Chapter 9

Sensitivity - Robustness

Why robustness?
• The plant characteristics may be variable or time-
varying. System modeling techniques identify the plant
within a certain model class and with a certain amount
of inaccuracy. So there always exists a plant uncer-
tainty, which can not be described exactly by the math-
ematical models.
• Control systems need to be made robust against this
plant variability and uncertainty.
Why frequency response analysis?
• There is a clear connection between frequency response
plots and experimentally obtained data.
• Easy to learn for trained engineers familiar with classi-
cal frequency response methods.

& %

ESAT–SCD–SISTA CACSD pag. 245

 Sensitivity

Definition
Sensitivity of a quantity α to changes in a quantity β:
∆ ∂α β ∂α/∂β
Sβα = · = .
∂β α α/β
⇒
A measure of the relative change in α due to a relative
change in β.

Sensitivity for a SISO system:

r yc
k(s) gt (s)

gt(s): plant, k(s): controller.

& %

ESAT–SCD–SISTA CACSD pag. 246

The closed-loop transfer function from r to yc:

gt(s)k(s)
h(s) = .
1 − gt(s)k(s)
The sensitivity of h(s) to changes in gt(s):
∂h(s) gt(s) 1
Sgh(s) = =
∂gt(s) h(s) 1 − gt(s)k(s)
Plant variations will have a small or a large effect on the
closed-loop transfer function according to the size of the
sensitivity function (1 − gt(s)k(s))−1.
⇒
The robustness of the closed-loop system to plant variations
is improved by making the sensitivity function small, said
otherwise, by making the “loop gain” gt(s)k(s) large.

& %

ESAT–SCD–SISTA CACSD pag. 247

Sensitivity for a MIMO system:

Consider the following two control systems :

equivalent
controller
r yo
K(s) Gt (s)
r yc
K(s) Gt (s)
G(s)

The equivalent open loop, G(s): model,

Gt(s): plant with uncertainties, K(s): controller.

Since

yc = Gt(s)K(s)(I − Gt(s)K(s))−1 r,
yo = Gt(s)K(s)(I − G(s)K(s))−1 r.

it follows that yc = yo for all r if the system Gt(s) and the

model G(s) are identical.

& %

ESAT–SCD–SISTA CACSD pag. 248

Suppose the plant Gt(s) depends on a parameter δ. Then,
∂yc
= (I − Gt(s)K(s))−1
∂δ
∂Gt(s)
× K(s)(I − Gt(s)K(s))−1 r,
∂δ
∂yo ∂Gt(s)
= K(s)(I − G(s)K(s))−1 r.
∂δ ∂δ
Assume G(s) = Gt(s)|δ=δnom , then
 
∂yc  −1 ∂yo 

= (I − Gt(s)K(s)) .
∂δ  δ=δnom ∂δ  δ=δnom
⇒
The sensitivity operator
∆
S(s) = (I − Gt(s)K(s))−1

determines how changes in the plant affect the output of the

closed-loop scheme given changes in the nominally equiva-
lent open-loop scheme.

For matrices :

if y = Su
⇓
||y|| ≤ σmax(S)||u||
& %

ESAT–SCD–SISTA CACSD pag. 249

Again, the closed-loop scheme will be more or less sensitive
to changes in the output depending on the “size” of S(s):

∂yc(jω) ∂yo(jω)

∂δ
≤ σ̄(S(jω))
∂δ
.
δ=δnom δ=δnom

where σ̄(S(jω)) is the largest singular value of the matrix

S(jω), a measure for the matrix size, and k · k is the
2-norm of a vector.

A feedback or controller design objective might be:

σ̄(S(jω)) < 1, ∀ω

which ensure that the closed-loop scheme is uniformally

less sensitive to changes in system parameters than the
open-loop scheme.

Complementary sensitivity:

∆
T (s) = G(s)K(s)S(s) = G(s)K(s)(I − G(s)K(s))−1
Note that S + T = I. The complementary sensitivity will
be used later.

