Nonlinear Dyn (2023) 111:22079–22099

https://doi.org/10.1007/s11071-023-08705-2

ORIGINAL PAPER

Model reduction for nonlinearizable dynamics via

delay-embedded spectral submanifolds
Joar Axås · George Haller

Received: 11 January 2023 / Accepted: 26 June 2023 / Published online: 16 July 2023
© The Author(s) 2023

Abstract Delay embedding is a commonly employed Keywords Invariant manifolds · Reduced-order

technique in a wide range of data-driven model reduc- modeling · Spectral submanifolds · Delay embedding ·
tion methods for dynamical systems, including the Embedding theorems
dynamic mode decomposition, the Hankel alterna-
tive view of the Koopman decomposition (HAVOK),
nearest-neighbor predictions and the reduction to spec- 1 Background
tral submanifolds (SSMs). In developing these applica-
tions, multiple authors have observed that delay embed- Much recent effort in nonlinear dynamics has focused
ding appears to separate the data into modes, whose on data-driven model reduction methods. Such algo-
orientations depend only on the spectrum of the sam- rithms return a simplified model of the system dynam-
pled system. In this work, we make this observation ics based on sampled trajectories from experiments or
precise by proving that the eigenvectors of the delay- simulations. Commonly pursued objectives for devel-
embedded linearized system at a fixed point are deter- oping these methods include dimensionality reduc-
mined solely by the corresponding eigenvalues, even tion, sparsity, and interpretability. Prevalent methods
for multi-dimensional observables. This implies that include the proper orthogonal decomposition (POD)
the tangent space of a delay-embedded invariant man- [4,51] and the dynamic mode decomposition (DMD)
ifold can be predicted a priori using an estimate of the [46,66,67], which fit models to data under various lin-
eigenvalues. We apply our results to three datasets to earity assumptions.
identify multimodal SSMs and analyse their nonlinear Linear models cannot, however, capture characteris-
modal interactions. While SSMs are the focus of our tically nonlinear (or nonlinearizable) phenomena. Such
study, these results generalize to any delay-embedded phenomena include the coexistence of, and the transi-
invariant manifold tangent to a set of eigenvectors at tion between, isolated and compact stationary states,
a fixed point. Therefore, we expect this theory to be such as fixed points, limit cycles, and invariant tori
applicable to a number of data-driven model reduction [55]. To address this shortcoming, the Sparse identi-
methods. fication of nonlinear dynamics (SINDy) algorithm fits
a sparse nonlinear model to training data using a library
of nonlinear functions [12]. However, the choice of this
library depends on the user [7] and the coordinate sys-
J. Axås · G. Haller (B)
Institute for Mechanical Systems, ETH Zürich, Leonhardstrasse
tem used. Additionally, the size of the library scales up
21, 8092 Zürich, Switzerland quickly with the problem dimensionality [45]. While
e-mail: [email protected] neural networks can pattern-match nonlinear phenom-

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22080 J. Axås, G. Haller

ena [19,21,64], the models they return are often diffi- sure modeling [56], and nearest-neighbor prediction
cult to interpret and generalize poorly outside the range [58,70]. In addition, delay embedding has been exten-
of training data [50]. sively employed in SSM-based model reduction from
In the last few years, spectral submanifolds (SSMs) data [16,17]. For SSMs, a closer understanding of the
have appeared as an alternative for model reduction in delay embedding map improves fits to data and pro-
intrinsically nonlinear systems. An SSM is the unique duces more accurate reduced-order models [5]. This
smoothest invariant manifold tangent to a nonresonant has motivated our present study on how invariant man-
spectral subspace emanating from a fixed point [14] ifolds can be efficiently and accurately reconstructed
or a periodic or quasiperiodic orbit [33]. Therefore, in delay coordinates.
an attracting SSM is the ideal candidate for a low- The main driver behind the introduction of delay
dimensional model of a nonlinear system [32,60]. Con- embedding as a tool in dynamical systems was the
cepts related to SSMs include nonlinear normal modes discovery that it could reconstruct strange attractors
(NNMs) defined either as sets of periodic motions in from scalar measurements of chaotic systems [20,54].
conservative systems [43,62,74] or invariant manifolds Floris Takens’s celebrated embedding theorem [72] and
[68,69] and invariant spectral foliations [71]. Here, we its later extension [65] show that, in principle, delay
will apply SSMs, as they are unique, exist under well- embedding recovers invariant sets from the full state
defined conditions in dissipative systems, can have space in a suitable observable space under generic
arbitrary dimensions, and can include internally res- assumptions. In practice, however, the choice of the
onant modes. timelag and embedding dimension is critical to obtain
After the computation of an SSM, we can project robust models [10,15,76]. The many methods for
either equations or data onto it to reduce the system choosing delay parameters for chaotic attractor recon-
to a high-fidelity low-dimensional model. Automated struction include minimization of the mutual informa-
model reduction to SSMs from equations [37] can suc- tion between subsequent samples [30], minimization
cessfully predict responses to small harmonic forcing of false nearest neighbors [1,42], and a Monte Carlo
[36,59,61] and bifurcations of those responses [47,48], decision tree search formulation [44].
and has also been extended to constrained mechani- Recent work has also explored the geometric struc-
cal systems [49]. Recently, Ref. [16] developed a data- ture of delay embedded invariant sets in an effort
driven method which identifies the SSM geometry and to improve model order reduction. For periodic data,
its reduced dynamics from trajectories in an observable singular value decomposition (SVD) on the delay-
space [18]. This approach also transforms the SSM- embedded snapshot matrix has been shown to con-
reduced dynamics to a normal form, which describes verge to a Fourier analysis [8]. The number of delays
the dynamics as sparsely as possible while maintaining required to recover such periodic orbits equals the num-
essential nonlinearities [31]. SSM-based model reduc- ber of coefficients of the Fourier spectrum [57]. Fitting
tion has since been applied to both numerical and a linear map between subsequent snapshots of such a
experimental datasets in fluid and structural dynam- delay-embedded periodic orbit produces a companion
ics [17,41] and control [3]. Reference [5] shows how matrix, whose eigenvectors are given by the inverse
to improve the computational efficiency of data-driven Vandermonde matrix [24,63].
SSM identification through a simplified formulation of Furthermore, connections to convolutional coordi-
the algorithm. nates [40] and the Frenet-Serret frame [35] have been
Delay embedding is the method of reconstructing made, and an interpretation of SVD modes in delay
invariant sets by viewing a select number of measure- coordinates as principal component trajectories has
ments separated by a timelag as independent observ- been proposed [26]. For the special case of an observed
ables. This method is routinely used to aid data- signal composed of oscillating sinusoidal functions, the
driven model identification in nonlinear dynamical sys- observable space contains invariant spaces determined
tems. Examples of model reduction methods based on by the signal frequencies [5]. Recently, it was shown
delay embedding include the extended dynamic mode that subsequent components of the DMD modes of
decomposition (DMD) [25,75], the Hankel alternative delay-embedded linear systems are related by a multi-
view of the Koopman decomposition (HAVOK) [11], plication of the corresponding eigenvalue [9].
the eigensystem realization algorithm (ERA) [39], clo-

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22081

In this work, we explore the local dynamics close ẋ = Ax + g(x), x ∈ Rn , g ∼ O(|x|2 ),

to a fixed point of a differentiable nonlinear delay- g : Rn → Rn . (1)
embedded system. We show that the linear part of the
delay-embedded dynamics depends solely on the corre- Let us denote the flow map of the system by
sponding eigenvalues, and not on the observable func- F t (x 0 ) := x(t, x 0 ), with x(t, x 0 ) denoting the trajec-
tion or the full state space eigenvectors. In particular, tory of (1) starting from x 0 at time 0. We assume that
the eigenvectors in the observable space are given by A ∈ Rn×n is diagonalizable and that the real parts of its
the columns of the Vandermonde matrix of the expo- eigenvalues are either all strictly negative or all strictly
nential of the eigenvalues multiplied by the timelag. positive. We take d eigenvectors of A and denote their
Unlike available previous work, we do not attempt a lin- span by E, i.e., a d-dimensional spectral subspace of
earization of the nonlinear dynamics, nor do we restrict Rn . In this step, we often choose the d slowest eigendi-
our attention to periodic orbits. Instead, our results rections.
imply that the nonlinear delay-embedded system has Provided that the d eigenvalues corresponding to E
an SSM whose tangent space coincides with the col- are non-resonant with the remaining n − d eigenvalues
umn space of this Vandermonde matrix. We exploit this of A, the nonlinear system has a unique smoothest,
structure to aid the data-driven identification of SSMs invariant manifold M tangent to E at the origin, i.e.,
in three mechanical examples. We believe that these T0 M = E [14]. Following [32], we call M a spectral
results enhance the understanding of delay embedding submanifold (SSM). In case of a resonance between E
in reduced-order modeling and also reveal new oppor- and the rest of the spectrum of A, the d-dimensional
tunities for SSM-based model reduction. SSM does not exist in general, and we must then include
The structure of this paper is the following. First, the resonant modal subspace into E to obtain a higher-
Sect. 2 briefly introduces SSM theory and summarizes dimensional SSM. If all eigenvalues of A are stable, the
a method for fast SSM-based data-driven modeling. slowest SSM attracts nearby trajectories, which makes
Section 3 outlines a new theory for delay-embedding it suitable for model order reduction.
tangent spaces of invariant manifolds and discusses The open-source numerical package SSMTool com-
their application to SSM-based model reduction. In putes SSMs from arbitrary finite-dimensional nonlin-
Sect. 4, we use these results to identify SSMs in exam- ear systems [36,37]. More recently, the SSMLearn
ples of a 2-degree-of-freedom oscillator, simulations package was developed to find SSMs in data from non-
of multimodal vibrations in a von Kármán beam, and linear dynamical systems [16,17]. Here, we will apply
experiments of complex behavior in a sloshing tank. the simplified data-driven SSM algorithm fastSSM
In Sect. 5, we draw conclusions from these examples introduced by Ref. [5].
and discuss possible further extensions of our theory.
Finally, Appendix 1 contains the proofs of the results
presented in Sect. 3. 2.2 Fast data-driven model order reduction to spectral
submanifolds

