Chapter 4

Transfer Function and Stability

There is nothing permanent except change.

Heraclitus

We use what was learned in Chapter 3, to keep an input - output representation of an

LTI system in the Laplace domain. Along the way we will discover that this is a great
tool for system analysis. We will discuss some typical problems that can be attacked. One
is the computation of the (transient) response of a system given a particular input. In
contrast, network analysis with frequency functions only gives you the steady state behavior
for harmonic functions, i.e., sinusoid in, sinusoid out. The inverse problem may also be of
interest, given inputs and corresponding outputs, what can be said about the behavior of
the system? In many instances we may not be interested in the actual form of the response,
but only care if the response will converge to zero in the long run (die out), or if the output
grows without bound, or in a critical case, will keep oscillating like a frictionless pendulum.
In these cases we are only interested in the long term behavior. What can the Laplace
transform tell us about this? This is the all-important system stability problem. Finally,
we may also be interested in sensitivity problems. How drastically does the input - output
behavior of a system change when we change some of its parameters. After a brief discussion
of these problems, we investigate first and second order systems in detail. There is no need
to go to higher orders, as the previous two constitute ‘typical’ behaviors. We will discover
that higher order systems behave as superpositions of these simple blocks. With this we
will be in a good position to characterize the long term behavior of a system. We will see
that this depends on the location of the roots of some polynomial in some broader sense.
Stability analysis does not require the precise knowledge of the roots, and that is a good
thing, because finding roots of a higher degree polynomial is not a simple task. We will

82
Erik I. Verriest Chapter 4: Transfer Function and Stability 83

provide an elegant algorithm for stability testing in general. The proof of the method is
quite simple, however we will defer this until we have some more mathematical background.

4.1 Finite Dimensional LTI-System: Transfer Function

A continuous time LTI System is called finite dimensional if its input-output behavior can
be described in the time domain by an LTI-ODE, say
a(D)y = b(D)u
The system is said to have order n if the degree of the monic polynomial a(·) is n. Usually
one assumes that the input is applied at t = 0, meaning that Π− u = 0. Recall that if the
system is initially at rest, then y and all its derivative are zero at t = 0− . The fact that we
have to take 0− stems from the fact that the output may not be smooth (it or its derivatives
may have jumps at t = 0) and we want to define the natural initial conditions as those before
any input, including impulsive inputs, are applied. With the above two assumptions, the
(one sided) Laplace transform of the LTI-ODE gives
a(s)Y (s) = b(s)U(s),
from which
b(s)
Y (s) = U(s).
a(s)
b(s)
The function H(s) = a(s) is known as the transfer function of the system. Note that for a sys-
tem described by an LTI-ODE, this is necessarily a rational function of s, i.e., the quotient of
two polynomials. The roots (if any) of the numerator polynomial, b(s), are called the zeros of
the system. The roots of the denominator polynomial,a(s) are called the poles of the system.

Let the system now be probed by an impulse. The response of a system that is initially
at rest, to the impulsive input, is appropriately called the impulse response of the system.
Since the Laplace transform of the impulse is L− [δ] = 1, the impulse response has Laplace
transform equal to H(s).
Consequently, we get:
[transfer function] = L [impulse response].
Whereas the differential equation a(D)y = b(D)u gives a destructive or implicit description
of the system, its Laplace transform, Y (s) = H(s)U(s) is an explicit representation of the
system, but in the Laplace (or s-) domain.
Erik I. Verriest Chapter 4: Transfer Function and Stability 84

4.2 Basic System Analysis Problems

1. Response of a system with known causal impulse response to a given input.
Since in the time domain, the response of an LTI system that is initially at rest to an
input is given by the convolution y = h ∗ u, it follows at once that
σs L(h ∗ u) = H(s)U(s)
provided that the regions of convergence of U and H overlap. The region of convergence
for the output is then the intersection of RoC(U) and RoC(H). Thus the Laplace
transform maps convolution to the usual product. This is quite some simplification.
2. Response of a system with known input - output ODE to a given input and
initial condition.
Of particular importance is the step response. The unilateral Laplace transform is here
the one of interest. (Why?)
3. Stability
Perhaps we are not interested in the detailed behavior of the system in finite time as
above, but are only interested to know if ultimately the response of the system to arbi-
trary initial conditions will diverge (grow without bound), converge to an equilibrium
or display sustained oscillations, or if the system will have bounded output if also the
input remains bounded.
4. System identification.
Given an (observed) input - output pair for a system, can one deduce the dynamical
equations for the system?
5. Sensitivity Analysis.
How do parameters variations influence the behavior of a system?

