Professional Documents
Culture Documents
Jain, Breunung, Haller - 2019 - Fast Computation of Steady-State Response For High-Degree-Of-Freedom Nonlinear Systems
Jain, Breunung, Haller - 2019 - Fast Computation of Steady-State Response For High-Degree-Of-Freedom Nonlinear Systems
https://doi.org/10.1007/s11071-019-04971-1
ORIGINAL PAPER
Received: 30 October 2018 / Accepted: 20 April 2019 / Published online: 8 May 2019
© Springer Nature B.V. 2019
123
314 S. Jain et al.
ear oscillations. Among these, the classic NNM defini- hulst [21]) after a transformation to amplitude-phase
tion of Rosenberg [2] targets periodic orbits in a two- coordinates in the case of small damping, nonlineari-
dimensional subcenter-manifold [3] in the undamped ties and forcing. They consider single as well as multi-
limit of the oscillatory system. By contrast, Shaw and harmonic forcing of multi-degree of freedom systems
Pierre [4] define NNMs as the invariant manifolds and obtain the solution in terms of a multi-frequency
tangent to modal subspaces at an equilibrium point Fourier expansion. Their formulas become involved
(cf. Avramov and Mikhlin [5] for a review) allowing even for a single oscillator, and thus, condensed formu-
application to dissipative systems. Haller and Ponsioen las or algorithms are unavailable for general systems.
[6] distinguish these two notions for dissipative sys- As conceded by Mitroposkii and Van Dao [19], the
tems under possible periodic/quasi-periodic forcing, by series expansion is formal, as no attention is given to
defining an NNM as a near-equilibrium trajectory with the actual existence of a periodic response. Existence
finitely many frequencies, and introducing a spectral is indeed a subtle question in this context, since the
submanifold (SSM) as the smoothest invariant man- envisioned periodic orbits would perturb from a non-
ifold tangent to a spectral subbundle along such an hyperbolic fixed point.
NNM. Vakakis [22] relaxes the small nonlinearity assump-
Alternatively, ROMs obtained using heuristic tion and describes a perturbation approach for obtain-
projection-based techniques are also used to approx- ing the periodic response of a single-degree-of-freedom
imate steady-state response of high-dimensional sys- Duffing oscillator subject to small forcing and small
tems. These include sub-structuring methods such damping. A formal series expansion is performed
as the Craig–Bampton method [7] (cf. Theodosiou around a conservative limit, where periodic solutions
et al. [8]), proper orthogonal decomposition [9] (cf. are explicitly known (elliptic Duffing oscillator). This
Kerchen et al. [10]), reduction using natural modes approach only works for perturbations of integrable
(cf. Amabili [11], Touzé et al. [12]) and the modal- nonlinear systems.
derivative method of Idelsohn and Cardona [13] (cf. Formally applicable without any small parameter
Sombroek et al. [14], Jain et al. [15]). A common assumption is the harmonic balance method. Intro-
feature of these methods is their local nature: they duced first by Kryloff and Bogoliuboff [23] for single-
seek to approximate nonlinear steady-state response in harmonic approximation of the forced response, the
the vicinity of an equilibrium. Thus, high-amplitude method has been gradually extended to include higher
oscillations are generally missed by these approaches. harmonics and quasi-periodic orbits (cf. Chua and
The methods reviewed here are fundamentally heuris- Ushida [24] and Lau and Cheung [25]). In the harmonic
tic as the relationship between the full system and balance procedure, the assumed steady-state solution
its simplified approximation is generally unknown is expanded in a Fourier series which, upon substi-
and has to be tested in each application. Though we tution, turns the original differential equations into a
focus on finite-dimensional dynamical systems in this set of nonlinear algebraic equations for the unknown
work, the same limitations are shared by many trunca- Fourier coefficients after truncation to finitely many
tion/approximation methods when applied to infinite- harmonics. The error arising from this truncation, how-
dimensional systems as well (cf. Malookani and van ever, is not well understood. For the periodic case,
Horssen [16] for the case of Galerkin truncation applied Leipholz [26] and Bobylev et al. [27] show that the
to string vibrations). solution of the harmonic balance converges to the actual
On the analytic side, perturbation techniques relying solution of the system if the periodic orbit exists and
on a small parameter have been widely used to approx- the number of harmonics considered tends to infin-
imate the steady-state response of nonlinear systems. ity. Explicit error bounds are only available as func-
Nayfeh et al. [17,18] give a formal multiple-scales tions of the (a priori unknown) periodic orbit (cf.
expansion applied to a system with small damping, Bobylev et al. [27], Urabe [28], Stokes [29] and García-
small nonlinearities and small forcing. Their results Saldaña and Gasull [30]). The quantities involved, how-
are detailed amplitude equations to be worked out on ever, generally require numerical integration to obtain.
a case-by-case basis. Mitropolskii and Van Dao [19] For quasi-periodic forcing, such error bounds remain
apply the method of averaging (cf. Bogoliubov and unknown to the best of our knowledge.
Mitropolsky [20] or, more recently, Sanders and Ver-
123
Fast computation of steady-state response 315
The shooting method (cf. Keller [31], Peeters et al. integral equation whose zeros are the steady-state
[32] and Sracic and Allen [33]) is also broadly used responses of the mechanical system. Along with a
to compute periodic orbits of nonlinear system. In this phase condition to ensure uniqueness, the same integral
procedure, the periodicity of the sought orbit is used equation can also be used to obtain the (quasi-) peri-
to formulate a two-point boundary value problem. The odic response in conservative, autonomous mechanical
solutions are initial conditions on the periodic orbit. systems.
Starting from an initial guess, one corrects the initial While certain elements of the integral equations
conditions iteratively until the boundary value prob- approach outlined here for periodic forcing have been
lem is solved up to a required precision. The iterated already discussed outside the structural vibrations liter-
correction of the initial conditions, however, requires ature, our treatment of quasi-periodic forcing appears
repeated numerical integration of the equation of varia- to be completely new. We do not set any concep-
tions along the current estimate of the periodic orbit, as tual bounds on the number of independent frequen-
well as numerical integration of the full system. Albeit cies allowed in such a forcing, which enables one to
the shooting method has moderate memory require- apply the results to more complex excitations mimick-
ments relative to that of harmonic balance due to its ing stochastic forcing.
smaller Jacobian, this advantage is useful only for very First, we derive a Picard iteration approach with
high-dimensional systems with memory constraints. In explicit convergence criteria to solve the integral
practice, shooting is limited by the capabilities of the equations for the steady-state response iteratively
time integrator used and can be unsuitable for solutions (Sect. 3.1). This fast iteration approach is particularly
with large Floquet multipliers, as observed by Seydel appealing for high-dimensional systems, since it does
[34]. Furthermore, the shooting method is only appli- not require the construction and inversion of Jacobian
cable to periodic steady-state solutions, not to quasi- matrices, and for non-smooth systems, as is does not
periodic ones. rely on derivatives. At the same time, this Picard itera-
The shooting method uses a time-march-type inte- tion will not converge near external resonances. Apply-
gration, i.e., the solution at each time step is solved ing a Newton–Raphson scheme to the integral equa-
sequentially after the previous one. In contrast, collo- tion, however, we can achieve convergence of the iter-
cation approaches solve for the solution at all time steps ation even for near-resonant forcing (Sect. 3.2). We
in the orbit simultaneously. Collocation schemes miti- additionally employ numerical continuation schemes
gate all the drawbacks of the shooting method but can to obtain forced response and backbone curves of
be computationally expensive for large systems since nonlinear mechanical systems (Sect. J.1). Finally, we
all unknowns need to be solved together over the full illustrate the performance gain from our newly pro-
orbit. Popular software packages, such as AUTO [35], posed approach on several multi-degree-of-freedom
MATCONT [36] and the po toolbox of coco [37], also mechanical examples (Sect. 4), using a MATLAB® -
use collocation schemes to continue periodic solutions based implementation.1
of dynamical systems. Renson et al. [38] provide a thor-
ough review of the commonly used methods for com-
putation of periodic orbits in multi-degree-of-freedom 2 Setup
mechanical systems.
Constructing particular solutions using integral We consider a general nonlinear mechanical system of
equations is textbook material in physics or vibration the form
courses for impulsive forcing the (system is at rest at
the initial time, prior to which the forcing is zero). Mẍ + Cẋ + Kx + S(x, ẋ) = f(t), (1)
Solving this problem with a classic Duhamel integral
will produce a particular solution that approaches the where x(t) ∈ Rn is the vector of generalized dis-
steady-state response asymptotically. This approach, placements; M, C, K ∈ Rn×n are the symmetric
therefore, suffers from the slow convergence we have mass, stiffness and damping matrices; S is a non-
already discussed for direct numerical integration. linear, Lipschitz continuous function such that S =
In this paper, assuming either periodicity or quasi-
periodicity for the external forcing, we derive an 1 Available at https://www.georgehaller.com.
123
316 S. Jain et al.
O |x|2 , |x| |ẋ| , |ẋ|2 ; f is a time-dependent, multi- structural dynamics systems, but the following treat-
frequency external forcing. Specifically, we assume ment also allows for time dependence in S or R. Specif-
that f(t) is quasi-periodic with a rationally incommen- ically, all the following results hold for nonlineari-
surate frequency basis ∈ Rk , k ≥ 1 which means ties with explicit time dependence as long as the time
dependence is quasi-periodic (cf. Eq. (2)) with the same
f(t) = f̃(t), κ, = 0, κ ∈ Zk − {0}, (2) frequency basis as that of the external forcing f(t).
with 2π
λ j = i , ∈ Z, (7)
ẋ 0 M M 0 T
z= , B= , A= ,
x M C 0 −K
are satisfied for all eigenvalues λ1 , . . . , λ2n defined in
0 0 (5), then there exists a unique T -periodic response to
R(z) = , F(t) = .
