• It is an equivalence relation
• Invariants : eigenvalues with their algebraic and geometric multiplicities
• Canonical form : By placing as columns in M the generalized
eigenvectors of A1 we get its Jordan Canonical form.
λi 1 ⋯ 0
J = diag {J ij } 0 λ ⋱ ⋮
i =1,…, v J ij = i ∈ ℂvij ×vij
j =1,…, vi ⋮ ⋱ 1
0 ⋯ 0 λ
i
A real alternative of M, J is also possible.
Similarity – Geometrical interpretation
Consider the homogeneous system of the form
ρ x1 (t ) = A1 x1 (t ) ⇔ ( ρ I n − A1 ) x1 (t ) = 0n ,1
where ρ := d / dt A1 ∈ ℝ n×n .
x1 (t ) = Mx2 (t )
x1 (t ) = Mx2 (t )
It is easy to see that the above change of coordinates results in a new description
of the form
ρ x2 (t ) = A2 x2 (t ) + B2 u(t )
y (t ) = C2 x2 (t ) + D2 u(t )
where A2 = M −1 A1M , B2 = M −1B1 , C2 = C1M and D2 = D1 .
sI p − A2 B2 M −1 0 sI p − A1 B1 M 0
=
−C
2
D2 0 I l −C1 D1 0 I m
M 0 sI p − A2 B2 sI p − A1 B1 M 0
0 =
I l −C2 D2 −C1 D1 0 I m
x1 ( t ) M 0 x2 ( t ) x1 ( 0 ) M 0 x2 ( 0 )
= =
−u ( t ) 0 I m −u ( t ) y2 ( t ) 0 I m y1 ( t )
X u2
Controllability Observability Yu
X u1
Commute
Unimodular equivalence
Definition [Gantmacher 1959, Rosenbrock 1970] Two polynomial matrices
Ai ( s ) ∈ ℝ[ s ] p×q , i = 1,2 are unimodular equivalent, if there exist unimodular
matrices (determinant constant) U ( s ) ∈ ℝ[ s ] p× p ,V ( s ) ∈ ℝ[ s ]q×q , such that
A1 ( s ) = U ( s ) A2 ( s )V ( s ).
U (s) A( s ) V ( s) S AC( s ) ( s )
Unimodular equivalence – Canonical form
Theorem [Gantmacher 1959] Given a polynomial matrix A( s ) , there exist
unimodular matrices U ( s ) , V ( s ) such that
A( s ) = U ( s ) S Aℂ( s ) ( s )V ( s )
where
S Aℂ( s ) ( s ) = diag{ε1 ( s ),…, ε r ( s ), 0}
vi
mij
is the Smith form of A( s ) and ε i ( s ) = ∏ ( s − λ j ) are monic polynomials
j =1
such that ε i ( s ) | ε i +1 ( s ) , i = 1,…, r − 1.
mij
The roots of ε i ( s ) are the finite zeros of A( s ) , whereas the factors ( s − λ j ) of
ε i ( s ) , are the finite elementary divisors of A( s ) .
The finite elementary divisor structure and therefore the zeros and their
multiplicities form a complete set of invariants for unimodular equivalence.
∃M ∈ℝ p× p
( sI p − A1 ) = U ( s ) ( sI p − A2 )V ( s ) ⇔ ( sI p − A1 ) = M ( sI p − A2 ) M −1
Finite Jordan pairs
Theorem ([Gohberg et. al.], pp.184) Let (C s0 ∈ R r×n , J s0 ∈ R n×n ) be a matrix pair where
J s0 is a Jordan matrix with unique eigenvalue s0 . Then the following conditions are
necessary and sufficient in order that (C s0 , J s0 ) be a Jordan pair of A( s )
corresponding to s0 : s0 1 0 ⋯ 0
(a) det A ( s ) has a zero at s0 of multiplicity n , 0 s0 1 ⋯ 0
(b) ∈ ℝ n×n
J s 0 := ⋮ ⋮ ⋱ ⋱ ⋮
Cs0
0 0 0 ⋯ 1
C J
s s 0 0 0 ⋯ s0
rank 0 0 = n
⋮
A ( s ) = Aq s q + ⋯ + A1s + A0
Cs J sn−1
(c) 0 0
AqCs0 J sq0 + ⋯ + A1Cs0 J s0 + A0Cs0 = 0
Taking a Jordan pair (C si ∈ R r×ni , J si ∈ R ni ×ni ) for every eigenvalue si , i = 1, 2,.., k of
A ( s ) we define a finite Jordan pair ( C ∈ R r×n , J ∈ R n×n ) of A ( s ) as
(
C = Cs1 C s2 ⋯ C sk ) ; J = blockdiag J s1 ( J s2 ⋯ J sk )
where n = n1 + n2 + ⋯ nk is the total number of finite zeros (eigenvalues) of A ( s ) (order
accounted for).
