MCE693/793: Analysis and

Control of Nonlinear Systems

Introduction to Describing Functions

Hanz Richter
Mechanical Engineering Department
Cleveland State University
Introduction

Frequency domain methods are essential for analysis and design of linear
systems. Stability, robustness and performance can be studied using classical
and modern tools such as
1. Bode, Nyquist and Nichols plots
2. Singular values and sensitivity plots (multivariable systems)
3. H∞ control for performance
4. µ-synthesis analysis and control for robust performance
5. Quantitative feedback theory (frequency domain design on the Nichols
chart)
Describing functions provide an approximate method for frequency-domain
analysis of nonlinear systems.

We will focus on limit cycle detection/creation or avoidance.

2 / 31
Basic Idea

In linear systems, transfer functions become frequency response functions

when the Laplace variable s is restricted to the imaginary axis.

That is, the Fourier transform is the Laplace transform restricted to the
imaginary axis. Frequency response functions in linear systems are functions of
frequency only:

s−1 1 − w2 2w
G(s) = , G(jw) = 2
+ 2
j
s+1 1+w 1+w
In control systems containing nonlinearities, we seek to find a frequency
response function for the nonlinear block. Due to nonlinearity, this function
will be amplitude-dependent.

0 φ(e)
e G(s)
−
{
{
N (A, w) G(jw)

3 / 31
Memoryless vs. Memory, Soft vs. Hard Nonlinearities

A memoryless nonlinearity produces an output which depends only on the

current value of the input. The nonlinearity must then be a function of its
argument: φ = φ(e(t)). Example: saturation.

The output of a nonlineariry with memory depends on the current and past
values of the inputs. The nonlinearity is not a function. Sometimes the
terminology “multiple-valued function” is used. Example: hysteresis.

A soft nonlinearity is continuous. The terminology is vague, sometimes soft

nonlinearities are also required to be differentiable. Example: φ(e) = e3 .

A hard nonlinearity has discontinuities. In some cases, a continuous function

which is not differentiable everywhere is called a hard nonlinearity. Example:
saturation.

4 / 31
Describing Function Concept

When a sinuosoidal function is applied to a nonlinearity, harmonic distortion

occurs and the output is no longer a sinusoid. Assuming the output remains
periodic (always true for static nonlinearities, still true for some others), there
is a Fourier (trigonometric) decomposition.

The decomposition contains sinuosoidal components with frequencies which

are integer multiples of the input frequency, and possibly a bias term (zero for
odd functions). In particular, the fundamental component will have an
amplitude which depends nonlinearly on the input amplitude.

The principle behind describing functions is to retain only this fundamental

output component as an approximation. A “transfer function” is formed by
taking the ratio of the output component (expressed as a complex
exponential) to the input sinusoid (also expressed as a complex exponential).

This defines the sinusoidal input describing function (SIDF).

5 / 31
Calculating a simple SIDF

As a first example, consider the cubic nonlinearity: φ(u) = u3 . Let the input
be u(t) = A sin(wt) and express this as the complex exponential
U (A, w) = Aejwt = A cos(wt) + Aj sin(wt).

The output of the nonlinearity is φ(u) = A3 sin3 (wt). We must find the
fundamental component of the Fourier series expansion of this function. Recall
that the expansion has the form:
∞
a0 X
φ(t) = + an cos( 2πnt
T ) + b n sin( 2πnt
T ), where
2
n=1

x0 +T
2
Z
an = φ(t) · cos( 2πnt
T ) dt
T x0
Z x0 +T
2
bn = φ(t) · sin( 2πnt
T ) dt
T x0

and a0 /2 is the mean of φ(t) in one period.

6 / 31
Calculating a simple SIDF...

Because φ is odd, an = 0 for all n. We only need to calculate b1 . With

T = 2π/w, the result is
3 3
b1 = A
4
3A3 jwt
The output sinusoid is then Φ = 4 e Therefore the describing function for
the cubic nonlinearity is

3A2 0
N (A, w) = N (A) = e
4
All odd static nonlinearities will have a purely real and frequency-independent
SIDF.

7 / 31
Example: Frequency-dependent SIDF

Following the introductory example in S&L on application of the DF, consider

the nonlinearity φ(x, ẋ) = −x2 ẋ associated with the van der Pol equation.

Assume X = Aejwt . Then the output of the nonlinearity is

φ(x, ẋ) = −A3 w sin2 (wt) cos(wt). Using trigonometric identities this is the
same as
A3 w
φ(x, ẋ) = − (cos(wt) − cos(3wt))
4
By the definition of SIDF, we retain only the fundamental component.
Therefore
A3 w j(wt−π/2)
Φ= e
4
A2 w −jπ/2 2
The SIDF is then N (A, w) = Φ/U = 4 e = − A4w j.

