Linear Multivariable Theory

Linear Multivariable Control Theory

D Viswanath

()

Notes

October 8, 2016

1 / 36

Linear Multivariable Theory

Outline

Introduction

Geometric Control Preliminaries

Results and Discussion

Conclusions and Future work

Publications

References

()

Notes

October 8, 2016

2 / 36

Introduction

Overview
Geometric term is used since the setting is linear state space and the mathematics used is chiefly linear algebra in abstract (geometric) style.
The concepts of controllability and observability are the geometric properties
of distinguished state subspaces.
The geometric approach does not directly look for a feedback law (say, u = F x)
as a solution to the synthesis problem. Instead it first characterizes solvability
as a verifiable property of some state space, say, . Then F is calculated from
.
Thus an intractable nonlinear problem in F is converted to a straightforward
quasilinear one in F .

()

Notes

October 8, 2016

3 / 36

Geometric Control Preliminaries

Linear Spaces
A linear (vector) space consists of an additive group, of elements called vectors,
together with an underlying field of scalars.
The field F of real numbers R or complex numbers C is considered.
Linear spaces are denoted by capital scripts X, Y, ...and their elements (vectors) by lower case Roman letters x, y,...; and the field elements (scalars) by
lower case Roman or Greek letters.
Properties of vector addition and multiplication of vectors by scalars apply, i.e.,
if x1 , x2 X and c1 , c2 F then
c1 x1 X, c1 (x1 + x2 ) = c1 x1 + c1 x2 , (c1 + c2 )x1 = c1 x1 + c2 x1 ,
(c1 c2 )x1 = c1 (c2 x1 ).
If k is a positive integer, k denotes the set of integers {1, 2, 3, ..., k}.

()

Notes

October 8, 2016

4 / 36

Geometric Control Preliminaries

Linear Spaces
Let x1 , x2 , ..., xk X where X is defined over F. Their span is the set of all
linear combinations of the xi , with coefficients in F.
X is f inite dimensional if there exists a (finite) k and a set {xi , i k; xi
X}.
If X 6= 0, the least k for which this happens is the dimension of X written
d(X). When X = 0, d(X) := 0. If k = d(X) 6= 0, a spanning set {xi , i k} is
a basis for X.
A set of vectors {xi X, i m} is (linearly) independent (over F) if for all
sets of scalars {ci F, i m}, the relation
m
X

ci xi = 0

(1)

i=1

implies ci = 0 for all i m. If the xi are independent, and if x Span{xi , i

m}, then the representation x = c1 x1 + c2 x2 + ... + cm xm is unique.

()

Notes

October 8, 2016

5 / 36

Geometric Control Preliminaries

Linear Subspace
A linear subspace S of the linear space X is a non-empty subset of X which is
a linear space under the operations of vector addition and scalar multiplication.
S X and for all x1 , x2 S and c1 , c2 F, we have c1 x1 + c2 x2 S.
The notation S X means that S is a subset of X.
If xi X; (i k), then Span{xi , i k} is a subspace of X.
Geometrically, a subspace may be pictured as a hyperplane passing through
the origin of X.
Thus the vector 0 S for every subspace S X.
0 d(S) X, with d(S) = 0 (resp.d(X)) if and only if S = 0.(resp.X)

()

Notes

October 8, 2016

6 / 36

Geometric Control Preliminaries

Linear Subspace
If R, S X, subspaces R + S X and R
to
R+S
\
R S

S X can be defined according

:= {r + s : r R, s S}

(2)

:= {x : x R & x S}

(3)

R + S isTthe span of R and S and may be much larger than the set-theoretic
union.R S is generally not a subspace.
T
As the zero subspace 0 R and 0 S, it is always true that 0 R S 6= .
In other words, two subspaces of X are never disjoint in the set-theoretic
sense.

