Chapter 3

Control Theory
1 Introduction
To control an object is to influence its behaviour so as to achieve some desired objective.
The object to be controlled is usually some dynamic process (a process evolving in time), for
example, the UK economy, illness/disease in a patient, air temperature in this lecture theatre.
The controls are inputs to the process which can influence its evolution (for example, UK interest
rates, medication in patient treatment, radiators in a heating system).
The evolution of the process may be monitored by observations or outputs (for example, UK
inflation, medical biopsies, temperature sensors in a heating system).
A central concept in control theory is that of feedback: the use of information available from the
observations or outputs to generate the controls or inputs so as to cause the process to evolve in
some desirable manner.

Input Dynamical process Output

Much of control theory is focussed on the question ? .

We start with a mathematical description of the dynamic process linking the output to the input.

2 Transfer functions
We restrict attention to linear, autonomous, single-input, single-output systems of the form

ẋ(t) = Ax(t) + bu(t), x(0) = ξ, A ∈ Rn×n , b ∈ Rn (1)

T n
y(t) = c x(t) + du(t), c ∈ R , d ∈ R, (2)

where u is a scalar-valued input and y is a scalar-valued output. The variable x(t) is referred to
as the state of the system. By the variation of parameters formula, the solution of (1) is
Z t
t 7→ x(t) = (exp At)ξ + (exp A(t − τ ))bu(τ ) dτ .
0

Therefore, the system output (2) is given by

Z t
t 7→ y(t) = cT (exp At)ξ + cT (exp A(t − τ ))bu(τ ) dτ + du(t).
0

Taking Laplace transform,

ŷ(s) = cT (sI − A)−1 ξ + cT (sI − A)−1 b + d û(s).

With an initial value ξ = 0, we have

ŷ(s) = G(s)û(s), G(s) := cT (sI − A)−1 b + d. (3)

Thus, with zero initial data, the function G relates the (transformed) input û to the (transformed)
output ŷ. G is the transfer function of the system.
Recalling that the Dirac delta function δ has Laplace transform 1 and writing f (t) = cT (exp At)b,
we see that
G(s) = L{f }(s) + dL{δ}(s) = L{f + dδ}(s)
and so G is the Laplace transform of the function

t 7→ g(t) := cT (exp At)b + dδ(t), (4)

and, for ξ = 0, y = g ⋆ u. The function g is referred to as the impulse response of system (1-2),
so called because the output response of the system with ξ = 0 and impulsive input u(t) = δ(t) is
given by
y(t) = (g ⋆ δ)(t) = g(t).
Recalling that an invertible matrix M has inverse given by
1
M −1 = adj M, where adj M is the adjugate of M,
|M |

we have
1
cT adj(sI − A)b + d|sI − A| .


G(s) =
|sI − A|
Since |sI − A| is a polynomial and the entries of adj(sI − A) are also polynomials, we see that the
transfer function G is a rational function, that is a ratio of two polynomial functions:

N (s)
G(s) = , N and D polynomial.
D(s)

Moreover, since d and the entries of A, b and c are real numbers, the polynomials N and D have
real coefficients.
Some notation and terminology:
R[s] denotes the set of polynomials in the complex variable s with real coefficients.
R(s) := {P/Q| P, Q ∈ R[s], Q 6= 0}, the set of rational functions with real coefficients.
A rational function R = P/Q ∈ R(s) is said to be (a) proper if deg(P ) ≤ deg(Q) and (b) strictly
proper if deg(P ) < deg(Q).
Remarks.
(i) R ∈ R(s) is proper if, and only if, lim|s|→∞ |R(s)| < ∞;
(ii) R ∈ R(s) is strictly proper if, and only if, lim|s|→∞ |R(s)| = 0.
Examples.

s2 + 1 s s3 + 7
is proper, is strictly proper, is not proper.
2s2 + 5 3s2 +4 s2 + 2

The transfer function G(s) = cT (sI − A)−1 b + d is such that

 
1 T 1
−1
lim G(s) = lim c I− A b+d =d
|s|→∞ |s|→∞ s s

and so is proper. Moreover, G is strictly proper if, and only if, d = 0.

