Sie sind auf Seite 1von 119

a

r
X
i
v
:
0
7
0
7
.
0
9
4
0
v
2


[
m
a
t
h
.
D
S
]


1
2

J
u
l

2
0
0
7
SOBOLEV REGULARITY OF SOLUTIONS
OF THE COHOMOLOGICAL EQUATION
GIOVANNI FORNI

CONTENTS
1. Introduction 1
2. Fractional weighted Sobolev spaces 9
2.1. Weighted Sobolev spaces 9
2.2. Fractional Sobolev norms 14
2.3. Local analysis 20
2.4. Smoothing operators 24
3. The cohomological equation 37
3.1. Distributional solutions 37
3.2. Invariant distributions and basic currents 43
3.3. Basic cohomology 47
4. Cocycles over the Teichmller ow 55
4.1. The Kontsevich-Zorich cocycle. 55
4.2. Distributional cocycles 71
4.3. Lyapunov exponents 79
5. Smooth solutions 92
5.1. The general case 92
5.2. Ergodic integrals 96
5.3. The generic case 105
References 117
1. INTRODUCTION
We prove the sharpest results available on the loss of regularity for so-
lutions of the cohomological equation for translation ows. For any given
translation surface and for the directional ow in almost all directions the
Date: February 1, 2008.
1991 Mathematics Subject Classication. 37A20, 37C10, 37C15.
Key words and phrases. Invariant distributions, cohomological equation, Teich muller
ow, Kontsevich-Zorich cocycle, distributional cocycles.
The author gratefully acknowledges support from National Science Foundation grant
DMS-0244463.
1
2 GIOVANNI FORNI

smallest loss Sobolev regularity available to the Fourier analysis methods
developed in [For97] is essentially 3 + (for any > 0). We remark that
this is the best result available for the ow of rational polygonal billiards
in almost all directions. The motivation for improving the estimate on the
loss of regularity obtained in [For97] was provided by a question of Marmi,
Moussa and Yoccoz [MMY05]. We also remark that their results, on the
related problem of solutions of the cohomological equation for almost all
interval exchange transformations (IETs), do not apply to rational billiards
for the well-known reason that IETs induced by rational billiard ows form
a zero measure set in the space of all IETs.
For almost all translation surfaces in every stratum of the moduli space,
we prove the rened, optimal result that the loss of Sobolev regularity for
the directional ow in almost all directions is 1 + (for any > 0). In
this case, in fact we prove that for any function of Sobolev order s > 1,
the solution and its derivatives up to order k < s 1 are L

functions
on the surface. We also determine precisely the Sobolev orders of the dis-
tributional obstructions to the existence of solutions (rst constructed in
[For97]) in terms of the Kontsevich-Zorich Lyapunov exponents [Zor96],
[Kon97]. As a consequence we are able to determine the exact codimen-
sion of coboundaries for every Sobolev regularity class of the transfer func-
tion. For instance, the codimension of coboundaries with square-integrable
transfer functions (in the space of functions of Sobolev order s > 1) is ex-
actly equal to the genus of the surface. For such coboundaries the transfer
function is actually in L

.
These results implies quite immediately corresponding results for interval
exchange transformations, which improve on the loss of regularity estab-
lished in [MMY05]. We should point out that in that paper the authors are
mostly concerned with Diophantine conditions on interval exchange trans-
formations for which the cohomological equation admits smooth solution,
while we have not investigated this question at all. The reason is that the
full measure sets of systems for which our results (as well as [MMY05])
are determined by several conditions which always include Oseledec regu-
larity (or rather a weaker coherence property) with respect to Kontsevich-
Zorich renormalization cocycle. The Oseledecs theorem is invoked to en-
sure that the set of regular (coherent) IETs has full measure. Any sub-
stantial progress over [MMY05] would have to succeed in characterizing
explicitly a full measure set of regular (coherent) points without relying
on the Oseledecs theorem. To the best of our knowledge this goal is still
beyond reach.
There are several motivations for this work. The study of cohomological
equations is a relevant part of the theory of (smooth) dynamical systems
directly connected to basic questions such as triviality of time-changes for
SOBOLEV REGULARITY OF SOLUTIONS 3
ows, asymptotic of ergodic averages and the smooth conjugacy problem
via linearization and Nash-Moser implicit function theorem. In the hyper-
bolic case (for dynamical systems with exponential divergence of nearby
orbits) such there are extensive, deep results on the cohomological equa-
tion pioneered in the work of Livsic [Liv71], later developed by several au-
thors (see [GK80], [CEG84], [dlLMM86]). The completely different case
of (Diophantine) linear ows on the torus is well-known, since the cohomo-
logical equation for such systems is closely related to the linearized equa-
tion in the classical KAM theory for Hamiltonian ows. This in an example
of elliptic dynamics (no divergence of nearby orbits) which can be studied
to a great extent by the classical theory of Fourier series. It is characterized
by the small divisors appearing in the Fourier coefcients of solutions,
which lead to a loss of regularity. It is not difcult to see that the optimal
loss of Sobolev regularity for the full measure set of Roth ows is 1 + and
that for such ows any zero average function of Sobolev order s > 1 is an
L
2
coboundary. It can be proved by our methods (and by the Gottschalk-
Hedlund theorem [GH55]) that the transfer function is in fact continuous.
We have not been able to locate this result in the literature, however it is
well within reach of the methods of [Her83], Chap. VI, 3. However, only
the measure zero case of rotation numbers of constant type seems to have
been explicitly considered there.
For systems with intermediate behavior, that is, for elliptic systems with
singularities or for parabolic systems (characterized by polynomial diver-
gence of nearby orbits) much less is known. The author discovered in
[For97] that the cohomological equation for generic translation ows (or
equivalently for generic IETs) has nitely smooth solutions for sufciently
smooth data under nitely many distributional conditions. In other terms,
on one hand the problem shares a typical feature of small divisors prob-
lems, namely the nite loss of regularity of solutions with respect to the
data; on the other hand, a new phenomenon appears: the existence of in-
nitely many independent distributional obstructions (of increasing order)
which are not given by invariant measures. In [For97] only a rough esti-
mate for the loss of derivatives is explicitly obtained ( 9). Our goal in this
paper is to improve such estimate as much as possible. In joint papers with
L. Flaminio the authors have investigated the existence of smooth solutions
of the cohomological equation for horocycle ows (on surfaces constant
negative curvature) [FF03], for generic nilows on quotients of the Heisen-
berg group [FF06] and generic nilows on general nilmanifolds [FF07]. In
all cases the fundamental features of nite loss derivatives and of the exis-
tence of innitely many independent distributional obstructions have been
established (although the structure of the space of invariant distributions is
4 GIOVANNI FORNI

signicantly different for IETs, horocycle ows and nilows). For horo-
cycle ows and for Heisenberg nilows it was possible to estimate that the
loss of Sobolev regularity is 1 + (for any > 0) and to establish the con-
jectural relation that the Sobolev order of the distributional obstructions be
related to the Lyapunov exponents of the distribution under the appropri-
ate renormalization dynamics. In this paper we prove analogous results for
generic translation ows. We should point out that for generic nilows on
general nilmanifolds the loss of regularity and the regularity of the distri-
butional obstructions seem to depend on the depth and rank of the nilpotent
group considered, although no lower bound was established in [FF07].
Let q be a holomorphic orientable quadratic differential on a Riemann
surface M of genus g 1. The horizontal and vertical measured foliations
(in the Thurstons sense) associated to a holomorphic quadratic differen-
tial q on M are dened as F
q
= (q
1/2
) = 0 (the horizontal foliation)
and F
q
= (q
1/2
) = 0. Such foliations are well-dened even in the
case that there is no globally dened square root of the quadratic differen-
tial. The horizontal foliation is endowed with the transverse measure given
by [(q
1/2
)[, the vertical foliation is endowed with the transverse measure
given by [(q
1/2
)[. The quadratic differential is called orientable if the hori-
zontal and vertical foliations are both orientable. Orientability is equivalent
to the condition that the quadratic differential is globally the square of a
holomorphic (abelian) differential. The structure induced by an orientable
holomorphic quadratic differential (or by a holomorphic abelian differen-
tial) can also be described as follows. There is a at metric R
q
associated
with any quadratic differential q on M. Such a metric has conical singular-
ities as the nite set
q
= p M[q(p) = 0. If q is orientable there exists
a (positively oriented) parallel orthonormal frame S
q
, T
q
of the tangent
bundle TM[M
q
such that S
q
is tangent to the horizontal foliation F
q
and T
q
is tangent to the vertical foliation F
q
everywhere on M
q
. In
other terms, the at metric R
q
has trivial holonomy. In another equivalent
formulation, any orientable holomorphic quadratic differential determines
a translation structure on M, that is, an equivalence class of atlases with
transition functions given by translation of the euclidean plane (see for in-
stance the excellent survey [MT05], 1.8. For a given orientable quadratic
differential q on a Riemann surface M, we will consider the one-parameter
family of vector elds on M
q
dened as
(1.1) S

:= cos S
q
+ sin T
q
, S
1
.
The vector eld S

is a parallel normalized vector eld in the direction at


angle S
1
with the horizontal. We remark that it is not dened as the
singular set
q
of the at metric. hence the ow it generates is dened
(almost everywhere) on the complement of the union of all separatrices
SOBOLEV REGULARITY OF SOLUTIONS 5
of the orbit foliation (a measure foliation). The singularities of the orbit
foliation are all saddle-like, but the saddles are degenerate if the order of
zero of the quadratic differential at the singularity is strictly greater than 2.
In fact, since the quadratic differential is supposed to be orientable it has
zeroes of even order and the orbit foliations of the vector elds (1.1) has m
stable and m unstable separatrices at any zero of order 2m.
Our goal is to investigate the loss of (Sobolev) regularity of solutions of the
cohomological equation S

u = f for Lebesgue almost all S


1
. The
author proved in [For97] that if the function f is sufciently regular, satis-
es a nite number of independent distributional conditions (which include
conditions on the jets at the singularities) then there exists a nitely smooth
solution (unique up to additive constants). The loss of regularity was esti-
mated in that paper to be no more than 9 derivatives in the Sobolev sense.
If q is any orientable quadratic differential, the regularity of functions on the
translation surface (M, q) is expressed in terms of a family H
s
q
(M)[s R
of weighted Sobolev spaces. Such spaces were introduced in [For97] for
all s Z as follows. Let
q
be the standard (degenerate) volume form
on M of the at metric R
q
. The space H
0
q
(M) is the space L
2
(M,
q
) of
square-integrable functions. For k N, the space H
k
q
(M) is the subspace
of functions f H
0
q
(M) such that the weak derivatives S
i
q
T
j
q
f H
0
q
(M)
[and T
i
q
S
j
q
f H
0
q
(M)] for all i + j k and the space H
k
q
(M) is the
dual Hilbert space H
k
q
(M)

. In 2 of this paper we introduce weighted


Sobolev spaces with arbitrary (real) exponents by methods of interpolation
theory. Although the Sobolev norms we construct do not form an inter-
polation family in the sense of (holomorphic) interpolation theory, they do
satisfy a standard interpolation inequality. The weighted Sobolev spaces
combine standard Sobolev smoothness conditions on M
q
with restric-
tions on the jet of the functions at the singular set
q
M.
As discovered in [For97], for functions f C

0
(M
q
) the space of all
distributional obstructions to the existence of a solution u C

0
(M
q
)
of the cohomological equation S

u = f coincides for almost all S


1
with the innite dimensional space of all S

-invariant distributions:
(1.2) I
q,
(M
q
) := D D

(M
q
) [ S

D = 0 in D

(M
q
) .
For data f H
s
q
(M) of nite Sobolev differentiability, a complete set of
obstructions is given for almost all S
1
by the nite dimensional sub-
space of invariant distributions
(1.3) I
s
q,
(M) := D H
s
q
(M) [ S

D = 0 in H
s
q
(M) .
The goal of this paper is to prove optimal estimates on the Sobolev reg-
ularity of solutions of the cohomological equation S

u = f and on the
6 GIOVANNI FORNI

dimension of the spaces I
s
q,
(M) of invariant distributions for all s > 0. In
5.1, Theorem 5.1, we prove
Theorem A1. Let q be any orientable holomorphic quadratic differential.
Let k N be any integer such that k 3 and let s > k and r < k 3.
For almost all S
1
(with respect to the Lebesgue measure), there exists
a constant C
r,s
() > 0 such that the following holds. If f H
s
q
(M) is such
that D(f) = 0 for all D I
s
q,
(M), the cohomological equation S

u = f
has a solution u H
r
q
(M) satisfying the following estimate:
(1.4) [u[
r
C
r,s
() [f[
s
.
The dimensions of the spaces of invariant distributions can be estimated as
follows (see 3.3, Corollary 3.20 and Theorem 3.21):
Theorem A2. Let q be any orientable holomorphic quadratic differential.
Let k N be any integer such that k 3 and let k < s k +1. For almost
all S
1
(with respect to the Lebesgue measure),
(1.5) 1 + 2(k 2)(g 1) dimI
s
q,
(M) 1 + (2k 1)(g 1) .
The proof of the above results is essentially based on the harmonic analysis
methods developed in [For97]. We remark that no other methods are known
for the case of an arbitrary orientable quadratic differential.
We prove much sharper results for almost all orientable quadratic differ-
entials. The moduli space of orientable holomorphic quadratic differen-
tials q on some Riemann surface M
q
with a given pattern of zeroes, that
is, with zeroes of (even) multiplicities = (k
1
, . . . , k

) at a nite set

q
= p
1
, . . . , p

M
q
is a stratum M

of the moduli space M


g
of
all holomorphic quadratic differential. Let M
(1)

M
(1)
g
be the subsets
of quadratic differential of total area equal to 1. It was proved by H. Ma-
sur [Mas82] and W. Veech [Vee86] that each stratum M
(1)

carries an ab-
solutely continuous probability measure
(1)

, invariant under the action of


the Teichmller geodesic ow, which is ergodic when restricted to each
connected component of M
(1)

(the connected components of strata of ori-


entable quadratic differentials were classied in [KZ03]). In fact, there is
natural action of the group SL(2, R) on the moduli space M
(1)
g
such that the
Teichmller geodesic ow corresponds to the action of the diagonal sub-
group of SL(2, R) and the measure
(1)

is SL(2, R) invariant.
In [Kon97] M. Kontsevich introduced a renormalization cocycle for trans-
lation ows, inspired to the Rauzy-Veech-Zorich cocycle for interval ex-
change transformations. The Kontsevich-Zorich cocycle is a dynamical
system on an orbifold vector bundle over M
(1)

with ber the rst cohomol-


ogy H
1
(M
q
, R) of the Riemann surface carrying the orientable holomorphic
SOBOLEV REGULARITY OF SOLUTIONS 7
quadratic differential q M
(1)

. The action of such a dynamical system is


(by denition of a cocycle) linear on the bers and projects onto the Teich-
mller geodesic ow on the base M
(1)

. Since the cocycle is symplectic, for


any probability measure on a stratumM
(1)

, the Lyapunov spectrum of the


Kontsevich-Zorich cocycle takes the form:
(1.6)

g
0

g+1
=

2g
=

1
.
In addition, it is not difcult to prove that

1
= 1. A probability measure
on a stratum M
(1)

, invariant under the Teichmller geodesic ow, will be


called (a) SO(2, R)-absolutely continuous if it induces absolutely contin-
uous measures on every orbit of the circle group SO(2, R) SL(2, R);
(b) KZ-hyperbolic if all the Lyapunov exponents in (1.6) are non-zero. It
is immediate that all SL(2, R)-invariant measures are SO(2, R)-absolutely
continuous. It was rst proved in [For02] that the measure
(1)

is KZ-
hyperbolic. A different proof that also reaches the stronger conclusion that
the exponents (1.6) are all distinct has been given more recently by A. Avila
and M. Viana [AV05] who have thus completed the proof of the Zorich-
Kontsevich conjectures [Zor96], [Kon97] on the Lyapunov spectrum of the
Kontsevich-Zorich cocycle (and its discrete counterparts).
Our sharpest results are proved for almost all quadratic differentials with re-
spect to any SO(2, R)-absolutely continuous, KZ-hyperbolic, Teichmller
invariant, probability measure on any stratum M
(1)

of orientable quadratic
differentials. The smoothness informations on the solutions is stronger
than just Sobolev L
2
regularity and it is naturally encoded by the following
spaces. For any k N, let B
k
q
(M) be the space of all functions u H
k
q
(M)
such that S
i
q
T
j
q
u = T
i
q
S
j
q
u L

(M) for all pairs of integers (i, j) such that


0 i + j k. The space B
k
q
(M) is endowed with the norm dened as
follows: for any u B
k
q
(M),
(1.7) [u[
k,
:=
_

i+jk
[S
i
q
T
j
q
u[
2

_
1/2
=
_

i+jk
[T
i
q
S
j
q
u[
2

_
1/2
.
For s [k, k +1), let B
s
q
(M) := B
k
q
(M) H
s
q
(M) endowed with the norm
dened as follows: for any u B
s
q
(M),
(1.8) [u[
s,
:=
_
[u[
2
k,
+ [u[
2
s
_
1/2
.
In 5.3, Theorem 5.19, we prove the following:
TheoremB1. Let be any SO(2, R)-absolutely continuous, KZ-hyperbolic
probability measure on any stratum M
(1)

of orientable quadratic differen-


tials. Let s > 1 and let r < s 1. For -almost all q M
(1)

and for
8 GIOVANNI FORNI

almost all S
1
(with respect to the Lebesgue measure), there exists a
constant C
r,s
() > 0 such that the following holds. If f H
s
q
(M) is such
that D(f) = 0 for all D I
s
q,
(M), the cohomological equation S

u = f
has a solution u B
r
q
(M) satisfying the following estimate:
(1.9) [u[
r,
C
r,s
() [f[
s
.
The regularity of invariant distributions can be precisely determined as fol-
lows (see 4.3, Corollary 4.33, and 5.3, Theorem 5.19). For any q M
(1)

and any distribution D D

(M
q
), the weighted Sobolev order is the
number
(1.10) O
H
q
(D) = infs R
+
[D H
s
q
(M) .
Let I
q,
(M) := I
s
q,
(M)[s 0 denote the space of all S

-invariant dis-
tribution of nite Sobolev order and let

I
q,
(M) I
q,
(M) be the subspace
of invariant distributions vanishing on constant functions. It follows imme-
diately by the denitions that I
q,
(M) = C

I
q,
(M).
TheoremB2. Let be any SO(2, R)-absolutely continuous, KZ-hyperbolic
probability measure on any stratum M
(1)

of orientable quadratic differen-


tials. For -almost all q M
(1)

and for almost all S


1
(with respect to
the Lebesgue measure), the space

I
q,
(M) has a basis D
i,j
() such that
(1.11) O
H
q
(D
i,j
()) =

i
(j +1) , i 2, . . . , 2g 1, j N0 .
In addition, the basis D
i,j
() can be generated fromthe nite dimensional
subsystem D
2
, . . . , D
2g1
by the following differential relations:
(1.12) D
i,j
() = T
j

D
i,0
() , i 2, . . . , 2g 1, j N 0 .
The above Theorems B1 and B2 are proved by methods based on renor-
malization, which were inspired by the work of Marmi-Moussa-Yoccoz
[MMY05]. However, our approach differs from theirs since we explicitly
assume that the Lyapunov exponents of the Kontsevich-Zorich cocycle are
all non-zero. The main idea of the argument, as in [MMY05], is to prove
uniform estimates for ergodic integrals of weakly differentiable functions,
then apply a version of Gottshalk-Hedlund theorem. The asymptotics of er-
godic averages of functions in H
1
(M) was studied by the author in [For02],
where the Kontsevich-Zorich conjectures on the deviation of ergodic aver-
ages for smooth functions (formulated in [Kon97]) were proved. The ap-
proach of [For02] is based on the analysis of distributional cocycles over the
Teichmller ow which extend the Kontsevich-Zorich cocycles. The esti-
mates proved in [For02] are (barely) not strong enough to yield the required
uniform boundedness of ergodic averages under the appropriate distribu-
tional conditions. In 4.2 of this paper we have recalled the denition of
SOBOLEV REGULARITY OF SOLUTIONS 9
distributional cocycles, and we have strengthened the estimates proved in
[For02] under the slightly stronger (and correct) assumption that the func-
tions considered belong to H
s
q
(M) for some s > 1.
Another important techical issue that separates Theorems B1 and B2 from
the less precise Theorems A1 and A2 is related to interpolation theory in
the presence of distributional obstructions. In the general case, we have not
been able to overcome the related difculties, hence the lack of precision
of Theorem A1 and A2 for intermediate Sobolev regularity. In the generic
case of Theorems B1 and B2 we have been able to prove a remarkable linear
independence property of invariant distributions which makes interpolation
possible in the construction of solutions of the cohomological equation.
We introduce the following denition (see Denition 5.11). A nite system
D
1
, . . . , D
J
H

q
(M) of nite order distributions is called -regular
(with respect to the family H
s
q
(M) of weighted Sobolev spaces) if for
any (0, 1] there exists a dual system u
1
(), . . . , u
J
() H

q
(M)
(that is, the identities D
i
(u
j
()) =
ij
hold for all i, j 1, . . . , J and all
(0, 1]) such that the following estimates hold. For all 0 r and all
> 0, there exists a constant C

r
() > 0 such that, for all i, j 1, . . . , J,
(1.13) [u
j
()[
r
C

r
()
O
H
(D
j
)r
.
A nite system D
1
, . . . , D
J
H
s
q
(M) of nite order distributions will
be called regular if it is -regular for any s. A nite dimensional
subspace I H
s
q
(M) of nite order distributions will be called -regular
[regular] if it admits a -regular [regular] basis.
We have proved that the spaces of distributional obstructions for the coho-
mological equation are regular in the above sense (see Theorem 5.18).
Theorem C. Let be any SO(2, R)-absolutely continuous, KZ-hyperbolic
probability measure on any stratum M
(1)

of orientable quadratic differen-


tials. For -almost all q M
(1)

, for almost all S


1
and for all s > 0,
the space I
s
q,
(M) H
s
q
(M) of S

-invariant distributions is regular.


2. FRACTIONAL WEIGHTED SOBOLEV SPACES
In [For97] we have introduced a natural scale of weighted Sobolev spaces
with integer exponent associated with any orientable holomorphic quadratic
differential q on a Riemann surface M (of genus g 2). In this section we
extend the denition of weighted Sobolev spaces to arbitrary (real) expo-
nents by methods of interpolation theory.
2.1. Weighted Sobolev spaces. Let
q
:= p
1
, . . . , p

M be the
set of zeros of the holomorphic quadratic differential q, of even orders
(k
1
, . . . , k

) respectively with k
1
+ + k

= 4g 4. Let R
q
:= [q[
1/2
be
10 GIOVANNI FORNI

the at metric with cone singularities at
q
induced by the quadratic differ-
ential q on M. With respect to a holomorphic local coordinate z = x + iy,
the quadratic differential q has the form q = (z)dz
2
, where is a locally
dened holomorphic function, and, consequently,
(2.1) R
q
= [(z)[
1/2
(dx
2
+ dy
2
)
1/2
,
q
= [(z)[ dx dy .
The metric R
q
is at, it is degenerate at the nite set
q
of zeroes of q and,
if q is orientable, it has trivial holonomy, hence q induces a structure of
translation surface on M.
The weighted L
2
space is the standard space L
2
q
(M) := L
2
(M,
q
) with
respect to the area element
q
of the metric R
q
. Hence the weighted L
2
norm [ [
0
are induced by the hermitian product , )
q
dened as follows:
for all functions u,v L
2
q
(M),
(2.2) u, v)
q
:=
_
M
u v
q
.
Let F
q
be the horizontal foliation, F
q
be the vertical foliation for the holo-
morphic quadratic differential q on M. The foliations F
q
and F
q
are mea-
sured foliations (in the Thurstons sense): F
q
is the foliation given (locally)
by the equation (q
1/2
) = 0 endowed with the invariant transverse measure
[(q
1/2
)[, F
q
is the foliation given (locally) by the equation (q
1/2
) = 0
endowed with the invariant transverse measure [(q
1/2
)[. If the quadratic
differential q is orientable, since the metric R
q
is at with trivial holonomy,
there exist commuting vector elds S
q
and T
q
on M
q
such that
(1) The frame S
q
, T
q
is a parallel orthonormal frame with respect to
the metric R
q
for the restriction of the tangent bundle TM to the
complement M
q
of the set of cone points;
(2) the vector eld S
q
is tangent to the horizontal foliation F
q
, the vector
eld T
q
is tangent to the vertical foliation F
q
on M
q
[For97].
In the following we will often drop the dependence of the vector elds S
q
,
T
q
on the quadratic differential in order to simplify the notations. We have:
(1) L
S

q
= L
T

q
= 0 on M
q
, that is, the area form
q
is invariant
with respect to the ows generated by S and T;
(2)
S

q
= (q
1/2
) and
T

q
= (q
1/2
), hence the 1-forms
S
:=
S

q
,

T
:=
T

q
are smooth and closed on M and
q
=
T

S
.
It follows from the area-preserving property (1) that the vector eld S, T
are anti-symmetric as densely dened operators on L
2
q
(M), that is, for all
functions u, v C

0
(M
q
), (see [For97], (2.5)),
(2.3) Su, v)
q
= u, Sv)
q
, respectively Tu, v)
q
= u, Tv)
q
.
SOBOLEV REGULARITY OF SOLUTIONS 11
In fact, by Nelsons criterion [Nel59], Lemma 3.10, the anti-symmetric op-
erators S, T are essentially skew-adjoint on the Hilbert space L
2
q
(M).
The weighted Sobolev norms [ [
k
, with integer exponent k > 0, are the
euclidean norms, introduced in [For97], induced by the hermitian product
dened as follows: for all functions u, v L
2
q
(M),
(2.4) u, v)
k
:=
1
2

i+jk
S
i
T
j
u, S
i
T
j
v)
q
+T
i
S
j
u, T
i
S
j
v)
q
.
The weighted Sobolev norms with integer exponent k < 0 are dened to
be the dual norms.
The weighted Sobolev space H
k
q
(M), with integer exponent k Z, is the
Hilbert space obtained as the completion with respect to the norm [ [
k
of
the maximal common invariant domain
(2.5) H

q
(M) :=

i,jN
D(

S
i

T
j
) D(

T
i

S
j
) .
of the closures

S,

T of the essentially skew-adjoint operators S, T on
L
2
q
(M). The weighted Sobolev space H
k
q
(M) is isomorphic to the dual
space of the Hilbert space H
k
q
(M), for all k Z.
Since the vector elds S, T commute (innitesimally) on M
q
, the fol-
lowing weak commutation identity holds on M.
Lemma 2.1. ([For97], Lemma 3.1) For all functions u,v H
1
q
(M),
(2.6) Su, Tv)
q
= Tu, Sv)
q
.
By the anti-symmetry property (2.3) and the commutativity property (2.6),
the frame S, T yields an essentially skew-adjoint action of the Lie algebra
R
2
on the Hilbert space L
2
q
(M) with common domain H
1
q
(M). If
q
,= ,
the (at) Riemannian manifold (M
q
, R
q
) is not complete, hence its
Laplacian
q
is not essentially self-adjoint on C

0
(M
q
). By a theoremof
Nelson [Nel59], 9, this is equivalent to the non-integrability of the action
of R
2
as a Lie algebra (to an action of R
2
as a Lie group).
Following [For97], the Fourier analysis on the at surface M
q
will be based
on a canonical self-adjoint extension
F
q
of the Laplacian
q
, called the
Friedrichs extension, which is uniquely determined by the Dirichlet hermit-
ian form Q : H
1
q
(M) H
1
q
(M) C. We recall that, for all u, v H
1
q
(M),
(2.7) Q(u, v) := Su, Sv)
q
+ Tu, Tv)
q
.
Theorem 2.2. ([For97], Th. 2.3) The hermitian form Q on L
2
q
(M) has the
following spectral properties:
12 GIOVANNI FORNI

(1) Q is positive semi-denite and the set EV(Q) of its eigenvalues is a
discrete subset of [0, +);
(2) Each eigenvalue has nite multiplicity, in particular 0 EV(Q) is
simple and the kernel of Q consists only of constant functions;
(3) The space L
2
q
(M) splits as the orthogonal sum of the eigenspaces.
In addition, all eigenfunctions are C

(real analytic) on M.
The Weyl asymptotics holds for the eigenvalue spectrum of the Dirichlet
form . For any > 0, let N
q
() := card EV(Q) / , where each
eigenvalue EV(Q) is counted according to its multiplicity.
Theorem 2.3. ([For97], Th. 2.5) There exists a constant C > 0 such that
(2.8) lim
+
N
q
()

= vol(M, R
q
) .
Let

q
:= S
q
T
q
be the Cauchy-Riemann operators induced by the holo-
morphic orientable quadratic differential q on M, introduced in [For97],
3. Let M

q
L
2
q
(M) be the subspaces of meromorphic, respectively anti-
meromorphic functions (with poles at
q
). By the Riemann-Roch theorem,
the subspaces M

q
have the same complex dimension equal to the genus
g 1 of the Riemann surface M. In addition, M
+
q
M

q
= C, hence
(2.9) H
q
:=
_
M
+
q
_

_
M

q
_

= u L
2
q
(M) [
_
M
u
q
= 0 .
Let H
1
q
:= H
q
H
1
q
(M). By Theorem 2.2, the restriction of the hermitian
form to H
1
q
is positive denite, hence it induces a norm. By the Poincar
inequality (see [For97], Lemma 2.2 or [For02], Lemma 6.9), the Hilbert
space (H
1
q
, Q) is isomorphic to the Hilbert space (H
1
q
, , )
1
).
Proposition 2.4. ([For97], Prop. 3.2) The Cauchy-Riemann operators

q
are closable operators on the common domain C

0
(M
q
) L
2
q
(M) and
their closures (denote by the same symbols) have the following properties:
(1) the domains D(

q
) = H
1
q
(M) and the kernels N(

q
) = C;
(2) the ranges R

q
:= Ran(

q
) =
_
M

q
_

are closed in L
2
q
(M);
(3) the operators

q
: (H
1
q
, Q) (R

, , )
q
) are isometric.
Let E = e
n
[ n N H
1
q
(M) C

(M) be an orthonormal basis of the


Hilbert space L
2
q
(M) of eigenfunctions of the Dirichlet form (2.7) and let
: N R
+
0 be the corresponding sequence of eigenvalues:
(2.10)
n
:= Q(e
n
, e
n
) , for each n N.
SOBOLEV REGULARITY OF SOLUTIONS 13
The Friedrichs weighted Sobolev norm | |
s
of order s R
+
is the norm
induced by the hermitian product dened as follows: for all u, v L
2
q
(M),
(2.11) (u, v)
s
:=

nN
(1 +
n
)
s
u, e
n
)
q
e
n
, v)
q
.
The inner products (2.4) and (2.11) induce equivalent Sobolev norms on
the weighted Sobolev space H
k
q
(M), for all k Z
+
. In fact, the following
result, a sharp version of Lemma 4.2 of [For97], holds:
Lemma 2.5. For each k Z
+
there exists a constant C
k
> 1 such that,
for any orientable holomorphic quadratic differential q on M and for all
functions u H
k
q
(M),
(2.12) C
1
k
[u[
k
|u|
k
C
k
[u[
k
.
Proof. By Proposition 2.4, (3), and Lemma 2.1, for all u H
k+1
q
(M),
k N, the following identity holds (see (4.4) in [For97]):
(2.13) [u[
2
k+1
= [u[
2
0
+

i+jk
Q(S
i
T
j
u, S
i
T
j
u) =
= [u[
2
0
+

i+jk
S
i
T
j

q
u, S
i
T
j

q
u)
q
= [u[
2
0
+ [

q
u[
2
k
.
Hence in particular u H
k+1
q
(M), k N, implies

q
u H
k
q
(M). If
k 1, a second application of the identity (2.13) yields
(2.14) [u[
2
k+1
= [u[
2
1
+ [

q
u[
2
k1
.
The statement then follows by induction on k N. For k = 0 it is immedi-
ate and for k = 1, by the identity (2.13),
(2.15) [u[
2
1
= [u[
2
0
+ Q(u, u) =

nN
(1 +
n
) [u, e
n
)
q
[
2
.
For k > 1, by induction hypothesis we can assume that the norms [ [
k1
and | |
k1
are equivalent, that is, there exists a constant C
k1
> 1 such
that, for all u H
k1
q
(M),
(2.16) C
1
k1
[u[
k1
|u|
k1
C
k1
[u[
k1
.
Since by (2.3) for all u, v H
1
q
(M),

q
u, v)
q
= u,

q
v)
q
, the adjoint
operator (

q
)

q
on H
1
q
(M). By Proposition 2.4, (3), we have
(2.17) |

q
u|
2
k1
=

nN
(1 +
n
)
k1
[

q
u,

q
e
n
)
q
[
2
=
=

nN
(1 +
n
)
k1
[Q(u, e
n
)[
2
=

nN
(1 +
n
)
k1

2
n
[u, e
n
)
q
[
2
.
14 GIOVANNI FORNI

There exists a constant C
k+1
> 1 such that, for all 0,
(2.18) C
2
k+1
(1 +)
k+1
1 + + C
2
k1

2
(1 +)
k1
1 + + C
2
k1

2
(1 +)
k1
C
2
k+1
(1 +)
k+1
.
By (2.14), (2.15), (2.16), (2.17) and (2.18), the estimate
(2.19) C
1
k+1
[u[
k+1
|u|
k+1
C
k+1
[u[
k+1
follows, thereby completing the induction step.
2.2. Fractional Sobolev norms. Let q be any orientable quadratic differ-
ential on M. For all s 0, let
(2.20)

H
s
q
(M) := u L
2
q
(M) /

nN
(1 +
n
)
s
[u, e
n
)
q
[
2
< +,
endowed with the hermitian product given by (2.11) and, for any s > 0, let

H
s
q
(M) be the dual space of the Hilbert space

H
s
q
(M). The spaces

H
s
q
(M)
will be called the Friedrichs (fractional) weighted Sobolev spaces.
Let H
1
H
2
be Hilbert spaces such that H
1
embeds continuously into H
2
with dense image. For all [0, 1], let [H
1
, H
2
]

be the (holomorphic)
interpolation space of H
1
H
2
in the sense of Lions-Magenes [LM68],
Chap. 1, endowed with the canonical interpolation norm. By the results
of [LM68], Chap. 1, 2, 5, 6 and 14, we have the following:
Lemma 2.6. The Friedrichs weighted Sobolev spaces form an interpolation
family

H
s
q
(M)
sR
of Hilbert spaces: for all r, s R with r < s,
(2.21)

H
(1)r+s
q
(M) [

H
r
q
(M),

H
s
q
(M)]

.
The family H
s
q
(M)
sR
of fractional weighted Sobolev spaces will be de-
ned as follows. Let [s] N denote the integer part and s [0, 1) the
fractional part of any real number s 0.
Denition 2.7. (1) The fractional weighted Sobolev norm [ [
s
of or-
der s 0 is the euclidean norm induced by the hermitian product
dened as follows: for all functions u, v H

q
(M),
(2.22) u, v)
s
:=
1
2

i+j[s]
(S
i
T
j
u, S
i
T
j
v)
{s}
+ (T
i
S
j
u, T
i
S
j
v)
{s}
.
(2) The fractional weighted Sobolev norm [ [
s
of order s < 0 is
dened as the dual norm of the weighted Sobolev norm [ [
s
.
(3) The fractional weighted Sobolev space H
s
q
(M) of order s Ris de-
ned as the completion with respect to the norm[ [
s
of the maximal
common invariant domain H

q
(M).
SOBOLEV REGULARITY OF SOLUTIONS 15
It can be proved that the weighted Sobolev space H
s
q
(M) is isomorphic to
the dual space of the Hilbert space H
s
q
(M), for all s R.
Denition 2.8. For any distribution D on M
q
, the weighted Sobolev
order O
H
q
(D) and the Friedrichs weighted Sobolev order

O
H
q
(D) are the
real numbers dened as follows:
(2.23)
O
H
q
(D) :=infs R[ D H
s
q
(M) ;

O
H
q
(D) :=infs R[ D

H
s
q
(M) .
The denition of the fractional weighted Sobolev norms is motivated by the
following basic result.
Lemma 2.9. For all s 0, the restrictions of the Cauchy-Riemann oper-
ators

q
: H
1
q
(M) L
2
q
(M) to the subspaces H
s+1
q
(M) H
1
q
(M) yield
bounded operators

s
: H
s+1
q
(M) H
s
q
(M)
(which do not extend to operators

H
s+1
q
(M)

H
s
q
(M) unless M is the
torus). On the other hand, the Laplace operator
(2.24)
q
=
+
q

q
=

q

+
q
: H
2
q
(M) L
2
q
(M)
yields a bounded operator

s
:

H
s+2
q
(M)

H
s
q
(M), dened as the re-
striction of the Friedrichs extension
F
q
:

H
2
q
(M) L
2
q
(M).
Proof. The restrictions

s
: H
s+1
q
(M) H
s
q
(M) of the Cauchy-Riemann
operators are well-dened and bounded for all s > 0 by denition of the
Sobolev spaces H
s
q
(M).
The operators

s
: H
s+1
q
(M) H
s
q
(M) do not extend to bounded oper-
ators

H
s+1
q
(M)

H
s
q
(M) unless M
q
is a at torus. In fact, every nite
combination f of eigenfunctions of the Dirichlet form belongs to

H
s
q
(M),
for all s R, but

q
f , H
1
q
(M) in all cases because of the presence of
obstructions in the Taylor expansion of eigenfunctions at the singular set

q
,= . In fact, if
+
q
e
n
H
1
q
(M) (or

q
e
n
H
1
q
(M)) for all n N,
then
+
q
e
n
H
1
q
(M) and

q
e
n
H
1
q
(M), for all n N, since the eigen-
functions e
n
can be chosen real. It follows that e
n
H
2
q
(M), for all n N,
hence H
2
q
(M) =

H
2
q
(M) is the domain of the Friedrichs extension
F
q
of
the Laplacian
q
of the metric R
q
. Thus the Laplacian
q
is self-adjoint on
the domain H
2
q
(M), hence the metric R
q
has no singularities and M is the
torus. In fact, the space H
2
q
(M) is the domain of the closure of the Lapla-
cian
q
on the common invariant domain H

q
(M). If H
2
q
(M) =

H
2
q
(M),
then
q
is essentially self-adjoint on H

q
(M) and by [Nel59], Th. 5 or Cor.
9.1, the action of the commutative Lie algebra spanned by S, T integrates
to a Lie group action. Hence, the singularity set
q
= and M is the torus.
16 GIOVANNI FORNI

Finally, the Friedrichs extension
F
q
, dened on

H
2
q
(M), has a bounded
restriction

s
:

H
s+2
q
(M)

