WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY

INTERESTING?

AMIE WILKINSON
arXiv:1608.02843v1 [[Link]] 9 Aug 2016

Introduction
At the 2014 International Congress of Mathematicians in Seoul, South Korea, Franco-
Brazilian mathematician Artur Avila was awarded the Fields Medal for “his profound
contributions to dynamical systems theory, which have changed the face of the field,
using the powerful idea of renormalization as a unifying principle.”1 Although it is not
explicitly mentioned in this citation, there is a second unifying concept in Avila’s work
that is closely tied with renormalization: Lyapunov (or characteristic) exponents. Lya-
punov exponents play a key role in three areas of Avila’s research: smooth ergodic theory,
billiards and translation surfaces, and the spectral theory of 1-dimensional Schrödinger
operators. Here we take the opportunity to explore these areas and reveal some under-
lying themes connecting exponents, chaotic dynamics and renormalization.
But first, what are Lyapunov exponents? Let’s begin by viewing them in one of their
natural habitats: the iterated barycentric subdivision of a triangle.
When the midpoint of each side of a triangle is connected to its opposite vertex
by a line segment, the three resulting segments meet in a point in the interior of the
triangle. The barycentric subdivision of a triangle is the collection of 6 smaller triangles
determined by these segments and the edges of the original triangle:

Figure 1. Barycentric subdivision.

As the process of barycentric subdivision starts with a triangle and produces triangles,
it’s natural to iterate the process, barycentrically subdividing the six triangles obtained
at the first step, obtaining 36 triangles, and so on, as in Figure 2.
Notice that as the subdivision gets successively finer, many of the triangles produced
by subdivision get increasingly eccentric and needle-like. We can measure the skinniness
of a triangle T via the aspect ratio α(T ) = area(T )/L(T )2 , where L(T ) is the maximum
Date: August 10, 2016.
1[Link]
1
2 AMIE WILKINSON

Figure 2. Iterating barycentric subdivision, from [54].

of the side lengths; observe that similar triangles have the same aspect ratio. Suppose
we fix a rule for labeling the triangles in a possible subdivision 1 through 6, roll a six-
sided fair die and at each stage choose a triangle to subdivide. The sequence of triangles
T1 ⊃ T2 ⊃ . . . obtained have aspect ratios α1 , α2 , . . ., where αn = α(Tn ).
Theorem 0.1 ([54], see also [16]). There exists a real number χ ≈ 0.0446945 > 0 such
that almost surely:
1
lim log αn = −2χ.
n→∞ n

In other words, if triangles are chosen successively by a random toss of the die, then
with probability 1, their aspect ratios will tend to 0 in the nth toss at an exponential
rate governed by exp(−2nχ). The same conclusion holds with the same value of χ if the
initial equilateral triangle is replaced by any marked triangle. This magical number χ is
a Lyapunov exponent. We return to this example at the end of the next section.
Lyapunov exponents make multiple appearances in the analysis of dynamical systems.
After defining basic concepts and explaining examples in Section 1, we describe in Sec-
tions 2–4 a sampling of Avila’s results in smooth ergodic theory, Teichmüller theory
and spectral theory, all of them tied to Lyapunov exponents in a fundamental way. We
explore some commonalities of these results in Section 5. Section 6 is devoted to a
discussion of some themes that arise in connection with Lyapunov exponents.
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 3

1. Cocycles, exponents and hyperbolicity

Formally, Lyapunov exponents are quantities associated to a cocycle over a measure-
preserving dynamical system. A measure-preserving dynamical system is a triple (Ω, µ, f ),
where (Ω, µ) is a probability space, and f : Ω → Ω is a measurable map that preserves
the measure µ, meaning that µ(f −1 (X)) = µ(X) for every measurable set X ⊂ Ω. We
say that (Ω, µ, f ) is ergodic if the only f -invariant measurable sets have µ-measure 0 or
1. Equivalently, (Ω, µ, f ) is ergodic if the only functions φ ∈ L2 (Ω, µ) satisfying φ◦f = φ
are the constant functions. Any f -invariant measure µ can be canonically decomposed
into ergodic invariant measures, a fact that allows us to restrict attention to ergodic
measures in some contexts, simplifying statements. The measures in such a decomposi-
tion are called ergodic components, and there can be uncountably many of them. The
process of ergodic decomposition is a bit technical to describe; we refer the reader to
[50] for more details.

1.1. Examples of measure-preserving systems. Here is a list of three types of

measure-preserving systems that we will refer to again in the sections that follow.

Rotations on the circle. On the circle Ω = R/Z, let fα (x) = x + α (mod 1), where
α ∈ R is fixed. The map fα preserves the Lebesgue-Haar measure µ (that assigns to an
interval I ∈ R/Z its length |I|). The map fα is ergodic with respect to µ if and only if
α is irrational. This has a straightforward proof: consider the equation φ ◦ fα = φ, for
some φ ∈ L2 (R/Z, µ), and solve for the Fourier coefficients of φ.
When α = p/q is rational, every point ω ∈ Ω is periodic, satisfying f q (ω) = ω. Each
ω then determines an ergodic fα -invariant probability measure νω obtained by averaging
the Dirac masses along the orbit of ω:
1 
νω := δω + δfα (ω) + · · · + δfαq−1 (ω) .
q
Each νω is an ergodic component of the measure µ, and hence there are uncountably
many such components.
When α is irrational, µ is the unique fα -invariant Borel probability measure. A
homeomorphism of a compact metric space that has a unique invariant measure is called
uniquely ergodic — a property that implies ergodicity and more. Unique ergodicity is
mentioned again in Section 2 and is especially relevant to the discussion of Schrödinger
operators with quasiperiodic potentials in Section 4.
There is nothing particularly special about the circle, and the properties of circle
rotations listed here generalize easily to rotations on compact abelian groups.

Toral automorphisms. Let Ω = T2 := R2 /Z2 , the 2-torus. Fix a matrix A ∈ SL(2, Z).
Then A acts linearly on the plane by multiplication and preserves the lattice Z2 , by virtue
of having integer entries and determinant 1. It therefore induces a map fA : T2 → T2 of
the 2-torus, a group automorphism. The area µ is preserved because det(A) = 1. Such
an automorphism is ergodic with respect to µ if and only if the eigenvalues of A are
λ and λ−1 , with |λ| > 1. This can be proved by examining the Fourier coefficients of
φ ∈ L2 (T2 , µ) satisfying φ ◦ fA = φ: composing with fA permutes the Fourier coefficients
of φ by the adjoint action of A, and the assumption on A implies that if φ is not constant
4 AMIE WILKINSON

there must be infinitely many identical coefficients, which violates the assumption that
φ ∈ L2 (T2 , µ).2
In contrast to the irrational rotation fα , the map fA has many invariant Borel prob-
ability measures, even when fA is ergodic with respect to the area µ. For example, as
we have just seen, every periodic point of fA determines an ergodic invariant measure,
and fA has infinitely many periodic points. This is a simple consequence of the Pigeon-
hole Principle, using the fact that A ∈ SL(2, Z): for every natural number q, the finite
collection of points {( pq1 , pq2 ) : p1 , p2 ∈ {0, . . . , q − 1}} ⊂ T2 is permuted by fA , and so
each element of this set is fixed by some power of fA .
Bernoulli shifts. Let Ω = {1, . . . , k}N be the set of all infinite, one sided strings
ω = (ω1 , ω2 , · · · ) on the alphabet {1, . . . , k}. Endowed with the product topology, the
space Ω is compact, metrizable, and homeomorphic to a Cantor set. The shift map
σ : Ω → Ω is defined by σ(ω)k = ωk+1 . In other words, the image of the sequence
ω = (ω1 , ω2 , · · · ) is the shifted sequence σ(ω) = (ωP 2 , ω3 , · · · ). Any nontrivial probability
vector p = (p1 , . . . , pk ) (i.e. with pi ∈ (0, 1), and i pi = 1)
defines a product measure
µ = pN supported on Ω.3 The triple Ω, σ, µ = (p1 , . . . , pk )N is called a Bernoulli shift,
and µ is called a Bernoulli measure. It is not hard to see that the shift σ preserves µ
and is ergodic with respect to µ.
The shift map σ : Ω → Ω manifestly has uncountably many invariant Borel probabil-
ity measures, in particular the Bernoulli measures (p1 , . . . , pk )N , but the list does not
end there. In addition to periodic measures (supported on orbits of periodic strings
(ω1 , . . . , ωq , ω1 , . . . , ωq , . . .)), there are σ-invariant probability measures on Ω encoding
every measure preserving continuous dynamical system4 — in this sense the shift is a
type of universal dynamical system.

1.2. Cocycles. Let Md×d be the d2 -dimensional vector space of d × d matrices (real or
complex). A cocycle is a pair (f, A), where f : Ω → Ω and A : Ω → Md×d are measurable
maps. We also say that A is a cocycle over f . For each n > 0, and ω ∈ Ω, we write
A(n) (ω) = A(f n−1 (ω))A(f n−2 (ω)) · · · A(f (ω))A(ω),
where f n denotes the n-fold composition of f with itself. For n = 0, we set A(n) (ω) = I,
and if both f and the values of the cocycle A are invertible, we also define, for n ≥ 1:
A(−n) (ω) = (A(n) (f −n+1 (ω)))−1 = A−1 (f −n+1 (ω)) · · · A−1 (ω).
One comment about the terminology “cocycle:” while A is colloquially referred to
as a cocycle over f , to fit this definition into a proper cohomology theory, one should
reserve the term cocycle for the function (n, ω) → A(n) (ω) and call A the generator of
this (1-)cocycle. See [11] for a more thorough discussion of this point.

