38.07.01 · dynamics / smooth-ergodic-theory

The Oseledets Multiplicative Ergodic Theorem and Lyapunov Exponents

shipped3 tiersLean: none

Anchor (Master): Katok-Hasselblatt 1995 *Introduction to the Modern Theory of Dynamical Systems* (Cambridge) Supplement and Ch. S; Barreira-Pesin 2007 *Nonuniform Hyperbolicity: Relative Stability and Smooth Ergodic Theory* (Cambridge) Ch. 1-3 (Lyapunov exponents, the Oseledets theorem, regularity); Arnold 1998 *Random Dynamical Systems* (Springer Monographs) Ch. 3-4 (the MET for random and continuous cocycles); Viana 2014 *Lectures on Lyapunov Exponents* (Cambridge) Ch. 4 (Oseledets via subadditivity, Kingman input)

Intuition Beginner

Take a small ball of nearby starting points in some system — say nearby weather states, or nearby positions of a planet — and let them all evolve by the same fixed rule. After one step the ball is no longer a ball: the rule stretches it in some directions and squashes it in others, turning the sphere into a thin ellipsoid. Keep going. The directions that get stretched grow longer and longer; the directions that get squashed shrink. A Lyapunov exponent is simply the long-run growth rate along one of these directions: how fast, on average, a tiny displacement gets amplified or damped per step.

Why does this matter? Because the largest growth rate is the precise mathematical meaning of "sensitive dependence on initial conditions" — the butterfly effect. If the top rate is positive, two nearby starts drift apart exponentially fast, and long-term prediction is hopeless no matter how good your measurements. If every rate is negative, nearby starts pull together and the system is stable. The exponents are a fingerprint of the dynamics that does not depend on which coordinates you happen to use.

The remarkable fact, due to Oseledets, is that these growth rates are not a vague average over all directions: at almost every point the tangent space splits cleanly into a few special subspaces, each with its own exact growth rate. A displacement that lies inside one of these subspaces grows at precisely that subspace's rate, with no ambiguity. The whole space of small displacements is sorted, once and for all, into "this many directions growing at rate $λ_{1}$ , this many at the slower rate $λ_{2}$ ," and so on.

The takeaway: repeatedly applying a linear rule to a vector makes its length grow or shrink exponentially, and the exponent depends only on which special direction the vector points along. Oseledets' theorem says these special directions and their exact rates exist at almost every point, turning the messy picture of stretching and folding into a finite list of numbers.

Visual Beginner

Picture a unit circle of displacement vectors at a point. Apply the linear rule once, twice, three times: the circle becomes a progressively longer and thinner ellipse, its long axis growing like $e^{λ_{1} n}$ and its short axis shrinking like $e^{λ_{2} n}$ .

The ellipse is the image of the displacement circle; its axes grow and shrink at the two Lyapunov rates. The two lines $E_{1}$ and $E_{2}$ are the Oseledets directions: a vector exactly on $E_{2}$ keeps shrinking at the slow rate, but the tiniest component along $E_{1}$ eventually takes over, which is why almost every vector grows at the top rate $λ_{1}$ .

Worked example Beginner

We track growth rates for the simplest repeated linear rule: a diagonal stretch-and-squash applied over and over.

Step 1. The rule. At each step, multiply the horizontal part of a displacement by $3$ and the vertical part by $1/2$ . As a matrix this is $(30 0 1/2)$ . Applying it $n$ times multiplies the horizontal part by $3^{n}$ and the vertical part by $(1/2)^{n}$ .

Step 2. A purely horizontal displacement. Start with the vector $(1, 0)$ . After $n$ steps it is $(3^{n}, 0)$ , with length $3^{n}$ . The growth rate is the log of the length divided by $n$ : $\frac{1}{n} lo g (3^{n}) = lo g 3 \approx 1.099$ . So a horizontal displacement grows at rate $lo g 3$ . This is the top Lyapunov exponent $λ_{1} = lo g 3$ .

Step 3. A purely vertical displacement. Start with $(0, 1)$ . After $n$ steps it is $(0, (1/2)^{n})$ , with length $(1/2)^{n}$ . The growth rate is $\frac{1}{n} lo g ((1/2)^{n}) = lo g (1/2) = - lo g 2 \approx - 0.693$ . So a vertical displacement shrinks at rate $- lo g 2$ . This is the second exponent $λ_{2} = - lo g 2$ .

Step 4. A slanted displacement. Start with $(1, 1)$ . After $n$ steps it is $(3^{n}, (1/2)^{n})$ . For large $n$ the horizontal part $3^{n}$ dwarfs the vertical part, so the length is essentially $3^{n}$ and the growth rate is again $lo g 3$ . The horizontal (fast) direction wins.

What this tells us: the two special directions are horizontal (growing at $λ_{1} = lo g 3$ ) and vertical (shrinking at $λ_{2} = - lo g 2$ ). A vector exactly on the vertical line keeps the slow rate; any vector with a sliver of horizontal part ends up at the fast rate. That sorting of directions into exact rates is exactly what Oseledets' theorem guarantees, even when the rule changes from step to step instead of staying fixed.

Check your understanding Beginner

Exercise (easy, multiple choice).

A Lyapunov exponent measures:

A. The total distance a single trajectory travels B. The long-run exponential rate at which a small displacement grows or shrinks per step C. The number of fixed points of the system D. The area of the region the system fills

Hint

Think about the ellipse picture: each axis grows or shrinks like $e^{λn}$ , and $λ$ is the rate.

Answer

B. The long-run exponential rate of growth or shrinkage of a small displacement.

Feedback-correct: a Lyapunov exponent is exactly $\frac{1}{n} lo g (length of the displacement after n steps)$ in the long run, the exponential rate along a direction. Feedback-wrong: A measures arc length of one orbit, not separation of nearby orbits; C counts equilibria; D is a volume, not a rate. Only B captures the exponential separation rate.