2.1 Poles and zeros
A number z ∈ C is a zero of R ∈ R(s) if lims→z R(s) = 0. A number p ∈ C is a pole of R ∈ R(s)
if |R(s)| → ∞ as s → p.
Examples.
s−1
(i) If R(s) = , then 1 is a zero and 2 is a pole.
s−2
s−1
(ii) If R(s) = 2 , then
s − 3s + 2
s−1 1
R(s) = =
(s − 1)(s − 2) s−2
and so R has pole 2 and no zeros.
(s − 1)2
(iii) If R(s) = , then
s2 − 3s + 2
(s − 1)2 s−1
R(s) = =
(s − 1)(s − 2) s−2
and so R has zero 1 and and pole 2.
Two polynomials are said to be coprime if they have no common linear factors.
Examples.
(i) P (s) = s − 1 and Q(s) = s2 − 5s + 6 = (s − 2)(s − 3) have no common linear factors and so are
coprime.
(ii) P (s) = s − 1 and Q(s) = s2 − 3s + 2 = (s − 1)(s − 2) have (s − 1) as a common linear factor
and so are not coprime.
Let R = P/Q ∈ R(s). By cancelling all common linear factors in P and Q, we may write R = P̃ /Q̃
where P̃ and Q̃ are coprime. It then follows that (i) z ∈ C is a zero of R if, and only if, z is a root
of P̃ , and (ii) p ∈ C is a pole of R if, and only if, p is a root of Q̃.
Proposition 2.1. Let G ∈ R(s) be a transfer function given by

G(s) = cT (sI − A)−1 b + d, A ∈ Rn×n , b ∈ Rn , c ∈ Rn , d ∈ R .

(i) If p ∈ C is a pole of G, then p is an eigenvalue of A.

(ii) The converse of statement (i) is not true in general.

Proof. (i) Assume p ∈ C is a pole of G. Seeking a contradiction, suppose that p is not an

eigenvalue of A. Then |pI − A| =
6 0 and so
1
cT adj(pI − A)b + d =: q ∈ C .


|pI − A|
it follows that
1
cT adj(sI − A)b + d = |q| < ∞


lim |G(s)| = lim
s→p s→p |sI − A|
which contradicts the fact that p is a pole of G. Therefore, p ∈ spec(A).
(ii) Assume p ∈ spec(A). If c = 0 or b = 0, then G(s) = d which has no poles (and so p ∈ C cannot
be a pole). A less trivial counterexample is the following. Let
   
−1 1 1
A=  , b = c =   , d = 0.
0 2 0
then   
1   s−2 1 1 s−2 1
G(s) = 1 0    = =
(s + 1)(s − 2) 0 s+1 0 (s + 1)(s − 2) s + 1

and so 2 is an eigenvalue of A but not a pole of G.

3 Stability
3.1 Asymptotic stability
Consider the homogeneous system

ẋ(t) = Ax(t), A ∈ Rn×n (5)

Clearly, t 7→ x(t) = 0 is a solution of (5): this solution is referred to as the equilibrium solution.
System (5) is said to be asymptotically stable if, for all ξ, the solution of the initial-value problem
ẋ = Ax, x(0) = ξ, approaches the equilibrium as t → ∞:

x(t) = (exp At)ξ → 0 as t → ∞ .

Recall that every solution of (5) is a linear combination of functions of the form

tk−1
 
λt
t 7→ e vk + tvk−1 + · · · v1 , k = 1, ..., m,
(k − 1)!

where λ ∈ spec(A) with associated generalized eigenvector v (of order m) and

v1 = (A − λI)m−1 v, v2 = (A − λI)m−2 v, · · · , vm−1 = (A − λI)v, vm = v.

It immediately follows that every solution of (5) approaches the equilibrium as t → ∞ if, and only
if, Re (λ) < 0 for all λ ∈ spec(A).
Proposition 3.1. System (5) is asymptotically stable if, and only if, Re (λ) < 0 for all λ ∈
spec(A).

