You are on page 1of 5

736

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 5, MAY 2004

Characterization of PID and Lead/Lag Compensators Specifications Satisfying Given


Franco Blanchini, Antonio Lepschy, Stefano Miani, and Umberto Viaro
AbstractThis note deals with the problem of characterizing a class of second-order three-parameter controllers [including proportional integral-derivative (PID) and lead/lag compensators] satisfying given closed-loop specifications. Design characterizations of similar form as in the recent work on PID control, are derived for a larger class of compensators using simple geometric considerations. Specifically it is shown that, given the value of one parameter: i) the region of the plane defined by the other two parameters where the considered constraint is satisfied, consists of the union of disjoint convex sets whose number can be bounded by means of the pancake-cutting formula, and ii) the closed-loop pole distribution can be related to them. An example illustrates how the method can be applied to design a PID controller in the case of bounded sensitivity.

analytic expression of the boundary of the feasible regions is also derived. The special case of PID controllers satisfying given H1 specifications on the sensitivity function is then examined in more detail (Section IV). An example illustrates how the method can be used for design purposes. In the following sections, reference is made to the usual one-degree-of-freedom unity-feedback system for controlling a plant with strictly proper rational transfer function:

P (s) =

N (s ) D (s )

(1)

where N (s) and D(s) are coprime polynomials and D(s) is monic and of degree n. It is assumed that the controller transfer function has the form

C (s ) =

x + ys + zs2 dx + s + dzs2

(2)

, parameter space, proportionalintegral-derivative Index Terms (PID) control, sensitivity.

I. INTRODUCTION AND PRELIMINARIES The control literature shows a renewed interest in the synthesis of low-order controllers [1][6], which is mainly related to the exploitation of robust control techniques, with special regard to the so-called parametric approach [7], in a field that is important for industrial applications [8]. In particular, on the basis of a generalization of the HermiteBiehler theorem it has been proved [1] that, for any assigned value of the proportional gain kP of a proportionalintegral-derivative (PID) controller, the stability region of the plane whose coordinates are the derivative gain kD and the integral gain kI , is the intersection of the admissible solutions of sets of linear equations. By pursuing this kind of characterization, Ho [2], [3] has recently shown that the set of admissible (kD ; kI )-values where the closed-loop transfer function satisfies a given H1 specification, is formed by the union of disjoint convex sets (and that the related design procedure can be carried out by solving a family of linear programming problems, rather than via a brute-force search). In this note, we investigate the problem of characterizing the set of parameters which meet the H1 specification, using simple analytic and graphic tools. On the basis of the properties of linear fractional transformations (LFTs) it is possible to jointly consider both PID controllers and lead/lag networks. The proposed analysis is based on intuitive geometric considerations (Section II) that allow us to easily characterize the admissible parameter regions, to provide an upper bound on the number of disjoint convex sets where the H1 bound is not exceeded, and to show that the closed-loop pole distribution with respect to the imaginary axis is the same for all points inside each of them (possibly after splitting some of them into convex subsets to take account of pole transitions from the LHP to the RHP, or vice versa, at zero or at infinity, if any) (Section III); an
Manuscript received January 24, 2003; revised August 6, 2003 and December 5, 2003. Recommended by Associate Editor U. Jonsson. F. Blanchini is with the Department of Mathematics and Computer Science, University of Udine, 33100 Udine, Italy (e-mail: blanchini@uniud.it). A. Lepschy is with the Department of Information Engineering, University of Padova, 35131 Padova, Italy. S. Miani and U. Viaro are with the Department of Electrical, Managerial, and Mechanical Engineering, University of Udine, 33100 Udine, Italy. Digital Object Identifier 10.1109/TAC.2004.825961

where x; y; z are (real) parameters to be determined according to certain H1 specifications, whereas d is a (real) parameter fixed a priori (for instance, in order to obtain the desired steady-state precision). When d = 0, the controller takes the usual PID structure; in this case, x corresponds to the integral gain kI ; y to the proportional gain kP , and z to the derivative gain kD . When d = 1; C (s) takes the form of the classic unit-gain lead/lag network, usually written as

