Important Information
Warranty
The media on which you receive National Instruments software are warranted not to fail to execute programming instructions, due to defects
in materials and workmanship, for a period of 90 days from date of shipment, as evidenced by receipts or other documentation. National
Instruments will, at its option, repair or replace software media that do not execute programming instructions if National Instruments receives
notice of such defects during the warranty period. National Instruments does not warrant that the operation of the software shall be
uninterrupted or error free.
A Return Material Authorization (RMA) number must be obtained from the factory and clearly marked on the outside of the package before any
equipment will be accepted for warranty work. National Instruments will pay the shipping costs of returning to the owner parts which are covered by
warranty.
National Instruments believes that the information in this document is accurate. The document has been carefully reviewed for technical accuracy. In
the event that technical or typographical errors exist, National Instruments reserves the right to make changes to subsequent editions of this document
without prior notice to holders of this edition. The reader should consult National Instruments if errors are suspected. In no event shall National
Instruments be liable for any damages arising out of or related to this document or the information contained in it.
EXCEPT AS SPECIFIED HEREIN, NATIONAL INSTRUMENTS MAKES NO WARRANTIES, EXPRESS OR IMPLIED, AND SPECIFICALLY DISCLAIMS ANY WARRANTY OF
MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. CUSTOMER’S RIGHT TO RECOVER DAMAGES CAUSED BY FAULT OR NEGLIGENCE ON THE PART OF NATIONAL
INSTRUMENTS SHALL BE LIMITED TO THE AMOUNT THERETOFORE PAID BY THE CUSTOMER. NATIONAL INSTRUMENTS WILL NOT BE LIABLE FOR DAMAGES RESULTING
FROM LOSS OF DATA, PROFITS, USE OF PRODUCTS, OR INCIDENTAL OR CONSEQUENTIAL DAMAGES, EVEN IF ADVISED OF THE POSSIBILITY THEREOF. This limitation of
the liability of National Instruments will apply regardless of the form of action, whether in contract or tort, including negligence. Any action against
National Instruments must be brought within one year after the cause of action accrues. National Instruments shall not be liable for any delay in
performance due to causes beyond its reasonable control. The warranty provided herein does not cover damages, defects, malfunctions, or service
failures caused by owner’s failure to follow the National Instruments installation, operation, or maintenance instructions; owner’s modification of the
product; owner’s abuse, misuse, or negligent acts; and power failure or surges, fire, flood, accident, actions of third parties, or other events outside
reasonable control.
Copyright
Under the copyright laws, this publication may not be reproduced or transmitted in any form, electronic or mechanical, including photocopying,
recording, storing in an information retrieval system, or translating, in whole or in part, without the prior written consent of National
Instruments Corporation.
National Instruments respects the intellectual property of others, and we ask our users to do the same. NI software is protected by copyright and other
intellectual property laws. Where NI software may be used to reproduce software or other materials belonging to others, you may use NI software only
to reproduce materials that you may reproduce in accordance with the terms of any applicable license or other legal restriction.
Trademarks
MATRIXx™, National Instruments™, NI™, ni.com™, and Xmath™ are trademarks of National Instruments Corporation. Refer to the Terms of
Use section on ni.com/legalfor more information about National Instruments trademarks.
Other product and company names mentioned herein are trademarks or trade names of their respective companies.
Members of the National Instruments Alliance Partner Program are business entities independent from National Instruments and have no agency,
partnership, or joint-venture relationship with National Instruments.
Patents
For patents covering National Instruments products, refer to the appropriate location: Help»Patents in your software, the patents.txtfile
on your CD, or ni.com/patents.
WARNING REGARDING USE OF NATIONAL INSTRUMENTS PRODUCTS
(1) NATIONAL INSTRUMENTS PRODUCTS ARE NOT DESIGNED WITH COMPONENTS AND TESTING FOR A LEVEL OF
RELIABILITY SUITABLE FOR USE IN OR IN CONNECTION WITH SURGICAL IMPLANTS OR AS CRITICAL COMPONENTS IN
ANY LIFE SUPPORT SYSTEMS WHOSE FAILURE TO PERFORM CAN REASONABLY BE EXPECTED TO CAUSE SIGNIFICANT
INJURY TO A HUMAN.
(2) IN ANY APPLICATION, INCLUDING THE ABOVE, RELIABILITY OF OPERATION OF THE SOFTWARE PRODUCTS CAN BE
IMPAIRED BY ADVERSE FACTORS, INCLUDING BUT NOT LIMITED TO FLUCTUATIONS IN ELECTRICAL POWER SUPPLY,
COMPUTER HARDWARE MALFUNCTIONS, COMPUTER OPERATING SYSTEM SOFTWARE FITNESS, FITNESS OF COMPILERS
AND DEVELOPMENT SOFTWARE USED TO DEVELOP AN APPLICATION, INSTALLATION ERRORS, SOFTWARE AND HARDWARE
COMPATIBILITY PROBLEMS, MALFUNCTIONS OR FAILURES OF ELECTRONIC MONITORING OR CONTROL DEVICES,
TRANSIENT FAILURES OF ELECTRONIC SYSTEMS (HARDWARE AND/OR SOFTWARE), UNANTICIPATED USES OR MISUSES, OR
ERRORS ON THE PART OF THE USER OR APPLICATIONS DESIGNER (ADVERSE FACTORS SUCH AS THESE ARE HEREAFTER
COLLECTIVELY TERMED “SYSTEM FAILURES”). ANY APPLICATION WHERE A SYSTEM FAILURE WOULD CREATE A RISK OF
HARM TO PROPERTY OR PERSONS (INCLUDING THE RISK OF BODILY INJURY AND DEATH) SHOULD NOT BE RELIANT SOLELY
UPON ONE FORM OF ELECTRONIC SYSTEM DUE TO THE RISK OF SYSTEM FAILURE. TO AVOID DAMAGE, INJURY, OR DEATH,
THE USER OR APPLICATION DESIGNER MUST TAKE REASONABLY PRUDENT STEPS TO PROTECT AGAINST SYSTEM FAILURES,
INCLUDING BUT NOT LIMITED TO BACK-UP OR SHUT DOWN MECHANISMS. BECAUSE EACH END-USER SYSTEM IS
CUSTOMIZED AND DIFFERS FROM NATIONAL INSTRUMENTS' TESTING PLATFORMS AND BECAUSE A USER OR APPLICATION
DESIGNER MAY USE NATIONAL INSTRUMENTS PRODUCTS IN COMBINATION WITH OTHER PRODUCTS IN A MANNER NOT
EVALUATED OR CONTEMPLATED BY NATIONAL INSTRUMENTS, THE USER OR APPLICATION DESIGNER IS ULTIMATELY
RESPONSIBLE FOR VERIFYING AND VALIDATING THE SUITABILITY OF NATIONAL INSTRUMENTS PRODUCTS WHENEVER
NATIONAL INSTRUMENTS PRODUCTS ARE INCORPORATED IN A SYSTEM OR APPLICATION, INCLUDING, WITHOUT
LIMITATION, THE APPROPRIATE DESIGN, PROCESS AND SAFETY LEVEL OF SUCH SYSTEM OR APPLICATION.
Download from Www.Somanuals.com. All Manuals Search And Download.
Conventions
The following conventions are used in this manual:
[ ]
Square brackets enclose optional items—for example, [response]. Square
brackets also cite bibliographic references.
»
The » symbol leads you through nested menu items and dialog box options
to a final action. The sequence File»Page Setup»Options directs you to
pull down the File menu, select the Page Setup item, and select Options
from the last dialog box.
This icon denotes a note, which alerts you to important information.
bold
Bold text denotes items that you must select or click in the software, such
as menu items and dialog box options. Bold text also denotes parameter
names.
italic
Italic text denotes variables, emphasis, a cross-reference, or an introduction
to a key concept. Italic text also denotes text that is a placeholder for a word
or value that you must supply.
monospace
Text in this font denotes text or characters that you should enter from the
keyboard, sections of code, programming examples, and syntax examples.
This font is also used for the proper names of disk drives, paths, directories,
programs, subprograms, subroutines, device names, functions, operations,
variables, filenames, and extensions.
monospace bold
Bold text in this font denotes the messages and responses that the computer
automatically prints to the screen. This font also emphasizes lines of code
that are different from the other examples.
monospace italic
Italic text in this font denotes text that is a placeholder for a word or value
that you must supply.
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Commonly Used Nomenclature ......................................................................1-2
Related Publications........................................................................................1-3
Controllability and Observability Grammians ................................................1-7
Hankel Singular Values...................................................................................1-8
Singular Perturbation.......................................................................................1-11
Chapter 2
Additive Error Reduction
Truncation of Balanced Realizations.............................................................................2-2
Related Functions............................................................................................2-11
truncate( ).......................................................................................................................2-11
Related Functions............................................................................................2-11
redschur( )......................................................................................................................2-12
Algorithm ........................................................................................................2-12
Related Functions............................................................................................2-14
ophank( )........................................................................................................................2-14
Restriction........................................................................................................2-14
Algorithm ........................................................................................................2-15
Behaviors.........................................................................................................2-15
© National Instruments Corporation
v
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Contents
Onepass Algorithm ......................................................................................... 2-18
Multipass Algorithm ...................................................................................... 2-20
Impulse Response Error.................................................................................. 2-22
Chapter 3
Algorithm........................................................................................................ 3-4
Securing Zero Error at DC.............................................................................. 3-8
Algorithm........................................................................................................ 3-14
right and left.................................................................................................... 3-15
Error Bounds................................................................................................... 3-20
Chapter 4
Controller Reduction....................................................................................... 4-2
Controller Robustness Result.......................................................................... 4-2
Fractional Representations.............................................................................. 4-5
wtbalance( ) ................................................................................................................... 4-10
Algorithm........................................................................................................ 4-12
Related Functions............................................................................................ 4-15
Xmath Model Reduction Module
vi
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
fracred( ) ........................................................................................................................4-15
Restrictions......................................................................................................4-15
Algorithm ........................................................................................................4-18
Related Functions............................................................................................4-21
Chapter 5
Utilities
hankelsv( )......................................................................................................................5-1
Related Functions............................................................................................5-2
Algorithm ........................................................................................................5-2
Related Functions............................................................................................5-4
compare( )......................................................................................................................5-4
Chapter 6
Tutorial
Plant and Full-Order Controller.....................................................................................6-1
ophank( )..........................................................................................................6-9
wtbalance.........................................................................................................6-12
fracred..............................................................................................................6-20
Appendix A
Appendix B
Technical Support and Professional Services
Index
© National Instruments Corporation
vii
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
1
Introduction
This chapter starts with an outline of the manual and some useful notes. It
also provides an overview of the Model Reduction Module, describes the
functions in this module, and introduces nomenclature and concepts used
throughout this manual.
Using This Manual
This manual describes the Model Reduction Module (MRM), which
provides a collection of tools for reducing the order of systems.
Readers who are not familiar with Parameter Dependent Matrices (PDMs)
should consult the Xmath User Guide before using MRM functions and
tools. Although several MRM functions accept both PDMs and matrices as
information that is useful for simulation, plotting, and signal labeling.
Document Organization
•
Chapter 1, Introduction, starts with an outline of the manual and some
useful notes. It also provides an overview of the Model Reduction
nomenclature and concepts used throughout this manual.
•
•
•
Chapter 2, Additive Error Reduction, describes additive error
and perturbation of balanced realizations.
Chapter 3, Multiplicative Error Reduction, describes multiplicative
error reduction presenting considerations for using multiplicative
rather than additive error reduction.
Chapter 4, Frequency-Weighted Error Reduction, describes
frequency-weighted error reduction problems, including controller
reduction and fractional representations.
© National Instruments Corporation
1-1
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
•
•
Chapter 5, Utilities, describes three utility functions: hankelsv( ),
stable( ), and compare( ).
Chapter 6, Tutorial, illustrates a number of the MRM functions and
their underlying ideas.
Bibliographic References
Throughout this document, bibliographic references are cited with
bracketed entries. For example, a reference to [VODM1] corresponds
to a paper published by Van Overschee and De Moor. For a table of
bibliographic references, refer to Appendix A, Bibliography.
Commonly Used Nomenclature
This manual uses the following general nomenclature:
•
•
Matrix variables are generally denoted with capital letters; vectors are
represented in lowercase.
G(s) is used to denote a transfer function of a system where s is the
Laplace variable. G(q) is used when both continuous and discrete
systems are allowed.
•
•
H(s) is used to denote the frequency response, over some range of
frequencies of a system where s is the Laplace variable. H(q) is used
to indicate that the system can be continuous or discrete.
A single apostrophe following a matrix variable, for example, x’,
denotes the transpose of that variable. An asterisk following a matrix
variable, for example, A*, indicates the complex conjugate, or
Hermitian, transpose of that variable.
Conventions
This publication makes use of the following types of conventions: font,
format, symbol, mouse, and note. These conventions are detailed in
Chapter 2, MATRIXx Publications, Online Help, and Customer Support,
of the MATRIXx Getting Started Guide.
Xmath Model Reduction Module
1-2
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
Related Publications
For a complete list of MATRIXx publications, refer to Chapter 2,
MATRIXx Publications, Online Help, and Customer Support, of the
MATRIXx Getting Started Guide. The following documents are particularly
useful for topics covered in this manual:
•
•
•
•
•
•
•
•
•
•
MATRIXx Getting Started Guide
Xmath User Guide
Control Design Module
Interactive Control Design Module
Interactive System Identification Module, Part 1
Interactive System Identification Module, Part 2
Model Reduction Module
Optimization Module
Robust Control Module
Xμ Module
MATRIXx Help
Model Reduction Module function reference information is available in
the MATRIXx Help. The MATRIXx Help includes all Model Reduction
functions. Each topic explains a function’s inputs, outputs, and keywords
in detail. Refer to Chapter 2, MATRIXx Publications, Online Help, and
Customer Support, of the MATRIXx Getting Started Guide for complete
instructions on using the help feature.
Overview
The Xmath Model Reduction Module (MRM) provides a collection of tools
for reducing the order of systems. Many of the functions are based on
state-of-the-art algorithms in conjunction with researchers at the Australian
National University, who were responsible for the original development of
some of the algorithms. A common theme throughout the module is the use
of Hankel singular values and balanced realizations, although
considerations of numerical accuracy often dictates whether these tools are
used implicitly rather than explicitly. The tools are particularly suitable
when, as generally here, quality of approximation is measured by closeness
of frequency domain behavior.
© National Instruments Corporation
1-3
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
As shown in Figure 1-1, functions are provided to handle four broad tasks:
•
•
•
Model reduction with additive errors
Model reduction with multiplicative errors
Model reduction with frequency weighting of an additive error,
including controller reduction
•
Utility functions
Functions
Additive Error
Model Reduction
Multiplicative
Model Reduction
Frequency Weighted
Model Reduction
balmoore
redschur
ophank
truncate
balance
mreduce
bst
mulhank
wtbalance
fracred
Utility Functions
hankelsv
stable
compare
Figure 1-1. MRM Function
Xmath Model Reduction Module
1-4
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
Certain restrictions regarding minimality and stability are required of the
input data, and are summarized in Table 1-1.
Table 1-1. MRM Restrictions
balance( )
A stable, minimal system
balmoore ( )
A state-space system must be stable and minimal,
having at least one input, output, and state
bst( )
A state-space system must be linear,
continuous-time, and stable, with full rank along
the jω-axis, including infinity
compare( )
fracred( )
hankelsv( )
mreduce( )
Must be a state-space system
A state-space system must be linear and continuous
A system must be linear and stable
A submatrix of a matrix must be nonsingular
for continuous systems, and variant for discrete
systems
mulhank( )
A state-space system must be linear,
continuous-time, stable and square, with full
rank along the jω-axis, including infinity
ophank( )
A state-space system must be linear,
continuous-time and stable, but can be nonminimal
redschur( )
A state-space system must be stable and linear,
but can be nonminimal
stable ( )
No restriction
truncate( )
wtbalance( )
Any full-order state-space system
A state-space system must be linear and
continuous. Interconnection of controller and plant
must be stable, and/or weight must be stable.
Documentation of the individual functions sometimes indicates how the
restrictions can be circumvented. There are a number of model reduction
methods not covered here. These include:
•
•
Padé Approximation
Methods based on interpolating, or matching at discrete frequencies
© National Instruments Corporation
1-5
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
•
•
•
L2 approximation, in which the L2 norm of impulse response error (or,
by Parseval’s theorem, the L2 norm of the transfer-function error along
the imaginary axis) serves as the error measure
Markov parameter or impulse response matching, moment matching,
covariance matching, and combinations of these, for example,
q-COVER approximation
Controller reduction using canonical interactions, balanced Riccati
equations, and certain balanced controller reduction algorithms
Nomenclature
This manual uses standard nomenclature. The user should be familiar with
the following:
•
•
•
sup denotes supremum, the least upper bound.
The acute accent (´) denotes matrix transposition.
A superscripted asterisk (*) denotes matrix transposition and complex
conjugation.
•
•
•
λmax(A) for a square matrix A denotes the maximum eigenvalue,
presuming there are no complex eigenvalues.