4.3 Transient Response

Consider an LTI system with input-output behavior characterized by the implicit equation
a(D)y = b(D)u
Suppose the system is initially at rest, i.e. at 0− y and all its derivatives are zero. Assume
that at time 0, a unit step is applied, i.e., u = u−1 . We find by Laplace transforming:
1
a(s)Y (s) = b(s)
s
Erik I. Verriest Chapter 4: Transfer Function and Stability 85

Hence,
b(s) H(s)
Y (s) = = ,
sa(s) s
where H(s) is known as the transfer function of the system. Consider the following example:
s−a
H(s) = s2 +3s+2 . Using the partial fraction expansion, we find
s−a −a/2 a + 1 −(a + 2)/2
Y (s) = = + + .
s(s + 1)(s + 2) s s+1 s+2
The inverse transform yields
 
a −t a + 2 −2t
y(t) = − + (a + 1)e − e u−1 (t).
2 2
Note that the limit for t → ∞ is −a/2. The final value theorem, which may be applied
here (Why?) gives
 
a s s
y(∞) = lim[sY (s)] = lim − + (a + 1) − (a + 2) = −a/2,
s→0 s→0 2 s+1 s+2
which is consistent.

If the same system is initially at rest but an impulse is applied at time t = 0 instead,
then its response is:
s−a −(a + 1) a + 2
Y (s) = H(s) = = + .
(s + 1)(s + 2) s+1 s+2
In the time domain, the transient response is
 
−t a + 2 −2t
y(t) = −(a + 1)e + e u−1 (t).
2
Note that if we take the derivative of the response to the step, then
 ′  
′ a −t a + 2 −2t a −t a + 2 −2t
ystep (t) = − + (a + 1)e − e u−1 (t) + − + (a + 1)e − e u0 (t)
2 2 2 2
By properties of generalized functions, this is
 
′
 −t −2t
 a a+2
ystep (t) = −(a + 1)e + (a + 2)e u−1 (t) + − + (a + 1) − u0 (t)
2 2
which is exactly the response to the impulse. Is this fact surprising?
Erik I. Verriest Chapter 4: Transfer Function and Stability 86

4.4 First and Second Order Systems

The reason why we consider these simple systems in detail is because they already charac-
terize the behavior of more complicated systems by reductionism. They are “prototypical”.
The behavior of more general systems can be reduced to a comination of first and second
order systems.‘ In this section we introduce the notions of time constant, natural frequency,
and damping ratio.

4.4.1 First order systems

K
The generic transfer function is simply H(s) = s+p . This function has a pole at s = −p. By
inverse Laplace transforming, we see that the impulse response is
h(t) = Ke−pt u−1 (t).
If p > 0, the pole lies on the negative real axis, and the impulse response decays exponen-
tially with time constant τ = 1/p. Note that for t = τ , the exponential decayed 63.2% of its
initial value. For t = 4τ , and t = 5τ respectively only about 1 percent and 1 promile is left
of the initial value, K.
To make a good quick sketch of an exponential, it helps to realize that the tangent at t = 0
intersects the time axis at t = τ . (Prove this fact).

If p < 0, the pole lies on the positive real axis, and the impulse response grows without
bound. The border case p = 0 gives for the impulse response Ku−1 (t), i.e. the step. This is
not surprising as the system is now simply an integrator.

How does a first order system respond when a unit step is applied at the input. We
obtain, for the step response in the Laplace domain
K 1 K/p K/p
Y (s) = = − .
s+ps s s+p
The corresponding time domain function is
K
y(t) = (1 − e−pt )u−1(t).
p
For positive p this means that y(t) approaches the value K/p asymptotically. Indeed since
RoC is here Re s > −p and thus zero belongs to the RoC, which implies that the final
theorem applies. The latter gives
K K
y(∞) = lim s = .
s→0 (s + p)s p
Erik I. Verriest Chapter 4: Transfer Function and Stability 87

We say that K/p is the steady state gain of the system. If p < 0, the final value theorem
does not apply. In fact we see that the step response grows without bound.

4.4.2 Second order systems

Consider the system
A
H(s) = .
s2 + a1 s + a2
If the denominator can be factored as s2 + a1 s + a2 = (s + r1 )(s + r2 ), with r1 and r2 both
real and disjoint then this system is a parallel combination of the two first order systems (as
seen by a simple partial fraction expansion)
   
A 1 A 1
H(s) = + .
r2 − r1 s + r1 r1 − r2 s + r2
Hence the impulse response is the sum of two decaying exponentials if r1 and r2 are both
positive.
If at least one of r1 or r2 is negative, the impulse response will exhibit an exponential growth.

If r1 = r2 = r, which happens when a21 = 4a2 , then this impulse response is h(t) = Ate−rt .
If r > 0, the response decreases to zero for t → ∞, but for small values of t (how small?) it
will rise approximately linearly with t. The response is extremal when (te−rt )′ = 0. This is
e−r1 t − r1 te−r1 t = 0, or t = 1/r1 = τ , the time constant. Note that the extremal magnitude
is |h(τ )| = |A|τ
e
. If r ≤ 0, the impulse response will grow without bound.