C (s ) =

in this case, z=x = 1 2 ; y=x = 1 + 2 ; (1 0 y )=x = 12 . Many closed-loop transfer functions of practical interest can be written in the form

1 + (1 + 2 + 12 )s + (1 2 )s2

1 + (1 + 2 )s + (1 2 )s

(3)

F (s ) =

A (s )C (s )P (s ) + B (s ) 1 + C (s )P (s )

(4)

where A(s) and B (s) are suitable stable rational functions. For instance, the sensitivity function S (s) is obtained for A(s) = 0 and B (s) = 1, while the complementary sensitivity function (or total transfer function) T (s) is obtained for A(s) = 1 and B (s) = 0. Expression (4), however, may also represent a weighted sum of S (s) and T (s). Remark 1.1: Note that (4) does not include the control sensitivity function C (s)=[1 + C (s)P (s)] as long as A(s) and B (s) are proper and stable. However, the extension of the results to this transfer function is straightforward and omitted for brevity. The design criterion referred to in the following sections consists in determining C (s), i.e., its free parameters x; y; z , in such a way that the considered closed-loop transfer function satisfies the constraint

!0

sup

jF (j!)j 

(5)

where is an assigned positive real number. Let us observe, in this regard, that assuming constant entails no restriction because any weighting function can be absorbed into A(s) and B (s) (in fact, could even be set equal to 1). II. GEOMETRIC INTERPRETATION For s = j! , the controller transfer function becomes
(x

C (j! ) =

0 !2 z) + j!y d(x 0 ! 2 z ) + j!

(6)

which points out that parameter y influences only the imaginary part of its numerator, and parameters x and z influence only the real parts of

0018-9286/04$20.00 2004 IEEE

Authorized licensed use limited to: LUNGHWA UNIV OF SCIENCE AND TECHNOLOGY. Downloaded on November 7, 2008 at 14:34 from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 5, MAY 2004

737

both its numerator and its denominator where they appear in the linear combination

 := x 0 !2 z:

(7)

In view of (6) and (7), the considered closed-loop function for s = j! can be written as

F (j!; ; y ) =

w1 (j!; y ) + w2 (j!; y ) w3 (j!; y ) + w4 (j!; y )

(8)

which shows that (8) is an LFT with respect to  according to certain (complex) coefficients w1 ; w2 ; w3 ; w4 depending, in general, on ! and y . It follows that, for fixed ! and y , as  varies over IR function (8) describes a circumference 8y;! . Therefore, condition

jF (j!; ; y)j 

(9)

can be interpreted geometrically as follows: given the values of ! and y , the admissible values of , if any, correspond to the intersection of the aforementioned circumference 8y;! with the disk 0 of radius centered at the origin, as shown in Fig. 1. Denoting by l (! ) and L (! ), respectively, the least and the largest value of  corresponding to the intersections, if any, with the circular boundary of the disk centered at the origin, four situations are clearly possible. i) The admissible values of  belong to the interval 3(!) = [l (!); L (!)], which means that jF (j!; 1; y)j > . ii) The admissible values of  are outside 3(! ), which means that jF (j!; 1; y)j < . iii) No admissible value exists, i.e., the circumference is external to the disk 0. iv) All values of  are admissible, i.e., the circumference is internal to the disk 0. Since  depends linearly on the parameters x and z according to (7), in case i) the admissible parameter values are those inside the stripe of the (z; x)-plane between the straight lines

Fig. 1. Intersection of circumference 8

with disk 0.

