.
1.3. A class of pseudo-differential operators. Let Y be a smooth man-

ifold. We study families of operators on Y smoothly varying with a param-
eter t. (We will take Y to be ∂Ω and t the normal distance.)
Definition 1.2. We shall say P ∈ ΨDOm,p (Y ; R+ ) if it is a family of
pseudo-differential operators of order m on Y , varying smoothly up to t = 0,
and such that
X p
P = tp−j Pj
j=0
with Pj a smooth family of operators on Y of order m − j. We will allow

operators on bundles also.
Definition 1.3. If P ∈ ΨDOm,p (Y ; R+ ) then the symbol of P at t = 0 is
the vector
(σm−j (Pj ))pj=0
evaluated at t = 0. This makes invariant sense as a vector of functions on
the cotangent bundle of Y.
The following is immediate.
Proposition 1.4. If B is a smooth family of operators on Y of order k and
P ∈ ΨDOm,p (Y ; R+ ) then BP, P B ∈ ΨDOm+k,p (Y ; R+ ) and the symbols of
P B, BP at t = 0 are just the product of the symbols. We also have that
[Dt , P ] ∈ ΨDOm,p−1
p−1
and that the principal symbol at the boundary is −i((p − j)σm−j (Pj ))j=0 .
1.4. The symbol of the difference. Now suppose we have two different
sets of parameters, (ε, µ) and (ε0 , µ0 ), with associated admittance maps Λ and
Λ0 ; let B and B 0 be the associated factorization operators from Proposition
1.1. We shall consistently use 0 to denote operators associated to ε0 , µ0 .
Suppose that Λ = Λ0 ; from [7] we know that on ∂Ω ε = ε0 , µ = µ0 and the
same is true of the first normal derivatives. We shall prove inductively that
if ε and ε0 are known to agree to order l ≥ 2 at the boundary, and similarly
for µ and µ0 , then the fact that Λ = Λ0 implies that the parameters agree to
order l + 1. To this end we write
(1.2) ε/ε0 = 1 + xl3 eε , µ/µ0 = 1 + xl3 eµ ,
with eε , eµ smooth up to the boundary. We will calculate eε and eµ at the
boundary from the principal symbol of the difference Λ0 − Λ which we write
in terms of the difference B 0 −B. We fix a point p on the boundary and work
in coordinates which are geodesic normal coordinates at p in the boundary
and extended normally off the boundary.
Proposition 1.5. Suppose (ε, µ) and (ε0 , µ0 ) are equal to order l ≥ 2 at the
boundary. Writing B 0 = B + F we have F ∈ ΨDO0,l−1 , and if (fj )j=0,−(l−1)
is the principal symbol of F at the boundary, then in our chosen local bound-
ary normal coordinates
    0 
0 0 iξ1 |ξ |x 0 iξ1
l!  
(1.3) f−(l−1) = l 0 l eε 0 0 iξ2  + eµ  0 |ξ 0 |x iξ2  + r0 ,
2 |ξ |x 0 0 |ξ | 0 0 0 0
x
where r0 vanishes to second order at x1 , x2 = 0.

Proof. With B 0 = B + F we have
∆0 + i[Dx3 , B] + M + P B + B 2 + R = 0, and
∆0 + i[Dx3 , B + F ] + M 0 + P 0 (B + F ) + (B + F )2 + R0 = 0.
Subtracting yields,
(1.4)
i[Dx3 , F ] + P 0 F + F 2 + F B + BF = (M 0 − M ) + (R0 − R) + (P 0 − P )B.
So if we can construct F satisfying (1.4), then B 0 = B + F as we know B 0
is unique (modulo smoothing).
The terms on the right hand side of (1.4) result from −ω 2 εµE,
d(E · d log ε) and ∗(d log µ ∧ ∗dE) (see Section 1.1). Specifically, the terms
involving differentiation in the tangential variables make up (M 0 − M ), and
those involving no differentiation make up (R0 − R); the coefficients of differ-
entiation with respect to the normal variable x3 are what comprise (P 0 − P ),
which then multiplies the first order pseudo-differential operator B. We com-
pute the contribution of these terms to the right hand side of (1.4) modulo
ΨDO1,l . Now
ω 2 εµ − ω 2 ε0 µ0 = ω 2 ε0 µ0 (xl3 (eε + eµ ) + x2l 0,l
3 eε eµ ) ∈ ΨDO .
We also have,
ε
d(E · d log ε0 ) − d(E · d log ε) = −d E · d log 0
ε
= −d(E · d log(1 + xl3 eε )),
and
lxl−1
3 eε dx3 xl3 deε
d[log(1 + xl3 eε )] = + .
1 + xl3 eε 1 + xl3 eε
Thus
lxl−1
3 eε E3 xl3 E · deε
E · d log(1 + xl3 eε ) = + ,
1 + xl3 eε 1 + xl3 eε
and so
l(l − 1)xl−2
3 eε E3 lxl−1
3 E3
d[E · d log(1 + xl3 eε )] = l
dx3 + deε
1 + x3 eε 1 + xl3 eε
lxl−1
3 eε lxl−1 l−1 2
3 (E · deε − x3 eε E3 )
+ dE 3 + dx3
1 + xl3 eε (1 + xl3 eε )2
−1
x2l
3 (lx3 eε E3 + E · deε ) xl3
− l
de ε + d(E · deε )
(1 + x3 eε )2 1 + xl3 eε
= T1 + T2 + T3 + T4 + T5 + T6 , say.
We compute the contribution of d(E · d log(1 + xl3 eε )) to the right hand side
of (1.4) modulo ΨDO1,l ; this means we can ignore operators not involving
Dx3 if they are in ΨDO1,l , and so we can drop the terms T2 , T4 , T5 . The
coefficients of operators in Dx3 are what make up P 0 − P which, in the right
hand side of (1.4), is multiplied by B. The resulting operator (P 0 − P )B is
in ΨDO1,l if the coefficients vanish to order l. Thus we may drop T6 .
We are left with
1
l−2 l−1

d(E · d log(1 + xl3 eε )) = l(l − 1)x3 eε E 3 dx3 + lx3 eε dE3 + T
1 + xl3 eε
where the contribution of T in (1.1) results in operators in ΨDO1,l . The

dE3 will have two parts - the coefficient of Dx3 is absorbed into P 0 − P and
then multiplied by B. Thus modulo ΨDO1,l the symbolic contribution of
−d(E · d log(1 + xl3 eε )) to the right hand side is
   
l−2 0 0 0 l−1 0 0 iξ1
−l(l − 1)x3 eε  lx3 eε 
l
0 0 0 − le
0 0 iξ2  .
1 + x3 eε 0 0 1 1 + x3 ε 0 0 −|ξ 0 | x
This leaves the term ∗(d log(µ) ∧ ∗dE). When we take the difference we get
− ∗ (d log(1 + xl3 eµ ) ∧ ∗dE). The only term from this which does not result
in something in the right hand side of (1.1) absorbable into ΨDO1,l is
−lxl−1
3 eµ
∗ (dx3 ∧ ∗dE).
1 + xl3 eµ
Now the Hodge star operator in these coordinates is equal to the flat star
operator, ∗f , in these coordinates plus an error. We now assume that we
have chosen coordinates normal about some point in the boundary and then
extended normally away from the boundary. Then

 
X
(1.5) ∗ = I + x3 D + xi xj ∗ Dij  ∗f ,
i,j<3
with D, Dij smooth homomorphisms of the form bundles. The x3 D will give
us an element of ΨDO1,l which is therefore ignorable. We first compute in
flat coordinates the value of ∗f (dx3 ∧ ∗f dE); this is equal to
∂E1 ∂E3 ∂E2 ∂E3
− dx1 + dx1 − dx2 + dx2 .
∂x3 ∂x1 ∂x3 ∂x2
As before the differentiation in x3 becomes part of P 0 − P and we therefore
have to drop the ξ3 in the symbol and replace it by the principal symbol of
B which is −|ξ 0 |x , so we conclude that the contribution to the forcing on the
right hand side is
   
0 0 iξ1 1 0 0
−lx3l−1 eµ 
0 0 iξ2  + |ξ 0 |x 0 1 0 + σ(G1 ) + σ(G2 ).
1 + xl3 eµ 0 0 0 0 0 0
Here G1 ∈ ΨDO1,l and G2 = xi xj Gij with Gij ∈ ΨDO1,l−1 .
P
j,j<3
So we have that i[Dx3 , F ]+P 0 F +F 2 +F B +BF = G with G ∈ ΨDO1,l−1
from (1.4). We first show that this means F ∈ ΨDO0,l−1 and then compute
the symbol of F at our chosen point.
We know F is a pseudo-differential operator of order zero as B, B 0 have
the same principal symbol. The principal symbol of the left hand side is
−2|ξ 0 |x σ0 (F ) and so we pick F0 to have principal symbol −1/(2|ξ 0 |x )σ1 (G)
which vanishes to order l − 1 at x3 = 0, so we can certainly take F0 ∈
ΨDO0,l−1 . Putting F1 = F − F0 we then obtain a similar equation but
with right hand side in ΨDO0,l−2 and thus solve to get F1 ∈ ΨDO−1,l−2 .
Repeating we get F0 + F1 + · · · + Fl−1 ∈ ΨDO0,l−1 solving up to an error
in ΨDO−l . This can then be removed in the same way B was originally
constructed so we conclude that F = B − B 0 can be constructed to be in
ΨDO0,l−1 and therefore by uniqueness is actually in ΨDO0,l−1 .
Now recall that the principal symbol of F at x3 = 0 is a well-defined
object and is a vector of matrices. Let (fj )j=0,−(l−1) be this vector. Then
F B + BF has symbol (−2|ξ 0 |x fj ) and i[Dx3 , F ] ∈ ΨDO0,l−2 has symbol
−(l−2)
((l − 1 + j)fj )j=0 . So from (1.4), we obtain that
− 2|ξ 0 |x f0 =
     
0 0 iξ1 1 0 0 0 0 iξ1
− leµ 0 0 iξ2  + |ξ 0 |x 0 1 0 − leε 0 0 iξ2  + r
0 0 0 0 0 0 0 0 −|ξ 0 |x
where r vanishes to second order at x1 , x2 = 0. We also have

 
0 0 0
−2|ξ 0 |x f−1 + (l − 1)f0 = −l(l − 1)eε 0 0 0 ,
0 0 1
and
−2|ξ 0 |x f−j + (l − j)f1−j = 0.
We can thus iteratively determine f−(l−1) , the principal symbol of F re-
stricted to x3 = 0, which is what we want to know. Explicitly,
    0 
0 0 iξ1 |ξ |x 0 iξ1
l!  
f−(l−1) = l 0 l eε 0 0 iξ2  + eµ  0 |ξ 0 |x iξ2  + r0 ,
2 |ξ |x 0
0 0 |ξ | 0 0 0
x
where r0 vanishes to second order at x1 , x2 = 0.
1.5. Recovering the coefficients. In [7] it was shown that the param-
eters and their first derivatives are determined on the boundary from the
admittance map. As in the previous section we assume that (ε, µ) and
(ε0 , µ0 ) are known to agree to order l ≥ 2 on the boundary. In this section
we show that the difference of the admittance maps, Λ0 − Λ, determines eε
and eµ (see (1.2)). We fix a point p in the boundary of the domain, take
normal coordinates in the boundary about it and then extend these off the
boundary normally.
Before proceeding to the proof, we present a useful lemma.
Lemma 1.6. Suppose that P1 and P2 are pseudo-differential operators of
the form
Pj = Aj + f (x0 )Rj
where Aj , Rj are mth -order pseudo-differential operators, j = 1, 2, and
f (p) = 0. Then we can recover the principal symbol of A1 − A2 at p (even
if it is of lower order than P1 − P2 ).
This follows from the fact that in any coordinate system the total left
symbol of f (x0 )(R1 − R2 ) will vanish at p. The usefulness lies in our ability
to discard terms in otherwise complicated computations.
From the definition of Λ (0.1), Maxwell’s equations (1.1), and the expan-
sion for the Hodge star operator (1.5), the admittance map at the point p
is

E2 1 ∂3 E1 − ∂1 E3
7→
−E1 x =0 iωµ ∂3 E2 − ∂2 E3 x =0
3 3
plus terms which have coefficients vanishing at x1 = x2 = 0; we shall be

considering the difference Λ0 − Λ and so from the above lemma, the symbol
of Λ0 − Λ at p does not involve these terms.
We wish to express the right hand side in terms of E1 and E2 and operators
solely on the boundary. We know that δ(εE) = 0 and hence that d∗(εE) = 0.
As above,
 
X
∗ = Id +x3 D + xi xj Dij  ∗f
i,j
with D, Dij variable matrices (all sums are taken over i, j ≤ 2). So we have
that
  
X
dε ∧ Id +x3 D + xi xj Dij  ∗f E 
ij
  
X
+ εd Id +x3 D + xi xj Dij  ∗f E  = 0.
ij
Evaluating at x3 = 0 this becomes,

X
(1.6) dε ∧ (∗f E) + xi xj dε ∧ (Dij ∗f E) + ε(d ∗f E)
ij
X
+ ε(dx3 ∧ (D ∗f E)) + ε d(xi xj Dij ∗f E) = 0.
ij
Proposition 1.7. If (U, x) is a local geodesic normal coordinate patch to

the boundary of Ω near p, then
3
∂Ej X
= Bjl El , j = 1, 2, 3
∂x3
l=1
where Bjl are the components of the matrix of operators B.
The proof of this is as in [7] Proposition 2, and [6] Proposition 1.2. So
dividing by ε and rearranging, we conclude that,
   
X X
(1.7) Id + xi xj Dij  (B33 + ∂3 log ε)E3 + D + xi xj ∂3 Dij  E3
i,j i,j
 
X
= − Id + xi xj Dij  (∂1 log ε)E1 + (∂2 log ε)E2
i,j

+ ∂1 E1 + ∂2 E2 + B31 E1 + B32 E2
   
X X
− ∂1  xi xj Dij  E1 − ∂2  xi xj Dij  E2 = T, say.
i,j i,j
Let J be the coefficient of E3 , and let K be a parametrix for J which is

certainly elliptic at x1 , x2 = 0. We also have that the principal symbol of K
is just −|ξ 0 |−1
x Id.
So E3 = KT and
∂3 E1 − ∂1 E3 = B11 E1 + B12 E2 + (B13 − ∂1 )KT.
Extracting the components acting on E2 , we have
iωµΛ11
 
X
= B12 − (B13 − ∂1 )K Id + xi xj Dij  (∂2 log ε + ∂2 + B32 )
i,j
 
X
+∂2  xi xj Dij 
i,j
plus terms vanishing at x1 = x2 = 0.

Now J, and hence K, depends on the derivatives of ε and µ. However,
the dependence in J − J 0 comes only from B33 − B330 for
 
X h ε i
J − J 0 = Id + 0
xi xj Dij  B33 − B33 + ∂3 log 0
ε
i,j
and as we are working at x3 = 0, ε = ε0 . Furthermore, at x3 = 0, B33 and

0 can differ by an operator of order at most −(l − 1). If two first order
B33
operators are equal up to order −(l − 1) then using
a−1 − b−1 = a−1 (b − a)b−1
we conclude that their parametrices agree to order −(l + 1). So we have
that K 0 − K is of order −(l + 1), where K 0 is the analogue of K for ε0 , µ0 ,
and its principal symbol at p will be −|ξ 0 |−2
x σ−(l−1) (F33 ).
And so
iωµ(Λ011 − Λ11 ) = (B12 + F12 )
 
X
− (B13 + F13 − ∂1 )K 0 Id + xi xj Dij  (∂2 log ε0 + ∂2 + B32 + F32 )
i,j
 
X
+∂2  xi xj Dij  − B12
i,j
 
X
+ (B13 − ∂1 )K Id + xi xj Dij  (∂2 log ε + ∂2 + B32 )
i,j
 
X
+∂2  xi xj Dij 
i,j
 
X
= F12 − (B13 − ∂1 )K 0 Id + xi xj Dij  F32 − (B13 − ∂1 )(K 0 − K)
i,j
   
X X
· Id + xi xj Dij  (∂2 + B32 ) + ∂2  xi xj Dij 
i,j i,j
 
X
− F13 K 0 Id + xi xj Dij  (∂2 log ε0 + ∂2 + B32 + F32 )
i,j
 
X
+∂2  xi xj Dij 
i,j
  
X
− (B13 − ∂1 ) (K 0 − K) Id + xi xj Dij  ∂2 log ε0
i,j
  
0
X ε 
+K Id + xi xj Dij  ∂2 log .
ε
i,j
The principal symbol of this at p will come from

F12 + ∂1 K 0 F32 + ∂1 (K 0 − K)∂2 − F13 K 0 (∂2 + F32 ).
Using the results of the previous section, if m = −(l − 1), this has principal
symbol at the point p equal to
σm (F12 ) − iξ1 |ξ 0 |−1 0 −2
x σm (F32 ) + ξ1 ξ2 |ξ |x σm (F33 )
+|ξ 0 |−1
x σm (F13 )(iξ2 + σm (F32 ))
which by (1.3) is equal to
l! −l!
(ξ ξ e − ξ1 ξ2 (eε + eµ )) =
l+1 1 2 ε
ξ1 ξ2 eµ .
l 0
2 |ξ |x 2 |ξ 0 |l+1
l
x
Computing the remaining components of the principal symbol of Λ0 − Λ at
p, we obtain
eµ −ξ1 ξ2 −ξ22

−l!
0= .
2l |ξ 0 |l+1
x iωµ ξ12 ξ1 ξ2
Thus we are able to determine eµ from Λ0 − Λ. Since we are assuming
knowledge of the map Λ, we know also its inverse; restricted to the space
1
2
T HDiv (∂Ω), this is the impedance map which maps the tangential component
of the magnetic field at the boundary to that of the electric field at the
boundary. Interchanging the roles of ε and µ in our analysis, we find that
computing the principal symbol of (Λ0 )−1 − Λ−1 yields
eε −ξ1 ξ2 −ξ22

l!
0=
2l |ξ 0 |l+1
x iωε ξ12 ξ1 ξ2
and we may conclude that eε and eµ are determined at the boundary from
the assumed knowledge of the boundary maps.
We remark the unexpected fact that the inverse of the first order non-
elliptic pseudo-differential operator exists, and that it is also a first order
pseudo-differential operator; of course, the symbol of the inverse is not the in-
1
2
verse of the symbol. This is a result of working on the restricted T HDiv (∂Ω)
spaces rather than on Sobolev spaces - see [10] for further discussion.
2. Chiral Media.
Suppose now that Ω is a body with chirality described by a smooth function
β. We shall apply the techniques of the previous section to show that if two
bodies have the same admittance map, then the chirality together with the
other three material parameters of the bodies must coincide to infinite order
at the boundary.
2.1. The equations for chiral media. Working again with one-forms,
and taking the formulation of [8] which is a change of variables in the Born-
Fedorov formulation, Maxwell’s equations for a chiral body take the form
∗dH = −iω(εE + βH)
(2.1)
∗dE = iω(µH − βE)
together with
(2.2) δ(εE + βH) = δ(µH − βE) = 0.
We write (2.1) as

E β −µ E E
∗d = −iω = −iωX
H ε β H H
where we introduce the notation

β −µ
X= .
ε β
We have

E E
∗d ∗ d = −iω ∗ d X
H H

E 2 2 E
= −iω ∗ dX ∧ −ω X
H H
where we have used (2.1) again. But

E 1 −1 E
=− X ∗d
H iω H
so

E −1 E 2 2 E
∗d ∗ d = ∗ dX ∧ X ∗ d −ω X
H H H
and thus

E E −1 E 2 2 E
−∆ +d∗d∗ − ∗ dX ∧ X ∗ d +ω X = 0.
H H H H
But (2.2) implies that

E −1 E
d∗d∗ = −d X (dX) ·
H H
where the inner product on one-forms is that induced from the metric, and
where the inner product is taken componentwise within the matrix multi-
plication. Combining these,

E −1 E −1 E
−∆ − d X (dX) · − ∗ dX ∧ X ∗ d
H H H

2 2 E
+ω X =0
H
which can be written in the form

E 2 0
E
N = Dx3 I − ∆x3 I − N (x, Dx ) − iQ(x)Dx3 − S(x)
0 =0
H H
where N consists of all terms involving first order differentiation in x1 and
x2 , Q is the coefficient matrix of ∂/∂x3 , and S consists of all zero order
terms.
As in Proposition 1.1, there is C(x, Dx0 ) ∈ ΨDO1 such that
N = (Dx3 I − iQ(x) − iC)(Dx3 I + iC),
with principal symbol −|ξ 0 |x times the identity, and with this choice of prin-
cipal symbol, C is unique modulo smoothing.
2.2. The symbol of C 0 − C. Suppose that we have two bodies (Ω; ε, µ, β)

and (Ω; ε0 , µ0 , β 0 ) for which the admittance maps are identical, and for which
the parameters are equal up to order l at the boundary, with l ≥ 2. As
before we shall use 0 to denote all operators associated with the parameters
ε0 , µ0 , β 0 . We shall prove that equality of the admittance maps implies that
in fact the parameters must agree to order l + 1, and thus inductively prove
that they are equal to infinite order at the boundary. We assume that we
have chosen coordinates normal about some point p in the boundary, and
have extended these normally away from the boundary to a neighborhood

of p. If we write C 0 = C + F then we have
∆0 + i[Dx3 , C] + N + QC + C 2 + S = 0,
∆0 + i[Dx3 , C + F ] + N 0 + Q0 (C + F ) + (C + F )2 + S 0 = 0
and subtracting yields
(2.3)
i[Dx3 , F ] + Q0 F + F 2 + F C + CF = (N 0 − N ) + (S 0 − S) + (Q0 − Q)C.
The terms on the right hand side of this expression come from

0 −1 0 E −1 E
(2.4) d (X ) (dX ) · − d X (dX) · ,
H H

0 0 −1 E −1 E
(2.5) ∗ dX ∧ (X ) ∗ d − ∗ dX ∧ X ∗ d , and
H H

E E
(2.6) ω2X 2 − ω 2 (X 0 )2 .
H H
Specifically, the terms involving differentiation in the tangential variables
make up (N 0 − N ), and those involving no differentiation make up (S 0 −
S); the coefficients of differentiation with respect to the normal variable x3
are what comprise (Q0 − Q), which then multiplies the first order pseudo-
differential operator C. We compute modulo ΨDO1,l and make use of the
following lemma.
Lemma 2.1. (i) If A is a matrix (of 0, 1, or 2-forms),
then terms from
∂ E
(2.4), (2.5) and (2.6) of the form xl3 A ∧ contribute terms in
∂x3 H
ΨDO1,l to the right hand side of (2.3).
(ii) If D is a differential operator
of order one in x1 and x2 , and of order
E
zero in x3 , then xl3 D ∈ ΨDO1,l .
H

l−1 E
(iii) If A is a matrix of functions, then x3 A ∈ ΨDOl−1,0 ⊂ ΨDO1,l .
H
Proof. Claims (ii) and (iii) are immediate. For (i), we must simply observe
that C is a first order pseudo-differential operator in x1 and x2 , and depends
smoothly on x3 as a parameter. Therefore, in the right hand side of (2.3)
we have xl3 AC ∈ ΨDO1,l .
We put

0 eβ −eµ
X −X = xl3 E = xl3
eε eβ
and endeavor to show that E = 0 at the boundary.
Proposition 2.2. Suppose (ε, µ, β) and (ε0 , µ0 , β 0 ) are equal to order l ≥ 2

at the boundary. Writing C 0 = C + F , we have F ∈ ΨDO0,l−1 , and if
(fj )j=0,−(l−1) is the principal symbol of F at the boundary then in our chosen
local boundary normal coordinates,

−l! W11 W12
f−(l−1) = l 0 l + r0 , where
2 |ξ |x (εµ + β 2 ) W21 W22
 0 
|ξ |x (βeβ + εeµ ) 0 iξ1 (2βeβ + εeµ + µeε )
W11 =  0 |ξ 0 |x (βeβ + εeµ ) iξ2 (2βeβ + εeµ + µeε ) ,
0 0 |ξ 0 |x (βeβ + µeε )
 0 
|ξ |x (µeβ − βeµ ) 0 iξ1 (2µeβ − 2βeµ )
W12 =  0 |ξ 0 |x (µeβ − βeµ ) iξ2 (2µeβ − 2βeµ ) ,
0 0 |ξ 0 |x (µeβ − βeµ )
 0 
|ξ |x (βeε − εeβ ) 0 iξ1 (2βeε − 2εeβ )
W21 =  0 |ξ 0 |x (βeε − εeβ ) iξ2 (2βeε − 2εeβ ) , and
0 0 |ξ 0 |x (βeε − εeβ )
 0 
|ξ |x (βeβ + µeε ) 0 iξ1 (2βeβ + εeµ + µeε )
W22 =  0 |ξ 0 |x (βeβ + µeε ) iξ2 (2βeβ + εeµ + µeε )
0 0 |ξ 0 |x (βeβ + εeµ )
and where r0 vanishes to second order at x1 , x2 = 0.
Proof. We first analyze (2.4):

0 −1 0 −1
E
d (X ) dX − X dX ·
H

−1 0 0 −1 0 −1 0
E
= d X d(X − X) + (X ) (X − X )X dX ·
H

E
= d X −1 d(xl3 E) − xl3 (X 0 )−1 EX −1 dX 0 ·
H

l−2 −1 E3 l−1 −1 E3
= l(l − 1)x3 X E dx3 + lx3 d(X E)
H3 H3

−1 dE3 −1 E
+ lxl−1
3 X E + lxl−1
3 X dE · dx3
dH3 H

E E
+ xl3 d(X −1 )dE · + xl3 X −1 d dE ·
H H

l−1 0 −1 −1 0 E l 0 −1 −1 0 E
− lx3 (X ) EX dX · dx3 − x3 d((X ) EX )dX ·
H H

l 0 −1 −1 0 E
− x3 (X ) EX d dX ·
H
9
X
= Ti , say.
i=1
From Lemma 2.1 terms T2 and T4 , . . . , T9 contribute terms in ΨDO1,l to the

right hand side of (2.3). Writing

l−1 −1 ∂1 E3 dx1 + ∂2 E3 dx2 l−1 −1 ∂3 E3 dx3
T3 = lx3 X E + lx3 X E
∂1 H3 dx1 + ∂2 H3 dx2 ∂3 H3 dx3
and recalling that the coefficient of the term involving ∂x3 multiplies C in
the right hand side of (2.3), we have the contribution from (2.4) equal to

l−2 −1 Y1 0 l−1 −1 Z1 0
(2.7) l(l − 1)x3 X E + lx3 X E
0 Y1 0 Z1
where

  
0 0 0 0 0 ∂1
Y1 = 0 0 0 , and Z1 = 0 0 ∂2 
0 0 1 0 0 C33
with C33 the 33-component of C. Next,

0 0 −1 E −1 E
∗ dX ∧ (X ) ∗ d − ∗ dX ∧ X ∗ d
H H

E
= ∗ dX 0 ∧ (X 0 )−1 (X − X 0 )X −1 + X −1 ∗ d

H

E
− ∗ dX ∧ X −1 ∗ d
H

0 0 −1 l −1 E l−1 −1 E
= ∗ dX ∧ (X ) x3 EX ∗ d + ∗ lx3 Edx3 ∧ X ∗ d
H H

E
+ ∗ xl3 dE ∧ X −1 ∗ d .
H
The first and third of these terms contribute elements of 1,l
PΨDO by Lemma
2.1. For the second term we write ∗ = (Id +x3 D + i,j≤2 xi xj Dij )∗f to
obtain

−1 E −1 E
l−1
∗ lx3 Edx3 ∧ X ∗ d l−1
= ∗f lx3 Edx3 ∧ X ∗f d + T + T0
H H
where T results in terms in ΨDO1,l by Lemma 2.1 again, and T 0 vanishes
to second order at x1 , x2 = 0. Computing in terms of the flat star operator,
the contribution from (2.5) is thus
(2.8)
 
−C33 0 ∂1
−1 Z2 0 E
lxl−1
3 EX + T 0 , where Z2 =  0 −C33 ∂2  .
0 Z2 H
0 0 0
The final consideration is the contribution from (2.6). This is

2 2 0 2 E 2
0 0 0
E
ω (X − (X ) ) = ω X(X − X ) − (X − X)X
H H

E
= ω 2 xl3 EX 0 − xl3 XE

H
which is in ΨDO1,l .
As before F ∈ ΨDO0,l−1 , and we now compute the symbol of F at p. Let
the vector (fj )j=0,−(l−1) be the principal symbol of F at x3 = 0. From (2.3),
(2.7) and (2.8) we have

0 −1 σ(Z1 ) 0 −1 σ(Z2 ) 0
−2|ξ |x f0 = lX E + lEX
0 σ(Z1 ) 0 σ(Z2 )
and

0 −1 σ(Y1 ) 0
−2|ξ |x f−1 + (l − 1)f0 = l(l − 1)X E .
0 σ(Y2 )
In general,

−l! −1 σ(Y1 ) 0
f−j = X E
(l − j − 1)!2j |ξ 0 |jx 0 σ(Y1 )
X −1 E σ(Z1 ) EX −1 σ(Z2 )

0 0
+ +
2|ξ 0 |x 0 σ(Z1 ) 2|ξ 0 |x 0 σ(Z2 )
and in particular, computing f−(l−1) we obtain the expression in the state-
ment of the proposition.
2.3. Proof that E = 0. From [7], we know that for a chiral body, the
parameters and their first normal derivatives are determined by the admit-
tance map. Suppose that (ε, µ, β) and (ε0 , µ0 , β 0 ) agree to order l ≥ 2 at the
boundary. We show now that if Π = Π0 , then the parameters must agree
to order l + 1. We fix p in the boundary and work in our chosen boundary
normal coordinates near p.
In order to be able to express Π in terms of the constructed operator C,
we shall need to write E3 and H3 in terms of E1 , E2 , H1 and H2 . From
(2.2) we know that

E E
dX ∧ ∗ + Xd ∗ = 0.
H H
P
Writing ∗ = (Id +x3 D + i,j≤2 xi xj Dij )∗f and computing at x3 = 0, we
have
 
X
(2.9) Id + xi xj Dij 
i,j≤2

∂1 βE1 + ∂2 βE2 + ∂3 βE3 − ∂1 µH1 − ∂2 µH2 − ∂3 µH3
·
∂1 εE1 + ∂2 εE2 + ∂3 εE3 + ∂1 βH1 + ∂2 βH2 + ∂3 βH3
 

X ∂1 E 1 + ∂2 E 2 + ∂3 E 3 E3
+ X Id + xi xj Dij  + DX
∂1 H1 + ∂2 H2 + ∂3 H3 H3
i,j≤2
 

X E
+ Xd  xi xj Dij ∧ ∗f
 = 0.
H
i,j≤2
Making use of the fact that ∂/∂3 |x3 =0 = C, (2.9) can be written

E3 R1
J =
H3 R2
where
 

X ∂ 3 β −∂3 µ C33 C36
J = Id + xi xj Dij  +X
∂3 ε ∂3 β C63 C66
i,j≤2
  
X
+ D + ∂3  xi xj Dij  X
i,j≤2
and
 

X ∂1 βE1 + ∂2 βE2 − ∂1 µH1 − ∂2 µH2
R = − Id + xi xj Dij 
∂1 εE1 + ∂2 εE2 + ∂1 βH1 + ∂2 βH2
i,j≤2

(∂1 + C31 )E1 + (∂2 + C32 )E2 + C34 H1 + C35 H2
+X
C61 E1 + C62 E2 + (∂1 + C64 )H1 + (∂2 + C65 )H2
P P
(∂1 i,j≤2 xi xj Dij )E1 + (∂2 i,j≤2 xi xj Dij )E2
−X P P .
(∂1 i,j≤2 xi xj Dij )H1 + (∂2 i,j≤2 xi xj Dij )H2
We similarly have expressions for J 0 and R0 .

We now consider the admittance map Π. In fact we are considering the
1
2
admittance map together with its inverse (when restricted to T HDiv (∂Ω))
resulting in the 4 × 4 system below which we shall continue to refer to as Π.
By (2.1) and (1.5) this map is


β
 1  

(∂3 E1 − ∂1 E3 ) E

 µ 2
  iωµ  
E2
  
 1  β 
 
(∂3 E2 − ∂2 E3 )  − E1 

 −E  
1  Π  iωµ  µ 
(2.10)  7 →
− +
   
 −1
 
 H2    β 
(∂
 iωε 3 1
 H − ∂ H )
1 3  − H2 
 
−H1 x =0
 ε 
3
−1
   β 

(∂3 H2 − ∂2 H3 ) H1

iωε x3 =0 ε x3 =0
plus terms with coefficients vanishing at x1 = x2 = 0; since we shall be
computing the symbol of Π0 − Π, we do not need to consider such terms by
Lemma 1.6.
Let K be such that KJ = Id modulo smoothing (and similarly define
K 0 ). Note that the principal symbols of K and K 0 at p are both equal to
−|ξ 0 |−1
x X
−1 since X = X 0 and X −1 = (X 0 )−1 there. Then
∂3 E1 − ∂1 E3 + iωβE2 = C11 E1 + C12 E2 + C14 H1 + C15 H2

+ (C13 − ∂1 )(KR)1 + C16 (KR)2 + iωβE2
and similarly, in terms of the second body,
∂3 E1 − ∂1 E3 + iωβE2
= (C11 + F11 )E1 + (C12 + F12 )E2
+ (C14 + F14 )H1 + (C15 + F15 )H2
+ (C13 + F13 − ∂1 )(K 0 R0 )1 + (C16 + F16 )(K 0 R0 )2 + iωβE2 .
Here (KR)j is the j th component of the vector (KR), and similarly for
(K 0 R0 )j . Looking at the difference of these two, the first component of the
image of (E2 , −E1 , H2 , −H1 )0 under Π0 − Π (times iωµ) is
F11 E1 + F12 E2 + F14 H1 + F15 H2 + (C13 + F13 − ∂1 )(K 0 R0 )1
− (C13 − ∂1 )(KR)1 + (C16 + F16 )(K 0 R0 )2 − C16 (KR)2
= F11 E1 + F12 E2 + F14 H1 + F15 H2 + F13 (K 0 R0 )1 + F16 (K 0 R0 )2
+ C13 (K 0 R0 − KR)1 + C16 (K 0 R0 − KR)2
+ ∂1 [(K − K 0 )R]1 − ∂1 [K 0 (R0 − R)]1 .
Now modulo smoothing,
 

0 F33 F36
X
0 0 0
K − K = K(J − J)K = K Id + xi xj Dij X
 K0
F63 F66
i,j≤2
(recall that we compute at x3 = 0). Thus if m = −(l − 1), acting on the

vector (E2 , −E1 , H2 , −H1 )0 , the principal symbol of (K − K 0 )R at the point
p is

−10 σm (F33 ) σm (F36 ) iξ2 −iξ1 0 0
σm−1 ((K − K )R) = 0 2 .
|ξ |x σm (F63 ) σm (F66 ) 0 0 iξ2 −iξ1
Next,

0 F31 E1 + F32 E2 + F34 H1 + F35 H2
R − R = −X
F61 E1 + F62 E2 + F64 H1 + F65 H2
which has symbol of order m equal to zero by Proposition 2.2, and

0 0 1 iξ2 −iξ1 0 0
σ0 (K R ) = 0 .
|ξ |x 0 0 iξ2 −iξ1
We now put all these results together. Recall that the principal symbol of
C is −|ξ 0 |x times the identity matrix, and we have the expression for fm in
Proposition 2.2. The first row of iωµ(Π0 − Π) thus has principal symbol at
p

iξ2 ξ1 ξ2 iξ1 ξ12
|ξ 0 |x σm (F13 ) + |ξ 0 |2x σm (F33 ), −σm (F11 ) − |ξ 0 |x σm (F13 ) − |ξ 0 |2x σm (F33 ),

iξ2 ξ 1 ξ2 iξ1 ξ12
σ (F
|ξ 0 |x m 16 ) + σ (F
|ξ 0 |2 m 36
), −σ (F
m 14 ) − σ (F
|ξ 0 |x m 16 ) − σ (F
|ξ 0 |2 m 36
)
x x
−l!
= l+1
−ξ1 ξ2 (βeβ + εeµ ), −ξ22 (βeβ + εeµ ),
0
2 |ξ |x (εµ + β 2 )
l
−ξ1 ξ2 (µeβ − βeµ ), −ξ22 (µeβ − βeµ ) .

Computing the principal symbol at p of the remaining rows of Π0 − Π = 0

and using Proposition 2.2, we find that in fact
−ξ1 ξ2 −ξ22
 
1
 
0  0
−l!  iωµ −1 
 ξ12 ξ1 ξ2 
0= −1 EX 2
 .
2l |ξ 0 |l+1 −ξ1 ξ2 −ξ2 
  
x 0 0
iωε ξ12 ξ1 ξ2
Evaluating, for example, at ξ1 = 1, ξ2 = 0, we obtain
1
 
µ 0 
−1
 1  EX = 0
0 −
ε
and so E = 0 as desired.
We remark that one could also argue that the values of eβ , eµ are obtain-
able from the first row and it therefore follows by symmetry that the value
of e is deducible from the third row and hence that all three values are
determined.
References
[1] M.S. Joshi, Recovering the total singularity of a potential from backscattering data,
preprint.
[2] , Recovering asymptotics of Coulomb-like potentials, preprint.
[3] M.S. Joshi and A. Sa Barreto, Recovering asymptotics of short range potentials, to
appear in Commun on Math. Phys.
[4] , Recovering asymptotics of metrics from fixed energy scattering data, preprint.
[5] A. Lakhtakia, V.K. Varadan and V.V. Varadan, 1989, Time-Harmonic Electromag-
netic Fields in Chiral Media, Lecture Notes in Physics, 335 (1989), Berlin, Springer-
Verlag.
[6] J. Lee and G. Uhlmann, Determining anisotropic real-analytic conductivities by
boundary measurements, Comm. Pure Appl. Math., 42 (1989), 1097-1112.
[7] S. McDowall, 1997, Boundary determination of material parameters from electromag-
netic boundary information, Inverse Problems, 13 (1997), 153-163.
[8] , An electrodynamic inverse problem in chiral media, Trans. AMS, 1997, to
appear.
[9] P. Ola, L. Päivärinta and E. Somersalo, An inverse boundary problem in electrody-
namics, Duke Math. J. 70 (1993), 617-653.
[10] E. Somersalo, Layer stripping for time harmonic Maxwell’s equations with high fre-
quency, Inverse Problems, 10(2) (1994), 449-466.
[11] E. Somersalo, D. Isaacson and M. Cheney, 1992, A linearized inverse boundary value
problem for Maxwell’s equations, J. Comput. Appl. Math., 42 (1992), 123-136.
Received July 14, 1998. The authors would like to thank the Fields Institute, Toronto,
where the work for this paper was initiated.
University of Cambridge
Cambridge CB2 1SB
England
E-mail address: joshi@dpmms.cam.ac.uk
University of Rochester
Rochester, NY 14627
E-mail address: mcdowall@math.rochester.edu
Vol. 193, No. 1, 2000
ON THE COMPOSITION OF A PRIME

TRANSCENDENTAL FUNCTION AND A PRIME
POLYNOMIAL
Tuen-Wai Ng and Chung-Chun Yang
Let f , g be transcendental entire functions and p, q be

nonlinear polynomials with deg p 6= 3, 6. Suppose that f
and p are prime and f (p(z)) = g(q(z)), then f = g ◦ L and
p = L−1 ◦ q, where L is a linear polynomial. Similar results
for p(f (z)) = q(g(z)) are also obtained.
1. Introduction and Main Results.

