IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO.

4, APRIL 2003 687

A New Approach to Digital PID Controller Design this third variable can be exactly determined. By sweeping or gridding
over the third variable, the complete stabilizing set can be determined
L. H. Keel, J. I. Rego, and S. P. Bhattacharyya constructively. The solution shows that the stabilizing set for any plant,
when it is nonempty, consists of unions of convex polygons in the space
of the PID gains.
Abstract—In this note, we present a new approach to the problem of
designing a digital proportional-integral-derivative (PID) controller for a The aforementioned solution technique is extended to solve two pre-
given but arbitrary linear time invariant plant. By using the Tchebyshev viously unsolved design problems. The first problem is related to dead-
representation of a discrete-time transfer function and some new results beat control wherein one places all closed loop characteristic roots at
on root counting with respect to the unit circle, we show how the digital the origin so that the transients are zeroed out in a finite number of
PID stabilizing gains can be determined by solving sets of linear inequali-
ties in two unknowns for a fixed value of the third parameter. By sweeping
steps. In general, deadbeat control is not possible using PIDs and a rea-
or gridding over this parameter, the entire set of stabilizing gains can be sonable goal is to place the closed loop characteristic roots as close to
recovered. The precise admissible range of this parameter can be predeter- the origin as possible so that the transient error decays quickly. Such
mined. This solution is attractive because it answers the question of whether designs have been advocated in the literature on sampled data control
there exists a stabilizing solution or not and in case stabilization is possible systems (see specifically [15, p. 292]). We show how the stabilization
the entire set of gains is determined constructively. Using this character-
ization of the stabilizing set we present solutions to two design problems: solution obtained by us can be exploited to give a constructive deter-
1) maximally deadbeat design, where we determine for the given plant, the mination of such “maximally” deadbeat designs. The second problem
smallest circle within the unit circle wherein the closed loop system charac- involves the determination of the maximum delay in the loop that a
teristic roots may be placed by PID control and 2) maximal delay tolerance, given plant under PID control can be made to tolerate. We show how
where we determine, for the given plant the maximal-loop delay that can
our solution can also be extended to determine this maximum delay for
be tolerated under PID control. In each case, the set of controllers attaining
the specifications is calculated. Illustrative examples are included. a given plant.
Index Terms—Deadbeat control, digital PID controller, stability,
Tchebyshev representation. II. TCHEBYSHEV REPRESENTATION AND ROOT CLUSTERING
The stabilization results to be developed later in the note require
us to determine the complex plane image of polynomials and rational
I. INTRODUCTION
functions on a circle of radius  centered at the origin.
There is renewed interest in proportional-integral-derivative (PID)
controllers (see [1], [2]) because of two reasons. First, they are exten- A. Tchebyshev Representation of Real Polynomials and Rational
sively used in applications in all industries (see [3], [4, Ch. 6]). Second, Functions
despite the existence of some results [5], [6] modern optimal control Let us consider a polynomial P (z ) = an z n + 1 1 1 + a0 with real
methods are not suitable to deal with fixed structure and fixed order coefficients. The image of P (z ) evaluated on the upper half of the circle
controllers (see [7, p. 3]). Thus, there is much that remains to be done C of radius , centered at the origin is
to modernize PID design methods.
Here, we develop some new results on discrete time PID controllers. P (z ) : z = ej ; 0 : (1)

