inner outer and spectral factorizations
TRANSCRIPT
Inner-Outer and Spectral Factorizations
SOLO HERMELIN
Updated: 27.05.11
1
Table of Content
SOLO Inner-Outer and Spectral Factorizations
2
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
3
Given the Linear Time-Invariant System 00 xxuBxAx
and the Quadratic Cost Function:
TT
T
TTc PPRRdt
u
x
RS
SPuxJ
&02
1
0
The Optimal Regulator that Minimizes the Cost Function Jc is given by the following procedure:Define the Hamiltonian of the Optimization Problem:
uBxAu
x
RS
SPuxuxH T
T
TT
2
1,,
The Euler-Lagrange Equations are:
TTTT
T
T
T
BxSRuBxSuRu
H
uSxPAx
H
td
d
uBxAH
td
xd
10
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
4
TTTT
T
T
T
BxSRuBxSuRu
H
uSxPAx
H
td
d
uBxAH
td
xd
10
B
A
s
In
Kc
xu
or:
xA
x
SRBASRSP
BRBSRBAxHTTT
TT
11
11
TTT
TT
HSRBASRSP
BRBSRBAA
11
11
:
We want to find X (t) such that: , then txtXt
XBSRKxKxXBSRu TTcc
TT 11 :
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
5
xA
x
SRBASRSP
BRBSRBAxHTTT
TT
11
11
We want to find X (t) such that: , then txtXt
txtXtxtXt
xXBRBxSRBAXxXxXSRBAxSRSP TTTTT 1111
txxSRSPXBRBXSRBAXXSRBAX TTTTT 01111
TTTTT SRSPXBRBXSRBAXXSRBAX 1111 We obtain the following Differential Riccati Equation for :X
0
0
1
11
11
1111
PXBSRXBSAXXA
X
IAIX
X
I
SRBASRSP
BRBSRBAIX
SRSPXBRBXSRBAXXSRBA
TTTTTT
HTTT
TT
TTTTT
If a Steady-state Solution exists for t→∞ then and:0X
Continuous Algebraic
RiccatiEquation(CARE)
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
B
A
s
In
Kc
xu
Theorem 1 A Unique Stabilizing( A+BKc=[A-BR-1(ST+BTX)] is Stable) Solution of CARE, X=XT, is obtained iff:a.(A,B) is Stabilizableb.Re[λi(AH)]≠0 for all i=1,2,…,2nThe Unique Stabilizing Solution is denoted X=Ric [AH] Proof
a. It is clear that the condition that (A,B) is Stabilizable is a necessary condition to stabilize A+BKc.
b. Rewrite
IX
I
XBSRBA
BRBXBSRBA
IX
I
SRBASRSP
BRBSRBAA
TTT
TTT
TTT
TT
H
0
0
0:
1
11
11
11
If λi is an eigenvalue of AH, - λi is also; hence AH has n eigenvalues λi s.t. Re[λi] ≤ 0 and n eigenvalues λj s.t. Re[λj] ≥ 0. If AH has no eigenvalue of jω axis, then we can find a solution X s.t. all the eigenvalues of A-BR-1(ST+BTX)=A+BKc are the stable eigenvalues of AH.
011
1 PXBSRXBSAXXA TTTTTT
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
B
A
s
In
Kc
xuTheorem 1 A Unique Stabilizing( A+BKc=[A-BR-1(ST+BTX)] is Stable) Solution of CARE is obtained iff:a.(A,B) is Stabilizableb.Re[λi(AH)]≠0 for all i=1,2,…,2nThe Unique Stabilizing Solution is denoted X=Ric [AH]
Proof (continue -1)c. Assume that there are two Stabilizing Solutions X1=X1
T and X2 =X2T that satisfy the
CARE 0
0
21
222
11
111
PXBSRXBSAXXA
PXBSRXBSAXXA
TTTTTT
TTTTTT
Subtracting those two equations we obtain:
0
0
11
211
2
21
211
11
21211
2121
XBRBXXBRBX
XBRBXXBRBXSRBXXXXBRSAXXXXATT
TTTTT
or: 011
212121 XBSRBAXXXXXBSRBA TTTTT
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
B
A
s
In
Kc
xuTheorem 1 A Unique Stabilizing( A+BKc=[A-BR-1(ST+BTX)] is Stable) Solution of CARE is obtained iff:a.(A,B) is Stabilizableb.Re[λi(AH)]≠0 for all i=1,2,…,2nThe Unique Stabilizing Solution is denoted X=Ric [AH]
Proof (continue -2)c. Assume that there are two Stabilizing Solutions X1=X1
T and X2 =X2T that satisfy the
CARE
This is a Sylvester Matrix Equation and since both X1 and X2 are Stabilizing Solutions we must have:
011
212121 XBSRBAXXXXXBSRBA TTTTT
jiXBSRBAXBSRBA
njXBSRBA
niXBSRBATT
jTT
iTTj
TTi
,0ReRe,,10Re
,,10Re2
11
1
21
11
Therefore the solution of the Sylvester Matrix Equation is Unique, and by substitution we can see that X1 = X2 is the Unique Solution.
q.e.d.
