Professional Documents
Culture Documents
QL Factorization in State-Space
QL-Factorization
The QL-factorization of a matrix T is a powerful tool for solving linear systems of equations or linear least squares problems as u = T y . The QL-factorization of a given matrix T amounts to computing T = QL, QT Q = 1n , L: lower triangular.
Note that L is a lower-triangular matrix corresponding with a causal system (in contrast with conventional QR decomposition, where R is supposed to be an upper triangular matrix. We assume that T is a tall matrix (i.e. T Rmn , m n) and that it has full column rank, i.e. det T T T = 0. Having determined the QL-factorization we can easily determine the inverse or the pseudo-inverse of T as T = L1 QT , L is square, det L = 0
Using the time-varying state-space methodology we are interested in calculating the QL-factorization of a given matrix T in terms of its state-space realizations directly. The procedure for computing the QL factorization then goes as follows: 1. Determine a state-space realization for T 2. Compute the QL factorization in state-space 3. Invert the factors in state-space 4. Arrive at T = L1 QT Note that for solving least squares problems it is not necessary to actually compute the product L1 QT , it is more economic to keep the factorized form an apply the matrices QT and L1 subsequently to the vector y and QT y , respectively.
Lecture 7
this computational process may require us to compute the corresponding factorizations and the necessary inversions, which turn out to be computationally expensive. Alternatively, we can determine realizations directly by reading o the matrix entries from T and creating a simple computational state-space structure for T as a starting point. Lets have a look at a simple example. Example 1 We consider the simple example of a 4 2 matrix T as shown in Figure 1. We can easily check that the computational structure shown in the gure realizes the computation for T u = y . The state-space realization matrices for this structure can also be easily read o as A1 C1 and A4 C4 B4 D4 = tT 4 . B1 D1 = 1 tT 1 , A2 C2 B2 D2 = 1 tT 2 , A3 C3 B3 D3 = 1 tT 3
tT 1
y1
tT k
T =
[]
tT 2
y2
[]
tT 3
y3
y = Tu
[]
tT 4
y4
Example 2 Yet another slightly more complicated example is shown in Figure 2. For this example we partition the u1 to give us input vector as u2 y=T u1 u2 ,
Lecture 7
where we have the corresponding partitioning of the matrix T given as T = T (1) T (2) .
For each of these two matrices we assume to have state-space realizations T (1) Ak (1) Ck
(1)
Bk (1) Dk
(1)
T (2)
Ak (2) Ck
(2)
Bk (2) Dk
(2)
which can be combined into one realization matrix as (1) (1) Ak 0 Bk 0 (2) (2) Tk = 0 Ak 0 Bk . (1) (1) (2) (2) Ck Ck Dk Dk
Using this generic partitioning of the realization matrix along with the realizations determined in the previous section we can easily read of a time-varying realization to consist of the matrices 1 0 0 0 1 0 T1 = 0 0 , T2 = 0 0 , tT tT 1 2 1 0 0 0 0 T3 = 0 0 1 , T4 = 0 1 0 , tT tT tT tT 3 1 4 2 0 0 0 0 T5 = 0 1 0 , T6 = 0 0 . tT tT 3 4
Determine QL-Factorization
We start out with a time-varying state-space realization for T , such that we can represent the coecient matrix as T = D + C (1n ZA)1 ZB , = A C B D .
Since T is the product of two matrices Q and L we aim at state-space realizations for them, that is we represent both matrices in terms of state-space models Q = D Q + C Q (1n ZAQ )1 ZB Q , Q = AQ CQ A CL BQ DQ B DL .
L = D L + C L (1n ZAL )1 ZB L ,
L =
Figure 3 shows the computational structure we aim for, that is, this structure represents the QLfactorization of the matrix T . To this end we need to determine the realizations for Q and for L. This represents the completion of step 2 in our operational procedure, where we represent the matrix T as the product N Q 2Q 1L N L 2L 1, T =Q
Lecture 7
u1
[]
y1
y2
u2
[]
y3
T =
y4
y=T
u1 u2
[] []
y5
y6
Figure 2: Direct State-Space Realization for the matrix T (taken from [5])
Lecture 7
N . . . Q 2Q 1L N . . . L 2L 1 T =Q
u1 L1 Q1 y1
u2
L2
Q2
y2
u3
L3
Q3
y3
u4
L4
Q4
y4
u5
L5
Q5
y5
Lecture 7
Q Bk
Q Dk
k = L CL k
AL k
L Bk
1 .. . 1
L Dk
k provides for an embedding of the k -th realization matrix in the form where the factor T Ak Bk 1 . . . k = T . 1 Ck Dk 1 Example We can check this factorized representation for T looking at a small example for N = 3. For the lower 4 4 semi-separable matrix we have D1 C 2 B1 D2 . T = C3 A2 B1 C3 B2 D3 C4 A3 A2 B1 C4 A3 B2 C4 B3 D4 We check the factored form by directly calculating 4 T 3 T 2 T 1 = T =T A3 . . B3 . A2 A4 . . . B 4 . 1 . 1 . . C2 = 1 1 . . . 1 C3 D3 C4 D4 . 1 . A4 A3 . . A4 B3 B4 A2 . B2 . . 1 . 1 C2 = . 1 D2 C3 . D3 1 C 4 A3 C4 B3 D4 . A4 A3 A2 . A 4 A3 B 2 A4 B 3 B 4 A1 B 1 . C1 D 1 . 1 . = C2 D2 1 C 3 A2 . C3 B2 D3 C4 A2 A2 C4 A3 B2 C4 B3 D4 .