Construction of the smooth solution space
The column span of Ψ = Ce Jt defines a solution space B C of
A ( ρ ) β ( t ) = 0, ρ := d / dt (Σ)
with dimension :
dim B C = n := total number of finite zeros of A( s ) (order accounted for)
B C is known as the smooth solution space of (Σ) defined by (Gohberg et. al. 1982) and
(Vardulakis 1991).
103 ρ 2 − 145 ρ + 4 3
ρ 2 − 52 ρ β1 ( t ) 0
=
10
1 2 2
− 10 ρ − 5 ρ
− 10 ρ − 5 ρ + 4 2 ( ) 0
1 2 6
β t
A( ρ ) β (t )
e 2 t ( t − 1) 2 t t − 1 k −1 k − 2
2 1
1 −1 0 2
t e2t 2 t 1 k 1
→ B = e 1 , e t + 1 B = 2 1 , 2 k + 2
C f
Ce =
Jt
e = 2t
1 1 e e ( t + 1)
2t
The column span of Ψ = CJ k defines a solution space B f of
A (σ ) β ( k ) = 0, σβ ( k ) = β ( k + 1)
with dimension :
dim B f = n := total number of finite zeros of A( s ) (order accounted for)
B f is known as the forward solution space of (A(σ)) defined by (Gohberg et. al.
1982).
Unimodular equivalence – Geometrical interpretation
Theorem [Polderman and Willems 1998] Consider two homogeneous
systems of the form
A1 ( ρ ) β1 ( t ) = 0, A2 ( ρ ) β 2 ( t ) = 0
p×q
where ρ := d / dt and A1 ( ρ ) , A2 ( ρ ) ∈ ℝ [ ρ ] . A1 ( ρ ) , A2 ( ρ ) define the
same smooth behavior B if and only if there exists a unimodular matrix
q× q
V ( ρ ) ∈ ℝ [ ρ ] such that V ( ρ ) A1 ( ρ ) = A2 ( ρ ) .
AL ( s ) BL ( s ) p× p p ×m
L = P ( s) =
− I 0
L
: A ( s ) ∈ ℝ [ s ] , rankAL ( s ) = p , BL ( s ) ∈ ℝ [ s ]
p p ,m
AR2 ( s ) AR1 ( s )
•Same output decoupling zeros, = T (s)
− BR ( s ) − BR ( s ) R
•Same transfer function 2 1
−1 −1
BR2 ( s ) AR2 ( s ) = BR1 ( s ) AR1 ( s )
Strict System Equivalence (PMDs)– Rosenbrock 1970
otherwise called Unimodular Strict System Equivalent
Ai ( ρ ) β ( t ) = Bi ( ρ ) u ( t ) i=1,2
y ( t ) = Ci ( ρ ) β ( t ) + Di ( ρ ) u ( t )
Let
Ai ( s ) − Bi ( s )
Pi ( s ) = , i = 1,2
Ci ( s ) Di ( s )
be the system matrices of two PMDs Σi , where Ai ( s ) ∈ ℝ[ s ]r ×r , Bi ( s ) ∈ ℝ[ s ]r ×l
, Ci ( s ) ∈ ℝ[ s ]m×r , Di ( s ) ∈ ℝ[ s ]m×l .