()

Notes

October 8, 2016

7 / 36

Geometric Control Preliminaries

Linear Subspace-Lattice
The family of all subspaces is partially ordered by subspace inclusion ().
Under the operations + and , this forms a lattice,Tnamely R + S is the smallest subspace containing both R and S, while R S is the largest subspace
contained in both R and S.
A lattice diagram may picture inclusion relations among subspaces in which
the nodes represent subspaces and a rising branch from R to S means R S.
Thus shown below is the lattice diagram for arbitrary R and S X:-

()

Notes

October 8, 2016

8 / 36

Geometric Control Preliminaries

Linear Subspace-Lattice
Let R, S, T X and suppose R S.
Then
\
\
\
R (S + T) = R S + R T
\
= S + R T

(4)
(5)

Equation 5 is called the modular distributive rule. The lattice in which this
rule holds i called modular.
If no inclusion, i.e., R, S, T X, is postulated, then
\
\
\
R (S + T) = R S + R T
(6)
It is also true that
S

\
\
\
(R + T) = S R + S T

(7)

\
\
\
(R + S) = T R + T S

(8)

and by symmetry

()

Notes

October 8, 2016

9 / 36

Geometric Control Preliminaries

Linear Subspace - Linear Independence

T
Two subspaces R, S X are linearly independent if R S = 0.
A family of k subspaces is independent if
\
Ri (R1 + ... + Ri1 + Ri+1 + ... + Rk ) = 0

(9)

for all i k.
An independent set of vectors cannot include the zero vector. But any independent family of subspaces remains independent if one or more zero subspaces
are adjoined.
The following statement are equivalent:(a) The family {Ri , i k} is independent.
Pk
TP
(b) i=1 (Ri
j6=i Rj ) = 0.
Pk
T Pi1
(c) i=2 (Ri
j=1 Rj ) = 0.
(d) Every vector x R1 +...+Rk has a unique representation x = r1 +...+rk
with ri Ri .

()

Notes

October 8, 2016

10 / 36

Geometric Control Preliminaries

Linear Subspace - Linear Independence

()

Notes

October 8, 2016

11 / 36

Geometric Control Preliminaries

Maps and Matrices

Let X and Y be linear spaces over F.
A function : X Y is a linear transformation (or map, for short), if
(c1 x1 + c2 x2 ) = c1 (x1 ) + c2 (x2 )

(10)

for all x1 , x2 X and c1 , c2 F.

Maps will usually be denoted by Roman capitals A, B, ....
With X and Y fixed, the set L(X, Y)of all linear maps C : X Y are
considered.
This set is turned into a linear space over F by the natural definitions of
addition, and multiplication by a scalar:
(C1 + C2 )x
(cC1 )x

:= C1 x + C2 x
:= c(C1 x)

(11)
(12)

for all x X, c F and C1 , C2 L(X, Y).

The dimension d(L(X, Y)) = d(X)d(Y).
()

Notes

October 8, 2016

12 / 36

Geometric Control Preliminaries

Maps and Matrices

Usually C L(X, Y) can be written as C : X Y.
Let xi , i n be a basis for X and yj , j p be a basis for Y.
If C : X Y is a map, then
Cxi = c1i y1 + c2i y2 + ... + cpi yp , i n

(13)

for uniquely determined elements cji F.

It can be seen that if x X, then Cx is completely determined by the Cxi ;
linearity does the rest.
The array

c11
..
M at C = .

cp1

. . . c1n
..
.

(14)

. . . cpn

is the matrix of C relative to the given basis pair.

()

Notes

October 8, 2016

13 / 36

Geometric Control Preliminaries

Maps and Matrices

Fundamentally, M at C can be considered as a function p n F.
The symbol Fpn denotes the class of all p n matrices with elements in F.
It is turned into a linear space over F, of dimension pn, by the usual
operations of matrix addition and multiplication of matrices by scalars.
Let C : X Y be a map.
X is the domain of C and Y is the co-domain; the array size of M at C is
thus d(X) d(Y).
The kernel (or null space) of C is the sub-space
Ker C := {x : x X & Cx = 0} X

(15)

while the image or range of C is the subspace

Im C

()

:= {y : y Y & x X, y = Cx
= {Cx : x X} Y

Notes

(16)
(17)

October 8, 2016

14 / 36

Geometric Control Preliminaries

()

Notes

October 8, 2016

15 / 36

Geometric Control Preliminaries

Linear Invariance
Let A : X X and let S X have the property AS S.
S is said to be A invariant.

()

Notes

October 8, 2016

16 / 36

Disturbance Decoupling

Disturbance Decoupling Problem(DDP)

Consider the system
x(t)

=
z(t) =

Ax(t) + Bu(t) + Sq(t); t 0

Dx(t); t 0

(18)
(19)

The term q(t) represents a disturbance which is assumed not to be directly

measurable by the controller.
The DDP problem is to find (if possible) state feedback F such that q(.) has
no influence on the controlled output z(.).
Assume that q(.) belongs to a rich function class Q since the specific features
of the disturbance are not known.
Setting Q := R with the continuous R -valued functions on [0, inf), we
assume that S : Q X.