2In higher dimensions, a matrix A ∈ SL(d, Z) similarly induces an automorphism f of Tn := Rn /Zn .

A
The same argument using Fourier series shows that fA is ergodic if and only if A does not have a root
of 1 as an eigenvalue.
3The product measure has a simple description in this context: setting C (j) := {ω ∈ Ω : ω = j}, the
i i
measure µ = (p1 , . . . , pk )N is defined by the properties: µ(Ci (j)) = pj , and µ(Ci (j) ∩ Ci0 (j 0 )) = pj pj 0 ,
for any i 6= i0 ∈ N and j, j 0 ∈ {1, . . . , k}.
4Subject to some constraints involving invertibility and entropy.
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 5

A fruitful way of viewing a cocycle A over f is as a hybrid dynamical system (f, A) : Ω×

Md×d → Ω × Md×d defined by
(f, A)(ω, B) = (f (ω), A(ω)B).
Note that the nth iterate (f, A)n of this hybrid map is the hybrid map (f n , A(n) ). The
vector bundle Ω × Md×d can be reduced in various ways to obtain associated hybrid
systems, for example, the map (f, A) : Ω × Rd → Ω × Rd defined by (f, A)(ω, v) =
(f (ω), A(ω)v). Thus a natural generalization of a cocycle over f is a map F : B → B,
where π : B → Ω is a vector bundle, and F acts linearly on fibers, with π ◦ F = f ◦
π. We will use this extended definition of cocycle to define the derivative cocycle in
Subsection 1.4.
1.3. Lyapunov exponents. Let f : Ω → Ω be a measurable map (not necessarily pre-
serving a probability measure). We say that a real number χ is a Lyapunov exponent
for the cocycle A : Ω → Md×d over f at the point ω ∈ Ω if there exists a nonzero vector
v ∈ Rd , such that
1
(1) lim log kA(n) (ω)vk = χ.
n→∞ n
Here k · k is a fixed norm on the vector space space Md×d . The limit in (1), when it
exists, does not depend on the choice of such a norm (exercise).
Oseledets proved in 1968 [55] that if (f, Ω, µ) is a measure-preserving system, and A
is a cocycle over f satisfying the integrability condition log kAk ∈ L1 (Ω, µ), then for
µ-almost every ω ∈ Ω and for every nonzero v ∈ Rd the limit in (1) exists. This limit
assumes at most d distinct values χ1 (ω) > χ2 (ω) > · · · > χk(ω) (ω). Each exponent χi (ω)
is achieved with a multiplicity di (ω) equal to the dimension of the space of vectors v
Pk(ω)
satisying (1) with χ = χi , and these multiplicities satisfy i=1 di (ω) = d.
If the cocycle A takes values in SL(d, R), then, since log det(A(ω)) ≡ log(1) = 0,
Pk(ω)
we obtain that i=1 di (ω)χi (ω) = 0. Thus if A takes values in SL(2, R), then the
exponents are of the form −χ(ω) ≤ 0 ≤ χ(ω).
If (Ω, µ, f ) is ergodic, then the functions k(ω), χi (ω) and di (ω), are constant µ-almost
everywhere. In this case, the essential values χ1 , . . . , χk ∈ R are called the Lyapunov
exponents of A with respect to the ergodic measure µ.
1.4. Two important classes of cocycles. Random matrix cocycles encode the behav-
ior of a random product of matrices. Let {A1 , . . . , Ak } ⊂ Md×d be a finite collection
of matrices. Suppose we take a k-sided die and roll it repeatedly. If the die comes up

Figure 3. A 30-sided Dungeons and Dragons die.

with the number j, we choose the matrix Aj , thus creating a sequence Aω1 , Aω2 , . . .,
where ω = (ω1 , ω2 , . . .) ∈ {1, . . . , k}N . This process can be packaged in a cocycle A over
6 AMIE WILKINSON

a measure preserving system (Ω, µ, σ) by setting Ω = {1, . . . , k}N , µ = (p1 , . . . , pk )N ,

where pj is the probability that the die shows j on a roll, and setting σ to be the shift
map. The cocycle is defined by A(ω) = Aω1 . Then A(n) (ω) is simply the product of the
first n matrices produced by this process.
More generally, suppose that η is a probability measure on the set of matrices Md×d .
The space Ω = Md×dN of sequences (M , M , . . . ) carries the product measure η N , which is
1 2
invariant under the shift map σ, where as above σ(M1 , M2 , . . . ) = (M2 , M3 , . . . ). There
is a natural cocycle A : Ω → Md×d given by A((M1 , M2 , . . . )) = M1 . The matrices
A(n) (ω), for ω ∈ Ω are just n-fold random products of matrices chosen independently
with respect to the measure η.
In the study of smooth dynamical systems the derivative cocycle is a central player.
Let f : M → M be a C 1 map on a compact d-manifold M . Suppose for simplicity that
the tangent bundle is trivial: T M = M × Rd . Then for each x ∈ M , the derivative
Dx f : Tx M → Tf x M can be written as a matrix Dx f ∈ Md×d . The map x 7→ Dx f is
called the derivative cocycle. The Chain Rule implies that if A = Df is a derivative
cocycle, then Dx f n = A(n) (x).
The case where T M is not trivializable is easily treated: either one trivializes T M over
a suitable subset of M , or one expands the definition of cocycle as described at the end
of Subsection 1.2: the map Df : T M → T M is an automorphism of the vector bundle
T M , covering the map f . Lyapunov exponents for the derivative cocycle are defined
analogously to (1). We fix a continuous choice of norms {k · kx : Tx M → R≥0 : x ∈ M },
for example the norms given by a Riemannian metric (more generally, such a family of
norms is called a Finsler). Then χ is a Lyapunov exponent for Df at x ∈ M if there
exists v ∈ Tx M such that
1
(2) lim log kDx f n vkf n (x) = χ.
n→∞ n
Since M is compact, the Lyapunov exponents of Df do not depend on the choice of
Finsler. The conclusions of Oseledets’s theorem hold analogously for derivative cocycles
with respect to any f -invariant measure on M .
A simple example of a derivative cocycle is provided by the toral automorphism
fA : T2 → T2 described above. Conveniently, the tangent bundle to T2 is trivial, and the
derivative cocycle is the constant cocycle Dx fA = A.

1.5. Uniformly hyperbolic cocycles. A special class of cocycles whose Lyapunov ex-
ponents are nonzero with respect to any invariant probability measure are the uniformly
hyperbolic cocycles.
Definition 1.1. A continuous cocycle A : Ω → Md×d over a homeomorphism f : Ω → Ω
of a compact metric space Ω is uniformly hyperbolic if there exists an integer n ≥ 1,
and for every ω ∈ Ω, there is a splitting Rd = E u (ω) ⊕ E s (ω) into subspaces that depend
continuously on ω, such that for every ω ∈ Ω:
(i) A(ω)E u (ω) = E u (f (ω)), and A(ω)E s (ω) = E s (f (ω)),
(ii) v ∈ E u (ω) =⇒ kA(n) (ω)vk ≥ 2kvk, and
(iii) v ∈ E s (ω) =⇒ kA(−n) (ω)vk ≥ 2kvk.
The definition is independent of choice of norm k · k; changing norm on Md×d simply
changes the value of n. The number 2 in conditions (ii) and (iii) can be replaced by
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 7

any fixed real number greater than 1; again this only changes the value of n. Notice
that measure plays no role in the definition of uniform hyperbolicity. It is a topological
property of the cocycle. For short, we say that (f, A) is uniformly hyperbolic.
A trivial example of a uniformly hyperbolic cocycle is the constant cocycle A(ω) ≡
A0 ∈ SL(2, R), where A0 is any matrix whose eigenvalues λ > λ−1 satisfy λ > 1. Here
the splitting R2 = E u (ω) ⊕ E s (ω) is the constant splitting into the sum of the λ and
λ−1 eigenspaces of A0 , respectively. For a constant cocycle, the Lyapunov exponents are
defined everywhere and are also constant; for this SL(2, R) cocycle, the exponents are
± log λ.
A nontrivial example of a uniformly hyperbolic cocycle is any nonconstant, continuous
A : Ω → SL(2, R) with the property that the entries of A(ω) are all positive, for any
ω ∈ Ω. In this case the splitting is given by
\ \
E u (ω) := A(n) (f −n (ω)) (C+ ) , and E u (ω) := A(−n) (f n (ω)) (C− ) ,
n≥0 n≥0

where C+ denotes the set of (x, y) ∈ R2 with xy ≥ 0, and C− is the set of (x, y) with
xy ≤ 0. For an example of this type, the Lyapunov exponents might not be everywhere
defined, and their exact values with respect to a particular invariant measure are not
easily determined, although they will always be nonzero where they exist (exercise). In
this example and the previous one, the base dynamics f : Ω → Ω are irrelevant as far as
uniform hyperbolicity of the cocycle is concerned.
Hyperbolicity is an open property of both the cocycle A and the dynamics f : if (f, A)
is uniformly hyperbolic, and fˆ and  and both uniformly close (i.e. in the C 0 metric) to
f and A, then (fˆ, Â) is uniformly hyperbolic. The reason is that, as in the example just
presented, uniform hyperbolicity is equivalent to the existence of continuously varying
cone families C+ (ω), C− (ω) ⊂ Rd , jointly spanning Rd for each ω ∈ Ω, and an integer
n ≥ 1 with the following properties: A(n) (ω) (C+ (ω)) ⊂ C+ (f n (ω)); A(−n) (ω) (C− (ω)) ⊂
C− (f −n (ω)); vectors in C+ (ω) are doubled in length by A(n) (ω); and vectors in C− (ω) are
doubled in length by A(−n) (ω). The existence of such cone families is preserved under
small perturbation.

