Comparison of two algorithms for Harmonic Balance

and path continuation

Lukas Woiwode, Nidish Narayanaa Balaji, Jonas Kappauf, Fabia Tubita, Louis
Guillot, Christophe Vergez, Bruno Cochelin, Aurélien Grolet, Malte Krack

To cite this version:

Lukas Woiwode, Nidish Narayanaa Balaji, Jonas Kappauf, Fabia Tubita, Louis Guillot, et al.. Com-
parison of two algorithms for Harmonic Balance and path continuation. Mechanical Systems and
Signal Processing, 2020, 136, pp.106503. �10.1016/j.ymssp.2019.106503�. �hal-02424746�

HAL Id: hal-02424746

https://hal.science/hal-02424746
Submitted on 28 Dec 2019

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
*Manuscript
Click here to view linked References

1
2
3 Comparison of two algorithms for Harmonic Balance and path continuation
4
5
6 Lukas Woiwoded , Nidish Narayanaa Balajie , Jonas Kappaufc , Fabia Tubitaf , Louis Guillotb , Christophe
7 Vergezb , Bruno Cochelinb , Aurélien Groleta , Malte Krackd
8 a Ecole Nationale Supérieure d’Arts et Métiers Lille - LISPEN, Lille, FRANCE
9
b Aix Marseille University, CNRS, Centrale Marseille, LMA, Marseille, FRANCE
10
11 c University of Kassel, Kassel, GERMANY
12 d University of Stuttgart, Stuttgart, GERMANY
13
e Rice University, Houston TX, USA
14
15 f Ecole Centrale de Lyon, Lyon, FRANCE
16
17
18
19
20
21 Abstract
22
23 In this work we apply and compare two algorithms for setting up Harmonic Balance equations and numerical
24 continuation of the solution path for harmonically driven mechanical systems. The first algorithm relies
25 on a predictor-corrector scheme and an Alternating Frequency-Time approach (AFT-PreCo). The second
26 algorithm relies on a high-order Taylor series expansion of the solution path (asymptotic numerical method)
27 and classical Harmonic Balance formulated entirely in the frequency domain (cHB-ANM). We conclude
28 that the cHB-ANM is suited for smooth nonlinearities, for instance geometrically nonlinear finite element
29 models. Here, cHB-ANM avoids aliasing errors and convinces with a numerically robust adjustment of the
30 continuation step length and a continuous representation of the solution path. For non-smooth nonlinearities
31 such as stick-slip friction or unilateral constraints, AFT-PreCo is better suited, and convinces with high
32 numerical robustness and efficiency.
33
34 Keywords: Harmonic Balance Method, non-linear vibration, periodic solution, continuation method,
35 friction damping, geometric nonlinearity
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59 To be submitted October 28, 2019
60
61
62
63
64
65
 1
2
3
4 Symbols and abbreviations
5 Latin letters Greek letters
6 ARMS root-mean-square value of the amplitude α step length
7 C vector of constant functions αnom nominal step length (AFT-PreCo)
8 D coefficient matrix of δ Dirac delta distribution
9 velocity-proportional forces εα ANM threshold
10 E HN forward discrete Fourier transform matrix εω frequency error at resonance peak
11
(N samples, H harmonics) εA amplitude error at resonance peak
12
13 E −1
NH inverse discrete Fourier transform matrix εreg regularization parameter
14 (N samples, H harmonics) εtol error tolerance (algebraic equations)
15 f generic function η normalized angular frequency
16 f ex vector of excitation forces κ linear stiffness of Coulomb slider
17 f nl vector of nonlinear forces ρ friction limit force
18 fu unilateral spring force τ normalized time
19 fρ friction force wj weight functions
20 g coordinate of Coulomb slider Ω fundamental angular osc. frequency
21 H harmonic truncation
√ order
22 i imaginary unit −1 Sub-, superscripts, operators
23 J X0 Jacobian at expansion point X 0 || magnitude
24 K coefficient matrix of kk norm (Euclidian if not specified)
25 coordinate-proportional forces  averaged value
26 L vector of linear functions ˜ vector of time samples
27 ˙
M coefficient matrix of  derivative with respect to time
28
29 acceleration-proportional forces 0 derivative with respect to
30 N number of time samples per period normalized time
31 NDOF number of degrees of freedom ˆH vector of Fourier coefficients
32 Nemp empirical rule for N (H harmonics)
33 Neq number of equations ˆ
(0) 0-th Fourier coefficient
34 Nfact number of Jacobian factorizations ˆ c (k) k ≥ 1, k-th Fourier coefficient of
35 Nnewt number of Newton iterations cosine function
36 Npt number of continuation steps ˆ s (k) k ≥ 1, k-th Fourier coefficient of
37 P Taylor series order sine function
38 q vector of generalized coordinates
39
Q vector of bilinear functions Abbreviations
40
r residual of differential equation system AFT Alternating Frequency-Time (scheme)
41
42 R residual of algebraic equation system ANM asymptotic numerical method
43 Raux residual of auxiliary equations cHB classical Harmonic Balance (method)
44 Rtot Concatenation of R and Raux DFT discrete Fourier transform
45 S dynamic stiffness matrix HB Harmonic Balance (method)
46 tcomp computation time iFFT/FFT inverse-/Fast Fourier Transform
47 T period of oscillation PreCo predictor-corrector (method)
48 x vector of unknowns SDOF single degree of freedom (system)
49 xaux vector of auxiliary unknowns MDOF multiple degree of freedom (system)
50 X x extended by free parameter
51 (excitation frequency)
52 X0 Taylor series expansion point
53 Xp Taylor series coefficients
54
55
56
57
58
59 2
60
61
62
63
64
65
 1
2
3 1. Introduction
4
5 Harmonic Balance (HB ) [1–3] is a method for approximating periodic solutions of nonlinear ordinary
6 differential and differential-algebraic equations. HB can be interpreted as a Fourier-Galerkin method, where
7 the (time-)dependent variables are represented by a truncated Fourier series, and the residual is required to
8 be orthogonal to the Fourier basis functions [4]. This leads to an algebraic equation system for the sought
9 Fourier coefficients of the approximation. HB avoids the costly computation of possibly long transients, and
10
reasonable accuracy is often achieved already for a low truncation order. This permits to reduce the com-
11
putational effort often by several orders of magnitude compared to numerical integration of the initial value
12
13 problem [3]. Consequently, HB greatly gained importance in the field of (nonlinear) mechanical systems.
14 For example, vibration problems of jointed structures with frictional contact interactions were thoroughly
15 analyzed with HB, e. g. [5–9]. Besides, there are various review articles addressing the application of HB
16 to a particular realm, e. g. on rotor dynamics [10] and brake squeal [11]. It is, however, not limited to
17 vibration problems in mechanical systems, HB is also applied in other disciplines such as fluid dynamics [12]
18 and electrical circuits [13].
19 A crucial task within HB is to determine the Fourier coefficients of the nonlinear terms. In some cases, in
20 particular in the case of polynomial terms, an algebraic expansion is possible using trigonometric identities,
21 or the continuous integral for the Fourier coefficients can be expressed in closed form. This is called classical
22 HB (cHB ). A popular and more versatile alternative is to approximate the continuous integral using the
23 discrete Fourier transform, which leads to the Alternating Frequency-Time (AFT ) HB [3, 14]. An important
24 parameter of this method is the number of time samples per oscillation period. In the special case of poly-
25 nomial nonlinearities, the method is exact beyond a certain number of samples. Otherwise aliasing errors
26 occur.
27 It is often of interest to determine how the solution evolves with a free parameter. For this task, numerical
28 path continuation is commonly applied. The task of numerical path continuation methods is to recover the
29
solution branch accurately and overcome potential turning points with respect to the free parameter. Par-
30
ticularly popular is the predictor-corrector (PreCo) method. From the current solution point, a prediction
31
32 is made, for instance by going a certain step length in the direction of the tangent to the solution branch. A
33 Newton-type method is then usually applied to iteratively correct the prediction and return to the solution
34 branch. An interesting alternative is the asymptotic numerical method (ANM ) [15–18]. Here, the solution
35 path is expanded into a high-order Taylor series with respect to an arc length parameter. By remaining
36 within the range of utility of the truncated power series, correction steps can be avoided. A challenge within
37 the ANM is to efficiently calculate the high-order Taylor series coefficients. For this, the algebraic equation
38 system is commonly recast to nonlinear terms of only quadratic order, by introducing auxiliary variables
39 and additional equations.
40 The ANM can only be applied when the derivatives for the Taylor series expansion exist, i. e., it is only appli-
41 cable to sufficiently smooth nonlinearities. However, many mechanical systems are modeled with non-smooth
42 force-deformation relations, such as systems with stick-slip friction, unilateral constraints or materials un-
43 dergoing phase transformations. A regularization is then usually required to approximate the non-smooth
44 relations by smooth ones in order to apply the ANM.
45 In this work, we focus on two algorithms, AFT-PreCo and cHB-ANM. The goal is to identify and understand
46 the individual strengths and weaknesses of these algorithms, both with regard to computational efficiency
47 and robustness, as well as required level of experience to properly select critical method parameters. We
48
analyzed various benchmark problems of harmonically driven mechanical systems, with either smooth or
49
non-smooth nonlinearities, and different numbers of degrees of freedom. In Section 2, we recap the compu-
50
51 tational methods and the theoretical background required to understand their similarities and differences.
52 In Section 3, we present selected results of representative benchmark problems. In Section 4, we develop
53 guidelines towards the recommended application range of either method.
54
55
56
57
58
59 3
60
61
62
63
64
65
 1
2
3 2. Recap of Harmonic Balance and continuation methods
4
5 2.1. Harmonic Balance
6
In this work, we consider mechanical systems described by the equation of motion
7
8
r(t) = M q̈(t) + D q̇(t) + Kq(t) + f nl (q(t)) − f ex (t) = 0 , (1)
9
10
where q ∈ RNDOF ×1 is the vector of generalized coordinates, overdots denote differentiation with respect
11
to time t, M , D, K ∈ RNDOF ×NDOF are coefficient matrices of acceleration-, velocity-, displacement-
12
13 proportional forces, and f nl , f ex ∈ RNDOF ×1 are nonlinear and external forces, respectively. For ease of
14 notation, we assume that f nl depends only on q; it is straight-forward to extend the following developments
15 to the case where it depends also on q̇ (and higher-order derivatives with respect to time [3]). We consider
16 T -periodic external forces, f ex (t + T ) = f ex (t), and seek T -periodic responses q(t). The requirement that
17 q(T ) = q(0) where q(t) has to satisfy Eq. (1) corresponds to a periodic boundary value problem on the
18 domain t ∈ [0, T ].
19
20 2.1.1. Method of mean weighted residuals, Fourier methods
21 The methods we apply in this work to solve the periodic boundary value problem can be interpreted as
22 methods of mean weighted residuals, with a truncated Fourier series as ansatz,
23
24 H
X
25 q(t) = q̂(0) + q̂ c (k) cos(kΩt) + q̂ s (k) sin(kΩt) , (2)
26 k=1
27
28 with the angular frequency Ω = 2π/T and the harmonic truncation order H. Substitution of the ansatz
29 (2) into Eq. (1) generally produces a non-zero residual r(t) 6= 0. The unknown Fourier coefficients are
30 determined as the solution of the algebraic equation system
31
32 ZT
1
33 r(t)wj (t)dt = 0 j = 1, . . . , 2H + 1 , (3)
T
34 0
35
36 which requires that the residual vanishes in a weighted average sense, with the weight functions wj (t);
37 i. e., the residual is orthogonal to the truncated Fourier basis. When the 2H + 1 Fourier basis functions
38 {1, cos(Ωt), . . . , sin(HΩt)} are used as weights, Eq. (3) requires that the Fourier coefficients of the residual
39 associated with the sine-/cosine vanish up to order H; i. e., r̂(0) = r̂ c (1) = . . . = r̂ s (H) = 0. This algorithm
40 was already introduced in 1965 by Urabe [4] and is now commonly known as Harmonic Balance (HB). HB
41 can thus be interpreted as a Galerkin method where Fourier basis functions are used both as ansatz and
42 weight functions. When Dirac delta distributions δ(t − tj ) with the shifts tj = jT /N , j = 0, . . . , N − 1 are
43 used as weights, Eq. (3) requires that the residual vanishes at these collocation points; i. e., r(t0 ) = . . . =
44 r(tN −1 ) = 0. Together with the Fourier ansatz in Eq. (2), this is known as Trigonometric Collocation.
45 When we stack all Fourier coefficients of q, r in vectors q̂ H , r̂ H , we can summarize the HB equations as
46
47 r̂ H (q̂ H ) = S(Ω)q̂ H + f̂ nl,H − f̂ ex,H = 0 , (4)
48
49
where S(Ω)q̂ H accounts for the linear terms in Eq. (1) with the dynamic stiffness matrix S, and f̂ nl,H ,
50
51 f̂ ex,H are the vectors of Fourier coefficients of the nonlinear and external forces, respectively.
52
53 2.1.2. Classical and Alternating-Frequency-Time Harmonic Balance
54 A challenge within HB is to determine the Fourier coefficients of the nonlinear terms. In some cases,
55 for example when f nl is a polynomial in q (and more generally also in q̇), closed-form expressions can be
56 established, either by algebraic expansion of f nl into a Fourier series (using trigonometric identities), or by
57 analytical integration of the Fourier coefficients (e. g. using discrete convolutions). In general, the discrete
58
59 4
60
61
62
63
64
65
 1
2
3 Fourier transform (DFT) can be useful to approximate the continuous integrals for the Fourier coefficients,
4 e. g.,
5
6 ZT N −1
1 1 X
7 f̂ nl,c (k) = f nl (q (t)) cos(kΩt)dt ≈ f (q(tj )) cos(kΩtj ) , (5)
8 T N j=0 nl
0
9
10 where q(tj ) is evaluated in accordance with Eq. (2) (sampled) at the equidistant time instants tj = jT /N .
11 When we stack all samples of q(tj ) in a vector q̃ N , we can summarize this as
12 f̂ nl,H ≈ E HN f̃ nl,N E −1


