Subspace System Identification Theory and Applications: Lecture Notes

SUBSPACE SYSTEM IDENTIFICATION
Theory and applications
Lecture notes
Dr. ing.
David Di Ruscio
Telemark Institute of Technology

Email: [email protected]
Porsgrunn, Norway
January 1995
6th edition December 11, 2009 Telemark University College

Kjølnes Ring 56
N-3914 Porsgrunn, Norway
2
Forord
The material in this report/book is meant to be used as lecture notes in practi-

cal and theoretical system identification and the focus is on so called subspace
system identification and in particular the DSR algorithms. Some of the ma-
terial and some chapters are based on published papers on the topic organized
in a proper manner.
On central topic is a detailed description of the method for system iden-
tification of combined Deterministic and Stochastic systems and Realization
(DSR), which is a subspace system identification method which may be used
to identify a complete Kalman filter model directly from known input and out-
put data, including the system order. Several special methods and variants of
the DSR method may be formulated in order to be used for the identification
of special systems, e.g., deterministic systems, stochastic systems, closed loop
system identification etc.
Furthermore basic theory as realization theory for dynamic systems based
on known markov parameters (impulse response matrices), practical topics as
effect of scaling, how to treat trends in the data, selecting one model from
different models based on model validation, ordinary least squares regression,
principal component regression as well as partial least squares regression, etc.
Parts of the material is used in the Mater Course, SCE2206 System Identi-
fication and Optimal Estimation at Telemark University College.
The material is also believed to be useful for students working with main
thesis on the subject and some of the material may also be used on a PhD
level. The material may also be of interest for the reader interested in system
identification of dynamic systems in general.
ii PREFACE
Contents
1 Preliminaries 1
1.1 State space model . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Inputs and outputs . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Data organization . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3.1 Hankel matrix notation . . . . . . . . . . . . . . . . . . . 2
1.3.2 Extended data vectors . . . . . . . . . . . . . . . . . . . . 3
1.3.3 Extended data matrices . . . . . . . . . . . . . . . . . . . 4
1.4 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Extended output equation . . . . . . . . . . . . . . . . . . . . . . 5
1.5.1 Extended vector output equation . . . . . . . . . . . . . . 5
1.5.2 Extended output matrix equation . . . . . . . . . . . . . . 6
1.6 Observability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.7 Extended state space model . . . . . . . . . . . . . . . . . . . . . 8
1.8 Controllability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.9 ARMAX and extended state space model . . . . . . . . . . . . . 10
2 Realization theory 13
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Deterministic case . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.1 Model structure and problem description . . . . . . . . . 13
2.2.2 Impulse response model . . . . . . . . . . . . . . . . . . . 14
2.2.3 Determination of impulse responses . . . . . . . . . . . . 14
2.2.4 Redundant or missing information . . . . . . . . . . . . . 17
2.2.5 The Hankel matrices . . . . . . . . . . . . . . . . . . . . . 18
2.2.6 Balanced and normalized realizations . . . . . . . . . . . . 20
iv CONTENTS
2.2.7 Error analysis . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.3 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 Combined Deterministic and Stochastic System Identification

and Realization - DSR: A subspace approach based on observa-
tions 1 27
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2 Preliminary Definitions . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.1 System Definition . . . . . . . . . . . . . . . . . . . . . . 29
3.2.2 Problem Definition . . . . . . . . . . . . . . . . . . . . . . 29
3.2.3 Matrix Definitions . . . . . . . . . . . . . . . . . . . . . . 30
3.2.4 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Extended State Space Model . . . . . . . . . . . . . . . . . . . . 31
3.4 System Identification and Realization . . . . . . . . . . . . . . . 32
3.4.1 Identification and Realization of System Dynamics . . . . 32
3.4.2 Realization of the Deterministic Sub-system . . . . . . . . 37
3.4.3 Realization of the Stochastic Sub-system . . . . . . . . . . 39
3.5 Implementation with QR Decomposition . . . . . . . . . . . . . . 44
3.5.1 Realization of A and D . . . . . . . . . . . . . . . . . . . 45
3.5.2 Realization of B and E . . . . . . . . . . . . . . . . . . . 46
3.5.3 Realization of C and ∆ . . . . . . . . . . . . . . . . . . . 47
3.5.4 Special Remarks . . . . . . . . . . . . . . . . . . . . . . . 49
3.6 Comparison with Existing Algorithms . . . . . . . . . . . . . . . 50
3.6.1 PO-MOESP . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.6.2 Canonical Variate Analysis (CVA) . . . . . . . . . . . . . 52
3.6.3 N4SID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.6.4 Main Differences and Similarities . . . . . . . . . . . . . . 54
3.7 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.7.1 Example 1: Monte Carlo Simulation . . . . . . . . . . . . 55
3.7.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
1
ECC95 paper extended with proofs, new results and theoretical comparison with existing
subspace identification methods. Also in Computer Aided Time Series Modeling, Edited by
Masanao Aoki, Springer Verlag, 1997.
CONTENTS v
3.7.3 Example 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.7.4 Example 4: Modeling the crude oil prize . . . . . . . . . . 66
3.7.5 Example 5: Modeling PH . . . . . . . . . . . . . . . . . . 72
3.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4 On the DSR algorithm 77

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 BASIC SYSTEM THEORETIC DESCRIPTION . . . . . . . . . 79
4.3 PROBLEM DESCRIPTION . . . . . . . . . . . . . . . . . . . . 80
4.4 BASIC DEFINITIONS . . . . . . . . . . . . . . . . . . . . . . . . 81
4.5 ORTHOGONAL PROJECTIONS . . . . . . . . . . . . . . . . . 82
4.6 MATRIX DEFINITIONS . . . . . . . . . . . . . . . . . . . . . . 83
4.7 BASIC MATRIX EQUATION IN SUBSPACE IDENTIFICATION 84
4.8 MATRIX EQUATION WITH STATES ELIMINATED . . . . . . 85
4.9 MATRIX OF INSTRUMENTAL VARIABLES W . . . . . . . . 86
4.10 Subspace identification of OL : autonomous systems . . . . . . . . 87
4.11 Subspace identification of OL : deterministic systems . . . . . . . 88
4.12 Subspace identification of OL+1 : “general” case . . . . . . . . . . 89
4.13 BASIC PROJECTIONS IN THE DSR ALGORITHM . . . . . . 90
4.14 COMPUTING PROJECTIONS BY QR DECOMPOSITION . . 91
4.15 COMPUTING PROJECTIONS WITH PLS . . . . . . . . . . . . 94
4.16 MODEL ORDER SELECTION BY EXAMPLE . . . . . . . . . 95
4.17 PROPERTIES OF THE ALGORITHM . . . . . . . . . . . . . . 96
4.18 COMPARISON WITH CLASSICAL APPROACHES . . . . . . 97
4.19 INDUSTRIAL APPLICATIONS . . . . . . . . . . . . . . . . . . 98
4.20 SOFTWARE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.21 Analysis and modeling of a scrubber . . . . . . . . . . . . . . . . 100
4.22 MONTE CARLO SIMULATION OF MIMO SYSTEM . . . . . . 101
4.23 STATISTICAL ANALYSIS . . . . . . . . . . . . . . . . . . . . . 103
4.24 IDENTIFICATION AND CONTROL OF TMP PROCESS . . . 104
4.25 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5 Subspace Identification for Dynamic Process Analysis and Mod-

eling 115
vi CONTENTS
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

5.2 Preliminary definitions . . . . . . . . . . . . . . . . . . . . . . . . 116
5.2.1 System definition . . . . . . . . . . . . . . . . . . . . . . . 116
5.2.2 Problem definition . . . . . . . . . . . . . . . . . . . . . . 117
5.3 Description of the method . . . . . . . . . . . . . . . . . . . . . . 117
5.4 Modeling the power . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.4.1 Refiner variables and model . . . . . . . . . . . . . . . . . 119
5.4.2 Numerical results . . . . . . . . . . . . . . . . . . . . . . . 120
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6 Dynamic Subspace Modeling (DSR) and Static Multivariate

Analysis and Regression (PCA,PCR,PLS) 129
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.2 System description and data organization . . . . . . . . . . . . . 131
6.2.1 Combined deterministic and stochastic description . . . . 131
6.2.2 Purely static description . . . . . . . . . . . . . . . . . . . 132
6.3 Dependent input or output data . . . . . . . . . . . . . . . . . . 134
6.4 The combined deterministic and stochastic problem . . . . . . . 135
6.4.1 Data compression . . . . . . . . . . . . . . . . . . . . . . 135
6.4.2 Identification of system dynamics . . . . . . . . . . . . . . 135
6.5 The static problem . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.5.1 Data compression . . . . . . . . . . . . . . . . . . . . . . 138
6.5.2 The residual . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.5.3 Effective rank analysis . . . . . . . . . . . . . . . . . . . . 139
6.5.4 Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.6 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . 140
7 On Subspace Identification of the Extended Observability Ma-

trix 143
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.3 Relationship between past and future data matrices . . . . . . . 144
7.4 The observability matrix . . . . . . . . . . . . . . . . . . . . . . . 144
CONTENTS vii
7.5 Standard deviation of the estimates . . . . . . . . . . . . . . . . . 148

7.6 PLS for computing the projections . . . . . . . . . . . . . . . . . 148
7.7 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.9 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.10 Appendix: proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.10.1 Proof of Equation (7.4) . . . . . . . . . . . . . . . . . . . 153
7.10.2 Proof I of Equation (7.10) . . . . . . . . . . . . . . . . . . 154
7.10.3 Proof II of Equation (7.10) . . . . . . . . . . . . . . . . . 155
7.10.4 Proof of Equation (7.22) . . . . . . . . . . . . . . . . . . . 156
7.10.5 Including constraints: regularization . . . . . . . . . . . . 157
7.10.6 Some useful results . . . . . . . . . . . . . . . . . . . . . . 158
7.10.7 Proof III of Equation (7.10) . . . . . . . . . . . . . . . . . 159
7.10.8 Prediction of future outputs . . . . . . . . . . . . . . . . . 159
8 Subspace identification of the states and the Kalman filter gain161

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
8.2 Notation and definitions . . . . . . . . . . . . . . . . . . . . . . . 162
8.2.1 System and matrix definitions . . . . . . . . . . . . . . . . 162
8.2.2 Hankel matrix notation . . . . . . . . . . . . . . . . . . . 163
8.2.3 Projections . . . . . . . . . . . . . . . . . . . . . . . . . . 164
8.3 Subspace system identification . . . . . . . . . . . . . . . . . . . 165
8.3.1 Subspace identification of the states . . . . . . . . . . . . 165
8.3.2 The extended observability matrix . . . . . . . . . . . . . 171
8.3.3 Identification of the stochastic subsystem . . . . . . . . . 174
8.3.4 SID of the deterministic subsystem . . . . . . . . . . . . . 177
8.4 Closed loop subspace identification . . . . . . . . . . . . . . . . . 179
8.4.1 Closed loop subspace identification: Using a filter in the
feeback loop! . . . . . . . . . . . . . . . . . . . . . . . . . 180
8.4.2 Closed loop subspace identification: Using the controller . 182
8.4.3 Closed loop subspace identification: Indirect method . . . 183
8.4.4 Closed loop subspace identification: Direct method . . . . 185
8.5 A new subspace identification method for closed and open loop
systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
viii CONTENTS
8.6 Further remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

8.6.1 Choice of algorithm parameters . . . . . . . . . . . . . . . 188
8.6.2 Choice of input signal . . . . . . . . . . . . . . . . . . . . 188
8.6.3 N4SID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
8.7 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . 189
8.7.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
8.7.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
8.7.3 Example 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
8.7.4 Example 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.7.5 Example 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
9 Effect of Scaling and how to Treat Trends 201

9.1 The data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
9.2 The data matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 201
9.3 Scaling the data matrices . . . . . . . . . . . . . . . . . . . . . . 202
9.3.1 Proof of Step 3 in Algorithm 9.3.1 . . . . . . . . . . . . . 203
9.3.2 Numerical Examples . . . . . . . . . . . . . . . . . . . . . 204
9.4 How to handle trends and drifts . . . . . . . . . . . . . . . . . . . 204
9.5 Trends and low frequency dynamics in the data . . . . . . . . . . 207
9.6 Time varying trends . . . . . . . . . . . . . . . . . . . . . . . . . 208
9.7 Constant trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
10 Validation 211
10.1 Model validation and fit . . . . . . . . . . . . . . . . . . . . . . . 211
10.1.1 Criteria based on the simulated error . . . . . . . . . . . . 211
10.1.2 Criteria based on the prediction error . . . . . . . . . . . 212
11 Input experiment design 213

11.1 Experiment design for dynamic systems . . . . . . . . . . . . . . 213
12 Special topics in system identification 219

12.1 Optimal predictions . . . . . . . . . . . . . . . . . . . . . . . . . 219
13 On the Partial Least Squares Algorithm 221

CONTENTS ix
13.1 Notation, basic- and system-definitions . . . . . . . . . . . . . . . 221

13.2 Partial Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . 222
13.2.1 Interpretation of the Partial Least Squares algorithm . . . 222
13.2.2 The Partial Least Squares decomposition . . . . . . . . . 233
13.2.3 The Partial Least Squares regression and prediction . . . 235
13.2.4 Computing projections from the PLS decomposition . . . 237
13.2.5 A semi standard PLS implementation . . . . . . . . . . . 238
13.2.6 The number of components a . . . . . . . . . . . . . . . . 238
13.3 Further interpretations of the PLS algorithm . . . . . . . . . . . 239
13.3.1 General results . . . . . . . . . . . . . . . . . . . . . . . . 239
13.3.2 Special case results . . . . . . . . . . . . . . . . . . . . . . 240
13.4 A covariance based PLS implementation . . . . . . . . . . . . . . 241
13.5 A square root PLS algorithm . . . . . . . . . . . . . . . . . . . . 242
13.6 Undeflated PLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
13.7 Total Least Squares and Truncated Total Least Squares . . . . . 244
13.7.1 Total Least Squares . . . . . . . . . . . . . . . . . . . . . 244
13.7.2 Truncated Total Least Squares . . . . . . . . . . . . . . . 245
13.8 Performance and PLS . . . . . . . . . . . . . . . . . . . . . . . . 246
13.9 Optimal solution . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
13.10Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . 247
13.11Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
13.12Appendix: proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
13.12.1 Proof of Equation (13.115) . . . . . . . . . . . . . . . . . 249
13.12.2 The PLS algorithm in terms of the W weight matrix . . . 250
13.13Matlab code: The PLS1 algorithm . . . . . . . . . . . . . . . . . 252
A Proof of the ESSM 255

A.1 Proof of the ESSM . . . . . . . . . . . . . . . . . . . . . . . . . . 255
A.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
A.3 Selector matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
d
A.4 A study of the predictor ZJ|L+1 for YJ|L+1 . . . . . . . . . . . . . 257
x CONTENTS
B Linear Algebra and Matrix Calculus 259

B.1 Trace of a matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
B.2 Gradient matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 259
B.3 Derivatives of vector and quadratic form . . . . . . . . . . . . . . 260
B.4 Matrix norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
B.5 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
B.6 Kronecer product matrices . . . . . . . . . . . . . . . . . . . . . . 261
C Optimization of the DSR algorithm 263

C.1 Existence of an optimal solution . . . . . . . . . . . . . . . . . . 263
C.2 Optimized algorithm for computing the B and E matrices . . . . 264
D D-SR Toolbox for MATLAB 269

D.1 The DSR function for use with MATLAB . . . . . . . . . . . . . 270
Chapter 1
Preliminaries
1.1 State space model
Consider a process which can be described by the following linear, discrete time
invariant state space model (SSM)
xk+1 = Axk + Buk + Cvk (1.1)

yk = Dxk + Euk + F vk (1.2)
where the integer k ≥ 0 is discrete time, xk ∈ Rn is the state vector, uk ∈ Rr is

the input vector, vk ∈ Rl is an external input vector and yk ∈ Rm is the output
vector. The constant matrices in the SSM are of appropriate dimensions. A
is the state transition matrix, B is the input matrix, C is the external input
matrix, D is the output matrix and E is the direct input to output matrix and
F is the direct external input to output matrix.
The following assumptions are stated:
• The pair (D, A) is observable.

£ ¤
• The pair (A, B C ) is controllable.
In some cases it is not desired to identify the direct input to output matrix
E. If E is apriori known to be zero then usually the other model matrices can
be identified with higher accuracy if E is not estimated. For this reason, define
the integer structure parameter g as follows
½
def 1 when E 6= 0m×r
g = (1.3)
0 when E = 0m×r
1.2 Inputs and outputs
First of all note that the definition of an input vector uk in the system identi-
fication theory is different from the vector of manipulable inputs uk in control
2 Preliminaries
theory. For example as quoted by Ljung (1995) page 61 and Ljung (1999)
page 523. Note that the inputs need not at all be control signals: anything
measurable, including disturbances, should be treated as input signals.
Think over the physics and use all significant measurement signals which is
available as input signals in order to model the desired output signals. With
significant measurement signals we mean input variables that describes and
observes some of the effects in the outputs.
For example if one want to model the room temperature, one usually gets
better fit when both the heat supply and the outside temperature is used as
input signals, than only using the heat supply, which in this case is an manip-
ulable input variable.
An example from the industry of using measured output signals and ma-
nipulable variables in the input vector uk is presented in Di Ruscio (1994).
Hence, input signals in system identification can be manipulable input vari-
ables as well as non-manipulable measured signals, including measurements,
measured states, measured properties, measured disturbances etc. The mea-
surements signals which are included in the input signal may of course be cor-
rupted with noise. Even feedback signals, i.e., input signals which are a function
of the outputs, may be included as inputs.
An important issue is to ask what the model should be used for. Common
problems are to:
• Identify models for control.

Important aspects in this case is to identify the dynamics and behaviour
from the manipulable input variables and possibly measured disturbances
upon the controlled output variables. The user must be careful to include
possibly measured states as input signals in this case. The reason for this
is that loss of identifibility of the behaviour from the manipulable inputs
to the outputs may be the result.
• Identify models for prediction of property and quality variables.
Generally speaking, the prediction and fit gets better when more measured
signals are included as inputs and worse when more outputs are added.
• Identify filters.
This problem is concerned with extraction of signals from noisy measure-
ments.
1.3 Data organization
1.3.1 Hankel matrix notation
Hankel matrices are frequently used in realization theory and subspace sys-
tem identification. The special structure of a Hankel matrix as well as some
notations, which are frequently used throughout, are defined in the following.
1.3 Data organization 3
Definition 1.1 (Hankel matrix) Given a (vector or matrix) sequence of data
st ∈ Rnr×nc ∀ 0, 1, 2, . . . , t0 , t0 + 1, . . . , (1.4)
where nr is the number of rows in st and nc is the number of columns in st .

Define integer numbers t0 , L and K and define tne the matrix St as follows
 
st0 st0 +1 st0 +2 ··· st0 +K−1

def  st0 +1 st0 +2 st0 +3 ··· st0 +K 

St0 |L =  .
. .. .. .. ..  ∈ RLnr×Knc . (1.5)
 . . . . . 
st0 +L−1 st0 +L st0 +L+1 · · · st0 +L+K−2
which is defined as a Hankel matrix because of the special structure. The integer
numbers t0 , L and K are defined as follows:
• t0 start index or initial time in the sequence st0 which is the upper left
block in the Hankel matrix.
• L is the number of nr-block rows in St0 |L .
• K is the number of nc-block columns in St0 |L .
A Hankel matrix is symmetric and the elements are constant across the anti-
diagonals. We are usually working with vector sequences in subspace system
identification, i.e., st is a vector in this case and hence,nc = 1.
Example 1.1 Given a vector valued sequence of observations
st ∈ Rnr ∀ 1, 2, . . . , N (1.6)
with N = 10. Choose parameters t0 = 3, L = 2 and use all observations to

define the Hankel data matrix S3|2 . We have
· ¸
s3 s4 s5 s6 s7 s8 s9
S3|2 = ∈ RLnr×7 . (1.7)
s4 s5 s6 s7 s8 s9 s10
The number of columns is in this case K = N − t0 = 7.
1.3.2 Extended data vectors
Given a number L known output vectors and a number L + g known input

vectors. The following extended vector definitions can be made
 
yk
 
def  yk+1 
yk|L =  .  ∈ RLm (1.8)
 .. 
yk+L−1
4 Preliminaries
where L is the number of block rows. The extended vector yk|L is defined as the
extended state space vector. The justification for the name is that this vector
satisfy an extended state space model which will be defined later.
 
uk
 uk+1 
 
def  
uk|L+g =  ...  ∈ R(L+g)m (1.9)
 
 uk+L+g−2 
uk+L+g−1
where L + g is the number of block rows. The extended vector uk|L+g is defined
as the extended input vector.
1.3.3 Extended data matrices
Define the following output data matrix with L block rows and K columns.
Known data matrix of output variables

z
 }| {
yk yk+1 yk+2 ··· yk+K−1
 yk+1 yk+2 yk+3 ··· yk+K 
def  
Yk|L =  .. .. .. .. ..  ∈ RLm×K (1.10)
 . . . . . 
yk+L−1 yk+L yk+L+1 · · · yk+L+K−2
Define the following input data matrix with L + g block rows and K columns.
Known data matrix of input variables

z
 }| {
uk uk+1 uk+2 · · · uk+K−1
 uk+1 uk+2 uk+3 · · · uk+K 
 
def  .. .. .. . . .. 
Uk|L+g =  . . . . .  ∈ R(L+g)r×K
(1.11)
 
 uk+L+g−2 uk+L+g−1 uk+L+g · · · uk+L+K+g−3 
uk+L+g−1 uk+L+g uk+L+g+1 · · · uk+L+K+g−2
1.4 Definitions
Associated with the SSM, Equations (1.1) and (1.2), we make the following
definitions:
• The extended observability matrix (Oi ) for the pair (D, A) is defined as
 
D
 
def  DA 
Oi =  .  ∈ Rim×n (1.12)
 .. 
DAi−1
where the subscript i denotes the number of block rows.

1.5 Extended output equation 5
• The reversed extended controllability matrix Cid for the pair (A, B) is
defined as
def £ ¤
Cid = Ai−1 B Ai−2 B · · · B ∈ Rn×ir (1.13)
where the subscript i denotes the number of block columns.
• A reversed extended controllability matrix Cis for the pair (A, C) is defined
similar to Equation (8.4),
def £ ¤
Cis = Ai−1 C Ai−2 C · · · C ∈ Rn×il (1.14)
i.e., with B substituted with C in Equation (8.4.
• The lower block triangular Toeplitz matrix (Hid ) for the quadruple matri-
ces (D, A, B, E)
 
E 0 0 ··· 0
 DB E 0 ··· 0 
 
def  DAB DB E ··· 0 
Hid =   ∈ Rim×(i+g−1)r
(1.15)
 .. .. .. .. .. 
 . . . . . 
DAi−2 B DAi−3 B DAi−4 B · · · E
where the subscript i denotes the number of block rows and i + g − 1 is

the number of block columns.
• A lower block triangular Toeplitz matrix His for the quadruple (D, A, C, F )
is defined as
 
F 0 0 ··· 0
 DC F 0 ··· 0 
 
def  ··· 0 
His =  DAC DC F  ∈ Rim×il (1.16)
 .. .. .. . . 
. . .. 
 . . .
DAi−2 C DAi−3 C DAi−4 C · · · F
1.5 Extended output equation
1.5.1 Extended vector output equation
The state space model, Equations (1.1) and (1.2), can in general be written as
the following extended vector output equation
d s
yk|L = OL xk + HL uk|L+g−1 + HL vk|L (1.17)
Proof: Stack L outputs defined by Equation (1.2) and substitute for the state
vector given by Equation (1.1). QED.
6 Preliminaries
1.5.2 Extended output matrix equation
The extended vector output equation can in general be written as the following
extended output matrix equation
d s
Yk|L = OL Xk + HL Uk|L+g−1 + HL Ek|L (1.18)
where
def £ ¤
Xk = xk xk+1 · · · xk+K−1 ∈ Rn×K (1.19)
Proof: The proof follows from the extended vector output equation (1.17).
QED.
1.6 Observability
Theorem 1.6.1 (Observability)

The system (1.1) and (1.2) is completely observable if the extended observability
matrix OL ∈RLm×n defined by Equation (8.3) satisfy
rank(OL ) = n
for all numbers L of block rows given by

½
def n − rank(D) + 1 when m < n
L ≥ Lmin = (1.20)
1 when m ≥ n
Proof:
We have from equation (1.17) that
d s
OL xk = yk|L − HL uk|L+g−1 − HL vk|L
Assume that the extended observability matrix has full rank. The state vector
can then be computed from
xk = (OLT OL )−1 OL
T d
(yk|L − HL s
uk|L+g−1 − HL vk|L )
if rank(OL ) = n and the right hand side is known. If rank(OL ) < n then only
a part of the state vector (xk ) can be observed. QED.
We will define OL as the extended observability matrix when the number
of block rows satisfy L > Lmin and only the sl observability matrix when the
number of block rows is L = Lmin .
The row size of the extended observability matrix OL is Lm. The minimal
number of block rows Lmin in the multiple output case follows from an argu-
mentation for when Lm ≥ n and for when rank(OL ) = n. In the single output
1.6 Observability 7
case with (m = rank(D) = 1). Then we must have L ≥ n in order for OL not
to loose rank.
The lower bound for the number of block rows in the observability matrix
is the ceiling function of n/m which is defined as the integer rounded towards
+∞ of the integer to integer ratio n/m. The upper bound must be n due to
the Cayley-Hamilton theorem. Hence, for an observable system the minimum
number of block rows in the observability matrix is bounded by
n
d e ≤ Lmin ≤ n (1.21)
m
It can be shown, Kalman, Falb and Arbib (1969), Ch.2, p. 38, that the rank of
the matrix OL must increase by at least one (maximum m) as each block row
is added to the observability matrix, until the maximal rank (n) is reached. If
rank(D) = m then it will be sufficient to include at most n − m block rows DA,
· · ·, DAn−m to see whether the maximal rank of OL can reach n. Hence, for an
observable system with rank(D) = m it is sufficient with
Lmin = n − m + 1
block rows in the observability matrix to ensure that rank(OLmin ) = n.

Finally to see that Lmin m ≥ n consider the multiple output case when
m = rank(D) and m ≤ n. From Theorem 1.6.1 we must have
Lmin m = (n − m + 1)m ≥ n
⇓
(n − m)(m − 1) ≥ 0
which prove that Lmin m ≥ n when m ≤ n.
Example 1.2 (Observability)

Given a system with m = 2 and n = 4
 
· ¸ 0 0 0 0
1 0 0 0  0 0 1 0 
D= ,A = 
 0
. (1.22)
0 1 0 0 0 0 1 
0 0 0 0
For this system the rank of the observability matrix O3 = [D; DA; DA2 ] will
increase by one each time the block row DA and DA2 is added until the maximal
rank 4 is reached, i.e., Lmin = 3 for this system. Changing elements in A so
that a14 = 1 and a24 = 1, i.e., for
 
0 0 0 1
 0 0 1 1 
A=  0 0 0 1 ,
 (1.23)
0 0 0 0
gives the necessary number of block rows, L, equal to only dn/me=2. This
means that rank(O2 ) = 3 in this last case. ¢
8 Preliminaries
1.7 Extended state space model
An extended state space model (ESSM) will be defined in this section.

The importance of the ESSM for subspace system identification is that it
gives us an equation where the states is eliminated. It leads us to an algorithm
for which the system model can be recovered from the known input and output
data. The importance of the ESSM for model predictive control is that it
facilitates a simple and general method for building a prediction model.
Theorem 1.7.1 (Extended state space model)

The SSM, Equations (1.1) and (1.2), can be described with an equivalent ex-
tended state space model (ESSM)
yk+1|L = ÃL yk|L + B̃L uk|L+g + C̃L vk|L+1 (1.24)
where the extended state space vector yk|L and the extended input vector uk|L+g
are defined in equations (1.8) and (1.9). The extended external input vector
vk|L+1 is defined similarly.
The integer number L which define the number of block rows in the extended
state vector must satisfy
L ≥ Lmin
where the minimal number of block rows is defined by

½
def n − rank(D) + 1 when m < n
Lmin = (1.25)
1 when m ≥ n
The matrices in the ESSM are given as

def T
ÃL = OL A(OL OL )−1 OL
T
∈ RLm×Lm (1.26)
def £ ¤ £ ¤
B̃L = OL B HLd d 0
− ÃL HL Lm×r ∈ RLm×(L+g)r (1.27)
def £ ¤ £ s ¤
C̃L = OL C HLs − ÃL HL 0Lm×l ∈ RLm×(L+1)l (1.28)
Proof: See Di Ruscio (1994).

The ESSM has some properties:
1. Minimal ESSM order. The minimal ESSM order Lmin , defined in

Equation (1.25), defines the minimal prediction horizon.
Proof: Same as proving observability, in this case (OL T OLmin )−1 is
min
non-singular and the existence of the ESSM can be proved.
2. Uniqueness. Define M as a non-singular matrix which transform the

state vector x to a new coordinate system. Given two system realiza-
tions, (A, B, C, D, F ) and (M −1 AM, M −1 B, M −1 C, DM, F ). The two
1.8 Controllability 9
realizations gives the same ESSM matrices for all L ≥ Lmin .

Proof: The ESSM matrices are invariant under state (coordinate) trans-
formations in the SSM.
3. Eigenvalues. The ESSM transition matrix (ÃL ) have the same (n) eigen-
values as the SSM transition matrix (A) and Lm − n eigenvalues equal to
zero.
Proof: From similarity, Equation 8.91.
In the subspace identification algorithm, Di Ruscio (1995), it is shown that

the ESSM can be identified directly from a sliding window of known data.
However, when the inputs used for identification are poor from a persistent
excitation point of view, it is better to only identify the minimal order ESSM
as explained in the paper. Note also that Lmin coincides with the demand for
persistent excitation of the inputs, i.e., the inputs must at least be persistent
exiting of order Lmin in order to recover the SSM from known input and output
data.
Assume that a minimal order ESSM is given and defined as follows
yk+1|Lmin = ÃLmin yk|Lmin + B̃Lmin uk|Lmin + C̃Lmin vk|Lmin +1 (1.29)
If the prediction horizon is chosen greater than Lmin it is trivial to construct
the following ESSM from the minimal ESSM.
By using yk+1|L−Lmin = yk+1|L−Lmin in addition to Equation (1.29) we ob-
tain the following ESSM model matrices
· ¸
0(L−Lmin )m×m I(L−Lmin )m×(L−1)m
ÃL = (1.30)
0Lmin m×(L−Lmin )m ÃLmin
· ¸
0(L−Lmin )m×Lr
B̃L = (1.31)
0Lmin m×(L−Lmin )r B̃Lmin
· ¸
0(L−Lmin )m×(L+1)l
C̃L = (1.32)
0Lmin m×(L−Lmin )l C̃Lmin
This last formulations of the ESSM matrices is attractive from an identifi-

cation point of view.
It will also have some properties with respect to noise retention compared to
using the matrices defined by Equations (8.91) to (8.93) directly in the model
predictive control algorithms which will be presented in a later chapter.
1.8 Controllability
Theorem 1.8.1 (Controllability)

The linear system (1.1) described by the pair (A, B) is completely controllable if
the extended controllability matrix CJd ∈Rn×Jr defined by Equation (8.4) satisfy
rank(CJd ) = n,
10 Preliminaries
for all numbers, J, of block rows given by

½
def n − rank(B) + 1 when r < n,
J ≥ Jmin = (1.33)
1 when r ≥ n.
1.9 ARMAX and extended state space model
Given a polynomial (ARMAX) model in the discrete time domain as follows
yk+L|1 = Ã1 yk|L + B̃1 uk|L+g + C̃1 vk|L+1 (1.34)
where the polynomial (ARMAX) model matrices are given by
def £ ¤
Ã1 = A1 A2 · · · AL ∈ Rm×Lm ,
def £ ¤
B̃1 = B0 B1 · · · BL−1+g BL+g ∈ Rm×(L+g)r , (1.35)
def £ ¤
C̃1 = C0 C1 · · · CL CL+1 ∈ Rm×(L+1)m .
where Ai ∈ Rm×m for i = 1, . . . , L, Bi ∈ Rm×r for i = 0, . . . , L + g, Ci ∈ Rm×m

for i = 0, . . . , L + 1. This model can be formulated directly as an ESSM which
have relations to the SSM matrices. We have
 
0 I ··· 0 0
 0 0 ··· 0 0 
 
 .. . .. .. . 
ÃL =  . .. . . ..  ∈ RLm×Lm (1.36)
 
 0 0 ··· 0 I 
A1 A2 · · · AL−1 AL
 
0 0 ··· 0 0 0
 0 0 ··· 0 0 0 
 
 .. .. .. .. .. .. 
B̃L =  . . . . . .  ∈ RLm×(L+g)r (1.37)
 
 0 0 ··· 0 0 0 
B0 B1 · · · BL+g−2 BL+g−1 BL+g
 
0 0 ··· 0 0 0
 0 0 ··· 0 0 0 
 
 .. . . . . . 
C̃L =  . .. .. .. .. ..  ∈ RLm×(L+1)m (1.38)
 
 0 0 ··· 0 0 0 
C0 C0 · · · CL−1 CL CL+1
Example 1.3 (Conversion of first-order ARMAX to ESSM)

Consider the ARMAX model
yk = ayk−1 + buk−1 + f ek + cek−1 (1.39)

1.9 ARMAX and extended state space model 11
An ESSM with L = 3 can be constructed as follows.
yk+1 = yk+1
yk+2 = yk+2
yk+3 = ayk+2 + buk+2 + f ek+3 + cek+2
which gives
 
          e
yk+1 0 1 0 yk 0 0 0 uk 0 0 0 0  k 
ek+1
 yk+2  =  0 0 1   yk+1  +  0 0 0   uk+1  +  0 0 0 0  
 ek+2


yk+3 0 0 a yk+2 0 0 b uk+2 0 0 c f
| {z } | {z }| {z } | {z } | {z } | {z } ek+3
yk+1|3 Ã3 yk|3 B̃ uk|3 C̃
| {z }
3 3
ek|L+1
(1.40)
The model is proper, i.e. g = 0. This can also be seen from the ARMAX model.
There is no direct feed-through term from the input uk to the output yk .
12 Preliminaries
Chapter 2
Realization theory
2.1 Introduction
A method for the realization of linear state space models from known system
input and output time series is studied. The input-output time series are usually
obtained from a number of independently input experiments to the system.
2.2 Deterministic case
2.2.1 Model structure and problem description
Assume that a system can be illustrated as shown in Figure (2.2.1). y ∈ <ny
Figure 2.1: Dynamic system with inputs and outputs
is the system outputs which are measured, u ∈ <nu is the system inputs which
can be manipulated, v̄ ∈ <nv is unknown system disturbances, and w̄ ∈ <ny is
unknown output (measurement) disturbances.
Assume that the system can be described by the discrete state space model
xi+1 = Axi + Bui + fi (v̄i ) (2.1)

yi = Dxi + gi (w̄i ) (2.2)
where i is discrete time, ie. an integer, x ∈ <nx is the state vector, and x0 is the
initial state. A, B and D are time invariant matrices of appropriate dimensions,
where (A, B) is a controllable pair and (D, A) is an observable pair.
The problem investigated in this work is to estimate the state space model
matrices A, B and D from known input and output time series. The input-
14 Realization theory
output time series are usually obtained from a number of independently input
experiments to the system. f (·) and g(·) are non-linear functions.
The derivation of the method will be based on the assumption that the
functions f and g are linear, ie.
fi (v̄i ) = C v̄i gi (w̄i ) = w̄i (2.3)
This is not a restriction, but a choice, which make it possible to apply stochastic
realization theory to estimate the noise covariance matrices, in addition to A,
B and D.
2.2.2 Impulse response model
A linear state space model xx+1 = Axk + Buk and yk = Dxk + Euk with the
initial state x0 given, can be described as the following impulse response model
k
X
yk = DAk x0 + DAk−i Bui−1 + Euk . (2.4)
i=1
We define the matrix
Hk−i+1 = DAk−i B ∈ Rny×nu , (2.5)
as the impulse response matrix at time instant k − i + 1. hence, the output, yk ,

at time instant k is defined in terms of the impulse response matrices H1 = DB,
H2 = DAB, . . ., Hk = DAk−1 B.
2.2.3 Determination of impulse responses
To determine the impulse responses a set of process experiments have to be

performed. uji is defined as the control input at time instant i for experiment
number j, and yij is the corresponding process output. The derivation is based
on information from mk = nu+1 experiments, however, this is not a restriction.
The derivation is based on the assumption that the noise vectors f and g
are the same for each experiment. In this case, exact impulse responses are the
result of this method.
In most cases, the noise vectors are different for each experiment, for exam-
ple when the disturbances are white noise. In this case errors are introduced,
and a crude estimate may occur. However, these errors may be reduced.
Time instant i = 1
x1 = Ax0 + C v̄0 + Bu0 (2.6)

y1 = D(Ax0 + C v̄0 ) + w̄1 + DBu0 (2.7)
2.2 Deterministic case 15
or in matrix form
· ¸
£ ¤ 1
y1 = z1 H1 (2.8)
u0
where
H1 = DB (2.9)
z1 = D(Ax0 + C v̄0 ) + w̄1 (2.10)
For nu + 1 experiments
· ¸
£ ¤ ones(1, nu + 1)
Y1 = z1 H1 (2.11)
U0
where
£ ¤ £ ¤
Y1 = y11 y12 · · · y1nu+1 U0 = u10 u20 · · · unu+1
0 (2.12)
Time instant i = 2
x2 = A(Ax0 + C v̄0 ) + C v̄1 + ABu0 + Bu1 (2.13)

2
y2 = D(A x0 + AC v̄0 + C v̄1 ) + w̄2 + DABu0 + DBu1 (2.14)
or in matrix form
 
£ ¤ 1
y2 = z2 H2 H1  u 0  (2.15)
u1
where
H2 = DAB (2.16)
2
z2 = D(A x0 + AC v̄0 + C v̄1 ) + w̄2 (2.17)
For nu + 1 experiments
 
£ ¤ ones(1, nu + 1)
Y2 = z2 H2 H1  U0  (2.18)
U1
where
£ ¤ £ ¤
Y2 = y21 y22 · · · y2nu+1 U1 = u11 u21 · · · unu+1
1 (2.19)
Time instants i = 1, ..., 2n

The results, for time instants 1 to 2n, may be stacked in the following matrix
system, which have a suitable structure for the determination of the unknown
matrices Hi and the vectors zi . Note that the information from time instant
i = 0 is not necessary at this point.
 
    ones(1, nu + 1)
Y1 z1 H1 0 0 ··· 0  
 Y2   z2 H2 H1 0 U0
   0 

 

 Y3   z3 H3 H2 H1 · · · 0   U 1 
 =  U2  (2.20)
 ..   .. .. . . . . . . .
.   
 .   . . . . . .  .. 
. 
Y2n z2n H2n · · · H3 H2 H1
U2n−1
This linear matrix equation may be solved recursively. Start with i = 1 and
solve for [z1 H1 ]. Then successively solve the i th equation for [zi Hi ] for
i = 1, ..., 2n. We have
· ¸ i−1
X
£ ¤ ones(1, nu + 1)
zi Hi = (Yi − Hi−k Uk ) (2.21)
U0
k=1
or
i−1
X · ¸−1
£ ¤ ones(1, nu + 1)
zi Hi = (Yi − Hi−k Uk ) (2.22)
U0
k=1
because the matrix is non singular by the definition of the input experiments
uj j = 1, ..., nu + 1.
We have now determined the impulse responses Hi and the vectors zi for
all i = 1, ..., 2n which contain information of noise and initial values. Hi and zi
can be expressed by
Hi = DAi−1 B (2.23)
i
X
zi = DAi x0 + DAi−k C v̄k−1 + w̄i (2.24)
k=1
Observe that the sequence zi ∀ i = 1, ..., 2n is generated by the following state

space model.
si+1 = Asi + C v̄i s0 = x0 (2.25)

zi = Dsi + w̄i (2.26)
where for the moment, z0 , is undefined. Define e = x − s, then
ei+1 = Aei + Bui e0 = 0 (2.27)

yi = Dei + zi (2.28)
The problem is now reduced to estimate the matrices (A, B, D), which satisfy
the model, Equations (2.27) and (2.28), from known impulse respenses Hi . In
addition, initial values and noise statistics can be estimated from the model
given by Equations (2.25) and (2.26) where zi and the matrices A, and D are
known.
z0 must be defined before continuing. At time instant i = 0 we have y0 =
Dx0 + w̄0 . For nu + 1 experiments
Y0 = Z0 (2.29)
£ ¤ £ ¤
Y0 = y01 y02 ··· y0nu+1 Z0 = z01 z02 · · · z0nu+1 (2.30)
We have assumed that the initial values are equal for all experiments. In this
case z0 is chosen as one of the columns in Y0 . z0 can even be computed from
· ¸−1
£ ¤ ones(1, nu + 1)
z0 0 = Y0 (2.31)
U0
However, this last method can have numerical disadvantages. It is a method

for constructing a right inverse for ones(1, nu + 1), and is nothing but the mean
of the columns of the matrix Y0 .
If the initial values are different for some or all nu + 1 experiments, then,
there are obviously many choices for z0 . One choice is the mean of the columns
of the matrix Y0 , which is consistent with the case of equal initial values, ie.
nu+1
X j
1
z0 = y0 (2.32)
nu + 1
j=1
2.2.4 Redundant or missing information
Assume that the number of experiments, m, are different from nu + 1. In this

case we have
i−1
X
£ ¤
zi Hi Ũ0 = (Yi − Hi−k Uk ) (2.33)
k=1
Assume that m ≥ nu + 1 and rank(Ũ0 Ũ0T ) = m, then, the solution of Equation

(2.33) with respect to Hi and zi is given by
i−1
X
£ ¤
zi Hi = (Yi − Hi−k Uk )Ũ0T (Ũ0 Ũ0T )−1 (2.34)
k=1
Assume that m < nu + 1 or rank(Ũ0 Ũ0T ) < m. An approximate solution is

determined by
i−1
X
£ ¤
zi Hi = (Yi − Hi−k Uk )Ṽ S̃ + Ũ T (2.35)
k=1
where
Ũ0 = Ũ S̃ Ṽ T (2.36)
is the singular value decomposition. In any cases the solutions discussed above
minimizes the 2 - norm of the error residual.
Note the connection to the theory of multivariate calibration, Martens and
Næss (1989), where usually a steady state model of the form y = Hu + z is
chosen to fit the data. The gain H is computed in the same way as outlined
above. Our method can thus be viewed as a dynamic extension of the theory
of multivariate calibration.
2.2.5 The Hankel matrices
The (block) Hankel matrices, which contains information of the system matri-
ces, can now be constructed from the impulse responses Hi ∀ i = 1, ..., L + J.
 
H1 H2 H3 · · · HJ
 H2 H3 H4 HJ+1 
 
 H5 · · · HJ+2 
H =  H3 H4  <ny(L+1)×nu·J (2.37)
 .. .. . . . . .
. 
 . . . . . 
HL+1 HL+2 · · · · · · HL+J
We extract the following submatrices of interests.

 
H1 H2 H3 · · · HJ
 H2 H3 H4 HJ+1 
 
 H3 H4 H · · · HJ+2 
H1|L = Hn =  5  <ny·L×nu·J (2.38)
 .. .. . .. . . . ... 
 . . 
HL HL+1 · · · · · · HL+J−1
 
H2 H3 H3 · · · HJ+1
 H3 H4 H4 HJ+2 
 
 H4 H5 H5 · · · HJ+3 
H2|L = HA =   <ny·L×nu·J (2.39)
 .. .. .. .. .. 
 . . . . . 
HL+1 HL+2 ··· ··· HL+J
 
H1
 H2 
 
 H3 
HB =   <ny·L×nu (2.40)
 .. 
 . 
HL
£ ¤
HD = H1 H2 H3 · · · HJ <ny×nu·J (2.41)
The realization theory is based on the observation that these matrices are closely
related to the observability matrix OL and the controllability matrix CJ given
by.
 
D
 DA 
 
 DA2 
OL =   <ny·L×nx (2.42)
 .. 
 . 
DAL−1
£ ¤
CJ = B AB A2 B · · · AJ−1 B <nx×nu·J (2.43)
The following factorization is the basic for the realization of the number of
states nx and the system matrices A, B and D.
Hn = H1|L = OL CJ (2.44)
At this stage it must be pointed out that the factorization of the Hankel ma-
trix by numerical methods, usually results in observability and controllability
matrices for a model representation in a different co-ordinate system than the
underlying system.
The number of states nx is estimated as the rank of Hn = H1|L . To obtain
a proper estimate we must ensure that
L ≥ nx − rank(D) + 1, J ≥ nx − rank(B) + 1. (2.45)
The reason for this is that, in this case, rank(OL ) = nx when the pair (D, A) is
observable, and similarly, rank(CJ ) = nx when the pair (A, B) is controllable.
OL and CJ may be determined by a suitable factorization of the known block
Hankel matrix, Hn = H1|L . The system matrices is then chosen to satisfy the
following three matrix relations.
H1|L = HA = OL ACJ , HB = OL B, HD = DCJ (2.46)
The system matrices are then estimated by

T
A = (OL OL )−1 OL
T
HA CJT (CJ CJT )−1 , (2.47)
T
B = (OL OL )−1 OL
T
HB , (2.48)
T T −1
D = HD CJ (CJ CJ ) . (2.49)
Assume for the moment that v̄ = w̄ = 0. The initial values satisfy, in this case
zn = OL Ax0 (2.50)
and can be estimated by
x0 = A−1 (OL
T
O)−1 OLT zn (2.51)
if A is non-singular. If A is singular then the information from time instant

i = 0, y0 = Dx0 + w̄0 , or z0 = y0 = Dx0 by assumption, can be added to the
algorithm to avoid the inversion. We have
T
x0 = (OL OL )−1 OLT z̃n (2.52)
where z̃n = [z0T , zn−1

T ]T is the vector where z is augmented to the first n − 1
0
elements in zn .
Note that the estimated model matrices usually are related to a different co-
ordinate system than the underlying model. Let x = T x̃ be the transformation
from the estimated state vector x̃ to the underlying state vector x. The observ-
ability and controllability matrices are given by ÕL = OL T and C̃J = T −1 CJ .
Which means that the transformation T is given by
T
T = (OL OL )−1 OLT ÕL , T −1 = C̃J CJT (CJ CJT )−1 . (2.53)
2.2.6 Balanced and normalized realizations
Perform the singular value decomposition, SVD, of the finite block Hankel ma-
trix Hn .
H1|L = Hn = OL CJ = U SV T = U S1 S2 V T . (2.54)
The order of the state space model is equal to the number of the non zero
singular values which is the same as the rank of Hn , ie. nx = rank(Hn ).
A reduced model is directly determined by choosing a subset of the non zero
singular values.
The following choices for the factorization of the Hankel matrix into the
product of the observability and controllability matrices results directly from
the SVD.
OL = U S1 , CJ = S2 V T (internally balanced),
OL = U, CJ = SV T (output normal), (2.55)
OL = U S, CJ = V T (input normal).
These factorizations are called internally balanced, output normal and input
normal, respectively, according to the definitions by Moore (1981). See also
Silverman and Bettayeb (1980). The meaning of these definitions will be made
clear later in this section.
If the internally balanced realization, as in Aoki (1990), is used, then the
system matrices are estimated by

A = S1−T U T HA V S2−T 
B = S1−T U T HB (internally balanced) (2.56)
−T 
D = HD V S 2
The output normal and input normal realizations are related to the co-ordinate
system, in which the internally balanced system is presented, by a transforma-
tion. These realizations are shown below for completeness.

A = U T HA V S −1 
B = U T HB (output normal) (2.57)

D = HD V S −1

A = S −1 U T HA V 
B = S −1 U T HB (input normal) (2.58)

D = HD V
The initial values may in all cases be estimated from Equation (2.51) if v̄ =
w̄ = 0.
Note that the matrix multiplications to perform B and D are not necessary,
because B is equal to the first nx x nu submatrix of CJ = U S1 and D is
equal to the first ny x nx submatrix of OL = S2 V T , in the internally balanced
realization.
Note that the SVD is not unique, in the sense that it may exist other
orthogonal matrices U and V satisfying Hn = U SV T . One can always change
sign for one column in U and in the corresponding column in V . However, S
is of course unique, if the singular values are ordered in descending order of
magnitude along the diagonal. This means that the estimated system matrices
are not unique, because the estimated matrices are dependent on the choice of
U and V , this is contradictory to the statement of uniqueness in Aoki (1990)
pp. 109.
However, the estimates have the properties that the L - observability and
the J - controllability grammians defined by
L
X
Wo,L = OLT OL = A(i−1)T DT DA(i−1) (2.59)
i=1
XJ
Wc,J = CJ CJT = A(i−1) BB T A(i−1)T (2.60)
i=1
satisfy
Wo,L = Snx Wc,J = Snx (internally balanced)

Wo,L = I 2
Wc,J = Snx (output normal) (2.61)
2
Wo,L = Snx Wc,J = I (input normal)
This can be shown by substituting the L - observability and the J - controlla-

bility matrices given in (2.55) into the grammians, Equations (2.59) and (2.60).
A model which satisfy one of the properties in (2.61) is called internally
balanced, output normal or input normal, respectively. This definitions are due
to Moore (1981).
The model produced by the method of Aoki (1990) is by this definition in-
ternally balanced. A model is called internally balanced if the corresponding
observability and controllability grammians are equal to the same diagonal ma-
trix. Particularly, a diagonal matrix with the non zero singular values of the
Hankel matrix on the diagonal, in this case. This imply that the new system is
”as controllable as it is observable”.
Note that the estimates (2.56) are not internally balanced due to the infinite
observability and the infinite controllability grammians Wo and Wc , which are
the grammians for a stable system when n → ∞ in Equations (2.59) and (2.60).
These grammians satisfy the discrete Lyapunov matrix equations
AT Wo A − Wo = −DT D, (2.62)
T T
AWc A − Wc = −BB , (2.63)
if A is stable.
The estimate of the process model is usually obtained from finite Hankel
matrices, i.e., for finite L and J. Note also that usually L = J. This model
estimate may be transformed to a model representation with balanced or nor-
malized grammians which satisfy the Lyapunov equations (2.62) and (2.63).
See Moore (1981), Silverman and Bettayeb (1980) and Laub (1980) for the
determination of such a transformation.
The statement that the internally balanced estimates are not unique will
now be clarified. Suppose that the model (A, B, D) is transformed to a new
co-ordinate system with x = T x̃. The grammians for the new system is then
given by
W̃o = T T Wo T, (2.64)
W̃c = T −1 Wc T −T . (2.65)
Assume that both realizations is internally balanced so that
T T Wo T = T −1 Wc T −T = Snx , (2.66)
Wo = Wc = Snx . (2.67)
This implies that
T −1 Snx
2 2
T = Snx , (2.68)
(2.69)
which constrain T to be a diagonal matrix with ±1 in each diagonal entry. The

argumentation shows that a model which is internally balanced is not unique.
However, the set of such models are constrained within sign changes.
2.2.7 Error analysis
When the order of the state space model is chosen less than the rank of the
Hankel matrix, errors is introduced.
The output sequence may be written as
i
X
yi = Hi−k uk−1 + zi (2.70)
k=1
where the impulse responses Hi and the noise terms zi are known. See Section
(2.2.3).
2.3 Numerical examples 23
The estimated state space model can be written in input output form as
i
X
ŷi = Ĥi−k uk−1 + ẑi (2.71)
k=1
where Ĥi are the impulse responses computed from the estimated state space
model, and ẑi is the output from an estimated model of the noise and initial
values based on the computed sequence zi .
The estimation error is given by
i
X
εi = ∆Hi−k uk−1 + ∆zi (2.72)
k=1
where ∆zi = zi + ẑi , which states that the output sequence can be written
yi = ŷi + εi (2.73)
2.3 Numerical examples
2.3.1 Example 1
Table 2.1: Input-output time series

t 1 2 3 4 5 6 7 8 9 10 11 12 13 14
u 0.1 0.1 0.1 0.1 5 6 7 8 9 10 0.01 12 13 0.1
y 0.10 0.01 0.05 0.06 0.01 0.01 0.01 0.02 0.01 0.01 0.01 0.01 0.01 0.01
2.4 Concluding remarks
Some (preliminary) comments are stated in the following items.
• It is not necessary that the time series are zero mean. If a linearized
model representation of a stable system around steady state values is the
goal, then, the mean steady state values of the inputs and outputs at time
instants i = 0 shuld be subtracted from the input and output time series.
• The method works for unstable systems, provided the time series are
finite. The method is not based on time series covariance estimation in
the usual way, ie. estimating Λk = E(yi+k yiT ) = DAk−1 M by applying
series methods or Fourier transform techniques (M contain information
of noise statistics), or cross covariance estimation, ie. estimating Hk =
E(yi+k uTi ) = DAk−1 BE(ui uTi ).
• The method works for non-minimum phase systems. If the time series are
generated from a state-output system (A, B, D), then the algorithm real-
ize an equivalent state-output system (Ã, B̃, D̃) with respect to impulse
responses, poles and zeros, even if the system is non-minimum phase. The
co-ordinate system in which the realized model is presented, is of course
generally different from the underlying co-ordinate system. The Markov
parameters are estimated from information of both u and y, not only y as
in standard realizing algorithms such as Aokis method. Recall that trans-
mission zeros cannot be observed from the outputs alone, which means
that realizing algorithms which are based on the estimation of Markov
parameters from y alone do not work for non-minimum phase systems.
• If the time series are generated from a transition-output system (A, B, D, E),
an essentially equivalent state-output system (Ã, B̃, D̃), with respect to
poles, zeros and impulse responses, is realized by this method. The in-
stantaneous dynamics from u to y in the transition-output system, are
realized with eigenvalues equal to zero in Ã, which means that states are
added which are infinitely fast. The spectrum of Ã is essentially equiva-
lent to the spectrum of A, exept for som zero poles added. The zeros of
the realization (Ã, B̃, D̃) is generally different from the zeros of an under-
lying system with realization (A, B, D, E). But the realization (Ã, B̃, D̃)
can be changed to a realization (Ā, B̄, D̄, Ē), with the same zeros as the
underlying system (A, B, D, E). This means that if the underlying system
(A, B, D, E) is non-minimum phase (zeros outside the unit-circle in the
z-plane), the realization (Ã, B̃, D̃) can lock like a minimum phase system
(zeros inside the unit-circle in the z-plane), hence; If there are zero eigen-
values (or zero in a relative sense) in the realization (Ã, B̃, D̃), change to
(A, B, D, E) system before computing zeros!
References
Aoki M. (1987). State Space Modeling of Time Series. Springer-Verlag Berlin,

Heidelberg.
Aoki M. (1990). State Space Modeling of Time Series. Springer-Verlag Berlin,

Heidelberg.
Laub, A. J. (1980). Computation of ”balancing” transformations. Proc.