2 Model reduction to spectral submanifolds The objective of dynamics-based machine learning is

to reconstruct SSMs from data, and then use SSM-
Here, we outline previous results on rigorous model reduced models for predictions of the full system
order reduction to SSMs in smooth nonlinear systems. response [16]. Here, we use fastSSM [6] to identify
We also summarize fastSSM, the algorithm we use the SSM from snaphots of trajectories in an observable
here to identify SSMs from data. space. The procedure consists of two steps: manifold
geometry detection and normal form computation. The
summary below follows Ref. [5], to which we refer for
2.1 Spectral submanifold theory further details. Whereas that reference differentiates
between the algorithm for cubic polynomial approx-
Consider a nonlinear, autonomous dynamical system of imations of two-dimensional SSMs and its extension
class C l , l ∈ {N+ , ∞, a}, where a denotes analyticity, to arbitrary order and dimension, here, we will simply
in the form refer to both algorithms as fastSSM.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22082 J. Axås, G. Haller

The SSM is parametrized in the graph style, that into (6) to obtain
is, we construct M as a graph over the spectral sub-
Dζ (Hζ 1:h )Nζ 1:h = G(Hζ 1:h )1:r . (8)
space E. The data consists of snapshots y(ti ) ∈ R p in
a p-dimensional observable space. For each trajectory The matrices H and N are computed by solving (8)
we construct the snapshot matrix Y ∈ R p×N from N recursively at increasing orders with SSMTool [36].
snapshots as This procedure requires that the training data lies
⎡ ⎤ sufficiently close to the SSM, which can be achieved
| | |
Y = ⎣ y(t1 ) y(t2 ) . . . y(t N ) ⎦ (2) by removing initial transients from the input signal, as
| | | identified by a spectral analysis on the training data
[17]. Since the SSM built over the slowest d modes
Let T ∈ R p×d be a matrix whose columns approxi- is unique and attracting, this method ensures relevant
mately span the SSM tangent space. In fastSSM, the training data.
standard procedure is to obtain T through SVD on the
snapshot matrix Y . However, T can also be prescribed
if the tangent space is known a priori. Denoting by 3 Delay-embedding the tangent spaces of invariant
(·)† the Moore–Penrose pseudoinverse, we project each manifolds
snapshot yi onto this subspace to obtain d-dimensional
reduced coordinates ξ as Here, we show how tangent spaces of invariant man-
ξ = T y.†
(3) ifolds at a fixed point can be analytically recovered
when the observable space arises from delay embed-
We write Ξ ∈ Cd×N
for the projection of the snapshot ding of a signal. We also describe how the recovered
matrix onto the tangent space. tangent spaces facilitate the reconstruction of spectral
Next, we seek to approximate the embedding of M submanifolds in such observable spaces.
as the graph of a multivariate polynomial of order m
from the data:
y(ξ ) = Mξ 1:m , 3.1 Theoretical results
(4)
M = [M 1 , M 2 , . . . , M m ], M i ∈ R p×di ,
For the dynamical system (1), we define a scalar observ-
where di is the number of d-variate monomials at order able μ(x(t)), where μ : Rn → R is a differentiable
i and the superscript in (·)1:l denotes a vector of all function that returns a measured feature of system (1),
monomials from order 1 up to l. We obtain the man- such as a displacement coordinate. In order to recon-
ifold parametrization coefficients M ∈ R p×d1:m by a struct features of the full phase space from the observ-
polynomial regression, which yields the solution able, we use delay embedding. We stack p consecutive
M = Y (Ξ 1:m )† . (5) measurements separated by a timelag τ > 0 to create
an observable space of dimension p. This yields a tra-
The reduced dynamics are approximated by an O(r ) jectory in the form y(t) = S(x(t)) ∈ R p , where we
polynomial regression, with a coefficient matrix G ∈ define the sampling map
Cd×d1:r , in the form ⎡ ⎤
μ(x)
ξ̇ ≈ Gξ 1:r , G = Ξ̇ (Ξ 1:r )† . (6) ⎢ μ(F τ (x)) ⎥
⎢ ⎥
⎢ ⎥
S : Rn → R p , x → ⎢ μ(F (x)) ⎥ .
2τ
Finally, we compute the normal form [31] of the (9)
SSM-reduced dynamics up to order h. This amounts to ⎢ .. ⎥
⎣ . ⎦
a near-identity polynomial transformation with coeffi-
μ(F ( p−1)τ (x))
cients H ∈ Cd×d1:h from the new coordinates ζ ∈ Cd
such that An important question is how invariant sets of sys-
ξ = Hζ 1:h
= ζ + H 2:h ζ 2:h
, tem (1) in Rn are reproduced in the observable space
(7) R p . In particular, when the full state space trajectory
ζ̇ = Nζ 1:h
= Λζ + N 2:h ζ 2:h
. x(t) resides on a d-dimensional invariant manifold M,
The normal form and the reduced dynamics are con- will y(t) also do so? Takens’s embedding theorem gives
jugate dynamical systems. Therefore, we substitute (7) an affirmative answer. It states that if μ is generic and

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22083

no small integer multiple of τ coincides with the period manifold M̃ at the fixed point can be written
of any possible periodic orbit of (1) lying in M, then
Tq M̃ = range V . (13)
for
Proof See Appendix A.1.
p ≥ 2d + 1, (10)
This result is illustrated in Fig. 1. Note that the
the manifold M will have a diffeomorphic copy M̃
observable function must have full rank, as spelled out
in R p via the mapping (9) [72]. Whereas Takens’s the-
in the following remark.
orem was formulated only for scalar observable func-
∂μ
tions, this result has since been extended to multi- Remark 1 For (13) to hold, we must have ∂z k
|0 = 0
dimensional μ as long as the total observable space ∀k ∈ {1, . . . , d}, which defines the genericity of μ. If
dimension exceeds 2d [22]. the gradient of the observable function is orthogonal to
Both the nonlinear geometry and dynamics of M any of the eigenvectors e1 , . . . , ed , the sampling map
and the observable function influence the geometry of S will not be an embedding of M.
M̃. It is therefore difficult to predict its geometry for a
This should be kept in mind particularly when dealing
general flow map. Around the fixed point q = S(0) ∈
with symmetries of engineering structures, as we will
R p , however, the O(1) expansion of M̃, i.e., its tangent
show in our examples below.
space Tq M̃, can be directly determined, as we will
show next. Note that since the flow map is the identity Theorem 2 The columns of V are eigenvectors of the
at the origin, q lies on the diagonal in the observable linearized delay-embedded system at the fixed point.
space, with each of its identical components given by Indeed, the dynamics in the observable space can be
μ(0). written
We start by rewriting (1) in modal coordinates:
ẏ = V ΛV † ( y − q) + o(| y − q|) (14)
−1
ż = f (z) = Λz + E g(E z), (11)
Proof See Appendix A.2.
where E = [e1 , . . . , en ] contains the eigenvectors of A
and Λ = diag(λ1 , . . . , λn ) the corresponding eigenval- Corollary 1 In the observable space R p , the timelag
ues, which we assume to be distinct. We define modal τ and the eigenvalues λk fully determine the tangent
coordinates z ∈ Cn by letting z = E −1 x. Whereas the space and the linear part of the dynamics. In particular,
observable function is defined as a function of x, it is the linear dynamics are independent of both the full
notationally convenient to instead define it as a function eigenvectors and the observable function.
of z, as μ(x) = μ(E z). In the following, we will demonstrate how this struc-
Let M be a d-dimensional invariant manifold of (1) ture can be exploited for parametrizing spectral sub-
intersecting the origin 0 ∈ Rn , where it is tangent to manifolds from data, when the corresponding eigen-
a set of d eigenvectors e1 , e2 , . . . , ed of A with corre- values are approximately known.
sponding eigenvalues λ1 , . . . , λd . We define the Van- Finally, we treat the case of a multi-dimensional
dermonde matrix V ∈ C p×d of the d eigenvalues gov- observable function. Then, the tangent space is influ-
erning the linearized dynamics on M as V jk = eλk jτ , enced by the relative dependency of each component
i.e., μ of the observable function on each modal coordinate
⎡ ⎤
1 1 ... 1 zk .
⎢ eλ1 τ eλ2 τ ... eλd τ ⎥
⎢ 2λ τ ⎥ Theorem 3 For a multi-dimensional observable μ :
⎢ 2λ τ ... e dτ
2λ ⎥
V =⎢ e e 2
1
⎥. Rn → Rq with components μ1 , . . . , μq , the tangent
⎢ .. .. . . ⎥
⎣ . . . . .
. ⎦ space Tq M̃ ⊂ R pq at the fixed point q can be expressed
e ( p−1)λ 1 τ e ( p−1)λ 2 τ ... e ( p−1)λd τ as
⎡  ⎤
(12) V diag ∂μ ∂
1
⎢ z 0 ⎥
⎢ .. ⎥
Theorem 1 Under the assumptions of a generic observ- Tq M̃ = range ⎢ ⎥. (15)
⎣ . ⎦
able function μ : Rn → R and distinct eigenvalues ∂μq
V diag ∂ z
λ1 = · · · = λd , the tangent space of the observable 0