S(x, ẋ) f(t) (4), given by
The first-order form in (3) ensures that the coeffi- T
cient matrices A and B are symmetric, if the matri-
z(t) = V G(t − s, T )V−1 F(s) ds, (8)
ces M, C and K are symmetric, as is usually the case 0
in structural dynamics applications (cf. Gérardin and
Rixen [39]). We assume that the coefficient matrix of where G(t, T ) is the diagonal matrix of periodic
the linear system Green’s functions for the modal displacement vari-
ables, defined as
Bż = Az + F(t) (4) G(t, T ) = diag (G 1 (t, T ), . . . , G 2n (t, T )) ∈ C2n×2n ,
λjT
e
can be diagonalized using the eigenvectors of the gen- G j (t, T ) = eλ j t + h(t) , j = 1, . . . , 2n,
eralized eigenvalue problem 1 − eλ j T
(9)
A − λ j B v j = 0, j = 1, . . . , 2n, (5) with the Heaviside function h(t) given by
via the linear transformation z = Vw, where w ∈ C2n 1 t ≥0
h(t) := .
represents the modal variables and V = [v1 , . . . , 0 t <0
v2n ] ∈ C2n×2n is the modal transformation matrix con- Proof We reproduce the proof for completeness in
taining the eigenvectors. The diagonalized linear ver- “Appendix A”.
123
Fast computation of steady-state response 317
The Green’s functions defined in (9) turn out to play there exists a unique quasi-periodic steady-state
a key role in describing periodic solutions of the full, response to (4) with the same frequency basis . This
nonlinear system as well. We recall this in the following steady-state response is given by
result (see eg. Bobylev et al. [27]).
Tκ
Theorem 1 (i) If z(t) is a T −periodic solution of the z(t) = V G(t − s, Tκ )V−1 F(s) ds. (14)
0
nonlinear system (3), then z(t) must satisfy the integral κ∈Zk
equation
Furthermore, z(t) is quasi-periodic with Fourier expan-
T
sion
z(t) = V G(t − s, T )V−1 [F(s) − R(z(s))] ds.
0
(11) z(t) = V H(Tκ )V−1 Fκ eiκ,t , (15)
κ∈Zk
(ii) Furthermore, any continuous, T −periodic solution
z(t) of (11) is a T −periodic solution of the nonlinear where H(Tκ ) is the diagonal matrix of the amplification
system (3). factors, defined as
H(Tκ ) = diag (H1 (Tκ ), . . . , H2n (Tκ )) ∈ C2n×2n ,
Proof We reproduce the proof for completeness in 1
“Appendix B”. The term V−1 [F(t) − R(z(t))] is treated H j (t, T ) = , j = 1, . . . , 2n . (16)
i κ, − λ j
as a periodic forcing term in (6) for a T -periodic z(t)
and Lemma 1 is used to prove (i). Statement (ii) is then Proof The proof is a consequence of the linearity of
a direct consequence of the Leibniz rule.
Remark 3 The maximum of H j (Tκ ) can be bounded
The above classic results on periodic steady-state solu- by the constant h max , defined as
tions extend to quasi-periodic steady-state solutions
max
H j (Tκ )
123
318 S. Jain et al.
Tκ
response and backbone curves. The special case of pro-
z(t) = V G(t − s, Tκ )V−1
0 portional damping and purely position-dependent non-
κ∈Zk
linearities further alleviates these computational chal-
× [F(s) − R(z(s))] ds . (18) lenges by reducing the dimensionality to half, as we
discuss in the following section.
(ii) Furthermore, any continuous, quasi-periodic
solution z(t) of (18), with frequency basis , is a quasi-
periodic solution of the nonlinear system (3).
2.3 Special case: structural damping and purely
Proof The proof is analogous to that for the periodic geometric nonlinearities
case (cf. Theorem 1). Again, the term F(t) − R(z(t))
is treated as a quasi-periodic forcing term.
123
Fast computation of steady-state response 319
Specifically, the jth mode (y j ) of equation (25) is cus- Proof See “Appendix D”.
tomarily expressed in the vibrations literature as
The periodic Green’s function L j (t, T ) for a single-
degree-of-freedom, underdamped harmonic oscillator
ÿ j + 2ζ j ω0, j ẏ j + ω0,
2
j y j = ϕ j (t), j = 1, . . . , n, has already been derived in the controls literature (see,
(26) e.g., Kovaleva [43], p. 19., formula (1.40), or Babit-
sky [44] p. 90). They also note a simplification when
uj Ku j the periodic forcing function has an odd symmetry
where ω0, j = are the undamped natural fre-
uj Mu j
with respect to half the period (e.g., sinusoidal forc-
uj Cu j ing), in which case the integral can be taken over just
quencies; ζ j = 2ω0, j
1
are the modal damping
uj Mu j
half the period with another Green’s function. Koval-
u F(t) eva [43] also lists the Green function without damping
coefficients; and ϕ j (t) = uj Mu are the modal par-
j j for the case of a multi-degree-of-freedom system with-
ticipation factors. The eigenvalues for the correspond- out damping, in transfer-function notation. In summary,
ing full system in phase space can be arranged as fol- formula (30) does not seem to appear in the vibrations
lows literature, but earlier controls literature has simpler
forms of it (single-degree-of-freedom modal form with
λ2 j−1,2 j = −ζ j ± ζ j2 − 1 ω0, j , j = 1, . . . , n. damping, or multi-dimensional form without damping
(27) in modal coordinates), albeit for the underdamped case
only.
With the constants Kovaleva [43] also observes for undamped multi-
degree-of-freedom systems that an integral equation
α j := Re(λ2 j ), ω j := |Im(λ2 j )|, β j := α j + ω j , with this Green’s function can be written out for non-
γ j := α j − ω j, j = 1, . . . , n, (28) linear systems, and then refers to Rosenwasser [45] for
existence conditions and approximate solution meth-
we can restate Lemma 1 specifically for linear systems ods. Chapter 4.2 of Babitsky and Krupenin [46] also
with proportional damping as follows. discusses this material in the context of the response of
linear discontinuous systems, citing Rosenwasser [45]
Lemma 3 For T -periodic forcing f(t), i.e., f(t + T ) = for a similar formulation. We formalize and generalize
f(t), t ∈ R, T > 0 and under the non-resonance con- these discussions as a theorem here:
ditions (7), there exists a unique T-periodic response Theorem 3 (i) If x(t) is a T-periodic solution of the
for system (24), given by nonlinear system (22), then x(t) must satisfy the inte-
T
gral equation
x(t) = U L(t − s, T )U f(s) ds, (29)
0 T
x(t) = U L(t − s, T )U [f(s) − S(x(s))] ds ,
where L(t, T ) is the diagonal Green’s function matrix 0
for the modal displacement variables defined as (31)
j = 1, . . . , n (30)
Remark 5 Once a solution to (31) is obtained for the
and ϕ(s) = [ϕ1 (s), . . . , ϕn (s)] is the forcing vector position variables x (cf. Sect. 3 for solution methods),