JACC, San Francisco, FA8-E.
Martens, H. and T. Næss (1989). ”Multivariate Calibration”. John Wiley and

Sons Ltd. Chichester.
Moore, B. C. (1981). Principal Component Analysis in Linear Systems: Con-

trollability, Observability, and Model Reduction. IEEE Trans. on Auto-
matic Control, Vol. AC-26, pp. 17-31.
2.4 Concluding remarks 25
Silverman L. M. and M. Bettayeb (1980). Optimal approximation of linear

systems. Proc. JACC, San Francisco, FA8-A.
Chapter 3
Combined Deterministic and

Stochastic System
Identification and Realization
- DSR: A subspace approach
based on observations 1
Abstract
The paper presentes a numerically stable and general algorithm for identifica-
tion and realization of a complete dynamic linear state space model, including
the system order, for combined deterministic and stochastic systems from time
series. A special property of this algorithm is that the innovations covariance
matrix and the Markov parameters for the stochastic sub-system are determined
directly from a projection of known data matrices, without e.g. recursions of
non-linear matrix Riccatti equations. A realization of the Kalman filter gain
matrix is determined from the estimated extended observability matrix and the
Markov parameters. Monte Carlo simulations are used to analyze the statistical
properties of the algorithm as well as to compare with existing algorithms.
1
ECC95 paper extended with proofs, new results and theoretical comparison with existing
subspace identification methods. Also in Computer Aided Time Series Modeling, Edited by
Masanao Aoki, Springer Verlag, 1997.
28 Subspace identificatio
3.1 Introduction
System identification can be defined as building mathematical models of sys-

tems based on observed data. Traditionally a set of model structures with some
free parameters are specified and a prediction error (PE) criterion measuring
the difference between the observed outputs and the model outputs is optimized
with respect to the free parameters. In general, this will result in a non linear
optimization problem in the free parameters even when a linear time invari-
ant model is specified. A tremendous amount of research has been reported,
resulting in the so called prediction error methods (PEM).
In our view the field of subspace identification, Larimore (1983) and (1990),
Verhagen (1994), Van Overschee and De Moor (1994), Di Ruscio (1994), not
only resolves the problem of system identification but also deals with the addi-
tional problem of structure identification. In subspace identification methods a
data matrix is constructed from certain projections of the given system data.
The observability matrix for the system is extracted as the column space of this
matrix and the system order is equal to the dimension of the column space.
Descriptions of the advantages of subspace identification methods over tra-
ditional PEM can be found in Viberg (1995) and in Van Overschee (1995).
Aoki (1990) has presented a method for the realization of state space linear
discrete time stochastic models on innovations form. See also Aoki (1994) for
some further improvements of the method. This method has many interesting
numerical properties and many of the numerical tools involved in the method
are common with the tools used by the subspace identification methods. The
method is based on the factorization of the Hankel matrix, constructed from
covariance matrices of the output time series, by singular value decomposition.
The states are presented relative to an internally balanced coordinate system,
Moore (1981), which has some interesting properties when dealing with model
reduction.
The subspace identification methods which are presented in the above refer-
ences are using instrumental variables constructed from past input and output
data in order to remove the effect of noise from future data. It should be pointed
out that the method by Aoki also uses instrumental variables constructed from
past data.
The method for system identification and state space model realization
which is presented in this work is believed to be a valuable tool for the analysis
and modeling of observed input and output data from a wide range of systems,
in particular combined deterministic and stochastic dynamical systems. Only
linear algebra is applied in order to estimate a complete linear time invariant
state space model.
Many successful applications of data based time series analysis and modeling
methods are reported. One industrial application is presented in Di Ruscio and
Holmberg (1996). One particularly important application is the estimation of
econometric models, see Aoki (1990).
3.2 Preliminary Definitions 29
The remainder of the paper is organized as follows. Section 3.2 gives a def-
inition of the system and the problem considered in this work. In Section 3.3
the data is organized into data matrices which satisfy an extended state space
model or matrix equation. Section 3.4 shows how the system order and the
model matrices can be extracted from the known data matrices. A numerically
stable and efficient implementation is presented in Section 3.5. Section 3.6 gives
a comparison of the method presented in this work with other published meth-
ods. Real world numerical examples and theoretical Monte Carlo simulations
are presented in Section 3.7 and some concluding remarks follow in Section 3.8.
3.2 Preliminary Definitions
3.2.1 System Definition
Assume that the underlying system can be described by a discrete-time, time

invariant, linear state space model (SSM) of the form
xk+1 = Axk + Buk + Cek (3.1)

yk = Dxk + Euk + ek (3.2)
where the integer k ≥ 0 is discrete-time, x ∈ Rn is the state vector with initial

value x0 , y ∈ Rm is the system output, u ∈ Rr is the system input, e ∈ Rm
is an unknown innovations process of white noise, assumed to be covariance
stationary, with zero mean and covariance matrix E(ek eTk ) = ∆. The constant
matrices in the SSM are of appropriate dimensions. A is the state transition
matrix, B is the external input matrix, C is the Kalman gain matrix, D is
the output matrix and E is the direct control input to output (feed-through)
matrix. We will assume that (D, A) is an observable pair.
The innovations model, Equations (3.1) and (3.2), is discussed in e.g. Faurre
(1976) and Aoki (1990).
3.2.2 Problem Definition
The problem investigated in this paper is to identify a state space model, in-
cluding the system order (n), for both the deterministic and the stochastic part
of the system, i.e. the quadruple matrices (A, B, D, E) and the double matrices
(C, ∆) respectively, directly from known system input and output data vectors
(or time series) defined as
¾
uk ∀ k = 0, . . . , N − 1
Known data vectors
yk ∀ k = 0, . . . , N − 1
In continuous time systems the matrix E in Equation (3.2) is usually zero. This
is not the case in discrete time systems due to sampling. However, E can be
forced to be zero by including a structure constraint. This will be commented
on later.
3.2.3 Matrix Definitions
Associated with the SSM, Equations (3.1) and (3.2), we make the following
definitions:
• The extended observability matrix (Oi ) for the pair (D, A) is defined as
 
D
 DA 
def  
Oi =  ..  ∈ Rim×n (3.3)
 . 
DAi−1
where subscript i denotes the number of block rows. The matrix Oi

is denoted the extended observability matrix when the number of block
rows i is grater than the minimal number of block rows needed in the
observability matrix, in order to check if the system is observable.
• The reversed extended controllability matrix (Cid ) for the pair (A, B) is
defined as
def £ ¤
Cid = Ai−1 B Ai−2 · · · B ∈ Rn×ir (3.4)
where subscript i denotes the number of block columns.

A matrix Cis for the pair (A, C) is defined similarly to Equation (3.4), i.e.
with C substituted for B in the above definition.
• The lower block triangular Toeplitz matrix (Hid ) for the quadruple ma-
trices (D, A, B, E)
 
E 0 0 ··· 0
 DB E 0 ··· 0 
 
def  DAB DB E ··· 0 
Hid =   ∈ Rim×ir (3.5)
 .. .. .. .. .. 
 . . . . . 
where subscript i denotes the number of block rows.

A lower block triangular Toeplitz matrix His for the quadruple matrices
(D, A, C, F ) is defined as
 
F 0 0 ··· 0
 DC F 0 ··· 0 
 
def  DAC DC F ··· 0 
His =   ∈ Rim×im (3.6)
 .. .. .. .. .. 
 . . . . . 
where F = I for the output model formulation, Equation (3.2).

3.3 Extended State Space Model 31
3.2.4 Notation
The projection A/B of two matrices A and B is defined as AB T (BB T )† B where

† denotes the Moore-Penrose pseudo-inverse of a matrix.
3.3 Extended State Space Model
The state space model, Equations (3.1) and (3.2), can generally be written as
the following extended state space model (ESSM) (Di Ruscio (1994))
Yk+1|L = ÃYk|L + B̃Uk|L+1 + C̃Ek|L+1 (3.7)
where the known output and input data matrices Yk|L and Uk|L+1 are defined
as follows
 
yk yk+1 yk+2 ··· yk+K−1
 yk+1 y yk+3 ··· yk+K 
def  k+2 
Yk|L =  . .. .. .. ..  ∈ RLm×K (3.8)
 .. . . . . 
yk+L−1 yk+L yk+L+1 · · · yk+L+K−2
 
uk uk+1 uk+2 ··· uk+K−1
 uk+1 uk+2 uk+3 ··· uk+K 
 
def  .. .. .. .. .. 
Uk|L+1 =  . . . . .  ∈ R(L+1)r×K(3.9)
 
 uk+L−1 uk+L uk+L+1 · · · uk+L+K−2 
uk+L uk+L+1 uk+L+2 · · · uk+L+K−1
The unknown data matrix Ek|L+1 of innovations noise vectors is defined as

 
ek ek+1 ek+2 ··· ek+K−1
 ek+1 ek+2 ek+3 ··· ek+K 
 

def .. .. .. .. .. 
Ek|L+1 =  . . . . .  ∈ R(L+1)m×K(3.10)
 
 ek+L−1 ek+L ek+L+1 · · · ek+L+K−2 
ek+L ek+L+1 ek+L+2 · · · ek+L+K−1
The scalar integer parameter L defines the number of block rows in the data
matrices and the ESSM model matrices. The number of columns in Yk|L , Uk|L+1
and Ek|L+1 are K = N − L − k + 1. Each column in these matrices can be
interpreted as extended output, input and noise vectors, respectively. K can
be viewed as the number of samples in these extended time series. We also
have that L < K < N . L is the only necessary parameter which has to be
specified by the user. L is equal to the number of block rows in the extended
observability matrix (OL ∈ RLm×n ), which will be determined by the algorithm.
For a specified L, the maximum possible order of the system to be identified
is n ≤ Lm (if rank(D) = m, i.e. m independent outputs), or n ≤ Ld where
1 ≤ d = rank(D) ≤ m, i.e. d independent output variables.
The parameter L can be interpreted as the identification horizon. This

means that L is the horizon used to recover the present state space vector xk .
The matrices in the extended state space model, Equation (3.7), are related
to the underlying state space model matrices as follows
T
Ã = OL A(OL OL )−1 OL
T
(3.11)
£ ¤
B̃ = OB − ÃE1 E1 − ÃE2 E2 − ÃE3 · · · EL−1 − ÃEL EL
£ ¤ £ ¤
= OL B HLd − Ã HLd 0Lm×r (3.12)
£ ¤
C̃ = OL C − ÃF1 F1 − ÃF2 F2 − ÃF3 · · · FL−1 − ÃFL FL
£ ¤ £ ¤
= OL C HLs − Ã HLs 0Lm×m (3.13)
The matrices Ei and Fi , i = 1, ..., L, are block columns in the Toeplitz matrices
HLd and HLs defined in Equations (3.5) and (3.6), i.e.
£ ¤
HLd = E1 E2 · · · EL (3.14)
s
£ ¤
HL = F1 F2 · · · FL (3.15)
The importance of the ESSM, Equation (3.7), is that the state vector pre-
liminary is eliminated from the problem. Hence, the number of unknowns is
reduced. The ESSM also gives us the relationship between the data matrices
and the model matrices which at this stage are unknown.
This paper is concerned with the problem of reconstructing the system order
and system matrices in the state space model, (3.1) and (3.2), from the known
data matrices Yk|L and Uk|L+1 which satisfy Equation (3.7). We refer to Di
Ruscio (1994) and (1995) for a proof of the above results, which are the basis
for the method presented in this work.
Note that the matrices HLd and HLs satisfy the matrix equation
£ ¤ £ ¤
Yk|L = OL Xk + HLd 0Lm×r Uk|L+1 + HLs 0Lm×m Ek|L+1 (3.16)
where
£ ¤
Xk = xk xk+1 xk+2 · · · xk+K−1 ∈ Rn×K (3.17)
is a matrix of state vectors. Equation (3.16) is frequently used in other sub-

space identification methods, e.g. in Van Overschee and De Moor (1994) and
Verhagen (1994).
3.4 System Identification and Realization
3.4.1 Identification and Realization of System Dynamics
The basic step in the algorithm is to identify the system order and the extended
observability matrix from known data. In order to do so we will in this section
3.4 System Identification and Realization 33
derive an autonomous matrix equation from which the system dynamics can be
identified. We will show how the system order n and the extended observability
matrix OL are identified. A realization for the system matrices A and D and
the ESSM transition matrix Ã are then computed.
The term B̃Uk|L+1 can be removed from Equation (3.7) by post-multiplying
⊥
with a projection matrix Uk|L+1 ⊥
such that Uk|L+1 UL|L+1 = 0. The projection
matrix can e.g. be defined as follows
⊥ T T
Uk|L+1 = IK×K − Uk|L+1 (Uk|L+1 Uk|L+1 )−1 Uk|L+1 (3.18)
⊥
Hence, Uk|L+1 is the orthogonal projection onto the null-space of Uk|L+1 . A
numerically well posed way of computing the projection matrix is by use of the
singular value decomposition (SVD). The projection matrix is given by the left
singular vectors of Uk|L+1 which are orthogonal to the null-space. However, in
order to solve the complete system identification and realization problem, it is
more convenient to use the QR decomposition for computing the projection,
as will be shown in Section 3.5. Note that a projection matrix onto the null-
space of Uk|L+1 exists if the number of columns K in the data-matrices satisfies
K > L + 1.
⊥
Post-multiplying Equation (3.7) with the projection matrix Uk|L+1 gives
T
Yk+1|L − Yk+1|L Uk|L+1 T
(Uk|L+1 Uk|L+1 )−1 Uk|L+1 =
T
Ã(Yk|L − Yk|L Uk|L+1 T
(Uk|L+1 Uk|L+1 )−1 Uk|L+1 )+
T
C̃(Ek|L+1 − Ek|L+1 Uk|L+1 T
(Uk|L+1 Uk|L+1 )−1 Uk|L+1 ) (3.19)
Note that the last noise term in equation (3.19) is per definition zero as the
number of samples approaches infinity, i.e.
1
lim C̃ E UT =0 (3.20)
K→∞ K k|L+1 k|L+1
Hence, we have the following result
⊥ ⊥
Yk+1|L Uk|L+1 = ÃYk|L Uk|L+1 + C̃Ek|L+1 (3.21)
The noise term C̃Ek|L+1 can be removed from Equation (3.21) by post-multiplying
with K1 WiT where Wi is defined as a matrix of ”instrumental” variables which
are uncorrelated with Ek|L+1 , i.e., we are seeking for a matrix with the following
property
1
lim E WT = 0 (3.22)
K→∞ K k|L+1 i
An additional property is that Wi should be sufficiently correlated with the
informative part in the ESSM in order not to destroy information about e.g.
the system order.
An intuitive good choice is to use past data as instruments to remove future
noise. This choice ensures that the instruments are sufficiently correlated with
the informative part of the signals and sufficiently uncorrelated with future
noise.
Define J as the number of time instants in the past horizon which is used
for defining the instrumental variable matrix. Define L as the number of time
instants in the horizon necessary for identifying the state at time instant k, that
is xk , as well as the extended observability matrix of the system. Define M as
a prediction horizon. However, we will restrict ourself to the case M = 1 in this
work. These horizons are illustrated schematically in Figure 3.1.
Past horizon Future horizon Prediction

for instruments for identification horizon
¾ J - ¾ L -¾ M -
-
J −1 k k+L−1 k+L+M −1
Figure 3.1: Illustration of horizons involved in the DSR algorithm. Usually

k = J. That is, the end of the past is the beginning of the future.
Some alternative instruments, Wi , that will remove the noise and which
satisfy Equation (3.22) are as follows. Define
Wi ∈ Rni×K ∀ i = 1, 2, 3 (3.23)
where the row dimension ni is the number of instrumental variables and
· ¸
Y0|J
W1 = W2 = Y0|J W3 = Ul|J ∀ l≥0 (3.24)
Ul|J
The choice of instruments is unimportant in the deterministic case, i.e. when
the process as well as the observation noise are identically zero which also
means that ek = 0 for all discrete time instants. However, the “optimal” choice
of instruments is important in the combined deterministic and stochastic case.
Assume that Ul|J with l > 0 is chosen as the instrumental variable matrix.
This means that not only past inputs but also future inputs are used to remove
future noise (first time instant in the future horizon satisfy J ≤ k). Our ex-
perience from Monte Carlo simulations indicates that this is not an “optimal”
choice. Note also that the future inputs are already used in the projection
⊥
matrix Uk|L+1 . Hence, it make sense to use only past inputs as instruments,
i.e. the choice U0|J . It can also be shown (Verhagen, 1994) that by using only
past inputs as instruments, only the deterministic part of the model can be
recovered.
Past outputs are assumed to be uncorrelated with future noise. This gives
a first constraint on the discrete time instant k, i.e. k ≥ J. We have
1 T
lim Ek|L+1 Y0|J =0 ∀ k≥J (3.25)
K→∞ K
This statement can be proved from Equations (3.16) and (3.20). By incorpo-
rating past outputs as instruments we are also able to recover the stochastic
part of the model. Note that the states which are exited from the known inputs
are not necessarily the same as those which are exited from the unknown pro-
cess noise variables. It is necessary that all states are exited from both known
and unknown inputs and that they are observable from the output, in order to
identify them.
Hence, the following past inputs and past outputs instrumental variable
matrix is recommended to remove future noise from the model
· ¸
Y0|J
W1 = ∈ RJ(m+r)×K ∀ J ≥ 1 (3.26)
U0|J
A consistent equation for Ã is then given by the following autonomous matrix

equation
Zk+1|L = ÃZk|L ∀ k≥J (3.27)
where
def 1 ⊥ T
Zk+1|L = K Yk+1|L Uk|L+1 Wi ∈ RmL×ni (3.28)
def 1 ⊥ T
Zk|L = K Yk|L Uk|L+1 Wi ∈ RmL×ni (3.29)
Equation (3.27) is consistent because Wi , given by Equations (3.23) and (3.24),

satisfies Equation (3.22). See also Equation (3.25).
We can now prove that the column space of the matrix Zk|L coincides with
the column space of the extended observability matrix OL , when the identifica-
tion (future) horizon parameter L is chosen great enough to observe all states
and the past horizon parameter J is chosen adequately. Using Equations (3.16)
and (3.29) with the past inputs and past outputs instrumental variable matrix
gives
1 ⊥ 1 T
Zk|L = Yk|L Uk|L+1 W1T = OL Xk Uk|L+1
⊥
W ∈ RmL×J(m+r) (3.30)
K K 1
Assume that both the row and column dimensions of Zk|L are greater or equal
to the number of states, i.e., Lm ≥ n and J(m + r) ≥ n, and that L is chosen
such that the system is observable. The dimension of the column space of the
left hand side matrix must be equal to the system order, i.e. rank(Xk ) = n.
Hence,
1 1 T
rank(Zk|L ) = rank( Y U ⊥ W T ) = rank(OL Xk Uk|L+1
⊥
W ) = n (3.31)
K k|L k|L+1 1 K 1
The row constraints have a theoretical lower limit. From system theory we
know that a number of L ≥ n − rank(D) + 1 observations of the output is
sufficient in order to observe the states of a linear system, Kalman, Falb and
Arbib (1969), p. 37. However, the theoretical lower limit is the ceiling function
L = dn/me, defined as the integer ratio n/m rounded towards plus infinity.
From the column dimension we must ensure that the past horizon J for
defining the instrumental variable matrix must satisfy J(m + r) ≥ n. Hence,
the theoretical lower limit is J = dn/(m + r)e.
The maximum system order which can be specified by the user for a specified
choice of the parameter L is n = Lm. In this case the observability matrix can
be estimated as the column space of Zk|L only if the past horizon parameter J
is chosen such that J(m + r) ≥ Lm. A reasonable choice is therefore J = L.
Monte Carlo simulation experiments shows a relatively constant statistical
behavior of the estimates as a function of the past horizon parameter J. Hence,
we simply recommend putting J = L.
We have the following algorithm for analysis and modeling of system dy-
namics.
Algorithm 3.4.1 (System order, n, and the pair (D, A))

Given the positive integer parameters L and J and the matrices Zk+1|L and
Zk|L with k ≥ J which satisfy the autonomous matrix equation
Zk+1|L = ÃZk|L (3.32)

where
T
T
(3.33)
and OL is the extended observability matrix for the pair (A, D).
1. The system order n
Determine the Singular Value Decomposition (SVD)
Zk|L = U SV T (3.34)
where U ∈ RmL×mL , S ∈ RmL×ni and V ∈ Rni×ni are given by
· ¸
£ ¤ Sn 0 £ ¤
U = U1 U2 S= V = V1 V2 (3.35)
0 S2
where Sn ∈ Rn×n and n is the number of “non-zero” singular values of Zk|L ,
which is equal to the system order. n is determined by inspection of the “non-
zero” diagonal elements of S or SS T . The term U2 S2 V2T represents the error
by estimating the system order as the n first principal singular values.
2. The extended observability matrix OL for the pair (D, A)
The (extended) observability matrix can be taken directly as the first left part
in U , i.e. U1 . We have
OL = U (1 : Lm, 1 : n) = U1 (3.36)
3. The system matrix A
The system matrix A can be determined as
· −1 ¸
T Sn
A = OL Zk+1|L V = U1T Zk+1|L V1 Sn−1 (3.37)
0
4. The system output matrix D

The matrix D can be taken as the m × n upper sub-matrix in the observability
matrix OL , i.e.
D = U (1 : m, 1 : n) (3.38)
5. The extended system matrix Ã

We have
T
Ã = OL A(OL OL )−1 OLT = Zk+1|L V1 Sn−1 U1T (3.39)
Note that it can be difficult to estimate the system order as the non-zero or large
singular values of the matrix Zk|L when the signal to noise ratio is “small”. In
practice, the model order is chosen by a trial and error procedure combined with
both model validation and physical knowledge of the data generating process.
A procedure for chosing the model order from some statistical test is possibile,
see e.g. Ch. 9.6 in Aoki (1990).
TO =
We have chosen OL = U1 in Step 2 for simplicity, because we have OL L
In×n in this case. This gives an output normal realization when L → ∞. The
1
algorithm can also be formulated with the choice OL = U1 Sn2 which gives a bal-
anced realization when L → ∞. OL T O is equal to the observability grammian
L
as L tends to infinity because, in this case, DAL−1 tends to zero. A third choice
is OL = U1 Sn which gives an input normal realization. These definitions are
attributable to Moore (1981). These choices only represent different scalings
of the column space and give similar state space model matrices. The scaling
does not affect the statistical properties of the algorithm.
3.4.2 Realization of the Deterministic Sub-system
At this stage the system matrices A and D as well as the extended observability
matrix OL are known, see Section 3.4.1. In order to obtain a complete realiza-
tion for the deterministic part of the system we need to compute the system
matrices B and E.
There are many alternatives for extracting the B and E matrices. See e.g.
Section 3.5.2 for an alternative to the method presented below.
A consistent estimate of the B̃ matrix can be computed from
1 1
B̃ U UT = (Y T
− ÃYk|L )Uk|L+1 (3.40)
K k|L+1 k|L+1 K k+1|L
where Ã is determined from Algorithm 3.4.1, because

1
lim C̃ E UT =0 ∀ k≥0 (3.41)
K→∞ K k|L+1 k|L+1
For known data and system matrices A and D, Equation (3.40) can be written
as an over determined set of linear equations in the unknown system matrices
B and E .
T
We will in the rest of this section discuss the simpler solution when Uk|L+1 Uk|L+1
is non-singular. The matrix B̃ can be computed directly from Equation (3.40)
T
in this case. We refer to Section 3.5.2 for the case when Uk|L+1 Uk|L+1 is singu-
lar. The system matrices B and E can be extracted from B̃. E is given directly
as the lower right m × r sub-matrix of B̃. We have
E = B̃(m(L − 1) + 1 : mL, rL + 1 : r(l + 1)) (3.42)
B is given as a function of the block columns in B̃ and the matrices Ã and OL .
Ã and OL are known from Algorithm 3.4.1. Define
B̃i = B̃(1 : mL, r(i − 1) + 1 : ri) 1≤i≤L+1 (3.43)
as block column number i of B̃. We have
L+1
X
OL B = Ãi−1 B̃i (3.44)
i=1
and
L+1
X
B= Ai−1 (OLT OL )−1 OL
T
B̃i (3.45)
i=1
One strategy for recursively extracting the B, E, HLd and OL B matrices is as

follows.
Algorithm 3.4.2 Determination of HLd , OL B, E and B from known B̃, OL

and A.
Ã = OL A(OL T O )−1 O T
L L
EL = B̃L+1
E = EL (m(L − 1) + 1 : mL, 1 : m)
for i = 1, . . . , L
EL−i = B̃L−i+1 + ÃEL−i+1
if i < L
EL−i (1 : r(L − i − 1), 1 : r) = zeros(r(L − i − 1), r)
end
end
OL B = E0
B = (OL T O )−1 O T O B
L L L
The lower block triangular Toeplitz matrix HLd is given by the block columns
E1 , · · ·, EL according to Equations (3.14) and (3.5).
3.4.3 Realization of the Stochastic Sub-system
This section is mainly concerned with the problem of identifying the stochastic
part of the system. However, for natural reasons most of the results in Sections
3.4.1 and 3.4.2 are extended and resolved in parallel.
The ESSM, Equation (3.7), gives a relation between the future data matrices
and the SSM matrices. The following extension of the ESSM can be proved
Yk|L+1 = OL+1 Ak OJ† Y0|J

· ¸
£ d d
¤ U0|k
+ OL+1 Ck HL+1
Uk|L+1
h i· U0|J
¸
k † d
− OL+1 A OJ HJ 0(L+1)m×(k+L+1−J)r
UJ|k+L+1−J
· ¸
£ s s
¤ E0|k
+ OL+1 Ck HL+1
Ek|L+1
h i· E0|J
¸
k † s
− OL+1 A OJ HJ 0(L+1)m×(k+L+1−J)r (3.46)
EJ|k+L+1−J
Assume k = J for the sake of simplicity. Then we have
YJ|L+1 = OL+1 AJ OJ† Y0|J

h i· U ¸
0|J
+ OL+1 CJd − OL+1 AJ OJ† HJd d
HL+1
UJ|L+1
h ·
i E ¸
0|J
+ OL+1 CJs − OL+1 AJ OJ† HJs s
HL+1 (3.47)
EJ|L+1
This last equation nicely shows the connection between the past and future
data matrices and the unknown model matrices. As we will see, it also yields
some important results. We have the following projection
 ⊥
UJ|L+1
YJ|L+1  U0|J 
Y0|J
  ⊥ 
UJ|L+1
 E0|J  U0|J  
 
h i Y0|J

 
= OL+1 CJs − OL+1 AJ OJ† HJs s
HL+1   ⊥  (3.48)
 UJ|L+1 
 
 EJ|L+1  U0|J  
Y0|J
The projection of future inputs, past inputs and outputs onto the null-space of
future noise is equal to EJ|L+1 . We have separated the deterministic part of
the system from the data by the projection in Equation (3.48). Hence we have
the following theorem concerning the stochastic part of the system
Theorem 3.4.1 (Realization of the Toeplitz matrix HL+1 s )

s
The lower triangular Toeplitz matrix HL+1 with Markov parameters for the
stochastic sub-system is given by the projection of past inputs, past outputs,
and future inputs onto the null-space of the future outputs, i.e.
 ⊥
UJ|L+1
def
s
ZJ|L+1 = YJ|L+1  U0|J 
Y0|J
 ⊥  ⊥
UJ|L+1 UJ|L+1
s E
= PL+1 0|J
 U0|J  + H s EJ|L+1  U0|J  (3.49)
L+1
Y0|J Y0|J
where
s
PL+1 = OL+1 CJs − OL+1 AJ OJ† HJs (3.50)
and when N → ∞, then
 ⊥
UJ|L+1 · ¸⊥
s def   s U0|J s
ZJ|L+1 = YJ|L+1 U0|J = PL+1 E0|J + HL+1 EJ|L+1 (3.51)
Y0|J
Y0|J
s
which is a linear problem in HL+1 .
s
In order to see that this is a linear problem in the unknown matrix HL+1 , one
s
only needs to write up the equations obtained by using the structure of HL+1 ,
Equation (3.15), and a partition of the last term on the right hand side into
sub-matrices.
This result with proof is to our knowledge new. However, the same solution
resulting from a QR decomposition was presented in Di Ruscio (1995b).
From Theorem 3.4.1 and Equation (3.47) we can immediately state the
following important result. The extended observability matrix can be recovered
from the column space of the matrix on the left hand side of the following matrix
Equation (3.52).
Theorem 3.4.2 (Realization of the extended observability matrix OL+1 )
Given the following matrix equation

 
UJ|L+1
def
ZJ|L+1 = (YJ|L+1 /  U0|J )U ⊥
J|L+1 =
Y0|J
 
Y0|J
h i
 U
 0|J

  ⊥
OL+1 AJ OJ† d J † d s J
CJ − A OJ HJ CJ − A OJ HJ † s  UJ|L+1 U
 J|L+1
 E0|J /  U0|J  
Y0|J
(3.52)
then the column space of the matrix ZJ|L+1 coincides with the column space of
the extended observability matrix OL+1 and the system order n of the SSM is
given as the dimension of the column space.
The proof of Theorem 3.4.2 and Equation (3.52) is simple. From Theorem 3.4.1,
Equation (3.49), we have
s s s
YJ|L+1 − YJ|L+1 /Ỹ = PL+1 E0|J − PL+1 E0|J /Ỹ + HL+1 EJ|L+1 (3.53)
where
 
UJ|L+1
def
Ỹ =  U0|J  (3.54)
Y0|J
and for the sake of simplicity the other matrices are defined according to Equa-
tions (3.49) and (3.50). Substituting Equation (3.53) into equation (3.47) in
order to remove the stochastic term HL+1s EJ|L+1 gives
YJ|L+1 /Ỹ = OL+1 AJ OJ† Y0|J + PL+1

d s
U0|J + PL+1 d
E0|J /Ỹ + HL+1 UJ|L+1 (3.55)
⊥
Hence, an extra projection UJ|L+1 on the right hand side is necessary in order to
d
remove the deterministic term HL+1 UJ|L+1 and in order to recover the extended
observability matrix. The matrix on the right hand side of Equation (3.52) is
proportional with the extended observability matrix OL+1 . Hence, the column
space of the matrix ZJ|L+1 coincides with OL+1 . The dimension of the column
space of the matrix ZJ|L+1 is the order n of the SSM. QED.
From Theorem 3.4.2 we immediately have the following corollary concerning
the system dynamics.
Corollary 3.4.1 (Identification of system dynamics)

From ZJ|L+1 defined in Theorem 3.4.2, Equation (3.52) we have the following
relationship
ZJ+1|L = ÃZJ|L (3.56)
where
def T
T
and where ZJ+1|L is the L last (m × K) block rows in ZJ|L+1 and ZJ|L is the
L first (m × K) block rows in ZJ|L+1 , i.e.
   
UJ|L+1 UJ|L+1
def def
ZJ+1|L = (YJ+1|L /  U0|J )U ⊥
J|L+1 , ZJ|L = (YJ|L /
 U0|J )U ⊥
J|L+1
Y0|J Y0|J
Moreover, the column space of ZJ|L coincides with the column space of the
extended observability matrix OL and the dimension of the column space is equal
to the order n of the SSM. A realization of n, OL , A, D and Ã is determined
by using Equation (3.56) in combination with Algorithm 3.4.1.
The matrix on the left hand side of Equation (3.52) can be shown to be
equivalent to
ZJ|L+1 = YJ|L+1 Wc (3.57)

def ⊥
Wc = UJ|L+1 W1T (W1 UJ|L+1
⊥
W1T )−1 W1 UJ|L+1
⊥
Comparing Equation (3.57) with Equation (3.30) shows that these matrices are
related. We have shown that the column space of both equations coincides with
the column space of the extended observability matrix. The difference can be
viewed as a different column weighting matrix Wc , on the right hand side of
Equation (3.57). Equation (3.30) can be viewed as a special case factorization
⊥
of equation (3.57) with k = J and Wc = UJ|L+1 W1T /K. Later in Section 3.6,
we will show that this last Equation (3.57) is extremely important and useful
in order to partly compare other subspace identification methods.
From Theorem 3.4.1 we have the following result concerning the stochastic
part of the system.
Theorem 3.4.3 (Realization of ∆ and C̃)

Assume that the number of system input and output observations, N → ∞.
Define
 ⊥
UJ|L+1
def
s
ZJ|L+1 = YJ|L+1  U0|J  (3.58)
Y0|J
then we have
s s
ZJ+1|L = ÃZJ|L + C̃EJ|L+1 (3.59)
and
s s
ZJ+1|L (ZJ|L )⊥ = C̃EJ|L+1 (ZJ|L
s
)⊥
where
def £ ¤ £ ¤
C̃ = OL C HLs − Ã HLs 0Lm×m
The covariance matrix of the innovations ∆ is estimated directly from the col-
umn space of the left hand side matrix Equation (3.58), e.g.
s s
ZJ+L|1 (ZJ|L )⊥ = F EJ+L|1 (ZJ|L
s
)⊥ = F EJ+L|1 (3.60)
and
∆ = FFT (3.61)
when
T
EJ+L|1 EJ+L|1 = Im×m
4
The importance of Theorem 3.4.3 is that it shows that the innovations covari-
ance matrix can be estimated directly from the column space of the data ma-
trix, Equation (3.58). The factorization of the left hand side matrix of Equation
T
(3.60) into the product F EJ+L|1 and EJ+L|1 EJ+L|1 = Im×m can be performed
by the QR decomposition. See also Section 3.5. The Kalman filter gain matrix
C can be extracted when A and D is known.
Finally, we have the following Theorem 3.4.4 for the realization of the de-
terministic part of the system.
Theorem 3.4.4 (Realization of B̃)

Given
 
UJ|L+1
def
d
ZJ|L+1 = YJ|L+1 /  U0|J  (3.62)
Y0|J
then we have
d d
ZJ+1|L = ÃZJ|L + B̃UJ|L+1 (3.63)
and
d d
ZJ+1|L (ZJ|L )⊥ = B̃UJ|L+1 (ZJ|L
d
)⊥ (3.64)
where
def £ ¤ £ ¤
B̃ = OL B HLd − Ã HLd 0Lm×r
Theorem 3.4.4 is proved from Equation (3.55). Remark that the SSM matrix
E can be estimated directly from Equation (3.64). This is so because of the
structure of the ESSM matrix B̃.
All the projections in this section, Equations (3.49), (3.52) and (3.62) can
be effectively computed from a QR decomposition, either directly from the
projections defined in this section, or as will be shown in Section (3.5).
We will conclude this section by pointing out the relationship between the
data matrices and the projection matrices. The data matrix with future system
outputs and the projection matrices, (3.49) and (3.62) are related as
d s
YJ|L+1 = ZJ|L+1 + ZJ|L+1
Z d represents the outputs from the deterministic part of the system. Z s rep-
resents the outputs from the stochastic part of the system. YJ|L+1 is the data
matrix with future outputs from the combined deterministic and stochastic
system.
3.5 Implementation with QR Decomposition
We will here use the QR decomposition in order to compute the column space
of the projection matrices derived in Section 3.4. The QR decomposition is
also used in the subspace identification methods by Verhagen (1994) and Van
Overschee and De Moor (1994).
Define the following QR decomposition
1
√ Ỹ = RQ =
K
    
Uk|L+1 R11 0 0 0 Q1
1  Wi     
√   =  R21 R22 0 0   Q2  (3.65)
K  Yk|L
  R31 R32 R33 0   Q3 
Yk+1|L R41 R42 R43 R44 Q4
where
R ∈ R(r(L+1)+ni+2mL)×(r(L+1)+ni+2mL) (3.66)
(r(L+1)+ni+2mL)×K
Q ∈R (3.67)
Note that this decomposition could perhaps more precisely have been defined
as a lower Left, Q-orthogonal (LQ) decomposition. See Golub and Van Loan
(1983) for the computation.
The QR decomposition can be viewed as a data compression step. The
data matrix Ỹ which usually has a large number of columns is compressed to
a usually much smaller lower triangular matrix R which contains all relevant
information of the system for which the data was generated. As we will show,
the orthogonal Q matrix is not needed in the algorithm.
Note that the first (L − 1)m rows in Yk+1|L are equal with the last (L − 1)m
rows in Yk|L . This means that Yk+1|L can be substituted for Yk+L|1 in the QR
decomposition, Equation (3.65). This is utilized in the efficient implementation
of the DSR algorithm. However, for the sake of simplicity we will present the
results according to Equation (3.65).
By definition, the instrumental variable matrix Wi is uncorrelated with
Ek|L+1 . We can therefore remove the noise matrix Ek|L+1 from Equation (3.7)
by post-multiplying with K1 WiT . We have from (3.65) that
· ¸
1 1 £ ¤ T
R21
lim Ek|L+1 WiT = ( lim Ek|L+1 QT1 QT2 ) T =0 (3.68)
K→∞ K K→∞ K R22
Post-multiplying Equation (3.7) with [QT1 QT2 ], using (3.68) and substituting for
the corresponding Rij sub-matrices from (3.65) gives
£ ¤ £ ¤ £ ¤
R41 R42 = Ã R31 R32 + B̃ R11 0 (3.69)
3.5 Implementation with QR Decomposition 45
which gives one matrix equation for Ã and one for both Ã and B̃. We will in the
next Sections 3.5.1 and 3.5.2 show how the order n and the system quadruple
(A, B, D, E) are computed from (3.69).
The stochastic part of the system, defined by the matrices C and ∆, is
computed from
R43 − ÃR33 = C̃Ek|L+1 QT3 (3.70)
R44 = C̃Ek|L+1 QT4 (3.71)
This will be shown in Section 3.5.3.
Note that the QR decomposition compresses the possible large data matri-
ces into a number of (smaller) matrices which contain the information of the
system. It is also interesting that the matrices (information) which define the
deterministic part (A, B, D, E) and the stochastic part (∆, C) are separated by
the QR decomposition. The user must specify the parameter k ≥ J in Equation
3.65. See Figure 3.1 for a definition of the horizons involved. We recommend
putting k = J. The matrix W1 with l = 0, Equation (3.24), is recommended
for Wi in 3.65.
3.5.1 Realization of A and D
We have from Equation (3.69) that

R42 = ÃR32 (3.72)
and we choose
Zk+1|L = R42 (3.73)
T
Zk|L = R32 = U SV (3.74)
in Algorithm (3.4.1) in order to determine A, D and the extended observability
matrix OL . The system order is determined by inspection of the dominant
singular values of S or SS T .
Note that the first (L − 1)m rows in R42 are equal to the last (L − 1)m rows
in R32 . This is utilized in the efficient implementation of the DSR algorithm,
in order to reduce the computational work.
Note also that if Ã is computed as the projection of R32 onto R42 then Ã
takes a special canonical form. This is due to the common rows.
The A matrix may also be determined as follows. From the extended ob-
servability matrix we have
def
O1 = OL (1 : (L − 1)m, 1 : n) (3.75)
def
O2 = OL (m + 1 : Lm, 1 : n) (3.76)
then
A = (O1T O1 )−1 O1T O2 (3.77)
However, we must put L =: L + 1 in this case if the extended observability

matrix is estimated as the left singular vectors in (3.74). This will increase the
computational work. This last method is the so called shift invariance method
for computing the transition matrix A from the extended observability matrix,
Kung (1978). The shift invariance method is used in the subspace algorithms
(N4SID) by Van Overschee and De Moor (1994) and (MOESP) by Verhagen
(1994). The parameter which defines the number of block rows in the N4SID
and MOESP algorithms is denoted I. This parameter is related to the DSR
parameter L as I = L + 1. This is one of the differences between the DSR
algorithm and the N4SID and MOESP algorithms.
This means that N4SID and MOESP computes a number Im = (L + 1)m
of singular values. However, the system order can only be chosen according
to Lm of these singular values, i.e. the maximum system order which can
be chosen for a specified parameter I = L + 1 is n = Lm. For comparison,
the DSR algorithm computes only a number Lm of singular values and the
maximum system order which can be chosen for a specified parameter L is
n = Lm. Hence, the DSR algorithm seems to be more consistent with respect
to choosing the system order as the number of non-zero singular values.
The shift invariance method can be included in the DSR algorithm, but
at a higher computational expense. This strategy can be described as follows.
The extended observability matrix can be estimated from the column space of
the matrix formed from R32 and R42 . Compute the SVD
· ¸ · ¸· ¸· T ¸ · ¸
R32 U11 U12 Sn 0 V1 U11 Sn V1T
= =
R42 U21 U22 0 0 V2T U21 Sn V1T
We then have that A and D are determined from the shift invariance method,
e.g. from the left singular vectors as follows
OL = U11 (3.78)
OL A = U21 (3.79)
T
A = (U11 U11 )−1 U11
T
U21 (3.80)
D = U11 (1 : m, 1 : n) (3.81)
3.5.2 Realization of B and E
We have from Equation (3.69) that
B̃R11 = R41 − ÃR31 (3.82)
B̃ can be determined directly from (3.82) if the input u is persistently exiting

of order L + 1. R11 is non singular in this case. We have
T T −1
B̃ = (R41 − ÃR31 )R11 (R11 R11 ) (3.83)
The B and E matrices are then extracted from B̃ as pointed out in Section
3.4.2.
At this stage, the system order is identified (Algorithm 3.4.1). It is possible

to determine B and E if the input is only persistently exiting of order p + 1
where Lmin ≤ p ≤ L, directly without recomputing the algorithm with L =
Lmin or L = p. The minimal observability index, for a given system order, is
Lmin = n − rank(D) + 1 when n ≥ rank(D) and Lmin = 1 when n ≤ rank(D).
Define
Op = OL (1 : mp, 1 : n) (3.84)
Ãp = Op A(OpT Op )−1 OpT (3.85)
p
R11 = R11 (1 : r(p + 1), 1 : r(p + 1)) (3.86)
p
R31 = R31 (1 : mp, 1 : r(p + 1)) (3.87)
p
R41 = R41 (1 : mp, 1 : r(p + 1)) (3.88)
We then have
p p p
B̃p R11 = R41 − Ãp R31 (3.89)
This result is a consequence of Equation (3.7) with L substituted with p. Note

that the minimal observability index, for a given system order, is p = n −
rank(D) + 1 when n ≥ rank(D). The B and E matrices are then extracted
from B̃p as shown in Section 3.4.2 and Algorithm 3.4.2 with L substituted with
p.
3.5.3 Realization of C and ∆
Corollary 3.5.1 (Realization of C and ∆)

Consider the lower left triangular matrix R, determined by the QR decompo-
sition in Equation (3.65). An estimate of the square root of the innovations
covariance matrix is given by the m × m lower right sub-matrix of R, i.e.
F = R44 (m(L − 1) + 1 : mL, m(L − 1) + 1 : mL) (3.90)
and the estimate of the innovations covariance matrix is
∆ = FFT (3.91)
Furthermore, when J > 1, then an estimate of the Kalman filter gain matrix C
can be computed from
OL CF = R43 (1 : mL, 1 : m) (3.92)
If F is non-singular, then we have

T
C = (OL OL )−1 OL
T
R43 (1 : mL, 1 : m)F −1 (3.93)
An estimate of the lower left block triangular Toeplitz matrix HLs for the stochas-
tic subsystem (D, A, CF, F ) is given by
HLs = R43 (1 : mL, m + 1 : m(L + 1)) (3.94)

The estimate of the lower left block triangular Toeplitz matrix for the stochastic
subsystem (D, A, C, I), according to Equation (3.15), can be formed from the
block columns F1 , · · · , FL which can be computed from
£ ¤
F1 F F2 F · · · FL F = R43 (1 : mL, m + 1 : m(L + 1)) (3.95)
4
The stochastic subsystem is identified separately from the deterministic sub-

system. The necessary separation into deterministic and stochastic subsystems
are implicitly done by the QR decomposition.
The first (L − 1)m rows in Yk+1|L are equal to the (L − 1)m last rows in
Yk|L . Q4 is uncorrelated with Uk|L+1 , Wi , Yk|L and with the (L − 1)m first rows
in Yk+1|L . The first (L − 1)m rows in Yk+1|L QT4 and Ek|L+1 QT4 are therefore
zero. We then have from (3.65) and the structure of C̃, given by (3.13) and
(3.15), that
 
0 0 ··· 0
 0 0 0 
 
Yk+1|L QT4 = C̃Ek|L+1 QT4 = R44 =  . . . ..  (3.96)
 .. . . 
0 0 ··· F
where R44 ∈ RLm×Lm . Hence, the square root of the innovations noise process
covariance matrix is estimated directly as the m × m lower left matrix, denoted
F , in the lower triangular matrix R from the QR decomposition, Equation
(3.65). Note that F also is lower left triangular and can be compared to a
Cholesky factorization of ∆. This result is believed to be of some importance.
The result (3.91) is then clarified.
The matrices Qi , i = 1, 2, 3, 4, are orthogonal matrices and we have
 
F 0 ··· 0
 0 F 0 
1  
T  .. . . 
E Q = . . .
. .  ∈ R(L+1)m×Lm (3.97)
K k|L+1 3  
 0 0 ··· F 
0 0 ··· 0
Another strategy is then to compute OCF from Equations (3.70) and (3.97),
and an algorithm similar to Algorithm 3.4.2. This is formulated in the following
Corollary 3.5.2.
Corollary 3.5.2 (Realization of C)

Given the sub-matrices R43 and R33 from the QR decomposition in Equation
(3.67) and the ESSM transition matrix Ã. Define according to Equation (3.70)
˜ def
CF = R43 − ÃR33 (3.98)
˜ , e.g. by a procedure similar
The matrix OL CF can then be computed from CF
to Algorithm 3.4.2.
4
This strategy is found from Monte Carlo simulations to be the best one
when the past horizon parameter is J = 1, but no significant difference is found
when J > 1.
3.5.4 Special Remarks
One advantage of the QR implementation of the algorithm is that potential ill-

conditioning of the covariance matrices are concentrated in a certain triangular
matrix. This ill-conditioning usually results from ill-conditioned noise processes
(process noise and measurements noise) and from rounding-off errors. Note that
the triangular matrix R is the square root of the covariance matrix (H = K1 Ỹ Ỹ T
where Ỹ is defined in (3.65)) and that the triangular matrix is computed without
ever computing the covariance matrix. The method can therefore be defined as
numerically stable.
The QR decomposition is not unique. The R matrix is post-multiplied
by a diagonal permutation matrix E such that R := RE has positive diagonal
elements. The diagonal elements of E are equal to the sign of the corresponding
diagonal elements of R which was the result from the QR decomposition. Note
also that Q := EQ and EE = I. This gives a more unique coordinate system
for the estimated (A, B, D, E) quadruple. This scaling is also one of the reasons
for the simple solutions for the C and ∆ matrices in Section 3.5.3. The scaling
ensures that the diagonal blocks of (3.97) gets the same sign.
Note that common rows in the data matrices Yk|L and Yk+1|L can be removed
prior to obtaining the QR decomposition in equation (3.67). It is also clear
from the above that the orthogonal Q matrix (QQT = I) is not needed in the
algorithm. This will reduce the computational effort considerably. In fact, the
QR factorization works on a data matrix of size only r(2L + 1) + m(2L + 1) × K
and not of size r(2L + 1) + 3mL × K as indicated in (3.65).
Another strategy for determining R is to first compute H = K1 Ỹ Ỹ T where
Ỹ is defined in (3.65) and then the SVD, H = U SV T , followed by a QR de-
1
composition of U S 2 in order to obtain the lower triangular matrix, R. This
strategy reduced the number of flops and increased the accuracy of the R ma-
trix when MATLAB was used for the computations. However, no significant
difference in the estimated models was observed. This strategy can be numeri-
cally ill-conditioned due to possible rounding errors when forming the product
of rows in Ỹ with columns in Ỹ T in order to compute the correlation matrix
Ỹ Ỹ T . This strategy is therefore not recommended.
Equation (3.82) is defined by the triangular factors for Equation (3.40) with
k ≥ J.
T T
B̃Uk|L+1 Uk|L+1 = Yk+1|L Uk|L+1 − ÃYk|L Uk|L+1 (3.99)
It is also possible to extract the triangular factors for

T T
B̃U0|L+1 U0|L+1 = Y1|L U0|L+1 − ÃY0|L U0|L+1 (3.100)
directly from the QR decomposition (3.65) when k = J = L. The first block

row in UL|L+1 is equal to the last block row in U0|L+1 . Hence, the matrices in
Equation (3.100) can be defined from the lower triangular R matrix, equation
(3.65). It is therefore natural to choose
£ 0
¤ £ 0
¤ £ 0
¤
B̃ R11 R11 = R41 R41 − Ã R31 R31 (3.101)
for defining B̃ and an equation for computing the B and E system matrices.
Equation (3.101) consists of the triangular factors for both K1 Uk|L+1 Uk|L+1
T and
1 T
K U0|L+1 U0|L+1 . Equation (3.101) has effect for systems where the input signal
is poor with frequencies, but gives no additional effect compared to (3.82) for
e.g. white noise inputs.
Note that the stochastic part of the model is determined from QR and
SV Decompositions only. The Markov parameters and the square root of the
innovations covariance matrix are determined from a QR decomposition only.
The Kalman filter gain matrix is determined from the Markov parameters and
the extended observability matrix. No matrix Lyapunov or non-linear matrix
Riccati equations have to be solved.
The method has in this work been illustrated for systems which are not
strictly proper (only proper), i.e. the case when E 6= 0 in the underlying model
(3.2). The method can also be implemented to handle proper systems, i.e.
systems where E is known to be zero. This can be done by deleting the last
block row in Uk|L+1 and the last block column in B̃, see Equation (3.7).
3.6 Comparison with Existing Algorithms
A comparison with the DSR algorithm and three different subspace algorithms
will be given, namely N4SID, Van Overschee and De Moor (1994), CVA,
Larimore (1983), (1990) and PO-MOESP, Verhagen (1994).
The first and common step in subspace identification algorithms is to es-
timate the extended observability matrix from the column space of a known
data matrix. We will therefore concentrate our discussion on the similarities
and differences in which these methods estimate the extended observability ma-
trix. We will only briefly discuss how the system matrices are estimated by the
different methods.
It is shown in Van Overschee and De Moor (1995) that the different methods
are related through certain row and column weightings with the N4SID data
matrix as the key matrix.
We will below present a different approach, with the matrix ZJ|L+1 defined
in Section 3.4.3, Theorem 3.4.2 and Equation (3.57), as the key matrix.
Multiplying Equation (3.57) from left with an extra row weighting matrix
Wr and using the SVD as discussed in Algorithm 3.4.1 gives
Wr ZJ|L+g = Wr YJ|L+g Wc = U1 Sn V1T + U2 S2 V2T (3.102)

