2015 IEEE 54th Annual Conference on Decision and Control (CDC)

December 15-18, 2015. Osaka, Japan

Multivariable Controller Synthesis using SISO Design Methods

Daniel N. Mohsenizadeh, Lee H. Keel and Shankar P. Bhattacharyya

Abstract— This paper proposes a new method to design controller designs. This results in the designed controller,
multivariable controllers for linear Multi-Input Multi-Output denoted by Cd , to be of diagonal form. Therefore, many
(MIMO) control systems using the Smith-McMillan form. The robust control methods that are well-suited for SISO systems,
Smith-McMillan form of the transfer function matrix of a
MIMO plant is an equivalent diagonal transfer function matrix such as PID control [9], [10], can be effectively used toward
using which the problem of multivariable controller synthesis MIMO control design. The multivariable controller C, to
can be reduced to multiple Single-Input Single-Output (SISO) be connected to the original MIMO plant, can be obtained
controller designs. If the designed SISO controllers satisfy by transforming back the designed diagonal controller Cd
certain relative degree conditions, then the corresponding through the same unimodular matrices used to calculate the
multivariable controller, to be connected to the MIMO plant,
will be proper. In this paper we show how such multivariable Smith-McMillan form of the plant. The designed controller
controllers can be designed to satisfy closed loop stability and Cd must be such that the multivariable controller C becomes
reference tracking. We also provide some illustrative examples. proper and provides stability and asymptotic tracking. This
imposes a set of constraints on the relative degree of each
element of Cd . We explore these conditions and also provide
I. INTRODUCTION some illustrative design examples.
Many methods have been proposed to design multivariable This paper is organized as follows. In Section II we
controllers for the class of Linear Time-Invariant (LTI) Multi- provide some preliminary material on the calculation of
Input Multi-Output (MIMO) systems. A broad overview of the Smith-McMillan form of a rational matrix. We present
linear multivariable control system design can be found in our main result in Section III. Some illustrative examples
[1], [2], [3]. Matrix Fraction Decomposition (MFD) is a are given in Section IV. Finally, we summarize with our
technique to write a rational matrix describing the dynamics concluding remarks in Section V.
of a MIMO system, known as the transfer function matrix,
as two coprime polynomial matrices [4]. Using MFD allows II. PRELIMINARIES
us to develop state-space realizations of the transfer function
matrix, enabling us to use a wide range of state-space design Let us denote by P (s) the transfer function matrix of a
methods. Many approaches are also provided for frequency linear MIMO plant. Suppose that P (s) is an n × n matrix,
domain design. The generalization of the Nyquist criterion 
p11 (s) . . . p1n (s)

to MIMO system is given in [5]. A variety of techniques on .. .. ..
P (s) =  , (1)
 
this topic are also presented in [6], [7]. Another approach . . .
that is often used to deal with MIMO control problems is to pn1 (s) · · · pnn (s)
decouple the different channels of a MIMO system through
state feedback or output feedback [8]; however, there are where each transfer function pij (s), i, j = 1, 2, , . . . , n is
systems that may not be decoupled in this fashion or even if rational and proper. P (s) can be written as
they are, the controller may be of higher order. 1
In this paper, we present an alternative approach to de- P (s) = N (s), (2)
d(s)
sign multivariable controllers. We use the equivalent Smith-
McMillan form of the transfer function matrix of a linear where d(s) is the least common multiple of the denominators
MIMO plant which itself is a decoupled representation of of all the elements in P (s), and N (s) is a polynomial
the MIMO plant and is a diagonal matrix obtained after per- matrix. Pre-multiplication and post-multiplication of N (s)
forming multiplications by appropriate unimodular matrices. by appropriate choices of unimodular matrices results in an
In such a framework, the original MIMO control design prob- equivalent diagonal polynomial matrix S(s), known as the
lem reduces to multiple Single-Input Single-Output (SISO) Smith form,

