Lyapunov Exponents: Definition, Computation, and the Chaotic Double Pendulum

Intuition Beginner

Set up two identical double pendulums. Release the first from an initial angle of 90 degrees for the upper arm and 90 degrees for the lower arm. Release the second from 90.001 degrees and 90 degrees -- a difference of one thousandth of a degree, far smaller than any human can see or any physical apparatus can control. For the first few seconds the two pendulums swing in unison. Then they start to drift. After ten or twenty seconds the motions are completely different: one pendulum spins clockwise while the other flips counter-clockwise. The tiny difference at the start has grown to dominate the entire state of the system.

This is sensitive dependence on initial conditions, the defining feature of deterministic chaos. The Lyapunov exponent is the number that tells you how fast this divergence happens. If the Lyapunov exponent is $\lambda$, then a tiny initial separation $\delta$ between two trajectories grows over time $t$ approximately as $\delta e^{\lambda t}$. When $\lambda$ is positive, the separation grows exponentially: the system is chaotic. When $\lambda$ is zero or negative, nearby trajectories stay close or converge: the motion is regular.

Think of the Lyapunov exponent as a speedometer for chaos. A larger positive value means faster divergence and more unpredictable dynamics. For the double pendulum at low energy (small swings), the Lyapunov exponent is zero and the motion is regular. At high energy (large swings, flips, and tumbling), the Lyapunov exponent becomes positive and the motion is chaotic. The transition between regular and chaotic motion as you increase the energy is one of the central phenomena in classical mechanics.

An $n$-dimensional system has $n$ Lyapunov exponents -- one for each independent direction in phase space. This is the Lyapunov spectrum. The largest one is the maximal Lyapunov exponent (MLE), and it is the single most important number for diagnosing chaos: if the MLE is positive, the system is chaotic. The other exponents tell you about contraction and conservation of volumes in phase space.

Visual Beginner

Quantity	Symbol	Meaning	Chaos signature
Maximal Lyapunov exponent	${\lambda }_1$	Fastest rate of trajectory divergence	${\lambda }_1>0$
Lyapunov spectrum	${\lambda }_1\ge {\lambda }_2\ge \cdots \ge {\lambda }_n$	Divergence rates in all directions	At least one positive
Lyapunov sum	${\lambda }_1+\cdots +{\lambda }_n$	Volume change rate	Zero (Hamiltonian)
Prediction horizon	$t_p\sim {\lambda }_1^{-1}\log (1/{\delta }_0)$	Time to lose predictability	Decreases as ${\lambda }_1$ grows

Worked example Beginner

The doubling map is the simplest chaotic system with a computable Lyapunov exponent. It acts on numbers between 0 and 1 by doubling and keeping only the fractional part: $x\to 2x(\mathrm{mod}1)$.

Start two trajectories at $x_0=0.3$ and $x_0^{\prime }=0.300001$. The initial separation is ${\delta }_0=0.000001$. After each iteration:

• Step 1: $x_1=0.6$, $x_1^{\prime }=0.600002$. Separation: $0.000002=2\times {\delta }_0$.
• Step 2: $x_2=0.2$, $x_2^{\prime }=0.200004$. Separation: $0.000004=4\times {\delta }_0$.
• Step 3: $x_3=0.4$, $x_3^{\prime }=0.400008$. Separation: $0.000008=8\times {\delta }_0$.

The separation doubles each step. After $n$ steps, ${\delta }_n=2^n{\delta }_0$. The Lyapunov exponent is the exponential growth rate: $\lambda =\log 2\approx 0.693$. This means the separation grows by a factor of $e^{0.693}=2$ per iteration.

The prediction horizon is the time at which the initial uncertainty has grown to the size of the entire system. Starting from ${\delta }_0=10^{-6}$ and the system size being 1:

$$t_p\approx \frac{\log (1/{\delta }_0)}{\lambda }=\frac{\log (10^6)}{\log 2}=\frac{6\log 10}{\log 2}\approx 20\text{ iterations}.$$

After about 20 iterations, the two trajectories are completely unrelated. The Lyapunov exponent sets a hard limit on how far into the future you can predict.

Check your understanding Beginner

Formal definition Intermediate+

Let ${\varphi }^t:M\to M$ be a flow on an $n$-dimensional manifold $M$. Fix a trajectory ${\varphi }^t(x_0)$ and let $\delta x(t)$ be an infinitesimal perturbation to the initial condition, evolved by the linearised equations (the tangent flow $D{\varphi }^t$).

The maximal Lyapunov exponent

The maximal Lyapunov exponent (MLE) at $x_0$ is

$${\lambda }_1(x_0)=\underset{t\to \infty }{\lim }\frac{1}{t}\log \frac{\Vert D{\varphi }^t(x_0)\cdot \delta x(0)\Vert }{\Vert \delta x(0)\Vert },$$

provided the limit exists and is independent of the initial perturbation $\delta x(0)$ (for generic perturbations). This measures the asymptotic exponential rate of stretching along the most expanding direction.

