A NULLSTELLENSATZ FOR AMOEBAS
A NULLSTELLENSATZ FOR AMOEBAS
A NULLSTELLENSATZ FOR AMOEBAS
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong><br />
KEVIN PURBHOO<br />
Abstract<br />
The amoeba of an affine algebraic variety V ⊂ (C ∗ ) r is the image of V under<br />
the map (z 1 ,...,z r ) ↦→ (log |z 1 |,...,log |z r |). We give a characterisation of the<br />
amoeba, based on the triangle inequality, which we call testing for lopsidedness. We<br />
show that if a point is outside the amoeba of V , there is an element of the defining ideal<br />
which witnesses this fact by being lopsided. This condition is necessary and sufficient<br />
for amoebas of arbitrary codimension as well as for compactifications of amoebas<br />
inside any toric variety. Our approach naturally leads to methods for approximating<br />
hypersurface amoebas and their spines by systems of linear inequalities. Finally, we<br />
remark that our main result can be seen as a precise analogue of a Nullstellensatz<br />
statement for tropical varieties.<br />
Contents<br />
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407<br />
2. The case r = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411<br />
3. The hypersurface case . . . . . . . . . . . . . . . . . . . . . . . . . . 417<br />
4. Approximating a hypersurface amoeba by linear inequalities . . . . . . 426<br />
5. More general amoebas . . . . . . . . . . . . . . . . . . . . . . . . . . 431<br />
6. Tropical varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438<br />
Appendix. Details of calculations . . . . . . . . . . . . . . . . . . . . . . . 441<br />
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445<br />
1. Introduction<br />
1.1. Statement of results<br />
Let V ⊂ (C ∗ ) r be an algebraic variety, defined by an ideal I ⊂<br />
C[z 1 ,z −1<br />
1 ,...,z r,zr −1 ].<br />
Definition 1.1 (Gel’fand, Kapranov, and Zelevinsky; see [GKZ])<br />
The amoeba of V is defined to be the image of V under the map Log : (C ∗ ) r → R r<br />
DUKE MATHEMATICAL JOURNAL<br />
Vol. 141, No. 3, c○ 2008 DOI 10.1215/00127094-2007-001<br />
Received 24 July 2006. Revision received 5 April 2007.<br />
2000 Mathematics Subject Classification. Primary 14Q15; Secondary 14Q10, 14M25.<br />
Author’s research partially supported by Natural Science and Engineering Research Council of Canada.<br />
407
408 KEVIN PURBHOO<br />
defined at the point z = (z 1 ,...,z r ) by<br />
Log(z) = (log |z 1 |,...,log |z r |).<br />
We denote the amoeba of V by either A V or A I .IfV = Z f is a hypersurface,<br />
the zero locus of a single function f , we also use the notation A f . We refer the<br />
reader to Mikhalkin’s survey article [M] for a broad discussion of amoebas and their<br />
applications.<br />
In this article, we address the following fundamental question: given a point<br />
a ∈ R r and an ideal I ⊂ C[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ],whenisa ∈ A I ? This problem was<br />
previously studied by Theobald [T], who gave a practical answer for certain families<br />
of amoebas. Here we give a general answer to this question. We first consider the case<br />
where I =〈f 〉 is the ideal of a hypersurface. From this, we deduce a characterisation<br />
theorem for arbitrary ideals which is the analytic counterpart to a fundamental theorem<br />
for tropical varieties.<br />
Consider f ∈ C[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ], and consider a point a ∈ R r . Write f as a<br />
sum of monomials f (z) = m 1 (z) +···+m d (z).Definef {a} to be the list of positive<br />
real numbers<br />
f {a} := {∣ ∣ m1<br />
(<br />
Log −1 (a) )∣ ∣ ,...,<br />
∣ ∣md<br />
(<br />
Log −1 (a) )∣ ∣ } .<br />
Note that since the m i are monomials, this is well defined, even though Log is not<br />
injective.<br />
Definition 1.2<br />
We say that a list of positive numbers is lopsided if one of the numbers is greater than<br />
the sum of all the others.<br />
Equivalently, a list of numbers {b 1 ,...,b d } is not lopsided if it is possible to choose<br />
complex phases φ i (|φ i |=1), so that ∑ φ i b i = 0. This follows from the triangle<br />
inequality. We also define<br />
LA f := { a ∈ R r ∣ ∣ f {a} is not lopsided<br />
}<br />
.<br />
One can easily see that if a ∈ A f ,thenf {a} cannot be lopsided; in other words,<br />
LA f ⊃ A f .Indeed,iff (z) = 0, thenm 1 (z) +···+m d (z) = 0, soitisgivinga<br />
way to assign complex phases to the list {|m 1 (z)|,...,|m d (z)|} = f {Log(z)} such<br />
that the sum is zero. Thus one can think of LA f as a crude approximation to the<br />
amoeba A f .<br />
Example 1.3<br />
Suppose that f (z 1 ,z 2 ) = 1 + z 1 z 2 + z2 2, and let a ∈ R2 . For any complex<br />
phases φ 1 ,φ 2 , there exist (z 1 ,z 2 ) ∈ Log −1 (a) such that φ 1 |z 1 z 2 | = z 1 z 2 and
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 409<br />
φ 2 |z2 2|=z2 2 . Thus a ∈ A f<br />
A f = LA f .<br />
if and only if {1, |z 1 z 2 |, |z2 2 |} is nonlopsided; that is,<br />
In the above example, we have enough freedom to choose the phases of the monomials<br />
m i (z) for z ∈ Log −1 (a) so that LA f = A f . However, this works only because f<br />
has very few nonzero terms. In general, LA f can be quite different from A f (see<br />
Figure 1). Nevertheless, we show that for a suitable multiple of f , we can use this<br />
lopsidedness test to get very good approximations for A f .<br />
Let n be a positive integer. We consider the polynomials<br />
˜f n (z) =<br />
∏n−1<br />
k 1 =0<br />
These ˜f n are cyclic resultants<br />
···<br />
∏n−1<br />
k r =0<br />
f (e 2πi k 1/n z 1 ,...,e 2πi k r /n z r ).<br />
˜f n (z) = Res ( Res(...Res(f (u 1 z 1 ,...,u r z r ),u n 1 − 1) ...,un r−1 − 1),un r − 1)<br />
and as such can be practically computed. When n = 2 k , this can be done reasonably<br />
efficiently as follows. Define polynomials h i by h 0 := f ,andlet<br />
h i (z 1 ,...,z 2 [i] ,...,z r):= h i−1 (z 1 ,...,z [i] ,...,z r ) h i−1 (z 1 ,...,−z [i] ,...,z r ),<br />
where [i] ∼ = i (mod r). Then ˜f n (z) = h kr (z1 n,...,zn r<br />
), and the recursion computes<br />
this in O(n 2(r2−r) )-time, which is proportional to the square of the number of terms of<br />
˜f n .<br />
Our main result for amoebas of hypersurfaces is roughly the following. The<br />
precise version is stated and proved in Section 3.2.<br />
THEOREM 1 (Rough version)<br />
As n →∞, the family LA ˜f n<br />
converges uniformly to A f . There exists an integer N<br />
such that to compute A f to within ε, it suffices to compute LA ˜f n<br />
for any n ≥ N.<br />
Moreover, N depends only on ε and the Newton polytope (or degree) of f and can be<br />
computed explicitly from these data.<br />
This leads us to the following characterisation of the amoeba of a general subvariety<br />
of (C ∗ ) r .<br />
THEOREM 2<br />
Let I ⊂ C[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ] be an ideal. A point a ∈ R r is in the amoeba A I if<br />
and only if for every g ∈ I, g{a} is not lopsided.<br />
Phrased another way, if a point a is outside the amoeba A I , a polynomial f ∈ I may<br />
witness this fact by being lopsided at a. Theorem 2 then states that there is always<br />
a witness. We actually show something slightly stronger in both Theorems 1 and 2.
410 KEVIN PURBHOO<br />
Figure 1. The image on the left depicts LA f ⊃ A f , while the image on the right depicts<br />
SA f ⊃ A f . Here SA f is not homotopic to A f , and in general,<br />
LA f need not be either.<br />
We show that there is a witness f such that f {a} is superlopsided according to the<br />
following definition.<br />
Definition 1.4<br />
Let d ′ ≥ d ≥ 2. We say that a list of positive numbers {b 1 ,...,b d+1 } is<br />
d ′ -superlopsided if there exists some i such that b i >d ′ b j for all j ≠ i. Ifd ′ = d,<br />
we simply say the list is superlopsided.<br />
As before, we also define<br />
SA f := { a ∈ R r ∣ ∣ f {a} is not superlopsided } .<br />
If a list of positive numbers is superlopsided, it is certainly lopsided; hence<br />
SA f ⊃ LA f ⊃ A f (see Figure 1). David Speyer [S] observed that each component<br />
of the complement of SA f is given by a system of linear inequalities, making it easier<br />
than LA f to compute explicitly. Hence Theorem 1 actually prescribes a method for<br />
approximating A f to within ε by systems of linear inequalities. Similar ideas lead<br />
to a method for approximating the spine of a hypersurface amoeba. We discuss these<br />
constructions in Section 4.<br />
The motivation for these results comes from tropical algebraic geometry, and<br />
from this viewpoint, lopsidedness (rather than superlopsidedness) is the more natural<br />
condition to consider. In tropical algebraic geometry, we work with the semiring<br />
R trop (⊙, ⊕). This is a semiring whose underlying set is R but whose operations are<br />
given by<br />
• a ⊙ b := a + b,<br />
• a ⊕ b := max(a,b).<br />
The operations ⊙ and ⊕ are known as tropical addition and tropical multiplication. One<br />
can easily check that they satisfy the usual commutative, associative, and distributive<br />
laws; however, there are no additive inverses.
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 411<br />
A polynomial g ∈ R trop [x 1 ,...,x r ] is therefore a piecewise linear function on<br />
R r ;if<br />
g(x) = ⊕<br />
c k1 ,...,k r<br />
⊙ x k 1<br />
1 ⊙···⊙xk r<br />
r ,<br />
k 1 ,...,k r<br />
then, translated into the usual operations on R,<br />
g(x) = max{c k1 ,...,k r<br />
+ k 1 x 1 +···+k r x r }.<br />
The tropical variety associated to g is then defined to be the singular locus of this<br />
piecewise linear function. A tropical variety associated to a single polynomial g in<br />
this way is called a tropical hypersurface.<br />
Thus there is a simple Nullstellensatz ∗ for tropical hypersurfaces. A point x is<br />
outside the tropical variety of g if there is a single monomial term of g which is<br />
strictly larger than each of the others when evaluated at the point x. In terms of the<br />
tropical operations, this term is strictly greater than the tropical sum of the other terms<br />
(cf. Definition 1.2).<br />
More generally, the principal results in this article can be seen as an analytic<br />
analogue of a theorem for tropical varieties of arbitrary codimension (see [EKL,Theorem<br />
2.2.5], [SS, Theorem 2.1], [St, Theorem 9.17]), also known as nonarchimedean<br />
amoebas. We discuss this connection in Section 6.<br />
2. The case r = 1<br />
2.1. A heuristic argument<br />
The idea of the one-variable case is simple enough. Suppose that f (z) = ∏ d<br />
i=1 (α i −z),<br />
and for sake of argument, assume that the absolute values of the α i are all distinct,<br />
say, |α 1 | > ···> |α d | > 0.Then<br />
˜f n (z) =<br />
d∏<br />
i=1<br />
(α n i<br />
− z n )<br />
=±z nd ∓ (α n 1 + { ···})z n(d−1)<br />
± (α n 1 αn 2 + { ···})z n(d−2)<br />
∓ ···+<br />
− (α n 1 ···αn d−1 + { ···})z n<br />
+ α n 1 ···αn d .<br />
∗ We use the term in the literal sense of being a statement about zeros; these results are not an analogue of<br />
Hilbert’s Nullstellensatz.
412 KEVIN PURBHOO<br />
For n large, the terms { ···} are small in comparison with the other terms, and so<br />
this is approximately<br />
g n (z) =±(z d ) n ∓ (α 1 z d−1 ) n ± (α 1 α 2 z d−2 ) n ∓···−(α 1 ···α d−1 z) n + (α 1 ···α d ) n .<br />
Suppose that |α k+1 | < |z| < |α k |. Consider g n (z)/(α 1 ···α d−k z d−k ) n .Asn →∞,<br />
every term tends to zero except for the constant term, which is ±1. Thus, for n large,<br />
there is a single term in g n (z), and likewise in ˜f n , which is much bigger in absolute<br />
value than all the others.<br />
2.2. The one-variable lemmas<br />
We now formalise this heuristic argument in a way that is useful in proving Theorem 1.<br />
At the crux of the heuristic argument are the following three key facts about ˜f n .<br />
(1) It has no roots inside a certain annulus. (In the heuristic argument, the annulus<br />
is {z ∈ C ||α k+1 | < |z| < |α k |}.)<br />
(2) The only nonzero terms that appear are of the form cz nk .<br />
(3) The number of terms is not too large. (This approach fails if instead of ˜f n (z),<br />
we try to use ˜f n (z) D with D ≫ n.)<br />
To get a result that we can apply to the multivariable case, we need to be able to<br />
make a uniform statement about polynomials with these properties. This is precisely<br />
captured by the next two lemmas. By applying Lemma 2.2 directly to the family of<br />
functions ˜f n (z), one immediately obtains a complete proof of Theorem 1 in the r = 1<br />
case.<br />
LEMMA 2.1<br />
Let A ={z | β 0 < |z|
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 413<br />
conclusion (2.1) remains valid. Here maxdeg(f ) and mindeg(f ) refer, respectively,<br />
to the largest and smallest exponents that appear in f (z). The notation in our proof<br />
assumes that f (z) is actually a polynomial.<br />
Proof<br />
We can write<br />
f (z) =<br />
d∏<br />
(z n + α n i ),<br />
where |α 1 |≥···≥|α d |. We adopt the convention that α 0 = 0 and α d+1 =∞.<br />
Since f n (z) has no roots in A, wehave<br />
i=1<br />
α k+1 ≤ β 0
414 KEVIN PURBHOO<br />
If l
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 415<br />
In particular, for each l ≠ k, wehave<br />
|m ′ l (z 0)| ≤(1 + γ n ) d − 1. (2.5)<br />
However, we can do slightly better than this by noting that the smallest power of<br />
γ which appears on the right-hand side of inequality (2.3) isγ |k−l| . Thus, whereas<br />
(2.5) tells us that |m ′ l (z 0)| < ∑ ( d<br />
)<br />
w≥1 w γ nw , in fact, we have<br />
|m ′ l (z 0)| < ∑ ( d<br />
γ<br />
w)<br />
nw<br />
w≥|k−l|<br />
< ∑<br />
w≥|k−l|<br />
(dγ n ) w<br />
. (2.6)<br />
w!<br />
(Although (2.3) is a better estimate than (2.6), the latter proves to be more useful to<br />
us.)<br />
For m k (z 0 ), we have the estimate<br />
∣ ( ∑ ) |m ′ k (z s<br />
0)| =<br />
1
416 KEVIN PURBHOO<br />
Remark 2.1<br />
Note that we have actually determined which term is the special term m k .Thisis<br />
done in (2.2). If f n is a polynomial, then n(d − k) is the number of roots (counted<br />
with multiplicity) of f n inside the disc {|z| ≤β 0 }.Iff n is a Laurent polynomial,<br />
n(d − k) − mindeg(f n ) is the number of roots inside {0 < |z| ≤β 0 }.<br />
LEMMA 2.2<br />
As before, let A ={z | β 0 < |z|
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 417<br />
γ → 1. The best general answer for the question is of the same form; that is,<br />
n log γ −1 ≤ (D 0 + D 1 ) log n + log (c 0 c 1 ).<br />
One can see this by performing the requisite analysis on the polynomials h n (z) =<br />
(z n + 1) c 0n D 0<br />
. As a general heuristic, the more closely the roots of f n are packed, the<br />
larger n has to be; thus the family of polynomials h n (z), where every root has as high<br />
a multiplicity as possible, is where we expect our worst-case behaviour to occur.<br />
Suppose we want n large enough to guarantee that h n {log |z 0 |} is (c 1 n D 1 )-superlopsided<br />
for |z 0 | < γ < 1. Write h n (z) = 1 + c 0 n D 0 zn + ···. We know that<br />
1 is the dominant term as n gets large (since 1 = lim n→∞ h n (z 0 )); thus we need<br />
(c 1 n D 1 )(c 0n D 0 zn 0 ) < 1 or, equivalently,<br />
n log γ −1 ≥ (D 0 + D 1 ) log n + log(c 0 c 1 ).<br />
If we only want to guarantee that f n {a} is lopsided for a ∈ K, we need n large<br />
enough so that<br />
∑<br />
|m l (z 0 )|≤(1 + γ n ) d − 1 < 2 − (1 + γ n ) d ≤|m k (z 0 )|<br />
l≠k<br />
or, equivalently, (1 + γ n ) c 0n D 0<br />
≤ 3/2. This holds if we have<br />
(<br />
n log γ −1 c<br />
)<br />
0<br />
≥ D 0 log n + log .<br />
log 3/2<br />
So n needs to be only about half as big to guarantee that f n {a} is lopsided as it does<br />
to guarantee that f n {a} is superlopsided. Again, we can see that this is fairly close to<br />
the best answer by considering (z n + 1) c 0n D 0<br />
.<br />
3. The hypersurface case<br />
3.1. Preliminaries<br />
In this section, we prove our main theorem characterising the amoeba of a hypersurface.<br />
If f (z) ∈ C[z 1 ,z −1<br />
1 ,...,z r,z −1 ], we consider the Laurent polynomials<br />
˜f n (z) =<br />
∏n−1<br />
k 1 =0<br />
r<br />
···<br />
∏n−1<br />
k r =0<br />
f (e 2πi k 1/n z 1 ,...,e 2πi k r /n z r ).<br />
Theorem 1 states that for ε>0 and for a point a ∈ R r in the complement of the<br />
amoeba A f whose distance from A f is at least ε, we can choose n large enough so<br />
that ˜f n {a} is superlopsided. Moreover, the theorem gives an upper bound on how large<br />
n needs to be, based only on ε and the Newton polytope of f .
418 KEVIN PURBHOO<br />
The idea behind the proof of Theorem 1 is to look at the family of ˜f n (z) and<br />
interpret this as a function of a single variable z i . At the point ζ = (ζ 1 ,...,ζ r ) ∈ C r ,<br />
we define<br />
˜f i,ζ<br />
n (z) := ˜f n (ζ 1 ,...,ζ i−1 ,z,ζ i+1 ,...,ζ n ).<br />
We apply Lemma 2.2 to these and find a single dominant term in this polynomial of<br />
one variable. Then by an averaging argument, we show that this implies that ˜f n has a<br />
single dominant term.<br />
First, however, we need a few simple observations.<br />
PROPOSITION 3.1<br />
We have A f = A ˜f n<br />
.<br />
Proof<br />
The cyclic resultant ˜f n (z) is a product of terms g u1 ,...,u r<br />
(z) = f (u 1 z 1 ,...,u r z r ), where<br />
u n i = 1.Since|u i |=1, A gu1 ,...,ur<br />
= A f , and so A ˜f n<br />
= ⋃ A gu1 ,...,ur<br />
= A f . <br />
We also need to know some information about the number and degree of the terms<br />
which appear in ˜f n . First, note the following important fact.<br />
PROPOSITION 3.2<br />
The only monomials that appear in ˜f n are of the form cz nk 1<br />
1 ···z nk r<br />
r<br />
. In particular, the<br />
only terms appearing in ˜f<br />
n i,ζ(z)<br />
are of the form cznk .<br />
Proof<br />
Let C n denote the cyclic group of roots of z n −1.Now, ˜f n is manifestly invariant under<br />
the group action of (C n ) r acting on C[z 1 ,z −1<br />
1 ,...,z r,zr −1 ] by (u 1 ,...,u n ) · g(z) =<br />
g(u 1 z 1 ,...,u n z n ). Thus each monomial of ˜f n must be invariant under this action.<br />
The only monomials with this property are of the form cz nk 1<br />
1 ···z nk r<br />
r<br />
.<br />
The statement about ˜f<br />
n i,ζ (z) follows immediately. <br />
Recall that if g ∈ C[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ], then its Newton polytope, denoted (g),is<br />
the subset of R r defined as the convex hull of the exponent vectors of the monomials<br />
which appear in g.<br />
For any polytope , letd() be any upper bound on (#{Z r ∩ m})/m r .In<br />
general, it is not easy to find a tight upper bound for this number. If one can compute<br />
the Ehrhart polynomial of explicitly, then an easy upper bound is the sum of the<br />
positive coefficients. Otherwise, it is possible to bound the coefficients of the Ehrhart<br />
polynomial in terms of the volume of (see [BM]). Using these estimates, for each r<br />
one can compute constants A and B such that (#{Z r ∩ m})/m r
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 419<br />
Clearly, we have ( ˜f n ) = n r (f ). This gives us an upper bound on the number<br />
of terms that ˜f n can have.<br />
PROPOSITION 3.3<br />
Let d = d((f )). Then ˜f n has at most dn r2 −r terms.<br />
Proof<br />
By Proposition 3.2, the number of terms in ˜f n is at most the number of integral points<br />
in (1/n)( ˜f n ) = n r−1 (f ). This is less than or equal to dn r2−r .<br />
<br />
Finally, we need to know something about maxdeg( ˜f i,ζ<br />
n<br />
c i (f ):= max x i<br />
(<br />
(f )<br />
)<br />
− min xi<br />
(<br />
(f )<br />
)<br />
,<br />
where x i denotes the ith coordinate function on R r .<br />
) − mindeg( ˜f<br />
n<br />
i,ζ).Let<br />
PROPOSITION 3.4<br />
We have maxdeg( ˜f i,ζ<br />
n<br />
Proof<br />
We have<br />
maxdeg( ˜f i,ζ<br />
n<br />
) − mindeg( ˜f i,ζ<br />
n ) = c i(f )n r .<br />
) − mindeg( ˜f i,ζ<br />
n<br />
) = max x (<br />
i ( ˜f n ) ) (<br />
− min x i ( ˜f n ) )<br />
(<br />
= n r max x i ( ˜f n ) ) (<br />
− n r min x i ( ˜f n ) )<br />
= d i n r . <br />
3.2. Proof of Theorem 1<br />
Armed with these facts and Lemmas 2.1 and 2.2, we are now in a position to precisely<br />
state and prove our main result for amoebas of hypersurfaces.<br />
THEOREM 1<br />
Let ε>0. Suppose that a = (a 1 ,...,a r ) ∈ R r \ A f is a point in the amoeba<br />
complement whose distance from A f is at least ε. Letd = d((f )), and let c =<br />
max{c i (f ) | 1 ≤ i ≤ r}.<br />
(1) If n is large enough so that<br />
then ˜f n {a} is lopsided.<br />
nε ≥ (r − 1) log n + log ( (r + 3)2 r+1 c ) , (3.1)
420 KEVIN PURBHOO<br />
(2) If n is large enough so that<br />
( (16 ) )<br />
nε ≥ (r 2 − 1) log n + log cd , (3.2)<br />
3<br />
then ˜f n {a} is superlopsided. (In fact, it is (dn r2 −r )-superlopsided.)<br />
The key to reducing to the one-variable case is the following basic result from complex<br />
analysis.<br />
LEMMA 3.5<br />
Let f (z) be a Laurent polynomial, and write f (z) = ∑ −→ m−→ j j<br />
(z), wherem−→ j<br />
(z) =<br />
m j1 ,...,j r<br />
(z) = b j1 ,...,j r<br />
z j 1<br />
1 ···zj r<br />
r . Suppose that for all ζ ∈ Log−1 (a), we have |f (ζ)| ≤<br />
M. Then for each −→ l, |m−→ l<br />
(ζ)| ≤M.<br />
Proof<br />
We integrate the equations M ≥|f (ζ)| over the set Log −1 (a 1 ,...,a r ):<br />
M ≥ 1<br />
(2π) r ∫ 2π<br />
θ 1 =0<br />
1<br />
≥ ∣<br />
(2πi)<br />
∫|z r 1 |=1<br />
∫ 2π<br />
··· ∣ ∑ m−→ j<br />
(e a 1+iθ 2<br />
,...,e a r +iθ r ) ∣ dθ 1 ···dθ r<br />
−→ j<br />
θ r =0<br />
=|m−→ l<br />
(e a 1<br />
,...,e a r<br />
)|<br />
∫<br />
∑<br />
···<br />
|z r |=1<br />
m−→ j<br />
(e a 1 z 1,...,e a r z r)<br />
−→ z l 1<br />
1 ···z l r<br />
r<br />
j<br />
dz 1<br />
z 1 ···dz 1<br />
z 1<br />
∣ ∣∣<br />
=|m−→ l<br />
(ζ)|.<br />
<br />
Proof of Theorem 1<br />
Let γ = e −ε ,andletA i ={z | γe a i<br />
< |z| ε. But since z 0 ∈ A i ,<br />
‖Log(ζ ′ ) − a‖ =|log(z 0 ) − a i |
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 421<br />
fixed i, it must be the same monomial term that dominates in each ˜f<br />
n<br />
i,ζ,<br />
independent<br />
of the choice of ζ. Let this be the z nk i<br />
-term, and let −→ k = (k 1 ,...,k r ).<br />
Write<br />
˜f n (z) = ∑ −→ j<br />
m−→ j<br />
(z),<br />
where m−→ j<br />
(z) is the monomial b−→ j<br />
z nj 1<br />
1 ···z nj r<br />
r<br />
.<br />
Let M =|m−→ k<br />
(ζ)|. Note that this does not depend on the particular choice of ζ.<br />
Let<br />
µ = max { |m−→ l<br />
(ζ)| ∣ ∣ −→ l ≠ −→ k } .<br />
We wish to show that µ
422 KEVIN PURBHOO<br />
We now prove statement (1). Although the approach is essentially the same, it is<br />
slightly more difficult and hence requires some additional lemmas (Calculations A.2<br />
and A.3, which are found in the appendix). The reason for this is that we cannot get<br />
these bounds by appealing directly to Lemma 2.2. Instead, we use Lemma 2.1,which<br />
gives better estimates for the coefficients of ˜f<br />
n i,ζ.<br />
Write ˜f i,ζ<br />
maxdeg ˜f<br />
n<br />
i,ζ<br />
that<br />
n (z) = ∑ j mi j (z), where mi j (z) = b jz nj . Then for each i,<br />
(z) − mindeg ˜f<br />
n<br />
i,ζ(z)<br />
≤ cnr . So by Lemma 2.1, there is some k i such<br />
|m i l (ζ)|<br />
∑<br />
|m i k i<br />
(ζ)| < w≥|k i −l| (cnr−1 γ n ) w /w!<br />
,<br />
2 − e cnr−1 γ n<br />
and in fact, it is the same k i for all choices of ζ.<br />
As before, let M =|m−→ k<br />
(ζ)|, andletσ = ∑ −→ j ≠<br />
−→ k<br />
|m−→ j<br />
(ζ)|. Wehaveforanyζ<br />
that |m i k i<br />
(ζ)|
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 423<br />
and by Calculation A.3 (see the appendix), this becomes<br />
σ<br />
( e<br />
(r+2)cn r−1 γ n<br />
M + σ < − 1<br />
)<br />
2r . (3.5)<br />
2 − e cnr−1 γ n<br />
Assume now that (3.1) holds. By Calculation A.2 (see the appendix), n is large<br />
enough so that the right-hand side is less than 1/2. Thus we have σdcan be used in place of d in Theorem 1.<br />
Thus we see that it suffices to take n so that<br />
( (16 ) )<br />
nε ≥ (r 2 − 1) log n + log cα . <br />
3
424 KEVIN PURBHOO<br />
3.4. Accuracy of bounds<br />
Just as was the case in Lemma 2.2, the bounds on n given in Theorem 1 are not quite<br />
optimal; there are a number of places in which the inequalities can obviously be made<br />
tighter. However, as ε → 0, the bounds are at least asymptotically correct.<br />
To see this for the superlopsided case, we can consider the example<br />
f (z) = (1 − z 1 ) D1 ···(1 − z r ) D r<br />
.<br />
The amoeba A f<br />
compute<br />
is the union of all coordinate hyperplanes in R r . We can easily<br />
˜f n (z) = (1 − z n 1 )D 1n r−1 ···(1 − z n r )D r n r−1<br />
= 1 − D 1 n r−1 z n 1 −···−D rn r−1 z n r +···.<br />
If our point is a = (a 1 ,...,a r ), with each −ε D i (D 1 ···D r )n r2 −r<br />
nε > (r 2 − 1) log n + log ( (D 1 ···D r )max{D i } ) .<br />
In contrast, (3.2) says that in this example, we should take n so that<br />
( (16 )<br />
)<br />
nε > (r 2 − 1) log n + log (D 1 + 1) ···(D r + 1) max{D i } .<br />
3<br />
Our bound (3.1) for lopsidedness appears to be slightly less satisfactory. In the<br />
above example, to guarantee lopsidedness one needs n large enough so that<br />
This holds when<br />
(1 − z n 1 )D 1n r−1 ···(1 − z n r )D r n r−1 − 1 < 1.<br />
(ea 1 n + 1) D 1n r−1 ···(e a r n + 1) D r n r−1 < 2<br />
⇔D 1 n r−1 log(1 + e a 1n ) +···+D r n r−1 log(1 + e a r n ) < log 2.
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 425<br />
Noting that a i > −ε, and approximating log(1 + x) ∼ x, this condition becomes<br />
(D 1 +···+D r )n r−1 e −εn < log 2<br />
( D1 +···+D r<br />
⇔ nε > (r − 1) log n + log<br />
log 2<br />
If we take D 1 =···=D r = D, then this simplifies to<br />
( rD<br />
)<br />
nε > (r − 1) log n + log .<br />
log 2<br />
In contrast, Theorem 1 tells us that it is sufficient to take n so that<br />
nε > (r − 1) log n + log ( (r + 3)D ) + (r + 1) log 2.<br />
Again, this shows that the bounds in Theorem 1 are asymptotically correct, at least<br />
for any fixed r. We suspect, however, that the correct general answer does not have<br />
this last term, or any term which is linear in r.<br />
3.5. Other cyclic resultants<br />
Instead of the family ˜f n , one may wish to consider a more general family of cyclic<br />
resultants. Let n 1 ,...,n r be positive integers, and consider<br />
˜f n1 ,...,n r<br />
(z) =<br />
n∏<br />
1 −1<br />
k 1 =0<br />
n∏<br />
r −1<br />
···<br />
k r =0<br />
)<br />
.<br />
f (e 2πi k 1/n 1<br />
z 1 ,...,e 2πi k r /n r<br />
z r ).<br />
Unfortunately, it is not true that the family SA ˜f n1<br />
converges uniformly to A<br />
,...,nr<br />
f<br />
as n 1 ,...,n r →∞. Trouble occurs if some of the n i are significantly larger than<br />
others. For example, consider the amoeba of f (z 1 ,z 2 ) = (1 − z 1 )(1 − z 2 ) at a point<br />
(a 1 ,a 2 ) ∈ R 2 with a 1 < 0, a 2 < 0. Then<br />
˜f n1 ,n 2<br />
(z 1 ,z 2 ) = (1 − z n 1<br />
1 )n 2<br />
(1 − z n 2<br />
2 )n 1<br />
= 1 − n 2 z n 1<br />
1 − n 1z n 2<br />
2 +···.<br />
If n 2 ∼ e −a 1n 1<br />
, then the first two terms above will have the same order of magnitude.<br />
Thus ˜f n1 ,n 2<br />
{(a 1 ,a 2 )} is not superlopsided, even if n 1 (and hence n 2 ) are large. It is not<br />
even lopsided.<br />
However, if we restrict ourselves to the situation in which each n i is bounded by<br />
some polynomial in each of the other n j , then a statement analogous to Theorem 1<br />
is true. For example, we can let n i be any polynomial function of a single parameter<br />
n. We do not compute explicit bounds for approximating the amoeba to within ε in
426 KEVIN PURBHOO<br />
this more general situation; however, the answer depends on these polynomials. It is<br />
certainly still true that SA ˜f n1<br />
converges uniformly to A<br />
,...,nr<br />
f , as this argument really<br />
only depends on the fact that degrees of ˜f n1 ,...,n r<br />
are growing only polynomially, while<br />
the terms are becoming suitably sparse.<br />
4. Approximating a hypersurface amoeba by linear inequalities<br />
4.1. Locating the dominant term<br />
Theorem 1 tells us that for n sufficiently large, one term of ˜f n dominates, but it does<br />
not specify which one. The answer depends on in which component of the amoeba<br />
complement our point a lies. Since ˜f n {a} varies continuously with a, it depends only<br />
on the component of the amoeba complement.<br />
Now, the number of components is relatively small compared to the number of<br />
terms of ˜f n . There is a natural injective map<br />
ind : components of R r \A f ↩→ (f ) ∩ Z r<br />
(cf. [FPT, Definition 2.1]). This is called the index of the component; a complete<br />
definition is given below. So only a few of the terms of ˜f n can possibly be dominant<br />
terms. Fortunately, it is relatively simple to determine which these are. The Newton<br />
polytope of ˜f n is n r (f ), and candidates for dominant term are, in fact, what one<br />
expects them to be; namely, they are the images of the integral points of (f ) under<br />
this scaling.<br />
PROPOSITION 4.1<br />
Let a ∈ R r \A f , and let ind(a) = −→ k = (k 1 ,...,k r ) be the corresponding point<br />
in (f ). If ˜f n {a} is lopsided, then the term of ˜f n (z) which dominates has exponent<br />
vector n r−→ k (i.e., it is the (z nr k 1<br />
1 ···z nr k r<br />
)-term).<br />
r<br />
In order to make complete sense of the statement, we need to know a definition of<br />
the index −→ k . There are a number of equivalent definitions, but the simplest for our<br />
purposes is the following.<br />
Let ζ ∈ Log −1 (a). For each i ∈{1,...,r}, consider the polynomial,<br />
f i,ζ (z) = f (ζ 1 ,...,ζ i−1 ,z,ζ i+1 ,...,ζ r ).<br />
If f is a polynomial, then k i is the number of roots (with multiplicity) of f i,ζ inside<br />
the open disc {|z|
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 427<br />
ζ, this number is independent of ζ.Iff is a Laurent polynomial, then<br />
k i = #roots of f i,ζ (z) inside { 0 < |z|
428 KEVIN PURBHOO<br />
is the candidate for the dominant term in this component, and m−→ j<br />
(z) = b−→ j<br />
z j 1<br />
1 ···zj r<br />
r<br />
are the other monomials.<br />
The corresponding component of R r \SA ˜f n<br />
is the set<br />
Log ({ z ∣ ∣ |M −→ k<br />
(z)| >D|m−→ j<br />
(z)|, ∀ −→ j }) ,<br />
where D + 1 is the number terms in ˜f n . Equivalently, this is the set of x ∈ R r such<br />
that<br />
log|B−→ k<br />
|+n r k 1 x 1 +···+n r k r x r > log D + log|b−→ j<br />
|+j 1 x 1 +···+j r x r (4.1)<br />
for all −→ j . This is a system of linear inequalities in the variable x, so the solutions<br />
to these equations are a convex polyhedron that approximates the component of the<br />
amoeba to within ε. If there is no component of the R r \A f corresponding to −→ k ,then<br />
this system of equations has no solutions. Conversely, if this system of inequalities<br />
has no solutions, then this component of the amoeba (if it exists) is not large enough<br />
to contain a ball of radius ε.<br />
Thus we can realise any component of the R r \A f as an increasing union of<br />
convex polyhedra. This gives an independent proof of the basic fact (see [FPT]) that<br />
the components of the R r \A f are convex. We must admit, however, that there are<br />
simpler proofs of this fact.<br />
Note that in Theorem 1, we actually show that ˜f n is (dn r2−r )-superlopsided. Thus<br />
we can, in fact, take D = dn r2−r in (4.1), and the set of solutions to this system of<br />
inequalities still approximates the component of R r \A f to within ε.<br />
In practice, it rapidly becomes impractical to get arbitrarily good approximations<br />
to the amoeba by linear inequalities in this way, particularly for r > 2, since the<br />
number of inequalities is O(n r2−r ).Forr = 2, this is more manageable, though for<br />
purposes of simply drawing the amoeba, Theobald’s numerical method for drawing<br />
planar amoebas (see [T]) is probably faster. It is therefore natural to wonder whether<br />
some smaller subset of these inequalities can suffice. Although the answer is yes,<br />
it is unfortunately not easy to give an a priori answer as to which inequalities are<br />
needed. As n →∞, the terms m−→ j<br />
that are “near” to M−→ k<br />
become more relevant<br />
than the terms that are farther; however, this is only heuristic, and moreover, since we<br />
are approximating a piecewise smooth region by polyhedra, the number of relevant<br />
inequalities also approaches infinity. On the other hand, one practical use of Theorem 1<br />
is to find components of R r \A f , and here the heuristic that nearby terms are the<br />
most relevant can be helpful. One can first look for a value of x = Log(z) such that<br />
|M−→ k<br />
(z)| ≫|m−→ j<br />
(z)| for nearby terms m−→ j<br />
, and if one exists, check that x satisfies<br />
all inequalities (4.1). The efficiency of such an algorithm is commensurate with the<br />
computation of ˜f n .
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 429<br />
4.3. Approximating the spine<br />
One of the primary tools for studying amoebas has been the Ronkin function N f ,<br />
defined in [R]. For f ∈ C[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ], N f is defined to be the pushforward<br />
of log|f | under the map Log:<br />
N f (x) := 1 log|f (z 1 ,...,z r )| dz 1 ···dz r<br />
.<br />
(2πi)<br />
∫Log r −1 (x) z 1 ···z r<br />
Ronkin shows in [R] thatN f is a convex function, and it is affine-linear precisely on<br />
the components of R r \A f . When restricted to a single component of E of R r \A f ,<br />
∇N f = ind(E).<br />
Passare and Rullgård [PR] use this function to define the spine of the amoeba as<br />
follows. For each component C of R r \A f , extend the locally affine-linear function<br />
of N f | E to an affine-linear function N E on all of R r .Let<br />
N ∞ f<br />
{<br />
(x) = max NE (x) } .<br />
E<br />
This is a convex piecewise linear function on R r , superscribing N f .Thespine of the<br />
amoeba A f is defined to be the set of points where Nf<br />
∞ is not differentiable and is<br />
denoted S f .<br />
The spine of the amoeba S f is a strong deformation retract of A f (see [PR], [Ru]).<br />
Also, note that S f is actually a tropical hypersurface, as defined in the introduction;<br />
that is, it is the singular locus of the maximum of a finite set of linear functions, where<br />
the gradient of each linear function is a lattice vector.<br />
Now, observe that<br />
1<br />
n r log| ˜f n (z)| = 1 n r<br />
n∑<br />
···<br />
k 1 =1<br />
n∑<br />
log|f (e 2πi k1/n z 1 ,...,e 2πi k r /n z r )|<br />
k r =1<br />
can be thought of as a Riemann sum for N f . In particular, we may expect<br />
(1/n r ) log| ˜f n (z)| to converge pointwise to N f (Log(z)). This is certainly true, provided<br />
that log| ˜f n (z)| is bounded on Log −1 (x), which is the case when x ∈ R r \A f .<br />
Suppose that x = Log(z) is in the component of R r \A f of index −→ k ∈ (f ).<br />
Assume that x has distance at least δ from the amoeba, where δ>0 is fixed. For any<br />
ε>0, we can find n sufficiently large so that<br />
˜f n (z) = M−→ k<br />
(z) + ∑ −→ j<br />
m−→ j<br />
(z),<br />
where each m−→ j<br />
is relatively small; that is,<br />
∑<br />
|m −→ j<br />
(z)|
430 KEVIN PURBHOO<br />
(see Corollary 3.7). Thus we have<br />
log|M−→ k<br />
(z)|+log(1 − ε) ≤ log| ˜f n (z)| ≤log|M−→ k<br />
(z)|+log(1 + ε).<br />
Thus we see that as n →∞, the values of (1/n r ) log| ˜f n (z)|,<br />
1<br />
n r log|M −→ k<br />
(z)| = 1 n r log|B −→ k<br />
|+k 1 x 1 +···+k r x r ,<br />
and N f (x) = N E (x) = c−→ k<br />
+ k 1 x 1 +···+k r x r all converge on N f (x).<br />
We can use this fact to obtain good approximations for the spine of the amoeba.<br />
For each n, we consider the function M ∞ : R r → R given by<br />
M ∞ (x) := max log|M−→ k<br />
(z)|, (4.2)<br />
where the maximum is taken over all components of R r \A f . This is a piecewise<br />
linear function. We define the approximate spine of the amoeba LS f,n to be the set<br />
of points where M ∞ (x) is not smooth. Equivalently, LS f,n is the set of points where<br />
the maximum in equation (4.2) is attained by two distinct values of −→ k .<br />
PROPOSITION 4.2<br />
We have the following relationships:<br />
(1) LS f,n ⊂ LA ˜f n<br />
;<br />
(2) lim n→∞ LS f,n = S f .<br />
Proof<br />
Statement (1) is true because on the component of R r \LA ˜f n<br />
of index −→ k , |M−→ k<br />
(z)| ><br />
|M−→ l<br />
(z)| for any other −→ l ∈ (f ). Thus the maximum value in equation (4.2) cannot<br />
be attained by two distinct −→ k if x ∈ R r \LA ˜f n<br />
.<br />
Statement (2) follows from the fact that (1/n r ) log|M−→ k<br />
(z)|−N f (x) is a constant<br />
function and is less than ε for n large. Let E 1 and E 2 be components of R r \A f of index<br />
−→ k 1 and −→ k 2 , respectively. Consider the hyperplane H ⊂ R r , where log|M−→ k 1<br />
(z)| and<br />
log|M−→ k 2<br />
(z)| coincide, and the hyperplane H ′ , where N E1 (x) and N E2 (x) coincide. The<br />
two hyperplanes H and H ′ are parallel, and their distance apart is most εK, where<br />
K is some constant depending only on −→ k 1 and −→ k 2 . As there are only finitely many<br />
−→ k ∈ (f ) ∩ Z r , these distances can be made uniformly small.<br />
<br />
For practical reasons, we may use an alternate definition of M ∞ , in which one takes<br />
the maximum in equation (4.2) over only those components that appear in R r \SA ˜f n<br />
.<br />
If we do, statement (2) is still true, and statement (1) is true for large n; for small n,<br />
we must settle for saying that LS f,n ⊂ SA ˜f n<br />
.
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 431<br />
One may hope to be able to simplify this construction by taking the maximum in<br />
equation (4.2) over all −→ k ∈ (f ) ∩ Z, rather than just those that actually correspond<br />
to components. It appears, however, that this does not give the same answer. With this<br />
alternate definition of M ∞ ,theapproximate spine has false chambers for all n; that<br />
is, the complement of this approximate spine has components that do not correspond<br />
to components of R r \A f . We may still hope that these false chambers shrink to zero<br />
volume as n gets large. Unfortunately, experimental evidence suggests that the limit of<br />
these false chambers, as n →∞, can sometimes contain a ball of positive radius, and<br />
so this method does not produce a good approximation of the spine. An interesting<br />
open question is whether the limiting behaviour of these false chambers is somehow<br />
captured by the Ronkin function.<br />
5. More general amoebas<br />
5.1. Amoebas of higher codimension varieties in (C ∗ ) r<br />
The higher-codimension statement (Theorem 2) follows fairly quickly from the hypersurface<br />
statement. Let V ⊂ C r be a variety that is the zero locus of an ideal<br />
I =〈f 1 ,...,f k 〉⊂C[z 1 ,z −1<br />
1 ,...,z r,zr −1 ].<br />
PROPOSITION 5.1<br />
For every a ∈ R r , there exists f a ∈ I such that<br />
Proof<br />
For any Laurent polynomial<br />
Z fa ∩ Log −1 (a) = V ∩ Log −1 (a).<br />
g(z) = ∑ b−→ j<br />
z j 1<br />
1 ···zj r<br />
r<br />
∈ C[z 1 ,z −1<br />
1 ,...,z r,z −1<br />
r<br />
],<br />
−→ j<br />
let ḡ denote its complex conjugate<br />
ḡ(z) = ∑ −→ j<br />
¯b−→ j<br />
z j 1<br />
1 ···zj r<br />
r .<br />
We define f a to be<br />
f a (z) :=<br />
k∑<br />
i=1<br />
f i (z 1 ,...,z r ) ¯f i (e 2a 1<br />
z −1<br />
1 ,...,e2a r<br />
z −1<br />
r<br />
).
432 KEVIN PURBHOO<br />
Clearly, f a is a Laurent polynomial and is in I. Moreover, if we restrict z to Log −1 (a),<br />
then z i ¯z i = e 2a i<br />
,so<br />
f a (z) =<br />
=<br />
=<br />
k∑<br />
f i (z 1 ,...,z r ) ¯f i (¯z 1 ,...,¯z r )<br />
i=1<br />
k∑<br />
f i (z 1 ,...,z r )f i (z 1 ,...,z r )<br />
i=1<br />
k∑<br />
|f i (z)| 2 .<br />
i=1<br />
Thus f a (z) = 0 if and only if f i (z) = 0 for all i.<br />
<br />
This result is also true for ideals in C[z 1 ,...,z r ]: one can find a suitable monomial<br />
m(z) such that<br />
m(z 1 ,...,z r ) ¯f i (e 2a 1<br />
z −1<br />
1 ,...,e2a r<br />
z −1<br />
r<br />
)<br />
is a polynomial for all i, and a similar argument holds if<br />
f a (z) =<br />
k∑<br />
i=1<br />
f i (z 1 ,...,z r ) ( m(z 1 ,...,z r ) ¯f i (e 2a 1<br />
z −1<br />
1 ,...,e2a r<br />
z −1<br />
r<br />
) ) .<br />
As an immediate consequence of Proposition 5.1, we have the following.<br />
COROLLARY 5.2<br />
For any ideal I ⊂ C[z 1 ,z −1<br />
1 ,...,z r,z −1<br />
r<br />
],<br />
A I = ⋂ f ∈I<br />
A f .<br />
It is now a simple task to prove our second main result.<br />
THEOREM 2<br />
Let I ⊂ C[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ] be an ideal. A point a ∈ R r is in the amoeba A I if<br />
and only if g{a} is not (super)lopsided for every g ∈ I.<br />
Proof<br />
If a ∈ A I ,thenf {a} cannot be lopsided for any f ∈ I since a ∈ A f for every<br />
f ∈ I. On the other hand, suppose that a /∈ A I . Then by Proposition 5.1, ifwetake
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 433<br />
g = f a ∈ I, thena /∈ A g . By Theorem 1, ifn is sufficiently large, then ˜g n {a} is<br />
(super)lopsided, and ˜g n ∈ I.<br />
<br />
Remark 5.1<br />
In summary, we have three coincident sets, any of which can be used to define the<br />
amoeba A V of a variety V = V (I):<br />
(1) A V = Log(V );<br />
(2) A V = ⋂ f ∈I A f ;and<br />
(3) A V ={a ∈ R r | f {a} is not lopsided for all f ∈ I}.<br />
In Section 6, we see that this is precisely analogous to a theorem for tropical algebraic<br />
varieties.<br />
If a point a is in R r \A I , the proof of Theorem 2 also tells us where to look for a<br />
witness to this fact: namely, we should look at ˜f an {a} for all n. For some sufficiently<br />
large n, this list is lopsided.<br />
One unfortunate misfeature of this proof is that it requires us to use a different<br />
g for every point a ∈ R r \A I . Thus this statement is purely local. It does not give<br />
any clues as to how to produce a global uniform approximation to A I .However,in<br />
general, we cannot expect there to be any finite set of elements g i ∈ I such that if<br />
a /∈ A I , then some ˜g i n is lopsided for n sufficiently large. If it were so, this would<br />
imply that A I is always an intersection of finitely many hypersurface amoebas, and<br />
this is certainly not true for dimensional reasons if dim V
434 KEVIN PURBHOO<br />
More concretely, if we identify T ′ with (S 1 ) r′ , we can write<br />
χ(µ 1 ,...,µ r ′) =<br />
( r ′<br />
∏<br />
i=1<br />
∏r ′<br />
)<br />
j ,..., µ A ir<br />
j ,<br />
where A ij are the integer entries of a matrix A—the matrix representation of ˆχ—and<br />
µ A i1<br />
( r∑<br />
Log ′ (z) = A Log(z) = A 1j log|z j |,...,<br />
j=1<br />
i=1<br />
r∑<br />
j=1<br />
)<br />
A r ′ j log|z j |, .<br />
We can also take a matrix A with integer entries as our starting point and construct<br />
T ′ , χ, and the map Log ′ so that (5.1) holds.<br />
Let I ⊂ C[z 1 ,z −1<br />
1 ,...,z r,zr −1 ] denote the ideal of V. Let Ṽ ⊂ (C ∗ ) r+r′ denote<br />
the variety of the ideal<br />
Ĩ = I + J ⊂ C[z 1 ,z −1<br />
1 ,...,z r,z −1<br />
r<br />
,w 1 ,w −1<br />
1 ,...,w r ′,w−1 r ], ′<br />
where J = 〈 w i − ∏ r<br />
〉<br />
j=1 zA ij<br />
j . Now, consider the projection of Ṽ onto the w-coordinates<br />
(C ∗ ) r′ . The image of Ṽ under this projection is a variety V ′ . Standard techniques of<br />
elimination theory allow us to compute its ideal I ′ (see, e.g., [CLO]).<br />
PROPOSITION 5.3<br />
We have the following relationships: Log ′ (V ) = A(A V ) = A V ′.<br />
Proof<br />
A point in Ṽ is simply a pair (z, w), where z ∈ V and w i = ∏ r<br />
A Ṽ = { (x, y) ∈ R r+r′ ∣ ∣ y = Ax, x ∈ AV<br />
}<br />
.<br />
Projecting onto the w-coordinates, we obtain<br />
A V ′ = { ∣<br />
y ∈ R r′ (x, y) ∈ A Ṽ for some x }<br />
={Ax | x ∈ A V }<br />
= A Log(V )<br />
j=1 zA ij<br />
j<br />
. Thus we have<br />
= Log ′ (V ). <br />
It is interesting to note that this construction is closely related to the cyclic resultants<br />
used in the proof of Theorem 1. Suppose that V = Z f is a hypersurface, and suppose<br />
that χ : T ′ = T → T is the map χ(t) = t n . In this case, A = nI is a multiple of<br />
the identity matrix, and the variety V ′ is the zero locus of the function ˜f n . Intuitively,<br />
we should think that the linear transformation is zooming in on the amoeba A;aswe
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 435<br />
zoom in, Theorem 1 tells us that we see more and more detail in the approximations<br />
LA and SA.<br />
5.3. Compactified amoebas<br />
The most natural generalisation of amoebas in the compact setting is to subvarieties of<br />
projective toric varieties. Each projective toric variety is a compactification of (C ∗ ) r<br />
with an (S 1 ) r -action which extends the (S 1 ) r -action on (C ∗ ) r . It also carries a natural<br />
symplectic form ω, for which the (S 1 ) r -action is Hamiltonian. We may therefore use<br />
the moment map for this Hamiltonian action to replace the map Log.<br />
Our goal in this section is to give a concrete description of this more general setting<br />
and observe that our results still hold. This follows fairly easily from the noncompact<br />
case. Our construction of toric varieties and their moment maps roughly follows a<br />
combination of [F]and[A].<br />
Let ⊂ R r be an r-dimensional lattice polytope; that is, the vertices of have<br />
integral coordinates. To every such , we can associate the following data:<br />
(1) a set of lattice points A = ∩ Z r ;<br />
(2) a semigroup ring C[A]; ifA = { −→ k 1 ,..., −→ k d }, this is defined to be the<br />
quotient ring<br />
C[s −→ k 1<br />
,...,s −→ k d<br />
]/J,<br />
where each s −→ k 1<br />
has degree 1 and J is generated by all (homogeneous) relations<br />
of the form<br />
whenever<br />
s −→ k i1 ···s<br />
−→ k ip − s<br />
−→ k j1 ···s<br />
−→ k jp = 0<br />
−→ k i1 + ··· + −→ k ip = −→ k j1 + ··· + −→ k jp ;<br />
note that C[A] carries an action of the complex torus T = (C ∗ ) r ,givenby<br />
(λ 1 ,...,λ r ) · s −→k = λ k 1<br />
1 ···λk r<br />
r s−→k ;<br />
(3) a toric variety X = Proj(C[A]);<br />
(4) a projective embedding φ : X↩→ P d−1 = Proj(C[t 1 ,...,t d ]), induced by the<br />
map on rings C[t 1 ,...,t d ] → C[A] given by t i ↦→ s −→ k i<br />
;<br />
(5) a symplectic form ω = φ ∗ (ω P d−1), where ω P d−1 is the Fubini-Study symplectic<br />
form on P d−1 ;
436 KEVIN PURBHOO<br />
(6) a moment map µ for the (S 1 ) r -action on (X, ω); we can, in fact, write down<br />
the moment map µ explicitly:<br />
µ(x) =<br />
1<br />
∑ d<br />
i=1 |s−→ k i(x)| 2<br />
d∑<br />
|s −→ k i<br />
(x)| 2−→ k i ;<br />
to evaluate the right-hand side, we must choose a lifting of x to ˜X =<br />
Spec(C[A]); however, since this expression is homogeneous of degree zero<br />
in the s −→ k i<br />
, it is, in fact, well defined.<br />
It is well known that µ(X) = and that if Y is any other projective toric variety<br />
with µ Y (Y ) = , thenY ∼ = X as toric varieties.<br />
Let I ⊂ C[A] be a homogeneous ideal, and let V = Proj(C[A]/I) be its variety<br />
inside X.<br />
Definition 5.2 (Gel’fand, Kapranov, and Zelevinsky; see [GKZ])<br />
The compactified amoeba of V is µ(V ) ⊂ . We denote the compactified amoeba of<br />
V by either A V or A I (or by A f if I =〈f 〉 is principal).<br />
Let f ∈ C[A] be a homogeneous polynomial of degree w. We can again decompose<br />
f as a sum of monomials; that is, write f = ∑ l<br />
i=1 m i, where each m i is a T-weight<br />
vector in C[A]. Each of these m i is a well-defined function on ˜X. Leta ∈ . We<br />
define f {a} := {|m 1 (ã)|,...,|m l (ã)|}, where ã is any preimage of a in the composite<br />
map ˜X → X → . Of course, f {a} depends on the choice of lifting under ˜X → X,<br />
but only up to rescaling. Thus the notions of f {a} being lopsided or superlopsided are<br />
still well defined. We define LA f and SA f to be the set of points a ∈ such that<br />
f {a} is nonlopsided and nonsuperlopsided, respectively.<br />
Let V ◦ denote the intersection of V with the open dense subset of X on which<br />
T acts locally freely. (A finite quotient of T acts freely.) We can identify this open<br />
dense subset with (C ∗ ) r and therefore consider A V ◦. As both Log and µ| (C ∗ ) r are<br />
submersions with fibres (S 1 ) r , it follows that A V ◦ is diffeomorphic to A V ∩ ◦ ,<br />
where ◦ denotes the interior of . Letψ : ◦ → R r denote this diffeomorphism.<br />
Moreover, any face ′ of corresponds to a toric subvariety X ′ ⊂ X. And<br />
A V ∩ ′ = A V ∩X ′ (see [GKZ]).<br />
Thus, for every point a ∈ , we can determine whether a is in the compactified<br />
amoeba A V as follows. First, we determine the face ′ ⊂ for which a ∈ ( ′ ) ◦ .<br />
Then ψ ′ identifies ( ′ ) ◦ with R r′ in such a way that A V ∩ ( ′ ) ◦ is identified with<br />
A (V ∩X ′ ) ◦.Wethenhavea ∈ A V if and only if ψ ′(a) ∈ A (V ∩X ′ ) ◦.<br />
LEMMA 5.4<br />
The map ψ is uniformly continuous.<br />
i=1
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 437<br />
Proof<br />
The projective embedding φ induces a map (C ∗ ) r ↩→ (C ∗ ) d−1 , which is defined by<br />
monomials. This induces a linear map from R r → A (C ∗ ) r ⊂ Rd−1 .<br />
We also have a map from the moment polytope d−1 of (P d−1 ) to which is the<br />
projection induced by the inclusion of tori T ⊂ T d−1 .<br />
The composite<br />
R r −→ R d−1<br />
ψ d−1<br />
−−−→ d−1 −→ <br />
is ψ . Since the first and last maps are uniformly continuous, it suffices to show that<br />
ψ d−1 is uniformly continuous.<br />
This is fairly straightforward. Write µ d−1 = (µ 1 ,...,µ d−1 ).Forz ∈ (P d−1 ) ◦ ,<br />
we may write z = (1,z 1 ,...,z d ) and µ j (z) =|z j | 2 / ( 1 + ∑ d−1<br />
i=1 |z i| 2) .If|log|z j |−<br />
log|z<br />
j ′ ||
438 KEVIN PURBHOO<br />
COROLLARY 5.6<br />
We have A I = ⋂ f ∈I A f . In particular,<br />
A I = { a ∈ ∣ ∣ f {a} is not lopsided, ∀f ∈ I<br />
}<br />
.<br />
It is noted that Corollary 5.6 holds for all toric varieties with a moment map, not just<br />
the compact ones. However, the statement of uniform convergence in Corollary 5.5<br />
does not hold in general for noncompact toric varieties. For example, if one considers<br />
the toric variety C r , with the standard moment map µ(z) = (|z 1 | 2 ,...,|z r | 2 )/2, the<br />
convergence of the family LA ˜f n<br />
is almost never uniform. One can even see this in<br />
the simple example f (z) = (1 − z 1 ) ···(1 − z r ). The failure is that Lemma 5.4 does<br />
not hold; the map log|x| ↦→ |x| 2 /2 is not uniformly continuous, and so the uniform<br />
convergence does not carry over.<br />
It is most unfortunate that Proposition 5.3 does not easily carry over to the compact<br />
case. The use of elimination theory appears to be well suited only to the study of (C ∗ ) r<br />
with its particular standard symplectic form.<br />
6. Tropical varieties<br />
In this section, we show that Theorem 2 is the analytic counterpart to a theorem for<br />
tropical varieties. We have already seen examples of tropical hypersurfaces. Tropical<br />
varieties, in general, can be thought of as a generalisation of amoebas, where one<br />
replaces the norm |·|: C → R with a valuation in some nonarchimedean field. For<br />
this reason, tropical varieties are also known as nonarchimedean amoebas.<br />
Let K be an algebraically closed field with valuation v. For our purposes, a<br />
valuation on K is a map v : K → R trop , which satisfies the following conditions:<br />
• v(xy) = v(x) ⊙ v(y);<br />
• v(x + y) ≤ v(x) ⊕ v(y).<br />
This differs from the usual definition of a valuation in two purely cosmetic ways. First,<br />
a valuation is traditionally given as a map to v : K → R; we have simply translated it<br />
into the operations of R trop . Second, this is (−1) times the usual notion of a valuation.<br />
Our reasons for making these cosmetic changes becomes abundantly clear by the end<br />
of this section.<br />
To every f ∈ K[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ], we can associate a tropical polynomial as<br />
follows. If f = ∑ −→ b−→ k ∈A k<br />
z k 1<br />
1 ···zk r<br />
r , write<br />
f τ (x) = ⊕ −→ k ∈A<br />
v(b−→ k<br />
) ⊙ x −→ k<br />
= max<br />
−→ k ∈A<br />
{<br />
v(b −→ k<br />
) + x · −→ k } ,
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 439<br />
and call it the tropicalisation of f . We denote the tropical hypersurface associated to<br />
f τ by T f .<br />
If a ∈ R r trop<br />
, we can assign a weight to every monomial m ∈<br />
K[z 1 ,z −1<br />
1 ,...,z r,z −1 ]. Define the weight of m at a to be<br />
r<br />
wt a (m) := m τ (a).<br />
If f (z) = ∑ d<br />
i=1 m i(z), where m i are monomials, let<br />
f {a} τ = { wt a (m 1 ),...,wt a (m d ) } .<br />
Recall that in R trop , a list of numbers {b 1 ,...,b r } is (tropically) lopsided if the<br />
maximum element of this list does not occur twice (in which case, the maximum<br />
element is greater than the tropical sum of all the other elements). Thus f {a} τ is<br />
lopsided if and only if a /∈ T f .<br />
Let I ⊂ C[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ] be an ideal, and let<br />
V = V (I) = { z ∈ (K ∗ ) ∣ r f (z) = 0, ∀f ∈ I }<br />
be its affine variety. Let val :(K ∗ ) r → R r trop<br />
be the map<br />
val(z) = ( v(z 1 ),...,v(z n ) ) .<br />
The following theorem, as stated, most closely resembles the formulation in [SS],<br />
though variants of it have also appeared in [EKL]and[St].<br />
THEOREM 3 (Speyer and Sturmfels [SS, Theorem 2.1])<br />
The following subsets of R r trop coincide:<br />
(1) the closure of the set val(V );<br />
(2) the intersection of all tropical hypersurfaces ⋂ f ∈I T f j; and<br />
(3) the set of points a ∈ R r trop such that f {a} τ is not lopsided for all f ∈ I.<br />
This set is called the tropical variety of the ideal I.<br />
In fact, a stronger result than Theorem 3 (as stated here) is shown in [SS]. Let<br />
k denote the residue field of K. IfI ⊂ K[z 1 ,...,z r ], then one can construct an<br />
initial ideal of I, in k[z 1 ,...,z r ], corresponding to any weight a ∈ R trop . One can<br />
equivalently describe the tropical variety of I as the set of points a ∈ R trop such<br />
that the associated initial ideal contains no monomial. Thus the tropical variety is<br />
a subcomplex of the Gröbner complex, and there are algorithms to compute it (see<br />
[BJS+]).