Reλi(A) and |λi(A)| for a square matrix A denote an arbitrary real part
and an arbitrary magnitude of an eigenvalue of A.
X( jω) for a transfer function X(·) denotes:
∞
sup
[λmax[X*(jω)X(jω)]]1/2
–∞ < ω < ∞
•
•
An all-pass transfer-function W(s) is one where X( jω) = 1 for all ω;
to each pole, there corresponds a zero which is the reflection through
the jω-axis of the pole, and there are no jω-axis poles.
An all-pass transfer-function matrix W(s) is a square matrix where
W′(–jω)W( jω) = I
P > 0 and P ≥ 0 for a symmetric or hermitian matrix denote positive
and nonnegative definiteness.
•
•
P1 > P2 and P1 ≥ P2 for symmetric or hermitian P1 and P2 denote
P1 – P2 is positive definite and nonnegative definite.
A superscripted number sign (#) for a square matrix A denotes the
Moore-Penrose pseudo-inverse of A.
Xmath Model Reduction Module
1-6
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
•
An inequality or bound is tight if it can be met in practice, for example
1 + logx – x ≤ 0
is tight because the inequality becomes an equality for x = 1. Again,
if F(jω) denotes the Fourier transform of some f (t)∈ L2 , the
Heisenberg inequality states,
f(t)2dt
∫
------------------------------------------------------------------------------------------
1 ⁄ 2 ≤ 4π
1 ⁄ 2
t2 f(t) 2dt
ω2 F(jω) 2dω
∫
∫
and the bound is tight since it is attained for f(t) = exp + (–kt2).
Commonly Used Concepts
This section outlines some frequently used standard concepts.
Controllability and Observability Grammians
Suppose that G(s) = D + C(sI–A)–1B is a transfer-function matrix with
Reλi(A)<0. Then there exist symmetric matrices P, Q defined by:
PA′ + AP = –BB′
QA + A′Q = –C′C
These are termed the controllability and observability grammians of the
realization defined by {A,B,C,D}. (Sometimes in the code, WC is used for
P and WO for Q.) They have a number of properties:
•
P ≥ 0, with P > 0 if and only if [A,B] is controllable, Q ≥ 0 with Q > 0
if and only if [A,C] is observable.
∞
∞
•
•
P = eAtBB′eA′tdt and Q = eA′tC′CeAtdt
∫
∫
0
0
With vec P denoting the column vector formed by stacking column 1
of P on column 2 on column 3, and so on, and ⊗ denoting Kronecker
product
[I ⊗ A + A ⊗ I]vecP = –vec(BB′)
•
The controllability grammian can be thought of as measuring the
difficulty of controlling a system. More specifically, if the system is in
norm of u) required to bring it to the state x0 is x0P –1x0; so small
eigenvalues of P correspond to systems that are difficult to control,
while zero eigenvalues correspond to uncontrollable systems.
© National Instruments Corporation
1-7
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
•
•
The controllability grammian is also E[x(t)x′(t)] when the system
·
–
x = Ax + Bw has been excited from time ∞ by zero mean white
noise with E[w(t)w′(s)] = Iδ(t – s).
The observability grammian can be thought of as measuring the
information contained in the output concerning an initial state.
·
If x = Ax, y = Cx with x(0) = x0 then:
∞
y′(t)y(t)dt = x′0Qx0
∫
0
Systems that are easy to observe correspond to Q with large
eigenvalues, and thus large output energy (when unforced).
• lyapunov(A,B*B')produces P and lyapunov(A',C'*C)
produces Q.
–1
For a discrete-time G(z) = D + C(zI-A) B with |λi(A)|<1, P and Q are:
P – APA′ = BB′
Q – A′QA = C′C
The first dot point above remains valid. Also,
∞
∞
•
•
P =
AkBB′A′k and Q =
AkC′CA′k
∑
∑
k = 0
k = 0
with the sums being finite in case A is nilpotent (which is the case if
the transfer-function matrix has a finite impulse response).
[I–A⊗ A] vec P = vec (BB′)
lyapunov( )can be used to evaluate P and Q.
Hankel Singular Values
If P, Q are the controllability and observability grammians of a
transfer-function matrix (in continuous or discrete time), the Hankel
Singular Values are the quantities λi1/2(PQ). Notice the following:
•
All eigenvalues of PQ are nonnegative, and so are the Hankel singular
values.
•
The Hankel singular values are independent of the realization used to
calculate them: when A,B,C,D are replaced by TAT–1, TB, CT–1 and D,
then P and Q are replaced by TPT′ and (T–1)′QT–1; then PQ is replaced
by TPQT–1 and the eigenvalues are unaltered.
The number of nonzero Hankel singular values is the order or
McMillan degree of the transfer-function matrix, or the state
dimension in a minimal realization.
Xmath Model Reduction Module
1-8
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
•
Suppose the transfer-function matrix corresponds to a discrete-time
system, with state variable dimension n. Then the infinite Hankel
matrix,
CB CAB CA2B
CAB CA2B
CA2B
H =
has for its singular values the n nonzero Hankel singular values,
together with an infinite number of zero singular values.
The Hankel singular values of a (stable) all pass system (or all pass matrix)
are all 1.
Slightly different procedures are used for calculating the Hankel singular
values (and so-called weighted Hankel singular values) in the various
functions. These procedures are summarized in Table 1-2.
Table 1-2. Calculating Hankel Singular Values
(balance( ))
For a discussion of the balancing algorithm, refer to
the Internally Balanced Realizations section; the
Hankel singular values are given by
diag(R1/2) = HSV
balmoore( )
For a discussion of the balancing algorithm, refer to
the Internally Balanced Realizations section; the
matrix SH yields the Hankel singular values through
diag(SH)
hankelsv( )
ophank( )
real(sqrt(eig(p*q)))
Calls hankelsv( )
redschur( )
Computes a Schur decomposition of P*Q and then
takes the square roots of the diagonal entries
bst( )
Same as redschur( )except either P or Q can be
mulhank( )
wtbalance( )
fracred( )
a weighted grammian
© National Instruments Corporation
1-9
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
Internally Balanced Realizations
Suppose that a realization of a transfer-function matrix has the
controllability and observability grammian property that P = Q = Σ for
some diagonal Σ. Then the realization is termed internally balanced. Notice
that the diagonal entries σi of Σ are square roots of the eigenvalues of PQ,
that is, they are the Hankel singular values. Often the entries of Σ are
assumed ordered with σi ≥ σi+1
.
As noted in the discussion of grammians, systems with small (eigenvalues
of) P are hard to control and those with small (eigenvalues of) Q are hard
to observe. Now a state transformation T = α I will cause P = Q to be
replaced by α2P, α–2Q, implying that ease of control can be obtained at the
expense of difficulty of observation, and conversely. Balanced realizations
are those when ease of control has been balanced against ease of
observation.
Given an arbitrary realization, there are a number of ways of finding a
state-variable coordinate transformation bringing it to balanced form.
A good survey of the available algorithms for balancing is in [LHPW87].
One of these is implemented in the Xmath function balance( ).
The one implemented in balmoore( )as part of this module is more
sophisticated, but more time consuming. It proceeds as follows:
1. Singular value decompositions of P and Q are defined. Because P and
Q are symmetric, this is equivalent to diagonalizing P and Q by
orthogonal transformations.
P = UcSc U′
c
Q = UoSo U′o
2. The matrix,
H = S10 ⁄ 2UHSHV1H⁄ 2
is constructed, and from it, a singular value decomposition is obtained:
H = UHSHV′H
3. The balancing transformation is given by:
T = U0S0–1 ⁄ 2UHSH1 ⁄ 2
The balanced realization is T–1AT, T–1B, CT.
Xmath Model Reduction Module
1-10
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
This is almost the algorithm set out in Section II of [LHPW87]. The one
difference (and it is minor) is that in [LHPW87], lower triangular Cholesky
factors of P and Q are used, in place of UcSc1/2 and UOSO1/2 in forming H
in step 2. The grammians themselves need never be formed, as these
Cholesky factors can be computed directly from A, B, and C in both
continuous and discrete time; this, however, is not done in balmoore.
The algorithm has the property that:
T′QT′ = T–1P(T–1)′ = SH
Thus the diagonal entries of SH are the Hankel singular values.
The algorithm implemented in balance( )is older, refer to [Lau80].
A lower triangular Cholesky factor Lc of P is found, so that LcLc′ = P.
Then the symmetric matrix Lc′QLc is diagonalized (through a singular
value decomposition), thus L′cQLc = VRU′, with actually V = U. Finally,
the coordinate basis transformation is given by T = LcVR–1/4, resulting in
T′QT = T–1P(T–1)′ = R1/2
.
Singular Perturbation
A common procedure for approximating systems is the technique of
Singular Perturbation. The underlying heuristic reasoning is as follows.
Suppose there is a system with the property that the unforced responses of
the system contain some modes which decay to zero extremely fast. Then
an approximation to the system behavior may be attained by setting state
variable derivatives associated with these modes to zero, even in the forced
case. The phrase “associated with the modes” is loose: exactly what occurs
is shown below. The phrase “even in the forced case” captures a logical
flaw in the argument: smallness in the unforced case associated with initial
conditions is not identical with smallness in the forced case.
Suppose the system is defined by:
·
x1
x1
x2
B1
B2
A11 A12
A21 A22
=
+
u
·
x2
x1
y =
+ Du
C1 C2
x2
(1-1)
© National Instruments Corporation
1-11
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
and also:
–1
Reλ (A )<0 and Reλi(A11 – A12A A21) < 0.
22
i
22
Usually, we expect that,
–1
22
Reλi(A22) « Reλi(A11 – A12A A21)
in the sense that the intuitive argument hinges on this, but it is not necessary.
·
Then a singular perturbation is obtained by replacing x2 by zero; this
means that:
–1
22
–1
22
A21x1 + A22x2 + B2u = 0
Accordingly,
x2 = –A A21x1 – A B2u
or
–1
–1
·
x1 = (A11 = A12A22A21)x1 + (B1 – A12A22B2)u
–1
22
–1
22
(1-2)
Equation 1-2 may be an approximation for Equation 1-1. This means that:
•
•
The transfer-function matrices may be similar.
If Equation 1-2 is excited by some u(·), with initial condition x1(to), and
by,
•
•
x1(to) and x2(to) = –A–122A21x1(to) –A22 B2u(to),
then x1(·) and y(·) computed from Equation 1-1 and from Equation 1-2
should be similar.
If Equation 1-1 and Equation 1-2 are excited with the same u(·), have
the same x1(to) and Equation 1-1 has arbitrary x2, then x1(·) and y(·)
computed from Equation 1-1 and Equation 1-2 should be similar after
a possible initial transient.
As far as the transfer-function matrices are concerned, it can be verified that
they are actually equal at DC.
Xmath Model Reduction Module
1-12
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
Similar considerations govern the discrete-time problem, where,
x1(k + 1)
x1(k)
B1
B2
A11 A12
A21 A22
=
+
u(k)
x2(k + 1)
x2(k)
x1(k)
y(k) =
+ Du(k)
C1 C2
x2(k)
can be approximated by:
x1(k + 1) = [A11 + A12(I – A22)–1A21]x1(k) +
[B1 + A12(I – A22)–1B2]u(k)
yk = [C1 + C2(I – A22)–1A21]x1(k) +
[D + C2(I – A22)–1B2]u(k)
mreduce( )can carry out singular perturbation. For further discussion,
refer to Chapter 2, Additive Error Reduction. If Equation 1-1 is balanced,
singular perturbation is provably attractive.
Spectral Factorization
Let W(s) be a stable transfer-function matrix, and suppose a system S with
transfer-function matrix W(s) is excited by zero mean unit intensity white
noise. Then the output of S is a stationary process with a spectrum Φ(s)
related to W(s) by:
Φ(s) = W(s)W′(–s)
(1-3)
Evidently,
Φ( jω) = W(jω)W*( jω)
is Φ( jω) with Φ( jω) = |W( jω)|2.
In the matrix case, Φ is singular for some ω only if W does not have full
rank there, and in the scalar case only if W has a zero there.
Spectral factorization, as shown in Example 1-1, seeks a W(jω), given
Φ(jω). In the rational case, a W(jω) exists if and only if Φ(jω) is
© National Instruments Corporation
1-13
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
nonnegative hermitian for all ω. If Φ is scalar, then Φ(jω)≥0 for all ω.
Normally one restricts attention to Φ(·) with limω→∞Φ(jω)<∞. A key result
is that, given a rational, nonnegative hermitian Φ(jω) with
limω→∞Φ(jω)<∞, there exists a rational W(s) where,
•
•
•
W(∞)<∞.
W(s) is stable.
W(s) is minimum phase, that is, the rank of W(s) is constant in Re[s]>0.
In the scalar case, all zeros of W(s) lie in Re[s]≤0, or in Re[s]<0 if Φ(jω)>0
for all ω.
In the matrix case, and if Φ(jω) is nonsingular for some ω, it means that
W(s) is square and W–1(s) has all its poles in Re[s]≤ 0, or in Re[s]<0 if Φ(jω)
is nonsingular for all ω.
Moreover, the particular W(s) previously defined is unique, to within right
multiplication by a constant orthogonal matrix. In the scalar case, this
means that W(s) is determined to within a ±1 multiplier.
Example 1-1
Example of Spectral Factorization
Suppose:
ω2 + 1
ω2 + 4
s + 1
-----------
Then Equation 1-3 is satisfied by W(s) =
minimum phase.
, which is stable and
s + 2
s – 1
s + 2
s – 3 s – 1
s + 2 s + 2
–sTs + 1
-----------
----------- -----------
,
-----------
e
Also, Equation 1-3 is satisfied by
and
and
, and
s + 2
so forth, but none of these is minimum phase.
bst( )and mulhank( )both require execution within the program of
a spectral factorization; the actual algorithm for achieving the spectral
factorization depends on a Riccati equation. The concepts of a spectrum
and spectral factor also underpin aspects of wtbalance( ).
Xmath Model Reduction Module
1-14
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
Low Order Controller Design Through Order Reduction
The Model Reduction Module is particularly suitable for achieving low
order controller design for a high order plant. This section explains some of
the broad issues involved.
Most modern controller design methods, for example, LQG and H∞, yield
that, to obtain a low order controller using such methods, one must either
follow a high order controller design by a controller reduction step,
or reduce an initially given high order plant model, and then design a
controller using the resulting low order plant, with the understanding that
the controller will actually be used on the high order plant. Refer to
Figure 1-2.
High Order Plant
High Order Controller
Plant
Reduction
Controller
Reduction
Low Order Plant
Low Order Controller
Figure 1-2. Low Order Controller Design for a High Order Plant
Generally speaking, in any design procedure, it is better to postpone
approximation to a late step of the procedure: if approximation is done
early, the subsequent steps of the design procedure may have unpredictable
effects on the approximation errors. Hence, the scheme based on high order
Controller reduction should aim to preserve closed-loop properties as far
as possible. Hence the controller reduction procedures advocated in this
module reflect the plant in some way. This leads to the frequency weighted
reduction schemes of wtbalance( )and fracred( ), as described in
Chapter 4, Frequency-Weighted Error Reduction. Plant reduction logically
should also seek to preserve closed-loop properties, and thus should involve
the controller. With the controller unknown however, this is impossible.
Nevertheless, it can be argued, on the basis of the high loop gain property
within the closed-loop bandwidth that is typical of many systems, that
© National Instruments Corporation
1-15
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 1
Introduction
multiplicative reduction, as described in Chapter 4, Frequency-Weighted
Error Reduction, is a sound approach. Chapter 3, Multiplicative Error
Reduction, and Chapter 4, Frequency-Weighted Error Reduction, develop
these arguments more fully.
Xmath Model Reduction Module
1-16
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
2
Additive Error Reduction
This chapter describes additive error reduction including discussions of
truncation of, reduction by, and perturbation of balanced realizations.
Introduction
Additive error reduction focuses on errors of the form,
G( jω) – Gr( jω)
∞
where G is the originally given transfer function, or model, and Gr is the
reduced one. Of course, in discrete-time, one works instead with:
G(ejω) – Gr(ejω)
∞
As is argued in later chapters, if one is reducing a plant that will sit inside
a closed loop, or if one is reducing a controller, that again is sitting in a
closed loop, focus on additive error model reduction may not be
appropriate. It is, however, extremely appropriate in considering reducing
the transfer function of a filter. One pertinent application comes specifically
from digital filtering: a great many design algorithms lead to a finite
impulse response (FIR) filter which can have a very large number of
coefficients when poles are close to the unit circle. Model reduction
provides a means to replace an FIR design by a much lower order infinite
impulse response (IIR) design, with close matching of the transfer function
at all frequencies.
