Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Notes of robust control<br />
• Zeros (continuous-time systems)<br />
• Parametrization (continuous-time systems)<br />
• H2 control (continuous-time systems)<br />
• H∞ control (continuous-time systems)<br />
• H∞ filtering in discrete-time (discrete-time systems)
.<br />
ZEROS AND POLES OF<br />
MULTIVARIABLE SYSTEMS<br />
• Transmission zeros<br />
SUMMARY<br />
Rank property and time-domain characterization<br />
• Invariant zeros<br />
Rank property and time-domain characterization<br />
• Decoupling zeros<br />
• Infinite zero structure
.<br />
(Transmission) Zeros and poles of a rational<br />
matrix<br />
Basic assumption:<br />
−1<br />
G ( s)<br />
= C(<br />
sI − A)<br />
B + D,<br />
p outputs,<br />
[ G(<br />
s)<br />
] = min( p,<br />
m)<br />
r<br />
normal rank<br />
=<br />
Smith-McMillan form of G(s)<br />
⎡ f1(<br />
s)<br />
⎢<br />
⎢<br />
0<br />
M ( s)<br />
= ⎢ M<br />
⎢<br />
⎢ 0<br />
⎢<br />
⎣ 0<br />
εi(<br />
s)<br />
fi<br />
( s)<br />
= , i = 1,<br />
2,<br />
L,<br />
r<br />
ψ ( s)<br />
εi divides εi+1<br />
ψi+1 divides ψi<br />
i<br />
f<br />
2<br />
0<br />
( s)<br />
M<br />
0<br />
0<br />
L<br />
L<br />
O<br />
L<br />
L<br />
f<br />
r<br />
0<br />
0<br />
M<br />
( s)<br />
0<br />
0⎤<br />
0<br />
⎥<br />
⎥<br />
M⎥<br />
⎥<br />
0⎥<br />
0⎥<br />
⎦<br />
• The transmission zeros are the roots of e i(s), i=1,2,…,r<br />
m inputs<br />
• The (transmission) poles are the roots of y i(s), i=1,2,…,r
.<br />
Example<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡<br />
=<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡−<br />
=<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎣<br />
⎡<br />
=<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎣<br />
⎡<br />
−<br />
−<br />
=<br />
1<br />
1<br />
,<br />
0<br />
1<br />
0<br />
0<br />
0<br />
0<br />
5<br />
13<br />
,<br />
0<br />
2<br />
1<br />
0<br />
,<br />
0<br />
1<br />
1<br />
1<br />
0<br />
5<br />
0<br />
0<br />
0<br />
0<br />
7<br />
10<br />
0<br />
0<br />
1<br />
0<br />
D<br />
C<br />
B<br />
A<br />
Hence<br />
)<br />
5<br />
)(<br />
2<br />
(<br />
1<br />
)<br />
2<br />
)(<br />
3<br />
(<br />
)<br />
1<br />
)(<br />
3<br />
(<br />
)<br />
(<br />
−<br />
−<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡<br />
−<br />
−<br />
+<br />
−<br />
=<br />
s<br />
s<br />
s<br />
s<br />
s<br />
s<br />
s<br />
G<br />
)<br />
5<br />
)(<br />
2<br />
(<br />
1<br />
0<br />
3<br />
)<br />
(<br />
,<br />
1<br />
−<br />
−<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡ −<br />
=<br />
=<br />
s<br />
s<br />
s<br />
s<br />
M<br />
r<br />
There is one transmission zero at s=3 and two poles in s=2 and s=5.
Theorem 1<br />
.<br />
Rank property of transmission zeros<br />
The complex number λ is a transmission zero of G(s) iff there exists<br />
polynomial vector z(s) such that z(λ)≠0 and<br />
⎧ lim G ( s)<br />
z(<br />
s)<br />
= 0<br />
⎪<br />
s → λ<br />
⎨<br />
lim G(<br />
s)'<br />
z(<br />
s)<br />
= 0<br />
⎪<br />
⎪⎩<br />
s → λ<br />
p ≥ m<br />
p ≤ m<br />
Remark<br />
In the square case (p=m), it follows<br />
det<br />
[ G ( s)<br />
]<br />
∏<br />
= r<br />
r<br />
i = 1<br />
∏<br />
i = 1<br />
ε<br />
ψ<br />
i<br />
i<br />
( s)<br />
( s)<br />
if<br />
if<br />
so that it is possible to catch all zeros and poles if and only if there are not<br />
cancellations.<br />
Proof of the theorem<br />
Assume that p≥ m and let M(s) be the Smith-Mc Millan form of G(s), i.e. G(s)=L(s)M(s)R(s) for suitable polynomial<br />
and unimodular matrices L(s) (of dimension p) and R(s) (of dimensions m). It is possible to write:<br />
where E(s)=diag{εi(s)} and Ψ(s)=diag{ψi(s)}. Now assume that λ is a zero of the polynomial eκ(s). Recall that R(s) is<br />
unimodular and take the polynomial vector z(s)=[R -1 (s)]k, i.e. the k-th column of the inverse of R(s). Since ψk(λ)≠0, it<br />
follows that G(s)z(s)=L(s)E(s)Ψ −1 (s)R(s)z(s)=L(s)E(s) [Ψ −1 (s)]k= L(s)[E(s)]k/ψ k (s)=[L(s)]kε k (s) /ψ k (s). This quantity<br />
goes to zero as s tends to λ. Vice-versa, if z(s) is such that G(s)z(s) goes to zero as s tends to λ, then the limit as s tends<br />
to λ of E(λ)Y -1 −1<br />
⎡E( s)<br />
Ψ ( s)<br />
⎤<br />
M<br />
( s)<br />
= ⎢ ⎥<br />
⎣ 0 ⎦<br />
(s)R(λ)z(λ) being zero means that at least one polynomial εk(s) should vanish at s= λ.<br />
The proof in the opposite case (p≤m) follows the same lines and therefore is omitted.
.<br />
Example<br />
G<br />
⎡1<br />
s)<br />
= ⎢<br />
⎣0<br />
( s −1)(<br />
s + 2)<br />
⎤ 1<br />
( s −1)<br />
⎥<br />
⎦ ( s −1)(<br />
s + 2)<br />
( 2<br />
Hence,<br />
⎡1<br />
G(<br />
s)<br />
= ⎢<br />
⎣0<br />
so that, with<br />
⎡ 1<br />
0⎤⎢<br />
( s −1)(<br />
s +<br />
⎢<br />
1<br />
⎥<br />
⎦⎢<br />
0<br />
⎢⎣<br />
⎡ − ( s −1)(<br />
s + 2)<br />
⎤<br />
z(<br />
s)<br />
= ⎢<br />
⎥<br />
⎣ 1 ⎦<br />
it follows<br />
lim G(<br />
s)<br />
z(<br />
s)<br />
= 0<br />
s →1<br />
2)<br />
Notice that it is not true that G(s)z(1) tends to zero as s tends to one. .<br />
⎤<br />
0<br />
⎥ 1<br />
⎥<br />
s −1<br />
⎥ ( s −1)(<br />
s + 2)<br />
s + 2⎥⎦
.<br />
Time-domain characterization of transmission<br />
zeros and poles<br />
Let Σ=(Α,Β, C ,D) be the state-space system with transfer function<br />
G(s)=C(sI-A) -1 B+D. Let Σ′=(A′,C′,B′,D′) the “transponse” system whose<br />
transfer function is G(s) ′=B′(sI-A′) -1 C′+D′. Now, let<br />
Σˆ<br />
⎧ Σ<br />
= ⎨<br />
⎩ Σ '<br />
Finally, let δ i (t) denote the i-th derivative of the impulsive “function”.<br />
Theorem 2<br />
p<br />
p<br />
≥<br />
≤<br />
m<br />
m<br />
Gˆ<br />
( s )<br />
⎧ G ( s )<br />
⎨<br />
⎩ G ( s )'<br />
(a) The complex number λ is a pole of Σ iff the exists an impulsive input<br />
such that the forced output yf(t) of Σ is<br />
(b) The complex number λ is a zero of Σ iff the exists an<br />
exponential/impulsive input<br />
such that the forced output yf(t) of Σ is<br />
=<br />
u ( t )<br />
y<br />
f<br />
y f<br />
Zeros blocking property. The proof of the theorem is not reported here for space limitations.<br />
∑<br />
i = 0<br />
= ν<br />
α<br />
p<br />
p<br />
i<br />
δ<br />
≥<br />
≤<br />
i<br />
m<br />
m<br />
( t )<br />
λt<br />
( t ) = y 0e , ∀ t ≥<br />
λ t<br />
i<br />
u ( t ) u 0 e + α<br />
iδ<br />
( t ) , u 0<br />
= ∑<br />
i = 0<br />
ν<br />
( t ) = 0 , ∀ t ≥<br />
0<br />
0<br />
≠<br />
0
.<br />
Invariant zeros<br />
Given a system (A,B,C,D) with transfer function G(s)=C(sI-A) -1 B+D, the<br />
polynomial matrix<br />
P ( s )<br />
=<br />
⎡ sI − A −<br />
⎢<br />
⎣ C D<br />
is called the system matrix.<br />
Consider the Smith form S(s) of P(s):<br />
⎡α1(<br />
s)<br />
⎢<br />
⎢<br />
0<br />
S(<br />
s)<br />
= ⎢ M<br />
⎢<br />
⎢ 0<br />
⎢<br />
⎣ 0<br />
αi divides αi+1.<br />
α<br />
2<br />
0<br />
( s)<br />
M<br />
0<br />
0<br />
L<br />
L<br />
O<br />
L<br />
L<br />
αν<br />
0<br />
0<br />
M<br />
( s)<br />
0<br />
0⎤<br />
0<br />
⎥<br />
⎥<br />
M⎥<br />
⎥<br />
0⎥<br />
0⎥<br />
⎦<br />
The invariant zeros are the roots of αi(s), i=1,2,…,ν<br />
−−−−−−−−−−−−−−−−−<br />
Notice that<br />
P ( s)<br />
=<br />
⎡sI −<br />
⎢<br />
⎣ C<br />
A<br />
so that the rank of P(s) is ϖ=n+rank(G(s))<br />
B<br />
⎤<br />
⎥<br />
⎦<br />
0⎤<br />
⎡( sI − A )<br />
I<br />
⎥ ⎢<br />
⎦ ⎣ 0<br />
−1<br />
0 ⎤ ⎡sI<br />
−<br />
⎥<br />
G ( s)<br />
⎢<br />
⎦ ⎣ 0<br />
A<br />
− B ⎤<br />
I<br />
⎥<br />
⎦
Theorem 3<br />
.<br />
Rank property of invariant zeros<br />
The complex number λ is an invariant zero of (A,B,C,D) iff P(s) looses<br />
rank in s=λ, i.e. iff there exists a nonzero vector z such that<br />
⎧ P ( λ ) z = 0<br />
⎪<br />
⎨<br />
P<br />
⎪<br />
⎪⎩<br />
( λ )' z = 0<br />
if<br />
if<br />
Notice that if T is the transformation for a change of basis in the state-space, then<br />
P ( s ) =<br />
⎡ sI − A − B<br />
⎢<br />
⎣ C D<br />
⎤<br />
⎥<br />
⎦<br />
=<br />
⎡T<br />
⎢<br />
⎣ 0<br />
Therefore, the invariant zeros are invariant with respect to a change of basis.<br />
p<br />
p<br />
≥<br />
≤<br />
m<br />
m<br />
0⎤<br />
⎡sI − A −<br />
I<br />
⎥ ⎢<br />
⎦ ⎣ C D<br />
⎤ ⎡T<br />
⎥ ⎢<br />
⎦ ⎣<br />
Proof of Theorem 3<br />
Consider the case p≥m and let P(s)=L(s)S(s)R(s), where R(s) and R(s) are suitable polynomial<br />
unimodular matrices. Hence, if P(λ)z=0 for some z≠0, then S(λ )R(λ)z=0 so that S(λ)y=0, with y≠0.<br />
Hence, λ must be the root of at least one polynomial αi(s). Vice-versa, if αk(λ) for some k, then<br />
z=[R -1 (λ)]k is such that P(λ)z=L(λ)S(λ)R(λ)z= L(λ)S(λ)R(λ)[R -1 (λ)]k=L(λ)[S(λ)]k=0.<br />
B<br />
−1<br />
0<br />
0 ⎤<br />
⎥<br />
I ⎦
.<br />
Time-domain characterization of invariant zeros<br />
Let Σ=(Α,Β, C ,D) be the state-space system with transfer function<br />
G(s)=C(sI-A) -1 B+D. Let Σ′=(A′,C′,B′,D′) the “transponse” system whose<br />
transfer function is G(s) ′=B′(sI-A′) -1 C′+D′. Now, let<br />
⎧ Σ<br />
Σˆ<br />
= ⎨<br />
⎩Σ<br />
'<br />
Theorem 4<br />
p ≥<br />
p ≤<br />
m<br />
m<br />
G s<br />
Gˆ<br />
⎧ ( )<br />
( s)<br />
= ⎨<br />
⎩G<br />
( s)'<br />
The complex number λ is an invariant zero Σ iff at leat one of the two<br />
following conditions holds:<br />
(i) λ is an eigenvalue of the unobservable part of Σ<br />
(ii) there exist two vectors x0 and u0≠0 such that the forced output of Σ<br />
corresponding to the input u(t)=u0e λt , t≥0 and the initial state<br />
x(0)=x0 is identically zero for t≥0.<br />
p ≥<br />
p ≤<br />
Zeros blocking property. The proof of the theorem is not reported here for space limitations.<br />
m<br />
m
.<br />
Trasmission vs invariant zeros<br />
Theorem 5<br />
(i) A transmission zero is also an invariant zero.<br />
(ii) Invariant and transmission zeros of a system in minimal form do<br />
coincide.<br />
Proof of point (i). Let p≥m and let λ be a transmission zero. Then, there exists an<br />
exponential/impulsive input<br />
λt<br />
ν<br />
∑<br />
i=<br />
0<br />
( i)<br />
u ( t)<br />
= u e + α δ ( t)<br />
such that the forced output is zero, i.e.<br />
y<br />
f<br />
0<br />
t<br />
i<br />
ν<br />
A(<br />
t−τ)<br />
⎡ λτ<br />
( i)<br />
⎤ λt<br />
( t)<br />
= C∫<br />
e B⎢u<br />
0e<br />
+ ∑ α iδ<br />
( τ)<br />
⎥dτ<br />
+ Du0e<br />
= 0,<br />
t > 0 . Letting<br />
⎣<br />
i=<br />
0 ⎦<br />
0<br />
⎡α0<br />
⎤<br />
⎢ ⎥<br />
ν<br />
[ ] ⎢<br />
α1<br />
x =<br />
⎥<br />
0 B AB L A B , it follows<br />
⎢ M ⎥<br />
⎢ ⎥<br />
⎣αν<br />
⎦<br />
Ce<br />
t<br />
At<br />
x0<br />
C<br />
A(<br />
t−τ<br />
) λτ<br />
λt<br />
e Bu0e<br />
dτ<br />
+ Du0e<br />
+ ∫<br />
0<br />
=<br />
0,<br />
t > 0<br />
so that λ is an invariant zero. The proof in the case p≤m is the same, considering the transponse<br />
system.<br />
Proof of point (ii). Let p≥m. We have already proven that a trasmission zero is an invariant zero.<br />
Then, consider a minimum system and an invariant zero λ. Hence Av+Bw=λw, Cv+Dw=0. Notice<br />
that w≠0 since the system is observable. Moreover, thanks to the reachability condition there exists<br />
a vector<br />
⎡α<br />
0 ⎤<br />
⎢ ⎥<br />
ν<br />
[ ] ⎢<br />
α1<br />
v = B AB L A B ⎥<br />
⎢ M ⎥<br />
⎢ ⎥<br />
⎣αν<br />
⎦<br />
Since λ is an invariant zero there exists an initial state x(0)=v and an input u(t)=we λt , such that<br />
Ce<br />
t<br />
At A<br />
λ<br />
v + C∫<br />
e<br />
By noticing that<br />
C<br />
0<br />
( t−τ<br />
) λτ<br />
t<br />
Bwe dτ<br />
+ Dwe = 0,<br />
t > 0<br />
Ce<br />
At<br />
ν<br />
∑<br />
t<br />
At i<br />
A(<br />
t−τ)<br />
( i)<br />
v = Ce A Bαi<br />
= Ce B αiδ<br />
( τ)<br />
dτ<br />
, it follows,<br />
i=<br />
0<br />
t<br />
ν<br />
A( t−τ)<br />
⎡ λτ<br />
λ<br />
∫ e B⎢u<br />
0e<br />
+ ∑<br />
0 ⎣<br />
i=<br />
0<br />
trransmission zero.<br />
∫<br />
0<br />
ν<br />
∑<br />
i=<br />
0<br />
( i)<br />
⎤<br />
t<br />
α iδ<br />
( τ)<br />
⎥dτ<br />
+ Du0e<br />
= 0,<br />
t > 0,<br />
which means that λ is a<br />
⎦
Associated Smith forms<br />
.<br />
S<br />
P C<br />
C<br />
Decoupling zeros<br />
⎡ sI − A ⎤<br />
( s ) = ⎢ ⎥ PB ( s)<br />
= [ sI − A − B]<br />
⎣ C ⎦<br />
( s )<br />
=<br />
⎡<br />
⎢<br />
⎣<br />
diag<br />
C { α ( s ) } ⎤<br />
B<br />
i<br />
⎥⎦ S B ( s ) =<br />
diag { α i ( s ) }<br />
0<br />
[ 0 ]<br />
The output decoupling zeros are the roots of α1 C (s) α2 C (s)… αn C (s), the<br />
input decoupling zeros are the roots of αi B (s) α2 B (s)…αn B (s) and the inputoutput<br />
decoupling zeros are the common roots of both.<br />
Lemma<br />
(i) A complex number λ is an output decoupling zero iff there exists<br />
z≠0 such that PC(λ)z=0.<br />
(ii) A complex number λ is an input decoupling zero iff there exists<br />
w≠0 such that P B(λ)′w=0.<br />
(iii) A complex number λ is an input-output decoupling zero iff there<br />
exist z≠0 and w≠0 such that PC(λ)z=0, PB(λ)′w=0.<br />
(iv) For square systems the set of decoupling zeros is a subset of the<br />
set of invariant zeros.<br />
(v) For square invertible systems the transmission zeros of the inverse<br />
system coincide with the poles of the system and the invariant<br />
zeros of the inverse system coincide with the eigenvalues of the<br />
system.
.<br />
Infinite zero structure<br />
Mc Millan form over the ring of causal rational matrices<br />
Given the transfer function G(s) there exists two unimodular matrices in<br />
that ring (bi-proper rational functions) such that<br />
⎡M<br />
( s)<br />
0⎤<br />
G(<br />
s)<br />
= V ( s)<br />
⎢ ⎥W<br />
( s),<br />
M ( s)<br />
= diag<br />
,<br />
⎣ 0 0⎦<br />
where δ1, δ2, …, δr are integers such that δ1≤ δ2≤…≤ δr. They describes the<br />
structure of the infinity zeros.<br />
• Inversion<br />
• Model matching<br />
• ……<br />
Example: compute the finite and infinite zeros of:<br />
⎡s<br />
+ 1<br />
⎢<br />
⎢s<br />
− 1<br />
G(<br />
s)<br />
= ⎢ 1<br />
⎢<br />
⎢ 1<br />
⎢<br />
⎣ s<br />
0<br />
1<br />
s<br />
0<br />
s<br />
3<br />
s + 1 ⎤<br />
⎥<br />
s − 1 ⎥<br />
s + 1 ⎥<br />
s ⎥<br />
2<br />
− 3s<br />
+ s + 5⎥<br />
3 ⎥<br />
s ( s − 3)<br />
⎦<br />
{ } r<br />
−δ1<br />
−δ2<br />
−δ<br />
s , s , L s
PARAMETRIZATION OF STABILIZING<br />
CONTROLLERS<br />
Summary<br />
• BIBO stability and internal stability of feedback systems<br />
• Stabilization of SISO systems. Parameterization and<br />
characterization of stabilizing controllers<br />
• Coprime factorization<br />
• Stabilization of MIMO systems. Parameterization and<br />
characterization of stabilizing controllers<br />
• Strong stabilization
Stability of feedback systems<br />
d<br />
r e u y<br />
C(s) P(s)<br />
The problem consists in the analysis of the asymptotic stability<br />
(internal stability) of closed-loop system from some characteristic<br />
transfer functions. This result is useful also for the design.<br />
THEOREM 1<br />
The closed loop system is asymptotically stable if and only if the<br />
transfer matrix G(s) from the input to the output<br />
is stable.<br />
G(<br />
s)<br />
⎡r<br />
⎤<br />
⎢ ⎥<br />
⎣d<br />
⎦<br />
−1<br />
⎡ ( I + P(<br />
s)<br />
C(<br />
s))<br />
⎢<br />
−1<br />
⎣(<br />
I + C(<br />
s)<br />
P(<br />
s))<br />
C(<br />
s)<br />
⎡e⎤<br />
⎢ ⎥<br />
⎣u⎦<br />
−1<br />
− ( I + P(<br />
s)<br />
C(<br />
s))<br />
P(<br />
s)<br />
⎤<br />
⎥<br />
( I + C(<br />
s)<br />
P(<br />
s))<br />
⎦<br />
= −1
Stability of interconnected SISO systems<br />
The closed loop system is asymptotically stable if and only if the<br />
three transfer functions<br />
S(<br />
s)<br />
=<br />
1<br />
are stable.<br />
+<br />
1<br />
, R(<br />
s)<br />
C(<br />
s)<br />
P(<br />
s)<br />
=<br />
1<br />
+<br />
C(<br />
s)<br />
, H ( s)<br />
C(<br />
s)<br />
P(<br />
s)<br />
P(<br />
s)<br />
C(<br />
s)<br />
P(<br />
s)<br />
• Notice that if S(s), R(s), H(s) are stable, then also the<br />
complementary sensitivity<br />
C(<br />
s)<br />
P(<br />
s)<br />
T ( s)<br />
=<br />
= 1 − S(<br />
s)<br />
1+<br />
C(<br />
s)<br />
P(<br />
s)<br />
is stable.<br />
• The stability of the three transfer functions are made to avoid<br />
prohibited (unstable) cancellations between C(s) and P(s).<br />
Example<br />
Let C(s)=(s-1)/(s+1), P(s)=1/(s-1) . Then S(s)=(s+1)/(s+2),<br />
R(s)=(s-1)/(s+2) e H(s)=(s+1)/(s 2 +s-2)<br />
=<br />
1<br />
+
Comments<br />
The proof of Theorem 1 can be carried out pursuing different<br />
paths. A simple way is as follows. Let assume that C(s) and P(s)<br />
are described by means of minimal realizations C(s)=(A,B,C,D),<br />
P(s)=(F,G,H,E). It is possible to write a realization of the closedloop<br />
system as Σcl=(Acl,Bcl,Ccl,Dcl). Obviously, if Acl is<br />
asymptotically stable, then G(s) is stable. Vice-versa, if G(s) is<br />
stable, asymptotic stability of Acl follows form being<br />
Σcl=(Acl,Bcl,Ccl,Dcl) a minimal realization. This can be seen<br />
through the well known PBH test.<br />
In our example, it follows<br />
A<br />
cl<br />
G(<br />
s)<br />
=<br />
0<br />
⎡<br />
⎢<br />
⎣−<br />
⎡<br />
⎢<br />
= ⎢<br />
⎢<br />
⎢⎣<br />
1<br />
s<br />
s<br />
s<br />
s<br />
+<br />
+<br />
−<br />
+<br />
−<br />
−<br />
1<br />
2<br />
1<br />
2<br />
2<br />
1<br />
⎤<br />
⎥<br />
⎦<br />
,<br />
( s<br />
−<br />
−<br />
B<br />
cl<br />
=<br />
⎡<br />
⎢<br />
⎣<br />
1<br />
1<br />
( s + 1)<br />
1)(<br />
s + 2)<br />
s + 1<br />
s + 2<br />
⎤<br />
⎥<br />
⎥<br />
⎥<br />
⎥⎦<br />
1<br />
0<br />
⎤<br />
⎥<br />
⎦<br />
,<br />
SMM<br />
><br />
⎡−<br />
= ⎢<br />
⎣−<br />
Hence, Acl is unstable and the closed-loop system is reachable and<br />
observable. This means that G(s) is unstable as well.<br />
C<br />
cl<br />
⎡<br />
⎢<br />
⎢<br />
⎣<br />
( s<br />
−<br />
1<br />
1<br />
−<br />
1<br />
1)(<br />
s<br />
0<br />
0<br />
2<br />
+<br />
⎤<br />
⎥<br />
⎦<br />
,<br />
2)<br />
D<br />
cl<br />
( s<br />
=<br />
+<br />
⎡<br />
⎢<br />
⎣<br />
1<br />
1<br />
0<br />
1)(<br />
s<br />
−<br />
0<br />
1<br />
⎤<br />
⎥<br />
⎦<br />
1)<br />
⎤<br />
⎥<br />
⎥<br />
⎦
Parametrization of stabilizing controllers<br />
1° case: SISO systems and P(s) stable<br />
d<br />
r e u<br />
C(s) P(s)<br />
THEOREM 2<br />
The family of all controllers C(s) such that the closed-loop system<br />
is asymptotically stable is:<br />
C(<br />
s)<br />
=<br />
1<br />
−<br />
Q(<br />
s)<br />
P(<br />
s)<br />
Q(<br />
s)<br />
where Q(s) is proper and stable and Q(∞)P(∞) ≠1.
Proof of Theorem 2<br />
Let C(s) be a stabilizing controller and define<br />
Q(s)=C(s)/(1+C(s)P(s)). Notice that Q(s) is stable since it is the<br />
transfer function from r to u. Hence C(s) can be written as<br />
=Q(s)/(1-P(s)Q(s)), with Q(s) stable and, obviously, the stability<br />
of both Q(s) and P(s) implies that Q(∞)P(∞)=C(∞)/(1+C(∞)P(∞))<br />
≠1.<br />
Vice-versa, assume that Q(s) is stable and Q(∞)P(∞)≠1. Define<br />
C(s)=Q(s)/(1-P(s)Q(s)). It results that<br />
S(s)=1/(1+C(s)P(s))=1-P(s)Q(s),<br />
R(s)=C(s)/(1+C(s)P(s))=Q(s),<br />
H(s)=P(s)/(1+C(s)P(s))=P(s)(1-P(s)Q(s))<br />
are stable. This means that the closed-loop system is<br />
asymptotically stable.