440 KEVIN PURBHOO<br />
One can easily see that Theorem 3 is precisely analogous to the summary given in<br />
Remark 5.1. The proofs of these results, however, are extremely different. An obvious<br />
question, therefore, is whether analogous statements can be made in other contexts.<br />
The following is a general context in which one may hope for such a theorem to<br />
be true. Suppose that K is an algebraically closed field, and let S(⊙, ⊕, ≤) be a totally<br />
ordered semiring. Suppose that ‖·‖ K : K ∗ → S satisfies the following conditions:<br />
(1) ‖xy‖ =‖x‖ K ⊙‖y‖ K for all x,y ∈ K;<br />
(2) for all a,b ∈ S, wehave<br />
a ⊕ b = max { ‖x + y‖ K<br />
∣ ∣ ‖x‖K = a, ‖y‖ K = b } .<br />
In particular, condition (2) implies that ‖x + y‖ K ≤‖x‖ K ⊕‖y‖ K for all x,y ∈ K.<br />
Thus ‖·‖ K is an S-valued norm.<br />
Let f ∈ K[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ], and write f = ∑ d<br />
i=1 m i as a sum of monomials.<br />
For any point a ∈ S r ,letζ be such that ‖ζ‖ K = a. Wedefine<br />
f {a} := { ‖m 1 (ζ)‖ K ,...,‖m d (ζ)‖ K<br />
}<br />
.<br />
As ‖·‖ K is multiplicative, this is independent of the choice of ζ. SinceS is totally<br />
ordered, we can define a list of elements of S to be lopsided if and only if one number<br />
is greater than the sum of all the others.<br />
Let V ⊂ (K ∗ ) r be a variety defined by an ideal I ⊂ K[z 1 ,z −1<br />
1 ,...,z r,zr<br />
−1 ].We<br />
can consider the following sets:<br />
• The closure of {(‖z 1 ‖ K ,...,‖z r ‖ K ) | z ∈ V };<br />
• {a ∈ S r | f {a} is not lopsided, ∀f ∈ I}.<br />
The question is whether these two sets are equal for a particular (K,S,‖·‖ K ).<br />
In this article, we primarily discuss the example in which K = C, S = R + ,and<br />
‖·‖ C =|·|, and we show that they are equal. We have also just seen that this is true<br />
if K is a nonarchimedean field with ‖·‖ K as its valuation and S = R trop .<br />
Many (though not quite all) of the elements of the proof of Theorem 1 are valid in<br />
a more general context. Suppose that in addition to being a totally ordered semiring,<br />
S is a Q + -module (i.e., we can make sense of such things as (2/3)a for a ∈ S). For<br />
example, R trop is a Q + -module with trivial Q + -action.<br />
Define a binary operation ⊖ on S by<br />
a ⊖ b := min{c ∈ S | c ⊕ b ≥ a}<br />
whenever this set is nonempty. (We need not overly concern ourselves with the fact that<br />
a precise minimum may not exist: one can always get around the problem by treating<br />
this set as a Dedekind cut.) Then the triangle inequality ‖x − y‖ K ≥‖x‖ K ⊖‖y‖ K<br />
is valid (assuming that ‖x‖ K ≥‖y‖ K ). To see this, note that a ≤ a ′ implies that
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 441<br />
a ⊖ b ≤ a ′ ⊖ b; thus<br />
‖x‖ K ⊖‖y‖ K ≤ (‖x − y‖ K ⊕‖y‖ K ) ⊖‖y‖ K .<br />
Clearly, ‖x − y‖ K ∈{c ∈ S | c ⊕‖y‖ K ≤‖x − y‖ K ⊕‖y‖ K }, which implies that<br />
‖x − y‖ K ≥ (‖x − y‖ K ⊕‖y‖ K ) ⊖‖y‖ K .<br />
A closer examination of the proofs of Lemma 2.1 and Calculation A.3 now reveal<br />
that they are also valid (almost word for word) for a general (K,S,‖·‖ K ). We can<br />
also prove Lemma 3.5, in general, by replacing the integral over the torus<br />
1<br />
(2πi) r ∫|z 1 |=1<br />
∫ ( ∑<br />
···<br />
|z r |=1<br />
m−→ j<br />
(e a 1 z 1,...,e a r z r)<br />
−→ z l 1<br />
1 ···z l r<br />
r<br />
j<br />
by a discrete average over a finite subgroup of the torus<br />
1<br />
N r<br />
) dz1<br />
z 1 ···dz 1<br />
z 1<br />
∑ ∑ ( ∑ m−→ j<br />
(e a 1<br />
···<br />
z 1,...,e a r z r) )<br />
.<br />
z 1 :z1 N =1 z r :zr N =1 −→ z l 1<br />
1 ···z l r<br />
r<br />
j<br />
If N is suitably large, this discrete average has the same effect as the integral (i.e.,<br />
picking out a single term from the polynomial). In fact, we can follow the proof of<br />
Theorem 1(1), up to and including inequality (3.5). All that remains is to show that<br />
the right-hand side of (3.5) becomes sufficiently small as n gets large. Unfortunately,<br />
in general, this is not always true for all γ
442 KEVIN PURBHOO<br />
(1) c 1 n D 1 ((1 + γ n ) c 0n D 0<br />
− 1) < 1/2;<br />
(2) (1 + γ n ) c 0n D 0<br />
< 3/2.<br />
Proof<br />
We have<br />
Also,<br />
( (8 )<br />
n log γ −1 ≥ D 0 + D 1 log n + log 0 c 1<br />
3)c<br />
⇔ γ −n ≥<br />
( 8<br />
3)<br />
c 0 n D 0<br />
c 1 n D 1<br />
⇔ c 0 n D 0<br />
γ n 3<br />
≤ . (A.1)<br />
8c 1 n D 1<br />
3<br />
(<br />
1<br />
)<br />
≤ − 1 (<br />
1<br />
) 2<br />
8c 1 n D 1 2c 1 n D 1 2 2c 1 n D 1<br />
(<br />
< log 1 + 1 )<br />
2c 1 n D 1<br />
(A.2)<br />
( 3<br />
< log .<br />
2)<br />
(A.3)<br />
Using (A.1), (A.2), and the fact that log(1 + γ n )
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 443<br />
CALCULATION A.2<br />
Assume that c ≥ 1, 0
444 KEVIN PURBHOO<br />
Putting together (A.6)and(A.8),<br />
e (r+2)cnr−1 γ n − 1<br />
< (2−1−r )(2+ r)/(2 + r + 2 −1−r )<br />
2 − e cnr−1 γ n (2 + r)/(2 + r + 2 −1−r )<br />
= 2 −1−r . <br />
CALCULATION A.3<br />
For x>0 and s ∈ Z + ,<br />
∑<br />
( )<br />
w0 + s − 1 ∑ x w<br />
s − 1 w!
A <strong>NULLSTELLENSATZ</strong> <strong>FOR</strong> <strong>AMOEBAS</strong> 445<br />
References<br />
[A] M. AUDIN, The Topology of Torus Actions on Symplectic Manifolds, Progr. Math. 93,<br />
Birkhäuser, Basel, 1991. MR 1106194 433, 435<br />
[BM] U. BETKE and P. MCMULLEN, Lattice points in lattice polytopes, Monatsh. Math. 99<br />
(1985), 253 – 265. MR 0799674 418<br />
[BJS+] T. BOGART, A. N. JENSEN, D. SPEYER, B. STURMFELS,andR. R. THOMAS, Computing<br />
tropical varieties, J. Symbolic Comput. 42 (2007), 54 – 73. MR 2284285 439<br />
[CLO] D. COX, J. LITTLE,andD. O’SHEA, Using Algebraic Geometry, Grad. Texts in Math.<br />
185, Springer, New York, 1998. MR 1639811 434<br />
[EKL] M. EINSIEDLER, M. KAPRANOV,andD. LIND, Non-archimedean amoebas and tropical<br />
varieties, preprint, arXiv:math/0408311v2 [math.AG] 411, 439<br />
[FPT] M. <strong>FOR</strong>SBERG, M. PASSARE,andA. TSIKH, Laurent determinants and arrangements of<br />
hyperplane amoebas, Adv. Math. 151 (2000), 45 – 70. MR 1752241 426, 428<br />
[F] W. FULTON, Introduction to Toric Varieties, Ann. of Math. Stud. 131, Princeton Univ.<br />
Press, Princeton, 1993. MR 1234037 435<br />
[GKZ] I. M. GEL’FAND, M. M. KAPRANOV, andA. V. ZELEVINSKY, Discriminants,<br />
Resultants, and Multidimensional Determinants, Math. Theory Appl., Birkhäuser,<br />
Boston, 1994. MR 1264417 407, 436<br />
[M] G. MIKHALKIN, “Amoebas of algebraic varieties and tropical geometry” in Different<br />
Faces of Geometry, Int. Math. Ser. (N.Y.) 3, Kluwer/Plenum, New York, 2004,<br />
257 – 300. MR 2102998 408<br />
[PR] M. PASSARE and H. RULLGÅRD, Amoebas, Monge-Ampère measures, and<br />
triangulations of the Newton polytope, Duke Math. J. 121 (2004), 481 – 507.<br />
MR 2040284 429<br />
[R] L. I. RONKIN, “On zeros of almost periodic functions generated by functions<br />
holomorphic in a multicircular domain” (in Russian) in Complex Analysis in<br />
Modern Mathematics, FAZIS, Moscow, 2001, 239 – 251. MR 1833516 429<br />
[Ru] H. RULLGÅRD, Polynomial amoebas and convexity, preprint, 2001. 429<br />
[S] D. SPEYER, personal communication, 2003. 410<br />
[SS] D. SPEYER and B. STURMFELS, The tropical Grassmannian, Adv. Geom. 4 (2004),<br />
389 – 411. MR 2071813 411, 439<br />
[St] B. STURMFELS, Solving Systems of Polynomial Equations, CBMS Regional Conf. Ser.<br />
in Math. 97, Amer. Math. Soc., Providence, 2002. MR 1925796 411, 439<br />
[T] T. THEOBALD, Computing amoebas, Experiment. Math. 11 (2002), 513 – 526.<br />
MR 1969643 408, 428<br />
Department of Combinatorics and Optimization, University of Waterloo, Waterloo, Ontario<br />
N2L 3G1, Canada; kpurbhoo@math.uwaterloo.ca
ENDOSCOPIC LIFTING IN CLASSICAL GROUPS<br />
AND POLES OF TENSOR L-FUNCTIONS<br />
DAVID GINZBURG<br />
Abstract<br />
In this article, we introduce a new construction of endoscopic lifting in classical<br />
groups. To do that, we study a certain small representation and use it as a kernel<br />
function to construct the liftings. As an application of the construction, we study<br />
the relations of poles of tensor L-function with certain liftings and certain period<br />
integrals.<br />
Contents<br />
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447<br />
2. Notation and basic definitions . . . . . . . . . . . . . . . . . . . . . . 450<br />
3. The cuspidality of the lift . . . . . . . . . . . . . . . . . . . . . . . . . 465<br />
4. The nonvanishing of the lift . . . . . . . . . . . . . . . . . . . . . . . . 474<br />
5. The unramified computations . . . . . . . . . . . . . . . . . . . . . . . 484<br />
6. Liftings and poles of tensor L-functions . . . . . . . . . . . . . . . . . 488<br />
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501<br />
1. Introduction<br />
This article presents a new construction for endoscopic liftings for classical groups. We<br />
consider five cases, all of which are described in Definition 2. For an example of this<br />
construction, consider the following case. Corresponding to the homomorphism of L-<br />
groups SO 2n+1 (C)×SO 2m (C) ↦→ SO 2(n+m)+1 (C), the Langlands conjectures predict a<br />
lifting from automorphic representations of the group Sp 2n (A)×SO 2m (A) to automorphic<br />
representations of Sp 2(n+m) (A). Thus, our goal in this article is the following: given<br />
two cuspidal generic irreducible representations on the groups Sp 2n (A) and SO 2m (A),<br />
we construct a generic cuspidal representation defined on Sp 2(n+m) (A) which corresponds<br />
to the above lifting. These types of liftings are examples of what is known as<br />
endoscopic lifting. Other examples of these liftings were constructed in [GRS7] using<br />
the descent method (for more on L-functions and liftings, see [L1], [A], [B]).<br />
DUKE MATHEMATICAL JOURNAL<br />
Vol. 141, No. 3, c○ 2008 DOI 10.1215/00127094-2007-002<br />
Received 12 October 2006. Revision received 18 May 2007.<br />
2000 Mathematics Subject Classification. Primary 11F70; Secondary 22E55.<br />
Author’s work partially supported by Israel Science Foundation grant 162/06.<br />
447
448 DAVID GINZBURG<br />
The method we use to construct these liftings is what we referred to in [G2]asthe<br />
theta lifting method. By that we mean that we construct a representation, specifically<br />
a residue of a certain Eisenstein series, defined on a certain group M(A) and then use<br />
it as a kernel function in order to construct our lifting. For example, in the above case,<br />
we start with a cuspidal representation τ = τ(ɛ) of the group GL 2m (A) whichisa<br />
functorial lift from a cuspidal representation ɛ of SO 2m (A) as constructed in [CKPS].<br />
We then use this representation to construct a residue representation ɛ defined on the<br />
group M(A) = Sp 2m(2n+1) (A). Using a suitable unipotent integration, we obtain a copy<br />
of the group Sp 2n × Sp 2(n+m) embedded inside M. Starting with a generic cuspidal<br />
representation σ of Sp 2n (A), we pair it against the above kernel representation, thus<br />
obtaining an automorphic representation defined on Sp 2(n+m) (A). More precisely, the<br />
automorphic representation that we obtain is the space generated by all functions of<br />
the form<br />
∫ ∫<br />
( )<br />
f (h) =<br />
ϕ σ (g)θ ɛ u(g, h) ψU (u) dudg.<br />
Sp 2n (F )\Sp 2n (A) U(F )\U(A)<br />
Here U is a unipotent group with a character ψ U defined on it. The functions ϕ σ and θ ɛ<br />
are vectors in the representation space of σ and ɛ , respectively; also, h ∈ Sp 2(n+m) (A)<br />
(for more details, see Section 2.3). It should be mentioned that our construction works<br />
if we replace the cuspidal representation ɛ on GL 2m (A) with a generic automorphic<br />
representation that is a constituent of an induced representation from cuspidal data.<br />
Sections 3 – 5 contain the setup of the construction and the basic properties of the<br />
lifting. Since the ideas behind the proofs are similar in all cases, we concentrate on one<br />
example, namely, the example mentioned above. Since most of the proofs of the basic<br />
properties are now quite standard, we have allowed ourselves a certain sketchiness in<br />
some of the technical details.<br />
Section 2.2 is devoted to the construction of the representation ɛ and the study<br />
of its basic properties. As mentioned above, this representation is defined as a residue<br />
of an Eisenstein series. In Section 2.3, we define the various lifts that we intend to<br />
construct and set up the global integral that we use for this construction. Section 3 is<br />
devoted to the proof of the cuspidality of the lift; this is done via direct calculations<br />
of the various constant terms. As usual in these types of construction, there is an<br />
obstruction for the lift to be cuspidal (see the beginning of Section 3 for more details).<br />
In Section 4, we prove the nonvanishing of the lift by showing that the representation<br />
of Sp 2(n+m) (A) obtained by the above integral is, in fact, a generic representation,<br />
under the assumption that σ is generic. As a consequence of Theorem 3, the main<br />
theorem in that section, we obtain that the lift is indeed nonzero. Section 5 is devoted<br />
to proving that the lift constructed is indeed functorial. We do this by computing the<br />
standard local L-function of the lift using the basic identity we derived in Theorem 3.
ENDOSCOPIC LIFTING 449<br />
In the last section, Section 6, we apply our construction to relate liftings with<br />
period integrals and poles of L-functions. We consider two cases. In the first case, we<br />
characterize the existence of a simple pole to standard tensor L-functions. (The main<br />
statement is Conjecture 1, together with Theorem 6.) More precisely, we prove the<br />
following theorem.<br />
MAIN THEOREM<br />
Let π denote an irreducible generic cuspidal representation of the orthogonal group<br />
SO 2(n+m)+1 (A), and let ɛ denote an irreducible generic cuspidal representation of<br />
SO 2m+1 (A). Letτ(ɛ) denote the lift of ɛ to GL 2m (A) whose existence was proved<br />
in [CKPS]. Assume that τ(ɛ) is a cuspidal representation. Then the following are<br />
equivalent.<br />
(1) The partial tensor L-function L S (π × ɛ, s) = L S (π × τ(ɛ),s) has a simple<br />
pole at s = 1. HereS is a finite set of places including the archimedean ones<br />
such that outside of S, all data is unramified.<br />
(2) There is a choice of data such that the period integral P(π, τ(ɛ)), defined in<br />
the proof of Theorem 6, is not zero for some choice of data.<br />
(3) There is a generic cuspidal representation σ of SO 2n+1 (A) such that π is the<br />
weak endoscopic lift from σ and ɛ.<br />
Two implications of this theorem follow from other references (see the first paragraph<br />
following Conjecture 1). The implication that statement (2) implies statement (3) is<br />
proved in Theorem 6 using our construction of the lifting.<br />
The second application that we discuss in Section 6 is more conjectural. Motivated<br />
by the low-rank cases that we describe in that section, we give a conjecture as to when<br />
the L-function associated to the tensor product of two Spin representations can have<br />
poles at certain values. Although the main conjecture is stated in Conjecture 3, we<br />
have more evidence in a special case of that conjecture stated in this form.<br />
CONJECTURE 2<br />
Let π and ɛ denote irreducible cuspidal generic representations of the groups<br />
GSO 2(m+4) (A) and GSp 2m (A). The following statements are equivalent.<br />
(1) The partial L-function L S (π × ɛ, Spin 2(m+4) × Spin 2m+1 ,s) has a simple pole<br />
at s = 1.<br />
(2) The period integral Q(π, ɛ), described in Section 6, is not zero for some choice<br />
of data.<br />
(3) There exists a cuspidal generic representation ν of the exceptional group G 2 (A)<br />
such that π is the weak functorial lift from ν and from ɛ.<br />
In other words, this conjecture ties the poles of the L-function of a tensor product<br />
of Spin representations, with a certain period integral and with a lifting related to<br />
the exceptional group G 2 . Indeed, Conjecture 2 is related to the following lifting.
450 DAVID GINZBURG<br />
Let ν denote a cuspidal automorphic representation of the exceptional group G 2 (A),<br />
and let ɛ denote a cuspidal representation of GSp 2m (A). Corresponding to the L-<br />
groups homomorphism G 2 (C) × GSpin 2m+1 (C) ↦→ GSpin 2m+8 (C), the Langlands<br />
conjectures predict a lifting from ν and ɛ to an automorphic representation defined<br />
on the group GSO 2m+8 (A). This lifting is not an endoscopic lifting. As we state in<br />
Theorem 9, Conjecture 2 is, in fact, a theorem in the case where m = 1; thiswas<br />
provedin[GH2]. Our contribution to this conjecture is the implication that statement<br />
(2) implies (3), which we prove in Theorem 10. To prove this theorem, we use our<br />
construction of the lifting and extend it to similitude groups.<br />
It is worth mentioning that the conjectures that we make and the theorems that<br />
we prove are also important in the study of the Langlands conjectures. These results<br />
tie the existence of poles of L-functions to a certain lifting. In addition, the image<br />
of the lift is characterized by a certain period integral. To prove these results, it is<br />
necessary to combine the so-called Rankin-Selberg method together with what we<br />
designate the theta lifting method. In our study of Conjecture 1, the fact that the<br />
corresponding L-function had an integral representation using the Rankin-Selberg<br />
method is crucial; Eulerian Rankin-Selberg integrals are a relatively rare phenomena.<br />
Indeed, the L-functions studied in Conjecture 2 do not have, as far as we know,<br />
such an integral representation except when m = 1. All this indicates that in order<br />
to study such conjectures, one also has to consider Rankin-Selberg integrals that<br />
are not Eulerian. It is conceivable that the study of residues of L-functions can be<br />
done by non-Eulerian integrals. This is clearly a step toward studying conjectures of<br />
the type stated in Conjecture 2. Another example of this phenomena was studied in<br />
[BFG, page 290], where a global period is given which conjecturally characterizes<br />
the image of the cubic Shimura lift for the group PGL 3 . The given period integral<br />
involves the minimal representation of the group SO 8 , which is a residue of a degenerate<br />
Eisenstein series. A possible way to study this period integral is to replace this<br />
minimal representation with the degenerate Eisenstein series and to unfold the integral.<br />
This way, one obtains a non-Eulerian integral whose residue, at least conjecturally,<br />
characterizes a possible lift.<br />
Finally, we point out two possible extensions of our construction. First, we can<br />
consider representations σ that are not generic. Another possible extension is to replace<br />
the representation ɛ with a representation that is nearly equivalent to it. In both cases,<br />
we expect at least some of the results stated in this article to still be valid. We hope to<br />
address these cases in the near future.<br />
2. Notation and basic definitions<br />
2.1. General notation<br />
Let M denote a split classical group of type B,C, orD. In terms of matrices,<br />
we represent these groups with respect to the following forms. Let J n denote the
ENDOSCOPIC LIFTING 451<br />
(n × n)-matrix, with 1’s along the second diagonal. For the orthogonal groups, we<br />
use the form corresponding to the matrix J n for a suitable value of n. The symplectic<br />
groups are represented with respect to the form ( J n<br />
)<br />
−J n . For the symplectic groups,<br />
we denote Mat 0 r×r ={A ∈ Mat r×r : A t J n = J n A}. For orthogonal groups, we denote<br />
Mat 0 r×r ={A ∈ Mat r×r : A t J n =−J n A}. Given a matrix X ∈ Mat n×n , we denote<br />
X ∗ = J n (X t ) −1 X.<br />
When M = Sp 2n , we use ˜M to denote its double cover.<br />
Given a parabolic subgroup of M, we refer to it as a standard parabolic subgroup of<br />
M if it contains the Borel subgroup of M which consists of upper triangular matrices.<br />
Let F denote a global field, and let A denote its ring of adeles. Let denote an<br />
automorphic representation defined on M(A).<br />
Given a subgroup U of M, we denote<br />
∫<br />
θ U (g) = θ(ug) du.<br />
U(F )\U(A)<br />
Here θ is a vector in the space of the representation .<br />
Let ψ denote a nontrivial additive character of the group F \A. Letψ U denote a<br />
character of the group U(F )\U(A). We denote<br />
∫<br />
θ U,ψ U<br />
(g) = θ(ug)ψ U (u) du.<br />
U(F )\U(A)<br />
2.2. Definitions and properties of some small representations<br />
In this section, we define and study the properties of certain small representations in<br />
the symplectic and split orthogonal groups obtained by residues of Eisenstein series.<br />
For basic definitions and properties of Eisenstein series, we refer the reader to [L2]or<br />
[MW]. The results of this section are quite straightforward and follow [G2], [GRS9],<br />
and [GRS7] and the references listed in those articles. There are five cases to consider.<br />
We construct the representation that we need in one case, while the other cases are<br />
mentioned at the end of this section, and their proofs are similar.<br />
Let τ denote a cuspidal representation of GL 2m (A). In the notation of<br />
[G2, Section 2], our goal is to study the multivariable Eisenstein series E τ (g, ¯s)<br />
defined on the group Sp 4mn (A). The main result is the following.<br />
PROPOSITION 1<br />
The Eisenstein series has a simple pole at the point ¯s = ¯s 0 in the sense that the limit<br />
lim¯s→¯s0 (¯s − ¯s 0 )E τ (g, ¯s) is nonzero. Here ¯s 0 is defined as in [G2, Section 2]. As a<br />
function of g ∈ Sp 4mn (A), we denote this representation by τ .
452 DAVID GINZBURG<br />
Proof<br />
The proof comes by analyzing the constant term along the unipotent group V of<br />
the Eisenstein series. Recall from [G2] thatE τ (g, ¯s) is associated to the induced<br />
(τ ⊗···⊗τ)δ¯s¯Q . Here Q = (GL 2m ×···×GL 2m )V , where<br />
V is the unipotent radical of Q. Thus, in the notation of Section 2.1, we consider the<br />
constant term Eτ V (g, ¯s). The first statement is that<br />
representation Ind Sp 4mn (A)<br />
Q(A)<br />
E V τ (g, ¯s) = ∑ w∈W 0<br />
M w (f τ , ¯s)(g),<br />
where W 0 and M w are defined as in [G2, (2.3)] and in the text following that formula.<br />
To prove this, we argue as in [GRS7, Sections 2.1, 2.2]. In fact, our Eisenstein series<br />
is a special case of the Eisenstein series considered in [GRS7, Section 2]. The point<br />
is that for any Weyl element w not in W 0 , when considering the conjugation wV w −1 ,<br />
we end up integrating the cuspidal representation τ along a unipotent radical of a<br />
parabolic subgroup of GL 2m . Hence we get zero.<br />
In terms of matrices, the set W 0 can be described as follows. Let ˜W denote the<br />
Weyl group of Sp 2n ; in matrices, we may choose a set of representatives to be all<br />
permutation matrices of Sp 2n whose nonzero elements consist of 1’s and –1’s. Take<br />
such an element, and replace each 1 by the identity matrix of size 2m; follow a similar<br />
process with each −1.<br />
Arguing as in [G2, (2.4) – (2.11)], we deduce that the following limits lim¯s→¯s0 (¯s −<br />
¯s 0 )M w (f τ , ¯s)(g) are, in fact, zero, except when we take w, which corresponds to the<br />
long Weyl element in ˜W. In other words, the above limits are all zero, except when<br />
we take w to have blocks of identity matrix of size 2m on the second diagonal. In this<br />
case, we obtain from [G2, (2.9)] the fact that M w (f τ , ¯s)(g) is equal to<br />
∏<br />
n∏ L S (τ,ρ i ,ζ i )<br />
M w,ν (f τν , ¯s)(g ν )<br />
L S (τ,ρ i ,ζ i + 1) LS τ (¯s),<br />
ν∈S<br />
i=1<br />
where the notation is as defined in [G2, (2.10), (2.11)]. Computing the limit in this<br />
case, we get a nonzero factor.<br />
<br />
The next step is to determine which Fourier coefficients the representation τ supports<br />
and which it does not. This is best expressed in terms of the structure of the unipotent<br />
orbits associated with the group Sp 4mn . The idea of relating representations with<br />
unipotent orbits is not new (e.g., see [Sp]). We mainly use the description of unipotent<br />
orbits as given in [CM]. The way in which we associate unipotent orbits with Fourier<br />
coefficients is explained in detail in [G3]; we do not review it here. As in [G2,<br />
Definition 3], we have the following.
ENDOSCOPIC LIFTING 453<br />
Definition 1<br />
Let π denote an automorphic representation of G = Sp 4mn . We denote by O G (π)<br />
the set of all unipotent classes of H with the following property. A unipotent class<br />
O ∈ O G (π) if, for all unipotent classes Õ > O, the representation π does not have a<br />
nonzero Fourier coefficient corresponding to Õ. When there is no confusion, we write<br />
O(π) for O G (π).<br />
The main result of Section 2 is the following theorem.<br />
THEOREM 1<br />
We have O( τ ) = ((2m) 2n ).<br />
Proof<br />
To prove this theorem, we need to prove two things. First, we need to prove that given a<br />
unipotent orbit O of Sp 4mn which is greater or not related to ((2m) 2n ), the representation<br />
τ has no nonzero Fourier coefficient that corresponds to O. Then we need to show<br />
that τ has a nonzero Fourier coefficient that corresponds to the unipotent orbit<br />
((2m) 2n ). The first part follows as in [GRS9, Sections 2, 3], but by replacing n with<br />
m and k with n. Indeed, the proof is local. We show that the unramified constituents<br />
of the residue cannot support any functional that is induced from a global Fourier<br />
coefficient corresponding to a unipotent orbit that is bigger than or not related to<br />
((2m) 2n ). It follows from [GRS9, Definition 2.1, Lemma 3.1] that the local unramified<br />
constituents are the same as in our case defined above. By applying [GRS9, Lemma 3.3,<br />
Proposition 3.6], we may deduce that a similar result holds for τ defined here.<br />
Thus we need only prove the second part, namely, that τ has a nonzero Fourier<br />
coefficient that corresponds to the unipotent orbit ((2m) 2n ).Let̂Q denote the standard<br />
parabolic subgroup of Sp 4mn whose Levi part is GL 2n ×···×GL 2n .LetU denote its<br />
unipotent radical. In terms of matrices, we have<br />
⎧⎛<br />
⎞<br />
⎫<br />
I 2n X 1 ∗<br />
I 2n X 2 ∗<br />
. ..<br />
⎪⎨<br />
Xm−1 ∗<br />
U =<br />
I 2n Y<br />
⎪⎬<br />
I 2n Xm−1<br />
∗ : X i ∈ Mat 2n×2n ,Y∈ Mat 0 2n×2n .<br />
.<br />
⎜<br />
.. ⎟<br />
⎝<br />
⎪⎩<br />
I 2n X1<br />
∗ ⎠<br />
⎪⎭<br />
I 2n<br />
On U(F )\U(A), we define a character ψ U as follows. For Y = (Y i,j ),defineψ U (u) =<br />
ψ(tr(X 1 +···+X m−1 ) + tr ′ (Y )), where tr ′ (Y ) = Y 1,1 +···+Y n,n .<br />
For a vector θ τ in the space of τ , consider the Fourier coefficient θ U,ψ U<br />
τ<br />
(g).<br />
Following [G3], it is not hard to check that this Fourier coefficient does indeed
454 DAVID GINZBURG<br />
correspond to the unipotent orbit ((2m) 2n ). One can also check that the stabilizer<br />
inside the Levi part of ̂Q is the group SO 2m , which is the stabilizer of this unipotent<br />
orbit (see [CM, Theorem 6.1.3], which is due to Springer and Steinberg). Our goal is<br />
to prove that θ U,ψ U<br />
τ<br />
(g) is nonzero for some choice of data. We assume that it is zero<br />
for every choice of data, and we derive a contradiction.<br />
Let w denote the Weyl element defined as follows. For 1 ≤ i ≤ 2m and 0 ≤ j ≤<br />
2n − 1,the(2mj + i, 2n(i − 1) + j)-entry of w is ±1, and all other entries are zero.<br />
The entry is −1 if 2mj + i>2mn and 2n(i − 1) + jiand z i,i = I 2m . Finally, for i>j, the matrix z i,j is upper unipotent<br />
with zero along the diagonal. Next, the matrix r 4mn = (r i,j ), where r i,j ∈ Mat 2m×2m ,<br />
is such that r i,j = 0 if i>j,andr i,i = I 2m .Fori
ENDOSCOPIC LIFTING 455<br />
and only if the integral<br />
∫<br />
θ τ (vy 4mn )ψ m (y 4mn ) dv dy 4mn<br />
is zero for every choice of data.<br />
Recall that V is the unipotent radical of the parabolic subgroup Q which was<br />
defined in the beginning of the proof of Proposition 1. Arguing as in [GRS5, Theorem 1,<br />
pages 889 – 894], we conclude that the above Fourier coefficient is nonzero for some<br />
choice of data. Thus, we obtain a contradiction.<br />
<br />
We also need the following local result.<br />
PROPOSITION 2<br />
Let F be a nonarchimedean local field. For this proposition only, let θ τ denote the<br />
local constituent of τ at the place F . Assume that θ τ is unramified. Let l denote a<br />
functional defined on θ τ which satisfies the property l(vy 4mn ω) = ψm −1(y<br />
4mn)l(ω) for<br />
all v ∈ V , for all y 4mn that were defined in the proof of Theorem 1, and for all vectors<br />
ω ∈ θ τ . Then the space of such functionals is 1-dimensional.<br />
Proof<br />
Let L denote the maximal unipotent subgroup of Sp 4mn . Thus, every element in L<br />
has a unique factorization as vy 4mn .From[G2, Lemma 3.1], we deduce that θ τ is a<br />
quotient of the induced representation Ind Sp 4mn<br />
̂Q (χ 1 ⊗···⊗χ m ) δ 1/2 . Here ̂Q 1<br />
̂Q is the<br />
standard parabolic subgroup of Sp 4mn whose Levi part is GL 2n ×···×GL 2n ,and<br />
χ i are certain unramified characters. To prove the proposition, we apply the Bruhat<br />
theory; thus, it is enough to prove the following. We say that g ∈ ̂Q\Sp 4mn /L is<br />
not admissible if we can find an element y 4mn as above, so that ψ m (y 4mn ) ≠ 1 and<br />
gy 4mn g −1 ∈ ̂Q. Otherwise, we say that g is admissible. To prove the proposition, it is<br />
enough to show that there is at most one admissible double coset. This proves that the<br />
space of such functionals is at most 1-dimensional. From Theorem 1, it follows that<br />
the space of such functionals is exactly 1-dimensional.<br />
We can choose elements in ̂Q\Sp 4mn /L as Weyl elements modulo from the left by<br />
Weyl elements in GL 2n ×···×GL 2n . We choose the Weyl elements to be permutation<br />
matrices. Let w be such a Weyl element. Consider the first n rows of w. We claim that<br />
for w to be admissible, we must have w i,2ni ≠ 0 whenever 1 ≤ i ≤ n. Indeed, if not,<br />
then from the definition of ψ m , we can find a matrix in y 4mn so that ψ m (y 4mn ) ≠ 0<br />
and wy 4mn w −1 ∈ ̂Q. Indeed,iffor1 ≤ i ≤ n we have w i,j ≠ 0 and j ≠ 2ni, then<br />
the matrix x(r) = I + re j,j+1 satisfies wx(r)w −1 ∈ ̂Q and ψ(x(r)) ≠ 1. Here I is<br />
the (4mn)-identity matrix, and e p,q is the (4mn)-matrix with 1 at the (p, q)-entry and<br />
zero elsewhere.
456 DAVID GINZBURG<br />
Next, consider the rows 2n + i for 1 ≤ i ≤ 2n. We claim that if w is admissible,<br />
then w 2n+i,2ni−1 ≠ 0. Indeed, suppose that w 2n+i,j ≠ 0 for some 1 ≤ i ≤ 2n and<br />
that j ≠ 2ni − 1. Then, using the Weyl group of GL 2n ×···×GL 2n if needed, we<br />
can find l>2n + i so that w l,j+1 ≠ 0. This means that x(r) = I + re j,j+1 satisfies<br />
wx(r)w −1 ∈ ̂Q.Sincex(r) is in L, and since ψ m (x(r)) ≠ 1,thenw is not admissible.<br />
Thus w 2n+i,2ni−1 ≠ 0 for all 1 ≤ i ≤ 2n. Continuing by induction, we see that if w is<br />
admissible, then w is also uniquely determined.<br />
<br />
Returning to the global situation, the following proposition follows immediately from<br />
Theorem 1.<br />
PROPOSITION 3<br />
Let θ U,ψ U<br />
τ<br />
(g) denote the Fourier coefficient corresponding to the unipotent orbit<br />
((2m) 2n ), as was described in Theorem 1. Letg ∈ SO 2n (A), which are the adelic<br />
points of the stabilizer of this Fourier coefficient (see the proof of Theorem 1). Then<br />
θ U,ψ U<br />
τ<br />
(g) = θ U,ψ U<br />
τ<br />
(e). In other words, θ U,ψ U<br />
τ<br />
(g) is invariant under all g ∈ SO 2n (A).<br />
Proof<br />
The idea is similar to the one sketched in [GRS8, Theorem 2.1]. Assume first that<br />
n ≥ 2. As mentioned above, the stabilizer of the character ψ U is the split orthogonal<br />
group SO 2n (A).Letx(r) denote the 1-parameter subgroup of SO 2n (A) corresponding<br />
to the highest-weight root vector in this group. If θ U,ψ U<br />
τ<br />
(g) were not left-invariant<br />
under elements g ∈ SO 2n (A), it would follow that for some a ∈ F ∗ , the integral<br />
∫<br />
F \A<br />
( )<br />
θ U,ψ U<br />
τ x(r)g ψ(ar) dr<br />
is not zero for some choice of data. But as mentioned in [GRS8], this last integral<br />
corresponds to a unipotent orbit that is strictly greater than ((2m) 2n ). Indeed, to show<br />
this, let U 0 denote the unipotent subgroup of U defined as follows. Let u = (u i,j ) ∈ U.<br />
We consider all matrices u ∈ U so that u i,j = 0 for all pairs (i, j) ∈{(2nk − 1, 2nk +<br />
1), (2nk − 1, 2nk + 2), (2nk, 2nk + 1), (2nk, 2nk + 2) : 1 ≤ k ≤ m}. If we restrict<br />
ψ U to U 0 , we obtain a character of U 0 (F )\U 0 (A) which we continue to denote by ψ U .<br />
Next, we define another unipotent subgroup of Sp 4mn , which we denote by U 1 ,<br />
which contains U 0 . We define this group by the unipotent group generated by U 0 ,<br />
and all 1-parameter unipotent matrices I 4mn + re<br />
i,j ′ , where (i, j) ∈{(2nk + 1, 2nl −<br />
2), (2nk + 1, 2nl − 1), (2nk + 2, 2nl − 2), (2nk + 2, 2nl − 1) : 0 ≤ k ≤ m − 1; 1 ≤<br />
l ≤ m},ande<br />
i,j ′ = e i,j − e 4mk−j+1,4mk−i+1 .<br />
Consider the above integral. We now perform certain Fourier expansions. Let<br />
y(r 1 ,r 2 ,r 3 ,r 4 ) = I 4mn + r 1 e ′ 1,2n−1 + r 2e ′ 1,2n + r 3e ′ 2,2n−1 + r 4e ′ 2,2n ,
ENDOSCOPIC LIFTING 457<br />
and let<br />
z(r 1 ,r 2 ,r 3 ,r 4 ) = I 4mn + r 1 e ′ 2n−1,2n+1 + r 2e ′ 2n,2n+1 + r 3e ′ 2n−1,2n+2 + r 4e ′ 2n,2n+2 .<br />
Notice that the group of matrices generated by z(r 1 ,r 2 ,r 3 ,r 4 ) is a subgroup of U.<br />
Expanding the above integral along the group y(r 1 ,r 2 ,r 3 ,r 4 ) with r i integrated over<br />
points in A modulo points in F , we obtain<br />
∑<br />
∫<br />
θ U,ψ U<br />
τ<br />
α i ∈F<br />
(F \A) 5<br />
(<br />
x(r)y(r1 ,r 2 ,r 3 ,r 4 )g ) ψ(<br />
ar +<br />
4∑ )<br />
α i r i dr dr i .<br />
From the fact that θ τ (g) is an automorphic function, it follows that it is left-invariant<br />
under the rational points. Using that, conjugating with the matrices z(α 1 ,α 2 ,α 3 ,α 4 )<br />
from left to right, and collapsing summation with integration, the above integral is<br />
equal to<br />
∫<br />
∫<br />
A 4 F \A<br />
i=1<br />
θ U 1,1,ψ U<br />
(<br />
τ x(r)z(r1 ,r 2 ,r 3 ,r 4 )g ) ψ(ar) dr dr i .<br />
Here U 1,1 is the unipotent subgroup of U 1 generated by all unipotent matrices of the<br />
form I 4mn + re<br />
i,j ′ , where (i, j) ∈{(2n − 1, 2n + 1), (2n − 1, 2n + 2), (2n, 2n +<br />
1), (2n, 2n + 2)} and the subgroup of U which consists of all matrices u = (u i,j ) ∈ U<br />
such that u i,j = 0 for all (i, j) ∈{(1, 2n − 1), (1, 2n), (2, 2n − 1), (2, 2n)}. Clearly,<br />
it is enough to prove the vanishing of the inner integral in the above integration. We<br />
continue this process inductively, along the corresponding subgroups of U 0 and U 1 ,<br />
and we finally obtain that it is enough to prove the vanishing of the integral<br />
∫<br />
θ U 1,ψ U<br />
( )<br />
τ x(r)g ψ(ar) dr,<br />
F \A<br />
where we view ψ U as a character of U 1 by its restriction to U 0 . It follows from the<br />
definition of the correspondence between unipotent orbits and Fourier coefficients, as<br />
explained in [GRS8] or[G3], that this last integral is associated with the unipotent<br />
orbit ((2m + 1) 2 (2m) 2n−4 (2m − 1) 2 ). This unipotent orbit is, of course, greater than<br />
((2m) 2n ). Hence the above integral is zero for every choice of data. This produces a<br />
contradiction; therefore, the above left-invariant property holds.<br />
The case where n = 1 is still true but treated differently. Indeed, in that case, the<br />
stabilizer of ψ U is just a torus, and by using certain Fourier expansions, similar to the<br />
ones performed above, we obtain our result.
458 DAVID GINZBURG<br />
To state the following lemma, let N m denote the standard unipotent radical subgroup<br />
of the maximal parabolic of Sp 4mn whose Levi part is GL 2m ×Sp 4m(n−1) .LetV (GL 2m )<br />
denote the standard maximal unipotent subgroup of GL 2m .Wedenotebyψ V (GL2m ) the<br />
Whittaker character of V (GL 2m ).LetN 0 m denote the subgroup of N m defined as<br />
⎧⎛<br />
⎞<br />
⎨ I 2m Y Z<br />
N 0 m = ⎝ I<br />
⎩<br />
4m(n−1) Y ∗ ⎠,Y ∈ Mat 2m×4m(n−1) : Y 2m,i = 0,<br />
I 2m<br />
⎫<br />
⎬<br />
Z ∈ Mat 0 2m×2m : Z 2m,1 = 0<br />
⎭ .<br />
Denote<br />
⎛<br />
X<br />
˜X = ⎝<br />
I 4m(n−1)<br />
X ∗ ⎞<br />
⎠, X ∈ V (GL 2m ).<br />
We have the following.<br />
LEMMA 1<br />
For every choice of data, we have the identity<br />
∫ ∫<br />
θ τ (y˜X)ψ V (GL2m )(X) dy dX =<br />
Nm 0 (F )\N m 0 (A) ∫ ∫<br />
N m (F )\N m (A)<br />
θ τ (y˜X)ψ V (GL2m )(X) dy dX,<br />
where X is integrated over V (GL 2m )(F )\V (GL 2m )(A).<br />
Proof<br />
We start by considering the Fourier expansion of the left-hand-side integral, along<br />
the unipotent group I 4mn + re 2m,4m(n−1)+1 . The contribution to the expansion from the<br />
nontrivial characters is zero. Indeed, each term corresponding to a nontrivial character<br />
corresponds to the unipotent orbit ((2m + 2)1 4mn−2m−2 ). From Theorem 1, it follows<br />
that this Fourier coefficient is zero. Thus we are left with only the trivial contribution.<br />
Hence, the left-hand side of the above identity equals<br />
∫ ∫<br />
θ τ (y˜X)ψ V (GL2m )(X) dy dX,<br />
Nm 1 (F )\N m 1 (A)<br />
where now
ENDOSCOPIC LIFTING 459<br />
⎧⎛<br />
⎞<br />
⎫<br />
⎨ I 2m Y Z<br />
⎬<br />
N 1 m = ⎝ I<br />
⎩<br />
4m(n−1) Y ∗ ⎠,Y ∈ Mat 2m×4m(n−1) : Y 2m,i = 0,Z ∈ Mat 0 2m×2m⎭ .<br />
I 2m<br />
Next, we expand the above integral along the unipotent group N m /Nm 1 with points in A<br />
modulo points in F . This is an abelian group, and Sp 4m(n−1) (F ) acts on this expansion<br />
with two orbits. Here Sp 4m(n−1) is embedded in Sp 4mn as g ↦→ diag(I 2m ,g,I 2m ).The<br />
trivial orbit produces the integral on the right-hand side of the identity at the statement<br />
of the lemma. Thus we need to prove that the nontrivial orbit contributes zero. In other<br />
words, we need to prove that the integral<br />
∫ ∫<br />
θ τ (y˜X)ψ V (GL2m )(X)ψ m (y) dy dX (1)<br />
N m (F )\N m (A)<br />
is zero for every choice of data. Here<br />
⎛⎛<br />
⎞⎞<br />
I 2m Y Z<br />
ψ m (y) = ψ m<br />
⎝⎝<br />
I 4m(n−1) Y ∗ ⎠⎠ = ψ(Y 2m,1 ).<br />
I 2m<br />
This is done in a way similar to [GRS9, Lemma 3.3]. Indeed, if not zero, the above<br />
integral induces a local functional that is nonzero on each of the constituents of τ .<br />
However, as in the above reference, one can show that the local unramified constituent<br />
of τ cannot support such a functional. We omit the details.<br />
<br />
For the construction of the lifting defined in Section 2.3, we need to consider residues<br />
of other Eisenstein series which are defined on other classical groups. As we did<br />
above, we need to study their properties. Since the arguments are exactly the same,<br />
we define only the residues and indicate the corresponding unipotent orbit attached to<br />
these representations. We start with the following step.<br />
(1) Let G = Sp 2m(2n+1) .Letɛ denote a cuspidal generic irreducible representation<br />
of the split orthogonal group SO 2m (A). Letτ = τ(ɛ) denote the functorial lift of ɛ to<br />
GL 2m (A), aswasprovedin[CKPS]. We assume that τ is cuspidal. Let µ(ɛ) denote<br />
the lift of ɛ to Sp 2m given by the theta representation. It follows from [GRS2] that<br />
µ(ɛ) is generic. Let Q denote the standard parabolic subgroup of G whose Levi part<br />
is GL 2m ×···×GL 2m × Sp 2m . Here GL 2m occurs n times. Let E τ(ɛ),µ(ɛ) (g, ¯s) denote<br />
the Eisenstein series defined on G and associated with the induced representation<br />
Ind G(A)<br />
Q(A) (τ(ɛ) ⊗ ··· ⊗ τ(ɛ) ⊗ µ(ɛ))δ¯s Q . Write an element in the Levi part of Q as<br />
g = diag(g 1 ,...,g n ,h,gn ∗,...,g∗ 1 ). Then we define δ¯s Q = ∏ n<br />
i=1 |g i| s i δ1/2 Q (g). Asin<br />
the case of the Eisenstein series which we considered in Section 2.2, it is not hard to
460 DAVID GINZBURG<br />
check that up to a product of local intertwining operators, the poles of the Eisenstein<br />
series are determined by<br />
∏n−1<br />
i=1<br />
L S (τ × τ,s i − s i+1 ) L S (τ × µ(ɛ),s n )<br />
L S (τ × τ,s i − s i+1 + 1) L S (τ × µ(ɛ),s n + 1) L τ (¯s),<br />
where L τ (¯s) is a product of partial L-functions that are holomorphic and do not vanish<br />
at the point s i = n − i + 1.Let¯s 0 = (n, n − 1,...,1). It follows from the above that<br />
the Eisenstein series has a pole at that point.<br />
If we denote by ɛ the residue of this Eisenstein series at the point ¯s 0 ,then<br />
O G ( ɛ ) = ((2m) 2n+1 ). Thus, the difference between this Eisenstein and the series that<br />
we studied at the beginning Section 2.2 is the use of the representation µ(ɛ) defined<br />
on Sp 2m (A). This difference is technical in its nature, and the statements proved above<br />
follow easily using similar arguments.<br />
(2) We have a similar situation on the double cover of the symplectic group. Denote<br />
G = ˜Sp 2m(2n+1) .Letɛ denote a generic cuspidal representation of ˜Sp 2m (A), andlet<br />
τ = τ(ɛ) denote the functorial lift to GL 2m from ɛ. By that we mean the following. It is<br />
well known (see, e.g., [GRS2]) that every generic cuspidal representation of ˜Sp 2m (A)<br />
has a functorial lift to a generic cuspidal representation of SO 2m+1 (A) or to a generic<br />
cuspidal representation of SO 2m−1 (A). Using the result of [CKPS], we can thus deduce<br />
that ɛ has a functorial lift to a cuspidal representation of GL 2m (A) or GL 2(m−1) (A).<br />
Henceforth, we assume that the given cuspidal representation ɛ has a functorial lift<br />
to a cuspidal representation τ = τ(ɛ) of GL 2m (A). In this context, we can therefore<br />
view the group Sp 2m (C) as the “L group” of ˜Sp 2m (see also [S]).<br />
We can form the Eisenstein series Ẽ τ(ɛ),ɛ (g, ¯s) as in case (1) and prove similar<br />
results.<br />
(3) Let G denote the split orthogonal group SO 2m(2n+1) .Letɛ denote a generic<br />
irreducible cuspidal representation of SO 2m (A), andletτ = τ(ɛ) denote its lift<br />
to GL 2m . We assume that τ is cuspidal. Let Q denote the standard parabolic subgroup<br />
of G whose Levi part is GL 2m × ··· × GL 2m × SO 2m .LetE τ(ɛ),ɛ (g, ¯s)<br />
denote the Eisenstein series defined on G and associated with the induced representation<br />
Ind G(A)<br />
Q(A) (τ ⊗···⊗τ ⊗ ɛ)δ¯s Q . Write an element in the Levi part of Q as<br />
g = diag(g 1 ,...,g n ,h,gn ∗,...,g∗ 1 ). Then we define δ¯s Q = ∏ n<br />
i=1 |g i| s i δ1/2 Q (g). Asin<br />
cases (1), (2), it follows that up to a product of local intertwining operators, the poles<br />
of the Eisenstein series are determined by<br />
∏n−1<br />
i=1<br />
L S (τ × τ,s i − s i+1 ) L S (τ × ɛ, s n )<br />
L S (τ × τ,s i − s i+1 + 1) L S (τ × ɛ, s n + 1) L τ (¯s),<br />
where L τ (¯s) is a product of partial L-functions that are holomorphic and do not vanish<br />
at the point s i = n − i + 1.Let¯s 0 = (n, n − 1,...,1). It follows from the above that<br />
the Eisenstein series has a pole at that point.