A meromorphic function F (z) is said to has a factorization with left factor
f and right factor g provided
(1) F (z) = f (g(z)),

where f is meromorphic and g is entire (g may be meromorphic when f is
rational). A nonlinear meromorphic function F (z) is called prime (pseudo -
prime) if every factorization of form (1) implies that either f is bilinear or g
is linear (either f is rational or g is a polynomial). Clearly, a prime function
is an analogy of a prime number. Over the past thirty years, many classes
of prime or pseudo-prime functions have been obtained (see [2]).
As an analogue of the unique factorizability of natural numbers, one can
also define that concept for entire functions. Suppose an entire function
F has two factorizations f1 ◦ f2 ◦ · · · ◦ fm (z) and g1 ◦ g2 ◦ · · · ◦ gn (z) into
nonlinear entire factors. If m = n and if there exist linear polynomials Lj
(j = 1, 2, 3, . . . , n − 1) such that the relations
(2)
f1 (z) = g1 ◦ L−1
1 , f2 (z) = L1 ◦ g2 ◦ L−1
2 , ...... , fn (z) = Ln−1 ◦ gn (z)
hold simultaneously, then the two factorizations are called equivalent. If any
two factorizations of F (z) into nonlinear, prime entire factors are equivalent
to each other, then F is called uniquely factorizable in entire sense.
As far as just polynomial factors are concerned, it is easy to exhibit
functions which are not uniquely factorizable in entire sense, for instance,
z3 ◦ z2 = z2 ◦ z3.
Therefore, the following question is not without interest.
131
132 TUEN-WAI NG AND CHUNG-CHUN YANG
Problem (A). Suppose f and g are prime entire functions and one of
them is transcendental, will F (z) = f ◦g(z) be uniquely factorizable in entire
sense?
2
Counter-example. Take f (z) = z 2 , g(z) = zez , f1 (z) = ze2z and g1 (z) =
z 2 . All of them are prime functions (see [2]) and f ◦ g = f1 ◦ g1 are two
2
nonequivalent factorizations of z 2 e2z .
In this paper, we shall consider the following problems. Let f and p be
two prime entire functions where f is transcendental and p is a polynomial.
Suppose that f ◦ p = g ◦ q or p ◦ f = q ◦ g. Under what conditions on the
entire functions g, q will these factorizations be equivalent?
From the above counterexample, it is clear that two factorizations of a
function F = h ◦ k = h1 ◦ k1 may not be equivalent. Therefore, we need to
have some further assumptions on these factors h, h1 , k and k1 .
With this in mind, we have come up with the following results. The
functions f, g, p and q considered below are all entire and nonlinear.
Theorem 1. Let f , p be two non-periodic prime entire functions and p be
a polynomial. Suppose that p ◦ f = q ◦ g and both f, g are transcendental.
Then p = q ◦ L−1 and f = L ◦ g, where L is a linear polynomial.
Theorem 2. Let f , p be two prime entire functions and f be transcendental.
Suppose that p ◦ f = q ◦ g and both p, q are polynomials. Then p = q ◦ L−1
and f = L ◦ g, where L is a linear polynomial.
Theorem 3. Let f , p be two prime entire functions and f be transcendental.
Suppose that f ◦ p = g ◦ q and both p, q are polynomials with deg p 6= 3, 6.
Then f = g ◦ L and p = L−1 ◦ q, where L is a linear polynomial.
Theorem 1, 2 and 3 deal with the relationships between polynomials p
and q, transcendental functions f and g when we have factorizations of the
form p ◦ f = q ◦ g or f ◦ p = g ◦ q. It is natural to investigate the case
f ◦ p = q ◦ g.
Theorem 4. Let f and g be two transcendental entire functions, p and q be
two nonlinear polynomials with degree n and m respectively. If f ◦ p = q ◦ g
and p is not a right factor of g, then deg p ≤ deg q. In particular, the
conclusion is true when g is prime.
z3
Remark 1. Let f (z) = ez , g(z) = e 2 , p(z) = z 3 and q(z) = z 2 . Then
f ◦ p = q ◦ g and deg p > deg q. Therefore, the condition that p is not a
right factor of g is essential.
Definition 1. Let F (z) be an nonconstant entire function. An entire func-
tion g(z) is a generalized right factor of F (denoted by g ≤ F ) if there exists
a function f ,which is analytic on the image of g, such that F = f ◦ g. If
such f is entire, g will be a right factor of F (denoted by g|F ).
ON THE COMPOSITION ... 133
Definition 2. If h ≤ f and h ≤ g, we say that h is a generalized common

right factor of f and g. If g ≤ F and f ≤ F , we say that F is a generalized
common left multiple of f and g.
The existence and uniqueness problems of the greatest generalized com-
mon right factor and the least generalized common left multiple for a given
pair of entire functions were solved by A. Eremenko and L.A. Rubel as
follows.
Lemma 1 ([4]). Any pair of non-constant entire functions has (up to a
linear factor) a unique greatest generalized common right factor h, greatest in
the sense that any generalized common right factor of f and g is a generalized
right factor of h.
Lemma 2 ([4]). Suppose that f and g have a generalized common left mul-
tiple. Then f and g have (up to a linear factor) a unique least generalized
common left multiple F , least in the sense that F is a generalized right factor
of any generalized common left multiple of f and g.
The proof of Theorem 1 is mainly based on the following lemma.
Lemma 3 ([10]). Let f and g be two entire functions. Suppose that there
exist two nonconstant complex functions k and R such that F = R◦f = k ◦g
is meromorphic. If g is transcendental and R is rational, then there exists
a transcendental entire function h satisfying h ≤ f and h ≤ g.
Proof of Theorem 1. By Lemma 3, there exists a transcendental entire func-
tion h satisfying h ≤ f and h ≤ g. Hence, f = h1 ◦ h and g = h2 ◦ h, where
h1 , h2 are analytic on the image of h. If the image of h is C − {a}, then
h = a + ek for some entire function k. Without loss of generality, we may
assume a = 0 so that f (z) = h1 (ew )◦k(z). The primeness of f will force k to
be linear. This contradicts the assumption that f is not a periodic function.
So the image of h must be the whole plane. This implies that both h1 , h2
are entire and p ◦ h1 = q ◦ h2 on C. Since f = h1 ◦ h is prime, h1 must be
linear. From p ◦ h1 = q ◦ h2 , h2 must also be linear as p is prime. Take
L = h1 ◦ h−12 and we are done.
The proof of Theorem 2 is similar, we simply apply Lemma 4 below
instead of Lemma 3.
exist two nonconstant polynomials p and q such that p ◦ f (z) = q ◦ g(z).
Then there exist an entire function h and rational functions U (z) and V (z)
such that
f (z) = U ◦ h(z), g(z) = V ◦ h(z).
To prove Theorem 4, we need the following lemma which can be used to
prove Lemma 3.
exist two nonconstant functions h1 and h2 so that F = h1 (f (z)) = h2 (g(z))
and F is meromorphic. Suppose further that there exist k ≥ 2 distinct points
z1 , ....., zk such that F 0 (zi ) 6= 0, ∞ for all i and
(
f (z1 ) = f (z2 ) = ......f (zk )
g(z1 ) = g(z2 ) = ......g(zk ).
Then, there exists an entire function h(z) (independent of k and zi0 s) with
h ≤ f , h ≤ g and h(z1 ) = h(zi ) for all 2 ≤ i ≤ k.
Proof of Theorem 4. By Lemma 1, there exists a generalized greatest
common right factor k of p and g. Since, p is a polynomial, k is actually the
greatest common right factor of p and g. Let p1 and g1 be entire functions
such that p = p1 ◦ k and g = g1 ◦ k. Hence, f ◦ p1 = q ◦ g1 on C and p1 , g1
do not have any nonlinear common right factor. p1 is nonlinear as p is not
a right factor of g. If we can show that deg p1 ≤ deg q1 , then deg p ≤ deg q.
Therefore, we may assume that p and g do not have any nonlinear common
right factor. Suppose that n > m. Define E = {p(z)|F 0 (z) = 0},where
F = f ◦ p. Then E is a countable set. Therefore, we can choose A ∈ C − E
so that the equation p(z) = A has n ≥ 2 distinct roots z1 , . . . , zn . Since
f (A) = f (p(zi )) = q(g(zi )), g(zi ) are roots of the equation q(z) = f (A)
which has at most m roots. n > m implies that there exist two distinct
roots zi , zj such that g(zi ) = g(zj ). Note that p(zi ) = p(zj ) = A and
F 0 (zi ), F 0 (zj ) 6= 0. By Lemma 5, there exists an entire function h with
h ≤ p, h ≤ g and h(zi ) = h(zj ). Clearly h is a polynomial. Hence, there
exists a nonlinear h such that h|p and h|g. This is impossible and we must
have n ≤ m.
In Theorem 3, we only assume that p and q are polynomials. If we further
restrict p and q to have deg p = deg q ≥ 3, then the conclusion of Theorem
3 can be drawn directly from the following lemma.
Lemma 6 ([5]). Let p and q be two polynomials with the same degree. Sup-
pose there exist entire functions f and g such that f ◦ p = g ◦ q. Then one
of the following two cases holds:
(a) p(z) = L ◦ q(z) where L is a linear polynomial.
(b) p(z) = (r(z))2 + a and q = b(r(z) + c)2 + d, where a, b, c, d are complex
numbers.
The above type of results were first investigated by I.N. Baker and F.
Gross in [1] and then L. Flatto in [5]. Finally, S.A. Lysenko in [8] gives an
algebraic necessary and sufficient condition for the existence of meromorphic
f and g satisfy f ◦ p = g ◦ q.
The proof of Theorem 3 is based on a method developed by S.A. Lysenko
in [8] which depends on a fundamental result of local holomorphic dynamics.
2. Local holomorphic dynamics.

Let X be a Riemann surface and let f : (X, a) → (X, a) denote a mapping
defined in some neighbourhood of a point a on X with f (a) = a. A germ
of a mapping f : (X, a) → (X, a) is defined to be the equivalent class of
all mappings which coincide with f in some neighbourhood of a and it is
denoted by [f ]. We say that f is conformal at a if f is analytic in some
neighbourhood of a and f 0 (a) 6= 0. In this case f will have an inverse f −1
in a neighbourhood of a. Let Γ(X, a) be the set of all germs of conformal
mapping (X, a) → (X, a). We define [f ]◦[g] by [f ◦g]. Note that if [f ] = [f1 ],
then f ≡ f1 on any region for which both f and f1 are analytic. Hence, the
binary operation ◦ is well-defined. Clearly, the inverse of [f ] under ◦ is [f −1 ].
Therefore, (Γ(X, a), ◦) is a group. Note that two germs in (Γ(X, a), ◦) are
the same if they have the same Talyor series expansions about a. Therefore,
from time to time, we shall simply denote the germ [f ] by its Talyor series.
a−1
For example, elements of Γ(CP1 , ∞) are of the form a1 z + a0 + +
z
a−2
+ · · · with a1 6= 0. While elements of Γ(C, 0) are of the form a1 z +
z2
a2 z + a3 z 3 + · · · with a1 6= 0.
2
We simply denote Γ(CP1 , ∞) by Γ.

Definition 3. Let p be a nonconstant polynomial. Since p−1 ({∞}) = {∞},
we can define a group Tp = {g ∈ Γ | p ◦ g = p}. Then, it can be shown that
Tp is a cyclic subgroup of Γ and its order equals to deg p.
Example 1. Tz n = {λz | λn = 1} and T(z+1)m = {δz + δ − 1 | δ m = 1}.
Tp is so-called a discrete invariant subgroup of Γ. In fact, we have the
following definition.
Definition 4. A subgroup G of Γ is discrete invariant if there exists a non-
constant function F , meromorphic in a punctured neighbourhood of infinity
in C, such that F (g(z)) = F (z) for all g ∈ G.
In [11], A.A. Shcherbakov proved that if G ⊂ Γ is discrete invariant, then
G is a solvable group.
We also need anothern important necessary condition for Go ⊂ Γ to be
a−1 a−2
discrete. Define Γ1 = g ∈ Γ | g = z + a0 + + 2 + · · · and Γ0 =
z z
n a−1 a−2 o
g∈Γ|g=z+ + 2 + · · · . Clearly, Γ1 /Γ0 is isomorphic to (C, +).
z z
Lemma 7 ([8]). Let G ⊂ Γ, G1 = G ∩ Γ1 and G0 = G ∩ Γ0 . If G is discrete
invariant, then G1 /G0 is isomorphic to a discrete subgroup of (C, +).
Example 2. Let f ,g be nonconstant meromorphic functions and p, q be
nonconstant polynomials. Suppose that F (z) = f (p(z)) = g(q(z)), then the
group generated by Tp and Tq , denoted by [Tp , Tq ], is a discrete invariant
subgroup of Γ. Hence, [Tp , Tq ] is solvable. If we take p(z) = z n , q(z) =

(z + 1)m and G = [Tz n , T(z+1)m ], then G1 ⊂ {Tb (z) = z + b | b ∈ C} and
G0 = {z}. Now G1 ∼
= G1 /G0 which is isomorphic to a discrete subgroup of
(C, +).
Tp and [Tp , Tq ] are the main objects we shall study. The following two
lemmas which were proved by using Galois Theory will be needed in the
proof of Theorem 3.
Lemma 8 ([8]). Let p and q be two nonconstant polynomials. Define Hp,q
= { σ ∈ Tp | ρσ = σρ for all ρ ∈ Tq }. Then Hp,q = Tp1 , where p1 is a right
factor of p.
Lemma 9 ([8]). If [Tp , Tq ] is finite, then there exist two nonconstant ratio-
nal functions R1 , R2 such that R1 ◦ p(z) = R2 ◦ q(z).
If [Tp , Tq ] is infinite, then [Tp , Tq ] must be non-Abelian as both Tp and
Tq are cyclic. Moreover, if [Tp , Tq ] is also solvable, then we can construct
some groups that are isomorphic to [Tp , Tq ]. These groups come from local
holomorphic dynamics and are easier to deal with.
Definition 5. Let w be a holomorphic vector field on V ⊂ C. Associated
with w, it is well known that there exists a unique local phase flow gw :
U × V → C which is a solution of the Cauchy problem
d
(3) gw (t, z) = w(gw (t, z)), gw (0, z) = z ,
dt
where U ⊂ < is a sufficiently small neighbourhood of 0. For brevity, we
denote gw (t, z) by gwt (z) the time-t transformation for the flow of the holo-
morphic vector field w. Moreover, we have the following important property:

t+s t s
(4) gw (z) = gw (gw (z)),
in the sense that if one side of (4) is defined, so is the other, and they are
t (z) for all t ∈ C, then g t (z) (possibly
equal. If we extend the definition of gw w
divergent) will be a formal solution of Equation (3), which will be denoted
t
as gc
w (z).
Definition 6. If f : V → W is a bijective conformal mapping, then the

forward image f∗ w of the vector field w on V is defined as
(f∗ w)(z) = f 0 (f −1 (z)) × w(f −1 (z)),
for all z ∈ W .
Let k be a natural number. We denote by gzt k+1 the time-t transforma-
∂
tion for the flow of the holomorphic vector field z k+1 ∂z . Express gzt k+1 as
a0 (t) + a1 (t)z + a2 (t)z 2 + · · · and substitute it into Equation (3). Comparing
the coefficient of the constant term, we have a00 (t) = ak+1 0 (t), a0 (0) = 0.
Hence, a0 (t) ≡ 0 on some neighbourhood of zero. By repeating this pro-

cess, it is easy to check that gzt k+1 (z) = z + tz k+1 + · · · . Therefore, for
each sufficiently small real t, gzt k+1 (z) is conformal in some neighbourhood
of zero with gzt k+1 (0) = 0. Note that for complex number |t| < 1, we have
gzt 2 (z) = z + tz 2 + t2 z 3 + t3 z 4 + · · · is conformal in some neighbourhood of
zero.
Now, we consider the set of germs
G(k) = {λgzt k+1 : (C, 0) → (C, 0) | λ ∈ C∗ = C − {0}, t ∈ C}.
We shall show that G(k) under composition is a group. For brevity, denote
λgzt k+1 by (λ, t). For any µ ∈ C∗ , let µ(z) = µz, it is easy to check that
µ−1 ◦ gµt ∗ w ◦ µ satisfies condition (3) and hence gµt ∗ w ◦ µ = µ ◦ gw
t . Similarly,
−k
we have gzt k+1 = gµµ∗ z k+1
t
. Now,
−k −k
(5) gzt k+1 ◦ µ = gµµ∗ z k+1
t
◦ µ = µ ◦ gzµk+1t .
(4) and (5) imply that G(k) is a group under composition. From (4) and
(5), the multiplication table for G(k) has the following form:
(λ, t) × (µ, s) = (λµ, tµ−k + s).
With the above formula, it is easy to prove that the subgroup C(k) = {λz =
λgz0k+1 ∈ G(k)|λk = 1} is the center of G(k) (i.e., set of element commutes
with all elements of G(k)).
Definition 7. Let G and G1 be two groups of germs of conformal mappings
(C, 0) → (C, 0). G and G1 is said to be formally equivalent if there exists
an isomorphism K : G → G1 and a formal series b h whose constant term is
zero and the linear term is nonzero, such that for any f ∈ G,
−1 ◦ f ◦ b
hd h = Kf
d.
The hat over a symbol stands for the corresponding formal series.
Now, we can state the main lemma as follows.
Lemma 10 ([3]). A finitely generated non-Abelian solvable group of all
germs of conformal mapping (C, 0) → (C, 0) is formally equivalent to a
finitely generated subgroup of G(k) for some k.
Remark 2. Let J(z) = 1/z and G be a subgroup of Γ(CP1 , ∞). Then
J −1 GJ = {J −1 ◦ g ◦ J|g ∈ G} is a subgroup of Γ(C, 0). Clearly G and
J −1 GJ are isomorphic and from now on, we shall identify G with J −1 GJ
frequently. For example, Tz n is identified with J −1 Tz n J = {λz | λn = 1} =
{λgz02 | λn = 1} and T(z+1)m is identified with J −1 T(z+1)m J = {δz + δ(δ −
1)z 2 + δ(δ − 1)2 z 3 + δ(δ − 1)3 z 4 + · · · | δ m = 1} = {δgzδ−1
2 | δ m = 1}.
3. Proof of Theorem 3.
Let F (z) = f (p(z)) = g(q(z)). From Example 2, we know that [Tp , Tq ] is
solvable. We shall consider two cases: i) [Tp , Tq ] is finite and ii) [Tp , Tq ] is
infinite.
Suppose that [Tp , Tq ] is finite, then by Lemma 9, there exist two noncon-
stant rational functions R1 , R2 such that R1 ◦ p(z) = R2 ◦ q(z). Express Ri
Pi
as Q i
, where Pi and Qi are polynomials and do not have any common zero.
Without loss of generality, we may assume that P1 is nonconstant. Since Pi
and Qi do not have any common zero, we have F1 = P1 (p(z)) = AP2 (q(z))
for some nonzero constant A. By Lemma 2, there exists a nonconstant entire
function F2 , which is the least generalized common left multiple of p and
q, such that F2 ≤ F1 and F2 ≤ F . From F2 ≤ F1 , it follows that F2 is a
polynomial and hence F2 |F1 and F2 |F . Now, we can let F2 = h◦p = k ◦q for
some polynomials h, k. Note that F2 |F which implies h|f . Since f is prime
and transcendental, h must be linear. Therefore, p = h−1 ◦ k ◦ q, where
h−1 ◦ k is linear because p is prime and q is nonlinear. So, we are done for
case i).
If [Tp , Tq ] is infinite, then it is non-Abelian as both Tp , Tq are finite order

cyclic groups. Since [Tp , Tq ] is also solvable, it follows from Lemma 10 that
[Tp , Tq ] is formally equivalent to a subgroup of G(k) for some natural number
k. Let d = lcm(n, m) where n = deg p and m = deg q. Let λgzt k+1 and
µgzsk+1 be the generators of Tp and Tq respectively. From the multiplication
table of G(k), λn = 1 and µm = 1. Hence, all elements of [Tp , Tq ] are in
Gd (k) = {λgzt k+1 ∈ G(k)|λd = 1}. Therefore, [Tp , Tq ] is actually formally
equivalent to a subgroup of Gd (k).
By Lemma 8 and the fact that p is prime, Hp,q = Tp or Tid . If Hp,q = Tp ,

then [Tp , Tq ] must be abelian which is impossible. So, we have Hp,q = Tid =
{z}. It is easy to check that if h ∈ Gk (k) is an element of finite order, then
h ∈ C(k). Hence, Tp ∩ Gk (k) ⊂ C(k). Note that C(k) is the center of G(k)
and so Tp ∩ Gk (k) ⊂ Hp,q = {z}. Now, we claim that g = gcd(n, k) = 1. Let
n
(λ, t) be a generator of Tp . Then, it is very easy to check that (λ, t) g is an
n
element of Tp ∩ Gk (k). Therefore, (λ, t) g = (1, 0) and hence ng = n. We get
g = gcd(n, k) = 1.
We first consider the case that q is prime. Then, we also have gcd(m, k) =
1. So, if d = lcm(n, m), then gcd(d, k) = 1. We define a map f : Gd (k) →
Gd (1) by f (λgzt k+1 ) = λk gzt 2 . Clearly, f is a group homorphism and sur-
jective. The condition that gcd(d, k) = 1 implies that f is also injective.
Therefore [Tp , Tq ] is isomorphic to a subgroup of Gd (1).
Let λgzt 2 and δgzs2 be the elements of Gd (1) corresponding to generators

of Tp and Tq respectively. Note that
(1, 0) = id = λgzt 2 ◦λgzt 2 · · ·◦λgzt 2 (n times) = (λn , t(λ−(n−1) +· · ·+λ−1 +1)).
So, λ (respectively δ) is a primitive nth root of unity (respectively a primitive
mth root of unity).
By choosing a suitable number r, we have (1, r) × (λ, t) × (1, −r) = (λ, 0).
Therefore, with this conjugation, we may assume t = 0 and this implies
that s 6= 0, for otherwise [Tp , Tq ] will be abelian. By using the automor-
phism λgzt 2 → λgzct2 (c 6= 0) of Gd (1), we may also assume that s = δ − 1.
Hence the generators are of the form λgz02 and δgzδ−1 2 . From Remark 2, we
know that they generate Tz n and T(z+1)m respectively. Therefore [Tp , Tq ]
is isomorphic to G = [Tz n , T(z+1)m ]. From Example 2, G1 ∼ = (G1 /G0 ) ∼
=
([Tp , Tq ] ∩ Γ1 )/([Tp , Tq ] ∩ Γ0 ) which is isomorphic to a discrete subgroup of
(C, +) by Lemma 7.
Suppose Tb ∈ G1 , then Tδb is also in G1 . It is because z + δb = (δz +
δ − 1) ◦ (z + b) ◦ (δ −1 z + δ −1 − 1). Similarly, Tλb ∈ G1 and hence Tb ∈ G1 ,
where is a d th root of unity with d = lcm (n,m). Since G1 is isomorphic
to a nontrivial discrete subgroup of (C, +), it is easy to show that either
G1 = {Tna | n ∈ Z} or G1 = {Tnb+mc | n, m ∈ Z} for some a, b, c ∈ C and
b/c being irrational (see [9], p. 63). We consider the first case: Ta ∈ G1 ,
which implies T2 a+a ∈ G1 . Hence, T2a cos 2π = T−1 ×(2 a+a) ∈ G1 . Thus, 2
d
cos 2πd is some integer which can only be 0, ±1 or ±2. So, it follows that
d ∈ {2, 3, 4, 6}. With similar argrument, we can have the same conclusion
for the second case.
If n = m = 3, 4, 6, then it follows from Lemma 6 that p = L ◦ q, where L
is linear. Hence, [Tp , Tq ] = Tp is finite, which is a contradiction.
If n = m = 2, without loss of generality we√may assume√that p(z) = z 2
and q(z) = (z + c)2 . Then we have F1 = cos z ◦ p = cos( z − c) ◦ q. By
Lemma 2, there exists a nonconstant entire function F2 , which is the least
generalized common left multiple of p and q, such that F2 ≤ F√1 and F2 ≤ F .
Let F2 = h ◦ p = k ◦ q, it follows that h ≤ f and h ≤ cos z. Thus h is
not periodic. By similar argrument used in the proof of Theorem 1, we have
h|f . Since f is prime, h is linear or h = L ◦ f for some linear function L. h is
−1
√ p = h ◦ k ◦ q which is impossible again. Therefore, h =√L ◦ f .
linear implies
Hence, cos z has a prime transcendental right factor f . Write cos z as
h1 ◦ f. Thus cos z = h1 ◦ f (z 2 ). From Theorem 3.10 in [2], f (z 2 ) = cos nz
√ √
which implies f (z) = cos nz . This is impossible as cos nz is not a prime
function.
Now, we can assume that n 6= m and hence d 6= 2, 3. d = 4 implies that
one of n, m equals to 2. We may assume without loss of generality that
n = 2 and p = z 2 , q(z) = z 4 + a3 z 3 + a2 z 2 + a1 z. Since f (p(z)) = f (p(−z)),
g(q(z)) = g(q(−z)), and because q is prime, Lemma 6 implies that q(z) =

L ◦ q(−z). Note that L is linear, then a3 = a1 = 0 and hence q is not prime
which is impossible. If d = 6, n can only be 2,3 or 6. The case for n = 2
can be treated similiar as above and the case n = 3, 6 are excluded from our
considerations.
For general q, we can express q as q2 ◦ q1 where q1 is prime. From the
above discussion, we have f = g ◦ q2 ◦ L−1 and p = L ◦ q1 . Thus, f is prime
implies that q2 is linear and we are done.
4. Further discussions.
In Theorem 3, we assume that both the right factors p, q have polynomial
growth. We can also restrict the left factors f, g to have comparable growth
rate and ask the following question.
Problem (B). Let f and p be two prime entire functions and p is a
polynomial. Suppose that F = f ◦p = g ◦q and both f, g are transcendental.
Are the two factorizations of F equivalent?
This problem is closely related to Problem C below (proposed by C.C.
Yang, see e.g., [7], p. 124), which remains unsolved for more than a decade.
Problem (C). Let f be a pseudo-prime transcendental meromorphic func-
tion and p be a polynomial of degree ≥ 2. Must f (p(z)) be pseudo-prime?
If the answer to Problem C is positive, then the function q in Problem B
must be a polynomial and this reduces to the case handled in Theorem 3.
One may try to solve Problem C for the special case that p(z) = z n , where
n is a prime number.
Similarly, we can ask:
Problem (D). Let f be a pseudo-prime transcendental meromorphic func-
tion and p a polynomial of degree ≥ 3, which has no quadratic right factor.
Must p(f (z)) be pseudo-prime?
In [12], G.D. Song and J. Huang proposed the above problem and solved
it for the case that p(z) = z n with n being an odd number. We proved in
[10] that it is true if f is not of the form H ◦ q, where H is an entire periodic
function and q is a polynomial. One may try to solve Problem D for deg p
is odd first.
Finally, we ask whether the answer of Problem A is yes if both f and g
are assumed to be transcendental?
References
[1] I.N. Baker and F. Gross, On factorizing entire functions, Proc. London Math. Soc.,
18(3) (1968), 69-76.
[2] C.T. Chuang and C.C. Yang, Fix-points and factorization of meromorphic functions,
World Scientific, Singapore, 1990.
[3] P.J. Elizarov, Yu.S. Il’yashenko, A.A. Shcherbakov and S.M. Voronian, Finitely gen-
erated groups of germs of one-dimensional conformal mappings and invariants for
complex singular points of analytic foliations of the complex plane, Advances in So-
viet Math., 14, Amer. Math. Soc., Providence, RI, (1993), 57-105.
[4] A. Eremenko and L.A. Rubel, The arithmetic of entire functions under composition,
Advances in Mathematics, 124 (1996), 334-354.
[5] L. Flatto, A theorem on level curves of harmonic functions, J. London Math. Soc., 1
(1969), 470-472.
[6] W.H.J. Fuchs and G.D. Song, On a conjecture by M. Ozawa concerning factorization
of entire functions, Ann. Acad. Sci. Fennicae Ser. A. I. Math., 10 (1985), 173-185.
[7] Y.Z. He and C.C. Yang, On pseudo-primality of the product of some pseudo-prime
meromorphic functions, Analysis of one complex variable, edited by C.C. Yang, 113-
124, World Scientific, Singapore, 1985.
[8] S.A. Lysenko, On the functional equation f (p(z)) = g(q(z)), where p and q are “gener-
alized” polynomials and f and g are meromorphic functions, Izvestiya: Mathematics,
60 (1996), 89-110.
[9] T.T. Moh, Algebra, World Scientific, Singapore, 1992.
[10] T.W. Ng and C.C. Yang, Certain criteria on the existence of a transcendental entire
common right factor, Analysis, 17, 387-393, 1997.
[11] A.A. Shcherbakov, On the denseness of orbits of the pseudogroup of conformal map-
pings and a generalization of the Huday-Verenov theorem, Vestnik Mosk. Gos. Univ.
Ser. 1, Mat.-Mekh., 2 (1982), 10-15; English transl. in Moscow Univ. Math. Bull., 37
(1982).
[12] G.D. Song and J. Huang, On pseudo-primality of the n-th power of prime entire
functions, Kodai Math. J., 10 (1987), 42-48.
Received July 29, 1998 and revised October 28, 1998. This research was partially sup-
ported by a UGC grant of Hong Kong (project no. HKUST710/96P).
Department of Pure Mathematics and Mathematical Statistics

University of Cambridge
16 Mill Lane, Cambridge CB2 1SB
England
E-mail address: ntw@dpmms.cam.ac.uk
Hong Kong University of Science and Technology
Clear Water Bay
Kowloon, Hong Kong
China
E-mail address: mayang@ust.hk
Vol. 193, No. 1, 2000
A PALEY–WIENER THEOREM
FOR THE INVERSE SPHERICAL TRANSFORM
Angela Pasquale
A Paley–Wiener theorem for the inverse spherical trans-

form is proved for noncompact semisimple Lie groups G which
are either of rank one or with a complex structure. Let K
be a fixed maximal compact subgroup of G. For each K-bi-
invariant function f in the Schwartz space on G, consider the
function f˜ defined on a fixed Weyl chamber a+ by f˜(H) :=
Q mα /2
∆(H) f (exp H). Here ∆(H) := α∈Σ+ (sinh α(H)) ,
+
where Σ is the set of positive restricted roots and mα is
the multiplicity of the root α. The K-bi-invariant functions
f whose spherical transform has compact support are identi-
fied as those for which f˜ extends holomorphically and with a
specific growth to a certain subset of the complexification ac
of a. The proof of the theorem in the rank-one case relies on
the explicit inversion formula for the Abel transform.
Introduction.
The classical Fourier transform is an isomorphism F of the Schwartz space
S(Rn ) onto itself. The space D(Rn ) of compactly supported C ∞ functions on
Rn is dense in S(Rn ), and the classical Paley–Wiener theorem characterizes
its image under F: a function f ∈ S(Rn ) is the image under F of a C ∞
function with support in the Euclidean ball {x ∈ Rn : |x| ≤ r} if and only
if it extends to Cn as an entire function of exponential type r and rapidly
decreasing. This is to say that given any integer N ≥ 0 there exists a
constant σN > 0 so that for all z ∈ Cn
|f (z)| ≤ σN (1 + |z|)−N er|=z| .
Since Rn is self-dual, the same theorem also applies to the inverse Fourier
transform. So the functions in S(Rn ) whose image under F is supported in
{x ∈ Rn : |x| ≤ r} are exactly those extending as entire functions on Cn of
exponential type r and rapidly decreasing.
Let G be a noncompact semisimple Lie group with a maximal compact
subgroup K. We refer to Section 1 for the notation and the basic definitions.
The spherical transform S is the analogue of the Fourier transform for K-
bi-invariant functions on G. Generalizing the notion of rapid decrease used
143
144 ANGELA PASQUALE
to define S(Rn ), Harish-Chandra defined the Schwartz space S(K\G/K)

for the K-bi-invariant functions on G. It contains the set D(K\G/K) of
the K-bi-invariant compactly supported C ∞ functions on G as a dense sub-
space. The spherical transform is an isomorphism of S(K\G/K) onto the
subspace SW (a∗ ) of the Weyl group invariants in the Schwartz space over a∗ .
In this setting, a Paley–Wiener theorem for the spherical transform has been
proved by Helgason [Hel66] for G of rank one or with a complex structure.
The proof for G arbitrary has been completed by Gangolli [Gan71]. As in
the classical case, the Helgason–Gangolli Paley–Wiener theorem character-
izes the image under S of the elements of D(K\G/K) as those functions in
SW (a∗ ) having an entire extension of exponential type and rapidly decreas-
ing, and the rate of growth is determined by the size of the support. But,
unlike the classical case, a Paley–Wiener theorem for the inverse spherical
transform cannot be deduced from that of the spherical transform. The
following question is therefore quite natural: What are the functions in
S(K\G/K) whose spherical transform has compact support?
This paper provides the answer when G is either of rank one or with a
complex structure. The characterization is given in terms of holomorphic
extendibility and growth conditions, and the rate of growth is determined
by the size of the support of the image. The precise statement is given
in Section 2. In the rank-one case the proof of the theorem relies on the
explicit formulas for the Abel transform and its inverse as given by Rouvière
[Rou83]. In the complex case, the theorem is an easy consequence of the
explicit expression of the (elementary) spherical functions.
After the fundamental works of Helgason and Gangolli, a number of au-
thors have proved Paley–Wiener type theorems for the spherical or related
transforms, e.g. [Koo75] and [Bra96] in the rank-one case. Not only
a Paley–Wiener theorem for the inverse spherical transform has not been
considered so far, but also the various estimates required in its proof for
the rank-one case are different from those considered in other Paley–Wiener
type theorems. The main difficulties are in the proof of the sufficiency of the
stated condition, where we also need a detailed analysis of the holomorphic
extendibility across given vertical segments of the complex space.
Acknowledgement. The material presented in this paper is part of the

author’s doctoral dissertation at the University of Washington, Seattle. The
author would like to thank her thesis advisor Prof. R. Gangolli for his
continuous guide and encouragement. The paper has been written while the
author was financially supported by the Dutch Organization for Scientific
Research (NWO).
A PALEY–WIENER THEOREM 145
1. Notation and preliminaries.

In the following, G denotes a connected noncompact real semisimple Lie
group with finite center, and K denotes a fixed maximal compact subgroup
of G. g and k (⊂ g) are the Lie algebras of G and K, respectively. p is the
orthogonal complement of k in g with respect to the Cartan-Killing form B
of g. The dimension of any maximal abelian subspace of p is a constant,
called the (real) rank of G. We fix a maximal abelian subspace a of p. a∗
denotes the (real) dual space of a. gc is the complexification of g and ac is
the complexification of a in gc .
The set of the restricted roots of the pair (g, a) is indicated by Σ. It
consists of all α ∈ a∗ for which the vector space gα := {X ∈ g : [H, X] =
α(H)X for every H ∈ a } contains nonzero elements. mα := dimR gα is the
multiplicity of the restricted root α. Σ+ denotes the set of the positive
restricted roots corresponding to a choice of a Weyl chamber a+ of a.
The restriction of the exponential map of G to a is an analytic diffeomor-
phism onto the abelian subgroup A := exp a. The inverse diffeomorphism
is denoted by log. The action on a of the Weyl group W of the pair (g, k)
induces actions of W on a∗ by duality, on A via the exponential map, and
on ac by complexP linearity.
Set n := α∈Σ+ gα . N := exp n is a simply connected nilpotent subgroup
of G. The map (k, a, n) 7−→ kan is an analytic diffeomorphism of the product
manifold K × A × N onto G, and the resulting decompostion G = KAN is
called the Iwasawa decomposition of G.
Every element x of G can be written as x = k1 ak2 for some k1 , k2 ∈ K
and a ∈ A. a is uniquely determined up to conjugation by elements of W .
This property will be referred to as the Cartan decomposition of G, written
G = KAK.
The Cartan-Killing form B is positive definite on p × p, so hX, Y i :=
B(X, Y ) defines Euclidean structures in p and in a ⊂ p. We extend this
inner product to a∗ by duality, that is we set hλ, µi := hHλ , Hµ i if Hγ is
the unique element in a such that hHγ , Hi = γ(H) for all H ∈ a. Set
|H| = hH, Hi1/2 . For x ∈ G define |x| := |H| if x = k1 exp Hk2 with
k1 , k2 ∈ K and H ∈ a.
If (V, h·, ·i) is a Euclidean space, the Schwartz space on V is the set S(V )
of all rapidly decreasing C ∞ functions on V : a C ∞ function f on V be-
longs to S(V ) provided for every differential operator D on V with constant
coefficients and for every integer N ≥ 0
τD,N (f ) := sup (1 + |v|)N |Df (v)| < ∞,
v∈V
where |v| := hv, vi1/2 . S(V ) is a Fréchet space in the topology defined by
the seminorms τD,N . SW (a) and SW (a∗ ) respectively denote the sets of all
rapidly decreasing C ∞ functions on the Eucidean spaces a and a∗ that are
146 ANGELA PASQUALE
W -invariant. Using the exponential map, the space SW (A) of W -invariant

rapidly decreasing C ∞ functions on A = exp a can be similarly defined.
A function f on G is said to be K-bi-invariant if f (k1 xk2 ) = f (x) for all
x ∈ G and k1 , k2 ∈ K. Because of the Cartan decomposition G = KAK, a
K-bi-invariant function is uniquely determined by its W -invariant restriction
to A. Let D(G) denote the set of the left-invariant differential operators on
G. The Schwartz space S(K\G/K) of K-bi-invariant functions over G is the
set of all K-bi-invariant C ∞ functions on G satisfying the following property:
For every D ∈ D(G) and every integer N ≥ 0
τD,N (f ) := sup (1 + |x|)N d(x)|Df (x)| < ∞.
x∈G
Here d denotes the K-bi-invariant analytic function on G defined by
(1.1)
m
Y sinh α(H) 2α
d(x) := if x = k1 exp Hk2 with k1 , k2 ∈ K, H ∈ a.
+
α(H)
α∈Σ
The seminorms τD,N define a Fréchet topology on S(K\G/K).

Let DK (G) be the set of all left-invariant differential operators on G
which are right-invariant under K. The (elementary) spherical functions
on G are the K-bi-invariant eigenfunctions ϕ of every differential operator
D ∈ DK (G), normalized by the condition ϕ(e) = 1 (e is the identity element
in G). R
Let dk be the Haar measure on K normalized so that K dk = 1. Harish-
Chandra proved that for λ ∈ a∗c the functions
Z
(1.2) ϕλ (x) = e (iλ−ρ)(H(xk)) dk x ∈ G,
K
exhaust the set of spherical functions on G. Here
1 X
(1.3) ρ := mα α
2 +
α∈ Σ
and, for g ∈ G, H(g) is the unique element of a such that exp H(g) is the
A-component of g in the Iwasawa decomposition G = KAN . Moreover,
ϕλ = ϕλ0 if and only if λ = w.λ0 for some w in the Weyl group W . ϕλ (x)
is a real analytic function of x ∈ G and a W -invariant entire function of
λ ∈ a∗c .
The spherical transform, the Abel transform and the Euclidean Fourier
transform are respectively the isomorphisms
∼
S : S(K\G/K) −→ SW (a∗ )
∼
A : S(K\G/K) −→ SW (A)
∼
F : SW (A) −→ SW (a∗ )
defined (up to a constant multiple) by

Z
Sf (λ) := f (x) ϕ−λ (x) dx, f ∈ S(K\G/K), λ ∈ a∗ ,
G
Z
Af (a) := eρ (log a) f (an) dn, f ∈ S(K\G/K), a ∈ A,
N
Z
Fg(λ) := g(a) e−iλ(log a) da, g ∈ SW (A), λ ∈ a∗ .
A
The Haar measures dx, dn and da respectively on G, N and A can be
normalized so that
(1.4) S = F ◦ A.
Via the Cartan decomposition of G, the spherical transform can be given
by integration over the Weyl chamber a+ : For f ∈ S(K\G/K) and λ ∈ a∗
Z
(1.5) Sf (λ) = C f (exp H) ϕ−λ (exp H)[∆(H)]2 dH
a+
where ∆ is the function on a+ defined by
Y mα
(1.6) ∆(H) := [sinh α(H)] 2
α∈Σ+
and C is a constant depending on the normalization of the measures.
An explicit analytic formula for the inverse Abel transform is not available
for arbitrary groups G. The inversion formula for the case G = SO0 (n, 1)
was given first by Takahashi [Tak63]. In 1968, Gangolli solved the complex
case [Gan68]. For the general rank-one case the explicit formula was deter-
mined with different approaches by Eaton [Eat73], Koornwinder [Koo75],
Lohoué and Rychener [LR82], and Rouvière [Rou83]. These formulas will
be described in Sections 3 and 4. For more information on the inversion
formulas of the Abel transform we refer to [Bee88].
2. Statement of the results.

As already observed, a K-bi-invariant function is uniquely determined by its
W -invariant restriction to A. Via the exponential map, the Schwartz space
S(K\G/K) is therefore identified with a subset of SW (a). In the Paley–
Wiener theorems we are going to state, the elements of f ∈ S(K\G/K)
whose spherical transform has compact support supp Sf will be character-
ized in terms of the holomorphic extendibility and growth of the function
H 7→ ∆(H)f (exp H) over suitable subsets of the complexification ac of a.
2.1. The rank-one case. In the rank-one case a further simplification can
be made, as we can intrinsically identify a with R. Indeed, if G is of rank one,
the set Σ+ of the positive restricted roots consists of at most two elements:
α and, possibly, 2α. Let H be the element of a satisfying α(H) = 1. The
148 ANGELA PASQUALE
choice of H and α and the exponential map allow us to identify a, a∗ and A

with R. The Weyl group reduces to {−1, 1} acting on R by multiplication,
so SW (a), SW (a∗ ) and SW (A) become S+ (R), the set of even functions in the
Schwartz space of R.
Formulas (1.1) and (1.3) respectively become
mα m2α
sinh t 2 sinh(2t) 2
(2.1) d(t) = , t ∈ R,
t 2t
1
(2.2) ρ = mα + m2α .
2
One can show that, under the above identifications, S(K\G/K) corresponds
to the set S+ρ (R) of all even C ∞ functions on R such that for every differential
operator D on R with constant coefficients and for every integer N ≥ 0
sup (1 + |t|)N d(t)|Df (t)| < ∞ .
t∈R
We fix the following constants related to the multiplicities mα and m2α :

( (
1 if m2α = 0 2 if mα is even and m2α = 0
(2.3) j := and J :=
2 if m2α 6= 0 1 otherwise.
We shall frequently use the property that if 2α is a restricted root, then mα
is even and m2α is odd. The case mα even and m2α = 0 corresponds to
the particular situation in which all Cartan subalgebras of G are conjugate
(G ∼= SO0 (2n + 1, 1)).
We shall use the following notation:
<+ := {z ∈ C : <z > 0}
=j := i R \ πj Z

(j = 1, 2)
<+
j
+
:= < ∪ =j (j = 1, 2)
∃j := C \ z : <z ≤ 0, =z ∈ πj Z

(j = 1, 2)
Sj := C \ z : =z ∈ πj Z

(j = 1, 2)
Sj := z ∈ C : =z ∈ − πj , πj

(j = 1, 2),
<z and =z being respectively the real and the imaginary part of z ∈ C.
For z ∈ Sj \ (−∞, 0], set (cf. Formula (1.6))
mα m2α
(2.4) ∆(z) = (sinh z) 2 (sinh(2z)) 2 .
1/2
If mjα is odd, (sinh(jz))mjα /2 stands for (sinh(jz))(mjα −1)/2 [sinh(jz)]+ ,
1/2
where [ · ]+ denotes the single-valued holomorphic branch of the square
√
root function determined on C \ (−∞, 0] by 1 = 1. If mα is even and
m2α = 0, then ∆(z) = (sinh z)mα /2 is an entire function on C.
The Paley–Wiener theorem for the inverse spherical transform on rank-

one groups is given by the following theorem.
Theorem 2.1. Let f ∈ S+ρ (R), and let f˜(t) := ∆(t)f (t) for t ∈ (0, ∞).
Then Sf is compactly supported, with supp Sf ⊂ [−r, r], if and only if f˜(t),
t ∈ (0, ∞), extends to a holomorphic function F̃ on ∃j such that:
(1) For every integer N ≥ 0 there is a constant τN > 0 so that for all
z ∈ <+j
mα mjα −J
|F̃ (z)| ≤ τN (1 + |coth z|)(j−1) 2 (1 + |coth(jz)|) 2 (1 + |z|)−N er|=z| .
(2) The function F defined by
F̃ (z)
F (z) := for z ∈ Sj \ (−∞, 0]
∆(z)
extends to be an even holomorphic function on the horizontal strip Sj ,
and F (t) = f (t) for all t ∈ R.
Observe that the growth estimate of F̃ is given on <+ j , not on the en-
tire complex plane. Moreover, a single-valued holomorphic extension of the
function f is generally only obtained in the strip Sj . But, when mα is even
and m2α = 0, the function f has actually a meromorphic extension to all
of C, with poles at most on iπZ \ {0}, with estimated growth on the whole
complex plane. Indeed, in this case, mα /2 is a positive integer. The func-
tion F (z) := F̃ (z)(sinh z)−mα /2 is therefore holomorphic on ∃j . Because of
2, F (z) is an even extension of f (t), t ∈ R, to Sj . Hence it holomorphically
extends to C \ {iπk : k = ±1, ±2, . . . } by setting F (z) := F (−z) if <z < 0.
The growth condition for F̃ stated in 1 becomes: For every integer N ≥ 0
there is a constant τN > 0 such that for all z ∈ <+ j
mα
−1
|F̃ (z)| ≤ τN (1 + |coth z|) 2 (1 + |z|)−N er|=z| .
It follows, in particular, that the function (sinh z)mα −1 F (z) is bounded near
the points iπk (k = ±1, ±2, . . . ). Thus (sinh z)mα −1 F (z) is entire, and
F (z) is a meromorphic extension of f to C.
In the case mα even and m2α = 0, Theorem 2.1 can be therefore equiva-
lently stated as follows.
Theorem 2.2. Suppose mα is even and m2α = 0. Let f ∈ S+ρ (R). Then
supp Sf ⊂ [−r, r] if and only if f (t), t ∈ R, extends to an even meromorphic
function F on C satisfying: For every integer N ≥ 0 there is a constant
τN > 0 such that for all z ∈ C
mα
−1
|∆(z)F (z)| ≤ τN (1 + |coth z|) 2 (1 + |z|)−N er|=z| .
150 ANGELA PASQUALE
Theorem 2.1 is proved in Section 3. The proof is based on the relation

S = F ◦ A, applying the classical Paley–Wiener theorem to F and then
using the explicit inversion formulas for the Abel transform A and for the
inverse Abel transform A−1 .
Observe that when mα = 2 (i.e. G ∼ = SL(2, C) ∼ = SO0 (3, 1)), Theorem
2.2 says that z 7−→ ∆(z)F (z) is an even entire function of exponential type
r and rapidly decreasing. As we shall see shortly, this property generalizes
to arbitrary complex groups.
2.2. The complex case. When G admits a complex structure, all the
restricted roots α have multiplicity mα = 2. The function ∆ in (1.6) becomes
Y
(2.5) ∆(H) = sinh α(H), H ∈ a.
α∈Σ+
Its entire extension to ac will be also denoted by ∆:

Y
∆(H̃) := sinh α̃(H̃), H̃ ∈ ac ,
α∈Σ+
α̃ being the complex linear extension of α ∈ Σ+ to ac .

