Mechatronics 42 (2017) 11–24

Contents lists available at ScienceDirect

Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics

A comparative overview of frequency domain methods for nonlinear

systemsR,RR
David Rijlaarsdam a,∗, Pieter Nuij a, Johan Schoukens b, Maarten Steinbuch a
a
Eindhoven University of Technology, Department of Mechanical Engineering, Control Systems Technology, PO Box 513, WH -1.129, 5600 MB, Eindhoven, The
Netherlands
b
Vrije Universiteit Brussel, Department of Fundamental Electricity and Instrumentation, K430, Pleinlaan 2, 1050 Brussels, Belgium

a r t i c l e i n f o a b s t r a c t

Article history: The widespread acceptance of frequency domain techniques for linear and time invariant systems has
Received 23 November 2015 been an impetus for the extension of these methodologies toward nonlinear systems. However, differ-
Revised 7 November 2016
ences and equivalences between alternative methods have been less addressed. This paper provides a
Accepted 29 December 2016
comparative overview of four classes of frequency domain methods for nonlinear systems: Volterra based
Available online 10 January 2017
models, nonlinear frequency response functions / Bode plots, describing functions and linear approxima-
Keywords: tions in the presence of nonlinearities. Each method is introduced using consistent nomenclature and
Frequency domain methods terminology, which allows for comparison in terms of system and signal classes for which the methods
Nonlinear systems are valid as well as the type of (nonlinear) effects captured by each model. Summarizing, the paper aims
Modeling to connect, and make different frequency domain methods for nonlinear systems accessible, by providing
Generalized FRF a comparative overview of such methodologies, accompanied by illustrative (experimental) examples.
Describing functions
Linear approximation © 2017 Elsevier Ltd. All rights reserved.
Nonlinear FRF
Volterra systems
Convergent systems
Fading memory

1. Introduction tain degree of nonlinear behavior could be ignored because it did

not impair system performance. In many present day dynamic sys-
For linear and time invariant (LTI) systems, frequency domain tems however nonlinear behavior limits performance or is deliber-
techniques resulted in a widespread acceptance in the engineering ately introduced to improve performance. See for example [45,89].
community for analysis, modeling and controller design [9]. Corre- In both cases nonlinear effects can no longer be neglected. Hence,
spondingly, the Frequency Response Function (FRF) and represen- the widespread acceptance of frequency domain techniques for LTI
tations such as the Bode, Nyquist and Nichols plot have become systems has been an impetus for the extension of these method-
standard engineering tools and have proven to be indispensable ologies toward nonlinear systems.
for the modeling and design of dynamic systems in industry. In- Although different approaches have been independently devel-
creased performance requirements of these systems result in a dif- oped to analyze and model nonlinear systems in the frequency
ferent attitude toward nonlinear system behavior. Previously a cer- domain, differences and equivalences between alternative methods
have been minorly addressed. In this paper an overview and com-
parison of the following four well established approaches is pre-
R
This work is carried out as part of the Condor project, a project under the super- sented:
vision of the Embedded Systems Institute (ESI) and with FEI company as the indus-
trial partner. This project is partially supported by the Dutch Ministry of Economic • Volterra series based approaches (Section 3)
Affairs under the BSIK program. This work was supported in part by the Fund for • Frequency response function and Bode plot for nonlinear sys-
Scientific Research (FWO-Vlaanderen), by the Flemish Government (Methusalem), tems (Section 4)
and by the Belgian Government through the Interuniversity Poles of Attraction (IAP • Describing functions (Section 5)
VI/4) Program.
RR
This paper was recommended for publication by Associate Editor Hamid Reza
• Linear approximations in the presence of nonlinearities
Karimi. (Section 6)
∗
Corresponding author.
E-mail addresses: [email protected], [email protected] (D. Ri- Table 1 provides a comparative overview of the main charac-
jlaarsdam). teristics of these modeling approaches and constitutes one of the

http://dx.doi.org/10.1016/j.mechatronics.2016.12.008
0957-4158/© 2017 Elsevier Ltd. All rights reserved.
12 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24

Table 1
Overview of frequency domain modeling approaches for nonlinear systems. Legend: dark grey: effect captured / information available,
light grey: effect partially captured / requires additional processing, white: effect not captured / information not available. For a description
of the different frequency domain effects, see [98].

main results of this paper. This comparison is further motivated in harmonics of the input frequency. Moreover, when a nonlinear sys-
the sequel. tem is subject to a multisine input, input frequencies may combine,
First, a brief introduction to the effects of nonlinearities in the producing new frequencies which are not present in the input sig-
frequency domain is presented in Section 2. Next, Sections 3 – nal and spectral components at a given frequency in the output
6 introduce each model class. This includes a definition of the sys- may depend on other frequencies in the input. These effects are
tem and signal class for which the model types are defined. The summarized in Table 1 and they are used in the following to dis-
notation and terminology used in the different sections are consis- cuss different frequency domain models for nonlinear systems. To
tent, which allows for comparison between different approaches. further illustrate the effects listed in Table 1, consider the follow-
A brief overview of relevant literature is provided for each ing example.
method (see Table 4, Appendix B, for a summary). Moreover, each
Example 1 (effects of nonlinearities in the frequency do-
section is concluded by either an analytical example or an experi-
main). Consider the following harmonic signal with nonzero spec-
mental case study, illustrating the introduced approach. Finally, in
tral contributions at 1 [Hz] and 5 [Hz].
Section 7 a comparison of the properties of the different model
types is provided as well as an assessment of their applicability in u(t ) = a cos(1 · 2π t ) + b cos(5 · 2π t )
practice. Finally, note that the required nomenclature and prelim- Now consider a static nonlinear mapping y(t ) = 4u3 (t ), which
inaries are presented in Appendix A to increase readablity of the yields:
paper.    
3a3 + 6ab2 cos(1 · 2π t ) + a3 + 3a2 b cos(3 · 2π t )
y(t ) =      
2. Nonlinearities in the frequency domain gain comp./exp. & desensitization harmonics & intermodulation
 
+ 6a2 b + 3b3 cos(5 · 2π t ) + 3a2 b cos(7 · 2π t )
A fundamental property of LTI systems is that it cannot shift      
energy from one frequency to the other. Hence, the response to gain comp./exp. & desensitization intermodulation
a sinusoidal input with a particular frequency is again sinusoidal
+ 3ab2 cos(9 · 2π t ) + 3ab2 cos(11 · 2π t )
with the same frequency as the input signal. The phase shift and      
gain relating the input and output are characterized by the FRF at intermodulation intermodulation
that particular frequency and as superposition holds for LTI sys- + b3 cos(15 · 2π t )
tems, the response to more general input signals is fully captured
   (1)
by the FRF as well. However, for nonlinear systems the superpo- harmonics
sition principle does not hold and the response to even a sim- The expression in (1) clearly shows the different effects indi-
ple sinusoid can be a complex, multi harmonic signal. However, cated in Table 1. First of all, multiples of the input frequencies
the increased richness in the output allows for effective identifica- 1 [Hz] and 5 [Hz] appear at 3 [Hz] and 15 [Hz] (harmonics). More-
tion and reduction/utilization of nonlinear effects. See for example over, the 1 [Hz] and 5 [Hz] components in (1) depend nonlinearly
[23,70,72,74,85]. Hence, classical frequency domain approaches for on the input amplitude (gain compression / expansion) and depend
LTI systems cannot be straightforwardly applied when nonlineari- on the input at other frequencies as well (desensitization). Finally,
ties are present and additional analysis is required to investigate if, the nonlinear mapping combines input frequencies and their sum
when and how similar methodologies can be used when analyzing and difference frequencies to new frequencies at 3,7,9 and 11 [Hz]
nonlinear systems. (intermodulation).
In the frequency domain, nonlinear effects manifest themselves These effects are visualized in Fig. 1 where the nonlinear de-
in different ways. The gain of a nonlinear system may, for example, formation of the signal u(t) is depicted for a = 2 and b = 3. Fig. 1a
depend on the amplitude of the input and the output may contain and b show the deformation of the signal in the time domain. The
 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24 13

Fig. 1. Time and frequency domain representation of a nonlinearly deformed harmonic signal.

