Journal of Computational and Applied Mathematics 132 (2001) 71–81

www.elsevier.nl/locate/cam

A fast algorithm for subspace state-space system identi&cation

via exploitation of the displacement structure
Nicola Mastronardia; b; ∗; 2 , Daniel Kressnerc;3 , Vasile Simaa; d;4 , Paul Van Doorene;3; 5 ,
Sabine Van Hu2ela;1; 2; 3
a
Department of Electrical Engineering, ESAT-SISTA=COSIC, Katholieke Universiteit Leuven,
Kardinaal Mercierlaan 94, 3001 Heverlee, Belgium
b
Dipartimento di Matematica, Universit)a della Basilicata, via N. Sauro 85, 85100 Potenza, Italy
c
Department of Mathematics, University of Chemnitz, Chemnitz, Germany
d
National Institute for Research & Development in Informatics, Bd. Mare4sal Averescu, Nr. 8–10,
71316 Bucharest 1, Romania
e
Department of Mathematical Engineering, Universit8e Catholique de Louvain,
Avenue Georges Lemaitre 4 B-1348 Louvain-la-Neuve, Belgium
Received 31 January 2000; received in revised form 17 July 2000

Abstract
Two recent approaches (Van Overschee, De Moor, N4SID, Automatica 30 (1) (1994) 75; Verhaegen, Int. J. Control
58(3) (1993) 555) in subspace identi&cation problems require the computation of the R factor of the QR factorization
of a block-Hankel matrix H , which, in general has a huge number of rows. Since the data are perturbed by noise, the
involved matrix H is, in general, full rank. It is well known that, from a theoretical point of view, the R factor of the
QR factorization of H is equivalent to the Cholesky factor of the correlation matrix H T H , apart from a multiplication
by a sign matrix. In Sima (Proceedings Second NICONET Workshop, Paris-Versailles, December 3, 1999, p. 75), a fast
Cholesky factorization of the correlation matrix, exploiting the block-Hankel structure of H , is described. In this paper
we consider a fast algorithm to compute the R factor based on the generalized Schur algorithm. The proposed algorithm
allows to handle the rank–de&cient case.
c 2001 Elsevier Science B.V. All rights reserved.

∗
Corresponding author.
E-mail address: [email protected] (N. Mastronardi).
1
This author is a Senior Research Associate with the F.W.O. (Fund for Scienti&c Research-Flanders).
2
This work is supported by UE Programme “Training and Mobility of Researchers” project (contract
ERBFMRXCT970160) entitled “Advanced Signal Processing of Medical Magnetic Resonance Imaging and Spectroscopy”.
3
This work is supported by the Belgian Programme on Interuniversity Poles of Attraction (IUAP-4=2 & 24), initiated
by the Belgian State, Prime Minister’s OIce for Science, and by a Concerted Research Action (GOA) project of the
Flemish Community, entitled “Mathematical Engineering for Information and Communications Systems Technology”.
4
This work is supported by the European Community BRITE-EURAM III Thematic Networks Programme NICONET
(project BRRT–CT97-5040).
5
This work is partly supported by the National Science Foundation contract CCR-97-96315.

0377-0427/01/$ - see front matter
c 2001 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 7 - 0 4 2 7 ( 0 0 ) 0 0 6 0 0 - 2
72 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81

Keywords: Generalized Schur algorithm; Hankel and block-Hankel matrices; Subspace identi&cation; QR decomposition;
Singular-value decomposition

1. Introduction

Subspace-based system identi&cation has become very popular in the last decade [4,9,15]. The
success of this state-space identi&cation approach is mainly due to the fact that it relies on a simple
matrix decomposition for which reliable numerical algorithms are available. Its major drawback, on
the other hand, is that large “data” matrices are involved and that this may lead to high computing
and storage costs. We now brieOy recall the basic formulation of the problem. Let uk and yk
be the m-dimensional input vector and the l-dimensional output vector, respectively, of the linear
time-invariant state-space model
xk+1 = Axk + Buk + wk ;

yk = Cxk + Duk + vk ;
where xk is the n-dimensional state vector at time k, {wk } and {vk } are state and output disturbances
or noise sequences, and A, B, C and D are unknown real matrices of appropriate dimensions.
For non-sequential data processing, one chooses N 2(m + l)s and constructs the N × 2(m + l)s
T T
matrix H = [U2s; N Y2s; N ], where U2s; N and Y2s; N are block-Hankel matrices de&ned in terms of the
input and output data, respectively,
 