in modal coordinates. the corresponding velocity ẋ can be recovered as
123
320 S. Jain et al.
T
where
ẋ(t) = U J(t − s, T )U [f(s) − S(x(s))] ds,
0 Q(Tκ ) = diag (Q 1 (Tκ ), . . . , Q n (Tκ )) ∈ Cn×n ,
where J(t − s, T ) = diag (J1 (t − s, T ), . . . , is the diagonal matrix of the amplification factors,
Jn (t − s, T )) ∈ Rn×n is the diagonal Green’s matrix which are explicitly given by
whose diagonal elements are given by
⎧
α T α T α T
⎪
⎪ αjt e j ω j cos ω j (T +t)−e j cos ω j t +α j sin ω j (T +t)−e j sin ω j t
⎪
⎪
e
+
⎪
⎪ ωj 2α T α T
1+e j −2e j cos ω j T ζj < 1
⎪
⎪
⎪
⎪
⎨
h(t) ω j cos
ω j t + α j sin ω j t,
J j (t, T ) = eα j (T +t) 1−eα j T (1+α j t )+α j T , (32)
⎪
⎪ + h(t) eα j t + α j teα j t , ζj = 1
⎪
⎪ αjT 2
⎪
⎪ 1−e
⎪
⎪ β (T +t)
⎪
⎪ βje j
γ (T +t)
γje j βt γ
⎩ (β −γ )
1
βjT − γjT + h(t) β j e − γ j e
j j t , ζj > 1
j j 1−e 1−e
(ii) Furthermore, any continuous quasi-periodic solu- In contrast to dissipative systems, which have isolated
tion x(t) to (33), with frequency basis , is a quasi- (quasi-) periodic solutions in response to (quasi-) peri-
periodic solution of the nonlinear system (1). odic forcing, unforced conservative systems will gener-
ally exhibit families of periodic or quasi-periodic orbits
Proof This theorem is just a special case of Theorem 2. (cf. Kelley [47] or Arnold [48]). The calculation of
123
Fast computation of steady-state response 321
2. Any given (quasi-) periodic solution z(t) to the Clearly, a fixed point of the mapping G P in (39) cor-
autonomous system (36) is a part of a family of responds to a periodic steady-state response of system
(quasi-) periodic solutions, with an arbitrary phase (1) by Theorem 1. Starting with an initial guess z0 (t)
shift θ ∈ R. for the periodic orbit, the Picard iteration applied to the
mapping (37) is given by
Nonetheless, Theorems 1–4 still hold for system (36)
with the external forcing function set to zero. Special
care needs to be taken, however, in the numerical imple- z+1 = G P (z ) , ∈ N. (40)
mentation of these results for unforced mechanical sys-
tems, as we shall discuss in “Appendix J.1.1”. To derive a convergence criterion for the Picard itera-
tion, we define the sup norm ·0 = maxt∈[0,T ] |·| and
consider a δ−ball of C 0 -continuous and T -periodic
3 Iterative solution of the integral equations functions centered at z0 :
We would like to solve integral equations of the form Cδz0 [0, T ] := z : [0, T ] → R2n | z ∈ C 0 [0, T ],
(cf. Theorems 1 and 3) z(0) = z(T ), z − z0 0 ≤ δ . (41)
T
z(t) = VG(t − s, T )V−1 [F(s)
0 We further define the first iterate under the map G P as
− R(z(s))] ds, t ∈ [0, T ] (37) T
E(t) = G P (z0 )(t) = VG(t − s, T )V−1
to obtain periodic solutions, or integral equations of the 0
form (cf. Theorem 2 and 4) × [F(s) − R(z0 (s))] ds, t ∈ [0, T ], (42)
z(t) = V H(Tκ )V−1 (Fκ − Rκ {z}) eiκ,t (38) and denote with L zδ0 a uniform-in-time Lipschitz con-
κ∈Zk stant for the nonlinearity R(z) with respect to its argu-
ment z within Cδz0 [0, T ]. With that notation, we obtain
to obtain quasi-periodic solutions of system (3). In the the following theorem for the convergence of a Picard
following, we propose iterative methods to solve these iteration performed on (37)
equations. First, we discuss a Picard iteration and then
subsequently a Newton–Raphson scheme. Theorem 5 If the conditions
1
L zδ0 < , (43)
3.1 Picard iteration a V V−1 Γ (T )
E0
δ≥ , (44)
Picard [49] proposed an iteration scheme to show local 1 − V V−1 L zδ0 Γ (T )
existence of solutions to ordinary differential equa-
tions, which is also used as practical iteration scheme to hold for some real number a ≥ 1, then the mapping
approximate the solutions to boundary value problems G P defined in Eq. (37) has a unique fixed point in the
in numerical analysis (cf. Bailey et al. [50]). We derive space (41) and this fixed point can be found via the
explicit conditions on the convergence of the Picard successive approximation
iteration when applied to Eqs. (37), (38). T
z+1 (t) = G P (z ) (t) = VG(t − s, T )V−1
0
3.1.1 Periodic response × [F(s) − R(z (s))] ds, ∈ N (45)
We define the right-hand side of the integral equation Proof The proof relies on the Banach fixed point theo-
(11) as the mapping G P acting on the phase space vec- rem. We establish that the mapping (37) is well defined
tor z, i.e., on the space (41). Subsequently, we prove that under
T
conditions (43), (44), the mapping (37) is a contraction.
z(t) = G P (z) (t) := VG(t − s, T )V−1
0 We detail all this in “Appendix G”.
× [F(s) − R(z(s))] ds, t ∈ [0, T ]. (39)
123
322 S. Jain et al.
Remark 8 If the nonlinearity R(z) is not only Lips- of the system are realistic and affect the convergence
chitz but also of class C 1 with respect to z, then condi- positively. At the same time, low damping values in
tion (43) can be more specifically written as such systems are also typical and affect the conver-
gence negatively.
1 An advantage of the Picard iteration approach we
max max
D R j (z)
< .
1≤ j≤n |z−z0 |≤δ aΓ (T ) V V−1 have discussed is that it converges monotonically, and
(46) hence, an upper estimate for the error after a finite num-
ber of iterations is readily available as the sup norm of
Remark 9 In case of geometric (purely position depen- the difference of the last two iterations. This can be
dent) nonlinearities and proportional damping (cf. Sect. exploited in numerical schemes to stop the iteration
2.3), we can avoid iterating in the 2n−dimensional once the required precision is achieved.
phase space by defining the iteration as
T 3.1.2 Quasi-periodic response
x+1 (t) = L P (x ) (t) := UL(t − s, T )U
0 We now consider the existence of a quasi-periodic solu-
× [f(s) − S(x)] ds, t ∈ [0, T ]. (47) tion under a Picard iteration of Eq. (18), which has
The existence of the steady-state solution and the con- apparently been completely absent in the literature. We
vergence of the iteration (47) can be proven analo- rewrite the right-hand side of the integral equation (18)
gously. as the mapping
z(t) = G Q (z) (t) := V H(Tκ )V−1
Babistky [44] derives via transfer functions an itera-
κ∈Zk
tion similar to (45) but without an explicit convergence
proof. He asserts that the iteration is sensitive to the × (Fκ − Rκ {z}) e iκ,t
, (48)
choice of the initial conditions z0 . We can directly con- where we have made use of the Fourier expansion
firm this by examining condition (44). Indeed, the norm defined in Remark 4.
of the initial error E0 is small for a good initial guess. We consider a space of quasi-periodic functions with
Therefore, the δ-ball in which the condition (43) on the the frequency base . Similarly to the periodic case (cf.
Lipschitz constant needs to be satisfied can be selected Sect. 3.1.1), we restrict the iteration to a δ−ball Cδz0 ()
small. centered at the initial guess z0 with radius δ, i.e.,
When no a priori information about the expected
steady-state response is available, we can select z0 (t) ≡ Cδz0 () := z(θ ) : Tk → R2n | z ∈ C 0 ,
0. Then, the term E(0, t) is equal to the forced response
z − z0 0 ≤ δ , (49)
of the linear system [cf. Eq. (42)]. In this case, the Lip-
schitz constant needs to be calculated for a δ−ball cen- where the sup norm ·0 = maxθ∈Tk |·| is the uniform
tered at the origin. supremum norm over the torus Tk . We then have the
The constant Γ (T ) [cf. Eq. (10)] affects the con- following theorem.
vergence of the iteration (45). Larger damping (i.e.,
smaller eRe(λ j )T ), larger distance of the forcing fre- Theorem 6 If the conditions
quency 2π/T from the natural frequencies (i.e., larger 1
L zδ0 < , (50)
|1−eλ j T |), and higher forcing frequencies (i.e., smaller a V V−1 h max
T ) all make the right-hand side of (43) larger and E0
δ≥ , (51)
hence are beneficial to the convergence of the itera- 1 − 2 V V−1 L zδ0 h max
tion. Likewise, a good initial guess (i.e., smaller
E
0)
and smaller nonlinearities (i.e., smaller
DS j (x)
) all hold for some real number a ≥ 1, then the mapping
make the left-hand side of (43) smaller and hence are G Q defined in Eq. (48) has a unique fixed point in the
similarly beneficial to the convergence of the iteration. space (49) and this fixed point can be found via the
In the context of structural vibrations, higher frequen- successive approximation
cies, smaller forcing amplitudes, and forcing frequen-
cies sufficiently separated from the natural frequencies z+1 (t) = G Q (z ) (t), ∈ N. (52)
123
Fast computation of steady-state response 323
Proof The is analogous to the proof of Theorem 5. We convergence criteria for the Picard iteration will not be
first establish that the mapping (48) is well defined in satisfied for near-resonant forcing and low damping.
the space (49). In “Appendix H”, we detail that the map- However, even if the Picard iteration fails to converge,
ping (48) is a contraction under the conditions (50) and one or more periodic orbits may still exist.
(51). A common alternative to the contraction mapping
approach proposed above is the Newton–Raphson
Remark 10 In case of geometric (position-dependent) scheme (cf., e.g., Kelley [51] ). An advantage of this
nonlinearities and proportional damping, we can reduce iteration is its quadratic convergence if the initial guess
the dimensionality of the iteration (52) by half, using is close enough to the actual solution of the problem.
(34). This results in the following, equivalent Picard This makes this procedure also appealing for a con-
iteration: tinuation setup. We first derive the Newton–Raphson
x+1 (t) = L Q (x ) (t) := U Q(Tκ )U scheme to periodically forced systems and afterward
to quasi-periodically forced systems.
κ∈Zk
× [fκ − Sκ {x }] e iκ,t
, ∈ N. (53)
3.2.1 Periodic case
The existence of the steady-state solution and the con-
vergence of the iteration (53) can be proven analo-
To set up a Newton–Raphson iteration, we reformulate
gously.
the fixed point problem (37) with the help of a func-
As in the periodic case, the convergence of the itera- tional F P whose zeros need to be determined:
tion (52) depends on the quality of the initial guess and T
the constant h max [cf. Eq. (17)], which is the maximum F P (z) := z − G P (z) = z − VG(t − s, T )V−1
0
amplification factor. Low damping results in a higher
× [F(s) − R(z(s))] ds = 0. (54)
amplification factor [cf. Eq. (17)] and will therefore
affect the iteration negatively, which is similar to the Starting with an initial solution guess z0 , we formulate
criterion derived in the periodic case. the iteration for the zero of the functional F P as
ds s=0
case of unforced conservative systems. Thus, in such T
cases, more advanced iterative schemes equipped with = μl + VG(t − s, T )V−1 DR (z(s))|z=zl
continuation algorithms are desirable, such as the ones 0
μl (s)ds. (56)
we describe next.