corresponding frequency domain representations are depicted in where h p (τ1 , τ2 , . . . , τ p ) : R p → R is the pth order Volterra kernel
Fig. 1c and d. For a description of the different frequency domain and yp (t) is referred to as the corresponding pth order output. The
effects, see also [98]. Volterra series model (2) is defined for the class of Volterra sys-
tems, which is defined as follows.
Before starting the formal introduction, application and com-
parison of frequency domain methods that aim to capture these Definition 1 (VS : Volterra system). A dynamical system is called
nonlinear effects one may read Appendix A, which introduces the a Volterra system if there exists a Volterra series representation
necessary nomenclature. Next, the first of the four model classes, (2) that converges uniformly around a given working point, with
i.e. the Volterra series based approach, is addressed. probability 1 to y(t) for all bounded inputs u ∈ G , i.e.
 τ
lim E{|y(t ) − y℘(t )|} = 0 ∀ τ ∈ R≥0
℘→∞ 0
3. Volterra series based approaches where the expected value E{·} is the ensemble average over the
considered class of random inputs.
Modeling of nonlinear phenomena is often achieved by iden-
tifying a local approximation of the behavior by an approxima- (see: [78,82])
tive series. In case of continuous functions, a polynomial or Tay- The frequency domain representation corresponding to (2) is
lor approximation is often used and the existence and conver- obtained by applying the multiple Fourier transform to the pth or-
gence of such series is analyzed in the Stone–Weierstrass theo- der Volterra kernel. This yields the definition of the pth order Gen-
rem [87,88,94]. An extension of the series approximation for non- eralized Frequency Response Function (GFRF).
linear dynamical systems is based on the work of Vito Volterra
Definition 2 (T p ( p ): Generalized frequency response function
[92,93] which was further developed by Norbert Wiener during
(GFRF)). Consider a Volterra system according to Definition 1.
World War II [95]. This series approximation is referred to as the
Then, its pth order GFRF is denoted T p ( p ) : R p → C, with  p =
Volterra series [76,78]. In this section, the frequency domain repre-
(ξ1 , ξ2 , . . . , ξ p ) ∈ R p . The GFRF is defined as the p-dimensional
sentation of the Volterra series developed in the 1950s [18] is dis-
Fourier transform of the pth order Volterra kernel in (2), i.e.
cussed, which is referred to as the generalized frequency response
 ∞  ∞ p
function.
The Volterra series approximation generalizes the polynomial T p ( p ) = ··· h p ( τ1 , . . . , τ p ) e−2π iξm τm dτm (3)
−∞ −∞
m=1
approximation of ’static functions’ over a given interval, to the ap-
proximation of dynamical systems around a given working point (see: [6,16,18,78])
[27]. This series expansion captures the input-output dynamics of
The concept of the GFRF was first introduced in [18] as a fre-
a (nonlinear) dynamical system in a set of Volterra kernels that are
quency domain equivalent of the Volterra kernel, and represents
a generalization of the impulse response of LTI systems. Provided
a generalization of the FRF for LTI systems. Although identifica-
that a Volterra series exists and can be identified [10,21,39,77], the
tion of the FRF for LTI systems is well understood, identification
response up to order ℘, i.e. y℘(t ) ∈ R of such system to an input
of the GFRF is nontrivial. Nevertheless, several approaches exist to
u(t ) ∈ R is given by:
obtain the GFRF for different classes of nonlinear systems. First of
 ∞  ∞ p all, for some nonlinear models the GFRF can be computed analyt-
y p (t ) = ··· h p ( τ1 , τ2 , . . . , τ p ) u(t − τm ) dτm ically. Examples of such approaches are: rational nonlinear models
−∞ −∞ m=1 [100], nonlinear, autoregressive moving average models with ex-
℘ ogenous inputs (NARMAX) [58] and nonlinear integro-differential
y℘(t ) = y p (t ) (2) equations [5]. Moreover, the GFRF of a known NARMAX model may
p=1 also be obtained by application of the probing method [2,6,8] and
14 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24

more complex dynamical systems may be modeled by a series of Russian literature [12] and relates to the existence of stable and
piecewise valid GFRFs [42]. Finally, the GFRF of quadratic and cubic unique limit solutions of nonlinear systems. The notion of conver-
nonlinear systems can be obtained in a numerically efficient man- gent systems is introduced in [50,54] as follows:
ner by application of a symbolic expression of the output of such
Definition 3 (CS : Uniformly convergent systems). A time invariant
system as discussed in [41].
system (11) is said to be uniformly convergent for a class of input
In general, a parametric representation is often used to analyze
signals W if for every w ∈ W:
the GFRF. For example, in [33], a parameter extraction operator is
discussed that allows to study the influence of the parameters of 1. All solutions xw (t) are well-defined for all t ∈ [t0 ∞) and all ini-
a nonlinear model on the GFRFs. This operator is applied to an- tial conditions x0 ∈ Rn ;
alyze the relationship between different order GFRFs [28] and to 2. There exists a unique solution x̄w (t ) defined and bounded for
distinguish between input and output nonlinearities in a class non- all t ∈ (−∞ + ∞ );
linear models [31]. Moreover, it is used to characterize and effi- 3. The solution x̄w (t ) is uniformly globally asymptotically stable.
ciently compute the generalized output frequency response func-
tion as discussed in [29,30,57]. Where the GFRF describes the re- To emphasize the dependence on the input w(t), the limit solu-
lation between the input and output of Volterra systems, the gen- tion is denoted by x̄w (t ).
eralized output frequency response function relates the parameters
(see: [50,54])
of a nonlinear model to the output spectrum of the nonlinear sys-
In [51,53] convergence is discussed for interconnected and
tem [36–38].
piecewise affine systems, which allows to analyze tracking and
For Volterra systems, the GFRF allows for spectral analysis of
synchronization in this type of nonlinear system [97]. For uni-
the nonlinear dynamics. In [35], it is shown that it is possible to
formly convergent nonlinear systems, it is possible to relate a si-
predict the output frequency range for an arbitrary input signal
nusoidal input signal to the corresponding steady-state solution
using the GFRF. Moreover, a spectral analysis of block structured
and output of the system. These mappings are referred to as the
dynamical systems in terms of the GFRF is presented in [26] and
nonlinear state and output frequency response function (NSFRF
the GFRF can be used to attain a user defined frequency domain
and NOFRF) and are collectively referred to as the nonlinear fre-
performance for a nonlinear system as shown in [32,34]. Finally,
quency response function (NFRF). The NFRF was first introduced in
application of the GFRF allows to compute bounds on the output of
[55] and is defined as follows:
Volterra systems [3,4] and assess convergence of the corresponding
Volterra series [21,40]. Definition 4 (NS/O (γ sin(ωt ), γ cos(ωt ), ω ): Nonlinear state and
Although the GFRF allows for detailed spectral analysis of the output FRF (NFRF)). Consider the system (11) subject to a sinu-
behavior of Volterra systems their interpretation is nontrivial. In soidal input1 w(t ) = γ sin(ωt ) with frequency ξ = ω/2π and am-
[98] an interpretation of the nonlinear effects in Table 1 in terms plitude γ and assume that:
of the GFRF is presented. Moreover, techniques to visualize the
GFRFs up to order three are introduced in [7,99]. Finally, to illus- 1. f(x, w) is locally Lipschitz with respect to x;
trate the derivation of the GFRF the following example, taken from 2. The system is uniformly convergent for the class of bounded
[5], presents the computation of the GFRF for the Duffing oscillator continuous input signals;
(continued in [24,43]). 3. The system is input to state stable.

Then, there exist a continuous function

Example 2 (GFRF of the forced Duffing oscillator [5]). The GFRFs NS (γ sin(ωt ), γ cos(ωt ), ω ) : R3 → Rn , called the Nonlinear
can be analytically computed for nonlinear systems described by State Frequency Response Function (NSFRF), such that for any
nonlinear integro-differential equations. A well known example of sinusoidal input the steady state solution of (11) is given by:
such nonlinear system is the Duffing oscillator which dynamics are
given by: x̄w (t ) = NS (γ sin(ωt ), γ cos(ωt ), ω ) (5)
Furthermore, the Nonlinear Output Frequency Response Function
ÿ(t ) + 2ζ ωn y˙ + ωn2 y(t ) + ωn2 ε y3 (t ) = u(t ) (4) (NOFRF) is denoted by NO (γ sin(ωt ), γ cos(ωt ), ω ) : R3 → Rm and
relates a sinusoidal input to the corresponding steady state output.
with ζ , ωn , ε ∈ R and the input and output are denoted by u(t) and Using the definition of the NSFRF in (5), the NOFRF is defined as:
y(t) respectively. As shown in [5] the first three GFRFs of the sys-
tem in (4) equal: NO (γ sin(ωt ), γ cos(ωt ), ω ) = g(NS (γ sin(ωt ), γ cos(ωt ), ω ))
(see: [55])
1
T 1 ( ξ1 ) =
ωn2 + 4π ζ ωn iξ1 − 4π 2 ξ12 Numerically efficient methods are required to compute the pe-
riodic solutions required to derive the NFRF. In [1,52], examples
T 2 ( ξ1 , ξ2 ) = 0
of such methods are presented. When using the NFRF, a method
3 3
for visualization of the dynamics is required for analysis and con-
T3 (ξ1 , ξ2 , ξ3 ) = −εωn2 T1 ξ T 1 ( ξm ) troller design. For LTI systems, the Bode plot visualizes the FRF
=1 m=1
and is widely applied in both analysis and controller design. Using
the NOFRF, the concept of the Bode plot is extended to nonlinear
where T1 (ξ1 ) relates to the linear part of the dynamics in (4) and
systems in [55]. Therefore, consider the following definition of the
the cubic nonlinearity appears in T3 (ξ1 , ξ2 , ξ3 ).
nonlinear Bode plot.