If a2 > 0, it helps to define

√
ωn = a2
a1
ζ = √
2 a2
A
K = .
a2
The parameters K, ζ and ωn (gain, damping and natural frequency) are characteristic for the
behavior. Indeed, from
Kωn2 Kωn2
H(s) = =
s2 + 2ζωn s + ωn2 (s + ζωn )2 + ωn2 (1 − ζ 2 )
we identify the following cases for this standard form:
Erik I. Verriest Chapter 4: Transfer Function and Stability 88

Case 1: ζ = 0. We directly recognize in

ωn
H(s) = Kωn
s2 + ωn2
a scaled version of the transform of sin ωn t. The impulse response in this case is the sustained
oscillation
h(t) = Kωn sin ωn t.
Case 2: 0 < ζ < 1 p
The denominator has two complex conjugate roots: s = −ζωn ± jωn 1 − ζ 2 . The inverse
Laplace transform is
Kωn p
h(t) = p e−ζωn t sin ωn 1 − ζ 2 t
1 − ζ2
which is oscillatory with decaying
p amplitude. Note that the frequency is not the natural
frequency ωn , but rather ωd = ωn 1 − ζ 2, called the damped natural frequency. We say that
the response is underdamped.

Case 3: ζ = 1
This is the case of the double root, and corresponds to

h(t) = Kωn2 te−ωn t .

This response type is called critically damped.

Case 4: ζ > 1   
p p
2 2
In this case the denominator factors as s + ωn (ζ + ζ − 1) s + ωn (ζ − ζ − 1) . Its
roots are real and disjoint, both of them being negative. The response is a combination of
two decaying exponentials, and can be written as
Kωn p
h(t) = p e−ωn t sinh(ωn ζ 2 − 1 t).
ζ2 − 1

This response type is called overdamped.

Case 5: ζ < 0
In this case the impulse response (or its envelope) grows exponentially fast.

Exercise: Compute for the system with transfer function

Kω 2
H(s) =
s2 + 2ζω + ω 2
Erik I. Verriest Chapter 4: Transfer Function and Stability 89

the response to a unit step in the underdamped, critically damped, and overdamped case.

Problem:
A system has impulse response: h(t) = Ke−αt sin ωt.
Determine its transfer function and determine its (unit) step response.
Determine the steady state gain, peak time and maximal overshoot if K = 1, α = 12 , and
ω = 2.
Determine a formula for the envelope of the step response.

4.5 Stability
In this section stability is discussed. There are essentially two types of stability: Stability
w.r.t. initial conditions, usually referred to as Lyapunov stability, and bounded input, bounded
output (BIBO) stability. The first is the ability of the system to return to a nominal behavior
(usually the equilibrium x = 0 with u = 0), once such behavior has been disturbed. The
second is the property that the response of the system remains bounded for all possible
inputs that are bounded. For linear systems, the two notions of stability are related. In
order to analyze Lyapunov stability, we first define the modes of a system, and characterize
their stability. Then we define the (Lyapunov) stability of an entire system. A useful
criterion is given (without proof) that allows to decide whether a system, given by its transfer
function, is stable or not. We discuss how such a criterion can be manipulated to obtain
more quantitative information about the speed of the transients of a system. Finally we
treat the BIBO stability.

4.5.1 Stable and unstable modes

Suppose a system has a rational transfer function H(s) = b(s)/a(s). We know then that the
input - output behavior of the system is characterized by the implicit ODE

a(D)y = b(D)u.

It follows that if the system has arbitrary initial conditions specified at t = 0− , then the
output function has the unilateral Laplace transform of the form

b(s) c(s)
Y (s) = U(s) + ,
a(s) a(s)
Erik I. Verriest Chapter 4: Transfer Function and Stability 90

for some polynomial c(·) of degree strictly less than the degree of a(·). (Why?) Suppose,
that in addition the denominator polynomial a(s) has all distinct roots, pi ; i = 1, . . . , n. Note
that, since a(s) has real coefficients, these roots are either real, or form complex conjugate
pairs. The zeros of the denominator polynomial of a rational function are called poles. (The
name refers to the fact that the graph of the modulus, |H(s)|, looks like a pole erected above
pi in the complex plane. If H(s) is strictly proper, then the partial fraction expansion of the
transfer function is: n
X Ai
H(s) = ,
i=1
s − pi
where the coefficients may be computed by solving the n equations in n unknowns, obtained
once one puts the above right hand side on a common denominator, or one can use the
residue-method. The latter method is a lot simpler if n is large. The coefficients are directly
obtained as
Ai = [H(s)(s − pi )]s=pi
Let’s try to understand this: H(s) is singular, i.e., becomes unbounded, at the pole pi .
Therefore by multiplying with the factor (s − pi ), one can ‘tame’ the singularity. Indeed, the
resulting function H(s)(s − pi ) is now well behaved in the neighborhood of pi . What remains
after this singularity is so removed, is appropriately called the ‘residue’, and is therefore
obtained as above. Now this only explains why Ai is called the residue, and not why it is
the correct coefficient in the partial fraction expansion. For this, substitute the exact partial
fraction expansion in the formula for the residue:
" n # n    
X Aj X s − pi s − pi
[H(s)(s − pi )]s=pi = (s − pi ) = Aj + Ai = Ai .
j=1
s − p j
j6=i
s − p j s=p i
s − p i s=p i
s=pi

Now the reason why we want to take a partial fraction expansion is that it gives a
Ai
simple additive decomposition for H(s), and each term of the form s−p i
is known to have
pi t
as inverse Laplace transform Ai e . Additivity is preserved in the time domain, so that the
corresponding impulse response is
n
X
h(t) = Ai epi t .
i=1

Likewise the strictly proper rational part due to the initial conditions may also be expanded
in a partial fractions. This also contributes terms of the form
n
X Ci
,
i=1
s − pi
Erik I. Verriest Chapter 4: Transfer Function and Stability 91

where the Ci depend on the initial conditions.