Parametrization of stabilizing controllers<br />
2° case: SISO systems and generic P(s)<br />
It is always possible to write (coprime factorization)<br />
P ( s)<br />
=<br />
N(<br />
s)<br />
M ( s)<br />
where N(s) and M(s) are stable rational coprime functions, i.e.<br />
such that there exist two stable rational functions X(s) e Y(s)<br />
satisfying (equation of Diofanto, Aryabatta, Bezout)<br />
N(<br />
s)<br />
X ( s)<br />
+ M ( s)<br />
Y ( s)<br />
= 1<br />
THEOREM 3<br />
The family of all controllers C(s) such that the closed-loop system<br />
is well-posed and asymptotically stable is:<br />
C(<br />
s)<br />
=<br />
X ( s)<br />
Y(<br />
s)<br />
+<br />
−<br />
M ( s)<br />
Q(<br />
s)<br />
N(<br />
s)<br />
Q(<br />
s)<br />
where Q(s) is proper, stable and such that Q(∞)N(∞) ≠ Y(∞).
Comments<br />
The proof of the previous theorem can be carried out following<br />
different ways. However, it requires a preliminary discussion on<br />
the concept of coprime factorization and on the stability of<br />
factorized interconnected systems.<br />
d<br />
r e z1 u z2 y<br />
Mc(s) -1<br />
LEMMA 1<br />
Let P(s)=N(s)/M(s) e C(s)=Nc(s)/Mc(s) stable coprime<br />
factorizations. Then the closed-loop system is asymptotically<br />
stable if and only if the transfer matrix K(s) from [r′ d′]′ to [z1′<br />
z2′]′ is stable.<br />
K(<br />
s)<br />
M(s) -1<br />
Nc(s) N(s)<br />
⎡<br />
⎢<br />
⎣<br />
r<br />
d<br />
=<br />
⎤<br />
⎥<br />
⎦<br />
⎡<br />
⎢<br />
⎣−<br />
M<br />
c<br />
N<br />
( s)<br />
c<br />
( s)<br />
⎡<br />
⎢<br />
⎣<br />
z<br />
z<br />
1<br />
2<br />
⎤<br />
⎥<br />
⎦<br />
N(<br />
s)<br />
M ( s)<br />
⎤<br />
⎥<br />
⎦<br />
−<br />
1
Proof of Lemma 1<br />
Define four stable functions X(s),Y(s),Xc(s),Yc(s) such that<br />
X ( s)<br />
N(<br />
s)<br />
+ Y ( s)<br />
M ( s)<br />
= 1<br />
X c c c c<br />
( s)<br />
N ( s)<br />
+ Y ( s)<br />
M ( s)<br />
= 1<br />
To proof the Lemma it is enough to resort to Theorem 1, by<br />
noticing that the transfer functions K(s) and G(s) (from [r′ d′]′ to<br />
[e′ u′]′) are related as follows:<br />
K(<br />
s)<br />
=<br />
⎡<br />
⎢<br />
⎣−<br />
Y ( s)<br />
c<br />
X ( s)<br />
X<br />
c<br />
( s)<br />
Y ( s)<br />
⎤<br />
⎥<br />
⎦<br />
G(<br />
s)<br />
⎡<br />
− ⎢<br />
⎣−<br />
⎡I<br />
0⎤<br />
⎡ 0 − N ⎤<br />
G(<br />
s)<br />
=<br />
⎢ ⎥ + ⎢ ⎥ K(<br />
s)<br />
0 I N 0<br />
⎣<br />
⎦<br />
⎣<br />
c<br />
⎦<br />
0<br />
X<br />
X<br />
0<br />
c<br />
⎤<br />
⎥<br />
⎦
Proof of Theorem 3<br />
Assume that Q(s) is stable and Q(∞)N(∞) ≠ Y(∞). Moreover,<br />
define C(s)=(X(s)+M(s)Q(s))/(Y(s)-N(s)Q(s)). It follows that<br />
1=N(s)X(s)+M(s)Y(s)=N(s)(X(s)+M(s)Q(s))+M(s)(Y(s)-<br />
N(s)Q(s))<br />
so that the functions X(s)+M(s)Q(s) e Y(s)-N(s)Q(s) defining C(s)<br />
are coprime (besides being both stable). Hence, the three<br />
characterist transfer functions are:<br />
S(s)=1/(1+C(s)P(s))=M(s)(Y(s)-N(s)Q(s)),<br />
R(s)=C(s)/(1+C(s)P(s))=M(s)(X(s)+M(s)Q(s)),<br />
H(s)=P(s)/(1+C(s)P(s))=N(s) (Y(s)-N(s)Q(s))<br />
Since they are all stable, the closed-loop system is asymptotically<br />
stable as well (Theorem 1).<br />
Vice-versa, assume that C(s)=Nc(s)/Mc(s) (stable coprime<br />
factorization) is such that the closed-loop system is well-posed<br />
and asymptotically stable. Define Q(s)=(Y(s)Nc(s)-<br />
X(s)Mc(s))/(M(s)Mc(s)+N(s)Nc(s)). Since the closed-loop system<br />
is asymptotically stable and (Nc,Mc) are coprime, then, in view of<br />
Lemma 1 it follows that<br />
Q(s) = (Y(s)Nc(s)-X(s)Mc(s))/(M(s)Mc(s)+N(s)Nc(s)) =<br />
= [0 I]K(s)[Y(s)′-X(s)′]′<br />
is stable. This leads to<br />
C(s)=(X(s)+M(s)Q(s))/(Y(s)-N(s)Q(s)).<br />
Finally, Q(∞)N(∞) ≠ Y(∞), as can be easily be verified.
Coprime Factorization<br />
1° case: SISO systems<br />
Lemma 1 makes reference to a factorized description of a transfer<br />
function. Indeed, it is easy to see that it is always possible to write<br />
a transfer function as the ratio of two stable transfer functions<br />
without common divisors (coprime). For example<br />
with<br />
It results<br />
with<br />
s + 1<br />
G(<br />
s)<br />
= = N(<br />
s)<br />
M ( s)<br />
( s − 1)(<br />
s + 10)(<br />
s − 2)<br />
N(<br />
s)<br />
=<br />
( s<br />
+<br />
1<br />
10)(<br />
s<br />
+<br />
1)<br />
N(<br />
s)<br />
X(<br />
s)<br />
+ M(<br />
s)<br />
X ( s)<br />
=<br />
X ( s)<br />
=<br />
Euclide’s Algorithm<br />
4(<br />
16s<br />
− 5)<br />
,<br />
( s + 1)<br />
,<br />
Y ( s)<br />
M ( s)<br />
1<br />
=<br />
( s<br />
( s<br />
=<br />
+<br />
+<br />
−<br />
( s − 1)(<br />
s<br />
( s + 1)<br />
15)<br />
10)<br />
1<br />
−<br />
2<br />
2)
Coprime Factorization<br />
1° case: MIMO systems<br />
In the MIMO case, we need to distinguish between right and left<br />
factorizations.<br />
Right and left factorization<br />
Given P(s), find four proper and stable transfer matrices such that:<br />
− 1<br />
−1<br />
P( s)<br />
= Nd<br />
( s)<br />
M d ( s)<br />
= M s(<br />
s)<br />
N s(<br />
s)<br />
Nd ( s),<br />
M d ( s)<br />
right coprime X d ( s)<br />
Nd<br />
( s)<br />
+ Yd<br />
( s)<br />
M d(<br />
s)<br />
Ns ( s),<br />
M s(<br />
s)<br />
left coprime N s(<br />
s)<br />
X s(<br />
s)<br />
+ M s(<br />
s)<br />
Ys<br />
( s)<br />
Double coprime factorization<br />
Given P(s), find eight proper and stable transfer functions such<br />
that:<br />
− 1<br />
−1<br />
( i) P(<br />
s)<br />
= Nd<br />
( s)<br />
M d(<br />
s)<br />
= M s(<br />
s)<br />
Ns<br />
( s)<br />
( ii)<br />
⎡<br />
⎢<br />
⎣<br />
Y<br />
d<br />
( s)<br />
X<br />
( s)<br />
M<br />
( s)<br />
( s)<br />
− N ( s)<br />
M ( s)<br />
N ( s)<br />
Y ( s)<br />
s<br />
Hermite’s algorithm<br />
Observer and control law<br />
d<br />
s<br />
⎤ ⎡ d −<br />
⎥ ⎢<br />
⎦ ⎣ d<br />
X<br />
s<br />
s<br />
⎤<br />
⎥<br />
⎦<br />
=<br />
I<br />
=<br />
I<br />
=<br />
I
Construction of a stable double coprime<br />
factorization<br />
Let (A,B,C,D) a stabilizable and detectable realization of P(s) and<br />
conventionally write P=(A,B,C,D) .<br />
THEOREM 4<br />
Let F e L two matrices such that A+BF e A+LC are stable. Then<br />
there exists a stable double coprime factorization, given by:<br />
Md=(A+BF,B, F,I), Nd=(A+BF,B,C+DF,D)<br />
Ys=(A+BF,L,-C-DF,I), Xs=(A+BF,L,F,0)<br />
Ms=(A+ LC,L,C,I), Ns=(A+LC,B+LD,C,D)<br />
Yd=(A+LC,B+LD,- F,I), Xd=(A+LC,L,F,0)
Proof of Theorem 4<br />
The existence of stabilizing F e L is ensured by the assumptions of<br />
stabilizability and detectability of the system. To check that the 8<br />
transfer functions form a stable double coprime factorization<br />
notice first that they are all stable and that Md e Ms are biproper<br />
systems. Finally one can use matrix calculus to verify the theorem.<br />
Alternatively, suitable state variables can be introduced. For<br />
example, from<br />
x&<br />
y<br />
u<br />
=<br />
=<br />
=<br />
Ax<br />
Cx<br />
Fx<br />
+<br />
+<br />
+<br />
Bu<br />
Du<br />
v<br />
=<br />
=<br />
( A<br />
( C<br />
+<br />
+<br />
control<br />
BF ) x<br />
DF)<br />
x<br />
law<br />
+<br />
+<br />
Bv<br />
Dv<br />
one obtains y=P(s)u=Nd(s)v, u=Md(s)v, so that P(s)Md(s)=Nd(s).<br />
Similarly,<br />
x&<br />
y<br />
=<br />
=<br />
Ax<br />
Cx<br />
⎧ =<br />
⎨<br />
⎩ η =<br />
θ &<br />
A<br />
C<br />
+<br />
+<br />
θ<br />
θ<br />
Bu<br />
Du<br />
+<br />
+<br />
Bu<br />
Du<br />
+<br />
−<br />
L<br />
η<br />
y<br />
observer<br />
implies η=Ns(s)u-Ms(s)y=Ns(s)u-Ms(s)P(s)y=0 (stable autonomous<br />
dynamic) so that Ns(s)=Ms(s)P(s). Analoguously one can proceed<br />
for the rest of the proof.
Parametrization of stabilizing controllers<br />
3° case: MIMO systems and generic P(s)<br />
− 1<br />
−1<br />
( i) P(<br />
s)<br />
= Nd<br />
( s)<br />
M d ( s)<br />
= M s(<br />
s)<br />
Ns<br />
( s)<br />
( ii)<br />
⎡ Yd<br />
( s)<br />
X d ( s)<br />
⎤ ⎡M d ( s)<br />
− X s(<br />
s)<br />
⎤<br />
⎢<br />
=<br />
N s s M s s<br />
⎥ ⎢<br />
Nd<br />
s Ys<br />
s<br />
⎥<br />
⎣−<br />
( ) ( ) ⎦ ⎣ ( ) ( ) ⎦<br />
THEOREM 5<br />
The family of all proper transfer matrices C(s) such that the<br />
closed-loop system is well-posed and asymptotically stable is:<br />
C(<br />
s)<br />
= ( X<br />
= ( Y<br />
d<br />
s<br />
C(s) P(s)<br />
( s)<br />
+ M<br />
( s)<br />
Q<br />
( s)<br />
− Q ( s)<br />
N ( s))<br />
s<br />
d<br />
s<br />
d<br />
( X<br />
I<br />
( s))(<br />
Y ( s)<br />
− N<br />
−1<br />
s<br />
d<br />
d<br />
( s)<br />
Q<br />
( s)<br />
+ Q ( s)<br />
M<br />
s<br />
d<br />
( s))<br />
s<br />
−1<br />
( s))<br />
where Qd(s) [Qs(s)] is stable and such that Nd(∞)Qd(∞)≠Ys(∞)<br />
[Qs(∞)Ns(∞)≠Yd(∞)].
Comments<br />
The proof of Theorem 5 follows the same lines as that of Theorem<br />
3. Thowever, the generalization of Lemma 1 to MIMO is required.<br />
d<br />
r e z1 u z2 y<br />
Mc(s) -1<br />
LEMMA 2<br />
Let P(s)=Nd(s)Md(s) -1 and C(s)=Nc(s)Mc(s) -1 stable right coprime<br />
factorizations. Then the closed-loop system is asymptotically<br />
stable if and only if the transfer matrix K(s) from [r′ d′]′ to [z1′<br />
z2′]′ is stable.<br />
⎡ M c(<br />
s)<br />
K(<br />
s)<br />
= ⎢<br />
⎣−<br />
Nc<br />
( s)<br />
Md(s) -1<br />
Nc(s) Nd(s)<br />
N<br />
M<br />
d<br />
d<br />
( s)<br />
⎤<br />
( s)<br />
⎥<br />
⎦<br />
−1
Proof of Lemma 2<br />
It is completely similar to the proof of Lemma 1.<br />
Proof of Theorem 5<br />
Assume that Q(s) is stable and define<br />
C(s)=(Xs(s)+Md(s)Q(s))(Ys(s)-Nd(s)Q(s)) -1<br />
It results that<br />
I=Ns(s)Xs(s)+Ms(s)Ys(s)<br />
=Ns(s)(Xs(s)+Md(s)Q(s))+Ms(s)(Y(s)s-Nd(s)Q(s))<br />
so that the functions Xs(s)+Md(s)Q(s) e Ys(s)-Nd(s)Q(s) defining<br />
C(s) are right coprime (besides being both stable). The four<br />
transfer matrices characterizing the closed-loop are:<br />
(1+PC) -1 = (Ys-NdQ)Ms<br />
-(1+PC) -1 P= (Ys-NdQ)Ns<br />
(1+CP) -1 C=C(I+PC) -1 = (Xs+MdQ)Ms<br />
(1+CP) -1 =I-C(I+PC) -1 C= I-(Xs+MdQ)Ns<br />
They are all stable so that the closed-loop system is asymptotically<br />
stable.<br />
Vice-versa, assume that C(s)=Nc(s)Mc(s) -1<br />
(right coprime<br />
factorization) is stabilizing. Notice that the matrices<br />
⎡ Yd<br />
( s)<br />
X d(<br />
s)<br />
⎤ ⎡M<br />
d ( s)<br />
Nc<br />
( s)<br />
⎤<br />
⎢<br />
and<br />
≈<br />
N s s M s s<br />
⎥ ⎢<br />
Nd<br />
s M c s<br />
⎥<br />
⎣−<br />
( ) ( ) ⎦ ⎣ ( ) ( ) ⎦<br />
have stable inverse. Hence,<br />
⎡ Yd<br />
( s)<br />
X d ( s)<br />
⎤ ⎡M<br />
d ( s)<br />
Nc<br />
( s)<br />
⎤ ⎡I<br />
Yd<br />
( s)<br />
N c(<br />
s)<br />
X d ( s)<br />
Nc<br />
( s)<br />
⎤<br />
⎢<br />
⎥ ⎢<br />
⎥ = ⎢<br />
⎥<br />
⎣−<br />
N s ( s)<br />
M s ( s)<br />
⎦ ⎣ N d ( s)<br />
M c ( s)<br />
⎦ ⎣0<br />
N s ( s)<br />
N c(<br />
s)<br />
+ M s ( s)<br />
M c ( s)<br />
⎦<br />
(*)<br />
has stable inverse as well. Then,<br />
K
Q(s)=(Yd(s)Nc(s)-Xd(s)Mc(s))(Ms(s)Mc(s)+Ns(s)Nc(s)) -1<br />
is well–posed and stable. Premultiplying equation (*) by<br />
[Md(s) –Xs(s);Nd(s) Ys(s)]<br />
it follows<br />
Nc(s)Mc(s) -1 =(Xs(s)+Md(s)Qd(s))(Ys(s)-Nd(s)Qd(s)) -1 .
Observation<br />
Taking Q(s)=0 we have the so-called central controller<br />
C0(s) =Xs(s)Ys(s) -1<br />
which coincides with the controller designed with the pole<br />
assignment technique.<br />
Taking r=d=0 one has:<br />
ξ&<br />
= Aξ<br />
+ Bu + L(<br />
Cξ<br />
+ Du − y)<br />
u = Fξ<br />
LEMMA 3<br />
−1<br />
−1<br />
−1<br />
[ ( I − Y ( s)<br />
N ( s)<br />
Q(<br />
s))<br />
] Y ( )<br />
−1<br />
C( s)<br />
= C0<br />
( s)<br />
+ Yd<br />
( s)<br />
Q(<br />
s)<br />
s d<br />
s s<br />
u y<br />
P<br />
Yd -1<br />
C0<br />
Nd<br />
Q<br />
Ys -1
Strong Stabilization<br />
The problem is that of finding, if possible, a stable controller<br />
which stabilizes the closed-loop system.<br />
Example<br />
s −1<br />
P(<br />
s)<br />
=<br />
( s − 2)(<br />
s + 2)<br />
is not stabilizable with a stable controller. Why?<br />
s − 2<br />
P(<br />
s)<br />
=<br />
( s −1)(<br />
s + 2)<br />
is stabilizable with a stable controller. Why?<br />
Stabilizazion of many plants<br />
Two-step stabilization<br />
C(s) P(s)
Interpolation<br />
Let consider a SISO control system with P(s)=N(s)/M(s) e<br />
C(s)=Nc(s)/Mc(s) (stable coprime stabilization). Then the closedloop<br />
system is asymptotically stable if and only if<br />
U(s)=N(s)Nc(s)+M(s)Mc(s) is an unity (stable with stable inverse).<br />
Since we want that C(s) be stable, we can choice Nc(s)=C(s) and<br />
Dc(s)=1. Then,<br />
U ( s)<br />
− M ( s)<br />
C(<br />
s)<br />
=<br />
N ( s)<br />
Of course, we must require that C(s) be stable. This fact depends<br />
on the role of the right zeros of P(s). Indeed, if N(b)=0, with<br />
Re(b)≥0 (b can be infinity), then the interpolation condition must<br />
hold true:<br />
U ( b)<br />
= M ( b)<br />
Consider the first example and take M(s)=(s-2)(s+2)/(s+1) 2 ,<br />
N(s)=(s-1)/(s+1) 2 . Then, it must be: U(1)=-0.75, U(∞)=1.<br />
Obviously this is impossible.<br />
Consider the second example and take M(s)=(s-1)/(s+2) e N(s)=(s-<br />
2)/(s+2) 2 . It must be U(1)=0.25, U(∞)=1, which is indeed possible.
THEOREM 6<br />
Parity interlacing property (PIP)<br />
• P(s) è strongly stabilizable if and only if the number of poles of<br />
P(s) between any pair of real right zeros of P(s) (including<br />
infinity) is an even number.<br />
• We have seen that the PIP is equivalent to the existence of the<br />
existence of a unity which interpolates the right zeros of P(s).<br />
Hence, if the PIP holds, the problem boils down to the<br />
computation of this interpolant.<br />
• In the MIMO case, Theorem 6 holds unchanged. However it is<br />
worth pointing out that the poles must be counted accordingly<br />
with their Mc Millan degree and the zeros to be considered are<br />
the so-called blocking zeros.
REFERENECES<br />
J.G.Truxal, Control Systems Synthesis, Mc-Graw Hill, NY, 1955.<br />
J.R. Ragazzini, G.E. Franklin, Sampled-data Control Systems,<br />
Mc-Graw Hill, NY, 1958.<br />
M. Vidyasagar, Control System Synthesis: a Factorization<br />
Approach, MIT Press, Cambridge, MA, 1985.<br />
J.C. Doyle, B.A. Franklin, A.R. Tannenbaum, Feedback Control<br />
Theory, Mc Millan, NY, 1992.<br />
D.C. Youla, J.J. Bongiorno, N.N. Lu, Single-loop feedback<br />
stabilization of linear multivariable dynamical plants”,<br />
Automatica, 10, p. 159-175, 1974.<br />
G. Zames, “Feedback and optimal sensitivity in model reference<br />
transformation, multiplicative seminorms, approximate inverses,<br />
IEEE Trans. AC-26, p. 301-320, 1981.<br />
V. Kucera, Discrete Linear Control: The Polynomial Equation<br />
Approach, Wiley, NY, 1979.