The Lyapunov spectrum

The full Lyapunov spectrum ${\lambda }_1\ge {\lambda }_2\ge \cdots \ge {\lambda }_n$ is obtained by considering the growth rates of all $k$-dimensional volume elements ($k=1,2,\dots ,n$). Equivalently, one tracks the growth of each principal axis of an infinitesimal ellipsoid transported by $D{\varphi }^t$. The $k$-th Lyapunov exponent is:

$${\lambda }_k=\underset{t\to \infty }{\lim }\frac{1}{t}\log {\sigma }_k(t),$$

where ${\sigma }_k(t)$ is the $k$-th largest singular value of $D{\varphi }^t(x_0)$. The existence of these limits for almost every point (with respect to an ergodic invariant measure) is guaranteed by the Oseledets multiplicative ergodic theorem (see Master section).

Properties of the Lyapunov spectrum

1. Volume contraction: For a flow that contracts volumes at rate $\alpha$, the sum of all Lyapunov exponents equals $\alpha$: ${\lambda }_1+\cdots +{\lambda }_n=\alpha$. For a Hamiltonian system, $\alpha =0$ (Liouville's theorem).

2. Symplectic pairing: For a Hamiltonian system with $n$ degrees of freedom, the Lyapunov exponents come in pairs $\pm {\lambda }_i$ for $i=1,\dots ,n$. There are also two exponents equal to zero (along the flow direction and the energy surface).

3. Ergodic invariance: If $\mu$ is an ergodic invariant measure, the Lyapunov exponents are constant $\mu$-almost everywhere. Different ergodic measures can give different Lyapunov spectra.

Numerical computation: the Wolf algorithm

Computing Lyapunov exponents numerically requires care because the tangent vectors grow exponentially and quickly exceed floating-point range. The standard approach (Wolf, Swift, Swinney, and Vastano, 1985) proceeds as follows:

1. Integrate the trajectory ${\varphi }^t(x_0)$ and simultaneously integrate the tangent equations (the variational equations for $D{\varphi }^t$) starting from an orthonormal basis $\{e_1,\dots ,e_n\}$.

2. At regular time intervals $\tau$, apply Gram-Schmidt orthonormalisation to the evolved basis $\{D{\varphi }^{\tau }{e}_1,\dots ,D{\varphi }^{\tau }{e}_n\}$. This prevents the vectors from collapsing onto the most expanding direction.

3. Record the logarithms of the normalisation factors at each step. The $k$-th Lyapunov exponent is the long-time average of the logarithm of the $k$-th Gram-Schmidt factor divided by $\tau$.

$${\lambda }_k\approx \frac{1}{N\tau }\sum _{j=1}^N\log r_k^{(j)},$$

where $r_k^{(j)}$ is the length of the $k$-th vector after subtracting its projection onto the previous $k-1$ vectors at the $j$-th renormalisation step.

The double pendulum equations

The double pendulum consists of two rigid rods of lengths $l_1,l_2$ and masses $m_1,m_2$, with the upper end of the first rod pivoted at a fixed point and the upper end of the second rod pivoted at the lower end of the first rod. The configuration is described by the angles ${\theta }_1$ (first rod from vertical) and ${\theta }_2$ (second rod from vertical).

The Lagrangian is:

$$L=\frac{1}{2}(m_1+m_2)l_1^2{\dot{\theta }}_1^2+\frac{1}{2}{m}_2{l}_2^2{\dot{\theta }}_2^2+m_2{l}_1{l}_2{\dot{\theta }}_1{\dot{\theta }}_2\cos ({\theta }_1-{\theta }_2)+(m_1+m_2)g{l}_1\cos {\theta }_1+m_{2}g{l}_2\cos {\theta }_2.$$

The equations of motion are a coupled pair of second-order nonlinear ODEs. Setting $m_1=m_2=m$ and $l_1=l_2=l$ for simplicity, and defining the state vector $\mathbf{y}=({\theta }_1,{\theta }_2,{\omega }_1,{\omega }_2)$ where ${\omega }_i={\dot{\theta }}_i$, the system becomes $\dot{\mathbf{y}}=F(\mathbf{y})$ with $F$ given by the Euler-Lagrange equations:

$${\dot{\omega }}_1=\frac{-g(2m)\sin {\theta }_1-mg\sin ({\theta }_1-2{\theta }_2)-2\sin ({\theta }_1-{\theta }_2)m({\omega }_2^{2}l+{\omega }_1^{2}l\cos ({\theta }_1-{\theta }_2))}{2l(m(3-\cos (2{\theta }_1-2{\theta }_2)))},$$ $${\dot{\omega }}_2=\frac{2\sin ({\theta }_1-{\theta }_2)({\omega }_1^{2}l(2m)+g(2m)\cos {\theta }_1+{\omega }_2^{2}lm\cos ({\theta }_1-{\theta }_2))}{2l(m(3-\cos (2{\theta }_1-2{\theta }_2)))}.$$