As the ultimate behavior (convergence, divergence, oscillation) of this function is deter-
mined or characterized by the poles, which in turn are obtained as the zeros of a(s), one
refers to a(s) as the characteristic polynomial, and to the equation

a(s) = 0

as the characteristic equation. Each of the elementary terms is called a mode of the system.
In summary then, the modes are determined by the poles and vice versa.
Our analysis is not quite complete, since we assumed that all roots of the characteristic
equation were disjoint. If this is not the case, a strictly proper rational transfer function will
have a partial fraction expansion, involving powers of the polar factors, 1/(s − pi ). More
precisely, if pi appears with multiplicity mi > 1, then there will terms of the form
1 1 1
, , ... , .
(s − pi ) (s − pi )2 (s − pi )m

From the Laplace transform theory, we know that the corresponding time functions are

t pi t tm−1 pi t
epi t , e , ... , e .
1! (m − 1)!

Consider the system with characteristic polynomial a(s). Its roots are the poles of the
system, and determine for instance the partial fraction expansion of the transfer function.
Now a pole at p is associated with a mode ept of the system. Hence if the real part of p is
positive, this function will grow without bound, while if ℜp < 0, the function converges to
0. In the first case we have instability, in the second asymptotic stability. If this pole occurs
with higher multiplicity, m, we get to multiply the corresponding exponential by a suitable
polynomial of degree m − 1 in t to get the new modes. However, the dominant behavior
is still captured by the exponential factor. Instability or asymptotic stability are thus still
determined by the sign of the real part of the pole.

The situation is different for poles on the imaginary axis. Simple poles correspond with
pure oscillation (a pole at 0 corresponds with the limit case: frequency 0, i.e., “DC”). Mul-
tiple poles on the imaginary axis correspond to [polynomial × oscillation], which always
behaves in a divergent way. Consequently, simple poles on the imaginary axis correspond to
marginally stable modes, whereas poles of higher multiplicity on the imaginary axis corre-
spond to unstable modes.
Erik I. Verriest Chapter 4: Transfer Function and Stability 92

4.5.2 Stable and Unstable Systems

A system is asymptotically stable if all its modes are asymptotically stable. A system is
unstable if it has at least one unstable mode (equivalently, its transfer function has a pole
in the right half plane, or a pole of higher multiplicity on the imaginary axis). A system is
marginally stable if all its poles are in the closed left half plane, and those on the imaginary
axis have multiplicity one.

Obviously, if one were to know the exact location of all the roots of the characteristic
polynomial of a system, stability or the lack of it could be decided. Now for polynomials of
second degree, the explicit formula for its roots is probably remembered by everyone. Sim-
ilar formulas exist (Cardano) for the polynomial of third and fourth degree: the roots are
obtainable from the coefficients with the operations of addition, multiplications and radicals
(taking roots). Mathematicians sought for a long time to extend these formulas. Among
them was Lagrange, who in 1770 unified the steps in solving equations for n ≤ 4, and showed
that this unified method failed for the quintic. In 1824, Abel proved conclusively that the
determination of the roots of a general quintic in terms of radicals of sums and products of
the coefficients was impossible. It was finally shown by Evariste Galois, that it is not possible
to obtain such general formulas for the roots of an n-th order equation for n ≥ 5. He did this
by making the crucial connection between group theory and polynomial equations. In fact
this work was the origin of what is now known as Galois theory, which has many applications
to other areas of mathematics.

All this does not mean that the roots do not exist! By the fundamental theorem of
algebra we know that every polynomial of degree n has exactly n roots (over the complex
field). Moreover, if the coefficients of the polynomial are all real, then these roots are either
real, or come in complex conjugate pairs. In addition, the roots are continuous functions of
the coefficients of the polynomial. This does not contradict Galois’s statement, which refers
to a specific class of such functions (obtained by sums products and radicals only).

Fortunately, such detailed information as the exact root location is not necessary. In
order to investigate the stability, it suffices to know whether all roots of the characteristic
equation are inside the left half plane. Their precise location in there is immaterial.

