WorksPhysics and Astronomy
Computation and stability analysis of periodic orbits using finite differences, Fourier or Chebyshev spectral expansions in time
Abstract
We analyse and compare several algorithms to compute numerically periodic solutions of high-dimensional dynamical systems and investigate their Floquet stability without building the monodromy matrix. The solution and its perturbation are discretised in time either using finite differences, Fourier-Galerkin or Chebyshev expansions. The resulting nonlinear set of equations describing the periodic orbit is solved using a Newton-Raphson algorithm. The linearised equations determining the stability lead to a generalised eigenvalue problem. Unlike the Fourier-Galerkin method, the use of Chebyshev polynomials or finite differences has the advantage that the relevant Floquet exponents are directly given without the well known issue of having to sort out the eigenvalues. The speed of convergence of these three methods is illustrated with examples from the Lorenz system, the Langford system and a two-dimensional thermal convection flow inside a differentially heated cavity. This last example demonstrates the potential of the newly proposed Chebyshev expansion for large-scale problems arising from the discretisation of the incompressible Navier-Stokes equations.
A. Gesla
Sorbonne Université, F-75005 Paris, France
LISN-CNRS, Université Paris-Saclay, F-91400 Orsay, France
Y. Duguet
LISN-CNRS, Université Paris-Saclay, F-91400 Orsay, France
P. Le Quéré
LISN-CNRS, Université Paris-Saclay, F-91400 Orsay, France
L. Martin Witkowski
Université Claude Bernard Lyon 1, CNRS, Ecole Centrale de Lyon, INSA Lyon, LMFA, UMR5509, 69622 Villeurbanne, France
Mathematics Subject Classification (2020) MSC 37M20
Introduction
Identification of periodic orbits (POs) of nonlinear dynamical systems, together with the computation of their stability, is crucial in many domains of physics and engineering. Periodic solutions are often the simplest unsteady solutions and emerge, for instance, from the local destabilisation of a fixed point solution of the system. In dissipative dynamical systems featuring attractors, periodic orbits can be stable or unstable. Deterministic chaotic attractors are known to be organised around an infinity of unstable periodic orbits ranked by their stability [10], which makes periodic orbits and their stability characteristics crucial also in chaotic settings.
Many numerical approaches to the determination of periodic orbits in ordinary differential equations (ODEs) exist. Popular choices include shooting methods and their variants (see § 2) and delayed feedback methods [32]. Other methods include adjoint-based methods [25, 14, 3, 29]. One of the most widely used alternative methods is a Fourier-based Galerkin method known as Harmonic Balance Method (HBM)(e.g. Cochelin and Vergez [9]). HBM approximates the periodic orbit as a truncated Fourier series. Due to the spectral expression of the cycle, high accuracy or equivalently fast convergence can be achieved. Historically, HBM was used mainly for low-dimensional ODE systems in mechanical and electrical engineering. Recent studies have considered such methods for higher-dimensional systems stemming from the spatial discretisation of partial differential equations (PDEs), in particular from fluid mechanics [5, 33, 36, 19]. The linear stability of an HBM solution is usually analyzed by forming the so-called Hill matrix and analyzing its eigenvalues in the search for unstable Floquet exponents. As pointed out by several authors (see e.g. Lazarus and Thomas [26]), a major drawback of using Hill’s matrix is that only a subset of the computed eigenvalues are relevant. As a consequence, it is necessary to filter out the spurious exponents from the relevant ones. Various strategies have been proposed either in low-dimensional [26, 30, 4] (Wu et al. 2022) or high-dimensional systems [31, 22].
The Fourier-Galerkin method referred to here combines the Harmonic Balance method for the search of the periodic solution in combination with Hill’s method to investigate its stability. As the key idea of this article, it is possible to use another type of temporal discretisation than the Fourier basis in the HBM equations. The first example described here corresponds to a finite difference discretisation. The present work furthermore proposes a description where the whole periodic orbit and its perturbation are expanded as a weighted sum of Chebyshev polynomials. The stability of the orbit emerges naturally from a generalised eigenvalue problem which involves the Jacobian matrix already computed in the Newton-Raphson method to find the orbit itself, which yields the most unstable Floquet multipliers.
Several studies have already made use of Chebyshev polynomials in the context of periodic orbits. The earliest use of Chebyshev polynomials for the computation of periodic motion concern orbital motion [23]. Zhou et al. (2001) also used a Chebyshev expansion of a PO in forced and an unforced problems. Zhou et al. (2013) uses the Chebyshev expansion to approximate the periodic orbit but the stability properties of the orbit are analyzed by forming a monodromy matrix. The authors claim higher accuracy when using a Chebyshev basis compared to a Fourier basis. The examples reported hereafter do not support this claim. A previous study based on a Chebyshev expansion for the linear stability analysis (LSA) of periodic orbits is the work by Woiwode and Krack (2023). The authors used Chebyshev polynomials to construct the monodromy matrix built on top of the orbit expressed with Fourier modes, which differs from the method presented here. They also report that the HBM is usually superior to the shooting methods when a smooth periodic solution is considered, yet it suffers from an excessive number of modes in the decomposition in the presence of sharp gradients. It is the aim of the present paper to revisit the use of Chebyshev polynomials in a few examples of dynamical systems with periodic solutions, in comparison with other more standard approaches.
The article is structured as follows. Section 2 recapitulates the theoretical basis concerning the periodic solutions of the dynamical systems and the Floquet theory used to study the stability of these solutions. The different numerical techniques presented include finite differences, the Harmonic Balance Method, and a method based on Chebyshev expansion. Section 3 presents examples of periodic orbits in the Lorenz and Langford systems, together with the linear stability analysis of these orbits. A direct application to a periodic solution in a high dimensional system of a fluid flow inside a differentially heated cavity is shown in section 4. Finally, section 5 summarises the presented work.