A1 ( s )
• Output Decoupling Zeros
− C
1 ( s )
• Input-Output Decoupling Zeros
• Order of Σ n = deg A1 ( s )
• Rosenbrock degree {
d r = max deg Pj11, j22 ,..., qjɶqɶ ( s )
i ,i ,...,i
}
• System Zeros { i ,i ,...,i
}
gcd Pj11, j22 ,..., qjɶqɶ ( s ) , q = qɶ = greatest value of q s.t. Pj11, j22 ,..., qjɶqɶ ( s ) ≠ 0
i ,i ,...,i
• System Poles A1 ( s )
−1
• Transfer function matrix G1 ( s ) = C1 ( s ) A1 ( s ) B1 ( s ) + D1 ( s )
• Transmission Zeros-Poles
Invariants – Geometric interpretation
• Karcanias, 1976. ‘Geometric Theory of zeros and its use in feedback analysis’,
Ph.D. Thesis, Electrical Engineering Department, UMIST, U.K.
• MacFarlane and Karcanias 1976, Poles and zeros of linear multivariable
systems: a survey of the algebraic, geometric and complex variable theory, Int. J.
Control, Vol. 24, 33-74.
• Desoer CA, Schulman JD (1974) Zeros and poles of matrix transfer functions
and their dynamical interpretations. IEEE Trans, vol CAS–21, No1, pp 3–8.
• Karcanias and Kouvaritakis, 1979 The output zeroing problem and its
relationship to the invariant zero structure: a matrix pencil approach", Int. J.
Control, Vol. 30, pp.395-415
• F.M. Callier, V.H.L. Cheng, and C.A. Desoer, Dynamic interpretation of poles
and zeros of transmission for distributed multivariable systems. IEEE Trans.
Circuits and Systems, CAS-28:300-306, April 1981
• C. B. Schrader and M. K. Sain, 1989, Research on System Zeros: A Survey,
International Journal of Control, Volume 50, Number 4, pp. 1407-1433
• Bourles H., Fliess M. (1997) "Finite poles and zeros of linear systems: an
intrinsic approach", Internal. J. Control, vol. 68, pp. 897-922
Special forms of Strict System Equivalence in different
system descriptions – Smith 1981
s+2 0
(
s + 5s + 6 )
2
0 s
+ 3
A1 ( s )
A2 ( s )
1 0 −1 1 s + 2 0 1 s + 3
= 0 1 s + 2
2
+ + s +
3 − s − 2 s + 3
0 s 5 s 6
Aɶ1 ( s ) U (s) A2 ( s ) V (s)
s + 2 1 1 0 s +2 0 1 s + 3
s + 3 1 =
0 s + 5s + 6 0
2
s + 3 1 s + 2
Vardulakis,
Antonis-Ioannis
Nicos Karcanias
Clive Pugh
• G.C. Verghese and T. Kailath, 1981, IEEE Transactions on Automatic Control,26, pp. 434-439
• A. I. G. Vardulakis, D. N. J. Limebeer, and N. Karcanias, “Structure and smith-MacMillan form of
a rational matrix at infinity,” International Journal of Control, vol. 35, no. 4, pp. 701–725, 1982.
Biproper equivalence
Although unimodular equivalence operations preserve the finite zero
structure of polynomial matrices, they will in general destroy the
corresponding structure at infinity.
Definition [P. D. P. Van Dooren 1979, Vardulakis et. al. 1982] Two rational
matrices Ai ( s ) ∈ ℝ( s ) p×q , i = 1,2 , are equivalent (at infinity) if there exist
biproper rational matrices U ( s ) ∈ ℝ pr ( s ) p× p ,V ( s ) ∈ ℝ pr ( s ) q×q , such that
A1 ( s ) = U ( s ) A2 ( s )V ( s )
i.e. lim s→∞ U ( s ) = U ∈ ℝ p× p ,det U ≠ 0 (similar for V(s)).
The indices qi, i=1,r form a complete set of invariants for biproper
equivalence.
1 0 s s2 + s + 1 − 1s s + 1 − 1
s 2 + s + 1 s2 0
s2
=
1
s 2 − 1 1 −1 −s − 1 1 1 0 1
s3 s s2
The infinite elementary divisors of A( s) are given by (Hayton et.al.1988 & Vardulakis 1991):
µ
s j , j = 1, 2,3,.., r
µ j = q − q j , j = 1, 2,..., k with q1 = q ≥ q2 ≥ ⋯ ≥ qk
where qi are the orders of the poles of A( s ) at s = ∞ defined above.
