You are on page 1of 4

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, AUGUST 1968 41 3

An Off-Axis Circle Criterion for the Stability of


Feedback Systems with a Monotonic Nonlinearity
YO-SUNG CHO AND KUMPATI S. NARENDRA,
SENIOR MEMBER, IEEE Fig. 1. Nonlinear feedback system.

Abstract-An off-axis circle criterion is stated in this paper as a


sufficient condition for the stability of a nonlinear feedback system FGe(t) e(t) = x(t) + (1)
with a single monotonic nonlinearity. The criterion provides a geo- Ge(t) = y ( l ) .
metric
method for determining sector
a (KI<[F&)-F(~~)]/
CyI-y2)<K2)within which the nonlinearity may lie for asymptotic Here G is assumed to be a linear time-invariant operator and F ( . )
a nonlinear operator. The operators are assumed to satisfy the fol-
stability. The relation between the criterion stated and the well-
known Popov criterion and the circle criterion for time-varging sys- lowing conditions.
tems is also discussed. Two examples are included to clarify various G: The operator G is linear and time invariant and is defined by
aspects of the criterion and its relation to existing results.
Ge(t) = ' g ( f - r)e(T)dT -k gl(t)
I. IXTRODUCTION 0

Ever since Popov [l ] stated the conditions for the stability of a where the impulse response function g(t) is in and the LnL,
nonlinear feedback system in the frequency domain, there has been function gl(t), representing the effect of initial conditions, is
considerable interest in deriving similar frequency domain conditions in L.The Laplace transform of g ( t ) is
for systems with other types of nonlinearities. At the present time,
frequency conditions for determining the stability of systems with G(s) = -g(t)e-W for Re [s] > 0.
monotonic and odd monotonic nonlinearities are well established- 0

see Zames [ 2 ] , [3], Brockett andWillems 141, Narendra andNeuman F: The operator F ( . ) is nonlinear and time invariant and satisfies
[5]. These conditions are expressed in terms of the transfer function the following conditions.
G(s) of the linear part of the system and a multiplier Z ( s ) having a
1) F b ) =0, if y = O .
specific form so that G(s)Z(s) is positive real. However, it is seldom
2) KI~[F~~)-F(y2)]/(yl-y~)~K~forallrealnumbersyl
apparent how such a multiplier can be found. In this paper a geo-
metric condition is derived for the existence of an RLor an RC mul- and y2 such that yl f y 2 , and K1 and K2 are real constants satis-
tiplier for the stability of feedback systems with monotonic non- fying the condition - m < K 1<K2 < m . Note that if R 1 2 0 ,
linearities. I t is shown that if any circle can be drawn to intersect the then F ( . ) is a monotonically increasing function.
negative real axis at - l/KI and - 1/Kg on the Nyquist plane so that Assuming that the solution to (1) exists and belongs to Le, the
the Nyquist plot of G(i') lies outside the circle and does not com- problem of interest may now be stated as follows. If the input to the
pletely encircle it, then the nonlinear feedback system is asympto- feedback system x ( t ) belongs to b,determine the conditions on the
tically stable for all real valuedfunctions F ( . ) satisfying K1< linear part G of the system so that
[ F ( r l )-FO.~)]/(r1-y2) <K2, where K1 and K Pare positive constants.
The center of such a circle need not lie on the real axis. 1 ) YO)E Lz
2) lim y(t) I I 40.
11. DEFIVITIOSS "I 00