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22084 J. Axås, G. Haller

Fig. 1 Delay embedding of the tangent space T0 M of an invari- on the flow map, but its tangent space, Tq M̃, is directly given by
ant manifold M. The full state space manifold M (left) has a dif- the eigenvalues at the fixed point, independent of the geometry
feomorphic copy M̃ in the observable space (right) by Takens’s of M and the observable function μ
theorem. The shape of the reconstructed manifold M̃ depends

Proof See Appendix A.3. We have seen that the gradient of a multi-dimensional
observable function μ enters the expression for the tan-
When the observable function is a set of displace- gent space (15). While an expression for this gradient
ments, the linearized multi-dimensional observable is typically not available in experiments, mode shapes
function ∂μ
∂ z 0 corresponds to the mode shapes of the Ê = [ê1 , . . . , êd ] ∈ Rq×d are often known from the-
system in terms of those displacements. Therefore, if ory or obtained experimentally. Here, each mode shape
the mode shapes and eigenvalues of the observed sys- êk ∈ Rq describes how the eigenvector ek ∈ Rn is
tem are known, we can directly compute the tangent observed. Specifically, they are related by
space of M̃. In the special case of a scalar observable, ∂μ
the tangent space is independent of the observable func- êk = ck ek , (17)
∂x 0
tion and we do not need any information about the mode
where ck ∈ C is a nonzero constant that only rescales
shapes.
the eigenvectors. We select T as the columnwise Kro-
necker product of the Vandermonde matrix and the
observable mode shapes, i.e.,
3.2 Delay-embedded spectral submanifold ⎡ ⎤
reconstruction V diag ê1
⎢ .. ⎥
T := ⎣ . ⎦. (18)
These theoretical results can be exploited as a con- V diag êd
straint to aid SSM identification from data. In the case
of a scalar signal and with the eigenvalues of interest A sketch of the geometry is shown in Fig. 2. In the case
λ1 , . . . , λd approximately known, we select the matrix of unknown mode shapes, it may be possible to first
representation of the tangent space T appearing in (3) project low-amplitude data onto the delay-embedded
as the Vandermonde matrix (12), i.e., eigenvectors and then extract the observable mode
shapes via SVD, although we do not explore this idea
T := V . (16) further in this work.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22085

total delay embedding to be as small as possible. While

solving (19) gives some guidance, a suitable choice of τ
and p will also depend on the nonlinearity of the system
in the data range and the amount of signal noise.

4 Applications

We now apply our method to three datasets: two from

simulations and one from experiments. The eigenvalues
in these examples are known either from theory or sim-
ulations. We infer the delay-embedded tangent space
Fig. 2 Schematic illustration of a delay-embedded invariant accordingly before parametrizing the SSM. The exam-
manifold M̃ in the case of a q-dimensional observable function ples include an oscillator chain, a clamped-clamped
μ. The tangent space Tq M̃ is the range of the matrix T , defined as beam and tank sloshing.
the columnwise Kronecker product of the Vandermonde matrix
V and the mode shapes Ê in terms of the observable

4.1 Two-degree-of-freedom oscillator with nonlinear

Prescribing T and projecting the data onto its springs
columns yields modal reduced coordinates. This diag-
onalization of the system simplifies the learning of the Our first example is a simple 4-dimensional mechan-
geometry and the reduced dynamics of the SSM via the ical system, in which we will construct the slow 2-
algorithm outlined in Sect. 2.2. dimensional SSM in a delay-embedded space. This
Choosing proper delay-embedding parameters to examples serves as a simple illustration of the invari-
reconstruct nonlinear systems can be a challenge. For ance of the SSM tangent space, the meaning of generic-
the linear part of the system, however, our results ity in a specific case, and the handling of mode shapes
suggest picking the timelag τ and embedding dimen- for multi-dimensional observable functions.
sionality p so as to obtain numerically favorable We consider an oscillator chain of two masses, both
reduced coordinates along the SSM. We ideally want attached with linear springs to each other and to the
the columns of the Vandermonde matrix (12) to be ground. In addition, the spring connecting the left mass
orthogonal in order to maximize the signal-to-noise to the ground has a quadratic softening nonlinearity and
ratio in each of the observed modes. To this end, we the spring connecting the masses is of cubic hardening
formulate a minimization problem, type. The masses and linear spring stiffnesses are set
to 1, the softening parameter is −2 and the hardening
(κ  , p  ) = argmin V (κΔt, p) V (κΔt, p) − I (19)
, parameter is 1. Each of the springs also has a linear
κ, p∈N+ F
damping coefficient of 0.03. Figure 3a shows the con-
in which the columns of V are normalized and  · F figuration. The sampling time is Δt = 0.1 s.
denotes the Frobenius norm. Since the timelag is an We compute an initial condition on the slow 2D SSM
integer multiple of the sampling timestep, τ = κΔt, for the single training trajectory using SSMTool. The
(19) defines an optimization over a set of discrete vari- trajectory in the full phase space and the SSM are shown
ables which can be solved simply by brute force. in Fig. 3b. The first two eigenvectors span the tangent
Bearing in mind the nonlinear part of the system, space of the SSM.
however, an optimal choice of delay parameters is not as Next, we delay embed the trajectory with a time-
straightforward. Increasing the timelag and embedding lag τ = 15Δt and embedding dimension p = 5.
dimension tends to curve the SSM, requiring higher Our observable function is the first mass displacement,
orders of approximation and in extreme cases folding μ(x) = x1 . We obtain reduced coordinates by projec-
the manifold, so that it can no longer be parametrized tion of the trajectory data onto the columns of the Van-
as a graph. Taking into account the nonlinear part of the dermonde matrix V as predicted by our theory. Next,
SSM reconstruction, therefore, we typically want the we seek the 2D SSM geometry in this observable space

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22086 J. Axås, G. Haller

delay-embedded SSM for various observable func-

tions, μ(x) = x1 (Fig. 3c), μ(x) = x2 (Fig. 4a), and
μ(x) = ẋ1 + ẋ2 (Fig. 4b). These different observable
functions clearly produce different SSM geometries,
but the eigenvectors and tangent spaces of the mani-
folds all agree.
One exception is when we observe the distance
between the masses, μ(x) = x2 − x1 (Fig. 4c). In this
case, the delay-embedded trajectory no longer lies on
an invariant manifold, as is evident by the nonsmooth
cusp in the data at the origin. The reason is that this
observable is non-generic precisely in the sense of our
theory; it coincides with the mode shape of the sec-
ond, fast mode of the full system. This means that the
observable function acts orthogonally to the slow SSM
at the fixed point and thus the delay mapping is not an
embedding, by Remark 1. In practical terms, the slow
SSM is the nonlinear continuation of the first vibra-
tion mode, which in this example corresponds to both
masses moving in unison. During such a movement,
the distance between the masses is constant, and so if
we only observe x2 − x1 , we do not obtain any infor-
mation about the ongoing dynamics, and thus cannot
reproduce the geometry of the SSM.
Next, we pick μ(x) = x2 and use fastSSM to
obtain a model of the SSM dynamics. We approximate
the reduced dynamics as a cubic order polynomial and
compute its normal form as
   
ρ̇1 −0.0014 ρ1 3 − 0.0148 ρ1
= . (20)
θ̇1 1.0025 − 0.0919 ρ1 2
The trajectory projected onto the columns of V is
shown in Fig. 5a. Integrating the obtained normal form
and mapping back to the observable space yields a good
reconstruction of the training data, as shown in Fig. 5b.
Finally, following Theorem 3, we demonstrate how
Fig. 3 a Setup for the two-degree-of-freedom oscillator example
with two nonlinear springs. b The slow 2D SSM (gray) in the to determine the tangent space of the SSM at the fixed
full state space, along with its tangent space (red). c The delay- point when the observable is a vector. When we choose
embedded SSM in the observable space. (Color figure online) μ(x) = [x1 , x2 ] , unlike for a scalar observable func-
tion, the tangent space orientation is influenced not only
by the eigenvalues, but also by the shape of the first
mode. Recall that the first mode shape corresponds to
as a 5th order polynomial in the projected coordinates
the masses moving in unison, i.e.,
using fastSSM. Figure 3c shows the SSM in the first  
1
three coordinates of the observable space. Indeed, the ê1 = ê2 = . (21)
tangent space of this observable space is identical to 1
the column space of V . Then, by (18), we obtain vectors spanning the tangent
Corollary 1 predicts that this tangent space will space as the columns of the matrix
   
be independent of the observable function, provided V diag(ê1 ) V
T= = , (22)
that it is generic. To illustrate this, we plot the V diag(ê2 ) V

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22087

Fig. 4 a, b Changing the observable function leads to different

SSM geometries, but the tangent space remains the same as in Fig. 5 a Projection of the data onto the delay-embedded tangent
Fig. 3c. c A nongeneric observable function however, observing space predicted by our theory. b fastSSM predicts a model that
only the second mode, does not embed the manifold successfully reconstructs the decay of the trajectory. c A view of
the SSM in a delay-embedded space from a multi-dimensional
observable, with the tangent space predicted by our theory

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22088 J. Axås, G. Haller

where V is the Vandermonde matrix (12). A view of

the SSM and its tangent space in this 10-dimensional
observable space is shown in Fig. 5c.
The relation of this mode shape to the observable
function is given by (17). In particular, we can compute
the derivative of the observable function with respect
to the modal coordinates as
   
∂μ1 ∂μ1
∂z 1 (0) ∂z 2 (0) c1 c2
∂μ2 ∂μ2 = , (23)
∂z 1 (0) ∂z 2 (0)
c1 c2

where c1 , c2 ∈ C are nonzero constants depending on

the scaling of the eigenvectors. For simplicity, in (21)
we chose c1 = c2 = 1, such that T is the Vandermonde
matrix vertically stacked twice.