HN q̂ H , (6)
13
14 where E HN is the DFT matrix (for approximation of the Fourier coefficients from the samples), and E −1 HN
15 is the inverse DFT matrix (for interpolation of the trigonometric polynomial). This procedure was also
16 already proposed by Urabe [19] and is now commonly known as Alternating-Frequency-Time scheme. For
17 a given frequency-domain representation of the approximation, q̂ H , the time samples q(tj ) are determined
18 first. The nonlinear forces f nl are then evaluated in the time domain. Finally, the Fourier coefficients f̂ nl,H
19
are approximated using the DFT.
20
In some cases, the sampling procedure in Eq. (6) (and Eq. (5)) is exact; i. e., the DFT yields the exact value
21
22 of the continuous integral for the Fourier coefficients. This is the case if f nl , evaluated at the Fourier ansatz
23 (2), is a Fourier series of finite truncation order Hf and N ≥ 2Hf + 1. This, in turn, happens if f nl is a
24 polynomial in q (and more generally also in q̇). This is shown in detail in Appendix A. The polynomial
25 can be broken down to sums of products. A product of two Fourier series, truncated to orders H1 and H2 ,
26 respectively, is again a Fourier series, with truncation order H1 + H2 . Moreover, any linear operation and
27 differentiation with respect to time yields a Fourier series with the same truncation order. The Fast Fourier
28 Transform (FFT) is a computationally efficient DFT algorithm. Using the AFT scheme can in fact be the
29 fastest and still exact opportunity to calculate f̂ nl,H even in the case of polynomials [20]. If the number
30 of samples, N , is not sufficiently high (N < 2Hf + 1), aliasing errors occur. Aliasing is inevitable if the
31 considered periodic function has an infinite sequence of non-zero Fourier coefficients (no finite truncation
32 order), see e. g. [3].
33 To summarize, the equation systems of classical HB (cHB) and Alternating-Frequency-Time (AFT) HB
34 are:
35
36 cHB: r̂ H (q̂ H ) = 0, (7)
37 AFT HB: E HN r̃ N E −1


HN q̂ H = 0. (8)
38
39 We here utilized that the sampling procedure is exact for the linear terms in Eq. (1). For large N , the
40 aliasing error typically becomes negligible compared to other errors (e. g. truncation error, tolerances of
41 numerical root finding methods), and the solution of Eq. (8) should converge to the solution of Eq. (7).
42
43 2.1.3. Another interpretation of AFT HB
44 We can split Eq. (1) into the differential-algebraic equation system,
45
46 0 = M q̈ + D q̇ + Kq + v − f ex , (9)
47 0 = v − f nl (q) , (10)
48 by introducing the auxiliary variables v. Suppose that we apply cHB with truncation order Hq only to the
49
differential part (9), and Trigonometric Collocation with N = 2Hv +1 collocation points to the algebraic part
50
51 (10). cHB applied to Eq. (9) simply leads to S(Ω)q̂ Hq + v̂ Hq − f̂ ex,Hq . Trigonometric Collocation applied to
   
52 Eq. (10) leads to E −1
Hv N v̂ H v
= f̃ nl,N E −1
Hq N q̂ H q
. From this, we can follow v̂ H v
= E H v N f̃ nl,N E −1
Hq N q̂ H q
.
53  
54 For Hv ≥ Hq one can truncate this to v̂ Hq = E Hq N f̃ nl,N E −1 Hq N q̂ Hq , which is fully equivalent to Eq. (8).
55 Thus, AFT HB can be interpreted as applying classical Harmonic Balance to the linear part of the equation
56 of motion and Trigonometric Collocation to the nonlinear terms, but generally distinct truncation orders
57 Hq 6= Hv for the two sets of variables.
58
59 5
60
61
62
63
64
65
 1
2
3 2.2. Numerical path continuation
4
5 It is often of interest to determine how the periodic solution evolves under variation of a free parameter.
6 In this work, we focus on the excitation frequency Ω as free parameter, i. e., we analyze the frequency
7 response. Thus, we are interested in the solutions X = [xT , Ω]T of the algebraic equation system
8
9 R(X) = 0 , (11)
10
11 where R, x ∈ RNeq ×1 are the residual vector and the vector of unknowns, Neq is the number of equations
12 (and unknowns). The solutions form continuous branches under variation of the free parameter. We can
13 directly consider the equation systems (7)-(8) and set x = q̂ H and e. g. R = r̂ H for cHB. The task of
14 numerical path continuation methods is to recover the solution branch accurately and overcome potential
15 turning points with respect to the free parameter (cf. Fig. 1a).
16
17 2.2.1. Predictor-Corrector (PreCo) method
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37 Figure 1: Illustration of continuation methods: (a) solution branch, (b) predictor-corrector (PreCo) method, (c) asymptotic
38 numerical method (ANM)
39
40
41 Suppose we already know a solution point X 0 . To obtain a next point on the solution branch, we can make
42 a prediction,
43
44 X(α) ≈ X 0 + αX 1 , (12)
45
in the direction of the tangent to the solution branch, X 1 , with the step length α. The tangent with unit
46
47 length is defined by
48 J X0
49 z }| {
50 ∂R 
X1 = 0 , XT
1 X1 = 1 . (13)
51 ∂X X 0
52
53 X 1 is well-defined if the Jacobian J X 0 has rank Neq ; then X 0 is called a regular solution point. We usually
54 have R(X 0 + αX 1 ) 6= 0, and a correction is needed to decrease the residual to a tolerable level and obtain
55 an acceptable next solution point. Once the next solution point is found, the current continuation step is
56 completed, and the process repeats from this point, until a termination criterion is met (e. g. the parameter
57
58
59 6
60
61
62
63
64
65
 1
2
3 end value is reached). It is common to introduce an additional equation, which defines where along the
4 solution branch the next point is located. A popular choice is [21–23]
5
6 
R (X)

7 Rtot (X) = T = 0. (14)
(X − X 0 ) (X − X 0 ) − α2
8
9 Here, the last equation requires that the next solution point lies on the hypersphere of radius α around the
10 previous solution point X 0 (cf. Fig. 1b). The number of equations is now equal to that of the unknowns, and
11 a Newton-type procedure is commonly applied, which corresponds to the iterative solution of the equations
12
13

∂Rtot     
14 X (j+1) − X (j) = −Rtot X (j) , (15)
∂X X (j)