The question is then, what information about the root location can one infer from knowl-
edge of the coefficients of a polynomial?
Erik I. Verriest Chapter 4: Transfer Function and Stability 93

4.6 Stability Criteria

4.6.1 Necessary Condition for Stability
Here we prove a first useful result: A necessary condition for asymptotic stability is that all
the coefficients of the characteristic polynomial are strictly positive. Note that the condition
is not sufficient.
It is nonetheless a useful result, as we can directly decide from inspection that a system will
not be stable if it has nonpositive (zero or negative) coefficient.
Indeed, any stable complex conjugate pole pair −σ ± jω (σ > 0) contributes a factor
((s + σ)2 + ω 2) = s2 + 2σs + (σ 2 + ω 2) to the characteristic polynomial, and a real stable pole
at −r, r > 0, contributes a factor (s + r). The product of polynomials with strictly positive
coefficients is a polynomial with strictly positive coefficients. A simple counterexample can
be constructed showing that this condition is not sufficient: Consider p(s) = (s2 −s+p)(s+q)
where both p and q are positive. The factor s2 − s + p has its roots in the open right half
plane, but we see that p(s) = s3 + (q − 1)s2 + (p − q)s + pq, which has positive coefficients
for p > q > 1. It can be shown that a polynomial with nonnegative coefficients cannot have
roots on the positive real axis [3].

Consider now the polynomial sn + p1 sn−1 + p2 sn−2 + . . . + pn . If we choose the coeffi-

cients at random, say in [−B, B] and let B increase to infinity, then the ‘volume’ extended
by all polynomials which have positive coefficients is a fraction 1/2n of the total volume in
coefficient space. The fraction of asymptotically stable polynomials must then still be less
as shown above. Clearly as n goes to infinity (complicated systems) the fraction of stable
systems must be vanishingly small. In a way, we could state this as theorem: “Complexity
implies instability.” The control engineer has to be really clever to stabilize large scale (in
the sense of order) systems. The reason why we can find stable systems at all stems from the
fact that many systems are not simply formed by “choosing coefficients in a uniform random
way.” For instance, if we arbitrarily assemble systems with nR resistors, nC capacitors and
nL inductors, attach a voltage source somewhere as input, and check as output the current
in some branch, then the differential equation so obtained, which will have at most an order
nC + nL , is guaranteed to have all nonnegative coefficients.

4.6.2 Routh-Hurwitz Criterion

In order to investigate the stability, it suffices to know whether all roots of the characteristic
equation are inside the left half plane. Their precise location in there is immaterial. Whereas
Erik I. Verriest Chapter 4: Transfer Function and Stability 94

it is not possible to obtain general formulas for the roots of an nth order equation for n ≥ 5
in terms of radicals, as shown by Galois, it is nevertheless possible to determine the number
of roots in the right half plane. One forms a table, the Routh array, and checks the signs
of certain numbers in the array. The number of sign changes corresponds to the number of
unstable poles. This test was independently discovered by Routh and Hurwitz.

Hurwitz proved that the polynomial

a(s) = a0 sn + a1 sn−1 + a2 sn−2 + · · · + an−1 s + an ,

with real coefficients and a0 > 0, has all its roots in the open left half plane if

D1 = a1 > 0
a1 a0
D2 = >0
a3 a2
a1 a0 0
D3 = a3 a2 a1 >0
a5 a4 a3
..
.
a1 a0
a3 a2 a1 a0
Dn = a5 a4 a3 a2 a1 a0 > 0.
..
.
a2n−1 a2n−2 ··· ··· an

The coefficients ai = 0 for i > n. His proof uses a decomposition in continued fractions (See
E.A. Guillemin, The Mathematics of Circuit Analysis, Wiley 1949.)
Routh proved the criterion in a more practical form, using what is now known as the “Routh
table”. The construction of the Routh table proceeds as follows: There are n + 1 rows, which
we shall identify by sn , sn−1 , . . . , s1 and finally s0 . Place the coefficients of the polynomial
a(s) alternatingly in the first two rows, as follows:
Erik I. Verriest Chapter 4: Transfer Function and Stability 95

Routh Array

sn a0 a2 a4 · · ·
sn−1 a1 a3 a5 · · ·

If you run out of entries, enter 0. It helps to relabel the entries

sn A1 A2 A3 · · ·
sn−1 B1 B2 B3 · · ·
and construct the third row from it
sn A1 A2 A3 · · ·
sn−1 B1 B2 B3 · · ·
sn−2 C1 C2 C3 · · ·
The elements in this row are obtained from the two previous as follows
B1 A2 − B2 A1 B1 A3 − B3 A1 B1 A4 − B4 A1
C1 = , C2 = , C3 = .
B1 B1 B1
Now construct a 4-th row (the sn−3 row from the previous ones (the sn−1 and sn−2 row in
the same way: i.e., if this the D-row, set
C1 B2 − C2 B1 C1 B3 − C3 B1 C1 B4 − C4 B1
D1 = , D2 = , D3 = .
C1 C1 C1
Proceeding this way, the last row (the s0 -row) will only contain at most one nonzero element.

The Routh criterion states:

The number of roots with positive real parts of a(s) is equal to the number of sign changes
in the first column of the Routh array.

Examples:

i) Second order system: a(s) = a0 s2 + a1 s + a2

s2 a0 a2
s1 a1
a1 a2
s0 a1
= a2
If all ai have the same sign, this polynomial has both of its roots in the left half plane.
Erik I. Verriest Chapter 4: Transfer Function and Stability 96

ii) Third order system: a(s) = a0 s3 + a1 s2 + a2 s + a3

s3 a0 a2
s2 a1 a3
a1 a2 −a0 a3
s1 a1
0
s0 a3

Note that if a0 , a1 and a3 are positive, then one additional conditional condition is
required: a1 a2 − a0 a3 > 0.

iii) a(s) = 8s4 + 2s3 + 3s2 + s + 5

s4 8 3 5
s3 2 1
6−8 10
s2 2=−1 2
=5
−1−10 0
s1 −1
= 11 −1 = 0
55
s0 11
= 5

The first column, [8, 2, −1, 11, 5]′, has two sign changes: Two roots of this polynomial
live in the right half plane.