1.6. Anosov diffeomorphisms. A diffeomorphism f : M → M whose derivative cocy-

cle is uniformly hyperbolic is called Anosov. Again, one needs to modify this definition
when the tangent bundle T M is nontrivial: the splitting of Rd in the definition is replaced
by a splitting T M = E u ⊕E s into subbundles — that is, a splitting Tx M = E u (x)⊕E s (x)
into subspaces, for each x ∈ M , depending continuously on x. The norm k · k on the
space Md×d is replaced by a Finsler, as in the discussion at the end of Subsection 1.4.
Since M is assumed to be compact, the Anosov property does not depend on the choice
of Finsler.
Anosov diffeomorphisms remain Anosov after a C 1 -small perturbation, by the open-
ness of uniform hyperbolicity of cocycles. More precisely, the C 1 distance dC 1 (f, g)
between two diffeomorphisms is the sum of the C 0 distance between f and g and the C 0
distance between Df and Dg; thus if f is Anosov and dC 1 (f, g) is sufficiently small, then
Dg is hyperbolic, and so g is Anosov. Such a g is often called a C 1 small perturbation
of f .
8 AMIE WILKINSON
 
2 1
The toral automorphism fA : → T2 T2 ,
with A = is Anosov; since the de-
1 1
rivative cocycle is constant, the splitting R2 = E u (x) ⊕ E s (x), for x ∈ T2 does not
depend on x: as above, E u (x)
√ is the expanding eigenspace for A corresponding to the
s
larger eigenvalue λ = (3 + 5)/2 > 1, and E (x) √ is the contracting eigenspace for A
corresponding to the smaller eigenvalue λ−1 = (3 − 5)/2 < 1. In this example, we can
choose n = 1 to verify that uniform hyperbolicity holds in the definition. The Lyapunov
exponents of this cocycle are ± log λ.
The map g : T2 → T2 given by
(3) g (x, y) = (2x + y +  sin(2π(x + y)), x + y)
is a C 1 small perturbation of fA if  is sufficiently small, and so g is Anosov for small
. Moreover, since det D(x,y) g ≡ 1, the map g preserves the area µ on T2 ; we shall see
in the next section that g is ergodic with respect to µ. The two Lyapunov exponents
of g with respect to the ergodic measure µ are ± log λ , where λ < λ for  6= 0 small.
There are several ways to see this: one way to prove it is to compute directly using the
ergodic theorem (Theorem 2.1) that  7→ λ is a smooth map whose local maximum is
achieved at  = 0.
In contrast to f , the exponents of g are not constant on T2 but depend on the in-
variant measure. For example, the averaged Dirac measures νx1 = δ(0,0) and νx2 =
 
1
3 δ 1 1 + δ
( , ) 1 + δ 1
(0, ) ( ,0) corresponding to the fixed point x1 = (0, 0) and the periodic
2 2 2 2
point x2 = ( 12 , 21 ) respectively, are both invariant and ergodic under g , for any . Di-
rect computation with the eigenvalues of the matrices Dx1 g and Dx2 g3 shows that the
Lyapunov exponents with respect to νx1 and νx2 are different for  6= 0.

1.7. Measurably (nonuniformly) hyperbolic cocycles. We say that a cocycle A

over (Ω, µ, f ) is measurably hyperbolic if for µ-a.e. point ω ∈ Ω, the exponents χj (ω)
are all nonzero. Since the role played by the measure is important in this definition, we
sometimes say that µ is a hyperbolic measure for the cocycle A, or A is hyperbolic with
respect to µ.
Uniformly hyperbolic cocycles over a homeomorphism f are hyperbolic with respect to
any f -invariant probability measure µ (exercise). An equivalent definition of measurable
hyperbolicity that neatly parallels the uniformly hyperbolic condition is: A is hyperbolic
with respect to the f -invariant measure µ if there exist a set X ⊂ Ω with µ(X) = 1 and
splittings Rd = E u (ω) ⊕ E s (ω) depending measurably on ω ∈ X, such that for every
ω ∈ X there exists an integer n = n(ω) ≥ 1 such that conditions (i)-(iii) in Definition 1.1
hold. The splitting into unstable and stable spaces is not necessarily continuous (or even
globally defined on Ω) and the amount of time n(ω) to wait for doubling in length to
occur depends on ω; for these reasons, measurably hyperbolic cocycles are often referred
to as “nonuniformly hyperbolic.” 5 In this nonuniform setting it is possible for a cocycle
to be hyperbolic with respect one invariant measure, but not another.

5The terminology is not consistent across fields. In smooth dynamics, a cocycle over a measurable sys-
tem that is measurably hyperbolic is called nonuniformly hyperbolic, whether it is uniformly hyperbolic
or not. In the spectral theory community, a cocycle is called nonuniformly hyperbolic if it is measurably
hyperbolic but not uniformly hyperbolic.
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 9

T b z

0 1
a
T !z

Figure 4. The upper half plane H is the space of marked triangles, up

to Euclidean similarity. The triangle T on the left corresponds to the
point z ∈ H on the right.

A measurably hyperbolic cocycle lurks behind random barycentric subdivision. The

random process generating the triangles in iterated barycentric subdivision can be en-
coded in a cocycle as follows. First, we identify upper half plane H ⊂ C with the space
of marked triangles (modulo Euclidean similarity) by sending a triangle T with vertices
cyclically labeled a, b, c to a point z ∈ H by rescaling, rotating and translating, sending
a to 0, b to 1 and c to z. See Figure 4.

T (5) T (4)

T (6) T (3)

T (1) T (2)

z
2(1+z)
3
1+z
3
0 1 1 0 1
2

T T (1)

Figure 5. A Möbius transformation that selects the first triangle in

barycentric subdivision.

Labelling cyclically the triangles T (1) , T (2) , . . . , T (6) in the subdivision as in Figure
5, the Möbius transformation B(z) = 2(z + 1)/3 sends the marked triangle T to the
marked triangle T (1) . The involutions pictured in Figure 6 generate the symmetric
group S3 , whose nontrivial elements we label P1 , . . . , P5 . A bit of thought shows that
10 AMIE WILKINSON

the transformations B, BP1 , . . . , BP5 are the 6 maps of the plane sending T to the
rescaled triangles in the subdivision.
Fixing an identification of the lower half plane with the upper half plane via z 7→ z,
the action of these 6 transformations on H are identified with the projective
√ action
√  of
2/ 6 2/√6
6 elements A1 , . . . , A6 of P GL(2, R), where B is identified with , P1
    0 3/ 6
1 0 0 1
is identified with , P2 is identified with , and P3 is identified with
1 −1 1 0
 
−1 1
.
0 1

P3
H

P2 P1

0 1 1
2

Figure 6. Three involutions generating the symmetric group S3 : inver-

sion in the circles |z| = 1 and |z − 1| = 1, and reflection across the line
Re(z) = 21 .

Thus random barycentric subdivision is governed by a random matrix cocycle over

a Bernoulli shift. If repeated rolls of the die generate the sequence ω1 , ω2 , . . . with

i
ωi ∈ {1, . . . , 6}, then the nth triangle Tn (ω) generated is the projective image of
1
(n)
under A (ω), and an exercise shows that the aspect ratio of Tn (ω) is given by:
α(Tn (ω)) = k(0, 1) · A(n) (ω)k−2 .
We thus obtain the formula
1 1
lim log α(Tn (ω)) = −2 lim log (0, 1) · A(n) (ω) ,

n→∞ n n→∞ n

and out pops the Lyapunov exponent χ = limn→∞ n1 log (0, 1) · A(n) (ω) .6

Numerical simulation gives the value χ ≈ 0.0446945, but the fact that this number is
positive follows from a foundational result of Furstenberg (stated precisely as a special
case in Theorem 6.1) that underlies some of Avila’s results as well. The upshot is that
a random product of matrices in SL(2, R) (or P GL(2, R)) cannot have exponents equal
to 0, except by design. In particular, if the matrices do not simultaneously preserve a
6As with eigenvalues, the Lyapunov exponents for left and right matrix multiplication coincide.
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 11

collection of one or two lines, and the group generated by the matrices is not compact,
then the exponents with respect to any nontrivial Bernoulli measure will be nonzero.
These conditions are easily verified for the barycentric cocycle. Details of this argument
about barycentric subdivision can be found in [54] and the related paper [16].
The barycentric cocycle is not uniformly hyperbolic, as can be seen by examining the
sequence of triangles generated by subdivision on Figure 2: at each stage it is always
possible to choose some triangle with aspect ratio bounded below, even though most
triangles in a subdivision will have smaller aspect ratio than the starting triangle. For
products of matrices in SL(d, R), uniform hyperbolicity must be carefully engineered,
but for random products, measurable hyperbolicity almost goes without saying.
A longstanding problem in smooth dynamics is to understand which diffeomorphisms
have hyperbolic derivative cocycle with respect to some natural invariant measure, such
as volume (See [19]). Motivating this problem is the fact that measurable hyperbolicity
produces interesting dynamics, as we explain in the next section.