ENDOSCOPIC LIFTING 461<br />
If we denote by ɛ the residue of this Eisenstein series at the point ¯s 0 ,then<br />
O G ( ɛ ) = ((2m) 2n (2m − 1)1). Indeed, recall that unipotent orbits for the orthogonal<br />
groups are parameterized by partitions such that even numbers occur with even<br />
multiplicity. Therefore, the Whittaker coefficients for automorphic representations of<br />
SO 2m (A) are attached to the unipotent orbit ((2m − 1)1). Arguing as in Theorem 1,<br />
the above statement regarding O G ( ɛ ) follows.<br />
(4) Let G denote the split orthogonal group SO (2n+1)(2m+1)+1 .Letɛ denote an<br />
irreducible generic cuspidal representation of Sp 2m (A), andletµ = µ(ɛ) denote its<br />
theta lift to SO 2(m+1) (A). Assume that ɛ lifts to a cuspidal representation τ = τ(ɛ) of<br />
GL 2m+1 (A). LetQ denote the standard parabolic subgroup of G whose Levi part is<br />
GL 2m+1 ×···×GL 2m+1 × SO 2(m+1) .LetE τ(ɛ),µ(ɛ) (g, ¯s) denote the Eisenstein series<br />
defined on G and associated with the induced representation Ind G(A)<br />
Q(A)<br />
(τ ⊗ ··· ⊗τ ⊗<br />
µ)δ¯s<br />
Q . Write an element in the Levi part of Q as g = diag(g 1,...,g n ,h,gn ∗,...,g∗ 1 ).<br />
Then we define δ¯s Q = ∏ n<br />
i=1 |g i| s i δ1/2 Q (g). As in cases (1) – (3), it follows that up<br />
to a product of local intertwining operators, the poles of the Eisenstein series are<br />
determined by<br />
∏n−1<br />
i=1<br />
L S (τ × τ,s i − s i+1 ) L S (τ × µ(ɛ),s n )<br />
L S (τ × τ,s i − s i+1 + 1) L S (τ × µ(ɛ),s n + 1) L τ (¯s),<br />
where L τ (¯s) is a product of partial L-functions that are holomorphic and do not vanish<br />
at the point s i = n − i + 1.Let¯s 0 = (n, n − 1,...,1). It follows from the above that<br />
the Eisenstein series has a pole at that point.<br />
If we denote by ɛ the residue of this Eisenstein series at the point ¯s 0 ,then<br />
O G ( ɛ ) = ((2m + 1) 2n+1 1).<br />
(5) Let G denote the split orthogonal group SO 4m(n+1) .Letτ = τ(ɛ) denote a<br />
cuspidal representation of GL 2m (A) which is a lift from a cuspidal generic irreducible<br />
representation ɛ of SO 2m+1 .LetQ denote the standard parabolic subgroup of G whose<br />
Levi part is GL 2m ×···×GL 2m , where GL 2m occurs n + 1 times. Let E τ(ɛ) (g, ¯s)<br />
denote the Eisenstein series defined on G and associated with the induced representation<br />
Ind G(A)<br />
Q(A) (τ ⊗···⊗τ)δ¯s Q . Here the character δ¯s Q is defined as follows. Write<br />
an element in the Levi part of Q as g = diag(g 1 ,...,g n+1 ,gn+1 ∗ ,...,g∗ 1 ).Thenwe<br />
define δ¯s Q = ∏ n+1<br />
i=1 |g i| s i δ1/2 Q (g). As in the case of the Eisenstein series considered in<br />
Section 2.2, it is not hard to check that up to a product of local intertwining operators,<br />
the poles of the Eisenstein series are determined by<br />
n∏ L S (τ × τ,s i − s i+1 ) L S( τ, ∧2 )<br />
,s n+1<br />
L S (τ × τ,s i − s i+1 + 1) L S( τ, ∧2 ,s n+1 + 1 )L τ (¯s),<br />
i=1
462 DAVID GINZBURG<br />
where L τ (¯s) is a product of partial L-functions that are holomorphic and do not vanish<br />
at the point s i = n − i + 2.Let¯s 0 = (n + 1,n+ 2,...,1). It follows from the above<br />
that the Eisenstein series has a pole at that point. If we denote by ɛ the residue of<br />
this Eisenstein series at the point ¯s 0 ,thenO G ( ɛ ) = ((2m) 2(n+1) ).<br />
These are the five residue representations that we consider. As mentioned above,<br />
one can prove analogous statements to those that we proved in detail in the beginning<br />
of Section 2.2.<br />
2.3. Definition of the lifts<br />
Let π denote an irreducible cuspidal generic representation defined on H (A),whereH<br />
is a split classical group of type B,C,orD. More specifically, in the remainder of this<br />
article, H denotes one of the split algebraic groups Sp 2(n+m) , SO 2(n+m)+1 , SO 2(n+m) ,<br />
SO 2(n+m+1) or the metaplectic group ˜Sp 2(n+m) . We define the following.<br />
Definition 2<br />
We say that π is a weak endoscopic lift from two generic automorphic representations<br />
σ and ɛ, defined on the groups H 1 (A) and H 2 (A),ifπ is the weak functorial lift from<br />
σ and ɛ corresponding to the homomorphism of L-groups as given by one of the<br />
following cases:<br />
(1) if H = Sp 2(n+m) ,H 1 = Sp 2n ,andH 2 = SO 2m , then the homomorphism of<br />
L-groups is given by SO 2n+1 (C) × SO 2m (C) ↦→ SO 2(n+m)+1 (C);<br />
(2) if H = ˜Sp 2(n+m) ,H 1 = ˜Sp 2n ,andH 2 = ˜Sp 2m , then the homomorphism of<br />
L-groups is given by Sp 2n (C) × Sp 2m (C) ↦→ Sp 2(n+m) (C) (see Section 2.2(2));<br />
(3) if H = SO 2(n+m) ,H 1 = SO 2n ,andH 2 = SO 2m , then the homomorphism of<br />
L-groups is given by SO 2n (C) × SO 2m (C) ↦→ SO 2(n+m) (C);<br />
(4) if H = SO 2(n+m+1) ,H 1 = Sp 2n ,andH 2 = Sp 2m , then the homomorphism of<br />
L-groups is given by SO 2n+1 (C) × SO 2m+1 (C) ↦→ SO 2(n+m+1) (C);and<br />
(5) if H = SO 2(n+m)+1 ,H 1 = SO 2n+1 ,andH 2 = SO 2m+1 , then the homomorphism<br />
of L-groups is given by Sp 2n (C) × Sp 2m (C) ↦→ Sp 2(n+m) (C).<br />
Our goal is to construct representations π in the above five cases. In each of the<br />
above cases, we introduce a group M and a representation defined on M(A).<br />
The representation corresponds to one of the five cases introduced at the end of<br />
Section 2.2. We also introduce a unipotent subgroup U of M and a character ψ U<br />
defined on U(F )\U(A). With this data, we construct an integral that is used to define<br />
the lifting.<br />
(a) Let M = Sp 2m(2n+1) or its double cover, where m, n are two natural numbers<br />
greater than or equal to 1. Let ɛ denote the automorphic representation of M(A) as<br />
constructed in Section 2.2(1), (2). Thus, O M ( ɛ ) = ((2m) 2n+1 ). This means that ɛ<br />
has no nonzero Fourier coefficient corresponding to any unipotent orbit of M which
ENDOSCOPIC LIFTING 463<br />
is greater than or not related to ((2m) 2n+1 ).LetO M = ((2m − 1) 2n 1 2(m+n) ). It follows<br />
from [CM, Theorem 6.1.3] that the stabilizer of this orbit is the group Sp 2n ×Sp 2(n+m) .<br />
We associate to O M a Fourier coefficient. This association is described in detail in<br />
[G3]. Let P (O M ) denote the standard parabolic subgroup of M whose Levi part is<br />
× Sp 2(m+2n) . We denote its unipotent radical by U(O M ),orsimplybyU. If<br />
m = 1,wetakeU to be the trivial group. In terms of matrices, we can identify U with<br />
the unipotent subgroup of M given by<br />
GL m−1<br />
2n<br />
U = U(O M ) = U p,q,r<br />
⎧⎛<br />
⎞⎫<br />
I p x 1 ∗<br />
. .. . .. ∗<br />
I p x r ∗<br />
I p y 1 y 2 z<br />
⎪⎨<br />
I<br />
=<br />
q 0 y ∗ ⎪⎬<br />
2<br />
I q y1<br />
∗ , (2)<br />
I p xr<br />
∗ . .. . ..<br />
⎜<br />
⎟<br />
⎝<br />
I p x1<br />
⎪⎩<br />
∗ ⎠<br />
⎪⎭<br />
I p<br />
where r = m − 2,p = 2n, andq = m + 2n. In the preceding display, x i ∈<br />
Mat p×p ,y j ∈ Mat p×q ,andz ∈ Mat 0 p×p ={A ∈ Mat p×p : A t J p = J p A}. Also, the<br />
∗ indicates arbitrary entries such that the above matrix is in M.<br />
To define the character ψ U , we identify the group U/[U,U] with the additive<br />
group<br />
X = Mat 2n×2n ⊕···⊕Mat 2n×2n ⊕ Mat 2n×2(2n+m) ,<br />
where Mat 2n×2n appear m − 2 times.<br />
Write an element x ∈ X as x = (x 1 ,...,x m−2 ,y), where we have x i ∈ Mat 2n×2n<br />
and y ∈ Mat 2n×2(2n+m) . Write y = (ȳ 1 , ȳ 2 , ȳ 3 ), where ȳ 1 , ȳ 3 ∈ Mat 2n×(n+m) and<br />
ȳ 2 ∈ Mat 2n×2n .Givenu ∈ U, write it as u = xu ′ , where x ∈ X and u ′ ∈ [U,U]. We<br />
define<br />
ψ U (u) = ψ U (xu ′ ) = ψ U (x) = ψ ( tr(x 1 +···+x m−2 + ȳ 2 ) ) .<br />
As mentioned above, the stabilizer of ψ U inside the Levi part of P (O M ) is given by<br />
Sp 2n × Sp 2(n+m) . The embedding is given as follows. Given that (g, h) ∈ Sp 2n ×<br />
Sp 2(n+m) , we embed it inside GL m−1<br />
2n × Sp 2(m+2n) as (g, h) ↦→ (g,...,g,(g, h)). We<br />
use the same embedding when M is the double cover of the symplectic group.
464 DAVID GINZBURG<br />
To define the lift that we intend to study, let σ denote an irreducible cuspidal<br />
representation of H 1 (A), where H 1 is as in Definition 2(1), (2). In this case, we let π<br />
denote the automorphic representation defined on H (A) generated by the space of all<br />
functions<br />
∫ ∫<br />
( )<br />
f (h) =<br />
ϕ σ (g)θ ɛ u(g, h) ψU (u) dudg. (3)<br />
H 1 (F )\H 1 (A) U(F )\U(A)<br />
Here ϕ σ is a vector in the space of σ ,andθ ɛ is a vector in the space of ɛ .<br />
(b) Let M = SO 2m(2n+1) , where m, n are two natural numbers such that n ≤ m.<br />
Let ɛ denote the automorphic representation of M(A) which was constructed in<br />
Section 2.2(3). Thus, O M ( ɛ ) = ((2m) 2n (2m − 1)1).LetO M = ((2m − 1) 2n 1 2(m+n) ).<br />
It follows from [CM, Theorem 6.1.3] that the stabilizer of this orbit is the group<br />
SO 2n × SO 2(n+m) .LetP (O M ) denote the standard parabolic subgroup of M whose<br />
Levi part is GL m−1<br />
2n × SO 2(m+2n) . We denote its unipotent radical by U(O M ),orsimply<br />
by U.Ifm = 1,wetakeU to be the trivial group. In term of matrices, we can identify<br />
U with the unipotent subgroup of M givenby(2) sothatr = m − 2,p = 2n, and<br />
q = m + 2n and so that the following conditions are satisfied. First, z ∈ Mat 0 p×p =<br />
{A ∈ Mat p×p : A t J p =−J p A}. Second, the ∗ indicates that the entries are such<br />
that U is a subgroup of SO 2m(2n+1) .<br />
Since we can identify U/[U,U] with the group X as defined in case (a), we<br />
define the character ψ U as in that case. The stabilizer of ψ U is SO 2n × SO 2(n+m) ,and<br />
it is embedded inside M in a way similar to the embedding of the stabilizer in case<br />
(a). The definition of the representation π which we construct is given by the space<br />
generated by the functions<br />
∫ ∫<br />
( )<br />
f (h) =<br />
ϕ σ (g)θ ɛ u(g, h) ψU (u) dudg.<br />
H 1 (F )\H 1 (A) U(F )\U(A)<br />
Here σ is a cuspidal representation of the group H 1 = SO 2n ,andθ ɛ is a vector in the<br />
space of ɛ . The group U and the character ψ U are defined above.<br />
(c) Let M = SO (2m+1)(2n+1)+1 , and assume that m ≥ n. Let ɛ denote the<br />
automorphic representation of M(A), as was constructed in Section 2.2(4). Thus,<br />
O M ( ɛ ) = ((2m + 1) 2n+1 1). LetP (O) denote the standard parabolic subgroup<br />
of M whose Levi part is GL m 2n × SO 2(n+m+1). We denote its unipotent radical by<br />
U = U(O). In term of matrices, we write this unipotent group as in (2), where<br />
r = m − 1,p = 2n, andq = m + n + 1. In this case, x i ∈ Mat 2n×2n ,y 1 ,y 2 ∈<br />
Mat 2n×(m+n+1) ,andz ∈ Mat 0 2n×2n = {A ∈ Mat 2n×2n : A t J 2n = −J 2n A}. We<br />
define ψ U (u) = ψ(tr(x 1 + ··· + x l−1 )). LetH 4n(m+n+1)+1 denote the Heisenberg<br />
group with 4n(m + n + 1) + 1 variables. From the definition of U, it follows<br />
that there is a projection l from U onto the Heisenberg group H 4n(m+n+1)+1
ENDOSCOPIC LIFTING 465<br />
defined as follows. Identify elements of H 4n(m+n+1)+1 with triples (x,y,z ′ ),where<br />
x,y ∈ Mat 2n×(m+n+1) and z ′ ∈ Mat 1×1 .Givenu ∈ U in the above coordinates, the<br />
projection is given by l(u) = (y 1 ,y 2 , tr ′ z). Here, for z = (z i,j ) ∈ Mat 0 2n×2n ,wedefine<br />
tr ′ z = z 1,1 +···+z n,n . The stabilizer of the character ψ U is Sp 2n × SO 2(n+m+1) .It<br />
is embedded inside GL m 2n × SO 2(n+m+1) as (g, h) ↦→ (g, g, . . . , g, h). Letσ denote<br />
a cuspidal representation of H 1 (A) = Sp 2n (A). We define the representation π of<br />
H (A) = SO 2(n+m+1) (A) as the space generated by all functions<br />
∫ ∫<br />
f (h) =<br />
ϕ σ (g)θ ψ ( ) ( )<br />
Sp 4n(m+n+1)<br />
l(u)(g, h) θɛ u(g, h) ψU (u) dudg.<br />
H 1 (F )\H 1 (A) U(F )\U(A)<br />
(4)<br />
Here θ ψ Sp 4n(m+n+1)<br />
is the theta function defined on H 4n(m+n+1)+1 · ˜Sp 4n(m+n+1) (for basic<br />
definitions regarding the theta representation, see [P]). Also, ϕ σ is a vector in the space<br />
of σ ,andθ ɛ is a vector in the space of ɛ . The embedding of Sp 2n ×SO 2(m+n+1) inside<br />
Sp 4n(m+n+1) is given by the tensor product.<br />
(d) To describe the last case in Definition 2, letM = SO 4m(n+1) , and, as before,<br />
assume that m ≥ n. Let ɛ denote the automorphic representation of M(A) which<br />
was defined in Section 2.2(5). Thus O M ( ɛ ) = ((2m) 2(n+1) ).LetO M = ((2m −<br />
1) 2n+1 1 2(m+n)+1 ).LetP (O) denote the parabolic subgroup of M whose Levi part is<br />
GL m−1<br />
2n+1 × SO 2(m+2n+1). We denote by U = U(O) the unipotent radical subgroup of<br />
P (O). We use the matrix description given in formula (2), adopted as in (b) to the<br />
even orthogonal group, with r = m − 2,p = 2n + 1, andq = m + 2n + 1. Thus<br />
U/[U,U] can be identified with the group<br />
X = Mat (2n+1)×(2n+1) ⊕···⊕Mat (2n+1)×(2n+1) ⊕ Mat (2n+1)×2(m+2n+1) ,<br />
where Mat (2n+1)×(2n+1) appears m − 2 times. Write an element x ∈ X as x =<br />
(x 1 ,...,x m−2 ,y), where x i ∈ Mat (2n+1)×(2n+1) and y ∈ Mat (2n+1)×2(2n+m+1) .<br />
Given u ∈ U, write it as u = xu ′ , where x ∈ X and u ′ ∈ [U,U]. Wedefine<br />
ψ U (u) = ψ U (xu ′ ) = ψ U (x) = ψ ( tr(x 1 +···+x m−2 ) + tr ′ y) ) .<br />
Here ψ(tr ′ y) = ψ(y 1,1 +···+y n,n +y n+1,m+2n+1 +y n+1,m+2n+2 +y n+2,2(m+2n+1)−n+1 +<br />
···+y 2n+1,2(m+2n+1) ). One can verify that the stabilizer of the character ψ U is the group<br />
SO 2n+1 × SO 2(n+m)+1 and that its embedding is similar to that of the corresponding<br />
groups in the previous cases.<br />
To define π, we start with a cuspidal representation σ of H 1 (A) = SO 2n+1 (A),<br />
and we use a similar integral representation as defined in formula (3).<br />
3. The cuspidality of the lift<br />
We continue with the notation of Section 2. Let ɛ denote an automorphic representation<br />
of the group M(A), as constructed in Section 2. In this section, we discuss the
466 DAVID GINZBURG<br />
cuspidality of the lift. Since the computations are quite similar in all five cases, we<br />
concentrate only on the first case.<br />
Let M = Sp 2m(2n+1) .Letσ denote an irreducible cuspidal representation of<br />
Sp 2n (A). In this case, the lift is given by the integrals<br />
∫ ∫<br />
( )<br />
f (h) =<br />
ϕ σ (g)θ ɛ u(g, h) ψU (u) dudg,<br />
H 1 (F )\H 1 (A) U(F )\U(A)<br />
where H 1 = Sp 2n and h ∈ H = Sp 2(m+n) . (The group U and the character ψ U were<br />
described in Section 2.)<br />
As often happens in the constructions of liftings using small representations, the<br />
image of the lift is not always cuspidal. Usually, there is an obstruction for the lift to<br />
be cuspidal. To understand when this can happen, assume, for example, that n ≥ m.<br />
Assume that σ itself is an endoscopic lift from two cuspidal representations, from a<br />
cuspidal representation σ ′ on Sp 2(n−m) and from a cuspidal representation ɛ ′ on SO 2m .<br />
For example, if ɛ ′ = ɛ, then it is not expected that the representation π is cuspidal. In<br />
this case, there is an obstruction for the lift to be cuspidal, which is basically expressed<br />
in terms of lifts to groups of smaller rank.<br />
To simplify notation, we prove Theorem 2 only when n ≤ m. Whenn>m,<br />
the formal computations of the constant terms are similar. Since we assumed that the<br />
cuspidal representation ɛ lifts to a cuspidal representation τ(ɛ) on GL 2m , the image<br />
of the lift can fail to be cuspidal only in the case where n = m. Thus, in this case, for<br />
the image to be cuspidal we have to assume that a certain integral is zero. In our case,<br />
the integral we need to assume to be zero is given by integral (7), defined in the proof<br />
of Theorem 2. One can interpret this integral as a lift to a group of a lower rank.<br />
The proof of the cuspidality of the lift requires a manipulation of Fourier expansions<br />
performed on the automorphic functions θ ɛ . Here the function θ ɛ liesinthe<br />
space of the residue representation ɛ . At each step, one has to check that the integrals<br />
converge absolutely. These justifications are now quite standard; the main reference<br />
required is [MW, I.2.10].<br />
THEOREM 2<br />
With the above notation, let π denote the automorphic representation of Sp 2(m+n) (A)<br />
generated by the space of functions f (h) defined above. Assume that n ≤ m. Inthe<br />
case where n = m, assume also that the integral (7) is zero for every choice of data.<br />
Then π is a cuspidal representation.<br />
Proof<br />
For 1 ≤ j ≤ m + n, letV j denote the standard unipotent radical of the maximal<br />
parabolic subgroup of H whose Levi part is GL j × Sp 2(m+n−j) . Thus, we prove that
ENDOSCOPIC LIFTING 467<br />
f V j<br />
(h) is zero for every choice of data. In other words, we need to consider the<br />
integrals<br />
∫ ∫ ∫<br />
( )<br />
ϕ σ (g)θ ɛ u(g, v) ψU (u) dv dudg. (5)<br />
H 1 (F )\H 1 (A) U(F )\U(A) V j (F )\V j (A)<br />
The group V j is embedded inside H as the group of all matrices of the form<br />
⎧⎛<br />
⎞<br />
⎫<br />
⎨ I j X ′ Y ′<br />
⎬<br />
V j = ⎝ I<br />
⎩ 2(m+n−j) X ′ ∗⎠ : X ′ ∈ Mat j×2(m+n−j) ,Y ′ ∈ Mat 0 j×j⎭ .<br />
I j<br />
Let w denote the Weyl element of M defined by<br />
⎛<br />
I j<br />
I k1 w =<br />
⎜ I 2(m+2n−j) ⎟<br />
⎝<br />
I k1<br />
⎠ , k 1 = 2n(m − 1).<br />
I j<br />
Conjugating in integral (5) the argument of the function θ ɛ by w, and using the<br />
left-invariance property of this function by rational points, we obtain<br />
∫<br />
ϕ σ (g)θ ɛ<br />
(<br />
t(Z, Y, R)u(g, 1) w t(X)w ) ψ m,n,j (u) dudY dZ dR dXdg.<br />
⎞<br />
Here<br />
⎛<br />
I j Z Y 1 Y 2 R<br />
⎞<br />
I k1 Y ∗ ⎛<br />
2<br />
I j<br />
I k2 Y ∗ 1<br />
X I k1<br />
t(Z, Y, R) =<br />
I 2n , t(X) =<br />
I k3 I k2 Z ∗<br />
⎜<br />
⎝<br />
I k1<br />
⎜<br />
⎟<br />
⎝<br />
I k1 ⎠<br />
X ∗<br />
I j<br />
⎞<br />
,<br />
⎟<br />
⎠<br />
I j<br />
where k 1 = 2n(m − 1),k 2 = m + n − j,andk 3 = 2(m + 2n − j). Here Y = (Y 1 ,Y 2 ),<br />
and all matrices are such that the above two matrices are in M. In the above integral,<br />
these matrices are integrated over Mat r1 ×r 2<br />
(F )\Mat r1 ×r 2<br />
(A) with the appropriate<br />
values of r i . Next, we integrate u ∈ U 2n,m+2n−j,m−2 (F )\U 2n,m+2n−j,m−2 (A),<br />
where the group U p,q,r wasdefinedin(2). The unipotent group U 2n,m+2n−j,m−2<br />
is a subgroup of Sp 2(2mn+m−j) . We view it as a subgroup of M, by embedding<br />
it as all matrices of the form diag(I j ,u,I j ). For the group U p,q,r , we defined
468 DAVID GINZBURG<br />
immediately following (2) a character of ψ Up,q,r which was denoted there by ψ U .<br />
In the above integral, we write ψ m,n,j for the character ψ U2n,m+2n−j,m−2 . Finally, we have<br />
(g, 1) w = w(g, 1)w −1 = diag(I j ,g,...,g,I m+n−j ,g,I m+n−j ,g,...,g,I j ), where<br />
the dots indicate that g occurs m − 1 times.<br />
Let α(S) = α(S 1 ,S 2 ) denote the unipotent subgroup of M defined by<br />
⎛<br />
α(S 1 ,S 2 ) =<br />
⎜<br />
⎝<br />
I j<br />
⎞<br />
S<br />
I 2n I k4 S ∗<br />
⎟<br />
I 2n<br />
⎠ ,<br />
I j<br />
k 4 = 2m(2n + 1) − 2(2n + j).<br />
Here S = (S 1 0 k2 S 2 0 k1 +k 2 −2n) ∈ Mat j×k4 , where S 1 ∈ Mat j×2n(m−2) ,S 2 ∈ Mat j×2n ,<br />
and 0 p represents a zero matrix of size j ×p. We now expand the above integral along<br />
the group α(S), where S is integrated over points in (A) modulo points in F .Wehave<br />
∫ ∑<br />
∫<br />
(<br />
ϕ σ (g)θ ɛ α(S) t(Z, Y, R)u(g, 1) w t(X)w )<br />
δ<br />
× ψ m,n,j (u)ψ δ (S) dS dudY dZ dR dXdg,<br />
where δ is summed over all characters of the group S(F )\S(A). We can identify this<br />
group of characters with all matrices Mat j×2n(m−1) (F ).Sinceθ ɛ is left-invariant under<br />
rational points, it is left-invariant under all matrices {t(δ) : δ ∈ Mat 2n(m−1)×j }.We<br />
conjugate the matrix t(δ) across α(S) t(Z, Y, R) u(g, 1) w from left to right. It follows<br />
from a matrix multiplication that after we change variables in u, the character ψ δ (S)<br />
is canceled. Thus we obtain, after conjugation, the matrix t(X + δ), and we can then<br />
collapse the summation over δ with the integration over X. We obtain<br />
∫<br />
ϕ σ (g)θ ɛ<br />
(<br />
u1 u(g, 1) w t(X)w ) ψ m,n,j (u) du 1 dudXdg. (6)<br />
Here u 1 ∈ U 1 , which is defined as the unipotent group of M generated by all matrices<br />
of the form<br />
⎛<br />
I j<br />
˜Z Ỹ<br />
u 1 = ⎜<br />
⎝ I k5<br />
˜Z ∗ ⎟<br />
⎠ , k 5 = 2(2mn + m − j),<br />
I j<br />
⎞<br />
where ˜Z = (0 Z 1 ) ∈ Mat j×2(2mn+m−j) with Z 1 ∈ Mat j×2(2mn+m−n−j) and Ỹ is such that<br />
the group U 1 is in M. Also, the matrix t(X) is integrated over X ∈ Mat 2n(m−1)×j (A).<br />
The variable u is integrated as before.
ENDOSCOPIC LIFTING 469<br />
Next, we expand the above integral along the group of all matrices of the form<br />
⎛<br />
β(R) =<br />
⎜<br />
⎝<br />
I j<br />
R<br />
I 2n<br />
I k4<br />
I 2n R ∗<br />
I j<br />
⎞<br />
⎟<br />
⎠ ,<br />
R ∈ Mat j×2n.<br />
First, we consider the contribution to the Fourier expansion from the constant term.<br />
For 1 ≤ j ≤ m + n, letU j (M) denote the unipotent radical of the standard maximal<br />
parabolic subgroup of M whose Levi part is GL j × Sp 2(2mn+m−j) . Since the group<br />
generated by U 1 and by {β(R) :R ∈ Mat j×2n } equals U j (M), it follows that we<br />
obtain the constant term θ U j (M)<br />
ɛ as an inner integration. From the definition of the<br />
representation ɛ and from the fact that j ≤ m + n ≤ 2m, it follows that if n
470 DAVID GINZBURG<br />
expansion. Here the embedding of this group inside Sp 2m(2n+1) (F ) is<br />
(k, g) ↦→ diag(k,g,...,g,I m+n−j ,g,I m+n−j ,g,...,g,k ∗ ).<br />
Since χ is nontrivial, we may assume after a conjugation by a suitable element of<br />
GL j (F ) × Sp 2n (F ) that for R = (R i,j ),wehaveχ(R) = ψ(R 1,1 +···). Assume first<br />
that χ is such that χ(R) = ψ(R 1,1 + R j,2n +···), where the dots indicate that χ is<br />
trivial on all entries R 1,l for l>1 and also trivial on R l,2n with l
ENDOSCOPIC LIFTING 471<br />
As with the group T 0 , we suppress j from the notation. Notice that S 0 is, in fact, a<br />
subgroup of the group of matrices of the form β(R). The function θ ɛ is left-invariant<br />
under rational points. Thus, given a character ν as above, one can find an element<br />
s 0 (ν) ∈ S 0 (F ) such that when we conjugate the above integral by that element from<br />
left to right, we can collapse summation and integration in such a way that we obtain<br />
the integral<br />
∫<br />
θ ɛ<br />
(<br />
t0 β(R)u 1 u ) ψ m,n,j (u)χ(R) dt 0 dR du 1 du<br />
as inner integration to the above integral. Here the variable R is integrated over the<br />
group S 0 (A)B ′ (F )\B ′ (A), where B ′ ={β(R) :R ∈ Mat j×2n }. Thus, to prove that<br />
(10) is zero, it is enough to prove that the above integral is zero.<br />
We proceed by induction. For 1 ≤ l ≤ m − 1, we define the following 2m − 1<br />
families of abelian unipotent subgroups of Sp 2m(2n+1) . Here a = 2(mn + m + n):<br />
T l =<br />
S l =<br />
{ 2nl+j<br />
∑<br />
I +<br />
k=1<br />
{ 2nl+j<br />
∑<br />
I +<br />
k=1<br />
}<br />
r k e ′ 2n(l−1)+j+1,k : r 1 = r 2ns+j+1 = 0; 0 ≤ s ≤ l − 1 ,<br />
}<br />
p k e ′ k,2nl+j+1 : p 1 = p 2ns+j+1 = 0; 0 ≤ s ≤ l − 1 ,<br />
{ a−j<br />
∑<br />
}<br />
T m = I+ r k e ′ 2mn+m−n+1,k : r 1 = r 2ns+j+1 = r 2mn+m−n+1 = 0; 0 ≤ s ≤ l−1 ,<br />
k=1<br />
{ a−j<br />
∑<br />
}<br />
S m = I + p k e ′ k,a−j+1 : p 1 = p 2ns+j+1 = p 2mn+m−n+1 = 0; 0 ≤ s ≤ l − 1 ,<br />
k=1<br />
{ a−j+2nl<br />
∑<br />
T m+l = I + r k e ′ a+2n(l−1)−j+1,k : r 1 = r 2ns+j+1 = 0;<br />
k=1<br />
}<br />
r 2mn+m−n+1 = r a+2nq−j+1 = 0; 0 ≤ s ≤ m − 2; 0 ≤ q ≤ l − 1 ,<br />
{ a−j+2nl<br />
∑<br />
S m+l = I + p k e ′ k,a+2nl−j+1 : p 1 = p 2ns+j+1 = 0;<br />
k=1<br />
}<br />
p 2mn+m−n+1 = p a+2nq−j+1 = 0; 0 ≤ s ≤ m − 2; 0 ≤ q ≤ l − 1 .<br />
All of the above groups depend on the parameter j, which we omit from the notation.
472 DAVID GINZBURG<br />
Notice that S l are all subgroups of the group U 1 U m,n,j defined above. Recall that<br />
in the above integral, we integrate along these two groups. We inductively expend the<br />
above integral along these groups. More precisely, we start by expending the above<br />
integral along T 1 (F )\T 1 (A). Then by arguing as above with the groups T 0 and S 0 ,we<br />
use S 1 to collapse summation and integration. Next, we proceed similarly with the<br />
pair T 2 and S 2 and so on. Performing this process 2m times, we obtain the integral<br />
∫<br />
θ ɛ<br />
(<br />
t0 t 1 ···t 2m−1 β(R)u 1 u ) ψ m,n,j (u)χ(R) dt l dR du 1 du.<br />
Here the variables t l are integrated over T l (F )\T l (A), and the variables β(R)u 1 u are<br />
integrated over<br />
S 0 (A) ···S 2m−1 (A)β(R)(F )U 1 (F )U m,n,j (F )\β(R)(A)U 1 (A)U m,n,j (A).<br />
Let w 0 denote the following Weyl element of Sp 2m(2n+1) .Letw 0 [i, k] denote the (i, k)-<br />
entry of w 0 .Wesetw 0 [1, 2] = w 0 [l,2n(l−2)+j+1] = w 0 [m+1, 2mn+m−n+1] =<br />
w 0 [m + l,2(mn + m + n) + 2n(l − 2) − j + 1] = 1, where 2 ≤ l ≤ m. All other<br />
entries of the first 2m rows are zero. For w 0 to be symplectic, this determines the last<br />
2m rows of w 0 uniquely. At the rest of the rows, we choose the entries of w 0 to be<br />
such that w 0 is a monomial matrix in Sp 2m(2n+1) such that all nonzero entries are 1 or<br />
−1. Clearly, θ ɛ is left-invariant by w 0 . Using that, if we conjugate from left to right<br />
by w 0 , then it is not hard to check that we obtain integral (9) with l = 2m as inner<br />
integration. As explained above, this integral is zero. This proves that the contribution<br />
to (8) of summands that correspond to characters of type A is zero.<br />
The second type of characters are those not of type A and such that the stabilizer<br />
inside GL j (F ) contains the unipotent radical of a parabolic subgroup of GL j .This<br />
can happen in the following situation. If χ is not of type A, then we can choose it<br />
to be as follows. Write R = (R i,k ). Then there exists a number l < j such that<br />
χ(R) = ψ(R l,1 + R l+1,2 +···+R j,j−l−1 ). We refer to such characters as characters<br />
of type B. For these characters, we can further expend their contribution to integral<br />
(8). In this case, we get the sum<br />
∑<br />
∫<br />
(<br />
θ ɛ β ′ (P )β(R)u 1 u ) ψ m,n,j (u)χ(R)µ(P ) dP dR du 1 du. (11)<br />
µ<br />
Here<br />
⎛<br />
β ′ (P ) =<br />
⎜<br />
⎝<br />
I j−l<br />
⎞<br />
P<br />
I l I b ⎟<br />
I l P ∗ ⎠ , b = 2m(2n + 1) − 2j,<br />
I j−l
ENDOSCOPIC LIFTING 473<br />
and µ is summed over all characters of this group. Now, we argue as in the case of<br />
characters χ as above. More precisely, if µ is trivial, we argue as we did right before<br />
equation (8) with the case of the trivial orbit, and we show that it has zero contribution<br />
to (11). If µ is nontrivial, then we can choose it to be of the form µ(P ) = µ(P 1,1 +···),<br />
where the dots indicate that µ is trivial on all entries P 1,k , where k>1. Now, we argue<br />
as we did with characters χ of type A. In other words, we define suitable groups T i<br />
and S i as above and show that we obtain integral (9) with l = m as inner integration.<br />
Thus, we obtain zero contribution in (8) also from characters χ of type B.<br />
The last type of characters that we need to handle in (8) are those not of type A and<br />
for which the stabilizer in GL j does not contain a unipotent radical. This can happen<br />
only if j
474 DAVID GINZBURG<br />
values of m ′ and n ′ . The variable m(Y ) is integrated over L(A), where L is a certain<br />
unipotent group. Finally, the matrix w 0 is a suitable Weyl element.<br />
The key point here is that when we replace the variable g in (14) byug, where<br />
u ∈ U j (Sp 2n )(A), we can conjugate it to the left. Using the integration over the<br />
unipotent group V 2m(2n+1),2mj (M)(A) by changing variables, we obtain the fact that<br />
(14) is left-invariant by the group U j (Sp 2n )(A). As explained above, this implies that<br />
the type-C characters also contribute zero to integral (8).<br />
Combining all of this completes the proof of the cuspidality.<br />
<br />
4. The nonvanishing of the lift<br />
In this section, we prove that the lift constructed in Section 2.3 is nonzero. We do this<br />
by computing the Whittaker coefficient of the lift and showing that it is not zero. In<br />
particular, this proves that the image of the lift contains cuspidal representations that<br />
are generic. As before, we give the details of one case. (The other cases are similar.)<br />
These types of computations are now quite familiar; there are many examples of them<br />
in the literature (see, e.g., [GJ], [GRS4]). Therefore, we indicate the necessary steps<br />
of these computations, in some places sketchily.<br />
We consider the first case introduced in Definition 2. In that case, the lift is given<br />
in terms of integral (3), where H 1 = Sp 2n . The group U and the character ψ U are<br />
described there explicitly. Let V (Sp 2k ) denote the maximal standard unipotent radical<br />
for the group Sp 2k . This group consists of upper unipotent matrices. Let ψ V (Sp2k ),a<br />
denote the Whittaker character defined on the group V (Sp 2k ). In more detail, if<br />
v = (v i,j ) ∈ V (Sp 2k ), then we define ψ V (Sp2k ),a = ψ(v 1,2 +···+v k−1,k + av k,k+1 ),<br />
where a ∈ (F ∗ ) 2 \F ∗ .<br />
Recall from Section 2 that the definition of the representation ɛ depends on a<br />
generic automorphic representation µ(ɛ) of Sp 2m . Thus, there exists an a ∈ F ∗ such<br />
that µ(ɛ) has a nonzero Whittaker-Fourier coefficient ψ V (Sp2m ),a. Using the notation of<br />
(3), our goal in this section is to compute the integral<br />
∫<br />
f (vh)ψ V (Sp2(m+n) ),a(v) dv. (15)<br />
V (Sp 2(m+n) )(F )\V (Sp 2(m+n) )(A)<br />
As with the cuspidality condition studied in Section 3, to prove the nonvanishing<br />
of (15) requires that we perform a certain quantity of Fourier expansions. The<br />
convergence of each of the integrals is justified using [MW, I.2.10].<br />
We start by introducing certain notation. For 1 ≤ i ≤ m 2 ,letMat col,i<br />
m 1 ×m 2<br />
denote<br />
the group of all matrices whose last i columns are zero. Similarly, for 1 ≤ i ≤ m 1 ,let<br />
Mat row,i<br />
m 1 ×m 2<br />
denote the group of all matrices whose last i rows are zero.
ENDOSCOPIC LIFTING 475<br />
by<br />
Let ̂R 1 denote the group of all unipotent matrices inside V(Sp 2m(2n+1) ) defined<br />
⎧⎛<br />
⎪⎨<br />
̂R 1 =<br />
⎜<br />
⎝<br />
⎪⎩<br />
I 2n(m−1)<br />
⎞<br />
R<br />
⎛<br />
I n+m I 2n ⎟<br />
I n+m R ∗ ⎠ ,R= ⎜<br />
⎝<br />
I 2n(m−1)<br />
R m−1<br />
R m−2<br />
.<br />
R 1<br />
⎞<br />
⎟<br />
⎠ : R i ∈ Mat col,i<br />
2n×(n+m)<br />
Also, let S denote the group of all unipotent matrices inside Sp 2m(2n+1) defined by<br />
⎧⎛<br />
⎪⎨<br />
S =<br />
⎜<br />
⎝<br />
⎪⎩<br />
I 2n(m−1)<br />
S 2 I n+m S 1 S 3<br />
I 2n S1<br />
∗<br />
I n+m<br />
S2<br />
∗<br />
⎫<br />
⎪⎬<br />
.<br />
⎞<br />
⎫<br />
⎟<br />
⎠ ,S 2 = ( )<br />
⎪⎬<br />
0 S 2,m−1 ··· S 2,3 S 2,2 .<br />
⎪⎭<br />
I 2n(m−1)<br />
Here S 1 ∈ Mat row,1<br />
(n+m)×2n ,S 2,i ∈ Mat row,i<br />
(n+m)×2n ,andS 3 ∈ Mat 0 (n+m)×(n+m)<br />
are such that the<br />
first column of S 3 is zero. The center of the group S, which we denote by Ŝ, consists of<br />
all matrices in S such that S 1 and S 2 are zero. Thus, Ŝ is an abelian unipotent subgroup<br />
of V (Sp 2m(2n+1) ).<br />
Returning to integral (15) as defined by integral (3), we start by expanding it along<br />
the group Ŝ(A)S(F )\S(A). We obtain<br />
∑<br />
∫<br />
( )<br />
ϕ σ (g)θ ɛ su(g, vh) ψU (u)ψ V (Sp2(m+n) ),a(v)ψ α (s) ds dv dudg.<br />
α<br />
Here g is integrated over Sp 2n (F )\Sp 2n (A), the variable u is integrated over the<br />
group U(F )\U(A) (see (3)), the variable s is integrated over Ŝ(A)S(F )\S(A),<br />
and v is integrated as in (15). The variable α is summed over all characters of<br />
the group Ŝ(A)S(F )\S(A). Notice that ̂R 1 is a subgroup of U. Sinceθ ɛ is leftinvariant<br />
over rational points, conjugating in the above integral by a suitable rational<br />
matrix in ̂R 1 (F ), and after a suitable collapsing of summation and integration, we<br />
obtain<br />
∫<br />
(<br />
ϕ σ (g)θ ɛ su(1,v)̂r1 (g, h) ) ψ U (u)ψ V (Sp2(m+n) ),a(v) ds dudv d̂r 1 dg.<br />
⎪⎭<br />
Here the variable u is integrated over U(F )̂R 1 (A)\U(A), and the variable ̂r 1 is integrated<br />
over ̂R 1 (A). All other variables are integrated as before.
476 DAVID GINZBURG<br />
For 1 ≤ i ≤ m − 1, denote by ν i the Weyl element of Sp 2m(2n+1) :<br />
⎛<br />
ν i =<br />
⎜<br />
⎝<br />
I 2n(m−i−1)<br />
I m+n−i<br />
I 2n<br />
I 2(2ni+i−n)<br />
I 2n<br />
I m+n−i<br />
⎞<br />
.<br />
⎟<br />
⎠<br />
I 2n(m−i−1)<br />
Denote ˜w 1 = ν m−1 ν m−2 ···ν 1 .Since˜w 1 is an element in Sp 2m(2n+1) (F ), the function<br />
θ ɛ (z) is left-invariant under this element. Thus θ ɛ (su(1,v)z) = θ ɛ (u ′˜w 1 z), where<br />
u ′ = ˜w 1 su(1,v)˜w −1<br />
1 . For a natural number p, letU p denote the standard unipotent<br />
radical of the standard parabolic subgroup of Sp 2m(2p+1) whose Levi part is GL 3p+1 ×<br />
GL m−2<br />
2p+1 × Sp 2p+2 . Thus, the above integral is equal to<br />
∫<br />
(<br />
ϕ σ (g)θ ɛ (v1 ,v 2 ) ′ n u n˜w 1̂r 1 (g, h) ) (<br />
ψ n (v1 ,v 2 ) ′ n ,u n)<br />
dvi du n d̂r 1 dg.<br />
Here u n ∈ U n ,and<br />
⎛<br />
(v 1 ,v 2 ) ′ n = ⎜<br />
⎝<br />
⎞<br />
⎛ ⎞<br />
1 x<br />
⎟<br />
⎠ , v 1 ∈ V (GL n+1 ),v 2 = ⎝ I 2n<br />
⎠, x∈ A,<br />
1<br />
v 1<br />
I q<br />
v 2<br />
I q<br />
v ∗ 1<br />
(16)<br />
where q = (m − 2)(2n + 1) + 2n. To describe ψ n , it is convenient, for reasons that<br />
become clear later, to describe ψ p for any natural number p.<br />
To do so, write v 1 = (v 1 (i, j)) ∈ V (GL p+1 ). Then for v 2 as above, the restriction<br />
of ψ p to (v 1 ,v 2 ) ′ p is given by ψ p((v 1 ,v 2 ) ′ p ) = ψ(v 1(1, 2) +···+v 1 (n, n + 1) + x).<br />
Next, identify the group U p modulo its commutator with the group of all matrices<br />
(X 1 ,X 2 ,...,X m−1 ), where X 1 ∈ Mat (3p+1)×(2p+1) ;for2 ≤ i ≤ m − 2, wehave<br />
X i ∈ Mat (2p+1)×(2p+1) and X m−1 ∈ Mat (2p+1)×(2p+2) . Write<br />
X 1 =<br />
( )<br />
X1,1<br />
,<br />
X 1,2<br />
where X 1,1 ∈ Mat p×(2p+1) and X 1,2 ∈ Mat (2p+1)×(2p+1) . Also, write X m−1 =<br />
(X m−1,1 X m−1,2 ), where X m−1,1 ∈ Mat (2p+1)×(2p+1) and X m−1,2 ∈ Mat (2p+1)×1 .Ifwe<br />
identify an element u p ∈ U p /[U p ,U p ] with (X 1 ,X 2 ,...,X m−1 ), then the restriction<br />
of ψ p to U p is given by ψ Up (u p ) = ψ(tr(X 1,2 + X 2 +···+X m−2 + X m−1,1 )).
ENDOSCOPIC LIFTING 477<br />
For 1 ≤ j ≤ n, letL j ={I 2m(2n+1) + l j,1 e j,n+2 ′ + ··· + l j,2ne j,3n+1 ′ },where<br />
e<br />
p,q ′ = e p,q − e 2m(2n+1)−q+1,2m(2n+1)−p+1 and e p,q is the (2m(2n + 1) × 2m(2n + 1))-<br />
matrix with 1 at the (p, q)-position and zero elsewhere. Thus, we can identify the<br />
group L j with Mat 1×2n . Expand the above integral along the group L 1 (F )\L 1 (A).<br />
From the embedding of Sp 2n inside Sp 2m(2n+1) , we deduce that Sp 2n (F ) actsonthe<br />
character group of L 1 (F )\L 1 (A) with two orbits, one trivial and the other nontrivial.<br />
Thus, the expansion of the above integral breaks into a sum of two terms: first, the<br />
one corresponding to the nontrivial orbit and which we denote by I 1 ; then, the one<br />
corresponding to the trivial orbit and which we denote by I 2 .InI 2 , we expand along<br />
the group L 2 (F )\L 2 (A). Once again, this expansion breaks into a sum of two terms<br />
according to the action of Sp 2n (F ). Continuing this process, we obtain that the above<br />
integral can be written as a finite sum of terms corresponding to the above Fourier<br />
expansions. Below, we compute I 1 . Proceeding in a similar way, one obtains the fact<br />
that all other terms contribute zero to the expansion. Indeed, this happens since we<br />
either obtain constant terms that ɛ does not support, or we obtain Fourier coefficients<br />
of θ ɛ corresponding to unipotent orbits greater than or not related to ((2m) 2n+1 ).By<br />
adopting Theorem 1 to this case, these Fourier coefficients are zero.<br />
Thus, the above integral equals I 1 , which is equal to<br />
∫<br />
(<br />
ϕ σ (g)θ ɛ (v1 ,v 2 ) ′ n l 1u n˜w 1̂r 1 (g, h) ) (<br />
ψ n (v1 ,v 2 ) ′ n ,u n)<br />
ψ(l1,1 ) dv i dl 1 du n d̂r 1 dg,<br />
where we identified l 1 ∈ L 1 with (l 1,1 ,...,l 1,2n ). All variables are integrated<br />
as in the previous integral, except the variable g, which is integrated over<br />
Sp 2(n−1) (F )Z n (F )\Sp 2n (A). Here Z n is the standard unipotent radical of the maximal<br />
parabolic subgroup of Sp 2n whose Levi part is GL 1 × Sp 2(n−1) . Indeed, the group<br />
Sp 2(n−1) (F )Z n (F ) is the stabilizer inside Sp 2n (F ) of the nontrivial orbit in the above<br />
expansion.<br />
If n>1, letx β1 (1) = I 2m(2n+1) − e ′ 2,n+2 .Whenn = 1, letx β 1<br />
(1) = I 6m −<br />
∑ m<br />
i=1 e′ 3i−1,3i . The function θ ɛ is left-invariant under x β1 (1). Hence we can conjugate<br />
it from left to right. After a suitable change of variables in l 1 , we obtain<br />
∫<br />
ϕ σ (g)θ ɛ<br />
(<br />
(v1 ,v 2 ) ′ n l 1u n x β1 (1)˜w 1̂r 1 (g, h) )˜ψ n<br />
(<br />
(v1 ,v 2 ) ′ n ,u n,l 1<br />
)<br />
dvi dl 1 du n d̂r 1 dg.<br />
Here ˜ψ n is defined as follows. First, the restriction to v 2 and to u n is defined as<br />
in ψ n .Onl 1 , we define it to be ψ(l 1,1 ), and the restriction to v 1 = (v 1 (i, j)) is<br />
ψ(v 1 (2, 3) +···+v 1 (n, n + 1)).<br />
Recall that for the symplectic group, one can choose representatives of Weyl<br />
elements to consist of permutation matrices having 1’s and −1’s. Let ˜w 2 (n) denote<br />
the following Weyl element in Sp 2m(2n+1) . We write it as a permutation matrix, as<br />
above, and we indicate for which entries it is nonzero (i.e., ±1). First, for the first
478 DAVID GINZBURG<br />
2m rows, we have a nonzero entry at the (1, 1)-position and for 2 ≤ i ≤ 2m at the<br />
(i, (2i − 3)n + i)-positions. Next, in the last 2m rows, we have a nonzero entry at the<br />
(4mn + i, (2i + 1)n + i)-position for all 1 ≤ i ≤ 2m − 1 and a nonzero entry at the<br />
(2m(2n + 1), 2m(2n + 1))-entry. Finally, the rows between the 2m + 1 row and<br />
the 4mn row form a matrix of size 2m(2n − 1) × 2m(2n + 1) given by the matrix<br />
⎛<br />
⎞<br />
0 I n 0<br />
0 I<br />
M<br />
I<br />
M<br />
. .. .<br />
I<br />
⎜<br />
M<br />
⎟<br />
⎝<br />
I 0 ⎠<br />
0 I n 0<br />
Here the zero represents a column of zeros, I is the identity matrix of size 2n − 2,and<br />
M is the (1 × 3)-matrix defined by M = (010). In the above integral, we conjugate<br />
by ŵ 2 (n) and obtain<br />
∫ (( )( )( Z Y X I I2m<br />
ϕ σ (g)θ ɛ I q Y ∗ A I q<br />
Z ∗ B A ∗ I<br />
(v 1 ,v 2 ) ′ n−1 u n−1<br />
)<br />
I 2m<br />
)<br />
× ˜w 2 (n)x β1 (1)˜w 1˜r 1 (g, h)<br />
(<br />
ψ n−1 (v1 ,v 2 ) ′ n−1 ,u n−1)<br />
ψV (GL2m)(Z) d (···). (17)<br />
Here Z ∈ V (GL 2m ), the standard maximal unipotent subgroup of GL 2m .Thevariable<br />
X is integrated over the group Mat 0 2m×2m with the condition that X i,j = 0 if i>j.<br />
The variable B is integrated over the group ̂B, which is defined as the subgroup of<br />
Mat 0 2m×2m with the condition that B i,j = 0 if i>jand also that B i,i+1 = 0 for all<br />
1 ≤ i ≤ 2m − 1.Letq = 2m(2n − 1). The variable A is integrated over the subgroup<br />
 of Mat 2m(2n−1)×2m defined as follows. Write<br />
A =<br />
(<br />
A1<br />
A 2<br />
)<br />
,<br />
where A 1 ,A 2 ∈ Mat m(2n−1)×2m . Then we integrate over all A 1 with the condition that<br />
the first two columns are zero and that the (i, j)-entry is zero for all 3 ≤ j ≤ m + 1<br />
and i>(2j − 3)n − j + 1. The matrix A 2 is integrated over all matrices such that<br />
all m + 1 rows are zero and the (j,2m − i)-entry is zero for all 0 ≤ i ≤ m − 2 and<br />
j ≥ (2i + 3)n − (i + 2). Finally, the variable Y is integrated over the subgroup Ŷ of
ENDOSCOPIC LIFTING 479<br />
Mat 2m×2m(2n−1) , defined as follows. Denote by Mat ′ 1×2m(2n−1)<br />
the row vectors such that<br />
all entries except the (1, 2m(2n − 1))-entry are zero. Recall that the first two columns<br />
of  are zero. Thus, by ignoring the first two columns, we can identify  with a<br />
subgroup  ′ of Mat 2m(2n−1)×(2m−2) . With this notation, the variable Y is an element in<br />
the group<br />
⎧ ⎛ ⎞<br />
⎫<br />
⎨ Y 1<br />
⎬<br />
Ŷ =<br />
⎩ Y = ⎝Y 2<br />
⎠,Y 1 ∈ Mat 1×2m(2n−1) ,Y 2 ∈ J 2m(2n−1) Â ′ J 2m−2 ,Y 3 ∈ Mat ′ 1×2m(2n−1)<br />
⎭ .<br />
Y 3<br />
Finally, the variables (v 1 ,v 2 ) ′ n−1 vary over the group of matrices as defined in (16),<br />
replacing n by n − 1, and the variable u n−1 ∈ U n−1 , a group defined right before (16).<br />
In integral (17), all the variables described so far are integrated over their groups of<br />
definition with points in A modulo points in F . The variables˜r 1 and g are integrated<br />
as before.<br />
Denote by ̂R 2 the subgroup of Sp 2m(2n+1) which consists of all matrices<br />
⎧⎛<br />
⎞<br />
⎫<br />
⎨ I<br />
⎬<br />
̂R 2 = ⎝A<br />
I<br />
⎩<br />
q<br />
⎠,A∈ Â, B ∈ ̂B<br />
B A ∗ ⎭ .<br />
I<br />
We consider the inner integration to integral (17) given by the integral<br />
⎛⎛<br />
⎞ ⎛ ⎞ ⎞<br />
∫ Z Y X I<br />
θ ɛ<br />
⎝⎝<br />
I q Y ∗ ⎠ ⎝A<br />
I q<br />
⎠ x⎠ ψ V (GL2m )(Z) dZ dY dXdAdB. (18)<br />
Z ∗ B A ∗ I<br />
We consider the Fourier expansion of (18) along the abelian unipotent group that<br />
consists of matrices in Sp 2m(2n+1) (A) of the form<br />
k 2 (r) = k 2 (r 1 ,...,r 3n−1 ) = I 2m(2n+1) +<br />
( 3n−2<br />
∑ )<br />
r i e ′ 2,2m+i<br />
+ r 3n−1 e ′ 2,4mn+1 ,<br />
where each r j is integrated over F \A. Thus, (18) is equal to<br />
⎛ ⎛<br />
⎞⎛<br />
⎞ ⎞<br />
∑<br />
∫<br />
Z Y X I<br />
θ ɛ<br />
⎝k 2 (r) ⎝ I q Y ∗ ⎠⎝A<br />
I q<br />
⎠ x⎠<br />
α j ∈F<br />
Z ∗ B A ∗ I<br />
i=1<br />
× ψ V (GL2m )(Z)ψ(α 1 r 1 +···+α 3n−1 r 3n−1 ) d(···). (19)<br />
Let l 3 (α) = l 3 (α 1 ,...,α 3n−1 ) = I 2m(2n+1) − ∑ 3n−2<br />
i=1 α ie ′ 2m+i,3 − α 3n−1e ′ 4mn+1,3 .Then<br />
l 3 (α) ∈ ̂R 2 (F ). Using the left-invariance properties of θ ɛ , we conjugate by l 3 (α) from
480 DAVID GINZBURG<br />
left to right. Changing variables and collapsing summation with integration, integral<br />
(19) is equal to<br />
⎛⎛<br />
⎞ ⎛ ⎞ ⎞<br />
∫ Z Y 1 X 1 I<br />
θ ɛ<br />
⎝⎝<br />
I q Y1<br />
∗ ⎠ ⎝A<br />
I q<br />
⎠ x⎠ ψ V (GL2m )(Z) d(···). (20)<br />
Z ∗ B A ∗ I<br />
Here Y 1 is integrated over the group Ŷ 1 generated by Ŷ and the group<br />
k 2 (r 1 ,...,r 3n−2 , 0), where r i ∈ A and, similarly, X 1 is in the group ̂X 1 generated<br />
by ̂X and the group of matrices k 2 (0,...,0,r 3n−1 ), where r 3n−2 ∈ A. Both variables<br />
are integrated in their groups with points in A modulo points in F . The variables A<br />
and B are not changed, but now we integrate the group l 3 (r 1 ,...,r 3n−1 ) with points<br />
in A, and all other variables in ̂R 2 are integrated with points in A modulo points in F .<br />
Next, we define the following group of unipotent matrices:<br />
{<br />
( 5n−3<br />
∑ )<br />
k 3 (r) = k 3 (r 1 ,...,r 5n−1 ) = I 2m(2n+1) + r i e ′ 3,2m+i<br />
i=1<br />
+ r 5n−2 e ′ 3,4mn+1 + r 5n−1e ′ 3,4mn+2<br />
Consider the Fourier expansion of (20) along the unipotent group {k 3 (r)} with points<br />
in A modulo points in F . Then, using<br />
5n−2<br />
∑<br />
l 4 (α) = l 4 (α 1 ,...,α 5n−1 ) = I 2m(2n+1) − α i e ′ 2m+i,4 −α 5n−2e ′ 4mn+1,4 −α 5n−1e ′ 4mn+2,4 ,<br />
i=1<br />
we obtain, after a suitable collapsing of summation and integration, the integral<br />
⎛⎛<br />
⎞ ⎛ ⎞ ⎞<br />
∫ Z Y 2 X 2 I<br />
θ ɛ<br />
⎝⎝<br />
I q Y2<br />
∗ ⎠ ⎝A<br />
I q<br />
⎠ x⎠ ψ V (GL2m )(Z) d(···). (21)<br />
Z ∗ B A ∗ I<br />
Here Y 2 is integrated over the group Ŷ 2 generated by Ŷ 1 and the group<br />
{k 2 (r 1 ,...,r 5n−3 , 0, 0)}; similarly, X 2 is in the group ̂X 2 generated by ̂X and the<br />
group {k 2 (0,...,0,r 5n−2 ,r 5n−1 )}. As for the variables A and B, we integrate the<br />
groups {l 3 (r 1 ,...,r 3n−1 )} and {l 4 (r 1 ,...,r 5n−1 )} over A; all other variables in ̂R 2 are<br />
integrated with points in A modulo points in F .<br />
We continue this process by induction, showing that (18) is equal to<br />
⎛⎛<br />
⎞ ⎛ ⎞ ⎞<br />
∫ Z Y 2m−2 X 2m−2 I<br />
θ ɛ<br />
⎝⎝<br />
I q Y2m−2<br />
∗ ⎠ ⎝A<br />
I q<br />
⎠ x⎠ ψ V (GL2m )(Z) d(···), (22)<br />
Z ∗ B A ∗ I<br />
}<br />
.