u1 u2 u3 : : : uN
 u2 u3 u4 : : : uN +1 
 
 u3 u4 u5 : : : uN +2 
U2s; N = ;
 .. .. .. .. 
 . . . . 
u2s u2s+1 u2s+2 : : : uN +2s−1
 
y1 y2 y3 : : : yN
 y2 y3 y4 : : : yN +1 
 
 y3 y4 y5 : : : yN +2 
Y2s; N = :
 .. .. .. .. 
 . . . . 
y2s y2s+1 y2s+2 : : : yN +2s−1
Then the R factor of a QR factorization H = QR is used for data compression [11,14,16]. In [12], a
fast Cholesky factorization of the correlation matrix, exploiting the block-Hankel structure of H , is
described. In this paper we consider a fast algorithm to compute the R factor of the QR factorization
of H based on the generalized Schur algorithm, exploiting its displacement structure [6]. Fast parallel
algorithms for QR factorization are described, e.g., in [3].
The paper does not deal with sequential or recursive processing of the input–output data; a detailed
treatment of a recursive on-line identi&cation technique, and comparisons with other techniques are
included, e.g., in [2]. However, the software we developed for the fast QR approach can also cope
with sequential data processing.
 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81 73

The paper is organized as follows. In Section 2 the generalized Schur algorithm to compute the
Cholesky factor of a symmetric positive-de&nite matrix is described and in Section 3 this algorithm
is applied to the matrix H . The rank-de&cient case is described in Section 4 and some numerical
experiments are reported in Section 5.

2. The Schur algorithm for positive-semi-denite matrices

2.1. The Cholesky factor and generator of A

We summarize here the key properties of the generalized Schur algorithm to compute the Cholesky
factor of a (symmetric) positive-semi-de&nite (psd) matrix, which will be used in the next section.
More details can be found in [7,8]. Let A be a psd matrix of order n, then we de&ne its displacement
A = A − Z T AZ;
using a generalized shift matrix Z of the same dimension. Here we only require the shift matrix Z
to be strictly upper triangular (and hence nilpotent) and we specialize later on to a particular choice
of Z. We call the rank of A the displacement rank  of A and we assume that it is signi&cantly
·
smaller than n. Let the symmetric matrix A have p positive eigenvalues and q =  − p negative
eigenvalues then it has a factorization
   
· · I 0 · Gp
A = G T G;  = p ; G= : (1)
0 −Iq Gq
The matrix G is called the generator of A and since Z is nilpotent one can reconstruct A via the
formula
n−1
A= (Z i )T G T GZ i :
i=0

The generator and intermediate results derived from transformations of the generator, allow to re-
construct the Cholesky factor of the psd matrix A
 
r1; 1 r1; 2 : : : rn; n
 r2; 2 : : : r2; n 
 
A = RT · R; R =  . . ..  :
 . . 
rn; n
Notice that if A has rank r ¡ n then so will the factor R which will have its last n − r rows equal
to zero:  
r1; 1 : : : : : : rn; n
A = RT · R; R = 
 .. ..  :
. . 
rr; r : :rr; n
If the leading r × r principal submatrix of A is nonsingular then ri; i ; i = 1; : : : ; r are all nonzero.
Otherwise the “pro&le” of the trapezoidal matrix R indents to the right each time the nullity of the
i × i principal submatrix of A increases. In our application, A is a product of the type H T H which
clearly is a psd matrix. Its Cholesky factor R is, up to a sign matrix D = diag(±1; : : : ; ±1), also
the RH factor of the QH RH decomposition of H : RH = DR. Hence, both the problem of computing
74 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81

the RH factor of the QR factorization of H and that of computing the Cholesky factor of H T H
are equivalent in exact arithmetic. We discuss now the computation of the Cholesky factor R of A
starting from the generator of A.
One easily shows that the generator G is not unique. We say that the generator G̃ is proper if
its &rst column is zero except possibly its leading element. The following theorem holds for proper
generators [7].