15
16  
17 starting from the initial guess X (0) = X 0 + αX 1 , until kRtot X (j) k ≤ εtol . For well-behaved problems,
18 the continuation step length α can be selected such that typically only 3 to 10 Newton iterations are needed.
19 The required derivatives ∂Rtot /∂X are commonly calculated analytically. This is straight-forward with cHB
20 where a closed-form expression is available for the residual as function of the unknowns. This is also easy
21 with the AFT scheme thanks to the linearity of forward and inverse DFT [24]. Various other predictors and
22 path parametrization strategies are possible, e. g. normal flow, the interested reader is referred to [3, 23],
23 and also [25] where the performance of different parametrizations is compared.
24
25
26 2.2.2. Asymptotic numerical method (ANM)
27
The predictor in Eq. (12) of the PreCo method can be viewed as first-order part of a Taylor series
28
expansion around X 0 . The ANM extends this to a higher-order predictor (order P ),
29
30 P
X
31 X(α) ≈ αp X p , (16)
32 p=0
33
34 where X p are the series coefficients. The domain of utility [0, αmax ] of the expansion in Eq. (16) can be
35 determined. To this end, the residual R is also expanded in a Taylor series. It is then assumed that the
36 residual is dominated by the (P + 1)-th order term, which yields
37
38   P 1+1
εα
39 αmax = , (17)
kRP +1 k
40
41 where εα is a user-defined tolerance (ANM threshold ) and RP +1 is the (P + 1)th-order coefficient of the
42 Taylor series expansion of the residual R [17, 26]. The idea behind this is to choose the continuation step
43
length αmax in such a way that the residual remains within the specified tolerance εtol , with the intent to
44
completely avoid correction iterations. Ideally, the ANM threshold εα should be set to the tolerance εtol .
45
46 Since the expansion point already has a non-zero residual in practice, and Eq. (17) is only an approximation
47 of the true radius of convergence (truncating the residual to the (P + 1)th-order term of its Taylor series),
48 kR(X(αmax ))k > εtol in general, such that correction steps are occasionally needed. To avoid this, the ANM
49 threshold εα is commonly set to a smaller value, εα < εtol . This is further discussed in the result sections
50 of this paper. Once αmax is determined, X(αmax ) is evaluated, which then serves as the next expansion
51 point and so forth, cf. Fig. 1c. The ANM thus yields a section-wise power series expansion of the solution
52 path, which together form an accurate continuous representation of the entire solution branch. The power
53 series expansion of the solution path can be useful for bifurcation analysis [27]. Experience has shown that
54 for most problems, an order P of around 20 yields reasonable computational performance [17], which is a
55 compromise between the effort of computing higher order terms and the reward of having an enlarged radius
56 of convergence, and this value was selected throughout this study.
57 An essential ingredient of the ANM is a procedure to compute the series coefficients X p . For this purpose,
58
59 7
60
61
62
63
64
65
 1
2
3 automatic differentiation can be used. However, this is known to be computationally costly when more than
4 a few hundreds of equations are considered [28]. The computation of the series coefficients is particularly
5
efficient when the algebraic equation system is in quadratic form,
6
7 R(X) = C + L(X) + Q(X, X) , (18)
8
9 where C, L, Q are constant, linear and bilinear (quadratic) functions. For a given solution point X 0 and
10 tangent X 1 determined as in Eq. (13), the higher-order series coefficients can then be determined successively
11 from the linear equations (see e. g. [15])
12
13 p−1
X
14 J X0 X p + Q(X i , X p−i ) = 0 p = 2, . . . , P . (19)
15 i=1
16
Note that the sought X p only occur in the first term, while the sum only involves X i with i < p.
17
18 The equation systems (7)-(8) are generally not in quadratic form. However, if f nl is an analytic function in
19 q (and more generally also in q̇), the equation system can usually be recast into quadratic form. To this
20 end, extra unknowns xaux are introduced besides the main unknowns q̂ H , along with additional equations
21 Raux = 0,
22    
q̂ H r̂ H
23 x= , R= = 0. (20)
24 xaux Raux
25 This recast is particularly simple if the nonlinear terms, f nl , in the equation of motion are polynomials
26 or rational functions [15]. One can then introduce auxiliary variables v defined by algebraic equations
27
r aux (q, v) = 0 in such a way that the differential-algebraic equation system contains only quadratic nonlinear
28
terms in q and v. Both q and v are then represented by truncated Fourier series. Application of HB (either
29
30 the cHB or the AFT variant) leads to a quadratic algebraic equation system with regard to the unknowns
31 q̂ H , v̂ H . The described formalism has been extended to some transcendental functions [16, 17], but this is
32 not relevant for the benchmark problems within the present study.
33 Note that the X p are the solutions of linear equation systems (19) that share the same coefficient matrix,
34 J X 0 , for each p. Hence, this matrix only needs to be factorized once per expansion point X 0 . Moreover,
35 if one defines every new auxiliary variable only by the main and lower-index auxiliary variables, it can be
36 shown that the Jacobian block associated to the extra unknowns xaux is triangular (and thus its inverse
37 is easy to determine). A block elimination can then be applied to solve Eq. (19) such that the critical
38 problem dimension reduces to that of the main variables [17]. For the efficient evaluation of C, L and Q
39 in large-dimensional problems, a sparse tensorial formalism can be used. The associated tensor lists can
40 be generated automatically using polarization formulae. The user then only has to define the equation of
41 motion (1), and the algebraic equations r aux (q, v) = 0 defining the auxiliary variables v [17].
42 HB theoretically does not fail if f nl is only differentiable up to a certain order (C m function with m < ∞),
43 i. e. representing a non-smooth function. However, the convergence of the Fourier series is generally slower
44 than in the case of analytic functions [3]. Urabe [4] established sufficient conditions for the convergence of
45 HB to an exact solution. Here, the nonlinear forces, along with their first-order derivatives with respect
46
to q (and more generally also q̇), were only required to be continuously differentiable. As one can easily
47
see from the left part of Eq. (5), finite differentiability of f nl with respect to q implies (by the chain rule)
48
49 finite differentiability of R(X(α)) with respect to α. Hence, the ANM can only be applied to a certain
50 order. More importantly, cHB can usually not be applied, and the system cannot be brought into quadratic
51 form. To apply cHB and the ANM properly, a regularization is needed, where the non-smooth functions
52 are approximated by analytic ones. This introduces additional errors and comes with the task to properly
53 choose the parameter(s) of the regularization.
54
55 2.3. Morphological box of methods
56 In principle, any continuation algorithm (PreCo method vs. ANM) can be combined with any Fourier
57 method (cHB vs. AFT scheme). On the other hand, the ANM only makes sense if one has an efficient
58
59 8
60
61
62
63
64
65
 1
2
3 procedure for computing the required Taylor series coefficients up to high order. Here, the aforementioned
4 recast of the equation of motion (1) into a differential-algebraic equation system with auxiliary variables v
5
is particularly useful. Both the main and the auxiliary variables are then approximated by H-order Fourier
6
series. To evaluate the residual of the quadratically recast system in Eq. (18), one has to determine the
7
8 Fourier coefficients of the nonlinear terms. This can, in principle, also be done using the discrete Fourier
9 transform. For at most quadratic nonlinear terms, the discrete Fourier transform yields the exact value of
10 the Fourier coefficients for N ≥ 3H + 1, as shown in Appendix A. In this sense, cHB and the AFT variant
11 (applied to the quadratically recast problem) are theoretically fully equivalent. However, the numerical
12 implementation of the nonlinear terms differs (discrete convolutions within cHB vs. FFT/iFFT within the
13 AFT scheme), leading to different computational effort in general.
14 In many vibration problems, the generalized coordinates q are dominated by only one or a few harmonics.
15 In contrast, more harmonics are commonly required to represent the nonlinear terms f nl in the equation
16 of motion Eq. (1), where f nl depends on the generalized coordinates q. It is therefore expected that more
17 harmonics are required to accurately approximate the nonlinear forces f nl , and also the auxiliary variables
18 v. Considering the quadratically recast system, it is therefore interesting to use a higher truncation order
19 Hv > Hq for the auxiliary variables than for the generalized coordinates. Analogously, considering instead
20 the initial system of equations of motion (1), it is interesting to apply the AFT scheme with a large number
21 of samples N  2H + 1.
22 In this work, we focus on two algorithms:
23
24 • AFT-PreCo: HB is applied to the initial equation of motion (1) using the AFT scheme (harmonic
25 truncation order H, number of time samples N ); continuation is done with tangent predictor (Eq. (12)),
26 pseudo arc-length constraint (Eq. (14)) and Newton-type correction iterations (Eq. (15))
27
28 • cHB-ANM: classical HB is applied to the quadratically recast differential-algebraic equation system
29 (harmonic truncation order H used for both generalized coordinates q and auxiliary variables v);
30 continuation is done with P -order Taylor series predictor (Eq. (16)), where the step length α is set to
31 αmax in accordance with Eq. (17)
32
33 It must be emphasized that a crucial difference between these two algorithms is whether HB is applied to
34 the ordinary differential equation of motion or a quadratic recast, and, thus, how the nonlinear terms are
35 represented. The problem dimension of cHB-ANM is larger due to the quadratic recast. To reach the same
36 accuracy, it is expected that cHB-ANM needs a higher harmonic truncation order H to properly represent
37 the auxiliary variables. These two aspects could lead to a larger computational effort for cHB-ANM as
38 compared with AFT-PreCo. On the other hand, the high-order expansion of the solution path within the
39 ANM might permit larger step lengths. Moreover, only a single Jacobian matrix has to be factorized for each
40 continuation step, as compared with typically 3 to 10 for the Newton iterations within the PreCo method.
41 These two aspects could (over-)compensate the aforementioned computational disadvantages. This will be
42 analyzed for several benchmark problems in the next section.
43 The algorithms were implemented in Matlab. The point of departure were the open source tools MAN-
44 LAB 4.0 (cHB-ANM) [29, 30] and NLvib 1.0 (AFT-PreCo) [3, 31]. Our intent is obviously to compare the
45 algorithms, not the tools. We therefore placed these implementations into a unified framework. The com-
46
putation of the initial solution point, pre- and post-processing tasks were excluded from the computational
47
effort comparisons. Where appropriate, the same linear algebra operations (e. g. matrix factorizations) were
48
49 used. Where appropriate, loops were vectorized. Nevertheless, a certain implementation bias cannot be