3.6 Comparison with Existing Algorithms 51
where g is an integer parameter and

⊥def
⊥
W1T )−1 W1 UJ|L+1
⊥
The extended observability matrix can be estimated as, e.g.
OL+g = Wr−1 U1 (3.103)
where Wr is a non-singular matrix. The matrix Wc is symmetric and can be

written as
def⊥ 1 1
⊥ ⊥
W1T )− 2 (W1 UJ|L+1 W1T )− 2 W1 UJ|L+1
⊥
(3.104)
| {z }| {z }
Wc2 (Wc2 )T
From the above factorization of the matrix Wc we have at least four matrices
(Wci , i = 1, · · · , 4) which all have essentially the same column space as Wc , i.e.
Wc1 = UJ|L+1
⊥ W1T (W1 UJ|L+1
⊥ W1T )−1 W1 UJ|L+1
⊥
1
Wc2 ⊥
= UJ|L+1 W1T (W1 UJ|L+1
⊥ W1T )− 2
Wc3 ⊥
= UJ|L+1 W1T (W1 UJ|L+1
⊥ W1T )−1 K
1
Wc4 ⊥ T
= UJ|L+1 W1 K
Wc5 ⊥
= UJ|L+1
The matrix Wc5 is sufficient only for purely deterministic systems and is shown
for the sake of completeness.
These column weighting matrices are used in the DSR algorithm which
is presented in this work. The past horizon parameter J is usually chosen as
J ≥ L. The parameter g = 0 in Equation (3.102) and the row weighting matrix
is the Lm × Lm identity matrix, denoted Wr = ILm . Algorithm 3.4.1 is used in
order to identify the extended observability matrix OL from the column space
of the matrix ZJ|L .
We will now illustrate the DSR algorithm’s similarity and difference to two
published algorithms, CVA by Larimore (1990) and PO-MOESP by Verhagen
(1994).
3.6.1 PO-MOESP
The PO-MOESP algorithm in Verhagen (1994) estimates the extended ob-

servability matrix OL+1 from Equations (3.102) and (3.103) with the following
matrices
⊥ 
Wc = UL+1|L+1 W1T (W1 UL+1|L+1
⊥ W1T )−1 W1 UL+1|L+1
⊥

Wr = I(L+1)m PO-MOESP(3.105)

g=1
From Theorem 3.4.2 and the factorization in Equation (3.104) we conclude that
the two algorithms PO-MOESP and DSR estimate the extended observability
matrix from a known data matrix which essentially has the same column space.
The only difference is that PO-MOESP estimates the extended observability
matrix OL+1 (of larger size than DSR does) in order to use the shift invariance
method for extracting the system matrix A as explained in Section 3.5.1.
Using the triangular factors from the QR decomposition in Section 3.5, we
have
T T −1
R32 R22 (R22 R22 ) R22 Q2 = (U1 Sn V1T + U2 S2 V2T )Q2 (3.106)
where the orthogonal matrix Q2 is not needed because the column space can
be estimated as the matrix U1 .
A major difference is that the PO-MOESP algorithm does not estimate the
stochastic part of the model. DSR estimates the Kalman gain and innovations
covariance matrix directly from the data as shown in Sections 3.4.3 and 3.5.3.
3.6.2 Canonical Variate Analysis (CVA)
The CVA algorithm in Larimore (1990) estimates the extended observability

matrix OL+1 from Equations (3.102) and (3.103) with the following matrices
1 
⊥
Wc = UL+1|L+1 W1T (W1 UL+1|L+1
⊥ W1T )− 2 
⊥ T − 12 CVA (3.107)
Wr = (YL+1|L+1 UL+1|L+1 YL+1|L+1 ) 

g=1
As we can see, the column weighting matrix Wc used by the CVA algorithm
fits into the factorization in Equation (3.104). A difference is that the CVA
algorithm uses a row weighting matrix Wr .
The only difference is that the DSR algorithm takes the SVD of a matrix
of size only Lm × J(r + m) where usually J = L, in order to identify OL . The
other methods take the SVD of a matrix of size (L + 1)m × J(r + m), in order
to identify OL+1 . This is to separate out the sub-matrices OL and OL A from
OL+1 . See also Section 3.5.1 for a discussion.
From Theorem 3.4.2 and the factorization in Equation (3.104) we conclude
that the two algorithms CVA and DSR essentially have the same column
space.
An interpretation of the CVA algorithm is that the system order is esti-
⊥
mated as the number of principal angles between the matrix YL+1|L+1 UL+1|L+1
⊥
and W1 UL+1|L+1 different from π/2. The principal angles can be effectively
computed using the SVD, see e.g. Van Overschee (1995), p. 29 and Golub and
Van Loan (1989), p. 428.
By using the triangular factors as shown in Section 3.5 we get the following
method for computing the principal angles
1 1
T T −2 T T −2
(R32 R32 + R33 R33 ) R32 R22 (R22 R22 ) = U1 Sn V1T + U2 S2 V2T (3.108)
3.6 Comparison with Existing Algorithms 53
The system order is here identified as the number of singular values equal to
one.
The next step in the CVA algorithm is to define a memory which defines
a valid sequence of system states. The system matrices can then be estimated
from a least squares problem.
3.6.3 N4SID
The N4SID algorithm is different. The following weighting matrices are used
⊥ 
Wc = UL+1|L+1 W1T (W1 UL+1|L+1
⊥ W1T )−1 W1 
Wr = I(L+1)m N4SID (3.109)

g=1
The column weighting matrix used in the N4SID algorithm does not generally
have the same column space as Wc or any of the column weighting matrices
which result from Equation (3.104) and Theorem 3.4.2. This is possibly the
reason for why N4SID gives bad results for deterministic input signals.
In Viberg (1995) it is pointed out that only the difference between the PO-
MOESP and N4SID algorithms is the extra projection UL+1|L+1 ⊥ . However, it
is also claimed that the resulting subspace estimates should therefore have very
similar properties. From Theorem 3.4.2 and the above discussion we conclude
that this conclusion in Viberg (1995) is wrong. This is illustrated in example
2, Section 3.7.2.
From the above discussion we have the following relationship between the
column weighting matrix Wc in Equation (3.104) and the matrix Wc in (3.109)
used by N4SID.
Wc = WcN4sid UL+1|L+1
⊥
(3.110)
In Theorem 3.4.2, Equations (3.52) and (3.55), it is proved that the extra
⊥
projection UL+1|L+1 is necessary in order to remove the deterministic term
d
HL+1 UL+1|L+1 and establish the data matrix ZL+1|L+1 which has the same
column space as the extended observability matrix. See also Example 2, Section
3.7.2, for an illustration.
The N4SID method computes the SVD of the data matrix defined in Equa-
tion (3.102) with the above matrices Wr and Wc , Equation (3.109). If the
triangular factors as shown in Section 3.5 are used then we have
· ¸ · ¸
T T −1 Q1 Q1
R32 R22 (R22 R22 ) [R21 R22 ] = (U1 Sn V1T + U2 S2 V2T ) (3.111)
Q2 Q2
The orthogonal matrices Q1 and Q2 are not used. The system order is identified
as the number of non-zero singular values and the extended observability matrix
OL+1 is estimated from the column space. The rest of the N4SID algorithm
can briefly be described as follows. From the extended observability matrix
OL+1 and the system data, a valid sequence of system states are estimated
and a least squares problem is solved in order to construct the system matrices
A, B, D, E. The covariance matrices for the process and measurements noise
are then identified from a residual, and a Riccati equation is solved in order
to construct the Kalman filter gain matrix C and the innovations covariance
matrix ∆.
The DSR method does not use state sequences and the Kalman gain matrix
C and the innovations covariance matrix ∆ are constructed directly from the
data, without recursions of non-linear matrix Riccati equations.
3.6.4 Main Differences and Similarities
Both of the algorithms N4SID and CVA estimate in the first instance a se-
quence of states. When the states are known, the state space model matrices
can be determined by simple linear regression. Both methods must solve a
matrix Riccati equation in order to identify the Kalman filter gain and the
innovations covariance matrices.
The DSR algorithm is based on first writing up an extended state space
model (ESSM) where the unknown states are eliminated from the problem.
The ESSM shows us the relationship between the known data matrices and the
SSM matrices. Hence, the DSR algorithm does not have any problems with
unknown states; unknown initial values etc. The state space model matrices are
then extracted from the ESSM. DSR estimates the Kalman gain and innova-
tions covariance matrices (stochastic part of the model) directly from the data,
without recursions of non-linear matrix equations, e.g. the Riccati equation.
The PO-MOESP algorithm does not estimate the stochastic part of the
model.
We have shown that the CVA, PO-MOESP and DSR algorithms gives
consistent estimates of the extended observability matrix. The algorithms fit
into the same Theorem 3.4.2. We have shown that the N4SID algorithm in
general does not give consistent estimates of the extended observability matrix.
However, it will give consistent results if an extra projection of future inputs is
included. From the above discussion we have the following relationship between
the column weighting matrix Wc in Equation (3.104) and the matrices used by
PO-MOESP, CVA and N4SID. See (3.105), (3.107) and (3.109).
Wc = WcPO-MOESP = WcCVA (WcCVA )T = WcN4sid UL+1|L+1

⊥
These are the most important similarities and differences between the method
presented in this work and previously published methods.
3.7 Numerical Examples 55
3.7 Numerical Examples
3.7.1 Example 1: Monte Carlo Simulation
A single input single output (SISO) system with one state is chosen to com-
pare the algorithm presented in this paper, entitled DSR (Deterministic and
Stochastic System Identification and Realization), with two other algorithms,
CVA (which stands for Canonical Variate Analysis, Larimore (1983), (1990))
and the prediction error method implemented in the MATLAB function AR-
MAX (i.e., in the system identification toolbox, Ljung (1991)).
xk+1 = 0.9xk + 0.5uk + 0.6ek (3.112)

yk = 1.0xk − 1.0uk + ek (3.113)
Three types of input signals were used. One input equal to a sum of four sinusoid
signals, u1 , one input equal to a white noise signal with unit covariance, u2 ,
and one equal to a sine, u3 .
k k k
u1 uk = 0.2(sin( ) + sin( ) + sin( ) + sin(k))
25 10 5
u2 White noise, unit covariance
u3 uk = sin(k)
For each input the time series (yk , uk ) was generated by simulating the model
with 100 different white noise sequences ek , also with unit variance.
The DSR algorithm parameter L was changed from 1 to 5 and the CVA
parameter I from 2 to 6. For each L and I, the mean and standard deviation of
the parameters of the 100 different estimated models are presented in Tables 3.1
to 3.6. The results obtained by the ARMAX algorithm are also shown in the
tables. See Ljung (1991) for the description of the parameters nn = [1, 2, 1, 0]
which is used as arguments for ARMAX.
The true deterministic system quadruple is denoted (a, b, d, e) := (0.9, 0.5,
1, −1) and the deterministic steady state gain and deterministic zero are de-
noted H d (1) = 4.0 and pd (1) = 1.4, respectively. The parameters in the stochas-
tic part of the model are (c, ∆) := (0.6, 1). The stochastic steady state gain
and stochastic zero are denoted H s (1) = 7.0 and ps (1) = 0.3, respectively. The
signal to noise ratio is approximately 0.4, hence, the identification problem is
not simple.
The CVA algorithm sometimes estimated systems with negative (b,d) pa-
rameters, i.e. sometimes an estimated quadruple (a,b,d,e) and sometimes (a,-
b,-d,e). This happened with the algorithm parameter I = 3 and with a random
input signal. It is believed that this can be avoided by using a scaling similar
to that presented in Section 3.5.4.
The results are very good for both the DSR and the CVA algorithms, see
Tables 3.1 and 3.2. There are small differences in the estimated models for
both methods when N is large, see Tables 3.1 and 3.2. This indicates that the
asymptotic statistical distribution of the parameter estimates is the same. The

example indicates that both the DSR and CVA algorithms are insensitive to
variation in the number of block rows.
The DSR algorithm is found to be marginally better than the CVA when
the number of samples is small (for this example and with N = 200 and N = 500
samples) see Tables 3.3 to 3.6.
It is also interesting to observe that the results from DSR are as good
as the results from the ARMAX function, even for a simple SISO system.
This indicates that DSR gives asymptotically statistical optimal results for this
example, both for purely deterministic inputs (u1 and u3 ) and stochastic input
sequences (u2 ). Note that the prediction error method (ARMAX function) is
based on iterative optimization but that the DSR algorithm is based on SVD
and QR decompositions only. Prediction error methods are rather complicated
for MIMO systems while the DSR algorithm is very simple.
Figures 3.2 to 3.8 are included in order to illustrate the asymptotic mean
and variance properties of the DSR algorithm for varying horizon parameters
L and J. The figures illustrate that the algorithm is numerically robust and
that the estimates are consistent and fairly insensitive for the parameters L and
J.
Mean of pole estimates
0.904
0.902
0.9
0.898
0.896
5
4 6
5
3 4
2 3
2
L 1 1
J
Figure 3.2: The mean value of the pole estimates for a Monte Carlo simulation
with varying past horizon parameter J and identification parameter L. The
number of samples in each simulation was N = 15000 and the number of sim-
ulations for each pair L, J was 100. The input was a sum of 5 sinusoid signals
(u1 ). The maximum estimate was 0.9003 for L = 3 and J = 4. The minimum
estimate was 0.8990 for L = 5 and J = 2. The closest estimate to the actual
pole (a = 0.9) was 0.89997 for L = 3 and J = 5.
Normalized standard deviation of pole estimates
0.8
0.6
0.4
0.2
0
5
4 6
5
3 4
2 3
2
L 1 1
J
Figure 3.3: The standard deviation multiplied (normalized) with N 1/2 of the
pole estimates for a Monte Carlo simulation with varying past horizon param-
eter J and varying identification parameter L. The number of samples in each
simulation was N = 15000 and the number of simulations for each pair L, J was
100. The input was a sum of 5 sinusoid signals (u1 ). The maximum standard
deviation of the pole estimates was 0.6265/N 1/2 at L = 5 and J = 3. The
minimum standard deviation was 0.4005/N 1/2 at L = 2 and J = 3.
Table 3.1: Mean of the parameters of 100 different estimated models: input
type u1 , N = 10000. The models was scaled so that d = 1.
Alg., Par. a b e pd (1) H d (1) c ∆ ps (1) H s (1)

DSR, L=1 0.8998 0.497 -1.002 1.396 3.9598 0.6088 1.0045 0.2910 7.0910
DSR, L=2 0.8995 0.497 -1.002 1.398 3.9516 0.6003 0.9989 0.2992 6.9867
DSR, L=3 0.8998 0.497 -1.002 1.398 3.9604 0.5992 0.9985 0.3006 6.9940
DSR, L=4 0.8996 0.497 -1.002 1.398 3.9572 0.5992 0.9981 0.3004 6.9875
DSR, L=5 0.8998 0.496 -1.002 1.398 3.9644 0.5989 0.9977 0.3009 6.9976
CVA, I=2 0.8996 0.497 -1.002 1.398 3.9545 0.6008 1.0046 0.2988 7.0015
CVA, I=3 0.8996 0.497 -1.003 1.398 3.9572 0.5995 1.0000 0.3001 6.9849
CVA, I=4 0.8998 0.495 -1.003 1.396 3.9471 0.5990 0.9994 0.3008 6.9930
CVA, I=5 0.8996 0.498 -1.001 1.400 3.9714 0.5992 0.9993 0.3003 6.9830
CVA, I=6 0.8996 0.498 -1.002 1.399 3.9659 0.5993 0.9993 0.3003 6.9872
ARMAX 0.8995 0.498 -1.002 1.399 3.9604 0.5995 0.3000 6.9769
Mean of C estimates
0.64
0.63
0.62
0.61
0.6
0.59
0.58
0.57
5
4 6
5
3 4
2 3
2
L 1 1
J
Figure 3.4: The mean value of the Kalman filter gain estimates for a Monte
Carlo simulation with varying past horizon parameter J and identification pa-
rameter L. The estimates are computed by the algorithm in Corollary 3.5.1.
The number of samples in each simulation was N = 15000 and the number of
simulations for each pair L, J was 100. The input was a sinusoid signal (u3 ).
The actual value is C = 0.6. The mean of the estimates at the flat region
described by 3 ≤ J ≤ 6 and 1 ≤ L ≤ 5 was 0.6000. This indicates that the
estimates are consistent for past horizon parameters J > 2 independently of
the choice of L, but clearly biased for J = 1. The estimates for J = 2 should
be further investigated.
Table 3.2: Standard deviation of the parameters of 100 different estimated

models: input type u1 , N = 10000.

DSR, L=1 0.0054 0.027 0.068 0.045 0.3020 0.0101 0.0153 0.0130 0.3170
DSR, L=2 0.0052 0.026 0.068 0.043 0.3063 0.0099 0.0152 0.0121 0.3021
DSR, L=3 0.0054 0.026 0.068 0.043 0.3190 0.0102 0.0152 0.0128 0.3211
DSR, L=4 0.0056 0.026 0.068 0.044 0.3236 0.0105 0.0151 0.0135 0.3309
DSR, L=5 0.0059 0.026 0.068 0.043 0.3328 0.0111 0.0151 0.0145 0.3447
CVA, I=2 0.0052 0.027 0.068 0.044 0.2959 0.0101 0.0153 0.0124 0.2994
CVA, I=3 0.0051 0.029 0.069 0.047 0.3423 0.0097 0.0152 0.0118 0.3009
CVA, I=4 0.0053 0.033 0.070 0.052 0.3920 0.0096 0.0152 0.0118 0.3155
CVA, I=5 0.0055 0.038 0.070 0.057 0.4456 0.0096 0.0151 0.0119 0.3236
CVA, I=6 0.0058 0.037 0.07 0.056 0.4554 0.0096 0.0152 0.0122 0.3391
ARMAX 0.0051 0.027 0.066 0.042 0.2925 0.0095 0.0115 0.2983
Normalized standard deviation of C estimates
1.4
1.2
0.8
0.6
0.4
5
4 6
5
3 4
2 3
2
L 1 1
J
kalman filter gain estimates for a Monte Carlo simulation with varying past
horizon parameter J and varying identification parameter L. The estimates
are computed by the algorithm in Corollary 3.5.1. The number of samples in
each simulation was N = 15000 and the number of simulations for each pair
L, J was 100. The input was a sinusoid signal (u3 ). The following parameters
is found from the region with consistent estimates. The maximum standard
deviation of the pole estimates was 1.2135/N 1/2 at L = 4 and J = 5. The
minimum standard deviation was 0.9344/N 1/2 at L = 2 and J = 5. The mean
of all standard deviations was 1.0315/N 1/2 .
type u2 , N = 200. The models was scaled sa that d = 1.

DSR, L=1 0.8861 0.507 -0.992 1.405 4.0874 0.6093 0.9961 0.2768 7.0385
DSR, L=2 0.8884 0.502 -0.995 1.399 3.9833 0.6045 0.9699 0.2839 6.9509
DSR, L=3 0.8886 0.502 -0.997 1.398 3.9781 0.6094 0.9476 0.2792 7.0266
DSR, L=4 0.8891 0.5 -0.993 1.399 4.0110 0.6128 0.9218 0.2763 7.1207
DSR, L=5 0.8902 0.5 -0.994 1.399 4.0741 0.6139 0.8991 0.2763 7.2495
CVA, I=2 0.8865 0.508 -0.992 1.406 3.9629 0.6027 1.0102 0.2838 6.8462
CVA, I=3 0.8865 0.503 -0.995 1.399 3.8955 0.6072 0.9953 0.2793 6.8840
CVA, I=4 0.8859 0.500 -0.997 1.396 3.8534 0.6111 0.9916 0.2749 6.9022
CVA, I=5 0.8852 0.499 -0.997 1.394 3.8399 0.6122 0.9844 0.2730 6.9141
CVA, I=6 0.8840 0.501 -0.931 1.396 3.7969 0.6145 0.9801 0.2695 6.8643
ARMAX 0.8864 0.504 -0.994 1.399 3.8848 0.5987 0.9971 0.2877 6.7799
Mean of F estimates
1.04
1.02
0.98
0.96
5
4 6
5
3 4
2 3
2
L 1 1
J
Figure 3.6: The mean value of the square root of the innovations variance
estimates F for a Monte Carlo simulation with varying past horizon parameter
J and identification parameter L. The estimates are computed by the algorithm
in Corollary 3.5.1. The number of samples in each simulation was N = 15000
and the number of simulations for each pair L, J was 100. The input was a
sinusoid signal (u3 ). The actual parameter value is F = 1. The mean of all the
estimates in the figure is 0.9996 with a standard deviation of 7.4 · 10−4 . This
indicates that the estimates are consistent.
Normalized standard deviation of F estimates
0.8
0.6
0.4
0.2
0
5
4 6
5
3 4
2 3
2
L 1 1
J
innovations variance estimates F for a Monte Carlo simulation with varying past
horizon parameter J and varying identification parameter L. The estimates are
computed by the algorithm in Corollary 3.5.1. The number of samples in each
simulation was N = 15000 and the number of simulations for each pair L, J was
100. The input was a sinusoid signal (u1 ). The minimum standard deviation
was 0.6016/N 1/2 at L = 4 and J = 6.
Mean of C estimates
0.64
0.63
0.62
0.61
0.6
0.59
0.58
0.57
5
4 6
5
3 4
2 3
2
L 1 1
J
Figure 3.8: The mean value of the Kalman filter gain estimates for a Monte
Carlo simulation with varying past horizon parameter J and identification pa-
rameter L. The estimates are computed by the algorithm in Corollary 3.5.2.
The number of samples in each simulation was N = 15000 and the number of
simulations for each pair L, J was 100. The input was a sinusoid signal (u3 ).
The actual value is C = 0.6. The bias for J = 1 as shown in Figure 3.4 when
the estimates are computed as in Corollary 3.5.2 is approximately eliminated.


DSR, L=1 0.0379 0.096 0.073 0.132 2.3261 0.0101 0.0907 0.0884 2.4984
DSR, L=2 0.0351 0.091 0.072 0.122 1.8766 0.0131 0.0889 0.0915 1.9967
DSR, L=3 0.0357 0.093 0.079 0.123 1.8585 0.0158 0.0887 0.0964 2.0583
DSR, L=4 0.0363 0.095 0.084 0.123 1.9096 0.0183 0.0872 0.1046 2.1884
DSR, L=5 0.0361 0.098 0.089 0.121 1.9666 0.0210 0.0865 0.1095 2.4102
CVA, I=2 0.0354 0.096 0.073 0.133 1.9081 0.0104 0.0936 0.0844 1.9693
CVA, I=3 0.0351 0.099 0.074 0.137 1.8680 0.0106 0.0959 0.0839 2.0161
CVA, I=4 0.0357 0.099 0.077 0.138 1.8757 0.0115 0.0931 0.0871 2.0303
CVA, I=5 0.0367 0.099 0.078 0.135 1.9511 0.0119 0.0951 0.0904 2.1405
CVA, I=6 0.0376 0.098 0.079 0.133 1.8513 0.0123 0.0955 0.0974 2.0864
ARMAX 0.0348 0.092 0.068 0.123 1.7760 0.0723 0.0933 0.0867 1.8969
type u3 , N = 500. The models was scaled so that d = 1.

DSR, L=1 0.8952 0.493 -1.001 1.392 4.0591 0.6013 1.0006 0.2939 7.0975
DSR, L=2 0.8938 0.494 -1.002 1.390 3.9881 0.5961 0.9879 0.2977 6.9587
DSR, L=3 0.8934 0.493 -1.003 1.389 3.9382 0.6000 0.9784 0.2934 6.9510
DSR, L=4 0.8931 0.493 -1.002 1.389 3.9376 0.6018 0.9695 0.2914 6.9572
DSR, L=5 0.8936 0.493 -1.002 1.390 3.9738 0.6030 0.9612 0.2906 7.0060
CVA, I=2 0.8944 0.493 -1.002 1.391 3.9908 0.5944 1.0047 0.2999 6.9579
CVA, I=3 0.8933 0.492 -1.001 1.389 3.9558 0.5959 0.9998 0.2974 6.9273
CVA, I=4 0.8931 0.492 -1.000 1.389 3.9105 0.5966 0.9992 0.2965 6.8970
CVA, I=5 0.8930 0.492 -1.000 1.389 3.9064 0.5967 0.9986 0.2963 6.8926
CVA, I=6 0.8931 0.491 -1.999 1.389 3.8941 0.5972 0.9977 0.2960 6.8930
ARMAX 0.8936 0.493 -1.002 1.39 3.9354 0.5952 0.9980 0.2983 6.8954


DSR, L=1 0.0277 0.066 0.077 0.102 1.6168 0.0409 0.0644 0.0542 1.4790
DSR, L=2 0.0274 0.066 0.075 0.099 1.6257 0.0442 0.0637 0.0580 1.4551
DSR, L=3 0.0268 0.066 0.077 0.099 1.5491 0.0456 0.0638 0.0583 1.4129
DSR, L=4 0.0272 0.066 0.076 0.100 1.5699 0.0470 0.0626 0.0600 1.4067
DSR, L=5 0.0275 0.067 0.076 0.100 1.6054 0.0479 0.0640 0.0609 1.4341
CVA, I=2 0.0269 0.066 0.076 0.100 1.5547 0.0417 0.0653 0.0542 1.4149
CVA, I=3 0.0275 0.066 0.076 0.099 1.5903 0.0431 0.0642 0.0560 1.4462
CVA, I=4 0.0268 0.067 0.075 0.099 1.5164 0.0415 0.0645 0.0532 1.3767
CVA, I=5 0.0267 0.066 0.075 0.099 1.5091 0.0420 0.0645 0.0539 1.3780
CVA, I=6 0.0262 0.067 0.076 0.101 1.5003 0.0433 0.0660 0.0539 1.3668
ARMAX 0.0260 0.066 0.075 0.099 1.5207 0.0421 0.0643 0.0535 1.3581
3.7.2 Example 2
We will in this example investigate the problem with colored input signals and
the N4SID algorithm.
Consider the same SISO one state example as in Example 1, Equations
(3.112) and (3.113). Two different input signals were chosen, one equal to a
pure sinusoid signal, input type u3 , and one equal to a white noise sequence with
unit variance, input type u2 . The inputs are the same as defined in Example
1. The number of samples was fixed to N = 500. The standard deviation of
the innovation was varied from ∆0.5 = 0 to ∆0.5 = 0.01 in order to investigate
the sensitivity for noise. The number of block rows in the data matrices was
chosen as L = 2.
The extended observability matrix O3 was estimated from the column space
of the matrix ZL+1|L+1 , Equation (3.102). The dimension of the column space
is estimated as the number of “non zero” singular values, see Figure 3.9. We
have also introduced the normalized singular value (s1 − s2 )/s1 as shown in
Figure 3.10. This means that when (s1 − s2 )/s1 = 1 then the number of states
is n = 1.
The conclusions to be drawn from Figures 3.9 and 3.10 are that the DSR
algorithm gives reasonable estimates for both the system order and the actual
pole (the CVA and PO-MOESP give essentially the same results for this
example) and that the N4SID algorithm does not work at all for this system
with a pure deterministic sinusoid input signal (uk = sin(k)). However, note
that when the input was changed to a white noise sequence (input type u3 )
then the two algorithms gave essentially the same singular values as well as
pole estimates.
3.7.3 Example 3
A two input two output system with the following model matrices is considered.
   
1.5 1 0.1 0 0
A =  −0.7 0 0.1  B =  0 1  (3.114)
0 0 0.85 1 0
· ¸ · ¸
3 0 −0.6 0 0
D = E= (3.115)
0 1 1 0 0
 
0 0.1 · ¸
1 0
C =  0.1 0  ∆= (3.116)
0 1
0 0.2
The system have a deterministic zero at −3, i.e. the system is non-minimum
phase. The eigenvalues of A are 0.85, 0.75 ± 0.3708i.
The deterministic and stochastic gain matrices are given by
· ¸ · ¸
d 16 15 s 2.5 4.7
H (1) = H (1) =
2.6667 −2.5 −0.25 1.1833
DSR: singular values as a function of varying noise

8
6
singular values
0
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
N4SID: singular values as a function of varying noise

50
40
singular values
30
20
10
0
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
standard deviation of the innovation
Figure 3.9: The singular values for model order selection as a function of varying
innovations noise level. The input to the system was a pure sinusoid signal
(input type u3 ). The other parameters are as described in Example 2. The
singular values from the DSR algorithm are shown in the upper figure and for
the N4SID algorithm in the lower. The actual system order is n = 1. As we
can see, the system order is fairly well detected by the DSR algorithm, while
the N4SID algorithm does not work at all for this system with a sinusoid input
signal.
DSR and N4SID: normalized singular values

1 DSR
normalized singular values
0.8
0.6
0.4 N4SID
0.2
0
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
DSR and N4SID: pole estimates

3
2
pole estimates
1 DSR
0 N4SID
−1
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
standard deviation of the innovation
Figure 3.10: This figure shows the normalized singular values (s1 − s2 )/s1 and
the pole estimates as a function of varying innovations noise level for the system
in Example 2. The input to the system was a pure sinusoid signal (input type
u3 ). The other parameters are as described in Example 2. The normalized
singular values from both the DSR and N4SID algorithms are shown in the
upper figure. The pole estimates are shown in the lower figure. The actual
system order is n = 1 and the actual pole is 0.9. As we can see, the system
order and the pole is fairly well estimated by the DSR algorithm, while the
N4SID algorithm does not work at all for this system with a sinusoid input
signal.
The algorithm gives exact results when ∆ = 0. Hence, this result is not pre-
sented. The time series yk , uk was generated by simulating the model with
one particular random noise process ek with covariance ∆. The input was
u = [u2 u1 ]T . The DSR parameter was fixed to L = 6. The following estimates
are obtained by DSR.
· ¸ · ¸
15.6575 14.6168 2.3625 4.5583
Ĥ d (1) = Ĥ s (1) =
2.5576 −2.4188 −0.2436 1.1275
· ¸
ˆ = 1.0531 −0.0244
∆
−0.0244 0.9859
System Singular Values Condition numbers

5 20
10 10
0 15
10 10
−5 10
10 10
−10 5
10 10
−15 0
10 10
0 5 10 15 0 5 10 15
System order System order
System Singular Values Condition numbers
2 3
10 10
1
10 2
10
0
10
1
−1
10
10
−2 0
10 10
0 5 10 15 0 5 10 15
Figure 3.11: Singular values (si , i = 1, ..., 12) and condition numbers ( ss1i , i =
1, ..., 12) for the system order selection in Example 3. The noise free determin-
istic case with ∆ = 0, upper figure. The case with ∆ = I, lower figure.
3.7.4 Example 4: Modeling the crude oil prize
The idea of this example is to investigate if the crude oil prize can be modeled by
a combined deterministic and stochastic model with the USD vs. NKR exchange
rate treated as an exogenous input variable. It is assumed that the exchange
rate is managed and exactly known. There are of course other variables than
the exchange rate which influences upon the crude oil prize. These variables
are in this example assumed to be unknown stochastic variables.
N KR
A number of 504 observations of the crude oil prize ([ barrel ]) in the period
from 1993 to 1995 is shown in Figure 3.13. The data is from Norges Bank
(the Bank of Norway). The crude oil prize at observation k is defined as yk
and also defined as an output variable. At the same time instants the USD vs.
U SD
NKR exchange rate ([ N KR ]) in the period from 1993 to 1995 is shown in Figure
3.14. The exchange rate at day k is defined as uk and also defined as an input
variable.
Trace of mean square errors: dsr pem robust
0.3 DSR
0.25 PEM
0.2 ROBUST
0.15
0.1
0.05
0
1000 2000 3000 4000 5000 6000 7000 8000
[number of observations]
Figure 3.12: Results from a Monte Carlo simulation in order to investigate the
accuracy of the innovation estimates. The model in Example 3 is simulated 100
times with varying number of samples. The DSR estimates are computed from
Theorem 3.4.3. The PEM algorithm converged only for number of samples
N < 3000.
The time series are separated into a trend and a series of variations. Define
dyk = yk − y0 (k) (3.117)
duk = uk − u0 (k) (3.118)
where dyk is the variation in the crude oil prize around the trend y0 (k) and
duk is the variation in the exchange rate around the trend u0 (k). The exchange
rate trend is simply chosen as the first observation, i.e. u0 (1). Inspection
of the series shows that there is an immediately fall in the crude oil prize at
N KR
observation number k = 368 of about 18 [ barrel ]. This drop in the crude oil prize
can hardly be explained by changes in the exchange rate (uk ). The explain-
nation is more likely an economical decision (from a OPEC meeting etc.). The
trend is therefore chosen as shown in Figure 3.13. However, note that it is not
necessary to remove trends from the time series when the algorithm for Subspace
Identification (4SID), (Di Ruscio (1995) algorithm name DSR) is used.
The 200 first observations in the time series dyk and duk was used as inputs
to the DSR algorithm. The algorithm was also executed with the parameter L =
5 (number of block rows in the extended observability matrix). The singular
values and condition numbers used for estimating the order of the state space
model is shown in Figure 3.15. It is clear from the figure that a first order model
(n = 1) is a good choice. The estimated model parameters are A = 0.9669,
B = 160.6463, C = 1.8302, D = 0.4759, E = −719.6302 and ∆ = 2.9844 and
the model
xk+1 = 0.9669xk + 160.6463duk + 1.8302ek (3.119)
ˆ
dy k = 0.4759xk − 719.6302duk + ek (3.120)
where ek is a zero mean, Gaussian distributed stochastic sequence with covari-

ance
∆ = E(ek eTk ) = 2.9844 (3.121)
and unknown realization.

First, the deterministic part of the estimated model was simulated with all
504 exchange rate observations as inputs. The simulated deterministic model
output (dykd ) and the actual observations (dyk ) are illustrated in Figure 3.16.
The error between the actual and simulated time series can also be measured
as (the covariance estimate)
200
1 X
J1 = (dyk − dykd )2 = 23.996 (3.122)
200
i=1
504
X
1
J2 = (dyk − dykd )2 = 38.88 (3.123)
504
i=1
The total combined deterministic and stochastic model was simulated. The
stochastic input (innovations process) was generated with the MATLAB func-
tion e = randn(501, 1) ∗ sqrt(∆) also with seed zero. Note that a different
noise realization (different seed) will give a slightly different result. However,
the results are illustrated in Figure 3.17.
The innovations model can be written as an optimal prediction model for
the output (the crude oil prize). Define the innovations process as ek = yk − ȳk
where ȳk is the optimal prediction of the output yk . Then we have
xk+1 = Axk + Buk + C(yk − Dxk − Euk ) (3.124)

ȳk = Dxk + Euk (3.125)
The optimal prediction of the crude oil prize variations is illustrated in Figure
3.18.
Even if the model is generated only by the 200 first observations it is ca-
pable of roughly predicting the slowly variations in all the 504 crude oil prize
observations.
dy: crude oil prize in [NKR/barrel] (trend removed)

20
10
−10
−20
−30
0 50 100 150 200 250 300 350 400 450 500
y0: crude oil prize trend [NKR/barrel]

130
125
120
115
110
105
100
0 50 100 150 200 250 300 350 400 450 500
[days]
N KR
Figure 3.13: The crude oil prize yk = dyk + y0 (k) [ barrel ] in the period 1993 to
1995 is separated into crude oil prize variations dyk (upper figure) and trend
y0 (k) (lower figure).
du: exchange rate [USD/NKR] (trend removed)
0.02
0.01
−0.01
0 50 100 150 200 250 300 350 400 450 500
u0: exchange rate trend [USD/NKR]

0.2
0.18
0.16
0.14
0.12
0.1
0 50 100 150 200 250 300 350 400 450 500
[days]
U SD
Figure 3.14: The exchange rate uk = duk + u0 (k) [ N KR ] in the period 1993 to
1995 is separated into rate variations duk (upper figure) and trend u0 (k) (lower
figure).
Singular Values Condition Numbers

2 2
10 10
1
10
1
10
0
10
−1 0
10 10
1 2 3 4 5 1 2 3 4 5
Figure 3.15: The singular values (si , i = 1, ..., 5) and condition numbers
( ss1i , i = 1, ..., 5) used to investigate the order of the state space model in Ex-
ample 4. The model order is estimated as the number of large singular values
or the number of small condition numbers.
Actual and simulated crude oil prize variations [NKR/barrel]

15
10
simulated
−5
−10
−15
−20
actual
−25
−30
0 50 100 150 200 250 300 350 400 450 500
[days]
Figure 3.16: Actual and simulated (deterministic part of model) crude oil prize.
Actual and simulated crude oil prize variazions [NKR/barrel]
20
15
simulated
10
−5
−10
−15
−20
−25
actual
−30
0 50 100 150 200 250 300 350 400 450 500
Figure 3.17: Actual and simulated (combined deterministic and stochastic

model) crude oil prize. This model was generated from the 200 first samples of
the input output samples, only. The algorithm parameter was L = 5.
Actual and optimal prediction of crude oil prize variations [NKR/barrel]
10
−10
−20
−30
0 50 100 150 200 250 300 350 400 450 500
[days]
Difference between actual and optimal prediction of crude oil prize
10
−5
−10
0 50 100 150 200 250 300 350 400 450 500
[days]
Figure 3.18: Actual and simulated optimal prediction of the crude oil prize is
shown in the upper part of the figure. The difference between the actual and
the optimal prediction of the crude oil prize is shown in the lower part of the
figure. This model was generated from the 200 first samples of the input output
samples, only. The algorithm parameter was L = 5.
3.7.5 Example 5: Modeling PH
The time series in this example is from a tray (plate) column scrubber at Union
Co., Skien, Norway (printed with permission). Liquid travels down (vertical)
the column over a series of trays. Gas travels up through contacting devices in
each tray. There is a PH control loop at each tray. A suspension of slurry is
used as input variable (here defined as uk ) to control the PH (output variable
here defined as yk ). The time series shown in Figure 3.19 are from an open loop
experiment at one tray in the column.
The sampling time was ∆t = 4 [sec]. All the N = 901 samples was used for
identification. The algorithm parameter was chosen as L = 2. Note that L = 1,
L = 2 and L = 3 gives about the same model and that a first order model was
detected from inspection of the L singular values (when L ≥ 2).
The estimated model is
xk+1 = 0.9881xk + 5.09 · 10−4 uk + 0.4358ek (3.126)
−5
ŷk = 0.7113xk + 4.76 · 10 uk + ek (3.127)
where ek is a zero mean, Gaussian distributed stochastic sequence with covari-
ance
∆ = E(ek eTk ) = 5.26 · 10−4 (3.128)
The estimated deterministic gain (gain from uk to ŷk ) is 0.031 and the estimated
∆t
time constant is T = − ln(a) = 5.6 [min]. Inspection of the actual time series
shows that these parameters are reasonable.
Singular values and principal angles for model order selection are shown
in Figure 3.20. A simulation of the deterministic part of the model and the
measured PH is shown in Figure 3.21.
3.8 Conclusions
A method for subspace identification and realization of state space models on

innovations form directly from given input and output data is presented.
The method determines both the deterministic part and the stochastic part
of the model.
The algorithm gives exact results in the deterministic case and consistent
results when the system is influenced by noise.
The stochastic part of the model is computed from standard linear algebra
decomposition methods and no matrix equations (e.g. Riccati or Lyapunov
equations) need to be solved.
3.8 Conclusions 73
y: PH measurement
0.4
0.2
−0.2
−0.4
0 10 20 30 40 50 60
u: valve position
20
10
−10
−20
0 10 20 30 40 50 60
[min]
Figure 3.19: The measured PH (output from the process) upper figure. The
manipulated input to the process (valve position) is shown in the lower figure.
Singular Values Pricipal angles (CVA)

0 0
10 10
−1
10
−1
10
−2
10
−3 −2
10 10
1 2 3 4 5 1 2 3 4 5
Figure 3.20: Singular values and principal angles for inspection of model order.
The principal angles are computed by the CVA algorithm. The DSR parame-
ter L = 5 was chosen. The figure shows that a reasonable model order is n = 1.
See Example 5 for details.
Measured and simulated PH

0.4
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
−0.5
0 10 20 30 40 50 60
[min]
Figure 3.21: Measured and simulated (deterministic model simulation) PH. The
model was generated from all the N = 901 samples. The algorithm parameter
was L = 2.
References
Aoki, M. (1990). State Space Modeling of Time Series. Second, Revised and
Enlarged Edition. Springer-Verlag Berlin, Heidelberg.
Aoki, M. (1994). Two Complementary Representations of Multiple Time Se-

ries in State-Space Innovation Forms. Journal of Forecasting, Vol. 13,
pp. 69-90.
Di Ruscio, D. (1994). Methods for the identification of state space models from
input and output measurements. SYSID 94, The 10th IFAC Symposium
on System Identification, Copenhagen, July 4 - 6.
Di Ruscio, D. (1995). A method for the identification of state space models

from input and output measurements. Modeling, Identification and Con-
trol, Vol. 16, no. 3. Program commercial available by Fantoft Process
AS, Box 306, N-1301 Sandvika.
Di Ruscio, D. (1995b). A method for identification of combined determin-

istic and stochastic systems. Proceedings of the third European Control
Conference, ECC95, Roma, September 5-8, pp. 429-434.
Di Ruscio, D. and A. Holmberg (1996). Subspace identification for dynamic

process analysis and modeling. Control Systems 96, Halifax, Nova Scotia,
May 1996.
Golub, G. H. and C. F. Van Loan (1983). Matrix Computations. North Oxford

Academic Publishers Ltd.
3.8 Conclusions 75
Larimore, W. E. (1983). System identification, reduced order filtering and

modeling via canonical variate analysis. Proc. of the American Control
Conference, San Francisco, USA, pp. 445-451.
Larimore, W. E. (1990). Canonical Variate Analysis in Identification, Filtering

and Adaptive Control. Proc. of the 29th Conference on Decision and
Control, Honolulu, Hawaii, December 1990, pp. 596-604.
Ljung, L. (1991). System Identification Toolbox. The Mathworks, Inc.
Faurre, P. L. (1976). Stochastic realization algorithms. In: System Identifica-

tion: Advances and Case Studies, (eds. R. K. Mehra and D. G. Lainiotis),
Academic Press.
Kalman, R. E., P. L. Falb and M. A. Arbib (1969). Topics in mathematical

system theory. McGraw-Hill Book Company.
Kung, S. Y. (1978). A new identification and Model Reduction Algorithm via

Singular Value Decomposition. Conf. on Circuits, Systems and Comput-
ers, Pacific Grove, CA, November 1978, pp. 705-714.

Van Overschee, P. and B. De Moor (1994). N4SID: Subspace Algorithms for

the Identification of Combined Deterministic Stochastic Systems. Auto-
matica, vol. 30, No. 1, pp.75-94.
Van Overschee, P. (1995). Subspace Identification: theory-implementation-

application. PhD thesis, Katholieke Universiteit Leuven, Belgium.
Van Overschee, P. and B. De Moor (1995). A Unifying Theorem for Three

Subspace System Identification Algorithms. Automatica, vol. 31, No. 12,
pp. 1853-1864.
Verhagen, M. (1994). Identification of the deterministic part of MIMO state

space models given on innovations form from input output data. Auto-
matica, vol. 30, No. 1, pp. 61-74.
Viberg, M. (1995). Subspace-Based Methods for the Identification of Linear

Time-invariant Systems. Automatica, vol. 31, No. 12, pp. 1835-1851.
Corrections
1. Index error in the matrices with zeroes on the right hand side (rhs) of
Equation (3.46).
In the third rhs term, change 0Lm×(k+L+1−J)r to 0(L+1)m×(k+L+1−J)r . In
the fifth rhs term, change 0Lm×(k+L+1−J)r to 0(L+1)m×(k+L+1−J)m . The
error is changed.
2. Index error on the right hand side of Equation (3.49). Change E0|L+1 to
E0|J . The error is changed.
⊥
5. The matrix UL+1|L+1 is missing on the right hand side in Equation (3.109).
The column weighting matrix Wc in the N4SID algorithm should be
⊥
Wc = UL+1|L+1 W1T (W1 UL+1|L+1
⊥
W1T )−1 W1
The error is changed.

Chapter 4
On the DSR algorithm
4.1 Introduction
The theory of subspace identification (SID) methods will be presented in gen-

eral. The theory and application of one particular SID method will be presented
in some detail.
A SID method can be viewed as a realization based approach to estimat-
ing state space models from input and output data. This is a most effective
and useful method, in particular for multivariable input and output (combined
deterministic and stochastic) systems.
A lower Left Q-orthogonal (LQ) decomposition is often used in subspace
identification methods in order to compute certain projection matrices and
subspaces of the known data matrices and to estimate the system order and
the extended observability matrix of the dynamic system.
The dynamics of the system can be extracted from the column space R(Z)
of one particular projected matrix Z which is computed from the input and
output data matrices Y , U and a method for computing subspaces, e.g., LQ
decomposition, singular value decomposition. An alternative method (to com-
pute the projection matrices, subspaces and the column space R(Z)) which
is based on the Partial Least Squares (PLS) method (decomposition) is also
presented.
Two examples are presented in order to compare different SID methods.
First: a Monte Carlo simulation experiment of a MIMO system is presented in
order to compare the numerically reliability of one particular subspace method
with two other subspace methods presented in the literature.
Second: a real world example from the pulp and paper industry is presented
in order to compare the quality of the methods. For this example there are three
input variables in the U data matrix and two output variables in the Y data
matrix. The data was collected from an experiment design. The quality of
the different models and validation aspects are addressed. The estimated state
78 On the DSR algorithm
space model is then used in a model predictive control strategy. Simulation

results are presented.
4.2 BASIC SYSTEM THEORETIC DESCRIPTION 79
4.2 BASIC SYSTEM THEORETIC DESCRIPTION
Underlying system described by a State space

model (SSM)
xk+1 = Axk + Buk + Gvk (4.1)

yk = Dxk + Euk + wk (4.2)
• Process noise vk , measurement noise wk are zero-mean and white.
• Colored process noise model included in SSM.
• The system is driven by both known input variables uk , unknown input

variables vk , wk and unknown initial values x0 .
• Description of combined deterministic and stochastic linear systems.
• Linearized approximation of non-linear systems.
Innovations form of state space model (ISSM)
is the starting point for the development of the algorithm.


Algorithm  x̄k+1 = Ax̄k + Buk + Cek
developed y = Dx̄ + Euk + F ek (4.3)
 k T k
from ISSM E(ek ek ) = I, x̄0 = E(x0 )
Traditional innovations form of SSM (Kalman filter)
x̄k+1 = Ax̄k + Buk + K²k (4.4)

yk = Dx̄k + Euk + ²k (4.5)
where
²k = F e k Innovations process
K = CF −1 Kalman filter gain (4.6)
E(²k ²Tk ) = F F T Innovations covariance matrix
Why use ISSM (4.3) ?
• Does not have to invert matrix F .
• Innovations ek has unit covariance matrix.
• ISSM a more unique representation of the system (4.1) and (4.2).

4.3 PROBLEM DESCRIPTION
From known system input and output data

¾
uk ∀ k = 1, 2, ..., N
Known data
yk ∀ k = 1, 2, ..., N
find the system order n, the initial state vector x0 and matrices in the ISSM
(up to within a similarity transformation)
½
ABC find system order and
D E F x0 matrices from known data
4.4 BASIC DEFINITIONS 81
4.4 BASIC DEFINITIONS
Fundamental subspaces of a matrix Z ∈ Rm×K
• R(Z):
– The column space of Z.

– I.e., the number of independent columns of Z.
– R(Z) is often called the range of Z.
– dim[R(Z)] = n
• N (Z):
– The null space of Z.

– dim[N (Z)] = K − n
dim[R(Z)] + dim[N (Z)] = K Number of columns
rank(Z) = dim[R(Z)] = n
In subspace identification:
• the observability matrix is estimated from the column space of a matrix

Z.
• the system order is identified as the dimension of the column space of Z.

4.5 ORTHOGONAL PROJECTIONS
• A matrix Y can be decomposed into two matrices with orthogonal row

spaces.
Y = Y /P + Y P ⊥
• Projection of the row space of Y onto the row space of P .
Y /P = Y P T (P P T )† P (4.7)
• Projection of the row space of Y onto the orthogonal complement of the

row space of P .
Y P ⊥ = Y − Y P T (P P T )† P (4.8)
Yf P ⊥ Yf
6 ©*
©©©
©
©©
©
©©
©
©© -
Yf /P P
Figure 4.1: Two dimensional illustration of orthogonal projections.
Some useful results
Lemma 4.5.1 The following equality is true

· ¸
U
U/ =U (4.9)
W

· ¸⊥
U
U =0 (4.10)
W
4.6 MATRIX DEFINITIONS 83
4.6 MATRIX DEFINITIONS
• The extended observability matrix for (D, A)

 
D
 
def  DA 
Oi =  .  ∈ Rim×n (4.11)
 .. 
DAi−1
Subscript i denotes the number of block rows.
• The reversed extended controllability matrix for (A, B)

def £ ¤
Cid = Ai−1 B Ai−2 B · · · B ∈ Rn×ir (4.12)
Subscript i denotes the number of block columns.
• The reversed extended controllability matrix for (A, C)

def £ ¤
Cis = Ai−1 C Ai−2 C · · · C ∈ Rn×ir (4.13)
Subscript i denotes the number of block columns.
• The lower block triangular Toeplitz matrix for (D, A, B, E)

 
E 0 0 ··· 0
 DB E 0 ··· 0 
 
d def  DAB DB E ··· 0 
Hi =   ∈ Rim×ir (4.14)
 .. .. .. .. .. 
 . . . . . 
i−2 i−3
DA B DA B DA B · · · i−4 E
• A lower block triangular Toeplitz matrix for (D, A, C, F )

 
F 0 0 ··· 0
 DC F 0 ··· 0 
 
def  ··· 0 
His =  DAC DC E  ∈ Rim×im(4.15)
 .. .. .. .
.. .  
 . . . . .
i−2 i−3
DA C DA C DA C · · · F i−4
4.7 BASIC MATRIX EQUATION IN SUBSPACE

IDENTIFICATION
Yk|L = OL Xk + HLd Uk|L + HLs Ek|L (4.16)
where
z
 }| {
yk yk+1 yk+2 ··· yk+K−1
 yk+1 yk+2 yk+3 ··· yk+K 
 
Yk|L =  .. .. .. .. ..  ∈ RLm×K (4.17)
 . . . . . 
yk+L−1 yk+L yk+L+1 · · · yk+L+K−2

z }| {
uk uk+1 uk+2 · · · uk+K−1
 uk+1 uk+2 uk+3 · · · uk+K 
 
Uk|L =  .. .. .. . . ..  ∈ RLr×K (4.18)
 . . . . . 
uk+L−1 uk+L uk+L+1 · · · uk+L+K−2
Unknown matrix of noise variables

z
 }| {
ek ek+1 ek+2 · · · ek+K−1
 ek+1 ek+2 ek+3 · · · ek+K 
 
Ek|L =  .. .. .. . . ..  ∈ RLm×K (4.19)
 . . . . . 
ek+L−1 ek+L ek+L+1 · · · ek+L+K−2
Unknown matrix of state vectors

z£ }| {¤
Xk = xk xk+1 xk+2 · · · xk+K−1 ∈ Rn×K (4.20)
4.8 MATRIX EQUATION WITH STATES ELIMINATED 85
4.8 MATRIX EQUATION WITH STATES ELIMI-

NATED
Define extended state space model (ESSM)
Yk+1|L = ÃL Yk|L + B̃L Uk|L+1 + C̃L Ek|L+1 (4.21)
where
ÃL = OL A(OLT OL )−1 OL

T
(4.22)
£ ¤ £ ¤
B̃L = OL B HLd − ÃL HLd 0Lm×r (4.23)
£ ¤ £ ¤
C̃L = OL C HLs − ÃL HLs 0Lm×m (4.24)
HLd and HLs are the lower left block triangular Toepliz matrices.
NOTE
Known B̃L → E given as lower right sub-matrix in B̃L

Known B̃L , OL and ÃL → B given
4.9 MATRIX OF INSTRUMENTAL VARIABLES

W
Need instrument to “remove” noise term in e.g., the ESSM.
unknown
↓
YJ+1|L = ÃL YJ|L + B̃L UJ|L+1 + C̃L EJ|L+1
YJ|L = OL XJ + HLd UJ|L + HLs EJ|L
↑
need instrument to
remove this term
1. Instrument should be “uncorrelated” with the noise matrix EJ|L+1 ,
lim EJ|L+1 /W = 0, lim EJ|L+1 W ⊥ = EJ|L+1 (4.25)

K→∞ K→∞
2. Instrument should be sufficiently correlated with informative part of ESSM

in order not to destroy information, e.g about system order.
rank(XJ /W ) = n
Define the matrix of instrumental variables as

· ¸ ¾
Y0|J instrument with J “past”
W = (4.26)
U0|J inputs and outputs
¾
J ≥L ESSM with “future”
(4.27)
YJ+1|L = ÃL YJ|L + B̃L UJ|L+1 + C̃L EJ|L+1 inputs and outputs
Interpretation of the horizons
• J: Past horizon for instruments.
• L: Future horizon used to predict the number of states.
How to chose parameters J and L
• Choose J = L. Seems no reasons for choosing J > L.
• Min. number of block rows for observability are Lmin = n − m + 1. Chose

L ≥ Lmin .
• System order n ≤ min(Lm, J(m + r)).

