Lp -GENERIC COCYCLES HAVE ONE-POINT LYAPUNOV

SPECTRUM

ALEXANDER ARBIETO AND JAIRO BOCHI

Abstract. We show the sum of the first k Lyapunov exponents of linear

cocycles is an upper semicontinuous function in the Lp topologies, for any
1 ≤ p ≤ ∞ and k. This fact, together with a result from Arnold and Cong,
implies that the Lyapunov exponents of the Lp -generic cocycle, p < ∞, are all
equal.

1. Introduction
The Lyapunov exponents of products of random matrices, has been the focus
of much study since the pioneering work of Furstenberg and Kesten [9]. A major
problem in the theory is to describe the Lyapunov spectrum of “typical” cocycles.
In the case of products of independent identically distributed matrices, for exam-
ple, for “most” choices of the probability distribution, all Lyapunov exponents are
different, see e.g. [8, p. 78] or [10, theorem 6.9].
Here we consider the general setting of linear cocycles. Fix a probability space
(X, F, µ) and an ergodic invertible measure preserving transformation T : X →
X. Take a measurable map A : X → GL(d, R) (satisfying an integrability con-
dition). Then, by the theorem of Oseledets, the Lyapunov exponents of A (with
multiplicity)
λ1 (A) ≥ . . . ≥ λd (A)
are defined: they are the possible values of the limits
1
lim log kA(T n−1 x) · · · A(T x)A(x)vk,
n→±∞ n

for µ-almost every x ∈ X and v ∈ Rd r {0}.

One may then consider some class of cocycles A and ask “how large” is the
subset of those cocycles with simple spectrum (i.e., with all Lyapunov exponents
different), or with one-point spectrum (all exponents equal). By a “large” set, we
usually mean a residual (dense Gδ ) set, in some topology. Of course, the answer
to the vague question above depends on the class of the cocycles considered and
of the topology. For instance, Knill [11] proved that in the space of all bounded
measurable cocycles with values in SL(2, R), the set of cocycles with two different
Lyapunov exponents is dense in the uniform topology. This set contains the open
subset of uniformly hyperbolic cocycles. It was showed by Bochi [5] that the set
of cocycles that either are uniformly hyperbolic or have both exponents equal to
Date: November 1, 2002.
A. A. is supported by CNPq and J. B. is supported by CNPq-Profix.
1
2 A. ARBIETO, J. BOCHI

zero is a residual subset. This result also holds for continuous cocycles, while it
is an open question whether Knill’s theorem does.
Knill’s result was extended (through different techniques) to cocycles with val-
ues in GL(d, R), for any d ≥ 2, by Arnold and Cong [4]. Also, Bochi and Viana
generalized Bochi’s result to very general matrix groups in [6].
In this paper we consider the space of GL(d, R)-valued cocycles, endowed with
the weaker topologies Lp , 1 ≤ p < ∞ (see section 2). In this case, the set of
cocycles with simple Lyapunov spectrum is still dense, as proved by Arnold and
Cong [3]. But it ceases to contain an open set: they showed that the set of cocyles
with one-point spectrum is also a dense subset. In this note we show that the
set of cocycles with one-point spectrum is residual, see theorem 1 below. That is,
generic Lp -cocycles, p < ∞, have all Lyapunov exponents equal.
The results of Arnold and Cong imply that the Lyapunov exponents do not
depend continuously on the cocycle. However, we show that the Lyapunov expo-
nents have a semicontinuity property. Using this and Arnold and Cong’s density
result we obtain theorem 1.
It is interesting to mention that the typical behavior of the spectrum may
change drastically when cocycles more regular than just continuous are considered.
Bonatti and Viana [7] considered an open class of Hölder-continuous cocycles over
an uniformly hyperbolic base dynamics. In this setting, they showed that generic
cocyles have at least two different Lyapunov exponents. Even more, the set of
cocyles with one-point spectrum has infinite codimension. Some of the techniques
employed in [7] originate from the theory of products of i.i.d. matrices (e.g. [10]),
where, as we have mentioned, simple Lyapunov spectrum prevails.

