Non-Autonomous Inertial Manifolds

• Non-autonomous inertial manifolds are time-dependent, finite-dimensional invariant graph families that capture the dynamics of evolution processes in differential equations.
• They are constructed via Lyapunov–Perron and frequency-domain methods under sharp spectral gap conditions, ensuring exponential tracking of system trajectories.
• This framework applies to stochastic, parabolic, and Banach-space systems, transferring properties like periodicity and almost periodicity from the forcing functions.

Searching arXiv for recent and foundational papers on non-autonomous inertial manifolds to ground the article. Non-autonomous inertial manifolds are time-dependent invariant families of finite-dimensional graphs associated with evolution processes or cocycles generated by non-autonomous differential equations. In contrast to an autonomous inertial manifold, which is a single invariant set for a semigroup, the non-autonomous object is parameterized by time and, in random settings, by the sample point, typically as $M(\tau)$ or $M(\tau,\omega)$. Its characteristic features are a finite-dimensional graph representation, invariance under the process or cocycle, and exponential tracking or attraction of trajectories. In recent work this framework has been developed for semilinear stochastic evolution equations, abstract Banach-space evolution processes, asymptotically compact cocycles, and non-autonomous parabolic equations treated through frequency-domain and quadratic-Lyapunov methods (Wang, 2014, Czaja et al., 31 Jul 2025).

1. Evolution-process and cocycle formulation

A standard abstract setting is the non-autonomous semilinear problem

$$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$

posed on a Banach space $X$, where the linear part generates an evolution process $L(t,\tau)$. In the recent sharp-gap formulation, the linear process is assumed to admit an exponential splitting with projections $Q(t)$, exponents $\gamma>\rho$, and the associated graph family is required to satisfy invariance, forward exponential attraction, pullback exponential attraction, and a Lipschitz graph property over the bundle $Q(\tau)X$ (Czaja et al., 31 Jul 2025).

In the stochastic Hilbert-space setting, the basic model is the semilinear equation

$$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$

where $A$ is symmetric, positive, and has compact inverse, $M(\tau,\omega)$0 is nonlinear, $M(\tau,\omega)$1 is a deterministic time-dependent forcing, and $M(\tau,\omega)$2 is a two-sided $M(\tau,\omega)$3-valued Wiener process. The novelty is that the inertial manifold must depend on the initial time $M(\tau,\omega)$4 as well as $M(\tau,\omega)$5, so one studies a family $M(\tau,\omega)$6. A central structural point is that the deterministic forcing $M(\tau,\omega)$7 is what makes the manifold genuinely non-autonomous, while the stochastic forcing is encoded by a random transformation through a stationary Ornstein–Uhlenbeck process (Wang, 2014).

The cocycle viewpoint is equally central in non-autonomous parabolic theory. For equations of the form

$M(\tau,\omega)$8

the dynamics is organized as a cocycle $M(\tau,\omega)$9 over a shift flow on the time base. This formulation is used precisely because time dependence in $$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$0 prevents reduction to a purely autonomous semigroup picture (Anikushin, 2020).

2. Lyapunov–Perron construction and graph representation

The classical construction of a non-autonomous inertial manifold is a backward-time fixed-point argument in a weighted trajectory space. For the stochastic equation above, after subtracting the stationary Ornstein–Uhlenbeck process $$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$1, one obtains the random PDE

$$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$2

The Lyapunov–Perron method is carried out in

$$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$3

and for fixed $$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$4 and $$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$5 the fixed-point operator is

$$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$6

The unique fixed point $$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$7 yields a graph map

$$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$8

and hence

$$u_t=A(t)u+f(t,u), \qquad t>\tau,\quad u(\tau)=\eta\in X,$$9

For the original stochastic equation, the manifold is shifted by the Ornstein–Uhlenbeck process: $X$0 The map $X$1 is measurable in $X$2 and Lipschitz in $X$3 (Wang, 2014).

An analogous Banach-space construction uses the backward weighted space

$X$4

together with the Lyapunov–Perron operator

$X$5

Its fixed point $X$6 determines the graph map

$X$7

so the manifold is the graph over $X$8 with values in $X$9 (Czaja et al., 31 Jul 2025).

3. Analytic hypotheses and sufficient conditions

In the Hilbert-space random setting, the construction is based on a spectral-gap scheme. The nonlinearity satisfies

$L(t,\tau)$0

for some $L(t,\tau)$1, the deterministic forcing obeys

$L(t,\tau)$2

and one imposes the gap condition

$L(t,\tau)$3

for some $L(t,\tau)$4, followed by

$L(t,\tau)$5

The separation $L(t,\tau)$6 is the key ingredient in the fixed-point argument (Wang, 2014).

For general non-autonomous Banach-space processes, the recent sharp formulation replaces earlier non-sharp hypotheses by

$L(t,\tau)$7

The paper emphasizes that this condition is sharp. In the case $L(t,\tau)$8, it becomes

$L(t,\tau)$9

and in the symmetric case $Q(t)$0,

$Q(t)$1

which is identified with the classical optimal threshold appearing in the sharp theory of Miklavčič and Romanov and also in Latushkin–Layton (Czaja et al., 31 Jul 2025).

A different route is provided by the Frequency Theorem. For non-autonomous parabolic equations, the decisive hypothesis is a frequency inequality equivalent to the existence of a quadratic Lyapunov functional $Q(t)$2. In this framework, the classical spectral-gap condition appears as a special case of the resolvent estimate

$Q(t)$3

and the resulting quadratic inequality yields the semi-dichotomy and cone properties needed for inertial-manifold construction in the cocycle setting (Anikushin, 2020).