We define a norm on the complexification ac = a ⊕ ia of a by setting
1/2
|H̃| := |<H̃|2 + |=H̃|2 .
An entire function g on ac is said to be of exponential type r > 0 and rapidly

decreasing provided for every integer N ≥ 0 there is a constant σN > 0 such
that
|g(H̃)| ≤ σN (1 + |H̃|)−N er|=H̃| , H̃ ∈ ac .
For r > 0, Br := {λ ∈ a∗ : |λ| ≤ r} denotes the closed ball around 0 with
radius r in a∗ .
Theorem 2.3. Suppose G has a complex structure. Let f ∈ S(K\G/K).
Then Sf is compactly supported, with supp Sf ⊂ Br , if and only if f (exp H),
H ∈ a, extends to a W -invariant meromorphic function F on ac such that
H̃ 7→ ∆(H̃)F (H̃) is an entire function on ac of exponential type r and
rapidly decreasing.
Theorem 2.3 is proved in Section 4.
3. Proof of Theorem 2.1.

Because of the classical Paley–Wiener theorem, the Fourier transform of a
function g ∈ S+ (R) has support in the interval [−r, r] if and only if it extends
as an even entire function on C which is of exponential type r and rapidly
decreasing. We indicate the set of such functions by H+ r (R).
Under the identifications of A and a∗ with R described in Section 2.1, the

Euclidean Fourier transform F : SW (A) −→ SW (a∗ ) reduces to the classical
Fourier transform on S+ (R). Because of the relation S = F ◦A, our problem
is therefore to describe, for every r > 0, the subset of S+ρ (R) which is the
image of H+ r (R) under the inverse Abel transform A−1 .
For every l ≥ 0, let S+l (R) be the set of the even C ∞ functions f on R
such that for every differential operator D on R with constant coefficients
and for every integer N ≥ 0
sup (1 + |t|)N coshl t|Df (t)| < ∞.
t∈R
Since (1 + |t|)−1 cosh t

≤ sinh t/t ≤ cosh t, t ∈ R, this definition is consis-
tent with our previous definition of S+ρ (R) when l = ρ (cf. Formula (2.2)).
Moreover, S+0 (R) = S+ (R).
As shown by Rouvière [Rou83], the Abel transform A : S+ρ (R) −→ S+ (R)
can be expressed as a composition of elementary transformations A1 and A2 .
Definition 3.1. 1 Let l ≥ 0. For j = 1, 2, Aj is the integral operator from
j
l+
S 2 (R)
+ to S+l (R) defined by
Z +∞
Aj f (t) := Φf [coshj t + x2 ]1/j dx, t ∈ R,
−∞
where Φf (cosh t) := f (t).
A1 (resp. A2 ) can be interpreted as partial Abel transform associated with
one-parameter subgroups of N generated by elements of gα (resp. g2α ).
Theorem 3.2. 2 Up to a constant multiple,
m2α
A = Am
1 ◦ A2
α
.
1 d
For j = 1, 2 define Dj := . Then Dj maps S+l (R) into S+l+j (R)
sinh(jt) dt
for every l ≥ 0. Rouvière proved the following theorem.
Theorem 3.3. 3 For j = 1, 2
j
l+ 2
(3.1) Dj ◦ Aj = Aj ◦ Dj on S+ (R)
(3.2) A2j ◦ Dj2 = −πDj on S+l (R).
1
[Rou83], p. 274. See also pp. 283, 286. Rouvière denotes the operator on S+ρ (R)
corresponding to A by F . The relation between our operators A1 and A2 and Rouvière’s
√ √
operators F0 and F00 is A1 = πF0 , A2 = πF00 .
2
[Rou83], Théorème 1, p. 275. See also Théorème 5, p. 283.
3
[Rou83], Théorème 1, p. 275, and the computations before it. See also Lemme 6,
p. 283, and Lemme 7, p. 286. Observe that Rouvière uses the differential operator
−1 d
.
sinh(jt) dt
152 ANGELA PASQUALE
Up to a constant multiple, A−1 = Am

2
2α
◦ D2m2α ◦ Am mα
1 ◦ D1 .
α
A−1 is therefore an integro-differential operator, which reduces to a dif-

ferential operator when mα is even and m2α = 0. According to the various
possibilities for mα and m2α , there are three cases:
mα /2
A−1 = D1 (mα even, m2α = 0),
(mα +1)/2
A−1 = A1 ◦ D1 (mα odd, m2α = 0),
−1 (m +1)/2 m /2
A = A2 ◦ D2 2α ◦ D1 α (mα even, m2α odd).
(All the equalities are given up to constant multiples.)
3.1. Necessity. To prove the necessity of the condition stated in Theo-

rem 2.1, we consider the “complexifications” of the operators Aj and Dj ,
and we apply them to the holomorphic extension of the functions in H+ r (R)
as prescribed by the formulas for A . −1
Let j > 0, and let Dj be the differential operator on functions on C defined

1 d
by Dj := . Observe that Dj maps even (resp. odd) functions
sinh(jz) dz
into even (resp. odd) functions. The next Proposition 3.7 describes how
the (n, m)-th iterate D2n D1m of D2 , D1 acts on entire functions which are of
exponential type and rapidly decreasing.
Lemma 3.4. Let r, s ≥ 0, and let N be a nonnegative integer. Let g be an
entire function satisfying for all z ∈ C
|g(z)| ≤ σ(1 + |z|)−N es|<z|+r|=z|
for some constant σ > 0. Then there is a constant σ̃ > 0 such that for all
z∈C
|g 0 (z)| ≤ σ̃(1 + |z|)−N es|<z|+r|=z| .
Consequently, for every positive j, Dj g is a meromorphic function, with at
most simple poles on i πj Z, satisfying
|Dj g(z)| ≤ σ̃|sinh(jz)|−1 (1 + |z|)−N es|<z|+r|=z| .

Proof. Fix z ∈ C. Let γ denote the rectangular contour γ of vertices

1 1 1
<z± 2√ 2
, =z± 2√ 2
. For ζ ∈ γ, we have |ζ − z| ≥ 2√ 2
and 2(1 + |ζ|) ≥
1 + |z|. Cauchy’s Formula therefore gives
|g(ζ)|
Z
0 1
|g (z)| ≤ |dζ|
2π γ |ζ − z|2
Z
4
≤ |g(ζ)| |dζ|
π γ
Z
4σ
≤ (1 + |ζ|)−N es|<ζ|+r|=| |dζ|
π γ
2 N +2 σ
Z
1 1
s(|<z|+ √ )+r(|=z|+ √ )
≤ (1 + |z|)−N e 2 2 2 2 |dζ|
π γ
−N s|<z|+r|=z|
= σ̃(1 + |z|) e ,
N +4
√
2√
where σ̃ := 2π
e(r+s)/2 2 σ.
Lemma 3.5. Let r ≥ 0, and let n > 0 and N ≥ 0 be integers. Let l(z) be a
meromorphic function on C satisfying the following properties.
i. (sinh z)2n−1 l(z) is an entire function.
ii. There is a constant ν > 0 such that for all z ∈ C \ iπZ
(1 + |coth z|)n−1
|l(z)| ≤ ν (1 + |z|)−N er|=z| .
|sinh z|n
Then
1. (sinh z)2n+1 D1 l(z) is an entire function.
2. There is a constant ν̂ > 0 (depending on r, n, N ) such that for all
z ∈ C \ iπZ
(1 + |coth z|)n
|D1 l(z)| ≤ ν̂ (1 + |z|)−N er|=z| .
|sinh z|n+1
3. sinh(2z)(sinh z)2n D2 l(z) is an entire function.
4. There is a constant ν̂ > 0 (depending on r, n, N ) such that for all
z ∈ C \ i π2 Z
n
−1 1 + |coth z|
|D2 l(z)| ≤ ν̂|sinh(2z)| (1 + |z|)−N er|=z| .
|sinh z|
Proof. Apply Lemma 3.4 to the entire function g(z) := (sinh z)2n−1 l(z),
using the following inequalities:
(3.3) √1 (1
2
+ |coth ζ|) ≤ e|<ζ| |sinh ζ|−1 ≤ 1 + |coth ζ|, ζ ∈ C \ iπZ.
1
For 3 and 4, observe that D2 = D1 , which gives
2 cosh z
sinh(2z)(sinh z)2n D2 l(z) = (sinh z)2n+1 D1 l(z).

Lemma 3.6. Let r ≥ 0, and let m, n > 0 and N ≥ 0 be integers. Let l(z)
be a meromorphic function on C satisfying the following properties.
i. (sinh(2z))2n−1 (sinh z)2m l(z) is an entire function.
154 ANGELA PASQUALE
ii. There is a constant ν > 0 such that for all z ∈ C \ i π2 Z

(1 + |coth(2z)|)n−1 1 + |coth z| m

|l(z)| ≤ ν (1 + |z|)−N er|=z| .
|sinh(2z)|n |sinh z|
Then
1. (sinh(2z))2n+1 (sinh z)2m D2 l(z) is an entire function.
2. There is a constant ν̂ > 0 (depending on r, n, m, N ) such that for all
z ∈ C \ i π2 Z
(1 + |coth(2z)|)n 1 + |coth z| m

|D2 l(z)| ≤ ν̂ (1 + |z|)−N er|=z| .
|sinh(2z)|n+1 |sinh z|
Proof. Apply Lemma 3.4 to g(z) := (sinh(2z))2n−1 (sinh z)2m l(z), using In-
equality (3.3) together with |coth z| ≤ 2|coth(2z)| + 1.
Proposition 3.7. Let g be an entire function on C which is of exponential
type r > 0 and rapidly decreasing. Let j = 1, 2. Then for every positive
integers n and m:
(j−1)m
1. (sinh(jz))2n−1 (sinh z)2(j−1)m Djn D1 g(z) is an entire function.
2. For every integer N ≥ 0 there is a constant νN > 0 (depending also on
j, n, m) such that for all z ∈ C \ i πj Z
(j−1)m
(1 + |coth(jz)|)n−1 er|=z|

(j−1)m 1 + |coth z|
|Djn D1 g(z)| ≤ νN .
|sinh(jz)|n |sinh z| (1 + |z|)N
(j−1)m
3. If g is even, then Djn D1 g is even and extends to be holomorphic
at 0.
Proof. Suppose first j = 1, and prove 1 and 2 inductively on n. The case
n = 1 follows from Lemma 3.4 (with j = 1 and s = 0), and the inductive
step is provided by Lemma 3.5, Parts 1 and 2 (with l(z) = D1n g(z)). Suppose
then j = 2, and prove 1 and 2 inductively on n for m arbitrarily fixed. The
case n = 1 is obtained from Lemma 3.5, Parts 3 and 4 (with l(z) = D1m g(z)).
The inductive step is given by Lemma 3.6 (with l(z) = D2n D1m g(z)).
If g is even and holomorphic near 0, then g 0 (0) = 0. Hence Dj g is even
and extends to be holomorphic at 0 by setting Dj g(0) = 1j g 00 (0).
We now want to determine the image under the operators Aj (j = 1, 2) of

(j−1)m
the functions h(t) := Djn D1 g(t), t ∈ R, described by Proposition 3.7.
j
l+ 2
If h ∈ S+ (R), then Aj h is a function in S+l (R) that can be written as
Z ∞
Φh [coshj t + x2 ]1/j dx,

Aj h(t) = 2 t > 0.
0
Substitute the variable x ∈ (0, ∞) with the variable w ∈ (0, ∞) defined by

the relation coshj t + x2 = coshj (t + w). Then
x2 = coshj (t + w) − coshj t = 2j sinh t + w2 sinh w2 ,

and
Z∞
sinh (j(t + w))
q
j
Aj h(t) = 2 h(t + w) 1/2 dw, t > 0.
sinh j t + w2 sinh j w2

0
1/2
Since the map z 7→ [sinh(jz)]+ is well defined and holomorphic on Sj \
(−∞, 0], we are led to the following definition.
Definition 3.8. For j = 1, 2, let Acj denote the integral transform given,
for all functions h for which it is well defined, by
Z∞
sinh(j(z + w))
q
j
Acj h(z) := 2 h(z + w) 1/2 dw,
sinh j z + w2 sinh j w2 +

0
z ∈ Sj \ (−∞, 0].
To study the operator Acj we need the following theorem.
Theorem 3.9. 4 Let U be an open subset of C, and let Ψ(z, w) be a con-

tinuous function on U × (0, ∞). Assume:
i. For every w ∈ (0, ∞), Ψ(z, w) is holomorphic in U .
ii. For every compact subset K of U there exists a function MK (w) which
is integrable in (0, ∞) and such that for all z ∈ K and w ∈ (0, ∞)
|Ψ(z, w)| ≤ MK (w).
R∞
Then ψ(z) := 0 Ψ(z, w) dw is holomorphic on U .
Lemma 3.10. Let h be an even holomorphic function on Sj with the fol-
lowing property: For every δ ∈ (0, πj ) there exists a constant Cδ such that
for all z with |=z| ≤ δ
|h(z)| ≤ Cδ |sinh(jz)|−1/2 (1 + |z|)−2 .

Then Acj h(z), z ∈ Sj \(−∞, 0], is well-defined and it extends to an even holo-
morphic function on Sj , which we also denote by Acj h. Moreover, Acj h(t) =
Aj h(t) for all t ∈ R.
4
[Lan93], Lemma 1.1, Chapter XV, p. 392, and [LR70], p. 368, for the M -test.
156 ANGELA PASQUALE
Proof. Set
sinh(j(z + w))
Ψ(z, w) := h(z + w) 1/2 .

By assumption, Ψ(z, w) is holomorphic in z ∈ Sj \ (−∞, 0] for every fixed

w ∈ (0, ∞). For every integer m ≥ 2, let S̃m := {z ∈ C : |=z| ≤ πj − m
1
, <z ≥
−m} \ {z ∈ C : <z < 0, |=z| < 1/m}. Let χA (w) denote the characteristic
function of A ⊂ (0, ∞). Since

sinh(j(z + w)) w + coth j w ,

sinh j z + w sinh j w ≤ coth j z + 2

2
2 2
there is a constant C > 0 so that for (z, w) ∈ S̃m × (0, ∞)
(3.4)
h 1/2 −2 i
|Ψ(z, w)| ≤ C χ(0,m] (w) coth j w2 + χ(m,∞) (w) 1 + (w − m) .
Theorem 3.9 thus guarantees that Acj h(z) is holomorphic on Sj \ (−∞, 0].
We now prove that Acj h is even by showing that Acj h(iy) = Acj h(−iy)
for all y ∈ 0, πj . Let Ly denote the horizontal half-line in <+

j from iy to
infinity. The change of variables u = iy + w gives
q Z
j sinh(ju)
Acj h(iy) = 2 h(u) 1/2
du
[dj (iy, u)]+
Ly
where

u+iy u−iy
dj (iy, u) := sinh j 2 sinh j 2 = 21 [cosh(ju) − cos(jy)].
1/2
By assumption h is holomorphic in the horizontal strip Sj . [dj (iy, ·)]+ is
holomorphic in the domain
Dyπ obtained from Sj by removing the vertical
π

segments −i j , −iy and iy, i j . Therefore the function
sinh(ju)
f (y, u) := h(u) 1/2
[dj (iy, u)]+
is a holomorphic function of u ∈ Dy .
Let R > 1, and let γ R = 5k=1 γkR be the closed curve in Dy pictured in
S
Figure 1 (γ5R is the quarter of circle centered at iy with radius y/R).

i π /j
iy R+iy
γR
4
γ5R γ3R
γ1R
0
γ2R R
Dy
- i π /j
Figure 1. The contour of integration γ.

By Cauchy’s Theorem
Z
f (y, u) du = 0.
γR
Because of the growth condition of h, when u = R + it, t ∈ (0, y),

|f (y, u)| ≤ Cy |sinh(ju)|−1/2 (1 + |u|)−2 |sinh(ju)||dj (iy, u)|−1/2
≤ Cy0 (1 + R)−2 .
Z
Hence lim f (y, u) du = 0. If u ∈ Dy is close to iy, then there is a
R→∞ γ R
3
−1/2
constant C0 = C0 (y) > 0 such that |f (y, u)| ≤ C0 u−iy

2
. Hence f (y, ·)
is integrable along the segment (0, iy) on the imaginary axis with
Z Zy
lim f (y, u) du = −i f (y, it) dt.
R→∞ γ R
1
0
y iθ
Also, if u = iy + Re , θ ∈ (−π/2, 0), then
Z Z0 iθ −1/2 r
ye y y
|f (y, u)| |du| ≤ C0
2R dθ = C0 π −→ 0 as R → ∞,
R 2R
γ5R −π/2
Z Z q
so lim f (y, u) du = 0. Since lim f (y, u) du = − 2j Acj h(iy), then
R→∞ γ R R→∞ γ R
5 4
q
j
Acj h(iy) = 2 [I1 (y) + I2 (y)]
158 ANGELA PASQUALE
where
Zy Zy
sinh(ijt)
I1 (y) = −i f (y, it) dt = −i h(it) 1/2
dt,
[dj (iy, it)]+
0 0
Z∞ Z∞
sinh(jt)
I2 (y) = f (y, t) dt =1/2
dt. h(t)
[dj (iy, t)]+
0 0
Consider now −y and the closed curve ΓR = 5k=1 ΓR

S
k in Dy which is
symmetric to γ R with respect to the real axis (cf. Figure 2).
i π /j
Dy
Γ2R = γ2R R
0
R
Γ1
Γ5
R Γ3R
R
Γ4
- iy R+iy
- i π /j
Figure 2. The contour of integration Γ.

Computations analogous to those made above show
q Z q
j
c
Aj h(−iy) = − 2 lim f (−y, u) du = 2j [I1 (−y) + I2 (−y)]
R→∞
ΓR
4
where
Z Z0
I1 (−y) = lim f (−y, u) du = i f (−y, it) dt
R→∞
ΓR −y
1
Z0
sinh(ijt)
=i h(it) 1/2
dt,
[dj (−iy, it)]+
−y
Z Z∞ Z∞
sinh(jt)
I2 (−y) = lim f (−y, u) du = f (−y, t) dt = h(t) 1/2
dt.
R→∞ [dj (−iy, t)]+
ΓR 0 0
2
Since dj (−iy, t) = dj (iy, t), then I2 (−y) = I2 (y). Since h is even and
dj (iy, it) = dj (−iy, −it), then
Zy Z0
sinh(ijt) sinh(ijt)
I1 (y) = −i h(it) 1/2
dt = i h(it) 1/2
dt = I1 (−y).
[dj (iy, it)]+ [dj (−iy, it)]+
0 −y
Thus Acj h(−iy) = Acj h(iy). Acj h is even, so we can extend it to Sj \ {0}
by setting Acj h(z) := Acj h(−z) if <z < 0. Moreover, (3.4) shows that Acj h
remains bounded on S̃1 ∩ <+ j and hence that it holomorphically extends to
c
Sj . Finally, Aj h and Aj h are continuous even functions of t ∈ R: Since they
agree on (0, ∞), they must agree on all R.
To extend Aj h outside Sj , we need to make its integrand single-valued.
The key observation is that the map
1/2
sinh(jz)
z 7−→

is well-defined and holomorphic on ∃j for every fixed w ∈ (0, ∞).

Definition 3.11. For j = 1, 2, let A+ j denote the integral operator given,
for all functions h for which it is well defined, by
A+
j h(z)
Z∞ " #1/2
sinh(jz)
q
j
:= 2 h(z + w) sinh(j(z + w)) dw,
sinh j z + w2 sinh j w2

0 +
z ∈ ∃j .
The next proposition determines the holomorphic extension of Aj h when
(j−1)m
h = Djn D1 g is given by Proposition 3.7.
Proposition 3.12. Let j = 1, 2 and let n, m be positive integers. Suppose
jn+(j−1)m
h ∈ S+ (R) extends to an even meromorphic function on C with the
following properties.
i. (sinh(jz))2n−1 (sinh z)2(j−1)m h(z) is an entire function.
ii. For every integer N ≥ 0 there is a constant νN > 0 such that for all
z ∈ C \ i πj Z
(j−1)m
(1 + |coth(jz)|)n−1

1 + |coth z|
|h(z)| ≤ νN (1 + |z|)−N er|=z| .
|sinh(jz)|n |sinh z|
iii. h is holomorphic at z = 0.
Then
160 ANGELA PASQUALE
1. A+ +
j h is holomorphic in ∃j , and Aj h(t) = (sinh(jt))
1/2 A h(t) for all
j
t ∈ (0, ∞).
2. For every integer N ≥ 0 there is σN > 0 (depending also on n, m, j)
such that for all z ∈ <+
j
n−1 (j−1)m
1 + |coth(jz)| 1 + |coth z|
|A+
j h(z)| ≤ σN (1 + |z|)−N er|=z| .
|sinh(jz)| |sinh z|
3. Acj h is an even holomorphic extension of Aj h(t), t ∈ R, to Sj .
A+j h(z)
4. Acj h(z) = 1/2
on Sj \ (−∞, 0].
[sinh(jz)]+
Proof. Let
" #1/2
sinh(jz)
Ψ(z, w) = h(z + w) sinh(j(z + w)).
sinh j z + w2 sinh j w2

+
Because of Condition i on h, Ψ is continuous on ∃j ×(0, ∞) and holomorphic
in ∃j for every fixed w ∈ (0, ∞). If z = t + iy ∈ <+ and w ∈ (0, ∞), then
|sinh(j(z + w))| ≥ sinh(jt), and |coth(j(z + w))| ≤ coth(jt). Condition ii
(for N = 0) therefore gives the following estimate:
|Ψ(z, w)|
1 + |coth(j(z + w))| n−1 1 + |coth(z + w)| (j−1)m er|=z|

≤ ν0 1/2
|sinh(j(z + w))| |sinh(z + w)|
sinh j w2
1 + coth(jt) n−1 1 + coth t (j−1)m ery

≤ ν0 1/2
sinh(jt) sinh t
sinh j w2
−1/2
= function bounded on compact subsets of <+ · sinh j w2

.
If z = t + iy ∈ Sj , then sin(jy) 6= 0 6= sin y. Condition ii on h (with N = 0)
gives, for some constant ν00 > 0,
|Ψ(z, w)|
sinh(jt) 1/2 ery

≤ ν00 |sin(jy)|2(1−n) |sin y|2(1−j)m 1 + 1/2
sin(jy)
sinh j w2
−1/2
= ( function bounded on compact subsets of Sj ) · sinh j w2

.
−1/2
Since w 7→ sinh j w2

is integrable on (0, ∞), Theorem 3.9 implies
that A+j h is holomorphic on ∃ +
j = < ∪ Sj .
If (z, w) ∈ <+ j ×(0, ∞),√we have |z+w| ≥ |z|, |sinh(j(z+w))| ≥ |sinh(jz)|,
and |coth(j(z + w))| ≤ 2 + |coth(jz)|. The growth condition for h then
implies: For every integer N ≥ 0 and z ∈ <+

j
|A+
j h(z)|
√ !n−1 √ !(j−1)m
1+ 2 + |coth(jz)| 1+ 2 + |coth z|
q
j
≤ 2 νN
|sinh(jz)| |sinh z|
Z∞
−N r|=z|
−1/2
sinh j w2

· (1 + |z|) e dw
0
n−1 (j−1)m
1 + |coth(jz)| 1 + |coth z|
≤ σN (1 + |z|)−N er|=z| ,
|sinh(jz)| |sinh z|
with
q Z∞
j −1/2
σN := 3[n−1+(j−1)m]/2 sinh j w2

2 νN dw.
0
Property 3 is a consequence of Lemma 3.10. In fact, h is holomorphic and

even on Sj , and if |<z| ≥ 1 and |=z| < π/j, then
(1 + |coth(jz)|)n−1 1 + |coth z| (j−1)m

|h(z)| ≤ ν2 (1 + |z|)−2 er|=z|
|sinh(jz)|n |sinh z|
≤ ν20 |sinh(jz)|−1/2 (1 + |z|)−2 .
A+
j h
Finally, Property 4 follows immediately because Acj h and 1/2
[sinh(jz)]+
are both holomorphic on Sj \ (−∞, 0] and agree with Aj h on (0, ∞).
Proof of Theorem 2.1 (Necessity). Let g ∈ H+ r (R) and let f := A−1 g. Sup-
pose first that mα is even and m2α = 0 (that is j = 1 and J = 2). Then (up
m /2
to a constant multiple) f = D1 α g, and Proposition 3.7 (with j = 1 and
n = mα /2) proves that f extends to an even meromorphic function F on C
satisfying the condition stated in Theorem 2.2.
Suppose now that J = 1, i.e. that either mα is odd (so j = 1) or mα is
even and m2α is odd (j = 2). Then (up to a constant multiple)
(j−1)m
f = A−1 g = Aj Djn D1 g
with n = (mjα + 1)/2 and m = mα /2. Because of Proposition 3.7, we can
(j−1)m
apply Proposition 3.12 to the function h = Djn D1 g.
For z ∈ ∃j , set
F̃ (z) := (sinh z)(j−1)mα /2 (sinh(jz))(mjα −1)/2 A+

j h(z).
162 ANGELA PASQUALE
Observe that the exponents (j − 1)mα /2 and (mjα − 1)/2 are nonnegative
integers, so F̃ (z) is holomorphic on ∃j . Moreover, for t ∈ (0, ∞),
F̃ (t) = (sinh t)(j−1)mα /2 (sinh(jt))(mjα −1)/2 A+
j h(t)
= (sinh t)(j−1)mα /2 (sinh(jt))(mjα −1)/2 (sinh(jt))1/2 Aj h(t)

= (sinh t)mα /2 (sinh(2t))m2α /2 f (t).
The growth condition of A+ +
j h on <j given by Proposition 3.12 determines
the growth estimate 2 for F̃ .
1/2
Let (sinh(jz))mjα /2 := (sinh(jz))(mjα −1)/2 [sinh(jz)]+ if mjα is odd. For
z ∈ Sj \ (−∞, 0],
F̃ (z)
F (z) : =
(sinh z)mα /2 (sinh(2z))m2α /2
(sinh z)(j−1)mα /2 (sinh(jz))(mjα −1)/2 A+
j h(z)
=
(sinh z)mα /2 (sinh(2z))m2α /2
A+j h(z)
= 1/2
[sinh(jz)]+
= Acj h(z).
(j−1)m
Condition 1 then follows from the equality Acj h(t) = Aj Djn D1 g(t) =
f (t) (t ∈ R) and from Proposition 3.12.
3.2. Sufficiency. Before completing the proof of Theorem 2.1, we give, fol-
lowing Rouvière, the explicit form of the Abel transform Af of a function
f ∈ S+ρ (R). Let dX (resp. dX 0 ) denote the Lebesgue measure on gα (resp.
g2α ) corresponding to the Euclidean structure induced by the inner product
(X, Y ) := −B(X, θY ), where B is the Cartan-Killing form and θ is the Car-
tan involution of g. Via SU (2, 1)-reduction, Rouvière proved the following
theorem.
Theorem 3.13. 5 Let f ∈ S+ρ (R). Then there is a constant C so that
Z 1/2
Af (t) = C Φf (cosh t + |X|2 )2 + |X 0 |2 dX dX 0 , t ∈ R,
gα ×g2α
where Φf (cosh t) := f (t). When m2α = 0, disregard the variable X 0 and the
integration over g2α .
For a fixed normalization of the Haar measure dn of N , the constant C
can be explicitely determined as a function of the multiplicities mα and m2α .
5
[Rou83], p. 272,(8). See also p. 283.
For our purposes it is more appropriate to have a different expression for

Af (t), t ∈ (0, ∞). We first pass to spherical coordinates on gα and on g2α ,
and then perform a change of variables in the integral that takes Φf back
to f . Finally, we replace f by f˜. The cases m2α = 0 and m2α 6= 0 are kept
separated.
Case m2α = 0 : For all t ∈ (0, ∞)
Z
Af (t) = C Φf (cosh t + |X|2 ) dX
gα
Z∞
mα −1
mα /2
Φf (cosh t + r2 ) rmα −1 dr

= 2 Cπ Γ 2
0
Z∞
mα −1
= C0 w w

f (t + w) sinh t + 2 sinh 2
2
sinh(t + w) dw
0
Z∞ " # m2α −1
sinh t + w2 sinh w

= C0 ˜
f (t + w) 2
dw,
sinh(t + w)
0
π mα /2 mα −1
where C 0 = 2

2 Γ 2 C.
Case m2α 6= 0 : For all t ∈ (0, ∞)
Af (t)
Z 1/2
=C Φf (cosh t + |X|2 )2 + |X 0 |2 dX dX 0
gα ×g2α
Z∞Z∞
4 C π (mα +m2α )/2 1/2 mα −1 m2α −1
= Φf (cosh t + r2 )2 + s2 r s dr ds
Γ m2α Γ m22α

0 0
Z∞Z∞
mα −1
= C 00 w w

f (t + w + v) sinh t + 2 sinh 2
2
sinh(t + w)
0 0
m2α
· [sinh(2(t + w) + v) sinh v] 2 −1 sinh(2(t + w + v)) dw dv
Z∞Z∞ "
w
w
# m2α −1
00 sinh t + sinh
=C f˜(t + w + v) 2 2
sinh(t + w + v)
0 0
m2α
sinh(2(t + w) + v) sinh v 2 −1 sinh(t + w)

· dw dv,
sinh(2(t + w + v)) sinh(t + w + v)
m /2 −1
where C 00 = 2 π2 α π m2α /2 Γ m2α Γ m22α

C.
164 ANGELA PASQUALE
In the above integrals, the variable r ∈ (0, ∞) has been replaced by the
variable w ∈ (0, ∞) defined by the relation
cosh t + r2 = cosh(t + w),
and the variable s ∈ (0, ∞) has been replaced by the variable v ∈ (0, ∞)
defined by the relation
cosh2 (t + w) + s2 = cosh2 (t + w + v).
The constants C 0 and C 00 do not affect the result we want to prove. We
therefore disregard them.
The idea to prove the sufficiency of the conditions in Theorem 2.1 is
the following. The hypothesis on f imposed by Theorem 2.1 involve the
holomorphic extension F̃ to ∃j of the function f˜(t) := ∆(t)f (t), t ∈ (0, ∞).
The function f˜ also appears in the integrand of Af (t), t ∈ (0, ∞). We
formally extend Af to AF on ∃j by replacing f˜(t) by F̃ (z), and the variable
t by the variable z in the remaining hyperbolic sines. A little extra care is
required when dealing with square roots. The growth condition for F̃ on
<+j is used to prove that AF is holomorphic on (some open neighborhood
of) <+j . Condition 2 in Theorem 2.1 is employed to show that AF is even,
which allows us to extend it to C \ i πj Z. Finally, the growth condition is
used again, to prove either that AF is bounded near each point in i πj Z (and
hence it is entire) or that AF is rapidly decreasing with exponential growth
r.
Definition 3.14. Let f ∈ S+ρ (R) satisfy the conditions stated in Theo-
rem 2.1. For z ∈ ∃j , formally define
Z∞
mα
−1
AF (z) := F̃ (z + w) [s1 (z, w)] 2 dw, if m2α = 0,
0
Z∞Z∞
mα m2α
−1 −1
AF (z) := F̃ (z + w + v) [s2 (z, w, v)] 2 [s3 (z, w, v)] 2
0 0
sinh(z + w)
· dw dv, if m2α 6= 0,
sinh(z + w + v)
where
sinh z + w2 sinh w2

s1 (z, w) := ,
sinh(z + w)
sinh z + w2 sinh w2

s2 (z, w, v) := ,
sinh(z + w + v)
sinh(2(z + w) + v) sinh v
s3 (z, w, v) := ,
sinh(2(z + w + v))
mjα mjα −1
−1 1/2
and [∗] 2 := [∗] 2 [1/∗]+ if mjα is odd.
Remark 3.15. By definition, the function AF extends, up to a constant
multiple, Af (t), t ∈ (0, ∞). Observe that the square roots appearing in
the formula when mjα is odd are well-defined single-valued holomorphic
functions of z ∈ ∃j for all v, w ∈ (0, ∞).
Lemma 3.16. Define
(
∗ x−1 if x ∈ (0, 1)
s (x) :=
1 if x ∈ [1, ∞).
Then there is a constant C > 0 such that for every ζ = a + ib with a ≥ 0

and for every x ∈ (0, ∞)

sinh(ζ + 2x)
max , 1 + |coth(ζ + x)| ≤ C s∗ (x).
sinh(ζ + x) sinh x
Lemma 3.17. Let Ψ(z, w) and Ψ(z, w, v) denote respectively the integrands
of AF (z) when m2α = 0 and m2α 6= 0. Then, for every integer N ≥ 0 there
are constants ηN , µN > 0 such that for all z ∈ <+
j and v, w ∈ (0, ∞)
|Ψ(z, w)| ≤ ηN (s∗ (w))(2−J)/2 (1 + |z + w|)−N er|=z| ,

|Ψ(z, w, v)| ≤ µN (s∗ (w))3/4 (s∗ (v))3/4 (1 + |z + w + v|)−N er|=z| .
Proof. Observe first that there is a constant C > 0 so that for all (ζ, x) ∈
<+j × (0, ∞) and j = 1, 2

sinh(jζ + x) sinh x
(3.5) sinh(jζ + 2x) (1 + |coth(jζ + 2x)|) ≤ C.

When mα > 1 and m2α = 0, the exponent (mα /2) − 1 of the function
s1 (z, w) in Definition 3.14 is positive. The growth condition for F̃ gives: For
every integer N ≥ 0 there is a constant τN > 0 such that
|Ψ(z, w)|
(m −J)/2
≤ τN |s1 (z, w)|(mα /2)−1 1 + |coth(z + w)| α (1 + |z + w|)−N er|=z|
(mα /2)−1 (2−J)/2
≤ τN |s1 (z, w)| 1 + |coth(z + w)| 1 + |coth(z + w)|
· (1 + |z + w|)−N er|=z|
≤ ηN (s∗ (w))(2−J)/2 (1 + |z + w|)−N er|=z| .
166 ANGELA PASQUALE
If mα = 1 and m2α = 0, then J = 1 and

1/2
sinh(z + w)
|Ψ(z, w)| = F̃ (z + w)

sinh z + w2 sinh w

2
≤ τN0 (1 + |z + w|)−N er|=z| (s∗ (w/2))1/2

≤ ηN (s∗ (w))1/2 (1 + |z + w|)−N er|=z| .
Suppose now m2α > 1. In this case the exponents of both functions
s2 (z, w, v) and s3 (z, w, v) are positive. The growth condition for F̃ gives:
For every integer N ≥ 0 there is a constant τN > 0 such that
|Ψ(z, w, v)|

(mα /2)−1
sinh(z + w)
(m2α /2)−1
≤ τN |s2 (z, w, v)| |s3 (z, w, v)|
sinh(z + w + v)
· (1 + |coth(z + w + v)|)mα /2 (1 + |coth(2(z + w + v))|)(m2α −1)/2
· (1 + |z + w + v|)−N er|=z|
≤ τN0 [ |s2 (z, w, v)| (1 + |coth(z + w)|) ](mα /2)−1 (1 + |coth(z + w + v)|)
· [|s3 (z, w, v)| (1 + |coth(2(z + w + v))|)](m2α /2)−1
· (1 + |coth(2(z + w + v))|)1/2 (1 + |z + w + v|)−N er|=z|
≤ τN00 (1 + |coth(z + w + v)|) (1 + |coth(2(z + w + v))|)1/2
· (1 + |z + w + v|)−N er|=z| by Inequality (3.5)
≤ µN (s∗ (w))3/4 (s∗ (v))1/4 (s∗ (2v))1/2 (1 + |z + w + v|)−N er|=z|
by Lemma 3.16
≤ µN (s∗ (w))3/4 (s∗ (v))3/4 (1 + |z + w + v|)−N er|=z| .
When m2α = 1, the exponent of s3 (z, w, v) is −1/2. Hence
|Ψ(z, w, v)|
= |F̃ (z + w + v)||s2 (z, w, v)|(mα /2)−1
sinh(2(z + w + v)) 1/2 sinh(z + w)

·
sinh(2(z + w) + v) sinh v sinh(z + w + v)
≤ τN0 |s2 (z, w, v)|(mα /2)−1 (1 + |coth(z + w + v)|)mα /2 (s∗ (v))1/2
· (1 + |z + w + v|)−N er|=z|
≤ τN00 [ |s2 (z, w, v)| (1 + |coth(z + w + v)|) ](mα /2)−1
· (1 + |coth(z + w + v)|)(s∗ (v))1/2 (1 + |z + w + v|)−N er|=z|
≤ µN (s∗ (w))3/4 (s∗ (v))1/4 (s∗ (v))1/2 (1 + |z + w + v|)−N er|=z|
by Inequality (3.5) and Lemma 3.16

≤ µN (s∗ (w))3/4 (s∗ (v))3/4 (1 + |z + w + v|)−N er|=z| .

Proposition 3.18. Let f and AF be as in Definition 3.14. Then AF is
holomorphic in <+ and continuous on <+ j . Moreover, for every integer
N ≥ 0 there is a constant σN ≥ 0 such that for all z ∈ <+
j
|AF (z)| ≤ σN (1 + |z|)−N er|=z| .

Proof. Let Ψ(z, w) and Ψ(z, w, v) be as in Lemma 3.17. The assumption on
F̃ and Remark 3.15 ensure that they are holomorphic functions of z ∈ ∃j and
continuous functions on ∃j × (0, ∞) and ∃j × (0, ∞) × (0, ∞), respectively.
The estimates in Lemma 3.17, Theorem 3.9 and the Dominated Convergence
Theorem prove that the function AF is holomorphic in <+ and continuous
in <+j .
To determine the growth of AF on <+ j , observe that if M is an even
integer ≥ 4, l ∈ (0, 1), and <ζ ≥ 0, then there is a constant C > 0 so that
Z∞
(3.6) (s∗ (x))l (1 + |ζ + x|)−M dx ≤ C (1 + |ζ|)−M +1 .
0
Indeed, there exist constants C1 and C2 such that
Z1 Z1
∗ −M −M
l
(s (x)) (1 + |ζ + x|) dx ≤ (1 + |ζ|) x−l dx ≤ C1 (1 + |ζ|)−M
0 0
and6
Z∞ Z∞
(s (x)) (1 + |ζ + x|) dx = (1 + |ζ + x|)−M dx
∗ l −M
1 1
Z∞
≤ (1 + |ζ|2 + x2 )−M/2 dx
0
1 · 3 · 5 · · · (M − 3) π
= (1 + |ζ|2 )−(M −1)/2
2 · 4 · 6 · · · (M − 2) 2
≤ C2 (1 + |ζ|)−M +1 .
6
[Dwi61], Formula 856.21. The formula can be applied because of the assumption that
M/2 is an integer ≥ 2.
168 ANGELA PASQUALE
For every integer N ≥ 0, choose

(
N + 4 + 2(j − 1) if N is even
M :=
N + 3 + 2(j − 1) if N is odd.
Then M is an even integer ≥ 4.
If m2α = 0 (j = 1), the estimate in Lemma 3.17 yields
Z∞
|AF (z)| ≤ ηM er|=z|
(s∗ (w))(2−J)/2 (1 + |z + w|)−M dw
0
≤ ηM (1 + |z|)−M +1 er|=z|
0
by (3.6)
≤ σN (1 + |z|)−N er|=z| .
R∞
If m2α 6= 0 (j = 2), set Ψ1 (z, v) = Ψ(z, w, v) dw. Then
0
Z∞
∗
|Ψ1 (z, v)| ≤ µM (s (v)) 3/4 r|=z|
e (s∗ (w))3/4 (1 + |z + w + v|)−M dw
0
0 ∗
≤ µM (s (v)) 3/4
(1 + |z + v|)−M +1 er|=z| by (3.6)
≤ µ0M (s∗ (v))3/4 (1 + |z + v|)−M +2 er|=z| ,
and, since M − 2 is again an even integer ≥ 4,
Z∞
|AF (z)| ≤ |Ψ1 (z, v)| dv
0
Z∞
0
≤ µM e r|=z|
(s∗ (v))3/4 (1 + |z + w|)−M +2 dv
0
≤ µM (1 + |z|)−M +3 er|=z|
00
by (3.6)
≤ σN (1 + |z|)−N er|=z| .

We now prove that, under the above assumptions, AF (iy) is a real an-
alytic function of y on the interval Ik := k πj , (k + 1) πj for every integer
k. Thus AF extends holomorphically across each vertical segment iIk . The
proof is an application of the classical criterion for which a C ∞ function g
is real analytic on an open interval I ⊂ R if and only if for every compact
K ⊂ I there is a constant M > 0 such that
h
d g h+1
dy h (y) ≤ M h!

for all y ∈ K and all integers h ≥ 0.