50 excluded, i. e., when a different person implements the methods, the computational efficiency might differ
51 to some degree. Thus, the computational effort comparisons must be regarded with care; only substantial
52 changes (e. g. an order of magnitude) should be considered relevant.
53
54 2.4. Summary of error sources
55 It is useful to recall at this point the causes of potential errors made by the described algorithms,
56 and by what parameters these are affected, as listed in Tab. 1. The PreCo algorithm just generates a
57 sequence of solution points, without information about the course of the solution branch between two
58
59 9
60
61
62
63
64
65
 1
2
3 adjacent points. Hence, when the step length is large, important features (e.g. super-harmonic regime,
4 primary resonance peak) along the solution branch might be resolved poorly or overlooked completely.
5
Then, approximation quality can be increased by recovering a continuous representation by interpolation
6
(here quadratic). However, the level of accuracy is then not just driven by HB approximation quality
7
8 (given by εtol ), but also by an additional interpolation error (related to α). In contrast, the ANM yields
9 a continuous representation of the solution branch; the accuracy of this representation is limited by the
10 truncation error (controlled by Taylor series order P ), and the assumed range of utility (controlled by the
11 ANM threshold εα ).
12
13 Table 1: Error sources in AFT-PreCo and cHB-ANM algorithms
14 error source control parameter
15
truncation harmonic order H
16
aliasing (for non-polynomial
17 number of samples per period N
18 nonlinearities within the AFT scheme)
19 regularization (to approximate
parameter εreg
20 non-smooth terms by analytic ones)
21 numerical tolerance threshold εtol , ANM: threshold εα
22
23 interpolation PreCo: step length α
24
25
26
27 3. Results for representative benchmark problems
28
29 In the course of the presented study, various benchmark problems were analyzed, with different types of
30 nonlinearities and different underlying linear base structures having different numbers of degrees of freedom.
31 In the following, we present only selected representative results of this comprehensive study. Three different
32 types of nonlinearity are presented: polynomial stiffness, unilateral stiffness and dry friction.
33
34 3.1. System with polynomial stiffness
35 First, we consider the equations of motion [32]
36
37 NX
DOF
π
38 qk00 + 0.1k 2 qk0 + k 4 qk + k 2 j 2 qk qj2 = fˆc (1) sin(k ) cos(ητ ) k = 1, . . . , NDOF . (21)
39 j=1
2
40
41 Herein, 0 denotes derivative with respect to normalized time τ . For NDOF = 1, Eq. (21) simplifies to the
42 equation of motion of the harmonically driven and damped Duffing oscillator. For NDOF > 1, Eq. (21) can
43 be interpreted as the equations of motion of a pinned-pinned beam: It is derived e. g. in [32], considers the
44 von-Karman beam theory and neglects the longitudinal inertia. The resulting momentum balance condensed
45 to the transverse direction is then projected onto the basis composed of the first NDOF lowest-frequency
46 odd-numbered (linear) normal modes. This results in the set of nonlinear ordinary differential equations
47 (21), where qk are the associated modal coordinates of the mass-normalized modes. In summary, the cubic
48 polynomial stiffness terms in Eq. (21) model the geometric stiffening caused by the midplane stretching
49
due to bending under axial constraints. The harmonic forcing with normalized angular frequency η is
50
applied as a concentrated load at the center of the beam with magnitude fˆc (1). Only the odd-numbered
51
modes receive external forcing, because the remaining modes have a node at the center of the symmetric
52
53 beam. Yet, both even- and odd-numbered modes are considered in the following analyses. For the quadratic
54 recast, one auxiliary variable vk is introduced per modal coordinate along with the additional equation
55 raux,k = vk − qk2 = 0.
56 As mentioned before, the sampling procedure within the AFT scheme yields the Fourier coefficients of
57 polynomial nonlinear terms without aliasing errors (exactly) for sufficiently many samples per period, N .
58
59 10
60
61
62
63
64
65
 1
2
3 In the considered examples, q(t) only responds with odd-numbered harmonics. As shown in Appendix A,
4 N ≥ 4H +1 is sufficient to completely avoid aliasing errors in the relevant H lowest-order Fourier coefficients
5
of the cubic nonlinear terms. For the results depicted in Fig. 2-6, the number of time samples was set to
6
N = 4H + 1 accordingly.
7
8
9 3.1.1. Effect of PreCo step length α and ANM threshold εα
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
Figure 2: Duffing oscillator, effect of nominal continuation step length of the AFT-PreCo method on amplitude-frequency
34 curves; (b), (c), (d) illustration of cyclic jumping between solution branches in super-harmonic regime for large step lengths
35 (do not represent zooms into (a))
36
37
38 In the following, we first focus on the case NDOF = 1 (Duffing oscillator), and the excitation level is set
39 ˆ
to fc (1) = 1.5. The amplitude ARMS is defined as the root-mean-square value
40
41 v
u H
1X 2
q
42 u
ARMS = q 2 = tq̂ 2 (0) + q̂c (k) + q̂s2 (k) . (22)
43 2
k=1
44
45 Since an exact solution is not available, the reference solution for this benchmark is generated by cHB-
46 ANM with a large number of harmonics (here H = 200 chosen). This seems justified since for polynomials
47 (C ∞ function), the truncated Fourier series is known to converge rapidly towards the reference solution
48
with increasing H. Besides that the continuous representation of the solution path allows an analytical
49
determination of the resonance peak.
50
51 The step length α within the PreCo method is generally a critical parameter and has to be specified by the
52 user, at least to a certain extent. For the considered PreCo method, only a nominal step length αnom is
53 specified by the user. The lower and upper bounds are set to αnom /5 and 5αnom , respectively. This was
54 empirically determined as reasonable tradeoff between adaptability in ranges with strong gradient changes
55 of the solution path and probability that important features of the solution path are not overlooked. If the
56 number of Newton iterations for consecutive continuation steps exceeded 9, the step length is halved, and
57 it is doubled if this number fell below 6.
58
59 11
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26 Figure 3: Duffing oscillator, effect of nominal step length αnom on AFT-PreCo: (a) number of continuation steps, (b) averaged
27 number of Newton iterations, (c) computation time; effect of ANM threshold εα : (d) number of continuation steps, (e) averaged
28 number of Newton iterations, (f) computation time
29
30
For very large nominal step lengths, the PreCo method might jump between solution branches. As a
31
32 consequence, important features of the bifurcation diagram might be overlooked, or the algorithm cycles
33 back and forth on the solution path as illustrated in Fig. 2b-d. In general, a very large step length may also
34 lead to divergence or unreasonably slow convergence of the Newton method. For moderately large nominal
35 step lengths, such problems do no longer occur. However, the solution path is only poorly represented in
36 ranges with rapid gradient changes, as indicated in the zooms within Fig. 2a.
37 In Fig. 3a-c, the effect of the nominal step length αnom (AFT-PreCo) on the number of solution points (equal
38 to the number of continuation steps), average number of Newton iterations and computation time is shown
39 (H = 20). For small step lengths, the number of continuation steps Npt and the time tcomp for computing
40 the whole branch are approximately proportional to 1/αnom . At the same time, the average number of
41 Newton iterations, N newt , increases with the nominal step length. The proper choice of the nominal step
42 length requires some level of experience and knowledge of the considered problem. The value αnom = 10−2
43 yields a reasonable tradeoff between computational effort, robustness and accuracy in this case. For the
44 analyzed benchmarks in this article the nominal step length was set to αnom = 10−2 , if not explicitly stated
45 otherwise.
46 The performance of the cHB-ANM algorithm depends on the ANM threshold εα , as illustrated in Fig. 3d-
47 f. As one can easily deduce from Eq. (17), the larger εα , the larger the step lengths, leading to fewer
48
continuation steps Npt . Throughout this work, the tolerance εtol (residual norm) is set to 10−10 . For this
49
benchmark, setting εα = εtol leads to occasional Newton iterations. This comes from the accumulation of
50
51 errors over the continuation steps. To avoid Newton iterations, we set εα = εtol /10 = 10−11 in the following
52 if not explicitly stated otherwise. It is interesting to see that the computational effort can be further reduced
53 when the ANM threshold is further increased. The ANM then degenerates to a PreCo algorithm with a
54 high-order predictor and conventional Newton corrector iterations.
55
56
57
58
59 12
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26 Figure 4: Duffing oscillator, effect of harmonic truncation order H: (a) overview of amplitude-frequency curve, (b) zoom into
27 the super-harmonic resonance regime; N = 4H + 1 for AFT-PreCo
28
29
30 3.1.2. Effect of harmonic truncation order H
31 In Fig. 4, the amplitude-frequency curves are depicted for different harmonic truncation orders H. To
32 better analyze the convergence behavior, error measures are defined. Note that the overall goal of each
33 algorithm is to determine a whole solution branch, not just a single point. It is therefore appropriate to
34 introduce global error measures that are representative for a complete branch. Because of its high tech-
35 nical importance, we consider only the point of maximum amplitude (resonance peak), max ARMS (η), and
36 determine the relative errors with respect to amplitude and frequency, εA and εω , respectively. Various alter-
37 native error measures were analyzed as well, including the amplitude ARMS for a fixed frequency, and it was
38 found that considering only the resonance peak yields representative results. To determine the maximum
39 amplitude point, the solution path was expressed as a continuous function. For cHB-ANM, the available
40 Taylor series expansion with respect to the arc length is used. For AFT-PreCo, a quadratic interpolation
41 polynomial is constructed between the three solution points with largest amplitude. The convergence of
42 these error measures with the harmonic truncation order is depicted in Fig. 5.
43
As expected, the results of both methods converge to the reference (cHB-ANM with H = 200) for sufficiently
44
large harmonic truncation order H. An exponential convergence rate is theoretically expected for systems
45
46 with analytic nonlinearities [33]. This would correspond to an error which has a slope in the log-log plots
47 (Fig. 5a-b) that approaches minus infinity in the asymptotic limit case for large H. Due to finite numerical
48 precision, however, a round-off plateau is reached in practice. For both methods, this appears for about
49 H = 101 . With cHB-ANM, the level εA ≈ εω ≈ 10−12 is reached for an ANM threshold εα = 10−11 . For a
50 much higher ANM threshold εα = 10−6  εtol , the ANM degenerates to a predictor-corrector method with
51 P -order predictor. This raises the error plateau to about 10−9 . To reach a similar order of the interpolation
52 error with AFT-PreCo, the nominal continuation step length αnom must be decreased to a relatively small
