Chapter 6

Control Design begins with the understanding of specifications or

performance to be attained
6.1 Introduction
6.1.1 Control Design Specifications
Step Response:-In classical control systems, step response characteristics are usually the main design
specifications. Overshoot, M, raise time, Tr, steady state error, eo, and settling time, Ts, are given as
maximum allowable values. The design is expected to result these performance indicators below the
given allowable values.
Pole location for a desired performance:- The performance of closed loop control systems with a
known closed loop transfer function pole location can be used as specification. Hence the control
design objective will be locating the important poles at the specified locations. For example, location
on the left hand side of the complex plane ensures stability while the distance from the pole to the
imaginary axis may indicate testability margin. The performance resulting from pole locations can
generally interpreted to step response performance which is in time domain.
Loop shaping using frequency domain:- In loop shaping, the gain of the loop transfer function (GK)
and the feedback phase shift determine the performance of the closed loop system. The loop gain
(GK) magnitude and phase plots as a function of frequency are considered to mark the stable
operating range with phase and gain margins to indicate the stability level. The bandwidth of the
closed loop transfer function can also be indicated. Again, this can be related to the time domain
step response through appropriate approximations.
Linear Quadratic Gaussian (LQG) control (Quantitative expression of tracking error and the control
signal):- LQG is a sort of optimal control in which the tracking error squared is minimized with a
minimum control signal squared in presence of Gaussian disturbance and measurement noise.

H-infinity control, Shaping the singular values of closed-loop transfer functions as a function of
frequency (Optimal approach):- Classical loop-shaping is possible for SISO system using loop-gain
(L(j)=G(j)K(j)) as a function of frequency as described above. In case of multivariable system
singular values of closed-loop transfer functions such as sensitivity and complementary sensitivity

Page 1 of 30
functions are used instead of the loop-shaping in the SISO system. Minimization of the singular
values of important closed loop transfer functions (like S, T, KS) is called H∞ control.
For a closed-loop system with reference r, output disturbance d and measurement noise n, we have
the expression for output y, the control signal u and the error, e=y-r, as;
y ( s )  T ( s ) r ( s )  S ( s ) d ( s )  T ( s ) n( s )
u ( s)  K ( s) S ( s)[r ( s)  n( s)  d ( s)]
e( s)  y ( s)  r ( s)   S ( s){r ( s)  d ( s)}  T ( s)n( s)
Then, the closed-loop objectives in addition to stabilizing G(s) are;
1. For disturbance rejection, the maximum singular value as a function of frequency in the effective

operating frequency range (  (S (s)) ) has to be small.

2. For noise attenuation, the maximum singular value of the complementary sensitivity function as

the function of frequency (  (T (s)) in the effective frequency range has to be small.
3. For reference tracking, the value of the minimum and maximum singular value of complementary

sensitivity function (  (T )   (T )  1 ) approximately equal to 1.

4. For control energy reduction make the maximum singular value of KS,  ( K (s)S (s)) small.

5. For robust stability with multiplicative representation of uncertainty as G p  ( I  )G make the

maximum singular value of complementary sensitivity function (  (T (s)) ) small.

From the above five requirements, we can see that disturbance rejection requirement number 1 and
noise attenuation requirement number 2 cannot be satisfied simultaneously at the same frequency.
Fortunately, disturbance rejection is usually requirement of low frequency while noise attenuation is
requirement at high frequency.

Exercise 6.1.1: Interpret the above requirements in terms of the loop gain L(s) = G(s)K(s)

In general, the stability constraint in the closed loop transfer function gain shaping for MIMO system
is more difficult than the closed-loop shaping.
6.1.2 Control Design considered in this chapter
There are several approaches to the design of multivariable control system. If the plant is with weak
coupling, where each input may affect one output significantly while their effect on other outputs is
weak or marginal, then, one can use classical control system design approach for each pair of input

Page 2 of 30
output. Mathematical tools like RGA can be utilized to identify and pair the input and outputs. Then,
controllers for the paired input outputs can be designed in the conventional classical control design
methods. It may also be possible to decouple around selected operating frequency like zero
frequency or cut-off frequency using pre and post compensators. Again the decoupled system pairs
can be designed as a single input single output control system independently.

In this chapter, we will consider other three control design technics which we can utilize directly for
both multivariable system and SISO control system design. The three controller design techniques
are linear quadratic Gaussian (LQG) control, which is generally a sort of optimal control minimizing
the error and the control signal, H-Infinity control which is again an optimal control and to some
extent the extension of loop-shaping in the SISO system and the third called internal model control.

These techniques are valid both for single input single output (scalar) and multivariable systems with
appropriate adjustments.

