Automation, Vol.311.No. I. pp. 75-93, 1994
00
Printedin Great Britain.
N4SID*" Subspace Algorithms for the
Identification of Combined Deterministic-
Stochastic Systemst
P E T E R VAN O V E R S C H E E $ and B A R T DE MOOR§
The asymptotic consistency is analyzed of two new subspace algorithms that
identify combined deterministic-stochastic state space models from given
input-output data.
Key Words--System identification: Kalman filters: difference equations: OR and singular value
decomposition; multivariable systems: state space methods.
Abstract--Recently a great deal of attention has been given
to numerical algorithms for subspace state space system
identification (N4SID). in this paper, we derive two new
N4SID algorithms to identify mixed deterministic-stochastic
systems. Both algorithms determine state sequences through
the projection of input and output data. These state
sequences are shown to be outputs of non-steady state
Kalman filter banks. From these it is easy to determine the
state space system matrices. The N4SID algorithms are
always convergent (non-iterative) and numerically stable
since they only make use of QR and Singular Value
Decompositions. Both N4S1D algorithms are similar, but the
second one trades off accuracy for simplicity. These new
algorithms are compared with existing subspace algorithms in
theory and in practice.
1. INTRODUCTION
THE GREATER PART OF the systems identification
literature is concerned with computing polyno-
mial models, which are however known to
typically give rise to numerically ill-conditioned
mathematical problems, especially for Multi-
Input Multi-Output systems. Numerical algo-
rithms for subspace state space system identifica-
tion (N4SID*) are then viewed as the better
* Numerical algorithms for Subspace State Space System
Identification. Read as a Californian license plate: enforce it.
t Received 6 March 1992; revised 2 September 1992:
revised 15 January 1993; received in final form 19 March
1993. The original version of this paper was presented
at the 12th IFAC Congress which was held in Sydney,
Australia, during 19-23 July 1993. The Published Pro-
ceedings of this IFAC Meeting may be ordered from:
Pergamon Press Limited, Headington Hill Hall, Oxford
OX3 0BW, U.K. This paper was recommended for publica-
tion in revised form by the Guest Editors. Corresponding
author: Peter van Overschee. Fax +32-16-221855; E-mail
vanoverschee@esat.kuleuven.ac.be.
ESAT, Department of Electrical Engineering, Katho-
lieke Universiteit Leuven, Kardinaal Mercierlaan 94, 3001
Heverlee, Belgium. Research Assistant of the Belgian
National Fund for Scientific Research.
§ Research Associate of the Belgian National Fund for
Scientific Research.
75
0(105-1098/94 $6.00+0.
(~ 1993PergamonPressLtd
alternatives. This is especially true for high-order
multivariable systems, for which it is not trivial
to find a useful parameterization among all
possible parametrizations. This parametrization
is needed to start up the classical identification
algorithms (see e.g. Ljung, 1987), which means
that a priori knowledge of the order and of the
observability (or controllability) indices is
required.
With N4SID algorithms, most of this a priori
parametrization can be avoided. Only the order
of the system is needed and it can be determined
through inspection of the dominant singular
values of a matrix that is calculated during the
identification. The state space matrices are not
calculated in their canonical forms (with a
minimal number of parameters), but as full state
space matrices in a certain, almost optimally
conditioned basis (this basis is uniquely deter-
mined, so that there is no problem of
identifiability). This implies that the observ-
ability (or controllability) indices do not have to
be known in advance.
Another major advantage is that N4SID
algorithms are non-iterative, with no non-linear
optimization part involved. This is why they do
not suffer from the typical disadvantages of
iterative algorithms, e.g. no guaranteed conver-
gence, local minima of the objective criterion
and sensitivity to initial estimates.
For classical identification, an extra para-
metrization of the initial state is needed when
estimating a state space system from data
measured on a plant with a non-zero initial
condition. A final advantage of the N4SID
algorithms, is that there is no difference between
zero and non-zero initial states.
Most commonly known subspace methods are
76 P. VAN OVERSCHEE and B. DE MOOR

realization algorithms of e.g. Kung (1978), Input-Output

where a discrete-time state space model is
computed from a block Hankel matrix with
data Uk,Yk I Classical
Markov parameters. It is unfortunate that the N4SID Identification
theory here relies on Markov parameters as a
starting point, something rather difficult to 1 1
measure or compute in practice (e.g. think of
unstable systems).
Kalman states ] I SYstemmatrieeI
An alternative direct identification scheme for Least Kalma.n
purely deterministic systems is described by, e.g. Squares filter
i

Moonen et al. (1989), Moonen and Ramos System matrices I Kalman states
(1991), where a state space model is computed I
directly from a block Hankel matrix constructed
from the input-output data. In a first step, a FIG. I. The left-hand side shows the N4S1D approach: first
the (Kalman) states, then the system matrices. The
state vector sequence is computed as an interface right-hand side is the classical approach: first the system, and
between a 'past' and a 'future'. Once the state then an estimate of the states.
vector sequence is known, the system matrices
are computed from a set of linear equations.
Similar data-driven identification schemes for
purely stochastic identification are well known,
data. This projection retains all the information
(see e.g. Arun and Kung (1990) and the
(deterministic and stochstic) in the past that is
references therein). Less well known is that
useful to predict the future. Then, the state
these algorithms can compute extremely biased
space matrices are determined from this state
results. This problem was studied and solved by
sequence. Figure 1 shows how these N4SID
Van Overschee and De Moor (1991a, b).
algorithms differ from the classical identification
The problem addressed in this paper is that of
schemes.
identifying a general state space model for com-
The connection of the two new N4SID
bined deterministic-stochastic systems directly
algorithms with the existing algorithms described
from the input-output data. Some papers in
above will also be indicated.
the past have already treated this problem but
This paper is organized as follows: the
from a different viewpoint. In Larimore (1990)
problem description and the mathematical tools
for instance, the problem is treated from a pure
can be found in Section 2. In Section 3 the main
statistical point of view. There is no proof of
projection is defined. Section 4 introduces a
correctness (in a sense of the algorithms being
closed form formula for the non-steady state
asymptotically unbiased) whatsoever. In De
Kalman filter estimation problem. This result is
Moor et al. (1991) and Verhaegen (1991) the
related to the results of Section 3 to find the
problem is split up into two subproblems:
interpretation of the main projection as a
deterministic identification followed by a stoch-
sequence of outputs of a non-steady state
astic realization of the residuals. In Moonen et
Kalman filter bank. Section 5 introduces a first
al. (1992) the problem is solved for double
N4SID algorithm that identifies the system
infinite block Hankel matrices, which implies
matrices exactly. In Section 6 accuracy is traded
practical computational problems.
off for simplicity in a second approximate N4SID
In this paper, we will derive two N4SID
algorithm. Section 7 shows how these N4SID
algorithms that determine the deterministic and
algorithms can be implemented in a numerically
stochastic system at the same time. The
reliable way, using the Q R and the Singular
connection with classical system theory (Kalman
Value Decomposition (SVD). Section 8 investi-
filter) will be used to prove the exactness
gates the connection with other existing algo-
(unbiasedness for an infinite number of measu-
rithms. Finally Section 9 will treat some
rements) of Algorithm 1, or the degree of
comparative examples. The conclusions can be
approximation (calculation of the bias for an
found in Section 10.
infinite number of measurements) of Algorithm
2.
The approach adopted here is similar to the 2. PRELIMINARIES
identification schemes of Moonen et aL (1989) In this section, we describe the linear time
for the purely deterministic case and Van invariant system we want to identify. We also
Overschee and De Moor (1991a, b) for the introduce the input and output block Hankel
stochastic case. First a state sequence is matrices, the past and future horizon as well as
determined from the projection of input-output the input-output equations.
 N4SID: Identification of combined systems 77

2.1. System description well as completely decoupled input-output

Consider the following combined dynamics•
deterministic-stochastic model to be identified The main problem of this paper can now be
stated: given input and output measurements
x~.+l = A x k + B u k -I- wk, (1)
u, . . . . . UN and Yl . . . . . YN (N--*~. ), and the fact
Yk = Cxk + Duk + Ok, (2) that these two sequences are generated by an
unknown combined deterministic-stochastic
with

*E
[(b:)
(w$ v~) = (S~.),
S')
R" dikl -> 02 (3)
model of the form described above, find A, B,
C, D, Q', R "~, S' (up to within a similarity
transformation)•
In the next two sections, we will define some
and A , Q " e N ..... , B e N ...... , C e R I×', D e more useful properties and notations for the
1~1.... , S ~ e N "×t and R ~ e N t×t. The input vectors deterministic and the stochastic subsystem•
u, e N ''×~ and output vectors YkelW ×' are 2.1.1. The deterministic subsystem• Associated
measured• o, e N~×' and wk e I~"×' on the other with the deterministic subsystem (4)-(5), we
hand are unmeasurable, Gaussian distributed, define the following matrices:

(c/
zero mean, white noise vector sequences• • The extended (i > n) observability matrix F,
{A, C} is assumed to be observable, while (where the subscript i denotes the number
{A, (B (Q.,)it2)} is assumed to be controllable• of block rows)
This system (1)-(2) is split up in a
deterministic and stochastic subsystem, by
splitting up the state (xk) and output (y~) in a CA
deterministic (.,i) and stochastic (.") component:
-i - CA:
xk = x a+x~., Yk = y d + y ~ . The deterministic
state (x d) and output (yd) follow from the
deterministic subsystem, which describes the CAi-ll
influence of the deterministic input (Uk) on the
deterministic output • The reversed extended controllability matrix
Q A}~ (where the subscript i denotes the
Xd+I = A x d + BUk, (4) number of block columns)
and
yd = Cx a + Dut. (5) Ad~'(Ai-'B A'-2B ... AB B).

