- ~-

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39, NO. 3, MARCH 1994 565

or [5] J. L. mob, “Martingales and one-dimensional diffusion,” Trans. Amer.

Math. Soc., vol. 78, pp. 168-208, 1955.
[61 I. I. Gikhman and A. V. Skorohod, Stochastic Diflerential Equation.
New York: Springer-Verlag, 1972.
[7] L. Arnold, Stochastic Differential Equations: Theory and Applications.
New York Wiley, 1974.
[81 K. Nishioka, “On the stability of two-dimensional linear stochastic
systems,” Kodui Math. Sem. Rep. 27, pp. 211-230, 1976.
[9] F. Kozin, and S. Prodromou, “Necessary and sufficient conditions for
almost sure sample stability of linear It0 equations,” SIAM J. Appl.
Math., vol. 21, no. 1, Nov. 1971.
or [lo] R. R. Mitchell and F. Kozin, “Sample stability of second-order linear
differentialequations with wide band noise coefficients,” SIAM J. Appl.
Math, vol. 22, no. 4, Dec. 1974.
[I11 W. Wedig, “Pitchfork and bifurcationsin stochastic systems-Effective
methods to calculate Lyapunov exponents,” In: P. Kree and W. Wedig,
Eds., Effective Stochastic Analysis. New York: Springer-Verlag, 1989.
c): If ‘Po E [-7r/4, ./2] [I21 F. Kozin and Z. Y. Zhang, “The exact almost sure stability for a
specific class of nonlinear Ito differential equations,” IUTM Symp.,
StuttgdGennany 1989; “Nonlinear Dynamics in Engineering Sys-
tems,” Prof. W. Schiehlen, Ed., New York Springer-Verlag. 1990, pp.
In this case, the stability indicator is 159-164.
[131 Z. Y. Zhang, “New developments in almost sure sample stability of
Q ( 9 ) d t = Q(-n/4) = -1. nonlinear stochastic dynamic systems,” Ph.D. dissertation, Dep. of
Electr. Comp. Sci. Eng., Polytechnic Univ., 1990.
d): If ‘PO E [37r/4, 3 ~ / 2 ] ,cp(t) is recurrent and ergodic. The
function J is evaluated by (5.7). The stable point is shown in Fig.
5at(‘=1.
In summary, it has been shown in this example that the stability
property is dependent on the initial state of the system, which An Iterative Algorithm for Pole
is the same as the nonlinear deterministic counterpart. If p(0) E Placement by Output Feedback
[-a/4, 7r/2] and C > 1,the system will be stable in a s . for practical
engineering situations (2< 20). T. H. Lee, Q. G. Wang, and E. K. Koh

V. CONCLUSIONS Abssfract-In this note, an iterative algorithm is proposed for solving

In this note, K h a s ~ n s k i i ’ theorem
s on stability of linear stochastic the pole assignment by output feedback problem for an nth-order 1-
differential equations has been extended to a class of nonlinear input m-output linear system. The algorithm uses a full rank feedback
stochastic dynamic systems. Based on a modified version of Feller’s gain matrix and requires only linear equations of low dimension to be
solved at each iteration. Application of the proposed algorithm allows
criteria, the sample stability property has been determined exactly, all n closed-loop poles to be almost exactly assigned provided that the
which constructs the first solutions to the almost sure sample stability iterations converge so that deviations of the resulting closed-loop poles
problem for nonlinear stochastic dynamic systems. An application of from the respective desired values become less than some user-specified
the modified criteria to the time-averaging problem will be presented tolerance, and m x 1 2 n. Examples are provided in the note to illustrate
the applicability of the proposed method.
in another work.