2. Smooth ergodic theory

Smooth ergodic theory studies the dynamical properties of smooth maps from a statis-
tical point of view. A natural object of study is a measure-preserving system (M, vol, f ),
where M is a smooth, compact manifold without boundary equipped with a Riemannian
metric, vol is the volume measure of this metric, normalized so that vol(M ) = 1, and
f : M → M is a diffeomorphism preserving vol. It was in this context that Boltzmann
originally hypothesized ergodicity for ideal gases in the 1870’s. Boltzmann’s non-rigorous
formulation of ergodicity was close in spirit to the following statement of the pointwise
ergodic theorem for diffeomorphisms:
Theorem 2.1. If f is ergodic with respect to volume, then its orbits are equidistributed,
in the following sense: for almost every x ∈ M , and any continuous function φ : M → R:
1
Z
n−1


(4) lim φ(x) + φ(f (x)) + · · · φ(f (x)) = φ dvol.
n→∞ n M
As remarked previously, an example of an ergodic diffeomorphism is the rotation fα
on R/Z, for α irrational. In fact this transformation has a stronger property of unique
ergodicity, which is equivalent to the property that the limit in (4) exists for every
x ∈ R/Z.7 While unique ergodicity is a strong property, the ergodicity of irrational
rotations is fragile; the ergodic map fα can be perturbed to obtain the non-ergodic map
fp/q , where α ≈ p/q.
Another example of an ergodic diffeomorphism, at the opposite extreme of the rota-
tions in more than one sense,is theautomorphism fA of the 2-torus induced by multi-
2 1
plication by the matrix A = with respect to the area µ. In spirit, this example
1 1
is closely related to the Bernoulli shift, and in fact its orbits can be coded in such a way
as to produce a measure-preserving isomorphism with a Bernoulli shift. As observed in
the previous section, ergodicity of this map can be proved using Fourier analysis, but
there is a much more robust proof, due to Anosov [1], who showed that any C 2 Anosov
diffeomorphism preserving volume is ergodic with respect to volume.
7This is a consequence of Weyl’s equidistribution theorem and can be proved using elementary anal-
ysis. See, e.g. [37].
12 AMIE WILKINSON

2.1. Ergodicity of Anosov diffeomorphisms and Pesin theory. Anosov’s proof of

ergodicity has both analytic and geometric aspects. For the map fA , it follows several
steps:
(1) The expanding and contracting subbundles E u and E s of the splitting T T2 =
E u ⊕ E s are tangent to foliations W u and W s of T2 by immersed lines. These
lines are parallel to the expanding and contracting eigendirections of A and wind
densely around the torus, since they have irrational slope. The leaves of this pair
of foliations are perpendicular to each other, since A is symmetric.
(2) A clever application of the pointwise ergodic theorem (presented here as a special
case in Theorem 2.1) shows that any φ ∈ L2 (T2 , µ) satisfying φ ◦ f = φ is, up to
a set of area 0, constant along leaves of the foliation W u , and (again, up to a set
of area 0) constant along leaves of W s . This part of the argument goes back to
Eberhard Hopf’s study of geodesics in negatively curved surfaces in the 1930’s.
(3) Locally, the pair of foliations W u and W s are just (rotated versions of) intervals
parallel to the x and y axes. In these rotated coordinates, φ(x, y) is a measurable
function constant a.e. in x and constant a.e. in y. Fubini’s theorem then implies
that such a φ must be constant a.e. This conclusion holds in local charts, but
since T2 is connected, φ must be constant.
(4) Since any f -invariant function φ ∈ L2 (T2 , µ) is constant almost everywhere with
respect to µ, we conclude that fA is ergodic with respect to µ.
The same proof works for any smooth, volume-preserving Anosov diffeomorphism —
in particular, for the maps g defined in (3) — if one modifies the steps appropriately.
The foliations by parallel lines in step (1) are replaced by foliations by smooth curves
(or submanifolds diffeomorphic to some Rk , in higher dimension). Step (2) is almost the
same, since it uses only volume preservation and the fact that the leaves of W u and W s
are expanded and contracted, respectively. Step (3) is the most delicate to adapt and
was Anosov’s great accomplishment, and Step (4) is of course the same.
The ergodicity of Anosov diffeomorphisms is also a consequence of the much stronger
property of being measurably encoded by a Bernoulli shift. This so-called Bernoulli
property implies that Anosov diffeomorphisms are mixing with
R respect to volume, mean-
2 n 2
R
ing that for any L function φ, we have M φ · φ ◦ f → ( M φ) as n → ∞. Visually,
sets are mixed up by Anosov diffeomorphisms: see Figure 4, which explains why fA is
sometimes referred to as the “cat map.” The map g will similarly do a number on a cat.
The proof of the Bernoulli property for Anosov diffeomorphisms builds the Anosov-Hopf
proof of ergodicity. The Anosov-Hopf argument has the additional advantage that it can
be adapted to prove ergodicity for systems that are not Bernoulli.
As explained in Subsection 1.6, any C 1 -small perturbation of an Anosov diffeomor-
phism is Anosov, and C 2 Anosov diffeomorphisms preserving volume are ergodic. Hence
volume-preserving C 2 Anosov diffeomorphisms are stably ergodic: the ergodicity cannot
be destroyed by a C 1 small perturbation, in marked contrast with the irrational rotation
fα .
The Anosov condition thus has powerful consequences in smooth ergodic theory. But,
like uniformly hyperbolic matrix products, Anosov diffeomorphisms are necessarily con-
trived. In dimension 2, the only surface supporting an Anosov diffeomorphism is the
torus T2 , and conjecturally, the only manifolds supporting Anosov diffeomorphisms be-
long to a special class called the infra-nilmanifolds. On the other hand, every smooth
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 13

Figure 7. The action of fA on a cat, from [2]. A cat is drawn in a square

fundamental domain for T2 at the lower left. Its image under A is shown
in the parallelogram, and it is reassembled into another fundamental
domain to show its image under fA . The image of the cat under fA2 is
depicted at right.

manifold supports a volume-preserving diffeomorphism that is hyperbolic with respect

to volume, as was shown by Dolgopyat and Pesin in 2002 [29].
In the 1970’s Pesin [57] introduced a significant innovation in smooth ergodic theory:
a nonuniform, measurable analogue of the Anosov-Hopf theory. Under the assumption
that a volume preserving diffeomorphism f is hyperbolic with respect to volume, Pesin
showed that volume has at most countably many ergodic components with respect to
f . Starting with Oseledets’s theorem, and repeatedly employing Lusin’s theorem that
every measurable function is continuous off of a set of small measure, Pesin developed an
ergodic theory of smooth systems that has had numerous applications. Some limitiations
of Pesin theory are: first, that it begins with the hypothesis of measurable hyperbolicity,
which is a condition that is very hard to verify except in special cases, and second,
without additional input, measurable hyperbolicity does not imply ergodicity, as the
situation of infinitely many ergodic components can and does occur [28].

2.2. Ergodicity of “typical” diffeomorphisms. The question of whether ergodicity

is a common property among volume-preserving diffeomorphisms of a compact manifold
M is an old one, going back to Boltzmann’s ergodic hypothesis of the late 19th Century.
We can formalize the question by fixing a differentiability class r ∈ [1, ∞] and considering
the set Diff rvol (M ) of C r , volume-preserving diffeomorphisms of M . This is a topological
space in the C r topology, and we say that a property holds generically in Diff rvol (M )
(or C r generically, for short) if it holds for all f in a countable intersection of open and
dense subsets of Diff rvol (M ).8

8Since Diff r (M ) is a Baire space, properties that hold generically hold for a dense set, and two
vol
properties that hold generically separately hold together generically.
14 AMIE WILKINSON

Di↵1vol (M )

di↵eomorphisms with entropy 0

ergodic, measurably hyperbolic di↵eomorphisms

Figure 8. Generically, positive entropy implies ergodicity and measur-

able hyperbolicity.

Oxtoby and Ulam [56] proved in 1939 that a generic volume-preserving homeomor-
phism of a compact manifold is ergodic. At the other extreme, KAM (Kolmogorov-
Arnol’d-Moser) theory, introduced by Kolmogorov in the 1950’s [43], implies that er-
godicity is not a dense property, let alone a generic one, in Diff ∞
vol (M ), if dim(M ) ≥ 2.
r
The general question of whether ergodicity is generic in Diff vol (M ) remains open for
r ∈ [1, ∞), but we now have a complete answer for any manifold when r = 1 under the
assumption of positive entropy. Entropy is a numerical invariant attached to a measure
preserving system that measures the complexity of orbits. The rotation fα has entropy
0; the Anosov map fA has positive entropy log(λ). By a theorem of Ruelle, positivity of
entropy means that there is some positive volume subset of M on which the Lyapunov
exponents are nonzero in some directions.

Theorem 2.2 (Avila, Crovisier, Wilkinson [7]). Generically in Diff 1vol (M ), positive en-
tropy implies ergodicity and moreover measurable hyperbolicity with respect to volume.