It is well known ([8, p. 71]) that, with u = 0 cos  ,

First, the complex plane image of a real polynomial or rational func-
tion over a circle of radius  centered at the origin, is determined and
p
expressed in terms of Tchebyshev polynomials [8]. In [9], Tchebyshev P (ej ) = R(u; )+ j 1 0 u2 T (u; ) = : Pc (u; ) (2)
polynomials are used in robust control problems related to discrete time
systems. They have also been used to develop a discrete time version where
of Foster’s theorem and subsequently used to give necessary and suf-
ficient conditions for Schur stability [10], [11]. In this note, a formula
R(u; ) = an cn (u; )+ an01 cn01 (u; )+ 1 1 1 + a1 c1 (u; )+ a0
1 1 1 + a1 s1(u; )
is first developed for root counting with respect to circular regions in
terms of this Tchebyshev representation. This formula which differs T (u; ) = an sn (u; )+ an01 sn01 (u; )+
from the root counting formulas given in[12], [13] is an extension of and
an initial result presented in [14], and constitutes the generalization of
Hermite Bieler type results for Schur stability. Using these results, we
show how the PID controller can be reparametrized so that the stabi- ck (u; ) = k ck (u); sk (u; ) = k sk (u); k = 0; 1; 2 . . .
lizing set is obtained as the solution of sets of linear inequalities in two
and ck (u) and sk (u) are real polynomials in u satisfying the recursive
variables for a fixed value of the third variable. The admissible range of
relations
c0 (u)
Manuscript received January 18, 2002; revised July 18, 2002. Recommended sk (u) = 0 k ; k = 1; 2; . . . (3)
k
by Associate Editor Y. Ohta. This work was supported in part by NASA 2
ck+1 (u) = 0 uck (u) 0 (1 0 u )sk (u); k = 1; 2; . . . (4)
under Grant NCC-5228 and by the National Science Foundation under Grant
HRD-9706268. This research was done while J. I. Rego was a visiting scholar,
supported by CNPQ, at the Department of Electrical Engineering, Texas A&M
and are known as the Tchebyshev polynomials of the first and second
University, College Station, Texas. kind, respectively. The complex plane image of P (z ) as z traverses the
L. H. Keel is with the Center of Excellence in Information Systems, Tennessee upper half of the circle C can be obtained by evaluating Pc (u; ) as u
State University, Nashville, TN 37203 USA (e-mail: [email protected]). runs from 01 to +1.
J. I. Rego is with the Department of Electrical Engineering, Federal Univer- Now, let Q(z ) be a ratio of two real polynomials P1 (z ) and P2 (z ).
We compute the image of Q(z ) on C and write it as the corresponding
sity of Rio Grande Do Norte, Natal 59078-370, Brazil.
S. P. Bhattacharyya is with the Department of Electrical Engineering,
Tchebyshev representation Qc (u; ) as follows. Let
p
Texas A&M University, College Station, TX 77843 USA (e-mail: bhatt@
eesun1.tamu.edu).
Digital Object Identifier 10.1109/TAC.2003.809768 j
Pi (z ) z =0u+j p10u = Ri (u; ) + j 1 0 u2 Ti (u; ) (5)

0018-9286/03$17.00 © 2003 IEEE

688 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 4, APRIL 2003

for i = 1; 2. Then, it is easy to show that Lemma 2: Let Q z ( ) = ( ( )) ( ( ))

P1 z = P2 z where the real polyno-
R(u;) () ()
mials P1 z and P2 z have i1 and i2 roots, respectively, in the interior
of the circle C and no roots on the circle. Then
R1 (u; )R2 (u; )+(1 0 u2 )T1 (u; )T2 (u; )
Q(ej ) = 1 [Q ()] = (i 0 i ) = 10 [Q (u)] :
R22 (u; )+(1 0 u2 )T22 (u; )
+1
0 1 2 1

T (u;)
p
+ j 1 0 u (TR ((u;u;))+(1
R (u; ) 0 R (u; )T (u; ))
2 B. Root Counting and Tchebyshev Representation
1 2 1 2
0 u )T (u; )
2 2 2 () p P 2z and its Tchebyshev repre-
Let us begin with a real polynomial
( ) = ( )+ 1 ( ) 0 u T u;  as developed before.
2 2
= : Qc (u; ): (6) sentation PC u;  R u; 
...
Henceforth, let 31 t1 ; ( )
; tk denote the real distinct zeros of T u; 
This representation will be needed in a later section on the solution of ( 1 1)
of odd multiplicity, for u 2 0 ; , ordered as follows:
the maximally deadbeat problem.
01 < t < t < 1 1 1 < tk < +1: 1 2