H2 CONTROL<br />
SUMMARY<br />
• Definition and characterization of the H2 norm<br />
• State-feedback control (LQ)<br />
• Parametrization, mixed problem, duality<br />
• Disturbance rejection, Estimation (Kalman filter), Partial<br />
Information<br />
• Conclusions and generalizations
The L2 space<br />
L2 space: The set of (rational) functions G(s) such that<br />
1<br />
2π<br />
∞<br />
∫<br />
−∞<br />
[ G(<br />
− jω)'G(<br />
jω<br />
] dω<br />
< ∞<br />
trace )<br />
(strictly proper – no poles on the imaginary axis!)<br />
With the inner product of two functions G(s) and F(s):<br />
∞<br />
∫<br />
−∞<br />
1<br />
G ( s),<br />
F(<br />
s)<br />
= trace [ G(<br />
− jω<br />
)' F(<br />
jω)<br />
] dω<br />
2π<br />
The space L2 is a pre-Hilbert space. Since it is also complete, it is indeed a<br />
Hilbert space. The norm, induced by the inner product, of a function G(s)<br />
is<br />
G(<br />
s)<br />
2<br />
⎡ ∞<br />
⎢ 1<br />
= ∫ ⎢ trace<br />
2π<br />
⎢<br />
⎣ −∞<br />
[ G(<br />
− jω<br />
)'G(<br />
jω)<br />
]<br />
dω<br />
⎤<br />
⎥<br />
⎥<br />
⎥<br />
⎦<br />
1/<br />
2
The subspaces H2 and H2 ^<br />
The subspace H2 is constituted by the functions of L2 which are analytic in<br />
the right half plane. (strictly proper – stable !)<br />
The subspace H2 ⊥ is constituted by the functions of L2 which are analytic<br />
in the left half plane. (stricty proper – untistable !)<br />
Note: A rational function in L2 is a strictly proper function without poles<br />
on the imaginary axis. A rational function in H2 is a strictly proper<br />
function without poles in the closed right half plane. A rational function in<br />
H2 ⊥ is a strictly proper function without poles in the closed left half plane.<br />
The functions in H2 are related to the square integrable functions of the<br />
real variable t in (0,∞]. The functions in H2 ⊥ are related to the square<br />
integrable functions of the real variable t in (-∞,0].<br />
A function in L2 can be written in an unique way as the sum of a function<br />
in H2 and a function in H2: G(s)=G1(s)+G2(s). Of course, G1(s) and G2(s)<br />
are orthogonal, i.e. =0, so that<br />
2<br />
2<br />
G s)<br />
=<br />
G<br />
2 1(<br />
s)<br />
+<br />
2<br />
( G ( s)<br />
2<br />
2<br />
2
Consider the system<br />
x& ( t)<br />
= Ax(<br />
t)<br />
+ Bw(<br />
t)<br />
z(<br />
t)<br />
= Cx(<br />
t)<br />
x(<br />
0)<br />
= 0<br />
System theoretic interpretation<br />
of the H2 norm<br />
And let G(s) be its transfer function. Also consider the quantity to be<br />
evaluated:<br />
J<br />
1<br />
=<br />
m<br />
∑∫<br />
i=<br />
1 0<br />
∞<br />
z<br />
( i)<br />
'(<br />
t)<br />
z<br />
( i)<br />
( t)<br />
dt<br />
where z (i) represents the output of the system forced by an impulse input at<br />
the i-th component of the input.<br />
J<br />
1<br />
1<br />
2π<br />
=<br />
∞<br />
m<br />
i=<br />
1 0<br />
m<br />
− ∞ i=<br />
1<br />
∞<br />
∑∫<br />
∫∑<br />
z<br />
( i)<br />
'(<br />
t)<br />
z<br />
( i)<br />
1<br />
( t)<br />
dt =<br />
2π<br />
− ∞ i=<br />
1<br />
Z<br />
1<br />
ei'G(<br />
− jω)'G(<br />
jω)<br />
eidω<br />
=<br />
2π<br />
∞<br />
m<br />
∫∑<br />
( i)<br />
( − jω)'Z<br />
∞<br />
∫<br />
−∞<br />
trace<br />
( i)<br />
( jω)<br />
dω<br />
=<br />
[ G(<br />
−jω)'G(<br />
jω)<br />
]<br />
= G(<br />
s)<br />
2<br />
2
Computation of the norm<br />
Lyapunov equations<br />
“Control-type Lyapunov equation”<br />
“Filter-type Lyapunov equation”<br />
G(<br />
s)<br />
A o o<br />
' P + P A + C'C<br />
= 0<br />
Pr r<br />
2<br />
2<br />
A'<br />
+ P A + BB'<br />
= 0<br />
=<br />
J<br />
1<br />
=<br />
m<br />
∑∫<br />
∞<br />
∞<br />
z<br />
( i)<br />
'(<br />
t)<br />
z<br />
( i)<br />
( t)<br />
dt =<br />
i=<br />
1 0<br />
0 i=<br />
1<br />
∞ ⎡<br />
m ⎤<br />
= trace ⎢e<br />
A't<br />
C'Ce<br />
At<br />
Beiei'<br />
B'⎥dt<br />
∫ ⎢ ∑ ⎥<br />
0 ⎢⎣<br />
i=<br />
1 ⎥⎦<br />
⎡ ∞<br />
⎤<br />
⎢ A't<br />
At ⎥<br />
= trace ⎢B'<br />
e C'<br />
Ce dt B⎥<br />
= trace ∫ ⎢<br />
⎣ 0<br />
⎥<br />
⎦<br />
⎡ ∞<br />
⎤<br />
⎢<br />
trace C e<br />
At<br />
BB'<br />
e<br />
A't<br />
⎥<br />
= ⎢<br />
dt C'⎥<br />
= trace ∫ ⎢<br />
⎣ 0<br />
⎥<br />
⎦<br />
m<br />
∫∑<br />
trace<br />
[ B'<br />
P B]<br />
[ CP C']<br />
r<br />
[ e 'B<br />
'e<br />
A't<br />
C'Ce<br />
At<br />
i<br />
Bei]<br />
0<br />
dt<br />
=
Other interpretations<br />
Assume now that w is a white noise with identity intensity and consider<br />
the quantity:<br />
J<br />
2<br />
= lim E ( z'(<br />
t)<br />
z(<br />
t))<br />
t→∞ It follows:<br />
J<br />
2<br />
t<br />
t<br />
⎡ ( )<br />
( ) ( )' '<br />
'(<br />
)<br />
lim trace E<br />
⎡<br />
Ce<br />
A t−τ<br />
Bw d w B e<br />
A t−<br />
C'd<br />
⎤⎤<br />
= ⎢<br />
⎥<br />
E =<br />
0<br />
0<br />
t ⎣<br />
⎢∫<br />
τ τ ∫ σ<br />
σ<br />
σ<br />
⎣<br />
⎥<br />
→∞<br />
⎦⎦<br />
⎡<br />
= lim trace ⎢<br />
t→∞<br />
⎣<br />
⎡<br />
= lim trace ⎢<br />
t→∞<br />
⎣<br />
t t<br />
∫∫<br />
0 0<br />
Finally, consider the quantity<br />
J<br />
3<br />
and let again w(.) be a white noise with identity intensity. It follows:<br />
Notice that J3=(||G(s)||2) 1/2 since<br />
t<br />
∫<br />
0<br />
Ce<br />
1<br />
= lim<br />
T →∞ T<br />
Ce<br />
A(<br />
t−τ)<br />
A(<br />
t−τ)<br />
E<br />
T<br />
∫<br />
0<br />
BE<br />
BB'e<br />
[ w(<br />
τ)<br />
w(<br />
σ )']<br />
A'(<br />
t−τ<br />
)<br />
z'(<br />
t)<br />
z(<br />
t)<br />
dt<br />
⎤<br />
C'dτ<br />
⎥<br />
⎦<br />
B'e<br />
=<br />
A'(<br />
t−σ<br />
)<br />
G(<br />
s)<br />
Aτ A'τ<br />
[ CP(<br />
t)<br />
C']<br />
, P(<br />
t)<br />
= e BB'<br />
e dτ<br />
E(<br />
z'(<br />
t)<br />
z(<br />
t))<br />
= trace<br />
∫<br />
t<br />
0<br />
⎤<br />
C'dτ<br />
dσ<br />
⎥<br />
⎦<br />
P&<br />
1<br />
( t)<br />
= AP ( t)<br />
+ P(<br />
t)<br />
A'+<br />
BB'<br />
, AY + YA'+<br />
BB'<br />
= 0,<br />
Y = lim ∫ P(<br />
τ ) dτ<br />
T →∞ T<br />
2<br />
2<br />
=<br />
T<br />
0
x&<br />
Optimal Linear Quadratic Control (LQ)<br />
versus Full-Information H2 control<br />
= Ax + Bu<br />
Assume that<br />
And notice that these conditions are equivalent to<br />
Find (if any) u(⋅) minimizing J<br />
The above matrix result is the well known Schur Lemma. The proof is as follows. If H≥0, then<br />
Vice-versa, assume that H≥0 and R>0, and suppose by contradiction that W is not positive<br />
semidefinite, then Wx=νx, for some ν<br />
0,<br />
W<br />
0<br />
= Q − SR<br />
[ ] 0<br />
1<br />
'≥<br />
0<br />
− S
Optimal Linear Quadratic Control (LQ)<br />
versus Full-Information H2 control<br />
Let C11 a factorization of W=C11′C11 and define<br />
C<br />
u<br />
1<br />
=<br />
R<br />
Then, it is easy to verify that<br />
∞<br />
J = ∫ ( x'Qx<br />
+ 2u'<br />
Sx + u'<br />
Ru<br />
) dt = ∫ z'zdt<br />
= z<br />
0<br />
Moreover, the free motion of the state can be considered as a state motion<br />
caused by an impulsive input. Hence, with w(t)=imp(t), let<br />
The (LQ) problem with stability is that of finding a controller<br />
fed by x and w and yielding u that minimizes the H2 norm of the transfer<br />
function from w to z.<br />
• LQS problem<br />
⎡ C<br />
⎢ 1/<br />
⎣R<br />
11<br />
= − 2<br />
1/<br />
2<br />
u,<br />
⎤<br />
,<br />
S'<br />
⎥<br />
⎦<br />
D<br />
z = C x + D<br />
1<br />
12<br />
x(<br />
0)<br />
⎡0⎤<br />
= ⎢<br />
I<br />
⎥<br />
⎣ ⎦<br />
1<br />
= 0<br />
12<br />
u<br />
x& = Ax + B u + B w<br />
2<br />
z = C x + D<br />
u<br />
12<br />
1<br />
u<br />
ξ = Fξ<br />
+ G x + G<br />
&<br />
1<br />
1<br />
21<br />
2<br />
w<br />
= Hξ<br />
+ E x + E w<br />
∞<br />
0<br />
2<br />
2
x&<br />
= Ax + B w + B u<br />
z = C x + D<br />
1<br />
⎡x<br />
⎤<br />
y = ⎢ ⎥<br />
⎣w⎦<br />
1<br />
12<br />
u<br />
Full information<br />
w z<br />
u x<br />
2<br />
Problem: Find the minimun value of ||Tzw||2 attainable by an<br />
admissible controller. Find an admissible controller minimizing<br />
||Tzw||2. Find a set of all controllers generating all ||Tzw||20<br />
(Ac,C1c) detectable<br />
Ac = A-B2(D12′D12) -1 D12′C1 stable invariant zeros of (A,B2,C1,D12)<br />
C1c = (I-D12 (D12′D12) -1 D12′)C1<br />
P<br />
K
Solution of the Full Information Problem<br />
Theorem 1<br />
There exists an admissible controller FI minimizing ||Tzw||2 . It is<br />
given by<br />
F<br />
2<br />
-1 ( 'D<br />
) ( B ' + D 'C<br />
)<br />
= −<br />
P<br />
D12 12 2 12 1<br />
where P is the positive semidefinite and stabilizing solution of the<br />
Riccati equation<br />
A'<br />
P+<br />
PA−(<br />
PB + C 'D<br />
)( D 'D<br />
)<br />
A<br />
= A−B<br />
( D<br />
2<br />
'D<br />
1<br />
')<br />
12<br />
( B 'P+<br />
D<br />
−1<br />
12<br />
'C<br />
) = A+<br />
B F<br />
−1<br />
cc 2 12 12 2 12 1<br />
2 2<br />
The minimum norm is:<br />
2 2<br />
Tzw α , α = P<br />
2<br />
cB1<br />
= trace(<br />
B<br />
2<br />
1'<br />
PB1<br />
) , Pc<br />
12<br />
( B 'P+<br />
D 'C<br />
) + C 'C<br />
= 0<br />
2<br />
12<br />
1<br />
1<br />
1<br />
stable<br />
= ( s)<br />
= ( C + D F )( sI − A−B<br />
F )<br />
The set is given by<br />
Q(s)<br />
u w<br />
where Q(s) is a stable strictly proper system, satisfying<br />
Q<br />
2 2 2<br />
( s)<br />
< γ −α<br />
2<br />
F2<br />
x<br />
1<br />
12<br />
2<br />
2<br />
2<br />
−1
Proof of Theorem 1<br />
The assumptions guarantee the existence of the stabilizing solution to the<br />
Riccati equation P. Let v=u-F2x so that u=F2x+v, where v is a new input.<br />
Hence, z(s)=Pc(s)B1w(s)+U(s)v, where U(s)=Pc(s)B2+D12. It follows that<br />
Tzw(s)=Pc(s)B1+U(s)Tvw(s). The problem is recast as to find a controller<br />
minimizing the norm from w to z of the following system<br />
where Pv is given by:<br />
x&<br />
= Ax + B w + B u<br />
v = −F<br />
x + u<br />
2<br />
⎡x<br />
⎤<br />
y = ⎢ ⎥<br />
⎣w⎦<br />
1<br />
2<br />
w v<br />
Pv<br />
u x<br />
K<br />
Notice that Tzw(s) is strictly proper iff Tvw(s) is such. Exploiting the<br />
Riccati equality it is simple to verify that U(s) ~ U(s)=D12′D12 and that<br />
U(s) ~ 2 2 2<br />
Pc(s) is antistable. Hence, ||Tzw(s)|| =||Pc(s)B1|| + ||Tvw(s)|| . Hence<br />
2<br />
2<br />
2<br />
the optimal control is v=0, i.e. u=F2x.<br />
2 2<br />
Finally, take a controller K(s) such that ||Tzw(s)|| < γ . From this controller<br />
2<br />
and the system it is possible to form the transfer function Q(s)=Tvw(s). Of<br />
2 2 2<br />
course, it is ||Q(s)|| < γ -α . It is enough to show that the controller<br />
2<br />
yielding u(s)=F2x(s)+v(s)=F2x(s)+Q(s)w(s) generates the same transfer<br />
function Tzw(s). This computation is left to the reader.<br />
w
x&<br />
= Ax + B w + B u<br />
z = C x + D<br />
1<br />
y = C<br />
2<br />
x + w<br />
Disturbance Feedforward<br />
1<br />
12<br />
u<br />
Same assumptions as FI+stability of A-B1C2<br />
C2=0 → direct compensation<br />
C2≠0 → indirect compensation<br />
A-B1C2 B1 B2 y u<br />
R(s) = 0 0 I<br />
I 0 0<br />
-C2 I 0<br />
2<br />
KDF<br />
R(s)<br />
KFI(s)<br />
The proof of the DF theorem consist in verifying that the transfer function from w to z achieved by<br />
the FI regulator for the system A,B1,B2,C1, D12 is the same as the transfer function from w to z<br />
obtained by using the regulator KDF shown in the figure.
x&<br />
= Ax + B w + ( B u)<br />
z = C x + u<br />
1<br />
1<br />
y = C2x<br />
+ D21w<br />
w y<br />
Output Estimation<br />
w z<br />
u y<br />
2<br />
P11(s)<br />
P21(s)<br />
P<br />
K(s)<br />
K(s)<br />
u =<br />
−zˆ<br />
Assumptions<br />
(A,C2) detectable<br />
D21D21′> 0 Af = A-B1 D21′(D21D21′) -1 C2<br />
B1f = B1(I-D21′ (D21D21′) -1 D21)<br />
(Af,Bf) stabilizable<br />
(A-B2C1) stable<br />
z
Solution of the<br />
Output Estimation problem<br />
Theorem 2<br />
There exists an admissible controller (filter) minimizing ||Tzw||2. It<br />
is given by<br />
ξ&<br />
= Aξ<br />
+ B<br />
u<br />
where<br />
L = −(<br />
PC + B<br />
2<br />
2<br />
2<br />
D'<br />
21<br />
)<br />
( ) -1<br />
D D<br />
21<br />
21<br />
and P is the positive semidefinite and stabilizing solution of the<br />
Riccati equation<br />
AP+<br />
PA'−(<br />
PC '+<br />
BD<br />
A<br />
= −C<br />
ξ<br />
1<br />
2<br />
= A−(<br />
PC '+<br />
B D<br />
2<br />
u + L<br />
1<br />
2<br />
21<br />
( C<br />
')(<br />
D<br />
')(<br />
D<br />
D<br />
D<br />
−1<br />
( C P+<br />
D B ')<br />
+ B B '=<br />
0<br />
= A+<br />
L C<br />
−1<br />
ff 2 1 21 21 21<br />
2 2<br />
2<br />
ξ<br />
21<br />
− y)<br />
21<br />
')<br />
')<br />
2<br />
21<br />
1<br />
1<br />
stable<br />
1
Solution of the<br />
Output Estimation problem (cont.)<br />
The optimal norm is<br />
2<br />
2<br />
Tzw = 2<br />
f<br />
2<br />
−1<br />
C1Pf<br />
( s)<br />
= trace(<br />
C1PC1'),<br />
P ( s)<br />
= ( sI−<br />
A−L2C2<br />
) ( B1<br />
+ L2D21)<br />
The set of all controllers generating all ||Tzw||2 < γ is given by<br />
M(s)=<br />
Aff -B2C1 L∞ -B2<br />
C1 0 I<br />
C2 I 0<br />
u y<br />
M<br />
Where Q(s) is proper, stable and such that ||Q(s)|| 2 2 <<br />
γ 2 −||C1Pf(s)|| 2 2<br />
Q
Proof of Theorem 2<br />
It is enough to observe that the “transpose system”<br />
λ&<br />
= A'λ<br />
+ C'<br />
μ<br />
ϑ<br />
= B 'λ<br />
+ D<br />
=<br />
B<br />
1<br />
2<br />
'λ<br />
+ ς<br />
1<br />
ς<br />
21<br />
+ C<br />
'ν<br />
2<br />
'ν<br />
has the same structure as the system defining the DF problem.<br />
Hence the solution is the “transpose” solution of the DF problem.<br />
It is worth noticing that the H2 solution does not depend on the<br />
particular linear combination of the state that one wants to<br />
estimate.
Given the system<br />
x& = Ax + ς + B<br />
1<br />
y = Cx + ς<br />
2<br />
22<br />
u<br />
Filtering<br />
where [z′1 z′2]′ are zero mean white Gaussian noises with intensity<br />
W<br />
⎡W11<br />
W12<br />
⎤<br />
= ⎢ ,<br />
12'<br />
⎥ W<br />
⎣W<br />
W22⎦<br />
22<br />
> 0<br />
find an estimate of the linear combination Sx of the state such as<br />
to minimize<br />
J<br />
= lim<br />
t →∞<br />
Letting<br />
C<br />
B<br />
1<br />
11<br />
= −S,<br />
B<br />
11<br />
y = W<br />
E<br />
'=<br />
W<br />
[ ( Sx(<br />
t)<br />
− u(<br />
t))'(<br />
Sx(<br />
t)<br />
− u(<br />
t))<br />
]<br />
C<br />
−1/<br />
2<br />
22<br />
11<br />
2<br />
y,<br />
= W<br />
−W<br />
12<br />
−1<br />
22<br />
W<br />
C,<br />
−1<br />
22<br />
W<br />
D<br />
z = C x + u,<br />
1<br />
12<br />
21<br />
',<br />
=<br />
⎡ς<br />
⎢<br />
⎣ς<br />
[ 0<br />
1<br />
2<br />
B<br />
1<br />
I],<br />
= [ B<br />
⎤ ⎡B<br />
⎥ = ⎢<br />
⎦ ⎣ 0<br />
11<br />
11<br />
W<br />
12<br />
W<br />
W<br />
−1/<br />
2<br />
22<br />
]<br />
−1/<br />
2<br />
12W22<br />
1 / 2<br />
W22<br />
⎤<br />
⎥w<br />
⎦<br />
the problem is recast to find a controller that minimizes the H2<br />
norm from w to z.
z = C x + D<br />
y = C x + D<br />
The partial information problem<br />
(LQG)<br />
x&<br />
= Ax + B w + B u<br />
1<br />
2<br />
12<br />
21<br />
u<br />
w<br />
Assumptions: FI + DF.<br />
1<br />
w z<br />
u y<br />
2<br />
P<br />
K(s)
Theorem 3<br />
There exists an admissible controller minimizing ||Tzw||2. It is<br />
given by<br />
ξ&<br />
= Aξ<br />
+ B u + L ( C ξ − y)<br />
u = −F<br />
ξ<br />
The optimal norm is<br />
2<br />
2<br />
2<br />
2<br />
Tzw P 2 c s)<br />
L2<br />
+ C 2 1Pf<br />
( s)<br />
= P ( s)<br />
B<br />
2<br />
c 1 + F 2 2Pf<br />
= ( ( s)<br />
= γ<br />
The set of all controllers generating all ||Tzw||2 < γ is given by<br />
S(s)=<br />
2<br />
2<br />
2<br />
A +B2F2+L2C2 L2 -B2<br />
-F2 0 I<br />
C2 I 0<br />
2<br />
2<br />
2<br />
2<br />
0<br />
u y<br />
S<br />
Where Q(s) is proper, stable and such that ||Q(s)|| 2 2 < γ 2 −γο 2<br />
Proof<br />
Separation principle: The problem is reformulated as an Output Estimation problem for the estimate<br />
of -F2 x.<br />
Q
Important topics to discuss<br />
• Robustness of the LQ regulator (FI)<br />
• Robustness of the Kalman filter (OE)<br />
• Loss of robustness of the LQG regulator (Partial Information)<br />
• LTR technique
.<br />
THE H¥<br />
CONTROL PROBLEM<br />
SUMMARY<br />
• Definition and characterization of the H∞ norm<br />
• Description of the uncertainty: standard problem<br />
• State-feedback control<br />
• Parametrization, mixed problem, duality<br />
• Disturbance rejection, Estimation, Partial Information<br />
• Conclusions and generalizations
.<br />
G ( s)<br />
20<br />
10<br />
0<br />
-10<br />
-20<br />
50<br />
0<br />
-50<br />
-100<br />
-150<br />
10 0<br />
What the H¥ norm is?<br />
rational<br />
stable<br />
proper<br />
G (s)<br />
= G(<br />
s)<br />
=<br />
∞<br />
sup Re( s ) ≥ 0<br />
Bode Diagrams<br />
From: U(1)<br />
Frequency (rad/sec)<br />
max G(<br />
jω)<br />
Ω = σ ( Ω)<br />
= λmax(<br />
Ω * Ω)<br />
= λmax(<br />
ΩΩ*)<br />
For the frequency-domain definition we can consider a transfer function without poles on the<br />
imaginary axis. It is easy to understand that the norm of a function in L∞ can be recast to the norm<br />
of its stable part given through the so-called inner-outer factorization. The time-domain definition,<br />
given in the next page, requires, for obvious reasons, the stability.<br />
10 1<br />
ω<br />
10 2
G(s) = D+C(sI-A) -1 B<br />
x&<br />
= Ax + Bw<br />
.<br />
Time-domain characterization<br />
z = Cx + Dw<br />
A = asymptotically stable<br />
G(s)<br />
∞<br />
=<br />
sup<br />
w∈L<br />
There does not exist a direct procedure to compute the infinity<br />
norm. However, it is easy to establish whether the norm is<br />
bounded from above by a given positive number γ.<br />
The symbol indifferently the space of square integrable or the space of the strictly proper rational<br />
function without poles on the imaginary axis. If w(.) is a white noise, the infinity norm represents<br />
the square root of the maximal intensity of the output spectrum.<br />
2<br />
z<br />
w<br />
2<br />
2
.<br />
BOUNDED REAL LEMMA<br />
Let γ be a positive number and let A be asymptotically stable.<br />
THEOREM 1<br />
The three following conditions are equivalent each other:<br />
(i) ||G(s)||∞ < γ<br />
(ii) ||D||
Comments<br />
Notice that as γ tends to infinity, the Riccati equation becomes a<br />
Lyapunov equation that admits a (unique) positive semidefinite<br />
solution, thanks to the system stability. This is obvious if one<br />
notices that the infinity norm of a stable system is finite.<br />
Proof of Theorem 1<br />
For simplicity, let consider the case where the system is strictly proper, i.e. D=0. The equation and<br />
the inequality are<br />
Also denote: G(s) ~ =G(-s)′.<br />
.<br />
PBB'P<br />
A'<br />
P + PA + 2 + C'C<br />
= ( < ) 0<br />
γ<br />
Points (ii)→(i) and (iii) →(i).<br />
Assume that there exists a positive semidefinite (definite) solution P of the equation (inequality).<br />
We can write:<br />
PBB'<br />
P<br />
( sI + A')<br />
P − P(<br />
sI − A)<br />
+ + 2 C'C<br />
= ( < ) 0<br />
γ<br />
Premultiply to the left by B′(sI+A′) -1 e to the right by (sI-A) -1 B, it follows<br />
G<br />
so that ||G(s)||∞ < γ.<br />
~<br />
( s)<br />
G(<br />
s)<br />
= γ<br />
T ( s)<br />
= γI<br />
−γ<br />
−1<br />
2<br />
I − T<br />
~<br />
( s)<br />
T(<br />
s)<br />
B'<br />
P(<br />
sI − A)<br />
Points (i)→(ii)<br />
Assume that ||G(s)||∞ < γ. We now proof that the Hamiltonian matrix<br />
⎡ A<br />
H = ⎢<br />
⎣−<br />
C'C<br />
−1<br />
−2<br />
γ BB'⎤<br />
⎥<br />
− A'<br />
⎦<br />
does not have eigenvalues on the imaginary axis. Indeed, if, by contradiction jω is one such<br />
eigenvalue, then
( jω<br />
− A)<br />
x − γ<br />
( jω<br />
+ A')<br />
y + C'Cx<br />
= 0<br />
.<br />
−2<br />
BB'<br />
y = 0<br />
Hence Cx = -γ -2 C(jωI-A) -1 BB′(jωI+A) -1 C′Cx, so that G(-jω)′G(jω)Cx=0 and Cx=0. Consequently,<br />
y=0 and x=0, that is a contradiction. Then, since the Hamiltonian matrix does not have imaginary<br />
eigenvalues, it must have 2n eigenvalues, n of them having negative real parts. The remaining on<br />
eigenvalues are the complex conjugate of the previous ones. This fact follows from the matrix being<br />
Hamiltonian, i.e. satisfying JH+H′J=0, where J=[0 I;-I 0]. Let take the n-dimensional subspace<br />
generated by the (generalized) eigenvectors associated with the stable eigenvalues and let choice a<br />
matrix S=[X* Y*]* whose range coincides with such a subspace. For a certain asymptotically stable<br />
matrix T (restriction of H) it follows HS=ST. We now proof that X*Y=Y*X. Indeed let define<br />
V=X*Y-Y*X=S*JS and notice that VT=S*JST=S*JHS=-S*H′JS=-T*S*JS=-T*V, so that<br />
VT+T*V=0. The stability of T yields (by the well known Lyapunov Lemma) that the unique<br />
solution is V=0, i.e. X*Y=Y*X. We now proof that X is invertible. Indeed from HS=ST it follows<br />
that AX+γ -2 BB′Y=XT and –A′Y-C′CX=YT. Premultiplying the first equation by Y* yields<br />
Y*AX+γ -2 Y*BB′Y =Y*XT=X*YT. Hence if, by contradiction, v∈Ker(X) then v*Y*BB′Yx=0 so<br />
that B′Yv=0. From the first equation we have XTv=0 and -A′Yv=YTv. In conclusion, if v∈Ker(X)<br />
then Tv∈Ker(X). By induction it follows T k v∈Ker(X), and again<br />
-A′YT k v =YT k+1 v, k nonnegative. Take the monic polynomial of minimum degree h(T) such that<br />
h(T)v=0 (notice that it always exists) and write h(T)=(λI-T)m(T). Since T è asymptotically stable, it<br />
results Re(λ)
.<br />
Worst Case<br />
The “worst case” interpretation of the H∞ norm is given by the<br />
following result:<br />
THEOREM 2<br />
Let A be stable, ||G(s)||∞ < γ and let x0 be the initial state of the<br />
system. Then,<br />
where P is the solution of the BRL Riccati equation.<br />
Proof of Theorem 2<br />
2 2 2<br />
sup w ∈L<br />
z −γ<br />
w = x<br />
2<br />
2 0'<br />
Px<br />
2 0<br />
Consider the function V(x)=x′Px and its derivative along the trajectories of the system. Letting<br />
Δ=(γ 2 I-D′D) -1 we have:<br />
V&<br />
= x'(<br />
A'<br />
P + PA+<br />
PBw + B'PB)<br />
x<br />
= −x'C'<br />
Cx − x'<br />
( PB + C'D)<br />
Δ(<br />
B'P<br />
+ D'<br />
C)<br />
x = −z'<br />
z<br />
+ w'(<br />
D'C<br />
+ B'<br />
P)<br />
x + x'(<br />
C'D<br />
+ PB)<br />
w+<br />
w'D'<br />
Dx +<br />
− x'<br />
( PB + C'<br />
D)<br />
Δ(<br />
B'P<br />
+ D'C)<br />
x =<br />
2<br />
−1<br />
= −z'<br />
z + γ w'<br />
w − ( w − w )'Δ<br />
( w−<br />
w )<br />
ws<br />
where wws = Δ(B’P+D’C)x is the worst disturbance. Recalling that the system is asymptotically<br />
stable and taking the integral of both hands, the conclusion follows.<br />
ws
.<br />
Observation: LMI<br />
Schur Lemma<br />
0<br />
'<br />
'<br />
'<br />
'<br />
'<br />
'<br />
2<br />
><br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡<br />
−<br />
+<br />
+<br />
−<br />
−<br />
−<br />
D<br />
D<br />
I<br />
C<br />
D<br />
P<br />
B<br />
D<br />
C<br />
PB<br />
C<br />
C<br />
PA<br />
P<br />
A<br />
γ<br />
( ) 0<br />
'<br />
)<br />
'<br />
'<br />
(<br />
'<br />
)<br />
'<br />
(<br />
'<br />
0<br />
1<br />
2<br />
<<br />
+<br />
+<br />
−<br />
+<br />
+<br />
+<br />
><br />
−<br />
C<br />
C<br />
DC<br />
P<br />
B<br />
D<br />
D<br />
I<br />
D<br />
C<br />
PB<br />
PA<br />
P<br />
A<br />
P<br />
γ
.<br />
x& = ( A + LΔN)<br />
x ,<br />
⎧ x&<br />
⎪<br />
⎨ z<br />
⎪<br />
⎩w<br />
=<br />
=<br />
=<br />
Complex Stability Radius and<br />
quadratic stability<br />
Ax + Lw<br />
Nx<br />
Δx<br />
Δ<br />
≤ α<br />
The system is said to be quadratically stable if there exists a<br />
solution (Lyapunov function) to the following inequality, for<br />
every Δ in the set<br />
THEOREM 3<br />
The system is quadratically stable if and only if ||N(sI-A) -1 L||∞
Proof of Theorem 3<br />
First observe that the following inequality holds:<br />
( A + LΔN)'P<br />
+ P(<br />
A + LΔN)<br />
=<br />
.<br />
A'<br />
P + PA + N'<br />
Δ'ΔN<br />
+ PLL'<br />
P − ( N'<br />
Δ'−PL)(<br />
ΔN<br />
− L'<br />
P)<br />
≤ A'P<br />
+ PA + PLL'<br />
P + N'<br />
Nα<br />
2<br />
= α<br />
2<br />
( A'<br />
X<br />
XLL'<br />
X<br />
+ XA + −2<br />
α<br />
where Pα 2 =X. Hence, if there exists X>0 satisfying<br />
+ N'<br />
N)<br />
then A+LΔN is asymptotically stable for every Δ, ||Δ|| ≤α, with the<br />
same Lyapunov function (quadratic stability). This happens if<br />
||N(sI-A) -1 L||∞0, we have that there exists<br />
s=jω that violates (*), a contradiction.
