Stochastic Lorenz 63 System Overview

• Stochastic Lorenz 63 system is a class of SDE models that incorporate random perturbations into the classical chaotic Lorenz 63 equations to study noise effects.
• Noise types—additive, component-wise, and multiplicative—critically influence invariant measures, bifurcations, and Lyapunov stability in the system.
• Advanced methods such as OLIM, Fokker–Planck analysis, and Bayesian inference enable robust parameter recovery, rare event prediction, and closure modeling.

The stochastic Lorenz 63 system refers to a class of stochastic differential equation (SDE) models that generalize the classical, deterministic Lorenz 63 system by introducing random perturbations (typically, additive or multiplicative noise) to one or more components of the system. Such stochastic models are used to study the effect of intrinsic or extrinsic randomness on nonlinear, chaotic dynamics, with key implications for statistical physics, climate modeling, and uncertainty quantification.

1. SDE Formulations of the Stochastic Lorenz 63 System

The canonical deterministic Lorenz 63 ODE is given by

$$\begin{aligned} \dot X &= \sigma (Y-X), \ \dot Y &= X (\rho - Z) - Y, \ \dot Z &= X Y - \beta Z, \end{aligned}$$

where $(X,Y,Z)\in\mathbb{R}^3$ and $(\sigma,\rho,\beta)>0$ are the Prandtl number, Rayleigh number, and geometric parameter, respectively.

Stochastic generalizations of this system include:

• Additive noise (fully-coupled):

$$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$

where $B^i_t$ are independent standard Brownian motions, and $\gamma_i$ scales the noise intensity in each direction (Foldes et al., 2020, Allawala et al., 2016).

• Component-wise additive noise:

In some studies, noise is only added to the $Z$-component:

$$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \epsilon\,dW_t, \end{aligned}$$

with $W_t$ a one-dimensional Wiener process and $\epsilon$ the intensity (Zelati et al., 2020).

• Multiplicative noise:

For state-dependent noise,

$(X,Y,Z)\in\mathbb{R}^3$0

(2206.13154, Geurts et al., 2017).

2. Invariant Measures and Noise-Induced Bifurcation Phenomena

For small noise intensity, the stochastic Lorenz system exhibits unique invariant probability measures with regular density that reflect the deterministic attractor structure. Notable results include:

• Degenerately-damped system: For vanishing vertical damping ($(X,Y,Z)\in\mathbb{R}^3$1), the system possesses a unique invariant measure if and only if noise directly forces the $(X,Y,Z)\in\mathbb{R}^3$2-component. If noise is applied only to $(X,Y,Z)\in\mathbb{R}^3$3 and $(X,Y,Z)\in\mathbb{R}^3$4, or if $(X,Y,Z)\in\mathbb{R}^3$5, the system is transient/null-recurrent and no invariant probability measure exists (Foldes et al., 2020). This shows extreme sensitivity of ergodic behavior to both damping and noise structure.
• Noise-induced bifurcation for $(X,Y,Z)\in\mathbb{R}^3$6:
  • For $(X,Y,Z)\in\mathbb{R}^3$9, the unique invariant law is concentrated on $(\sigma,\rho,\beta)>0$0 with a Gaussian law for $(\sigma,\rho,\beta)>0$1.
  • For $(\sigma,\rho,\beta)>0$2, a genuine bifurcation occurs: exactly two ergodic invariant laws coexist, one supported on $(\sigma,\rho,\beta)>0$3, the other with a smooth density off this plane.
  • The critical threshold is characterized by a Lyapunov exponent $(\sigma,\rho,\beta)>0$4 changing sign, with $(\sigma,\rho,\beta)>0$5 for small $(\sigma,\rho,\beta)>0$6 and $(\sigma,\rho,\beta)>0$7 for large $(\sigma,\rho,\beta)>0$8 (Zelati et al., 2020).
• Full (additive) noise case: For generic ($(\sigma,\rho,\beta)>0$9) nondegenerate additive noise, there always exists a unique invariant probability measure, with exponential mixing under mild conditions (Allawala et al., 2016). The stationary measure washes out fine fractal structure as noise increases.

3. Statistical and Geometrical Properties of Stochastic Attractors

The interplay of deterministic chaos and stochastic forcing creates random attractors whose geometry depends sensitively on noise strength and type.