Lemma 3.19. Let j = 1, 2, z ∈ Sj and a, b, c ∈ [0, ∞) with b 6= 0. Define
sinh(jz + a + b) sinh b
s(j, z, a, b, c) := .
sinh(jz + a + 2b + c)
Then there are functions l(j, z) and m(j, z) which are bounded on compact
subsets of Sj such that
|s(j, z, a, b, c)| ≤ l(j, z)
|s(j, z, a, b, 0)|−1 ≤ m(j, z)s∗ (b)
for all j, z, a, b, c (s∗ is the function defined in Lemma 3.16).
Lemma 3.20. Let f and AF be as in Proposition 3.18. Then AF (iy) is a
real analytic function of y ∈ Ik := k πj , (k + 1) πj for every integer k.
Proof. Observe first that if s1 (z, w), s2 (z, w, v), s3 (z, w, v) are as in Defini-
tion 3.14 and if s(j, z, a, b, c) is as in Lemma 3.19, then
sinh z + w2 sinh w2

s(1, z, 0, w/2, 0) = = s1 (z, w)
sinh(z + w)
sinh z + w2 sinh w2

s(1, z, 0, w/2, v) = = s2 (z, w, v)
sinh(z + w + v)
sinh(2(z + w) + v) sinh v
s(2, z, 2w, v, 0) = = s3 (z, w, v).
sinh(2(z + w + v))
Moreover, for z = t + iy ∈ Sj and w, v ∈ (0, ∞) we have

sinh(z + w) cosh t
sinh(z + w + v) ≤ |sin y| .

Set
S1 (z, w) := [s1 (z, w)](mα /2)−1
sinh(z + w)
S2 (z, w, v) := [s2 (z, w, v)](mα /2)−1 [s3 (z, w, v)](m2α /2)−1 .
sinh(z + w + v)
Then S1 and S2 are holomorphic functions of z = t + iy ∈ Sj . Suppose first
mα 6= 1 and m2α 6= 1. If l(j, z) is the function in Lemma 3.19, we have
(3.7) |S1 (z, w)| ≤ [l(1, z)](mα /2)−1
cosh t
(3.8) |S2 (z, w, v)| ≤ [l(1, z)](mα /2)−1 [l(2, z)](m2α /2)−1 .
| sin y|
Observe that the right-hand sides of (3.7) and (3.8) are bounded on the
compact subsets of Sj and do not depend on w, v.
170 ANGELA PASQUALE
Let k be an arbitrarily fixed integer. For simplicity, the dependence on the

choice of k will be omitted in the following notation. For every δ ∈ (0, π/4j),
consider the open half-strip
n o
Sj,δ := z ∈ C : <z > 0, =z ∈ k πj + 2δ, (k + 1) πj − 2δ
and its left edge
Ij,δ = i k πj + 2δ, (k + 1) πj − 2δ .

Let Qj,δ be the open rectangle {z ∈ Sj,δ : <z π< 2}. Then the closed rectangle
π π
Rj,δ of vertices (2+δ)+i j k+δ , (2+δ)+i j (k+1)−δ , −δ+i j (k+1)−δ
and −δ + i πj k + δ contains Qj,δ and is entirely contained in Sj .

Let K be the supremum of the right-hand sides of (3.7) and (3.8) over
z = t + iy ∈ Rj,δ . By Cauchy’s Inequalities, for every integer h ≥ 0,
h h
∂ K ∂ K
(3.9) ∂y h S1 (iy, w) ≤ δ h h! and ∂y h S2 (iy, w, v) ≤ δ h h!

for all iy ∈ Ij,δ .

Suppose z ∈ Sj,δ \Qj,δ . Then the circle
n Γ centered at z with radius δ is en-
o
tirely contained in the subset D := ζ ∈ C : <ζ ≥ 1, =ζ ∈ πj k, πj (k + 1)
of <+ +
j . The growth estimate of F̃ on <j with N = 2j gives for all ζ ∈ D
|F̃ (ζ)|
≤ τ2j (1 + |coth ζ|)(j−1)mα /2 (1 + |coth(jζ)|)(mjα −J)/2 (1 + |ζ|)−2j er|=ζ|
≤ τ 0 (1 + |ζ|)−2j
for some constant τ 0 (depending on k). For ζ ∈ Γ, |ζ| ≥ |z| − δ ≥ |z| − 1 ≥
|z|/2, so (1 + |ζ|)−2j ≤ 22j (1 + |z|)−2j . Applying Cauchy’s Integral formula,
we obtain, for all h ≥ 0,
|F̃ (ζ)| |dζ| τ0
Z
≤ h! 22j h (1 + |z|)−2j .
(h)
F̃ (z) ≤ h!
|ζ − z| h+1 2π δ
Γ
Therefore, for some constant τ 00 ,

00
h
∂
≤ h! τ (1 + w2 + (j − 1)v 2 )−j

(3.10)
∂y h F̃ (iy + w + (j − 1)v) δh
for all integers h ≥ 0, iy ∈ Ij,δ and w, v ∈ (0, ∞) with w + (j − 1)v ≥ 2.
Since F̃ is holomorphic on ∃j , we can conclude that for every integer h ≥ 0
there is a constant M̃ > 0 such that for all iy ∈ Ij,δ and w, v ∈ (0, ∞)
h
∂ M̃ 2 2 −j
(3.11) ∂y h F̃ (iy + w + (j − 1)v) ≤ δ h h! (1 + w + (j − 1)v ) .

Let Ψ(z, w) = F̃ (z + w)S1 (z, w) and Ψ(z, w, v) = F̃ (z + w + v)S2 (z, w, v).

Then for every integer h ≥ 0 and all iy ∈ Ij,δ , w, v ∈ (0, ∞)
h X h h−n n
∂ h ∂ ∂
∂y h Ψ(iy, w) ≤ F̃ (iy + w) n S1 (iy, w)

n ∂y h−n ∂y
n=0
h
!
X h! M̃ 2 −1 K
≤ (h − n)! (1 + w ) n!
n!(h − n)! δ h−n δn
n=0
M̃ K
≤ (h + 1)h! h (1 + w2 )−1
δ
h
2
≤ h! M̃ K(1 + w2 )−1
δ
and, similarly,
h h
∂ 2
M̃ K(1 + w2 + v 2 )−2 .

∂y h Ψ(iy, w, v) ≤ h! δ

If mα = 1 or m2α = 1, then for all z = t + iy ∈ Sj and w, v ∈ (0, ∞)

|S1 (z, w)| = |s1 (z, w)|−1/2 ≤ m(1, z)1/2 (s∗ (w/2))1/2 ,

(mα /2)−1 −1/2 sinh(z + w)

|S2 (z, w, v)| = |s2 (z, w, v)| |s3 (z, w, v)| sinh(z + w + v)
cosh t ∗
≤ (l(1, z))(mα /2)−1 (m(2, z))1/2 (s (v))1/2
|sin y|
where l(j, z) and m(j, z) are as in Lemma 3.19. If K is an upper bound
for (2m(1, z))1/2 and (l(1, z))(mα /2)−1 (m(2, z))1/2 cosh t|sin y|−1 over all z =
t + iy ∈ Rj,δ , then
h h
∂ K ∗
∂ K
∂y h S1 (iy, w) ≤ δ h h!(s (w))
1/2
and h S2 (iy, w, v) ≤ h h!(s∗ (v))1/2
∂y δ
for all iy ∈ Ij,δ and w, v ∈ (0, ∞). Computations as above therefore give
h h
∂
≤ h! 2 M̃ K(s∗ (w))1/2 (1 + w2 )−1

∂y h Ψ(iy, w) δ
h h
∂ 2
M̃ K(s∗ (v))1/2 (1 + w2 + v 2 )−2 .

∂y h Ψ(iy, w, v) ≤ h! δ

Differentiation under integral sign then proves that, for any multiplicities
mα and m2α , AF (iy) is C ∞ on each Ij,δ and that, for some constant M > 0,
h
d h+1
dy h AF (iy) ≤ M h! .

172 ANGELA PASQUALE
Since the sets Ij,δ cover iIk , AF (iy) is a real analytic function of y ∈ Ik .
Our last step is to prove that AF is even. We need the following lemma.
Lemma 3.21. Let j = 1, 2, and let n > 0 and m ≥ 0 be integers. Suppose
h is an even holomorphic function on Sj with the following property. For
every δ ∈ (0, π/j) there is a constant Cδ > 0 such that for all z with |=z| ≤ δ
|h(z)| ≤ Cδ |sinh(jz)|−n/2 |sinh z|−(j−1)m/2 (1 + |z|)−2[n+1+(j−1)m] .
Then h satisfies the hypothesis of Lemma 3.10.
Moreover, for every δ ∈ (0, π/j) there is a constant Cδ0 > 0 such that
whenever |=z| ≤ δ
|Acj h(z)| ≤ Cδ0 |sinh(jz)|−(n−1)/2 |sinh z|−(j−1)m/2 (1 + |z|)−2[n+(j−1)m] .

Proof. Since h is holomorphic on Sj , the estimate on |=z| ≤ δ describes the
growth of h(z) only for large values of |<z|, where
|sinh(jz)|−n/2 |sinh z|−(j−1)m/2 ≤ |sinh(jz)|−1/2 .
It is therefore clear that h satisfies the hypothesis of Lemma 3.10.
For z ∈ Sj \ (−∞, 0] and w ∈ (0, ∞), let
sinh(j(z + w))
Ψ(z, w) := h(z + w) 1/2 .

Because of Lemma 3.16, if |=z| ≤ δ and <z ≥ 0, then

sinh(j(z + w)) 1/2

|Ψ(z, w)| ≤ Cδ
|sinh(j(z + w))|−(n−1)/2
sinh j w2 sinh j w2

· |sinh(z + w)|−(j−1)m/2 (1 + |z + w|)−2[n+1+(j−1)m]

≤ Ĉδ (s∗ (w))1/2 |sinh(jz)|−(n−1)/2 |sinh z|−(j−1)m/2
· (1 + |z + w|)−2[n+1+(j−1)m] .
Formula (3.6) therefore implies: For |=z| ≤ δ and <z ≥ 0
|Acj h(z)| ≤ Ĉδ |sinh(jz)|−(n−1)/2 |sinh z|−(j−1)m/2
Z∞
· (s∗ (w))1/2 (1 + |z + w|)−2[n+1+(j−1)m] dw
0
≤ Cδ |sinh(jz)|−(n−1)/2 |sinh z|−(j−1)m/2 (1
0
+ |z|)−2[n+(j−1)m]
for some constant Cδ0 . Since Acj h(z) is even, this estimate holds also for
<z ≤ 0.
Proposition 3.22. In the assumptions of Proposition 3.18, AF is an even

holomorphic function on a neighborhood of <+
j .
Proof. The growth condition for F̃ and the fact that 1+|coth(jz)| is bounded
for large |<z| imply that the even holomorphic function F satisfies the hy-
pothesis for h in Lemma 3.21 with n = mjα and m = mα . The same
lemma also ensures that the application of Ac2 to F m2α -times and of Ac1 to
(Ac2 )m2α F mα -times is legittimate and gives an even holomorphic function
on Sj .
F holomorphically extends f (t), t ∈ R, to Sj . Hence (Ac1 )mα (Ac2 )m2α F (z)
m2α
holomorphically extends Am 1 A2
α
f (t), t ∈ R, to Sj . Up to constant mul-
tiples, AF (z) holomorphically extends Af (t), t ∈ (0, ∞) to some neighbor-
mα m2α
hood U of <+ j , and, because of Theorem 3.2, Af (t) = A1 A2 f (t) on
R. Thus, up to a constant, AF (z) must agree with (A1 ) (Ac2 )m2α F on
c mα
Sj ∩ U , and, therefore, it is even.

Proof of Theorem 2.1 (Sufficiency). Proposition 3.18 and Lemma 3.20 proved
that AF is holomorphic in a neighborhood of <+ j . Because of Proposi-
tion 3.22, AF has to be even, so we can extend it holomorphically to C\i πj Z
by setting AF (z) := AF (−z) if <z < 0. The growth condition proved in
Proposition 3.18 therefore holds on all C\i πj Z. In particular, AF is bounded
near each point in i πj Z, and therefore it extends to be entire. By continuity,
the growth condition can be extended to C to become: For every integer
N ≥ 0 there is a constant σN > 0 such that for all z ∈ C
|AF (z)| ≤ σN (1 + |z|)−N er |=z|.
Thus: If f satisfies the conditions stated in Theorem 2.1, then Af (t), t ∈ R,
extends to be an even entire function AF on C which is of exponential type
r and rapidly decreasing.
4. Proof of Theorem 2.3.

The study of the spherical transform on complex groups is greatly simplified
by an explicit formula for the elementary spherical functions. Let (a∗c )0 :=
{λ ∈ a∗c : hα, λi =6 0 for all α ∈ Σ}, where h·, ·i denotes the C-bilinear
extension to ac of the inner product in a∗ induced by the Cartan-Killing
∗
form. Then, for every λ ∈ (a∗c )0 and H ∈ a

+ π(ρ) X
(4.1) ∆(H) ϕλ (exp H) = 2−|Σ | (det w)eiwλ(H)
π(iλ)
w∈W
where ∆ is given by (2.5), W is the Weyl group, |Σ+ | denotes the cardinality
of Σ+ , and π is the polynomial function on a∗c defined by
Y
(4.2) π(ν) := hα, νi.
α∈Σ+
174 ANGELA PASQUALE
Formula (4.1), due to Harish-Chandra, relates the spherical transform of

a function f ∈ S(K\G/K) to the Fourier transform of the function H 7−→
∆(H)f (exp H) on a. Indeed, since
∆(wH) = (det w)∆(H)
for all H ∈ a and w ∈ W , we obtain from Formula (1.5): Up to a constant

multiple, for all λ ∈ (a∗ )0 := (a∗c )0 ∩ a∗ ,
Z
Sf (λ) = f (exp H)ϕ−λ (exp H)[∆(H)]2 dH
a+
Z
1
= ∆(H)f (exp H)∆(H)ϕ−λ (exp H) dH
|W |
a
+
2−|Σ |
Z
π(ρ) X
= (det w)∆(H)f (exp H)e−iwλ(H) dH
|W | π(−iλ)
w∈W a
Z
+ π(ρ)
= 2−|Σ | ∆(H)f (exp H)e−iλ(H) dH.
π(−iλ)
a
Therefore, up to a constant depending only on |Σ+ | and on the normalization

of the measures,
(4.3) π(−iλ)Sf (λ) = F(∆(f ◦ exp))(λ)
for all λ ∈ (a∗ )0 . By continuity, (4.3) holds for all λ ∈ a∗ .
Proof of Theorem 2.3. Since π(−iλ) is a polynomial in λ, Formula (4.3) to-

gether with the classical Paley–Wiener theorem prove that Sf is compactly
supported, with supp Sf ⊂ Br , if and only if ∆(H)f (exp H), H ∈ a, extends
to an entire function F̃ on ac which is rapidly decreasing and of exponential
F̃ (H̃)
type r. Set F (H̃) := ∆( H̃)
, H̃ ∈ ac . Then F is a meromorphic function on
C with singularities at most on the set
[
{H̃ ∈ ac : ∆(H̃) = 0} = {H̃ ∈ ac : α(<H̃) = 0 and α(=H̃) ∈ iZ}.
α∈Σ+
Since F̃ agrees with ∆(H)f (exp H) on a, F extends to be holomorphic on

the set {H ∈ a : α(H) = 0 for some α ∈ Σ+ } by setting F (H) := f (exp H).
So F extends f (exp H), H ∈ a. In particular, F must be W -invariant.
We conclude this section with a remark on the Abel transform. For com-
plex groups an explicit formula for the inverse Abel transform is available.
As proved by Gangolli7 , if f ∈ S(K\G/K), then, up to a constant multiple,

Y 1
(4.4) f (exp H) = ∂(Hα ) Af (exp H)
+
sinh α(H)
α∈Σ
where Hα ∈ a is uniquely determined by the condition α(H) = hHα , Hi for

all H ∈ a, and ∂(Hα ) is the corresponding differential operator on a.
Theorem 1.2 can be also proved using Formula (4.4) to characterize the
functions f whose Abel transform extends to a rapidly decreasing entire
function of exponential type r. However, doing so, we would not free our-
selves from the use of the explicit espression for the elementary spherical
functions. In fact, the only known general procedure to get (4.4) is to
take the inverse Fourier transform of both sides of (4.3), using the prop-
erty S = F ◦ A. Note that this is not the case for the rank-one groups.
In fact, the explicit formulas for A and A−1 we used has been determined
by Rouvière (and others) directly, without assuming any knowledge of the
elementary spherical functions on the group.
References
[Bee88] R.J. Beerends, The Abel transform and shift operators, Comp. Math., 66 (1988),
145-197.
[Bra96] W.O. Bray, Generalized spectral projections on symmetric spaces of noncompact
type: Paley–Wiener theorems, J. Funct. Anal., 135 (1996), 206-232.
[Dwi61] H.B. Dwight, Tables of integrals and other mathematical data, Macmillan, New
York, 1961.
[Eat73] T.R. Eaton, Asymptotic behaviour of spectra of compact quotients of symmetric
spaces of rank one of the noncompact type, University of Washington, Seattle,
1973.
[Gan68] R. Gangolli, Asymptotic behaviour of spectra of compact quotients of certain sym-
metric spaces, Acta Math., 12 (1968), 151-192.
[Gan71] , On the Plancharel formula and the Paley–Wiener theorem for spherical
functions on semisimple Lie groups, Ann. of Math., 93 (1971), 150-165.
[GV88] R. Gangolli and V.S. Varadarajan, Harmonic Analysis of Spherical Functions on
Real Reductive Groups, Springer-Verlag, 1988.
[HC58] Harish-Chandra, Spherical functions on a semisimple Lie group, II, Amer. J. of
Math., 80 (1958), 553-613; Collected papers, Vol. II, 479-539, Springer-Verlag,
1983.
[Hel66] S. Helgason, An analogue of the Paley–Wiener theorem for the Fourier transform
on certain symmetric spaces, Math. Annalen, 165 (1966), 297-308.
[Hel84] , Groups and Geometrical Analysis, Academic Press, Orlando, 1984.
[Koo75] R. Koornwinder, A new proof of a Paley–Wiener Theorem for the Jacobi trans-
form, Ark. Mat., 13 (1975), 145-159.
7
[Gan68], p. 164. See also [Rou83], Thèoréme 5, p. 287.
176 ANGELA PASQUALE
[Lan93] S. Lang, Complex Analysis, Springer-Verlag, third edition, 1993.

[LR70] N. Levinson and R. Redheffer, Complex variables, Holden-Day, San Francisco,
1970.
[LR82] N. Lohoué and T. Rychener, Die Resolvente von ∆ auf symmetrischen Räumen
vom nichtkompakten Typ, Comment. Math. Helvetici, 57 (1982), 445-468.
[Nar73] R. Narasimhan, Analysis on real and complex manifolds, Masson, Paris; American
Elsevier Pub. Co., New York, 1973.
[Pas96] A. Pasquale, A Paley–Wiener Theroem for the Inverse Spherical Transform on
certain Symmetric Spaces, University of Washington, Seattle, 1996.
[Rou83] F. Rouvière, Sur la transformation d’Abel des groupes de Lie semisimples de rang
un, Ann. Scuola Norm. Sup. Pisa, 10 (1983), 263-290.
[Tak63] R. Takahashi, Sur les représentations unitaires des groupes de Lorentz générelisés,
Bull. Soc. Math. France, 91 (1963), 289-433.
Received July 15, 1998.
TU-Clausthal
Erzstrasse 1
38678 Clausthal-Zellerfeld
Germany
E-mail address: mapa@math.tu-clausthal.de
Vol. 193, No. 1, 2000
ON THE ACTION OF THE GROUP OF

DIFFEOMORPHISMS OF A SURFACE ON SECTIONS OF
THE DETERMINANT LINE BUNDLE
Doug Pickrell
Let Σ denote a closed oriented surface. There is a nat-

ural action of the group Diff + (Σ) on sections of the chiral
determinant line over the space of gauge equivalence classes
of connections. The question we address is whether this action
is unitarizable. We introduce a SDif f -equivariant regulariza-
tion, and we prove the existence of, and explicitly compute,
the limit as the regularization is removed. The SDif f uni-
tary representations that arise, both by regularization and
after removing the regularization, appear to be new.
0. Introduction.
Let Σ denote a closed oriented surface, and let D denote the group of
orientation-preserving diffeomorphisms of Σ. Let K denote a connected
compact Lie group, A the space of K-connections in the trivial bundle
P = (Σ × K → Σ), and C the space of gauge equivalence classes of K-
connections. It is well-known that the orientation of Σ induces a Aut(P )-
invariant symplectic structure on A. Ideally we would want to consider the
action of D in the metaplectic representation, corresponding to a polariza-
tion coming from a choice of complex structure for Σ, but it is easy to check
that D is not implemented in this representation (see §1); instead D per-
mutes these representations. Roughly speaking, our goal is to understand
whether the gauge invariant sectors of these representations (the conformal
blocks of conformal field theory) can be coherently incorporated into a single
unitary representation for D.
To be more precise, there exists a nontrivial D-equivariant Hermitian line
bundle L → C (the projection to C of the prequantum line bundle on A).
The issue we address is whether, in the simplest case K = T, a submodule
of the natural action
(0.1) D × Ω0 (L) → Ω0 (L)
can be unitarized. Heuristically the invariant inner product is given by
Z
s1 , s2 → hs1 , s2 idV,
C
177
178 DOUG PICKRELL
where “dV ” is a fictitious D-invariant measure on C arising from the D-

invariant symplectic structure on A.
The submodule arises in the following way. Fix a complex structure on
Σ. We can then identify L with a determinant line bundle for ∂¯ coupled
to gauge potentials. There is a canonical section det which is fixed by the
subgroup Aut(Σ) ⊂ D. The submodule is the spherical representation for
the pair (D, Aut(Σ)) generated by det.
To obtain a unitary structure we choose a positive area form on Σ, and
introduce a regularization, in this case the Yang-Mills measure on (a com-
pletion of) C,
1 −1 R
hFA ∧∗FA i
(0.2) dνT ([A]) = e 2T dV ([A]).
Z(T )
For each T < ∞, we rigorously obtain a SDiff-invariant unitary structure
Z
|L|−2
(0.3) s1 , s2 → hs1 , s2 iT = s1 s̄2 dνT ,
|L|−2
where dνT denotes the renormalized coupling of the Yang-Mills measure
and Hermitian structure of L. In the genus = 0 case, the spherical function
(0.4) φT : D/Aut(Σ) → C : σ → hσ · det, detiT
is given by the formula
(0.5) φT (σ) = det2 (1 + T (∂ ∂¯ − ∂µ∂)−1 ∗)−1/2 ,
where µ is the complex dilatation of σ −1 (see §1), and det2 denotes the
Hilbert-Schmidt regularized determinant. Although φT is defined on all of
D, only its restriction to SDiff is positive definite.
The main result of this paper is that one can remove the regularization.
After renormalizing φT , one can take the limit T → ∞ to obtain (in the
genus = 0 case) a positive definite function
detζ (4∆(1 − ∂¯−1 µ∂)−1 )−1/2
(0.6) φ : SDiff → C : σ →
detζ (4∆)−1/2
(the precise meaning of this formula is explained in §3). One can heuristically
arrive at this formula for φ by using ζ-function determinants to directly
“evaluate” certain integrals. On the other hand it is not at all transparent
that |φ| is bounded by 1 on SDiff, let alone that φ is a positive definite
function on SDiff; this follows by first evaluating the regularized integrals
and then taking the limit T → ∞. From a purely technical point of view,
one of the most interesting aspects of this work is the following: There are
two regularizations involved in arriving at (0.6), one probabilistic and one of
the ζ variety, and it is fascinating to see how they balance out (see Remarks
(2.28) and (3.30)).
DIFFEOMORPHISMS ACTING ON THE DETERMINANT LINE BUNDLE 179
To go from (0.5) to (0.6), we establish a result of independent interest

concerning the multiplicative anomaly for ζ-function determinants. If E is a
classical, elliptic, invertible pseudo-differential operator on an n dimensional
compact manifold, with symbol satisfying the Agmon-Niremberg condition,
and if C has order < −n, so that it represents a trace class operator, then
it is known that
(0.7) detζ (Eeτ C ) = detζ (E)eτ tr(C) ,
as one would naively suspect ([KV]). We consider the borderline case in
which C has order −n, in which case the trace is replaced by a regularized
trace, depending upon E; see Lemma (3.10) for the precise statement.
The formula (0.6) defines an extension of φ to all of D, but this extension
cannot possibly be positive definite; for (0.6) defines a holomorphic function
of µ, hence cannot possibly be bounded by 1. This strongly suggests that the
action (0.1) is not unitarizable, so that the answer to our original question
is negative. We should mention that this question cannot be decided by
considering the Lie algebra action, for det cannot possibly be a differentiable
vector in a unitary representation for D (see §1).
The theory of unitary representations for various types of diffeomorphism
groups is basically at the stage of searching for interesting examples; we re-
fer to [I] for an account of this. It appears that the unitary representations
of SDiff which we have constructed, both the regularized representations
and their limit, are new (but, as pointed out by the referee, it is difficult to
judge this from the spherical function alone). Because we assume K = T is
abelian, the Yang-Mills measure (0.2) is essentially Gaussian, so that it is a
fairly mundane object. However, we should point out that in 2 dimensions,
the Yang-Mills measure is well-defined mathematically for any compact Lie
group K (see [Sen] or [Pi] and references there), so that the corresponding
representations should exist in this more general (and technically challeng-
ing) setting. In terms of mathematical physics, the regularized representa-
tions amount to a geometric formulation of the coupling of Yang-Mills fields
to massless fermions (see [Pi]). The unitary structure that should arise when
the regularization is removed should restrict to a canonical scalar product
on the spaces of conformal blocks in conformal field theory (or equivalently
the space of states for Chern-Simons theory), the existence of which has
been proven in some nonabelian cases by Gawedzki (see [G] and references
there). As we mentioned previously, understanding whether these blocks fit
together coherently to give a unitary representation of D was the original
motivation for this work.
The outline of the paper is as follows. In §1 we briefly review what we
need to know from the theory of quasiconformal mappings, and the relation
of this to the symplectic action of D on the space of all connections. In
§2 and §3 we deal with the regularized and limiting representations in the
180 DOUG PICKRELL
genus = 0 case. In the last section we indicate the relatively minor changes
necessary to deal with the cases of positive genus.
Acknowledgement. I thank Lennie Friedlander and John Palmer for use-

ful discussions on determinants. I also thank the referee for useful comments.
1. The Quotient D/Aut(Σ).

Given a closed oriented surface Σ, there is a natural symplectic structure ω
on the real vector space Ω1 (Σ) of one-forms on Σ:
Z
(1.1) ω(θ, η) = θ ∧ η.
Σ
The choice of a complex structure j on Σ induces a positive polarization:

(1.2) (Ω1 )C = Ω0,1 ⊕ Ω1,0 .
Let σ denote an orientation-preserving diffeomorphism of Σ. Following stan-
dard practice, with respect to the decomposition (1.2), we write

1 −1 ∗ A B
(1.3) D → Sp(Ω , ω) : σ → (σ ) = ,
B̄ Ā
where we have extended (σ −1 )∗ by complex linearity to (Ω1 )C .

Given a general symplectic space (V, ω), it does not make sense to speak
of a bounded symplectic transformation. However, given a positive polar-
ization,
(1.4) V C = V + ⊕ V −,
where V + is a complex Hilbert space with respect to the form iω(θ, η̄), and
V − is the conjugate Hilbert space, we can define
Sp(L) (V + ⊕ V − , ω)

A B
(1.5) = g = ∈ Sp(V, ω) : B ∈ L
B̄ Ā

A B + − ∗ ∗ ∗ ∗
(1.6) = ∈ L(V ⊕ V ) : A A − B̄ B̄ = 1, A B − B̄ Ā = 0 ,
B̄ Ā
where “L” stands for bounded operators, B̄ is the conjugate of B, and the
adjoint is computed with respect to the Hilbert space structure of V + ⊕
V − . We then have a fibration over a (non-Riemannian) symmetric bounded
domain
(1.7) 0 → U (V + ) → Sp(L) → {Z ∈ L(V + , V − ) : Z = Z̄ ∗ , Z ∗ Z < 1} → 0,

A
where A → , g → Z,
Ā

A B 1 A 1 W
(1.8) g= = .
B̄ Ā Z 1 Ā − ZAW 1
It is more common to consider proper operator ideals, especially Hilbert-
Schmidt operators in representation-theoretic contexts, in place of L above,
but our point will be that L is natural in our context.
To apply this to our context, let V denote the completion of Ω1 with
respect to our chosen complex structure on Σ,
Z
1
(1.9) V = θ ∈ Ωmeas : θ ∧ ∗θ < ∞ .
The form ω extends continuously to V , and we let V + denote the completion

of Ω0,1 .
Proposition 1.10. The map (1.3) extends to a representation of the group
of quasi-conformal homeomorphisms
Dqc → Sp(L) (V + ⊕ V − , ω).
In the genus= 0 case, the fibration (1.7) induces an exact sequence
0 → Aut(Σ) → Dqc → Ω0L∞
<1
(T ∗0,1 ⊗ T 1,0 ) → 0,
where σ → µ, the complex dilatation of σ −1 , and Z̄(σ) = µ = −W (σ), where

µ is viewed as a multiplication operator (see Remark (2) below); in general
we have an exact sequence
0 → Aut(Σ) → Dqc → Ω0L∞
<1
(T ∗0,1 ⊗ T 1,0 ) → M(Σ) → 0,
where M(Σ) is the moduli space of Riemann surface structures for Σ, with
basepoint j.
Remarks 1.11. (1) Note that there is a Aut(Σ)-equivariant mapping
(1.12) Ω0 (T ∗0,1 ⊗ T 1,0 ) → Ω0 : µ → µ ⊗ µ̄,
so that there is a Aut(Σ)-invariant L∞ structure on Ω0 (T ∗0,1 ⊗ T 1,0 ). In

(1.10) “L∞<1 ” denotes the unit ball with respect to this Banach structure.
(2) Given µ ∈ Ω0L∞ (T ∗0,1 ⊗ T 1,0 ), there is an associated bounded “multi-
plication operator”
(1.13) µ : V − → V + : θ → µ ⊗ θ.
∂
Using the coordinate expressions, θ = f dz, µ = µ̃dz̄ ⊗ ∂z , and µ ⊗ θ = µ̃f dz̄,
it is obvious that |µ|L = |µ|L∞ .
182 DOUG PICKRELL
(3) A homeomorphism σ is quasi-conformal provided (1) in holomorphic

coordinates σ has locally square-integrable partial derivatives, and (2) with
respect to the decomposition T C = T 1,0 ⊕ T 0,1 , if
¯

C ∂σ ∂σ
(1.14) (dσ) = ¯ ∗ (∂σ)∗ ,
(∂σ)
then the complex dilatation
¯
∂σ
(1.15) µσ = ∈ Ω0 (T ∗0,1 ⊗ σ ∗ T 1,0 ⊗ (T ∗1,0 ⊗ σ ∗ T 1,0 )−1 ) = Ω0,1 (T 1,0 )
∂σ
satisfies |µσ |L∞ < 1.
Proof of (1.10). Suppose σ is quasi-conformal. Then σ −1 is quasi-conformal,
and it is straightforward to show that
W (σ) = −µ, Z(σ) = µ̄.
For example to see that W (σ) = −µ, suppose that θ ∈ Ω1,0 . Since W =
A−1 B, we must show that
(1.16) Bθ = −Aµθ, i.e., (σ ∗ θ)0,1 = −(σ ∗ (µθ))0,1 .
In local coordinates w = σ(z), θ = f dw, µ = µσ−1 = zzww̄ dw̄ ⊗ ∂
∂w , and

zw zw̄ wz wz̄ 1
(1.17) ◦σ = .
z̄w z̄w̄ w̄z w̄z̄ 1
The LHS of (1.16) equals f (σ(z))wz̄ dz̄, and the RHS of (1.16) equals

zw̄
(1.18) − ◦ σ(z)w̄z̄ f (σ(z))dz̄.
zw
Thus (1.16) follows from the (1, 2) entry of (1.17).
Thus σ quasi-conformal implies Z = µ̄ is bounded and |Z|L < 1. Since
AA∗ = (1−ZZ ∗ )−1 , A and B are also bounded, and it follows that σ ∈ Sp(L) .
Now suppose that we are given µ with |µ|L∞ < 1. The basic fact in the
theory of quasi-conformal mappings is that there exists a (possibly different)
complex structure j 0 on Σ and a quasi-conformal mapping σ : (Σ, j 0 ) →
(Σ, j) such that µ = µσ−1 (see Theorem 1.1 of Chapter V of [L]). In the
genus = 0, case j 0 is equivalent to j, and we obtain the first exact sequence
in (1.10); in general we obtain the second sequence in (1.10), where µ maps
to the equivalence class of j 0 . This completes the proof.
In the introduction we mentioned that the representations we consider
do not induce representations of Lie algebras. One way to see this, in the
genus = 0 case, is as follows.
Proposition 1.19. Suppose that Σ = Ĉ, the Riemann sphere. Suppose that
π : D → U (H)
is a unitary representation, and suppose also that there exists a smooth vector
v ∈ H which is fixed by Aut(Σ) = P SL(2, C). Then v is fixed by all of D.
Proof. Given that v is fixed by Aut(Σ), there is a mapping
(1.20) D/Aut(Σ) → P(H) : [g] → P(g · v).
Since v is a smooth vector, this is a differentiable mapping, and hence we
can pull the Fubini-Study metric back to obtain a D-invariant (possibly
incomplete) Riemannian structure on D/Aut(Σ). But by considering the
isotropy action of Aut(Σ) at the base point, this means that we have a
Aut(Σ)-invariant unitary structure for the natural action
(1.21) Aut(Σ) × Ω0,1 (T 1,0 ).
But it is known that this particular representation for P SL(2, C) is not uni-
tarizable (see e.g. Chapter 1 of [K]). Since sl(2, C) is a maximal closed
subalgebra of vect(Σ), it follows that D must act trivially on v. This com-
pletes the proof.
2. SDiff(S 2 ) unitary representations.

Throughout this section and the next, Σ will denote a closed oriented surface
of genus = 0.
Let A denote the set of continuous T-connections in the trivial bundle
Σ × T → Σ, and let K denote the C 1 gauge group. Since genus(Σ) = 0, K
is connected, hence K = exp(Ω0C 1 (Σ, iR)). Define
Z
0
(2.1) Ω (L) = f : A → C : f (A + dξ) = exp i A ∧ dξ f (A),

∀A ∈ A, ξ ∈ Ω0C 1 (Σ, iR) .
This is the space of sections of a line bundle L → A/K, hence the notation. It
is easy to check that the natural action by pullback leads to a representation
(2.2) D × Ω0 (L) → Ω0 (L).
To unitarize this representation, it would suffice to have a D-invariant
measure on the base space A/K, since L has a natural Hermitian structure.
But such a measure does not exist. The point of this subsection is that this
naive idea can be made to work, provided that we choose an area form on
Σ, and restrict the representation to SDiff, the area-preserving diffeomor-
phisms.
So fix a smooth area form on Σ. We can then consider the Yang-Mills
measure, corresponding to temperature T , which is given heuristically by
184 DOUG PICKRELL
the expression
Z
1 1
(2.3) dνT −1 Y M (A) = exp F ∧ ∗F DA,
Z 2T
where F = dA is curvature. In this abelian context it is well-known that
one can interpret the projection of the heuristic expression (2.3) to A/K as
a Gaussian measure, and this Gaussian is SDiff-invariant. By definition the
“projection” is the Gaussian measure νT corresponding to the real Hilbert
space structure
Z
1 1
(2.4) F ·T F = F · F = − F ∧ ∗F,
T T
R
where F is an iR-valued two-form with F = 0. Since SDiff acts orthogo-
nally on this Hilbert space, νT is an SDiff-invariant measure. The question
that we must address is whether we can use this measure to integrate func-
tions of the form s1 s̄2 , for s1 , s2 ∈ Ω0 (L).
The most interesting sections are obtained in the following way. Choose
a complex structure for Σ. We then have an isomorphism
(2.5) A → Ω0,1
C0
: A → a = A0,1 , A = a − a∗ .
It is easy to check that the function
Z
∗
(2.6) det : A → C : A → exp i (a ∧ (a + Ha))
is in Ω0 (L), where (because genus(Σ) = 0) H is the well-defined unitary

0th -order operator
0,1 1,0
(2.7) H = ∂ ∂¯−1 : ΩL 2 → ΩL2 .
H is often referred to as either the Hilbert or Beurling transform. The

rationale for the notation is that, relative to our choice of complex structure,
L can be identified with the determinant line bundle with Quillen metric for
the family of Fredholm operators ∂¯κ1/2 + a, a ∈ Ω0,1 , and the canonical
section corresponds to the function (2.6).
In order to have a better grasp of the section det, it is useful to consider
the parameterization of A
(2.8) ¯
a = ∂x, x ∈ Ω0C 1 (Σ, C)/C.
A gauge transformation A → A + dξ corresponds to a → a + ∂ξ,¯ i.e., adding
a iR-valued function to x. Hence we can parameterize the space A/K by
x ∈ Ω0C 1 (Σ, R). The relation between (R-valued) x and the curvature F is
simply
(2.9) ¯
F = i ∗ ∆x = 2∂ ∂x.
In terms of x the projection of the Yang-Mills measure with temperature

T to A/K is the Gaussian measure corresponding to the real Hilbert space
structure
Z Z
1 1
(2.10) x ·T x = − F ∧ ∗F = ∗∆x ∧ ∆x.
T T
From this we can spot the basic source of difficulty: This Gaussian measure
is not supported on x ∈ C 1 ; rather, it is only supported on x ∈ C 0 ∩
W 1− , for > 0. Hence we cannot integrate functions of the form s1 s̄2 in
a straightforward manner. Essentially what we must do is show that the
bundle L can be SDiff-equivariantly extended to the support of νT .
In terms of (C-valued) x, the section det is given by the expression
Z Z
(2.11) det|x = exp i ∂x ¯ ∧ 2∂(Re(x)) = exp i x ∧ F ,
where the first equality makes sense if x ∈ C 1 , the second if x ∈ C 2 . Given

σ ∈ D we obtain a transition function

σ · det
(2.12) cA/K (σ, x) =
det x
for the line bundle L over A/K, where we are viewing (R-valued) x ∈ C 1 /R
as a global coordinate. We can also interpret cA/K as a cocycle, i.e., it
satisfies the equation
(2.13) cA/K (σ1 ◦ σ2 ) = cA/K (σ1 )σ1 · cA/K (σ2 ),
for σ1 , σ2 ∈ D.
To understand why cA/K extends to the support of νT , and why the ex-
tension is an SDiff-cocycle, we must have an explicit formula for it. Let j
denote the original complex structure, and let j 0 = σ · j denote the trans-
formed complex structure; we do the same for the corresponding D-bar
operators, ∗-operators, and so on.
Lemma 2.14. Suppose that σ ∈ D, and let µ = µσ−1 .
(a) (∂¯0 )0,1 = (1 − µµ̄)−1 (∂¯ − µ∂).
(b) (∂¯0 )1,0 = µ̄(∂¯0 )0,1 .
(c) a0 = (1 + µ̄)(1 − µ conj)−1 a, where conj denotes conjugation.
(d) x0 = (∂¯ − µ∂)−1 (∂¯ + µ∂)x, for R-valued x; x0 = x, for iR-valued x.
¯ 1+∂¯¯−1
(e) ∂ 0 ∂¯0 = 2∂ ∂(
−1 µ∂ −1 ¯
+ 1+∂ µ̄∂ )−1 .
1−∂ µ∂ R1−∂ −1 µ̄∂¯
(f) σ · det ( det)∗ = exp(i (1− ∂¯−1 µ∂)−1 x∧F ) = exp(−S(µ)F ·F ). where
S(µ) = i ∗ (∂(∂¯ − µ∂))−1 .
(g) cA/K (σ, F ) = exp(−(S(µ) − S(0))F · F ).
(h) The real and imaginary parts of S(µ) are symmetric, with respect to
the real Hilbert space structure (2.4).
186 DOUG PICKRELL
(i) Re(S(µ)) = i ∗ (∂ 0 ∂¯0 )−1 = 2 ∗ ∗0 ∆0−1 (the Laplacian on 2-forms); hence

if σ ∈ SDiff, ReS(µ) = 2∆0−1 .
Proof of (2.14). In j-holomorphic coordinates suppose that w = σ(z), f =
f (w). Then
(2.15) ∂¯0 f = σ −1∗ (∂¯z {f ◦ σ})
w
= (fw̄ w̄z̄ ◦ σ −1 + fw wz̄ ◦ σ −1 )dz̄
= (fw̄ − µfw )w̄z̄ ◦ σ −1 z̄w̄ dw̄ + (fw̄ − µfw )w̄z̄ ◦ σ −1 z̄w dw.
Using (1.17), we see that

−1 zw 1
(2.16) w̄z̄ ◦ σ z̄w̄ = z̄w̄ = .
zw z̄w̄ − zw̄ z̄w 1 − µµ̄
Similarly w̄z̄ ◦ σ −1 z̄w = µ̄(1 − µµ̄)−1 . This proves (a) and (b).
We have ∂x ¯ = a, ∂¯0 x0 = a0 , where
A = a − a∗ = a0 − a0∗ .
This implies
(2.17) a = (∂¯0 x0 )0,1 − conj((∂¯0 x0 )1,0 )
= (1 − µµ̄)−1 (1 − µconj)(∂¯ − µ∂)x0
= (1 + µconj)−1 (∂¯ − µ∂)x0 .
Hence
(2.18) a0 = ∂¯0 {(∂¯ − µ∂)−1 (1 + µconj)a} = (1 + µ̄)(1 − µconj)−1 a,
which proves (c), and also
(2.19) (∂¯ − µ∂)x0 = (1 + µconj)∂x,
¯
which implies (d).

We have
1 + ∂¯−1 µ∂
(2.20) (∂¯ − µ∂)−1 (∂¯ + µ∂) = .
1 − ∂¯−1 µ∂
For R-valued x
(2.21) ¯ = F = 2∂ 0 ∂¯0 Re(x0 ),
2∂ ∂x
hence (e) follows from (2.20) and (d).
For R-valued x,
2
R
(x0 +x)∧F i ¯−1 µ∂ x∧F
R
(2.22) σ · det( det)∗ = ei =e 1−∂ ,
which is (f). Part (g) follows immediately from (f).
In part (h) it suffices to show that (∂ ∂¯ − ∂µ∂)∗ is symmetric, and this is
immediate from integration by parts.
We have S(0) = 2∆ (acting on 2-forms), and by (f) | det| = exp(−2∆F ·F ).

