Proceedings of the American Control Conference
Philadelphia, Pennsylvania June 1998
Harmonic Analysis of Nonlinear and Uncertain Systems
A. Rantzer
Dept. of Automatic Control
Box 118
S-22100 Lund
SWEDEN
e m d : rantzerQcontrol.1th.se
Abstract
The theory of integral quadratic constraints is used
for performance analysis of nonlinear and uncertain
systems. Periodic oscillations induced by sinusodal
excitation are studied in particular.
1. Introduction
Stability criteria based on Lyapunov functions, dissi-
pativity and absolute stability have been developed
over several decades. However, a new perspective on
the theory has recently emerged with the develop-
ment of new numerical methods. For linear time-
invariant systems with uncertainty, efficient com-
putational tools have been developed based on the
notion structured singular value, [l, 31. For non-
linear and time-varying systems, the search for a
quadratic Lyapunov function can be written as a con-
vex optimization problem with linear matrix inequal-
ity (LMI) constraints. Such problems can be solved
with great efficiency using interior point methods.
A large variety of results of this kind were recently
unified and generalized using the notion integral
quadratic constraint (IQC) [Z].In particular, the
general computational problem to find multipliers
that prove stability was stated as a linear matrix
inequality.
This paper is devoted t o analysis of robust per-
formance in the presence of inputs described by
quadratic constraints. In particular, we consider si-
nusodal excitation of a system involving saturation.
The analysis is performed in three steps. Firstly, it is
proved for a range of frequencies of excitation, that
the output of the system converges to a peridic sig-
nal, with a period length related t o the period of
excitation. Secondly, uniqueness of the limiting or-
bit is proved. Finally, the magnitudes of the output
harmonics are estimated. In all three steps, there
are different restrictions on what integral quadratic
constraints that can be used.
$10.000 1998 AACC
0-7803-4530-4198
A. Megretski
35-418 EECS MIT
Cambridge MA 02139
USA
e m d : amegQmit .edu
Notation
The notation Lge is used for the linear space of all
functions f : ( 0 , ~+ ) Rn which are square inte-
grable on any finite interval. The subspace consist-
ing of square integrable functions is denoted L i and
the corresponding norm is denoted Ilfll.
The set of proper rational transfer matrices G =
G ( s ) of size k by m is denoted by RLk,"". This is
a subspace of CL?", the set of all matrix functions
that are bounded and continuous on the imaginary
axis. The subset of stable functions G E RLkX" is
denoted by RHk,"". Each element G E RL,"" tY is
associated with a corresponding causal LTI operator
G : LE + LE, defined by
( G f ) ( t )= D f ( t ) + f 0
C e Z A B f (-
t z)dz
where G ( s ) = C ( s l - A ) - l B + D.An element G E
RLk,"" is called strictly proper if D = 0.
The word "operator" will be used to denote an in-
put/output system. Mathematically, it simply means
any function (possibly multi-valued) from one signal
space LEe into another: an operator A : Lie -+ L2",
is defined by a subset C La, x LE such that for
every U E L& there exists w E L g with ( u , ~ E) SA.
The notation
w = A(u) (1)
means that ( U , w ) E SA.
An operator A is said t o be causal if the set of
past projections PTW of possible outputs w = A(u)
corresponding t o a particular input U does not depend
on the future U -PTUof the input, i.e. PTA = PTAPT
for all T 2 0. The operator A is bounded if there
exists C such that
l l P ~ W l l5 CllP~ull b'T > O,W = A(u), U E La,
The gain 11A11 of A is then defined as the infimium of
all C for which the inequality holds.
3654
 2. Stability via Integral Quadratic
Constraints
The object for stability analysis is interconnections
of the form
U = G ( w )+ f
(2)
w = A(u) + e
We say that the interconnection of the two operators
G : L g + Lie and A : Lie -+ LE is well posed if the
set of all solutions to (2) defines a causal operator
(G,A] : ( f , e ) H ( u , w ) . The interconnection is called
stable if in addition [G,A] is bounded.
For ll E CL2mX'+mdefine an to be the quadratic
form
Assume that a feedblack system has been proved to
be stable and that a number of integral quadratic
constraints are available:
Some of these constraints may describe nonlinear or
uncertain components, while others represent prop-
erties of external signals. Using this information, we
would like to make a conclusion about the system
performance, represented by another quadratic in-
equality.
This is the standard setup for the so called S-
procedure, observing that ( 5 ) follows fiom (4) if there
exists numbers TI,.. . ,z n 2 0 such that

