Sei sulla pagina 1di 13

Electrical and Computer Engineering Department University of Maryland, College Park ENEE 660 System Theory Fall 2008

Professor John S. Baras Solution to Homework Set #3 Problem 1


Given n n matrix A and n m matrix B controllability fails to hold when rank[B | AB | A2 B | An1 B] < n This implies that all n n submatrices of this n nm matrix also have rank < n. That is they (nm)! are all singular. There are nm = n!(nmn)! such n n submatrices of the controllability matrix. n Then nm det(Mk ) = 0, k = 1, , n for all such submatrices. But these are polynomial equations in the parameters, which here are the 2 elements of the matrices A and B. Therefore the set of all matrices (A, B) in IRn IRnm which fail to form a controllable pair lie on a variety, dened by these polynomial equations from the determinants of all n n submatrices of the controllability matrix. Hence controllability is generic. Observability of (A, C) is the same as controllability of (AT , C T ). So by the same argument observability is generic also. Minimality of (A, B, C) is equivalent to controllability of (A, B) and observability of (A, C). So (A, B, C) fails to be minimal on the subset of the parameter space 2 IRn I nm IRnp dened as the simultaneous solution set of the polynomial equations obtained R by asking that the determinants of all n n submatrices of the controllability or of observability matrices vanish. So minimality property of (A, B, C) is also generic.

Problem 2
Since the pair [A(), B()] is controllable, there exists for any tf a control utf () such that the solution of x(t) = A(t)x(t) + B(t)u(t) x(to ) = xo satises x(tf ) = 0. Then consider the control u(t) = utf (t), to t tf 0, tf < t 1

The corresponding cost is J() = u


tf to

[xT (t)Q(t)x(t) + uTf R(t)utf (t)]dt t

which is nite since all functions are bounded functions of t. Clearly the optimal cost satises Jopt = min J(u) J() u
u

Note also that for any control u(), J(u) is a nonnegative number since it is the integral of a nonnegative function, due to the assumptions on Q(), R(). Next observe that from the general solution of a linear system x() is continuously dependent on u(). J(u) depends continuously on x() and u() (in fact it is quadratic in x() and u()). So J(u) depends continuously on u(). Since 0 J(u) J() these considerations imply that there exists u an optimal control.

Problem 3
0 1 ,b = 1 0 k a constant in [0, 1]. Let A = 1 0 ,c = 0 . Let C be the controllability matrix for (A, b + kc), for 1

C = [b + kc, A(b + kc), , An1 (b + kc)] = detC = 1 + k 2 = 0 for 0 k 1. So controllability of {A, b + kc} holds for each constant k in [0, 1]. Now let 0 k(t) 1. Then, W (0, t) = = =
t 0 t 0 t 0

1 k k 1

cos sin sin cos

1 k()

1 k()

cos sin sin cos

d d d

cos + k() sin sin + k() cos

cos + k() sin sin + k() cos

(cos + k() sin )2 (cos + k() sin )( sin + k() cos ) (cos + k() sin )( sin + k() cos ) ( sin + k() cos )2

Select k() = tan , [0, ]. 4 Then 0 k() 1, and W (0, t) =


t 0 1 cos2

0 0

d,

which for 0 t is clearly not of full rank. 4 This is a COUNTEREXAMPLE to the statement in the problem. So, the fact that {A, b + kc} is a controllable pair for all constant k [0, 1] does not imply that {A, b + k(t)c} is controllable for any function k with values in [0, 1]. You cannot determine the controllability of a time variant system by looking at the controllability of sampled versions of the pair (A(), B()).

Problem 4
(a) The state vector is
x() =

r() r() () () () ()

and the control vector is

ur () u() = u () u ()

For the given f (x, u) we have


f1 x f2 x f3 x f4 x f5 x f6 x

= 0, 1, 0, 0, 0, 0 = 2 cos2 + 2 + = 0, 0, 0, 1, 0, 0 = 2 r2 r
u , 2 , 0, 2r/r r mr2 cos mr sin sin + 2 cos , 2 sec2 u (mr cos )2 , 2 tan 2k , 0, 0, 2r cos2 , 2r2 cos sin r3

+ 2r, 0

= 0, 0, 0, 0, 0, 1 =
2r r2

u , 0, 0, 2 cos sin , 2 cos , 2r/r mr2

f (x, u) A(t) = |xo (t),uo (t) = x

0 1 2 0 3o 0 0 0 2o /ro 0 0 0 0

0 0 0 0 2ro o 0 0 1 0 0 0 0 0 0 0 2 0 0 o

0 0 0 0 1 0

2 3 2 3 2 where ro o = k and therefore o + 2k/ro = 3o .