© National Instruments Corporation
2-1
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
Truncation of Balanced Realizations
A group of functions can be used to achieve a reduction through truncation
of a balanced realization. This means that if the original system is
·
x1
x1
x2
B1
B2
A11 A12
A21 A22
=
+
u
·
x2
y =
x + Du
C1 C2
(2-1)
and the realization is internally balanced, then a truncation is provided by
·
x1 = A11x1 + B1u
y = C1x1 + Du
The functions in question are:
• balmoore( )
• balance( )(refer to the Xmath Help)
• truncate( )
• redschur( )
One only can speak of internally balanced realizations for systems which
are stable; if the aim is to reduce a transfer function matrix G(s) which
contains unstable poles, one must additively decompose it into a stable part
and unstable part, reduce the stable part, and then add the unstable part back
in. The function stable( ), described in Chapter 5, Utilities, can be used
to decompose G(s). Thus:
G(s)
=
=
=
Gs(s) + Gu(s)(Gs(s) stable, Gu(s) unstable)
found by algorithm (reduction of Gs(s))
Gsr(s) + Gu(s) (reduction of G(s))
Gsr(s)
Gr(s)
Xmath Model Reduction Module
2-2
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
A very attractive feature of the truncation procedure is the availability
of an error bound. More precisely, suppose that the controllability and
observability grammians for [Enn84] are
Σ1
0
P = Q = Σ =
(2-2)
0 Σ2
with the diagonal entries of Σ in decreasing order, that is, σ1 ≥ σ2 ≥ ···. Then
the key result is,
G( jω) – Gr( jω) ≤ 2trΣ2
∞
with G, Gr the transfer function matrices of Equation 2-1 and Equation 2-2,
respectively. This formula shows that small singular values can, without
great cost, be thrown away. It also is valid in discrete time, and can be
improved upon if there are repeated Hankel singular values. Provided that
the smallest diagonal entry of Σ1 strictly exceeds the largest diagonal entry
of Σ2, the reduced order system is guaranteed to be stable.
Several other points concerning the error can be made:
•
•
The error G( jω)–Gr( jω) as a function of frequency is not flat; it is zero
at ω = ∞, and may take its largest value at ω = 0, so that there is in
general no matching of DC gains of the original and reduced system.
The actual error may be considerably less than the error bound at all
frequencies, so that the error bound formula can be no more than an
advance guide. However, the bound is tight when the dimension
reduction is 1 and the reduction is of a continuous-time
transfer-function matrix.
•
With g(·) and gr(·) denoting the impulse responses for impulse
responses of G and Gr and with Gr of degree k, the following L1 bound
holds [GCP88]
g – gr ≤ (4 2k + 1)trΣ2
1
This bound also will apply for the L∞ error on the step response.
It is helpful to note one situation where reduction is likely to be difficult (so
that Σ will contain few diagonal entries which are, relatively, very small).
Suppose G(s), strictly proper, has degree n and has (n – 1) unstable zeros.
(2n – 1)π/2. Much of this change may occur in the passband. Suppose Gr(s)
has degree n–1; it can have no more than (n – 2) zeros, since it is strictly
© National Instruments Corporation
2-3
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
proper. So, even if all zeros are unstable, the maximum phase shift when ω
moves from 0 to ∞ is (2n – 3)π/2. It follows that if G(jω) remains large in
magnitude at frequencies when the phase shift has moved past (2n – 3)π/2,
approximation of G by Gr will necessarily be poor. Put another way, good
approximation may depend somehow on removing roughly cancelling
pole-zeros pairs; when there are no left half plane zeros, there can be no
rough cancellation, and so approximation is unsatisfactory.
As a working rule of thumb, if there are p right half plane zeros in the
passband of a strictly proper G(s), reduction to a Gr(s) of order less than
p + 1 is likely to involve substantial errors. For non-strictly proper G(s),
having p right half plane zeros means that reduction to a Gr(s) of order less
than p is likely to involve substantial errors.
An all-pass function exemplifies the problem: there are n stable poles and
n unstable zeros. Since all singular values are 1, the error bound formula
indicates for a reduction to order n – 1 (when it is not just a bound, but
exact) a maximum error of 2.
Another situation where poor approximation can arise is when a highly
oscillatory system is to be replaced by a system with a real pole.
Reduction Through Balanced Realization Truncation
This section briefly describes functions that reduce( ), balance( ),
and truncate( )to achieve reduction.
• balmoore( )—Computes an internally balanced realization of a
system and optionally truncates the realization to form an
approximation.
• balance( )—Computes an internally balanced realization of a
system.
• truncate( )—This function truncates a system. It allows
examination of a sequence of different reduced order models formed
from the one balanced realization.
• redschur( )—These functions in theory function almost the same
as the two features of balmoore( ). That is, they produce a
state-variable realization of a reduced order model, such that the
transfer function matrix of the model could have resulted by truncating
a balanced realization of the original full order transfer function
matrix. However, the initially given realization of the original transfer
function matrix is never actually balanced, which can be a numerically
hazardous step. Moreover, the state-variable realization of the reduced
Xmath Model Reduction Module
2-4
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
order model is not one in general obtainable by truncation of an
internally-balanced realization of the full order model.
Figure 2-1 sets out several routes to a reduced-order realization. In
continuous time, a truncation of a balanced realization is again balanced.
This is not the case for discrete time, but otherwise it looks the same.
Full Order Realization
balmoore
(with both steps)
balmoore
(with first step)
balance
redschur
truncate
Balanced Realization of
Nonbalanced
Reduced Order Model
(in continuous time)
Realization of
Reduced Order Model
Reduced Order Model Transfer Function
Figure 2-1. Different Approaches for Obtaining the Same Reduced Order Model
Singular Perturbation of Balanced Realization
Singular perturbation of a balanced realization is an attractive way to
produce a reduced order model. Suppose G(s) is defined by,
·
x1
x1
x2
B1
B2
A11 A12
A21 A22
=
+
u
·
x2
y =
x + Du
C1 C2
© National Instruments Corporation
2-5
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
with controllability and observability grammians given by,
Σ1
0
P = Q = Σ =
0 Σ2
in which the diagonal entries of Σ are in decreasing order, that is,
σ1 ≥ σ2 ≥ ···, and such that the last diagonal entry of Σ1 exceeds
the first diagonal entry of Σ2. It turns out that Reλi(A–1)<0 and
22
Reλ (A11–A12A–1A21)< 0, and a reduced order model Gr(s) can be
22
i
defined by:
–1
–1
·
x = (A11 – A12A22A21)x + (B1 + –A12A22B2)u
–1
22
–1
22
y = (C1 – C2A A21)x + (D – C2A B2)u
The attractive feature [LiA89] is that the same error bound holds as for
balanced truncation. For example,
G( jω) – Gr( jω) ≤ 2trΣ2
∞
Although the error bounds are the same, the actual frequency pattern of
the errors, and the actual maximum modulus, need not be the same for
reduction to the same order. One crucial difference is that balanced
truncation provides exact matching at ω = ∞, but does not match at DC,
while singular perturbation is exactly the other way round. Perfect
matching at DC can be a substantial advantage, especially if input signals
are known to be band-limited.
Singular perturbation can be achieved with mreduce( ). Figure 2-1 shows
the two alternative approaches. For both continuous-time and discrete-time
reductions, the end result is a balanced realization.
Hankel Norm Approximation
In Hankel norm approximation, one relies on the fact that if one chooses an
approximation to exactly minimize one norm (the Hankel norm) then the
infinity norm will be approximately minimized. The Hankel norm is
defined in the following way. Let G(s) be a (rational) stable transfer
Xmath Model Reduction Module
2-6
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
function matrix. Consider the way the associated impulse response maps
inputs defined over (–∞,0] in L2 into outputs, and focus on the output over
[0,∞). Define the input as u(t) for t < 0, and set v(t) = u(–t). Define the
output as y(t) for t > 0. Then the mapping is
∞
y(t) = CexpA(t + r)Bv(r)dr
∫
0
if G(s) = C(sI-A)–1B. The norm of the associated operator is the Hankel
G
norm
of G. A key result is that if σ1 ≥ σ2 ≥ ···, are the Hankel singular
H
values of G(s), then G = σ1.
H
To avoid minor confusion, suppose that all Hankel singular values of G are
ˆ
distinct. Then consider approximating G by some stable G of prescribed
ˆ
degree k much that
is minimized. It turns out that
G – G
H
ˆ
infGˆ of degree k G – G = σk + 1(G)
H
ˆ
and there is an algorithm available for obtaining G . Further, the
ˆ
ˆ
optimum which is minimizing
does a reasonable job
G
G – G
H
ˆ
of minimizing
∞, because it can be shown that
G – G
ˆ
G – G
≤
σj
∞
∑
j = k + 1
ˆ
where n = deg G, with this bound subject to the proviso that G and G are
allowed to be nonzero and different at s = ∞.
ˆ
The bound on G – G is one half that applying for balanced truncation.
However,
•
•
It is actual error that is important in practice (not bounds).
The Hankel norm approximation does not give zero error at ω = ∞
or at ω = 0. Balanced realization truncation gives zero error at ω = ∞,
and singular perturbation of a balanced realization gives zero error
at ω = 0.
There is one further connection between optimum Hankel norm
ˆ
approximation and L∞ error. If one seeks to approximate G by a sum G + F,
ˆ
with stable and of degree k and with F unstable, then:
G
ˆ
infGˆ of degree k and F unstable G – G – F = σk + 1(G)
∞
© National Instruments Corporation
2-7
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
ˆ
Further, the G which is optimal for Hankel norm approximation also is
optimal for this second type of approximation.
In Xmath Hankel norm approximation is achieved with ophank( ).
The most comprehensive reference is [Glo84].
balmoore( )
[SysR,HSV,T] = balmoore(Sys,{nsr,bound})
The balmoore( )function computes an internally-balanced realization of
a continuous system and then optionally truncates it to provide a balance
reduced order system using B.C. Moore’s algorithm.
When balmoore( )is being used to reduce a system, its objective mirrors
that of redschur( ), therefore, if the same Sysand nsrare used for both
algorithms, the reduced order system should have the same transfer
function (though in general the state-variable realizations will be different).
When balmoore( )is being used to balance a system, its objective, like
that of balance, is to generate an internally-balanced state-variable
realization. The implementations are not identical.
balmoore( )only can be applied on systems that have a stable A matrix,
and are controllable and observable, (that is, minimal). Checks, which are
rather time-consuming, are included. The computation is intrinsically not
well-conditioned if Sysis nearly nonminimal. The first part of
balmoore( )serves to find a transformation matrix T such that the
controllability and observability grammians after transformation are equal,
and diagonal, with decreasing entries down the diagonal, that is, the system
representation is internally balanced. (The condition number of T is a
measure of the ill-conditioning of the algorithm. If there is a problem with
ill-conditioning, redschur( )can be used as an alternative.) If this
common grammian is Σ, then after transformation:
(continuous)
Σ A′ + A Σ = –BB′ Σ A + A′Σ = –C′C
(discrete)
Σ – A Σ A = –BB′ Σ - A′ Σ A = –C′C
Singular Values of Sys. In the second part of balmoore( ), a truncation
Xmath Model Reduction Module
2-8
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
of the balanced system occurs, (assuming nsris less than the number of
states). Thus, if the state-space representation of the balanced system is
A11 A12
A21 A22
B1
B2
A =
B =
C =
C1 C2
with A11 possessing dimension nsr× nsr, B1 possessing nsrrows and C1
possessing nsrcolumns, the reduced order system SysRis:
(continuous)
(discrete)
·
x1(k + 1) = A11x1(k) + B1u(k)
x1 = A11x1 + B1u
y = C1x + Du
y(k) = C1x1(k) + Du(k)
The following error formula is relevant:
(continuous)
[C(jωI – A)–1] – [C1(jωI – A1)–1(B1 + D)]
∞
≤ 2[σnsr + 1 + σnsr + 2 + ... + σns]
(discrete)
[C(ejωI – A)–1B + D] – [C1(ejωI – A1)–1B1 + D]
∞
≤ 2[σnsr + 1 + σnsr + 2 + ... + σns]
It is this error bound which is the basis of the determination of the order
of the reduced system when the keyword boundis specified. If the error
bound sought is smaller than 2σns, then no reduction is possible which is
consistent with the error bound. If it is larger than 2trΣ, then the constant
transfer function matrix D achieves the bound.
For continuous systems, the actual approximation error depends on
frequency, but is always zero at ω = ∞. In practice it is often greatest at
ω = 0; if the reduction of state dimension is 1, the error bound is exact, with
the maximum error occurring at DC. The bound also is exact in the special
case of a single-input, single-output transfer function which has poles and
zero alternating along the negative real axis. It is far from exact when the
poles and zeros approximately alternate along the imaginary axis (with the
poles stable).
© National Instruments Corporation
2-9
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
The actual approximation error for discrete systems also depends on
frequency, and can be large at ω = 0. The error bound is almost never tight,
that is, the actual error magnitude as a function of ω almost never attains
the error bound, so that the bound can only be a guide to the selection of the
reduced system dimension.
In principle, the error bound formula for both continuous and discrete
systems can be improved (that is, made tighter or less likely to overestimate
the actual maximum error magnitude) when singular values occur with
multiplicity greater than one. However, because of errors arising in
calculation, it is safer to proceed conservatively (that is, work with the error
bound above) when using the error bound to select nsr, and examine the
actual error achieved. If this is smaller than required, a smaller dimension
for the reduced order system can be selected.
mreduce( )provides an alternative reduction procedure for a balanced
realization which achieves the same error bound, but which has zero error
at ω = 0. For continuous systems there is generally some error at ω = ∞,
because the D matrix is normally changed. (This means that normally the
approximation of a strictly proper system through mreduce( )will not be
strictly proper, in contrast to the situation with balmoore( ).) For discrete
systems the D matrix is also normally changed so that, for example, a
system which was strictly causal, or guaranteed to contain a delay (that is,
D = 0), will be approximated by a system SysRwithout this property.
The presentation of the Hankel singular values may suggest a logical
dimension for the reduced order system; thus if σk » σk + 1, it may be
sensible to choose nsr = k.
With mreduce( )and a continuous system, the reduced order system
SysRis internally balanced, with the grammian diag[σ1, σ2, ...,σnsr], so
that its Hankel Singular Values are a subset of those of the original system
Sys. Provided σnsr > σnsr + 1, SysRalso is controllable, observable, and
stable. This is not guaranteed if σnsr = σnsr + 1, so it is highly advisable to
avoid this situation. Refer to the balmoore( ) section for more on the
balmoore( )algorithm.
With mreduce( )and discrete systems, the reduced order system SysRis
not in general balanced (in contrast to balmoore( )), and its Hankel
singular values are not in general a subset of those of Sys. Provided
σnsr > σsrn + 1, the reduced order system SysRalso is controllable,
highly advisable to avoid this situation. For additional information about
the balmoore( )function, refer to the Xmath Help.
Xmath Model Reduction Module
2-10
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
Related Functions
balance(), truncate(), redschur(), mreduce()
truncate( )
SysR = truncate(Sys,nsr,{VD,VA})
The truncate( )function reduces a system Sysby retaining the first
nsrstates and throwing away the rest to form a system SysR.
If for Sysone has,
A11 A12
A21 A22
B1
B2
A =
B =
C =
C1 C2
the reduced order system (in both continuous-time and discrete-time cases)
is defined by A11, B1, C1, and D. If Sysis balanced, then SysRis an
well be used after an initial application of balmoore( )to further reduce
a system should a larger approximation error be tolerable. Alternatively, it
may be used after an initial application of balance( )or redschur( ).
If Syswas calculated from redschur( )and VA,VDwere posed as
arguments, then SysRis calculated as in redschur( )(refer to the
redschur( ) section).
truncate( )should be contrasted with mreduce( ), which achieves a
reduction through a singular perturbation calculation. If Sysis balanced,
the same error bound formulas apply (though not necessarily the same
errors), truncate( )always ensures exact matching at s = ∞ (in the
continuous-time case), or exacting matching of the first impulse response
coefficient D (in the discrete-time case), while mreduce( )ensures
matching of DC gains for Sysand SysRin both the continuous-time and
discrete-time case. For a additional information about the truncate( )
function, refer to the Xmath Help.
Related Functions
balance(), balmoore(), redschur(), mreduce()
© National Instruments Corporation
2-11
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
redschur( )
[SysR,HSV,slbig,srbig,VD,VA] = redschur(Sys,{nsr,bound})
The redschur( )function uses a Schur method (from Safonov and
Chiang) to calculate a reduced version of a continuous or discrete system
without balancing.
Algorithm
The objective of redschur( )is the same as that of balmoore( )when
the latter is being used to reduce a system; this means that if the same Sys
and nsrare used for both algorithms, the reduced order system should have
the same transfer function matrix. However, in contrast to balmoore( ),
redschur( )do not initially transform Systo an internally balanced
realization and then truncate it; nor is SysRin balanced form. The fact that
there is no balancing offers numerical advantages, especially if Sysis
nearly nonminimal.
Sysshould be stable, and this is checked by the algorithm. In contrast to
balmoore( ), minimality of Sys(that is, controllability and
observability) is not required.
If the Hankel singular values of Sysare ordered as σ1 ≥ σ2 ≥ ... ≥ σns ≥ 0,
then those of SysRin the continuous-time case are σ1 ≥ σ2 ≥ ... ≥ σnsr > 0.