2. Statement of the results

Let (X, F, µ) be a probability space and let T : X → X an automorphism
preserving the measure µ. Let d ≥ 2 be an integer and denote by G the set of
all (mod 0 equivalence classes of) measurable maps A : X → GL(d, R), where
GL(d, R) is endowed with the Borel σ-algebra.
We shall consider the set GIC of all maps A : X → GL(d, R) in G satisfying the
integrability condition:
Z
log+ kA±1 (x)k dµ(x) < +∞.
X

Under this condition, the multiplicative ergodic theorem of Oseledets (see [13]
or [1]) gives us the Lyapunov exponents (with multiplicity) λ1 (A, x) ≥ . . . ≥
λd (A, x) of the cocycle A.
Provided T is ergodic, these functions are constant almost everywhere. Then
we say that A ∈ GIC has one-point Lyapunov spectrum if all Lyapunov exponents
are equal, and we denote the set of those cocycles by GOPS . If the cocycle A takes
values in SL(d, R) then A ∈ GOPS if and only if all Lyapunov exponents are zero.
Let 1 ≤ p ≤ ∞. Following Arnold and Cong [3], we endow the set G with a
“Lp -like” topology.
 Lp -GENERIC COCYCLES HAVE ONE-POINT LYAPUNOV SPECTRUM 3

Let k·k be an operator norm on the set M(d, R) of d × d matrices. For any
measurable A : X → M(d, R), let
Z 1/p
p
kAkp = kA(x)k dµ(x) if p < ∞,
X

and
kAk∞ = ess sup kA(x)k.
x∈X
We have 0 ≤ kAkp ≤ ∞. Next, for A, B ∈ G, let
τp (A, B) = kA − Bkp + kA−1 − B −1 kp .
Then set
τp (A, B)
ρp (A, B) = .
1 + τp (A, B)
Here it is understood that
kA − Bkp = ∞ or kA−1 − B −1 kp = ∞ ⇔ ρp (A, B) = 1.
According to [3], ρp is a metric on G (and therefore on GIC ). Moreover, (G, ρp )
(and hence (GIC , ρp )) is complete.
Remark 1. By the Hölder inequality, we have ρ1 (A, B) ≤ ρp (A, B) for all A,
B ∈ G, 1 ≤ p ≤ ∞.
Remark 2. If A ∈ GIC and B ∈ G with ρp (A, B) < 1 then B ∈ GIC ; see [3].
Arnold and Cong proved in [3, theorem 4.5] that the set GOPS is dense in GIC
for any metric ρp , 1 ≤ p < ∞. We improve this result by showing:
Theorem 1. Assume T is ergodic. Then GOPS is a residual subset of GIC in the
Lp -topology, for any 1 ≤ p < ∞.
Let k = 1, . . . ,d. We will study the following quantities:
Z
Λk (A) = (λ1 (A, x) + · · · + λk (A, x)) dµ(x).
X
In fact, our main result is:
Theorem 2. Let 1 ≤ p ≤ ∞ and endow the set GIC with the metric ρp . Then:
(a) the maps Λk : GIC → R are upper semicontinuous for all k = 1, . . . , d, that is
for every A ∈ GIC and ε > 0 there exists 0 < δ < 1 such that if ρp (A, B) < δ
then Λk (B) < Λk (A) + ε;
(b) the function Λd : GIC → R is continuous.
From theorem 2 and Arnold and Cong’s density result, we shall deduce:
Theorem 3. Assume T is ergodic. Let 1 ≤ p < ∞, and endow the set GIC with
the metric ρp . Take k ∈ {1, . . . , d − 1}. Then the map Λk : GIC → R is continuous
at some A ∈ GIC if and only if A ∈ GOPS .
4 A. ARBIETO, J. BOCHI

Since the set of continuity points of any upper semicontinuous map is residual
(see [12]), theorem 1 is an immediate corollary of theorems 2 and 3.
Our theorems extend some of the results from [2]. For a description of the
continuity points of the functions Λk in the case of bounded or continuous cocycles,
see [6].