Page 3 of 30
 6.2 Linear Quadratic Gaussian (LQG) Control
LQG control is basically a control system which is a trade-off between minimizing the tracking error
and minimizing the control signal for linear control system in presence of Gaussian noise disturbance
and measurement noise. In general, to minimize the tracking error, a control system requires large
control signal and vice versa. However, control signal cannot be increased at will; because saturation
has to be avoided and also control signal represents energy, which represents cost and has to be
limited. Therefore; the control design keeps the balance between the two objectives using
appropriate weights, as control parameters, to the two objectives.
6.2.1 The LQG Problem formulation
Assume that plant dynamics are known, linear and the measurement noise and disturbance signals
are stochastic with known statistical properties as given below.
x  Ax  Bu  Nwd
y  Cx  wn
Where wd and wn are the disturbance (process noise) and measurement noise inputs respectively,
which are usually assumed to be uncorrelated zero-mean Gaussian stochastic processes with
constant power spectral density matrices W and V respectively. That is, wd and wn are white noise
processes with covariance given by
E wd (t ) wd ( ) T   W (t   )
E wn (t ) wn ( ) T   V (t   )

   
And E wd (t )wn ( )T  0, E wn (t )wd ( )T =0 Where E is expectation operator while  (t-) is delta

function.

 wd  W Rwv 
Let the intensity of the white noise disturbance and noise   be RT V 
.
 wn   wv
Refer to appendix I of this chapter for the signal measurement techniques.
LQG control then is determination of control signal u which minimizes the cost function J shown
below. The name LQG is from L-Linear, Q-Quadratic system model and G-Gaussian white noise
modeling of disturbance and noise.

Page 4 of 30
There are two objectives formulated in the LQG cost function equation; namely, square error
minimization which is represented by XTQX, and optimum required control signal squared which is
represented by UTRU.
 
 
T
1
J  E  Lim  x T Qx  u T Ru dt 
 T  T 0 

Where Q and R are approximately chosen constant weighting matrices (design parameters) such that
Q = QT ≥ 0 and R = RT > 0. Q and R vales determine the weight we give to minimize the error and the
control signal respectively. Furthermore, elements of Q and R determine the relative weight given to
vector elements of error and the control signal.
6.2.2 Solution to LQG Problems
Solution to the LQG problem can be done by the separation theorem or Certainty Equivalence
Principle, which is simple and elegant requiring two steps.
The first step is determining the optimal control to a deterministic linear quadratic regulator (LQR)
problem; that is LQG problem without disturbance and noise (wd and wn). The solution to this
problem can be written in terms of the simple state feedback law, assuming availability of x for
feedback.
x  Ax  Bu
y  Cx

u(t )   K r x(t )
Where Kr is a constant matrix which is easy to compute.

 
The second step is to find an optimal estimate x of the state x so that E x  x T x  x  is minimized.
 

The optimal state estimate is given by a Kalman filter and is independent of Q and R. The required
 
solution to the LQG problem is then found by replacing x by x , to give u(t )   K r x (t ) .
  
x  Ax  Bu  K f ( y  Cx )

Where Kf is observer gain.

Optimal State Feedback (LQR) Solution

The LQR problem where all the states are known, is the deterministic initial value problem; Given the
system x  Ax  Bu with a given non-zero initial state x(0), find the input signal u(t) which the
system back to the zero state x=0 in an optimal manner, i.e. by minimizing the deterministic cost.

Page 5 of 30
 
J r   ( x(t )T Qx(t )  u (t )T Ru (t ))dt
0

The optimal control signal value (solution) for any initial state is given by u(t )   K r x(t ) where

K r  R 1 B T X

And X  X T  0 is the unique positive-semi definite solution of the algebraic Riccati equation
AT X  XA  XBR 1 B T X  Q  0

Kalman Filter (Solution in presence of noise and estimated state) Solution

The Kalman filter has the structure of an ordinary state estimate or observer, as shown below.
  
x  Ax  Bu  K f ( y  Cx )


The optimal choice of Kf , which minimizes the error, E x  x T x  x  , is given by
 

K f  (YC T  NRwv )V 1

Where, again, Y  Y T  0 is the unique positive semidefinite solution of the algebraic Riccati
equation.
YAT  AY  (YC T  NRwv )V 1 (YC T  NRwv )T  NWN T  0

Rwv is cross covariance between vn and vw.

The LQG control with estimated state variable feedback system can be described by the following
equation and plotted as in the figure.

x̂  Axˆ  Bu  K f ( y  Cxˆ )

u  K r x

xˆ  sI  A BU  sI  A K f y  sI  A K f Cxˆ

1 1 1

xˆ  BU  K f y  K f Cxˆ
 ( sI  A) 1

Solution of LQG is labrous for MIMO system and is in general done using computers.

Page 6 of 30
 u u

u y
PLANT
3 1 2

B
4 -

y C
Fig. Block diagram of linear quadratic Gaussian control

Page 7 of 30
Exercise 6.2.1: Consider the following SISO system.
x = ax + u
y = ax

Note the LQR: Cost function J r   ( x(t ) T Qx(t )  u (t ) T Ru (t ))dt whose optimal solution is
0

u(t)=–krx(t). k r  R 1 B T X such that X = X T ³ 0 ; where X is unique positive semi definite

solution for Riccati algebraic equation. AT X + XA - XBR-1BT X +Q = 0

a) Letting the control parameters, Q and R, respectively for error and control signal,
determine the optimal LQR controller in terms of Q and R.
b) Let initial value x(0)=10, a=1, Q = 1 and R=100, calculate the cost function.
c) Repeat b) for Q=100 and R=1 with other parameters remaining the same. Explain the two
results, that is result of b) and c).
d) Consider now, an LQG control for the system with Kalman filter constant K f. Determine
KLQG, which is loop gain from a point on control signal back to the same point on the
control signal.