Formal definition Intermediate+

Throughout, $(X, B, μ, T)$ is a measure-preserving system 38.04.01 on a probability space, and $d \geq 1$ is a fixed dimension. We write $GL (d, R)$ for the group of invertible real $d \times d$ matrices, $∥ \cdot ∥$ for the operator norm on matrices and the Euclidean norm on $R^{d}$ , and $\land^{k} R^{d}$ for the $k$ -th exterior power, on which a matrix $A$ acts as $\land^{k} A$ with $∥ \land^{k} A ∥$ equal to the product $σ_{1} (A) \dots σ_{k} (A)$ of the $k$ largest singular values of $A$ .

Definition (linear cocycle). A measurable linear cocycle over $(X, T)$ is a measurable map $A : X \to GL (d, R)$ . Its iterates are $A^{(n)} (x) = A (T^{n - 1} x) A (T^{n - 2} x) \dots A (T x) A (x), A^{(0)} (x) = I,$ the product of the matrices encountered along the orbit, in order. These satisfy the cocycle identity $A^{(n + m)} (x) = A^{(m)} (T^{n} x) A^{(n)} (x),$ which is exactly the statement that $A^{(n)}$ is the derivative-style chain rule for the skew product $(x, v) \mapsto (T x, A (x) v)$ on $X \times R^{d}$ . The leading geometric example is the derivative cocycle of a smooth map $f$ of a manifold preserving a measure $μ$ : take $A (x) = D_{x} f$ acting on the tangent space (in a fixed measurable frame), so $A^{(n)} (x) = D_{x} f^{n}$ by the chain rule.

Definition (integrability). The cocycle is log-integrable (satisfies the Oseledets integrability condition) if $\int_{X} lo g^{+} ∥ A (x) ∥ d μ (x) < \infty and \int_{X} lo g^{+} ∥ A (x)^{- 1} ∥ d μ (x) < \infty,$ where $lo g^{+} t = max (lo g t, 0)$ . The first condition controls forward growth; the second, needed for the two-sided theorem, controls backward growth.

Definition (Lyapunov exponent of a vector). For $x \in X$ and $0 \neq = v \in R^{d}$ , the forward Lyapunov exponent is $λ (x, v) = n \to \infty lim sup \frac{1}{n} lo g ∥ A^{(n)} (x) v ∥.$ The function $v \mapsto λ (x, v)$ takes finitely many values $λ_{1} (x) > \dots > λ_{r (x)} (x)$ , and ${v : λ (x, v) \leq λ_{i} (x)} \cup {0}$ is a linear subspace $V_{i} (x)$ (the Lyapunov filtration): a $lim sup$ of log-growth is subadditive in $max$ over sums of vectors, forcing the level sets to be nested subspaces. The number $λ_{i} (x)$ is the $i$ -th Lyapunov exponent and $d_{i} (x) = dim V_{i} (x) - dim V_{i + 1} (x)$ its multiplicity.

Definition (Oseledets splitting and regularity). A point $x$ is Lyapunov-Perron regular (forward and backward) if there is a measurable splitting $R^{d} = E_{1} (x) \oplus \dots \oplus E_{r} (x)$ , equivariant under the cocycle in the sense $A (x) E_{i} (x) = E_{i} (T x)$ , such that for every $0 \neq = v \in E_{i} (x)$ , $n \to \pm \infty lim \frac{1}{n} lo g ∥ A^{(n)} (x) v ∥ = λ_{i},$ with a genuine two-sided limit (not merely $lim sup$ ), and the exponents satisfy the trace identity $\sum_{i} d_{i} λ_{i} = lim_{n} \frac{1}{n} lo g ∣ det A^{(n)} (x) ∣$ . The subspaces $E_{i} (x)$ are the Oseledets subspaces; the splitting refines the filtration via $V_{i} (x) = ⨁_{j \geq i} E_{j} (x)$ , and on regular points the forward filtration $V_{∙}$ and the analogous backward filtration are transverse, which is what upgrades $lim sup$ to $lim$ .

Counterexamples to common slips Intermediate+

The exponents are not the eigenvalue logs of $A (x)$ . Lyapunov exponents are growth rates of the product $A^{(n)} (x)$ , not of a single matrix. Two matrices each with eigenvalues on the unit circle (e.g. rotations by different angles composed with shears) can have a strictly positive top exponent for their product. The relevant object is the singular-value growth of the long product, governed by Kingman, not the spectrum of $A (x)$ .
The Oseledets splitting is not the eigenspace decomposition and need not be orthogonal. The subspaces $E_{i} (x)$ are generally neither orthogonal nor continuous in $x$ — only measurable. Expecting an orthonormal Oseledets basis, or continuity, is the standard error; the angles between the $E_{i} (x)$ can degenerate along the orbit (though only subexponentially, which is what regularity guarantees).
Forward existence gives only a filtration, not a splitting. The one-sided theorem produces the nested subspaces $V_{1} ⊋ V_{2} ⊋ \dots$ where each $λ (x, v) = λ_{i}$ for $v \in V_{i} ∖ V_{i + 1}$ , but the complementary subspaces $E_{i}$ realising a genuine two-sided limit require the invertible (two-sided) theorem. Claiming a canonical $E_{i}$ from forward data alone conflates the filtration with its splitting.
Integrability is essential. Without $\int lo g^{+} ∥ A ∥ d μ < \infty$ the Kingman input fails and the limits can be $+ \infty$ on a positive-measure set; the exponents need not exist. The condition is the multiplicative analogue of the $L^{1}$ hypothesis in Birkhoff's theorem 37.02.03.
The top exponent is a $sup$ over directions, attained generically, not on a special line. $λ_{1} (x) = max_{v} λ (x, v)$ is achieved by almost every $v$ (those with a component outside $V_{2} (x)$ ); the slower exponents are attained only on the measure-zero union of lower filtration pieces. Reading $λ_{1}$ off a fixed coordinate axis fails unless that axis avoids $V_{2}$ .