& %

ESAT–SCD–SISTA CACSD pag. 250

 Robust Stability Analysis

Nyquist stability

Consider the following feedback system

w u
G(s)

y d
K(s)

Definition for internal stability :

Suppose G(s) and K(s) are propera rational transfer func-
tion matrices and let H(s)
" #denote " the# closed-loop transfer
w u
function matrix from to :
d y
" #−1
I −K(s)
H(s) = .
−G(s) I
Then
• The feedback loop is called well posed if H(s) is proper.
• The feedback loop is internally stable if H is stable.
a
H(s) is proper iff lims→∞ H(s) = D and D is finite. If D = 0, H(s) is called strictly proper.

& %

ESAT–SCD–SISTA CACSD pag. 251

Nyquist stability criterion for MIMO systems:
Let G(s) and K(s) be given proper rational transfer func-
tions that form a well-posed closed loop, and let G(s) and
K(s) have nG and nK poles (counting multiplicities) re-
spectively in the closed-right half plane. The feedback
loop is internally stable if and only if the Nyquist diagram
Γ = det(I − G(s)K(s)) makes nG + nK anti-clockwise
encirclements around the origin (without crossing it) for
s ∈ DR , where DR is the Nyquist contour:
jω

R
r
x

x: poles of G(s) or K(s) on the imaginary axis.

R is large enough to contain all the closed right half plane
poles of G(s), K(s) and H(s).

& %

ESAT–SCD–SISTA CACSD pag. 252

Small Gain Theorem
Gain of a system H
||Hw||
||H||gn = sup
||w||6=0 ||w||

here ||.|| is your favorite, well-chosen norm.

Not all norms will do : not every system norm is a gain.
Some gains for instance have the property that they are
finite if and only if the system is stable :

||H||gn < ∞ ⇔ H is stable.

The H∞-norm is an example.

Sub-multiplicativity
For all gains, the gain of a cascade is not larger than the
product of the gains

||H1H2||gn ≤ ||H2||gn||H1||gn.

This is called sub-multiplicativity.

& %

ESAT–SCD–SISTA CACSD pag. 253

Small Gain Theorem
Consider the following system :

w + H1 z
+ +

H2
Assume the problem is well-posed i.e. det(I − H1H2) 6= 0.
Then the transfer matrix from w to z is

Gzw = (I − H1H2)−1H1

Now the small gain theorem states :

If ||H1||gn||H2||gn < 1, then

||H1||gn
||Gzw ||gn ≤
1 − ||H1||gn ||H2||gn

If ||.||gn is such that finiteness means stability, then the

small gain condition ||H1||gn||H2||gn < 1 implies stability
of Gzw .

& %

ESAT–SCD–SISTA CACSD pag. 254

Proof :
we have

z = H1(w + H2z) = H1w + H1H2z

The triangle inequality delivers

||z|| ≤ ||H1w|| + ||H1H2z||

Using the gain definitions & sub-multiplicativity

||z|| ≤ ||H1||gn||w|| + ||H1H2||gn||z||

≤ ||H1||gn||w|| + ||H1||gn||H2||gn||z||

so that

||z||(1 − ||H1||gn||H2||gn) ≤ ||H1||gn||w||.

From the small gain condition

||H1||gn||H2||gn < 1

we find
||H1||gn
||z|| ≤ ||w||.
1 − ||H1||gn||H2||gn

& %

ESAT–SCD–SISTA CACSD pag. 255

Additive model error
Consider the control system:

∆a(s)

G(s)

K(s)

Feedback loop with additive model error

G(s): Nominal system transfer function,

∆a(s): Additive stable perturbation transfer function,
K(s): Controller transfer function designed to ensure the
internal stability of the nominal closed-loop.

How large can σ̄(∆a(jω)) become before the closed-loop

becomes unstable?

