6.2.

Weighted H2 and H∞ Performance 85

and good robustness and good sensor noise rejection require in some frequency range,
typically some high-frequency range (ωh , ∞),

σ(P K)  1, σ(KP )  1, σ(K) ≤ M

where M is not too large. These design requirements are shown graphically in Figure 6.2.
The specific frequencies ωl and ωh depend on the specific applications and the knowledge
one has of the disturbance characteristics, the modeling uncertainties, and the sensor
noise levels.

HH
@ HH6
@ HH
@ @
XX @ σ(L)
XX @ @
Z @ @
Z @
Z @ @
Z S @
Z @
Z S
S Z
H Z ωh
HH Z -
ωl HH ZZ log ω

S @HH
σ(L) S @ HH
S @ H
HH
S @ HH
S @
S @

Figure 6.2: Desired loop gain

6.2 Weighted H2 and H∞ Performance

In this section, we consider how to formulate some performance objectives into math-
ematically tractable problems. As shown in Section 6.1, the performance objectives of
a feedback system can usually be specified in terms of requirements on the sensitivity
functions and/or complementary sensitivity functions or in terms of some other closed-
loop transfer functions. For instance, the performance criteria for a scalar system may
be specified as requiring 
|S(jω)| ≤ ε, ∀ω ≤ ω0 ,
|S(jω)| ≤ M, ∀ω > ω0
where S(jω) = 1/(1 + P (jω)K(jω)). However, it is much more convenient to reflect
the system performance objectives by choosing appropriate weighting functions. For
86 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

example, the preceding performance objective can be written as

|We (jω)S(jω)| ≤ 1, ∀ω
with 
1/ε, ∀ω ≤ ω0
|We (jω)| =
1/M, ∀ω > ω0
To use We in control design, a rational transfer function We (s) is usually used to ap-
proximate the foregoing frequency response.
The advantage of using weighted performance specifications is obvious in multi-
variable system design. First, some components of a vector signal are usually more
important than others. Second, each component of the signal may not be measured in
the same units; for example, some components of the output error signal may be mea-
sured in terms of length, and others may be measured in terms of voltage. Therefore,
weighting functions are essential to make these components comparable. Also, we might
be primarily interested in rejecting errors in a certain frequency range (for example, low
frequencies); hence some frequency-dependent weights must be chosen.

6
ũ d˜i d˜
? ?
Wu Wi Wd
6 di d
- Wr r-e - K u -?
e - P -?
e y- We e-

− 6

?n
e Wn 
ñ

Figure 6.3: Standard feedback configuration with weights

In general, we shall modify the standard feedback diagram in Figure 6.1 into Fig-
ure 6.3. The weighting functions in Figure 6.3 are chosen to reflect the design objectives
and knowledge of the disturbances and sensor noise. For example, Wd and Wi may be
chosen to reflect the frequency contents of the disturbances d and di or they may be used
to model the disturbance power spectrum depending on the nature of signals involved
in the practical systems. The weighting matrix Wn is used to model the frequency
contents of the sensor noise while We may be used to reflect the requirements on the
shape of certain closed-loop transfer functions (for example, the shape of the output
sensitivity function). Similarly, Wu may be used to reflect some restrictions on the con-
trol or actuator signals, and the dashed precompensator Wr is an optional element used
to achieve deliberate command shaping or to represent a nonunity feedback system in
equivalent unity feedback form.
6.2. Weighted H2 and H∞ Performance 87

It is, in fact, essential that some appropriate weighting matrices be used in order
to utilize the optimal control theory discussed in this book (i.e., H2 and H∞ theory).
So a very important step in the controller design process is to choose the appropriate
weights, We , Wd , Wu , and possibly Wn , Wi , Wr . The appropriate choice of weights for a
particular practical problem is not trivial. In many occasions, as in the scalar case, the
weights are chosen purely as a design parameter without any physical bases, so these
weights may be treated as tuning parameters that are chosen by the designer to achieve
the best compromise between the conflicting objectives. The selection of the weighting
matrices should be guided by the expected system inputs and the relative importance
of the outputs.
Hence, control design may be regarded as a process of choosing a controller K such
that certain weighted signals are made small in some sense. There are many different
ways to define the smallness of a signal or transfer matrix, as we have discussed in
the last chapter. Different definitions lead to different control synthesis methods, and
some are much harder than others. A control engineer should make a judgment of the
mathematical complexity versus engineering requirements.
Next, we introduce two classes of performance formulations: H2 and H∞ criteria.
For the simplicity of presentation, we shall assume that di = 0 and n = 0.