The controllable modes of {A, B} can be either • The lower block triangular Toeplitz matrix
stable or unstable. The stochastic state (x~) and H~'
output (y~) follow from the stochastic sub- H~"
system, which describes the influence of the
noise sequences (wk and ok) on the stochastic
0 0 " i\ )
output CB D 0 ...
x~ +, = Ax'~ + Wk, (6) CAB CB D . .
and
y~ = Cx~ + ok. (7)
CAi-2B
The controllable modes of { A , ( Q ' ) Irz} are
assumed to be stable. 2.1.2. The stochastic subsystem. For the
The deterministic inputs (uk) and states (x d) stochastic subsystem (6)-(7) we define
and the stochastic states (xl) and outputs ( y i )
are assumed to be quasi-stationary (as defined in e~ %' E[x~(x~)']
dcf ~ x t
Ljung, 1987). Note that even though the G = E[xT(yk) ]
deterministic subsystem can have unstable An d~, E[y~(y~)'].
modes, the excitation (Uk) has to be chosen in
such a way that the deterministic states and With equations (3), (6) and (7) and through
output are finite for all time. Also note that since stability of the controllable modes of the system
the systems {A, B} and {A, (Q.,-)i/--,} are not {A, (Q,).2}, we find easily that the following
assumed to be controllable, the deterministic equations are satisfied:
and stochastic subsystem may have common as
p, = AP~A, + Q,
G = A P'C' + S" (8)
*E denotes the expected value operator and 6,~ the
Kronccker index. An = CP'C' + R'.
78 P. VAN OVERSCHEE and B. DE MOOR

This set of equations describes the set of all The deterministic and stochastic state matrices
possible stochastic realizations that have the are defined as
same second order statistics as a given stochastic sddef z d
sequence y~,. We call them the positive real
, =tx, x",+, x:+2 . . . x:+/_,),
equations. More details can be found in Faure x s di e=f t/ X is X is+ l X is+ 2 ... X is+ j _ l ) .

(1976). For the deterministic subsystem we define

It is also easy to derive that
1 { U')li-' ~
def ,. s" , I'cmi-IG i>0 lim - ~
Ai = E[yk+i(y'~.) ] = A~) i = 0. j--~] / /(U~)ji-, U~J2i-i I (xg)')
/ G'(A')-i-IC ' i<0
\ X,, /

Associated with the stochastic subsystem, we

define the following matrices: -
w'/R"R,21s:R_2_s
Oe.~tl .Ot
,2 R~ S'~I=
• The matrix M \ S, $2 P"/ kS I P"}'
A~%t(Ai-~G Ai-ZG ... AG G).
where we use the assumption that the limit exists
• The block Toeplitz covariance matrix L," (quasi-stationarity of uk and xka).
For the stochastic subsystem we find that, due
/ At)A_, A_. ".. A,_i\
to stationarity of y~,, the following equalities
hold true:

I Ai-2 Ai-3 "'" A. / lim 1 {~.,

-
Y(%li_ I ~tt~,,,s
]t~-oli-,sx, (Yit2i-,)')
s
j ~ = J \ lei]2/-,
• The block Toeplitz cross covariance matrix
H~
= ( L,'. (H,'.)"~
\H~ L~ /" (9)

The Matrix input-output equations are defined in

H~ ~f I Ai Ai- t ••• A,
the following Theprem (De Moor, 1988):

Theorem 1.
= F i AT. Yoli_l = F i x d _ F H ia Uoji_ 1 21- Y os l i _ l , (10)
2.2.Block Hankel matrices and input-output Yq~-~ =FiXd + HiUilzi-, + Yq21-~, d s (11)
equations and
g/-- id
Input and output block Hankel matrices are - A Xo + mid Uoli-i. (12)
defined as The theorem is easy to prove by recursive
/ Uo u, u., ...u/_,) substitution into the state space equations.

?.'..2 .2 iii .u'. , 3. T H E MAIN P R O J E C T I O N

\Ui--| Ui Ui+l ''' Ui+j_ 2
In this section, we introduce the projection of
/ Yo Yl Y2 "'" Yj-, \ the future outputs onto the past and future
yoli_,,,~Y, Y.2. .Y.3. ::: .Y./. ), inputs and the past outputs. The results can be
described as a function of the system matrices
,;,?i y, y., y..2.
and the input-output block Hankel matrices.
We define the matrices Zi and Zi+~ as
where we presume that j---->~ throughout the
paper. The subscripts of U and Y denote the .3>
subscript of the first and last element of the first
column. The block Hankel matrices formed with
the output yl of the stochastic subsystem are Zi+l = Yi+lJ2i-i .4>
defined as Y~;li-~ in the same way.
Somewhat loosely we denote the 'past' inputs where A/B =AB'(BB')-IB. T h e r o w space of
with Uoli_I or Uoji and the 'future' inputs with A/B is equal to the projection of the row space
ad2i_ l o r Si+lJ2i-|- A similar notation applies for of A onto the row space of B.
the past and future outputs. This notational Formula (13) corresponds to the optimal
convention is useful when explaining concepts. prediction of Y/12i-i given U,,izi_l and Yoli-i in a
 N4SID: Identification of combined systems 79

sense that and

II Y,I~;-, - Z,ll~-
is minimized constrained to
row space Z~ ~ row space \ Y . l i - i / .
So, intuitively, the kth row of Zi would
correspond to a k step ahead prediction of the
output. This intuition will become clearer in
Section 4.
These projections (Zi and Zi+l) are useful in
determining the combined system, since (as we
will show in T h e o r e m 2) the linear combinations
to be made of the input-output block Hankel
matrices to generate the matrices Z~ and Z~÷~ are
functions of the system matrices (A, B, C, D,
Q', S ", R'). Moreover, the system matrices can
be retrieved from these linear combinations, as
will be explained in Section 5.
It is tedious though straightforward to prove
the following theorem which delivers formulas
for the linear combinations to be n.'ade of the
rows of the input-output block Hankel matrices
to generate the matrices Z~ and Z~+~.
Theorem 2. Main projection.
• If the deterministic input uk and state x d are
uncorrelated with the stochastic output y~.:
lim 1 y,~l,_,U,. = 0 lim 1 Yi~l,_,(X.,),= 0
j-=j j-=j
1 following recursive formulas:
l i m - Y~I2,-,U'. = 0 lim 1 Y~lz,_.(X.a) ' = 0 ,
Qi = zilPi -1
xi=Ai(pd - SR -i S , )Fi+
t A~s (19)
W, = r , ( Pa - SR-'S')F~ + LT. (20)
A proof can be found in Appendix A. In the
next section, we give an interpretation of these
projections•
4. A BANK OF KALMAN FILTERS
In this section, we show how the sequences ~';
and )(i+t can be interpreted in terms of states of
a bank of j non-steady state Kalman filters,
applied in parallel to the data. This inter-
pretation will lead to a formula that will prove to
be extremely useful when determining the
system matrices from the data.
As stated before, it may come as no surprise
that there is a connection between the states 2,.
defined by the projection Z~ and some optimal
prediction of the outputs Yil2~-~-
To establish this connection, we need one
more theorem that states how the nonsteady
state Kalman filter state estimate -fk can be
written as a linear combination of u. . . . . . Uk-~,
Y. . . . . . Yk-~ and the initial state estimate 2..
Theorem 3. Kalman filter. Given 2., P.,
u. . . . . . Uk-i, y. . . . . . Yk-t and all the system
matrices (A, B, C, D, Q', S", R'), then the non-
steady state Kalman filter state -fk defined by the
$k = A $ k - i + B u k _ i + K k _ t

where the subscript, denotes past or future, × (Yk-l -- C2k-i -- Duk-i), (21)
• if the input is 'persistently exciting of order Kk-, = (APk-,C' + G ) ( A . + CPk-,C')-'. (22)
2i (Ljung, 1987)': rank U.12~_~= 2mi, and
• if the stochastic subsystem is not identically Pk = A P k - t A ' - (APk-~C' + G)
zero (the purely deterministic case will be x (I~,+ CPk_,C')-'(APk_,C'+ G)' (23)
treated in Section 6.1•3).
Then (for j---~ oo) can be written as

Z, = r,2, +HiUilzi-.,
" (15) No
z,+, = r,_,2,+, +H}L,U,+,I2,_,, (16)
with xk = ( A k - Qkrk I/,'/- Qkn'[I Q')l "'- '
X~ = ( A ' - Q,r, I a:'- Q,HI' I Q,)
{ S R - ' U,,i2,-.' ~
u.,_, !'
~ Y.

