Oseledets Theorem

Seminar in Homogenous Dynamics

February 2024

Definition: Let (𝑋, ℬ, 𝜇, 𝑇) be a probability preserving system (𝑇∗ 𝜇 = 𝜇). A

sequence of functions 𝑔𝑛 : 𝑋 → ℝ of measurable functions is called subadditive,
relative to 𝑓 if 𝑔𝑚+𝑛 ≤ 𝑔𝑚 + 𝑔𝑛 ∘ 𝑇 𝑚 for all 𝑚, 𝑛 ≥ 1

Examples: 1) If 𝑓: 𝑋 → ℝ is measurable, then the ergodic sum 𝑔𝑁 = ∑𝑁−1 𝑛

𝑛=0 𝑓 ∘ 𝑇 is

an additive sequence, and particularly subadditive.

2) Given a measurable 𝐴: 𝑋 → 𝐺𝐿𝑑 ℝ, we define 𝐴𝑛 = (𝐴 ∘ 𝑓 𝑛−1 ) ⋅ (𝐴 ∘ 𝑓 𝑛−2 ) ⋅ … ⋅ 𝐴

and 𝜑𝑛 (𝑥) = log ||𝐴𝑛 (𝑥)|| with the norm being the operator norm. This is
subadditive since the operator norm satisfies ||𝐴𝐵|| < ||𝐴|| ⋅ ||𝐵|| for any two
𝐴, 𝐵 ∈ 𝑀𝑑 (ℝ).

3) In the previous example if 𝑋 is a smooth manifold, then the differential map of

𝑓, 𝐷𝑥 (𝑓): 𝑇𝑥 𝑋 → 𝑇𝑓(𝑥) 𝑋 between tangent spaces satisfies

𝐷𝑓𝑛(𝑥) (𝑓) = 𝐷𝑓𝑛−1 (𝑥) (𝑓) ⋅ … ⋅ 𝐷𝑥 (𝑓)

Any 𝑓 with differential of full rank gives us a map 𝐷(𝑓): 𝑋 → 𝐺𝐿𝑑 (ℝ), where 𝑑 =
dim 𝑋.

Remark: A map 𝐹: 𝑋 × ℝ𝑑 → 𝑋 × ℝ𝑑 with 𝐹(𝑥, 𝑣) = (𝑓(𝑥), 𝐴(𝑥)𝑣) for 𝐴 as in

example 2 is called a linear cocycle and 𝐹 𝑛 (𝑥, 𝑣) = (𝑓 𝑛 (𝑥), 𝐴𝑛 (𝑥)𝑣)

Theorem 1 (Kingman's subadditive ergodic theorem): Let (𝑋, ℬ, 𝜇, 𝑇) be a p.p.s.

and 𝑔𝑛 : 𝑋 → ℝ a subadditive sequence with 𝑔1+ ∈ 𝐿1 (𝑋, ℬ, 𝜇). Then the limit 𝑔 ≔
1
lim 𝑔𝑛 exists 𝜇 −a.e. and it is an invariant function. Moreover, we have that
𝑛→∞ 𝑛
1 1
∫ 𝑔𝑑𝜇 = lim 𝑛 ∫ 𝑔𝑛 𝑑𝜇 = inf ∫ 𝑔𝑛 𝑑𝜇.
𝑛→∞ 𝑛→∞ 𝑛

The following theorem is a direct consequence of the previous one (but was
originally proved without using the subadditive ergodic theorem):
Theorem 2 (Furstenberg, Kesten): Let 𝐴 be as above and assume that log + (𝐴±1 ) ∈
𝐿1 . Then the following limits exist:

1 1
𝜆+ (𝑥) = lim log‖𝐴𝑛 (𝑥)‖, 𝜆− (𝑥) = lim log‖𝐴𝑛 (𝑥)−1 ‖−1
𝑛→∞ 𝑛 𝑛→∞ 𝑛

and they are invariant and integrable with

1
∫ 𝜆+ 𝑑𝜇 = lim ∫ log‖𝐴𝑛 (𝑥)‖ 𝑑𝜇
𝑛→∞ 𝑛

1
∫ 𝜆− 𝑑𝜇 = lim ∫ log‖𝐴𝑛 (𝑥)−1 ‖−1 𝑑𝜇
𝑛→∞ 𝑛

𝜆± (𝑥) are called the extremal Lyapunov exponents

The theorem of Furstenberg and Kesten provides us with growth rate for ||𝐴𝑛 (𝑥)||.
Oseledets theorem, also known as the multiplicative ergodic theorem, provides
growth rates for ||𝐴𝑛 (𝑥)𝑣|| for all 𝑣 ∈ ℝ𝑑 . Here we only give the statement of the
general case and provide a proof of the 2-dimensional case:

Theorem (Oseledets): Let (𝑋, ℬ, 𝜇, 𝑇) be a probability preserving system. And let

𝐴: 𝑋 → 𝐺𝐿2 be a measurable function with log + ||𝐴±1 || ∈ 𝐿1 . For almost every 𝑥 ∈
𝑋 there exist 𝑘(𝑥) ∈ ℕ, numbers 𝜆1 (𝑥) > ⋯ > 𝜆𝑘(𝑥) , and a sequence of subspaces

ℝ𝑑 = 𝑉𝑥1 ⊃ ⋯ ⊃ 𝑉𝑥𝑘(𝑥) ⊃ {0} such that for all 1 ≤ 𝑖 ≤ 𝑘:

𝑖
1) 𝑘(𝑇(𝑥)) = 𝑘(𝑥), 𝜆𝑖 (𝑇(𝑥)) = 𝜆𝑖 (𝑥), 𝐴(𝑥)𝑉𝑥𝑖 = 𝑉𝑇(𝑥)
2) The maps 𝑘, 𝜆𝑖 , 𝑉𝑥𝑖 are measurable.
1
3) For all 𝑣 ∈ 𝑉𝑥𝑖 ∖ 𝑉𝑥𝑖+1, 𝜆𝑖 (𝑥) = lim 𝑛 log‖𝐴𝑛 (𝑥)𝑣‖
𝑛→∞

Remarks: 1) When 𝜇 is ergodic, 𝑘(𝑥), 𝜆𝑖 (𝑥), dim 𝑉𝑥𝑖 are constant a.e.

2) The result does not depend on the choice of norm or basis on each of the
subspaces because all norms on the Euclidean space are equivalent and constants
disappear. Change of basis is equivalent to taking another norm.

Examples: 1) Take a matrix 𝐴 ∈ 𝑆𝐿𝑑 ℝ and take the cocycle 𝐹: 𝑋 × ℝ𝑑 → 𝑋 × ℝ𝑑

with 𝐹(𝑥, 𝑣) = (𝑓(𝑥), 𝐴𝑣) and 𝐹 𝑛 (𝑥, 𝑣) = (𝑓 𝑛 (𝑥), 𝐴𝑛 𝑣). If 𝐴 is diagonalizable with
eigenvalues 𝜆1 ≥ ⋯ ≥ 𝜆𝑛 and eigenspaces ℝ𝑑 = 𝑉 𝜆1 ⊕ … ⊕ 𝑉 𝜆𝑑 , for every vector
𝑣 with a component in 𝑉 𝜆1 the norm ‖𝐴𝑛 𝑣‖ grows like 𝜆𝑛 ‖𝑣‖. On the other hand,
if 𝑣 ∈ 𝑉 𝜆2 ⊕ … ⊕ 𝑉 𝜆𝑑 then the norm has a different rate of growth. We can
iteratively use the largest eigenvalue to determine the rate of growth outside the
sum of the eigenspaces with smaller eigenvalues. The theorem approves this
intuitive argument in a much general setting.

2) Consider products of random matrices. Namely, take 𝐴0 , 𝐴1 ∈ 𝐺𝐿𝑑 ℝ , 𝑋 =

{𝐴0 , 𝐴1 }ℤ and 𝜇 = 𝜇𝑝ℤ for some 𝑝 ∈ (0, 1). As usual, we take 𝑇: 𝑋 → 𝑋 to be the shift
map. Also define 𝐴: 𝑋 → 𝐺𝐿𝑑 ℝ by (𝑎𝑛 ) ↦ 𝑎0 . Then the cocycle defined by 𝐴, 𝑇 is
𝐹: 𝑋 × ℝ𝑑 → 𝑋 × ℝ𝑑 so that 𝐹 𝑚 ((𝑎𝑛 ), 𝑣) = ((𝑎𝑛+𝑚 ), 𝑎𝑚−1 ⋅ … ⋅ 𝑎0 𝑣). The theorem
provides us with the "growth rate" of this random walk.