Key theorem with proof Intermediate+

Theorem (Furstenberg-Kesten; existence of the top exponent). Let $A : X \to GL (d, R)$ be a log-integrable cocycle over an ergodic measure-preserving system $(X, B, μ, T)$ . Then for $μ$ -almost every $x$ the limit $λ_{1} = n \to \infty lim \frac{1}{n} lo g ∥ A^{(n)} (x) ∥ = n \geq 1 in f \frac{1}{n} \int_{X} lo g ∥ A^{(n)} (x) ∥ d μ (x)$ exists, is finite, and is $μ$ -a.e. equal to the constant $λ_{1}$ (the top Lyapunov exponent). More generally, for each $1 \leq k \leq d$ , $n \to \infty lim \frac{1}{n} lo g ∥ \land^{k} A^{(n)} (x) ∥ = λ_{1} + \dots + λ_{k} a.e.,$ where $λ_{1} \geq λ_{2} \geq \dots \geq λ_{d}$ are the exponents listed with multiplicity.

Proof. Fix $k$ and define $a_{n} (x) = lo g ∥ \land^{k} A^{(n)} (x) ∥$ . The exterior power is a homomorphism, $\land^{k} (B A) = (\land^{k} B) (\land^{k} A)$ , so submultiplicativity of the operator norm gives, using the cocycle identity $A^{(n + m)} (x) = A^{(m)} (T^{n} x) A^{(n)} (x)$ , $∥ \land^{k} A^{(n + m)} (x) ∥ \leq ∥ \land^{k} A^{(m)} (T^{n} x) ∥ \cdot ∥ \land^{k} A^{(n)} (x) ∥.$ Taking logs yields the subadditive cocycle relation $a_{n + m} (x) \leq a_{n} (x) + a_{m} (T^{n} x),$ which is precisely the hypothesis of Kingman's subadditive ergodic theorem 37.02.03, with the stationarity $a_{m} \circ T^{n}$ supplied by measure-preservation. Log-integrability gives $\int a_{1}^{+} d μ < \infty$ : indeed $∥ \land^{k} A (x) ∥ \leq ∥ A (x) ∥^{k}$ , so $a_{1}^{+} (x) \leq k lo g^{+} ∥ A (x) ∥$ , integrable by hypothesis. A lower bound $\int a_{n} d μ \geq - c n$ follows from $∥ \land^{k} A^{(n)} ∥ \geq ∣ det A^{(n)} ∣^{k / d}$ together with $lo g ∣ det A^{(n)} (x) ∣ = \sum_{j = 0}^{n - 1} lo g ∣ det A (T^{j} x) ∣$ and the integrability of $lo g ∣ det A ∣$ (bounded by $lo g^{+} ∥ A ∥ + lo g^{+} ∥ A^{- 1} ∥$ up to constants), so the time constant $in f_{n} \frac{1}{n} \int a_{n} d μ =: c_{k}$ is finite. Kingman's theorem then gives $\frac{1}{n} a_{n} (x) = \frac{1}{n} lo g ∥ \land^{k} A^{(n)} (x) ∥ n \to \infty c_{k} a.e. and in L^{1},$ with $c_{k}$ constant by ergodicity. Setting $Λ_{k} = c_{k}$ and $λ_{k} = Λ_{k} - Λ_{k - 1}$ (with $Λ_{0} = 0$ ) produces a list $λ_{1} \geq \dots \geq λ_{d}$ ; the monotonicity $λ_{1} \geq \dots \geq λ_{d}$ is the concavity of $k \mapsto Λ_{k}$ , which holds because singular values are nonincreasing and $∥ \land^{k} A^{(n)} ∥ = σ_{1} \dots σ_{k}$ for the singular values $σ_{1} \geq \dots \geq σ_{d}$ of $A^{(n)}$ . The case $k = 1$ is the top-exponent statement. $□$

Bridge. The Furstenberg-Kesten theorem builds toward the full Oseledets splitting and appears again in the volume-growth formula of the next section, where the same exterior-power trick recovers each exponent as a difference of singular-value sums. The foundational reason the limits exist is that submultiplicativity of operator norms becomes subadditivity of logs, and this is exactly the structure Kingman's theorem 37.02.03 was built to handle — the multiplicative ergodic theorem is the multiplicative shadow of the subadditive one, just as Birkhoff is its additive boundary case. Putting these together, Furstenberg-Kesten generalises the scalar law of large numbers to noncommuting matrix products: the central insight is that singular-value growth, not eigenvalue spectrum, is the invariant that survives the noncommutativity, and the bridge from the top exponent $λ_{1}$ to the entire spectrum $λ_{1} > \dots > λ_{r}$ is the passage from $∥ A^{(n)} ∥$ to the exterior-power norms $∥ \land^{k} A^{(n)} ∥$ . This is dual to the eigenvalue picture in the special case of a constant cocycle $A (x) \equiv A$ , where the exponents are the logs of the moduli of the eigenvalues of $A$ and Oseledets reduces to the Jordan decomposition.

Exercises Intermediate+

Exercise 4 (medium, symbolic).

Let $A^{(n)} (x)$ have singular values $σ_{1}^{(n)} \geq \dots \geq σ_{d}^{(n)} > 0$ . Show that $∥ \land^{k} A^{(n)} (x) ∥ = σ_{1}^{(n)} \dots σ_{k}^{(n)}$ , and deduce that $\frac{1}{n} lo g ∥ \land^{k} A^{(n)} ∥ \to λ_{1} + \dots + λ_{k}$ identifies $λ_{k} = lim_{n} \frac{1}{n} lo g σ_{k}^{(n)}$ .

Hint

Use the singular value decomposition $A^{(n)} = U Σ V^{*}$ ; the exterior power $\land^{k}$ acts diagonally on the basis $e_{i_{1}} \land \dots \land e_{i_{k}}$ with eigenvalue $σ_{i_{1}} \dots σ_{i_{k}}$ .