Equation (56) is a linear integral equation in μl , where
zl is the known approximation to the solution of (54)
3.2 Newton–Raphson iteration at the lth iteration step.
123
324 S. Jain et al.
F Q (z) := z − G Q (z) = z − VH(Tκ )V−1 involves the inversion of the corresponding linear oper-
κ∈Zk ator in Eq. (56) or (59), which is costly for large sys-
× (Fκ − Rκ {zl }) eiκ,t . (57) tems.
Regarding the first issue above, the tangent stiffness
Analogous to the periodic case, the Newton–Raphson is often available in finite-element codes for structural
scheme seeks to find a zero of F Q via the iteration: vibration. Nonetheless, there are many quasi-Newton
schemes in the literature that circumvent this issue by
zl+1 = zl + ν l , offering either a cost-effective but inaccurate approx-
(58) imation of the Jacobian (e.g., the Broyden’s method
−F Q (zl ) = DF Q (zl )ν l .
[52]), or avoid the calculation of the Jacobian alto-
To obtain the correction step ν l , the linear system of gether (cf. Kelley [51]).
equations For the second challenge above, one can opt for
the iterative solution of the linear system (56) or (59),
−F Q (zl ) = D F Q (zl )ν l
which would circumvent operator inversion when the
system size is very large. A practical strategy for
= νl + V H(Tκ )V−1 (Fκ −{DR(zl )ν l }κ ) eiκ,t obtaining force response curves of high-dimensional
κ∈Zk
systems would be to use the Picard iteration away
(59) from resonant forcing and switch toward the Newton–
needs to be solved for ν l . The Fourier coefficients Raphson approach when the Picard iteration fails. Even
{DR(zl )ν l }κ in (59) are then given by the formula though the Newton–Raphson approach has better rate
of convergence (quadratic) as compared to the Picard
{DR(zl (t))ν l }κ approach (linear), the computational complexity of a
τ
1 single Picard iteration is an order of magnitude lower
= lim DR(z(t))ν l (t)e−iκ,t dt .
τ →∞ 2τ −τ than that of Newton–Raphson method (simply because
of the additional costs of Jacobian evaluation and inver-
Remark 11 As noted in Introduction, the results in sion involved in the correction step). In our experi-
Sects. 3.1.2 and 3.2.2 hold without any restriction on ence with high-dimensional systems, when the Picard
the number of independent frequencies allowed in the iteration converges, it is significantly faster than the
forcing function f(t). This enables us to compute the Newton–Raphson in terms of CPU time, even though
steady-state response for arbitrarily complicated forc- it takes significantly more number of iterations to con-
ing functions, as long as they are well approximated verge.
by a finite Fourier expansion. Thus, the treatment of Both the Picard and the Newton–Raphson iter-
random-like steady-state computations is possible with ation are efficient solution techniques for specific
the methods proposed here. forcing functions. However, to obtain the steady-
state response as a function of forcing amplitudes
and frequencies, numerical continuation (cf. Dankow-
Discussion of the iteration techniques and numerical icz and Schilder [37], Doedel et al. [35], Dhooge et al.
solution [36]) is required. We discuss numerical continuation in
the context of the proposed integral equations approach
The Newton–Raphson iteration offers an alternative in “Appendix J.1”.
to the Picard iteration, especially when the system is In the case of multiple coexisting equilibrium posi-
forced near resonance, and hence, the convergence of tions in the unforced-damped system, the persistence
the Picard iteration cannot be guaranteed. At the same of these solutions as k-dimensional tori under small
time, the Newton–Raphson iteration is computation- (quasi-) periodic forcing can be deduced by the gen-
ally more expensive than the Picard iteration for two eral results of Haro and de la Llave [53]. Depending
reasons: First, the evaluation of the Gateaux deriva- on the specific application, a selection of these solu-
tives (56) and (59) can be expensive, especially if the tions can be numerically continued with the described
Jacobian of the nonlinearity is not directly available. techniques. To explore bifurcation phenomena, such as
Second, the correction step μl or ν l at each iteration merging of such solutions, advanced continuation tech-
123
Fast computation of steady-state response 325
We consider a two-degree-of-freedom oscillator shown which is the same nonlinearity considered by Szalai
in Fig. 1. The nonlinearity S is confined to the first et al. [57]. The integral-equation-based steady-state
spring and depends on the displacement of the first mass response curves are shown in Fig. 2 for a full fre-
only. The equations of motion are quency sweep and for different forcing amplitudes. As
123
326 S. Jain et al.
123
Fast computation of steady-state response 327
Table 1 Comparison of
Forcing Computation time [seconds (# continuation steps)]
computational performance
amplitude (A)
for different continuation po-toolbox of Integral eq. with Integral eq. with
approaches coco in-house continuation coco continuation
0.3
maximal displacement q 1
0.25
0.2
0.15
0.1
0.05
0
1.9
1.8
1.7
1.05
1
Ω2 1.6
0.9
0.95
Ω1
(a) (b)
Fig. 4 a Response curve for Example 1 with the nonlinearity anteed region of convergence for the Picard iteration. The white
(62) and the forcing (63), and the non-dimensional parameters region is the domain where this iteration fails, and we employ
m = 1, k = 1 and c = 0.02; b number of iterations needed on the Newton–Raphson scheme. (Color figure online)
the construction of Fig. 4a. Red curves bound the a priori guar-
tion of the two forcing frequencies, which are always isfied and, accordingly, the iteration is guaranteed to
selected to be incommensurate; otherwise, the forcing converge. Since these conditions are only sufficient
would not be quasi-periodic. We nevertheless connect for convergence, the iteration converges also for fre-
the resulting set of discrete points with a surface in quency pairs within the red curves. The number of
Fig. 4a for better visibility. iterations required increases gradually and within the
To carry out the quasi-periodic Picard iteration (53), white region bounded by green lines, the Picard iter-
the infinite summation involved in the formula has to be ation fails. In such cases, we proceed to employ the
truncated. We chose to truncate the Fourier expansion Newton– Raphson scheme (cf. Sect. 3.2.2).
once its relative error is within 10−3 . If the iteration (53)
did not converge, we switched to the Newton–Raphson
scheme described in Sect. 3.2.2. In that case, we only
kept the first three harmonics as Fourier basis. 4.1.3 Non-smooth nonlinearity
Figure 4b shows the number of iterations needed
to converge to a solution with this iteration procedure. As noted earlier, our iteration schemes are also appli-
Especially away from the resonances, a low number cable to non-smooth system as long as they are still
of iterations suffices for convergence to an accurate Lipschitz. We select the nonlinearity of the form
result. Also included in Fig. 4b are the conditions (50)
and (51), which guarantee the convergence for the iter-
ation to the steady-state solution of system (1). Out- αsign(q1 )(|q1 | − β), for |q1 | > β,
S(q1 ) = (64)
side the two red curves, both (50) and (51) are sat- 0, otherwise,
123
328 S. Jain et al.
(a) (b)
Fig. 5 a Graph of the non-smooth nonlinearity (64); b response curve for Example 1 with the nonlinearity (64) and the forcing (61);
parameters: m = 1, k = 1 and c = 0.02, α = β = 0.1
123
Fast computation of steady-state response 329
123
330 S. Jain et al.
t
Acknowledgements We are thankful to Harry Dankowicz and + e(t−s) ψ(s) ds
Mingwu Li for clarifications and help with the continuation pack- 0
t −1
age coco [37]. We also acknowledge helpful discussions with
Mark Mignolet and Dane Quinn. = eT I − eT + I e(t−s) ψ(s) ds
0
T −1
Compliance with ethical standards + eT I − eT e(t−s) ψ(s) ds
t
Conflict of interest The authors declare that they have no con-
T −1
flict of interest. = eT I − eT + h(t − s)I e(t−s) ψ(s) ds ,
0
$ %& '
G(t−s,T )
(68)
A Proof of Lemma 1
where G(t, T ) is a diagonal matrix with the entries
The general solution of (6) is given by the classic vari-
given by (9). Using the linear modal transformation
ation of constants formula
z = Vw, we find the unique T -periodic solution to
t system (4) in the form
t
w(t) = e w(0) + e(t−s) ψ(s) ds. (65) T
0
z(t) = V G(t − s, T )ψ(s) ds.
A T -periodic solution w0 (t) of (6) must satisfy (65), 0
resulting in
T B Proof of Theorem 1
T
w0 (T ) = e w0 (0) + e(T −s) ψ(s) ds = w0 (0).
0
If z(t) is a T -periodic solution of (3), then it satisfies
(66)
the linear inhomogeneous differential equation
#
Since the matrix I − eT is invertible due to the non- Bż = Az + F(t) − R(z(t)),
resonance condition (7), this allows us to solve Eq. (66)
for a unique initial condition w0 (0) as (cf. Burd [40], where we view F(t) − R(z(t)) as a T −periodic forcing
Chapter 2) term. Thus, according to Lemma 1, we have
T
−1 T z(t) = V G(t − s, T )V−1 [F(s) − R(z(s))] ds,
w0 (0) = I − eT e(T −s) ψ(s) ds. (67) 0
0
as claimed in statement (i).