Exercise 6.2.2: Room temperature modeling and Control for LQG Control
Consider a room as heated by radiator supplied by hot water. The radiator is installed on the wall of
the room and is continuously supplied by the hot water. Considering the following heat balances.
Radiator heat balance:
Change in radiator temperature = heat transferred from hot water to radiator – heat transferred to
room air. Let, x1 is radiator surface temperature, x2 is room air temperature, U is hot water temp
(heat) supply to radiator through hot water (flow rate assumed constant) and w is outside
temperature (to be considered as noise). Then;
x1  K 3 (U  x1 )  K 4 ( x1  x2 ) or

x1  ( K 3  K 4 ) x1  K 4 x 2  K 3U

Room Heat Balance:

Change in room temperature = Heat transfer from radiator to room – heat transfer from
room to outside.
x 2  K1 ( x1  x2 )  K 2 ( x2  ) or

Page 8 of 30
 x 2  K1 x1  ( K1  K 2 ) x2  K 2)
Overall model of the room temperature heating is then given by;

 x1   ( K 3  K 4 ) K4   x1   K 3  0
 x         U   
 2  K1  ( K1  K 2 )  x 2   0  K 2 
Characterization of outside temp disturbance as white noise:
Think of outside temperature as disturbance as periodic with period of 24 Hours. The
frequency in radiance of the fundamental, with largest amplitude, is   2 T . Therefore the

spectrum of the signal has peak at *   12 rad. This disturbance can be modeled as output
of a filter with pole at  j * input being white noise.
1
 v w , vw being white noise.
p  0.01p  0.068
2

p 2  0.01pw  0.068w  vw

Let p 2  x3  x3  p = x 4  x4  

Then, p 2  0.01pw  0.068w  vw

x 3  0.01x3  0.068  vw
x 4  x3

These two states can be added to the previous system, resulting a 4th order system.

 x1   ( K 3  K 4 ) K4 0 0   x1   K 3  0 
 x   K1  ( K1  K 2 ) 0    
K 2   x2   0  0 
 2    U   vw
 x 3   0 0  0.01  0.068  x3   0  1
        
 x 4   0 0 1 0   x4   0  0 
Measurement noise Inclusion:
Let us say measurement of room temperature for feedback is measured at two points. One of
the measurement instrument has calibration error n1, resulting y1  x2  n1 . The other

measurement is distributed by wind through window resulting y2  x2  v2 , v2 white noise.

1
Let n1  v n , vn being white noise.  pn1  v n  x 5  pn1  x5  n1
p

x 5  v n

Page 9 of 30
  x1   ( K 3  K 4 ) K4 0 0   x1   K 3  0 0 
 x     0 0 
 2  K1  ( K1  K 2 ) 0 K 2   x 2   0    v w 
 x 3    0 0 
 0.01  0.068 3 x   0 U   1 0  
  v
 x   0 0 1 0   x4   0 
 0 0  n 
   4
   
 x 5   0 0 0 0   x5   0  0 1

 x1 
x 
 y1  0 1 0 0 1   0 0 v w 
2

 y   0 1 0 0 0  x 3   0 1   v 

 2  x    n 
 4
 x5 

This is the overall model of the system which can be solved by LQG control.
Now, let K1=K2=1, K3=1/8 and K4 = 1/5. Let mean outside temp be 180C, and set room temp be
200C, radiator surface temp 220C, heating water temp 350C.

0.0001 0 0 
10 0 
Typical, W    ,V  0 0.1 0  ,
 0 1  0 0 0.001

(c) Design LQG control for the room temperature with weight Q=1 and R=0.01 and noise
characteristics as given above.
(d) Simulate for step response and indicate the performance.
6.2.3 Loop Transfer Recovery (LTR)
State feedback loop gain
Consider state feedback control system shown in figure below.

r y
G
- x
Kr

Fig. Feed back from measured state

If the states are measured for feedback the optiaml state feedback as descussed above would be;
u  K r x
x  Ax  Bu x
The loop gain from the point of brocken line on the figure back to the same pint can be shown to be
L  K r ( sI  A) 1 B .

Page 10 of 30
LTR is procedure of regaining ideal loop gain properties of K r ( sI  A) 1 B . LTR requires equal
number of inputs and outputs.
From this loop gain, the sensitivity, complementary function can be determined. For SISO system, the
phase margin and gain margins and other performance indices can be determined.
LTR can also be used to design LQG probmlem as summarized below.
Summary of LTR
Chose Kr as optimal LQ-state feedback for penality matrices Q and R which give a suitable dynamics
for the closed loop system. Veryfy that the ideal loop gain K r ( sI  A) 1 B gives good sensitivity and
robustness.

1. Chose observer gain Kf =B as optimal (KALMAN) gain. Increase  until the system has
desired sensitivity and robustness. Sufficiently large  gives result close to ideal one provided
system is minimum phase.
2. Another way to achieve Kf =B with large is to chose N = B and W = V with a large . As 

increases the Riccati equation results K f   B provided A-KfC is stable.