In S&L the input is taken as (−x), so their N (A, w) has the opposite sign

8 / 31
Example: Memory Relay with Deadzone

In class, we obtain the SIDF for a relay having the characteristic shown below:

out

in
−δ δ

−D

q
4D
Show that N (A, w) = πA 4Dδ
1 − ( Aδ )2 − j πA 2 . Because the nonlinearity is not

dynamic, frequency is absent.

9 / 31
Limit Cycle Analysis with Describing Functions

If N (A, w) were a truly linear transfer function, we could investigate the

presence of oscillations by using the idea of marginal stability and the Nyquist
criterion. For example, linear analysis can provide the value of K at which the
following system oscillates:
Nyquist Diagram Root Locus
1 15

0.8
10
0.6

)
-1
0.4
5

Imaginary Axis (seconds

Imaginary Axis
0.2

0 0

-0.2
-5
-0.4
0 16
k (s+2)(s2 +2s+8) -0.6
-10
−
-0.8

-1 -15
-1 0 1 2 -20 -10 0 10
Real Axis Real Axis (seconds -1
)

10 / 31
Review of the Nyquist Stability Criterion

Complex Mapping Theorem: Let F (s) : C 7→ C be a mapping defined on a

region of the complex plane and let C be a closed curve contained in that
region. Assume:
1. F (s) has no more zeroes than poles and C does not pass through any of
them.
2. F (∞) is finite.
3. C encloses Z zeroes and P poles of F (s)
Suppose C is parameterized with s = s(t) for t = [0, T ] so that the curve is
traversed clockwise. Then the image curve F (s) has N = Z − P clockwise
encirclements of the origin.

(Animation)

11 / 31
Review of the Nyquist Stability Criterion...

Nyquist’s idea was to take C as an infinite half-circle encompassing the entire

Im axis and the right half of the complex plane (where unstable poles lie).

Since F (∞) is a constant, only the Im axis portion determines how/if F (s)
encircles the origin.

Also, the Complex Mapping Theorem can be “shifted”: if 1 + F (s) has Z − P

clockwise encirclements of the origin, then F (s) has the same number of
encirclements of (−1, 0).

12 / 31
Review of the Nyquist Stability Criterion...

In a unity negative feedback control system, the open-loop transfer function is

L(s) = G(s)K(s) and the closed-loop transfer function is

L(s)
T (s) =
1 + L(s)

The closed-loop system is stable whenever L(s) doesn’t have poles inside the
Nyquist contour.

Key idea: when L(s) is a rational function (L(s) = n(s)

d(s) with n(s) and d(s)
polynomials in s), the characteristic equation for T (s) has the form:

n(s)
L(s) = = −1
d(s)

*Note*: The closed-loop poles are the zeroes of 1 + L(s) and the open-loop
poles (poles of L(s)) are also poles of 1 + L(s).

13 / 31
Review of the Nyquist Stability Criterion...

Apply the “shifted” Complex Mapping Theorem to 1 + L(s):

■ N is the number of clockwise origin encirclements of 1 + L(s), or
equivalently (we actually use this form) the number of clockwise
encirclements of (-1,0). Note we evaluate L(s) on the Im axis only (this is
called a Nyquist plot)
■ Z is the number of zeroes of 1 + L(s) inside the Nyquist contour (unstable
closed-loop poles)
■ P is the number of poles of 1 + L(s) (equivalently of L(s) and we use
this) inside the Nyquist contour (unstable open-loop poles)

Then Z = N + P

14 / 31
Example

We use the Nyquist criterion to predict the stability of a unity feedback control
system with K(s) = k and G(s) = (s−2)(s28+2s+8) . Can the system be
stabilized with proportional control?
Nyquist Diagram
0.25

0.2

0.15

0.1
Imaginary Axis
0.05

-0.05

-0.1

-0.15

-0.2

-0.25
-1 -0.8 -0.6 -0.4 -0.2 0
Real Axis

Confirm analysis using the Routh-Hurwitz criterion.

15 / 31
Practical Significance of the Nyquist Stability Criterion

Many systems are open-loop stable, so P = 0 is known even without a math

model. We can use the Nyquist criterion to predict the critical proportional
gain for closed-loop stability without having to model the system.

Since P = 0, the system can be operate safely in open-loop. We can use a

dynamic signal analyzer or offline signal processing to obtain Nyquist plot
data.