It follows from the construction of the Routh array, that multiplication of any row by a
positive constant will not change the criterion. This can be useful to reduce large numbers.
Here is an example: Let a(s) = s5 + 4s4 + 11s3 + 16s2 + 5s + 8. The table starts as

s5 1 11 5
s4 4 16 8

At this point, reduce the second row to [1, 4, 2], and continue the array

s4 1 4 2
s3 7 3
25
s2 7
2
−23
s1 25
s0 2

Since there are two sign changes, this a(s) will have two roots in the right half plane.
Erik I. Verriest Chapter 4: Transfer Function and Stability 97

Two special case my occur: a zero may appear in the first column. This is discussed in the
next section. The other special case is the case when two consecutive rows are proportional.
Take for instance a(s) = s7 + 4s6 + 5s5 + 5s4 + 6s3 + 9s2 + 8s + 2.

s7 1 5 6 8
s6 4 5 9 2
15 15 30
s5 4 4 4
s4 1 1 2
s3 0 0

We see that now we cannot generate the s2 row since each term would be 0/0. This problem
originates because the s5 and s4 rows are proportional. Now we need to say why we labeled
the rows by the decreasing powers of s. These rows actually refer to polynomials. For
instance the s7 row refers actually to the polynomial s7 +5s5 +6s3 +8s, the s6 row corresponds
to the polynomial 4s6 + 5s4 + 9s2 + 2. These polynomials have alternating parity (odd or
even). When it happens that two rows in the Routh array are proportional, the polynomial
corresponding to the second proportional row (this is here : s4 + s2 + 2) is a factor of the
original polynomial. To proceed further, differentiate this 4-th degree polynomial to get
d
the requisite 3-d degree polynomial. Thus we use, corresponding to ds (s4 +s2 +2) = 4s3 +2s,
the new s3 row [4, 2]. Now continue the array to get

s3 4 2
2 1
s 2
2
1
s −14
s0 2

There are two roots in the right half plane.

4.6.3 Homotopy
The main (and quite intuitive) idea behind this is that the roots of a polynomial are contin-
uous functions of the coefficients of that polynomial. Hence small changes in the coefficients
will induce small changes in the roots. Consequently, if one of the entries in the first column
of the Routh array is zero, it means that by a small change in a coefficient, one can make
this entry nonzero. Typically by changing the coefficient in the other direction, the entry
will change sign. Suppose we had a sequence (· · · , +, 0, +, · · ·). If by a small increase (by
ǫ > 0) in some coefficient, c, we can change this sign sequence to (· · · , +, +, +, · · ·), and by a
small decrease in the same coefficient to (· · · , +, −, +, · · ·), then the number of sign changes
in the two cases differs by two. Consequently, this means that two poles must cross from
Erik I. Verriest Chapter 4: Transfer Function and Stability 98

the left half plane to the right half plane if the coefficient goes from c + ǫ to c − ǫ for some
positive ǫ. Hence for ǫ = 0, the original polynomial, it must mean that two poles were on
the imaginary axis. What conclusion should be drawn if we had originally a sign sequence
(· · · , +, 0, −, · · ·)? What if the zero occurred in the last entry?

As an example, consider the family of polynomials aǫ (s) = s3 + (1 + ǫ)s2 + (1 + ǫ)s + 1.

The Routh array is
s3 1 1+ǫ
2
s 1+ǫ 1
1 ǫ(2+ǫ)
s 1+ǫ
s0 1
For small ǫ the first column is approximately [1, 1, 2ǫ, 1]. We see that if ǫ is positive, there
are no sign changes, while if ǫ is negative there are two. In the first case all roots are in the
open left hand plane, in the latter, two roots lie in the open right plane. Since the roots
must be continuous functions of the coefficients, that can only mean that there are two roots
that cross the jω axis precisely when ǫ = 0. Indeed, in this case this is easily verified by
the exact factorization aǫ (s) = (s2 + ǫs + 1)(s + 1). The roots crossing the jω axis form a
complex conjugate pair.

Here is another example: a(s) = s6 + 3s5 + 2s4 + 6s3 + 3s2 + 6s + 3.

s6 1 2 3 3
s5 3 6 6
s4 0 1 3
At this point, if we try to compute the next row, we get nonsense, i.e.,
−3 −9
s3 0 0

Let’s use the homotopy: introduce a small perturbation in one of the coefficients, so that a
zero no longer appears in the first column. We change the coefficient of s4 , to get a1 (s) =
s6 + 3s5 + 2.1s4 + 6s3 + 3s2 + 6s + 3.
Observe that we also ‘simplified’ the s5 , s4 and s3 rows
s6 1 2.1 3 3
s5 1 2 2
4
s 1 10 30
s3 −2 −7
s2 13 60
s1 22
13
s0 60
Erik I. Verriest Chapter 4: Transfer Function and Stability 99

The perturbed polynomial a1 (s) has two roots in the right half plane.

