Lecture notes on Controllability
By
N. Sukavanam
Department of Mathematics
IIT Roorkee
Controllability of finite dimensional control systems
Dynamical System
A System is a collection of objects which are related by interactions and produce various outputs in response to
different inputs.
If a system changes with respect to time then it is termed as Dynamical System
For example Electro-mechanical machines such as motor car, aircraft, or spaceships, Biological systems such as the
human body, Economic structures of countries or regions, Population growth in a region are dynamical systems
A mathematical model of a dynamical system is represented by
d xi
=f i ( t , x 1 , x 2 ,. . . , x n ) :i=1, 2,. .. ,n
dt
where
fi:R
n +1
R ; i=1,2, , n
x i ; i=1,2, , n
(1)
are some nonlinear functions.
are called state variables of the dynamical system.
Example: An equation of the form
dx
=x , x ( 0 )=x 0 , 0 t T
dt
is a linear dynamical system. The solution is given by
x ( t )=exp ( t )x 0
Example :
The system described by
x 1=x1 2 x 1 x 2
x 2=2 x 2+ x1 x 2
is a nonlinear dynamical system.
The above nonlinear system represents a mathematical model of prey-predator population known as Volterra-Lotka
model.
Consider the equation
x ' =f ( t , x ) ,t >0 : x ( 0 )=(2)
where f
is assumed to be integrable with respect to time,
Theorem 1 :Assume that the nonlinear function
is Lipschitz continuous in the 2
nd
variable and cotinuous in the
st
1 variable. Then equation (2) has a unique solution.
Control system
If a dynamical system is controlled by suitable inputs to obtain desired output then it is called a Control System
d xi
=f i ( t , x 1 , x 2 ,. . . , x n ,u 1 , u2 , .. . ,u n ) :i =1,2, ... , n
dt
Here
x i ; i=1,2, , n
represent the state of the system and
ui ;i=1,2, , n
are the control variables.
Example: Consider the control system
dx
=x +u ; x ( 0 ) =x 0 0 t T ,
dt
The solution is
x ( t )=exp ( t ) x 0+ exp (s ) u(s) ds
0
Now the control function
position at
t=T .
u(t) (input) can be chosen suitably so that the solution (output) has a desired final
If the desired value of x ( T )=x 1
u (t)=
u(t ) can be choosen as
the
exp ( (st 0 ))
[ x 0 + x 1 exp ( (tt 0 ))]
T
Example: Suppose a particle is to be moved along a straight line and let its distance from an initial point 0 be s (t )
at time t . For simplicity assume that the particle is controlled only by a force
only factor of interest are the particles position
which describe the state of the object at time
x 1 ( t )=s(t)
and velocity
u(t )
per unit mass. Suppose that
x 2 ( t )=s (t ) . Then the equations
are
x 1=x 2
x 2=u
or in matrix notation
x = Ax+ Bu
where
[]
x= x 1
x2
[ ]
A= 0 1
0 0
[]
B= 0 .
1
In practice there will be a limit on the values of u for obvious reason, and it also be necessary to impose restrictions on
the magnitudes of velocity and acceleration.
It may then be required to start from rest at 0 and reach some fixed point in the least possible time, or perhaps
with minimum consumption of fuel. The mathematical problems are firstly to determine whether such objectives are
achievable with the selected control variable, and if so, to find appropriate expressions for
and / or
x 1 and
x2
as a function of time
Solution of linear system
Consider the system
x = Ax ,
with
x ( t 0 )= x0 .
(3)
The solution is given by
tt 0
A ()
x ( t )=exp
We define the exponential matrix by
exp ( At )=I +tA +
The series on the right converges for all finite
and all
t 2 A 2 t 3 A3
+
+
2!