(a)

(b)

Fig. 2. Regions associated with intervals of the (a) first and (b) second type.

x 0 ! 2 z 0  l (! ) = 0 x 0 ! 2 z 0  L (! ) = 0

(10) (11)

and in case ii) those outside the same stripe. Let us now analyze what happens as ! varies from 0 to 1. The set
:= f! : !  0g can be subdivided into consecutive intervals Jk := f! : !k  !  !k+1 g within each of which only one of the four aforementioned situations holds. Clearly, if there is at least one value of ! such that situation iii) holds, then no (z; x)-pair will satisfy specification (5), which turns out to be too tight for the considered value of y . Conversely, the intervals where situation iv) holds can be disregarded because they do not impose any restriction. Therefore, the analysis can be limited to the intervals where either situation i) or situation ii) occurs. In the first case, by indicating with Jk the relevant interval, the region of the (z; x) parameter plane where condition (9) is satisfied 8! 2 Jk , is defined by

Remark 2.1: The intersection of these stripes is a single convex set, being the intersection of convex sets [see Fig. 2(a)]. Similarly, in the second case, by indicating again with Jk the relevant interval, the parameter region where (9) is satisfied 8! 2 Jk , is defined by either x 0 ! 2 z or x 0 !2 z

  l (! ) ;   L (! )

2 Jk

(13)

 l (! )  x 0 ! 2 z

  L (! );

! 2 Jk

(12)

which is the intersection of the stripes included between lines (10) and (11), 8! 2 Jk .

and is given by the intersection of all half-planes external to the stripes between lines (10) and (11), 8! 2 Jk . Remark 2.2: It is immediately seen that the intersection of the halfplanes defined by the first of inequalities (13), 8! 2 Jk , forms a convex set, and that the same is true for the intersection of the half-planes defined by the second of inequalities (13). These two sets are clearly disjoint [see Fig. 2(b)]. Simple continuity arguments show that any point not belonging to either sets is internal to, at least, one stripe corresponding to a value of ! 2 Jk and, thus, does not satisfy constraint (5). By taking into account (8), l (! ) and L (! ) can easily be computed as the roots of the second-degree polynomial equation resulting from jF (j!; ; y )j = . It follows that, given the nature of coefficients wi (j!; y ), they are differentiable functions of ! in the interior of each interval of the type i) and ii). The boundaries of

Authorized licensed use limited to: LUNGHWA UNIV OF SCIENCE AND TECHNOLOGY. Downloaded on November 7, 2008 at 14:34 from IEEE Xplore. Restrictions apply.

738

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 5, MAY 2004

the (z; x)-regions defined by (12) or (13) are the envelopes of the families of the straight lines (10) or, respectively, (11), 8! 2 Jk (whose slope depends quadratically on ! ). Therefore, the analytic expression of the boundary of the region defined by the first of inequalities (12) [or (13)] can be determined from

On the basis of these considerations, the following proposition can be stated. Proposition 3.2: The overall number of convex components for any given value of y is less than or equal to

x 0 ! 2 z 0  l (! ) = 0

@ [x 0 ! 2 z 0 l (! )] @!

:=

2 + up + 2 up

2:

(21)

=0

(14)

which can be rewritten in the usual parametric form

= 0 1 @  l (! ) 2! @!

x=0

! @  (! ) + l (! ): 2 @! l

(15)

The equations for the boundary of the region defined by the second of inequalities (12) [or (13)] have the same form with L (! ) replacing l (! ). Clearly, (15) represents the lower boundary of the admissible region defined by (12) and the upper boundary of the lower part of the admissible region defined by (13) (see Fig. 2). III. CHARACTERIZATION OF PARAMETER REGIONS For any given value of y , the feasible values of x and z are those satisfying (9) for all !  0. In view of the considerations made in Section II, the following result holds. Proposition 3.1: The set A of all the feasible parameter pairs (z; x) corresponding to a given value of y is the union of disjoint convex sets (henceforth, the convex components Ck ). Proof: This is an immediate consequence of the fact that A is obtained from the intersection of the admissible parameter regions corresponding to all ! -intervals Jk and that, according to Remarks 2.1 and 2.2, these regions consist of disjoint convex sets. The determination of the intervals Jk that partition
entails finding the values of ! for which the related circumference 8y;! intersects (twice) the boundary of disk 0. This clearly occurs when the rational equation

