Linearized Frobenius–Perron Operators

• Linearized Frobenius–Perron operators are operator-theoretic tools that linearize nonlinear density evolution via approximations, perturbation derivatives, and spectral techniques.
• They are applied to both deterministic and stochastic systems to reveal invariant densities, stability properties, and asymptotic behavior through spectral insights.
• Finite-dimensional implementations, such as matrix mappings and Ulam’s method, enable practical numerical spectral computations and analysis of transfer operators.

Searching arXiv for relevant papers on linearized Frobenius–Perron operators, perturbations, generators, and spectral computation. arXiv search query: "linearized Frobenius-Perron operator perturbation generator spectral pollution transfer operator" Linearized Frobenius–Perron operators are operator-theoretic representations that approximate, differentiate, or otherwise localize the action of a Frobenius–Perron (transfer) operator on densities. In the available literature, this notion appears in several distinct but related senses: as an infinitesimal iteration $f(a,\delta)=a+\delta F(a)$ associated with an ODE or PDE; as the infinitesimal generator of a transfer semigroup; as a Fréchet derivative of the operator with respect to perturbations of the underlying drift; as a linearization about an equilibrium density yielding discrete Vlasov-type equations; and as a finite-dimensional approximation used for spectral computation and data-driven analysis (Cirier, 2011, Koltai, 2011, Koltai et al., 2018, Tzenov, 7 Aug 2025, Herwig et al., 22 Jul 2025).

1. Classical setting and operator-theoretic framework

For a measure-preserving transformation $\varphi:\Omega\to\Omega$ on a probability space $(\Omega,\Sigma,\mu)$, the Perron–Frobenius operator

$P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$

is defined by

$$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$

Its adjoint on $L^\infty$ is the Koopman operator $Tf=f\circ\varphi$ (Gerlach, 2016). In continuous-time deterministic control-affine dynamics $\dot x=f(x)+g(x)u$, the Perron–Frobenius semigroup propagates densities by

$$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$

with infinitesimal generator

$\mathcal{P}_f\psi=-\nabla\cdot(f(x)\psi(x)).$

This linear action on densities is the basis for convex formulations of nonlinear optimal control and for finite-dimensional operator approximations derived from data (Huang et al., 2020).

The term “linearized” is therefore not restricted to a single construction. In some works it refers to replacing the nonlinear state dynamics by a first-order iteration and then studying the induced transfer operator; in others it denotes an operator-valued derivative or a discrete generator; in still others it denotes a matrix representation acting on harmonic, polynomial, or indicator-function bases. A plausible implication is that the subject is best understood as a family of operator-localization procedures rather than as a single canonical object.

2. Infinitesimal iterations and generators

A foundational route to linearization begins with the infinitesimal iteration associated with an ODE or PDE,

$\varphi:\Omega\to\Omega$0

where $\varphi:\Omega\to\Omega$1 and $\varphi:\Omega\to\Omega$2 is small. This iteration is used to associate the deterministic flow $\varphi:\Omega\to\Omega$3 with a Perron–Frobenius operator and thereby to study evolution of densities under the dynamics (Cirier, 2011). In the partially linear case, $\varphi:\Omega\to\Omega$4 is assumed to have a block structure, and near a zero $\varphi:\Omega\to\Omega$5 of $\varphi:\Omega\to\Omega$6 the iteration can be approximated by a map of the form

$\varphi:\Omega\to\Omega$7

Within that framework, nondegenerate Hessians lead only to fixed points or regular cycles as invariant measures, whereas degeneracy of the Hessian can produce nontrivial asymptotic densities supported on random ovals or related sets (Cirier, 2011).

A second route is semigroup theory. For outflow systems, where trajectories may leave a compact state space $\varphi:\Omega\to\Omega$8 and are then terminated, the outflow transfer semigroup $\varphi:\Omega\to\Omega$9 acts on $(\Omega,\Sigma,\mu)$0. A discrete infinitesimal generator is introduced by

$(\Omega,\Sigma,\mu)$1

where $(\Omega,\Sigma,\mu)$2 projects onto piecewise-constant functions over a uniform box covering (Koltai, 2011). The resulting semigroup converges pointwise in $(\Omega,\Sigma,\mu)$3 to the true Frobenius–Perron semigroup for every fixed time. This provides a rigorous justification for matrix-based approximations of transfer dynamics in open systems (Koltai, 2011).

A closely related formulation uses the generator exponential. For a vector field $(\Omega,\Sigma,\mu)$4, the Perron–Frobenius generator is

$(\Omega,\Sigma,\mu)$5

If $(\Omega,\Sigma,\mu)$6, then $(\Omega,\Sigma,\mu)$7, but in general

$(\Omega,\Sigma,\mu)$8

When the generators commute, the exponentials factor exactly; otherwise, the leading noncommutative correction is controlled by Lie brackets, and for divergence-free $(\Omega,\Sigma,\mu)$9 and $P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$0 one has the cubic approximation

$P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$1

with error determined by nested Lie brackets (Tallapragada, 11 Nov 2025).