Then the stability of the closed-loop system can be studied using data from
the open-loop system only. Specifically, we can find the critical proportional
gain that scales the Nyquist plot to touch the (-1,0) point.

Also, since pure phase lags (such as those introduced by delays) rotate the
Nyquist plot, we can determine the maximum amount of delay that the system
can tolerate before going unstable.

16 / 31
Nyquist-like analysis with Describing Functions

Several conditions have to be satisfied for a DF-based frequency analysis to

produce correct results (see S&L Sect. 5.1.3):
1. There’s a single nonlinearity in the loop (unless 2 or more can be blended
into one)
2. The nonlinearity is time-invariant (a relay with a time-dependent deadzone
wouldn’t)
3. Higher harmonics produced by the nonlinearity are attenuated by the linear
dynamics following the nonlinearity (the linear TF must be low-pass).
4. The nonlinearity is odd, so that the Fourier expansion of its output has
zero bias.
Notes:
■ Even when these conditions hold, the DF technique may produce
erroneous results, particularly when predicting limit cycles.
■ There are extended DF techniques relaxing assumptions 1,2 and 4, but
they are cumbersome.
17 / 31
Example: van der Pol oscillator

See S&L Sect. 5.1.1. The van der Pol equation is re-arranged as a negative
unity feedback loop, where the nonlinear block is x2 ẋ, followed by the unstable
transfer function
µ
G(s) = 2
s − µs + 1
2
We had determined that the SIDF of the nonlinear block is N (A, w) = j A4w .
We can solve the characteristic equation for A and w:

1 + N (A, w)G(jw) = 0

The solution is A = 2 and w = 1. Remarkably, the solution is independent of

µ.

The characteristic equation solution corresponds to a Nyquist plot of

N (2, w)G(jw) crossing through (−1, 0) at a frequency of w = 1, as we can
verify by obtaining a computer plot.

18 / 31
Observations

We have solved the characteristic equation directly, but in the general case it
won’t be easy to find solutions for both A and w. A more exciting graphical
method will be used.

The existence of the limit cycle was correctly predicted (by the existence of
solutions for A and w), but the accuracy in predicting w is reduced as µ is
increased.

Likewise, the limit cycle predicted by the DF is sinusoidal (circular phase

orbit), while the shape of the actual cycle is not circular unless µ = 0.

19 / 31
Observations...

The stability of the limit cycle can also be studied using the DF method. In
S&L, the poles of the closed-loop characteristic equation are shown to be
r
1 2 1 2 2
− µ(A − 4) ± µ (A − 4)2 − 1
8 64
If A > 2, the poles are stable, indicating L.C. stability (approaching the L.C.).
If A < 2, the poles are unstable, indicating that trajectories grow towards the
L.C. This indicates the L.C. is stable.

Alternatively, from the Nyquist diagram, A < 2 causes the plot to enclose
(−1, 0) (trajectories grow towards A = 2), while A > 2 stays clear of (−1, 0)
(trajectories decrease towards A = 2).

20 / 31
Extended Nyquist Criterion

To allow non-unity loops (for instance with sensor dynamics H(s)), the
Nyquist criterion is simply applied to L(s) = K(s)G(s)H(s)

Note that S&L uses L(s) = G(s)H(s), takes K(s) = K (constant) and uses
the critical point −1/K, but K is allowed to be a complex number. But
Figure 5.22 does not include K or H anymore. Our version:
■ The characteristic equation is 1 + N (A, w)L(jw) = 0.
■ Include everything linear in L(s) (L = KGH)
■ Limit cycle detection is based on L(jw) = −1/N (A, w)
■ Follow S&L Sect. 5.4.2 by thinking of their G as our L.

21 / 31
Limit Cycle Detection with DFs

Since N (A, w) is both amplitude and frequency dependent in general, there

will be a family of Nyquist plots of −1/N (A, w), where each curve has fixed w
and is parameterized by A.

Graph L(jw) and the family of −1/N (A, w) and look for intersections where
frequencies coincide.

When N (A, w) = N (A) (frequency-independent), there is only one curve

parameterized by A, so the process is simpler.

22 / 31
Example

Consider the nonlinear 3rd-order diff.eq.:

d3 x 2 2
3
+ ẍ + 2(1 + kx ) ẋ + 3(1 + x )x = u
dt
with u = u0 (constant). There is an additional condition for a periodic
solution to be possible with u0 6= 0 (harmonic balance).
Represent the equation as a unity negative feedback loop with a cascade of
linear dynamics and nonlinearity:

u0 1 x
s3 +s2 +2s+3
−
φ(x)
2k ẋx2 + 3x3

23 / 31
Example...