Answer

Write $A^{(n)} = U Σ V^{*}$ with $Σ = diag (σ_{1}^{(n)}, \dots, σ_{d}^{(n)})$ and $U, V$ orthogonal. Then $\land^{k} A^{(n)} = (\land^{k} U) (\land^{k} Σ) (\land^{k} V^{*})$ , and $\land^{k} U$ , $\land^{k} V^{*}$ are orthogonal on $\land^{k} R^{d}$ , so the operator norm equals that of $\land^{k} Σ$ . On the orthonormal basis ${e_{i_{1}} \land \dots \land e_{i_{k}} : i_{1} < \dots < i_{k}}$ , $\land^{k} Σ$ is diagonal with entries $σ_{i_{1}}^{(n)} \dots σ_{i_{k}}^{(n)}$ ; the largest is $σ_{1}^{(n)} \dots σ_{k}^{(n)}$ since the singular values are ordered. Hence $∥ \land^{k} A^{(n)} ∥ = σ_{1}^{(n)} \dots σ_{k}^{(n)}$ . Taking $\frac{1}{n} lo g$ and using the Furstenberg-Kesten limits $Λ_{k} = λ_{1} + \dots + λ_{k}$ , the difference $Λ_{k} - Λ_{k - 1} = lim_{n} \frac{1}{n} lo g σ_{k}^{(n)} = λ_{k}$ .

Exercise 5 (medium, symbolic).

Show that the set ${v : λ (x, v) \leq c} \cup {0}$ is a linear subspace of $R^{d}$ for each $c$ , where $λ (x, v) = lim sup_{n} \frac{1}{n} lo g ∥ A^{(n)} (x) v ∥$ .

Hint

For nonzero $u, w$ , bound $∥ A^{(n)} (x) (u + w) ∥$ by $∥ A^{(n)} (x) u ∥ + ∥ A^{(n)} (x) w ∥$ and take $\frac{1}{n} lo g$ of a sum of two exponentially-growing quantities.

Answer

Scaling $v$ by a nonzero constant leaves $λ (x, v)$ unchanged, so the set is closed under scalar multiplication (and contains $0$ by convention). For additivity, let $λ (x, u) \leq c$ and $λ (x, w) \leq c$ . By the triangle inequality $∥ A^{(n)} (x) (u + w) ∥ \leq ∥ A^{(n)} (x) u ∥ + ∥ A^{(n)} (x) w ∥$ , and for two sequences $p_{n}, q_{n} > 0$ one has $lim sup_{n} \frac{1}{n} lo g (p_{n} + q_{n}) = max (lim sup_{n} \frac{1}{n} lo g p_{n}, lim sup_{n} \frac{1}{n} lo g q_{n})$ , because the larger term dominates exponentially. Hence $λ (x, u + w) \leq max (λ (x, u), λ (x, w)) \leq c$ . So the set is closed under addition, hence a subspace. This nesting over $c$ produces the Lyapunov filtration $V_{1} \supseteq V_{2} \supseteq \dots$ .

Exercise 6 (medium, numeric).

A cocycle over a two-point system alternates between $B = (2001)$ and $C = (1003)$ along the orbit $x, T x, x, T x, \dots$ with $A (x) = B$ , $A (T x) = C$ , each visited with frequency $1/2$ . Compute the top Lyapunov exponent. Give it as $\frac{1}{2} lo g 6$ in decimal to three places.

Hint

Over two steps the product is $C B = diag (2, 3)$ ; its norm growth per step is half the log of the largest diagonal entry of the two-step product, but track both diagonal entries' geometric mean rate.

Answer

Because $B$ and $C$ are diagonal they commute, and over $2 n$ steps $A^{(2 n)} (x) = (C B)^{n} = diag (2^{n}, 3^{n})$ , with norm $3^{n}$ . Then $\frac{1}{2 n} lo g 3^{n} = \frac{1}{2} lo g 3 \approx 0.549$ is the top exponent along the second axis. The first axis gives $\frac{1}{2} lo g 2 \approx 0.347$ . The top exponent is $λ_{1} = \frac{1}{2} lo g 3 \approx 0.549$ . (The hint's $\frac{1}{2} lo g 6$ is the sum $λ_{1} + λ_{2} = \frac{1}{2} lo g 6 \approx 0.896$ , the area growth rate; the per-axis rates are $\frac{1}{2} lo g 3$ and $\frac{1}{2} lo g 2$ .)

Exercise 7 (hard, symbolic).

Prove the trace identity: for a log-integrable cocycle over an ergodic system, $\sum_{i} d_{i} λ_{i} = lim_{n} \frac{1}{n} lo g ∣ det A^{(n)} (x) ∣ = \int_{X} lo g ∣ det A ∣ d μ$ a.e., where the sum is over distinct exponents with multiplicities $d_{i}$ .

Hint

Apply Birkhoff's theorem to $f (x) = lo g ∣ det A (x) ∣$ , using $lo g ∣ det A^{(n)} (x) ∣ = \sum_{j = 0}^{n - 1} f (T^{j} x)$ ; separately, $\sum_{i} d_{i} λ_{i} = Λ_{d} = lim_{n} \frac{1}{n} lo g ∥ \land^{d} A^{(n)} ∥$ .

Answer

The determinant is multiplicative, so $det A^{(n)} (x) = \prod_{j = 0}^{n - 1} det A (T^{j} x)$ and $lo g ∣ det A^{(n)} (x) ∣ = \sum_{j = 0}^{n - 1} f (T^{j} x)$ with $f = lo g ∣ det A ∣$ , which is integrable since $∣ lo g ∣ det A ∣∣ \leq d (lo g^{+} ∥ A ∥ + lo g^{+} ∥ A^{- 1} ∥)$ . Birkhoff's pointwise theorem 37.02.03 gives $\frac{1}{n} lo g ∣ det A^{(n)} (x) ∣ \to \int f d μ = \int lo g ∣ det A ∣ d μ$ a.e. (constant by ergodicity). On the other hand, $∥ \land^{d} A^{(n)} ∥ = ∣ det A^{(n)} ∣$ (the top exterior power is the determinant), so by Furstenberg-Kesten $\frac{1}{n} lo g ∣ det A^{(n)} ∣ \to Λ_{d} = λ_{1} + \dots + λ_{d} = \sum_{i} d_{i} λ_{i}$ , the full sum of exponents with multiplicity. Equating the two limits gives $\sum_{i} d_{i} λ_{i} = \int lo g ∣ det A ∣ d μ$ . For a volume-preserving cocycle ( $∣ det A ∣ \equiv 1$ ) the exponents sum to zero.

Exercise 8 (hard, symbolic).