The first kind of IED, i.e. the ones with degrees µ j = q − q j > 0, j = 2,3,.., k we call
infinite pole IEDs, since are connected with the orders of the poles q j , j = 1, 2,..., k of
A( s ) at s = ∞ .
1 0 s s2 + s + 1 −1 − s −1 − s − s 2 s 2−2 0
=
s − s3 1 −s 2 −s 2 − s 1 s 0 s 2+2
Ṽ s
0 s
Ũs Ãs S Ãs
Connection between the Smith form at infinity and
the Smith form at C
Define the dual polynomial matrix of A ( s )
1
Aɶ ( s ) = s q A = A0 s q + A1s q −1 + ⋯ + Aq
s
The infinite elementary divisors of A( s) are given by (Hayton et.al.1988 & Vardulakis 1991):
µ j = q, j = k + 1, k + 2,.., v
µ j = q + qˆ j , j = v + 1,τ + 2,..., r with qˆv +1 ≤ qˆv + 2 ≤ ⋯ ≤ qˆr
where qˆi are the orders of the zeros of A( s) at s = ∞ defined above.
The second kind of IED, i.e. the ones with degrees µ j = q + qˆ j > 0, j = v + 1, v + 2,..., r we
call infinite zero IEDs, since are connected with the orders of the zeros
qˆ j , j = v + 1, v + 2,..., r of A( s ) at s = ∞ .
1 0 s s2 + s + 1 −1 − s −1 − s − s 2 s 2−2 0
=
s − s3 1 −s 2 −s 2 − s 1 s 0 s 2+2
Ṽ s
0 s
Ũs Ãs S Ãs
Infinite Jordan pairs
Theorem (Gohberg et. al.1982 ,pp.185) Let (C ∞ ∈ R r× µ
, J ∞ ∈ R µ ×µ
) be a matrix pair
where J ∞ is a Jordan matrix with unique eigenvalue s0 = 0 . Then, the following
conditions are necessary and sufficient in order that ( C∞ , J ∞ ) be an infinite Jordan pair
of A ( s ) : 1
Aɶ ( s ) = s q A = A0 s q + A1s q −1 + ⋯ + Aq
(a) det Aɶ ( s ) has a zero at s0 = 0 of multiplicity µ , s
(b)
C∞
C J
rank ∞ ∞ = µ
⋮
µ −1
(c) C∞ J ∞
A0C∞ J ∞q + ⋯ + Aq −1C∞ J ∞ + AqC∞ = 0
( )
Taking a finite Jordan pair C∞i ∈ R r×µi , J ∞i ∈ R µi ×µi for different algebraic multiplicities
of the eigenvalue si = 0, i = 1, 2,.., r of Aɶ ( s ) we define an infinite Jordan pair
(C ∞ ∈ R r×µ , J ∞ ∈ R µ ×µ ) of A ( s ) as
C∞ = ( C∞1 C∞ 2 ⋯ C∞r ) ∈ R r×µ ; J ∞ = diag ( J ∞1 J ∞ 2 ⋯ J ∞r ) ∈ R µ ×µ
where µ = µ1 + µ 2 + ⋯ + µ r is the total number of infinite elementary divisors
(eigenvalues) of A ( s ) (order accounted for).
Infinite Jordan pairs
Theorem (Gohberg et. al.1982 ,pp.185) Let (C∞ ∈ R r× µ
, J ∞ ∈ R µ ×µ
) be a matrix pair
where J ∞ is a Jordan matrix with unique eigenvalue s0 = 0 . Then, the following
conditions are necessary and sufficient in order that ( C∞ , J ∞ ) be an infinite Jordan pair
of A ( s ) :
(a) det Aɶ ( s ) has a zero at s0 = 0 of multiplicity µ ,
(b)
C∞
C J
rank ∞ ∞ = µ
⋮
µ −1
(c) C∞ J ∞
A0C∞ J ∞q + ⋯ + Aq −1C∞ J ∞ + AqC∞ = 0
It is easily seen from the above that ( C∞ , J ∞ ) is an infinite Jordan pair of A ( s ) iff it is
a finite Jordan pair of its dual polynomial matrix Aɶ ( s ) corresponding to its zero
(eigenvalue) at s = 0 .