ENDOSCOPIC LIFTING 481<br />
where now we integrate Y 2m−2 over the group of all matrices in Mat 2m×2m(2n−1) with the<br />
condition that Y 2m−2 (2m, i) = 0 for all 1 ≤ i ≤ 2m(2n − 1) − 1. The variable X 2m−2<br />
is integrated over the group Mat 0 2m×2m with the condition that X 2m−2(2m, 1) = 0,<br />
and the variables A and B are integrated over ̂R 2 (A). In this last step, that is, when<br />
we move from the (m − 1)th step to the mth step, we also need to use the smallness<br />
properties of the representation ɛ . Indeed, in this step, we first need to consider the<br />
Fourier coefficient along I 2m(2n+1) + re m+1,4mn+m−1 . The smallness property of the<br />
representation implies that the contribution to the expansion of all the nontrivial terms<br />
is zero. This follows from the fact that we obtain as an inner integration a Fourier<br />
coefficient of θ ɛ which corresponds to the unipotent orbit ((2m + 2)1 4mn−2 ). It thus<br />
follows from Theorem 1, adopted to this case, that these Fourier coefficients are zero.<br />
Applying Lemma 1, adopted to this case, to integral (22), it follows that integral<br />
(17) is equal to<br />
⎛⎛<br />
⎞<br />
⎞<br />
∫<br />
2m<br />
ϕ σ (g)θ N m,ψ ⎝⎝I ɛ (v 1 ,v 2 ) ′ n−1 u n−1<br />
⎠˜r 2˜w 2 (n)x β1 (1)˜w 1˜r 1 (g, h) ⎠<br />
I 2m<br />
× ψ n−1<br />
(<br />
(v1 ,v 2 ) ′ n−1 u n−1)<br />
d(···). (23)<br />
Here N m is the standard unipotent radical of the parabolic subgroup of Sp 2m(2n+1)<br />
whose Levi part is GL 2m<br />
1<br />
× Sp 2m(2n−1) . Also, for x ∈ Sp 2m(2n+1) (A), wehave<br />
∫<br />
θ N m,ψ<br />
ɛ<br />
(x) = θ ɛ (yx)ψ Nm (y) dy,<br />
N m (F )\N m (A)<br />
where, for y = (y i,j ) ∈ N m ,wesetψ Nm (y) = ψ(y 1,2 +···+y 2m−1,2m ). Also, in<br />
integral (23), the variable ˜r 2 is integrated over ̂R 2 (A), and all other variables are<br />
integrated as in integral (17).<br />
At this point, we can continue by induction on n. Indeed, notice that the Fourier<br />
coefficient θ N m,ψ<br />
ɛ<br />
is actually a composition of a constant term and a Whittaker coefficient.<br />
Indeed, the integral over N m can be computed in two steps: first, by computing<br />
the constant term along the unipotent radical of the parabolic subgroup of Sp 2m(2n+1)<br />
whose Levi part is GL 2m × Sp 2m(2n−1) ; then, by computing the Whittaker coefficient<br />
along the group GL 2m . From the definition of the representation ɛ , it follows that<br />
the above constant term, viewed as a representation of GL 2m (A) × Sp 2m(2n−1) (A),<br />
defines a cuspidal representation on the GL 2m -part, and on Sp 2m(2n−1) we obtain<br />
a representation defined on Sp 2m(2n−1) (A) which has properties similar to those of<br />
the residue representation on Sp 2m(2n+1) . Thus, on the group Sp 2m(2n−1) , we obtain a<br />
representation that corresponds to the unipotent orbit ((2m) 2n−1 ). In other words, we<br />
obtain a representation that has no nonzero Fourier coefficients corresponding to any
482 DAVID GINZBURG<br />
unipotent orbit greater than or not related to ((2m) 2n−1 ). Hence, in (23), we can now<br />
repeat the same process that we did above, but this time over the integration over<br />
(v 1 ,v 2 ) ′ n−1 u n−1.<br />
Hence, repeating this process n − 1 times, we obtain the fact that integral (23)is<br />
equal to<br />
∫<br />
ϕ σ (g)θ N,ψ<br />
ɛ<br />
(˜rn+1˜w 2 (1)x βn (1)˜r n˜w 2 (2)x βn−1 (1) ···˜r 2˜w 2 (n)x β1 (1)˜w 1˜r 1 (g, h) ) d(···).<br />
(24)<br />
Here ˜r i is integrated over ̂R i , which is defined similarly to the definition of<br />
̂R 2 . The Weyl elements ˜w 2 (i) are defined similarly to ˜w 2 (n), and we define<br />
x βi (1) = diag(I 2m(i−1) ,x<br />
β ′ i<br />
(1),I 2m(i−1) ). Here, for 1 ≤ i ≤ n − 1, wesetx<br />
β ′ i<br />
(1) =<br />
I 2m(2n−2i+3) − e 2,n−i+3 ′ and x′ β n<br />
(1) = I 6m − ∑ m<br />
i=1 e′ 3i−1,3i . Also, the variable g is<br />
integrated over V (Sp 2n )(F )\Sp 2n (A). Finally, the group N is the standard maximal<br />
unipotent subgroup of Sp 2m(2n+1) ,and<br />
∫<br />
θ N,ψ<br />
ɛ<br />
(x) = θ ɛ (yx)ψ N (y) dy,<br />
N(F )\N(A)<br />
where ψ N is defined as follows. For y = (y i,j ) ∈ N,define<br />
( 2m−1<br />
ψ N (y) = ψ<br />
∑<br />
i=1<br />
2m−1<br />
∑<br />
y i,i+1 +<br />
i=1<br />
2m−1<br />
∑<br />
y 2m+i,2m+i+1 +···+<br />
i=1<br />
y 2m(n−1)+i,2m(n−1)+i+1<br />
+ y 2mn+1,2mn+2 +···+y m(2n+1)−1,m(2n+1) + ay m(2n+1),m(2n+1)+1<br />
).<br />
Next, in integral (24), we factor the group V (Sp 2n ) from the g integration, obtaining<br />
the following basic identity.<br />
THEOREM 3<br />
Let W σ denote the Whittaker coefficient of ϕ σ . With the above notation, the integral<br />
∫<br />
f (vh)ψ V (Sp2(m+n) ),a(v) dv<br />
V (Sp 2(m+n) )(F )\V (Sp 2(m+n) )(A)<br />
is equal to the integral<br />
∫<br />
W σ (g)θ N,ψ<br />
ɛ<br />
(˜rn+1˜w 2 (1)x βn (1)˜r n˜w 2 (2)x βn−1 (1) ···˜r 2˜w 2 (n)x β1 (1)˜w 1˜r 1 (g, h) ) d(···),<br />
where g is integrated over V (Sp 2n )(A)\Sp 2n (A), and all other variables are integrated<br />
as in integral (24).
ENDOSCOPIC LIFTING 483<br />
From this, we deduce the following.<br />
THEOREM 4<br />
Let σ denote a generic irreducible cuspidal representation of the group Sp 2n (A).Then<br />
the lift to Sp 2(m+n) (A) is generic. In particular, the lift is nonzero.<br />
Proof<br />
The proof is quite standard. A similar example for such a process is given, for example,<br />
in [GS, Section 7]. The idea is to use the identity established in Theorem 3. More<br />
precisely, suppose that the integral<br />
∫<br />
f (vh)ψ V (Sp2(m+n) ),a(v) dv<br />
V (Sp 2(m+n) )(F )\V (Sp 2(m+n) )(A)<br />
is zero for every choice of data. The idea is to prove that this implies that W ϕσ (e) θɛ<br />
N,ψ (e)<br />
is zero for every choice of data. However, by our assumption that σ is generic, this<br />
produces a contradiction.<br />
From both our vanishing assumption and from Theorem 3, it follows that the<br />
integral<br />
∫<br />
W σ (g)θ N,ψ<br />
ɛ<br />
(˜rn+1˜w 2 (1)x βn (1)˜r n˜w 2 (2)x βn−1 (1) ···˜r 2˜w 2 (n)x β1 (1)˜w 1˜r 1 (g, h) ) d(···)<br />
is zero for every choice of data. We may assume that h = e. The idea is to follow the<br />
same steps that we performed to derive the identity in the statement of Theorem 3.<br />
So, let φ denote a Schwartz function defined on the unipotent group Ŝ(A). Since the<br />
above integral is zero for every choice of data, we thus obtain that for every choice of<br />
data, the integral<br />
∫<br />
W σ (g)θ N,ψ<br />
ɛ<br />
(˜rn+1˜w 2 (1)x βn (1)˜r n˜w 2 (2)x βn−1 (1)<br />
×···×˜r 2˜w 2 (n)x β1 (1)˜w 1˜r 1 (g, e)S ) φ(S) d(···)<br />
is zero. By conjugating the S variable inside the function θɛ<br />
N,ψ<br />
by changing variables, we obtain that the integral<br />
∫<br />
W σ (g)θ N,ψ<br />
ɛ<br />
(˜rn+1˜w 2 (1)x βn (1)˜r n˜w 2 (2)x βn−1 (1)<br />
×···×˜r 2˜w 2 (n)x β1 (1)˜w 1˜r 1 (g, e) )̂φ(˜r 1 ) d(···)<br />
from left to right and<br />
is zero for every choice of data. Here ̂φ is the Fourier transform of φ. Sinceφ is an<br />
arbitrary Schwartz function, it follows that ̂φ is also arbitrary. Thus, we deduce that
484 DAVID GINZBURG<br />
for every choice of data, the integral<br />
∫<br />
W σ (g)θ N,ψ<br />
ɛ<br />
(˜rn+1˜w 2 (1)x βn (1)˜r n˜w 2 (2)x βn−1 (1) ···˜r 2˜w 2 (n)x β1 (1)˜w 1 (g, e) ) d(···)<br />
is zero. The process is clearly inductive, and we omit the details.<br />
<br />
5. The unramified computations<br />
In this section, we prove that the generic part of the lift introduced in Section 4<br />
is functorial. From Section 3, it follows that the lift, under the assumption stated<br />
in Theorem 2, is cuspidal. Therefore, we can write the lift as a sum of cuspidal<br />
representations. From the computations of Section 4, it follows that at least one<br />
irreducible representation of this summand is generic. In this section, we prove that<br />
this summand is a functorial lift, as predicted by Definition 2.<br />
We use the identity established in Theorem 3; we concentrate on this case. The<br />
other cases are done in a similar way. Our method of proving the correspondence is the<br />
same in [GRS4]. At the end of this section, we suggest a different approach sketched<br />
out in [G2, Section 6]. This other approach has the advantage of not requiring the<br />
existence of a Whittaker coefficient, and it has the potential to work in other cases<br />
as well. In fact, using this method implies that any irreducible summand of the lift is<br />
functorial.<br />
In this section, F denotes a local nonarchimedean field. Given a group G, we<br />
denote by G(F ) the F -points of G; when there is no confusion, we omit F from the<br />
notation.<br />
Let π and σ denote a generic irreducible representation of Sp 2(m+n) and Sp 2n ,<br />
respectively. We denote by ɛ the irreducible constituent of the residue representation<br />
constructed in Section 2. Here τ = τ(ɛ) is a generic irreducible representation of<br />
GL 2m which we assume to be the local lift of a generic cuspidal representation ɛ of<br />
SO 2m . Assume that all representations are unramified, and let W π and W σ denote<br />
the unramified vector of each of these two representations. Also, let θ ɛ denote the<br />
unramified vector in the space of ɛ .<br />
Assume that the above vectors satisfy the corresponding local version of the<br />
identity stated in Theorem 3. By this we mean the following. Clearly, the global<br />
Whittaker coefficient is factorizable. Also, it follows from Proposition 2, adopted to<br />
this case, that at the unramified places the global Fourier coefficient given by θɛ<br />
N,ψ<br />
(here, in the meaning of Theorem 3) induces a local functional that is unique. Choose<br />
such a functional, and if we view the residue representation at the place F as a<br />
quotient of an induced representation, we may realize this functional evaluated at the<br />
unramified vector as a function that we denote by θɛ N,ψ (x), where x ∈ Sp 2m(2n+1) .<br />
This function is fixed under the standard maximal compact subgroup of Sp 2m(2n+1) ;it<br />
is normalized so that its value at the identity is 1. In this notation, we assume that the
ENDOSCOPIC LIFTING 485<br />
following identity holds:<br />
∫<br />
W π (h) = W σ (g)θ N,ψ<br />
ɛ<br />
(˜rn+1˜w 2 (1)x βn (1)˜r n˜w 2 (2)x βn−1 (1)<br />
×···×˜r 2˜w 2 (n)x β1 (1)˜w 1˜r 1 (g, h) ) d(···). (25)<br />
We now define the local lift that corresponds to our case. Assume that π is the<br />
irreducible constituent of Ind Sp 2(m+n)<br />
B(Sp 2(m+n) ) χδ1/2 B(Sp 2(m+n) ), that σ is the irreducible constituent<br />
of Ind Sp 2n<br />
B(Sp 2n ) µδ1/2 B(Sp 2n ),andthatτ = τ(ɛ) is the irreducible constituent of<br />
Ind GL 2m<br />
B(GL 2m ) νδ1/2 B(GL 2m ). Here, for a given group G, we denote by B(G) the standard Borel<br />
subgroup of G. The representations χ,µ,andν are unramified characters of the<br />
corresponding Borel subgroups. Thus, χ is determined by n + m unramified characters<br />
(χ 1 ,...,χ n+m ) of F ∗ , µ by (µ 1 ,...,µ n ),andν by (ν 1 ,...,ν m ,ν −1<br />
1 ,...,ν−1 m ).<br />
Indeed, the character ν is as described since τ(ɛ) is the local lift of ɛ. To say that π<br />
is the local endoscopic lift from σ and ɛ is to say that the sets (χ 1 ,...,χ n+m ) and<br />
(µ 1 ,...,µ n ,ν 1 ,...,ν m ) are the same.<br />
The main result in this section is the following.<br />
THEOREM 5<br />
Let π, σ, and ɛ be as above, and suppose that integral (25) holds for all unramified<br />
data. Then π is the local endoscopic lift of σ and ɛ.<br />
Proof<br />
For a complex variable s, letL(π, s) denote the standard local L-function attached<br />
to π. Thisisa2(n + m) + 1 degree L-function. Similarly, we denote by L(σ, s) and<br />
L(ɛ, s) the standard L-functions of these two representations. The first L-function is<br />
of degree 2n + 1, and the second is of degree 2m. To prove the theorem, we prove<br />
that L(π, s) = L(σ, s)L(ɛ, s).<br />
Let h(t) = diag(t,1,...,1,t −1 ) be a torus element in Sp 2(n+m) . It follows from<br />
[GRS3, Theorem 3.1] that<br />
∫<br />
( )<br />
W π h(t) |t| 2s−(n+m+1/2) L(π, 2s − 1/2)<br />
dt = , (26)<br />
ζ (4s − 1)<br />
F ∗<br />
where ζ (s) denotes the local zeta function. Thus, to prove our result, we need to prove<br />
that the integral<br />
∫<br />
W σ (g)θ N,ψ<br />
ɛ<br />
(˜rn+1˜w 2 (1)x βn (1)˜r n˜w 2 (2)x βn−1 (1) ···˜r 2˜w 2 (n)<br />
× x β1 (1)˜w 1˜r 1 (g, h(t)) ) |t| 2s−(n+m+1/2) d(···)
486 DAVID GINZBURG<br />
is equal to<br />
L(σ, 2s − 1/2)L(ɛ, 2s − 1/2)<br />
.<br />
ζ (4s − 1)<br />
In the above integral, we integrate t over F ∗ . Parameterize the maximal torus of Sp 2n<br />
as a = diag(a 1 ,...,a n ,an<br />
−1,...,a−1<br />
1 ). Performing the Iwasawa decomposition in the<br />
above integral, we obtain<br />
∫<br />
W σ (a)θ N,ψ<br />
ɛ<br />
(˜rn+1˜w 2 (1)x βn (1)˜r n˜w 2 (2)x βn−1 (1) ···˜r 2˜w 2 (n)<br />
Here,<br />
× x β1 (1)˜w 1˜r 1 l ′ (a,t) ) δ B(Sp2n )(a) −1 |t| 2s−(n+m+1/2) d(···).<br />
l ′ (a,t) = diag(a,...,a,t,I n+m−1 ,a,I n+m−1 ,t −1 ,a,...,a),<br />
where the above-defined torus a occurs overall 2m − 1 times. Next, in the above<br />
integral, we conjugate the torus l(a,t) from right to left. We need to keep track of the<br />
factors obtained from the change of variables. Since, eventually, only the variables t<br />
and a 1 are important, the others turn out to be units; we keep track of these two only.<br />
From the definition of the Weyl elements ˜w 2 (j) and the groups ̂R i , it follows that<br />
the t-variable commutes with all ̂R i , except when i = 1. In that case, we obtain a<br />
factor of |t| −2n(m−1) from the change of variables. The variable a 1 contributes a factor<br />
of |a 1 | −4mn(m−1) from the change of variables in the ˜r 2 -variable. Since there is also a<br />
factor of |a 1 | −2n from the δ B(Sp2n )(a) −1 factor, the above integral is equal to<br />
∫<br />
( )<br />
W σ (a)θ N,ψ<br />
ɛ l(a,t)˜rn+1˜w 2 (1)x βn (a n )˜r n˜w 2 (2)x βn−1 (a n−1 ) ···˜r 2˜w 2 (n)x β1 (a 1 )˜w 1˜r 1<br />
Here,<br />
×|a 1 | −4mn(m−1)−2n |t| 2s−(2mn+m−n+1/2) p(a 2 ,...,a n ) d(···).<br />
l(a,t) = diag(t,a 1 I 2m−1 , 1,a 2 I 2m−1 ,...,1,a n I 2m−1 ,<br />
I 2m ,a −1<br />
n<br />
I 2m−1, 1,...,a −1<br />
1 I 2m−1,t −1 ),<br />
and p(a 2 ,...,a n ) is a product of factors of the form |a i | p i<br />
. Notice that for all 1 ≤<br />
i ≤ n, the element x βi (a i ) is in the standard maximal compact subgroup of Sp 2m(2n+1) .<br />
Indeed, from the properties of W σ (a), it follows that its value is nonzero if and only if<br />
|a i /a i+1 |≤1 and |a n |≤1. Hence we may assume that |a i |≤1 for all i. By arguing<br />
in a way similar to the proof of Theorem 4, we deduce that we get zero contribution to<br />
the integral unless all variables˜r i are in the standard maximal compact group. Hence,
ENDOSCOPIC LIFTING 487<br />
using the fact that |a i |≤1, the above integral is equal to<br />
∫<br />
( )<br />
W σ (a)θ N,ψ<br />
ɛ l(a,t) |a1 | −4mn(m−1)−2n |t| 2s−(2mn+m−n+1/2) p(a 2 ,...,a n ) d(···).<br />
From the left-invariant property of θɛ<br />
N,ψ , it follows that we get zero contribution unless<br />
|a i |≥1 for all 2 ≤ i ≤ n. This, with the fact that |a i |≤1, implies that |a i |=1 for<br />
all 2 ≤ i ≤ n. Thus, the above integral is equal to<br />
∫<br />
(<br />
W σ (a)θ N,ψ<br />
ɛ l(a1 ,t) ) |a 1 | −4mn(m−1)−2n |t| 2s−(2mn+m−n+1/2) da 1 dt,<br />
where a = diag(a 1 , 1,...,1,a −1<br />
1 ) and by l(a 1,t) we now mean that we consider all<br />
the matrices l(a,t) as above, with the conditions a i = 1 for all 2 ≤ i ≤ n. From the<br />
fact that θ ɛ is a vector in the residue representation, it follows that<br />
(<br />
θ N,ψ<br />
ɛ l(a1 ,t) ) (<br />
= W ɛ l1 (a 1 ,t) ) |t| (4mn−2n+1)/2 |a 1 | (4mn−2n+1)(2m−1)/2 ,<br />
where l 1 (a 1 ,t) = diag(t,a −1<br />
1 I 2m−1), which is a matrix inside GL 2m . This follows<br />
from the definition of l(a,t) and from the fact that |a i |=1 for all 2 ≤ i ≤ n. Also,<br />
W ɛ denotes the Whittaker function of ɛ. Sinceɛ has a trivial central character, then<br />
W ɛ (l 1 (a 1 ,t)) = W ɛ (l 1 (1,a −1<br />
1 t)). Plugging this into the above equation, and changing<br />
variables t ↦→ ta 1 , we obtain<br />
∫<br />
(<br />
W σ (a)W ɛ l1 (1,t) ) |a 1 | 2s−n−1/2 |t| 2s−m da 1 dt.<br />
This integral factorizes to a product of two integrals. Using (26) and the well-known<br />
local Whittaker integral for the standard L-function for GL 2m (see [JPS]), the above<br />
integral equals<br />
L(σ, 2s − 1/2)L(ɛ, 2s − 1/2)<br />
.<br />
ζ (4s − 1)<br />
Therefore, we have the identity L(π, s) = L(σ, s)L(ɛ, s) for all values of s.Fromthe<br />
definition of the standard local L-function, it follows that the sets (χ 1 ,...,χ n+m ) and<br />
(µ 1 ,...,µ n ,ν 1 ,...,ν m ) are the same. <br />
Another approach to prove the unramified correspondence is the one sketched in [G2,<br />
Section 6]. The idea is that the global integral that defines the lifting (e.g., the integral<br />
(3)), once we know it to be nonzero, induces a global nonzero integral given by<br />
∫ ∫ ∫<br />
( )<br />
f π (h)ϕ σ (g)θ ɛ u(g, h) ψU (u) dudg. (27)<br />
H (F )\H (A) H 1 (F )\H 1 (A) U(F )\U(A)
488 DAVID GINZBURG<br />
Here f π (h) is a vector in the space of π.Letπ ′ denote an irreducible summand of π.<br />
Assume that v is a local finite nonarchimedean place such that at that place, all data<br />
is unramified. As in Proposition 2, adopted to our case, we know that if (θ ɛ ) v is the<br />
local unramified constituent of ɛ at the local finite place v,then(θ ɛ ) v is a quotient of<br />
Ind M̂Q χ. Here ̂Q is a certain parabolic subgroup of H ,andχ is an unramified character<br />
of ̂Q. Thus, integral (27) induces a nonzero element in the space<br />
Hom H ×H1<br />
(<br />
(Ind<br />
M̂Q χ) U,ψ U<br />
,π ′ ⊗ σ ) ,<br />
where (···) U,ψU denotes the twisted Jacquet module with respect to ψ U . Arguing as<br />
in [G2], one can show that if the above Hom space is nonzero, then any two of the<br />
representations π ′ ,σ,andɛ determine the third one uniquely.<br />
6. Liftings and poles of tensor L-functions<br />
6.1. Endoscopic lifting and poles of the standard tensor L-functions<br />
Let π denote an irreducible generic cuspidal representation of H (A),andletɛ denote<br />
an irreducible generic cuspidal representation of H 2 (A). Here and in what follows,<br />
H,H 1 ,andH 2 are as defined in Definition 2(1) – (5). Since ɛ is generic, we can<br />
use the result of [CKPS] to lift it to an automorphic representation τ = τ(ɛ) on<br />
GL k (A), where k is determined by m. We assume that τ(ɛ) is a cuspidal representation<br />
of GL k (A). Denote by L S (π × ɛ, s) the standard tensor L-function for the group<br />
H (A) × H 2 (A). Here S denotes a finite set of places, including the archimedean<br />
ones, such that outside of S, all data is unramified. From the references below, we<br />
know that these L-functions can have at most a simple pole at s = 1. Unfortunately,<br />
we do not know a period condition, defined on H and H 2 only, which characterizes<br />
the pole of these L-functions. We do, however, have a natural candidate for such a<br />
period, in terms of the representations π and τ(ɛ). This follows from the fact that<br />
L S (π × ɛ, s) = L S (π × τ(ɛ),s) and from the fact that we do have a good Rankin-<br />
Selberg theory for L S (π × τ(ɛ),s). We now review the global constructions for the<br />
tensor product L-functions, in each of the five cases. In each case, we also introduce<br />
the global period integral, which is related to the pole at s = 1 of this L-function.<br />
As before, we assume that n ≤ m. This guarantees that the constructions that we<br />
had in the previous sections do indeed give us nonzero cuspidal representations when<br />
n
ENDOSCOPIC LIFTING 489<br />
(1) For general values of m and n, this case was studied in [GRS3]. The case where<br />
m = n was studied in [GP]. In this case, let Ẽ τ (h, s) denote the Eisenstein series defined<br />
on the group ˜Sp 4m (A) associated with the induced representation Ind ˜Sp 4m (A)<br />
˜P m<br />
τδ s (A) P m<br />
.<br />
Here P m is the standard parabolic subgroup of Sp 4m whose Levi part is GL 2m .Let<br />
θ ψ Sp 2(n+m)<br />
denote the theta representation defined on the group ˜Sp 2(n+m) (A). The global<br />
integral is then given by (see [GRS3, page 210])<br />
∫<br />
∫<br />
ϕ π (h)θ ψ (<br />
Sp 2(n+m)<br />
l(u)h<br />
)Ẽτ (uh, s)ψ Um−n (u) dudg.<br />
Sp 2(n+m) (F )\Sp 2(n+m) (A) U m−n (F )\U m−n (A)<br />
(28)<br />
Here U m−n is a unipotent group (denoted as H n · V 2k,k−n−1 in [GRS3]), defined as<br />
follows. Using the notation of Section 2.2, we consider the unipotent orbit of Sp 2(n+m)<br />
defined by O = ((2m − 2n)1 2(m+n) ).Asexplainedin[G3], to this unipotent orbit we<br />
can associate a unipotent group, denoted by U m−n and a character ψ Um−n . These are<br />
the unipotent group and the character that we use in the above integral. The function<br />
l is the projection from the group U m−n onto the Heisenberg group of 2(n + m) + 1<br />
variables. Arguing as in [GRS6], this Eisenstein series can have at most a simple pole<br />
at s 0 = (k + 2)/(2(k + 1)). We denote by Ẽ τ (h) a vector in the residue representation<br />
at that point. Taking the residue at s 0 in (28), we denote by P(π, τ(ɛ)) the family of<br />
integrals<br />
∫<br />
Sp 2(n+m) (F )\Sp 2(n+m) (A) U m−n (F )\U m−n (A)<br />
∫<br />
ϕ π (h)θ ψ Sp 2(n+m)<br />
(<br />
l(u)h<br />
)Ẽτ (uh)ψ Um−n (u) dudg.<br />
(2) This case is similar to case (1). The only difference is that now, π is an<br />
irreducible generic cuspidal representation defined on ˜Sp 2(n+m) (A), and the Eisenstein<br />
series is defined on the symplectic group Sp 4m (A). The corresponding global integral<br />
wasdefinedin[GRS3, page 210]. Denoting by E τ (h) a vector in the corresponding<br />
residue representation, we denote by P(π, τ(ɛ)) the family of integrals<br />
∫<br />
∫<br />
˜ϕ π (h)θ ψ ( )<br />
Sp 2(n+m)<br />
l(u)h Eτ (uh)ψ Um−n (u) dudh.<br />
Sp 2(n+m) (F )\Sp 2(n+m) (A) U m−n (F )\U m−n (A)<br />
(3) This case was considered in [G1] and also in [So]. The case when m = n was<br />
studied in [GP]. We consider the Eisenstein series E τ (h, s) defined on SO 4m+1 (A),<br />
which corresponds to the induced representation Ind SO 4m+1(A)<br />
P m (A)<br />
τδm s . Here P m is the<br />
parabolic subgroup of SO 4m+1 whose Levi part is GL 2m .LetO = ( (2(m − n) +<br />
1)1 2(m+n)) .In[G3], we attached to this unipotent orbit a unipotent group U m−n and a
490 DAVID GINZBURG<br />
character ψ Um−n . With this notation, the global integral is given by<br />
∫ ∫<br />
ϕ π (h)E τ (uh, s)ψ Um−n (u) dudh.<br />
SO 2(m+n) (F )\SO 2(m+n) (A) U(F )\U(A)<br />
If Re(s) > 1/2, this Eisenstein series can have at most a simple pole at a unique point<br />
s 0 . Denoting the residue by E τ (h), we consider the period integral<br />
P ( π, τ(ɛ) ) ∫<br />
∫<br />
=<br />
ϕ π (h)E τ (uh)ψ Um−n (u) dudh.<br />
SO 2(m+n) (F )\SO 2(m+n) (A) U m−n (F )\U m−n (A)<br />
(4) This case is similar to case (3). The only difference is in the rank of the<br />
groups in question. Once again, we can consider a period integral that is obtained by<br />
considering the residue of a global Rankin-Selberg integral. As above, we denote this<br />
period by P(π, τ(ɛ)).<br />
(5) This case was also considered in [G1] andin[So]. When m = n, itwas<br />
studied in [GP]. We consider the Eisenstein series E τ (g, s) defined on SO 4m (A) which<br />
τδm s . Here P m is the parabolic<br />
subgroup of SO 4m whose Levi part is GL 2m .Form ≥ n + 1, letO = ( (2(m − n) −<br />
1)1 2(m+n)+1) .In[G3], we attached to this unipotent orbit a unipotent group U m−n and a<br />
character ψ Um−n .Whenm = n,wesetU m−n to be the trivial group. With this notation,<br />
when m ≥ n + 1, the global integral is given by<br />
corresponds to the induced representation Ind SO 4m(A)<br />
P m (A)<br />
∫<br />
SO 2(n+m)+1 (F )\SO 2(n+m)+1 (A) U m−n (F )\U m−n (A)<br />
∫<br />
ϕ π (h)E τ (uh, s)ψ Um−n (u) dudh.<br />
If Re(s) > 1/2, this Eisenstein series can have at most a simple pole at a unique point<br />
s 0 . Denoting the residue by E τ (g), we consider the period integral<br />
P ( π, τ(ɛ) ) ∫<br />
∫<br />
=<br />
ϕ π (h)E τ (uh)ψ Um−n (u) dudh.<br />
SO 2(n+m)+1 (F )\SO 2(n+m)+1 (A) U m−n (F )\U m−n (A)<br />
When m = n, wedefine<br />
P ( π, τ(ɛ) ) =<br />
∫<br />
SO 2(n+m) (F )\SO 2(n+m) (A)<br />
ϕ π (h)E τ (h) dh.<br />
One of the main goals of this article is to study the following.
ENDOSCOPIC LIFTING 491<br />
CONJECTURE 1<br />
Let π denote an irreducible generic cuspidal representation of the group H (A), and let<br />
ɛ denote an irreducible cuspidal representation of H 2 (A) according to Section 6.1(1) –<br />
(5). Assume that τ = τ(ɛ) is a cuspidal representation. Then the following are<br />
equivalent.<br />
(1) The partial tensor L-function L S (π × ɛ, s) = L S (π × τ(ɛ),s) has a simple<br />
pole at s = 1.HereS is a finite set of places, including the archimedean ones,<br />
such that outside of S, all data is unramified.<br />
(2) There is a choice of data such that the period integral P(π, τ(ɛ)) is not zero<br />
for some choice of data.<br />
(3) There is a generic cuspidal representation σ of H 1 (A) such that π is the weak<br />
endoscopic lift from σ and ɛ.<br />
Two parts of the conjecture are, in fact, a theorem. The implication that (1) implies<br />
(2) follows from the usual Rankin-Selberg theory. Indeed, it follows from the above<br />
references that when we unfold the global integrals, we represent the above tensor<br />
product L-functions. It also follows from the above references that for any finite<br />
place, data can be chosen so that the integral is not zero. From this, it follows that<br />
if the partial L-function L S (π × τ(ɛ),s) has a simple pole at s = 1, the period<br />
integral P(π, τ(ɛ)) is not zero for some choice of data. The implication that (3)<br />
implies (1) follows from the definition of the weak lift. Indeed, if we assume (3), then<br />
L S (π × ɛ, s) = L S (σ × ɛ, s)L S (ɛ × ɛ, s). Since all data are generic, we know from<br />
[CKPS] that all representations have a lift to an automorphic representation of GL.<br />
By the result of [Sh], we know that the tensor product L-function of two automorphic<br />
representations does not vanish at s = 1. From this, it follows that L S (π × ɛ, s) has a<br />
simple pole at s = 1.<br />
We note that the implication that (2) implies (1) in Conjecture 1 should, in<br />
principle, follow also from the Rankin-Selberg integral representations given above.<br />
In this part, we study the implication that (2) implies (3). We use the lifting studied<br />
in the previous sections to prove this implication in one case. The other four cases<br />
stated in Conjecture 1 are different. The main problem is that the representations ɛ<br />
involve a representation µ(ɛ) defined on a classical group. Therefore, another step<br />
is required. This step requires the study of the descent of a certain residue of an<br />
Eisenstein series and to prove that this descent is by itself a residue. At this point, it is<br />
not clear how to prove this.<br />
THEOREM 6<br />
Let H and H i be as in Definition 2(5). In other words, suppose that H = SO 2(n+m)+1 ,<br />
that H 1 = SO 2n+1 , and that H 2 = SO 2m+1 . Then Conjecture 1 holds.
492 DAVID GINZBURG<br />
Proof<br />
The proof of the theorem relies on Fourier expansions and uses the fact that O M ( ɛ ) =<br />
((2m) 2(n+1) ). These Fourier expansions are similar to those in [GRS8, proof of<br />
Lemma 2.4].<br />
Let π denote an irreducible generic representation of H (A). We consider the<br />
cases when m ≥ n + 1. The case when m = n, with the assumption that π is cuspidal,<br />
is similar.<br />
It is given that the period integral<br />
P ( π, τ(ɛ) ) ∫ ∫<br />
=<br />
ϕ π (h)E τ (uh)ψ Um−n (u) dudh<br />
H (F )\H (A) U m−n (F )\U m−n (A)<br />
is nonzero for some choice of data. We need to construct a generic cuspidal representation<br />
σ defined on H 1 (A) such that π is the endoscopic lift from σ and ɛ. Let<br />
σ ′ denote the automorphic representation of H 1 (A) generated by all functions of the<br />
form<br />
∫ ∫<br />
( )<br />
f (g) =<br />
ϕ π (h)θ ɛ u(g, h) ψU (u) dudh.<br />
H (F )\H (A) U(F )\U(A)<br />
Here the function θ ɛ is a vector in the space of the representation ɛ defined<br />
on M(A) = SO 4m(n+1) (A). This representation was constructed at the end of<br />
Section 2.2(5). Similarly, the function ϕ π is a vector in the space of π.<br />
By arguing as in Section 3, we can prove that σ ′ is a cuspidal representation of<br />
H 1 (A). Below, we prove that σ ′ is generic. Assuming that, let σ denote an irreducible<br />
cuspidal generic summand of σ ′ . Then it follows from Section 5 that π is the weak<br />
endoscopic lift from σ and ɛ.<br />
To prove that σ ′ is generic, for this proof only let R denote the standard maximal<br />
unipotent subgroup of H 1 = SO 2n+1 .Letψ R denote the Whittaker character defined<br />
on R(F )\R(A) as follows. If r = (r i,j ),thenψ R (r) = ψ(r 1,2 + ··· + r n,n+1 ).<br />
Assuming that P(π, τ(ɛ)) is nonzero for some choice of data, we prove that the<br />
integral<br />
∫ ∫ ∫<br />
( )<br />
ϕ π (h)θ ɛ u(r, h) ψU (u)ψ R (r) dr dudh (29)<br />
H (F )\H (A) U(F )\U(A) R(F )\R(A)<br />
is nonzero for some choice of data. From this, the theorem follows.<br />
Suppose that (29) is zero for every choice of data. We derive a contradiction. Recall<br />
that we may choose Weyl elements of H to be permutation matrices with zeros and 1’s.<br />
Given a Weyl element w, we denote by w[i, j] its (i, j)-entry. For a Weyl element to be<br />
in M, it is enough that if w[i, j] = 1,thenw[4m(n+1)−i+1, 4m(n+1)−j +1] = 1.
ENDOSCOPIC LIFTING 493<br />
From this, it follows that a Weyl element in M is determined uniquely by the 1’s located<br />
in the first 2m(n+1) rows. Let w denote the Weyl element of M defined as follows. For<br />
all 1 ≤ j ≤ n and for all 1 ≤ i ≤ m,letw[(2j − 2)m + i, (i − 1)(2n + 1) + j] = 1.<br />
Let a = 2mn + 2n + 3m + 1. Forall1 ≤ j ≤ n and all 1 ≤ i ≤ m − j, set<br />
w[(2j − 1)m + i, a + (i − 1)(2n + 1) + j] = 1; andform − j + 1 ≤ i ≤ m, set<br />
w[(2j − 1)m + i, a + (i − 2)(2n + 1) + j + 1] = 1. For1 ≤ i ≤ m − n − 1, set<br />
w[2mn + i, 2n 2 + 2n + 1 + (i − 1)(2n + 1)],andforall2mn + m − n ≤ i ≤<br />
2mn(n + 1), setw[i, i] = 1.<br />
In (29), we conjugate the argument of θ ɛ by the Weyl element w, andwe<br />
obtain<br />
⎛ ⎞⎛<br />
⎞⎛<br />
⎞ ⎞<br />
∫<br />
Z Y X A<br />
2mn<br />
ϕ π (h)θ ɛ ⎝ I 4m Y ∗ ⎠⎝BI 4m<br />
⎠⎝I u ′ h ⎠ w⎠ ˜ψ(Z, Y, B, u ′ ) d(···).<br />
Z ∗ CB ∗ A ∗ I 2mn<br />
(30)<br />
Here u ′ is integrated over U m−n (F )\U m−n (A), andh is integrated over H (F )\H (A).<br />
The character ˜ψ restricted to u ′ is equal to ψ Um−n . As matrices A, Z ∈<br />
Mat 2mn×2mn ,C,X ∈ Mat 0 2mn×2mn ={L ∈ Mat 2mn×2mn : J 2mn L t =−LJ 2mn },Y ∈<br />
Mat 2mn×4m ,andB ∈ Mat 4m×2mn are defined as follows. We start with the matrices<br />
A and Z. Write A = (A i,j ) and Z = (Z i,j ), where A i,j ,Z i,j ∈ Mat 2m×2m .First,<br />
we have A i,j = 0 if ij.Next,for1 ≤ i ≤ n, welet<br />
A i,i = I 2m ,andZ i,i is a matrix in the group of all upper triangular matrices in GL 2m .<br />
The precise definition of A i,j for i>jand Z i,j if j>iis not as important here as the<br />
relation between these two matrices. The relation is as follows. Let Z i,j (l,q) denote<br />
the (l,q)-entry of the matrix Z i,j .ThenifZ i,j (l,q) is a nonzero entry, A j,i (q,l) is<br />
zero, and vice versa. For example, if m = 2, then a possible configuration for Z i,j and<br />
A j,i is<br />
⎛<br />
⎞<br />
⎛<br />
⎞<br />
0 ∗ ∗ ∗<br />
∗ ∗ ∗ ∗<br />
Z i,j = ⎜0 0 0 ∗<br />
⎟<br />
⎝0 0 0 0⎠ , A j,i = ⎜0 ∗ ∗ ∗<br />
⎟<br />
⎝0 ∗ ∗ ∗⎠ ,<br />
0 0 0 0<br />
0 0 ∗ ∗<br />
where ∗ indicates an arbitrary nonzero entry. A similar situation holds with the pair of<br />
matrices B and Y and the pair X and C. Also, we mention that some of the entries in<br />
the variables A, B,orC lie across. By that, we mean that some variables are embedded<br />
inside M as a product of several 1-parameter unipotent subgroups which corresponds<br />
to root vectors. Finally, all variables in the above integral are integrated with points in<br />
A modulo points in F .<br />
At this point, we start with a sequence of Fourier expansions, similarly to [GRS8,<br />
proof of Lemma 2.4]. Using the fact that O M ( ɛ ) = ((2m) 2(n+1) ), in the same way as
494 DAVID GINZBURG<br />
in the above reference, we obtain the fact that integral (30) is equal to<br />
⎛<br />
⎛<br />
⎞<br />
I<br />
⎛<br />
⎞ 2m<br />
∫<br />
Z 1 Y 1 X 1<br />
Z 2 Y 2 X 2<br />
ϕ π (h)θ ɛ ⎝<br />
⎜ I 4mn Y1<br />
∗ ⎠<br />
⎜ I 4m Y ∗ ⎝<br />
Z1<br />
∗ 2 ⎟<br />
⎝<br />
Z2<br />
∗ ⎠<br />
I 2m<br />
⎛<br />
⎞<br />
I 2m<br />
⎛<br />
⎞<br />
A 2 A 1<br />
×<br />
⎜ B 2 I 4m<br />
⎝<br />
⎟ B 1 I 4mn<br />
⎠<br />
⎝ C 2 B2 ∗ A ∗ ⎠<br />
2 C 1 B1 ∗ A ∗ 1<br />
I 2m<br />
⎞<br />
⎛<br />
⎞<br />
2mn × ⎝I<br />
u ′ h ⎠<br />
⎟<br />
˜ψ 1 (Z 1 ,Y 2 ,B 2 ,u ′ ) d(···),<br />
I 2mn<br />
⎠<br />
where the variables X 2 ,Y 2 ,Z 2 and A 2 ,B 2 ,andC 2 are matrix variables inside Sp 4mn ,<br />
defined similarly to the matrix variables X, Y, Z and A, B, C. Also, the character<br />
˜ψ 1 on these variables is the restriction of ˜ψ. These variables are integrated with<br />
points in A modulo points in F . The variable Y 1 is integrated over Mat 2m×4mn and<br />
Z 2 over Mat 0 2m×2m . The variable Z 1 is integrated over all upper triangular matrices<br />
of size 2m. All these three variables are integrated with points in A modulo points<br />
in F . The character ˜ψ 1 restricted to Z 1 is the Whittaker character. In other words, if<br />
Z 1 = (Z 1 (i, j)), then˜ψ 1 (Z 1 ) = ψ(Z 1 (1, 2) +···+Z 1 (2m − 1, 2m)). Finally, the<br />
variables A 1 ,B 1 ,andC 1 are integrated with points in A.<br />
Continuing this process inductively, this time with the corresponding matrices<br />
inside Sp 4mn , the above integral is equal to<br />
⎛ ⎛<br />
⎞ ⎛<br />
⎞ ⎞<br />
∫<br />
A<br />
2mn<br />
ϕ π (h)θ ɛ<br />
⎝v ⎝B<br />
I 4m<br />
⎠ ⎝I<br />
u ′ h ⎠ w⎠ ˜ψ 2 (v, B, u ′ ) d(···).<br />
C B ∗ A ∗ I 2mn<br />
(31)<br />
Here the variables A, B,andC are integrated with points in A and also v ∈ V , where<br />
V is the standard unipotent radical of the standard parabolic subgroup of Sp 4m(n+1) ,<br />
whose Levi part is GL 2mn<br />
1<br />
× Sp 4m . The variable v is integrated over V (F )\V (A).The<br />
character ˜ψ 2 (v) is defined as follows. Write v = (v i,j ).Then˜ψ 2 (v) is defined as<br />
ψ(v 1,2 +···+v 2m−1,2m + v 2m+1,2m+2 +···+v 4m−1,4m<br />
+···+ v 2m(n−1)+1,2m(n−1)+2 +···+v 2mn−1,2mn ).
ENDOSCOPIC LIFTING 495<br />
To summarize, we conclude that σ ′ is not zero if and only if integral (31) is<br />
nonzero for some choice of data. Arguing as in [GS] or[GJ], we first deduce that<br />
integral (31) is nonzero for some choice of data if and only if the integral<br />
⎛ ⎛<br />
⎞ ⎞<br />
∫<br />
2mn<br />
ϕ π (h)θ ɛ<br />
⎝v ⎝vI<br />
u ′ h w⎠ w⎠ ˜ψ 2 (v, u ′ ) d(···) (32)<br />
I 2mn<br />
is nonzero for some choice of data. But from the definition of θ ɛ , arguing as in [GRS6],<br />
integral (32) is nonzero for some choice of data if and only if the integral<br />
∫<br />
∫<br />
ϕ π (h)E τ(ɛ) (uh)ψ Um−n (u) dudh<br />
SO 2(n+m)+1 (F )\SO 2(n+m)+1 (A) U m−n (F )\U m−n (A)<br />
is not zero for some choice of data. However, as stated at the beginning of the proof,<br />
this integral is exactly the definition of P(π, τ(ɛ)) in this case. From this, the theorem<br />
follows.<br />
<br />
Remark. Our construction can be extended inductively to the cases when π is a<br />
lift from k-distinct cuspidal representations of the classical groups corresponding to<br />
suitable L-group homomorphisms. It is also interesting to mention that the number of<br />
representations that occur are related to the poles of a certain L-function, the details<br />
of which we give in the case of the odd orthogonal group.<br />
For 1 ≤ i ≤ k, letɛ i denote a generic cuspidal representation of SO 2ni +1(A).<br />
Assume that n 1 +···+n k = n, and assume that all ɛ i are distinct. Then the Langlands<br />
conjectures predict that there exists a cuspidal generic representation π of SO 2n+1 (A),<br />
which is a lift from the k representations ɛ i . Clearly, our method produces these liftings<br />
inductively. We refer to the number k as the endoscopic number of π.<br />
Let ϖ 2 denote the second fundamental representation of Sp 2n (C), which is the<br />
L-group of SO 2n+1 (A). With the above assumptions, we have the identity<br />
ζ S (s)L S (π, ϖ 2 ,s) = L S (τ 1 ⊗···⊗τ k ,<br />
2∧ ) k∏<br />
,s = L<br />
(τ S i ,<br />
i=1<br />
2∧ ) ∏<br />
,s L S (τ i ⊗τ j ,s).<br />
i
496 DAVID GINZBURG<br />
THEOREM 7<br />
The irreducible cuspidal generic representation π of SO 2n+1 (A) has an endoscopic<br />
number k if and only if the partial L-function L S (π, ϖ 2 ,s) has a pole of order k − 1<br />
at s = 1.<br />
We mention that the endoscopic number does not determine the groups from which<br />
the cuspidal representation π is lifted. For example, if π is defined on SO 9 (A) and<br />
has an endoscopic number 1, it can be a lift either from SO 5 (A) × SO 5 (A) or from<br />
SO 7 (A) × SO 3 (A).<br />
6.2. Liftings and poles of tensor Spin L-functions<br />
In Section 6.1, we related the poles of standard tensor L-functions to the endoscopic<br />
liftings, as described in Definition 2. In this section, we relate poles of other L-<br />
functions to liftings and period integrals related to our global construction. Since we<br />
do not have a good theory of L-functions related to Spin representations, this part is<br />
somewhat more speculative than the previous one.<br />
To motivate the general conjecture, we start by considering some low-rank examples.<br />
Consider the following special case given in Definition 2(4). Let π denote an<br />
irreducible generic cuspidal representation of SO 2m+8 (A).Letɛ denote an irreducible<br />
generic cuspidal representation of Sp 2m (A). In Conjecture 1, we stated a criterion<br />
where there exists an irreducible generic cuspidal representation σ of Sp 6 (A) such<br />
that π is the endoscopic lift from σ and ɛ. Suppose further that σ is a lift from a<br />
generic cuspidal representation ν of the exceptional group G 2 (A). In this section, we<br />
state a conjecture analogous to this situation.<br />
More precisely, let π denote an irreducible generic cuspidal representation of<br />
GSO 2m+8 (A), andletɛ denote an irreducible generic cuspidal representation of<br />
GSp 2m (A). The question we study is: when can we find an irreducible generic cuspidal<br />
representation ν of G 2 (A) such that π is a lift from ν and ɛ corresponding to<br />
the homomorphism of the L-groups G 2 (C) × GSpin 2m+1 (C) ↦→ GSpin 2m+8 (C)?Itis<br />
convenient to summarize this by the following diagram:<br />
GSO 2m+8 (A)<br />
↑<br />
GSp 6 (A) × GSp 2m (A)<br />
↑<br />
G 2 (A) × GSp 2m (A)<br />
GSO 2m+8 (C)<br />
↑<br />
GSpin 7 (C) × GSpin 2m+1 (C)<br />
↑<br />
G 2 (C) × GSpin 2m+1 (C)<br />
The left-hand side of this diagram describes the lifting on the group level, and the<br />
right-hand side of the diagram describes the homomorphism of L-groups which corresponds<br />
to that lifting. In other words, let π denote an irreducible generic cuspidal<br />
representation π of GSO 2m+8 (A). In Section 6.1, we stated a conjecture when π is
ENDOSCOPIC LIFTING 497<br />
an endoscopic lift from cuspidal generic representations of GSp 6 (A) and GSp 2m (A).<br />
Now, we pose another conjecture: when is π actually a lift from G 2 (A) × GSp 2m (A)?<br />
That is, we want to find cuspidal representations of G 2 (A) and GSp 2m (A) so that π is<br />
a lift from these two representations. This lift is the one associated with the L-group<br />
homomorphism given in the right-hand side of the above diagram.<br />
We have an answer to this question in the cases when m = 0 and m = 1. These<br />
two cases were considered in [GH1] and[GH2].<br />
We start with the case when m = 0, which was studied in [GH1]. We first state<br />
the result and then explain the notation.<br />
THEOREM 8([GH1, Theorem 4.3])<br />
Let π denote an irreducible generic cuspidal representation of GSO 8 (A). The following<br />
statements are equivalent.<br />
(1) The partial L-functions L S (π, St, s) and L S (π, Spin 8 ,s) both have a simple<br />
pole at s = 1.<br />
(2) The period integral Q(π, ɛ) is not zero for some choice of data.<br />
(3) There exists a cuspidal generic representation ν of G 2 (A) such that π is the<br />
weak functorial lift from ν.<br />
In this case, the representation ɛ is the identity representation. The L-functions considered<br />
in the first part are the Standard and one of the Spin representations of the<br />
group GSpin 8 (C). Both are of degree 8. The period Q(π, ɛ) is defined as follows. Let<br />
V 2 denote the standard unipotent radical subgroup of the standard maximal parabolic<br />
subgroup of GSO 8 whose Levi part is GL 2 × GSO 4 .LetH 9 denote the Heisenberg<br />
group with nine variables. Then V 2 is isomorphic to H 9 . We denote this isomorphism<br />
by l. Let ψ Sp 8<br />
denote the theta representation of ˜Sp 8 (A). WedefineQ(π, ɛ) to be the<br />
integral<br />
∫<br />
∫<br />
ϕ π (vk)θ ψ ( )<br />
Sp 8<br />
l(v)k dv dk,<br />
SL 2 (F )×SO 4 (F )\SL 2 (A)×SO 4 (A) V 2 (F )\V 2 (A)<br />
where ϕ π is a vector in the space of π and θ ψ Sp 8<br />
is a vector in the space of ψ Sp 8<br />
.<br />
Next, we consider the case where m = 1, which was studied in [GH2]. As above,<br />
we first state the result. We have the following.<br />
THEOREM 9([GH2, Main Theorem])<br />
Let π and ɛ denote irreducible cuspidal generic representations of the groups<br />
GSO 10 (A) and GL 2 (A). The following statements are equivalent.<br />
(1) The partial L-function L S (π ×ɛ, Spin 10 ×Spin 3 ,s) has a simple pole at s = 1.<br />
(2) The period integral Q(π, ɛ) is not zero for some choice of data.<br />
(3) There exists a cuspidal generic representation ν of G 2 (A) such that π is the<br />
weak functorial lift from ν and from ɛ.