Daniel N. Mohsenizadeh is an associate research scientist at the De- S(s) = Yy (s) . . . Y2 (s)Y1 (s) N (s) U1 (s)U2 (s) . . . Uu (s) .
partment of Electrical and Computer Engineering, Texas A&M University, | {z } | {z }
College Station, TX 77843, USA [email protected] Y (s) U (s)
Lee H. Keel is with the Department of Electrical and Computer (3)
Engineering, Tennessee State University, Nashville, TN 37203, USA
[email protected] A unimodular polynomial matrix is a square polynomial ma-
Shankar P. Bhattacharyya is with the Department of Electrical and
Computer Engineering, Texas A&M University, College Station, TX 77843, trix whose inverse is also a polynomial matrix. A necessary
USA [email protected] and sufficient condition for this is that the determinant of

978-1-4799-7886-1/15/$31.00 ©2015 IEEE 2680

the polynomial matrix be a constant. Therefore, S(s) can be Let us denote the Smith-McMillan form of the transfer
written as function matrix P (s) by Pd (s). Thus, we can write
S = diag[01 (s), 02 (s), . . . , 0n (s)], (4) Pd (s) = Y (s)P (s)U (s), (12)
where or
0i (s)|0i+1 (s), i = 1, 2, . . . , n − 1, (5) P (s) = Y −1 (s)Pd (s)U −1 (s), (13)
meaning that each polynomial 0i (s) divides 0i+1 (s) for where Y (s) and U (s) are appropriate unimodular matrices.
i = 1, 2, . . . , n − 1. Dividing each diagonal element of Note that Y −1 (s) and U −1 (s) are unimodular polynomial
S(s), 0i (s), by d(s) and performing all possible cancellations matrices as well. Using (11) and (13) we obtain
yields coprime polynomials i (s) and ψi (s) such that
y(s) = Y −1 (s)Pd (s)U −1 (s)u(s), (14)
i (s) 0 (s)
= i , i = 1, 2, . . . , n, (6) and by pre-multiplying (14) by Y (s),
ψi (s) d(s)
and Y (s)y(s) = Pd (s) U −1 (s)u(s), (15)
| {z } | {z }
yd (s) ud (s)
i (s)|i+1 (s),
where we define
ψi+1 (s)|ψi (s), i = 1, 2, . . . , n − 1. (7)
ud (s) := U −1 (s)u(s), (16)
The Smith-McMillan form of the transfer function matrix
P (s) will be yd (s) := Y (s)y(s). (17)
A diagonal controller Cd (s) is now to be designed for the
 
1 (s) 2 (s) n (s)
Pd (s) = diag , ,..., . (8) equivalent diagonal system Pd (s) as depicted in Fig. 3. Since
ψ1 (s) ψ2 (s) ψn (s)
the Smith-McMillan form Pd (s) is of diagonal form, one can
The poles of P (s) are the roots of the following polynomial, design a SISO controller for each single loop, considering
known as the pole polynomial p(s), each diagonal element of Pd (s), independently. The transfer
p(s) := ψ1 (s)ψ2 (s) · · · ψn (s). (9) function matrix of the designed controller Cd (s) can be
written as
Similarly, the roots of the zero polynomial z(s),
Cd (s) = diag C1d (s), C2d (s), . . . , Cnd (s) ,
 
(18)
z(s) := 1 (s)2 (s) · · · n (s), (10)
where each designed controller Ckd (s), k = 1, 2, . . . , n has
are the zeros of P (s). a relative degree rkd .

III. MAIN RESULT rd ed ud yd

+ Cd(s) Pd(s)
Consider the linear MIMO plant P shown in Fig. 1 where -

u1 y1
u2 y2 Fig. 3. Closed loop system with the designed diagonal controller
P(s)
...
...

un yn
Let us define
Fig. 1. A linear MIMO plant rd (s) := Y (s)r(s). (19)
From the closed loop system in Fig. 3 we have
the inputs and outputs are related by
ed (s) = rd (s) − yd (s), (20)
y(s) = P (s)u(s). (11)
or using (17) and (19),
The control problem is to design a multivariable controller
C(s) so that the closed loop system, shown in Fig. 2, is ed (s) = Y (s)r(s) − Y (s)y(s) = Y (s)(r(s) − y(s))
stable and the output signals track the reference signals. = Y (s)e(s). (21)
The inputs and outputs of Cd are related through
r e u y
+ C(s) P(s)
- ud (s) = Cd (s)ed (s), (22)
which can be written as