\Yk -
(17)
\ Y.li- i / (24)
where
2 , + , = (A '+' - Q , + , F , + , [ a f + , - Q,+,H,~+, I Q,+,)
Q~. = Zklp~-I, (25)
/ S R - i Uopi- ; I x, = A~P,,F~, + a L (26)
x~ U.li" J (18) and
= rka,r + (27)
\ ~:'li /
80 P. VAN OVERSCHEEand B. DE MOOR

The proof of this theorem and some details written as

concerning the special form of the Kalman filter
fi” = SR-‘U,,lri-,.
equations (21)-(23) can be found in Van
Overschee and De Moor (1992). Let us just All this is clarified in Fig. 2.
indicate that the error covariance matrix The expressions for P” and 2” can be
pk ‘if E[(xk - ik)(xk -a,)‘] is given by P’ + Pk, interpreted (somewhat loosely) as follows:
with P” the state covariance matrix from if we had no information at all about the
Lyapunov equation (8). initial state, then the initial state estimate
Note that the limiting solution (k + 03) of (23) would be 2” = 0 and the initial error
is -Pm, where P, is the state covariance matrix covariance would be equal to the expected
of the forward innovation model (Faure, 1976). variance of the state: P” = E[x~x~] = P” +
Hence the limiting error covariance matrix is P”. Now, since the inputs are possibly
pz = P” - P%, which is the smallest state error correlated, we can derive information
covariance matrix we can obtain (in the sense of about Xi” out of the inputs U,,lzi-,. This is
nonnegative definiteness). done by projecting the (unknown) exact
Also note that the expressions for vlk and xk initial state sequence Xi+ Xfl onto the
(26)-(27) are equal to the expressions of vi and row space of the inputs U,,l,,-,
Xi (19)-(20) with P” - SR-‘S’ substituted by p0.
2” = (Xi; + X;,)IU,,l,i-, = SR-‘U,,lzi-,.
If we now combine the results of Theorem 2
and 3, we find an interpretation of the sequences This extra information on the initial state
X, and Xii,, in terms of states of a bank of of the Kalman filter also implies that the
non-steady state Kalman filters, applied in error covariance matrix reduces from
parallel to the data. More specifically, compare P” + p” to:
formulas (17), (18) and (24):
(1) The j columns of Xj are equal to the p” = p” + p’ - ;i;; R”(jy)[
outputs of a bank of j non-steady state
Kalman filters in parallel. The (p + 1)th =P”+P‘--SR-‘S’.
column of Xi for instance, is equal to the
non-steady state Kalman filter state Pi+,, of These are exactly the same expressions for
the Kalman filter (21)-(23) with initial Xi” and p” as we found above.
error covariance matrix at starting time p It can also be seen that when the inputs
are uncorrelated (white noise), the pro-
&p,+P”=p”-SR-‘S’+P” jection of Xz + Xf, onto the inputs U,,l,;-,
UP
is zero, which implies that there is no
f,,=SR-’ ... information about the initial state 2”
contained in the inputs $I~,-,.
( Up+?;-I 1. The state sequence Xi+, has a similar
Notice that pp is independent of the interpretation. The pth column of Xi+, is
column index, so it is denoted with P”. equal to the non-steady state Kalman filter
In this way, all the columns can be state estimate of the same (in a sense of
interpreted as Kalman filter states. The the same initial conditions) non-steady
initial states of the j filters together can be state Kalman filter bank as discussed

1
UO uj-l

( )I/01;-1

&Ii-l
W-1

Yo
Ui+p-1
YP
uitj-2
Y,-1

Yi-1 Yi+p-1 Yi+j-2

T_
i; [ *i
. ii+, .. -++,-I
1
FIG. 2. Interpretation of the sequence k, as a sequence of non-steady state Kalman filter state estimates
based upon i measurements of uk and ya.
 N4SID: Identification of combined systems 81

above, but now the filter has iterated one deterministic subsystem, directly from the given
step beyond the estimate of the pth inputs Uk and outputs y,. The stochastic
column of X,-. This is valid for all columns subsystem can be determined in an approximate
p = l . . . . . j. sense.
(2) We define the residuals ~a of the
projection as 5.1. The projections
First, the projections Z; and Zi+~ (13)-(14)
~ i . Yil21-i
. Zi
. Yil2i-I
. -- Fif(i HidUil2i-i. have to be calculated. In Section 7 we will
(28) describe a numerically stable way to do this.
Since Z~ is the result of the projection of For convenience, we rewrite these projections
Y,.121_l on the row space of Uot2i-, and as follows:
Y01~-~, the residuals of this projection ( ~ )
will always satisfy: ~iU[~12i_l= 0,
~iY[~li_ I = 0 and ~iZ~ = 0. Also, since X'i (33)
\ l i x mi
can be written as a linear combination of
\ Y"I~-~ /
Uol~-i and Yoli-i (see formula (17)), we
find: ~iX't = 0. Zi+ I
(3) Since the corresponding columns of ,,~'i and L~+z ':+' /
,~i+~ are state estimates of the same (in a I(i- I) × m ( i + I I(i--I)xm(i--I) I(i--I)xl(i+|)/
sense of the same initial conditions)
non-steady state Kalman filter at two U(,Ii )
consecutive time instants, we can write
x Ui+ll2i_! (34)
(see formula (21))
Yol/
-.~,+ t = AX, + BU,], + K,(Y..I,- C.,.(", - DU, I, ).
with, from (15)-(18)
(29)
It is also trivial that L~ = F;([A i - Q,F,]S(R-'), I,,/+ A~'- Q,,'/),
Yih = C . f ( i dr DU, I, + (Y,I i - CS, - DUd, ). (35)
(30) L 2,= H~i + F,[A'- QiF,]S(R-')...÷,I~.,,. (36)
If we inspect the formula for ~i a little bit L~ = FiOi, (37)
closer (28), we see that its first row is
equal to Y/i / - C f ( i -DUil i. And since we with (R-~)q,, denoting the submatrix from
know that the row space of ~s (and thus column 1 to column mi.
also the first l rows of ~;) is perpendicular The expressions for L~÷l, L~+l and L3÷t are
to Uol2/-i, Yoli-i and ,('i, we find (together similar, but with shifted indices.
with (29) and (30))
5.2. Determination of Fj and n
/ U,,12~-,~ l An important observation is that the column
"ff~+'=mf(~+BU~l~+~ Y " I ' - ' ] ' (31) space of the matrices L~ and L 3 coincides with
the column space of F i. This implies that Fi and
',ki/ the order of the system n can be determined
from the column space of one of these matrices.
gI/--C2'+DV~I'+~ ~ * - ' 1 ' (32) The basis for this column space actually
determines the basis for the states of the final
\2,/ (identified) state space description.
where (.)± indicates a matrix whose row Let us mention two other possible matrices
space is perpendicular to the row space of that have the same column space as Fi
(.). These formulas will prove to be
L] + L;. LT, (38)
extremely useful in the next section where
and
we determine the system matrices from Z~
and Zi+l. L ~ U,,{,_,
(L~[ ,)(y.ii_, ). (39)
This summarizes the whole interpretation as a
bank of non-steady state Kalman filters. It should be mentioned that, for i---,oo the first
one (38) will lead to a deterministic subsystem
5. IDENTIFICATION SCHEME that is balanced (see also Moonen and Ramos,
In this section, we derive an N4SID algorithm 1991) while the second one (39) leads to a
to identify exactly (unbiased for j---~oo) the deterministic system that is frequency weighted
82 P. VAN OVERSCHEE and B. DE M o o r

(with the input spectrum) balanced (Enns, 1984) where we define

together with a stochastic subsystem of which the
forward innovation model is balanced in a
deterministic sense. We will not expand any X22]
more on this, but keep this for future work.
We can now determine Fi, Fi-t and the order ~f ( ~ - AF*~(CF,~(kFi~_DDB)F*~_tH----idAF*~(
_tCF*~(kHi-~o
O -~
) ]l-
n as follows: let T be any rank deficient matrix of
which the column space coincides with that of Fi.
• Calculate the Singular Value Decomposition \Fi_lB/ \Hi-i//
(44)
Observe that the matrices B and D a p p e a r
linearly in the matrices ~t2 and ~2~.-
• Since T is of rank n,
the n u m b e r of singular Let I-[ be a matrix whose row space coincides
values different from zero will be equal to with that of
the order of the system. r:z,
• The column spaces of F~ and UIZI,~2 Uil2i_l]'
coincide.* So, Fi can be put equal to U,Z~ n.
• With F~ defined as F~ without the last I rows then (from (43))
(l is the number of outputs), we get:
(r:_,zi+, r*,z, /rI
Fi- I = Fi.
In the following, we will take T equal to the Obviously, this is a set of linear equations in the
expression in formula (39), but one can replace unknowns A, C, 2gl,_,Yg22.
this with any other matrix of which the column Another point of view is that one could solve
space coincides with F~. the least squares problem
(r: r:z, )ll
5.3. Determinationof the system matrices A,cmi,~x.. \ Yili /I \C1~22 ~ [IF"
We now assume that Fi, Fi_t and n are (45)
determined as described in the previous section,
and are thus known. From (15) and (16) it Either way, from (43) we find (term by term).
follows that:t
Term 1. A and C exactly.
)(i = F[(Zi - HidUilzi_,). (40)
X,+, = Fi*-t(Zi+, - Hd-, Ui+,I2i-,). (41) Term 2. Xi2 and ~22 from which B and D can be
unraveled by solving a set of linear equations,
In these formulas, the only unknowns on the analogous to the one described in De M o o r
right-hand side are the matrices H a and H,-d_j. (1988). Note that in (44), B and D a p p e a r
From (31) and (32) we also know that linearly. Hence if A, C, Fi, Fi-1, ~t2 and ~c22 are

(2)lzi known, solving for B and D is equivalent with

solving a set of linear equations.