Theorem 1. Let
 
a a
A = 11 12
a21 A22
be a positive-semi-deBnite matrix with proper generator
 
g̃1; 1 g̃1; 2 · · · g̃1; n
 
  G̃ 1
 0 g̃2; 2 · · · g̃2; n  ·
G̃ = 
 .. .. .. 
=

 ;
 . . . G̃ 2
0 g̃; 2 · · · g̃; n
then G̃ 1 is the Brst row of the Cholesky factor R of A. Furthermore; the generator matrix for the
T
Schur complement  = A − G̃ 1 G̃ 1 is given by Ĝ; where
 
0 0 G̃ 1 Z
 = ; Ĝ =  :
0 A22 − a21 a−1
11 a12 G̃ 2

We observe that the &rst column of Ĝ is zero, which needs to be the case since the &rst column
and row of  are zero.
The generalized Schur algorithm just consists of a recursive use of this Theorem: via a transfor-
mation $ (de&ned below) the generator G of the current matrix A is put in proper form G̃: This
yields the current row of the Cholesky factor and the generator of the Schur complement is trivially
obtained from a shift Z applied to the &rst row of the generator. We refer to [7] for more details.
The complexity of this algorithm is that of the transformation $ since the shift Z does not imply
any operations. In the next section we describe brieOy the construction of $.

2.2. Reduction of the generator to proper form

The &rst row of the Cholesky factor R of A is thus obtained from a proper generator G of A.
Reducing a non-proper generator of A to a proper generator G̃; is obtained by applying a transfor-
mation $ to the generator G. In order not to change the product G T G it suIces to choose $ to
be -unitary, i.e.,
$T $ = ;
since then
T
G̃ G̃ = ($G)T ($G) = G T G:
 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81 75

Typically, the matrix $ is constructed as follows:

 
% &
 Ip−1  Hp
:  
$= : : (2)
& %  Hq
Iq−1

The blocks Hp and Hq of the second factor are p×p and q×q Householder transformations reducing
the &rst column of G as follows:
 
x11 x12
 0 
 
 . 
 .. X 
 22 
Hp Gp  0 
 
= : (3)
Hq Gq  y11 y12 
 
 0 
 
 .. 
 . Y  22
0
:
The &rst factor only transforms the rows containing x11 and y11 and eliminates y11 provided ( =
−y11 =x11 is smaller than one in modulus:
     
% & x x x̃ x̃
: 11 12 = 11 12 :
& % y11 y12 0 ỹ12
: :
This 2 × 2 transformation is constructed from % = 1= 1 − (2 , & = (= 1 − (2 and is called a
hyperbolic rotation since it satis&es %2 − &2 = 1 [1]. Also, note that x̃11 = x11 1 − (2 . When it is
implemented in factored form

1 0 √1 0 1 (
1−(2 ;
( 1 − (2 0 1 01

one shows that the generalized Schur algorithm is backward stable and that it has the same complexity
as the unfactored implementation [13]. Hyperbolic Householder transforms [10] may also be used.
2 2
It follows already from (3), that the (1; 1) element of the psd matrix A equals a11 = x11 − y11 ¿0.
Therefore, if a11 = 0 the above transformation can be performed. On the other hand, if a11 = 0 then
the whole row a12 must be zero since otherwise A would not be psd. Since a12 = x11 x12 − y11 y12
this also implies that [x11 ; x12 ] = ±[y11 ; y12 ] and that both these rows can just be deleted from the
generator [5]. In other words, if a11 = 0 a simpli&cation can be performed to the current generator
G̃. The complexity of the reduction of G to proper form G̃ is essentially that of the Householder
transformations Hp and Hq which costs 4(p + q)n Oops. If r = rank A steps are performed, this
algorithm thus requires a total of 4r(p + q)n Oops. We point out that this is an overestimate since
the number of nonzero columns n of the generator decreases at each step and that potentially the
number of rows p + q may decrease as well.
76 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81

3. Fast computation of the R factor of the QR factorization of H

We show here how to compute the generator G of A = H T H where H ∈ RN; 2(m+l)s is the
block-Hankel matrix described in the &rst section with blocks of sizes 1 × m and 1 × l:
 T T T 
u1 u2 : : : u2s y1T y2T : : : y2s T

 T .. .. .. . . .. 
 u2 . .
 . y2T . . . . . 
:
H = . .. .. . . . . .. 
 .. . . . . . . . . . . . 
T T T T
uN : : : : : : uN +2s−1 yN : : : : : : yN +2s−1
The shift matrices used in this context are the matrices
   
0 m Im 0 l Il
 .   . 
: 
 0m . . 
 : 
 0l . .  :
; Z =
Zm =  .  ; Zl =  .  Zm ⊕ Zl : (4)
 .. I   .. I 
m l
0m 0l
The following theorem then gives a construction of a generator for A.