See Figure 8. This result was proved in dimension 2 by Mañé-Bochi [51, 17] and
dimension 3 by M.A. Rodriguez-Hertz [58]. Positive entropy is an a priori weak form
of chaotic behavior that can be confined to an invariant set of very small measure, with
trivial dynamics on the rest of the manifold. Measurable hyperbolicity, on the other
hand, means that at almost every point all of the Lyapunov exponents of the derivative
cocycle Df are nonzero. Conceptually, the proof divides into two parts:
(1) C 1 generically, positive entropy implies nonuniform hyperbolicity. One needs
to go from some nonzero exponents on some of the manifold to all nonzero
exponents on almost all of the manifold. Since the cocycle and the dynamics are
intertwined, carrying this out is a delicate matter. This relies on the fact that
the C 1 topology is particularly well adapted to the problem. On the one hand,
constructing C 1 -small perturbations with a desired property is generally much
easier than constructing C 2 small perturbations with the same property. On the
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 15

other hand, many useful dynamical properties such as uniform hyperbolicity are
C 1 open.
(2) C 1 generically, measurable hyperbolicity (with respect to volume) implies ergod-
icity. This argument uses Pesin theory but adds some missing input needed to
establish ergodicity. This input holds C 1 generically. For example, the posi-
tive entropy assumption generically implies existence of a dominated splitting;
this means that generically, a positive entropy diffeomorphism is something in-
termediate between an Anosov diffeomorphism and a nonuniformly hyperbolic
one. There is a continuous splitting T M = E u ⊕ E s , invariant under the de-
rivative, such that for almost every x ∈ M , there exists an n = n(x) such
that for every ξ u ∈ E u (x), kDx f n (ξ u )k ≥ 2kξ u k, and for every ξ s ∈ E s (x),
kDx f −n (ξ s )k ≥ 2kξ s k.
The proof incorporates techniques from several earlier results, most of which have been
proved in the past 20 years [6, 18, 20, 59]. Also playing an essential technical role
in the argument is a regularization theorem of Avila: every C 1 diffeomorphism that
preserves volume can be C 1 approximated by a C 2 volume-preserving diffeomorphism
[5]. The fact that this regularization theorem was not proved until recently highlights
the difficulty in perturbing the derivative cocycle to have desired properties: you can’t
change Df without changing f too (and vice versa). This is why completely general
results analogous to Furstenberg’s theorem for random matrix products are few and far
between for diffeomorphism cocycles.

3. Translation surfaces
A flat surface is any closed surface that can be obtained by gluing together finitely
many parallelograms in R2 along coherently oriented parallel edges, as in Figure 9. Two

Figure 9. A flat surface with a distinguished “South,” also known as a

translation surface (courtesy Marcelo Viana). Parallel edges of the same
color are identified.
flat surfaces are equivalent if one can be obtained from the other by cutting, translat-
ing, and rotating. A translation surface is a flat surface that comes equipped with a
well-defined, distinguished vertical, “North” direction (or, “South” depending on your
orientation). Two translation surfaces are equivalent if one can be obtained from the
other by cutting and translating (but not rotating).
 16
Asymptotic cycles AMIE WILKINSON

Fix a translation surface Σ of genus g > 0. If one picks an angle θ and a point x
on Σ, and follows the corresponding straight ray through Σ, there are two possibilities:
either it terminates in a corner, or it can be continued for all time. Clearly for any θ,
and almost every starting point (with respect to area), the ray will continue forever. If
it continues forever, either it returns to the initial point and direction and produces a
Given
PSfrag any long geodesic segment in a given direction,
closed curve, or it continues on a parallel course without returning. A version of the
replacements
Pigeonhole Principle for area (Poincaré recurrence) implies that for almost every point
“close” it to get an element of :
and starting direction, the line will come back arbitrarily close to the starting point.

Figure 10. Closing up a ray that comes back close to itself (courtesy
Kerckhoff,Marcelo
Masur,Viana)Smillie: The geodesic flow in almost
every direction is uniquely ergodic.
Kerckhoff-Masur-Smillie [42] proved more: for a fixed Σ, and almost every θ, the ray
Thenthrough any point x is dense in Σ, and in fact is equidistributed with respect to area.
Such a directionconverges uniformly
θ is called uniquely ergodic, as it isto some
uniquely ergodic in the same sense
whenforthe
that fα length
is, for irrational α. , where the asymptotic
Suppose we start with a uniquely ergodic cycle
direction and wait
the successive times that this ray returns closer and closer to itself. This produces
depends only
a sequence on the
of closed curvessurface and athe
γn which produces direction.
sequence of cycles [γn ] in homology
H1 (Σ, Z) ' Z .
2g

Unique ergodicity of the direction θ implies that there is a unique c1 ∈ H1 (Σ, R) such
that for any starting point x:
Lyapunov exponents ofTeichmüller flows – p. 10/61
[γn ]
lim = c1 ,
n→∞ `(γn )

where `(γ) denotes the length in Σ of the curve γ.

Theorem 3.1 (Forni, Avila-Viana, Zorich [32, 13, 65, 66]). Fix a topological surface S
of genus g ≥ 1, and let Σ be almost any translation surface modelled on S.9 Then there
exist real numbers 1 > ν2 > . . . > νg > 0 and a sequence of of subspaces L1 ⊂ L2 ⊂ · · · Lg
of H1 (Σ, R) with dim(Lk ) = k such that for almost every θ, for every x, and every γ in
direction θ, the distance from [γ] to Lg is bounded, and
log dist([γ], Li )
lim sup = νi+1 ,
`(γ)→∞ log(`(γ))
for all i < g.

9“Almost any” means with respect to the Lebesgue measure on possible choices of lengths and di-
rections for the sides of the pentagon. This statement can be made more precise in terms of Lebesgue
measure restricted to various strata in the moduli space of translation surfaces.
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 17

This theorem gives precise information about the way the direction of [γn ] converges
to its asymptotic cycle c1 : the convergence has a “directional nature” much in the way
a vector v ∈ Rd converges to infinity under repeated application of a matrix
 
λ1 ∗ · · · ∗
 0 λ2 · · · ∗ 
A=  0 ··· ··· ∗ ,


0 0 · · · λd
with λ1 > λ2 > · · · λd > 1.

Figure 11. A local picture of the Teichmüller flow (courtesy Marcelo Viana).

The numbers νi are the Lyapunov exponents of the Kontsevich-Zorich (KZ) cocycle
over the Teichmüller flow. The Teichmüller flow Ft acts on the moduli space M of
translation surfaces (that is, translation surfaces modulo cutting and translation) by
stretching in the East-West direction and contracting in the North-South direction.
More precisely, if Σ is a translationsurface,then Ft (Σ) is a new surface, obtained by
et 0
transforming Σ by the linear map . Since a stretched surface can often be
0 e−t
reassembled to obtain a more compact one, it is plausible that the Teichmüller flow has
recurrent orbits (for example, periodic orbits). This is true and reflects the fact that the
flow Ft preserves a natural volume that assigns finite measure to M. The Kontsevich-
Zorich cocycle takes values in the symplectic group Sp(2g, R) and captures homological
data about the cutting and translating equivalence on the surface.
Veech proved that ν2 < 1 [60], Forni proved that νg > 0 [32], and Avila-Viana
proved that the numbers ν2 , ν3 , . . . , νg−1 are all distinct [13]. Zorich established the
connection between exponents and the so-called deviation spectrum, which holds in
greater generality [65, 66]. Many more things have been proved about the Lyapunov
exponents of the KZ cocycle, and some of their values have been calculated which are
(until recently, conjecturally) rational numbers! See [31, 25].
Zorich’s result reduces the proof of Theorem 3.1 to proving that the the exponents
ν1 , . . . , νg are positive and distinct. In the g = 1 case where Σ is a torus, this fact has
a simple explanation. The moduli space M is the set of all flat structures on the torus
(up to homothety), equipped with a direction. This is the quotient SL(2, R)/SL(2, Z),
which is the unit tangent bundle of the modular surface H/SL(2, Z). The (continuous
 t 
e 0
time) dynamical system Ft on Ω is given by left multiplication by the matrix .
0 e−t
The cocycle is, in essence, the derivative cocycle for this flow (transverse to the direction
of the flow) This flow is uniformly hyperbolic (i.e. Anosov), and its exponents are ν1 =
log(e) = 1 and −ν1 = log(e−1 ) = −1.
18 AMIE WILKINSON

Figure 12. The moduli space H2 /P SL(2, Z) of flat structures on the

torus, which is a punctured sphere with two cone points, correspond-
ing to the hexagonal and square lattices in R2 . Pictured at right is a
fundamental domain in H2 . (Courtesy Carlos Matheus.)

The proof for general translation surfaces that the exponents ν1 , . . . , νg are positive
and distinct is considerably more involved. We can nonetheless boil it down to some
basic ideas.
(1) The Teichmüller flow itself is nonuniformly hyperbolic with respect to a natural
volume (Veech [60]), and can be coded in a way that the dynamics appear almost
random.
(2) Cocycles over perfectly random systems (for example i.i.d. sequences of matrices)
have a tendency to have distinct, nonzero Lyapunov exponents. This was first
proved by Furstenberg in the 2 × 2 case [34] and later by Guivarc’h-Raugi [36]
(see also [35]).
(3) Cocycles over systems that are nonrandom, but sufficiently hyperbolic and with
good coding, also tend to have distinct, nonzero Lyapunov exponents. This
follows from series of results, beginning with Ledrappier in the 2 × 2 case [46],
and in increasing generality by Bonatti-Viana [22], Viana [61], and Avila-Viana
[12].