4.10 Subspace identification of OL : autonomous systems 87
4.10 Subspace identification of OL : autonomous sys-

tems
Consider the matrix equation
Zk|L = OL Xk ∈ RLm×K (4.28)
Assumptions
• Data matrix Zk|L known.
• The pair (D, A) observable.
• rank(Xk ) = n.
Column space of product
• The column space of OL Xk is contained in the column space of OL .
R(OL Xk ) ⊆ R(OL ) (4.29)
Sylvesters inequality gives
rank(OL ) + rank(Xk ) − n ≤ rank(Zk|L ) ≤ min{rank(OL ), rank(Xk )} (4.30)
Conclusions
• The column space of Zk|L coincide with the column space of OL .
R(Zk|L ) = R(OL ) (4.31)
• The rank of Zk|L (which is equal to the dimension of the column space)
is equal to the number of states n.
rank(Zk|L ) = dim[R(Zk|L )] = n (4.32)
• Can estimate both the system order n and the extended observability
matrix OL from the matrix Zk|L .
4.11 Subspace identification of OL : deterministic sys-

tems
Deterministic SSM can be written as the matrix equation
Yk|L = OL Xk + HLd Uk|L (4.33)
• Data-matrices Yk|L and Uk|L are known.
Results
def ⊥ ⊥
Zk|L = Yk|L Uk|L = OL Xk Uk|L (4.34)
and
⊥ )=n
rank(Zk|L ) = n if rank(Xk Uk|L (4.35)
⊥ loose rank below n ?

• Xk Uk|L
Analysis
A matrix Xk can be decomposed into two matrices with orthogonal row spaces.
⊥
Xk = Xk /Uk|L + Xk Uk|L
⊥
Xk Uk|L Xk
6 ©*
©©
©©
© ©©
©©
©
©© -
Xk /Uk|L Uk|L
• Xk /Uk|L : Projection of the row space of Xk on the row space of Uk|L .

⊥ : Projection of the row space of X onto the orthogonal comple-
• Xk Uk|L k
ment of the row space of Uk|L .
4.12 Subspace identification of OL+1 : “general” case 89
Conclusion
⊥ ) = n if X 6= U
rank(Xk Uk|L k k|L (4.36)
4.12 Subspace identification of OL+1 : “general” case
Define from the known data

 
UJ|L+1
def
ZJ|L+1 = (YJ|L+1 /  U0|J )U ⊥
J|L+1 , (4.37)
Y0|J
ZJ|L+1 is related to the extended observability matrix as
ZJ|L+1 = OL+1 XJa (4.38)
where
   
UJ|L+1

 X0 /  U0|J  

 
UJ|L+1  Y0|J 
£ J ¤  ⊥
XJa = XJ /  U0|J  U⊥
J|L+1 = A CJd CJ 
s
 U U
 J|L+1
  0|J  
Y0|J  UJ|L+1 
 E0|J /  U0|J  
Y0|J
• The column space of ZJ|L+1 coincides with column space of OL+1 .
• The system order n is given as the dimension of the column space.

4.13 BASIC PROJECTIONS IN THE DSR ALGO-

RITHM
“Autonomous system” for system dynamics, A and D
ZJ|L+1 = OL+1 XJa (4.39)
ZJ+1|L = ÃL ZJ|L (4.40)

where
 
UJ|L+1
def
ZJ|L+1 = (YJ|L+1 /  U0|J )U ⊥
J|L+1 (4.41)
Y0|J
Hence,
(A, D) = arg min k ZJ+1|L − ÃL ZJ|L k2F (4.42)
A,D
“Deterministic system” for B and E
d d
ZJ+1|L = ÃL ZJ|L + B̃L UJ|L+1 (4.43)
where
 
UJ|L+1
def
d
ZJ|L+1 = YJ|L+1 /  U0|J  (4.44)
Y0|J
Hence,
d d
(B, E) = arg min k ZJ+1|L − ÃL ZJ|L − B̃L UJ|L+1 k2F (4.45)
B,E
“Stochastic system” for C and F
s s
ZJ+1|L = ÃL ZJ|L + C̃L EJ|L+1 (4.46)
where
 ⊥
UJ|L+1
def
s
ZJ|L+1 = YJ|L+1  U0|J  (4.47)
Y0|J
Hence,
s s
(C, F ) = arg min k ZJ+1|L − ÃL ZJ|L − C̃L EJ|L+1 k2F (4.48)
C,F
4.14 COMPUTING PROJECTIONS BY QR DECOMPOSITION 91
4.14 COMPUTING PROJECTIONS BY QR DE-

COMPOSITION
DATA COMPRESSION BY QR-DECOMPOSITION
Compute the QR decomposition (or lower Left triangular, Q-orthogonal

(LQ))
“large” “small and square” “large”

K (2L+1)(r+m) K
z }| { z
 }| { z }| {
UJ|L+1 R11 0 0 0 Q1
1 def 1  W   R21 R22 0 0   Q2 
√ Ỹ = √  =   
  Q3  = RQ
(4.49)
K K YJ|L   R31 R32 R33 0
YJ+1|L R41 R42 R43 R44 Q4
| {z } | {z }
needed in algorithm not needed
POINTS
• Compress large and known data matrices into (usually) a “smaller” matrix
R.
• All relevant system information in lower left triangular part of matrix R.

Q is orthogonal.
1
• R is the “square root” of the covariance matrix K Ỹ Ỹ T .
1
• R is computed without never computing the covariance matrix K Ỹ Ỹ T .
• QR decomposition make algorithm numerically robust, e.g. due to round-

ing off errors (usually in situations when N is large), and ill-conditioned
noise realizations.
SEPARATION INTO
• AUTONOMOUS SYSTEM
• DETERMINISTIC SYSTEM AND
• STOCHASTIC SYSTEM
I.e., “data” separated into three parts by the QR-decomposition
1. One part for analyzing and determination of system dynamics, i.e. the
system order n and state equation system matrix A
R42 = ÃL R32 (4.50)
2. One part for determination of the deterministic part of the system
R41 − ÃL R42 = B̃L R11 (4.51)
3. One part for determination of the stochastic part of the system
R44 = C̃L EL QT4 (4.52)

R43 − ÃR33 = C̃L EL QT3 (4.53)
4.14 COMPUTING PROJECTIONS BY QR DECOMPOSITION 93
EXTRACTING THE SYSTEM MATRICES

(FROM R)
SOLUTION:
From sub-matrices in R Extract

½ order n and matrices in the ISSM
ABC
R →
DEF
1. System order: n and matrix D
SVD 
z · }| {
¸· T ¸  n System order by inspection
£ ¤ S1 0 
V1 of “non zero” singular values.
R32 = U1 U2
0 S2 V2T   O = U1 Extended observability matrix.
 L
D = OL (1 : m, :) Matrix in ISSM output eq.
2. Matrix A
¾
A = U1T R42 V1 S1−1 System matrix in ISSM state eq.
R42 = ÃL R32
ÃL = OA(OT O)−1 OT System matrix in ESSM.
3. Matrices B and E
· ¸
B
R41 − ÃL R31 = B̃L R11 → cs( ) = N † cs(R41 − ÃL R31 )
E
4. Matrices C and F
R44 → F and E(²k ²Tk ) = F F T

R43 − ÃL R33 = C̃L EL QT3 → C
where EL QT3 block diagonal with F on the diagonal, see paper eq. (66).
4.15 COMPUTING PROJECTIONS WITH PLS
• PLS: Partial Least Squares, Wold (1966).
• Projections in DSR algorithm can be computed by the use of PLS.
• Some effect for purely exiting input signals.
• Regularization. Bias and variance.

4.16 MODEL ORDER SELECTION BY EXAMPLE 95
4.16 MODEL ORDER SELECTION BY EXAMPLE

2 0
10 10
0
10 −2
10
−2
10
−4
10
−4
10
−6
−6
10 10
−8
10 −8
10
−10
10
−10
10
−12
10
−12
−14 10
10
−16 −14
10 10
0 5 10 0 5 10
Figure 4.2: Model order selection for process with two inputs and two outputs
and three states (no noise case).

2 0
10 10
1
10
0 −1
10 10
−1
10
−2 −2
10 10
0 5 10 0 5 10
Figure 4.3: Model order selection for process with two inputs and two outputs
and three states (noisy case).
4.17 PROPERTIES OF THE ALGORITHM

• Numerically robust implementation. The complete model is determined
from QR and SV D only.
• Practical and easy to use implementation.
• Only one scalar parameter to be specified by the user.
¤ the number of block rows in the extended observability matrix.
Choose
L≥1 (4.54)
• Can choose system order

n ≤ min(Lm, J(m + r)) (4.55)
• QR-decomposition of a data matrix with row size only

(2L + 1)(r + m) (4.56)
• Do not solve any Riccati equation in order to determine

– C - the Kalman filter gain matrix
– ∆ - the innovations process noise covariance matrix
• “Exact” results for deterministic systems.
• Consistent results for combined deterministic and stochastic systems (or
pure stochastic systems).
• No problems with non-zero initial values (of the states).
4.18 COMPARISON WITH CLASSICAL APPROACHES 97
4.18 COMPARISON WITH CLASSICAL APPROACHES
1. Parameterizations:
• Classical PE approaches need a certain user-specified model param-

eterization, so-called canonical forms.
There has been an extensive amount of research to determine canon-
ical forms.
Many problems with canonical forms (minimal parameterizations).
Can lead to numerically ill-conditioned problems, meaning that the
parameters in the canonical form model is extremely sensitive to
small perturbations.
• DSR method need no parameterization.
2. Convergence:
• PE approaches is iterative. Many hard to deal with problems such

as
problems with lack of convergence; no convergence; slow conver-
gence;
local minima; numerical instability;
• DSR method is non-iterative, there are no problems with conver-
gence and the method is numerically robust.
3. Speed:
DSR method faster than classical PE approaches (because the method
is non-iterative).
4.19 INDUSTRIAL APPLICATIONS
A short description is given of applications of the DSR identification method

for analysis and modeling of five different industrial data sets.
1. Thermo Mechanical Pulping (TMP) refiner. Models successfully

identified between almost all input and output variables in the process.
2. Paper machine. Dynamic model with 2 input and 2 output variables

identified.
3. Plate column scrubber. 4 input and 5 output variables basic for mod-
eling. (student project)
4. Raw oil distillation column. 4 input and 3 output variables are used
for identifying a dynamic modeling.
5. Rotary oven. 4 input and 2 output variables are used for identifying a
dynamic model.
4.20 SOFTWARE 99
4.20 SOFTWARE
• DSR Toolbox for MATLAB. Fantoft Prosess A.S, P.O. Box 306, 1301
Sandvika, Norway.
• DSR Windows 3.x, 95 or NT program. Fantoft Prosess A.S, P.O. Box

306, 1301 Sandvika, Norway.
• Deterministic and Stochastic system identification and Realization (DSR).
How to use matlab script
Known process input and output data

¾
uk ∀ k = 1, 2, ..., N
known
yk ∀ k = 1, 2, ..., N
Organize into data matrices

z  }| T {
y0
 yT 
 1 
Y = .  ∈ RN ×m (4.57)
 .. 
T
yN −1

z  }| T {
u0
 uT 
 1 
U = .  ∈ RN ×r (4.58)
 .. 
uTN −1
• Computing the dynamic model
£ ¤
A,B,D,E,C,F ,x0 = DSR(Y, U, L) (4.59)
• One scalar parameter L.
• Choose system order n ≤ min(Lm, J(m + r)).

4.21 Analysis and modeling of a scrubber
• Time series from a tray (plate) column scrubber, Union Co., Skien, Nor-
way.
• Sampling time 4 [sec]. N = 901 samples used for identification.
y: PH measurement
0.4
0.2
−0.2
−0.4
0 10 20 30 40 50 60
u: valve position
20
10
−10
−20
0 10 20 30 40 50 60
[min]
Figure 4.4: The measured PH in upper figure. The input (valve position) in
lower figure.

0 0
10 10
−1
10
−1
10
−2
10
−3 −2
10 10
1 2 3 4 5 1 2 3 4 5
Figure 4.5: Principal directions and principal angles for dynamic order selection.
4.22 MONTE CARLO SIMULATION OF MIMO SYSTEM 101
4.22 MONTE CARLO SIMULATION OF MIMO

SYSTEM
System data:
Number of inputs: r=2

Number of outputs: m = 2
Number of states: n=3
Input experiment signal:
· t
¸ ¾
sin( 14 ) + sin( 7t ) + sin( 3t )
ut = t t Case 1: results in Figure 4.6
(sin( 25 ) + sin( 10 ) + sin( 5t ) + sin(t))0.2
· t
¸ ¾
(sin( 14 ) + sin( 7t ) + sin( 3t ))0.05
ut = t t Case 2: results in Figure 4.7
(sin( 25 ) + sin( 10 ) + sin( 5t ) + sin(t))0.1
Noise covariance matrices:
· ¸
2 0
E(wk ) = 0, E(wk wkT ) = (4.60)
0 2
· ¸
1 0
E(vk ) = 0, E(vk vkT ) =
0 1
Simulation parameters:
Number of samples: N = 5000

Number of experiments: M = 100
Algorithms and parameters used:
DSR, L=2
CVA, Larimore (1990), I=3
ROBUST, Van Overschee (1995), I = 3
• For purely exiting input signals, choose DSR parameter L small.
• L is the horizon used to predict the number of states.
• For choice L = 2, must choose n ≤ Lm = 4.
• CVA and ROBUST: The smallest parameter is I = 3 in order to identify

a 3rd order model.
DSR_lq ROBUST
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0.6 0.7 0.8 0.6 0.7 0.8
DSR_pls CVA
0.5 0.5
0.4 0.4
0.3 0.3
Im
0.2 0.2
0.1 0.1
0 0
0.6 0.7 0.8 0.6 0.7 0.8
Re
Figure 4.6: Pole estimates of a system with two inputs, two outputs and three
states. DSR parameters L = J = 2 for both the LQ and PLS method for
computing the projection matrices. Parameter I = 3 for ROBUST and CVA.
The exact poles are λ1,2 = 0.75 ± 0.3708i and λ3 = 0.85.
4.23 STATISTICAL ANALYSIS 103
DSR_lq ROBUST
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1
DSR_pls CVA
0.5 0.5
0.4 0.4
0.3 0.3
Im
0.2 0.2
0.1 0.1
0 0
0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1
Re
states. DSR parameters L = J = 2 for both the LQ and PLS method for
computing the projection matrices. Parameter I = 3 for ROBUST and CVA.
The exact poles are λ1,2 = 0.75 ± 0.3708i and λ3 = 0.85.
4.23 STATISTICAL ANALYSIS
• The method (DSR) is consistent.
• Efficiency is analyzed with Monte Carlo simulation
Remarks
• DSR is “more” efficient than n4sid, for a “small” number of observations.
• PEM did not converge sufficiently when initial-values (to PEM) was esti-
mated with the functions in the matlab ident toolbox.
• Refinement of the prediction error approach is possible, but at a higher

computational expense.
Conclusions:
• This example shows a larger variability in the complex pole estimates

from CVA andROBUST than the corresponding pole estimates from
DSR.
• CVA andROBUST estimated the real pole with reasonable accuracy.
• The estimates from DSR is close to the estimates from PEM.

Trace of mean square errors: dsr n4sid pem

1
0.9
0.8
0.7 DSR
0.6 N4SID
0.5 PEM
0.4
0.3
0.2
0.1
0
200 400 600 800 1000 1200 1400 1600 1800 2000
[number of observations]
Figure 4.8: The trace of the covariance matrix of the difference between the
estimated and the true parameter vector. PEM did not converge sufficiently
when the number of samples was small and large.
4.24 IDENTIFICATION AND CONTROL OF TMP

PROCESS
• The process is a Thermo Mechanical Pulping (TMP) plant at Union Co,

Skien, Norway. A key part in this process is a Sunds Defibrator double
(rotating) disk RGP 68 refiner.
• Subspace identification and MPC of a TMP refiner are addressed.
– Real process data for model identification.

– MPC algorithm simulation results only.
• The refiner is modeled by a MIMO ( 3-input, 2-output and 6-state) state

space model.
Description of the process variables
Input and output time series from a TMP refiner are presented in Figures 4.9
and 4.10. The time series is the result of a statistical experimental design .
The manipulable input variables

 u1 : Plug screw speed, [rpm]
Refiner input variables u : Flow of dilution water, [ kg
s ]
 2
u3 : Plate gap, [mm]
4.24 IDENTIFICATION AND CONTROL OF TMP PROCESS 105
The following vector of input variables is defined

 
u1
uk =  u2  ∈ R3 . (4.61)
u3 k
The output variables

The process outputs used for identification is defined as follows
½ MW h
y1 : Refiner specific energy, [ 1000kg ]
Refiner output variables
y2 : Refiner consistency, [%]
The following vector of process output variables is defined

· ¸
y1
yk = ∈ R2 . (4.62)
y2 k
Subspace identification
• Input and output time series from a TMP refiner are presented in Figures
4.9 and 4.10. The time series is the result of a statistical experimental
design .
• The problem is to identify the SSM matrices (A, B, C, D, E, F ) (up to

within a similarity transformation), including the system order (n), di-
rectly from known system input and output data vectors (or time series).
• The SSM is assumed to be on innovations form (Kalman filter).
The known process input and output data vectors from the TMP process can
be defined as follows
¾
uk ∀ k = 1, . . . , N
Known data
yk ∀ k = 1, . . . , N
For the TMP refiner example we have used:
N = 1500 [samples] for model identification

860 [samples] for model validation
• Trends should preferably be removed from the data and the time series
should be adjusted for time delays. Data preprocessing is not necessary
but it often increase the accuracy of the estimated model.
The following constant trends (working points) are removed from the refiner
input and output data.
 
52.3 · ¸
0   0 1.59
u = 7.0 , y = (4.63)
34.3
0.58
4.24 IDENTIFICATION AND CONTROL OF TMP PROCESS 107
Organize the process output and input data vectors yk and uk as follows

z  T }| {
y1
 yT 
 2 
Y =  .  ∈ RN ×m (4.64)
 .. 
T
yN

z  T }| {
u1
 uT 
 2 
U =  .  ∈ RN ×r (4.65)
 .. 
uTN
The problem of identifying a complete (usually) dynamic model for the process
can be illustrated by the following function (similar to a matlab function).
£ ¤
A, B, C, D, E, F = DSR(Y, U, L) (4.66)
where the sixfold matrices (A, B, C, D, E, F ) are the state space model matrices.
• The algorithm name DSR stands for Deterministic and Stochastic model
Realization and identification.
• L is a positive integer parameter which should be specified by the user.
• The parameter L defines an upper limit for the unknown system order n.
n ≤ Lm.
• The user must chose the system order by inspection of a plot with Singular
Values (SV) or Principal Angles (PA). The system order n is identified as
the number of “non-zero” SV’s or “non-zero” PA’s.
• Note that the Kalman filter gain is given by K = CF −1 and that the
covariance matrix of the noise innovation process is given by E(vk vkT ) =
FFT.
• L = 3 and n = 6 where chosen in this example.

Model predictive control of TMP refiner
The following weighting matrices are used in the control objective.

 
· ¸ 0.5 0 0
100 0
Qi = , Ri =  0 0.1 0  , Pi = 0r , ∀ i = 1, · · · , L.(4.67)
0 100
0 0 100
The following horizons are used

½
L = 6, the prediction horizon
Horizons in MPC algorithm
J = 6, the identification horizon
The following constraints are specified

   
50.0 54.0
umin
k =  6.0  , umax k =  8.0  ∀ k > 0 (4.68)
0.5 0.7
  
−0.5 0.5
∆umin
k =  −0.15  , ∆umax
k =  0.15  ∀ k > 0 (4.69)
−0.05 0.05
· ¸ · ¸
1.5 1.7
ykmin = , ykmax = ∀ k>0 (4.70)
33 40
Simulation results are illustrated in Figures 4.11 and 4.12.

4.25 References 109
4.25 References
Di Ruscio, D. (1997). A method for identification of combined deterministic

stochastic systems. In: Applications of Computer Aided Time Series Mod-
eling, Lecture Notes in Statistics 119, Eds. M. Aoki and A. M. Havenner,
Springer Verlag, ISBN 0-387-94751-5.

May 1996.




Academic Press.



matica, vol. 30, No. 1, pp.75-94.


pp. 1853-1864.

matica, vol. 30, No. 1, pp. 61-74.

Wold, H. (1966). Non-linear estimation by iterative least squares procedures.

Researsch papers in statistics, Ed. David, F. Wiley, New York, pp. 411-
444.
u_1: dosage screw speed setpoint [rpm]. Actual dosage screw speed.
60 60
u_1^pv
u_1
55 55
50 50
0 500 1000 1500 2000 0 500 1000 1500 2000
u_2: dilution water setpoint [kg/s]. Actual dilution water.
8 u_2^pv 8
u_2
7.5 7.5
7 7
6.5 6.5
0 500 1000 1500 2000 0 500 1000 1500 2000
u_3: plate gap setpoint [mm]. Actual plate gap.
0.8 0.8
u_3^pv
u_3
0.6 0.6
0.4 0.4
0 500 1000 1500 2000 0 500 1000 1500 2000
Time [samples] Time [samples]
Figure 4.9: Input time series from a TMP plant at Union Bruk. The inputs are
from an experimental design. The manipulable input variables are u1 , u2 and
u3 These inputs are set-points to local input controllers. The outputs from the
pv
local controllers (controlled variables) are shown to the left and denoted u1 ,
pv pv
u2 and u3 .
4.25 References 111
Spec. energy: Actual and model output #1: First 1500 samples used by DSR.
0.2
0.1
0
y_1
−0.1
−0.2
−0.3
0 500 1000 1500 2000 2500
Time [samples]
Cons.: Actual and model output #2: First 1500 samples used by DSR.
0.1
0.05
y_2
−0.05
−0.1
0 500 1000 1500 2000 2500
Time [samples]
Figure 4.10: Actual and model output time series. The actual output time
series is from a TMP plant at Union Bruk. The corresponding input variables
are shown in Figure 4.9.
Specific energy [MW/tonn]: r_1 and y_1

1.7
1.65
y_1
1.6
1.55
1.5
0 20 40 60 80 100 120
Consistency [%]: r_2 and y_2

40
38
y_2
36
34
32
0 20 40 60 80 100 120
[samples]
Figure 4.11: Simulation results of the MPC algorithm applied on a TMP refiner.
The known references and process outputs are shown.
Dosage screw speed [rpm]: u_1

54
52
50
0 20 40 60 80 100 120
Dilution water [kg/s]: u_2
8
6
0 20 40 60 80 100 120
Plate gap [mm]: u_3
0.7
u_3
0.6
0.5
0 20 40 60 80 100 120
[samples]
The optimal control inputs are shown.
Motor power [MW]: y_3

16.2
16
15.8
y_3
15.6
15.4
15.2
0 20 40 60 80 100 120
Production [1000 kg/h]: y_4

10
9.9
9.8
y_4
9.7
9.6
9.5
0 20 40 60 80 100 120
[samples]
4.25 References 113
Optimal predictor and deterministic model (scrubber)

0.4
0.2
−0.2
−0.4
0 100 200 300 400 500 600 700 800 900

[min]
Actual measured output
0.4
0.2
−0.2
−0.4
0 100 200 300 400 500 600 700 800 900

samples
Figure 4.14: Simulated model and optimal prediction of process output (upper)
and actual PH (lower).
Chapter 5
Subspace Identification for

Dynamic Process Analysis and
Modeling
Abstract
A recently developed subspace identification method for robust data analysis

and modeling of dynamic and static systems from known observations will be
presented. The method is used in order to analyze and modeling time series
from an Thermo Mechanical Pulping refiner line at Norske Skog, Norway.
Keywords: Subspace identification; time series; TMP refiners; power; specific
energy; system identification; practical results.
5.1 Introduction
The subspace method discussed in this work, [3] - [6], has a number of ad-
vantages compared to the traditional prediction error approaches for system
identification. The method is practical and easy to use, it is non-iterative, al-
ways convergent and completely data driven. See also [8] and the survey paper
[9].
The method detects system dynamics (system order) and it can be viewed
as a generalization of traditional multivariate data analysis and modeling meth-
ods such as Principal Component Analysis (PCA) and Regression (PCR) and
Partial Least Squares (PLS). These last methods all deals with analyzing and
modeling of static relationship between variables and are well known methods
in e.g., chemometrics and econometrics. However, these methods are insuffi-
cient in order to modeling data from dynamical systems, i.e., systems where
the direction of time is important.
This implies that PCA, PCR and PLS deals with a special case analysis and
116 Subspace Identification for Dynamic Process Analysis and Modeling
modeling problem of the problem which is solved by the subspace methods.

The first step in the subspace method is a data compression step. The pos-
sibly large number of observations (data) are compressed to a much smaller ma-
trix by a standard lower Left Q-orthogonal (LQ) decomposition. This smaller
lower triangular matrix contains the information of the system (process and sen-
sors) which generated the data. The system dynamics as well as the complete
model for the system is extracted from this smaller matrix. A short description
of how to use the method will be given in this paper.
Note that the method works for systems where the outputs are driven by
both known input variables and unknown process disturbance variables. The
unknown variables may be high frequency or colored noise variables. The
method is believed to be a valuable tool for analyzing and modeling of data from
industrial processes, in particular dynamic processes. The subspace method is
used on some important processes in the pulp and paper industry in Norway,
e.g. a plate column scrubber; a paper machine; the Thermo Mechanical Pulping
(TMP) process.
The process variables measured on a two stage TMP refiner line at Norske
Skog, Skogn, Norway are analyzed and modeled by the method. In particular,
a state space model for the refiner motor power is constructed from observed
data (time series). This results will be presented in the paper. The model for
the 1st stage refiner has three input variables, i.e., plug screw speed, dilution
water and plate gap. The power model will be presented in this paper.
The rest of this paper is organized as follows. Some preliminary system and
problem definitions are stated in Section 5.2. A description of the subspace
method is presented in Section 5.3. An application of the subspace method
[6] is presented in Section 5.4. The actual process is a two-stage TMP refiner
line at Norske Skog, Skogn, Norway. A state space model for the TMP refiner
motor power is identified. Some conclusions follows in Section 5.5.
5.2 Preliminary definitions
5.2.1 System definition

invariant, linear state space model (SSM) of the form
xk+1 = Axk + Buk + Cek (5.1)

where the integer k ≥ 0 is discrete-time, x ∈ Rn is the state vector with initial

value x0 , y ∈ Rm is the system output, u ∈ Rr is the system input, e ∈ Rm
stationary, with zero mean and covariance matrix E(ek eTk ) = ∆. The constant
matrices in the SSM are of appropriate dimensions. A is the state transition
5.3 Description of the method 117
matrix, B is the external input matrix, C is the Kalman gain matrix, D is

the output matrix and E is the direct control input to output (feed-through)
matrix. We will assume that (D, A) is an observable pair.
The innovations model, Equations (5.1) and (5.2), is discussed in e.g. [7],
[1].
5.2.2 Problem definition
The problem investigated in this paper is to identify a state space model, in-
cluding the system order (n), for both the deterministic part and the stochastic
part of the system, i.e., the quadruple matrices (A, B, D, E) and the double ma-
trices (C, ∆), respectively, directly from known system input and output data
vectors (or time series) defined as
¾
uk ∀ k = 0, . . . , N − 1
Known data
yk ∀ k = 0, . . . , N − 1
Note that the variables can be linearly dependent as well as independent. In

continuous time systems the matrix E in Equation (5.2) is usually zero. This
is not the case in discrete time systems due to sampling. However, E can be
forced to be zero by including a structure constraint in the algorithm, [6].
5.3 Description of the method
Given sequences with process input and output raw data. The first step in
a data modeling procedure is usually to analyze the data for trends and time
delays. Trends should preferably be removed from the data and the time series
should be adjusted for time delays. The trend of a time series can often be
estimated as the sample mean which represents some working point. Data pre-
processing is not necessary but it usually increase the accuracy of the estimated
model.
The simplicity of the subspace method [6] will be illustrated in the following.
Assume that the data is adjusted for trends and time delays. Organize the
process output and input data vectors yk and uk as follows

z  T }| {
y0
 yT 
 1 
Y = .  ∈ <N ×m (5.3)
 .. 
T
yN −1

z  T }|  {
u0
 uT 
 1 
U = .  ∈ <N ×r (5.4)
 .. 
uTN −1
The problem of identifying a complete (usually) dynamic model for the process
can be illustrated by the following function (similar to a matlab function).
£ ¤
A, B, C, D, E, ∆ = DSR(Y, U, L) (5.5)
where the sixfold matrices (A, B, C, D, E, ∆) are the state space model matrices
in Equations 5.1 and 5.2. The algorithm name DSR stands for Deterministic
and Stochastic model Realization and identification, see [3]- [6] for details. L
is a positive integer parameter which should be specified by the user. The
parameter L defines an upper limit for the unknown system order n. The user
must chose the system order by inspection of a plot with Singular Values (SV)
or Principal Angles (PA). The system order n is identified as the number of
“non-zero” SV’s or “non-zero” PA’s. See Figure 5.1 for an illustration of the
order selection.
0 0
10 10
−1
10
−1
10
−2
10
−3 −2
10 10
1 2 3 4 5 1 2 3 4 5
Figure 5.1: Singular Values (SV) and Principal Angles (PA) for dynamic order
selection computed from time series from a PH-control loop in a plate column
scrubber at Union Co., Skien, Norway. We have specified L = 5. A reasonable
choice for the system order is n = 1, since the first SV is large compared to the
other SV’s.
5.4 Modeling the power
The main refiner motor power is believed to be one of the most important refiner
measurements which is available. As an example, the motor power is the most
5.4 Modeling the power 119
important variable in order to estimate the consistency, the specific energy and
the steam which is generated in the refiner. The motor power is an output
from the refining process. The power should be modeled as a function of the
refiner input variables. A combined mechanistic and empirical power model is
discussed in [2]. For a CD refiner we have the following input variables; dosage
screw speed ns ; plate gap for the plane zone s1 ; plate gap for the conical zone
s2 ; flow rate of dilution water to the refiner inlet or plane zone qw1 and to the
conical zone qw2 . The power model can be formulated as
y = y(ns , qw1 , qw2 , s1 , s2 ) (5.6)
where y is the refiner motor power. Another procedure is to identify a model for
the refiner motor power, directly from measured input and output data. This
approach will be discussed in the rest of this section.
5.4.1 Refiner variables and model
The manipulable input variables are defined as follows


 u1 Plug screw speed, [rpm]
u ∈ <3 := u Total flow of dilution water, [ kg
s ]
(5.7)
 2
u3 Plate gap, conical zone, [mm]
The linearized power model is assumed to be on innovations form, Equation

5.1 and 5.2. The innovations model can be written as an optimal predictor for
the power output yt given all past process inputs and outputs. The optimal
prediction of the power can be computed from
xt+1 = Axt + But + C(yt − Dxt − Eut ) (5.8)

ŷt = Dxt + Eut (5.9)
where ŷt is the optimal prediction of the present output yt given all past outputs
(..., yt−2 , yt−1 ) and the present and all past inputs (..., ut−1 , ut ). The problem
is to find the model matrices (A, B, C, D, E) and the covariance matrix of the
innovations noise process, ∆ = cov(et eTt ).
We will also mention that the combined deterministic and stochastic model,
Equations 5.1 and 5.2, can be separated into two parts. One deterministic part
(with output ytd ) which is driven by the known process input variables ut and
one stochastic part (with output yts ) driven by the unknown innovations noise
process et . The actual output is then yt = ytd + yts . See e.g., [2] p. 36.
We will in this work focusing on modeling the 1st stage refiner motor power.
It is often not clear whether a process variable should be defined as an input
or output variable to a model. An example of this is the problem of modeling
the 2nd stage refiner. It is assumed that the 1st stage motor power yt contains
information of the disturbances which affects the 2nd stage refiner. Hence, the
inclusion of the output yt as an additional input to the power model for the
2nd stage refiner is reasonable. We will not elaborate this further.
5.4.2 Numerical results
The results in this section is based on the time series from Norske Skog, Skogn,
Norway. The following trends (working point, nominal values) are removed
from the time series prior to the numerical analysis.
y0 = 17.69 nominal motor power
u10 = 44.08 nominal plug screw speed
u20 = 7.15 nominal sum of dilution water
u30 = 0.83 nominal plate gap CD zone
Case 1: Input variables u1 and u2 .
The following MISO 1st stage refiner power model is determined by the use of
the DSR algorithm. The parameter to DSR was L = 1.
· 1 ¸
uk
xk+1 = Axk + B + Cek (5.10)
u2k
· 1 ¸
uk
yk = Dxk + E + ek (5.11)
u2k
where
£ ¤
A = 0.9557 B = £ 0.0227 −0.0558 ¤
D = 1.0 E = 0.0198 −0.0335 (5.12)
C = 0.4057 ∆ = 0.0472
The steady state gain for the deterministic part, Hd (1), and for the stochastic
part, Hs (1), is given by
Hd (1) = D(I − A)−1 B + E
£ ¤
= 0.5328 −1.2919 (5.13)
−1
Hs (1) = D(I − A) C + I = 10.14 (5.14)
The signal to nose ratio is given by
cov(y d )
S/N = = 2.45 (5.15)
cov(y s )
The prediction error
J = cov(y s ) = cov(y − y d ) = 0.1367 (5.16)
See Figures 5.2 - 5.10 for illustrations.
Case 2: Input variables u1 , u2 and u3 .

The following MISO refiner power model is determined by the use of the DSR
algorithm. The parameter to DSR is L = 1.
 1 
uk
xk+1 = Axk + B  u2k  + Cek (5.17)
3
uk
 1 
uk
yk = Dxk + E u2k  + ek
 (5.18)
u3k
5.5 Conclusion 121
where
£ ¤
A = 0.9539 B = £ 0.0236 −0.0526 0.0267¤
D = 1.0 E = 0.0207 −0.1459 0.014 (5.19)
C = 0.4090 ∆ = 0.0474
The steady state gain for the deterministic part, Hd (1), and for the stochastic
part, Hs (1), is given by
Hd (1) = D(I − A)−1 B + E

£ ¤
= 0.5315 −1.2860 −1.9500 (5.20)
−1
Hs (1) = D(I − A) C + I = 9.87 (5.21)
The signal to nose ratio is given by
cov(y d )
S/N = = 2.48 (5.22)
cov(y s )
The prediction error
J = cov(y s ) = cov(y − y d ) = 0.1354 (5.23)
The results from this model is not plotted. The influence from the plate
gap is approximately zero compared to the influence from the screw speed and
the dilution water. And the model is virtually the same as the model presented
above in Case 1.
5.5 Conclusion
A state space model for a TMP refiner motor power is constructed from time
series. The model have three input variables, plug screw speed, dilution water
and plate gap.
Bibliography
[1] M. Aoki, “State Space Modeling of Time Series,” Springer-Verlag, 1990.
[2] D. Di Ruscio, Topics in model based control with application to the thermo
mechanical pulping process, Dr. Ing. thesis. Report no. 93-46-W. Division
of Engineering Cybernetics, NTH, N-7034 Trondheim, 1993.
[3] D. Di Ruscio, “A Method for identification of combined deterministic and

stochastic system,” In Proceedings of the European Control Conference,
ECC95, Roma, Italy, September 5-8, 1995.
[4] D. Di Ruscio, “Methods for the identification of state space models from
input and output measurements,” SYSID 94, The 10th IFAC Symposium
on System Identification, Copenhagen, July 4 - 6, 1994. Also in: Modeling,
Identification and Control, Vol. 16, no. 3, 1995.
[5] D. Di Ruscio, “Identification and Realization of Combined Deterministic

and Stochastic Systems,” To be included in: Lecture Notes in Time Series
Analysis, (edited by M. Aoki, UCLA,USA), 1996.
[6] “DSR: Algorithm for combined deterministic and stochastic system identi-
fication and realization of state space models from observations,” Program
commercial available by Fantoft Process AS, Box 306, N-1301 Sandvika.
[7] P. L. Faurre, “Stochastic Realization Algorithms,” In: System Identifi-

cation: Advances and Case Studies, (edited by R. K. Mehra and D. G.
Lainiotis), Academic Press, 1976.
[8] W. E. Larimore, “Canonical Variate Analysis in Identification, Filtering

and Adaptive Control,” Proc. of the 29th Conference on Decision and Con-
trol, Honolulu, Hawaii, december 1990, pp. 596-604.
[9] M. Viberg, “Subspace-Based Methods for the Identification of Linear Time

Invariant Systems,” Automatica, vol. 31, No. 12, pp. 1835-1851.
124 BIBLIOGRAPHY
Actual power: output y

2.5
1.5
0.5
Power
-0.5
-1
-1.5
-2
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
samples
Figure 5.2: 1 st stage refiner motor power, R27.
Screw speed: input u1

3.5
2.5
1.5
u1
0.5
-0.5
-1
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
samples
Figure 5.3: Plug screw speed, u1 .

BIBLIOGRAPHY 125
Dilution water: input u2

0.7
0.6
0.5
0.4
0.3
u2
0.2
0.1
-0.1
-0.2
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
samples
Figure 5.4: Flow of Dilution water, u2 .
Deterministic part of power model

2
1.5
1
Power
0.5
-0.5
-1
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
samples
Figure 5.5: Deterministic part of the power model output, y d .

126 BIBLIOGRAPHY
Power and deterministic part of power model

2.5
1.5
0.5
Power
-0.5
-1
-1.5
-2
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
samples
Figure 5.6: Deterministic part of the power model and actual measured power.
Power model
2.5
1.5
0.5
Power
-0.5
-1
-1.5
-2
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
samples
Figure 5.7: The power model, sum of deterministic part and stochastic part.
BIBLIOGRAPHY 127
Actual power and power model

2.5
1.5
0.5
Power
-0.5
-1
-1.5
-2
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
samples
Figure 5.8: The power model and the actual measured power.
Difference between power and deterministic part of power model

2.5
1.5
0.5
MW
-0.5
-1
-1.5
-2
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
samples
Figure 5.9: The residual between the actual measured power and the determin-
istic part of the estimated power model, i.e. y s = y − y d .
128 BIBLIOGRAPHY
Spectral plot of ys and yr

1
10
0
10
-1
10
-2
10
-3
10 -2 -1 0 1
10 10 10 10
frequency (rad/sec)
Figure 5.10: The estimated frequency function of the actual stochastic part of
the power model, i.e. y s = y − y d , and ŷ d .
Chapter 6
Dynamic Subspace Modeling

(DSR) and Static Multivariate
Analysis and Regression
(PCA,PCR,PLS)
Abstract
It is shown that a solution to the multivariate analysis and regression problem

is a special case of a recently developed subspace method for analysis and iden-
tification of combined deterministic and stochastic dynamical systems. This
method detects system dynamics. A numerically stable solution to the multi-
variate analysis and regression problem is also presented. Keywords: System
identification; subspace methods; dynamic systems; combined deterministic and
stochastic systems; multivariate analysis and regression; PCR; PLS; data anal-
ysis and modeling.
6.1 Introduction
The theory of multivariate analysis and regression is a research field in many

disciplines, e.g. in chemistry and economics. Econometrics and chemometrics
are well established disciplines. A problem in many disciplines, including eco-
nomics and chemistry, is to postulate relationships between variables. It is the
purpose of econometrics to postulate the relationship between economic vari-
ables. In the same way one can state that a purpose of chemometrics is to
postulate relationships between chemical variables. This problem is of great
importance in general.
Solution procedures such as Partial Least Squares (PLS) and Principal Com-
ponent regression (PCR) are effective tools for analysis and regression of pa-
Dynamic Subspace Modeling (DSR) and Static Multivariate Analysis and
130 Regression (PCA,PCR,PLS)
rameters in systems which is static. I.e. these tools are effective to analyze and
modeling variables that are static related.
On the other side it can be very ineffective to use these methods to ana-
lyze, modeling and identify (input output) relationships between variables in a
dynamical system.
One goal of statistical planning of system experiments is to produce data
which is rich of system information. The system input variables which are
manipulable are perturbed in such a way that some invert-ability requirement
of the manipulable variables is satisfied. However, it is important that all the
parameters in the system is exited (in order to identify them). The system
input and output variables are observed over the experimental period.
Usually all systems are dynamic in nature. In order to use PLS and PCR
methods, to analyze the system input and output behavior, one have to wait for
all the transients to die out in order to e.g. identify the gain from the variables
which are manipulated to the observed variables. All the observations in the
transient period are wasted. The result of this identification procedure usually
gives a wrong conclusion about the gain. The system output is in addition to
the manipulable variables driven by noise. It is known from the identification
theory that the system gain can be identified before the transients in the system
have died out. PLS and PCR methods can not detect dynamics and handle
dynamical systems in a proper way. For such systems, methods which can
detect the dynamics, including the system order are prefered.
A so called subspace method for identification of combined deterministic
and stochastic systems directly from the data is presented in Di Ruscio (1994),
Di Ruscio (1994a) and Di Ruscio (1995). The starting point for this method is
to allow a more general model structure for the system in which the standard
static multivariate analysis and regression problem is a special case.
A fundamental property of this method is that the system dynamics (system
order) are identified from the data. If the system order is identified to be zero
then a typical static regression problem results, in which PLS or PCR techniques
can be applied. The system order is not the same as the number of principal
components in PCR and number of factors (components) in PLS. The number
of principal components and PLS factors are measures of the effective rank of a
sequence of input (or output) variable observations. It is more or less the same
as what is called persistently exiting analysis of the input signal in classical
identification theory.
Section 6.2 gives a system description of both the static and the dynamic
modeling problem and how the data is ordered into certain data matrices for
which informations about the underlying system can be retrieved.
Section 6.4 gives a review of the solution to the more general dynamic
system analysis and modeling problem. The first step is to compress the data
down to the square root of some covariance matrices. The next step is to
identify the system dynamics (system order). If the system order is detected
or specified to be zero then, the covariance matrices for the static problem are
6.2 System description and data organization 131
directly extracted. This method gives a numerically stable way of identifying

the system order and the matrices in a so called combined deterministic and
stochastic dynamic model.
The static analysis and regression modeling problem will be discussed in
Section 6.5. In Lindgren, Geladi and Wold (1993) a covariance based PLS
method is presented. The first step in this method is to compute the covariance
matrices which is an ill-conditioned step. The next step is then to use a modified
PLS method which works on the covariance matrices.
Another solution to this special static analysis and modeling problem will be
presented in this paper. This includes a numerically stable method to compress
the data down to the square root of the covariance matrices, without explicitly
forming the covariance matrices. In the next step, the standard PLS, SVD or
PCR methods can be used.
6.2 System description and data organization
We will in this section give a system description. A system description is impor-

tant because it is the fundament of the method to be presented. A poor system
description usually results in limitations of any method and algorithm. We will
in Section 6.2.1 give a description of the more general combined deterministic
and stochastic system. Section 6.2.1 gives a description of the less general static
system which is the fundament for the traditional analysis and regression tools
as PCA, PCR and PLS. The latter is a special case of the system description
in Section 6.2.2.
6.2.1 Combined deterministic and stochastic description

invariant, linear state space model of the form
xk+1 = Axk + Buk + Cek (6.1)

where k ≥ 0 is discrete-time, i.e. an integer, x ∈ <n is the state vector with

initial value x0 , y ∈ <m is the system output, u ∈ <r is the system input, e ∈ <m
stationary, with zero mean and covariance matrix E(ek eTk ) = ∆. A, B, C,
D and E are constant matrices of appropriate dimensions, where (D, A) is an
observable pair.
A solution to the identification problem of constructing a state space model,
including the system order (n), for both the deterministic part and the stochas-
tic part of the system i.e. the quadruple (A, B, D, E) and the double (C, ∆),
respectively, directly from known input and output time series uk and yk .
Given the system output and input time series, yk ∀k = 0, . . . , N − 1 and
uk ∀k = 0, . . . , N − 1, respectively. The state space model, Equations (6.1) and
(6.2), can generally be written as
Yk+1 = ÃYk + B̃Uk + C̃Ek (6.3)
where
 
yk yk+1 yk+2 ··· yk+K−1
 yk+1 yk+2 yk+3 ··· yk+K 
 
Yk =  .. .. .. .. ..  ∈ <Lm×K (6.4)
 . . . . . 
yk+L−1 yk+L yk+L+1 · · · yk+L+K−2
 
uk uk+1 uk+2 ··· uk+K−1
 uk+1 uk+2 uk+3 ··· uk+K 
 
 .. .. .. .. .. 
Uk =  . . . . .  ∈ <(L+1)r×K
(6.5)
 
 uk+L−1 uk+L uk+L+1 · · · uk+L+K−2 
uk+L uk+L+1 uk+L+2 · · · uk+L+K−1
 
ek ek+1 ek+2 ··· ek+K−1
 ek+1 ek+2 ek+3 ··· ek+K 
 
 
Ek =  ... ..
.
..
.
..
.
..
.  ∈ <(L+1)m×K(6.6)
 
 ek+L−1 ek+L ek+L+1 · · · ek+L+K−2 
ek+L ek+L+1 ek+L+2 · · · ek+L+K−1
The number of columns in Yk , Uk and Ek are K = N −L−k+1. Each column in

these matrices can be interpreted as extended output, input and noise vectors,
respectively. K can be viewed as the number of samples in these extended time
series. We also have that L < K < N . L is a user specified parameter which
is the number of block rows in the extended observability matrix which will be
determined.
6.2.2 Purely static description
A model suitable for PLS and PCR problems can be written as the following
linear steady state equation.
ytT = uTt E T + eTt (6.7)
where yt ∈ <m is a vector of known output variables at observation number t.

ut ∈ <r is a vector of known input variables and et ∈ <m is a vector of unknown
white noise variables, assumed to be covariance stationary, with zero mean and
covariance matrix E(et eTt ) = ∆.
Assume a number of N observations, i.e. t = 1, ..., N . The N equations
can then be formed into the following matrix equation with notations which is
common in multivariate regression theory.
Y = XB + E (6.8)
6.2 System description and data organization 133
where Y ∈ <N ×m is a known matrix of observations. The vector ytT is also

called a vector of responses at observation number t. X ∈ <N ×r is a known
matrix of input variables. Each row in X (i.e. uTt ) is also called a vector
of predictor variables at observation number t. The matrix B ∈ <r×m is the
matrix of unknown parameters (also called regression coefficients).
Both data matrices Y and X can be corrupted with (measurements) noise.
The principal components, detected as a first step in the PCR method, are the
number of linearly independent columns in the data matrix X. This possibly
rank deficiency problem can e.g. be detected as the number of “non-zero”
singular values which results from the Singular Value Decomposition (SVD) of
the covariance matrix estimate N1 X T X.
The (principal) components in the PLS method are computed by taking
advantage of both Y and X variables.
6.3 Dependent input or output data
We will in this section illustrate how a sequence of dependent output (or input)
variables can be projected to a sequence of dependent variables.
Assume that a sequence of possibly dependent input and (or) output vari-
ables are known
¾
Uk ∈ <s ∀ k = 1, 2, ..., N
known dependent data
Yk ∈ <p ∀ k = 1, 2, ..., N
We will illustrate the analysis only for the output variables because the anal-
ysis for the input variables ar similar. Define the data-matrices of possibly
dependent output variables as follows
p
z }| {
Y1T
 YT 
 2 
Y= .  (6.9)
 .. 
YNT
The effective rank of this data matrix can be analyzed by the SVD.
p
z· }| ¸{ · ¸
m
z}|{ S1 0 £ ¤T
Y= V1 V2 (6.10)
U1 U2 0 S2
The number m independent output variables are determined by inspection of

the non-zero or large singular values. A data matrix of dependent output vari-
ables can then be determined as
m
z }| {
y1T
 yT 
 2 
Y =  .  = YV1 S1−1 (6.11)
 .. 
T
yN
A sequence of dependent outputs (and inputs) can then be written as
¾
uk ∈ <r ∀ k = 1, 2, ..., N
known independent data
yk ∈ <m ∀ k = 1, 2, ..., N
6.4 The combined deterministic and stochastic problem 135
6.4 The combined deterministic and stochastic prob-

lem
6.4.1 Data compression
Define the following QR decomposition

    
UL R11 0 0 0 Q1
1 1   
W1   R21 R22 0 0   
√ Ỹ = √  =   Q2  = RQ (6.12)
K K YL   R31 R32
 R33 0   Q3 
YL+1 R41 R42 R43 R44 Q4
where
R ∈ <(r(L+1)+ni+2mL)×(r(L+1)+ni+2mL) (6.13)
(r(L+1)+ni+2mL)×K
Q ∈< (6.14)
The matrix W1 is a matrix of instrumental variables which are uncorrelated
with the noise variables in EL and sufficiently correlated with the informative
part of the data.
· ¸
Y0
W1 = (6.15)
U0
We can now remove the noise matrix EL from Equation (6.3) by post multiply-
ing with K1 W1T . We have from (6.12) that
· T ¸
1 T 1 £ T T
¤ R21
lim EL W1 = ( lim EL Q1 Q2 ) T =0 (6.16)
K→∞ K K→∞ K R22
Post multiplying Equation (6.3) with [QT1 QT2 ], using (6.16) and substituting for
the corresponding Rij sub-matrices from (6.12) gives
R42 = ÃR32 (6.17)
R41 = ÃR31 + B̃R11 (6.18)
The system order n and the system quadruple (A, B, D, E) are computed from
(6.18).
The stochastic part of the system, defined by the matrices C and ∆, is
computed from
R43 − ÃR33 = C̃Ek QT3 (6.19)
R44 = C̃Ek QT4 (6.20)
6.4.2 Identification of system dynamics
Identification of dynamics by principal component analysis

Given the matrices R42 and R32 which satisfy
R42 = ÃR32 (6.21)
where
Ã = OA(OT O)−1 OT (6.22)
and O is the extended observability matrix for the pair (A, D).
Determine the SVD
R32 = U SV T (6.23)
where
· ¸
£ ¤ mL×mL Sn 0 £ ¤
U= U1 U2 ∈< S= ∈ <mL×ni V = V1 V2 ∈ <ni×ni
(6.24)
0 0
where Sn ∈ <n×n and n is the number of non-zero singular values of R32 , which
is equal to the system order. n is determined by inspection of the non-zero
diagonal elements of S or SS T . If the system order is detected or specified to
zero, then we have a static modeling problem, and the rest of this algorithm is
skipped.
2. The extended observability matrix O for the pair (D, A).
The (extended) observability matrix can be taken directly as the first left part
in U , i.e. U1 . We have
O = U (1 : Lm, 1 : n) (6.25)

· −1 ¸
T Sn
A = O Zk+1 V = U1T Zk+1 V1 Sn−1 (6.26)
0
matrix O, i.e.
D = U (1 : m, 1 : n) (6.27)

We have
Ã = OA(OT O)−1 OT = R42 V S −1 OT = R42 V1 Sn−1 U1T (6.28)
System order from PLS factor analysis

Given the matrices R42 and R32 which satisfy
R42 = ÃR32 (6.29)

6.4 The combined deterministic and stochastic problem 137
where
Ã = OA(OT O)−1 OT (6.30)
and O is the extended observability matrix for the pair (A, D).
Determine the PLS weight matrices W and C, the loading matrix P and the
“score” matrix T which satisfy
R32 = P T T (6.31)
T
R42 = CT (6.32)
T T T is a diagonal matrix. We have W T R32 R32

T W = W T P T T T P T W . Hence,
the singular values of R32 is the square root of the eigenvalues of P T T T P T .