Proof: The maximal number of convex components is achieved when the number of intervals Jk giving rise to a situation of type (ii) in Section II, i.e., a situation in which inequalities (13) hold, is in turn maximal. As already said, these intervals define two disjoint regions of the (z; x)-plane separated by a stripe included between the two branches of a hyperbola-like curve (which may degenerate into a straight line). The considered problem is therefore equivalent to the determination of the maximum number of pieces in which it is possible to divide the plane for a given number of cuts. Expression (21), usually called the pancake-cutting formula [9], supplies the number of pieces when the number of cuts is equal to the upper bound up . A. Root Distribution So far, we have considered only magnitude constraints on F (s), which may also be satisfied by an unstable system. In this section we consider its pole distribution. Now, we assume that A1) y 6= 0 (the only values of practical interest); A2) d 6= 0 when N (0) = 0 (no pole-zero cancellation at the origin in C (s)P (s)); A3) A(j! ) 6= B (j! ); 8! . The pole location of F (s) in (4) depends on the roots of

py (x; z ; s) := (x + ys + zs2 )N (s) + (dx + s + dzs2 )D(s) = 0: (22)

jF (j!; ; y)j2 = 2

(16)

admits (two, possibly coincident) real solutions in terms of  (which have previously been denoted by l and L ). By recalling (8), (16) is equivalent to the second-degree polynomial equation in  whose coefficients are functions of ! 2 . By indicating with Ri and Ii ; i = 1; . . . ; 4, the real and imaginary parts of wi , i.e., wi = Ri + jIi , the discriminant 1(! 2 ) of (17) turns out to be (except for a positive proportionality factor)

jw1  + w2 j2 0 2 jw3  + w4 j2 = 0

(17)

1(!2 ) = [(R1 R2 + I1 I2 ) 0 2 (R3 R4 + I3 I4 )]2 1 (!2 ) (18) 0 (jw1 j2 0 2 jw3 j2 )(jw2 j2 0 2 jw4 j2 ) =: n d 1 (! 2 ) where d1 (! 2 ) is the least common denominator of all rational addenda in (18) and n1 (! 2 ) the corresponding numerator. It follows that there
can be at most

! ) where the sign of (18) may change; consequently, the number J of intervals Jk where (18) is nonnegative (so that (17) admits real solutions) cannot exceed the bound up

:= deg[n1 (!2 )] + deg[d1 (!2 )] (19) nonnegative real values of ! 2 (and, thus,  nonnegative real values of


Let us note first that one root goes from the right half-plane (RHP) to the left half-plane (LHP), or vice versa, through the origin when x passes from positive to negative values, or vice versa, and through the point at infinity when the degree of py (x; z ; s) decreases; this occurs for z = 0 if d 6= 0 and, if d = 0 and deg[N (s)] = deg[D(s)] 0 1 = n 0 1, for z = 01=Nn01 where Nn01 denotes the coefficient of sn01 in N (s). Now, let Di denote a) for d = 0 and deg[N (s)] < n 0 1, the intersection of a convex component Ck with either x > 0 or x < 0; b) for d = 0 and deg[N (s)] = n 0 1, the intersection of a convex component Ck with one of the four open quadrants generated by the straight lines x = 0 and z = 01=Nn01 ; c) for d 6= 0, the intersection of a convex component Ck with one of the four open quadrants generated by the straight lines x = 0 and z = 0. The following result holds. Proposition 3.3: The root distribution of (22) is the same for all (z; x)-points in Di . Proof: Let us assume, by contradiction, that the root distribution at a point (z1 ; x1 ) 2 Di is different from that at another point (z2 ; x2 ) 2 Di . The convexity of Di ensures that (z1 ; x1 ) and (z2 ; x2 ) can be connected with a continuous curve (actually, even a straight-line z; x ^) segment) entirely contained in Di . Therefore, there exists a point (^ on this line and a frequency ! ^ such that

py (^ z; x ^; j ! ^ ) = (^ x + jy ! ^ 0z ^! ^ 2 )N (j ! ^) ^ ) = 0: (23) + (dx ^ + j! ^ 0 dz ^! ^ 2 )D (j !

