ING FUNCTION ANALYSIS OF

NONLINEAR SYSTEMS WITH PARAMETRIC UNCERTAINTIES

S. T. Impram and N. Munro

Control Systems Centre, UMIST, England

Keywords: nonlinear systems, describing hctions, the diversity of circumstances under which the method
parameter uncertainty, robust stability. has been studied. However, almost all of the work done
to date, has been restricted to systems with known
structure. Unfortunately, in real life, theres mostly a
ABSTRACT mismatch between the actual system and its assumed
model since a precise and complete mathematical
Periodic phenomena, such as limit cycles, are among the description of a process is rarely available.
most prominent features of nonlinear control systems.
Predictions as to whether or not oscillations of this kind Within the context of uncertain systems, describing
are likely to exist can be made using the describing functions have received much less attention. The only
function approach. Here, some of the now well papers that appear in the literature are [3] and [lo].
established robustness results regarding linear systems While the latter deals with the prediction of limit cycles
are combined with the describing function method to within a p-synthesis framework, the former treats
analyze the stability of autonomous uncertain systems systems subject to structured uncertainties arising from
with separable nonlinearities. An example is provided to known parameter variations in both the linear and
substantiate the mathematical derivations and nonlinear elements. By overbounding the unknown
theoretical conclusions. coefficients of the system characteristic equation and
subsequently utilizing Kharitonovs Theorem [5], the
authors perform an overall stability test which, unless
1. INTRODUCTION the numerator of the linear subsystem is constant, gives
conservative results.
The stability analysis of autonomous nonlinear control
systems can often be converted into a problem of
investigating the existence of sustained oscillations
known as limit cycles. Apart from a limited number of
cases where such oscillations are deliberately formed, in
most control systems they are undesirable and must be
predicted. Although, a precise knowledge of the
waveform of a limit cycle is usually not mandatory, the
knowledge of its existence, as well as that of its
approximate nature is critical. Among the techniques
used for this purpose, the describing function method
stands out because of the ease with which it can be
incorporated into the frequency domain analysis tools
available for linear systems. The basic idea of this
method is to represent a separable nonlinear element by
a kind of transfer function derived from its effects on N(A,~,P) k G(jwq)
sinusoidal input signals or combinations of these. Even
though it is an approximation technique, the important
role of Nyquist plots in classical control make it an
indispensable component of the bag of tools of
practicing engineers. Figure 1:Uncertain nonlinear system.

The describing function method is covered in many

books, particularly detailed treatments being found in 2. DESCRIBING FUNCTION ANALYSIS
[ l ] and [4]. Due to its popularity, researchers are
constantly trying to extend the range of problems to To be able to develop the basic version of the describing
which it applies. Publications in which describing function method, the system has to satisfy the following
functions are employed to study oscillations in conditions:
nonlinear feedback systems exist in abundance. The athere is only one nonlinear element which is time
reader if referred to [6], [8]and [9] to get a flavour of
p and q are vectors of uncertain parameters

UKACC International Conference on CONTROL 98, 1-4 September 1998,

112 Conference Publication No. 455,O IEE, 1998
invariant and has odd symmetry, of G(jw,q) and -1/N(A,p). In the absence of any
the linear part possesses low-pass properties. intersection on the other hand, all one can conclude is
The second condition constitutes a standing assumption that there will probably not be sustained oscillations in
and has crucial effects on the reliability of any the system. Note that, because of the approximate nature
prediction made. It implies that higher harmonics in the of the describing function method, it is not usually
output will be filtered out significantly. possible to replace the word probably with
certainly.
In order to represent a nonlinearity with input x( t) and
output w(t) by a describing function, consider the
simple periodic signal x(t) = Asin(ot) with fixed A 3. POLYNOMIAIS~UNCERTAINPARAMETJCRS
and o . Because of the first condition stated above,
In the analysis of systems subject to parameter
w(t) will then be also a periodic function (of period variations, the issue of uncertainty structure frequently
2n/w), which can be expanded in a Fourier series as arises. To formalize the concept, consider a polynomial
follows: expressed as

The coefficients ak and bk are computed fiom where q is the vector of uncertain parameters. The
components of q are assumed to vary in prescribed
intervals defining an axis parallel box Q. The
uncertainty structure is manifested via the coefficient
ak = L2jw(t)cos(kot)d(ot) (2)
no functions ak(q) .