3.2 Bounded-input bounded-output stability

Consider the control system with zero initial state x(0) = 0

ẋ(t) = Ax(t) + bu(t), x(0) = 0, A ∈ Rn×n , b ∈ Rn 

(6)
y(t) = cT x(t) + du(t), c ∈ Rn , d ∈ R



with impulse response t 7→ g(t) := cT (exp At)b(t) + dδ(t). This system is said to be bounded-input,
bounded-output (BIBO) stable if it has the property that the output function y = g ⋆ u is bounded
whenever the input function u is bounded. Furthermore, it can be shown that BIBO stability
admits the following characterization.
Definition. System (6) is BIBO stable if, and only if, there exists a constant γ > 0 such that,
for each bounded input function u, the corresponding output function y = g ⋆ u satisfies

sup |y(t)| ≤ γ sup |u(t)| .

t≥0 t≥0

BIBO stability of (6) is intimately connected with the poles of its transfer function.
Theorem 3.2. System (6) is BIBO stable if, and only if, each pole of the transfer function
G(s) = cT (sI − A)−1 b + d has negative real part.
Proof. For notational convenience, write g0 (t) = cT (exp At)b, in which case the impulse response
is given by g(t) = g0 (t) + dδ(t) (with Laplace Transform G) and y = g ⋆ u = g0 ⋆ u + du.
Necessity: Assume that (6) is BIBO stable. For each m ∈ N, define a function um by
 

 +1, g 0 (m − t) ≥ 0 
, 0≤t≤m


um (t) = −1, g0 (m − t) < 0 


0, t>m



and so 
 |g0 (m − t)|, 0 ≤ t ≤ m

g0 (m − t)um (t) =
 0,

t > m.
Clearly, for all m ∈ N, supt≥0 |um (t)| = 1 and so by BIBO stability, |(g ⋆ um )(m)| ≤ γ. Therefore,
Z m Z m Z m
|g0 (t)| dt = |g0 (m − t)| dt = g0 (m − t)um (t) dt = (g0 ⋆ um )(m)
0 0 0
= (g ⋆ um )(m) − dum (m) ≤ |(g ⋆ um )(m)| + |d| ≤ γ + |d| ∀ m ∈ N.

Therefore, Z ∞
|g0 (t)| dt ≤ γ + |d|.
0

Write G0 = L{g0 }. Note that G = G0 + d and so G and G0 have the same set of poles. Let s ∈ C
be such that Re (s) ≥ 0. Then
Z ∞ Z ∞ Z ∞
∞ > γ + |d| ≥ |g0 (t)| dt ≥ |g0 (t)|e−tRe (s) dt ≥ g0 (t)e−st dt
0 0 0

and so |G0 (s)| ≤ γ + |d| for all s ∈ C with Re (s) ≥ 0. Therefore, G0 has no poles p with Re (p) ≥ 0.
We have now shown that BIBO stability of (6) implies that every pole of its transfer function has
negative real part. It remains to prove the reverse implication.
Sufficiency. If G has no poles, then G = d and the BIBO property holds with γ = |d|. Now,
assume that G has at least one pole. Let pi , i = 1, ..., l denote the distinct poles of G and let mi
denote their multiplicities. Assume that Re (pi ) < 0, i = 1, ..., l. Expanding G0 in partial fractions,
we have
l Xmi
X cij
G0 (s) =
i=1 j=1
(s − pi )j

for some constants cij ∈ C. Note that cij /(s − pi )j is the Laplace Transform of the function gij
given by
cij tj−1 epi t
gij (t) = , i = 1, ..., l, j = 1, ..., mi
(j − 1)!
and so
mi
l X
X
g0 (t) = gij (t) .
i=1 j=1

Since Re (pi ) < 0, i = 1, ..., l, it follows that

Z ∞
|g0 (t)| dt =: L < ∞ .
0
Let u be a bounded input and write U := supt≥0 |u(t)|. The corresponding output satisfies

|y(t)| = |(g ⋆ u)(t)| ≤ |(g0 ⋆ u)(t)| + |du(t)| = |(u ⋆ g0 )(t)| + |d||u(t)|

Z t Z t
≤ |u(t − τ )g0 (τ )| dτ + |d||u(t)| ≤ U |g0 (τ )| dτ + U |d|
0 0
≤ [L + |d|]U ∀ t ≥ 0.