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
Proof (1)
(1) The Matrix YTZ is Symmetric.(2) If Y-1 exists, the X=ZY-1 is the Solution of CARE such that the Matrix ((A-BR-1ST)-BR-1BT X) has the Eigenvalues λ1,λ2,…,λn. X is Symmetric.(3) If (A,B) Stabilizable and Re[λi(AH)]≠0 for all i=1,2,…,2n (Theorem 1) Y-1 exists.
Theorem 2 Let the columns of the Matrix be the Eigenvectors
of AH corresponding to the Stable Eigenvalues λ1,λ2,…,λn, and J the corresponding Jordan Matrix. Then
nxnnxn RZYRZ
Y
,2
Define TTT SRSPQBRBWSRBAE 111 :,:,:
TTTT
TT
HEQ
WE
SRBASRSP
BRBSRBAA
11
11
:
JZZEQY
JYZWYEJ
Z
Y
Z
Y
EQ
WEJ
Z
Y
Z
YA
TTHWe have
ZYJZWZZEYZYJWZEYJYZWYE TTTTTTTTTTT
JZYQYYZEYJZYZEYQYYJZZEQYY TTTTTTTTTT
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
Proof (1) (continue – 1)
Since –YTQTY-ZTWZ is symmetric so is (YTZ)J+JT(YTZ) or
ZYJJZYZWZQYYZWZZYJJZYQYYJZYQYYZEY
ZYJZWZZEY TTTTTTTTTT
TTTT
TTTTT
TTTTTTTTTTT ZYJJZYZYJJZYZYJJZY
Because J represents the stable Eigenvalues Re[λi(J)+λj(J)]≠0 for all I,j=1,…,n, the Unique Solution of the Sylvester Matrix Equation is
0TTTTTTT ZYZYJJZYZY Sylvester Matrix Equation
SymmetricisZYZYZYZYZY TTTTTTT 0
(1) The Matrix YTZ is Symmetric.(2) If Y-1 exists, the X=ZY-1 is the Solution of CARE such that the Matrix ((A-BR-1ST)-BR-1BT X) has the Eigenvalues λ1,λ2,…,λn. X is Symmetric.(3) If (A,B) Stabilizable and Re[λi(AH)]≠0 for all i=1,2,…,2n (Theorem 1) Y-1 exists.
Theorem 2 Let the columns of the Matrix be the Eigenvectors
of AH corresponding to the Stable Eigenvalues λ1,λ2,…,λn, and J the corresponding Jordan Matrix. Then
nxnnxn RZYRZ
Y
,2
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
Proof (2)
JZZEQY
JYZWYET
111 YJYYZWEYJYZWYE1111111 YJZYZWYZEYZYJYYZWEYZ
111 YJZYZEQYJZZEQY TT
01111 QYZWYZEYZYZET
Define TTT SRSPQBRBWSRBAE 111 :,:,:
Therefore X=ZY-1 is the Unique Stabilizing Solution of the CARE
01111 TTTTT SRSPXBRBXSRBAXXSRBA q.e.d.
(1) The Matrix YTZ is Symmetric.(2) If Y-1 exists, the X=ZY-1 is the Solution of CARE such that the Matrix ((A-BR-1ST)-BR-1BT X) has the Eigenvalues λ1,λ2,…,λn. X is Symmetric.(3) If (A,B) Stabilizable and Re[λi(AH)]≠0 for all i=1,2,…,2n (Theorem 1) Y-1 exists.
Theorem 2 Let the columns of the Matrix be the Eigenvectors
of AH corresponding to the Stable Eigenvalues λ1,λ2,…,λn, and J the corresponding Jordan Matrix. Then
nxnnxn RZYRZ
Y
,2
Assume Y-1 exists:
Linear Quadratic Regulator (LQR) Problem
SOLO Inner-Outer and Spectral Factorizations
Proof (2) (continue – 1)
q.e.d.