()

Notes

October 8, 2016

17 / 36

Disturbance Decoupling

Disturbance Decoupling
Let u(t) = F x(t) in equation 18 with state feedback F introduced.
The system represented by equations 18 and 19 are said to be disturbance
decoupled relative to the pair q(.), z(.), if for each initial state x(0) X, the
output z(t), t 0, is the same for every q(.) Q.
Thus disturbance decoupling means that the forced response
Z t
z(t) = D
e(ts)(A+BF ) Sq(s)ds = 0
(20)
0

for all q(.) Q and t 0.

()

Notes

October 8, 2016

18 / 36

Disturbance Decoupling

Let K := Ker D and S := Im S.

Then from equation 20 the following result is obtained.
Lemma 1. The system 18 and 19 is disturbance decoupled if and only if
hA + BF |Si K.
In algebraic terms the realization of disturbance decoupling amounts to the
following.
Disturbance Decoupling Problem(DDP). Given
A : X X, B : U X, S X, and K X,
find (if possible) F : X U such that
hA + BF |Si K

(21)

The subspace on the left on eqn.21 is (A + BF )-invariant and if eqn.21 is true,

it belongs to K.
Intuitively, DDP will be solvable if and only if the largest subspace having
these properties contains S.
()

Notes

October 8, 2016

19 / 36

Disturbance Decoupling

(A, B)-Invariant Subspaces

Let A : X X and B : U X.
A subspace V X is A, B-invariant if there exists a map F : X U such that
(A + BF )V V

(22)

The class of A, B invariant spaces of X are denoted by T(A, B; X), or

simply T(X) where A and B are fixed.
It can be observed that any A invariant subspace is automatically A, B
invariant with F = 0.

()

Notes

October 8, 2016

20 / 36

Disturbance Decoupling

Let V be A, B invariant and state feedback F chosen so as to satisfy

Equation 22.
The disturbance-free closed loop system
x = Ax + Bu; u = F x

(23)

then has the property that if x(0) x0 V,

x(t) = et(A+BF ) x0 V

(24)

for all t, i.e., V is invariant under the motion.

Thus V has the property that if x(0) V then there exists a control u(t)(t
0) such that x(t) V for all t 0.
In other words, the state x() can be held in V by suitable choice of u().

()

Notes

October 8, 2016

21 / 36

Disturbance Decoupling

An explicit test to determine whether a given subspace is A, B invariant is

given by the following result.
Lemma 4.2 Let V X and B = Im B. Then V T(A, B; X) if and only if
AV V + B

(25)

Proof. Suppose V T(A, B; X) and let v V. By equation 23, (A +

BF )v = w for some w V, or
Av = w BF v V + B

(26)

Conversely, suppose equation 25 is true. Let v1 , v2 , ..., v be a basis for V.

By Eqn.25 there exists wi V and ui U (i ) such that
Avi = wi Bui , i

(27)

F0 vi = ui , i

(28)

Defining F0 : V U by
and let F be any extension of F0 to X.
Then (A + BF )vi = wi V, i.e.,(A + BF )V V, so that V T(A, B; X).
()

Notes

October 8, 2016

22 / 36

Disturbance Decoupling

If V T(A, B; X) then the class of maps F : X U, such that (A+BF )V V

is written as F(A, B; V), or simply F(V).
The notation F F(V) is read F is a friend of V.
F(V) if and only if
From the proof of Lemma 4.2, if F F(V) then F
1

(F F )V B V.
T
F )|V = 0 if B is monic and B V = 0.
In particular, (F

()

Notes

October 8, 2016

23 / 36

Disturbance Decoupling

Proposition 4.1
Let (A, B) be controllable and V T(A, B; X), with d(V) = v.
V
If F0 F(A, B; V) and is a symmetric set of n v complex numbers, there
exists F : X U, such that
F |V = F0 |V
and
(A + BF ) = [(A + BF )|V]
Proof

()

Notes

(29)
]^

(30)

October 8, 2016

24 / 36

Disturbance Decoupling

Lemma 4.3
The class of subspaces T(A, B; X) is closed under the operation of subspace
addition.
Proof. From Lemma 4.2, if V1 , V2 T(X), then
A(V1 + V2 )

= AV1 + AV2

(31)

V1 + V2 + B

(32)

Hence V1 + V2 T(X).