4. Hofstadter’s butterfly
Pictured in Figure 13 is the spectrum of the operator Hxα : `2 (Z) → `2 (Z) given by
[Hxα u](n) = u(n + 1) + u(n − 1) + 2 cos(2π(x + nα))u(n),
where x is a fixed real number called the phase, and α ∈ [0, 1] is a parameter called the
frequency. The vertical variable is α, and the horizontal variable is the spectral energy
parameter E, which ranges in [−4, 4]. We can read off the spectrum of Hxα by taking a
horizontal slice at height α; the black region is the spectrum.
In an influential 1976 paper, Douglas Hofstadter of Gödel, Escher, Bach fame dis-
covered this fractal picture while modelling the behavior of electrons in a crystal lattice
 From properties (i) and (iv), it follows that a call that unit cell the N th unit cell. Every unit cell
graph of the spectrum need only include values of has a "local variable" P which runs from 0 to 1.
& between + 4 and -4, and values of e in any unit in particular, P is defined to be the fractional part
interval. We shall look at the interval [0, 1]. Fur of rt, usually denoted as (a). At P=O and P= I,
thermore, as a consequence of pxoperties, the there is one band which stretches across the full
graph inside the above-defined rectangular region width of the cell, separating it from its upper and
must have two axes of reflection, namely the hor- lower neighbors; this band is therefore called a
izontal line z= &, and the vertical line &=0. A "
"cell wall. It turns out that eex'tain rational val-
plot of spectrum(o. ), with n along the vertical axis, ues of I3 play a very important role in the descrip-
WHAT ARE
appears in Fig. 1. (Only
LYAPUNOV EXPONENTS,
rational valuesAND WHY ARE
of a with tion THEY INTERESTING?
of the structure of a unit cell; these are 19
the
denominator less than 50 are shown. ) "pure cases"

FIG. 1. Spectrum inside

a unit cell. & is the hori-
zontal variable, ranging
between+4 and -4, and
p=(n) is the vertical vari-
able, ranging from 0 to 1.

Figure 13. Hofstadter’s butterfly, from [38].

under the force of a magnetic field [38]. This operator plays a central role in the theory
of the integer quantum Hall effect developed by Thouless et al., and, as predicted theo-
retically, the butterfly has indeed appeared in von Klitzing’s QHE experiments. Similar
operators are used in modeling graphene, and similar butterflies also appear in graphene
related experiments (see, e.g. [27]).
Some properties of the butterfly have been established rigorously. For example, Avila
and Krikorian proved:
Theorem 4.1 (Avila-Krikorian, [9]). For every irrational α ∈ [0, 1], the α-horizontal
slice of the butterfly has measure 0.
Their proof complements and thus extends the earlier result of Last [45], who proved
the same statement, but for a full measure set of α satisfying an arithmetic condition.
In particular, we have:
Corollary 4.2. The butterfly has measure 0.
Other properties of the butterfly, for example its Hausdorff dimension, remain un-
known.
The connection between the spectrum of this operator and cocycles is an interesting
one. Recall the definition of the spectrum of Hxα :
σ(Hxα ) := {E ∈ C : Hxα − E · Id is not invertible}.
The eigenvalues are those E so that the eigenvalue equation Hxα u = Eu admits `2 (Z)
solutions, i.e. those E such that Hxα − E · Id is not injective.
The following simple observation is key. A sequence (un : n ∈ Z) ⊂ CZ (not necessarily
in `2 (Z)) solves Hxα u = Eu if and only if
   
n un un+1
AE (fα (x)) = , n ∈ Z,
un−1 un
20 AMIE WILKINSON

where fα : R/Z → R/Z is the translation mentioned above, and

 
E − 2 cos(2πx) −1
(5) AE (x) = ,
1 0
which defines an SL(2, R)-cocycle over the rotation fα , an example of a Schrödinger
cocycle. Using the cocycle notation, we have
   
(n) u0 un
AE (x) = , n ∈ Z.
u−1 un−1
Now let’s connect the properties of this cocycle with the spectrum of Hxα . Suppose
for a moment that the cocycle AE over fα is uniformly hyperbolic, for some value of E.
Then for every x ∈ R/Z there is a splitting R2 = E u (x) ⊕ E s (x) invariant under cocycle,
(mn)
with vectors in E u (x) expanded under AE (x), and vectors in E s (x) expanded under
(−mn)
AE (x), both by a factor of 2m . Thus no solution u to Hxα u = Eu can be polynomially
bounded simultaneously in both directions, which implies E is not an `2 eigenvalue of
Hxα . It turns out that the converse is also true, and moreover:
Theorem 4.3 (R. Johnson, [39]). If α is irrational, then for every x ∈ [0, 1]:
(6) σ(Hxα ) = {E : AE is not uniformly hyperbolic over fα }.
For irrational α, we denote by Σα the spectrum of σ(Hxα ), which by Theorem 4.3
does not depend on x. Thus for irrational α, the set Σα is the α-horizontal slice of the
butterfly.
The butterfly is therefore both a dynamical picture and a spectral one. On the one
hand, it depicts the spectrum of a family of operators Hxα parametrized by α, and on
the other hand, it depicts, within a 2-parameter family of cocycles {(fα , AE ) : (E, α) ∈
[−4, 4] × [0, 1]}, the set of parameters corresponding to dynamics that are not uniformly
hyperbolic.
Returning to spectral theory, we continue to explore the relationship between spec-
trum and dynamics. If α is irrational, then fα is ergodic, and Oseledets’s theorem implies
that the Lyapunov exponents for any cocycle over fα take constant values over a full
measure set. Thus the Lyapunov exponents of AE over fα take two essential values,
− −
χ+ +
E ≥ 0, and χE ; the fact that det(AE ) = 1 implies that χE = −χE ≤ 0. Then either
+
AE is nonuniformly hyperbolic (if χE > 0), or the exponents of AE vanish.
Thus for fixed α irrational, the spectrum Σα splits, from a dynamical point of view,
into two (measurable) sets: the set of E for which AE is nonununiformly hyperbolic, and
the set of E for which the exponents of AE vanish. On the other hand, spectral analysis
gives us a different decomposition of the spectrum:
σ(Hxα ) = σac (Hxα ) ∪ σsc (Hxα ) ∪ σpp (Hxα )
where σac (Hxα ) is the absolutely continuous spectrum, σpp (Hx ) is the pure point spectrum
(i.e., the closure of the eigenvalues), and σsc (Hxα ) is the singular continuous spectrum.
All three types of spectra have meaningful physical interpretations. While the spectrum
σ(Hxα ) does not depend in x (since α is irrational), the decomposition into subspectra can
depend on x.10 It turns out that the absolutely continuous spectrum does not depend
on x, so we can write Σac,α for this common set.
10In fact, the decomposition is independent of a.e. x, just not all x.
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 21

The next deep relation between spectral theory and Lyapunov exponents is the fol-
lowing, which is due to Kotani:
Theorem 4.4 (Kotani, [44]). Fix α irrational. Let Z be the set of E such that the
Lyapunov exponents of AE over fα vanish. Let Z ess denote the essential closure of Z,
i.e. the closure of the Lebesgue density points of Z. Then
Σac = Z ess .
Thus Lyapunov exponents of the cocycle are closely related to the spectral type of the
operators Hx . For instance, Theorem 4.4 implies that if AE is nonuniformly hyperbolic
over fα for almost every E ∈ Σα , then Σac,α is empty: Hxα has no absolutely continuous
spectrum.
We remark that Theorems 4.3 and 4.4 hold for much broader classes of Schrödinger
operators over ergodic measure preserving systems. For a short and self-contained proof
of Theorem 4.3, see [64]. The spectral theory of one-dimensional Schrödinger operators
is a rich subject, and we’ve only scratched the surface here; for further reading, see the
recent surveys [40] and [26].
Avila’s very recent work, some of it still unpublished, provides further fascinating
connections of this type, linking the spectral properties of quasiperiodic operators with
analytic potentials to properties of the Lyapunov exponents of their associated cycles
[3, 4].

5. Spaces of dynamical systems and metadynamics

Sections 2, 3 and 4 are all about families of dynamical systems. In Section 2, the family
is the space of all volume preserving diffeomorphisms of a compact manifold M . This
is an infinite dimensional, non-locally compact space, and we have thrown up our hands
and depicted it in Figure 8 as a blob. Theorem 2.2 asserts that within a residual set of
positive entropy systems (which turn out to be an open subset of the blob), measurable
hyperbolicity (and ergodicity) is generic.
In Section 3, the moduli space M of translation surfaces can also be viewed as a
space of dynamical systems, in particular the billiard flows on rational polygons, i.e.,
polygons whose corner angles are multiples of 2π. In a billiard system, one shoots a ball
in a fixed direction and records the location of the bounces on the walls. By a process
called unfolding, a billiard trajectory can be turned into a straight ray in a translation
surface.11 The process is illustrated in Figure 14 for the square torus billiard.
The moduli space M is not so easy to draw and not completely understood (except
for g = 1). It is, however, a finite dimensional manifold and carries some nice structures,
which makes it easier to picture than Diff(M ). Theorem 3.1 illustrates how dynamical
properties of a meta dynamical system, i.e. the Teichmüller flow Ft : M → M, are tied
to the dynamical properties of the elements of M. For example, the Lyapunov exponents
of the KZ cocycle over Ft for a given billiard table with a given direction describe how
well an infinite billiard ray can be approximated by closed, nearby billiard paths.
In Section 4, we saw how the spectral properties of a family of operators {Hxα :
α ∈ [0, 1]} are reflected in the dynamical properties of families of cocycles {(fα , AE ) :
11Not every translation surface comes from a billiard, since the billiards have extra symmetries. But
the space of billiards embeds inside the space of translation surfaces, and the Teichmüller flow preserves
the set of billiards.
22 AMIE WILKINSON

Figure 14. Unfolding billiards in a square to get lines in a torus (cour-

tesy Diana Davis).