Substituting the unique initial condition from (67) into Now, let z(t) be a continuous, T −periodic solu-
the general solution (65) provides us an explicit expres- tion to (11). After introducing the notation χ (t) =
sion of the unique T -periodic solution w0 (t) to sys- V−1 [F(t) − R(z(t))], we have
tem (6) as
T
−1 T z(t) = V G(t − s, T )χ (s) ds,
w0 (t) = et I − eT e (T −s) ψ(s) ds 0
0 −1 T
t
= Vet I − eT e(T −s) χ (s) ds
+ e(t−s) ψ(s) ds 0
0 t
−1
=e T
I − eT
T
e(t−s) ψ(s) ds
+V e(t−s) χ (s) ds , (69)
0
0
t
+ e(t−s) ψ(s) ds where (69) is a direct consequence of (68). By the con-
0
−1 t tinuity of z(t), χ(t) is also at least C 0 (F is at least C 0
= eT I − eT e(t−s) ψ(s) ds and R is Lipschitz). Thus, for any t ∈ [0, T ], the right-
0
−1 T hand side of (69) can be differentiated with respect to
+ eT I − eT e(t−s) ψ(s) ds t according to the Leibniz rule, to obtain
t
123
Fast computation of steady-state response 331
−1 T
dz(t) 1 1
= Ve t
I−e T
e(T −s) χ (s) ds = eλ j t ψκ, j
dt 1 − eλ j Tκ i κ, − λ j
0 κ∈Zk
t
+ V e(t−s) χ(s) ds + Vχ(t) e(iκ,−λ j )t − 1 + eλ j Tκ (e(iκ,−λ j )Tκ
0
− e(iκ,−λ j )t )
= VV−1 z(t)
1 1
+ Vχ (t) [using (69)] = eλ j t ψκ, j
i κ, − λ j 1 − eλ j Tκ
= VV−1 z(t) + VV−1 [F(t) − R(z(t))] ,
κ∈Zk
(1 − eλ j Tκ )(e(iκ,−λ j )t ) − 1 + eiκ,Tκ
which implies 1
= ψκ, j eiκ,t .
dz(t) i κ, − λ j
B = Az(t) + F(t) − R(z(t)), κ∈Zk
dt
as claimed in statement (ii).
λ j Tκ Tκ
e
+ eλ j t ψ κ, j e−λ j s eiκ,s ds y j (t) ( j) y j (0)
1 − e λ j Tκ = N (T ) , j = 1, . . . , n.
κ∈Zk
t
ẏ j (t) ẏ j (0)
1 1 (72)
= eλ j t λ
ψκ, j
1−e j Tκ i κ, − λj
κ∈Zk
s=t Since F(t) is uniformly bounded for all times and all
e(iκ,−λ j )s
e(iκ,−λ j )s
tial condition y j (0), ẏ j (0) for the unique T -periodic
s=t
123
332 S. Jain et al.
solution of (71) is obtained by imposing periodicity, Thus, we can explicitly compute the particular periodic
i.e., y j (0) = y j (T ) for j = 1, . . . , n and is given by solution given in (74) using (75) as
⎛ ⎞
( j) ( j) ( j) ( j)
y j (0) 1 T 1 − N22 (T ) N12 (T − s) + N12 (T )N22 (T − s)
= ⎝ ⎠ ϕ j (s) ds.
y j (0) 1 − Trace N (T ) + det N( j) (T )
( j)
0
( j) ( j) ( j) ( j)
1 − N11 (T ) N22 (T − s) + N21 (T )N12 (T − s)
(73)
Finally, the unique periodic response y j (t), ẏ j (t) is
obtained by substituting the initial condition (73) into
the Duhamel’s integral formula (71) as
⎛ ⎞
T ( j) ( j) ( j) ( j)
y j (t) N( j) (t) 1 − N22 (T ) N12 (T − s) + N12 (T )N22 (T − s)
= ⎝ ⎠ ϕ j (s) ds
ẏ j (t) 1 − Trace N( j) (T ) + det N( j) (T ) 0 ( j) ( j) ( j) ( j)
1 − N11 (T ) N22 (T − s) + N21 (T )N12 (T − s)
t
0
+ N( j) (t − s) ds, j = 1, . . . , n . (74)
0 ϕ j (s)
⎧ , -
⎪
⎪ α t ω j cos ω j t − α j sin ω j t sin ω j t
⎪
⎪ e j
, ζj < 1
⎪
⎪ ωj
−α 2j sin ω j t − ω2j sin ω j t ω j cos ω j t + α j sin ω j t
⎪
⎪ , -
⎪
⎨ eα j t − α teα j t teα j t
N( j) (t) =
j
, ζj = 1 , j = 1, . . . , n . (75)
⎪
⎪
α
−α j te j
2 t e j + α j teα j t
α t
⎪
⎪ , -
⎪
⎪ 1 β j eγ j t − α j eβ j t eβ j t − eγ j t
⎪
⎪
⎩ β j −γ j γ β eγ j t − eβ j t β eβ j t − γ eγ j t ,
⎪ ζj > 1
j j j j
123
Fast computation of steady-state response 333
⎧
⎪
⎪ α t
α T α T
e j sin ω j (T +t)−e j sin ω j t
⎪
⎪ e j
+ h(t) sin ω j t , ζj < 1
⎪
⎪ ωj 2α T α T
1+e j −2e j cos ω j T
⎪
⎪
⎪
⎨ eα j (T +t) 1−eα j T t+T
L j (t, T ) = + h(t)teα j t , ζj = 1 ,
⎪
⎪ 1−e j
α T 2
⎪
⎪
⎪
⎪ β (T +t) γ T
1−e j −e j
γ (T +t) β T
βt
⎪
⎪
e j 1−e j
γ
⎩
1
⎪ (β j −γ j ) + h(t) e − e j j t , ζj > 1
1−e j −e j +e( j j )
γ T β T γ +β T
⎧
α T α T α T
⎪
⎪ α t e j ω j cos ω j (T +t)−e j cos ω j t +α j sin ω j (T +t)−e j sin ω j t
⎪
⎪
e j
+
⎪
⎪ ωj 2α T α T
1+e j −2e j cos ω j T
⎪
⎪ , ζj < 1
⎪
⎪
⎪
⎪ h(t) 1
cos ω t + α sin ω t
⎨ ωj j j j
J j (t, T ) = eα j (T +t) 1−eα j T (1+α j t )+α j T
⎪
⎪ 2 + h(t) eα j t + α j teα j t , ζ j = 1, j = 1, . . . , n.
⎪
⎪ α T
⎪
⎪
1−e j
⎪
⎪ γ j (T +t)
⎪
⎪
β (T +t)
βje j
γ T
1−e j −γ j e j
β T
1−e j βt
⎪
⎪ (β j −γ j )
1
+ h(t) β j e − γ j e
j γ j t , ζj > 1
⎩ 1−e j −e j +e( j j )
γ T β T γ +β T
⎧ α (t+T ) #
⎪ e j
⎪ 2 ω j cos ω j (T + t) + α j sin ω j (T + t) −
⎪
⎪ 2α j T αj T
⎪
⎪
ω j 1+e −2e cos ω j T , ζj < 1
⎪
⎪
⎪
⎪ 2eα j T ω j cos ω j t + α j sin ω j t + e2α j T α j sin ω j (t − T ) + ω j cos ω j (t − T )
⎪
⎪ α (t+T )
⎪
⎪ α e j αjT α T
⎪ jα j T 2 t + T − 2teα j T + 1 − 2e (Tα−t
⎪
(e j −1))
ζ j = 1, j = 1, . . . , n.
⎨ (e −1) T
(e j −1)
dL j
(t, T ) = β j −(γ j +β j )e
γj T β j (T +t)
+ (γ j +β j )e
βjT
−γ j e
γ j (T +t)
dT ⎪
⎪ 1
e
+
⎪
⎪ (β −γ ) (γ j +β j )T
⎪
⎪
j j
1−e
γj T
−e
βj T
+e
⎪
⎪ ⎤
⎪
⎪ eβ j (T +t) 1−eγ j T −eγ j (T +t) 1−eβ j T γ j eγ j T +β j eβ j T −(γ j +β j )e(γ j +β j )T , ζj > 1
⎪
⎪
⎪
⎪ ⎥
⎪
⎪ 2 ⎦
⎩ 1−e j −e j +e( j j )
γ T β T γ +β T
The derivative with respect to the time period T of the We derive an estimate for the sup norm of the inte-
first-order periodic Green’s function G given in (9) is gral.of the operator norm
of the Green’s function, i.e.,
for 0 G j (t − s, T ) ds defined in equation (9). For
simply given by T
123
334 S. Jain et al.
1
Under the non-resonance condition (7), the linear map
G j (t − s, T )
=
eλ j (t−s)
1 − eλ j T
T
Υ P : P2n → P2n , Υ P p = V G(t − s, T )V−1 p(s) ds
1
≤
eλ j (t−s)
0
1 − e λ j T
is well defined, i.e., Υ P maps T -periodic functions into
max(
eλ j t
, 1)
≤
, 0 ≤ s ≤ t < T. T -periodic functions. Indeed, for any p ∈ P2n , let
1 − e λ j T
q = Υ P p. We have
q(t) = V G(t − s, T )V−1 p(s) ds
G j (t − s, T )
0
λ j T
T
λ (t−s) e
=V G(t + T − (s + T ), T )V−1 p(s) ds
=
e
j
1−e λ j T
0
λ j T
λ j T
T
e
λ (t−T ) e
=V G(t + T − (s + T ), T )V−1 p(s + T ) ds
≤ max
,
e j
1−e λ j T
1−e λ j T
0
λ t
2T
max(
e j
, 1) =V G(t + T − σ, T )V−1 p(σ ) dσ
≤
, 0 ≤ s ≤ t < T.