Theorem 2. Given the QR factorization of the Brst block columns:

   
u1T y1T q1T
 u2T y2  
T  T 
  q2 
 . ..  =  .  R1 ; (5)
 .. .   .. 
uNT yNT qNT
where R1 can be assumed upper trapezoidal of row rank k6m + l and qi ∈ Rk ; deBne the product
[ Cu; 1 : : : Cu; 2s Cy; 1 : : : Cy; 2s ] = [ q1 : : : qN ]H: (6)
T
Then a generator G for H H is given by
G = [ Gu Gy ];  = Ik+1 ⊕ −Ik+1 ; (7)
where
   
Cu; 1 Cu; 2 : : : Cu; 2s Cy; 1 Cy; 2 : : : Cy; 2s
 0 uNT +1 : : : uNT +2s−1   0 yNT +1 : : : yNT +2s−1 
   
Gu =  ; Gy =  :
 0 Cu; 2 : : : Cu; 2s   0 Cy; 2 : : : Cy; 2s 
0 u1T : : : u2s−1
T
0 y1T : : : y2s−1
T

Proof. In order to prove the result we consider the displacement matrix H T H :

 T T T T T T

U2s; N U2s; N − Zm U2s; N U2s; N Zm U2s; N Y2s; N − Zm U2s; N Y2s; N Zl
 ; (8)
T T T T T T
Y2s; N U2s; N − Zl Y2s; N U2s; N Zm Y2s; N Y2s; N − Zl Y2s; N Y2s; N Zl

which ought to be equal to

 
GuT Gu GuT Gy
 : (9)
GyT Gu GyT Gy
 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81 77

It follows from (5; 6) that R1 = [Cu; 1 Cy; 1 ] and, hence,

[Cu; 1 Cy; 1 ]T [ Cu; 1 : : : Cu; 2s Cy; 1 : : : Cy; 2s ] = RT1 [ q1 : : : qN ]H;
which are the &rst block rows of the sub-blocks of (8). This thus veri&es the &rst block rows and
block columns of the equality between (8) and (9). The rest easily follows from the block-Hankel
structure of H .

Note that if the &rst block columns of H in (5) have full rank then R1 is square invertible
and k = m + l. If moreover the whole matrix H has full column rank, then the generalized Schur
algorithm will not encounter any singularities. But since the low rank case is of particular interest
here, singularities in the generalized Schur algorithm will be encountered and hence lead to a lower
complexity of the algorithm.
The above theorem also shows that the displacement rank of H T H is at most 2(k +1)62(m+l+1),
with the same number of positive and negative generators. Hence, the generalized Schur algorithm
to compute the R factor requires about (8Nrk) Oops. To compute the generator G of H T H , the QR
factorization (5) requires (6N (m + l)2 ) Oops and product (6) requires less than (4Nk(m + l)s) Oops.
We recall that k6(m + l) and r62(m + l)s but that equality is obtained when no rank de&ciency
is detected. The most time-consuming steps are then clearly the generalized Schur algorithm and
product (6).

4. The generalized Schur algorithm for rank-decient matrices

Our description of the generalized Schur algorithm allows to handle rank-de&cient matrices H T H .
In this case, we can drop some rows of the generator during the algorithm. For this, we need a
:
tolerance, say ) = *||H T H || where * is the requested relative accuracy. Referring to the description
2 2
of Section 2, we test if x11 − y11 6): We then check as well if the leading row a12 of the current
Schur complement is small. If so, the currently computed row of the Cholesky factor is neglectable
and we delete the two corresponding rows of the generator. It is possible that a12 is much larger
than ) although a11 6). In this case, the deletion of a row of the Cholesky factor will yield residual
errors ||H T H − RT R|| of the same size. This is analyzed in this section. From the &rst example, we
can conclude that the described procedure works accurately when it is applied to a matrix H with
a suEciently large gap between signiBcant singular values and negligible ones. On the other hand,
a loss of accuracy in the computed factor R is observed when the distribution of the small singular
values of H shows a uniform and slow decrease. The relative accuracy * is chosen equal to 10−13
in both examples.