A restriction of the algorithm is that σnsr > σnsr + 1 is required for both
continuous-time and discrete-time cases. Under this restriction, SysRis
guaranteed to be stable and minimal.
The algorithms depend on the same algorithm, apart from the calculation
of the controllability and observability grammians Wc and Wo of the
original system. These are obtained as follows:
(continuous)
(discrete)
Wc A′ + AWc = –BB′
WoA + A′Wo = C′C
Wc – AWcA′ = BB′
Wo – A′WoA = C′C
The maximum order permitted is the number of nonzero eigenvalues of
Xmath Model Reduction Module
2-12
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
Next, Schur decompositions of WcWo are formed with the eigenvalues of
WcWo in ascending and descending order. These eigenvalues are the square
of the Hankel singular values of Sys, and if Sysis nonminimal, some can
be zero.
V′AWcWoVA = Sasc
V′DWcWoVD = Sdes
The matrices VA, VD are orthogonal and Sasc, Sdes are upper triangular. Next,
submatrices are obtained as follows:
0
Insr
Vlbig = VA
Vrbig = VD
Insr
0
and then a singular value decomposition is found:
UebigSebigVebig = V′lbigVrbig
From these quantities, the transformation matrices used for calculating
SysRare defined:
1 ⁄ 2
Slbig = VlbigUebig
S
S
ebig
1 ⁄ 2
ebig
Srbig = VrbigVebig
and the reduced order system is:
AR = S′lbigASrbig
BR = S′lbig
B
AR = CSrbig
D
An error bound is available. In the continuous-time case it is as follows. Let
G( jω) and GR( jω) be the transfer function matrices of Sysand SysR,
respectively.
For the continuous case:
G( jω) – GR( jω) ≤ 2(σnsr + 1 + σnsr + 2 + ... + σns)
∞
© National Instruments Corporation
2-13
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
For the discrete-time case:
G(ejω) – GR(ejω) ∞ ≤ 2(σnsr + 1 + σnsr + 2 + ... + σns)
When {bound}is specified, the error bound just enunciated is used to
choose the number of states in SysRso that the bound is satisfied and nsr
is as small as possible. If the desired error bound is smaller than 2σns,
no reduction is made.
In the continuous-time case, the error depends on frequency, but is always
zero at ω = ∞. If the reduction in dimension is 1, or the system Sysis
single-input, single-output, with alternating poles and zeros on the real
axis, the bound is tight. It is far from tight when the poles and zeros
approximately alternate along the jω-axis. It is not normally tight in the
discrete-time case, and for both continuous-time and discrete-time cases,
it is not tight if there are repeated singular values.
The presentation of the Hankel singular values may suggest a logical
dimension for the reduced order system; thus if σk » σk + 1, it may be
sensible to choose nsr = k.
Related Functions
ophank(), balmoore()
ophank( )
[SysR,SysU,HSV] = ophank(Sys,{nsr,onepass})
The ophank( )function calculates an optimal Hankel norm reduction
of Sys.
Restriction
This function has the following restriction:
•
Only continuous systems are accepted; for discrete systems use
makecontinuous( )before calling bst( ), then discretize the
result.
Sys=ophank(makecontinuous(SysD));
SysD=discretize(Sys);
Xmath Model Reduction Module
2-14
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
Algorithm
The algorithm does the following. The system Sysand the reduced order
system SysRare stable; the system SysUhas all its poles in Re[s] > 0. If
the transfer function matrices are G(s), Gr(s) and Gu(s) then:
•
•
Gr(s) is a stable approximation of G(s).
and optimal in a certain sense.
Of course, the algorithm works with state-space descriptions; that of G(s)
can be minimal, while that of Gr(s) cannot be.
These statements are explained in the Behaviors section. If onepassis
specified, reduction is calculated in one pass. If onepassis not called or is
set to 0 {onepass=0}, reduction is calculated in (number of states of
Sys - nsr) passes. There seems to be no general rule to suggest which
setting produces the more accurate approximation Gr. Therefore, if
accuracy of approximation for a given order is critical, both should be tried.
As noted previously, if an approximation involving an unstable system is
desired, the default {onepass=1}is specified.
Behaviors
The following explanation deals first with the keyword {onepass}.
Suppose that σ1, σ2,..., σns are the Hankel Singular values of S, which has
transfer function matrix G(s). Suppose that the singular values are ordered
so that:
σ1 = σ2 = ... = σn > σn + 1... = σn + 1... = σn > σn + 1...
1
1
1
2
2
> σn
= σn
m – 1 + 1
m – 1 + 2
m
Thus, there are n1 equal values, followed by n2 – n1 equal values, followed
by n3 – n2 equal values, and so forth.
The order nsrof Gr(s) cannot be arbitrary when there are equal Hankel
singular values. In fact, the orders shown in Table 2-1 for the strictly stable
Gr (all poles in Re[s]<0) and strictly unstable Gu (all poles Re[s]>0) are
possible (and there are no other possibilities).
© National Instruments Corporation
2-15
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
Table 2-1. Orders of G
Number of
Eliminated States
(Retaining Gu)
Number of
Eliminated States
(Discarding Gu)
Order of
Gr nsr
Order of
Gu nsu
0
ns – n1
ns – n2
ns – n3
⇓
n1
n2 – n1
n3 – n2
⇓
ns
ns – n1
ns – n2
⇓
n1
n2
⇓
nm – 1
0
ns – nm – 1
ns – nm – 1
By abuse of notation, when we say that G is reduced to a certain order, this
corresponds to the order of Gr(s) alone; the unstable part of Gu(s) of the
approximation is most frequently thrown away. The number of eliminated
states (retaining Gu) refers to:
(# of states in G) – (# of states in Gr) – (# of states in Gu)
This number is always the multiplicity of a Hankel singular value. Thus,
when the order of Gr is ni – 1 the number of eliminated states is ni – ni – 1 or
the multiplicity of σn + 1 = σni.
i – 1
For each order ni – 1 of Gr(s), it is possible to find Gr and Gu so that:
G( jω) – Gr( jω) – Gu( jω) ≤ σn
∞
i
(Choosing i = 1 causes Gr to be of order zero; identify n0 = 0.) Actually,
among all “approximations” of G(s) with stable part restricted to having
degree ni – 1 and with no restriction on the degree of the unstable part, one
can never obtain a lower bound on the approximation error than σn ; in the
i
scalar or SISO G(s) case, the Gr(s) which achieves the previous bound is
unique, while in the matrix or MIMO G(s) case, the Gr(s) which achieves
the previous bound may not be unique [Glo84]. The algorithm we use to
find Gr(s) and Gu(s) however allows no user choice, and delivers a single
pair of transfer function matrices.
The transfer function matrix Gr( jω) alone can be regarded as a stable
chosen, (and the algorithm ensures that it is), then:
G( jω) – Gr( jω) ≤ σn + σn + ... + σns
(2-3)
∞
i
i + 1
Xmath Model Reduction Module
2-16
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
Thus, the penalty for not being allowed to include Gu in the approximation
is an increase in the error bound, by σn + 1 + ... +σns. A number of
i
theoretical developments hinge on bounding the Hankel singular values of
Gr(s) and Gu(–s) in terms of those of G(s). With Gr(s) of order ni – 1, there
holds:
σk(Gr ) ≤ σk(G)k = 1, 2, …, ni – 1
The transfer function matrix Gu(s), being unstable, does not have Hankel
singular values; however, Gu(–s) (which is stable) does have Hankel
singular values. They satisfy:
σk[Gu(–s)] ≤ σn + k(G)
i
In most cases, the Hankel singular values of G(s) are distinct. If,
accordingly,
∞
then Gr has degree (i – 1), Gu has degree ns – i and
G – Gr = σi + σi + 1 + ... + σns
(2-4)
∞
Observe that the bound (Equation 2-3 or Equation 2-4), which is not
necessarily obtained, is one half that applying for both balanced truncation
(as implemented by balmoore( )or, effectively, by redschur( )); it
also is one half that obtained when applying mreduceto a balanced
realization. In general, the D matrices of G and Gr are different, that is,
G(∞) ≠ Gr(∞) (in contrast to balmoore( )and redschur( )). Similarly,
G(0) ≠ Gr(0) in general (in contrast to the result when mreduceis applied
to a balanced realization). The price paid for obtaining a smaller error
bound overall through Hankel norm reduction is that one no longer
(normally) secures zero error at ω = ∞ or ω = 0.
Two special cases should be noted. If nsr= 0 then Gr(s) is a constant only.
This constant can be added onto Gu(s), so that G(s) is then being
approximated by a totally unstable transfer function matrix, with error σ1;
this type of approximation is known as Nehari approximation. The second
special case arises when nsr= nm – 1 (or NS – 1 if the smallest Hankel
singular value has multiplicity 1). In this case, Gu(s) becomes a constant,
© National Instruments Corporation
2-17
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
being approximated by a stable Gr(s) with the actual error (as opposed to
just the error bound) satisfying:
G(s) – Gr(s) = σns
∞
Note Gr is optimal, that is, there is no other Gr achieving a lower bound.
Onepass Algorithm
The first steps of the algorithm are to obtain the Hankel singular values of
G(s) (by using hankelsv( )) and identify their multiplicities. (Stability of
G(s) is checked in this process.) If the user has specified nsrand this does
not coincide with one of 0,n1,n2, ... an error message is obtained; generally,
all the σi are different, so the occurrence of error messages will be rare.
The next step of the algorithm is to calculate the sum G(s) = Gr(s) + Gu(s),
following [SCL90]. (A separate function ophred( )is called for this
purpose.) The controllability and observability grammians P and Q are
found in the usual way.
AP + PA′ = –BB′
QA + A′Q = –C′C
and then a singular value decomposition is obtained of the
matrixQP – σ2n I:
i
V1′
SB 0
= QP – σ2n I
U1 U2
i
V2′
0 0
There are precisely ni – ni – 1 zero singular values, this being the multiplicity
of σn . Next, the following definitions are made:
i
U1′
A11 A12
A21 A22
=
(σ2n A′ + QAP)(V1V2)
i
U2′
U1′
B1
=
QB
U2′
B2
[C1 C2] = CP[V1 V2]
Xmath Model Reduction Module
2-18
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
and finally:
#
–1
˜
A = SB (A11 – A–12A22A21)
#
–1
˜
B = SB (B1 – A–12A22B2)
#
#
˜
C = C1 – C2A22A21
#
˜
D = D – C2A22B2
˜
These four matrices are the constituents of the system matrix of G(s),
where:
˜
G(s) = Gr(s) + Gu(s)
Digression:
This choice is related to the ideas of [Glo84] in the following way;
˜
in [Glo84], the complete set is identified of G(s) satisfying
˜
G(jω) – G(jω) = σn
∞
i
˜
with G having a stable part of order ni – 1. The set is parameterized in
terms of a stable transfer function matrix K(s) which has to satisfy
C2 + K(s)B′2 = 0
I – K′(–jω)K( jω) ≤ 0 for all ω
with C2, B2 being two matrices appearing in the course of the algorithm
of [Glo84], and enjoying the property C′2C2 = B2B2′ . The particular
choice
K(s) = –C2(C′2C2)#B2
in the algorithm of [Glo84] and flagged in corollary 7.3 of [Glo84] is
equivalent to the previous construction, in the sense of yielding the
˜
same , though the actual formulas used here and in [Glo84] for the
Gs
construction procedure are quite different. In a number of situations,
including the case of scalar (SISO)G(s), this is the only choice.
The next step of the algorithm is to call stable( ), to separate G(s) into
˜
˜
its stable and unstable parts, call them
and
, stable( )will
Gu(s)
G(s)
˜
˜
always assign the matrix to
, and the final step of the algorithm is
Gr(s)
D
© National Instruments Corporation
2-19
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
˜
to choose the D matrix of Gr(s), by splitting D between Gr(s) and Gu(s).
This is done by using a separate function ophiter( ). Suppose Gu(s) is
the unstable output of stable( ), and let K(s) = Gu(–s). By applying the
multipass Hankel reduction algorithm, described further below, K(s) is
reduced to the constant K0 (the approximation), which satisfies,
K(s) – K0 ≤ σ1(K) + ... + σ σn – n (K)
∞
s
i
≤ σn + 1(G) + ... + σn (G)
i
s
that is, if it is larger than,
ns
Gu(–s) – K0
≤
σk(G)
∞
∑
k = ni + 1
then one chooses:
˜
Gr = Gr + K0
˜
Gu = Gu + K0
This ensures satisfaction of the error bound for G – Gr given previously,
because:
˜
˜
˜
G – Gr
=
G – Gr – Gu + (Gu – K0)
∞
∞
˜
˜
= G – Gr – Gu ∞ + K – K0
∞
≤ σn (G) + σn + 1(G) + ... + σn (G)
i
i
s
Multipass Algorithm
We now explain the multipass algorithm. For simplicity in first explaining
the idea, suppose that the Hankel singular values at every stage or pass are
distinct.
1. Find a stable order ns – 1 approximation Gn – 1(s) of G(s) with:
G( jω) – Gns – 1jω = σns(G)
∞
(This can be achieved by the algorithm already given, and there is no
unstable part of the approximation.)
Xmath Model Reduction Module
2-20
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
2. Find a stable order ns – 2 approximation Gns – 2 of Gns – 1(s), with
Gns – 1( jω) – Gns – 2( jω) = σns – 1(Gns – 1
)
∞
.
.
.
3. (Step ns–nr):
Find a stable order nsr approximation of Gnsr + 1
,
with Gnsr + 1( jω) – Gnsr( jω) = σnsr + 1(Gnsr + 1
)
∞
Then, because σi(Gns – 1) ≤ σi(G) for i < ns,
σi(Gns – 2) ≤ σi(Gns – 1
)
for i ≤ s – i, ..., this being a property of the algorithm, there follows:
G( jω) – Gnsr( jω) ≤ σnsr + 1(Gnsr + 1) + ... + σns(G)
ns
≤
σi(G)
∑
i = nsr + 1
The only difference that arises when singular values have multiplicity in
excess of 1 is that the degree reduction at a given step is greater. Thus, if
σns(G) has multiplicity k, then G(s) is approximated by Gns – k(s) of degree
ns – k.
No separate optimization of the D matrix of Gnsr is required. The
approximation error bound is the same as for the first algorithm. The actual
approximation error may be different. Notice that this second algorithm
does not calculate an unstable Gu(s) such that
G(jω) – Gnsr(jω) – Gu(jω) = σnsr + 1
∞
Discrete-Time Systems
No special algorithm is presented for discrete-time systems. Any stable
discrete-time transfer-function matrix G(z) can be used to define a stable
continuous-time transfer-function matrix by a bilinear transformation, thus
α + s
⎛
⎝
⎞
⎠
-----------
H(s) = G
α – s
when α is a positive constant.
© National Instruments Corporation
2-21
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
We use sysZ to denote G(z) and define:
bilinsys=makepoly([-1,a]/makepoly([1,a])
as the mapping from the z-domain to the s-domain. The specification is
reversed because this function uses backward polynomial rotation. Hankel
norm reduction is then applied to H(s), to generate, a stable reduced order
approximation Hr(s) and unstable Hu(s) such that:
H – Hr – Hu = σi
H – Hr = σi + σn + 1 + ... + σns
i
Here, the sni are the Hankel singular values of both G(z) and H(s); they are
the same:
z – 1
z + 1
⎛
⎞
⎠
-----------
Gr(z) = Hr α
⎝
z – 1
z + 1
⎛
⎞
⎠
-----------
α
Gu(z) = Hu
⎝
We then implement the s-domain equivalent with:
sysS=subsys(sysZ,bilinsys)
There is no simple rule for choosing α; the choice α = 1 is probably as good
as any. The orders of Gr and Gu are the same as those of Hr and Hu,
respectively. The error formulas are as follows:
G(ejω) – Gr(ejω) – Gu(ejω) = σn
∞
i
G(ejω) – Gr(ejω) ∞ ≤ σn + σn + 1 + ...σns
i
i
Impulse Response Error
If Gr is determined by the first (single-pass) algorithm, the impulse
response error (for t > 0) between the impulse responses of G and Gr can
be bounded. As shown in Corollary 9.9 of [Glo84], if Gr is of degree i – 1
and the multiplicity of the ith larger singular value σi of G is r, then:
σj[G – Gr] ≤ σiG for j = 1, 2, ..., 2i – 2 + r
≤ σj – i + 1(G) for j = 2i – 1 + r, ...,ns + i – 1
Xmath Model Reduction Module
2-22
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 2
Additive Error Reduction
It follows by a result of [BoD87] that the impulse response error for t > 0
satisfies:
ns
g(t) – gr(t) 1 ≤ 2 (2i – 2 + r)σi(G) +
σj(G)
∑
i + r
Evidently, Hankel norm approximation ensures some form of
approximation of the impulse response too.
Unstable System Approximation
A transfer function G(s) with stable and unstable poles can be reduced by
applying stable( )to separate G(s) into a stable and unstable part. The
former is reduced and then the unstable part can be added back on. For
additional information about the ophank( )function, refer to the Xmath
Help.