Term 3. The residuals of the least squares

(42) solution (43) can be written as
If we now substitute the expressions for )(i and
Xi+l (40)-(41) in this formula, we get [wi q
p= Zi
.% = \ Viii]'
Ell, )=-(A)r,zi+/X,.',
(r~_,zi.,
~)Uil2,-, where W I. and V. I. are block Hankel matrices:~
Term 1 Term 2 with as entries wk and ok: the process and
measurement noise. This is clearly indicated by
Uol2i-! equation (42).
+ Zi , (43) The system matrices R s, S s and Q" are
YC,
The block Hankel matrices of the residual p have only
° The factor ZIIn is introduced for symmetry reasons. one block row. The notation is introduced to be consistent
t At denotes the Moore-Penrose pseudo-inverse. with previous notations.
 N4SID: Identification of combined systems 83

determined approximately from p as follows: (ii) B, D follow from A, C and ~ z , ~22

>__( Q, s.,) through a set of linear equations.

1 (SS) ' R'" (iii) (\(S,), I R,) *-'1 (PtP~ P,P'2~

The approximation is due to the fact that the
bank of Kalman filters for finite i is not in steady
state (see for instance the Riccati difference
equation (23)). As i grows larger, the ap-
proximation error grows smaller. For infinite i
the stochastic subsystem is determined exactly
(unbiased). More details on this purely stochas-
tic aspect can be found in Van Overschee and
De Moor (1991a, b).
5.4. N4SID Algorithm 1
In this section, we summarize the first N4SID
algorithm step by step, not yet paying attention
to the fine numerical details, which will be
treated in Section 7.
(1) Determine the projections
/{ .,,,,-,)
zi=Yir2i-,/ ! Algorithm 1. Another advantage of this
IX Yoli- I /
P ~ ' I ~ '
"li x mi I li x mi i li x li"
The deterministic subsystem will be
identified exactly (as j---,oo, independent
of i). The approximation of the stochastic
subsystem is still dependent on i and
converges as i---* oo.
6. A S I M P L E A P P R O X I M A T E SOLUTION
In this section, we introduce an N4SID
algorithm that is very similar to the 'exact' (as
/.--, oo) algorithm of the previous section
(Algorithm 1). The algorithm we present now
finds a good approximation to the state Xi, and
to the system matrices, without having to go
through the complicated Step 4 for the
determination of B and D. This results in a
simple and elegant algorithm with a slightly
lower computational complexity as compared to
simplified N4SID algorithm is that it is very
closely related to existing algorithms (Larimore,
1990). This means that the analysis of this
simplified algorithm can also be applied to the
\ ~)1i- I / other algorithms, and can thus contribute to a

Z i + I = Y/+l]2i_ I
/( ,,,, ) Ui+ll2i-i .
better understanding of the mechanism of these
algorithms. A disadvantage is that the results are
not exact (unbiased) for finite i (except for
Y.li special cases), but an estimate for the bias on the
solutions can be calculated.
(2) Determine the Singular Value Decom- If we could determine the state sequences Xi
position and -~'i+~ directly from the data, the matrices A,
B, C and D could be found as the least squares
solution of (31)-(32).
\ Y.li- t / Unfortunately, it is not possible to separate
The order is equal to the number of the effect of the input HidUi]2i-i from the effect
non-zero singular values. of the state Fi-~'i in formula (15), by just
dropping the term with the linear combinations
I'~i = UI'~.II/2 and r i _ I = UI'~~II/2.
of Udzi_~(L~) in the expression for Zi. We would
(3) Determine the least squares solution (Pt automatically drop a part of the initial state if we
and P2 are residuals) did this (see for instance (36)). So, it is not
possible to obtain an explicit expression for FiXi
j n mi and I"i_lZYi+l, without knowledge of H/d, which
n(Ft-,Zi+l] _n(~{I, ff{,2~ would require knowledge of the system matrices.
It is however easy to find a good approxima-
tion of the state sequences, directly from the
J J
data. If we use this approximation in (31) and
(32), we obtain a second very elegant and simple
mi Udzi_~/+ lkp2/" N4SID algorithm that calculates approximations
(4) The system matrices are determined as of the system matrices.
follows:
6.1. The approximate states
A*--;~ll, In this section, we derive an approximate
(i)
C*--:~21. expression for the states ,~'i and Xi+j. This

AUTO 30:1-G
84 P. VAN OVERSCHEE and B. DE MOOR
approximation can be calculated directly from
input-output data.
The approximate state sequences are calcu-
lated by dropping the linear combinations of
U~l~_l out of Zi, and the linear combinations of
U,+112,-~ out of Z;+~. In this way, we obtain
Kalman filter states of a different Kalman filter
compared to the one that produces X~ (in a sense
of different initial conditions). We call the
resulting matrices ri,f( i and I"i_ i,e~'i+ I
Fi,f(i = Zi - L 2iUil2i-,
and
Fi_lf(i+ I = Zi+ a - L2+tUi+ll2i_t.
From (33) and (34) we derive
L = ([A'- Q,F,]S(R-'),I,.,
+ a",-
\ Y"Ii- 1]
)t'~+, = ([A '+' _ Q,+tF,+,lS(R-'),Im~,+, ~
~ + , / U.IA
+ A",+,-Q,+,n,+,
The matrices ["if( i and 1-'i_l.e~i+l can be obtained
directly from the data without any knowledge of
the sytem matrices, and can be interpreted as
oblique projections as described in Van Over-
schee and De Moor (1992).
The state sequence ,('~ is generated by a bank
of non-steady state Kalman filters, with:
po = pd _ S R - I S , + p~ and ,~-o= S(R-l)llmiUoli_ j.
The state sequence ,('~+~ on the other hand, is
generated by a bank of non-steady state Kalman
filters, with: 15o= p d _ S R - I S , + p~ and ,(.o=
S(R-I)qm(i+~)Uoti. So clearly, both sequences do
not belong to the same bank of Kalman filters,
and the useful formulas (31)-(32) are not valid
for these two sequences.
Still, we will see below, that ,l~'i and )(~+l are
very close to ~'/, and that ) ( / = ~'/, -~'i+t = -'~'i+~ if
at least one of the following conditions is satisfied:
• i -.-~ 00.
• The deterministic input u~ of the combined
deterministic-stochastic system is white
noise.
• The system is purely deterministic.
For these three special cases, we will analyze the
difference between Xi (17) and Xi (48), defined
as 6Xi:
6X,~r ff, -ff, = [A' - Q,F,]S(R-')m,+,I2.,,U,12i_,.
6.1.1. i---, oo. Define the term between square
brackets in (49) as ~ A ~ - Q;Fi. Now, it is
easy to prove (Van Overschee and De Moor,
1992) that
i--I
= I-[ ( A - KkC). (50)
k =0
Since the non-steady state Kalman filter
converges to a stable closed-loop system
(Anderson and Moore, 1979), we find that ~
grows smaller when i grows larger. It is also clear
that: lim ~ = 0. In the limit, equation (49) thus
i~tm
becomes (with S(R-t)mi+ ~12,,iUd2;_~ finite)
(46)
lim 6Xi = lim ~iS(R-t)mi+ll2,,iUil2,._l = O.
i~ i~oo
(47) The same holds for tSXi+~. So, we can conclude
that, for i--->o0, there is no difference between
the state sequences ~'i and X,-. Actually, when
i---~0o, the non-steady state Kalman filter bank
converges to a steady state Kalman filter bank,
(48) i.e. the Kalman filters converge (the Riccati
difference equation (23) converges to an
algebraic Riccati equation). To obtain the effect
of i---~ oo on a computer, in theory we would need
a test like.
II6X, ll -_ II -'
V'n X j V'-n x j
where em,c, is the machine precision. This is a
condition that can only be checked, after the
identification is done. Appendix B indicates how
this quantity can be calculated.
In practice, it turns out that i does not have to
be that large. The identification results are
already very good for reasonably small i (---10).
This will become apparent in the examples of
Section 9.
6.1.2. Uk white noise. With the deterministic
input uk white noise, we find: S = 0 and R = 12,,,~
with 12,, the 2mi x 2mi identity matrix. So, for a
white noise input, we find for any i (see (49)):
6.1.3. Purely deterministic system. From Mo-
onen et al. (1989), we know that generically for
deterministic systems we have (if there is no
'rank-cancellation', see De Moor (1988))
rank = mi + n
and (51)
( Uil21-1~
rank \Y~l2 ~ - t / = mi + n,
which implies that
(49)
rank
(u.12,-,]
Yoli-I / = 2mi + n.
This rank deficiency implies that for purely
deterministic systems, the proof of Theorem 2
 N4SID: Identification of combined systems 85

breaks down (see Appendix A: ~-~ cannot be attention to the fine numerical details, which will
calculated). The following theorem is an be treated in Section 7.
alternative for Theorem 2 for purely determinis- (1) Determine the projections
tic systems:

Theorem 4. Purely deterministic systems. If the z,: I/ !

