Local Random Center Manifolds

• Local random center manifolds are local invariant graphs defined over center subspaces, capturing the behavior near nonhyperbolic equilibria in random dynamical systems.
• They are constructed using Lyapunov–Perron fixed-point methods and exponential trichotomy frameworks, leveraging Oseledets splittings for stability analysis.
• Approximation schemes, such as small-noise and rough Taylor-like expansions, enable effective reductions in both finite and infinite-dimensional stochastic and rough PDE settings.

Searching arXiv for recent and foundational papers on local random center manifolds, including stochastic, rough, infinite-dimensional, and deterministic comparison results. arxiv_search(query="local random center manifolds stochastic rough differential equations center manifold random dynamical systems", max_results=10) Local random center manifolds are local invariant manifolds for random or stochastic dynamical systems near stationary trajectories or nonhyperbolic equilibria, typically represented as random graphs over a center subspace or center Oseledets bundle. In the random dynamical systems setting, they are fibered over a metric dynamical system $(\Omega,\mathcal F,\mathbb P,\theta)$, depend measurably on $\omega$, and are local because the graph is defined only on a small, often tempered, random neighborhood of the center direction (Ren et al., 2012). In rough-path and rough-PDE formulations, the same notion appears as a random manifold in the sense of random dynamical systems, with the cocycle generated pathwise by a rough differential equation or a rough partial differential equation (Kuehn et al., 2018, Kuehn et al., 2021).

1. Definitions and basic geometric form

A standard formulation starts from a random dynamical system $\phi(t,\omega,x)$ with a center splitting $E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$. In the stochastic RDE framework of the small-noise paper, a random center manifold $M(\omega)$ is defined by two properties: invariance,

$$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$

and graph representation over the center space,

$M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$

where

$$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$

with $h^c(\cdot,v)$ measurable for each $v$ (Ren et al., 2012). The tangency condition $\omega$0 is the random analogue of classical center-manifold tangency at a deterministic equilibrium.

In rough PDEs the same structure is stated as

$\omega$1

with $\omega$2 Lipschitz, differentiable at $\omega$3, and satisfying

$\omega$4

The result is explicitly local: there exists a tempered-from-below random radius $\omega$5 such that

$\omega$6

(Kuehn et al., 2021). In measurable fields of Banach spaces, the graph may be encoded by a parameterization $\omega$7 rather than by a fixed complement map, but the geometric content is the same: the manifold is modeled on the random center space $\omega$8 and is tangent to it at the stationary trajectory (Varzaneh et al., 2023).

Setting	Manifold form	Random feature
Stochastic/RDS	$\omega$9	measurable dependence on $\phi(t,\omega,x)$0
Rough PDE	$\phi(t,\omega,x)$1	tempered random radius $\phi(t,\omega,x)$2
Banach-field cocycle	$\phi(t,\omega,x)$3	random stationary trajectory and random center space

Locality is not a secondary feature. In essentially all cited frameworks, the original nonlinearities are only locally Lipschitz or are made effectively local by truncation. The manifold is therefore invariant only inside a random neighborhood where the cut-off cocycle agrees with the original cocycle (Kuehn et al., 2021, Varzaneh et al., 2023).

2. Linear spectral structure and random splittings

The linear backbone of the theory is a stable/center/unstable splitting, but the way it is obtained depends on the framework. In finite-dimensional stochastic RDEs, one starts with the random linear system

$\phi(t,\omega,x)$4

assumes $\phi(t,\omega,x)$5, and obtains Lyapunov exponents

$\phi(t,\omega,x)$6

together with Oseledets subspaces $\phi(t,\omega,x)$7, yielding

$\phi(t,\omega,x)$8

where

$\phi(t,\omega,x)$9

The paper calls the resulting estimates an exponential trichotomy (Ren et al., 2012).

For infinite-dimensional random dynamical systems on a separable Hilbert space, the linear theory is built from a multiplicative ergodic theorem for compact cocycles. The outcome is again an Oseledets splitting, and when Lyapunov exponents are separated around $E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$0, one obtains an exponential trichotomy

$E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$1

with tempered random constants $E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$2 and bounds

$E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$3

$E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$4

$E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$5

(Chen et al., 2013). This places the center manifold in the same Lyapunov-theoretic setting as stable and unstable random manifolds.

A notable generalization replaces exponential rates by arbitrary admissible bounds. In the generalized-trichotomy framework, a linear random dynamical system $E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$6 admits a measurable $E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$7-invariant splitting with

$E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$8

and estimates

$E^s(\omega)\oplus E^c(\omega)\oplus E^u(\omega)$9

where $M(\omega)$0 need not be exponential (Bento et al., 2024). This extends center-manifold theory to tempered exponential, integral exponential, summable exponential, quotient-type nonexponential, and polynomial-type growth regimes.