These equations are integrated numerically (e.g. RK4 or adaptive Runge-Kutta) to produce trajectories. The tangent equations $D{\varphi }^t$ are obtained by differentiating $F$ with respect to $\mathbf{y}$ and integrating the resulting linear system alongside the trajectory. For the double pendulum, $DF$ is a $4\times 4$ matrix of partial derivatives that depends on the current state. The Lyapunov exponents are extracted from the long-time behaviour of the tangent map using the Wolf algorithm.

Numerical results for the double pendulum

For the standard double pendulum ($m_1=m_2=m$, $l_1=l_2=l$), numerical computation reveals:

• Low energy ($E\ll mgl$): The Lyapunov spectrum is approximately $(0,0,0,0)$ -- all exponents are zero, and the motion is regular (two coupled oscillators).
• Moderate energy ($E\sim mgl$): One pair of exponents becomes nonzero: ${\lambda }_1\approx -{\lambda }_4>0$ while ${\lambda }_2\approx {\lambda }_3\approx 0$. The system is partially chaotic.
• High energy ($E\gg mgl$): The maximal Lyapunov exponent grows with energy. Typical values for $E/(mgl)\approx 10$ give ${\lambda }_1\sim 1$ (in units where $t$ is measured in $\sqrt{l/g}$).

The transition from regular to chaotic motion is not sharp: as energy increases, chaotic regions in phase space grow at the expense of regular islands, following the pattern described by KAM theory 09.08.01 and Poincare sections 09.08.02.

Worked examples Intermediate+

Computing the Lyapunov exponent of the logistic map

The logistic map $x_{n+1}=r{x}_n(1-x_n)$ is a one-dimensional map, so it has a single Lyapunov exponent:

$$\lambda =\underset{N\to \infty }{\lim }\frac{1}{N}\sum _{n=0}^{N-1}\log |f^{\prime }(x_n)|=\underset{N\to \infty }{\lim }\frac{1}{N}\sum _{n=0}^{N-1}\log |r(1-2x_n)|.$$

For $r=4$, the logistic map is chaotic. The invariant measure is $\mu (x)=1/(\pi \sqrt{x(1-x)})$ on $[0,1]$. The Lyapunov exponent is:

$$\lambda ={\int }_0^1\log |4(1-2x)|\cdot \frac{dx}{\pi \sqrt{x(1-x)}}=\log 2\approx \mathrm{0.693.}$$

This can also be verified numerically: iterate $x_{n+1}=4x_n(1-x_n)$ starting from any generic $x_0$, accumulate $\log |4(1-2x_n)|$ at each step, and average. After $10^4$ iterations the estimate converges to $\lambda \approx 0.693$.

For $r=3.5$, the orbit converges to a period-4 cycle. The Lyapunov exponent is $\lambda \approx -0.486<0$, confirming regular (periodic) motion.

The Henon map Lyapunov spectrum

For the Henon map $x^{\prime }=1-a{x}^2+y$, $y^{\prime }=bx$ with $a=1.4$, $b=0.3$ (dissipative), the Lyapunov spectrum has two exponents. The Jacobian determinant is $-b=-0.3$, so ${\lambda }_1+{\lambda }_2=\log |b|=\log (0.3)\approx -1.204$. Numerical computation gives ${\lambda }_1\approx 0.42$ and ${\lambda }_2\approx -1.62$, confirming ${\lambda }_1+{\lambda }_2\approx -1.20$. The positive ${\lambda }_1$ confirms chaos.

The double pendulum: a numerical protocol

To compute the Lyapunov spectrum of the double pendulum:

1. Choose parameters: $m_1=m_2=1$, $l_1=l_2=1$, $g=9.81$.
2. Set initial conditions: e.g. ${\theta }_1=2.0$ rad, ${\theta }_2=2.0$ rad, ${\omega }_1={\omega }_2=0$.
3. Compute the total energy: $E\approx 19.62$ J (well into the chaotic regime).
4. Integrate the equations of motion using RK4 with time step $dt=0.001$ s.
5. Simultaneously integrate the tangent equations (4 coupled linear ODEs with state-dependent coefficients) starting from a $4\times 4$ identity matrix.
6. Every $\tau =1.0$ s, apply Gram-Schmidt to the 4 column vectors and record the logarithms of the normalisation factors.
7. Average over $N=10^4$ renormalisation steps (total integration time: $10^4$ s).