• Fractal dimension scaling: The scale-dependent instantaneous dimension $$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$0, computed via multivariate empirical mode decomposition (MEMD) and extreme value theory, reveals that:
  • For deterministic Lorenz-63, $$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$1 at large scales.
  • With strong noise, $$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$2 at large scale due to phase space filling.
  • Multiplicative noise yields higher $$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$3 at small scales than additive noise, reflecting the enhancement of local instability by state-dependent fluctuations (2206.13154).
• Probability density computation: The stationary Fokker–Planck equation,

$$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$4

is solvable numerically to obtain the invariant density. Even moderate noise smooths deterministic fractal features (Allawala et al., 2016). Cumulant expansion methods capture low-order statistics efficiently but do not resolve the full distribution.

4. Rare Event Geometry and Quasipotential Analysis

Noise-induced transitions between metastable states or attractor basins are governed by large-deviation principles:

• Quasipotential: For non-gradient SDEs, the quasipotential $$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$5 is defined by a Hamilton–Jacobi PDE or via Freidlin–Wentzell action minimization. For the stochastic Lorenz-63,

$$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$6

(Cameron et al., 2018).

• Numerical computation: The ordered line integral method (OLIM) efficiently computes $$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$7 in high dissipation settings, combining full 3D PDE solutions with dimension-reduced 2D radial meshes when the rotational component dominates. Knowledge of $$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$8 yields asymptotically exponential predictions for exit times from basins:

$$\begin{aligned} dX_t &= \sigma(Y_t - X_t)\,dt + \gamma_1 dB^1_t, \ dY_t &= [X_t (\rho - Z_t) - Y_t]\,dt + \gamma_2 dB^2_t, \ dZ_t &= [X_t Y_t - \beta Z_t]\,dt + \gamma_3 dB^3_t, \end{aligned}$$9

and geometry of most-likely transition paths (Cameron et al., 2018).

5. Linear Stability, Lyapunov Exponents, and Closure Modeling

Stochastic perturbations can fundamentally alter the stability spectrum of the system:

• Lyapunov exponents: For multiplicative "fluctuation-dissipation" noise, the sum of Lyapunov exponents becomes path-dependent and can diverge from the deterministic value; for "SALT" (stochastic advection by Lie transport) noise, the contraction rate remains unchanged from the deterministic system (Geurts et al., 2017). The Cayley–QR algorithm extends the computation to stochastic settings.
• Stochastic closure and generative modeling: Recent approaches embed learned drift and diffusion via neural SDEs to achieve stable, accurate coarse-grained surrogates of Lorenz-63. Parametric and generative methods (e.g., guided flow-matching diffusion models) can stabilize linear instabilities and rectify model error by learning state-dependent diffusion that regularizes dynamics, as evidenced by reduction in Wasserstein and Hellinger distances to fine-scale truth (Williams et al., 13 Apr 2025).

6. Structure Discovery and Statistical Inference

Model identification from data is facilitated by Bayesian equation selection under SDE frameworks. For the Lorenz-63 system:

• The true governing equations can be rediscovered by expanding the drift over a dictionary and applying spike-and-slab priors. Linchpin MCMC yields efficient sampling, enabling parameter recovery and uncertainty quantification even from sparse, noisy observations (Gupta et al., 2021). This approach supports robust inference of both structure (via high posterior inclusion probability for the correct terms) and parameters (with credible intervals containing ground truth).

7. Stochastic and Statistical Stability under Physical and Random Perturbations

Beyond canonical SDE noise, physical or impulsive random perturbations are used to model, e.g., anthropogenic atmospheric forcing. Piecewise-deterministic Markov process (PDMP) models where the vector field is randomly perturbed at section crossings preserve statistical properties under small perturbations. Uniform Lasota–Yorke inequalities and Keller–Liverani theory imply stochastic stability of the invariant measure as perturbation amplitude vanishes (Gianfelice, 2023).

---

Major SDE Model	Formulation	Key Phenomenon/Result
Full additive noise	$B^i_t$0	Unique invariant measure; loss of fractal geometry as $B^i_t$1
Single-component noise	$B^i_t$2	Bifurcation: noise-induced transition from 1 to 2 invariant measures
Multiplicative noise	$B^i_t$3, etc.	Modified Lyapunov spectrum, higher local dimension, structure-dependent stability
PDMP (piecewise-deterministic)	Discrete kicks to vector field	Statistical and stochastic stability of invariant measure

A plausible implication is that the qualitative and quantitative behavior of Lorenz-63 under stochastic forcing depends crucially not only on the overall noise intensity, but also on the structure, direction, and regularity of the stochastic perturbation. This governs both statistical steady-state properties and the dynamical features of rare events, bifurcations, and model closure accuracy.