()

Notes

October 8, 2016

25 / 36

Disturbance Decoupling

If B is a family of subspaces of X, the largest or supremal element V of B

is defined as that member of B (when it exists) which contains every member
of B.
Thus, V B and if V B then V V .
V is unique.
We write
V = sup {V : V B}

(33)

In simple terms, this can be written as V = sup B.

()

Notes

October 8, 2016

26 / 36

Disturbance Decoupling

Lemma 4.4
Let B be a nonempty class of subspaces of X, closed under addition. Then B
contains a supremal element V .
Proof. Since X is finite-dimensional there is an element V B of greatest
dimension.

If V B, V + V B and so d(V ) d(V + V ) d(V ); i.e., V = V + V

hence V V and so V is supremal.

()

Notes

October 8, 2016

27 / 36

Disturbance Decoupling

Let K X be arbitrary, and let T(A, B; K) denote the subclass of (A, B)

invariant subspaces contained in K such that
T(A, B; K) := {V : V T(A, B; X) & V K

(34)

With A and B fixed, T(A, B; K) can be written as T(K).

Trivially, 0 K, so T(K) 6= .
Since K is a subspace Lemma 4.3 implies that T(K) is closed under addition.
Then, Lemma 4.4 guarantees the existence of the supremal element
V := sup T(K)

()

Notes

(35)

October 8, 2016

28 / 36

Disturbance Decoupling

Theorem 4.1
Let A : X X and B : U X. Every subspace K X contains a unique
supremal (A, B)-invariant subspace T(K).
Thus when K = Ker D and z = Dx, a choice of feedback control F F(V )
where V = sup T(Ker D) renders the system maximally unobservable from
z.

()

Notes

October 8, 2016

29 / 36

Disturbance Decoupling

Theorem 4.2
DDP is solvable if and only if
V S

(36)

V := sup T(A, B; K)

(37)

where
Proof. (If) From Lemma 4.2, F F(V ), i.e., (A + BF )V V .
Using equation 36,
hA + BF |Si hA + BF |V i = V K

(38)

(Only if) If F solves DDP, the subspace

V := hA + BF |Si

(39)

clearly belongs to T(K), and therefore

V V S

()

Notes

(40)

October 8, 2016

30 / 36

Disturbance Decoupling

Algorithm for V : Theorem 4.3

An algorithm by which V can be computed efficiently in a finite number of
steps can help in determining V without the need for checking condition 36
of Theorem 4.2.
Theorem 4.3. Let A : X X, B : U X, and K X. Define the sequence
V according to
V0
V

= K
\
= K A1 (B + V1 ); = 1, 2, ..

(41)

Then V V1 and for some k d(K),

V = sup T(A, B; K)

(42)

for all k.

()

Notes

October 8, 2016

31 / 36

Disturbance Decoupling

Principle of Solution to DDP

If V T(X) and q() = 0, then by suitable choice of u() the system state
x() can always be held in V if it starts there.
If q 6= 0 but Im S V, then the contribution to x(t)

by the disturbance Sq(t)

(i.e., the first-order-effect of q() on x()) is also localized to V.
Under these conditions, the integrated contribution to x() by q() can be
controlled to remain in V.
This contribution is unobservable at z just when V Ker D, and so it is
enough to work with V .
In actual operation with a control of form u = F x + Gv, where F F(V )
and v() is a new external input, the system state will generally not remain in
V ; however linearity ensures that the contribution to x() from q() is held in
V , which is all that is required to decouple q() from z().

()

Notes

October 8, 2016

32 / 36

Disturbance Decoupling

()

Notes

October 8, 2016

33 / 36

Disturbance Decoupling

()

Notes

October 8, 2016

34 / 36

Disturbance Decoupling

()

Notes

October 8, 2016

35 / 36

Disturbance Decoupling

References
[1]

Morgan Jr., The synthesis of linear multivariable systems by state variable

feedback, Proc. 1964 JACC, Stanford, California, pp. 468-472., 1964.

[2]

Wonham, W.M., Linear Multivariable Control - A Geometric Approach,

Springer-Verlag, New York, 1979.

[3]

Isidori, A., Nonlinear Control Systems., Springer-Verlag, London, 3rd Ed., 1995.

()

Notes

October 8, 2016

36 / 36