Transverse Lyapunov Exponent in Turbulent DA

• Transverse Lyapunov Exponent is a metric that quantifies the exponential growth or decay of perturbations perpendicular to a synchronization manifold in coupled dynamical systems.
• It distinguishes the critical scale for data assimilation by identifying the wavenumber cutoff (k_c*) above which small-scale turbulence synchronizes with true flow dynamics.
• The framework links chaos synchronization and turbulence theory, offering a principled approach to target unstable transverse modes for optimal, cost-efficient algorithm design.

A transverse Lyapunov exponent (TLE) quantifies the stability of synchronization manifolds in coupled dynamical systems, encoding the exponential rate at which infinitesimal perturbations orthogonal (transverse) to a chosen invariant manifold grow or decay. In the context of fluid dynamics, specifically data assimilation (DA) in Navier–Stokes (NS) turbulence, the TLE provides a rigorous, dynamically intrinsic criterion for multiscale synchronization: it determines the critical small-scale length above which assimilation ensures the reconstructed turbulent fields are synchronized to the true flow. This concept originates in chaos synchronization theory and now underpins a comprehensive framework for understanding and optimizing DA in large-scale turbulent systems (Inubushi et al., 2023).

1. Transverse Lyapunov Exponent in Drive–Response Systems

Let $X$ denote a Hilbert space, with the coupled system

• Drive (“base”): $\dot{x}=F(x)$,
• Response (“fiber,” forced by $x$): $\dot{y}=G(x,y)$.

Assuming an invariant manifold $M= \{ (x,y)\ |\ y=h(x)\}$, its transverse stability is governed by perturbations $\delta\eta$ off $M$. The transverse Lyapunov exponent is

$$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$

when this limit exists. If $\lambda_\perp<0$, $M$ is locally stable (synchronizing); $\dot{x}=F(x)$0 indicates instability (desynchronization transverse to $\dot{x}=F(x)$1).

In DA for NS turbulence, $\dot{x}=F(x)$2 corresponds to the “true” turbulent solution $\dot{x}=F(x)$3; $\dot{x}=F(x)$4 to the small-scale component $\dot{x}=F(x)$5 of the assimilated solution $\dot{x}=F(x)$6, with large-scale modes of $\dot{x}=F(x)$7 clamped to $\dot{x}=F(x)$8. The DA manifold is $\dot{x}=F(x)$9.

2. Projected Navier–Stokes Dynamics and Variational Equation

The full 3D NS equations in a periodic domain $x$0 ($x$1): $x$2 Projecting onto “large” ($x$3) and “small” ($x$4) scales, define $x$5, $x$6, with $x$7; $x$8, $x$9.

Continuous DA enforces $\dot{y}=G(x,y)$0. The evolution of $\dot{y}=G(x,y)$1 (“fiber”) is

$\dot{y}=G(x,y)$2

The DA manifold is $\dot{y}=G(x,y)$3.

Transverse stability of $\dot{y}=G(x,y)$4 is revealed by the variational equation for $\dot{y}=G(x,y)$5 about $\dot{y}=G(x,y)$6 in background $\dot{y}=G(x,y)$7: $\dot{y}=G(x,y)$8 The TLE at wavenumber cutoff $\dot{y}=G(x,y)$9 is therefore

$M= \{ (x,y)\ |\ y=h(x)\}$0

as computed by integrating the above for $M= \{ (x,y)\ |\ y=h(x)\}$1 with periodic renormalization (Inubushi et al., 2023).

3. Limiting Cases: Connection to Maximal Lyapunov Exponent and Viscous Decay

• $M= \{ (x,y)\ |\ y=h(x)\}$2 (no projection): $M= \{ (x,y)\ |\ y=h(x)\}$3. The variational equation reduces to the full NS tangent equation:

$M= \{ (x,y)\ |\ y=h(x)\}$4

yielding $M= \{ (x,y)\ |\ y=h(x)\}$5, the maximal Lyapunov exponent of the NS attractor.

• $M= \{ (x,y)\ |\ y=h(x)\}$6 (smallest scales only): Inertial (nonlinear) terms negligible for high $M= \{ (x,y)\ |\ y=h(x)\}$7: $M= \{ (x,y)\ |\ y=h(x)\}$8, each Fourier mode decays as $M= \{ (x,y)\ |\ y=h(x)\}$9. In Kolmogorov units $\delta\eta$0:

$\delta\eta$1

The mapping $\delta\eta$2 interpolates between $\delta\eta$3 at $\delta\eta$4 and $\delta\eta$5 at large $\delta\eta$6.

4. Critical Wavenumber, Critical Length, and Reynolds Scaling

Define the critical cutoff $\delta\eta$7 where

$\delta\eta$8

• For $\delta\eta$9, $M$0, and the DA manifold $M$1 is stable: small scales synchronize to large scales.
• For $M$2, $M$3, instability occurs and DA fails.

In a two-term model (Kolmogorov units),

$M$4

so setting to zero gives

$M$5

and the critical physical length

$M$6

As $M$7 grows slowly with Reynolds number (Re), $M$8 decreases slowly with Re. Numerically, for $M$9, $$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$0 so that $$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$1.

5. Implications for Data Assimilation and Algorithmic Design

TLEs provide a model-based, intrinsic criterion for the minimum length scale $$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$2 at which turbulent flows must be observed or assimilated to guarantee small-scale synchronization. Since $$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$3 changes sign only once (“blowout bifurcation”), it is optimal to tune the assimilation cutoff $$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$4 slightly beyond $$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$5, ensuring convergence while minimizing observational or nudging cost.

A full TLE spectrum—or computed covariant Lyapunov vectors (CLVs) restricted to small-scale modes—enables the systematic design of adaptive DA: nudging can target only unstable transverse CLVs. In the presence of noise or model error, the same linearized framework applies. Noise-stabilized DA requires that the largest (mean) transverse Lyapunov exponent be negative, suggesting dissipative feedback should be added selectively to the unstable subspace. This approach links DA theory directly with chaos synchronization and blowout bifurcations.

6. Quantitative Framework and Unification With Empirical DA Results

The TLE formalism quantifies the critical length $$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$6 and encodes the Reynolds number dependence through $$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$7. This perspective unifies previously empirical observations—such as $$\lambda_\perp := \lim_{t\rightarrow\infty} \frac{1}{t}\ln \|\delta\eta(t)\| ,$$8 found in diverse DA schemes—and enables the principled design of next-generation DA algorithms that stabilize only the necessary directions in phase space. TLEs thus establish a bridge between turbulence theory, synchronization theory, and optimal DA of complex spatiotemporal systems (Inubushi et al., 2023).