498 DAVID GINZBURG<br />
In the above, the L-function is the tensor product L-function of the two Spin representations.<br />
Its degree is 32. In this case, ɛ is a cuspidal representation defined<br />
on GL 2 (A). Hence the L-group is GL 2 (C), andtheSpin 3 -representation is just the<br />
standard representation of that group. In this case, the period integral Q(π, ɛ) is given<br />
by<br />
∫<br />
ϕ π (k)θ SO10 (k)θ ɛ (k) dk,<br />
SO 10 (F )\SO 10 (A)<br />
where θ ɛ is a vector in the space of the representation ɛ constructed in Section 2.2(4),<br />
with m = n = 1.<br />
Motivated by Theorems 8 and 9, we state the conjecture for general values of m.<br />
CONJECTURE 2<br />
Let π and ɛ denote irreducible cuspidal generic representations of the groups<br />
GSO 2(m+4) (A) and GSp 2m (A). The following statements are equivalent.<br />
(1) The partial L-function L S (π × ɛ, Spin 2(m+4) × Spin 2m+1 ,s) has a simple pole<br />
at s = 1.<br />
(2) The period integral Q(π, ɛ), defined below, is not zero for some choice of data.<br />
(3) There exists a cuspidal generic representation ν of G 2 (A) such that π is the<br />
weak functorial lift from ν and from ɛ.<br />
Here the L-function is the tensor Spin L-function whose degree is 2 2m+3 .Sincewe<br />
have no theory for these L-functions (except the case when m = 1), it is hard to say<br />
much about the relations between the first part and the others. However, we can present<br />
our reasoning for why we expect that (3) implies (1) in Conjecture 2. To do that, let<br />
η m+4 denote the (m + 4)th fundamental representation of the group Spin 2(m+4) (C).<br />
This is the Spin representation of that group. For 1 ≤ i ≤ m, letϖ i denote the ith<br />
fundamental representation of Spin 2m+1 (C), andfor1 ≤ j ≤ 3, letµ j denote the<br />
jth fundamental representation of Spin 7 (C). Given two representation ω 1 and ω 2 of<br />
the complex groups K 1 (C) and K 2 (C), respectively, we denote by (ω 1 |ω 2 ) K1 ×K 2<br />
the<br />
corresponding representation of K 1 (C) × K 2 (C). We omit the reference to K 1 and K 2<br />
since the group to which we are referring is clear from the context.<br />
To motivate the relation with the L-function in Conjecture 2, we show that<br />
when we restrict the representation η m+4 to the group G 2 (C) × Spin 2m+1 (C),<br />
then the representation (η m+4 |ϖ m ) Spin2(m+4) ×Spin 2m+1<br />
↓ G2 ×Spin 2m+1<br />
contains the identity<br />
representation. Indeed, it follows from [K] that branching down, we obtain the fact<br />
that η m+4 ↓ Spin7 ×Spin 2m+1<br />
= (µ 3 |ϖ m ) Spin7 ×Spin 2m+1<br />
. We have the identity<br />
ϖ m ⊗ ϖ m = 1 ⊕ 2ϖ m ⊕ m−1<br />
i=1 ϖ i,
ENDOSCOPIC LIFTING 499<br />
from which it follows that<br />
(η m+4 |ϖ m ) ↓ Spin7 ×Spin 2m+1<br />
= (µ 3 |0) ⊕ (µ 3 |2ϖ m ) ⊕ m−1<br />
i=1 (µ 3|ϖ i ). (33)<br />
The representations on the right-hand side of this equality are representations of the<br />
group Spin 7 (C)×Spin 2m+1 (C).Let(01) denote the second fundamental representation<br />
of G 2 (C). Its degree is 7. Then µ 3 restricted to G 2 (C) gives us 1 ⊕ (01). From this,<br />
we obtain the fact that<br />
(η m+4 |ϖ m ) ↓ G2 ×Spin 2m+1<br />
= (0|0) ⊕ (01|0) ⊕ (0|2ϖ m ) ⊕ (01|2ϖ m )<br />
⊕ m−1<br />
i=1 [(0|ϖ i) ⊕ (01|ϖ i )],<br />
where the representations of the right-hand side are representations of the group<br />
G 2 (C) × Spin 2m+1 (C).<br />
Assume that Conjecture 2(3) holds. Then the above branching decomposition<br />
induces a factorization of the partial L-function L S (π × ɛ, Spin 2(m+4) × Spin 2m+1 ,s),<br />
which contains at least one zeta factor. One expects that for generic representations,<br />
the other partial L-function must be nonzero at s = 1. Moreover, if the representations<br />
ν and ɛ are in general position, one would expect that these L-functions also must be<br />
holomorphic at s = 1. This would then imply that at s = 1, the pole of the above<br />
L-function would actually be a simple pole. This then motivates the implication that<br />
(3) implies (1) in Conjecture 2.<br />
Next, we introduce the period integral Q(π, ɛ). To simplify things, we introduce<br />
the period not in the similitude groups but on the orthogonal and symplectic groups<br />
themselves. The adoption to the similitude groups is quite simple but requires some<br />
more notation. Let P denote the standard parabolic subgroup of M ′ = SO 2(3m+2)<br />
whose Levi part is GL m−1<br />
2 × SO 2(m+4) ,andletU ′ denote its unipotent radical. In terms<br />
of matrices, this unipotent group is described in (2) with r = m − 2,p = 1, and<br />
q = m + 4. As explained in Section 2.2(c), one can define a projection l from the<br />
group U ′ onto the Heisenberg group H 4(m+4)+1 .LetK = SO 2(m+4) × SL 2 .Wedefine<br />
Q(π, ɛ) to be the period integral<br />
∫ ∫<br />
ϕ π (k 1 )θ ψ (<br />
Sp l(u)(k1<br />
4(m+4)<br />
,k 2 ) ) θ ɛ( ′ u(k1 ,k 2 ) ) ψ U ′(u) dudk 1 dk 2 . (34)<br />
K(F )\K(A) U ′ (F )\U ′ (A)<br />
Here k 1 ∈ SO 2(m+4) ,andk 2 ∈ SL 2 . The function θ<br />
ɛ ′ is a vector in the space of the<br />
representation ′ ɛ defined on the group M′ (A) as follows. Recall the representation<br />
ɛ defined on the group M = SO 14m+8 in Section 2.2(4), with n = 3. This representation<br />
is a residue representation that was attached to the induced representation<br />
from the parabolic subgroup Q whose Levi part is GL 3 2m+1 × SO 2(m+1).Todefinethe<br />
representation ′ ɛ , we start with the parabolic subgroup Q′ of M ′ whose Levi part
500 DAVID GINZBURG<br />
is GL 2m+1 × SO 2(m+1) . Then we attach to the corresponding induced representation<br />
an Eisenstein series whose residue we denote by ′ ɛ<br />
. All other notation is as in<br />
Section 2.3(c). Arguing as in [GRS1, pages 114 – 115], by choosing suitable Schwartz<br />
functions in the representation θ ψ Sp 4(m+4)<br />
, the integral (34) converges absolutely. We now<br />
prove the following theorem.<br />
THEOREM 10<br />
Conjecture 2(2) implies Conjecture 2(3).<br />
Proof<br />
The case when m = 1 was shown in [GH2] in complete detail. For general values<br />
of m, the proof is similar and follows similar steps as in the proof of Theorem 6.<br />
Therefore, we sketch only the main steps.<br />
The key ingredient is to use the construction of the lifting from the group Sp 6 ×<br />
Sp 2m to the group SO 2(m+4) . This lifting was introduced in Section 2.3(c), where<br />
we now take n = 3. LetM = SO 14m+8 ,andletH = SO 2(m+4) . Starting with the<br />
representations π and ɛ as above, we construct an automorphic representation σ ′ ,<br />
defined on the group Sp 6 (A), as the space generated by all functions<br />
∫ ∫<br />
f (g) =<br />
ϕ π (h)θ ψ ( ) ( )<br />
Sp 12(m+4)<br />
l(u)(g, h) θτ(ɛ) u(g, h) ψU (u) dudh.<br />
H (F )\H (A) U(F )\U(A)<br />
As in Section 3, we can prove that σ ′ is a cuspidal representation of the group Sp 6 (A).<br />
In fact, the computations are very similar to those in [GH2, Lemma 8], where the case<br />
m = 1 was studied in detail. If nonzero, then it follows from Section 5 that if σ is<br />
any nonzero irreducible summand of σ ′ , π is an endoscopic lift from σ and ɛ. Hence<br />
we need to know when σ ′ is nonzero and when it is a lift from the exceptional group<br />
G 2 (A). For that, we use the result from [GJ], described as follows. Let σ denote an<br />
irreducible cuspidal representation of Sp 6 (A). Suppose that for some vector ϕ σ in the<br />
space of σ , the period integral<br />
⎛⎛<br />
⎞ ⎛ ⎞⎞<br />
∫ ∫ I 1 X Y k 1<br />
ϕ ψ σ (g) =<br />
ϕ σ<br />
⎝⎝<br />
I 2 X ∗ ⎠ ⎝ k 1<br />
⎠⎠ ψ(tr X) dXdY dk 1<br />
SL 2 (F )\SL 2 (A) (F \A) I 7 2 k 1<br />
is nonzero. Here X ∈ Mat 2×2 ,andY ∈ Mat 0 2×2<br />
. Then it follows from [GJ] that there<br />
exists a cuspidal generic representation ν of the exceptional group G 2 (A) such that σ<br />
is the weak functorial lift from ν.<br />
Performing steps similar to those in the proof of Theorem 6, we deduce that the<br />
above integral is nonzero if and only if Q(π, ɛ) is nonzero.
ENDOSCOPIC LIFTING 501<br />
To motivate a possible generalization of Conjecture 2, we first extend the branching rule<br />
(33). Indeed, let η m+l denote the (m+l)th fundamental representation of Spin 2(m+l) (C),<br />
and for 1 ≤ i ≤ m,letϖ i denote the ith fundamental representation of Spin 2m+1 (C).<br />
For 1 ≤ j ≤ l − 1, letµ j denote the jth fundamental representation of Spin 2l−1 (C).<br />
Generalizing the branching rule (33) by using, for example, the method of [K], we<br />
obtain<br />
(η m+l |ϖ m ) ↓ Spin2l−1 ×Spin 2m+1<br />
= (µ l−1 |0) ⊕ (µ l−1 |2ϖ m ) ⊕ m−1<br />
i=1 (µ l−1|ϖ i ). (35)<br />
Suppose that Spin 2l−1 has a subgroup such that when restricting the Spin representation<br />
to this group, we get a fixed vector. Then the representation (η m+l |ϖ m ) ↓ Spin2l−1 ×Spin 2m+1<br />
has a fixed vector. Motivated by that, we state the following.<br />
CONJECTURE 3<br />
Let π and ɛ denote irreducible cuspidal generic representations of the groups<br />
GSO 2(m+l) (A) and GSp 2m (A). The following statements are equivalent.<br />
(1) The partial L-function L S (π × ɛ, Spin 2(m+l) × Spin 2m+1 ,s) has a simple pole<br />
at s = 1.<br />
(2) The period integral Q ′ (π, ɛ) is not zero for some choice of data.<br />
(3) There exists a cuspidal generic representation ν of GSp 2(l−1) (A) such that π is<br />
the weak functorial lift from ν and from ɛ.<br />
References<br />
[A]<br />
J. ARTHUR, “Automorphic representations and number theory” in 1980 Seminar<br />
on Harmonic Analysis (Montréal, 1980),CMSConf.Proc.1, Amer. Math.<br />
Soc., Providence, 1981, 3 – 54. MR 0670091 447<br />
[B]<br />
A. BOREL, “Automorphic L-functions” in Automorphic Forms, Representations<br />
and L-functions, Part 2 (Corvallis, Ore., 1977), Proc. Sympos. Pure Math.<br />
33, Amer. Math. Soc., Providence, 1979, 27 – 61. MR 0546608 447<br />
[BFG] D. BUMP, S. FRIEDBERG,andD. GINZBURG, On the cubic Shimura lift for PGL 3 ,<br />
Israel J. Math. 126 (2001), 289 – 307. MR 1882041 450<br />
[CKPS] J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO [PIATETSKII-SHAPIRO],and<br />
F. SHAHIDI, Functoriality for the classical groups, Publ. Math. Inst. Hautes<br />
Études Sci. 99 (2004), 163 – 233. MR 2075885 448, 449, 459, 460, 488,<br />
491<br />
[CM] D. H. COLLINGWOOD and W. M. MCGOVERN, Nilpotent Orbits in Semisimple Lie<br />
Algebras, Van Nostrand Reinhold Math. Ser., Van Nostrand Reinhold, New<br />
York, 1993. MR 1251060 452, 454, 463, 464<br />
[GS] W. T. GAN and G. SAVIN, Real and global lifts from PGL 3 to G 2 ,Int.Math.Res.<br />
Not. 2003, no. 50, 2699 – 2724. MR 2017248 483, 495
502 DAVID GINZBURG<br />
[GP] S. GELBART and I. PIATETSKI-SHAPIRO [PIATETSKII-SHAPIRO],“L-functions for<br />
G × GL(n)” in Explicit Constructions of Automorphic L-Functions, Lecture<br />
Notes in Math. 1254, Springer, Berlin, 1987, 53 – 136. MR 0892097 489,<br />
490<br />
[G1] D. GINZBURG, L-functions for SO n<br />
× GL k<br />
, J. Reine Angew. Math. 405 (1990),<br />
156 – 180. MR 1041001 489, 490<br />
[G2] ———, A construction of CAP representations in classical groups,Int.Math.<br />
Res. Not. 2003, no. 20, 1123 – 1140. MR 1963483 448, 451, 452, 455,<br />
484, 487, 488<br />
[G3] ———, Certain conjectures relating unipotent orbits to automorphic<br />
representations, Israel J. Math. 151 (2006), 323 – 356. MR 2214128 452,<br />
453, 457, 463, 489, 490<br />
[GH1] D. GINZBURG and J. HUNDLEY, Multivariable Rankin-Selberg integrals for<br />
orthogonal groups,Int.Math.Res.Not.2004, no. 58, 3097 – 3119.<br />
MR 2098700 497<br />
[GH2] ———, On spin L-functions for GSO 10 , J. Reine Angew. Math. 603 (2007),<br />
183 – 213. MR 2312558 450, 497, 500<br />
[GJ]<br />
D. GINZBURG and D. JIANG, Periods and liftings: From G 2 to C 3 , Israel J. Math.<br />
123 (2001), 29 – 59. MR 1835288 474, 495, 500<br />
[GRS1] D. GINZBURG, S. RALLIS,andD. SOUDRY, On the automorphic theta<br />
representation for simply laced groups, Israel J. Math. 100 (1997), 61 – 116.<br />
MR 1469105 500<br />
[GRS2] ———, Periods, poles of L-functions and symplectic-orthogonal theta lifts,<br />
J. Reine Angew. Math. 487 (1997), 85 – 114. MR 1454260 459, 460<br />
[GRS3] ———, L-functions for symplectic groups, Bull. Soc. Math. France 126 (1998),<br />
181 – 244. MR 1675971 485, 489<br />
[GRS4] ———, Lifting cusp forms on GL 2n to ˜Sp 2n : The unramified correspondence,<br />
Duke Math. J. 100 (1999), 243 – 266. MR 1722953 474, 484<br />
[GRS5] ———, On a correspondence between cuspidal representations of GL 2n and<br />
˜Sp 2n ,J.Amer.Math.Soc.12 (1999), 849 – 907. MR 1671452 454, 455<br />
[GRS6] ———, On explicit lifts of cusp forms from GL m to classical groups, Ann. of<br />
Math. (2) 150 (1999), 807 – 866. MR 1740991 489, 495<br />
[GRS7] ———, Endoscopic representations of ˜Sp 2n<br />
,J.Inst.Math.Jussieu1 (2002),<br />
77 – 123. MR 1954940 447, 451, 452<br />
[GRS8] ———, On Fourier coefficients of automorphic forms of symplectic groups,<br />
Manuscripta Math. 111 (2003), 1 – 16. MR 1981592 456, 457, 492, 493<br />
[GRS9] ———, “Construction of CAP representations for symplectic groups using the<br />
descent method” in Automorphic Representations, L-Functions and<br />
Applications: Progress and Prospects (Columbus, Ohio, 2003), Ohio State<br />
Univ. Math. Res. Inst. Pub. 11, de Gruyter, Berlin, 2005, 193 – 224.<br />
MR 2192824 451, 453, 459<br />
[J]<br />
H. JACQUET, “On the residual spectrum of GL(n)” in Lie Group Representations,<br />
II (College Park, Md., 1982/1983), Lecture Notes in Math. 1041, Springer,<br />
Berlin, 1984, 185 – 208. MR 0748508
ENDOSCOPIC LIFTING 503<br />
[JPS] H. JACQUET, I. I. PIATETSKII-SHAPIRO [PIATETSKI-SHAPIRO],andJ. A.<br />
SHALIKA, Rankin-Selberg convolutions, Amer. J. Math. 105 (1983),<br />
367 – 464. MR 0701565 487<br />
[Ji]<br />
D. JIANG, On the fundamental automorphic L-functions of SO(2n + 1), Int.<br />
Math.Res.Not.2006, no. 64069. MR 2211151 495<br />
[K]<br />
R. C. KING, Branching rules for classical Lie groups using tensor and spinor<br />
methods,J.Phys.A8 (1975), 429 – 449. MR 0411400 498, 501<br />
[L1]<br />
R. P. LANGLANDS, “Problems in the theory of automorphic forms” in Lectures in<br />
Modern Analysis and Applications, III, Lecture Notes in Math. 170,<br />
Springer, Berlin, 1970, 18 – 86. MR 0302614 447<br />
[L2] ———, On the Functional Equations Satisfied by Eisenstein Series, Lecture<br />
Notes in Math. 544, Springer, Berlin, 1976. MR 0579181 451<br />
[MW] C. MOEGLIN and J.-L. WALDSPURGER, Décomposition spectrale et séries<br />
d’Eisenstein: Une paraphrase de l’Écriture, Progr. Math. 113, Birkhaüser,<br />
Basel, 1994. MR 1261867 451, 466, 474<br />
[P]<br />
P. PERRIN,“Représentations de Schrödinger, indice de Maslov et groupe<br />
metaplectique” in Noncommutative Harmonic Analysis and Lie Groups<br />
(Marseille, 1980), Lecture Notes in Math. 880, Springer, Berlin, 1981,<br />
370 – 407. MR 0644841 465<br />
[S] G. SAVIN, Local Shimura correspondence, Math. Ann. 280 (1988), 185 – 190.<br />
MR 0929534 460<br />
[Sh] F. SHAHIDI, On certain L-functions, Amer. J. Math. 103 (1981), 297 – 355.<br />
MR 0610479 491<br />
[So]<br />
D. SOUDRY, Rankin-Selberg convolutions for SO 2l+1 × GL n : Local theory, Mem.<br />
Amer. Math. Soc. 105 (1993), no. 500. MR 1169228 489, 490<br />
[Sp]<br />
T. A. SPRINGER, A construction of representations of Weyl groups, Invent. Math.<br />
44 (1978), 279 – 293. MR 0491988 452<br />
School of Mathematical Sciences, Sackler Faculty of Exact Sciences, Tel-Aviv University,<br />
Ramat-Aviv 69978, Israel; ginzburg@post.tau.ac.il
A CHARACTERIZATION OF SUBSPACES AND<br />
QUOTIENTS OF REFLEXIVE BANACH SPACES<br />
WITH UNCONDITIONAL BASES<br />
W. B. JOHNSON and BENTUO ZHENG<br />
Abstract<br />
We prove that the dual or any quotient of a separable reflexive Banach space with the<br />
unconditional tree property (UTP) has the UTP. This is used to prove that a separable<br />
reflexive Banach space with the UTP embeds into a reflexive Banach space with an<br />
unconditional basis. This solves several longstanding open problems. In particular,<br />
it yields that a quotient of a reflexive Banach space with an unconditional finitedimensional<br />
decomposition (UFDD) embeds into a reflexive Banach space with an<br />
unconditional basis.<br />
1. Introduction<br />
It has long been known that Banach spaces with unconditional bases as well as their<br />
subspaces are much better behaved than general Banach spaces and that many of the<br />
reflexive spaces (including L p (0, 1), 1
506 JOHNSON and ZHENG<br />
There is, of course, quite a lot known concerning problems (a) and (b). For<br />
example, Pełczyński and Wojtaszczyk [15, Theorem 1.1] proved that if X has an<br />
unconditional expansion of identity (i.e., a sequence (T n ) of finite-rank operators such<br />
that ∑ T n converges unconditionally in the strong operator topology to the identity<br />
on X), then X is isomorphic to a complemented subspace of a space that has an<br />
unconditional finite-dimensional decomposition (UFDD). Later, Lindenstrauss and<br />
Tzafriri [11, Theorem 1.g.5] showed that every space with a UFDD embeds (not<br />
necessarily complementably) into a space with an unconditional basis. As regards<br />
reflexive spaces, it was shown in [4, Theorem 3.1] using a result from [1, Lemma 1]<br />
(and answering a question from that article), that if a reflexive Banach space embeds<br />
into a space with an unconditional basis, then it embeds into a reflexive space with<br />
an unconditional basis. As regards the quotient problem mentioned above, Feder [3,<br />
Theorem 4] gave a partial solution by proving that if X is a quotient of a reflexive<br />
space that has a UFDD and X has the approximation property, then X embeds into a<br />
space with an unconditional basis.<br />
It is well known and easy to see that if a Banach space X embeds into a space<br />
with an unconditional basis, then X has the unconditional subsequence property;<br />
that is, there exists a K > 0 such that every normalized weakly null sequence in<br />
X has a subsequence that is K-unconditional. In fact, failure of the unconditional<br />
subsequence property is the only known criterion for proving that a given reflexive<br />
space does not embed into a space with an unconditional basis. However, in Section<br />
3, we construct a Banach space that has the unconditional subsequence property<br />
but does not embed into a Banach space that has an unconditional basis. This is not<br />
surprising, given previous examples of Odell and Schlumprecht [12]. Moreover, Odell<br />
and Schlumprecht have taught us that when a subsequence property is replaced with<br />
the corresponding “branch of a tree” property (see [12, introduction]), the result is a<br />
stronger property that sometimes can be used to give a characterization of spaces that<br />
embed into a space with some kind of structure. The relevant property for us here is<br />
the unconditional tree property (UTP), and Odell and Schlumprecht’s beautiful results<br />
are essential tools for us in applying it.<br />
We use standard Banach space theory terminology, such as can be found in [11].<br />
2. Main results<br />
Definition 2.1<br />
Let [N]
A CHARACTERIZATION OF SUBSPACES AND QUOTIENTS 507<br />
ordered subset of the tree under the tree order. We say that X has the C-UTP if every<br />
normalized weakly null tree in X has a C-unconditional branch for some C>0 and<br />
that X has the UTP if X has the C-UTP for some C>0.<br />
Remark 2.2<br />
Odell, Schlumprecht, and Zsákprovedin[14, Proposition 2.2] that if every normalized<br />
weakly null tree in X admits a branch that is unconditional, then X has the C-UTP<br />
for some C > 0. A simpler proof appears in the preprint of Haydon, Odell, and<br />
Schlumprecht [5]. There is, therefore, no ambiguity when using the term UTP.<br />
Given a finite-dimensional decomposition (FDD) (E n ), (x n ) is said to be a block<br />
sequence with respect to (E n ) if there exists a sequence of integers 0 = m 1
508 JOHNSON and ZHENG<br />
to zero. Then there are blockings (B<br />
n ′ ) of (B n) and (C<br />
n ′ ) of (C n) such that for any further<br />
blockings ( B˜<br />
n ) of (B<br />
n ′ ) with B˜<br />
n = ⊕ k n+1 −1<br />
i=k n<br />
B<br />
i ′ and ( C˜<br />
n ) of (C<br />
n ′ ) with C˜<br />
n = ⊕ k n+1 −1<br />
i=k n<br />
C<br />
i<br />
′<br />
and for any x ∈ B˜<br />
n ,thereisay ∈ ˜C n−1 ⊕ C˜<br />
n such that ‖Tx− y‖ ≤δ n ‖x‖.<br />
Proof<br />
Let (δ i ) be a sequence of positive numbers decreasing to zero. Let ( ˜δ i ) be another<br />
sequence of positive numbers that go to zero so fast that ∑ ∞<br />
˜<br />
j=i<br />
δ j
A CHARACTERIZATION OF SUBSPACES AND QUOTIENTS 509<br />
([12, Theorem 3.3]) there is a blocking (F n ) of the (E n ) which is a USB FDD. Then we<br />
use the “killing the overlap” technique of [6] to get a further blocking (G n ) so that any<br />
norm 1 vector y is a small perturbation of the sum of a skipped block sequence (y i )<br />
with respect to (F n ) and y i ∈ G i−1 ⊕ G i .LetQ : Y ↦→ X be the quotient map. Using<br />
Lemma 2.5 and passing to a further blocking, without loss of generality, we assume<br />
that QG i is essentially contained in H i−1 + H i , where (H i ) is the corresponding<br />
blocking of (V n ).Let(x A ) be a normalized weakly null tree in X. We then choose a<br />
branch (x Ai ) so lacunary that (x Ai ) is a small perturbation of a block sequence of (H n ),<br />
and for each i there is at least one H ki between the essential support of x Ai and x Ai+1 .<br />
Let x = ∑ a i x Ai with ‖x‖ =1. Considering a preimage y of x under the quotient<br />
Q from Y onto X (with ‖y‖ =1), by our construction we can essentially write y as<br />
the sum of (y i ), where (y i ) is a skipped block sequence with respect to (F n ).Since<br />
(F n ) is a USB, (y i ) is unconditional. By passing to a suitable blocking (z i ) of (y i ) and<br />
then using Lemma 2.5, it is not hard to show that Qz i is essentially equal to a i x Ai .<br />
Noticing that (z i ) is unconditional, we conclude that (x Ai ) is also unconditional.<br />
For the general case where X and Y do not have an FDD, we have to embed them<br />
into some superspaces with FDD. The difficulty is that when we decompose a vector<br />
in Y as the sum of disjointly supported vectors in the superspace, we do not know that<br />
the summands are in Y . The same problem occurs for vectors in X. This makes the<br />
proof rather technical and requires many computations.<br />
THEOREM 2.8<br />
Let X be a quotient of a separable reflexive Banach space Y with the UTP. Then X<br />
has the UTP.<br />
Proof<br />
By Zippin’s result (see [17]), Y embeds isometrically into a reflexive space Z with<br />
an FDD. A key point in the proof is that Odell and Schlumprecht proved (see [13,<br />
Proposition 2.4]) that there is a further blocking (G n ) of the FDD for Z, δ = (δ i ),and<br />
a C>0 such that every δ-skipped block sequence (y i ) ⊂ Y with respect to (G i ) is<br />
C-unconditional. Let λ be the basis constant for (G n ).<br />
Since X is separable, we can regard X as a subspace of L ∞ .Letɛ>0. We may<br />
assume that<br />
∑<br />
(a)<br />
j>i δ j
510 JOHNSON and ZHENG<br />
Let (x A ) be a normalized weakly null tree in X. Thenwelet(E n ) and (F n ) be<br />
blockings of (G i ) and (v i ), respectively, which satisfy the conclusions of Lemmas 2.5<br />
and 2.6. Using the “killing the overlap” technique (see [13, Proposition 2.6]), we can<br />
find a further blocking ( E˜<br />
n = ⊕ l(n+1)<br />
i=l(n)+1 E i)<br />
so that for every y ∈ SY , there exist<br />
(y i ) ⊂ Y and integers (t i ) with l(i − 1)
A CHARACTERIZATION OF SUBSPACES AND QUOTIENTS 511<br />
This gives an estimate of the second term. For the third term, we have<br />
∥<br />
∥a i x Ai −<br />
( k 2i−1−1<br />
∑<br />
P˜<br />
j<br />
)x∥ < ∥<br />
( k 2i−1−1<br />
∑<br />
P˜<br />
∥<br />
j<br />
)(a i x Ai − x) ∥ + ∥a i x Ai −<br />
( k 2i−1−1<br />
∑<br />
∥<br />
P˜<br />
∥∥<br />
j<br />
)x Ai<br />
j=k 2i−2<br />
j=k 2i−2<br />
< 2<br />
(k 2i−2 δ k2i−2 + ∑ )<br />
δ j + 2δ i<br />
j≥k 2i−1<br />
j=k 2i−2<br />
< 2(δ k2i−2 −1 + δ k2i−1 −1) + 2δ i < 4δ i . (2.4)<br />
For the first term, let Q j be the canonical projection from X onto F j ,andletJ 1 =<br />
[t k2i−3 +1,t k2i−1 +1],J 2 = [l k2i−2 + 1,l k2i−1 ],andJ 1 ′ = (t k 2i−3 +1,t k2i−1 +1). Thenwehave<br />
(<br />
∥<br />
∑ ) ( ∑ )<br />
∥Q P j y − Q j Qy∥<br />
j∈J 1 j∈J 2 (<br />
∥ ∑ ) ( ∑ )<br />
( ∑ ) ( ∑ )<br />
≤ ∥Q P j y − Q j Qy∥ + ∥ Q j Qy − Q j Qy∥<br />
j∈J 1 j∈J 1 j∈J 1 j∈J 2 (<br />
∥<br />
∑ ) ( ∑ )<br />
= ∥Q P j y − Q j Qy∥ + ∥ ∑ ( ∑ )∥ ∥∥<br />
Q j ai x Ai<br />
j∈J 1 j∈J 1 j∈J 1 −J 2 (<br />
∥<br />
∑ ) ( ∑ )<br />
< ∥Q P j y − Q j Qy∥ + 4δ i<br />
j∈J 1 j∈J 1 ( ≤<br />
∑ ) ( ∑ )∥ ∥( ∥∥ ∥∥<br />
∑ ) ( ∑ )∥ ∥∥<br />
∥ Q j Q P j y + Q j Q P j y + 4δi<br />
j/∈J 1 j∈J 1 j∈J 1 j/∈J 1 ( < ∥<br />
∑ ( ∑ )∥ ∥( ∥∥ ∥∥<br />
∑ ( ∑ )∥ ∥∥<br />
Q P j y + Q j<br />
)Q<br />
+ 6δi<br />
j/∈J 1<br />
Q j<br />
)<br />
j∈J ′ 1<br />
j∈J ′ 1<br />
j/∈J 1<br />
P j y<br />
< 2λδ i + 2λδ i + 6δ i = (4λ + 6)δ i . (2.5)<br />
From inequalities (2.2)–(2.5), we conclude that<br />
‖Qz i − a i x Ai ‖ < (4λ + 12)δ i .<br />
Let (ɛ i ) ⊂{−1, 1} N .LetI ⊂ N be the set of indices i ∈ N for which ‖y i ‖
512 JOHNSON and ZHENG<br />
Remark 2.9<br />
If the original space Y has the (1 + ɛ)-UTP for any ɛ>0, then any quotient of Y has<br />
the (1 + ɛ)-UTP for any ɛ>0.<br />
The following is an elementary lemma, which is used later. We omit the standard<br />
proof.<br />
LEMMA 2.10<br />
Let X be a Banach space, and let X 1 ,X 2 be two closed subspaces of X.IfX 1 ∩X 2 ={0}<br />
and X 1 + X 2 is closed, then X embeds into X/X 1 ⊕ X/X 2 .<br />
In [7, Theorem 4.4], Johnson and Rosenthal proved that any separable Banach space<br />
X admits a subspace Y so that both Y and X/Y have an FDD. The proof uses<br />
Markuschevich bases; a Markuschevich basis for a separable Banach space X is a<br />
biorthogonal system {x n ,xn ∗} n∈N for which the span of the x n ’s is dense in X and the<br />
xn ∗ ’s separate the points of X.By[11, Theorem 1.f.4], every separable Banach space X<br />
has a Markuschevich basis {x n ,xn ∗} n∈N so that [xn ∗ ] contains any designated separable<br />
subspace of X ∗ . The following lemma is a stronger form of the result of Johnson and<br />
Rosenthal, which follows from the original proof. For the convenience of the reader,<br />
we give a sketch of the proof. We use [x i ] i∈I to denote the closed linear span of (x i ) i∈I .<br />
LEMMA 2.11<br />
Let X be a separable Banach space. Then there exists a subspace Y with FDD (E n ) such<br />
that for any blocking (F n ) of (E n ) and for any sequence (n k ) ⊂ N, X/span{(F nk ) ∞ k=1 }<br />
admits an FDD (G n ). Moreover, if X ∗ is separable, (E n ) and (G n ) can be chosen to<br />
be shrinking.<br />
Proof<br />
Let {x i ,xi ∗} be a Markuschevich basis for X so that [x∗ i<br />
] is a norm-determining<br />
subspace of X ∗ and even [xi ∗] = X∗ if X ∗ is separable. Then we can choose inductively<br />
finite sets σ 1 ⊂ σ 2 ⊂··· and η 1 ⊂ η 2 ⊂··· so that σ = ⋃ ∞<br />
n=1 σ n and η = ⋃ ∞<br />
n=1 η n<br />
are complementary infinite subsets of the positive integers and for n = 1, 2,...,<br />
(i) if x ∗ ∈ [xi ∗] i∈η n<br />
, there is an x ∈ [x i ] i∈ηn ∪σ n+1<br />
such that ‖x‖ =1 and |x ∗ (x)| ><br />
(1 − 1/(n + 1))‖x ∗ ‖;<br />
(ii) if x ∈ [x i ] i∈σn , there is an x ∗ ∈ [xi ∗] i∈σ n ∪η n<br />
such that ‖x ∗ ‖=1 and |x ∗ (x)| ><br />
(1 − 1/(n + 1))‖x‖.<br />
Once we have this, by [7, proof of Theorem 4] we have it that [x i ] ⊥ i∈σ is the w∗ -<br />
closure of [xi ∗] i∈η. PutY = [xi ∗]⊥ i∈η<br />
= [x i] i∈σ . By the analogue of [7, Proposition<br />
2.1(a)], we deduce that X/Y has an FDD and that ([x i ] i∈σn ) ∞ n=1 forms an FDD for<br />
Y . In order to prove Lemma 2.11, it is enough to prove that for any blocking ( n ) of<br />
(σ n ) or any subsequence (σ nk ) of (σ n ) (this, of course, needs the redefining of (η n )),
A CHARACTERIZATION OF SUBSPACES AND QUOTIENTS 513<br />
(i) and (ii) still hold. But this is more or less obvious because if n = ⋃ k n<br />
i=k n−1 +1 σ i,<br />
then we define n = ⋃ k n<br />
i=k n−1 +1 η i and it is easy to check that { n , n } satisfy (i)<br />
and (ii). For a subsequence (σ nk ),ifwelet k = σ nk and define k = ⋃ n k+1 −1<br />
i=n k<br />
η i ,<br />
then { n , n } satisfy (i) and (ii). The rest is exactly the same as in [7, proof of Theorem<br />
4.4].<br />
<br />
The next lemma shows that for a reflexive space with a USB FDD, its dual also has a<br />
USB FDD.<br />
LEMMA 2.12<br />
Let X be a reflexive Banach space with a USB FDD (E n ). Then there is a blocking<br />
(F n ) of (E n ) such that (F ∗ n ) isaUSBFDDforX∗ .<br />
Proof<br />
Without loss of generality, we assume that (E n ) is monotone. Let (δ i ) be a sequence<br />
of positive numbers decreasing fast to zero. By the “killing the overlap” technique,<br />
we get a blocking (F n ) of (E n ) with F n = ∑ k n<br />
i=k n−1 +1 E i so that given any x = ∑ x i<br />
with x i ∈ E i , ‖x‖ =1, there is an increasing sequence (t n ) with k n−1
514 JOHNSON and ZHENG<br />
where C is the unconditional constant associated with the USB FDD (E n ).Ifwelet<br />
∑<br />
δi <br />
1 − ɛ<br />
C(1 + ɛ) .<br />
Therefore (xi ∗ ) is unconditional with unconditional constant less than (1 + 3ɛ)C if ɛ<br />
is sufficiently small. Hence (Fn ∗ ) is a USB FDD.<br />
<br />
THEOREM 2.13<br />
Let X be a separable reflexive Banach space. Then the following are equivalent.<br />
(a) X has the UTP.<br />
(b) X embeds into a reflexive Banach space with a USB FDD.<br />
(c) X ∗ has the UTP.<br />
Proof<br />
It is obvious that (b) implies (a). If we can prove that (a) implies (b) and that X satisfies<br />
(b), then by Lemma 2.12, X ∗ is a quotient of a reflexive space with a USB FDD. So,<br />
by Theorem 2.8, X ∗ has the UTP. Hence we need only show that (a) implies (b). Let<br />
X 1 be a subspace of X with an FDD (E n ) given by Lemma 2.11.By[13, Proposition<br />
2.4], we get a blocking (F n ) of (E n ) so that (F n ) is a USB FDD. Let Y 1 = [F 4n ],and<br />
let Y 2 = [F 4n+2 ].Then(F 4n ) and (F 4n+2 ) form UFDDs for Y 1 and Y 2 . By Lemma 2.11,<br />
X/Y i has an FDD. Since X has the UTP, by Theorem 2.8 we know that X/Y i has<br />
the UTP. By using [13, Proposition 2.4] again, we know that X/Y i has a USB FDD.<br />
Noticing that Y 1 ∩ Y 2 ={0} and that Y 1 + Y 2 is closed, by Lemma 2.10 we have that<br />
X embeds into X/Y 1 ⊕ X/Y 2 . Hence X embeds into a reflexive space with a USB<br />
FDD.<br />
<br />
COROLLARY 2.14<br />
Let X be a separable reflexive Banach space with the UTP. Then X embeds into a<br />
reflexive Banach space with an unconditional basis.<br />
Proof<br />
By Theorem 2.13, X embeds into a reflexive space Y with a USB FDD (E n ).We<br />
prove that Y embeds into a reflexive space with a UFDD. Then, as mentioned in the<br />
introduction, Y embeds into a reflexive space with an unconditional basis, and so X<br />
does, too.<br />
By Lemma 2.12, there is a blocking (F n ) of (E n ) such that (Fn ∗ ) is a USB FDD for<br />
Y ∗ .Now,letY 1 = ⊕ F 4n ,andletY 2 = ⊕ F 4n+2 .ThenwehaveY 1 ∩ Y 2 ={0}, and<br />
Y 1 + Y 2 is closed because (F 2n ), being a skipped blocking of (E n ), is unconditional.<br />
By Lemma 2.10, Y embeds into Y/Y 1 ⊕ Y/Y 2 .Since(Y/Y i ) ∗ is isomorphic to Yi<br />
⊥ ,it
A CHARACTERIZATION OF SUBSPACES AND QUOTIENTS 515<br />
is enough to prove that Yi<br />
⊥ has a UFDD. Let G ∗ n = F 4n−3 ∗ ⊕ F 4n−2 ∗ ⊕ F 4n−1 ∗ .Itiseasy<br />
to see that (G ∗ n ) forms an FDD for Y 1 ⊥. Noticing that (G n) is a skipped blocking of<br />
(Fn ∗), we conclude that (G n) is unconditional. Similarly, we can show that Y2 ⊥ admits<br />
a UFDD. This finishes the proof.<br />
<br />
COROLLARY 2.15<br />
Let X be a quotient of a reflexive Banach space with a UFDD. Then X embeds into a<br />
reflexive Banach space with an unconditional basis.<br />
Proof<br />
Combine Theorem 2.8 and Corollary 2.14.<br />
<br />
We mention again that in 1974, Davis, Figiel, Johnson, and Pełczyński proved in [1]<br />
that a reflexive Banach space X that embeds into a Banach space with a shrinking<br />
unconditional basis embeds into a reflexive space X with an unconditional basis.<br />
The next year, Figiel, Johnson, and Tzafriri [4] got a stronger result by removing the<br />
shrinkingness of the unconditional basis in the hypothesis. Our next corollary gives a<br />
parallel result for quotients.<br />
COROLLARY 2.16<br />
Let X be a separable reflexive Banach space. If X is a quotient of a Banach space<br />
with a shrinking unconditional basis, then X is isomorphic to a quotient of a reflexive<br />
Banach space with an unconditional basis.<br />
Proof<br />
Since X is a quotient of a Banach space with a shrinking unconditional basis, X ∗ is a<br />
subspace of a Banach space with an unconditional basis. Hence by [4, Theorem 3.1],<br />
X ∗ is isomorphic to a subspace of a reflexive Banach space with an unconditional<br />
basis. Therefore, X is isomorphic to a quotient of a reflexive Banach space with an<br />
unconditional basis.<br />
<br />
Remark 2.17<br />
Corollary 2.16 is different from the result of [4] in that the shrinkingness in our result<br />
cannot be removed. The reason is more or less obvious since every separable Banach<br />
space is a quotient of l 1 , which has an unconditional basis.<br />
Gluing Theorem 2.13 and Corollaries 2.14, 2.15, and2.16 together, we have the<br />
following long list of equivalences.<br />
THEOREM 2.18<br />
Let X be a separable reflexive Banach space. Then the following are equivalent.
516 JOHNSON and ZHENG<br />
(a)<br />
(b)<br />
(c)<br />
(d)<br />
(e)<br />
(f)<br />
(g)<br />
(h)<br />
(i)<br />
X has the UTP.<br />
X is isomorphic to a subspace of a Banach space with an unconditional basis.<br />
X is isomorphic to a subspace of a reflexive space with an unconditional basis.<br />
X is isomorphic to a quotient of a Banach space with a shrinking unconditional<br />
basis.<br />
X is isomorphic to a quotient of a reflexive space with an unconditional basis.<br />
X is isomorphic to a subspace of a quotient of a reflexive space with an<br />
unconditional basis.<br />
X is isomorphic to a subspace of a reflexive quotient of a Banach space with a<br />
shrinking unconditional basis.<br />
X is isomorphic to a quotient of a subspace of a reflexive space with an<br />
unconditional basis.<br />
X is isomorphic to a quotient of a reflexive subspace of a Banach space with a<br />
shrinking unconditional basis.<br />
3. Example<br />
In this section, we give an example of a reflexive Banach space for which there exists<br />
a C>0 such that every normalized weakly null sequence admits a C-unconditional<br />
subsequence, while for any D>0 there is a normalized weakly null tree such that<br />
every branch is not D-unconditional. The construction is an analogue of Odell and<br />
Schlumprecht’s example (see [12, Example 4.2]).<br />
We first construct an infinite sequence of reflexive Banach spaces X n . Each X n is<br />
infinite-dimensional and has the property that for ɛ>0, every normalized weakly null<br />
sequence has a (1 + ɛ)-unconditional basic subsequence, while there is a normalized<br />
weakly null tree for which every branch is at least C n -unconditional and C n goes to<br />
infinity when n goes to infinity. Then the l 2 -sum of X n ’s is a reflexive Banach space<br />
with the desired property.<br />
Let [N] ≤n be the set of all subsets of the positive integers with cardinality less than<br />
or equal to n. Letc 00 ([N] ≤n ) be the space of sequences with finite support indexed<br />
by [N] ≤n , and denote its canonical basis by (e A ) A∈[N] ≤n. Let(h i ) be any normalized<br />
conditional basic sequence that satisfies a block lower l 2 -estimate with constant 1,for<br />
example, the boundedly complete basis of James’s space (see [2, Problem 6.41]). Let<br />
∑<br />
aA e A be an element of c 00 ([N] ≤n ).Let(β k ) m k=1 be disjoint segments. By “a segment<br />
in [N] ≤n ,” we mean a sequence (A i ) k i=1 ∈ [N]≤n with A 1 ={n 1 ,n 2 ,...,n l },A 2 =<br />
{n 1 ,n 2 ,...,n l ,n l+1 },...,A k ={n 1 ,n 2 ,...,n l ,...,n l+k−1 } for some n 1
A CHARACTERIZATION OF SUBSPACES AND QUOTIENTS 517<br />
Let X = ( ∑ X n<br />
)2 .LetC M be the unconditional constant of (h i ) M i=1 . It is clear that C M<br />
tends to infinity when M goes to infinity. The normalized weakly null tree (e A ) A∈[N] ≤M<br />
in X M has the property that every branch of it is 1-equivalent to (h i ) M i=1 since (h i) has<br />
a block lower l 2 -estimate with constant 1. So what is remaining is to verify that for<br />
every ɛ>0, every normalized weakly null sequence in X has a (1 + ɛ)-unconditional<br />
basic subsequence. Actually, we prove that there is a subsequence which is (1 + ɛ)-<br />
equivalent to the unit vector basis of l 2 . By a gliding-hump argument, it is not hard to<br />
verify the following fact.<br />
Fact<br />
Let (Y k ) be a sequence of reflexive Banach spaces, and let Y = ( ∑ Y k<br />
)l 2<br />
. If for every<br />
ɛ>0,k ∈ N, every normalized weakly null sequence in Y k has a subsequence that is<br />
(1+ɛ)-equivalent to the unit vector basis of l 2 , then for every ɛ>0, every normalized<br />
weakly null sequence in Y has a subsequence that is (1 + ɛ)-equivalent to the unit<br />
vector basis of l 2 .<br />
Considering this fact, it is enough to show that for every ɛ > 0,k ∈ N, every<br />
normalized weakly null sequence in X k has a subsequence that is (1 + ɛ)-equivalent<br />
to the unit vector basis of l 2 . We prove this by induction.<br />
For k = 1, X 1 is isometric to l 2 , so the conclusion is obvious.<br />
Assume that the conclusion is true for X k . By the definition of X k+1 , X k+1 is<br />
isometric to ( ∑ (R ⊕ X k ) ) l 2<br />
(where R ⊕ X k has some norm so that {0} ⊕X k is<br />
isometric to X k ). Hence by hypothesis and the fact mentioned above, it is easy to see<br />
that the conclusion is true in X k+1 . This finishes the proof.<br />
Remark 3.1<br />
The proof of the corresponding induction step in [12, Example 4.2] is more complicated<br />
than the very simple induction argument in the previous paragraph. Schlumprecht<br />
realized after [12] was published that the induction could be done this simply (see<br />
[16]), and his argument works in our context.<br />
Acknowledgments. The authors thank the referees for useful corrections, especially<br />
for pointing out the imprecision in the initial construction of the example in Section<br />
3. This article is based in part on the doctoral dissertation of Zheng, which is being<br />
prepared at Texas A&M University under Johnson’s direction.<br />
References<br />
[1] W. J. DAVIS, T. FIGIEL, W. B. JOHNSON, andA. PEŁCZYŃSKI, Factoring weakly<br />
compact operators, J. Functional Analysis 17 (1974), 311 – 327. MR 0355536<br />
506, 515
518 JOHNSON and ZHENG<br />
[2] M. FABIAN, P. HABALA, P. HÁJEK, V. MONTESINOS SANTALUCÍA, J. PELANT,andV.<br />
ZIZLER, Functional Analysis and Infinite-Dimensional Geometry, CMS Books<br />
Math./Ouvrages Math. SMC 8, Springer, New York, 2001. MR 1831176 516<br />
[3] M. FEDER, On subspaces of spaces with an unconditional basis and spaces of<br />
operators, Illinois J. Math. 24 (1980), 196 – 205. MR 0575060 506<br />
[4] T. FIGIEL, W. B. JOHNSON, andL. TZAFRIRI, On Banach lattices and spaces having<br />
local unconditional structure, with applications to Lorentz function spaces,<br />
J. Approximation Theory 13 (1975), 395 – 412. MR 0367624 506, 515<br />
[5] R. HAYDON, E. ODELL,andT. SCHLUMPRECHT, Small subspaces of L p ,preprint,<br />
arXiv:0711.3919v1 [math.FA] 507<br />
[6] W. B. JOHNSON, On quotients of L p which are quotients of l p , Compositio Math. 34<br />
(1977), 69 – 89. MR 0454595 509<br />
[7] W. B. JOHNSON and H. P. ROSENTHAL, On ω ∗ -basic sequences and their applications<br />
to the study of Banach spaces, Studia Math. 43 (1972), 77 – 92. MR 0310598<br />
512, 513<br />
[8] W. B. JOHNSON and A. SZANKOWSKI, Complementably universal Banach spaces,<br />
[9]<br />
Studia Math. 58 (1976), 91 – 97. MR 0425582<br />
W. B. JOHNSON and M. ZIPPIN, “On subspaces of quotients of ( ∑ G n<br />
)l p<br />
and<br />
( ∑ )<br />
Gn ”inProceedings of the International Symposium on Partial Differential<br />
c 0<br />
Equations and the Geometry of Normal Linear Spaces (Jerusalem, 1972), Israel J.<br />
Math. 13 (1972), 311 – 316. MR 0331023 508<br />
[10] ———, Subspaces and quotient spaces of ( ∑ )<br />
G n l p<br />
and ( ∑ G n , Israel J. Math. 17<br />
)c 0<br />
(1974), 50 – 55. MR 0358296 507, 508<br />
[11] J. LINDENSTRAUSS and L. TZAFRIRI, Classical Banach Spaces, I: Sequence Spaces,<br />
Ergeb. Math. Grenzgeb. 92, Springer, Berlin, 1977. MR 0500056 506, 507, 512<br />
[12] E. ODELL and T. SCHLUMPRECHT, Trees and branches in Banach spaces,Trans.Amer.<br />
Math. Soc. 354, no. 10 (2002), 4085 – 4108. MR 1926866 506, 509, 516, 517<br />
[13] ———, A universal reflexive space for the class of uniformly convex Banach spaces,<br />
Math. Ann. 335 (2006), 901 – 916. MR 2232021 508, 509, 510, 514<br />
[14] E. ODELL, T. SCHLUMPRECHT,andA. ZSÁK, On the structure of asymptotic l p spaces,<br />
to appear in Q. J. Math. 507<br />
[15] A. PEŁCZYŃSKI and P. WOJTASZCZYK, Banach spaces with finite dimensional<br />
expansions of identity and universal bases of finite dimensional subspaces, Studia<br />
Math. 40 (1971), 91 – 108. MR 0313765 506<br />
[16] T. SCHLUMPRECHT, private communication, 2006. 517<br />
[17] M. ZIPPIN, Banach spaces with separable duals,Trans.Amer.Math.Soc.310, no. 1<br />
(1988), 371 – 379. MR 0965758 509<br />
Johnson<br />
Department of Mathematics, Texas A&M University, College Station, Texas 77843, USA;<br />
johnson@math.tamu.edu<br />
Zheng<br />
Department of Mathematics, University of Texas at Austin, Austin, Texas 78712, USA;<br />
btzheng@math.utexas.edu
DEGREE GROWTH OF MEROMORPHIC<br />
SURFACE MAPS<br />
SÉBASTIEN BOUCKSOM, CHARLES FAVRE, and MATTIAS JONSSON<br />
Abstract<br />
We study the degree growth of iterates of meromorphic self-maps of compact Kähler<br />
surfaces. Using cohomology classes on the Riemann-Zariski space, we show that the<br />
degrees grow similarly to those of mappings that are algebraically stable on some<br />
bimeromorphic model.<br />
0. Introduction<br />
Let X be a compact Kähler surface, and let F : X X be a dominant meromorphic<br />
mapping. Fix a Kähler class ω on X, normalized by (ω 2 ) X = 1, and define the degree<br />
of F with respect to ω to be the positive real number<br />
deg ω (F ):= (F ∗ ω · ω) X = (ω · F ∗ ω) X ,<br />
where (·) X denotes the intersection form on H 1,1<br />
R<br />
(X). WhenX = P2 and ω is the<br />
class of a line, this coincides with the usual algebraic degree of F . One can show that<br />
deg ω (F n+m ) ≤ 2deg ω (F n )deg ω (F m ) for all m, n. Hence the limit<br />
λ 1 := lim<br />
n→∞<br />
deg ω (F n ) 1/n<br />
exists. We refer to it as the asymptotic degree of F . It follows from standard arguments<br />
(see Proposition 3.1) thatλ 1 does not depend on the choice of ω, thatλ 1 is invariant<br />
under bimeromorphic conjugacy, and that λ 2 1 ≥ λ 2, where λ 2 is the topological degree<br />
of F .<br />
MAIN THEOREM<br />
Assume that λ 2 1 >λ 2. Then there exists a constant b = b(ω) > 0 such that<br />
deg ω (F n ) = bλ n 1 + O(λn/2 2 ) as n →∞.<br />
DUKE MATHEMATICAL JOURNAL<br />
Vol. 141, No. 3, c○ 2008 DOI 10.1215/00127094-2007-004<br />
Received 28 August 2006. Revision received 22 May 2007.<br />
2000 Mathematics Subject Classification. Primary 32H50; Secondary 14E05, 14C17.<br />
Boucksom’s work supported in part by the Japanese Society for the Promotion of Science.<br />
Jonsson’s work supported in part by National Science Foundation grant DMS-0449465, the Swedish Research<br />
Council, and the Gustafsson Foundation.<br />
519
520 BOUCKSOM, FAVRE, and JONSSON<br />
The dependence of b on ω can be made explicit (see Remark 3.7). For the polynomial<br />
map F (x,y) = (x d ,x d y d ) on C 2 (with ω the standard Fubini-Study form), one has<br />
λ 2 = λ 2 1 = d2 , deg ω (F n ) = nd n ; hence the assertion in the main theorem may fail<br />
when λ 2 1 = λ 2.<br />
Degree growth is an important component in the understanding of the complexity<br />
and dynamical behavior of a self-map and has been studied in a large number of works<br />
in both mathematics and physics literature. It is connected to topological entropy<br />
(see, e.g., [Fr], [G1], [G2], [DS]), and controlling it is necessary in order to construct<br />
interesting invariant measures and currents (see, e.g., [BF], [FS], [RS], [S]). Even in<br />
simple families of mappings, degree growth exhibits a rich behavior (see, e.g., the articles<br />
by Bedford and Kim [BK1], [BK2], which also contain references to the physics<br />
literature).<br />
In [FS], Fornaess and Sibony connected the degree growth of rational self-maps to<br />
the interplay between contracted hypersurfaces and indeterminacy points. In particular,<br />
they proved that deg(F n ) is multiplicative if and only if F is what is now often called<br />
(algebraically) stable. This analysis was extended to slightly more general maps in [N].<br />
Bonifant and Fornaess [BF] showed that only countably many sequences (deg(F n )) ∞ 1<br />
can occur, but in general, the precise picture is unclear.<br />
For bimeromorphic maps of surfaces, the situation is quite well understood since<br />
the work of Diller and Favre [DF]. Using the factorization into blowups and blowdowns,<br />
they proved that any such map can be made stable by a bimeromorphic change<br />
of coordinates. This reduces the study of degree growth to the spectral properties of<br />
the induced map on the Dolbeault cohomology H 1,1 . In particular, it implies that λ 1 is<br />
an algebraic integer and that deg(F n ) satisfies an integral recursion formula and gives<br />
a stronger version of our main theorem when λ 2 1 > 1(= λ 2).<br />
In the case that we consider, namely, (noninvertible) meromorphic surface maps,<br />
there are counterexamples to stability when λ 2 1 = λ 2 > 1 (see [F]). It is an interesting<br />
(and probably difficult) question whether counterexamples also exist with<br />
λ 2 1 >λ 2 > 1.<br />
Instead of looking for a particular birational model in which the action of F n<br />
on H 1,1 can be controlled, we take a different tack and study the action of F on<br />
cohomology classes on all modifications π : X π → X at the same time. This idea<br />
already appeared in the study of cubic surfaces in [M] and was recently used by Cantat<br />
as a key tool in his investigation of the group of birational transformation of surfaces<br />
(see [C2]). In the context of noninvertible maps, Hubbard and Papadopol [HP] used<br />
similar ideas, but their methods apply only to a quite restricted class of maps.<br />
Here we show that F acts (functorially) by pullback F ∗ and pushforward F ∗ on<br />
the vector space W := lim H 1,1<br />
←−<br />
R<br />
(X π) and on its dense subspace C := lim H 1,1<br />
−→<br />
R<br />
(X π).<br />
Compactness properties of W imply the existence of eigenvectors having eigenvalue<br />
λ 1 and certain positivity properties.