Related Functions
stable(), redschur(), balmoore()
© National Instruments Corporation
2-23
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
3
Multiplicative Error Reduction
This chapter describes multiplicative error reduction presenting
two reasons to consider multiplicative rather than additive error reduction,
one general and one specific.
Selecting Multiplicative Error Reduction
The general reason to use multiplicative error reduction is that many
specifications are given using decibels; ±1 db corresponds to a
multiplicative error of about 12%. Specifications regarding phase shift also
can be regarded as multiplicative error statements: ±0.05 radians of phase
shift is like 5% multiplicative error also.
The more specific reason arises in considering the problem of plant
approximation, with a high order (possibly very high order) plant being
initially prescribed, with no controller having been designed, and with a
requirement to provide a simpler model of the plant, possibly to allow
controller design. Consider the arrangement of Figure 3-1, controller C(s)
designed for G(s)j with G(s) = (I + Δ)G(s).
Δ
ˆ
C
G
Figure 3-1. Controller C(s) Designed for Multiplicative Error Reduction
ˆ
The full order plant is G = (I + Δ)G, and the reduced order model is G.
–1
ˆ ˆ
Since
, this means that Δ is the multiplicative error.
(G – G)G = Δ
Another way one could measure the multiplicative error would be as
ˆ ˆ
(G – G)G
product gives two more possibilities again.
The following multiplicative robustness result can be found in [Vid85].
© National Instruments Corporation
3-1
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
Multiplicative Robustness Result
ˆ
ˆ ˆ –1
Suppose C stabilizes G, that Δ = (G – G)G has no jω-axis poles, and
ˆ
that G has the same number of poles in Re[s] ≥ 0 as . If for all ω,
G
–1
ˆ
ˆ
Δ( jω) G( jω)C( jω)[I + G( jω)C( jω)] < 1
(3-1)
then C stabilizes G.
This result indicates that if a controller C is designed to stabilize a nominal
ˆ
or reduced order model G , satisfaction of Equation 3-1 ensures that the
controller also will stabilize the true plant G.
In reducing a model of the plant, there will be concern not just to have this
type of stability property, but also concern to have as little error as possible
ˆ
between the designed system (based on ) and the true system (based
G
on G). Extrapolation of the stability result then suggests that the goal
should be not just to have Equation 3-1, but to minimize the quantity on the
left side of Equation 3-1, or its greatest value:
–1
ˆ
ˆ
max{ Δ( jω) G( jω)C( jω)[I + G( jω)C( jω)]
}
ω
However, there are difficulties. The principal one is that if we are reducing
the plant without knowledge of the controller, we cannot calculate the
measure because we do not know C(jω). Nevertheless, one could presume
that, for a well designed system, GC(I + GC)–1 will be close to I over the
operating bandwidth of the system, and have smaller norm than 1 (tending
to zero as ω→∞ in fact) outside the operating bandwidth of the system.
This suggests that in the absence of knowledge of C, one should carry out
multiplicative approximation by seeking to minimize:
ˆ
ˆ
max Δ(jω)
=
Δ(jω)
∞
ω
This is the prime rationale for (unweighted) multiplicative reduction of a
plant.
Two other points should be noted. First, because frequencies well beyond
–1 will be small, it is in a sense,
ˆ
ˆ
the closed-loop bandwidth,
GC(I + GC)
of maxω Δ(jω) as the index is convenient, because it removes a
requirement to make assumptions about the controller, but at the same time
it does not allow Δ(jω) to be made even smaller in the closed-loop
Xmath Model Reduction Module
3-2
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
bandwidth at the expense of being larger outside this bandwidth, which
would be preferable.
ˆ ˆ –1
(G – G)G
Second, the previously used multiplicative error is
. In the
ˆ ˆ –1
algorithms that follow, the error
δ = (G – G)G
appears. It is easy to
check that:
Δ(jω)
1 – Δ(jω)
∞
-------------------------------
δ(jω)
≤
∞
∞
∞
and
δ(jω)
∞
------------------------------
Δ(jω)
≤
1 – δ(jω)
∞
This means that if either bound is small, so is the other, with the bounds
approximately equal.
Two algorithms for multiplicative reduction are presented: bst( ),
a mnemonic for balanced stochastic truncation, and mulhank( ).
Roughly, they relate to one another in the same way that redschur( )
and ophank( )relate, that is, one focuses on balanced realization
truncation and the other on Hankel norm approximation. Some of the
similarities and differences are as follows:
•
When the errors are small, the error bound formula for bst( )is
about one half of that for bst( ).
•
With bst( ), the actual multiplicative error as a function of frequency
goes to zero as ω→∞ (or, after using an optional transformation given
in the algorithm description, to zero as ω→ 0); with mulhank( ), the
error tends to be more uniform as a function of frequency.
• bst( )can handle nonsquare reduction, while mulhank( )cannot.
•
Both algorithms are restricted to stable G(s); both preserve right half
plane zeros, that is, these zeros are copied into the reduced order
object; both have difficulties with jω-axis zeros of G(s), but these
difficulties are not insuperable.
bst( )
[SysR,HSV] = bst(Sys,{nsr,left,right,bound,method})
The bst( )function calculates a balanced stochastic truncation of Sysfor
the multiplicative case.
© National Instruments Corporation
3-3
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
The objective of the algorithm is to approximate a high-order stable transfer
function matrix G(s) by a lower-order Gr(s) with either inv(g)(g-gr)or
(g-gr)inv(g)minimized, under the condition that Gr is stable and of the
prescribed order.
Restrictions
This function has the following restrictions:
•
•
•
The user must ensure that the input system is stable and nonsingular at
s = infinity.
The algorithm may be problematic if the input system has a zero on the
jω-axis.
Only continuous systems are accepted; for discrete systems use
makecontinuous( )before calling bst( ), then discretize the
result.
Sys=bst(makecontinuous(SysD));
SysD=discretize(Sys);
Algorithm
The modifications described in this section allow you to circumvent the
previous restrictions.
The objective of the algorithm is to approximate a high order stable transfer
function matrix G(s) by a lower order G (s) with, in the square G(s) case,
r
(G – Gr)G–1
G–1(G – Gr)
either
∞ or
∞ (approximately) minimized,
under the constraint that Gr is stable and of prescribed order nsr. In case
G is not square but has full row rank, the algorithm seeks to minimize:
(G – Gr) (GG*)–1(G – Gr)
*
∞
Recall that X*(s) = X′(–s) so that when s = jω,
X*( jω) = X*( jω)
When G is not square but has full column rank, the algorithm seeks to
minimize:
(G – Gr)(G*G)–1(G – Gr)
*
∞
Xmath Model Reduction Module
3-4
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
These cases are secured with the keywords rightand left, respectively.
If the wrong option is requested for a nonsquare G(s), an error message will
result.
The algorithm has the property that right half plane zeros of G(s) remain as
right half plane zeros of G (s). This means that if G(s) has order nsr with n+
r
zeros in Re[s] > 0, Gr(s) must have degree at least n+, else, given that it has
n+ zeros in Re[s] > 0 it would not be proper, [Gre88].
The conceptual basis of the algorithm can best be grasped by considering
the case of scalar G(s) of degree n. Then one can form a minimum phase,
stable W(s) with |W(jω)|2 = |G(jω)|2 and then an all-pass function (the phase
function) W–1(–s) G(s). This all pass function has a mixture of stable and
unstable poles, and it encodes the phase of G(jω). Its stable part has n
Hankel singular values σi with σi ≤ 1, and the number of σi equal to 1 is the
same as the number of zeros of G(s) in Re[s] > 0. State-variable realizations
of W,G and the stable part of W–1(–s)G(s) can be connected in a nice way,
and when the stable part of W–1(–s)G(s) has a balanced realization, we say
that the realization of G is stochastically balanced. Truncating the balanced
realization of the stable part of W–1(–s)G(s) induces a corresponding
truncation in the realization of G(s), and the truncated realization defines an
approximation of G. Further, a good approximation of a transfer function
encoding the phase of G somehow ensures a good approximation, albeit in
a multiplicative sense, of G itself.
Algorithm with the Keywords right and left
The following description of the algorithm with the keyword rightis
based on ideas of [GrA86] developed in [SaC88]. The procedure is almost
the same when leftis specified, except the transpose of G(s) is used; the
algorithm finds an approximation in the same manner as for right, but
transposes the approximation to yield the desired Gr(s).
1. The algorithm checks
•
•
•
That the system is state-space, continuous, and stable
That a correct option has been specified if the plant is nonsquare
That D is nonsingular; if the plant is nonsquare, DD´ must be
nonsingular
© National Instruments Corporation
3-5
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
2. With G(s) = D + C(sI – A)–1B and stable, with DD´ nonsingular and
G(jω) G'(–jω) nonsingular for all ω, part of a state variable realization
of a minimum phase stable W(s) is determined such that
W´(–s)W(s) = G(s)G´(–s) with
W(s) = DW + CW(sI – Aw)–1BW
The state variable matrices in W(s) are obtained as follows. The
controllability grammian P associated with G(s) is first found from
AP + PA´ + BB´=0 then AW = A, BW = PC´ + BD´.
When G(s) is square, the algorithm checks to see if there is a zero or
singularity of G(s) close to the jω-axis (the zeros are given by the
eigenvalues of A – BD–1C and are computed reliably with the aid of
schur( )). If one is found, you are warned that results may be
unreliable. Next, a stabilizing solution Q is found for the following
Riccati equation:
QA + A′Q + (C – B′WQ)′(DD′)–1(C – B′WQ) = 0
The singriccati( )function is used; failure of the nonsingularity
condition on G(jω)G´(–jω) will normally result in an error message
that no stabilizing solution exists. To obtain the best numerical results,
singriccati( )is invoked with the keyword {method="schur"}.
Although DW, CW are not needed for the remainder of the algorithm,
they are simply determined in the square case by
DW = D′ CW = D–1(C – BW′ Q)
with minor modification in the nonsquare case. The real point of the
algorithm is to compute P and Q; the matrix Q satisfies (square or
nonsquare case).
QA + A′Q + C′WCW = 0
P, Q are the controllability and observability grammians of the transfer
function CW(sI – A)–1B. This transfer function matrix, it turns out, is
the strictly proper, stable part of θ(s) = W–T(–s)G(s), which obeys the
matrix all-pass property θ(s)θ´(–s) = I, and is the phase matrix
associated with G(s).
3. Compute ordered Schur decompositions of PQ, with the eigenvalues
of PQ is ascending and descending order. Obtain the phase matrix
Hankel singular values, that is, the Hankel singular values of the
Xmath Model Reduction Module
3-6
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
strictly proper stable part of θ(s), as the square roots of the eigenvalues
of PQ. Call these quantities νi. The Schur decompositions are,
V′APQVA = Sasc
V′DPQVD = Sdes
where VA, VD are orthogonal and Sasc, Sdes are upper triangular.
4. Define submatrices as follows, assuming the dimension of the reduced
order system nsris known:
0
Insr
Vlbig = VA
Vrbig = VD
Insr
0
Determine a singular value decomposition,
UebigSebigVebig = V′lbigVrbig
and then define transformation matrices:
–1 ⁄ 2
Slbig = VlbigUebig
S
S
ebig
–1 ⁄ 2
ebig
Srbig = VrbigVebig
The reduced order system Gr is:
AR = S′lbigASrbig
BR = S′lbig
B
AR = CSrbig
DR = D
where step 4 is identical with that used in redschur( ), except
the matrices P, Q which determine VA, VD and so forth, are the
controllability and observability grammians of CW(sI – A)–1B rather
than of C(sI – A)–1B, the controllability grammian of G(s) and the
observability grammian of W(s).
The error formula [WaS90] is:
vi
G–1(G – Gr) ∞ ≤ 2
------------
(3-2)
∑
1 – vi
All νi obey νi ≤ 1. One can only eliminate νi where νi < 1. Hence, if nsris
chosen so that νnsr + 1 = 1, the algorithm produces an error message. The
algorithm also checks that nsrdoes not exceed the dimension of a minimal
© National Instruments Corporation
3-7
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
state-variable representation of G. In this case, the user is effectively asking
for Gr = G. When the phase matrix has repeated Hankel singular values,
they must all be included or all excluded from the model, that is,
ν
nsr = νnsr + 1 is not permitted; the algorithm checks for this.
The number of νi equal to 1 is the number of zeros in Re[s]>0 of G(s), and
as mentioned already, these zeros remain as zeros of Gr(s).
If erroris specified, then the error bound formula (Equation 3-2) in
conjunction with the νi values from step 3 is used to define nsrfor step 4.
For nonsquare G with more columns than rows, the error formula is:
ns
(G – Gr) (G*G)–1(G – Gr) 1 ⁄ 2 ≤ 2
------------
vi
*
∞
∑
1 – vi
i = nsr + 1
If the user is presented with the νi, the error formula provides a basis for
intelligently choosing nsr. However, the error bound is not guaranteed to
be tight, except when nsr = ns – 1.
Securing Zero Error at DC
The error G–1(G – Gr) as a function of frequency is always zero at ω = ∞.
When the algorithm is being used to approximate a high order plant by a
low order plant, it may be preferable to secure zero error at ω = 0. A method
for doing this is discussed in [GrA90]; for our purposes:
1. We need a bilinear transformation of sys = 1/z. Given G(s) we generate
H(s) through:
bilinsys=makepoly([b3,b4]/makepoly([b1,b2])
sys=subsys(sys,bilinsys)
2. Reduce with the previous algorithm:
[sr,nsr,hsv] = bst(sys)
3. Use the bilinear transformation s = 1/z again:
[sr1,nsr1] = bilinear(sr,nsr,[0,1,1,0])
The νi are the same for G(s) and H(s) = G(s–1). The error bound formula is
the same; H is stable and H(jω)H'(–jω) of full rank for all ω including
ω = ∞ if and only if G has the same property; right half plane zeros of G are
still preserved by the algorithm. The error G–1(G – Gr), though now zero at
Xmath Model Reduction Module
3-8
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
Hankel Singular Values of Phase Matrix of Gr
The νi, i = 1,2,...,ns have been termed above the Hankel singular values of
the phase matrix associated with G. The corresponding quantities for Gr are
νi, i = 1,..., nsr.
Further Error Bounds
The introduction to this chapter emphasized the importance of the error
measures
(G – Gr)G–r 1 ∞ or G–r 1(G – Gr)
(G – Gr)G–1
for plant reduction, as opposed to
or G–1(G – Gr)
∞
∞
The BST algorithm ensures that in addition to (Equation 3-2), there holds
[WaS90a].
ns
vi
G–r 1(G – Gr) ∞ ≤ 2
------------
∑
1 – vi
i = nsr + 1
which also means that for a scalar system,
ns
⎛
⎞
Gr
-----
vi
⎜
⎟
20log
≤ 8.69 2
------------ dB
10 G
⎜
⎝
∑
⎟
1 – vi
⎠
i = nsr + 1
and, if the bound is small:
ns
vi
------------
phase(G) – phase(Gr) ≤
radians
∑
1 – vi
i = nsr + 1
Reduction of Minimum Phase, Unstable G
For square minimum phase but not necessarily stable G, it also is possible
to use this algorithm (with minor modification) to try to minimize (for Gr
of a certain order) the error bound
(G – Gr)G–r 1
∞
© National Instruments Corporation
3-9
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
which also can be relevant in finding a reduced order model of a plant.
The procedure requires G again to be nonsingular at ω = ∞, and to have no
jω-axis poles. It is as follows:
1. Form H = G–1. If G is described by state-variable matrices A, B, C, D,
then H is described by A – BD–1C, BD–1, –D–1C, D–1. H is square,
stable, and of full rank on the jω-axis.
2. Form Hr of the desired order to minimize approximately:
H–1(H – Hr)
∞
3. Set Gr = H–1
.
r
Observe that
H–1(H – Hr) = I – H–1Hr
= I – GG–r 1
= (Gr – G)G–r 1
The reduced order Gr will have the same poles in Re[s] > 0 as G, and
be minimum phase.
Imaginary Axis Zeros (Including Zeros at ∞)
We shall now explain how to handle the reduction of G(s) which has a rank
drop at s = ∞ or on the jω-axis. The key is to use a bilinear transformation,
[Saf87]. Consider the bilinear map defined by
z – a
s = ------------------
– bz + 1
s + a
z = --------------
bs + 1
˜
where 0 < a < b–1 and mapping G(s) into G(s) through:
s – a
– bs + 1
˜
⎛
⎝
⎞
⎠
-------------------
G(s) = G
s + a
bs + 1
˜
⎛
⎝
⎞
⎠
--------------
G(s) = G
Xmath Model Reduction Module
3-10
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
The values of G(s), as shown in Figure 3-2, along the jω-axis are
˜
the same as the values of
around a circle with diameter defined by
G(s)
[a – j0, b–1 + j0] on the positive real axis.
a
b–1
˜
G(s)
G(s)
values
values
˜
Figure 3-2. Bilinear Mapping from G(s) to (Gs) (Case 1)
˜
Also, the values of G(s), as shown in Figure 3-3, along the jω-axis are
the same as the values of G(s) around a circle with diameter defined by
[–b–1 + j0, –a + j0].
b–1
-a
˜
G(s)
G(s)
values
values
˜
Figure 3-3. Bilinear Mapping from G(s) to (Gs) (Case 2)
We can implement an arbitrary bilinear transform using the subsys( )
function, which substitutes a given transfer function for the s- or z-domain
operator.
s – a
– bs + 1
˜
⎛
⎝
⎞
⎠
------------------
To implement G(s) = G
use:
gtildesys=subsys(gsys,makep([-b,1]/makep([1,-a])
s + a
s + 1
˜
⎛
⎝
⎞
⎠
-----------
To implement G(s) = G
use:
gsys=subsys(gtildesys,makep([b,1]/makep([1,a])
Note The systems substituted in the previous calls to subsys invert the function
specification because these functions use backward polynomial rotation.