3) Applying the theorem in the context given in example 3 from the previous
discussion gives us a direction on the tangent space for which the cocycle has
exponential growth rate of 𝜆, and determines the rate for all the other directions
(This is my intuition for the 2-dimensional case).

Before we get to the proof of the 2-dimensional case, we prove a lemma:

Lemma 1: Let (𝑋, ℬ, 𝜇, 𝑇) be a p.p.s. and 𝑓 ∈ 𝐿1 then for 𝜇-a.e. 𝑥 ∈ 𝑋,

1
lim 𝑓(𝑇 𝑛 (𝑥)) = 0
𝑛→∞ 𝑛

Proof: Let 𝜀 > 0 and denote 𝐴𝑛 = {𝑥 ∈ 𝑋: |𝑓(𝑇 𝑛 (𝑥))| ≥ 𝑛𝜀}. Therefore,

∞
|𝑓(𝑥)|
𝜇(𝐴𝑛 (𝜀)) = 𝜇({𝑥 ∈ 𝑋: |𝑓(𝑥)| > 𝑛𝜀}) = ∑ 𝜇 ({𝑥 ∈ 𝑋: 𝑘 ≤ < 𝑘 + 1})
𝜀
𝑘=𝑛

∞ ∞
|𝑓(𝑥)| |𝑓|
∑ 𝜇(𝐴𝑛 (𝜀)) = ∑ 𝑘 𝜇 ({𝑥 ∈ 𝑋: 𝑘 ≤ < 𝑘 + 1}) ≤ ∫ 𝑑𝜇 < ∞
𝜀 𝑋 𝜀
𝑛=1 𝑘=1

By Borel-Cantelli 𝜇(limsup𝐴𝑛 (𝜀)) = 0. For every 𝑥 ∉ limsup𝐴(𝜀) there is some

𝑝 > 1 such that for every 𝑛 ≥ 𝑝 such that |𝑓(𝑇 𝑛 (𝑥))| < 𝑛𝜀. To conclude the proof,
1 1
notice that 𝐴 = ⋃𝑖≥1 limsup𝐴 ( 𝑖 ) has measure zero and lim 𝑛 𝑓(𝑇 𝑛 (𝑥)) = 0 on its
𝑛→∞

complement.
Theorem: Let (𝑋, ℬ, 𝜇, 𝑇) be a probability preserving system. And let 𝐴: 𝑋 → 𝐺𝐿2
be a measurable function with log + ||𝐴±1 || ∈ 𝐿1 . Then for 𝜇-a.e. 𝑥 ∈ 𝑋 one of the
following holds:

1
1) 𝜆− (𝑥) = 𝜆+ (𝑥) and for all 𝑣 ∈ ℝ2 ∖ {0}, 𝜆± (𝑥) = lim 𝑛 log||𝐴𝑛 (𝑥)𝑣||
𝑛→∞

2) 𝜆+ (𝑥) > 𝜆− (𝑥) and there exists a 1-dimensional subspace 𝐸𝑥 ⊂ ℝ2 such

that:
1 𝜆 (𝑥), 𝑣 ∈ 𝐸𝑥 ∖ {0}
lim log||𝐴𝑛 (𝑥)𝑣|| = { −
𝑛→∞ 𝑛 𝜆+ (𝑥), 𝑣 ∈ ℝ2 ∖ 𝐸𝑥

And 𝐴(𝑥)𝐸𝑥 = 𝐸𝑇(𝑥)

Note that the existence of 𝜆± (𝑥) follows of the theorem of Furstenberg-Kesten.