Suppose also that T (u; ) has p zeros at u = 01 and let f i (x ) denote

B. Interlacing Conditions for Root Clustering and Schur Stability
0
The formulas of the last section can be used to derive conditions for the ith derivative to f (x) evaluated at x = x . Let us also define 0
root clustering in circular regions, that is for the roots to lie strictly
within a circle of radius . For Schur stability we simply take  = 1. 01 if x < 0
As before let P (z ) be a real polynomial of degree n and Sgn[x] = 0 if x = 0
p 1 if x > 0:
P (e ) = R (; ) + j T(; ) = R(u; ) + j 1 0 u T (u; ) (7)
j 2
We begin with the following.
where u = 0 cos  , and R(u; ) and T (u; ) are real polynomials of Theorem 2: Let P (z ) be a real polynomial with no roots on the
degree n and n 0 1, respectively, in u, for fixed . circle C and suppose that T (u; ) has p zeros at u = 01. Then, the
Theorem 1: P (z ) has all its zeros strictly within C if and only if number of roots i of P (z ) in the interior of the circle C is given by
1) R(u; ) has n real distinct zeros ri , i = 1; 2; 1 1 1 ; n in (01,1); 1
2) T (u; ) has n 0 1 real distinct zeros tj , j = 1; 2; 1 1 1 ; n 0 1 i = 2 Sgn T (01; ) (Sgn [R(01; )]
p ( )

in (01,1); k
3) the zeros ri and tj interlace +2 (01)j Sgn [R(tj ; )] + (01)k +1
Sgn [R(+1; )] : (9)
01 < r < t < r < t < 1 1 1 < tn0 < rn < +1: j =1

(; )+ j T(; ) and define i ,

1 1 2 2 1
Proof: Recall that P (ej ) = R
Proof: Let tj = 0 cos j , j 2 (0;  ), j = 1; 2; . . . ; n 0 1 or i = 1; 1 1 1 ; k through ti = 0 cos i , for i 2 [0; ]. Let  := 0, 0

j = 0 cos
0 tj ; j = 1; 2; . . . ; n 0 1
1
= 0; n =  t := 01 and k :=  , and note that the i , i = 0; 1. . . ; k + 1 are
zeros of T(; ). The proof depends on the following elementary and
0 0 +1

and let i = 0 cos0 ri , i = 1; 2; . . . ; n, i 2 (0;  ).

1
easily verified facts which are first stated. (In the following, i denotes +

Then, ( ; ; 1 1 1 ; n ) are the n + 1 zeros of T(; ) = 0 and the point immediately to the right of i ).
( ; ; 1 1 1 ; n0 ) are the n zeros of R (; ) = 0. The condition (c)
0 1

1 [()] = i
1 2 1

means that i and j satisfy: (a) 0

(b) 1 [()] = 1 [()] + 1 [()] + 1 1 1 + 1 [()]

0= < < 1 1 1 < n01 < n = :
< < (8) 0 0

[()] = 2 Sgn T i ;  Sgn R(i ; )

0 1 1 2
(c) 1
 +
Conditions 1)–3) imply that the plot of P (e ) for  2 [0; ] turns
j

counterclockwise through exactly 2n quadrants and this condition is 0Sgn R(i ; ) ; i = 0; 1; . . . ; k

+1
equivalent to P (z ) having n zeros inside the circle C . 555 (d) Sgn T i ;  = 0Sgn T i ;  ;
+
i = 0; 1; 1 1 1 ; k
+
+1

III. ROOT COUNTING FORMULAS (e) Sgn T(0 ; ) = Sgn T (01; )

+ p ( )

A. Phase Unwrapping and Root Distribution (f) Sgn R(i ; ) = Sgn [R(ti ; )] ; i = 0; 1; 1 1 1 ; k:
Let P () := ArgP(ej ) denote the phase of P (z ) evaluated at Using (a)–(f), we have
z = ej and let 1 [P ()] denote the net change in or unwrapped
phase of P (ej ) as  increases from 1 to 2 . Similarly notation
i =10 [()]
applies to the rational function Q(z ) with Tchebyshev representation =10 [()] + 1 1 1 + 1 [()] by (a) and (b)
QC (u; ): let Q (u) = ArgQC (u; ) denote the phase of QC (u; ) = 2 Sgn T(0+ ; ) Sgn R(0; ) 0 Sgn R (1; )
and 1uu [Q (u)] the net change in or unwrapped phase of QC (u; )
as u increases from u1 to u2 . + 1 1 1 + Sgn T(k+ ; ) Sgn R(k ; )
Lemma 1: Let P (z ) have i roots in the interior of the circle C and 0 Sgn R(; ) by (c)
no roots on the circle. Then 
= Sgn T 0 ;  Sgn R(0; ) 0 Sgn R (1 ; )
 +