Definition 5 (G(ω, γ ): Nonlinear Bode plot). Consider the system

(11), with scalar output y(t ) ∈ R and subject to a sinusoidal in-
4. Nonlinear frequency response function and Bode plot
put w ∈ S with frequency ξ = ω/2π and assume conditions 1–3 in
In this section a class of frequency response functions for non-
linear systems is introduced that is based on the concept of con- 1
In the remainder of this section the angular frequency ω = 2πξ is used for ease
vergent systems. The idea of convergent systems originates from of notation.
 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24 15

Definition 4 are satisfied. Then, the nonlinear Bode plot G(ω, γ ) :

R2 → R is defined as:

1
G (ω , γ ) = sup NO (γ sin(ωt ), γ cos(ωt ), ω )
γ t∈[− π π
)
ω ω

where NO (· ) is the NOFRF as in Definition 4.

(see: [55])

Remark. The nonlinear Bode plot represents a conservative esti-

mate of the system gain in the sense that the ratio between the
input amplitude and the spectral component of the output at the
input frequently may be smaller than the value shown in the non-
linear Bode plot.

The definition of the nonlinear Bode plot in Definition 5, al-

though based on the norm of a time domain signal, relates directly
to the gain of the FRF as visualized in the classical Bode plot for LTI
systems. However, where the classical Bode plot provides informa-
tion about the phase shift as well, such phase information cannot Fig. 2. Nonlinear Bode plot (7) of system (6).
be defined for nonlinear systems. Moreover, where system norms
can be computed from the FRF and Bode plot for LTI systems, this
is generally not possible for nonlinear systems based on the NFRF 5. Describing functions
and nonlinear Bode plot. The concept of the nonlinear Bode plot
is used to extend the notion of the sensitivity function to nonlin- For LTI systems the FRF fully captures the systems dynamics
ear systems in [56]. These generalized sensitivity functions have as the response to a single sinusoid can be extended to arbitrary
been applied for the design of variable gain control for optical stor- inputs using the principle of superposition. For nonlinear systems
age drives and wafer scanners [25,96] and feed forward design for superposition does not hold and the FRF fails to capture the com-
wafer scanners [22]. Finally, the following example is taken from plete dynamics. Hence, an alternative approach is required to ex-
[55] and concludes this section by illustrating the computation of tend the concept of the FRF to nonlinear systems. Describing Func-
the NFRF and nonlinear Bode plot for a given nonlinear system. tions (DF) aim to do so by describing the response to a given type
of input signal based on some frequency domain mapping from in-
Example 3 (nonlinear state and output FRF [55]). Consider the put to output. Although different types of describing functions ex-
following dynamical system, subject to a sinusoidal input u(t ) = ist (Fig. 3), their definition is generally similar to that of the FRF
γ sin(ωt ) with frequency ξ = ω/2π and amplitude γ . for LTI systems. The main difference between different describing
x˙ 1 = −x1 + x22 y = x1 functions originates from the signal and system class for which
(6)
x˙ 2 = −x2 + u they are defined and the type of nonlinear effects captured. In this
section a selection of describing functions is introduced, compared
As shown in [55], condition 1–3 in Definition 4 are satisfied for
and an experimental example is provided illustrating the identifi-
this system. Hence, the NSFRF exists and is unique. Next, us-
cation of one of the describing functions.
ing the results from [55] and denoting v(t ) = [sin(ωt ) cos(ωt )]T ,
In Fig. 3, three classes of describing functions are depicted:
the NSFRF, which describes the steady state solution x̄u (t ) =
the sinusoidal describing function, the generalized describing func-
[x̄1,u (t ) x̄2,u (t )]T of (6), equals:
  tion and the higher order sinusoidal describing function. The dif-
γ vT (t ) (ω ) v(t ) ferent branches represent different subclasses of DF such that for
NS (γ sin(ωt ), γ cos(ωt ), ω ) = γ 1
2 (ω ) v (t ) any two consecutive DF, the lower one is a special case of the
  upper one. The classical (two) sinusoidal input describing func-
2ω 4 + 1 ω 3 − 2ω tion (SIDF) models the response to (two) sinusoidal inputs at the
( ω ) = ( 1 + ω ) ( 1 + 4ω )
2 −2 2 −1
1
ω 3 − 2ω 2ω 4 + 5ω 2 corresponding excitation frequencies. The Generalized Describing
  Function (GDF) on the other hand, is based on the Volterra series
2 (ω ) = (1 + ω )
2 −1
1 −ω approximation of nonlinear systems and allows to model the re-
Using these results, and as y(t ) = g(x(t )) = x1 (t ), the correspond- sponse to multisine inputs. Finally, the Higher Order Sinusoidal De-
ing NOFRF, which describes the output corresponding to the steady scribing Functions (HOSDF) model the response to sinusoidal sig-
state solution ȳu (t ) = x̄1,u (t ) of (6), equals: nals at both the fundamental excitation frequency and higher har-
monics of this frequency. These describing functions are defined
NO (γ sin(ωt ), γ cos(ωt ), ω ) = γ 2 vT (t ) 1 (ω ) v(t ) for different classes of input signals and nonlinear systems and
Finally, the nonlinear Bode plot is obtained by (numerically) com- capture a different subset of nonlinear effects. In this section three
puting: describing functions are introduced in detail: the SIDF, GDF and
the Higher Order Sinusoidal Input Describing Function (HOSIDF).
1
G (ω , γ ) = sup γ 2 vT (t ) (ω ) v(t ) (7) First, consider the SIDF, which is defined for sinusaoidal input
γ t∈[− π π
)
1
signals and is defined as follows:
ω ω

and the result is depicted in Fig. 2. Definition 6 (DS (ξ , γ ): Sinusoidal input describing function
(SIDF)). Consider a nonlinear system2 (11), with y(t ) ∈ R, subject
Summarizing, the NSFRF relates any sinusoidal input to the cor-
to a sinusoidal input u ∈ S with frequency ξ 0 and amplitude γ and
responding steady state solution and the NOFRF yields the cor-
responding steady state output. Finally, the nonlinear Bode plot
yields the maximum gain of the nonlinear system for a sinusoidal 2
The authors in [17] define these describing functions for systems that ’have a
input with a given amplitude and excitation frequency. single loop configuration with separable linear and nonlinear elements’, here de-
16 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24

Fig. 3. Overview of a selection of different describing functions. The signal and system classes refer to the definitions used by the authors in the corresponding references.

define the systems steady state output y(t) and Fourier transforms ated by the order pth order GFRF, T p ( p ) by:
of the input and output U (ξ ), Y (ξ ) ∈ C. Then the sinusoidal in-  ∞  p−1
∞
put describing function DS (ξ , γ ) : R>0 × R → C is defined as:
Y p ( ξ ) = . . . Tp ξ1 , . . . , ξ p−1 , ξ − ξm ·
−∞ −∞
Y  ( ξ0 , γ ) m=1
D S ( ξ0 , γ ) =
U  ( ξ0 , γ ) p−1 p−1
U ξ− ξm U (ξ ) dξ1 . . . dξ p−1
(see: [17]) m=1 =1

The SIDF models the response to a sinusoidal input by relating (see: [60])
the spectral components in the input and output at the fundamen-
The GDF allows to model the response of Volterra systems to
tal excitation frequency. Hence, the SIDF captures gain compression
multisine input signals [59,60]. However, it is specific to the se-
/ expansion (see Table 1), but fails to model desensitization as the
lected input signal, which limits the validity of the model to the
input is a single sinusoid. Moreover, intermodulation is not cap-
excitation signal used. As the GDF allows for broadband input sig-
tured for the same reason and generation of harmonic components
nals, both gain compression / expansion and desensitization are
is not detected as only the response at the excitation frequency in
captured by the GDF. However, intermodulation and the genera-
considered.
tion of harmonics are not modeled as the GDF only models the
The GDF, on the other hand, captures a wider range of nonlin-
response at frequencies present in the input signal.
ear phenomena and allows for a broader class of input signals, but
Finally, the Higher Order Sinusoidal Input Describing Functions
is defined only for Volterra systems. Consider the following defini-
model the response of uniformly convergent nonlinear systems to
tion of the GDF:
a sinusoidal input at harmonics of the excitation frequency and are
Definition 7 (DG ( ξ , ς ): Generalized descr. function defined as follows:
(GDF)). Consider a Volterra system (11) subject to multisine Definition 8 (Hk (ξ , γ ): Higher order sinusoidal input describing
input u ∈ M scaled by a factor ς and output y(t). Moreover, define function (HOSIDF)). Consider a uniformly convergent, time invari-
the corresponding Fourier transforms of the input and output ant nonlinear system (11) subject to a sinusoidal input u ∈ S with
U (ξ ), Y (ξ ) ∈ C. Then, the generalized describing function frequency ξ 0 and amplitude γ . Next, define the systems steady
DG (ξ , ς ) : R2 → C is defined as: state output y(t) and single sided spectra of the input and output

℘ U (ξ ), Y (ξ ) ∈ C. Then, the kth higher order sinusoidal input de-
ς p−1 Y p (ξ ) scribing function Hk (ξ , γ ) : R>0 × R → C, k = 0, 1, 2, . . . is defined
p=1
DG ( ξ , ς ) = as:
U ( ξ ) Y (kξ0 , γ )
H k ( ξ0 , γ ) = (8)
where the GDF is defined for all ξ ∈ {ξ ∈ R| |U (ξ )| = 0} and the U k ( ξ0 , γ )
uni-dimensional spectrum Y p (ξ ) is related to the output gener- 
k
where U k (ξ0 , γ ) = U ( ξ0 , γ ) .
=1
(see: [46,71])
noted by SL. However, this describing function can be computed for all system
classes considered in this paper. Remark. Note that H1 (ξ0 , γ ) = DS (ξ0 , γ ) ∀ ξ0 > 0.
 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24 17

A typical series of HOSIDFs is depicted in Fig. 5a. Given the odd

nature of the friction nonlinearity in this case, the even HOSIDFs
are low. Fig. 5a hints that the system behaves more linearly for
increasing value of γ . This can be observed from the combina-
tion of a decreasing gradient ∂ H ˆ 1 /∂γ and a decreasing value of
Hˆ k (ξ , γ ) ∀ k > 1 for all measured frequencies. Finally, in Fig. 5b
|Hˆ 1 | is depicted as a function of input power and frequency.