ACKNOWLEDGMENT I. INTRODUCTION
The problem of pole assignment by constant gain output feedback
The work described in this note was completed just before the death for a general MIMO system has been the subject of much study.
in April 1990 of Prof. F. Kozin, Dr. Zhang’s Ph.D. advisor. Prof. In [l], Davison showed that if the system is both controllable and
Kozin’s strong support and encouragement, and the many stimulating observable, then m poles of the closed-loop system are assignable
discussions that led to this work, are gratefully acknowledged. Thanks almost arbitrarily by gain output feedback where m is the number
are expressed to Profs. P. Voltz and L. Birenbaum for their help with of independent outputs. Extensions of this result were given by
the manuscript, and to Dr. W. S. Wong and the reviewers for the Davison and Chatterjee in [2] and Sridhar and Lindorff in [3] where
improvement generated by their comments. they essentially showed that under the same conditions as that of
[l], max(l, m) closed-loop poles are assignable almost arbitrarily
REFERENCES by gain output feedback, where I is the number of independent
inputs. Later, these results were generalized in the seminal work of
[l] R. Z. Khasminskii, “Necessary and sufficient conditionsfor the asymp-
-
totic stability of linear stochastic systems,” Th.Prob. Appl., vol. 1, pp. Kimura [4] who showed that under the condition m E - 1 2 n, +
144-147, 1967. where n is the number of state variables, the system is always
[2] -Stochastic Stability of Differential Equations. Sijthoff & Noord- assignable by gain output feedback provided that a slight modification
hoff, 1980.
[31 W. Feller, “The parabolic differential equations and the associated Manuscript received August 19, 1992; revised October 6, 1992.
semigroups of transformations,”Ann. Math., vol. 55, pp. 468-519, 1952. The authors are with the Department of Electrical Engineering, National
[4] -“Diffusion processes in one dimension,” Trans. Am. Math. Soc., University of Singapore, Kent Ridge (051l), Singapore.
vol. 77, pp. 1-31, 1954. IEEE Log Number 9213582.

0018-9286/94$04.00 0 1994 IEEE

Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:26:51 UTC from IEEE Xplore. Restrictions apply.
 566 EEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39, NO. 3, MARCH 1994

of the poles to be assigned is tolerable. However, while these results where e? is the ith row of the ( m x m ) identity matrix, and r,
is
added much to our understanding of the pole placement by output equal to K except for the j t h column which has all 0 elements. Then
feedback problem, because of its nonlinear nature, there has not the closed-loop characteristic polynomial (3) may be rewritten as
been complete computational procedures proposed for solving for the
required feedback gain matrix. There exist some interesting partial p c ( s )= d e t [ s I - A - B r , C - BkJeTC]
solutions to the problem which assume a unity rank dyadic form for = det [SI- A - B r , C ]
the feedback gain matrix K, and thus reduces the problem to one . det [I- ($1- A - BF,C)-' B&,eTC]
of finding a linear solution to K. Unfortunately, such approaches = det[sI - A - B Z J C ]
possess certain inherent limitations because of the assumptions made
regarding the form of K. In [ 5 ] , it is shown that for an nth-order -eT~adj(s~-~ B-Z ~ C ) B ~ ,
+
2-input m-output linear dynamical system, not more than ( m 2 - 1) :=~T,(s) -NJ(S)k3. (4)
poles can be freely assigned under this assumption. Further, when
n, the dimension of the state-space, is such that m + 2 5 n, the In the above, for a given r,,note that (4)is linear in k, so that if
+
remaining ( n - m - 2 1) poles cannot be assigned and may move all the other columns are known (or specified), then the j t h column
can be appropriately computed. Thus (4)provides a basis for the
to undesirable locations. In more recent work, several authors [6]-[9]
have proposed algorithms which retain the use of a dyadic form for construction of a suitable iterative algorithm for pole placement by
K, but provide additional flexibility by assigning some of the poles output feedback. To ensure that (4)can be solved for each j , consider
and using root-locus techniques to find the best locations for the a partition of A into m subsets
unassigned poles. A = {A,}:= { A , i } U { A t 2 } U . . . U { A z m }
In this note, we propose an iterative algorithm for solving the pole
:=R1UA2U***UAm (5)
assignment by output feedback problem where the feedback matrix
to be used is not restricted to be of a unity rank dyadic form; it such that A, will be used for solution of k,. Based on the discussion
is a general full rank matrix K so that no design freedom is lost. above, the following iterative algorithm for pole assignment by static
Only linear equations of low dimension are required to be solved output feedback can now be proposed.
at each iteration, and application of the proposed algorithm allows
all n closed-loop poles to be almost exactly assigned provided that Iterative Algorithm
the iterations converge so that deviations of the resulting closed-loop
Step I : Set K = KO;
poles from the respective desired values become less than some user-
specified tolerance, and m x 1 > n. This is in contrast to existing Step 2: For j = 1, 2 , . 3 . ,m, solve the following system of linear
equations:
solutions utilizing K in unity-rank dyadic form where, regardless of
+
the values of m and 2, not more than (m I - 1) closed-loop poles
may be freely assigned. In the note, examples are also provided for
illustrations.