1 [P ()] = i: 2

0
0 R (1 ; ) 0 Sgn R (2 ; ) + 1 1 1 + (01)k
Sgn
Proof: From geometric considerations it is easily seen that each 2 Sgn R (k ; ) 0 Sgn R (; )
2 1 [ ( )]
by (d)
interior root contributes  to 20 P  and, therefore, because of 
the symmetry of roots about the real axis the interior roots contribute = Sgn[T (p)(01; )] Sgn R (0; ) 0 2Sgn R (1; )
2
1 [ ( )]
i to 0 P  . 555 + 2Sgn R( ; ) + 1 1 1 + (01)k Sgn R(k ; )
2

+(01)k Sgn R(; ) by (e)

We state the corresponding result for a rational function. The proof
+1
is similar to the previous lemma and is omitted.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 4, APRIL 2003 689

Sgn T (p) (01; )

fSgn [R(01; )] 0 2Sgn [R(t ; )] Multiplying the characteristic polynomial by z 01 N (z 01 ), we have

= 1
2
+2Sgn [R(t ; )] + 1 1 1 + (01) 2Sgn [R(tk ; )]
k
2 01
z  (z )N (z
01 ) = (z 0 1)D(z )N (z 01 )
+(01) 01 )N (z )N (z 01 ):
k +1
Sgn [R(+1; )] by (f) +(K2 z + K1 + K0 z
from which the result follows. 555
The previously derived result can now be extended to the case of ra- Using the Tchebyshev representations, we have
tional functions. Let Q(z ) = (P1 (z ))=(P2 (z ))pwhere Pi (z ), i = 1; 2
are real rational functions. Let Ri (u; ) + j 1 0 u2 Ti (u; ), i = z
01  (z )N (z 01 ) = 0 (u + 1)P1 (u) 0 (1 0 u2 )P2 (u)
1; 2 denote the Tchebyshev representations of Pi (z ), i = 1; 2 and 0 [(pK 0 + K2 )u 0 K ] P (u) 1 3
QC (u; ) denote the Tchebyshev representation of Q(z ) on the circle
C . Let R(u; ), T (u; ) be defined by +j 1 0 u2 0
[ (u + 1)P2 (u)
+P1 (u) + (K2 0 K )P (u)] 0 3
R(u; ) = R1 (u; )R2 (u; ) + (1 0 u )T (u; )T (u; )
2
= R(u; K0 ; K1 ; K2 )
p
1 2

0 R (u; )T (u; ):

T (u; ) = T1 (u; )R2 (u; ) 1 +j 1 0 u T (u; K ; K ):
2
0 2

Suppose that T (u; ) has p zeros at u = 01 and let t 1 1 1 tk denote the 1

Now, let K3 := K2 0 K0 and rewrite R(u; K0 ; K1 ; K2 ) and
real distinct zeros of T (u; ) of odd multiplicity ordered as follows:
T (u; K0 ; K2 ) as follows:
01 < t < t2 < 1 1 1 < tk < +1:
1
R(u; K0 ; K1 ; K2 ) = 0 (u +1)P (u) 0 (1 0 u )P (u) 1
2
2

Theorem 3: Let Q(z ) = (P1 (z ))=(P2 (z )) where Pi (z ), i = 1; 2 0 [(2K 0 K )u 0 K ] P (u) 2 3 1 3 (14)

T (u; K ) =P (u) 0 (u +1)P (u)+ K P (u):
are real polynomials with i1 and i2 zeros, respectively, inside the circle
C and no zeros on it. Then 3 1 2 3 3 (15)