Numerical identification of a periodic orbit and stability analysis
General concepts
We consider a nonlinear system governed by the following equation :
where is the state vector of the system and is the number of degrees of freedom, and is a smooth enough function of . A periodic orbit of the system satisfies the relation:
for a value of and for all times . The smallest value of defines the period of the PO. Eq. (2) does not define the solution in a unique way as any solution shifted in time is also a solution. Imposing a phase condition selects one solution (see chapter 10 in Kuznetsov et al. [24]). In the current work the time derivative of one of the components of will be fixed to zero to form such a phase condition.
Basic elements of Floquet theory
Let be a periodic solution of (1). We consider a perturbation to in the form . Linearising (1) around leads to the non-autonomous linear O.D.E.
where the linear operator is the Jacobian matrix associated with evaluated along the periodic solution. This operator is time-periodic with the same period as . The stability of the nonlinear solution to infinitesimal perturbations is directly related to the spectral properties of . According to Floquet theory (Floquet [16]), in finite dimension , a general solution to (3) can be written under the form
where the ’s are periodic functions of of period , and the ’s and ’s are complex numbers. The exponents , are by definition the Floquet exponents of and they are defined modulo . Their real parts, interpreted as Lyapunov exponents, determine the stability of the periodic solution . In particular, is unstable in the Floquet sense if and only if at least one of the Floquet exponents has a strictly positive real part. Alternatively one may consider the characteristic Floquet multipliers , which are uniquely defined. The orbit is then unstable if the modulus of its largest characteristic multiplier is larger than unity, i.e. .
In practice, for the computation of the Lyapunov exponents, consider an orthonormal basis at . For each vector , , integrating the linear O.D.E. (3) over exactly the period of leads to the monodromy matrix . Exploiting the periodicity of the ’s, we have , with a time-independent operator. The Lyapunov exponents are directly extracted as the real part of eigenvalues , of , ranked by decreasing value ([13, 8]). In practice however, for large , due to the cost of accurately integrating times (3), forming the monodromy matrix can be computationally prohibitive both in terms of CPU and storage. We are hence looking for a way to extract all the leading Floquet exponents without directly forming the monodromy matrix.
General background
There is no general way to find all periodic orbits of a nonlinear dynamical system [24]. The simplest method, if a stable PO exists, is to use a numerical time integrator and solve an initial value problem (IVP). Provided the initial point belongs to the basin of attraction of the requested PO, numerical time integration will converge towards this specific PO. However for weakly stable orbits convergence may be slow, while unstable POs will not be found using a numerical time integrator alone.
If time is rescaled such that the PO is defined over a time interval of length unity, the period becomes a parameter of the system and finding a PO can be recast as a boundary value problem (BVP). Equations (1) and (2) become
while equation (3) becomes
More specifically, eq. (5) is a two-point BVP where periodic boundary conditions are imposed. This observation paves the way for alternate strategies to find PO owing to the numerous numerical techniques dedicated to two-point BVPs.
One of the most popular techniques for solving two-point BVPs is the shooting method and its variants, see e.g. Atkinson [2]. It is based on sequences of a repeated IVP solver, iteratively refining the initial value and the period to the point where periodicity is satisfied over the whole unit interval. The iterative process converges when the initial value lies on a stable or unstable periodic orbit and matches the end value within some prescribed accuracy. Overviews of shooting algorithms applied to large multidimensional systems can be found in Sánchez et al. [34]; Viswanath [41]; Frantz et al. [17]. However, shooting methods may fail if the PO is too unstable. Even if the initial value is close to a PO point, numerical integration can make a trajectory deviate rapidly from the desired PO. The perturbation to the PO can undergo such a large amplification over a period that numerical convergence can be hampered [12].
More robust alternate techniques to solve two-point BVPs can be used. The entire PO can be discretised as a finite set of points. The governing equations are satisfied at each of these points. As an alternative, the PO can also be expanded as a series on a well-chosen functional basis. The most straightforward option, given the temporal periodicity of the solution, is to consider a Fourier expansion, yet this is not the only possible choice. Instead of iterating only the initial condition and the value of the period as in iterative shooting methods, each iteration to reach the PO involves computing all the points of the (discretised) orbit -or all coefficients of the expansion- simultaneously. The size of the linear system solved at each iteration scales linearly with times the number of points on the orbit or the number of retained coefficients. Variants of these approaches combine highly accurate polynomial expansions between discrete points of the orbit. This technique is implemented for example in the software package AUTO [11].
In the approach retained here, we discretize the entire PO. We introduce and compare three different discretizations to compute the periodic base solution and predict its stability in the Floquet sense.
Finite differences discretisation
Eq. (5) is discretised at times corresponding to time intervals. In the simplest form the time derivative in Eq. (5) is approximated using a second-order accurate Crank-Nicolson finite difference scheme:
where is the component of expressed at the th timestep, is the index of the temporal position on the orbit and is the time interval between two states and . For simplicity, is assumed constant equal to . This leads to a discrete set of equations:
Equation (8a) is written for all points on the orbit and all degrees of freedom , resulting in scalar equations. Equation (8b) imposes periodicity for degrees of freedom ( scalar equations). Equation (8c) closes the system for the extra unknown using a phase condition (one scalar equation). The exact number of equations and variables is therefore . Instead of (8c), other phase conditions can be considered (see chapter 7 in Seydel [35]). Here, the time derivative of the first degree of freedom () is imposed to be zero to second-order accuracy at the midpoint between and . The resulting algebraic nonlinear system (8) is solved by Newton-Raphson’s method.