Real stability radius<br />
A= stable (eigenvalues with strictly negative real part)<br />
r ( A,<br />
B,<br />
C)<br />
= inf<br />
r<br />
Linear algebra problem: compute<br />
μr<br />
Proposition<br />
= inf<br />
s∈<br />
jω<br />
= inf<br />
s∈<br />
jω<br />
m×<br />
p<br />
{ Δ : Δ ∈ R and A + BΔC<br />
is unstable}<br />
m×<br />
p<br />
inf{<br />
Δ : Δ ∈ R and det(<br />
sI − A − BΔC)<br />
= 0}<br />
m×<br />
p<br />
inf{<br />
Δ : Δ ∈ R<br />
−1<br />
and det(<br />
I − ΔC(<br />
sI − A)<br />
B)<br />
= 0}<br />
[ ] 1 −<br />
inf<br />
m×<br />
p<br />
{ Δ : Δ ∈ R and det(<br />
I − Δ ) = 0}<br />
( M ) = M<br />
μ<br />
⎛ ⎡ Re M − γ Im M ⎤⎞<br />
⎜<br />
⎟<br />
⎜ ⎢<br />
⎥⎟<br />
⎝ ⎣ Im M Re ⎦⎠<br />
r ( M ) = inf σ 2<br />
γ∈(<br />
0,<br />
1]<br />
−1<br />
γ<br />
M<br />
Taking M=G(s)=C(sI-A) -1 B it follows<br />
⎛ ⎡ Re G(<br />
jω)<br />
− γ Im G(<br />
jω)<br />
⎤ ⎞<br />
sup inf σ ⎜<br />
⎟<br />
2<br />
∈ ⎜ ⎢<br />
⎥ ⎟<br />
ω γ ( 0,<br />
1]<br />
⎝ ⎣γ<br />
Im G(<br />
jω)<br />
Re G(<br />
jω)<br />
⎦ ⎠<br />
−1<br />
rr ( A,<br />
B,<br />
C)<br />
= −1
Example<br />
0 5 10 15 20 25 30<br />
0<br />
0.5<br />
1<br />
1.5<br />
2<br />
2.5<br />
3<br />
0 5 10 15 20 25 30<br />
0<br />
0.2<br />
0.4<br />
0.6<br />
0.8<br />
1<br />
1.2<br />
1.4<br />
1.6<br />
1.8<br />
2<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡<br />
=<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎣<br />
⎡<br />
=<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎣<br />
⎡<br />
−<br />
−<br />
−<br />
−<br />
−<br />
−<br />
−<br />
−<br />
=<br />
4175<br />
.<br />
0<br />
3834<br />
.<br />
0<br />
6711<br />
.<br />
0<br />
0535<br />
.<br />
0<br />
0668<br />
.<br />
0<br />
0077<br />
.<br />
0<br />
5297<br />
.<br />
0<br />
0346<br />
.<br />
0<br />
8310<br />
.<br />
0<br />
6793<br />
.<br />
0<br />
5194<br />
.<br />
0<br />
6789<br />
.<br />
0<br />
3835<br />
.<br />
0<br />
0470<br />
.<br />
0<br />
9347<br />
.<br />
0<br />
2190<br />
.<br />
0<br />
,<br />
11<br />
5<br />
.<br />
14<br />
9<br />
5<br />
.<br />
33<br />
38<br />
60<br />
40<br />
167<br />
13<br />
17<br />
12<br />
41<br />
20<br />
30<br />
20<br />
79<br />
C<br />
B<br />
A<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡−<br />
=<br />
Δ<br />
4996<br />
.<br />
0<br />
1214<br />
.<br />
0<br />
1214<br />
.<br />
0<br />
4996<br />
.<br />
0<br />
worst
.<br />
Entropy<br />
Consider a stable system G(s) with state space realization<br />
(A,B,C,0), and assume that ||G(s)||∞
Comments<br />
It is easy to check that the γ-entropy measure is not a norm, but<br />
can be considered as a generalization of the square of the H2 norm<br />
Indeed the γ-entropy can be written as<br />
Graph of f(x 2 ) parametrized in γ>1. For large γ the function f(x 2 )<br />
gets closer to x 2 (red line).<br />
.<br />
I<br />
G(<br />
s)<br />
γ<br />
m<br />
1<br />
( G)<br />
= f ( σ<br />
2π<br />
∫∑<br />
f ( x<br />
2<br />
2<br />
2<br />
) = −γ<br />
f<br />
1<br />
=<br />
2π<br />
5<br />
4.5<br />
4<br />
3.5<br />
3<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
− ∞ i=<br />
1<br />
2<br />
+∞<br />
ln<br />
+∞<br />
( 1<br />
m<br />
∫∑<br />
− ∞ i=<br />
1<br />
σ<br />
x<br />
−<br />
γ<br />
2<br />
i<br />
2<br />
i<br />
2<br />
2<br />
[ G(<br />
jω)<br />
] dω<br />
[ G(<br />
jω)<br />
]<br />
)<br />
) dω<br />
0<br />
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1<br />
x
Comments<br />
An interpretation for the γ-entropy for SISO systems is as follows.<br />
Consider the feedback configuration:<br />
.<br />
w z<br />
G(s)<br />
Δ(s)<br />
where w(.) is a white noise and Δ(s) is a random transfer function<br />
with Δ(jω1) independent on Δ(jω2) and uniformly distributed on<br />
the disk of radius γ -1 in the complex plane.<br />
Hence the expectation value over the random feedback transfer<br />
function is<br />
2<br />
⎛ G(<br />
s)<br />
⎞<br />
⎜<br />
⎟<br />
E = Iγ<br />
( G)<br />
Δ<br />
⎜ 1 G(<br />
s)<br />
( s)<br />
⎟<br />
⎝<br />
− Δ 2 ⎠
Proof of the Proposition<br />
First notice that f(x 2 ) ≥ x 2 so that the conclusion that Iγ(G) ≥||G(s)|| 2 2 follows immediately.<br />
Now, let<br />
Of course it is ri≥β>1. Then<br />
so that<br />
which is the conclusion. In order to prove that the entropy can be computed from the stabilizing<br />
solution of the Riccati equation, recall (Theorem 1) that, since ||G(s)||∞
.<br />
z = Cx + Du<br />
ACTUATOR DISTURBANCE<br />
x&<br />
= Ax + B(<br />
w + u)<br />
The optimal H2 state-feedback controller is<br />
u<br />
= F x,<br />
2<br />
F<br />
2<br />
= −(<br />
D'D)<br />
Proposition<br />
The optimal H2 control law ensures an H∞ norm of the closed loop<br />
system (from w to z) lower than 2||D||.<br />
Proof. Consider the equation of the Bounded Real Lemma for the closed-loop system, i.e.<br />
−2<br />
( A + BF2<br />
)'Q<br />
+ Q(<br />
A + BF2<br />
) + γ QBB'Q<br />
+ ( C + DF2<br />
)'(<br />
C + DF2)<br />
< 0<br />
Now take Q=P/α and consider both equations in P. It follows:<br />
( 1<br />
This condition is verified by choosing α≤1 and the minimum γ such that R≤0, i.e. the minimum γ<br />
such that<br />
α-α 2<br />
The term α-α 2 is maximized by α=0.5, for which α-α 2 2 −2<br />
2 −2<br />
2<br />
( α − α ) I − γ D'<br />
D ≥ α −α<br />
− γ σ ( D)<br />
= 0<br />
=0.25, so that the minimum γ is<br />
−1<br />
0 = A'<br />
P + PA + C'C<br />
− F D'<br />
DF<br />
2<br />
( B'<br />
P + D'C)<br />
−1<br />
−1<br />
−1<br />
−1<br />
−2<br />
( I − D(<br />
D'D)<br />
D')<br />
C + PBRB'<br />
P < 0,<br />
R = ( D'<br />
D)<br />
( 1−<br />
α ) + α<br />
−1<br />
−α ) C'<br />
I γ<br />
γ =<br />
2σ( D)<br />
2
The Small Gain Theorem<br />
THEOREM 4<br />
Assume that G1(s) is stable. Then:<br />
(i) The interconnected system is stable for each stable G2(s) with<br />
||G2(s)||∞α −1 then there exists a stable G2(s) with<br />
||G2(s)||∞
(iii) Proof of Theorem 4<br />
Point (i).<br />
If ||G2(s)||∞0, and write the<br />
singular value decomposition of G1(jω), i.e.<br />
G1(jω)=U(jω)Σ(jω)V ~ (jω) where Σ(jω)=[S(jω)′ 0]′ and S(jω) is<br />
square with dimension m. Moreover, take a stable G2(s) such that<br />
G2(jω)=ρV(jω)[I 0]U ~ (jω). Notice that G2 ~ (jω)G2(jω)=ρ 2
The H¥<br />
design problem<br />
du dc<br />
c 0 y u up K(s) G(s)<br />
c<br />
Find K(s) in such a way to guarantee:<br />
• Stability<br />
• Satisfactory performances<br />
Design in nominal conditions<br />
Uncertainties description<br />
Design under uncertaint conditions<br />
dr
G(s) = Gn(s)<br />
Nominal Design<br />
The performances are expressed in terms of requirements on some<br />
transfer functions<br />
du dc<br />
c 0 y K(s) u up G(s)<br />
c<br />
Characteristic functions:<br />
• Sensitivity Sn(s)=(I+Gn(s)K(s)) -1<br />
dc → c, c 0 →y, -dr→y<br />
• Complementary sensitivity<br />
Tn(s)=Gn(s)K(s)(I+Gn(s)K(s)) -1<br />
c 0 →c, -dr→c<br />
• Control sensitivity<br />
Vn(s)=K(s)(I+Gn(s)K(s)) -1<br />
c 0 →up, -dr→up, -dc→up<br />
dr
Shaping Functions<br />
In the SISO case, the requirement to have a “small” transfer<br />
function φ(s) can be well expressed by saying that the absolute<br />
value |φ(jω)|, for each frequency ω, must be smaller than a<br />
given function θ(ω) (generally depending on the frequency):<br />
φ( jω)<br />
< θ ( ω),<br />
∀ω<br />
Analogously, in the MIMO case, one can write<br />
This can be done by choosing a matrix transfer function W(s),<br />
stable with stable inverse (“shaping” function) , such that<br />
∀ω<br />
σ<br />
[ φ ( jω)<br />
] < θ ( ω),<br />
ω<br />
σ ∀<br />
−1<br />
[ W ( jω)<br />
] = θ ( ω)<br />
← W ( s)<br />
φ ( s)<br />
< 1<br />
A general requirement is that the sensitivity function is small at<br />
low frequency (tracking) whereas the complementary sensitivity<br />
function is small at high frequiency (feedback disturbances<br />
attenuation). Mixed sensitivity:<br />
⎡W1<br />
( s)<br />
Sn<br />
( s)<br />
⎤<br />
⎢<br />
2(<br />
) ( )<br />
⎥<br />
⎣W<br />
s Tn<br />
s ⎦<br />
∞<br />
< 1<br />
∞
Description of the uncertainties<br />
The nominal transfer function Gn(s) of the process G(s) belongs<br />
to a set G which can be parametrized through a transfer<br />
function Δ(s) included in the set<br />
D<br />
α<br />
For instance,<br />
{ Δ(<br />
s)<br />
Δ(<br />
s)<br />
∈ H , Δ(<br />
s)<br />
< α }<br />
= ∞<br />
∞<br />
• G = {G(s) | G(s) = Gn(s)+Δ(s)}<br />
• G = {G(s) | G(s) = Gn(s)(I+Δ(s))}<br />
• G = {G(s) | G(s) = (I+Δ(s))Gn(s)}<br />
• G = {G(s) | G(s) = (I-Δ(s)) -1 Gn(s)}<br />
• G = {G(s) | G(s) = (I-Gn(s)Δ(s)) -1 Gn(s)}
Examples<br />
• G = {G(s) | G(s) = Gn(s)+Δ(s)}<br />
Uncertaint unstable zeros<br />
s − 2 ε<br />
G(<br />
s)<br />
=<br />
+<br />
( s + 2)(<br />
s + 1)<br />
( s + 2)(<br />
s + 1)<br />
• G = {G(s) | G(s) = Gn(s)(I+Δ(s))}<br />
Unmodelled high frequency poles or unstable zeros<br />
G(<br />
s)<br />
1 ⎛ εs<br />
⎞<br />
= ⎜1−<br />
⎟,<br />
( s + 1)<br />
⎝ ( 1+<br />
εs)<br />
⎠<br />
• G = {G(s) | G(s) = (I-Δ(s)) -1 Gn(s)}<br />
Unmodelled unstable poles<br />
⎛ 10 ⎞<br />
G(<br />
s)<br />
= ⎜1−<br />
⎟<br />
⎝ ( 1+<br />
s)<br />
⎠<br />
• G = {G(s) | G(s) = (I-Gn(s)Δ(s)) -1 Gn(s)}<br />
Uncertain unstable poles<br />
−1<br />
⎛ 1 ⎞<br />
G( s)<br />
= ⎜1−<br />
ε<br />
⎟<br />
⎝ s −1<br />
⎠<br />
G(<br />
s)<br />
1<br />
( s + 10)<br />
−1<br />
1<br />
( s −1)<br />
1 ⎛ 2 ⎞<br />
= ⎜1−<br />
⎟<br />
( s + 1)<br />
⎝ ( 1+<br />
s)<br />
⎠
Design in nominal conditions<br />
Example of mixed performances<br />
z1<br />
W 1<br />
du dc<br />
w = c 0 y K(s) u up Gn(s) c z2<br />
W2 w z<br />
P<br />
u y<br />
K<br />
⎡ z1<br />
⎤<br />
z =<br />
⎢ , w = c<br />
z<br />
⎥<br />
⎣ 2 ⎦<br />
W5=1<br />
0<br />
⎡W<br />
=<br />
⎢<br />
⎢<br />
0<br />
⎢⎣<br />
I<br />
1<br />
dr<br />
−W<br />
G<br />
P 2<br />
1<br />
W G<br />
− G<br />
n<br />
n<br />
n<br />
⎤<br />
⎥<br />
⎥<br />
⎥⎦
Design under uncertain condition<br />
Robust stability and nominal performances<br />
z1<br />
W 1<br />
z2<br />
K(s) Gn(s) dc<br />
c 0 y u c<br />
w z<br />
P<br />
u y<br />
K<br />
⎡ z1⎤<br />
z = ⎢ w =<br />
z<br />
⎥,<br />
⎣ 2⎦<br />
W4<br />
dc<br />
Δ<br />
W5<br />
⎡−W<br />
=<br />
⎢<br />
⎢<br />
0<br />
⎢⎣<br />
− I<br />
dc=w<br />
dr<br />
−W<br />
G<br />
P 4<br />
1<br />
1<br />
W G<br />
− G<br />
n<br />
n<br />
n<br />
⎤<br />
⎥<br />
⎥<br />
⎥⎦
Design under uncertain condition<br />
Robust stability and robust performances<br />
z1<br />
W 1<br />
z2<br />
K(s) Gn(s) dc<br />
c 0 y u c<br />
1) Robust sensitivity performance<br />
2) Robust stability<br />
dc=w<br />
Result: (take W5=1)<br />
1) and 2) are both achieved if the controller is such that<br />
W4<br />
Δ<br />
W5<br />
−1<br />
[ I −W<br />
( s)<br />
Δ(<br />
s)<br />
W ( s)<br />
V ( s)<br />
] , ∀ Δ(<br />
s)<br />
< 1<br />
W1 ( s)<br />
S(<br />
s)<br />
= W1(<br />
s)<br />
Sn(<br />
s)<br />
5<br />
4 n<br />
W4 n<br />
∞<br />
( s)<br />
V ( s)<br />
W5(<br />
s)<br />
∞<br />
< 1<br />
⎡W1<br />
( s)<br />
Sn<br />
( s)<br />
⎤<br />
⎢<br />
4(<br />
) ( )<br />
⎥<br />
⎣W<br />
s Vn<br />
s ⎦<br />
∞<br />
< 1<br />
dr<br />
∞
The proof of the result follows from 1) and 2) and on the following<br />
Lemma<br />
Let X(s) and Y(s) in L∞ and Ψ a generic L∞ function such that ||Y∞||
The Standard Problem<br />
w z<br />
u y<br />
All the situations studied so far can be recast in the so-called<br />
standard problem: Find K(s) in such a way that:<br />
• The closed-loop system is asymptotically stable<br />
•<br />
T(<br />
z,<br />
w,<br />
s)<br />
∞<br />
< γ<br />
P<br />
K<br />
Existence of a feasible K(s)and parametrization of all controllers<br />
such that the closed-loop norm between w and z be less than g.
x&<br />
= Ax+<br />
B w + B u<br />
z = C x + D<br />
1<br />
⎡x<br />
⎤<br />
y = ⎢ ⎥<br />
⎣w⎦<br />
1<br />
12<br />
u<br />
Ac = A-B2(D12′D12) -1 D12′C1<br />
C1c = (I-D12 (D12′D12) -1 D12′)C1<br />
Full information<br />
w z<br />
u x<br />
2<br />
P<br />
K<br />
w<br />
Assumptions<br />
(A,B2) stabilizzabile<br />
D12′D12>0<br />
(Ac,C1c) rivelabile
Solution of the Full Information Problem<br />
THEOREM 5<br />
There exists a controller FI, feasible and such that ||T(z,w,s)||∞ < γ<br />
if and only if there exists a positive semidefinite and stabilizing<br />
solution of the Riccati equation<br />
−1<br />
B1B<br />
1'<br />
A'<br />
P+<br />
PA−(<br />
PB2<br />
+ C1'D12)(<br />
D12'D12)<br />
( B2'<br />
P+<br />
D12'C1)<br />
+ P P+<br />
C1'C<br />
2<br />
1 = 0<br />
γ<br />
−1<br />
B1B<br />
1'<br />
A−B2<br />
( D12'D12')<br />
( B2'P+<br />
D12'C1)<br />
+ P stable<br />
2<br />
γ<br />
F<br />
∞<br />
=<br />
−<br />
F∞<br />
-γ<br />
u w<br />
Q(s)<br />
-2 B1′P<br />
-1 ( 'D<br />
) ( B 'P<br />
+ D 'C<br />
)<br />
D12 12 2 12 1<br />
Q(s) is a stable system, proper and satisfying ||Q(s)||∞ < γ.<br />
x
Comments<br />
• As (γ→∞) the central controller (Q(s)=0) coincides with the H2<br />
optimal controller<br />
• The parametrized family directly includes the only static<br />
controller u=F∞x.<br />
• The proof of the previous result, in its general fashion, would<br />
require too much time. Indeed, the necessary part requires a<br />
digression on the Hankel-Toeplitz operator. If we drop off the<br />
result on the parametrization and limit ourselves to the static<br />
state-feedback problem, i.e. u=Kx, the following simple result<br />
holds:<br />
THEOREM 6<br />
There exists a stabilizing control law u=Kx such that the norm of<br />
the system (A+B2K,B1,C1+D12K) is less than γ if and only if there<br />
exists a positive definite solution of the inequality:<br />
A'<br />
P + PA − ( PB<br />
2<br />
+ C'<br />
D<br />
B1B1<br />
'<br />
+ P P + C1'<br />
C1<br />
< 0<br />
2<br />
γ<br />
12<br />
)( D<br />
12<br />
'D<br />
12<br />
)<br />
−1<br />
( B<br />
2<br />
' P + D<br />
12<br />
'C<br />
1<br />
)
Observation: LMI<br />
P > 0<br />
A'P<br />
+ PA − ( PB<br />
Ac = A-B2(D12′D12) -1 D12′C1<br />
+ C'D<br />
B1B1'<br />
+ P P + C1'C<br />
1 < 0<br />
2<br />
γ<br />
X<br />
> 0<br />
C1c = (I-D12 (D12′D12) -1 D12′)C1<br />
2<br />
12<br />
⎡ − XAc<br />
'−Ac<br />
X − B1B1'+<br />
γ<br />
⎢<br />
⎣ C1c<br />
X<br />
)( D<br />
2<br />
12<br />
'D<br />
12<br />
)<br />
−1<br />
( B<br />
2<br />
Schur Lemma<br />
X= γ 2 P -1<br />
B<br />
2<br />
B<br />
2<br />
'<br />
'P<br />
+ D<br />
XC '⎤<br />
1c<br />
> 0<br />
2 ⎥<br />
γ I ⎦<br />
12<br />
'C<br />
1<br />
)
Proof of Theorem 6<br />
Assume that there exists K such that A+B2K is stable and the<br />
norm of the system (A+B2K,B1,C1+D12K) is less than γ. Then,<br />
from the we know that there exists a positive definite solution of<br />
the inequality<br />
( A+ B K)'P<br />
+ P(<br />
A + B K)<br />
+ γ<br />
+ ( C<br />
1<br />
2<br />
+ D<br />
12<br />
Now, defining<br />
F<br />
K)'(<br />
C<br />
1<br />
+<br />
D<br />
12<br />
2<br />
K)<br />
< 0<br />
−2<br />
-1 ( 'D<br />
) ( B ' + D 'C<br />
)<br />
= −<br />
P<br />
D12 12 2 12 1<br />
PB B 'P<br />
+<br />
1<br />
1<br />
(*)<br />
the Riccati inequality can be equivalently rewritten as<br />
A'<br />
P + PA − ( PB<br />
P<br />
B<br />
1 1<br />
2<br />
γ<br />
B<br />
'<br />
1<br />
2<br />
1<br />
+ C'<br />
D<br />
12<br />
)( D<br />
' D<br />
) +<br />
so concluding the proof. Vice-versa, assume that there exists a<br />
positive definite solution of the inequality. Then, inequality (*)<br />
is satisfied with K=F, so that with such a K, the norm of the<br />
closed-loop transfer function is less than γ (BRL).<br />
12<br />
12<br />
)<br />
−1<br />
P + C 'C<br />
+ ( K − F)'(<br />
K − F)<br />
< 0<br />
( B ' P + D<br />
2<br />
12<br />
'C<br />
1
Define:<br />
A<br />
C<br />
W<br />
R<br />
R<br />
Parametrization of all algebraic<br />
“state-feedback” controllers<br />
THEOREM 7<br />
The set of all controllers u=Kx such that the H∞ norm of the<br />
closed-loop system is less than γ is given by:<br />
K<br />
W<br />
1<br />
=<br />
=<br />
=<br />
[ A B2<br />
]<br />
[ C1<br />
D12<br />
]<br />
[ W W ]<br />
= W<br />
> 0<br />
2<br />
1<br />
'W<br />
−1<br />
1<br />
⎡− ARW<br />
−WAR'−<br />
B1B<br />
⎢<br />
⎣ CRW<br />
'<br />
2<br />
1<br />
'<br />
WCR<br />
'⎤<br />
2 ⎥ > 0<br />
γ I ⎦
Proof of Theorem 7<br />
We prove the theorem in the simple case in which C1′D12=0 and<br />
D12′D12=I. The LMI is equivalent to<br />
K = W 'W<br />
W<br />
1<br />
AW<br />
> 0<br />
1<br />
2<br />
1<br />
−1<br />
1<br />
+ W A'+<br />
B B '+<br />
W B<br />
1<br />
1<br />
2<br />
2<br />
'+<br />
B W<br />
2<br />
2<br />
'+<br />
γ<br />
If there exists W satisfying such an inequality, it follows that the<br />
inequality<br />
Is satisfied with K=W2′W1 -1 and P=γ 2 W1 -1 . The result follows from<br />
the BRL. Vice-versa, assume that there exists P>0 satisfying this<br />
last inequality. The conclusion directly follows by letting<br />
W1==γ 2 P -1 e and W2=W1K′.<br />
−2<br />
WC<br />
'CW<br />
+ γ<br />
1<br />
1<br />
1<br />
1<br />
−2<br />
W W '<<br />
0<br />
−2<br />
P( A + B2<br />
K)<br />
+ ( A + B2K)'P<br />
+ γ<br />
PB1B<br />
1'P<br />
+ C1'C1<br />
+ KK'<<br />
0<br />
2<br />
2
Example<br />
⎡ 0<br />
x&<br />
= ⎢<br />
⎣−1<br />
1 ⎤ ⎡0⎤<br />
⎥x<br />
+ ⎢ ⎥(<br />
w + u)<br />
−1⎦<br />
⎣1⎦<br />
⎡1<br />
z = ⎢<br />
⎣0<br />
0⎤<br />
⎡0⎤<br />
⎥x<br />
+ ⎢ ⎥u<br />
0⎦<br />
⎣1⎦<br />
γinf = 0.75, γ = 0.76<br />
||T(z1,w)||∞<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
robustness<br />
⎡0<br />
ΔA<br />
= ⎢<br />
⎣Ω<br />
0<br />
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1<br />
H2<br />
H<br />
Ω<br />
0⎤<br />
0<br />
⎥<br />
⎦
x&<br />
=<br />
z<br />
ξ<br />
=<br />
Ax + B w + B<br />
C<br />
1<br />
x +<br />
1<br />
D<br />
12<br />
= Lx + Mu<br />
u<br />
Mixed Problem H2/H¥<br />
2<br />
u<br />
The problem consists in finding u=Kx in such a way that<br />
min T ( ξ, w,<br />
s,<br />
K)<br />
: T ( z,<br />
w,<br />
s,<br />
K)<br />
≤ γ<br />
2<br />
K<br />
• For instance, take the case z=ξ. The problem makes sense since<br />
a blind minimization of the H∞ infinity norm may bring to a<br />
serious deterioration of the H2 (mean squares) performances.<br />
• Obviously, the problem is non trivial only if the value of γ is<br />
included in the interval (γinf γ2), where γing is the infimun<br />
achievable H∞ norm for T(z,w,s,K) as a function of K, whereas<br />
γ2 is the H∞ norm of T(z,w,s,KOTT) where KOTT is the optimal<br />
unconstrained H2 controller.<br />
• This problem has been tackled in different ways, but an analytic<br />
solution is not available yet. In the literature, many sub-optimal<br />
solutions can be found (Nash-game approach, convex<br />
optimization, inverse problem, …)<br />
∞
x&<br />
=<br />
z<br />
ξ<br />
=<br />
Ax + B w + B<br />
C<br />
1<br />
x +<br />
D<br />
= Lx + Mu<br />
Post-Optimization procedure<br />
1<br />
12<br />
u<br />
2<br />
u<br />
simplifying assumptions L′M=0, M′M=I<br />
Let Ksub a controller such that ||T(z,w,s,Ksub)||∞
THEOREM 8<br />
Post-Opt Algorithm<br />
Each controller Kα of the family (α-con) is a stabilizing controller<br />
and is such that<br />
T ( ξ,<br />
w,<br />
s,<br />
K<br />
( 1<br />
d<br />
−α<br />
)<br />
dα<br />
α<br />
)<br />
2<br />
= T ( ξ,<br />
w,<br />
s,<br />
K2−<br />
T ( ξ,<br />
w,<br />
s,<br />
K<br />
α<br />
)<br />
2<br />
≤ 0<br />
Interpretation: For α=0, the controller is K0=Ksub. Hence,<br />
||T(z,w,s,K0)||∞
Proof of Theorem 8<br />
First notice that the equation (α-RIC) canm be rewritten as:<br />
( 2 α α α 2<br />
A + B K )'P<br />
+ P ( A + B K ) + K 'K<br />
+ L'<br />
L = 0 ( α − RIC)<br />
so that<br />
α<br />
||T(ξ,w,s,Kα)||2 =(Trace(B1′PαB1)) 1/2<br />
The fact that the norm ||T(ξ,w,s,Kα)||2 is symmetric with respect to<br />
α=1 follows directly (α-RIC) by inspection. Taking the derivative<br />
with respect to α one obtains<br />
F<br />
α<br />
' Xα<br />
+ XαFα<br />
− 2(<br />
1−α<br />
) Λα'Λα<br />
= 0<br />
where Xα is the derivate of Pα with respect to α and<br />
Fα=Aα-B2αB2α′Pα, Λα=Ksub+PαB2<br />
Matrix Fα is stable since Pα is the stabilizing solution of (α-RIC).<br />
Hence, for the Lyapunov Lemma the conclusion follows.<br />
α<br />
α
Example<br />
⎡ 0<br />
x&<br />
= ⎢<br />
⎣−1<br />
1 ⎤ ⎡0⎤<br />
⎥x<br />
+ ⎢ ⎥(<br />
w + u)<br />
−1⎦<br />
⎣1⎦<br />
⎡1<br />
z = ⎢<br />
⎣0<br />
0⎤<br />
⎡0⎤<br />
⎥x<br />
+ ⎢ ⎥u<br />
0⎦<br />
⎣1⎦<br />
γinf = 0.75, γ2 = 0.8415, γ=0.8<br />
Ksub=Kcen=[-0.6019 -0.7676]<br />
robustness<br />
a*=0.56, Ka*=[-0.4981 -0.5424]<br />
Performances with respect to W<br />
1.1<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.64<br />
0.63<br />
0.62<br />
0.61<br />
0.6<br />
0.59<br />
0.58<br />
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2<br />
Ka<br />
Kcen<br />
K2<br />
0.4<br />
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1<br />
1<br />
0.95<br />
0.9<br />
0.85<br />
0.8<br />
⎡0<br />
ΔA<br />
= ⎢<br />
⎣Ω<br />
0⎤<br />
0<br />
⎥<br />
⎦<br />
0.75<br />
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2<br />
3<br />
2.5<br />
2<br />
1.5<br />
1<br />
K2<br />
Kcen<br />
0.5<br />
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1<br />
Ka
x&<br />
= Ax + B w + B u<br />
z = C x + D<br />
1<br />
y = C<br />
2<br />
x + w<br />
Disturbance Feedforward<br />
1<br />
12<br />
u<br />
Same assumptions as FI+stability of A-B1C2<br />
C2=0 → direct compensation<br />
C2≠0 → indirect compensation<br />
A-B1C2 B1 B2 y u<br />
R(s) = 0 0 I<br />
I 0 0<br />
-C2 I 0<br />
2<br />
KDF<br />
R(s)<br />
KFI(s)<br />
The proof of the DF theorem consist in verifying that the transfer function from w to z achieved by<br />
the FI regulator for the system A,B1,B2,C1, D12 is the same as the transfer function from w to z<br />
obtained by using the regulator KDF shown in the figure.