Together with (g) we now see that
(2.23) |σ · det| = exp(−2∆0 F ·0 F )
= exp(−2 ∗ ∗0 ∆0 F · F ) = exp(−ReS(µ)F · F ).
This proves (i).
Given a real Hilbert space H and a symmetric Hilbert-Schmidt operator
S on H (possibly with values in HC ), one can interpret “Sv · v − trS” as a
random variable with respect to the Gaussian measure dν associated to H.
More precisely, if {n } is an orthonormal basis for H (hence also HC ), and
Pn the orthogonal projection onto the C-span of the first n basis elements,
then the regularization of the symmetric form defined by S is the L2 (dν)
limit
(2.24) Sreg = lim (Pn Sx · x − tr(Pn S)).
n→∞
This definition is independent of the choice of orthonormal basis. Moreover
Z
(2.25) e−Sreg dν = det((1 + 2S)e−2S )−1/2 ,
and if S 0 is also Hilbert-Schmidt and symmetric, and O ∈ O(H), then

(S + S 0 )reg = Sreg + Sreg
0
,
(2.26) O · (Sreg ) = (OSO−1 )reg

(see Section 4 of [R]).
In our context the Hilbert-Schmidt operator is ∆−1 , or some 0th -order
perturbation. (Note that ∆−1 , while not trace class, is in the Dixmier trace
class L+
1 , so that the regularization in our case is necessary, but very mild.)
Definition 2.27. Given σ ∈ D,

cT (σ) = exp(−T (S(µ) − S(0))reg )
(a random variable with respect to νT ).
Remarks 2.28. (a) The appearance of T in the above definition is poten-
tially confusing; it appears because we are regularizing with respect to νT ,
the Gaussian corresponding to the inner product T1 F · F .
(b) In the next section we will see that (2.27) is fine for σ ∈ SDiff, but
needs adjustment for general µ; see Remark (3.30) below, expecially (3.32).
The point is that one cannot indiscrimately subtract a trace!
Proposition 2.29. cT is an SDiff-cocycle with respect to νT , i.e.,
cT (σ1 ◦ σ2 ) = cT (σ1 )σ1 · cT (σ2 )
as random variables with respect to νT , ∀σ1 , σ2 ∈ SDiff.
188 DOUG PICKRELL
Proof. This follows immediately from (2.13) and (2.26), since SDiff acts
orthogonally with respect to νT .
Although it is somewhat artificial in our current context, we can now view
L as a measureable line bundle over the support of νT , using the transition
functions cT (σ), σ ∈ SDiff.
|L|−2
Definition 2.30. We define a measure dνT having values in the line
bundle |L|−2 by
|L|−2 s1 s2 ∗ −T S(0)reg
s1 s̄2 dνT = e dνT ,
det det
where s1 , s2 are sections of L over the support of νT .
|L|−2
dσ∗ | det|2 dνT
Proposition 2.31. (a) |L|−2
= |cT (σ)|2 .
d| det|2 dνT
|L|−2
(b) dνT is SDiff-invariant.
Proof of (2.31). Part (a) follows immediately from the definition of the mea-
sure
|L|−2
(2.32) | det|2 dνT = e−T S(0)reg dνT
and the fact that σ∗ (S(0)reg ) = Re(S(µ)reg ), by (i) of (2.14).
Part (b) is a restatement of (a).
We can now define a SDiff-invariant unitary structure on sections of L
over the support of νT by
Z
|L|−2
(2.33) hs1 , s2 i = s1 s̄2 dνT .
We let πT denote the unitary representation of SDiff acting on sections

of L over the support of νT . We can equivalently view πT as the cocycle
representation
|L|−2

2 2
(2.34) SDiff × L | det| dνT
given by σ · f = (f ◦ σ −1 )cT (σ) (where the function f corresponds to the

section s = f det).
Proposition 2.35. The spherical function for the representation πT of
SDiff corresponding to the vector det is given by
Z
|L|−2
hπT (σ) · det, deti = σ · det( det)∗ dνT
Z −1/2
= e−T S(µ)reg dνT = det(1 + 2T S(µ))e−2T S(µ) .
This follows from (2.25).

Remarks 2.36. (a) There is a simple explanation for why the vector det
is not a smooth vector for the unitary action of SDiff. Given v ∈ Vect (Σ),
d
v · det|A = etv · det|A = det|A (d(log det)|A (−Lv A))
dt t=0 Z
¯−1 ¯−1 0,1
= i det|A (iv {∂ F } − ∂ {(iv F ) }) ∧ F .
The operator
F → ∗(iv {∂¯−1 F } − ∂¯−1 {(iv F )0,1 })
has order = −1, hence generally it is not Hilbert-Schmidt, and it is not
possible to regularize the corresponding quadratic form.
(b) The partition function of abelian Yang-Mills (at temperature T ) cou-
pled to fermions (with coupling constant 1) is the integral
Z
|L|−2 −1
Z(T ) = | det|2 dνT = ( det(1 + 4T ∆−1 )e−4T ∆ )−1/2 .
An interesting question is whether one can compute the corresponding quan-

tity in the nonabelian case.
3. The limit T → ∞.
As in the preceding section, Σ denotes a genus = 0 oriented surface, and we
consider the representation (2.2). We fix a complex structure, so that we
have a distinguished section det. We also fix an area form, so that we can
use the Yang-Mills construction of the preceding section. Our objective is
to prove that the SDiff-submodule generated by the section det is unitary
with respect to the Hermitian inner product
Z
1 |L|−2
(3.1) hσ1 · det, σ2 · deti = lim cT (σ2−1 σ1 )| det|2 dνT ,
T →∞ Z(T )
where Z(T ) is the partition function in (b) of (2.36).

Proposition 3.2. Suppose that σ ∈ SDiff, µ = µσ−1 . Then
det((1 + 2T S(µ))e−2T S(µ) )−1/2 detζ (4∆)1/2
lim = ,
T →∞ det((1 + 2T S(0))e−2T S(0) )−1/2 detζ (4∆(1 − ∂¯−1 µ∂)−1 )1/2
and this defines a positive definite function on SDiff.
To explain the meaning of the RHS of (3.2), and before undertaking the
proof, we need some preparatory remarks concerning determinants (we will
use [Sh] and [KV] as basic references). Suppose that E is an invertible
elliptic classical pseudo-differential operator of positive real order d, on a
compact manifold of dimension n, and suppose that its principal symbol
σd satisfies the Agmon-Nirenberg condition, i.e., there is a ray Lθ extending
from the origin in the complex plane such that at each point of the manifold,
190 DOUG PICKRELL
the spectrum of σd intersects trivially with Lθ . In our context E will always

have the property that we can choose Lθ to be the negative real axis, as in
§10 of [Sh]. In this situation one can form the complex powers E −s , s ∈ C,
and the ζ-function
(3.3) ζE (s) = tr(E −s ), Re(s) > n/d,
has a meromorphic extension to all of C which is regular in a neighborhood of
s = 0 (§10 and §13 of [Sh], respectively). We then define the ζ determinant
in the usual way, detζ E = exp(−ζ 0 (0)). In particular this is the meaning of
the determinants on the RHS of (3.2), where we have thrown out the zero
eigenvalue (note that the symbol of ∆(1− ∂¯−1 µ∂)−1 is ρ|ξ|2 (1− µ̃(z)ξ¯−1 ξ)−1 ,
where µ̃ and ρ are local representations for µ and the area form, respectively;
since |µ̃| < 1, the negative real axis is a spectral cut).
Now initially suppose also that C is a pseudo-differential operator of order
< −n, so that C represents a trace class operator. In this case we have
detζ Eeτ C = ( detζ E) eτ trC .

(3.4)
To see this, note that (Eeτ C )−s C is a trace class operator for all s ∈ C,
hence the trace of this holomorphic family is an entire function of s. Also
∂ ∂
(3.5) ζ τ C (s) = tr(Eeτ C )−s = −str((Eeτ C )−s C), Re(s) > n/d.
∂τ Ee ∂τ
By analytic continuation this equality is valid in all of C, in particular in a
neighborhood of s = 0. Hence
∂ 0
(3.6) ζ τ C (0) = −trC,
∂τ Ee
implying that
∂
(3.7) detζ (Eeτ C ) = detζ (Eeτ C )trC,
∂τ
which integrates to (3.4).
Now suppose that order (C) ≤ −n. In this case C ∈ L+ 1 (the dual of the
τ C −s
Macaev ideal; see [C]), and (Ee ) C is trace class for Re(s) > 0. The
trace of this family has a meromorphic extension to the complex plane, and
in a neighborhood of s = 0, we have
Res(C)
(3.8) tr((Eeτ C )−s C) = + h(s),
s
where Res (C) is the noncommutative residue and h(s) is holomorphic. The
value of h at s = 0 is called the finite part of the trace; we will write “FPtr”
for the finite part:
(3.9) FPtr((Eeτ C )−s C) = h(0).
Note that if order (C) < −n, then Res(C) = 0 and h(0) = trC.
Lemma 3.10. Suppose that E is classical, elliptic, invertible and satisfies

the Agmon-Nirenberg condition as above, and C has order ≤ −n. Then
−s
(a) detζ (Eeτ C ) = detζ (E) eτ F Ptr(E C) .
(b) The linear functional C → FPtr(E −s C) depends only upon the prin-
cipal symbol of E.
Proof of (3.10). We have
(3.11) FPtr(Eeτ C )−s C = FPtr((Eeτ C )−s C − E −s e−sτ C C)
+ FPtr(E −s e−sτ C C − E −s C) + FPtr(E −s C).
We must show that the first two terms on the RHS are zero. As s → 0,
(3.12)
(Eeτ C )−s C − E −s e−sτ C C = −s(log(Eeτ C ) − log E − τ C)C + O(s2 )
(3.13) E −s e−sτ C C − E −s C = −sτ C 2 + O(s2 )
where we have used a spectral cut near Lθ to form the logarithms (see
(2.8) of [KV] for the definition of the logarithms and basic properties).
The (non-classical) pseudo-differential operator log(Eeτ C ) − log E − τ C has
order ≤ −1 + , for any > 0, because the log and 0th order terms cancel
out in the calculation of the symbol, and we are assuming that E is classical.
Thus the RHSs of both (3.12) and (3.13) are of the form −sT +O(s2 ), where
T is trace class. It follows that the first two terms on the RHS of (3.11) are
zero.
As before
∂
(3.14) ζ τ C (s) = −str((Eeτ C )−s C), Re(s) > 0.
∂τ Ee
Since the first two terms on the RHS of (3.12) are zero, as s → 0,
∂
(3.15) ζ τ C (s) = −FPtr(E −s C) + O(s).
∂τ Ee
Thus
∂
(3.16) detζ (Eeτ C ) = detζ (Eeτ C )FPtr(E −s C),
∂τ
and this completes the proof of (a)
The proof of (b) follows the same pattern as for (a). Suppose that
order (F ) < order (E), where E + F is also invertible. Then we can write
E + F = E(1 + B), where order (B) ≤ −1, and since the spectrum of 1 + B
is discrete, we can find a spectral cut to form the compex powers of 1 + B.
Then
(3.17) FPtr((E(1 + B))−s C) = FPtr(E −s (1 + B)−s C) = FPtr(E −s C),
where the first equality uses (3.12) with log(1 + B) (which is of order ≤ −1)
in place of C, and the second uses (3.13) in the same way.
192 DOUG PICKRELL
Proof of (3.2). In the preceding section we viewed S(µ) as an operator on

a Hilbert space of two-forms. In this proof we will replace S(µ) by ∗S(µ)∗,
so that we can view it as an operator on functions. Also recall that ∆0 =
σ ◦ ∆ ◦ σ −1 .
Fix T > 0. Let E denote an operator as in (3.10) (for us E = ∆ or ∆0 ).
The manifold Σ has dimension 2, the operator S(µ) has order −2, and the
operator (1 + 2T S(µ))e−2T S(µ) is of the form 1+ order(−4). Hence by (3.4)
and (3.10),

1 2
(3.18) detζ E det((1 + 2T S(µ))e−2T S(µ) )
T
= detζ (E(T, µ)e−2T S(µ) )
= detζ (E(T, µ)) exp(−2T FPtr(E(T, µ)−s S(µ))),
where E(T, µ) = T1 E 2 (1 + 2T S(µ)).
Now we claim that
−s !
−s 1 2 −s
(3.19) FPtr(E(T, µ) S(µ)) = FPtr E (1 + 2T S(µ)) S(µ)
T
−s !
1 2
= FPtr E S(µ)
T
= Re s(S(µ)) ln T + FPtr(E −s S(µ)).
The first two equalities follow from (b) of (3.10). The third equality follows
simply from T s = 1 + slnT + O(s2 ),
Res(S(µ))
(3.20) tr(E −s S(µ)) = + FPtr(E −s S(µ)) + O(s) as s → 0
s
and the fact that changing the argument in a meromorphic function from s
to 2s does not change the constant term in a Laurent expansion.
From (3.18) and (3.19) we therefore conclude that

1 2
(3.21) detζ E det((1 + 2T S(µ))e−2T S(µ) )
T

1 2 −s
= detζ E + 2E 2 S(µ) T −T Re s(S(µ)) e−T F Ptr(E 2S(µ)) .
T
Since ∆ and ∆0 have the same spectrum, (3.21) implies that
detζ ( T1 ∆2 + ∆2 S(µ))
(3.22) 1 02 02
= exp(−T FPtr((∆−s − ∆0−s )S(µ))).
detζ ( T ∆ + ∆ S(µ))
Since the LHS of (3.22) has a finite limit as T → ∞, we conclude that
(3.23) FPtr(∆−s S(µ)) = FPtr(∆0−s S(µ)).
Let C(σ) = S(µ) − S(0), where µ = µσ−1 . Using (g) of (2.14), we see that
(2.13) translates into the equality of operators
(3.24) C(σ1 ◦ σ2 ) = C(σ1 ) + σ1 C(σ2 )σ1−1 ,
for σ1 , σ2 ∈ SDiff. It follows from (3.23) and (3.24) that
(3.25) SDiff → C : σ → FPtr(∆−s C(σ))
is a homomorphism of groups. But the only such homomorphism of groups
is the trivial map. We therefore conclude that
(3.26) FPtr(∆−s (S(µ) − S(0))) = 0,
provided that µ corresponds to a σ ∈ SDiff. [Note: There is a more direct
way to see that the real part of (3.26) is zero, for by (i) of (2.14) and (3.23),
(3.27) ReFPtr(∆−s (S(µ) − S(0))) = 2FPtr(∆0−s ∆0−1 − ∆−s ∆−1 ) = 0.]
Now the principal symbol of S(µ) is of the form
(3.28) ρ|ξ|−2 (1 − ξ µ̃ξ)−1 ,
where ρ and µ̃ are the local representations for µ and our fixed area ele-
ment, respectively. Since the density for Re s(S(µ)) can be computed by
integrating over |ξ| = 1, it follows that Re s(S(µ)) is independent of µ. Now
combining this with (3.21) we have
det((1 + 2T S(µ))e−2T S(µ) )
(3.29)
det((1 + 2T S(0))e−2T S(0) )
detζ T1 ∆2 + ∆2 2S(µ) −2T F Ptr(∆−s S(µ)−∆−s S(0))

= e
detζ T1 ∆2 + 4∆
for any µ. By (3.26), if µ corresponds to σ ∈ SDiff, then the exponential
term vanishes. One can now take the limit T → ∞. This completes the
proof.
Remarks 3.30. It is very unlikely that (3.26) vanishes for a general µ,
hence the limit in (3.2) probably does not exist for general µ. But this
makes perfectly good sense for the following reason. In the previous section,
following standard practice in probability theory, we regularized
(3.31) σ · det( det)∗ = exp(−S(µ)F · F )
by subtracting a trace, that is we replaced S(µ)F ·F by “S(µ)F ·F −trS(µ)”,
where this had to be properly interpreted as a random variable with respect
to νT . The proof of (3.2) makes it clear that this is inadequate; we should
actually have replaced (3.31) by
(3.32) exp(−T FPtr(∆−s S(µ))) exp(−T S(µ)reg(νT ) ),
that is, we should have added the trace back in! If we make this change,
then the limit in (3.2) exists for all µ.
194 DOUG PICKRELL
4. The Cases genus (Σ) > 0.

We briefly indicate the changes necessary to handle the case in which
genus (Σ) is positive.
First, it is no longer the case that the gauge group is connected; in fact,
we can identify
(4.1)
0 → K0 → K → π0 (K) → 0

exp π
0 → Ω0 (Σ; iR)/2πiZ −→ Map(Σ, T) −→
0
H 1 (Σ, 2πiZ) → 0
where π0 (g) = [g −1 dg] ∈ H 1 (Σ, 2πiZ). We define
(4.2)
Ω0 (L) = {f : A → C : f (A + g −1 dg) = c(g, A)f (A), ∀A ∈ A, g ∈ K}
1
A∧g −1 dg). The group D acts naturally on Ω0 (L).
R
where c(g, A) = exp( 2πi
Remarks 4.3. (a) Ω0 (L) is the space of sections of a D-equivariant line
bundle over C, i.e., c satisfies the cocycle identity
(4.4) c(g1 g2 , A) = c(g1 , A)c(g2 , A + g1−1 dg1 );
this depends crucially on the fact that
Z
1 −1 −1
(4.5) exp g2 dg2 ∧ g1 dg1 = 1.
2πi
(b) There is a refinement of (4.2), which depends upon the additional
choice of a spin structure on Σ. A spin structure determines a function
(4.6) (−1)q : H 1 (Σ, 2πiZ) → {±1}
1
R
satisfying (−1)q(λ1 +λ2 ) = (−1)q(λ1 ) (−1)q(λ2 ) e 4πi λ1 ∧λ2 , where q of a simple
loop is 0 or 1, depending upon whether the spin structure restricted to the
loop is trivial or nontrivial. We then define Ω0 (Lq ) to consist of functions
satisfying
Z
−1 q(π0 (g)) 1 −1
(4.7) f (A + g dg) = (−1) exp A ∧ g dg f (A).
4πi
The (−1)q factor compensates for the fact that the square root of the LHS
of (4.5) may not be 1. In this case we have a representation of Dq on Ω0 (Lq ).
To define the Yang-Mills measure, we must fix an area form as before.
As in the genus = 0 case, the support of the Yang-Mills measure is thicker
than the space of gauge equivalence classes of continuous T-connections, C.
To complete the space C in a D-equivariant way, one can adopt the point of
view that a connection is a parallel transport functor, hence that C consists
of holonomy functors. This is described in [Pi], see especially §2.5 for the
abelian case. The resulting completion of C has the structure of a principal
bundle
C completion ← H 1 (Σ, T)
(4.8) R ↓
{F ∈ Ω : F = 0}completion
2
over a completion of the space of curvatures, where

(4.9) H 1 (Σ, T) = Hom(π1 (Σ), T).
The Yang-Mills measure, again denoted by dνT , is the unique H 1 (Σ, T)-
invariant probability measure on the completion of C which projects to the
Gaussian corresponding to the inner product (2.4).
Now fix a complex structure for Σ. The section det is characterized in
the following:
Lemma 4.10. There is a unique section det of L with det(0) = 1 such
that Z
1 ∗
exp a∧a det
2πi
is a holomorphic function of a.
Proof. There is a canonical splitting of the sequence
(4.11) 0 → Ω0 /C → Ω0,1 → H 0,1 → 0
(4.12) ¯ 0 ⊕ ker(∂),
Ω0,1 = ∂Ω ¯ + a0
a = ∂x
¯ uniquely determines x ∈ Ω0 /C. Note that x depends
because ∂a = ∂ ∂x
holomorphically on a. The function
(4.13)
Z Z
i ∗ ¯ −1 i ∗
exp {a ∧ a − ∂a ∧ (∂ ∂) ∂a} = exp (a0 ∧ a0 + x ∧ F )
2π 2π
satisfies the transformation property (4.2) for g = exp(ξ) ∈ K0 , the identity
component of K. It follows that the section det has the form
Z
i ∗
(4.14) det(A) = exp (a0 ∧ a0 + x ∧ F ) Θ(a),
2π
where Θ is a holomorphic function which is K0 -invariant, hence G 0 -invariant,
where G is the complexification of K. Thus Θ descends to a function on H 0,1 .
For g ∈ K, not necessarily in the identity component, as in (4.12) we can
write uniquely
(4.15) g −1 dg = dξ + θ, g −1 ∂g
¯ = ∂ξ¯ + θ0,1
196 DOUG PICKRELL
where ∂θ0,1 = 0, ξ ∈ Ω0 (Σ, iR)/iR, θ ∈ H 1 (Σ, 2πiZ). Then

(4.16)
Θ(a + g −1 ∂g)
¯
det(A + g −1 dg)
= i ¯ ∧ (a + g −1 ∂g)
¯ ∗ − ∂(a + g −1 ∂g)
¯ ∧ x(a + g −1 ∂g)}
¯
R
exp 2π {(a + g −1 ∂g)
Z
i −1
1,0 0,1 ∗ 1,0

= exp A ∧ g dg − −a ∧ θ + θ ∧ (a − θ ) Θ(a)
2π
Z
i 0,1 0,1 1,0

= exp a ∧ 2θ − θ ∧ θ Θ(a).
2π
Thus Θ : H 0,1 → C satisfies
Z
0,1 i 1,0 0,1 1,0

(4.17) Θ(a + λ ) = exp a ∧ 2λ −λ ∧λ Θ(a)
2π
for all a ∈ H 0,1 , λ ∈ H 1 (Σ, 2πiZ). This implies that Θ is essentially Rie-
mann’s theta function
Z
1 X i 1,0 0,1
(4.18) Θ(a) = exp λ ∧ (a + 2λ ) .
Z 2π
λ

Remark 4.19. If we choose a spin structure q as in (b) of (4.3), i.e., a square
root of the canonical bundle, then we can identify Lq with the determinant
line of the family of Fredholm operators {∂¯κ1/2 + a : a ∈ Ω0,1 }, equipped
with the Quillen metric. In this case the section det is identified with the
canonical section, and (4.18) is replaced by the normalized theta function
Z
1 X q i 1,0 0,1
(4.20) Θq (a) = (−1) exp λ ∧ (a + 2λ ) .
Z 4π
Now suppose that σ ∈ D. As in (2.14) we denote transformed objects
using primes.
Lemma 4.21.
(a) (∂¯0 )0,1 = (1 − µµ̄)−1 (∂¯ − µ∂).
(b) (∂¯0 )1,0 = µ̄(∂¯0 )0,1 .
(c) a0 = (1 + µ̄)(1 − µ conj)−1 a, where conj denotes conjugation.
(c)0 a00 = f (µ, a0 ), where f depends linearly on a0 .
(d) x0 = (∂ ∂¯ − ∂µ∂)−1 {(∂ ∂¯ + ∂µ∂)x + ∂[µā0 + A00,1 01,0
0 − µA0 ]}, for R-valued
x, where A0 = a0 − ā0 .
¯ ∂¯ − ∂µ∂)−1 (∂ ∂¯ − ∂µ∂) + ∂[µā0 + A00,1 − µĀ01,0 ]+
(e) ∂ 0 ∂¯0 = 2∂ ∂((∂ 0 0
conjugate)−1 .
(f)
σ · det( det)∗
= exp(−S(µ)F · F )
Z
i ¯ −1
h
00,1 01,0
i
· exp (∂ ∂ − ∂µ∂) ∂ µā0 + A0 − µĀ0 ∧F
2π
Z
i
· exp (a00 ∧ ā00 + a0 + ā0 ) Θ0 (a00 )Θ̄(a0 ),
2π
where S(µ) = 2π i
∗ (∂ ∂¯ − ∂µ∂)−1 .
(h) The real and imaginary parts of S(µ) are symmetric, with respect to
the real Hilbert space structure (2.4).
(i) Re(S(µ)) = i ∗ (∂ 0 ∂¯0 )−1 = 2 ∗ ∗0 ∆0−1 (on 2-forms); hence if σ ∈ SDiff,
ReS(µ) = 2∆0−1 .
Proof of (4.21). If we view a as the projection (Ω1 )C → Ω0,1 , then using
(1.10),
A(σ −1 ) B(σ −1 )

0 ∗ A(σ) B(σ) 1
(4.22) a = σ∗ ◦ a ◦ σ =
B̄(σ) Ā(σ) 0 B̄(σ −1 ) Ā(σ −1 )
A(σ)A∗ (σ) −A(σ)B̄ ∗ (σ)

=
B̄(σ)A∗ (σ) −B̄(σ)B̄ ∗ (σ)
(1 − µµ̄)−1 −(1 − µµ̄)−1 µ

=
µ̄(1 − µµ̄)−1 −µ̄(1 − µµ̄)−1 µ
(1 − µµ̄)−1

1 1 −µ
= .
µ̄ 1 0 1
If we apply this to a one-form of the form df = ∂f ¯ + ∂f , then we obtain
(a) and (b). If we apply this to a one-form of the form A = a − ā, then we
obtain (c).
Similarly, if we view a0 as the projection onto ker(∂ : Ω0,1 → Ω2 ), then
A(σ)a0 A∗ (σ) −A(σ)a0 B̄ ∗

0 ∗
(4.23) a0 = σ∗ ◦ a0 ◦ σ = .
B̄(σ)a0 A∗ (σ) −B̄(σ)a0 B̄ ∗
If σ ∈ Aut(Σ), then σ ◦ a0 ◦ σ ∗ = a0 ; this implies (c)0 (although we would
dearly love to have an explicit formula).
Assuming that x is real, as in the proof of (2.13),
(4.24) ¯ + a0 = (1 + µ conj)−1 (∂¯ − µ∂)x0 + (a0 − ā0 )0,1
a = ∂x 0 0
0 00,1
=⇒ (∂¯ + µ∂)x + a0 + µā0 = (∂¯ − µ∂)x + A − µA01,0 .
0 0
By applying ∂ to both sides and solving for x0 , we obtain (d).

The other parts follow as in the proof of (2.14).
198 DOUG PICKRELL
Given σ ∈ SDiff, we define
(4.25) cT (σ) = exp(−T (S(µ) − S(0))reg + T L(µ, a0 ) ·T F )

Z 0
i Θ
· exp (a00 ∧ ā00 − a0 ∧ ā0 ) ,
2π Θ
(a random variable with respect to νT ), where

i
(4.26) L(µ, a0 ) = ∗ (∂ ∂¯ − ∂µ∂)−1 (∂[µa0 + A00,1 01,0
0 − µA0 ])
2π
(a smooth two-form with vanishing integral). As in (2.29), cT is an SDiff-
cocycle with respect νT .
|L|−2
We define the measure dνT having values in the line bundle |L|−2 by
|L|−2 s1 s2 ∗ −T S(0)reg i R a0 ∧ā0 2

(4.27) s1 s̄2 dνT = e e 2π |Θ| dνT ,
det det
where s1 , s2 are sections of L over the support of νT . We then have (2.31),
and we can define the representation πT . The spherical function is given by
(4.28) Z
|L|−2
φT (σ) = hπT (σ) · det, deti = σ · det( det)∗ dνT
Z
1 0 2 2
= e−T Sreg +T L·T F − 4π (|A0 | +|A0 | ) Θ0 Θ̄dνT
Z
−2T S −1/2 T −1 1 0 2 2
e 2 (1+2T S) L·L e− 4π (|A0 | +|A0 | ) Θ0 Θ̄

= det (1 + 2T S)e
where S = S(µ), L = L(µ, A0 ), |A0 |2 = − A0 ∧ ∗A0 , and the integral is

R
with respect to the translation invariant measure on the torus H 1 (Σ, T).
Proposition 4.29. In terms of the notations of the preceding paragraph,

the limit
φT (σ)
lim
T →∞ φT (0)
detζ (4∆)1/2
Z
¯ 1 0 2 +|A |2 )
= e2πi∗(∂ ∂−∂µ∂)L·L e− 4π (|A0 | 0
Θ0 Θ̄
detζ (2∆2 (∂ ∂¯ − ∂µ∂)−1 )1/2
exists and defines a positive definite function on SDiff.
This follows from (3.2).

References
[C] A. Connes, The Dixmier trace and the Wodzicki residue, preprint; incorporated into
Noncommutative Geometry, Academic Press, 1994.
[G] K. Gawedzki, Lectures on conformal field theory, School of Mathematics, IAS,
Princeton, 1996, to appear.
[I] R.S. Ismagilov, Representations of Infinite-Dimensional Groups, AMS Translations
of Mathematical Monographs, 152 (1996).
[K] A. Knapp, Representation Theory of Semisimple Groups, Princeton University
Press, 1986.
[KV] M. Kontsevich and S. Vishik, Geometry of determinants of elliptic operators, in
Functional Analysis on the Eve of the 21st Century, 1 (New Brunswick, NJ, 1993),
Prog. Math., 131, Birkhauser, (1995), 173-197.
[L] O. Lehto, Univalent Functions and Teichmuller Spaces, Springer-Verlag, 1986.
[Pi] D. Pickrell, On Y M2 measures and area-preserving diffeomorphisms, J. Geom. and
Physics, 19 (1996), 315-367.
[R] R. Ramer, On nonlinear transformations of Gaussian measures, J.F.A., 15 (1974),
166-187.
[Sen] A. Sengupta, Gauge theory on compact surfaces, Memoirs of the AMS, 126(600)
(1997).
[Sh] M.A. Shubin, Pseudodifferential Operators and Spectral Theory, Springer-Verlag,
1987.
Received December 1, 1997 and revised March 10, 1999.
University of Arizona
Tucson, AZ 85721
E-mail address: Pickrell@math.arizona.edu
Vol. 193, No. 1, 2000
ON CENTRAL EXTENSIONS OF GYROCOMMUTATIVE

GYROGROUPS
Krzysztof Rózga
Central extensions of gyrocommutative gyrogroups

(K-loops) are studied in order to clarify the status of a co-
cycle equation introduced by Smith and Ungar. A sufficient
and necessary conditions under which a central invariant ex-
tension is a gyrocommutative gyrogroup are formulated in
terms of a 2-cochain f (x, y). In particular, it is shown that for
central invariant extensions of gyrocommutative gyrogroups
defined by Cartan decompositions of simple Lie algebras, the
corresponding f (x, y) satisfies the cocycle equation, provided
an extension is a gyrocommutative gyrogroup.
1. Introduction.
There has been a renewal of an interest in loop theory in recent years, con-
cerning a special non-associative loop structure called a gyrocommutative
gyrogroup, known also under the name of a K-loop. It began with a paper
by A. Ungar [15], who pointed it out that the addition law of relativistic ve-
locities leads to an interesting algebraic structure on a unit ball in Rn , which
he originally called a K-loop. Further properties of these structures under
the name of gyrogroups and gyrocommutative gyrogroups were summarized
in [16]. Concerning terminology, see also remarks in [8].
Independent studies by A. Kreuzer [9] and H. Karzel, and H. Wefelscheid
[4] clarified the status of gyrocommutative gyrogroups within the framework
of loop theory and provided some important constructions generalizing an
example of Ungar.
In fact, gyrocommutative gyrogroups or similar structures were contem-
plated, although not explicitly under that name, by M. Kikkawa [5], in
relation to symmetric spaces and by P. Miheev and L. Sabinin in relation
to so called odular structures, [10]. That first aspect of gyrocommutative
gyrogroups has been discussed by Y. Friedman and A. Ungar in [2] and
recently by W. Krammer and H.K. Urbantke in [8].
One can obtain more examples of gyrocommutative gyrogroups by means
of their extensions. Such extensions are in fact extensions of loops. These
were discussed a long time ago by R.H. Bruck in [1]. They generalize exten-
sions of groups (see for example [12]). As for groups, the simplest among
201
202 K. RÓZGA
their extensions are the ones with an abelian kernel, which include central
extensions as well [12]. They are constructed, in principle, via a 2-cocycle,
i.e. a 2-cochain f (x, y) subject to a cocycle equation,
(1) f (x1 , x2 ) · f (x1 · x2 , x3 ) = f (x2 , x3 ) · f (x1 , x2 · x3 ),
under the assumption that the group under consideration acts trivially on
an abelian group.
Central invariant extensions of gyrocommutative gyrogroups, in a narrower
sense, were defined in [14] and then employed with a particular purpose of
reconstructing from a gyrocommutative gyrogroup of relativistic velocities,
the Lorentz group, i.e. SO0 (1, n); more specifically a standard matrix rep-
resentation of that group. As far as central extensions are concerned it was
assumed that the operators generating the left associant of an extension (in
terminology of [10]) or the structure group of an extension (in terminology
of [9]) are central (in the sense of [14]). That assumption resulted in an
equation for the corresponding 2-cochain f (x, y) which the authors called a
cocycle equation. Thus, in principal the cocycle equation was incorporated
into a definition of a central extension of a gyrocommutative gyrogroup. It
reads,
(2) f (x1 , x2 ) · f (x2 · x1 , x3 ) = f (x2 , x3 ) · f (x1 , x2 · x3 ).
That brings us to a question whether such an extension always has to

arise from a 2-cochain satisfying the cocycle equation of [14]?
To answer that question we discuss matters in a context of central exten-
sions of loops according to R.H. Bruck [1]. Basic definitions and properties
of relevant concepts are presented in Sections 2 and 3. A multiplicative
notation instead of an additive one, preferred by the authors of [14], is em-
ployed. Definitions concerning extensions of loops are, except of small mod-
ifications, faithful copies of the ones in [1]. Definitions of a gyrogroup and a
gyrocommutative gyrogroup (K − loop) are equivalent to the ones of [16]; see
Appendix. They emphasize the role of three identities, known as a Bol iden-
tity, an A-loop identity and an inverse automorphic identity (see for example
[10]). In Section 4 a discussion is restricted to invariant, in the sense of [14],
central extensions of a gyrocommutative gyrogroup M with a trivial action
of M on an abelian group G. Also there, necessary and sufficient conditions
under which a 2-cochain f (x, y) determines an extension of M which is a
gyrogroup or a gyrocommutative gyrogroup (Theorems 16 and 17) are pro-
vided. Theorem 20 of Section 5 is a generalization of the fact pointed out in
[14], that central extensions of gyrocommutative gyrogroups corresponding
to symmetric 2-cochains are gyrocommutative gyrogroups again.
In Section 6 we discuss gyrocommutative gyrogroups determined by Car-
tan decompositions of noncompact semisimple and in particular simple Lie
CENTRAL EXTENSIONS OF GYROGROUPS 203
algebras. It turns out that any central invariant extension of the latter struc-
ture, if it is a gyrocommutative gyrogroup, then it arises from a 2-cochain
which satisfies a cocycle equation of [14], (2), (Theorem 29).
2. Definitions of basic concepts.

Let G be an abelian group, M a loop, [11], and χ a function, χ : M →
Aut(G), which satisfies the following properties,
(3) 1χ = idG ,
where 1 and idG are units of M and Aut(G) respectively, and
(4) (x · y)χ = xχ · yχ,
for all x, y M.
Given g G and x M, we denote the value of an automorphism xχ at g
by gx.
Definition 1. A (G, M, χ)-extension (E,θ) is a pair consisting of a loop E
and a homomorphism θ of E onto M, such that,
(i) ker(θ) ⊂ A(E), where A(E) is the associator of E, i.e. a subset A(E)
of E, such that (e1 · e2 ) · e3 = e1 · (e2 · e3 ) if at least one of e1 , e2 , e3 is
in A(E),
(ii) the center of ker(θ) equals to G,
(iii) g · e = e · (g x) for all g ∈ G, e ∈ E and x = eθ.
A (G, M, χ)-extension, (E, θ), is called central if ker(θ) = G.
Definition 2. A normalized 2-cochain f is a function f : M 2 → G, with
values f (x1 , x2 ), taking the value 1 whenever one of x1 or x2 is 1.
The 3-coboundary δf of f is the following normalized 3-cochain,
(5) δf (x1 , x2 , x3 )
= [f (x1 , x2 )x3 ] · [f (x2 , x3 )]−1 · f (x1 · x2 , x3 ) · [f (x1 , x2 · x3 )]−1 .
Definition 3. Let f be a normalized 2-cochain. We define a central
(G, M, χ) extension, (E, θ), as follows.
(i) E is the set of all ordered pairs (x, g), where x ∈ M and g ∈ G,
(ii) (x, g) · (y, h) = (x · y, f (x, y) · (g y) · h),
(iii) (x, g)θ = x.
We denote this central extension by (G, M, χ, f ).
Remark 1. In order to fulfill the conditions of Definition 1 G has to be
identified with its homomorphic image in E under a natural injective homo-
morphism which sends g ∈ G into (1, g).
It is known, [1], that each central (G, M, χ)-extension is equivalent to at
least one (G, M, χ, f )-extension.
204 K. RÓZGA
Further discussion will concern extensions corresponding to a trivial func-

tion χ, i.e. the one that assigns to each x ∈ M the identity automorphism
of G. Such extensions will be referred to as (G, M ) or (G, M, f )-extensions
respectively.
To define a gyrogroup and a gyrocommutative gyrogroup we point out
certain important identities.
Definition 4. Let M be a loop. The following identities,
(6) Lx ◦ Ly ◦ Lx = Lx·(y·x) ,
(7) `(x, y) ◦ Lz = L`(x,y)z ◦ `(x, y),

where `(x, y) = L−1
x·y ◦ Lx ◦ Ly , and
(8) (x · y)−1 = x−1 · y −1 ,

are called a B, an A and an I-identity respectively. B stands for Bol, A for
an A-loop, and I for inverse automorphic. (For any z ∈ M , z −1 means its
right inverse.)
For purposes of this paper we introduce the following terminology. A loop
will be called a B-loop, a BA-loop or a BAI-loop, if a B, a B and an A or
B, A and I-identities hold.
Remark 2. There are other equivalent forms of those identities. In partic-
ular, (7) means that the mapping `(x, y) is an automorphism of M for all
x, y ∈ M . We list below two identities equivalent to (6) and (7) correspond-
ingly,
(9) x · (y · (x · z)) = (x · (y · x)) · z,
(10) l(x, y)(z · w) = (l(x, y)z) · l(x, y)w.

Now we define a gyrogroup and a gyrocommutative gyrogroup.
Definition 5. Any BA-loop is called a gyrogroup and any BAI-loop is
called a gyrocommutative gyrogroup.
We have also this natural definition.
Definition 6. A central (G, M ) extension, (E, θ), of a gyrocommutative gy-
rogroup is called a B, a BA or a BAI-extension if E is a B, a BA or a
BAI-loop respectively.
3. Central extensions of gyrocommutative gyrogroups.

We study now (G, M ) central extensions of a gyrocommutative gyrogroup M.
We follow the method of Bruck, [1].
Let u(x) be a normalized transversal function of E. Thus, for each x ∈ M ,

u(x) is an element of E, such that
(11) u(x)θ = x,
and
(12) u(1) = 1 .
Then every e ∈ E has a unique representation of the form e = u(x) · g,
with g ∈ G and x = eθ. Moreover,
(13) u(x) · g · u(y) · h = u(x) · u(y) · g · h = u(x · y) · f (x, y) · g · h,
where f is a normalized, because of (12), 2-cochain. (There is no need for
additional parentheses in that identity since g, h ∈ G ⊂ A(E), and G is
abelian.) Then the mapping u(x) · g 7→ (x, g) gives the equivalence of E and
(G, M, f ). (See [1].)
Thus, for all x, y ∈ M ,
(14) u(x) · u(y) = u(x · y) · f (x, y).
We make now few simple observations.
Lemma 7. Let (E, θ) be a central (G, M ) extension and u(x) a normalized
transversal function of E. Then for all g ∈ G and e ∈ E,
(15) g · e = e · g,
and
(16) (e · g)−1 = g −1 · e−1 = e−1 · g −1 .
If E is a B-loop, then
(17) (u(x))−1 = u(x−1 ) · [f (x, x−1 )]−1 = [f (x, x−1 )]−1 · u(x−1 ),
(18) f (x, x−1 ) = f (x−1 , x),
(19) `(u(a), u(b))u(c) = u(`(a, b)c) · ga,b,c ,

where
(20) ga,b,c = [f (a, b) · f ((a · b)−1 , a · b)]−1
· f ((a · b)−1 , a · (b · c)) · f (a, b · c) · f (b, c)
and
(21) `(u(a), u(b))g = g,
for all g ∈ G.
206 K. RÓZGA
Proof. Indeed, (15) follows directly from (iii) of Definition 1. (16) is a con-
sequence of the fact that G ⊂ A(E) and (15). e−1 is understood here as the
right inverse of e in the loop E.
For B-loops right and left inverses coincide [11]. Due to (14) we have,
(22) u(x) · u(x−1 ) = f (x, x−1 ),
and consequently
(23) u(x−1 ) = [u(x)]−1 · f (x, x−1 ).
Next,
(24) u(x−1 ) · u(x) = f (x−1 , x),
(25) u(x) = [u(x−1 )]−1 · f (x−1 , x)

and due to (16),
(26) [u(x)]−1 = [f (x−1 , x)]−1 · u(x−1 ).
Then from (23) and (26) we get (18) and (17). To prove (19) we employ
(15) and the other just proven identities. The method of proving (21) is
similar.
We prove now this fact.
Proposition 8. Let M be a gyrocommutative gyrogroup. The central exten-
sion E = (G, M, f ) is a B-extension, if and only if,
(27)
f (x, y · (x · z)) · f (y, x · z) · f (x, z) = f (x · (y · x), z) · f (x, y · x) · f (y, x).
Proof. It is clear, due to (9) and (13), (see also [1]), that a B-identity for E
is equivalent to,
(28) u(x) · (u(y) · (u(x) · u(z))) = (u(x) · (u(y) · u(x))) · u(z).
Now, using (14), (15), the fact that G ⊂ A(E) and commutativity of G,
one obtains,
(29) u(x) · (u(y) · (u(x) · u(z)))

= u(x · (y · (x · z))) · f (x, y · (x · z)) · f (y, x · z) · f (x, z)
and
(30) (u(x) · (u(y) · u(x))) · u(z)

= u((x · (y · x)) · z) · f (x · (y · x), z) · f (x, y · x) · f (y, x).
That shows equivalence of (28) and (27), since M is a B-loop.
As far as BA-extensions are concerned we have,
Proposition 9. Let M be a gyrocommutative gyrogroup. The central ex-

tension E = (G, M, f ) is a BA-extension if and only if it is a B-extension
and,
(31) f (z, w) · gx,y,z·w = f (`(x, y)z, `(x, y)w) · gx,y,z · gx,y,w .
Proof. An A − identity for E is equivalent to (see (10)),
(32) `(u(x), u(y)) · (u(z) · u(w)) = [`(u(x), u(y))u(z)] · `(u(x), u(y))u(w).
To put it into an equivalent form of (31), it suffices to apply (19), (20)
and (21). Indeed, one can rewrite then (32) into,
(33) u(`(x, y)(z · w)) · f (z, w) · gx,y,z·w
= u(`(x, y)z · `(x, y)w) · f (`(x, y)z, `(x, y)w) · gx,y,z · gx,y,w ,
which due to the fact that M is an A-loop is equivalent to (31).
Finally we have a Proposition concerning central BAI-extensions.
Proposition 10. Let M be a gyrocommutative gyrogroup. The central ex-
tension E = (G, M, f ) is a BAI-extension if and only if it is a BA-extension
and,
(34) f (x−1 , x) · f (y −1 , y) = f (x−1 , y −1 ) · f (x, y) · f ((x · y)−1 , x · y).
Proof. An I-identity for E is equivalent to,
(35) (u(x) · u(y))−1 = (u(x))−1 · (u(y))−1 .
Next one derives, by a rather straightforward process, which employs
Lemma 7, these identities,
(36) [u(x) · u(y)]−1 = u((x · y)−1 ) · [f (x, y) · f ((x · y)−1 , x · y)]−1 ,
and
(37)
[u(x)]−1 · [u(y)]−1 = u(x−1 · y −1 ) · f (x−1 , y −1 ) · [f (x−1 , x) · f (y −1 , y)]−1 .
Feeding them back into (35) and making use of the fact that M satisfies
an I-identity, one infers equivalence of (35) and (34).
4. Central invariant extensions of gyrocommutative gyrogroups.