Fig. 6 von Kármán beam: schematic first, second, and third

mode shapes. The scalar observable must be generic in the sense
4.2 6D SSM in a nonlinear finite-element model of a that it must have contributions from all modes of interest. For
beam instance, the midpoint displacement is not excited by the second
mode. Instead, we choose the shown transverse displacement at
1/4 of the beam length
We train an SSM-reduced model with data from numer-
ical simulations of a finite-element (FE) representation
of a clamped-clamped von Kármán nonlinear beam 50 mm, and thickness 20 mm. The sampling time is
[38]. This example was previously studied in Refs. Δt = 0.0955 s.
[5,16], which identified the slowest 2D SSM in the In order for the sampling map to be an embedding,
delay-embedded observable space, predicted the forced the observable function must be generic. By Remark
response and analyzed the radius of convergence of the 1, for our purposes this implies that the observable
analytical normal form. Here, thanks to our results on function μ must have significant contributions from all
the tangent spaces of delay-embedded SSMs, we can modes z k that we wish to model. In this example, our
extend the analysis to the six-dimensional SSM ema- theory allows us to define the genericity and therefore
nating from the three slowest modes of the linear part suitability of a particular observable. For example, the
of the system. midpoint displacement chosen as observable function
Each node in the FE model has three degrees of in Refs. [5,16] was sufficient to model the 2D SSM,
freedom: axial deformation u, transverse deflection w, but cannot be employed for higher-dimensional SSMs.
and rotation w  . The von Kármán axial strain is given This is because the antisymmetric shape of the sec-
by ond mode has zero displacement at the midpoint (see
Fig. 6), i.e.,
1 
= u  (x) + − zw  (x).
2
11 w (x) (24) ∂μ ∂μ
2 (0) = (0) = 0. (26)
∂z 3 ∂z 4
The axial stress is given by
Instead, we choose the transverse displacement of the
σ =E 11 + c ˙11 , (25) beam at one fourth of the total length, μ = w(l/4), as
this degree of freedom has nonzero contributions from
where E = 70 GPa denotes the Young’s modulus all three mode shapes.
and c = 1.0 × 106 Pa · s the material rate of vis- For our data-driven modeling objectives, we need
cous damping. Based on a convergence analysis, we set training data containing the first three modes. To gener-
the number of elements to 12, resulting in a 33-degree ate initial conditions for such trajectories, we use linear
of freedom mechanical system, i.e., a 66-dimensional combinations of the mode shapes of the system com-
phase space. We set the beam length to 1000 mm, width puted from its linear part. Since the SSM is normally

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22089

attracting, these trajectories will quickly approach it

and we can use them to train our reduced-order model.
With this method, we produce three trajectories close to
the 6D SSM with different initial conditions, of which
we use two as training data and one as test data. For
validation purposes, we also pick the individual mode
shapes as initial conditions and use as test data. The
individual modal contributions in these initial condi-
tions were chosen as follows:

Initial Mode Type

Condition 1 2 3

1 1 0 0 Test
2 0 1 0 Test
3 0 0 1 Test
4 0.8 −0.8 0.8 Train
5 −0.1 0.8 0.8 Train
6 −0.6 −0.2 −0.8 Test

We choose the delay embedding parameters guided

by the observations in Sect. 3.2. Setting κ = 1 such
that the timelag τ = Δt and the embedding dimension
to p = 50 gives a local optimum of the function (19)
with the computed eigenvalues, while still keeping the
maximal delay κ p moderate to prevent folding of the
embedding.
Figure 7a shows the delay embedding of the single-
mode trajectories 1–3, corresponding to the first three
modes, in three of the 50 delay coordinates. These tra-
jectories visualize the orientations of the correspond-
ing eigenspaces in the observable space. Indeed, min-
imization of (19) corresponds to making these planes
orthogonal, simplifying their identification.
Figure 7b similarly displays the delay embedding of
the first training trajectory along with a visualization of
the columns of the Vandermonde matrix as vectors. Our
delay theory predicts that projection of the data onto
these vectors yields modal coordinates, as verified in
Fig. 7c. This space will serve as the reduced coordinates
of the SSM.
We project onto these eigenvectors and use fastSSM
to approximate the geometry of the 6D SSM with a
third-order polynomial. For the reduced dynamics, we
Fig. 7 a The trajectories with single modal contributions visual-
use a third-order approximation. We compute the nor- ize the modal subspaces in the delay-embedded space for the von
mal form of this reduced dynamics up to seventh order Kármán beam. The third mode data has been scaled by a factor
for our dynamics model. The terms up to third order in 2 to increase visibility. b The same delay-embedded view of the
first training trajectory, along with the delay-embedded eigenvec-
polar form are found by fastSSM to be of the form
tors. c After projection of this trajectory onto the eigenvectors,
the modal structure becomes clear
123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22090 J. Axås, G. Haller

⎛ ⎞ ⎛ ⎞
ρ̇1 0.3058 ρ1 3 + 2.088ρ1 ρ2 2 − 3.091ρ1
⎜ ρ1 θ̇1 ⎟ ⎜ 102.0 ρ1 3 + 82.70 ρ2 2 ρ1 + 657.2ρ1 ⎟
⎜ ⎟ ⎜ ⎟
⎜ ρ̇2 ⎟ ⎜ −2.705 ρ1 2 ρ2 + 1.723 ρ2 3 − 23.72ρ2 ⎟
⎜ ⎟=⎜ ⎟
⎜ ρ2 θ̇2 ⎟ ⎜ 95.64 ρ1 2 ρ2 + 115.6 ρ2 3 + 1812ρ2 ⎟
⎜ ⎟ ⎜ ⎟
⎝ ρ̇3 ⎠ ⎝ −8.968ρ3 ρ1 2 − 13.27ρ3 ρ2 2 − 88.47ρ3 ⎠
ρ3 θ̇3 115.9 ρ1 2 ρ3 + 85.04 ρ2 2 ρ3 + 3558ρ3
+O(5).
(27)
We transform the initial conditions from the observ-
able space to the normal form and integrate our model to
predict signal decay. This produces a normalized mean
trajectory error (as defined in [16]) of 2.2 % on the test
data. The predictions for trajectories 5 and 6 are shown
in Fig. 8a, b, while Fig. 8c shows the trajectories in
the normal form amplitude coordinates. In this figure,
trajectories 1, 2, and 3 lie close to the axes, since their
initial conditions lie along the respective modal coor-
dinates. Similarly, for example, trajectory 6 is close to
the ρ1 –ρ3 plane, having been initialized mostly with
contributions from modes 1 and 3.
We also visualize our reduced-order model (27) by
plotting the instantaneous frequency and damping as
predicted by the normal form (27) for varying ampli-
tudes of mode 1 and 2. For instance, our model predicts
hardening of the first mode with respect to both the first
and the second modal amplitudes (Fig. 9a), a decrease
in the instantaneous damping of mode 1 with respect to
mode 2 (Fig. 9b), and independence of the third instan-
taneous frequency with respect to itself (Fig. 9c). The
predictions for each of the trajectories are included for
reference.

4.3 Multimodal sloshing of water in a tank

For our final example, we apply our results to sloshing

experiments. Sloshing models have a wide range of
industrial applications, including fluid container inter-
action with ship motion [52], road transportation of flu-
ids [29], damping devices in towers [34], and fuel tank
design in spacecraft [2,23]. A tank partially filled with
water exhibits several nonlinear phenomena under hori-
zontal harmonic excitation [73]. On the one hand, inten-
sified fluid motion can alter the instantaneous damping
Fig. 8 a, b Predictions from fastSSM for the decaying trajec-
and frequency of the first sloshing mode [27]. On the tories 5 and 6 for the von Kármán beam c Phase portrait of the
other hand, increasing the amplitude further activates trajectories after transformation to the normal form
several nonlinearly coupled modes of the system and
gives rise to a range of different wave motions [28,53].