x&<br />
= Ax + B w + ( B u)<br />
z = C x + u<br />
1<br />
1<br />
y = C2x<br />
+ D21w<br />
w y<br />
Output Estimation<br />
w z<br />
u y<br />
2<br />
P11(s)<br />
P21(s)<br />
P<br />
K(s)<br />
K(s)<br />
ξ<br />
Assumptions<br />
(A,C2) detectable<br />
D21D21′> 0 Af = A-B1 D21′(D21D21′) -1 C2<br />
B1f = B1(I-D21′ (D21D21′) -1 D21)<br />
(Af,B1f) stabilizable<br />
(A-B2C1) stable<br />
u<br />
ξˆ = −<br />
z
Solution of the<br />
Output Estimation problem<br />
THEOREM 9<br />
There exists a feasible controllor (Filter) such that ||T(z,w,s)||∞ < γ<br />
if and only if there exists the stabilizing positive semidefinite<br />
solution of the Riccati equation<br />
M(s)=<br />
A<br />
C<br />
I<br />
L<br />
AΠ<br />
+ ΠA′<br />
− ( ΠC<br />
A − ( ΠC<br />
2 f<br />
f<br />
ff<br />
∞<br />
=<br />
=<br />
=<br />
Aff L∞ -B2 γ -2 ΠC1′<br />
C1 0 I<br />
C2f If 0<br />
( )<br />
( )<br />
( )<br />
( ) 1 -<br />
-1<br />
A − ΠC2'<br />
D21D21'<br />
C2<br />
-1<br />
D21D21'<br />
C2<br />
-1<br />
D21D21'<br />
−<br />
−(<br />
ΠC<br />
'+<br />
B D ')<br />
D D '<br />
=<br />
2<br />
'+<br />
B D<br />
1<br />
2<br />
21<br />
2<br />
'+<br />
B<br />
')(<br />
D<br />
1<br />
1<br />
21<br />
D<br />
21<br />
21<br />
D<br />
')(<br />
D<br />
21<br />
')<br />
−1<br />
21<br />
C1'C<br />
1<br />
C2<br />
Π 2<br />
γ<br />
21<br />
B<br />
2<br />
C<br />
1<br />
stable<br />
u y<br />
M<br />
Q(s) is a proper stable system satisfying ||Q(s)||∞ < γ.<br />
21<br />
D<br />
21<br />
')<br />
−1<br />
( C Π +<br />
2<br />
D<br />
21<br />
C1'C<br />
1<br />
B1<br />
) + Π Π + B1<br />
B<br />
2<br />
1'<br />
= 0<br />
γ<br />
Q
Proof of Theorem 9<br />
It is enough to observe that the “transpose system”<br />
λ&<br />
= A'λ<br />
+ C'<br />
μ<br />
ϑ<br />
= B 'λ<br />
+ D<br />
=<br />
B<br />
1<br />
2<br />
'λ<br />
+ ς<br />
1<br />
ς<br />
21<br />
+ C<br />
'ν<br />
2<br />
'ν<br />
Has the same structure as the system defining the DF problem.<br />
Hence the solution is the “transpose” solution of the DF problem.<br />
It is worth noticing that the solution depends on the particular<br />
linear combination of the state that one wants to estimate.
Example<br />
⎡ 0<br />
x&<br />
= ⎢<br />
⎣−1<br />
y =<br />
1 ⎤ ⎡0⎤<br />
x + w<br />
−1<br />
+ Ω<br />
⎥ ⎢<br />
1<br />
⎥<br />
⎦ ⎣ ⎦<br />
[ 1 0]<br />
x + w2<br />
1<br />
damping ξ = ( 1−<br />
Ω)<br />
/ 2<br />
Let consider six filters:<br />
Filter K1: Kalman filter. The noises are assumed to be white<br />
uncorrelated gaussian noises with identity intensities, namely<br />
W1=1, W2=1.<br />
Filter K2: Kalman filter. The noises are assumed to be white<br />
uncorrelated gaussian noises with W1=1, W2=0.5.<br />
Filter K3: Kalman filter. The noises are assumed to be white<br />
uncorrelated gaussian noises with W1=0.5, W2=1.<br />
Filters H1,H2,H3: H∞ filters with γ=1.1, γ=1.01, γ=1.005 (notice<br />
that with γ=1 the stabilizing solution of the Riccati equation does<br />
not exist).<br />
Other techniques of Robust filtering are possible<br />
Norm H2<br />
2.2<br />
2<br />
1.8<br />
1.6<br />
1.4<br />
1.2<br />
1<br />
0.8<br />
H3<br />
0.6<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9<br />
H2<br />
H1<br />
Norm H∞<br />
7<br />
6<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Assumptions: FI + DF.<br />
Partial Information<br />
w z<br />
u y<br />
Theorem 10<br />
There exists a feasible such that ||T(z,w,s)||∞ < γ if and only if<br />
• There exists the stabilizing positive semidefinite solution of<br />
−1<br />
B1B<br />
1'<br />
A'<br />
P+<br />
PA−(<br />
PB2<br />
+ C1'D12)(<br />
D12'<br />
D12)<br />
( B2'<br />
P + D12'C1<br />
) + P 2 P+<br />
C1'C1<br />
= 0<br />
γ<br />
• There exists the stabilizing positive semidefinite solution of<br />
•<br />
x&<br />
= Ax + B w + B u<br />
z = C x + D<br />
y = C x + D<br />
−1<br />
C1'C1<br />
AΠ+<br />
ΠA'−(<br />
ΠC2<br />
'+<br />
B1D21')(<br />
D21D21')<br />
( C2Π+<br />
D21B1')<br />
+ Π Π+<br />
B1B<br />
1'=<br />
0<br />
2<br />
γ<br />
max λ ( PΠ)<br />
< γ<br />
i<br />
1<br />
2<br />
i<br />
1<br />
12<br />
21<br />
u<br />
w<br />
2<br />
2<br />
P<br />
K(s)
Structure of the regulator<br />
u y<br />
S∞<br />
Q<br />
Afin B1fin B2fin<br />
S∞(s) = F∞ 0 (D21D21′) -1/2<br />
C2fin (D21D21′) -1/2 0<br />
Afin = A−B2(D12′D12) -1 D12C1−B2(D12′D12) -1 B2′P+γ -2 B1B1′P+<br />
+ (I−γ -2 ΠP) -1 L∞(C2+γ -2 D21B1′P)<br />
B1fin = −(I−γ -2 ΠP) -1 L∞<br />
B2fin = (I−γ -2 ΠP) -1 (B2+γ -2 ΠC1′D12)(D12′D12) -1/2<br />
C2fin = −(D21D21′) -1 (C2+γ -2 D21B1′P)
Comments<br />
It is easy to check that the central controller (Q(s)=0) is described<br />
by:<br />
ξ&<br />
= A ξ + B u + Z L C ξ − y + D w*)<br />
+ B w*<br />
u<br />
=<br />
F<br />
∞<br />
ξ<br />
2<br />
∞<br />
∞ ( 2<br />
21 1<br />
This closely resembles the structure of the optimal H2 controller.<br />
Notice, however, that the well-known separation principle does<br />
not hold for the presence of the worst disturbance w*=B1′Pξ.<br />
Proof of Theorem 10 (sketch)<br />
Define the variables r e q as:<br />
w = r + γ -2 B1′Px<br />
q = u − F∞x<br />
In this way, the system becomes<br />
x&<br />
= ( A + γ<br />
q<br />
y<br />
= −F<br />
= ( C<br />
∞<br />
2<br />
−2<br />
x + u<br />
+ γ<br />
B B 'P)<br />
x + B r + B<br />
−2<br />
1<br />
D<br />
1<br />
21<br />
B ')<br />
x + D<br />
1<br />
1<br />
21<br />
w<br />
2<br />
u<br />
This system has the same structure as the one of the OE problem.<br />
Hence, the solution can be obtained from the solution of the OE<br />
problem, by recognizing that the solution Πt of the relevant<br />
Riccati equation is related to the solutions P e Π above in the<br />
following way:<br />
Π = Π(<br />
I −γ<br />
P<br />
t<br />
−2 Π<br />
)<br />
−1
⎡P<br />
P = ⎢<br />
⎣P<br />
11<br />
21<br />
P<br />
P<br />
12<br />
22<br />
⎤<br />
⎥<br />
⎦<br />
T ( z,<br />
w,<br />
s)<br />
= P<br />
11<br />
T ( z,<br />
w,<br />
s)<br />
= T ( s)<br />
−T<br />
1<br />
Operatorial Approach<br />
w z<br />
P<br />
u y<br />
( s)<br />
+ P<br />
2<br />
12<br />
( s)<br />
K<br />
( I − K(<br />
s)<br />
P ( s)<br />
)<br />
( s)<br />
Q(<br />
s)<br />
T<br />
( s)<br />
K(<br />
s)<br />
P<br />
( s)<br />
The functions T1, T2, T3 depend on a double coprime factorization<br />
of P22. The function Q(s) is the Youla-Kucera parameter (stable<br />
and proper function).<br />
• Inner-outer factorization and reduction of the problem to a<br />
distance Nehari problem<br />
3<br />
22<br />
−1<br />
21
Laurent and Hankel operators<br />
• Let F(s) ∈ L∞ . The map ΛF: L2 → L2 defined as<br />
ΛF : G(s) → ΛFG(s) = F(s)G(s)<br />
is called the Laurent operator (with symbol F).<br />
• Consider G(s)∈L2 and let G(s) =Gs(s)+Ga(s), with Gs(s) ∈H2 and Ga(s)<br />
∈ H2 ⊥ . The map Π s: L2 → H2 defined as<br />
Πs : G(s) → Π sG(s) = Gs(s)<br />
is called the stable projection.<br />
• Consider G(s)∈L2 and let G(s) =Gs(s)+Ga(s), with Gs(s) ∈H2 and Ga(s)<br />
∈ H2 ⊥ . The map Π a: L2 → H2 ⊥ defined as<br />
Π a : G(s) → Π aG(s) = Ga(s)<br />
is called the untistable projection.<br />
• Let F(s)∈L∞. The map ΓF: H2 ⊥ → H2 defined as<br />
ΓF : G(s) → ΓFG(s) = Π sΛFG(s)<br />
is called the Hankel operator (with symbol F).<br />
Notice that ΛF is linear and ||ΛF||=||F(s)||∞ so that is a bounded operator. It is immediate to check that<br />
the projections Πs and Πa are indeed linear operators. The Hankel operator is the result of the<br />
composition of the Laurent operator and the stable projection, i.e. ΓF=ΠsΛF . Notice also that only<br />
the stable part of F(s) contributes to ΓFG(s). Indeed, since G(s) since G(s)∈Η2 ⊥ it follows<br />
ΓFG(s)=ΠsΛFG(s)= ΠsF(s)G(s)= Πs[(Fs(s)+Fa(s))G(s)]= ΠsFs(s)G(s)=ΓFsG(s)
The adjoint operators<br />
• The adjoint Laurent operator (with symbol F) is the adjoint operator<br />
with symbol F ~ , i.e.<br />
Λ*F = ΛF ~<br />
• The adjoint Hankel operator (with symbol F) is<br />
Γ*F = Π aΛ ∗ F<br />
Laurent: G1(s)∈ L2 and G2(s)∈ L2 . Then<br />
G<br />
1<br />
G , F<br />
1<br />
,<br />
*<br />
1<br />
Λ FG<br />
2 = Λ FG1,<br />
G 2 =<br />
2 ∫<br />
+∞<br />
π −∞<br />
~<br />
G<br />
2<br />
=<br />
G , Λ<br />
1<br />
F<br />
~<br />
G<br />
2<br />
trace<br />
[ G ( − jω)'<br />
F(<br />
− jω)'G<br />
( jω)<br />
]<br />
Hankel: H(s) ∈ H2 and G(s)∈ H2 ⊥ . Recall that =0 if a∈H2 and b∈<br />
H2 ⊥ . Then<br />
G, Γ<br />
G, F<br />
*<br />
F<br />
~<br />
H<br />
H<br />
=<br />
=<br />
Γ G, H<br />
F<br />
G, Π<br />
a<br />
F<br />
~<br />
=<br />
H<br />
Π FG, H<br />
=<br />
s<br />
G, Π Λ<br />
a<br />
=<br />
F<br />
~<br />
H<br />
1<br />
FG, H<br />
=<br />
2<br />
dω<br />
=
Time-domain interpretation of the Hankel<br />
operator<br />
Let F(s)=C(sI-A) -1 B, with A stable, (A,C) observable and (A,B) reachable.<br />
Hence, the inverse Laplace transform of F(s) is<br />
⎧ 0<br />
( t)<br />
= ⎨<br />
⎩Ce<br />
f At<br />
t ≤ 0<br />
t > 0<br />
Now, consider the map Γf : L2(−∞,0) → L2(0, +∞)<br />
Defined by<br />
⎧ 0<br />
⎪ 0<br />
Γf<br />
: u( t)<br />
→ Γf<br />
u(<br />
t)<br />
= y(<br />
t)<br />
= ⎨ At −Aτ<br />
Ce ( )<br />
⎪ ∫ e Bu τ dτ<br />
⎩ −∞<br />
B<br />
t ≤ 0<br />
t > 0<br />
This operator maps past input to future outputs. Indeed, the past inputs (with x(−∞)=0) contribute<br />
to form the state x(0) from which the free output can be calculated. Γf is the time-domain<br />
counterpart of ΓF. Γf can be seen as the composition of two operators Ψo (the observability<br />
operator) and Ψr (the reachability operator)<br />
Γ<br />
f<br />
Ψ<br />
Ψ<br />
r<br />
o<br />
= Ψ<br />
r<br />
Ψ<br />
o<br />
,<br />
Ψ<br />
: u(<br />
t)<br />
→ Ψ u(<br />
t)<br />
: = x =<br />
r<br />
r<br />
n<br />
: L ( −∞,<br />
0)<br />
→ R ,<br />
⎧ 0<br />
: x → Ψox<br />
: = y(<br />
t)<br />
= ⎨ At<br />
⎩Ce<br />
x<br />
2<br />
t < 0<br />
t ≥ 0<br />
→ L<br />
The operator Ψo is injective thanks to the observability of (A,C). Indeed, from Ψo x=0 it follows<br />
x=0.<br />
Conversely, the operator Ψr is surjective thanks to the reachability of (A,B). Indeed, given any point<br />
x of R n it follows that Ψru(t)=x, where u(t)=0, for t>0 and u(t)=B′e -A′t Pr -1 for t≤0, where Pr is the<br />
unique solution of the Lyapunov equation A′Pr+PrA+BB′=0<br />
0<br />
∫ ∞<br />
−<br />
e<br />
− Aτ<br />
Ψ<br />
o<br />
Bu(<br />
τ)<br />
dτ<br />
: R<br />
n<br />
2<br />
( 0,<br />
∞)
Time-domain interpretation of the adjoint<br />
Hankel operator<br />
The adjoint reachability and observability operators are as follows:<br />
Ψ<br />
Ψ<br />
*<br />
r<br />
*<br />
o<br />
− ⎧<br />
* B'e<br />
: x → Ψr<br />
x = u(<br />
t)<br />
= ⎨<br />
⎩ 0<br />
: y(<br />
t)<br />
→ Ψ<br />
*<br />
o<br />
y(<br />
t)<br />
= x =<br />
C'<br />
y(<br />
τ ) dτ<br />
Therefore the adjoint Hankel operator in time-domain is<br />
Γ<br />
Γ<br />
*<br />
f<br />
*<br />
f<br />
: L<br />
2<br />
[ 0,<br />
∞)<br />
→ L ( −∞,<br />
0]<br />
: y(<br />
t)<br />
→ Γ<br />
*<br />
f<br />
2<br />
The rank of Ψr and Ψo is n = McMillan degree of C(sI-A) -1 B. Moreover,<br />
since Ψo is injective and Ψr surjective, it turns out that Γf (and ΓF) has rank<br />
n. Therefore the self adjoint operatopr Γ * FΓF has rank n, so that it admits<br />
eigenvalues.<br />
∞<br />
∫<br />
0<br />
e<br />
⎧<br />
⎪B'e<br />
y(<br />
t)<br />
= u(<br />
t)<br />
= ⎨<br />
⎪<br />
⎩<br />
A't<br />
x<br />
−A'τ<br />
∞<br />
A't<br />
∫<br />
0<br />
e<br />
t ≤ 0<br />
t > 0<br />
A'τ<br />
0<br />
C'<br />
y(<br />
τ)<br />
dτ<br />
t ≤ 0<br />
t > 0
Reachability and observability grammians<br />
Notice that Pr and Po are the unique (positive definite) solutions of the<br />
Lyapunov equations:<br />
APr+PrA′+BB′=0<br />
PoA+A′Po+C′C=0<br />
The operator Γ * FΓF and the matrix PrPo share the same nonzero<br />
eigenvalues. To see this notice that Γ*fΓf = Ψ * rΨ * oΨoΨr=PrPo<br />
The Hankel norm of the system is<br />
______________<br />
Example:<br />
Consider the stable part and take a minimal realization:<br />
It results<br />
Γ<br />
Ψ<br />
Ψ<br />
F<br />
r<br />
*<br />
o<br />
Ψ<br />
Ψ<br />
=<br />
*<br />
r<br />
o<br />
Γ<br />
=<br />
=<br />
f<br />
∞<br />
∫<br />
0<br />
∞<br />
∫<br />
0<br />
e<br />
e<br />
=<br />
Aτ<br />
A'τ<br />
BB'e<br />
C'Ce<br />
λ<br />
A'τ<br />
Aτ<br />
dτ<br />
= P<br />
r<br />
dτ<br />
= P<br />
( P P )<br />
max r o<br />
5(<br />
s −1)(<br />
s + 5)(<br />
s − 3)<br />
22.<br />
5 2.<br />
5(<br />
3s<br />
− 5)<br />
F(<br />
s)<br />
=<br />
= 5 + +<br />
2<br />
2<br />
( s + 1)<br />
( s − 7)<br />
s − 7 ( s + 1)<br />
⎡ 0 1 ⎤ ⎡0⎤<br />
A = ⎢ , = , =<br />
1 2<br />
⎥ B ⎢<br />
1<br />
⎥ C<br />
⎣−<br />
− ⎦ ⎣ ⎦<br />
[ −12.<br />
5 7.<br />
5]<br />
⎡0.<br />
25 0 ⎤ ⎡303.<br />
125 78.<br />
125⎤<br />
Pr<br />
= ⎢ ,<br />
,<br />
0 0.<br />
25<br />
⎥ Pr<br />
= ⎢<br />
78.<br />
125 53.<br />
125<br />
⎥ λ<br />
⎣ ⎦ ⎣<br />
⎦<br />
max<br />
(P P ) = 81.3827, Γ<br />
r<br />
o<br />
o<br />
F<br />
=<br />
9.<br />
0212
Formulae for the minimization problem via<br />
the operatorial approach<br />
All stable functions<br />
)<br />
(<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
21<br />
3<br />
12<br />
2<br />
21<br />
12<br />
11<br />
1<br />
s<br />
P<br />
s<br />
M<br />
s<br />
T<br />
s<br />
M<br />
s<br />
P<br />
s<br />
T<br />
s<br />
P<br />
s<br />
Y<br />
s<br />
M<br />
s<br />
P<br />
s<br />
P<br />
s<br />
T<br />
=<br />
=<br />
+<br />
=<br />
I<br />
s<br />
X<br />
s<br />
N<br />
s<br />
Y<br />
s<br />
M<br />
s<br />
M<br />
s<br />
N<br />
s<br />
Y<br />
s<br />
X<br />
s<br />
N<br />
s<br />
M<br />
s<br />
M<br />
s<br />
N<br />
s<br />
P<br />
=<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡<br />
−<br />
−<br />
=<br />
=<br />
−<br />
−<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
)<br />
(<br />
)<br />
(<br />
1<br />
1<br />
22<br />
[ ] [ ]<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
)<br />
(<br />
ˆ<br />
)<br />
(<br />
1<br />
s<br />
M<br />
s<br />
Q<br />
s<br />
Y<br />
s<br />
N<br />
s<br />
Q<br />
s<br />
X<br />
s<br />
K −<br />
−<br />
=<br />
−<br />
K<br />
P
Scalar case<br />
It is possible to perform an inner-outer factorization of T4=T2T3=T4iT4o so<br />
that, letting F=T4i -1 T1 and X=T4oQ we have:<br />
T ( s)<br />
− T ( s)<br />
Q(<br />
s)<br />
T ( s)<br />
T<br />
1<br />
4i<br />
The problem is then reduced to a Nehari problem.<br />
Now, let<br />
F<br />
( s)<br />
Theorem 11<br />
The function<br />
is such that<br />
−1<br />
2<br />
T ( s)<br />
− T<br />
1<br />
F(<br />
s)<br />
= F<br />
a<br />
PQβ<br />
= λ<br />
max<br />
40<br />
( s)<br />
= C(<br />
sI − A)<br />
A'<br />
P + PA = BB',<br />
a<br />
3<br />
( PQ)<br />
β,<br />
∞<br />
=<br />
( s)<br />
Q(<br />
s)<br />
( s)<br />
+ F ( s)<br />
+ D,<br />
−1<br />
f ( s)<br />
= B'(<br />
sI − A')<br />
X<br />
0<br />
λ<br />
( s)<br />
= F ( s)<br />
−<br />
s<br />
−1<br />
max<br />
B<br />
β,<br />
∞<br />
T ( s)<br />
− T ( s)<br />
Q(<br />
s)<br />
1<br />
=<br />
F<br />
AQ + QA'<br />
= C'C<br />
χ = [ λ<br />
a<br />
4<br />
F ( s)<br />
− X ( s)<br />
max<br />
( s)<br />
∈ H<br />
( PQ)]<br />
⊥<br />
2<br />
−1<br />
Pβ<br />
g(<br />
s)<br />
= C(<br />
sI − A)<br />
( PQ)<br />
g(<br />
s)<br />
f ( s)<br />
0<br />
F( s)<br />
X ( s)<br />
= infx<br />
( s ) ∈H<br />
− ∞<br />
F(<br />
s)<br />
− X ( s)<br />
,<br />
−1<br />
∞<br />
∞<br />
χ<br />
=<br />
F ( s)<br />
∈ H<br />
s<br />
2
Proof of Theorem 11<br />
From F(s) construct f(s), g(s) and λmax(PQ). Let be X o (s) be the optimal<br />
solution and define h(s)=(F(s)-X o (s))f(s). Notice that h(s) ∈L2, since F(s)-<br />
X o (s) ∈L∞ and f(s) ∈ H2. It follows:<br />
h(<br />
s)<br />
− Γ<br />
〈 h(<br />
s),<br />
h(<br />
s)<br />
〉 + 〈 Γ<br />
Now, taking into account that Γ * F ~ f(s) ∈H2 ⊥ and X o f(s)∈H2, it follows<br />
〈 h(<br />
s),<br />
Γ<br />
〈 Π<br />
a<br />
So that<br />
*<br />
~<br />
F<br />
*<br />
~<br />
F<br />
f ( s)<br />
2<br />
2<br />
*<br />
~<br />
F<br />
f ( s)<br />
〉<br />
F(<br />
s)<br />
f ( s),<br />
Γ<br />
h(<br />
s)<br />
− Γ<br />