H2 Performance
Assume, for example, that the disturbance d˜ can be approximately modeled as an
impulse with random input direction; that is,
˜ = ηδ(t)
d(t)

and
E(ηη ∗ ) = I
where E denotes the expectation. We may choose to minimize the expected energy of
˜
the error e due to the disturbance d:
n o Z ∞ 
2 2 2
E kek2 = E kek dt = kWe So Wd k2
0

In general, a controller minimizing only the above criterion can lead to a very large
control signal u that could cause saturation of the actuators as well as many other
undesirable problems. Hence, for a realistic controller design, it is necessary to include
the control signal u in the cost function. Thus, our design criterion would usually be
something like this:
n o

We So Wd
 2

E kek2 + ρ2 kũk2 =
2 2
ρWu KSo Wd
2

with some appropriate choice of weighting matrix Wu and scalar ρ. The parameter ρ
clearly defines the tradeoff we discussed earlier between good disturbance rejection at
88 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

the output and control effort (or disturbance and sensor noise rejection at the actuators).
Note that ρ can be set to ρ = 1 by an appropriate choice of Wu . This problem can
be viewed as minimizing the energy consumed by the system in order to reject the
disturbance d.
This type of problem was the dominant paradigm in the 1960s and 1970s and is
usually referred to as linear quadratic Gaussian control, or simply as LQG. (Such prob-
lems will also be referred to as H2 mixed-sensitivity problems for consistency with the
H∞ problems discussed next.) The development of this paradigm stimulated extensive
research efforts and is responsible for important technological innovation, particularly
in the area of estimation. The theoretical contributions include a deeper understanding
of linear systems and improved computational methods for complex systems through
state-space techniques. The major limitation of this theory is the lack of formal treat-
ment of uncertainty in the plant itself. By allowing only additive noise for uncertainty,
the stochastic theory ignored this important practical issue. Plant uncertainty is par-
ticularly critical in feedback systems. (See Paganini [1995,1996] for some recent results
on robust H2 control theory.)

H∞ Performance
Although the H2 norm (or L2 norm) may be a meaningful performance measure and
although LQG theory can give efficient design compromises under certain disturbance
and plant assumptions, the H2 norm suffers a major deficiency. This deficiency is due
to the fact that the tradeoff between disturbance error reduction and sensor noise error
reduction is not the only constraint on feedback design. The problem is that these
performance tradeoffs are often overshadowed by a second limitation on high loop gains
— namely, the requirement for tolerance to uncertainties. Though a controller may
be designed using FDLTI models, the design must be implemented and operated with
a real physical plant. The properties of physical systems (in particular, the ways in
which they deviate from finite-dimensional linear models) put strict limitations on the
frequency range over which the loop gains may be large.
A solution to this problem would be to put explicit constraints on the loop gain in
the cost function. For instance, one may chose to minimize

sup kek2 = kWe So Wd k∞

kd˜k2 ≤1

subject to some restrictions on the control energy or control bandwidth:

sup kũk2 = kWu KSo Wd k∞

kd̃k2 ≤1

Or, more frequently, one may introduce a parameter ρ and a mixed criterion
n o
  2

We So Wd
sup kek22 + ρ2 kũk22 = ρWu KSo Wd

kd̃k2 ≤1 ∞
6.3. Selection of Weighting Functions 89

Alternatively, if the system robust stability margin is the major concern, the weighted
complementary sensitivity has to be limited. Thus the whole cost function may be
 
We So Wd

ρW1 To W2
∞

where W1 and W2 are the frequency-dependent uncertainty scaling matrices. These

design problems are usually called H∞ mixed-sensitivity problems. For a scalar system,
an H∞ norm minimization problem can also be viewed as minimizing the maximum
magnitude of the system’s steady-state response with respect to the worst-case sinusoidal
inputs.