input is persistently exciting and the stochastic
subsystem is zero we have / \ Yoli-i /

= (r/[af - A'rlnf] I r,Air,-t)(~]-

Yoli- l ]
= 1~-;-4 ~ 1 .-.t~l/u~12,_,/,
= r;x, = r,2,. klixmi I lixmi I lixli/ ~ l
\ Yoh-i /
A proof can be found in Van Overschee and De and
Moor (1992). This implies that for deterministic
systems, 6X~ is also equal to zero.
Zi+l = Y/+ll2i-t Ui+II2/-i
6.2. The algorithm Y.h
We know from (31) and (32) that the least
= ( LL'+'
/ 2 +[ ' ~
squares solution ~ of
I(i--I)×m(i+l) I(i--I)xm(i--I)

is equal to: X
I(i--I) x l ( i + l ) Yoli "
~= C D"
(2) Determine the Singular Value
We also know that from the residuals of this
Decomposition
least squares problem, we can approximately
calculate the stochastic subsystem (exactly if
i~oo). Unfortunately, it is impossible to
calculate the states ~'~ and -~'~+l directly from the
data, without any knowledge of the system The order is equal to the number of
matrices. In the previous section, we have seen non-zero singular values.
that for some special cases ,~'i=,~'i. In the Fi = UI~-I/2, and Fi-t = Uj__.._Z]/2.
general case, we have X~ = ,~ if i is reasonably
large. This is because (Section 6.1.1) (3) Determine the states -'~i and Xi+~
i--I , ,
//L...\
6X, = 1-I (A - KkC)S(R-'),.,+q2.,,U,12i_ ,. = Fi ( L i [ \ Yoli- i /
k =0
and
Consequently, for the least squares solution .~ of

min ('~'~+'~ - [ "~'~ ~ 2

(52)
(4) Determine the least squares solution
we have
j n m j j
n(X,+,~ _n{.~,! .~lz 5 x n ( X, ~ +nCo,~
Contrary to ~'i, Xi can be calculated directly l\ Y,I, i - l\~e~, ~e=/ m \ ~ / l\pd"
from the data, without any knowledge of the (5) The system matrices are (approximately)
system matrices. The solution is exact (.~ = &~) determined as follows:
for the cases of Section 6.1.1(i--*~), Section
6.1.2 (input is white noise) and Section 6.1.3
(purely deterministic), since then ~'~ =~';. In
Appendix B, an approximate expression for
( ~ - . ~ ) is derived. This is the bias on the
solution.
(ii) \(S~),[R,/ J \P2P'~ p2P'2]'
where in the general case the approxima-
6.3. N4SID algorithm 2 tion of the deterministic and stochastic
In this section, we summarize the second subsystem depends on i (even when j---~ oo)
N4SID algorithm step by step, not yet paying and converges as i, j---~ oo.
86 P. VAN OVERSCHEE and B. DE MOOR

7. A NUMERICALLY STABLE AND EFFICIENT where we use the shorthand notation

IMPLEMENTATION R4:6.1:3 for the submatrix of R consisting of
In this section, we describe how N4SID block rows 4-6 and block columns 1-3.
Algorithms 1 and 2 can be implemented in a (3) Calculate the projections (see also the
numerically stable and efficient way. We make note at the end of this section).
extensive use of the QR decomposition and
SVD. Zi = R5:6.1:4R-1:4'
' h4 ( ~'l,~i_--1I)
In Sections 5.4 and 6.3 we described step by
step, two N4SID algorithms that determine the
system matrices from given input-output data.
In this section, we will show how these two =(L~ H[L,~)I~
\v,,,,-,/I,
algorithms can be implemented in a numerically and
stable and efficient way.
Steps 1 through 4 are common to both Zi+l = R6:6.I:5R~..I.l:5(U~II~1 )
algorithms.
(1) Construct the block Hankel matrix ~(*

~ -- ( ~ : , l : : ~ : ) / V f .
= (el+, IEL, ILL,)U~+,I~_, -
Yoli
2 ( m + I)i × j
(4) Determine Fi and n through the SVD of
(2) Calculate the R factor of the RQ Fi)(i
factorization of (L] 0 3
Li)Rl:4.1:,Ql:,
~= R x Q'
=(U, U2)(~ ' Z2)(Qt:4V)
0 •
2(,n+l)ix2(tn+l)i 2(m+l)ixl

Note that this SVD can be calculated

with Q'Q = 1 and R lower triangular. It is
through the SVD of (L] 0 L3)RI:4.1:4,
important to note that in the final
since Q,:4 is an orthonormal matrix. The
calculations, only the R factor is needed,
rank is determined from the dominant
which lowers the computational com-
plexity significantly. singular values of this decomposition (X~),
Partition this factorization in the follow- and Fi can be chosen as
ing way: Fi = U, ZI a, Fi_, = F~,
where the underbar means deleting the
J last l rows (1 is the number of outputs).
mi/u,l,,\ The two algorithms described in Sections 5
m [ UA, ~ and 6 now differ in Steps 5 and 6.

m(i- I)[ Ui~,1"-,-,I = N4S1D Algorithm 1.

(5) We find for the left-hand side and
li ~Y"l, I I
right-hand side of equation (43), written
as a function of the original Q matrix
I(i- I) \E÷fl2i-i/
( r*,z, )= ["X-"~'"R
, ~,, .~:6.,:4~c v
mi m m(i- 1) ti t t(i- 1) ] \Uil2i_l \ R2:3.1:4 ]~1:4,
mi and
In R2, R2: 0 0 0 0 Q~ I/2 t
~--- h5 t
m(i- 1) R~, R~ R~ 0 0 0 '3
,

li Ral ga2 R43 R44 0 0 [Q~| (6) Now the least squares problem (45) can be
l RsI R52 R5" Rs4 R55 1 ' ~ ' ~ ! rewritten as
/(i- 1) R., R~2 R, 3 R~ R6~ R~/ \Q min (211t2(UI)tR6:6"I:4~
X \ Rs:5. h4 /
/Z-t~,R \112
* The scalar k/i-is used to be conform with the definition of
E. \ R2:3.h4 /IIF
 N4SID: Identification of combined systems 87

and solved in a least squares sense for Xc. With the singular value decomposition
Note that the Q matrix has completely
disappeared from these final formulas. Uoli-i ~ /Rt.i 1.4\
The first n columns of ~ are A and C (yo[i_l] =tg4141114)Ql:4t
stacked on top of each other. The next
=(Ul U2)( S' 0 V'' '
columns determine B and D as described
in Section 5. The residuals of this solution
we find as an appropriate basis
determine the stochastic system as de-
scribed in Section 5. V'I ) Q t
I'I= l~ 1:4-
~J~'2:3.1:4 •
N4SID Algorithm 2.
The rest of the algorithm is very similar to the
(5) We have
general case.
We should note that in the generic real life
case, the problem of rank deficiency is a pure
academic one.
-----(Z -In , i, 0 L 3i)R1:4.1:4"~,
I UI(L
\ R2:2.1:4 ] h~: 1:4, a n d 8. C O N N E C T I O N WITH EXISTING ALGORITHMS

8.1. Instrumental variable method

Y/li ] This method was described by De Moor et al.
(1991), Verhaegen (1991). The basic idea will be
Z'("Z(U,)t(L)+, 0 L3+,)R,:s.,:.s'~o,
shortly repeated here. We start from the
= R5:5.1:5 ]~l:S. input-output equation (11): Yilzi_ 1= ['iXd +
d s
(6) Now the least squares problem (52) can be HiUd~-i+ Yd2i-I. Projection of this equation
rewritten as on the row space perpendicular to that of GI2,-t,
gives
min ( Z-(I'z(G)t(L]+I 0 L3+I)RI:5.,:4)
~" R5:5,1:4
Y/12i_ I/Uil2i-I
.1_ = FiX di/Uil2i-i
± + Yilzi-l.
s (53)

_,,~(Zl - I t 2 U,(L,
t I
0 L,3)R 1:4.1:4) 2.
\ R2:2,1:4
Once again, the Q matrix has completely
disappeared from these final formulas.
From 5~ and the residuals we can find the
system matrices as described in Section 6.
Note.
• The projection of Step 3 is written as a
function of U and Y, and not as a function
of Q, because we need to be able to drop
the linear combinations (L~) of the rows of
Uil2i_ I in Zi to find the sequence FiXi and
the linear combination (L~+I) of the rows of
Ui+ll~i-i in Zi+l to find the sequence
Fi-iXi+l (see formulas (46)-(47)).
• The possible rank deficiency of the row
space we are projecting on (in Step 3) has to
be taken into account (see Section 6.1.3).
This occurs when R1:4.1:4 and R~:.S.l:.S are
rank deficient. It can be checked by
inspection of the singular values of R1:4.1:4.
If one of them turns out to be zero (purely
deterministic case), a basis 1-I for the space
we are projecting on has to be selected. This
basis should contain the row vectors of
Uil2i_J explicitly. This makes it easy to drop
the linear combinations o f GI2i-~ (L~z) in Zi
to obtain the sequence FiXi.
Note that Y~12i-I/U02i-I = Y~'12i-I, since
Y~12i_t/Uil2i_ I = 0 as j---~ oo. Projection of (53) on
t IIF
Uoli-i gives: [ Yiiz~-ll Uii2i-i]l
± U,,li-i =
r i[Xi/Uil21_l]/Uoii_l,
d ±
since
s
Yi[2i_l/Uoli_ I =0. It
can be seen from this equation that the column
.L
space of [Yd2~_ffUd2i_~]/Uoli_t is equal to that of
I~/(only the deterministic controllable modes are
observed). So, from the column space of 1-~,, A
and C can be derived. B and D are then found
from the fact that: ( r i d) ± [ ril2,-,/ uil:i-,lUii2i-,
* =
(I~/)±H/a which leads to a set of linear equations
in B and D. Finally, the stochastic subsystem is
identified from the difference between the
measured output and a simulation of the
identified deterministic subsystem with the given
input uk. This difference is almost equal to the
output of the stochastic subsystem (y~). The
stochastic subsystem can then be identified with
for instance one of the algorithms described by
Arun and Kung (1990) and Van Overschee and
De Moor (1991a, b).
A clear disadvantage of this algorithm is the
fact that the deterministic and stochastic
subsystem are identified separately. This in-
volves more calculations. Two different A
matrices will also be identified (one for the
deterministic and one for the stochastic system),
even though a lot of the dynamics could be
shared between the two subsystems.
88 P. VAN OVERSCHEE and B. DE MOOR