6. Linear approximation in the presence of nonlinearities

In many applications, the use of LTI models is common practice,

mainly because methods to identify and interpret such models are
well developed [44] and many tools exist that allow controller de-
sign and synthesis based on LTI models [86]. However, most real
life systems are not linear and it is therefore required to assess the
validity of a linear approximation (see for example [14,63,79,82]).
In this section, the estimation of a linear model in the presence of
Fig. 4. Industrial high precision stage used in a transmission electron microscope.
nonlinearities is addressed.
Consider a nonlinear system as depicted in Fig. 6a and assume
the system can be transformed to a structure as in Fig. 6b. In
Fig. 6b, the output Y (ξ ) is assumed to exist of a component
Remark. Compared to the original definition in [46], the ampli-
that is generated by an LTI system Bu (ξ ), a distortion S u (ξ ) due
tude dependency |Hk | ∝ γ k inherent to the original HOSIDF model
to nonlinearities and an output disturbance N (ξ ). Note that the
structure is removed in [71] such that the HOSIDFs reveal only sys-
transformation in Fig. 6 is nontrivial and a (unique) transformation
tem characteristics.
may not exist. In the sequel, the existence and invariance of this
transformation is addressed. However, for now, consider the repre-
The HOSIDFs model the response to a sinusoidal input signal at
sentation in Fig. 6b and note that the nonlinear system is repre-
harmonics of the fundamental input frequency. Therefore both gain
sented by the combination of an LTI model Bu (ξ ) and a nonlinear
compression / expansion and the generation of harmonics are cap-
disturbance S u (ξ ), both of which depend on the power spectrum
tured in this describing function. However, as the excitation sig-
of the input signal.
nal is a single sinusoid, both desensitization and intermodulation
Bu (ξ ) is referred to as the Best Linear Approximation (BLA)
are not captured by the HOSIDFs. The HOSIDFs yield valuable in-
[14,82]. A BLA can be computed for a large class of input signals
formation about nonlinearities that can be used for modeling and
and nonlinear systems, but additional assumptions are required to
control purposes [47,48,71,72]. Finally, related studies as presented
assure invariance of the approximation with respect to a given
in [49] use the concept of the HOSIDF to find an input signal
class of input signals. In this section the BLA is therefore intro-
that yields a sinusoidal output signal, similar to Volterra based ap-
duced in two steps. First, the formal definition of the BLA is pro-
proaches presented in [32,34].
vided and then sufficient conditions are provided for a class of
A comparison of the describing functions discussed in this sec-
nonlinear systems and input signals for which the BLA exists and
tion is provided in Table 1. Next, this section is concluded by an
is invariant.
example that illustrates the identification of the HOSIDFs in prac-
The transformation in Fig. 6 is defined by the pair (Bu , S u ),
tice.
which is defined along the lines of [15,82].

Example 4 (experimental identification of the HOSIDFs). Consider Definition 9 ((Bu , S u ): Best Linear Approximation). Consider a
an industrial high precision stage (Fig. 4) used in a Transmis- nonlinear system (11), with y(t ) ∈ R and assume the transforma-
sion Electron Microscope (TEM). At the high resolution and re- tion in Fig. 6 exists, such that:
producibility that is required in a TEM, nonlinear effects start to
effect the performance of the system significantly. This motivated Y ( ξ ) = Bu ( ξ )U ( ξ ) + S u ( ξ ) + N ( ξ )
the measurements presented in this example. The motion stage is then, the Best Linear Approximation (BLA) is defined by the pair
a single input, single output system with the voltage to the current (Bu , S u ), such that:
amplifier as the input and the position of the stage as its output.
 2
The identification is performed in open loop as in [46]. Bu (ξ ) = arg min E{ y(t ) − B(ξ )u(t ) } (9)
The system is excited with a sinusoidal input and the response B (ξ )
is measured using a SigLab 20 − 42 dynamic signal analyzer. To
where the expected value E{·} is the ensemble average over the
measure the response of the system an encoder system is used
considered class of inputs.
that allows high resolution position measurements over the entire
stroke of the set-up. To measure a single HOSIDF Hk (ξ0 , γ ) for a (see: [15])
given frequency / amplitude combination, several sinusoidal input Definition 9 does not assure the existence of the BLA, i.e. of the
signals u ∈ S are applied. Repeating the experiment multiple times minimizer in (9). Moreover, although it is clear that the BLA de-
yields the average HOSIDF and variance on the average which are pends on the power spectrum of the input signal, invariance of the
denoted by H ˆ k (ξ0 , γ ) and ς 2 (ξ 0 , γ ) respectively.
BLA with respect to a class of input signals needs to be addressed
By repeating each experiment ten times with frequencies rang- as well. Therefore, consider the following class of excitation signals.
ing from
√ 5 [Hz] to 300 [Hz] in steps of 5 [Hz] and input signal pow-
ers γ / 2, ranging from 0.07 [V] to 1.41 [V] (logarithmically spaced) Definition 10 (Geq : Gaussian Riemann equivalent sig-
the HOSIDFs are identified. All measurements have been performed nals). Consider a power density spectrum Su (ξ ), which is piecewise
with a sampling frequency of 5120 [Hz] and a block length of 2048 continuous and has a finite number of discontinuities and consider
points, resulting in leakage free measurements. an excitation signal u(t) that equals:
18 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24

Fig. 5. Measured HOSIDF of the motion stage in a TEM.

Fig. 6. Equivalent representations of a nonlinear system.

1. A Gaussian noise excitation G with power density spectrum all bounded inputs u ∈ Geq , i.e.
Su (ξ ), or  τ  
1
2. A periodic, real valued signal u(t ) ∈ R such that U (±ξk ) = lim E |y(t ) − y℘(t )|2 = 0 ∀ τ ∈ R≥0
Ak e±iϕk , k ∈ N>1 , ξk ∈ R>0 and U (ξ ) = 0 otherwise. Moreover,
℘→∞ τ 0

Ak and ϕ k are random and mutually independent, E{eiϕk } = 0 where the expected value E{·} is the ensemble average over the
and either E{A2k } = fA2 (ξk ) or A2k = fA2 (ξk ) with fA (ξ ) a piecewise considered class of random inputs.
continuous function with a finite number of discontinuities.
Remark. This class of systems is referred to as special Wiener sys-
Then, the sets of signals 1 and 2 are Gaussian Riemann equiva- tems as it is closely related to the classical Wiener systems [78] for
lent if: which the requirement on continuity of the Fourier transforms is
 ζ2 not imposed and which considers only Gaussian inputs G ⊂ Geq in
E{|U (ξk )|2 } = Su (ξ ) dξ + O(NK−1 ) ∀ ζi ∈ R>0 the convergence criterium.
k∈K
ζ1
Using Definitions 10 and 11, the following sufficient conditions
with K = {k ∈ N>1 |ζ1 < ξk < ζ2 }, NK the number of elements in K, for the existence and invariance of the BLA are derived in [14].
0 < ζ 1 < ζ 2 and where E{·} denotes the expected value.
Lemma 1 (Invariance and existence of the BLA). Consider a nonlin-
(proof: [82]) ear system (11). If the system:
To assure that the minimizer in (9) and thus the BLA exists, the
class of nonlinear systems is restricted as follows (along the lines 1. Belongs to the class of special Wiener systems WS ;
of [14]): 2. Is subject to Gaussian Riemann equivalent input signals, i.e. u
∈ Geq ,
Definition 11 (WS : special Wiener system). System (11) is called a
special Wiener system if there exists a Volterra series representa- then, the best linear approximation (Definition 9) exists and the
tion (2), with continuous multidimensional Fourier transforms (3), pair (Bu , S u )is asymptotically invariant under inputs in the equiv-
that converges in mean square sense, with probability 1 to y(t) for alence class Geq . Moreover, if condition 1–2 are satisfied, the BLA
 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24 19

equals:
E {Y  ( ξ ) U 

( ξ )}
Bu ( ξ ) = (10)
E {U  ( ξ ) U 
 ( ξ )}

where  indicates the complex conjugate and the expected value

E{·} is the ensemble average over the considered class of inputs which
includes averaging over the effect of the measurement noise N .