Note that choosing the optimal observer gain Kf directly instead of through Kf = B gives the
advantage that A-KfC is guaranted to be stable for any .
Exersise 6.2.3 Exercise on Loop Transfer Recovery
(1-1) Referring to figure above, determine the loop gain from point 1 back to 1.
xˆ  BU  K f y  K f Cxˆ
 BK r xˆ  K f Cxˆ  K f y u   K r xˆ

( I  BK r  K f C) xˆ  K f y

Multiplying both sides by -1

( 1  BK r  K f C ) xˆ  K f y

xˆ  ( 1  BK r  K f C ) 1 K f y

Multiplying by –Kr
 K r xˆ   K r ( 1  BK r  K f C ) 1 K f y

u

y  CBU   K r xˆ   K r ( 1  BK r  K f C ) 1 K f CBu
  
u K LQG G

Page 11 of 30
That is the loop gain K LQGG from point 1-1 and is given by K LQG   K r ( 1  BK r  K f C ) 1 K f and

G  CB .

(2-2) Determine gain between 2 and 2 as indicated on the figure

From the above calculation; U   K r ( 1  BK r  K f C ) 1 K f y

Multiplying by system transfer function G,

K r ( 1  BK r  K f C ) 1 K f y
  G 
GU
y   
K LQG

(3-3) between 3 and back to 3

x  Ax  Bu  X  sI  A BU
1

Multiplying by –Kr

 K r X   K r sI  A BU
1
 
u

Controller is given by K  K r sI  A1 B

(4-4) As exercise show that from point 4 to 4 is given by CK f which represents filter.

xˆ  Axˆ  Bu  K f ( y  Cxˆ)

xˆ  Bu  K f ( y  Cxˆ ) Multiplying by C;

Cxˆ  C
Bu
  CK f ( y  Cx)
ˆ
y

 ( y  Cxˆ)  CK f ( y  Cxˆ)  CK f is filter.

Exercise 6.2.4: A second order system LQG problem

 2 3 
 s  2 u (t )  w(t )
Given the following transfer function y (t )   s  1
1 1 
 
s 1 s 1
(a) Express in state space representation
(b) Solve an LQG problem with design parameters indicated.
Solution: (a)
The least common pole polynomial of the system can be determined as  ( s)  ( s  1) 2  ( s  2)
resulting a third order system.
The followings can be defined as state variables.
Page 12 of 30
 1 1 1
x1  u1 (t ) , x 2  u 2 (t ) , x3  u 2 (t )
s 1 s 1 s2
This results the following state space representation of the system without the disturbance w(t).

 x1   1 0 0 1 0
 x    0  1 0  x  0 1u
 2    
 x 3   0 0  2 0 1

2 0 3
y x
1 1 0
1
Let the measurement noise w be modeled as w(t )  v(t ) where v(t) is white noise.
s  0.001
x 4  0.001x4  v(t ) this same appear in the two outputs. Hence we have similar
x 5  0.001x5  v(t )

 1 0 0 0 0  1 0 0 0
 0 1 0 0 0   0 1 0 0
  
x   0 0 2 0 0  x  0 1  u  0 0 v
     
0 0 0  0.001 0  0 0 1 0
 0 0 0 0  0.001 0 0 0 1

 2 0 3 1 0
y x
1 1 0 0 1
Solution (b)
Given R1=I, R2=0.001 and N=C, Q=I and R=0.1I
Simulate on MATLAB for step response.
6.2.4 Summary of LQG control
(1) System plant G is given or known being determining factor for optimal control
(2) Weighting matrices Q and R also determine the result and trade-off between control efforts
and control error. It also determines how different components of control error are weighted
together in R.
(3) Disturbance model and reference signal model can also be considered as design variables. The
size of error and control signal depends on the disturbance model.
(4) Measurement variables y are controlled outputs, disturbed by measurement noise. Result of
the system depends on Q and R.

Page 13 of 30
 (5) Once control has been computed, all relevant properties of closed loop system have to be
tested. That is calculated and plotted Gc, T and S in the frequency domain. It may also be
necessary to test time to reference signals and disturbance by simulation. If results are not
satisfactory, other penalty weighting matrices Q and R have to be tried. Description of
disturbances may also be required to be changed.

Page 14 of 30
 6.3 Internal Model Controllers
Internal Model Control (IMC) encapsulates the process to be controlled in addition to its model as
shown in Figure below. The controller K is derived from transfer function Q and the plant model.
𝒅𝒚
K plant

r u + y
+
Q G
- +

- +
G
Model

Fig. 1 Internal Model Based Control Design

6.3.1 Controller K, Sensitivity S and Complementary Sensitivity functions from the IMC
The IMC philosophy relies on the internal model principle, which states: Accurate control can be
achieved only if the control systems encapsulates (either implicitly or explicitly) some representation
of the process to be controlled. As indicated in the figure, the controller K shown in broken line block
is made of Q and plant model G with input r and y. It’s output is control signal u. The transfer
function from r an y to u can be written as;
U  Qr  QY  QGU

( I  QG )U  QR  QY )

U  ( I  QG ) 1 QR  ( I  QG ) 1 QY

K  ( I  QG ) 1 Q or

K  Q( I  GQ) 1

Note that the two are equal from matrix identity.