6.3 Selection of Weighting Functions

The selection of weighting functions for a specific design problem often involves ad hoc
fixing, many iterations, and fine tuning. It is very hard to give a general formula for
the weighting functions that will work in every case. Nevertheless, we shall try to give
some guidelines in this section by looking at a typical SISO problem.
Consider an SISO feedback system shown in Figure 6.1. Then the tracking error is
e = r − y = S(r − d) + T n − SP di . So, as we have discussed earlier, we must keep |S|
small over a range of frequencies, typically low frequencies where r and d are significant.
To motivate the choice of our performance weighting function We , let L = P K be a
standard second-order system
ωn2
L=
s(s + 2ξωn )
It is well-known from the classical control theory that the quality of the (step) time
response can be quantified by rise time tr , settling time ts , and percent overshoot
100Mp%. Furthermore, these performance indices can be approximately calculated as

0.6 + 2.16ξ 4 − √ πξ
tr ≈ , 0.3 ≤ ξ ≤ 0.8; ts ≈ ; Mp = e 1−ξ2 , 0 < ξ < 1
ωn ξωn

The key points to note are that (1) the speed of the system response is proportional to
ωn and (2) the overshoot of the system response is determined only by the damping ratio
ξ. It is well known that the frequency ωn and the damping ratio ξ can be essentially
captured in the frequency domain by the open-loop crossover frequency and the phase
margin or the bandwidth and the resonant peak of the closed-loop complementary
sensitivity function T .
Since our performance objectives are closely related to the sensitivity function, we
shall consider in some detail how these time domain indices or, equivalently, ωn and ξ
are related to the frequency response of the sensitivity function

1 s(s + 2ξωn )
S= = 2
1+L s + 2ξωn s + ωn2
90 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

1
10

0
10
sensitivity function

−1
10

−2
10 −1 0 1
10 10 10
normalized frequency

Figure 6.4: Sensitivity function S for ξ = 0.05, 0.1, 0.2, 0.5, 0.8, and 1 with normalized
frequency (ω/ωn )

The frequency
√ response of the sensitivity function S is shown in Figure √ 6.4. Note that
|S(jωn / 2)| = 1. We can regard the closed-loop bandwidth ωb ≈ ωn / 2, since beyond
this frequency the closed-loop system will not be able to track the reference and the
disturbance will actually be amplified.
Next, note that
p
α α2 + 4ξ 2
Ms := kSk∞ = |S(jωmax )| = p
(1 − α2 )2 + 4ξ 2 α2
q p
where α = 0.5 + 0.5 1 + 8ξ 2 and ωmax = αωn . For example, Ms = 5.123 when
ξ = 0.1. The relationship between ξ and Ms is shown in Figure 6.5. It is clear that the
overshoot can be excessive if Ms is large. Hence a good control design should not have
a very large Ms .
Now suppose we are given the time domain performance specifications then we can
determine the corresponding requirements in frequency domain in terms of the band-
width ωb and the peak sensitivity Ms . Hence a good control design should result in
a sensitivity function S satisfying both the bandwidth ωb and the peak sensitivity Ms
requirements, as shown in Figure 6.6. These requirements can be approximately repre-
sented as  
 s 
|S(s)| ≤   , s = jω, ∀ ω
s/Ms + ωb 
6.3. Selection of Weighting Functions 91

4.5

3.5
peak sensitivity

2.5

1.5

1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
damping ratio

Figure 6.5: Peak sensitivity Ms versus damping ratio ξ

1/|We|

Ms

ωb |S(jω)|

Figure 6.6: Performance weight We and desired S

92 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

Or, equivalently, |We S| ≤ 1 with

s/Ms + ωb
We = (6.5)
s
The preceding discussion applies in principle to most control design and hence the
preceding weighting function can, in principle, be used as a candidate weighting function
in an initial design. Since the steady-state error with respect to a step input is given by
|S(0)|, it is clear that |S(0)| = 0 if the closed-loop system is stable and kWe Sk∞ < ∞.
Unfortunately, the optimal control techniques described in this book cannot be used
directly for problems with such weighting functions since these techniques assume that all
unstable poles of the system (including plant and all performance and control weighting
functions) are stabilizable by the control and detectable from the measurement outputs,
which is clearly not satisfied if We has an imaginary axis pole since We is not detectable
from the measurement. We shall discuss in Chapter 14 how such problems can be
reformulated so that the techniques described in this book can be applied. A theory
dealing directly with such problems is available but is much more complicated both
theoretically and computationally and does not seem to offer much advantage.