Geometric interpretation of finite and infinite zeros in
continuous time systems
qˆr −1
The column span of Ψ = ∑ Cɶ Jɶ δ ( ) ( t ) where
∞ ∞, z
i
∞, z
i
i =0
Definition [Gantmacher 1959] Two regular matrix pencils sEi − Ai , i = 1,2, with
Ei , Ai ∈ ℝ p× p , are strictly equivalent if there exist non singular matrices
M , N ∈ ℝ p× p such that
sE1 − A1 = M ( sE2 − A2 ) N .
{
Lε ( s ) = blockdiag Lε1 , Lε 2 ,..., Lε r } {
Lη ( s ) = blockdiag Lη1 , Lη2 ,..., Lηl }
Lε i ( s ) = sM ε i − N ε i ∈ R[ s ]ε i ×(ε i +1) , Lηi ( s ) = sM ηTi − NηTi ∈ R[ s ](ηi +1)×ηi
1 0 ⋯ 0 0 1 ⋯ 0
M ε i = ⋮ ⋱ ⋱ ⋮ ∈ R ε i ×(ε i +1) ; Nηi = ⋮ ⋱ ⋱ ⋮ ∈ Rηi ×(ηi +1) .
0 ⋯ 1 0 0 ⋯ 0 1
ε i (ηi ) are the right (left) Kronecker indices
Restricted system equivalence (Rosenbrock 1974)
Let the descriptor state space representations of the form
ρ E1 x1 (t ) = A1 x1 (t ) + B1u(t )
y (t ) = C1 x (t ) + D1u (t )
where ρ := d / dt , E1 , A1 ∈ ℝ p× p , B1 ∈ ℝ p×m , C1 ∈ ℝ l × p , D1 ∈ ℝ m×l ,
x1 (t ) : ℝ → ℝ p is the state vector and u (t ) : ℝ → ℝ l , y (t ) : ℝ → ℝ m are
respectively the input and output vectors.
Premultiplying by M x1 (t ) = Nx2 (t )
It is easy to see that the above change of coordinates results in a new description
of the form
ρ Ex2 (t ) = A2 x2 (t ) + B2u (t )
y (t ) = C2 x2 (t ) + D2u (t )
where E2 = ME1 N , A2 = MA1 N , B2 = MB1 , C2 = C1 N and D2 = D1 .
Restricted system equivalence (Rosenbrock 1974)
5 or in system matrix terms
sE2 − A2 B2 M 0 sE1 − A1 B1 N 0
−C =
2 D2 0 I l −C1 D1 0 I m
M −1 0 sE2 − A2 B2 sE1 − A1 B1 N 0
=
0 I l −C 2 D2 −C1 D1 0 I m
x1 ( t ) N 0 x2 ( t ) E1 x1 ( 0 ) M −1 0 E2 x2 ( 0 )
= =
( ) 0
− u t I m −u ( t ) 2( ) 0
y t I m y1 ( t )
Controllability Observability X u2
Yu
x1 ( 0 ) = Nx2 ( 0 ) ⇒ E1 x1 ( 0 ) = E1 Nx2 ( 0 ) = M −1 E2 x2 ( 0 ) X u1
Commute
Strong equivalence – Verghese et.al. 1981
Disadvantage of strict equiv. : connect matrix pencils of same dimension
Strong equivalence
sE1 − A1 0
( sE2 − A2 ) = M N
0 I
Operations of strong equivalence (can eliminate non-dynamic variables)
sE2 − A2 B2 M 0 sE1 − A1 B1 N R
=
−C
2 D2 Q I l −C1 D1 0 I m
det M ≠ 0,det N ≠ 0 QE1 = 0 = E1 R
Definition [Anderson et. al. 1985] Two descriptor systems are constant system
equivalent if there exist invertible matrices M , N and constant matrices X , Y of
appropriate dimensions such that
I 0 0 I 0 0
M 0 N Y
X I 0 sE1 − A1 − B1 = 0 sE2 − A2 − B2 0 I
m
0 C D 0 C D l
1 1 2 2
Complete equivalence – Pugh et.al. 1987
Strict equivalence preserves both the finite and infinite elementary
divisors.