(E, α) ∈ [−4, 4] × [0, 1]}. Theorems about spectral properties thus have their dynamical
counterparts. For example, Theorem 4.3 tells us that the butterfly is the complement
of those parameter values where the cocycle (fα , AE ) is uniformly hyperbolic. Since
uniform hyperbolicity is an open property in both α and E, the complement of the
butterfly is open. Corollary 4.2 tells us that the butterfly has measure 0. Thus the set of
parameter values in the square that are hyperbolic form an open and dense, full-measure
subset. In fact, work of Bourgain-Jitomirskaya [24] implies that the butterfly is precisely
the set of parameter values (E, α) where the Lyapunov exponents of (fα , AE ) vanish for
some x.12 These results in some ways echo Theorem 2.2, within a very special family of
dynamics.
The Hofstadter butterfly is just one instance of a low-dimensional family of dynamical
systems containing interesting dynamics and rich structure. A similar picture appears in
complex dynamics,13 in the 1 (complex) parameter family of dynamical systems {pc (z) =
z 2 + c : c ∈ C}. The Mandelbrot set consists of parameters c for which the map pc has
a connected Julia set Jc :

Figure 15. The Mandelbrot Set

12which automatically means for all x in case of irrational α.
13
Another field in which Avila has made significant contributions, which we do not touch upon here.
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 23

Note that in this conformal context, uniform hyperbolicity of the derivative cocycle
of pc on Jc just means that there exists an n such that |(pnc (z))0 | ≥ 2, for all z ∈ Jc . It
is conjectured that the set of parameters c such that pc is uniformly hyperbolic on Jc is
(open and) dense in the Mandelbrot set.

6. Themes
We conclude by summarizing a few themes, some of which have come up in our
discussion.

Nonvanishing exponents sometimes produce chaotic behavior. The bedrock

result in this regard is Anosov’s proof [1] that smooth Anosov flows and diffeomorph-
isms are mixing (and in particular ergodic). Another notable result is Anatole Katok’s
proof [41] that measurable hyperbolicity of diffeomorphism with respect to some mea-
sure µ produces many periodic orbits — in particular, the number of orbits of period n
grows exponentially in n. Pesin theory provides a sophisticated tool for exploiting mea-
suable hyperbolicity to produce chaotic behavior such as mixing and even the Bernoulli
property.

Exponents can carry geometric information. We have not discussed it here, but
there are delicate relationships between entropy, exponents and Hausdorff dimension of
invariant sets and measures, established in full generality by Ledrappier-Young [47, 48].
The expository article [63] contains a clear discussion of these relationships as well as
some of the other themes discussed in this paper. The interplay between dimension,
entropy and exponents has been fruitfully exploited in numerous contexts, notably in
rigidity theory. Some examples are: Ratner’s theorem for unipotent flows, Elon Linden-
strauss’s proof of Quantum Unique Ergodicity for arithmetic surfaces, and the Einsiedler-
Katok-Lindenstrauss proof that the set of exceptions to the Littlewood conjecture has
Hausdorff dimension 0. See [53, 49, 30].

Vanishing exponents sometimes present an exceptional situation that can be

exploited. Both Furstenberg’s theorem and Kotani theory illustrate this phenomenon.
Here’s Furstenberg’s criterion, presented in a special case:
Theorem 6.1 (Furstenberg, [34]). Let (A1 , . . . , Ak ) ⊂ SL(2, R), and let G be the small-
est closed subgroup of SL(2, R) containing {A1 , . . . , Ak }. Assume that:
(1) G is not compact.
(2) There is no finite collection of lines ∅ 6= L ⊂ R2 such that M (L) = L, for all
M ∈ G.
Then for any probability vector p = (p1 , . . . pk ) on {1, . . . , k} with pi > 0, for all i,
there exists χ+ (p) > 0, such that for almost every ω ∈ {1, . . . , k}N (with respect to the
Bernoulli measure pN ):
1
lim log kA(n) (ω)k = χ+ .
n→∞ n

One way to view this result: if the exponent χ+ vanishes, then the matrices either
leave invariant a common line or pair of lines, or they generate a precompact group.
Both possibilities are degenerate and are easily destroyed by perturbing the matrices.
One proof of a generalization of this result [46] exploits the connections between entropy,
24 AMIE WILKINSON

dimension and exponents alluded to before. This result was formulated more completely
in a dynamical setting by [21] as an “Invariance Principle,” which has been further
refined and applied in various works of Avila and others. See for example [14, 10, 15].
For general SL(d, R) cocycles, vanishing of exponents is still an exceptional situation,
but even more generally, the condition di ≥ 2 — that an exponent has multiplicity
greater than 1 — is also exceptional. This statement was rigorously established for
random matrix products by Guivarc’h-Raugi [36] and undleries the Avila-Viana proof
of simplicity of spectrum for the KZ cocycle. See the discussion at the end of Section 3.
The same ideas play a role in Margulis’s proof of superrigidity for higher rank lattices
in semisimple Lie groups. See [52].

Continuity and regularity of exponents is a tricky business. In general, Lya-

punov exponents do not depend smoothly, or even continuously, on the cocycle. Un-
derstanding the exact relationship between exponents, cocycles, measures and dynamics
is an area still under exploration, and a few of Avila’s deepest results, some of them
still in preparation with Eskin and Viana, lie in this area. The book [62] is an excellent
introduction to the subject.

Acknowledgments. Effusive thanks to Artur Avila, Svetlana Jitomirskaya, Curtis

McMullen, Zhenghe Zhang, and Anton Zorich for patiently explaining a lot of math to
me, to Clark Butler, Kathryn Lindsay, Kiho Park, Jinxin Xue and Yun Yang for catching
many errors, and to Diana Davis, Carlos Matheus, Curtis McMullen and Marcelo Viana
for generously sharing their images. I am also indebted to Bryna Kra for carefully
reading a draft of this paper and suggesting numerous improvements.

References
[1] D. Anosov, Geodesic flows on closed Riemannian manifolds of negative curvature. Trudy Mat.
Inst. Steklov. 90 (1967).
[2] Arnold, V. I.; Avez, A., Ergodic problems of classical mechanics. Translated from the French
by A. Avez. W. A. Benjamin, Inc., New York-Amsterdam 1968.
[3] A. Avila, Global theory of one-frequency Schrödinger operators, Acta Math. 215 (2015), 1–54.
[4] A. Avila. KAM, Lyapunov exponents and spectral dichotomy for one-frequency Schrödinger
operators. In preparation.
[5] A. Avila, On the regularization of conservative maps. Acta Math. 205 (2010), 5–18.
[6] A. Avila, J. Bochi, Nonuniform hyperbolicity, global dominated splittings and generic prop-
erties of volume-preserving diffeomorphisms. Trans. Amer. Math. Soc. 364 (2012), no. 6,
2883–2907.
[7] A. Avila, S. Crovisier, A. Wilkinson, Diffeomorphisms with positive metric entropy, preprint.
[8] A. Avila, S. Jitomirskaya, The Ten Martini Problem, Ann. Math. 170 (2009), 303–342.
[9] A. Avila, R. Krikorian, Reducibility or non-uniform hyperbolicity for quasiperiodic
Schrödinger cocycles. Ann. Math. 164 (2006), 911-940.
[10] A. Avila, J, Santamaria, Jimmy, M. Viana, Holonomy invariance: rough regularity and appli-
cations to Lyapunov exponents. Astérisque 358 (2013), 13-74.
[11] A. Avila, J. Santamaria, M. Viana, A. Wilkinson, Cocycles over partially hyperbolic maps.
Astérisque 358 (2013), 1-12.
[12] A. Avila, M. Viana, Simplicity of Lyapunov spectra: a sufficient criterion. Port. Math. 64
(2007), 311-376.
[13] A. Avila, M. Viana, Simplicity of Lyapunov spectra: proof of the Zorich-Kontsevich conjecture.
Acta Math. 198 (2007), no. 1, 1-56.
[14] A. Avila, M. Viana, Extremal Lyapunov exponents: an invariance principle and applications.
Invent. Math. 181 (2010), no. 1, 115-189.
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 25