However, the zero eigenvalues should be reflected in the diagonal matrix T T T .
The system order (n) is the number of non-zero singular values of R32 . n
is determined by inspection of the non-zero singular values of R32 or the the
number of non-zero diagonal elements in T T T . If the system order is detected
or specified to zero, then we have a static modeling problem, and the rest of
this algorithm is skipped.
Extract the following sub-matrices according to the specified order
W =: W (:, 1 : n) ∈ <mL×n C =: C(:, 1 : n) ∈ <mL×n P =: P (:, 1 : n) ∈ <mL×n

(6.33)
Only the matrices W , P and C are needed in the rest of the algorithm.
2. The extended observability matrix O for the pair (D, A).
The (extended) observability matrix can be taken directly as,
O = W (1 : Lm, 1 : n) (6.34)

A = W T C(W T P )−1 = OT C(OT P )−1 (6.35)

matrix O, i.e.
D = W (1 : m, 1 : n) (6.36)

We have
Ã = W AW T = OAOT (6.37)
4
Note that the weight matrix W is not really necessary. We can use
O = P (6.38)
T −1 T
A = (P P ) P C (6.39)
because C = P A when O = P .
Further research
The effect of taken the system matrix as A = C(1 : n, 1 : n)P (1 : n, 1 :
n)−1 should be investigated, because, in this case, A is represented in a special
canonical form, e.g. with certain zeros and ones in addition to the necessary
number of parameters as elements in the matrix.
6.5 The static problem
The solution to the multivariate analysis and regression problem which will be
presented in this section follows as a special case of the solution to the more
general combined deterministic and stochastic dynamic problem which will be
reviewed in Section 6.4.
The solution presented in this section should be of interest because it only
works on the square root of the covariance matrices involved, e.g. the square
root of N1 [Y X]T [Y X] and thereby can be defined as numerically stable.
6.5.1 Data compression
The row dimension (number of observations) of Y and X is often much larger

then the column sizes (number of variables). These large data matrices can be
compressed into some much smaller matrices which contains the information
of the system. Note that the procedure of directly forming the covariance
matrix N1 [Y X]T [Y X] is ill-conditioned due to rounding-off errors (usually in
situations when N is large). We will in this section compress the data, down
to the square root of the covariance matrix. (without explicitly forming the
covariance matrix).
Define the following standard QR decomposition
· T ¸ · ¸· ¸
1 1 X R11 0 Q1
√ Ỹ = √ T = = RQ (6.40)
N N Y R21 R22 Q2
where
R11 ∈ <r×r R21 ∈ <m×r R22 ∈ <m×m (6.41)

R ∈ <(r+m)×(r+m) Q ∈ <(r+m)×N (6.42)
The solution to the total multivariate problem is given by the triangular factors
R11 , R21 and R22 , only. The orthogonal matrix Q is not needed. This will
6.5 The static problem 139
reduce the computational effort and storage considerably, especially when the
number of observations N is large compared to the number of variables.
We have directly the following equation for the regression coefficients
R21 = BT R11 (6.43)
In order to solve this equation for B, standard PLS or PCR methods can be
applied.
6.5.2 The residual
The lower triangular matrix R22 is the square root of the residual covariance
matrix. The covariance estimate of the noise (or residuals) is given by
ˆ = 1 E T E = R22 R22
∆ T
(6.44)
N
6.5.3 Effective rank analysis
The principal components
The lower triangular matrix R11 is the square root of the covariance matrix
1 T
N X X. The singular values of R11 is equal to the square root of the eigen-
values of N1 X T X. Hence, the principal component analysis can be made by
a singular value decomposition of R11 . The number of principal components
is the number of “non-zero” singular values of R11 . The “non-zero” singular
values are determined by inspection. We have
· ¸
T
£ ¤ Sa 0 £ ¤T
R11 = U SV = U1 U2 V1 V2 (6.45)
0 S0
The number of principal components a, is equal to the row size of Sa ∈ <a×a .

S0 contains the zero singular values or singular values which are put to zero on
the diagonal. Hence
R11 = U1 Sa V1T (6.46)
PLS factors
From an iterative procedure (e.g. by the Power method for computing eigen-
values and eigenvectors) the PLS method compute the weight matrices W and
C and the loading matrix P , and decomposes X and Y into
T
R11 = TPT (6.47)
T T
R21 = TC (6.48)
It is possible to determine r factors from the PLS algorithm corresponding to

the columns in W . The first factor and the first column in W is the maximum
eigenvalue and eigenvector of the covariance matrix N1 X T Y Y T X, respectively.
The effective rank index (a) is bounded by 1 ≤ a ≤ r. One idea is to estimate
the rank index as the number of “non-zero” PLS factors followed by cross
validation. The “non-zero” factors are estimated by inspection.
6.5.4 Regression
The principal component regression
T
B̂ = R11 (R11 R11 )−1 R21
T
(6.49)
T −1 T
B̂ = (R11 ) R21 (6.50)
If the number of principal components is less than r, then, we are using the
pseudo-inverse in the expressions for the regression matrix, i.e. we do not invert
zero or almost zero singular values. Hence
BT = R21 V1 Sa−1 U1T (6.51)
PLS regression
The PLS regression is of the form
B = W (P T W )−1 C T (6.52)
where the matrices W ∈ <r×a , P ∈ <r×a , C ∈ <m×a , results from the PLS
algorithm. Note that W T W = I.
6.6 Concluding remarks
A method for analyzing and modeling input and output relationships between
variables in possibly dynamical systems is presented. This method is viewed
against methods such as PCR and PLS which are frequently used for modeling
and analysis of static relationships between variables.
References
Di Ruscio, D. (1994). Methods for the identification of state space models from
input and output measurements. SYSID 94, The 10th IFAC Symposium
on System Identification, Copenhagen, July 4 - 6.
Di Ruscio, D. (1995a). A method for the identification of state space mod-

els from input and output measurements. Modeling, Identification and
Control, Vol.16 No. 3.
6.6 Concluding remarks 141
Di Ruscio, D. (1995b). A Method for Identification of Combined Determin-

istic and Stochastic Systems: Robust implementation. Third European
Control Conference ECC95, September 5-8, 1995.
Lindgren, F., Geladi, P. and S. Wold (1993). The kernal algorithm for PLS.
Journal of chemometrics, Vol. 7, 45-59.
Chapter 7
On Subspace Identification of
the Extended Observability
Matrix
7.1 Introduction
The lower Left Q-orthogonal (LQ) (or the Q-orthogonal, upper Right (QR))
decomposition is used in order to compute certain projection matrices in the
Subspace IDentification (SID) algorithms by Verhagen (1994), Van Overschee
and De Moor (1994) and Di Ruscio (1995), (1997). In this paper the Partial
Least Squares (PLS) method, Wold (1966), is used to compute the projection
matrices. The PLS method is a simple alternative.
7.2 Definitions
7.2.1 Notation
The projection of (the row space of) a matrix A onto (the row space of) another
matrix B is denoted A/B and defined as
def
A/B = AB T (BB T )† B, (7.1)
where † denotes the Moore-Penrose pseudo-inverse. Also define
def
B ⊥ = I − B T (BB T )† B (7.2)
so that BB ⊥ = 0. The projection of the row space of A onto the orthogonal
complement of the row space of B is then AB ⊥ . Hence A is decomposed into
two terms as
A = A/B + AB ⊥ , (7.3)
where the two matrix terms A/B and AB ⊥ are orthogonal to each other.
144 On Subspace Identification of the Extended Observability Matrix
7.3 Relationship between past and future data ma-

trices
Theorem 7.3.1 (relationship between past and future data matrices)
Define Y0|J , U0|J and E0|J as the past output, the past input and the past noise
data matrices, respectively. Similarly, define YJ|L+1 , UJ|L+1 and EJ|L+1 as the
future output, the future input and the future noise data matrices, respectively.
Then we have the following relationship between the past and the future data
matrices
 
UJ|L+1 · ¸
£ d ¤ £ s ¤ EJ|L+1
YJ|L+1 = HL+1 PL+1 ÃL+1  U0|J
d J + H s
L+1 PL+1 (7.4)
E0|J
Y0|J
where
ÃJL+1 = OL+1 AJ OJ† (7.5)
d
PL+1 = OL+1 (CJd − A OJ† HJd )
J
(7.6)
s
PL+1 = OL+1 (CJs − AJ OJ† HJs ) (7.7)
Proof:
See Appendix 7.10.1.
7.4 The observability matrix
One important result concerning the estimation of the extended observability

matrix is presented in Theorem 4.2, Equation (52), in the paper by Di Ruscio
(1996). The theorem with a new proof is presented in the following.
Define the following matrix from the known data

 
UJ|L+1
def
ZJ|L+1 = (YJ|L+1 /  U0|J )U ⊥
J|L+1 , (7.8)
Y0|J
ZJ|L+1 is related to the extended observability matrix as
ZJ|L+1 = OL+1 XJa (7.9)
where
 
Y0|J
h i
 U
 0|J

  ⊥
def  U
XJa = AJ OJ† CJd − AJ OJ† HJd CJs − AJ OJ† HJs  UJ|L+1 (7.10)
 J|L+1
 
E0|J / U0|J  
Y0|J
7.4 The observability matrix 145
The column space of the matrix ZJ|L+1 coincides with the column space of the
extended observability matrix OL+1 and the system order n of the SSM is given
as the dimension of the column space.
4
Proof:
See Appendix 7.10.2 and 7.10.3 for the proof of the relationship, Equations
(7.8), (7.9) and (7.10).
We will in the following give a brief discussion of rank(ZJ|L+1 ). XJa ∈
Rn×K is defined in (7.10). From (7.8) we have that ZJ|L+1 = OL+1 XJa . The
superscript a stands for autonomous. XJa can be interpreted as a state sequence
of an autonomeous system. The matrix ZJ|L+1 is proportional to the extended
observability matrix OL+1 . The rank of ZJ|L+1 is equal to the rank of OL+1
(i.e., rank(ZJ|L+1 ) = n) if rank(XJa ) = n and the pair (A, D) is observable.
This result follows from Sylvesters inequality. We have
rank(OL+1 ) + rank(XJa ) − n ≤ rank(OL+1 XJa ) ≤ min{rank(OL+1 ), rank(XJa )}
which gives
rank(XJa ) ≤ rank(ZJ|L+1 ) ≤ min{n, rank(XJa )}
Hence, the question is whether XJa lose rank below n.

We will in the following discuss some alternative formulations of XJa . We
have from (7.10) that
   
UJ|L+1
 Y0|J
 − HJd U0|J − HJs E0|J /  U0|J  

h i
 Y0|J 
 ⊥
XJa = AJ OJ† CJd CJs 
 
U0|J

U (7.11)
 J|L+1
 UJ|L+1 
 
 E / U  
0|J 0|J
Y0|J
By using Equation (7.33), this can be written as

  ⊥ 
UJ|L+1
 
 X0 + OJ† HJs E0|J  U0|J  
 
£ ¤ Y0|J 
  ⊥
XJa = AJ CJd s
CJ  U  UJ|L+1 (7.12)
  0|J  
 UJ|L+1 
 
 E0|J /  U0|J  
Y0|J
An alternative description is found by using Equations (7.36) and (7.68)

   
UJ|L+1
 X0 /  U0|J  
 
 Y 
£ J ¤ 0|J  ⊥
a
XJ = A CJ CJ  d s  U0|J U (7.13)
   J|L+1
 U 
 J|L+1 
 E / U  
0|J 0|J
Y0|J
This expression relates XJa to the reversed extended controllability matrices CJd
and CJs for the pairs (A, B) and (A, C), respectively.
At this stage our statement is that rank(XJa ) = n if:
• the pair (A, [B, C]) is controllable,

• the states X0 is sufficiently correlated with the instruments (past inputs
and outputs) U0|J and Y0|J ,
• the input is persistent exiting of a sufficient high order,
• the past noise E0|J is sufficiently correlated with the instruments (past
data) U0|J and Y0|J .
We have the following special case results.
Lemma 7.4.1 Assume J → ∞, K → ∞, J(m + r) ≤ Lm and that A is stable.

Then
 
UJ|L+1
ZJ|L+1 = (YJ|L+1 /  U0|J )U ⊥
J|L+1 =
Y0|J
 
U
 0|J 
£ ¤ UJ|L+1  ⊥
OL+1 CJd CJs   E0|J /  U0|J

  UJ|L+1
Y0|J
(7.14)
and
rank(ZJ|L+1 ) = n (7.15)
if and only if the pair A, [BC] is controllable.
4
Still another description is found by using (7.34)

 ⊥
UJ|L+1
XJa = (XJ + (AJ OJ† HJs − CJs )E0|J  U0|J  )U ⊥
J|L+1 (7.16)
Y0|J
7.4 The observability matrix 147
From (7.16) we have that XJa = XJ UJ|L+1

⊥ when the system is deterministic.
a
Hence, for a deterministic system rank(XJ ) = n if the inputs are persistently
exiting of a sufficiently high order.
A simple formulation can be found by using (7.33) and (7.67)
 
UJ|L+1
XJa = XJ /  U0|J  U⊥
J|L+1 (7.17)
Y0|J
Using the definition of the projection A/B, and providing the inverse exist, and
define the matrix of instruments as
· ¸
U0|J
W1 = (7.18)
Y0|J
then we have
XJa = XJ UJ|L+1
⊥
W1T (W1 UJ|L+1
⊥
W1T )−1 W1 UJ|L+1
⊥
(7.19)
The covariance can then be written as
1 1
X a (X a )T = ⊥
XJ UJ|L+1 W1T (W1 UJ|L+1
⊥
W1T )−1 W1 UJ|L+1
⊥
XJT (7.20)
K −1 J J K −1
For purely deterministic systems or for systems with a large signal to noise
ratio we have
1 1
XJa (XJa )T = ⊥
XJ UJ|L+1 XJT (7.21)
K −1 K −1
which is positive definite. The inputs has (partly) generated the states. The
states is therefore sufficiently correlated with the inputs.
Note also that Equation (7.8) can be expressed as follows
· ¸
def UJ|L+1 ⊥
ZJ|L+1 = (YJ|L+1 / )UJ|L+1
W1
⊥
= YJ|L+1 UJ|L+1 W1T (W1 UJ|L+1
⊥
W1T )−1 W1 UJ|L+1
⊥
⊥ ⊥
= (YJ|L+1 UJ|L+1 )/(W1 UJ|L+1 ) (7.22)
T
when UJ|L+1 UJ|L+1 ⊥
and W1 UJ|L+1 W1T are non-singular matrices.
Proof:
See Appendix 7.10.4.

7.5 Standard deviation of the estimates
The subspace methods DSR, CVA and MOESP are shown to compute con-
sistent (pole) estimates. However, in some circumstances the estimates shows
a larger variability (standard deviation) than the estimates from PEM. This
problem is addressed in this section.
N4SID (Van Overschee and De Moor (1995)) works in general only for
white noise inputs. It is shown in Di Ruscio (1996a) that an extra projection
is needed in order for the N4SID algorithm to give consistent estimates in the
case of colored inputs. A robustified algorithm is presented in Van Overschee
(1995). However, as we understand it, the ROBUST algorithm estimate the
poles similarly as the MOESP algorithm.
In the case of rich exiting input signals, e.g. such as many real world inputs
and of course white noise inputs etc., the subspace methods DSR, CVA and
MOESP computes pole estimates which is comparable with those from PEM,
both with respect to bias and variance.
For poorly exciting input signals the variance of the subspace pole estimates
is often larger than the variance of the PEM pole estimates. See e.g. Viberg
(1995).
It is therefore important to clarify the reason for the sensitivity of the esti-
mates from subspace methods in the case of poorly exciting input signals. The
question of how we can influence upon and reduce the variance of the subspace
parameter estimates is important.
In order to improve the estimates in case of poorly exiting input signals we
suggest to investigate the following ideas.
1. PCR, PLS or some other form of regularization for computing the pro-
jections and pseudo-inverse matrices.
2. Using regularization when computing the pseudo-inverse matrices.
3. Scaling
We will in the next section show how the Partial Least Squares (PLS) method
can be used to compute the projection matrices.
7.6 PLS for computing the projections
The LQ decomposition is used in the SID algorithms by Verhagen (1994), Van

Overschee and De Moor (1994) and Di Ruscio (1995), (1997).
In the following a new method for computing the projection matrices which
is based on the Partial Least Squares (PLS) method is presented. Based on
numerical examples, we have found that this method may reduce the variability
of the pole estimates in case of poorly exciting input signals. The reason for this
7.6 PLS for computing the projections 149
is that the projection matrices are computed without computing the pseudo-
inverse of badly conditioned matrices. This results is to our knowledge new.
For simplicity in the following, define
 
UJ|L+1
def
Ỹ =  U0|J  ∈ Rp×K (7.23)
Y0|J
where the number of rows in Ỹ are p = (L + 1 + J)r + Jm.

A 1st PLS decomposition is computed as follows. The arguments to the PLS
algorithm are the data matrix of future outputs YJ|L+1 and the data matrix with
past inputs and outputs and future inputs Ỹ . The PLS algorithm is iterative.
The number of iterations is equal to the number of components a1 which is
bounded by 1 ≤ a1 ≤ p. Normally, a1 can be chosen as the effective rank of
Ỹ . The problem of choosing the number of components should be discussed in
some details.
We have the following decompositions
Ỹ = PT + G (7.24)
YJ|L+1 = CT + F (7.25)
where T ∈ Ra1 ×K is an orthonormal matrix, i.e., T T T = Ia1 , P ∈ Rp×a1 is a

loading matrix for Ỹ , C ∈ R(L+1)m×a1 is a loading matrix for YJ|L+1 . G ∈ Rp×K
and F ∈ R(L+1)m×K are residual matrices.
It can be shown that the projection of the row space of YJ|L+1 onto the row
space of Ỹ is equal to
d def
ZJ|L+1 = YJ|L+1 /Ỹ = CT (7.26)
d
We now have to remove the effect of future outputs from ZJ|L+1 in order to
recover the extended observability matrix of the system. From (7.24) we have
that
UJ|L+1 = P1 T + G1 (7.27)
where P1 ∈ R(L+1)r×a1 is equal to the first (L + 1)r rows in matrix P , similarly

G1 ∈ R(L+1)r×K is equal to the first (L + 1)r rows in the residual matrix G.
d
The projection of the row space of matrix ZJ|L+1 onto the orthogonal com-
plement of the row space of UJ|L+1 can then be computed as
def
d ⊥
ZJ|L+1 = ZJ|L+1 UJ|L+1 = C − CP1T (P1 P1T )† P1 (7.28)
However, in case of a badly conditioned matrix P1 P1T a second PLS can be

computed with matrices C and P1 as arguments. We have the following decom-
positions
P1 = P2 T2 + G2 (7.29)
C = C2 T2 + F2 (7.30)
where T2 ∈ Ra2 ×a1 is orthonormal. P2 ∈ R(L+1)r×a2 is a loading matrix for

P1 , C2 ∈ R(L+1)m×a2 is a loading matrix for C. G2 ∈ R(L+1)r×a1 and F2 ∈
R(L+1)m×a1 are residual matrices. From this we have the alternative to (7.28)
def
d ⊥
ZJ|L+1 = ZJ|L+1 UJ|L+1 = C − C2 T2T . (7.31)
The number of components a2 used to compute the 2nd PLS decomposition

should be bounded by 1 ≤ a2 ≤ a1 .
Usually one can put a1 = rank(Ỹ ) for the 1st PLS decomposition and
a2 = rank(P1 ) for the 2nd PLS decomposition. In practical situations it may
be difficult to compute the (effective) rank.
One should note that the effect of choosing a1 and a2 smaller than the
ranks of Ỹ and UJ|L+1 , respectively, may be that the variance of the estimates
is reduced. However, a bias may be included in the estimates. The problem of
choosing a1 and a2 is in general a trade off between bias and variance. This is
a form of regularization.
Monte Carlo simulations shows that the estimates when choosing a1 = p and
a1 = rank(Ỹ ), i.e., choosing the number of components in the 1st PLS equal to
the row size of Ỹ or equal to the rank, and using (7.28) gives approximately the
same results as using the LQ decomposition for computing the same projections.
This will be illustrated in the next section. The effect of choosing a1 < rank(Ỹ )
is that a bias in general is introduced in the estimates, however, the variance
may be reduced.
7.7 Numerical example
The results from a Monte Carlo simulation of a MIMO system are illustrated
in Figure (7.1). The number of components was chosen as a1 = rank(Ỹ ) in
the PLS decomposition used to compute the matrix ZJ|L+1 . The number of
samples was N = 5000. The input signal was chosen as
· t
¸
(sin( 14 ) + sin( 7t ) + sin( 3t ))k1
ut = t t (7.32)
(sin( 25 ) + sin( 10 ) + sin( 5t ) + sin(t))k2
where k1 and k2 are positive scalar parameters. Our experience is that in case
of porely exiting input signals the parameter L in the DSR algorithm (which
is the number of block rows in the extended observability matrix and which
can be interpreted as the horizon used to predict the number of states) should
be chosen as small as possible in order to reduce the variance of the estimates.
For the DSR method we have the condition n ≤ Lm. For this example n = 3
and m = 2. Hence, the smallest possible parameter is L = 2. The results from
the DSR methods obtained from using both the LQ and PLS approaches are
illustrated in Figure 7.1.
The results from the ROBUST method by Van Overschee (1995) is also
presented in Figure 7.1. The parameter for the number of block rows in this
7.8 Conclusion 151
method is I. The smallest possible parameter was found to be I = 3 which was

used in the simulations.
The results from the prediction error method PEM, Ljung (1991), are also
presented in Figure 7.1.
This example shows a larger variability in the complex pole estimates from
ROBUST than the corresponding pole estimates from DSR. The estimates
from DSR is close to the estimates from PEM. An open question is: can the
variance of the pole estimates from the SID algorithms be further reduced ?
DSR_lq ROBUST
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0.6 0.7 0.8 0.6 0.7 0.8
DSR_pls PEM
0.5 0.5
0.4 0.4
0.3 0.3
Im
0.2 0.2
0.1 0.1
0 0
0.6 0.7 0.8 0.6 0.7 0.8
Re
states. DSR parameters L = J = 2 for both the LQ and PLS method for com-
puting the projection matrices. ROBUST parameter I = 3. PEM stands for
the prediction error method in the System Identification Toolbox for MATLAB.
7.8 Conclusion
It is shown that the partial least squares method is an alternative to the stan-
dard LQ decomposition for computing the range space of the extended observ-
ability matrix.
7.9 References
May 1996.




Academic Press.



matica, vol. 30, No. 1, pp.75-94.


pp. 1853-1864.

matica, vol. 30, No. 1, pp. 61-74.

Wold, H. (1966). Non-linear estimation by iterative least squares procedures.

Researsch papers in statistics, Ed. David, F. Wiley, New York, pp. 411-
444.
7.10 Appendix: proofs 153
7.10 Appendix: proofs
7.10.1 Proof of Equation (7.4)
We have
d s
YJ|L+1 = OL+1 XJ + HL+1 UJ|L+1 + HL+1 EJ|L+1 (7.33)
The relationship between the “future” sequence of states, XJ , and the “past”
sequence of states, X0 , is given by
XJ = AJ X0 + CJd U0|J + CJs E0|J (7.34)
Substituting Equation (7.34) into equation (7.33) in order to eliminate XJ gives
YJ|L+1 = OL+1 (AJ X0 + CJd U0|J + CJs E0|J ) + HL+1

d s
UJ|L+1 + HL+1 EJ|L+1(7.35)
An expression for the sequence of past states, X0 , is determined from
Y0|J = OJ X0 + HJd U0|J + HJs E0|J (7.36)
Equation (7.36) gives
X0 = OJ† (Y0|J − HJd U0|J − HJs E0|J ) (7.37)
Substituting Equation (7.37) into Equation (7.35) in order to eliminate X0 gives
YJ|L+1 = OL+1 AJ OJ† (Y0|J − HJd U0|J − HJs E0|J ) + OL+1 CJd U0|J + OL+1 CJs E0|J
d s
+ HL+1 UJ|L+1 + HL+1 EJ|L+1 (7.38)
which can be written as
YJ|L+1 = OL+1 AJ OJ† Y0|J + (OL+1 CJd − OL+1 AJ OJ† HJd )U0|J
+ (OL+1 CJs − OL+1 AJ OJ† HJs )E0|J
d s
+ HL+1 UJ|L+1 + HL+1 EJ|L+1 (7.39)
which is identical to Equation (7.4), Theorem 7.3.1. QED.

7.10.2 Proof I of Equation (7.10)
Projection I: (remove effect of future noise, EJ|L+1 )

From Equation (7.4) we have
   
UJ|L+1 £ ¤ U J|L+1
YJ|L+1 /  U0|J  = Hd d J
L+1 PL+1 ÃL+1
 U0|J 
Y0|J Y0|J
 
UJ|L+1
s E
+PL+1 0|J /  U0|J  + dE1 (7.40)
Y0|J
where
 
UJ|L+1
def s
dE1 = HL+1 EJ|L+1 /  U0|J  (7.41)
Y0|J
and
 
UJ|L+1
s
lim dE1 = HL+1 ( lim EJ|L+1 /  U0|J ) = 0(L+1)m×K (7.42)
K→∞ K→∞
Y0|J
Rearranging the terms on the right hand side in Equation (7.40) gives
 
UJ|L+1
YJ|L+1 /  U0|J =
Y0|J
 
U0|J
 
£ d ¤  Y0|J   
J s
PL+1 ÃL+1 PL+1   UJ|L+1  + H d UJ|L+1 + dE1(7.43)
 L+1
 E0|J /  U0|J  
Y0|J
Projection II: (remove effect of future inputs, UJ|L+1 )
 
UJ|L+1
(YJ|L+1 /  U0|J )U ⊥
J|L+1 =
Y0|J
 
U0|J
 
£ d ¤  Y0|J    ⊥
PL+1 ÃL+1 PL+1 
s
 UJ|L+1 U
 J|L+1 + dE2 (7.44)
 E0|J /  U0|J  
Y0|J
where
def ⊥
dE2 = dE1 UJ|L+1 (7.45)
Equation (7.44) is equal to Equation (7.10) in Theorem 7.4.1. Note that
limK→∞ dE2 = 0(L+1)m×K . QED.
7.10.3 Proof II of Equation (7.10)
From Theorem 7.3.1, Equation (7.4) we have

 
£ ¤ UJ|L+1
YJ|L+1 = d
HL+1 d
PL+1 ÃJL+1  U0|J  + PL+1
s s
E0|J + HL+1 EJ|L+1(7.46)
Y0|J
The terms defined by the future inputs and future noise, HL+1d U
J|L+1 and
s
HL+1 EJ|L+1 , respectively, does not have the same column space as the ex-
tended observability matrix. These terms has to be removed. We have
 ⊥  ⊥
UJ|L+1 UJ|L+1
YJ|L+1  U0|J  − PL+1
s
E0|J  U0|J  = HL+1
s
EJ|L+1 + dE1 (7.47)
Y0|J Y0|J
We have assumed that the noise is zero mean and white. Future noise is there-
fore uncorrelated with past noise and past and future inputs. Hence we have
limK→∞ dE1 = 0(L+1)m×K .
s
Substituting (7.47) into (7.46) in order to eliminate the term HL+1 EJ|L+1
gives
 
UJ|L+1
YJ|L+1 /  U0|J =
Y0|J
 
U0|J
 
£ ¤  Y0|J   
d
PL+1 ÃJL+1 PL+1
s  UJ|L+1  + H d UJ|L+1 + dE1(7.48)
  L+1
 E0|J /  U0|J  
Y0|J
⊥
Post-multiplying with UJ|L+1 gives
 
UJ|L+1
def
ZJ|L+1 = YJ|L+1 /  U0|J  U⊥ ⊥
J|L+1 = OL+1 Xa + dE1 UJ|L+1 (7.49)
Y0|J
where
 
Y0|J
h i
 U
 0|J

  ⊥
def  U
Xa = AJ OJ† CJd − AJ OJ† HJd CJs − AJ OJ† HJs  UJ|L+1  J|L+1
 
E0|J / U0|J  
Y0|J
Hence, ZJ|L+1 has the same column space as the extended observability matrix.
For simplicity define

 
UJ|L+1 · ¸
def  = U J|L+1
Ỹ = U0|J (7.50)
W1
Y0|J
Theorem 7.4.1, Equation (7.10), yields

⊥
ZJ|L+1 = (YJ|L+1 /Ỹ )UJ|L+1 = YJ|L+1 Ỹ T (Ỹ Ỹ T )−1 Ỹ UJ|L+1
⊥
(7.51)
Notice that
· ¸ · ¸
⊥ UJ|L+1 ⊥ 0(L+1)r×K
Ỹ UJ|L+1 = UJ|L+1 = ⊥ (7.52)
W1 W1 UJ|L+1
The inverse can be written

" #−1 · ¸
T
UJ|L+1 UJ|L+1 UJ|L+1 W1T
T −1 def I11 I12
(Ỹ Ỹ ) = T = (7.53)
W1 UJ|L+1 W1 W1T I21 ∆−1
For simplicity define Uf = UJL+1 . We have
I12 = −(Uf UfT )−1 Uf W1T ∆−1 (7.54)

∆ = W1 W1T − W1 UfT (Uf UfT )−1 Uf W1T = W1 Uf⊥ W1T (7.55)
where we have used index f for future. ∆ is the Schur complement of Uf UfT .
See e.g. Kailath (1980). We now have
· ¸· ¸
£ ¤ I11 I12 0(L+1)r×K
ZJ|L+1 = Yf UfT Yf W1T ⊥
I21 ∆−1 W1 UJ|L+1
· ¸
£ ¤ I12 ⊥
= Yf UfT Yf W1T W1 UJ|L+1
∆−1
= (Yf W1T ∆−1 + Yf UfT I12 )W1 UJ|L+1
⊥
= (Yf − Yf UfT (Uf UfT )−1 Uf )W1T ∆−1 W1 UJ|L+1

⊥
⊥
= Yf UJ|L+1 W1T ∆−1 W1 UJ|L+1
⊥
(7.56)
7.10.5 Including constraints: regularization
From Equation 7.4 we have
Yf Ỹ T = M Ỹ Ỹ T + PL+1
s
Ep Ỹ T + HL+1
s
Ef Ỹ T
= M (Ỹ Ỹ T + Λ) + PL+1
s
Ep Ỹ T + HL+1
s
Ef Ỹ T − M Λ (7.57)
and
Yf Ỹ T (Ỹ Ỹ T + Λ)−1 = M + PL+1

s
Ep Ỹ T (Ỹ Ỹ T + Λ)−1 + (HL+1
s
Ef Ỹ T − M Λ)(Ỹ Ỹ T + Λ)−1
(7.58)
and
Yf Ỹ T (Ỹ Ỹ T + Λ)−1 Ỹ = M Ỹ + PL+1

s
Ep Ỹ T (Ỹ Ỹ T + Λ)−1 Ỹ + dE1 (7.59)
where
s
dE1 = (HL+1 Ef Ỹ T − M Λ)(Ỹ Ỹ T + Λ)−1 Ỹ (7.60)
Rearranging Equation (7.59) gives
Yf Ỹ T (Ỹ Ỹ T + Λ)−1 Ỹ =
 
£ ¤ U 0|J
d
PL+1 ÃJL+1 PL+1
s  Y0|J  + H d Uf + dE(7.61)
1
L+1
T T
E0|J Ỹ (Ỹ Ỹ + Λ) Ỹ −1
and finally
Yf Ỹ T (Ỹ Ỹ T + Λ)−1 Ỹ Uf⊥ =

 
£ ¤ U0|J
d
PL+1 ÃJL+1 s
PL+1  Y0|J  U ⊥ + dE2 (7.62)
f
E0|J Ỹ T (Ỹ Ỹ T + Λ)−1 Ỹ
where
Uf⊥ = I − UfT (Uf UfT + Λ2 )−1 Uf (7.63)

dE2 = dE1 Uf⊥ + error term (7.64)
A non-zero matrix Λ will in general include a bias in the estimates and also
affect the variance. Consider the following block-diagonal matrix
 
Λ11 0 0
Λ= 0 Λ22 0  (7.65)
0 0 Λ33
The term M Λ in the error equation 7.60 is then

d s
M Λ = HL+1 Λ11 + PL+1 Λ22 + ÃJL+1 Λ33 (7.66)
Notice that Λ11 = 0(L+1)r×(L+1)r gives consistent estimates.

7.10.6 Some useful results

· ¸
Uf
Uf / = Uf (7.67)
W1

· ¸⊥
Uf
Uf =0 (7.68)
W1
Proof of Lemma 7.10.1:
From the definition of A/B, Equation (7.1), we have

· ¸ · ¸−1 · ¸
Uf £ ¤ Uf UfT Uf W1T Uf
Uf / = Uf UfT W1T (7.69)
W1 W1 UfT W1 W1T W1
The inverse of a partitioned matrix is (Kailath (1980))

· ¸−1
Uf UfT Uf W1T
=
W1 UfT W1 W1T
· ¸
(Uf UfT )−1 + (Uf UfT )−1 Uf W1T ∆−1 W1 UfT (Uf UfT )−1 −(Uf UfT )−1 Uf W1T ∆−1
−∆−1 W1 UfT (Uf UfT )−1 ∆−1
(7.70)
where ∆ is the Schur complement of Uf UfT . See A.22. Using the expression for
the inverse, Equation (7.70), then we have that
· ¸−1
£ ¤ Uf UfT Uf W1T £ ¤
Uf UfT Uf W1T = I 0 (7.71)
W1 UfT W1 W1T
Substituting (7.71) into (7.67) gives

· ¸ · ¸
Uf £ ¤ Uf
Uf / = I 0 = Uf (7.72)
W1 W1
and Lemma 7.10.1 is proved. QED.
Proof of Lemma 7.10.2:
From the definition of AB ⊥ , Equation (7.2), we have

· ¸⊥ · ¸
Uf Uf
Uf = Uf − Uf / =0 (7.73)
W1 W1
where we have used Lemma 7.10.1 with proof. QED.

7.10.7 Proof III of Equation (7.10)
The ESSM constructed from future data is

YJ+1|L = ÃL YJ|L + B̃L UJ|L+1 + C̃L EJ|L+1 (7.74)
Using Lemma 7.10.1 we get
d d
ZJ+1|L = ÃL ZJ|L + B̃L UJ|L+1 + C̃L dEJ|L+1 (7.75)
where
 
UJ|L+1
def
d
ZJ+1|L = YJ+1|L /  U0|J  (7.76)
Y0|J
 
UJ|L+1
def
d
ZJ|L = YJ|L /  U0|J  (7.77)
Y0|J
 
UJ|L+1
def
dEJ|L+1 = EJ|L+1 /  U0|J  (7.78)
Y0|J
Due to assumptions (ergodic noise process) the error term dEJ|L+1 tends to
zero w.p.1 as the number of columns K tends to infinity.
Hence,
d d
(A, B, D, E) = arg min k ZJ+1|L − ÃZJ|L − B̃UJ|L+1 kF (7.79)
A,B,D,E
7.10.8 Prediction of future outputs
We have the following relationship

d s
YJ|L+1 = ZJ|L+1 + ZJ|L+1 (7.80)
From Theorem 4.3, Di Ruscio (1997), we have that the innovations can be
estimated from
s s
F EJ+L|1 = ZJ+L|1 (ZJ|L )⊥ (7.81)
from the above we have that a predictor for the future outputs is given by
d s
ŷJ+L|1 = ZJ+L|1 (ZJ|L )⊥ (7.82)
This indicates that
s
yJ+L|1 = YJ+L|1 (ZJ|L )⊥ (7.83)
The predictor can simply be written
d
ŷJ+L|1 = ZJ+L|1 (7.84)
Chapter 8
Subspace identification of the

states and the Kalman filter
gain
8.1 Introduction
A complete subspace identification (SID) algorithm are discussed and derived

in this paper. The derivation presented is different from the other published pa-
pers on subspace identification, Van Overschee and De Moor (1994), Larimore
(1990), Viberg (1995) and Van Overschee (1995) and the references therein,
because we are using general input and output matrix equations which de-
scribes the relationship between the past and the future input and output data
matrices.
One of the contributions in this paper is that it is shown that the Kalman
filter model matrices, including the Kalman gain and the noise innovations
process, of a combined deterministic and stochastic system can be identified
directly from certain projection matrices which are computed from the known
input and output data, without solving any Riccati or Lyapunov matrix equa-
tions. This subspace method and results was presented without proof in Di
Ruscio (1995) and Di Ruscio (1997). One contribution in this paper is a com-
plete derivation with proof. A new method for computing the matrices in the
deterministic part of the system is presented. This method has been used in
the DSR Toolbox for Matlab, Di Ruscio (1996), but has not been published
earlier.
Furthermore, it is pointed out that the states, in general (i.e. for colored
input signals), only can be computed if the complete deterministic part of the
model is known or identified first. This is probably the reason for which the state
based subspace algorithms which are presented in the literature does not work
properly for colored input signals. The SID algorithm in Verhagen (1994) works
for colored input signals. The stochastic part of the model is not computed by
162 Subspace identification of the states and the Kalman filter gain
this algorithm. The N4SID algorithm in Van Overschee and De Moor (1994)
works well and only for white input signals. The stochastic part of the model is
computed by solving a Riccati equation. However, the robust modification in
Van Overschee and De Moor (1995) works well also for colored input signals.
The rest of this paper is organized as follows. Some basic matrix definitions
and notations are presented in Section 8.2. The problem of subspace identi-
fication of the states for both colored and white input signals is discussed in
Section 8.3.1. The subspace identification of the extended observability matrix,
which possibly is the most important step in any SID algorithm, are discussed
in Section 8.3.2. It is proved that the Kalman filter gain matrix and the noise
innovations process can be identified directly from the data in Section 8.3.3.
A least squares optimal method for computing the deterministic part of the
combined deterministic and stochastic system is presented in Section 8.3.4.
The problem of using subspace methods for closed loop systems are pointed
out and some solutions to the problem are pointed out in section 8.4.
The main contribution in this paper is a new method for subspace system
identification that works for closed loop as well as open loop systems. The
method are based on the theory in Section 3 and is presented in Section 8.5.
This method is probably one of the best for closed loop subspace system iden-
tification.
Some topics and remarks related to the algorithm are presented in Section
8.6. Numerical examples are provided in Section 8.7 in order to illustrate the
behaviour of the algorithm both in open and closed loop. Some concluding
remarks follows in Section 8.8.
8.2 Notation and definitions
8.2.1 System and matrix definitions
Consider the following state space model on innovations form
x̄k+1 = Ax̄k + Buk + Cek , (8.1)

yk = Dx̄k + Euk + F ek , (8.2)
where ek is white noise with covariance matrix E(ek eTk ) = Im . One of the
problems addressed and discussed in this paper is to directly identify (subspace
identification) the system order, n, the state vector x̄k ∈ Rn , and the matrices
(A, B, C, D, E, F ) from a sequence of known input and output data vectors, uk ,
∈ Rr and yk , ∈ Rm , respectively. A structure parameter, g, is introduced so
that g = 1 when E is to be identified and g = 0 when E is a-priori known
to be zero. This should be extended to a structure matrix G with ones and
zeroes, the ones pointing to the elements in E which are to be estimated. This
is not considered further here. Based on (8.1) and (8.2) we make the following
definitions for further use:
8.2 Notation and definitions 163
Definition 8.1 (Basic matrix definitions)

The extended observability matrix, Oi , for the pair (D, A) is defined as
 
D
 DA 
def  
Oi =  ..  ∈ Rim×n , (8.3)
 . 
DAi−1

The reversed extended controllability matrix, Cid , for the pair (A, B) is
defined as
def £ ¤
Cid = Ai−1 B Ai−2 B · · · B ∈ Rn×ir , (8.4)
where the subscript i denotes the number of block columns. A reversed extended
controllability matrix, Cis , for the pair (A, C) is defined similar to (8.4), i.e.,
def £ ¤
Cis = Ai−1 C Ai−2 C · · · C ∈ Rn×im , (8.5)
i.e., with B substituted with C in (8.4). The lower block triangular Toeplitz
matrix, Hid , for the quadruple matrices (D, A, B, E)
 
E 0m×r 0m×r ··· 0m×r
 DB E 0m×r ··· 0m×r 
 
def  DAB DB E ··· 0m×r 
Hid =   ∈ Rim×(i+g−1)r , (8.6)
 .. .. .. .. .. 
 . . . . . 
where the subscript i denotes the number of block rows and i + g − 1 is the
number of block columns. Where 0m×r denotes the m × r matrix with zeroes.
A lower block triangular Toeplitz matrix His for the quadruple (D, A, C, F ) is
defined as
 
F 0m×m 0m×m ··· 0m×m
 DC F 0m×m ··· 0m×m 
 
def  DAC DC F ··· 0m×m 
His =   ∈ Rim×im . (8.7)
 .. .. .. .. .. 
 . . . . . 
8.2.2 Hankel matrix notation
Hankel matrices are frequently used in realization theory and subspace sys-
tem identification. The special structure of a Hankel matrix as well as some
matching notations, which are frequently used througout, are defined in the
following.
Definition 8.2 (Hankel matrix) Given a (vector or matrix) sequence of data
st ∈ Rnr×ns ∀ t = 0, 1, 2, . . . , t0 , t0 + 1, . . . , (8.8)
where nr is the number of rows in st and nc is the number of columns in st .

Define integer numbers t0 , L and K and define the matrix St as follows
 
st0 st0 +1 st0 +2 · · · st0 +K−1
def 
 st0 +1 st0 +2 st0 +3 · · · st0 +K 

St0 |L =  . .. .. . . ..  ∈ RLnr×Knc . (8.9)
 .. . . . . 
st0 +L−1 st0 +L st0 +L+1 · · · st0 +L+K−2
which is defined as a Hankel matrix because of the special structure. The integer
numbers t0 , L and K are defined as follows:
• t0 start index or initial time in the sequence, st0 , which is the upper left
block in the Hankel matrix.
• L is the number of nr-block rows in St0 |L .
• K is the number of nc-block columns in St0 |L .
A Hankel matrix is symmetric and the elements are constant across the anti-
diagonals. We are usually working with vector sequences in subspace system
identification, i.e., st is a vector in this case and hence, nc = 1. Examples of
such vector processes, to be used in the above Hankel-matrix definition, are the
measured process outputs, yt ∈ Rm , and possibly known inputs, ut ∈ Rr . Also
define
def £ ¤T
yj|i = yjT yj+1 T T
· · · yj+i−1 ∈ Rim , (8.10)
which is refereed to as an extended (output) vector, for later use.
8.2.3 Projections
Given two matrices A ∈ Ri×k and B ∈ Rj×k . The orthogonal projection of the
row space of A onto the row space of B is defined as
A/B = AB T (BB T )† B. (8.11)
The orthogonal projection of the row space of A onto the orthogonal comple-
ment of the row space of B is defined as
AB ⊥ = A − A/B = A − AB T (BB T )† B. (8.12)
The following properties are frequently used

· ¸
A
A/ = A, (8.13)
B
8.3 Subspace system identification 165
· ¸⊥
A
A/ = 0. (8.14)
B
Prof of (8.13) and (8.14) can be found in e.g., Di Ruscio (1997b). The Moore-
Penrose pseudo-inverse of a matrix A ∈ Ri×k where k > i is defined as A† =
AT (AAT )−1 . Furthermore, consistent with (8.12) we will use the definition
B ⊥ = Ik − B T (BB T )† B, (8.15)
throughout the paper. Note also the properties that (B ⊥ )T = B ⊥ and B ⊥ B ⊥ =
B⊥.
8.3 Subspace system identification
8.3.1 Subspace identification of the states
Consider a discrete time Kalman filter on innovations form, i.e.,

x̄k+1 = Ax̄k + Buk + Kεk , (8.16)
yk = Dx̄k + Euk + εk , (8.17)
where x̄k ∈ Rn is the predicted state in a minimum variance sense, εk ∈ Rm
is the innovations at discrete time k, i.e., the part of yk ∈ Rm that cannot
be predicted from past data (i.e. known past inputs and outputs) and the
present input. Furthermore, ȳk = Dx̄k + Euk is the prediction of yk , and
εk is white noise with covariance matrix ∆ = E(εk εTk ). Here εk = F ek is
the innovations and the model (8.1) and (8.2) is therefore equivalent with the
Kalman filter (8.16) and (8.17). Furthermore, we have that K = CF −1 and
∆ = E(εk εTk ) = F F T , when F is non-singular, i.e., when the system is not
deterministic and when the Kalman filter exists.
A well known belief is that the states is a function of the past. Let us have
a lock at this statement. The predicted state at time k := t0 + J, i.e. x̄t0 +J of
a Kalman filter with the initial predicted state at k := t0 , i.e. x̄t0 given, can be
expressed as
x̄t0 +J = C̃Js yt0 |J + C̃Jd ut0 |J + (A − KD)J x̄t0 , (8.18)
where C̃Js = CJ (A − KD, K) is the reversed extended controllability matrix of

the pair (A − KD, K), C̃Jd = CJ (A − KD, B − KE) is the reversed extended
controllability matrix of the pair (A − KD, B − KE) and x̄t0 is the initial
predicted state (estimate) at the initial discrete time t0 . See (8.5) for the
definition of the reversed controllability matrix. J is the past horizon, i.e., the
number of past outputs and inputs used to define the predicted state (estimate)
x̄t0 +J at the discrete time t0 + J.
Using (8.18) for different t0 , i.e. for t0 , t0 + 1, t0 + 2, . . ., t0 + K − 1, gives
the matrix equation
Xt0 +J = C̃Js Yt0 |J + C̃Jd Ut0 |J + (A − KD)J Xt0 , (8.19)
where
£ ¤
Xt0 +J = x̄t0 +J x̄t0 +J+1 · · · x̄t0 +J+K−1 ∈ Rn×K , (8.20)
£ ¤
Xt0 = x̄t0 x̄t0 +1 · · · x̄t0 +K−1 ∈ Rn×K . (8.21)
where K is the number of columns in the data matrices. Note that K also is
equal to the number of vector equations of the form (8.18) which is used to
form the matrix version (8.19). Note also that the state matrix Xt0 can be
eliminated from (8.19) by using the relationship
Yt0 |J = OJ Xt0 + HJd Ut0 |J+g−1 + HJs Et0 |J , (8.22)
which we have deduced from the innovations form, state space model (8.1) and
(8.2). Putting t0 =: t0 + J in (8.22) gives
Yt0 +J|L = OL Xt0 +J + HLd Ut0 +J|L+g−1 + HLs Et0 +J|L . (8.23)
Using (8.19) to eliminate Xt0 +J in (8.23) gives a matrix equation which re-
lates the future data matrices Yt0 +J|L , Ut0 +J|L+g−1 , Et0 +J|L and the past data
matrices Yt0 |J , Ut0 |J , Et0 |J .
The data is usually defined at time instant (or number of observations)
k = 1, 2, . . . , N . Hence, t0 = 1 in this case. However, we are often defining
t0 = 0 which corresponds to data defined at k = 0, 1, . . . , N − 1. The bar used
to indicate predicted state is often omitted. Hence, for simplicity of notation,
we define the following equations from (8.19), (8.22) and (8.23),
Y0|J = OJ X0 + HJd U0|J+g−1 + HJs E0|J , (8.24)
XJ = C̃Js Y0|J + C̃Jd U0|J + (A − KD)J X0 , (8.25)
YJ|L = OL XJ + HLd UJ|L+g−1 + HLs EJ|L , (8.26)
for lather use. Furthermore, (8.26) and (8.25) gives

 
£ d ¤ UJ|L+g−1
YJ|L = HL OL C̃Jd OL C̃Js  U0|J  + OL (A − KD)J X0 + HLs EJ|L .
Y0|J
(8.27)
Equation (8.27) is important for understanding a SID algorithm, because, it

gives the relationship between the past and the future. Note also the terms
in (8.27) which are ”proportional” with the extended observability matrix OL .
From (8.27) we see that the effect from the future inputs, UJ|L+g−1 , and the
future noise, EJ|L , have to be removed from the future outputs, YJ|L , in order
to recover the subspace spanned by the extended observability matrix, OL . A
variation of this equation, in which the term X0 is eliminated by using (8.22)
or (8.24) is presented in Di Ruscio (1997b). Note also that (8.25) and (8.24)
gives
· ¸
£ ¤ U0|J
XJ = PJu PJy − PJe E0|J , (8.28)
Y0|J
PJu = C̃Jd − (A − KD)J OJ† HJd , (8.29)

PJy = C̃Js + (A − KD)J OJ† , (8.30)
PJe = (A − KD)J OJ† HJs , (8.31)
where we for the sake of simplicity and without loss of generality have put
g = 1. Equation (8.28) is useful because it shows that the future states XJ
is in the range of a matrix consisting of past inputs, U0|J , and past outputs,
Y0|J (in the deterministic case or when J → ∞). Note that we have introduced
the notation, PJu , in order to represent the influence from the past inputs upon
the future. Combining (8.28) and (8.26) gives an alternative to (8.27), i.e. the
”past-future” matrix equation,
 
£ ¤ UJ|L+g−1
YJ|L = HLd OL PJu OL PJy  U0|J  − OL PJe E0|J + HLs EJ|L . (8.32)
Y0|J
The two last terms in (8.32) cannot be predicted from data, i.e., because E0|J
and EJ|L are built from the innovations process ek .
It is important to note that a consistent estimate of the system dynamics
can be obtained by choosing L and N properly. Choosing Lmin ≤ L where
Lmin = n + rank(D) − 1 and letting N → ∞, is in general, necessary conditions
for a consistent estimate of the dynamics. See Section 8.3.2 for further details.
On the other side, it is in general, also necessary to let J → ∞ in order to
obtain a consistent estimate of the states. The reason for this is that the term
(A − KD)J = 0 in this case. Hence, the effect of the initial state matrix X0 on
the future states XJ has died out. We have the following Lemma
Lemma 8.3.1 (Subspace identification of the states)

Let K → ∞ in the data matrices. The projected state matrix is defined as
d
ZJ|L
  z  }| {
UJ|L+g−1 UJ|L+g−1
XJ /  U0|J  = O† (YJ|L /  U0|J  −HLd UJ|L+g−1 )
L
Y0|J Y0|J
 
UJ|L+g−1
= C̃Js Y0|J + C̃Jd U0|J + (A − KD)J X0 /  U0|J .
Y0|J
(8.33)
Consider the case when

 
UJ|L+g−1
(A − KD)J X0 /  U0|J  = 0, (8.34)
Y0|J
which is satisfied when J → ∞ and (A − KD) is stable. This gives

 
UJ|L+g−1
XJ /  U0|J  = XJ , (8.35)
Y0|J
and hence we have, in general, the following expression for the future states
d
ZJ|L
z  }| {
UJ|L+g−1
†
XJ = OL (YJ|L /  U0|J  −HLd UJ|L+g−1 ). (8.36)
Y0|J
Proof 8.1 The proof is divided into two parts.
Part 1
The relationship between the future data matrices is given by
YJ|L = OL XJ + HLd UJ|L+g−1 + HLs EJ|L . (8.37)

 
UJ|L+g−1
Projecting the row space of each term in (8.37) onto the row space of  U0|J 
Y0|J
gives
   
YJ|L /  U0|J  = OL XJ /  U0|J  + HLd UJ|L+g−1 + dE1 (8.38)
Y0|J Y0|J
where the error term is given by