Example 1. Consider the matrix H = [U T |Y T ]; with Y = U , where the &rst row and the last column
of U are
[ 40 39 38 · · · 3 2 1 2 2 3 ];

[ 3 2 2 1 2 3 4 5 6 7 ]T ;
respectively. The rank of the matrix H is 5 and ||H T H ||1 = 6:31 × 105 .
78 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81

Table 1
Numerical results for Example 1

RM RS Backward error RM Backward error RS Numerical rank

no. of Oops no. of Oops
31660 9489 1:51 × 10−16 5:20 × 10−15 5

Fig. 1. Distribution of the singular values, in logarithmic scale, of the matrix considered in Example 1.

In Table 1, the results of the computation of the R factor of the matrix H by means of the
standard QR and the generalized Schur algorithm are shown. We denote by RM ; RS , backward error
R∗ , and numerical rank, the R factor of the QR factorization of H computed by the matlab function
triu(qr(H)) and by the generalized Schur algorithm, the backward error of H T H de&ned as
||H T H − RT∗ R∗ ||1
;
||H T H ||1
and the rank of H detected by the generalized Schur algorithm, respectively. In this case, the R factor
is accurately computed by the generalized Schur algorithm, because of the big di2erence between
the signi&cant singular values and the negligible ones of H (see Fig. 1).

Example 2. This is the fourth application considered in the next section. In Fig. 2, we can see that
the distribution of the small singular values of the involved matrix H slightly decreases. We point
out that the correlation matrix H T H computed by matlab is not numerically s.p.d. because of the
nearly rank de&ciency of H . Furthermore, ||H T H ||1 = 3:99 × 104 . So, in this case the fast Cholesky
factorization, exploiting the block-Hankel structure of H and described in [12], can not be used. In
 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81 79

Fig. 2. Distribution of the singular values, in logarithmic scale, of the matrix considered in Example 2.

Table 2
Numerical results for Example 2

RM RS Backward error RM Backward error RS Numerical rank

no. of Oops no. of Oops
12:49 × 106 40:61 × 104 1:27 × 10−14 2:92 × 10−2 18

Table 2, we can see that, although the generalized Schur algorithm is very fast w.r.t. the standard
QR algorithm, the achieved accuracy is not satisfactory.

5. Numerical results

In this section results computing the R matrix by means of the generalized Schur algorithm are
summarized. The data sets considered are publicly available on the DAISY web site
http://www.esat.kuleuven.ac.be/sista/daisy:
At each iteration of the generalized Schur algorithm, two Householder matrices and one modi&ed
hyperbolic rotation are computed in order to reduce the generator in proper form. All the numerical
results have been obtained on a Sun workstation Ultra 5 using Matlab 5.3.
Table 3 gives a summary description of the applications considered in our comparison, indicating
the number of inputs m, the number of outputs l, the number of block rows s, the total number of
data samples used t and the number of rows of H .
80 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81

Table 3
Summary description of applications

Appl. no. Application m l s N

1 Glass tubes 2 2 20 1361
2 Labo dryer 1 1 15 970
3 Glass oven 3 6 10 1227
4 Mechanical Outter 1 1 20 960
5 Flexible robot arm 1 1 20 984
6 Evaporator 3 3 10 6285
7 CD player arm 2 2 15 2018
8 Ball and beam 1 1 20 960
9 Wall temperature 2 1 20 1640

Table 4
Comparative results for the computation of the R factor

Appl. no. Application RM RS Backward error RS Rel. residual

no. of Oops no. of Oops
1 Glass tubes 6:76 × 107 2:61 × 106 2:20 × 10−15 8:73 × 10−14
2 Labo dryer 7:01 × 106 3:36 × 105 8:30 × 10−15 9:48 × 10−13
3 Glass oven 7:63 × 107 6:38 × 106 3:73 × 10−15 7:91 × 10−12
4 Mechanical Outter 1:25 × 107 4:89 × 103 2:92 × 10−2 0:39 × 100
5 Flexible robot arm 1:25 × 107 4:76 × 105 4:13 × 10−15 3:38 × 10−5
6 Evaporator 1:82 × 108 1:11 × 107 6:26 × 10−15 5:13 × 10−14
7 CD player arm 5:77 × 107 2:59 × 106 5:01 × 10−15 2:08 × 10−8
8 Ball and beam 1:22 × 107 4:67 × 105 7:59 × 10−15 6:79 × 10−13
9 Wall temperature 4:67 × 107 1:64 × 106 2:45 × 10−14 3:76 × 10−12