IMC can be used to design an internally stable controller for stable K and Q. A stable appropriate Q
can be selected so that the sensitivity function, the complementary sensitivity function and the
controller can written as linear function of Q and checked for performance of the IMC.

Page 15 of 30
Extraction of the controller K, sensitivity and complementary sensitivity function can be solved from
the selected stable transfer function Q as follows for a stable system G to result an internally stable
control system.
Q  K ( I  GK ) 1  KS

( I  GK ) 1  I  GQ  S

( I  KG ) 1  I  QG = S I

G ( I  KG ) 1  G ( I  QG )

GS I  G( I  QG)
6.3.2 Controller Design using IMC
The controller design using IMC can be considered as having two steps. The first step is choosing
appropriate stable Q depending on the plant G while the second step is calculating the controller K
from the linear relation between K, G and Q as derived above.
An ideal choice of Q is Q  G 1 , which would result S = 0, K=∞ and loop gain equal to I. However, this
is practically impossible.
Furthermore, there are other concerns like,
 RHP-zero of G, which can result RHP-pole of Q;
 If G is proper, its inverse results non-causal transfer function.
The followings can be used to select Q in various cases with n chosen based on the order difference
between numerator and denominator of G.
Case I: G has more poles than zeros and the inverse results physically non realizable system. For
such cases choose;
1
Q( s )  G 1 ( s)
(s  1) n

Where n is chosen so that Q(s) can be realizable (numerator degree < denominator degree). The
number  is design parameter that can be adjusted to desired bandwidth of the closed loop system.
Case II: G has an unstable zero (non-minimum phase)
The existence of such a zero results cancelation of RHP-pole when the system and its inverse are
multiplied during calculation of K and results non internally stable system.
If the RHP-zero is given by ( s  1) in the numerator, we can approach in two ways.

(a) Ignore the RHP-zero factor when Q  G 1 is formed.

Page 16 of 30
 (b) Replace this factor with (  s  1) when Q  G 1 is formed. This results only a slight phase
error with no magnitude error.
In both cases the original G is used when the controller is formed as the linear combination of G
and Q.
Case III: G has time delay (the factor e  ST )
For this case also we have two possibilities.
(a) Ignore this factor when forming the Q  G 1 and use e  ST when forming the controller K.
1  sT / 2
(b) Approximate this factor with e  ST  . In this option, we handle the RHP-zero as
1  sT / 2
described in case II above. The same approximation can be used for e  ST when forming K.
Example 6.3.1: For SISO system
1 s  1
Let the system be G ( s )  . Then, Q ( s ) 
s  1 s  1
 1
Which results K  (1  )
 s
Accidentally, this is a PI controller.
Example 2: SISO with RHP-zero
 s  1
G (s) 
s  1

Q1 ( s)  (s  1)
(s  1)

Q2 ( s)  (s  1)
( s  1)
The corresponding controllers are;
s  1 s  1
K1 (s)  for Q1 ( s) and K 2 ( s) 
(   ) s 2s

Example 6.3.2: SISO system with Delay e  sT

Let G  G1e  sT ,
Case 1, ignoring the delay and selecting appropriate Q for G1 without the delay. The resulting
Q( s )
controller will be; K 
1  Q( s)G1 ( s)e  sT

Page 17 of 30
6.4 H∞ Control
In LQG control formulation and controller design use of plant model is necessary. The performance
from the control, hence, depends on the plant model certainty. On the other hand, in H∞ control,
robustness is considered.
6.4.1 General Linear Control Configuration
To utilize H∞ Control, a control system has to be formulated into general linear control configuration
as shown in figure below.

Fig. General Linear Control Configuration

Where; P – is generalized plant model. It will include G and Gd and the interconnection structure
between the plant and the controller. In addition, if P is being used to formulate a design problem,
then it will also include weighting functions. K - is controller

w – is weighted exogenous inputs; commands, disturbances and noise.

z – weighted exogenous outputs; “error” signals to be minimized, e.g.; y-r.
u – Control signals signal for the general configuration, e.g., commands,
v - measured plant outputs, measured disturbances, etc. For the special case of one degree of
freedom controller with perfect measurements we have v = r – y.

The overall control objective of the H∞ control is to minimize (usually make less or equal to 1) the
𝑯_∞ norm of the transfer functions (like S, KS, T) from w to z. The problem then is, find a controller
K which based on the information in v, generates and input u, which counteracts the influence of w
on z, thereby minimizing the norm from w to z. Output v is feedback to controller K to generate
control signal u resulting robustness in H-infinity control.

Exercise 6.4.1: Consider a unity feedback control system block diagram shown below and convert it
the generalized linear control configuration.

Page 18 of 30
 Figure _ Unity feedback Control system
The inputs to the system are r, d, and n which can represented by vector w in the generalized

d 
configuration. Let the vector w be represented as; w   r  . Similarly, let the output z be the error
 n 

between y and r (z=e-y-r).