3!
n n
matrices
that
( At )
exp
exp ( 0 )=I and
so that it represents the solution of (3).
d
State Transition Matrix
The solution of (3) subject to the above initial condition is often written in the form
x ( t )= ( t , t 0 ) x 0
having finite elements. It is clear
where
( t ,t 0 )=exp A (tt 0 ) is called the state transition matrix, since it relates the state at any initial time
t 0 to the state at any other time t . The state transition
matrix has the following properties:
(1)
d
( t , t 0 )= A ( t ,t 0 )
dt
(2)
( t , t )=I
(3)
( t 0 ,t )= ( t , t 0 )
(4)
( t ,t 0 )= ( t , t 1 ) ( t 1 ,t 0 )
Time varying systems
x = A ( t ) x ( t ) , x ( t 0 )=x 0 (4)
It is then easy to verify by direct differentiation that
x ( t )= ( t , t 0 ) x 0
is the solution of (4) with initial condition
x ( t 0 )= x0 .
where ( t ,t 0 )= X (t ) X ( t 0 ) (5)
and X ( t )
is the unique
n n matrix satisfying
dX
=A ( t ) X ( t ) , X ( 0 ) =I .(6)
dt
In time varying case we can no longer define matrix exponential of A, but there is a result corresponding to the fact
exp( At )
that
is nonsingular when
is constant.
X (t )
Theorem : 2 The matrix
is nonsingular.
dY
YA(t ),
Y (0) I
dt
Y (t )
Proof: Define a matrix
as the solution of
Such a matrix exists and is unique.
Now
d
& YX&
(YX ) YX
dt
YAX YAX 0
Y (t ) X (t )
So
is equal to a constant matrix, which must be the unit matrix because of the condition at
X (t )
t 0
. Hence
Y (t )
is nonsingular and its inverse is in fact
(t , t0 )
It is most interesting that although in general it is not possible to obtain an analytic expression for
, this
matrix possesses precisely the same properties as those for the constant case. A further correspondence with the time
invariant case is the following generalization of constant matrix case.
Peano-Baker series : It can be seen that RHS of (5) is equal to the power series known as Peano-Baker Series given
below.
t
t t1
t n1
( t , t 0 )=I + A ( s ) ds ++ A ( t 1 ) A ( t 2 ) A ( t n ) d t n d t 1 +
t0
t 0 t0
t0
Finite dimensional linear control systems
Consider the linear control system defined as
x Ax Bu
x ( 0) x 0
x (t) Rn
where
u(t) R m
and
(7)
for each t
[0,T], A is
nn
and B is
n m
matrices.
The solution of the system is written as
t
x ( t )=exp ( At ) x 0+ exp ( A (ts ) ) B u ( s ) ds
0
Controllability: The system is said to be controllable in the time interval [0,T] if for any
and
find a control vector function u(t) such that the solution of (7) corresponding to u(t) satisfies
x (T )=x 1 .
Kalman Condition (1963)
The system (1) is controllable iff rank of the matrix .1
n = [B AB A2B.An-1B]
2.The system (1) is controllable iff the controllability Grammian matrix
T
( t , s )BB
U=
( t , s ) ds
Example: Consider the control system
dx
=x+u ; x ( t 0 )=x 0 , 0 t T
dt
is nonsingular.
x R n
in
x (0)=x 0
one can
and
The solution is
Now the control function
t0
u(t) (input) can be chosen suitably so that the solution (output) has a desired final
t=T .