53 value. This reflects the advantage of the ANM of yielding a high-order representation of the whole solution
54 path.
55 For the same H, one might think that both methods yield the same truncation error. However, the results
56 generally differ. It is interesting to note that a reasonable accuracy can be achieved with the AFT HB
57 already for H = 1 near the primary resonance and for H = 3 near the 1 : 3 super-harmonic resonance
58
59 13
60
61
62
63
64
65
 1
2
3 (cf. Fig. 4). In contrast, the corresponding results for cHB considerably deviate from the reference. This can
4 be attributed to the fact that the nonlinear terms are represented in a different way when AFT HB is applied
5
immediately to the equations of motion, as compared to applying cHB to the quadratically recast system.
6
In AFT HB, the nonlinear terms are represented by the N time samples (recall that N = 4H + 1 here, where
7
8 H is the harmonic truncation order). In cHB, the nonlinear terms are represented by the 2H + 1 Fourier
9 coefficients (since the nonlinear terms are also approximated by a H-order Fourier series). Near the primary
10 resonance, the main variable q1 is dominated by the fundamental harmonic, and consequently q12 should be
11 dominated by the zeroth and the second harmonic. Therefore, the defining equation raux,1 = v1 − q12 = 0
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
Figure 5: Duffing oscillator, effect of harmonic order H: relative error of (a) amplitude and (b) frequency of the resonance
36 peak, (c) computation time; N = 4H + 1 for AFT-PreCo; cHB-ANM? shall indicate that in this case εα  εtol such that the
37 ANM degenerates to a predictor-corrector method with P -order predictor
38
39 for the auxiliary variable v1 is not represented well for H = 1. In general, q1 only responds with odd har-
40 monics in this benchmark problem, such that the auxiliary variable v1 should only respond with the zeroth
41 and higher even harmonics. Thus, the error for the quadratically recast system generally changes when H
42 is changed, while additional higher even harmonics are not relevant when only the equation of motion is
43
directly considered (as done with AFT-PreCo). For instance, the AFT-PreCo error remains the same when
44
increasing H from 1 to 2, but decreases for H = 3 (highlighted by pink diamonds in Fig. 5a-b).
45
46 The computational effort is depicted in Fig. 5c. To estimate the time complexity of the analyzed methods,
47 we provide asymptotes with different scaling behavior (e. g. O(H), O(H log H))) in corresponding plots
48 throughout this study. Here, the computation time increases significantly with H, for large H slightly
49 super-linearly for both methods. The number of Jacobian factorizations, Nfact , is also given for some points
50 in Fig. 5c. Recall that in the case of the ANM this equals the number of continuation steps, i. e., the number
51 of Taylor series expansion points, Nfact = Npt , and the number of continuation steps (equal to the num-
52 ber of generated solution points) times the average number of Newton iterations in the case of the PreCo
53 method, Nfact = Npt N newt . Here, the number of continuation steps for the PreCo method is about 300,
54 which is slightly larger than that of the ANM. The larger step lengths within the ANM are explained by the
55 higher order of the predictor. In accordance, N newt ≈ 2 − 3 here. Hence, the ANM requires fewer Jacobian
56 factorizations, here by a factor of about 4 to 9. But apparently this does not lead to substantially shorter
57 computation times. This is explained by the larger problem dimension: The quadratic recast introduces
58
59 14
60
61
62
63
64
65
 1
2
3 an auxiliary variable v1 (here NDOF = 1). Thus, we have Neq = 2(2H + 1) + 1 for cHB-ANM vs. only
4 Neq = (2H + 1) + 1 for AFT-PreCo.
5
As mentioned before, the computational effort should be viewed in the light of implementation bias. The
6
efficiency of each algorithm is affected by how well it is implemented (use of parallelization, vectorization of
7
8 loops within Matlab, etc.) for the given computing architecture. Although we unified the implementation
9 where appropriate, we are convinced that the implementation bias can easily account for an order of magni-
10 tude of computational effort. Here, the computation time is always in the same order of magnitude for both
11 methods (for αnom = 10−2 and εα = 10−11 ). However, we conclude that the cHB-ANM is computationally
12 more robust for this benchmark problem, since it is less sensitive to the threshold εα , while AFT-PreCo is
13 quite sensitive to the nominal step length αnom .
14
15 3.1.3. Effect of number of degrees of freedom NDOF
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40 Figure 6: System with polynomial stiffness: (a) amplitude-frequency curves, (b) computation time
41
42 Now the effect of the number of modal degrees of freedom, NDOF ≥ 1, is analyzed, and the excitation
43 level is set to fˆc (1) = 10. In Fig. 6a, the frequency response is illustrated for NDOF = 5 and H = 9.
44 Compared to the Duffing oscillator, additional primary and secondary resonances occur, as expected.
45 In Fig. 6b, the computation time is depicted vs. NDOF . Analogously to the above discussed case of
46 NDOF = 1, the quadratic recast now introduces one auxiliary variable per modal coordinate. Thus, we
47 have Neq = 2NDOF (2H + 1) + 1 for cHB-ANM vs. only Neq = NDOF (2H + 1) + 1 for AFT-PreCo. The
48 problem dimension of cHB-ANM is, hence, approximately twice that of AFT-PreCo. A change of slope can
49 be seen in Fig. 6b for AFT-PreCo at about NDOF = 10. This indicates that the bottleneck of the algorithms
50 changed. For large problem dimensions, the factorization of the Jacobian becomes the bottleneck. The
51
effort for this grows quadratically with NDOF for both methods. Consequently, AFT-PreCo is about one
52
order of magnitude faster than cHB-ANM for NDOF = 10, but this difference reduces to a factor of about 4
53
54 up to NDOF = 99 (1400 s for AFT-PreCo and about 6000 s for cHB-ANM).
55
56
57
58
59 15
60
61
62
63
64
65
 1
2
3 3.2. System with unilateral stiffness
4
5 We now consider the equation of motion
6
7 q 00 + 0.1q 0 + q + fu (q) − 0.2 cos (ηt) = 0 (23)
8
with the unilateral stiffness term fu (q) = 100 max (q − 1, 0). If q < 1, the spring is in separated state,
9
10 fu = 0. If q > 1, the spring is in contact state, and a force proportional to the spring deformation q − 1
11 is built up. fu is only continuous in q, its derivative has jumps, it is therefore non-smooth. As discussed
12 before, cHB cannot be applied since the the Fourier coefficients of the nonlinear force cannot be expressed
13 analytically 1 . Also, it is not reasonable to apply the ANM directly, since the high-order derivatives required
14 to expand the Taylor series simply do not exist. To apply cHB-ANM, fu must be regularized to an analytic
15 function. Here, we use
16 s 2
17 100(q − 1) + 100(q − 1)
18 fu (q) ≈ fu,εreg (q) = (−) + εreg . (24)
2 2
19
20 The smaller the regularization parameter εreg , the better the approximation of the unilateral stiffness term.
21 For the quadratic recast, one
22  auxiliary variable fu,εreg is introduced, along with one additional equation
raux = fu,εreg − 100 (q − 1) fu,εreg − εreg = 0.
23
24
The reference solution was developed as follows: Without regularization, the differential equation (23)
25
is piecewise linear. Hence, an analytical expression for the general solution in the separated and the contact
26
27 state can be easily derived. The main challenge is then to determine the contact state transition times
28 (contact↔separation). The condition of a contact state transition is q − 1 = 0. Substituting each gen-
29 eral solution into this equation yields an implicit equation for the next contact state transition time. This
30 equation is nonlinear and transcendental, and numerical rood finding is therefore used to determine each
31 transition time. This procedure is well-known and often called stitching. To find periodic solutions, it is
32 required that the coordinate and velocity at the end of an excitation period are the same as those at the
33 beginning; i. e., the problem is cast into a two-point periodic boundary value problem (shooting method).
34 Finally, the periodic solutions are continued using the PreCo method.
35
36 3.2.1. Effect of regularization parameter εreg
37 We studied the influence of the regularization parameter on the cHB-ANM solution, H = 20, by applying
38 different values down to εreg = 10−10 . In this benchmark, cHB-ANM was able to continue the whole branch
39 even for very steep regularizations, when setting the error tolerances properly. Moreover, the computational
40 effort showed no significant dependence on εreg . For the following analyses, we choose εreg = 10−9 , which
41
yields a highly accurate approximation of the non-smooth force law.
42
It is important to understand that the quadratic recast of fu given below Eq. (24) generates an ambiguity:
43
44 Besides the desired solution with the + sign in Eq. (24), which approximates the unilateral stiffness behavior,
45 we get an artificial solution with the (−) sign in Eq. (24), which gives fu,εreg (q) ≈ 0. During the continuation
46 procedure, this ambiguity comes to light by bifurcation points, cf. Fig. 7, where the unilateral spring would
47 be activated. The artificial solution yields the quasi-linear behavior indicated as dashed curve in Fig. 7.
48 The cHB-ANM implementation requires the user to manually select one of the emanating branches to be
49 further continued. When the quasi-linear branch is continued further, we encountered additional bifurcation
50 points not depicted in Fig. 7. To avoid these kind of branching problems in subsequent analyses, we apply a
51 predefined rule to automatically choose the correct (physical) solution path. The rule comprises that as soon
52 as cHB-ANM detects a bifurcation point, the emanating branch has to be continued in positive direction
53 with regard to excitation frequency η. Note that for switching reliably the branch at a bifurcation point we
54 had to reduce the ANM threshold to εreg = 10−13 .
55
56
57
58
59 16
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22 Figure 7: SDOF oscillator with unilateral stiffness: bifurcations and non-physical solutions caused by the selected regularization
and quadratic recast
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48 Figure 8: SDOF oscillator with unilateral stiffness, effect of samples per period N applied to the non-smooth system: error of
49 AFT HB using (a) H = 1 and (b) H = 7, (c) computation time for AFT-PreCo
50
51
52 3.2.2. Effect of number of samples per period N
53 In contrast to the system with polynomial stiffness, the sampling procedure within the AFT scheme is
54 not exact beyond a certain value N : The unilateral stiffness function fu (q) generates an infinite sequence
55
56 1 A piecewise analytical integration is possible. However, the transition times are the roots of high-order polynomial equa-
57 tions, which can only be approximated numerically, see e. g. [34].
58
59 17
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24 Figure 9: SDOF oscillator with unilateral stiffness: amplitude-frequency curves for (a) different harmonic truncation orders H
25 and (b) different numbers of time samples N ; reference and AFT-PreCo consider the non-regularized, cHB-ANM the regularized
force law
26
27
28 of (decaying but) nonzero Fourier coefficients fˆu (k) already for harmonic input q. Hence, aliasing errors