8.2. Intersection algorithms calculate the states Xi and Xi+~ as intersections

In the literature, a couple of algorithms, which
we call 'intersection algorithms' have been
published by Moonen et al. (1989, 1992). It is
interesting to note the analogy between these
algorithms and Algorithm 2 described in Section
6. We show that the row space of A'i can also be
found as the intersection between the row spaces
of two Hankel matrices ~'~ and ~2. Define
,9 = row space ~ fq row space ~2 with
(see Moonen et al., 1992 for example).
8.3. Interpretation of general projection methods
Larimore (1990) shows that viewed from a
statistical point of view, the state (memory) M; is
the solution to
[Y;12/-I \ Yoli_l ,/j

where p[A [B] is the probability distribution of

A given B.
and For Gaussian processes, these distributions are

./y
I
if('2 = |
\
il2i-I/\"
" YOI;-I
Uil2i-i
"
characterized by the first two moments, and we
can replace p[A [B] by the expectation opera-
tion E[A I B]. This expectation operator can in
its turn be replaced by the projection operator
(Papoulis, 1984). We then get
Facts.
• The subspace `9 is n dimensional. l( Uil2i-I ~
Yil2i-l/\ Yi)l;__lll= Yil2i-I/\ Mi 1" (54)
Proof. We have: rank ~1 = tel, with tel = (m + By substitution it is easy to verify that all of the
l)i in the generic combined deterministic- following three alternatives: Mi = X;, Mi = -~'i
stochastic case. From equation (15), we find
and Mi = f(;/U~zi-t satisfy equation (54).
that: rank ~2 = mi + n. Finally, if the input is
Actually, every matrix Mi satisfying: row space
persistently exciting: rank (~'t ~z)' = rt + mi.
M ; = r o w space [Zi-ff2Uilzi-i and rank Mi = n
With Grassmann's Theorem, we find: dim [ ~ N
(with ff~ ~ Rzi×"), will also satisfy equation (54)
Y(z]= rank ~t + rank ~z - rank (Ygl ~2) ' = ( r , ) +
and give rise to an n dimensional memory. This
(mi + n) - (xl + mi) = n.
ambiguity is due to the fact that the exact
• The intersection subspace `9 is equal to the
influence of the input variable Uilz;-~ on the
row space of Xi.
output is not known. In the frame work of linear
theory, the most natural choice for Mi is of the
Proof. one for which
--,,~; lies in the row space of ,gt'~, since it is
written as a linear combination of Uop_~
Y,'12i-I Yop-I / = FiMi + Hi U;12i-i
and Yoli_I (see Formula (48)).
- - , ( ' ; lies also in the row space of Y(2 since we
is satisfied (f~ = H~). This is because this choice
know from (46) that
corresponds to the linear state space equations
Mi+ 1=AM; + BUd;,
and
So, ~'i lies in the intersection of the row Y~i / ( U°I2i-' ~ = CMi + DUd;.
spaces of ~ and ~ . Since both the I \ Y~Ji-I I
intersection subspace `9 and the row space This leads to the choice M; = ,~';.
of A'i are n dimensional, they must coincide.
This derivation indicates strong similarities
between N4SID Algorithm 2 and the existing 9. EXAMPLES
intersection algorithm mentioned above. When
j-.-,0o both algorithms will calculate the same
9.1. A simple example
solution (if the same state space basis is used). In This simple simulation example illustrates the
practice however (finite j), they calculate slightly concepts explained in this paper. We consider
different models. In the references given above, the single input single output single state system
in forward innovation form (Faure, 1976)
only purely deterministic systems and the
asymptotic case i - - , ~ are treated. The inter- xk+m = 0.9490xk + 1.8805uk - 0.1502ek,
pretation of Kaiman filter states is totally absent,
as is the effect of finite i. and
The similarities indicate an alternative way to Yk = 0.8725xk - 2.0895uk + 2.5894ek.
 N4SID: Identification of combined systems 89

with e, a unit energy (or= 1), zero mean, sequences was calculated (Monte-Carlo
Gaussian distributed stochastic sequence. experiment).
Effect of i. Figure 3 shows the results as a function of i for
We first investigate the effect of the number of both N4SID algorithms. The results for Algo-
block rows (i) of the block Hankel matrices. The rithm 1 are represented with a dotted line (:) and
input Uk to the system is the same for all circles (O). The results for Algorithm 2 are
experiments and is equal to a filtered unit energy represented with a d o t t e d - d a s h e d (-.) line and
( o = 1), zero mean, Gaussian distributed stoch- the stars (*). The exact values are indicated with
astic sequence added to a Gaussian white noise a dashed line.
sequence ( o = 0 . 1 ) . The filter is a second order Figure 3a shows the eigenvalue of the system
Butterworth filter, with cut-off frequency equal as a function of i. Clearly the estimates are
to 0.025 (sampling time 1). The number of data accurate, and there is hardly any difference
used for identification is fixed at 1000. One between the two algorithms. Figure 3b shows the
hundred different disturbance sequences e, were deterministic zero of the system (A - B D - ~ C ) as
generated. For each of these sequences and for a function of i. The difference between the two
each i in the interval [2, 15], two models were N4SID algorithms is clearly visible. Algorithm 1
identified using N4SID Algorithm 1 (Section 5) estimates a zero that is close to the exact
and 2 (Section 6). Also for each of these models deterministic zero, independently of i.
the bias was calculated using the expression in Algorithm 2 on the other hand is clearly biased
Appendix B. Then the mean of all these for small i. Figure 3c shows this bias
quantities over the 100 different disturbance (dashed-dotted line) and the estimated bias

Eigenvalue of A Deterministic zero

0.95 1.76

0.949
' h i ., i 1.75 --i t ........ -i,.-'*. .................. i ........................
. .; .
0.949 '° i ; w,• "q ~,--* .
..... ":.,7-:,--/ .... -y.~,~-:~.-:z
i' V :, ° 1.74 ......... ; ... t~.,,.:... . L ..~".'~ .................

0.948 i ............ i . . . . . . . ..d .e °~ " '

,,.,.f..2. . . . . . . . . . - - , i - . . . . . ~.~_.~,,::~
0.948 t 1.73 i i •
5 10 15 5 10 15
i

Bias on deterministic zero Bias on D

0 • : f~-.& .4 "
0 ! ! .io..-'~..-o :
' ~" "'v o'"
: , 8 ' - e " . 6 ' " ," 6~.~.e..~,' • -¢"
-0.005 i f" "~" °"
i . '" .."

-0.01
-0.01 ..... 17.'.( L~............................................... ;"

-0.015
-0.02!
-0.02

-0.025 i -0.03
5 10 15 5 I0 15
i

Stochastic zero Standard deviation of A

I0-I :: ::::::: : :'.:::I:::":; I 1:1~: : : :'::~: : :',:::~::; :l:~t
1
1
:, . - . . . ~ - ~ . -
0.9 ...... ! -~..-., .... .~.*" ........... 4 !:i!; ......... " ,-,
10-2
t,"
0.8
!!: :,!: : :.i :i:i i !! :.-.!:~%;.!i ::-!!i !!!k-!