4.7 Supplement
(This section may be skipped at a first reading). The utility of this Routh-Hurwitz criterion
goes much further than one may expect. For instance it is possible to adapt the criterion to
find the number of roots of a polynomial to the left or right of Re s = α, and hence also the
number of roots inside a vertical strip bounded by α < Re s < β. We will develop this in
several steps below

Shift operator applied to a point set

Let S = {s1 , . . . , sn } be a set of points in C. Then denote by Tα S, for α ∈ C, the shifted set
Tα S = {s1 − α, . . . , sn − α}.
Likewise we define the operation on infinite sets. Let R+ denote the open right half plane
(RHP), then
Tα R+ = Tα { s | Re s > 0 }
{z α} | Re s > 0}
= { s| −
=s′
= { s | Re (s′ + α) > 0}
′

= { s′ | Re s > −Re α}.

Most of the time we will be interested in shifts by a real number, α ∈ R. In that case,

Tα R+ = { s | Re s > −α}.
Denote this half plane by R(−α)+ .

Given a finite set S, the number of elements in it is called the cardinality of the set and
this is denoted by card S.

Lemma: If S1 and S2 are finite sets, then

card (Tα S1 ∩ S2 ) = card (S1 ∩ T−α S2 ).

Proof: (Tα S1 ∩ S2 ) = Tα (S1 ∩ T−α S2 ), and the cardinality of a set does not change by
translation of the set. ✷
Erik I. Verriest Chapter 4: Transfer Function and Stability 100

Root set of a polynomial

If p is a polynomial, then let R(p) denote the set of roots of p.
R(p) = {s | p(s) = 0}.
Lemma: If p is a polynomial and α ∈ R, then R(T−α p) = Tα R(p). We say that the root
set is contravariant, meaning that upon commutation α changes to −α in the index of the
shift operator.
def
Proof: If p(s) = 0, then s ∈ R(p), and pα (s) = p(s + α) = 0 implies s + α ∈ R(p), and
thus s ∈ R(p) − α. Hence, R(T−α p) = R(pα ) = Tα R(p). ⋄

Define Rα+ as the domain in the complex plane to the right of α: Rα+ = { s | Re s > α }.
Let Rα+ (p) = R(p) ∩ Rα+ denote the set of roots of the polynomial p in the open half plane
to the right of α, and for simplicity let us denote R0+ (p) = R+ (p).

Note that the standard Routh - Hurwitz algorithm determines card R+ (p).

Theorem: If p is a polynomial and α ∈ R, then Rα+ (p) = R+ (T−α p), and therefore its
cardinality can be detected by the Routh - Hurwitz test on T−α p.
Proof:
Rα+ (p) = R(p) ∩ Rα+
= R(p) ∩ T−α R+
= Tα R(p) ∩ R+
= R(T−α p) ∩ R+
= R+ (T−α p). ⋄
Examples:
1. Consider the polynomial p(s) = s2 − 3s + 2. The number of roots with real part
larger than α ∈ R follow from the Routh Hurwitz test on the polynomial p(s + α) =
(s + α)2 − 3(s + α) + 2 = s2 + (2α − 3)s + α2 − 3α + 2. We find that the number of sign
changes in the sequence (1, 2α − 3, α2 − 3α + 2) equals the requisite number Rα+ (p).
i.e., 
 2 if α < 1
Rα+ (p) = 1 if α ∈ (1, 2) ,

0 if α > 2
which is easily verified since p has roots 1 and 2.
Erik I. Verriest Chapter 4: Transfer Function and Stability 101

2. For the polynomial p(s) = s3 + 2s2 + as − 3, determine R1+ (p). Noting that p(s + 1) =
(s + 1)3 + 2(s + 1)2 + a(s + 1) − 3 = s3 + 5s2 + (a + 7)s + a, the Routh - Hurwitz test
on the polynomial q(s) = s3 + 5s2 + (a + 7)s + a gives
s3 1 a+7
2
s 5 a
s1 4a + 35
s0 a
R1+ (p) = R+ (q) number of sign changes of the sequence (1, 5, 4a + 35, a), and thus

1 if a < 0
R1+ (p) = .
0 if a > 0
This must indicate that for a = 0, one root must cross the vertical through 1. Since
complex roots must come as conjugate pairs, this indicates that for a = 0 the polyno-
mial has a zero at s = 1. This is easily verified: s3 + 2s2 − 3 = (s − 1)(s2 + 3s + 3).

If, for α ∈ R, we let Rα− denote the half plane to the left of α, then

Theorem: Rα− (p) = R(−α)+ (Rp), where R is the parity or reversal operator.

Note that Rα− (p) is also deg p − Rα+ (p), provided that p has no roots on the line Re s = α.