Proof: We begin with a reduction to the case where 𝐴 takes values in 𝑆𝐿2 . We justify
this step by checking that for any 𝐴(𝑥) ∈ 𝐺𝐿𝑑 we can normalize 𝐴(𝑥) to obtain a
matrix 𝐵(𝑥) ∈ 𝑆𝐿𝑑 . If the integrability assumptions hold for 𝐴(𝑥), then they also
hold for 𝐵(𝑥). Also, the corresponding one-dimensional lines specified by the
theorem are the same for both cocycles, and the Lyapunov exponents of the one
are just a shift of the exponents of the other by some additive constant dependent
on 𝑥.

We show an equivalent definition for the operator norm taken with respect to the
𝐿2 norm on ℝ𝑛

Proposition: ‖𝐴‖=√𝜆 where 𝜆 is the largest eigen value of 𝐴𝑡 𝐴

Proof: 𝐴𝑡 𝐴 is symmetric and hence has an orthonormal basis of eigenvectors

{𝑒1 , … 𝑒𝑛 }. Let 𝜆1 ≥ ⋯ ≥ 𝜆𝑛 be the corresponding eigenvalues. Then for any 𝑥 ∈
ℝ𝑛 ∖ {0} if 𝑥 = ∑𝑛𝑖=1 𝛼𝑖 𝑒𝑖 we have ‖𝐴(𝑥)‖2 = ∑𝑛𝑖=1 𝜆𝑖 𝑎𝑖2 ≤ 𝜆1 ⋅ ‖𝑥‖, so ‖𝐴‖ ≤ √𝜆1.
We also have ‖𝐴‖2 ≥ ‖𝐴𝑒1 ‖2 = 𝜆1 which ends the proof.

We prove two preliminary claims about 𝑆𝐿2 :

Claim 1: For any 𝐴 ∈ 𝑆𝐿2 , ‖𝐴‖ = ||𝐴−1 || where the operator norm is taken with
respect to Euclidean 𝐿2 norm on ℝ2 .
Proof: The characteristic polynomial of 𝐴𝑡 𝐴 is 𝑥 2 − 𝑡𝑟(𝐴𝑡 𝐴)𝑥 + det 𝐴𝑡 𝐴 and it is
easy to see that the same polynomial is obtained for (𝐴−1 )𝑡 (𝐴−1 ) for 𝐴 ∈ 𝑆𝐿2 . From
the previous proposition we the equality of the norms.

−1
Claim 2: Let 𝐴 ∈ 𝑆𝐿2 , then there exist vectors 𝑠, 𝑢 ∈ ℝ2 such that ||𝐴𝑠|| = ||𝐴||
and ||𝐴𝑢|| = ||𝐴||. If ‖𝐴‖ ≠ 1 then 𝑢, 𝑣 are orthogonal and unique up to sign.

Proof: Existence of both vectors follows from the fact that the continuous function
𝑥 ↦ ||𝐴𝑥|| achieves maximal and minimal values on the compact set 𝑆 1 . Let 𝑠 ∈ ℝ2
be the most contracted vector. Pick some unit vector 𝑢 ∈ ℝ2 orthogonal to 𝑠. Then,
for every 𝑥 ∈ ℝ2 , 𝑥 = 𝛼1 𝑠 + 𝛼2 𝑢 with |𝛼1 |2 + |𝛼2 |2 = 1.

We have ||𝐴𝑥|| ≤ |𝛼1 |||𝐴𝑠|| + |𝛼2 |||𝐴𝑢|| ≤ (|𝛼1 | + |𝛼2 |) ⋅ ||𝐴𝑢|| ≤ ||𝐴𝑢||.
Therefore, ||𝐴|| = ||𝐴𝑢|| as required.

Now, fix some 𝑥 ∈ 𝑋 for which theorem 2 holds.

1
Notice that 𝜆+ (𝑥) + 𝜆− (𝑥) = lim 𝑛 log(‖𝐴𝑛 (𝑥)‖ ⋅ ‖𝐴𝑛 (𝑥)−1 ‖−1 ) = [Link],
𝑛→∞

we denote 𝜆(𝑥): = 𝜆+ (𝑥) = −𝜆− (𝑥).

We prove that of 𝜆(𝑥) = 0 then (1) holds and if 𝜆(𝑥) > 0 then (2) holds.