In the idealized framework of Kharitonov [ 5 ] , each

element of q enters into only one coefficient of (7). The
result is an interval polynomial family. The next level of
complication is the affine linear uncertainty structure.
For a nonlinear inputJoutput characteristic with odd Each coefficient of (7) is given by
symmetry, the constant term a. is zero. Furthermore,
due to the low-pass filtering hypothesis, only the
fundamental component of w( t) needs to be taken into
account: where bk is a column vector and Ck is a scalar. To
deal with more realistic situations, multilinear
wl(t) = al cos(ot) + bl sin(ot) (4) uncertainty structures are of paramount importance. The
coefficients of (7) are said to depend multilinearly on
The describing function can now be expressed as the the elements of q if they attain an affine linear structure
complex ratio of (4) to the input sinusoid: when all but one of the uncertain parameters is fixed. A
rather interesting uncertainty structure which is not as
frequently considered in the robust control field, gives
rise to the so called Complex Coefficient Kharitonov
Since ( 5 ) is a scalar quantity, without loss of generality, Polynomials:
any parametric uncertainty in the nonlinearity can easily
be allowed for by varying the describing function in an
interval and denoting it with N(A, p) .

With the describing function representing the frequency The parameters qk and rk lie within known intervals
response of the nonlinear component, the characteristic and respectively define axis parallel uncertainty
equation of the uncertain system becomes bounding sets Q and R.Polynomial families of this type
are usually motivated from modelling considerations.

The existence of a limit cycle is thus predicted if there is 4. NONLINEAR SYSTEMS WITH UNCERTAIN
an intersection between the loci, in the complex plane, PARAMETERS

The polynomial families summarized in the previous

In the sequel, the dependence of N(.) on o is dropped by section can be used to construct a rich class of uncertain
restricting the nonlinearity to be static.

113
rational functions, and hence, transfer functions for exactly. However, to avoid any ambiguity, it has to be
linear systems. It is the purpose of this section to study immediately said that this exactness is only valid if
the behaviour, in particular, the stability of such systems A(s, U) and B(s, v) have either an interval or an affine
connected in the configuration of Figure 1. linear uncertainty structure.
Let the parameters of the nonlinear element vary such The reason for ruling out the multilinear dependency
that the governing describing function is approximately comes from the theory of convex analysis and can be
given by best understood by referring to (12). Let

A( j o *,Q) = {A(j o * ,u) : q = [U v] E Q)

with the imaginary part vanishing in the case of no denote the value set of the corresponding family at the
memory. Consider the uncertain linear system in the
fixed frequency a*.Similarly, B( jo*,Q) is the image
form G(s,q) = A(s,u)/B(s,v) where
ofQunder B(jo*,v). Itisknownthatboth A(jo*,Q)
and B( j o *,Q) are convex shapes, the extreme points of
k=1 which are mapped from the vertices of Q [2]. The range
B(s, v) = be(v)se of [ n R , n i ] on the other hand, is a portion of the real
e=i
axis between n i and nfR. The result of multiplying
with m s n . The uncertain parameters are such that this straight line segment with A(jw*,Q) is also a
q = [U v], i.e. variations in the numerator coefficients convex set. Therefore, the image of the characteristic
do not effect the denominator polynomial, and vice equation (12), which essentially is the addition of two
versa. With a single valued nonlinearity, the convex sets, is itself convex. However, it has to be
characteristicequation of the system is stressed that this result cannot be attributed only to the
convexity of the sets involved in the addition, but rather,
to the fact that it has been obtained by the convex value
B(s, v) +[nR ,n&]A(s,u) = 0 (12)
sets of two polynomials, namely B(s,v) and
Note that, unless A(s,u) is simply a constan?, (12) [ n i ,n &]A(s, U), which have no common parameter.
belongs to the class of polynomials with multilinear
coefficients. For notational simplicity, let A(s, U) and Since the image of Q under a multilinear mapping
B(s,v) have an interval uncertainty structure with the would not necessarily be convex, the foregoing
conclusions cannot be drawn. Said another way, if
elements of U and v varying such that they remain
A(s,u) and B(s,v) obey ( l l ) , but their coefficients
positive. Equation (12) then becomes
depend multilinearly on the uncertain parameters, there