Fig. 2. A MIMO closed loop system with a multivariable controller U −1 (s)u(s) = Cd (s)Y (s)e(s), (23)

2681
using the definitions in (16) and (21). Pre-multiplying (23) and its inverse Y −1 (s), which exists and is a unimodular
by U (s) yields polynomial matrix as well, in the following form

u(s) = U (s)Cd (s)Y (s) e(s), (24) Y −1 (s) = L0 + L1 s + · · · . (31)

| {z }
C(s) We now have
where I = Y (s)Y −1 (s)
= (Y0 + Y1 s + · · · )(L0 + L1 s + · · · )
 
C11 (s) ... C1n (s)
C(s) := U (s)Cd (s)Y (s) =  .. .. .. = Y0 L0 + (Y0 L1 + Y1 L0 )s + · · · . (32)
,
 
. . .
Cn1 (s) · · · Cnn (s) Equation (32) is valid if and only if
(25)
Y0 L0 = I,
is a non-diagonal multivariable controller to be connected to Y0 L1 + Y1 L0 = 0,
P (s) (see Fig. 2). From (25) it is clear that
..
. (33)
Cd (s) = U −1 (s)C(s)Y −1 (s). (26)
Therefore, the first condition in (33), Y0 L0 = I, proves that
We introduce the following 4 lemmas leading us toward Y (0) = Y0 is a full rank matrix.
the design of a diagonal controller Cd that guarantees stabil- Lemma 3: The output signal y(t) tracks the reference
ity, asymptotic tracking and properness of the corresponding signal r(t) in the original system (see Fig. 2) if and only
multivariable controller C. if the output yd (t) tracks the reference rd (t) (see Fig. 3).
Lemma 1: The multivariable controller C(s) stabilizes the Proof: If yd (t) tracks rd (t), then
MIMO plant P (s) if and only if the designed diagonal
controller Cd (s) stabilizes the equivalent diagonal plant lim ed (t) = 0, (34)
t→∞
Pd (s). or by the final value theorem, we have
Proof: The necessary and sufficient condition to prove
this is to show that both closed loop systems have the same lim sed (s) = 0. (35)
s→0
characteristic polynomials. The characteristic polynomial of
the closed loop system attained by connecting the designed Substituting ed (s) by (21) gives
diagonal controller Cd (s) to the equivalent diagonal plant lim sY (s)e(s) = 0. (36)
Pd (s) (see Fig. 3) is the numerator of δd (s), s→0

According to Lemma 2, Y (0) is a full rank matrix; thus, (36)

δd (s) := det[I + Pd (s)Cd (s)], (27) is valid if and only if
or using the forms in (8) and (18), lim se(s) = 0, (37)
s→0
n  
Y k (s) d which is
δd (s) = det 1 + Ck (s) , (28)
ψk (s)
k=1 lim e(t) = 0, (38)
t→∞
because I + Pd (s)Cd (s) is diagonal. Using (12) and (26), for the original closed loop system, implying that the output
we can write y(t) tracks the reference r(t). The reverse direction of the
proof follows immediately.
δd (s) = det([I + Pd (s)Cd (s)]
Remark 1: For ydk (t) to track rdk (t), the controller Ckd (s)
= det[I + Y (s)P (s) U (s)U −1 (s) C(s)Y −1 (s)] can be designed as
| {z }
I
αk (s)
= det[I + Y (s)P (s)C(s)Y −1
(s)] Ckd (s) = , (39)
βk (s)γk (s)
−1 −1
= det[Y (s)Y (s) + Y (s)P (s)C(s)Y (s)]
where the characteristic polynomial
−1
= det(Y (s)[I + P (s)C(s)]Y (s))
ψk (s)βk (s)γk (s) + k (s)αk (s), (40)
= det[I + P (s)C(s)] =: δ(s), (29)
is Hurwitz, and βk (s) has roots at the poles of rdk (s), and
where the numerator of δ(s) is the characteristic polynomial the product βk (s)γk (s) has no RHP cancellations with k (s).
of the original closed loop system in Fig. 2.
Lemma 2: If Y (s) is a unimodular polynomial matrix, Let us write the unimodular matrix Y (s), in (12), as
then Y (0) is a full rank matrix. 
y11 (s) . . . y1n (s)