Proof. (see3 [14]) 

Remark. Although Lemma 1 assures invariance of the BLA, the BLA

still specific to the fixed power spectrum of the set of input signals
that is used.

Remark. In the sequel, the BLA is considered under the conditions

in Lemma 1 which assures both invariance and existence of the
linear approximation but does not provide any condition about the
quality of the approximation. Fig. 7. Linear approximation identified with different powers of the input signal (–)
urms = 0.10 [V ], (- grey) urms = 0.50 [V ] and (- black) urms = 1.0 [V ].
Experimental identification of the BLA and the detection of
nonlinear effects in the frequency domain requires careful de-
Table 2
sign of the excitation signal and signal processing. An overview
Estimated variables (see [70]).
of techniques for the detection of nonlinearities is presented in
[91]. Moreover, excitation signal design is for example addressed Variable Description

in [19,20,79] and signal processing techniques are discussed in for Bu Best linear approximation
example [63,66,67,84]. Yˇ Output spectrum
The concept behind the detection of nonlinear effects in the fre- σN2
Variance on Yˇ due to noise
σS
2
Variance on Yˇ due to nonlinearities
quency domain is often based on two approaches: P( ) Power at non-excited odd (o) and even (e) lines
o/e

1. The presence of nonlinearities will cause a change in linear ap-

proximation of the dynamics when the input signal changes.
Hence, a variance analysis on an ensemble of appropriately se-
WS systems. Then, application of Gaussian Riemann equivalent in-
lected input signals allows to quantify the effects of nonlinear-
put signals, yields an invariant estimate of the BLA for a given
ities [13,90].
power spectrum of the input signal. In the following, random odd
2. Nonlinearities can be detected by analyzing the power appear-
multisine excitation signals are used [70], which are Gaussian Rie-
ing in the output spectrum at spectral lines that are not excited
mann equivalent.
by the input signal [62,68].
The identification procedure consists of exciting the system
Combining these approaches allows quantification of nonlinear with several periods of different realizations of a random multi-
effects and characterization of the nature of the nonlinearities. A sine. Computation of the averages and variances then yields the
detailed discussion of these methods is outside the scope of this BLA and an estimate of nonlinear influences based on: 1. Variance
work but an example is presented at the end of this section that information and 2. Energy appearing at non excited lines. The sig-
illustrates both methods. nal processing involved is discussed in [70] and yields the esti-
Both the identification of the BLA and means to detect nonlin- mated variables in Table 2.
earities are often discussed for single input, single output systems Ten realizations of the odd random phase multisine have been
that are identified in open loop. However, extensions for multi in- generated and the response has been measured for ten consecu-
put, multi output systems are presented in [81,83]. Moreover, iden- tive periods. Furthermore, this experiment is repeated for twenty
tification of the BLA and estimation of nonlinear effects in closed different rms values of the multisines which are logarithmically
loop is addressed in for example [61,65,69]. Finally, the following scaled between 0.3 [V] and 5.0 [V]. Measurements are performed
example concludes this section and illustrates the identification of using a SigLab measurement system with a measurement fre-
the BLA in practice. quency of 2560 [Hz] and a block length of 8192 measurement
points. The multisine signal is comprised of sinusoid with frequen-
Example 5 (Experimental identification of a high precision mo-
cies kξ0 = 0.3125 [Hz], k ∈ N. Finally, signal processing as in [70] is
tion stage). Consider again the high precision motion stage from a
applied to each measurement set, yielding a typical output spec-
transmission electron microscope as in Example 4. In the following
trum as depicted in Fig. 8a.
both the BLA of this system is identified and the effects of nonlin-
Fig. 8a shows that the power generated by nonlinear effects has
earities on the dynamics are quantified.
an average level that is 10 [dB] lower than the power generated in
First consider a conventional identification experiment using
the output spectrum by the BLA of the system. Analysis of the non-
band limited white noise with three different root mean square
excited lines shows that the nonlinearities have both an odd and
(rms) values. The estimates depicted in Fig. 7 clearly indicate non-
even nature, but the odd components dominate by almost 20 [dB].
linear behavior as the FRF changes with the rms value of the input
Furthermore, it becomes clear that the variation due to nonlinear-
signal. This indicates that further analysis of the validity of a linear
ities is of the same order of magnitude as the odd nonlinearities
model is required. Therefore, the BLA is identified in a setting that
that are detected. Finally the variation due to process and / or
allows to quantify nonlinear effects and assess the validity of this
measurement noise is almost 30 [dB] lower than the variation due
linear approximation.
to nonlinear effects.
In this example the BLA is identified under the assumptions
Fig. 8b depicts the BLA of the system as a function of both the
in Lemma 1. Hence, the system is assumed to be in the class of
rms value and frequency of the input. This figure indicates that for
high input power, the system behaves more linear. This becomes
3
See also [63,64,80]. clear from the reducing gradient ∂ |Bu |/∂ urms .
20 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24

Fig. 8. Estimated output spectrum and BLA for different excitation levels.

7. Comparison On a side note, in [11] a concept similar to convergence, called

Fading Memory (FM), is introduced. On the one hand, convergence
In this section a comparison between the different frequency relates to the existence, stability and uniqueness of a steady state
domain models and the corresponding signal and system classes solution, which is independent of initial conditions/ Hence, con-
is provided. First, the different system and signal classes for which vergence is concerned with the input to state properties of a state
the models are defined are compared. Next, a qualitative compari- space realization. FM, on the other hand, relates to the dependence
son between the frequency domain models is provided yielding an of the steady state solution on the input signals themselves and is
overview of the applicability of the different models and the non- concerned with the input to output properties of an operator. For
linear effects that can be detected using each of the model types. an operator that has a realization of some sort, the notion of FM
may allow to proof the existence of a unique steady state solution
for a subset of input signals and initial conditions [11]. However,
the following two arguments illustrate why FM and convergence
System and signal classes are not equivalent. First of all, an operator may not have a state
space realization (11). Hence, even if an operator has FM, a state
In Fig. 9 an overview of the signal and system classes discussed space realization may not exist, such that convergence cannot be
in this paper is provided. Both random and deterministic signals claimed. Secondly, an operator which has a realization (11) that in-
are considered. The relation between different signal classes is as cludes unstable non-observable dynamics may still have FM. How-
follows: By definition, the class of sinusoidal signals (Definition 16) ever, the corresponding realization cannot be convergent.
is a subset of the class of multisine signals (Definition 15). Simi-
larly, the class of Gaussian signals (Definition 14) is a subset of the
class of Gaussian Riemann equivalent signals (Definition 10). Frequency domain models
Next, consider the comparison of the different system classes
depicted in Fig. 9b. The classes of Volterra and special Wiener sys- This section provides a comparison of the frequency domain
tems can directly be compared as both prescribe the same type methods for nonlinear systems considered. First, consider the
of mathematical model. The classes of Volterra and special Wiener overview presented in Table 3. This table depicts the applicability
systems partially overlap, but neither is a subset of the other, e.g. of each model type for each combination of input signal and class
Volterra systems with discontinuous Fourier transforms are in the of nonlinear systems. For all model types, apart from the SIDF, the
class of special Wiener systems. The class of Wiener systems in- definitions used in literature are used to generate Table 3.
cludes the classes of Volterra and special Wiener systems as the The GFRF is applicable for all signal types considered, but
continuity condition on the Fourier transform is relaxed in this is limited to modeling Volterra systems. The corresponding GDF
case. is applicable to Volterra systems as well, but valid for multi-
The class of convergent systems can not be straightforwardly be sine inputs only. A detailed analysis connecting the GFRF and the
related to the classes of Volterra and special Wiener systems. Con- HOSIDFs can be found in [73].
vergence is a property of a state space realization of a system and For convergent systems the HOSIDF and NFRF allow to model
concerns the existence and stability of a unique limit solution of the systems dynamics, but these are restricted to sinusoidal inputs.
the nonlinear system. This requires the definition of initial condi- The SIDF can be computed for all system classes considered in this
tions and a well defined notion of a limit solution solution. As the paper, but is restricted to sinusoidal inputs as well. Finally, the
Volterra series does not include the definition of an initial condi- BLA when considered under the conditions in Lemma 1, is defined
tion it can not straightforwardly be related to the notion of con- for special Wiener systems subject to Gaussian Riemann equivalent
vergence. signals.
 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24 21

Fig. 9. Comparison of the system and signal classes.

Table 3
Model types ordered by signal and system class (according to definitions used in references).

Signal Deterministic Random

system S (sinusoidal) M (multisine) Geq (Gaussian Riemann eq.)