H
s
q
(M), for all s 0. In fact, we have
(2.25)
F
q
u =

nN

n
u, e
n
)
q
e
n
, for all u

H
2
q
(M) ,
hence
F
q
u

H
s
q
(M) if u

H
s+2
q
(M), by denition of the Friedrichs
weighted Sobolev spaces

H
s
q
(M) (in terms of eigenfunction expansions for
the Dirichlet form).
Lemma 2.10. The fractional weighted Sobolev norms satisfy the following
interpolation inequalities. For any 0 r < s there exists a constant C
r,s
>
0 such that, for any [0, 1] and any function u H
s
q
(M),
(2.26) [u[
(1)r+s
C
r,s
[u[
1
r
[u[

s
.
Proof. The argument will be carried out in three steps: (1) the open interval
(r, s) does not contain integers; (2) the open interval (r, s) contains a single
integer; (3) the general case.
In case (1) there exists k N such that k r < s k + 1. The in-
terpolation inequality follows from the denition (2.22) of the euclidean
product which induces the fractional Sobolev norms, from the interpola-
tion inequality for Friedrichs weigthed Sobolev norms (which are by def-
inition interpolation norms) and from the Hlder inequality. In fact, since
0 r k s k 1 and (0, 1), the fractional part
(2.27) (1 )r + s = (1 )(r k) + (s k) ,
hence, by the interpolation inequality for Friedrichs norms (see for instance
[LM68], Chap. 1, 2.5), for all i + j k the following estimates hold:
(2.28)
[S
i
T
j
u[
{(1)r+s}
[S
i
T
j
u[
1
rk
[S
i
T
j
u[

sk
;
[T
i
S
j
u[
{(1)r+s}
[T
i
S
j
u[
1
rk
[T
i
S
j
u[

sk
.
By the denition (2.22) of the Sobolev norms, the interpolation inequality
(2.26) follows from (2.28) by Hlder inequality.
In case (2), there exists k N such that k 1 r < k < s k + 1. We
claim that, for any u H
s
q
(M),
(2.29) [u[
k
C
r,s
[u[
sk
sr
r
[u[
kr
sr
s
.
Let us prove that step (1) and the above claim (2.29) imply step (2). Let
= (1 )r + s. We will consider only the case when (r, k) since
the case when (k, s) is similar. By step (1) we have the inequality
(2.30) [u[

C
r
[u[
k
kr
r
[u[
r
kr
k
.
SOBOLEV REGULARITY OF SOLUTIONS 17
By the claim (2.29) it then follows that
(2.31) [u[

C
(1)
r,s
[u[
k
kr
+
r
kr
sk
sr
r
[u[
r
kr
kr
sr
s
.
It is immediate to verify that
k
k r
+
r
k r
s k
s r
=
s
s r
.
Let us turn to the proof of the claim(2.29). Since 1 rk < 0 < sk
1 and the weighted Sobolev norm [ [
s
coincides with the Friedrichs norm
| |
s
for any s [1, 1], by the interpolation inequality for the Friedrichs
Sobolev norms, the following estimates hold: for all i, j N with i +j k
and for any function u H
s
q
(M),
(2.32)
[S
i
T
j
u[
0
C
(2)
r,s
[S
i
T
j
u[
sk
sr
rk
[S
i
T
j
u[
kr
sr
sk
;
[T
i
S
j
u[
0
C
(2)
r,s
[T
i
S
j
u[
sk
sr
rk
[T
i
S
j
u[
kr
sr
sk
.
Since the operators S, T :

H
1
q
(M) L
2
q
(M) are well-dened, bounded
and have bounded linear extensions L
2
q
(M)

H
1
q
(M), by the fundamen-
tal theorem of interpolation (see for instance [LM68], Chap. 1, 5.1), the
operators S, T :

H
s
q
(M)

H
s1
q
(M) are well-dened and bounded for
any s [0, 1]. It follows that there exists a constant C
r
> 0 such that
(2.33)

i+jk
[S
i
T
j
u[
2
rk
(C

r
)
2

i+jk1
[S
i
T
j
u[
2
r(k1)
;

i+jk
[T
i
S
j
u[
2
rk
(C

r
)
2

i+jk1
[T
i
S
j
u[
2
r(k1)
The claim (2.29) then follows by Hlder inequality from (2.32) and (2.33).
In general, let k
1
< k
2
be positive integers such that
k
1
1 r < k
1
< k
2
< s k
2
+ 1 .
By Lemma 2.5, since the Friedrichs norms are interpolation norms, we have
that there exists a constant C
k
1
,k
2
> 0 such that, for all k N [k
1
, k
2
],
(2.34) [u[
k
C
k
1
,k
2
[u[
k
2
k
k
2
k
1
k
1
[u[
kk
1
k
2
k
1
k
2
.
By step (2) there exists a constant C
(3)
r,s
> 0 such that
(2.35)
[u[
k
1
C
(3)
r,s
[u[
1
k
1
+1r
r
[u[
k
1
r
k
1
+1r
k
1
+1
;
[u[
k
2
C
(3)
r,s
[u[
sk
2
sk
2
+1
k
2
1
[u[
1
sk
2
+1
s
.
18 GIOVANNI FORNI

The estimates in (2.35) imply, by bootstrap-type estimates based on (2.34)
for k = k
1
+ 1 and k = k
2
1, that there exists a constant C
(4)
r,s
> 0 such
that
(2.36)
[u[
k
1
C
(4)
r,s
[u[
k
2
k
1
k
2
r
r
[u[
k
1
r
k
2
r
k
2
;
[u[
k
2
C
(4)
r,s
[u[
sk
2
sk
1
k
1
[u[
k
2
k
1
sk
1
s
.
By (2.36) and again by bootstrap, there exists a constant C
(5)
r,s
> 0 such that
(2.37)
[u[
k
1
C
(5)
r,s
[u[
sk
1
sr
r
[u[
k
1
r
sr
s
;
[u[
k
2
C
(5)
r,s
[u[
sk
2
sr
r
[u[
k
2
r
sr
s
,
Let (r, s). We have proved the interpolation inequality for the subcases
= k
1
and = k
2
. Let us prove that the general case can be reduced to
these subcases. If (r, k
1
), by step (1) there exists C

r
> 0 such that
(2.38) [u[

r
[u[
k
1

k
1
r
r
[u[
r
k
1
r
k
1
.
The interpolation inequality in this case follows immediately from (2.37)
and (2.38). If (k
2
, s), the argument is similar. If (k
1
, k
2
), then by
step (1), there exists C
[]
> 0 such that
(2.39) [u[

C
[]
[u[
1{}
[]
[u[
{}
[]+1
.
The interpolation inequality then follows from (2.34), (2.37) and (2.39).

Let H
s
(M), s R, denote a family of standard Sobolev spaces on the
compact manifold M (dened with respect to a Riemannian metric). The
comparison lemma below claries to some extent the relations between the
different scales of fractional Sobolev spaces.
Lemma 2.11. The following continuous embedding and isomorphisms of
Banach spaces hold:
(1) H
s
(M) H
s
q
(M)

H
s
q
(M) , for 0 s < 1;
(2) H
s
(M) H
s
q
(M)

H
s
q
(M) , for s = 1;
(3) H
s
q
(M)

H
s
q
(M) H
s
(M) , for s > 1.
For s [0, 1], the space H
s
(M) is dense in H
s
q
(M) and, for s > 1, the
closure of H
s
q
(M) in

H
s
q
(M) or H
s
(M) has nite codimension.
Proof. By denition H
0
(M) = L
2
(M) and H
0
q
(M) =

H
0
q
(M) = L
2
q
(M).
Since the area form induced by any quadratic differential is smooth on M,
which is a compact surface, it follows that L
2
(M) L
2
q
(M). The em-
bedding H
1
q
(M)

H
1
q
(M) follows by Lemma 2.5 and the embedding
SOBOLEV REGULARITY OF SOLUTIONS 19

H
1
q
(M) H
1
q
(M) holds since the eigenfunctions of the Dirichlet form
are in H
1
q
(M). The isomorphism H
1
(M) H
1
q
(M) is proved in [For02],
6.2. Hence (2) is proved and (1) follows by interpolation.
Let s > 1. If [s] = 2k is even, there exists a constant A
k
> 0 such that, for
all functions u H

q
(M), we have
(2.40) |u|
2
s
= |(I
F
q
)
k
u|
2
{s}
A
2
k

i+j2k
|S
i
T
j
u|
2
{s}
= C
2
k
[u[
s
.
If [s] = 2k +1 is odd, we argue as follows. The Cauchy-Riemann operators

q
: H
1
q
(M) L
2
q
(M) are bounded and extend by duality to bounded
operators

0
: L
2
q
(M) H
1
q
(M). Hence, by the fundamental theorem
of interpolation (see [LM68], Chap. 1, 5.1), for all [0, 1] the Cauchy-
Riemann operators have bounded restrictions
(2.41)

:

H

q
(M)

H
1
q
(M) .
It follows that there exists a constant B
k
> 0 such that, for all functions
u H

q
(M), we have
(2.42)
|u|
2
s
= |(I
F
q
)
k+1
u|
2
{s}1
B
2
k

i+j2k+1
|S
i
T
j
u|
2
{s}
= C
2
k
[u[
s
.
Thus the embeddings H
s
q
(M)

H
s
q
(M) hold for all s > 1.
It was proved in [For02], 6.2, that H
k
q
(M) H
k
(M), for all k Z
+
. We
prove below the stronger statement that

H
s
q
(M) H
s
(M), for all s R
+
.
Let R be a smooth Riemannian metric on M conformally equivalent to the
degenerate metric R
q
and let H
s
R
(M), s 0, denote the Sobolev spaces
of the Riemannian manifold (M, R) which are dened as the domains of
the powers of the essentially self-adjoint Laplacian
R
of the metric. Since
M is compact, the Sobolev spaces H
s
R
(M) H
s
(M) are independent, as
topological vector spaces, of the choice of the Riemannian metric R, for all
s R. We claim that

H
2k
q
(M) H
2k
R
(M), for all k Z
+
. In fact, there
exists a smooth non-negative real-valued function W on M (vanishing only
at
q
) such that W
F
q

R
. Let W be the unique function such that
the area forms of the metrics are related by the identity
q
= W
R
. If
u

H
2
q
(M), then
F
q
u L
2
q
(M), so that
(2.43)
R
u = W
F
q
u L
2
(M,
R
) .
Let us assume that

H
2k2
q
(M) H
2k2
R
(M) and let u

H
2k
q
(M). We have
(2.44)
k
R
u =
k1
R
W
F
q
u = [
k1
R
, W]
F
q
u + W
k1
R

F
q
u .
Since the commutator [
k1
R
, W] and
k1
R
are differential operators of or-
der 2k 2 on M and
F
q
u H
2k2
R
(M) by the induction hypothesis, the
20 GIOVANNI FORNI

function
k
R
u L
2
(M,
R
). The claim is therefore proved. It follows by
interpolation that

H
s
q
(M) H
s
R
(M), for all s 1. Thus (3) is proved.
For s [0, 1], the space C

0
(M
q
) H
s
q
(M) is dense in H
s
q
(M).
For s > 1, the subset C

(M)

H
s
q
(M) is dense in

H
s
q
(M), since the
eigenfunctions of the Dirichlet form (hence all nite linear combinations)
belong to C

(M) and the space C

(M) is dense in H
s
(M). Finally, the
subspace C

(M) H
k
q
(M) C

(M) can be described, for any k N,


as the kernel of a nite number of distributions of nite order supported on
the nite set
q
(see [For02], (7.9)), hence for any k > s the closure of
H
k
q
(M) H
s
q
(M) in

H
s
q
(M) or in H
s
(M) has nite codimension.
2.3. Local analysis. For each p M and all k Z
+
, let H
k
q
(p),

H
k
q
(p),
and H
k
(p) the spaces of germs of fuctions at p which belong to H
k
q
(M),

H
k
q
(M) and H
k
(M), endowed with the respective direct limit topologies.
More precisely, a germ of function f at p belongs to the space H
k
q
(p),

H
k
q
(p)
or H
k
(p) iff it can be realized by a function F on M which belongs to the
space H
k
q
(M),

H
k
q
(M) or H
k
(M) respectively and the open sets in H
k
q
(p),

H
k
q
(p) or H
k
(p) are dened as the images of open sets in H
k
q
(M),

H
k
q
(M)
or H
k
(M) under the natural maps H
k
q
(M) H
k
q
(p),

H
k
q
(M)

H
k
q
(p) or
H
k
(M) H
k
(p).
By Lemma 2.11 we have the inclusions
(2.45)
H
0
(p) H
0
q
(p)

H
0
q
(p) ;
H
1
q
(p) =

H
1
q
(p) = H
1
(p) ;
H
k
q
(p)

H
k
q
(p) H
k
(p) .
If p ,
q
, since there is an open neighbourhood D
p
of p in M isomorphic
to a at disk and the operator
F
q
is elliptic of order 2 on D
p
(isomorphic
to the at Laplacian), all the inclusions in (2.45) are identities. We will
describe precisely the inclusions H
k
q
(p)

H
k
q
(p) H
k
(p) for k > 1.
Let p
q
be a zero of (even) order 2m of the (orientable) quadratic dif-
ferential q on M. There exists a unique canonical holomorphic coordinate
z : D
p
C, dened on a neighbourhood D
p
of p M, such that z(p) = 0
and q(z) = z
2m
dz
2
. With respect to the canonical coordinate the Cauchy-
Riemann operators

q
can be written in the following form:
(2.46)
+
q
=
2
z
m

z
and

q
=
2
z
m

z
.
SOBOLEV REGULARITY OF SOLUTIONS 21
Let C

(p) be the space of germs at p M of smooth complex-valued


functions on M and for any u C

(p) let
(2.47) u(z, z) =

i,jN
a
ij
(u, p)z
i
z
j
.
be its (formal) Taylor series at p (with respect to the canonical coordinate).
Lemma 2.12. Let p
q
be a zero of (even) order 2m of the (orientable)
quadratic differential q on M. For any k N, a germ
u C

(p)

H
k
q
(p) a
ij
(u, p) = 0 , for all i + j (k 1)(m+ 1) ,
except all pairs (i, j) for which one of the following conditions holds:
(2.48)
(1) i N (m+ 1) , j N (m + 1) ;
(2) i N (m+ 1) , j , N (m + 1) and i < j ;
(3) i , N (m+ 1) , j N (m + 1) and i > j .
Proof. Let u C

(p). For any n N, there is a local Taylor expansion


u(z, z) =

i+jn
a
ij
(u, p)z
i
z
j
+ R
u
n
(z, z)
where the remainder R
u
n
is a smooth function vanishing at order n at p.
A straightforward calculation (based on formulas (2.46) yields that any
smooth function R vanishing at p at order n belongs to the space H
k
q
(p)

H
k
q
(p) if n > (k 1)m. It follows that u

H
k
(p) iff its Taylor polynomial
of any order n > (k 1)m does. The argument can therefore be reduced to
the case of polynomials.
It follows from formulas (2.46) that, for all N and all (i, j) N N,
there exists a complex constant c
m,
ij
such that
(2.49)

q
(z
i
z
j
) = c
m,
ij
z
i(m+1)
z
j(m+1)
.
The area form of the quadratic differential q can be written as
(2.50)
q
= [z[
2m
dx dy
with respect to the canonical coordinate z := x+y. Hence, straightforward
computations in polar coordinates yield that, if c
m,
ij
,= 0,
(2.51)

q
(z
i
z
j
) H
0
q
(p) i + j 2(m+ 1) > (m+ 1) ;

q
(z
i
z
j
) H
1
q
(p) i + j 2(m+ 1) > 0 .
If c
m,
ij
= 0, then either i N (m+1) and i < (m+1) or j N (m+1)
and j < (m+ 1). It follows that, if i, j , N (m + 1), then
(2.52) z
i
z
j


H
k
q
(p) i + j (k 1)(m+ 1) > 0 .
22 GIOVANNI FORNI

If i N (m + 1), i = h(m + 1), and j , N (m + 1), then conditions
(2.51) apply for all h, hence (2.52) holds if k 2h, while if k > 2h,
(2.53) z
i
z
j


H
k
q
(p)
h
q
(z
i
z
j
) H
1
q
(p) j > i .
Similarly, if j N (m+1), j = h(m+1), and i , N (m+1), then (2.52)
holds if k 2h, while if k > 2h,
(2.54) z
i
z
j


H
k
q
(p)
h
q
(z
i
z
j
) H
1
q
(p) i > j .
It follows immediately from (2.52), (2.53) and (2.54) that the conditions
listed in the statement of the lemma are sufcient. The necessity follows
from the following argument. For any r
1
< r
2
, let D(r
1
, r
2
) M be the an-
nulus (centered at p) dened by the inequalities r
1
< [z[ < r
2
. The system
of Laurent monomials z
i
z
j
[i, j Z is orthogonal in

H
k
(D
r
1
,r
2
). In fact, a
computation in polar coordinates shows that, for all (i, j) ,= (i

, j

) ZZ,
(2.55)
_
D(r
1
,r
2
)

q
(z
i
z
j
)

q
( z
i

z
j

)
q
= 0 ,
hence
(2.56) |

i+jn
a
ij
z
i
z
j
|
2
k
=

i+jn
[a
ij
[
2
|z
i
z
j
|
2
k
.
It follows that only Laurent monomials z
i
z
j


H
k
(p) can appear in the
Taylor expansion of a function f C

(p)

H
k
(p).
Lemma 2.13. Let p
q
be a zero of order 2mof the quadratic differential
q on M. For any k N, a germ
u C

(p) H
k
q
(p) a
ij
(u, p) = 0 , for all i + j (k 1)(m+ 1) ,
except all pairs (i, j) N (m + 1) N (m+ 1).
Proof. The proof is similar to that of Lemma 2.12 above. In fact, formulas
(2.49) are replaced by the following formulas. For all N and all i N,
there exists a complex constant c
m,
i
such that
(2.57) (
+
q
)

q
)

z
i
z
j
= c
m,
i
c
m,
j
z
i(m+1)
z
j(m+1)
.
As in the proof of Lemma 2.12, it follows by a straightforward computation
in polar coordinates that, if c
m,
i
c
m,
j
,= 0,
(2.58) (
+
)

z
i
z
j
H
0
q
(p) i +j (+)(m+1) > (m+1) .
Since there exists N such that c
m,
i
= 0 iff i N (m + 1), either
i, j N (m + 1), in which case z
i
z
j
H
k
q
(p), or there exists (, )
such that + = k and c
m,
i
c
m,
j
,= 0, in which case z
i
z
j
H
k
q
(p) iff
i + j > (k 1)(m+ 1).
SOBOLEV REGULARITY OF SOLUTIONS 23
Let p
q
be a zero of order 2m
p
of the orientable quadratic differential q.
Let T
p
N N be the set of (i, j) such that
(2.59)
i N (m
p
+ 1) , j , N (m
p
+ 1) and i < j or
i , N (m
p
+ 1) , j N (m
p
+ 1) and i > j,
For any (i, j) T
p
, let
ij
p
be the linear functional (distribution) on C

(M)
dened as follows. Let u(z, z) =

a
ij
(u, p)z
i
z
j
denote the Taylor expan-
sion of u C

(M) at p
q
with respect to the canonical coordinate
z : D
p
C for the differential q at p. Let
(2.60)
ij
p
(u) := a
ij
(u, p) .
It is clear from the denition that
ij
p
=
ji
p
for all (i, j) T
p
. A calculation
shows that, for any h NN(m
p
+1) we have the following representation
in terms of the Cauchy principal value: for any u C

(p),
(2.61)

h0
p
(u) =
1
4h
PV
_
M

q
u
z
h

q
;

0h
p
(u) =
1
4h
PV
_
M

q
u
z
h

q
.
(The above formulas can be derived one from the other by conjugation). In
addition, for any N, by formulas (2.49) there exist complex constants
c
mp,
0,h
,= 0 and c
mp,
h,0
,= 0 such that the following identities hold in the sense
of distributions:
(2.62)
c
mp,
0,h

(mp+1),(mp+1)+h
p
=

q
_

0,h
p
_
;
c
mp,
h,0

(mp+1)+h,(mp+1)
p
=

q
_

h,0
p
_
.
Let T
k
p
T
p
be the subset of (i, j) such that i + j (k 1)(m
p
+ 1).
Lemma 2.14. For each (i, j) T
k
p
, the functional
ij
p
has a unique (non-
trivial) continuous extension to the space

H
k
q
(p) and the following holds:
(2.63) H
k
q
(p) = u

H
k
q
(p) [
ij
p
(u) = 0 for all (i, j) T
k
p
.
Proof. The functions z
h
and z
h
L
2
q
(D) for all 1 h m
p
and the
operator
F
q
:

H
2
q
(p) L
2
q
(p) is bounded. Hence the linear functionals

0h
p
and
h0
p
are continuous on

H
2
q
(p) for all 1 h m
p
. Similarly, the
distributions PV(z
h
) and PV( z
h
) H
1
(D
p
) for all m
p
< h < 2(m
p
+
1) and the operator
F
q
:

H
3
q
(p) H
1
(p) is bounded. Hence the linear
functionals
0h
p
and
h0
p
are continuous on

H
3
q
(p) for all m
p
< h < 2(m
p
+
1). Since the space H
k
q
(p) is equal to the closure in

H
k
q
(p) of the subspace
C

(p) H
k
q
(p), the statement for k = 2, k = 3 follows from Lemmas 2.12
and 2.13.
24 GIOVANNI FORNI

We complete the argument by induction on k N. The Friedrichs extension
denes bouned operators
F
q
:

H
k+1
q
(p)

H
k1
q
(p) and its dual
F
q
:

H
k+1
q
(p)

H
k1
q
(p). By the induction hypothesis, since all functionals
in T
k1
p
extend (uniquely) to bounded functionals in

H
k+1
q
(p), it follows
that all functionals in
F
q
(T
k1
p
) extend (uniquely) to bounded functionals
in

H
k1
q
(p) and the following holds. For any u

H
k+1
q
(p),
(2.64)
F
q
u H
k1
q
(p)
ij
p
(u) = 0 for all
ij
p

F
q
(T
k1
p
) .
Let E
k+1
q


H
k+1
q
(p) the closed nite-codimensional subspace dened as
(2.65) E
k+1
q
:= u

H
k+1
q
(p) [
ij
p
(u) = 0 for all
ij
p

F
q
(T
k1
p
) .
By formulas (2.62), any distribution
p
T
k+1
p

F
q
(T
k1
p
) is of the form

p
=
0h
p
or
p
=
h0
p
with 1 h k(m
p
+ 1). By formulas (2.61), such
distributions have a (unique) continuous extension to the subspace E
k+1
q

H
k+1
q
(p). In fact, the distributions PV(z
h
) and PV( z
h
) H
k+1
q
(D
p
),
for all 1 h < k(m
p
+ 1), and
F
q
(f) H
k1
q
(D) for all f E
k+1
q
.
Hence all distributions in the set T
k+1
p
have a continuous extension to the
space

H
k+1
q
(p) and the characterization (2.63) of the subspace H
k+1
q
(p)

H
k+1
q
(p) follows by Lemmas 2.12 and 2.13.

Let D
k
q
be the set of all continuous extensions to the space

H
k
q
(M) of the
functionals
ij
p
for all p
q
and all (i, j) T
k
p
. Let
(2.66) D
q
:=
_
kN
D
k
q
.
Theorem 2.15. The (closed) kernel of the system D
k
q
on

H
k
q
(M) coincides
with the subspace H
k
q
(M), that is,
H
k
q
(M) = u

H
k
q
(M) [ (u) = 0 , for all D
k
q
.
2.4. Smoothing operators. We will establish below ner results on the
Sobolev regularity of the distributions in D
q
. The key step will be to con-
struct smoothing operators for the scale of Sobolev spaces H
k
q
(p) [ k N.
For any p
q
, let z : D
p
C be a canonical coordinate for the (ori-
entable) quadratic differential q such that p D
p
and z(p) = 0. For any
(i, j) N N, let Z
ij
p
C

(M) be a function such that


(2.67) Z
ij
p
(z) z
i
z
j
on D
p
.
Lemma 2.16. Let p
q
be a zero of order 2m of the quadratic differen-
tial q on M. There exists a one-parameter family K
p
() [ (0, 1] of
bounded operators K
p
() : C

(M) H

q
(M) such that the following
SOBOLEV REGULARITY OF SOLUTIONS 25
estimates hold. For each k N, there exists a constant C
k
> 0 such that,
for any (i, j) N N and for all (0, 1]:
(2.68)
[K
p
()(Z
ij
p
) Z
ij
p
[
k
C
k

1+
i+j
m+1
k
, for k < 1 +
i + j
m + 1
;
[K
p
()(Z
ij
p
)[
k
C
k
[ log [
1/2
, for k = 1 +
i + j
m + 1
;
[K
p
()(Z
ij
p
)[
k
C
k

[k(1+
i+j
m+1
)]
, for k > 1 +
i + j
m + 1
.
Proof. Let z : D
p
C be a canonical coordinate at p
q
dened on an
open neighbourhood D
p
M such that D
p

q
= p. There exists r
1
> 0
such that D(r
1
) z(D
p
), where D(r
1
) is the euclidean disk centered at
the origin of radius r
1
> 0 . Let 0 < r
2
< r
1
and let : C R be
any non-negative smooth function identically zero on the closure of D(r
2
)
and identically equal to 1 outside D(r
1
). Let D

p
D
p
be any relatively
compact neighbourhood of p in D
p
such that D(r
1
) z(D

p
). For any
(0, 1], let us dene, for all F C

(M),
(2.69) K
p
() (F) (x) :=
_
(

1
m+1
z(x)) F (z(x)) , for x D
p
;
F(x) , for x , D

p
.
Clearly, the denition (2.69) is well-posed for all (0, 1] and the func-
tions K
p
() (F) H

q
(M), since the rescaled functions

: C R
dened as

(z) := (

1
m+1
z) , for z C,
are smooth and have support away from the origin. Since the functions

are bounded uniformly with respect to the parameter (0, 1], for any
function F C

(M) L
2
q
(M), by the dominated convergence theorem,
(2.70) [K
p
() (F) F[
0
0 , as 0
+
.
Let us denote for convenience, for any (a, b) N N,

ab

(z) := [(
+
q
)
a
(

q
)
b
](

1
m+1
z) , for z C.
The functions
ab

are smooth and bounded on C, uniformly with respect to


> 0. For any (a, b) such that (a, b) ,= (0, 0), the function
ab

has compact
support contained in the euclidean annulus centered at the origin of inner
radius r
1

1
m+1
and outer radius r
2

1
m+1
. For a = b = 0,
ab

has
support outside the euclidean disk of radius r
1

1
m+1
and is identically equal
to 1 outside the disk of radius r
2

1
m+1
.
For any (i, j) N N, let K
ij

:= K
p
()
_
Z
ij
p
_
H

q
(M). A calculation
shows that the following formulas hold. For all m, i and N and all
26 GIOVANNI FORNI

a N such that 0 a , let
(2.71) C
m,i
,a
=
_

a
_
a

=1
[i (m+ 1)] .
For any (, ) N N, the derivative (
+
q
)

q
)

(K
ij

) (z) is given on
D
p
p by the sum
(2.72)

a=0

b=0
C
m,i
,a
C
m,j
,b

ab

(z)
(a+b)
z
i(a)(m+1)
z
j(b)(m+1)
.
If i + j > ( + 1)(m+ 1), since the functions
z
i(a)(m+1)
z
j(b)(m+1)
L
2
q
(D
p
) ,
for all 0 a and 0 b , and
ab

is bounded, by change of
variables we obtain that for (a, b) ,= (0, 0) there exists a constant C
a,b
> 0
such that
(2.73) [
ab

(z)
(a+b)
z
i(a)(m+1)
z
j(b)(m+1)
[
0
C
a,b

1+
i+j
m+1
(+)
,
and that similarly, for a = b = 0, there exists a constant C
0
> 0
(2.74) [(

(z) 1) z
i(m+1)
z
j(m+1)
[
0
C
0

1+
i+j
m+1
(+)
.
If i +j < [( a) + ( b) 1](m+ 1), there exists a constant C

a,b
> 0
such that
(2.75) [
ab

(z)
(a+b)
z
i(a)(m+1)
z
j(b)(m+1)
[
0
C

a,b

1+
i+j
m+1
(+)
,
since the function
ab

is bounded and supported outside the euclidean disk


of radius r
1

1
m+1
centered at the origin.
If i + j = [( a) + ( b) 1](m+ 1), a similar calculation yields
(2.76)
[
ab

(z)
(a+b)
z
i(a)(m+1)
z
j(b)(m+1)
[
0
C

a,b

1+
i+j
m+1
(+)
[ log [
1
2
.
By formula (2.72) for the Cauchy-Riemann iterated derivatives, the required
estimates (2.68) follow immediately from estimates (2.73), (2.74), (2.75)
and (2.76).
We derive below estimates for the local smoothing family K
p
() con-
structed in Lemma 2.16 with respect to the fractional weighted Sobolev
norms. Let p
q
be any zero of order 2m
p
of the quadratic differential
SOBOLEV REGULARITY OF SOLUTIONS 27
q on M. For each pair (i, j) N N, let e
ij
p
: R
+
[0, 1] denote the
function dened as follows:
(2.77) e
ij
p
(s) :=
_

_
s , if [s] = 1 +
i+j
mp+1
;
1 s , if [s] =
i+j
mp+1
;
0 , otherwise .
Theorem 2.17. The family K
p
() [ (0, 1] of local smoothing oper-
ators K
p
() : C

(M) H

q
(M), dened in (2.69), has the following
properties. For each s R
+
, there exists a constant C
s
> 0 such that, for
any pair (i, j) N N and all (0, 1]:
[K
p
()(Z
ij
p
) Z
ij
p
[
s
C
s

1+
i+j
mp+1
s
[ log [
e
ij
p
(s)
2
, for s < 1 +
i + j
m
p
+ 1
;
[K
p
()(Z
ij
p
)[
s
C
s
[ log [ [ log [
e
ij
p
(s)
2
, for s = 1 +
i + j
m
p
+ 1
;
[K
p
()(Z
ij
p
)[
s
C
s

[s(1+
i+j
mp+1
)]
[ log [
e
ij
p
(s)
2
, for s > 1 +
i + j
m
p
+ 1
.
Proof. For any k N, the function Z
ij
p
H
k
q
(M), if k < 1+(i+j)/(m+1),
since z
i
z
j
H
k
q
(p). By Lemma 2.16, there exists a constant C
k
> 0 such
that, for all (0, 1],
(2.78)
[K
ij
p
() K
ij
p
(/2)[
k
[(K
ij
p
() Z
ij
p
) (K
ij
p
(/2) Z
ij
p
)[
k
;
[K
ij
p
() Z
ij
p
[
k
+[K
ij
p
(/2) Z
ij
p
[
k
C
k

1+
i+j
m+1
k
.
If k 1 + (i + j)/(m+ 1),
(2.79) [K
ij
p
() K
ij
p
(/2)[
k
[K
ij
p
()[
k
+ [K
ij
p
(/2)[
k
hence, by Lemma 2.16,
(2.80)
[K
ij
p
() K
ij
p
(/2)[
k
2C
k
[ log [
1/2
, if k = 1 +
i + j
m+ 1
;
[K
ij
p
() K
ij
p
(/2)[
k
2C
k

1+
i+j
m+1
k
, if k > 1 +
i + j
m+ 1
.
As a consequence, by the interpolation inequality (Lemma 2.10), for every
s R
+
there exists a constant C
s
> 0 such that, for all (0, 1],
[K
ij
p
() K
ij
p
(/2)[
s
C
s

1+
i+j
m+1
s
, if [s], [s] + 1 ,= 1 +
i + j
m + 1
;
[K
ij
p
() K
ij
p
(/2)[
s
C
s

1+
i+j
m+1
s
[ log [
1{s}
2
, if [s] = 1 +
i + j
m+ 1
;
[K
ij
p
() K
ij
p
(/2)[
s
C
s

1+
i+j
m+1
s
[ log [
{s}
2
, if [s] =
i + j
m + 1
.
28 GIOVANNI FORNI

If s < 1 + (i + j)/(m+ 1) and [s] ,= (i + j)/(m+ 1), for all n N,
(2.81) [K
ij
p
(/2
n
) K
ij
p
(/2
n+1
)[
s
C
s

1+
i+j
m+1
s
2
n(1+
i+j
m+1
s)
,
hence, for any xed (0, 1], the sequence K
ij
p
(/2
n
)
nN
is Cauchy,
and therefore convergent, in the Hilbert space H
s
q
(M). By Lemma 2.16
K
ij
p
(/2
n
)
nN
converges to Z
ij
p
in H
[s]
q
(M). Since H
[s]
q
(M) H
s
q
(M),
by uniqueness of the limit Z
ij
p
H
s
q
(M) and K
ij
p
(/2
n
)
nN
converges to
Z
ij
p
in H
s
q
(M). The estimate (2.81) also implies that
(2.82)
[K
ij
p
() Z
ij
p
[
s

nN
[K
ij
p
(/2
n
) K
ij
p
(/2
n+1
)[
s
C

1+
i+j
m+1
s
.
If s < 1+(i +j)/(m+1) and [s] = (i +j)/(m+1), by a similar argument
we again get that Z
ij
p
H
s
q
(M) and
(2.83) [K
ij
p
() Z
ij
p
[
s
C

1+
i+j
m+1
s
[ log [
{s}
2
.
If s 1 +(i +j)/(m+1) and [s], [s] +1 ,= 1 +(i +j)/(m+1) we argue
as follows. For each 1/2, we have
(2.84) [K
ij
p
(2) K
ij
p
()[
s
C
s
(2)
1+
i+j
m+1
s
,
hence if 2
n
, for all 0 k < n,
(2.85) [K
ij
p
(2
k+1
) K
ij
p
(2
k
)[
s
C
s
2
(k+1)(1+
i+j
m+1
s)

1+
i+j
m+1
s
.
It follows that, there exists a constant C

s
> 0 such that
(2.86) [K
ij
p
(2
n
) K
ij
p
()[
s
C

s
2
n(1+
i+j
m+1
s)

1+
i+j
m+1
s
.
For every (0, 1], let n() be the maximum n N such that 2
n
1.
By this denition it follows that 1/2 < 2
n()
1. Since
sup
1/21
[K
ij
p
()[
s
sup
1/21
[K
ij
p
()[
[s]+1
< + ,
it follows that, there exists a constant C

s
> 0 such that
(2.87)
[K
ij
p
()[
s
C

s

1+
i+j
m+1
s
, if s > 1 +
i + j
m + 1
;
[K
ij
p
()[
s
C

s
[ log [ , if s = 1 +
i + j
m + 1
.
By a similar argument, for s > 1 + (i + j)/(m+ 1) we have
(2.88)
[K
ij
p
()[
s
C

s

1+
i+j
m+1
s
[ log [
1{s}
2
, if [s] = 1 +
i + j
m+ 1
;
[K
ij
p
()[
s
C

s

1+
i+j
m+1
s
[ log [
{s}
2
, if [s] =
i + j
m + 1
,
SOBOLEV REGULARITY OF SOLUTIONS 29
while for s = 1 + (i + j)/(m+ 1) we have
(2.89)
[K
ij
p
()[
s
C

s
[ log [ [ log [
1{s}
2
, if [s] = 1 +
i + j
m+ 1
;
[K
ij
p
()[
s
C

s
[ log [ [ log [
{s}
2
, if [s] =
i + j
m + 1
.

Theorem 2.17 implies in particular the following smoothness results.