Definition 11. A central extension E = (G, M, f ) of a loop M is called
invariant if
(38) f (`(a, b)x, `(a, b)y) = f (x, y)
for all a, b, x, y ∈ M .
Definition 12. For any 2-cochain f we define a 3-cochain ∆f,
(39) ∆f (x, y, z) = f (x, y) · [f (y, z)]−1 · f (y · x, z) · [f (x, y · z)]−1 .
208 K. RÓZGA
The expression for ∆f is different from the one for δf , (5). Indeed, those
two are related by,
(40) δf (x, y, z) = ∆f (x, y, z) · f (x · y, z) · [f (y · x, z)]−1 .
However, it is ∆f rather than δf , which is important in analysis of
identities of Propositions 8-10. We have now a sequence of a Proposition
and two Theorems that correspond to Propositions 8-10 in case of central
invariant extensions.
Proposition 13. Let M be a gyrocommutative gyrogroup. The central in-
variant extension E = (G, M, f ) is a B-extension if and only if,
(41) ∆f (x, y · x, `(y, x)z) · ∆f (y, x, z) = 1.
Proof. We rewrite (27) as,
(42)
f (v, u · (v · w)) · f (u, v · w) · f (v, w) = f (v · (u · v), w) · f (v, u · v) · f (u, v).
From (39) we infer that
(43) f (u, v · w) · f (v, w) · ∆f (u, v, w) = f (u, v) · f (v · u, w).
Hence (42) can be put into,
(44) f (v, u · (v · w)) · f (v · u, w) = f (v · (u · v), w) · f (v, u · v) · ∆f (u, v, w).
However,
(45) u · (v · w) = (u · v) · `(u, v)w
and
(46) f (v · u, w) = f (`(u, v)(v · u), `(u, v)w) = f (u · v, `(u, v)w),
since M is a gyrocommutative gyrogroup and (38) holds. Consequently, (44)
can be rewritten as,
(47) f (x, (y · x) · `(y, x)z) · f (y · x, `(y, x)z)

= f (x · (y · x), z) · f (x, y · x) · ∆f (y, x, z).
We simplify (47) further by means of (43) applied to the left-hand side of
(47) with u = x, v = y · x and w = `(y, x)z . The result is,
(48) f ((y · x) · x, `(y, x)z)

= f (x · (y · x), z) · ∆f (y, x, z) · ∆f (x, y · x, `(y, x)z).
Finally, we observe that
(49) f (x · (y · x), z) = f (`(y, x)`(x, y · x)((y · x) · x), `(y, x)z)
= f ((y · x) · x, `(y, x)z),
where the last inference is based on the identity,

(50) `(y, x)`(x, y · x) = `(y, x)`(x, y) = idM ,
satisfied in any gyrocommutative gyrogroup M (see [16], [14]). Feeding (49)
back into (48) reduces (48) to a desired form of (41). It is clear that an
outlined process of inference of (41) from (27) can be reversed. Therefore
(41) and (27) are indeed equivalent.
Before we discuss BA-extensions, we need an identity.
Lemma 14. Let M be a gyrocommutative gyrogroup and E = (G, M, f ) an
invariant central B-extension. Then,
(51) f ((a · b)−1 , a · (b · c)) · f (b · a, c) = f (a · b, (a · b)−1 ).
Proof. We employ (42) with v = (a · b)−1 , u = a · b and w = (a · b) · `(a, b)c.
That leads to,
(52) f ((a · b)−1 , (a · b) · `(a, b)c) · f (a · b, `(a, b)c) = f (a · b, (a · b)−1 ).
However,
(53) f ((a · b)−1 , (a · b) · `(a, b)c) = f ((a · b)−1 , a · (b · c)),
and
(54) f (a · b, `(a, b)c) = f (`(a, b)(b · a), `(a, b)c) = f (b · a, c).
Feeding (53) and (54) back into (52) one arrives at (51).
Now we prove this Lemma.
Lemma 15. Let Φ : M → G be a homomorphism of a loop M into a group
G. Then for all x, y, z ∈ M ,
(55) Φ(`(x, y)z) = Φ(z).
Proof. Indeed,
(56) Φ(x · (y · z)) = Φ(x) · Φ(y) · Φ(z)
and
(57) Φ(x · (y · z)) = Φ((x · y) · `(x, y)z) = Φ(x) · Φ(y) · Φ(`(x, y)z).
Hence (55) follows.
Theorem 16. Let M be a gyrocommutative gyrogroup. The central invari-
ant extension E = (G, M, f ) is a BA-extension if and only if
(58) ∆f (x, y · x, z) · ∆f (y, x, z) = 1
and
(59) ∆f (x, y, z · w) = ∆f (x, y, z) · ∆f (x, y, w).
210 K. RÓZGA
Proof. Let E be a BA-extension. We show first that the condition (31) of

Proposition 9 implies (59). Indeed, due to (38), (31) reads,
(60) g x,y,z·w = gx,y,z · gx,y,w .
Making use of (39), with u = a, v = b, and w = c in (20), one obtains,
(61)
f ((a · b)−1 , a · b) · ∆f (a, b, c) · ga,b,c = f (b · a, c) · f ((a · b)−1 , a · (b · c)).
Employing now (61) and (51) in (60), one arrives at (59). Now, hav-
ing established that, we employ Lemma 15 to prove that (41) implies (58).
Conversely, if E is a central invariant extension and satisfies (58) and (59),
then by Lemma 15 it satisfies (41) as well. Thus E is a B-extension. Conse-
quently, (61) holds and because of (51) it reads ga,b,c = [∆f (a, b, c)]−1 . Next,
(59) can be rewritten into (60) which for invariant extensions is (31).
Theorem 17. Let M be a gyrocommutative gyrogroup. The central invari-
ant extension E = (G, M, f ) is a BAI-extension if and only if,
(62) ∆f (x, y · x, z) · ∆f (y, x, z) = 1,
(63) ∆f (x, y, z · w) = ∆f (x, y, z) · ∆f (x, y, w)

and
(64) f (x, y) · ∆f (y, x, x−1 ) = f (y, x) · ∆f (x · y, x−1 , y −1 ).
Proof. Due to Theorem 16 it suffices to prove that a central invariant BA-
extension is a BAI-extension if and only if (64) holds. We prove that (64)
is equivalent to (34).
From (43) with u = x−1 · y −1 , v = x and w = y, we get,
(65) f (x−1 · y −1 , x · y) · f (x, y) · ∆f (x−1 · y −1 , x, y)

= f (x−1 · y −1 , x) · f (y −1 , y).
Next, using (65) and an I-identity for M in (34), one arrives at,
(66) f (x−1 , x) · ∆f (x−1 · y −1 , x, y) = f (x−1 , y −1 ) · f (x−1 · y −1 , x).
Employing again (43), with u = y −1 , v = x−1 , w = x, we arrive at,
(67) f (y −1 , x−1 ) · f (x−1 · y −1 , x) = f (x−1 , x) · ∆f (y −1 , x−1 , x).
Feeding it back into (66) leads to,
(68) f (x−1 , y −1 ) · ∆f (y −1 , x−1 , x) = f (y −1 , x−1 ) · ∆f (x−1 · y −1 , x, y),
which is equivalent to (64).
5. Central symmetric and invariant extensions of a

gyrocommutative gyrogroup.
Definition 18. A central extension E = (G, M, f ) is called symmetric if,
(69) f (x, y) = f (y, x).
We prove this Proposition.
Proposition 19. Let M be a gyrocommutative gyrogroup and let E =
(G, M, f ) be a central, symmetric, invariant, BA-extension. Then, E is
a BAI-extension. In particular,
(70) ∆f (z, y, x) = (∆f (x, y, z))−1 ,
(71) ∆f (x, y, y) = 1,
and
(72) ∆f (a, b, c) = ∆f (c, a, b) = ∆f (b, c, a).
Proof. We prove (70)-(72) first. In order to prove (70) it suffices to apply
a definition of ∆f, and (69). Next we employ identities, (58) and (59) of
Theorem 16. In particular, in (58) we substitute x = a, y = b/a and z = c.
Since M is a B-loop (even a gyrocommutative gyrogroup), then (see [10],
[13]),
(73) b/a = a−1 · ((a · b) · a−1 ).
Feeding that back into (58), applying (70) and (59), one arrives at,
(74) ∆f (a, b, c) = ∆f (c, a, b) · ∆f (c, a, a−1 ).
Putting b = 1, one infers that ∆f (c, a, a−1 ) = 1, which due to (59) is
equivalent to ∆f (c, a, a) = 1. That in turn is equivalent to (71). Now, (74)
reads,
(75) ∆f (a, b, c) = ∆f (c, a, b),
which implies (72).
Next, it is not difficult to prove the principal assertion of the Proposition.
Indeed, it suffices to prove the identity (64) of Theorem 17. That identity
is a straightforward consequence of (69), (59) and (70)-(72).
We close this section with a Theorem.
Theorem 20. Let M be a gyrocommutative gyrogroup and E = (G, M, f ) a
symmetric, invariant, BA-extension. E is a BAI-extension if and only if,
(76) ∆f (x, y, z · w) = ∆f (x, y, z) · ∆f (x, y, w),
(77) ∆f (z, y, x) = (∆f (x, y, z))−1 ,

212 K. RÓZGA
(78) ∆f (x, y, y) = 1,
and
(79) ∆f (x, y, z) = ∆f (z, x, y) = ∆f (y, z, x).
Proof. Indeed, necessity of these conditions follows from Theorem 16 and
Proposition 19 . Conversely, it is not difficult to infer from them (58).
Thus, according to Theorem 16, E is a BA-extension and by Proposition
19, it is a BAI-extension.
6. Gyrocommutative gyrogroups determined by Cartan

decompositions of noncompact semisimple Lie algebras.
For definitions of related concepts the reader is referred to [3].
Let g be a noncompact semisimple Lie algebra over R and let g = t + p,
be a Cartan decomposition of g. Then the mapping θ : T + X 7→ T − X,
where T ∈ t and X ∈ p is an involutive automorphism of g.
We say that a pair (G, H) is associated with (g, θ) if G is a connected Lie
group with Lie algebra g and H is a Lie subgroup of G with Lie algebra t.
We shall refer to (G, H) as to a pair of a noncompact type.
Theorem 21 (Helgason [3]). Let (G, H) be a pair of a noncompact type
associated with (g, θ). Then:
(i) There exists an involutive, analytic automorphism Θ of G whose fixed
point set is H and whose differential at the identity of G is θ. In
particular H is closed.
(ii) The mapping ϕ : p × H → G, defined by
(80) ϕ(X, h) = (exp X)h
is a diffeomorphism.
Remark 3. According to that Theorem, G = P H, where P = exp p, is an
exact decomposition of G, i.e. any element g ∈ G has a unique representation
g = ph, where p ∈ P and h ∈ H. We refer to those p and h as to P - and
H-factors of g respectively.
Next, there is a natural binary operation ? on P determined by the exact
decomposition of G. Indeed, ∀p1 , p2 ∈ P, one defines p1 ? p2 to be a unique
P -factor of p1 p2 . Thus,
(81) p1 p2 = (p1 ? p2 )h(p1 , p2 ),
where h(p1 , p2 ) is an H − f actor of p1 p2 . It turns out that this operation
provides a gyrocommutative gyrogroup structure on P . The origins of this
fact can be traced in papers of M. Kikkawa [5]; see also [10], [2] and for the
most recent discussion of the subject, [8]. Notice also that the inverse in the
group G of an element p ∈ P , which we denote by p−1 , is an element of P and
it is the inverse of p in the loop (P, ?). Indeed, 1 = pp−1 = (p?p−1 )h(p, p−1 ).
Consequently, p ? p−1 = 1. The same is true for p−1 ? p.
Theorem 22. Let (G, H) be a pair of a noncompact type associated with
(g, θ), G = P H the corresponding exact decomposition of G and ? a binary
operation on P determined by the decomposition. Then (P, ?) is a smooth
gyrocommutative gyrogroup.
The following fact will be useful in a sequel.
Lemma 23. Let n be a positive integer, n ≥ 2 and p1 , . . . , pn ∈ P . Then,
(82) pn pn−1 · · · p2 p1 = qn hn ,
where qn , hn are elements of P and H respectively, determined recursively
by,
(83) q1 = p1 , h1 = 1,
and (see (81)),
(84) qk+1 = pk+1 ? qk , hk+1 = h(pk+1 , qk )hk ,
for k = 1, . . . , n − 1.
Proof. By induction. For n = 2 it follows from (81). Suppose the assertion
is true for n = m. Take n = m + 1. Then,
pm+1 pm · · · p1 = pm+1 qm hm = (pm+1 ? qm )h(pm+1 , qm )hm ,
where the last equality is inferred from (81). Hence qm+1 = pm+1 ? qm , and
hm+1 = h(pm+1 , qm )hm . Thus the assertion is true for n = m + 1.
Now we can prove the following Proposition.
Proposition 24. Let (G, H) be a pair of noncompact type associated with
(g, θ), where g is simple. Let G = P H be the corresponding exact decompo-
sition of G. Then, the group G is generated by the set P and the subgroup
H is generated by the set
(85) S = {h(p1 , p2 ) : p1 , p2 ∈ P },
in the sense that any element of G or H is a product of a finite number of
elements of P or S respectively.
Proof. Indeed, the homogeneous space G/H of the pair (G, H) is reductive
(see [7], p. 27), because Ad(h)p ⊂ p, for all h ∈ H (see [3]). Then, ([7],
p. 27), l = p + [p, p] is a nontrivial ideal of g and the corresponding connected
normal subgroup of G is generated by the set P = {exp(X) : X ∈ p}.
However, since g is simple l = g and G itself is generated by the set P .
To prove the second part of this Proposition, assume that h ∈ H. Then
according to the first part of the Proposition there exists a finite sequence
p1 , . . . , pn of elements of P , such that h = pn · · · p1 . Employing Lemma
214 K. RÓZGA
23 we infer that h = qn hn , where qn ∈ P, hn ∈ H and hn is a product of

elements from S. Moreover, since G = P H is an exact decomposition, we
must have qn = 1 and h = hn .
We need yet another definition ([10]).
Definition 25. Let M be a loop. Then its left associant, asl (M ) is a sub-
group of the group of left multiplications of M , LM lt(M ), generated by the
set, {`(x, y) : x, y ∈ M }.
A relation between asl (P ) and H is clarified in the following Proposition.
Proposition 26. Under the assumptions of Proposition 24, the group
asl (P ) is identical to the group of conjugations of P by elements of H.
Proof. Indeed, for all p1 , p2 , p3 ∈ P,
`(p1 , p2 )p3 = (p1 ? p2 )−1 ? (p1 ? (p2 ? p3 )).
However, employing (81) one obtains,
p1 p2 p3 = (p1 ? (p2 ? p3 ))h,
(p1 ? p2 )−1 = h12 p−1 −1
2 p1 ,
where h, h12 ∈ H and h12 = h(p1 , p2 ) ∈ S. (Concerning the inverses of

elements of P see Remark 3.) Hence,
h12 p3 = (p1 ? p2 )−1 (p1 ? (p2 ? p3 ))h = [`(p1 , p2 )p3 ]h1 ,
where h1 is yet another element of H.
Consequently, `(p1 , p2 )p3 = (h12 p3 h−1 −1
12 )(h12 h1 ). Since the left-hand side
of the previous equation is an element of P , and the factors of its right-hand
side are in P and H respectively, therefore h12 h−1 1 = 1 and,
`(p1 , p2 )p3 = h12 p3 h−1

12 .
Notice that h12 depends on p1 and p2 only. Therefore `(p1 , p2 ) is indeed
a conjugation of P by h12 ∈ H. Consequently, any element of asl (P ) is a
conjugation of P by an element of H. Conversely, given h ∈ H, it can be
represented as a product of elements of the form h(p, q), where p, q ∈ P, (85).
Therefore a conjugation by h is a product of conjugations by such elements.
Since a conjugation by h(p, q) equals to `(p, q), therefore a conjugation by
h is an element of asl (P ).
Now we arrive at this result.
Proposition 27. Under the assumptions of Proposition 24, let Φ : P → A,
be a homomorphism of a gyrocommutative gyrogroup P into an abelian group
A. Then there exists a homomorphism Ψ : G → A, such that Ψ |P = Φ and
for all h ∈ H, Ψ(h) = 1.
Proof. Indeed, we define the mapping Ψ : G → A by, Ψ(g) = Φ(p) for

all g ∈ G, where p is a unique P -factor of g. Next we verify that Ψ is a
homomorphism. Indeed, for p1 , p2 ∈ P and h1 , h2 ∈ H, there exists, due to
(81), h ∈ H such that,
p1 h1 p2 h2 = p1 ph2 1 h1 h2 = (p1 ? ph2 1 )h,
where ph2 1 = h1 p2 h−1
1 .
Consequently,
(86) Ψ(p1 h1 p2 h2 ) = Φ(p1 ? ph2 1 ) = Φ(p1 )Φ(ph2 1 ).
However, according to Proposition 26, the conjugation of P by h1 is an
element of asl (P ). Then, employing Lemma 15 we can rewrite (86) into,
Ψ(p1 h1 p2 h2 ) = Φ(p1 )Φ(p2 ) = Ψ(p1 h1 )Ψ(p2 h2 ).

Finally we point out the following fact.
Proposition 28. Let ψ : g → a, be a homomorphism of a simple Lie algebra
g, with a Cartan decomposition g = t + p, into an abelian Lie algebra a, such
that ψ |t = 0. Then, ψ = 0.
Proof. Indeed, ker ψ 6= {0} and ker ψ is an ideal of g. But g is simple.
Therefore ker ψ = g.
We arrive now at the main assertion of this section.
Theorem 29. Let (G, H) be a pair of a noncompact type associated with
(g, θ), where g is simple. Let G = P H be the corresponding exact decomposi-
tion of G and ? a binary operation on P determined by that decomposition.
Let E = (A, P, f ), where A is an abelian Lie group and f (p, q) a smooth
2-cochain, be a central, invariant extension of P . Then:
(i) E is a BA-extension if and only if ∆f (p, q, r) = 1, for all p, q, r ∈ P .
(ii) E is a BAI-extension if and only if ∆f (p, q, r) = 1 and f (p, q) =
f (q, p), for all p, q ∈ P.
Proof. The conditions (59) of Theorem 16 or (63) of Theorem 17 are crucial.
Indeed, let ∀p, q ∈ P, Φp,q be a mapping of P into A, defined by
Φp,q (r) = ∆f (p, q, r),
for all r ∈ P . Now, if that mapping is a homomorphism of a gyrocommutative
gyrogroup P into a group A, then it can be extended to a homomorphism
Ψp,q of the group G into A such that Ψp,q |P = Φp,q and Ψp,q (h) = 1, for all
h ∈ H (Proposition 27). Ψp,q induces, in turn, a homomorphism ψp,q of Lie
algebras g and a (ψp,q is a differential of Ψp,q at 1 ∈ G). It must be a trivial
homomorphism, by Proposition 28; ∀p, q ∈ P, ∀Z ∈ g, ψp,q Z = 0. However,
the group G is generated by elements of the form exp Z, where Z ∈ g, and
216 K. RÓZGA
Ψp,q (exp Z) = exp ψp,q Z = 1. Thus Ψp,q is itself trivial. Consequently, if

∆f (p, q, r) satisfies (59) of Theorem 16, then ∆f = 1, which already implies
identity (58). Similarly, (63) of Theorem 16, implies ∆f = 1, (62) and
reduces (64) into, f (p, q) = f (q, p).
Conversely, given that ∆f = 1, one infers using Theorem 16 that E is a
BA-extension. Likewise, assuming ∆f = 1 and the identity f (p, q) = f (q, p),
one employs Theorem 17 to infer that E is a BAI-extension.
Remark 4. The equation, ∆f = 1 (2) has been called in [14] a cocycle

equation. Together with a condition of invariance for a cochain f , they were
employed there to study central extensions of gyrocommutative gyrogroups.
Our slightly more general treatment of central extensions shows that at
least for a class of gyrocommutative gyrogroups discussed in this section
(Theorem 29), a cocycle equation emerges in a natural way. This class
of gyrocommutative gyrogroups is determined by Cartan decompositions of
simple Lie algebras. Many of classical Lie algebras are of that type. (For
their list see [6], Appendix C.) In particular, a gyrocommutative gyrogroup,
which has been one of the main objects studied in ([14]), and which arises
from the pair (SO0 (1, n), SO(n)), belongs to that class.
Appendix.
For completeness of presentation we clarify here how the definitions of gy-
rogroup and gyrocommutative gyrogroup used in this paper correspond to
the original ones.
According to [16] a gyrogroup is a grupoid (M, ·) satisfying the following
axioms.
(G1) There is an element 1 ∈ M , such that,
1 · x = x,
for all x ∈ M.
(G2) For each x ∈ M there is an a ∈ M such that
a · x = 1.
(G3) For any x, y, z ∈ M there exists a unique element gyr[x, y]z ∈ M
such that
x · (y · z) = (x · y) · gyr[x, y]z.
(G4) If gyr[x, y] denotes the map of M into M , given by z 7→ gyr[x, y]z
then gyr[x, y] ∈ Aut(M, ·).
(G5) For all x, y ∈ M
gyr[x, y] = gyr[x · y, y].
From these axioms one can infer (see [16]) that the equation x · y = z has
a unique solution for x or y, given the other two elements. Thus (M, ·) is a
loop, [11].
Proposition 30. Let (M, ·) be a loop. (M, ·) is a gyrogroup if and only if
it is an A-loop and it is a left Bol loop, [11].
Proof. Indeed, suppose that (M, ·) is a gyrogroup. Then (G3) can be rewrit-
ten into gyr[x, y]z = L−1 x·y Lx Ly z = `(x, y)z. Hence gyr[x, y] = `(x, y). There-
fore (G3) and (G4) are equivalent to the statement that (M, ·) is an A-
loop. To prove Bol-identity we calculate Lx Ly Lx z = x · (y · (x · z)) =
x · ((y · x) · `(y, x)z) = (x · (y · x)) · `(x, y · x)`(y, x)z. However, (see [16]),
`(x, y · x) = `(x, y) and `(y, x) = (`(x, y))−1 . Hence Lx Ly Lx = Lx·(y·x) ,
which is a Bol identity, (6).
Conversely, let (M, ·) be an A − loop and a Bol loop. Then (G1) −
(G4) hold automatically. It suffices to prove (G5). We obtain, L(x·y)·y =
L(x·y)·(x−1 ·(x·y)) = Lx·y Lx−1 Lx·y , where we have employed the fact that for
Bol loops `(x, x−1 ) = `(x−1 , x) = Lx−1 Lx = idM , [11]. Therefore
L−1 −1 −1 −1 −1
(x·y)·y Lx·y Ly = Lx·y Lx−1 Lx·y Lx·y Ly = Lx·y Lx Ly , which is equivalent to
`(x · y, y) = `(x, y), which is (G5).
Now, according to [16], (M, ·) is a gyrocommutative gyrogroup if it is a

gyrogroup, which satisfies an additional axiom,
(G6) x · y = gyr[x, y](y · x).
As it has been indicated in [16], this identity in a gyrogroup is equivalent
to an inverse automorphic identity, (8).
Therefore the following is true.
Proposition 31. Let (M, ·) be a loop. (M, ·) is a gyrocommutative gyro-
group if and only if it is an A-loop, a left Bol loop and it satisfies an inverse
automorphic identity.
Acknowledgments. I would like to acknowledge the support and hospital-

ity of the Department of Mathematics at NDSU during my sabbatical year.
My special thanks go to Abraham Ungar.
References
[1] R.H. Bruck, An extension theory for a certain class of loops, Bull. Am. Math. Soc.,
57 (1951), 11-26.
[2] Y. Friedman and A. Ungar, Gyrosemidirect product structure of bounded symmetric
domains, Res. Math., 26 (1994), 28-38.
[3] S. Helgason, Differential Geometry and Symmetric Spaces, Pure and Applied Math-
ematics, Academic Press, New York, 1962.
218 K. RÓZGA
[4] H. Karzel and H. Wefelscheid, Groups with an involutory antiautomorphism and K-

loops; Application to space-time-world and hyperbolic geometry, I, Res. Math., 23
(1993), 338-354.
[5] M. Kikkawa, Geometry of homogeneous Lie loops, Hiroshima Math. J., 5 (1975), 141-
179; ‘On some quasigroups of algebraic models of symmetric spaces, II’, Mem. Fac.
Lit. Sci. Shimane Univ. Nat. Sci., 7 (1974), 29-35.
[6] A.W. Knapp, Lie Groups Beyond an Introduction, Progress in Mathematics, Vol. 140,
Brikhauser, Boston, 1996.
[7] O. Kowalski, Generalized Symmetric Spaces, Lecture Notes in Mathematics, Vol. 805,
Springer-Verlag, Berlin, 1980.
[8] W. Krammer and H.K. Urbantke, K-loops, gyrogroups and symmetric spaces,
preprint, 1997.
[9] A. Kreuzer, Beispiele endlicher und unendlicher K-loops, Res. Math., 23 (1993),
355-362.
[10] P.O. Miheev and L.V. Sabinin, Quasigroups and Differential Geometry, 357-430; in
‘Quasigroups and Loops. Theory and applications’ (O. Chein, H.O. Pflugfelder and
J.D.H. Smith, eds.), Sigma Series in Pure Mathematics, Vol. 8, Heldermann-Verlag,
Berlin, 1990.
[11] H.O. Pflugfelder, Quasigroups and loops: Introduction, Heldermann, Berlin, 1990.
[12] D.J.S. Robinson, A course in the theory of groups, Springer-Verlag, New York, 1996.
[13] B.L. Sharma, Left loops which satisfy left Bol identity, Proc. Amer. Math. Soc., 61
(1976), 189-195.
[14] J.D.H. Smith and A. Ungar, Abstract space-times and their Lorentz groups, J. Math.
Phys., 37 (1996), 3073-3098.
[15] A.A. Ungar, The relativistic noncommutative nonassociative group of velocities and
the Thomas rotation, Res. Math., 16 (1989), 168-179.
[16] , Thomas Precession: Its underlying gyrogroup axioms and their use in hyper-
bolic geometry and relativistic physics, Found. Phys., 27 (1997), 881-951.
Received February 24, 1998 and revised August 19, 1998. This work has been supported
by NFS grant #OSR-9452892 from NDSU EPScoR.
University of Puerto Rico

Mayaguez PR 00681-9018
E-mail address: rozga@math.upr.clu.edu
Vol. 193, No. 1, 2000
GLOBAL EXISTENCE AND DECREASING PROPERTY OF

BOUNDARY VALUES OF SOLUTIONS TO PARABOLIC
EQUATIONS WITH NONLOCAL BOUNDARY
CONDITIONS
Sangwon Seo
It is shown that the local solution of parabolic equation

with nonlocal boundary condition representing entropy can
be extended to whole time domain for weights with large L1
norms. When the weight is identically zero on some part
of the boundary, it is shown that the boundary values can
decrease even when the other weights are some large.
1. Introduction.
This paper is concerned with the investigation of large time behavior of
solutions to parabolic initial value problem subject to nonlocal boundary
condition which describes the entropy in a quasi-static theory of thermoe-
lasticity, namely,
ut (x, t) = ∆u(x, t) + µu, (x, t) ∈ Ω × (0, T ),
Z
(1.1) u(z, t) = f (z, y)u(y, t) dy, z ∈ ∂Ω, 0 < t < T,
Ω
u(x, 0) = u0 (x), x ∈ Ω,
where u0 (x) are assumed to be continuous on Ω and µ is a constant. The
function f (z, y) is defined for z ∈ ∂Ω and y ∈ Ω and continuous functions
in y ∈ Ω for each z ∈ ∂Ω. Since for each z ∈ ∂Ω, f (z, y) plays the weight
of integration in (1.1), the function f (z, y) is called weights throughout this
paper. Denote DT = Ω × (0, T ) and DT ∪ ΓT = Ω × [0, T ). The variable
z stands for a generic point of boundary ∂Ω. The large time behavior
of the solution u to Problem (1.1) is studied by taking an upper bound
for u in Section 2. It is shown that the solution of Problem (1.1) with
large weights f (z, ·) for each z ∈ ∂Ω has an exponential lower bound in
Section 3. Moreover, it is shown that the difference between maximum and
minimum value on the boundary ∂Ω can decrease if the weights are zero
on a nonempty subset of ∂Ω in z and zero on the boundary in y, that is,
f (z, y) = 0, z ∈ Γ ⊂ ∂Ω for all y ∈ Ω and f (z, y) = 0, y ∈ ∂Ω.
219
220 SANGWON SEO
A function u(x, t) is called a subsolution of Problem (1.1) on DT if u ∈

C 2,1 (DT )
∩ C(DT ∪ ΓT ) satisfies
ut ≤ ∆u + µu on DT ,
Z 1
(1.2) u(z, t) ≤ f (z, y)u(y, t) dy, z ∈ ∂Ω, 0 < t < T,
0
u(x, 0) ≤ u0 (x), x ∈ Ω.
A supersolution is defined by reversing inequalities in (1.2). Using the no-
tions of super and subsolution, we state the following:
Theorem 1.1. Comparison principle: Let f (z, y) be nonnegative and con-
tinuous in y on Ω for each z ∈ ∂Ω. Let also u and v be subsolution and
supersolution of Problem (1.1), respectively, and u(x, 0) < v(x, 0) for x ∈ Ω.
Then
u<v in DT .
As a corollary, u0 (x) > 0(< 0) implies u(x, t) > 0(< 0). Moreover, the
local existence and uniqueness can be written as:
Theorem 1.2. Let u0 in (1.1) be continuous. Then for a small T , there is
a unique solution of Problem (1.1) in C 2,1 (DT ) ∩ C(DT ∪ ΓT ).
For the proofs of the comparison principle and the local existence theorem,
see [4]. Here one can see that the results hold without any restriction to
constant µ in the parabolic equation in (1.1). In this paper, we are only
interested in the positive solutions of Problem (1.1), and hence we assume
that u0 > 0. Throughout this paper, L1 and L2 norm of a function on Ω is
denoted by k · k1 and k · k2 , respectively.
2. Global existence.
Under the convexity assumption on the domain Ω and the following assump-
tions on the weights;
(2.1) kf (z, ·)k1 < 1 for each z ∈ ∂Ω,
it was known ([3]) that the maximum modulus maxx∈Ω |v(x, t)| of solution
v to the parabolic equation,
vt = ∆v + νv, ν ≤ 0,
subject to the nonlocal boundary conditions in (1.1) decreases, moreover, it
was shown the exponential decay of the solution; there are some constants
c1 and γ > 0 such that
(2.2) max |v(x, t)| ≤ c1 e−γt .
x∈Ω
GLOBAL EXISTENCE AND DECREASING PROPERTY 221
On the other hand, the author of [2] set up an example of one dimensional
problem with µ = 0 and weights that violate (2.1), such that its maximum
modulus of solution increases exponentially. In this paper, basic assumption
on the weights is that
f (z, y) ≥ 0 for y ∈ Ω,
for each z ∈ ∂Ω. One of the main concern is to find an upper bound of
solutions in (1.2) with nonnegative weights in the integrals of boundary
conditions which guarantees the global existence. For Problem (1.1) with
arbitrary nonnegative weights, the author could not find other results about
global existence. For convenience, let the domain Ω = {x = (x1 , x2 ) ∈ R2 :
|x| < 1}. The results derived here can be extended to higher dimension
without difficulties.
Theorem 2.1. For the solution u to Problem (1.1) with nonnegative

weights f , there is a sufficiently smooth function S(x, t) such that
(2.3) u(x, t) < S(x, t), t > 0.
The choice of the function S(x, t) depends on the weights and supx∈Ω u0 (x).
Proof. Let S = p + q with p and q such that
(2.4) p(x, t) = Aeκt+a(r−1) and q(x, t) = Aeκt−a(r+1) , x ∈ Ω, t > 0,

1/2
where r = |x| = x21 + x22 . The positive constants A, a and κ will be
chosen later. Then p and q satisfy
p(z, t) = Aeκt , q(z, t) = Aeκt e−2a for each z ∈ ∂Ω,

1/2
√ e−2a

1 1
kp(·, t)k2 = 2π − + Aeκt ,
2a 4a2 4a2
1/2
√
−2a
e 1 1 −4a
kq(·, t)k2 = 2π − + e Aeκt .
4a2 2a 4a2
If a satisfies
( 1/2 1/2 )
√

1 + 2a 1
(2.5) 2π + max {kf (z, ·)k2 } ≤ 1,
4a2 2a z∈∂Ω
then one has
(2.6) {kp(·, t)k2 + kq(·, t)k2 } kf (z, ·)k2 < Aeκt for each z ∈ ∂Ω.
222 SANGWON SEO
Note that
∂S
(0, t) = 0, t > 0,
∂r
n o
ar−1 ea(r−1) − e−a(r+1) ≤ ar−1 ea(r−1) ,
n o
lim ar−1 ea(r−1) = a2 e−a ,
r→0
ea(r−1)
+ e−a(r+1) ≥ 2e−a .
Since ar−1 ea(r−1) is decreasing for r < 1/a and increasing for r > 1/a, one
has ar−1 ea(r−1) ≤ max{a2 e−a , a}. Choose the constant a so large that (2.5)
holds and,
a2 e−a < a, ar−1 ea(r−1) ≤ a,
and take an integer k ≥ 2 so that 2ak e−a > a. Now let
κ ≥ a2 + ak + µ.
Then, by the choice of κ, k and a, we see that S = p + q satisfies
St = κS
≥ ak S + a2 S + µS
n o
= Aeκt ak ea(r−1) + e−a(r+1) + a2 S + µS
≥ Aeκt ak (2e−a ) + a2 S + µS
≥ Aeκt a + a2 S + µS
≥ Aeκt ar−1 ea(r−1) + a2 S + µS

≥ Aeκt ar−1 ea(r−1) − e−a(r+1) + a2 S + µS
∂S ∂ 2 S
= r−1 + 2 + µS
∂r ∂r
= ∆S + µS on DT ,
and by the inequality (2.6), we obtain for z ∈ ∂Ω,
S(z, t) = p(z, t) + q(z, t) = A eκt + eκt e−2a

> Aeκt
> kf (z, ·)k2 (kp(·, t)k2 + kq(·, t)k2 )
Z
≥ f (z, y)(p + q) dy
Ω
Z
= f (z, y)S(y, t) dy, 0 < t < T.
Ω
Note that above inequalities hold for arbitrary positive constant A. After
choosing a and κ, let A satisfy 2Ae−a > supx∈Ω u0 (x). Then one has
n −a(r+1)
o
S(x, 0) = A ea(r−1)+e ≥ sup u0 (x).
x∈Ω
Hence S(x, t) is a supersolution to (1.1), and thus inequality (2.3) holds by

Theorem 1.1.
By Theorem 2.1, we have a supersolution for any T > 0. Hence the local
solution u on DT from Theorem 1.2 is bounded in DT for arbitrary T > 0,
and thus u can be extended to the whole time domain.
3. Decreasing property of boundary values.

In this section, boundary behavior of the solution to Problem (1.1) is studied.
The difference of largest and smallest boundary values grows exponentially
(inequality (3.4)). When the weights are identically zero on some part of
boundary, it is shown that the difference can be nonincreasing in Theorem
3.2.
From now on, it is assumed that µ ≥ 0. A lower bound of the solution of
Problem (1.1) with weights satisfying for each z ∈ ∂Ω,
(3.1) f (z, y) ≥ 0, y ∈ Ω and kf (z, ·)k1 > 1,
can be obtained by the following Theorem:
Theorem 3.1. If µ ≥ 0 and u is a solution to Problem (1.1) and the weight
f satisfies (3.1), then there are positive constants c2 and γ such that
(3.2) u(x, t) ≥ c2 eγt on DT .
Proof. Let v = u−1 , where u is the positive solution to Problem (1.1). By

denoting F (z) = kf (z, ·)k1 , we see that for arbitrary T > 0,
vt = u−2 (−∆u − µu)
|∇u|2
= ∆v − 2 − µv
u3
≤ ∆v − µv on DT ,
and for t > 0,
Z −1
v(z, t) = f (z, y)u(y, t) dy
Ω
Z
−1 f (z, y) −1
≤ F (z) u (y, t) dy.
Ω F (z)
224 SANGWON SEO
Since kf (z, ·)k1 /F 2 (z) < 1 for each z ∈ ∂Ω and µ ≥ 0, one has v ≤ c1 e−γt
for some positive constants c1 and γ by (2.2) and Theorem 1.1. Therefore,
the solution u to Problem (1.1) satisfies (3.2) with c2 = c−1
1 .
For weights satisfying (3.1), consider maximum and minimum of bound-
ary values;
M (t) = max u(z, t), m(t) = min u(z, t).
z∈∂Ω z∈∂Ω
Let M (t) = u(α, t), m(t) = u(β, t) for some α, β ∈ ∂Ω. Then, if
(3.3) f (α, y) > f (β, y) for each y ∈ Ω,
the difference δ(t) = M (t) − N (t) satisfies
Z
δ(t) = {f (α, y) − f (β, y)}u(y, t) dy
(3.4) Ω
≥ c{kf (α, ·)k1 − kf (β, ·)k1 }eγt ,

for some positive constant c by Theorem 3.1. Hence the difference of the
boundary values become exponentially large if (3.1) and (3.3) are satisfied.
This increasing property fails when some weights are identically zero, that
is, for some nonempty set Γ ⊂ ∂Ω, f (z, ·) ≡ 0 for z ∈ Γ. To see this, assume
that, for each z ∈ ∂Ω, f (z, y) ∈ Cy2 (Ω) and that
(3.5) f (z, y) ≡ 0 for each z ∈ ∂Ω, y ∈ ∂Ω.
By the assumption (3.5) and using integration by parts, we see that, for
ξ ∈ Γc ,
du(ξ, t)
dt
Z Z
=− ∇n f (ξ, ω)u(ω, t) dω + {∆y f (ξ, y) + µf (ξ, y)} u(y, t) dy
∂Ω Ω
(3.6) Z Z
= ∆y f (ξ, y) + µf (ξ, y) − f (ω, y)∇n f (ξ, ω) dω u(y, t) dy
Ω ∂Ω
Z
= Iξ (f ; y)u(y, t) dy.
Ω
Therefore, if Iξ (f ; y) ≥ 0 for each y ∈ Ω, then the boundary value u(ξ, t)
increases, and if Iξ (f ; y) ≤ 0, then the boundary value decreases. Since we
are interested in the weights for which the value u(ξ, t) decreases, consider
the nonnegative function g on Ω such that
(3.7) ∆g = λg in Ω,
and
(3.8) g=0 on ∂Ω.
Define the characteristic function on ∂Ω = Γ ∪ Γc by

0 for z ∈ Γ
(3.9) χ(z) =
1 for z ∈ Γc ,
and let
(3.10) h(z, y) = χ(z)g(y).
Theorem 3.2. If Γ is sufficiently large in the sense that
Z
(3.11) ∇n g(ω) dω is sufficiently large,
Γ
then, for sufficiently small µ ≥ 0, the boundary values of the solution to
Problem (1.1) with weights given by (3.10) are nonincreasing.
Proof. For ξ ∈ ∂Ω, Iξ (h, y) satisfies
Z
Iξ (h; y) = χ(ξ)∆g(y) + µχ(ξ)g(y) − χ(ω)χ(ξ)g(y)∇n g(ω) dω
∂Ω
Z
(3.12) = χ(ξ)∆g(y) + µχ(ξ)g(y) − χ(ξ) g(y)∇n g(ω) dω
Γc
Z
= χ(ξ) λ + µ − ∇n g(ω) dω g(y).
Γc
If the nonnegative constant µ is sufficiently small, then we can set Iξ (h; y)
as nonpositive by taking Γ as sufficiently large in the sense of (3.11) while
kh(z, ·)k1 are large for z ∈ Γc . This can be done because the eigenvalue
λ is fixed negative number and g is a smooth function. Therefore, by the
identities (3.6), we get the conclusion.
On the other hand, since g(y) is nonnegative for every y ∈ Ω, ∇n g(ω) in
(3.12) is nonpositive. Thus, if
µ > −λ,
then Iξ (h; y) > 0 on arbitrary proper subset Γc of ∂Ω. Therefore, u(ξ, t) is
increasing for ξ ∈ Γc .
References
[1] W.A. Day, Extensions of a property of the heat equation to linear thermoelasticity
and other theories, Quart. Appl. Math., 40(3) (1982/83), 319-330.
[2] , A decreasing property of solutions of parabolic equations with applications to
thermoelasticity, Quart. Appl. Math., 40(4) (1982/83), 468-475.
[3] A. Friedman, Monotonic decay of solutions of parabolic equations with nonlocal bound-
ary conditions, Quart. Appl. Math., 44(3) (1986), 401-407.
226 SANGWON SEO
[4] Keng Deng, Comparison principle for some nonlocal problems, Quart. Appl. Math.,
50 (1992), 517-522.
Received July 29, 1998 and revised September 3, 1998. This work is partially supported
by KOSEF.
Seoul National University

Seoul 151-742
Korea
E-mail address: sangwon@math.snu.ac.kr
Vol. 193, No. 1, 2000
GENERALIZED QUADRANGLES WEAKLY EMBEDDED

OF DEGREE 2 IN PROJECTIVE SPACE
Hendrik Van Maldeghem
In this paper, we classify all generalized quadrangles weakly

embedded of degree 2 in projective space. More exactly, given
a (possibly infinite) generalized quadrangle Γ = (P, L, I) and
a map π from P (respectively L) to the set of points (respec-
tively lines) of a projective space PG(V ), V a vector space
over some skew field (not necessarily finite-dimensional), such
that:
(i) π is injective on points,
(ii) if x ∈ P and L ∈ L with x I L, then xπ is incident with
Lπ in PG(V ),
(iii) the set of points {xπ | x ∈ P} generates PG(V ),
(iv) if x, y ∈ P such that y π is contained in the subspace of
PG(V ) generated by the set {z π | z is collinear with x in
Γ}, then y is collinear with x in Γ,
(v) there exists a line of PG(V ) not in the image of π and
which meets P π in precisely 2 points,
then we show that Γ is a Moufang quadrangle and we can ex-
plicitly describe the weak embedding of Γ in PG(V ). This
completes the classification of all weak embeddings of ar-
bitrary generalized quadrangles (using the classification of
Moufang quadrangles).
1. Introduction.
Weakly embedded polar spaces were introduced by Lefèvre-Percsy, see e.g.,
[4] (although she had a stronger notion of weak embedding, but it was
proved to be equivalent with the present one by Thas & Van Maldeghem
[11, Lemma 2]). In the same paper, she proves that the number of points
of a weakly embedded polar space Γ on a secant line (i.e., a line of the pro-
jective space not belonging to the polar space and meeting P π in at least
two points) is a constant (and hence does not depend on that line). Fol-
lowing Thas & Van Maldeghem [11], we call this constant the degree of the
weak embedding. In [3], Lefèvre-Percsy classifies the finite weakly embed-
ded generalized quadrangles (which are the nondegenerate polar spaces of
rank 2) in PG(3, q). All those thick weak embeddings have automatically
227
228 H. VAN MALDEGHEM
degree > 2. In Thas & Van Maldeghem [12], all weakly embedded general-
ized quadrangles in finite projective space are classified. Also, Steinbach &
Van Maldeghem [9] classify the weakly embedded generalized quadrangles
of degree > 2 in arbitrary projective space. In the present paper, we com-
plete the classification of all weakly embedded generalized quadrangles in
any projective space by considering the case of degree 2. This has been an
open problem for almost twenty years and it is a far-reaching generalization
of a result of Dienst [2], who classifies all full embeddings of generalized
quadrangles in arbitrary projective space. A full embedding satisfies con-
ditions (i), (ii) and (iii) (of the abstract) and the additional condition that
every point in PG(V ) of the image of every line of the quadrangle is also
the image of a point of the quadrangle. A direct and elementary argument
then shows that a full embedding also satisfies condition (iv). Hence every
full embedding is also a weak embedding. Dienst’s result says that only the
classical Moufang quadrangles turn up with their natural embedding in a
(possibly degenerate) polarity, see Tits [15]. Asking for a further general-
ization (i.e., embeddings satisfying only conditions (i), (ii) and (iii) in the
abstract above and calling this a lax embedding) is not reasonable, as is
evidenced by the fact that one can then laxly embed a freely constructed
finitely generated generalized quadrangle (in the sense of Tits [18]) in some
projective space; see Section 8 below. Hence, our result is the best one
can do and finishes the general problem. It also provides new and indepen-
dent proofs for the full case (Dienst [2]) and the finite case (Thas & Van
Maldeghem [12]).
Finding new techniques was essential because, unlike the finite case, there
are generalized quadrangles which can be weakly embedded in projective
space, but which do not admit a full embedding. In fact, all Moufang quad-
rangles can, up to duality, be weakly embedded, except for the exceptional
ones (see below, also for a list of the generalized quadrangles, Γ say, such that
Γ and the dual generalized quadrangle ΓD is weakly embeddable). Hence
the classification of weakly embedded general quadrangles requires methods
which are different from those used in the finite or full case. One of the
tools we use is the classification of all Moufang quadrangles, recently fin-
ished by Tits & Weiss [20], but not yet available in the literature. Without
invoking this classification, our result remains true if restricted to all known
generalized quadrangles. But we emphasize the fact that the part of the
classification that we use, namely, that every Moufang quadrangle is in a
well defined list (to the classes of Moufang quadrangles enumerated in Tits
[16], one has to add the so-called exceptional quadrangles of type F4 , discov-
ered by Richard Weiss in February 1997, and proved to be of exceptional
type by Mühlherr & Van Maldeghem [5]), is completely finished; the yet
unfinished parts in the manuscript of Tits & Weiss [20] merely concern the
existence problem, which does not affect our proof.
WEAKLY EMBEDDED QUADRANGLES 229
Note that results of Steinbach [7] and Thas & Van Maldeghem [11] treat
the same kind of question for polar spaces with some additional conditions.
In all cases, the assumptions imply that the polar space is classical (i.e.,
arises from a vector space with form). In the present paper, we hypoth-
esize an arbitrary generalized quadrangle weakly embedded of degree 2 in
arbitrary projective space and prove that it must belong to the class of so-
called Moufang quadrangles. Then we have to treat several classes (amongst
them the classical cases). In the course of our proof, we slightly improve
the result of Steinbach & Van Maldeghem [9] in that we determine when a
weak embedding π : Γ → PG(V ) is obtained from a full embedding in a
subspace of V (defined over a skew subfield of K) by extending the ground
field. In Section 8, we put together the results of the present paper with
those of Steinbach & Van Maldeghem [9] and list all generalized quadrangles
weakly embedded in projective space. We then also show that no further
generalization is possible.
So the eventual determination of all weakly embedded generalized quad-
rangles of degree 2 requires some knowledge about the classification of
Moufang quadrangles. We will introduce notation and repeat some known
results in the next section.
2. Definitions and Notation.