VS (Volterra) GFRF GFRF GFRF

GDF GDF
SIDF
WS (Special Wiener) SIDF BLA
CS (Convergent) HOSIDF
NFRF
SIDF

Moreover, Table 1 provides an overview of the information puts are also required, but detection must appear at non-excited
present in each model type. To fully reconstruct the response spectral lines as well. Hence, the GFRF is able to capture inter-
of a nonlinear system to a given input signal the corresponding modulation and the signal processing related to the BLA can be
model needs to model both the response at excited and non ex- adapted to detect intermodulation. For the detection of higher har-
cited frequencies and both phase and gain information needs to monics a sinusoidal excitation suffices, but detection must appear
be recoverable from the model. This is the case for the GFRF and at non-excited harmonic spectral lines as captured in the HOSIDF.
this model therefore allows to fully reconstruct the response of a Moreover, the GFRF models the generation of harmonics and these
given Volterra system to a class of input signals. The HOSDIF and can be detected by adapting the signal processing related to the
NFRF also allow to fully reconstruct the steady state response of BLA as well. Finally, the appearance of harmonics in the systems
a uniformly convergent system to any given sinusoidal input sig- output are captured by the NFRF, but the effects of gain compres-
nal although the representation of the nonlinear dynamics in both sion / expansion and the generation of harmonics are hard to dis-
model types is different. However, the nonlinear Bode plot corre- tinguish in the corresponding nonlinear Bode plot.
sponding to the NFRF does not allow to reconstruct the full re-
sponse as all phase information is lost in this representation. Fi- Applicability
nally, the SIDF, GDF and BLA do not allow to fully reconstruct the
output of a nonlinear system as information about spectral compo- In the preceding sections an overview and comparison of four
nents at non-excited frequencies is not available in these models. main approaches to the frequency domain analysis of nonlinear
Finally, Table 1 also presents an overview of the ability of the systems was presented. To conclude this paper, a discussion con-
different models to detect the different effects of nonlinearities in cerning the application of the different methods is presented. The
the frequency domain. As becomes clear, all methods considered following reflects the personal opinion of the author and may serve
are able to model gain compression and expansion. Detection of as a guideline for those seeking to use any of the frequency do-
desensitization however, requires multi-tone inputs and detection main methods discussed in this paper.
at exited spectral lines. Hence, only the GFRF, GDF and BLA are able First of all, the GFRF provides a useful tool for the analysis of
to model desensitization. To model intermodulation, multi-tone in- Volterra systems and allows to model the response to an exten-
sive class of input signals. However, the practical application of the
22 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24

GFRF is limited as identification of high order kernels / GFRFs is continuous time, causal, time invariant and are assumed to have a
difficult. Moreover, even if a model is known, the GFRF may be realization of the form:
used for prediction, but the model structure is complex, which lim- x˙ (t ) = f (x(t ), w(t )) x(t0 ) = x0
(11)
its application to practical analysis or control design. Secondly, the y(t ) = g(x(t ), w(t ))
nonlinear Bode plot provides means to assess the gain of a non-
linear system and may be used for sensitivity analysis as well [96]. where x(t ) ∈ Rn are the states, y(t ) ∈ Rm the outputs, w ∈ B p the
However, this model is only valid for sinusoidal inputs and the cor- input of the system and f : Rn × R → Rn and g : Rn × R → Rm . A
responding NFRF can be hard to identify in practice. solution of (11), corresponding to an input w(t), is denoted by
Thirdly, the HOSIDF are easily identifiable and provide a fast xw (t). The corresponding limit solution (if it exists) is denoted
method to assess nonlinear effects in practice. However, although by x̄w (t ), such that limt→∞ x̄w (t ) − xw (t ) = 0 ∀ x0 ∈ Rn . Further-
the HOSIDF provide some means to assess performance degrading more, the output corresponding to the limit solution is denoted by
nonlinear effects, this model is valid for sinusoidal inputs only. This ȳw (t ) = g(x̄w (t ), w(t )). Finally, a system (11) is called Linear and
limits their (real-time) application when non sinusoidal inputs are Time Invariant (LTI) if it satisfies the principles of superposition
required. Finally, the BLA provides means to attain a well defined and homogeneity [101].
linear approximation of the nonlinear dynamics as well as a mea-
Signal classes
sure of the quality of this linear approximation. This yields a use-
ful method to identify LTI approximations of weakly nonlinear sys-
The analysis presented in the sequel makes use of several signal
tems. However, if severe nonlinearities are present or small non-
classes and transforms. Hence, consider a real valued signal z(t ) ∈
linearities negatively influence the performance of the system, the
R and define the corresponding Fourier transform as follows:
BLA provides insufficient information to address or model these ef-
fects. Definition 12 (Z (ξ ): Fourier transform). Consider a signal z(t).
Summarizing, if an LTI approximation of the systems dynamics Then its Fourier transform Z (ξ ) is given by:
is required, identification of the BLA is recommended as it is easy  ∞
to identify, yields a widely valid linear approximation and pro- Z ( ξ ) = z(t )e−2π iξ t dt (12)
−∞
vides a corresponding quality measure. If more information about √
the nonlinear behavior is required, the nonlinear Bode plot allows with ξ ∈ R the frequency in [Hz], i = −1 and the corresponding
for a gain and sensitivity analysis of the nonlinear system. Finally, scaling of the transform equal to one. The corresponding inverse
if even more information about the nonlinearity is required, the Fourier transform is then given by:
HOSIDFs provide gain and phase information about the nonlinear  ∞
behavior, allowing analysis and control of the nonlinear effects. z(t ) = Z  ( ξ ) e 2 π iξ t d ξ (13)
−∞

8. Conclusion Moreover, for real valued signals, all information available in

the Fourier transform is captured in the single sided spectrum, i.e.
This paper addresses the comparison of four well established
Definition 13 (Z (ξ ): single sided spectrum). Consider a signal
frequency domain methods for nonlinear systems: Volterra beased
z(t ) ∈ R and its Fourier transform Z (ξ ). Then, the single sided
models, nonlinear frequency response functions, describing func-
spectrum Z (ξ ), equals:
tions and linear approximations in the presence of nonlinearities. 
The main contributions are summarized in Tables 1 and 3, which 2Z ( ξ ) ξ >0
compare the information captured in each modal and the signal / Z (ξ ) = Z ( ξ ) ξ =0
system classes for which each model is valid. The consistent in- 0 ξ <0
troduction of each model, accompanied by (experimental) exam-
In the following, several signal classes are used. The classes of
ples and their comparison provides insight into the differences and
Gaussian, multisine and sinusoidal signals are of particular impor-
similarities between alternative approaches and their applicability
tance and are therefore defined below. First consider the sequence
in practice.
z(tk ), which equals a Gaussian signal if it is defined as follows:
Appendix A. Nomenclature and definitions Definition 14 (G : Gaussian signal). Let G denote the set of Gaus-
sian signals. A continuous time signal z(t) is said to be a Gaus-
In the following, signals are real-valued, scalar and denoted by sian random process if, for every set of fixed times {τ n }, the ran-
non-capitalized roman letters, e.g. u(t ) ∈ R. Frequency domain rep- dom variables x(τ n ) follows a multidimensional normal distribu-
resentations of time domain signals, such as the Fourier transform tion with a user defined power density spectrum Sz (ξ ).
and other spectra are denoted by corresponding capitalized calli-
graphic font, e.g. U (ξ ) ∈ C. Furthermore, transfer functions are de- Next, consider a real valued signal z(t) which is composed of a
noted by Fractur font, e.g. H(ξ ) ∈ C (unless specified otherwise). finite number of sinusoidal components. In general such signal is
Moreover, signal classes are denoted by single blackboard bold cap- referred to as a multisine and in the special case where the sinu-
ital letter, e.g. S and system classes by multiple, overlined black- soidal components are harmonically related, the resulting signal is
board bold capital letter, e.g. VS . The sets of real, natural and com- called harmonic, i.e.
plex numbers are denoted as usual by R, N and C respectively and Definition 15 (M : Multisine). Let M denote the set of
finally, differentiation with respect to time is denoted by ˙ , i.e. multisine signals. A signal z(t) is a multisine if z(t ) =
x˙ (t ) = dt
d
x(t ). K
k=1 γk cos (2π ξk t + ϕk ), t ∈ R, for some ξk , γk ∈ R>0 , and ϕk ∈ R.
Furthermore, a signal z(t) is called harmonic if it is a multisine
System class
and ξk = kξ0 , k ∈ N for some ξ0 ∈ R>0 .