3. Linearization about invariant states and equilibrium densities

Linearized Frobenius–Perron operators also arise by expanding about invariant or equilibrium densities. In the partially linear density approach, the asymptotic density is obtained as $P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$2 after linearization near each fixed point. For the Lorenz system, the analysis yields two families of random ovals around the nontrivial fixed points, with a communication tunnel given by a plane relation; for Navier–Stokes equations written in the form $P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$3, a similar mechanism leads to families of beta-distributed ellipsoids and more complex invariant sets in higher dimension (Cirier, 2011). This suggests that “linearized” need not mean collapse to a purely local deterministic Jacobian picture; in these constructions, the linearized transfer analysis can instead reveal random invariant sets and asymptotic density profiles.

In beam–beam interaction for circular colliders, the Frobenius–Perron operator for each beam propagates the phase-space distribution from turn to turn under a symplectic twist map. Writing

$P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$4

with $P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$5 an equilibrium distribution and $P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$6 a perturbation, linearization gives

$P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$7

The paper identifies these linearized Frobenius–Perron operators as a discrete form of the linearized Vlasov equations, leading to a matrix mapping technique for coherent beam–beam instabilities (Tzenov, 7 Aug 2025). Under an isolated coherent beam–beam resonance, stability is characterized in closed form by the condition

$P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$8

Away from structural resonance, renormalization-group reduction shows that the map propagator describes a random walk of the angle variable, implying the existence of an equilibrium distribution depending only on the action variable (Tzenov, 7 Aug 2025).

A more general ergodic perspective identifies asymptotic operator convergence with convergence in the measure algebra. For a bimeasurable measure-preserving transformation, strong convergence of $P:L^1(\Omega,\Sigma,\mu)\to L^1(\Omega,\Sigma,\mu)$9 in $$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$0 for all $$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$1 is equivalent to convergence of $$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$2 in the measure algebra for every measurable set $$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$3, and the limit is $$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$4, the conditional expectation onto the invariant $$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$5-algebra (Gerlach, 2016). In exact systems this collapses to convergence onto constants, and the corresponding uniform mixing-like property is stronger than classical mixing (Gerlach, 2016).

4. Perturbation theory and operator derivatives

A distinct sense of linearization is differential sensitivity of the transfer operator itself. For time-inhomogeneous ordinary stochastic differential equations, the Perron–Frobenius and Koopman operators are Fréchet differentiable with respect to the drift. If $$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$6 denotes the solution of the perturbed SDE and $$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$7 the unperturbed one, then for time-$$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$8 observables the Koopman derivative at $$\int_A Pf\,d\mu=\int_{\varphi^{-1}(A)}f\,d\mu \qquad \forall f\in L^1,\ A\in\Sigma.$$9 is

$L^\infty$0

The corresponding Perron–Frobenius operator derivative is given via the derivative of the transition kernel, and isolated eigenvalues, eigenfunctions, singular values, and singular vectors depend continuously and Fréchet-differentiably on the drift (Koltai et al., 2018).

An important limitation is explicit in this literature: such differentiability does not generally hold for deterministic systems in the $L^\infty$1 operator norm topology, and the smoothing action of elliptic noise is crucial (Koltai et al., 2018). This is one of the principal conceptual boundaries of linearized transfer-operator theory. In stochastic dynamics, linearization can be formulated as operator differentiation. In deterministic dynamics, the same level of norm-topology regularity may fail, so one often resorts instead to generator approximations, spectral regularization, or discretized semigroups.

A related perturbative construction appears in the generator-exponential framework for primitive flows and parameter perturbations. If $L^\infty$2 and

$L^\infty$3

then the perturbed Perron–Frobenius operator can be approximated from the separate operators associated with $L^\infty$4 and $L^\infty$5, avoiding full recomputation from new trajectory data (Tallapragada, 11 Nov 2025). In that sense, linearization is operational: it is a device for rapidly updating transfer operators under small model changes.

5. Spectral theory, generalized spectra, and spectral pollution

Much of the modern literature treats linearized Frobenius–Perron operators as spectral objects. For a bounded linear operator $L^\infty$6 on a Banach space,

$L^\infty$7

and the $L^\infty$8-pseudospectrum is defined by

$L^\infty$9

This viewpoint is central to residual-based methods for transfer operators, because spurious eigenvalues introduced by finite-dimensional approximations are often better interpreted through pseudospectral stability than through raw eigenspectra alone (Herwig et al., 22 Jul 2025).

A residual-based framework extending Residual DMD to Frobenius–Perron operators introduces a kernel residual

$Tf=f\circ\varphi$0

computed as the minimal eigenvalue of a well-posed $Tf=f\circ\varphi$1 matrix (Herwig et al., 22 Jul 2025). Under mild density assumptions on the kernel RKHS in $Tf=f\circ\varphi$2, the limiting residual supplies a necessary condition for true spectrality. The same framework is extended to Hardy–Hilbert spaces $Tf=f\circ\varphi$3, Sobolev spaces $Tf=f\circ\varphi$4, and their duals, making explicit that spectral features may appear in computations even when the corresponding eigenfunctions lie outside the chosen $Tf=f\circ\varphi$5 space (Herwig et al., 22 Jul 2025). A common misconception is therefore that any numerically persistent eigenvalue is automatically “true” in the ambient space being used. The Blaschke-map examples show that the validity of an eigenvalue may depend decisively on the function space.