 
UJ|L+g−1
dE1 = HLs EJ|L /  U0|J . (8.39)
Y0|J
It make sense to assume that future noise matrix EJ|L is uncorrelated with past
data and the future inputs, hence, we have that (w.p.1)
lim dE1 = 0. (8.40)

K→∞
Part 2
Equation (8.25) gives the relationship between the future state matrix XJ and
the past data matrices. Projecting the row space of each term in this equation
UJ|L+g−1

onto the row space of U0|J  gives
Y0|J
   
XJ /  U0|J  = C̃Js Y0|J + C̃Jd U0|J + (A − KD)J X0 /  U0|J  (. 8.41)
Y0|J Y0|J
Letting J → ∞ (or assuming the last term to be zero) gives
 
UJ|L+g−1
XJ /  U0|J  = C̃Js Y0|J + C̃Jd U0|J . (8.42)
Y0|J
Letting J → ∞ and assuming the system matrix (A − KD) for the predicted
outputs to be stable in (8.25) shows that
XJ = C̃Js Y0|J + C̃Jd U0|J . (8.43)
Comparing (8.42) and (8.43) gives
 
UJ|L+g−1
XJ = XJ /  U0|J . (8.44)
Y0|J
Using (8.44) in (8.38) and solving for XJ gives (8.36). 2
The condition in (8.35) is usually satisfied for large J, i.e., we have that
limJ→∞ (A − KD)J = 0 when A − KD is stable. Note also that the eigen-
values of A − KD usually are close to zero for “large” process noise (or “small”
measurements noise). Then, (A − KD)J is approximately zero even for rela-
tively small numbers J. We will now discuss some special cases
Lemma 8.3.2 (SID of states: white input)

Consider a combined deterministic and stochastic system excited with a white
input signal. Then
 
UJ|L+g−1
†
XJ = OL YJ|L /  U0|J  U⊥
J|L+g−1 (8.45)
Y0|J
when J → ∞.
Proof 8.2 This result follows from the proof of Lemma 8.3.1 and (8.36) and
using that
⊥
XJ UJ|L+g−1 = XJ (8.46)
when uk is white and, hence, X0 /UJ|L+g−1 = 0. 2
Lemma 8.3.3 (SID of states: pure stochastic system)

Consider a stochastic system. Then we simply have that
†
XJ = OL YJ|L /Y0|J (8.47)
when J → ∞ or when (A − KD)J X0 /Y0|J = 0 is satisfied.
Proof 8.3 This result follows from the proof of Lemma 8.3.1 by putting the
measured input variables equal to zero. 2
Lemma 8.3.1 shows that it is in general (i.e. for colored input signals)
necessary to know the deterministic part of the system, i.e., the Toepliz matrix
HLd in (8.36), in order to properly identify the states. This means that the
matrices B and E in addition to D and A has to be identified prior to computing
the states. I.e. we need to know the deterministic part of the model. However,
a special case is given by Lemma 8.3.2 and Equation (8.45) which shows that
the states can be identified directly when the input signals is white. Note also
that the extended observability matrix OL is needed in (8.36) and (8.45). OL
can be identified directly from the data. This is proved in the next Section
8.3.2, and this is indeed the natural step in a SID algorithm.
In the case of a white input signal or when J → ∞ then, HLd , and the state
matrix, XJ , can be computed as by the N4SID algorithm, Van Overschee and
De Moor (1996). From (8.32) and (8.28) we have the following lemma
Lemma 8.3.4 (States, XJ , and Toepliz matrix HLd : N4SID)

The following LS solution
 †
£ d ¤ U J|L+g−1
HL OL PJu OL PJy = YJ|L  U0|J  + dE. (8.48)
Y0|J
holds in:
i) The deterministic case, provided the input is PE of order J + L + g − 1. The

error term, dE = 0, in this case.
ii) When J → ∞, and the input is PE of infinite order. The error term,
dE = 0, in this case.
iii) A white uk gives a consistent estimate of HLd irrespective of J > 0. How-
ever, OL PJu and OL PJy are not consistent estimates in this case. The first
mL × (L + g)r part of the error term, dE, is zero in this case.
Hence, under conditions i) and ii), OL PJu and OL PJy can be computed as in
(8.48). Then the states can be consistently estimated as
· ¸
£ ¤ U0|J
XJ = OL† OL PJu OL PJy , (8.49)
Y0|J
†
provided conditions i) and ii) are satisfied, and OL is known.
Proof 8.4 The PE conditions in the lemma· are due

¸ to the existence of the LS
UJ|L+g−1
solution, i.e., the concatenated matrix has to be of full row rank.
U0|J
From (8.32) we have that the error term in the LS problem is
 †  †
dE = (−OL PJe E0|J + HLs EJ|L )  U0|J  = −OL PJe E0|J  U0|J  (8.50)
.
Y0|J Y0|J
It is clear from (8.31) that the error term dE = 0 when J → ∞. This proves
condition i) in the lemma. Furthermore, the error term, dE = 0, in the deter-
ministic case because E0|J = 0 in this case. This proves condition ii). Analyzing
the error term, dE, for a white input shows that the error term is of the form
£ ¤†
dE = 0mL×(L+g)r dE2 dE3 , (8.51)
where the dE2 and dE3 are submatrices in dE different from zero. Note that
dE2 = 0 for strictly proper systems, g = 0, when uk is white. This proves
condition iii).
The states can then be computed by using (8.28) or (8.43), provided condi-
tions i) or ii) are satisfied. 2
One should note that in the N4SID algorithm the past horizon is put equal to
the future horizon (N4SID parameter i). In order for the above lemma to give
the same results as in the N4SID algorithm we have to put i = L + 1, J = L + 1
and g = 1, i.e so that J + L = 2L + 1 = 2i. Note that this last result does
not hold in general. It holds in the deterministic case or when J → ∞. The
extended observability matrix OL can be computed as presented in the next
section.
8.3.2 The extended observability matrix
An important first step in the SID algorithm is the identification of the system
order, n, and the extended observability matrix OL+1 . The reason for searching
for OL+1 is that we have to define A from the shift invariance property, Kung
(1978), or a similar method, e.g. as in Di Ruscio (1995). The key is to compute
a special projection matrix from the known data. This is done without using
the states. We will in this section show how this can be done for colored input
signals.
Lemma 8.3.5 (SID of the extended observability matrix)

The following projections are equivalent
 
UJ|L+g
ZJ|L+1 = (YJ|L+1 /  U0|J )U ⊥
J|L+g (8.52)
Y0|J
· ¸
⊥ U0|J ⊥
ZJ|L+1 = (YJ|L+1 UJ|L+g )/( UJ|L+g ) (8.53)
Y0|J
· ¸
U0|J ⊥
ZJ|L+1 = YJ|L+1 /( UJ|L+g ) (8.54)
Y0|J
Furthermore, ZJ|L+1 is related to the extended observability matrix OL+1 as
ZJ|L+1 = OL+1 XJa , (8.55)
where the “projected states” XJa can be expressed as

 
UJ|L+g
XJa = (XJ /  U0|J )U ⊥
J|L+g (8.56)
Y0|J
 
UJ|L+g
= (C̃Jd U0|J + C̃Js Y0|J − (A − KD)J X0 /  U0|J )U ⊥
J|L+g (8.57)
Y0|J
 ⊥
UJ|L+g
= (XJ − (A − KD)J X0  U0|J  )U ⊥
J|L+g (8.58)
Y0|J
 ⊥
UJ|L+g
= (XJ + (A − KD)J OJ† HJs E0|J  U0|J  )U ⊥
J|L+g (8.59)
Y0|J
Furthermore, the column space of ZJ|L+1 coincide with the column space of
OL+1 and n = rank(ZJ|L+1 ) if rank(XJa ) = n.
Proof 8.5 The proof is divided into two parts. In the first part (8.52) and
(8.55) with the alternative expressions in (8.56) to (8.58) are proved. In the
second part the equivalence with (8.52), (8.53) and (8.54) are proved.
Part 1 Projecting
 the row space of each term in (8.26) with L := L + 1 onto
UJ|L+g−1
the row space of  U0|J  gives
Y0|J
   
YJ|L+1 /  U0|J  = OL+1 XJ /  U0|J  + HL+1
d
UJ|L+g−1 + dE1(8.60)
,
Y0|J Y0|J
where we have used (8.13). Then, w.p.1
lim dE1 = 0, (8.61)

K→∞
where the error term, dE1 , is given by (8.39) with L := L + 1 . Removing the
effect of the future input matrix, UJ|L+g−1 , on (8.60) gives (8.52) and (8.55)
with XJa as in (8.56).
Furthermore,
 projecting
 the row space of each term in (8.25) onto the row
UJ|L+g−1
space of  U0|J  gives
Y0|J
   
XJ /  U0|J  = C̃Js Y0|J + C̃Jd U0|J + (A − KD)J X0 /  U0|J (8.62)
Y0|J Y0|J
From (8.25) we have that
C̃Js Y0|J + C̃Jd U0|J = XJ − (A − KD)J X0 . (8.63)
Combining (8.60), (8.62) and (8.63) gives (8.52) and (8.57)-(8.58).

Part 2 It is proved in Di Ruscio (1997) that
· ¸
UJ|L+g ⊥
ZJ|L+1 = YJ|L+1 / UJ|L+g
W
⊥
= YJ|L+1 UJ|L+g W T (W UJ|L+g
⊥
W T )−1 W UJ|L+g
⊥
, (8.64)
where
· ¸
U0|J
W = . (8.65)
Y0|J
⊥
Using that UJ|L+g ⊥
UJ|L+g ⊥
= UJ|L+g in (8.64) proves the equivalence between
(8.53), (8.54) and (8.52). 2
Lemma 8.3.6 (Consistency: Stochastic and deterministic systems)

Let J → ∞, then
⊥
ZJ|L+1 = OL+1 XJ UJ|L+g , (8.66)
where ZJ|L+1 is defined as in Lemma 8.3.5. A sufficient condition for consis-

tency, and that OL+1 is contained in the column space of ZJ|L+1 , is that there
are no pure state feedback.
Proof 8.6 Letting J → ∞ in (8.58) gives (8.66). This can also be proved
by using (8.44) in (8.56). Furthermore, if there are pure state feedback then
⊥
XJ UJ|L+g will lose rank below the normal rank which is n. 2
Lemma 8.3.7 (Deterministic systems)

For pure deterministic systems we have that (8.66) can be changed to
⊥ ⊥
ZJ|L+1 =: YJ|L+1 UJ|L+g = OL+1 XJ UJ|L+g . (8.67)
The extended observability matrix OL+1 can be computed from the column space
⊥
of YJ|L+1 UJ|L+g . Furthermore, one can let J = 0 in the deterministic case.
Proof 8.7 This follows from (8.66) and Lemma 8.3.5 by excluding the projec-
tion which removes the noise. 2
Lemma 8.3.8 (Stochastic systems)

For pure stochastic systems we have that (8.66) can be changed to
ZJ|L+1 =: YJ|L+1 /Y0|J = OL+1 XJ . (8.68)
The extended observability matrix OL+1 can be computed from the column space
of YJ|L+1 /Y0|J .
Proof 8.8 This follows from (8.66) and Lemma 8.3.5 by excluding the input
matrices from the equations and definitions. 2
8.3.3 Identification of the stochastic subsystem
We will in this section prove that, when the extended observability matrix is
known (from Section 8.3.2), the kalman filter gain matrix can be identified
directly from the data. Furthermore, it is proved that the noise innovations
process can be identified directly in a first step in the DSR subspace algorithm.
This result was first presented in Di Ruscio (1995) without proof. Some results
concerning this is also presented in Di Ruscio (2001) and (2003).
Lemma 8.3.9 (The innovations)

Define the following projection from the data
   ⊥
UJ|L+g UJ|L+g
s
ZJ|L+1 = YJ|L+1 − YJ|L+1 /  U0|J  = YJ|L+1  U0|J  . (8.69)
Y0|J Y0|J
Then w.p.1 as J → ∞
s s
ZJ|L+1 = HL+1 EJ|L+1 . (8.70)
Hence, the Toeplitz matrix HL+1s (with Markov matrices F , DC, . . ., DAL−1 C)
for the stochastic subsystem is in the column space of √1K ZJ|L+1
s since K1 EJ|L+1 EJ|L+1
T =
IL+1×L+1 .
Proof 8.9 The relationship between the future data matrices is given by
YJ|L = OL XJ + HLd UJ|L+g−1 + HLs EJ|L . (8.71)
 
UJ|L+g−1
Projecting the row space of each term in (8.71) onto the row space of  U0|J 
Y0|J
gives
   
YJ|L /  U0|J  = OL XJ /  U0|J  + HLd UJ|L+g−1 + dE1 , (8.72)
Y0|J Y0|J
then, w.p.1
lim dE1 = 0, (8.73)

K→∞
where dE1 is given in (8.39). Furthermore,

 
UJ|L+g−1
lim XJ /  U0|J  = XJ , (8.74)
J→∞
Y0|J
where we have used Equations (8.44) and (8.38). From (8.71), (8.72) and
(8.74) we have that
 
UJ|L+g−1
YJ|L − YJ|L /  U0|J  = HLs EJ|L . (8.75)
Y0|J
Putting L := L + 1 in (8.75) gives (8.69). 2
Note the following from Lemma 3.9. The innovations can be identified
directly as for g = 1
 
UJ|1
s
ZJ|1 = F EJ|1 = YJ|1 − YJ|1 /  U0|J  (8.76)
Y0|J
or for g = 0 when E = 0
· ¸
s U0|J
ZJ|1 = F EJ|1 = YJ|1 − YJ|1 / (8.77)
Y0|J
One should note that (8.77) holds for both open and closed loop systems. For
closed loop systems it make sense to only consider systems in which the direct
feed-through matrix, E, from the input uk to the output yk is zero. This result
will be used in order to construct a subspace algorithm whic gives consisten
results for close loop systems, se Section 5.
It is now possible to directly identify the matrices C and F in the innovations
model (8.1) and (8.2) and K and ∆ in the Kalman filter (8.16) and (8.17). Two
methods are presented in the following. The first one is a direct covariance
based method for computing K and ∆ and the second one is a more numerically
reliable “square root” based method for computing C and F .
Lemma 8.3.10 (correlation method for K and ∆) Define the projection

s
matrix ZJ|L+1 as in (8.69) and define the correlation matrix
1 s
∆L+1 = Z (Z s )T = HL+1
s s
(HL+1 )T . (8.78)
K J|L+1 J|L+1
s
where the Toepliz matrix HL+1 can be partitioned as
· ¸
s F 0m×Lm
HL+1 = , (8.79)
OL C HLs
where C = KF . Hence, (8.78) can be written as

· ¸ · ¸
∆11 ∆12 FFT F (OL C)T
∆L+1 = = . (8.80)
∆21 ∆22 OL CF T OL C(OL C)T + HLs (HLs )T
From this we have
E(εk εTk ) = F F T = ∆11 (8.81)
and
†
K = CF −1 = OL ∆21 ∆−1
11 . (8.82)
Lemma 8.3.11 (square-root method for C and F ) The LQ decomposition

of √1K ZJ|L+1
s gives
1 s
√ ZJ|L+1 = R33 Q3 . (8.83)
K
s
Then, the Toeplitz matrix HL+1 , and the Markov matrices F , DC, . . ., DAL−1 C,
are given directly by
s
HL+1 = R33 . (8.84)
F can be taken as one of the diagonal m × m block matrices in R33 , e.g. the
lower left sub-matrix, i.e.
F = R33 (Lm + 1 : (L + 1)m, Lm + 1 : (L + 1)m), (8.85)
or as the mean of all the diagionals. Furthermore,
OL C = R33 (m + 1 : (L + 1)m, 1 : m). (8.86)
The system matrix C is given by

† †
C = OL OL C = OL R33 (m + 1 : (L + 1)m, 1 : m). (8.87)
The Kalman filter gain matrix and the innovations covariance matrix are given
by
K = CF −1 (8.88)
† −1
= OL R33 (m + 1 : (L + 1)m, 1 : m)R33 (Lm + 1 : (L + 1)m, Lm + 1 : (L + 1)m)
T
∆ = FF . (8.89)
8.3.4 SID of the deterministic subsystem
The parameters in the B and E matrices can be computed from an optimal

least squares problem. A solution to this is given in the following
Lemma 8.3.12 (Extended state space model)

The states can be eliminated from the state space model (8.1) and (8.2) to yield
the so called Extended State Space Model (ESSM)
YJ+1|L = ÃL YJ|L + B̃L UJ|L+g + C̃L EJ|L+1 , (8.90)
where
def T
ÃL = OL A(OL OL )−1 OLT ∈ RLm×Lm , (8.91)
def £ ¤ £ ¤
B̃L = OL B HLd − ÃL HLd 0Lm×r ∈ RLm×(L+g)r , (8.92)
def £ ¤ £ ¤
C̃L = OL C HLs − ÃL HLs 0Lm×m ∈ RLm×(L+1)m . (8.93)
Proof 8.10 Putting J =: J + 1 in (8.37) and substituting XJ+1 = AXJ +

BUJ + CEJ into this equation gives
£ ¤ £ ¤
YJ+1|L = OL AXJ + OL B HLd UJ|L+g + OL C HLs EJ|L+1 . (8.94)
Equation (8.37) can be solved for XJ when (A, D) is observable, i.e.,
XJ = OL† (YJ|L − HLd UJ|L+g−1 − HLs EJ|L ), (8.95)

†
where OL = (OLT OL )−1 OL
T is the More-Penrose pseudo inverse of O . Substi-
L
tuting (8.95) into (8.94) gives (8.90)-(8.93). 2
Lemma 8.3.13 (Projection matrix for the deterministic subsystem)

Define the projection matrix
 
UJ|L+g
d
ZJ|L+1 = YJ|L+1 /  U0|J . (8.96)
Y0|J
d
This matrix can be partitioned into matrices ZJ+1|L d
and ZJ|L which satisfy the
deterministic model
d d
ZJ+1|L = ÃL ZJ|L + B̃L UJ|L+g . (8.97)
Proof 8.11 This follows from (8.90) and (8.13) and that
 
UJ|L+g
lim C̃L EJ|L+1 /  U0|J  = 0, (8.98)
K→∞
Y0|J
when the future inputs, UJ|L+g , the past data, U0|J and YJ|J , are all independent
of the future noise term, C̃L EJ|L+1 . 2
Lemma 8.3.14 (SID of the system matrices B and E)

Define from (8.97) the linear equation
Y = B̃L U, (8.99)
where
d d
Y = ZJ+1|L − ÃL ZJ|L , (8.100)
U = UJ|L+g . (8.101)
From (8.99) a least squares problem
min k Y − B̃L (B, E)U k2F (8.102)
B,E
for the unkown elements in B and E are defined as

· ¸
B
vec(Y) = X vec( ), (8.103)
E
which can be solved for the unknown parameters as
· ¸
B
vec( ) = X † vec(Y), (8.104)
E
where X † = (X T X )−1 X T . The matrix X is defined as follows
L+g
X
def
X = RiT ⊗ (Ei−1 − ÃL Ei ) ∈ RLmK×(n+gm)r , (8.105)
i=1
where ⊗ denotes the Kronecker tensor product. The matrices Ri and Ei are
defined in the following. The matrices
Ri ∈ Rr×K ∀ i = 1, · · · , L + g, (8.106)
are r-block rows in the U ∈ R(L+g)r×K matrix. I.e. extracted from
 
R1
 
U =  ... . (8.107)
RL+g
The matrices
Ei−1 ∈ RLm×(n+mg) ∀ i = 1, · · · , L + g, (8.108)
are defined as follows:
£ ¤
E0 = OL 0Lm×m , , EL+1 = 0Lm×(n+m) , (8.109)
     
0m×n Im×m 0m×n 0m×m 0m×n 0m×m
 D 0m×m   0m×n Im×m   0m×n 0m×m 
     
 DA 0m×m   D 0m×m   0m×n 0m×m 
E1 =   , E2 =   , EL =  (8.110)
.
 .. ..   .. ..   .. .. 
 . .   . .   . . 
DAL−2 0m×m DAL−3 0m×m 0m×n Im×m
The matrix Im×m denotes the m × m identity matrix.
8.4 Closed loop subspace identification 179
Proof 8.12 From (8.92) we have that B̃L is a linear function of B and E when
A and D are given. The matrix X in the LS problem (8.103) is defined from
B̃L U by using the identity
vec(AXB) = (B T ⊗ A)vec(X). (8.111)
Note that the number of columns in Y and U, which is defined in (8.100) and
(8.101), can be reduced to K =: (L + g)r by post-multiplying both (8.100)
T
and (8.101) with UJ|L+g . However, this does not affect the estimates of B and
E but will in general reduce the computation. Another variant, which should
be preferred for numerical reasons, is to define Y and U from the R matrix
provided by the RQ/LQ decomposition. This will also reduce the number of
columns to K =: (L + g)r in Lemma 8.3.14.
Note that only a matrix of size (n+gm)r ×(n+gm)r has to be inverted (i.e.,
the matrix X T X in the LS solution (8.104)) in order to solve for the unknown
parameters in B and E. This method combined with the LQ decomposition is
found to be very efficient. The method in Lemma 8.3.14 require only that the
input is exciting of order n + gm, and hence, independent of the user specified
parameters L and J. This is consistent with the lower bound on the order
of persistence of excitation for consistent estimation of an nth order possibly
proper (g = 1) linear system.
Note that the alternative strategy of first solving for B̃L in (8.97) and then
extracting B and E would require the inversion of an (L + g)r × (L + g)r matrix
T
UJ|L+g UJ|L+g . This matrix may be singular for colored input signals, and hence
is not preferred.
8.4 Closed loop subspace identification
We have in Section 8.3.2 shown that the extended observability matrix OL+1
can be estimated from the column space of the projection matrix ZJ|L+1 as
defined in (8.52). Let us look at the error term in this projection. We have
ZJ|L+1 = OL+1 XJa + dZ, (8.112)
The error term dZ is given by

· ¸
s UJ|L+g ⊥
dZ = HL+1 (EJ|L+1 / )UJ|L+g
W
s ⊥
= HL+1 EJ|L+1 UJ|L+g W T (W UJ|L+g
⊥
W T )−1 W UJ|L+g
⊥
,
s
≈ −HL+1 EJ|L+1 /UJ|L+g W T (W UJ|L+g
⊥
W T )−1 W UJ|L+g
⊥
, (8.113)
where W is defined in (8.65). We have in the last expression in (8.113) used

that EJ|L+1 W T /K ≈ 0 when the number of columns K tends to infinity. The
remaining projection in the error term is then EJ|L+1 /UJ|L+g . This term will
also be approximately zero for open loop and many closed loop problems, which
will be pointed out in the following. However, the term EJ|L+1 /UJ|L+g may be
non-zero and cause biased estimates for feedback systems in which the control
is directly proportional to the innovations noise. We will in the next sections
discuss how to overcome this problem. We will also stress that biased estimates
may be more reliable than estimates from an unbiased algorithm because the
variance may be small. This is illustrated in the section of examples.
8.4.1 Closed loop subspace identification: Using a filter in the

feeback loop!
Since we are allowing the inputs to be colored the question whether it is possible
with feedback in the inputs have to be pointed out. An external (dither) signal,
i.e., a reference signal, should be used to excite the system when collecting data
for subspace identification. The natural excitations from process disturbances
are often insufficient. The SID algorithm, e.g. DSR, works perfect for closed
loop deterministic systems. Our simulation results also shows that the results
may be good even for combined deterministic and stochastic systems, however,
the results depends on the dither signal or the signal to noise ratio. If the
signal to noise ratio is low then there may be a bias in the estimates. However,
the variance may be small. There may also exist an ”optimal” dither signal
which gives very accurate subspace estimates (small bias and small variance)
even for systems with a large signal to noise ratio. This will be illustrated in
Example 8.7.4. A white noise or random binary signal in the reference usually
gives very good closed loop identification results. Furthermore, a minimum of
measurements noise is, as always, to be preferred in order to obtain good closed
loop estimates.
It is believed that SID of systems with state feedback or feedback from
Kalman filter states would work well, provided an external dither signal is
introduced in the loop. The reason for this is that the states are ”noise-free”
and not correlated with the innovations noise. There are no problems by using
subspace identification methods in these cases.
The key is to make the term EJ|L+1 /UJ|L+g small, which is equivalent to
making the error term (8.39) small.
The (open loop) subspace identification methods may give biased estimates
for closed loop systems as in Figures 8.1 and 8.3 when the signal-to-noise ratio
is low. The reason for this is that the error term in (8.39) is not zero when
the future inputs, UJ|L+g−1 , are correlated with the future noise, EJ|L . Most of
our simulations shows that the bias in the DSR algorithm, due to noisy closed
loop data, is less than the bias in the other algorithms as N4SID, SUBID (Van
Overschee and De Moor (1996)), MOESP (Verhagen (1994)).
One of our solutions to the bias problem is to include a filter, e.g. a first
order low-pass filter, in the feedback path of the control system as illustrated
v_k w_k
r_k u_k y_k

h_c(z) h_p(z)
Figure 8.1: Standard feedback system with reference signal, rk , controller repre-
sented with uk = hc (z)(rk −yk ) and the system represented with hp (z). Possibly
process and measurements noise are represented with vk and wk , respectively.
in Figure 8.2. This filter will reduce or eliminate the feedback problem when
using subspace identification algorithms. Hence, the input and output data,
uk and yk , can be used directly to identify a state space model of the system.
The reason for this is that the input is no longer directly proportional to the
measurements noise wk in the output yk = Dxk + wk (or the innovations ek
in the output yk = Dxk + ek ). This solution to the feedback problem will be
illustrated in Example 8.7.3.
v_k w_k
r_k u_k y_k

h_c(z) h_p(z)
h_f(z)
Figure 8.2: Feedback system with a filter in the feeback loop in order to elim-
inate problems with feeback in the data when using subspace identification
algorithms. Here, uk is the control/input signal, yk the output signal and rk ,
is the reference signal. The controller is represented with uk = hc (z)(rk − ȳk )
where ȳk = hf (z)yk is the filtered output. The system is represented with the
transfer function hp (z). The controller is represented with the transfer function
hc (z) and the filter is represented with the transfer function hf (z). Possibly
Note that the control system in Figure 8.2 is a special case of the more
general control configuration in Figure 8.3.
Note that SID algorithms are very useful for model predictive control in
which the control input signal often is the reference signal to some local con-
troller for the process. Hence, the SID algorithm are used to identify the closed
loop model from the reference signal to the output. In this case we do not have
problems with feedback in the data.
v_k w_k
r0_k r_k u_k y_k

r_k
h_r(z) h_p(z)
h_c(z)
Figure 8.3: Alternative feedback system with reference signal, rk , controller rep-
resented with uk = rk −hc (z)yk and the system represented with hp (z). Possibly
8.4.2 Closed loop subspace identification: Using the controller
The knowledge of the controller or the reference signal can be used to obtain
consistent subspace identification algorithms for closed loop systems. Consider
a linear state space model of the controller in Figure 8.1, i.e.
xck+1 = Ac xck + Bc (rk − yk ), (8.114)
uk = Dc xck + Ec (rk − yk ), (8.115)
where Ac ∈ Rnc ×nc , Bc ∈ Rnc ×m , Dc ∈ Rr×nc and Ec ∈ Rr×m is the state space
model matrices of the controller and xck ∈ Rn is the controller state vector. Note
also that the results which follows also holds for the control strategy in Figure
8.3.
We will in the following assume that the controller is linear and that the
quadruple (Ac , Bc , Dc , Ec ) matrices in addition to the input and output data,
uk and yk , are known. One should also note that the linear controller matrices
can be exactly identified if rk , uk and yk are known. The problem of identi-
fying the controller is deterministic and one can in this case perfectly identify
(Ac , Bc , Dc , Ec ) by using rk − yk as inputs and uk as outputs by using e.g. the
DSR subspace algorithm (provided that noise-free rk ’s and uk ’s are given).
Consider the following matrix equation obtained from the state space model
(8.114) and (8.115) of the controller in Figure 8.1, i.e.
UJ|L = OLc XJc + HLc (RJ|L − YJ|L ). (8.116)
We will now adopt the idea in Van Overschee and De Moor (1997) and define a
signal/matrix MJ|L from (8.116) which is orthogonal to the future noise matrix
EJ|L as
def
MJ|L = UJ|L + HLc YJ|L = OL XJc + HLc RJ|L (8.117)
The signal/matrix 8.117 was introduced by Van Overschee and De Moor (1997)
in order to solve the bias problem in the subspace identification algorithms.
Note also that a similar signal can be defined from the control system in
Figure 8.3. The only difference is that the right hand side of (8.117) becomes
RJ|L − OL XJc in this case. This matrix is also orthogonal to the future noise
and the closed loop subspace identification algorithm which will be presented
in the following thus holds for both control systems as presented in Figures 8.1
and 8.3. The main point of introducing MJ|L is that
EJ|L /MJ|L = 0, (8.118)
which holds if EJ|L is orthogonal to both XJc and RJ|L .
8.4.3 Closed loop subspace identification: Indirect method
We will in this section derive a consistent version of the closed loop subspace
identification algorithm which is presented in Van Overschee and De Moor
(1997). We have the following consistent projection lemma for closed loop
subspace system identification
Lemma 8.4.1 (Closed loop SID)

Given the following closed loop projection
 
MJ|L+g
ZJ|L+1 = (YJ|L+1 /  U0|J )M ⊥
J|L+g (8.119)
Y0|J
Then, ZJ|L+1 is related to the extended observability matrix OL+1 as
ZJ|L+1 = T −1 OL+1 XJa , (8.120)
where T ∈ R(L+1)m×(L+1)m is a lower block Toepliz matrix given by

d c
T = I(L+1)m + HL+1 HL+g when g = 1 (8.121)
£ d c
¤
T = I(L+1)m + HL+1 HL+g 0(L+1)m×m when g = 0 (8.122)
Furthermore, the system order is given by
n = rank(ZJ|L+1 ). (8.123)
From the SVD
ZJ|L+1 = U1 S1 V1T + U2 S2 V2T , (8.124)
where the n large/dominant singular values are contained on the diagonal in S1

and the other zero/smaller singular values on the diagonal of S2 . Furthermore,
U1 ∈ R(L+1)m×n and U2 ∈ R(L+1)m×((L+1)m−n) .
From this we have the estimate
T −1 OL+1 = U1 . (8.125)
Furthermore, the ”autonomous” states are determined as XJa = S1 V1T .

Proof 8.13 A simple proof is given as follows. From (8.26) we have

d s
YJ|L+1 = OL+1 XJ + HL+1 UJ|L+g + HL+1 EJ|L+1 . (8.126)
d Hc
Adding HL+1 L+g YJ|L+g on both sides of (8.126) and using the definition in
(8.117) gives
d s
T YJ|L+1 = OL+1 XJ + HL+1 MJ|L+g + HL+1 EJ|L+1 . (8.127)
 
MJ|L+g
where T is as in (8.121) and (8.122). Since the matrix  U0|J  is uncor-
Y0|J
related (orthogonal) to the future noise matrix EJ|L+1 we have that
d
T ZJ|L+1 = OL+1 XJd + HL+1
d
MJ|L+g (8.128)
where
 
MJ|L+g
d
ZJ|L+1 =: YJ|L+1 /  U0|J , (8.129)
Y0|J
and
 
MJ|L+g
d
XJ|L+1 =: XJ /  U0|J , (8.130)
Y0|J
The lower triangular matrix T is non-singular if Im + EEc is non-singular.

Hence,
d
ZJ|L+1 = T −1 OL+1 XJd + T −1 HL+1
d
MJ|L+g (8.131)
⊥
Post-multiplication of (8.131) with MJ|L+g proves (8.119) -(8.121).
Furthermore we have that
U2T ZJ|L+1
d
= U2T T −1 HL+1
d
MJ|L+g (8.132)
where U2 is the left singular vectors from the SVD of ZJ|L+1 which is re-
lated to the ”zero/small” singular values. Equation (8.132) is obtained by
pre-multiplying (8.131) with U2T and using that T −1 OL+1 = U1 and U2T U1 = 0.
Equation (8.132) is a linear equation of the elements in the lower block
triangular Toepliz matrix T −1 HL+1
d . The solution to this problem is an im-
portant step in the closed loop subspace algorithm and therefore needs further
discussion. In order to do this we write (8.132) as
Y = UKMJ|L+g , (8.133)
where
Y =: U2T ZJ|L+1
d
(8.134)
U =: U2T (8.135)
def
K = T −1 HL+1
d
∈ R(L+1)m×(L+g)r (8.136)
The matrix K is a lower block triangular Toepliz matrix with m × r blocks Ki

∀ i = 1, . . . , L + 1. Hence, we can solve (8.132) (or equivalent (8.133) in a least
squares optimal sense for the parameters in K. An algorithm for doing this is
presented in Van Overschee and De Moor (1996b) Appendix C.
The problem of solving (8.132) is very similar to the least squares problem
in Section 8.3.4 for determining B and E for the open loop subspace problem.
c
Since HL+g d
is known the matrix HL+1 is simply obtained from K as (when
g = 1)
d c
HL+1 = K(I(L+1)r − HL+g K)−1 . (8.137)
Finally, the extended observability matrix of the system can be obtained as

(when g = 1)
d c
OL+1 = (I(L+1)m + HL+1 HL+g )U1 . (8.138)
The system matrices A and D are then computed from OL+1 . The system
d
matrices B and E are computed from HL+1 and OL .
The stochastic part of the system is determined very similar to the theory
in Section 8.3.3. The difference is that the projection matrix now is given by
 
MJ|L+g
s
ZJ|L+1 = YJ|L+1 − YJ|L+1 /  U0|J  = T −1 HL+1
s
EJ|L+1 (8.139)
Y0|J
d
The projection matrices ZJ|L+1 , ZJ|L+1 s
and ZJ|L+1 can effectively be computed
by the LQ/QR decomposition.
The above ideas is among other details used to construct a MATLAB func-
tion, dsr cl.m, for consistent closed loop subspace identification.
8.4.4 Closed loop subspace identification: Direct method
A drawback with the above algorithm is that T and HL+1d have to be identified
before the extended observability matrix OL+1 and the system matrices A and
D could be identified. We will in this section present a solution to the closed
loop subspace identification problem which is more consistent with the DSR
subspace algorithm in which OL+1 and A and D is identified directly in a first
step. We have the following lemma
Lemma 8.4.2 (Direct closed loop SID)

The extended observability matrix OL+1 is obtained from the following projection
d c d
ZJ|L+1 = ZJ|L+1 (MJ|L+g − HL+1 ZJ|L+1 )⊥
d c s
= ZJ|L+1 (UJ|L+g + HL+1 ZJ|L+1 )⊥ = OL+1 XJa (8.140)
d
where ZJ|L+1 s
and ZJ|L+1 are defined in (8.129) and (8.139), respectively.
d ) can be obtained from
Furthermore, B and E (or also HL+1
d
ZJ|L+1 = OL+1 XJd + HL+1
d c
(MJ|L+g − HL+1 d
ZJ|L+1 ) (8.141)
or from (as in the standard DSR algorithm) the equation

d c d
ZJ+1|L+1 − ÃL ZJ|L = B̃L (MJ|L+g − HL+1 ZJ|L+1 ) (8.142)
Proof 8.14 From (8.128) we have that

d c d
(I + HL+1 HL+g )ZJ|L+1 = OL+1 XJd + HL+1
d
MJ|L+g . (8.143)
Rearranging (8.143) gives

d
ZJ|L+1 = OL+1 XJd + HL+1
d c
(MJ|L+g − HL+g d
ZJ|L+1 ) (8.144)
The closed loop subspace algorithm which is presented in this section is very
similar to the open loop subspace identification algorithm which is presented in
Section 8.3. The only difference is that the projection matrices are modified to
incorporate the Markov parameters of the controller (the Toepliz matrix HL+g c .
8.5 A new subspace identification method for closed

and open loop systems
It was presented in Di Ruscio (1995) and proved in Di Ruscio (2001) that the
noise innovation process could be identified directly from the data in a first
step. This approach is valid for both open and closed loop systems.
We will in the following consider a closed loop system in which it make sense
to assume that E = 0. Putting g = 0 and letting J → ∞ in Equation ( 8.27)
gives
· ¸
£ d s
¤ U0|J
YJ|1 = D C̃J C̃J + F EJ|1 . (8.145)
Y0|J
Hence, the innovations can simply be identified as (for g = 0)

· ¸
s U0|J
ZJ|1 = F EJ|1 = YJ|1 − YJ|1 / (8.146)
Y0|J
8.5 A new subspace identification method for closed and open loop systems 187
This result is also obtained directly from Lemma 3.9, and Di Ruscio (2001). It
is clear that the above approach is valid for both open and closed loop systems
since the past data, U0|J and Y0|J , is uncorrelated with the future noise EJ|1 .
The innovations is known and the DSR algorithm is used directly with the
innovations as extra inputs in order to identify the ”deterministic” model
· ¸
£ ¤ uk
xk+1 = Axk + B K , (8.147)
εk
yk − εk = Dxk . (8.148)
We now simply solve a deterministic subspace identification problem in order

to obtain the system matrices A, B, K and D. Any efficient subspace method,
e.g. the DSR method as presented earlier in this paper, can be used.
The DSR algorithm for closed loop systems is sketched simply as follows:
1. From the known input and output data uk and yk ∀ k = 0, 1, . . . , N − 1

s = FE
and a large past horizon, J, the future innovations matrix ZJ|1 J|1 is
computed from (8.146). The projection involved can be computed directly
from the definition or preferably from a QR decomposition.
2. The innovations sequence εk ∀ k = J, J +1, . . . , N −1 in the Kalman filter

is given directly from the corresponding data in ZJ|1 s . From the known
innovations sequence we form the Hankel matrix EJ|L+1 .
3. Define the known input and output data sequence uk and yk ∀ k = J, J +

1, . . . , N − 1 and form the matrices UJ|L and YJ|L+1 .
4. The data from step 2 and 3 and a specified future horizon, L, are used
to identify the state space model by an efficient subspace identification
method. As a rule of thumb, chose L small such that 1 ≤ n ≤ Lm. The
theory in Sections 3.2 and 3.4 can be used directly by zeroing out the
matrices U0|J and Y0|J from the projections.
Hence, we simply obtain the model matrices in the Kalman filter from the
projection equations
· ¸⊥
UJ|L
ZJ|L+1 = YJ|L+1 ≈ OL+1 XJa (8.149)
EJ|L+1
which gives OL+1 , A and D and

· ¸
£ ¤ UJ|L
YJ+1|L = ÃL YJ|L + B̃L C̃L (8.150)
EJ|L+1
gives B and K according to the theory in Section 3.4. This algorithm is imple-
mented in the DSR e MATLAB function in the D-SR Toolbox for MATLAB.
8.6 Further remarks
8.6.1 Choice of algorithm parameters
There are two parameters in the algorithm, i.e., L and J. L is interpreted as

the identification-horizon used to predict the number of states. J is interpreted
as the horizon (into the past) used to define instruments from the data which
are used to remove noise. The system order, n, which is specified or identified,
is bounded by the user-specified parameter L, i.e. so that, 1 ≤ n ≤ mL where
m is the number of outputs. Hence, Lm singular values are computed by the
algorithm and the user may chose the system order, n, by inspection of the, n,
non-zero singular values.
A rule of thumb is that L should be chosen as small as possible if the inputs
are poorly exciting. The minimum identification-horizon, Lmin , so that the pair
n
(D, A) is observable and rank(OL ) = n is bounded by, d m e ≤ Lmin ≤ n −
min
rank(D)+1, where d·e is the ceiling function, i.e., rounding towards plus infinity.
If the outputs are independent, then, we suggest to use Lmin = n − m + 1 when
n ≥ m and Lmin = 1 when n = m. If the inputs are rich, e.g. white, then
this point is not critical. In practice, it is suggested that model validation
on independent data is taken into consideration when choosing the ”optimal”
settings for L.
The past horizon, J, may for combined deterministic and stochastic systems
and for pure stochastic systems usually be chosen as J = L + 1 or J = L. Note
that the estimates of C and the Kalman filter gain matrix K = CF −1 usually
becomes better when J increases. For pure deterministic systems we may chose
J = 1. The instruments Y0|J and U0|J can also be removed from the projections,
i.e., and putting J = 0, in this case.
The theory in this paper is the basis for the D-SR Toolbox for MATLAB
which are available upon request. The toolbox consists of MATLAB functions
for subspace system identification of both open and closed loop systems.
8.6.2 Choice of input signal
The subspace identification methods tends to be more sensitive to the input

signal compared to e.g, the Prediction Error Method (PEM). This means that
there may exist colored input signals which gives subspace estimates which are
as optimal (efficient and consistent) as the PEM estimates. On the other side
there may exist colored input signals where the subspace methods gives poorer
results compared to the PEM. An optimal experiment for the subspace methods
is in general not a white noise input, but rather a colored input signal where the
frequency spectrum is optimized to excite the parameters in the system as well
as possible. Our experience is also that an input signal which are minimizing
the condition number of the Hankal matrix U0|J or UJ|L+g , is usually not an
optimal input signal.
8.6.3 N4SID
The N4SID algorithms in Van Overschee and De Moor (1994) are using an
oblique-projection
⊥
Oi = Yi|i Ui|i WpT (Wp Ui|i
⊥
WpT )−1 Wp , (8.151)
· ¸
U0|i
Wp = , (8.152)
Y0|i
for the identification of the extended controllability matrix Oi , i.e., Oi is esti-
mated from the column space of Oi in (8.151), e.g. using the SVD. Comparing
(8.64) with (8.151) shows that the extra projection matrix Ui|i⊥ are missing on
the left hand side of (8.151). Hence, we conclude that in general

Oi 6= Oi Xia . (8.153)
The consequence of this is that the subspace identification theorems in Van
Overschee and De Moor (1994), (1996), (1997) which are using the oblique
projection, to our understanding, are wrong.
⊥ on the left hand side of (8.151) removes the
The extra projection matrix Ui|i
influence of the future inputs on the future outputs, and is necessary in order to
obtain accurate/consistent subspace estimates for colored input signals. Hence,
a consistent projection is therefore
⊥
Zi|i = Oi Ui|i = Oi Xia Ui|i
⊥
. (8.154)
Hence, the extra projection can not be considered as a weighting matrix but
as a projection matrix. One should note that the parameter i used by N4SID
is related to the parameter L in DSR as i = L + 1. Furthermore, in N4SID
the past horizon is put equal to the future horizon. Hence, this corresponds to
putting J = L + 1 = i in DSR.
8.7.1 Example 1
Given the system (8.1) and (8.2) with the following matrices and vectors
· ¸ · ¸ · ¸
0 1 0.25 0.5
A= , B= , C= , (8.155)
−0.7 1.5 0.625 0.5
£ ¤
D = 1 0 , E = 1, F = 1. (8.156)
The following colored input signals where used for identification
k
u1k = sin(k) + sin( ), (8.157)
2
k k
u2k = sin(k) + sin( ) + sin( ), (8.158)
2 3
u3k = withe noise with variance E(u2k ) = 1. (8.159)
The number of samples was N = 1000. The system was simulated 100 times,
each time with the same input but with a different noise realization ek . However,
with the same covariance E(e2k ) = 1. The DSR parameters where chosen as
L = 2, J = 3. The model structure parameter where g = 1. The poles of the
100 identified system matrices are illustrated in Figure 8.4, 8.5 and 8.6. From
this we conclude that the method presented in this paper is almost as efficient
as the PEM method, for the inputs which are considered. However, the N4SID
oblique-method gives an unacceptable bias in the pole estimates for input u1k ,
i.e., because the estimated poles are unstable. From Figure 8.5 we see that the
bias is eliminated but the results from N4SID are highly uncertain. However,
the results from N4SID are nearly the same as DSR for the white input signal
u3k . We also see that the variance of the subspace estimates may be smaller for
a colored input signal, Figure 8.5, than for the white noise input, Figure 8.6.
Eigenvalues: DSRc1 Eigenvalues: DSRc2
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.4 0.6 0.8 1 0.4 0.6 0.8 1
Eigenvalues: PEM Eigenvalues: N4SID

0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.4 0.6 0.8 1 0.4 0.6 0.8 1
Figure 8.4: Identified poles for a Monte carlo simulation. The exact pole is
marked with a cross. Input signal u1k where used.
8.7.2 Example 2
We consider the following system
xk+1 = Axk + Buk + Cvk , (8.160)

yk = Dxk + wk , (8.161)
where the system matrices are the same as in Example 8.7.1. The process noise,
vk , and
q the measurements noise, wk , q
are both white noise with standard devia-
√ √
tion E(vk ) = 0.02 = 0.1458 and E(wk2 ) = 0.002 = 0.0447, respectively.
2
The system is operating in closed loop. The input to the system is generated
by the following discrete time PI-controller
uk = Kp (rk − yk ) + zk , (8.162)

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9
marked with a cross.Input signal u2k where used.

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9
marked with a cross.Input signal u3k where used in this case.
where the controller state, zk , is defined by
Kp
zk+1 = zk + (rk − yk ). (8.163)
Ti
The proportional constant is Kp = 0.2, the integral time is Ti = 5 and the
reference, rk , is taken as the binary signal in Figure 8.7.
The number of samples was N = 1000. The system was simulated 100 times,
each time with the same reference, rk , but with a different noise realizations
vk and wk , but with the same variance. The DSR parameters where chosen as
L = J = 3 and the structure parameter where g = 0. The subspace algorithms
works perfect in the deterministic case. However, the algorithm gives a small
bias in the estimates in the case of noise. The bias is negligible for this example.
The pole estimates are presented in Figure 8.8.
Reference, rk, input, uk, and output, yk
0.5
rk
−0.5
−1
0 100 200 300 400 500 600 700 800 900 1000
0.5
uk
−0.5
0 100 200 300 400 500 600 700 800 900 1000
1
yk
−1
−2
0 100 200 300 400 500 600 700 800 900 1000
Discrete time
Figure 8.7: The reference signal, rk , the input, uk and the output yk for two
particular noise realizations vk and wk , used in Example 8.7.2.
8.7.3 Example 3
Consider the same closed loop example as in Example 2. We will in this exam-
ple illustrate the bias-problem when using subspace identification algorithms
directly from input and output data collected in closed loop (Figure 8.1). Fur-
thermore, we will illustrate that the feedback problem can be eliminated by
including a low-pass filter in the feedback as in Figure 8.2.
The process noise, vk , q
and the measurements noise, wk ,qare both white noise
√ √
with standard deviation E(vk2 ) = 0.05 = 0.2236 and E(wk2 ) = 0.01 =
0.1, respectively. The pole estimates after a Monte carlo simulation is presented
in Figure 8.9. We can clearly see a bias in the estimates from the (open loop)
subspace identification algorithms. The bias in the DSR estimates is smaller
Eigenvalues: DSR Eigenvalues: DSRc2

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9
Figure 8.8: The pole estimates from the closed loop data as described in Ex-
ample 8.7.2.
than the bias in the estimates from both SUBID (Van Overschee and De Moor
(1996)) and N4SID. This is also the conclusion from many other simulation
experiments.
Consider now the feedback system in Figure 8.2. We are using a PI-
controller as in Example 8.7.2 where the filtered output is used as input to
the controller. The controller equations are as follows.
uk = Kp (rk − ȳk ) + zk , (8.164)
where the controller state, zk , is defined by

Kp
zk+1 = zk + (rk − ȳk ). (8.165)
Ti
The filter is a 1. order low-pass filter of the form
ȳk+1 = ȳk + Kf (yk − ȳk )

= (1 − Kf )ȳk + Kf yk , (8.166)
with filter constant Kf = 0.1. The initial filter output is taken as ȳ0 = y0 . Pole
estimates after a Monte Carlo simulation is illustrated in Figure 8.10. We see
that the pole estimates now are consistent.
8.7.4 Example 4
We will in this example search for an optimal experiment in the reference.

Consider the reference rk = sin(ωk) for varying frequency ω. The following
investigation shows that the bias in the DSR pole estimates is a function of the
frequency and that the bias reach a minimum for a particular frequency.
Eigenvalues: DSR Eigenvalues: SUBID

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9
Figure 8.9: The pole estimates from the closed loop data as described in Exam-
ple 8.7.3. The control system is as in Figure 8.1 with the same reference signal
as in Figure 8.7.
Eigenvalues: DSR Eigenvalues: SUBID

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9
Figure 8.10: The pole estimates from the closed loop data as described in
Example 8.7.3 with a filter in the feedback. The control system is as in Figure
8.2.
The reference signal which gave the smallest bias in the pole estimates is
1
for this example found to be for ω = 1.526 , i.e.,
1
rk = sin( k). (8.167)
1.526
The pole estimates from an Monte carlo experiment is illustrated in Figure 8.11.
The results are very interesting because, as we see, the pole estimates from the
DSR subspace identification method is more reliable than the pole estimates
from the prediction error method PEM.
The process noise, vk ,q
and the measurements noise,q
wk , are both white noise
√ √
with standard deviation E(vk2 ) = 0.1 = 0.01 and E(wk2 ) = 0.1 = 0.01,
respectively. The DSR parameters is L = 5, g = 0 and J = 6.
Eigenvalues: DSR Eigenvalues: MOESP
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9
Eigenvalues: DSRcl Eigenvalues: PEM

0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0.6 0.7 0.8 0.9 0.6 0.7 0.8 0.9
Example 8.7.4 with reference signal rk = sin(k/1.526). The control system is
as in Figure 8.1.
8.7.5 Example 5
Consider the following system
xk+1 = 0.9xk + 0.5uk + 0.6ek (8.168)

yk = xk + ek (8.169)
with ek white noise with unit variance. The controller is taken as
uk = Kp (yk − rk ), (8.170)
with Kp = 0.6 and a reference signal
rk = sin(0.5k) + sin(k). (8.171)

A Monte carlo simulation study is performed with N = 2000 data points and
M = 100 different noise realizations. The results are illustrated in Figures 8.12
and 8.13. As we can see, both DSR e and PEM gives consisten results. It shows
also that the subspace method, DSR e, is as efficient that PEM is. The DSR
function results in a smaller bias than MOESP. The parameters L = 1 and
J = 6 was used for the DSR e function. Parameters L = 2 and J = 6 for DSR
and i = L + 1 = 3 for MOESP. The DSR e function is implemented along the
lines in Section 5.
The DSR algorithm gives usually less bias than MOESP for closed loop data,
se Figures 8.12 and 8.13. It is very interestin that the DSR e algorithm gives
parameter estimates which are as optimal as the corresponding PEM estimates.
Eigenvalues: DSR Eigenvalues: MOESP
0.5 0.5
0 0
−0.5 −0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Eigenvalues: DSRe Eigenvalues: PEM

0.5 0.5
0 0
−0.5 −0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Example 8.7.5 with reference signal rk = sin(k)+sin(0.5k). The control system
is as in Figure 8.1.
8.8 Conclusion
The extended observability matrix OL+1 can be computed directly from the
column space of a projection matrix ZJ|L+1 , which is defined in terms of the
known data. There are in general two projections involved in order to define
ZJ|L+1 . One projection is used to remove the effect of noise and one projection is
used to remove the effect of future inputs from the future outputs. A necessary
condition for a consistent estimate of OL+1 is that the number of columns K
in the data matrices tends to infinity.
The states are not needed in order to compute the extended observability
matrix and, hence, to identify the system dynamics, i.e., the number of states
n and the system matrices A and D.
An additional condition for a consistent state estimate is that the past
8.8 Conclusion 197
B parameter DSR B parameter MOESP

0.54 0.6
0.52 0.5
0.5 0.4
0.48 0.3
0.46 0.2
0.44 0.1
0.42 0
0.4 −0.1
0 20 40 60 80 100 0 20 40 60 80 100
B parameter DSRe B parameter PEM

0.65 0.65
0.6 0.6
0.55 0.55
0.5 0.5
0.45 0.45
0.4 0.4
0 20 40 60 80 100 0 20 40 60 80 100
Realizations Realizations
Figure 8.13: The estimates of the B parameter from the closed loop data as
described in Example 8.7.5 with reference signal rk = sin(k) + sin(0.5k). The
control system is as in Figure 8.1.
horizon J has to tend to infinity. Furthermore, for colored input signals both
the extended observability matrix OL and the lower block triangular Toepliz
matrix HLd has in general to be known in order to properly computing the states.
The stochastic part of the model, i.e., the Kalman filter gain matrix and
the innovations covariance matrix can be identified directly from the data, i.e.
s
from the projection matrix ZJ|L+1 , without solving any Riccati or Lyapunov
matrix equations.
The deterministic part of the model can be identified from an optimal least
d
squares problem defined from the projection matrix ZJ|L+1 , the extended ob-
servability matrix OL+1 , A and D.
d
Furthermore the necessary projections, ZJ|L+1 , ZJ|L+1 s
and ZJ|L+1 , which
are needed in order to compute a complete state space model realization for
the sixtuple matrices (A, B, D, E, C, F ) (and/or K and ∆ = E(εk εTk )), can
 
UJ|L+g
 U0|J 
be computed throug a numerically stable LQ decomposition of   Y0|J
.