Let B p denote the space of piecewise continuous, bounded func- Finally, a special type of multisine signal is the single sine or
tions4 R≥t0 → R, t0 ∈ R. In the following all systems considered are sinusoidal signal, which is defined as follows:

Definition 16 (S : Sinusoidal signal). Let S ⊂ M denote the set

4
Where R≥t0 = {t ∈ R|t ≥ t0 }. of sinusoidal signals. A signal z(t) is sinusoidal if z(t ) =
 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24 23

Table 4 [20] Guillaume P, Schoukens J, Pintelon R, Kollár I. Chrest-factor minimization

Suggested references for each research area. using nonlinear Chebyshev approxmiation methods. IEEE T Instrum Meas
1991;40(6):982–9.
GFRF NFRF DF BLA [21] Guo L, Guo Y, Billings S, Coca D, Lang Z. A volterra series representation for a
Background [16] [54] [17] [63] class of nonlinear infinite dimensional systems with periodic boundary con-
[18] ditions. Syst Control Lett 2013;62(2):115–23.
[22] Heertjes M, Hennekens D, Steinbuch M. Mimo feed-forward design in wafer
[76]
scanners using a gradient approximation-based algorithm. Control Eng Pract
[78]
2010;18:495–506.
[92] [23] Ho C, Lang Z-Q, Billings S. Design of vibration isolators by exploiting
[93] the beneficial effects of stiffness and damping nonlinearities. J Sound Vib
[95] 2014;333(12):2489–504.
Introduction [98] [55] [17] [63] [24] Ho C, Lang Z-Q, Billings S. A frequency domain analysis of the effects
[46] of nonlinear damping on the duffing equation. Mech Syst Signal Process
[59] 2014;45(1):49–67.
Analysis [33] [52] [47] [81] [25] Hunnekens B, Heertjes M, van de Wouw N, Nijmeijer H. Performance opti-
[38] [56] [60] [64] mization of piecewise affine variable-gain controllers for linear motion sys-
[80] tems. Mechatronics 2014;24:648–60.
[82] [26] Jing X. Frequency domain analysis and identification of block-oriented non-
[71] [14] linear systems. J Sound Vib 2011;330(22):5427–42.
[27] Jing X, Lang Z. Frequency domain analysis and design of nonlinear systems
Applications [6] [96] [48] [13]
based on Volterra series expansion. Springer; 2015.
[7] [22] [49] [68]
[28] Jing X, Lang Z, Billings S. Mapping from parametric characteristics to gen-
[8] [25] [75]
eralized frequency response functions of non-linear systems. Int J Control
2008;81(7):1071–88.
[29] Jing X, Lang Z, Billings S. Output frequency response function-based analysis
for nonlinear volterra systems. Mech Syst Signal Pr 2008;22(1):102–20.
γ cos (2π ξ0 t + ϕ0 ), t ∈ R, for some γ , ξ0 ∈ R>0 , ϕ0 ∈ R. Finally, [30] Jing X, Lang Z, Billings S. Determination of the analytical parametric relation-
Sξ0 ⊂ S denotes the subset of sinusoidal signals with frequency ξ 0 . ship for output spectrum of Volterra systems based on its parametric charac-
teristics. J Math Anal Appl 2009;351(2):694–706.
[31] Jing X, Lang Z, Billings S. Parametric characteristic analysis for generalized
Appendix B. Literature overview frequency response functions of nonlinear systems. Circ Syst Signal Process
2009;28(5):699–733.
[32] Jing X, Lang Z, Billings S. Output frequency properties of nonlinear systems.
Table 4 provides an overview of the main literature related to Int J Nonlinear Mech 2010;45(7):681–90.
each model. [33] Jing X, Lang Z, Billings S, Tomlinson G. The parametric characteris-
tic of frequency response functions for nonlinear systems. Int J Control
2006;79(12):1552–64.
References [34] Jing X, Lang Z, Billings S, Tomlinson G. Frequency domain analysis for sup-
pression of output vibration from periodic disturbance using nonlinearities. J
[1] Alcorta-García E, Rodríguez-Alfaro L, Diaz-Romero D, Posadas-Castillo C. Sound Vib 2008;314(3–5):536–57.
Characterization of frequency response functions for nonlinear convergent [35] Lang Z-Q, Billings S. Output frequencies of nonlinear systems. Int J Control
systems. In: Proc conference on control and fault-tolerant systems; 2010. 1997;67(5):713–30.
p. 630–5. [36] Lang Z-Q, Billings S. Evaluation of output frequency responses of nonlinear
[2] Bedrosian E, Rice S. Output properties of Volterra systems (nonlinear sys- systems under multiple inputs. IEEE T Circuits II 20 0 0;47(1):28–38.
tems with memory) driven by harmonic and Gaussian inputs. Proc IEEE [37] Lang Z-Q, Billings S. Output frequency characteristics of nonlinear systems.
1971;59(12):1688–707. Int J Control 1996;64(6):1049–67.
[3] Billings S, Lang Z-Q. A bound for the magnitude characteristics of nonlin- [38] Lang Z, Billings S, Yue R, Li J. Output frequency response function of nonlinear
ear output frequency response functions: part 2. Practical computation of the Volterra systems. Automatica 2007;43(5):805–16.
bound for systems described by the nonlinear autoregressive model with ex- [39] Lesiak C, Krener A. Existence and uniqueness of Volterra seris vor nonlinear
ogenous input. Int J Control 1996;65(3):365–84. systems. IEEE T Automat Contr 1978;AC-23(6):1090–5.
[4] Billings S, Lang Z-Q. A bound for the magnitude characteristics of nonlinear [40] Li L, Billings S. Analysis of nonlinear oscillators using volterra series in the
output frequency response functions: part 1. Analysis and computation. Int J frequency domain. J Sound Vib 2011;330:337–55.
Control 1996;65(2):309–28. [41] Li L, Billings S. Estimation of generalized frequency response func-
[5] Billings S, Peyton Jones JC. Mapping non-linear integro-differential equations tions for quadratically and cubically nonlinear systems. J Sound Vib
into the frequency domain. Int J Control 1990;52:863–79. 2011;330(3):461–70.
[6] Billings S, Tsang K. Spectral analysis for non-linear systems, part i: parametric [42] Li L, Billings S. Piecewise Volterra modeling of the duffing oscillator in the
non-linear spectral analysis. Mech Syst Signal Pr 1989;3(4):319–39. frequency-domain. Mech Syst Signal Pr 2012;26(1):117–27.
[7] Billings S, Tsang K. Spectral analysis for non-linear systems, part ii: inter- [43] Li L, Billings S. Piecewise Volterra modeling of the duffing oscillator in the
pretation of non-linear frequency response functions. Mech Syst Signal Pr frequency-domain. Mech Syst Signal Process 2012;26(0):117–27.
1989;3(4):341–59. [44] Ljung L. System identification, theory for the user. Prentice hall; 1999.
[8] Billings S, Tsang K, Tomlinson G. Spectral analysis for non-linear systems, part [45] Meyer Y, Cumunel G. Active vibration isolation with a mems device. Effects of
iii: case study examples. Mech Syst Signal Pr 1990;4(1):3–21. nonlinearities in control efficiency. Smart Matter Struct 2015;24(8):085004.
[9] Bode H. Network analysis and feedback amplifier design. Van Nostrand; 1945. [46] Nuij P, Bosgra O, Steinbuch M. Higher-order sinusoidal input describing func-
[10] Boyd S, Tang Y, Chua LO. Measuring volterra kernels. IEEE T Circuits Syst tions for the analysis of non-linear systems with harmonic responses. Mech
1983;CAS-30(8):571–7. Syst Signal Pr 2006;20(8):1883–904.
[11] Boyd S, Chua LO. Fading memory and the problem of approximat- [47] Nuij P, Steinbuch M, Bosgra O. Measuring the higher order sinusoidal input
ing nonlinear operators with volterra series. IEEE T Circuits Syst describing functions of a non-linear plant operating in feedback. Control Eng
1985;CAS-32(11):1150–61. Pract 2008;16(1):101–13.
[12] Demidovich B.. Lectures on stability theory; 1967. Nauka, Moscow (in Rus- [48] Nuij P, Steinbuch M, Bosgra O. Experimental characterization of the
sian). stick/sliding transition in a precision mechanical system using the third or-
[13] D’haene T, Pintelon R, Schoukens J, Van Gheem E. Variance analysis of fre- der sinusoidal input describing function. Mechatronics 20 08;18(2):10 0–10.
quency response function measurements using periodic excitations. IEEE T [49] Nuij P, Steinbuch M, Bosgra O. Non-parametric identification of higher order
Instrum Meas 2005;54(4):1452–6. sinusoidal output describing functions. In: Proc international conference on
[14] Schoukens J, Lataire J, Pintelon R, Vandersteen G, Dobrowiecki T. Robustness noise and vibration engineering; 2008. p. 2361–71.
issues of the best linear approximation of a nonlinear system. IEEE Trans In- [50] Pavlov A, Pogromsky A, van de Wouw N, Nijmeijer H. Convergent
strum Meas 2009;58(5):1737–45. dynamics, a tribute to Boris Pavlovich Demidovich. Syst Control Lett
[15] Enqvist M, Ljung L. Linear approximations of nonlinear fir systems for sepa- 2004;52(3–4):257–61.
rable input processes. Automatica 2005;41(3):459–73. [51] Pavlov A, Pogromsky A, Van De Wouw N, Nijmeijer H. On convergence prop-
[16] Eykhoff P. System identification - parameter and state estimation. John Wiley erties of piecewise affine systems. Int J Control 2007;80(8):1233–47.
& Sons; 1974. [52] Pavlov A, van de Wouw N. Fast computation of frequency response functions
[17] Gelb A, Vander Velde W. Multiple input describing functions and nonlinear for a class of nonlinear systems. In: Proc conference on decision and control;
system design. McGraw Hill; 1968. 2008. p. 1180–6.
[18] George D. Continuous nonlinear systems. Technical Report MIT, USA; 1959. [53] Pavlov A, van De Wouw N, Nijmeijer H. Convergent systems: analysis and
[19] Godfrey Keditor. Perturbation signals for system identification. Prentice hall; synthesis. In: Lect Notes control inf sci, 322; 2005. p. 131–46.
1993.
24 D. Rijlaarsdam et al. / Mechatronics 42 (2017) 11–24