29 cannot be avoided within the AFT scheme. In Fig. 8a-b, the influence of N on the relative error of the
30 amplitude εA is depicted. For this we solved the AFT HB residual at a fixed excitation frequency close to
31 the resonance peak. This way, interpolation errors inevitable with the PreCo method are avoided. The error
32 decreases with N but does not reach any plateau at the numerical precision level in the depicted range.
33 N also influences the computational effort, as shown in Fig. 8c. For large N , the computational effort is
34
dominated by the calculation of the Fourier coefficients of the nonlinear force. The bottleneck of the algo-
35
rithm are then the forward and inverse fast Fourier transforms, the effort of which increases approximately
36
37 as N log(N ).
38 In Fig. 9b, results are shown for different N when using the AFT-PreCo algorithm with H = 100. Aliasing
39 errors may lead to divergence and numerical oscillations of the HB approximation. In this case, N ≥ 603 was
40 sufficient to ensure convergence. Even in the converged results, numerical oscillations occur, caused by the
41 aliasing-induced non-physical energy transfer from high to low frequencies. To better see these oscillations,
42 the nominal continuation step length was decreased to αnom = 10−4 . This effect decreases for higher sam-
43 pling rates. To the authors’ knowledge, there is no rigorous theory available to a priori determine a useful
44 value for N that ensures convergence and reduces the aliasing error below a specified threshold. Clearly,
45 the optimal choice of N depends on the mechanical system, the considered dynamic regime (affected e. g.
46 by the excitation level). For the following analyses, we considered two different settings for N : a fixed,
47 high number N = 213 and the empirical rule N = Nemp (H) = min(500 + 25 · H, 2000). A lower bound of
48 Nemp (0) = 500 ensures convergence and reasonable accuracy for low numbers of H, while the upper bound
49 Nemp (H ≥ 60) = 2000 avoids unnecessary computational effort for very high H.
50
51 3.2.3. Effect of harmonic truncation order H
52
In Fig. 9a, the amplitude-frequency curve is depicted for different harmonic truncation orders H. The
53
54 error measures, εA and εω (relative errors of resonance peak amplitude and frequency), introduced for the
55 system with polynomial stiffness are again considered to illustrate the convergence behavior in Fig. 10a-b.
56 Again, we observe that for the same (small) H, cHB-ANM yields larger errors than AFT-PreCo, since the
57 auxiliary variable is only represented by the 2H + 1 Fourier coefficients within cHB, whereas the nonlinear
58
59 18
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26 Figure 10: SDOF oscillator with unilateral stiffness, effect of harmonic truncation order H: relative error of (a) amplitude and
27 (b) frequency of the resonance peak, (c) computation time
28
29
30 forces are represented with more samples N > 2H + 1 within the AFT scheme. Using AFT-PreCo with
31 Nemp (H) leads to a more oscillatory decrease of the error than with N = 213 . This oscillatory behavior can
32 already be inferred from Fig. 9b. For N = 2000, the numerical oscillations have still a relative amplitude
33 of around 10−4 both in the amplitude and the frequency direction, respectively. This numerical oscillation
34 amplitude is reflected in the error level depicted in Fig. 10a-b.
35 Compared to the system with polynomial stiffness, the truncation error decreases much more slowly. For
36 about H = 101 , the errors decrease below 1% for this benchmark problem, while H = 1 or 2 was sufficient
37 for the system with cubic stiffness. While the error of the cHB-ANM still decreases (in average) beyond
38 H > 100, the AFT-PreCo reaches a plateau. As opposed to the case of the system with polynomial stiffness,
39 this plateau is driven by both interpolation and aliasing errors.
40 As can be seen in Fig. 10c, AFT-PreCo is faster than cHB-ANM when using the empirical rule Nemp (H),
41 but largely slower when setting N = 213 in the depicted range. In both cases, the number of Jacobian
42 factorizations Nfact are larger for AFT-PreCo compared to cHB-ANM (cf. Section 3.1.2). But here the
43 selected number of AFT time samples N decides whether the lower number of Jacobian factorizations Nfact
44
(but also larger problem size) within the cHB-ANM can be compensated or not. Overall, the differences are
45
again not considered substantial (less than one order of magnitude). One should remark the ANM threshold
46
47 was set to εα = 10−13 , as opposed to the other benchmarks, not only to avoid Newton iterations but also
48 to ensure that the physical solution branch is continued, cf. Section 3.2.1.
49
50 3.3. System with dry friction
51 We now consider a single degree-of-freedom oscillator with elastic dry friction element, described with
52 the equation of motion
53
54 q 00 + 0.02q 0 + q + fρ − fˆc (1) cos (ηt) = 0 , (25)
55
56
57
58
59 19
60
61
62
63
64
65
 1
2
3 where the friction force fρ is defined by the differential law
4
5
(
κdq |fρ + κdq| < ρ
6 dfρ = . (26)
7 0 otherwise
8
9 The elastic dry friction element can be interpreted as a linear spring with stiffness κ connected to a Coulomb
10 slider. If the magnitude of the force in the spring remains smaller than ρ, the Coulomb slider is sticking,
11 and the friction element acts as preloaded linear spring. Once the magnitude of the force reaches ρ, the
12 Coulomb slider slides in the direction of q 0 and the friction element produces a constant force ρ in the
13 opposite direction. The current value of fρ not only depends on q and q 0 , but also on their time history
14 (hysteresis), hence, fρ cannot be expressed as a unique function of q and q 0 . It is, however, still possible to
15 account for the differential law in Eq. (26) within the AFT scheme [35]: One starts with arbitrary initial
16 conditions, e. g., fρ = 0, and determines the evolution of fρ in a time-discrete way until the steady hysteresis
17 cycle is reached (two periods were sufficient here). Thus, only q has to be expanded in a truncated Fourier
18 series within AFT-PreCo.
19 As in the case of the unilateral stiffness, neither cHB nor the ANM can be directly applied to Eqs. (25)-(26),
20 since the Fourier coefficients of the nonlinear force cannot be expressed in closed form, and the system is
21 non-smooth, respectively. Therefore, a regularization is applied, which is setup in two steps: First, the
22 coordinate of the Coulomb slider,
23 p g, is introduced, and the signum-function describing the Coulomb law is
approximated as sgn g 0 ≈ g 0 / εreg + (g 0 )2 ,
24
25 q 00 + 0.02q 0 + q + κ (q − g) − fˆc (1) cos (ηt) = 0 , (27)
26 0
27 g
κ (g − q) + ρ p = 0. (28)
28 εreg + (g 0 )2
29
30 For the quadratic recast, the four auxiliary variables g, u, fρ and v4 are introduced, and Eq. (28) is replaced
31 by the four equations
32
33 κ(g − q) + fρ = 0, (29)
0
34 u−g = 0, (30)
35 v4 fρ − ρu = 0, (31)
36
37 v42 2
− u − εreg = 0. (32)
38
39 3.3.1. Effect of regularization parameter εreg
40 In Fig. 11, the amplitude-frequency curves are depicted for different values of the regularization parameter
41 εreg , and three different excitation levels fˆc (1). Here and in the following, we set κ = 3 and ρ = 1. As for
42 the system with unilateral stiffness, the shooting method is used as reference, with analytical integration of
43 the piecewise linear system, Eqs. (25)-(26), and direct calculation of the transition times between stick and
44 slip friction (not regularized ).
45 In contrast to the unilateral stiffness benchmark, we observed that as the regularization becomes steeper,
46 i. e. with decreasing regularization parameter εreg , the number of harmonics in the cHB-ANM has to be
47 increased to continue the whole branch. This is analyzed in more detail in Section 3.3.3. In Fig. 11, H = 20,
48
and cHB-ANM continues the whole branch only if εreg (fˆc (1) = 0.5) ≥ 3 · 10−4 . (Using H = 100 steeper
49
50 regularizations down to εreg (fˆc (1) = 0.5) = 5 · 10−5 are possible.) For the higher excitation level, we even
51 need εreg (fˆc (1) = 1.1) ≥ 5.5 · 10−3 . Compared to the unilateral stiffness, the remaining regularization error
52 is not negligible, and particularly severe for higher excitation levels (more sliding friction). For εreg = 10−5
53 (finest considered εreg ), the regularization error is substantial for the highest excitation level only. We also
54 applied AFT-PreCo to the regularized (but not quadratically recast) system, and did not encounter any
55 convergence problems down to εreg = 10−5 for N ≥ 45 (H = 20). We thus conclude that the convergence
56 problems are associated with the representation of the nonlinear terms induced by applying cHB to the
57 quadratically recast differential-algebraic equation system.
58
59 20
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24 Figure 11: SDOF oscillator with elastic dry friction element, effect of regularization εreg on amplitude frequency curve for
25 different forcing levels; reference considers the non-regularized force law
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
Figure 12: SDOF oscillator with dry friction, effect of samples per period N applied to the non-smooth system: error of AFT
50
HB using (a) H = 1 and (b) H = 7, (c) computation time for AFT-PreCo
51
52
53 3.3.2. Effect of number of samples per period N
54 If the non-smooth friction law in Eq. (26) is used, the AFT scheme inevitably leads to aliasing errors, as
55
in the case of the unilateral stiffness. In Fig. 12a-b we analyzed the influence of N on the relative error of
56
the amplitude εA for a fixed frequency η = 1.7. Here, aliasing effects can be minimized using much fewer
57
58
59 21
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26 Figure 13: SDOF oscillator with elastic dry friction element, effect of harmonic truncation order H on amplitude-frequency
27 curves: (a) some successful and (b) some failed computations
28
29
30 time samples N compared to the unilateral stiffness benchmark. While N ≈ 60 (H = 1) is sufficient to
31 decrease the error to kεA − εA (N = 2 · 104 )k = 10−3 , about N ≈ 400 (H = 1) were required to reduce the
32 error to about the same level in the unilateral stiffness benchmark.
33 Analogous to the benchmark with the unilateral stiffness, we use two settings of N for the following analyses:
34 a fixed, high number N = 211 and the empirical rule N = Nemp (H) = min(50 + 10 · H, 500). Compared to
35 the benchmark with the unilateral stiffness, much smaller sampling rates are thus used for the benchmark
36 with dry friction.
37
38 3.3.3. Effect of harmonic truncation order H
39 For the results depicted in the following, the intermediate excitation level fˆc (1) = 0.5 is considered, and
40 the regularization parameter is set to εreg = 3 · 10−4 (smallest feasible with cHB-ANM for H = 20). In
41 Fig. 13, the amplitude-frequency curve is depicted for different harmonic truncation orders H. The conver-
42 gence behavior is illustrated in Fig. 14a-b, in terms of the error measures introduced earlier.
43
Remarkably, H = 1 with Nemp = 60 is sufficient to achieve reasonable accuracy (errors < 1%) with AFT-
44
PreCo applied to the non-smooth system. A plateau is reached for about H ≥ 13, at an error level
45
46 εA ≈ 5 · 10−5 . Because of aliasing, Nemp (H) leads to a slightly larger error εA than for N = 211 in a wide
47 range, and to more pronounced numerical oscillations. Note that as opposed to εA , εω stays almost constant.
48 The reason for this is that, by coincidence, the PreCo continuation generates a point very close to the actual