The Jacobian computed at each Newton-Raphson iteration is a square bordered block matrix of dimension of the form:
where and are the identity and null submatrices and is the null column vector in . The submatrices , have size . The submatrices are obtained as the first-order partial derivative of the left hand side of (8a) with respect to for all (with fixed). Similarly the submatrices are obtained by deriving the left hand side of (8a) with respect to for all (with and fixed). The row vector is given by . Finally, the are column vectors of size .
This specific structure of the Jacobian matrix is obtained by choosing the ordering of the unknowns as in . As the Jacobian matrix is sparse and bordered, efficient techniques exist to solve the linear system at each Newton step.
Computation of the stability of the periodic orbit solution
The stability of the periodic solution is found by introducing explicitly the complex Floquet multipliers in an eigenvalue problem which comes directly as a by-product of the computation of the periodic orbit. This approach was suggested by Fairgrieve and Jepson [13] in order to avoid forming the monodromy matrix. Let us assume that the periodic solution is encoded in a solution vector and is perturbed in the direction of one of the Floquet vectors , itself associated with a Floquet multiplier . Then the following constraint on has to be satisfied:
The orbit is stable along the direction spanned by if , neutrally stable if and unstable if . In order to use the Jacobian matrix (9) in the eigenvalue problem with no extra computation, equation (10), following Fairgrieve and Jepson [13], is rewritten as:
Linearisation of (8) around the solution leads to the generalised eigenvalue problem
The square matrix of dimension is extracted directly from the matrix (9) by removing the last line and the last column. The square matrix of dimension is defined by:
This generalized eigenvalue problem involves the matrix , the very same one as considered by Fairgrieve and Jepson [13]. Its block structure is shown in Eq. (15) and contains as only non-null elements the identity matrix, denoted as . In practice we used the routine of Matlab to find the eigenvalues of the generalised eigenvalue problem (12) if the size of the linear system is moderate. For larger systems, more specific routines are used as detailed in § 5. Note that when the matrices and are real, the eigenvalues and the eigenvectors to (12) are either real-valued or come in complex conjugate pairs. eigenvalues, among the eigenvalues output by the generalised eigensolver, are infinite. The relevant Floquet multipliers , , are computed from the finite eigenvalues , . This is one of the main advantages of this method compared to, say, Hill’s method (discussed in next section): the number of meaningful eigenvalues is the one expected from theory, namely . In particular there is no need to sort out the physically meaningful eigenvalues from spurious ones resulting from enlarging the problem from dimension to . Here we did not make any assumption on these eigenvalues (e.g. the existence of a neutral one as implied by the continuous dynamics, or the relation between them) and directly report them as output by the eigenvalue solver.
Fourier-Galerkin method
A natural choice to cope with the periodicity of the PO is the use of trigonometric functions as a functional basis in a spectral expansion. In Harmonic Balance Method, the periodic orbit is expanded as a standard truncated Fourier series:
where is finite and the ’s are complex coefficients. When is real-valued, the property that, for all , and are complex conjugate, can be used to reduce the number of real-valued degrees of freedom down to . As is common with spectral expansions, HBM allows for fast convergence provided the target solution is smooth enough. The complex coefficients are found by introducing the decomposition (14) into the governing equations (5). This yields the residual
A set of algebraic equations for the ’s is generated by requesting that the residual be orthogonal to each Fourier function. We define the products
where the ’s have dimension , and request that , .
The phase condition, similar to (8c), reads
Like for the finite differences method, these equations are solved using a standard Newton-Raphson method. The Jacobian matrix reads
where
and is a row vector in defined by
Compared to the system arising from the finite difference discretisation, here the Jacobian matrix is no longer sparse.
Computation of the stability of the periodic orbit solution
In order to retrieve the stability of the orbit, Hill’s method can be used. The details of the method are described by Guillot et al. [20] and we only summarize the main steps here. Let us assume again that the periodic solution is perturbed in the direction of one of the Floquet vectors itself associated with a Floquet exponent . The Floquet decomposition, together with the Fourier decomposition of the periodic function involved, leads in this particular case to:
In practice it is assumed that although this is no strict request. The temporal growth or decay of the perturbation is governed by the exponential term for the exponent with largest real part. As in the previous subsection, all the entries of for are stored in a vector of size . The expression (22) is introduced into the linearised equation (6). This yields the eigenvalue problem for
where is the Jacobian matrix (20) with the last column and last row removed. Unlike the generalised eigenvalue problem (14) arising in the finite difference case, it admits finite eigenvalues rather than only .
One well-known problem associated with Hill’s method is the difficulty of identifying which Floquet exponents are meaningful (Lazarus and Thomas [26]). As the method is expected to output an infinity of identical Floquet multipliers that are identical modulo . For finite truncation however, this exact degeneracy is lost and one is left with a large set of exponents to analyse. In particular the prediction of the exponent with largest real part can not be trusted (Kern [21]). This is a major limitation when predicting the stability of a given periodic orbit.
A way out is to express the periodic orbit and especially the perturbation as a weighted sum of functions from another functional basis, one that is not by construction periodic. We consider in this study the basis of Chebyshev polynomials. Expanding a smooth enough function on a Chebyshev basis is known to minimise the pointwise error of the approximation of analytic functions (Boyd [6]). This makes the Chebyshev basis a relevant choice over any other polynomial basis.
Chebyshev polynomial expansion
We begin by expanding the periodic orbit solution as a finite sum of Chebyshev polynomials:
where the number of polynomials retained in the truncation is . Since the polynomials are defined for a proper rescaling of the governing equations is required. Unlike Fourier trigonometric functions, the Chebyshev polynomials are not periodic by construction, hence the periodicity condition for the PO must now be enforced. The decomposition (24) is introduced into the governing equation (5) leading to:
where the Chebyshev polynomials of the second kind are used to express the derivative of . The time derivative has to be transformed as , where the original time is transformed into .