Suppose the cocycle is conformal, $A (x) = ρ (x) O (x)$ with $ρ (x) > 0$ scalar and $O (x)$ orthogonal. Show all $d$ Lyapunov exponents coincide and equal $\int_{X} lo g ρ d μ$ , so the Oseledets splitting collapses to a single block ( $r = 1$ , $E_{1} = R^{d}$ ).

Hint

For any unit vector $v$ , $∥ A^{(n)} (x) v ∥ = (\prod_{j < n} ρ (T^{j} x)) ∥ O^{(n)} (x) v ∥$ and the orthogonal product preserves norm.

Answer

Write $A^{(n)} (x) = (\prod_{j = 0}^{n - 1} ρ (T^{j} x)) O^{(n)} (x)$ , where $O^{(n)} (x) = O (T^{n - 1} x) \dots O (x)$ is orthogonal (a product of orthogonal matrices), using that the scalars commute past the orthogonal factors. For any nonzero $v$ , $∥ A^{(n)} (x) v ∥ = (\prod_{j < n} ρ (T^{j} x)) ∥ O^{(n)} (x) v ∥ = (\prod_{j < n} ρ (T^{j} x)) ∥ v ∥$ since orthogonal maps are isometries. Hence $\frac{1}{n} lo g ∥ A^{(n)} (x) v ∥ = \frac{1}{n} \sum_{j < n} lo g ρ (T^{j} x) + \frac{1}{n} lo g ∥ v ∥$ , and Birkhoff 37.02.03 sends the first term to $\int lo g ρ d μ$ and the second to $0$ . The limit is independent of $v$ , so every direction has the same exponent $λ = \int lo g ρ d μ$ ; there is a single exponent of multiplicity $d$ , and the Oseledets splitting is the whole space $E_{1} (x) = R^{d}$ . Conformality erases all directional distinction.

Advanced results Master

Theorem 1 (Oseledets multiplicative ergodic theorem, one-sided; Oseledets 1968). Let $A : X \to GL (d, R)$ be a log-integrable cocycle over a measure-preserving system $(X, B, μ, T)$ . Then for $μ$ -a.e. $x$ there exist a number $r (x) \in {1, \dots, d}$ , exponents $λ_{1} (x) > \dots > λ_{r (x)} (x)$ with multiplicities $d_{i} (x)$ summing to $d$ , and a filtration $R^{d} = V_{1} (x) ⊋ V_{2} (x) ⊋ \dots ⊋ V_{r (x)} (x) ⊋ V_{r (x) + 1} (x) = {0}$ by measurable subspaces with $A (x) V_{i} (x) = V_{i} (T x)$ , such that $n \to \infty lim \frac{1}{n} lo g ∥ A^{(n)} (x) v ∥ = λ_{i} (x) for v \in V_{i} (x) ∖ V_{i + 1} (x) .$ The matrices $(A^{(n)} (x)^{*} A^{(n)} (x))^{1/ (2 n)}$ converge to a positive-definite matrix $Λ (x)$ whose eigenvalues are $e^{λ_{i} (x)}$ with multiplicities $d_{i} (x)$ ; the $V_{i} (x)$ are the spans of the eigenspaces of $Λ (x)$ for eigenvalues $\leq e^{λ_{i} (x)}$ . On an ergodic system the data $r, λ_{i}, d_{i}$ are a.e. constant ^{[Oseledets 1968]}.

Theorem 2 (two-sided Oseledets theorem and the splitting). If in addition $T$ is invertible and $\int lo g^{+} ∥ A^{- 1} ∥ d μ < \infty$ , then a.e. $x$ is Lyapunov-Perron regular: the forward filtration $V_{∙} (x)$ and the backward filtration $V_{∙} (x)$ (built from $A^{(- n)}$ ) are transverse, and their intersection produces the Oseledets splitting $R^{d} = E_{1} (x) \oplus \dots \oplus E_{r} (x)$ with $E_{i} (x) = V_{i} (x) \cap V_{i} (x)$ , $dim E_{i} (x) = d_{i} (x)$ , equivariant $A (x) E_{i} (x) = E_{i} (T x)$ , and $n \to \pm \infty lim \frac{1}{n} lo g ∥ A^{(n)} (x) v ∥ = λ_{i} for every 0 \neq = v \in E_{i} (x),$ a genuine two-sided limit. The angles $∠ (E_{i} (T^{n} x), E_{j} (T^{n} x))$ decay subexponentially: $\frac{1}{n} lo g sin ∠ \to 0$ , the tempering property that makes the splitting "as good as block-diagonal" on the level of exponential rates ^{[Oseledets 1968]}.

Theorem 3 (Raghunathan's exterior-power proof). The MET follows from the Furstenberg-Kesten convergence of $\frac{1}{n} lo g ∥ \land^{k} A^{(n)} (x) ∥$ for all $k$ together with the convergence of the associated singular directions. Raghunathan reduced the existence of the limiting matrix $Λ (x)$ to applying the subadditive ergodic theorem to each exterior power and an analytic argument that the singular subspaces of $A^{(n)} (x)$ converge; this is the cleanest modern route, replacing Oseledets' original singular-value-decomposition estimates ^{[Raghunathan 1979]}.

Theorem 4 (filtration vs. splitting; non-uniqueness one-sided). The one-sided theorem (Theorem 1) determines the filtration $V_{∙} (x)$ canonically but does not canonically determine a complement to each $V_{i + 1} (x)$ inside $V_{i} (x)$ : any measurable equivariant choice of complement gives a valid "fast/slow" decomposition, and only the two-sided data of Theorem 2 pins the splitting uniquely. This is the structural reason that non-invertible smooth maps (endomorphisms) carry an Oseledets filtration but not, in general, an Oseledets splitting, and why Pesin theory for endomorphisms is technically heavier than for diffeomorphisms ^{[Ruelle 1979]}.