Proof: Let us write Y (s) as .. .. ..
Y (s) =  , (41)
 
. . .
Y (s) = Y0 + Y1 s + · · · , (30) yn1 (s) · · · ynn (s)

2682
where ykj (s), k, j = 1, 2, . . . , n is a polynomial of degree The least common multiple of the denominators of all the
dykj . Similarly, each entry of U (s) in (12) is a polynomial, elements in P (s) is
uik (s), i, k = 1, 2, . . . , n of degree duik .
The following lemma characterizes the required relative d(s) = (s + 1)(s + 2). (46)
degree for each designed SISO controller Ckd (s), k = Thus, we can write P (s) as
1, 2, . . . , n, so that the corresponding multivariable controller  
C(s) is proper. 1 4 −s − 2
P (s) = −1 . (47)
Lemma 4: If the relative degrees rkd of the designed SISO (s + 1)(s + 2) 2s + 4 2
controllers Ckd (s), k = 1, 2, . . . , n, satisfy
| {z }
N (s)

min {rkd − duik − dykj } ≥ 0, The Smith form of N (s) can be obtained by multiplying
k=1,2,...,n
N (s) by the following unimodular matrices Y (s) and U (s),
∀i, j = 1, 2, . . . , n, (42)
1
1 s+2
   
4 0 4 −s − 2 4
where duik and dykj are the degree of polynomials in the S(s) = −1
−s − 2 2 2s + 4 0 1
unimodular matrices U (s) and Y (s), respectively, then the | {z }| {z 2 }| {z }
Y (s) N (s) U (s)
corresponding multivariable controller C(s) will be proper.  
For any k that uik (s)Ckd (s)ykj (s) = 0, then the correspond- 1 0
= . (48)
ing rkd − duik − dykj term in (42) should be neglected. 0 s2 + 4s + 3
Proof: Performing the matrix multiplication For the above unimodular matrices, we have du11 = du21 =
U (s)Cd (s)Y (s), given in (25), in terms of matrices’ du22 = dy11 = dy12 = dy22 = 0 and du12 = dy21 = 1.
elements gives the (i, j) element of C(s) as Dividing each element of S(s), in (48), by d(s), in (46),
n
X and performing all possible cancellations gives the Smith-
Cij (s) = uik (s)Ckd (s)ykj (s), (43) McMillan form of P (s) as
k=1  1 
(s+1)(s+2) 0
c
which has at least the relative degree rij , Pd (s) = s+3 . (49)
0 s+2
c
rij = min {rkd − duik − dykj }. (44) Now, two SISO controllers C1d (s) and C2d (s) can be designed
k=1,2,...,n
for two single loops corresponding to the diagonal elements
If for a specific k, uik (s)Ckd (s)ykj (s) = 0, then the corre- of Pd (s), to satisfy closed loop stability and reference signal
sponding rkd − duik − dykj term in (44) needs to be neglected. tracking. Let us consider the following controller
c
C(s) will be proper if rij ≥ 0, ∀i, j = 1, 2, . . . , n. "
3
#
s(s+4) 0
Now, based on Lemmas 1 to 4, we can present our main Cd (s) = 1 , (50)
0
theorem. 2s(s+1)

Theorem 1: Given a rational proper transfer function ma-

where both relative degrees r1d and r2d are 2. Based on the
trix P (s), of a linear MIMO plant, if a diagonal controller
statement of Lemma 4, if r1d and r2d satisfy the following set
Cd (s) is designed such that it stabilizes the equivalent Smith-
of inequalities,
McMillan form of P (s), denoted by Pd (s), and the output
yd (t) tracks the reference rd (t), and Cd (s) satisfies the rela- min {rkd − du1k − dyk1 } ≥ 0,
k=1,2
tive degree conditions given in (42), then the corresponding
multivariable controller C(s) will be proper and stabilize min {rkd − du1k − dyk2 } ≥ 0,
k=1,2
P (s), and the output y(t) will track the reference r(t).
min {rkd − du2k − dyk1 } ≥ 0,
Proof: The proof follows immediately from Lemmas 1 k=1,2
to 4. min {rkd − du2k − dyk2 } ≥ 0, (51)
k=1,2