0.7 ..... ~ :;iii ........... i..~:}'.~

i i 10-3 .... ;;iii , , i;iiii
5 10 15 10 z 103 104
i
FIG, 3. (a) Eigenvalue of A as a function of i ( N 4 S I D A l g o r i t h m I (*) and 2 (O)). (b) Deterministic zeros
as a function of i ( N 4 S I D Algorithm I (*) and 2 (O)). (c) Calculated (*) and estimated ( O ) bias on the
deterministic zero as a function of i ( N 4 S I D Algorithm 2). (d) Calculated (*) and estimated (O) bias on D as
a function of i ( N 4 S I D Algorithm 2). (e) Stochastic zero as a function of i ( N 4 S I D A l g o r i t h m 1 (*) and 2
(O)). (f) Standard deviation of A as a function o f j ( N 4 S I D A l g o r i t h m I).
90 P. VAN OVERSCHEE and B. DE M o o a

(dotted line) using the expressions of Appendix Singular Values

I0 ~ . . . . . . . . , ........... , ............ .. .
B as a function of i. As expected, the bias grows : !!!!!!!!!
!!!?!!!!!!!~!!!!
!!!!!!!!!!!!!
!-!!!i!!!!!!
!!!!!!
!!!!!!!!!!!i
!?.'!!!!
!!!!!!
!!!!!!
!!!!~!!!!!!!
!!!!!!
!!!!!!:::
smaller as i grows larger. Figure 3d shows the . . . . . . . . . . ~ ................. , ............. ! .................... ~ ..............

calculated and estimated bias on the identified D

matrix. This matrix seems to be a lot more
sensitive to i than A. i!!!
.:::,:: ::i.:~: ::2 :.:..::i.!::.:.:/.::...:::.i:::/:::::::..:::2!.:::..:::::.:::
Finally Fig. 3e shows the estimated stochastic
zero as a function of i (the eigenvalues of ......... ;.........~........................i..................... ~...............................................
A - KC, with K the steady state Kalman gain). 101 .::::: :::::::::::::::::::::::::::::::::: :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
. . . . . . . . . . . . . . . v . . ~ ...................... , .......................................................................

The convergence is very slow (the exact zero is

0.9996). More details can be found in Van
Overschee and De Moor (1991a, b). For this
example, we can conclude that both algorithms i0 o .........................
:. :.. i...................
~...................................

do a good job of estimating A and C. B and D ................

are estimated accurately with N4SID Algorithm
I, but not with Algorithm 2 (for small i). The
bias can be calculated through. The accuracy of
the stochastic subsystem is strongly dependent
on i for both algorithms.
Effect of j.
Throughout the paper, we assumed that j--~ oo.
The effect of finite j is shown in Fig. 3f. This
figure shows the standard deviation (stars) of the
estimate of A as a function of j (i = 5). For each
j, 100 Monte-Carlo experiments were done. The
standard deviation is proportional to 1/V~
(dashed line). So, the accuracy of the results
depend on j as 1/V~.
9.2. A glass oven
The glass oven has three inputs (two burners
and one ventilator) and six outputs (tempera-
ture). The data have been pre-processed:
detrending, peak shaving, delay estimates,
normalization (Backx, 1987).
Using 700 data points, five different models
are identified (the 300 following points are used
as validation data):
(1) A state space model with N4SID Algo-
rithm I of Section 5.
(2) A state space model with N4SID Algo-
rithm 2 of Section 6.
(3) A state space model with the algorithm of
Section 8.1.
(4) An A R X model (Ljung, 1991) followed by
a balanced truncation.
(5) A prediction error model (Ljung, 1991).
The results are summarized in Table 1.
-i......L...:i i..:L...i:i :.ii~.:.'..:.'.i
ir.:.,~i.:i
::i .:.:.:..
10-t i ,
10 15 20 25
Order
FIG. 4. O r d e r decision based on the dominant singular
values.
The first row indicates the chosen order.
Figure 4 shows the singular values that led to the
system order five for N4SID Algorithms 1 and 2.
The gap in the singular values is clearly visible.
For the A R X model, the parameters 'na/nb'
indicate that 'na' was a 6 x 6 matrix with as
entries 4, and 'nb' a 6 x 3 matrix with as entries
4 (see also Ljung, 1991). In this way, the
resulting state space model was of order 33. It
was reduced to a fourth order model using
balanced truncation of the deterministic sub-
system. This is basically the same philosophy as
for the N4SID algorithms, but the last ones do
not calculate the intermediate model explicitly.
For the prediction error model, we used the
identified model of Algorithm 1 in the
controllability canonical form (controllability
indices: 2, 2, 1. Number of parameters: 93) as an
initial value (the initial values obtained from
'canform' or 'canstart' of Ljung (1991) did not
converge). We restricted the number of itera-
tions to three (no improvement with more
iterations).
The third row shows the number of floating
point operations. Algorithm 2 needs about 5%
less computation (simpler algorithm). The

TABLE 1. COMPARISON OF FIVE IDENTIFICATION ALGORITHMS

Algorithm 1 2 3 4 5

Order 5 5 3/3 4/4 5

i 5 5 5 -- --
Flops 14,355,148 13,59(I ,364 24,056,676 41,744,504 330,529,255
Pred. determ. 47.41 47.95 48.56 45.72 47.34
Pred. Kalman 10.76 10.76 14.17 28.98 10.87
 N4SID: Identification of combined systems
calculation of the A R X model takes three times
as much computation. The calculation of the
prediction error model (only three iterations)
takes about 25 times as much computation! It
should be mentioned that none of the algorithms
was optimized in the n u m b e r of operations (nor
ours, nor the ones in the Matlab toolbox), but
these figures still give an idea of the order of
magnitude of the n u m b e r of computations. The
fourth row is the error (in per cent) between the
measured validation output and the simulated
output using only the deterministic subsystem.
The fifth row shows the error between measured
and Kalman filter one step ahead prediction.
With [Y~k']i the ith validation output channel
and [Y~]i the ith simulated output channel, the
error is defined as (N,, = 300)
1 " kS.= ([Yk]i -- [Yk],)
. . . . O/~.
E 100 kEL
The best deterministic prediction is obtained
with the A R X model, which is logical since it
came about by balanced truncation of the
deterministic subsystem. The other models
perform a little worse (but all four about the
same).
The Kalman filter error is the smallest though
for the first two N4SID algorithms. The
projection retains the deterministic and
stochastic information in the past useful to
predict the future. This implies that the resulting
models will perform best when used with a
Kalman filter. The prediction error model does
not improve this Kalman filter prediction error
at all. There is even a slight decline which is
probably due to numerical errors since during
the iteration procedure a warning for a badly
conditioned matrix was given. This means that,
even though there is no optimization involved,
the results obtained with the N4SID algorithms,
for this industrial M I M O process are close to
optimal.
For this example, the N4SID algorithms thus
calculate a state space model without any a
priori fixed parametrization, and this in a fast
and numerically reliable way.
Finally, we should mention that, as we found
out in discussion with Lennart Ljung, the
prediction error method can be made to work
better with some extra manipulations. The main
problem with the prediction error method is that
the outputs are pairwise collinear. This results in
a very hard, ill conditioned optimization
problem, for every possible parametrization.
This problem can be solved as follows: first
91
estimate an initial model. Now, in the
optimization step, only the p a r a m e t e r s of the C
and D matrix corresponding to three outputs
can vary (output 1, 4 and 6 for this example).
The optimization is better conditioned now and
the resulting model has a similar iaerformance as
the subspace models for this example. The
number of floating point operations stays
extremely high though and this method is rather
complicated and requires quite some insight
from the identifiers.
10. CONCLUSIONS
In this p a p e r two new N4SID algorithms to
identify combined deterministic-stochastic linear
systems have been derived. The first one
calculates unbiased results, the second one is a
simpler biased approximation. The connection
between these N4SID algorithms and the
classical Kalman filter estimation theory has
been indicated. This allows interpretations and
proofs of correctness of the algorithms.
In future work, we will discuss other
connections with linear system theory: model
reduction properties, frequency interpretations
and connections with robust control.
The open problems that remain to be solved
are the connection with maximum likelihood
algorithms (Ljung, 1987) and the problem of
finding good asymptotic statistical error
estimates.
Acknowledgments--We would like to thank Lennart Ljung
and Mats Viberg for their help and useful comments.
REFERENCES
Anderson, B. D. O. and J. B. Moore (1979). Optimal
Filtering. Prentice-Hall, Englewood Cliffs, NJ.
Arun, K. S. and S. Y. Kung (1990). Balanced approximation
of stochastic systems. SIAM J. Matrix Analysis and
Applications, ll, 42-68.
Backx, T. (1987). Identification of an industrial process: a
Markov parameter approach. Doctoral Dissertation,
Technical University Eindhoven, The Netherlands.
De Moor, B. (1988). Mathematical concepts and techniques
for modeling of static and dynamic systems. Doctoral
Dissertation, Department of Electrical Engineering, Kath.
Universiteit Leuven, Belgium.
De Moor, B., P. Van Overschee and J. Suykens (1991).
Subspace algorithms system identification and stochastic
realization. Proceedings MTNS. Kobe, Japan. part 1,
589-594.
Enns, D. (1984). Model reduction for control system design.
Ph.D., dissertation, Dep. Aeronaut. Astronaut., Stanford
University, Stanford, CA.
Faure, P. (1976). Stochastic realization algorithms. In R.
Mehra and D. Lainiotis (Eds), System Identification:
Advances and Case Studies. Academic Press, NY.
Kailath, T. (19811). Linear Systerr~. Prentice-Hall, Engle-
wood Cliffs, NJ.
Kung, S. Y. (1978). A new identification and model
reduction algorithm via singular value decomposition.
Proc. of the 12th Asilomar Conference on Circuits, Systems
and Computers, Pacific Grove, 705-714.
92 P. VAN OVERSCHEE a n d B . D E MOOR