Corollary 2.18. Let z : D
p
C be a canonical coordinate for an ori-
entable quadratic differential q at a zero p
q
of order 2m. For each
(i, j) N N, the function
(2.90) z
i
z
j
H
s
q
(p) , for all s < 1 +
i + j
m
p
+ 1
.
Corollary 2.19. Let p
q
be a zero of order 2m
p
and let (i, j) T
p
. The
distribution
ij
p
has the following regularity properties:
(2.91)

ij
p


H
s
q
(p) for s > 1 +
i + j
m
p
+ 1
,

ij
p
, H
s
q
(p) for s < 1 +
i + j
m
p
+ 1
.
Proof. By the formulas (2.61) and by Lemma 2.14, for any h N N
(m
p
+ 1) and for any N, there exist constants C
mp,
h0
,= 0 and C
mp,
0h
,= 0
such that the following identities hold in the dual Hilbert space

H
k
q
(p) for
any integer k 1 +h/(m
p
+ 1):
(2.92)
C
mp,
h0

h0
p
=
+1
q
_
z
(mp+1)h
z
(mp+1)
_
,
C
mp,
0h

0h
p
=
+1
q
_
z
(mp+1)
z
(mp+1)h
_
.
By Corollary 2.18, if (m
p
+ 1) h > 0, for all s < 2 + 1 h/(m
p
+ 1),
z
(mp+1)h
z
(mp+1)
and z
(mp+1)
z
(mp+1)h
H
s
q
(p)

H
s
q
(p) .
Hence
h0
p
,
0h
p


H
s
q
(p) for all s > 1 +h/(m
p
+1). By formulas (2.62) it
then follows that
ij
p


H
s
q
(p) for all s > 1 +(i +j)/(m
p
+1) as claimed.
Let (i, j) T
p
and s < 1+(i+j)/(m
p
+1). By Corollary 2.18, the function
z
i
z
j
H
s
q
(p). Since by denition H

q
(p) is dense in H
s
q
(p) for any s > 0,
the functional
ij
p
0 on H

q
(p) and
ij
p
(z
i
z
j
) = 1, it follows that
ij
p
does
not extend to a bounded functional on H
s
q
(p).
30 GIOVANNI FORNI

Let p
q
be a zero of (even) order 2m
p
the quadratic differential q on M.
For every s R
+
, let T
s
p
T
p
be the subset dened as
(2.93) T
s
p
:= (i, j) T
p
[ i + j < (s 1)(m
p
+ 1) .
Let D
s
q


H
s
q
(M) be the set of distributions dened as follows:
(2.94) D
s
q
:=
ij
p
[ p
q
and (i, j) T
s
p
.
Corollary 2.20. The closure of the subspace H
s
q
(M) in

H
s
q
(M) is a subset
of the (closed) kernel of the system D
s
q
on

H
s
q
(M), that is,
(2.95) H
s
q
(M) u

H
s
q
(M) [ (u) = 0 , for all D
s
q

The reverse inclusion holds if the following sufcient condition is satised:


(2.96) s , 1 + (i + j)/(m
p
+ 1) [ p
q
and (i, j) T
p
.
Proof. Since H

q
(M) is dense in H
s
q
(M)

H
s
q
(M),
H

q
(M) u

H
s
q
(M) [ (u) = 0 , for all D
s
q

and D
s
q


H
s
q
(M), it follows that
H
s
q
(M) u

H
s
q
(M) [ (u) = 0 , for all D
s
q
.
Conversely, if condition (2.96) is satised, by Corollary 2.18 the subspace
u C

(M) [ (u) = 0 , for all D


s
q
H
s
q
(M) .
Since C

(M)

H
s
q
(M) is dense in

H
s
q
(M), the result follows.
The regularity result proved in Corollary 2.18 extends to a certain subset of
all pairs (i, j) Z Z if the functions z
i
z
j
are interpreted as distributions
in the sense of the Cauchy principal value:
(2.97) PV
_
z
i
z
j
_
(v) := PV
_
M
z
i
z
j
v
q
, for all v C

(p) .
The most general regularity result for the distributions (2.97) is based on the
following generalization of Corollary 2.18 to include logarithmic factors.
Lemma 2.21. Let z : D
p
C be a canonical coordinate for an orientable
holomorphic quadratic differential q at a zero p
q
of order 2m. For
each (i, j, h) N N N, the function
z
i
z
j
log
h
[z[ H
s
q
(p) , for all s < 1 +
i + j
m
p
+ 1
.
SOBOLEV REGULARITY OF SOLUTIONS 31
Proof. Simple calculations show that log [z[ L
2
q
(M) and that by formulas
(2.46) the following identities hold on D
p
p:
(2.98)
+
log [z[ =
1
z
m+1
and

log [z[ =
1
z
m+1
.
It follows that, for each (i, j, h) NNNand each (, ) NN, there
exists a nite sequence of non-zero constants C
1
, . . . , C
h
, which depend on
(i, j, h, , , m), such that the following identity holds on D
p
p:
(2.99) (
+
)

_
z
i
z
j
log
h
[z[
_
= z
i(m+1)
z
j(m+1)
h

=0
C

log

[z[ .
For all (i, j, h) NNN, let L
ijh
p
C

(M p) be any function such


that L
ijh
p
(z) = z
i
z
j
log
h
[z[ for all z D
p
. By (2.99), the function
L
ijh
p
H
k
q
(M) , if k N and k < 1 +
i + j
m
p
+ 1
Let K
p
() [ (0, 1] be the family of local smoothing operators dened
by formulas (2.69). By computations similar to those carried out in the
proof of Lemma 2.16, based on formulas (2.99), it is possible to prove that
for each k N, there exists a constant C
k
> 0 such that for all (0, 1]:
[K
p
()(L
ijh
p
) L
ijh
p
[
k
C
k

1+
i+j
m+1
k
[ log [
h
, for k < 1 +
i + j
m+ 1
;
[K
p
()(L
ijh
p
)[
k
C
k
[ log [
1/2
[ log [
h
, for k = 1 +
i + j
m+ 1
;
[K
p
()(L
ijh
p
)[
k
C
k

[k(1+
i+j
m+1
)]
[ log [
h
, for k > 1 +
i + j
m+ 1
.
Reasoning as in the proof of Theorem 2.17, we can derive similar estimates
for fractional Sobolev norms. For each (i, j) NN, let e
ij
p
: R
+
[0, 1]
be the function dened in formula (2.77). By the interpolation Lemma 2.10,
for any s < 1 + (i + j)/(m+ 1) there exists a constant C
s
> 0 such that
[K
p
()(L
ijh
p
) K
p
(/2)(L
ijh
p
)[
s
C
s

1+
i+j
m+1
s
[ log [
h+
e
ij
p
(s)
2
.
It follows that the sequence K
p
(/2
n
)(L
ijh
p
)
nN
is Cauchy and therefore
converges in H
s
q
(M). By uniqueness of the limit
L
ijh
p
H
s
q
(M) , for all s < 1 +
i + j
m
p
+ 1
.
32 GIOVANNI FORNI

In addition, the following estimates hold. For each s R
+
there exists a
constant C

s
> 0 such that for all (0, 1]:
[K
p
()(L
ijh
p
)[
s
C

s
[ log [ [ log [
h+
e
ij
p
(s)
2
, for s = 1 +
i + j
m
p
+ 1
;
[K
p
()(L
ijh
p
)[
s
C

[s(1+
i+j
mp+1
)]
[ log [
h+
e
ij
p
(s)
2
, for s > 1 +
i + j
m
p
+ 1
.

Theorem 2.22. Let z : D


p
C be a canonical coordinate for an ori-
entable holomorphic quadratic differential q at a zero p
q
of order
2m
p
. For each (i, j) Z Z such that (1) i j , Z (m
p
+ 1) or (2)
i > (m
p
+ 1) or (3) j > (m
p
+ 1), the distribution
(2.100) PV
_
z
i
z
j
log
h
[z[
_
H
s
q
(p) , for all s < 1 +
i + j
m
p
+ 1
.
Proof. For all (i, j) Z Z such that i j , Z (m
p
+ 1) and for any
h Z the following formulas hold for all functions v H

q
(p):
(2.101)
(a) PV
_
M

+
(z
i
z
j
log
h
[z[)v
q
= PV
_
M
z
i
z
j
log
h
[z[
+
v
q
;
(b) PV
_
M

(z
i
z
j
log
h
[z[)v
q
= PV
_
M
z
i
z
j
log
h
[z[

v
q
.
Formulas (2.101) also hold in case (a) if i > (m
p
+1), j Z, and in case
(b) if j > (m
p
+ 1), i Z, for all germs v C

(p).
By taking into account the formulas (2.46) for the Cauchy-Riemann opera-
tors with respect to a canonical coordinate, it follows from formulas (2.101)
by induction on h N that if
PV(z
i
z
j
log
h
[z[) H
s
q
(p) , for all h N,
then, if i j , Z (m
p
+ 1) or i > (m
p
+ 1) and j Z,
PV(z
i
z
j(mp+1)
log
h
[z[) H
s1
q
(p) , for all h N,
and, if i j , Z (m
p
+ 1) or j > (m
p
+ 1) and i Z,
PV(z
i(mp+1)
z
j
log
h
[z[) H
s1
q
(p) , for all h N.
Thus, the statement of the theorem can be derived from Corollary 2.18, by
an induction argument based on formulas (2.101).
Corollary 2.23. Let z : D
p
C be a canonical coordinate for an ori-
entable holomorphic quadratic differential q at a zero p
q
of order 2m
p
.
SOBOLEV REGULARITY OF SOLUTIONS 33
If (i, j) , Z (m
p
+ 1) Z (m
p
+ 1), the distribution
(2.102) PV
_
z
i
z
j
log
h
[z[
_
, H
s
q
(p) , for s > 1 +
i + j
m
p
+ 1
,
and, if i j Z (m
p
+ 1) and both i (m + 1) and j (m + 1),
(2.103) PV
_
z
i
z
j
log
h
[z[
_
, H

q
(p) .
Proof. We argue by contradiction. Assume there exists (i, j) Z Z such
that (i, j) , Z (m
p
+ 1) Z (m
p
+ 1) and PV(z
i
z
j
) H
r
q
(p) for
some r > 1 + (i + j)/(m
p
+ 1). By taking Cauchy-Riemann derivatives
if necessary, we can assume that i 0 and j 0. By Theorem 2.22, the
distribution
PV
_
z
i(m+1)
z
j(m+1)
_
H
s
q
(p) , for all s < 1
i + j
m
p
+ 1
.
It follows that, for any positive smooth function C

0
(D
p
) identically
equal to 1 on a disk D

p
D
p
centered at p
q
, the principal value
PV
_
M
(z) z
i
z
j
z
i(m+1)
z
j(m+1)

q
is nite, which can be proved to be false by a simple computation in (ge-
odesic) polar coordinates. This contradiction proves the rst part of the
statement.
If i j Z (m
p
+1) and both i (m+1) and j (m+1), we argue
as follows. It is not restrictive to assume that i j, hence the function
z
ij
H

q
(M). However, by a computation in polar coordinates, since
i (m
p
+ 1),
PV
_
M
(z) z
i
z
j
log [z[ z
ij

q
= +.
It follows that PV
_
z
i
z
j
log
h
[z[
_
, H

q
(p), hence the second part of the
statement is also proved.
We conclude this section with a fundamental smoothing theorem for the
1-parameter family of weighted Sobolev spaces.
Theorem 2.24. For each k N, there exists a family S
k
() [ (0, 1]
of bounded operators S
k
() : L
2
q
(M) H
k
q
(M) such that the following
estimates hold. For any s, r [0, k] and for any > 0, there exists a
constant C
k
r,s
() > 0 such that, for all u H
s
q
(M) and for all (0, 1]:
(2.104)
[S
k
()(u) u[
r
C
k
r,s
()|u|
s

sr
, if s > r ;
[S
k
()(u)[
r
C
k
r,s
()|u|
s

sr
, if s r .
34 GIOVANNI FORNI

Proof. For each p
q
, let z : D
p
C be a canonical coordinate dened
on a disk D
p
(centered at p) such that D
p

q
= p. For each (i, j) T
p
,
let Z
ij
p
C

(M) be a (xed) smooth extension, as in (2.67), of the locally


dened function z
i
z
j
C

(p). Let P
k
be the linear operator dened as
follows:
(2.105) P
k
(f) := f

pq

(i,j)T
k
p

ij
p
(f) Z
ij
p
, for all f

H
k
q
(M) .
The operator P
k
:

H
k
q
(M) H
k
q
(M) is well-dened and bounded. It is
well-dened by Lemma 2.14 and Theorem 2.15. It is bounded since, for all
p
q
and all (i, j) T
p
, the functions Z
ij
p


H

q
(M) by Lemma 2.12
and, for all (i, j) T

p
, the distributions
ij
p


H
k
q
(M) by Corollary 2.19.
In fact, the condition (i, j) T

p
implies i + j < (k 1)(m
p
+ 1).
For each p
q
, let K
p
() [ (0, 1] be the family of local smoothing
operators constructed in Lemma 2.16. Let S
k

[ (0, 1] be the one-


parameter family of bounded linear operators S
k

:

H
k
q
(M) H
k
q
(M)
dened as follows. For all f

H
k
q
(M), we let
(2.106) S
k

(f) := P
k
(f) +

pq

(i,j)T
k
p

ij
p
(f) K
p
()
_
Z
ij
p
_
.
By denition the following identity holds for all f

H
k
q
(M):
S
k

(f) f =

pq

(i,j)T
k
p

ij
p
(f)
_
K
p
()
_
Z
ij
p
_
Z
ij
p

.
Since for all p
q
the condition (i, j) T
p
implies i+j , N (m
p
+1), by
Lemma 2.16 the following estimate holds. For each N such that k,
there exists a constant C
k

> 0 such that, for any f



H
k
q
(M),
(2.107) |S
k

(f) f|

C
k

pq

(i,j)T
k
p

1+
i+j
mp+1

[
ij
p
(f)[ .
In fact, for each p
q
and each (i, j) T
p
, since Z
ij
p
H

q
(M), which
implies K
p
()
_
Z
ij
p
_
Z
ij
p
H

q
(M), if < 1 + (i + j)/(m
p
+ 1), by
Lemma 2.16, there exist constants C

> 0, C

> 0 such that


|K
p
()
_
Z
ij
p
_
Z
ij
p
|

[K
p
()
_
Z
ij
p
_
Z
ij
p
[


1+
i+j
mp+1

,
while, if > 1 + (i + j)/(m
p
+ 1), since Z
ij
p


H

q
(M),
|K
p
()
_
Z
ij
p
_
Z
ij
p
|

|K
p
()
_
Z
ij
p
_
|

+|Z
ij
p
|


1+
i+j
mp+1

.
The scale of Friedrichs Sobolev spaces admits a standard family of smooth-
ing operators T

[ > 0 such that the operator T

: L
2
q
(M)

H

q
(M)
SOBOLEV REGULARITY OF SOLUTIONS 35
is dened, for each > 0, by the following truncation of Fourier series. Let
e
n
[ n N be an orthonormal basis of eigenfunctions of the Friedrichs
Laplacian
F
q
and let : N R
+
0 be the corresponding sequence of
eigenvalues. Then
T

(u) :=

2
n1
u, e
n
)
q
e
n
, if u =

nN
u, e
n
)
q
e
n
.
If u

H
s
q
(M), then T

(u) u in

H
s
q
(M) (as 0
+
) and the following
estimates hold. For all r R
+
, there exists a constant C
r,s
(q) > 0 such that
(2.108)
|T

(u) u|
r
C
r,s
(q) |u|
s

sr
, if s r ;
|T

(u)|
r
C
r,s
(q) |u|
s

(rs)
, if r s .
If p
q
and (i, j) T
p
, the distribution
ij
p


H
s
ij
q
(M) for any s
ij
> 0
such that i+j < (s
ij
1)(m
p
+1). Hence there exists a constant C
ij
p
(q) > 0
such that, by estimates (2.108), for all u

H
s
q
(M),
(2.109) [
ij
p
(T

(u)) [ C
ij
p
(q) |u|
s
max1,
ss
ij
.
If u H
s
q
(M), then
p
ij
(u) = 0, for all p
q
and all (i, j) T
s
p
. Hence, if
s
ij
s and i + j < (s
ij
1)(m
p
+ 1),
(2.110) [
ij
p
(T

(u)) [ = [
ij
p
(T

(u) u) [ C
ij
p
(q) |u|
s

ss
ij
.
The following estimates hold. Let s R
+
and N. For any > 0, there
exists a constant C
,s
() > 0 such that, for all (0, 1] and all u H
s
q
(M),
(2.111) |S
k

(u) T

(u)|

C
,s
() |u|
s

s
.
In fact, if p
q
and (i, j) T
k
p
,
(2.112) [
ij
p
(T

(u)) [ C
ij
p
(q) |u|
s

ss
ij
;
for any s
ij
> 1 + (i + j)/(m
p
+ 1) s, if (i, j) T
k
p
T
s
p
, and for any
1+(i+j)/(m
p
+1) < s
ij
s, if (i, j) T
s
p
. The claim (2.111) then follows
from (2.107). By estimates (2.108) and (2.111), for any > 0, there exists
a constant C

,s
() > 0 such that, for all (0, 1] and for all u H
s
q
(M),
(2.113) |S
k

(u) u|

,s
() |u|
s

s
.
Let S
k
() [ (0, 1] be the family of operators S
k
() : L
2
q
(M)
H
k
q
(M) dened as follows: for each (0, 1],
(2.114) S
k
() := S
k

.
By estimate (2.113), for any > 0, there exists a constant C

,s
() > 0 such
that, for all (0, 1] and for all u H
s
q
(M),
(2.115) |S
k
()(u) S
k
(/2)(u)|

,s
() |u|
s

s
.
36 GIOVANNI FORNI

Since S
k
()(u) H
k
q
(M) for all (0, 1], by the interpolation inequality
proved in Lemma 2.10 it follows that, for any r [0, k] and for any > 0
there exists C
r,s
() > 0 such that, for all (0, 1] and for all u H
s
q
(M),
(2.116) [S
k
()(u) S
k
(/2)(u)[
r
C
r,s
() |u|
s

sr
.
It follows that, for every n N and for every (0, 1],
(2.117)
[S
k
(/2
n
)(u) S
k
(/2
n+1
)(u)[
r
C
r,s
() |u|
s
2
n(sr)

sr
.
If s > r and 0 < < s r, the sequence S
k
(/2
n
)(u)
nN
is Cauchy and
therefore convergent to the function u H
s
q
(M) H
r
q
(M) in the Hilbert
space H
r
q
(M). It follows that, for all (0, 1] and for all u H
s
q
(M),
(2.118) [S
k
()(u) u[
r
C
r,s
() |u|
s

sr
.
If s r and s r 0 < , we argue as follows. By estimate (2.116), for
each 1/2 we have
(2.119) [S
k
(2)(u) S
k
()(u)[
r
C
r,s
() |u|
s
(2)
sr
,
hence if 2
n
, for all 0 k < n,
(2.120)
[S
k
(2
k+1
)(u) S
k
(2
k
)(u)[
r
C
r,s
() |u|
s
2
(k+1)(sr)

sr
,
It follows that, there exists a constant C

r,s
() > 0 such that
(2.121) [S
k
(2
n
)(u) S
k
()(u)[
r
C

r,s
() |u|
s
2
n(sr)

sr
,
For every (0, 1], let n() be the maximumn Nsuch that 2
n
1. By
this denition it follows that 1/2 < 2
n()
1, hence by estimate (2.113)
there exists a constant C
k
> 0 such that, for all u H
s
q
(M),
[S
k
(2
n()
)(u)[
r
sup
1/21
|S
k
()(u)|
k
< C
k
|u|
s
.
It follows that there exists C

r,s
() > 0 such that, for all (0, 1] and for
all u H
s
q
(M),
(2.122) [S
k
()(u)[
r
C

r,s
() |u|
s

sr
,

By Lemma 2.11 and Theorem 2.24, we have the following comparison es-
timate for the (Friedrichs) weigthed Sobolev norms :
Corollary 2.25. For any 0 < r < s there exists constants C
r
> 0 and
C
r,s
> 0 such that, for all u H
s
q
(M), the following inequalities hold:
C
1
r
|u|
r
[u[
r
C
r,s
|u|
s
.
SOBOLEV REGULARITY OF SOLUTIONS 37
Finally, we derive a crucial interpolation estimate for the dual weighted
Sobolev norms:
Corollary 2.26. Let 0 s
1
< s
2
. For any s
1
r < s s
2
there exists a
constant C
r,s
> 0 such that for any distribution u H
s
1
q
(M) the following
interpolation inequality holds:
(2.123) [u[
s
C
r,s
[u[
s
2
r
s
2
s
1
s
1
[u[
rs
1
s
2
s
1
s
2
.
Proof. Let k N be any integer larger than s
2
> s
1
and let S
k
() :
L
2
q
(M) H
k
q
(M) be the family of smoothing operators constructed above.
By Theorem 2.24, since 0 r s
1
< s s
1
and any r s
2
< s s
2
0,
there exists a constant C
k
r,s
> 0 such that the following holds: for any
u H
s
1
q
(M) 0, any v H
s
q
(M) and for all (0, 1],
(2.124)
[u, v)[ [u[
s
1
[v S
k
()(v)[
s
1
+ [u[
s
2
[S
k
()(v)[
s
2
C
k
r,s

rs
1
[u[
s
1
+
rs
2
[u[
s
2
[v[
s
.
The interpolation inequality (2.123) then follows by taking
=
_
[u[
s
2
[u[
s
1
_ 1
s
2
s
1
(0, 1] .

3. THE COHOMOLOGICAL EQUATION


3.1. Distributional solutions. In this section we give a streamlined ver-
sion of the main argument of [For97] (Theorem 4.1) with the goal of estab-
lishing the sharpest bound on the loss of Sobolev regularity within the reach
of the methods of [For97]. We were initially motivated by a question of
Marmi, Moussa and Yoccoz who found for almost all orientable quadratic
differentials a loss of regularity of 1 + BV (they nd bounded solutions
for absolutely continuous data with rst derivative of bounded variation un-
der nitely many independent compatibility conditions and corresponding
results for higher smoothness) [MMY03], [MMY05] . The results of this
section, as those of [For97], hold for all orientable quadratic differentials.
There is a natural action of the circle group S
1
SO(2, R) on the space
Q(M) of holomorphic quadratic differentials on a Riemann surface M:
r

(q) := e
i
q , for all (r

, q) SO(2, R) Q(M) .
Let q

denote the quadratic differential r

(q) and let S

, T

denote the
frame (introduced in 2.1) associated to the quadratic differential q

for any
38 GIOVANNI FORNI

S
1
. We have the following formulas:
(3.1)
S

= cos
_

2
_
S
q
+ sin
_

2
_
T
q
=
e
i

2
2

+
q
+
e
i

2
2

q
;
T

= sin
_

2
_
S
q
+ cos
_

2
_
T
q
=
e
i

2
2i

+
q

e
i

2
2i

q
;
Denition 3.1. Let q be an orientable quadratic differential. A distribution
u

H
r
q
(M) will be called a (distributional) solution of the cohomological
equation S
q
u = f for a given function f

H
s
q
(M) if
u, S
q
v) = f, v) , for all v H
r+1
q
(M)

H
s
q
(M) .
Let

H
s
q
(M)

H
s
q
(M), H
s
q
(M) H
s
q
(M) (for any s R) be the subspaces
orthogonal to constant functions, that is
(3.2)

H
s
q
(M) := f

H
s
q
(M) [ f, 1)
s
= 0 ,
H
s
q
(M) := f H
s
q
(M) [ (f, 1)
s
= 0 .
The spaces

H
s
q
(M)

H
s
q
(M) and H
s
q
(M) H
s
q
(M) coincide with the
subspaces of functions of zero average for s 0, and with the subspaces of
distributions vanishing on constant functions for s < 0.
Theorem 3.2. Let r > 2 and p (0, 1) be such that and rp > 2. There
exists a bounded linear operator
U :

H
1
q
(M) L
p
_
S
1
,

H
r
q
(M)
_
such that the following holds. For any f

H
1
q
(M) there exists a full
measure subset F
r
(f) S
1
such that u := U(f)()

H
r
q
(M) is a distri-
butional solution of the cohomological equation S

u = f for all F
r
(f).
Proof. We claim that for any r > 2, any p (0, 1) such that pr > 2 and
any f

H
1
q
(M), there exists a measurable function A
q
:= A
q
(p, r, f)
L
p
(S
1
, R
+
) such that the following estimates hold. Let S
1
be such that
A
q
() < +. For all v H
r+1
q
(M) we have
(3.3) [f, v)[ A
q
() |S

v|
r
.
In addition, the following bound for the L
p
norm of the function A
q
holds.
There exists a constant B
q
(p) > 0 such that
(3.4) [A
q
[
p
B
q
(p) |f|
1
.
Assuming the claim, we prove the statement of the theorem. In fact, by the
estimate (3.3) the linear map given by
(3.5) S

v f, v) , for all v H
r+1
q
(M) ,
SOBOLEV REGULARITY OF SOLUTIONS 39
is well dened and extends by continuity to the closure of the range

R
r
()
of the linear operator S

in

H
r
q
(M). Let U(f)() be the extension uniquely
dened by the condition that U(f)() vanishes on the orthogonal comple-
ment of

R
r
() in

H
r
q
(M). By construction, for almost all S
1
the linear
functional u := U(f)()

H
r
q
(M) yields a distributional solution of the
cohomological equation S

u = f whose norm satises the bound


|U(f)()|
r
A
q
() .
By (3.4) the L
p
norm of the measurable function U(f) : S
1


H
r
q
(M)
satises the required estimate
[U(f)[
p
:=
__
S
1
|U(f)()|
p
r
d
_
1/p
B
q
(p) |f|
1
.
We turn now to the proof of the above claim. Let R

q
=
_
M

q
_

be the
(closed) ranges of the Cauchy-Riemann operators

q
: H
1
q
(M) L
2
q
(M)
(see Proposition 2.4). Following [For97], we introduce the linear operator
U
q
: R

q
R
+
q
dened as
(3.6) U
q
:=

q
(
+
q
)
1
.
By Proposition 2.4, (3), the operator U
q
is a partial isometry on L
2
q
(M),
hence by the standard theory of partial isometries on Hilbert spaces, it has
a family of unitary extensions U
J
: L
2
q
(M) L
2
q
(M) parametrized by
isometries J : M
+
q
M

q
(see formulas (3.10)-(3.12) in [For97]). By
denition the following identities hold on H
1
q
(M) (see formulas (3.13) in
[For97]) :
(3.7) S

=
e
i

2
2
_
U
J
+ e
i
_

+
q
=
e
i

2
2
_
U
1
J
+ e
i
_

q
.
The proof of estimate (3.3) is going to be based on properties of the re-
solvent of the operator U
J
. In fact, the proof of (3.3) is based on the re-
sults, summarized in [For97], Corollary 3.4, concerning the non-tangential
boundary behaviour of the resolvent of a unitary operator on a Hilbert
space, applied to the operators U
J
, U
1
J
on L
2
q
(M). The Fourier analysis
of [For97], 2, also plays a relevant role through Lemma 4.2 in [For97] and
the Weyls asymptotic formula (Theorem 2.3).
Following [For97], Prop. 4.6A, or [For02], Lemma 7.3, we prove that there
exists a constant C
q
> 0 such that the following holds. For any distribution
f H
1
q
(M) there exist (weak) solutions F

L
2
q
(M) of the equations

q
F

= f such that
(3.8) [F

[
0
C
q
|f|
1
.
40 GIOVANNI FORNI

In fact, the maps given by
(3.9)

q
v f, v) , for all v H
1
q
(M) ,
are bounded linear functionals on the (closed) ranges R

q
L
2
q
(M) (of the
Cauchy-Riemann operator

q
: H
1
q
(M) L
2
q
(M). In fact, the functionals
are well-dened since f vanishes on constant functions, that is, on the ker-
nel of the Cauchy-Riemann operators, and it is bounded since by Poincar
inequality (see [For97], Lemma 2.2 or [For02], Lemma 6.9) there exists a
constant C
q
> 0 such that, for any v H
1
q
(M) H
1
q
(M),
(3.10) [f, v)[ |f|
1
|v|
1
C
q
|f|
1
[

q
v[
0
.
Let

be the unique linear extension of the linear map (3.9) to L


2
q
(M)
which vanishes on the orthogonal complement of R

q
in L
2
q
(M). By (3.10),
the functionals

are bounded on L
2
q
(M) with norm
|

| C
q
|f|
1
.
By the Riesz representation theorem, there exist two (unique) functions
F

L
2
q
(M) such that
v, F

)
q
=

(v) , for all v L


2
q
(M) .
The functions F

are by construction (weak) solutions of the equations

q
F

= f satisfying the required bound (3.8).


The identities (3.7) immediately imply that
(3.11)

q
v, F

)
q
= 2e
i

2
R

J
(z)S

v, F

)
q
(z + e
i
)R

J
(z)

q
v, F

)
q
,
where R
+
J
(z) and R

J
(z) denote the resolvents of the unitary operators U
J
and U
1
J
respectively, which yield holomorphic families of bounded opera-
tors on the unit disk D C.
Let r > 2 and let p (0, 1) be such that pr > 2. Let E = e
k

kN
be the
orthonormal Fourier basis of the Hilbert space L
2
q
(M) described in 2. By
Corollary 3.4 in [For97] all holomorphic functions
(3.12) R

k
(z) := R

J
(z)e
k
, F

)
q
, k N,
belong to the Hardy space H
p
(D), for any 0 < p < 1. The corresponding
non-tangential maximal functions N

k
(over cones of arbitrary xed aper-
ture 0 < < 1) belong to the space L
p
(S
1
, d) and for all 0 < p < 1 there
exists a constant A
,p
> 0 such that the following inequalities hold:
(3.13) [N

k
[
p
A
,p
[e
k
[
0
[F

[
0
= A
,p
[F

[
0
A
,p
C
q
|f|
1
.
SOBOLEV REGULARITY OF SOLUTIONS 41
Let
k

kN
be the sequence of the eigenvalues of the Dirichlet form Q
introduced in 2. Let w

H
r
q
(M). We have
(3.14) R

J
(z)w, F

)
q
=

k=0
w, e
k
)
q
R

k
(z) ,
hence, by the Cauchy-Schwarz inequality,
(3.15) [R

J
(z)w, F

)
q
[
_

k=0
[R

k
(z)[
2
(1 +
k
)
r
_
1/2
|w|
r
,
Let N

() be the functions dened as


(3.16) N

() :=
_

k=0
[N

k
()[
2
(1 +
k
)
r
_
1/2
.
Let N

(w) be the non-tangential maximal function for the holomorphic


function R

J
(z)w, F

)
q
. By formulas (3.15) and (3.16), for all S
1
and
all functions w

H
r
q
(M), we have
(3.17) N

(w)() N

() |w|
r
.
The functions N

L
p
(S
1
, d) for any 0 < p < 1. In fact, by formula
(3.13) and (following a suggestion of Stephen Semmes) by the triangular
inequality for L
p
spaces with 0 < p < 1, we have
(3.18) [N

[
p
p
(A
,p
C
q
)
p
_

k=0
1
(1 +
k
)
pr/2
_
|f|
p
1
< + .
The series in formula (3.18) is convergent by the Weyl asymptotics (Theo-
rem 2.3) since pr/2 > 1.
By taking the non-tangential limit as z e
i
in the identity (3.11),
formula (3.17) implies that, for all S
1
such that N

( ) < +,
[

v, F

)
q
[ N

( ) |S

v|
r
,
hence the required estimates (3.3) and (3.4) are proved wih the choice of
the function A
q
() := N
+
( ) or A
q
() := N

( + ) for all S
1
.
Since the claim is proved the result follows.
Theorem 3.3. Let r > 2. For almost all S
1
(with respect to the
Lebesgue measure), there exists a constant C
r
() > 0 such that, for all
f

H
r1
q
(M) such that
_
M
f
q
= 0, the cohomological equation S

u = f
has a distributional solution u

H
r
q
(M) satisfying the following estimate:
|u|
r
C
r
() |f|
r1
.
42 GIOVANNI FORNI

Proof. Let E = e
k

kN
be the orthonormal Fourier basis of the Hilbert
space L
2
q
(M) described in 2. Let r > 2 and p (0, 1) be such that
pr > 2. By Theorem 3.2, for any k N 0 there exists a function
with distributional values u
k
:= U(e
k
) L
p
_
S
1
,

H
r
q
(M)
_
such that the
following holds. There exists a constant C
q
:= C
q
(p, r) > 0 such that
(3.19)
__
S
1
|u
k
()|
p
r
d
_
1/p
C
q
|e
k
|
1
= C
q
(1 +
k
)
1/2
.
In addition, for any k N 0, there exists a full measure set F
k
S
1
such that, for all F
k
, the distribution u := u
k
()

H
r
q
(M) is a
(distributional) solution of the cohomological equation S

u = e
k
.
Any function f

H
r1
q
(M) such that
_
M
f
q
= 0 has a Fourier decompo-
sition in L
2
q
(M):
f =

kN\{0}
f, e
k
)
q
e
k
.
A (formal) solution of the cohomological equation S

u = f is therefore
given by the series
(3.20) u

:=

kN\{0}
f, e
k
)
q
u
k
() .
By the triangular inequality in

H
r
q
(M) and by Hlder inequality, we have
|u

|
r

_
_

kN\{0}
|u
k
()|
2
r
(1 +
k
)
r1
_
_
1/2
|f|
r1
,
hence by the triangular inequality for L
p
spaces (with 0 < p < 1) and by
the estimate (3.19),
(3.21)
_
S
1
|u

|
p
r
d C
p
q
_
_

kN\{0}
1
(1 +
k
)
pr/2
_
_
|f|
p
r1
.
Since pr/2 > 1 the series in (3.21) is convergent, hence by Chebyshev
inequality for the space L
p
(S
1
), there exists a full measure set B S
1
such that, for all B, formula (3.20) yields a well-dened distribution
u



H
r
q
(M) and there exists a constant C
q
() > 0 such that
(3.22) |u

|
r
C
q
() |f|
r1
.
The set F = F
k
Bhas full measure and for all F, for all k N0,
the distribution u
k
()

H
r
q
(M) is a solution of the equation S

u = e
k
.
It follows that u



H
r
q
(M) is a solution of the cohomological equation
S

u = f which satises the required bound (3.22).


SOBOLEV REGULARITY OF SOLUTIONS 43
We nally derive a result on distributional solutions of the cohomological
equation for distributional data of arbitrary regularity:
Corollary 3.4. For any s R there exists r > 0 such that the following
holds. For almost all S
1
(with respect to the Lebesgue measure), there
exists a constant C
r,s
() > 0 such that, for all F

H
s
q
(M) orthogonal to
constant functions, the cohomological equation S

U = F has a distribu-
tional solution U

H
r
q
(M) satisfying the following estimate:
|U|
r
C
r,s
() |F|
s
.
Proof. Since F

H
s
q
(M) is orthogonal to constant functions, for every
k N there exists f
k


H
s+2k
q
(M), orthogonal to constant functions, such
that (I
F
q
)
k
f
k
= F. In fact, the family of Friedrichs Sobolev spaces

H
s
q
(M)[s R is dened in terms of the Friedrichs extension
F
q
of the
Laplace operator
q
of the at metric determined by the quadratic differ-
ential. Let n N be the minimum integer such that := s + 2n + 1 > 2.
By Theorem 3.3, for almost all S
1
(with respect to the Lebesgue mea-
sure) the cohomological equation S

u = f
n
has a distributional solution
u

q
(M) satisfying the following estimate:
(3.23) |u

() |f
n
|
1
.
Let U

:= (I
F
q
)
n
u



H
2n
q
(M). It follows immediately from the
estimate (3.23) and from the denitions that
|U

|
2n
= |u
n
|

() |f
n
|
1
= C

() |F|
s
.
Finally U



H
2n
q
(M) is a distributional solution of the cohomological
equation S

U = F, for almost all S


1
. In fact, for any v H
+2n+1
q
(M),
the function S

v H
+2n
q
(M), hence (I
q
)
n
S

v = S

(I
q
)
n
v and,
since the distribution u

q
(M) is a solution of the cohomological
equation S

u = f
n
, for almost all S
1
, and (I
F
q
)
n
v H
+1
q
(M),
U

, S

v) = (I
F
q
)
n
u

, S

v) = u

, S

(I
q
)
n
v)
= f
n
, (I
F
q
)
n
v) = (I
F
q
)
n
f
n
, v) = F, v) ,
as required by the denition of distributional solution of the cohomological
equation (Denition 3.1).
3.2. Invariant distributions and basic currents. Invariant distributions
yield obstructions to the existence of smooth solutions of the cohomolog-
ical equation. We derive below from Theorem 3.2 a sharp version of the
main results of [For02], 6, about the Sobolev regularity of invariant distri-
butions. We then recall the structure theorem proved in that paper on the
space of invariant distributions (see [For02], Th. 7.7).
44 GIOVANNI FORNI

Invariant distributions for the horizontal [respectively vertical] vector eld
of an orientable quadratic differential q are closely related to basic currents
(of dimension and degree equal to 1) for the horizontal [vertical] foliation F
q
[ F
q
]. The notion of a basic current for a measured foliation on a Riemann
surface has been studied in detail in [For02], 6, in the context of weighted
Sobolev spaces with integer exponent. We outline below some of the basic
constructions and results on basic currents and invariant distributions which
carry over without modications to the more general context of fractional
weighted Sobolev spaces. Finally, we derive from Theorem 3.2 a result on
the Sobolev regularity of basic currents (or invariant distributions) which
improves upon a similar result proved in [For02] (see Theorem 7.1 (i)).
Let M be a nite subset. The space D(M) will denote the standard
space of de Rham currents on the open manifold M , that is the dual of
the Frchet space
c
(M ) of differential forms with compact support in
M . A homogenous current of dimension d N (and degree 2 d)
on M is a continuous linear functional on the subspace
d
c
(M ) of
diferential forms of degree d. The subspace of homogeneous currents of
dimension d on M will be denote by D
d
(M ).
Let q be an orientable quadratic differential on a Riemann surface M. Let

q
be the (nite) set of its zeroes. In [For02] we have introduced the fol-
lowing space
q
(M) of smooth test forms on M.
Denition 3.5. For any p M of (even) order k = 2m N (m = 0 if p ,

q
), let z : U
p
C be a canonical complex coordinate on a neighbourhood
U
p
of p M, that is a complex coordinate such that z(p) = 0 and q = z
k
dz
2
on U
p
. Let
p
: U
p
C be the (local) covering map dened by

p
(z) :=
z
m+1
m+ 1
, z C.
The space
q
(M) is dened as the space of smooth forms on M such
that the following holds: for all p M, there exists a smooth form
p
on
a neighbourhood of 0 C such that =

p
(
p
) on U

p
U
p
. The space

q
(M) is the direct sum of the subspaces
d
q
(M) of homogeneous forms of
degree d 0, 1, 2. The spaces
d
q
(M), for any d 0, 1, 2, and
q
(M)
can be endowed with a natural Frchet topology modeled on the smooth
topology in every coordinate neighbourhood.
Lemma 3.6. For any orientable quadratic differential q Q(M), the space
of functions
0
q
(M) is dense in the space H

q
(M) endowed with the inverse
limit Frchet topology induced by the family of weighted Sobolev norms.
Proof. By denition, the MacLaurin series of any f
q
(M) with respect
to a canonical complex coordinate z for q at every p
q
(of order 2m
p
)
SOBOLEV REGULARITY OF SOLUTIONS 45
has the following form:
(3.24) f(z) =

h,kN
f
hk
z
h(mp+1)
z
k(mp+1)
.
By Lemmas 2.12 and 2.13, f H

q
(M). Thus
q
(M) H

q
(M).
Let F H

q
(M). By Lemma 2.11 the function F C

(M) and by
Lemmas 2.12 and 2.13 its MacLaurin series has the form (3.24) at every
p
q
(of order 2m
p
). By Borels thereorem and by a partition of unity
argument, there exists a function f
q
(M) such that F f C

(M)
vanishes at innite order at
q
. Let U

be the open neighbourhood of


q
which is the union of a nite number of disjoint geodesic disks D

(p) of
radius (0,
0
), each centered at a point p
q
. Let

: M [0, 1] be
a smooth function such that (a)

0
(M
q
), (b)

1 on M U

and (c) for each (i, j) N N there exists a constant C


ij
> 0 such that,
for all (0,
0
),
max
xM
[S
i
T
j

(x)[
C
ij

i+j
.
If can be proved that, since F f vanishes at innite order at
q
,
f +

(F f) F in H

q
(M) , as 0
+
,
which implies, since by construction f +

(F f)
0
q
(M), that F
belongs to the closure of
0
q
(M) in H

q
(M).
Denition 3.7. The space S
q
(M) D(M
q
) of q-tempered currents
(introduced in [For02], 6.1) is the dual space of the Frchet space
q
(M).
A homogeneous q-tempered current of dimension d (and degree 2 d) is a
continuous functional on the subspace
d
q
(M)
q
(M) of homogeneous
forms of degree d 0, 1, 2. The space of homogeneous currents of di-
mension d (and degree 2 d) will be denoted by S
d
q
(M).
For any quadratic differential q on M, there is a natural operator

, which
maps the space D
0
(M
q
) of currents of dimension 0 and degree 2 on
the non-compact manifold M
q
(which is naturally identied with the
space of distributions on M
q
) bijectively onto the space D
2
(M
q
) of
currents of dimension 2 and degree 0 on M
q
. The operator

: D
0
(M
q
) D
2
(M
q
)
is dened as follows. Let
q
be the smooth area form associated with the
(orientable) quadratic differential q on M. It is a standard fact in the theory
of currents that any distribution U on the 2-dimensional surface M
q
can
be written as U = U

q
for a unique current U

of dimension 2 and degree


0. Since
q

2
q
(M), the map

extends to a bijective map

: S
0
q
(M) S
2
q
(M) .
46 GIOVANNI FORNI

Denition 3.8. A distribution D D
2
(M
q
) is horizontally [vertically]
quasi-invariant if SD = 0 [TD = 0] in D
2
(M
q
). A distribution
D S
2
q
(M) is horizontally [vertically] invariant if SD = 0 [TD = 0] in
S
2
q
(M). The space of horizontally [vertically] quasi-invariant distributions
will be denoted by I
q
(M
q
) [I
q
(M
q
)] and the subspace of horizontally
[vertically] invariant distributions will be denoted by I
q
(M) [I
q
(M)].
Denition 3.9. For any s R
+
, let
(3.25)
I
s
q
(M
q
) := I
q
(M
q
) H
s
q
(M) ;
I
s
q
(M) := I
q
(M) H
s
q
(M) .
The subspaces I
s
q
(M) H
s
q
(M) of horizontally [vertically] invariant
distributions can also be dened as follows:
(3.26)
I
s
q
(M) := D H
s
q
(M) [ SD = 0 in H
s1
q
(M) ;
[I
s
q
(M) := D H
s
q
(M) [ TD = 0 in H
s1
q
(M)] .
The subspaces of horizontally [vertically] invariant distributions which can
be extended to bounded functionals on Friedrichs weighted Sobolev spaces
will be denoted by
(3.27)