1/|We|

Ms

ωb |S(j ω)|

Figure 6.7: Practical performance weight We and desired S

Now instead of perfect tracking for step input, suppose we only need the steady-
state error with respect to a step input to be no greater than  (i.e., |S(0)| ≤ );
then it is sufficient to choose a weighting function We satisfying |We (0)| ≥ 1/ε so that
kWe Sk∞ ≤ 1 can be achieved. A possible choice of We can be obtained by modifying
the weighting function in equation (6.5):

s/Ms + ωb
We = (6.6)
s + ωb ε
6.3. Selection of Weighting Functions 93

Hence, for practical purpose, one can usually choose a suitable ε, as shown in Figure 6.7,
to satisfy the performance specifications. If a steeper transition between low-frequency
and high-frequency is desired, the weight We can be modified as follows:
 √ k
s/ k Ms + ωb
We = √ (6.7)
s + ωb k ε

for some integer k ≥ 1.

The selection of control weighting function Wu follows similarly from the preceding
discussion by considering the control signal equation

u = KS(r − n − d) − T di

The magnitude of |KS| in the low-frequency range is essentially limited by the allowable
cost of control effort and saturation limit of the actuators; hence, in general, the max-
imum gain Mu of KS can be fairly large, while the high-frequency gain is essentially
limited by the controller bandwidth (ωbc ) and the (sensor) noise frequencies. Ideally,
one would like to roll off as fast as possible beyond the desired control bandwidth so
that the high-frequency noises are attenuated as much as possible. Hence a candidate
weight Wu would be
s + ωbc /Mu
Wu = (6.8)
ωbc

1/|Wu |
Mu

|KS(jω)|

ω bc

ε1

Figure 6.8: Control weight Wu and desired KS

However, again the optimal control design techniques developed in this book cannot
be applied directly to a problem with an improper control weighting function. Hence
94 PERFORMANCE SPECIFICATIONS AND LIMITATIONS

we shall introduce a far away pole to make Wu proper:

s + ωbc /Mu
Wu = (6.9)
ε1 s + ωbc

for a small ε1 > 0, as shown in Figure 6.8. Similarly, if a faster rolloff is desired, one
may choose
 √ k
s + ωbc / k Mu
Wu = √
k ε s + ω
(6.10)
1 bc

for some integer k ≥ 1.

The weights for MIMO problems can be initially chosen as diagonal matrices with
each diagonal term chosen in the foregoing form.

6.4 Bode’s Gain and Phase Relation

One important problem that arises frequently is concerned with the level of perfor-
mance that can be achieved in feedback design. It has been shown in Section 6.1 that
the feedback design goals are inherently conflicting, and a tradeoff must be performed
among different design objectives. It is also known that the fundamental requirements,
such as stability and robustness, impose inherent limitations on the feedback properties
irrespective of design methods, and the design limitations become more severe in the
presence of right-half plane zeros and poles in the open-loop transfer function.
In the classical feedback theory, Bode’s gain-phase integral relation (see Bode [1945])
has been used as an important tool to express design constraints in scalar systems. This
integral relation says that the phase of a stable and minimum phase transfer function
is determined uniquely by the magnitude of the transfer function. More precisely, let
L(s) be a stable and minimum phase transfer function: then
Z ∞
1 d ln |L| |ν|
∠L(jω0 ) = ln coth dν (6.11)
π −∞ dν 2

|ν| e|ν|/2 + e−|ν|/2

where ν := ln(ω/ω0 ). The function ln coth = ln |ν|/2 is plotted in Fig-
2 e − e−|ν|/2
ure 6.9.
|ν|
Note that ln coth decreases rapidly as ω deviates from ω0 and hence the integral
2
d ln |L(jω)|
depends mostly on the behavior of near the frequency ω0 . This is clear
dν
from the following integration:
 
Z  1.1406 (rad), α = ln 3  65.3o , α = ln 3
1 α |ν|
ln coth dν = 1.3146 (rad), α = ln 5 = 75.3o , α = ln 5
π −α 2  
1.443 (rad), α = ln 10 82.7o , α = ln 10.