We used in the previous section the option to represent a matrix T as the product N T N 1 T N 2 T 1 , T =T
. 1
B2 D2
1 A1 C1 . . 1 . . . = 1 1 . 1
A1 C1 . . . . 1 1 .
B1 D1
. 1
1 . 1
B1 D1
1 D1 C2 B1 C 3 A2 B 1 C4 A3 A2 B1
D3 C4 B3
D2 C3 B2 C4 A3 B2
D4
where we have used that A1 = [], C1 = [], A4 = [] and B4 = [], producing a zero-dimensional rst block-column and rst block-row.
Lecture 7
Inversion of Factors in State-Space In the next step we take the individual stages of the structure and determine the inverse realization by local inversion. Once we have computed the state-space realizations for Q and L we can easily determine the state-space realizations of the inverse systems via S k =
L L 1 L AL Ck k Bk (Dk ) L 1 L (Dk ) Ck L L 1 Bk (Dk ) L 1 (Dk )
AS k S Ck
S Bk S Dk
where the realizations Q k represent an anti-causal system. Anti-causal realization corresponds to a backward recursion xk yk = =
Q T T (A Q k ) xk+1 + (Ck ) uk Q T Q T (Bk ) xk+1 + (Dk ) uk
k = N, N 1, . . . 1 Concatenating the inverse realizations produces the structure shown in Figure 4, which is a state-space realization for the inverse matrix T 1 given in terms of the factors QT and L1 , or more accurately, in terms of the product N S 2 S 1 Q T Q T Q T T = S 1 2 1. Note in Figure 4 how the anti-causality of the realization for QT creates an upward ow of the state signals.
Yk L Ck
0 L Dk
(1)
R The matrices Yk and Ak have the same number of columns, Dk and Dk have also the same number R R of columns, while Dk has full row rank. The matrices Yk or Dk may be zero-dimensional ([]) and hence the matrix entry 0 may also vanish. The recursion starts out with Yk+1 = [] and continues for k = N, N 1, . . . , 1. For practical purposes we rewrite Equation 1 by bringing the factor Q to the left side to arrive at
AQ k Q Ck
Q Bk Q Dk
Yk+1 Ak Ck
Yk+1 Bk Dk
Yk L Ck
0 L Dk
Note that this amounts to applying an appropriately chosen sequence of Givens rotations from the left with the purpose to eliminate the 12-block and to generate the lower triangular shape on the right-hand side of the equation. This is very similar to the conventional algorithm for computing the QR factorization
Lecture 7
N . . . S 2 S 1 Q T T T T = S 1 Q2 . . . QN
y1 QT 1 S1 u1
y2
QT 2
S2
u2
y3
QT 3
S3
u3
y4
QT 4
S4
u4
y5
QT 5
S5
u5
Lecture 7
(as shown in [2]) except that we create a lower triangular matrix instead of an upper triangular. This requires a slight change in the elimination sequence. As a result of performing this recursive computation scheme for all values of k we arrive at a realization matrix for the lower matrix L L k = Ak L Ck Bk L Dk .
Details of the QL Factorization Process Working out the rst steps of the recursive algorithm produces the intermediate matrices Q Q T BN YN +1 BN AN YN +1 AN N 1 T 1 = T T Q 1 1 NT = Q Q CN DN CN DN YN 0 AN 1 B N 1 1 1 .. .. 1 = TN 2 T . . CN 1 1 D N 1 L L 1 CN DN YN AN 1 YN BN 1 1 .. 1 , = TN 2 T . L CN 1 DN 1 L L L CN AN 1 CN BN 1 DN
T where we made us of Equation 1. Pre-multiplication with the next factor Q N 1 produces the intermediate result YN 1 0 AN 2 B N 2 1 1 T T . T .. QN 1 QN T = D N 1 C N 2 N 3 T1 L 1 CL DN N 1 1 L L L 1 C N AN 1 CN BN 1 DN YN 1 AN 2 YN 1 BN 2 1 L T = CN 2 DN 2 N 3 T1 L L L CN 1 C N 1 B N 2 DN 1 L L L L CN AN 1 CN AN 1 BN 2 CN BN 1 DN Continuing the recursive computation will eventually produce Y1 T L C D1 1 L L L C 2 A1 C2 B1 D2 T T Q 1 . . . QN T = . . . . . . . . . L L CN AN 1 . . . A1 CN AN 1 . . . A2 B1 . . . . ... ..
L DN
Lecture 7
10
Literatur
[1] G. Strang. Computational Science and Engineering. Wellesley-Cambridge Press, 2007. [2] G. Golub, Ch. van Loan. Matrix Computations. John Hopkins, 1992. [3] T. Kailath. Linear Systems. Prentice Hall, 1980. [4] P. Dewilde, A.-J. van derVeen. Time-Varying Systems and Computations. Kluwer Academic Publishers, 1998. [5] L. Tong,A.-J. van der Veen, P. Dewilde, Y. Sung. Blind Decorrelating RAKE Receivers for Long-Code WCDMA. IEEE Trans. Signal Processing, Vol.51, No.6, pp. 1642-1655, June 2003.