[54] Pavlov A, van de Wouw N, Nijmeijer H. Uniform output regulation of nonlin- [77] Sandberg I. Mathematical foundation of associated expansions for mildly non-
ear systems - a convergent dynamics approach. Systems and control: founda- linear systems. IEEE T Circuits Syst 1983;CAS-30(7):441–55.
tions & applications. Birkhauser; 2006. [78] Schetzen M. The Volterra and Wiener theories of nonlinear systems. John Wi-
[55] Pavlov A, van de Wouw N, Nijmeijer H. Frequency response functions for non- ley & Sons, Inc.; 1980.
linear convergent systems. IEEE T Autom Control 2007;52(6):1159–65. [79] Schoukens J, Dobrowiecki T. Design of broadband excitation signals with a
[56] Pavlov A, Van De Wouw N, Pogromsky A, Heertjes M, Nijmeijer H. Frequency user imposed power spectrum and amplitude distribution. In: Proc IEEE in-
domain performance analysis of nonlinearly controlled motion systems. In: strumentation and measurement technologie conference, 2; 1998. p. 1002–5.
Proc. CDC; 2007. p. 1621–7. [80] Schoukens J, Dobrowiecki T, Pintelon R. Parametric and nonparametric identi-
[57] Peng Z, Lang Z, Zhao L, Billings S, Tomlinson G, Guo P. The force transmis- fication of linear systems in the presence of nonlinear distortions, a frequency
sibility of mdof structures with a non-linear viscous damping device. Int J domain approach. IEEE Trans Autom Control 1998;43(2):176–90.
Non-Lin Mech 2011;46(10):1305–14. [81] Dobrowiecki T, Schoukens J. Measuring a linear approximation to weakly
[58] Peyton Jones J, Billings S. Recursive algorithm for computing the frequency nonlinear mimo systems. Automatica 2007;43(10):1737–51.
response of a class of non-linear difference equation models. Int J Contr [82] Schoukens J, Pintelon R, Dobrowiecki T, Rolain Y. Identification of linear sys-
1989;50(5):1925–40. tems with nonlinear distortions. Automatica 2005;41(3):491–504.
[59] Peyton Jones J, Billings S. Interpretation of non-linear frequency response [83] Schoukens J, Pintelon R, Vandersteen G, Rolain Y. Design of excitation signals
functions. Int J Control 1990;52(2):319–46. for nonparametric measurements on mimo-systems in the presence of non-
[60] Peyton Jones J, Billings S. Describing functions, Volterra series, and the linear distortions. In: Proc international instrumentation and measurement
analysis of non-linear systems in the frequency domain. Int J Control technology conference; 2010. p. 1–5.
1991;53(4):871–87. [84] Schoukens J, Rolain Y, Pintelon R. Analysis of windowing/leakage effects in
[61] Pintelon R, Barbé K, Vandersteen G, Schoukens J. Improved (non-)parametric frequency response function measurements. Automatica 2006;42(1):27–38.
identification of dynamic systems excited by periodic signals. Mech Syst Sig- [85] Setiadi A, Rijlaarsdam D, Nuij P, Schoukens J, Steinbuch M. Frequency domain
nal Pr 2011;25(7):2683–704. based real-time performance optimization of Lur’e systems. Mech Syst Signal
[62] Pintelon R, Rolain Y, Vandersteen G, Schoukens J. Experimental characteriza- Process 2014;42(1):58–70.
tion of operational amplifiers: a system identification approach - ii: calibra- [86] Skogestad S, Postlethwaite I. Multivariable feedback control. John Wiley &
tion and measurements. IEEE T Instrum Meas 2004;53(3):863–76. Sons; 2005.
[63] Pintelon R, Schoukens J. System identification: a frequency domain approach. [87] Stone M. The generalized Weierstrass approximation theorem. Math Mag
IEEE Press; 2001. 1948;21(4):167–84.
[64] Pintelon R, Schoukens J. Measurement and modelling of linear systems in the [88] Stone M. The generalized Weierstrass approximation theorem. Math Mag
presence of non-linear distortions. Mech Syst Signal Pr 2002;16(5):785–801. 1948;21(5):237–54.
[65] Pintelon R, Schoukens J. The best linaer appoximation of nonlinear systems [89] Sun X, Jing X. Nonlinear vibration isolator achieving high-static-low-dynamic
operating in feedback. In: Proc IEEE I2MTC; 2012. p. 2092–7. stiffness and tunable anti-resonance frequency band. Mech Syst Signal Pro-
[66] Pintelon R, Schoukens J, Vandersteen G, Barbe K. Estimation of nonparamet- cess 2016;80:166–88.
ric noise and frf models for multivariable systems-part i: theory. Mech Syst [90] Vanhoenacker K, Dobrowiecki T, Schoukens J. Design of multisine excitations
Signal Pr 2010;24(3):573–95. to characterize the nonlinear distortions during frf-measurements. IEEE T In-
[67] Pintelon R, Schoukens J, Vandersteen G, Barbe K. Estimation of nonparamet- str Meas 2001;50(5):1097–102.
ric noise and frf models for multivariable systems-part ii: extensions, appli- [91] Vanhoenacker K, Schoukens J, Swevers J, Vaes D. Summary and comparing
cations. Mech Syst Signal Pr 2010;24(3):596–616. overview of techniques for the detection of non-linear distortions. In: Proc
[68] Pintelon R, Vandersteen G, De Locht L, Rolain Y, Schoukens J. Experimental ISMA; 2002. p. 1241–55.
characterization of operational amplifiers: a system identification approach - [92] Volterra V. Sopra le funzioni che dipendono de altre funzioni. Rend R
i: theory and simulations. IEEE T Instrum Meas 2004;53(3):854–62. Academia dei Lincei 2 Sem 1887. 141–146 and 153–158
[69] Pintelon R, Vandersteen G, Schoukens J, Rolain Y. Improved (non-)parametric [93] Volterra V. Theory of functionals and of integral and integro-differential equa-
identification of dynamic systems excited by periodic signals, the multivariate tions. Dover Publications Inc; 1959.
case. Mech Syst Signal Pr 2011;25(8):2892–922. [94] Weierstrass K. Über die analytische darstellbarkeit sogenannter willkür-
[70] Rijlaarsdam D. Frequency domain based performance optimization of sys- licher functionen einer reellen veränderlichen. Sitzungsberichte der Königlich
tems with static nonlinearities. Dept. Mech. Eng., Eindhoven Univ. Ph.D. thesis Preußischen Akademie der Wissenschaften zu Berlin 1985;II. 633–639 and
Technol., Eindhoven, The Netherlands; 2012. 789–805
[71] Rijlaarsdam D, Nuij P, Schoukens J, Steinbuch M. Spectral analysis of block [95] Wiener N. Response of a nonlinear system to noise. Tech. Rep.. Radiation Lab
structured nonlinear systems and higher order sinusoidal input describing MIT (restricted, declassified July 1946, no. PB-1-58087); 1942.
functions. Automatica 2011;47(12):2684–8. [96] Wouwvan de Wouw N, Pastink H, Heertjes M, Pavlov A, Nijmeijer H. Perfor-
[72] Rijlaarsdam D, Nuij P, Schoukens J, Steinbuch M. Frequency domain based mance of convergence-based variable-gain control of optical storage drives.
nonlinear feed forward control design for friction compensation. Mech Syst Automatica 2008;44(1):15–27.
Signal Pr 2012;27(2):551–62. [97] Wouwvan de Wouw N, Pavlov A. Tracking and synchronisation for a class of
[73] Rijlaarsdam D, Oomen T, Nuij P, Schoukens J, Steinbuch M. Uniquely PWAsystems. Automatica 2008;44(11):2909–15.
connecting frequency domain representations of given order polynomial [98] Yue R, Billings S, Lang Z-Q. An investigation into the characteristics of non–
Wiener–Hammerstein systems. Automatica 2012;48(9):2381–4. linear frequency response functions. Part 1: understanding the higher dimen-
[74] Rijlaarsdam D, Setiadi A, Nuij P, Schoukens J, Steinbuch M. Frequency do- sional frequency spaces. Int J Control 2005;78(13):1031–44.
main-based nonlinearity detection and compensation in lure systems. Int J [99] Yue R, Billings S, Lang Z-Q. An investigation into the characteristics of
Robust Nonlinear Control 2013;23:1168–82. non-linear frequency response functions. part 2: new analysis methods
[75] Rijlaarsdam D, van Loon S, Nuij P, Steinbuch M. Nonlinearities in industrial based on symbolic expansions and graphical techniques. Int J Control
motion stages - detection and classification. In: Proc American control con- 2005;78(14):1130–49.
ference; 2010. p. 6644–9. [100] Zhang H, Billings S, Zhu Q. Frequency response functions for nonlinear ratio-
[76] Rugh W. Nonlinear system theory: the Volterra / Wiener approach. Johns nal models. Int J Control 1995;61(5):1073–97.
Hopkins University Press; 1981. [101] Zhou K, Doyle J, Glover K. Robust and optimal control. Prentice Hall; 1996.