n. PROBLEM FORMULATION
AND PROPOSED ITERATIVE ALGORITHM
where A i j E A,. For each kj computed, update the previous values
Consider the following nth-order 2-input m-output linear system
with the current computed value;
x = AX + BU +
Step 3: Compute the eigenvalues i i of (A B K C ) , and if
Y=CX
C(A1- <E
i=l
to which a static output feedback
where E > 0 is some specified tolerance, then stop. Otherwise, go
U=KY (2) to step 2.
Remark I : Note especially the use of a full rank gain matrix K
is applied. The problem under consideration is to choose K such
in our proposed approach. This differs from the commonly used
that the closed-loop poles have the specified values {A*}: = A. The
unity-rank K in dyadic form. No design freedom is lost in our
closed-loop characteristic polynomial in this case is
approach.
p c ( s ) = det [SI- A - B K C ] . Remark 2: The expression (4)and the resulting algorithm given
(3)
above has been developed by considering the columns of K. It should
Since there does not exist a general solution to the problem of pole- be obvious that a similar development can be carried out by consider-
placement by output feedback, we consider the construction of an ing the rows of K instead, leading to an iterative algorithm that solves
iterative algorithm for finding the feedback gain matrix K such that for K row-wise. This turns out to be a useful practical observation,
the closed-loop poles are essentially assigned at the desired values. for there are examples which we have attempted where the column-
To achieve the above objective, note, first that a general K can be wise algorithm for solving for K did not attain convergence while
expressed in terms of its column vectors k, as the row-wise algorithm did.
Further, it should be mentioned that the asymptotic values of the
I
convergent sequences of computed feedback gain matrices arising
K =KImx, = xkzeT out of the iterative algorithm are dependent on the initialization used
Z=1
in Step 1; thus not all convergent sequences tend to the solution
of the pole assignment problem (possibly a problem associated with
= + kje; local minima). Step 3 of the algorithm computes the deviation of the
resulting closed-loop poles from the respective desired values and
compares this against some user-specified tolerance. The algorithm

Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:26:51 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39, NO. 3, MARCH 1994 567

terminates with an acceptable value for the feedback gain matrix K solution by fixing certain elements of K. The least squares algorithm
only when the required tolerance is met. The speed of convergence is with singular value decomposition may also be used.
dependent on the initialization and also the system being considered.
It will be seen in the examples provided subsequently that the number
111, ILLUSTRATIVE
EXAMPLES
of steps required can differ widely between two examples with only
small differences in dimensions. Example 1: Consider the 4th-order 2-input 2-output linear system
Remark 3: It is well known that pole assignment by output feed- drawn from [9]:

lo 11
back is a nonlinear problem. This fact is actually evident from the ro1 o 01
development above. The nonlinear nature of the problem makes it
0 0 1 0
difficult to solve, and since the proposed iterative algorithm involves A= 0 0
a system of nonlinear algebraic equations, it is hard to provide 0 0 0 0
general results on solvability. Also, it is worthwhile to note that
most available methods for the solution of nonlinear equations are rl 01
essentially based on computational methods and iterations. In the
light of these observations, the algorithm presented above provides
the means for a very reasonable iterative approach to achieving a
solution. In each iteration, only linear equations of low dimension
have to be solved.
Some observations concerning the choice of A should also be made
here which would serve to illustrate the difficulty of the problem
considered and its nonlinear nature. For a given choice of A, it is
known that there exist some controllable and observable systems
whose poles cannot be assigned exactly to A even when a full rank
feedback gain matrix is used. An example is given in Appendix A
where the solution for K has a singular point if a desired closed-
loop pole is specified at s = -1. The nonlinear nature of the
problem is also evident from the closed-form solution for the feedback
gain matrix given there for that simple example. Thus, in using the
proposed iterative algorithm, it may be necessary to consider a few
different choices of A to avoid such singular points.
Remark 4: The case where m x 1 2 n is especially interesting as
the proposed iterative algorithm provides possibilities not available
with existing solutions which restrict K to unity-rank dyadic form.
In this case, it can be seen that the partition (5) can always be chosen
so that the (T x +matrix

ro 1 0 0 0 0
0 0 1 0 0 0
0 0 0 1 0 0
A=
0 0 0 0 1 0
has T 5 I for all j. Thus for every j = 1, 2 , . ,m, the system of 0 0 0 0 0 1
linear equations (6) is never overdetermined. This implies that under -0 0 0 0 0 0
these conditions, there exist sufficient degrees of freedom for finding

i Hj
a solution to the problem so that if the iterative algorithm converges
and deviations of the resulting closed-loop poles from the respective
desired values become less than the user-specified tolerance, then B = l
essentially all n poles are exactly assigned. This may be compared
with other solutions utilizing K in unity-rank dyadic form where, 0 0 1
regardless of the values of m and I, not more than (m 1 - 1) + 1 0 0 0 0 0
closed-loop poles may be freely assigned.
The condition m x I 2 n mentioned here simply ensures that c= [o 1 0 0 0 01.
there exist sufficient degrees of freedom for finding a solution in the In this case, the desired closed-loop poles were set at -o.5, -2,
proposed iterative algorithm. It is clearly not sufficient as shown by -2.5, -3, -3.5, and -4, The elements of were initially all set
the simple example given in Appendix A. Reference [4] provides an to -40. The solution for K converged after approximately 30000
elegant treatment of conditions ensuring existence of solutions for the
iterations yielding
pole-placement
~- by- output
- feedback problem.
Remark 5: For implementation of the algorithm, a prion" distribu-
tion of A over individual subsets A; is required. If m x 1 2 n but m
does not divide n, then certain Ais may have different numbers of
elements. It may thus happen that (6) is possibly underdetermined. In
this case, since there are redundancies in the elements of the feedback
-3.27253
-8.38933
-75.3207
-12.2223
-92.9134
-290.325 I
and the closed-loop poles are actually assigned at -0.5, -2,
gain matrix K, it is sometimes simpler computationally to attempt a -2.49819, -3.00285, -3.491, and -4.00277.

Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:26:51 UTC from IEEE Xplore. Restrictions apply.
 568 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 39, NO. 3, MARCH 1994

Example 3: In this final example, we consider a 5th-order 3-input For the output feedback with
2-output system (so that m x 1 > n) where U=KY
1 0 0 0 where
[k’
A=k ::: ;I
0 0 0 1 0

0 0 0 0 0
1 0 0
K =
k3
kz]
k4
straightforward algebra shows that the closed-loop characteristic
polynomial is given by:
p c ( s )= det[sl - A - BKC]

B=

c = [o
1; E 1I
0 1

1 0 0 0 0
1 0 0 01.
0

pc(S):=
+
+
= s4 s3(-k1 - k2) s2(-kl - k4)
+
(-k3
+
4 - k 3 - k4 klk4 - k2k3)
+
kik4 - kZk3).
+

If the desired closed-loop characteristic polynomial is given by

S4 + P3S3 + P2S2 +Pis +PO
(7)