Now, using the generalized configuration we have . In the unity feedback system,
z=-y-r and v=r-n-y. Therefore; we can write the following.

d 
z  I I 0 G   r  P P12 
Which can be written as; v    I , which means P   11 ,
   I I  G   n   P21 P22 
 
u 
P11  I  I 0 , P12  G , P21  G , P21   I I  I

Fig. ____ Generalized Linear Control Configuration

 y  r r 
Exercise 6.4.2: Now consider two degree of freedom where z    , v  y  .
 u   m
Page 19 of 30
The plant model, therefore, can be given as P or as state matrices of the state space representation
of the system.

x  Ax  B1u  B2 w , z  C1 x  D11u  D12 w , v  C2 x  D21u  D22 w and hence P can also be

A B1 B2 
expressed as P   C1 D11 D12 
C 2 D21 D22 

The closed loop transfer function from w to z is given by the linear fractional transformation which
can be expressed as;

z  Fl ( P, K ) w(t ) where; Fl ( P, K )  P11  P12 K ( I  P22 K ) 1 P21

H∞ control involves the minimization of the H∞ norm of Fl ( P, K ) .

The followings are assumptions in the solution of H∞ problems.

(1) ( A, B2 , C2 ) stablizable and detectable.

(2) ( D12 D21 ) have full rank.

 A  jI B2 
(3)  has full column rank for all w.
 C1 D12 

 A  jI B1 
(4)  has full row rank for all w.
 C2 D21 

(5) D11=0 and D22=0

0 
(6) D12    and D11  0 I 
I 
(7) D12T C1  0 , B1 D21
T
0

(8) ( A, B1 ) is stablizable and ( A, C1 ) is detectable.

6.4.2 Solution for H∞ Control

As stated earlier, once the problem is formulated to the general linear control configuration, H∞
optimal control problem is to find stabilizing controller K which minimizes the peak of the maximum
singular value of the closed loop transfer function over the effective frequency.

Page 20 of 30
 Fl ( P, K ) 
 max  ( Fl ( P, K )( j))


Usually, optimal value (the minimum possible) determination is time consuming. Suboptimal values
may be sufficient and simpler.

The sub-optimal control problem is given as acceptable maximum value, which we can denote as 
and hence, the control design is determining controller K, which minimizes the closed loop transfer
function iteratively until lower or equal to the allowable as shown below is attained.

Fl ( P, K ) 


Glover-Doyle Algorithm for Solution of H∞ Problem

The Glover-Doyle algorithm solves a family of stabilizing controllers such that Fl ( P, K )   . In

general, commercial algorithms are used to solve the H∞ problems. The problem can be given in the
generalized linear control configuration and or state space representation. It also to be noted that
the system to be fulfilled as listed above have to be valid.

The solution of H∞ control requires the solving of two Riccati equations, one for controller and one
for the observer.

Control law: U  K c x

The state estimator equation by:

   
x  Ax  B2 u  B1 w  z  K e ( y  y)

 
w   2 B1T X  x

  
y  C 2 x   2 D21 B1T X  x

The controller Kc is as for LQG case while the estimator gain is z  K e instead of Kc in LQG.

z   ( I   2Y X  ) 1

X   0 and Y  0 are solutions to Riccati equations. Refer text for Riccati equation, conditions
and other details.

Page 21 of 30
 6.4.3 Mixed Sensitivity H∞ Control and Weighting matrices

Mixed Sensitivity

Mixed sensitivity H∞ control refers to minimization of sensitivity function shaping along with one or
more other closed loop transfer functions such as complementary sensitivity function (T) and
controller times sensitivity (KS).

By so doing, we are defining corresponding objectives as follows.

i) Controlling over sensitivity (making it small), the controller disturbance rejection at low
frequency with high gain in low frequency and low gain at high frequency is targeted.
ii) Controlling KS, the control system robustness and low control signal can be achieved.
Similarly,
iii) Control over complementary sensitivity function result control over measurement noise,
especially at high frequency.

Weighting matrices

The closed loop transfer function Fl ( P, K ) is defined in terms of the closed loop transfer functions S,

T, KS and the weighting matrices Wx which are used to penalize the corresponding control. In other
words, weighting matrices Wx are introduced at the outputs to force the transfer functions to be
shaped “small” for all frequencies simultaneously, which is otherwise impossible. The Wx are design
parameters.

The closed loop transfer function can be defined in terms of the weighting matrices as follows.

 Ws S 
Ws S 
Fl ( P, K )   for three outputs; Fl ( P, K ) WT 
 WT   
Wu KS 

The objectives of weighting matrices selection is first to suppress the singular value of the sensitivity function
as much as possible for the largest possible bandwidth by large loop gains. Secondly, the singular values of the
complementary sensitivity function must be suppressed by 20dB at cut-off frequency and second order fall off
(-40dB) for frequency above the cut-off frequency. The zero dB cross-over frequency of Ws must be sufficiently
bellow the zero dB cross-over frequency of WT.

Page 22 of 30
 0.01s  1
I , Wu  0.001I , WT   2 I
1 1000
Example; Ws  0.01
1  100 s

Example 6.4.3: Regulating Problem formulation for mixed sensitivity

d
+ y
r+ e

Consider the unity feedback system block diagram with output disturbance as shown. Formulate the
system for mixed sensitivity H∞ control such that disturbance is rejected and control signal is
minimized.