DEGREE GROWTH OF MEROMORPHIC SURFACE MAPS 521<br />
Following [DF], we then study the spectral properties of F ∗ and F ∗ under the<br />
assumption that λ 2 1 >λ 2. The space W is too big for this purpose, and we introduce a<br />
subspace L 2 that is the completion of C with respect to the (indefinite) inner product<br />
induced by the cup product, which is of Minkowski type by the Hodge index theorem.<br />
The main theorem then follows from the spectral properties of F ∗ and its adjoint F ∗<br />
on L 2 .<br />
Using a different method, polynomial mappings of C 2 were studied in detail by<br />
Favre and Jonsson in [FJ4]: in that case, λ 1 is a quadratic integer. However, our main<br />
theorem for polynomial maps does not immediately follow from the analysis in [FJ4];<br />
the methods of the two articles can be viewed as complementary.<br />
The space W above can be thought of as the Dolbeault cohomology H 1,1 of the<br />
Riemann-Zariski space of X. While we do not need the structure of the latter space<br />
in this article, the general philosophy of considering all bimeromorphic models at the<br />
same time is very useful for handling asymptotic problems in geometry, analysis, and<br />
dynamics (see [BFJ], [C1], [M], [FJ1], [FJ2], [FJ3]). In the present setting, it allows<br />
us to bypass the intricacies of indeterminacy points: heuristically, a meromorphic map<br />
becomes holomorphic on the Riemann-Zariski space.<br />
The article is organized in three sections. In the first, we recall some definitions<br />
and introduce cohomology classes on the Riemann-Zariski space. In the second, we<br />
study the actions of meromorphic mappings on these classes. Finally, Section 3 deals<br />
with the spectral properties of these actions under iteration, concluding with the proof<br />
of the main theorem.<br />
Remark on the setting. We choose to state our main result in the context of a complex<br />
manifold because the study of degree growth is particularly important for applications<br />
to holomorphic dynamics. However, our methods are purely algebraic, so that our main<br />
result actually holds in the case when X is a projective surface over any algebraically<br />
closed field of any characteristic, and ω = c 1 (L) for some ample line bundle. In this<br />
(X) with the real Néron-Severi vector space and work<br />
with the suitable notion of pseudoeffective and nef classes, as defined in [L, Sections<br />
1.4, 2.2].<br />
context, one has to replace H 1,1<br />
R<br />
1. Classes on the Riemann-Zariski space<br />
Let X be a complex compact Kähler surface (for background, see [BHPV]), and write<br />
(X) := H 1,1 (X) ∩ H 2 (X, R).<br />
H 1,1<br />
R<br />
1.1. The Riemann-Zariski space<br />
By a blowup of X we mean a bimeromorphic morphism π : X π → X, where X π is a<br />
smooth surface. Up to isomorphism, π is then a finite composition of point blowups.<br />
If π and π ′ are two blowups of X, we say that π ′ dominates π and write π ′ ≥ π if<br />
there exists a bimeromorphic morphism µ : X π ′ → X π such that π ′ = π ◦ µ. The
522 BOUCKSOM, FAVRE, and JONSSON<br />
Riemann-Zariski space of X is the projective limit<br />
X := lim ←−π<br />
X π .<br />
While suggestive, the space X is, strictly speaking, not needed for our analysis, and<br />
we refer to [ZS, Chapter 6, Section 17], [V, Section 7] for details on its structure.<br />
1.2. Weil and Cartier classes<br />
When one blowup π ′ = π ◦ µ dominates another one π, we have two induced linear<br />
maps, µ ∗ : H 1,1<br />
R (X 1,1<br />
π ′) → HR (X π) and µ ∗ : H 1,1<br />
R (X π) → H 1,1<br />
R<br />
(X π ′), which satisfy<br />
the projection formula µ ∗ µ ∗ = id. This allows us to define the following spaces.<br />
Definition 1.1<br />
The space of Weil classes on X is the projective limit<br />
W (X) := lim ←−π<br />
H 1,1<br />
R (X π)<br />
with respect to the pushforward arrows. The space of Cartier classes on X is the<br />
inductive limit<br />
with respect to the pullback arrows.<br />
C(X) := lim −→π<br />
H 1,1<br />
R (X π)<br />
The space W (X) is endowed with its projective limit topology, that is, the coarsest<br />
topology for which the projection maps W (X) → H 1,1<br />
R<br />
(X π) are continuous. There is<br />
also an inductive limit topology on C(X), but we do not use it.<br />
Concretely, a Weil class α ∈ W (X) is given by its incarnations α π ∈ H 1,1<br />
R<br />
(X π),<br />
compatible by pushforward; that is, µ ∗ a π ′ = α π whenever π ′ = π ◦ µ. The topology<br />
on W (X) is characterized as follows: a sequence (or net † ) α j ∈ W(X) converges to<br />
α ∈ W (X) if and only if α j,π → α π in H 1,1<br />
R (X π) for each blowup π.<br />
The projection formula recalled above shows that there is an injection C(X) ⊂<br />
W (X), so that a Cartier class is, in particular, a Weil class. In fact, if α ∈ H 1,1<br />
R<br />
(X π) is<br />
a class in some blowup X π of X,thenα defines a Cartier class, also denoted α, whose<br />
incarnation α π ′ in any blowup π ′ = π ◦ µ dominating π is given by α π ′ = µ ∗ α.<br />
We say that α is determined in X π . (It is then also determined in X π ′ for any blowup<br />
dominating π.) Each Cartier class is obtained that way. The space C(X) is dense in<br />
W (X): ifα is a given Weil class, the net α π of Cartier classes determined by the<br />
incarnations of α on all models X π tautologically converges to α in W(X).<br />
† A net is a family indexed by a directed set (see [Fo]).
DEGREE GROWTH OF MEROMORPHIC SURFACE MAPS 523<br />
Remark 1.2<br />
The spaces of Weil classes and Cartier classes are denoted Z • (X) and Z • (X) by<br />
Manin [M]. He views these classes as living on the bubble space lim X −→ π rather than<br />
the Riemann-Zariski space lim X ←− π .<br />
1.3. Exceptional divisors<br />
This section can be skipped on a first reading, the main technical issue being Proposition<br />
1.6, which is used for the proof of Theorem 3.2.<br />
The spaces C(X) and W (X) are clearly bimeromorphic invariants of X. Once<br />
the model X is fixed, an alternative and somewhat more explicit description of these<br />
spaces can be given in terms of exceptional divisors.<br />
Definition 1.3<br />
The set D of exceptional primes over X is defined as the set of all exceptional prime<br />
divisors of all blowups X π → X modulo the following equivalence relation: two<br />
divisors E and E ′ on X π and X π ′ are equivalent if the induced meromorphic map<br />
X π X π ′ sends E onto E ′ .<br />
When X is a projective surface, D is the set of divisorial valuations on the function<br />
field C(X) whose center on X is a point.<br />
If E ∈ D is an exceptional prime and X π is any model of X, one can consider the<br />
center of E on X π , denoted by c π (E). It is a subvariety defined as follows. Choose<br />
ablowupπ ′ ≥ π so that E appears as a curve on X π ′.Thenc π (E) is defined as the<br />
image of E ⊂ X π ′ by the map X π ′ → X π . It does not depend on the choice of π ′ and<br />
is either a point or an irreducible curve. In this 2-dimensional setting, there is a unique<br />
minimal blowup π E such that c π (E) is a curve if and only if π ≥ π E . (In particular,<br />
c πE (E) is a curve.)<br />
Using these facts, one can construct an explicit basis for the vector space C(X)<br />
as follows (cf. [M, Proposition 35.6]). Let α E ∈ C(X) be the Cartier class determined<br />
by the class of E on X πE . Write R (D) for the direct sum ⊕ D<br />
R or, equivalently, for<br />
the space of real-valued functions on D with finite support.<br />
PROPOSITION 1.4<br />
The set {α E | E ∈ D} is a basis for the vector space of Cartier classes α ∈ C(X)<br />
which are exceptional over X, that is, whose incarnations on X vanish. In other<br />
words, the map H 1,1<br />
R (X) ⊕ R(D) → C(X), sending α ∈ H 1,1<br />
R<br />
(X) to the Cartier class<br />
it determines, and E ∈ D to α E is an isomorphism.<br />
We now describe W (X) in terms of exceptional primes. If α ∈ W(X) is a given<br />
Weil class, let α X ∈ H 1,1<br />
R<br />
(X) be its incarnation on X. For each π, the Cartier class
524 BOUCKSOM, FAVRE, and JONSSON<br />
α π − α X is determined on X π by a unique R-divisor Z π exceptional over X. IfE is<br />
a π-exceptional prime, we set ord E (α) := ord E (Z π ) so that Z π = ∑ E ord E(Z π )E.<br />
It is easily seen to depend only on the class of E in D. LetR D denote the (product)<br />
space of all real-valued functions on D. We obtain a map W(X) → H 1,1<br />
R (X) × RD ,<br />
which is easily seen to be a bijection, and even naturally a homeomorphism, as the<br />
following straightforward lemma shows.<br />
LEMMA 1.5<br />
Anetα j ∈ W (X) converges to α ∈ W (X) if and only if α j,X converges to α X in<br />
H 1,1<br />
R (X) and ord E(α j ) → ord E (α) for each exceptional prime E ∈ D.<br />
A result of Zariski (cf. [Ko, Theorem 3.17], [FJ1, Proposition 1.12]) states that the<br />
process of successively blowing up the center of a given exceptional prime E ∈ D<br />
starting from any given model must stop after finitely many steps with the center<br />
becoming a curve. In other words, if X = X 0 ← X 1 ← X 2 ← ··· is an infinite<br />
sequence of blowups such that the center of each blowup X n ← X n+1 meets c Xn (E),<br />
then X n must dominate X πE for n large enough. Using this result, we record the<br />
following fact, which is used later on in the article.<br />
PROPOSITION 1.6<br />
Let X = X 0 ← X 1 ← X 2 ← ··· be an infinite sequence of blowups, and for each<br />
n, suppose that α n ∈ C(X) is a Cartier class that is determined in X n+1 and whose<br />
incarnation on X n is zero. Then α n → 0 in W (X) as n →∞.<br />
Proof<br />
In view of Proposition 1.5, we have to show that for every given exceptional prime<br />
E ∈ D, ord E (α n ) converges to zero as n →∞. In fact, we claim that ord E (α n ) = 0<br />
for n ≥ n(E) large enough. Indeed, according to Zariski’s result, there are two<br />
possibilities: either there exists N such that c XN (E) is a curve, or there exists N such<br />
that the center of the blowup X n+1 → X n does not meet c Xn (E) for all n ≥ N.Inthe<br />
first case, it is clear that ord E (α n ) = 0 for n ≥ N since α n is exceptional over X N .<br />
In the second case, the center of E on X n does not meet the exceptional divisor of<br />
X n → X n−1 for n>N, which supports the exceptional class α n ; thus ord E (α n ) = 0<br />
for n>Nas well.<br />
<br />
1.4. Intersections and L 2 -classes<br />
For each π, the intersection pairing H 1,1<br />
R (X π)×H 1,1<br />
R (X π) → R is denoted by (α·β) Xπ .<br />
It is nondegenerate and satisfies the projection formula: (µ ∗ α · β) Xπ = (α · µ ∗ β) Xπ ′ if<br />
π ′ = π ◦ µ. It thus induces a pairing W (X) × C(X) → R which is denoted simply<br />
by (α · β).
DEGREE GROWTH OF MEROMORPHIC SURFACE MAPS 525<br />
PROPOSITION 1.7<br />
The intersection pairing induces a topological isomorphism between W(X) and C(X) ∗<br />
endowed with its weak-∗ topology.<br />
Proof<br />
A linear form L on C(X) = lim H 1,1<br />
−→π<br />
R<br />
(X π) is the same thing as a collection of<br />
linear forms L π on H 1,1<br />
R<br />
(X π), compatible by restriction. Now, such a collection is by<br />
definition an element of the projective limit lim H 1,1<br />
←−π<br />
R<br />
(X π) ∗ , which is identified to<br />
W (X) via the intersection pairing. This shows that the intersection pairing identifies<br />
W (X) with the dual of C(X) endowed with its weak-∗ topology.<br />
<br />
The intersection pairing defined above restricts to a nondegenerate quadratic form on<br />
C(X), denoted by α ↦→ (α 2 ). However, it does not extend to a continuous quadratic<br />
form on W (X). For instance, if z 1 ,z 2 ,...is a sequence of distinct points on X and π n<br />
denotes the blowup of X at z 1 ,...,z n , with exceptional divisor F n = E 1 +···+E n ,<br />
we have (F 2 n ) =−n, but{F n}∈C(X) converges in W (X). We thus introduce the<br />
maximal space to which the intersection form extends.<br />
Definition 1.8<br />
The space of L 2 -classes L 2 (X) is defined as the completion of C(X) with respect to<br />
the intersection form.<br />
The usual setting in which to perform a completion is that of a definite quadratic form<br />
on a vector space, which is not the case of the intersection form on C(X). However,<br />
the Hodge index theorem implies that it is of Minkowski type, and it is easy to show<br />
that the completion exists in that setting.<br />
Let us be more precise. If ω ∈ C(X) is a given class with (ω 2 ) > 0, the intersection<br />
form is negative definite on its orthogonal complement ω ⊥ :={α ∈ C(X) | (α·ω) = 0}<br />
(X π). Wehavean<br />
orthogonal decomposition C(X) = Rω ⊕ ω ⊥ , and we then let L 2 (X) := Rω ⊕ ω ⊥ ,<br />
where ω ⊥ is the completion in the usual sense of ω ⊥ endowed with the negative<br />
definite quadratic form (α 2 ). Note that tω ⊕ α ↦→ t 2 − (α 2 ) is then a norm on L 2 (X)<br />
which makes it a Hilbert space, but this norm depends on the choice of ω. However,<br />
the topological vector space L 2 (X) does not depend on the choice of ω.<br />
In fact, the completion can be characterized by the following universal property:<br />
if (Y, q) is a complete topological vector space with a continuous nondegenerate<br />
quadratic form of Minkowski type, any isometry T : C(X) → Y continuously<br />
extends to L 2 (X) → Y .<br />
as a consequence of the Hodge index theorem applied to each H 1,1<br />
R
526 BOUCKSOM, FAVRE, and JONSSON<br />
The intersection form on L 2 (X) is also of Minkowski type, so that it satisfies<br />
the Hodge index theorem: if a nonzero class α ∈ L 2 (X) satisfies (α 2 ) > 0, then the<br />
intersection form is negative definite on α ⊥ ⊂ L 2 (X).<br />
Remark 1.9<br />
The direct sum decomposition C(X) = H 1,1<br />
R<br />
(X) ⊕ R(D) of Proposition 1.4 is orthogonal<br />
with respect to the intersection form. Furthermore, the intersection form is negative<br />
definite on R (D) ,and{α E | E ∈ D} forms an orthonormal basis for −(α 2 ).Indeed,the<br />
center of E ∈ D on the minimal model X πE on which it appears is necessarily the last<br />
exceptional divisor to have been created in any factorization of π E into a sequence of<br />
point blowups; thus it is a (−1)-curve.<br />
Using this, one sees that L 2 (X) is isomorphic to the direct sum H 1,1<br />
R<br />
(X)⊕l2 (D) ⊂<br />
W (X), where l 2 (D) denotes the set of real-valued, square-summable functions E ↦→<br />
a E on D.<br />
The different spaces that we have introduced so far are related as follows.<br />
PROPOSITION 1.10<br />
There is a natural continuous injection L 2 (X) → W (X), and the topology on L 2 (X)<br />
induced by the topology of W (X) coincides with its weak topology as a Hilbert space.<br />
If α ∈ W (X) is a given Weil class, then the intersection number (απ 2 ) is a<br />
decreasing function of π, and α ∈ L 2 (X) if and only if (απ 2 ) is bounded from below,<br />
in which case, (α 2 ) = lim π (απ 2 ).<br />
Proof<br />
The injection L 2 (X) → W(X) is dual to the dense injection C(X) ⊂ L 2 (X). By<br />
Proposition 1.7,anetα k ∈ L 2 (X) converges to α ∈ L 2 (X) in the topology induced by<br />
W (X) if and only if (α k · β) → (α · β) for each β ∈ C(X). SinceC(X) is dense in<br />
L 2 (X), thisimpliesthatα k → α weakly in L 2 (X).<br />
For the last part, one can proceed using the abstract definition of L 2 (X) as a<br />
completion, but it is more transparent to use the explicit representation of Remark 1.9.<br />
For any π, wehaveα π = α X + ∑ E∈D π<br />
(α · α E )α E , where D π ⊂ D is the set of<br />
exceptional primes of π. Then(απ 2 ) = (α2 X ) − ∑ E∈D π<br />
(α · α E ) 2 , which is decreasing<br />
in π. It is then clear that α ∈ L 2 (X) if and only if (απ 2 ) is uniformly bounded from<br />
below and (α 2 ) = lim(απ 2 ).<br />
<br />
1.5. Positivity<br />
Recall that a class in H 1,1<br />
R<br />
(X) is pseudoeffective (psef) if it is the class of a closed<br />
positive (1, 1)-current on X. It is numerically effective (nef) if it is the limit of Kähler
DEGREE GROWTH OF MEROMORPHIC SURFACE MAPS 527<br />
classes. Any nef class is psef. The cone in H 1,1<br />
R<br />
(X) consisting of psef classes is strict:<br />
if α and −α are both psef, then α = 0.<br />
If π ′ = π ◦ µ is a blowup dominating some other blowup π,thenα ∈ H 1,1<br />
R (X π)<br />
is psef (nef) if and only if µ ∗ α ∈ H 1,1<br />
R<br />
(X π ′) is psef (nef ). On the other hand, if<br />
α ′ ∈ H 1,1<br />
R (X π ′) is psef (nef ), then so is µ ∗α ′ ∈ H 1,1<br />
R<br />
(X π). (For the nef part of the last<br />
assertion, it is important that we work in dimension two.)<br />
Definition 1.11<br />
AWeilclassα ∈ W (X) is psef (nef ) if its incarnation α π ∈ H 1,1<br />
R (X π) is psef (nef )<br />
for any blowup π : X π → X.<br />
We denote by Nef(X) ⊂ Psef(X) ⊂ W (X) the convex cones of nef and psef classes.<br />
The remarks above imply that a Cartier class α ∈ C(X) is psef (nef ) if and only if<br />
α π ∈ H 1,1<br />
R<br />
(X π) is psef (nef ) for one (or any) X π in which α is determined. We write<br />
α ≥ β as a shorthand for α − β ∈ W (X) being psef.<br />
PROPOSITION 1.12<br />
The nef cone Nef(X) and the psef cone Psef(X) are strict, closed, convex cones in<br />
W (X) with compact bases.<br />
Proof<br />
The nef (resp., psef ) cone is the projective limit of the nef (resp., psef ) cones of each<br />
H 1,1<br />
R<br />
(X π). These are strict, closed, convex cones with compact bases, so the result<br />
follows from the Tychonoff theorem.<br />
<br />
Nef classes satisfy the following monotonicity property.<br />
PROPOSITION 1.13<br />
If α ∈ W (X) is a nef Weil class, then α ≤ α π for each π. In particular, α π ≠ 0 for<br />
each π unless α = 0.<br />
Proof<br />
By induction on the number of blowups, it suffices to prove that α π ′ ≤ µ ∗ α π when<br />
π ′ = π ◦ µ and µ is the blowup of a point in X π .Butthenµ ∗ α π = α π ′ + cE, where<br />
E is the class of the exceptional divisor and c = (α π ′ · E) ≥ 0. To get the second<br />
point, note that α π = 0 for some π implies that α ≤ 0. On the other hand, α ≥ 0 as<br />
α is nef. Since Psef(X) is a strict cone, we infer that α = 0.<br />
<br />
PROPOSITION 1.14<br />
The nef cone Nef(X) is contained in L 2 (X).Ifα i ≥ β i , i = 1, 2, are nef classes, then<br />
we have (α 1 · α 2 ) ≥ (β 1 · β 2 ) ≥ 0.
528 BOUCKSOM, FAVRE, and JONSSON<br />
Proof<br />
If α ∈ W (X) is nef, each incarnation α π is nef, and thus (απ 2 ) ≥ 0,sothatα ∈ L2 (X)<br />
by Proposition 1.10 with (α 2 ) = inf π (απ 2 ) ≥ 0. To get the second point, note that<br />
(α 1 · α 2 ) ≥ (α 1 · β 2 ) since α 2 − β 2 is psef and α 1 is nef and, similarly, (α 1 · β 2 ) ≥<br />
(β 1 · β 2 ). <br />
These two propositions together show that if ω ∈ C(X) is a Cartier class determined<br />
by a Kähler class down on X,then(α · ω) > 0 for any nonzero nef class α ∈ W(X).<br />
PROPOSITION 1.15<br />
We have 2(α · β) α ≥ (α 2 ) β for any nef Weil classes α, β ∈ W(X). In particular, if<br />
ω ∈ C(X) is determined by a Kähler class on X normalized by (ω 2 ) = 1, we have,<br />
for any nonzero nef Weil class α,<br />
(α 2 )<br />
ω ≤ α ≤ 2(α · ω) ω. (1.1)<br />
2(α · ω)<br />
Proof<br />
The second assertion is a special case of the first one. To prove the first one, we may<br />
assume that (α · β) > 0,orelseα and β are proportional by the Hodge index theorem,<br />
and the result is clear. It is a known fact (see the remark after [B, Theorem 4.1]) that<br />
if γ ∈ C(X) is a Cartier class with (γ 2 ) ≥ 0, then either γ or −γ is psef. In view of<br />
Proposition 1.10, the same result is true for any γ ∈ L 2 (X). Apply this to γ = α − tβ,<br />
where t = ((α · α)/2(α · β)).As(γ · γ ) ≥ 0 and (γ · α) ≥ 0, γ must be psef. <br />
1.6. The canonical class<br />
The canonical class K X is the Weil class whose incarnation in any blowup X π is the<br />
canonical class K Xπ . It is not Cartier and does not even belong to L 2 (X). However,<br />
K Xπ ′ ≥ K Xπ whenever π ′ ≥ π, andK X is the smallest Weil class dominating all the<br />
K Xπ . This allows us to intersect K X with any nef Weil class α in a slightly ad hoc<br />
way: we set (α · K X ):= sup π (α π · K Xπ ) Xπ ∈ R ∪{+∞}.<br />
2. Functorial behavior<br />
Throughout this section, let F : X Y be a dominant meromorphic map between<br />
compact Kähler surfaces. Following [M, Section 34.7], we introduce the action of F<br />
on Weil and Cartier classes. We then describe the continuity properties of these actions<br />
on the Hilbert space L 2 (X).<br />
For each blowup Y ϖ of Y , there exists a blowup X π of X such that the induced<br />
map X π → Y ϖ is holomorphic. The associated pushforward H 1,1<br />
R (X π) → H 1,1<br />
R (Y ϖ )<br />
and pullback H 1,1<br />
R (Y ϖ ) → H 1,1<br />
R<br />
(X π) are compatible with the projective and injective<br />
systems defined by pushforwards and pullbacks that define Weil and Cartier classes,
DEGREE GROWTH OF MEROMORPHIC SURFACE MAPS 529<br />
respectively, so we can consider the induced morphisms on the respective projective<br />
and inductive limits.<br />
Definition 2.1<br />
Given F : X Y as above, we denote by F ∗ : W (X) → W(Y) the induced<br />
pushforward operator and by F ∗ : C(Y) → C(X) the induced pullback operator.<br />
Concretely, if α ∈ W (X) is a Weil class, the incarnation of F ∗ α ∈ W(Y) on a given<br />
blowup Y ϖ is the pushforward of α π ∈ H 1,1<br />
R<br />
(X π) by the induced map X π → Y ϖ for<br />
any π such that the latter map is holomorphic. Similarly, if β ∈ C(Y) is a Cartier class<br />
determined on a blowup Y ϖ , its pullback F ∗ β ∈ C(X) is the Cartier class determined<br />
on X π by the pullback of β ϖ ∈ H 1,1<br />
R<br />
(Y ϖ ) by the induced map X π → Y ϖ , whenever<br />
the latter is holomorphic.<br />
These constructions are functorial, that is, (F ◦ G) ∗ = F ∗ ◦ G ∗ and (F ◦ G) ∗ =<br />
G ∗ ◦ F ∗ , and are compatible with the duality between C and W since this is true for<br />
each holomorphic map X π → Y ϖ . In other words, for any α ∈ W(X) and β ∈ C(Y),<br />
we have (F ∗ α · β) = (α · F ∗ β).<br />
We also see that F ∗ preserves nef and psef Weil classes and that F ∗ preserves<br />
nef and psef Cartier classes. Indeed, the pullback and pushforward by a surjective<br />
holomorphic map both preserve nef and psef (1, 1)-classes in dimension two.<br />
Remark 2.2<br />
If π : X π → X and ϖ : Y ϖ → Y are arbitrary blowups, then the pullback operator<br />
H 1,1<br />
R (Y ϖ ) → H 1,1<br />
R (X π) usually associated to the meromorphic map X π Y ϖ is<br />
(Y ϖ ), followed by the projection<br />
of C(X) onto H 1,1<br />
R (X π). Similarly, the pushforward operator H 1,1<br />
R (X π) → H 1,1<br />
R (Y ϖ )<br />
usually associated to X π Y ϖ is given by the restriction of F ∗ : W(X) → W(Y)<br />
given by the restriction of F ∗ : C(Y) → C(X) to H 1,1<br />
R<br />
to H 1,1<br />
R (X π), followed by the projection of W (Y) onto H 1,1<br />
R (Y ϖ ).<br />
The intersection forms on C(X) and C(Y) are related by F ∗ as follows: (F ∗ β 2 ) =<br />
e(F )(β 2 ), where e(F ) > 0 is the topological degree of F . In view of the universal<br />
property of completions mentioned in Section 1.4, we get the following.<br />
PROPOSITION 2.3<br />
The pullback F ∗ : C(Y) → C(X) extends to a continuous operator F ∗ :L 2 (Y) →<br />
L 2 (X), so that ((F ∗ β) 2 ) = e(F )(β 2 ) for each β ∈ L 2 (Y). By duality, the pushforward<br />
F ∗ : W (X) → W (Y) induces a continuous operator F ∗ :L 2 (X) → L 2 (Y), so that<br />
(F ∗ α · β) = (α · F ∗ β) for any α, β ∈ L 2 (X).<br />
Next, we show that the pullback F ∗ : C(Y) → C(X) continuously extends to Weil<br />
classes and—dually—that the pushforward F ∗ : W (X) → W(Y) preserves Cartier<br />
classes.
530 BOUCKSOM, FAVRE, and JONSSON<br />
In doing so, we repeatedly use a consequence of the result of Zariski mentioned<br />
in Section 1. Namely, given F : X Y and a blowup π : X π → X, there exists<br />
ablowupY ϖ of Y such that the induced meromorphic map X π Y ϖ does not<br />
contract any curve to a point.<br />
LEMMA 2.4<br />
Suppose that π : X π → X and ϖ : Y ϖ → Y are two blowups such that the induced<br />
meromorphic map X π Y ϖ does not contract any curve to a point. Then for each<br />
Cartier class β ∈ C(Y), the incarnations of F ∗ β and F ∗ β ϖ on X π coincide.<br />
Proof<br />
Any Cartier class is a difference of nef Cartier classes, so we may assume that β is<br />
nef and determined in some blowup ϖ ′ dominating ϖ .Pickπ ′ dominating π so that<br />
the induced map X π ′ → Y ϖ ′ is holomorphic. Set α := F ∗ (β ϖ − β).Thenα ∈ C(X)<br />
is psef and determined in X π ′. We must show that α π = 0. Ifα π ≠ 0, thenα ≥ λC,<br />
where λ>0 and C is the class of an irreducible curve on X π .Now,C is not contracted<br />
by X π Y ϖ , so the incarnation of F ∗ α on Y ϖ is nonzero. But this is a contradiction<br />
since this incarnation equals e(F )(β ϖ − β) ϖ = 0.<br />
<br />
COROLLARY 2.5<br />
The pullback operator F ∗ : C(Y) → C(X) continuously extends to F ∗ : W(Y) →<br />
W (X) and preserves nef and psef Weil classes.<br />
More precisely, if X π is a given blowup of X and Y ϖ is a blowup of Y such that<br />
the induced meromorphic map X π Y ϖ does not contract curves, then for any Weil<br />
class γ ∈ W (Y), one has (F ∗ γ ) π = (F ∗ γ ϖ ) π .<br />
COROLLARY 2.6<br />
The pushforward operator F ∗ : W (X) → W (Y) preserves Cartier classes. More<br />
precisely, if α ∈ C(X) is a Cartier class determined on some X π ,thenF ∗ α is Cartier,<br />
determined on Y ϖ as soon as the induced meromorphic map X π Y ϖ does not<br />
contract curves.<br />
Proof<br />
For any β ∈ C(Y), the incarnations of F ∗ β and F ∗ β ϖ on X π coincide by Corollary<br />
2.5. Hence<br />
(F ∗ α · β) = (α · F ∗ β) = (α · F ∗ β ϖ ) = (F ∗ α · β ϖ ) = ( (F ∗ α) ϖ · β ) .<br />
As this holds for any Cartier class β ∈ C(Y), we must have F ∗ α = (F ∗ α) ϖ<br />
Proposition 1.7.<br />
by
DEGREE GROWTH OF MEROMORPHIC SURFACE MAPS 531<br />
3. Dynamics<br />
Now, consider a dominant meromorphic self-map F : X X of a compact Kähler<br />
surface X. Write λ 2 = e(F ) for the topological degree of F .Ifω ∈ Nef(X) is a nef<br />
Weil class such that (ω 2 ) > 0, we define the degree of F with respect to ω as<br />
deg ω (F ):= (F ∗ ω · ω) = (ω · F ∗ ω).<br />
This coincides with the usual notion of degree when X = P 2 and ω is the Cartier class<br />
determined by a line on P 2 .<br />
PROPOSITION 3.1<br />
The limit<br />
λ 1 := λ 1 (F ):= lim<br />
n→∞<br />
deg ω (F n ) 1/n (3.1)<br />
exists and does not depend on the choice of the nef class ω ∈ Nef(X) with (ω 2 ) > 0.<br />
Moreover, λ 1 is invariant under bimeromorphic conjugacy and λ 2 1 ≥ λ 2.<br />
The result above is well known, but we include the proof for completeness. We call<br />
λ 1 the asymptotic degree of F .Itisalsoknownasthefirst dynamical degree and can<br />
be computed (see [DF]) as λ 1 = lim n→∞ ρn<br />
1/n , where ρ n is the spectral radius of F n<br />
acting on H 1,1<br />
R<br />
(X) by pullback or pushforward (cf. Remark 2.2).<br />
Proof of Proposition 3.1<br />
Upon scaling ω, we can assume that (ω 2 ) = 1. By(1.1), we then have G ∗ ω ≤<br />
2(G ∗ ω · ω) ω for any dominant mapping G : X X. Applying this with G = F m<br />
yields<br />
deg ω (F n+m ) = (F n∗ F m∗ ω · ω) ≤ 2(F n∗ ω · ω)(F m∗ ω · ω) = 2deg ω (F n )deg ω (F m ).<br />
This implies (see, e.g., [KH, Proposition 9.6.4]) that the limit in (3.1) exists. Let us<br />
temporarily denote it by λ 1 (ω).Ifω ′ ∈ C(X) is another nef class with (ω ′ 2 ) > 0,then<br />
it follows from (1.1) thatω ′ ≤ Cω for some C>0. By Proposition 1.14, thisgives<br />
deg ω ′(F n ) = (F n∗ ω ′ · ω ′ ) ≤ C 2 (F n∗ ω · ω) = C 2 deg ω (F n ).<br />
Taking nth roots and letting n →∞shows that λ 1 (ω ′ ) ≤ λ 1 (ω), and thus λ 1 (ω ′ ) =<br />
λ 1 (ω) by symmetry, so that λ 1 is indeed independent of ω. It is then invariant by<br />
bimeromorphic conjugacy since X and all the spaces attached to it are.<br />
Finally, Proposition 1.14 yields F n∗ ω ≤ 2(F ∗n ω · ω) ω, which implies that<br />
e(F ) n = e(F n ) = (F n∗ ω 2 ) ≤ 4(F n∗ ω · ω) 2 = 4deg ω (F n ) 2 ,<br />
and letting n →∞yields λ 2 = e(F ) ≤ λ 2 1 .
532 BOUCKSOM, FAVRE, and JONSSON<br />
3.1. Existence of eigenclasses<br />
To begin, we do not assume that λ 2 1 >λ 2.<br />
THEOREM 3.2<br />
Let F : X X be any dominant meromorphic self-map of a smooth Kähler surface<br />
X with asymptotic degree λ 1 . Then we can find nonzero nef Weil classes θ ∗ and θ ∗<br />
with F ∗ θ ∗ = λ 1 θ ∗ and F ∗ θ ∗ = λ 1 θ ∗ .<br />
Note that by Proposition 1.14, both classes θ ∗ ,θ ∗ belong to L 2 (X).<br />
Proof<br />
We use the pushforward and pullback operators<br />
S π : H 1,1<br />
R (X π) → H 1,1<br />
R (X π) and T π : H 1,1<br />
R (X π) → H 1,1<br />
R<br />
(X π),<br />
usually associated to the meromorphic map X π X π induced by F for a given<br />
blowup π : X π → X. Thus S π (resp., T π ) is the restriction to H 1,1<br />
R (X π) of F ∗ :<br />
C(X) → C(X) (resp., F ∗ : C(X) → C(X)) followed by the projection C(X) →<br />
H 1,1<br />
R<br />
(X π) (cf. Remark 2.2). These operators are typically denoted F ∗ and F ∗ in the<br />
literature, but here that notation conflicts with the corresponding operators on C(X)<br />
or W (X).<br />
The spectral radius ρ π > 0 of T π can be computed as follows: if θ ∈ H 1,1<br />
R (X π)<br />
is any nef class with (θ 2 ) > 0, then(Tπ nθ · θ)1/n → ρ π as n →∞.<br />
LEMMA 3.3<br />
We have λ 1 ≤ ρ π ′ ≤ ρ π for all π ′ ≥ π.<br />
Proof<br />
Let θ ∈ C(X) be a given nef class determined on X π ′ with (θ 2 ) > 0,sothatθ ≤ θ π by<br />
Proposition 1.13.ThenT π ′θ is the incarnation on X π ′ of the nef class F ∗ θ on X π ′,and<br />
T π θ π is the incarnation on X π of the nef class F ∗ θ π ≥ F ∗ θ; thus F ∗ θ ≤ T π ′θ ≤ T π θ π<br />
holds by Proposition 1.13. By induction, we get F n∗ θ ≤ Tπ n ′θ ≤ T π nθ π for all n; hence<br />
(F n∗ θ ·θ) 1/n ≤ (Tπ n ′θ ·θ)1/n ≤ (Tπ nθ π ·θ π ) 1/n by Proposition 1.14,andλ 1 ≤ ρ π ′ ≤ ρ π<br />
follows by letting n →∞.<br />
<br />
Now, the set of nef classes in H 1,1<br />
R<br />
(X π) is a closed convex cone with compact basis<br />
invariant by T π ; thus a Perron-Frobenius-type argument (see [DF, Lemma 1.12])<br />
establishes the existence of a nonzero nef class θ(π) ∈ H 1,1<br />
R (X π) with T π θ(π) =<br />
ρ π θ(π).
DEGREE GROWTH OF MEROMORPHIC SURFACE MAPS 533<br />
If we identify θ(π) with the nef Cartier class that it determines, this says that the<br />
nef Cartier classes F ∗ θ(π) and ρ π θ(π) have the same incarnation on X π .Wehave<br />
thus obtained approximate eigenclasses, and now the plan is to get the desired class<br />
θ ∗ as a limit of classes of the form θ(π). We then explain how to modify the argument<br />
to construct θ ∗ .<br />
We normalize θ(π) by (θ(π) · ω) = 1 for a fixed class ω ∈ C(X) determined by<br />
aKähler class on X with (ω 2 ) = 1, so that the θ(π) all lie in a compact subset of the<br />
nef cone Nef(X) by Proposition 1.12.<br />
Let X = X 0 ← X 1 ← ··· be an infinite sequence of blowups so that the lift of<br />
F as a map from X n+1 to X n is holomorphic for n ≥ 0.<br />
For each n, letρ n denote the spectral radius of T n on H 1,1<br />
R<br />
(X n) as above,<br />
and pick a nonzero nef Cartier class θ n ∈ C(X) determined on X n and such that<br />
T n θ n = ρ n θ n .ThenF ∗ θ n is a Cartier class determined in X n+1 , and by definition,<br />
T n θ n is the incarnation of this class in X n . Therefore F ∗ θ n and ρ n θ n coincide on<br />
X n . By Proposition 1.6, it follows that F ∗ θ n − ρ n θ n converges to zero in W(X) as<br />
n →∞.<br />
We have seen above that ρ n is a decreasing sequence. Let ρ ∞ := lim ρ n ,sothat<br />
ρ ∞ ≥ λ 1 by Lemma 3.3. Since the θ n lie in a compact subset of Nef(X), we can find<br />
a cluster point θ ∗ for the sequence θ n , which is also a nef Weil class with (θ ∗ · ω) = 1.<br />
Since F ∗ θ n − ρ n θ n converges to zero in W (X), it follows that F ∗ θ ∗ = ρ ∞ θ ∗ .<br />
To complete the proof, we show that ρ ∞ = λ 1 . In fact, if α ∈ W(X) is any<br />
nonzero nef eigenclass of F ∗ with F ∗ α = tαfor some t ≥ 0,thent ≤ λ 1 .Indeed,we<br />
have α ≤ Cω for some C>0 by Proposition 1.15, and it follows that (F n∗ ω · ω) ≥<br />
C −1 (F n∗ α · ω) = C −1 t n (α · ω). Takingnth roots and letting n →∞yields λ 1 ≥ t.<br />
In order to construct θ ∗ , we modify the above argument as follows. Let S π :<br />
H 1,1<br />
R (X π) → H 1,1<br />
R<br />
(X π) be the pushforward operator defined above. As F ∗ and F ∗ are<br />
adjoint to each other with respect to the intersection pairing, it follows that S π and T π<br />
are adjoint with respect to Poincaré duality on H 1,1<br />
R<br />
(X π), so that they have the same<br />
spectral radius ρ π . By a Perron-Frobenius-type argument, there exists a nonzero nef<br />
class ϑ(π) ∈ H 1,1<br />
R<br />
(X π) such that S π ϑ(π) = ρ π ϑ(π).<br />
Now, pick X = X 0 ← X 1 ←··· to be an infinite sequence of blowups such that<br />
the lifts of F from X n to X n+1 do not contract any curves. For each n, we get a nef class<br />
ϑ n ∈ C(X) determined on X n normalized by (ϑ n · ω) = 1. By Corollary 2.6, the class<br />
F ∗ ϑ n is determined in X n+1 ,soF ∗ ϑ n and ρ n ϑ n coincide in X n . Proposition 1.6 then<br />
shows that F ∗ ϑ n − ρ n ϑ n converges to zero in W (X) as n →∞; hence θ ∗ ∈ Nef(X)<br />
can be taken to be any cluster value of ϑ n .<br />
<br />
Remark 3.4<br />
When K X is not psef (i.e., if X is rational or ruled) we may also achieve (θ ∗ ·K X ) ≤ 0.<br />
To see this, first note that F ∗ K X ≤ K X as classes in W(X) since K Xπ ′ − F ∗ K Xπ
534 BOUCKSOM, FAVRE, and JONSSON<br />
is represented by the effective zero divisor of the Jacobian determinant of the map<br />
X π ′ → X π induced by F , assuming that this is holomorphic. Now, for each blowup<br />
X π ,letC π be the set of nef classes α ∈ H 1,1<br />
R<br />
(X π) such that (α · K X ) ≤ 0.ThenC π is<br />
a closed convex cone with compact basis and is not reduced to zero since K X is not<br />
psef. It is, furthermore, invariant by S π . Indeed, if α ∈ H 1,1<br />
R<br />
(X π) is a nef class, we<br />
have<br />
(S π α · K X ) = (F ∗ α · K Xπ ) ≤ (F ∗ α · K X ) = (α · F ∗ K X ) ≤ (α · K X ).<br />
We can thus assume that the nonzero eigenclasses ϑ n in the proof of Theorem 3.2<br />
belong to C n , and we get (θ ∗ · K X ) ≤ 0.<br />
The same argument does not work for θ ∗ since F ∗ K X ≤ K X does not hold in<br />
general.<br />
3.2. Spectral properties<br />
Theorem 3.2 asserts the existence of eigenclasses for F ∗ and F ∗ with eigenvalue λ 1 .<br />
We now further analyze the spectral properties under the assumption that λ 2 1 >λ 2.<br />
THEOREM 3.5<br />
Assume that λ 2 1 >λ 2. Then the nonzero nef Weil classes θ ∗ ,θ ∗ ∈ L 2 (X) such that<br />
F ∗ θ ∗ = λ 1 θ ∗ and F ∗ θ ∗ = λ 1 θ ∗ are unique up to scaling. We have (θ ∗ · θ ∗ ) > 0 and<br />
(θ ∗2 ) = 0. We rescale them so that (θ ∗ · θ ∗ ) = 1. LetH ⊂ L 2 (X) be the orthogonal<br />
complement of θ ∗ and θ ∗ , so that we have the decomposition L 2 (X) = Rθ ∗ ⊕Rθ ∗ ⊕H.<br />
The intersection form is negative definite on H, and ‖α‖ 2 :=−(α 2 ) defines a Hilbert<br />
norm on H. The actions of F ∗ and F ∗ with respect to this decomposition are as<br />
follows.<br />
(i) The subspace H is F ∗ -invariant, and<br />
⎧<br />
F n∗ θ ∗ = λ n 1 θ ∗ ,<br />
( ⎪⎨ λ2<br />
) (<br />
nθ∗<br />
F n∗ θ ∗ = + (θ 2 ∗<br />
λ ) λn 1<br />
1 −<br />
1<br />
with h ⎪⎩<br />
n ∈ H, ‖h n ‖=O(λ n/2<br />
2 ),<br />
‖F n∗ h‖=λ n/2<br />
2 ‖h‖ for all h ∈ H.<br />
( λ2<br />
λ 2 1<br />
) n<br />
)<br />
θ ∗ + h n<br />
(ii) The subspace H is not F ∗ -invariant in general, but<br />
⎧<br />
F∗ ⎪⎨<br />
nθ ∗ = λ n 1 θ ∗,<br />
(<br />
F∗ nθ λ2<br />
) nθ ∗ =<br />
∗ ,<br />
λ 1<br />
⎪⎩ ‖F∗ nh‖≤Cλn/2<br />
2 ‖h‖ for some C>0 and all h ∈ H.
DEGREE GROWTH OF MEROMORPHIC SURFACE MAPS 535<br />
COROLLARY 3.6<br />
For any Weil class α ∈ L 2 (X), we have<br />
(<br />
1<br />
(λ2 ) ) n/2<br />
F n∗ α = (α · θ ∗ )θ ∗ + O<br />
λ n 1<br />
λ 2 1<br />
and<br />
(<br />
1<br />
(λ2 ) ) n/2<br />
F n ∗ α = (α · θ ∗ )θ ∗ + O .<br />
λ n 1<br />
λ 2 1<br />
Proof<br />
The decomposition of α in L 2 (X) = Rθ ∗ ⊕ Rθ ∗ ⊕ H is given by<br />
α = ( (α · θ ∗ ) − (α · θ ∗ )(θ 2 ∗ )) θ ∗ + (α · θ ∗ )θ ∗ + α 0 , (3.2)<br />
where α 0 ∈ H. The result follows from (3.2) using Theorem 3.5(i), (ii).<br />
<br />
Proof of the main theorem<br />
Applying Corollary 3.6 to α = ω (which is nef and hence in L 2 (X))gives<br />
deg ω (F n ) = (F n∗ ω · ω) = (ω · θ ∗ )(ω · θ ∗ )λ n 1 + O(λn/2 2 ).<br />
This completes the proof with b := (ω · θ ∗ )(ω · θ ∗ ).<br />
<br />
Proof of Theorem 3.5<br />
Using Theorem 3.2, we may find nonzero nef Weil classes θ ∗ ,θ ∗ such that F ∗ θ ∗ = λ 1 θ ∗<br />
and F ∗ θ ∗ = λ 1 θ ∗ . Fix two such classes for the duration of the proof. In the end, we<br />
see that they are unique up to scaling.<br />
The proof amounts to a series of simple arguments using general facts for transformations<br />
of a complete vector space endowed with a Minkowski form. We provide<br />
the details for the benefit of the reader.<br />
First, note that λ 1 F ∗ θ ∗ = F ∗ F ∗ θ ∗ = λ 2 θ ∗ ,sothatF ∗ θ ∗ = (λ 2 /λ 1 )θ ∗ .Since<br />
F ∗ θ ∗ = λ 1 θ ∗ and λ 2 1 >λ 2, it follows that θ ∗ and θ ∗ cannot be proportional.<br />
Applying the relation (F ∗ α 2 ) = λ 2 (α 2 ) to α = θ ∗ yields λ 2 1 (θ ∗2 ) = λ 2 (θ ∗2 ),and<br />
thus (θ ∗2 ) = 0 since λ 2 1 >λ 2. By the Hodge index theorem, θ ∗ and θ ∗ would thus<br />
have to be proportional if they were orthogonal. We infer that (θ ∗ · θ ∗ ) > 0, andwe<br />
rescale θ ∗ so that (θ ∗ · θ ∗ ) = 1.<br />
Let us first prove the properties in (i) for the pullback. As both θ ∗ and θ ∗ are<br />
eigenvectors for F ∗ , the space H is invariant under F ∗ . Using (3.2) and the invariance<br />
properties of θ ∗ and θ ∗ ,weget<br />
F ∗ θ ∗ = λ 2<br />
λ 1<br />
θ ∗ + λ 1<br />
(<br />
1 − λ 2<br />
λ 2 1<br />
)<br />
(θ 2 ∗ )θ ∗ + h 1 , (3.3)
536 BOUCKSOM, FAVRE, and JONSSON<br />
where h 1 ∈ H. Inductively, (3.3)gives<br />
F n∗ θ ∗ =<br />
( λ2<br />
λ 1<br />
) nθ∗<br />
+ λ n 1<br />
(<br />
1 −<br />
( λ2<br />
λ 2 1<br />
) n<br />
)<br />
(θ 2 ∗ )θ ∗ + h n , (3.4)<br />
where h n+1 = F ∗ h n + (λ 2 /λ 1 ) n h 1 ∈ H. Using the fact that ‖F ∗ h‖ 2 = λ 2 ‖h‖ 2 on<br />
H, weget‖h n+1 ‖≤λ 1/2<br />
2 ‖h n ‖+(λ 2 /λ 1 ) n ‖h 1 ‖, which is easily seen to imply that<br />
‖h n ‖=O(λ n/2<br />
2 ) since ∑ k (λ1/2 2 /λ 1 ) k < +∞. This concludes the proof of (i).<br />
Let us now turn to the pushforward operator. The first two equations are clear. As<br />
θ ∗ may not be an eigenvector for F ∗ , H need not be invariant by F ∗ , but since F ∗ h is<br />
orthogonal to θ ∗ for any h ∈ H, we can write F∗ nh = a nθ ∗ +g n with a n = (F n∗ θ ∗ ·h)<br />
and g n ∈ H. We have seen that F n∗ θ ∗ = h n modulo θ ∗ ,θ ∗ with ‖h n ‖=O(λ n/2<br />
2 );<br />
thus |a n |=|(h n · h)| ≤Cλ n/2<br />
2 ‖h‖. On the other hand, we have (gn 2) = (F n∗ g n · h);<br />
thus ‖g n ‖ 2 ≤ λ n/2<br />
2 ‖g n ‖‖h‖, and this shows that ‖F∗ nh‖≤Cλn/2<br />
2 ‖h‖. <br />
Remark 3.7<br />
It follows from the proof of the main theorem that there exist nef classes α ∗ ,α ∗ ∈<br />
H 1,1<br />
R<br />
(X) such that for any Kähler classes ω, ω′ on X, wehave<br />
(<br />
deg ω (F n )<br />
deg ω ′(F n ) = (α∗ · ω) X (α ∗ · ω) (λ2 ) ) n/2 X<br />
+ O .<br />
(α ∗ · ω ′ ) X (α ∗ · ω ′ ) X<br />
Indeed, we can take α ∗ and α ∗ as the incarnations in X of θ ∗ and θ ∗ , respectively.<br />
Remark 3.8<br />
When F is bimeromorphic, we have θ ∗ (F ) = θ ∗ (F −1 ); hence (θ∗ 2 ) = 0. However,in<br />
general, we may have (θ∗ 2 ) > 0. For example, let F be any polynomial map of C2<br />
whose extension to P 2 is not holomorphic but does not contract any curve. If ω is the<br />
class of a line on P 2 ,thendeg ω (F ) > √ λ 2 > 1. On the other hand, F ∗ ω = deg ω (F )ω<br />
by Corollary 2.6,soλ 1 = deg ω (F ), θ ∗ = ω and (θ∗ 2) = 1.<br />
Remark 3.9<br />
The case when θ ∗ (or θ ∗ ) is Cartier is very special. For example, when F is bimeromorphic,<br />
it follows from [DF, Theorem 0.4] that θ ∗ (or, equivalently, θ ∗ ) is Cartier if<br />
and only if F is biholomorphic in some birational model. In the general noninvertible<br />
case, similar rigidity results are expected (see [C1] for work in this direction).<br />
Note also that F being algebraically stable in some birational model does not<br />
imply that the eigenclasses are Cartier. We do not know whether having a Cartier<br />
eigenclass implies algebraic stability in some model, but having a Cartier eigenclass<br />
has many of the same consequences as stability: λ 1 is an algebraic integer, and the<br />
sequence of degrees (deg ω F n ) ∞ 1 satisfies a linear recurrence relation.<br />
λ 2 1
DEGREE GROWTH OF MEROMORPHIC SURFACE MAPS 537<br />
Acknowledgments. We thank Serge Cantat and Jeff Diller for many useful remarks<br />
and the referees for careful readings of the article.<br />
References<br />
[BHPV] W. P. BARTH, K. HULEK, C. A. M. PETERS,andA. VAN DE VEN, Compact Complex<br />
Surfaces, 2nd ed., Ergeb. Math. Grenzgeb. (3) 4, Springer, Berlin, 2004.<br />
MR 2030225 521<br />
[BK1] E. BED<strong>FOR</strong>D and K. H. KIM, On the degree growth of birational mappings in higher<br />
dimension, J. Geom. Anal. 14 (2004), 567 – 596. MR 2111418 520<br />
[BK2] ———, Periodicities in linear fractional recurrences: Degree growth of birational<br />
surface maps, Michigan Math. J. 54 (2006), 647 – 670. MR 2280499 520<br />
[BF] A. M. BONIFANT and J. E. <strong>FOR</strong>NAESS, Growth of degree for iterates of rational maps in<br />
several variables, Indiana Univ. Math. J. 49 (2000), 751 – 778. MR 1793690<br />
520<br />
[B] S. BOUCKSOM, Divisorial Zariski decompositions on compact complex manifolds,<br />
Ann. Sci. École Norm. Sup. (4) 37 (2004), 45 – 76. MR 2050205 528<br />
[BFJ] S. BOUCKSOM, C. FAVRE,andM. JONSSON, Differentiability of volumes of divisors<br />
and a problem of Teissier, to appear in J. Algebraic Geom., preprint,<br />
arXiv:math/0608260v2 [math.AG] 521<br />
[C1] S. CANTAT, Caractérisation des exemples de Lattès et de Kummer, preprint, 2006.<br />
521, 536<br />
[C2] ———, Sur les groupes de transformations birationnelles des surfaces, preprint, 2007.<br />
520<br />
[DF] J. DILLER and C. FAVRE, Dynamics of bimeromorphic maps of surfaces,Amer.J.<br />
Math. 123 (2001), 1135 – 1169. MR 1867314 520, 521, 531, 532, 536<br />
[DS] T.-C. DINH and N. SIBONY, Une borne supérieure pour l’entropie topologique d’une<br />
application rationnelle, Ann. of Math. (2) 161 (2005), 1637 – 1644.<br />
MR 2180409 520<br />
[F] C. FAVRE, Les applications monomiales en deux dimensions, Michigan Math. J. 51<br />
(2003), 467 – 475. MR 2021001 520<br />
[FJ1] C. FAVRE and M. JONSSON, The Valuative Tree, Lecture Notes in Math. 1853, Springer,<br />
Berlin, 2004. MR 2097722 521, 524<br />
[FJ2] ———, Valuations and multiplier ideals,J.Amer.Math.Soc.18 (2005), 655 – 684.<br />
MR 2138140 521<br />
[FJ3] ———, Valuative analysis of planar plurisubharmonic functions, Invent. Math. 162<br />
(2005), 271 – 311. MR 2199007 521<br />
[FJ4] ———, Eigenvaluations, Ann. Sci. École Norm. Sup. (4) 40 (2007), 309 – 349.<br />
MR 2339287 521<br />
[Fo] G. B. FOLLAND, Real Analysis: Modern Techniques and Their Applications, 2nd ed.,<br />
Pure Appl. Math. (N.Y.), Wiley, New York, 1999. MR 1681462 522<br />
[FS] J.-E. <strong>FOR</strong>NAESS and N. SIBONY, “Complex dynamics in higher dimension, II” in<br />
Modern Methods in Complex Analysis (Princeton, 1992), Ann. of Math. Stud.<br />
137, Princeton Univ. Press, Princeton, 1995, 135 – 182. MR 1369137 520<br />
[Fr] S. FRIEDLAND, Entropy of polynomial and rational maps, Ann. of Math. (2) 133<br />
(1991), 359 – 368. MR 1097242 520
538 BOUCKSOM, FAVRE, and JONSSON<br />
[G1] V. GUEDJ, Entropie topologique des applications méromorphes, Ergodic Theory<br />
Dynam. Systems 25 (2005), 1847 – 1855. MR 2183297 520<br />
[G2] ———, Ergodic properties of rational mappings with large topological degree,<br />
Ann. of Math. (2) 161 (2005), 1589 – 1607. MR 2179389 520<br />
[HP] J. HUBBARD and P. PAPADOPOL, Newton’s method applied to two quadratic equations<br />
in C 2 viewed as a global dynamical system, to appear in Mem. Amer. Math. Soc.<br />
191 (2008), no. 891, preprint, 2000. 520<br />
[KH] A. KATOK and B. HASSELBLATT, Introduction to the Modern Theory of Dynamical<br />
Systems, Encyclopedia Math. Appl. 54, Cambridge Univ. Press, Cambridge, 1995.<br />
MR 1326374 531<br />
[Ko] J. KOLLÁR, “Singularities of pairs” in Algebraic Geometry (Santa Cruz, Calif., 1995),<br />
Proc. Symp. Pure Math. 62, Part 1, Amer. Math. Soc., Providence, 1997,<br />
221 – 287. MR 1492525 524<br />
[L] R. LAZARSFELD, Positivity in Algebraic Geometry, I: Classical Setting: Line Bundles<br />
and Linear Series, Ergeb. Math. Grenzgeb. (3) 48, Springer, Berlin, 2004.<br />
MR 2095471 521<br />
[M] Y. MANIN, Cubic Forms: Algebra, Geometry, Arithmetic, 2nd ed., North-Holland Math.<br />
Lib. 4, North-Holland, Amsterdam, 1986. MR 0833513 520, 521, 523, 528<br />
[N] V.-A. NGUYEN, Algebraic degrees for iterates of meromorphic self-maps of P k , Publ.<br />
Mat. 50 (2006), 457 – 473. 520<br />
[RS] A. RUSSAKOVSKII and B. SHIFFMAN, Value distribution for sequences of rational<br />
mappings and complex dynamics, Indiana Univ. Math. J. 46 (1997), 897 – 932.<br />
MR 1488341 520<br />
[S] N. SIBONY, “Dynamique des applications rationnelles de P k ”inDynamique et<br />
géométrie complexes (Lyon, France, 1997), Panor. Synthèses 8, Soc. Math.<br />
France, Montrouge, 1999, 97 – 185. MR 1760844 520<br />
[V] M. VAQUIÉ, “Valuations” in Resolution of Singularities (Obergurgl, Austria, 1997),<br />
Progr. Math. 181, Birkhäuser, Basel, 2000, 539 – 590. MR 1748635 522<br />
[ZS] O. ZARISKI and P. SAMUEL, Commutative Algebra, Vol. II, reprint of the 1960 ed.,<br />
Grad. Texts in Math. 29, Springer, New York, 1975. MR 0389876 522<br />
Boucksom<br />
Institut de Mathématiques, CNRS-Université Paris 7, F-75251 Paris CEDEX 05, France;<br />
boucksom@math.jussieu.fr<br />
Favre<br />
Institut de Mathématiques, CNRS-Université Paris 7, F-75251 Paris CEDEX 05, France;<br />
favre@math.jussieu.fr<br />
Jonsson<br />
Department of Mathematics, University of Michigan, Ann Arbor, Michigan 48109-1109, USA;<br />
mattiasj@umich.edu
DISTORTION OF HAUSDORFF MEASURES AND<br />
IMPROVED PAINLEVÉ REMOVABILITY <strong>FOR</strong><br />
QUASIREGULAR MAPPINGS<br />
K. ASTALA, A. CLOP, J. MATEU, J. OROBITG, and I. URIARTE-TUERO<br />
Abstract<br />
The classical Painlevé theorem tells us that sets of zero length are removable for<br />
bounded analytic functions, while (some) sets of positive length are not. For general<br />
K-quasiregular mappings in planar domains, the corresponding critical dimension is<br />
2/(K + 1). We show that when K>1, unexpectedly one has improved removability.<br />
More precisely, we prove that sets E of σ -finite Hausdorff (2/(K + 1))-measure are<br />
removable for bounded K-quasiregular mappings. On the other hand, dim(E) =<br />
2/(K + 1) is not enough to guarantee this property.<br />
We also study absolute continuity properties of pullbacks of Hausdorff measures<br />
under K-quasiconformal mappings: in particular, at the relevant dimensions<br />
1 and 2/(K + 1). For general Hausdorff measures H t , 0 < t < 2, we reduce<br />
the absolute continuity properties to an open question on conformal mappings (see<br />
Conjecture 2.3).<br />
1. Introduction<br />
A homeomorphism φ : → ′ between planar domains , ′ ⊂ C is called K-<br />
quasiconformal if it belongs to the Sobolev space W 1,2<br />
loc () and satisfies the distortion<br />
inequality<br />
max |∂ αφ| ≤K min|∂ α φ| a.e. in . (1.1)<br />
α<br />
α<br />
It has been known since the work of Ahlfors [3] that quasiconformal mappings preserve<br />
sets of zero Lebesgue measure. It is also well known that they preserve sets of zero<br />
Hausdorff dimension since K-quasiconformal mappings are Hölder continuous with<br />
exponent 1/K (see Mori [22]). However, these maps do not preserve Hausdorff<br />
DUKE MATHEMATICAL JOURNAL<br />
Vol. 141, No. 3, c○ 2008 DOI 10.1215/00127094-2007-005<br />
Received 28 July 2006. Revision received 13 April 2007.<br />
2000 Mathematics Subject Classification. Primary 30C62; Secondary 35J15, 35J70.<br />
Astala supported in part by Academy of Finland projects 106257, 110641, and 211485.<br />
Clop supported in part by European Union project Conformal Structures and Dynamics (CODY).<br />
Clop, Mateu, and Orobitg supported in part by projects MTM2004-00519 (Spain), Acción Integrada HF2004-<br />
0208 (Spain), and 2005-SGR-00774 (Generalitat de Catalunya).<br />
Uriarte-Tuero supported by Academy of Finland projects 209371 and 203949.<br />
539
540 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
dimension in general, and it was in the article of Astala [4] where the precise bounds<br />
for the distortion of dimension were given. For any compact set E with dimension t<br />
and for any K-quasiconformal mapping φ, wehave<br />
1<br />
( 1<br />
K t − 1 )<br />
≤<br />
2<br />
1<br />
dim(φ(E)) − 1 ( 1<br />
2 ≤ K t − 1 )<br />
. (1.2)<br />
2<br />
Furthermore, these bounds are optimal (i.e., equality may occur in either estimate).<br />
The fundamental question that we study in this article is whether the estimates<br />
(1.2) can be improved to the level of Hausdorff measures H t . In other words, if φ is<br />
aplanarK-quasiconformal mapping, 0
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 541<br />
The constant C K depends only on K if h is normalized at infinity requiring h(z) =<br />
z + O(1/z). For the area, the corresponding estimate was shown in [4]. In fact, as we<br />
see later, a counterpart of (1.5)forthet-dimensional Hausdorff content M t is the only<br />
missing detail for proving the absolute continuity φ ∗ H t ′<br />
≪ H t for general t. Toward<br />
solving (1.3), we conjecture that actually,<br />
M t( h(E) ) ≤ C M t (E), 0
542 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
mapping f : \ E → C admits a K-quasiregular extension to . In this definition,<br />
as in the analytic setting, we may replace L ∞ () by BMO() to get a close variant<br />
of the problem.<br />
The sharpness of the bounds in equation (1.2) determines the index 2/(K + 1)<br />
as the critical dimension in both the L ∞ and the BMO quasiregular removability<br />
problems. In fact, Iwaniec and Martin [15] previously conjectured that in R n , n ≥<br />
2, sets with Hausdorff measure H n/(K+1) (E) = 0 are removable for bounded K-<br />
quasiregular mappings. A preliminary positive answer for n = 2 was described in [6].<br />
Generalizing this, in the present article, we show that, surprisingly, for K>1, one<br />
can do even better. We have the following improved Painlevé removability.<br />
THEOREM 1.2<br />
Let K>1, and suppose that E is any compact set with<br />
H 2/(K+1) (E) σ -finite.<br />
Then E is removable for all bounded K-quasiregular mappings.<br />
The theorem fails for K = 1 since, for instance, the line segment E = [0, 1] is not<br />
removable for bounded analytic functions.<br />
For the converse direction, the article [4] finds for every t>2/(K + 1) non-Kremovable<br />
sets with dim(E) = t. We also make an improvement here and construct<br />
compact sets with dimension precisely equal to 2/(K + 1) yet not removable for some<br />
bounded K-quasiregular mappings (for details, see Theorem 5.1).<br />
Theorems 1.1 and 1.2 are closely connected via the classical Stoïlow factorization,<br />
which tells (see [6], [18]) that in planar domains, K-quasiregular mappings are<br />
precisely the maps f representable in the form f = h ◦ φ, where h is analytic and φ<br />
is K-quasiconformal. Indeed, the first step in proving Theorem 1.2 is to show that for<br />
a general K-quasiconformal mapping φ, one has<br />
H 2/(K+1) (E) σ -finite ⇒ H 1( φ(E) ) σ -finite.<br />
However, this conclusion is not enough since there are rectifiable sets of finite length<br />
(such as E = [0, 1]) which are nonremovable for bounded analytic functions. Therefore,<br />
in addition, we need to establish that such “good” sets of positive analytic capacity<br />
actually behave better also under quasiconformal mappings. That is, we show that up<br />
to a set of zero length,<br />
F 1-rectifiable ⇒ dim ( φ −1 (F ) ) > 2<br />
K + 1<br />
(for details and a precise formulation, see Corollary 3.2).