Generalized spectral theory provides another response to this difficulty. For symbolic dynamical systems, a rigged Hilbert space $Tf=f\circ\varphi$6 enables analytic continuation of the resolvent through continuous spectrum. In the one-sided full $Tf=f\circ\varphi$7-shift, the generalized eigenvalues are $Tf=f\circ\varphi$8, and the iteration admits an asymptotic expansion whose convergence rate is determined by the generalized spectrum (Chiba et al., 2021). Here “linearization” takes the form of replacing continuous-spectrum dynamics by a resonance expansion in a Gelfand triplet.

Beta-transformations furnish explicit examples of isolated non-leading eigenvalues. On $Tf=f\circ\varphi$9, the Perron–Frobenius operator for $\dot x=f(x)+g(x)u$0 has essential spectral radius $\dot x=f(x)+g(x)u$1, and the set of $\dot x=f(x)+g(x)u$2 for which there is at least one non-leading eigenvalue is open and dense in $\dot x=f(x)+g(x)u$3 (Suzuki, 2024). Each non-leading eigenvalue is Hölder continuous as a function of $\dot x=f(x)+g(x)u$4, yet continuous but non-differentiable, and eigenfunctionals corresponding to non-leading eigenvalues cannot be expressed by any complex Borel measure on the interval (Suzuki, 2024). For quadratic Parry-type greedy $\dot x=f(x)+g(x)u$5-expansions, the transfer operator $\dot x=f(x)+g(x)u$6 admits explicit eigenfunctions $\dot x=f(x)+g(x)u$7 and $\dot x=f(x)+g(x)u$8 with $\dot x=f(x)+g(x)u$9 and $$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$0, leading to the asymptotic formula

$$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$1

for sufficiently smooth normalized $$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$2 (Cornean et al., 24 Feb 2025).

6. Numerical realizations and applied uses

Finite-dimensional realizations of Frobenius–Perron operators are central to numerical practice. In one-dimensional iterated function systems, a family of positive linear operators

$$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$3

is studied on $$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$4, where $$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$5 has a strictly positive $$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$6 eigenfunction $$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$7 with eigenvalue $$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$8, and the Hausdorff dimension of the invariant set is the unique $$[\mathbb{P}_t\psi](x)=\psi(\phi_{-t}(x)) \left|\frac{\partial \phi_{-t}(x)}{\partial x}\right|,$$9 satisfying $\mathcal{P}_f\psi=-\nabla\cdot(f(x)\psi(x)).$0 (Falk et al., 2016). A collocation method with continuous piecewise linear functions constructs matrices $\mathcal{P}_f\psi=-\nabla\cdot(f(x)\psi(x)).$1 and $\mathcal{P}_f\psi=-\nabla\cdot(f(x)\psi(x)).$2 such that

$\mathcal{P}_f\psi=-\nabla\cdot(f(x)\psi(x)).$3

yielding rigorous upper and lower bounds converging to $\mathcal{P}_f\psi=-\nabla\cdot(f(x)\psi(x)).$4 as the mesh size tends to zero (Falk et al., 2016).

In data-driven control, the Perron–Frobenius generator appears as the linear PDE constraint in an infinite-dimensional convex optimization problem over densities, while finite-dimensional approximations are obtained through the dual Koopman operator and the Naturally Structured DMD method, which enforces positivity and Markov properties (Huang et al., 2020). In Bayesian filtering for nonlinear stochastic systems, recursion is written as

$\mathcal{P}_f\psi=-\nabla\cdot(f(x)\psi(x)).$5

and Ulam’s method approximates the transfer operator by a row-stochastic matrix

$\mathcal{P}_f\psi=-\nabla\cdot(f(x)\psi(x)).$6

The posterior is then represented by a convex combination of indicator functions, and a low-rank filter is obtained by spectral decomposition of the discretized operator (Liu et al., 2023).

Set-oriented transport analysis uses the Perron–Frobenius operator to define a covariance-based finite-time Lyapunov exponent,

$\mathcal{P}_f\psi=-\nabla\cdot(f(x)\psi(x)).$7

thereby connecting density evolution, principal stretching, and phase-space transport (Tallapragada, 2011). Recent residual-based transfer-operator methods extend these numerical uses to high-dimensional and kernelized settings, with case studies including Blaschke maps and alanine dipeptide, where the leading real eigenvalues of the transfer operator were used to approximate slow dynamics and metastable states (Herwig et al., 22 Jul 2025).

Across these settings, linearized Frobenius–Perron operators serve three persistent functions: they convert nonlinear dynamics into linear density evolution; they expose asymptotic structure through spectra, generators, or perturbative derivatives; and they support finite-dimensional algorithms whose validity depends sharply on the function space, regularity class, and spectral interpretation adopted.