= ( + 1) 2

(20)

Note first that, by the definition of Di , the degree of py (x; z ; s) is the same at all points of this parameter set, so that no root can cross the imaginary axis at infinity as the parameter point travels along the con^ must be finite. sidered path; it follows that !

Authorized licensed use limited to: LUNGHWA UNIV OF SCIENCE AND TECHNOLOGY. Downloaded on November 7, 2008 at 14:34 from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 5, MAY 2004

739

Fig. 3. Convex components (left) and step responses (right) for k

= 2 and = 3.

Also, it is easily seen that

D(j ! ^ ) 6= 0 and dx ^ + j! ^ 0 dz ^! ^2

6= 0

(24)

because: If D(j ! ^ ) = 0, by the coprimeness of N (s) and D(s) (23) ^ + jy ! ^0z ^! ^ 2 = 0 and, thus, ! would imply x ^ = 0 [since y 6= 0 by ^ = 0, which is in contrast with the definition of Assumption A1)] and x Di ; and if dx ^ + j! ^ 0 dz ^! ^ 2 = 0, then ! ^ = 0 and dx ^ = 0, which would ^ = 0, which is again in contrast with the definition of imply either x Di , or d = 0, but in this case, if x ^ 6= 0, then N (0) = 0, in contrast with Assumption A2). Therefore, (23) is equivalent to

1 + C (j ! ^ )P (j ! ^ ) = 0:
Now, from the H1 constraint (5) it follows, in particular, that

(25)

If we adopt the further assumption that P (j! ) 6= 0; 8! , feasible parameter regions, i.e., regions where jS (j! )j  ; 8! , always exist because only situations of type (ii) and (iv) in Section II may arise; indeed, in this case jS (j! )j ! 0; 8! , as  ! 1, so that the related circumference 8y;! (see Fig. 1) passes through the origin 8! . Also, by indicating with Jk each of the disjoint ! -intervals (called active intervals in the sequel), if any, at whose frequencies circumference 8y;! intersects the boundary of disk 0 for (! ) = l (! ) and (! ) = L (! ), the parameter region where (9) is satisfied, is defined by (13) and consists of two disjoint convex sets. By indicating with RP (! ); IP (! ), and MP (! ), respectively, the real part, the coefficient of the imaginary part and the magnitude of P (j! ), (17) particularizes to
2 MP (!)2 + 2IP (!)! 2 (!)y2 + (1 0 02 ) + 2RP (!)y !2 + MP

A (j ! ^ )C (j ! ^ )P (j ! ^ ) + B (j ! ^) 1 + C (j ! ^ )P (j ! ^)

(26)

=0

(27)

and (18) to:

and this is in contradiction with (25) which would imply that the denominator of the fraction at the left-hand side of (26) is equal to zero ^ )(01)+ B (j! ^ ) that is different from and its numerator is equal to A(j ! zero since, by Assumption A3), A(j! ) 6= B (j! ); 8! . Clearly, to determine the entire region of the three-dimensional parameter space with the same pole distribution, it is necessary to sweep over y 2 (01; 1). However, if only the stability regions are of interest, then the bounds on y provided, e.g., in [6] and [3] can profitably be exploited. The present approach can also be adopted when H1 specifications on different closed-loop functions must simultaneously be met: the admissible regions will then be formed by the intersection of all the regions corresponding to every specification. IV. PID CASE WITH BOUNDED SENSITIVITY In this section, we deal with the special case of a PID controller [d = 0 in (2)] with a prescribed bound on the sensitivity function S (s) [A(s) = 0 and B (s) = 1 in (4)].