1 − e λ j T
T
= q(t + T ) ,
The upper bounds on the Green’s function in the two
intervals are equal and we therefore obtain i.e., q ∈ P2n .
T Since the space (41) consists of periodic functions,
G(t − s, T ) ds we know that it is well defined in the space Cδz0 [0, T ].
0 0 Therefore, by the Banach fixed point theorem, the inte-
T gral equation (37) has a unique solution if the map-
= max G(t − s, T ) ds
t∈[0,T ] 0 ping H is a contraction of the complete metric space
T
Cδz0 [0, T ] into itself for an appropriate choice of the
max(
eλ j t
, 1)
≤ max max
ds radius δ > 0 and the initial guess z0 .
t∈[0,T ] 0 1≤ j≤n
1 − e λ j T
To find a condition under which this holds, we first
T max(
eλ j T
, 1) note that for z−z0 0 ≤ δ, Eq. (37) gives
≤ max
=: Γ (T ).
1≤ j≤n
1 − e λ j T
T
|H(z(t))| =
G Proof of Theorem 5
In the following, we derive conditions under which the T
=
VG(t − s, T )V−1 [F(s) − R(z0 (s))] ds
mapping H defined in equation (37) is a contraction 0 0
mapping. We rewrite (37) as T
+
[R(z0 (s)) − R(z(s))] ds
VG(t − s, T )V −1
z(t) = Υ P (F(t) − R(z(t))) 0
−1 z0
0
T ≤ E (z0 , t)0 + V V L δ z − z0 0
T
:= VG(t − s, T )V−1 [F(s) − R(z(s))] ds,
0 G(t − s, T ) ds
0
where Υ P is a linear map representing the convolution ≤ E (z0 , t)0 + V V−1 L zδ0 δ
operation with the Green’s function. Specifically, we T
H(t − s, T ) ds
define the space of n-dimensional periodic T -periodic
0 0
functions as 2
≤ E (z0 , t)0 + δ V V−1 L zδ0 Γ (T ),
Pn := {p : R → Rn , p ∈ C 0 , p(t) = p(t+T )∀t ∈ R}. where L zδ0 denotes a uniform-in-time Lipschitz con-
(78) stant for the function S(z) with respect to its argument
123
Fast computation of steady-state response 335
z within the ball |z − z0 | ≤ δ, and Γ (T ) is the constant Furthermore, we note that under the non-resonance
defined in (10). The initial error term E(t) is defined in condition (7), the linear map
Eq. (42). Taking the sup norm of both sides, we obtain
Υ Q : Q2n → Q2n , Υ Q q
that H(z)0 ≤ δ, and hence
Tκ
H : Cδz0 [0, T ] → Cδz0 [0, T ] =V G(t − s, Tκ )V−1 q(s) ds
0
κ∈Zk
holds, whenever condition (44) holds. is well defined, i.e., Υ Q maps any quasi-periodic func-
Similarly, for two functions z, z̃ ∈ Cδz0 [0, T ], Eq. tion q with frequency base vector to quasi-periodic
(37) gives the estimate functions with the same frequency base vector . This
is a direct consequence of the linearity of the Υ Q and
|H(z(t)) − H(z̃(t))| definition of Υ P in “Appendix G”.
T
#
Since the mapping (48) is well defined in the space
≤
iκ,t
+ Rκ {z0 } − Rκ {z}] e
H Proof of Theorem 6
≤ V −1
H(Tκ )V [Fκ − Rκ {z0 }] e iκ,t
We show here that the mapping H defined in the quasi-
κ∈Zk
periodic case (cf. equation (48)) is a contraction on the 0
space (49) if the conditions (50) and (51) hold. The con-
+ V −1
H(Tκ )V [Rκ {z0 } − Rκ {z}] e iκ,t
vergence estimate for the iteration (52) is then similar
κ∈Zk
in spirit to the periodic case (cf. “Appendix G”). 0
−1
We rewrite (38) as ≤ E(z0 , t)0 + V V h max
z(t) = Υ Q (F(t) − R(z(t)))
:= V H(Tκ )V−1 (Fκ − Rκ {z}) eiκ,t , × [R {z
κ 0 } − R κ {z}] e iκ,t
κ∈Zk
κ∈Zk 0
−1
where Υ Q is a linear map representing the convolution ≤ E(z0 , t)0 + V V h max
operation with the Green’s function. Similarly to the
× R(z0 (s), s) − R(z(s), s)0
periodic case, we define the space of n-dimensional
quasi-periodic functions with frequency base vector ≤ E(z0 , t)0 + δ V V−1 h max L zδ0 ,
as
where we have used that the Fourier series of the
Q2n := {p : T → R , p ∈ C }.
k n 0
(80) nonlinearity k Rκ {z}eiκ,t converges to the func-
123
336 S. Jain et al.
|H(z(t)) − H(z̃(t))|
I Explicit expressions for Fourier coefficients
#
VH(Tκ )V −1
Rκ {z} − Rκ {z̃} e iκ,t
in Remark 6
κ∈Zk
Tκ
w j (t) = L j (t − s, Tκ )ψ j,κ eiκ,s ds
0
κ∈Zk
#
Tκ eα j (t−s) eα j Tκ sin ω j (Tκ + t − s) − eα j Tκ sin ω j (t − s)
= + h(t − s) sin ω j (t − s) ψ j,κ eiκ,s ds
0 ωj 1 + e2α j Tκ − 2eα j Tκ cos ω j Tκ
κ∈Zk
t eα j (t−s)
= sin ω j (t − s)ψ j,κ eiκ,s ds
0 ωj
κ∈Zk
#
Tκ eα j (t−s) eα j Tκ sin ω j (Tκ + t − s) − eα j Tκ sin ω j (t − s)
+ ψ j,κ eiκ,s ds
0 ωj 1 + e2α j Tκ − 2eα j Tκ cos ω j Tκ
κ∈Zk
t , -
eα j t eα j (Tκ +t) 1
= ψ j,κ sin ω j (t − s)e(iκ,−α j )s ds + ψ j,κ
ωj 0 ωj 1 + e2α j Tκ − 2eα j Tκ cos ω j Tκ
k κ∈Z k κ∈Z
Tκ Tκ
× e(iκ,−α j )s sin ω j (Tκ + t − s) ds − eα j Tκ e(iκ,−α j )s sin ω j (t − s) ds
0 0
s=t
eα j t e(iκ,−α j )s ω j cos ω j (t − s) + (i κ, − α j ) sin ω j (t − s)
= ψ j,κ
κ∈Z
k s=0
⎡
⎤
eα j (Tκ +t) e(iκ,−α j )s ω j cos ω j (Tκ + t − s) + (i κ, − α j ) sin ω j (Tκ + t − s)
s=Tκ
+ ⎣ s=0 ⎦ ψ j,κ
ωj 1 + e2α j Tκ − 2eα j Tκ cos ω j Tκ (i κ, − α j )2 + ω2j
κ∈Zk
⎡
⎤
eα j (2Tκ +t) e(iκ,−α j )s ω j cos ω j (t − s) + (i κ, − α j ) sin ω j (t − s)
s=Tκ
− ⎣ s=0 ⎦ ψ j,κ
ωj 1 + e 2α j Tκ − 2eα j Tκ cos ω T (i κ, − α )2 + ω2
κ∈Zk j κ j j
1
= ψ j,κ eiκ,t .
(i κ, − α j )2 + ω2j
κ∈Zk
123
Fast computation of steady-state response 337
Finally, for the overdamped configuration (ζ j > 1), we Equation (82) needs to be solved for the unknown coef-
obtain ficients cm j ∈ R2n for all j ∈ {1, . . . , m}. Note that
Tκ
w j (t) = L j (t − s, Tκ )ψ j,κ eiκ,s ds
0
κ∈Zk
eβ j (Tκ +t−s) 1 − eγ j Tκ − eγ j (Tκ +t−s) 1 − eβ j Tκ
Tκ
= 1 − eγ j Tκ − eβ j Tκ + e(γ j +β j )Tκ
0 (β j − γ j )
κ∈Zk
h(t − s) eβ j (t−s) − eγ j (t−s)
+ ψ j,κ eiκ,s ds
(β j − γ j )
t t
ψ j,κ
= eβ j t e(iκ,−β j )s ds − eγ j t e(iκ,−γ j )s ds
(β j − γ j ) 0 0
κ∈Zk
. . Tκ
ψ j,κ 1 − eγ j T eβ j (T +t) 0 e(iκ,−β j )s ds − 1 − eβ j T eγ j (T +t) 0 e(iκ,−γ j )s ds
Tκ
+
κ∈Zk
(β j − γ j )(1 − eγ j Tκ − eβ j Tκ + e(γ j +β j )Tκ )
1
= ψ j,κ eiκ,t .