position at
If the desired value of x ( T )=x 1
u (t)=
x ( t )=exp ( (tt 0)) x0 + exp ( ( st 0) ) u( s) ds
the
u(t ) can be choosen as
exp ( (st 0 ))
[ x 0 + x 1 exp ( (tt 0 ))]
T
Theorem 3: The constant system
x& Ax Bu
is completely controllable if and only if the
n nm
(8)
controllability matrix
U B, AB, A2 B,....., An 1B
(9)
n
has rank
Proof: Necessity: We suppose that (8) is completely controllable and wish to prove
assuming
rankU n
rankU n
, which leads to a contradiction. For then there will exist a constant row
. This is done by
q0
vector
that
qB 0, qAB 0,....., qAn 1 B 0
(10)
such
x(t ) exp( At ) x0 exp( A ) Bu ( )d
0
x (0) x0
The solution of (8) subjected to
t t1
Put
is
x(t1 ) 0
and
we get
0 exp( At1 ) x0 exp( A ) Bu ( ) d
0
t1
t1
x0 exp( A ) Bu ( ) d
exp( At1 )
Since
is nonsingular
exp( A )
Since
r ( A)
can be expressed as some polynomial
in
( n 1)
having degree at most
we get
t1
x0 ( r0 I r1 A ..... rn 1 An 1 ) Bu ( ) d
0
qx0 0
q
Multiplying on the left by
, using (10), we get
q0
x0
Since (8) is completely controllable this must hold for any vector
assumption that
rankU n
Sufficiency: we now assume
, which implies
, this contradicting the
rankU n
x (0) x0
, again the solution of (8) subjected to
x(t ) exp( At ) x0 exp( A ) Bu ( ) d
0
t t1
Put
t1
exp( At1 ) x(t1 ) x0 exp( A ) Bu ( )d
0
is
t1
exp( At1 ) x(t1 ) x0 r ( A) Bu ( ) d
0
or
(11)
Where the coefficients of
will be the functions of
. Carrying out the integration in (11) will produce
exp( At1 ) x(t1 ) x0 ( s0 B s1 AB ..... sn 1 An 1B)
(12)
Since
rankU n
x0
it follows that for any given
si
it will be possible to choose the
u ( )
(and hence by implication
x(t1 ) 0
) so that the right hand side of (12) is identical zero, giving
. This establishes that (8) is completely
controllable since the condition of the definition are satisfied.
Theorem-4 gives a criterion for determining whether a constant linear system is completely controllable, but gives no
help in determining a control vector which will carry out a required alteration of states. We now give an explicit
expression for such a vector for both constant and time varying systems.
S1
Theorem 4: The system
x&(t ) A(t ) x(t ) B(t )u (t )
(13)
is completely controllable if and only if the
nn
symmetric controllability matrix
t1
U (t0 , t1 ) (t0 , )B( ) BT ( )T (t0 , )d
t0
(14)
(t , t0 ) X (t ) X 1 (t0 )
where
is nonsingular. In this case the control
u (t ) BT (t )T (t0 , t )U 1 (t0 , t1 )[ x0 (t0 , t1 ) x f ]
(15)
t0 t t1
Defined on
x(t1 ) x f
x(t0 ) x0
, transfers
to
U (t0 , t1 )
Proof of Sufficiency: If
is assumed nonsingular then the control defined by (15) exists. It is then straight
S1
forward to show that
is completely controllable.
x(t0 ) x0
The solution of (13) with initial condition
is
x(t ) (t , t0 ) x0 (t0 , ) B( )u ( ) d
t0
Substituting (15) for u(t) in the above equation we get
x(t ) (t , t0 ) (t0 , t1 ) x f
t t1
x(t1 ) (t1 , t0 ) (t0 , t1 ) x f x f
At
S1
Hence
is completely controllable.
S1
U (t0 , t1 )
Necessity: we need to show that if
is an arbitrary constant column
is completely controllable then
is nonsingular. First notice that if
U
vector then from (14) since
is symmetric we can construct the quadratic form
t1
U T ( , t0 ) ( , t0 )d
T
t0
t1
|| || d
2
e
t0
( , t0 ) BT ( )T (t0 , )
where
(16)
U (t0 , t1 )
, so that
T U (t0 , t1 ) 0
such that
. Eqn. (16) with
is positive semi-definite. Suppose there exists some
when
then implies
t1
|| || d 0
2
e
t0
which implies that
( , t0 ) 0, t0 t1
.