1(!2 ) = !2

2 4 IP (! ) 0 M P (! )y 2

2 2 0 (1 0 02 )MP (! ) 0 2M P (! )R P (! )y

: (28)

The upper bound on the maximal number of active intervals Jk can be determined according to (19)(21). Example: Let us consider the plant transfer function shown in (29) at the bottom of the page, whose Nyquist plot is quite involved due to the presence of two pairs of complex poles and a pair of complex zeros close to the imaginary axis. For kP = y = 2 and = 3, the sign of (28) changes for ! = 0; 2:933; 3:559; 5:764; 6:069; 8:224; 8:488, so that the active intervals turn out to be: J1 = [2:933; 3:559]; J2 = [5:764; 6:069]; J3 = [8:224; 8:488]. Therefore, the number of convex components Ck cannot exceed (32 +3+2)=2 = 7, which is indeed their actual number. Fig. 3 (left) shows (not shaded) the 5 disjoint regions of the upper half-plane (kI > 0) where jS (j!)j < 3 for the given proportionality gain (they are separated by shaded regions where this condition is not satisfied).

P (s ) =

2s5

+ 2:7255s4

10s3 + 9s2 + 362:4s + 36:16 + 138:4292s3 + 156:471s2 + 637:6472s + 360:1779

(29)

Authorized licensed use limited to: LUNGHWA UNIV OF SCIENCE AND TECHNOLOGY. Downloaded on November 7, 2008 at 14:34 from IEEE Xplore. Restrictions apply.

740

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 49, NO. 5, MAY 2004

As already said, to find the closed-loop pole distribution, it is enough to consider only one point inside each Di . Clearly, the controller parameters must be chosen within the stability regions. Usually, for small values of kI the step response reaches the steady-state value slowly (long settling time); therefore it is reasonable to choose a point in the stability region corresponding to the highest values of the integral gain. Since in the presence of sudden changes of the reference input the derivative gain kD should be kept small, a good choice corresponds to a point to the left of this region. Fig. 3 (right) depicts the responses to a step reference signal when the parameter point lies: a) in the left corner of the upper stability region (kD = 9:00; kI = 612:49), and b) in the left corner of the lower stability region (kD = 0:0594; kI = 1:984). V. CONCLUSION By means of simple geometrical considerations, a characterization of the regions of the parameter space for a class of standard controllers, including (but not limited to) PID controllers, satisfying assigned H closed-loop specifications has been obtained. Such a characterization is in line with previous results on PID control [1][3] based on generalizations of the Hermite-Biehler theorem. In particular, it has been shown that, for any fixed value of one controller parameter [i.e., y in (2)], the admissible region in the plane of the other two parameters [i.e., x and z in (2)] consists of disjoint convex sets (the convex components) whose number is bounded by the pancake-cutting formula. The boundaries of the convex components can easily be drawn using conventional computer graphics methods, even if their parametric equations can be obtained analytically. Moreover, it has been proved that the closed-loop pole distribution is the same at all points in the intersection Di of the interior of each convex component Ck with a suitable quadrant or half-plane (cf. Section III-A). The procedure has been particularized to the case of PID controllers with bounded sensitivity. By taking into account the shape of the feasible sets, a criterion for choosing the values of the controller parameters has been suggested. Finally, as far as the computational complexity of the suggested graphic procedure is concerned, it requires drawing lines (10) and (11) for a number of values of ! , say n! , for every value of y . By indicating with ny the number of values of y , the order of complexity is therefore O(n! ny ). Of course n! depends not only on the frequency discretization but also on the number J of active intervals Jk , which in turn requires finding the real zeros of polynomials n1 (!2 ) and d1 (! 2 ) in (18).

Necessary Conditions for Schur-Stability of Interval Polynomials


Richard Greiner
AbstractNew bounds on the coefficient diameter of real Schur-stable interval polynomials are given using techniques from complex analysis. They can be used to unmask interval polynomials at low computational cost as being non-Schur-stable. Index TermsD-stability, interval family, polynomials, robust stability, Schur-stability, Schwarz lemma.