Corollary: The number of roots of p in the strip (α, β) is found by computing Rα+ (p) −
Rβ+ (p), provided that p has no roots on the lines Re s = α and Re s = β.

Example
Let’s revisit Example 2 above. How many roots does p have in the vertical strip 0 < Re s < 1?
The Routh-Hurwith test for p gives
s3 1 a
2
s 2 −3
s1 2a + 3
s0 −3
For all a there is one sign change. Hence there ia always a root in the RHP, R+ (p) = 1.
From example 2 above, R1+ (p) = 1 or 0, depending on the sign of a. Hence, for a > 0, there
are no roots of p in this strip, while for a < 0, there will be one root in the vertical strip. It
must then be a real root.
Erik I. Verriest Chapter 4: Transfer Function and Stability 102

4.8 BIBO Stability

We assume now that the system was initially at rest, so that all initial conditions of the
system are zero. Let for all t, |u(t)| ≤ Bu for t > 0 and u(t) = 0 for t < 0. It follows from
the convolution representation that
Z t
|y(t)| = h(τ )u(t − τ ) dτ
0
Z t
≤ |h(τ )u(t − τ )| dτ
0
Z t
≤ Bu |h(τ )| dτ.
0
Rt
Clearly, if 0
|h(τ )| dτ remains bounded for all t, or
Z ∞
|h(τ )| dτ < ∞,
0
then the systemR is BIBO stable.
Conversely, if |h(τ )| dτ = ∞, then there exists a bounded input yielding an unbounded
output. Indeed, choosing an arbitrary t0 , and letting u(t) = sgn (h(t0 − t)), where sgn is the
sign function sgn(x) = x/|x| for x 6= 0, and +1 when x = 0, then
Z ∞
y(t0 ) = h(τ )u(t0 − τ ) dτ
0
Z ∞
= h(τ ) sgn (h(τ )) dτ
0
Z ∞
= |h(τ )| dτ
0
is unbounded.
Since h(t) = L−1
− H(s), it follows that for rational H(s), BIBO stability holds if and only if
all poles of H(s) belong to the open left half plane.

Remark: The above criterion still holds if the system has arbitrary initial conditions, un-
der the additional assumption that b(·) and a(·) have no common factors. For example,
the system, given in destructive form by D2 y − y = Du − u has a stable impulse response
h(t) = e−t , hence is BIBO stable. However, it is not stable with respect to initial conditions.
The unstable mode happened to correspond to a common unstable factor s − 1 in numerator
and denominator of b(s)/a(s).
Erik I. Verriest Chapter 4: Transfer Function and Stability 103

4.9 Sensitivity
Consider now a system with transfer function H(s, θ) = b(s,θ)a(s,θ
where θ is some parameter
H
of interest. The sensitivity function Sθ (s) is the relative sensitivity of H(s) with respect
to changes in θ. First assume a nominal value for the parameter: θ = θ0 , resulting in
H0 (s) = H(s, θ0 ). Consider now the new parameter value θ = θ0 +∆θ, where ∆θ is sufficiently
small. The corresponding evaluation of the transfer function at s is

∂H(s, θ0 )
H(s, θ0 + ∆θ) = H(s, θ0 ) + ∆θ + h.o.t
∂θ
Hence, taking limits
∆H(s, θ0 ) ∂H(s, θ0 )
→
∆θ ∂θ
The relative change in the transfer function evaluated at s is
∆H(s,θ0 )
H(s,θ0 ∂H(s, θ0 ) θ0
SθH (s) = ∆θ
= .
θ0
∂θ H(s, θ0 )

In general, this is complex valued. Often times we are just interested in the magnitude of
the above and define this as the sensitivity function.

Example: Consider a simple series RC circuit. If the applied voltage is the input, and the
voltage across the capacitor the output, then the transfer function is

1/Cs 1
H(s) = = .
R + 1/Cs 1 + RCs

The sensitivity with respect to variations in the resistance is

RCs τs
SRH (s) = − =− .
1 + RCs 1 + τs
Hence, for a fixed time constant τ = RC the sensitivity with respect to the tolerance on R
does not depend on how τ is generated as the product of R and C. In the high frequency
limit, the sensitivity approaches −1 or, if we use the modulus in the definition, 1. The low
frequency limit is zero.
Bibliography

[1] I. Stewart, “Galois Theory,” Chapman & Hall, 1989.

[2] G. Verriest, “Leçons sur la Théorie des Equations selon Galois,” Editions J. Gabay,
Paris, 1997 (reprint of original 1939 Gauthier Villars edition).

[3] Erik I. Verriest and Nak-seung Patrick Hyun. Roots of Polynomials with Positive Coef-
ficients. Proceedings of the 2018 International Symposium on the Mathematical Theory
of Networks and Systems, (MTNS 2018), Hong Kong, July 2018.

[4] Interesting web sites for some historical info:

http://www-groups.dcs.st-andrews.ac.uk/ history/Mathematicians/Galois.html
http://www-groups.dcs.st-andrews.ac.uk/ history/Mathematicians/Hurwitz.html
http://www-groups.dcs.st-andrews.ac.uk/ history/Mathematicians/Routh.html

104