Typical result: ${\lambda }_1\approx 4.9$, ${\lambda }_2\approx 0.0$, ${\lambda }_3\approx 0.0$, ${\lambda }_4\approx -4.9$. The pairing ${\lambda }_1=-{\lambda }_4$ and ${\lambda }_2={\lambda }_3=0$ is the expected pattern for a Hamiltonian system with 2 degrees of freedom (two symplectic pairs, with one pair at zero corresponding to the energy surface and the flow direction). The positive ${\lambda }_1$ confirms chaotic dynamics.

Key derivation Intermediate+

Deriving the Lyapunov exponent for a one-dimensional map

For a one-dimensional map $x_{n+1}=f(x_n)$, two trajectories starting at $x_0$ and $x_0+{\delta }_0$ evolve as:

$$x_n+{\delta }_n\approx f^n(x_0+{\delta }_0)\approx f^n(x_0)+(f^n{)}^{\prime }(x_0)\cdot {\delta }_0.$$

The perturbation after $n$ steps is ${\delta }_n=(f^n{)}^{\prime }(x_0)\cdot {\delta }_0$. By the chain rule:

$$(f^n{)}^{\prime }(x_0)=f^{\prime }(x_0)\cdot f^{\prime }(x_1)\cdots f^{\prime }(x_{n-1})=\prod _{k=0}^{n-1}{f}^{\prime }(x_k).$$

The exponential growth rate of the perturbation is:

$$\lambda =\underset{n\to \infty }{\lim }\frac{1}{n}\log |(f^n{)}^{\prime }(x_0)|=\underset{n\to \infty }{\lim }\frac{1}{n}\sum _{k=0}^{n-1}\log |f^{\prime }(x_k)|.$$

For a chaotic system, this limit exists and is positive for almost every initial condition (with respect to the natural invariant measure). The Lyapunov exponent converts the product of derivatives along the orbit into an average of logarithmic derivatives, which is amenable to both analytical and numerical computation.

Bridge

The Lyapunov exponent is the bridge between the geometric picture of chaos (stretching and folding in the horseshoe 09.08.02) and its quantitative measurement. The horseshoe stretches nearby points apart at an exponential rate; the Lyapunov exponent is precisely that exponential rate. The Smale-Birkhoff theorem guarantees the existence of a horseshoe when there is a transverse homoclinic intersection, and the resulting Lyapunov exponent is $\log 2$ per horseshoe iterate (the stretching factor of the horseshoe is 2). For physical systems like the double pendulum, the Lyapunov exponent serves as the primary diagnostic: it is positive in the chaotic regime and zero in the regular regime, providing a clear, computable boundary between order and chaos.

Exercises Intermediate+

Lean formalization Intermediate+

Mathlib has no theory of Lyapunov exponents. The closest infrastructure is the derivative of a flow (the tangent map), which exists in the ODE theory within Mathlib/Analysis/ODE/, but the following are all absent:

1. Oseledets multiplicative ergodic theorem: the existence of Lyapunov exponents as limits, the Oseledets splitting $T_{x}M=E_1\oplus \cdots \oplus E_k$, and the measurability of the Lyapunov filtration.

2. Pesin stable manifold theorem: the existence of smooth stable manifolds at points with negative Lyapunov exponents, and the local product structure.

3. Pesin entropy formula: the equality $h_{\mu }(f)=\int {\sum }_{{\lambda }_i>0}{\lambda }_{i}d\mu$ relating the metric entropy to the sum of positive Lyapunov exponents.

4. Kaplan-Yorke formula: the conjectured relation between Lyapunov exponents and the information dimension of an attractor.

5. Wolf algorithm: numerical computation of Lyapunov exponents requires verified numerical ODE integration, which Mathlib does not support.

lean_status: none reflects the gap. This unit ships without a Lean module and is reviewer-attested.

Advanced results Master

The Oseledets multiplicative ergodic theorem

The existence of Lyapunov exponents as well-defined limits is guaranteed by one of the deepest results in ergodic theory.

Theorem (Oseledets 1968). Let $f:M\to M$ be a $C^1$ diffeomorphism of a compact Riemannian manifold $M$, and let $\mu$ be an $f$-invariant probability measure. Then for $\mu$-almost every $x\in M$, there exists a splitting of the tangent space $T_{x}M=E_1(x)\oplus E_2(x)\oplus \cdots \oplus E_k(x)$ and real numbers ${\lambda }_1>{\lambda }_2>\cdots >{\lambda }_k$ such that:

1. For every $v\in E_i(x)\setminus \{0\}$: ${\lim }_{n\to \pm \infty }\frac{1}{n}\log \Vert D{f}^n(x)\cdot v\Vert ={\lambda }_i$.
2. The splitting is $Df$-invariant: $Df(x)E_i(x)=E_i(f(x))$.
3. The numbers ${\lambda }_i$ and the dimensions $\dim E_i(x)$ are $\mu$-measurable and constant on each ergodic component.