2.1. Generalized quadrangles. A generalized quadrangle Γ = (P, L, I )
is a point-line incidence geometry (where P is the set of points and L the
set of lines) satisfying the following two axioms:
(i) Each point is incident with t + 1 lines; each line is incident with s + 1
points; two distinct points are never incident with two distinct lines
(here s, t ≥ 1, possibly infinite).
(ii) If x is a point and L is a line not incident with x, then there is a unique
pair (y, M ) ∈ P × L for which x I M I y I L.
The pair (s, t) is usually called the order of Γ. If s, t > 1, then the quad-
rangle is said to be thick. Furthermore, we use standard terminology such
as collinear points, concurrent lines, etc. Also, there is a duality for gen-
eralized quadrangles: Every statement has a dual, i.e., if one interchanges
the names point and line (and the numbers s and t), then a (usually new)
statement is obtained. The dual of Γ is denoted by ΓD . Further, the line M
(respectively the point y) of (ii) is called the projection of L onto x (respec-
tively of x onto L). A subquadrangle Γ0 of Γ is called an ideal subquadrangle,
if all lines of Γ incident with a point of Γ0 belong to Γ0 as well. Dually, one
defines the notion of a full subquadrangle. In a generalized quadrangle, a
line L is called regular if for every line M not meeting L, the two lines L
and M are contained in a full subquadrangle with two lines per point.
Generalized quadrangles were introduced by Tits in [14]. For more infor-

mation, we refer to the monograph of Payne & Thas [6], to Thas [10], or
Van Maldeghem [21] (in the latter also the infinite case is covered).
There is no hope of classifying all generalized quadrangles (the situation is
more or less the same as for projective planes), as there are (many variations
of) free constructions of such geometries, see e.g., Tits [18]. Nevertheless, if
one imposes some extra conditions, then classification is possible. Two such
conditions are related to our Main Result, namely, the Moufang condition,
and the condition of being weakly embedded in a projective space.
2.2. Moufang quadrangles. Let Γ = (P, L, I ) be a thick generalized

quadrangle. We denote by Γ(a) the set of elements of Γ incident with the
element a (point or line). A point-elation is an automorphism of Γ fixing
the set Γ(x) ∪ Γ(y) ∪ Γ(L) elementwise, where x, y, x 6= y, are two distinct
points incident with the line L. Such a collineation is also called an (x, L, y)-
elation. If for some line M I x, M 6= L, the group of all (x, L, y)-elations
acts transitively on Γ(M ) \ {x}, then we say that (x, L, y) is a Moufang
path. Dually, one defines line-elations and Moufang paths (L, x, M ). Let
x, y ∈ P, L, M ∈ L. If the paths (x, L, y), for all choices of x I L I y, x 6= y
(respectively the paths (L, x, M ) for all choices of L I x I M , L 6= M ) are
Moufang paths, then we say that Γ is a half-Moufang quadrangle and that
all point-elation groups (respectively line-elation groups) act transitively. If
all paths (x, L, y) and all paths (L, x, M ) are Moufang paths, then we say
that Γ is a Moufang quadrangle.
The standard examples of Moufang quadrangles are the classical quad-
rangles, i.e., generalized quadrangles related to a vector space with a form.
Namely, let W be a (left) vector space over some skew field L endowed
with either a pseudo-quadratic form or a (σ, )-hermitian form in the sense
of Tits [15, §8]. Suppose that the from is nondegenerate and of Witt in-
dex 2. Define the geometry Γ with points and lines the 1-dimensional and
2-dimensional subspaces of W , where the form vanishes, and symmetrized
inclusion as incidence. Then Γ is a generalized quadrangle. (We neglect
the case that W is a 4-dimensional vector space endowed with an ordinary
quadratic form, in wich Γ is not thick.)
When Γ arises from a (σ, 1)-quadratic from with σ 6= 1, then we will call
such a quadrangle a hermitian quadrangle. When Γ arises from an ordinary
quadratic form, we have an orthogonal quadrangle. The dual of such a Γ is
a dual hermitian or a dual orthogonal quadrangle, respectively.
Proportional pseudo-quadratic froms give rise to the same generalized
quadrangle. We see as follows that a hermitian quadrangle also arises from
a (σ 0 , −1)-quadratic form. Choose a ∈ L such that c := a − aσ 6= 0. Then
0
qc is a (σ 0 , −1)-quadratic form, where tσ = c−1 tσ c for t ∈ L. Furthermore,
0
1 ∈ {t + tσ | t ∈ L}.
The mixed quadrangles are certain subquadrangles of orthogonal quad-

rangles defined over a (non-perfect) field of characteristic 2, see Section 7
below. There is a further class of Moufang quadrangles not related to
pseudo-quadratic or (σ, )-hermitian forms, the so-called exceptional quad-
rangles. We will not need an explicit description of these quadrangles (see
the reduction in 4.4).
The classification of Moufang quadrangles by Tits & Weiss [20] yields
that up to duality every Moufang quadrangle is isomorphic to a hermitian,
an orthogonal, a mixed or an exceptional quadrangle.
For later use, we say a generalized quadrangle Γ is a symplectic quad-
rangle, if Γ arises from a 4-dimensional vector space over a commutative
field endowed with an alternating form. A symplectic quadrangle is a dual
orthogonal quadrangle (associated to a 5-dimensional vector space), but it
is convenient to have a separate name for it.
For any (σ, )-hermitian form f , the radical of f is Rad(W, f ) = {v ∈ W |
f (v, w) = 0 for all w ∈ W }.
Let Γ be a generalized quadrangle and p a point in Γ. If a collineation
fixes every point collinear with p, then we call that collineation a central
collineation or a central elation. Dually, one defines an axial elation or axial
collineation. Every Moufang quadrangle contains, up to duality, nontrivial
central elations. This can easily be deduced from the main result of Tits
[19].
2.3. Weak embeddings of generalized quadrangles. Let PG(V ) be

a projective space, where V is a vector space over some skew field (not
necessarily finite-dimensional). The subspace of V generated by vectors
v1 , v2 , . . . , vn will be denoted by hv1 , v2 , . . . , vn i; we also sometimes write
Kv for hvi, when the vector space is defined over K and we want to make
this clear.
Let Γ be a generalized quadrangle with point set P, line set L and inci-
dence relation I. Then we say that Γ is weakly embedded in PG(V ) if there
exists a map π from P (respectively L) to the set of points (respectively
lines) of PG(V ), such that the following conditions are satisfied:
(i) π is injective on points,
(ii) if x ∈ P and L ∈ L with x I L, then xπ is incident with Lπ in PG(V ),
(iii) the set of points {xπ | x ∈ P} generates PG(V ),
(iv) if x, y ∈ P such that y π is contained in the subspace of PG(V ) gen-
erated by the set {z π | z is collinear with x in Γ}, then y is collinear
with x in Γ.
Condition (iv) may be replaced by the (a priori weaker) condition (see
Thas & Van Maldeghem [13], Corollary 1):
(iv)0 for each point x ∈ P, the set {z π | z is collinear with x in Γ} does not
generate PG(V ).
Another equivalent statement is: The subspace spanned by {z π | z collinear

with x in Γ} is a hyperplane of PG(V ).
The map π is called the weak embedding. It will sometimes be convenient
to see a weak embedding as an injective morphism from the point-line geom-
etry Γ to the geometry of 1- and 2-dimensional subspaces of a vector space
(and then to write π(x) instead of xπ for a point x). Also, for a given weak
embedding π, we will denote by Γπ the quadrangle whose points and lines
are the images under π of the points and lines of Γ. The quadrangle Γπ is a
subgeometry of PG(V ).
Let π be a weak embedding of Γ. A line of PG(V ) which intersects the
set of points of Γπ in at least two elements, and which is not a line of Γπ , is
called a secant line. It has been shown by Lefèvre-Percsy [4] that the number
of points of Γπ on a secant line is a constant, and we call that constant the
degree. In this paper, we will mainly be concerned with weakly embedded
quadrangles of degree 2.
A full embedding π of a generalized quadrangle Γ in PG(V ) is a weak
embedding such that all points of PG(V ) on a line of Γπ are also points of
Γπ (this definition has been justified in the Introduction).
3. Main Result.
Main Result. Let π be a weak embedding of degree 2 of a thick generalized
quadrangle Γ in the projective space PG(V ), where V is a vector space over
the skew field K (not necessarily finite-dimensional). Then Γ is a Moufang
quadrangle. Up to isomorphism, we have the following cases:
(1) Γ is an orthogonal quadrangle (arising from an ordinary quadratic

form) and the weak embedding is induced by an injective semi-linear
mapping. In other words, there is a (commutative) subfield F of K, a
subspace V0 of V (viewed as vector space over F) and a quadratic form
Q : V0 → F of Witt index 2, such that for each point x of Γ there is a
unique point Fx0 of the associated quadric with π(x) = Kx0 , and every
point of the quadric arises in this way.
If the radical of the corresponding bilinear form has (vector) dimen-
sion at most 1 (which happens for instance whenever charK 6= 2), then
an F-basis of V0 is a K-basis of V . Hence, in this case, π is obtained
from a full embedding in PG(V0 ) by extending the ground field.
(2) Γ is a mixed quadrangle and the weak embedding is induced by an in-
jective semi-linear mapping. In other words, there is a (commutative)
subfield F of K, a subspace V0 of V (viewed as vector space over F)
and a quadratic form Q : V0 → F of Witt index 2, such that for each
point x of Γ there is a unique point Fx0 of the associated quadric with
π(x) = Kx0 . Not every point of the quadric arises in this way.
(3) Γ is the unique generalized quadrangle of order (2, 2) (symplectic quad-

rangle W (2) over GF(2)) and the weak embedding is the universal
weak embedding in the sense of Thas & Van Maldeghem [12], with
charK 6= 2.
In particular this means that Γ can never be a hermitian quadrangle, nor

can Γ be isomorphic or dual to an exceptional quadrangle.
The Main Result will follow from the reduction in 4.4, and from Lemma
5.1, Lemma 5.5, Lemma 6.3, Lemma 6.4 and Lemma 7.2.
In the case where Γ is an orthogonal quadrangle weakly embedded (of
degree 2) in PG(V ), let Γ be associated to a vector space W over the field L
and to the ordinary quadratic form q on W . In Case (1) of the Main Result,
there exists an embedding α : L → K and an injective semi-linear mapping
ϕ : W → V (with respect to α) such that π(Lw) = Kϕ(w) for all points Lw
of Γ (which means that the weak embedding π : Γ → PG(V ) is induced by
a semi-linear mapping). In particular, Γ is fully embedded in the projective
space PG(ϕ(W )), where ϕ(W ) is a vector space over the (commutative)
subfield α(L) of K.
A central tool in the classification is the notion of regular lines (as de-
fined above) in generalized quadrangles. Indeed, it is easily seen (as we will
show in Lemma 4.2) that every line of a weakly embedded quadrangle of de-
gree 2 is regular. Once the Moufang condition is proved, this will facilitate
considerably the rest of the proof.
The paper is organized as follows. In the next section we reduce the
problem to Moufang quadrangles and discuss which classes of Moufang
quadrangles have regular lines. To prove the Moufang condition, we show
that any generalized quadrangle weakly embedded of degree 2 admits ax-
ial elations, see Lemma 4.2. (Compare the approach in Steinbach & Van
Maldeghem [9] and Steinbach [8], where the weakly embedded generalized
quadrangles admit central elations.) In Section 5, we handle the case of or-
thogonal quadrangles (Cases (1) and (3) of the Main Result). In Section 6,
we show that the only weak embeddings of degree 2 of the dual hermitian
quadrangles are given by semi-linear mappings of orthogonal quadrangles.
In Section 7, we deduce from Steinbach & Van Maldeghem [9, (6.1.3)] that
for any weakly embedded mixed quadrangle Γ of degree 2, Γπ is part of the
null set of a quadratic form of Witt index 2 in a subspace V0 of V (over a
subfield F of K). This is Case (2) of the Main Result. In Section 8, we state
the complete classification of generalized quadrangles weakly embedded in
projective space independent of the degree of the weak embedding. Then
we show that a further generalization to lax embeddings is not possible.
Furthermore, we mention a corollary on generalized quadrangles Γ with the
property that both Γ and the dual ΓD are weakly embeddable in projective
space.
We remark that part of the proof of the Main Result is contained in other
papers. Indeed, the classification of weak embeddings of generalized quad-
rangles arising from a vector space with a form has been done in Steinbach
[7] and Steinbach & Van Maldeghem [9], independently of the degree of the
weak embedding. The latter reference also covers the mixed case. It is only
when it became clear to us that for degree 2, the Moufang condition can
be proved, that a complete classification came into reach. This reduction to
the Moufang quadrangles is the crux of the proof. It is based on a lemma
of Steinbach [8].
4. Reduction to Moufang quadrangles.

For any point p of a generalized quadrangle Γ weakly embedded in PG(V ),
we denote by ξp the unique hyperplane of PG(V ) spanned by the points xπ ,
with x collinear with p. A special linear transformation of V is an element
of SL(V ) (the subgroup of the group of all invertible linear mappings from
V to V , which is generated by the transvections).
The following lemma is due to Steinbach [8, Proposition 2.1]. We phrase
it a little differently, according to our needs.
Lemma 4.1 ([8]). Let Γ be a generalized quadrangle weakly embedded of
degree 2 in PG(V ), where V is a vector space over some skew field K. Let
L be any line of Γ. Except for the universal weak embedding of W (2), there
is a unique subspace ξL of codimension 2 contained in all ξp with p on L.
The subspace ξL in Lemma 4.1 is ξL = ξp ∩ ξq , where p and q are different
points on L.
Lemma 4.2. Let Γ be a generalized quadrangle weakly embedded of degree
2 in PG(V ), where V is a vector space over some skew field K. Then all
lines of Γ are regular. Also, Γ is a half-Moufang quadrangle. More exactly,
for every line L of Γ, and every pair (x, x0 ) of collinear points of Γ such that
xx0 is concurrent with L and x, x0 not on L, there exists an axial elation
with axis L mapping x to x0 . Moreover, this elation is induced by a special
linear transformation of V .
Proof. First, if the weak embedding is the universal one of W (2) (with
charK 6= 2), then it is proved in Van Maldeghem [21, Section 8.6] that
all collineations of Γ are induced by special linear transformations of V . Al-
ternatively, this also follows from Thas & Van Maldeghem [13]. Hence from
now on, we may assume that for any line L of Γ, there is a unique subspace
ξL of codimension 2 contained in all ξy , for all points y of L. To simplify
notation, we identify Γ with Γπ in this proof.
Let M be any line of Γ not concurrent with L. Let PG(3, K) be the
3-dimensional projective space generated by L and M . Then PG(3, K)
intersects Γ in a full subquadrangle Γ0 weakly embedded in PG(3, K). Let
t + 1 be the number of lines of Γ0 through a point of Γ0 , then considering

u⊥ ∩ v ⊥ for two noncollinear points of Γ0 , we obtain a line of PG(3, K),
which is not a line of Γ0 , meeting Γ0 in precisely t + 1 points. Since the
degree of the weak embedding is 2, we obtain t + 1 = 2 and L is a regular
line.
Let a be the intersection of the lines xx0 and L. Then a 6= x, x0 , since
x, x0 are not on L. Let b be any other point of Γ on L. Let L0 be any line
of Γ through b, with L0 6= L. Let y and y 0 be the unique points of Γ on L0
collinear with x and x0 , respectively. Notice that y and y 0 are not contained
in ξa . Hence there exists a unique collineation θ1 of PG(V ) fixing all points
of ξa , stabilizing all subspaces through b, and mapping y to y 0 . Similarly,
there is a unique collineation θ2 of PG(V ) fixing all points of ξb , stabilizing
all subspaces through a, and mapping x to x0 . We put θ = θ1 θ2 . Clearly
θ fixes all points of ξL (= ξa ∩ ξb ), it maps x to x0 , and it maps y to y 0 .
Moreover, every subspace of PG(V ) containing L is stabilized by θ, since it
is stabilized by both θ1 and θ2 .
Now let z be any point of Γ. We show that z θ is a point of Γ. This is
clear if z lies on L. Now suppose that z is incident with xy. Since the line
L is regular, there is a unique point z 0 on x0 y 0 collinear in Γ with z and the
line zz 0 meets L in, say, the point c of Γ. Since θ fixes the plane hL, zi, it
maps the intersection z of xy and hL, zi onto the intersection z 0 of x0 y 0 and
hL, zi = hL, z 0 i.
Notice that the restriction of θ to ξc fixes all points in the hyperplane ξL
of ξc (see Lemma 4.1); hence it must fix all subspaces of ξc through some
point of ξc . Clearly this point must belong to zz 0 ; also it must belong to
L, since every subspace through L is stabilized. Hence this point is c. We
conclude that θ leaves invariant every line of Γ meeting L.
Now let z be collinear with x, but z not on ax. We may suppose that z
is not collinear with y. Let again c 6= a be the projection of z onto L. Also,
let w be the projection of d onto xz, where d is a point of Γ on L different
from a and c, and let w0 be the projection of x0 onto dw. By the regularity
of lines, the line x0 w0 meets the line cz in, say, the point z 0 . Since wx is the
unique line of PG(V ) through x meeting both cz and dw, and since w0 x0 is
the unique line of PG(V ) through x0 meeting both cz and dw, the image of
wx under θ is w0 x0 . It follows that wθ = w0 and z θ = z 0 .
Now let z be arbitrary, but not collinear with either x or a. We may
assume that z is not on L. Let Nz be the projection of L onto z. Let N be
any line on z different from Nz and from the projection of ax onto z. Let w
be the projection of x onto N . Then w is collinear with x and not on ax,
hence wθ belongs to Γ by the previous step. Let Na be the projection of
N onto a. As before, N θ is the unique line of PG(V ) through wθ meeting
both Na and Nz (which are both fixed by θ). By the regularity of L, N θ is
a line of Γ and it meets Nz in the point z 0 of Γ, with clearly z 0 = z θ .
Now let z be collinear with a. Then we interchange the roles of (a, x, x0 )

and (b, y, y 0 ), and we conclude that also in this case z θ belongs to Γ.
Hence θ preserves Γ and induces clearly an axial elation in Γ with axis L.
The lemma is proved.
2 in PG(V ), for some skew field K. Then Γ is a Moufang quadrangle and
the little projective group of Γ is induced by PSL(V ).
Proof. We can copy word for word the proof of Lemma 4.0.2 of Steinbach
& Van Maldeghem [9].
Reduction 4.4. We have shown that, in order to prove the Main Result, we
have to classify the weak embeddings of degree 2 of Moufang quadrangles.
Candidates must have regular lines. Using the classification of Moufang
quadrangles as carried out by Tits & Weiss [20], Van Maldeghem discusses
the regularity of points and lines of Moufang quadrangles in [21], Table
5.1. It follows that the only Moufang quadrangles with regular lines are
the mixed quadrangles, the orthogonal quadrangles, and the duals of some
hermitian quadrangles. The classification of weakly embedded mixed and
orthogonal quadrangles has been treated in general (for every degree) in
Steinbach & Van Maldeghem [9]. Sections 5 and 7 take care of them. In
Section 6, we look at the dual hermitian case.
5. Weakly embedded orthogonal quadrangles of degree 2.

In this section, we handle the weakly embedded orthogonal quadrangles of
degree 2. Let Γ be a thick orthogonal quadrangle with natural embedding
in the projective space PG(W ), where W is a vector space over the com-
mutative field L. So Γ arises from an ordinary quadratic form q on W . Let
there be given a weak embedding π : Γ → PG(V ), where V is a vector space
over the skew field K. Then π is induced by a semi-linear mapping (with
respect to the embedding α : L → K) by Steinbach [7], Steinbach & Van
Maldeghem [9, (5.1.1)] (apart from the universal weak embedding of W (2)).
We show that there is a (commutative) subfield F of K and a subspace
V0 of V (viewed as vector space over F) such that the point set of Γπ is the
point set of a projective quadric of Witt index 2 in PG(V0 ).
Lemma 5.1. Let Γ be an orthogonal quadrangle arising from the vector
space W over the field L, endowed with the ordinary quadratic form q. As-
sume that Γ is weakly embedded of degree 2 in the projective space PG(V ),
where V is a vector space over the skew field K (not the universal weak
embedding of W (2)).
Then there exists a commutative subfield F of K, a subspace V0 of V (over
F) and a quadratic form Q : V0 → F of Witt index 2, such that for each point
x of Γ there is a point Fx0 of the associated quadric with π(x) = Kx0 . Every
point of the quadric arises in this way.
Proof. By Steinbach [7], Steinbach & Van Maldeghem [9, (5.1.1)], the weak
embedding π is induced by a semi-linear mapping. Hence there exists an
embedding α : L → K and a semi-linear mapping ϕ : W → V such that
π(Lw) = Kϕ(w), for all points Lw of Γ.
Since the degree of π is 2, we may conclude as follows that ϕ is injective.
By Steinbach & Van Maldeghem [9, Subsection 5.3], we know that kerϕ ⊆
Rad(W, f ), where f is the bilinear form associated to the quadratic form q
defining Γ. If 0 6= r ∈ Rad(W, f ) with ϕ(r) = 0, then there exist vectors
x, y ∈ W such that Lx, Ly and L(x + y + r) are (noncollinear) points of
Γ. But then π(L(x + y + r)) is a third point on the secant line spanned by
π(Lx) and π(Ly), a contradiction.
We consider the subspace V0 := ϕ(W ) of V (over the commutative subfield
F := α(L) of K). We define a quadratic form Q : V0 → F of Witt index
2, by Q(ϕ(w)) := α(q(w)), for w ∈ W . This yields an orthogonal space
isomorphic to W . Now Lemma 5.1 is obvious.
The proof of Lemma 5.1 shows that the semi-linear mapping ϕ inducing
π is injective. Hence for each L-basis B of W , the set ϕ(B) is an F-basis of
V0 . Our next aim is to decide when an F-basis of V0 is a K-basis of V in
Lemma 5.1.
General setting 5.2. We start with a more general setting (to prove the
next lemma in full generality). Let W be a (left) vector space over some
skew field L endowed with one of the following nondegenerate forms of Witt
index 2:
(a) a pseudo-quadratic form q on W (with associated (σ, )-hermitian form
f ),
(b) a (σ, )-hermitian form f on W with Λmin := {c − cσ | c ∈ L} =
{c ∈ L | cσ = −c} =: Λmax .
By Γ we denote the associated generalized quadrangle. Let π : Γ →
PG(V ) be a weak embedding, where V is a vector space over the skew field
K. We assume that π is induced by a semi-linear mapping ϕ : W → V (with
respect to the embedding α : L → K). (By Steinbach [7], Steinbach & Van
Maldeghem [9, (5.1.1)] this is true up to few exceptions.) We set F := α(L)
and V0 := ϕ(W ).
Recall that for any point p of the generalized quadrangle Γ, we denote
by ξp the unique hyperplane of PG(V ) spanned by the points π(x), with x
collinear with p. A vector w ∈ W , on which the form vanishes (i.e., q(w) = 0
or f (w, w) = 0, respectively), is called singular. This happens if and only if
Lw is a point of Γ.
Lemma 5.3. In the general setting of 5.2 the following holds: If Lp is a

point of Γ and w ∈ W such that ϕ(w) is contained in ξLp , then f (w, p) = 0.
Proof. If w is singular, then the claim follows from the weak embedding
axiom 2.3 (iv). Hence we may assume that w is nonsingular and w 6∈
Rad(W, f ). Then there exist singular points Lx, Ly with f (x, y) = 1 such
that w ∈ Lx + Ly. We may assume f (x, p) 6= 0 and f (y, p) 6= 0. Indeed,
if f (x, p) = 0, then ϕ(x) and ϕ(w), and therefore also ϕ(y), are contained
in ξLp . This yields f (y, p) = 0 by the weak embedding axiom and hence
f (w, p) = 0.
Set H1 = hp, p0 i with p0 singular and f (p, p0 ) = 1. Then W = H1 ⊥ H1 ⊥
with H1 ⊥ spanned by singular points. We write Lx = L(λp + p0 + a1 ),
Ly = L(µp + p0 + a2 ) with λ, µ ∈ L and a1 , a2 ∈ H1 ⊥ . Then there are
0 6= c, d ∈ L with w = c(λp + p0 + a1 ) + d(µp + p0 + a2 ). Applying ϕ,
this yields that (α(c) + α(d))ϕ(p0 ) is contained in ξLp . Thus c + d = 0 and
f (w, p) = 0.
Lemma 5.4. In the general setting of 5.2 the following holds: If
dim Rad(W, f ) ≤ 1, n ∈ N and w1 , . . . , wn ∈ W such that ϕ(w1 ), . . . , ϕ(wn )
are linearly independent over F, then ϕ(w1 ), . . . , ϕ(wn ) are linearly inde-
pendent over K. In particular, an F-basis of V0 is a K-basis of V in this
case.
Proof. We use induction on n, the case n = 1 is trivial. Let n ≥ 2 and
w1 , . . . , wn ∈ W such that ϕ(w1 ), . . . , ϕ(wn ) are linearly independent over
F. Then w1 , . . . , wn are linearly independent over L. There exists a sin-
gular point p in W with hw1 , . . . , wn i 6⊆ p⊥ . (Otherwise hw1 , . . . , wn i ⊆
Rad(W, f ), since W is spanned by its singular points. But this is a con-
tradiction to the assumption on dim Rad(W, f ).) Write hw1 , . . . , wn i =
hu1 , . . . , un i where u1 , . . . , un−1 ∈ p⊥ and un 6∈ p⊥ . Then X := hϕ(w1 ), . . . ,
ϕ(wn )iK = hϕ(u1 ), . . . , ϕ(un )iK . Further, X 6⊆ ξp by Lemma 5.3. We have
hϕ(u1 ), . . . , ϕ(un−1 )iK ⊆ X ∩ ξp and the latter is a hyperplane of X. Since
hϕ(w1 ), . . . , ϕ(wn )iF = hϕ(u1 ), . . . , ϕ(un )iF , we see that ϕ(u1 ), . . . , ϕ(un )
are linearly independent over F. Hence by induction ϕ(u1 ), . . . , ϕ(un−1 ) are
linearly independent over K and n − 1 ≤ dim (X ∩ ξp ) ≤ n − 1. But then
X = (X ∩ ξp ) ⊕ hϕ(un )iK is n-dimensional.
For weakly embedded orthogonal quadrangles, we have shown:
Lemma 5.5. Let Γ be an orthogonal quadrangle arising from the vector
space W over the field L, endowed with the ordinary quadratic form q (with
associated bilinear form f ). Let π : Γ → PG(V ) be a weak embedding of
degree 2, where V is a vector space over the skew field K. Assume that π is
induced by the semi-linear mapping ϕ : W → V (with respect to α : L → K).
If Rad(W, f ) has dimension at most 1 (which is true whenever charK 6= 2),
then for any L-basis B of W , the set ϕ(B) is a K-basis of V .
Hence weak embeddings of degree 2 of orthogonal quadrangles are as

described in Cases (1) and (3) of the Main Theorem. The next example
shows that we cannot drop the assumption on the dimension of the radical
Rad(W, f ).
Example 5.6. Let L be a nonperfect field of characteristic 2. We consider
the symplectic quadrangle over L and pass to an isomorphic quadrangle Γ
arising from a quadratic form in vector space dimension 4 + dimL2 L, see
Cohen [1, (3.23), (3.27)]. If we extend scalars to the algebraic closure K of
L, then Γ is weakly embedded in vector space dimension 5.
6. Weakly embedded dual hermitian quadrangles of degree 2.

In this section, we handle the case of weakly embedded dual hermitian quad-
rangles of degree 2. We prove that there are exactly two possibilities, oc-
curring from exceptional isomorphisms between certain dual hermitian and
orthogonal quadrangles.
Notation 6.1. Let ∆ be a hermitian quadrangle with natural embedding
in the projective space PG(W ), where W is a (left) vector space over the
skew field L. This means that ∆ consists of the singular points and lines
with respect to a (σ, −1)-quadratic form q with 1 ∈ Λmin . (Recall that for
= −1, we have Λmin = {c + cσ | c ∈ L} and Λmax = {c ∈ L | c = cσ }.)
If Λmin = Λmax (which can only fail when charL = 2 and σ fixes the center
Z(L) of L elementwise), then ∆ coincides with the polar space arising from
the (σ, −1)-hermitian form f associated to q. If x, y ∈ W with q(x) = 0,
q(y) = 0 and f (x, y) = 1, then (x, y) is called a hyperbolic pair.
Remark 6.2. For any generalized quadrangle Γ weakly embedded of de-
gree 2 in a projective space PG(V ) (with weak embedding π), we use the
following method to calculate image points: Recall from Lemma 4.2 that
each line of Γ is regular. We write the nine points of a 3 × 3-grid of points of
Γ in a 9-tuple consisting of the first, second and third row. Here · stands for
a point without name. If (x1 , x2 , x3 ; y1 , y2 , ·; z1 , z2 , z3 ) is such a 3 × 3-grid,
then π(z3 ) = hπ(z1 ), π(z2 )i ∩ hπ(x3 ), π(y1 ), π(y2 )i.
2 in the projective space PG(V ), where V is a vector space over the skew
field K. Assume that ΓD is a hermitian quadrangle with natural embedding
in the projective space PG(W ), where W is a vector space over the skew
field L (see Notation 6.1).
If the vector space dimension of W is 4, then Λmin is a field and
(i) L is a commutative separable quadratic extension of Λmin (here Λmin =
Λmax ), or
(ii) L is a quaternion division ring over Λmin with σ its standard involution
(here Λmin ⊂ Λmax in characteristic 2).
In both cases Γ is isomorphic to the orthogonal quadrangle associated to

the vector space L×Λmin 4 (over Λmin ) with the quadratic form (x0 , x1 , x2 , x3 ,
x4 ) 7→ x0 x0 σ − x1 x3 + x2 x4 , σ as above.
Notice that x0 x0 σ = x0 · 1 · x0 σ ∈ Λmin , since 1 ∈ Λmin . The associated
bilinear form is b(x, y) = x0 y0 σ + y0 x0 σ − x1 x3 − y1 y3 + x2 x4 + y2 y4 with
trivial radical. The lemma reduces the dual hermitian case with dimW = 4
to certain orthogonal quadrangles considered in the previous section.
Proof. We apply the method of Remark 6.2. The following is inspired by

the proof of Tits [15, (10.2)]. By Tits [15, (10.5), (10.9)], the result follows
from the assertion that all maps from Λmin to Λmin of the form λ 7→ cσ λc,
where 0 6= c ∈ L, commute with each other.
The expression cσ λc occurs in the following construction. We write W =
hv1 , w1 i ⊥ hv2 , w2 i with hyperbolic pairs (vi , wi ), i = 1, 2. For each point x
and each line L of the hermitian quadrangle ΓD , we set ρx (L) := hx, L∩x⊥ i.
For λ ∈ Λmin and 0 6= c ∈ L, we obtain
ρcw1 +v2
hv1 , λv2 + w2 i 7→ hcw1 + v2 , c−σ v1 + λv2 + w2 i
ρ w2
7→ hw2 , −λcw1 + c−σ v1 i
ρw1 +v2
7→ hw1 + v2 , c−σ w2 − λcw1 + c−σ v1 i
ρv1
7 → hv1 , cσ λcv2 + w2 i
=: τc (hv1 , λv2 + w2 i).
In view of the above, we have to show that
(∗) τd (τc (hv1 , λv2 + w2 i)) = τc (τd (hv1 , λv2 + w2 i))

for λ ∈ Λmin , 0 6= c, d ∈ L.
We prove (∗) in several steps by calculation in Γ. Let 0 6= c ∈ L be fixed.

We choose notation such that the apartment (v1 , w2 , w1 , v2 ) of the hermi-
tian quadrangle ΓD corresponds to the apartment (p, q, t, z) of Γ (this means
that v1 corresponds to the line pq, further hv1 , w2 i corresponds to q and so
on). The line hv1 , λv2 + w2 i corresponds to some point on pq. The points
cw1 + v2 , w1 + v2 correspond to lines through z (different from pz and zt).
By ea and a, respectively, we denote the projections of q onto these lines.
We choose coordinates in Γ as follows. Let y1 be a third point on pz. We
set y2 := qa ∩ y1 ⊥ , e := zt ∩ y2 ⊥ , f := pq ∩ e⊥ . Then y2 6= a, q, e 6= z, t and
f 6= p, q. We choose vectors p0 , q 0 , t0 , z 0 ∈ V such that
π(p) = hp0 i, π(z) = hz 0 i, π(y1 ) = hp0 − z 0 i,
π(q) = hq 0 i, π(a) = ha0 i, π(y2 ) = ha0 − q 0 i,
π(t) = ht0 i, π(e) = ht0 + z 0 i.
Then there exists 0 6= α ∈ K such that π(f ) = hq 0 − αp0 i. Replacing p0 by

α−1 p0 , z 0 by α−1 z 0 and t0 by α−1 t0 , we may assume that π(f ) = hq 0 − p0 i.
Let c1 := qt ∩ y1 ⊥ . Then π(c1 ) = hq 0 + t0 i, using the method of Remark 6.2
with (p, q, f ; y1 , c1 , ·; z, t, e).
Set ye2 := qe a ∩ y1 ⊥ , ee := zt ∩ ye2 ⊥ , fe := pq ∩ ee⊥ . As above, we choose
00
vectors v in V spanning the images under π of p, . . . , fe. Then π(y1 ) =
hp0 − z 0 i = hp00 − z 00 i and π(c1 ) = hq 0 + t0 i = hq 00 + t00 i. Hence there exist
0 6= α, β ∈ K such that p00 = αp0 , z 00 = αz 0 , q 00 = βq 0 , t00 = βt0 . We set
γ := α−1 β. Then γ is a scalar in K depending on c ∈ L (via e a). Note that
γ is unique and well defined since one can calculate that π(f ) = hγq 0 − p0 i.
e
We denote by rλ the point on pq corresponding to the line hv1 , λv2 + w2 i
of the hermitian quadrangle ΓD . Let µ := µ(λ) ∈ K with π(rλ ) = hµp0 + q 0 i.
We name by u1 , . . . , u4 the points of Γ that correspond to the lines in the
hermitian quadrangle ΓD occurring in the above calculation of τc (hv1 , λv2 +
w2 i). Then
a ∩ rλ ⊥ , u2 = qt ∩ u1 ⊥ , u3 = za ∩ u2 ⊥ , u4 = pq ∩ u3 ⊥ .
u1 = ze
With β, γ from above, we obtain
π(u1 ) = hµz 0 + β −1 a00 i, using (rλ , p, q; ·, y1 , ye2 ; u1 , z, e
a),
π(u2 ) = hµγt0 − q 0 i, using a, z; ·, ye2 , ee; u2 , q, t),
(u1 , e
(∗∗)
π(u3 ) = hµγz 0 + a0 i, using (u2 , q, t; ·, y2 , e; u3 , a, z),
π(u4 ) = hµγp0 + q 0 i, using (u3 , z, a; ·, y1 , y2 ; u4 , p, q).
This yields that there is a scalar µ ∈ K depending on λ ∈ Λmin and a
scalar γ ∈ K depending on 0 6= c ∈ L, such that the mapping hv1 , λv2 +
w2 i 7→ τc (hv1 , λv2 + w2 i) in the hermitian quadrangle ΓD reads in Γ as
hµp0 + q 0 i 7→ hµγp0 + q 0 i. Let 0 6= d ∈ L and denote by δ the corresponding
scalar in K. To prove (∗), we have to show that the following equation holds
in Γ:
h(µγ)δp0 + q 0 i = h(µδ)γp0 + q 0 i for all µ = µ(λ), λ ∈ Λmin .
We see that γδ = δγ as follows. If we let vary λ in Λmin \ {0} in the line
hv1 , λv2 + w2 i of the hermitian quadrangle ΓD , then the corresponding point
rλ in Γ reaches all points on pq, different from p, q. Hence there exists
λ0 ∈ Λmin with rλ0 = f . Since π(f ) = hq 0 − p0 i, we have µ(λ0 ) = −1.
The calculation in (∗∗) with µ = −1 and γ and δ, respectively, shows that
there are points r, g on pq and qt, respectively, with π(r) = h−γp0 + q 0 i and
π(g) = h−δt0 − q 0 i. The projection s of r onto zt satisfies π(s) = ht0 + γz 0 i
(use (r, p, q; ·, y1 , c1 ; s, z, t)). The projection h of g onto pz satisfies π(h) =
hp0 − δz 0 i (use (f, p, q; ·, h, g; e, z, t)). The calculation of the remaining point
j in the grid (r, p, q; j, h, g; s, z, t) yields γδ = δγ.
Hence (∗) holds, and we obtain Lemma 6.3. This finishes the determina-
tion of Γ in the case where W is 4-dimensional.
The final step in this section is to show that in the situation of Lemma
6.3 the vector space W is necessarily 4-dimensional.
2 in the projective space PG(V ), where V is a vector space over the skew
field K. Assume that ΓD is a hermitian quadrangle with natural embedding
in the projective space PG(W ), where W is a vector space over the skew
field L (see Notation 6.1).
Then the vector space W is 4-dimensional. Hence the possibilities for Γ
and ΓD are determined by Lemma 6.3.
Proof. We identify Γ with Γπ in this proof. Since Γ is weakly embedded
of degree 2 in PG(V ), all lines of Γ are regular (see Lemma 4.2). Hence
ΓD has regular points. By the discussion in Van Maldeghem [21], Table
5.1, we may conclude that W = H1 ⊥ H2 ⊥ Rad(W, f ), where H1 , H2 are
hyperbolic lines and f is the underlying (σ, −1)-hermitian form.
We set ∆ = ΓD . Let ∆0 be the full subquadrangle of ∆ obtained by
intersecting the natural embedding of ∆ with PG(H1 ⊥ H2 ). Our aim is to
show that ∆0 = ∆, then W = H1 ⊥ H2 is a 4-dimensional vector space.
Let Γ0 = (∆0 )D . Then Γ0 is an ideal subquadrangle of Γ. Further, Γ0 is
weakly embedded in PG(V ). To prove this, let p be a point of Γ0 . Let ξp be
spanned by the set of points of Γ collinear with p in Γ and let ξp0 be spanned
by the set of points of Γ0 collinear with p in Γ0 . Then ξp0 is contained in ξp ,
which is a hyperplane of PG(V ). If a point x of Γ is in ξp , then x is collinear
with p in Γ. Since Γ0 is an ideal subquadrangle of Γ, the line on p and x is
a line of Γ0 and x is collinear with p in Γ0 . Hence ξp0 = ξp . For q in Γ0 , q not
collinear with p in Γ, this yields PG(V ) = hξp , qi = hξp0 , qi and the point set
of Γ0 generates PG(V ).
Hence we may apply Lemma 6.3 for the weak embedding π : Γ0 → PG(V ).
We obtain that Λmin is a field and L is quadratic or quaternion over Λmin .
Now let ∆00 be a symplectic subquadrangle of ∆0 , obtained by restricting
scalars to Λmin . Then ∆00 is an ideal subquadrangle of ∆0 . Hence Γ00 = (∆00 )D
is a full subquadrangle of Γ0 and Γ00 is weakly embedded in PG(V 0 ), where
the latter is spanned by the set of points of Γ00 . (Let p be a point of Γ00 .
If PG(V 0 ) is contained in ξp00 , notation as above, then every point of Γ00 is
collinear in Γ with p. This is a contradiction, since Γ00 contains ordinary
quadrangles.)
Now PG(V 0 ) meets Γ in a full subquadrangle Γ000 , which is weakly embed-
ded in PG(V 0 ). (Let p be a point of Γ000 . With notation as above, PG(V 0 ) is
not contained in ξp000 . If p is in Γ00 , then ξp00 = ξp000 is a hyperplane of PG(V 0 ).)
This implies that Γ00 is an ideal subquadrangle of Γ000 . But all points of Γ00
(a dual symplectic, hence mixed quadrangle) are regular, hence Γ000 has also
regular points. Since all lines of Γ000 are regular (this even holds for Γ), Γ000
is a mixed quadrangle.
Now the symplectic quadrangle ∆00 is a full subquadrangle of the dual of

Γ000 , with the latter a mixed quadrangle. That implies Γ000 = Γ00 (indeed,
with the notation of the next section, any full subquadrangle of the mixed
quadrangle Q(L0 , L2 ; Λ0 , Λ2 ) can be written as Q(L0 , L20 ; Λ0 , Λ20 ), with L20 ⊆
Λ0 ⊆ L0 ⊆ Λ0 ⊆ L0 and L0 ⊆ L and Λ0 ⊆ Λ, by Tits [17]; the latter is a
symplectic quadrangle only if Λ20 = L20 = L0 = Λ0 , implying Λ2 = L2 = L0 ).
So Γ00 is a full subquadrangle of Γ. Since it is also a full subquadrangle
of Γ0 , we deduce that Γ0 is a full subquadrangle of Γ. But it is also an ideal
subquadrangle. Consequently, Γ0 = Γ (cp. Van Maldeghem [21], (1.8.2)).
Passing to the dual, we obtain ∆0 = ∆.
7. Weakly embedded mixed quadrangles of degree 2.