From (1) YTZ is Symmetric, so
YZZYZY TTTT
TTTTTTTT ZYXZXYZYZYYYZZY 11
1 YZX
TTTT XYZZYX 1 X is Symmetric
(1) The Matrix YTZ is Symmetric.(2) If Y-1 exists, the X=ZY-1 is the Solution of CARE such that the Matrix ((A-BR-1ST)-BR-1BT X) has the Eigenvalues λ1,λ2,…,λn. X is Symmetric.(3) If (A,B) Stabilizable and Re[λi(AH)]≠0 for all i=1,2,…,2n (Theorem 1) Y-1 exists.
Theorem 2 Let the columns of the Matrix be the Eigenvectors
of AH corresponding to the Stable Eigenvalues λ1,λ2,…,λn, and J the corresponding Jordan Matrix. Then
nxnnxn RZYRZ
Y
,2
References
SOLO
13
S. Hermelin, “Robustness and Sensitivity Design of Linear Time-Invariant Systems”,PhD Thesis, Stanford University, 1986
G.Stein, J.Doyle, B.Francis, “Advances in Multivariable Control”, ONR/Honeywell Workshop, 1984
Inner-Outer and Spectral Factorizations
T. Kailath, A.H. Sayed, B. Hassibi, “Linear Estimation”. Prentice Hall, 2000
14
SOLO
TechnionIsraeli Institute of Technology
1964 – 1968 BSc EE1968 – 1971 MSc EE
Israeli Air Force1970 – 1974
RAFAELIsraeli Armament Development Authority
1974 – 2013
Stanford University1983 – 1986 PhD AA
Sylvester Matrix Equation : Anxn Xnxm+Xnxm Bmxm=Cnxm
SOLO
15
Consider the Sylvester Matrix Equationwhere Anxn, Bmxm, Cnxm are given matrices. Then there exists a Unique Solution Xnxm if and only if
Matrices
James Joseph Sylvester(1814 – 1887)
nxmmxmnxmnxmnxn CBXXA
njiAA nxnjnxni ,,1,0ReRe
0
dteCeX tBnxm
tAnxm
mxmnxn
Note : If B=AT : Lyapunov EquationThe Necessary and Sufficient Condition for the existence of a Unique Solution is
nxnT
nxnnxnnxnnxn CAXXA
In particular if λi(A)=- λj(A) = j ω a Unique Solution does not exist.
Aleksandr Mikhailovich Lyapunov
1857 - 1918
mjniBA mxmjnxni ,,1,,,10
the Unique Solution is given byIf mjniBA mxmjnxni ,,1,,,10ReRe
Sylvester Matrix Equation : Anxn Xnxm+Xnxm Bmxm=Cnxm
SOLO Matrices
Proof
Let rewrite
0
21
22221
11211
21
22221
11211
21
22221
11211
21
22221
11211
21
22221
11211
nmnn
m
m
mmmm
m
m
nmnn
m
m
nmnn
m
m
nnnn
n
n
nxmmxmnxmnxmnxn
ccc
ccc
ccc
bbb
bbb
bbb
xxx
xxx
xxx
xxx
xxx
xxx
aaa
aaa
aaa
CBXXA
nmnn
m
m
nnnn
n
n
nxmnxn
xxx
xxx
xxx
aaa
aaa
aaa
XA
21
22221
11211
21
22221
11211
nxmnxnc
mnxn
nxn
nxn
Xvec
m
nm
m
n
n
nxn
nxn
nxn
nxmcnxnm XAvec
cA
cA
cA
c
x
x
c
x
x
c
x
x
A
A
A
XvecAI
nxmc
2
1
1
2
2
12
1
1
11
00
00
00
nxmc Xvec
xmn
m
nm
m
n
n
mcnxmc
c
x
x
c
x
x
c
x
x
cccvecXvec
1
1
2
2
12
1
1
11
21 :
Define the VectorizationOperator vec:
Sylvester Matrix Equation : Anxn Xnxm+Xnxm Bmxm=Cnxm
SOLO Matrices
Proof (continue – 1)
Let rewrite
0
21
22221
11211
21
22221
11211
21
22221
11211
21
22221
11211
21
22221
11211
nmnn
m
m
mmmm
m
m
nmnn
m
m
nmnn
m
m
nnnn
n
n
nxmmxmnxmnxmnxn
ccc
ccc
ccc
bbb
bbb
bbb
xxx
xxx
xxx
xxx
xxx
xxx
aaa
aaa
aaa
CBXXA
mxmnxmc
mmmmm
mm
mm
Xvec
m
nm
m
n
n
nmmnmnm
nmnm
nmnn
nxmcnT
mxm BXvec
cbcbcb
cbcbcb
cbcbcb
c
x
x
c
x
x
c
x
x
IbIbIb
IbIbIb
IbIbIb
XvecIB
nxmc
2211
2222112
1221111
1
2
2
12
1
1