We observe the parameter separation previously achieved: K3 appears

i1 0i 2 =
1
2
Sgn T (p) (01; ) (Sgn [R(01; )]
only in the imaginary part and K1 , K2 , K3 appear linearly in the real
k part. Thus, by applying root counting formulas to the rational func-
+2 0
( 1)
j
Sgn [R(tj ; )] + (01)k+1 Sgn [R(+1; )] : (10) tion on the left, and imposing the stability requirement yields linear
j =1 inequalities in the parameters for fixed K3 . The solution is completed
by sweeping over the range of K3 for which an adequate number of
Proof: The proof is based on the representation of QC (u; ) de-
real roots tk exist. We illustrate with an example.
veloped in (6). Since the denominator of (6) is strictly positive for
u 2 [01; +1], it follows that the phase unwrapping can be computed
Example 1:
from the numerator. The rest of the proof is similar to the proof for the 1
polynomial case and is left to the reader. 555 G(z ) =
z2 0 0:25
Then
IV. PARAMETER SEPARATION AND STABILIZATION WITH
PID CONTROLLERS
RD (u) = 2u
2
0 1:25; TD (u) = 02u; RN (u) = 1; TN (u) = 0
Consider the control system in unity feedback configuration wherein
the plant is a represented by its discrete time transfer function G(z ) = P1 (u) = 2u
2
0 1:25; P (u) = 02u; P (u) = 1:
2 3

(N (z ))=(D (z )) with N (z ), D (z ) being polynomials with real coeffi-

cients and with degree D(z ) = n and degree N (z )  n. The closed Recall (14). Since G(z ) is of order 2 and C (z ), the PID controller, is
loop system is stable iff the characteristic polynomial, denoted by  (z ), of order 2, the number of roots of  (z ) inside the unit circle is required
is Schur stable. The general formula of a discrete PID controller, using to be 4 for stability. From Theorem 2
backward differences to preserve causality, is
ii 0i = (i + iN ) 0 (l + 1) (16)
1 z 0z 1 + KTD 1 z 0z 1
2

C (z ) = KP + KI T i i

KP + KI T + T z + 0KP 0
K 2 2K K
T z + T
: where i and iN are the numbers roots of  (z ) and the reverse poly-
z (z 0 1)
=
nomial of N (z ), respectively, and l is the order of N (z ). Since the
Therefore, we can use required i is 4, iN = 0, and l = 0, i1 0 i2 is required to be 3. To
illustrate the example in detail, we first fix K3 = 1:3. Then, the real
K2 z 2 + K1 z + K0 roots of T (u; K3 ) in (01, 1) are 00.4736 and 00.0264. Furthermore,
C (z ) =
z (z 1) 0 (11) Sgn[T (01)] = 1, and from Theorem 2, i1 0 i2 = 3 requires that
where
Sgn [T (01)] (Sgn [R(01)] 0 2Sgn [R(00:4736)]
1
2
KP = 0 K 0 2K
1 0 KI =
K0 + K1 + K2
T
+2Sgn [R( 0:0264)] 0 0 Sgn [R(1)]) = 3:
KD = K0 T : (12)
In this example, we have Sign[R(tj )], j = 0; 1; 2; 3, and each
The characteristic polynomial becomes Sign[R(1)] may assume the value +1 or 01 since 0 is excluded as we
are testing for stability. This leads to 24 = 16 possible strings which
 (z ) = z (z 0 1)D(z) + 2
K2 z + K1 z + K0 N (z ): (13) need to be tested. In general, it is easy to devise a sorting algorithm to
690 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 4, APRIL 2003

Fig. 1. Stability regions in (K , K , K ), and (K , K , K ) spaces.

pick out the feasible strings. Here, we have only one valid sequence V. MAXIMALLY DEADBEAT CONTROL VIA PID CONTROLLERS
satisfying the aforementioned equation, namely
An important design technique in digital control is deadbeat control
Sgn [R(01)] Sgn [R(00:4736)] Sgn [R(00:0264)] wherein one places all closed-loop poles at the origin. If this is used in
1 01 1
conjunction with integral control the tracking error is zeroed out in a
finite number of sampling steps. Deadbeat control requires in general
Sgn [R(1)] 2(i1 0 i2 ) that we be able to control all the poles of the system. However, such
01 6: a pole placement design is in general not possible when a lower order
controller is used. Thus, we are motivated to design a PID controller
From this valid sequence, we have the following set of linear inequali- that places the closed-loop poles as close to the origin as possible. The
ties: transient response of such a system will decay out faster than any other