The general formulation of Zames [Z], [3] will be adopted here. I I I. PRELIMINXRP
RESCLTS
LZe which is an extension of the L
z space of square integrable func-
tions is defined as follows. Let RL be theclass of complex valued functions Z(s) having the
form
L2: A function n(t) defined in the interval t E [0, m ] belongs to s+a,
L if j 0 * x 2 ( t ) d t < 3c. The norm of r(t) in b is denoted by Z(s) =
n=o
-
s + an
llxll?= {j-oU"x2(t)dt] 1'2.
La:A function x ( t ) belongs to Laif every finite time truncation of where s is a complex variable and O<cu0<,f30<al<,f31 . * , and let
x ( t ) belongs to LP. If x , @ ) denotes the truncated function, then RC be the class of functions Z ( s ) for which Z-l(s) is in RL.
x,(t) =r(t)for t < 7 and r E T , where T is the subinterval of the Theorem 1 and Lemma 1 stated in this section are used for deriv-
reals [0, m ), and rr(t)= O elsewhere. An extended norm denoted ing the circle criterion. Theorem 1 is a minor extension of a multiplier
by ilxll?cis assigned to each r(t) in Le as follows. theorem originally presented in Zames [Z], and later proved using a
Liapunov approach in Brockett and Rillems [4] and Narendra and
Neuman [SI. While all the methods essentially lead to thesame con-
ditions for proving the stability of systemswithmonotonic non-
linearities, Brockett and Willems [4]and Narendra and Neuman [j]
The feedback system whose stability is considered in this paper has use a differential equation framework which is distinctly different
~- from the approach used in Zames [2]. The approach used in this
the form shown in Fig. 1. If the operators G and F map 6 , into
itself, then the feedback equations for the system for t>_O may be paper is similar to that used in Zames [2], [3].
formulated as Lemma 1 provides the specific form of the multiplier Z(s) in RL
or RC that is needed for the proof of the geometric criterion.

ManuscriptreceivedDecember 2. 1966;revised May 1. 1967. November 28. Theorem 1 (Skeleton TheoTem)


1967. andApril 18,. 1968.
Y.S. Cho is wlth Honeywell Inc., Waltham. Mass. In the feedback system of Fig. 1, let G and F satisfy the condi-
K. S. Narendra is with the Dept. of Engineering and AppliedScience, Yale Uni-
versity, New Haven. Conn. tions stated in Section 11. Then if there exists a multiplier Z(s) in RL
414 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, AUGUST 1968

or RC with the property

the system is stable in the following sense:


where /I 1, and 11 111 denote the n o m s in L, and L1, respectively.
1) x(t) E L?implies y(t) E & Since
2)
1- m
I A4 I -0.
Lentme 1 (Construction of 1Multifiliers) and
Let [a, b ] beanysubinterval of (0, 00). e anyconstant in
(-7r/2, 7r/2), and e > O a constant.
1) If 020, then there exists a function Z ( L ) in RL such that it follows that
I arg { Z(L)1 - e I < e, for all w in [a, b].
2) If B<O, then there exists a function Z(&) in RC with the
property that
+ 2c(l - f) logp.
Now a and p can be chosen small enough and X large enough so that
I arg ( z ( ~ w-) )e I < e, for all o in [a, 61. forany w in [a,b ] , tan-' (w/wO)--w/2<e/3, tan-' (w/wSH)<E/3,
A heuristic explanation of the method of proof may be outlined as fol- and 2c(l -c) log p<e/3. Thereupon, the inequality larg { Z ( & ) } -el
lows. The multiplier Z(&) will be defined by a function having alter- < e is satisfied, and consequently Lemma 1 is true.

nating poles and zeros on the negative real axis. I t will then beshown VI. MAIX RESULTS
that arg { Z ( i u ) } can be expressed as a convolution of a "step func-
tion" and a "smoothing kernel," and this fact will be used to obtain The following three theorems are direct consequences of Theorem
a bound on larg { Z & ) } -01 using Holders inequality. 1 and Lemma 1 and constitute the main results of this paper.
Proof of Lemma 1: Let c = 0 / ( ~ / 2 ) . Let a > O and p > 1 be constants Theorem 2
and N a positive integer with the property a <a and ap"+c>b. Let
the multiplier Z(&) be defined by Consider the feedback equations (1) (see Fig. 1). If 1) the Nyquist
plot of the linear part of the system G ( i w ) lies entirely to theright of
a straight line passing through the point (- (1 /&)+ 6, 0) where 6 > 0
and if 2) 0 5 [FF(yl)--F(y?)]/(3"-y?)~K2for all y1fy2, then the sys-
tem is as)-mptotically stable, Le.,lim \ y ( t ) \ =O.
Then t.00