Assume that 𝜆(𝑥) = 0, (𝜆+ (𝑥) = 𝜆− (𝑥)) then for any 𝑣 ∈ ℝ2 ,

‖𝐴𝑛 (𝑥)‖−1 ‖𝑣‖ = ‖𝐴𝑛 (𝑥)−1 ‖−1 ‖𝑣‖ ≤ ‖𝐴𝑛 (𝑥)𝑣‖ ≤ ‖𝐴𝑛 (𝑥)‖‖𝑣‖

‖𝑢‖ ‖𝑣‖
because ‖𝐴𝑛 (𝑥)−1 ‖ = sup {‖𝐴𝑛 (𝑥)𝑢‖ : 𝑢 ∈ ℝ2 ∖ {0}} ≥ ‖𝐴𝑛 (𝑥)‖

And hence

1 1 1
log‖𝐴𝑛 (𝑥)‖−1 ‖𝑣‖ ≤ log‖𝐴𝑛 (𝑥)𝑣‖ ≤ log‖𝐴𝑛 (𝑥)‖‖𝑣‖
𝑛 𝑛 𝑛

Which shows the result for the first case in the theorem.

Now, assume that 𝜆(𝑥) > 0, then for 𝑛 large enough ‖𝐴𝑛 (𝑥)‖ > 1. By claim 2 we
can choose 𝑢𝑛 (𝑥), 𝑠𝑛 (𝑥) ∈ ℝ2 such that ‖𝐴𝑛 (𝑥)𝑠𝑛 (𝑥)‖ = ‖𝐴𝑛 (𝑥)‖−1 and
‖𝐴𝑛 (𝑥)𝑢𝑛 (𝑥)‖ = ‖𝐴𝑛 (𝑥)‖. We proceed with the following lemmas:

1
Lemma 2: limsup 𝑛 log|sin ∠(𝑠𝑛 (𝑥), 𝑠𝑛+1 (𝑥))| < −2𝜆(𝑥)
Proof: We denote 𝛼𝑛 = ∠(𝑠𝑛 (𝑥), 𝑠𝑛+1 (𝑥)) and decompose

𝑠𝑛 (𝑥) = sin 𝛼𝑛 𝑢𝑛+1 (𝑥) + cos 𝛼𝑛 𝑠𝑛+1 (𝑥)

Then we use the properties of 𝑢𝑛 (𝑥), 𝑠𝑛 (𝑥) to get:

|sin 𝛼𝑛 | ⋅ ‖𝐴𝑛+1 (𝑥)‖ ≤ ‖𝐴𝑛+1 (𝑥)𝑠𝑛 (𝑥)‖ ≤ ‖𝐴(𝑇 𝑛 (𝑥))‖‖𝐴𝑛 (𝑥)‖−1

Which implies

1 1 1 1
log|sin(𝛼𝑛 )| ≤ log‖𝐴(𝑇 𝑛 (𝑥))‖ − log‖𝐴𝑛 (𝑥)‖ − log‖𝐴𝑛+1 (𝑥)‖
𝑛 𝑛 𝑛 𝑛

After taking limsup of both sides, the definition of 𝜆(𝑥) and lemma 1 complete
the proof.

Lemma 3: The sequence (𝑠𝑛 (𝑥))𝑛 is a Cauchy sequence.

Proof: Using the previous lemma, we obtain ‖𝑠𝑛 (𝑥) − 𝑠𝑛+1 (𝑥)‖ ≤ 2|sin 𝛼𝑛 | ≤
2𝑒 𝑛(−2𝜆(𝑥)+𝜀) for all 𝜀 > 0 such that −2𝜆(𝑥) + 𝜀 < 0. Therefore, using the triangle
inequality we can bound the differences between elements of the sequence:

‖𝑠𝑛 (𝑥) − 𝑠𝑛+𝑘 (𝑥)‖ ≤ 𝐶𝑒 𝑛(−2𝜆(𝑥)+𝜀)

For a suitable constant 𝐶 > 0 and for arbitrary large enough 𝑛, 𝑘 > 0.