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 543<br />
The article is structured as follows. In Section 2, we deal with the quasiconformal<br />
distortion of Hausdorff measures and of other set functions. In Section 3, we study the<br />
quasiconformal distortion of 1-rectifiable sets. Section 4 gives the proof for the improved<br />
Painlevé removability theorem for K-quasiregular mappings and other related<br />
questions. Finally, in Section 5, we describe a construction of nonremovable sets.<br />
2. Absolute continuity<br />
There are several natural ways to normalize the quasiconformal mappings φ : C → C.<br />
In this article, we mostly use the principal K-quasiconformal mappings (i.e., mappings<br />
that are conformal outside a compact set and are normalized by φ(z) − z = O(1/|z|)<br />
as |z| →∞).<br />
It is shown in Astala’s article [4] that for all K-quasiconformal mappings φ :<br />
C → C,<br />
|φ(E)| ≤C |E| 1/K , (2.1)<br />
where C is a constant that depends on the normalizations. By scaling, we may always<br />
arrange<br />
diam ( φ(E) ) = diam(E) ≤ 1, (2.2)<br />
and then C = C(K) depends only on K. In order to achieve the result (2.1), one<br />
first reduces to the case where the set E is a finite union of disks. Second, applying<br />
Stoïlow factorization methods, the mapping φ is written as φ = h ◦ φ 1 , where both<br />
h, φ 1 : C → C are K-quasiconformal mappings, so that φ 1 is conformal on E and<br />
h is conformal in the complement of the set F = φ 1 (E). Here, one obtains the right<br />
conclusion for φ 1 ,<br />
|φ 1 (E)| ≤C |E| 1/K ,<br />
by including φ 1 in a holomorphic family of quasiconformal mappings. Further, one<br />
shows in [4, page 50] that under the special assumption where h is conformal outside<br />
of F ,wehave<br />
|h(F )| ≤C |F |, (2.3)<br />
where the constant C still depends only on K.<br />
In searching for absolute continuity properties of other Hausdorff measures under<br />
quasiconformal mappings, such a decomposition seems unavoidable, and this leads<br />
one to look for counterparts of (2.3) for Hausdorff measures H t or Hausdorff contents<br />
M t . Here, we establish the result for the dimension t = 1.
544 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
LEMMA 2.1<br />
Suppose that E ⊂ C is a compact set, and let φ : C → C be a principal K-<br />
quasiconformal mapping, such that φ is conformal on C \ E. Then<br />
with constants depending only on K.<br />
M 1( φ(E) ) ≃ M 1 (E)<br />
In order to prove this result, some background is needed. The space of functions<br />
of bounded mean oscillation, BMO, is invariant under quasiconformal changes of<br />
variables (see [26]). More precisely, if φ is a K-quasiconformal mapping and f ∈<br />
BMO(C), thenf ◦ φ ∈ BMO(C) with BMO-norm<br />
‖f ◦ φ‖ ∗ ≤ C(K) ‖f ‖ ∗ .<br />
The space BMO(C) gives rise to a capacity,<br />
γ 0 (F ) = sup|f ′ (∞)|,<br />
where the supremum runs over all functions f ∈ BMO(C) with ‖f ‖ ∗ ≤ 1<br />
which are holomorphic on C \ F and satisfy f (∞) = 0. Here, f ′ (∞) =<br />
lim |z|→∞ z (f (z) − f (∞)). Observe that in this situation, ∂f defines a distribution<br />
supported on F , and actually, |〈∂f,1〉| = |f ′ (∞)|. It turns out (see [32]) that for any<br />
compact set E, wehave<br />
γ 0 (E) ≃ M 1 (E). (2.4)<br />
According to result of Král [17], in the class of functions f ∈ BMO(C) holomorphic<br />
on C \ E every f admits a holomorphic extension to the whole plane if and only if<br />
M 1 (E) = 0 (i.e., γ 0 characterizes those compact sets that are removable for BMO<br />
holomorphic functions). Because of these equivalences, to prove Lemma 2.1, it suffices<br />
to show that γ 0 (φ(E)) ≃ γ 0 (E).<br />
Proof<br />
Suppose that f ∈ BMO(C) is a holomorphic mapping of C \ E such that ‖f ‖ ∗ ≤ 1<br />
and f (∞) = 0. Then the function g = f ◦ φ −1 is in BMO(C), and‖g‖ ∗ ≤ C(K).<br />
On the other hand, g is holomorphic on C \ φ(E), and since φ is a principal K-<br />
quasiconformal mapping, g(∞) = 0,and<br />
|g ′ (∞)| = lim |zg(z)| = lim |φ(w) f (w)| =|f ′ (∞)|.<br />
|z|→∞ |w|→∞<br />
Hence γ 0 (E) ≤ C(K) γ 0 (φ(E)). The converse inequality follows by symmetry since<br />
the inverse φ −1 is also a principal mapping.
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 545<br />
Lemma 2.1 is a first step toward the results on absolute continuity, as presented in the<br />
following reformulation of Theorem 1.1.<br />
THEOREM 2.2<br />
Let E be a compact set, and let φ : C → C be K-quasiconformal, normalized by<br />
(2.2). Then<br />
M 1( φ(E) ) ≤ C ( M 2/(K+1) (E) ) (K+1)/(2K)<br />
,<br />
where the constant C = C(K) depends only on K. In particular, if H 2/(K+1) (E) = 0,<br />
then H 1( φ(E) ) = 0.<br />
Proof<br />
There is no restriction if we assume that E ⊂ D. We can also assume that φ is a<br />
principal K-quasiconformal mapping, conformal outside D.Now,sinceE is compact,<br />
for any ε>0 there is a finite covering of E by open disks D j , j = 1,...,m, such<br />
that<br />
n∑<br />
j=1<br />
r 2/(K+1)<br />
j ≤ M 2/(K+1) (E) + ε.<br />
By Vitali’s covering lemma, we can replace our covering by a new finite family of<br />
disjoint disks, also denoted D j = D(z j ,r j ), j = 1,...,m, such that E is contained<br />
in the union of 5D j = D(z j , 5r j ). Denote now = ⋃ n<br />
j=1 5D j.Asin[4], we use<br />
a decomposition φ = h ◦ φ 1 , where both φ 1 and h are principal K-quasiconformal<br />
mappings. Moreover, we may require that φ 1 be conformal in ∪ (C \ D) and that h<br />
be conformal outside φ 1 ().<br />
By Lemma 2.1, we see that<br />
M 1( φ(E) ) ≤ M 1( φ() ) = M 1( h ◦ φ 1 () ) ≤ C M 1( φ 1 () ) .<br />
Hence the problem has been reduced to estimating M 1 (φ 1 ()). For this, K-quasidisks<br />
have area comparable to the square of the diameter,<br />
diam ( φ 1 (5D j ) ) ≃ diam ( φ 1 (D j ) ) ( ∫ 1/2<br />
≃|φ 1 (D j )| 1/2 = J (z, φ 1 ) dA(z))<br />
D j<br />
with constants that depend only on K. Thus, using Hölder estimates twice, we obtain<br />
n∑<br />
diam ( φ 1 (5D j ) ) n∑<br />
≃ diam ( φ 1 (D j ) ) ( n∑<br />
∫<br />
≤ C(K) J (z, φ 1 ) p dA(z)<br />
D j<br />
j=1<br />
j=1<br />
j=1<br />
( n∑<br />
) 1−1/(2p),<br />
× |D j | (p−1)/(2p−1)<br />
j=1<br />
) 1/(2p)
546 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
as long as J (z, φ 1 ) p is integrable. But here we are in the special situation of [7, Lemma<br />
5.2]. Namely, as φ 1 is conformal in the subset , we may take p = K/(K − 1) and<br />
apply [7] to obtain<br />
n∑<br />
∫<br />
∫<br />
J (z, φ 1 ) p dA(z) ≤<br />
D j<br />
j=1<br />
<br />
J (z, φ 1 ) p dA(z) ≤ π.<br />
With the above choice of p, one has (p − 1)/(2p − 1) = 1/(K + 1). Hence we get<br />
n∑<br />
diam ( φ 1 (5D j ) ) ( n∑ ) (K+1)/(2K)<br />
≤ C(K) r 2/(K+1)<br />
j<br />
j=1<br />
j=1<br />
≤ C(K) ( M 2/(K+1) (E) + ε ) (K+1)/(2K)<br />
. (2.5)<br />
But ⋃ j φ 1(5D j ) is a covering of φ 1 (), so that actually, we have<br />
M 1( φ(E) ) ≤ CM 1( φ 1 () ) ≤ C(K) ( M 2/(K+1) (E) + ε ) (K+1)/(2K)<br />
.<br />
Since this holds for every ε>0, the result follows.<br />
<br />
At this point, we emphasize that for a general quasiconformal mapping φ, wehave<br />
J (z, φ) ∈ L p loc only for p < K/(K − 1). The improved borderline integrability<br />
(p = K/(K − 1)) under the extra assumption that φ | is conformal was shown in<br />
[7, Lemma 5.2]. This phenomenon is crucial for our argument since we are studying<br />
Hausdorff measures rather than dimension. Actually, the same procedure shows that<br />
inequality (2.5) works in a much more general setting. That is, still under the special<br />
assumption that φ 1 is conformal in ⋃ n<br />
j=1 D j,wehaveforanyt ∈ [0, 2],<br />
( n∑<br />
diam ( φ 1 (D j ) ) ) d<br />
1/d ( n∑ ) (1/t)(1/K),<br />
≤ C(K) diam(D j ) t (2.6)<br />
j=1<br />
where d = 2Kt/(2 + (K − 1)t). On the other hand, another key point in our proof is<br />
the estimate<br />
j=1<br />
M 1( h(E) ) ≤ C M 1 (E),<br />
valid whenever h is a principal K-quasiconformal mapping that is conformal outside<br />
E. We believe that finding the counterpart to this estimate is crucial for<br />
understanding distortion of Hausdorff measures under quasiconformal mappings. We<br />
make the following conjecture.
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 547<br />
CONJECTURE 2.3<br />
Suppose that we are given a real number d ∈ (0, 2]. Then for any compact set E ⊂ C<br />
and for any principal K-quasiconformal mapping h which is conformal on C \ E,we<br />
have<br />
with constants that depend only on K and d.<br />
M d( h(E) ) ≃ M d (E) (2.7)<br />
One may also formulate a convenient discrete variant, which is actually stronger than<br />
Conjecture 2.3.<br />
Question 2.4<br />
Suppose that we are given a real number d ∈ (0, 2] and a finite number of disjoint<br />
disks D 1 ,...,D n . If a mapping h is conformal on C \ ⋃ n<br />
j=1 D j and admits a K-<br />
quasiconformal extension to C, isitthentruethat<br />
n∑<br />
diam ( h(D j ) ) n∑<br />
d<br />
≃ diam(D j ) d (2.8)<br />
j=1<br />
with constants that depend only on K and d?<br />
We already know that (2.7)istrueford = 1 and d = 2; however, for Question 2.4,we<br />
know a proof only at d = 2. An affirmative answer to Conjecture 2.3, combined with<br />
the optimal integrability bound proving (2.6), would provide the absolute continuity of<br />
φ ∗ H d with respect to H t , where d = 2Kt/(2 + (K − 1)t), 0 ≤ t ≤ 2, andK ≥ 1.<br />
Therefore (2.7) would have important consequences in the theory of quasiconformal<br />
mappings.<br />
The positive answer to (2.7) for the dimension d = 1 was based on the equivalence<br />
(2.4) and the invariance of BMO. Actually, more is true. The space VMO, equal to<br />
the BMO-closure of uniformly continuous functions, is quasiconformally invariant<br />
as well. We may also describe VMO, vanishing mean oscillation, as consisting of<br />
functions f ∈ BMO for which<br />
lim 1<br />
|B|<br />
∫<br />
B<br />
j=1<br />
|f − f B |=0,<br />
as |B| +1/|B| → ∞. As we now see, the invariance of VMO has interesting<br />
consequences.
548 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
THEOREM 2.5<br />
Let E ⊂ C be a compact set, and let φ : C → C be a K-quasiconformal mapping. If<br />
H 2/(K+1) (E) is finite (or even σ -finite), then H 1 (φ(E)) is σ -finite.<br />
This result may be equivalently expressed in terms of the lower Hausdorff content.<br />
To understand this alternative formulation of Theorem 2.5, we first need some background.<br />
A measure function is a continuous nondecreasing function h(t), t ≥ 0, such<br />
that lim t→0 h(t) = 0.Ifh is a measure function and F ⊂ C, weset<br />
M h (F ) = inf ∑ j<br />
h(δ j ),<br />
where the infimum is taken over all countable coverings of F by disks of diameter δ j .<br />
When h(t) = t α , α>0,thenM h (F ) = M α (F ) equals the α-dimensional Hausdorff<br />
content of F . Moreover, the content M α and the measure H α have the same zero sets.<br />
We denote by F = F d the class of measure functions h(t) = t d ε(t), 0 ≤ ε(t) ≤ 1,<br />
such that lim t→0 ε(t) = 0. Thelowerd-dimensional Hausdorff content of F is then<br />
defined by<br />
M d ∗<br />
(F ) = sup M h (F ).<br />
h∈F d<br />
One has M d ∗ ≤ Md , but it can happen that M d ∗ (F ) = 0 < Md (F ). For instance, if F<br />
is the segment [0, 1] in the plane, then M 1 ∗ (F ) = 0,butM1 (F ) = 1. An old result of<br />
Sion and Sjerve [28] in geometric measure theory asserts that M d ∗<br />
(F ) = 0 if and only<br />
if F is a countable union of sets with finite d-dimensional Hausdorff measure. For a<br />
disk B,forM d ∗ (B) = Md (B), and for open sets U, M d ∗ (U) ≃ Md (U). We may now<br />
reformulate Theorem 2.5 as follows.<br />
THEOREM 2.6<br />
Let E ⊂ C be a compact set, and let φ : C → C be a principal K-quasiconformal<br />
mapping. If M 2/(K+1)<br />
∗<br />
(E) = 0, thenM 1 ∗<br />
(φ(E)) = 0.<br />
For the proof, for any bounded set F ⊂ C define first<br />
γ ∗ (F ) = sup|f ′ (∞)|, (2.9)<br />
where the supremum is taken over all functions f ∈ VMO, with ‖f ‖ ∗ ≤ 1, which<br />
are holomorphic on C \ F and satisfy f (∞) = 0. Again, here we may replace<br />
|f ′ (∞)| with |〈∂f,1〉|. TheVMO-invariance leads to the following analogue of<br />
Lemma 2.1.
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 549<br />
LEMMA 2.7<br />
Let E be a compact set. For any principal K-quasiconformal mapping φ : C → C,<br />
conformal on C \ E, we have<br />
γ ∗<br />
(<br />
φ(E)<br />
)<br />
≃ γ∗ (E).<br />
Proof<br />
Consider f ∈ VMOwhich is analytic in C\φ(E) and f (∞) = 0.Setg = f ◦φ.Then<br />
g ∈ VMO, g is analytic on C \ E, ‖g‖ ∗ ≤ C ‖f ‖ ∗ ,and|g ′ (∞)| =|f ′ (∞)| since φ<br />
is a principal K-quasiconformal mapping. Consequently, γ ∗ (φ(E)) ≤ Cγ ∗ (E). <br />
It was shown by Verdera [32] that this VMO-capacity is essentially the 1-dimensional<br />
lower content.<br />
LEMMA 2.8 ([32, page 288])<br />
For any compact set E, M 1 ∗ (E) ≃ γ ∗(E).<br />
With these tools, we are ready to prove Theorem 2.6.<br />
Proof of Theorem 2.6<br />
Naturally, the argument is similar to that in Theorem 2.2. Without loss of generality,<br />
we may assume that E ⊂ D and that φ is a principal K-quasiconformal mapping.<br />
Furthermore, we may assume that H 2/(K+1) (E) is finite, and for any δ, we have a finite<br />
family of disks D i such that E ⊂ ⋃ i D i, ∑ i diam(D i) 2/(K+1) ≤ H 2/(K+1) (E) + 1,<br />
and diam(D i )
550 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
Since K-quasiconformal mappings are Hölder continuous with exponent 1/K,<br />
M h( φ 1 () ) ≤ ∑ diam ( φ 1 (D j ) ) ε ( diam(φ 1 (D j )) ) ≤ ε(C K δ 1/K ) ∑ diam ( φ 1 (D j ) )<br />
j<br />
j<br />
≤ ε(C K δ 1/K ) ∑ j<br />
( ∫ ) (K−1)/(2K)|Dj<br />
J (z, φ 1 ) K/(K−1) dm(z)<br />
| 1/2K<br />
D j<br />
( ∑<br />
) (K+1)/(2K)<br />
≤ ε(C K δ 1/K ) C K diam(D j ) 2/(K+1)<br />
j<br />
≤ ε(C K δ 1/K ) C K<br />
(H 2/(K+1) (E) + 1) (K+1)/(2K).<br />
Finally, taking δ → 0, wegetM h (φ(E)) = 0. This holds for any h ∈ F, andthe<br />
theorem follows.<br />
<br />
One may think of extending the preceding results from the critical index 2/(K + 1)<br />
to arbitrary ones by using other capacities that behave like a Hausdorff content. For<br />
instance, the capacity γ α , associated to analytic functions contained in Lip(α) (see<br />
[23]), satisfies<br />
M 1+α (E) ≃ γ α (E),<br />
but unfortunately, the space Lip(α) is not invariant under a quasiconformal change<br />
of variables. Thus other procedures are needed. It turns out that the homogeneous<br />
Sobolev spaces provide suitable tools, basically, since Ẇ 1,2 (C) is invariant under quasiconformal<br />
mappings. Here, recall that for 0
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 551<br />
so that g ∈ Ẇ 1,2 (C). In other words, every K-quasiconformal mapping φ induces a<br />
bounded linear operator<br />
T : Ẇ 1,2 (C) → Ẇ 1,2 (C),<br />
T(f ) = f ◦ φ<br />
with norm depending only on K. As we have mentioned before, this operator T is also<br />
bounded on BMO(C) (see [26]). Moreover, Reimann and Rychener [27, page 103]<br />
proved that Ẇ 2/q,q (C), q>2, may be represented as a complex interpolation space<br />
between BMO(C) and Ẇ 1,2 (C). It follows that T is bounded on the Sobolev spaces<br />
Ẇ 2/q,q (C),q >2. More precisely, there exists a constant C = C(K,q) such that<br />
‖f ◦ φ‖ Ẇ 2/q,q (C) ≤ C‖f ‖ Ẇ 2/q,q (C) (2.10)<br />
for any K-quasiconformal mapping φ on C. These invariant function spaces provide<br />
us with related invariant capacities. Recall (e.g., see [1, pages 34, 46]) that for any<br />
pair α>0, p>1 with 0
552 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
and consequently, we can write<br />
f = 1 z ∗ µ = R(I 1 ∗ µ) = I 1−α ∗ R(I α ∗ µ),<br />
where R is a Calderón-Zygmund operator and ‖f ‖ Ẇ 1−α,q =‖R(I α ∗µ)‖ q ‖I α ∗µ‖ q .<br />
For the converse, let f = I 1−α ∗ g be an admissible function for γ 1−α,q .Wehave<br />
that, up to a multiplicative constant, T = ∂f is an admissible distribution for Ċ α,p<br />
because<br />
I α ∗ T = R t (g),<br />
where R t is the transpose of R. Thus Ċ α,p (E) 1/p ≥|〈T,1〉| = |f ′ (∞)|, and the proof<br />
is complete.<br />
<br />
We end up with new quasiconformal invariants built on the Riesz capacities.<br />
THEOREM 2.10<br />
Let φ : C → C be a principal K-quasiconformal mapping of the plane which is<br />
conformal on C \ E.Let1
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 553<br />
are compact sets F such that Ċ α,p (F ) = 0 and H h (F ) > 0 for some measure function<br />
h(t) = t p ε(t). Thus Theorem 2.10 does not help with Conjecture 2.3. Wehavetobe<br />
content with the following setup.<br />
Given 1
554 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
3. Distortion of rectifiable sets<br />
In general, if φ is a K-quasiconformal mapping and E is a compact set, it follows<br />
from (1.2) that<br />
dim(E) = 1 ⇒ 2 ≤ dim φ(E) ≤<br />
2K<br />
K + 1<br />
K + 1 . (3.1)<br />
Here, for both estimates, one may find mappings φ and sets E such that the equality<br />
is attained (see [4]). In [4], all examples come from nonregular Cantor-type constructions.<br />
Thus the extremal distortion of Hausdorff dimension is attained, at least, by sets<br />
irregular enough. The main purpose of this section is to prove that some irregularity<br />
is also necessary. Namely, we show that quasiconformal images of 1-rectifiable sets<br />
cannot achieve the maximal distortion of dimension.<br />
THEOREM 3.1<br />
Suppose that φ : C → C is a K-quasiconformal mapping, K>1.LetE ⊂ ∂D be a<br />
subset of the unit circle with dim(E) = 1. Then we have the strict inequality<br />
dim ( φ(E) ) > 2<br />
K + 1 .<br />
With a similar but easier argument, one may also prove that for such sets E, neither<br />
can dim(φ(E)) attain the upper bound in (3.1) (for details, see Remark 3.7).<br />
From Theorem 3.1, we obtain as an immediate corollary the following more<br />
general result.<br />
COROLLARY 3.2<br />
Suppose that E is a 1-rectifiable set, and let φ : C → C be a K-quasiconformal<br />
mapping, K>1.Then<br />
dim φ(E) > 2<br />
K + 1 .<br />
Recall that a set E ⊂ C is said to be 1-rectifiable if there exists a set E 0 of zero length<br />
such that E \ E 0 is contained in a countable union of Lipschitz curves; that is,<br />
E \ E 0 ⊂<br />
∞⋃<br />
j ([0, 1]),<br />
j=1<br />
where all j : [0, 1] → C are Lipschitz mappings. Alternatively (see [20]), 1-<br />
rectifiable sets can be viewed as subsets of countable unions of C 1 -curves, modulo a<br />
set of zero length. In particular, for any ε>0, there is a decomposition<br />
∞<br />
E \ E ′ 0 = ⋃<br />
E i ,<br />
i=1
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 555<br />
where E ′ 0 has zero length and each E i can be written as E i = f i (F i ) with f i : C → C<br />
a (1 + ε)-bi-Lipschitz mapping and F i ⊂ ∂D. From this and Theorem 3.1, we obtain<br />
Corollary 3.2.<br />
To prove Theorem 3.1, first some reductions may be made. Recall (see [18]) that<br />
every K-quasiconformal mapping φ can be factored as φ = φ n ◦···◦φ 1 , where each<br />
φ j is K j -quasiconformal, and K 1 K 2 ···K n = K. In particular, given ε>0, we can<br />
choose K j ≤ 1 + ε for all j = 1,...,nwhen n is large enough. On the other hand,<br />
recall that from the distortion of Hausdorff dimension (1.2), we have<br />
1<br />
dim φ(E) − 1 (<br />
2 ≤ K 1<br />
dim E − 1 )<br />
. (3.2)<br />
2<br />
If φ is such that equality in (3.2) holds for E, then every factor φ j above must give<br />
equality for the set E j = φ j−1 ◦···◦φ 1 (E) and K = K j . In particular, if the mapping<br />
φ 1 fails to satisfy the equality in (3.2), then so will φ. By combining these facts, we<br />
deduce that in order to prove Theorem 3.1, we can assume that K = 1 + ε with ε>0<br />
as small as we wish.<br />
For mappings with small dilatation, it is possible to achieve quantitative and more<br />
symmetric local distortion estimates. In particular, Theorem 3.1 follows from the next<br />
lower bounds for compression of dimension.<br />
THEOREM 3.3<br />
Suppose that φ : C → C is (1 + ε)-quasiconformal, and suppose that E ⊂ ∂D.Then<br />
for all ε>0 small enough,<br />
dim(E) ≥ 1 − c 0 ε 2 ⇒ dim ( φ(E) ) ≥ 1 − c 1 ε 2 , (3.3)<br />
where the constants c 0 ,c 1 > 0 are independent of ε.<br />
Our basic strategy toward this result is to reduce it to the properties of harmonic measure<br />
and conformal mappings admitting quasiconformal extensions. Indeed, denote<br />
by µ the Beltrami coefficient of φ,andleth be the principal solution to ∂h = χ D µ∂h.<br />
Then h is conformal outside the unit disk. Inside D, it has the same dilatation µ as φ<br />
and hence differs from this by a conformal factor. Consequently, we may find Riemann<br />
mappings f : D → := φ(D) and g : D → ′ := h(D) so that<br />
φ(z) = f ◦ g −1 ◦ h(z), z ∈ D. (3.4)<br />
Moreover, since the (1 + ε)-quasiconformal mapping G = g −1 ◦ h preserves the disk,<br />
reflecting across the boundary ∂D one may extend G to a (1 + ε)-quasiconformal<br />
mapping of C. At the same time, this procedure provides both f and g with (1 + ε) 2 -<br />
quasiconformal extensions to the entire plane C.
556 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
As the final reduction, we now find from (3.4) that for Theorems 3.1 and 3.3,itis<br />
sufficient to prove the following result.<br />
THEOREM 3.4<br />
Suppose that f : C → C is a (1 + ε)-quasiconformal mapping of C, conformal in<br />
the disk D.LetA ⊂ ∂D. There are constants c 0 , c 1 and γ 0 , γ 1 , independent of ε, such<br />
that for ε ≥ 0 small enough,<br />
(i) dim(A) ≥ 1 − c 0 ε 2 ⇒ dim(f (A)) ≥ 1 − c 1 ε 2 , and<br />
(ii) dim(A) ≤ 1 − γ 0 ε 2 ⇒ dim(f (A)) ≤ 1 − γ 1 ε 2 .<br />
Proof<br />
The conclusion (i) follows from Makarov’s fundamental estimates for the harmonic<br />
measure (see [19]; see also [24, page 231]). In his article [19], Makarov proves that<br />
for any conformal mapping f defined on D, for any Borel subset A ⊂ ∂D ,andfor<br />
every q>0, we have the lower bound<br />
dim ( f (A) ) ≥<br />
0<br />
q dim(A)<br />
β f (−q) + q + 1 − dim(A) . (3.5)<br />
Here, β f (p) stands for the integral means spectrum. That is, for a given p ∈ R, β f (p)<br />
is the infimum of all numbers β such that<br />
∫ 2π<br />
(<br />
|f ′ (re it )| p 1<br />
)<br />
dt = O , (3.6)<br />
(1 − r) β<br />
as r → 1 − .<br />
Hence we need estimates for β f (p), and here for mappings admitting K-<br />
quasiconformal extensions, one has qualitatively sharp bounds. Indeed, it can be<br />
shown (see [24, page 182]) that<br />
( K − 1<br />
) 2<br />
β f (p) ≤ 9 p<br />
2<br />
(3.7)<br />
K + 1<br />
for any p ∈ R. The constant 9 is not optimal but suffices for our purposes. Choosing<br />
q = 1 in (3.5) immediately gives the claim (i).<br />
For general conformal mappings, there is no bound for expansion of dimension<br />
(i.e., there is no upper bound analogue of (3.5)). Hence the proof of (ii) strongly<br />
uses the fact that the mappings considered have (1 + ε)-quasiconformal extensions.<br />
However, here also, this information is easiest to use in the form (3.7).<br />
We first need to introduce some further notation. The Carleson squares of the unit<br />
disk are defined as<br />
Q j,k = { z ∈ D :2 −k ≤ 1 −|z| < 2 −k+1 , 2 −k+1 πj ≤ arg(z) < 2 −k+1 π(j + 1) } .
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 557<br />
Given a point z ∈ D \{0}, letQ(z) denote the unique Carleson square that contains<br />
z. Then it follows from Koebe’s distortion theorem and quasisymmetry (see [6], [18])<br />
that if D(ξ,r) is a disk centered at ξ ∈ ∂D, wehave<br />
diam ( f (D(ξ,r)) ) ≃ diam ( f (Q(z)) ) ≃|f ′ (z)|(1 −|z|) for z = (1 − r)ξ, (3.8)<br />
whenever f : C → C is a K-quasiconformal mapping, conformal in D.<br />
Furthermore, assume that we are given a family of disjoint disks D i = D(ξ i ,r i )<br />
with centers ξ i ∈ ∂D, i ∈ N, on the unit circle. Then, write z i = (1 − r i )ξ i ,andfor<br />
any pair of real numbers 0
558 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
then<br />
(2 −k ) α/δ ≤ (1 −|z i |) α/δ < (1 −|z i |)|f ′ (z i )|≤2 −m .<br />
Hence the indices m with Im k nonempty lie on an interval m 0 ≤ m ≤ (α/δ)k.<br />
From Koebe, we also see that if i ∈ Im k ,then|f ′ (w)| p ∼ 2 p(k−m) for every<br />
w ∈ Q(z i ) with constants depending only on p. Combining this with (3.7) gives,for<br />
any τ>0,<br />
q k m ≤ C 2k((9/4)ε2 p 2 +τ+1−((k−m)/k)p) ,<br />
where C now depends on p and τ. We may take p = (k − m)/(10kε 2 ) and obtain<br />
q k m ≤ C 2(1+τ)k−(k/(10ε)2 )((k−m)/k) 2 .<br />
Since diam(f (D i )) is comparable to |f ′ (z i )| (1 −|z i |) ∼ 2 −m for i ∈ I k m ,<br />
∑<br />
i∈I b (α,δ)<br />
diam ( f (D i ) ) δ<br />
≤ C<br />
∞<br />
∑<br />
≤ C<br />
(α/δ)k<br />
∑<br />
q k m 2−mδ<br />
k=0 m=m 0<br />
∞∑<br />
k=0<br />
(α/δ)k<br />
∑<br />
m=m 0<br />
2 k(1+τ−m(δ/k)−(1/(100ε2 ))((k−m)/k) 2) . (3.10)<br />
One now needs to ensure that the exponent 1 + τ − m(δ/k) − (1/100ε 2 )((k − m)/k) 2<br />
is negative. In particular, we want the exponent to attain its maximum at m = (α/δ)k,<br />
andthisissatisfiedif<br />
α<br />
δ ≤ 1 − 1 2 (10ε)2 δ.<br />
Under the assumptions of the lemma, this is easy to verify. Similarly, one verifies that<br />
the specific choices of the lemma yield the maximum value<br />
1 + τ − α − 1 (1 − α ) 2<br />
< 0<br />
(10ε) 2 δ<br />
when τ is chosen small enough. It follows that the sum in (3.10) has a finite upper<br />
bound depending only on the constants M, N. This proves the lemma.<br />
<br />
The dimension bounds required in Theorem 3.4(ii) are now easy to establish. For<br />
every α>1 − γ 0 ε 2 , we have disjoint families of disks D j = D(z j ,r j ) centered on<br />
∂D and radius r j ≤ ρ → 0 uniformly small, so that A is covered by 5D j and the<br />
sums ∑ j diam(D j) α are uniformly bounded. On the image side, for each δ>0,<br />
∑<br />
diamf (5D i ) δ ≃ ∑<br />
i<br />
i<br />
diamf (D i ) δ =<br />
∑<br />
i∈I g (α,δ)<br />
diamf (D i ) δ +<br />
∑<br />
i∈I b (α,δ)<br />
diamf (D i ) δ .
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 559<br />
As soon as α
560 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
Moreover, for the dimension of quasicircles, Smirnov (unpublished) has obtained the<br />
upper bound<br />
dim(Ɣ) ≤ 1 +<br />
( K − 1<br />
) 2,<br />
K + 1<br />
answering a question in [4]. It is still unknown if this bound is sharp; the best known<br />
lower bounds so far (see [8]) give curves with dimension<br />
( K − 1<br />
) 2.<br />
1 + 0.69<br />
K + 1<br />
The arguments that we have used are related to the generalized Brennan conjecture,<br />
which says that<br />
β f (p) ≤ p2<br />
4<br />
( K − 1<br />
) 2<br />
for |p| ≤2 K + 1<br />
K + 1<br />
K − 1 , (3.12)<br />
whenever f is conformal in D and admits a K-quasiconformal extension to C. This<br />
connection suggests the following.<br />
Question 3.8<br />
Let E ⊂ R be a set with Hausdorff dimension 1, andletφ be a K-quasiconformal<br />
mapping. Then is it true that<br />
1 −<br />
( K − 1<br />
) 2 ( ) ( K − 1<br />
) 2?<br />
≤ dim φ(E) ≤ 1 + (3.13)<br />
K + 1<br />
K + 1<br />
The positive answer for the right-hand-side inequality follows from Smirnov’s unpublished<br />
work, while the left-hand side is only known up to some multiplicative<br />
constants. On the other hand, Prause [25] proves the left-hand-side inequality for the<br />
mappings that preserve the unit circle ∂D.<br />
4. Improved Painlevé theorems<br />
A compact set E is said to be removable for bounded analytic functions if, for any<br />
open set with E ⊂ , every bounded analytic function on \ E has an analytic<br />
extension to . Equivalently, such sets are described by the condition γ (E) = 0,<br />
where γ is the analytic capacity<br />
γ (E) = sup { |f ′ (∞)| : f ∈ H ∞ (C \ E),f(∞) = 0, ‖f ‖ ∞ ≤ 1 } .<br />
Finding a geometric characterization for the sets of zero analytic capacity was a longstanding<br />
problem. It was solved by David [12] for sets of finite length and, finally, by<br />
Tolsa [29] in the general case. The difficulties of dealing with this question motivated<br />
the study of related problems. In particular, we have the question of determining the
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 561<br />
removable sets for BMO analytic functions (i.e., those compact sets E such that every<br />
BMO-function in the plane, holomorphic on C \ E, admits an entire extension). This<br />
problem was solved by Král [17], who showed that a set E has this BMO-removability<br />
property if and only if H 1 (E) = 0. This was also proved independently by Kaufman<br />
[16].<br />
For the original case of bounded functions, the Painlevé condition H 1 (E) = 0<br />
can be weakened. As is well known, there are sets E with zero analytic capacity<br />
and positive length (e.g., see [14]). In fact, it is now known that among the compact<br />
sets E with 0 < H 1 (E) < ∞, precisely the purely unrectifiable ones are the removable<br />
sets for bounded analytic functions (see [12]). Moreover, if E has positive<br />
σ -finite length, this characterization still remains true, due to the countable semiadditivity<br />
of analytic capacity (see [29]).<br />
The preceding problems can be formulated also in the K-quasiregular setting.<br />
More precisely, a set E is said to be removable for bounded (resp., BMO) K-<br />
quasiregular mappings if every K-quasiregular mapping in C \ E which is in L ∞ (C)<br />
(resp., BMO(C)) admits a K-quasiregular extension to C. For simplicity, we use<br />
here the term K-removable for sets that are removable for bounded K-quasiregular<br />
mappings.<br />
Obviously, when K = 1, in both situations of L ∞ and BMO we recover the<br />
original analytic problem. Moreover, by means of the Stoïlow factorization, one can<br />
represent any bounded K-quasiregular function as a composition of a bounded analytic<br />
function and a K-quasiconformal mapping. The corresponding result also holds true<br />
for BMO since this space, like L ∞ , is quasiconformally invariant.<br />
Therefore, when we ask ourselves if a set E is K-removable, we just need to<br />
analyze how it may be distorted under quasiconformal mappings and then apply<br />
the known results for the analytic situation. With this basic scheme, it is shown in [4,<br />
Corollary 1.5] that every set with dimension strictly below 2/(K + 1) is K-removable.<br />
Indeed, the precise formulas for the distortion of dimension (1.2) ensure that for such<br />
sets, the K-quasiconformal images have dimension strictly smaller than 1.<br />
Iwaniec and Martin [15] had earlier conjectured that, more generally, sets of zero<br />
(2/(K + 1))-dimensional measure are K-removable. A preliminary answer to this<br />
question was found in [6], and actually, it was that argument that suggested Theorem<br />
2.2. Using our results from above, we can now prove that sets of zero (2/(K + 1))-<br />
dimensional measure are even BMO-removable.<br />
COROLLARY 4.1<br />
Let E be a compact subset of the plane. Assume that H 2/(K+1) (E) = 0. ThenE is<br />
removable for all BMO K-quasiregular mappings.<br />
Proof<br />
Assume that f ∈ BMO(C) is K-quasiregular on C \ E. Denote by µ the Beltrami<br />
coefficient of f ,andletφ be the principal solution to ∂φ = µ∂φ.ThenF =
562 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
f ◦ φ −1 is holomorphic on C \ φ(E) and F ∈ BMO(C). On the other hand, as we<br />
showed in Theorem 2.2, H 1 (φ(E)) = 0. Thus φ(E) is a removable set for BMO<br />
analytic functions. In particular, F admits an entire extension, and f = F ◦ φ extends<br />
quasiregularly to the whole plane.<br />
<br />
We believe that Corollary 4.1 is sharp, in the sense that we expect a positive answer<br />
to the following.<br />
Question 4.2<br />
Does there exist for every K ≥ 1, a compact set E with 0 < H 2/(K+1) (E) < ∞, such<br />
that E is not removable for some K-quasiregular functions in BMO(C)?<br />
Here, we observe that by [4, Corollary 1.5], for every t>2/(K + 1) there exists a<br />
compact set E with dimension t, nonremovable for bounded and hence, in particular,<br />
nonremovable for BMO K-quasiregular mappings.<br />
Next, we return back to the problem of removable sets for bounded K-quasiregular<br />
mappings. Here, Theorem 2.2 proves the conjecture of Iwaniec and Martin, that sets<br />
with H 2/(K+1) (E) = 0 are K-removable. However, the analytic capacity is somewhat<br />
smaller than length, and hence with Theorem 2.5 we may go even further. If a set<br />
has finite or σ -finite (2/(K + 1))-measure, then all K-quasiconformal images of E<br />
have at most σ -finite length. Such images may still be removable for bounded analytic<br />
functions if we can make sure that the rectifiable part of these sets has zero length. But<br />
for this, Theorem 3.1 provides exactly the correct tools. We end up with the following<br />
improved version of the Painlevé theorem for quasiregular mappings.<br />
THEOREM 4.3<br />
Let E be a compact set in the plane, and let K>1. Assume that H 2/(K+1) (E) is<br />
σ -finite. Then E is removable for all bounded K-quasiregular mappings.<br />
In particular, for any K-quasiconformal mapping φ, the image φ(E) is purely<br />
unrectifiable.<br />
Proof<br />
Let f : C → C be bounded, and assume that f is K-quasiregular on C \ E. Asin<br />
Corollary 4.1, we may find the principal quasiconformal homeomorphism φ : C → C,<br />
so that F = f ◦ φ −1 is analytic in C \ φ(E). If we can extend F holomorphically<br />
to the whole plane, we are done. Thus we have to show that φ(E) has zero analytic<br />
capacity.<br />
By Theorem 2.5, φ(E) has σ -finite length (i.e., φ(E) = ⋃ n F n, where each<br />
H 1 (F n ) < ∞). A well-known result due to Besicovitch (see, e.g., [20, page 205])
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 563<br />
assures us that each set F n can be decomposed as<br />
F n = R n ∪ U n ∪ B n ,<br />
where R n is a 1-rectifiable set, U n is a purely 1-unrectifiable set, and B n is a set of<br />
zero length. Because of the semiadditivity of analytic capacity (see [29]),<br />
γ (F n ) ≤ C ( γ (R n ) + γ (U n ) + γ (B n ) ) .<br />
Now, γ (B n ) ≤ C H 1 (B n ) = 0 and γ (U n ) = 0 since purely 1-unrectifiable sets<br />
of finite length have zero analytic capacity (see [12]). On the other hand, R n is<br />
a 1-rectifiable image, under a K-quasiconformal mapping, of a set of dimension<br />
2/(K + 1). Thus applying Theorem 3.1 and Corollary 3.2 to φ −1 shows that we must<br />
have H 1 (R n ) = 0. Therefore we get γ (F n ) = 0 for each n. Again, by countable<br />
semiadditivity of analytic capacity, we conclude that γ (φ(E)) = 0.<br />
<br />
As pointed out earlier, Theorem 4.3 does not hold for K = 1.Any1-rectifiable set such<br />
as E = [0, 1] of finite and positive length gives a counterexample. In the above proof,<br />
the improved distortion of 1-rectifiable sets is the decisive phenomenon allowing the<br />
result. In fact, such “good” behavior of rectifiable sets has further consequences. For instance,<br />
even strictly above the critical dimension 2/(K + 1) = 1 − (K − 1)/(K + 1),<br />
one may find removable sets, as soon as they have enough geometric regularity.<br />
COROLLARY 4.4<br />
There exists a constant c ≥ 1 such that if E ⊂ ∂D is compact and<br />
( K − 1<br />
) 2,<br />
dim(E) < 1 − c<br />
K + 1<br />
then E is removable for bounded and BMO K-quasiregular mappings, K = 1 + ε,<br />
whenever ε>0 is small enough.<br />
Proof<br />
This is a consequence of Corollary 3.6.Ifε>0 is small enough and K = 1 + ε,then<br />
the K-quasiconformal images of E always have dimension strictly below 1, sothat<br />
γ (φ(E)) = 0 for each K-quasiconformal mapping φ.<br />
<br />
In conjunction with Question 3.8, we have the following.<br />
Question 4.5<br />
Let K > 1. Then is every set E ⊂ ∂D with dim(E) < 1 − ((K − 1)/(K + 1)) 2<br />
removable for bounded and BMO K-quasiregular mappings?
564 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
5. Examples of extremal distortion<br />
Sections 2 and 3 provide a delicate analysis of distortion of 1-dimensional sets under<br />
quasiconformal mappings but still leave open the cases where dim(E) = 2/(K + 1)<br />
precisely, but E does not have σ -finite (2/(K + 1))-measure. Hence we are faced with<br />
the natural question: are there compact sets E, with dim(E) = 2/(K + 1), which are<br />
not removable for some bounded K-quasiregular mappings?<br />
In this last section, we give a positive answer and show that our results are sharp<br />
in quite a strong sense. Indeed, to compare with the analytic removability, recall first<br />
that by Mattila’s theorem [21, Theorem 3.8], if a compact set E supports a probability<br />
measure with µ(B(z, r)) ≤ rε(r) and<br />
∫<br />
ε(t) 2<br />
dt < ∞, (5.1)<br />
0 t<br />
then the analytic capacity γ (E) > 0. On the other hand, if the integral in (5.1) diverges,<br />
then there are compact sets E of vanishing analytic capacity supporting a probability<br />
measure with µ(B(z, r)) ≤ rε(r) (see [29]). In a complete analogy, we prove the<br />
following.<br />
THEOREM 5.1<br />
Let K ≥ 1. Suppose that h(t) = t 2/(K+1) ε(t) is a measure function such that<br />
∫<br />
ε(t) 1+1/K<br />
dt < ∞. (5.2)<br />
0 t<br />
Then there is a compact set E that is not K-removable and yet supports a probability<br />
measure µ with µ(B(z, r)) ≤ h(r) for every z and r>0.<br />
In particular, whenever ε(t) is chosen so that, in addition, for every α>0, we<br />
have t α /ε(t) → 0 as t → 0, then the construction gives a non-K-removable set E<br />
with dim(E) = 2/(K + 1).<br />
Proof<br />
We construct a compact set E and a K-quasiconformal mapping φ so that H h (E) ≃ 1<br />
and, at the same time, φ(E) has a positive and finite H h′ -measure for some measure<br />
function h ′ (t) = tε ′ (t), where<br />
h ′ (t) = tε ′ (t)<br />
with<br />
∫ 1<br />
0<br />
ε ′ (t) 2<br />
dt < ∞.<br />
t<br />
Then Mattila’s theorem [21, Theorem 3.8] shows that γ (φ(E)) > 0, so that there exist<br />
nonconstant bounded functions h holomorphic on C \ φ(E). Thus, with f = h ◦ φ,<br />
we see that E is not removable for bounded K-quasiregular mappings.
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 565<br />
We construct the K-quasiconformal mapping φ as the limit of a sequence φ N<br />
of K-quasiconformal mappings, and E is a Cantor-type set. To reach the optimal<br />
estimates, we need to change, at every step in the construction of E, both the size and<br />
the number m j of the generating disks.<br />
Without loss of generality, we may assume that for every α>0, t α /ε(t) → 0 as<br />
t → 0.<br />
Step 1. First, choose m 1 disjoint disks D(z i ,R 1 ) ⊂ D, i = 1,...,m 1 ,sothat<br />
c 1 := m 1 R 2 1 ∈ ( 1<br />
2 , 1 ).<br />
For R 1 small enough (i.e., for m 1 large enough), this is clearly possible. The function<br />
f (t) = m 1 h(tR 1 ) is continuous with f (0) = 0. Moreover, for each fixed t,<br />
f (t) = m 1 (tR 1 ) 2/(K+1) ε(tR 1 ) =<br />
ε(t √ c 1 /m 1 )<br />
(t √ c 1 /m 1 ) 2K/(K+1) t 2 c 1 → ∞<br />
as m 1 →∞. Hence, for any t < 1, we may choose m 1 so large that there exists<br />
σ 1 ∈ (0,t) satisfying m 1 h(σ K 1 R 1) = 1. A simple calculation gives<br />
m 1 σ 1 R 1 ε(σ K 1 R 1) (K+1)/(2K) (c 1 ) (1−K)/(2K) = 1. (5.3)<br />
Next, let r 1 = R 1 . For each i = 1,...,m 1 ,letϕ 1 i (z) = z i + σ K 1 R 1 z, and using<br />
the notation αD(z, ρ) := D(z, αρ),set<br />
D i := 1<br />
σ K 1<br />
ϕ 1 i (D) = D(z i,r 1 ),<br />
D ′ i := ϕ1 i (D) = D(z i,σ K 1 r 1) ⊂ D i .<br />
As the first approximation of the mapping, define<br />
⎧<br />
σ ⎪⎨<br />
1−K<br />
1 (z − z i ) + z i , z ∈ D<br />
i ′ ,<br />
g 1 (z) = ∣ z − z ∣<br />
i ∣∣<br />
1/K−1<br />
(z − zi ) + z i , z ∈ D i \ D<br />
i ′ r ,<br />
1<br />
⎪⎩ z, z /∈ ∪D i .<br />
This is a K-quasiconformal mapping, conformal outside of ⋃ m 1<br />
i=1 (D i \ D<br />
i ′ ). It maps<br />
each D i onto itself and D<br />
i ′ onto D′′<br />
i<br />
= D(z i ,σ 1 r 1 ), while the rest of the plane remains<br />
fixed. Write φ 1 = g 1 .<br />
Step 2. We have already fixed m 1 ,R 1 ,σ 1 ,andc 1 . Consider m 2 disjoint disks of radius<br />
R 2 , centered at z 2 j , j = 1,...,m 2, uniformly distributed inside of D, sothat<br />
c 2 = m 2 R 2 2 > 1 2 .