For flows ${\varphi }^t$, the analogous statement holds with $n$ replaced by $t$ and $D{f}^n$ by $D{\varphi }^t$. The Oseledets splitting is sometimes called the Lyapunov filtration and the subspaces $E_i(x)$ are the Oseledets subspaces.

The proof relies on the Furstenberg-Kesten theorem (1960), which establishes that the limits ${\lim }_{n\to \infty }\frac{1}{n}\log \Vert D{f}^n\Vert$ exist almost surely for matrix cocycles over measure-preserving transformations. Oseledets' contribution was to refine this to obtain the full splitting and the individual exponents by applying the Furstenberg-Kesten theorem to the exterior powers ${\bigwedge }^{k}D{f}^n$.

Pesin's entropy formula

The Kolmogorov-Sinai (KS) entropy $h_{\mu }(f)$ measures the rate of information production by the dynamics with respect to the invariant measure $\mu$. Pesin's formula connects this purely measure-theoretic quantity to the geometric Lyapunov exponents.

Theorem (Pesin 1977). Let $f$ be a $C^{1+\alpha }$ diffeomorphism ($\alpha >0$) of a compact Riemannian manifold $M$, and let $\mu$ be an $f$-invariant probability measure that is absolutely continuous with respect to the Riemannian volume. Then:

$$h_{\mu }(f)={\int }_M\sum _{{\lambda }_i(x)>0}{\lambda }_i(x)d\mu (x).$$

That is, the metric entropy equals the integral of the sum of positive Lyapunov exponents.

The converse direction is the Pesin entropy formula (sometimes called the Ruelle inequality in the forward direction): Ruelle (1978) proved that $h_{\mu }(f)\le \int {\sum }_{{\lambda }_i^{+}}d\mu$ for any invariant measure, and Pesin proved equality when $\mu$ is absolutely continuous. The Ledrappier-Young theorem (1985) characterises exactly which measures satisfy equality: the SRB (Sinai-Ruelle-Bowen) measures.

For the double pendulum with the Liouville measure on the energy surface, Pesin's formula gives the entropy in terms of the positive Lyapunov exponents computed numerically. The KS entropy quantifies the rate at which the double pendulum generates information about its initial conditions -- the rate at which prediction becomes impossible.

The Kaplan-Yorke conjecture

The Kaplan-Yorke conjecture (1979) relates the Lyapunov spectrum to the fractal dimension of an attractor. Given the ordered Lyapunov exponents ${\lambda }_1\ge {\lambda }_2\ge \cdots \ge {\lambda }_n$, define $j$ as the largest integer with ${\lambda }_1+\cdots +{\lambda }_j\ge 0$. The Lyapunov dimension is:

$$d_L=j+\frac{{\lambda }_1+\cdots +{\lambda }_j}{|{\lambda }_{j+1}|}.$$

Conjecture (Kaplan-Yorke). For "typical" attractors, the information dimension $d_I$ equals the Lyapunov dimension $d_L$.

The conjecture has been proved for uniformly hyperbolic systems (where $d_I=d_L$ holds for SRB measures by a result of Ledrappier, 1981) and for certain classes of two-dimensional maps. It remains open in full generality. The physical intuition is that the expanding directions stretch the attractor (contributing to its dimension) while the contracting directions compress it. The Lyapunov dimension interpolates between the integer dimensions by weighting the next direction by the ratio of the remaining expansion to the first contraction.

Relation to Kolmogorov-Sinai entropy

The KS entropy and the Lyapunov exponents are connected through a chain of inequalities and equalities:

1. Ruelle inequality: $h_{\mu }(f)\le \int {\sum }_{{\lambda }_i^{+}}d\mu$ for any invariant measure (1978).
2. Pesin equality: $h_{\mu }(f)=\int {\sum }_{{\lambda }_i^{+}}d\mu$ for smooth invariant measures (1977).
3. Margulis-Ruelle inequality: A refinement that accounts for the multiplicity of exponents.
4. Ledrappier-Young theorem: The equality $h_{\mu }(f)=\int {\sum }_{{\lambda }_i^{+}}d\mu$ holds if and only if $\mu$ has absolutely continuous conditional measures on the unstable manifolds (i.e., $\mu$ is an SRB measure).

For the double pendulum, the KS entropy (with respect to the Liouville measure on an energy surface) is given by Pesin's formula as the integral of the positive Lyapunov exponents. In the regular regime, all exponents are zero and the entropy is zero: no information is produced. In the chaotic regime, ${\lambda }_1>0$ and the entropy is positive: the system generates information at a rate proportional to ${\lambda }_1$.