S1
However, by assumption
x(t1 ) 0
v (t )
is completely controllable so there exists a control
, say, making
if
x(t0 ) x0
. Hence we have the solution of (13)
t1
x(t1 ) (t1 , t0 ) x0 (t0 , ) B( )u ( )d
t0
It implies that
t1
(t0 , ) B ( )u ( )d
t0
Therefore
t1
|| || vT ( ) BT ( )T (t0 , ) d
2
e
t0
t1
vT ( )( , t0 )d
t0
Which contradicts the assumption that
U (t0 , t1 )
. Hence
is positive definite and is therefore nonsingular.
( x f , t1 )
( x0 , t0 )
The control function (15) which transfers the system from
to
requires calculation of the transition
matrix and the controllability matrix (14). This is not too difficult for constant linear systems. Of course there will in
general be many other suitable control vectors which achieve the same result, but the expression (15) has an
additional interesting property:
u (t )
Theorem 5: If
( x f , t1 )
( x0 , t0 )
is any other control taking
t1
to
then
t1
|| u ( ) || d || u ( ) || d
2
e
2
e
t0
t0
u ( )
Where
is given by (15), provided
u
Proof: Since both
u u
u
and
satisfy (13), we obtain after subtraction
t1
0 (t0 , )B ( ) u ( ) u ( ) d
t0
Multiplication of this equation on the left by
x0 (t0 , t1 ) x f
U 1 (t0 , t1 )
and use of (15) gives
t1
( ) u ( ) u ( ) d 0
t0
(17)
Therefore
t1
t1
t0
t0
T
2
2
T
(u u) (u u )d || u ||e || u ||e 2u u d
t1
|| u ||e2 || u ||e2 d
t0
Using (17),
t1
|| u ||
t1
2
e
t0
d || u ||e2 || u u ||e2 d
t0
t1
|| u ||e2 d
t0
provided
u u
, as required.
This result can be interpreted as showing that the control (15) is optimal in the sense that it minimizes the integral
t1
|| u ( ) ||
2
e
t0
t1
d u12 u22 ..... um2 d
t0
( x f , t1 )
( x0 , t0 )
over the set of all controls which transfer
to
, and this integration can be thought of as a measure of
control energy involved.
If the system is not completely controllable, it would be misleading to call it uncontrollable, since the implication of
our definition is that for a non-completely controllable system there are only certain final states which cannot be
achieved by any choice of control. We can however modify the argument used in Theorem 5 to show how such final
S1
states can be attained when
is not completely controllable.
xf
Theorem 6: If, for a given
, there exists a constant column
vector
such that
U (t0 , t1 ) x0 (t0 , t1 ) x f
(18)
Then the control
u (t ) BT (t )T (t0 , t )
x(t1 ) x f
x(t0 ) x0
Transfers the system (13) from
to
x(t0 ) x0
Proof: The solution of the controlled system with
is
x(t ) (t , t0 ) x0 (t0 , ) B( )u ( )d
t0
Put
u (t ) BT (t )T (t0 , t )
x(t ) (t , t0 ) x0 (t0 , ) B( ) BT ( )T (t0 , ) d
t0
t
t t1
Put
x(t1 ) (t1 , t0 ) x0 (t0 , ) B( ) BT ( ) T (t0 , ) d
t0
t1
Using (14)
x(t1 ) (t1 , t0 ) x0 U (t0 , t1 )
Using (18), we get
x(t1 ) (t1 , t0 )(t0 , t1 ) x f
xf
, as required.
S1
Since the system
u (t )
U
is completely controllable,
is nonsingular and the expression for
in Theorem 7 reduces
xf
to (15). It can also be shown that the converse of this Theorem is true, namely that only states
holds can be reached.
for which (18)
x0 x(0)
Theorem 7: A given state
subspace spanned by the columns of
xf
can be transferred into another state
U B, AB, A2 B,....., An 1B
xf
x0
provided both
and
lie in the