01:3+ K1 +2K2 >0 0 0:9286+ K1 +0:9472 < 0

design and therefore the fastest possible convergence of the error under
PID control will be achieved.
1:1286+ K1 +0:0528K2 >0 0 0:2+ K1 0 2K2 < 0: The design scheme to be developed will attempt to place the closed
loop poles in a circle of minimum radius . Let S denote the set of PID
This set of inequalities characterize the stability region in (K1 , K2 ) controllers achieving such a closed loop root cluster. We show below
space for the fixed K3 = 1:3. By repeating this procedure for the range how S can be computed for fixed . The minimum value of  can be
of K3 , we obtain the the stability region shown in the left of Fig. 1. found by determining the value 3 for which S =  but S 6= ,
Consider the following relation:  > 3 .
Now, let us again consider the PID controller
KP 02 01 0 K0
1 1 1
KI = T T T K1 K2 z 2 + K1 z + K0
C (z ) =
KD T 0 0 K2 0
z (z 1)
(17)
02 01 0 0 1 01 K1
=
1 1 1 1 0 0 K2 : and the characteristic polynomial
T T T
T 0 0 0 1 0 K3
 (z ) = z (z 0 1)D(z) + (K2 z2 + K1 z + K0 )N (z): (18)
Using this relation, we plot the stabilizing region in (KP , KI , KD )
space in the right of Fig. 1. Note that
Remark 1: From (16) and Theorem 2, it is clear that a necessary
p 2
condition for stabilization is that T (u; K3 ) must have two real zeros
in (01, 1). In this example, this specifies the admissible range of K3
D(z ) jz=0u+jp10u = RD (u; )+ j 1 0 u TD (u; )

to be (00.75, 1.5). In a like manner, the range of K3 can always be p p 2

N (z )j = RN (u; )+ j 1 0 u TN (u; )
z =0u+j 10u
predetermined from the requirement on the number of real roots of
and
T (u; K3 ).
Remark 2: An alternative approach to determine the stabi-
lizing set is via D-decomposition. In this approach, one sets N ( z
2 01 )j
0
z = u+j
p10u jz=0u0jp10u
= N (z )
 (ej ; KP ; KI ; KD ) = 0 and determines the corresponding solution p
surfaces in the (KP , KI , KD ) space. These surfaces partition this = RN (u; ) 0 j 1 0 u2 TN (u; ):
space into disjoint open regions each with a fixed number of roots in
the interior of the unit circle. The stabilizing regions will then have to We now evaluate
be picked out by testing an arbitrary point from each region. On the
other hand, our approach directly determines the stabilizing regions. 2 01  (z )N (2 z 01 ) = 2 z 01
 z
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 4, APRIL 2003 691

Fig. 2. Stability regions with  = 0:275 and closed-loop poles of the selected PID gains.

Fig. 3. Maximally deadbeat design and arbitrary stabilization.

1 z (z 0 1)D(z) + (K2 z2 + K z + K )N (z) 1 0

2
N ( z
01 ) we have
 (z )
2
 z
01  (z )N (2 z 01 )
0u+j 10u
p
over the circle C
z=
= 0  (u + 1)P (u; )
2
1


2 01 2 01
z  (z )N ( z )
z=
p
0u j 0u
0  (1 0 u )P (u; ) 0 (2K  0 K )
3 2
2 2
2
3

2pu 0 K  P (u; )
+ 1
2
= 0  (u + 1)P (u; )
2 1 3

+ j 1 0 u  P (u; ) 0  (u + 1)P (u; )

1
3 2
0  (1 0 u )P (u; ) 0 (K + K  )
2
3 2 2 1 2
2 0 2

2pu 0 K  P (u; )
1
2
3
+K3 P3 (u; )] :