In the most abstract setting, the linearization is taken along a stationary trajectory $M(\omega)$1 of a cocycle on a measurable field of Banach spaces. The derivative cocycle $M(\omega)$2 satisfies a semi-invertible multiplicative ergodic theorem, and the center space is the Oseledets subspace associated with a zero Lyapunov exponent,

$M(\omega)$3

(Varzaneh et al., 2023). This shifts the center-manifold problem from deterministic equilibria to random stationary points and from fixed spaces to measurable fields of spaces.

The rough RDE formulation near a random fixed point fits into the same pattern. There the derivative cocycle

$M(\omega)$4

has an Oseledets splitting

$M(\omega)$5

and the center manifold is tied to the zero Lyapunov directions at the random stationary point (Varzaneh et al., 2023).

3. Construction methods

Across the literature, the dominant construction is a Lyapunov–Perron fixed-point argument, often after localization. In the stochastic RDE paper, the center-manifold graph satisfies a projected Lyapunov–Perron equation, and existence follows from a contraction argument in weighted Banach spaces under the condition

$M(\omega)$6

(Ren et al., 2012). In that setting, the center dynamics has at most weak growth, the stable dynamics decays forward, and the fixed-point space is weighted precisely to balance those rates.

For rough differential equations, the continuous Lyapunov–Perron map is not used directly. Rough-path norms are controlled on finite intervals, so the method is discretized in time. The rough center-manifold paper works in the weighted sequence space

$M(\omega)$7

and studies a discrete Lyapunov–Perron map $M(\omega)$8 built from rough convolutions on unit intervals. The key point is that the localizing cut-off produces a small Lipschitz constant $M(\omega)$9, and a gap condition then yields contraction (Kuehn et al., 2018). The rough PDE theorem uses the same basic idea, but in a scale of interpolation spaces and with controlled rough paths replacing classical mild solutions (Kuehn et al., 2021).

In the rough PDE setting, the mild equation is

$$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$0

and the analysis must compensate for the fact that the semigroup is not Hölder continuous at $$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$1 on the base Banach space. The resolution is to use controlled rough paths on interpolation spaces together with a random truncation $$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$2, producing $$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$3 and $$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$4, and then to solve a discretized Lyapunov–Perron fixed-point problem under the explicit gap condition (6.8) (Kuehn et al., 2021).

The measurable-field theorem organizes the same strategy in a different language. It introduces a weighted bi-infinite orbit space

$$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$5

cuts off the nonlinear remainder by

$$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$6

and constructs the center manifold from the unique fixed point of a Lyapunov–Perron map $$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$7 (Varzaneh et al., 2023). Invariance for the original cocycle is then only local: it holds as long as the orbit stays within the random neighborhood where the modified cocycle and the original cocycle coincide.

Under generalized trichotomies, the Lyapunov–Perron method becomes fully nonuniform. One works with a space $$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$8 of center trajectories and a space $$\phi(t,\omega,M(\omega))\subset M(\theta_t\omega),\qquad \forall t,$$9 of graph maps, defines a coupled operator $M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$0, and proves contraction when

$M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$1

(Bento et al., 2024). This condition plays the role of the classical spectral-gap-plus-small-Lipschitz assumption, but it is expressed through the bounds $M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$2 rather than through fixed exponents.

These constructions make the local nature of the theory explicit. The graph is usually global only for a truncated system. The local random center manifold for the original system is recovered inside the random domain where the truncation is inactive.

4. Approximation and reduced dynamics

Existence theorems are complemented by two distinct approximation programs. The first is a small-noise expansion for stochastic systems transformed into random differential equations. For the Stratonovich system

$M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$3

the random center manifold is expanded as

$M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$4

after transforming the SDE into a random differential equation using stationary Ornstein–Uhlenbeck processes (Ren et al., 2012). In the worked example

$M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$5

the deterministic center manifold is $M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$6, and the noise modifies the graph itself rather than merely perturbing trajectories near a fixed deterministic graph (Ren et al., 2012).

The second program treats rough differential equations directly, without transforming the noise away. The Taylor-like approximation theorem for rough center manifolds assumes that the local random center manifold graph exists and is smooth enough, and then approximates it by

$M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$7

where the coefficients $M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$8 are stationary solutions of auxiliary RDEs driven by the same geometric rough path as the original equation (Blessing et al., 1 Oct 2025). The error estimate is

$M(\omega)=\{(v,h^c(\omega,v)): v\in E^c\},$9

so the graph is approximated to order $$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$0 in the center variable (Blessing et al., 1 Oct 2025).

A conceptual difference separates these two approximation schemes. In the small-noise theory, the expansion parameter is the noise amplitude $$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$1, and the deterministic center manifold appears as the zeroth-order term (Ren et al., 2012). In the rough Taylor-like theory, the expansion is in the center coordinate $$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$2, and the coefficients are themselves random dynamical objects, namely stationary rough solutions (Blessing et al., 1 Oct 2025). This suggests two complementary computational viewpoints: perturbation in stochastic intensity and polynomial approximation in center amplitude.