Synthesis

The Lyapunov spectrum provides the quantitative fingerprint of chaotic dynamics. The Oseledets theorem guarantees that the exponents exist as well-defined limits for almost every trajectory. Pesin's formula connects the exponents to the metric entropy, establishing that chaos (positive Lyapunov exponents) is equivalent to information production (positive KS entropy). The Kaplan-Yorke conjecture links the exponents to the fractal geometry of attractors. Together, these results form the quantitative backbone of chaos theory: the horseshoe 09.08.02 provides the geometric mechanism, and the Lyapunov exponents provide the numbers that quantify how fast that mechanism operates.

Full proof set Master

Proposition 1 (Lyapunov exponent of the doubling map). The Lyapunov exponent of the doubling map $f(x)=2x(\mathrm{mod}1)$ is $\lambda =\log 2$ for Lebesgue-almost every $x\in [0,1)$.

Proof. The map $f$ is piecewise linear with $f^{\prime }(x)=2$ at every point where the derivative exists (i.e., all $x\ne 1/2$). The Lyapunov exponent is:

$$\lambda =\underset{n\to \infty }{\lim }\frac{1}{n}\sum _{k=0}^{n-1}\log |f^{\prime }(x_k)|=\underset{n\to \infty }{\lim }\frac{1}{n}\sum _{k=0}^{n-1}\log 2=\log 2.$$

The limit exists immediately (every term is $\log 2$) at every point whose orbit avoids $x=1/2$. The orbit of $1/2$ under $f$ is $\{1/2,0,0,\dots \}$ (a finite set), and its preimages are the dyadic rationals $k/2^n$, which form a countable (hence measure zero) set. For Lebesgue-almost every $x$, the orbit avoids $1/2$ and $\lambda =\log 2$. $\square$

Proposition 2 (Symplectic pairing of Lyapunov exponents). Let ${\varphi }^t$ be a Hamiltonian flow on a $2n$-dimensional symplectic manifold. Then the Lyapunov exponents satisfy ${\lambda }_{2n+1-i}=-{\lambda }_i$ for $i=1,\dots ,n$ (when ordered ${\lambda }_1\ge \cdots \ge {\lambda }_{2n}$).

Proof. The tangent map $D{\varphi }^t(x)$ is a symplectic matrix: $D{\varphi }^t(x{)}^{T}JD{\varphi }^t(x)=J$ where $J$ is the symplectic form. The symplectic eigenvalues come in reciprocal pairs: if $\sigma$ is a singular value of $D{\varphi }^t$, so is $1/\sigma$. This follows because if $D{\varphi }^{t}v=\sigma w$ for unit vectors $v,w$, then by symplecticity $D{\varphi }^t(Jv)=\frac{1}{\sigma }Jw$ (the symplectic adjoint). Taking logarithms and the limit $t\to \infty$: if the growth rate along $v$ is $\lambda$, the growth rate along $Jv$ is $-\lambda$. Therefore each exponent $+\lambda$ is paired with $-\lambda$. Combined with Liouville's theorem ($\sum {\lambda }_i=0$), the exponents are ${\lambda }_1\ge \cdots \ge {\lambda }_n\ge -{\lambda }_n\ge \cdots \ge -{\lambda }_1$. $\square$

Proposition 3 (Ruelle inequality). Let $f$ be a $C^1$ diffeomorphism of a compact manifold and $\mu$ an $f$-invariant probability measure. Then $h_{\mu }(f)\le {\int }_M{\sum }_{{\lambda }_i(x)>0}{\lambda }_i(x)d\mu (x)$.

Proof sketch (following Ruelle 1978). The metric entropy $h_{\mu }(f)$ is bounded by the volume growth rate of small balls under iteration: $h_{\mu }(f)\le {\mathrm{limsup}}_{n\to \infty }\frac{1}{n}\log \mu (B(x,ϵ){)}^{-1}$, averaged over $\mu$. The local volume expansion is controlled by the sum of the positive Lyapunov exponents via the Oseledets splitting: in the directions with ${\lambda }_i>0$, the volume of an infinitesimal parallelepiped grows as $e^{({\lambda }_1+\cdots +{\lambda }_j)n}$ where $j$ is the number of positive exponents. This volume growth bounds the rate at which $\mu$-measure of neighbourhoods changes, giving the inequality. $\square$

Proposition 4 (Lyapunov exponent of the logistic map at $r=4$). For the logistic map $f(x)=4x(1-x)$ with the invariant measure $\mu (x)=\frac{dx}{\pi \sqrt{x(1-x)}}$, the Lyapunov exponent is $\lambda =\log 2$.