Denote 𝑠(𝑥) = lim 𝑠𝑛 (𝑥)

𝑛→∞

1
Lemma 3: lim 𝑛 log‖𝐴𝑛 (𝑥)𝑠(𝑥)‖ = −𝜆(𝑥)
𝑛→∞

Proof: Denote 𝛽𝑛 = ∠(𝑠(𝑥), 𝑠𝑛 (𝑥)). We have 𝑠(𝑥) = cos 𝛽𝑛 𝑠𝑛 (𝑥) + sin 𝛽𝑛 𝑢𝑛 (𝑥).
And hence

1
limsup log‖𝐴𝑛 (𝑥)𝑠(𝑥)‖
𝑛
1 1
≤ max {limsup log(|cos 𝛽𝑛 |‖𝐴𝑛 (𝑥)𝑠𝑛 (𝑥)‖) , limsup log(|sin 𝛽𝑛 |‖𝐴𝑛 (𝑥)𝑢𝑛 (𝑥)‖) }
𝑛 𝑛
1
limsup log(‖𝐴𝑛 (𝑥)‖−1 ) ,
≤ max { 𝑛 }
1 1 𝑛
limsup log(|sin 𝛽𝑛 |) + limsup log(‖𝐴 (𝑥)𝑢𝑛 (𝑥)‖)
𝑛 𝑛
≤ max{−𝜆(𝑥), −2𝜆(𝑥) + 𝜆(𝑥) } = −𝜆(𝑥)
as required.

It remains to deal with vectors that are not on the line generated by 𝑠(𝑥).

1
Lemma 4: For any 𝑣 ∈ ℝ2 ∖ 𝑆𝑝𝑎𝑛(𝑠(𝑥)), lim 𝑛 log‖𝐴𝑛 (𝑥)𝑣‖ = 𝜆(𝑥)
𝑛→∞

Proof: 𝑣, 𝑠(𝑥) are not collinear, therefore, if 𝛾𝑛 = ∠(𝑣, 𝑠𝑛 (𝑥)) then |sin 𝛾𝑛 | > 0 for
𝑛 large enough. Again, we decompose to get 𝑣 = cos 𝛾𝑛 𝑠𝑛 (𝑥) + sin 𝛾𝑛 𝑢𝑛 (𝑥). Then,
a lower bound is ‖𝐴𝑛 (𝑥)𝑣‖ ≥ |sin 𝛾𝑛 |‖𝐴𝑛 (𝑥)𝑢𝑛 (𝑥)‖ − |cos 𝛾𝑛 |‖𝐴𝑛 𝑠𝑛 (𝑥)‖. Using
the fact that for any two real sequences 𝑎𝑛 , 𝑏𝑛 ,

1 1 1
liminf log(𝑎𝑛 + 𝑏𝑛 ) ≥ max {liminf log 𝑎𝑛 , liminf log 𝑏𝑛 }
𝑛 𝑛 𝑛
1
we obtain liminf 𝑛 log‖𝐴𝑛 (𝑥)𝑣‖ ≥ 𝜆(𝑥). On the other hand,
1
limsup 𝑛 log‖𝐴𝑛 (𝑥)𝑣‖ ≤ 𝜆(𝑥) , which end the proof.

We now show that the line spanned by 𝑠(𝑥) is invariant under the action:

Claim: 𝐴(𝑥)𝑠(𝑥) and 𝑠(𝑇(𝑥)) are collinear.

Proof: By the previous lemma, it suffices to check that

1
lim log‖𝐴𝑛 (𝑇(𝑥)) ⋅ 𝐴(𝑥)𝑠(𝑥)‖ ≠ 𝜆(𝑓(𝑥))
𝑛→∞ 𝑛

And indeed, by lemma 3,

1 1
lim log‖𝐴𝑛 (𝑓(𝑥)) ⋅ 𝐴(𝑥)𝑠(𝑥)‖ = lim log‖𝐴𝑛+1 (𝑥)𝑠(𝑥)‖ = −𝜆(𝑥)
𝑛→∞ 𝑛 𝑛→∞ 𝑛 + 1

From the invariance and positiveness of 𝜆 we gat that 𝜆(𝑇(𝑥)) = 𝜆(𝑥) ≠ −𝜆(𝑥).

This finishes the proof for the 2-dimensional case.