Kaplan–Yorke Dimension in Chaotic Systems

• Kaplan–Yorke Dimension is a measure estimating the fractal structure of chaotic attractors by partitioning phase space into expanding and contracting directions.
• It effectively bounds the Hausdorff dimension and often approximates the information dimension in various dissipative dynamical systems.
• Both analytical derivations and numerical methods, such as the Benettin–Shimada algorithm, are used to compute this dimension in diverse systems from low-dimensional maps to high-dimensional networks.

The Kaplan–Yorke dimension, also referred to as the @@@@3@@@@, provides a widely adopted quantitative framework for describing the effective fractal dimension of chaotic attractors in dissipative dynamical systems. It is formulated in terms of the Lyapunov spectrum, serving as an upper bound for the Hausdorff dimension, and, in many systems, closely approximates or coincides with the information dimension of the physical invariant measure. The Kaplan–Yorke dimension is foundational in the rigorous theory of dimension estimation and is central to both numerical diagnostics and analytical estimates in the study of chaotic flows, maps, and high-dimensional dissipative systems.

1. Definition and Theoretical Foundation

Given a dynamical system (either flow or discrete map) with associated Lyapunov exponents $$\lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_n$$, the Kaplan–Yorke (KY) dimension is defined as follows. Let $S_j = \sum_{i=1}^j \lambda_i$ and let $j$ be the largest integer such that $S_j \geq 0$ but $S_{j+1} < 0$. Then,

$D_{KY} = j + \frac{S_j}{|\lambda_{j+1}|}$

This construction partitions phase space according to the number of expanding (positive) Lyapunov directions, with the fractional part indicating the extent to which the next contracting direction is filled before net contraction dominates. The formula applies universally to both continuous-time flows and discrete-time maps, with the ordered spectrum arising from the Oseledec multiplicative ergodic theorem in the presence of an invariant probability measure (Kuznetsov, 2016, Leonov et al., 2015).

In rigorous terms, the Lyapunov dimension is formulated via the singular-value function of the linearized flow or map, as established by Douady–Oesterlé and further developed in the work of Leonov and others. The Kaplan–Yorke dimension provides an upper bound for the Hausdorff dimension of invariant sets and is invariant under smooth diffeomorphisms (Leonov et al., 2015, Kuznetsov, 2016).

2. Analytical and Numerical Computation

Analytically, the computation of $D_{KY}$ proceeds from either the Lyapunov exponents evaluated along trajectories or, in certain cases, from explicit formulas derived at equilibria or periodic orbits where the maximum local Kaplan–Yorke dimension is achieved. Notable closed-form results exist for classical dissipative systems such as the Lorenz and Glukhovsky–Dolzhansky models. For example, the attractor dimension in the G–D model is given by

$$\dim_L(K) = 3 - \frac{2(\sigma+b+1)}{\sigma+1 + \sqrt{(\sigma-1)^2 + 4\sigma r}}$$

for appropriate parameter regimes, providing exact agreement with the Lyapunov dimension at the global attractor (Kuznetsov et al., 2015).

Numerically, the Lyapunov spectrum is computed either by QR or SVD orthonormalization along tangent-space evolutions, widely known as the Benettin–Shimada or Eckmann–Ruelle algorithms. This is performed by periodic re-orthonormalization and accumulation of logarithmic scaling rates for tangent vectors, yielding estimates for $\{\lambda_i\}$. The KY dimension is then assembled as per the above formula (Kuznetsov, 2016, Carroll, 2019). In spatially extended or high-dimensional systems (e.g., PDEs, networks, or reservoir computers), this approach scales to the full phase-space dimension provided proper algorithmic controls on numerical stability (Carroll, 2019, Edson et al., 2019).

3. Relation to Fractal and Information Dimensions

The original conjecture of Kaplan and Yorke posited that the information dimension $D_1$ (as defined via the scaling of the entropy of box-count probabilities) of the invariant measure coincides with the Lyapunov dimension $D_{KY}$ for typical dissipative attractors. Direct computational comparisons of $D_{KY}$, $D_1$, and box-counting dimensions have been performed in both theoretical models and numerical experiments.

In the context of dissipative Baker maps, area-wise and point-wise estimates of the information dimension can disagree with the Kaplan–Yorke dimension by several percent (as observed for the N2 map), or agree to within numerical error as in the N3 map (Hoover et al., 2019). In uni-directional skew-product systems, such as coupled skinny Baker's maps, robust violation of the equality $D_1 = D_{KY}$ has been established when coupling is into the strongly stable direction, while prevalence of equality holds in generic bi-directional or weakly stable coupling regimes (Gröger et al., 2013). These results highlight both the success and limitations of the Kaplan–Yorke conjecture and motivate refined analyses of dimension equality mechanisms.