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22091

Fig. 10 a Experimental setup for tank sloshing (adapted from

[13]) b The first four sloshing mode shapes

Our training data comes from experiments described

in Ref. [13] with a rectangular tank of width w = 500
mm and thickness 50 mm, partially filled with water up
to a height of h = 400 mm. The tank was attached to a
horizontally moving platform harmonically excited by
a motor at different frequencies. Then, once the sys-
tem had reached a steady state, the motor was turned
off. Depending on the forcing frequency, this periodic
response exhibited planar, wave-breaking, or three-
periodic motion. The three-periodic forced state was
characterized by an increase in the response ampli-
tude every third forcing cycle, while the wave-breaking
response was defined as overturning of the water close
to the walls [13]. A camera detected the surface pro-
file h with the sampling time Δt = 0.01 s. Figure 10a
displays the experimental setup.
Fig. 9 Visualization of the normal form (27) for the von Kármán
beam with the trajectories for a instantaneous frequency and b
damping of mode 1, as well as c frequency of mode 3

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22092 J. Axås, G. Haller

While previous work successfully captured the

dynamics of the main sloshing mode using a 2D SSM
for the center of mass signal [16] and the full surface
profile [5], here, we model the decay from a multi-
modal state by identifying a 6D SSM, corresponding
to the nonlinear extension of the three dominant oscilla-
tory modes. We train on three decaying measurements:
trajectories 1 and 2 start at a three-periodic state, and
trajectory 3 starts at a wave-breaking state.
The observable vector μ is the surface profile mea-
sured at 1 771 points along the tank width. Since this
function is multi-dimensional, in order to apply our
theory on delay-embedded tangent spaces, we need an
estimate of the eigenvalues and linear mode shapes in
our observable. The eigenfrequencies can be computed
from potential theory [27] as
  
gπ h
ωk = k tanh π k , (28)
w w
which scales approximately with the square root of the
mode number k for our configuration. The first five
eigenfrequencies are [7.80, 11.1, 13.6, 15.7, 17.6] rad/s,
with an approximate 1:2 resonance between frequen-
cies 1 and 4. The mode shapes by the same theory are
êk = cos(kx/w), x ∈ [0, w], (29)
shown in Fig. 10b. For the tangent space, in principle,
we also need the linear damping of each mode. In prac-
tice, this real part of the eigenvalues has very little influ-
ence on V for limited delay embedding and we pick the
values [−0.05, −0.07, −0.08, −0.09, −0.1] based on
previous fits of the first mode and the assumption of
increasing damping with the mode number.
Based on (19), we delay-embed the data with timelag
τ = 5Δt and dimension p = 47. A projection of the
delay-embedded data onto the eigenvectors T predicted
by our theory appears to yield modal coordinates, as
indicated in Fig. 11a.
Consequently, the norm of these projections can be
used as a heuristic measure of the modal content in
the signal. This procedure should be used with caution,
since it does not take manifold curvature into account,
but it can be employed to provide an initial guess for
the SSM dimension. In Fig. 11b, we plot the absolute Fig. 11 a Projecting one of the sloshing trajectories onto the
value of the projection onto each modal subspace of T tangent space vectors unveils the modal structure. b By projecting
the trajectory onto the eigenvectors and taking the absolute value,
over time for trajectory 2. This plot shows that the first we can estimate the relative modal contributions in the signal c A
mode dominates, while the zoomed-in view (Fig. 11c) zoomed-in view of b indicates that modes 1, 2, and 4 dominate
indicates that the second and fourth modes appear to be
the most prevalent of the higher modes throughout the

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22093

decay. The third and fifth mode are present at first but
quickly die out. All amplitudes are decaying except for
the fourth mode, which instead initially grows. Based
on this analysis, we will identify a 6D SSM emanating
from the spectral subspace of modes 1, 2, and 4. This
choice also takes SSM theory into account, by which
the 1:2 resonance requires that the modal subspace of
the fourth mode is included in the spectral subspace
of the SSM. We choose to start our training data after
1.2 s, as the third and fifth modal amplitudes are small
thereafter and we expect the trajectory to lie sufficiently
close to the SSM.
With an SSM parametrization order m = 4, reduced
dynamics order r = 3, and normal form order h = 3,
we compute the SSM geometry and dynamics and inte-
grate our reduced-order model to predict the decay from
the various flow states. This yields a normalized mean
trajectory error (NMTE) [16] of 2.6 %. fastSSM suc-
cessfully detects and accounts for the internal reso-
nance by adding phase-dependent terms to the com-
puted normal form, which reads
ρ̇1 2 2 2
ρ1 = −0.056 − 0.0069 sin(ψ − 0.26)ρ4 − 0.0015ρ4 − 0.039ρ2 + 0.023ρ1
θ̇1 = 7.78 + 0.0069 cos(ψ − 0.26)ρ4 + 0.040ρ42 + 0.016ρ22 − 0.82ρ12
ρ̇2 2 2 2
ρ2 = −0.13 + 0.15ρ4 − 0.89ρ2 + 0.37ρ1
θ̇2 = 11.4 + 0.57ρ42 − 0.0085ρ22 − 2.2ρ12
ρ̇4 2 2 2 2
ρ4 = −0.30 − 0.29ρ4 + 0.67ρ2 − 0.27 sin(ψ + 1.4)ρ1 /ρ4 + 1.2ρ1
2 2 2
θ̇4 = 15.9 − 0.085ρ4 + 1.2ρ2 + 0.27 cos(ψ + 1.4)ρ1 /ρ4 − 2.0ρ1 2

(30)
where ψ = θ4 − 2θ1 and the subscripts denote the
corresponding mode number. Looking at the linear part,
we see that the eigenfrequencies are well captured.
Good agreement between the experimentally mea-
sured surface profile elevation at the tank’s leftmost
point and the delay-embedded SSM-reduced prediction
is shown for the first period-three initial state in Fig. 12a
and the wave-breaking state in Fig. 12b. Further, our 6D
reduced model can accurately predict the full surface
profile decay, with snapshots shown in Fig. 13.
We project the training trajectories onto the SSM and
transform them to the normal form in polar coordinates.
The development of the amplitudes in the normal form
are shown in Fig. 12c for each trajectory. This plot
suggests that (i) the wave-breaking motion (Traj. 3)
does not seem to have any significant content of the
second mode, (ii) the amplitude of the fourth mode Fig. 12 Tank sloshing: the prediction on the 6D SSM for the
decay of a trajectory 1 and b trajectory 3. c Phase portrait of the
indeed increases after motor detachment, as indicated amplitudes of the normal form coordinates on the SSM for each of
by the increase in the ρ4 component of trajectories 1 the trajectories shows the modal contributions and development
and 2, (iii) there is a small oscillation in these signals for different initial flow states
not captured by our model, which may be due to noise,

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22094 J. Axås, G. Haller

Fig. 13 The experimentally measured surface profile decay agrees with our 6D SSM model prediction for trajectory 2

insufficient separation of the modal subspaces, a mode 5 Conclusions

outside our model, or some other phenomenon.
We note that the combined higher modal content in We have shown that for a scalar observation of an
the signal is small—only about 10% with respect to the invariant manifold tangent to a spectral subspace at a
first mode. This is because the data is decaying from fixed point, the delay-embedded reconstruction of the
steady states induced by forcing near the first eigenfre- tangent space is dependent only on the corresponding
quency. Due to their symmetric shape, isolated forcing eigenvalues of the full system linearized at that point.
of the second and fourth modes is not possible with hor- In particular, we have proven that the columns of a
izontal harmonic excitation. The small activation of the Vandermonde matrix, given by repeated multiplication
higher modes implies that their fitting is relatively more of the exponential of the eigenvalues times the timelag,
sensitive to noise. Nevertheless, we are able to capture are eigenvectors for the linearized system in the observ-
these smaller oscillations on the SSM. The key technol- able space. Therefore, the Vandermonde matrix diago-
ogy allowing this enhancement is the enforcement of nalizes the linear part of the delay-embedded dynam-
the delay-embedded tangent space in our SSM recon- ics. We have also shown that when several quantities
struction, based on the theoretical eigenfrequencies and are measured and delay-embedded simultaneously, the
mode shapes. tangent space can be expressed given the Vandermonde
Our model is stable with respect to changes in start- matrix and the mode shapes expressed in the observable
ing time, manifold order, and delay parameters, but may function components. These results hold for any invari-
have dynamical features that are structurally unstable ant manifold tangent to a modal subspace with distinct
at a given, lower order of approximation, i.e., might eigenvalues, including, e.g., classic stable manifolds.
change under the addition of higher-order terms. It is Here, our focus was the application of this result to
also possible to neglect the relatively minor influence spectral submanifolds of hyperbolic fixed points.
of the fourth mode and instead fit a 4D SSM tangent In an attempt to exploit this uncovered structure, we
to the first and second modal subspaces. Since such have shown that for data-driven SSM model reduction,
a lower-dimensional model requires identification of when the eigenvalues are approximately known, we can
fewer coefficients, it is more robust with regard to noise analytically predict the tangent space of the embedded
and parameters, at the cost of a slightly larger average SSM a priori to achieve local modal decomposition and
NMTE of 3.9 %. Here, however, since our objective aid the reconstruction of the nonlinear reduced dynam-
was modal analysis of different flow states, we chose ics. We have found that even for small activation of
the more detailed 6D model, guided by the SSM res- higher modes, this trick helps modeling complex mul-
onance condition. In practice, this choice will depend timodal nonlinear dynamics on an SSM, which in turn
on the intended application of the model. allows for analysis of modal energy interchange and
instantaneous frequencies.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22095