& %

ESAT–SCD–SISTA CACSD pag. 256

Necessary and sufficient condition for the closed-loop sys-
tem to be internally stable:

σ̄(∆a(jω))σ̄(K(jω)(I − G(jω)K(jω))−1) < 1, ∀ω.

⇒
• Given the gain of the perturbation:
∆
γa = max σ̄(∆a(jω)),
ω

then the stability condition is

1
−1
σ̄(K(jω)(I − G(jω)K(jω)) ) < , ∀ω,
γa
which can be used as a design objective. And the design
problem can be formulated as a minimization problem:
 
min max σ̄(K(jω)(I − G(jω)K(jω))−1) .
K(s) ω

• The exact structure of ∆a(s) is not needed.

& %

ESAT–SCD–SISTA CACSD pag. 257

Multiplicative model error
Consider the control system:

∆m(s)

G(s)

K(s)

Feedback loop with multiplicative model error

G(s): Nominal system transfer function,

∆m(s): Multiplicative stable perturbation transfer func-
tion,
K(s): Controller transfer function designed to ensure the
internal stability of the nominal closed-loop.

How large can σ̄(∆m(jω)) become before the closed-loop

becomes unstable?

& %

ESAT–SCD–SISTA CACSD pag. 258

Necessary and sufficient condition for the closed-loop sys-
tem to be internally stable:

σ̄(∆m(jω))σ̄(G(jω)K(jω)(I − G(jω)K(jω))−1) < 1, ∀ω.

| {z }
T (jω)
⇒
• Given the gain of the perturbation:
∆
γm = max σ̄(∆m(jω)),
ω

then the stability condition is

1
σ̄(T (jω)) < , ∀ω
γm
which can be used as a design objective. And the design
problem can be formulated as a minimization problem:
 
min max σ̄(T (jω)) .
K(s) ω

• The exact structure of ∆m(s) is again not needed.

& %

ESAT–SCD–SISTA CACSD pag. 259

Feedback multiplicative model error
Consider the control system:

∆f (s)

G(s)

K(s)

Feedback loop with feedback multiplicative model error

G(s): Nominal system transfer function,

∆f (s): Feedback multiplicative perturbation transfer
function such that both ∆f (s), and (I − ∆f (s))−1 are
well-posed and stable.
K(s): Controller transfer function designed to ensure the
internal stability of the nominal closed-loop.

How large can σ̄(∆f (jω)) become before the closed-loop

becomes unstable?

& %

ESAT–SCD–SISTA CACSD pag. 260

Necessary and sufficient condition for the closed-loop sys-
tem to be internally stable:

σ̄(∆f (jω))σ̄((I − G(jω)K(jω))−1) < 1, ω ∈ R.

| {z }
S(jω)
⇒
• Given the gain of the perturbation
∆
γf = max σ̄(∆f (jω)),
ω

then the stability condition is

1
σ̄(S(jω)) < , ∀ω,
γf
which can be used as an design objective, and the design
problem can be formulated as
 
min max σ̄(S(jω)) .
K(s) ω

• The exact structure of ∆f (s) is not needed.

& %

ESAT–SCD–SISTA CACSD pag. 261

General case - small gain condition
Consider the following feedback loop:

∆(s)

Gc (s)

General uncertainty feedback loop

∆(s): Perturbation transfer function, stable. Gc(s): Plant

transfer function “seen” from the perturbation ∆(s), sta-
ble.

Model Error ∆(s) Gc(s)

Additive ∆a(s) K(s)(I − G(s)K(s)−1 )
Multiplicative ∆m(s) T (s)
Feedback Multplic. ∆f (s) S(s)

How large can σ̄(∆(jω)) become before the closed-loop be-

comes unstable?

& %

ESAT–SCD–SISTA CACSD pag. 262

Small Gain Condition:
Necessary and sufficient condition for the closed-loop sys-
tem to be internally stable (page 254) :

σ̄(∆(jω))σ̄(Gc(jω)) < 1, ∀ω.