3. Proofs
We are going to uses some basic facts about exterior powers and Lyapunov
exponents, see e.g. [1, chapter 3].
Let k ∈ {1, . . . , d}. We denote by ∧k Rd the k-th exterior power of Rd . If
A : Rd → Rd is a linear map, then it induces a linear map ∧k A : ∧k Rd → ∧k Rd .
An inner product in Rd induces an inner product in ∧k Rd , and the corresponding
operator norms satisfy
(1) k∧k Ak ≤ kAkk , ∀ A ∈ M(d, R).
We fix operator norms as above from now on.
Lemma 4. Let A, B ∈ GL(d, R) and k ∈ {1, . . . , d}. Then
log+ k∧k Bk ≤ log+ k∧k Ak + kkB − Ak.
Proof. Using (1),
 1/k
k∧k Bk1/k ≤ k∧k Ak + k∧k (B − A)k
≤ k∧k Ak1/k + k∧k (B − A)k1/k
≤ k∧k Ak1/k + kB − Ak.

Since log+ (x + y) ≤ log+ x + y for all x, y ≥ 0, the lemma follows.

We denote, for A ∈ G, x ∈ X and n ∈ N,

An (x) = A(T n−1 x) · · · A(T x)A(x).
A basic property of Λk from which we shall deduce its upper semicontinuity is the
following: For all A ∈ GIC ,
Z Z
1 1
(2) Λk (A) = lim log k∧k An (x)k dµ(x) = inf log k∧k An (x)k dµ(x).
n→∞ n n n

In what follows, all integrals are meant with respect to the measure µ. We shall
also need the following measure-theoretic result:
Lemma 5. Given f ∈ L1 (µ) with f ≥ 0, and η > 0, there exists K > 0 such that
for all h in L1 (µ) with h ≥ 0 and kh − f k1 < η, we have
Z
2η
h < 2η and, consequently, µ({h > K}) < .
{h>K} K
 Lp -GENERIC COCYCLES HAVE ONE-POINT LYAPUNOV SPECTRUM 5

Proof. Since f ≥ 0 is integrable, one can find γ > 0 such that

Z
Z ⊂ X measurable, µ(Z) < γ ⇒ f < η.
Z
Let K = γ −1 (kf k
1 + η). Given h ∈ L1 (µ) with h ≥ 0 and kh − f k1 < η, let
Eh = {h ≤ K}. Then
µ(Ehc ) ≤ K −1 khk1 < K −1 (kf k1 + η) = γ.
Therefore Z Z Z
h≤ f+ |h − f | ≤ η + kh − f k1 < 2η.
Ehc Ehc Ehc
This proves the first part of the lemma. The second part is an immediate conse-
quence.
Next we prove theorems 2 and 3.
Proof of theorem 2. We first prove part (a). By remark 1, we may assume p = 1.
Let A ∈ GIC and ε > 0 be given. We will denote
λ̂k (A, x) = λ1 (A, x) + · · · + λk (A, x)
Let us first consider the case where the following condition is satisfied:
(3) λ̂k (A, x) ≥ 0 for µ-a.e. x ∈ X.
From the subadditive ergodic theorem, we know that the convergence in
1
λ̂k (A, x) = lim log k∧k An (x)k,
n→+∞ n

takes place almost everywhere and also in L1 . Hence, using (3),

Z
1
lim log− k∧k An k = 0.
n→+∞ n X

We take N ∈ N such that (recall (2))

Z Z
1 1
log− k∧k AN k < ε and log k∧k AN k < Λk (A) + ε.
N X N X
Therefore
Z
1
(4) log+ k∧k AN k < Λk (A) + 2ε.
N X
+
Let f = log kAk and η = ε/N . Let K > 0 be given by lemma 5 applied to f
and η. Set
n o δ0
(5) δ 0 = min η, εe−K(N −1) and δ = .
1 + δ0
Now fix B ∈ G such that ρ1 (B, A) < δ. Therefore kB −Ak1 < δ 0 and, by remark 2,
B ∈ GIC . Let g = log+ kBk. By lemma 4, kg − f k1 ≤ kB − Ak1 < δ 0 ≤ η. We use
lemma 5 with h = f and h = g: Let
Ef = {f ≤ K} and Eg = {g ≤ K} ;
6 A. ARBIETO, J. BOCHI