Writing γ := L + |d|, it follows that

sup |y(t)| ≤ γU = γ sup |u(t)|

t≥0 t≥0

and so system (6) is BIBO stable.

An immediate consequence of Proposition 2.1 and Theorem 3.2 is the following.

Proposition 3.3. If Re (λ) < 0 for all λ ∈ spec(A), then (6) is BIBO stable.

3.3 Hurwitz stability criterion

Given a matrix A ∈ Rn×n a basic question is: how do we check in an efficient manner whether
or not the homogeneous system (5) is asymptotically stable. We know that (5) is asymptotically
stable if, and only if, every eigenvalue of A has negative real part. This, in turn, is equivalent to
the requirement that every root of the characteristic equation

|sI − A| = P (s) = sn + an−1 sn−1 + · · · + a1 s + a0 = 0

should have negative real part.

In the context of the control system (6), with (proper rational) transfer function

N (s)
G(s) = = cT (sI − A)−1 b + d, N and D coprime polynomials,
D(s)

an analogous question is: how do we check that (6) is BIBO stable. We know that (6) is BIBO
stable if, and only if, every pole of G has negative real part. This, in turn, is equivalent to the
requirement that every root of the polynomial equation

D(s) = 0

should have negative real part. Therefore, testing for asymptotic stability and BIBO stability
involves a study of the roots of polynomial equations.
Definition: A polynomial P ∈ R[s] is said to be stable if every root of the equation P (s) = 0
has negative real part.
A natural question: can we deduce stability or non-stability of a given polynomial P ∈ R[s] by
means that do not require computation of the roots? For example, the following proposition gives
a necessary condition for stability of P ∈ R[s] (and so, if a given P fails to satisfy the necessary
condition, then it cannot be stable).
Proposition 3.4. If the polynomial P ∈ R[s] given by
n
X
P (s) = am sm = an sn + · · · + a0 , an 6= 0 ,
m=0

is stable, then the coefficients am , m = 0, ..., n, are all non-zero and have the same sign.
The Hurwitz criterion provides a necessary and sufficient condition for stability of P ∈ R[s] in
terms of its coefficients.
Theorem 3.5. Let P ∈ R[s] be given by
n
X
P (s) = am sm = an sn + · · · + a0 , an 6= 0 .
m=0

Without loss of generality, assume an > 0. Let H be the n × n determinant

an−1 an 0 0 ··· 0 0 0
an−3 an−2 an−1 an ··· 0 0 0
an−5 an−4 an−3 an−2 · · · 0 0 0
.. .. .. .. . . .. .. ..
H := |hij | = . . . . . . . .
0 0 0 0 · · · a2 a3 a4
0 0 0 0 · · · a0 a1 a2
0 0 0 0 ··· 0 0 a0

(For 1 ≤ 2i − j ≤ n, the general element satisfies hij = an−(2i−j) , otherwise hij = 0.) Let Hm ,
m = 1, ..., n be the principal subdeterminants of H:

an−1 an 0
an−1 an
H1 = an−1 , H2 = , H3 = an−3 an−2 an−1 , ... , Hn = a0 Hn−1 .
an−3 an−2
an−5 an−4 an−3