I
s
q
(M
q
) := I
q
(M
q
)

H
s
q
(M) ;

I
s
q
(M) := I
q
(M)

H
s
q
(M) .
Let V
q
(M) be the space of vector elds X on M
q
such that the contrac-
tion
X
and the Lie derivative L
X

q
(M) for all
q
(M).
Denition 3.10. A current C D
1
(M
q
) is horizontally [vertically]
quasi-basic, that is basic for F
q
[F
q
] in the standard sense on M
q
, if
the identities
(3.28)
X
C = L
X
C = 0
hold in D(M
q
) for all smooth vector elds X tangent to F
q
[F
q
] with
compact support on M
q
. A q-tempered current C S
1
q
(M) is horizon-
tally [vertically] basic if the identities (3.28) holds in S
q
(M) for all vector
elds X V
q
(M), tangent to F
q
[F
q
] on M
q
. The vector spaces
of horizontally [vertically] quasi-basic (real) currents will be denoted by
B
q
(M
q
) [B
q
(M
q
)] and the subspace of horizontally [vertically]
basic (real) currents will be denoted by B
q
(M) [B
q
(M)].
Denition 3.11. For any s R, the Friedrichs weighted Sobolev space of
1-currents

W
s
q
(M) S
q
(M) and the weighted Sobolev space of 1-currents
W
s
q
(M) S
q
(M) are dened as follows:
(3.29)

W
s
q
(M) := S
q
(M) [ (
S
,
T
)

H
s
q
(M)

H
s
q
(M) ;
W
s
q
(M) := S
q
(M) [ (
S
,
T
) H
s
q
(M) H
s
q
(M) .
SOBOLEV REGULARITY OF SOLUTIONS 47
Denition 3.12. For any 1-current C D(M
q
), the weighted Sobolev
order O
W
q
(C) and the Friedrichs weighted Sobolev order

O
W
q
(C) are the
real numbers dened as follows:
(3.30)
O
W
q
(C) :=infs R[ D W
s
q
(M) ;

O
W
q
(C) :=infs R[ C

W
s
q
(M) .
Denition 3.13. For any s R, let
(3.31)
B
s
q
(M
q
) := B
q
(M
q
) W
s
q
(M) ;
B
s
q
(M) := B
q
(M) W
s
q
(M) .
The subspaces B
s
q
(M) W
s
q
(M) of horizontally [vertically] basic cur-
rents can also be dened as follows:
(3.32)

S
C = 0 in H
s
q
(M) and L
S
C = 0 in W
s1
q
(M) ;
[
T
C = 0 in H
s
q
(M) and L
T
C = 0 in W
s1
q
(M)] .
The subspaces of basic currents which can be extended to bounded func-
tionals on Friedrichs weighted Sobolev spaces will be denoted by
(3.33)

B
s
q
(M
q
) := B
q
(M
q
)

W
s
q
(M) ;

B
s
q
(M) := B
q
(M)

W
s
q
(M) .
According to Lemma 6.5 of [For02], the notions of invariant distributions
and basic currents are related (see also Lemma 6.6 in [For02]):
Lemma 3.14. A current C B
s
q
(M
q
) [C B
s
q
(M)] if and only if
the distribution C (q
1/2
) I
s
q
(M
q
) [C (q
1/2
) I
s
q
(M)]. A
current C B
s
q
(M
q
) [C B
s
q
(M)] if and only if the distribution
C (q
1/2
) I
s
q
(M
q
) [C (q
1/2
) I
s
q
(M)]. In addition, the map
(3.34)
D
q
: C C (q
1/2
) ;
[D
q
: C C (q
1/2
)] ;
is a bijection from the space B
s
q
(M
q
) [B
s
q
(M
q
)] onto the space
I
s
q
(M
q
) [I
s
q
(M
q
)], which maps the subspace B
s
q
(M) [ B
s
q
(M)]
onto the subspace I
s
q
(M) [I
s
q
(M)]. The map (3.34) also maps the space

B
s
q
(M) [

B
s
q
(M)] onto

I
s
q
(M) [

I
s
q
(M)].
3.3. Basic cohomology. Let Z(M ) D
1
(M ) denote the subspace
of all (real) closed currents, that is, the space of all (real) de Rham currents
C D
1
(M ) such that the exterior derivative dC = 0 in D(M ). Let
Z
q
(M) S
1
q
(M) be the subspace of all (real) closed q-tempered currents,
that is, the space of all q-tempered (real) currents C such that dC = 0 in
48 GIOVANNI FORNI

S
q
(M). It was proved in [For02], Lemma 6.2, that the natural de Rham
cohomology map
(3.35) j
q
: Z(M
q
) H
1
(M
q
, R)
has the property that the subspace of closed q-tempered currents is mapped
onto the absolute real cohomology of the surface, that is,
(3.36) j
q
: Z
q
(M) H
1
(M, R) H
1
(M
q
, R) .
It was also proved in [For02], Lemma 6.2, that quasi-basic and basic cur-
rents are closed, in the sense that the following inclusions hold:
(3.37)
B
q
(M
q
) Z(M
q
) ,
B
q
(M) Z
q
(M) .
The images of the restrictions of the natural cohomology map to the various
spaces of basic currents are called the horizontal [vertical] basic cohomolo-
gies , namely the spaces
(3.38)
H
1
q
(M
q
, R) := j
q
(B
q
(M
q
)) H
1
(M
q
, R) ;
H
1,s
q
(M
q
, R) := j
q
_
B
s
q
(M
q
)
_
H
1
q
(M
q
, R) ;
H
1
q
(M, R) := j
q
(B
q
(M)) H
1
(M, R) ;
H
1,s
q
(M, R) := j
q
_
B
s
q
(M)
_
H
1
q
(M, R) .
Following [For02], Theorem 7.1, we give below a description of the hori-
zontal [vertical] basic cohomologies for the orientable quadratic differential
q

, for any orientable holomorphic quadratic differential q on M and for al-


most all S
1
. The result we obtain below is stronger than Theorem 7.1
of [For02] since it requires weaker Sobolev regularity assumptions.
(Absolute) real cohomology classes on M can be represented in terms of
meromorphic (or anti-meromorphic) functions in L
2
q
(M) (see [For02], 2).
In fact, by the Hodge theory on Riemann surfaces [FK92], III.2, all real
cohomology classes can be represented as the real (or imaginary) part of
a holomorphic (or anti-holomorphic) differential on M. In turn, any ori-
entable holomorphic quadratic differential induces an isomorphismbetween
the space Hol
+
(M) [Hol

(M)] of holomorphic [anti-holomorphic] differ-


entials and the space of square-integrable meromorphic [anti-meromorphic]
functions. Let M
+
q
[M

q
] be the space of meromorphic [anti-meromorphic]
functions on M which belong to the Hilbert space L
2
q
(M) (see Proposi-
tion 2.4). Such spaces can be characterized as the spaces of all meromor-
phic [anti-meromorphic] functions with poles at
q
= q = 0 of orders
bounded in terms of the multiplicity of the points p
q
as zeroes of the
quadratic differential q. In fact, if p
q
is a zero of q of order 2m, that p
is a pole of order at most m for any m

q
.
SOBOLEV REGULARITY OF SOLUTIONS 49
Let q
1/2
be a holomorphic square root of q on M. Holomorphic [anti-
holomorphic] differentials h
+
[h

] on M can be written in terms of mero-


morphic [anti-meromorphic] functions in L
2
q
(M) as follows:
(3.39)
h
+
:= m
+
q
1/2
, m
+
M
+
q
;
h

:= m

q
1/2
, m

q
.
The following representations of real cohomology classes therefore hold:
(3.40)
c H
1
(M, R) c = [(m
+
q
1/2
)] , m
+
M
+
q
;
c H
1
(M, R) c = [(m

q
1/2
)] , m

q
.
The maps c

q
: M

q
H
1
(M, R) given by the representations (3.40) are
bijective and it is in fact isometric if the spaces M

q
are endowed with the
euclidean structure induced by L
2
q
(M) and H
1
(M, R) with the Hodge prod-
uct relative to the complex structure of the Riemann surface M. In fact, the
Hodge norm |c|
2
H
of a cohomology class c H
1
(M, R) is dened as fol-
lows:
(3.41) |c|
2
H
:=

2
_
M
h

if c = [(h

)] , h

Hol

(M) .
We remark that the Hodge norm is dened in terms of the complex structure
of the Riemann surface M (carrying a holomorphic quadratic differential
q Q(M)) but does not depend on the quadratic differential. If q Q(M)
is any orientable quadratic differential on M, by the representation (3.40),
we can also write:
(3.42)
|c
+
q
(m
+
)|
2
H
:=
_
M
[m
+
[
2

q
, for all m
+
M
+
q
;
|c

q
(m

)|
2
H
:=
_
M
[m

[
2

q
, for all m

q
.
The representation (3.39)-(3.40) can be extended to the punctured cohomol-
ogy H
1
(M
q
, R) as follows. For any nite set M, let Hol
+
(M )
[Hol

(M )] be the space of meromorphic [anti-meromorphic] differen-


tials with at most simple poles at . By Riemann surface theory, any real
cohomology class c H
1
(M , R) can be represented as the real (or
imaginary) part of a differential h
+
Hol
+
(M ) or h

Hol

(M ).
Let M be a nite set and let M
+
() [M
+
()] be the space of all
meromorphic [anti-meromorphic] functions which are holomorphic [anti-
holomorphic] on M . The spaces M

() can be identied with a sub-


space of the distributional space D
2
(M ). In fact, if q is any orientable
holomorphic quadratic differential on M, the spaces M

(
q
) identify with
subspaces of the space S
2
q
(M) of q-tempered distributions. The distribution
determined by a function M
+
(
q
) or M

(
q
) is dened by integration
50 GIOVANNI FORNI

(in the standard way) as a linear functional on C

0
(M
q
), which can be
extended to the space
0
q
(M) as follows:
(v) := PV
_
M
v
q
, for all v
0
q
(M) .
The Sobolev regularity of a distributions M

(
q
) depend on the order
of its poles. In fact, by Theorem 2.22 we have the following:
Lemma 3.15. Let M
+
(
q
) [ M
+
(
q
) ] be a meromorphic [anti-
meromorphic] function with poles at
q
. For any s R, the associated
distribution H
s
q
(M) if at every p
q
of order 2m
p
the function
has a pole of order < (m
p
+ 1)(s + 1).
We introduce the following notation: for all s > 0,
(3.43) M

s
(
q
) := M

(
q
) H
s
q
(M) .
There exist natural maps

q
: Hol

(M
q
) M

(
q
) dened as follows:
for all h

Hol

(M
q
),

+
q
(h
+
) = h
+
/q
1/2
[

q
(h

) = h

/ q
1/2
] .
By Lemma 3.15 the range of the maps

q
is contained in the weighted
Sobolev space H
s
q
(M) for all s > 0, hence there are well-dened maps
(3.44)

q
: Hol

(M
q
) M

s
(
q
) for all s > 0 .
The maps (3.44) are clearly injective and by Corollary 2.23 there exists
s
q
> 0 such that, for any s (0, s
q
), they are also surjective. Let
(3.45) M

(
q
) =

s>0
M

s
(
q
) = M

s
(
q
) , for any s (0, s
q
) .
The representation (3.40) of the absolute real cohomology generalizes to
the punctured real cohomology as follows.
(3.46)
c H
1
(M
q
, R) c = [(m
+
q
1/2
)] , m
+
M
+

(
q
) ;
c H
1
(M
q
, R) c = [(m

q
1/2
)] , m

(
q
) .
The following lemma, proved in [For02], Lemma 7.6, for weighted Sobolev
spaces with integer exponent, holds:
Lemma 3.16. Let s R
+
. Let C W
s
q
(M) be any real current of di-
mension (and degree) equal to 1, closed in the space D(M
q
) of currents
on M
q
. There exists a distribution U H
s+1
q
(M) and a meromorphic
differential h
+
Hol
+
(M
q
) such that
(3.47) dU

= (h
+
) C in W
s
q
(M) .
SOBOLEV REGULARITY OF SOLUTIONS 51
If C is closed in the space S
q
(M) of q-tempered currents there exists a dis-
tribution U H
s+1
q
(M) and a holomorphic differential h
+
Hol
+
(M)
such that the identity (3.47) holds.
The argument given in [For02], Lemma 7.6, in the case of integer order
s N extends the general case of order s R
+
. In fact, it follows from the
distributional identity (3.47) in S
q
(M) that the current U

H
s+1
q
(M) if
and only if the current C W
s
q
(M), for any s R
+
. Hence, Lemma 3.16
follows immediately from [For02], Lemma 7.6.
The construction of basic currents (or, equivalently, of invariant distribu-
tions) is based on the following method.
Lemma 3.17. Let q be an orientable holomorphic quadratic differential on
a Riemann surface M. Let m
+
M
+
s
(
q
) be a meromorphic function with
poles at
q
M. A distribution U H
s+1
q
(M) is a (distributional)
solution in D(M
q
) of the cohomological equation
(3.48) SU = (m
+
) [TU = (m
+
)] in D(M
q
) ,
if and only if the current C W
s
q
(M) uniquely determined by the identity
(3.49) dU

= (m
+
q
1/2
) + C
is horizontally [vertically] quasi-basic. If (m
+
q
1/2
) Hol
+
(M), the
distribution U H
s+1
q
(M) is a solution of the cohomological equation
(3.48) in the space S
q
(M) of q-tempered currents if and only if the cur-
rent C W
s
q
(M) uniquely determined by formula (3.49) is horizontally
[vertically] basic.
Proof. If formula (3.49) holds in D(M
q
), then C is closed in D(M

q
). If the differential (m
+
q
1/2
) is holomorphic and formula (3.49) holds
in S
q
(M), then C is closed in S
q
(M). The standard formula for the Lie
derivative of a current,
(3.50) L
X
C =
X
dC + d
X
C = 0 ,
holds in D(M
q
) for any vector eld X with compact support contained
in M
q
and it holds in S
q
(M) for any vector eld X V
q
(M) It follows
that a current C D(M
q
) is horizontally [vertically] quasi-basic if
and only if it is closed and
S
C = 0 [
T
C = 0] in D(M
q
) and it
is horizontally [vertically] basic if and only if C S
q
(M) is closed and

S
C = 0 [
T
C = 0] in S
q
(M). The distribution U H
s
q
(M) in formula
(3.49) is a solution of the cohomological equation (3.48) in D(M
q
)
or S
q
(M) if and only if
S
C = 0 [
T
C = 0] in D(M
q
) or S
q
(M)
respectively. As a consequence, the lemma is proved.
52 GIOVANNI FORNI

Let q be an orientable holomorphic quadratic differential on a Riemann sur-
face M (of genus g 1). Let
q
(M
q
, R) H
1
(M
q
, R) be the
codimension 1 subspaces dened as follows:
(3.51)

1
q
(M
q
, R) := c H
1
(M
q
, R) [ c [(q
1/2
)] = 0 ;

1
q
(M
q
, R) := c H
1
(M
q
, R) [ c [(q
1/2
)] = 0 .
Since the absolute cohomology can be regarded as a subspace of the punc-
tured cohomology it also is possible to dene the subspaces
(3.52)

1
q
(M, R) :=
1
q
(M
q
, R) H
1
(M, R) ;

1
q
(M, R) :=
1
q
(M
q
, R) H
1
(M, R) .
Theorem 3.18. For any s > 3 there exists a full measure set F
s
S
1
such
that the following holds. For any F
s
, the following inclusions hold
(3.53)

1
q

(M
q
, R) H
1,s
q

(M
q
, R) ,

1
q

(M, R) H
1,s
q

(M, R) .
Proof. Let m
+
M
+
1
(
q
) be any meromorphic function such that the
induced distribution PV(m
+
) H
1
q
(M). A computation shows that, for
all S
1
,
(3.54) PV
_
M
(m
+
)
q
= 0 [(m
+
q
1/2

)]
1
q

(M
q
, R) .
Under the zero-average condition (3.54), by Theorem 3.2 for any r > 2
there exists a full measure set F
r
(m
+
) S
1
such that, for all S
1
,
the cohomological equation S

U = (m
+
) has a distributional solution
U

(m
+
)

H
r
q
(M). Let U

(m
+
) be the current of dimension 2 corre-
sponding to the distribution U

(m
+
). Let C

(m
+
) W
r1
q
(M) be the
1-dimensional current determined by the identity
(3.55) dU

(m
+
) = (m
+
q
1/2

) C

(m
+
) .
By Lemma 3.17, we have thus proved that, for all meromorphic functions
m
+
M
+
1
(
q
) and for all F
r
(m
+
), there exists a quasi-basic current
C

(m
+
) B
r+1
q

(M
q
) such that
[C

(m
+
)] = [(m
+
q
1/2

)]
1
q

(M
q
, R) ;
in addition, whenever m
+
M
+
q
, the current C

(m
+
) B
r+1
q

(M) is basic
and has a cohomology class
[C

(m
+
)] = [(m
+
q
1/2

)]
1
q

(M, R) .
Let N be the cardinality of the set
q
M and let
m
+
1
, . . . , m
+
2g1
, . . . , m
+
2g+1

SOBOLEV REGULARITY OF SOLUTIONS 53


be a basis (over R) of the real subspace of M
+

(
q
) dened by the zero-
average condition (3.54). For any s > 3, let
F
+
s
:=
2g+1

i=1
F
s1
(m
+
i
) .
Clearly he set F
+
s
has full Lebesgue measure. We claim that for all F
+
s
the following inclusions hold:
(3.56)

1
q

(M
q
, R) H
1,s
q

(M
q
, R) ,

1
q

(M, R) H
1,s
q

(M R) .
The claim is proved as follows. For any c H
1
(M
q
, R) there exists a
unique meromorphic function m
+
M
+

(
q
) such that c = [(m
+
q
1/2

)].
The function m
+
M
+
q
for any c H
1
(M, R). If c
1
q

(M, R), the


distribution (m
+
) vanishes on constant functions as in (3.54), hence for
all F
+
s
, there exists a solution U H
s+1
q
(M) of the cohomological
equation S

U = (m
+
). The current C B
s
q

(M
q
) such that [C] = c

1
q

(M
q
, R) is then given by the identity (3.55). By the above discussion
the current C B
s
q

(M) for all c


1
q

(M, R).
By a similar argument it is possible to construct a full measure set F

s
such
that, for all F

s
, the following inclusions hold:
(3.57)

1
q

(M
q
, R) H
1,s
q

(M
q
, R) ,

1
q

(M, R) H
1,s
q

(M, R) .
Thus the set F
s
:= F
+
s
F

s
has the required properties since it has full
measure and the inclusions (3.53) hold.
By Lemma 3.14 and Theorem 3.18 the following holds:
Corollary 3.19. For any s > 3 there exists a full measure set F
s
S
1
such that, for all F
s
, the spaces I
s
q

(M) H
s
q
(M) of horizontally or
vertically quasi-invariant distributions have dimension at least 2g + 1
and the spaces I
s
q

(M) H
s
q
(M) of horizontally or vertically invariant
distributions have dimension at least 2g 1.
Corollary 3.20. For any s > 3 and for almost all S
1
,
(3.58) H
1,s
q

(M, R) =
1
q

(M, R) .
For any s > 4 and for almost all S
1
,
(3.59) H
1,s
q

(M
q
, R) = H
1
(M
q
, R) .
54 GIOVANNI FORNI

Proof. The inclusions H
1,s
q
(M, R)
1
q
(M, R) hold for any orientable
quadratic differential q on M and for any s > 0. In fact,
(3.60)
[C (q
1/2
)](1) =
S
C(
q
) = 0 , if C B
q
(M) ;
[C (q
1/2
)](1) =
T
C(
q
) = 0 , if C B
q
(M) .
Thus, the identity (3.58) follows immediately from Theorem 3.18.
By Theorem 3.18, in order to prove the identity (3.59) it is enough to prove
that for almost all S
1
the cohomology class [(q

)] B
q

(M
q
)
and the cohomology class [(q

)] B
q

(M
q
). By Lemma 3.17 the
argument is therefore reduced to the construction, for any s > 3 and for al-
most all S
1
, of a solution U H
s
q
(M) of the cohomological equation
S

U = 1 [T

U = 1] in D(M
q
). Such a construction can be carried out
as follows. Let
p
be the Dirac mass at any point p
q
. The distribution
F = 1
p
H
s
(M) H
s
q
(M) for any s > 1. We claim that for
any s > 3 and for almost all S
1
there exists a distributional solution
U H
s
q
(M) of the cohomological equation S

U = F [T

U = F] in
H
s1
q
(M). It follows that U is a solution of the cohomological equation
S

U = 1 [TU = 1] in D(M
q
). The above claim is proved as follows.
By [For97], Prop. 4.6, or [For02], Lemma 7.3, since F H
2
q
(M) van-
ishes on constant functions, there exists a distribution f H
1
q
(M) such
that
+
q
f = F (as well as a distribution f

H
1
q
(M) such that

q
f

= F).
By Theorem 3.2, for almost all S
1
and for all s > 2, there exists a so-
lution u H
s
q
(M) of the equation S

u = f [T

u = f ]. The distribution
U :=
+
q
u H
s
q
(M) for any s > 3 and solves the cohomological equation
S

U = F [T

U = F] in H
s1
q
(M).
The structure of the space of basic currents with vanishing cohomology
class with respect to the ltration induced by weighted Sobolev spaces with
integer exponent was described in [For02], 7. We extend below the results
of [For02] to fractional weighted Sobolev spaces.
Let
q
: B
q
(M
q
) B
q
(M
q
) be the linear maps dened as
follows (see [For02], formula (7.18

)):
(3.61)

q
(C) := d
_
C (q
1/2
)
_

, for C B
q
(M
q
) ;

q
(C) := d
_
C (q
1/2
)
_

, for C B
q
(M
q
) .
It can be proved by Lemma 3.14 and by the denition of the weighted
Sobolev spaces H
s
q
(M) and W
s
q
(M) that the above formulas (3.61) dene,
for all s R
+
, bounded linear maps
(3.62)

s
q
: B
s1
q
(M
q
) B
s
q
(M
q
) ;

s
q
: B
s1
q
(M) B
s
q
(M) .
SOBOLEV REGULARITY OF SOLUTIONS 55
We remark that a similar statement is false in general for the Friedrichs
Sobolev spaces of currents. The following result extends Theorem 7.7 of
[For02] to fractional weighted Sobolev spaces.
Theorem 3.21. For all s R
+
there exist exact sequences
(3.63)
0 R B
s1
q
(M
q
)

s
q

B
s
q
(M
q
)
jq

H
1
(M
q
, R) ;
0 R B
s1
q
(M)

s
q

B
s
q
(M)
jq

H
1
(M, R) .
Proof. The maps i
q
: R B
s1
q
(M) B
s1
q
(M
q
) dened as
i
q
() =
S
and i
q
() =
T
, for all R,
are clearly injective and the kernels Ker(
s
q
) = i
q
(R), for all s R
+
. In
fact, if a current C Ker(
s
+q
), then the distribution
_
C (q
1/2
)
_

R,
hence C (q
1/2
) R
q
and C R (q
1/2
) +R (q
1/2
). It follows
that C R (q
1/2
) since C B
q
(M
q
). Similarly, if the current
C Ker(
s
q
), it follows that C (q
1/2
) R
q
and C R (q
1/2
)
since C B
q
(M
q
). This proves the inclusions Ker(
s
q
) i
q
(R).
The opposite inclusions are immediate.
By Lemma 3.16 a current C B
s
q
(M
q
) has zero cohomology class,
that is, it is in the kernel of the cohomology map, if and only if there exists
a distribution U
C
H
s+1
q
(M) such that dU

C
= C in D(M
q
) and
this identity holds in S
q
(M), hence in W
s
q
(M), if C B
s
q
(M). It is
immediate to verify that U
C
I
s1
q
(M
q
) if and only if C B
s
q
(M
q
)
and that U
C
I
s1
q
(M) if and only if C B
s
q
(M). By Lemma 3.14 we
have thus proved that the map j
q
: B
s
q
(M
q
) H
1
(M
q
, R) has
kernel equal to the range
s
q
_
B
s1
q
(M)
_
and that the map j
q
: B
s
q
(M)
H
1
(M, R) has kernel equal to the range
s
q
_
B
s1
q
(M)
_
.
Finer results on invariant distributions and on smooth solutions of the co-
homological equation for directional ows can be obtained by combining
the results of this section with the renormalization method based on the
Teichmller ow and related cocycles, such as the Kontsevich-Zorich co-
cycle. Our goal is to improve upon the results of Marmi, Moussa and Yoc-
coz [MMY05] who have studied the cohomological equation for interval
exchange transformations solely by methods based on the renormalization
dynamics (the Rauzy-Veech-Zorich induction).
4. COCYCLES OVER THE TEICHMLLER FLOW
4.1. The Kontsevich-Zorich cocycle. The Kontsevich-Zorich cocycle is
a multiplicative cocycle over the Teichmller geodesic ow on the moduli
56 GIOVANNI FORNI

space of (orientable) holomorphic quadratic differentials on compact Rie-
mann surfaces. This cocycle appears in the study of the dynamics of interval
exchange transformations and of (translation) ows on surfaces for which
it represents a renormalization dynamics and of the Teichmller ow itself.
In fact, the study tangent cocycle of the Teichmller ow can be reduced to
that of the Kontsevich-Zorich cocycle.
Let T
g
and Q
g
be respectively the Teichmller spaces of complex (confor-
mal) structures and of holomorphic quadratic differentials on a surface of
genus g 1. We recall that the spaces T
g
and Q
g
can be described as fol-
lows. Let Diff
+
0
(M) is the group of orientation preserving diffeomorphisms
of the surface M which are isotopic to the identity (equivalently, it is the
connected component of the identity in the Lie group of all orientation pre-
serving diffeomorphisms of M). By denition
(4.1)
T
g
:= complex (conformal) structures /Diff
+
0
(M) ,
Q
g
:= holomorphic quadratic differentials /Diff
+
0
(M) .
Atheoremof L. Ahlfors, L. Bers and S. Wolpert states that T
g
has a complex
structure holomorphically equivalent to that of a Stein (strongly pseudo-
convex) domain in C
3g3
[Ber74], 6, or [Nag88], Chap. 3, 4 and Appendix
6. The space Q
g
of holomorphic quadratic differentials is a complex vector
bundle over T
g
which can be identied to the cotangent bundle of T
g
. Let

g
:= Diff
+
(M)/Diff
+
0
(M) be the mapping class group and let R
g
, M
g
be respectively the moduli spaces of complex (conformal) structures and
of holomorphic quadratic differentials on a surface of genus g 1. The
spaces R
g
and M
g
can be described as the quotient spaces:
(4.2) R
g
:= T
g
/
g
, M
g
:= Q
g
/
g
,
In case g = 1, the Teichmller space T
1
of elliptic curves (complex struc-
tures on T
2
) is isomorphic to the upper half plane C
+
and the Teichmller
space Q
1
of holomorphic quadratic differentials on elliptic curves is a com-
plex line bundle over T
1
(see [Nag88], Ex. 2.1.8). The mapping class group
can be identied with the lattice SL(2, Z) which acts on the upper half plane
C
+
in the standard way. The moduli space R
1
:= C
+
/SL(2, Z) is a non-
compact nite volume surface with constant negative curvature, called the
modular surface. The moduli space M
1
can be identied to the cotangent
bundle of the modular surface.
The Teichmller (geodesic) ow on M
g
can be dened as the geodesic ow
for a natural metric on R
g
called the Teichmller metric. Such a metric
measures the amount of quasi-conformal distorsion between two differ-
ent (equivalent classes of) complex structures in R
g
. In the higher genus
case, the Teichmller metric is not Riemannian, but only Finsler (that is,
SOBOLEV REGULARITY OF SOLUTIONS 57
the norm on each tangent space does not come from an euclidean product)
and, as H. Masur proved, does not have negative curvature in any reason-
able sense [Ber74], 3 (E). If g = 1, the Teichmller metric coincides with
the Poincar metric on the modular surface R
1
[Nag88], 2.6.5, in particular
it is Riemannian with constant negative curvature.
There is a natural action of the Lie group GL(2, R) on Q
g
(see also [HS05],
1.4 or [Mas05], 3) which is dened as follows. The map
q (F
q
, F
q
) , q Q
g
,
is a bijection between the space Q
g
and the space of all pairs of transverse
measured foliations. We recall that transversality for a pair (F
,
F

) of mea-
sured foliations is taken in the sense that F and F

have a common set of


(saddle) singularities, have the same index at each singularity and are trans-
verse in the standard sense on M . The set of common singularities
coincides with the set
q
of zeroes of the holomorphic quadratic differen-
tial q (F, F

). Any pair of transverse measured foliations is determined


locally by a pair (,

) of (locally dened) transverse real-valued closed


1-forms. The group GL(2, R) acts naturally by left multiplication on the
space of (locally dened) pairs of transverse real-valued closed 1-forms,
hence it acts on the space of all pairs of transverse measured foliations an
on the space of Q
g
of holomorphic quadratic differentials. Such an action
is equivariant with respect to the action of the mapping class group, hence
it passes to the quotient M
g
. The Teichmller ow G
t

tR
is given by
the action of the diagonal subgroup diag (e
t
, e
t
)
tR
on Q
g
(on M
g
). In
other terms, if we identify holomorphic quadratic differentials with pairs of
transverse measured foliations as explained above, we have:
(4.3) G
t
(F
q
, F
q
) := (e
t
F
q
, e
t
F
q
) .
In geometric terms, the action of the Teichmller ow on quadratic differ-
entials induces a one-parameter family of deformations of the conformal
structure which consist in contracting along vertical leaves (with respect to
the horizontal length) and expanding along horizontal leaves (with respect
to the vertical length) by reciprocal (exponential) factors. The reader can
compare the denition in terms of the SL(2, R)-action with the analogous
description of the geodesic ow on a surface of constant negative curvature
(such as the modular surface). In fact, in case g = 1 the above denition
reduces to the standard Lie group presentation of the geodesic ow on the
modular surface: the unit sub-bundle M
(1)
1
M
1
of all holomorphic qua-
dratic differentials of unit total area on elliptic curves can be identied with
the homogeneous space SL(2, R)/SL(2, Z) and the geodesic ow on the
modular surface is then identied with the action of the diagonal subgroup
of SL(2, R) on SL(2, R)/SL(2, Z).
58 GIOVANNI FORNI

We list below, following [Vee90], [Kon97], the main structures carried by
the moduli space M
g
of quadratic differentials:
(1) the moduli space M
g
is a stratied analytic orbifold; each stra-
tum M

M
g
(determined by the multiplicities = (k
1
, . . . , k

)
of the zeroes p
1
, . . . , p

of quadratic differentials) is GL(2, R)-


invariant, hence in particular G
t
-invariant;
(2) The total area function A : M
g
R
+
,
A(q) :=
_
M
[q[ ,
is SL(2, R)-invariant; hence the unit bundle M
(1)
g
:= A
1
(1) and
its strata M
(1)

:= M

M
(1)
g
are SL(2, R)-invariant and, in partic-
ular, G
t
-invariant.
Let M

be a stratum of orientable quadratic differentials, that is,


quadratic differentials which are squares of holomorphic 1-forms.
In this case, the natural numbers (k
1
, . . . , k

) are all even.


(3) The stratum of squares M

has a locally afne structure modeled


on the afne space H
1
(M,

; C), with

:= p
1
, . . . , p

. Local
charts are given by the period map q [q
1/2
] H
1
(M,

; C).
(4) The Lebesgue measure on the euclidean space H
1
(M,

; C), nor-
malized so that the quotient torus
H
1
(M,

; C) / H
1
(M,

; Z iZ)
has volume 1, induces an absolutely continuous SL(2, R)-invariant
measure

on M

. The conditional measure


(1)

induced on the
stratum M
(1)

is SL(2, R)-invariant, hence G


t
-invariant.
All the above structures (the stratication, the area function, the locally
afne structure on the strata of squares) lift to corresponding structures at
the level of the Teichmller space of quadratic differentials, equivariant un-
der the action of the mapping class group.
It was discovered by W. Veech [Vee90] that M
(1)

has in general several


connected components. The connected components for the strata of abelian
differentials (or equivalently of orientable quadratic differentials) have been
classied completely by M. Kontsevich and A. Zorich [KZ03]. A similar
classication for the case of strata of non-orientable holomorphic quadratic
differentials has been recently obtained by E. Lanneau in his thesis [Lan03].
Taking this phenomenon into account, the following result holds:
Theorem 4.1. [Mas82], [Vee86] The total volume of the measure
(1)

on
M
(1)

is nite and the Teichmller geodesic ow G


t

tR
is ergodic on each
connected component of M
(1)

.
SOBOLEV REGULARITY OF SOLUTIONS 59
We will describe below several results about the Lyapunov structure of var-
ious coycles over the Teichmller geodesic ow (including the tangent co-
cycle). We refer the reader to the recent and excellent survey [BP05] by
L. Barreira and Ya. Pesin (and references therein), which covers all the
relevant results on the theory of Lyapunov exponents, including Oseledecs
multiplicative ergodic theorem and the Oseledec-Pesin reduction theorem.
The Lyapunov spectrumof the Teichmller ow, with respect to any ergodic
invariant probability measure on the moduli space, has symmetries. In
fact, there exists non-negative numbers

1
= 1

g
such that
the Lyapunov spectrum of the Teichmller ow has the following form (see
5 in [Zor96], 7 in [Kon97] or 2.3 in [Zor99]):
(4.4)
2 (1 +

2
) (1 +

g
)
#()1
..
1 = = 1 (1

g
)
(1

2
) 0 = 0 (1

2
) (1

g
)
1 = = 1
. .
#()1
(1 +

g
) (1 +

2
) 2 .
In [Vee86] Veech proved that the Teichmller ow is non-uniformly hy-
perbolic, in the sense that all of its Lyapunov exponents, except one cor-
responding to the ow direction, are non-zero. By formulas (4.4) Veechs
theorem can be formulated as follows:
Theorem 4.2. [Vee86] The inequality
(4.5)

2
<

1
= 1 .
holds if is the absolutely continuous SL(2, R)-invariant ergodic proba-
bility measure on any connected component of a stratum M
(1)

M
(1)
g
of
orientable quadratic differentials.
M. Kontsevich and A. Zorich have interpreted the non-negative numbers
(4.6)

1
= 1

g
as the non-negative Lyapunov exponents of a symplectic cocycle over the
Teichmller ow, which we now describe.
The Kontsevich-Zorich cocycle
t

tR
is a cocycle over the Teichmller
owG
t

tR
on the moduli space M
g
, dened as the projection of the trivial
cocycle
(4.7) G
t
id : Q
g
H
1
(M, R) Q
g
H
1
(M, R)
onto the orbifold vector bundle H
1
g
(M, R) over M
g
dened as
(4.8) H
1
g
(M, R) :=
_
Q
g
H
1
(M, R)
_
/
g
.
The mapping class group
g
acts naturally on the cohomology by pull-back.
60 GIOVANNI FORNI

The Kontsevich-Zorich cocycle was introduced in [Kon97] as a continuous-
time version of the Zorich cocycle. The Zorich cocycle was introduced ear-
lier by A. Zorich [Zor96], [Zor97] in order to explain polynomial deviations
in the homological asymptotic behavior of typical leaves of orientable mea-
sured foliations on compact surfaces, a phenomenon he had discovered in
numerical experiments [Zor94]. We recall that the real homology H
1
(M, R)
and the real cohomology H
1
(M, R) of an orientable closed surface M are
(symplectically) isomorphic by Poincar duality.
Let M

M
g
be a stratum of orientable holomorphic quadratic differen-
tials and Q

Q
g
the pull-back of the stratumM

to the Teichmller space


of quadratic differentials. The relative Kontsevich-Zorich coycle

tR
is dened as the projection of the trivial cocycle
(4.9) G
t
id :
_
qQ
q H
1
(M,
q
; R)
_
qQ
q H
1
(M,
q
; R)
on the orbifold vector bundle H
1

(M,

; R) over M

dened as
(4.10) H
1

(M,

; R) :=
_
_
qQ
q H
1
(M,
q
; R)
_
/
g
.
By a similar construction it is possible to dene a punctured Kontsevich-
Zorich coycle
t

tR
dened as the projection of the trivial cocycle
(4.11) G
t
id :
_
qQ
q H
1
(M
q
; R)
_
qQ
q H
1
(M
q
; R)
onto the orbifold vector bundle H
1

(M

; R) over M

dened as
(4.12) H
1

(M

; R) :=
_
_
qQ
q H
1
(M
q
; R)
_
/
g
.
However, by Poincar-Lefschetz duality, there exists a natural isomorphism
(4.13) H
1
(M,

; R) H
1
(M

; R)

,
hence the punctured Kontsevich-Zorich cocycle
t

tR
is isomorphic to
the dual

tR
of the relative Kontsevich-Zorich cocycle.
Let H
1

(M, R) be the restriction to the stratumM


(1)

of the bundle H
1
g
(M, R)
dened in (4.8). Let K
1

(M,

; R) be the bundle dened as the kernel of


the natural bundle map H
1

(M,

; R) H
1

(M, R).
Lemma 4.3. The sub-bundle K
1

(M,

; R) is invariant under the rela-


tive Kontsevich-Zorich cocycle

tR
. The Lyapunov spectrum of the
restriction

t
[K
1

(M,

; R) consists of the single exponent 0 with mul-


tiplicity #(

) 1. In fact, there exists a Lyapunov norm on the bundle


K
1

(M,

; R) for which the cocycle is isometric.