Proof. The derivative is $f^{\prime }(x)=4-8x=4(1-2x)$. The Lyapunov exponent is:

$$\lambda ={\int }_0^1\log |4(1-2x)|\cdot \frac{dx}{\pi \sqrt{x(1-x)}}.$$

Substitute $x={\sin }^2\theta$ (so $dx=2\sin \theta \cos \theta d\theta$, $\sqrt{x(1-x)}=\sin \theta \cos \theta$, and $1-2x=\cos 2\theta$):

$$\lambda ={\int }_0^{\pi /2}\log |4\cos 2\theta |\cdot \frac{2\sin \theta \cos \theta d\theta }{\pi \sin \theta \cos \theta }=\frac{2}{\pi }{\int }_0^{\pi /2}\log (4|\cos 2\theta |)d\theta .$$

Splitting: $\frac{2}{\pi }{\int }_0^{\pi /2}\log 4d\theta +\frac{2}{\pi }{\int }_0^{\pi /2}\log |\cos 2\theta |d\theta =\log 4+\frac{2}{\pi }{\int }_0^{\pi /2}\log |\cos 2\theta |d\theta$.

The integral ${\int }_0^{\pi /2}\log |\cos 2\theta |d\theta =-\frac{\pi }{2}\log 2$ (a classical result). Therefore $\lambda =\log 4-\log 2=\log 2$. $\square$

Connections Master

• Hamiltonian mechanics 09.04.01 provides the symplectic structure that constrains the Lyapunov spectrum: the exponents come in pairs $\pm \lambda$ and their sum is zero (Liouville's theorem). The canonical coordinates $(q,p)$ define the tangent space in which the Lyapunov exponents are computed.

• Poincare sections and the horseshoe 09.08.02 provide the geometric mechanism (stretching and folding) that generates positive Lyapunov exponents. The horseshoe's stretching factor of 2 per iterate gives a Lyapunov exponent of $\log 2$.

• KAM theory 09.08.01 describes the regular islands where Lyapunov exponents are zero, surrounded by chaotic seas where the MLE is positive. The transition from regular to chaotic as energy increases corresponds to the destruction of KAM tori.

• Strange attractors 09.08.05 in dissipative systems are characterised by a positive MLE combined with a negative sum of exponents (volume contraction). The Kaplan-Yorke formula connects the Lyapunov spectrum to the fractal dimension of the attractor.

• Ergodic theory 38.03.01 develops the measure-theoretic framework (Oseledets theorem, Pesin formula, Ledrappier-Young theorem) that gives the Lyapunov exponents their rigorous meaning. The physics of the double pendulum provides the primary example.

Historical & philosophical context Master

Lyapunov and the stability of motion (1892)

Aleksandr Mikhailovich Lyapunov introduced his characteristic exponents in his 1892 doctoral dissertation "The General Problem of the Stability of Motion" at the University of Kharkov. Lyapunov's original motivation was the stability of rotating fluid masses (the figure of the Earth, studied earlier by Maclaurin, Jacobi, and Poincare). He developed what is now called Lyapunov's first method (or the indirect method): linearise the equations near an equilibrium and study the eigenvalues of the linearised system. For time-dependent or nonlinear systems where the eigenvalues alone do not determine stability, Lyapunov introduced the characteristic numbers (now called Lyapunov exponents) that measure the asymptotic growth rate of solutions.

Lyapunov's work was far ahead of its time. The connection between positive characteristic numbers and chaotic dynamics was not made until the 1960s-1970s, when the theory of dynamical systems matured. The modern definition of the Lyapunov spectrum (as opposed to Lyapunov's original characteristic numbers) is due to Oseledets (1968), who proved the existence of the spectrum for almost every trajectory with respect to an invariant measure.

Oseledets and the multiplicative ergodic theorem (1968)

Valentin Oseledets, a student of Sinai at Moscow State University, proved the multiplicative ergodic theorem in 1968. The theorem establishes that the Lyapunov exponents exist as well-defined limits for almost every point (with respect to an ergodic invariant measure) and that the tangent space splits into invariant subspaces corresponding to each exponent. Oseledets' proof uses the Furstenberg-Kesten theorem on the growth of norms of matrix products and a careful analysis of the exterior algebra of the tangent bundle. The result is one of the pillars of smooth ergodic theory, alongside the Pesin theory that builds on it.

Pesin theory (1977)

Yakov Pesin, working at the University of Syktyvkar and later at Penn State, developed the theory of smooth hyperbolic systems (now called Pesin theory) in a landmark 1977 paper. Pesin proved three fundamental results: (1) the existence of stable and unstable manifolds at almost every point with a nonzero Lyapunov exponent (the Pesin stable manifold theorem), (2) the entropy formula relating the KS entropy to the sum of positive Lyapunov exponents (Pesin's formula), and (3) the equivalence between positive Lyapunov exponents and the "hyperbolicity" of the system in a measure-theoretic sense. Pesin theory extends the classical theory of uniformly hyperbolic systems (Anosov, Smale) to the case of non-uniform hyperbolicity, where the expansion and contraction rates vary from point to point. This is precisely the situation for the double pendulum and most physical systems.