5. Infinite-dimensional, rough, and Banach-field settings

The theory is no longer confined to finite-dimensional stochastic ODEs. In a separable Hilbert space, the infinite-dimensional random-dynamical-systems paper develops stochastic center manifolds under exponential trichotomy and emphasizes applications to discretisations of nonlinear stochastic partial differential equations with space-time white noise (Chen et al., 2013). The resulting manifolds are random graphs

$$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$3

Lipschitz in the center variable and tangent to the center space at the origin (Chen et al., 2013).

Rough PDEs provide a genuinely pathwise, infinite-dimensional extension. The semilinear rough evolution equation

$$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$4

is studied on a monotone family of interpolation spaces $$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$5, and the center manifold is a random manifold in the sense of random dynamical systems (Kuehn et al., 2021). The theorem covers reaction-diffusion equations driven by nonlinear multiplicative noise and the Swift–Hohenberg equation. The local manifold is

$$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$6

with $$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$7 tempered from below (Kuehn et al., 2021). A decisive technical point is that rough integration is handled pathwise rather than transformed away, and the Lyapunov–Perron method is discretized because rough path norms must be controlled interval by interval.

The rough-differential-equation theory around random stationary points goes further in the stochastic direction. For stochastic semiflows induced by RDEs, the invariant objects are built around a random fixed point $$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$8, and the center manifold consists of points admitting bi-infinite local orbits with subexponential growth relative to the stationary trajectory (Varzaneh et al., 2023). The driving signal may be the rough-path lift of fractional Brownian motion with Hurst parameter $$h^c(\omega,\cdot):E^c\to E^s\oplus E^u,\qquad h^c(\omega,0)=0,\qquad Dh^c(\omega,0)=0,$$9, so the theory extends beyond semimartingale noise (Varzaneh et al., 2023).

The most abstract formulation is the theorem on measurable fields of Banach spaces. It proves a local center manifold theorem for nonlinear cocycles around stationary trajectories on fields $h^c(\cdot,v)$0, with the center manifold parameterized by the center Oseledets space $h^c(\cdot,v)$1 and locally invariant for the original cocycle (Varzaneh et al., 2023). This framework is designed to cover SDEs, SPDEs, rough differential equations, stochastic delay equations, and random PDE cocycles in one theorem (Varzaneh et al., 2023). The local random center manifold is therefore not tied to a single analytic technology; it is a common invariant-graph phenomenon appearing across several cocycle categories.

6. Deterministic prototypes, terminological distinctions, and non-uniqueness

A recurrent source of confusion concerns what “random” means. In the standard random-dynamical-systems sense, randomness comes from a cocycle over a base flow $h^c(\cdot,v)$2, and the manifold is a random invariant graph depending measurably on $h^c(\cdot,v)$3 (Varzaneh et al., 2023). By contrast, the paper on randomly coupled oscillators studies deterministic local center manifolds of ODEs with random but frozen connectivity matrices. For each realization of the random matrix, the ODE is deterministic and has its own deterministic local center manifold; the randomness enters through the statistics of the coefficients across realizations, especially through the eigenvector-overlap quantities

$h^c(\cdot,v)$4

(Moirogiannis et al., 2019). This is a different use of “random.”

Deterministic center-manifold results remain structurally relevant. For partially hyperbolic invariant compact sets of diffeomorphisms, a compact set $h^c(\cdot,v)$5 lies in a local center submanifold if and only if each strong stable and strong unstable leaf intersects $h^c(\cdot,v)$6 at exactly one point (Bonatti et al., 2014). That theorem is not random, but it isolates a geometric obstruction—strong stable or unstable connections—that has no direct analogue in fixed-point center-manifold theorems (Bonatti et al., 2014). Likewise, for differential equations with state-dependent delay, a local center manifold can be constructed as the local intersection of a local center-stable manifold and a local center-unstable manifold by an application of the Implicit Mapping Theorem (Stumpf, 2015). This suggests a general geometric principle: center manifolds may arise as intersections of partially hyperbolic invariant manifolds, even though the random literature usually constructs them by Lyapunov–Perron methods.

Localization itself has delicate analytic content. The corrigendum to “Center Manifolds without a Phase Space” shows that a pointwise cutoff can fail in weighted $h^c(\cdot,v)$7-spaces, and replaces it with a norm-based, translation-equivariant cutoff

$h^c(\cdot,v)$8

so that the localized nonlinearity maps exponentially weighted $h^c(\cdot,v)$9-spaces into themselves and has a small Lipschitz constant (Faye et al., 2020). That result is deterministic, but it is directly relevant methodologically: local random center manifolds also rely on cutoffs compatible with the topology of the working space.

Non-uniqueness is another point where deterministic and random theories diverge in emphasis. The deterministic partially hyperbolic theorem states explicitly that the local center submanifold is generally not unique (Bonatti et al., 2014). In several random Lyapunov–Perron theorems, by contrast, one obtains uniqueness inside a prescribed class of Lipschitz graph maps or weighted-orbit classes (Bento et al., 2024, Varzaneh et al., 2023). This does not remove the usual locality restrictions, but it clarifies the sense in which a local random center manifold is canonical once the cocycle class, growth rate $v$0, and truncation regime are fixed.