Definition [Pugh et. al. 1987] Two matrix pencils sEi − Ai , i = 1,2, with
Ei , Ai ∈ P( m, l ), are completely equivalent if there exist constant matrices M , N
of appropriate dimensions such that
M ( sE1 − A1 ) = ( sE2 − A2 ) N
where the compound matrices
sE1 − A1
[ M sE 2 − A2]
,
− N
have full rank and no zeros in ℂ ∪ {∞}.
Definition [Pugh et. al. 1987] Two descriptor systems are completely system
equivalent if there exist constant matrices M , N , X , Y of appropriate
dimensions such that
M 0 sE1 − A1 − B1 sE2 − A2 − B2 N Y
X I C = 0 I
m 1 D1 C 2 D 2 l
where the compound matrices
sE1 − A1
[ M sE2 − A2 ] ,
− N
have full rank and no zeros in ℂ ∪ {∞} .
sE1 − A1
• Finite and Infinite Output Decoupling Zeros
−C1
• Finite and Infinite Input-Output Decoupling Zeros
r
Definition [Hayton et. al. 1986] Let P1 , P2 be two descriptor systems. They are
said to be fundamentally equivalent if there exists :
a) a constant injective map
x2 ( s ) N Y x1 ( s )
= −u ( s )
− u ( s ) 0 I
b) a constant surjective map
E2 x2 ( 0 − ) M 0 E1 x1 ( 0 − )
= − y (s)
− y ( s ) X I
Ai ( ρ ) Bi ( ρ ) 0 β i ( t ) 0
−
i C ( ρ ) Di ( ρ ) I − u ( t ) = 0 u (t )
0 − I 0 y (t ) I
Ti ( ρ ) ξi ( t ) Ui
y ( t ) = ( 0 0 I ) ξi ( t )
Vi
Definition [Hayton et. al. 1990] The normalized system matrices Pi ( s ) i = 1,2
are normal full system equivalent if there exist polynomial matrices M ( s ) ,
N ( s ) , X ( s ) , Y ( s ) such that the relation
M ( s ) 0 T1 ( s ) − U1 T2 ( s ) − U 2 N ( s ) Y ( s )
X ( s) I V = 0
1 0 2V 0 I l
is a full equivalence relation.
Definition [Hayton et. al. 1990] The normalized system matrices Pi ( s ) i = 1,2
are full system equivalent if there exist polynomial matrices M ( s ) , N ( s ) ,
X ( s ) , Y ( s ) such that the relation
M ( s ) 0 A1 ( s ) B1 ( s ) A2 ( s ) B2 ( s ) N ( s ) Y ( s )
X ( s ) I −C ( s ) D ( s ) = −C ( s ) D ( s ) 0 I
1 1 2 2 l
is a full equivalence relation.
Two PMDs are fundamentally equivalent if and only if they are normal
full system equivalent.
Full Unimodular equivalence – Karampetakis and Vardulakis 1992
AR2 ( s ) AR2 ( s )
AR2 ( s )
where − BR2 ( s ) has no infinite zeros and δ M − BR2 ( s ) = δ M .
− BR ( s )
T (s) T (s) 2
R R
Full Unimodular equivalence – Karampetakis and Vardulakis 1992
( )
Theorem [Antoniou et. al. 1998] If CF ∈ ℝ r×n , J f ∈ ℝ n×n is a finite (resp.
( )
C∞ ∈ ℝ r×µ , J ∞ ∈ ℝ µ ×µ is an infinite) Jordan pair of A( s ) ∈ ℝ [ s ]
r ×r
, where
n = deg det ( A ( s ) ) µ
(resp. = rankcol C∞ J ∞( )
k µ −1
k =0
) then the columns of the matrix
( )
Ψ F k = CF J Fk , k = 0,1,..., N
( )
Ψ ∞ k = C∞ J ∞N − k , k = 0,1,..., N
are linearly independent solutions of A (σ ) β ( k ) = 0 for N ≥ n ( N ≥ µ ) .
dim B = n + µ = rq
r ×r
A ( s ) = A0 + A1s + ⋯ + Aq s ∈ ℝ [ s ] q
Divisor equivalence – Karampetakis et.al. 2004