Wolf and the numerical computation (1985)

Alan Wolf and collaborators published their algorithm for computing Lyapunov exponents from time series in 1985, making the Lyapunov exponent a practical tool for experimental physics and engineering. The Wolf algorithm (Gram-Schmidt renormalisation of the tangent vectors) remains the standard method. Its importance is that it works with experimental data, not just equations: given a time series of measurements, one can reconstruct the dynamics (via delay embedding, following Takens 1981) and compute the Lyapunov spectrum, providing a direct experimental test for chaos.

Philosophical implications

The Lyapunov exponent quantifies the tension between determinism and predictability. A deterministic system can have a positive Lyapunov exponent, meaning that exact knowledge of the present state determines the future exactly, but approximate knowledge determines the future approximately only for a finite time proportional to ${\lambda }^{-1}\log (1/ϵ)$ where $ϵ$ is the precision of the initial measurement. For the double pendulum, with ${\lambda }_1\sim 5$ (in SI units for typical parameters), an initial precision of $10^{-6}$ radians gives a prediction horizon of about $t_p\sim \log (10^6)/5\approx 2.8$ seconds. Beyond this time, the motion is unpredictable in practice, even though the equations are deterministic. The Lyapunov exponent converts the philosophical puzzle of deterministic chaos into a concrete, measurable number.

Bibliography Master

• Lyapunov, A. M., "The General Problem of the Stability of Motion," doctoral dissertation, University of Kharkov (1892). English translation: Taylor & Francis (1992). The original definition of characteristic numbers (Lyapunov exponents) and the first and second methods of stability analysis.

• Oseledets, V. I., "A multiplicative ergodic theorem. Lyapunov characteristic numbers for dynamical systems," Trans. Moscow Math. Soc. 19 (1968), 197-231. The Oseledets theorem establishing the existence and properties of the Lyapunov spectrum for almost every trajectory.

• Pesin, Ya. B., "Characteristic Lyapunov exponents and smooth ergodic theory," Russian Math. Surveys 32 (1977), 55-114. The Pesin entropy formula, stable manifold theorem, and the foundations of non-uniform hyperbolicity theory.

• Ruelle, D., "An inequality for the entropy of differentiable maps," Bol. Soc. Brasil. Mat. 9 (1978), 83-87. The Ruelle inequality bounding metric entropy by the integral of positive Lyapunov exponents.

• Ledrappier, F. & Young, L.-S., "The metric entropy of diffeomorphisms. Part I: Characterization of measures satisfying Pesin's entropy formula," Ann. Math. 122 (1985), 509-539. Characterisation of SRB measures via the equality in Pesin's formula.

• Kaplan, J. L. & Yorke, J. A., "Chaotic behavior of multidimensional difference equations," in Functional Differential Equations and Approximation of Fixed Points, Springer Lecture Notes in Mathematics 730 (1979), 204-227. The Kaplan-Yorke conjecture relating Lyapunov exponents to fractal dimension.

• Wolf, A., Swift, J. B., Swinney, H. L. & Vastano, J. A., "Determining Lyapunov exponents from a time series," Physica D 16 (1985), 285-317. The Wolf algorithm for numerical computation of the Lyapunov spectrum.

• Barreira, L. & Pesin, Ya. B., Lyapunov Exponents and Smooth Ergodic Theory (AMS, 2002). A modern comprehensive treatment of the Oseledets theorem, Pesin theory, and non-uniform hyperbolicity.

• Katok, A. & Hasselblatt, B., Introduction to the Modern Theory of Dynamical Systems (Cambridge, 1995). Ch. 6-8: Lyapunov exponents, entropy, and the relation between them.

• Strogatz, S. H., Nonlinear Dynamics and Chaos, 2nd ed. (CRC Press, 2015). Ch. 10: Lyapunov exponents for one-dimensional maps.

• Taylor, J. R., Classical Mechanics (University Science Books, 2005). Ch. 13: Lyapunov exponents and chaos in Hamiltonian systems.

• Goldstein, H., Poole, C. P. & Safko, J., Classical Mechanics, 3rd ed. (Pearson, 2002). Ch. 11: Classical chaos and Lyapunov exponents.

• Furstenberg, H. & Kesten, H., "Products of random matrices," Ann. Math. Statist. 31 (1960), 457-469. The Furstenberg-Kesten theorem on growth rates of random matrix products, the foundation of the Oseledets theorem.