Similarly,

f (x, u) B(t) = |xo (t),uo (t) = u

0
1 m

0 0 0 0

0 0 0
1 mro

0 0

0 0 0 0 0
1 mro

(b) The reduced system (r plane) is

d dt

r r

0 1 2 3o 0 0 0 0 2o /ro

0 0 0 2o ro 0 1 0 0

r r

1 m + 0

0 0 0
1 mro

ur r

For radial thrust only u = 0, ur = 0. 0 So, b = Then,


2 3

1 m . 0

0
1 [b, Ab, A b, A b] = m 0

1 m

0 0
2 mro o

0 2 o m 0

2o mro

o m 0 0
3 2o mro

It is easy to see that this matrix is singular. Therefore the system is not controllable when there is only radial thrust. In the case of tangential thrust only, ur = 0 and u = 0. 0 0 Now, b = . 0 The matrix
1 mro

[b, Ab, A bA b] =
2 3

0 0 0
1 mro

0
2o m 1 mro

2o m

0
3 2 4o mro

0 0 4o ro
2

2o m 0

It is easy to see that this matrix has full rank. So system is controllable using the tangential thrust only.

Problem 5
There was an error in the statement of the problem. I should have written reachability instead of controllability Grammian. Then for t1 W (to , t1 ) = (t1 , )B()B T ()T (t1 , )d
to

If we make a coordinate change x(t) = P (t)z(t) then we know that the transition matrix for the z(t) evolution is given by P 1 (t)(t, to )P (to ), while the corresponding B() matrix is P 1 (t)B(t), while the corresponding C() matrix is C(t)P (t). Therefore the corresponding reachability Grammian is Wz (to , t1 ) =
t1 to

P 1 (t1 )(t1 , )P ()P 1 ()B()B T ()P T ()P T ()T (t1 , )P T (t1 ) = = P 1 (t1 )W (to , t1 )P T (t1 ) (1)

Similarly for the observability Grammian M (to , t1 ) = we get Mz (to , t1 ) =


t1 to t1 to

T (, to )C T ()C()(, to )d

P T (to )T (, to )P T ()P T ()C T ()C()P ()P 1 ()(, to )P (to )d = = P T (to )M (to , t1 )P (to ) (2)

Set to = t , t1 = t in (1) and to = t, t1 = t + in (2). We get, Wz (t , t)Mz (t, t + ) = P 1 (t)W (t , t)P T (t)P T (t)M (t, t + )P (t) = = P 1 (t)W (t , t)M (t, t + )P (t) But then if x is an eigenvector of W (t , t)M (t, t + ) with eigenvalue , we consider z = P 1 (t)x . Then, Wz (t , t)Mz (t, t + )z = P 1 (t)W (t , t)M (t, t + )P (t)P 1 (t)x = P 1 (t)x = z So z is an eigenvector of Wz (t , t)Mz (t, t + ) with eigenvalue . Thus the eigenvalues of W (t , t)M (t, t + ) remain invariant under similarity transformations. (b) Both W and M are nonnegative denite. Let (W (t , t))1/2 5 (3)

be a symmetric square root of W (t , t). Then W (t , t)M (t, t + ) and W (t , t)1/2 M (t, t + )W (t , t)1/2 (4)

have the same eigenvalues except 0. Since M (t, t + ) is symmetric and nonnegative, and W (t , t)1/2 is the symmetric square root, it follows that the matrix in (4) is also symmetric and nonnegative denite. Hence its eigenvalues are nonnegative. (c) Let us assume that W (t , t) is nonsingular. Then W (t , t) = W (t , t)1/2 W (t , t)1/2 . Let, T (t) = W (t , t)1/2 . Then in the new coordinates W1 (t , t) = In and M1 (t, t + ) = W 1/2 (t , t)M (t, t + )W 1/2 (t , t) Let us assume M (t, t + ) is nonsingular. Then we can write M1 (t, t + ) = SS T where S is orthogonal and is diagonal. Let R(t) = S1/2 . In the new coordinates dened by R(t), W2 (t , t) = R1 (t)In RT (t) = 1/2 S T S1/2 = and M2 (t, t + ) = 1/2 S T S2 S T S1/2 = We can also treat the more general case when W (t , t) or M (t, t + ) or both are singular.