IV. ILLUSTRATIVE EXAMPLES then the multivariable controller C(s) will be proper. Ex-
panding the terms in (51) and substituting the values for
Example 1: A Two-Input Two-Output (TITO) Stable Plant du , dy and rd , it can be easily verified that the above set of
Consider the following transfer function matrix inequalities are satisfied. The transfer function matrix C(s)
"
4
#
−1 is
(s+1)(s+2) s+1
P (s) = , (45) 3 2
" #
2 −1 −(s +8s +14s+10) s+2
s+1 2(s+1)(s+2) 8s(s+1)(s+4) 4s(s+1)
C(s) = U (s)Cd (s)Y (s) = −(s+2) 1
,
2s(s+1) s(s+1)
representing a TITO stable plant and suppose that a mul-
(52)
tivariable controller is to be designed so that the closed
loop system is stable and the output signals track unit step which is proper. Connecting C(s) to P (s) and closing the
reference signals. loop, the closed loop system transfer function matrix H(s)

2683
 r1(t) r2(t) P (s) will be
2 2 " 1
#
(s2 −1) 0
1 1 Pd (s) = 3s2 +2s+3 . (57)
0 3(s2 −1)
y1(t)

y1(t)
0 0
Now, two SISO stabilizing controllers C1d (s) and C2d (s)
−1 −1 should be designed for the diagonal elements of Pd (s) in (57)
0 10 20 30 0 10 20 30 that also guarantee unit step tracking. Moreover, the relative
Time (seconds) Time (seconds)
degrees of these controllers need to satisfy the inequality
r1(t) r2(t) conditions given in the statement of Lemma 4. Consider the
2 2 following controller
" 2 #
5s +5s+1
1 1
s(.1s+1) 0
Cd (s) = , (58)
y2(t)

y2(t)

5
0 0 0 s(.1s+1)

−1 −1 where the relative degrees of C1d (s) and C2d (s) meet the in-
0 10 20 30 0 10 20 30 equality conditions in Lemma 4. The multivariable controller
Time (seconds) Time (seconds)
C(s) becomes
" 1.05s+.336 .415s+.166 #
Fig. 4. Step response of the closed loop system after connecting C(s) in s(.1s+1) s(.1s+1)
(52) to P (s) in (45) (solid line), and the step response of the open loop C(s) = −1.67(s−1) 3.33(s+1) , (59)
plant P (s) in (45) (dashed line). s(.1s+1) s(.1s+1)

which is strictly proper, and the closed loop system transfer

function matrix H(s) will be
becomes " h (s) h (s) #
11 12

H(s) = [I + P (s)C(s)]−1 P (s)C(s) H(s) = g(s)

h21 (s)
g(s)
h22 (s) , (60)
g(s) g(s)
" 3
#
s4 +7s3 +14s2 +8s+3 0
= −s(s+2)(s4 +10s3 +29s2 +32s+12) s+3 . where
2(s4 +7s3 +14s2 +8s+3)(2s3 +6s2 +5s+3) 2s3 +6s2 +5s+3
(53) g(s) = s8 + 20s7 + 198s6 + 1043s5 + 3094s4
The response of the closed loop system to unit steps is plotted + 3703s3 + 3873s2 + 2233s + 500,
in Fig. 4 and verifies tracking of step inputs. h11 (s) = 58s6 + 605s5 + 2688s4 + 3862s3
Example 2: A Two-Input Two-Output (TITO) Unstable Plant + 4420s2 + 1533s + 500,
Consider the following TITO unstable transfer function ma- h12 (s) = − s(17s5 + 143s4 − 290s3
trix − 543s2 + 273s + 400),
 4 −1 
s−1 s+1 h21 (s) = s(4s5 + 28s4 − 136s3
P (s) = 2 1 . (54)
s+1 (s−1) + 73s2 + 132s − 100),
Our objective is to design a multivariable controller C(s) to h22 (s) = 42s6 + 478s5 + 3105s4 + 3988s3
stabilize the closed loop system and also make the outputs + 4020s2 + 2533s + 500. (61)
track unit steps. Here, we have d(s) = (s + 1)(s − 1). Thus,
P (s) can be written as The response of this closed loop system to unit steps is
  plotted in Fig. 5. Since H(s)|s=0 = I, the closed loop system
1 4s + 4 −s + 1 tracks steps.
P (s) = , (55)
(s + 1)(s − 1) 2s − 2 s + 1
| {z }
N (s) V. CONCLUDING REMARKS
In this paper we presented an alternative approach to
and the Smith form of N (s) can be calculated as follows
design multivariable controllers for LTI MIMO systems
−1
 1
1 3s+1
  