*<br />
~<br />
( F(<br />
s)<br />
− X<br />
= 〈 h(<br />
s)<br />
− Γ<br />
f ( s),<br />
Γ<br />
= 〈 Π<br />
*<br />
~<br />
2<br />
⎡ o<br />
( F(<br />
s)<br />
− X ( s))<br />
−<br />
∞<br />
⎢⎣<br />
F<br />
F<br />
f ( s)<br />
o<br />
2<br />
2<br />
*<br />
~<br />
F<br />
*<br />
~<br />
F<br />
f ( s)<br />
〉 = 〈 Γ<br />
f ( s),<br />
h(<br />
s)<br />
− Γ<br />
f ( s)<br />
〉 − 2〈<br />
h(<br />
s),<br />
Γ<br />
[ F(<br />
s)<br />
− X<br />
a<br />
( s))<br />
f ( s)<br />
2<br />
2<br />
− λ<br />
max<br />
max<br />
*<br />
~<br />
F<br />
o<br />
( PQ)<br />
⎤<br />
⎥⎦<br />
f ( s)<br />
〉<br />
The Nehari theorem says that the minimum of ||F(s)-X(s)||∞ is equal to<br />
||ΓF~||=λmax(PQ) [the proof is omitted]. Hence, it follows ||F(s)-X o (s)||∞ =<br />
λmax(PQ) so that h(s)=Γ * F~f(s). On the other hand, as it is easy to check,<br />
Γ * F~f(s)=g(s). Therefore,<br />
*<br />
~<br />
*<br />
~<br />
F<br />
( s)]<br />
f ( s),<br />
Γ<br />
f ( s),<br />
Γ<br />
= 〈 h(<br />
s),<br />
h(<br />
s)<br />
〉 − 〈 Γ<br />
λ<br />
( PQ)<br />
o<br />
X F<br />
( s)<br />
=<br />
F(<br />
s)<br />
− Γ ~<br />
*<br />
~<br />
F<br />
*<br />
~<br />
F<br />
f ( s)<br />
f ( s)<br />
2<br />
2<br />
F<br />
*<br />
~<br />
F<br />
f ( s)<br />
〉<br />
f ( s),<br />
Γ<br />
2<br />
2<br />
g(<br />
s)<br />
f ( s)<br />
≤<br />
f ( s)<br />
〉 =<br />
f ( s)<br />
〉<br />
*<br />
~<br />
F<br />
=<br />
=<br />
f ( s)<br />
〉 =
⎡a<br />
⎢<br />
⎣a<br />
1<br />
2<br />
⎤<br />
⎥<br />
⎦<br />
Chain Scattering Representation<br />
⎡b<br />
= Σ⎢<br />
⎣b<br />
1<br />
2<br />
⎤<br />
⎥<br />
⎦<br />
a1 b1<br />
If Σ21 is invertible, then b1=Σ21 -1 (a2−Σ22b2) and<br />
⎡a<br />
⎢<br />
⎣b<br />
1<br />
1<br />
⎤<br />
⎥<br />
⎦<br />
⎡a<br />
= Chain(<br />
Σ)<br />
⎢<br />
⎣a<br />
1<br />
2<br />
⎤<br />
⎥<br />
⎦<br />
• Algebra of chain scattering representation<br />
a2<br />
Chain(<br />
a1<br />
b1<br />
Σ<br />
⎡Σ<br />
) = ⎢<br />
⎣<br />
b2<br />
−1<br />
−1<br />
12 − Σ11Σ21<br />
Σ22<br />
Σ11Σ21<br />
Σ −1<br />
−1<br />
− Σ21<br />
Σ22<br />
Σ21<br />
chain(Σ)<br />
b2<br />
a2<br />
⎤<br />
⎥<br />
⎦
Homographic Trasformation<br />
a1<br />
b1<br />
a2<br />
b2<br />
a1 b2<br />
b1<br />
a2<br />
Y is J-unitary<br />
Σ<br />
S<br />
Θ =<br />
chain(Σ)<br />
S<br />
1<br />
22<br />
21<br />
12<br />
11<br />
21<br />
1<br />
22<br />
12<br />
11<br />
1<br />
1<br />
)<br />
)(<br />
(<br />
)<br />
;<br />
(<br />
)<br />
(<br />
)<br />
;<br />
(<br />
,<br />
−<br />
−<br />
Θ<br />
+<br />
Θ<br />
Θ<br />
+<br />
Θ<br />
=<br />
Θ<br />
=<br />
Φ<br />
Σ<br />
Σ<br />
−<br />
Σ<br />
+<br />
Σ<br />
=<br />
Σ<br />
=<br />
Φ<br />
Φ<br />
=<br />
S<br />
S<br />
S<br />
HM<br />
S<br />
I<br />
S<br />
S<br />
LF<br />
b<br />
a<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎣<br />
⎡<br />
−<br />
=<br />
−<br />
Ψ<br />
=<br />
Ψ<br />
=<br />
Ψ<br />
Ψ<br />
I<br />
I<br />
J<br />
s<br />
s<br />
J<br />
s<br />
J<br />
s<br />
0<br />
0<br />
,<br />
)'<br />
(<br />
)<br />
(<br />
,<br />
)<br />
(<br />
)<br />
(<br />
~<br />
~
The Standard Problem via chain-scattering<br />
representation<br />
w z<br />
P<br />
u y<br />
We say that G(s) admits a J-unitary factorization if<br />
G(s)=Θ(s)Λ(s)<br />
Where Θ(s) is J-unitary and Λ(s) is unimodular (stable with stable<br />
inverse) .<br />
Theorem12<br />
Assume that P has a chain scattering representation G=chain(P)<br />
such that G is left invertible and does not have poles or zeros on<br />
the imaginary axis. Then, the H∞ problem is solvable if and only if<br />
G has a J-unitary factorization G=ΘΛ. The controllers K can be<br />
written as K=HM(Λ -1 ;S) for each proper and stable S.<br />
• J-spectral Factorization in control and filtering<br />
K
A robust stability problem<br />
k<br />
z = w,<br />
1 − gk<br />
Z δ<br />
w<br />
k<br />
q =<br />
1 − gk<br />
We want to maximize ||d||∞ (minimize ||q|∞). For the stability of the closed<br />
loop it must be<br />
(1) q stable<br />
(2) gq stable<br />
(3) (1+gq)g =stable<br />
Hence, if pi are the unstable poles of g (assumed to be simple and with<br />
Re(pi)>0) q must be such that<br />
(0) q minimum norm<br />
(1) q stable<br />
(2) q(pi)=0<br />
(3) (1+gq)(pi)=0<br />
Letting<br />
w<br />
s + pi<br />
q = qa = q∏i<br />
pi<br />
− s<br />
it follows that q w must be such that<br />
(0) q W minimim norm<br />
(1) q W stable<br />
(2) q W (pi)=-ag -1 (pi)<br />
g<br />
k
A simple example<br />
where<br />
z<br />
δ<br />
w<br />
u y<br />
g<br />
s + 2<br />
g =<br />
( s + 1)(<br />
s −1)<br />
Find k in such a way that ||d||¥ is maximized.<br />
Take<br />
s + 1<br />
a =<br />
1−<br />
s<br />
and find q w stable of minimum norm such that such that<br />
q w (1)=-ag -1 (1)=4/3. It follows q w =4/3 so that ||d||∞max =3/4 and<br />
w<br />
q 4(<br />
s + 1)<br />
= = −<br />
1 + q g ( 3s<br />
+ 5)<br />
k w<br />
k
In state-space<br />
A=[0 1;1 0], B1=[0;0], B2=[0;1], C1=[0 0], D12=1, C2=[2 1], D21=1.<br />
Stabilizing solution of:<br />
A′P+PA-PB2B2′P+PB1B1′P/γ 2 +C1′C1=0<br />
Stabilizing solution of:<br />
AQ+QA′-PC2′C2Q+PC1′C1Q/ γ 2 +B1B1′=0<br />
rs(PQ)
Operatorial approach<br />
P<br />
⎡0<br />
s)<br />
= ⎢<br />
⎣1<br />
Bezout<br />
1⎤<br />
g<br />
⎥<br />
⎦<br />
Model Matching<br />
Inner-outer factorization of t4=t2t3<br />
Unstable part fa of f=t4i -1 t1<br />
Hence, X 0 =0, Q=0 and<br />
s + 2 n<br />
= 2<br />
s −1<br />
m<br />
s + 2<br />
( s + 1)<br />
( g =<br />
n = 2<br />
s + 5/<br />
3<br />
− ny + mx = 1, x = , y = −4<br />
s + 1<br />
t<br />
t<br />
1<br />
2<br />
t<br />
=<br />
p<br />
= t<br />
4<br />
3<br />
11<br />
=<br />
+<br />
p<br />
12<br />
s −1<br />
s + 1<br />
myp<br />
21<br />
− 4 / 3(<br />
s −1)<br />
=<br />
s + 1<br />
2<br />
( s −1)<br />
= t4it4o<br />
, t4i<br />
= t4<br />
= , t 2 4o<br />
( s + 1)<br />
/ 3<br />
= 1<br />
− 4/<br />
3(<br />
s + 1)<br />
8/<br />
3 8/<br />
3<br />
f =<br />
= −4<br />
/ 3−<br />
, f a = −<br />
s −1<br />
s −1<br />
s −1<br />
K(<br />
s)<br />
=<br />
y<br />
x<br />
− 4/<br />
3(<br />
s + 1)<br />
4(<br />
s + 1)<br />
=<br />
= −<br />
s + 5/<br />
3 3s<br />
+ 5<br />
,<br />
m =<br />
s −1<br />
s + 1
REFERENCES<br />
V.M. Adamjan, D.Z. Arov and M.G. Krein. Infinite block Hankel<br />
matrices and related extension problem. American Mathematical Society<br />
Translation, 111: 133-156, 1978.<br />
Bala Subrahmanyam, Finite Horizon H∞ and Related Control Problems,<br />
Birkhäuser, Boston, 1995.<br />
Basar, T., and Bernhard, P., H∞ Optimal Control and Related Miinimax<br />
Design Prooblems: A Dynamic Game Approach, Birkhäuser, Boston,<br />
1995.<br />
Chen, B.M., Robust and Optimal Control, Springer-Verlag, 2000.<br />
Colaneri, P., Geromel, J.C., and Locatelli, A., Control Theory & Design:<br />
An H2 and H∞ Viewpoint, Academic Press, San Diego, 1997.<br />
Dahleh, M.A., and Diaz-Bobillo, I., Control of Uncertain Systems, A<br />
Linear Programming Approach, Prentice Hall, Englewood Cliffs, 1995.<br />
Davis, C., Kahan W.M., Weinberger, H.F., "Norm.preserving dilations<br />
and their applications to optimal error bounds", SIAM Journal of Control<br />
and Optimization, vol. 20, pp. 445-469, 1982.<br />
Doyle J.C., "Analysis of feedback systems with structured<br />
uncertainties", Proceedings of the IEEE, Pard D, vol. 129, pp. 242-250,<br />
1982Hazony, D., "Zero cancellation synthesis using impedence<br />
operators", IRE Trans. on Circuit Theory, vool. 8, pp. 114-120, 1961.
Doyle J.C., Lecture Notes in Advanced Multivariable Control, Lecture<br />
Note ONR/Honeywell Workshop, Minneapolis, 1984.<br />
Doyle, J.C., Francis, B.A., and Tannenbaum, A.R., Feedback Control<br />
Theory, MacMillan Publishing Co., New York, 1992.<br />
J.C. Doyle, K. Glover, P.P. Khargonekar, B. Francis, “State-space<br />
solution to standard H2-H∞ control problems”, IEEE Trans. Autom.<br />
Control, AC-34, pp. 831-847, 1989.<br />
D. Fragopoulos, H∞ synthesis theory using polynomial system<br />
representations”, PhD Thesis, University of Strathclyde, 1994.<br />
Francis B.A., and Zames, G., "On L∞- optimal sensitivity theory for<br />
SISO feedback systems", IEEE Trans. Automat. Contr., vol. AC-29, pp.<br />
9-16, 1984.<br />
B.A. Francis, “A course in H∞ control theory”, in Lecture Notes in<br />
Control and Information Science, Vol. 88, Springer-Verlag, Berlin,<br />
1987.<br />
Glover, K., "All optimal Hankel-norm approximations of linear<br />
multivariable systems and their L∞-error bounds, International Journal<br />
of Control, vol. 39, pp. 1115-1193, 1984.<br />
M. Green, K. Glover, D.J.N. Limebeer, J.C. Doyle, “A J-spectral<br />
factorization approach to H∞ control”, SIAM J. Contr. And<br />
Optimization, 28, pp. 1350-1371, 1990.<br />
M.J. Grimble and V. Kucera, “Polynomial methods for control systems<br />
design”, Springer Verlag 1996.<br />
Helton, J.W:, and Merino, O., Classical Control Using H∞ Methods:<br />
Theory, Optimization and Design, SIAM, Philadelphia, 1988.
Heyde, R.A., H∞ Aerospace Control Design: A VSTOL Flight<br />
Application, Springer-Verlag, New-York, 1995.<br />
Kimura, H., Okinishi, F., "Chain scattering approach to control system<br />
design", In Trends in Control, An European Perspective, A.Isidori Ed.,<br />
pp. 151-171, Springer Verlag,, Berlin, 1995.<br />
Kimura, H., "Conjugation, interpolation and model-matching in H∞,<br />
International Journal of Control, vol. 49, pp. 269-307, 1989.<br />
H. Kwakernaak, M. Sebek, “Polynomial J-spectral factorization”, IEEE<br />
Trans. Autom. Control, AC-39, pp. 315-328, 1994.<br />
McFarlane D., and Glover, K, "A loop shaping design procedure using<br />
H∞ synthesis", IEEE Trans. Automat. Control, vol. 37, pp. 759-769,<br />
1992.<br />
G. Mensma, “Frequency-domain methods in H∞ Control”, PhD Thesis,<br />
University of Twente, 1993.<br />
Mustafa, D. and Glover, K., Minimum Entropy H∞ Control, Lecure<br />
Notes in Control and Information Sciences, Vol. 16, Springer-Verlag,<br />
1990.<br />
R. Nevanlinna,, "Über beschränkte Funktionen, die in gegeben Punkten<br />
vorgeschriebenne Werte annehmen", Ann. Acad, Sci. Fenn., Vol. 13, pp.<br />
27-43, 1919.<br />
G. Pick, "Über Beschränkuungen analytischer Funktionen, welche durch<br />
vorgegebene Funktions wert bewirkt sind", Mathematische Ann., Vol.<br />
77, pp. 7-23, 1916.
Stoorgvogel, A., The H∞ Control Problem: A State Space Approach,<br />
Prentice Hall, Englewwod Cliffs, New Jersey, 1992.<br />
Van Keulen, B., H∞ Control for Distribuited Parameter Systems: A<br />
State Space Approach, Birkhäuser, Boston, 1993.<br />
W.M. Vidyasagar, “Control system synthesis: a factorization approach”,<br />
MIT Press, Cambridge, 1985.<br />
I. Yaesh, “Linear Optimal Control and Estimation in the Minimum H∞<br />
norm sense”, Phd thesis, Tel Aviv University, 1992.<br />
Zames, G., "Feedback and optimal sensitivity: Model refernce<br />
transformations, multiplicity seminorms, and approximation inverses",<br />
IEEE Trans. Automat. Contr., vol. AC-26, pp. 301-320, 1981.
H¥ FILTERING IN DISCRETE-TIME<br />
P.Colaneri<br />
Dipartimento di Elettronica e Informazione del Politecnico di<br />
Milano
• H¥ analisys<br />
SUMMARY<br />
H∞ norm. Small gain theorem. Algebraic and difference analysis Riccati<br />
equation<br />
• H¥ filtering (infinite horizon)<br />
Stochastic filtering. Basic prediction problem. J-spectral factorization.<br />
Duality. Krein spaces. Risk sensitivity.<br />
• Mixed H 2 /H¥ filtering<br />
Convex optimization. α procedure<br />
• H¥ filtering (finite horizon)<br />
Feasibility. Convergence. Gamma Switching<br />
• Conclusions
x(<br />
t + 1)<br />
= Ax ( t)<br />
+ Bw(<br />
t)<br />
z(<br />
t)<br />
= Cx(<br />
t)<br />
+ Dw(<br />
t)<br />
H¥ NORM<br />
−1<br />
G ( z)<br />
= C(<br />
zI − A)<br />
B + D<br />
2<br />
jϑ − jϑ<br />
yF<br />
G( z ) ∞ = sup λ max ( G( e ) G( e )' ) = sup = sup<br />
ϑ∈ 0 , π<br />
w l<br />
2<br />
w ( )<br />
z − 0. 5<br />
G( z)<br />
= 2<br />
z + 0. 5z + 0. 7<br />
2<br />
2<br />
∈ 2 w z ∈l2<br />
2<br />
G( z ) w( z )<br />
w( z )<br />
If w( t)<br />
is a white noise, the norm represents the square-root of the peak<br />
value of the spectrum of the output<br />
______________________________________________________________<br />
For the frequency-domain definition it is enough to require that G(z) don’t have poles on the unit circle (L ∞ ). It is easy<br />
to reckognize that the H ∞ norm of a function in L ∞ coicides with the norm of its stable part G o (z) in the inner-outer<br />
factorization G(z)=G o (z)G i (z). The time-domain definition requires the external stability of the system. The symbol l 2<br />
indicates the set of the square-summable discrete-time funtions in Z + . Noice that if w belongs to l 2 and the system is<br />
asymptotically stable, the forced y F is in l 2 as well.<br />
2<br />
2<br />
2<br />
2
SMALL GAIN THEOREM: robust stability<br />
Let G(z) and Δ(z) stable. Then the closed-loop system (Δ,G) is stable for<br />
each ||Δ(z)|∞≤ α iff ||G(z)||∞
Descriptor simplectic system:<br />
Theorem 2<br />
RICCATI EQUATION<br />
FOR THE H¥ ANALISYS<br />
A is stable and G( z)<br />
< γ iff there exists Q ≥ 0 satisfying<br />
∞<br />
• (i)<br />
2 −1<br />
Q = AQA '+ BB ' + ( AQC ' + BD ')( γ I − DD '− CQC ') ( AQC '+ BD ')'<br />
• (ii) γ 2<br />
I − DD' − CQC'<br />
> 0<br />
feasibility<br />
2 −1<br />
• (iii) A + ( AQC '+ BD ')( γ I − DD' −CQC')<br />
C<br />
stability
Notes:<br />
The Riccati equation (or its dual) is associated with the simplectic system written above and described by equations<br />
below.<br />
⎡I<br />
⎢<br />
⎣0<br />
2<br />
−1<br />
2<br />
−1<br />
− C'(<br />
γ I − DD'<br />
) C ⎤ ⎡A'+<br />
C'<br />
( γ I − DD')<br />
DB'<br />
η(<br />
t + 1)<br />
=<br />
2<br />
−1<br />
⎥ ⎢<br />
−2<br />
A + BD'<br />
( γ I − DD'<br />
) C⎦<br />
⎣ − B(<br />
I − γ D'D)<br />
B'<br />
0⎤<br />
⎥η(<br />
t)<br />
I ⎦<br />
For the existence of the stabilizing solution of the Riccati equation it is necessary that such a system does not have<br />
characteristic values with unitary modulus.<br />
From the equation it readily follows that:<br />
γ<br />
2<br />
I − G(<br />
z)<br />
G(<br />
z<br />
−1<br />
)'=<br />
Ψ(<br />
z)<br />
Ψ(<br />
z)<br />
= I − C(<br />
zI − A)<br />
−1<br />
2 [ γ I − DD'−CQC']<br />
( BD'+<br />
AQC'<br />
)( γ<br />
2<br />
Ψ(<br />
z<br />
−1<br />
)',<br />
I − DD'−CQC'<br />
)<br />
This explains the fact that γ2I − DD ' −CQC<br />
' must be positive. As will be clear from the proof, the theorem can be<br />
formulated for the inequality<br />
− Q + AQA' + ( AQC ' + BD' )( γ I − DD ' − CQC ' ) ( AQC ' + BD'<br />
)' ≤<br />
2 1 0<br />
without the stability condition (iii).<br />
−<br />
Notice that as γ→∞ (the loop in the figure opens) the Riccati equation boils down to the Lyapunov equation<br />
associated with the H 2 norm.<br />
Proof of Theorem 2<br />
Assume that there exists a solution Q of the Riccati equation. Such an equation can be transformed in the following<br />
way (the computations are left to the reader):<br />
γ<br />
2<br />
I − G ( z ) G ( z<br />
−1 )' = Ψ( z ) γ<br />
2<br />
I − DD ' −CQC<br />
' Ψ ( z<br />
−1<br />
) ',<br />
where<br />
Ψ( z) = I−C( zI− A) −1 ( BD' + AQC' )( γ<br />
2<br />
I −DD' −CQC'<br />
)<br />
−1<br />
.<br />
Hence, on the basis of the property γ 2<br />
I − DD ' − CQC ' > 0 and invertibiliy of Ψ ( e )<br />
jϑ we have that<br />
2 jϑ − jϑ jϑ 2<br />
− jϑ<br />
γ I − G ( e ) G ( e )' = Ψ ( e ) γ I − DD' − CQC ' Ψ ( e )' > 0 , ∀ϑ<br />
,<br />
i.e. the thesis.<br />
Vice-versa assume that G ( z ) ∞ < γ . We prove that the symplectic system<br />
⎡I<br />
⎢<br />
⎣0<br />
2<br />
−1<br />
2<br />
−1<br />
− C'(<br />
γ I − DD'<br />
) C ⎤ ⎡A'+<br />
C'<br />
( γ I − DD')<br />
DB'<br />
η(<br />
t + 1)<br />
=<br />
2<br />
−1<br />
⎥ ⎢<br />
−2<br />
A + BD'<br />
( γ I − DD'<br />
) C⎦<br />
⎣ − B(<br />
I − γ D'D)<br />
B'<br />
−1<br />
0⎤<br />
⎥η(<br />
t)<br />
I ⎦<br />
does not have characteristic values with unitary modulus. Indeed, assume, by contradiction, λ is such, i.e.