5 [b, ,b i ] s e +Cm [bk+nRak,b:+nia:]sk=

e=m+l k=l
0 (13)
is no guarantee that A(jo*,Q) and B(jo*,Q), let
alone the image of (12), will be convex.

Clearly, the coefficients (though not necessarily all) of So far the analysis has been for real describing
the characteristic equation have a multilinear functions. If instead (10) holds, it is still possible to
dependency on the uncertain parameters of the system. generate value sets without any conservatism provided
Needless to say, similar arguments hold also for rational the numerator of the linear plant is constant (assuming
functions with affine linear and multilinear coefficients of course, the denominator is interval or affine linear).
even if the describing function is complex.

A powerful tool to analyze the stability of such 5. EXAMPLE

polynomials is the Mapping Theorem [ 111. Although it
is known,in general, to provide only sufficient results, it Consider the feedback system given in Figure 1. The
has always been preferred to other overbounding nonlinear element has the saturation characteristic
methods. Part of the reason for this lies in the fact that it depicted in Figure 2. If A < a , the nonlinearity has the
gives the tightest possible bounds on the polynomial same effect as a simple linear gain k. For A 2 a , the
family. Yet, in cases where the numerator and describing function becomes
denominator of the linear plant obey (1 l), the Mapping

t)+ (:),/SI
Theorem is capable of determining these bounds
N(A, k) = %[arcsin(
7F
In the discussion to follow, unless otherwise stated, A(s, U) (14)
is assumed to involve at least one uncertain parameter.

114
Suppose that the, upper and lower output limits are oscillations in the system (Figure 5 ) .
constrained to be constant at all times, whereas the slope
k assumes*valuesin the interval [1,4]. This implies that
N(A7 k) E [0,41*

I I I I I I *
I I I I I I ,
-2 5
-2 -15 .1 0 5 0 05 1 15
Real A x i s

Figure 3: Loci of -1/N(A, k) and G(jo,q) .

Figure 2: Saturationnonlinearity.

The linear plant is such that the coefficients of the

numerator and denominator polynomials have an
interval uncertainty structure:

s2 + a s + b
(36,4 = (15)
cs4 +ds3 +es2 +fs+g

The uncertainty bounding box Q is defined by

q = [ a b c d e f g ] where

2 Ia I 8 31.4 Ie 532.6 Real Axis

97IbI103 49If I50 Figure 4: Zero Exclusion Test.

0.4 I c 51.6 59 Ig I 6 0
7.4 I d 2 8 . 6
08.

For graphical evaluation of the system behaviour, the 06 ~

frequency response of G(jw,q) has to be plotted. This 04 -

can easily be done using software packages such as [7]. 02-
m
Figure 3 shows the locus of -1/N(A, k) superimposed B 0-
onto the frequency templates of the linear subsystem. I
k02-
Since there is an intersection between the two, one can 0 4 ~

say that a limit cycle is likely to occur. This prediction

06.
can be M e r strengthened by investigating the stability
08 -
of the closed loop system. Using (6), (14) and (1 5), the
characteristic polynomial can be written as 11 " " " " ' 1
0 5 1 0 1 5 2 0 2 5 a 3 5 4 0 4 5 5 0
time (sec)

cs4 +ds3 +es2 +fs+g+N(A,k)(s2 +as+b)=O (16) Figure 5: Computer simulation results.