I. INTRODUCTION A polynomial with real or complex coefficients is called Schur-stable if all its zeros are inside the open unit disk := fz 2 : jz j < 1g. Consider a linear difference equation xk+1 = Axk , where A is a real (or complex) n 2 n matrix. It is well-known that all solutions of such a time discrete system are asymptotically stable if and only if the characteristic polynomial p of A is Schur-stable. A standard task for those systems is to decide whether it is robust asymptotically stable, i.e., whether it remains asymptotically stable unter perturbations of A. This also includes the case, occurring in practical applications, that the elements of A are known only approximately (e.g., because of gauge inaccuracy). In all these cases, one gets a family of polynomials
p(z )

+ an01 z

n01

111 +

a1 z

+ a0

(1)

REFERENCES
[1] A. Datta, M. T. Ho, and S. P. Bhattacharyya, Structure and Synthesis of PID Controllers, London, U.K.: Springer-Verlag, 2000. [2] M. T. Ho, Synthesis of PID controllers, in Proc. 40th IEEE Conf. Decision Control, Orlando, FL, Dec. 2001, pp. 255260. , Synthesis of PID controllers: A parametric approach, Au[3] tomatica, vol. 39, no. 6, pp. 10691075, June 2003. [4] W. K. Ho, T. H. Lee, H. P. Han, and Y. Hong, Self tuning IMC-PID control with interval gain and phase margins assignment, IEEE Trans. Contr. Syst. Technol., vol. 9, pp. 535541, May 2001. [5] C. Hwang and C. Y. Hsiao, Solution of a nonconvex optimization arising in PI/PID control design, Automatica, vol. 38, pp. 18591904, Nov. 2002. [6] M. T. Sylemez, N. Munro, and H. Baki, Fast calculation of stabilizing PID controllers, Automatica, vol. 39, no. 1, pp. 121126, Jan. 2003. [7] J. Ackermann et al., Robust Control. The Parameter Space Approach, 2nd ed. London, U.K.: Springer-Verlag, 2002. [8] K. J. strm and T. Hgglund, PID Controllers: Theory, Design, and Tuning. Research Triangle Park, NC: Instrum. Soc. Amer., 1995. [9] N. J. A. Sloane and S. Plouffe, The Encyclopedia of Integer Sequences. San Diego, CA: Academic, 1995.

with real coefficients ak ; k = 0; 1; . . . ; n 0 1, varying independently + 0 + in given intervals [a0 k ; ak ]; ak  ak , (respectively, complex coefficients ak varying independently in given rectangles or disks). Such families of polynomials are called interval polynomials (respectively, disk polynomials). Hence, the question of robust stability of a system is reduced to the question whether all polynomials of such an interval polynomial are Schur-stable. Despite the fact that Schur-stability of a single (real or complex) polynomial p can be verified by standard methods (e.g., by algorithms based on the SchurCohn criterion (cf. [14, Sec. 11.5], [11], [15], or the HermiteFujiwara theorem, cf. [12]), not much is known about efficient necessary or sufficient conditions for Schur-stability of real interval polynomials. This is in sharp contrast (cf. [8]) to the case of Hurwitzstability, where the zeros of p are required to have negative real part. The well-known theorem of Kharitonov (cf. [13], [1, Ch. 5], [5, Sec. 5.2]) states that a real interval polynomial is Hurwitz-stable if and only if four certain extreme polynomials are Hurwitz-stable. Similarly (cf. [1, Sec. 6.9], [5, Sec. 5.3]), for families of complex polynomials with coefficients in rectangles, eight extreme polynomials have to be checked to ensure Hurwitz-stability. In the Schur case for complex disk polynomials a simple criterion is available, too, bounding the sum of the diameters of the coefficent disks by the minimum modulus of the central polynomial on the unit circle (cf. [5, Sec. 3.5.2]).

Manuscript received August 11, 2003; revised November 15, 2003. Recommended by Associate Editor A. Datta. The author is with the Department of Mathematics and Computer Science at the University of Wrzburg, Am Hubland, 97074 Wrzburg, Germany (e-mail: greiner@mathematik.uni-wuerzburg.de). Digital Object Identifier 10.1109/TAC.2004.825963

0018-9286/04$20.00 2004 IEEE

Authorized licensed use limited to: LUNGHWA UNIV OF SCIENCE AND TECHNOLOGY. Downloaded on November 7, 2008 at 14:34 from IEEE Xplore. Restrictions apply.

You might also like