In order to generate a set of nonlinear algebraic equation (25) is evaluated at the Gauss-Lobatto points. We recall that Gauss-Lobatto points are defined by
and span non-uniformly the whole interval . Writing equation (25) at the points yields equations. Enforcing the periodicity condition
using and yields additional equations. The last equation comes from the phase condition: enforcing a zero derivative with respect to time at to one specified degree of freedom (here , the first component of ) leads to:
where the ’s are the weights for . The set of nonlinear algebraic equations is solved with the Newton-Raphson method.
The Jacobian matrix computed at each Newton-Raphson iteration is a square bordered block matrix of dimension of the following form, assuming the following ordering of the unknowns :
and are submatrices defined by:
The submatrices and and the vector are already defined in section 2.4. and are respectively column and row vectors in defined by:
Computation of the stability of the periodic orbit solution
Once the spectral coefficients of the periodic orbit and the period T are known, the linear stability of the orbit can be assessed. The procedure is closely related to that used for finite differences. We introduce a small perturbation to the solution:
where . For the perturbation we do not assume periodicity but rather
where is again interpreted as a Floquet multiplier controlling the stability of the periodic orbit. Inserting the condition (29) into the ansatz (28) leads to:
Eq. (29) is decomposed as as :
The decomposition (28) is introduced into equation (6). Evaluation at the first Gauss-Lobatto points yields the generalised eigenvalue problem
Owing to equation (31), the matrix is directly determined by the Jacobian matrix (27) computed at convergence (last iteration in Newton-Raphson solver) by removing the last row and the last column. Each eigenvector is ordered in the following way . The square matrix of dimension is defined by :
The eigenvalue problem (32) has eigenvalues. Despite the different approximation to the solution compared to Fairgrieve and Jepson [13], the matrix involved in (32) is the same as in that study. As a consequence, among the eigenvalues computed using Matlab’s eig() routine (or specific routines for large linear systems), eigenvalues are infinite and the Floquet multipliers are computed from the finite eigenvalues . The Floquet exponents , are directly deduced from the Floquet multipliers , via the relation . The periodic orbit is unstable if (i.e. ). The main property to exploit is that the eigenvalue problem (32) yields only finite multipliers that can directly be interpreted as numerical approximations of the true Floquet multipliers. This is in contrast to Hill’s method that yields finite exponents together with no obvious method to sort them out and select the meaningful ones.
Application to low-dimensional systems
In this section several examples of periodic orbits in autonomous dynamical systems are presented. The stability of the orbits, predicted by either real or complex Floquet multipliers, is computed successively using finite differences, Fourier and Chebyshev spectral time discretisations.
Figure 1. The unstable periodic orbit () for the Lorenz system , , as a state portrait on the left and in a component-wise representation over one period on the right. A stable fixed point is marked with . Representation obtained with HBM using modes. .
Lorenz system : three cases
The Lorenz system is governed by a system of three ordinary differential equations:
=24
For , , the system admits an unstable periodic orbit born in a subcritical Hopf bifurcation at of a fixed point Lorenz [28], shown in figure 1.
The stability of the orbit is found using each of the three methods presented in § 2. The resulting spectrum and the convergence of the unstable Floquet exponent are shown in figure 2. For the evaluation of the error on the exponents and we take as reference exponents () those extracted from Hill’s matrix for . The absolute error for each exponent is computed as . The values of the three Floquet exponents are approximately , , for stable, neutral and unstable perturbations, respectively, with the value for the neutral exponent assumed exactly zero.
First, as can be seen in figure 2, the spectrum associated with Hill’s matrix is populated by a large number of spurious eigenvalues, whereas for a three-dimensional ODE system such as the Lorenz system, only three eigenvalues are needed (one of them being necessarily neutral). In the absence of any strategy to filter the eigenvalues of Hill’s matrix it can thus be challenging to draw conclusions concerning the orbit’s stability.
Figure 2. Left: Subset of Floquet exponents of the orbit of the Lorenz system for found by Hill’s method for and . Only the Floquet exponents around and are displayed. The region of the complex plane around is omitted for clarity. Right: Corresponding Floquet multipliers . Hill’s method yields a set of exponents that have to be filtered to discard the spurious values.
The convergence of the Floquet exponents is listed in Table 1 and also illustrated graphically in Figure 3. Convergence is reported for each exponent provided the resolution is high enough. On one hand, both HBM and Chebyshev method
Figure 3. Error committed on each Floquet exponent for a periodic orbit of the Lorenz system for , , computed with Finite Differences, Hill’s method (Fourier) and Chebyshev method. Exponents are displayed in increasing order from left to right. Hill’s and Chebyshev methods show spectral convergence while Finite Differences only show second order convergence for the non-zero exponents. Results reported as a function of the number of real-valued degrees of freedom, namely , and respectively.
| 13 | -14.3895880 | 0.0437157935 | 0.692268000 | |
| 35 | -13.7967351 | 0.0460974317 | 0.681075444 | |
| 60 | -13.7412922 | 0.0463577558 | 0.679926770 | |
| 100 | -13.7232514 | 0.0464440859 | 0.679548972 | |
| 300 | -13.7142788 | 0.0464873253 | 0.679360332 | |
| 1 | -13.6751635 | 0.0468109907 | 0.677821008 | |
| 4 | -13.7131592 | 0.0464931094 | 0.679336744 | |
| 9 | -13.7131594 | 0.0464927338 | 0.679336764 | |
| 19 | -13.7131594 | 0.0464927338 | 0.679336764 | |
| 11 | 0.0438154932 | 0.678437981 | ||
| 13 | 0.0446843534 | 0.679357954 | ||
| 21 | 0.0464901602 | 0.679336823 | ||
| 31 | 0.0464927345 | 0.679336764 | ||
| 51 | 0.0464927338 | 0.679336764 |
Table 1. Three Floquet exponents and the period of one unstable periodic orbit of the Lorenz system for , , . , and are the numerical resolutions corresponding respectively to Finite Differences, Hill’s method and Chebyshev method. The Fourier-Galerkin method reaches machine precision with only modes, while the Chebyshev method reaches three significant digits with .