ao(h) + ~ k a k ( h5
) 0 v h E L:e[O, (6)
k
The operator A : Lie + LE is said to satisfy the
integral quadratic constraint (IQC) defined by ll if This simple idea will. be used several times below.

on(h)2 0 V h = ( u , A ( u ) )E LFm Pulse input

The following proposition was proved in [2]. Let us first estimate Lp-gain when the input is a
pulse of the form
THEOREM1
Let G E RHgm and let A be a bounded causal 1 t E [O,T]
u(t)= (7)
operator. Assume that 0 otherwise
(i) for every z E [0,1],
the interconnection of G and where T 1.TO.Such signals have norm
z A is well-posed.
(ii) for every z E [0,1],
the IQC defined by ll is
satisfied by z A.
(iii) there exists E > 0 such that while the energy in the low frequency interval IwI <
W O is estimated by

[ ] * n(jw) [ ] 5 -EI VW E R
(3)
5 J' T2dw = 2T2wo I:2TTow0
Then the interconnection of G and A is stable. 0

An alternative to the condition on(0,w ) +~]lwll'1.0 The result can be restated as follows.
is to assume existence of some GO E RH,""' such
that the inequlaity condition on G is satisfied by GO PROPOSITION
1
and the interconnection of GO and A is stable. The Signals of the form (7) with T 1.TOsatisfy the IQC:
conditions above are recovered with GO= 0.

3. Performance Analysis
In the stability analysis, no properties of the distur- 0
bance variables other than their La-norm were used.
The objective of this section is t o show how such in- This gives, via the S-procedure, a bound on the Lp-
formation can be used to improve the bounds on the gain of a transfer function G ( s ) applied to pulses of
system variables. the form (7).
3655
 THEOREM2
Suppose G E RHEm,while A : Lk, --+ LE and [G,A]
are causal, bounded and have fading memory. As-
sume f ( t )= Re[foezoot]and ~ ( t=)R e [ C T = o ~ k e z o k t ]
with z k IWk12 < cm and all > O different. If the
inequalities involving G and A in Theorem 1 hold,
then
e[fOezoot]
P o l + JILo12 + MolKol
lwol L
K OI If01 for 1 = m = 1

Figure 1 Feedback loop with sinusodal excitation

(8)
lWkl2 5

COROLLARY 1 Moreover, if the inequalities hold with, l7 replaced

Suppose that G E RH, and U satisfies (7). Then, by Hj for j = 1,... ,n, let Ki,Li,ML be defined
for every WO > 0 accordingly. Then, the bound l w k l < ylfol holds for
every y that together with some 21 > 0,. .. ,Zm > 0
satisfies inequality
< Too0 m m +
Il~1l2 - n I4<wo

Proof: Let if K = 0 and

Oo(U) = llG4I2 - r”ll.llZ 0 0 K{ Li 0

z= rz -ma&)>, /GI2 -
-,<,,” /GI2- rz
TO@OlZ 1 - ToWo/Z
(11)
Then 0 2 OO(U) + ?o~(u)
and the result follows by
the S-procedure. 13 if k 2 1. 0