Theorem 3
If 1) the Yyquistplot of G(iw) lies outside acircle passing through
Let be the variable defined by Q=log w ; let Q,= log wn, where
+
(- (1 /K?) 6', 0) and (- (1/KI) - 61, 0) where 61 > O without com-
pletely encircling it andif 2) K11: [F(yl)-FF(yn)]/(3'1-y~)IKnfor all
w n = a p n for n = 1 , 2, . . , N ; le:f(Q) be the function of st defined by yl #yp, then the system is asymptotically stable.
f ( ~ )= tan-' [exp ( G I ] . Theorem 4
1
Then arg { Z ( L ) can be written in the form If 1) the Nyquist plot of G(&) lies entirely inside a circle passing
B through (- (1/&) +
61, 0) and ((1/K1-)61, 0) where 61>0 and if 2)
arg ( ~ ( & ) j= n = O ( J ( Q - 4)- f ( ~- Q,)}. -K1I[FF(yl)-FF(yn)]/(3'1--y2)IK~for all yl#fyp, the systemis
asymptotically stable.
Let u ( Q ) and g(Q) be functions of n defined by Only Theorem 2 is proved here since i t can be shown that Theo-
rems 3 and 4 can be derived using simpletransformations from
I, for Q in [Q,, ],n
, Theorem 2 [ 6 ] , [SI.
0, Q elsewhere
Proof of T l e o r m t 2 (Fig. 2):
and
1) Case A:

Let the transfer function be defined as GI(&) =G(iw)+(l/Ks)


and let the locus of G(&) lie to the right of a straight line passing
+
through (- (1/K?) 6 , O ) and making an angle 8 with the imaginary
axis. The argumentof GI(&) lies in the range
7i 7i
-- - 0
2
< arg (Cl(iw)}< 5 -e.
Hence positive constants el and €2 exist such that

where the left and right equalities hold for frequencies w=u and
w = b. Let E be a positive constant such that E <SI and Q. By Lemma 1
Hence the following bound is obtained on arg { Z(iu)} -e. there exists a multiplier Z ( L ) in RL such that
SHORT PAPERS 415

Fig. 2. Proof of Theorem 2.

Modified Nyquist Plot (-1 and Nyqist Plot f+ of

a&?-rma
Fig. 4. Example 1.

(C)

Fig. 3. Admissible regions for F ( .) for given G ( i w ) .


(a) Theorem 2. (b) Theorem 3. ( c ) Theorem 4.
*ReG(.) -&

( a r g ( Z ( k ) ]- e l <e, fnra<w<b.

Hence the argument of Z(iw)G~(iw)


lies in the range
Nyquist plot of G ( s ) = ~
$(&s+25)
- - + e:
7r
-E < arg {Z(iw)Gl(iw)) < -2 - + 7r
ep E
Fig. 5. Example 2.
2
or

1 arg (Z(io)G~(io)}I < -


2
- 9.
7i
lows. If the lower limit of the monotonic gain KI is specified, the
upper limit IC? of the gain defined by K1< F(y)/y<Kz within which
Hence Re Gl(io)Z(iw>_6> O and by Theorem 1 the system is asymp- the nonlinear feedback system isasymptotically stablemaybe
totically stable. obtained by drawing the circle, which passing through (- l / K l , 0)
2) Case B: O > O > --7i/2. touches G(iw) at t w o points. The second intersection of the circle
Using similar arguments as in case A one can construct, again by with the negative real axis yields - ~ / K z .
Theorem 1 , a multiplier Z(iwj in RC such that 2) The criterion given by Theorem 3 is a sufficient condition and
by no means necessary for the stability of the nonlinear feedback
I arg ( Z ( i w ) G ~ ( i w ) )< 2
- Q, system. Since the theorem is proved for the most stringent case, the
actual stability range for a specific G(s) may be K1< F(y)/y<x2,
Since Re { Gl(iw)Z(iwj} >6>0, the system is asymptotically stable. where R2<K2. I n such cases a multiplier ZRLmay exist so that
iVhile a proof for Theorem 3 is not given here the approach may G(s)ZRI,(S)is positive real even when G(iw) does not lie to the right
be outlined as follows. If the Syquist plot of G ( L ) lies outside a of a straight line, or outside a circle.
circle passing through (- ~ / K I0 ,) and (- 1/ K z ,0 ) (Fig. 3(c)), it can 3) If the lower limit K 1= 0 , for a given linear system G(iw),gains
be shown [6], [SI using simple transformationsthatthe Nyquist K.v, K M , K pand
, K c represent the gains for which the feedback sys-
plot of K I / [ ~ + K I G ( L lies
) ] entirely to the right of a straight line tem is asymptotically stable with 1) constant gains, Kx=Nyquist
through (-KlK.z/(K?- K I ) ,0 ) . Hence by Theorem 2 the feedback gain: 2) monotonic gains KaIfobtained by the criterion stated in this
system can be shown to be asymptotically stable for all monotonic paper; 3) first and third quadrant nonlinearities, K p = P o p o ~gain,
nonlinear functions F ( . j in the sector and 4) nonlinear time-varying gain K c obtained by the circle cri-
terion [6]-[8].
The methodsfor obtaining these valuesin practical situations are
illustrated in Figs. 4 and 5 for Examples 1 and 2. I t is wellknown that
Kc<KP<KN. In most situations thegain K M is found t o lie between
v. S O M E C O M M E N T S O S T H E C I R C L E CRITERION
the Kyquist gain and the Popov gain, Le., K p < K x < K s .
Some comments can now be made regarding the applicability of 4) The circle criterion for time-varying systems calls for a circle
the criterion in practical situations as well as its relation to well- with its center on the negative real axis while the circle criterion for
h o w n existing criteria. monotonic nonlinearities calls for any circle touching G ( L ) at two
1) The circle criterion stated in Theorems 3 and 4 may be applied points. The two areidentical only in the particular case where G(iw)
in a practical situation in many ways. In particular, the maximum is symmetrical about the real axis. In general the monotonic circle
range of stability given by the criterion may be determined as fol- (C.u) criterion yields a much larger sector of stability for monotonic
416 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, AUGUST 1968

