Digital Signal Processing Reference
In-Depth Information
where
d
dt δ
PA δ
PC (
HH ) 1 CP
PC (
HH ) 1 C
P
=
A
δ
P
+ δ
δ
P
and
d
dt P 1
P 1
PP 1
P 1 A
A P 1
P 1 GG P 1
C (
HH ) 1 C
=−
=−
+
As a result
2Tr
GG P 1
d
dt g P
P 1
PP 1
P
P
) =−
(
δ
δ
P
)
2Tr
HH ) 1 CP
P 1
PP 1
C (
(
δ
δ
P
)
2Tr
GG P 1
P 1
PP 1
≤−
(
δ
δ
P
)
which proves that the Riccati flow is contracting in the sense of ( 3.6 ) for the natural
metric of the cone with
is a lower bound on the eigenvalues
of GG , that is assumed to be invertible, and p max is an upper bound on the eigenvalues
of P .
λ = μ/
p max where
μ
The contraction property of the Riccati equation is in fact due to the invariances
enjoyed by the natural metric of the cone. This can be understood the following way:
the equation writes
HH ) 1 CP . The two first
terms neither expand nor contract as it is the differential form of the transformation
γ B (
d
PA +
GG
PC (
dt P
=
AP
+
) which is an isometry for the distance d P + ( n ) . The addi-
P
) = (
I
+ τ
A
)
P
(
I
+ τ
A
d
tion of a positive matrix
Q where Q is a positive matrix is neither expanding
nor contracting in the Euclidian space, but it contracts for the natural metric as g P
tends to dilate distances when P becomes large. Finally
dt P
=
HH ) 1 CP
is a naturally contracting term which is paramount in the theory of the Kalman fil-
ter and observers (correction term). We have the following proposition, refining the
results of [ 12 ] in the case of time-independent coefficients:
d
PC (
dt P
=−
,
,
(
,
)
Proposition 1
Suppose the matrices A
C
G do not depend on the time t ,
A
C
P + (
)
is observable and G is full-rank. There exists a unique solution Q
n
of the
algebraic Riccati equation defined by
Φ(
Q
) =
0 where
Φ
is given by ( 3.3 ) . For any
R
P
+ (
n
)
,letS R be the ball center Q and radius d P + ( n ) (
Q
,
R
)
S R ={
P
P
+ (
n
),
d P + ( n ) (
Q
,
P
)
d P + ( n ) (
Q
,
R
) }
0 be the lowest eigenvalue of G G .
Let M R =
sup
{
P
2 ,
P
S R } <
, and
μ>
d
Let P 1 (
t
),
P 2 (
t
)
be two arbitrary solutions of the Riccati equation
dt P
= Φ(
P
)
initialized in S R . Then P 1 (
t
),
P 2 (
t
)
S R for any t
0 , and the following contraction
result holds
 
Search WWH ::




Custom Search