(β j − i κ, )(γ j − i κ, )
kκ∈Z
123
338 S. Jain et al.
exact solution z is twice continuously differentiable in equations of the second kind are considerably sim-
t, we also obtain an error estimate for the linear spline pler than for those of the first kind. We refer to
collocation approximate solution zm as Atkinson [56] for an exhaustive treatment of numeri-
cal methods for such integral equations. Our supple-
zm − z∞ ≤ Mz̈∞ 2 ,
mentary MATLAB® code provides the implementa-
where is the spacing between the uniform colloca- tion details of a simple yet powerful collocation-based
tion points and M is an order constant depending upon approach for the periodic case, and a Galerkin pro-
the Green’s function G. jection using a Fourier basis for the quasi-periodic
Alternatively, the Galerkin method can be chosen to case.
approximate the solution of Eqs. (37) and (38) using
Fourier basis with harmonics of the base frequencies, J.1 Numerical continuation
followed by a projection onto the same basis vectors,
given by A simple approach is to use sequential continuation,
in which the frequency (or time period) of oscillation
Tj
is treated as a continuation parameter (multi-parameter
zm (s)bm j (s) ds
0 in case of quasi-periodic oscillations). This parameter
Tj is varied in small steps and the corresponding (quasi)
= VH(Tκ )V−1 (Fκ − Rκ {zm }) periodic response is iteratively computed at each step.
0
κ∈Zk The solution in any given step is typically used as initial
e iκ,s
bm j (s) ds , j = 1, . . . , m, (83) guess for the next step. Such an approach will generally
fail near a fold bifurcation with respect to one of the
where bm j (t) = ei κ j ,t are the Fourier basis func- base frequencies . In such cases, more advanced con-
tions, with the corresponding time periods T j = tinuation schemes such as the pseudo-arc-length con-
2π tinuation are required.
κ j , . Due to the orthogonality of these basis func-
tion, system (83) simplifies to
The pseudo-arc-length approach. We briefly explain
cm j = VH(Tκ )V−1 (Fκ − Rκ {zm }) , the steps involved in numerical continuation for cal-
κ∈Zk culation of periodic orbits using the pseudo-arc-length
j = 1, . . . , m approach, which is commonly used to track folds in
a single-parameter solution branch. We seek the fre-
which is a system of nonlinear equations for the quency response curve of system (22) as a family of
unknown coefficients cm j . solutions to
In the frequency domain, this system of coefficient
equations are the same as that obtained from the multi- T
frequency harmonic balance method after finite trunca- N (x, T ) := x − UL (t − s; T ( p)) U
0
tion (see, e.g., Chua and Ushida [24] or Lau and Che-
(f(s) − S(x(s))) ds = 0, (84)
ung [25]). Our explicit formulas in (35), however,
should speed up the calculations relative to a general
with the fictitious variable p parameterizing the solu-
application of the harmonic balance method. The same
tion curve (x, T ) of equation (84). If (x, T ) is a regular
scheme applies in the periodic case. For both cases
solution for equation (84), then the implicit function
(periodic and quasi-periodic), the error due to the omis-
theorem guarantees the existence of a nearby unique
sion of higher harmonics in harmonic balance proce-
solution. The tangent vector t to the solution curve
dure is not well understood (see our review of the avail-
at (x, T ) is given by the null-space of the Jacobian
able results on the periodic case in Introduction ). For
DN |(x,T ) , i.e., t can be obtained by solving the sys-
the quasi-periodic case, no such error bounds are known
tem of equations
to the best of our knowledge.
Equations (37) and (38) are Fredholm integral equa-
DN |(x,T ) t = 0,
tions of the second kind (cf. Zemyan [55]). Fortu-
nately, the theory and solution methods for integral t, t = 1,
123
Fast computation of steady-state response 339
where the second equation specifies a unity constraint p(z) = ẋk (0) = 0,
on the length of the tangent vector to uniquely identify asserting that the initial velocity at the kth degree of
t. This tangent vector gives a direction on the solution freedom is zero. As the solution time period T is
curve along which p increases. Starting with a known unknown, Eqs. (84), (85) have an equal number of equa-
solution u0 := (x0 , T0 ) on the solution branch with the tions and variables (x, T ). Thus, there is no parameter
corresponding tangent direction vector t0 and a pre- among the variables, which can be used for the contin-
scribed arc-length step size p, we obtain a nearby uation of a given solution.
solution u by solving the nonlinear system of equations In the literature, this issue is avoided by introducing
a fictitious damping parameter, say d, and establishing
N (u) = 0, the existence of a periodic orbit if and only if d = 0,
u − u0 , t0 = p. i.e., in the conservative limit (Muñoz-Almaraz et al.
[59]). With this trick, we reformulate the integral equa-
This system can again be solved iteratively using, tion (84) (with forcing f(t) ≡ 0 ) to find periodic orbits
e.g., the Newton–Raphson procedure, with the Jaco- of the system
bian given by [DN , t0 ] . We need the derivative of
N (x, T ) with respect to T to evaluate this Jacobian, Mẍ + dKẋ + Kx + S(x) = 0. (86)
which can also be explicitly computed using the deriva-
tive of the Green’s function L or G with respect to T . Periodic solutions to this system are created through a
These expressions are detailed in “Appendix E” and Hopf bifurcation that occurs precisely at the conserva-
implemented in the supplementary MATLAB® code. tive limit of system (86) (d = 0). It can be shown that
Although this pseudo-arc-length continuation is able the damping parameter maintains a zero value along
to capture folds in single-parameter solution branches, the periodic solution-branch obtained at the bifurca-
it is not the state-of-the-art continuation algorithm. tion point. Advanced continuation algorithms can then
More advanced continuation schemes, such as the be used to detect such a Hopf bifurcation and to make
atlas algorithm of coco [37], enable continuation a switch to the periodic solution branch (cf. Galán-
with multi-dimensional parameters required for quasi- Vioque and Vanderbauwhede [60]).
periodic forcing. In this work, we have implemented A similar continuation procedure can be carried out
our integral equations approach with the MATLAB® - in our integral equation approach. The periodic solu-
based continuation package coco [37] . tion, however, technically does not arise from a Hopf
bifurcation, because the non-resonance conditions (7)
and (13) are violated at the Hopf bifurcation point.
J.1.1 Continuation of periodic orbits in the
Nonetheless, continuation of a given solution point on
conservative autonomous setting
the solution branch is possible using standard continu-
ation algorithms. In this conservative autonomous set-
As remarked earlier, conservative autonomous systems
ting, a given solution of system (84), (85) may again be
have internally parameterized periodic orbits a priori
continued in the (x, T, d) space using, e.g., the pseudo-
unknown period. Moreover, each (quasi-) periodic orbit
arc-length approach.
of such a system is part of a family of (quasi-) periodic
trajectories with the solutions differing only in their
phases. A unique solution is obtained using a phase References
condition, a scalar equation of the form
1. Avery, P., Farhat, C., Reese, G.: Fast frequency sweep
p(z(t)) = 0, (85) computations using a multi-point Padé-based reconstruc-
tion method and an efficient iterative solver. Int. J. Numer.
Methods Eng. 69, 2848–2875 (2007). https://doi.org/10.
which fixes a Poincare section transverse to the 1002/nme.1879
(quasi-) periodic orbit in the phase space. The choice 2. Rosenberg, R.M.: The normal modes of nonlinear n-degree-
of-freedom systems. J. Appl. Mech. 30, 7–14 (1962)
of this section is arbitrary but often involves setting the 3. Kelley, A.: On the Liapounov subcenter manifold. J. Math.
initial velocity of one of the degrees of freedom to be Anal. Appl. 18, 472–478 (1967). https://doi.org/10.1016/
zero, i.e., letting 0022-247X(67)90039-X
123
340 S. Jain et al.
4. Shaw, S., Pierre, C.: Normal modes for non-linear vibratory 22. Vakakis, A., Cetinkaya, C.: Analytic evaluation of periodic
systems. J. Sound Vib. 164, 85–124 (1993) responses of a forced nonlinear oscillator. Nonlinear Dyn.
5. Avramov, K.V., Mikhlin, Y.V.: Nonlinear normal modes 7, 37–51 (1995)
for vibrating mechanical systems. Rev. Theor. Dev. ASME 23. Kryloff, N., Bogoliuboff, N.: Introduction to Non-linear
Appl. Mech. Rev. 65, 060802 (2010) Mechanics. Princeton University Press, Princeton (1949)
6. Haller, G., Ponsioen, S.: Nonlinear normal modes and 24. Chua, L., Ushida, A.: Algorithms for computing almost
spectral submanifolds: existence, uniqueness and use in periodic steady-state response of nonlinear systems to
model reduction. Nonlinear Dyn. 86, 1493–1534 (2016). multiple input frequencies. IEEE Trans. Circuits Syst.
https://doi.org/10.1007/s11071-016-2974-z 28(10), 953–971 (1981)
7. Craig, R., Bampton, M.: Coupling of substructures for 25. Lau, S., Cheung, Y.: Incremental harmonic balance method
dynamic analysis. AIAA J. 6(7), 1313–1319 (1968). https:// with multiple time scales for aperiodic vibration of
doi.org/10.2514/3.4741 nonlinear systems. J. Appl. Mech. 50, 871–876 (1983)
8. Theodosiou, C., Sikelis, K., Natsiavas, S.: Periodic steady 26. Leipholz, H.: Direct Variational Methods and Eigenvalue
state response of large scale mechanical models with local Problems in Engineering. Vol. Mechanics of Elastic
nonlinearities. Int. J. Solids Struct. 46, 3565–3576 (2009). Stability, 5. Leyden: Noordhoff, (1977)
https://doi.org/10.1016/j.ijsolstr.2009.06.007 27. Bobylev, N.A., Burman, Y.N., Korovin, S.K.: Approxima-
9. Kosambi, D.: Statistics in function space. J. Indian Math. tion Procedures in Nonlinear Oscillation Theory. Walter de
Soc. 7, 76–78 (1943) Gruyter, Berlin (1994)
10. Kerschen, G., Golinval, J., Vakakis, A.F., Bergman, L.A.: 28. Urabe, M.: Galerkin’s procedure for nonlinear periodic
The method of proper orthogonal decomposition for dynam- systems. Arch. Ration. Mech. Anal. 20(2), 120–152 (1965)
ical characterization and order reduction of mechanical 29. Stokes, A.: On the approximation of nonlinear oscillations.