2 −1 2 −1<br />
λx − λC ' ( γ I − DD ' ) Cy = ( A' + C' ( γ I − DD ' ) DB ' ) x<br />
2 −1 −2<br />
λ( A + BD ' ( γ I − DD ' ) C ) y = y − B( I − γ D' D ) B' x<br />
for a certain [x 'y']'≠0 and define p=(γ 2 I-DD') -1 (DB'x+λCy). It results G(λ -1 )G(λ)'=γ 2 I, which is a contradiction unless<br />
p=0. On the other hand, p=0 implies A'x=λx, and, since A is stable, x=0. Finally, p=0 and x=0 imply (λA-I)y=0 so<br />
that y=0 ( A is stable). Hence, x=0 and y=0 is a contradiction.<br />
Since the symplectic system does not have characteristic values on the unit circle, there are n characteristic values<br />
inside (possible in zero) and n outside (the reciprocals) the unit circle (possibly at infinity). Grouping the 2ndimensional<br />
vectors associated with the n characteristic values inside the unit circle, one can construct a matrix [X'<br />
Y']', whose range is a ln-dimensional subspace. It follows<br />
⎡I<br />
⎢<br />
⎣0<br />
2<br />
−1<br />
2<br />
−1<br />
− C'<br />
( γ I − DD'<br />
) C ⎤⎡X<br />
⎤ ⎡A'+<br />
C'<br />
( γ I − DD')<br />
DB'<br />
⎥⎢<br />
⎥T<br />
=<br />
2<br />
−1<br />
⎢<br />
−2<br />
A + BD'<br />
( γ I − DD'<br />
) C⎦<br />
⎣Y<br />
⎦ ⎣ − B(<br />
I −γ<br />
D'D)<br />
B'<br />
0⎤⎡<br />
X ⎤<br />
⎥⎢<br />
⎥<br />
I ⎦⎣Y<br />
⎦<br />
where T is a matrix with eigenvalues inside the unit circle (stable matrix). Consider now the solutions of the system<br />
r(t+1)=Tr(t), i.e. r(t)=T t r(0) and left multiply by r(t) the two matrix equations above. Next, define p(t)=B'Xr(t) and<br />
q(t)=CYr(t). It follows<br />
p( t + 1)<br />
= A' p( t) + C ' v1( t ), v2 ( t ) = B' p( t ) + D ' v1( t )<br />
Aq ( t + 1)<br />
= q ( t ) − Bv2 ( t) ,<br />
1<br />
v1( t ) = ( Cq( t + 1)<br />
+ Dv ( t ))<br />
2 2<br />
γ<br />
from which, passing to the Z-transform<br />
−1 2 −1<br />
2<br />
G ( z ) ' G ( z ) v1( z ) = γ v1( z) , G ( z ) G ( z ) ' v2( z ) = γ v2 ( z ) .<br />
From the assumption it is v1 ( t ) = 0, v2 ( t ) = 0 , so that B'X =0 and C'Y=0. Hence, from the symplectic system it<br />
follows A'X=XT, AYT=Y. Assume that there exists a vector d such that Xd=0 and take the minimal degrre polynomial<br />
ϖ(T) of T such that ϖ(T)d=0. Such a polynomial can be written as ϖ(T)=(λI-T)μ(T). For the minimality of ϖ(T) it<br />
results μ(T)d≠0. Then, AYTd=λAYd=Yd. Since λ has modulus less than 1, λ-1 has modulus greater than one so that<br />
this equation contradicts the stability of A', unless Yd=0. On the other hand, this last conclusion, together with Xd=0<br />
contradicts the n-dimensionality of the range of [X 'Y']'.
THE DIFFERENCE EQUATION<br />
x( t + 1)<br />
= Ax( t ) + Bw( t )<br />
y( t ) = Cx( t ) + Dw( t )<br />
Q( t + 1)<br />
= AQ( t ) A' + ( AQ( t ) C' + BD' )( γ I − DD' − CQ( t ) C' ) ( AQ( t ) C' + BD' )' + BB'<br />
Q( 0 ) = Q > 0<br />
0<br />
Theorem 3<br />
N<br />
N<br />
2 −1<br />
J = ∑ y( t )' y( t ) − γ ∑w( t )' w( t ) − γ x( 0 )' Q x( 0 ) < 0, ∀( w(.), x(<br />
0 )) ≠ 0<br />
t=<br />
0<br />
t=<br />
0<br />
2<br />
2<br />
−<br />
0 1<br />
iff there exists Q(t)≥0 in [0,N] such that γ 2<br />
I − DD' − CQ( t ) C'<br />
><br />
0
Proof of Theorem 3<br />
Consider the cost.<br />
N<br />
N<br />
~<br />
2<br />
J ( N)<br />
= z(<br />
t)'<br />
z(<br />
t)<br />
−γ<br />
w(<br />
t)'<br />
w(<br />
t)<br />
+ x(<br />
N + 1)'<br />
PN<br />
+ 1x(<br />
N + 1)<br />
< 0,<br />
x(<br />
0)<br />
= 0,<br />
∀w(.)<br />
≠ 0<br />
∑<br />
t=<br />
0<br />
~<br />
∑<br />
t=<br />
0<br />
We proof the dual result: J ( N)<br />
< 0,<br />
x(<br />
0)<br />
= 0,<br />
∀w(.)<br />
≠ 0 iff there exists the solution of the equation<br />
P( t ) = A' P( t + 1)<br />
A +<br />
2 −1<br />
+ ( A' P( t + 1) B + C' D )( γ I − D' D − B' P ( t + 1 ) B ) ( A' P( t + 1 ) B + C' D )' + C' C ,<br />
P( N + 1) = PN + 1 > 0<br />
with γ 2<br />
I − D'D − B' P( t + 1) B><br />
0 , in [0,N+1]. Sufficient condition: if there exists the solution with the indicated<br />
properties, then pre-premultiplying the equation by x(t)' and post-multiplying by x(t), and summing all terms, it results<br />
N<br />
∑<br />
t=<br />
0<br />
(*)<br />
where<br />
z(<br />
t)'<br />
z(<br />
t)<br />
−γ<br />
2<br />
N<br />
∑<br />
t=<br />
0<br />
w(<br />
t)'w(<br />
t)<br />
+ x(<br />
N + 1)'<br />
P<br />
~ 1<br />
N + 1<br />
x(<br />
N + 1)<br />
= x(<br />
0)'P<br />
( 0)<br />
x(<br />
0)<br />
−<br />
2<br />
−<br />
w ( t)<br />
= ( I − D'<br />
D − B'<br />
P(<br />
t + 1)<br />
B)<br />
( A'<br />
P(<br />
t + 1)<br />
B + C'<br />
D)'<br />
x(<br />
t)<br />
γ .<br />
Hence , being x(0)=0 we have J ( N)<br />
< 0,<br />
∀w(.)<br />
≠ 0 .<br />
~<br />
~<br />
N<br />
∑<br />
t=<br />
0<br />
( w(<br />
t)<br />
−w~<br />
( t))'(<br />
w(<br />
t)<br />
− w~<br />
( t))<br />
Vice-versa, assume that J ( N)<br />
< 0,<br />
∀w(.)<br />
≠ 0 and take w(.)=0 in [0,N-1]. Being x(0)=0, with such a choice we have<br />
0 > w( N)' − γ 2I+<br />
D'D + BPN + 1B'<br />
w( N), ∀w( N)<br />
≠0<br />
so that<br />
− γ + + + ><br />
2<br />
I D' D BPN 1B'<br />
0.<br />
Furthermore, from<br />
~ ~<br />
J ( N)<br />
= J ( N −1)<br />
− ( w(<br />
N)<br />
− w~<br />
( N ))'(<br />
w(<br />
N)<br />
− w~<br />
( N))<br />
< 0,<br />
∀w(.)<br />
∈ ,<br />
it turns out taht<br />
~<br />
( N −1) > 0,<br />
∀w(.)<br />
∈ N<br />
[ 0,<br />
−1]<br />
J .<br />
By iterating this reasoning we have<br />
γ 2<br />
I − D'D − B' P( t + 1) B><br />
0<br />
in [0,N]. Theorem 3 is proven from (*) by noticing that<br />
−1 2<br />
P( t ) = Q( t ) γ .<br />
[ 0 N ]
STOCHASTIC FILTERING<br />
⎧ x(<br />
t + 1)<br />
= Ax(<br />
t)<br />
+ v1(<br />
t)<br />
⎪<br />
⎨ z(<br />
t)<br />
= Lx(<br />
t)<br />
⎪<br />
⎩ y(<br />
t)<br />
= Cx(<br />
t)<br />
+ v2(<br />
t)<br />
where v1( t), v2 ( t)<br />
are zero-mean white gaussian noises . Moreover,<br />
⎡v1<br />
( t)<br />
⎤ ⎡ V1<br />
E ⎢ =<br />
2(<br />
)<br />
⎥ ⎢<br />
⎣v<br />
t ⎦ ⎣V12<br />
'<br />
V<br />
V<br />
12<br />
2<br />
⎤<br />
⎥,<br />
⎦<br />
V<br />
2<br />
> 0<br />
The problem is to find an estimate z ˆ( t)<br />
in H∞ of a linear combination<br />
z ( t)<br />
= Lx(<br />
t)<br />
of the state, i.e. one wants to find an estimator based on the<br />
observations of { y( i),<br />
i ≤ t − N}<br />
) in such a way that the maximum of the<br />
output spectrum of the error ε = z − zˆ<br />
due to [ v 1'<br />
v2']'<br />
is minimized (or<br />
bounded from above by a certain positive scalar γ).<br />
Letting<br />
u( t)<br />
= −zˆ<br />
( t),<br />
ε ( t)<br />
= Lx(<br />
t)<br />
+ u(<br />
t),<br />
D =<br />
B =<br />
1/<br />
2 [ 0 V2<br />
] , C = C<br />
−1<br />
1/<br />
2<br />
−1/<br />
2<br />
[ ( V −V<br />
V V ')<br />
V V ]<br />
1<br />
12<br />
2<br />
12<br />
12<br />
2<br />
y(<br />
t)<br />
= V<br />
−1/<br />
2<br />
2<br />
y(<br />
t)<br />
and defining with w t<br />
( ) a zero mean white gaussian noise with identity<br />
covariance, it follows:
BASIC FILTERING PROBLEM<br />
S<br />
⎧ x(<br />
t + 1)<br />
= Ax(<br />
t)<br />
+ Bw(<br />
t)<br />
⎪<br />
S : ⎨ ε(<br />
t)<br />
= Lx(<br />
t)<br />
+ u(<br />
t)<br />
⎪<br />
⎩ y(<br />
t)<br />
= Cx(<br />
t)<br />
+ Dw(<br />
t)<br />
⎧ xˆ<br />
( t + 1)<br />
= A<br />
F : ⎨<br />
⎩ u(<br />
t)<br />
= C<br />
F<br />
F<br />
xˆ<br />
( t)<br />
+ B<br />
xˆ<br />
( t)<br />
+ D<br />
Basic problem: Find A F , BF<br />
, CF<br />
, DF<br />
such that<br />
F<br />
F<br />
e<br />
y(<br />
t − N)<br />
y(<br />
t − N)<br />
• (S,F) is asymptotically stable<br />
T < γ<br />
•<br />
ε w<br />
∞<br />
________________________________________________________________<br />
If z(t)=Lx(t) is the variable to be estimated and z ˆ( t)<br />
the estimated variable, it follows u( t)<br />
= −zˆ<br />
( t)<br />
. Hence ε( t )<br />
represents the filtering error z( t)<br />
= z(<br />
t)<br />
− zˆ<br />
( t)<br />
= ε(<br />
t)<br />
. If w(t) is a white noise, the problem is that of minimizing<br />
the peak value of the error spectrum If w(t) is a generic signal in l2 , the problem is that of minimizing the l2 “gain”<br />
between the error and the disturbance.
SOLUTION OF THE FILTERING PROBLEM<br />
DD'> 0<br />
Ipotesi: A = stable<br />
( A , B,<br />
C,<br />
D)<br />
does not have unstable invariant zeros<br />
Theorem 4<br />
There exists a filter F solving the basic problem iff ∃ Q≥0 solution of<br />
⎡ CQA'+<br />
DB'⎤<br />
⎡DD'+<br />
CQC'<br />
i)<br />
Q = AQA'+<br />
BB'−⎢<br />
⎥ ⎢<br />
⎣ LQA'<br />
⎦ ⎣ LQC'<br />
ii) V = γ<br />
2<br />
I − LQL'+<br />
LQC'(<br />
DD'+<br />
CQC ')<br />
⎡ CQA'+<br />
DB'⎤<br />
⎡DD'+<br />
CQC'<br />
iii)<br />
Aˆ<br />
= A − ⎢ ⎥ ⎢<br />
⎣ LQA'<br />
⎦ ⎣ LQC'<br />
'<br />
2<br />
'<br />
−1 CQL<br />
CQL'<br />
⎤<br />
2<br />
− γ I + LQL'<br />
⎥<br />
⎦<br />
'><br />
0<br />
CQL'<br />
⎤<br />
2<br />
−γ<br />
I + LQL'<br />
⎥<br />
⎦<br />
CENTRAL FILTER<br />
ξ ( t + 1)<br />
= Aξ(<br />
t)<br />
+ F ( y(<br />
t)<br />
−Cξ<br />
( t)),<br />
zˆ(<br />
t)<br />
= Lξ<br />
( t)<br />
+ F ( y(<br />
t)<br />
− Cξ<br />
( t)),<br />
F<br />
1<br />
2<br />
F<br />
1<br />
−1<br />
−1<br />
⎡ CQA+<br />
⎢<br />
⎣<br />
filter feasibility<br />
⎡C⎤<br />
⎢ ⎥<br />
⎣L⎦<br />
= LQC'(<br />
CQC'+<br />
DD')<br />
−1<br />
stable<br />
= ( AQC'+<br />
BD')(<br />
CQC'+<br />
DD')<br />
−1<br />
' DB'⎤<br />
LQA'<br />
⎥<br />
⎦<br />
_________________________________________________________________________<br />
The assumption on the zeros is equivalent to the stabilizability of the pair<br />
~ ~ ~<br />
−1<br />
~<br />
−1<br />
( A,<br />
B),<br />
A = A − BD'<br />
( DD'<br />
) C,<br />
B = B(<br />
I − D'(<br />
DD'<br />
) D)<br />
. The assumption of stability of A can be weaked to<br />
the assumption of detectability of (A,C) if one is only interested in the stability of the filter.
NOTES<br />
• If G≠0, one obtains the central filter<br />
ξ ( t + 1)<br />
= Aξ(<br />
t)<br />
+ F ( y(<br />
t)<br />
− Cξ(<br />
t))<br />
+ Gu(<br />
t)<br />
zˆ<br />
( t)<br />
= Lξ(<br />
t)<br />
+ F ( y(<br />
t)<br />
−Cξ<br />
( t))<br />
2<br />
1<br />
• As γ → ∞ the H 2 filtering is obtained, where F 1 e F 2 depend on the<br />
stabilizing solution of the standard Riccati equation:<br />
Q = AQA '+ BB' − ( CQA' + DB')'( DD' + CQC) ( CQA' + DB')<br />
• Notice that in the H ∞, filter the solution depends on L, i.e. on the<br />
particular linear combination of the state to be estimated. This is a<br />
difference between H2, and H ∞, design.<br />
• The “gains” F 1 and F 2 are related by:<br />
with<br />
F<br />
F<br />
2<br />
1<br />
=<br />
=<br />
LK<br />
−1<br />
−1<br />
[ A − BD'(<br />
DD')<br />
C]<br />
K − BD'(<br />
DD')<br />
K = QC'(<br />
CQC'+<br />
DD')<br />
In the uncorrelated case (DB'=0) these relations become<br />
Hence<br />
F = LK , F = AK<br />
2 1<br />
ξ ( t +<br />
1)<br />
= Aξ<br />
( t)<br />
+<br />
rˆ<br />
( t)<br />
= Lξ<br />
( t)<br />
+ LK<br />
−1<br />
AK[<br />
y(<br />
t)<br />
−Cξ<br />
( t)<br />
]<br />
[ y(<br />
t)<br />
− Cξ<br />
( t)<br />
]<br />
−1
BASIC ONE-STEP PREDICTION PROBLEM<br />
Assume that we want to estimate z ( t)<br />
= Lx(<br />
t)<br />
on the basis of the data<br />
y (i)<br />
, i≤ t.<br />
S<br />
P<br />
⎧ x(<br />
t + 1)<br />
= Ax(<br />
t)<br />
+ Bw(<br />
t)<br />
+ Gu(<br />
t)<br />
⎪<br />
S : ⎨ ε(<br />
t)<br />
= Lx(<br />
t)<br />
+ u(<br />
t)<br />
⎪<br />
⎩ y(<br />
t)<br />
= Cx(<br />
t)<br />
+ Dw(<br />
t)<br />
⎧<br />
~<br />
x(<br />
t + 1)<br />
= A<br />
~<br />
P x(<br />
t)<br />
+ BP<br />
y(<br />
t)<br />
P : ⎨<br />
u(<br />
t)<br />
C<br />
~<br />
⎩ = Px<br />
( t)<br />
Basic problem: Find AP , BP , CP<br />
such that:<br />
• (S,F) asymptotically stable<br />
• T rw ∞
SOLUTION OF THE ONE-STEP PREDICTION<br />
PROBLEM<br />
Assumption:<br />
A = stable, DD'> 0, ( A, B, C, D ) without unstable zeros<br />
Theorem 5<br />
There exists a one-step predictor P solving the problem iff ∃ Q≥0<br />
satisfying<br />
'<br />
⎡ CQA'+<br />
DB'⎤<br />
⎡DD'+<br />
CQC'<br />
i)<br />
Q = AQA'+<br />
BB'−⎢<br />
⎥ ⎢<br />
⎣ LQA'<br />
⎦ ⎣ LQC'<br />
CQL'<br />
⎤<br />
2 −γ<br />
I + LQL'<br />
⎥<br />
⎦<br />
⎡ CQA'+<br />
DB'⎤<br />
⎢ ⎥<br />
⎣ LQA'<br />
⎦<br />
2<br />
ii) γ I − LQL'><br />
0<br />
predictor feasibility<br />
iii )<br />
⎡ CQA + DB ⎤ ⎡DD<br />
+ CQC<br />
Aˆ<br />
' ' ' '<br />
= A − ⎢<br />
⎥ ⎢<br />
⎣ LQA ' ⎦ ⎣ LQC '<br />
'<br />
− γ<br />
CENTRAL PREDICTOR<br />
CQL ' ⎤<br />
⎥<br />
I + LQL '⎦<br />
η(<br />
t + 1)<br />
= Aη(<br />
t)<br />
+ F3(<br />
y(<br />
t)<br />
−Cη(<br />
t))<br />
~ z(<br />
t)<br />
= Lη(<br />
t)<br />
2<br />
−1<br />
−1<br />
⎡ C ⎤<br />
⎢ ⎥<br />
⎣−<br />
L⎦<br />
F = ( AQC' + BD' )( CQC' + DD' ) + A QL'V LQC' ( CQC' + DD'<br />
)<br />
3<br />
A = A − ( AQC' + BD' )( CQC' + DD' ) C<br />
X<br />
2 −1<br />
V = γ I − LQL' + LQC' ( DD' + CQC' ) CQL'<br />
stable<br />
−1 X<br />
−1<br />
−1 −1<br />
_________________________________________________________________________<br />
The assumption on the zeros is equivalent to the stabilizability of the pair<br />
~ ~ ~<br />
−1<br />
~<br />
−1<br />
( A,<br />
B),<br />
A = A − BD'<br />
( DD'<br />
) C,<br />
B = B(<br />
I − D'(<br />
DD'<br />
) D)<br />
. The assumption of stability of A can be weakened<br />
to that of detectability of (A,C), if one is only interested to the stability of the predictor.
NOTES<br />
• If G≠0, the so-called central predictor is obtained<br />
μ t + 1)<br />
= Aη(<br />
t)<br />
+ F<br />
~ z(<br />
t)<br />
= Lη(<br />
t)<br />
( 3<br />
( y(<br />
t)<br />
− Cη(<br />
t))<br />
+ Gu(<br />
t)<br />
3 → 1,<br />
i.e.<br />
the gain of the predictor and the gain of the filter coincide and depend<br />
on the stabilizing solution of the standard Riccati equation:<br />
• When γ → ∞ the H 2 predictor is recovered. Moreover, F F<br />
Q = AQA '+ BB' − ( CQA' + DB')'( DD' + CQC) ( CQA' + DB')<br />
• Notice that the H ∞ predictor depends on L, i.e. on the particular linear<br />
combination of the state to be estimated. This is a difference between<br />
H2, and H ∞, design.<br />
• The “gain” F3 of the predictor and the “gains” F1 and F2 of the filter<br />
are related by:<br />
F = F +<br />
−<br />
A − F C QL'V F<br />
3 1 1<br />
________________________________________________________________<br />
In the uncorrelated case (DB'=0) and under the assumption of reachability of (A,B), the Ricacti equation of<br />
Theorem 3 can be rewritten in the following way:<br />
⎡ −1<br />
−1<br />
1 ⎤<br />
Q = A⎢Q<br />
+ C'<br />
( DD'<br />
) C − L'<br />
L A'+<br />
BB'<br />
2 ⎥<br />
⎣<br />
γ ⎦<br />
Feasibility condition for the predictor γ 2<br />
I − LQL'><br />
0<br />
Feasibility condition for the filter<br />
Filter/Predictor<br />
−1 −2<br />
Q + C' C − L' Lγ<br />
> 0<br />
F<br />
1<br />
−1<br />
1 2<br />
−1<br />
−1<br />
−1<br />
−1<br />
−1<br />
1<br />
= A(<br />
Q + C'C<br />
) C',<br />
F2<br />
= L(<br />
Q + C'C)<br />
C',<br />
F3<br />
= A(<br />
Q + C'C<br />
− L'L)<br />
C'<br />
2<br />
γ<br />
−1
PREDICTOR VS FILTER<br />
The Riccati equation is the same, but the feasibility conditions are<br />
different<br />
System for the prediction error<br />
μ( t + ) = ( A − F C) μ(<br />
t) + ( F C − B) w( t)<br />
1 3 3<br />
e( t) = Lμ( t)<br />
feasibility (analysis) : γ 2<br />
I − LQL'<br />
><br />
System for the filtering error<br />
η(<br />
t + 1)<br />
= ( A − F C)<br />
η(<br />
t)<br />
+ ( F D − B)<br />
w(<br />
t)<br />
e(<br />
t)<br />
= ( L − F C)<br />
η(<br />
t)<br />
+ F Dw(<br />
t)<br />
feasibility (analysis) :<br />
γ<br />
2<br />
1<br />
2 −1<br />
2<br />
1<br />
0<br />
I − LQL' + LQC '( DD' + CQC') CQL' −( F − F )( DD' + CQC')( F − F )' ><br />
F = LQC'( CQC' + DD')<br />
2<br />
−1<br />
2 2 2 2 0<br />
________________________________________________________________<br />
1) At γ fixed, the fesibility condition of the predictor is more stringent than that of the filter. They coincide as γ<br />
tends to infinity, to witness the fact that the Kalman predictor and filter share the same existence conditions.<br />
2) Assuming the existence of the stabilizing solution of the Riccati equation satisfying the feasibility condition, the<br />
sufficient part of Theorem 3 is proved from the analysis result. Indeed, the analysis Riccati equation coincides<br />
with the synthesis Riccati equation once the gains F 1 , F 2 (for the filter) or F 3 (for the predictor) are substituted<br />
there.