Problem 6 (Optional, Extra Credit)


Let x1 = x x2 = x Then the corresponding state model is

d dt

x1 (t) x2 (t)

0 1 1 0

x1 (t) x2 (t)

0 1

u(t)

The transition matrix for this system is

eAt =

cos t sin t sin t cos t 6

So, the trajectory starting at x1 (t) x2 (t)


t 0

0 0

is given by 0 1
t 0 sin(t )u()d t 0 cos(t )u()d

cos(t ) sin(t ) sin(t ) cos(t )

u()d =

(5)

We want to analyze and compare the reachable set of the above system, in one unit of time, with controls (a) 0 u(t) 1 and piecewise continuous. (b) u(t) = 1 or 0 and piecewise constant (ono type control). Let us describe in detail the reachable set Rb , using controls (b). For this we need to know the trajectory of the system for u(t) = 1 and for u(t) = 0. For u(t) = 0

x1 (t) x2 (t)

cos(t to ) sin(t to ) sin(t to ) cos(t to ) 0 0

x1 (to ) x2 (to ) x1 (to ) x2 (to )

(6)

which are obviously circles, with center tion with angular velocity 1 rad/sec.
x2

and passing through

in a clockwise direc-

u(t)=0
(x1 (to ), x (t )) 2 o

to

t-t

(x1 (t), x (t))


2

(0,0)

x1

For u(t) = 1

x1 (t) x2 (t)

cos(t to ) sin(t to ) sin(t to ) cos(t to ) 7

x1 (to ) x2 (to )

1 cos(t to ) sin(t to )

(7)

or x1 (t) 1 x2 (t) cos(t to ) sin(t to ) sin(t to ) cos(t to ) x1 (to ) 1 x2 (to )

(8)

which is clearly a circle with center (1,0) passing through x1 (to ),x2 (to ),

x2 u(t)=1
(x1 (to ), x (t )) 2 o (x1 (t), x (t))
2 o

t-t

x1

in the clockwise direction with angular velocity 1 rad/sec. It is clear that in order to move away from the origin we have to start with u(t) = 1 and keep it there for a while and then we can start switching to 0 and back to 1 for one unit of time = 1sec. We apply rst controls of the type shown below
u(t)

1 0

We get the following reachable set R


x2
R sin1 F B D -cos1 A

1-cos1 1

x1

The curve ABC is part of the circle with center (1,0) and passing through the origin. It goes up to the point (1 cos 1, sin 1). All points on ABC are reachable by keeping the control u(t) = 1. At any time, 0 1 we can switch to u(t) = 0 for the remaining time. The trajectory will be as described by (6) and the end point will be

x1 (1) x2 (1)

cos(1 ) sin(1 ) sin(1 ) cos(1 )

1 cos sin

cos(1 ) cos 1 sin(1 ) + sin 1

(9)

which shows that ADC is the arc of a circle with center ( cos 1, sin 1) which passes through the origin. From simple geometrical considerations, notice that the line ( cos 1, sin 1)(1, 0) is perpendicular to the line AC. And therefore, since both circles have radius 1, the set R is symmetric with respect to the line AC. From the denition of a convex set, a set is convex if and only if every point of a straight line segment joining two points of the set is also a part of the set. Clearly, the set R is convex. Obviously Rb R Let us know discuss Ra also. We also obviously have Ra Rb Since 1 < , from (5) it follows that x1 x2
t 0 sin(t )u()d t 0 cos(t )u()d

(10)

(11)

Ra =

x =

for some 0 t 1 and some u(t)s.t.0 u(t) 1

(12)

So, for x Ra clearly x1 0, x2 0. Therefore, for x Ra , x1 1 cos t x2 sin t

for 0 t 1

(13)

So, R in the orthogonal polygon (0, 0)(1 cos 1, 0)(1 cos 1, sin 1)(0, sin 1) i.e., R AECF (14)

(in our gure above). Observe now that the set of control functions U which are piecewise continuous and 0 u(t) 1 is convex. Indeed, if u1 () U, u2 () U then, u1 () + (1 a)u2 () = u3 () belong to U for 0 < < 1. This is clear since u3 () is obviously piecewise continuous if u1 () and u2 () are. And 10