The polynomial P is stable if, and only if,

Hm > 0 ∀ m = 1, ..., n.
Example. Determine conditions on the coefficients which are necessary and sufficient for stability
of the cubic polynomial P ∈ R[s]:
P (s) = as3 + bs2 + cs + d, a > 0.
Solution. In this case,
b a 0
H= d c b
0 0 d
with principal subdeterminants

b a
H1 = b, H2 = = bc − ad, H3 = dH2 = d(bc − ad)
d c

Therefore, necessary and sufficient conditions for stability are: b > 0, d > 0 and bc − ad > 0.
Example. For what values of the real parameter k is the following polynomial stable ?
P (s) = s4 + 6s3 + 11s2 + 6s + k .
Solution. In this case,
6 1 0 0
6 11 6 1
H=
0 k 6 11
0 0 0 k
with principal subdeterminants

H1 = 6, H2 = 60, H3 = 36(10 − k), H4 = 36k(10 − k).

Therefore, by Theorem 3.5, the polynomial is stable if, and only if, 0 < k < 10.

3.4 BIBO stabilization by output feedback

Consider system (6), with d = 0 and transfer function G given by

G(s) = cT (sI − A)−1 b.

Suppose that this system is not BIBO stable. A question arises: can the system be rendered BIBO
stable through the use of output feedback?
Write u(t) = v(t) − ky(t), where v is a new input, k ∈ R and −ky(t) is an output feedback
component.

+ û
Input v̂ G Output ŷ
−


Taking Laplace transforms and with reference to the Figure, we have ŷ(s) = G(s)û(s) = G(s) v̂(s)−
kŷ(s) and so the transfer function Gk of the feedback system from new input v̂ to output ŷ is
given by
G(s)
Gk (s) = .
1 + kG(s)
The question now is: does there exist a value k such that the feedback system is BIBO stable?
Theorem 3.6. Assume (i) G has no zero with non-negative real part, and (ii) lim|s|→∞ sG(s) =:
L > 0. Then there exists k ∗ > 0 such that, for all k > k ∗ , Gk is the transfer function of a BIBO
system.

Proof. Write G(s) = N (s)/D(s), where N and D are coprime polynomials. Without loss of
generality, we may assume that D is a monic polynomial, that is, has leading coefficient 1. By
hypothesis (ii), G is strictly proper with deg(D) = deg(N ) + 1. Therefore, for some n ∈ N and real
constants a0 , ..., an , b0 , ..., bn , we have

N (s) = an sn + · · · + a1 s + a0 , an 6= 0, and D(s) = sn+1 + bn sn + · · · + b1 s + b0 .

By hypothesis (ii), it follows that an = L > 0. Consider the rational function

1 s D(s) s an D(s) − sN (s)

R(s) = − = − =
G(s) an N (s) an an N (s)
n+1
+ bn s + · · · + b0 − an sn+1 + an−1 sn + · · · + a0 s
n



an s
=

an an sn + · · · + a0
an bn − an−1 sn + terms of order < n


=
a2n sn + terms of order < n
an bn − an−1
→ < ∞ as |s| → ∞ .
a2n
Therefore, R is proper and its poles coincide with the zeros of G. Write
C+
0 := {s ∈ C| Re (s) ≥ 0} (the closed right-half complex plane).

By hypothesis (i), it follows that R has no pole in C+

0 . We may now conclude that

sup |R(s)| =: k ∗ < ∞ .

s∈C+
0

Let k > k ∗ . The proof is complete if we can show that

G(s) 1 1
Gk (s) = = =
1 + kG(s) (1/G(s)) + k R(s) + k + (s/an )

has no pole with non-negative real part. Seeking a contradiction, suppose that p ∈ C+
0 is a pole of
Gk . Then
p

R(p) + k + = 0 and so p = −an R(p) + k .
an
Now,
|Re (R(p))| ≤ sup |Re (R(s))| ≤ sup |R(s)| = k ∗ < k,
s∈C+
0 s∈C+
0

and so
0 < k − |Re (R(p))| ≤ k + Re (R(p)).
Thus, we arrive at a contradiction:


0 ≤ Re (p) = −an Re (R(p)) + k < 0.
Therefore, for all k > k ∗ , Gk has no pole with non-negative real part and so is the transfer function
of a BIBO stable system.
Example. Consider the case
   
1 0 1 1
   
A = 0 0 1  , b = 0 = c, d = 0.
   