8 4 4s + 4 −s + 1 8
using the Smith-McMillan form. We showed that the MIMO
S(s) = −s+1 2s+2 control design task can be accomplished through multiple
3 2s − 2 s + 1 0 1
| {z 3 } | {z }| {z } SISO controller designs which enables the designer to use
Y (s) N (s) U (s)
  the full power of classical techniques developed for SISO
1 0
= . (56) systems. We also determined conditions on the relative
0 s2 + 23 s + 1 degrees of the designed SISO controllers to guarantee the
u u u y y properness of the resulting multivariable controller as well
For this example, we have d11 = d21 = d22 = d11 = d12 =
u y y as stability and reference tracking.
0 and d12 = d21 = d22 = 1. The Smith-McMillan form of

2684
 r1(t) r2(t)
1.5 0.5

1
y1(t)

y1(t)
0
0.5

0 −0.5
0 10 20 30 0 10 20 30 40
Time (seconds) Time (seconds)

r1(t) r2(t)
0.2 1.5

0.1
1
y2(t)

y2(t)

0
0.5
−0.1

−0.2 0
0 10 20 30 40 0 10 20 30
Time (seconds) Time (seconds)

Fig. 5. Step response of the closed loop system after connecting C(s) in
(59) to P (s) in (54).

VI. ACKNOWLEDGEMENT
The work of L.H. Keel was supported in part by DoD
grant W911NF-08-0514.

R EFERENCES
[1] W. A. Wolovich, Linear Multivariable Systems. New York: Springer-
Verlag, 1974.
[2] T. Kailath, Linear Systems. Prentice-Hall, Englewood Cliffs, NJ,
1980.
[3] A. I. G. Vardulakis, Linear Multivariable Control: Algebraic Analysis
and Synthesis Methods. Chichester, Sussex, UK: John Wiley and
Sons Ltd., 1991.
[4] M. Vidyasagar, Control System Synthesis: A Factorization Approach,
ser. Synthesis Lectures on Control and Mechatronics. Morgan and
Claypool Publishers, 2011.
[5] I. Postlethwaite and A. G. J. MacFarlane, A Complex Variable Ap-
proach to the Analysis of Linear Multivariable Feedback Systems, ser.
Lecture Notes in Control and Information Sciences. Berlin: Springer-
Verlag, 1979, vol. 12.
[6] J. M. Maciejowski, Multivariable Feedback Design. Wokingham,
UK: Addison-Wesley Publishing Company, 1989.
[7] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control:
Analysis and Design. Chichester, Sussex, UK: John Wiley and Sons
Ltd., 1995.
[8] W. M. Wonham, Linear Multivariable Control: A Geometric Ap-
proach. New York: Springer-Verlag, 1979.
[9] S. P. Bhattacharyya, A. Datta, and L. H. Keel, Linear Control Theory:
Structure, Robustness, and Optimization. CRC Press, Boca Raton,
FL, 2009.
[10] L. H. Keel and S. P. Bhattacharyya, “A bode plot characterization of
all stabilizing controllers,” IEEE Transactions on Automatic Control,
vol. 55, no. 11, pp. 2650–2654, Nov 2010.

2685