then Z
2η
h < 2η and µ(Ehc ) < for h = f , g.
Ehc K
Set
−1
N\
G= T −i (Ef ∩ Eg ).
i=0
Then Gc has small measure:
4N η 4ε
(6) µ(Gc ) ≤ N µ(Efc ∪ Egc ) < = .
K K
1 + k N
R
We are going to bound the expression R N R X logR k∧ B k. To do so, we are
going to split the integral in two parts, X = Gc + G . For the first part, we have
Z N −1 Z N −1 Z
1 1 X k X
log+ k∧k B N k ≤ log+ k∧k Bk ≤ g.
N Gc N T i (Gc ) N T i (Gc )
i=0 i=0
For each i = 0, . . . ,N − 1 we have, by lemma 5 and relation (6),
Z Z Z
g= g+ g
T i (Gc ) Egc Eg ∩T i (Gc )

< 2η + Kµ(Eg ∩ T i (Gc )) ≤ 2ε + Kµ(Gc ) < 6ε .

Hence
Z
1
(7) log+ k∧k B N k ≤ 6kε.
N Gc

Next we estimate the second part. Using lemma 4 and (4) we get
Z Z Z
1 + k N 1 + k N k
log k∧ B k ≤ log k∧ A k + kB N − AN k
N G N G N G
(8) Z
k
≤ Λk (A) + 2ε + kB N − AN k.
N G
To estimate the integral on the right hand side, we proceed as follows. Take x ∈ G
and 1 ≤ i ≤ N − 1; then
kB i+1 (x) − Ai+1 (x)k ≤ kB(T i x)k kB i (x) − Ai (x)k + kB(T i x) − A(T i x)k kAi (x)k
≤ eK kB i (x) − Ai (x)k + eKi kB(T i x) − A(T i x)k.
Integrating over G and using kB − Ak1 < δ 0 , we get
Z Z
kB i+1 − Ai+1 k ≤ eK kB i − Ai k + eKi δ 0 .
G G
By induction, we obtain
Z
kB i − Ai k ≤ ieK(i−1) δ 0 ∀ i = 1, . . . , N.
G
Using this relation with i = N , (8) and (5), we get
Z
1
(9) log+ k∧k B N k ≤ Λk (A) + (2 + k)ε.
N G
 Lp -GENERIC COCYCLES HAVE ONE-POINT LYAPUNOV SPECTRUM 7

From (7) and (9), we conclude that

Z Z
1 k N 1
Λk (B) ≤ log k∧ B k ≤ log+ k∧k B N k ≤ Λk (A) + (2 + 7k)ε.
N X N X
This proves part (a) of the theorem in the case condition (3) is satisfied. (Replace
ε with ε/(2 + 7k) along the proof.)
Next we consider the general case. Again, let A ∈ GIC and ε > 0. For a > 0,
consider the T -invariant set La = {λ̂k (A, x) < −a}. Choose a large enough so
that
Z Z
+ k
(10) log k∧ Ak < ε and λ̂k (A, x) dµ(x) > −ε.
La La
The cocycle ea A restricted to Lca satisfies condition (3). If B ∈ G is ρ1 -
sufficiently close to A then ρ1 (ea A, ea B) will also be small and therefore, by the
case already considered,
Z Z
λ̂k (ea B, x) dµ(x) ≤ λ̂k (ea A, x) dµ(x) + ε,
Lca Lca
that is, Z Z
λ̂k (B, x) dµ(x) ≤ λ̂k (A, x) dµ(x) + ε.
Lca Lca
On the other hand, since La is invariant,
Z Z Z
1
λ̂k (B, x) dµ(x) = inf log+ k∧k B n k ≤ log+ k∧k Bk.
La n n La La
We may also assume that kB − Ak1 < ε. Then, lemma 4 and (10), we have
Z Z Z
λ̂k (B, x) dµ(x) ≤ log+ k∧k Ak + k kB − Ak
La La La
Z
≤ (1 + k)ε ≤ λ̂k (A, x) dµ(x) + (2 + k)ε.
La
We conclude that Λk (B) < Λk (A) + (3 + k)ε. This completes the proof of part (a)
of the theorem.
Part (b) is an easy consequence of part (a): The functions
Z
−1
(11) A 7→ Λk (A) = −Λk (A ) =
e (λd−k+1 (A, x) + · · · + λd (A, x)) dµ(x)
X