nonlinearities than
the circle (CC) criterion doestime-varying
for non- ACKNOWUZDGXEXT
linearities. The sectors for a specific example are shon-n in Fig. 5 .
The authors would like to thank thereviewer for suggestin: the
VI. EX4MPLE.S form of the multiplier used in the main lemma.
Two examples are given here to demonstrate theapplicability of REFEREXCES
the criterion derived in Section V and to compare it with Popov's V. hi. Popov. "Absolute stability of nonlinear systems of automatic control,'
criterion and the circle criterion for time-vaqing systems. Aubmalmn and Rmwfe Control. vol. 22. DD. 857-815. March 1962.
G. Zames. stability of nodin&,
the time-va&g~feedba& systems,"
Proc. A'd'l Electronics Conf.. vol. 20, pp.,725-730., Octobq 1964.
Example I -O ' n the input-output stability of tlme-vawmg nonhnear feedback svs-
tern;" pts. I and 11. I E E E T r a m . Aufomatic Control, vol. AC-11, pp. 258-
The transfer function of the linear part of the feedback system in 238. April 1966, and pp. 46M76, J u p 1966.
R. W. Brockett and J. W . Willems. Frequency domain stability criteria." pts.
Fig. 1 is given by I and 11. I E E E Tram. Aulomolrc Confrol. rol. AC-10, ~. ~p.~255-261, July1965.
and pp. 407-413, October 1965.
s+l I(.S. Karendra and C. P. Neumann. "Staplity of a class of dXerentia1 equa-
G(s) = __
tions with a s h l e monotone nonlinearitv. S I A M J. Cmdrol, vol. 4. DD, 295-
s(s + 0.1)(s2 + 0 . 5 + 9) 308. May 1966.-
K. S. Narendra and R. M. Goldwyn. 'A geometrical criterion for the stability
of certain nonlinear nonautonomous systems.> I E E E T ~ Q X CircuitS. Theory
The Nyquistplot G ( L ) as well asthe modified Nyquistplot (Correspondence). vol. CT-11, pp. 406408, September 1964.
I. W. San$berg. "On the L?-boundedness of solutions of nonlinear functional
(Re G versus w Im G) are shown in Fig. 4. The Nyquist gain for the equations, Bell Sys. Tech. 1..vol. 43, pp. 1581-1599. July 1964.
system is K.v=4.17 while the Popov gain K ~ = 0 . 6 3is obtained by -. '-4 frequency-domainconditionforthestabllity of feedbacksystems
containing a sinde time-varying nonlinear element,"BeU Srs. Tech. J . , vol. 43,
drawinga straight line tangential at two points tothe modified PP. 1601-1608. j U l y 1 9 ~ .
hi. A. Aiizerman and F. R. Gantmacher. Absolute SfabiIifyof Regulator Sysfems.
Nyquist plot. The straightline tangential toC(iw) a t two points inter- San Francisco. Calif.: Holden-Dav. 1964.
sects the real axis a t (-1/3.23, 0 ) yielding a K ~ = 3 . 2 3 .Hence the C.
C. T . k e e a n d A . Deroer, "Stability of single-loop nonlinear feedback ws-
terne. Electronic Research Laboratory,University of California. Berkeley,
system is asymptotically stable for all monotonic increasing non- Calif..Rept. ERL 66-13. May 1966.
linear functions in the sector
FbJ - F b 2 )
O< < 3.23.
Yl - y s
Ersmple 2 Optimal Control of the Linear Regulator with
Constant Disturbances
In thiscase
3(s + 1) C. D. JOHNSON
G(s) =
s2(s2 + s + 25) Abstracf-An optimal control problem for a linear regulator with
constant external disturbance is formulated. It is shown that, for a
I t is required to find a sector 1< [FbL)-F(y?)]/bl--y?) <K?within suitablyselected quadratic-type performance index, the optimal
which the system is asymptotically stable. The Kyquist plot G(&) is control is not an explicit function of the external disturbance. More-
shown in Fig. 5. over, the optimal control can be synthesized a s a time-invariant