systems: an overview. Nonlinear Dyn/ 41, 147–169 (2005). J. Diff. Equ. 12, 535–558 (1972)
https://doi.org/10.1007/s11071-005-2803-2 30. García-Saldaña, J.D., Gasull, A.: A theoretical basis for the
11. Amabili, M.: Reduced-order models for nonlinear vibra- harmonic balance method. J. Differ. Equ. 254, 67–80 (2013)
tions, based on natural modes: the case of the circular 31. Keller, H.B.: Numerical Methods for Two-point Boundary-
cylindrical shell. Philos. Trans. R. Soc. A 371, 20120474 value Problems. Blaisdell, Waltham (1968)
(2013). https://doi.org/10.1098/rsta.2012.0474 32. Peeters, M., Viguié, R., Sérandour, G., Kerschen, G.,
12. Touzé, C., Vidrascu, M., Chapelle, D.: Direct finite element Golinval, J.C.: Nonlinear normal modes, Part II: toward a
computation of non-linear modal coupling coefficients for practical computation using numerical continuation tech-
reduced-order shell models. Comput. Mech. 54, 567–580 niques. Mech. Syst. Signal. Process. 23, 195–216 (2009)
(2014). https://doi.org/10.1007/s00466-014-1006-4 33. Sracic, M., Allen, M.: Numerical Continuation of Periodic
13. Idelsohn, S.R., Cardona, A.: A reduction method for Orbits for Harmonically Forced Nonlinear Systems. Civil
nonlinear structural dynamic analysis. Comput. Methods Engineering Topics, Volume 4: Proceedings of the 29th
Appl. Mech. Eng. 49(3), 253–279 (1985). https://doi.org/ IMAC, 51– 69 (2011)
10.1016/0045-7825(85)90125-2 34. Seydel, R.: Practical Bifurcation and Stability Analysis.
14. Sombroek, C.S.M., Tiso, P., Renson, L., Kerschen, G.: Springer, New York (2010). ISBN 978-1-4419-1739-3
Numerical computation of nonlinear normal modes in 35. Doedel, E., Oldeman, B.: Auto-07p: Continuation and
a modal derivatives subspace. Comput. Struct. 195, 34– Bifurcation Software for ordinary differential equations,
46 (2018). https://doi.org/10.1016/j.compstruc.2017.08. urlhttp://indy.cs.concordia.ca/auto/
016 36. Dhooge, A., Govaerts, W., Kuznetsov, Y.: Matcont: a
15. Jain, S., Tiso, P., Rixen, D.J., Rutzmoser, J.B.: A quadratic MATLAB package for numerical bifurcation analysis of
manifold for model order reduction of nonlinear structural odes. ACM Trans. Math. softw. 29(2), 141–164 (2003)
dynamics. Comput. Struct. 188, 80–94 (2017). https://doi. 37. Dankowicz, H., Schilder, F.: Recipes for Continuation,
org/10.1016/j.compstruc.2017.04.005 SIAM (2013). ISBN 978-1-611972-56-6. https://doi.org/
16. Malookani, R.A., van Horssen, W.T.: On resonances and the 10.1137/1.9781611972573
applicability of Galerkin’s truncation method for an axially 38. Renson, L., Kerschen, G., Cochelin, B.: Numerical
moving string with time-varying velocity. J. Sound Vib. computation of nonlinear normal modes in mechanical
344, 1–17 (2015). https://doi.org/10.1016/J.JSV.2015.01. engineering. J. Sound Vib. 364, 177–206 (2016). https://
051 doi.org/10.1016/j.jsv.2015.09.033
17. Nayfeh, A.H.: Perturbation Methods. Wiley, Hoboken 39. Géradin, M., Rixen, D.J.: Mechanical Vibrations: Theory
(2004) and Application to Structural Dynamics, 3rd edn. Wiley,
18. Nayfeh, A.H., Mook, D.T., Sridhar, S.: Nonlinear analysis Chichester (2015). ISBN 978-1-118-90020-8
of the forced response of structural elements. J. Acoust. 40. Burd, V.M.: Method of Averaging for Differential Equations
Soc. Am. 55(2), 281–291 (1974) on an Infinite Interval: Theory and Applications. Chapman
19. Mitropolskii, YuA, Van Dao, N.: Appl. Asymptot. Methods and Hall, London (2007). Chapter 2
Nonlinear Oscil. Springer, New York (1997) 41. Schilder, F., Vogt, W., Schreiber, S., Osinga, H.M.: Fourier
20. Bogoliubov, N., Mitropolsky, Y.: Asymptotic Methods in methods for quasi-periodic oscillations. Int. J. Numer.
the Theory of Nonlinear Oscillations. Gordon and Breach Methods Eng. 67, 629–671 (2006). https://doi.org/10.1002/
Science Publication, New York (1961) nme.1632
21. Sanders, J.A., Verhulst, F.: Averaging Methods in Nonlinear 42. Mondelo González, J.M.: Contribution to the study of
Dynamical Systems. Springer, New York (1985) Fourier methods for quasi-periodic functions and the vicin-
123
Fast computation of steady-state response 341
ity of the collinear libration points. Ph.D. Thesis, University 55. Zemyan, S.M.: The Classical Theory of Integral Equations:
of Barcelona (2001). http://hdl.handle.net/2445/42084 A Concise Treatment. Birkhäuser, Basel (2012). https://doi.
43. Kovaleva, A.: Optimal Control of Mechanical Oscillations. org/10.1007/978-0-8176-8349-8. ISBN 978-0-8176-8348-
Springer, New York (1999) 1
44. Babistkiy, V.I.: Theory of vibro-impact systems and 56. Atkinson, K.: The Numerical Solution of Integral Equa-
applications. Springer, Berlin (1998) tions of the Second Kind (Cambridge Monographs on
45. Rosenwasser, E.N.: Oscillations of Non-linear Systems: Applied and Computational Mathematics). Cambridge
Method of Integral Equations. Nauka, Moscow (1969). (in University Press, Cambridge (1997). https://doi.org/10.
Russian) 1017/CBO9780511626340
46. Babitsky, V.I., Krupenin, V.L.: Vibration of Strongly 57. Szalai, R., Ehrhardt, D., Haller, G.: Nonlinear model
Nonlinear Discontinuous Systems. Springer, Berlin (2012) identification and spectral submanifolds for multi-degree-
47. Kelley, A.F.: Analytic two-dimensional subcenter manifolds of-freedom mechanical vibrations. Proc. R. Soc. A. 473,
for systems with an integral. Pacific J. of Math. 29, 335–350 20160759 (2017)
(1969) 58. Breunung, T., Haller, G.: Explicit backbone curves
48. Arnold, V.I.: Mathematical Methods of Classical Mechan- from spectral submanifolds of forced-damped nonlinear
ics. Springer, New York (1989) mechanical systems. Proc. R. Soc. A. 474, 20180083 (2018)
49. Picard, E.: Traité d’analyse, vol. 2. Gauthier-Villars, Paris 59. Muñoz-Almaraz, F.J., Freire, E., Galán-Vioque, J., Doedel,
(1891) E., Vanderbauwhede, A.: Continuation of periodic orbits in
50. Bailey, P., Shampine, L., Waltman, P. (Eds.): Nonlinear Two conservative and Hamiltonian systems. Phys. D: Nonlinear
Point Boundary Value Problems, Mathematics in Science Phenomena 181(1), 1–38 (2003). https://doi.org/10.1016/
and Engineering, Vol. 44, (1968) S0167-2789(03)00097-6
51. Kelley, C.T.: Solving Nonlinear Equations with Newton’s 60. Galán-Vioque, J., Vanderbauwhede, A.: Continuation of
Method. SIAM, New York (2003). ISBN: 978-0-89871- periodic orbits in symmetric Hamiltonian systems. In:
546-0. https://doi.org/10.1137/1.9780898718898 Krauskopf, B., Osinga, H.M., Galán-Vioque, J. (eds.)
52. Broyden, C.G.: A class of methods for solving Numerical Continuation Methods for Dynamical Sys-
nonlinear simultaneous equations. Math. Comput. tems. Springer, Dordrecht (2007). https://doi.org/10.1007/
19(92), 577–593 (1965). https://doi.org/10.1090/ 978-1-4020-6356-5_9
S0025-5718-1965-0198670-6
53. Haro, A., de la Llave, R.: A parameterization method
for the computation of invariant tori and their whiskers in Publisher’s Note Springer Nature remains neutral with regard
quasi-periodic maps: rigorous results. J. Differ. Equ. 228(2), to jurisdictional claims in published maps and institutional affil-
530–579 (2006). https://doi.org/10.1016/j.jde.2005.10.005 iations.
54. Kress, R.: Linear Integral Equations, Chapter 13, 3rd
edn. Springer, New York (2014). ISBN 978-1-4612-
6817-8
123