SOBOLEV REGULARITY OF SOLUTIONS 61
Proof. By de Rham theorem for the relative cohomology, the relative co-
homology complex H

(M,
q
; R) is isomorphic to the cohomology of the
complex of relative differential forms, which are dened as all differential
forms vanishing at
q
. The map H

(M,
q
; R) H

(M, R) is naturally
dened on the de Rham cohomology since every closed, exact relative form
is also a closed, respectively exact form on M. It follows that any class
c H
1
(M,
q
; R) which belongs to the kernel K
1
(M,
q
; R) of the map
H
1
(M,
q
; R) H
1
(M, R) can be represented by a differential 1-form
exact on M. Hence for any c K
1
(M,
q
; R) there exists a smooth func-
tion f
c
: M R, uniquely determined up to the addition of any function
vanishing at
q
, such that c = [df
c
] as a relative de Rham cohomology class.
As a consequence of the above discussion, the following formula yields a
well-dened euclidean norm | |
K
on K
1
(M,
q
; R):
(4.14) |c|
2
K
:=

p
1
,p
2
q
[f
c
(p
1
) f
c
(p
2
)[
2
, c = [df
c
] K
1
(M,
q
; R) .
The norm (4.14) induces an euclidean norm on the bundle K
1

(M,

; R)
which is invariant under the relative Kontsevich-Zorich cocycle.
Since the vector bundle H
1
g
(M, R) has a symplectic structure, given by the
intersection form on its bers, which are isomorphic to the cohomology
H
1
(M, R), the Lyapunov spectrum of the cocycle
t

tR
with respect to
any G
t
-invariant ergodic probability measure on M
(1)
g
is symmetric:
(4.15)

g
0

g+1
=

2g
=

1
.
The non-negative Kontsevich-Zorich exponents coincide with the numbers
(4.6) which appear in the Lyapunov spectrum (4.4) of the Teichmller ow.
This relation is explained for instance in [For05].
Zorich conjectured (see [Zor96]) that the exponents (4.15) are all distinct
and different from zero when is the canonical absolutely continuous er-
godic invariant probability measure on any connected component of a stra-
tum M
(1)

of orientable quadratic differentials. In other terms he conjec-


tured that the canonical measures are KZ-hyperbolic and and KZ-simple,
according to the following:
Denition 4.4. A G
t
-invariant ergodic probability measure on a stra-
tum of orientable quadratic differentials will be called KZ-hyperbolic if the
Kontsevich-Zorich cocycle (
t
, ) is non-uniformly hyperbolic, in the
sense that its Lyapunov exponents satisfy the inequalities
(4.16)

1
= 1

g
> 0 .
62 GIOVANNI FORNI

A KZ-hyperbolic measure on a stratum of orientable quadratic differen-
tials will be called KZ-simple if the Kontsevich-Zorich cocycle (
t
, ) is
simple, in the sense that all inequalities (4.16) are strict.
In [For02] we have proved the following result:
Theorem 4.5. ([For02], Th. 8.5) The absolutely continuous, SL(2, R)-
invariant, ergodic probability measure on any connected component of a
stratumM
(1)

M
(1)
g
of orientable quadratic differentials is KZ-hyperbolic.
In [For05], 7, we have given an example of a SL(2, R)-invariant measure
on M
3
, supported on the closed SL(2, R)-orbit of a particular genus 3
branched cover of the 2-torus, such that

2
=

3
= 0.
A proof of the simplicity of the Zorich and Kontsevich-Zorich cocycles,
which yields in particular a new independent proof of Theorem 4.5, has
been recently obtained by A. Avila and M. Viana [AV05] by methods com-
pletely different from ours.
Theorem 4.6. [AV05] The absolutely continuous, SL(2, R)-invariant, er-
godic probability measure on any connected component of a stratumM
(1)


M
(1)
g
of orientable quadratic differentials is KZ-simple.
The results of this paper do not depend in any way on the simplicity of the
Kontsevich-Zorich cocycle, while the non-uniform hyperbolicity is crucial
to the sharp estimates on the regularity of solutions of the cohomological
equation proven in 5.3. However, our results can be rened by taking into
account that the Kontsevich-Zorich exponents (4.16) are all distinct.
Our proof of Theorem 4.5 in [For02] yields in particular a new independent
proof of a strong version of Veechs Theorem 4.2. In fact, we have proved
that the strict inequality (4.5) holds for an arbitrary probability G
t
-invariant
measure on any stratum of orientable quadratic differentials. By combining
our methods with a recent result of J. Athreya [Ath06] on large deviations
of the Teichmller ow, it is possible to prove a similar strict upper bound
for the second exponent for Lebesgue almost all quadratic differentials in
any orbit of the circle group SO(2, R) on any stratum M

of orientable
quadratic differentials.
We recall below the variational formulas for the evolution of Hodge norm
of absolute cohomology classes under the Kontsevich-Zorich cocycle. Fol-
lowing 2 in [For02], such formulas can be written in terms of a natural
R-linear extension

U
q
of the partial isometry U
q
, dened in formula (3.6),
which plays a crucial role in the construction of distributional solutions of
the cohomological equation. Let

U
q
: L
2
q
(M) L
2
q
(M) be the R-linear
SOBOLEV REGULARITY OF SOLUTIONS 63
isometry dened as follows in terms of the partial isometry U
q
and of the
orthogonal projections

q
: L
2
q
(M) M

q
:
(4.17)

U
q
:= U
q
(I

q
)

q
.
Let q
t

tR
be the orbit of an orientable quadratic differential q Q
(1)

un-
der the Teichmller ow G
t

tR
. We remark that by the denition of the
Teichmller ow G
t

tR
, the area form
t
of the metric induced by the
quadratic differential q
t
is constant equal to
q
for all t R. Hence the
Hilbert space L
2
q
(M) is invariant under the action of the Teichmller ow
(in fact, it is invariant under the full SL(2, R) action). For each t R,
let M
t
the Riemann surface carrying q
t
Q
(1)

and let M

t
L
2
q
(M) be
the space of meromorphic, respectively anti-meromorphic, functions on the
Riemann surface M
t
. Such spaces are respectively the kernels of the ad-
joints of the Cauchy-Riemann operators

t
, associated to the holomorphic
quadratic differential q
t
. The dimension of M

t
is constant equal to the
genus g 1 of M (see 2.1) and it can be proved that M

t

tR
are smooth
families of g-dimensional subspaces of the xed Hilbert space L
2
q
(M).
By (3.40) there exists a one-parameter family m
+
t

tR
M
+
t
such that
(4.18) c
t
= c
+
qt
(m
+
t
) := [(m
+
t
q
1/2
t
)] H
1
(M
t
, R) .
Lemma 4.7. ([For02], Lemma 2.1) The ordinary differential equation
(4.19) u

=

U
qt
(u)
is well dened in L
2
q
(M) and saties the following properties:
(1) Solutions of the Cauchy problem for (4.19) exist for all times and
are uniquely determined by the initial condition;
(2) If u
t
L
2
q
(M) is any solution of (4.19) such that the initial condi-
tion u
0
M
+
q
, then u
t
M
+
t
for all t R.
(3) Let m
+
t
M
+
t
be the unique solution of (4.19) with initial condition
m
+
0
= m
+
M
+
q
. For all t R, we have
(4.20)
t
_
c
+
q
(m
+
)
_
= c
+
qt
(m
+
t
) .
It follows immediately from Proposition 2.4 that, for every u L
2
q
(M),
there exist functions v

H
1
q
(M) such that
(4.21) u =
+
q
v
+
+

q
(u) =

q
v

+
+
q
(u) .
The O. D. E. in formula (4.19) can be written explicitly, in terms of the
orthogonal decompositions (4.21), as follows. Let

t
: L
2
q
(M) M

t
denote the orthogonal projections in the (xed) Hilbert space L
2
q
(M). By
64 GIOVANNI FORNI

denition, the projections

t
coincide with the projections

q
for q = q
t
,
for any t R. A function u C
1
_
R, L
2
q
(M)
_
satises equation (4.19) iff
(4.22)
_
u
t
=
+
t
v
t
+

t
(u
t
) ;
d
dt
u
t
=

t
v
t

t
(u
t
) .
An immediate consequence of Lemma 4.7 is the following result on the
variation of the Hodge norm of cohomology classes under the action of the
Kontsevich-Zorich cocycle. Let B
q
: L
2
q
(M)L
2
q
(M) C be the complex
bilinear form given by
(4.23) B
q
(u, v) :=
_
M
u v
q
, for all u, v L
2
q
(M) .
Corollary 4.8. ([For02], Lemma 2.1) The variation of the Hodge norm
|c
t
|
H
, which coincides with the L
2
q
-norm [m
+
t
[
0
under the identication
(4.18), is given by the following formulas:
(4.24)
(a)
d
dt
[m
+
t
[
2
0
= 2 [B
q
(m
+
t
)] = 2
__
M
(m
+
t
)
2

q
_
;
(b)
d
2
dt
2
[m
+
t
[
2
0
= 4
_
[

t
(m
+
t
)[
2
0

__
M
(
+
t
v
t
) (

t
v
t
)
q
__
.
The second order variational formula (4.24), (b), is crucial in our proof of
lower bounds for the Kontsevich-Zorich exponents (see Theorem 4.5). The
rst order variational formula (4.24), (a), implies quite immediately an ef-
fective upper bound for the second exponent, which yields in particular the
average spectral gap result proved in [For02], Corollary 2.2 (a generaliza-
tion of Veechs Theorem 4.2 to arbitrary G
t
-invariant ergodic probability
measures):
Theorem 4.9. ([For02], Corollary 2.2) The inequality
(4.25)

2
<

1
= 1 .
holds for any G
t
-invariant ergodic probability measure on any connected
component of any stratum M
(1)

of orientable quadratic differentials.


The above spectral gap result implies the following unique ergodicity theo-
rem for measured foliations:
Corollary 4.10. For any stratum M
(1)

of orientable quadratic differentials,


the set of quadratic differentials q M
(1)

with minimal but not uniquely


ergodic horizontal [vertical] foliation has zero measure with respect to any
G
t
-invariant probability measure on M
(1)

.
SOBOLEV REGULARITY OF SOLUTIONS 65
In the remainder of this section we prove a pointwise spectral gap result
which holds for almost all quadratic differentials in any orbit of the circle
group SO(2, R) on any stratum M
(1)

of orientable quadratic differentials.


The argument is based on formula (4.24), (a) and, as mentioned above, on
a result of J. Athreya [Ath06] on large deviations of the Teichmller ow.
The upper second Lyapunov exponent of the Kontsevich-Zorich cocycle at
any (orientable) quadratic differentials q M
(1)

is dened as follows. Let


I
q
(M, R) H
1
(M, R) be the subspace of real dimension 2 dened as
(4.26) I
q
(M, R) := R (q
1/2
) +R (q
1/2
)
and let I

q
(M, R) be the symplectic orthogonal of I
q
(M, R) in H
1
(M, R),
with respect to the symplectic structure induced by the intersection form:
(4.27) I

q
(M, R) := c H
1
(M, R)[c [q
1/2
] = 0 .
The complementary sub-bundles I

(M, R) and I

(M, R) H
1
(M, R),
with bers at any q M
(1)

respectively equal to I
q
(M, R) and I

q
(M, R),
are invariant under the Kontsevich-Zorich cocycle. In fact, it is immediate
to verify that the sub-bundle I

(M, R) is invariant under the Kontsevich-


Zorich cocycle and that the Lyapunov spectrum of the restriction of the
Kontsevich-Zorich cocycle to I

(M, R) equals 1, 1 (both exponents


with multiplicity 1). Since the Kontsevich-Zorich cocycle is symplectic,
the symplectic orthogonal bundle I

(M, R) is also invariant. In addition, it


is not difcult to verify that 1 is the top (upper) exponent for the cocycle on
the full cohomology bundle H
1

(M, R).
The second upper (forward) exponent of the Kontsevich-Zorich cocycle is
the top upper (forward) Lyapunov exponent at any quadratic differential
q M
(1)

of the restriction of the cocycle to the sub-bundle I

(M, R):
(4.28)
+
2
(q) := limsup
t+
1
t
log |
t
[I

q
(M, R)|
H
.
Theorem 4.11. For any stratum M
(1)

of orientable quadratic differentials,


there exists a measurable function L

: M
(1)

[0, 1) such that for any


(orientable) quadratic differential q M
(1)

,
(4.29)
+
2
(q

) L

(q) < 1 , for almost all S


1
.
Proof. The argument follows closely the proof of Corollary 2.2 in [For02].
Under the isomorphism (3.40), the vector space I

q
(M, R) is represented
by meromorphic functions with zero average (orthogonal to constant func-
tions). It can be seen that the subspace of zero average meromorphic func-
tions is invariant under the ow of equation (4.20) or, equivalently (4.22).
66 GIOVANNI FORNI

By formula (a) in (4.24),
(4.30)
d
dt
log [m
+
t
[
2
0
= 2
B
q
(m
+
t
)
[m
+
t
[
2
0
.
Following [For02], we introduce a continuous function
+

: M
(1)

R
+
dened as follows: for any q M
(1)

,
(4.31)
+

(q) := max
[B
q
(m
+
)[
[m
+
[
2
0
[ m
+
M
+
q
0 ,
_
M
m
+

q
= 0 .
Since by the Schwarz inequality,
(4.32) [B
q
(m
+
t
)[ = [(m
+
t
, m
+
t
)
q
[ [m
+
t
[
2
0
,
the image of the function
+

is contained in the interval [0, 1]. We claimthat

(q) < 1 for all q M


(1)

. In fact,
+
(q) = 1 if and only if there exists
a non-zero meromorphic function with zero average m
+
M
+
q
such that
[(m
+
, m
+
)
q
[ = [m
+
[
2
0
. A well-known property of the Schwarz inequality
then implies that there exists u C such that m
+
= u m
+
. However,
it cannot be so, since m
+
would be meromorphic and anti-meromorphic,
hence constant, and by the zero average condition it would be zero.
It follows from formula (4.30) that, for any q M
(1)

,
(4.33)
1
t
log |
t
[I

q
(M, R)|
H

1
t
_
t
0

(G
s
(q)) ds .
Let q
0
M
(1)

. By the large deviation result of J. Athreya (see [Ath06],


Corollary 2.4) the following holds. For any < 1 there exists a compact
set K M
(1)

such that, for almost all q SO(2, R) q


0
,
(4.34) limsup
t
1
t
[0 s t [ G
s
(q) , K[ .
Let
K
:= max

(q) [ q K and, for any (t, q) R


+
M
(1)

, let
E
K
(t, q) := [0 s t [ G
s
(q) , K[ .
Since the function
+

is continuous and
+

(q) < 1 for all q M


(1)

, its
maximum on any compact set is < 1, in particular
K
< 1. The following
immediate inequality holds:
(4.35)
_
t
0

(G
s
(q)) ds (1
K
) E
K
(t, q) + t
K
.
SOBOLEV REGULARITY OF SOLUTIONS 67
It follows from (4.33), (4.34) and (4.35) that, for almost all q SO(2, R)
q
0
,
(4.36) limsup
t+
1
t
log |
t
[I

q
(M, R)|
H
(1
K
) +
K
< 1 .
The function L

: M
(1)

[0, 1) can be dened for every q


0
M
(1)

as
the essential supremum of the second upper Lyapunov exponent over the
orbit SO(2, R) q
0
of the circle group. Such a function is measurable by
denition and it is everywhere < 1 by the above argument.
For any Oseledec regular point q Q
(1)

of the Kontsevich-Zorich cocycle,


let E
+
q
(M, R)[E

q
(M, R)] H
1
(M, R) be the unstable [stable] subspace of
the Kontsevich-Zorich cocycle. Homology cycles which are Poincar duals
to cohomology classes in E

q
(M, R) are called Zorich cycles for the foli-
ation F
q
. It follows from Theorem 4.5 and from the symplectic property
of the cocycle that E

q
are transverse Lagrangian subspaces (with respect
to the intersection form), as conjectured by Zorich in [Zor96]. In [For02],
Theorem 8.3 (see also [For05], Theorem 8.2) we have proved the following
representation theorem:
Theorem 4.12. For Lebesgue almost all q M
(1)

, we have
(4.37) E
+
q
(M, R) = H
1,1
q
(M, R) , E

q
(M, R) = H
1,1
q
(M, R) .
(The Poincar duals of Zorich cycles for a generic orientable measured
foliation F are represented by basic currents for F of Sobolev order 1).
We prove below a conjectural relation between Lyapunov exponents of a
cohomology class under the Kontsevich-Zorich cocycle and the Sobolev
regularity of the basic current representing the cohomology class. This re-
sult answers in the afrmative a question posed by the author in [For02]
(Question 9.9) and, independently by M. Kontsevich (personal communi-
cation). In order to formulate our results in the greatest possible generality,
we introduce the following class of measures on the moduli space.
Denition 4.13. A measure on the moduli space M
g
of holomorphic qua-
dratic differentials will be called SO(2, R)-absolutely continous if induces
absolutely continous conditional measures on

()-almost every ber of


the bration : M
g
M
g
/SO(2, R) (with respect to the Haar/Lebesgue
measure class on each ber).
It is immediate that any SO(2, R)-invariant, hence a fortiori any SL(2, R)-
invariant measure is SO(2, R)-absolutely continous.
68 GIOVANNI FORNI

Lemma 4.14. Let be any SO(2, R)-absolutely continous, G
t
-invariant
ergodic probability measure on a stratum M
(1)

of orientable quadratic dif-


ferentials. For -almost all q M
(1)

, the unique basic current C B


1
q
(M)
[C B
1
q
(M)] which represents a cohomology class c E
+
q
(M, R)
[c E

q
(M, R)] of Lyapunov exponent (c) > 0 [(c) < 0] under the
Kontsevich-Zorich cocycle, has the following Sobolev regularity:
C W
s
q
(M) , for all s > 1 [(c)[ .
Proof. The argument follows the proof of Lemma 8.2 in [For02]. The in-
terpolation inequality (2.26) for fractional weighted Sobolev spaces allows
us to estimate the Sobolev regularity of the basic current constructed there.
Let q M
(1)

be any Oseledec regular point of the Kontsevich-Zorich co-


cycle and let c H
1
(M, R) be a cohomology class. Let q
t

tR
Q
(1)

denotes the lift to the Teichmller space of the orbit G


t
(q)
tR
of q under
the Teichmller ow. Let M
+
t
L
2
q
(M) be the space of meromorphic
functions on the Riemann surface M
t
carrying the quadratic differential
q
t
Q
(1)

. According to the representation formula (3.40), there exists a


(smooth) family m
+
t

tR
M
+
t
such that, for each t R,
(4.38)
t
(c) = [m
+
t
q
1/2
t
] .
By the variational formulas (4.20) and (4.22), there exists a (smooth) family
v
tR
H
1
(M) of zero-average functions such that
(4.39)
_
m
+
t
=
+
t
v
t
+

t
(m
+
t
) ,
d
dt
m
+
t
=

t
v
t

t
(m
+
t
) .
Since
t
(c) c H
1
(M, R) (by the denition (4.7) of the Kontsevich-
Zorich cocycle
t
over the Teichmller space), for each t R there
exists a unique zero average function U
t
L
2
q
(M) such that
(4.40) dU
t
= [m
+
t
q
1/2
t
] [m
+
0
q
1/2
] .
It follows that the family U
t

tR
is smooth and satises, by the variational
formulas (4.39), the following Cauchy problem in L
2
q
(M):
(4.41)
_
d
dt
U
t
= 2 (v
t
) ,
U
0
= 0 ,
We claim that, if c E
+
q
(M, R) [c E

q(M, R)] has Lyapunov exponent


(c) > 0 [(c) < 0], the set U
t
[ t 0 [ U
t
[ t 0] is a bounded subset
of the Hilbert space H
s
q
(M) for any s < [(c)[.
By Oseledecs theorem, for any 0 < < [(c)[, there exist a measurable
function K

> 0 on M
(1)

and such that


(4.42) |
t
(c)|
H
= [m
+
t
[
0
K

(q)[m
+
0
[
0
exp([t[) , t 0 [t 0] .
SOBOLEV REGULARITY OF SOLUTIONS 69
For any (orientable) quadratic differential q Q
(1)

, let |q| denote the


length of the shortest geodesic segment with endpoints in the set
q
(of
zeroes of q) with respect to the induced metric. By the Poincar inequality
proved in [For02], Lemma 6.9, there exists a constant K
g,
> 0 (depend-
ing on the genus g 2 of the Riemann surface M
q
and on the cardinality
:= #(
q
) of any q Q
(1)

) such that
(4.43) [v
_
M
v
q
[
0

K
g,
|q|
Q(v, v) , for all v H
1
q
(M) .
By the commutativity property (2.6) of the horizontal and vertical vector
elds, the Dirichlet form Q of the quadratic differential can be written as
Q(v, v) = [

q
v[
2
0
, for all v H
1
q
(M)
(where

q
are the Cauchy-Riemann operators introduced in 2).
By the Poincar inequality and by the orthogonality of the decompositions
in (4.39) with respect to the invariant euclidean structure on L
2
q
(M), we
have
(4.44) [v
t
[
0
K
g,
|q
t
|
1
[
+
t
v
t
[
0
K
g,
|q
t
|
1
[m
+
t
[
0
.
It follows by (4.41), (4.42) and (4.44) that there exists a measurable function
K

> 0 on M
(1)

such that if c E
+
q
(M, R) [c E

q
(M, R)] the following
inequality holds for all t 0 [ t 0]:
(4.45) [
d
dt
U
t
[
0
2 [v
t
[
0
K

(q) [m
+
0
[
0
|q
t
|
1
e
|t|
.
By (4.39) and (4.41) we also have the following straightforward estimate
for the norm of the funcion
d
dt
U
t
in the xed Hilbert space H
1
q
(M):
(4.46) [
d
dt
U
t
[
1
e
|t|
[m
+
t
[
0
K

(q)[m
+
0
[
0
e
(1)|t|
.
It follows from the inequalities (4.45) and (4.46), by the interpolation in-
equality proved in Lemma 2.10, that for any s (0, 1) there exists a mea-
surable function K
s

> 0 on the stratum Q


(1)

such that, if c E
+
q
(M, R)
[c E

q
(M, R)] the following inequality holds for all t 0 [ t 0]:
(4.47) [
d
dt
U
t
[
s
e
|t|
[m
+
t
[
0
K
s

(q)[m
+
0
[
0
e
(s)|t|
.
Since U
0
= 0, by Minkowskis integral inequality we nally obtain the
estimate:
(4.48) [U
t
[
s
K
s

(q)[m
+
0
[
0
_
|t|
0
e
(s)
|q

|
1
d .
70 GIOVANNI FORNI

By the logarithmic law for the Teichmller geodesic ow on the moduli
space, proved by H. Masur in [Mas93], the following estimate holds for
almost all quadratic differentials q M
(1)

(see [Mas93], Prop. 1.2):


(4.49) limsup

log |q

|
log [[

1
2
.
As a consequence, for -almost all q M
(1)

and for any s < , the integral


in formula (4.48) is uniformly bounded for t 0 [t 0]. Since for any
s < [(c)[, there exists (s, [(c)[), it follows that the family of func-
tions U
t
[ t 0 [U
t
[ t 0] is uniformy bounded in the Sobolev space
H
s
q
(M) for any s < [(c)[, as claimed.
For any s < [(c)[, let U
+
s
H
s
q
(M) [U

s
H
s
q
(M)] be any weak limit of
the family U
t
as t [as t ], which exists since all bounded
subsets of the separable Hilbert space H
s
q
(M) are sequentially weakly com-
pact. Since the functions U
t
have zero average for all t R and the sub-
space of zero average functions is closed in H
s
q
(M) for all s > 0, the weak
limit U
+
s
[U

s
] has zero average. By contraction of the identity (4.40) with
the horizontal vector eld S [with the vertical vector eld T] we have:
(4.50)
SU
t
= (m
+
0
) + e
t
(m
+
t
) , t 0 ,
[ TU
t
= (m
+
0
) + e
t
(m
+
t
) , t 0 , ]
and by taking the limit as t [as t +],
(4.51)
SU
+
s
= (m
+
0
) ,
[ TU

s
= (m
+
0
) .]
Since for almost all quadratic differential q Q
(1)

the horizontal foliation


[the vertical foliation] is ergodic, the solution U
+
L
2
q
(M) [U

L
2
q
(M)]
of the cohomological equation (4.51) is unique (if it exists). Hence there
exists a unique zero-average function U
+
L
2
q
(M) [U

L
2
q
(M)] , which
solves the cohomological equation (4.51), such that U
+
s
= U
+
[U

s
= U

]
for all s < [(c)[. As a consequence, U
+
H
s
q
(M) [U

H
s
q
(M)] for
all s < [(c)[. The current C
+
W
s1
q
(M) [C

W
s1
q
(M)] uniquely
determined by the identity
(4.52)
dU
+
= C
+
[m
+
0
q
1/2
] ,
[ dU

= C

[m
+
0
q
1/2
] , ]
is basic for the horizontal [vertical] foliation by Lemma 3.17 and represents
the cohomology class c E
+
q
(M, R) [c E

q
(M, R)] of Lyapunov expo-
nent (c) > 0 [(c) < 0].
SOBOLEV REGULARITY OF SOLUTIONS 71
4.2. Distributional cocycles. Let Q

(M) be the space of all orientable


quadratic differentials, holomorphic with respect to some complex struc-
ture on a closed surface M, with zeros of multiplicities = (k
1
, . . . , k

).
For any s R, there is a natural action of the group Diff
+
(M) of orientation
preserving diffeomorphisms on the trivial bundles
(4.53)
_
qQ(M)
q H
s
q
(M)
_
qQ(M)
q

H
s
q
(M) .
In fact, any diffeomorphism f Diff
+
(M) denes by pull-back an iso-
morphism f

:

H
s
q
(M)

H
s
f

(q)
(M) which maps the subspace H
s
q
(M)

H
s
q
(M) onto H
s
f

(q)
(M)

H
s
f

(q)
(M). The quotient bundles
(4.54)
_
_
qQ(M)
q H
s
q
(M)
_
/Diff
+
0
(M) ,
_
_
qQ(M)
q

H
s
q
(M)
_
/Diff
+
0
(M)
are well-dened bundles over the stratum Q

of the Teichmller space of


quadratic differentials. There is natural action of the mapping class group

g
on the bundles (4.54) induced by the action of Diff
+
(M) on the bundles
(4.53). The resulting quotient bundles
(4.55)
H
s

(M) :=
_
_
qQ(M)
q H
s
q
(M)
_
/Diff
+
(M) ,

H
s

(M) :=
_
_
qQ(M)
q H
s
q
(M)
_
/Diff
+
(M)
are well-dened bundles over the stratum M

of the moduli space. We also


introduce bundles of 1-currents over a stratum M

as follows:
(4.56)
W
s

(M) :=
_
_
qQ(M)
q W
s
q
(M)
_
/Diff
+
(M) ,

W
s

(M) :=
_
_
qQ(M)
q

W
s
q
(M)
_
/Diff
+
(M) .
Since, for all q Q

(M) and for all s R, the weighted Sobolev space of


1-currents W
s
q
(M) is isomorphic to the tensor product R
2
H
s
q
(M),
(4.57) W
s

(M) R
2
H
s

(M) ,

W
s

(M) R
2


H
s

(M)
72 GIOVANNI FORNI

Let G
s
t

tR
be the cocycle over the Teichmller ow, dened as the pro-
jection onto the bundle H
s

(M) of the trivial skew-product cocycle


(4.58) G
t
id :
_
qQ(M)
q H
s
q
(M)
_
qQ(M)
q H
s
q
(M) .
Let
s
t

tR
be the cocycle over the Teichmller ow, dened as the pro-
jection onto the bundle W
s

(M) of the trivial skew-product cocycle:


(4.59) G
t
id :
_
qQ(M)
q W
s
q
(M)
_
qQ(M)
q W
s
q
(M) .
Such cocycles can be described as the cocycles obtained by parallel trans-
port of distributions and 1-currents with respect to the trivial connection
along the orbits of the Teichmller ow. The following immediate identity
allows to express the cocycle
s
t

tR
in terms of the cocycle G
s
t

tR
:
(4.60)
s
t
:= diag(e
t
, e
t
) G
s
t
on W
s

(M) = R
2
H
s

(M) .
Lemma 4.15. For any s R, the spaces H
s

(M),

H
s

(M), W
s

(M) and

W
s

(M) are well-dened Hilbert bundles over a stratum M


(1)

of orientable
quadratic differentials. The ows G
s
t

tR
and
s
t

tR
are well-dened
smooth cocycles on the bundles H
s

(M) and W
s

(M) respectively.
Proof. Let q
0
Q

(M). There exists a neighbourhood S


0
M of the set

q
0
and a neighbourhood D
0
Q

(M) in the space of quadratic differen-


tials with zeros of multiplicities = (k
1
, . . . , k

) such that, for all q D


0
,

q
S
0
and the quadratic differential q is isotopic to q
0
on S
0
. Thus there
exists a smooth map f : D
0
Diff
+
(M) such that q = f

q
(q
0
) on S
0
. The
bundle H
s

(M) and

H
s

(M) are trivialized over D


0
by the map
(4.61)

H
s

(M)[D
0
D
0


H
s
q
0
(M);
(q, D)
_
q, (f
1
q
)

(D)
_
,
which maps the subspace H
s

(M)[D
0
onto D
0
H
s
q
0
(M). It follows that
H
s

(M) and

H
s

(M) are well-dened Hilbert bundles, hence so are W


s

(M)
and W
s

(M) by formula (4.57). The dynamical system G


s
t

tR
concides
with the product cocycle G
t
id on D
0
H
s
q
0
(M) with respect to the
trivialization (4.61), hence G
s
t

tR
and
s
t

tR
are well-dened smooth
cocycles. In fact, the weighted Sobolev spaces H
s
q
(M) and W
s
q
(M) are
invariant under the action of the Teichmller ow on the space Q

(M) (al-
though their Hilbert structure is not).
We point out that for any s > 0 there is no natural extension of the distribu-
tional cocycles G
s
t

tR
and
s
t

tR
respectively to the bundles

H
s

(M)
and

W
s

(M), since the action of the Teichmller ow on Q

does not
SOBOLEV REGULARITY OF SOLUTIONS 73
respect the domain of the Friedrichs Laplacian, hence the trivial cocycle
G
t
id is not dened on the bundles D
0


H
s
q
0
(M).
Let I
s
,+
(M)[I
s
,
(M)] H
s

(M) be the nite dimensional sub-bundle of


horizontally [vertically] invariant distributions. By denition, the ber of
the bundle I
s
,+
(M) [I
s
,
(M)] at any q M
(1)

coincides with the vector


space I
s
+q
(M) [I
s
q
(M)] of horizontally [vertically] invariant distributions.
Lemma 4.16. For any s 0 and for any G
t
-invariant ergodic probability
measure on a stratum M
(1)

M
(1)
g
of orientable quadratic differentials,
(1) the sub-bundles I
s
,
(M) H
s

(M) are G
s
t
-invariant, measurable
and of nite, almost everywhere constant rank;
(2) the cocycle G
s
t
[I
s
,
(M) satises the Oseledecs theorem.
Proof. (1) The horizontal and vertical vector elds determined by a qua-
dratic differential q Q

(M) are rescaled under the Teichmller geodesic


ow and the space S
q
(M) of q-tempered currents is invariant. Hence the
spaces I
q
(M) of all horizontally, respectively vertically, invariant distri-
butions are invariant. Since the Sobolev spaces H
s
q
(M) are invariant, the
spaces I
s
q
(M) are also invariant under the Teichmller geodesic ow.
For any q Q

(M), the space I


s
+q
(M) [I
s
q
(M)] can be characterized as
the perpendicular of the closure in H
s
q
(M) of the range of the Lie derivative
L
Sq
[L
Tq
] as a linear operator dened on the space H
s+1
q
(M):
(4.62)
I
s
q
(M) = S
q
u [ u H
s+1
q
(M)

;
[I
s
q
(M) = T
q
u [ u H
s+1
q
(M)

] .
Let q
0
Q

(M) and let f : D


0
Diff
+
(M) be a map dened on a
neighbourhood D
0
Q

(M) of q
0
which trivializes the bundles H
s

(M)
as in formula (4.61). For any xed v H
s+1
q
0
(M) the maps
(4.63) q (f
1
q
)

S
q
f

q
(v) and q (f
1
q
)

T
q
f

q
(v)
are well-dened and continuous on D
0
with values in H
s
q
0
(M). Since the
Hilbert spaces H
s
q
(M) are separable, the sub-spaces
(f
1
q
)

I
s
q
(M) H
s
q
0
(M)
are measurable functions of the quadratic differentials q Q

(M). In fact,
the orthogonal projections I
q
on (f
1
q
)

I
s
q
(M) can be constructed as fol-
lows. Let B
s+1
0
be an orthonormal basis for H
s+1
q
0
(M). If the horizontal
[vertical] foliation of q D
0
is minimal, the subset of H
s
q
(M) dened as
(4.64) B
s
q
:= S
q
f

q
(v)[v B
s+1
0
[ B
s
q
:= T
q
f

q
(v)[v B
s+1
0
]
74 GIOVANNI FORNI

is linerarly independent. Since the set of q D
0
with non-minimal hori-
zontal or vertical foliation has measure zero, the Gram-Schmidt orthonor-
malization algorithm applied to the system (f
1
q
)

B
s
q
[ (f
1
q
)

B
s
q
] yields
an orthonormal basis u
+
k
(q)
kN
[u

k
(q)
kN
] in H
s
q
0
(M) of the subspace
(4.65) (f
1
q
)

S
q
v [ v H
s+1
q
(M) [ (f
1
q
)

T
q
v [ v H
s+1
q
(M) ]
such that, for all k N, the functions u

k
: D
0
H
s
q
0
(M) are dened
-almost everywhere and are continuous on their domain of denition by
the continuity of the maps (4.63). Let D

k
(q)
kN
H
s
q
0
(M) be the (or-
thonormal) system dual to the orthonormal systemu

k
(q)
kN
H
s
q
0
(M).
Then we can write
(4.66) I
q
(D) = D

kN
D, D

k
(q))
s
D

k
(q) , for all D H
s
q
0
(M).
It is immediate to verify that formula (4.66) yields the orthogonal projec-
tions onto the subspaces (f
1
q
)

I
s
q
(M) for any q D
0
. Such projections
can therefore be obtained as a limit of operators which depend continuously
on q Q

(M) in the complement of the subsets of quadratic differential


with non-minimal horizontal [vertical] foliation. Since such a set has mea-
sure zero with respect to any G
t
-invariant ergodic probability measure, the
measurability of the sub-bundles of invariant distributions is proved.
The rank of the bundles I
s
,
(M) is nite by Lemma 3.14 and Theorem
3.21 and it is almost everywhere constant with respect to any ergodic G
t
-
invariant probability measure by denition of ergodicity.
(2) It follows from the Denition 2.7 of the weighted Sobolev norms and
from the fundamental theorem of interpolation (see [LM68], Chap. 1, 5.1)
applied to the interpolation family of Hilbert spaces H
s
q
(M)[s [0, 1]
that, for any s R and for any q Q

(M), the following estimates hold:


(4.67) |G
s
t
: H
s
q
(M) H
s
Gt(q)
(M)| e
|s| |t|
.
Since the bundles I
s
,
(M) are nite dimensional and measurable, the Os-
eledecs theorem applies to the cocycle G
s
t
[I
s
,
(M) with respect to any
ergodic G
t
-invariant probability measure.

The measurable dependence of the spaces of invariant distributions on the


quadratic differentials can also holds for distributional solutions of the co-
homological equation with arbitrary data. In order to formulate the result,
which will be relevant later on, we introduce the following:
SOBOLEV REGULARITY OF SOLUTIONS 75
Denition 4.17. For any s R and any r > 0, the range (in

H
s
q
(M))
of the horizontal [vertical] Lie derivative operator (on

H
r
q
(M)) is the sub-
space
(4.68)
R
r,s
q
(M) := S
q
U

H
s
q
(M) [U

H
r
q
(M)
[R
r,s
q
(M) := T
q
U

H
s
q
(M) [U

H
r
q
(M)]
The Green operators U
r,s
q
: R
r,s
q
(M) H
r
q
(M) are dened as follows:
for any distribution F R
r,s
q
(M) [ F R
r,s
q
(M)] :
(4.69)
U
r,s
q
(F) := U , if S
q
U = F and U I
r
q
(M)


H
r
q
(M) ;
[U
r,s
q
(F) := U , if T
q
U = F and U I
r
q
(M)


H
r
q
(M)] .
By Corollary 3.4, for any s R there exists r > 0 such that, for almost
all quadratic differentials in every circle orbit in a stratum M
(1)

, the Green
operators are bounded operators dened on the codimension 1 subspace

H
s
q
(M)

H
s
q
(M) of distributions vanishing on constant functions.
For any s R, let

H
s

(M)

H
s

(M) be the continuous sub-bundle of


distributions vanishing on constant functions:

H
s

(M) := (q, D)

H
s

(M) [ D, 1)
s
= 0 .
Lemma 4.18. Let be any SO(2, R)-absolutely continous, G
t
-invariant
ergodic probability measure on a stratum M
(1)

of orientable quadratic dif-


ferentials. For any s R, there exists r > 0 such that the the Green op-
erator U
r,s
q
yield a measurable bundle map U
r,s
,
:

H
s

(M) H
r

(M).
In particular, the operator norm |U
r,s
q
| yields a well-dened measurable
real-valued function of the quadratic differential q M
(1)

.
Proof. Let q
0
Q

(M) and let f : D


0
Diff
+
(M) be a map dened on a
neighbourhood D
0
Q

(M) of q
0
which trivializes the bundles

H
s

(M)
as in formula (4.61). The argument is similar to the one given in the proof
of the measurability of the sub-bundles of invariant distributions in Lemma
4.16. Let B
r+1
0
H

q
0
(M) be a basis for the Hilbert space H
r+1
q
0
(M). The
subset of H
r
q
(M) dened as
(4.70) B
r
q
:= S
q
f

q
(v)[v B
r+1
0
[ B
r
q
:= T
q
f

q
(v)[v B
r+1
0
]
is linerarly independent if the horizontal [vertical] foliation of q D
0
is
minimal, Hence, by the continuity of the maps (4.63), the Gram-Schmidt
orthonormalization applied to the system B
r
q
[ B
r
q
] yields an orthonormal
basis u
+
k
(q)
kN
[u

k
(q)
kN
] in H
r
q
(M) of the subspace
(4.71) S
q
v [ v H
r+1
q
(M) [T
q
v [ v H
r+1
q
(M) ]
76 GIOVANNI FORNI

such that, for all k N, the functions (f
1
)

k
: D
0
H
r
q
0
(M) are
dened -almost everywhere and continuous on their domain of deni-
tion. In fact, there exists bases v

kN
H

q
0
(M) of the Hilbert space
H
r+1
q
0
(M) such that the following holds. For each k N there exists a
function v

k
: D
0
span(v

1
, . . . , v

k
), dened -almost everywhere and
continuous on its domain of denition, such that for -almost all q D
0
the set v

k
(q)
kN
is a basis of the Hilbert space H
r+1
q
0
(M) and
u
+
k
(q) = S
q
(f

v
+
k
)(q) [u

k
(q) = T
q
(f

k
)(q)] .
Let D

k
(q)
kN
H
r
q
0
(M) be the system dual to the linearly independent
system (f
1
q
)

k
(q)
kN
H
r
q
0
(M). Such systems are in general not
orthonormal. Then for any F

H
s
q
0
(M), the following formula holds:
(4.72) (f
1
q
)

U
r,s
q
f

q
(F) :=

kN
F, v

k
(q)) D

k
(q) .
In fact, let U

k
(q) be the series on the right hand side of formula (4.72).
Since f

q
D

k
(q), u

k
(q)) =
kh
for any k, h N, it follows that
(4.73)
f

q
U
+
k
(q), S
q
f

q
v
+
k
(q)) = F, v
+
k
(q)) = f

q
(F), f

q
v
+
k
(q))
[f

q
U

k
(q), T
q
f

q
v

k
(q)) = F, v

k
(q)) = f

q
(F), f

q
v

k
(q))]
which implies that f

q
U
+
k
(q) [f

q
U

k
(q)] is a distributional solution of the
equation S
q
u = f

q
(F) [T
q
u = f

q
(F)]. Finally, since all the distributions
f

q
D
+
k
(q) [f

q
D

k
(q)] vanish on the orthogonal complement of the space
S
q
v [ v H
r+1
q
(M) [T
q
v [ v H
r+1
q
(M)] in H
r
q
(M), it follows that
f

q
U

k
(q) is orthogonal to the space I
r
q
(M) of invariant distributions, hence
U
r,s
q
f

q
(F) = f

q
U

k
(q). It is also immediate from formula (4.72) that, for
any F

H
s
q
0
(M), the distribution (f
1
q
)

U
r,s
q
f

q
(F) is a -measurable
function (dened almost everywhere) of the quadratic differential q D
0
with values in the Hilbert space H
r
q
0
(M). In fact, by construction the func-
tions D

k
: D
0
H
r
q
0
(M) are dened -almost everywhere and continu-
ous on their domain of denition.