1) The Nyquist gain Ks=8.30. Hence for all constant gains K linear function of the state plus the f i s t time integral of a certain
in the feedback path O<K<8.3, thesystem is as>mptoticallystable. other time-invariant linear function of the state.
2) A circle Cc with its center on the negative real axis, passing
through (- 1, 0 ) and touching G(iw), intersects the real axis a second IXTRODECTIOX
time a t - 1/2.22. Hence for all nonlinear time-varying gain functions The optimal regulator problem for linear dynamic systems canbe
in the feedback path satisfying 1 < F ( y , t)/y<2.22, the system is roughly stated as follows. In the class of piecewise continuous func-
asymptotically stable. tions, find a vector control u(t) which minimizes the functional1
3) A circle C M ,passing through (- 1 , O ) and touching the Syquist
plot a t two points, intersects the real axis a t - 1/6.66. Hence by the
circle criterion developed in this paper, the feedback system with
Jb1=
I
zJo
T
+
[(do,Qxx!t)) ( 4 % Ru(t))ldt (1)
monotonic increasing nonlinear functions in the feedback path is subject to the restrictions
stable for all F ( . ) satisfying
d = Ax +
Fu(l), . = d/dt (2)
x ( 0 ) = xo, ilxdl < (3)
x(T) = 0 (4
u(t) E 17, O $ t l T (5)
1711. CONCLUSION
A simple geometric criterion is derived in this paper as a sufficient where x is an n vector, the system state vector; u is an r vector,I I n ; Q
condition for determining the asymptotic stabilityof a feedback sys- and R are, respectively, nXtz and Y X Pnon-negative definite sym-
tem with asingle monotonic increasing nonlinear gain function in the metric matrices; A and F are n X n and n X r matrices, respectively;
feedback path. and C is a convex subset of the r-dimensional Euclidean space. The
Using the fundamental lemma derived, it is shown that if the terminal time T>O may be a fixed finite number or, alternatively,
Nyquist plot G(iw) lies entirely to theright of a straight line, an RL one may consider the limiting case T+m.
or RC multiplier Z(s) can always be found such that G(s)Z(s) is posi- From the point of view of design rationale, the term (x, Qx) in
tive real. This in turn insures the stabilityof the system for all mono- (1) is chosen to penalize deviations of the regulated state x@) from
tonic feedback gains in some range the desired equilibrium condition x ( t ) = 0 whereas the term (u,Ru)
discourages the use of excessively large control effort. One version of
this problem, the special case U=Er, R>O, Q>O, was rigorously
solved in the well-known 1960 paperby Kalman [l]. Since that
By suitable transformations,it canalso be shown that if the Nyquist time, several other special cases of the optimal regulator problem (1)
plot G(&) of the linear part lies entirely outside a circle intersecting
Manuscript received September 11 1967. revised December 28. 1967. and April
the negative real axis at (-l/Kl, 0) and (- l/K*, 0), then the non- 23. 1968. This research was supported'by tde XationalAeronautics and Space Ad-
linear system is asymptotically stable in the sector ministration under Contract NAS8-18008 with the Convair Division. General DY-
namics Corp.
The author is xcitb the Dept. of Electrical Engineering. Cniversity of Alabama
in Huntsville, Ala. Heisalso Consultant to the ConvairDivision, General Dymmics
corn.
1 (x,y ) denotes the inner productof x and y.

You might also like