[15] A. Avila, M. Viana, A. Wilkinson, Absolute continuity, Lyapunov exponents and rigidity I:
geodesic flows. J. Eur. Math. Soc. 17 (2015), 1435-1462.
[16] I. Bárány, A. F. Beardon, T. K. Carne, Barycentric subdivision of triangles and semigroups
of Möbius maps, Mathematika 43 (1996), 165–171.
[17] J. Bochi, Genericity of zero Lyapunov exponents. Ergodic Theory Dynam. Systems 22 (2002),
no. 6, 1667–1696.
[18] J. Bochi, C 1 -generic symplectic diffeomorphisms: partial hyperbolicity and zero centre Lya-
punov exponents. J. Inst. Math. Jussieu 9 (2010), no. 1, 49–93.
[19] J. Bochi, M. Viana, Lyapunov exponents: how frequently are dynamical systems hyperbolic?
Modern dynamical systems and applications 271–297, Cambridge Univ. Press, Cambridge,
2004.
[20] J. Bochi, M. Viana, The Lyapunov exponents of generic volume-preserving and symplectic
maps. Ann. Math. 161 (2005), no. 3, 1423–1485.
[21] C. Bonatti, X. Gómez-Mont, M. Viana, Généricité d’exposants de Lyapunov non-nuls pour
des produits déterministes de matrices. Ann. Inst. H. Poincaré Anal. Non Linéaire 20 (2003),
no. 4, 579–624.
[22] C. Bonatti and M. Viana. Lyapunov exponents with multiplicity 1 for deterministic products
of matrices. Ergod. Th. & Dynam. Sys, 24:1295–1330, 2004.
[23] J. Bourgain and M. Goldstein, On nonperturbative localization with quasi-periodic potential,
Ann. Math. 152 (2000), 835–879.
[24] J. Bourgain, S. Jitomirskaya, Continuity of the Lyapunov exponent for quasiperiodic operators
with analytic potential. J. Statist. Phys. 108 (2002), 1203–1218.
[25] J. Chaika, A. Eskin, Every flat surface is Birkhoff and Oseledets generic in almost every
direction. J. Mod. Dyn. 9 (2015), 1–23.
[26] D. Damanik, Schrödinger operators with dynamically defined potentials: a survey, to appear:
Erg. Th. Dyn. Syst.
[27] C. Dean, L. Wang, P. Maher, C. Forsythe, F. Ghahari, Y. Gao, J. Katoch, M. Ishigami, P.
Moon, M. Koshino, T. Taniguchi, K. Watanabe, K. L. Shepard, J. Hone, P. Kim, Hofstadter’s
butterfly and the fractal quantum Hall effect in moiré superlattices Nature 497 (2013), 598–
602.
[28] D. Dolgopyat, H. Hu, Y. Pesin, An example of a smooth hyperbolic measure with countably
many ergodic components Proceedings of Symposia in Pure Mathematics 69 (2001) 95–106.
[29] D. Dolgopyat, Y. Pesin, Every compact manifold carries a completely hyperbolic diffeomor-
phism. Ergodic Theory Dynam. Systems 22 (2002), 409–435.
[30] M. Einsiedler, A. Katok, E. Lindenstrauss, Invariant measures and the set of exceptions to
Littlewood’s conjecture. Ann. Math. (2) 164 (2006), 513–560.
[31] A. Eskin, M. Kontsevich, A. Zorich, Sum of Lyapunov exponents of the Hodge bundle with
respect to the Teichmüller geodesic flow. Publ. Math. Inst. Hautes Études Sci. 120 (2014),
207–333.
[32] G. Forni, Deviation of ergodic averages for area-preserving flows on surfaces of higher genus.
Ann. Math. 155 (2002), 1–103.
[33] H. Furstenberg, H. Kesten, Products of random matrices. Ann. Math. Statist. 31 (1960) 457–
469.
[34] H. Furstenberg, Noncommuting random products. Trans. Amer. Math. Soc. 108 (1963) 377–
428.
[35] I. Goldsheĭd, G. Margulis, Lyapunov exponents of a product of random matrices. (Russian)
Uspekhi Mat. Nauk 44 (1989), no. 5(269), 13–60; translation in Russian Math. Surveys 44
(1989), no. 5, 11–71.
[36] Y. Guivarc’h, A. Raugi, Propriétés de contraction d’un semi-groupe de matrices inversibles.
Coefficients de Liapunoff d’un produit de matrices aléatoires indépendantes. Israel J. Math.
65 (1989), no. 2, 165–196.
[37] H. Helson, Harmonic analysis. Second edition. Texts and Readings in Mathematics, 7. Hin-
dustan Book Agency, New Delhi, 2010.
[38] D. Hofstadter, Energy levels and wavefunctions of Bloch electrons in rational and irrational
magnetic fields. Physical Review B 14 (1976) 2239–2249.
26 AMIE WILKINSON

[39] R. Johnson, Exponential dichotomy, rotation number, and linear differential operators with
bounded coefficients, J. Differential Equations 61 (1986), 54–78.
[40] S. Jitomirskaya, C. A. Marx, Dynamics and spectral theory of quasi-periodic Schrödinger-type
operators, to appear: Erg. Th. Dyn. Syst.
[41] A. Katok, Lyapunov exponents, entropy and periodic orbits for diffeomorphisms. Inst. Hautes
Études Sci. Publ. Math. 51 (1980), 137–173.
[42] S. Kerckhoff, H. Masur, J. Smillie, Ergodicity of billiard flows and quadratic differentials. Ann.
Math. 124 (1986), 293–311.
[43] A. Kolmogorov, Théorie générale des systèmes dynamiques et mécanique classique. Proceedings
of the International Congress of Mathematicians (Amsterdam 1954) Vol. 1, 315–333
[44] S. Kotani, Ljapunov indices determine absolutely continuous spectra of stationary random
one-dimensional Schrödinger operators, Stochastic Analysis (Katata/Kyoto, 1982), (North-
Holland Math. Library 32, North-Holland, Amsterdam)(1984), 225–247.
[45] Y. Last, Zero measure spectrum for the almost Mathieu operator, Comm. Math Phys. 164,
421-432 (1994).
[46] F. Ledrappier. Positivity of the exponent for stationary sequences of matrices. In Lyapunov
exponents (Bremen, 1984), volume 1186 of Lect. Notes Math., pages 56–73. Springer-Verlag,
1986.
[47] F. Ledrappier, L.S. Young, The metric entropy of diffeomorphisms. I. Characterization of
measures satisfying Pesin’s entropy formula. Ann. Math. 122 (1985), 509–539.
[48] F. Ledrappier, L.S. Young, The Metric Entropy of Diffeomorphisms. II. Relations between
Entropy, Exponents and Dimension. Ann. Math. 122 (1985), 540–574.
[49] E. Lindenstrauss, Invariant measures and arithmetic quantum unique ergodicity. Ann. Math.
163 (2006), 165–219.
[50] R. Mañé, Ergodic theory and differentiable dynamics. Ergebnisse der Mathematik und ihrer
Grenzgebiete Springer-Verlag, Berlin, 1987.
[51] R. Mañé, Oseledets’s theorem from the generic viewpoint. Proc. Int. Congress of Mathemati-
cians (Warszawa 1983) Vol. 2, 1259–76.
[52] G. A. Margulis, Discrete subgroups of semisimple Lie groups. Ergebnisse der Mathematik und
ihrer Grenzgebiete 17. Springer-Verlag, Berlin, 1991.
[53] D. W. Morris, Ratner’s theorems on unipotent flows. Chicago Lectures in Mathematics. Uni-
versity of Chicago Press, Chicago, IL, 2005.
[54] C. McMullen, Barycentric subdivision, martingales and hyperbolic geometry, preprint, 2011.
[55] V. I. Oseledets, A multiplicative ergodic theorem. Characteristic Ljapunov, exponents of dy-
namical systems. (Russian) Trudy Moskov. Mat. Obs̆c̆. 19 (1968) 179-210.
[56] J. Oxtoby, S. Ulam, Measure-preserving homeomorphisms and metrical transitivity. Ann.
Math. (2) 42 (1941), 874–920.
[57] Y. Pesin, Characteristic Ljapunov exponents, and smooth ergodic theory. Uspehi Mat. Nauk
32 (1977), no. 4 (196), 55–112, 287.
[58] M.A. Rodriguez-Hertz, Genericity of nonuniform hyperbolicity in dimension 3. J. Mod. Dyn.
6 (2012), no. 1, 121–138.
[59] M. Shub, A. Wilkinson, Pathological foliations and removable zero exponents. Invent. Math.
139 (2000), no. 3, 495–508.
[60] W. A. Veech, The Teichmüller geodesic flow, Ann. Math. 124 (1986), 441530.
[61] M. Viana. Almost all cocycles over any hyperbolic system have nonvanishing Lyapunov expo-
nents. Ann. Math., 167 (2008) 643–680.
[62] M. Viana, Lectures on Lyapunov exponents. Cambridge Studies in Advanced Mathematics,
145. Cambridge University Press, Cambridge, 2014.
[63] L.-S. Young, Ergodic theory of differentiable dynamical systems. Real and complex dynamical
systems, 293–336, NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., 464, Kluwer Acad. Publ.,
Dordrecht, 1995.
[64] Z. Zhang, Resolvent set of Schrödinger operators and uniform hyperbolicity,
arXiv:1305.4226v2(2013).
 WHAT ARE LYAPUNOV EXPONENTS, AND WHY ARE THEY INTERESTING? 27

[65] A. Zorich, How do the leaves of a closed 1-form wind around a surface, “Pseudoperiodic
Topology”, V. Arnold, M. Kontsevich, A. Zorich (eds.), Translations of the AMS, Ser.2, vol.
197, AMS, Providence, RI (1999), 135–178.
[66] A. Zorich, Asymptotic Flag of an Orientable Measured Foliation on a Surface, in “Geometric
Study of Foliations”, World Scientific Pb. Co., (1994), 479–498.