Chebyshev expansions are characterised by spectral exponential-like convergence, although slower for the Chebyshev method for the same number of degrees of freedom. On the other hand, as shown in Figure 3, the finite differences method is only second order accurate in determining non-zero Floquet exponents. The second-order accuracy is consistent with the fact that the approximation of the time derivative is also here second order. The advantage of Chebyshev expansions for identifying the relevant Floquet exponents is balanced by its slower convergence compared to HBM.
We report now on the convergence of the methods for more demanding parameter values. For , , the Lorenz system (34) admits a chaotic attractor and hence an infinite number of unstable periodic orbits [10]. The periodic orbit with the smallest period among those identified, labelled AB by Viswanath [40], is shown in Figure 4. For this specific orbit, the period found is and the largest Lyapunov exponent is . The remaining exponents are easily deduced from . Since the tangent vector is a neutral direction, . Using the property that the rate of growth rate of infinitesimal volumes is independent of space and time [28], we evaluate it at any point on the periodic orbit where it is equal to the sum . We deduce that
Figure 4. Periodic orbit solution of the Lorenz system for ()(left) and ()(right) in blue, together with an unstable (left panel) or stable (right panel) eigenvector (solid red line) computed with the Chebyshev method. Green dot : initial position, red dot : position after one period.
. The three Floquet exponents found numerically are listed in Table 2 as a function of the temporal resolution used. For the finite difference method (not included in the table), the relative error drops below 1% for the period and for all Floquet exponents when .
| 11 | -14.661 | 0.99542 | 1.55867856 | |
| 15 | -14.661 | 0.99466 | 1.55865243 | |
| 19 | -14.661 | 0.99465 | 1.55865221 | |
| 29 | 0.9868981 | 1.55723896 | ||
| 49 | 0.9948799 | 1.55863779 | ||
| 69 | 0.9946627 | 1.55865221 | ||
| 89 | 0.9946504 | 1.55865221 | ||
| 109 | 0.9946500 | 1.55865221 |
Table 2. Three Floquet exponents and period of an unstable periodic orbit of the Lorenz system for , , for HBM/Hill’s () and Chebyshev () method.
The last example from the Lorenz system is a stable periodic orbit. For there is a periodic window [38]. A stable orbit for is shown in Figure 4. Its Floquet exponents are listed in Table 3. As expected, for this stable orbit all non-zero exponents are negative apart from which is neutral. Hill’s method with modes achieves four digits of accuracy for all Floquet exponents. Comparable accuracy on using the Chebyshev method requests, in comparison, as many as modes.
| 19 | ||||
| 21 | ||||
| 23 | ||||
| 25 | ||||
| 27 | ||||
| 49 | ||||
| 69 | ||||
| 89 | ||||
| 109 | ||||
| 129 | ||||
| 149 |
Table 3. Three Floquet exponents and the period of the stable periodic orbit of the Lorenz system for , , for HBM/Hill’s () and Chebyshev () method.
Lorenz system : convergence analysis
The convergence of finite differences and/or spectral methods is well documented in the literature. In particular finite differences are characterised by power-law convergence (i.e. the error decreases as inverse powers of the resolution ) whereas for smooth enough solutions spectral methods, including Fourier and Chebyshev expansions, are characterised by (asymptotically faster) exponential convergence [6].
It is observed that, for equivalent accuracy, the number of Chebyshev coefficients is larger than the number of Fourier coefficients for the cases corresponding to tables 2 and 3. One reason for the slower convergence of the Chebyshev expansion could be the limited smoothness of the function being approximated. To investigate the convergence properties of the expansions (14) and (22), the magnitude of Fourier and Chebyshev coefficients corresponding to the base solution and the least stable eigenvector are plotted in figure 5. For both cases geometric convergence is observed. Following Trefethen [39], the coefficients of the truncated Chebyshev expansion verify
with
and the difference between the truncated expansion and the function both converge geometrically according to
provided is analytic in a Bernstein ellipse around the interval with . Furthermore, the following inequality holds [37] for large enough :
which relates the approximation error to the last coefficient of the computed expansion. Therefore, since and are at least analytic in the large number of coefficients needed in practice is caused by the rich dynamics of the periodic orbit and its eigenmodes rather than by the lack of regularity of the solution.
The speed of convergence of the Fourier expansion with respect to the Chebyshev expansion can be explained by looking at the particular shape of the eigenvectors in Figure 5. The eigenvector shows a narrow bump in the middle of the interval, which is known to hamper the speed of convergence of the Chebyshev expansion by comparison with the Fourier expansion. The ratio observed in Table 3 to reach the same accuracy is consistent with the rule-of-thumb put forward by Boyd [6].
Figure 5. Periodic orbit solution (top) and least stable Floquet eigenvector (bottom) for the Lorenz system , , (cf. Figure 4 right) and the corresponding magnitudes of Fourier and Chebyshev coefficients for cases and . Whereas the base solution is periodic, its corresponding least stable eigenvector is not. Note that the Chebyshev coefficients fall to the eye on two distinct lines, which is related to a particular symmetry of the solution seen in the Fourier coefficients. Geometric convergence of the coefficients, marked by the constant slope in the lin-log plot, is characteristic for functions that are smooth in variable .