Proof of Theorem 4/5<br />
For simplicity we study the case where DB'=0, DD'=I, (A,B) reachable. Moreover, we limit the proof to filters<br />
(predictors) in observer form. This is a restriction, but the analysis of all filters (predictors) ensuring the norm<br />
bounded is far from being trivial and is not within our scope. Notice that if F 1 =AK e F 2 =LK are the characteristic<br />
gains of a generic filter, the transfer function from the disturbance to the error is<br />
T εw =(L-LKC)(zI-A+AKC) -1 (AKD-B)+LKD<br />
So that the analysis Riccati equation is<br />
−1 −2 −1<br />
Q = A( X − L' Lγ ) A' + BB' ,<br />
with the feasibility condition<br />
X = ( I − KC ) Q( I − KC )' + KK'<br />
(1)<br />
γ 2<br />
I − LXL'<br />
> 0<br />
and asymptotic stability of<br />
(2)<br />
2 −1<br />
A( I − KC ) + AXL' ( γ I − LXL' ) ( I − KC )' L'<br />
On the other hand, if F3 is the gain of a generic one-step predictor, the transfer function is<br />
(3)<br />
T εw =L(zI-A+F 3 C) -1 (F 3 D-B),<br />
and then the analysis Riccati equation is:<br />
−1<br />
−2<br />
−1<br />
Q = A(<br />
Q + C'C<br />
− L'<br />
Lγ<br />
) + BB'+<br />
( F<br />
Y = ( Q<br />
−1<br />
− L'<br />
Lγ<br />
−2<br />
)<br />
3<br />
− AYC'(<br />
I + CYC')<br />
−1<br />
)( I + CYC')(<br />
F<br />
3<br />
− AYC'(<br />
I + CYC')<br />
with the feasibility condition<br />
γ 2<br />
I − LQL ' > 0<br />
(5)<br />
and asymptotic stability of<br />
2 −1<br />
A − F3C + ( A + F3C ) QL' ( γ I − LQL' ) L'<br />
(6)<br />
Now, assume that there exists Q≥0 satisfying Q=A[Q -1 +C'C-L'Lγ -2 ] -1 +BB', and such that Q -1 +C'C-L'Lγ -2 >0 and<br />
A(Q -1 +C'C-L'Lγ -2 ) -1 Q -1 stable. Take the filter characterized by F 1 =AK, F 2 =LK, with K=(Q -1 +C'C) -1 C'. From (1) it<br />
results X=(Q-1 +C'C) -1 and hence<br />
Q = AXA' + BB' + AXL' ( γ2I − LXL' ) −1 LXA' = A( X− 1 − L' Lγ− 2 ) −1 A' + BB' = A( Q−1 + C' C− L' Lγ− 2 ) −1<br />
A' + BB'<br />
Therefore, the analysis Riccati equation admits a positive semidefinite solution with<br />
2 2 −1<br />
γ I − LXL' = γ I − LQL ' LQC ' ( I + CQC ' ) CQL'<br />
> 0,<br />
so that the feasibility condition is satisfied. Also the stability condition (3) is satisfied.<br />
Hence the norm is less than γ.<br />
Assume now that there exists Q≥0 satisfying Q=A[Q -1 +C'C-L'Lγ -2 ] -1 +BB', and such that Q -1 -L'Lγ -2 >0 and A(Q -<br />
1 +C'C-L'Lγ -2 ) -1 Q -1 stable. Take a predictor defined by F3 =A(Q -1 +C'C-L'Lγ -2 ) -1 C'. From (4) it results that<br />
F 3 =AYC'(I+CYC') -1 and hence Q=AYA'+BB'-AYC'(I+CYC') -1 CYA'=A(Q -1 +C'C-L'Lγ -2 ) -1 A'+BB'. The analysis<br />
Riccati equation admits a positive semidefinite solution and Q -1 -L'Lγ -2 >0 coincides with the feasibility condition<br />
(5) for the predictor. Also the stability condition (6) is satisfied .<br />
Hence the norm is less than γ.<br />
Vice-versa, assume that there exists a filter such that the norm is less than γ. Then, there exists Q solving (1)-(3). On<br />
the other hand, X=(Q -1 +C'C) -1 +(K-(Q -1 +C'C) -1 C')(I+CQC')(K-(Q -1 +C'C) -1 C')'≥(Q -1 +C'C) -1 and hence X -1 ≤Q -1 +C'C<br />
implies that Q solves the inequality Q≥A(Q -1 '+C'C-L'Lγ -2 ) -1 A’+BB' whereas γ 2 I-LXL'>0 with X>0 implies that<br />
Q -1 +C'C-L'Lγ -2 >0. The existence of Q solving Q=A(Q -1 +C'C-L'Lγ -2 ) -1 A'+BB' with Q -1 A'+C'C-L'Lγ -2 >0 and<br />
satisfying the stability condition follows form standard monotonicity results of the solutions of the Riccati equation.<br />
Finally, assume that there exists a predictor such that the norm is less than Then, there exists Q solving (4)-(6).<br />
Hence Q≥A(Q -1 A'+C'C-L'Lγ -2 ) -1 A+BB' with γ 2 I-LQL'>0 The existence of Q solving Q=A(Q -1 A'+C'C-L'Lγ -2 ) -<br />
1 A'+BB' with γ 2 I-LQL'>0 and satisfying the stability condition follows form standard monotonicity results of the<br />
solutions of the Riccati equation<br />
−1<br />
(4)<br />
)'
J-SPECTRAL FACTORIZATION<br />
The Riccati equation underlying the problem solution can be compactly<br />
written as:<br />
~ ~ ~ ~ −1<br />
~ ~<br />
[ AQC<br />
'+<br />
BD'][<br />
R + CQC<br />
']<br />
[ CQA'<br />
DB']<br />
Q = AQA'+<br />
BB'−<br />
+<br />
~ ⎡C⎤<br />
~ ⎡D⎤<br />
⎡DD'<br />
= ⎢ ⎥,<br />
D = ⎢ ⎥,<br />
R = ⎢<br />
⎣ L⎦<br />
⎣ 0 ⎦ ⎣ 0<br />
0 ⎤<br />
− γ I<br />
⎥<br />
⎦<br />
C 2<br />
Let define:<br />
H<br />
H<br />
1<br />
2<br />
( z)<br />
= D + C(<br />
zI − A)<br />
−<br />
( z)<br />
= L(<br />
zI − A)<br />
1<br />
B<br />
−1<br />
B<br />
⎡H1(<br />
z)<br />
H(<br />
z)<br />
= ⎢<br />
⎣H<br />
2(<br />
z)<br />
0⎤<br />
I<br />
⎥<br />
⎦<br />
⎡I<br />
= ⎢<br />
⎣0<br />
0 ⎤<br />
− γ I<br />
⎥<br />
⎦<br />
J 2<br />
Theorem 6<br />
Assume that A is stable. There exists a stable filter F(z) such that<br />
H2(z)-F(z)H1(z) is stable and<br />
H ( z)<br />
− F(<br />
z)<br />
H ( z)<br />
2<br />
H ( z)<br />
− F(<br />
z)<br />
H ( z)<br />
2<br />
1<br />
1<br />
∞<br />
stable<br />
< γ<br />
if and only if there exists a square Ω(z), stable with stable inverse, with<br />
stable [Ω(z) -1 ]11, solving the factorization problem<br />
H ( z)<br />
JH ( z<br />
)'=<br />
Ω(<br />
z)<br />
JΩ(<br />
z<br />
−1<br />
−1<br />
)'<br />
All feasible filters can be parametrized in the following way:<br />
−1<br />
F(<br />
z)<br />
= ( Ω21(<br />
z)<br />
− Ω22<br />
( z)<br />
U(<br />
z))(<br />
Ω11(<br />
z)<br />
− Ω12<br />
( z)<br />
U ( z))<br />
, U ( z)<br />
∈ H ∞,<br />
U ( z)<br />
< γ ∞
Notes:<br />
From z = H2<br />
( z)<br />
w,<br />
y = H1(<br />
z)<br />
w,<br />
zˆ<br />
= F(<br />
z)<br />
y , it results z − zˆ<br />
= ( H1(<br />
z)<br />
− F ( z)<br />
H2<br />
( z))<br />
w and then the<br />
filtering problem can be equivalently rewritten as a model matching problem.<br />
Proof of Theorem 6<br />
Assume that the exists a feasible filter. We limit ourselves to the case where the filter is in the observer form. From<br />
Theorem 4 we know that the exists a positive semidefinite solution Q of the Riccati equation. Moreover, such a<br />
solution is stabilizing and feasible. It is a matter of cumbersome computations to show that<br />
Ω(<br />
z)<br />
=<br />
V<br />
= γ<br />
2<br />
~<br />
−1<br />
~ ~ ~ ~ −1<br />
[ I + C(<br />
zI − A)<br />
( BD'+<br />
AQC<br />
')(<br />
R + CQC')<br />
]<br />
I − LQL'+<br />
LQC'<br />
( DD'+<br />
CQC'<br />
)<br />
−1<br />
CQL'<br />
1/<br />
2<br />
⎡ ( DD'+<br />
CQC'<br />
)<br />
⎢<br />
⎢LQC<br />
'(<br />
DD'+<br />
CQC'<br />
)<br />
⎣<br />
is a J-spectral factor. The stability of Ω(z) follows from stability of A. Stability of Ω(z) -1 and [Ω(z) -1 ]11 follows from<br />
Q being a stabilizing solution. Existence of V 1/2 follows from the feasibility condition.<br />
Vice-versa, assume that Ω(z) with the given properties exists and take the formula defining all filters F(z). It follows<br />
−1<br />
−1<br />
−1<br />
−1<br />
[ ( z)<br />
− I]<br />
= −(<br />
Ω ( z)<br />
− Ω ( z)<br />
Ω ( z)<br />
Ω ( z)<br />
)( I −U<br />
( z)<br />
Ω ( z)<br />
Ω ( z)<br />
) [ U ( z)<br />
I]<br />
Ω(<br />
z)<br />
F (*)<br />
22<br />
Notice that, thanks to the small gain theorem,<br />
( ) 1<br />
−1 −<br />
I<br />
− U(<br />
z)<br />
Ω ( z)<br />
Ω ( z)<br />
21<br />
11<br />
12<br />
is stable. Indeed, U(z) is stable and ||U(z)||∞
DUALITY'<br />
The prediction problem is dual with respect to the full-state control<br />
problem (x is measurable).<br />
The filtering problem is dual with respect to the full-information<br />
control problem (x and w are measurable).<br />
⎧ x(<br />
t + 1)<br />
= Ax(<br />
t)<br />
+ Bw(<br />
t)<br />
+<br />
⎪<br />
S : ⎨ ε ( t)<br />
= Lx(<br />
t)<br />
+ u(<br />
t)<br />
⎪<br />
⎩ y(<br />
t)<br />
= Cx(<br />
t)<br />
+ Dw(<br />
t)<br />
( ( ( (<br />
⎧ x(<br />
t + 1)<br />
= A'<br />
x(<br />
t)<br />
+ L'ε<br />
( t)<br />
+ C'<br />
y(<br />
t)<br />
S' R : ⎨ ( ( (<br />
⎩ w(<br />
t)<br />
= B'x(<br />
t)<br />
+ D'<br />
y(<br />
t)<br />
Theorem 7<br />
y t)<br />
K x(<br />
t)<br />
K ε ( t)<br />
( ( (<br />
= − − stabilizes S'R and T ( (<br />
wε<br />
< γ iff<br />
( 1<br />
2<br />
ξ( t + ) = Aξ( t) + K '( y( t) − Cξ( t))<br />
1 1<br />
− u( t) = Lξ( t) + K '( y( t) − Cξ( t))<br />
stabilizes S and Tεw ∞<br />
2<br />
< γ .<br />
∞<br />
OE<br />
FI-SF<br />
_________________________________________________________________<br />
Notice that if K2 is zero (state-feedback), the filter is indeed a predictor. The proof of Theorem 7 is immediate. It<br />
is enough to verify that for any K1 and K2 it results T ( (<br />
εw<br />
= Twε<br />
' . Actually, an extended version of this result<br />
allows to link any full-information controller K(z) for system S R ' to a filter F(z) for system S.
KREIN SPACES<br />
The filtering (prediction) problem can be interpreted as a Kalman<br />
filtering (prediction) problem in the Krein space (instead of Hilbert<br />
space).<br />
with<br />
~ ⎡C⎤<br />
C = ⎢ ⎥,<br />
⎣L⎦<br />
( )<br />
~<br />
( ) ( )<br />
) (<br />
⎧ x(<br />
t + 1)<br />
= Ax(<br />
t)<br />
+ v1<br />
t<br />
⎪<br />
⎨ z(<br />
t)<br />
= Lx t<br />
⎪<br />
⎩ y(<br />
t)<br />
= Cx<br />
t + v2<br />
t<br />
⎡⎡v1<br />
⎤<br />
E⎢⎢<br />
⎥<br />
⎣⎣v2⎦<br />
[ v ' v ']<br />
1<br />
2<br />
⎡BB'<br />
⎤<br />
=<br />
⎢<br />
⎥ ⎢<br />
DB'<br />
⎦ ⎢⎣<br />
0<br />
BD'<br />
DD'<br />
0<br />
0 ⎤<br />
0<br />
⎥<br />
⎥<br />
2<br />
−γ<br />
I⎥⎦
RISK SENSITIVITY<br />
x( t + ) = Ax ( t) + v ( t)<br />
1 1<br />
y( t) = Cx( t) + v ( t)<br />
where the noises are white gaussian noises with zero mean.<br />
The problem of the risk sensitive filter is to minimize<br />
⎡ 2 ⎛ ⎛<br />
minμ<br />
( ϑ)<br />
= min⎢−<br />
log⎜<br />
E⎜<br />
zˆ<br />
( y)<br />
zˆ<br />
( y)<br />
⎢<br />
⎜ ⎜<br />
e<br />
⎣<br />
ϑ<br />
⎝ ⎝<br />
The parameter θ is the risk parameter<br />
θ=0 risk-neutral Kalman<br />
θ>0 risk-seeking modified Kalman<br />
θ
MIXED H2/H¥ FILTERING<br />
It makes sense to distinguish between two classes of disturbances:<br />
those whose statistics is known and those which are only requested to<br />
be bounded. For instance, in general, the spectral characteristics of the<br />
transducer errors are known whereas those related to modelling errors<br />
are unknown. Consequently, we can write a mixed filtering problem<br />
associated to a variable z(t):<br />
x( t + ) = Ax ( t)<br />
+ Bw( t)<br />
= Ax ( t)<br />
+ B w ( t)<br />
+ B w ( t)<br />
z( t)<br />
y( t)<br />
1 1 1<br />
= Lx( t)<br />
= Cx( t)<br />
+ Dw( t)<br />
= Cx( t)<br />
+ D w ( t)<br />
+ D w ( t)<br />
ε ( t) = z(<br />
t)<br />
− zˆ<br />
( t)<br />
Mixed Problem<br />
Find a filter solving:<br />
1 1<br />
{ Tε , w ( F)<br />
s.<br />
t.<br />
Tε<br />
, w ( F)<br />
≤ γ }<br />
inf 1<br />
2<br />
2<br />
∞<br />
2 2<br />
2 2<br />
The solution to this problem is non trivial if γ < Tε, w ( Kalman)<br />
2 ∞ .<br />
Up to now there not exist a closed-form solution. It can be proven that<br />
the optimal mixed solution Fo is on the “boundary”, i.e. such that<br />
o<br />
Tε, w ( F )<br />
2 ∞ =γ.<br />
___________________________________________________________________<br />
Obviously, the mixed problem makes sense since there exist an infinite number of filters F (a parametrized family)<br />
such that ( F)<br />
≤ γ<br />
Tε , w<br />
2<br />
∞
OBSERVATIONS<br />
The gain F3 of the predictor ensuring the norm less than γ can be<br />
-1<br />
parametrized in the following way: F3=W1 W2, where (W1,W2) satisfyies an elliptic equation.<br />
Convex Optimization<br />
min tr(XW) ≥ min T ( W W )<br />
ε w<br />
1 1 −<br />
2<br />
2<br />
2<br />
⎡W1<br />
W2<br />
⎤<br />
, W = ⎢ ⎥<br />
⎣W2<br />
' W3<br />
⎦<br />
where X depends from the data. If w 1=w 2, then one obtains the<br />
central predictor. In this way one simply minimizes an upper bound<br />
of the real cost. The solution may be too conservative.<br />
Search for boundary values<br />
The boundary values are (norm equal to γ) on the tangent sets between<br />
the ellipsoid and the planes passing through the origin.<br />
− γ2<br />
W<br />
2<br />
γ2<br />
W<br />
1
a - PROCEDURE<br />
We can parametrize with a scalar variable a family of predictor<br />
(defined by the only gain K) which ensure the norm less than γ with<br />
minimal two norm.<br />
Idea<br />
||T εw(K)|| 2<br />
α<br />
γ ||T εw(K)|| ∞<br />
K( α) = ( 1−<br />
α) K + α AQ( α) C' + BD' ( CQ( α)<br />
C' + DD'<br />
)<br />
c<br />
2 −1<br />
Q = AQA' + BB' −( 2α<br />
− α )( AQC' + BD' )( CQC' + DD' ) ( AQC' + BD'<br />
)'<br />
A = A − KcC, B = B − KcD where Kc is such that ||Tεw(Kc)|| ∞
⎧ x(<br />
t + 1)<br />
= Ax(<br />
t)<br />
+ Bw(<br />
t)<br />
⎪<br />
S : ⎨ z(<br />
t)<br />
= Lx(<br />
t)<br />
⎪<br />
⎩ y(<br />
t)<br />
= Cx(<br />
t)<br />
+ Dw(<br />
t)<br />
Finite horizon problem in 1,T :<br />
FINITE HORIZON<br />
Find a linear causal filter based on su y( 1),<br />
y(<br />
2),<br />
L , y(<br />
t)<br />
, t ≤ T<br />
providing an estimate z ˆ( t)<br />
guaranteeing a level of attenuation γ>0,<br />
i.e. such that J F 0 is a weighting matrix (it playes the role played by the<br />
initial covariance matrix in the Kalman filter design). Putting the two equations together we get the difference<br />
Riccati equation: Q( t+ ) = A Q( t) −<br />
−<br />
1<br />
1<br />
+ C'C− L' L/ 2<br />
1<br />
−1<br />
γ + BB'<br />
, with Q( 1)<br />
= AΨ A' + BB'<br />
.<br />
.
zˆ<br />
( t)<br />
= Lη(<br />
t)<br />
SOLUTION OF THE FINITE HORIZON<br />
PROBLEM<br />
η(<br />
t + 1)<br />
= Aη(<br />
t)<br />
+ K ( t + 1)<br />
The gain is given by:<br />
K( t ) = Q( t ) C' I + CQ( t ) C'<br />
[ y(<br />
t + 1)<br />
− CAη(<br />
t)<br />
]<br />
−1<br />
where Q( t ) is the solution of the difference Riccati equation<br />
QUESTIONS<br />
Under which conditions it is possible to guarantee the existence of the<br />
time-varying filter for an arbitrarily long interval?<br />
Under which conditions it is possible to guarantee the convergence for<br />
t→∞ to the H∞ filter?<br />
____________________________________________________________<br />
The time-varying filter can be written as<br />
ξ ( t + 1)<br />
= Aξ<br />
( t)<br />
+ AK<br />
zˆ<br />
( t)<br />
= Lξ<br />
( t)<br />
+ LK(<br />
t)<br />
( t)<br />
[ y(<br />
t)<br />
− Cξ<br />
( t)<br />
]<br />
[ y(<br />
t)<br />
− Cξ<br />
( t)<br />
]<br />
The gains F1 ( t ) = AK ( t ), F2 ( t ) = LK ( t ) have the same structure as in the stationary case (uncorrelated noises).
1) A invertible<br />
FEASIBILITY AND CONVERGENCE<br />
2) There exists the stabilizing solution Qs of<br />
−1<br />
~ −1<br />
~ −1<br />
Q = A Q + C'<br />
R C A'+<br />
BB<br />
3) The pair (A,B) is reachable<br />
[ ] '<br />
Thanks to these assumptions, there exists the unique (positive<br />
semidefinite) solution of the Lyapunov equation<br />
−1<br />
~ −1<br />
~ −1<br />
−1<br />
~ −1<br />
~ −1<br />
~ ~ ~ −1<br />
~<br />
[ Qs<br />
+ C'<br />
R C]<br />
A'YA[<br />
Qs<br />
+ C'<br />
R C]<br />
− [ C'<br />
( R + CQ<br />
C')<br />
] −<br />
Y s<br />
= C<br />
Theorem 10<br />
(i) The solution Q(t) of the difference Riccati equation is feasible, ∀<br />
t>0, if, for some ε :<br />
{ [ [ ] ] } 1<br />
~ ~ ~ −1<br />
~<br />
−<br />
Y − C'(<br />
R + CQ<br />
C')<br />
C +<br />
0 < ( 1)<br />
< Q +<br />
I<br />
Q s s ε + +<br />
(ii) The solution Q(t) of the difference Riccati equation is feasible and<br />
converges to the stabilizing solution Qs if<br />
0 < Q( 1)<br />
< Q + Y + ε I<br />
s<br />
−1<br />
____________________________________________________________<br />
It is possible to see that the conditions introduced in the theorem are also necessary in two significant cases: L=0<br />
(γ→∞) and C=0. In the first case Y=0 and therefore one recovers the condition of convergence for the Kalman<br />
filter (Q(1)>0), whereas, in the second case the condition of convergence is equivalent to<br />
S(1) = Q(1) -1 +C'C- L'Lγ -2 >S a ,<br />
where S a is the antistabilizing solution of<br />
S=(AS -1 A'+BB') -1 +C'C-L'Lγ -2
g-SWITCHING<br />
If Q(1) is "too large", nothing can be said on the convergence of<br />
Q(t,γs) towards the stabilizing solution Qs(γ) associated with a given<br />
value of γ=γs.<br />
The convergence condition is: 0 < Q( 1)<br />
< Q ( γ ) + Y( γ ) + ε I .<br />
Theorem 11<br />
(i) The solution Y(γ) of the Lyapunov equation goes to zero for γ→∞<br />
(ii) The stabilizingsolution Qs(γ) of the Riccati equation increases as γ<br />
increases<br />
Thanks to point (i) of the theorem, if Q(1) does not satisfy the<br />
convergence condition, one can find a value of γ=γ 1>γ s such that the<br />
convergence condition is satisfied for γ=γ1. The solution of the Riccati<br />
equation with γ=γ1 converges to Q s(γ1). On the other hand, thanks to<br />
point (ii) ot the theorem, one can say that<br />
Qs(γ1)≤ Qs(γs)< Qs(γs)+(Y(γs)-1+εI)-1<br />
So that ε>0 and t ε>0 are such that<br />
Q(t ε,γ1)< Qs(γs)+(Y(γs)-1+εI)-1..<br />
Finally, imposing<br />
⎧γ<br />
1 , 1 ≤ t < tε<br />
γ ( t)<br />
= ⎨<br />
⎩γ<br />
s , t ≥ tε<br />
the solution Q(t,γ(t)) tends a Qs(γs) as t→∞.<br />
_______________________________________________________<br />
The γ-switching strategy modifies the cost:<br />
2<br />
T −<br />
T<br />
~ r(<br />
t)<br />
rˆ<br />
( t)<br />
⎡<br />
2<br />
⎤<br />
J F = ∑ − ⎢∑<br />
w(<br />
t)<br />
+ ( x(<br />
0)<br />
− xˆ<br />
0 )'Ψ(<br />
x(<br />
0)<br />
− xˆ<br />
2<br />
0 ) ⎥<br />
t = 1 γ<br />
( t)<br />
⎣ t=<br />
0<br />
⎦<br />
s<br />
−1
EXAMPLE<br />
A=0.5, B=[1 0], C=3, D=[0 1], L=2, T=1/12<br />
Q( t + 1) = 0. 25<br />
Q( t )<br />
Feasibility condition<br />
2<br />
γ<br />
Q( t ) <<br />
2<br />
4 − 9γ<br />
Algebraic Riccati equation<br />
Q = 0. 25<br />
Q<br />
Choice: γs=0.66<br />
1<br />
−1 −2<br />
+ 9 − 4γ<br />
1<br />
−1 −2<br />
+ 9 − 4γ<br />
+ 1<br />
+ 1 , Q(<br />
1) = 4<br />
Hence, Q s (γs)=1.5318 is the stabilizing solution<br />
Q( 1)<br />
< 5.4724 feasibility condition<br />
Q(1) 0.6576<br />
Integrating the equation with γ=γs it results Q( t )=4, 4.7167, 9.5394, -<br />
2.2089, 0.6066, ..... and then the feasibility is lost after three steps.<br />
Taking γ1 in such a way that 4
REFERENCES<br />
T.Basar, P.Bernhard, "H∞ -Optimal Control and related minimax design<br />
problems- a dynamic game approach", Birkhauser, 1991.<br />
P.Bolzern, P.Colaneri, G.De Nicolao, "Transient and asymptotic analysis<br />
of discrete-time H∞ filters", European Journal of Control, to appear.<br />
J.C.Geromel, P.L.D.Peres, S.R.Souza, "A convex approach to the mixed<br />
H2/H∞ control problem for discrete-time uncertain systems", SIAM J.<br />
Control and Optimization, Vol. 33, 6, pp. 1816-1833, 1995.<br />
M.Green, K.Glover, D.Limebeer, J.Doyle, "A J-spectral factorization<br />
approach to H∞ control", SIAM J. Control and Optimization, 25,6, pp.<br />
1350-1371.<br />
M.J.Grimble, A. El Sayed, "Solution of the H∞ optimal linear filtering<br />
problem for discrete-time systems", IEEE Trans. ASSP, 38, 7, pp. 1092-<br />
1104, 1990.<br />
W.Haddad, D.Bernstein, D.Mustafa, "Mixed H2/H∞ - regulation and<br />
estimation", Systems and Control Letters, 16, 1991.<br />
B.Hassibi, T.Kailath, "H∞ bounds for recursive-least square algorithms",<br />
33rd Conference on Decision and Control, Lake Buena Vista, pp. 3927-<br />
3938, 1994.<br />
B.Hassibi, T.Kailath, "H∞ optimality of the LMS algorithms", IEEE<br />
Trans. on Automatic Control, 44, 2, pp. 267-280, 1996.<br />
B.Hassibi, A.H.Sayed, T.Kailath, "Linear estimation in Krein spaces -<br />
Part II: Applications", IEEE Trans. AC, 41, pp. 34-49, 1996.<br />
B.Hassibi, A.H.Sayed, T.Kailath, "Linear estimation in Krein spaces -<br />
Part I: Theory", IEEE Trans. AC, 41, pp. 18-33, 1996.<br />
B.Hassibi, A.H.Sayed, T.Kailath, "H∞ Optimality of the LMS<br />
algorithm", IEEE Trans. ASSP, 44, pp. 267-1022, 1996.<br />
D.H.Jacobson, "Optimal stochastic linear systems with exponential<br />
performance criteria and their relation to deterministic games", IEEE<br />
Trans. Automatic Control, AC-18, 2, 1973.<br />
P.Lancaster, L.Rodman, "Algebraic Riccati Equations", Oxford Series<br />
Pub., 1995.<br />
D.J.Limebeer, M.Green, D.Walker, "Discrete-time H∞ control", Proc.<br />
28th Conference on Decision and Control, Tampa (USA), pp. 392-396,<br />
1989.
H.Rotstein, M.Sznaier, M.Idan, "H2/H∞ filtering theory and an<br />
aerospace application", Int. J. Robust and Nonlinbear Control, 6, pp.<br />
347-366, 1996.<br />
J.Speyer, J.Deyst, D.H.Jacobson, "Optimization of stochastic linear<br />
systems with additive measurement and process noise using exponential<br />
performance criteria", IEEE Trans. Automatic Coontrol, AC-19, pp.358-<br />
366, 1974.<br />
J.L.Speyer, C.Fan, R.N.Banavar, "Optimal stochastic estimation with<br />
exponential cost criteria", Proc. Conference on Decision and Control, pp.<br />
2293-2298, 1992.<br />
A.A.Stoorvogel, A.Saberi, B.M.Chen, "The discrete-time H∞ control<br />
problem with measurementt feedback", Int. J. Robust and Nonlinear<br />
Contr., 4, pp. 457-479, 1994.<br />
Y.Theodor, U.Shaked, N.Berman, "Time-domain H∞ identification",<br />
IEEE Trans. on Aut. Control, 41, pp. 1019-1022, 1996.<br />
M.Vidyasagar, "Control Sysem Synthesis: a facorization approach", MIT<br />
Press Series, 1985.<br />
P.Whittle, "Risk Sensitive Optimal Control", New York: Wiley, 1990.<br />
I.Yaesh, U.Shaked, "Game theory approach to optimal linear estimation<br />
in the minimum H∞ norm sense", 28th Conference on Decision and<br />
Control, Tampa, pp. 421-425, 1989.<br />
I.Yaesh, U.Shaked, "A transfer function approach to the problem of<br />
discrete-time systems: H∞-optimal linear control and filtering", IEEE<br />
Trans. on Automatic Control, 36, 11, pp. 1264-1271, 1991.<br />
I.Yaesh, U.Shaked, "Game theory approach to state-estimation of linear<br />
discrete-time processes and its relation to H∞-optimal estimation", Int. J.<br />
on Control, 55, 6, pp. 1443-1452, 1992.