Let Γ be a mixed quadrangle weakly embedded of degree 2 in the projective
space PG(V ), where V is a vector space over the skew field K. In this
section, we deduce from Steinbach & Van Maldeghem [9, (6.1.3)] that there
exists a commutative subfield F of K and a quadric of Witt index 2 over this
subfield, such that Γπ is part of that quadric (except for the universal weak
embedding of W (2)).
7.1. Definition of mixed quadrangles. First, we recall the definition

of a mixed quadrangle (introduced by Tits [17]), cp. Steinbach & Van
Maldeghem [9, (6.1.1)]. Let L be a (commutative) field of characteristic 2
and let
L2 ⊆ Λ0 ⊆ L0 ⊆ Λ ⊆ L,
where L0 is a subfield of L, Λ is a subspace of L considered as vector space
over L0 and Λ0 is a subspace of L0 considered as vector space over L2 . We
suppose that L and L0 are generated as rings by Λ and Λ0 , respectively .
A mixed quadrangle is a certain subquadrangle of the symplectic quad-
rangle W (L0 ). Passing from the symplectic quadrangle to an isomorphic
orthogonal quadrangle, a mixed quadrangle is a subquadrangle of the or-
thogonal quadrangle Q(W, q), associated to the vector space W := Λ × (L0 )4
with usual scalar multiplication and the (nondegenerate) quadratic form
q : W → L0 defined by q((x0 ; x1 , x2 , x3 , x4 )) = x0 2 + x1 x2 + x3 x4 for x0 ∈ L,
xi ∈ L0 (i = 1, . . . , 4).
The mixed quadrangle Q(L0 , L2 ; Λ0 , Λ2 ) consists of the points of Q(W, q)
spanned by vectors of the form
(0; 1, 0, 0, 0), (0; a, 0, 1, 0), (k; b, 0, k 2 , 1), (l; l2 + aa0 , 1, a0 , a),
where a, b, a0 ∈ Λ0 , k, l ∈ Λ.
Lemma 7.2. Assume that the mixed quadrangle Γ := Q(L0 , L2 ; Λ0 , Λ2 ) with
L 6= GF(2) is weakly embedded of degree 2 in the projective space PG(V ),
where V is a vector space over the skew field K.
Then the weak embedding is induced by an injective semi-linear mapping.

In other words, there exists a commutative subfield F of K, a subspace V0
of V (over F) and a quadratic form Q : V0 → F of Witt index 2, such that
for each point x of Γ there is a point Fx0 of the associated quadric with
π(x) = Kx0 .
Proof. Recall that W = Λ×(L0 )4 and that W is endowed with the quadratic
form q : W → L0 . By Steinbach & Van Maldeghem [9, (6.1.3)] there exists an
embedding α : L0 → K and a semi-linear mapping ϕ : W → V (with respect
to α) such that kerϕ ⊆ Λ (the kernel of the symplectic form associated to
q) and π(L0 w) = K(ϕ(w)) for all points L0 w of Γ.
Since the degree of π is 2, we have ϕ(k) 6= 0, for all 0 6= k ∈ Λ. Namely,
otherwise π(L0 (k; 0, 0, k 2 , 1)) is a third point on the secant line spanned by
π(L0 (0; 0, 0, 1, 0)) and π(L0 (0; 0, 0, 0, 1)), a contradiction. This yields that
the semi-linear mapping ϕ is injective.
The lemma is now obvious with F := α(L0 ), V0 := ϕ(W ) and the quadratic
form Q : V0 → F of Witt index 2 defined by Q(ϕ(w)) := α(q(w)), for
w ∈ W.
8. Appendix.
In this appendix, we present the list of all weakly embedded generalized
quadrangles, by putting together the results of the present paper and the
results in Steinbach & Van Maldeghem [9]. We also mention a corollary
on quadrangles Γ for which both Γ and its dual are weakly embeddable in
some projective space. Finally, we show that certain finitely generated free
generalized quadrangles admit lax embeddings in some finite dimensional
projective space.
Recall from Subsection 2.2 that for any skew field L with involutory anti-
automorphism σ and = ±1, we have Λmin := {c − cσ | c ∈ L} ⊆ Λmax :=
{c ∈ L | cσ = c}.
8.1. Classification of generalized quadrangles weakly embedded
in projective space. Let π be a weak embedding of a thick generalized
quadrangle Γ in the projective space PG(V ), where V is a vector space over
the skew field K (not necessarily finite-dimensional). Then Γ is a Moufang
quadrangle. Up to isomorphism, we have the following cases:
(1) Γ is a classical quadrangle arising as the geometry of points and lines
of PG(W ), W a (left) vector space over the skew field L, where one
of the following nondegenerate forms of Witt index 2 vanishes:
(a) a pseudo-quadratic form q : W → L/Λmin (with associated (σ, )-
hermitian form f ),
(b) a (σ, )-hermitian form f : W × W → L with Λmin = Λmax .
Furthermore, there exists a semi-linear mapping ϕ : W → V (with
respect to an embedding α : L → K) with kerϕ ⊆ Rad(W, f ) such that
π(Lw) = Kϕ(w) for all points Lw of Γ (i.e., the weak embedding π is

induced by a semi-linear mapping).
(2) There exists a quaternion skew field L with standard anti-automor-
phism σ and center Z such that Γ is isomorphic to the hermitian
quadrangle arising from the (left) vector space W = L4 endowed with
the (σ, −1)-quadratic form q : W → L/Z defined by q(x1 , x2 , x3 , x4 ) =
x1 xσ3 + x2 xσ4 + Z for x1 , x2 , x3 , x4 ∈ L (i.e., Γ is a quaternion quadran-
gle). Furthermore, the composition of some (nontrivial) automorphism
of Γ and the weak embedding π is induced by an injective semi-linear
mapping.
(3) Γ is a so-called special subquadrangle of a quaternion quadrangle Γ0
and the weak embedding π : Γ → PG(V ) may be extended to a weak
embedding π 0 : Γ0 → PG(V ) (for which Case (1) or (2) applies).
(4) There exists a nonperfect commutative field L of characteristic 2 with
L2 ⊆ Λ0 ⊆ L0 ⊆ Λ ⊆ L,
where L0 is a subfield of L, Λ is an L0 -subspace of L which generates

L as a ring and Λ0 is an L2 -subspace of L0 which generates L0 as a
ring, such that Γ is the mixed subquadrangle Q(L0 , L2 ; Λ0 , Λ2 ) of the
orthogonal quadrangle arising from the vector space W = Λ × (L0 )4
over L0 endowed with the quadratic form q : W → L0 defined by
q(x0 ; (x1 , x2 , x3 , x4 )) := x20 +x1 x2 +x3 x4 for x0 ∈ Λ and x1 , x2 , x3 , x4 ∈
L0 . Furthermore, the weak embedding π is induced by a semi-linear
mapping ϕ : W → V (with respect to an embedding α : L0 → K) with
kerϕ ⊆ Λ.
(5) Γ is the unique generalized quadrangle of order (2, 2) (symplectic quad-
rangle W (2) over GF(2)) and the weak embedding is the universal
weak embedding in the sense of Thas & Van Maldeghem [12], with
charK 6= 2.
A special subquadrangle of a quaternion quadrangle is exactly a quad-

rangle arising from the standard embedding of the dual of the quaternion
quadrangle as orthogonal quadrangle by intersecting with a suitable hyper-
plane, see Steinbach & Van Maldeghem [9, (7.2.1)].
If in Case (1) the radical of the corresponding (σ, )-hermitian form has
(vector) dimension at most 1 (which happens for instance whenever charK 6=
2), then an α(L)-basis of ϕ(W ) is a K-basis of V , see Lemma 5.4.
For further details on weak embeddings of degree 2 or degree > 2, we
refer to the Main Result of the present paper and the one by Steinbach &
Van Maldeghem [9].
As a corollary of Theorem 8.1, we obtain a list of all generalized quadran-
gles, Γ say, such that Γ and the dual generalized quadrangle ΓD are weakly
embeddable. Compare Tits [15, (10.10)] for the case of full embeddings in
polarities.
Proposition 8.2. Let Γ be a thick generalized quadrangle. If Γ and the
dual generalized quadrangle ΓD are weakly embeddable, then, up to duality,
Γ is one of the following:
(1) Γ is an orthogonal quadrangle and the weak embedding of ΓD is in
a projective 3-space. Further, Γ has a standard embedding in a d-
dimensional projective space with
(a) d = 4 (ΓD is a symplectic quadrangle),
(b) d = 5 (ΓD is a hermitian quadrangle),
(c) d = 6 (ΓD is a special subquadrangle of some quaternion quad-
rangle),
(d) d = 7 (ΓD is a quaternion quadrangle).
(2) Γ is any mixed quadrangle (possibly W (2) with universal weak embed-
ding).
For the free construction of a finitely generated generalized quadrangle,
see Tits [18, (4.4)]. It is also contained in Van Maldeghem [21, (1.3.13)].
We end by proving the following result.
Proposition 8.3. Let Γ be a freely constructed generalized quadrangle gen-
erated by the finite geometry Γ0 . If Γ0 can be laxly embedded in some projec-
tive space PG(V ), with V a vector space over any infinite skew field, then
Γ can be laxly embedded in PG(V ).
Proof. One step of the free construction is as follows: For each point-line
pair (x, L) with d(x, L) = 5, introduce a ‘new’ point y and a ‘new’ line M
with x I M I y I L (i.e., d(x, L) = 3 in the new geometry).
To prove the proposition, one has to show that, if x and L are a point
and a line, respectively, of some laxly embedded finite point-line geometry
Γ0 , then we can find a point y of PG(V ) on L off Γ0 such that the line xy
of PG(V ) is not a line of Γ0 . By the finiteness of Γ0 , this is clear.
Of course, one can extend in the obvious way the previous proposition to
quadrangles Γ generated by n points and lines, with n some infinite cardinal
number, and projective spaces PG(V ) over any field with m elements, m a
cardinal number with m > n. Indeed, in the free construction process, in
each step, no more than 2n2 = n new elements are introduced.
It is now clear that one can produce laxly embedded non-Moufang gener-
alized quadrangles. As generating structure Γ0 one can for instance choose
a usual pentagon, or a finite generalized hexagon or octagon laxly embed-
ded in the standard way in some projective space over some finite field and
then extend the field to an infinite field to obtain an embedding in PG(V ),
with V a vector space over an infinite field. In general, ‘free’ quadrangles
have zero probability of being Moufang quadrangles. In our examples, Γ0

does not contain an ordinary quadrangle, and consequently one can easily
see that the corresponding free generalized quadrangle Γ does not contain
a (3 × 3)-grid, or a dual such grid. This implies that Γ cannot admit any
central or axial elation, and hence cannot be a Moufang quadrangle.
References
[1] A.M. Cohen, Point-line spaces related to buildings, in F. Buekenhout, editor, ‘Hand-
book of Incidence Geometry: Buildings and Foundations’, Chapter 12, North-
Holland, Amsterdam (1995), 647-737.
[2] K.J. Dienst, Verallgemeinerte Vierecke in projektiven Räumen, Arch. Math., (Basel),
35 (1980), 177-186.
[3] C. Lefèvre-Percsy, Quadrilatères généralisés faiblement plongés dans PG(3, q), Euro-
pean J. Combin., 2 (1981), 249-255.
[4] C. Lefèvre-Percsy, Projectivités conservant un espace polaire faiblement plongé, Acad.
Roy. Belg. Bull. Cl. Sci., 67 (1981), 45-50.
[5] B. Mühlherr & H. Van Maldeghem, Exceptional Moufang quadrangles of type F4 ,
Canad. J. Math., 51 (1999), 347-371.
[6] S.E. Payne & J.A. Thas, Finite Generalized Quadrangles, Pitman, Boston, London,
Melbourne, 1984.
[7] A. Steinbach, Classical polar spaces (sub-)weakly embedded in projective spaces, Bull.
Belg. Math. Soc. Simon Stevin, 3 (1996), 477-490.
[8] , Generalized quadrangles with a thick hyperbolic line weakly embedded in pro-
jective space, in ‘Finite Geometry and Combinatorics’ (ed. F. De Clerck et al.), Third
International Conference at Deinze, 1997; Bull. Belg. Math. Soc. Simon Stevin, 5
(1998), 447-459.
[9] A. Steinbach & H. Van Maldeghem, Generalized quadrangles weakly embedded of
degree > 2 in projective space, Forum Math., 11 (1999), 139-176.
[10] J.A. Thas, Generalized Polygons, in F. Buekenhout, editor, ‘Handbook of Inci-
dence Geometry: Buildings and Foundations’, Chapter 9, North-Holland, Amsterdam
(1995), 383-431.
[11] J. A. Thas & H. Van Maldeghem, Orthogonal, symplectic and unitary polar spaces
sub-weakly embedded in projective space, Compositio Math., 103 (1996), 75-93.
[12] , Generalized quadrangles weakly embedded in finite projective space, J. Statist.
Plan. Inference, 73 (1998), 353-361.
[13] , Lax embeddings of generalized quadrangles in finite projective spaces, submit-
ted.
[14] J. Tits, Sur la trialité et certains groupes qui s’en déduisent, Inst. Hautes études Sci.
Publ. Math., 2 (1959), 13-60.
[15] , Buildings of spherical type and finite BN -pairs, Lecture Notes in Math., 386,
Springer-Verlag, Berlin, Heidelberg, New York, 1974.
[16] , Classification of buildings of spherical type and Moufang polygons: A survey,
in ‘Coll. Intern. Teorie Combin. Acc. Naz. Lincei’, Proceedings Roma 1973, Atti dei
convegni Lincei, 17, 229-246.
[17] , Quadrangles de Moufang, I, Preprint, 1976.

[18] , Endliche Spiegelungsgruppen, die als Weylgruppen auftreten, Invent. Math.,
43 (1977), 283-295.
[19] , Moufang polygons, I, Root data, Bull. Belg. Math. Soc. Simon Stevin, 1
(1994), 455-468.
[20] J. Tits & R. Weiss, Classification of Moufang Polygons, book in preparation.
[21] H. Van Maldeghem, Generalized Polygons, Birkhäuser, Basel, Monographs in Math-
ematics, 93, 1998.
Received August 18, 1998 and revised March 10, 1999. The second author is a Research
Director of the Fund for Scientific Research – Flanders (Belgium).
Justus-Liebig-Universität Gießen
D 35392 Gießen
Germany
E-mail address: anja.steinbach@math.uni-giessen.de
University Gent
9000 Gent
Belgium
E-mail address: hvm@cage.rug.ac.be
Vol. 193, No. 1, 2000
EMBEDDINGS OF Z2 -HOMOLOGY 3-SPHERES IN R5 UP

TO REGULAR HOMOTOPY
Masamichi Takase
Let F : M 3 ,→ R5 be an embedding of an (oriented) Z2 -

homology 3-sphere M 3 in R5 . Then F bounds an embedding
of an oriented manifold W 4 in R5 . It is well known that the
signature σ(W 4 ) of W 4 is equal to the µ-invariant of M 3 mod-
ulo 16. In this paper we prove that σ(W 4 ) itself completely
determines the regular homotopy class of F .
1. Introduction.
Let Imm[X, Y ] be the set of regular homotopy classes of immersions of a
manifold X in a manifold Y , and Emb[X, Y ] denote the subset of Imm[X, Y ]
consisting of all regular homotopy classes containing an embedding. Smale
[6] has given a 1-1 correspondence (the Smale invariant) s : Imm[S n , RN ] →
πn (VN,n ), where VN,n is the Stiefel manifold of all n-frames in RN . Hirsch
[2] has generalized this to the case of immersions of an arbitrary manifold
in an arbitrary manifold. These results solve the problem of the number of
regular homotopy classes in terms of homotopy theory, but do not succeed
in finding representatives for each class or determining which classes are
represented by an embedding.
According to Hughes [4], Imm[S n , RN ] has a group structure under con-
nected sum and the Smale invariant actually gives a group isomorphism. [4]
gives explicit generators of Imm[S 3 , R4 ] and Imm[S 3 , R5 ].
Hughes-Melvin [5] determine which classes of Imm[S n , Rn+2 ] are repre-
sented by an embedding, and prove that Emb[S n , Rn+2 ] is isomorphic to
Z if n ≡ 3 mod 4, and to 0 otherwise. Furthermore, [5] proves that the
regular homotopy class of an embedding S n ,→ Rn+2 (n ≡ 3 mod 4) can be
completely determined by the signature of its oriented “Seifert” manifold.
For example, in the case n = 3, we have the following diagram:
≈
s : Imm[S 3 , R5 ] −→ π3 (V5,3 ) ≈ Z
∪ ∪
3 5 ≈
Emb[S , R ] −→ 24Z
f 7−→ − 32 σ(V 4 )
where V 4 is an oriented Seifert manifold for f .
249
250 MASAMICHI TAKASE
This implies that there exist many n-knots which cannot be transformed
to the standard embedding even through a smooth deformation admitting
self-intersections (n ≡ 3 mod 4).
The purpose of this paper is to prove a similar statement for embeddings
of Z2 -homology 3-spheres in R5 . More precisely we prove that the regu-
lar homotopy class of an embedding of a Z2 -homology 3-sphere in R5 is
completely determined by the signature of its oriented Seifert manifold.
Throughout this paper, manifolds and immersions are of class C ∞ . The
symbol “≈” denotes an appropriate isomorphism betweeen algebraic objects;
“∼” and “∼r ” mean respectively “homotopic” and “regularly homotopic”.
We often do not distinguish between an immersion f and its regular homo-
topy class, both of which we denote by f .
The author is much grateful to Professor Yukio Matsumoto for his valu-
able advice and encouragement.
2. Preliminaries.
We recall some results of [9]. Let M n be a parallelizable n-manifold, and
f : M n # RN be an immersion. Fix a trivialization T M ∼ = M n × Rn ; we
n n
can associate to f a map df : M → VN,n from M to the Stiefel manifold
VN,n , where VN,n is identified with the set of all injective linear maps from
Rn to RN . df is essentially the differential of f . By Hirsch’s theorem [2],
the correspondence f 7→ df gives a bijection between Imm[M n , RN ] and the
homotopy set [M n , VN,n ]. Every oriented 3-manifold M 3 is parallelizable, so
Imm[M 3 , R5 ] ≈ [M 3 , V5,3 ].
We now study the set [M 3 , V5,3 ]. Since V5,3 is simply connected, we can
make use of the results of Whitney [8]. Let πi = πi (V5,3 ), then π1 = 0,
π2 ≈ π3 ≈ Z. Therefore we must consider the secondary difference.
Identify π2 and π3 with Z in the same way as [9, Proof of Theorem 2]. For
a map ξ : M 3 → V5,3 we can suppose ξ(M (1) ) = p ∈ V5,3 because π1 = 0, (p
is a point in V5,3 and M (q) denotes the q-skeleton of M ). So we can consider
the difference 2-cochain between ξ and the constant map to the point p.
Since ξ is defined over M 3 , this 2-cochain is actually a 2-cocycle. Let Cξ2
denote its cohomology class in H 2 (M 3 ; Z).
Next, for two maps ξ, η : M 3 → V5,3 with ξ|M (2) ∼ η|M (2) , denote by
∆3ξ,η the difference 3-cochain.
The following is an application of [8, Theorem 8A] to our special case of
mappings of M 3 in V5,3 (see also [9, proof of Theorem 2]).
Lemma 2.1 ([8, Theorem 8A], [9, Theorem 2]). Two maps ξ, η : M 3 →
V5,3 are homotopic if and only if
(a) Cξ2 = Cη2 ∈ H 2 (M 3 ; Z).
EMBEDDINGS OF Z2 -HOMOLOGY 3-SPHERES 251
(b) There is a 1-cocycle X 1 and a 2-cochain Y 2 such that ∆3ξ,η = 4X 1 ∪

Cξ2 + δY 2 .
3. Main results.
Let M 3 be a closed oriented 3-manifold. Let D3 be the 3-disk, which from
now on we will often identify with the northern hemisphere of the 3-sphere
S 3 . Fix an inclusion D3 ⊂ M 3 , and put M0 = M 3 − int D3 . Suppose
F0 : M 3 ,→ R5 is an embedding such that F0 |D3 coincides with the northern
part of the standard embedding S 3 ⊂ R5 . For an immersion f : S 3 # R5 ,
we can assume f |(the southern hemisphere) is standard, so define the map
]F0 : Imm[S 3 , R5 ] −→ Imm[M 3 , R5 ]
f 7−→ F0 ]f
where (F0 ]f )|M0 = F0 |M0 , and (F0 ]f )|D3 = f |D3 . The normal bundle of
F0 is trivial and if F0 is altered on D3 its normal bundle does not change.
So we can in fact define the map
]F0 : Imm[S 3 , R5 ] −→ Imm[M 3 , R5 ]0
where Imm[M 3 , R5 ]0 is the subset of Imm[M 3 , R5 ] consisting of all regu-
lar homotopy classes of immersions with trivial normal bundle. Note that
Emb[M 3 , R5 ] ⊂ Imm[M 3 , R5 ]0 .
Proposition 3.1. If H 2 (M 3 ; Z) has no elements of even order, then
]F0 : Imm[S 3 , R5 ] −→ Imm[M 3 , R5 ]0
is bijective.
Proof. Let νF be the normal bundle of an immersion F : M 3 # R5 . Since
there is the bundle map
νF −→ V5,5
↓ ↓
dF
M 3 −→ V5,3
and since the Euler class of the S 1 -bundle V5,5 → V5,3 is equal to 2Σ2 for a
generator Σ2 ∈ H 2 (V5,3 ; Z) ≈ Z, we have
νF is trivial,
⇔ the normal Euler class of F (denoted by χF ) is zero,
∗ ∗
⇔ dF (2χF ) = 2dF (χF ) = 0,
⇔ 2CdF 2 = 0,
2
⇔ CdF = 0.
Therefore, Imm[M03 , R5 ]0 ≈ H 3 (M0 ; Z) = 0 by [9, Theorem 2]. This means
that ]F0 is surjective from the covering homotopy property for immersion
spaces (see [7]).
We next prove the injectivity. For two immersions f, g : S 3 # R5 , by

Lemma 2.1,
F0 ]f ∼r F0 ]g,
⇔ d(F0 ]f ) ∼ d(F0 ]g),
⇔ ∆3 is a coboundary.
d(F0 ]f ),d(F0 ]g)
If we consider D3 as a 3-cell, ∆3 is a 3-cochain which assigns

d(F0 ]f ),d(F0 ]g)
s(f ) − s(g) ∈ π3 (V5,3 ) to D3 , and 0 ∈ π3 (V5,3 ) to other 3-cells by definition.
So clearly
∆3 is a coboundary,
d(F0 ]f ),d(F0 ]g)
⇔ s(f ) = s(g) ∈ π3 (V5,3 ),
⇔ f ∼r g : S 3 → R5 .
This completes the proof.
Remark 3.2. We already know, by a result of Wu ([9, Theorem 2]), that
for a closed oriented 3-manifold M 3 , Imm[M 3 , R5 ]0 ≈ Z t · · · t Z, i.e.,
the disjoint union of as many copies of Z as the number of elements c ∈
H 2 (M 3 ; Z) with 2c = 0. In particular, in our case where H 2 (M 3 ; Z) has no
elements of even order, this implies that Imm[M 3 , R5 ]0 ≈ Z.
We now investigate ]F0 restricted to Emb[M 3 , R5 ]. We want to show that
]F0 gives a bijection between Emb[S 3 , R5 ] and Emb[M 3 , R5 ].
Theorem 3.3. If H 1 (M 3 ; Z2 ) = 0, then
]F0 : Emb[S 3 , R5 ] −→ Emb[M 3 , R5 ]
is bijective.
Prop. 3.1
Furthermore, under the identification Imm[M 3 , R5 ]0 ≈ Imm[S 3 , R5 ]
Smale inv.
≈ Z,
Emb[M 3 , R5 ] ≈ 24Z
3 4
F 7−→ 2 (σ(WF ) − σ(WF40 ))
where WF4 stands for an oriented Seifert manifold for F , and σ(WF4 ) is its
signature.
Proof. Extend the embedding F0 : M 3 ,→ R5 to an embedding F0 : WF40 ,→
R5 . Take a suitable neighbourhood of M 3 in WF40 diffeomorphic to M 3 ×
[0, 1), and further extend F0 to an embedding (denoted again by F0 )
F0 : WF40 ∪ M 3 × (−1, 0] ,→ R5 .
M ×{0}
Let F : M 3 ,→ R5 be an embedding, and extend F to

F : WF4 ∪ M 3 × (−1, 0] ,→ R5 .
M ×{0}
in the same way as above.

Take a neighbourhood M00 of M0 in M 3 . Since M00 × (−1, 1) is paralleliz-
able,
Imm[M00 × (−1, 1), R5 ] ≈ [M00 × (−1, 1), V5,4 ] ≈ [M0 , SO(5)].
And it follows by obstruction theory that Imm[M00 × (−1, 1), R5 ] ≈

[M0 , SO(5)] consists of a unique element, because π2 (SO(5)) = 0,
H 3 (M0 ; π3 (SO(5))) = 0, and H 1 (M0 ; π1 (SO(5))) ≈ H 1 (M 3 ; Z2 ) = 0. There-
fore we can alter F by a regular homotopy (we use again the letter F to
represent the resulting immersion) so that
F |(M00 × (−1, 1))(x, t) = F0 |(M00 × (−1, 1))(x, −t), (x, t) ∈ M00 × (−1, 1).
Consider the manifold VF4 = WF40 ∪ WF4 (the orientation of VF4 is

M0 ×{0}
taken to be in accord with the one of WF40 ), whose boundary is S 3 . Using F
and F0 , construct a map from VF4 to R5 . This map is an immersion except
on S 2 = ∂M0 ⊂ ∂VF4 . Pushing a neighbourhood of S 2 into VF4 , we have an
immersion G of the whole VF4 in R5 (Figure 1).
Figure 1.
Figure 2.
Now clearly F ∼r F0 ](G|∂VF4 ) : M 3 # R5 (Figure 2). By Proposition

3.1, the regular homotopy class of F depends only on the regular homotopy
class of G|∂VF4 : S 3 # R5 . Since [5, Proof of Theorem and Corollary 2]
actually proves that if an immersion f : S 3 # R5 bounds an immersion of
an oriented 4-manifold V 4 then s(f ) is equal to − 32 σ(V 4 ), we can see
3
s G|∂VF4 = − σ(VF4 ) ∈ 24Z,

2
and G|∂VF4 ∈ Emb[S 3 , R5 ]. Thus, the map ]F0 gives a bijection from
Prop. 3.1
Emb[S 3 , R5 ] to Emb[M 3 , R5 ]. Therefore, identifying Imm[M 3 , R5 ]0 ≈
Smale inv.
Imm[S 3 , R5 ] ≈ Z, F ∈ Emb[M 3 , R5 ] corresponds to 23 σ(VF4 ) =
3
− 2 (σ(WF40 ) − σ(WF4 )) by Novikov additivity. This completes the proof.
Remark 3.4. We actually proved here that if an immersion F : M 3 # R5
bounds an immersion of an oriented 4-manifold WF4 then F corresponds to
3 4 4 3 5
2 (σ(WF ) − σ(WF0 )) under the above identification Imm[M , R ]0 ≈ Z.
Remark 3.5. Suppose M 3 is a Z2 -homology sphere. By Theorem 3.3,

we can choose F0 so that σ(WF40 ) = µ(M 3 )0 , where µ(M 3 )0 is the inte-
ger in {0, 1, · · · , 15} representing the µ-invariant µ(M 3 ) ∈ Z/16Z. Let
S : Imm[M 3 , R5 ]0 → Z denote the previous identification through this F0 ,
Imm[M 3 , R5 ]0 ≈Imm[S 3 , R5 ]≈Z. Then Theorem 3.3 implies that S(F ) =
3 4 3 0 3 5
2 (σ(WF ) − µ(M ) ) ∈ 24Z if F ∈ Emb[M , R ].
4. Realizing h-cobordisms in R5 .
In this section, we study the following problem. Suppose M1 , M2 are
two Z2 -homology 3-spheres which are mutually h-cobordant and let Si :
Imm[Mi , R5 ]0 → Z (i = 1, 2) denote the bijections as in Remark 3.5. Is it
possible to relate S1 to S2 ?
Let M1 , M2 be as above, and V be an h-cobordism between M1 and M2 .
Let Fi : Mi ,→ R5 be embeddings and Wi be oriented Seifert manifolds
for them (i = 1, 2). Abstractly each Mi bounds a simply connected spin

4-manifold Wi0 of signature σ(Wi0 ) = σ(Wi ) (taking a connected sum with
some copies of the ±K3-surface if necessary) (i = 1, 2)(see [3]). Consider
the closed manifold
Y = W10 ∪ V ∪ W20 .
M1 M2
Y is a simply connected spin 4-manifold of signature ±(σ(W10 ) − σ(W20 )),

since W10 ∪ V is homotopy equivalent to W10 and since each Mi admits a
M1
unique spin structure. By Cochran [1], Y can embed in R5 if σ(W10 ) =
σ(W20 ). Clearly this embedding restricted to each Mi is regularly homotopic
to Fi (i=1,2), using Theorem 3.3.
Conversely, suppose H : V ,→ R5 is an embedding. H can extend to
an immersion of W1 ∪ V in R5 for a Seifert manifold W1 for H|M1 , if the
trivialization of the normal bundle of H|M1 (for the construction of W1 ) is
suitably chosen. This, together with Theorem 3.3, implies that S1 (H|M1 ) =
S2 (H|M2 ) ∈ Z because σ(W1 ) = σ(W1 ∪ V ).
Thus, we have:
Proposition 4.1. Let Mi , Si (i = 1, 2) and V be as above. For embeddings

Fi : Mi ,→ R5 (i = 1, 2), S1 (F1 ) = S2 (F2 ) ∈ Z if and only if there is an
embedding H : V ,→ R5 with H|Mi ∼r Fi (i = 1, 2) (or equivalently, there
is an immersion H : V # R5 with H|Mi = Fi (i = 1, 2)).
References
[1] T. Cochran, Embedding 4-manifolds in S 5 , Topology, 23 (1984), 257-269.
[2] M. Hirsch, Immersions of manifolds, Trans. Amer. Math. Soc., 93 (1959), 242-276.
[3] , The imbedding of bounding manifolds in Euclidean space, Ann. of Math., 74
(1961), 494-497.
[4] J. Hughes, Bordism and regular homotopy of low-dimensional immersions, Pacific J.
Math., 156 (1992), 155-184.
[5] J. Hughes and P. Melvin, The Smale invariant of a knot, Comment. Math. Helv., 60
(1985), 615-627.
[6] S. Smale, A classification of immersions of spheres in euclidean spaces, Ann. of Math.,
69 (1959), 327-344.
[7] , A survey of some recent developments in differential topology, Bull. Amer.
Math. Soc., 69 (1963), 131-145.
[8] H. Whitney, Classification of the mappings of a 3-complex into a simply connected
space, Ann. of Math., 50 (1949), 270-284.
[9] W.T. Wu, On the immersion of C ∞ -3-manifolds in a Euclidean space, Sci. Sinica.,
13 (1962), 335-336.
Received February 1, 1998 and revised February 5, 1999.
Graduate School of Mathematical Sciences

University of Tokyo
3-8-1 Komaba, Meguro-ku
Tokyo 153
Japan
E-mail address: takase@ms.u-tokyo.ac.jp
Guidelines for Authors
Authors may submit manuscripts at pjm.math.berkeley.edu/about/journal/submissions.html

and choose an editor at that time. Exceptionally, a paper may be submitted in hard copy to
one of the editors; authors should keep a copy.
By submitting a manuscript you assert that it is original and is not under consideration
for publication elsewhere. Instructions on manuscript preparation are provided below. For
further information, visit the web address above or write to pacific@math.berkeley.edu or
to Pacific Journal of Mathematics, University of California, Los Angeles, CA 90095–1555.
Correspondence by email is requested for convenience and speed.
Manuscripts must be in English, French or German. A brief abstract of about 150 words or
less in English must be included. The abstract should be self-contained and not make any
reference to the bibliography. Also required are keywords and subject classification for the
article, and, for each author, postal address, affiliation (if appropriate) and email address if
available. A home-page URL is optional.
Authors are encouraged to use LATEX, but papers in other varieties of TEX, and exceptionally
in other formats, are acceptable. At submission time only a PDF file is required; follow
the instructions at the web address above. Carefully preserve all relevant files, such as
LATEX sources and individual files for each figure; you will be asked to submit them upon
acceptance of the paper.
Bibliographical references should be listed alphabetically at the end of the paper. All
references in the bibliography should be cited in the text. Use of BibTEX is preferred but
not required. Any bibliographical citation style may be used but tags will be converted to
the house format (see a current issue for examples).
Figures, whether prepared electronically or hand-drawn, must be of publication quality.
Figures prepared electronically should be submitted in Encapsulated PostScript (EPS) or
in a form that can be converted to EPS, such as GnuPlot, Maple or Mathematica. Many
drawing tools such as Adobe Illustrator and Aldus FreeHand can produce EPS output.
Figures containing bitmaps should be generated at the highest possible resolution. If there
is doubt whether a particular figure is in an acceptable format, the authors should check
with production by sending an email to pacific@math.berkeley.edu.
Each figure should be captioned and numbered, so that it can float. Small figures occupying
no more than three lines of vertical space can be kept in the text (“the curve looks like
this:”). It is acceptable to submit a manuscript will all figures at the end, if their placement
is specified in the text by means of comments such as “Place Figure 1 here”. The same
considerations apply to tables, which should be used sparingly.
Forced line breaks or page breaks should not be inserted in the document. There is no point
in your trying to optimize line and page breaks in the original manuscript. The manuscript
will be reformatted to use the journal’s preferred fonts and layout.
Page proofs will be made available to authors (or to the designated corresponding author)
at a Web site in PDF format. Failure to acknowledge the receipt of proofs or to return
corrections within the requested deadline may cause publication to be postponed.
Pacific Journal of Mathematics

Volume 193 No. 1 March 2000 Pacific
Simple connectivity of the Markov partition space
L. BADOIAN AND J.B. WAGONER
1
Journal of
Nilmanifolds and associated Lie algebras over the integers 5
B OHUMIL C ENKL AND R ICHARD P ORTER
Rigidity of compact minimal submanifolds in a sphere
S HARIEF D ESHMUKH
31 Mathematics
The cobordism of oriented three dimensional orbifolds 45
K.S. D RUSCHEL
Two-sided braid groups and asymptotic inclusions 57
J ULIANA E RLIJMAN
Characterization of Hardy spaces by singular integrals and ‘Divergence-Free’
2000
wavelets 79
J.E. G ILBERT , J.A. H OGAN AND J.D. L AKEY
Total determination of material parameters from electromagnetic boundary
Vol. 193, No. 1

information 107
M ARK S. J OSHI AND S TEPHEN R. M C D OWALL
On the composition of a prime transcendental function and a prime
polynomial 131
T UEN -WAI N G AND C HUNG -C HUN YANG
A Paley–Wiener theorem for the inverse spherical transform 143
A NGELA PASQUALE
On the action of the group of diffeomorphisms of a surface on sections of the
determinant line bundle 177
D OUG P ICKRELL
On central extensions of gyrocommutative gyrogroups 201
K RZYSZTOF RÓZGA
Global existence and decreasing property of boundary values of solutions to
parabolic equations with nonlocal boundary conditions 219
S ANGWON S EO
Volume 193 No. 1 March 2000
Generalized quadrangles weakly embedded of degree 2 in projective space 227
H ENDRIK VAN M ALDEGHEM
Embeddings of Z2 -homology 3-spheres in R5 up to regular homotopy 249
M ASAMICHI TAKASE

A Paley Wiener Theorem For The Inverse S PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Paley Wiener Theorem For The Inverse S PDF

Uploaded by

Copyright:

Available Formats

PACIFIC JOURNAL OF MATHEMATICS

Pacific Journal of Mathematics

Volume 193 No. 1 March 2000

SIMPLE CONNECTIVITY OF THE MARKOV PARTITION

L. Badoian and J.B. Wagoner

In Wagoner, 1987 the simplicial complex PA of Markov par-

Proposition. A closed path in PA with L edges can be spanned by a

2) If U → V and U and V are Markov partitions, then U ∩ V is a Markov

We can now deform a closed path of length L to a constant path as fol-

closed path of length 2L with vertices V0 , V1 , . . . , V2L . The number of trian-

E-mail address: wagoner@math.berkeley.edu

NILMANIFOLDS AND ASSOCIATED LIE ALGEBRAS

Bohumil Cenkl and Richard Porter

We consider the correspondence between nilmanifolds and

1. Nilmanifolds and corresponding Lie algebras over the reals

It is easy to find the first three homogeneous components of z. Namely,

has the structure of a Lie algebra L = LQ . L as a subset of GR has group

The group G, as a set, can be identified with Qk via the map e : G → Qk

2. Finitely-generated torsion-free nilpotent groups.

L Gi /Gi+1 are infinite cyclic.

is a canonical basis, and

basis for a finitely-generated torsion-free nilpotent group whose associated

and there is a similar formula when x is replaced by y = (y1 , . . . , y7 ).

From here it is immediate that for x and y such that x2 = 1, xj = 0, j 6= 2

However, it is possible to find a nilpotent torsion-free Lie algebra L0 on

Example 3. Let L0 be the Lie algebra over Z with generators e1 , . . . , e7

3. The spectral sequence for polynomial cochains.

where qjr,s ∈ Q and gj ∈ La+b . Thus,

((x1 g1 ) ∗ · · · ∗ (xk gk )) ∗ ((y1 g1 ) ∗ · · · ∗ (yk gk )) = (ρ1 (x, y)g1 ∗ · · · ∗ ρk (x, y)gk ).

|ξj + ψj (ξ1 , . . . , ξj−1 )| ≤ |ξj | = |gj |

|xj + ϕj (x1 , . . . , xj−1 )| ≤ |xj | = |gj |.

It follows that the statement is proved if we show that the product in L,

the result follows from the identities

where the a0ij s are nonnegative integers. P 0 (G) = Z. The differential d on

kuvk = kuk + kvk,

Lemma 2. The polynomial functions ρj and τj have their norms bounded

Using the norm k k, we define a filtration on P (G) by setting

Then du is given by formula (5) above. Note that

And according to Lemma 2,

which is ≤ i, by the assumption. Therefore, kduk ≤ i. Hence, du ∈ F i .

3.1. E0 and E1 -terms of the spectral sequence. Let {Er , dr } be the

Let p be the projection F i → E0 . The next step is to compute E1 . Since

when aj ≥ 2. But this is the differential on P 1 (A), when A is the Z-module

But from Lemma 1, it follows that

Hence, we must have

4. The spectral sequence for the Lie algebra cochains.

L∗ = HomA (L, A). Recall that there is a pairing

In particular, for every f ∈ r (L∗ ) and g ∈ s (L∗ ), f ∧ g ∈ r+s (L∗ ) and

kωj k = |gj | − 1, kqωj k = kωj k for q ∈ A,

Using this norm, we define an ascending filtration on (L∗ ) by setting

(L∗ ) lowers the filtration.

Proof. Since d is a derivation it suffices to verify that if kωj k = i, then

kdωj k = max{kωr k + kωs k}

4.1. E0 , E1 , and E2 terms of the spectral sequence for Lie algebra

hdωj , gr ∧ gs i = −hωj , [gr , gs ]i

lower central series with torsion-free quotients is G = G1 > G2 > G3 > 1.

is filtered and graded. The elements ω1 , . . . , ω7 have dimension one and

dω1 = dω2 = dω3 ,

Then an explicit computation shows that

d2 is determined by that part of the differential d that lowers the filtration

5. Construction of a group from a Lie algebra.

group G such that L(G) is rationally equivalent to L. However, G is not

where m1 , . . . , mk are arbitrary integers and

SinceS ≥ n(n− 2), that is khk2 ≤ n, above integral gives k∇hk = 0,

Ωn,orb (F − (H), F) Ωn,orb (F − (H), F̂) Ωn,orb (F − (H))