Langford system
The periodic orbits of the Lorenz system presented so far were either stable or unstable, but always with real Floquet multipliers. The question arises whether the Chebyshev expansion can also handle the analysis of periodic orbits, stable or unstable, characterised by complex Floquet multipliers. Whenever stability is lost due to a pair of non real Floquet multipliers crossing the unit circle in the complex plane, a Neimark-Sacker bifurcation occurs, leading to additional oscillations on top of the PO. A low-dimensional example of such dynamics is the three-dimensional Langford ODE system ([35]) given by
The system is parametrised by the real parameter . For , it possesses an exactly circular periodic orbit whose analytical expression is given by
with . This PO loses stability above . Due to its simple analytical expression it is possible to express analytically its stability exponents as:
where denotes the position of the plane containing the orbit (see Figure 6), its radius and to the complex Floquet exponent of the orbit. At (, ) the orbit loses stability to a complex pair of Floquet exponents and leaves place to a 2-torus. The periodic orbit, the torus and the instability mode computed using the Chebyshev method are all shown in Figure 6. A pair of unstable non-trivial Floquet exponents is reported in Table 4 depending on the resolution. Using modes six digits of the real part of the unstable Floquet exponent are computed accurately. The Chebyshev expansion is hence able to capture the oscillatory bifurcation.
Figure 6. Left: trajectory of the Langford system obtained using a time integrator for values of below (blue) and above (red). For the trajectory converges asymptotically to the stable perfectly circular periodic orbit. For the trajectory diverges towards a stable torus. Right: the unstable periodic orbit and its eigenvector for , obtained using the Chebyshev method. is marked with a green and with a red arrow.
| 17 | |||
| 21 | |||
| 25 | |||
| 29 | |||
| exact |
Table 4. Non-trivial Floquet exponents for the Langford system. The row labelled exact corresponds to the analytical value of the exponent calculated using (42).
Differentially Heated Cavity
In order to check the applicability of the Chebyshev expansion in time to a much higher-dimensional problem, we consider the buoyancy-driven flow inside an air-filled differentially heated cavity with perfectly conducting horizontal top and bottom boundaries. This configuration was shown experimentally to undergo a transition to unsteadiness for values of the Rayleigh number around [7]. It was much used in the late 1980’s as a test case to predict numerically the transition to unsteadiness. Different numerical algorithms have provided the critical value for the loss of stability of the steady solution at [42, 18, 45]. Numerical simulations of the unsteady equations, reported in Le Quéré [27], showed that the nonlinear unsteady solutions above belong to different branches of solutions according to their oscillation period (see figure 5 therein) and that the first branch of time-periodic solutions born at loses its stability at a value of around . This flow configuration thus constitutes a good candidate to test the applicability of algorithms aimed at computing the instability of periodic solutions of the Navier-Stokes equations.
Governing equations and numerical methods
The non-dimensional velocity , pressure and temperature obey the Navier-Stokes equations under the Boussinesq approximation:
with a Prandtl number . The square cavity is shown in Figure 7 together with the boundary conditions. For the sake of brevity we refer to Xin and Le Quéré [45] for a more detailed presentation of the physical and dimensionless parameters.

Figure 7. Square differentially heated cavity with perfectly conducting top and bottom walls. Temperature field of the steady base flow at together with the boundary conditions. No slip conditions is assumed at the walls. Gravity points downwards towards decreasing . The colormap spans the interval uniformly with 8 distinct colors. The solution verifies the symmetry .
The cavity is discretised with a Finite Volume method using cells resulting in degrees of freedom (see Faugaret et al. [15]). To account for the existence of thin boundary layers along the walls, non-uniform grids in and are used, the cell boundaries being distributed as :
For the given resolution, the steady base flow is converged to machine precision using a Newton-Raphson method. The base flow exhibits a symmetry with respect to the cavity center that reads :
while perturbations can be invariant under or under
Its linear stability is computed by forming an eigenvalue problem and solving it numerically using ARPACK. More details can be found in Gesla et al. [19].
Instability of the Hopf branch
The temperature field for the steady base flow is shown in Figure 7 for . At it loses its stability in a supercritical Hopf bifurcation that gives rise to a stable periodic orbit for . To trace out the amplitude of the cycle born at , an observable linked to vorticity is monitored. We use the -norm of the vorticity perturbation to define
where is the vorticity, is the equivalent quantity for the base flow, and the r.m.s. is applied over time. A bifurcation diagram is shown in Figure 8 as a function of . At the branch of periodic solutions is born. The periodic orbit is described at each subsequent value of with HBM. Continuation in is performed using standard pseudo-arclength continuation method [1]. Stability of the periodic solution can be investigated with Hill’s method in the case of a Fourier expansion in time or with Chebyshev expansion, as proposed in this work.
Figure 8. Bifurcation diagram for a square differentially heated cavity for the observable defined in (48). The steady base flow is stable for (red solid line) and unstable above (red dashed line). The supercritical Hopf branch of periodic orbits emanating from is stable for (blue solid line) and unstable above (blue dashed line).
The evolution of the most unstable Floquet exponent of the supercritical branch is investigated using the Chebyshev expansion in time with and is listed in Table 5. It shows that the periodic orbit loses its stability to a pair of complex conjugate Floquet exponents in a Neimark-Sacker bifurcation slightly above , thus confirming the result by Le Quéré [27].
Table 5. Evolution of the leading Floquet exponent and multiplier of the periodic branch in the differentially heated cavity computed with the Chebyshev method (). The loss of stability occurs for between and .