are lower-semicontinuous. In particular, the function Λd (·) = Λ

e d (·) is continuous.

Proof of theorem 3. Let A ∈ GOPS and ε > 0. If B is sufficiently ρp -close to A

then (recall (11))
Λ1 (B) < Λ1 (A) + ε/k and Λ e 1 (B) − ε/k.
e 1 (A) > Λ
Therefore (
≤ kΛ1 (B) < kΛ1 (A) + ε = Λk (A) + ε,
Λk (B)
≥ kΛ e 1 (A) − ε = Λk (A) − ε.
e 1 (B) > k Λ
8 A. ARBIETO, J. BOCHI

showing that Λk is continuous at A.

Conversely, take A0 ∈ GIC and assume Λk is continuous at A0 . By the theorem
of Arnold and Cong [3], there exists a sequence Bn in GOPS converging to A0 in
the ρp -metric. Since Λd is also continuous at A0 , we have
Λk (A0 ) Λk (Bn ) Λd (Bn ) Λd (A0 )
= lim = lim = .
k n→∞ k n→∞ d d
This implies A0 ∈ GOPS , because k < d.
As we have mentioned, theorem 1 follows from theorems 2 and 3.
Acknowledgement. We would like to thank the referee for his corrections.

References
[1] L. Arnold. Random Dynamical Systems. Springer-Verlag, 1998.
[2] L. Arnold and Nguyen Dinh Cong. Generic properties of Lyapunov exponents. Random
Comput. Dynam., 2: 335–345, 1994.
[3] L. Arnold and Nguyen Dinh Cong. On the simplicity of the Lyapunov spectrum of
products of random matrices. Erg. Th & Dynam. Sys., 17: 1005–1025, 1997.
[4] L. Arnold and Nguyen Dinh Cong. Linear Cocycles with simple Lyapunov Spectrum
are dense in L∞ . Erg. Th & Dynam. Sys., 19: 1389–1404, 1999.
[5] J. Bochi. Genericity of zero Lyapunov exponents. To appear in Erg. Th & Dynam. Sys.
[6] J. Bochi and M. Viana. The Lyapunov exponents of generic volume preserving and sym-
plectic systems. Preprint www.preprint.impa.br, 2002.
[7] C. Bonatti and M. Viana. Lyapunov exponents with multiplicity 1 for deterministic
product of matrices. Preprint www.preprint.impa.br, 2001.
[8] P. Bougerol and J. Lacroix. Products of random matrices with applications to
Schrödinger operators. Birkhäuser, 1985.
[9] H. Furstenberg and H. Kesten. Products of random matrices. Ann. Math. Statist., 31:
457–469, 1960.
[10] I. Goldsheid and G. Margulis. Lyapunov indices of a product of random matrices.
Russian Math. Surveys, 44: 11–71, 1989.
[11] O. Knill. Positive Lyapunov exponents for a dense set of bounded measurable SL(2, R)
cocycles. Erg. Th & Dynam. Sys., 12: 319–331, 1992.
[12] K. Kuratowski. Topology, vol. 1. Academic Press, 1966.
[13] V. I. Oseledets. A multiplicative ergodic theorem: Lyapunov characteristic numbers for
dynamical systems. Trans. Moscow Math. Soc., 19: 197–231, 1968.

Alexander Arbieto (alexande@impa.br) Jairo Bochi (bochi@impa.br)

IMPA, Estrada D. Castorina 110, Jardim Botânico, 22460-320 Rio de Janeiro, Brazil