As a concluding remark, the possibility of the

Making use of the Mapping Theorem and performing
describing function method making erroneous
the Zero Exclusion Test [2], yields Figure 4. Since at
predictions even under circumstances for which the
some frequency the value sets include the origin, (16) stability test is not conservative, is worthwhile
can be deemed to be unstable. Note that, despite the mentioning. In other words, there may well be cases
multilinear coefficients of ( 16), the stability test is where the value sets of the characteristic polynomial are
conclusive. This of course, is due to the structure of
exact, but the behaviour of the system does not agree
G(s,q) . Indeed, the computer simulation results for with the conclusions drawn from the stability results.
fixed q* E Q and k* E [1,4] clearly illustrate the Therefore, it must always be bom in mind that the

115
derivation procedure of a particular describing function Switched Circuits, IEEE Trans. Circuits Syst., 40,
involves various approximations at several stages, and 564-572.
whether or not the predictions will turn out to be correct [lo] Tierno J. E., 1997, Describing Function Analysis
depends heavily on the problem under consideration. in the Presence of Uncertainty, J. Guidance,
Contr.. Dynamics, 2,956-961.
[ll] Zadeh L. A. and Desoer C. A., 1963, Linear
6. CONCLUSION System Theory - A State Space Approach,
McGraw-Hill, New York, USA.
Predictions of limit cycles in uncertain control systems
with separable nonlinearities can be made by using the
describing fimction method together with some of the
robustness tools for linear systems. The overall picture
usually turns out to be the stability analysis of a
polynomial with uncertain coefficients.

Depending upon the inputloutput characteristic of the

nonlinear element, stability results for particular classes
of uncertain linear systems can be obtained without
conservatism. However, as the complexity of the
uncertainty structure increases, the results may be
inconclusive.

Moreover, if the linear part is not sufficiently7low-

pass, the approximate nature of the describing function
analysis is very likely to lead to an incorrect assessment
of the closed loop behaviour.

REFERENCES

[l] Atherton D. P., 1965, Nonlinear Control

Engineering, Van Nostrand Reinhold, London,
UK.
[2] Barmish B. R., 1994, New Tools for Robustness
of Linear Systems, Macmillan, New York, USA.
[3] Fadali M. S. and Chachavalvoong N., 1995,
Describing Function Analysis of Uncertain
Nonlinear Systems Using the Kharitonov
Approach, Proc. Amer. Contr. CodL,2908-2912.
[4] Gelb A. and Vander Velde W. E., 1968, Multiple
Input Describing Functions and Nonlinear System
Design, McGraw-Hill, New York, USA.
[ 5 ] Kharitonov V. L., 1979, Asymptotic Stability of
an Equilibrium Position of a Family of Systems of
Linear Differential Equations, Differential
Equations, 14,26-35.
[6] Klyde D. H., McRuer D. T. and Myers T. T., 1997,
Pilot-Induced Oscillation Analysis and Prediction
with Actuator Rate Limiting, J. Guidance. Contr.,
Dynamics, 3,81-89.
[7] Kontogiannis E. and Munro N., 1997, The
Parametric Systems Toolbox, Proc. 7th IFAC
S v m ~ .ComD. Aided Contr. Syst. Design, Gent,
Belgium.
[8] Leu M. C., Yang S. and Meyer A. U., 1993, Anti-
Windup Control of Second Order Plants with
Saturation Nonlinearity, ASME J. Dynamic Svst,
Measurement. Contr,, U , 7 15-720.
[9] Sanders S. R., 1993, On Limit Cycle and the
Describing Function Method in Periodically

116