Table 6 provides a comparison of the convergence of the largest Floquet exponent as a function of the number of temporal degrees of freedom in Fourier or Chebyshev expansion. For the present spatial discretisation at , two correct digits of the real part of the leading Floquet exponent and five correct digits of the period require Chebyshev modes. The linearly unstable base flow is shown in Figure 9, together with temperature perturbation snapshots along the first branch and its least stable Floquet eigenvector. As previously observed in § 4, the Fourier expansion converges faster than the Chebyshev expansion. The sorting issue of Hill’s method was avoided by first computing approximate values of the Floquet multipliers with the Chebyshev algorithm. In other words, taking advantage of the fast convergence properties of Fourier expansions is only possible once the correct eigenmodes have been unambiguously identified. This shows that both methods can be used in a complementary way.
Table 6. Convergence study for the least stable Floquet exponent and the period of the stable orbit of the differentially heated cavity obtained with Fourier method () and Chebyshev method (). .

Figure 9. Temperature contours of the base flow (top left), instantaneous field of the saturated periodic orbit (top right), instantaneous field of saturated perturbation of the periodic orbit (bottom left) and its least stable Floquet vector (bottom right) for . Note that the base flow and the saturated periodic orbit temperature fields obey the symmetry , while the Floquet vector obeys the other symmetry .
Summary
We have presented and compared three methods to compute periodic solutions of multi-dimensional systems of ordinary differential equations and to investigate their stability characteristics. Three different time discretisations investigated include finite differences and spectral discretisation using Fourier modes (also known as Harmonic Balance method) or Chebyshev polynomials. All three methods rely on computing the periodic orbit using Newton-Raphson iterations. The linear stability of the periodic solution is investigated by forming a generalised eigenvalue problem without the need to consider, to form let alone to store a monodromy matrix. When the PO obeys to a PDE involving spatial and time variables, the observation that the time dependency is recast into a two-point BVP is of practical interest. Time-dependence can be handled as any spatial variable and thus, the implementation of the three presented methods in an existing PDE steady-state solver requires no conceptual change.
While the finite-difference time discretisation is only algebraically accurate, both Fourier-Galerkin and Chebyshev methods share similar spectral convergence characteristics. In practice however, Fourier-Galerkin requires less modes to reach a given accuracy. The true advantage of the Chebyshev method is brought to the forefront when extracting the stability of the orbit. While the HBM is efficient to compute the periodic orbit, when coupled to Hill’s method, it is known to yield a spectrum with non-trivial sorting issues which are still the subject of ongoing research. The finite difference and Chebyshev methods output a directly usable spectrum, thereby avoiding the sorting issues associated with Hill’s method. This property is directly associated with the structure of the matrix appearing in the generalised eigenvalue problem governing the stability of the periodic orbit. The three methods were tested on selected stable or unstable periodic orbits of the Lorenz system, on a Neimark-Sacker bifurcation in the Langford system, as well as on a two-dimensional cavity flow governed by the incompressible Navier-Stokes equations. This demonstrates the potential of Chebyshev method in conjunction with Hill’s method for the Floquet analysis of time-periodic solutions of large-scale systems such as those coming from the discretisation of PDEs.
Acknowledgements
Acknowledgements The authors wish to thank Daniel Henry, Wietze Herreman, Simon Kern, Arnaud Lazarus and Laurette Tuckerman for fruitful discussions and many valuable comments.
Conflict of interest
The authors declare no conflict of interest.
Data availability statement
Data sets generated during the current study are available from the corresponding author on reasonable request.
References
- [1]Allgower, E. L. and Georg, K. (2012). Numerical Continuation Methods: An Introduction, volume 13 of Springer Ser. Comput. Math. Springer Science & Business Media, Cham.
- [2]Atkinson, K. (1991). An Introduction to Numerical Analysis. John Wiley & Sons, New York.
- [3]Azimi, S., Ashtari, O., and Schneider, T. M. (2022). Constructing periodic orbits of high-dimensional chaotic systems by an adjoint-based variational method. Phys. Rev. E, 105(1):014217.DOI
- [4]Bayer, F. and Leine, R. I. (2023). Sorting-free Hill-based stability analysis of periodic solutions through Koopman analysis. Nonlinear Dyn., 111(9):8439–8466.DOI
- [5]Bengana, Y., Loiseau, J.-C., Robinet, J.-C., and Tuckerman, L. (2019). Bifurcation analysis and frequency prediction in shear-driven cavity flow. J. Fluid Mech., 875:725–757.DOI
- [6]Boyd, J. P. (2001). Chebyshev and Fourier Spectral Methods. Courier Corporation, Mineola, New York.
- [7]Briggs, D. G. and Jones, D. N. (1985). Two-dimensional periodic natural convection in a rectangular enclosure of aspect ratio one. J. Heat Transfer, 107(4):850–854.
- [8]Chicone, C. (2006). Ordinary differential equations with applications. Springer, New York.
- [9]Cochelin, B. and Vergez, C. (2009). A high order purely frequency-based harmonic balance formulation for continuation of periodic solutions. J. Sound Vib., 324(1-2):243–262.
- [10]Cvitanovic, P., Artuso, R., Mainieri, R., Tanner, G., Vattay, G., Whelan, N., and Wirzba, A. (2005). Chaos: classical and quantum. ChaosBook.org (Niels Bohr Institute, Copenhagen 2005), 69:25.
- [11]Doedel, E., Paffenroth, R., Champneys, A., Fairgrieve, T., Kuznetsov, Y., Sandstede, B., and Wang, X. (2001). AUTO 2000: Continuation and bifurcation software for ordinary differential equations (with HomCont). Technical report, Caltech.
- [12]Duguet, Y., Pringle, C. C., and Kerswell, R. R. (2008). Relative periodic orbits in transitional pipe flow. Phys. Fluids, 20(11).
- [13]Fairgrieve, T. F. and Jepson, A. D. (1991). OK Floquet multipliers. SIAM J. Numer. Anal., 28(5):1446–1462.