49 resonance peak (for the particular choice of the continuation limits and the specified nominal step length).
50 Errors < 1% are never reached with cHB-ANM. The accuracy is mainly limited by the regularization error.
51 The cHB-ANM error stabilizes beyond about H = 26.
52 It is important to note that the harmonic truncation order has to be set to H ≥ 62 to ensure that the whole
53 solution branch is computed with cHB-ANM. Fig. 13b shows some cases where the computation of the
54 (physical) solution branch is unsuccessful. Branches of ghost solutions appear, which satisfy the condition
55 kRk < εtol , but are far from any reference solution. In other cases, the continuation gets stuck or turns
56 around. Interestingly and in contrast to the benchmark with unilateral stiffness, the ANM did not detect any
57 branching points. In general, the existence of a HB approximation is only guaranteed for sufficiently large
58
59 22
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26 Figure 14: SDOF oscillator with elastic dry friction element, effect of harmonic truncation order H: relative error of (a)
27 amplitude and (b) frequency of the resonance peak, (c) computation time
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52 Figure 15: SDOF oscillator with elastic dry friction element, generalized and auxiliary variables computed with H = 500 using
regularized force law (εreg = 3 · 10−4 ) close to resonance peak: (a) one period of vibration and (b) decay of corresponding
53 Fourier coefficients for k ≥ 1
54
55
56 H [4]. This explains why the cHB can fail to converge in some cases. Apparently, the harmonic truncation
57 order must be set to a comparatively large number, H ≥ 62, to ensure convergence. This number is cer-
58
59 23
60
61
62
63
64
65
 1
2
3 tainly driven by the required accuracy for representing certain auxiliary variables of the quadratically recast
4 problem. To provide further insight into this, the time evolution and the Fourier spectrum of the auxiliary
5
variables are depicted in Fig. 15. In the non-regularized case, the velocity of the Coulomb slider jumps at a
6
stick-to-slip transition. This results in the almost discontinuous behavior of u(τ ) (and consequently v4 (τ ))
7
8 in the regularized case, associated with a Gibbs-type phenomenon and slow decay of the corresponding
9 Fourier coefficients. In contrast, the friction force and the deformation of the spring κ are continuous (C 0 )
10 functions, even in the non-regularized case, yielding slightly better convergence of the Fourier series. Even
11 better is the Fourier representation of the coordinate q(τ ), which is a C 2 function.
12 We also observed that the ANM algorithm applied Newton correction iterations to reach some solution
13 points, even in those runs where the whole branch was successfully continued. This means that the high-
14 order Taylor series predictor produced unacceptably large residuals (although the ANM threshold εα is
15 already set to one tenth of the tolerance εtol ).
16 The non-negligible regularization error and the poor numerical robustness are already strong arguments
17 against applying cHB-ANM to the considered dry frictional system. Yet another argument is the compu-
18 tational effort (see Fig. 14c): AFT-PreCo takes less than 0.5 s to compute the frequency response of the
19 non-regularized problem with H = 1, Nemp = 60. This is about 25 times faster than cHB-ANM applied
20 to the regularized problem with H = 26. As opposed to the previous benchmarks, not only the increased
21 problem size (quadratic recast and higher number of harmonics) of cHB-ANM prevents better computational
22 performance compared to AFT-PreCo but also the number of Jacobian factorizations Nfact : 491 needed for
23 cHB-ANM vs. only 164 for AFT-PreCo when using H = 100. We observed, when applying the same regu-
24
larization according to Eq. (28) to AFT-PreCo the required number of Jacobian factorizations also increases
25
significantly with H (e. g. about Nfact (H = 100) = 714).
26
27
28 4. Concluding guidelines
29
30 In this work, we compared the performance of the cHB-ANM and the AFT-PreCo algorithm for harmon-
31 ically driven systems with either polynomial stiffness, unilateral spring or elastic dry friction nonlinearity.
32 On the basis of the results, we derive the following conclusions.
33 If the considered differential-algebraic equation system contains non-smooth terms, arising e. g. from contact
34 constraints, the AFT-PreCo algorithm is highly recommended. To apply cHB or the ANM, a regularization
35 would be necessary to approximate the non-smooth terms by analytic ones. Additionally, a very steep reg-
36 ularization can be required to limit the error made by this approximation, as in the case of the benchmark
37 problem with dry friction. A prohibitively high harmonic truncation order might then be needed to ensure
38
convergence of the cHB-ANM algorithm. In contrast, reasonable accuracy can often be achieved with the
39
AFT-PreCo already with a very low truncation order, e. g. H = 1 or H = 3. This is because the AFT-PreCo
40
algorithm treats only the main variables (usually the generalized coordinates) as unknowns, which can often
41
42 be represented with only a few Fourier terms, while the nonlinear terms are treated internally only, and
43 represented by (potentially a much larger number of) time samples.
44 There are two important drawbacks of the AFT-PreCo algorithm: first, the AFT scheme yields inevitable
45 aliasing errors for non-polynomial nonlinear terms. If the number of samples per period is too small, nu-
46 merical oscillations or even divergence are possible. To the authors’ knowledge there is no rigorous theory
47 available to select this number in accordance with a priori estimated aliasing errors, and empirical rules are
48 commonly used. Second, the PreCo continuation step length has to be properly selected by the user. Large
49 step lengths may lead to a poor resolution (leading to inevitable interpolation errors when behavior between
50 discrete solution points is of interest) or even jumps over important features of the solution path. Small
51 step lengths increase the computational effort without providing significant additional information.
52 If the considered differential equation system contains only analytic terms (which are not designed to approx-
53 imate non-smooth terms), the cHB-ANM algorithm is recommended. To compute the required coefficients
54 of the high-order Taylor series expansion efficiently, a quadratic recast of the differential equation system is
55 commonly applied (which has to be done by the user). For this recast, auxiliary variables and additional
56 equations need to be introduced, which increases the problem dimension. In some cases, this recast can
57
58
59 24
60
61
62
63
64
65
 1
2
3 introduce mathematical ambiguities (it is not always bijective) leading to artificial bifurcations, which re-
4 quires some user interaction. Recently developed algorithms can then be used to automatically set up the
5
harmonic balance equations and carry out the continuation. The important advantage of this procedure is
6
that the continuation step length is automatically chosen depending on the range of utility of the truncated
7
8 Taylor series. This makes the procedure numerically robust (the step length does not have to be selected by
9 the user) and computationally competitive (since correction iterations are avoided). The cHB-ANM algo-
10 rithm appears useful for smooth nonlinearities, for instance geometrically nonlinear finite element models.
11 For large numbers of degrees of freedom (NDOF ), the Jacobian factorization becomes the bottleneck of both
12 algorithms, the computational effort then increases approximately quadratically with NDOF .
13 In the future, it would be interesting to extend the cHB-ANM algorithm in such a way that different
14 variables are resolved with individual harmonic truncation orders. This would permit to combine a more
15 precise representation of the nonlinear terms (idea behind AFT HB) with the sophisticated continuation
16 framework of the ANM. Furthermore, the aforementioned conclusions on the performance of the cHB-ANM
17 (high harmonic truncation order needed, mathematical ambiguities) for initially non-smooth problems were
18 drawn for the considered regularization; it remains to be investigated, how different regularizations affect
19 the cHB-ANM performance. Moreover, it would be interesting to extend the comparison of cHB-ANM and
20 AFT-PreCo algorithms to bifurcation analyses.
21
22
23 Acknowledgement
24 This work presents results of the Tribomechadynamics Research Camp (TRC), formerly called Nonlinear
25 Dynamics of Coupled Structures and Interfaces summer school, an annual month long international research
26
collaboration for graduate students and postdoctoral researchers. For more information, visit http://tmd.
27
rice.edu.
28
29 The author Fabia Tubita is grateful for the funding from the European Union’s Horizon 2020 research and
30 innovation programme under the Marie Sklodowska-Curie grant agreement No. 721865.
31 The authors Bruno Cochelin, Louis Guillot and Christophe Vergez would like to thank the Labex MEC
32 (ANR-10-LABX-0092) and the A*MIDEX project (ANR-11-IDEX-0001-02), funded by the Investissements
33 d’Avenir French Government program managed by the French National Research Agency (ANR).
34
35 Appendix A. On the exactness of the AFT scheme for polynomials
36
37 The goal of the AFT scheme is to compute the Fourier coefficients of the nonlinear forces up to the
38 considered truncation order H. In this appendix, we show that the AFT gives the exact Fourier coefficients
39 in the special case of polynomial nonlinear forces for a sufficiently large number of samples N . To this end,
40 we consider a scalar function
41
42 f (q) = q P , (A.1)
43 i. e., a monomial in q of order P . The following argument can be extended to arbitrary polynomials in a
44 straight-forward way. In accordance with HB, we use as ansatz for q a Fourier series truncated to order H,
45
H
46 X
47 q= q̂(k)eikτ , (A.2)
48 k=−H
49 with the normalized time τ = Ωt. In this appendix, we use the complex-exponential representation of a
50 Fourier series, which is, of course, fully equivalent to the sine-cosine representation e. g. in Eq. (2). To
51 ensure that the sum in Eq. (A.2) yields a real result, we must have q̂(−k) = q̂(k).
52 One can easily see that substituting the H-order Fourier series in Eq. (A.2) into the P -order monomial in
53
Eq. (A.1) generates a P H-order Fourier series,
54
55 PH
X
56 f= fˆ(k)eikτ . (A.3)
57 k=−P H
58
59 25
60
61
62
63
64
65
 1
2
3 For instance, a cubic term (P = 3) generates at most Fourier terms with frequency 3Ω.
4
The Fourier coefficients fˆ(k) in Eq. (A.3) are defined by the integral
5
6 Z2π
7 1
fˆ(k) = f (q(τ ))e−ikτ dτ . (A.4)
8 2π
9 0
10
11 The discrete Fourier transform (DFT) generally gives an approximation of the continuous integral in
12 Eq. (A.4),
13 N −1
14 1 X ˜ 2π
fˆ(k) ≈ f (n)e−ik N n , (A.5)
15 N n=0
16
17 where f˜(n) are samples, f˜(n) = f (q(2πn/N )), and N is the number of samples. Eq. (A.5) can be identified
18 as applying the trapezoidal rule to approximate the continuous integral in Eq. (A.4) [3].
19 We now show that if f is a truncated Fourier series, Eq. (A.5) gives the exact value of the continuous integral
20
in Eq. (A.4), and therefore the exact Fourier coefficients, for sufficiently large N .
21
22 N −1 N −1 PH
!
1 X ˜ −ik 2π 1 X X 2π 2π
23 f (n)e N n
= fˆ(m)e N
im n
e−ik N n
24 N n=0 N n=0
m=−P H
25 PH N −1
26 X X 1 i2π m−k n
= fˆ(m) e N
27 N
m=−P H n=0
28
PH N −1
29 X X 1  i2π m−k n
30 = fˆ(m) e N