In Table 4 some results for the computation of the R factor of the QR factorization of H are
presented. Rel. residual denotes
|||RM | − |RS |||1
;
|||RM |||1
where RM and RS have been de&ned in Section 4.
The results in Table 4 are comparable with those described in [12], where the R factor is
obtained considering the Cholesky factorization of the correlation matrix H T H , and exploiting
the block-Hankel structure of H . The analysis of the fourth application is described in Example 2
of the previous section.

6. Conclusions

In this paper the generalized Schur algorithm to compute the R factor of the QR factorization of
block-Hankel matrices, arising in some subspace identi&cation problems, is described.
It is shown that the generalized Schur algorithm is signi&cantly faster than the classical QR factor-
ization. A rank-revealing implementation of the generalized Schur algorithm in case of rank-de&cient
matrices is also discussed. Algorithmic details and numerical results have been presented.
 N. Mastronardi et al. / Journal of Computational and Applied Mathematics 132 (2001) 71–81 81

References

[1] A.W. Bojanczyk, R.P. Brent, P. Van Dooren, F.R. De Hoog, A note on downdating the Cholesky factorization,
SIAM J. Sci. Statist. Comput. 1 (1980) 210–220.
[2] Y.M. Cho, G. Xu, T. Kailath, Fast recursive identi&cation of state space models via exploitation of displacement
structure, Automatica 30 (1) (1994) 45–59.
[3] J. Chun, T. Kailath, H. Lev-Ari, Fast parallel algorithms for QR and triangular factorization, SIAM J. Sci. Statist.
Comput. 8 (6) (1987) 899–913.
[4] B. De Moor, P. Van Overschee, W. Favoreel, Numerical Algorithms for Subspace State-Space System Identi&cation
— An Overview, in: Applied and Computational Control, Signals and Circuits, Vol. 1, BirkhWauser, Boston, 1999
(Chapter 6).
[5] K.A. Gallivan, S. Thirumalai, P. Van Dooren, V. Vermaut, High performance algorithms for Toeplitz and block
Toeplitz matrices, Linear Algebra Appl. 241–243 (1996) 343–388.
[6] T. Kailath, S. Kung, M. Morf, Displacement ranks of matrices and linear equations, J. Math. Anal. Appl. 68 (1979)
395–407.
[7] T. Kailath, A.H. Sayed, Displacement structure: theory and applications, SIAM Review 37 (1995) 297–386.
[8] T. Kailath, Displacement structure and array algorithms, in: T. Kailath, A.H. Sayed (Eds.), Fast Reliable Algorithms
for Matrices with Structure, SIAM, Philadelphia, 1999.
[9] M. Moonen, B. De Moor, L. Vandenberghe, J. Vandewalle, On- and o2-line identi&cation of linear state space
models, Int. J. Control 49 (1989) 219–232.
[10] C.M. Rader, A.O. Steihardt, Hyperbolic Householder transforms, SIAM J. Matrix Anal. Appl. 9 (1988) 269–290.
[11] V. Sima, Subspace-based algorithms for multivariable system identi&cation, Studies in Informatics and Control 5 (4)
(1996) 335–344.
[12] V. Sima, Cholesky or QR factorization for data compression in subspace-based identi&cation? Proceedings Second
NICONET Workshop, Paris–Versailles, December 3, 1999, pp. 75 –80.
[13] M. Stewart, P. Van Dooren, Stability issues in the factorization of structured matrices, SIAM J. Matrix Anal. Appl.
18 (1997) 104–118.
[14] P. Van Overschee, B. De Moor, N4SID: Two subspace algorithms for the identi&cation of combined
deterministic-stochastic systems, Automatica 30 (1) (1994) 75–93.
[15] P. Van Overschee, B. De Moor, Subspace Identi&cation for Linear Systems: Theory, Implementation, Applications,
Kluwer Academic Publishers, Dordrecht, 1996.
[16] M. Verhaegen, Subspace model identi&cation. Part 3: Analysis of the ordinary output-error state-space model
identi&cation algorithm, Int. J. Control 58 (3) (1993) 555–586.