On the structure of complex spectra and eigenfunctions of transfer and Koopman operators

Matheus M. Castro and Gary Froyland · 2025

Read On the structure of complex spectra and eigenfunctions of transfer and Koopman operators by Matheus M. Castro and Gary Froyland (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-05770.txt.

Introduction

Matheus Castro and Gary Froyland study in a 2025 paper how transfer and Koopman operators show turning motion. These math tools track how a system moves, and their spectra can show cycles. In many real cases the speed of turning shifts with where the system sits, which makes the spectra hard to read.

Working with a clean model plus a small amount of noise, the pair pin down the spectra in exact terms. They also track what happens as the noise fades toward zero. Their results yield simple ways to spot cycles and locate them in space. Tests with numbers back the claims and show the cycle facts hold up well across noise levels in the range studied.

Catalog record

CREATORMatheus M. Castro and Gary Froyland
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE126 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-05770

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

On the structure of complex spectra and eigenfunctions of transfer and Koopman operators Matheus M. Castro and Gary Froyland Abstract . Complex eigenspectra of transfer and Koopman operators describe rotational motion in dynamical systems. A particularly relevant situation in applications is when the rotation speed depends on the state-space position of the dynamics. We consider a canonical model of such dynamics in the presence of small noise, and provide precise characterisations of the eigenspectrum and eigenfunctions of the corresponding transfer operators. Further, we study the limiting behaviour of the eigenspectrum and eigenfunctions in the zero-noise limit, including their quadratic and linear response. Our results clarify the structure of transfer and Koopman operator eigenspectra, and provide new interpretations relevant to applications. Our theorems on support localisation of the eigenfunctions yield simple algorithms to detect the existence and state-space location of approximately cyclic motion with distinct periods. Our numerical results verify that information on the cycle periods and their locations determined by the operator eigendata is insensitive to noise level in the linear response regime. We believe that the dynamic mechanisms underlying the eigendata and their properties apply rather broadly and enhance our understanding of approximate cycle detection in dynamical systems with operator methods. 1. Introduction Over the last three decades, operator-theoretic techniques have proven to be highly successful methods of analysis for complex dynamical systems. The basis of the analysis is usually the dominant eigenspectrum of a transfer operator or Koopman operator and the corresponding eigenfunctions. In the autonomous setting, where the underlying dynamics does not change in time, large real spectral values correspond to almost-invariant sets (large regions of state space that have no collective state-space motion under the dynamics) [ 7 , 6 ] . This concept has been applied extensively in several fields, including molecular dynamics [ 27 ] to find stable molecular conformations, space mission design [ 5 ] to identify orbital trapping regions, and physical oceanography [ 13 , 15 , 25 ] to map ocean gyres and garbage patches. Regarding complex eigenvalues, the Halmos–von Neumann Theorem (see e.g. [ 8 , Theorem 17.11] ) states that each ergodic measure-preserving system with discrete spectrum is isomorphic to an ergodic rotation system, for example, an aperiodic rotation on a n -torus. This concept has been used extensively in the context of Koopman operators, e.g. [ 24 , 26 ] . The current paper takes the underlying Halmos–von Neumann idea and extends it in two ways. Firstly, we allow the rotation to depend on the location in the state space , rather than being fixed along a particular coordinate throughout the state space. Secondly, we make a detailed study of the effect of noise on the complex eigenvalues and eigenfunctions. The motivation for this study is that complex dynamical systems often possess many co-existing cycles or approximate cycles. For example, in the Earth system, climate drivers such as the El-Niño Southern Oscillation (ENSO) and the Madden–Julian Oscillation (MJO) are approximate co-existing cycles that influence the weather and climate. On even longer timescales, changing glacial cycles describe how northern-hemisphere continental ice sheets repeatedly accumulate and melt over millions of years. Recent work [ 10 , 11 ] has demonstrated that transfer operator methods are a flexible and reliable way of extracting cycles from models, observations, and even time series of spatial fields; for example extracting the ENSO cycle from a time series of observed sea-surface temperature fields. Long-lived persistent cycles typically give rise to complex eigenvalues in the dominant spectrum of the transfer operator. The magnitude of the eigenvalue quantifies the rate of temporal decay of the cycle, while the argument of the eigenvalue directly provides an estimate of the cycle’s period. Moreover, the corresponding eigenfunctions may be used to represent the cycle in terms of the original spatiotemporal time series [ 10 ] . These ideas may be illustrated with the Lorenz equations, where an estimate of the transfer operator is constructed and the complex eigenvalue with largest magnitude is computed. The left panel of Figure 1 displays the magnitude of the complex entries of the corresponding eigenvector on a blue colourscale, and the right panel shows the argument of the complex entries on a periodic colourscale. Figure 1. Left: Magnitude of the leading complex eigenvector of the normalised transfer operator for the Lorenz flow. A peak in magnitude occurs in the vicinity of the lowest-period periodic orbit of the Lorenz equations with a period of 1.5586 time units (shown in red, see Fig. 10a [ 14 ] and [ 9 ] ). Right: The argument of the complex eigenvector. In the periodic colormap, one sees a full rotation around each of the two wings as the argument proceeds from -\pi to \pi ; see [ 10 ] . In this example, the argument of the eigenvalue associated with this complex eigenvector corresponds to a cycle of period 0.770 . This period is roughly half that of the lowest-period periodic orbit of the Lorenz equations, which loops once around each wing (displayed in red in Figure 1 (left)). We make this comparison to emphasise that we are not aiming to detect single periodic orbits, but instead to capture large regions in the state space that revolve with an approximately constant period. The eigenvector detects that the dominant cycle in the Lorenz system consists of wide bands on each wing that rotate around each wing with a period of roughly 0.770 . The symmetry of the Lorenz equations under (x,y,z)\mapsto(-x,-y,z) , is reflected in the representation of the cycle in Figure 1 (right): the cycle phase given by the argument respects this symmetry, see [ 10 ] for details. Our focus in this work is to provide a formal explanation for the behaviour of the spectrum and eigenfunctions of the transfer operator in the presence of coexisting cycles with different periods. Roughly speaking, eigenfunctions corresponding to cycles with distinct periods should have supports that are largely disjoint , because large overlaps would need to simultaneously rotate at two distinct speeds. Our formal analysis concentrates on a mathematical model with multiple coexisting cycles, acting on a discrete collection of circles, between which trajectories may randomly jump under the influence of noise. Our state space M is the union of N circles: M:=\{1,\ldots,N\}\times\mathbb{S}^{1} , which we view as a cylinder that has been discretised along its central axis. The dynamics on the j^{\rm th} circular band of the discretised cylinder is a rotation with speed \alpha_{j} ; see Figure 2 . Figure 2. Illustration of noisy rotational dynamics on circular fibres. Each horizontal black line represents a copy of S^{1} . In one iteration of the dynamics, a point x on the j^{\rm th} fibre is (i) rotated by \alpha_{j} to x+\alpha_{j} , (ii) a random rotation drawn uniformly from the interval (-\delta,\delta) is added (shown as a red interval), and (iii) finally one moves from fibre j to fibre j^{\prime} according to the Markov chain with conditional probability W_{\varepsilon,j,j^{\prime}} . In this figure, fibres 1, 2, and 3 all rotate with the same speed, \alpha_{1}=\alpha_{2}=\alpha_{3} . Similarly, the fibres j and j+1 rotate at the same speed, as do the final group of fibres numbered \ldots,,N-1,N . We are particularly

Read the complete text