Proof: We prove the last statement first. From

Sinusodal input condition an,( U , A ( u ) ) 2 0 and the fading memory
To begin with, a steady state situation is considered Of A fol~owsthat
and it is assumed that all signals are periodic. Note
]:[
00
however, that this does not follow from boundedness
o s x [ : ] * n ~ ( i m k )
of the corresponding closed-loop operator. k=O
In a periodic steady-state situation, the following
result can be applied. Consider the system for all w = A ( u ) of the form ~ ( t=) Re[C;P,o~kezo’t]
and w ( t ) = R e [ x r = o ~ k e z o hwith
t] 1Uk12 < 00, xk
Gii Giz C k I W k I 2 < 00. Using the corresponding condition
on G and the equations IJO = G(ioo)wo fo, Uk = +
G(iWk)Wk O n e gets
w = A(u)

See Figure 1.For notational simplicity, assume that 0

Gzl is invertible and introduce for k 2 0

= [ [ ~ ] (?)/
*
0
L<
Mi
0 :I [El
Kj
3656
Hence, after multiplying (10) and (11) from right
and left by (y0,fo) and (yo,fo,yk) respectively, all
terms with j 4 (0,K) are non-negative and can
be removed because of the assumption on G. What
remains are the desired inequalities lwkl < ylfol.
Finally, (8) and (9) are obtained by eliminating the
parameters z1 and y analytically. 0
4 , 4,

COROLLARY 2
Let G(iw) = C(iw1 -A)-'B and let @ : R -+ R with
0 5 @(.)U 5 cu2 for all U. Suppose that A is Hurwitz,
ReG(iw) > l / c for all w and the system
i = h - B [ @ ( C x ) +aosinwot] y = Cx
has a solution of the from y ( t ) = cr=o
b k sin(wkt + Figure 2 Sinusodid excitation of system with satura-
O h ) . Then tion.

PROPOSITION
2
If @ is nondecreasing and odd with 0 5 @(u)u 5 cu2,
then
U 00

Proofi Theorem 2 can be applied with f = a0 sinwot, 0 5 Re L m ( G - i 3 ~ ) * ( 2 + - y^)dw (13)

A ( u ) = -@(U) and
A

for all y = U - w/c, w = @(U), = e-ioTG and

n(io) = [ O
-1 -2/c
-l ]
A

UT = with U ,w, y E Lz.

WT
U

G G G G - l Proof The integral. quadratic constraint

"=[ 1 0 1 [ 0 11 =[l/'G !l] 00

0 5 [_W^'H(iw)Fdw Vw = @(U) E LO

[ (2)'
-2(Re G-l+ lGl-2/c)
]l: = [ 1 2G-'/c + -2/c
was proved by Zames and Falb [5], assuming that @
The desired bounds are obtained from (8) and (9). U is nondecreasing arid odd and

4. A Case study
To concretize the statements in the previous section,
we will analyze the system
3E. = Ax - B[sat(Cx) + sin2nt/T] x ( 0 ) = xo (12)
where Th DroDosition follows from the observation t k t
I I