Since there are now redundancies in the number of elements in K, then it can be shown that the pole assignment problem has the solution
we attempt a solution with some of the elements predefined and fixed.
(See Remark 5.) The desired closed-loop poles were selected to be
ki = - 0 2 + Pi - Po
-5, -4, -3, and -2 f 2i. The elements of K were initially all set kz = -03 + P z - Pi + Po
to -11, and a row-wise algorithm (see R e m k 2) was used. In this
case, it is necessary to partition A into sets of unequal sizes, and with
the adoption of the row-wise algorithm, the partition k4 = -Pi + Po.
A = (-5) U {-4, -3) U (-2 + 2i, -2 - 2ij Some of the difficulties of the pole assignment by output feedback
problem are evident from this simple example:
was used. In the iterative computations, the (1,2) element was fixed The nonlinearity of the problem is evident from the solution for
and after 40 iterations, K converged to k i , k2, k3, k4.
-5. -11. 1 An admissible solution for k3 requires that kz # - 1, or

-3.00015, -2 f 2i. 0 =Pc(-l)

= (-1)4 + P3(-1)3 + P z ( - ~ )+~ Pi(-1) + PO
IV. CONCLUSIONS
or
In this note, the pole assignment by output feedback problem for
an nth-order I-input m-output linear system has been considered and -P3 + P2 - Pi +Po = -1
an iterative algorithm for solving the problem has been proposed.
The main advantage of the method is its ability to return a full rank which contradicts the requirement above. This means that the desired
feedback gain matrix K that does not lose any degrees of freedom in closed-loop characteristic polynomial cannot have a factor (s 1). +
Such restrictions on the location of the assignable closed-loop poles
the pole assignment design as compared with previous results which
are not obvious from the onset of the problem, and they illustrate the
restricts K to be of a unity rank dyadic form. The algorithm requires
difficulties involved.
only linear equations of low dimension to be solved at each iteration,
and allows all n closed-loop poles to be almost exactly assigned REFERENCES
provided that the iterations converge so that deviations of the resulting
closed-loop poles from the respective desired values become less than [l] E. J. Davison, “On pole assignment in linear systems with incomplete
some user-specified tolerance, and m x 1 2 n. Examples provided in state feedback,”fEEE Trans.Auromaf.Contr., vol. AC-15, pp. 348-351,
June 1970.
the note illustrate the applicability of the proposed method. [2] E. J. Davison and R. Chatterjee, “A note on pole assignment in linear
systems with incomplete state feedback,”IEEE Trans.Automat. Contr.,
APPENDIX vol. AC-16, pp. 98-99, Feb. 1971.
[3] B. Sridhar and P. D. Indorff, “Pole-placement with constant gain output
Consider the 4th-order 2-input 2-output linear system feedback,” fnf.J. Confr., vol. 18, pp. 993-1103, 1973.
yo 1 o 01 [4] H. Kimura, “Pole assignment by gain output feedback,” IEEE Trans.
0 0 1 0 Automat. Confr.,vol. AC-20, pp. 509-516, Aug. 1975.
[5] N.Munro and S. Novin Hirbod, “Pole assignment using full rank output
A = I 0 0 l] feedback compensator,”Int. J. Syst. Sci., vol. 10, pp. 285-306, 1979.

B=

c=
0
1

[h
0

[; ;
:I ::I.
0 0 0
0

1
[6] A. P. Paplinski and M. J. Gibbard, “On constraints in pole assignment
by static feedback,” fnf.J. Confr., vol. 42, pp. 1099-1111, 1985.
[7] C.-L. Chen, T.-C. Yang, aid N. Munro, “Output feedback pole-
assignment procedure,” fnt. J. Conrr., vol. 48,pp. 1503-1518, 1988.
[8] T. Rajgopalan and K. K. Appukuttan, “Pole assignment with constraint
using output feedback,” Aufomatica, vol. 24, pp. 423424, 1988.
[9] K. Kamar and K. K. Appukuttan, “Pole assignment for multi-input
multi-output systems using output feedbacks,” Aufomafica,vol. 27, pp.
1061-1062, 1991.

Authorized licensed use limited to: WASEDA UNIVERSITY LIBRARY. Downloaded on August 28,2023 at 05:26:51 UTC from IEEE Xplore. Restrictions apply.