   
1 −1 −1 0

Because of the structure of b and c, in order to determine G(s) = cT (sI − A)−1 b, we need only
compute the top left element E11 (s) of the matrix
 −1
s − 1 0 −1
 
(sI − A)−1 =  0 s −1  .
 
 
−1 1 s + 1

In particular,
s2 + s + 1 s2 + s + 1
E11 (s) = = 3 , and so G(s) = E11 (s).
|sI − A| s −s−1
The hypotheses of Theorem 3.6 hold, and so there exists k ∗ such that, for all k > k ∗ , Gk =
G/(1 + kG) is the transfer function of a BIBO stable system. We can compute k ∗ by appealing
to Theorem 3.2 and the Hurwitz criterion. In particular, if we can determine k ∗ such that the
denominator polynomial of Gk is stable for all k > k ∗ , then, by Theorem 3.2, Gk is the transfer
function of a BIBO stable system for all k > k ∗ . Now,
G(s) s2 + s + 1 s2 + s + 1
Gk (s) = = 3 2
= 3 2
.
1 + kG(s) s − s − 1 + k(s + s + 1) s + ks + (k − 1)s + (k − 1)
By the Hurwitz criterion, the denominator polynomial is stable if, and only if, k > 1 =: k ∗ .
3.5 Integral control: essentially Question 2 of Problem Sheet 11
Consider system (6), with transfer function given by

G(s) = cT (sI − A)−1 b + d

under integral control action Z t  

u(t) = k r(τ ) − y(τ ) dτ (7)
0
where k is a real parameter and r is a reference input. The control objective is to cause the output
y to approach the reference input r in the sense that y(t) − r(t) → 0 as t → ∞. Applying the
Laplace transform to (7), we have

k
û(s) = [ r̂(s) − ŷ(s)] (8)
s
and the overall controlled system (6-7) takes the form

+ k
r̂(s) G(s) ŷ(s)
s
−

Noting that
k  
ŷ(s) =
G(s) r̂(s) − ŷ(s) ,
s
the transfer function Fk from the reference input r̂ to output ŷ is given by
k
ŷ(s) G(s) kG(s)
Fk (s) = = s k = .
r̂(s) 1 + s G(s) s + kG(s)

Now assume
1
G(s) = ,
(s + 1)(s + 2)(s + 3)
in which case, we have
k
Fk (s) = .
+ s4 6s3
+ 11s2 + 6s + k
Now apply the Hurwitz criterion to the denominator polynomial. The Hurwitz determinant is

6 1 0 0
6 11 6 1
H=
0 k 6 11
0 0 0 k

with principal subdeterminants

H1 = 6, H2 = 60, H3 = 36(10 − k), H4 = 36k(10 − k).

Therefore, by the Hurwitz criterion, the polynomial is stable (and Fk is the transfer function of a
BIBO system) if, and only if, 0 < k < k ∗ := 10.
Finally, assume that the reference input is constant: r(t) = r0 for all t ≥ 0. Then r̂(s) = r0 /s.
Let k ∈ (0, k ∗ ) and let fk be such that Fk = L{fk }. Since Fk is the transfer function of a BIBO
stable system, it follows that fk : [0, ∞) → R is exponentially decaying and so, by the final-value
theorem, we may conclude

lim y(t) = lim (fk ⋆ r0 H)(t) = r0 lim (fk ⋆ H)(t) = r0 Fk (0) = r0 ,

t→∞ t→∞ t→∞

and so the control objective is achieved.

The above is a particular example of the following general result.
Theorem 3.7. Assume that (a) G is the transfer function of a BIBO stable system, (b) G(0) > 0,
Then there exists k ∗ > 0 such that, for all k ∈ (0, k ∗ ), the feedback system is BIBO stable. If,
in addition, the reference input is constant r(t) = r0 for all t ≥ 0, then, for all k ∈ (0, k ∗ ),
y(t) → r0 as t → ∞.