102 H ( iw) = 3 - 2eiWT-- e-i2oT is of this kind and a t the

C(s1 -A)-lB = G(s) = same time equal t o (1 - ,-CUT)* (2 + e-iwT) (1- e-ioT).
s3 + 2s2 + 2s + 1
U
The system is illustrated in Figure 2 together with
plots of a simulation. Analysis is performed in three It is natural to rewrite the system (12) with e(t) =
steps: sin2nt/T and f ( t ) := CeAtxo as
1. Prove convergence t o a periodic orbit.
y=Gw+f
2. Prove uniquness of the periodic orbit.
w = sat(--u) + e
3. Estimate the shape of the periodic orbit.
Different types of integral quadratic constraints are
used in the different steps of analysis. For the first
step, we use the following proposition.
3657
 min,Re[(2 + e-ZWT)(1+ G(iw))]
T In particular, the asymptotic periodic orbits are
Figure 3 According to the plotted test quantity, periodic
inputs with period time between 0.8 and 4.5 will result
in system oscillations of the same periad
Proposition 2 with c = 1 gives an IQC for the
operator w - WT = A ( u - U T ) = sat(u) -&(UT).
Hence, application of Theorem 1 shows that U - U T
and w - WT are square integrable provided that
0 < E < Re[(2 + e-ioT)(l + G(io))] U ER
It is easy t o check that this condition is fulfilled for
0.8 5 T 5 4.5. See Figure 3. For these values of
T, we can therefore conclude that all solutions tend
towards a T-periodic behavior.
However, it is also possible to make a conclusion
about excitation of higher frequency, i.e. for T < 0.8.
For such values, since 0.8 is less than half of 4.5
there will always exists an integer N such that N T
is in the interval [0.8,4.5].The previous argument
can therefore be applied with T replaced by N T ,
to conclude that all solutions tend towards N T -
periodicity. This opens the door for so called sub-
harmonic oscillations. For inputs with T > 4.5, no
conclusion can be drawn.
The second step of analysis is to verify uniqueness,
i.e. that the limiting periodic solutions are indepen-
dent of the initial state. For this purpose we note
that the “stability condition” Re G ( i o ) > -1 holds
for U > 0.6. This turns out t o be sufficient, since
0.6 .4.5 < 2n, so the limiting periodic orbits that we
have proved to exist, all have higher frequency than
0.6.
To make this argument rigorous, consider an input
with T < 4.5 and two corresponding solutions
w1 = sat(Gw1 + f1) w2 = sat(Gwg +f2)
with f k ( t ) = CeAtxkfor h = 1and h = 2. Both w1 and
wz tend towards periodic orbits of frequency higher
than 0.6, so
05 L‘[(~-E)(WI -WZ) +G(wl -WZ)](WI- w z ) d t
3658
for sufficiently large z.In addition, the saturation
satisfies the corresponding constraint incrementally:
o 2 l‘[sat(ul) - sat(uz) - (u1- uz>l[sat(ul>- sat(uz)ldt
Hence, the argument behind Theorem 1 shows that
there exists a constant y, independent of x1 and x2
such that
I’ IUl(t) - v2(t)12dtI ylx1-.2} v,r
identical and unique.
Finally, it is interesting to simulate the system
and compare the actual output with the bounds of
Corollary 2. With W O = 3, u1 = 3010, the conic
condition sat(u)[u - sat(u)] 2 0, gives the bounds
Ibol/laol i 4.66 Ib1l/laol 5 0 . 9 5
Simulation with gives
Ibol/laol = 1.53 Ibll/laol = 0.18 for a0 = 2
Ibol/laol M 3.33 Ibll/laol M 0 for very large a0
5. Conclusions
It has been shown that beyond stability analysis,
integral quadratic constraints can be used t o analyse
the response to particular input signals, such a
unit pulse or a sinusoid. We believe that that the
approach can be extended to a large number of other
cases, particularly in combination with appropriate
computer tools for numerical optimization.
6. References
J . C. Doyle. Analysis of feedback systems with
structured uncertainties. In IEE Proceedings,
volume D-129, pages 242-251,1982.
A. Megretski and A. Rantzer. System analysis via
Integral Quadratic Constraints. IEEE Zhnsac-
tions on Automatic Control, 47(6):819-830, June
1997.
A. Packard and J.C. Doyle. The complex struc-
tured singular value. Automatica, 29( 1):71-109,
1993.
A. Rantzer and A. Megretski. System analysis via
Integral Quadratic Constraints, Part 11. Techni-
cal Report TFRT-7559, Department of Automatic
Control, Lund Institute of Technology, Septem-
ber 1997. Invited for publication in Systems and
Control Letters.
G. Zames and P.L. Falb. Stability conditions for
systems with monotone and slope-restricted non-
linearities. SIAM Journal of Control, 6( 1):89-
108, 1968.