Our theory assumes a generic observable function, grants, or other support were received during the preparation of
which we describe in more detail in our first and second this manuscript.
example, and distinct eigenvalues. While the second Data availability The data supporting the findings of this
assumption is a generic one in a mathematical sense, it work can be openly accessed at www.github.com/haller-group/
is not always satisfied for engineering structures with SSMLearn/tree/main/fastSSM
symmetry. Vibrations in a square plate is an example
where our theory would fail, as it has repeated eigen- Declarations
values. Using a vector-valued observable may help in Conflict of interest The authors have no relevant financial or
differentiating between the modes in such a case. Fur- non-financial interests to disclose.
ther, while technically covered by the theory, possible
Open Access This article is licensed under a Creative Com-
practical difficulties related to the conditioning of the
mons Attribution 4.0 International License, which permits use,
Vandermonde matrix include highly different or very sharing, adaptation, distribution and reproduction in any medium
similar eigenvalues, or eigenvalues of different stability or format, as long as you give appropriate credit to the original
type. author(s) and the source, provide a link to the Creative Com-
mons licence, and indicate if changes were made. The images or
In our third example with data from experiments,
other third party material in this article are included in the article’s
we also devised a new heuristic scheme for using delay Creative Commons licence, unless indicated otherwise in a credit
embedding to study modal contents in a signal. With line to the material. If material is not included in the article’s Cre-
this method, that served as an initial guess, we pro- ative Commons licence and your intended use is not permitted by
statutory regulation or exceeds the permitted use, you will need
jected the data onto the respective prescribed modal
to obtain permission directly from the copyright holder. To view
subspaces, thereby implicitly assuming that the SSM a copy of this licence, visit http://creativecommons.org/licenses/
of each mode is nearly flat. An interesting develop- by/4.0/.
ment of this idea would be its use as a filter, which
could remove or keep certain frequencies in a signal.
Another idea would be to use the tangent space con-
dition as a verification or for iterative adjustment of A Appendix
the linear fit of the reduced dynamics. Finally, in anal-
ogy with a Fourier analysis, it would be possible to A.1 Proof of Theorem 1
estimate both instantaneous frequency and damping of
a signal by singular value decomposition of the tra- Let M be a d-dimensional invariant manifold of (1)
jectory in delay coordinates followed by analysis of containing the origin of Rn . The tangent space of M at
the Vandermonde matrix columns. This ties in with the origin can be written T0 M = span {ek }k∈K , where
several other observations made in the literature; for K ⊂ {1, . . . , n} is an index set labeling the d eigenvec-
example, for a linear system, these columns agree with tors ek spanning the spectral subspace from which the
the recently proposed notion of principal component manifold is emanating. For example, for a stable man-
trajectories [26]. Overall, we believe that our findings ifold, K = {k : Re λk < 0}. We also assume that the
shed more light on delay-embedding invariant mani- eigenvalues in question are distinct, i = k ⇔ λi = λk ,
folds and selecting delay parameters in particular. For for i, k ∈ K .
that reason, we expect these results to be of use for a To simplify the notation, we transform the full state
wide range of data-driven methods. space to modal coordinates (11). We rewrite the observ-
able function on the system x ∈ Rn as an observable on
Acknowledgements We are grateful to Kerstin Avila and Bas- the modal coordinate system z ∈ Cn as μ̂(z) = μ(E z).
tian Bäuerlein (U. Bremen) for sharing their experimental surface We denote by Φ t = E ◦ F t ◦ E −1 the flow in Cn . Con-
profile data from Ref. [13] with us.
sider the sampling map in modal coordinates
Author contributions JA and GH designed the research. JA ⎡ ⎤
carried out the research. JA wrote the paper. GH edited the paper
μ̂
⎢ μ̂ ◦ Φ τ ⎥
and lead the research team. ⎢ ⎥
⎢ 2τ ⎥
Ŝ = ⎢ μ̂ ◦ Φ ⎥ : Cn → R p , y = Ŝ(z). (31)
Funding Open access funding provided by Swiss Federal Insti- ⎢ .. ⎥
tute of Technology Zurich. The authors declare that no funds, ⎣ . ⎦
μ̂ ◦ Φ ( p−1)τ

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22096 J. Axås, G. Haller

Under the conditions of Takens’s embedding theorem, We compute the ODE on M̃, ẏ = f̃ ( y), as
the delay embedding map Ψ = Ŝ|M : M → M̃ is a d t
smooth embedding with a smooth inverse Ψ −1 : M̃ → f̃ = Φ̃ = DΨ ◦ f ◦ Ψ −1 , (36)
dt
M, and Ψ (0) = q.
where f is given by (11). The derivative of f̃ at the
In order to derive the tangent space Tq M̃, we com-
fixed point is, therefore,
pute the derivative of the embedding at 0:
⎡ ⎤ D f̃ (q) = DΨ (0) ◦ Λ ◦ DΨ −1 (q)
Dμ̂(0)
   −1
⎢ Dμ̂(Φ τ (0)) ◦ DΦ τ (0) ⎥ ∂ μ̂ ∂ μ̂
⎢ ⎥ = V diag Λ diag V †(37)
DΨ (0) = ⎢ .. ⎥ ∂z 0 ∂z 0
⎣ . ⎦
Dμ̂(Φ ( p−1)τ (0)) ◦ DΦ ( p−1)τ (0) = V ΛV † ,
⎡ ⎤ (32)
Dμ̂(0) where we used the commutative property of multipli-
⎢ Dμ̂(0) ◦ eΛτ ⎥ cation of diagonal matrices, which eliminates the lin-
⎢ ⎥
=⎢ .. ⎥. earized observable function terms, and the fact that
⎣ . ⎦ −1
DΨ −1 (q) = diag ∂∂μ̂z V † is well-defined. To
Dμ̂(0) ◦ eΛ( p−1)τ 0
see this, note that the derivative of the delay embed-
Now note that the jth component expressed in modal
ding map composed with its inverse
coordinates is
Dμ̂(0) ◦ eΛ jτ (z) DΨ (0) ◦ DΨ −1 (q)
   −1
 ∂ μ̂ ∂ μ̂ ∂ μ̂
= eλk jτ z k , j ∈ {1, . . . , p}. (33) = V diag diag V† (38)
∂z k 0 ∂z 0 ∂z 0
k∈K
We define the Vandermonde matrix V of the d eigen- = V V †,
values {λk }k∈K governing the linearized dynamics on maps all points in T0 M to themselves, since V has
M as V jk = eλk jτ . We conclude that the tangent space full rank under the assumption of distinct eigenvalues
of the observable manifold at the fixed point in modal {λk }k∈K .
coordinates can be written as Taylor-expanding the ODE on M̃ in the observable
Tq M̃ = { DΨ (0)z, z ∈ Cn } space yields
  
∂ μ̂ (34) ẏ = D f̃ (q)( y − q) + o(| y − q|) = V ΛV † ( y − q)
= range V diag = range V ,
∂z 0 +o(| y − q|). (39)
where the diagonal matrix acts only as a rescaling of
Therefore, under the assumptions of a generic observ-
each component of z. Therefore, one matrix representa-
able function and distinct eigenvalues, the tangent
tion of the tangent space of the manifold in the observ-
space Tq M̃ and the linearized dynamics D f̃ (q) in the
able space is V , which is independent both of the matrix
observable space R p are both fully determined by the
E of full system eigenvectors and the observable func-
timelag τ and the eigenvalues λk , k ∈ K .
tion μ.
∂ μ̂ In the special case that M = Rn , if p ≥ 2n +
Note that we must have ∂z |0 = 0 ∀k ∈ K , which
k 1, the entire phase space can be reconstructed. For a
defines the genericity of μ. In practice, this implies
linear system, p = n suffices, and the delay embedding
that the linearized observable function must contain
reduces to a linear operator.
contributions from all modal coordinates that we wish
to model. In addition, note that for the embedding of
the tangent space itself, i.e., the linear system, p = d A.3 Proof of Theorem 3
suffices.
For a vector-valued observable, μ : Rn → Rq , the
delay embedding map reads
A.2 Proof of Theorem 2 ⎡ ⎤
Ψ μ1
⎢ ⎥
The flow on M̃ is Ψ = ⎣ ... ⎦ : M → M̃ ⊂ R pq ,
t
Φ̃ = Ψ ◦ Φ t ◦ Ψ −1 . (35) Ψ μq