© National Instruments Corporation
3-11
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
Any zero (or rank reduction) on the jω-axis of G(s) becomes a zero (or rank
˜
reduction) in Re[s] > 0 of
, and if G(s) has a zero (or rank reduction)
G(s)
˜
at infinity, this is shifted to a zero (or rank reduction) of
at the point
G(s)
b–1, (in Re[s] > 0). If all poles of G(s) are inside the circle of diameter
˜
[–b–1 + j0, a + j0], all poles of
will be in Re[s] < 0, and if G(s) has no
G(s)
zero (or rank reduction) on this circle,
˜
will have no zero (or rank
G(s)
reduction) on the jω-axis, including ω = ∞.
If G(s) is nonsingular for almost all values of s, it will be nonsingular or
have no zero or rank reduction on the circle of diameter [–b–1 + j0, –a + j0]
for almost all choices of a,b. If a and b are chosen small enough, G(s) will
have all its poles inside this circle and no zero or rank reduction on it, while
˜
G(s) then will have all poles in Re[s] < 0 and no zero or rank reduction on
the jω-axis, including s = ∞. The steps of the algorithm, when G(s) has a
zero on the jω-axis or at s = ∞, are as follows:
s – a
– bs + 1
–1
˜
⎛
⎝
⎞
⎠
------------------
1. For small a,b with 0 < a < b , form G(s) = G
as shown for
gtildesys.
˜
˜
˜
2. Reduce G(s) to Gr(s), this being possible because G(s) is stable and
has full rank on s = jω, including ω = ∞.
s + a
bs + 1
˜
⎛
⎞
--------------
3. Form Gr(s) = Gr
as shown for gsys.
⎝
⎠
The error G–1(G – Gr) ∞ will be overbounded by the error
G (G – Gr) ∞, and Gr will contain the same zeros in Re[s] ≥ 0 as G.
–1
˜
˜
˜
If there is no zero (or rank reduction) of G(s) at the origin, one can take
a = 0 and b–1 = bandwidth over which a good approximation of G(s) is
needed, and at the very least b–1 sufficiently large that the poles of G(s)
lie in the circle of diameter [–b–1 + j0, –a + j0]. If there is a zero or rank
reduction at the origin, one can replace a = 0 by a = b. It is possible to take
b too small, or, if there is a zero at the origin, to take a too small. The user
will be presented with an error message that there is a jω-axis zero and/or
that the Riccati equation solution may be in error. The basic explanation is
˜
that as b → 0, and thus a → 0, the zeros of
approach those of G(s).
G
G(s)
˜
Thus, for sufficiently small b, one or more zeros of (s) may be identified
as lying on the imaginary axis. The remedy is to increase a and/or b above
the desirable values.
The procedure for handling jω-axis zeros or zeros at infinity will be
deficient if the number of such zeros is the same as the order of G(s)—for
example, if G(s) = 1/d(s), for some stable d(s). In this case, it is possible
Xmath Model Reduction Module
3-12
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
again with a bilinear transformation to secure multiplicative
approximations over a limited frequency band. Suppose that
s
˜
⎛
⎝
⎞
⎠
--------------
G(s) = G
εs + 1
˜
Create a system that corresponds to
with:
G(s)
gtildesys=subs(gsys,(makep([-eps,1])/makep([1,-]))
bilinsys=makep([eps,1])/makep([1,0])
sys=subsys(sys,bilinsys)
Under this transformation:
˜
•
Values of G(s) along the jω-axis correspond to values of G(s) around
a circle in the left half plane on diameter (–ε–1 + j0, 0).
˜
•
Values of G(s) along the jω-axis correspond to values of G(s) around
a circle in the right half plane on diameter (0, ε–1 + j0).
˜
Multiplicative approximation of G(s) (along the jω-axis) corresponds to
multiplicative approximation of G(s) around a circle in the right half plane,
touching the jω-axis at the origin. For those points on the jω-axis near the
circle, there will be good multiplicative approximation of G( jω). If it is
desired to have a good approximation of G(s) over an interval [–jΩ, jΩ],
then a choice such as ε–1 = 5 Ω or 10 Ω needs to be made. Reduction then
proceeds as follows:
˜
1. Form G(s).
˜
2. Reduce G(s) through bst( ).
˜
3. Form Gr(s) = –Gr(s ⁄ (1 – εs)) with:
gsys=subsys(gtildesys(gtildesys,
makep([-eps,-1])/makep[-1,-0]))
Notice that the number of zeros of G(s) in the circle of diameter
(0, ε–1 + j0)s
sets a lower bound on the degree of Gr(s)—for such zeros become right half
˜
plane zeros of G(s), and must be preserved by bst( ). Obviously, zeros at
s = ∞ are never in this circle, so a procedure for reducing G(s) = 1/d(s) is
available.
© National Instruments Corporation
3-13
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
There is one potential source of failure of the algorithm. Because G(s) is
˜
stable, G(s) certainly will be, as its poles will be in the left half plane circle
˜
on diameter (–ε = j0, 0). If Gr(s) acquires a pole outside this circle
(but still in the left half plane of course)—and this appears possible in
principle—Gr(s) will then acquire a pole in Re [s] > 0. Should this difficulty
be encountered, a smaller value of ε should be used.
Related Functions
redschur(), mulhank()
mulhank( )
[SysR,HSV] = mulhank(Sys,{nsr,left,right,bound,method})
The mulhank( )function calculates an optimal Hankel norm reduction of
Sysfor the multiplicative case.
Restrictions
This function has the following restrictions:
•
•
•
The user must ensure that the input system is stable and nonsingular at
s = infinity.
The algorithm may be problematic if the input system has a zero on the
jω-axis.
Only continuous systems are accepted; for discrete systems use
makecontinuous( )before calling mulhank( ), then discretize
the result.
Sys=mulhank(makecontinuous(SysD));
SysD=discretize(Sys);
Algorithm
The objective of the algorithm, like bst( ), is to approximate a high order
square stable transfer function matrix G(s) by a lower order Gr(s) with
either (G – Gr)G–1 or G–1(G – Gr) ∞ (approximately) minimized,
∞
under the constraint that Gr is stable and of prescribed order.
The algorithm has the property that right half plane zeros of G(s) are
zeros in Re[s] > 0, Gr(s) must have degree at least N+—else, given that it
has N+ zeros in Re[s] > 0 it would not be proper, [GrA89].
Xmath Model Reduction Module
3-14
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
The conceptual basis of the algorithm can best be grasped by considering
the case of scalar G(s) of degree n. Then one can form a minimum phase,
stable W(s) with |W(jω)|2 = |G(jω)|2 and then an all-pass function (the phase
function) W–1(–s) G(s). This all-pass function has a mixture of stable and
unstable poles, and it encodes the phase of G(jω). Its stable part has
n Hankel singular values σi with σi ≤ 1, and the number of σi equal to 1
is the same as the number of zeros of G(s) in Re[s]>0. State-variable
realizations of W,G and the stable part of W–1(–s)G(s) can be connected in
a nice way. The algorithm computes an additive Hankel norm reduction of
the stable part of W–1(–s)G(s) to cause a degree reduction equal to the
multiplicity of the smallest σi. The matrices defining the reduced order
object are then combined in a new way to define a multiplicative
approximation to G(s); as it turns out, there is a close connection between
additive reduction of the stable part of W–1(–s)G(s) and multiplicative
reduction of G(s). The reduction procedure then can be repeated on the new
phase function of the just found approximation to obtain a further reduction
again in G(s).
right and left
A description of the algorithm for the keyword rightfollows. It is based
on ideas of [Glo86] in part developed in [GrA86] and further developed
in [SaC88]. The procedure is almost the same when {left}is specified,
except the transpose of G(s) is used; the following algorithm finds an
approximation, then transposes it to yield the desired Gr(s).
1. The algorithm checks that G(s) is square, stable, and that the transfer
function is nonsingular at infinity.
2. With G(s) = D + C(sI–A)–1B square and stable, with D nonsingular
[rank(d)must equal number of rows in d] and G(jω) nonsingular for
all finite ω, this step determines a state variable realization of a
minimum phase stable W(s) such that,
W´(–s)W(s) = G(s)G´(–s)
with:
W(s) = Dw + Cw(sI–Aw)–1Bw
The various state variable matrices in W(s) are obtained as follows. The
controllability grammian P associated with G(s) is first found from
AP + PA´ + BB´ = 0, then:
Aw = ABw = PC´+BD´Dw = D´
close to the jω-axis. The zeros are determined by calculating the
© National Instruments Corporation
3-15
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
eigenvalues of A – B/D * C with the aid of schur( ). If any real part
of the eigenvaluesis less than eps, a warning is displayed.
Next, a stabilizing solution Q is found for the following Riccati
equation:
QA + A′Q + (C – B′wQ)′(DD′)–1(C – Bw′ Q) = 0
The function singriccati( )is used; failure of the nonsingularity
condition of G(jω) will normally result in an error message. To obtain
the best numerical results, singriccati( )is invoked with the
keyword method="schur".
The matrix Cw is given by C = D–1(C – B′wQ).
w
Notice that Q satisfies QA + A′Q + C′wCw = 0, so that P and Q are
the controllability and observability grammians of
F(s) = Cw(sI – A)–1B
This strictly proper, stable transfer function matrix is the strictly
proper, stable part (under additive decomposition) of
–T
θ(s)=W (–s)G(s), which obeys the matrix all pass property
θ(s)θ'(–s)=I. It is the phase matrix associated with G(s).
3. The Hankel singular values νi of F(s) = C (sI – A)–1B are
w
computed, by calling hankelsv( ). The value of nsris obtained if
not prespecified, either by prompting the user or by the error bound
formula ([GrA89], [Gre88], [Glo86]).
ns
vnsr + 1 ≤ G–1(G – Gr)
≤
(1 + vj) – 1
(3-3)
∞
∏
j = nsr + 1
(with νi ≥ νi + 1
≥
⋅⋅⋅ being assumed). If νk = νk + 1 = ... = νk + r for some
k, (that is, νk has multiplicity greater than unity), then νk appears once
only in the previous error bound formula. In other words, the number
of terms in the product is equal to the number of distinct νi less than
ν
nsr. There are restrictions on nsr. nsrcannot exceed the dimension
of a minimal realization of G(s); although νi ≥ i + 1⋅⋅⋅, nsr must obey
nnsr > nnsr+1; and while 1 ≥ νi for all i, it is necessary that 1>νnsr + 1. (The
number of νi equal to 1 is the number of right half plane zeros of G(s).
be equal to the number of νi equal to 1.) The software checks all these
conditions. The minimum order permitted is the number of Hankel
Xmath Model Reduction Module
3-16
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
singular values of F(s) larger than 1– ε (refer to steps 1 through 3 of the
Restrictions section). The maximum order permitted is the number of
nonzero eigenvalues of WcWo larger than ε.
4. Let r be the multiplicity of νns. The algorithm approximates
F(s) = Cw(sI – A)–1B
ˆ
by a transfer function matrix F(s) of order ns – r, using Hankel norm
approximation. The procedure is slightly different from that used in
ophank( ).
Construct an SVD of QP – v2nsI:
V1′
QP – v2NSI = U
V′ = [U1U2]
Σ1 0
Σ1 0
V2′
0 0
0 0
with Σ1 of dimension (ns – r) × (ns – r) and nonsingular. Also, obtain
an orthogonal matrix T, satisfying:
B2 + C′w2T = 0
where B2 and C′w2 are the last r rows of B and C′w , the state variable
–1
matrices appearing in a balanced realization of C′ws(I – A) B. It is
possible to calculate T without evaluating B B, Cw as it turns out (refer
to [AnJ]), and the algorithm does this. Now with
ˆ
ˆ
ˆ
ˆ
–1 ˆ
F(s) = DF + CF(sI – AF) BF
ˆ
ˆ
ˆ
ˆ
Fp(s) = CF(sI – AF)BF
there holds:
ˆ
2
AF = Σ–11U1′ [vnsA′ + QAP – vnsCw′ TB′]V1
BF = Σ–11U′1[QB + vnsCw′T]
ˆ
ˆ
CF = (CwP + vnsTB′)V′
ˆ
DF = –vnsT
© National Instruments Corporation
3-17
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
ˆ
ˆ
Note The expression Fp(s) is the strictly proper part of F(s). The matrix
–1
ˆ
is all pass; this property is not always secured in the multivariable case
vns [F(s) – F(s)]
when ophank( )is used to find a Hankel norm approximation of F(s).
ˆ
ˆ
5. The algorithm constructs G and W , which satisfy,
ˆ
ˆ
G(s) = G(s) – W′(–s)[F(s) – F(s)]
and,
–1
ˆ
W(s) = (I – vnsT′)(I – vnsT)
ˆ
{W(s) – [F(s) – F(s)]G′ + (–s)}
through the state variable formulas
–1 ˆ
′
ˆ
ˆ
ˆ
(G(s) = (D(I – vnsT))[DCF + BWUΣ1](sI – AF) BF)
and:
–1
ˆ
W(s) = (I – vnsT′)D′ + (I – vnsT′)(I – vnsT)
–1
′
ˆ
ˆ
ˆ
CF(sI – AF) [BFD′ + V1C′]
ˆ
ˆ
ˆ
Continue the reduction procedure, starting with G, W , F and
repeating the process till Gr of the desired degree nsris obtained.
^
ˆ
For example, in the second iteration,
is given by:
(s)
G
^
^
ˆ
ˆ
ˆ
ˆ
ˆ
G(s)= G(s) – W′ + (–s)[Fp(s) – F(s)]
(3-4)
Consequences of Step 5 and Justification of Step 6
A number of properties are true:
ˆ
•
G(s) is of order ns – r, with:
–1
ˆ
G (G – G) = vns
(3-5)
∞
Xmath Model Reduction Module
3-18
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
ˆ
ˆ
•
W(s) and Gs stand in the same relation as W(s) and G(s), that is:
ˆ
ˆ
ˆ
ˆ
–
–
W′(–s)W(s) = G(s)G′(–s)
ˆ ˆ ˆ ˆ
ˆ ˆ
With PA′F + AFP = –BFB′F , there holds
′
′
ˆ
BWˆ = PCGˆ + BGˆ DGˆ
or
ˆ
ˆ
ˆ
ˆ
BFD′ + V1C′ = P(DCF + B′WU1Σ1)′ + BF(I – vnsT′)D′
ˆ ˆ
ˆ ′ ˆ
ˆ
ˆ
–
With
there holds
ˆ
QAF + AFQ = –C′FCF
–1
CWˆ = DGˆ (CGˆ – B′Wˆ Q)
or
(I – v T′)(I – v T) CF = [D(I – vnsT)]–1
–1
ˆ
ns
ns
ˆ
ˆ
ˆ
{DCF + B′WU1(Σ1 – [BFD′ + V1C′]′Q)}
–
–
DWˆ = D′Gˆ
–1
ˆ
ˆ
ˆ
is the stable strictly proper part of (W (–s))G(s).
F
ˆ
ˆ
•
•
The Hankel singular values of Fp(and F ) are the first as – r Hankel
singular values of F,
P = Σ1 U1QV1 = V′1QU1Σ1–1
–1
′
ˆ
Q = V1PU1Σ1 = Σ1U′1PV1
ˆ
′
ˆ
Gs has the same zeros in Re[s] > 0 as G(s).
These properties mean that one is immediately positioned to repeat the
ˆ
reduction procedure on , with almost all needed quantities being on
Gs
hand.