YJ|L+1
d
However, it is in general faster to compute ZJ|L+1 , ZJ|L+1 s
and ZJ|L+1 directly
from the definitions. This means that the algorithm both can be implemented
as an correlation based method and a square root based method.
Finally, a method for subspace identification of closed loop systems which
gives unbiased estimates is presented. Simulation results shows that the esti-
mates are as efficient as those from the prediction error method, however, the
estimates are somewhat dependent of the parameters L and J.
References
Di Ruscio, D. (2003). Subspace system identification of the Kalman filter.

Modeling, Identification and Control. Vol24, No.3, pp. 125-157.
Di Ruscio, D. (2001). Subspace system identification of the Kalman filter gain.

The European Control Conference, ECC2001, Porto, Portogal 2001.

Di Ruscio, D. (1997b). On Subspace Identification of the Extended Observ-

ability Matrix. In the proceedings of the 36th Conference on Decision and
Control 1997, San Diego, California, December 6-14.
Di Ruscio, D. (1996). Combined Deterministic and Stochastic System Identifi-

cation and Realization: DSR-a subspace approach based on observations.
Modeling, Identification and Control, vol. 17, no.3.

stochastic systems: robust implementation. The Third European Control
Conference ECC95, September 5-8, Rome, Italy.
D-SR Toolbox for MATLAB. Copyright D. Di Ruscio. Available upon request,

[email protected].



matica, vol. 30, No. 1, pp.75-94.
Van Overschee, P. and B. De Moor (1996). Subspace Identification for Linear

Systems: theory-implementation-applications. Kluwer Academic Publish-
ers.
Van Overschee, P. and B. De Moor (1996b). Closed loop subspace system iden-
tification. Internal report. Katholieke Universiteit Leuven. Department
Elektrotechniek. ESAT-SISTA/TR 1996-521.
Van Overschee, P. and B. De Moor (1997). Closed loop subspace system

identification. In the proceedings of the 36th Conference on Decision and
Control 1997, San Diego, California, December 6-14.
8.8 Conclusion 199

matica, vol. 30, No. 1, pp. 61-74.

Chapter 9
Effect of Scaling and how to

Treat Trends
9.1 The data
We will consider some system input and output data vectors

¾
uk ∀ k = 1, · · · , N
Known system input and output data (9.1)
yk ∀ k = 1, · · · , N
We will assume that the data can be modeled by a linear discrete time
combined deterministic and stochastic state space model. The data does not
necessarily have to be stationary. Constant trends or drifts is represented with
non-stationary series which can be modeled by a state space model with poles
on the unit circle.
9.2 The data matrices
From the known input and output series (9.1) we define the data matrices as
follows
z  T }| {
y1
 yT 
 2 
Y =  .  ∈ RN ×m (9.2)
 .. 
T
yN

z  T }| {
u1
 uT 
 2 
U =  .  ∈ RN ×r (9.3)
 .. 
uTN
202 Effect of Scaling and how to Treat Trends
9.3 Scaling the data matrices
Consider the scaled system input and output data

¾
usk = Su uk ∀ k = 1, · · · , N
Known scaled input and output data. (9.4)
yks = Sy yk ∀ k = 1, · · · , N
The scaled input and output data vectors can be organized into data-matrices
similarly as in (9.2) and (9.3). However, the matrices which are identified should
be “unscaled” in order to satisfy the underlying (unscaled) system. A procedure
for incorporating scaling matrices is presented in Algorithm 9.3.1.
Algorithm 9.3.1 (Modeling procedure with scaled data matrices)

Step 1, scaling the data: From the data matrices, Equations (9.2) and (9.3),
and some scaling matrices Sy and Su , we define the scaled data matrices
Y s = Y SyT , (9.5)
s
U = U SuT . (9.6)
Step 2, identification: The problem of identifying a complete linear (usu-

ally) dynamic model for the process can be solved by the DSR function (see
D-SR Toolbox for Matlab).
£ ¤
A, B s , Ds , E s , C, F s = DSR(Y s , U s , L), (9.7)
where L is a positive scalar.
Step 3, unscaling: The matrices
B = B s Su , D = Sy−1 Ds , E = Sy−1 E s Su , F = Sy−1 F s , (9.8)
satisfy the SSM
xk+1 = Axk + Buk + Cesk , (9.9)
yk = Dxk + Euk + F esk , (9.10)
where esk has unit covariance.
Step 4, Kalman filter: The innovations form of the model (Kalman filter)
is given by
xk+1 = Axk + Buk + Kek , (9.11)
yk = Dxk + Euk + ek , (9.12)
where K is the Kalman filter gain
K = CF −1 . (9.13)
ek is the innovations with covariance matrix
E(ek eTk ) = F F T . (9.14)
4
Remark 9.1 One particular and common choice is to chose the scaling matri-
ces Sy and Su so that each column (variable) in the data-matrices Y s and U s
has unit variance.
9.3 Scaling the data matrices 203
9.3.1 Proof of Step 3 in Algorithm 9.3.1
The sixfold matrices (A, B s , C, Ds , E s , F s ) are the scaled state space model
matrices (in the state space model)
xk+1 = Axk + B s usk + Cesk , (9.15)

yks = D xk + s
E s usk + F s esk , (9.16)
where esk has unit covariance matrix, i.e., E(esk (esk )T ) = I. Substituting for yks
and usk given by (9.4) we get
B
z }| {
xk+1 = Axk + B s Su uk + Cesk , (9.17)
s s
Sy yk = D xk + E Su uk + F s esk . (9.18)
Assume that Sy is non-singular. Post multiply Equation (9.18) with Sy−1 we

get
yk = Sy−1 Ds xk + Sy−1 E s Su uk + Sy−1 F s esk . (9.19)

| {z } | {z } | {z }
D E F
Hence, we have
B = B s Su , D = Sy−1 Ds , E = Sy−1 E s Su , F = Sy−1 F s . (9.20)
To write the model on innovations (Kalman filter) form we define (from the
noise term in Equation (9.19))
ek = Sy−1 F s esk , (9.21)
where ek is the innovations with covariance matrix
E(ek eTk ) = Sy−1 F s (F s )T (Sy−1 )T , (9.22)
because esk has unit covariance.

From Equation (9.21) we get
esk = (F s )−1 Sy ek . (9.23)
Substituting for esk into equations (9.17) and (9.19) we get the innovations form
of the state space model
K
z }| {
xk+1 = Axk + Buk + C(F s )−1 Sy ek , (9.24)
yk = Dxk + Euk + ek , (9.25)
where the Kalman filter gain matrix K is overbraced. QED.

9.3.2 Numerical Examples
Example 9.1 (Monte Carlo simulation)

Monte Carlo simulations of a MIMO system with two inputs (r = 2), two
outputs (m = 2) and three states (n = 3) was worked out in order to analyze
the effect of scaling the input and output data on the pole estimates. Note that,
in this case, the minimal identification parameter L is Lmin = n − m + 1 = 2 if
the two output variables are independent. The output variables are independent
in this example. The DSR parameter L was varyed from L = 2 to L = 6.
Conclusions drawn when scaled data was used:
• Scaling the inputs gave no effect upon the pole estimates. The statistical
distribution of the pole estimates was unchanged.
• In general, output scaling of the type Y s = Y SyT where Sy is a gen-

eral non-singular matrix, can destroy the statistical distribution of the
pole estimates. However, this was only observed when L > Lmin where
Lmin = 2. The statistical distribution seams to depend on the chice of
scaling. A question is, does there exist an optimal scaling matrix Sy which
should be used when L > Lmin .
• Irrespective of scaling the output variables, the statistical distribution of

the pole estimates was unchanged when the minimal parameter L = Lmin =
2 was used.
• Irrespective of choice of L, the statistical distribution of the pole estimates

was unchanged when each column (variable) in Y was equally scaled, i.e.,
for scaling matrices Sy = diag(s, · · · , s) where s is a non-zero scalar.
Conclusions drawn when trended data was used:
• The system order increased by one when dtrended data (i.e., data with
constant trends removed) was used. A unit pole is included in the system
in order to handle the trends.
9.4 How to handle trends and drifts
Time series usually has a trend. Such trends can be nonzero constants or mean
values, low frequency noise or drifts. The trends are often aprori unknown and
time varying. Assume that the output yk from a system can be separated into
three parts. One deterministic part ykd which is driven from the known inputs
uk , one stochastic part yks driven from unknown inputs or disturbances and one
part yk0 which represents the trend.
yk = ykd + yks + yk0 . (9.26)

9.4 How to handle trends and drifts 205
In some applications we do not distinguish between the stochastic part yks and
the trend yk0 , i.e. we treat them simultaneously. The deterministic part, the
stochastic part and the trend may have common as well as separate and different
dynamics.
Constant non-zero trends and drifts can be described by a state space model
which has a pole on the unit circle. Low-frequency noise can be characterized
by a state space model which has a pole close to the unit circle. Such trends
can approximately be described by a random walk (drift)
x0k+1 = x0k + vk , (9.27)

yk0 = x0k , (9.28)
where vk is assumed to be a white and Gaussian distributed disturbance. This

is a non-stationary process because the system has a pole on the unit circle.
Trends can often be described by non-stationary processes which has a pole on
the unit circle.
Assume that the trend is purely time invariant (constant) and equal to y 0 ,
e.g., y 0 can be the sample mean of the series yk . The trend can then be described
by a 1st order state space model with a pole equal to one (integrator)
x0k+1 = x0k , x01 = y 0 , (9.29)

yk0 = x0k , (9.30)
where vk is assumed to be Gaussian distributed. The output yk0 from this system
is purely exited by the initial values y 0 . The initial values which are identified
are important because wrong initialization is not forgotten (because the system
is equal (or close) to an integrator).
It is not necessary to remove trends from the data when using DSR. Nonzero
constant trends or drifts are usually identified and represented with low fre-
quency dynamics, i.e. poles close to the unit circle.
One simple strategy is to first remove some constant trend from the data
and then identify the state space model matrices with initial values as follows,
£ ¤
A, B, D, E, C, F ,x0 = DSR(Y, U, L),
where Y and U are the data matrices adjusted for constant trends.
The constant trends which are removed can be the sample mean, the values
of the series at time zero k = 1 or the mean of the first, say, j samples of the
series. A simple strategy for identifying constant trends , which are frequently
used in practice, are given by
j
1X
u0 = uk , 1 ≤ j ≤ N, (9.31)
j
k=1
j
1X
y0 = yk , 1 ≤ j ≤ N. (9.32)
j
k=1
Nevertheless, even if this strategy is used, it is important to note that it can be

low-frequency dynamics in the estimated state space model. It is important to
note that this low-frequency dynamics often is incorporated in order to represent
trends or drifts (which are not properly identified) in the output time series, it
can be decoupled from the dynamics from the known inputs uk to the outputs
yk . In the following this is illustrated with some examples.
Example 9.2 (Using trended data)

Assume a purely steady state deterministic system given by
yk = Euk . (9.33)
where E ∈ Rm×r is the gain matrix of the system.

Assume that a sequence of N input and output samples are given. We can
then define the data matrices Y and U . (Note the relationship Y = U E T ).
Let us use a trended output data matrix for identification, i.e.,
Y s = Y − Y 0, (9.34)
where Y 0 is the trend (sample mean).

£ ¤
Using DSR for identification, i.e., A, B, D, E, C, F ,x0 = DSR(Y s , U, L),
we find the following 1st order state space model
xk+1 = xk , x1 = x0 , (9.35)
ŷks = Dxk + Euk . (9.36)
The estimated system has a unit pole. The reason for the unit pole is to handle
the trend. This means that both the trend Dxk = y 0 as well as the gain matrix
E are identified. We have that ŷks − Dxk = yk .
Example 9.3 (Non-stationary process)

Consider the system
yk = euk + yks , (9.37)
The output yk from this system consists of two parts. One purely steady state
deterministic part euk where e = −1 and one stochastic trend yks .
The trend is a drift which is described by
xk+1 = xk + vk , (9.38)
yks = xk + wk , (9.39)
where vk and wk is serially uncorrelated and Gaussian distributed with unit

variance. The system (9.38) is a so called “random walk” which has a pole on
the unit circle (a = 1). The process is therefore not stationary.
9.5 Trends and low frequency dynamics in the data 207
The input to the system is exactly known and described by the random walk
zk+1 = zk + µk , (9.40)
uk = zk , (9.41)
where µk is Gaussian distributed with unit variance and serially uncorrelated

with vk and wk .
A Monte Carlo simulation of the system was done. I.e., the system (9.37)-
(9.39) was simulated M = 500 times, each time with new realizations for vk and
wk but with the same input (9.41). The number of samples for each simulation
was N = 5000,
500 different 1st order state space models was estimated from the 500 se-
quences of input and output data (uk , yk , ∀, k = 1, · · · , N ).
The innovations model for the system is
xk+1 = axk + buk + cek , (9.42)

ŷk = xk + euk + ek . (9.43)
where a = 1, b = 0 and e = −1. The “exact” Kalman filter gain matrix is

c = 0.6180 and the variance of the innovations is E(e2k ) = 2.6180.
The mean and standard deviation of the 500 estimated set of parameters
(for the innovations model) are given by
¾
â = 0.9994 ± 0.0007, b̂ = 0.0000 ± 0.0009, ê = −1.0017 ± 0.0220,
(9.44)
ĉ = 0.6188 ± 0.0132, E(e2k ) = 2.6083 ± 0.0532.
Note that the estimated model parameters was scaled so that d = 1. The pa-
rameters was estimated by DSR with identification horizon L = 2 (used for
identification of the number of states) and past horizon J = 5 (used to define
instruments).
For comparison a purely steady state algorithm was used. The mean and
standard deviation of the gain estimates from PLS was
ª
ê = −0.9891 ± 1.2538 PLS with raw data (9.45)
ª
ê = −0.9758 ± 0.7409 PLS with centered data (9.46)
The mean of the parameters is not to bad. However, the standard deviation is
large which makes the estimates unreal-able.
9.5 Trends and low frequency dynamics in the data
There are basically two different approaches to deal with trends and slow dis-
turbances in the data:
The input time series: u_k

40
20
0
u_k
−20
−40
−60
−80
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
The output time series : y_k

100
50
y_k
−50
−100
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Time [samples]
Figure 9.1: This figure shows the input used in Example 9.3 as well as the
output (for one particular noise realization).
• Removing the disturbances by explicit pretreatment of the data.

Assume that input and output trends u0t and yt0 are given. A model is
then identified based on the input and output data which are adjusted for
the trends, i.e.
dyt = yt − yt0 , (9.47)

dut = ut − u0t . (9.48)
• Letting a noise model take care of the disturbances. In this case

the offset or trends are estimated as part of the modeling procedure. Some
additional states are included in order to identify the trends.
It is important to note that the trends must satisfy the model (or system).
This means that the trend either must be an equilibrium point or a steady state
point.
9.6 Time varying trends
Considder the data ut and yt wich satisfy the system
xt+1 = Axt + But + Cet , (9.49)

yt = Dxt + Eut + F et , (9.50)
and some trends u0t and yt0 which also satisfy the system
x0t+1 = Ax0t + Bu0t , (9.51)

yt0 = Dx0t + Eu0t . (9.52)
9.7 Constant trends 209
In this case the deviations du0t = ut − u0t and dyt0 = yt − yt0 will also satisfy the
system, i.e.
xt+1 − x0t+1 = A(xt − x0t ) + B(ut − u0t ) + Cet , (9.53)

yt − yt0 = A(xt − x0t ) + E(ut − u0t ) + F et , (9.54)
This means that the adjusted data dut and dyt satisfy the system model (A, B, D, E, CF, F )
with states dxt = xt − x0t .
9.7 Constant trends
Assume that the data are adjusted for some constant trends or working points
u0 and y 0 . It is important to note that such trends must satisfy the steady
state (static) relationship of the system, i.e.
y 0 = H d u0 (9.55)
where H d is the steady state deterministic gain matrix
H d = D(I − A)−1 + E. (9.56)
The data adjusted for the constant trends
dyt = yt − y 0 , (9.57)
0
dut = ut − u , (9.58)
will also satisfy the system given by (A, B, D, E, C, F ). This can relatively easy
be prooved in the frequency domain.
An approximation of these steady state trends wich are commonly used in
practice are the sample mean
N
0 1 X
y = yt , (9.59)
N
t=1
XN
1
u0 = ut . (9.60)
N
t=1
Another strategy which are found from practical experiences to work well in
most cases is to identify the working point from the mean of the first few
samples.
Chapter 10
Validation
10.1 Model validation and fit
In order to measure the quality of a model we have to define some criteria.

Consider a multivariable system with m output variables stacked on each other
in the output vector yt and r input variables stacked on each other in the input
vector ut .
We will in the following assume that we have a set of input and output time
series (or observations)
¾
ut
∀ t = 1, · · · , N (10.1)
yt
for model validation.
10.1.1 Criteria based on the simulated error
Assume that a deterministic model (i.e. some model matrices (A, B, D, E)

and a set of input and output validation data, as defined in (10.1), are given.
Simulation will then give us the simulated outputs
xdt+1 = Axdt + But (10.2)

ŷtd = Dxdt + Eut (10.3)
where the initial state vector xd1 is known/specified 1
The difference between each actual output and each simulated output can
be measured by the following criteria.
Definition 10.1 (Mean Square Error (MSE))

Define [yo ]t as (validation) output channel/number o and [ŷod ]t for the simulated
1
Note that the super-script d stands for deterministic and that the actual state vector xt
can be splitted into two parts, i.e., deterministic and stochastic states, satisfying xt = xdt + xst .
212 Validation
output number/channel o. The Mean Square Error for output number o (MSEo ),
i.e. the mean square error between the output number o and the simulated output
number o is defined as
N
1 X
MSEo = ([yo ]t − [ŷod ]t )2 ∀ o = 1, · · · , m (10.4)
N
t=1
where N is the number of observations (samples). This gives a vector of MSE

for the m output channels, i.e.
 
MSE1
 ..  m
MSE =  .  ∈ R (10.5)
MSEm
Definition 10.2 (Relative Mean Square Error (MSER))

Define [yo ]t as (validation) output channel/number o and [ŷod ]t for the simulated
output channel/number o. The Relative Mean Square Error for output number
o (MSERo ), i.e. the Mean Square Error between the output number o and the
simulated output number o Relative to the “energy” in the signal [yo ]t , is defined
as follows
1 PN d 2
N t=1 ([yo ]t − [ŷo ]t ) MSEo
MSERo = 1 PN
= 1 PN ∀ o = 1, · · · , m (10.6)
[y ]2 [y ]2
N t=1 o t N t=1 o t
where N is the number of observations (samples). This gives a vector of MSER

defined from the m output channels, i.e.
 
MSER1
 ..  m
MSER =  .  ∈ R (10.7)
MSERm
10.1.2 Criteria based on the prediction error
Assume that a combined deterministic and stochastic model (i.e. some model
matrices (A, B, D, E and Kalman filter gain matrix F ) and a set of input and
output validation data, as defined in (10.1), are given. Simulation will then
give us the optimal predictions as follows
xt+1 = Axt + But + K(yt − Dxt − Eut ) (10.8)

ŷt = Dxt + Eut (10.9)
where the initial state vector x1 is known/specified.

Chapter 11
Input experiment design
11.1 Experiment design for dynamic systems
Input experiment design for dynamic systems is discussed in , among others,

Goodwin and Payne (1972) and Ljung (1989). We will in the following focus on
Binary input Signals (BS) and Pseudo Random Binary input Signals (PRBS)
and on some simple criteria for measuring the quality of different experiment
design.
Assume that an input signal series
ut ∈ Rr ∀ 1 ≤ t ≤ N (11.1)
is given. From the input series (11.1) define the following input data matrix
with n + g block rows and K = N − n − k block columns.
z
 }| {
uk uk+1 uk+2 · · · uk+K−1
 uk+1 uk+2 uk+3 · · · uk+K 
 
def  .. .. .. . . .. 
Uk|n+g =  . . . . .  ∈ R(n+g)r×K
(11.2)
 
 uk+n+g−2 uk+n+g−1 uk+n+g · · · uk+n+K+g−3 
uk+n+g−1 uk+n+g uk+n+g+1 · · · uk+n+K+g−2
Definition 11.1 (Excitation condition)

The input signal ut defined by (11.1) is defined to be exciting of order n if and
only if the matrix Uk|n+g is non-singular, i.e.
rank(Uk|n+g ) = (n + g)r (11.3)
where g is a prescribed model structure parameter with values g = 0 if E = 0m×r

and g = 1 if E 6= 0m×r .
Definition 11.2 (Excitation condition)

The input signal ut defined by (11.1) is defined to be exciting of order n if and
214 Input experiment design
only if the matrix

1
Pn = U UT ∈ R(n+g)r×(n+g)r (11.4)
K k|n+g k|n+g
is non-singular.
A simple measure of the quality of the input signal (11.1) is the condition
number of Pn or Uk|n+g . An optimal input design is an input signal with mini-
mum condition number (cond(Un|n+g ) or cond(Pn )) subject to some constraints.
Following Ljung (1987) we have the following definition of a persistent ex-
iting input signal
Definition 11.3 (Persistent excitation)

The input signal ut defined by (11.1) is defined to be persistently exciting of
order n if and only if the matrix
lim Pn ∈ R(n+g)r×(n+g)r (11.5)

K→∞
is non-singular.
u1: prbs(1000,15,100) u2: prbs(1000,15,100)

1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
0 500 1000 0 500 1000
u3: prbs(1000,15,100) u4: prbs(1000,15,100)

1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
0 500 1000 0 500 1000
Figure 11.1: Illustration of a Pseudo Random Binary Signal (PRBS) experi-

ment design. The input signal series are designed by four separate calls to the
MATLAB function PRBS1.
Example 11.1 (optimal design)

Consider the problem of designing a binary single input experiment signal
ut ∀ 1 ≤ t ≤ N (11.6)
11.1 Experiment design for dynamic systems 215
u1: prbs(1000,40,200) u2: prbs(1000,40,200)

1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
0 500 1000 0 500 1000
u3: prbs(1000,40,200) u4: prbs(1000,40,200)

1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
0 500 1000 0 500 1000
Figure 11.2: Illustration of a Pseudo Random Binary Signal (PRBS) experi-

ment design. The input signal series are designed by four separate calls to the
MATLAB function PRBS1.
with the following data
N = 15 Total number of samples in ut

NT = 3 Total number of intervals in ut
Tmin = 2 Minimum sample interval
Tmax = 5 Minimum sample interval
umax
t =1 Maximum input amplitude
min
ut = −1 Minimum input amplitude
The NT = 3 sample intervals T1 , T2 and T3 satisfy the constraint
T1 + T2 + T3 = N. (11.7)
This gives only a number of two independent discrete parameters, say T1 and
T2 , because T3 can be computed as T3 = N − T1 − T2 .
It make sense to find those integer parameters T1 and T2 for which the con-
dition number of Uk|(n+g) is minimized. This is an integer variable optimization
problem.
The parameters T1 and T2 is bounded from up by Tmax = 5 and from below
by Tmin = 2. Hence, T1 and T2 can only have the discrete (integer) values 2, 3, 4
or 5. This gives a total number of MC = 4 · 4 = 16 different combinations.
The condition numbers of the matrices U1|n and Pn are presented in the fol-
lowing tables for system orders n = 5 and n = 6 and for all possibile experiment
design.
Design # T1 T2 T3 cond(P5 ) cond(U1|5 ) evaluation

1 2 2 11 74.729 8.645
2 2 3 10 42.695 6.534
3 2 4 9 22.123 4.704
4 2 5 8 22.123 4.704
5 3 2 10 44.488 6.670
6 3 3 9 34.487 5.873
7 3 4 8 15.156 3.893
8 3 5 7 15.156 3.893
9 4 2 9 22.407 4.734
10 4 3 8 12.159 3.487
11 4 4 7 10.472 3.236 minimum
12 4 5 6 10.472 3.236 minimum
13 5 2 8 22.407 4.734
14 5 3 7 12.159 3.487
15 5 4 6 10.472 3.236 minimum
16 5 5 5 10.472 3.236 minimum
Design # T1 T2 T3 cond(P6 ) cond(U1|6 ) evaluation
1 2 2 11 ∞ ∞
2 2 3 10 56.9469 7.5463
3 2 4 9 42.1114 6.4893
4 2 5 8 24.1949 4.9188
5 3 2 10 109.2796 10.4537
6 3 3 9 47.8914 6.9204
7 3 4 8 41.2023 6.4189
8 3 5 7 24.1949 4.9188
9 4 2 9 81.4547 9.0252
10 4 3 8 32.6288 5.7122
11 4 4 7 44.8400 6.6963
12 4 5 6 ∞ ∞
13 5 2 8 26.4365 5.1416
14 5 3 7 17.6257 4.1983 minimum
15 5 4 6 24.1949 4.9188
11.1 Experiment design for dynamic systems 217
First experiment design: cond(U)=3.24
3
2.5
2
u_k
1.5
1
0.5
0
2 4 6 8 10 12 14
Second experiment design: cond(U)=8.64
3
2.5
2
u_k
1.5
1
0.5
0
2 4 6 8 10 12 14
Discrete time [samples]
Figure 11.3: Illustration of two different binary input signal design. The design
in the upper part has a condition number cond(U1|5 ) = 3.236 and the design in
the lower part cond(U1|5 ) = 8.645. The input design in the upper part is better
conditioned for identification of a 5th (n = 5) order dynamic model compared
to the input design in the lower part. See example 11.1 for details.
Chapter 12
Special topics in system

identification
12.1 Optimal predictions
For some time series we observe that the system dynamics are close to an
integrator and in particular for stochastic single output systems we may find
that the A parameter is close to one. In such cases we also often find that the
optimal Kalman filter prediction, ȳk , is approximately equal to the previous
output yk−1 . The reason for this is presented in the following proposition
Proposition 12.1 (Prediction equal to previous output)

Consider a Kalman filter for a stochastic system xk+1 = Axk + vk and yk =
Dxk + wk , i.e.,
x̄k+1 = Ax̄k + K(yk − ȳk ), (12.1)
ȳk = Dx̄k . (12.2)
Assume that A = 1 (or A close to one) and DK = 1 (or DK close to one). In
this case we have that the optimal prediction, ȳk , at present time k is equal to
the previous output yk−1 , i.e.,
ȳk = yk−1 . (12.3)
Proof 12.1
From (12.1) and (12.2) with A = I we have that
∆x̄k = x̄k − x̄k−1 = K(yk−1 − ȳk−1 ), (12.4)
ȳk = yk−1 + D∆x̄k (12.5)
This gives
ȳk = ȳk−1 + DK(yk−1 − ȳk−1 ), (12.6)
which gives Equation (12.3) when DK = 1. QED
220 Special topics in system identification
The Kalman filter is constructed to minimize the covariance of the estima-

tion error, i.e. the parameters in the Kalman filter is found such that, e.g. the
trace of the matrix E((xk − x̄k )(xk − x̄k )T ) is minimized and usually also the
prediction error criterion is minimized. Furthermore, given different models one
should selected the best model from some validation criteria of the prediction
error criterion V = trace((E((yk − ȳk )(yk − ȳk )T )).
Chapter 13
On the Partial Least Squares

Algorithm
13.1 Notation, basic- and system-definitions
Define yk ∈ Rm as a vector of output variables at observation number k. The

output variables are some times referred to as response variables. Similarly a
vector xk ∈ Rr of input variables is defined. It is assumed that the vector of
output variables yk are linearly related to the vector of input variables xk as
follows
yk = B T xk + ek , (13.1)
where ek is a vector of white noise with covariance matrix E(ek eTk ) and k is the
number of observation.
With N observations k = 1, · · · , N we define an output (or response) data
matrix Y ∈ RN ×m and an input data matrix X ∈ RN ×r as follows
 T   T 
y1 x1
 ..   .. 
Y =  . , X =  . . (13.2)
T
yN xTN
The data matrices Y and X are assumed to be known.
The linear relationship (13.1) can be written as the following linear matrix
equation
Y = XB + E, (13.3)
where B is a matrix with regression coefficients. E ∈ RN ×m is in general an

unknown matrix of noise vectors, defined as follows
 T 
e1
 .. 
E =  . . (13.4)
T
eN
222 On the Partial Least Squares Algorithm
The linear relationship between the output (response) and the input data (or
regressors) is an important assumption and restriction for the PLS as well as
any LS algorithm to work.
We will in this work analyze systems with multiple output variables in the
data matrix Y . This is often referred to as multivariable (or multivariate)
systems.
If we only are interested in the matrix of regression coefficients B, then one
should note that it (for steady state systems) usually is sufficient to consider one
output at a time and only investigate single output systems. This means that
the multivariable LS problem can be solved from m single output LS problems,
i.e., each column in B is estimated from a separated LS problem.
Note also that (instead of modeling one output variable at a time) equation
(13.3) can be transformed to an equivalent model with one output in differ-
ent ways. Two possibile models with one output, which are equivalent to the
multivariable model (13.3) are preseted as follows
cs(Y ) = (Im ⊗ X)cs(B) + cs(E), (13.5)

cs(Y T) = (X ⊗ Im )cs(B T ) + cs(E T ), (13.6)
where cs(·) is the column string operator and ⊗ is the Kronecker product. cs(Y )
∈ RN m is a column vector constructed from Y by stacking each column in Y
on another. We also have (Im ⊗ X) ∈ RN m×rm and cs(B) ∈ Rrm .
Note that (13.6) can be constructed directly from (13.1) by first writing
(13.1) as
yk = (xTk ⊗ Im )cs(B T ) + ek (13.7)
and then combine all N equations (k = 1, · · · , N ) into a matrix equation of the

form (13.3).
However, for the sake of completeness we will in general consider multivari-
able (multiple output) systems of the form (13.3). One should also note that
the PLS algorithm can be used to compute projections. An example is the
problem of computing the projection of the row space of Y T onto the row space
of X T . In order to solve this problem we have to consider the general multiple
output model (13.3).
13.2 Partial Least Squares
13.2.1 Interpretation of the Partial Least Squares algorithm
Define a vector t from the column space of X as t = Xw where w is a weight

vector with constraint wT w = 1. The vector t is defined as the score vector for
X. Similarly, a score vector u for Y is defined from the column space of Y as
u = Y c where c is a weight vector with constraint cT c = 1.
13.2 Partial Least Squares 223
The PLS algorithm can be interpreted as the problem of computing weight

vectors w and c which maximize the correlation between the score vector t
and the score vector u. Hence, we have the problem of maximizing the function
f (w, c) = tT u with respect to w and c subject to the orthonormal constraints on
w and c. We can formulate the key part in the PLS algorithm as the following
optimization problem.
Problem 13.1 (PLS optimization problem) The key part in the PLS al-
gorithm can be formulated as an optimization problem for each component.
Maximize
f (w, c) = tT u,
with respect to w and c, where
t = Xw,
(13.8)
u = Y c,
subject to
wT w = 1,
cT c = 1.
4
The Lagrangian associated with this optimization problem is given by
L(w, c) = tT u + µ(1 − wT w) + σ(1 − cT c)

= wT X T Y c + µ(1 − wT w) + σ(1 − cT c), (13.9)
where µ and σ are Lagrange multipliers associated with the constraints. The
optimal solution is found by setting all possible derivatives of the Lagrangian
to zero. We have
∂L
∂c = Y T Xw − 2σc = 0, (13.10)
∂L T
∂w = X Y c − 2µw = 0, (13.11)
∂L
∂µ = 1 − wT w = 0, (13.12)
∂L T
∂σ = 1 − c c = 0. (13.13)
From Equations (13.10) and (13.11) we have
cT Y T Xw = 2σ, (13.14)
T T
w X Y c = 2µ. (13.15)
By transposing Equation (13.15) and comparing the result with Equation (13.14)
shows that the Lagrange multipliers are equal, i.e., µ = σ.
From this it is clear that w and c are given by the following SVD problem
cT Y T Xw = s̄, (13.16)
or equivalently (by transposing (13.16))
wT X T Y c = s̄, (13.17)
where s̄ = 2σ = 2µ is the singular value associated with X T Y or Y T X. The

weight vectors w and c can be taken as singular vectors of X T Y or Y T X.
Furthermore, one should take w and c as the singular vectors associated with
the maximal singular value of X T Y (or Y T X). The reason for this is that the
objective
f (w, c) = tT u = wT X T Y c = s̄ (13.18)
is maximized when s̄ is the maximum singular value of X T Y . Note also that

if r > m then we should compute the economy size SVD of X T Y otherwise we
should work with Y T X. Further details of this SVD problem will be presented
and commented on in the next section.
From Equation (13.10) we have c = Y T Xw/(2σ). Substituting this into
Equation (13.11) gives an eigenvector problem for w. Similarly, from Equa-
tion (13.11) we have w = X T Y c/(2µ). Substituting this into Equation (13.10)
gives an eigenvector problem for c. Hence, we have the following eigenvec-
tor/eigenvalue problems associated with the PLS algorithm
X T Y Y T Xw = λw, with wT w = 1, (13.19)

T T T
Y XX Y c = λc, with c c = 1, (13.20)
where the eigenvalue λ is related to the Lagrangian multipliers in (13.9) as

λ = 4σµ. Note also that λ is related to the maximum singular value s̄ of
(13.16) or (13.17) as λ = s̄2 .
Computing the orthonormal weight vectors w and c
From Section 13.2.1 it is clear that both weight vectors w and c can be computed
as the left and right singular vectors of the matrix X T Y , respectively. We will
in the following assume that r > m. Otherwise we should work with Y T X. In
this case c is the left singular vector and w is the right singular vector associated
with Y T X, respectively.
The key step in the PLS algorithm is the following singular value decompo-
sition
X T Y = USV T , (13.21)
where U ∈ Rr×r is a matrix of left orthonormal singular vectors, V ∈ Rm×m is a

matrix of right orthonormal singular vectors and S ∈ Rr×m is a diagonal matrix
of singular values. Note that a reduced singular value decomposition can be
computed when r > m. In this case U ∈ Rr×m , S ∈ Rm×m and V ∈ Rm×m .
Define the maximum (largest) singular value of X T Y as s̄. Hence, s̄ is equal
to the upper left (first diagonal) element of the matrix S of singular values. The
weight vector w is given by the left singular vector associated with the largest
singular value. Similarly, the weight vector c is given by the right singular
vector associated with the largest singular value. We have
w = U(:, 1), (13.22)

c = V(:, 1), (13.23)
s̄ = S(1, 1). (13.24)
The weight vectors are orthonormal, i.e., they satisfy wT w = 1 and cT c = 1.

The reason for this is that the columns of the SVD matrices U and V are
orthonormal.
Equation (13.21) is central in the PLS algorithm. This can be shown as
follows. From the SVD in (13.21) it is clear that the eigenvector problems in
(13.19) and (13.20) can be written as
X T Y Y T XU = USS T , (13.25)
T T T
Y XX Y V = VS S, (13.26)
where SS T ∈ Rr×r is a diagonal matrix with the eigenvalues of X T Y Y T X

along the diagonal and S T S ∈ Rm×m is a diagonal matrix with the eigenvalues
of Y T XX T Y along the diagonal.
After w and c are computed, the score vectors can be computed as follows
t = Xw, (13.27)
u = Y c. (13.28)
The maximum correlation between the score vectors t and u is
max f (w, c) = tT u = wT X T Y c = wT USV T c = s̄, (13.29)

w,c
where s̄ is the maximum singular value of X T Y , i.e. equal to the upper left
diagonal element in S. This is the solution to the PLS optimization problem
(13.8).
Remark 13.1 For one component, the maximum correlation between the t
score vector and the u score vector is equal to the maximum singular value
s̄ of X T Y , i.e., tT u = s̄.
We have given an interpretation of the PLS algorithm for one component

(also for the first component in the general PLS algorithm). The following
algorithm illustrates the central part of the general PLS algorithm.
Algorithm 13.2.1 (Central part of the PLS algorithm)
X1 = X
f or i = 1, a Loop for all components 1 ≤ a ≤ r
XiT Y = USV T The Singular Value Decomposition
wi = U(:, 1) The weight vector for X
ci = V(:, 1) The weight vector for Y
s̄i = S(1, 1) The largest singular value of X T Y
ti = Xi wi The score vector for X
ti tT
Xi+1 = (IN − tT
i
)Xi Rank one reduction of X
i ti
end
This results in a (score vector) matrix T ∈ RN ×a with orthogonal columns,

weight matrices C ∈ Rm×a and W ∈ Rr×a with orthonormal columns and a
diagonal matrix S ∈ Ra×a of maximum singular values s̄i of XiT Y . The matrices
are defined as follows
£ ¤ £ ¤ £ ¤
T = t1 · · · ta , C= c1 · · · ca , W = w1 · · · wa , (13.30)
 
s̄1 · · · 0
 .. .. ..  .
S= . . .  (13.31)
0 s̄a
The details in the algorithm will be discussed in the next sections.
Remark 13.2 For each new component a rank one reduced matrix Xi is used
in the SVD problem. Similarly it is common practice to use a rank one reduced
t tT
matrix Yi+1 = (IN − tiT ti )Yi . In this case the SVD of the matrix XiT Yi is
i i
computed at each iteration. However, rank one reduction of Y is not necessary.
Further details will be discussed in the next Section 13.2.1.
Definition 13.1 Define a score matrix U of score vectors for Y computed from
a rank one reduced matrix Yi as follows
£ ¤
U= u1 · · · ua ∈ RN ×a , (13.32)
where the score vectors are given by
ui = Yi ci ∈ RN . (13.33)
4
Remark 13.3 The maximum correlation matrix between the score vector ma-
trix T and the score vector matrix U is equal to a lower triangular matrix L,
i.e.,
T T U = L ∈ Ra×a , (13.34)
if the score matrix U is computed as in definition 13.1, i.e., from a rank one
reduced (undeflated) matrix Yi .
The diagonal element number i of L is equal to the maximum singular value
s̄i of XiT Y , i.e., lii = s̄i , where lii is diagonal element number i in matrix L.
This means that
tr(T T U ) = tr(S), (13.35)
where S is defined by (13.31).
Proof of remark 13.3
The statement that L given by (13.34) is a lower triangular matrix can be

proved as follows. The upper triangular elements of T T U = L are given by
ti tTi
li,i+1 = tTi ui+1 = tTi Yi+1 ci+1 = tTi (I − )Yi ci+1 = (tTi − tTi )Yi ci+1 = 0.
(13.36)
tTi ti
QED.
Definition 13.2 Define a score matrix Ũ for Y computed from an undeflated

Y matrix as follows
Ũ = Y C ∈ RN ×a . (13.37)
Each score vector is defined as
ũi = Y ci ∈ RN . (13.38)
4
Remark 13.4 The maximum correlation matrix between the score vector ma-
trix T and the score matrix Ũ is in general equal to a full matrix M defined
as
T T Ũ = M ∈ Ra×a , (13.39)
when matrix Ũ is computed without Y deflation, i.e., as in Definition 13.2.
The diagonal element number i of M is equal to the maximum singular value
s̄i of XiT Y , i.e., mii = s̄i , where ii is diagonal element number i in matrix M .
Hence
tr(T T Ũ ) = tr(S), (13.40)
where S is defined by (13.31).
4
Gram-Schmidt orthogonalization
We have shown how the weight vectors w and c, and the score vectors t and u
are computed (for one component) from the SVD of X T Y . See Section 13.2.1.
The next step in the PLS algorithm is a rank one reduction of X and Y .
(However, one should note that rank one reduction of Y is not necessary). This
step is very central in the PLS algorithm and should therefore be discussed in
detail. Define the following (rank one reduced) matrices from X, Y and the
t-score vector.
ti tTi
Xi+1 = (IN − )Xi , (13.41)
tTi ti
ti tTi
Yi+1 = (IN − )Yi , (13.42)
tTi ti
where X1 = X and Y1 = Y . In the PLS algorithm, the weight vectors w and c
and the score vectors t and u, for the next component, are computed based on
the rank one reduced matrices Xi+1 and Yi+1 .
The score vector ti+1 (for the next component) is computed based on Xi+1
and Yi+1 . This means that a new SVD of Xi+1 T Y
i+1 is computed. The weight
vectors wi+1 , ci+1 and the score vectors ti+1 , ui+1 can be computed as shown
in Section 13.2.1.
ti+1 is orthogonal to ti , i.e., tTi+1 ti = 0. The rank one reduction, Equation
(13.41) and (13.42), is similar to the Gram-Schmidt orthogonalization process.
The rank one reduction process can be written in terms of so called loading
vectors. The loading vector pi for Xi is defined as
tTi Xi
pTi = . (13.43)
tTi ti
The rank one reduction of X, Equation (13.41), can then be written as
Xi+1 = Xi − ti pTi . (13.44)
Using that ti = Xi wi we have the following alternative to (13.43),

wiT XiT Xi
pTi = . (13.45)
wiT XiT Xi wi
The rank one reduction of Xi , Equation (13.41), can then be written as
Xi+1 = Xi (I − wi pTi ). (13.46)
Equations (13.41), (13.44) and (13.46) can be used to formulate different PLS
implementations.
The loading vector c̄i for Y is defined as
tTi Yi
c̄Ti = . (13.47)
tTi ti
The rank one reduction of Yi , Equation (13.42), can then be written as
Yi+1 = Yi − ti c̄Ti . (13.48)
By using ti = Xi wi we can find alternative expressions for (13.47) and (13.48).

We will know show that the loading vector c̄i is related to the weight vector
ci , i.e. related to the SVD of XiT Yi . For convenience in the final PLS algorithm
we define a scaled weight (loading) vector c̄i for Yi as follows
s̄i
c̄i = ci , (13.49)
tTi ti
where s̄i is the largest singular value of XiT Yi , i.e. the upper left (diagonal)
element in the matrix S of singular values.
Proof of equation (13.49)
For the sake of simplicity the subscript i is neglected in the following proof.
The loading vector can be written as
Y Tt Y T Xw
c̄ = = , (13.50)
tT t tT t
where we have used that t = Xw. Using that Y T X = VS T U T we have
Y T Xw 1
c̄ = T
= T VS T U T w. (13.51)
t t t t
The matrix U with left singular vectors for X T Y has orthonormal columns.
The weight vector w is equal to the first column in U. Hence we have
 T   
w 1
 uT   0 
 2   
UT w =  .  w =  .  . (13.52)
.
 .   .. 
uTr 0
Using (13.52) we have that

      
s1 0 · · · 0 1 s1 s̄
 0 s2 · · · 0  0   0   0 
      
ST U T w =  . .. .. ..  .. = .. = .. . (13.53)
 .. . . .  .   .   . 
0 0 ··· sm 0 0 0
By using (13.53) we have that

 
s̄
£ ¤
 0 

VS T U T w = c v2 · · · vm  ..  = s̄c. (13.54)
 . 
0
Finally, we have that the loading vector c̄ can be written as

1 s̄
c̄ = VS T U T w = c. (13.55)
tT t tT t
QED.
Remark 13.5 Note that it is not necessary to compute the rank one reduction
of both matrices X and Y as shown in Equation (13.41) and (13.42), in order to
compute the SVD of XiT Yi for each component. This can be shown as follows
T ti tTi T ti tTi
Xi+1 Yi+1 = XiT (I − ) (I − )Yi . (13.56)
tTi ti tTi ti
The matrix I − ti tTi /(tTi ti ) is symmetric. We also have that (I − ti tTi /(tTi ti ))n =
I − ti tTi /(tTi ti ) for positive integers n. Hence
T ti tTi
Xi+1 Yi+1 = XiT (I − )Yi . (13.57)
tTi ti
This means that
T T
Xi+1 Yi+1 = Xi+1 Y = X T Yi+1 . (13.58)
Algorithm 13.2.2 (PLS algorithm which does not reduce Y)

In the following algorithm Y is not reduced at each iteration. The price to pay
is that the u-score vectors are not computed.
X1 = X
XiT Y = USV T The Singular Value Decomposition
wi = U(:, 1) The weight vector for X
ci = V(:, 1) The weight vector for Y
XiT ti
pi = Loading vector for X. (13.59)
tT
i ti
s1
di = T
ti ti
Scaling coefficient
c̄i = di ci Scaled weight (loading) vector for Y
ti tT
Xi+1 = (I − tT
i
i ti
end £ ¤ £ ¤
W =£ w1 · · · wa¤ , C̄ =
£ c̄1 · · · c̄¤a ,
P = p1 · · · pa , T = t1 · · · ta .
This algorithm decomposes X into X = T P T +G, where G is the residual matrix

which is equal to E = Xa+1 . Y is decomposed into Y = T C̄ T + F . The residual
can be computed as F = Y − T C̄ T . The matrix of regression coefficients can be
computed as BPLS = W (P T W )−1 C̄ T . Note also that the rank one reduction
of X can be computed as Xi+1 = Xi − ti pTi or Xi+1 = Xi (I − wi pTi ).
4
Monte carlo simulation with PLS used on dynamical systems indicates that
additional rank reduction of Y gives improved numerical precision. However,
the effect of removing the rank reduction step on Y with respect to numerical
precision should be further investigated.
The partial regression
We will here study the rank one reduction of X and Y which is carried out for
each component in the PLS algorithm. This will in turn give us an expression
for the PLS estimate of the regression coefficients.
Consider the rank one reduction (deflation) process of X and Y
tT X
Xi+1 = X − tpT , pT = , (13.60)
tT t
and
Yi+1 = Y − tc̄T , c̄T = dcT , (13.61)
where we have defined the scalar d as

s̄
d= . (13.62)
tT t
Using that t = Xw we have
Xi+1 = X(I − wpT ), (13.63)

T T
Yi+1 = Y − Xwc̄ = Y − Xwdc . (13.64)
Comparing (13.64) with the assumed linear relationship Y = XB + E between

X and Y , we have that the matrix of regression coefficients (for one component
only) is given by
s̄
B1 = wc̄T = wdcT = wcT . (13.65)
tT t
The prediction of Y (for one component only) is then given by
Ŷ1 = XB1 = Xwc̄T . (13.66)
Hence, if the rank one reduction of Y is carried out at each step, it can be
interpreted as the residual
Yi+1 = Y − Ŷ1 . (13.67)
Example 13.1 (PLS with two components)

Assume two components a = 2. The residual can be written as
Y3 = Y2 − t2 c̄T2 = Y2 − X2 w2 c̄T2 = Y2 − X2 B2 , (13.68)

where we have defined B2 = w2 c̄T2 . Substituting for Y2 = Y − XB1 and X2 =

X(I − w1 pT1 ) gives
Y3 = Y − XB1 − X(I − w1 pT1 )B2 = Y − X(B1 + (I − w1 pT1 )B2 ). (13.69)
The PLS estimate of the matrix of regression coefficients B can be written as
BPLS = B1 + (I − w1 pT1 )B2 = w1 c̄T1 + (I − w1 pT1 )w2 c̄T2 , (13.70)
where we have used that B1 = w1 c̄T1 and B2 = w2 c̄T2 . This can be written in
matrix form as follows
(P T W )−1
zW }| { z C̄}|T
z£ }| {¤ · ¸ {
1 −pT1 w2 £ ¤T
BPLS = w1 w2 c̄1 c̄2 . (13.71)
0 1
It can in general be shown that BPLS = W (P T W )−1 C̄ T as indicated. The

loading matrix C̄ can be written as
S̄
C z" }| #{
z£ }| {¤ s̄1
0
tT
1 t1
C̄ = c1 c2 s̄2 . (13.72)
0 tT
2 t2
The matrix of regression coefficients can then be written as
(P T W )−1 S̄
z }| z }| #{ z CT
z£ }|
W
{¤ · ¸{ " s̄1
0 }| {
T
1 −p1 w2 tT £ ¤T
BPLS = w1 w2 1 t1 c1 c2 . (13.73)
s̄2
0 1 0 tT
2 t2
It can in general be shown that BPLS = W (P T W )−1 S̄C T , where S̄ is a diagonal

matrix with s̄i /(tTi ti ) on the diagonal.
4
The PLS algorithm step for step
The following algorithm is presented in order to illustrate the computations

involved in the PLS algorithm. Refined PLS algorithms will be presented later.
Algorithm 13.2.3 (PLS computations)
X1 = X, Y1 = X
XiT Yi = U SV T The Singular Value Decomposition
wi = U (:, 1) The weight vector for X
ci = V (:, 1) The weight vector for Y
XiT ti
pi = tT
Loading vector for X.
i ti
s̄i
di = tT
Scaling coefficient
i ti
ūi = Yi ci /di Scaled score vector u = Y c for Y
c̄i = di ci Scaled weight (loading) vector for Y
ti tT
Xi+1 = (I − tT
i
i ti
T
ti ti
Yi+1 = (I − tT
)Yi Rank one reduction of Y
i ti
end
Note that the rank one reduction of Y only is needed when the u-score vector
is needed.
13.2.2 The Partial Least Squares decomposition
The rank one reduction step Xi+1 = Xi −ti pTi ∀ i = 1, · · · , a which is carried out
for each component in the iterative PLS algorithm, can be written as follows
Xa+1 = X − t1 pT1 − · · · ta pTa . (13.74)
The residual matrix is here defined as G = Xa+1 . A similar rank one reduction
for Y , i.e. Yi+1 = Yi − ti c̄Ti ∀ i = 1, · · · , a can be carried out. This results in
Ya+1 = Y − t1 c̄T1 − · · · ta c̄Ta . (13.75)
The residual is defined as F = Ya+1 . Note however that the reduction of Y is

not necessary in order to compute the decomposition.
Equations (13.74) and (13.75) can be written in matrix form. Hence, the
PLS algorithm decomposes X and Y as follows
X = T P T + G, (13.76)
T
Y = T C̄ + F, (13.77)
where G ∈ RN ×r and F ∈ RN ×m are residual matrices. T ∈ RN ×a is a matrix

of score vectors associated with the X matrix. Note that T is an orthogonal
matrix, i.e. T T T is a diagonal matrix. P ∈ Rr×a is a matrix of loading vectors
associated with X. C̄ ∈ Rm×a is a matrix of scaled weight vectors (or loading

vectors) associated with Y .
The loading matrix (scaled weight matrix) C̄ for Y can be decomposed into
C̄ = CSΛ−1 T −1
t = Y T Λt , (13.78)
where C is a matrix with orthonormal columns ci , i.e., cTi ci = 1. See the

Appendix, Equation (13.185) for further details. ci can be taken from the right
singular vectors of XiT Y . S is a diagonal matrix with the maximum singular
values s̄i of XiT Y on the diagonal. Λt is a diagonal matrix with tTi ti on the
diagonal. Note that ti = Xi wi and that
Λt = T T T. (13.79)
The loading matrix P for X can then be decomposed into
P = Pt Λ−1
t , (13.80)
where Pt is the unscaled loading matrix with columns formed from the unscaled
loading vectors, i.e.,
pt,i = XiT ti . (13.81)
We will in the next section use the above relations to formulate a PLS decom-
position with an orthonormal score vector matrix for X.
By combining Equations (13.80) and (13.81) we find that the P loading ma-
trix for X, in general can be expressed as P = X T T (T T T )−1 . This expression
for P can be proved by putting the gradient matrix of k G k2F with respect to
tr(P ) equal to zero. These results are presented in the following lemma.
Lemma 13.2.1 (PLS decomposition in terms of the T -score matrix)

The P loading matrix for X and the C̄ loading matrix for Y can in general be
expressed as
P = X T T (T T T )−1 , (13.82)
T T −1
C̄ = Y T (T T ) , (13.83)
(13.84)
where we have assumed that T T T is non-singular.