Solution:

y  Gu  d u   Ky

u  K (r  y)   Ky r 0  K ( I  GK ) 1 d

u  KSd
y  GKy  d  y  ( I  GK ) 1 d
y  Sd
z1  ws d  ws u w  w G 
P11   s  P12   s  P21  I  P22  G
z 2  0  wu  0 ,  w , ,
y  Id  Gu

Example 6.4.4: Tracking problem with disturbance equal to zero (Ws and W weighting matrices).

Page 23 of 30
w r yr e u  K (r  y ) re
r u, r u  K (r  Gu )  ( I  KG )u  Kr e  (r  y )  r  Gu
d 0 d 0 u  ( I  KG ) 1 Kr u  K (r  y )
 KSr e  r  GK (r  y )  e  Sr

z1  Ws ( Ir  GU ) z 2  0  WU V  w  GU

Example 6.4.5 Given the generalized linear system with augmented plant P(s)

(a) Write the equation for P

(b) Write the closed loop transfer function

Ws Y1
e
w +
Wu Y2

u -
G W Y3

Page 24 of 30
 LQG for SISO System
1
Consider a first order system with transfer function G ( s )  . Its state-space realization is
sa
𝒙̇ = 𝒂𝒙(𝒕) + 𝒖(𝒕), y(t) = x(t)
a) Determine the control gain considering LQR with measured available state variable. Assume
Q=1 and R=R.
b) Let the Q=Q while R=1. Determine the controller gain for LQG control.
Solution for a:

J r   ( x 2  Ru 2 )dt
0

The corresponding Riccati equation becomes, aX  Xa  XR 1 X  1  0

 X 2  2aRX  R  0 . Solution for this second order system is; X  aR  (aR) 2  R =

X  aR  aR 1  1
a2R
The optimal control is given by u   K r x .

X 1
Where K r   a  a 1 2 .
R a R

1 1
The closed loop system is then, x  ax  ax  a 1  2
x that is; x  a 1  2 x .
Ra a R

1 1
For stability, the closed loop pole is located at s  a 1  2
 0 . That is a 1  2 should be less
a R a R
than zero.
The root locus of the pole for stability starts at absolute value of a for R infinity and approaches
negative infinity along the real axis as R approaches zero.
Solution for b: Complete this??

Page 25 of 30
Appendix I: Signal Description
Linear systems specially when in multivariable form require proper description. The signals “Size”,
frequency content of the signals and expression of signals as an output of linear systems input being
“white noise” or an impulse are important characterization of signals.
1.1 Signal “size”
1.1.1 2-norm
The 2-norm of a signal u(t) can be obtained for both scalar and vector u(t) in the following
form.
 2
  u (t ) dt
2
Scalar u(t): 2-norm of u(t) expressed as u (t ) 2 

 u1 
2-norm of u(t)= u 2  , u (t )

  u T (t )u (t )dt . This results
2
Vector u(t): 2 
 ... 
 
  u T (t )u (t )dt =  u12 (t ) u 22 (t )  ...dt form can be described by 2-norm. If the
2
u (t ) 2  

1.1.2 Matrix measure of a vector signal

A vector signal can be measured as a matrix measure of the vector signal. Matrix measure has
various interesting characteristics and is given by the following expression for a vector signal
u(t).

Ru   u (t )u T (t )dt


 u1    u u1u 2 
2
Example: let u(t)=   then, Ru    1
 dt
 u u u 22 
u 2   1 2
Matrix measure of a vector signal corresponds to “covariance matrix” of the signal.
Characteristics of Ru:
i) If Ru is known for a signal u(t), we can calculate the size of a number of related (linear
combination) signals.

  Lu (t )u (t ) T LT dt  L  u (t )u (t ) T dtLT =LRuLT
2
Example: Let S(t) = Lu(t), then S 2

ii) The size of the kth component of u is given by the kth diagonal element of Ru. That is

 tr ( Ru ) that is trace of Ru.

2
u 2

1.1.3 Power Measure

Page 26 of 30
 Power measure is another measure of signal. It is average power measure which usually results
finite even when the energy signal size is infinite.
N
1
 lim  u(t )
2 2
u (t ) dt
e N  N
1

The size of a signal may be related to frequency content of the signals.

1.2 Spectral Description of Signals
Any signal, in general can be expressed as integral of all its frequency content. Hence, spectral
description refers to description of the signals in terms of their frequency content.
1.2.1 Energy Spectrum
For example u(t) can in general be expressed as;

U(t) u (t )   A( ) cos(t   ( ))d where A() is amplitude of the cosine function at frequency


 of the signal u(t).

Definition: The energy spectrum of u(t) denoted by u ( ) is then the expected value of A 2 ( ) .

An m dimensional signal u(t) can be associated with Hermitan mxm matrix function u ( ) called

its spectrum.

 u1   1 ( ) 12 ( )
Example: U    have  
u 2   21 ( )  2 ( ) 
The diagonal 1 and 2 are spectrum of u1 and u2 respectively. The off diagonal blocks 12 and

21 are the cross-spectrum of u1 and u2. u is hermitan and hence 12   21* .