Theorem 5 (gateway to nonuniform hyperbolicity; Pesin sets). When the cocycle is the derivative cocycle $A (x) = D_{x} f$ of a $C^{1 + α}$ diffeomorphism preserving $μ$ , the Oseledets splitting partitions the tangent space into $E^{s} (x) = ⨁_{λ_{i} < 0} E_{i} (x)$ , $E^{c} (x) = ⨁_{λ_{i} = 0} E_{i} (x)$ , and $E^{u} (x) = ⨁_{λ_{i} > 0} E_{i} (x)$ . On the set $Λ$ where all exponents are nonzero the system is nonuniformly hyperbolic, and Pesin's stable-manifold theorem constructs local stable and unstable manifolds tangent to $E^{s}, E^{u}$ whose sizes vary measurably (tempered) along orbits. This is the foundation of Pesin theory: entropy via the Pesin-Ruelle inequality $h_{μ} (f) \leq \int \sum_{λ_{i} > 0} d_{i} λ_{i} d μ$ with equality (Pesin's formula) for SRB / smooth invariant measures ^{[Pesin 1977]}.

Synthesis. The Oseledets theorem is one statement read at several depths, and the foundational reason its pieces cohere is that each is the multiplicative image of a single additive fact — the convergence of $\frac{1}{n} lo g ∥ \land^{k} A^{(n)} ∥$ supplied by Kingman 37.02.03. The top exponent is that fact at $k = 1$ ; the full spectrum is exactly the same fact swept across all $k$ , the difference $Λ_{k} - Λ_{k - 1}$ extracting $λ_{k}$ , and this is the central insight that noncommuting matrix products have a law of large numbers once one passes from spectrum to singular values. The filtration is what forward growth alone can see; the splitting is dual to it, obtained by intersecting forward with backward filtrations, and putting these together with regularity upgrades every $lim sup$ to a two-sided $lim$ — the bridge from measurable asymptotics to genuine invariant geometry. This generalises the eigenvalue/Jordan picture of a single matrix (the constant-cocycle case) to the random and nonautonomous setting, and it is exactly the linear-algebraic engine that Pesin theory mounts on the derivative cocycle: nonuniform hyperbolicity is the Oseledets splitting with no zero exponents, and the stable-manifold theory, the entropy formula, and the SRB-measure construction are its nonlinear consequences.

Full proof set Master

Proposition 1 (subadditivity of exterior-power log-norms). For a cocycle $A$ and each $k$ , the family $a_{m, n} (x) = lo g ∥ \land^{k} A^{(n - m)} (T^{m} x) ∥$ is subadditive: $a_{0, n} \leq a_{0, m} + a_{m, n}$ with $a_{m, n} = a_{0, n - m} \circ T^{m}$ .

Proof. The cocycle identity gives $A^{(n)} (x) = A^{(n - m)} (T^{m} x) A^{(m)} (x)$ , and $\land^{k}$ is a homomorphism, so $\land^{k} A^{(n)} (x) = \land^{k} A^{(n - m)} (T^{m} x) \cdot \land^{k} A^{(m)} (x)$ . Submultiplicativity of the operator norm yields $∥ \land^{k} A^{(n)} (x) ∥ \leq ∥ \land^{k} A^{(n - m)} (T^{m} x) ∥ \cdot ∥ \land^{k} A^{(m)} (x) ∥$ , and taking logs gives $a_{0, n} (x) \leq a_{0, m} (x) + a_{m, n} (x)$ with $a_{m, n} (x) = lo g ∥ \land^{k} A^{(n - m)} (T^{m} x) ∥ = a_{0, n - m} (T^{m} x)$ . The integrability $\int a_{0, 1}^{+} d μ \leq k \int lo g^{+} ∥ A ∥ d μ < \infty$ holds by $∥ \land^{k} A ∥ \leq ∥ A ∥^{k}$ . $□$

Proposition 2 (the limiting matrix exists; Oseledets' route). For a.e. $x$ , the symmetric positive-definite matrices $Λ_{n} (x) = (A^{(n)} (x)^{*} A^{(n)} (x))^{1/ (2 n)}$ converge to a positive-definite matrix $Λ (x)$ , whose eigenvalues are $e^{λ_{i} (x)}$ .

Proof (sketch following Oseledets/Raghunathan). The eigenvalues of $Λ_{n} (x)$ are $(σ_{j}^{(n)})^{1/ n}$ , the $n$ -th roots of the singular values of $A^{(n)} (x)$ . By Proposition 1 and Kingman 37.02.03, $\frac{1}{n} lo g (σ_{1}^{(n)} \dots σ_{k}^{(n)}) = \frac{1}{n} lo g ∥ \land^{k} A^{(n)} ∥ \to Λ_{k}$ a.e. for each $k$ , so each $\frac{1}{n} lo g σ_{k}^{(n)} = \frac{1}{n} lo g (σ_{1} \dots σ_{k}) - \frac{1}{n} lo g (σ_{1} \dots σ_{k - 1}) \to Λ_{k} - Λ_{k - 1} = λ_{k}$ , giving convergence of the eigenvalues of $Λ_{n} (x)$ to $e^{λ_{k}}$ . Convergence of the eigenspaces requires the gaps $λ_{i} > λ_{i + 1}$ : where consecutive limiting eigenvalues are distinct, the spectral projections of $Λ_{n} (x)$ onto the bands of singular values converge (a perturbation argument, since the eigenvalue gaps of $Λ_{n}$ stay bounded below in the relevant blocks), and the limiting projections cut out the filtration $V_{i} (x)$ . Equivariance $A (x) V_{i} (x) = V_{i} (T x)$ follows from $A^{(n)} (T x) A (x) = A^{(n + 1)} (x)$ and uniqueness of the limiting decomposition. $□$

Proposition 3 (limit along the filtration). For a.e. $x$ and $v \in V_{i} (x) ∖ V_{i + 1} (x)$ , $lim_{n} \frac{1}{n} lo g ∥ A^{(n)} (x) v ∥ = λ_{i} (x)$ .