566 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
Figure 1<br />
Then repeat the above procedure, and choose m 2 so large that the equation<br />
m 1 m 2 h(σ K 1 σ K 2 R 1R 2 ) = 1<br />
has a unique solution σ 2 ∈ (0, 1), as small as we wish. Then<br />
m 1 m 2 σ 1 σ 2 R 1 R 2 ε(σ K 1 σ K 2 R 1R 2 ) (K+1)/(2K) (c 1 c 2 ) (1−K)/(2K) = 1.<br />
Denote r 2 = R 2 σ 1 r 1 and ϕ 2 j (z) = z2 j + σ K 2 R 2 z, and define the auxiliary disks<br />
( 1<br />
)<br />
D ij = φ 1 ϕ 1<br />
σ2<br />
K i<br />
◦ ϕ 2 j (D) = D(z ij ,r 2 ),<br />
D ′ ij = φ (<br />
1 ϕ<br />
1<br />
i<br />
◦ ϕ 2 j (D)) = D ′ (z ij ,σ K 2 r 2)<br />
for certain z ij ∈ D, where i = 1,...,m 1 and j = 1,...,m 2 .Now,let<br />
⎧<br />
σ ⎪⎨<br />
1−K<br />
2 (z − z ij ) + z ij , z ∈ D<br />
ij ′ ,<br />
g 2 (z) = ∣ z − z ∣<br />
ij ∣∣<br />
1/K−1<br />
(z − zij ) + z ij , z ∈ D ij \ D<br />
ij ′ r ,<br />
2<br />
⎪⎩ z, otherwise.<br />
Clearly, g 2 is K-quasiconformal, conformal outside of ⋃ i,j (D ij \D<br />
ij ′ ), and maps each<br />
D ij onto itself and D<br />
ij ′ onto D′′<br />
ij = D(z ij ,σ 2 r 2 ), while the rest of the plane remains<br />
fixed. Define φ 2 = g 2 ◦ φ 1 .<br />
The induction step (see Figure 1). After step N − 1, wetakem N disjoint disks of<br />
radius R N , with union of D(z N l ,R N ) covering at least half of the area of D,<br />
c N = m N R 2 N > 1 2 . (5.4)
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 567<br />
As before, we may choose m N so large that m 1 ···m N h(σ K 1 ···σ K N R 1 ···R N ) = 1<br />
holds for a unique σ N , as small as we wish. Note that lim N→∞ σ N = 0 and<br />
m 1 ···m N σ 1 R 1 ···σ N R N ε(σ K 1 R 1 ···σ K N R N) (K+1)/(2K) (c 1 ···c N ) (1−K)/(2K) = 1.<br />
Then, denote ϕj N(z) = zN j + σN KR N z and r N = R N σ N−1 r N−1 . For any multi-index<br />
J = (j 1 ,...,j N ), where 1 ≤ j k ≤ m k , k = 1,...,N,let<br />
D J = φ N−1<br />
( 1<br />
σ K N<br />
)<br />
ϕ 1 j 1<br />
◦···◦ϕ N j N<br />
(D) = D(z J ,r N ),<br />
D ′ J = φ (<br />
N−1 ϕ<br />
1<br />
j 1<br />
◦···◦ϕ N j N<br />
(D) ) = D ′ (z J ,σ K N r N),<br />
and let<br />
⎧<br />
σ ⎪⎨<br />
1−K<br />
N (z − z J ) + z J , z ∈ D<br />
J ′ ,<br />
g N (z) = ∣ z − z ∣<br />
J ∣∣<br />
1/K−1<br />
(z − zJ ) + z J , z ∈ D J \ D<br />
J ′ r ,<br />
N<br />
⎪⎩ z, otherwise.<br />
Clearly, g N is K-quasiconformal, conformal outside of ⋃ J =(j 1 ,...,j N ) (D J \ D<br />
J ′ ),and<br />
maps D J onto itself and D<br />
J ′ onto D′′<br />
J<br />
= D(z J ,σ N r N ), while the rest of the plane<br />
remains fixed. Now, define φ N = g N ◦ φ N−1 .<br />
Since each φ N is K-quasiconformal and equals the identity outside the unit disk<br />
D, there exists a limit K-quasiconformal mapping<br />
φ = lim<br />
N→∞ φ N<br />
with convergence in W 1,p<br />
loc (C) for any p
568 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
Observe that we have chosen the parameters R N ,m N ,σ N so that<br />
m 1 ···m N h(s N ) = 1, (5.6)<br />
m 1 ···m N t N ε(s N ) (K+1)/(2K) (c 1 ···c N ) (1−K)/(2K) = 1. (5.7)<br />
Claim. We have H h (E) ≃ 1.<br />
Since diam(ϕ 1 j 1<br />
◦···◦ϕ N j N<br />
(D)) ≤ δ N → 0 when N →∞,wehave,by(5.6),<br />
∑<br />
H h (E) = lim H h δ<br />
(E) ≤ lim h ( diam(ϕ 1 j δ→0 δ→0<br />
1<br />
◦···◦ϕ N j N<br />
(D)) ) = m 1 ···m N h(s N ) = 1.<br />
j 1 ,...,j N<br />
For the converse inequality, take a finite covering (U j ) of E by open disks of diameter<br />
diam(U j ) ≤ δ, andletδ 0 = inf j (diam(U j )) > 0. Denote by N 0 the minimal integer<br />
such that s N0 ≤ δ 0 . By construction, the family (ϕ N 0<br />
j N0<br />
◦···◦ϕj 1 1<br />
(D)) j1 ,...,j N0<br />
is a covering<br />
of E with the M h -packing condition (see [20]). Thus<br />
∑<br />
h ( diam(U j ) ) ≥ C<br />
j<br />
∑<br />
Hence H h δ (E) ≥ C, and letting δ → 0, weget<br />
j 1 ,...,j N0<br />
h ( diam(ϕ N 0<br />
j N0<br />
◦···◦ϕ 1 j 1<br />
(D)) ) = C.<br />
C ≤ H h( φ(E) ) ≤ 1,<br />
proving our first claim.<br />
A similar argument, based this time on (5.7), gives that H h′ (φ(E)) ≃ 1 for a<br />
measure function h ′ (t) = tε ′ (t), as soon as for all indices N,<br />
ε ′ (t N ) = ε(s N ) (K+1)/(2K) (c 1 ···c N ) (1−K)/(2K) . (5.8)<br />
Claim. One can find a continuous and nondecreasing function ε ′ (t) satisfying (5.8)<br />
and<br />
∫ 1<br />
ε ′ (t) 2<br />
dt < ∞. (5.9)<br />
0 t<br />
Indeed, let us first choose a continuous nondecreasing function v(t) so that v(t) → 0<br />
as t → 0 and so that (5.2) still holds in the form<br />
∫<br />
ε(t) 1+1/K<br />
dt < ∞. (5.10)<br />
tv(t)<br />
0
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 569<br />
In the above inductive construction, we can then choose the σ j ’s so that<br />
v(σ1<br />
K ···σ N K) ≤ 2−N(1−1/K) for every index N. Now,(5.4)and(5.8) imply that<br />
ε ′ (t N ) 2 ≤ ε(s N ) 1+1/K 2 N(1−1/K) ≤ ε(s N) 1+1/K<br />
.<br />
v(s N )<br />
On the other hand, by (5.5), we also have t N−1 /t N ≤ s N−1 /s N , and so we may extend<br />
ε ′ (t), determined by (5.8) only at the t N ’s, so that it is continuous, nondecreasing, and<br />
satisfies<br />
∫<br />
0<br />
ε ′ (t) 2 dt<br />
t<br />
∫<br />
≤<br />
0<br />
ε(s) 1+1/K<br />
v(s)<br />
ds<br />
s < ∞.<br />
Hence the claim follows. Combining it with Mattila’s theorem [21, Theorem 3.8]<br />
completes the proof of the theorem.<br />
<br />
Lastly, let us note that if we do not care for the analytic capacity of the target set, a<br />
straightforward modification of Theorem 5.1, normalizing the disks of the construction<br />
so that m N t N η(t N ) = 1, gives the following.<br />
COROLLARY 5.2<br />
Let K ≥ 1, and let h(t) = tη(t) be a measure function such that<br />
• η is continuous and nondecreasing, η(0) = 0, and η(t) = 1 whenever t ≥ 1;<br />
• lim<br />
t→0<br />
(t α /η(t)) = 0 for all α>0.<br />
There exist a compact set E ⊂ D and a K-quasiconformal mapping φ such that<br />
dim(E) = 2<br />
K + 1<br />
and H h( φ(E) ) = 1. (5.11)<br />
Note added in proof. In a recent work, Bishop [10] has given a negative answer to<br />
Question 2.4. However, Conjecture 2.3 remains open.<br />
On the other hand, Uriarte-Tuero [31] has recently given a positive answer to<br />
Question 4.2.<br />
References<br />
[1] D. R. ADAMS and L. I. HEDBERG, Function Spaces and Potential Theory, Grundlehren<br />
Math. Wiss. 314, Springer, Berlin, 1996. MR 1411441 551, 553<br />
[2] L. V. AHL<strong>FOR</strong>S, Bounded analytic functions, Duke Math. J. 14 (1947), 1 – 11.<br />
MR 0021108 541
570 ASTALA, CLOP, MATEU, OROBITG, and URIARTE-TUERO<br />
[3] ———, Lectures on Quasiconformal Mappings, Wadsworth Brooks/Cole Math. Ser.,<br />
Wadsworth and Brooks/Cole Adv. Books Software, Monterey, Calif., 1987.<br />
MR 0883205 539<br />
[4] K. ASTALA, Area distortion of quasiconformal mappings, Acta Math. 173 (1994),<br />
37 – 60. MR 1294669 540, 541, 542, 543, 545, 554, 559, 560, 561, 562<br />
[5] K. ASTALA, T. IWANIEC, P. KOSKELA, andG. MARTIN, Mappings of BMO-bounded<br />
distortion, Math. Ann. 317 (2000), 703 – 726. MR 1777116<br />
[6] K. ASTALA, T. IWANIEC,andG. MARTIN, Elliptic partial differential equations and<br />
quasiconformal mappings in plane, manuscript. 542, 557, 561<br />
[7] K. ASTALA and V. NESI, Composites and quasiconformal mappings: New optimal<br />
bounds in two dimensions, Calc. Var. Partial Differential Equations 18 (2003),<br />
335 – 355. MR 2020365 546<br />
[8] K. ASTALA, S. ROHDE,andO. SCHRAMM, Dimension of quasicircles, in preparation.<br />
560<br />
[9] J. BECKER and C. POMMERENKE, On the Hausdorff dimension of quasicircles, Ann.<br />
Acad. Sci. Fenn. Ser. A I Math. 12 (1987), 329 – 333. MR 0951982 559<br />
[10] C. BISHOP, Distortion of disks by conformal maps, preprint, 2007. 569<br />
[11] L. CARLESON, Selected Problems on Exceptional Sets, Van Nostrand Math. Stud. 13,<br />
Van Nostrand, Princeton, 1967. MR 0225986<br />
[12] G. DAVID, Unrectifiable 1-sets have vanishing analytic capacity, Rev.Mat.<br />
Iberoamericana 14 (1998), 369 – 479. MR 1654535 541, 560, 561, 563<br />
[13] J. DUOANDIKOETXEA, Fourier Analysis, revision of the 1995 Spanish original, Grad.<br />
Stud. Math. 29, Amer. Math. Soc., Providence, 2000. MR 1800316<br />
[14] J. GARNETT, Analytic Capacity and Measure, Lecture Notes in Math. 297, Springer,<br />
Berlin, 1972. MR 0454006 561<br />
[15] T. IWANIEC and G. MARTIN, Quasiregular mappings in even dimensions, Acta Math.<br />
170 (1993), 29 – 81. MR 1208562 542, 561<br />
[16] R. KAUFMAN, Hausdorff measure, BMO, and analytic functions, Pacific J. Math. 102<br />
(1982), 369 – 371. MR 0686557 541, 561<br />
[17] J. KRÁL, “Analytic capacity” in Elliptische Differentialgleichungen (Rostock, East<br />
Germany, 1977), Wilhelm-Pieck-Univ., Rostock, East Germany, 1978, 133 – 142.<br />
MR 0540193 541, 544, 561<br />
[18] O. LEHTO and K. I. VIRTANEN, Quasiconformal Mappings in the Plane, 2nd ed.,<br />
Grundlehren Math. Wiss. 126, Springer, New York, 1973. MR 0344463 542,<br />
555, 557<br />
[19] N. G. MAKAROV, Conformal mapping and Hausdorff measures, Ark.Mat.25 (1987),<br />
41 – 89. MR 0918379 556<br />
[20] P. MATTILA, Geometry of Sets and Measures in Euclidean Spaces, Cambridge Stud.<br />
Adv. Math. 44, Cambridge Univ. Press, Cambridge, 1995. MR 1333890 554,<br />
562, 568<br />
[21] ———, On the analytic capacity and curvature of some Cantor sets with non σ -finite<br />
length, Publ. Mat. 40 (1996), 195 – 204. MR 1397014 564, 569<br />
[22] A. MORI, On an absolute constant in the theory of quasi-conformal mappings,J.Math.<br />
Soc. Japan 8 (1956), 156 – 166. MR 0079091 539
DISTORTION OF HAUSDORFF MEASURES AND REMOVABILITY 571<br />
[23] A. G. O’FARRELL, Hausdorff content and rational approximation in fractional<br />
Lipschitz norms, Trans. Amer. Math. Soc. 228 (1977), 187 – 206. MR 0432887<br />
550<br />
[24] C. POMMERENKE, Boundary Behaviour of Conformal Maps, Grundlehren Math. Wiss.<br />
299, Springer, Berlin, 1992. MR 1217706 556<br />
[25] I. PRAUSE, A remark on quasiconformal dimension distortion on the line, Ann. Acad.<br />
Sci. Fenn. Math. 32 (2007), 341 – 352. MR 2337481 559, 560<br />
[26] H. M. REIMANN, Functions of bounded mean oscillation and quasiconformal<br />
mappings, Comment. Math. Helv. 49 (1974), 260 – 276. MR 0361067 544, 551<br />
[27] H. M. REIMANN and T. RYCHENER, Funktionen beschränkter mittlerer Oszillation,<br />
Lecture Notes in Math. 487, Springer, Berlin, 1975. MR 0511997 551<br />
[28] M. SION and D. SJERVE, Approximation properties of measures generated by<br />
continuous set functions, Mathematika 9 (1962), 145 – 156. MR 0146331 548<br />
[29] X. TOLSA, Painlevé’s problem and the semiadditivity of analytic capacity, Acta Math.<br />
190 (2003), 105 – 149. MR 1982794 541, 560, 561, 563, 564<br />
[30] ———, Bilipschitz maps, analytic capacity, and the Cauchy integral, Ann. of Math.<br />
(2) 162 (2005), 1243 – 1304. MR 2179730 541<br />
[31] I. URIARTE-TUERO, Sharp examples for planar quasiconformal distortion of Hausdorff<br />
measures and removability, preprint, arXiv:0707.1184v3 [math.CV] 569<br />
[32] J. VERDERA, BMO rational approximation and one-dimensional Hausdorff content,<br />
Trans. Amer. Math. Soc. 297, no. 1 (1986), 283 – 304. MR 0849480 541, 544,<br />
549<br />
Astala<br />
Department of Mathematics and Statistics, University of Helsinki, FI-00014 Helsinki, Finland;<br />
kari.astala@helsinki.fi<br />
Clop<br />
Departament de Matemàtiques, Facultat de Ciències, Universitat Autònoma de Barcelona,<br />
08193 Bellaterra, Barcelona, Catalonia; albertcp@mat.uab.cat<br />
Mateu<br />
Departament de Matemàtiques, Facultat de Ciències, Universitat Autònoma de Barcelona,<br />
08193 Bellaterra, Barcelona, Catalonia; mateu@mat.uab.cat<br />
Orobitg<br />
Departament de Matemàtiques, Facultat de Ciències, Universitat Autònoma de Barcelona,<br />
08193 Bellaterra, Barcelona, Catalonia; orobitg@mat.uab.cat<br />
Uriarte-Tuero<br />
Department of Mathematics, University of Missouri–Columbia, Columbia, Missouri<br />
65211-4100, USA; ignacio@math.missouri.edu
SOME ASYMPTOTICS OF TOPOLOGICAL<br />
QUANTUM FIELD THEORY VIA SKEIN THEORY<br />
JULIEN MARCHÉ and MAJID NARIMANNEJAD<br />
Abstract<br />
For each oriented surface of genus g, we study a limit of quantum representations of<br />
the mapping class group arising in topological quantum field theory (TQFT) derived<br />
from the Kauffman bracket. We determine that these representations converge in the<br />
Fell topology to the representation of the mapping class group on H(), the space of<br />
regular functions on the SL(2, C)-representation variety with its Hermitian structure<br />
coming from the symplectic structure of the SU(2)-representation variety. As a corollary,<br />
we give a new proof of the asymptotic faithfulness of quantum representations.<br />
1. Introduction<br />
A topological quantum field theory (TQFT) in dimension 2 + 1 is an algebraic structure<br />
very close to topology: roughly speaking, it associates to each surface a finitedimensional<br />
vector space and to each cobordism a linear map between the vector<br />
spaces associated to the boundaries. Such theories have physical origins: they were<br />
introduced by Witten [W] in the 1980s from Chern-Simons actions and generated<br />
very rich mathematical developments. There are various rigorous constructions coming<br />
from quantum groups (see [RT]), geometric quantization, and many other areas.<br />
Unfortunately, such constructions remain complicated, and it is hard to make concrete<br />
computations.<br />
In this article, we prefer the approach of [BH+2], which defines TQFTs in a<br />
purely combinatorial way: using skein theory and the Kauffman bracket, the authors<br />
define a family of Hermitian TQFTs (V p , 〈·, ·〉 p ) corresponding for p = 2r to the<br />
SU(2)-theory at level r − 2. Despite the simple and very beautiful structure of these<br />
combinatorial TQFTs, the connection with geometry is less clear than from other<br />
approaches. In this article, we show that some connections can be found in a simple<br />
and direct way. From the axioms, a TQFT generates for any closed surface a family<br />
of representations of the extended mapping class group of , a central extension of<br />
the mapping class group by Z coming from p 1 -structures (see [MR]). In some sense,<br />
these representations carry the main topological meaning of TQFTs. Hence we link<br />
DUKE MATHEMATICAL JOURNAL<br />
Vol. 141, No. 3, c○ 2008 DOI 10.1215/00127094-2007-006<br />
Received 7 June 2006. Revision received 15 June 2007.<br />
2000 Mathematics Subject Classification. Primary 57M27; Secondary 57M50, 37E30.<br />
573
574 MARCHÉ and NARIMANNEJAD<br />
them with some geometrical representation. The basic idea for this comes from a<br />
general belief that when p goes to infinity, things become classical, by which we<br />
mean geometrical. This belief is based on the so-called semiclassical approximation.<br />
Hence we study the limit of ρ p , the quantum representations of Ɣ g on V p ().<br />
For this purpose, let us describe two classical spaces on which the mapping class<br />
group acts.<br />
Fix a closed oriented surface of genus g. We call multicurve an isotopy class of<br />
1-dimensional submanifold of without component bounding a disc. (The empty set<br />
is also considered as a multicurve.) The mapping class group of Ɣ g acts on the set of<br />
multicurves in a natural way. Call C() the C vector space generated by multicurves;<br />
we obtain a representation of Ɣ g on C(). This fundamental representation carries<br />
almost all information about the structure of Ɣ g . For instance, no nontrivial element<br />
of Ɣ g acts trivially on multicurves, except for the elliptic and hyperelliptic involutions<br />
in genus 1 and 2.<br />
Another very natural space on which the mapping class group acts is<br />
hom(π 1 (),G)/G, theG-character variety of π 1 () for a fixed Lie group G. Let<br />
us denote it by S(,G). Any element of the mapping class group of may be<br />
represented as an automorphism of π 1 (); its action on S(,G) is then obtained<br />
by left composition on hom(π 1 (),G). In that way, we also obtain an action of<br />
the mapping class group on any space of functions defined on the G-character<br />
variety.<br />
We are interested here in the cases where G = SU(2) and G = SL(2, C).<br />
These spaces have a rich structure; we use the natural symplectic structure ω on<br />
the smooth part of S(,SU(2)) (see [G1]) and the structure of an algebraic variety<br />
on S(,SL(2, C)). WedefineH() as the ring of regular functions on S(,<br />
SL(2, C)).<br />
Using the natural inclusion of S(,SU(2)) in S(,SL(2, C)), we can define a<br />
Hermitian form on H() by the formula<br />
∫<br />
〈f, g〉 = f gdV.<br />
S(,SU(2))<br />
Here, dV is the volume form on S(,SU(2)) induced by the symplectic form ω.<br />
The following theorem can be interpreted in terms of the Fell topology. For<br />
convenience, let us recall this notion briefly. Let G be a discrete group, and let<br />
ρ k : G → U(V k ) be a sequence of unitary representations of G into V k . One says that<br />
this sequence converges to the representation ρ : G → U(V ) in the Fell topology<br />
if, for any unit vector v ∈ V and any finite subset S ⊂ G, there is a sequence of<br />
unit vectors v k ∈ V k such that for all g ∈ S, the sequence 〈ρ k (g)v k ,v k 〉 converges to<br />
〈ρ(g)v, v〉.<br />
We obtain the following result.
SOME ASYMPTOTICS OF TQFT VIA SKEIN THEORY 575<br />
THEOREM<br />
Let be a closed oriented surface of genus g. For all even integers p, thereisa<br />
Ɣ g -equivariant map ϕ p : H() → End(V p ()) such that<br />
2<br />
) d(g)〈ϕp<br />
〈v, w〉 = lim (v),ϕ<br />
p→∞(<br />
p (w)〉 p for all v, w ∈ H().<br />
p<br />
Here, we have set d(1) = 1 and d(g) = 3g − 3 for g>1. The Hermitian form on<br />
End(V p ()) is defined by 〈x,y〉 p = Tr(xy ∗ ). This implies, in particular, that the quantum<br />
representations ρ p ⊗ ρ p converge in the Fell topology to ρ : Ɣ g → U(H()),<br />
the natural representation coming from the action of Ɣ g on S(,SL(2, C)).<br />
We obtain as a corollary a new proof of the following result of [A1](seealso[FWW])<br />
about asymptotic faithfulness of quantum representations.<br />
COROLLARY<br />
Let be a closed oriented surface of genus g. For any nontrivial h in Ɣ g which is not<br />
the elliptic (g = 1) or hyperelliptic (g = 2) involution, there is some even p 0 such<br />
that ρ p (h) ≠ Id for all even p ≥ p 0 .<br />
Proof<br />
One can associate to any curve γ on a regular function f γ on S(,SL(2, C))<br />
by the formula f γ (ρ) =−Tr ρ(γ ). For a disjoint union of curves, we associate the<br />
product of the functions associated to each component. In this way, we construct a<br />
map f from C() to H(). By a result of [B, Theorem 10] and [PS, Theorem 4.7],<br />
the map f is an isomorphism of vector spaces. Therefore, we can think of a regular<br />
function on S(,SL(2, C)) as a linear combination of multicurves.<br />
Recall that no element of Ɣ g acts trivially on C() except the identity and the<br />
elliptic and hyperelliptic involutions in genus 1 and genus 2. Hence we can suppose<br />
that there is some v in C() ≃ H() such that w = hv − v is nonzero. This implies<br />
that 〈w, w〉 is nonzero because the form 〈·, ·〉 is nondegenerate.<br />
In fact, if 〈w, w〉 =0, then the regular function on S(,SL(2, C)) associated<br />
to w satisfies<br />
∫<br />
|w| 2 dV = 0.<br />
S(,SU(2))<br />
As w is continuous, it must vanish on S(,SU(2)). Moreover, as it is holomorphic<br />
on the space S(,SL(2, C)) and zero on S(,SU(2)), it vanishes identically (see<br />
[G2, proof of Theorem 1.4.1]).<br />
Due to the equality 〈w, w〉 =lim p→∞ (2/p) d(g) 〈ϕ p w, ϕ p w〉 p , we can find even<br />
p 0 such that for all even p ≥ p 0 , ϕ p w ≠ 0. Hence ϕ p (hv) ≠ ϕ p (v),andρ p (h) cannot<br />
be the identity.
576 MARCHÉ and NARIMANNEJAD<br />
1.1. Proof of the theorem<br />
The heart of the proof is the construction of the map ϕ p , which is almost obvious but<br />
is fundamental. As the space H() is isomorphic to C(), to define the map ϕ p it is<br />
sufficient to construct ϕ p (γ ) ∈ V p () ⊗ V p () ∗ for any multicurve γ .<br />
For such a multicurve, we consider the cobordism × [0, 1] with the multicurve<br />
embedded as γ ×{1/2}. The TQFT naturally induces an element Z p ( × [0, 1],γ)<br />
in V p (∐ − ) = V p () ⊗ V p () ∗ . We call this element ϕ p (γ ). We remark that<br />
it defines a self-adjoint element of End(V p ()) as the pair made of the cobordism<br />
× [0, 1] and the curve γ inside is isomorphic to the same pair with opposite<br />
orientations and exchanged boundaries. This gives our fundamental map ϕ p ,whichis<br />
clearly equivariant because of the naturality of the construction.<br />
To prove the theorem, one has to compute the limit of the expression<br />
(2/p) d(g) 〈ϕ p (γ ),ϕ p (δ)〉 p for two multicurves γ and δ.<br />
We do this in two steps. In the first step, we assume that δ is empty. Using<br />
combinatorial techniques from [BH+2], we obtain for 〈ϕ p (γ ), 1〉 p an explicit formula<br />
resembling a Riemann sum. When we normalize it, it converges to an integral over a<br />
subspace of R d(g) ; we denote its value by 〈γ 〉. By linearity, we extend 〈·〉 to a map<br />
from C() to C.<br />
In the second step, we use the connection between the TQFT V p and the Kauffman<br />
skein module at A =−e iπ/p . We easily find that (2/p) d(g) 〈ϕ p (γ ),ϕ p (δ)〉 p converges<br />
to 〈γ · δ〉, where · is the multiplication induced on C() by its identification with the<br />
Kauffman skein algebra of × [0, 1] at A =−1 (see [PS]).<br />
On the other hand, it is well known that this multiplication on C() is isomorphic<br />
to the natural multiplication on H(), the space of regular functions on<br />
S(,SL(2, C)) (see [B], [PS]).<br />
It remains to identify the linear form on H() defined by f γ ↦→〈γ 〉. Suppose<br />
that γ is a multicurve. We choose curves C i on which decompose the surface<br />
into pants such that all components of γ are parallel to some C i .Itiswellknown<br />
that the maps f i = f Ci form a system of Poisson commuting functions on S(,<br />
SU(2)).<br />
As shown in [JW], the product of the maps (f i ):S(,SU(2)) → R d(g) is the<br />
moment map for an action of a torus of dimension d(g) on a dense open subset of<br />
S(,SU(2)). The authors use the Duistermaat-Heckman theorem to give an explicit<br />
formula for the volume form dV on S(,SU(2)). From their result, we deduce the<br />
following striking formula:<br />
∫<br />
〈γ 〉=<br />
S(,SU(2))<br />
The theorem follows from this formula.<br />
f γ dV.
SOME ASYMPTOTICS OF TQFT VIA SKEIN THEORY 577<br />
1.2. Remarks and perspectives<br />
The main motivation for this work came from the article [FK], which is about the<br />
asymptotics of quantum representations of the mapping class group of the torus. Our<br />
approach is different in the sense that we study the limit of V p ⊗ Vp ∗ instead of simply<br />
V p . We were also inspired by the ideas contained in the works [F]and[M]. Our work<br />
is, of course, related to the article [A1], where similar ideas appear, and also has some<br />
intersection with [BFK].<br />
Questions<br />
There are many questions naturally linked to our results.<br />
(1) How can we link our asymptotic result to the asymptotics considered in [FK]?<br />
(2) Can we apply our result or some refinements thereof to the problem of<br />
Andersen, Masbaum, and Ueno in [AMU, Question 1.1]? The Nielsen-<br />
Thurston classification of the elements of the mapping class group is directly<br />
related to their action on multicurves. As the quantum representations converge<br />
to this action, can we find some trace of this classification in quantum representations<br />
at finite level? Some ideas with respect to this question are developed<br />
in [A3].<br />
(3) In [BH+2], one can choose any primitive 4r root of unity to construct a TQFT.<br />
We have chosen roots converging to −1. Is it possible to develop the same<br />
asymptotics for roots of unity converging to different complex numbers?<br />
(4) Can we obtain a stronger convergence for the sequence involved in the theorem?<br />
For instance, what is the expansion of this sequence into powers of 1/p?<br />
The results of this article were recovered and generalized to the SU(n)-case via the<br />
theory of Toeplitz operators in a purely geometric framework (see [A2]).<br />
2. Review of TQFT<br />
This part is a quick and formal review of the TQFT constructed in [BH+2]whichwe<br />
give to fix notation and settings and to recall results that are used in this article. We<br />
refer the interested reader to that beautiful original article.<br />
Fixanevenintegerp = 2r. The complex number A =−e iπ/(2r) is a primitive<br />
4rth root of unity. One can construct from it a 2 + 1 TQFT.<br />
In the notation of [BH+2], we set κ = e −iπ/(2r)−iπ(2r+1)/12 and η =<br />
( √ 2/r) sin(π/r). WedefineC r ={0, 1,...,r − 2} to be the set of colors.<br />
Atriple(a,b,c) of elements of C r is called r-admissible if a + b + c is even,<br />
the triangle inequality |a − b| ≤c ≤ a + b is satisfied, and, moreover, we have<br />
a + b + c
578 MARCHÉ and NARIMANNEJAD<br />
2.1. The cobordism category<br />
A TQFT is a linear representation of a cobordism category. In our settings, the objects<br />
of our category are oriented surfaces with marked points and p 1 -structures.<br />
• A marking of a surface is a finite family (z j ,c j ) j∈J , where (z j ) is a family<br />
of distinct points in with, for all j ∈ J , a nonzero tangent direction v j at z j<br />
on . Forallj ∈ J , c j is a color in C r .<br />
• A p 1 -structure is a somewhat complicated object used to solve the so-called<br />
framing anomaly. Consider the map p 1 : BO → K(Z, 4) corresponding to<br />
the first Pontryagin class. Let X be its homotopy fiber, that is, the set of couples<br />
(x,γ) ∈ BO × C([0, 1],K(Z, 4)) satisfying γ (0) =∗,andγ (1) = p 1 (x).<br />
Let E be the universal stable bundle over BO,andletE X be its pullback over<br />
X. Ap 1 -structure on a manifold M is a fiber map from the stable tangent<br />
bundle of M to E X .<br />
In the notation of an object (,z,c), we do not mention the directions v j and the<br />
p 1 -structure, although they are present.<br />
Now, we define morphisms. Let ( 1 ,z 1 ,c 1 ) and ( 2 ,z 2 ,c 2 ) be two objects as<br />
defined above. A morphism is<br />
• an oriented 3-manifold M whose boundary is decomposed as ∂M =− 1 ∐ 2 ,<br />
where − means with opposite orientation;<br />
• a colored, banded trivalent graph G embedded in M whose restriction to the<br />
boundary is compatible with the marked points;<br />
• a p 1 -structure on M extending the p 1 -structure given on the boundary.<br />
A banded trivalent graph G in M is a graph with monovalent or trivalent vertices<br />
contained in an oriented surface SG ⊂ M such that<br />
(i) G meets ∂M transversally on the set of 1-valent vertices of G noted ∂G;<br />
(ii) the surface SG is a regular neighborhood of G in SG, andSG ∩ ∂M is a<br />
regular neighborhood of G ∩ ∂M in SG ∩ ∂M.<br />
A coloring of G is a map σ from the set of edges of G to C r such that the colors of<br />
the edges meeting at each vertex are r-admissible. The restriction of a banded graph<br />
G ⊂ M on ∂M gives marked points (z j ) j∈J with tangent directions (v j ) j∈J , whereas<br />
the restriction of a coloring gives colors (c j ) j∈J .<br />
Two morphisms are called equivalent if the corresponding manifolds are isomorphic,<br />
the banded graphs are isotopic, and the p 1 -structures are homotopic relative to<br />
the boundary.<br />
2.2. Main properties of TQFT<br />
Theorem 1.4 in [BH+2] states that for each integer p, there is a functor (V p ,Z p )<br />
from the precedent cobordism category to the category of finite-dimensional C vector<br />
spaces.
SOME ASYMPTOTICS OF TQFT VIA SKEIN THEORY 579<br />
This means that to every object (,z,c), we can associate a vector space<br />
V p (,z,c), and to any morphism (M,G) between two objects, we can associate<br />
a linear map Z p (M,G) between the vector spaces corresponding to the objects. By<br />
convention, V p (∅) = C; hence any closed manifold (M,G) acts as a scalar 〈M,G〉 p<br />
that is a 3-manifold invariant. Moreover, there is natural Hermitian form 〈·, ·〉 p on<br />
V p (,z,c) such that for any two morphisms (M 1 ,G 1 ) and (M 2 ,G 2 ) from ∅ to<br />
(,z,c),wehave〈Z p (M 1 ,G 1 ),Z p (M 2 ,G 2 )〉 p =〈M 1 ∪ (−M 2 ),G 1 ∪ G 2 〉 p .<br />
We give here some important results related to this construction.<br />
THEOREM 2.1 ([BH+1, Theorem 4.11])<br />
Let (,z,c) be a surface with marked points and p 1 -structure. Let H be a handlebody<br />
whose boundary is and with a p 1 -structure extending that of . LetG be a<br />
banded graph with monovalent or trivalent vertices in H such that monovalent vertices<br />
correspond to marked points z and such that H is a tubular neighborhood of G. For<br />
each coloring σ of G compatible with the coloring of the boundary, we denote by u σ<br />
the element induced by Z p in V p (,z,c).<br />
Then the elements u σ form an orthogonal basis of V p (,z,c), and if G does not<br />
contain any closed loop, we have<br />
∏<br />
〈u σ ,u σ 〉 p = η #v−#e ∏ v〈σ v〉<br />
e 〈σ e〉 .<br />
In this formula, v ranges over the set of vertices of G, and e ranges over the set of<br />
edges. Moreover, for any trivalent vertex v, σ v is the triple of colors of the edges<br />
adjacent to this vertex, and for any monovalent vertex v, σ v is the color of the edge<br />
incoming to it.<br />
We set 〈j〉 =(−1) j [j + 1] and 〈a,b,c〉 =(−1) α+β+γ ([α + β + γ + 1]![α]!<br />
[β]![γ ]!/([a]![b]![c]!)), whereα, β, and γ are defined by the equations a = β +<br />
γ,b = α + γ, and c = α + β.<br />
If G is reduced to a closed loop, then the formula is simply 〈u σ ,u σ 〉 p = 1.<br />
Remark 2.2<br />
We check that for our choice of root of unity A and for a surface without marked<br />
points, the Hermitian pairing on V p () is positive definite.<br />
2.3. Kauffman bracket and TQFT<br />
We define K(M) as the usual skein module of any oriented 3-manifold M. We refer<br />
to [PS] for a complete account, but we recall here what we need. Let A be some<br />
indeterminate. The Z[A, A −1 ]-module K(M) is the free module generated by isotopy<br />
classes of banded links in M including the empty link, ∅, quotiented by the submodule<br />
generated by the local relations of Figure 1.<br />
For any u ∈ C \{0}, wesetK(M,u) = K(M) ⊗ Z[A,A −1 ] C, where A acts on C<br />
by multiplication by u.
580 MARCHÉ and NARIMANNEJAD<br />
Figure 1. Kauffman relations<br />
The following proposition is a consequence of the construction of the TQFT.<br />
PROPOSITION 2.3 ([BH+2, Proposition 1.9])<br />
Let M be an oriented connected 3-manifold with p 1 -structure and boundary (without<br />
boundary or marked points). Then there is a surjective map from K(M,−e iπ/p )<br />
to V p ().<br />
This map is defined by sending the element L ⊗ 1 to Z p (M,L), whereL is<br />
considered as a banded link with color 1.<br />
3. Convergence of TQFT<br />
3.1. Settings<br />
Let be a closed oriented surface of genus g with p 1 -structure. We denote by Ɣ g the<br />
mapping class group of .Fixp = 2r.<br />
If h is an element of Ɣ g , we can construct a cobordism C h from to itself as<br />
× [0, 1], where we identify the first boundary component with using the identity<br />
and the second one using h. Ifh ′ is another element of Ɣ g , the cobordisms C h ◦ C h ′<br />
and C hh ′ are diffeomorphic. We should obtain a representation of Ɣ g on V p () by<br />
considering the linear map Z p (C h ). The problem is that we have not chosen any<br />
p 1 -structure on C h , and we cannot make a canonical choice.<br />
One way to get rid of this annoying fact is to consider the action of Ɣ g on V p ()⊗<br />
V p () ∗ = V p (∐ − ). The action of h on this space is given by Z p (C h ∐ − C h ),<br />
where we choose any p 1 -structure on C h and put the same one on −C h . The action<br />
does not depend any further on the p 1 -structure: in fact, if we change the p 1 -structure<br />
in a cobordism M, the linear map Z p (M) is changed by a multiple of κ, a root of<br />
unity. When we take the dual, the root becomes its conjugate. Hence the two anomalies<br />
cancel, and we get a true representation of Ɣ g .<br />
We thus obtain a sequence of representations (V p () ⊗ V p () ∗ ,Z p ⊗ Zp ∗) of<br />
Ɣ g , and we want to find their limit in some sense. The problem is that the spaces on
SOME ASYMPTOTICS OF TQFT VIA SKEIN THEORY 581<br />
which the mapping class group acts are a priori completely different. We need a way<br />
to compare them; this is suggested by Proposition 2.3.<br />
Given a multicurve γ in , one can give it a banded structure by taking a<br />
neighborhood of it in . We can consider the curve γ as a banded link in × [0, 1]<br />
by sending it to γ ×{1/2}. We use the same notation for the multicurve on and its<br />
associated banded link in × [0, 1].<br />
In [PS], it is shown that the Kauffman skein module K( × [0, 1]) is a free<br />
Z[A, A −1 ]-module with a basis of the isotopy classes of multicurves. It provides an<br />
isomorphism of vector spaces between C() and K( ×[0, 1],u) for any u in C\{0}.<br />
In particular, using Proposition 2.3, we get a surjective map<br />
ϕ p : C() ∼ → K( × [0, 1], −e iπ/p ) → V p (∐ − ).<br />
THEOREM 3.1<br />
Let be a closed oriented surface of genus g. There is a Hermitian pairing 〈·, ·〉 on<br />
C() such that for all x and y in C(), the following holds, where d(1) = 1 and<br />
d(g) = 3g − 3 for g>1:<br />
2<br />
) d(g)〈ϕp<br />
〈x,y〉= lim (x),ϕ<br />
p→∞(<br />
p (y)〉 p .<br />
p<br />
3.2. The trace function<br />
Definition 3.2<br />
Let be a closed oriented surface of genus g, andletγ be a multicurve on . We<br />
set Tr p (γ ) =〈 × S 1 ,γ〉 p . Here, γ is seen as a banded link with color 1 lying in the<br />
slice ×{1/2} of × S 1 .<br />
LEMMA 3.3<br />
Suppose that a surface is presented as the boundary of a handlebody H which<br />
retracts on a trivalent banded graph G as in Theorem 2.1. We choose meridian discs<br />
D e transverse to each edge of G and define C e = ∂D e ; the curves C e are disjoint on<br />
. We choose a nonnegative integer m e for each edge of G.<br />
Then we define γ as the multicurve on obtained by taking m e parallel copies<br />
of C e for each edge of G. We have<br />
Tr p (γ ) = ∑ σ<br />
∏<br />
e<br />
[ ( (σe + 1)π<br />
−2 cos<br />
r<br />
)] me.<br />
Here, σ ranges over r-admissible colorings of G, and e ranges over edges of G.
582 MARCHÉ and NARIMANNEJAD<br />
Figure 2. Action of a curve in TQFT<br />
Proof<br />
The proof is an easy consequence of the following fact from skein theory: a trivial<br />
curve colored with 1 and making a Hopf link with a curve colored with j may be<br />
removed and replaced by a factor −A 2j+2 −A −2j−2 =−2 cos((j + 1)π/r). We refer,<br />
for instance, to [BH+1, Lemma 3.2].<br />
We use the general trace formula of TQFT (see, e.g., [BH+2, (1.2)]). Let M<br />
be a cobordism from to , andletƔ be a colored banded graph in M. LetM <br />
be the closed manifold obtained from M by identifying the two copies of . Then<br />
〈M ,Ɣ〉 p = Tr Z p (M,Ɣ).<br />
Consider the basis u σ = Z p (H,G σ ) of V p () involved in Theorem 2.1. For<br />
each curve C e , the cobordism ( × [0, 1],C e ) acts on u σ by multiplication with<br />
−2 cos((σ e + 1)π/r), as suggested in Figure 2.<br />
∏<br />
e<br />
(<br />
Then the cobordism ( × [0, 1],γ) acts on u σ by multiplication with<br />
−2 cos((σe + 1)π/r) ) m e. The formula for Trp (γ ) comes now from the trace<br />
formula of TQFT.<br />
3.3. Limit of the trace function<br />
As before, fix a surface , presented as in Theorem 2.1, as the boundary of a handlebody<br />
H which retracts on a trivalent banded graph G.<br />
The number of edges of G is 3g − 3 if g>1 or 1 if g = 1. We denote this<br />
number by d(g) and consider the subset U g of R d(g) consisting of all maps τ from the<br />
set of edges of G to [0, 1] such that for all triples of incoming edges (e, f, g) of some<br />
vertex, the following relations are satisfied:<br />
−|τ f − τ g |≤τ e ≤ τ f + τ g ,<br />
−τ e + τ f + τ g ≤ 2.<br />
We use the formula of Lemma 3.3 to deduce the asymptotics of the trace function.<br />
LEMMA 3.4<br />
With the same hypothesis as in Lemma 3.3, letF γ : U g → R be the map defined by<br />
F γ (τ) = ∏ e (−2 cos(τ eπ)) m e<br />
. Then the following formula holds:<br />
2<br />
) d(g)<br />
∫<br />
lim Trp (γ ) = 2<br />
r→∞( g−d(g) F γ (τ) dτ.<br />
p<br />
U g
SOME ASYMPTOTICS OF TQFT VIA SKEIN THEORY 583<br />
Proof<br />
The formula for Tr p (γ ) looks like a Riemannian sum; hence the result is not a surprise.<br />
To obtain the precise result, we have to decompose U g into small pieces parametrized<br />
by r-admissible colorings σ .<br />
Given a positive integer r and any coloring σ from the set of edges of G to C r ,<br />
we define the set A r σ<br />
= ∏ e [σ e/r,σ e + 1/r) ⊂ R d(g) .Asσ runs over r-admissible<br />
colorings of G, these sets do not cover U g because of the parity condition. We have to<br />
pack some sets A r σ<br />
together, which we do in the following way.<br />
We denote by C 1 (G, Z 2 ) the Z 2 vector space of 1-cochains of G with Z 2 -<br />
coefficients. The subspace of 1-cycles is denoted by Z 1 (G, Z 2 ). Choose a subspace<br />
S of C 1 (G, Z 2 ) so that C 1 (G, Z 2 ) = S ⊕ Z 1 (G, Z 2 ). The subspace S has dimension<br />
d(g) − g. For an admissible coloring σ of G, wedefineBσ r = ⋃ ρ∈S Ar σ +ρ<br />
. Here, we<br />
have identified Z/2Z with the set {0, 1}. The sets Bσ r are disjoint and almost cover<br />
U g .<br />
Let us prove that they are disjoint. Suppose that we have σ + ρ = σ ′ + ρ ′ with σ<br />
and σ ′ admissible and ρ,ρ ′ in S; then consider these maps modulo 2. If we apply the<br />
boundary map, the admissible colorings vanish by definition, and we have ∂ρ = ∂ρ ′ .<br />
But ∂ induces a bijection from S onto its image; hence we have ρ = ρ ′ , and it follows<br />
that σ = σ ′ . Hence the sets Bσ r are actually disjoint. Moreover, the measure of Br σ is<br />
2 d(g)−g /r d(g) . It follows that ∑ ∫<br />
σ,r−admissible F γ (σ e + 1/r)(2 g−d(g) /r d(g) ) converges to<br />
U g<br />
F γ (τ) dτ, and the result is proved.<br />
<br />
3.4. Proof of Theorem 3.1<br />
Let be a closed oriented surface of genus g. We recall that C() and K(×[0, 1],u)<br />
are isomorphic as vector spaces for any u in C \{0}. The stacking product induces on<br />
K( × [0, 1]) a natural algebra structure that induces an algebra structure on C()<br />
for each u ∈ C \{0}. We consider the algebra structure obtained for u =−1.<br />
Fix γ and δ, two multicurves on . We aim to compute the limit of the sequence<br />
(2/p) d(g) 〈ϕ p (γ ),ϕ p (δ)〉 p as p goes to infinity. The right-hand side is the quantum<br />
invariant of two thickened surfaces with a multicurve inside, glued along their<br />
boundary. Instead of gluing the two boundaries simultaneously, we glue one and then<br />
the other. If we glue one boundary component, we obtain the stacking product of γ<br />
and δ. In the skein module for generic A, we have a decomposition γ · δ = ∑ i c iζ i for<br />
some multicurves ζ i and some Laurent polynomials c i in Z[A, A −1 ]. When evaluating<br />
this combination in V p (∐ − ), wehavetospecializeA to −e iπ/p . In formulas,<br />
we have ϕ p (γ · δ) = ∑ i c i(−e iπ/p )ϕ p (ζ i ). Then, we glue together the remaining<br />
boundary components and obtain 〈ϕ p (γ ),ϕ p (δ)〉 p = ∑ i c i(−e iπ/p )Tr p (ζ i ).<br />
The asymptotic formula becomes clear if we define the following linear form on<br />
C().
584 MARCHÉ and NARIMANNEJAD<br />
Definition 3.5<br />
Let γ be a multicurve on . Then, there is a pants decomposition associated to γ<br />
such that all components of γ are parallel copies of the boundary circles. We define<br />
〈γ 〉=2 g−d(g) ∫ U g<br />
F γ (τ) dτ, where F γ (τ) = ∏ e (−2 cos(τ eπ)) m e<br />
. The expression of<br />
〈γ 〉 as a limit shows that this definition does not depend on the pants decomposition.<br />
We extend 〈·〉 to a linear form on C().<br />
Coming back to our computation, we obtain lim p→∞ (2/p) d(g) 〈ϕ p (γ ),ϕ p (δ)〉 p =<br />
∑<br />
i c i(−1)〈ζ i 〉=〈γδ〉. Finally, we define a Hermitian form on C() by the formula<br />
〈x,y〉 = 〈xy〉, where the product corresponds to the skein module product<br />
for A = −1, and the conjugation corresponds to conjugation of coefficients<br />
in C(). We have proved the following result: for all x,y ∈ C(), wehave<br />
〈x,y〉=lim p→∞ (2/p) d(g) 〈ϕ p (x),ϕ p (y)〉 p .<br />
4. Geometric interpretation<br />
The heart of the following geometric interpretation is the theorem of [B, Theorem 10]<br />
and [PS, Theorem 4.7] stating that the algebra K( × [0, 1], −1) is isomorphic to<br />
H(), the ring of regular functions on the SL(2, C)-character variety of . Recall<br />
that the isomorphism is given by f γ (ρ) =−Tr(ρ(γ )) when γ is a connected curve<br />
on and ρ : π 1 () → SL(2, C) is a representation of π 1 ().<br />
This identifies C() with its algebra structure. It remains to identify the linear<br />
form 〈·〉 of Definition 3.5.<br />
Recall that the SL(2, C)-character variety contains the SU(2)-character variety,<br />
which carries a natural symplectic form ω defined in [AB], [G1, page 208]. Following<br />
[JW, page 154], we define S g to be the moduli space of irreducible representations of<br />
π 1 () on SU(2) and S g to be the moduli space of all representations. Then it is known<br />
that S g is a smooth 2d(g)-manifold with symplectic form ω obtained by symplectic<br />
reduction from the form ω(a,b) = (1/(4π 2 )) ∫ Tr(a ∧ b) for a,b ∈ 1 (,su(2)).<br />
We denote the volume form on S g by dV = ω d(g) /(d(g)!).<br />
PROPOSITION 4.1<br />
For all multicurves γ on , we have<br />
∫<br />
〈γ 〉= f γ dV.<br />
S g<br />
Proof<br />
We give a proof of this proposition by adapting the results of [JW].<br />
Fix a pants decomposition of associated to γ , and denote the set of curves<br />
bounding the pants by C e . We define the functions h e on S g with values in [0, 1] by<br />
the formula Tr ρ(C e ) = 2 cos(πh e (ρ)).
SOME ASYMPTOTICS OF TQFT VIA SKEIN THEORY 585<br />
Where the functions h e are not equal to 0 or 1, they Poisson commute, and their<br />
Hamiltonian flows define a torus action on S g . In fact, we have the following theorem.<br />
THEOREM 4.2 ([JW, Propositions 3.8, 4.1])<br />
Let Ug<br />
gen be the interior of U g in R d(g) , and let h = (h 1 ,...,h d(g) ):S g → R d(g) be<br />
the collection of the h e -functions.<br />
For x in S g such that h(x) = y ∈ Ug<br />
gen , the torus action identifies h −1 (y)<br />
with U(1) d(g) /Z 2g−2<br />
2 , where an element (ε v ) ∈ Z 2g−2<br />
2 acts on U(1) d(g) by the formula<br />
e 2iπx e ↦→ (−1)ε v+ε v′<br />
e 2iπx e<br />
,wherev and v ′ are the indices of the pants bounding C e .<br />
If we choose a Lagrangian submanifold L of S g transverse to the fibers of the<br />
torus action and which h maps diffeomorphically onto V ⊂ Ug<br />
gen , then we can define<br />
canonical coordinates on h −1 (V ) by setting x e = 0 on L and y e = h e .<br />
The volume form is given on h −1 (V ) by ∏ ∏<br />
dy e dxe .<br />
We come back to the integral of the function associated to γ on the moduli space<br />
S g . Recall that γ was adapted to the pants decomposition. This means that γ is the<br />
union of parallel curves C e with multiplicity m e . The function f γ is then defined by<br />
f γ (ρ) = ∏ e (− Tr ρ(C e)) m e<br />
= ∏ (<br />
e −2 cos(πhe (ρ)) ) m e<br />
= Fγ (h), where F γ is the<br />
function of Lemma 3.4.<br />
As this function depends only on the values of h, we can perform the integration<br />
on its fiber first. The fiber is isomorphic to U(1) d(g) /Z 2g−2<br />
2 . Hence U(1) d(g) is a<br />
Riemannian covering over the fiber and has volume equal to 1. To find the volume<br />
of the fiber, it is then sufficient to find the degree of this covering. Let G be the<br />
graph associated to the pants decomposition. The degree of the covering is equal to<br />
the dimension of the Z 2 -subspace of C 1 (G, Z 2 ) generated by the family of vectors<br />
u v = e a + e b + e c for each pant v bounding circles a,b, andc. This subspace is<br />
the image of the coboundary map d : C 0 (G, Z 2 ) → C 1 (G, Z 2 ). Its dimension is<br />
then complementary to the dimension of H 1 (G, Z 2 ),whichisg. We find that the<br />
dimension is d(g) − g; hence the covering has degree 2 d(g)−g , and the volume of the<br />
fiber is 2 g−d(g) .<br />
We finally obtain ∫ S g<br />
f γ dV = 2 ∫ g−d(g) U g<br />
F γ (τ) dτ =〈γ 〉, which completes the<br />
proof.<br />
<br />
Acknowledgment. We thank Gregor Masbaum for his remarks, encouragement, and<br />
simplification of the proof of Lemma 3.3.<br />
References<br />
[A1]<br />
J. E. ANDERSEN, Asymptotic faithfulness of the quantum SU(n) representations of<br />
the mapping class groups, Ann. of Math. (2) 163 (2006), 347 – 368.<br />
MR 2195137 575, 577
586 MARCHÉ and NARIMANNEJAD<br />
[A2]<br />
[A3]<br />
[AMU]<br />
[AB]<br />
[BH+1]<br />
[BH+2]<br />
[B]<br />
[BFK]<br />
[F]<br />
[FK]<br />
[FWW]<br />
[G1]<br />
[G2]<br />
[JW]<br />
[M]<br />
[MR]<br />
[PS]<br />
———, Asymptotics of the Hilbert-Smith norm of curve operators in TQFT,<br />
preprint, arXiv:math/0605291v1 [math.QA] 577<br />
———, The Nielsen-Thurston classification of mapping classes is determined by<br />
TQFT, preprint, arXiv:math/0605036v1 [math.QA] 577<br />
J. E. ANDERSEN, G. MASBAUM,andK. UENO, Topological quantum field theory<br />
and the Nielsen-Thurston classification of M(0, 4), Math. Proc. Cambridge<br />
Philos. Soc. 141 (2006), 477 – 488. MR 2281410 577<br />
M. F. ATIYAH and R. BOTT, The Yang-Mills equations over Riemann surfaces,<br />
Philos. Trans. Roy. Soc. London Ser. A 308 (1983), 523 – 615.<br />
MR 0702806 584<br />
C. BLANCHET, N. HABEGGER, G. MASBAUM,andP. VOGEL, Three-manifold<br />
invariants derived from the Kauffman bracket, Topology 31 (1992),<br />
685 – 699. MR 1191373 579, 582<br />
———, Topological quantum field theories derived from the Kauffman bracket,<br />
Topology 34 (1995), 883 – 927. MR 1362791 573, 576, 577, 578, 580, 582<br />
D. BULLOCK, Rings of SL 2 (C)-characters and the Kauffman bracket skein<br />
module, Comment. Math. Helv. 72 (1997), 521 – 542. MR 1600138 575,<br />
576, 584<br />
D. BULLOCK, C. FROHMAN,andJ. KANIA-BARTOSZYNSKA, The Yang-Mills<br />
measure in the Kauffman bracket skein module, Comment. Math. Helv. 78<br />
(2003), 1 – 17. MR 1966749 577<br />
M. H. FREEDMAN, A magnetic model with a possible Chern-Simons phase,<br />
Comm. Math. Phys. 234 (2003), 129 – 183. MR 1961959 577<br />
M. H. FREEDMAN and V. KRUSHKAL, On the asymptotics of quantum SU(2)<br />
representations of mapping class groups, Forum Math. 18 (2006), 293 – 304.<br />
MR 2218422 577<br />
M. H. FREEDMAN, K. WALKER,andZ. WANG, Quantum SU(2) faithfully detects<br />
mapping class groups modulo center, Geom. Topol. 6 (2002), 523 – 539.<br />
MR 1943758 575<br />
W. M. GOLDMAN, The symplectic nature of fundamental groups of surfaces,<br />
Adv. in Math. 54 (1984), 200 – 225. MR 0762512 574, 584<br />
———, “The complex-symplectic geometry of SL(2, C)-characters over<br />
surfaces” in Algebraic Groups and Arithmetic (Mumbai, 2001), TataInst.<br />
Fund. Res., Mumbai, 2004, 375 – 407. MR 2094117 575<br />
L. C. JEFFREY and J. WEITSMAN, Toric structures on the moduli space of flat<br />
connections on a Riemann surface: Volumes and the moment map,<br />
Adv. Math. 106 (1994), 151 – 168. MR 1279216 576, 584, 585<br />
G. MASBAUM, “Quantum representations of mapping class groups” in Groupes et<br />
géométrie, SMF Journ. Annu. 2003, Soc. Math. France, Montrouge, 2003,<br />
19 – 36. MR 2202283 577<br />
G. MASBAUM and J. D. ROBERTS, On central extensions of mapping class groups,<br />
Math. Ann. 302 (1995), 131 – 150. MR 1329450 573<br />
J. H. PRZYTYCKI and A. S. SIKORA, On skein algebras and Sl 2 (C)-character<br />
varieties, Topology 39 (2000), 115 – 148. MR 1710996 575, 576, 579,<br />
581, 584
SOME ASYMPTOTICS OF TQFT VIA SKEIN THEORY 587<br />
[RT]<br />
[R]<br />
[W]<br />
N. RESHETIKHIN and V. G. TURAEV, Invariants of 3-manifolds via link<br />
polynomials and quantum groups, Invent. Math. 103 (1991), 547 – 597.<br />
MR 1091619 573<br />
J. D. ROBERTS, Skeins and mapping class groups, Math. Proc. Cambridge Philos.<br />
Soc. 115 (1994), 53 – 77. MR 1253282<br />
E. WITTEN, Quantum field theory and the Jones polynomial, Comm. Math. Phys.<br />
121 (1989), 351 – 399. MR 0990772 573<br />
Marché<br />
Université Pierre et Marie Curie, Analyse Algébrique, Institut de Mathematiques de Jussieu,<br />
F-75252 Paris CEDEX 05, France; marche@math.jussieu.fr<br />
Narimannejad<br />
Université Denis Diderot, Topologie et Géométrie Algébriques, Institut de Mathematiques de<br />
Jussieu, F-75251 Paris CEDEX 05, France; nariman@math.jussieu.fr; current: Institut für<br />
Mathematik, Universität Zürich, CH-8057 Zürich, Switzerland;<br />
majid.narimannejad@math.unizh.ch