© National Instruments Corporation
3-19
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
Error Bounds
The error bound formula (Equation 3-3) is a simple consequence of
iterating (Equation 3-5). To illustrate, suppose there are three reductions
ˆ
ˆ
ˆ
→
→
→
G2 G3
, each by degree one. Then,
G
G
–1
–1
ˆ
ˆ
G (G – G3) = G (G – G)
–1
–1
ˆ ˆ
ˆ
ˆ
+ G GG (G – G2)
ˆ ˆ –1 ˆ ˆ –1
ˆ
ˆ
–1
+ G GG G2G2 (G2 – G3)
Also,
ˆ –1
G (G – G) + I
ˆ
–1
ˆ
G G
=
≤ 1 + vns
Similarly,
Then:
ˆ –1
ˆ
ˆ –1
ˆ
G G2 ≤ 1 + vns – 1, G2 G3 ≤ 1 + vns – 2
–1
ˆ
G (G – G3) ≤ vns + (1 + vns)vns – 1 + (1 + vns – 1)vns – 2
= (1 + vns)(1 + vns – 1)(1 + vns – 2) – 1
The error bound (Equation 3-3) is only exact when there is a single
reduction step. Normally, this algorithm has a lower error bound than
bst( ); in particular, if the νi are all distinct and vnsr + 1 « 1, the error
bounds are approximately
ns
ns
vi
vi
for mulhank( )
for bst(
2
∑
∑
i = nsr + 1
i = nsr + 1
Xmath Model Reduction Module
3-20
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
For mulhank( ), this translates for a scalar system into
ns
ˆ
–86.9
vi dB < 20log10 Gnsr ⁄ G
∑
ns
i = nsr + 1
< 8.69
vi dB
∑
i = nsr + 1
and
ns
phase error <
vi radians
∑
i = nsr + 1
The bounds are double for bst( ).
The error as a function of frequency is always zero at ω = ∞ for bst( )
(or at ω = 0 if a transformation s → s–1 is used), whereas no such particular
property of the error holds for mulhank( ).
Imaginary Axis Zeros (Including Zeros at ∞)
When G(jω) is singular (or zero) on the jω axis or at ∞, reduction can be
handled in the same manner as explained for bst( ).
The key is to use a bilinear transformation [Saf87]. Consider the bilinear
map defined by
z – a
s = ------------------
– bz + 1
s + a
z = --------------
bs + 1
˜
where 0 < a < b–1 and mapping G(s) into G(s) through
s – a
– bs + 1
˜
⎛
⎝
⎞
⎠
-------------------
G(s) = G
s + a
bs + 1
˜
⎛
⎝
⎞
⎠
--------------
G(s) = G
© National Instruments Corporation
3-21
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
˜
The values of G(s) along the jω-axis are the same as the values of G(s)
–1
around a circle with diameter defined by [a – j0, b + j0] on the positive
˜
real axis (refer to Figure 3-2). Also, the values of
along the jω-axis
G(s)
are the same as the values of G(s) around a circle with diameter defined by
[–b–1 + j0, –a + j0].
We can implement an arbitrary bilinear transform using the subsys( )
function, which substitutes a given transfer function for the s- or z-domain
operator, as previously shown.
s – a
– bs + 1
˜
⎛
⎝
⎞
⎠
-------------------
To implement G(s) = G
use:
gtildesys=subsys(gsys,makep([-b,1]/makep([1,-a])
s + a
bs + 1
˜
⎛
⎝
⎞
⎠
--------------
To implement G(s) = G
use:
gsys=subsys(gtildesys,makep([b,1]/makep([1,a])
Note The systems substituted in the previous calls to subsys invert the function
specification because these functions use backward polynomial rotation.
Any zero (or rank reduction) on the jω-axis of G(s) becomes a zero (or rank
˜
reduction) in Re[s] > 0 of G(s), and if G(s) has a zero (or rank reduction)
˜
at infinity, this is shifted to a zero (or rank reduction) of G(s) at the point
b–1, again in Re[s] > 0. If all poles of G(s) are inside the circle of diameter
˜
[–b–1 + j0, a + j0], all poles of
will be in Re[s] < 0, and if G(s) has no
G(s)
zero (or rank reduction) on this circle,
˜
will have no zero (or rank
G(s)
reduction) on the jω-axis, including ω = ∞.
If G(s) is nonsingular for almost all values of s, it will be nonsingular or
have no zero or rank reduction on the circle of diameter [–b–1 + j0, –a + j0]
for almost all choices of a,b. If a and b are chosen small enough, G(s) will
have all its poles inside this circle and no zero or rank reduction on it, while
˜
then will have all poles in Re[s] < 0 and no zero or rank reduction on
G(s)
the jω-axis, including s = ∞.
The steps of the algorithm, when G(s) has a zero on the jω-axis or at s = ∞,
are as follows:
s – a
–bs +1
˜
⎛
⎝
⎞
⎠
1. For small a,b with 0 < a < b–1, form G(s) = G
as shown for
-------------------
gtildesys.
˜
˜
˜
2. Reduce G(s) to Gr(s), this being possible because G(s) is stable and
s + a
˜
⎛
⎞
--------------
3. Form Gr(s) = Gr
as shown for gsys.
⎝
⎠
bs + 1
Xmath Model Reduction Module
3-22
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
G–1(G – Gr)
The error
will be overbounded by the error
∞
–1
˜
˜
˜
∞, and Gr will contain the same zeros in Re[s] ≥ 0 as G.
G (G – Gr)
If there is no zero (or rank reduction) of G(s) at the origin, one can take
a = 0 and b–1 = bandwidth over which a good approximation of G(s) is
needed, and at the very least b–1 sufficiently large that the poles of G(s)
lie in the circle of diameter [–b–1 + j0, –a + j0]. If there is a zero or rank
reduction at the origin, one can replace a = 0 by a = b. It is possible to take
b too small, or, if there is a zero at the origin, to take a too small. In these
cases an error message results, saying that there is a jω-axis zero and/or that
the Riccati equation solution may be in error. The basic explanation is that
˜
as b → 0, and thus a → 0, the zeros of
approach those of G(s). Thus,
G(s)
˜
for sufficiently small b, one or more zeros of
lying on the imaginary axis. The remedy is to increase a and/or b above the
may be identified as
G(s)
desirable values.
The previous procedure for handling jω-axis zeros or zeros at infinity will
be deficient if the number of such zeros is the same as the order of G(s); for
example, if G(s) = 1/d(s), for some stable d(s). In this case, it is possible
again with a bilinear transformation to secure multiplicative
approximations over a limited frequency band. Suppose that
s
˜
⎛
⎝
⎞
⎠
--------------
G(s) = G
εs + 1
˜
Create a system that corresponds to G(s) with:
gtildesys=subs(gsys,(makep([-eps,1])/makep([1,-]))
bilinsys=makep([eps,1])/makep([1,0])
sys=subsys(sys,bilinsys)
Under this transformation:
˜
•
Values of G(s) along the jω-axis correspond to values of G(s) around
a circle in the left half plane on diameter (–ε–1 + j0, 0).
˜
•
Values of G(s) along the jω-axis correspond to values of G(s) around
a circle in the right half plane on diameter (0, ε–1 + j0).
© National Instruments Corporation
3-23
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 3
Multiplicative Error Reduction
˜
Multiplicative approximation of G(s) (along the jω-axis) corresponds to
multiplicative approximation of G(s) around a circle in the right half plane,
touching the jω-axis at the origin. For those points on the jω-axis near the
circle, there will be good multiplicative approximation of G(jω). If a good
approximation of G(s) over an interval [–jΩ, jΩ] it is desired, then
ε
–1 = 5 Ω or 10 Ω are good choices. Reduction then proceeds as follows:
˜
1. FormG(s).
˜
2. Reduce G(s) through bst( ).
˜
3. Form Gr(s) = –Gr(s ⁄ (1 – εs)) with:
gsys=subsys(gtildesys(gtildesys,
makep([-eps,-1])/makep[-1,-0]))
Notice that the number of zeros of G(s) in the circle of diameter (0, ε–1 + j0)
sets a lower bound on the degree of Gr(s)—for such zeros become right half
˜
plane zeros of G(s), and must be preserved by bst( ). Zeros at s = ∞ are
never in this circle, so a procedure for reducing G(s) = 1/d(s) is available.
There is one potential source of failure of the algorithm. Because G(s) is
˜
stable, G(s) certainly will be, as its poles will be in the left half plane circle
on diameter (–ε–1 = j0, 0). If
acquires a pole outside this circle
˜
Gr(s)
(but still in the left half plane of course)—and this appears possible in
principle—Gr(s) will then acquire a pole in Re [s] >0. Should this difficulty
be encountered, a smaller value of ε should be used.
Related Functions
singriccati(), ophank(), bst(), hankelsv()
Xmath Model Reduction Module
3-24
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
4
Frequency-Weighted Error
Reduction
This chapter describes frequency-weighted error reduction problems. This
includes a discussion of controller reduction and fractional representations.
Introduction
Frequency-weighted error reduction means that the error is measured not,
as previously, by
E0
=
G(jω) – Gr(jω)
∞
but rather by
E1
=
G(jω) – Gr(jω)V(jω)
(4-1)
(4-2)
(4-3)
∞
or
or
E2
=
W(jω)[G(jω) – Gr(jω)]
∞
E3
=
W(jω)[G(jω) – Gr(jω)]V(jω)
∞
where W,V are certain weighting matrices. Their presence reflects a desire
that the approximation process be more accurate at certain frequencies
(where V or W have large singular values) than at others (where they
have small singular values). For scalar G(jω), all the indices above are
effectively the same, with the effective weight just |V(jω)|, |W(jω)|,
or |W(jω)V(jω)|.
When the system G is processing signals which do not have a flat spectrum,
and is to be approximated, there is considerable logic in using a weight. If
the signal spectrum is Φ(jω), then taking V(jω) as a stable spectral factor
© National Instruments Corporation
4-1
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
(so that VV* = Φ) is logical. However, a major use of weighting is in
controller reduction, which is now described.
Controller Reduction
Frequency weighted error reduction becomes particularly important in
reducing controller dimension. The LQG and H∞ design procedures lead to
controllers which have order equal to, or roughly equal to, the order of the
plant. Very often, controllers of much lower order will result in acceptable
performance, and will be desired on account of their greater simplicity.
of a controller will not necessarily ensure closeness of the behavior of the
two closed-loop systems formed from the original and reduced order
controller together with the plant. This behavior is dependent in part on the
plant, and so one would expect that a procedure for approximating
controllers ought in some way to reflect the plant. This can be done several
ways as described in the Controller Robustness Result section. The
following result is a trivial variant of one in [Vid85] dealing with robustness
in the face of plant variations.
Controller Robustness Result
Suppose that a controller C stabilizes a plant P, and that Cr is a (reduced
order) approximation to C with the same number of unstable poles. Then
Cr stabilizes P also provided
[C( jω) – Cr( jω)]P( jω)[I + C( jω)P( jω)]–1 ∞ < 1
or
[I + P( jω)C( jω)]–1(P( jω)[C( jω)Cr( jω)]) ∞ < 1
An extrapolation to this thinking [AnM89] suggests that Cr will be a good
approximation to C (from the viewpoint of some form of stability
robustness) if
EIS
EIS
4-2
=
(C – Cr)P(I + CP)–1
∞
or
=
(C – Cr)P(I + CP)–1
∞
Xmath Model Reduction Module
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
is minimized (and of course is less than 1). Notice that these two error
measures are like those of Equation 4-1 and Equation 4-2. The fact that the
plant ought to show up in a good formulation of a controller reduction
problem is evidenced by the appearance of P in the two weights.
It is instructive to consider the shape of the weighting matrix or function
P(Ι + CP)–1. Consider the scalar plant case. In the pass band, |PC| is likely
to be large, and if this is achieved by having |C| large, then |P(Ι + CP)–1|
will be (approximately) small. Also outside the plant bandwidth,
|P(Ι + CP)–1| will be small. This means that it will be most likely to take its
biggest values at frequencies near the unity gain cross-over frequency. This
means that the approximation Cr is being forced to be more accurate near
this frequency than well away from it—a fact very much in accord with
classical control, where one learns the importance of good loop shaping
round this frequency.
The above measures EIS and EOS are advanced after a consideration of
stability, and the need for its preservation in approximating C by Cr. If one
takes the viewpoint that the important thing to preserve is the closed-loop
transfer function matrix, a different error measure arises. With T, Tr
denoting the closed-loop transfer function matrices,
T – Tr = PC(I + PC) – PCr(I + PCr)–1
Then, to a first order approximation in C – Cr, there holds
T – Tr ≈ (I + PC)–1P(C – Cr)(I + PC)–1
The natural error measure is then
EM
=
(I + PC)–1P(C – Cr)(I + PC)–1
(4-4)
∞
and this error measure parallels E3 in Equation 4-3. Further refinement
again is possible. It may well be that closed-loop transfer function matrices
should be better matched at some frequencies than others; if this weighting
on the error in the closed-loop transfer function matrices is determined by
the input spectrum VV* = Φ, then one really wants (T – Tr)V to be small,
so that Equation 4-4 is replaced by
EMS
=
(I + PC)–1P(C – Cr)(I + PC)–1V
∞
© National Instruments Corporation
4-3
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
Most of these ideas are discussed in [Enn84], [AnL89], and [AnM89].
The function wtbalance( )implements weighted reduction, with five
choices of error measure, namely EIS, EOS, EM, EMS, and E1 with arbitrary
V(jω). The first four are specifically for controller reduction, whereas the
last is not aimed specifically at this situation.
Several features of the algorithms are:
•
Only the stable part of C is really reduced; the unstable part is copied
exactly into Cr.
•
A modification of balanced realization truncation underpins the
algorithms, namely (frequency) weighted balanced truncation,
although to avoid numerical problems, the actual construction of
a frequency weighted balanced realization of C is avoided.
•
Frequency weighted Hankel singular values can be computed,
and although no error bound formula is available (in contrast to the
unweighted problem), generally speaking there is little damage done in
reducing by a number of states equal to the number of (relatively) small
Hankel singular values.
The error measures themselves deserve certain comments:
•
The two stability based measures EIS and EOS are derived from a
sufficiency condition for stability, rather than a necessity and
sufficiency condition, and so capture stability a little crudely.
•
For any constant nonsingular N, the error measure EIS can be replaced
N(C – Cr)P(I + CP)–1N–1
by
∞ and the robustness result remains
valid. Use of an N may improve or worsen the quality of the
approximation.
•
•
Having T – Tr small normally ensures closeness of the closed-loop
impulse and step responses.
In classical control especially, constraints on the loop gain can be
imposed (Minimum value of gain in one band, maximum value of gain
in another band, for example). None of the methods presented directly
addresses the task of retaining satisfaction of these constraints after
reduction of a high order acceptable controller. However, judicious use
of a weight V can assist. Suppose that above the closed-loop bandwidth
there is an overbound constraint on the loop gain, which is violated
when a controller reduction is performed (but not with the original
controller). At these frequencies, roughly PC and PCr are small, so that
frequencies in the region in question will evidently encourage PCr to
better track PC.
Xmath Model Reduction Module
4-4
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
Fractional Representations
The treatment of jω-axis or right half plane poles in the above schemes is
crude: they are simply copied into the reduced order controller. A different
approach comes when one uses a so-called matrix fraction description
(MFD) to represent the controller, and controller reduction procedures
based on these representations (only for continuous-time) are found in
fracred( ). Consider first a scalar controller C(s) = n(s) ⁄ d(s). One
can take a stable polynomial d(s) of the same degree as d, and then
represent the controller as a ratio of two stable transfer functions, namely
–1
n(s)
d(s)
d(s)
d(s)
---------
---------
Now n ⁄ d is the numerator, and d ⁄ d the denominator. Write d ⁄ d as
1 + e ⁄ d. Then we have the equivalence shown in Figure 4-1.
e
--
d
n
--
d
C(s)
Figure 4-1. Controller Representation Through Stable Fractions
Evidently, C(s) can be formed by completing the following steps:
1. Construction of the one-input, two-output stable transfer function
matrix
n ⁄ d
G =
e ⁄ d
(which has order equal to that of d or d).
2. Interconnection through negative feedback of the second output to the
single input.
These observations motivate the reduction procedure:
•
Reduce G to Gr; notice that G is stable. Let Gr be
nr ⁄ dr
G =
er ⁄ dr
© National Instruments Corporation
4-5
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
•
Form the reduced controller by interconnecting using negative
feedback the second output of Gr to the input, that is, set
nr
Cr(s) = ---------------
dr + er
Nothing has been said as to how d should be chosen—and the end result
of the reduction, Cr(s), depends on dr. Nor has the reduction procedure
been specified.
When C(s) has been designed to combine a state estimator with a
stabilizing feedback law, it turns out that there is a natural choice for d(s).
As for the reduction procedure, one possibility is to use a weight based
on the spectrum of the input signals to G—and in case C(s) has been
determined by an LQG optimal design, this spectrum turns out to be white,
that is, independent of frequency, so that no weight (apart perhaps from
scaling) is needed. A second possibility is to use a weight based on a
stability robustness measure. These points are now discussed in more
detail.
To understand the construction of a natural fractional representation for
C(s), suppose that P(s) = C(sI – A)–1B and let KR, KE be matrices such
that A – BKR and A – KEC are stable. The controller
·
ˆ
ˆ
ˆ
x = Ax + Bu – KE(Cx – y)
ˆ
u = –KRx
ˆ
generates an estimate –KRx of the feedback control –KRx . The controller
can be represented as a series compensator
·
ˆ
ˆ
ˆ
ˆ
x = Ax + BKRx – KECx + KEy
ˆ
u = –KRx
(with compensator input y and output u). Allowing for connection with
negative feedback, the compensator transfer function matrix is:
C(s) = KR(sI – A + BKR + KEC)–1KE
Xmath Model Reduction Module
4-6
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
Matrix algebra shows that C(s) can be described through a left or right
matrix fraction description
C(s) = D–L1(s)NL(s) = NR(s)D–R1(s)
with DL, and related values, all stable transfer function matrices.
