Lyapunov Dimension in Chaotic Systems

• Lyapunov Dimension is a quantitative measure that characterizes the fractal geometry of invariant sets in chaotic dynamical systems.
• It is computed via the Kaplan–Yorke formula by linking ordered Lyapunov exponents with the effective (often non-integer) dimensionality of long-time dynamics.
• Computational approaches, such as SVD and QR-based methods, are essential in applications ranging from fluid turbulence to electronic circuits.

The Lyapunov dimension provides a quantitative measure of the fractal geometry of invariant sets—such as chaotic attractors—in finite- and infinite-dimensional dynamical systems. It operationalizes the connection between local instability given by Lyapunov exponents and the effective dimensionality (often non-integer) of the subset of state space occupied by the system’s long-time dynamics. The Lyapunov dimension, most commonly computed via the Kaplan–Yorke formula, serves as a rigorous upper bound to the Hausdorff dimension under broad circumstances. Analytical, computational, and variational approaches for its estimation or exact computation have played a fundamental role in the theory and numerical practice of nonlinear dynamics, with applications spanning fluid turbulence, atmospheric models, electronic circuits, and more.

1. Mathematical Foundations and Definitions

Let $\dot{x} = f(x)$ describe an autonomous ODE on $\mathbb{R}^n$ ($f\in C^1$). Let $\phi^t(x_0)$ be the flow and $D\phi^t(x_0)$ the fundamental matrix of the variational equation along a trajectory. Denote by $$\sigma_1(t,x_0) \geq \sigma_2(t,x_0) \geq \dots \geq \sigma_n(t,x_0)>0$$ the singular values of $D\phi^t(x_0)$. The finite-time Lyapunov exponents (FTLEs) at $(T,\,x_0)$ are defined as

$$\lambda_i(T, x_0) = \frac{1}{T} \ln \sigma_i(T, x_0),\quad i=1,\ldots,n.$$

The finite-time local Lyapunov dimension at $x_0$, based on the ordered FTLEs, is given by the Kaplan–Yorke formula: $\mathbb{R}^n$0 where $\mathbb{R}^n$1 and, by convention, $\mathbb{R}^n$2 if all sums are negative.

The global finite-time Lyapunov dimension of a compact invariant set $\mathbb{R}^n$3 is defined as

$\mathbb{R}^n$4

The (asymptotic) Lyapunov dimension of $\mathbb{R}^n$5 is

$\mathbb{R}^n$6

The Douady–Oesterlé theorem ensures that for any compact invariant $\mathbb{R}^n$7,

$\mathbb{R}^n$8

where $\mathbb{R}^n$9 is the Hausdorff dimension.

2. Rigorous Properties and Coordinate Invariance

Two central types of Lyapunov exponents are widely referenced: the Lyapunov characteristic exponents (LCEs)—growth rates of columns of the fundamental matrix—and the (singular-value) Lyapunov exponents (LEs)—growth rates of the singular values. The latter provide the minimal growth rates over all possible bases and are used in the definition of the Lyapunov dimension.

A fundamental invariance property holds: the spectrum of Lyapunov exponents and the associated Lyapunov (Kaplan–Yorke) dimension are invariant under smooth changes of variables (diffeomorphisms) and under the choice of fundamental matrix solution. This invariance is established via inequalities for the singular values under matrix products and conjugations, as shown in (Kuznetsov et al., 2014).

Regularity or irregularity of the linearization—i.e., whether the sum of exponents matches the logarithmic growth of $f\in C^1$0—does not affect the existence or uniqueness of the Lyapunov dimension via the singular value definition; the invariance extends to these situations as well (Kuznetsov et al., 2014).

3. Analytical Methods and the Kaplan–Yorke Formula

The Kaplan–Yorke formula, introduced heuristically to connect local exponents to the effective (possibly fractal) dimension of chaotic attractors, is rigorously justified via properties of singular-value truncation and covering arguments. For an ordered set $f\in C^1$1 (finite-time or asymptotic), the dimension is

$f\in C^1$2

At each $f\in C^1$3 and $f\in C^1$4 (or for the limiting exponents as $f\in C^1$5), this formula yields the local Lyapunov dimension (Kuznetsov et al., 2015, Leonov et al., 2015, Kuznetsov et al., 2018, Kuznetsov, 2016).

For a flow or map, the maximum over $f\in C^1$6 of the local dimension at a given time is taken, and then the infimum over $f\in C^1$7 or liminf as $f\in C^1$8 gives the Lyapunov dimension of $f\in C^1$9. The resulting dimension is always an upper bound for the Hausdorff dimension of $\phi^t(x_0)$0 (Leonov et al., 2015).

Kaplan–Yorke dimension is pervasively used as a practical proxy for fractal analysis in numerical and theoretical studies of low-order and spatially extended systems (Plan et al., 2017).