Let Z
s

(M) W
s

(M) be the innite dimensional sub-bundle of closed


currents over M
(1)

. By denition, the ber of the bundle Z


s

(M) at any
q M
(1)

coincides with the vector space of closed currents:


(4.74) Z
s
q
(M) := Z
q
(M) W
s
q
(M) .
The bundle Z
s

(M) and the sub-bundle E


s

(M) Z
s

(M) of exact currents


are smooth,
s
t
-invariant sub-bundles of the bundle W
s
q
(M). The quotient
cocycle, dened on the H
s
de Rham cohomology bundle, is isomorphic
SOBOLEV REGULARITY OF SOLUTIONS 77
to the Kontsevich-Zorich cocycle. The latter isomorphism is the essential
motivation for the formulas (4.58) and (4.60) which dene, respectively, the
cocycles G
s
t

tR
and
s
t

tR
. In fact, let
(4.75) j

: Z
s

(M) H
1

(M, R)
be the natural de Rham cohomology map onto the cohomology bundle
H
1

(M, R), dened as the restriction to the stratum M


(1)

of the cohomol-
ogy bundle H
1
g
(M, R) introduced in formula (4.8).
Let B
s
,
(M) Z
s

(M) be the sub-bundles with ber at q M


(1)

given by
the vector spaces B
s
q
(M) of F
q
-basic currents (dened in (3.31)).
Lemma 4.19. For any s 0 and for any G
t
-invariant ergodic probability
measure on a stratum M
(1)

M
(1)
g
of orientable quadratic differentials,
(1) the identity j


s
t
=
t
j

holds everywhere on Z
s

(M);
(2) the sub-bundles B
s
,
(M) Z
s

(M) are
s
t
-invariant, measurable
and of nite, almost everywhere constant rank;
(3) the cocycle
s
t
[B
s
,
(M) satises the Oseledecs theorem.
Proof. (1) It is an immediate consequence of the denition of the cocycle

s
t

tR
on the bundle Z
s

(M) of closed currents.


(2) The horizontal and vertical measured foliations of a quadratic differen-
tial q Q

(M) are projectively invariant under the Teichmller geodesic


ow and the space S
q
(M) of q-tempered currents is invariant. As a conse-
quence, the spaces B
+q
(M) and B
q
(M) of horizontally, respectively ver-
tically, basic currents are invariant. The Sobolev spaces W
s
q
(M) are also
invariant (although the Hilbert structure is not). Hence the spaces B
s
q
(M)
are invariant under the Teichmller geodesic ow.
By Lemma 3.14, the measurability of the bundles B
s
,
(M) W
s

(M) of
basic currents is equivalent to the measurability of the bundles I
s
,
(M)
H
s
q
(M), proved in Lemma 4.16.
The rank of the bundles B
s
,
(M) is nite by Theorem 3.21 and it is almost
everywhere constant with respect to any ergodic G
t
-invariant probability
measure by denition of ergodicity.
(3) It follows immediately from the identities (4.60) and from the bound
(4.67) that, for any s R and for any q Q

(M), the following estimates


hold:
(4.76) |
s
t
: W
s
q
(M) W
s
Gt(q)
(M)| e
(|s|+1)|t|
.
Since the bundles B
s
,
(M) are nite dimensional and measurable, the Os-
eledecs theorem applies to the cocycle
s
t
[B
s
,
(M) with respect to any
ergodic G
t
-invariant probability measure.
78 GIOVANNI FORNI

Lemma 4.16 can be generalized to the bundle of quasi-invariant distribu-
tions. Let I
s
,
(M

) H
s

(M) be the bundle over M


(1)

dened as fol-
lows: its ber at each q M
(1)

is the vector space I


s
q
(M
q
) H
s
q
(M)
of quasi-invariant distributions. An argument analogous to the proof of
Lemma 4.16 proves the following:
Lemma 4.20. For any s 0 and for any G
t
-invariant ergodic probability
measure on a stratum M
(1)

M
(1)
g
of orientable quadratic differentials,
(1) the sub-bundles I
s
,
(M

) H
s

(M) are G
s
t
-invariant, mea-
surable and of nite, almost everywhere constant rank;
(2) the cocycle G
s
t
[I
s
,
(M

) satises the Oseledecs theorem.


Lemma 4.19 can be generalized to the bundle of currents closed on the com-
plement of the singular set and to the sub-bundle of quasi-basic currents.
Let Z
s

(M

) W
s

(M) be the bundle over M


(1)

dened as follows:
its ber at each q M
(1)

is the vector space Z


s
(M
q
) W
s
q
(M) of
closed currents on the open manifold M
q
. The bundle Z
s

(M

) and
the sub-bundle E
s

(M

) Z
s

(M

) of exact currents are smooth,

s
t
-invariant sub-bundles of the bundle W
s
q
(M). The quotient cocycle, de-
ned on the H
s
de Rham punctured cohomology bundle, is isomorphic to
the punctured Kontsevich-Zorich cocycle. Let
(4.77) j

: Z
s

(M

) H
1

(M

, R)
be the natural de Rham cohomology map onto the punctured cohomology
bundle H
1

(M

, R) introduced in formula (4.12).


Let B
s
,
(M

) Z
s

(M

) be the sub-bundles with ber at q M


(1)

given by the vector spaces B


s
q
(M
q
) of quasi-basic currents for the
measured foliation F
q
(dened in (3.31)). An argument analogous to the
proof of Lemma 4.19 proves the following:
Lemma 4.21. For any s 0 and for any G
t
-invariant ergodic probability
measure on a stratum M
(1)

M
(1)
g
of orientable quadratic differentials,
(1) the identity j


s
t
=
t
j

holds everywhere on Z
s

(M

);
(2) the sub-bundles B
s
,
(M

) Z
s

(M

) are
s
t
-invariant,
measurable and of nite, almost everywhere constant rank;
(3) the cocycle
s
t
[B
s
,
(M

)
tR
satises the Oseledecs theorem.
It follows immediately from the denitions and from Lemma 3.14 that the
following cocycle isomorphisms hold:
(4.78)

s
t
[B
s
,
(M) e
t
G
s
t
[I
s
,
(M) ;

s
t
[B
s
,
(M

) e
t
G
s
t
[I
s
,
(M

) .
SOBOLEV REGULARITY OF SOLUTIONS 79
As a consequence of such an isomorphism parts (2) and (3) in Lemma 4.19
and Lemma 4.21 can be immediately derived fromLemma 4.16 and Lemma
4.20 respectively. In addition, the Lyapunov spectra and the Oseledecs
decomposition of the cocycles
s
t
[B
s
,
(M) [
s
t
[B
s
,
(M

)] and
G
s
t
[I
s
,
(M) [G
s
t
[I
s
,
(M

)] can be immediately derived from one


another. By part (1) in Lemma 4.19 and Lemma 4.21, by Corollary 3.20
and by the structure theorem for basic currents (Theorem 3.21), informa-
tions on the Lyapunov spectrum of the cocycles
s
t
[B
s
,
(M

) and

s
t
[B
s
,
(M) can be derived from that of the (punctured) Kontsevich-
Zorich cocycle.
For s = 1, by [For02], Lemma 8.1, and by the representation Lemma 4.14,
in the non-uniformly hyperbolic case it is possible to give a quite complete
description of the distributional cocycle on the bundle of basic currents.
Lemma 4.22. Let be any SO(2, R)-absolutely continuous, KZ-hyperbolic
measure on a stratum M
(1)

of orientable quadratic differentials.


(1) The cocycle
1
t

tR
has strictly positive [strictly negative] Lya-
punov spectrum, with respect to the measure on M
(1)

, on the in-
variant sub-bundle B
1
,+
(M) [B
1
,
(M)].
(2) The sum B
1

(M) := B
1
,+
(M) + B
1
,
(M) of bundles over M
(1)

is
direct and the restriction of the map j

to the sub-bundle B
1

(M) is
-almost everywhere injective.
(3) The coycle
1
t
[B
1

(M)
tR
is isomorphic to the Kontsevich-Zorich
cocycle
t

tR
on the real cohomology bundle H
1

(M, R), hence


it has the same Lyapunov spectrum.
4.3. Lyapunov exponents. We prove below that the the spaces of all hor-
izontally and vertically (quasi)-basic currents and invariant distributions
have well-dened Oseledec decompositions at almost all point of any stra-
tum of the moduli space (with respect to any G
t
-invariant ergodic measure).
We establish a fundamental relation between the Lyapunov exponents and
the Sobolev order of basic currents or distributions in each Oseledec sub-
space. We conclude with a crucial spectral gap result for the distributional
cocycle on the bundle of exact currents which is the basis for sharp es-
timates on the growth of ergodic averages, hence for the construction of
square-integrable solutions of the cohomological equation in 5.2.
We introduce the following:
Denition 4.23. A non-zero current C

B
s
,
(M

) will be called
(Oseledec) simple if it belongs to an Oseledec subspace of the cocycle

s
t
[B
s
,
(M

). A non-zero invariant distributionD

I
s

(M

)
80 GIOVANNI FORNI

will be called (Oseledec) simple if it belongs to an Oseledec subspace of
the cocycle G
s
t
[I
s

(M

).
The above denition is well-posed since for any r s the cocycles
r
t

[G
r
t
] are the restrictions of the cocycles
s
t
[G
s
t
]. It is also immedi-
ate to prove that the image of any Oseledec simple basic current under the
isomorphism D
q
: B
s
q
(M
q
) I
s
q
(M
q
) introduced in formula
(3.34) is an Oseledec simple invariant distribution.
Lemma 4.24. Let be any G
t
-invariant probability measure on M
(1)

and
let R

M
(1)

be the set of all holomorphic orientable quadratic differ-


entials which, for all s 0, are Oseledec regular point for the cocy-
cles
s
t
[B
s
,+
(M

) and
s
t
[B
s
,
(M

) over the Teichmller


ow (G
t
, ). The set R

has full measure and there exist measurable


functions L

: B
,
(M

) R (with almost everywhere constant


range) such that, for any q R

, the number L

(C

) is equal to the
Lyapunov exponent of the Oseledec simple current C

B
q
(M
q
)
of Sobolev order O
W
q
(C

) 0 (see Denition 3.12), with respect to the


cocycle
s
t
[B
s
,
(M

) over the ow (G
t
, ) for any s > O
W
q
(C

).
Proof. For any r s, the embeddings B
r
,
(M

) B
s
,
(M

) hold
and the cocycle
r
t

tR
on the bundle W
r

(M) coincides with restriction


of the cocycle
s
t

tR
, dened on the bundle W
s

(M). The Oseledecs


theorem holds for the cocycles
s
t
[B
s
,
(M

)
tR
on the measurable,
nite dimensional sub-bundles B
s
,
(M

) W
s

(M) for all s 0.


Let R

M
(1)

be the set of points which are Oseledec regular for the co-
cycles
k
t
[B
k
,
(M

)
tR
for all k N. By the Oseledec theorem the
set R

has full measure and, for any q R

and for all s r, the Lya-


punov exponents of any Oseledec simple current C

B
r
,
(M

) with
respect to the cocycles
r
t

tR
and
s
t

tR
are well-dened and coincide.
The common value L

(C

) of all Lyapunov exponents with respect to the


cocycles
s
t

tR
for s > O
W
q
(C

) is therefore well-dened.
By the isomorphisms (4.78) an analogous statement holds for the cocycles
on the bundles of invariant distributions:
Lemma 4.25. Let be a G
t
-invariant probability measure on M
(1)

. There
exist measurable functions l

: I
,
(M

) R(with almost everywhere


constant range) such that, for any quadratic differential q R

, the number
l

(D

) is equal to the Lyapunov exponent of the Oseledec simple invariant


distribution D

I
q
(M
q
) of Sobolev order O
H
q
(D

) 0 (see Def-
inition 2.8), with respect to the cocycle G
s
t
[I
s
,
(M

) over the ow
(G
t
, ) for any s > O
H
q
(D

). Let D
q
: B
q
(M
q
) I
q
(M
q
)
SOBOLEV REGULARITY OF SOLUTIONS 81
be the isomorphism introduced in formula (3.34), the following identities
hold:
(4.79) l

D
q
= L

1 , on B
q
(M
q
) .
Lyapunov exponents of basic currents impose restrictions on their Sobolev
regularity. In fact, we have:
Lemma 4.26. The following inequalities hold for any quadratic differential
q R

M
(1)

. For any Oseledec simple basic current C

B
q
(M
q
)
and any Oseledec simple invariant distribution D

I
q
(M
q
) the
Sobolev order functions satisfy the following lower bounds:
(4.80) O
W
q
(C

) [L

(C

) 1[ and O
H
q
(D

) [l

(D

)[ .
Proof. Since C

B
q
(M
q
), by the identities (4.60) and by the bound
(4.67) the following inequalities hold for any s r:
(4.81) [
s
t
(C

)[
s
[
r
t
(C

)[
r
e
r|t|t
[C

[
r
, for all t R.
On the other hand, by the Oseledecs theorem, for any q R

and any > 0


there exists a constant K

(q) > 0 such that


(4.82)
[
s
t
(C

)[
s
K

(q) e
(L

(C

)) t
[C

[
s
, for t 0 ;
[
s
t
(C

)[
s
K

(q) e
(L

(C

)+) t
[C

[
s
, for t < 0 ;
A comparison of the inequalities (4.81) and (4.82) proves the statement for
the case of basic currents. The statement for invariant distributions follows
immediately by Lemma 4.25 since the following identity holds:
O
H
q
D
q
= O
W
q
on B
q
(M
q
) .

A similar argument, based on the spectral gap for almost all quadratic dif-
ferentials in every circle orbit (see Theorem 4.11), proves the following:
Theorem 4.27. For any quadratic differential q M
(1)

there exists a real


number s(q) > 0 such that, for almost all S
1
and for all s < s(q)
(4.83)
dimB
s
q

(M
q
) = dimB
s
q

(M
q
) = 1 ;
dimI
s
q

(M
q
) = dimI
s
q

(M
q
) = 1 .
Proof. If there exists s (0, 1) such that (4.83) does not hold on a positive
measure subset of the circle, it follows that there exists a positive measure
set E
s
S
1
such that dimB
s
q

(M
q
) > 1. Hence, for any E
s
, there
exists a non-vanishing vertically (quasi)-basic current C

B
s
q

(M
q
)
such that C

q
1/2

= 0. We claim that for almost all E


s
, the current C

has non-zero cohomology class c

H
1
(M
q
, R). In fact, by Theorem
82 GIOVANNI FORNI

3.21, if C

has vanishing cohomology class, then there exists a basic current

B
s1
q

(M
q
) such that
1
q

(

C

) = C

. Since s < 1, if the vertical


foliation F
q

is ergodic, then the current



C

R (q

) which implies
C

= 0, a contradiction. Since by the Keane conjecture (see [Mas82] or


[Vee82]) the vertical foliation F
q

is uniquely ergodic for almost all S


1
,
the claim is proved. As in formula (4.81), we have
(4.84) [
s
t
(C

)[
s
e
s|t|t
[C

[
s
, for all t R.
Let
t

tR
be the punctured Kontsevich-Zorich cocycle, introduced in
4.1. By Lemma 4.19 the following identity holds:
[
s
t
(C

)] =
t
(c

) H
1
(M
q
, R) , for all t R.
The de Rham punctured cohomology bundle H
1

(M

, R) can be en-
dowed, for any s 0, with the quotient norm [ [
s

induced by the Sobolev


norm on the bundle W
s

(M). It follows by the above discussion and by


the estimate (4.84) that, for each E
s
, there exists a non-vanishing coho-
mology class c

H
1
q

(M
q
, R) such that
(4.85)
c

[(q
1/2

)] = c

[(q
1/2

)] = 0 ,
limsup
t+
1
[t[
log [
t
(c

)[
s

s 1 .
By Lemma 4.3 and by the Poincar-Lefschetz duality (4.13) between rel-
ative and punctured cohomology, the cohomology bundle H
1

(M, R) ad-
mits a complement in the punctured cohomology bundle H
1

(M

, R) on
which the punctured Kontsevich-Zorich cocycle is isometric (with respect
to a continuous norm). By the estimate in (4.85) on the upper Lyapunov ex-
ponent, it follows that, since s < 1, the cohomology class c

H
1
q

(M, R)
for all E
s
. The restriction of the punctured Kontsevich-Zorich cocycle

tR
to the bundle H
1

(M, R) concides with the Kontsevich-Zorich co-


cycle
t

tR
. Since the latter is symplectic, it follows from (4.85) that
its upper Lyapunov exponent on the symplectic subspace I

(M, R)
H
1
q

(M, R) (see formulas (4.26) and (4.27)) satises the estimate:


(4.86)
+
2
(q

) 1 s , for all E
s
.
By Theorem 4.11 the inequality
+
2
(q

) L

(q) < 1 holds for all q M


(1)

and for almost all S


1
. Since the set E
s
has positive measure, it follows
that s > 1 L

(q) > 0.

By Lemma 4.14, Lemma 4.22 and Lemma 4.26 we derive a fundamental


relation between the weighted Sobolev order and the Lyapunov exponents
of Oseledec simple basic currents and invariant distributions.
SOBOLEV REGULARITY OF SOLUTIONS 83
Denition 4.28. An Oseledec basis for the space I
s
q
(M) [ B
s
q
(M)] of
invariant distributions [of basic currents] is a basis contained in the union of
all Oseledec subspaces for the cocycle G
s
t
[I
s
,
(M) [
s
t
[B
s
,
(M)] at
any Oseledec regular point q M
(1)

. Since the cocycles


s
t
[I
s
,
(M) and

s
t
[B
s
,
(M) are isomorphic via the maps D
q
: B
s
q
(M) I
s
q
(M)
introduced in (3.34), any basis
s
q
I
s
q
(M) is Oseledec if and only if the
basis D
q
(
s
q
) B
s
q
(M) is Oseledec as well.
Theorem 4.29. For any SO(2, R)-absolutely continuous, KZ-hyperbolic
measure on a stratum M
(1)

of orientable quadratic differentials, for -


almost all q M
(1)

and for any Oseledec basis C

1
, . . . , C

g
B
1
q
(M):
(4.87) O
W
q
(C

i
) = 1 [L

(C

i
)[ = 1

i
, for all i 1, . . . , g .
Consequently, for any Oseledec basis D

1
, . . . , D

g
I
1
q
(M),
(4.88) O
H
q
(D

i
) = [l

(D

i
)[ = 1

i
, for all i 1, . . . , g .
The restriction of the cocycle
1
t

tR
to the invariant sub-bundle Z
1

(M) is
described by the following Oseledec-type theorem, which is a straightfor-
ward generalization of [For02], Theorem 8.7, based on Lemma 4.22.
Theorem 4.30. For any SO(2, R)-absolutely continuous, KZ-hyperbolic
measure on a stratumM
(1)

of orientable quadratic differentials, there exists


a measurable
1
t
-invariant splitting:
(4.89) Z
1

(M) = B
1
,+
(M) B
1
,
(M) E
1

(M) .
(1) The restriction of the cocycle
1
t

tR
to the bundle B
1

(M) :=
B
1
,+
(M)B
1
,
(M) is (measurably) isomorphic to the Kontsevich-
Zorich cocycle, hence it has Lyapunov spectrum (4.15). The sub-
bundle B
1
,+
(M) corresponds to the strictly positive exponents and
the sub-bundle B
1
,
(M) to the strictly negative exponents.
(2) The Lyapunov spectrum of restriction of the cocycle
1
t

tR
to the
innite dimensional bundle E
1

(M) of exact currents is reduced to


the single Lyapunov exponents 0 (in fact, the cocycle is isometric
with respect to a suitable continuous norm on E
1

(M)).
Informations on the Lyapunov structure of the restrictions of the cocycles

s
t

tR
to the sub-bundles B
s
,
(M) B
s
,
(M

) W
s

(M), for any


s 1, can be derived from Theorem 4.30 and from the structure theorem
for basic currents (see Theorem 3.21) combined with the following result:
Lemma 4.31. Let be any G
t
-invariant ergodic probability measure on
a stratum M
(1)

of orientable quadratic differentials. For any q R

, the
image under the map
q
: B
q
(M
q
) B
q
(M
q
) dened by
84 GIOVANNI FORNI

formulas (3.61) of any simple current C

B
s
q
(M
q
) is a simple
current
q
(C

) B
s+1
q
(M
q
). The following identities hold for the
Sobolev order map and the Lyapunov exponent map on B
q
(M
q
):
(4.90)
O
W
q

q
= O
W
q
1 , for all q M
(1)

;
L


q
= L

1 , for all q R

M
(1)

.
Proof. For any C

B
q
(M
q
), it follows immediately by the de-
nitions of the Sobolev spaces and of the maps
q
on B
q
(M
q
) that

q
(C
+
) W
s1
q
(M) if and only if C
+
(q
1/2
) H
s
q
(M), hence if and
only if C
+
W
s
q
(M). Similarly,
q
(C

) W
s1
q
(M) if and only if
C

(q
1/2
) H
s
q
(M), hence if and only if C

W
s
q
(M). The rst
identity in (4.90) is therefore proved. In fact, we have proved that the maps

s
q
send B
s
q
(M
q
) onto the subspace of cohomologically trivial currents
in B
s+1
q
(M
q
) while the space B
s
q
(M) is mapped onto the subspace of
cohomologically trivial currents in B
s+1
q
(M).
The following identity follows immediately from the denitions: for any
quadratic differential q M
(1)

, for each s 0 and all t R, we have


(4.91)
s
Gt(q)

s
t
= e
t
(
s+1
t

s
q
) on B
s
q
(M
q
) .
Since the maps
s
q
: B
s
q
(M
q
) B
s+1
q
(M
q
) are embeddings, it
follows from the identity (4.91) and from the Oseledecs theorem that the
current
s
q
(C

) B
s+1
q
(M
q
) is simple if C

B
s
q
(M
q
) is. It
also follows that the second indentity in (4.90) holds.
An analogous statement for invariant distributions can be derived. For any
q M
(1)

, let L
q
: S
q
(M) S
q
(M) denote the Lie derivative operators
on the space S
q
(M) of all q-tempered currents: for any C S
q
(M),
(4.92) L
q
(C) := L
Sq
(C) and L
q
(C) := L
Tq
(C) .
Lemma 4.32. For any q M
(1)

, the operators L
q
: I
q
(M) I
q
(M)
are well-dened and injective. Let be any G
t
-invariant ergodic probabil-
ity measure on M
(1)

. The following identities hold on the spaces I


q
(M):
(4.93)
O
H
q
L
q
= O
H
q
1 , for all q M
(1)

;
l

L
q
= l

1 , for all q R

M
(1)

.
Corollary 4.33. Let be any G
t
-invariant ergodic probability measure on
a stratum M
(1)

of orientable holomorphic quadratic differentials. For any


s 0 the (nite) Lyapunov spectrum of the cocycle
s
t
[B
s
,
(M) is a
nite subset of the countable set
(4.94)

i
j [1 < i < 2g, j N ;
SOBOLEV REGULARITY OF SOLUTIONS 85
hence the Lyapunov spectrum of the cocycle G
s
t
[I
s
,
(M) is a nite subset
of the countable set
(4.95) 0

i
(j + 1) [1 < i < 2g, j N
(each element of the sets (4.94) and (4.95) is taken with multiplicity one).
Proof. Let
1
,+
(M, R) [
1
,
(M, R)] be the continuous sub-bundles of the
cohomology bundle H
1

(M, R) over M
(1)

which bers at any quadratic dif-


ferential q M
(1)

are given by the spaces


1
+q
(M, R) [
1
q
(M, R)] de-
ned in formula (3.52). The sub-bundles
1
,
(M, R) are invariant under
the Kontsevich-Zorich cocycle
t

tR
and the Lyapunov spectrum of the
restriction
t
[
1
,
(M, R) consists of the set

i
[1 i 2g 1. By
formulas (3.60), the image of the bundle B
s
,
(M) under the cohomology
map j

: Z
1

(M) H
1

(M, R) is a sub-bundle of
1
,
(M, R) which is
invariant under the Kontsevich-Zorich cocycle. By Lemma 4.19 the cocy-
cle
s
t
[B
s
,
(M) is mapped under the cohomology map onto a restriction
of the Kontsevich-Zorich cocycle. Let
s
,
: B
s1
,
(M) B
s
,
(M) the
measurable bundle maps dened ber-wise for -almost all q M
(1)

as the
maps
s
q
: B
s1
q
(M) B
s
q
(M), dened in formula (3.61). The kernel
of the cohomology map j

on B
s
,
(M) is a
s
t
-invariant sub-bundle which
coincides with the range of the map
s
,
. By Lemma 4.31, the Lyapunov
spectrum (4.94) of
s
t
[B
s
,
(M) can therefore be derived by induction on
[s] N. The Lyapunov spectrum of (4.95) of G
s
t
[I
s
,
(M) can then be
derived by the isomorphism (4.78).
By Corollary 3.20 for any s > 3 the cohomology map is surjective for
almost all quadratic differentials in any circle orbit. The above result can be
rened as follows:
Corollary 4.34. Let be any SO(2, R)-absolutely continous, G
t
-invariant
ergodic probability measure on a stratum M
(1)

M
(1)
g
of orientable holo-
morphic quadratic differentials. For any s > 3, there exists an integer
vector h
s
:= (h
s
2
, . . . , h
s
2g1
) N
2g2
such that the Lyapunov spectrum of
the cocycle
s
t
[B
s
,
(M) is the (nite) set
(4.96)

i
j [1 < i < 2g, 0 j h
s
i
;
hence the Lyapunov spectrum of the cocycle G
s
t
[I
s
,
(M) is the (nite) set
(4.97) 0

i
(j + 1) [1 < i < 2g, 0 j h
s
i

(each element of the sets (4.96) and (4.97) is taken with multiplicity one).
The integer vector h
s
depends on the Sobolev regularity of basic currents in
the Oseledecs spaces related to the Lyapunov exponents

2
, . . . ,

2g1

86 GIOVANNI FORNI

which come from the Kontsevich-Zorich cocycle. Hence, in particular, in
case is the unique absolutely continuous, SL(2, R)-invariant ergodic mea-
sure on a connected component of a stratum, by Theorem4.29 the following
estimate for the numbers (h
s
2
, . . . , h
g
) holds: for all i 2, . . . , g,
h
s
i
= maxh[1

i
+ h < s , if s , N

i
.
The Oseledec-Pesin theory of the cocycles G
s
t
[I
s
,
(M) has crucial im-
plications for the theory of the cohomological equation. In particular, it
implies that (if is a KZ-hyperbolic measure) for -almost all q M
(1)

,
Oseledec bases of the spaces I
s
q
(M) of invariant distributions have strong,
quantitative linear independence properties.
Theorem 4.35. Let be any G
t
-invariant ergodic probability measure on
a stratum M
(1)

of orientable holomorphic quadratic differentials. For any


s > 0 and for -almost all q M
(1)

, let b
s
q
= D

1
, . . . , D

(s)
be a
Oseledec basis of the space I
s
q
(M) of invariant distributions and let
(4.98) (l

1
, . . . , l

(s)
) :=
_
l

(D

1
), . . . , l

(D

(s)
)
_
denote the Lyapunov spectrumof the cocycle G
s
t
[I
s
,
(M) over (G
t
, ).
For any > 0, there exists a measurable function K
s

: M
(1)

R
+
such
that the following holds. For every (0, 1], there exist linearly indepen-
dent systems of smooth functions u

1
(), . . . , u

(s)
() H

q
(M) such
that, for all i, j 1, . . . , J

(s) and for all 0 r s,


(4.99) D

i
_
u

j
()
_
=
ij
and [u

j
()[
r
K

(q)
|l

j
|r
.
Proof. By Corollary 4.33, for any G
t
-invariant ergodic probability measure
on M
(1)

and any s > 0, the Lyapunov exponents of the distributional


cocycle G
s
t
[I
s
,
(M) are all of the same sign, namely
(4.100) l

1
, . . . , l

(s)
R
+
0 .
It follows from the Oseledecs theorem that, for any > 0, there exists a
strictly positive measurable function C
(1)

: M
(1)

R
+
such that, for every
i 1, . . . , J

(s), for -almost all q M


(1)

and for all t 0,


(4.101) [G
s
t
(D

i
)[
s
C
(1)

(q) e
(|l

i
|)t
.
It also follows from the Oseledecs theorem that for any s > 0, for -almost
all q M
(1)

and any Oseledec basis b


s
q
I
s
q
(M), the distorsion of the
(Oseledec) basis G
s
t
(b
s
q
) H
s
Gt(q)
(M) grows subexponentially in time.
SOBOLEV REGULARITY OF SOLUTIONS 87
The distorsion of a basis b
s
q
:= D

1
, . . . , D

J
I
s
q
(M) is the number
(4.102) d
q
(b
s
q
) := sup

(s)
i=1
[c
i
[ [D

i
[
H
s
q
(M)
[

(s)
i=1
c
i
D

i
[
H
s
q
(M)
[ c C
J

(s)
.
The Oseledecs theorem implies that, if b
s
q
I
s
q
(M) is an Oseledec basis,
then for any > 0, there exists a measurable function C
(2)

: M
(1)

R
+
such that for -almost all q M
(1)

and for all t R,


(4.103) d
Gt(q)
_
G
s
t
(b
s
q
)
_
C
(2)

(q) e
|t|
.
By the estimates (4.101) and (4.103), it follows that, for any > 0, there
exists a measurable function C
(3)

: M
(1)

R
+
such that the following
holds. For every t 0, since H

q
(M) is dense in H
s
q
(M) for any s R,
there exists a system v

1
(t), . . . , v

(s)
(t) H
s
q
(M) such that, for -
almost all q M
(1)

and for all i, j 1, . . . , J

(s),
(4.104) D

i
_
v

j
(t)
_
=
ij
and [v

j
(t)[
H
s
G
t
(q)
(M)
C
(3)

(q) e
(|l

j
|)t
.
By the bound (4.67) on the norm |G
s
t
: H
s
q
(M) H
s
Gt(q)
(M)|, it follows
that, for any 0 r s and for all j 1, . . . , J

(s),
(4.105) [v

j
(t)[
H
r
q
(M)
C
(3)

(q) e
(r|l

j
|+)t
, for any t 0 .
It follows that the system u

1
(), . . . , u

(s)
() H

q
(M), dened for
any (0, 1] by the identities
(4.106) u

j
() := v

j
(log ) , j 1, . . . , J

(s) ,
satisises the required properties (4.99).
The following theorem, which can be interpreted as spectral gap result for
the cocycles G
0
t

tR
on the bundle H
0

(M) of square-integrable functions,


is the main technical result of the paper.
For any (, l) R
2
, we introduce the upper (forward) Lyapunov norm of
a distribution U H

q
(M) at a quadratic differential q M
(1)

as the
non-negative extended real number
(4.107) N
,l
q
(U) := sup
t+
e
lt
[G

t
(U)[

Theorem 4.36. Let be any SO(2, R)-absolutely continous, G


t
-invariant
ergodic probability measure on a stratum M
(1)

of orientable quadratic dif-


ferentials. For any any > 0 and any l < 1, there exist a real number
88 GIOVANNI FORNI

:= (, l) > 0 and a measurable function C
,l
: M
(1)

R
+
such that, for
-almost all q M
(1)

and all functions U L


2
q
(M) of zero average,
(4.108) N
,
q
(U) < C
,l
(q) [U[
0
+ N
1,l
q
(S
q
U) .
Proof. The outline of the argument is as follows. From the results on
the cohomological equation and on the Lyapunov spectrum of the coycles
G
r
t

tR
on the bundles of invariant distributions we will derive (exponen-
tial) estimates on the norms [G
r
t
(U)[
r
for a sufciently large r > 0, along a
sequence of suitable visiting times. Since [G
0
t
(U)[
0
is constant, as a conse-
quence of the invariance of the L
2
q
(M) norm under the Teichmller ow, the
interpolation inequality for dual weighted Sobolev norms (Corollary 2.26)
implies the required (exponential) estimates on [G

t
(U)[

for any > 0.


Let q M
(1)

and U L
2
q
(M) be a function of zero average. Let us as-
sume that there exist l
0
< 1 and N R
+
such that the Lyapunov norm
N
1,l
0
q
(S
q
U) N. By denition (4.107) of Lyapunov norms, for any t 0,
(4.109) [G
1
t
(S
q
U)[
1
Ne
l
0
t
.
By the spectral gap Theorem 4.9 and by Corollary 4.33, there exist C
1
> 0,
l
1
> 0 and a positive measure set P
1
M
(1)

such that, for all q P


1
and
all S
q
-invariant distributions D H
r
q
(M) I
r
q
(M), the following holds:
(4.110) [G
r
t
(D)[
r
C
1
e
l
1
t
[D[
r
.
By Lemma 4.18 on the measurability of the Green operators for the coho-
mological equation, there exists r > 0 such that the following holds. There
exists a constant C
2
> 0 and a set P
2
P
1
of positive measure such that
(4.111) |U
r,1
q
: H
1
q
(M) H
r
q
(M)| C
2
, for all q P
2
.
For -almost all q M
(1)

, there exists a sequence (t


n
)
nN
of visiting times
of the forward orbit G
t
(q) [ t 0 to the positive measure set P
2
such that,
for each n N, the function t
n
: M
(1)

R
+
is measurable and
(4.112) lim
n+
t
n
(q)
n
:= p >
log C
1
l
1
.
For each n N and any r R, let us introduce the notations:
G
n
:= G
tn
, q
n
:= G
n
(q) , S
n
, T
n
:= S
qn
, T
qn
,
G
r
n
:= G
r
tn
U
r
n
:= G
r
n
(U) , F
n
:= G
1
n
(S
q
U) ,
H
r
n
(M) := H
r
qn
(M) , H
r
n
(M) := H
r
qn
(M) , I
r
n
(M) := I
r
qn
(M) .
By the denition of the Teichmller ow G
t
, the orthonormal frame
S
n
, T
n
= e
tn
S
q
, e
tn
T
q
for all n N. Hence the following coho-
mological equation holds:
(4.113) S
n
U
r
n
= e
tn
F
n
H
1
n
(M) .
SOBOLEV REGULARITY OF SOLUTIONS 89
For each n N and any r R let [ [
r,n
denote the weighted Sobolev norm
on the space H
r
n
(M). It follows from equation (4.109) that
(4.114) [S
n
U
r
n
[
1,n
Ne
(1l
0
)tn
.
By the denition of the Green operators and by equation (4.111), there ex-
ists a solution G
n
H
r
n
(M), orthogonal to I
r
n
(M), of the cohomological
equation S
n
G
n
= S
n
U
r
n
H
1
n
(M) such that
(4.115) [G
n
[
r,n
C
2
Ne
(1l
0
)tn
.
It follows that there exists D
n
I
r
n
(M) H
r
n
(M) such that U
n
H
0
n
(M)
has the following orthogonal decomposition in the Hilbert space H
r
n
(M):
(4.116) U
r
n
= G
n
+ D
n
.
For each n N, let
n
: H
r
n
(M) I
r
n
(M) be the orthogonal projection
onto the subspace of S
n
-invariant distributions and let
n
:= t
n+1
t
n
. Since
by the denitions U
r
n+1
= G
r
n
(U
r
n
), for all n N, and
G
r
n
_
I
r
n
(M) H
r
n
(M)
_
I
r
n+1
(M) H
r
n+1
(M)
the following recursive identity holds:
(4.117) D
n+1
= G
r
n
(D
n
) +
n+1
G
r
n
(G
n
) .
Since [G
r
n
(G
n
)[
r,n+1
e
rn
[G
n
[
r,n
by the bound (4.67) on the norm of
G
r
t
: H
r
t
(M) H
r
Gt(q)
(M), it follows from (4.110) and (4.115) that
(4.118) [D
n+1
[
r,n+1
C
1
e
l
1
n
[D
n
[
r,n
+ C
2
Ne
rn(1l
0
)tn
,
which implies by induction a bound of the form
(4.119) [D
n+1
[
r,n+1
C
n
1
e
l
1
(t
n+1
t
0
)
[D
0
[
r,0
+ C
2
N
n1

j=0
C
j
1
e
s
n,j
with the sequence s
n,j
[n N, 0 j n given by the identity
(4.120) s
n,j
:= l
1
(t
n+1
t
nj+1
) + r
nj
(1 l
0
)t
nj
.
Since U L
2
q
(M) and G
r
n
(U) = G
0
n
(U) H
0
n
(M) = L
2
q
(M) H
r
n
(M),
the following bound holds:
(4.121) [G
r
n
(U)[
0,n
[U[
0
, for all n N.
It follows in particular from the decomposition (4.116) that
(4.122) [D
0
[
r,0
[U
r
0
[
r,0
[G
r
0
(U)[
0,0
[U[
0
.
90 GIOVANNI FORNI

The main step in the argument is the proof of the following claim. There
exist a positive measurable function C
3
: P
2
R
+
and a real number
l
3
:= l
3
(l
0
, l
1
) > 0 such that
(4.123)
n1

j=0
C
j
1
e
s
n,j
C
3
(q) e
l
3
n
, for all n N.
Let
1
< <
2
be positive real numbers such that
(4.124)
(a) (1 l
0
)
1
(l
1
+ r)(
2

1
) > 0 ,
(b) l
1

2
l
1
(
2

1
) > log C
2
.
By condition (4.112) on the sequence (t
n
)
nN
there exists a measurable
function n
0
: M
(1)

N such that
(4.125)
1
n t
n
(q)
2
n, for all n n
0
(q) .
It follows that, for any q M
(1)

, all n n
0
(q) and all j n,
(4.126)
s
n,j
[(l
1
+ r)(
2

1
) (1 l
0
)
1
] n
+ [(1 l
0
)
1
r(
2

1
) l
1

2
] j + l
1
(
2

1
) + r
2
.
Let K := C
2
e
(1l
0
)
1
r(
2

1
)l
1

2
. There are two cases to consider:
(4.127) (a) K 1 (b) K > 1 .
In case (a) we immediately obtain that, for all n n
0
(q),
(4.128)
n1

j=n
0
C
j
1
e
s
n,j
ne
l
1
(
2

1
)+r
2
e
[(1l
0
)
1
(l
1
+r)(
2

1
)] n
,
in case (b) we obtain instead that
(4.129)
n1

j=n
0
C
j
1
e
s
n,j

e
l
1
(
2

1
)+r
2
K 1
C
n
1
e
[l
1

2
l
1
(
2

1
)] n
.
It follows that there exist constants A > 0 and := (l
0
, l
1
) > 0 such that
(4.130)
n1

j=n
0
C
j
1
e
s
n,j
Ae
n
, for all n n
0
(q) .
By condition (4.112) on the sequence (t
n
)
nN
and by denition (4.120)
(4.131)
s
n,j
(q)
n
(1 l
2
) , as n +,
SOBOLEV REGULARITY OF SOLUTIONS 91
for -almost all q M
(1)

, uniformly with respect to j 0, . . . , n


0
(q)1.
Hence there exists a measurable function n
1
: M
(1)

N with n
1
n
0
and
positive constants B > 0 and := (l
0
, l
1
) > 0 such that
(4.132)
n
0
1

j=0
C
j
1
e
s
n,j
Be
n
, for all n n
1
(q) .
The claim (4.123) then follows from the estimates (4.130) and (4.132).
It then follows from the orthogonal decomposition (4.116), from the claim
(4.123), proved above, together with the upper bounds (4.115), (4.119),
(4.122) and the lower bound for visiting times in (4.125), that there exist a
measurable function C
4
: M
(1)

R
+
and a constant l
4
:= l
4
(l
0
, l
1
) > 0
such that
(4.133) [G
r
n
(U)[
r,n
C
4
(q)[U[
0
+ N e
l
4
n
, for all n n
1
(q) .
By the interpolation inequality for the scale of dual weighted Sobolev norms
(see Corollary 2.26), from the upper bounds (4.121) and (4.133) it follows
that for any > 0 there exist a measurable function C

: M
(1)

R
+
and a
constant l

:= l(, l
0
, l
1
) > 0 such that
(4.134) [G

n
(U)[
,n
C

(q)[U[
0
+ N e
l n
, for all n n
1
(q) .
It remains to prove that the latter bound implies the statement of the theo-
rem. For -almost all q M
(1)

and for all t t


0
(q), there exists a unique
n(t, q) N such that
(4.135) t
n(t,q)
(q) t < t
n(t,q)+1
(q) .
Let > 0 be xed and let
()
1
< <
()
2
be positive real numbers
such that l

(
()
2

()
1
) > 0. There exists a measurable function
n
()
2
: M
(1)

R such that n
(s)
2
n
1
and, for -almost q M
(1)

,
(4.136)
()
1
n t
n
(q)
()
2
n, for all n n
()
2
(q) .
Let t
()
1
: M
(1)

R
+
be a measurable function such that t
()
1
t
0
and
n(t, q) n
()
2
(q) if t t
()
1
(q), for -almost all q M
(1)

. It follows from
(4.135) and (4.136) that, for any t t
()
1
(q),
(4.137) n(t, q) + 1
t
n(t,q)+1
(q)

()
2
>
t

()
2
.
It follows by (4.136) and (4.137), for -almost q M
(1)

and all t t
()
1
(q),
(4.138)
n(q,t)
(
()
2

()
1
) (n(q, t) + 1) +
()
1
.
92 GIOVANNI FORNI

Since, for any n N and any t
n
t < t
n+1
,
(4.139)
[G

t
(U)[

= [G

ttn
G
t
1
(U)[

e
(ttn)
[G

n
(U)[

e
n
[G

n
(U)[

,
by the upper bounds (4.134), (4.137) and (4.138) the following estimate
holds. Let := (, l
0
, l
1
) be the real number dened as follows:
(4.140) :=
l

(
()
2

()
1
)

()
2
> 0 .
For -almost all q M
(1)

and all t t
()
1
(q)
(4.141) [G

t
(U)[

(q) e

()
1
+l
[U[
0
+ N e
t
.
Finally, for -almost all q M
(1)

and all 0 t t
()
1
(q),
(4.142) [G

t
(U)[

[U[
0
e
t
()
1
(q)
[U[
0
e
t
.
The desired estimate (4.108) immediately follows from the upper bounds
(4.141) and (4.142).