Proof. Decompose $v = \sum_{j \geq i} v_{j}$ in the limiting spectral basis of $Λ (x)$ with $v_{i} \neq = 0$ (this is what $v \in / V_{i + 1}$ means). Then $∥ A^{(n)} (x) v ∥^{2} = ⟨ A^{(n) *} A^{(n)} v, v ⟩ = ⟨ Λ_{n}^{2 n} v, v ⟩$ , and $Λ_{n}^{2 n} = A^{(n) *} A^{(n)}$ has eigenvalues $(σ_{j}^{(n)})^{2} \approx e^{2 n λ_{j}}$ along the converging eigenspaces. The component of $v$ in the eigenspace of exponent $λ_{i}$ is nonzero and dominates: $∥ A^{(n)} (x) v ∥^{2} ≍ e^{2 n λ_{i}} ∥ v_{i} ∥^{2} (1 + o (1))$ , the faster bands being absent (those are the $V_{j}$ with $j < i$ , to which $v$ is orthogonal in the limiting basis) and the slower bands negligible. Taking $\frac{1}{2 n} lo g$ gives $\frac{1}{n} lo g ∥ A^{(n)} (x) v ∥ \to λ_{i}$ . The control of the cross terms uses the subexponential decay of the angles between the $Λ_{n}$ -eigenspaces and their limits, which is the tempering supplied by the Kingman convergence being $L^{1}$ as well as a.e. $□$

Proposition 4 (top exponent as a variational quantity). $λ_{1} (x) = lim_{n} \frac{1}{n} lo g ∥ A^{(n)} (x) ∥ = max_{0 \neq = v} λ (x, v)$ , and the max is attained for a.e. $v$ (Lebesgue) on the unit sphere.

Proof. The operator norm is $∥ A^{(n)} (x) ∥ = σ_{1}^{(n)} = max_{∥ v ∥ = 1} ∥ A^{(n)} (x) v ∥$ , so $\frac{1}{n} lo g ∥ A^{(n)} ∥ = max_{v} \frac{1}{n} lo g ∥ A^{(n)} v ∥ \to λ_{1}$ by Furstenberg-Kesten, and $λ (x, v) \leq λ_{1}$ for all $v$ with equality whenever $v \in / V_{2} (x)$ . Since $V_{2} (x)$ is a proper subspace it has Lebesgue measure zero on the sphere, so the top exponent is attained for a.e. direction. The exceptional directions are exactly the lower filtration $V_{2} (x)$ , of positive codimension. $□$

Connections Master

The subadditive ergodic theorem of Kingman 37.02.03 is the analytic engine of the entire unit: applied to $lo g ∥ \land^{k} A^{(n)} ∥$ it delivers every Lyapunov exponent, and the multiplicative ergodic theorem is precisely its multiplicative reformulation. Where Birkhoff handles additive cocycles $\sum_{k < n} f \circ T^{k}$ , Kingman handles the submultiplicative matrix-norm growth, and Oseledets is the structural unpacking of that single convergence statement into an invariant splitting.
The measure-preserving-system framework 38.04.01 supplies the base dynamics over which the cocycle lives: the cocycle identity is the chain rule for the skew product $(x, v) \mapsto (T x, A (x) v)$ over $(X, T)$ , measure-preservation gives the stationarity that Kingman requires, and the induced-map and tower technology of that unit reappears when one localises the MET to Pesin sets of nearly-constant exponent. Ergodicity there is what collapses the exponents to constants here.
Lyapunov stability by the direct method 02.12.08 is the deterministic, single-equilibrium ancestor of the Lyapunov exponent: there a Lyapunov function certifies the sign of the linearisation's spectrum at one fixed point, here the Oseledets exponents are the orbit-averaged generalisation valid at almost every point of a measure rather than at an isolated equilibrium. The negative-exponent directions $E^{s}$ are the measurable, nonautonomous analogue of a stable eigenspace, and Pesin's stable manifolds are the nonlinear integral of that linear datum.
The ergodic-decomposition and entropy theory of dynamical systems 38.06.01 consumes the Oseledets data through the Pesin-Ruelle inequality $h_{μ} (f) \leq \int \sum_{λ_{i} > 0} d_{i} λ_{i} d μ$ and Pesin's entropy formula: positive Lyapunov exponents are the source of metric entropy for smooth systems, and the Margulis-Ruelle bound and its equality case (SRB measures) are stated in exactly the exponent-and-multiplicity language built here.
The hyperbolicity and structural-stability theory 38.03.01 is the uniform special case: when the Oseledets splitting is continuous and the angles are bounded below uniformly (not merely tempered) one recovers a uniformly hyperbolic splitting $T_{x} M = E^{s} \oplus E^{u}$ . Oseledets relaxes uniform hyperbolicity to its measurable, nonuniform shadow, which is why nonuniform hyperbolicity is "Anosov almost everywhere with measurable, possibly degenerating, constants."

Historical & philosophical context Master

Valery Oseledets proved the multiplicative ergodic theorem in his 1968 paper in the Trudy Moskovskogo Matematicheskogo Obshchestva ^{[Oseledets 1968]}, building on the 1960 work of Furstenberg and Kesten ^{[Furstenberg-Kesten 1960]} on products of random matrices, which had established the existence of the top exponent for stationary matrix products but not the full splitting. Oseledets' insight was that the entire Lyapunov spectrum and an equivariant decomposition of the tangent space could be extracted at once, by studying the convergence of the symmetric matrices $(A^{(n) *} A^{(n)})^{1/2 n}$ ; his original argument was a direct singular-value analysis. The theorem's centrality grew once Kingman's 1968 subadditive ergodic theorem ^{[Kingman 1968]} revealed the proper abstract engine, and M. S. Raghunathan gave in 1979 a streamlined exterior-power proof ^{[Raghunathan 1979]} that reduced the MET to applying Kingman across all exterior powers, the route now standard in the textbooks of Viana and Barreira-Pesin.

The theorem became the gateway to nonuniform hyperbolicity through Yakov Pesin's 1977 Russian Math. Surveys memoir ^{[Pesin 1977]}, which mounted the Oseledets splitting on the derivative cocycle of a smooth measure-preserving diffeomorphism and constructed stable and unstable manifolds wherever the exponents are nonzero, proving the entropy formula $h_{μ} (f) = \int \sum_{λ_{i} > 0} d_{i} λ_{i} d μ$ for smooth invariant measures. David Ruelle's 1979 IHÉS paper ^{[Ruelle 1979]} recast the theory in the language of differentiable dynamics and the Margulis-Ruelle inequality, and connected the exponents to the dimension theory and SRB measures that organise the statistical description of chaotic attractors. The numerical computation of Lyapunov exponents — by the algorithm of Benettin, Galgani, Giorgilli, and Strelcyn that periodically reorthonormalises an evolving frame — made the exponents the practical diagnostic of chaos across physics, from the Lorenz system to celestial mechanics.