11
21
22212
12111
mmmm
m
m
nmnn
m
m
mxmnxm
bbb
bbb
bbb
xxx
xxx
xxx
BX
21
22221
11211
21
22221
11211
Sylvester Matrix Equation : Anxn Xnxm+Xnxm Bmxm=Cnxm
SOLO Matrices
Proof (continue – 2)
We have
0 nxmcmxmnxmcnxmnxncnxmmxmnxmnxmnxnc CvecBXvecXAvecCBXXAvec
nxmcnxmcnT
mxmnxnm CvecXvecIBAI or
Let use the Jordan decomposition for Anxn and Bmxm11 & TTT BBB
TmxmAAAnxn SJSBSJSA
n
TTT
m
TT
TTT
I
AABBBAAA
I
BB
nBBBAAAmnT
mxmnxnm
SSSJSSJSSS
ISJSSJSIIBAI1111
11
111
111
1111
ATB
T
nTBATB
TT
ATBAm
TT
SS
AB
IJSS
ABB
SSJI
AABAB
DBCADCBA
BABASSSJSSJSSS
1 ABnBAmABnT
mxmnxnm SSIJJISSIBAIT
TT
This equation has a Unique Solution iff is Nonsingular. nTmxmnxnm IBAI
Sylvester Matrix Equation : Anxn Xnxm+Xnxm Bmxm=Cnxm
SOLO Matrices
Proof (continue – 3)
where
1 ABnBAmABnT
mxmnxnm SSIJJISSIBAIT
TT
lJ
J
J
J
00
00
00
2
1
i
i
i
i
i
xkki iiJ
0000
1000
0000
0010
0001
Ji are Upper-Triangular Matrices
nBAm IJJI T Therefore is also Upper Triangular with diagonal elements λi[A] + λj[B](i=1,…,n, j=1,…,m), that are the Eigenvalues of and therefore to the Similar Matrix . This Matrix is Nonsingular iff
nBAm IJJI T
nTmxmnxnm IBAI
mjniBA mxmjnxni ,,1,,,10
Sylvester Matrix Equation : Anxn Xnxm+Xnxm Bmxm=Cnxm
SOLO Matrices
Since 0lim,0lim,,1,,,10ReRe
tB
t
tA
tmxmjnxni eemjniBA
Proof (continue – 4)
Let rewrite
nxm
mSnnxm
nxm
m
nxnnxm
mxmnnxmnxmmxmnxmnxmnxn X
IAIX
X
I
AC
BIXCBXXA
00
nxmnxmtB
nxmtA
nxm CPeCetP mxmnxn 0:Define
By differentiation
mxmnxmnxmnxnmxmtB
nxmtAtB
nxmtA
nxnnxm BPPABeCeeCeAtd
tPdmxmnxnmxmnxn
Let Integrate the Differential Equation
mxmnxmnxmnxn
nxmnxmnxm BdtPdtPAdt
td
tPdPP
000
0
We have and where
0
dteCeX tBnxm
tAnxm
mxmnxn
0nxmP
mxmnxmnxmnxnmxmnxmnxmnxnnxm BXXABdtPdtPAC
00
Matrix Differential Riccati Equation
SOLO Inner-Outer and Spectral Factorizations
tDtXtCtXtBtXtXtAtX
where all matrices are nxn and A (t),B (t), C(t), D(t) are integrable over an interval t0 ≤ t ≤ tf.
This Matrix Differential Equation can be decompose in the following 2 Linear Differential Equations:
The Nonlinear Matrix Differential Riccati Equation is given by
tZtAtYtDtZ
tZtCtYtBtY
If in the interval t0 ≤ t ≤ tf the Matrix Y(t) is nonsingular, then the solution of the Riccati Differential Equation is 10
1 ttttYtZtX
To verify we have tYtd
tYdtYtY
td
dtY
td
dtZtY
td
tZdtX
td
d 11111 &
tXtCtXtBtXtXtAtD
tYtZtCtYtBtYtZtYtZtAtYtDtXtd
d
111
Matrix Differential Riccati Equation
SOLO Inner-Outer and Spectral Factorizations
tDtXtCtXtBtXtXtAtX
If yhe matrices are nxn and A ,B , C, D are constant, then
This Matrix Differential Equation can be decompose in the following 2 Linear Differential Equations:
The Nonlinear Matrix Differential Riccati Equation is given by
tZAtYDtZ
tZCtYBtY
If in the interval t0 ≤ t ≤ tf the Matrix Y(t) is nonsingular, then the solution of the Riccati Differential Equation is 10
1 ttttYtZtX
Define
AD
CBG
tZ
tYtW :&:
to obtain which have the solution tGWtW 00 tWetW ttG