n=0
N
m=−P H
31 (
32 PH
X 1 |m − k| mod N = 0
33 = fˆ(m) . (A.6)
m=−P H
0 otherwise
34
35
36 Herein, a b denotes the modulo operation which determines the remainder of the division of a by b. Thus,
37 the expression |m − k| mod N = 0 means that |m − k| is an integer multiple of N . To verify the last step of
38 Eq. (A.6), we use the geometric regression lemma:
39 N −1
40 X
sN = αz n ,
41
n=0
42
43 ⇒ sN − zsN = α − αz N ,
44 1 − zN
45 ⇒ sN = α . (A.7)
1−z
46
2π
47 With z = ei N ` and α = 1/N we obtain
48
49 −1
N
(
X 1  i 2π ` n 1 1 − ei2π` 1 |`| mod N = 0
50 e N = = . (A.8)
51 n=0
N N 1 − ei 2π
N ` 0 otherwise
52
53 The expression in the middle is actually indeterminate for |`| mod N = 0, but its limit for |`| mod N → 0
54 is well-defined and given on the right. As alternative to taking the limit, the sum can be easily calculated
2π
55 noticing that ei N ` = 1 if |`| is an integer multiple of N .
56 Now, the last line of Eq. (A.6) can be verified by setting ` = m − k. To avoid aliasing, we need to ensure
57 that N is sufficiently large so that |m − k| never reaches N (or any higher integer multiple), and thus
58
59 26
60
61
62
63
64
65
 1
2
3 |m − k| mod N = 0 is only achieved for m = k. To set up the HB equations, we need only the Fourier
4 coefficients of f up to order H. Thus, the largest possible value of |m − k| is obtained when m = +P H
5
and k = −H, or when m = −P H and k = +H. In both cases, |m − k| = (P + 1)H. As a consequence,
6
7 for N ≥ (P + 1)H the sum on the right hand side of Eq. (A.5) yields exactly fˆ(k) for k = −H, . . . , H (no
8 aliasing errors). It is important to emphasize that N = (P + 1)H is not sufficient to correctly compute the
9 higher-order Fourier terms of f , i. e., fˆ(k) with H < |k| ≤ P H. But these are not needed for setting up the
10 HB equations.
11 In summary, we can say that the AFT scheme for computing the Fourier coefficients of the nonlinear forces
12 is exact if N ≥ 3H + 1 for polynomial nonlinear terms of up to quadratic order. Analogously, it is exact if
13 N ≥ 4H + 1 for polynomial nonlinear terms of up to cubic order.
14
15
16 References
17 [1] E. Sarrouy, J.-J. Sinou, Non-linear periodic and quasi-periodic vibrations in mechanical systems-on the use of the harmonic
18 balance methods, in: Advances in Vibration Analysis Research, Intech, 2011.
19 [2] T. Detroux, L. Renson, G. Kerschen, The harmonic balance method for advanced analysis and design of nonlinear mechan-
ical systems, Nonlinear Dynamics, Volume 2: Proceedings of the 32nd IMAC, A Conference and Exposition on Structural
20 Dynamics (2014).
21 [3] M. Krack, J. Gross, Harmonic Balance for Nonlinear Vibration Problems, Springer, 2019. doi:10.1007/
22 978-3-030-14023-6.
23 [4] M. Urabe, Galerkin’s procedure for nonlinear periodic systems, Archive for Rational Mechanics and Analysis 20 (2) (1965)
24 120–152.
[5] E. P. Petrov, Stability analysis of multiharmonic nonlinear vibrations for large models of gas turbine engine structures
25 with friction and gaps, Journal of Engineering for Gas Turbines and Power 139 (2) (2016) 022508–022508–10. doi:
26 10.1115/1.4034353.
27 [6] L. Salles, L. Blanc, F. Thouverez, A. M. Gouskov, P. Jean, Dual time stepping algorithms with the high order harmonic
28 balance method for contact interfaces with fretting-wear, Paper GT2011-46488, Proc. of GT2011, ASME Turbo Expo
29 2011: Advancing Clean and Efficient Turbine Technology, June 7-10, Vancouver, Canada (2011).
[7] S. Zucca, C. M. Firrone, Nonlinear dynamics of mechanical systems with friction contacts: Coupled static and dynamic
30
multi-harmonic balance method and multiple solutions, Journal of Sound and Vibration 333 (3) (2014) 916–926. doi:
31 10.1016/j.jsv.2013.09.032.
32 [8] D. Süß, K. Willner, Investigation of a jointed friction oscillator using the multiharmonic balance method, Mechanical
33 Systems and Signal Processing 52–53 (2015) 73–87. doi:10.1016/j.ymssp.2014.08.003.
34 [9] D. Laxalde, F. Thouverez, Complex non-linear modal analysis for mechanical systems: application to turbomachinery
bladings with friction interfaces, Journal of Sound and Vibration 322 (4-5) (2009) 1009–1025. doi:10.1016/j.jsv.2008.
35
11.044.
36 [10] G. Jacquet-Richardet, M. Torkhani, P. Cartraud, F. Thouverez, T. Nouri Baranger, M. Herran, C. Gibert, S. Baguet,
37 P. Almeida, L. Peletan, Rotor to stator contacts in turbomachines. review and application, Mechanical Systems and Signal
38 Processing 40 (2) (2013) 401–420. doi:10.1016/j.ymssp.2013.05.010.
39 [11] N. Coudeyras, S. Nacivet, J.-J. Sinou, Periodic and quasi-periodic solutions for multi-instabilities involved in brake squeal,
Journal of Sound and Vibration 328 (4–5) (2009) 520–540. doi:10.1016/j.jsv.2009.08.017.
40
[12] K. C. Hall, J. P. Thomas, W. S. Clark, Computation of unsteady nonlinear flows in cascades using a harmonic balance
41 technique, AIAA Journal 40 (5) (2002) 879–886.
42 [13] R. J. Gilmore, M. B. Steer, Nonlinear circuit analysis using the method of harmonic balance—a review of the art. part
43 i. introductory concepts, International Journal of Microwave and Millimeter–Wave Computer–Aided Engineering 1 (1)
44 (1991) 22–37.
[14] T. M. Cameron, J. H. Griffin, An alternating frequency/time domain method for calculating the steady-state response of
45
nonlinear dynamic systems, Journal of Applied Mechanics 56 (1) (1989) 149–154. doi:10.1115/1.3176036.
46 [15] B. Cochelin, C. Vergez, A high order purely frequency-based harmonic balance formulation for continuation of periodic
47 solutions, Journal of Sound and Vibration 324 (1–2) (2009) 243–262. doi:10.1016/j.jsv.2009.01.054.
48 [16] S. Karkar, B. Cochelin, C. Vergez, A high-order, purely frequency based harmonic balance formulation for continuation
49 of periodic solutions: The case of non-polynomial nonlinearities, Journal of Sound and Vibration 332 (4) (2013) 968–977.
50 doi:10.1016/j.jsv.2012.09.033.
[17] L. Guillot, B. Cochelin, C. Vergez, A generic and efficient taylor series–based continuation method using a quadratic recast
51 of smooth nonlinear systems, International Journal for Numerical Methods in Engineeringdoi:10.1002/nme.6049.
52 [18] L. Guillot, B. Cochelin, C. Vergez, A taylor series-based continuation method for solutions of dynamical systems, Nonlinear
53 Dynamicsdoi:10.1007/s11071-019-04989-5.
54 [19] M. Urabe, A. Reiter, Numerical computation of nonlinear forced oscillations by galerkin’s procedure, Journal of Mathe-
55 matical Analysis and Applications 14 (1) (1966) 107–140.
[20] E. Moussi, S. Bellizzi, B. Cochelin, I. Nistor, Nonlinear normal modes of a two degrees-of-freedom piecewise linear system,
56 Mechanical Systems and Signal Processing (2013) 1–26.
57 [21] R. W. Klopfenstein, Zeros of nonlinear functions, Journal of the ACM (JACM) 8 (3) (1961) 366–373.
58
59 27
60
61
62
63
64
65
 1
2
3 [22] E. Riks, The application of newton’s method to the problem of elastic stability, Journal of Applied Mechanics 39 (4)
4 (1972) 1060–1065. doi:10.1115/1.3422829.
5 [23] R. Seydel, Practical Bifurcation and Stability Analysis: from Equilibrium to Chaos, Springer New York, 1994.
6 [24] A. Cardona, T. Coune, A. Lerusse, M. Geradin, A multiharmonic method for non-linear vibration analysis, International
7 Journal for Numerical Methods in Engineering 37 (9) (1994) 1593–1608.
[25] M. Jerschl, D. S, K. Willner, Numerical Continuation Methods for the Concept of Non-linear Normal Modes, 2014, pp.
8 19–26. doi:10.1007/978-3-319-04546-7_3.
9 [26] B. Cochelin, N. Damil, M. Potier-Ferry, The asymptotic-numerical method: an efficient perturbation technique for non-
10 linear structural mechanics, Revue Européenne des Éléments Finis 3 (2) (1994) 281–297. doi:10.1080/12506559.1994.
11 10511124.
12 [27] B. Cochelin, M. Medale, Power series analysis as a major breakthrough to improve the efficiency of asymptotic numerical
method in the vicinity of bifurcations, Journal of Computational Physics 236 (1) (2013) 594–607. doi:10.1016/j.jcp.
13
2012.11.016.
14 [28] I. Charpentier, B. Cochelin, Towards a full higher order ad-based continuation and bifurcation framework, Optimization
15 Methods and Software 33 (4-6) (2018) 945–962. doi:10.1080/10556788.2018.1428604.
16 [29] Manlab - an interactive path-following and bifurcation alnalysis software, available at https:\\manlab.lma.cnrs-mrs.fr.
17 [30] R. Arquier, B. Cochelin, C. Vergez, Manlab logiciel de continuation interactif (2005).
18 [31] Nlvib - a matlab tool for nonlinear vibration problems, available via https:\\www.ila.uni-stuttgart.de\nlvib.
[32] A. H. Nayfeh, D. T. Mook, Nonlinear Oscillations, Vol. 1979, John Wiley & Sons, New York.
19 [33] E. Tadmor, The exponential accuracy of fourier and chebyshev differencing methods, SIAM J. Numer. Anal. 23 (1) (1986)
20 1–10. doi:10.1137/0723001.
21 [34] M. Krack, L. Panning-von Scheidt, J. Wallaschek, A high-order harmonic balance method for systems with distinct states,
22 Journal of Sound and Vibration 332 (21) (2013) 5476–5488. doi:10.1016/j.jsv.2013.04.048.
23 [35] J. Guillen, C. Pierre, An efficient hybrid frequency-time domain method for the dynamics of large-scale dry-friction damped
structural systems, in: C. Glocker, F. Pfeiffer (Eds.), IUTAM Symposium on Unilateral Multibody Contacts, Vol. 72 of
24 Solid mechanics and its applications, Kluwer Academic Publishers, Dordrecht, 1999, pp. 169–178.
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59 28
60
61
62
63
64
65