If the two signals are uncorrelated, the off diagonal elements are zero, 12   21
*
 0.

Size of a signal from its spectrum, for example for u(t), is given by;
1 
Ru 
2 

u ( )d

Note on special signal white noise; “White noise” is a special signal having constant matrix
spectrum. The measure of size of white noise signal is infinite.
For signals with finite energy its size can be related in its frequency and time domain model.

u ( )  U ( j )U * ( j ) where, U ( j )   u (t )e  jt dt


Page 27 of 30
 1  
Ru  
2

 
 u ( )d   

u (t )u (t ) T dt

Parservals identity

1.2.2 Spectrum of Output signal of a linear System

If u(t) is input to a system with transfer function G(s) and y(t) is the corresponding output,
then the energy spectrum of y(t) denoted as  y ( ) is given by;

 y ( )  G( j )u ( )G * ( j )

Example:
y(t )  G( p)u(t )  H ( p)e(t )
u(t )  u (t )

Let us say u(t) and e(t) be uncorrelated scalar signals with spectra u ( ) and e ( )   .

u (t )  u ( ) 0   y(t ) G( p) H ( p) u(t )

 (t )    ,     , Z (t )   
 e (t )   0  e ( )    u(t )   1 0   e(t ) 

  y ( )  yu ( )
 z ( )   
 yu ( ) u ( ) 

G ( j ) H ( j ) u 0   G (  j ) 1 
=
 1 0   0 e   H ( j ) 0

G ( j )u G ( j )  H ( j )e H ( j ) G ( j )u 

=
  u G (  j ) u 

“Size” of output signal then is;

1

1

G ( j )u G ( j )  H ( j )e H ( j ) G ( j )u 
Rz 
2   z ( )d = Rz 

2  

 u G (  j ) u 
d

Spectrum and size of state;

x  Ax  Nv
1
If A is stable;  x  ( jI  A) 1 NRN T ( jI  A) T
x  ( sI  A) Nv
1 
The covariance matrix for x is  x  R x 
2 

 x ( )d

Theorem (state size): Let  x be defined by equations x  Ax  Nv

 x ()  ( jI  A) 1 NRN T ( jI  A) T

1 
 x  Rx 
2 

 x ( )d

Page 28 of 30
 Where all eigen values of A are strictly inside the stability region. Then  x is the solution to

the matrix equations.

A x   x AT  NRN T  0

Example:
 1 2 1
x    x   v where v is white noise with variance 1.
  1 0 0
A x   x AT  NRN T  0

 1 2  11  12   11  12   1  1 1 0 0 
 1 0         1 0   ; R=1
   21  22   21  22   2 0  0 0 0 

1 
 0
  x  2
1
0 
 4
1.3 Observer and Kalman Filters
From modern control state space theory of control, if the state variables are available from
measurement control signal can be constructed from the feedback states as u  Kx for
x  Ax  Bu .
x  Ax  BKx  ( A  BK ) x

Appropriate gain K can be determined a stable feedback control system by placing eigen values of
A-BK in the left hand side of the complex plane. However; it is not always easy to measure the
state variables and sometimes they are not accessible or nonphysical quantities. Therefore, state
variables are usually estimated from a known control signal u and the measured output y.
Error between the estimated and the actual states can be formulated as follows.
x  Ax  Bu  Nd
  
x  Ax  Bu  K f ( y  Cx  Du )

Subtracting one from the other;

  
x  x  A( x  x )  Nd  K f (Cx  Du  n  Cx  Du )
   
x  x  ( A  K f C )( x  x )  Nd  K f n , Let x  x  ~
x

x  ( A  K f C ) ~
~ x  Nd  K f n

From this equation we can see two effects of the estimator gain Kf.

Page 29 of 30
 (i) It determines how fast the error dies out in (A-KfC)
(ii) It determines the effect of the noise on error, amplifying the noise.
The selection of Kf is then, the trade-off between the two.

x  ( A  K f C ) ~
From the theorem above and the error state equation ( ~ x  Nd  K f n , we can also

have the expression for the size of the error.


R~x   ~
x~x T dt    ~x ( )d


Where;  ~x  ( jI  A  K f C ) 1 LRLT ( jI  A  K f C ) T , L  N  Kf 

 R1 R12  R R12   N T 
Let R   T
R2 
, then   N
LRL T
 
 K f  T1  
R2   K Tf 
 R12 R~x  R12

R~x  NR1 N T  K f R12T N T  NR12 K Tf  K f R2 K Tf

The magnitude P of the estimate ~

x ,for stable (A-KfC) is therefore the solution of the equation;
( A  K f C ) P  P( A  K f C ) T  R x  0 =

( A  K f C ) P  P( A  K f C ) T  NR1 N T  K f R2 K Tf  K f R12T N T  NR12 K Tf  0

XXXX
The minimal estimation error covariance matrix is given by E ( ~
x (t ) ~
x T (t )  P . The observer with

this choice of Kf is called the Kalman filter while the following algebraic equation is called the
Riccati equation.
AP  PA  ( PC T  NR12 ) R 1 ( PC T  NR12 ) T  NR1 N T  0

Page 30 of 30