⇒
• Essentially the small gain condition states that if a feed-
back loop consists of stable systems and the loop-gain
product is less than unity, then the feedback loop is in-
ternally stable. Otherwise there always exists a ∆(s)
such that the closed-loop system is unstable.
• Given the gain of the perturbation
∆
γ = max σ̄(∆(jω)),
ω

then the stability condition is

1
σ̄(Gc(jω)) < , ∀ω
γ
• There is no need to know the structure of the pertur-
bation ∆(s).

& %

ESAT–SCD–SISTA CACSD pag. 263

Example Tape drive control - Robustness analysis

Since only robust stability will be investigated here, all

the noise inputs, the reference input and the regulated
output are neglected and only the main loop (LQG design
here) will be considered:

∆a(s)

∆m(s) ∆f (s)

Tape drive G(s)

LQG controller K(s)

We analyze how large the additive, multiplicative and feed-

back multiplication perturbations (∆a(s), ∆m(s), ∆f (s))
would be before the system becomes unstable when they
are independently acting in the closed-loop.

& %

ESAT–SCD–SISTA CACSD pag. 264

Additive error.
Plot of σ̄(K(jω)(I − K(jω)G(jω))−1)
0.8

σ̄(K(jω)(I − K(jω)G(jω))−1)
0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 −2 −1 0 1 2
10 10 10 10 10
ω : rad/sec.

This plot shows that the the peak value of σ̄(K(jω)(I −

G(jω)K(jω))−1) is less than 0.8. Thus if the gain of the
additive error, γa = maxω (∆a(jω)), is less than 1/0.8 =
1.25, the closed-loop system will not be destabilized by this
error.

& %

ESAT–SCD–SISTA CACSD pag. 265

Multiplicative error.
Complementary sensitivity plot
1

0.8

0.6
σ̄(T )

0.4

0.2

0 −2
10
−1
10
ω : rad/sec.
0
10
1
10
2
10

This plot shows that the peak value of the maximum sin-
gular value of the complementary sensitivity is less than
1. Thus once the gain of the multiplicative error, γm =
maxω (∆m(jω)), is less than 1, the closed-loop system will
not be destabilized by this error.

& %

ESAT–SCD–SISTA CACSD pag. 266

Feedback multiplicative error.
Sensitivity plot
1.25

1.2

1.15

1.1
σ̄(S)

1.05

0.95

0.9

0.85 −2
10
−1
10
ω : rad/sec.
0
10
1
10
2
10

This plot shows that the peak value of the maximum

singular value of the sensitivity is less than 1.24. Thus
once the gain of the feedback multiplicative error, γf =
maxω (∆f (jω)), is less than 1/1.24 = 0.8065 (< 1) than
the closed-loop system will not be destabilized by this er-
ror. Note also that the peak value of the maximum singular
value of the sensitivity is about 1.24. This means that a
change in the open loop system will cause a change in the
closed-loop which is 1.24 times larger (in percentage).

& %

ESAT–SCD–SISTA CACSD pag. 267

Summary

• Using SVD as a vehicle to analyze the robustness of

systems with perturbations gives a simple way to gen-
eralize the frequency analysis methods from SISO to
MIMO systems. However, since the structures of the
perturbations are not considered, the results are often
conservative. Thus it is possible that some classes of
perturbations with specific structure (diagonal for in-
stance) might never destabilize the closed-loop systems
even if the gain conditions are not satisfied.
• So far we have not discussed the subject on robust con-
trol design, which is out of the scope of this course.
Robust control design for systems with unstructured
perturbations, namely H∞ control synthesis, has been
solved. The structure of the solution is similar to LQG
control: combine the optimal control law and optimal
estimator based on a similar separation principle as in
LQG design. However, the Riccati equations are differ-
ent.

& %

ESAT–SCD–SISTA CACSD pag. 268