The PLS decomposition, Equations (13.76) and (13.77), can in general be
expressed as
X = T (T T T )−1 T T X + G, (13.85)
T −1 T
Y = T (T T ) T Y + F. (13.86)
The T ∈ RN ×a score vector matrix for X can in general be expressed as
T = XW (P T W )−1 = XW (T T XW )−1 (T T T ). (13.87)
4
Orthonormal score vector matrix
In some applications, e.g. in subspace identification algorithms for dynamic

systems, orthonormal decompositions are prefered. The PLS decomposition,
Equations (13.76) and (13.77), can be rewritten as
X = To PoT + G, (13.88)
Y = To CoT + F, (13.89)
where
−1
To = T Λt 2 , (13.90)
1
−1
P o = P Λt = 2
Pt Λt 2 , (13.91)
1
−1
Co = C̄Λt = 2
CSΛt 2 . (13.92)
The scaled score vector To is orthonormal, i.e. ToT To = Ia . Note that Λt = T T T

is a diagonal matrix.
Note also the following alternative description of the orthonormal PLS de-
composition
To Po
z }| { z }| {
− 12 −1
X = T Λt Λt 2 PtT +G, (13.93)
−1 −1
Y = T Λt 2 Λt 2 SC T +F. (13.94)
| {z }
Co
13.2.3 The Partial Least Squares regression and prediction
We will in this section discuss some alternative methods for computing the PLS
estimate of the matrix of regression coefficients B.
From Equation (13.77) we have that the prediction of Y is given by
Ŷ = T C̄ T . (13.95)
Similarly, from Equation (13.76) we have that the prediction of X (or the
information in X used for modeling Y ) is given by
X̂ = T P T . (13.96)
The matrix of PLS regression coefficients is given by
B̂ = W (P T W )−1 C̄ T . (13.97)
Note that only the weight matrix W , loading matrix C̄ and the loading matrix
P are needed in order to compute B. Hence, the score vector matrices T and U
are not explicitly needed. T is implicitly computed in order to define P and C̄.
U need not to be computed in order to compute B̂. Note also that the matrix
P T W which have to be inverted is of size a × a. P T W is an upper triangular

matrix with ones on the diagonal. This should be utilized when computing B̂
for numerical efficiency.
The matrix of PLS regression coefficients can equivalently be written as
B̂ = W (PtT W )−1 SC T = W (T T XW )−1 SC T , (13.98)
where PtT W = T T XW is an upper triangular matrix with tTi ti on the diagonal.

Se also Lemma 13.2.1.
Proof of Equation (13.97)
The prediction of Y can be written as
Ŷ = X̂ B̂. (13.99)
Substituting for X̂ = T P T and B̂ given by (13.97) gives
Ŷ = T P T W (P T W )−1 C̄ T = T C̄ T , (13.100)
which is equivalent to Equation (13.95).

QED.
The matrix of regression coefficients can be computed recursively during the
PLS iterations. This is numerically preferable because we in this case do not
have to explicitly inverting the upper triangular matrix P T W . We have
a
X
B̂ = (Ir − wi−1 pTi−1 )wi c̄Ti , w0 = 0, p0 = 0. (13.101)
i=1
This formula can be proved from the PLS updating expressions Yi+1 = Yi −
Xi wi c̄Ti and Xi+1 = Xi (I − wi pTi ). We also have the following alternative
expression
Xa
s̄i
B̂ = (Ir − wi−1 pTi−1 )wi T cTi , w0 = 0, p0 = 0. (13.102)
i=1
ti ti
Remark 13.6 (Prediction) Note that the prediction of Y usually is com-

puted as
Ŷ = X B̂, (13.103)
when X is not used for computing B̂. Hence, Equation (13.103) is used for
validation and prediction.
13.2.4 Computing projections from the PLS decomposition
Consider the basic linear relationship
Y T = BT X T + ET . (13.104)
The projection of (the row space) of a matrix A onto (the row space) of a matrix
B is defined as follows
A/B = AB T (BB T )† B, (13.105)
where (·)† denotes the Moore-Penrose pseudo inverse. We also define the pro-
jection of the row space of A onto the orthogonal complement of the row space
of B as
AB ⊥ = A − AB T (BB T )† B. (13.106)
This implies that A is decomposed into two terms which are orthogonal to each
other, i.e.
A = A/B + AB ⊥ . (13.107)
Then we have the following projection of Y T onto X T .
Y T /X T = B T X T + E T /X T . (13.108)
The projection of E T onto X T is assumed to be zero. This is the case when

N → ∞ and the noise process ek is white and serially uncorrelated with xi .
The projection can then be computed as
Y T /X T = B̂ T X̂ T . (13.109)
Substituting for the decomposition X̂ = T P T and the estimate of the regression

matrix B̂ = W (P T W )−1 C̄ T into (13.109) gives
Y T /X T = C̄(P T W )−T W T P T T = C̄T T . (13.110)
Using the orthonormal PLS decomposition Y = To CoT and X = To PoT we have

the following alternative description
XT
z }| {
Y /X = Co Po (Po Po ) Po ToT .
T T T T −1
(13.111)
where we have used (13.110). Comparing (13.111) with (13.108) shows that
B T = Co PoT (Po PoT )−1 and B = (Po PoT )−1 Po CoT , (13.112)
when Po PoT ∈ Rr×r is non-singular.

Note also that the projection of Y T onto the orthogonal complement of X T
is equal to the residual GT from the PLS decomposition.
13.2.5 A semi standard PLS implementation
The version of the PLS algorithm which is presented in this section is denoted
semi standard because it is based on the SVD.
Algorithm 13.2.4 (Standard PLS implementation with SVD)
X1 = X, Y1 = Y
XiT Yi = U SV T The Singular Value Decomposition
wi = U (:, 1) The weight vector for X
YiT ti
c̄i = tT
Scaled weight vector for Y
i ti
Yi c̄i
ūi = T
c̄i c̄i
Score vector for Y
XiT ti
pi = tT
Loading vector for X
i ti
Xi+1 = Xi − ti pTi Updating of X, (rank one reduction of X)
Yi+1 = Yi − ti c̄Ti Updating of Y , (rank one reduction of Y )
end
The loading matrix P , the weight matrix W and the loading matrix C̄ are then
given by
£ ¤ £ ¤ £ ¤
P = p1 · · · pa , W = w1 · · · wa , C̄ = c̄1 · · · c̄a (13.113) .
The matrix of regression coefficients can then be computed as BPLS = W (P T W )−1 C̄ T .

Note that P T W is upper triangular with ones on the diagonal.
The algorithm also computes the score vector matrices
£ ¤ £ ¤
T = t1 · · · ta , Ū = ū1 · · · ūa . (13.114)
13.2.6 The number of components a
The number of variables in X is r. The number of PLS components a (or

factors) is bounded from above by r. Hence,
1 ≤ a ≤ r.
The problem of choosing the number of components a is in general a trade of

between, on the one hand, “rank decision” based on the score matrices U and
T or the largest singular values s̄i of XiT Yi , i = 1, · · · , r, and, on the other hand,
model validation on independent data.
The PLS algorithm is constructed in order to maximize the linear relation-
ship between the ui score for Yi and the ti score for Xi . Recall that the ui score
vector lies in the column space of Yi and that the ti score vector lies in the
13.3 Further interpretations of the PLS algorithm 239
column space of Xi . See Problem (13.1). In order to chose the effective number
of components a one can plot the score vector ui against the score vector ti for
all i = 1, · · · , a and choose a from the plot where there is no linear relationship
between ua and ta .
Another strategy is simply to choose a as the number of “non zero” signifi-
cant singular values s̄i ∀ i = 1, · · · , r.
The problem of choosing a can of course be done recursively during the
PLS iteration. One procedure is to plot ui against ti at each PLS iteration
and by inspection decide if a new component should be incorporated. Another
procedure is to at each PLS iteration check if s̄i is significant different from
zero.
We end this section by pointing out that it often make sense to be parsimo-
nious when choosing the number of components.
13.3 Further interpretations of the PLS algorithm
13.3.1 General results
The results in this sections are general in the sense that it holds for all com-
ponents a. However, we will restrict our results to 1 ≤ a ≤ r where r is the
number of X variables.
Remark 13.7 We have the following general expression for the Squared Pre-
diction Error (SPE) when the PLS algorithm is used
VN =k Y − X B̂ k2F =k Y − T C̄ T k2F (13.115)
which can be expressed as
VN = tr(Y T Y ) − 2tr(SΛ−1 T −1 T
t T Y C) + tr(CSΛt SC ) (13.116)
by using that C̄ = CS(T T T )−1 . Furthermore, (13.116) can be evaluated as
VN (B) =k Y − X B̂ k2F = tr(Y T Y ) − tr(S 2 Λ−1
t )
a
X s̄ 2
= tr(Y T Y ) − i
Tt
, (13.117)
t
i=1 i i
where we have assumed that Λt = T T T is non-singular. a is the number of

components which usually is bounded as 1 ≤ a ≤ r. For each new component
which is incorporated, the variance in the output variables is reduced by the
s̄2
positive term tTit . It is assumed that tTi ti 6= 0.
i i
4
Remark 13.8 Putting the gradient matrix of Equation (13.116) with respect
to C equal to zero, gives the following general PLS condition
SC T = T T Y. (13.118)
This expression is to our knowledge new.
We have also derived a quite different and general expression for the SPE
as follows.
Remark 13.9 Substituting Equation (13.118) into (13.116) shows that the
SPE can be written as
VN = tr(Y T Y ) − tr(Y T T Λ−1 T T T T

t T Y ) = tr(Y Y ) − tr(Y T0 T0 Y ). (13.119)
The first term on the right hand side is the sum of variances for each output
(response) variable (times N − 1). The second term is the sum of variances for
the predicted outputs (times N − 1), i.e.,
Ŷ = X B̂ = T \Y = T Λ−1 T T
t T Y = To To Y, (13.120)
where To is the orthogonal score matrix. Note that for finite N we have E(ŷk ŷkT ) ≈
1 T T T
N −1 Ŷ Ŷ and (N − 1)E(ŷk ŷk ) ≈ tr(Ŷ Ŷ ).
The second term on the right hand side of (13.119) can be computed from
the SVD of the matrix
−1
Λt 2 T T Y = ToT Y = Ũ S̃ Ṽ T . (13.121)
Then,
min(a,m)
X
T
VN = tr(Y Y ) − s̃2i . (13.122)
i=1
Note the difference in the second term on the right hand side of Equations
(13.117) and (13.122). The number of singular values s̄ is a but the number of
singular values s̃i is min(a, m).
4
13.3.2 Special case results
The results in this section holds in general only for one component, i.e., a = 1.
Remark 13.10 We have the following expression for the Squared Prediction
Error (SPE) when a = 1.
s̄2
VN =k Y − XB1 k2F = tr(Y T Y ) − , (13.123)
tT t
where T = t = Xw and s̄ is the maximum singular value of X T Y . We have

assumed that tT t is non-singular.
4
13.4 A covariance based PLS implementation 241
Remark 13.11 We have the following expression for the squared Frobenius
norm of u − t
k u − t k2F = tr(Y T Y ) + tr(X T X) − 2tr(W T X T Y C), (13.124)
where t = XW and u = Y C. Hence, in the PLS algorithm the weight vectors

are computed as
min k t − u k2F = max(tT u). (13.125)

W,C W,C
This gives an alternative interpretation of the basic PLS optimization Problem

13.1.
4
Remark 13.12 In methods to solve least squares problems it is often useful to

have some knowledge of the norm of the solution. An example is the Tikhonov
regularized solution or the least squares solution with a quadratic inequality side
constraint on the solution B̂. See Hansen (1992).
For PLS with one component (a = 1) we have
s̄ k X T Y kF
k B1 kF = = . (13.126)
tT t wT X T Xw
Note that, in this case, the above also holds for the 2-norm of the solution, i.e.,
by replacing k · kF in (13.126) with k · k2 .
13.4 A covariance based PLS implementation
From the basic linear Equation (13.3) we have
X T Y = X T XB + X T E (13.127)
which is the normal equation for the LS problem. We will in the following
present a version of the PLS algorithm which works with X T Y and X T X rather
than X and Y .
The weights w and c are computed from the SVD of X T Y . First, initialize
(X T X)i = X T X and (X T Y )i = X T Y , then compute the weights wi and ci
from the SVD of (X T Y )i as follows,
(X T Y )i = U SV T , wi = U (:, 1), ci = U (:, 1), s̄i = S(1, 1). (13.128)
Compute the loading vector pi as

(X T X)i wi
pi = , where zi = wiT (X T X)i wi . (13.129)
zi
The loading vector c̄i can be computed as
(X T Y )Ti wi s̄i
c̄i = = ci . (13.130)
zi zi
Finally, the correlation matrix X T Y and the covariance matrix X T X are up-
dated directly as
(X T X)i+1 = (I − wi pTi )T (X T X)i (I − wi pTi ), (13.131)

T
(X Y )i+1 = (I − wi pTi )T (X T Y )i . (13.132)
Equations (13.128) to (13.132) are computed for all i = 1, · · · , a components,

where 1 ≤ a ≤ r. This procedure results in the matrices W , C̄ and P . The ma-
trix of regression coefficients can then be computed as BPLS = W (P T W )−1 C̄.
Remark that the product P T W is upper triangular with ones on the diagonal.
This structure should be used for efficient computation of (P T W )−1 .
The covariance based PLS implementation is very simple. It is very fast
compared to the traditional PLS when the number of observations N is very
large compared to the number of input variables r and the number of output
variables m.
The initialization of the covariance matrix X T X and the correlation ma-
trix X T Y can be a problem due to rounding off errors when the number of
observations N is large. The algorithm which is presented in the next section
is numerically robust.
13.5 A square root PLS algorithm
Consider the QR decomposition of the concatenated matrix

· ¸
£ ¤ £ ¤ R11 R12
X Y = Q1 Q2 . (13.133)
0 R22
From this we have that X = Q1 R11 and Y = Q1 R12 + Q2 R22 . Substituting into
the linear equation Y = XB + E gives the normal equation
R12 = R11 B, (13.134)
where R11 ∈ Rr×r is upper triangular and R12 ∈ Rr×m .

A standard PLS algorithm can then be used to decompose R11 and R12 as
follows
R11 = T P T + G, (13.135)
T
R12 = T C̄ + F (13.136)
and to compute the weighting matrix W . The PLS estimate of B is then

BPLS = W (P T W )−1 C̄ T .
The above algorithm is numerically robust in the sense that it avoid possibly
rounding off errors when computing X T Y . The reason for this is that the square
root PLS algorithm instead works on X T Y = R11T R . However, the square root
12
PLS algorithm is not so fast as the covariance based PLS algorithm.
13.6 Undeflated PLS 243
13.6 Undeflated PLS
The PLS algorithm is iterative. At each iteration a rank one reduced (deflated)
matrix Xi is computed. Usually a rank one reduction (deflation) of Y is also
carried out. Deflation of Y is not necessary in order to compute the regression
matrix B. The reason is that XiT Yi = XiT Y . In the following an LS method
without deflation is presented. This method is denoted UDPLS. The results
from UDPLS are in general different from the results from PLS.
Problem 13.2 (UDPLS optimization problem)
Maximize
f (W, C) = tr(T T U ),
with respect to W and C, where
T = XW,
U = Y C,
subject to
W T W = I,
C T C = I.
The Lagrangian associated with this optimization problem is given by
L(w, c) = tr(W T X T Y C) + tr[ΓT (I − W T W )] + tr[ΣT (I − C T C)], (13.137)
where Γ and Σ are matrices with Lagrange multipliers associated with the con-
straints. The optimal solution is found by setting all possible gradient matrices
of the Lagrangian to zero. We have
∂L
= (C T Y T X − 2ΓT W T )T = 0, ⇒ X T Y C = 2W Γ. (13.138)
∂W
∂L
= (W T X T Y − 2ΣT C T )T = 0, ⇒ Y T XW = 2CΣ. (13.139)
∂C
These derivatives are restricted to symmetric matrices Γ and Σ.
For non-singular Σ we have C = 0.5Y T XW Σ−1 , which gives
X T Y Y T XW = 4W ΓΣ, (13.140)
Y T XX T Y C = 4CΣΓ. (13.141)
The key step is the following SVD
X T Y = USV T . (13.142)
This gives
X T Y Y T XU = USS T , (13.143)
T T T
Y XX Y V = VS S. (13.144)
Chose W and C as
W = U, (13.145)
C = V. (13.146)
Note that
T = XW, (13.147)
T T T
T U = W X Y C = S. (13.148)
The matrix of regression coefficients are given by
B̂UDPLS = W (T T T )−1 SC T = U(U T X T XU)−1 SV. (13.149)
where one should use only the a first components of T and U. The estimate
can also be written as
B̂UDPLS = U(U T X T XU)−1 U T X T Y. (13.150)
Note that the above method is equivalent to the ordinary PLS algorithm when
a = 1.
13.7 Total Least Squares and Truncated Total Least

Squares
Total Least Squares (TLS) and Truncated Total Least Squares (TTLS) are
methods for solving over-determined linear systems of equations. To be specific,
TLS is a technique to solve models of the type Y = XB + E, where not only
the left hand side Y is subject to errors but also X is subject to errors. This
model is known by statisticians as the errors in variables model. This model is
also known as a bi-linear model because Y is linear in both X and B but both
B and X is unknown, since X is subject to errors. Software for TLS and TTLS
are presented in Hansen (1992).
13.7.1 Total Least Squares
Consider the Total Least Squares (TLS) problem

min f (Z, B), (13.151)
Z,B
where
f (Z, B) = kX − Zk2F + kY − ZBk2F , (13.152)
where k · kF is the Frobenius norm. The solution to this optimization problem
can be found from the SVD of the compound (or concatenated) matrix [X Y ].
We have
· ¸· T ¸
£ ¤ £ ¤ S1 0 V11 V21 T
X Y = U1 U2 T V , (13.153)
0 S2 V12 22
13.7 Total Least Squares and Truncated Total Least Squares 245
where U1 ∈ RN ×r , U2 ∈ RN ×m , S1 ∈ Rr×r , S2 ∈ Rm×m , V11 ∈ Rr×r , V12 ∈

Rr×m , V21 ∈ Rm×r and V22 ∈ Rm×m . The solution is given by
−1
BTLS = −V12 V22 , (13.154)
T
Z = U1 S1 V11 . (13.155)
In order to prove this results one should note that kAk2F = tr(AT A). The results
in this section are from Moor and David (199?).
Proof of TLS solution
An argumentation/proof of the above TLS solution is given in the following.

First, note that the model (13.3) can be equivalently written as
· ¸
£ ¤ −B
X Y = E. (13.156)
Im
From the SVD (13.153) we have

· ¸T · ¸T
£ ¤ V11 V12
X Y = U1 S1 + U2 S2 . (13.157)
V21 V22
From this we have

· ¸
£ ¤ V12
X Y = U2 S2 (13.158)
V22
and
· ¸
£ ¤ −1
V12 V22 −1
X Y = U2 S2 V22 . (13.159)
Im
Comparing (13.159) with (13.156) we get the TLS estimate of the regression
matrix (13.154). From this we also have an estimate of the residual matrix E.
13.7.2 Truncated Total Least Squares
The TLS solution presented in the above Section can be extended to problems
where X is nearly rank deficient. A technique to solve this problem is Truncated
Total Least Squares (TTLS). See e.g., Fierro et al, Hansen (1992). Basically,
the idea£ is to only
¤ retain the 1 ≤ a ≤ r largest singular values of the compound
matrix X Y in the matrix S1 . The TTLS solution is similar to the TLS
solution, exept that we now have
V22 ∈ R(r−a+m)×m , V12 ∈ Rr×(r−a+m) (13.160)
and
†
BTTLS = −V12 V22 . (13.161)
Remark 13.13 Note also that the above TLS and TTLS methods can be mod-
ified to work directly on the normal Equation (13.127). In this case we take
the SVD of the concatenated matrix [X T X X T Y ]. The solution are then found
from the SVD and the following formulation of the normal equation
· ¸
£ T ¤ −B
X X XT Y = X T E. (13.162)
Im
13.8 Performance and PLS
Motivated by the TLS performance objective (13.152) we suggest the following

performance objective to be minimized
f (P, B) = kX − T P T k2F + kY − T P T Bk2F . (13.163)
Let us first study the second part which is a measure of the residual F . For one
component we have
kF k2F = kY − tpT Bk2F . (13.164)
Substituting for BPLS = w(pT w)−1 c̄T and c̄ = Y T t/(tT t) gives (or simply re-
place the right hand side of (13.164) with the Gram-Schmidt orthogonalization
process, Equation (13.42), which is central in the PLS algorithm)
ttT
kF k2F = k(I − )Y k2F . (13.165)
tT t
Using the relationship between the Frobenius norm and trace we have
ttT
kF k2F = tr[Y T (I − )Y ] (13.166)
tT t
13.9 Optimal solution
From the discussion in this work we have that the PLS solution is of the form
B = W (W T X T XW )−1 W T X T Y (13.167)
where W ∈ Rr×a and a is the number of components bounded by 1 ≤ a ≤ r.

W is a so called weighting matrix. Different regression methods gives different
weighting matrices.
The squared Frobenius norm of the residual Y − XB is given by
V (W ) = kY − XBk2F = tr(Y T Y ) − tr(Y T XW (W T X T XW )−1 W T X T Y

(13.168)
).
Assume first for simplicity that W is equal to a vector w. The squared Frobenius
norm of the residual is in this case given by
Y T XwwT X T Y wT X T Y Y T Xw
V (w) = kY − XBk2F = Y T Y − = Y T
Y − (13.169)
.
wT X T Xw wT X T Xw
where we also have assumed that Y is a vector.

The minimum weight vector w can be found by putting the gradient of V (w)
with respect to w equal to zero. The gradient is given by
dV (w) 2X T Y Y T Xw(wT X T Xw) − wT X T Y Y T Xw(2X T Xw)

=− (13.170)
dw (wT X T Xw)2
Putting the gradient equal to zero gives
wT X T Y Y T Xw T
X T Y Y T Xw = X Xw (13.171)
wT X T Xw
T T T
This is a generalized eigenvalue problem, i.e. λ1 = w wXT XYTYXwXw is a gener-
alized eigenvalue of the square matrices X T Y Y T X and X T X and w is the
corresponding generalized eigenvector.
From this we have that a solution in general (i.e. when W ∈ Rr×a and the
number of components satisfy 1 ≤ a ≤ r) can be computed by the following
generalized eigenvalue problem
X T Y Y T XW = X T XW Λ (13.172)
where Λ ∈ Rr×r is a diagonal matrix with the generalized eigenvalues on the

diagonal. W ∈ Rr×r is the corresponding generalized eigenvector matrix. W
and Λ can e.g. be computed by the MATLAB function eig(·, ·), i.e.
£ ¤
W, Λ = eig(X T Y Y T X, X T X) (13.173)
The weight matrix corresponding to the first a generalized eigenvalues is

then given by
W := W (:, 1 : a); (13.174)
Note that it is possibile to compute only the a first generalized eigenvectors.

The norm of the residual is given by
Xa
V (W ) = kY − XBk2F T T
= tr(Y Y ) − tr(Λ(1 : a, 1 : a)). = tr(Y Y ) − (13.175)
λi
i=1
Example 13.2
Consider the following example from Hansen (1992)
Y X E
z }| { z }| { z }|
B z }| {
0.27 0.16 0.10 · ¸{ 0.01
 0.25  =  0.17 0.11  1.00 +  −0.03  . (13.176)
1.00
3.33 2.02 1.29 0.02
The problem adressed is to find the best estimate of B from given data matrices
X and Y and knowledge of the model structure (13.3).
· ¸
7.01
BLS = , k BLS k2 = 10.94, k Y − XBLS k2 = 0.02. (13.177)
−8.40
· ¸
1.1703
BPLS = , k BPLS k2 = 1.3885, k Y − XBPLS k2 = 0.0322.
(13.178)
0.7473
· ¸
1.1703
BTTLS = , k BTTLS k2 = 1.3885, k Y − XBTTLS k2 = 0.0322.
(13.179)
0.7473
A major difficulty with the above ordinary least squares solution BLS in
(13.177) is that its norm is
√ significantly grater than the norm of the exact
solution, which is kBk2 = 2. One component (a = 1) was specified for the
PLS and TTLS algorithms. The PLS and TTLS solutions are almost similar
for this example.
13.11 Conclusion
The PLS estimate of the matrix of regression coefficients can be computed as

BPLS = W (W T X T XW )−1 W T X T Y, (13.180)
where W ∈ Rr×a is a weight matrix. If the weight matrix W is square and non-
singular then this expression reduces to the ordinary least squares estimate.
The PLS algorithm is in general not optimal with respect to (minimizing)
the Squared Prediction error, i.e., the PLS estimate of the regression coefficients
gives in general not minimum of k Y − XBPLS kF . This can be shown by
counterexample. The optimal solution can be written in terms of an optimal
weight matrix W which is derived and presented in Section (13.9).
References
De Moor, B, and J. David (199?). Total least squares and the algebraic Ric-
cati equation. Katholieke Universiteit Leuven, B-3001 Leuven, Belgium.
Internal report.
Fierro, R. D., G. H. Golub, P. C. Hansen and D. P. O’Leary (199?) Regular-
ization by truncated total least squares.
Hansen, P. C. (1992). Regulatrization Tools. A Matlab Package for Analysis
and Solution of Discrete Ill-Posed Problems. Danish Computing Centre
for Research and Education, DK-2800 Lyngby, Denmark.
Höskuldsson, A. (1996). Prediction Methods in Science and technology, COLOURSCAN
Warsaw, ISBN 87-985941-0-9.
13.12 Appendix: proofs
The Squared Prediction error (SPE) can be expressed as
VN (B) =k Y − T C̄ T k2F = tr{(Y − T C̄ T )T (Y − T C̄ T )}

= tr{(Y T − CSΛ−1 T −T T T
t T )(Y − T Λt S C(13.181)
)},
where we have used Equation (13.78). We have
VN (B) =k Y − T C̄ T k2F = tr(Y T Y ) − 2tr(CSΛt−1 T T Y ) + tr(CSΛ−1 T −T T T

t T T Λt S C(13.182)
),
and
VN (B) = tr(Y T Y ) − 2tr(SΛ−1 T −1 T T

t T Y C) + tr(Λt S C CS) (13.183)
We have from Definition 13.2 and Remark 13.4 that the diagonal of matrix
M = T T Y C = T T Ũ is equal to the diagonal of matrix S. We also have that
the matrix C T C have ones on the diagonal, S is a diagonal matrix and Λt is a
diagonal matrix. This gives
VN (B) = tr(Y T Y ) − 2tr(S 2 Λ−1 −1 2 T −1 2

t ) + tr(Λt S ) = tr(Y Y ) − tr(Λt S (13.184)
).
QED.
Note that a general condition for PLS is given by
SC T = T T Y. (13.185)
This condition is to our knowledge new. Equation (13.185) can be derived by

putting the gradient matrix dVN (B)/dtr(C) = 0, i.e., putting the gradient ma-
trix of (13.183) to zero. Using this condition we derive the following expression
for the decomposition of Y
Y = T Λ−1 T
t T Y + F. (13.186)
Hence, the projection of the row space of Y T onto the row space of T T is equal
to
Y T /T T = Y T T Λ−1 T
t T , (13.187)
and, similarly
T \Y = T Λ−1 T T
t T Y = To To Y. (13.188)
Using (13.185) the SPE can be written as
VN = tr(Y T Y ) − tr(Y T T Λt−1 T T Y ). (13.189)

−1
The second term on the right hand side is related to the SVD of Λt 2 T T Y .
Define
−1
Λt 2 T T Y = ToT Y = Ũ S̃ Ṽ T . (13.190)
Then,
min(a,m)
X
VN = tr(Y T Y ) − s̃2i . (13.191)
i=1
Note also the following alternative expression for the PLS regression matrix
BPLS = W (P T W )−1 (T T T )−1 T T Y. (13.192)
A proof is given as follows
T \Y = XBPLS = T Λ−1 T
t T Y =| P T} W (P T W )−1 (T T T )−1 T T Y . (13.193)
T{z
| {z }
X BPLS
13.12.2 The PLS algorithm in terms of the W weight matrix
The results in this section is of interest sinse it shows that the entire PLS
algorithm is dependent on the weight matrix W , only. We have the following
results.
The PLS prediction of Y can be expressed as
Ŷ = XW (W T X T XW )−1 W T X T Y. (13.194)
This result can be proved by combining Ŷ = T (T T T )−1 T T Y and T = XW (T T XW )−1 Λt .

Comparing this with Ŷ = XBPLS gives
BPLS = W (W T X T XW )−1 W T X T Y. (13.195)
Similarly as (13.194), the PLS prediction of X can be expressed as
X̂ = XW (W T X T XW )−1 W T X T X. (13.196)
The above result are important since it shows that the PLS estimate of
B can be computed from X, Y and the weight matrix W , only. The only
problem is how to compute the weight matrix W . The PLS techique is only
one approach.
Note also that if W is an arbitrary but square non-singular matrix, then
(13.195) reduces to the ordinary least squares (LS) solution. This shows that the
PLS estimate is equivalent to the LS solution when the number of components
is a = r.
An expression for the rank one reduction of X in PLS can be derived from
the equation for the residual
G = X − XW (W T X T XW )−1 W T X T X. (13.197)
Similarily, for the outputs
F = Y − XW (W T X T XW )−1 W T X T Y. (13.198)
The consequense of this is that the SPE is
VN (B) = k Y − Ŷ k2F =k Y − XBPLS k2F

= tr(Y T Y ) − tr(Ŷ T Ŷ )
= tr(Y T Y ) − tr(Y T XW (W T X T XW )−1 W T X T Y ). (13.199)
13.13 Matlab code: The PLS1 algorithm

13.14 Matlab code: The PLS2 algorithm 253


Appendix A
Proof of the ESSM
A.1 Proof of the ESSM
The common equation in subspace identification is derived from the linear state
space model and given by
d
ỹk|L = OL xk + HL ũk|L (A.1)
The scalar positive integer parameter L can be viewed as the horizon necessary
for the observability matrix OL to be of rank n. Hence L ≥ n − rank(D) + 1 in
order for the state space vector xk to be identified from Equation (A.1). Thus
it make sense to define L as the identification horizon.
Equation (A.1) can be used to predict M steps into the future. Assume
M ≥ 1. We have
d
ỹk+M |L = OL xk+M + HL ũk+M |L (A.2)
The state space vector in the future xk+M is related to the present state xk
through the state equation. Hence we have
xk+M = AM xk + CM
d
ũk|M (A.3)
Substituting Equation (A.3) into (A.2) gives
ỹk+M |L = OL AM xk + OL CM
d d
ũk|M + HL ũk+M |L (A.4)
The present state xk is determined from Equation (A.1) and given by
xk = O† (ỹk|L − HL
d
ũk|L ) (A.5)
Finally, substituting xk defined in equation (A.5) into (A.4) gives
ỹk+M |L = OL AM OL† (ỹk|L − HL

d d
ũk|L ) + OL CM d
ũk|M + HL ũk+M |L (A.6)
256 Proof of the ESSM
which can be written as
ỹk+M |L = ÃM
L ỹk|L + B̃L|L+M ũk|L+M + C̃L|L+M ẽk|L+M (A.7)
£ d d
¤ £ M d ¤
B̃L|L+M = OL CM HL − ÃL HL 0Lm×M r (A.8)
£ s s
¤ £ M s ¤
C̃L|L+M = OL CM HL − ÃL HL 0Lm×M m (A.9)
†
ÃM
L = OL AM OL (A.10)
The ESSM model correspond to Equation (A.7) with M = 1 and k ≥ L. The

past inputs and outputs, 0 ≤ k < L, is used as instruments to remove future
noise.
A.2 Discussion
An interesting result is obtained by choosing M = L. In this case the toepliz

matrix HLd can be determined directly if the inputs are sufficiently exiting.
Consider the case in which M = L, then
Yk+L|L = ÃLL Yk|L + B̃L Uk|2L + C̃L Ek|2L (A.11)

£ ¤
B̃L = OL CL − ÃL d
L HL HL
d (A.12)
£ ¤
C̃L = OL CL − ÃL s
L HL HL
s (A.13)
A.3 Selector matrices
Our experience is that the choice of the parameter L (number of block rows in
the extended observability matrix) should be chosen as small as possibile. I.e.
close to the minimal value in order to ensure observability.
The reason for this is twofold. First, the computational expence will increase
when L is increased. Second, for porely exiting input signals, we avoid (pseudo)
inversion of badly conditioned Hankel matrices when L is small. This will
usually influence on the variance of the parameter estimates.
Sy ỹk+M |L = ÃM
p Sy ỹk|L + B̃p Su ũk|L+M + C̃p Se ẽk|L+M (A.14)
ÃM
p = Sy ÃM
L Sy
†
(A.15)
B̃p = Sy B̃L Su† (A.16)
C̃p = Sy C̃L Se† (A.17)
d
A.4 A study of the predictor ZJ|L+1 for YJ|L+1 . 257
d
A.4 A study of the predictor ZJ|L+1 for YJ|L+1 .
A possible subspace based predictor for the matrix YJ|L+1 of future outputs is
d
given by the matrix ZJ|L+1 as defined in Di Ruscio (1997). I.e.,
· ¸
d UJ|L+g
ZJ|L+1 = YJ|L+1 /
W1
Bz
z }| {·
h i · U ¸h i UJ|L+g
¸
T
UJ|L+g W1T J|L+g T T
UJ|L+g W1 ) −1
= YJ|L+1 ( ,
W1 W1
(A.18)
where
· ¸
U0|J
W1 = , (A.19)
Y0|J
and Bz is the matrix of regression coefficients from the corresponding LS prob-
lem. For simplicity of notation we will skip the subscript indexes in the follow-
ing. First let us compute the regression coefficient matrix
· ¸
£ T T
¤ U U T U W T −1
Bz = Y U YW . (A.20)
WUT WWT
The inverse of a partitioned matrix is (Kailath (1980))
· ¸−1
UUT UWT
=
WUT WWT
· ¸
(U U T )−1 + (U U T )−1 U W T ∆−1 W U T (U U T )−1 −(U U T )−1 U W T ∆−1
−∆−1 W U T (U U T )−1 ∆−1
(A.21)
where ∆ is the Schur complement of U U T given by
∆ = W W T − W U T (U U T )−1 U W T = W U ⊥ W T (A.22)
where we have used index f for future. This gives
£ ¤
Bz = B1 B2 (A.23)
where
B1 = Y U T ((U U T )−1 + (U U T )−1 U W T ∆−1 W U T (U U T )−1 ) − Y W T ∆−1 W U T (U U T )−1
= (Y − Y U ⊥ W T (W U ⊥ W T )−1 W )U T (U U T )−1 , (A.24)
T T −1 T −1 T −1
B2 = −Y U (U U ) UW ∆ +YW ∆ ,
⊥ T ⊥ T −1
= Y U W (W U W ) . (A.25)
d
Hence, ZJ|L+1 is the sum of two parts, i.e.,
d
ZJ|L+1 = YJ|L+1 /W1 UJ|L+g + YJ|L+1 /UJ|L+g W1 = B1 U + B2 W
= (Y − Y U ⊥ W T (W U ⊥ W T )−1 W )/U + Y U ⊥ W T (W U ⊥ W T )−1 W.
(A.26)
258 Proof of the ESSM
Appendix B
Linear Algebra and Matrix

Calculus
B.1 Trace of a matrix
The trace of a n × m matrix A is defined as the sum of the diagonal elements

of the matrix, i.e.
n
X
tr(A) = aii (B.1)
i=1
We have the following trace operations on two matrices A and B of apropriate

dimensions
tr(AT ) = tr(A) (B.2)

tr(AB T ) = tr(AT B) = tr(B T A) = tr(BAT ) (B.3)
tr(AB) = tr(BA) = tr(B T AT ) = tr(AT B T ) (B.4)
tr(A ± B) = tr(A) ± tr(B) (B.5)
B.2 Gradient matrices
∂
∂X tr[X] = I (B.6)
∂
∂X tr[AX] = AT (B.7)
∂ T
∂X tr[AX ] = A (B.8)
∂
∂X tr[AXB] = AT B T (B.9)
∂ T
∂X tr[AX B] = BA (B.10)
∂ T
∂X tr[XX] = 2X (B.11)
∂ T
∂X tr[XX ] = 2X (B.12)
260 Linear Algebra and Matrix Calculus
∂
∂X tr[X n ] = n(X n−1 )T (B.13)
∂
∂X tr[AXBX] = AT X T B T + B T X T AT (B.14)
∂ AXB AXB T
∂X tr[e ] = (Be A) (B.15)
∂
tr[XAX T ] = 2XA, if A = AT (B.16)
∂X
∂
∂X T
tr[AX] = A (B.17)
∂ T T
∂X T
tr[AX ] = A (B.18)
∂
∂X T
tr[AXB] = BA (B.19)
∂ T T T
∂X T
tr[AX B] = A B (B.20)
∂
∂X T
tr[eAXB ] = BeAXB A (B.21)
B.3 Derivatives of vector and quadratic form
The derivative of a vector with respect to a vector is a matrix. We have the

following identities:
∂x
∂xT
=I (B.22)
∂ T
∂x (x Q) = Q (B.23)
∂ T
∂x (Qx) = Q (B.24)
(B.25)
The derivative of a scalar with respect to a vector is a vector. We have the
following identities:
∂
∂x (y T x) = y (B.26)
∂ T
∂x (x x) = 2x (B.27)
∂ T T
∂x (x Qx) = Qx + Q x (B.28)
∂
∂x (y T Qx) = QT y (B.29)
Note that if Q is symmetric then
∂
(xT Qx) = Qx + QT x = 2Qx. (B.30)
∂x
B.4 Matrix norms
The trace of the matrix product AT A is related to the Frobenius norm of A as

follows
kAk2F = tr(AT A), (B.31)
where A ∈ RN ×m .
B.5 Linearization 261
B.5 Linearization
Given a vector function f (x) ∈ Rm where x ∈ Rn . The derivative of the vector

f with respect to the row vector xT is defined as
 ∂f ∂f1 ∂f1

∂x1
1
∂x2 · · · ∂x n
 ∂f2 ∂f2 ∂f2 
∂f  ∂x1 ∂x2 · · · ∂x 
=  . n  m×n
∂xT  . .
. . . .
. ∈R (B.32)
 . . . . 
∂fm ∂fm ∂fm
∂x1 ∂x2 ··· ∂xn
Given a non-linear differentiable state space model
ẋ = f (x, u), (B.33)

y = g(x). (B.34)
A linearized model around the stationary points x0 and u0 is

˙ = Ax + Bu,
δx (B.35)
δy = Dx, (B.36)
where
∂f
A = |x ,u , (B.37)
∂xT 0 0
∂f
B = |x ,u , (B.38)
∂uT 0 0
∂g
D = |x ,u , (B.39)
∂xT 0 0
and where
x = x − x0 , (B.40)
u = u − u0 . (B.41)
B.6 Kronecer product matrices
Given a matrix X ∈ RN ×r . Let Im be the (m × m) identity matrix. Then
(X ⊗ Im )T = X T ⊗ Im , (B.42)
(Im ⊗ X)T = Im ⊗ X T . (B.43)

262 Linear Algebra and Matrix Calculus
Appendix C
Optimization of the DSR

algorithm
C.1 Existence of an optimal solution
For a given pair (D, A) there exists a canonical form (Dc , Ac ) where Dc consist
of only ones and zeroes and with a minimal number of free parameters in Ac .
Denote the free parameters as
θA = css(A) (C.1)
where css(·) is defined as the column structure string operator. See Di Ruscio
(1993) p. 121.
The problem of identifying the poles as well as the system matrices D and
A is reduced to solving the following optimization problem.
T
min kR42 − OL A(OL OL )−1 OT R32 kF (C.2)
θA
where k · kF is the matrix Frobenius norm.

The size of this optimization problem can be further reduced. Define the
SVD
R32 = U SV T (C.3)
The Frobenius norm is invariant under orthogonal transformations. See Golub

and Loan (1883) p. 15. Then we have
T
min kR42 V − OL A(OL OL )−1 OT U SkF (C.4)
θA
Monte Carlo simulations shows that this solution is optimal and different from
the original DSR algorithm.
Note also that A can be computed from the following least squares problem
£ T O )−1 O T R )T ⊗ O
¤
cs(R42 ) = ((OL L 32 L cs(A) (C.5)
264 Optimization of the DSR algorithm
C.2 Optimized algorithm for computing the B and

E matrices
In the DSR algorithm the system order and the matrices D and A are computed
in a first step. There are many alternatives for computing the system matrices
B and E from known D and A. One alternative to the one presented in Di
Ruscio (1995) is given in the following. A similar strategy for computing the B
and E matrices for SISO systems is presented in the covariance based method
in Di Ruscio (1993) p. 50. Monte Carlo simulations of combined deterministic
and stochastic systems shows that the algorithm improves the accuracy of the
B E estimates compared to the original DSR algorithm. The method is optimal
in a least squares sense, i.e., the parameters in B and E are computed from a
least squares optimal solution. Note that the algorithms gives the same result
in the purely deterministic case.
From the DSR algorithm we have the matrix equation
M = B̃L R11 , (C.6)
where
def
M = R41 − ÃL R31 ∈ RLm×(L+g)r (C.7)
and
T
ÃL = OL A(OL OL )−1 OLT ∈ RLm×Lm . (C.8)
The matrices R41 ∈ RLm×(L+g)r , R31 ∈ RLm×(L+g)r and R11 ∈ R(L+g)r×(L+g)r

in Equations (C.6) and (C.7) are submatrices from the lower Left, Q-orthogonal
(LQ) decomposition.
The matrix equation (C.6) can be solved for B̃L for known matrices M and
R11 provided R11 is non-singular. The matrices B and E can be computed from
B̃L when A and D are known. In Di Ruscio (1995) a variant of this strategy
which only have to invert a submatrix of R11 of size (n − rank(D) + 1 + g)r is
presented.
In the following an alternative and optimal method for computing the B
and E matrices are presented. For given system matrices D and A the matrix
B̃L is a function of the unknown matrices B and E. We will show that Equation
(C.6) can be written as a least squares problem for the unknown parameters in
B and E.
In the following a proper state space model with E 6= 0m×r is assumed.
From the definition we recall that g = 1 for this case. Define
L+g
X
def
N = RiT ⊗ (Ei−1 − ÃL Ei ) ∈ RLm(L+g)r×(n+gm)r (C.9)
i=1
where ⊗ denotes the Kronecker tensor product. The matrices Ri and Ei are
defined below. We then have the following least squares problem for computing
C.2 Optimized algorithm for computing the B and E matrices 265
the B and E matrices.

· ¸
B
cs(M ) = N cs( ) (C.10)
E
where cs(·) is the column string operator, i.e. a column vector constructed by
stacking the columns of a matrix on each other. The matrix N is determined
from the term B̃L R32 by extensive use of the identity
cs(AXB) = (B T ⊗ A)cs(X)
for the column string operation of the product of a triple matrices A, X and B
of appriopriate dimensions (Vetter (1970)).
Then we have the least squares solution
· ¸
B
cs( ) = N † cs(M ) (C.11)
E
where N † = (N T N )−1 N T is the Moore-Penrose pseudo-inverse of N .

The matrices
Ri ∈ Rr×(L+g)r ∀ i = 1, · · · , L + g
are block rows in the R11 ∈ R(L+g)r×(L+g)r matrix. The matrices
Ei−1 ∈ RLm×(n+mg) ∀ i = 1, · · · , L + g
are defined as follows
£ ¤
E0 = OL 0Lm×m , EL+1 = 0Lm×(n+m) , (C.12)
     
 D 0m×m   0m×n Im×m   0m×n 0m×m 
     
 DA 0m×m   D 0m×m   0m×n 0m×m 
E1 =   , E2 =   , EL =  (C.13)
 .
 .. ..   .. ..   .. .. 
 . .   . .   . . 
DAL−2 0m×m DAL−3 0m×m 0m×n Im×m
Some examples are presented in the following in order to illustrate the above
method for computing the B and E matrices.
Example C.1
Considder the equation
Z = H3d U0|3 (C.14)
and the problem of computing B and E from known matrices Z, U0|3 , D and
A. We have
H3d
z
 }| {  
E 0m×r 0m×r U1
Z =  DB E 0m×r   U2  (C.15)
DAB DB E U3
and
     
0m×n Im×m · ¸ 0m×n 0m×m · ¸ 0m×n 0m×m · ¸
B B B
Z= D 0m×m  U1 +  0m×n Im×m  U2 +  0m×n 0m×m  U
(C.16)
3
E E E
DA 0m×m D 0m×m 0m×n Im×m
Then we have
· ¸
B
cs(Z) = N cs( ) (C.17)
E
where
     
N = U1T ⊗  D 
0m×m  + U2T ⊗  0m×n Im×m  + U3T ⊗  0m×n 0m×m (C.18)
Example C.2
Considder the ESSM with L = 1 and g = 1, i.e.
Y1|1 = Ã1 Y0|1 + B̃1 U0|2 (C.19)
Assume that the input and output data as well as Ã1 are known. Define
M = Y1|1 − Ã1 Y0|1 (C.20)
Then we have the matrix equation
M = B̃1 U0|2 (C.21)
Problem: Compute B and E from known matrices M , Ã1 and U0|2 .

Solution: We have
B̃1
z£ · ¸
¤ }| £ ¤{ U1
M = ( DB E − Ã1 E 0m×r ) (C.22)
U2
The matrix B̃1 can be written as follows

£ ¤
B̃1 = DB − Ã1 E E (C.23)
From this we have
· ¸ · ¸
B B
M = (E0 − Ã1 E1 ) U1 + E1 U2 (C.24)
E E
where
£ ¤ £ ¤
E0 = D 0m×m , E1 = 0m×n Im×m . (C.25)
This gives the least squares problem
· ¸
B
cs(M ) = N cs( ) (C.26)
E
where
N = U1T ⊗ (E0 − Ã1 E1 ) + U2T ⊗ E1 (C.27)
C.2 Optimized algorithm for computing the B and E matrices 267
Example C.3
Considder the ESSM
Y1|2 = Ã2 Y0|2 + B̃2 U0|3 (C.28)
Assume that the input and output data as well as Ã2 are known. Define
M = Y1|2 − Ã2 Y0|2 (C.29)
Then we have
M = B̃2 U0|3 (C.30)
Problem: Compute B and E from known matrices M , Ã2 and U0|3 .

Solution: We have
B̃2  
z· ¸ }| · ¸{ U1
DB E 0m×r E 0m×r 0m×r
M =( − Ã2 )  U2  (C.31)
DAB DB E DB E 0m×r
U3
The matrix B̃2 can be written as follows

· · ¸ · ¸ · ¸ · ¸ · ¸ ¸
DB E E 0m×r 0m×r
B̃2 = − Ã2 − Ã2 (C.32)
DAB DB DB E E
From this we have
· ¸ · ¸ · ¸
B B B
M = (E0 − Ã2 E1 ) U1 + (E1 − Ã2 E2 ) U2 + E2 U3 (C.33)
E E E
where
· ¸ · ¸ · ¸
D 0m×m 0m×n Im×m 0m×n 0m×m
E0 = , E1 = , E2 = (C.34)
This gives the least squares problem
· ¸
B
cs(M ) = N cs( ) (C.35)
E
where
N = U1T ⊗ (E0 − Ã2 E1 ) + U2T ⊗ (E1 − Ã2 E2 ) + U3T ⊗ E2 (C.36)
Remark C.1
We conclude this section by commenting that the cs(·) operation of R42 = ÃL R32
is given by
cs(R42 ) = [((OL OL )−1 OL
T
R32 )T ⊗ OL ]cs(A) (C.37)
which define a least squares problem for computing the system matrix A for
known observability matrix. Note that the column space of R32 coincide with
the column space of the observability matrix, i.e. the observability matrix can
be estimated as the column space of R32 .
Appendix D
D-SR Toolbox for MATLAB

270 D-SR Toolbox for MATLAB
D.1 The DSR function for use with MATLAB
THE ACRONYM:
Deterministic and Stochastic system identification and Realization (DSR).
PURPOSE:
Given the output data matrix Y and the input data matrix U . DSR estimate
the system order n, the matrices A, B, D, E, CF, F and the initial state vector
x0 in the discrete time, combined deterministic and stochastic dynamic model,
on innovations form:
xt+1 = Axt + But + Cet , xt=0 = x0, (D.1)

yt = Dxt + Eut + et , (D.2)
where t is discrete time, C = CF · F −1 is the Kalman filter gain matrix and

E(et eTt ) = F F T is the innovations noise covariance matrix.
SYNOPSIS:
[A, B, D, E, CF, F, x0] = dsr(Y, U, L)

[A, B, D, E, CF, F, x0] = dsr(Y, U, L, g)
[A, B, D, E, CF, F, x0] = dsr(Y, U, L, g, J, M, n)
PARAMETERS ON INPUT:
Y − An (N × m) matrix with output data/observations. N is the number of
observations and m is the number of output variables.
U − An (N × r) matrix with input data/observations. r is the number of
input variables.
L − Integer. Number specifying the future horizon used for predicting the sys-
tem order. Choose L > 0 such that the assumed system order satisfy
n ≤ Lm. L = 1 is default.
OPTIONAL INPUT PARAMETERS:

DSR has four optional integer parameters. These parameters is for advanced
use. A description is given below.
g − Integer. Optional model structure parameter. g = 1 is default. If g = 0 then

a model with E = 0m×r is estimated. If g = 1 then E is estimated.
J − Integer. Number defining the past horizon used to define the instrumental
variables used to remove future noise. J = L is default and recommended.
M − Integer. With M = 1 (default) a simple method for computing CF and F
is used. A more computational expensive method is used when M 6= 1 .
n − Integer. Optional specification of model order, 0 < n ≤ Lm.

Subspace System Identification Theory and Applications: Lecture Notes

Uploaded by

Copyright:

Available Formats

Subspace System Identification Theory and Applications: Lecture Notes

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Subspace System Identification Theory and Applications: Lecture Notes

Uploaded by

Copyright:

Available Formats

SUBSPACE SYSTEM IDENTIFICATION

Theory and applications

Telemark Institute of Technology

6th edition December 11, 2009 Telemark University College

The material in this report/book is meant to be used as lecture notes in practi-

2.2.7 Error analysis . . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Combined Deterministic and Stochastic System Identification

4 On the DSR algorithm 77

5 Subspace Identification for Dynamic Process Analysis and Mod-

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

6 Dynamic Subspace Modeling (DSR) and Static Multivariate

7 On Subspace Identification of the Extended Observability Ma-

7.5 Standard deviation of the estimates . . . . . . . . . . . . . . . . . 148

8 Subspace identification of the states and the Kalman filter gain161

8.6 Further remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

9 Effect of Scaling and how to Treat Trends 201

11 Input experiment design 213

12 Special topics in system identification 219

13 On the Partial Least Squares Algorithm 221

13.1 Notation, basic- and system-definitions . . . . . . . . . . . . . . . 221

A Proof of the ESSM 255

B Linear Algebra and Matrix Calculus 259

C Optimization of the DSR algorithm 263

D D-SR Toolbox for MATLAB 269

1.1 State space model

xk+1 = Axk + Buk + Cvk (1.1)

where the integer k ≥ 0 is discrete time, xk ∈ Rn is the state vector, uk ∈ Rr is

• The pair (D, A) is observable.

1.2 Inputs and outputs

• Identify models for control.

1.3 Data organization

1.3.1 Hankel matrix notation

Definition 1.1 (Hankel matrix) Given a (vector or matrix) sequence of data

where nr is the number of rows in st and nc is the number of columns in st .

• L is the number of nr-block rows in St0 |L .

• K is the number of nc-block columns in St0 |L .

Example 1.1 Given a vector valued sequence of observations

with N = 10. Choose parameters t0 = 3, L = 2 and use all observations to

The number of columns is in this case K = N − t0 = 7.

1.3.2 Extended data vectors

Given a number L known output vectors and a number L + g known input

1.3.3 Extended data matrices

Known data matrix of output variables

Known data matrix of input variables

where the subscript i denotes the number of block rows.

where the subscript i denotes the number of block columns.

i.e., with B substituted with C in Equation (8.4.

where the subscript i denotes the number of block rows and i + g − 1 is

1.5 Extended output equation

1.5.1 Extended vector output equation

1.5.2 Extended output matrix equation

Theorem 1.6.1 (Observability)

for all numbers L of block rows given by

block rows in the observability matrix to ensure that rank(OLmin ) = n.

Example 1.2 (Observability)

1.7 Extended state space model

An extended state space model (ESSM) will be defined in this section.

Theorem 1.7.1 (Extended state space model)

yk+1|L = ÃL yk|L + B̃L uk|L+g + C̃L vk|L+1 (1.24)