4. Computational Approaches and Algorithms

The numerical computation of FTLEs and Lyapunov dimension proceeds by integrating the ODEs for both the system and its variational equations. Two main algorithms are standard (Kuznetsov et al., 2015, Kuznetsov et al., 2017, Kuznetsov et al., 2018, Leonov et al., 2015):

• SVD-based Method: Integrate the trajectory and variational system over $\phi^t(x_0)$1, construct $\phi^t(x_0)$2, and perform a singular value decomposition at each time point. The FTLEs are the (logarithmic) time-averaged singular values.
• QR-based Method (Periodic Re-orthonormalization/Benettin's Algorithm): Decompose time into intervals, at each interval QR-factor the variational flow, accumulate the logarithms of the norms of the orthogonalized vectors (diagonal $\phi^t(x_0)$3 entries), and average over time. This yields the (column-wise) LCEs, which agree with the LE only in special situations, but can be made to coincide with proper orthonormalization strategies and SVD sweeps (Kuznetsov et al., 2015, Kuznetsov et al., 2017, Leonov et al., 2015).

Numerical best practices include sufficiently small integration steps, frequent orthonormalization to avert loss of numerical independence, and long-time convergence checks for the FTLEs. For attractors with multiple basins or hidden attractors, careful domain sampling is necessary (Kuznetsov et al., 2015, Kuznetsov et al., 2015, Kuznetsov et al., 2018).

5. Analytical Estimation and Exactness

The Leonov direct Lyapunov method provides sharp analytical upper bounds (and in many cases exact formulas) for the Lyapunov dimension using symmetrized Jacobians and Lyapunov-like functions without explicit attractor localization (Kuznetsov, 2016, Leonov et al., 2015).

The method constructs a metric change and a scalar function $\phi^t(x_0)$4 so that the sum of the largest $\phi^t(x_0)$5 symmetrized eigenvalues plus $\phi^t(x_0)$6 times the next, plus $\phi^t(x_0)$7, is uniformly negative over $\phi^t(x_0)$8, implying that the Lyapunov dimension is less than $\phi^t(x_0)$9. This approach yields explicit formulas for classical examples, including the Lorenz system, Henon map, Tigan/Yang systems, and others: $D\phi^t(x_0)$0 for the global attractor of the Lorenz system under hyperbolicity hypotheses (Leonov et al., 2015, Kuznetsov et al., 2019).

In many self-excited attractors, the supremum of the local dimension is achieved at an unstable equilibrium or an unstable periodic orbit. In such cases, one obtains an "exact Lyapunov dimension" as the Kaplan–Yorke dimension of the eigenvalues of the Jacobian at that equilibrium (Kuznetsov et al., 2015, Kuznetsov et al., 2017, Kuznetsov et al., 2015, Leonov et al., 2015, Leonov et al., 2015).

6. Distinction Between Self-Excited and Hidden Attractors

A self-excited attractor is characterized by a basin intersecting a neighborhood of an unstable equilibrium; hidden attractors lack such an intersection. The computation of Lyapunov dimension for hidden attractors typically requires dedicated localization procedures (e.g., continuation, scanning absorbing sets, perpetual points) to identify representative trajectories. Both classes require scanning over grids and time windows due to non-ergodicity and potential for long transients. For hidden attractors, careful grid refinement and exhaustive search are often essential (Kuznetsov et al., 2015, Kuznetsov et al., 2015, Kuznetsov et al., 2018, Leonov et al., 2015).

In multistable settings or in the presence of hidden transient chaos, FTLE-based computation may overestimate dimension if not checked over sufficiently long orbits and multiple initial conditions. Explicit analytical bounds remain valid in all cases, but tight numerical evaluation requires fine sampling (Kuznetsov et al., 2015).

7. Applications, Limitations, and Open Problems

Lyapunov dimension has widespread application in characterizing the chaoticity and complexity of attractors in diverse nonlinear systems: hydrodynamic models, plasma-wave interaction (Rabinovich), electronic circuits (Chua), and spatially-extended and infinite-dimensional PDE models (e.g., elastic turbulence in the Oldroyd-B model, where rigorous upper bounds are connected to polymer stress gradients and the Weissenberg number) (Plan et al., 2017).

Estimation is limited to $D\phi^t(x_0)$1 smooth flows with invertible Jacobians and typically requires identification of an absorbing set or attractor localization. Finite-time diagnostics can be misleading for long transient chaos; high-precision and parallel sampling can partially mitigate these effects (Kuznetsov et al., 2015, Kuznetsov et al., 2018).

A key open conjecture is that, for a typical self-excited attractor $D\phi^t(x_0)$2, the Lyapunov dimension equals the local Kaplan–Yorke dimension at a corresponding unstable equilibrium: $D\phi^t(x_0)$3 for some equilibrium $D\phi^t(x_0)$4 in the basin. This remains unproven in full generality (Kuznetsov et al., 2015, Kuznetsov et al., 2017).

Future research aims to tighten these results in more general settings (multistability, non-smooth dynamics), refine the analytical–numerical interface, and develop algorithms robust to extended transient effects and high-dimensional settings (Kuznetsov et al., 2015).