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22097

⎡ ⎤
DΨ μ1 (0) 9. Bronstein, E., Wiegner, A., Shilo, D., Talmon, R.: The spa-
⎢ .. ⎥ tiotemporal coupling in delay-coordinates dynamic mode
DΨ (0) = ⎣ . ⎦, (40)
decomposition. Chaos An Interdiscip. J. Nonlinear Sci.
DΨ μq (0) 32(12), 123127 (2022). https://doi.org/10.1063/5.0123101
10. Broomhead, D., King, G.P.: Extracting qualitative dynam-
where Ψ μ is the delay embedding map corresponding ics from experimental data. Phys. D Nonlinear Phe-
to the th component of the observable function μ. The nom. 20(2), 217–236 (1986). https://doi.org/10.1016/
derivatives are given by 0167-2789(86)90031-X
  11. Brunton, S.L., Brunton, B.W., Proctor, J.L., Kaiser, E.,
∂ μ̂
DΨ μ (0) = V diag . (41) Kutz, J.N.: Chaos as an intermittently forced linear
∂z 0 system. Nat. Commun. (2017). https://doi.org/10.1038/
s41467-017-00030-8
In this case, the tangent space is not independent of the
12. Brunton, S.L., Proctor, J.L., Kutz, J.N.: Discovering govern-
observable function. Instead, it is affected by the rela- ing equations from data by sparse identification of nonlinear
tive dependency of each component μ of the observ- dynamical systems. Proc. Natl. Acad. Sci. 113(15), 3932–
able function on each modal coordinate z k . The tangent 3937 (2016)
13. Bäuerlein, B., Avila, K.: Phase lag predicts nonlinear
space can, however, be expressed as
⎡  ⎤ response maxima in liquid-sloshing experiments. J. Fluid
V diag ∂∂μ̂z1 Mech. 925, A22 (2021). https://doi.org/10.1017/jfm.2021.
⎢ 0 ⎥ 576
⎢ .. ⎥
Tq M̃ = range ⎢ . ⎥. (42) 14. Cabré, X., Fontich, E., de la Llave, R.: The parameterization
⎣  ⎦ method for invariant manifolds I: manifolds associated to
∂ μ̂q
V diag ∂ z non-resonant subspaces. Indiana Univ. Math. J. 52(2), 283–
0 328 (2003)
15. Casdagli, M., Eubank, S., Farmer, J., Gibson, J.: State space
reconstruction in the presence of noise. Phys. D Nonlin-
References ear Phenom. 51(1), 52–98 (1991). https://doi.org/10.1016/
0167-2789(91)90222-U
1. Abarbanel, H.D.I., Kennel, M.B.: Local false nearest neigh- 16. Cenedese, M., Axås, J., Bäuerlein, B., Avila, K., Haller,
bors and dynamical dimensions from observed chaotic data. G.: Data-driven modeling and prediction of non-linearizable
Phys. Rev. E 47, 3057–3068 (1993). https://doi.org/10. dynamics via spectral submanifolds. Nat. Commun. (2022).
1103/PhysRevE.47.3057 https://doi.org/10.1038/s41467-022-28518-y
2. Abramson, H. (ed.): The dynamic behavior of liquids in 17. Cenedese, M., Axås, J., Yang, H., Eriten, M., Haller, G.:
moving containers: with applications to space vehicle tech- Data-driven nonlinear model reduction to spectral submani-
nology/edited by H. Norman Abramson. NASA SP-106. folds in mechanical systems. Philos. Trans. R. Soc. A Math.
Scientific and Technical Information Division, National Phys. Eng. Sci. (2022). https://doi.org/10.1098/rsta.2021.
Aeronautics and Space Administration, Washington, D.C 0194
(1966) 18. Cenedese, M., Axås, J., Haller, G.: SSMLearn (2021). http://
3. Alora, J.I., Cenedese, M., Schmerling, E., Haller, G., www.georgehaller.com
Pavone, M.: Data-driven spectral submanifold reduction 19. Chen, S., Billings, S.A.: Neural networks for nonlin-
for nonlinear optimal control of high-dimensional robots ear dynamic system modelling and identification. Int. J.
(2022). https://doi.org/10.48550/ARXIV.2209.05712 Control 56(2), 319–346 (1992). https://doi.org/10.1080/
4. Awrejcewicz, J., Krys’ko, V.A., Vakakis, A.F.: Order Reduc- 00207179208934317
tion by Proper Orthogonal Decomposition (POD) Analysis, 20. Crutchfield, J.P.: Prediction and stability in classical
pp. 279–320. Springer, Berlin (2004) mechanics. Senior thesis in physics and mathematics, Uni-
5. Axås, J., Cenedese, M., Haller, G.: Fast data-driven model versity of California, Santa Cruz, June 1979; submitted to
reduction for nonlinear dynamical systems. Nonlinear APS First Annual Apker Undergraduate Thesis Competition
Dyn. 111(9), 7941–7957 (2022). https://doi.org/10.1007/ (1979)
s11071-022-08014-0 21. Daniel, T., Casenave, F., Akkari, N., Ryckelynck, D.: Model
6. Axås, J., Haller, G.: fastSSM: algorithm for fast computation order reduction assisted by deep neural networks (rom-net).
of spectral submanifolds from data (2022). https://github. Adv. Model. Simul. Eng. Sci. 7, 105786 (2020)
com/haller-group/SSMLearn/tree/main/fastSSM 22. Deyle, E.R., Sugihara, G.: Generalized theorems for non-
7. Bertsimas, D., Gurnee, W.: Learning sparse nonlin- linear state space reconstruction. PLoS ONE 6, 3 (2011).
ear dynamics via mixed-integer optimization. Nonlin- https://doi.org/10.1371/journal.pone.0018295
ear Dyn. 111, 6585–6604 (2022). https://doi.org/10.1007/ 23. Dodge, F.: The new “dynamic behavior of liquids in
s11071-022-08178-9 moving containers”. Southwest Research Institute, San
8. Bozzo, E., Carniel, R., Fasino, D.: Relationship between Antonio, Texas (2000). https://books.google.ch/books?
singular spectrum analysis and Fourier analysis: theory and id=RltitwAACAAJ
application to the monitoring of volcanic activity. Com- 24. Drmač, Z., Mezić, I., Mohr, R.: Data driven koopman
put. Math. Appl. 60(3), 812–820 (2010). https://doi.org/10. spectral analysis in vandermonde-cauchy form via the dft:
1016/j.camwa.2010.05.028 numerical method and theoretical insights. SIAM J. Sci.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