Bibliography Master

@article{Oseledets1968,
  author  = {Oseledets, Valery I.},
  title   = {A multiplicative ergodic theorem: Lyapunov characteristic numbers for dynamical systems},
  journal = {Trudy Moskovskogo Matematicheskogo Obshchestva},
  volume  = {19},
  year    = {1968},
  pages   = {179--210}
}

@article{FurstenbergKesten1960,
  author  = {Furstenberg, Harry and Kesten, Harry},
  title   = {Products of random matrices},
  journal = {Annals of Mathematical Statistics},
  volume  = {31},
  number  = {2},
  year    = {1960},
  pages   = {457--469}
}

@article{Kingman1968,
  author  = {Kingman, John F. C.},
  title   = {The ergodic theory of subadditive stochastic processes},
  journal = {Journal of the Royal Statistical Society, Series B},
  volume  = {30},
  number  = {3},
  year    = {1968},
  pages   = {499--510}
}

@article{Raghunathan1979,
  author  = {Raghunathan, Madabusi S.},
  title   = {A proof of Oseledec's multiplicative ergodic theorem},
  journal = {Israel Journal of Mathematics},
  volume  = {32},
  number  = {4},
  year    = {1979},
  pages   = {356--362}
}

@article{Pesin1977,
  author  = {Pesin, Yakov B.},
  title   = {Characteristic Lyapunov exponents and smooth ergodic theory},
  journal = {Russian Mathematical Surveys},
  volume  = {32},
  number  = {4},
  year    = {1977},
  pages   = {55--114}
}

@article{Ruelle1979,
  author  = {Ruelle, David},
  title   = {Ergodic theory of differentiable dynamical systems},
  journal = {Publications Math\'ematiques de l'IH\'ES},
  volume  = {50},
  year    = {1979},
  pages   = {27--58}
}

@book{KatokHasselblatt1995,
  author    = {Katok, Anatole and Hasselblatt, Boris},
  title     = {Introduction to the Modern Theory of Dynamical Systems},
  publisher = {Cambridge University Press},
  series    = {Encyclopedia of Mathematics and its Applications},
  volume    = {54},
  year      = {1995}
}

@book{BarreiraPesin2007,
  author    = {Barreira, Luis and Pesin, Yakov},
  title     = {Nonuniform Hyperbolicity: Relative Stability and Smooth Ergodic Theory},
  publisher = {Cambridge University Press},
  series    = {Encyclopedia of Mathematics and its Applications},
  volume    = {115},
  year      = {2007}
}

@book{Viana2014,
  author    = {Viana, Marcelo},
  title     = {Lectures on Lyapunov Exponents},
  publisher = {Cambridge University Press},
  series    = {Cambridge Studies in Advanced Mathematics},
  volume    = {145},
  year      = {2014}
}

@book{Arnold1998,
  author    = {Arnold, Ludwig},
  title     = {Random Dynamical Systems},
  publisher = {Springer},
  series    = {Springer Monographs in Mathematics},
  year      = {1998}
}

Prerequisites

38.04.01
37.02.03
02.12.08

Tier anchors

beginner: Katok-Hasselblatt 1995 *Introduction to the Modern Theory of Dynamical Systems* (Cambridge) Suppl. (informal: stretching directions of a repeatedly applied linear map); Strogatz 2015 *Nonlinear Dynamics and Chaos* 2e (Westview) §9.3 (Lyapunov exponents as the rate of separation of nearby trajectories)
intermediate: Katok-Hasselblatt 1995 *Introduction to the Modern Theory of Dynamical Systems* (Cambridge) Supplement (S.1-S.5: cocycles, Lyapunov exponents, the multiplicative ergodic theorem); Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) §10.2-10.3 (subadditive ergodic theorem and its consequences)
master: Katok-Hasselblatt 1995 *Introduction to the Modern Theory of Dynamical Systems* (Cambridge) Supplement and Ch. S; Barreira-Pesin 2007 *Nonuniform Hyperbolicity: Relative Stability and Smooth Ergodic Theory* (Cambridge) Ch. 1-3 (Lyapunov exponents, the Oseledets theorem, regularity); Arnold 1998 *Random Dynamical Systems* (Springer Monographs) Ch. 3-4 (the MET for random and continuous cocycles); Viana 2014 *Lectures on Lyapunov Exponents* (Cambridge) Ch. 4 (Oseledets via subadditivity, Kingman input)

References

Oseledets — A multiplicative ergodic theorem: Lyapunov characteristic numbers for dynamical systems · Trudy Moskovskogo Matematicheskogo Obshchestva 19 (1968), 179-210 (English: Trans. Moscow Math. Soc. 19, 197-231)
Furstenberg-Kesten — Products of random matrices · Annals of Mathematical Statistics 31 (1960), 457-469
Kingman — The ergodic theory of subadditive stochastic processes · Journal of the Royal Statistical Society Series B 30 (1968), 499-510
Raghunathan — A proof of Oseledec's multiplicative ergodic theorem · Israel Journal of Mathematics 32 (1979), 356-362
Pesin — Characteristic Lyapunov exponents and smooth ergodic theory · Uspekhi Matematicheskikh Nauk 32:4 (1977), 55-112 (Russian Math. Surveys 32:4, 55-114)
Ruelle — Ergodic theory of differentiable dynamical systems · Publications Mathématiques de l'IHÉS 50 (1979), 27-58
Katok-Hasselblatt — Introduction to the Modern Theory of Dynamical Systems · Cambridge University Press 1995, Supplement (cocycles, Lyapunov exponents, MET)

Estimated time

beginner: 20m
intermediate: 58m
master: 95m