5. SMOOTH SOLUTIONS
In this nal section we prove our main theorems on Sobolev regularity of
smooth solutions of the cohomological equation for translation ows. The
general result which we are able to prove for any translation surface (and
almost all directions) is a direct consequence of the Fourier analysis con-
struction of distributional solutions in 3.1 and of the uniqueness result
that follows from Theorem 4.27. The sharper result which we will prove for
almost all translation surfaces (and almost all directions) requires a deeper
analysis of ergodic averages of translation ows. In fact, the construction
of square integrable (bounded) solutions is based (as in [MMY05]) on the
Gottschalk-Hedlund theorem. The required bounds on the growth of er-
godic integrals are derived from the Oseledec-type result (Theorem 4.30)
and the spectral gap result (Theorem 4.36) for distributional cocycles, by
the methods developed in [For02] in the study of the deviation of ergodic
averages. The Oseledec-type theorem was in fact already proved in [For02],
while the spectral gap theorem for distributional cocycles is new.
5.1. The general case. In this section we prove a result on existence of
smooth solutions of the cohomological equation for translation ows which
holds for any orientable quadratic differential in almost all directions. Such
a result answers a question of Marmi, Moussa and Yoccoz who asked what
is the best, that is, the smallest regularity loss within reach of the Fourier
SOBOLEV REGULARITY OF SOLUTIONS 93
analysis methods of [For97]. As indicated in [MMY05], the answer is es-
sentially that the solution loses no more than 3+ derivatives (for any > 0)
with respect to the scale of weighted Sobolev spaces introduced in 2.1. We
recall that the regularity loss obtained in [MMY05] (which only holds for
almost all quadratic differentials) is essentially 1 + BV.
Theorem 5.1. Let q M
(1)

be any quadratic differential. Let k N be


any integer such that k 3 and let s > k and r < k 3. For almost
all S
1
(with respect to the Lebesgue measure), there exists a constant
C
r,s
() > 0 such that the following holds. If f H
s
q
(M) is such that
D(f) = 0 for all D I
s
q

(M), the cohomological equation S

u = f has a
solution u H
r
q
(M) satisfying the following estimate:
(5.1) [u[
r
C
r,s
() [f[
s
.
Proof. As a rst step, we prove that, under nitely many distributional con-
ditions on f H
s
q
(M) with s > 3, the cohomological equation S

u = f
has, for almost all S
1
, a solution u H
r
q
(M) for any r > 0, which
satises an estimate such as (5.1). Since
F
q
f H
s2
q
(M) with s 2 > 1,
by Theorem 3.3, for any r > 0 and for almost all S
1
, there exists a so-
lution U

H
2r
q
(M) of the cohomological equation S

U =
F
q
f, which
vanishes on constant functions and satises the bound
(5.2) |U|
2r
C
(1)
r,s
() |
F
q
f|
s2
C
(1)
r,s
() |f|
s
.
Let u

H
r
q
(M) be the unique distribution vanishing on constant functions
such that
F
q
u = U. Since the commutation relation S

q
v =
q
S

v holds
for any v H
3+r
q
(M), the following distributional equation holds:
(5.3)
q
(S

u f) = 0 in H
3r
q
(M) .
In addition, from the estimate (5.2) it follows immediately that
(5.4) [u[
r
C
(1)
r,s
() [f[
s
.
Let N
r



H
r
q
(M) be the subspace dened as follows:
N
r

:= u H
r
q
(M) [
q
S

u = 0 H
3r
q
(M) .
Since the kernel K
r
(
q
) H
1r
q
(M) (which is equal to the subspace
perpendicular to the range of the operator
q
: H
3+r
q
(M) H
1+r
q
(M)) and
the space I
r
q

(M) of S

-invariant distributions are both nite dimensional,


it follows that the subspace N
r

is nite dimensional (for almost all S


1
).
As a consequence, the space
(5.5) S

(N
r

) := S

u [ u N
r

K
r
(
q
)
94 GIOVANNI FORNI

is nite dimensional, hence closed in H
1r
q
(M). We claim that there exists
a constant C
(2)
r,s
() > 0 such that the following holds. For almost all S
1
,
there exists a unique distribution U

(f) H
r
q
(M) such that
(a) U

(f) is orthogonal to the subspace I


r
q

(M);
(b) S

(f) f K
r
(
q
) is orthogonal to the subspace S

(N
r

);
(c) U

(f) satises the following bound:


(5.6) [U

(f)[
r
C
(2)
r,s
() [f[
s
.
In fact, let u H
r
q
(M) be any solution of equation (5.3) which satises
the bound (5.4). Let u
1
H
r
q
(M) be the component of u orthogonal
in H
r
q
(M) to the subspace I
r
q

(M) of S

-invariant distributions. Since


u u
1
I
r
q

(M), the distribution u


1
still satises the equation (5.3). In
addition, the norm [u
1
[
r
[u[
r
, hence the bound (5.4) still holds. Since
N
r

is nite dimensional, there exists u


2
H
r
q
(M), a solution of (5.3),
such that S

u
2
f is equal to the component of S

u
1
f orthogonal to
S

(N
r

) in H
1r
q
(M). Let U

(f) be the component of u


2
orthogonal to
I
r
q

(M) in H
r
q
(M). By construction U

(f) satises the conditions (a) and


(b) above and it is uniquely determined by them. Let us prove condition
(c). The linear operator S

: N
r

(N
r

) is surjective and its domain is a


nite dimensional subspace, hence it has a bounded right inverse and there
exists K
r
:= K
r
() > 0 such that, for any u N
r

orthogonal to I
r
q

(M),
(5.7) [u[
r
K
r
[S

u[
1r
.
Since by denition U

:= U

(f) and u
1
H
r
q
(M) are both orthogonal to
I
r
q

(M) and U

u
1
N
r

, by the bound (5.7),


(5.8) [U

[
r
[U

u
1
[
r
+[u
1
[
r
K
r
[S

(U

u
1
) [
1r
+[u
1
[
r
.
Since again by denition S

= S

u
2
and S

u
2
f is equal to the com-
ponent of S

u
1
f orthogonal to S

(N
r

),
(5.9)
[S

(U

u
1
) [
1r
[S

u
1
f[
1r
+[S

u
2
f[
1r
2[S

u
1
f[
1r
2[u
1
[
r
+ 2[f[
1r
.
It follows from the bounds (5.8), (5.9) and from the bound (5.4) for the
distribution u
1
H
r
q
(M) that the required bound (5.6) holds with
C
(2)
r,s
() := [(2K
r
() + 1)C
(1)
r,s
()] + 1 .
Since S

(N
r

) K
r
(
q
) is nite dimensional, there exist a nite (linearly
independent) set
1
, . . . ,
K
of bounded linear (real-valued) functionals
on the Hilbert space H
1r
q
(M) such that
K
r
(
q
) S

(N
r

Ker(
1
) Ker (
K
) = 0 .
SOBOLEV REGULARITY OF SOLUTIONS 95
Let D
1
, . . . , D
K
H
s
q
(M) be the system of distributions dened as
follows: for each j 1, . . . , K,
(5.10) D
j
(f) :=
j
(S

(f) f) , for all f H


s
q
(M) .
The system D
1
, . . . , D
K
has by construction the following property: if
D
j
(f) = 0 for all j 1, . . . , K, then U

(f) is the required solution of


the cohomological equation S

u = f. In fact, under such conditions U

(f)
is by construction a solution orthogonal to constant functions. By Theorem
4.27, there exists s(q) > 0 such that for all 0 < s < s(q) the space I
s
q

(M)
is 1-dimensional for almost all S
1
. Thus, if r < s(q), for almost all
S
1
the solution u H
r
q
(M) of the cohomological equation S

u = f
is unique (if it exists). It follows that, for any s(q) > r > 0, the distribution
U

(f) is the unique solution in H


r
q
(M) of the cohomological equation,
which implies that U

(f) H
r
q
(M) for any r > 0.
We claim that D
1
, . . . , D
K
I
s
q

(M). In fact, by its denition in formula


(5.10), the distribution D
j
H
s
q
(M) for all j 1, . . . , K. In addition,
if there exists v H
s+1
q
(M) such that f = S

v , by conditions (a) and (b)


on the distribution U

(f) we have
S

(U

(f) v) = S

(f) f K
r
(
q
) S

(N
r

.
However, by denition S

(U

(f) v) S

(N
r

), hence S

(f) f = 0
which implies that D
1
(f) = = D
K
(f) = 0. Hence by denition all the
distributions of the system D
1
, . . . , D
K
are S

-invariant.
Finally we prove the statement of the theorem by induction on k N.
For k = 3 the statement holds by the previous argument. Let us assume
that the statement holds for k 3. By the induction hypothesis, for any
s > k+1 and any r < k2 there exists C
r,s
() > 0 such that, for almost all
S
1
and for any f H
s
q
(M) with D(f) = 0 for all D I
s
q

(M), there
exist solution u, u
S
and u
T
H
r1
q
(M) of the cohomological equations
S

u = f, S

u
S
= Sf and S

u
T
= Tf respectively such that u, u
S
and u
T
are orthogonal to constant functions and the following bound holds:
(5.11)
[u[
2
r1
+[u
S
[
2
r1
+[u
T
[
2
r1
C
r,s
()
_
[f[
2
s1
+[Tf[
2
s1
+[Sf[
2
s1
_
.
In fact, since f H
s
q
(M) is such that D(f) = 0 for all D I
s
q

(M) it
follows immediately that D(Sf) = D(Tf) = 0 for all D I
s1
q

(M).
Since S

commutes with S, T in the sense of distributions, it follows that


the distributions Su u
S
and Tu u
T
H
r1
q
(M) are S

-invariant. Let
s(q) > 0 be such that for all 0 < s < s(q) and for almost all S
1
the space I
s
q

(M) is 1-dimensional (see Theorem 4.27). If s(q) > 1 r,


since Su u
S
and Tu u
T
H
r1
q
(M) are S

-invariant and orthogonal


to constant functions, it follows that Su = u
S
and Tu = u
T
. The latter
96 GIOVANNI FORNI

identities imply that u H
r
q
(M) and by (5.11) the required bound (5.1) is
satised. The proof of the induction step is therefore completed.
5.2. Ergodic integrals. Optimal results on the loss of regularity for almost
all orientable quadratic differentials (and almost all directions) will be de-
rived from bounds on ergodic integrals by the Gottschalk-Hedlund theorem.
The required bounds will be proved along the lines of 9 in [For02] with
the key improvement given by the spectral gap Theorem 4.36.
The key idea of the argument given in [For02] consists in studying the dy-
namics of the distributional cocycle
1
t

tR
on the innite dimensional
(non-closed) sub-bundles

W
1

(M) of 1-dimensional currents gener-


ated by segments of leaves of the horizontal and vertical foliations.
Let T > 0. We will denote by
T
q
a positively oriented segment of length
T > 0 of a leaf of the measured foliation F
q
respectively. By the trace the-
orems for standard Sobolev spaces and by the comparison Lemma 2.11, the
vector spaces

q
generated by the sets
T
q
(p)[(p, T) M R are sub-
spaces of the weighted Sobolev space W
s
q
(M) of 1-dimensional currents
for any s 1 and the corresponding sub-bundles

are invariant under the


action of the cocycle
s
t

tR
. Let F
q
: MR M [F
q
: MR M]
be the (almost everywhere dened) ow of the horizontal [vertical] vec-
tor elds S
q
[T
q
] on M. The horizontal and vertical foliations F
q
co-
incide almost everywhere with the orbit foliations of the ows F
q
. Let

T
q

q
be a positively oriented segment with initial point p

M and
let := f
+

T
+ f

S
W
s
q
(M). The following identity holds:
(5.12)
T
q
() =
_
T
0
f

F
q
(p

, ) d ,
The ergodic averages of the functions f

H
s
q
(M) (for s 1) can
therefore be understood by studying the dynamics of the renormalization
cocycles
s
t

tR
on the sub-bundle

. In [For02] we have proved the


following basic estimate on the weighted Sobolev norm of the currents

T
q
W
1
q
(M). Let R
q
be the at metric with conical singularities associ-
ated to the quadratic differential q M
(1)

and let [[q[[ denote the R


q
-length
of the shortest saddle connection of the at surface (M, R
q
). (We recall that
a saddle connection is a segment joining conical points).
Lemma 5.2. ([For02], Lemma 9.2) There exists a constant K > 0 such
that, for all quadratic differentials q M
(1)

,
(5.13) [
T
q
[
W
1
q
(M)
K(1 +
T
[[q[[
) .
SOBOLEV REGULARITY OF SOLUTIONS 97
The above estimate is a trivial bound, linear with respect to the length
T > 0 of the orbit segment, and can be quite easily derived fromthe Sobolev
trace theoremfor rectangles in the euclidean plane R
2
. The number [[q[[ > 0,
which measures the pinching of the at surface, gives the order of magni-
tude of the edges of the largest at rectangle which can be embedded in the
at surface (M, R
q
) around an arbitrary regular point.
We recall below the terminology and the notations, introduced in 9 of
[For02], concerning return trajectories of translation ows.
Denition 5.3. A point p M is regular with respect to a measured folia-
tion F if it belongs to a regular (non-singular) leaf of F. For any quadratic
differential q Q(M). A point p M will be said to be q-regular if it is
regular with respect to the horizontal and vertical foliations F
q
.
The set of q-regular points is of full measure and it is equivariant under the
actions of the group Diff
+
(M) and of the Teichmller ow on Q(M).
Denition 5.4. Let p M be a q-regular point and let I
q
(p) be the vertical
[horizontal] segment of length [[q[[/2 centered at p. A forward horizontal
[vertical] return time of p M is a real number T := T
q
(p) > 0 such
that F
q
(p, T) I
q
(p). If T > 0 is a horizontal [vertical] return time for
a q-regular point p M, the horizontal [vertical] forward orbit segment

T
q
(p) will be called a forward horizontal [vertical] return trajectory at p.
There is a natural map from the set of horizontal [vertical] return trajectories
into the set of homotopically non-trivial closed curves.
Denition 5.5. The closing of any horizontal [vertical] trajectory segment

T
q
(p) is the piece-wise smooth homotopically non-trivial closed curve
(5.14)
T
q
(p) :=
T
q
(p)
_
p, F
q
(p, T)
_
,
obtained as the union of the trajectory segment
T
q
(p) with the shortest
geodesic segment
_
p, F
q
(p, T)
_
joining its endpoints p and F
q
(p, T).
Let T
(1)
q
(p) > 0 be the forward horizontal [vertical] rst return time of a
q-regular point p M, dened to be the real number
(5.15) T
(1)
q
(p) := minT > 0 [ F
q
(p, T) I
q
(p) .
The corresponding forward horizontal [vertical] trajectory
(1)
q
(p) with ini-
tial point p will be called the forward horizontal [vertical] rst return tra-
jectory at p. The following bounds hold for rst return times:
Lemma 5.6. ([For02], Lemma 9.2) There exists a measurable function
K
r
: M

R
+
such that, if T
(1)
q
(p) is the forward horizontal [vertical]
98 GIOVANNI FORNI

rst return time of a q-regular point p M, then
(5.16) [[q[[/2 T
(1)
q
(p) K
r
(q) .
The lower bound in (5.16) is an immediate consequence of the denition of
[[q[[ as the length of the shortest saddle connection, while the upper bound
depends essentially on the fact that the rst return map of a translation ow
to a transverse interval is an interval exchange transformation, hence the
return-time function is piece-wise constant (and bounded).
In 9.3 of [For02] we have constructed special sequences of horizontal [ver-
tical] return times for almost all quadratic differentials generated. Such spe-
cial return times are related to visiting times of the Teichmller ow, which
renormalizes translation ows, to appropriate compact subsets of positive
measure. Let q M
(1)

be a Birkhoff generic point of the Teichmller ow


G
t

tR
and let S

(q) M
(1)

be a smooth compact hypersurface of codi-


mension 1, such that q S

(q) and S

(q) is transverse to the Teichmller


ow. Let (t
k
) be the sequence of visiting times of the orbit G
t
(q) [ t R
to S

(q). Since, by denition, for any t R,


(5.17) (F
Gt(q)
, F
Gt(q)
) = (e
t
F
q
, e
t
F
q
) ,
if t := t
k
< 0 is a backward visiting time of the orbit G
t
(q)
tR
, any
forward rst return trajectory of the horizontal foliation F
Gt(q)
is a for-
ward return trajectory of the horizontal foliation F
q
, provided [t
k
[ is suf-
ciently large. In a similar way, if t := t
k
> 0 is a forward visiting time
of G
t
(q)
tR
, any forward rst return trajectory of the vertical foliation
F
Gt(q)
is a forward return trajectory of the vertical foliation F
q
.
By the closing of the return trajectories of the horizontal [vertical] foliation,
as in (5.14), we obtain closed currents of Sobolev order 1. The evolution of
such currents under the action of the Teichmller ow is therefore described
by the cocycles
s
t

tR
on the bundle Z
s

(M) for any s 1. In the case


s = 1 the dynamics of the cocycle
1
t
[Z
1

(M) is completely described


by Theorem 4.30. In the following we will analyze Lyapunov exponents
of closed currents given by the closing of horizontal [vertical] trajectories
of translation ows under the cocycles
s
t

tR
for s > 1. A complete
description of the Lyapunov structure of the cocycles
s
t
[Z
s

(M) for s >


1 is not relevant for our purposes and will not be attempted.
Let be a SO(2, R)-absolutely continuous, KZ-hyperbolic measure on a
stratumM
(1)

of orientable quadratic differentials (in the sense of Denitions


4.13 and 4.4). We recall that according to [For02] all the canonical abso-
lutely continuous, SL(2, R)-invariant measures on any connected compo-
nent of any stratum of orientable quadratic differentials are KZ-hyperbolic.
Adifferent and stronger proof which establishes that all the afore-mentioned
SOBOLEV REGULARITY OF SOLUTIONS 99
measures are KZ-simple has been given in [AV05] (see Theorem 4.6). The
latter theorem is not necessary for any of the results of this paper to hold.
By Theorem 4.30, for any s 1, there exists a measurable splitting
(5.18) Z
s

(M) = B
1
,+
(M) B
1
,
(M) E
s

(M) .
The Lyapunov exponents of the restriction of the distributional cocycle

s
t

tR
to the nite dimensional sub-bundles B
1
,+
(M) [B
1
,
(M)] are
equal to the non-negative [non-positive] exponents of the Kontsevich-Zorich
cocycle. It follows from the non-uniform hyperbolicity hypothesis for the
latter that all Lyapunov exponents of the cocycle
s
t
[B
1
,+
(M) are strictly
positive while all Lyapunov exponents of the cocycle
s
t
[B
1
,
(M) are
strictly negative. In particular, by Oseledecs theorem there exists a mea-
surable function
s

: M
(1)

R
+
and a real number

> 0 such that, for


-almost all q M
(1)

, for all C B
1
q
(M) and all t 0,
(5.19) [
s
t
(C)[
s

s

(q)[C[
s
e
t
.
In [For02], we have proved that for s = 1 the restriction
s
t
[E
s

(M) has
0 has the unique Lyapunov exponent. In fact, this cocycle is isometric with
respect to a continuous Lyapunov norm (see Theorem 4.30). For s > 1,
estimates on Lyapunov exponents and Lyapunov norms for the restriction

s
t
[E
s

(M) will be derived from Theorem 4.36. Let := s 1 > 0. For


any q M
(1)

, any l < 1 and any > 0, let


(5.20) N
,,l

(q) := sup
UH
0
q
(M)
N
,
q
(U)
[U[
0
+N
1,l
q
(S
q
U)
.
By Theorem 4.36 for any l < 1 there exists := (, l) > 0 such that
formula (5.20) denes a (measurable) function N
,,l

: M
(1)

R
+
.
Let
q
: Z
1
q
(M) B
1
q
(M) and E
q
: Z
1
q
(M) E
1
q
(M) be the projec-
tions determined by the splitting (5.18). For any s 1, the restrictions of
the projections
q
and E
q
to the subspace Z
s
q
(M) Z
1
q
(M) can be ex-
tended to projections
s
q
: W
s
q
(M) B
1
q
(M) and E
s
q
: W
s
q
(M)
E
s
q
(M), dened on the Sobolev space W
s
q
(M) of 1-dimensional currents,
by composition with the orthogonal projection W
s
q
(M) Z
s
q
(M) onto
the closed subspace of closed currents. Let d
s

: M
(1)

R
+
be the dis-
torsion of the splitting (5.18), that is, the function dened for -almost all
q M
(1)

as
(5.21) d
s

(q) := sup
CZ
s
q
(M)
[
q
(C)[
s
+[
q
(C)[
s
+[E
q
(C)[
s
[C[
s
.
Let s > 1 and l < 1. Let P
(1)

M
(1)

be a compact set satisfying the


100 GIOVANNI FORNI

following conditions:
(1) All q P
(1)

are Birkhoff generic points for the Teichmller ow


G
t
and Oseledec regular points for the cocycle
s
t
[B
1

(M);
(2) the set P
(1)

is transverse to the Teichmller ow and has positive


transverse measure;
(3) there exists a constant K
(1)
r
> 0 such that, for all q P
(1)

and all
q-regular p M, the rst return times T
(1)
q
(p) K
(1)
r
;
(4) the real-valued functions
s

, N
,,l

and
1

, introduced respectively
in formulas (5.19), (5.20) and (5.21), are bounded on P
(1)

.
By the ergodicity of the system (G
t
, ), by the Oseledecs theorem for
the cocycle
s
t
[B
1

(M) and by Lemma 4.6, it follows that the union of all


sets P
(1)

with the properties (1) (4) is a full measure subset of M


(1)

.
Denition 5.7. Let q M
(1)

and (t
k
)
kN
be the sequence of visiting times
of the backward orbit G
t
(q) [ t t
1
= 0 to a compact positive measure
set P
(1)

M
(1)

. A principal sequence (T
(k)
q
(p))
kN
of forward return times
for the horizontal foliation F
q
at a q-regular point p M is the sequence
(5.22) T
(k)
q
(p) := T
(1)
Gt(q)
(p) exp [t[ , t = t
k
.
For each k N, a horizontal principal (forward) return trajectory
(k)
q
(p)
at a q-regular point p M is the horizontal forward return trajectory at the
point p corresponding to a principal return time T
(k)
q
(p) > 0.
We remark that a horizontal principal return trajectory
(k)
q
(p) coincides
with the horizontal rst return trajectory at p of the quadratic differential
G
t
(q), t = t
k
< 0. A similar denition can be given for the vertical foliation
F
q
by considering forward visiting times of the Teichmller ow.
The following standard splitting lemma allows to reduce the analysis of
arbitrary regular trajectories to that of principal return trajectories.
Lemma 5.8. ([For02], Lemma 9.4) Under conditions (1) (3) there exists
a constant K
P
> 0 such that the following holds. Let q P
(1)

and, for any


T > 0, let
T
q
(p) be a forward trajectory at a q-regular point p M. There
exists a nite set of points p
(k)
j
[ 1 k n, , 1 j m
k

T
q
(p), such
that the principal return trajectories
(k)
q
(p
(k)
j
)
T
q
(p) do not overlap and
(5.23)

T
q
(p) =
n

k=1
m
k

j=1

(k)
q
(p
(k)
j
) + b
T
q
(p) ,
with m
k
K
P
e
|t
k+1
||t
k
|
, for 1 k n,
and the R
q
-length L
q
_
b
T
q
(p)
_
K
P
.
SOBOLEV REGULARITY OF SOLUTIONS 101
Proof. We recall the proof given in [For02], Lemma 9.4. The argument
is based on the following estimate on principal return times. By (5.22),
Lemma 5.6 and condition (3), there exists a constant K
pr
> 0 such that, for
all q P
(1)

, all q-regular points p M and all k N,


(5.24) K
1
pr
exp [t
k
[ T
(k)
q
(p) K
pr
exp [t
k
[ .
Let n = maxk N[ T
(k)
q
(p) T . The maximum exists (nite) by
(5.24). Let p
(n)
1
:= p. The sequence (p
(k)
j
) with the properties stated in
(5.23) can be constructed by a nite iteration of the following procedure.
Let p
(k)
j
be the last point already determined in the sequence and let
(5.25) p
(k)
j+
:= F
q
_
p
(k)
j
, T
(k)
q
(p
(k)
j
)
_

T
q
(p) .
Let then k

1, . . . , k be the largest integer such that


(5.26) F
q
_
p
(k)
j+
, T
(k

)
q
(p
(k)
j+
)
_

T
q
(p) .
If k

= k, let p
(k)
j+1
:= p
(k)
j+
. If k

< k, let m
k
:= j, m
h
= 0 (no points) for
all k

< h < k and p


(k

)
1
:= p
(k)
j+
. The iteration step is concluded. By (5.24)
it follows that
(5.27) K
1
pr
e
|t
k
|
m
k
T
(k+1)
q
(p
(k)
1
) K
pr
e
|t
k+1
|
.
The length of the remainder b
T
q
(p) has to be less than any upper bound for
the length of rst return times. Hence, by (5.27), Lemma 5.6 and condition
(3), the estimates in (5.23) are proved and the argument is concluded.
The result below gives a fundamental uniformestimate on ergodic integrals.
Theorem 5.9. Let be an SO(2, R)-absolutely continuous, KZ-hyperbolic
measure on a stratum M
(1)

of orientable quadratic differentials. For any


s > 1 there exists a measurable function
s

: M
(1)

R
+
such that, for
-almost q M
(1)

, for all q-regular p M and for all T > 0 ,


(5.28) [
s
q
_

T
q
(p)
_
[
s
+ [E
s
q
_

T
q
(p)
_
[
s

s

(q) .
Proof. Let P := P
(1)

be a compact set satisfying conditions (1) (4) listed


above. For -almost all q P, there exists a sequence (t
k
)
kN
of back-
ward return times of the Teichmller orbit G
t
(q) to the set P. By Lemma
5.8 the uniform estimate (5.28) can be reduced (exponential) estimates for
principal return trajectories
(k)
q
(p). We claim that for every s > 1 there
exist constants K
s
P
> 0 and
s
> 0 such that for -almost all q P, for all
q-regular p M and all k N,
(5.29) [
s
q
_

(k)
q
(p)
_
[
s
+[E
s
q
_

(k)
q
(p)
_
[
s
K
s
P
e
s|t
k
|
.
102 GIOVANNI FORNI

By Lemma 5.2, by conditions (3) and (4) on the set P, in particular the
bound on the distorsion, there exists a constant K
s
1
(P) > 0 such that the
following holds. For -almost all q P and all q-regular p M, the
closing
(1)
q
(p) of the rst return horizontal trajectory
(1)
q
(p) (see Denition
5.5) satises the following bound:
(5.30) [
q
_

(1)
q
(p)
_
[
1
+ [E
q
_

(1)
q
(p)
_
[
1
K
1
(P)
Since by denition of the principal return trajectories
(5.31)
(k)
q
(p) =
s
|t|
_

(1)
Gt(q)
(p)
_
, for any t = t
k
0 ,
it follows from the initial bound (5.30), from the invariance of the sub-
bundle B
1
,
(M) under the cocycles
s
t
, from the Lyapunov bound (5.19)
on the cocycle
s
t
[B
1
,
(M) and from condition (4) on the set P M
(1)

that there exists a constant K


s
2
(P) > 0 such that the following estimate
holds. For -almost all q P and for all q-regular p M,
(5.32) [
q
_

(k)
q
(p)
_
[
s
K
s
2
(P) e
|t
k
|
, for all k N.
A similar estimate holds for the projections of principal return trajectories
on the sub-bundle E
1

(M) of exact currents. In fact, there exist K


s
3
(P) > 0
and
s
> 0 such that, for -almost all q P and for all q-regular p M,
(5.33) [E
q
_

(k)
q
(p)
_
[
s
K
s
2
(P) e
s|t
k
|
, for all k N.
The estimate (5.33) is proved as follows. By denition of the bundle E
1

(M)
of exact currents, for all q M
(1)

and all q-regular p M, there exists a


unique function U
(1)
q
(p) H
0
q
(M) L
2
q
(M) such that
(5.34) E
q
_

(1)
q
(p)
_
= dU
(1)
q
(p) .
By the denitions of the distributional cocycles
s
t
and G
s
t
the follow-
ing identity of cocycles holds for any s 1:
(5.35)
s
t
d = d G
s1
t
on H
s+1
q
(M) .
The exponential estimate (5.33) will therefore follow from Theorem 4.36
if we can prove that there exist a positive number l < 1 and a constant
K
3
(P) > 0 such that, for -almost all q P and all q-regular p M, the
following holds:
(5.36) [U
(1)
q
(p)[
0
+ N
1,l
q
_
S
q
U
(1)
q
(p))

K
s
3
(P) .
Let [ [

be the norm on the bundle E


1

(M) of exact currents dened as


follows: for any q M
(1)

and C E
1
q
(M),
(5.37) [C[
E
:= [U[
L
2
q
(M)
, if C = dU with U H
0
q
(M) .
SOBOLEV REGULARITY OF SOLUTIONS 103
By the denitions of weighted Sobolev norms, it is immediate to prove that
the map d : H
0
q
(M) E
1
q
(M) is a continous bijective map of Hilbert
spaces. Hence, by denition (5.37) and by the open mapping theorem, for
each q M
(1)

there exists K(q) > 0 such that


(5.38) K(q)[ [
E
[ [
1
[ [
E
on E
1
q
(M) .
Since the sub-bundle H
0

(M) H
0

(M) and the norms [ [


E
, [ [
1
on
the bundle E
1

(M) are all continuous, the function K : M


(1)

R
+
can
be chosen continuous (see Lemma 9.3 in [For02] for a detailed argument).
Hence by the estimates (5.30) and by identity (5.34), there exists a constant
C
1
(P) > 0 such that, for -almost all q P and all q-regular p M,
(5.39) [U
(1)
q
(p)[
L
2
q
(M)
K(q)
1
[E
q
_

(1)
q
(p)
_
[
1
C
1
(P) .
The proof of the estimate (5.36) is completed as follows. Since
(1)
q
(p) is a
horizontal trajectory, by the splitting (5.18) and by the identity (5.34), the
following formula holds:
(5.40) S
q
U
(1)
q
(p) =
Sq
[
(1)
q
(p)
(1)
q
(p)]
Sq

q
[
(1)
q
(p)] .
Since the distribution
Sq
[
(1)
q
(p)
(1)
q
(p)] is given by integration along a
vertical arc of unit length, by the Sobolev embedding Lemma 5.2 and by the
logarithmic law of geodesics for the Teichmller geodesic ow on moduli
spaces (see [Mas93], Prop. 1.2, or 4.1, formula (4.49)) it follows that for
any l > 0 there exists a constant C
l
> 0 such that, for -almost all q M
(1)

and all q-regular p M, the Lyapunov norm


(5.41) N
1,l
q
_

Sq
[
(1)
q
(p)
(1)
q
(p)]
_
C
l
.
By the cocycle isomorphism(4.78), by the Lyapunov estimate (5.19) and by
condition (4) on the set P, since the distribution
Sq

q
[
(1)
q
(p)] I
1
q
(M),
there exists a constant C
2
(P) > 0 such that, for any 1 > l 1

, for
-almost all q M
(1)

and all q-regular p M, the Lyapunov norm


(5.42) N
1,l
q
_

Sq

q
[
(1)
q
(p)]
_
C
2
(P) .
The bound (5.36) follows immediately from the bound (5.39), from the
identity (5.40) and from the bounds (5.41) and (5.42).
As hinted above, the estimate (5.36) implies the required exponential esti-
mate (5.33) on the projections of principal return trajectories on the space
of exact currents. In fact, by Theorem 4.36 and by condition (4) on the set
P, for any 1 > l 1

and any > 0, there exist K

4
(P) > 0 and

s
:= (l, ) > 0 such that, for -almost all q P and all q-regular p M,
(5.43) N
,
q
_
U
(1)
q
(p)
_
K
s
4
(P) .
104 GIOVANNI FORNI

Let us introduce the following notation: for any k N, let
(5.44) q
k
:= G
t
k
(q) P and U
(k)
q
(p) := G
0
|t
k
|
_
U
(1)
q
k
(p)
_
L
2
q
(M) .
Since the splittings (5.18) are
s
t
-invariant (in particular for s = 1), by
the ddenition of principal return trajectories (5.31), by the identity (5.34)
and the cocycle identity (5.35), the following holds:
(5.45)
E
q
_

(k)
q
(p)
_
= E
q

1
|t
k
|
_

(1)
q
k
(p)
_
=
=
1
|t
k
|
_
dU
(1)
q
k
(p)
_
= dU
(k)
q
(p)
Let s > 1 and := s 1. It follows from (5.43) (by the denition (4.107)
of Lyapunov norms) that, for -almost all q P and all q-regular p M,
(5.46) [E
q
_

(k)
q
(p)
_
[
s
[G
0
|t
k
|
_
U
(1)
q
k
(p)
_
[

K
s
4
(P) e
s|t
k
|
.
The crucial exponential estimate (5.29) on projections of principal return
trajectories, claimed above, can now be derived from estimates (5.32) and
(5.33) together with the remark that, for any l > 0, there exists a constant
C
l
> 0 such that, for -almost all q P and all q-regular p M,
(5.47) [
(k)
q
(p)
(k)
q
(p)[
1
C
l
e
l|t
k
|
.
By the denition of principal return trajectories and closing, the latter esti-
mate follows immediately from (5.41), since
(k)
q
(p)
(k)
q
(p) are currents
of integration along a vertical q-regular arc.
The required estimate (5.28) nally follows from estimate (5.29) by the
trajectory splitting Lemma 5.8. In fact, for any s > 1 there exist positive
real numbers
(1)
s
<
(2)
s
and a measurable map k
s
: P N such that

(1)
s
>
(2)
s

(1)
s
and, for -almost all q P,
(5.48)
(1)
s
k [t
k
(q)[
(2)
s
k , for all k k
s
(q) .
Let p
(k)
j
[1 k n, 1 j m
k

T
(p) be the sequence of q-regular
points constructed in Lemma 5.8. Since
s

(1)
s
(
(2)
s

(1)
s
) > 0, by
estimates (5.29) and (5.48) , there exists a constant K
s
5
(P) > 0 such that,
for -almost all q P, all q-regular p M and all T > 0,
(5.49)
n

k=ks(q)
e
|t
k+1
||t
k
|
[
s
q
_

(k)
q
(p
(k)
j
)
_
[
s
K
s
5
(P) ;
n

k=ks(q)
e
|t
k+1
||t
k
|
[E
s
q
_

(k)
q
(p
(k)
j
)
_
[
s
K
s
5
(P) .
SOBOLEV REGULARITY OF SOLUTIONS 105
In addition, by Lemma 5.8 there exists a constant K
6
(P) > 0 such that
(5.50) L
q
_
ks(q)

k=1
m
k

j=1

(k)
q
(p
(j)
k
)

K
6
(P)
ks(q)

k=1
e
|t
k+1
(q)|
,
hence the estimate (5.28) follow from estimates (5.49), from the trajectory
splitting Lemma 5.8, from estimate (5.50) and nally from Lemma 5.2,
which yields a bound on weighted Sobolev norms of (horizontal and verti-
cal) trajectories in terms of their R
q
-length.
5.3. The generic case. The above Theorem 5.9 implies, by a standard
Gottschalk-Hedlund argument, the following sharp result on the existence
of a Green operator for the (horizontal) cohomological equation in the case
of generic orientable quadratic differentials.
Theorem 5.10. For any SO(2, R)-absolutely continuous, KZ-hyperbolic
measure on a stratum M
(1)

of orientable quadratic differentials and for


-almost all q M
(1)

, there exists a densely dened, anti-symmetric Green


operator U
q
: H
0
q
(M) H
0
q
(M) for the horizontal cohomological equa-
tion S
q
u = f. For any s > 1, the maximal domain of the Green operator
U
q
on the Hilbert space H
0
q
(M) L
2
q
(M) contains the dense subspace
(5.51) [I
s
q
(M)]

:= f H
s
q
(M)[D(f) = 0 for all D I
s
q
(M) ,
and there exists a measurable function C
s

: M
(1)

R
+
such that the fol-
lowing holds. For any f [I
s
q
(M)]

, the Green solution U


q
(f) L

(M)
has zero-average and satises the estimate:
(5.52) [U
q
(f)[
L

(M)
C
s

(q) [f[
s
.
Proof. Let u
T

TR
be the 1-parameter family of measurable functions de-
ned almost everywhere on M as follows: for all q-regular p M,
(5.53) u
T
(p) :=
1
T
_
T
0
_

0
f F
q
(p, s) ds d .
By Theorem 5.9 (and Lemma 5.2) for any s > 1 there exists a measurable
function C
s

: M
(1)

R
+
such that, for -almost all q M
(1)

, for all
q-regular p M and all T > 0,
(5.54) [
T
q
(p)
s
q
_

T
q
(p)
_
[
s
C
s

(q) .
We recall that the projection
s
q
: W
s
q
(M) B
1
q
(M) is dened as the
composition of the orthogonal projection W
s
q
(M) Z
s
q
(M) with the pro-
jection Z
s
q
(M) B
1
q
(M) determined by the splitting (5.18).
106 GIOVANNI FORNI

If f [I
s
(M)]

, the 1-form
f
:= f
T
W
s
q
(M) is such that C() = 0
for any horizontally basic current C B
1
q
(M). It follows from the estimate
(5.54) that, for -almost all q M
(1)

, for all q-regular p M and all T > 0,


(5.55) [u
T
(p)[ [
T
q
()[ = [
T
q
(p)
s
q
_

T
q
(p)
_
()[ C
s

(q) [f[
s
,
hence the family u
T

TR
is uniformly bounded in the Hilbert space L
2
q
(M)
for -almost all q M
(1)

. In addition, a computation shows that if the


horizontal ow F
q
is ergodic, as T +,
(5.56) S
q
u
T
= f
1
T
_
T
0
f F
q
(, ) d f in L
2
q
(M) ,
hence any weak limit u L
2
q
(M) of the family u
T

TR
is a solution of the
cohomological equation S
q
u = f. Since any function f [I
s
(M)]

has
zero average, it follows by the denition (5.53) that u
T
has zero average for
all T > 0, hence any weak limit of the family u
T

TR
has zero average.
By the ergodicity of the horizontal ow F
q
, the cohomological equation
S
q
u = f has a unique zero average solution in U
q
(f) L
2
q
(M). It follows
that the operator U
q
: [I
s
q
(M)]

H
0
q
(M) L
2
q
(M) is well-dened and
linear, for -almost all q M
(1)

. In addition, by the uniform bound (5.55)


the function U
q
(f) L

(M) and satises the required bound (5.52).