We note that the nonlinearity is odd and the linear transfer function has
low-pass properties.

Let x = x0 + A sin(wt). Work with trigonometric identities and leave out

second and higher harmonics to find:

2 A2
ẋx ≈ Aw cos(wt)(x20 + )
4
A 2 3 2
3 2 2
x ≈ 3(x0 + )A sin(wt) + x0 (x0 + A )
4 2
The output of the non-bias portion of nonlinearity is then

A2
φ(x, ẋ) = 9(x20 + )A [(2k/9)w cos(wt) + sin(wt)]
4
As an operator on y = sin(wt) this is:
2
 
A dy
9(x20 + )A (2k/9) + y
4 dt
24 / 31
Example...

That is
A2
Φ(s) = 9(x20 + )A [(2k/9)s + 1]
4
In the frequency domain:

A2
N (A, w) = 9(x20 + )A [(2k/9)jw + 1]
4
Suppose that u0 = 0. A DC harmonic balance will show that x0 must be zero.
To find potential limit cycles we solve (graphically or numerically):

G(jw) = −1/N (A, w)

Taking x0 = 0 and k = 6, a solution is found: w = 2.4495 and A = 1.1006.

25 / 31
Example...
Nyquist Diagram
1

w=3
0.8

0.6

0.4

0.2
Imaginary Axis

w=0
0

-0.2

-0.4

-0.6

-0.8

-1
-1 -0.5 0 0.5 1 1.5
Real Axis

26 / 31
Example...
Nyquist Diagram
0.3

0.25

0.2
Imaginary Axis

0.15

System: sysG
Real: -0.0283
Imag: 0.0925
0.1 Frequency (rad/s): 2.45

0.05

0
-0.06 -0.05 -0.04 -0.03 -0.02 -0.01 0
Real Axis

27 / 31
Example... DC Harmonic Balance

Once the nonlinearity has been approximated with the SIDF, substitution of
the candidate periodic solution into the diff. eq. gives:

−Aw3 cos(wt) − Aw2 sin(wt) + 2Aw cos(wt) + 3(x0 + A sin(wt)) =

A2 A2 3
u0 − 2kAw(x20
+ 2
) − 9(x0 + )A sin(wt) − 3x0 (x20 + A2 )
4 4 2
In particular, the DC components must be matched:
3 2
u0 = 3x0 (1 + x20 + A )
2
This provides information about the possible combinations of u0 and x0
associated with a periodic solution of amplitude A.
Coefficient matching for sin(wt) and cos(wt) on both sides of the equation is essentially the same as the frequency
domain analysis with the Nyquist criterion. We obtain a system of nonlinear equations.

28 / 31
Example: Limit Cycle Stability and Routh-Hurwitz
Analysis
The nonlinearity was replaced by the amplitude-dependent linear operator
N (A, s). A “characteristic equation” can be written for the closed-loop system:
   
3 2 2 1 2 2 1 2
s + s + 2 1 + k(x0 + A ) s + 3 1 + 3(x0 + A ) = 0
4 4

The classical Routh-Hurwitz criterion can provide an equation for the onset of
oscillations.

Using the critical stability condition and the DC harmonic balance leads to the
system of nonlinear equations:
1
(2k − 9)(x20 + A2 ) = 1
4
3
3x0 (1 + x20 + A2 ) = u0
2
From the first equation, it’s clear that k > 4.5 is necessary for a limit cycle to
exist. The R-H criterion can also indicate the stability of any limit cycles.

29 / 31
DF Analysis of Forced Sinusoidal Input Response -
Jump Resonance
Consider the nonlinear system with sinusoidal input:

ẋ = f (x) + u(t)
u(t) = U ejwt

Use SIDF as necessary to represent f as

f = N (X, w)x

with x = Xejwt and Xi = Ai ejφi , also U = [u1 u2 ...un ]T .

Then, for each frequency w, the following must hold:

X = [jwI − N (X, w)]−1 U

This is a system of 2n nonlinear equations with 2n unknowns that can be

solved for the Ai and φi . When more than one solution exists for each
frequency, the system exhibits jump resonance.

30 / 31
Example

Consider the mass-spring-damper system with nonlinear stiffness:

ẍ + 2ζ ẋ + x(1 + µx2 ) = u

Let the input be u(t) = U sin(wt). In class, we show that the output-input
amplitude ratio satisfies
X 1
=q
U (1 + 3µ 2
− w2 ) + (2ζw)2
4 X

Up to 3 solutions are found for X in a frequency range.

Amplitude ratio
B
A

frequency

31 / 31