- [14]Farazmand, M. (2016). An adjoint-based approach for finding invariant solutions of Navier–Stokes equations. J. Fluid Mech., 795:278–312.
- [15]Faugaret, A., Duguet, Y., Fraigneau, Y., and Martin Witkowski, L. (2022). A simple model for arbitrary pollution effects on rotating free-surface flows. J. Fluid Mech., 935:A2.
- [16]Floquet, G. (1883). Sur les équations différentielles linéaires à coefficients périodiques. Annales scientifiques de l’École Normale Supérieure, 2e série, 12:47–88.
- [17]Frantz, R. A., Loiseau, J.-C., and Robinet, J.-C. (2023). Krylov methods for large-scale dynamical systems: Application in fluid dynamics. Appl. Mech. Rev., 75(3):030802.
- [18]Gelfgat, A. Y. and Tanasawa, I. (1994). Numerical analysis of oscillatory instability of buoyancy convection with the Galerkin spectral method. Num. Heat Transf. A, 25:627–648.
- [19]Gesla, A., Duguet, Y., Le Quéré, P., and Martin Witkowski, L. (2024). Subcritical asymmetric solutions in rotor-stator flow. Phys. Rev. Fluids, 9:083903.
- [20]Guillot, L., Lazarus, A., Thomas, O., Vergez, C., and Cochelin, B. (2020). A purely frequency based Floquet-Hill formulation for the efficient stability computation of periodic solutions of ordinary differential systems. J. Comput. Phys., 416:109477.
- [21]Kern, J. S. (2023). Linear and non-linear dynamics of non-autonomous flows. PhD thesis, KTH Royal Institute of Technology.
- [22]Kern, J. S., Lupi, V., and Henningson, D. S. (2024). Floquet stability analysis of pulsatile flow in toroidal pipes. Phys. Rev. Fluids, 9(4):043906.DOI
- [23]Kolenkiewicz, R. and Carpenter, L. (1967). Periodic motion around the triangular libration point in the restricted problem of four bodies. Astron. J., 72:180.DOI
- [24]Kuznetsov, Y. A., Kuznetsov, I. A., and Kuznetsov, Y. (1998). Elements of Applied Bifurcation Theory, volume 112 of Appl. Math. Sci. Springer, Cham, Switzerland.
- [25]Lan, Y. and Cvitanović, P. (2004). Variational method for finding periodic orbits in a general flow. Phys. Rev. E, 69(1):016217.pubmed.ncbi.nlm.nih.gov/14995703
- [26]Lazarus, A. and Thomas, O. (2010). A harmonic-based method for computing the stability of periodic solutions of dynamical systems. C. R. Mécanique, 338(9):510–517.DOI
- [27]Le Quéré, P. (1994). Onset of unsteadiness, routes to chaos and simulations of chaotic flows in cavities heated from the side: a review of present status. In International Heat Transfer Conference Digital Library. Begel House Inc.DOI
- [28]Lorenz, N. (1963). Deterministic nonperiodic flow. J. Atmos. Sci., 20(2):130–141.
- [29]Parker, J. P. and Schneider, T. M. (2022). Variational methods for finding periodic orbits in the incompressible Navier–Stokes equations. J. Fluid Mech., 941:A17.DOI
- [30]Peletan, L., Baguet, S., Torkhani, M., and Jacquet-Richardet, G. (2013). A comparison of stability computational methods for periodic solution of nonlinear problems with application to rotordynamics. Nonlinear Dyn., 72:671–682.DOI
- [31]Pier, B. and Schmid, P. J. (2017). Linear and nonlinear dynamics of pulsatile channel flow. J. Fluid Mech., 815:435–480.DOI
- [32]Pyragas, K. (1992). Continuous control of chaos by self-controlling feedback. Phys. Lett. A, 170(6):421–428.
- [33]Rigas, G., Sipp, D., and Colonius, T. (2021). Nonlinear input/output analysis: application to boundary layer transition. J. Fluid Mech., 911:A15.DOI
- [34]Sánchez, J., Net, M., Garcia-Archilla, B., and Simó, C. (2004). Newton–Krylov continuation of periodic orbits for Navier–Stokes flows. J. Comput. Phys., 201(1):13–33.DOI
- [35]Seydel, R. (2009). Practical Bifurcation and Stability Analysis, volume 5 of Interdiscip. Appl. Math. Springer Science & Business Media, Cham.DOI
- [36]Sierra-Ausín, J., Citro, V., Giannetti, F., and Fabre, D. (2022). Efficient computation of time-periodic compressible flows with spectral techniques. Comp. Meth. Appl. Mech. Eng., 393:114736.DOI
- [37]Sinha, S. C. and Wu, D.-H. (1991). An efficient computational scheme for the analysis of periodic systems. J. Sound Vib., 151(1):91–117.DOI
- [38]Strogatz, S. H. (2018). Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering. CRC press, Boca Raton, FL.
- [39]Trefethen, L. N. (2019). Approximation theory and approximation practice. SIAM, Philadelphia, PA.
- [40]Viswanath, D. (2003). Symbolic dynamics and periodic orbits of the Lorenz attractor. Nonlinearity, 16(3):1035.DOI
- [41]Viswanath, D. (2007). Recurrent motions within plane Couette turbulence. J. Fluid Mech., 580:339–358.DOI
- [42]Winters, K. H. (1987). Hopf bifurcation in the double-glazing problem with conducting boundaries. J. Heat Transfer, 109(4):894–898.DOI
- [45]Xin, S. and Le Quéré, P. (2001). Linear stability analyses of natural convection flows in a differentially heated square cavity with conducting horizontal walls. Phys. Fluids, 13(9):2529–2542.DOI