A further extension replaces the classical stationary frequency condition by the Spatial Averaging Principle of Mallet-Paret and Sell in a nonstationary Hamiltonian system. In that setting, the central object is a stable Lagrangian bundle, and the spatial averaging assumption acts as a nonautonomous replacement for the spectral gap (Anikushin, 14 Mar 2025).

4. Invariance, exponential tracking, and finite-dimensional reduction

The defining dynamical property is invariance of the graph family under the evolution. For the stochastic cocycle $Q(t)$4 generated by the non-autonomous stochastic equation,

$Q(t)$5

More strongly, the manifold is asymptotically complete: for every initial value $Q(t)$6, there exists $Q(t)$7 such that

$Q(t)$8

The same paper proves that any tempered pullback random attractor $Q(t)$9, if it exists, is contained in the inertial manifold: $\gamma>\rho$0 This yields a finite-dimensional invariant set containing the long-term pullback dynamics (Wang, 2014).

In the cocycle theory for asymptotically compact systems, the analogous object is the principal leaf $\gamma>\rho$1, defined by amenable complete trajectories. Under the cone-squeezing and asymptotic compactness assumptions, $\gamma>\rho$2 is a Lipschitz admissible set, the cocycle maps it homeomorphically to $\gamma>\rho$3, and under the differentiability hypothesis (DIFF) it is a $\gamma>\rho$4-submanifold. The same framework reconstructs vertical leaves, and for every $\gamma>\rho$5 there exists a unique $\gamma>\rho$6 such that

$\gamma>\rho$7

In this sense, inertial manifolds are part of a broader foliation theory for non-autonomous cocycles (Anikushin, 2020).

A related unified theorem formulates inertial manifolds, saddle point property, and exponential dichotomy within a single invariant-graph construction for evolution processes. The same mechanism that produces the inertial manifold also produces stable and unstable graph representations and roughness of exponential dichotomy under perturbations (Carvalho et al., 2021).

5. Temporal structure: periodic and almost periodic families

A distinctive non-autonomous phenomenon is that the manifold may inherit temporal recurrence properties from the forcing. For the stochastic evolution equation with deterministic forcing $\gamma>\rho$8, if $\gamma>\rho$9 is almost periodic as a map $Q(\tau)X$0, then the inertial manifold is pathwise almost periodic: for every $Q(\tau)X$1, there exists a length $Q(\tau)X$2 such that any interval of length $Q(\tau)X$3 contains a shift $Q(\tau)X$4 satisfying

$Q(\tau)X$5

If $Q(\tau)X$6 is $Q(\tau)X$7-periodic, then the manifold is pathwise $Q(\tau)X$8-periodic: $Q(\tau)X$9 The paper explicitly interprets this as transfer of the temporal behavior of the deterministic forcing to the manifold itself (Wang, 2014).

This inheritance principle is consistent with the general cocycle viewpoint. In the frequency-theorem treatment of parabolic equations, time-dependent nonlinearities $$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$0 and continuous forcing $$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$1 are handled directly in the non-autonomous framework, and the existence theory is formulated for invariant families of finite-dimensional Lipschitz submanifolds rather than for a single autonomous manifold (Anikushin, 2020).

6. Regularity, computation, and limitations

Regularity beyond Lipschitz continuity has become a central issue. In the sharp-gap Banach-space theory, if $$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$2 is Fréchet differentiable in $$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$3 and $$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$4 is continuous, then the solution map $$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$5 is differentiable in the backward weighted space and the graph map has derivative

$$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$6

The conclusion is that the non-autonomous inertial manifold is $$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$7, not merely Lipschitz (Czaja et al., 31 Jul 2025).

The same theme appears in the quadratic-Lyapunov and cocycle-based approaches. The frequency-theorem program states that one can establish not only Lipschitz regularity and exponential tracking but also $$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$8-differentiability and normal hyperbolicity for the invariant family. In the asymptotically compact cocycle framework, the principal leaf is $$\frac{du}{dt}+Au=F(u)+g(t)+\frac{dW}{dt}, \qquad u(\tau)=u_\tau,$$9 under the differentiability assumption (DIFF) (Anikushin, 2020, Anikushin, 2020).

There is also a computational strand. A numerical non-autonomous inertial manifold reduction method treats a dissipative non-autonomous differential equation $A$0 by linearizing along a trajectory, performing a time-dependent orthogonal change of variables $A$1, splitting the transformed system into slow and fast coordinates, and computing the nonlinear decoupling transformation through a boundary value problem rather than a Lyapunov–Perron fixed-point iteration. The implementation uses smooth, time-dependent Householder reflectors and standard BVP solvers such as MATLAB’s bvp4c; the finite-$A$2 truncation error is estimated by

$A$3

with

$A$4

(Chung et al., 2015).

The literature also makes clear that non-autonomous inertial-manifold theory should not be conflated with every nearby invariant-manifold construction. The normal-form theory for non-autonomous PDEs in graded Fréchet spaces constructs approximate conjugacies and invariant center/stable/unstable subsets up to a residual $A$5, but it explicitly states that this is not a classical global inertial manifold theorem with exact exponential tracking (Hochs et al., 2019). More broadly, adjacent inertial-manifold results show that existence is not automatic: periodic vector reaction–diffusion–advection systems may have no finite-dimensional inertial manifold containing the attractor (Kostianko et al., 2017), and Kwak-transformed systems with Jordan cells require sharp non-self-adjoint spectral-gap conditions rather than the usual self-adjoint criterion (Kostianko et al., 2019). This suggests that, in the non-autonomous theory as well, dissipation, periodicity, or a variable transformation by themselves are not sufficient; the decisive issue is the presence of a verifiable mechanism such as a sharp gap condition, a frequency inequality, a cone-squeezing estimate, or spatial averaging.