0 u3 (t) + (1 ) = 1. So, if x1 , x2 belong to R , we must have t1 , t2 and u1 , u2 U s.t.

x1 =

t1 A(t1 ) 0 e

0 1 0 1

u1 () d

x2 =

t2 A(t2 ) 0 e

u2 () d

without loss of generality, assume t2 t1 . = (t2 t1 ))

Now, observe that I can rewrite x1 as (just set

x1 =

t2 A(t2 ) t2 t1 e

0 1 0 1

u1 ( (t2 t1 )) d

= where

t2 A(t2 ) 0 e

u3 ( ) d

u3 (t) = 0, 0 < t2 t1 = u1 ( (t2 t1 )), t2 t1 1. So, for 0 < < 1 0 1

x1 + (1 )x2 =

t2 A(t2 ) 0 e

(u3 () + (1 )u2 ())d

= and certainly

t2 A(t2 ) 0 e

0 1

u4 ()d

u4 () = u3 () + (1 )u2 () is admissible since U is convex. Therefore, R is also convex. 11 (15)

Let now x(t) be a trajectory in R emanating from (0,0). Now observe from (12) that for any time t [0, 1]

x2 (t) sin t x1 (t) cos t + cos t 1 = +


t 0 sin(

t 0 cos(

t) sin tu()d

t) cos tu()d + cos t 1 = + cos t 1

t 0 sin()u()d t 0 sin d

+ cos t 1 = 0

or x2 (t) sin t x1 (t) cos t + cos t 1 0, 0 t 1 or (x2 (t) sin t) sin t (x1 (t) + cos t) cos t + cos t 0, 0 t 1 or (x2 (t) sin t) sin t (x1 (t) (1 cos t)) cos t 0, 0 t 1 (16)

Now, (sin t, cos t) is a tangent unit vector to the curve ABC at the point (1 cos t, sin t) (see gure just below)
x2
(1-cost,sint) unit tangent vector with components (sint,cost)
outwards

C
1111 0000 1 0

1111 0000 unit normal 1111 0000 1111 0000 vector with
components (-cost,sint)

(x1 (t), x (t))= a generic trajectory


2 point

x1

So, ( cos t, sin t) is a normal unit vector pointing outwards. So (16) says: if at any time t [0, 1] you draw the vector from the corresponding point (1 cos t, sin t) of the curve ABC to the trajectory point (x1 (t), x2 (t)) it should point inwards! Conclusion: R ABCE (17)

Similarly, (sin(1 t), cos(1 t)) is a unit target vector to the curve ADC at the point (cos(1 t) cos 1, sin(1 t) + sin 1) and, therefore, (cos(1 t), sin(1 t)) 12

is a unit vector normal to ADC at the same point and pointing in the exterior of ABCDA, as in the gure just below.
x2
unit tangent vector (sin(1-t),cos(1-t))

(x1 (t), x (t)) C 2


1 0

(cos(1-t)-cos1,-sin(1-t)+sin1) unit normal vector (cos(1-t),-sin(1-t)) A

x1

So, again using (16), (x1 (t) + cos 1 cos(1 t)) cos(1 t) sin(1 t)(x2 (t) sin 1 + sin(1 t)) = x1 (t) cos(1 t) sin(1 t)x2 (t) + cos 1 cos(1 t) cos2 (1 t) + sin(1 t) sin 1 sin2 (1 t) = =
t 0 [sin(t

) cos(1 t) sin(1 t) cos(t )]u()d 1 + cos t 1)u()d 1 + cos t cos(t 1) cos(2t 1) 1 + cos t

t 0 sin(2t

= cos(t 1) cos t cos(t 1) + sin t sin(t 1) 1 + cos t = (1 cos t)(1 cos(1 t)) sin t sin(1 t) 0, for 0 t 1 So, again: if at any time t (0, 1) you draw the vector from the corresponding point (cos(1 t) cos 1, sin(1 t) + sin 1) of the curve ADC to the trajectory point (x1 (t), x2 (t)) it should point inwards! Conclusion: R ABCDA = R Now, from (10),(11),(18) we have R = R = Rb What we have proved is a special case of the famous Bang-Bang theorem of control theory that states that for general classes of dynamical systems the reachable sets with controls with values in a convex compact set U is exactly the same as the reachable set with controls with values in the extreme points of the set U . (18)

13

Potrebbero piacerti anche