+ j 1 0 u  P (u; ) 0  (u + 1)P (u; )

2 3 2
1 2
To determine the set of controllers achieving root clustering inside
+(K  0 K )P (u; )
2
2 0 3 a circle of radius , we proceed as before: Fix K3 , use the root
where counting formulas of Section IV, develop linear inequalities in K2 ,
K3 and sweep over the requisite range of K3 . This procedure is then
performed as  decreases until the set of stabilizing PID parameters
P1 (u; ) = RD (u; )RN (u; ) + (1 0 u )TD (u; )TN (u; )
2
just disappears. The following example illustrates this scheme.
P2 (u; ) = RN (u; )TD (u; ) 0 TN (u; )RD (u; ) Example 2: We consider the same plant used in Example 1. Fig. 2
2
P3 (u; ) = RN (u; ) + (1 0 u )TN (u; ):
2 2 (left) shows the stabilizing set in the PID gain space at  = 0:275.
For a smaller value of , the stabilizing region in PID parameter space
By letting disappears. This means that there is no PID controller available to push
all closed loop poles inside a circle of radius smaller than 0.275. From
K3 := K2 
2
0K 0 (19) this we select a point inside the region that is: K0 = 0:0048, K1 =
692 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 4, APRIL 2003

00:3195, K2 = 0:6390, and K3 = 0:0435. From the relationship in [5] D. S. Bernstein and W. M. Haddad, “LQG control with an H per-
(19), we have formance bound: a Riccati equation approach,” IEEE Trans. Automat.
Contr., vol. 34, pp. 293–305, Mar. 1989.
KP 01 022 2 K1 0:3099
[6] T. Iwasaki and R. E. Skelton, “All fixed order H controllers: observer-
based structure and covariance bounds,” IEEE Trans. Automat. Contr.,
KI = 1
T

T + T
1 0 1
T K2 = 0:3243 : vol. 40, pp. 512–516, Mar. 1995.
KD 0
2
 T 0T K3 0:0048
[7] P. Dorato, Analytic Feedback System Design: An Interpolation Ap-
proach. Pacific Grove, CA: Brooks Cole, 2000.
[8] G. Pólya and G. Szegó, Problems and Theorems in Analysis II. New
Fig. 2 (right) shows the closed-loop poles that lie inside the circle of York: Springer-Verlag, 1976.
radius  = 0:275. The roots are 0:2500 6 j 0:1118 and 0:2500 6 [9] M. Mansour, “Robust stability in systems described by rational func-
j 0:0387. tions,” in Control and Dynamic Systems, C. T. Leandes, Ed. New York:
Academic, 1992, vol. 51, pp. 79–128.
To illustrate further, we select several sets of stabilizing PID param- [10] J. F. Delansky and N. K. Bose, “Schur stability and stability domain
eters from the set obtained in Example 1 (i.e.,  = 1) and compare the construction,” Int. J. Control, vol. 49, no. 4, pp. 1175–1183, 1989.
step responses between them. Fig. 3 shows that the maximally dead- [11] N. K. Bose and J. F. Delansky, “Tests for robust Schur interval polyno-
beat design produces nearly deadbeat response. mials,” in Proc. 30th Midwest Symp. Circuits Systems, G. Glasford and
K. Jabbour, Eds., 1988, pp. 1357–1361.
[12] I. Yamada and N. K. Bose, “Algebraic phase unwrapping and zero dis-
VI. MAXIMUM DELAY TOLERANCE DESIGN tribution of polynomial for continuous time systems,” IEEE Trans. Cir-
cuits Syst. I, vol. 49, pp. 298–304, Mar. 2002.
In some control systems an important design parameter is the delay [13] I. Yamada, K. Kurosawa, H. Hasegawa, and K. Sakaniwa, “Algebraic
tolerance of the loop, that is the maximum delay that can be inserted multidimensional phase unwrapping and zero distribution of complex
into the loop without destabilizing it. In digital control a delay of k polynomials—characterization of multivariate stable polynomials,”
sampling instants is represented by z 0k . We use this to determine the
IEEE Trans. Signal Processing, vol. 46, pp. 1639–1664, June 1998.
[14] L. H. Keel and S. P. Bhattacharyya, “Root counting, phase unwrapping,
maximum delay that a control-loop under PID control can be designed stability and stabilization of discrete time systems,” Linear Alg. Ap-
to tolerate. This gives the limit of delay tolerance achievable for the plicat., vol. 351–352, pp. 501–518, 2002.
given plant under PID control. [15] J. Ackermann, Sampled-Data Control Systems: Analysis and Synthesis,
Let the plant be G(z ) = (N (z ))=(D(z )). We consider the problem Robust System Design, ser. Communications and Control Series. New
of finding the maximum delay L3 such that the plant can be stabilized
York: Springer-Verlag, 1985.