4. Behavior in High-Dimensional and Complex Systems

Extensive high-dimensional systems, such as PDEs (e.g., the Kuramoto–Sivashinsky equation), turbulent flows, or large dynamical networks, display rich scaling of $D_{KY}$ with system size and bifurcation parameters. In such cases, the Kaplan–Yorke dimension grows linearly with phase-space extent or system size, capturing the transition from low-dimensional chaos to turbulence.

For instance, in the Kuramoto–Sivashinsky PDE, the dimension scales approximately as $D_{KY} \propto 0.2 L$ with domain length $L$, confirming the extensive nature of chaos and the almost linear addition of active degrees of freedom per spatial subdomain (Edson et al., 2019). Similarly, in 2D Kolmogorov flows, $D_{KY}$ saturates above a transition, providing a quantitative measure of the large-scale degrees of freedom and scaling linearly with forcing wavenumber rather than with total available Fourier modes, with $D_{KY}^{\mathrm{sat}} \sim k_f$ (Vinograd et al., 9 Feb 2026).

In time-delay coupled chaotic networks, the Kaplan–Yorke dimension can exhibit discontinuous jumps—for example, at the transition to synchronization, $\Delta D_{KY} \approx (N-1)(\tau+1)$ in networks of $N$ Bernoulli maps with delay $\tau$, signaling macroscopic reduction in the effective attractor dimension (Zeeb et al., 2012).

5. Rigorous Foundations and Invariance Properties

The Lyapunov (Kaplan–Yorke) dimension is invariant under $C^1$ diffeomorphisms of phase space, a property critical for the reliability of both analytical estimates (e.g., by the Leonov method) and for interpretation of numerical results across different coordinate systems (Kuznetsov, 2016).

Rigorous bounding results due to Douady–Oesterlé and Constantin–Foias–Temam denote that the Hausdorff dimension of negatively invariant, compact sets satisfies $\dim_H K \leq d_L(\varphi^t, K)$, where $d_L$ is obtained either via the singular value function or Lyapunov exponents, and for flows as the infimum over time. These bounds hold without ergodicity assumptions and provide a mathematical underpinning for the broad applicability of the Kaplan–Yorke dimension in dynamical systems (Leonov et al., 2015, Kuznetsov, 2016).

6. Illustrative Examples and Physical Relevance

Explicit calculation of the Kaplan–Yorke dimension is possible for classical low-dimensional maps and flows. Table 1 compiles characteristic formulas and situations:

System	Lyapunov Exponents	$D_{KY}$ Formula
2D map / ODE	$\lambda_1 > 0 > \lambda_2$	$1 + \lambda_1 / |\lambda_2|$
Hénon map at fixed point	$\ln|d_1|, \ln|d_2|$	$1 + \frac{\ln|d_1|}{|\ln|d_2||}$
Lorenz / G–D system	Explicit spectrum from Jacobian at equilibrium	Model-dependent closed formula (see above)
Bernoulli map network (delay)	Bands from characteristic polynomial	$D_{KY} \approx N(\tau+1)$ or discontinuity

These analytic and numerical examples affirm the utility of $D_{KY}$ as an operationally effective and physically meaningful measure of attractor complexity in chaotic dynamics.

7. Limitations, Failure Modes, and Extensions

The Kaplan–Yorke formula provides an upper bound for Hausdorff and information dimension but may overestimate the physical dimension in systems with non-generic invariant measures, non-exact-dimensionality, or special structure (e.g., product or skew-product systems with weak or highly anisotropic coupling). Robust failure of the dimension equality $D_1 = D_{KY}$ can occur (see skew-product systems (Gröger et al., 2013) and non-coincidence in certain Baker maps (Hoover et al., 2019)). In high-dimensional systems, incomplete computation of Lyapunov exponents, neutral directions due to continuous symmetries, or numerical uncertainties may affect both the computation and interpretation of $D_{KY}$ (Edson et al., 2019).

Despite these caveats, the Kaplan–Yorke dimension remains a fundamental metric, both in theory and in practical computational diagnostics, for the characterization of chaos, the assessment of model complexity, and the development of reduced-order modeling strategies in nonlinear dynamical systems.