Larimore, W. E. (1990). Canonical variate analysis in and

identification, filtering, and adaptive control. Proceedings
of the IEEE 29th Conf. on Decision and Control, Hawaii.
596-6(M.
Ljung, L. (1987). System Identification: Theory for the User.
Prentice-Hall Information and System Sciences Series, = lim I(F,A'X~,+ F,A~U, + n ; ' v r + Y})
Englewood Cliffs, NJ.
Ljung, L. (1991). System Identification Toolbox. The x ((x~)'E + Up(H,)
' " ' + (Yv)' ' )
Mathworks, Inc.
= FiAipa~ + r,A;St(H~,) '
Moonen, M., B. De Moor, L. Vandenberghe and J.
Vandewallc (1989). On- and off-line identification of linear + FA:'s~E+ r,A;'R,,(HI)'
state space models. International Journal of Control, 49, + H~IS'E + HaiR~2(Hi) ' + H",
219- 232.
Moonen, M. and J. Ramos (1991). A subspace algorithm for = F,A'PaI~, + r,A;'R, ,(H,a)'
balanced state space system identification. ESAT/SISTA + r~A'S,(HI)'+ GA:!S',E
report 1991-(17, Kath. Universiteit Leuven, Dept. E.E.,
Belgium. Accepted for publication in IEEE Automatic + H;'R't2(H;t) ' + H;tS;E + r,A:.
Control (to appear 1993).
Moonen, M., P. Van Overschee and B. De Moor (1992). A Thus, we find
Subspace Algorithm for Combined Stochastic-
Deterministic State Space Identification. ESAT/SISTA
\li x mi It x m i li li/
report 1992-10, Kath. Universiteit Leuven, Dept. E.E., x

Belgium. [ F,A'S + r;A;l(RliRp)

Papoulis, A. (1984). Probability, Random Variables, and =[ + H;I(R~IzR..2) -
Stot'httYtic Processes. McGraw-Hill, NY.
Van Ovcrschcc, P. and B. De Moor (1991a). Subspace F,A i Pd E + EA,S + r,A;R,,(H,)
d d I
+ GA t S,(H;)
d I
]
algorithms for the stochastic identification problem. 30th +F,A;,S,tI~, + H;tR,2(H;;) , + H;,S~E j. (A.I)
IEEE Conference On Decision and Control, Brighton,
U.K., 1321-1326. The second part we need to express in terms of the system
Van Overschee, P. and B. De Moor (1991b). Subspace matrices is
Algorithms for the Stochastic Identification Problem.
ESAT/SISTA report 1991-26, Kath. Universiteit Leuven,
Dept. E.E., Belgium. Automatica, 29, 649-660. {u.] ,
Van Ovcrschec, P. and B. De Moor (1992). Two Subspace 90~',im ~ l uf/(u,,
Algorithms for the Identification of Combined ;~®I \ y . / u'rl "'-k~,T~,_)'
deterministic-Stochastic Systems. ESAT/SISTA report with
1992-34, Kath. Universiteit Leuven, Dept. E.E., Belgium. 1
Vcrhacgcn, M. (1991). A novel non-iterative MIMO state !im ~ YvlU;, U~-]
space model identification technique. Preprints 9th
IFAC/IFORS syn~posium on Identification and System = tim l ( r l x d + HdUp + Y~,)[Up U}]
parameter estimation, Budapest, Hungary, 1453-1458.
=r,S+HIIR,, R,,],
9~,.= l i m l ypY'p

APPENDIX A. PROOF OF THE PROJECTION = lira 1 (F;Xd + H:~Up+ Y~,)

THEOREM 1
x ((x;3'~ + U',,(H~)' + (Y;)')
P r o o f o f f o r m u l a (33).
= F,P"E + F,S,(H~)' + H~,S',~
Note. Duc to the complexity of the formulas, we use the
subscripts p (past) and f (future) throughout this appendix as + H'/R I j(H~I) ' + L",,
follows: For U, yd and Y', these subscripts denote
respectively the subscript 0 1i - 1 and i I 2i - 1. For X d and and ,~ll = R.
X' they denote the subscript 0 and i. Wc also make use of Note that 90 is guaranteed to be of full rank ((2m + I)i)
(9)-(12) without explicitly mentioning it. due to the persistently exciting input and the non-zero
Dclinc stochastic subsystem. To compute 90-t, we use the formula
for the inverse of a block matrix (Kailath, 1980)

Wc havc

,dr = ,lim-,} YIU;, - ~ - '.~-.,90[t' _, , (A.2)

= lira
1 t t
~ ( r , A ' X ; , + r , A ; U,, + H;'U~ + Y})U;, with ~0 = ~'-2 ~(~2.1~Ii190~1 "
-

So, in our case, this becomes

= F,A'S I + F,A;tR,I + H;IR',2,
: i v,u;
" 1

= lim I. (F,A'X;~ + r,A;'up + H;'U r + Y'I)U;

-V,7'(H~(I 0)+ F,S R - ' ) W-~ /
= F,A'S2 + F,A;tRt2 + H;tR22, (A.3)
 N 4 S I D : Identification o f c o m b i n e d s y s t e m s 93

with we have:
~,,: r , P " C + L 7 + r,S,(H'/)' + HiS',r', + HIR,,(Hia) '
6"f(i = i ~ (A - K , C ) S ( R I) . . . . . ij2aniUi}2~_l,
-(r,s+ ]('")i k =()
This quantity can only be calculated after the system is
= r A W - SR- ' S ' ) g + L;. identified. In the following it is thus assumed that the system
matrices A , B. C, D, Q", S', R ~ are kno%n. (R and Q are
From formulas (13) and (33) and the definitions of s~ and ~ , the submatrices of the R Q decomposition as defined in
we find (with/'---* ~) Section 7).

(L} L~ L~)=)Y,.I~_,(UI,I~_~_ , Y:,b- ,) • S R - i can be calculated as

SR -' = X",,U',,12 ,_ ,(U.l:,_ , U:,l~_ ,) - '

, )(u,,,2,_, = F~IY;',I,-, - H;'U.I,- ,IU:,I.~,-,(u,12,-, u'.p,_ ,)-'
=s~ -j. = r)[ 1",,I,-, U'.l~-, - H:~U,,I ~- , U'.I-'~ -, ](U.I:, -, UI,I.~.- ,) '

Using (A.I) and (A.3), we thus find

= r'~lr4:4.,:~- H:!R,:,.,:3IR,:'.,:~.

• For the calculation of the Kalman filter gains K~, we

(L~ L~)=F,XSR-' +F,A,a(/ 0)+n,a(0 t) also need the initial covariance estimate:
+ [r, AiS,(H:!)'+ F,AiSR - 'S, E
~ = pa + p,, _ S R " 'S'.
- r,A:!R,,(Hf)' + r, a i s ; r l , --We have approximately
+ n;!r;..(n~')' + H~,S'._U- r , a ' P " U
lira I ~ , ~ , = pa + p,
- r, A7 - F~A~tRI i(H~t) ' I-°~ ]

- r,A'S((H;')' - ria','s',r',- HIr;~(n'/)' and from the description of the algorithm (see Section 7)
- H",S'UIv/; '(n','(t 0) + r , s r - ' ) we have
2,2:
= (FiA~t HI!) + F I A i S R - t + Fi(A,(S R - i S,
J
- p a ) F ' , - AT)'t,O['(H~t(l O) + FiSR - t ) Thus we have
pa + p., = YI
= (r, a l - rixi~.'H:! hi)
+ F i A , S R - I _ rizilp [ 'FiSR - ' --SR IS' can be easily found as
= (r, la:' - QiH:!I H'/) + Film' - QiF, I S R - ' SR IS'¼SR I(Rl:3.,:3RSl:3.t:3)(SR- t )'
which is exactly the same as the first part of (33). = F~IR4:4.1::~- H~tR,:,.,3I
Once again, using (A.I) and (A.3), we find for L~ ' R'1:1.1:3~tH't)'l/FW
x 1R4:4.1:3- i I% t / '
L~ = [-F,A'S,( H:')' - riA','r, ,(Hi)' Care should be taken when subtracting both quantities
- H'.'R'~.(HI)' - r~A'SR -'S'I~, to obtain ~ , It is possible that due to the
approximations, ~b becomes negative definite. If this is
- F,A','(I O)S'U, H'/(O I)S'U
-

the case, it should be put equal to zero.

+ FiA'P'q", + F,A~ • Now the Kalman gains K, can be calculated using
+ FiA:'R,,(H:!)' + FiA'S,(H',')' formulas _(22)-(23).
The errors 6X, and 6Xi+ t can be easily calculated as
+ LA;~S',I", + H','R'..(HI)'+ H~S'U]V).' i-I
= F~(A'(P a - SR -'S')I~, + a~)~p,- ' 6"~i : H
k =|)
(A - K k C)S(R - I)m i ~. 1 1 2 , m R 2 : 3 . I::~Q I:.~.
= r,x,~,: i
and
= F,Q,
bf(i , i = [ I (A - Kn, C)S( R - t),,,(,, j)+ qz,,Ra:3.t:3a l: 3.
which is exactly the same as the second part of (33). V1 k --I)
The proof of formula (34) is completely analogous. El
The least squares solution of

rain + + b)D ll-"

,) Y,I, / \ U, li /lie
A P P E N D I X B. C A L C U L A T I O N O F T H E BIAS will thus be a better approximation for .se as £e is. The bias
For the calculation of the bias ~ - ~ the error on the can then be calculated as
states Xi and "~'i+i has to be calculated. From (49) and (50). a e - .9'= S e - _@.