by a PID controller. In other words, finding the maximum values of L3

[16] W. M. Wonham, Linear Multivariable Control: A Geometric Ap-
proach. New York: Springer-Verlag, 1985.
such that the stabilizing PID gain set for the plant

z 0L G(z ) = L for L = 0; 1; . . . ; L3
N (z )
; (20)
z D(z )

is not empty. Let Si be the set of PID gains that stabilizes the plant Stability Analysis of Swarms
z 0i G(z ). Then, it is clear that
Veysel Gazi and Kevin M. Passino
\ S
L i
i=0 i stabilizes z G(z ) for all i = 0; 1; . . . ; L: (21)
Abstract—In this note, we specify an “individual-based” contin-
uous-time model for swarm aggregation in -dimensional space and
study its stability properties. We show that the individuals (autonomous
VII. CONCLUDING REMARKS agents or biological creatures) will form a cohesive swarm in a finite time.
Moreover, we obtain an explicit bound on the swarm size, which depends
In this note, we have given a solution to the problem of stabilization only on the parameters of the swarm model.
of a digital control system using PID controllers. The solution is com- Index Terms—Biological systems, multiagent systems, stability analysis,
plete in the sense that a constructive yes or no answer to whether stabi- swarms.
lization is possible, is given and in case it is possible the entire set is de-
termined by solving sets of linear inequalies in two variables obtained
by gridding over the third variable. This approach is akin to the geo- I. INTRODUCTION
metric approach to synthesis and design advocated in [16]. These solu- For a long time, it has been observed that certain living beings tend
tion sets open up the possibility of improved and optimal design using to perform swarming behavior. Examples of swarms include flocks of
PID controllers. The questions of loop shaping, time domain response birds, schools of fish, herds of animals, and colonies of bacteria. It is
shaping, and robust designs are important candidates for research. known that such a cooperative behavior has certain advantages such

REFERENCES
Manuscript received August 3, 2001; revised March 27, 2001 and August
[1] A. Datta, M. T. Ho, and S. P. Bhattacharyya, Structure and Synthesis 23, 2002. Recommended by Associate Editor G. De Nicolao. This work was
of PID Controllers in Advances in Industrial Control. London, U.K.: supported by the DARPA MICA program and was performed at The Ohio State
Springer-Verlag, 2000. University, Columbus.
[2] H. Xu, A. Datta, and S. P. Bhattacharyya, “Computation of all stabilizing V. Gazi was with the Department of Electrical Engineering, The Ohio State
PID gains for digital control systems,” IEEE Trans. Automat. Contr., vol. University, Columbus, OH 43210 USA. He is now withthe Department of Elec-
46, pp. 647–652, Apr. 2001. trical and Electronics Engineering, Atilim University, 06836 Ankara, Turkey.
[3] K. Åström and T. Hägglund, PID Controllers: Theory, Design, and K. M. Passino is with the Department of Electrical Engineering, The Ohio
Tuning. Research Triangle Park, NC: Instrum. Soc. America, 1995. State University, Columbus, OH 43210 USA (e-mail: [email protected]
[4] G. C. Goodwin, S. F. Graebe, and M. E. Salgado, Control System De- state.edu).
sign. Upper Saddle River, NJ: Prentice-Hall, 2001. Digital Object Identifier 10.1109/TAC.2003.809765

0018-9286/03$17.00 © 2003 IEEE