It remains to be proven that the linear operator U
q
: H
0
q
(M) H
0
q
(M) is
anti-symmetric. For any f [I
s
q
(M)]

and v H
s+1
q
(M) H
s+1
q
(M),
since U
q
(f) L
2
q
(M) is a weak-solution of the equation S
q
u = f,
(5.57) U
q
(f), S
q
v)
q
= f, v)
q
= f, U
q
(S
q
v))
q
.
Since the subspace S
q
v H
s
q
(M)[v H
s+1
q
(M) is dense in the subspace
[I
s
q
(M)]

H
s
q
(M) and U
q
: [I
s
q
(M)]

L
2
q
(M) is continuous, it follows
from identity (5.57) by a density argument that
(5.58) U
q
(f), g)
q
= f, U
q
(g))
q
, for all f, g [I
s
q
(M)]

Optimal results on the regularity of solutions of the cohomological equation


for intermediate fractional regularity require smoothing and interpolation
techniques in the presence of distributional obstructions. The following def-
inition appears to capture the relevant condition on the obstructions which
allow for effective smoothing and interpolation results.
SOBOLEV REGULARITY OF SOLUTIONS 107
Denition 5.11. Let H
s
[s 0 be a 1-parameter family of normed spaces
such that the following embeddings hold:
(5.59) H

:=

s>0
H
s
H
s
H
r
H
0
, for all s r .
The order O
H
(D) (with respect to H
s
[s 0) of any linear functional
D (H

is the non-negative real number


(5.60) O
H
(D) := infs 0 [ D (H
s
)

.
A nite system of linear functionals D
1
, . . . , D
J
(H

will be called
-regular (with respect to the family H
s
[s 0) if, for any (0, 1]
there exists a dual systemu
1
(), . . . , u
J
() H

such that the following


estimates hold. For all 0 r and all > 0, there exists a constant
C

r
() > 0 such that, for all i, j 1, . . . , J,
(5.61) [u
j
()[
r
C

r
()
O
H
(D
j
)r
.
A nite system D
1
, . . . , D
J
(H
s
)

will be called regular if it is -


regular for any s. A nite dimensional subspace I (H

will be
called -regular [regular] if it admits a -regular [regular] basis.
It was proved in 1.4, in particular in Corollary 2.19, that for any s 0 the
closure H
s
q
(M)

H
s
q
(M) of the weighted Sobolev space as a subspace
of the Friedrichs weighted Sobolev space, is equal to the perpendicular
of a subspace D
s
q


H
s
q
(M), introduced in (2.94), of distributions sup-
ported on the singular set
q
M. It follows by Theorem 2.17 and by
Corollary 2.19 that all the spaces D
s
q
are regular with respect to the fam-
ily

H
s
q
(M)[s 0 of Friedrichs weighted Sobolev spaces. This regular-
ity result is the key to the existence of smoothing operators for the family
H
s
q
(M)[s 0 of weighted Sobolev spaces. In fact, their construction in
Theorem 2.24 is based on the regularity of the subspaces D
s
q


H
s
q
(M)
and on the existence of smoothings for the family of Friedrichs weighted
Sobolev spaces, which can be dened by appropriate truncations of the
Fourier expansion.
Our goal in this section is to implement a similar strategy in order to con-
struct smoothing families with values in the perpendicular of the subspaces
of invariant distributions. We prove below, under the hypothesis that the
Kontsevich-Zorich cocycle is non-uniformly hyperbolic, a preliminary re-
sult on the regularity of the spaces of horizontal [vertical] invariant distri-
butions with respect to the family of weighted Sobolev spaces. The basic
criterion for the regularity of spaces in invariant distributions is based on
Theorem 4.35 and the following notion:
108 GIOVANNI FORNI

Denition 5.12. Let be any G
t
-invariant ergodic probability measure on
any stratumM
(1)

M
(1)
g
of orientable holomorphic quadratic differentials.
For any q R

, a simple invariant distribution D

I
q
(M
q
) will be
called coherent (with respect to the family H
s
q
(M)[s 0) if its weighted
Sobolev order and its Lyapunov exponent are related:
O
H
q
(D

) = [l

(D

)[ .
A nite dimensional space of invariant distributions will be called coherent
(with respect to the family H
s
q
(M)[s 0) if it has a basis of simple
coherent elements.
By the denitions of coherence and regularity, the following result follows
immediately from Theorem 4.35:
Lemma 5.13. Let be any G
t
-invariant ergodic probability measure on
any stratum M
(1)

M
(1)
g
of orientable holomorphic quadratic differen-
tials. For any q R

, any coherent nite dimensional space of invariant


distributions is regular (with respect to the family H
s
q
(M)[s 0).
For any q M
(1)

and for any k N0, let J


k
q
(M) I
k
q
(M) be the space
of horizontally invariant distributions dened as follows:
(5.62) J
k
q
(M) :=
k1
j=0
L
j
Tq
_
I
1
q
(M)

.
Corollary 5.14. For any SO(2, R)-absolutely continuous, KZ-hyperbolic
measure on any stratum M
(1)

M
(1)
g
of orientable quadratic differen-
tials, for -almost all orientable quadratic differential q M
(1)

and for
any k N, the space J
k
q
(M) I
k
q
(M) is coherent, hence regular, with
respect to the family H
s
q
(M)[s 0 of weighted Sobolev spaces.
Proof. It follows immediately from Theorem 4.29 for the case k = 0 and
from Theorem 4.29 and Lemma 4.32 in the general case.
It was proved in Theorem 2.24 that the family H
s
q
(M)[s 0 of weighted
Sobolev spaces admits a family of smoothing operators. The existence of
a smoothing family together with the regularity of the distributional ob-
structions are the key elements of the interpolation theory for solutions of
the cohomological equation. We formalize below the basic construction of
smoothing projectors onto the perpendicular of a regular subspace.
Denition 5.15. Let H
s
[s 0 be a 1-parameter family of normed spaces
such that the embeddings (5.59) hold. A smoothing projection of degree
R
+
relative to a subspace I

(H

is a family P

()[ (0, 1]
of linear operators such that the operator P

() : H
0
[I

is
bounded for all (0, 1] and the following estimates hold. For any r,
SOBOLEV REGULARITY OF SOLUTIONS 109
s [0, ] and for any > 0, there exists a constant C

r,s
() > 0 such that,
for all u [I

(H
s
)

H
s
and for all (0, 1],
(5.63)
[P

()(u) u[
r
C

r,s
() [u[
s

sr
, if s > r ;
[P

()(u)[
r
C

r,s
() [u[
s

sr
, if s r .
A smoothing projection relative to the trivial subspace I

= 0 H

will
be called a smoothing of degree R
+
for the family H
s
[s 0.
The following result is a straightforward generalization of Theorem 2.24,
which implies the existence of smoothings of any nite degree for the family
H
s
q
(M)[s 0 of weighted Sobolev spaces.
Theorem 5.16. Let H
s
[s 0 be a 1-parameter family of normed spaces
such that the embeddings (5.59) hold. If the family H
s
[s 0 has a
smoothing of degree R
+
and the subspace I

(H

is -regular,
there exists a smoothing projection of degree R
+
relative to I

.
Proof. Let S

()[ (0, 1] be a smoothing of degree > 0 for the fam-


ily H
s
[s 0 and let D
1
, . . . D
J
be a -regular basis for the subspace
I

(H

. By the Denition 5.11 of regularity, for any (0, 1], there


exists a dual basis u
1
(), . . . , u
J
() H

such that the estimates (5.61)


hold for all 0 r and all > 0. For any u H
0
, we dene
(5.64) P

()(u) := S

()(u)
J

j=1
D
j
(S

()(u)) u
j
() .
We claim that the family P

()[ (0, 1] is a smoothing projection (of


degree > 0) relative to the subspace I

(H

. It follows immediately
from the denition that P

()(u) [I

for any u H
0
, hence
the operators P

() : H
0
[I

are well-dened, linear and


bounded. We claim that, for any s [0, ], for any j 1, . . . , J and
for any > 0, there exists a constant C
s
j
() > 0 such that, for all vectors
u [I

(H
s
)

H
s
, the following estimate holds:
(5.65) [D
j
(S

()(u)) [ C
s
j
() [u[
s

sO
H
(D
j
)
, for all (0, 1] .
In fact, if the order O
H
(D
j
) s, since D
j
(H
r
j
)

for any r
j
> O
H
(D
j
)
and the family S

()[ (0, 1] is a smoothing for H


s
[s 0, for any
> 0 there exists a constant A
s
j
() > 0 such that for all u H
s
and all
(0, 1],
(5.66) [D
j
(S

()(u)) [ [D
j
[

r
j
[S

()(u)[
r
j
A
s
j
() [u[
s

sr
j
/2
.
If O
H
(D
j
) < s, since D
j
(H
r
j
)

(H
s
)

, for any O
H
(D
j
) < r
j
< s
and the family S

()[ (0, 1] is a smoothing for H


s
[s 0, there
110 GIOVANNI FORNI

exists B
s
j
() > 0 such that, for all u [I

(H
s
)

and all (0, 1],


(5.67)
[D
j
(S

()(u)) [ = [D
j
(S

()(u) u) [
[D
j
[

r
j
[S

()(u) u[
r
j
B
s
j
() [u[
s

sr
j
/2
.
The estimate (5.65) follows immediately from estimates (5.66) and (5.67)
by taking r
j
(O
H
(D
j
), O
H
(D
j
) + /2) for all j 1, . . . , J in both
cases.
By estimate (5.65) (just proved) and by the estimates on u
1
(), . . . , u
J
()
in formula (5.61), for any r [0, ] and for any > 0, there exists a constant
C
(1)
r,s
() > 0 such that, for all u [I

(H
s
)

and all (0, 1],


(5.68)
J

j=1
[D
j
(S

()(u)) [ [u
j
()[
r
C
(1)
r,s
() [u[
s

sr
.
Finally, the required estimates (5.63) for the family P

()[ (0, 1], de-


ned in (5.64), follow from (5.68), since S

()[ (0, 1] is a smoothing


(of degree > 0) for the family H
s
[s 0.
The smoothness of the solutions of the cohomological equation is best ex-
pressed with respect to the following uniform norms.
Denition 5.17. For any k N, let B
k
q
(M) be the space of all functions
u H
k
q
(M) such that S
i
q
T
j
q
u = T
i
q
S
j
q
u L

(M) for all pairs of integers


(i, j) such that 0 i +j k. The space B
k
q
(M) is endowed with the norm
dened as follows: for any u B
k
q
(M),
(5.69) [u[
k,
:=
_

i+jk
[S
i
q
T
j
q
u[
2

_
1/2
=
_

i+jk
[T
i
q
S
j
q
u[
2

_
1/2
.
For s [k, k +1), let B
s
q
(M) := B
k
q
(M) H
s
q
(M) endowed with the norm
dened as follows: for any u B
s
q
(M),
(5.70) [u[
s,
:=
_
[u[
2
k,
+ [u[
2
s
_
1/2
.
Theorem 5.18. Let be a SO(2, R)-absolutely continuous, KZ-hyperbolic
measure on a stratum M
(1)

M
(1)
g
of orientable quadratic differentials.
For -almost all q M
(1)

and for any k N, the space I


k
q
(M) is coherent,
hence regular (with respect to the family H
s
q
(M)[s 0) and, for any
s > k, there exists a measurable function C
k,s

: M
(1)

R
+
such that
the following holds. For any function f [I
k
q
(M)]

H
s
q
(M) the Green
SOBOLEV REGULARITY OF SOLUTIONS 111
solution U
q
(f) of the cohomological equation S
q
u = f belongs to the
space B
k1
q
(M) and satises the estimates:
(5.71) [U
q
(f)[
k1
[U
q
(f)[
k1,
C
k,s

(q) [f[
s
Proof. For k = 1 the statement reduces to Theorem 5.10 and Corollary
5.14. Since f [J
2
q
(M)]

implies by denition that f and T


q
f [I
1
q
(M)]

,
by Theorem 5.10, for -almost all q M
(1)

and for any s > k = 2, the


cohomological equations S
q
u = f and S
q
u
T
= T
q
f have Green (zero
average) solutions u(f) and u
T
(f) L

(M) L
2
q
(M) which satisfy the
bounds,
(5.72)
[u(f)[

C
s1

(q) [f[
s1
,
[u
T
(f)[

C
s1

(q) [f[
s
.
It follows that the distribution u
T
(f) T
q
u(f) H
1
q
(M) is horizontally
invariant. Let D
1
, . . . , D
g
I
1
q
(M) be any regular basis such that
(5.73) O
H
q
(D
j
) = 1

j
, for all j 1, . . . , g .
It is no restrictive to assume that D
1
is the average. Since Green solutions
have zero average by denition, there exists F
2
(f), . . . F
g
(f) C such that
(5.74) u
T
(f) T
q
u(f) =
g

j=2
F
j
(f) D
j
.
It follows from the bounds (5.72) that the maps F
j
: [J
2
q
(M)]

C are
linear bounded functionals, for all j 2, . . . , g, dened on the closed
subspace [J
2
q
(M)]

H
s
q
(M). In fact, let u
1
(), . . . , u
g
() H
1
q
(M)
be any dual basis of the regular basis D
1
, . . . , D
g
as in Denition 5.11.
By Theorem 5.10, for any s > 2 and for any (
2
, . . . ,
g
) (0, 1]
g1
,
(5.75)
[F
j
(f)[ = [u
T
(f), u
j
(
j
))
q
[ + [u(f), T
q
u
j
(
j
))
q
[
C
s1

(q) [u
j
(
j
)[
0
[f[
s
+ [u
j
(
j
)[
1
[f[
s1
.
We claim that, for each j 2, . . . , g, the linear functional F
j
extends to
a horizontally invariant distribution
j
H
s
q
(M) such that
(5.76) O
H
q
(
j
) 1 +

j
, for all j 2, . . . , g .
In fact, since by Corollary 5.14 the space J
2
q
(M) is regular, by Theorem5.16
there exists a smoothing projection P

J
()[ (0, 1] of degree > 2
relative to the subspace J
2
q
(M) H
2
q
(M). Hence by denition, for any r,
s [0, ] and any > 0, there exists a constant C

r,s
() > 0 such that, for
112 GIOVANNI FORNI

all f [J
2
q
(M)]

H
s
q
(M) and for all (0, 1],
(5.77)
[P

J
()(f) f[
r
C

r,s
() [u[
s

sr
, if s > r ;
[P

J
()(f)[
r
C

r,s
() [f[
s

sr
, if s r .
Let (f
n
)
nN
be the sequence of functions dened as follows:
(5.78) f
n
:= P

J
(2
n
)(f) [J
2
q
(M)]

q
(M) , for all n N.
It follows from estimates (5.77) that, if s 1 < 1 +

j
< s
j
< s, for any
> 0, there exists a constant C

s
j
,s
() > 0 such that
(5.79)
[f
n+1
f
n
[
s1
C

s
j
,s
() [f[
s
j
2
n(s
j
s+1)
;
[f
n+1
f
n
[
s
C

s
j
,s
() [f[
s
j
2
n(ss
j
+)
.
Let u
(n)
1
, . . . , u
(n)
g
H
1
q
(M) be the sequence of dual basis of the regular
basis D
1
, . . . , D
g
I
1
q
(M) dened as follows: for each j 1, . . . , g,
(5.80) u
(n)
j
:= u
j
(2
n
) , for all n N.
By the estimate (5.61) in Denition (5.11) and by the identities (5.73), for
any > 0 there exists a constant C() > 0 such that, for all j 1, . . . , g,
(5.81)
[u
(n)
j
[
0
C() 2
n(1

j
)
;
[u
(n)
j
[
1
C() 2
n(

j
+)
.
For any s
j
> 1 +
j
, there exists s (2, 2 +

j
) and > 0 such that
(5.82) s
j
+ 1 s

j
2 :=
j
> 0 .
Hence by the estimate (5.61) and (5.75), for -almost q M
(1)

, there exists
a constant C
j
q
() := C
q
(, s, s
j
,
j
) > 0 such that
(5.83) [F
j
(f
n+1
f
n
)[ C
j
q
()2

j
n
[f[
s
j
.
By (5.77) and (5.78) the sequence f
n
f in H
s
q
(M) for any s < .
Since (5.75) holds for any s > 2, it follows from (5.83) that, for -almost
q M
(1)

,
(5.84) [F
j
(f)[
C
j
q
()
1 2

j
[f[
s
j
+ [P

J
(1)(f)[

.
Since the operator P

J
(1) : L
2
q
(M) H

q
(M) is bounded, we conclude that
for any s
j
> 1 +

j
there exists a continuous extension
s
j
j
H
s
j
q
(M)
of the linear functional F
j
: [J
2
q
(M)]

H
s
q
(M) C. By construction,
for any r
j
, s
j
> 1 +

j
, the extensions
r
j
j
=
s
j
j
(mod. J
2
q
(M)). Since
J
2
q
(M) is nite dimensional, for each j 2, . . . , g there exists a distri-
bution
j
H
2
q
(M), which extends the linear functional F
j
, such that
SOBOLEV REGULARITY OF SOLUTIONS 113
the Sobolev order O
H
q
(
j
) 1 +

j
as claimed in (5.76). Finally, we
prove that, by construction, the distributions
2
, . . . ,
g
H
2
q
(M) are hor-
izontally invariant. In fact, for any s > 2 and any v H
s+1
q
(M), the
function f
v
:= S
q
v [I
2
q
(M)]

H
s
q
(M), hence, by Theorem 5.10, the
cohomological equations S
q
u = f
v
and S
q
u
T
= T
q
f
v
have unique Green
solutions u(f
v
) and u
T
(f
v
) L

(M) respectively. By the ergodicity of


the horizontal foliation, since v and T
q
v L
2
q
(M) are also zero-average
solutions, the identities u(f
v
) = v and u
T
(f
v
) = T
q
v hold. It follows that
u
T
(f
v
) T
q
u(f
v
) = 0, hence by (5.74), for all j 2, . . . , g,
(5.85)
j
(S
q
v) = F
j
(S
q
v) = 0 , for all v H
s+1
q
(M) ,
thus
2
, . . .
g
I
s
q
(M) and the claim is completely proved.
For -almost all q M
(1)

and for all integers k 2, let



I
k
q
(M) I
k
q
(M)
be the subspaces dened as follows:
(5.86)

I
k
q
(M) :=
k2
_
h=0
L
h
Tq
_
J
2
q
(M)
g

j=2
C
j
_
.
It follows from the above construction that the following holds: for any
integer k 2, for any s > k and for any function f H
s
q
(M) [

I
k
q
(M)]

,
the Green solution U
q
(f) B
k1
q
(M) and satises the required estimate
(5.71). In fact, for k = 2 the statement follows by the construction of the
system of invariant distributions
2
, . . . ,
g
H
2
q
(M). For k 3, the
statement can be proved by induction. In fact, by the induction hypothesis,
for any s > k and any f H
s
q
(M)[

I
k
q
(M)]

, the cohomological equations


S
q
u = f has a unique solution u B
k2
q
(M) such that
(5.87) [u[
k2,
C
k1,s1

(q) [f[
s1
.
In addition, the function u
T
:= T
q
u H
k3
q
(M) is the unique solution
of the cohomological equation S
q
u
T
= T
q
f . Since T
q
f H
s1
q
(M)
[

I
k1
q
(M)]

, by the induction hypothesis, the following estimate holds:


(5.88) [T
q
u[
k2,
C
k1,s1

(q) [T
q
f[
s1
C
k1,s1

(q) [f[
s
.
Finally, by the Sobolev embedding theorem, there exists a continuous func-
tion

C

: M
(1)

R
+
such that the following estimate holds:
(5.89)
_
k1

i=1
[S
i
q
u[
2

_
1/2
=
_
k2

i=0
[S
i
q
f[
2

_
1/2

(q) [f[
s
.
The required estimate follows from (5.87), (5.88) and (5.89).
114 GIOVANNI FORNI

It remains to be proven that, for -almost q M
(1)

and all integers k 2,


the space I
k
q
(M) is coherent. From the above argument it follows that
(5.90) I
q
(M) =
+

k=2

I
k
q
(M) .
In fact, if f H

q
(M) is such that f [

I
k
q
(M)]

for all k N, there exists


u H

q
(M) such that S
q
u = f, hence f I
q
(M)

. The identity (5.90)


immediately implies that
(5.91) I
k
q
(M) =

I
k
q
(M) , for all integers k 2 .
The identities (5.91) in turn imply that the system
2
, . . . ,
g
constructed
above is linearly independent over J
2
q
(M), in particular
(5.92) dim
C
I
2
q
(M) = dim
C

I
2
q
(M) = 3g 2 .
In fact, by the above construction the following identity holds:
(5.93) I
q
(M) = I
1
q
(M) +
g

j=2
C
j
+ L
Tq
[I
q
(M)] .
Let D
q
: B
q
(M) I
q
(M) the isomorphism between the space B
q
(M) of
horizontally basic currents and the space of horizontally invariant distribu-
tions dened in (3.34). Let
q
: B
q
(M) B
q
(M) be the differential map
on the space of basic currents introduced in (3.61). From (5.93) it follows
immediately that
(5.94) B
q
(M) = B
1
q
(M) +
g

j=2
C D
1
q
(
j
) +
q
[B
q
(M)] .
By the structure theorem for (real) basic currents (Theorem 3.21), the coho-
mology map j
q
: B
q
(M) H
1
(M, R) vanishes on the space
q
[B
q
(M)].
On the other hand, by Corollary 3.20, for -almost all q M
(1)

, the co-
homology map on B
q
(M) has rank of codimension 1 in the homology
H
1
(M, R), hence of dimension 2g 1. Since B
1
q
(M) has dimension g
and the map D
q
: B
q
(M) I
q
(M) is an isomorphism, it follows that the
system
2
, . . . ,
g
is linearly independent over J
2
q
(M) and (5.92) holds.
We claim that, for -almost all q M
(1)

, the space

I
2
q
(M) = I
2
q
(M) is
coherent. For any Lyapunov exponent l < 0 of the cocycle
2
t
[I
2
q
(M),
let E
2
q
(l) I
2
q
(M) denote the corresponding Oseledec subspace and let
F
2
q
(l) E
2
q
(l) be the subspace of coherent distributions. Let
l
1
< < l
d
SOBOLEV REGULARITY OF SOLUTIONS 115
be the distinct Lyapunov exponents of the cocycle
2
t
[I
2
q
(M) on the the
Oseledec complement of the subspace J
2
q
(M). Since J
2
q
(M) is coherent
and
I
2
q
(M) =

I
2
q
(M) = J
2
q
(M)
d

i=1
E
2
q
(l
i
) ,
it is sufcient to prove the identities:
(5.95) F
2
q
(l
i
) = E
2
q
(l
i
) , for all i 1, . . . , d .
By Theorem 4.30 and Lemma 4.32, the Lyapunov spectrum of
2
t
[J
2
q
(M)
is the ordered set
0 >

2
1

g
1

2
2

g
2 .
Hence, by the description of the Lyapunov spectrum of
2
t
[I
2
q
(M) given
in Corollary 4.33, the set
l
1
, . . . , l
d
= 1

2
, . . . , 1

g
.
For any s 0, let I
q
(s) E
2
q
(l
1
) E
2
q
(l
d
) be the subset of horizontally
invariant distributions of Sobolev order less or equal to s 0. It follows
from Lemma 4.26 that the following inclusions hold:
(5.96) I
q
([l
i
[)

ji
E
2
q
(l
j
) , for all i 1, . . . , d .
By the estimate (5.76) on Sobolev orders of the distribuitons in the system

2
, . . . ,
g
, the following lower bounds hold:
(5.97) dim
C
I
q
([l
i
[)
i

j=1
dim
C
E
2
q
(l
j
) , for all i 1, . . . , d .
It follows from (5.96) and (5.97) that the inclusions in (5.96) are in fact
identities, for all i 1, . . . , d, and by Lemma 4.26 the claim (5.95) holds.
We have thus proved that the space I
2
q
(M) =

I
2
q
(M) is coherent. It follows
from denition (5.86) and Lemma 4.32 that the space

I
k
q
(M) is coherent for
any integer k 3. Since by (5.91) the identity I
k
q
(M) =

I
k
q
(M) holds for
all k 2, the space I
k
q
(M) is coherent and the proof is complete.
Theorem 5.19. Let be a SO(2, R)-absolutely continuous, KZ-hyperbolic
measure on a stratum M
(1)

M
(1)
g
of orientable quadratic differentials.
For -almost all q M
(1)

and for any s R


+
, the space I
s
q
(M) is coherent,
hence regular (with respect to the family H
s
q
(M)[s 0) and, for any
0 < r < s 1 there exists a measurable function C
r,s

: M
(1)

R
+
such
that the following holds. For any function f [I
s
q
(M)]

H
s
q
(M) the
116 GIOVANNI FORNI

Green solution U
q
(f) of the cohomological equation S
q
u = f belongs to
the space H
r
q
(M) and satises the estimates:
(5.98) [U
q
(f)[
r
C
r,s

(q) [f[
s
Proof. By Theorem 5.18, the subspace I
k
q
(M) is coherent for -almost all
q M
(1)

and for any k N. Since for any s < k the sub-bundle I


s
,+
(M)
I
k
,+
(M) is
k
t
-invariant, it follows that I
s
q
(M) is coherent, hence regular
by Lemma 5.13, for -almost all q M
(1)

.
By Theorem 5.16, for -almost all q M
(1)

there exists a smoothing


projection P

()[ (0, 1] of any given degree > 0 relative to the


subspace I

q
(M). Let s > 1 and let R
+
and k N be such that
> k + 1 s 1 > k. Let f [I
s
q
(M)]

H
s
q
(M). By Deni-
tion 5.15 of a smoothing projection, for any f [I
s
q
(M)]

H
s
q
(M), the
function P

()(f) [I

q
(M)]

q
(M) and satises the Sobolev bounds
(5.63). By Theorem 5.18 the cohomological equation S
q
u = P

()(f) has
a (unique) Green solution u() H
k1
q
(M) and there exists a measurable
function C

: M
(1)

R
+
such that , for all (0, 1],
(5.99)
[u() u(/2)[
k
C

(q) [P

()(f) P

(/2)(f)[

,
[u() u(/2)[
k1
C

(q) [P

()(f) P

(/2)(f)[
1
.
By the interpolation inequality proved in Lemma 2.10, for any r [k1, k]
there exists C
k,r
> 0 such that, for all (0, 1],
(5.100) [u()u(/2)[
r
C
k,r
[u()u(/2)[
kr
k1
[u()u(/2)[
rk+1
k
.
By the bounds (5.63), it follows from (5.99) and (5.99) that, for any > 0
there exists C

r,s
() > 0 such that
(5.101) [u() u(/2)[
r
C

r,s
() [f[
s

(kr)(s+1)

(rk+1)(s)
.
Since r < s 1, it is possible to choose R
+
, > 0 and k N so that
= (k r)(s + 1 ) + (r k + 1)(s ) > 0 .
It follows then from the bound (5.101), that the sequence u(2
n
) [ n N
is Cauchy in H
r
q
(M), hence it converges to a function u H
r
q
(M) of zero
average. Since P

()(f) f in H
s
q
(M) as 0
+
, it follows that the
function u H
r
q
(M) is the unique zero-average (Green) solution of the
cohomological equation S
q
u = f. The required Sobolev bound (5.98) also
follows from (5.101). In fact, by the interpolation inequality (Lemma 2.10),
by Theorem 5.18 and by the bounds (5.63) for the smoothing projection,
there exists a measurable function C

: M
(1)

R
+
such that,
(5.102) [u(1)[
r
C

(q) [f[
s
.
SOBOLEV REGULARITY OF SOLUTIONS 117
The bound (5.98) can then be derived from (5.101) and (5.102).
REFERENCES
[Ath06] J. Athreya, Quantitative Recurrence and Large Deviations for Teichmller
geodesic ow, Ph. D. thesis, University of Chicago, 2006.
[AV05] A. Avila and M. Viana, Simplicity of Lyapunov spectra: Proof of the Zorich-
Kontsevich conjecture, preprint on arXiv:math.DS/0508508, 2005.
[Ber74] L. Bers, Spaces of degenerating Riemann surfaces, Discontinuous groups and
Riemann surfaces, Proc. Conf. Univ. of Maryland, College Park, Md, 1973
(Princeton, NJ), Annals of Mathematical Studies, vol. 79, Princeton Univer-
sity Press, 1974, pp. 4355.
[BP05] L. Barreira and Y. Pesin, Smooth ergodic theory and non-uniformly hyper-
bolic dynamics, Handbook of Dynamical Systems (B. Hasselblatt and A. Ka-
tok, eds.), vol. 1B, Elsevier, 2005.
[CEG84] P. Collet, H. Epstein, and G. Gallavotti, Perturbations of geodesic ows on
surfaces of constant negative curvature and their mixing properties, Comm.
Math. Phys. 95 (1984), 61112.
[dlLMM86] R. de la Llave, J. M. Marco, and R. Moriyn, Canonical perturbation theory
of Anosov systems and regularity results for the Livsic cohomology equation,
Ann. of Math. 123 (1986), 537611.
[FF03] L. Flaminio and G. Forni, Invariant distributions and time averages for horo-
cycle ows, Duke Math. J. 119 (2003), no. 3, 465526.
[FF06] , Equidistribution of nilows and applications to theta sums, Ergodic
Theory Dynam. Systems 26 (2006), no. 02, 409433.
[FF07] , On the cohomological equation for nilows, Journal of Modern Dy-
namics 1 (2007), no. 1, 3760.
[FK92] H. M. Farkas and I. Kra, Riemann Surfaces, Springer-Verlag, New York, NY,
1992, second edition.
[For97] G. Forni, Solutions of the cohomological equation for area-preserving ows
on compact surfaces of higher genus, Ann. of Math. (2) 146 (1997), no. 2,
295344.
[For02] , Deviation of ergodic averages for area-preserving ows on surfaces
of higher genus, Ann. of Math. (2) 155 (2002), no. 1, 1103.
[For05] , On the Lyapunov exponents of the Kontsevich-Zorich cocycle, Hand-
book of Dynamical Systems (B. Hasselblatt and A. Katok, eds.), vol. 1B,
Elsevier, 2005.
[GH55] W. H. Gottschalk and G. A. Hedlund, Topological dynamics, AMS Collo-
quium Publications, vol. 36, AMS, Providence, R.I., 1955.
[GK80] V. Guillemin and Kazhdan, Some inverse spectral results for negaotively
curved 2-manifolds, Topology 19 (1980), 301312.
[Her83] M. R. Herman, Sur les courbes invariantes par les diffomorphismes de
lanneau, vol. 1, Asterisque, vol. 103-104, Socit Mathmatique de France,
Paris, France, 1983.
[HS05] P. Hubert and T. Schmidt, Afne diffeomorphisms and the Veech Dichotomy,
Handbook of Dynamical Systems (B. Hasselblatt and A. Katok, eds.),
vol. 1B, Elsevier, 2005.
118 GIOVANNI FORNI

[Kon97] M. Kontsevich, Lyapunov exponents and Hodge theory, The mathematical
beauty of physics, Saclay, 1996 (River Edge, NJ), Adv. Ser. Math. Phys.,
vol. 24, World Scientic, 1997, pp. 318332.
[KZ03] M. Kontsevich and A. Zorich, Connected components of the moduli space
of abelian differentials with prescribed singularities, Inv. Math. 153 (2003),
631678.
[Lan03] E. Lanneau, Connected components of the moduli spaces of quadratic differ-
entials, Thse de Doctorat, Universit de Rennes, 2003.
[Liv71] A. Livsic, Homology properties of U-systems, Math. Notes USSR Acad. Sci.
10 (1971), 758763, (in Russian).
[LM68] J. L. Lions and E. Magenes, Problmes aux limites non homognes et appli-
cations, vol. 1, Dunod, Paris, 1968.
[Mas82] H. Masur, Interval exchange transformations and measured foliations, Ann.
of Math. 115 (1982), 168200.
[Mas93] , Logarithmic law for geodesics in moduli space, Mapping Class
Groups and Moduli Spaces of Riemann surfaces, Proceedings, Gttingen
June-August 1991 (Providence, RI), Contemporary Mathematics, vol. 150,
AMS, 1993, pp. 229245.
[Mas05] , Ergodic Theory of Translation Surfaces, Handbook of Dynamical
Systems (B. Hasselblatt and A. Katok, eds.), vol. 1B, Elsevier, 2005.
[MMY03] S. Marmi, P. Moussa, and J.-C. Yoccoz, On the cohomological equation for
interval exchange maps, C. R. Math. Acad. Sci. Paris 336 (2003), 941948.
[MMY05] S. Marmi, P. Moussa, and J.-C. Yoccoz, The cohomological equation for Roth
type interval exchange maps, J. Amer. Math. Soc. 18 (2005), 823872.
[MT05] H. Masur and S. Tabachnikov, Rational billiards and at structures, Hand-
book of Dynamical Systems (B. Hasselblatt and A. Katok, eds.), vol. 1A,
Elsevier, 2005, pp. 10151089.
[Nag88] S. Nag, The Complex Analytic Theory of Teichmller Spaces, John Wiley &
Sons, Inc., New York, NY, 1988.
[Nel59] E. Nelson, Analytic vectors, Ann. of Math. (2) 70 (1959), 572615.
[Vee82] W. Veech, Gauss measures for transformations on the space of interval ex-
change maps, Ann. of Math. 115 (1982), 201242.
[Vee86] , The Teichmller geodesic ow, Ann. of Math. 124 (1986), 441530.
[Vee90] , Moduli spaces of quadratic differentials, J. DAnalyse Math. 55
(1990), 117171.
[Zor94] A. Zorich, Asymptotic ag of an orientable measured foliation on a surface,
pp. 479498, World Scientic, 1994.
[Zor96] , Finite Gauss measure on the space of interval exchange transfor-
mations. Lyapunov exponents, Ann. Inst. Fourier (Grenoble) 46 (1996), 325
370.
[Zor97] , Deviation for interval exchange transformations, Ergodic Th. and
Dynam. Syst. 17 (1997), 14771499.
[Zor99] , How Do the Leaves of a Closed 1-form Wind around a Surface?,
Pseudoperiodic Topology (Providence, RI) (M. Kontsevich V. I. Arnold and
A. Zorich, eds.), Amer. Math. Soc. Transl. (2), vol. 197, AMS, 1999, pp. 135
178.
SOBOLEV REGULARITY OF SOLUTIONS 119
DEPARTMENT OF MATHEMATICS, UNIVERSITY OF TORONTO, 40 ST. GEORGE ST.,
TORONTO, ONTARIO, CANADA M5S 2E4
E-mail address: forni@math.toronto.edu
LABORATOIRE DE MATHMATIQUES, UNIVERSIT DE PARIS-SUD, BTIMENT 425,
91405 ORSAY CEDEX, FRANCE
E-mail address: giovanni.forni@math.u-psud.fr