22098 J. Axås, G. Haller

Comput. 41(5), A3118–A3151 (2019). https://doi.org/10. 40. Kamb, M., Kaiser, E., Brunton, S.L., Kutz, J.N.: Time-delay
1137/18M1227688 observables for koopman: theory and applications. SIAM J.
25. Dylewsky, D., Barajas-Solano, D., Ma, T., Tartakovsky, Appl. Dyn. Syst. (2018)
A.M., Kutz, J.N.: Stochastically forced ensemble dynamic 41. Kaszás, B., Cenedese, M., Haller, G.: Dynamics-based
mode decomposition for forecasting and analysis of near- machine learning of transitions in couette flow. Phys. Rev.
periodic systems. IEEE Access 10, 33440–33448 (2022). Fluids 7, L082402 (2022)
https://doi.org/10.1109/ACCESS.2022.3161438 42. Kennel, M.B., Brown, R., Abarbanel, H.D.I.: Determining
26. Dylewsky, D., Kaiser, E., Brunton, S.L., Kutz, J.N.: Principal embedding dimension for phase-space reconstruction using
component trajectories for modeling spectrally continuous a geometrical construction. Phys. Rev. A 45, 3403–3411
dynamics as forced linear systems. Phys. Rev. E 105, 015312 (1992). https://doi.org/10.1103/PhysRevA.45.3403
(2022). https://doi.org/10.1103/PhysRevE.105.015312 43. Kerschen, G., Peeters, M., Golinval, J., Vakakis, A.: Nonlin-
27. Faltinsen, O., Timokha, A.: Sloshing. Cambridge University ear normal modes, part i: a useful framework for the struc-
Press, Cambridge (2009) tural dynamicist. Mech. Syst. Signal Process. 23(1), 170–
28. Faltinsen, O.M., Rognebakke, O.F., Lukovsky, I.A., Timo- 194 (2009). https://doi.org/10.1016/j.ymssp.2008.04.002
kha, A.N.: Multidimensional modal analysis of nonlinear 44. Kraemer, K.H., Gelbrecht, M., Pavithran, I., Sujith, R.I.,
sloshing in a rectangular tank with finite water depth. J. Marwan, N.: Optimal state space reconstruction via Monte
Fluid Mech. 407, 201–234 (2000). https://doi.org/10.1017/ Carlo decision tree search. Nonlinear Dyn. 108(2), 1525–
s0022112099007569 1545 (2022). https://doi.org/10.1007/s11071-022-07280-2
29. Fleissner, F., Lehnart, A., Eberhard, P.: Dynamic simulation 45. Kutz, J.N., Brunton, S.L.: Parsimony as the ultimate regu-
of sloshing fluid and granular cargo in transport vehicles. larizer for physics-informed machine learning. Nonlinear
Veh. Syst. Dyn. 48(1), 3–15 (2010). https://doi.org/10.1080/ Dyn. 107(3), 1801–1817 (2022). https://doi.org/10.1007/
00423110903042717 s11071-021-07118-3
30. Fraser, A.M., Swinney, H.L.: Independent coordinates for 46. Kutz, J.N., Brunton, S.L., Brunton, B.W., Proctor, J.L.:
strange attractors from mutual information. Phys. Rev. A Dynamic Mode Decomposition. SIAM, Philadelphia, PA
33, 1134–1140 (1986). https://doi.org/10.1103/PhysRevA. (2016)
33.1134 47. Li, M., Haller, G.: Nonlinear analysis of forced mechani-
31. Guckenheimer, J., Holmes, P.: Nonlinear Oscillations. cal systems with internal resonance using spectral submani-
Dynamical Systems and Bifircation of Vector Fields. folds, part II: bifurcation and quasi-periodic response. Non-
Springer, New York (1983) linear Dyn. 110(2), 1045–1080 (2022). https://doi.org/10.
32. Haller, G., Ponsioen, S.: Nonlinear normal modes and spec- 1007/s11071-022-07476-6
tral submanifolds: existence, uniqueness and use in model 48. Li, M., Jain, S., Haller, G.: Nonlinear analysis of forced
reduction. Nonlinear Dyn. 86(3), 1493–1534 (2016) mechanical systems with internal resonance using spectral
33. Haro, A., de la Llave, R.: A parameterization method for the submanifolds, part I: periodic response and forced response
computation of invariant tori and their whiskers in quasi- curve. Nonlinear Dyn. 110(2), 1005–1043 (2022). https://
periodic maps: rigorous results. J. Differ. Eqs. 228(2), 530– doi.org/10.1007/s11071-022-07714-x
579 (2006) 49. Li, M., Jain, S., Haller, G.: Model reduction for constrained
34. Hickey, J., Broderick, B., Fitzgerald, B., Moore, H.: Mitiga- mechanical systems via spectral submanifolds. Nonlinear
tion of wind induced accelerations in tall modular buildings. Dyn. 111(10), 8881–8911 (2023). https://doi.org/10.1007/
Structures 37, 576–587 (2022). https://doi.org/10.1016/j. s11071-023-08300-5
istruc.2022.01.037 50. Loiseau, J.C., Brunton, S.L., Noack, B.R.: From the POD-
35. Hirsh, S.M., Ichinaga, S.M., Brunton, S.L., Kutz, J.N., Brun- Galerkin Method to Sparse Manifold Models, pp. 279–320.
ton, B.W.: Structured time-delay models for dynamical sys- De Gruyter, Berlin (2020)
tems with connections to frenet-serret frame. Proc. R. Soc. 51. Lumley, J.L.: The structure of Inhomogeneous turbulent
A Math. Phys. Eng. Sci. 477, 2254 (2021). https://doi.org/ flows. In: Yaglom, A.M., Tartarsky, V.I. (eds.) Atmospheric
10.1098/rspa.2021.0097 Turbulence and Radio Wave Propagation, pp. 166–177.
36. Jain, S., Haller, G.: How to compute invariant manifolds and Nauka, Moscow (1967)
their reduced dynamics in high-dimensional finite element 52. Mitra, S., Hai, L.V., Jing, L., Khoo, B.C.: A fully cou-
models? Nonlinear Dyn. 107(2), 1417–1450 (2021). https:// pled ship motion and sloshing analysis in various container
doi.org/10.1007/s11071-021-06957-4 geometries. J. Mar. Sci. Technol. 17(2), 139–153 (2012).
37. Jain, S., Thurner, T., Li, M., Haller, G.: SSMTool: compu- https://doi.org/10.1007/s00773-012-0157-2
tation of invariant manifolds and their reduced dynamics in 53. Narimanov, G.S.: Movement of a tank partly filled by a fluid:
high-dimensional mechanics problems. (2021). https://doi. the taking into account of non-smallness of amplitude. Prikl.
org/10.5281/zenodo.4614201 Mat. Mekh. 21, 513–524 (1957). (In Russian)
38. Jain, S., Tiso, P., Haller, G.: Exact nonlinear model reduc- 54. Packard, N.H., Crutchfield, J.P., Farmer, J.D., Shaw, R.S.:
tion for a von Kármán beam: slow-fast decomposition and Geometry from a time series. Phys. Rev. Lett. 45, 712–716
spectral submanifolds. J. Sound Vib. 423, 195–211 (2018) (1980). https://doi.org/10.1103/PhysRevLett.45.712
39. Juang, J.N., Pappa, R.S.: An eigensystem realization algo- 55. Page, J., Kerswell, R.: Koopman mode expansions between
rithm for modal parameter identification and model reduc- simple invariant solutions. J. Fluid Mech. 879, 1–27 (2019)
tion. J. Guid. Control Dyn. 8(5), 620–627 (1985). https:// 56. Pan, S., Duraisamy, K.: Data-driven discovery of closure
doi.org/10.2514/3.20031 models. SIAM J. Appl. Dyn. Syst. 17(4), 2381–2413 (2018).
https://doi.org/10.1137/18m1177263

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Model reduction for nonlinearizable dynamics via delay-embedded spectral submanifolds 22099

57. Pan, S., Duraisamy, K.: On the structure of time-delay 70. Sugihara, G., May, R.M.: Nonlinear forecasting as a way
embedding in linear models of non-linear dynamical sys- of distinguishing chaos from measurement error in time
tems. Chaos Interdiscip. J. Nonlinear Sci. 30(7), 073135 series. Nature 344(6268), 734–741 (1990). https://doi.org/
(2020). https://doi.org/10.1063/5.0010886 10.1038/344734a0
58. Pikovsky, A.: Noise filtering in the discrete time dynamical 71. Szalai, R.: Invariant spectral foliations with applications to
systems. Sov. J. Commun. Technol. Electron 31(5), 911–914 model order reduction and synthesis. Nonlinear Dyn. 101,
(1986) 2645–2669 (2020)
59. Ponsioen, S., Jain, S., Haller, G.: Model reduction to spec- 72. Takens, F.: Detecting strange attractors in turbulence. In:
tral submanifolds and forced-response calculation in high- Rand, D., Young, L. (eds.) Dynamical Systems and Turbu-
dimensional mechanical systems. J. Sound Vib. 488, 115640 lence, Warwick 1980, pp. 366–381. Springer, Berlin (1981)
(2020) 73. Taylor, G.: An experimental study of standing waves. Proc.
60. Ponsioen, S., Pedergnana, T., Haller, G.: Automated com- R. Soc. Lond. Ser. A. Math. Phys. Sci. 218(1132), 44–59
putation of autonomous spectral submanifolds for nonlinear (1953)
modal analysis. J. Sound Vib. 420, 269–295 (2018) 74. Vakakis, A.: Non-linear normal modes (NNMs) and their
61. Ponsioen, S., Pedergnana, T., Haller, G.: Analytic prediction applications in vibration theory: an overview. Mech. Syst.
of isolated forced response curves from spectral submani- Signal Process. 11(1), 3–22 (1997). https://doi.org/10.1006/
folds. Nonlinear Dyn. 98, 2755–2773 (2019) mssp.1996.9999
62. Rosenberg, R.M.: The normal modes of nonlinear n-degree- 75. Williams, M.O., Kevrekidis, I.G., Rowley, C.W.: A data-
of-freedom systems. J. Appl. Mech. 29(1), 7–14 (1962). driven approximation of the koopman operator: extending
https://doi.org/10.1115/1.3636501 dynamic mode decomposition. J. Nonlinear Sci. 9, 1307–
63. Rowley, C.W., Mezić, I., Bagheri, S., Schlachter, P., Hen- 1346 (2015)
ningson, D.: Spectral analysis of nonlinear flows. J. Fluid 76. Yap, H.L., Rozell, C.: Stable takens’ embedding for linear
Mech. 641, 115–127 (2009) dynamical systems. pp. 2948–2953 (2010). https://doi.org/
64. Salmela, L., Tsipinakis, N., Foi, A., Billet, C., Dudley, 10.1109/TSP.2011.2160629
J.M., Genty, G.: Predicting ultrafast nonlinear dynam-
ics in fibre optics with a recurrent neural network. Nat.
Mach. Intell. 3(4), 344–354 (2021). https://doi.org/10.1038/ Publisher’s Note Springer Nature remains neutral with regard
s42256-021-00297-z to jurisdictional claims in published maps and institutional affil-
65. Sauer, T., Yorke, J., Casdagli, M.: Embedology. J. Stat. Phys. iations.
65, 579–616 (1991)
66. Schmid, P.: Dynamic mode decomposition of numerical and
experimental data. J. Fluid Mech. 656, 5–28 (2010)
67. Schmid, P.J.: Dynamic mode decomposition and its variants.
Annu. Rev. Fluid Mech. 54(1), 225–254 (2022). https://doi.
org/10.1146/annurev-fluid-030121-015835
68. Shaw, S., Pierre, C.: Non-linear normal modes and invariant
manifolds. J. Sound Vib. 150(1), 170–173 (1991). https://
doi.org/10.1016/0022-460x(91)90412-d
69. Shaw, S., Pierre, C.: Normal modes for non-linear vibratory
systems. J. Sound Vib. 164(1), 85–124 (1993)

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”),
for small-scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are
maintained. By accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use
(“Terms”). For these purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or
a personal subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or
a personal subscription (to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the
Creative Commons license used will apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data
internally within ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking,
analysis and reporting. We will not otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of
companies unless we have your permission as detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that
Users may not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to
circumvent access control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil
liability, or is otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by
Springer Nature in writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer
Nature journal content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates
revenue, royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain.
Springer Nature journal content cannot be used for inter-library loans and librarians may not upload Springer Nature journal
content on a large scale into their, or any other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any
information or content on this website and may remove it or features or functionality at our sole discretion, at any time with or
without notice. Springer Nature may revoke this licence to you at any time and remove access to any copies of the Springer Nature
journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express
or implied with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or
warranties imposed by law, including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be
licensed from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other
manner not expressly permitted by these Terms, please contact Springer Nature at

[email protected]