In particular:
DL = I + KR(sI – A + KEC)–1B
NL = KR(sI – A + KEC)–1KE
DR = I + C(sI – A + BKR)–1KE
For matrix C(s), the left and right matrix fraction descriptions are distinct
entities. It is the right MFD which corresponds to Figure 4-1; refer to
Figure 4-2.
C
-
+
-
1
s
+
+
KE
KR
P(s)
--
+
A – BKR
+
C(s)
P(s)
-
Figure 4-2. C(s) Implemented to Display Right MFD Representation
© National Instruments Corporation
4-7
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
The left MFD corresponds to the setup of Figure 4-3.
+
-
P(s)
Kr(sI – A +KEC)–1
B
KE
Figure 4-3. C(s) Implemented to Display Left MFD Representation
The setup of Figure 4-2 suggests approximation of:
Kr
G(s) =
(sI – A + BKr)–1KE
C
whereas that of Figure 4-3 suggests approximation of:
H(s) = KR(sI – A + KEC)–1
B KE
In the LQG optimal case, the signal driving KE in Figure 4-2 is white noise
(the innovations process); this motivates the possibility of using no
frequency dependent weighting in approximating G(s) [but observe that
after approximating, the signal will no longer be white noise, so that
frequency dependent weighting for H(s). These are two of the options
offered by fracred( ).
Two more fracred( )options depend on examining stability robustness
(the options are duals of one another). From the stability point of view, the
ˆ
set-up of Figure 4-3 is identical to that of Figure 4-4, with
.
P =
P I
Xmath Model Reduction Module
4-8
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
-
+
-
KR
C
(sI – A +BKR)–1KE
P(s) I
+
+
ˆ
C(s)
P(s)
-
Figure 4-4. Redrawn; Individual Signal Paths as Vector Paths
It is possible to verify that
–1 ˆ
(I + PG) P = [CsI – A + KEC B
–1
ˆ
I – C(sI – (A + KEC)–1KE)]
ˆ
–1 ˆ
and accordingly the output weight (I + PG) P = W can be used in an
error measure W(G – Gr) . It turns out that the calculations for frequency
weighted balanced truncation of G and subsequent construction of Cr(s) are
exceptionally easy using this weight.
The second fracred( )option is the dual of this. The error measure is
(H – Hr)V where:
I – KR(sI – A + BKR)–1B
C(sI – A + BKR)–1B
V =
It is possible to argue heuristically the relevance of these error measures
from a second point of view. It turns out that:
DL NL V1 –NR
I 0
=
–W1 W2 V2 Dr
0 I
© National Instruments Corporation
4-9
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
(Here, the Wi and Vi are submatrices of W,V.) Evidently,
NR
DR
V = I
and
W
= I
DL NL
Some manipulation shows that trying to preserve these identities after
approximation of DL, NL or NR, DR suggests use of the error measures
W(G – Gr)
[LAL90].
and (H – Hr)V . For further details, refer to
and
[AnM89]
∞
∞
In all four fracred( )options, it is possible to construct (weighted)
Hankel singular values, and to use them as a guide to the likely quality of
approximation. The patterns tend to be different for the four options.
The fracred( )options are normally different in outcome from the
wtbalance( )options. However, if the controller has been designed
by the loop transfer recovery method and is stable, then one of the
fracred( )options is essentially the same as one of the wtbalance( )
options, refer to [LiA90].
More precisely, if the LTR design is performed with input noise or process
noise weighting tending to infinity, reduction with fracred( )and
type="left stab", which uses the error measure (H – Hr)V , leads to
effectively the same reduction as wtbalance( )with the type="input
stab". If the LTR design is performed with state or output weighting
tending to infinity (in the index determining the state feedback law),
reduction with fracred( )and type="right stab"using the error
measure W(G – Gr) ∞ leads to effectively the same reduction as
wtbalance( )with type="output stab".
wtbalance( )
[SysCR,SysCLR,HSV] =wtbalance(Sys,SysC,type,{nscr,SysV})
The wtbalance( )function calculates a frequency weighted balanced
truncation of a system.
wtbalance( )has two separate uses:
with the plant P(s) known. Frequency-weighted balanced truncation is
used, with the weights involving P(s) and being calculated in a
predominantly standard way.
Xmath Model Reduction Module
4-10
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
•
Reduce the order of a transfer function matrix C(s) through
frequency-weighted balanced truncation, a stable frequency weight
V(s) being prescribed.
The syntax is more accented towards the first use. For the second use,
the user should set S = 0, NS = 0. This results in (automatically)
SCLR = NSCLR = 0. The user will also select the type="input
spec".
Let Cr(s) be the reduced order approximation of C(s) which is being
sought. Its order is either specified in advance, or the user responds to
a prompt after presentation of the weighted Hankel singular values.
Then the different types concentrate on (approximately) minimizing
certain error measures, through frequency weighted balanced
truncation. These are shown in Table 4-1.
Table 4-1. Types versus Error Measures
Type
Error Measure
[C – Cr]P[I + CP]–1
[I + PC]–1P[C – Cr]
[I + PC]–1P[C – Cr][I + PC]–1
"input stab"
"output stab"
"match"
∞
∞
∞
[I + PC]–1P[C – Cr][I + PC]–1V
"match spec"
"input spec"
∞
[C – Cr]V
∞
These error measures have certain interpretations, as shown in Table 4-2.
In case C(s) is not a compensator in a closed-loop and the error measure
V(jω)[C(jω) – Cr(jω)]
∞
is of interest, you can work with type="input spec"and C', V' in lieu
of C and V.
There is no restriction on the stability of C(s) [or indeed of P(s)] in the
algorithm, though if C(s) is a controller the closed-loop must be stabilizing.
Also, V(s) must be stable. Hence all weights (on the left or right of
C(jω) – Cr(jω) in the error measures) will be stable. The algorithm,
however, treats unstable C(s) in a special way, by reducing only the stable
into Cr(s).
© National Instruments Corporation
4-11
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
This rather crude approach to the handling of the unstable part of a
controller is avoided in fracred( ), which provides an alternative to
wtbalance( )for controller reduction, at least for an important family
of controllers.
Table 4-2. Error Measure Interpretation for wtbalance
Type
Error Measure Interpretations
"input stab"
"output stab"
"match"
A stability robustness argument, based on breaking the loop at the controller
output, indicates that if C is stabilizing for P and the error measure is less
than 1, then Cr is stabilizing for P. The smaller the error measure is, the
greater the stability robustness.
A similar stability robustness argument, but based on breaking the loop
at the controller input, indicates that if C is stabilizing for P and the error
measure is less that 1, then Cr is stabilizing for P. The smaller the error
measure is, the greater the stability robustness.
If T = PC(I + PC)–1 and Tr = PCr(I + PCr)–1 are the two closed-loop transfer
function matrices, then T – Tr to first order in C – Cr is given by
(I + PC)–1P[Cr– C][I + PC]–1, so that the error measure looks at matching
of the closed-loop transfer function matrix.
"match spec"
It may be important to match closed-loop transfer function matrices more
at certain frequencies than others; frequency weighting is achieved by
introducing V(s). Frequencies corresponding to larger values of |V(jω)| or
V(jω)V (jω) will be the frequencies at which T(jω) and Tr(jω) should have
*
smaller error.
"input spec"
This is the one error measure that is not associated with a plant, or
closed-loop of some kind. It simply allows the user to emphasize certain
frequencies in the reduction procedures.
Algorithm
The major steps of the algorithm are as follows:
1. Check dimension, syntax, stability of SysV, closed-loop stability, and
decomposition of C(s) into the sum of a stable part (poles in Re[s] < 0)
and unstable part (poles in Re[s] ≥ 0); stable( )is used for this
purpose.
2. Compute input (right) weight and/or output (left) weight as appropriate
for the specified type.
Xmath Model Reduction Module
4-12
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
3. Compute weighted Hankel Singular Values σi (described in more
detail later). If the order of Cr(s) is not specified a priori, it must be
input at this time. Certain values may be flagged as unacceptable for
various reasons. In particular nscr cannot be chosen so that
σnscr = σnscr + 1.
4. Execute reduction step on stable part of C(s), based on a modification
of redschur( )to accommodate frequency weighting, and yielding
stable part of Cr(s).
5. Compute Cr(s) by adding unstable part of C(s) to stable part of Cr(s).
6. Check closed-loop stability with Cr(s) introduced in place of C(s),
at least in case C(s) is a compensator.
More details of steps 3 and 4, will be given for the case when there is an
input weight only. The case when there is an output weight only is almost
the same, and the case when both weights are present is very similar, refer
to [Enn84a] for a treatment. Let
C(s) = Dc + Cc(sI – Ac)–1Bc
WS(s) = Dw + Cw(sI – Aw)–1Bw
be a stable transfer function matrix to be reduced and its stable weight.
Thus, W(s) may be P(I + CP)–1, corresponding to "input stab", and will
thus have been calculated in step 2; or it maybe an independently specified
stable V(s). Then
–1
sI – Ac BcCw
BcDw
Bw
Cs(s)W(s) = DcDw +
Cc DcCw
0
sI – Aw
The controllability grammian P satisfying
Ac′
C′wB′C A′w
0
BcDw
Bw
Ac BcCw
D′wB′c B′w
P
+
P +
= 0
0
Aw
is written as
Pcc Pcw
P′cw Pww
P =
© National Instruments Corporation
4-13
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
and the observability grammian Q, defined in the obvious way, is written as
Qcc Qcw
Q′cw Qww
Q =
It is trivial to verify that QccAc + A′cQcc = –C′ C so that Qcc is the
c
c
observability gramian of Cs(s) alone, as well as a submatrix of Q.
The weighted Hankel singular values of Cs(s) are the square roots of the
eigenvalues of PccQcc. They differ from the usual or unweighted Hankel
singular values because Pcc is not the controllability gramian of Cs(s) but
rather a weighted controllability gramian. The usual controllability
gramian can be regarded as E[x x′ ] when Cs(s) is excited by white noise.
c
c
′
The weighted controllability gramian is still
excited by colored noise, that is, the output of the shaping filter W(s), which
, but now C (s) is
E[xcxc]
s
is excited by white noise.
Small weighted Hankel singular values are a pointer to the possibility
of eliminating states from Cs(s) without incurring a large error in
[C(jω) – Cr(jω)]W(jω) ∞. No error bound formula is known, however.
The actual reduction procedure is virtually the same as that of
redschur( ), except that Pcc is used. Thus Schur decompositions of
PccQcc are formed with the eigenvalues in ascending and descending order
V′APccQccVA = Sasc
V′DPccQccVD = Sdes
The maximum order permitted is the number of nonzero eigenvalues of
PccQcc that are larger than ε.
The matrices VA, VD are orthogonal and Sasc and Sdes are upper triangular.
Next, submatrices are obtained as follows:
0
Inscr
Vlbig = VA
Vrbig = VD
Inscr
0
and then a singular value decomposition is formed:
UebigSebigVebig = V′lbigVrbig
Xmath Model Reduction Module
4-14
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
From these quantities the transformation matrices used for calculating
Csr(s), the stable part of Cr(s), are defined
–1 ⁄ 2
Slbig = VlbigVebig
S
ebig
–1 ⁄ 2
Srbig = VrbigVebig
S
ebig
and then
ACR = S′lbigACSrbig BCR = Sl′bigBC
ACR = CCSrbig BCR = DC
Just as in unweighted balanced truncation, the reduced order transfer
function matrix is guaranteed stable, the same is guaranteed to be true in
weighted balanced truncation when either a left (output) weight or a right
(input) weight is used. It is suspected to be true when both input and output
weights are present. The overall algorithm is not, however, at risk in this
case, since it is stability of the closed-loop system which is the key issue of
concern, (except for type="input spec", but here there is only a single
weight, and so the theory guarantees preservation of stability).
Related Functions
balance(), redschur(), stable(), fracred()
fracred( )
[SysCR,HSV] = fracred(Sys,Kr,Ke,type,{nscr,Qyy})
The fracred( )function uses fractional representations to calculate a
reduction of a continuous-time compensator comprising a state estimator
with state feedback law.
Restrictions
1. The closed-loop system (SCLR,NSCLR) is calculated from
sysol=scr*sys
# open loop system
syscl=feedback(sysol)
# closed loop system
considered by fracred( ).
© National Instruments Corporation
4-15
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
3. Only continuous systems are accepted; for discrete systems use
makecontinuous( )before calling bst( ), then discretize the
result.
Sys=fracred(makecontinuous(SysD));
SysD=discretize(Sys);
Defining and Reducing a Controller
Suppose P(s) = C(sI – A)–1B and A – BKR and A – KEC are stable (where
KR is a stabilizing state feedback gain and KE a stabilizing observer gain).
A controller for the plant P(s) can be defined by
·
ˆ
ˆ
ˆ
x = Ax + Bu – KE(Cx – y)
ˆ
u = –KRx
(with u the plant input and y the plant output). The associated series
compensator under unity negative feedback is
C(s) = KR(sI – A + BKR + KEC)–1KE
and this may be written as a left or right MFD as follows:
C(s) = [I + KR(sI – A + KEC)–1B]–1KR(sI – A + KEC)–1KE
(4-5)
(4-6)
C(s) = KR(sI – A + BKR)–1KE[I + C(sI – A + BKR)–1KE]–1
The reduction procedures "right perf"and "left perf"have similar
rationales. We shall describe "right perf", refer to [AnM89] and
[LiA86]. The first rationale involves observing that to reduce C(s), one
might as well reduce its numerator and denominator simultaneously, and
then form a new fraction Cr(s) of lower order than C(s).
This amounts to reducing
KR
E(s) =
(sI – A + BKR)–1KE
(4-7)
C
Xmath Model Reduction Module
4-16
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
to, for example,
KR
ˆ
Er(s) =
(sI – A)KE
C
through, for example, balanced truncation, and then defining:
–1
–1
–1
ˆ
ˆ
C(s) = KR(sI – A) KE[I + C(sI – A) KE]
–1
ˆ
= KR(sI – A + KEC) KE
For the second rationale, consider Figure 4-5.
C
-
+
-
1
--
+
+
KE
KR
P(s)
s
+
X
A – BKR
Figure 4-5. Internal Structure of Controller
Recognize that the controller C(s) (shown within the hazy rectangle in
Figure 4-5) can be constructed by implementing
KR(sI – A + BKR)–1KE
and
C(sI – A + BKR)–1KE
and then applying an interconnection rule (connect the output of the second
transfer function matrix back to the input at point X in Figure 4-5).
© National Instruments Corporation
4-17
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
Controller reduction proceeds by implementing the same connection rule
but on reduced versions of the two transfer function matrices.
When KE has been defined through Kalman filtering considerations, the
spectrum of the signal driving KE in Figure 4-5 is white, with intensity Qyy.
It follows that to reflect in the multiple input case the different intensities
on the different scalar inputs, it is advisable to introduce at some stage a
weight Q1yy⁄ 2 into the reduction process.
Algorithm
After preliminary checks, the algorithm steps are:
1. Form the observability and weighted (through Qyy) controllability
grammians of E(s) in Equation 4-7 by
P(A – BKR)′ + (A – BKR)P = –KEQyyKE′
(4-8)
(4-9)
Q(A – BKR) + (A – BKR)′Q = –K′RKR – C′C
2. Compute the square roots of the eigenvalues of PQ (Hankel singular
values of the fractional representation of Equation 4-5). The maximum
order permitted is the number of nonzero eigenvalues of PQ that are
larger than ε.
3. Introduce the order of the reduced-order controller, possibly by
displaying the Hankel singular values (HSVs) to the user. Broadly
speaking, one can throw away small HSVs but not large ones.
4. Using redschur( )-type calculations, find a state-variable
description of Er(s). This means that Er(s) is the transfer function
matrix of a truncation of a balanced realization of E(s), but the
redschur( )type calculations avoid the possibly numerically
difficult step of balancing the initially known realization of E(s).
Suppose that:
′
′
ˆ
A = Slbig(A – BKR)Srbig, KE = SlbigKE
5. Define the reduced order controller Cr(s) by
ACR = S′lbig(A – BKR – KEC)Srbig
(4-10)
Cr(s) = CCR(sI – ACR)–1BCR
Xmath Model Reduction Module
4-18
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
E(s) = KR(sI – A + KEC)–1
(4-11)
B KE
which is formed from the numerator and denominator of the MFD
in Equation 4-5. The grammian equations (Equation 4-8 and
Equation 4-9) are replaced by
Q(A – KEC) + (A – KEC)′Q = –K′RKR
redschur( )-type calculations are used to reduce E(s) and Equation 4-10
again yields the reduced-order controller. Notice that the HSVs obtained
from Equation 4-10 or the left MFD (Equation 4-5) of C(s) will in general
may be possible to reduce much more with the left MFD than with the right
MFD (or vice-versa) before closed-loop stability is lost.
As noted in the fracred( )input listing, type="left stab"and
"right stab"focus on a stability robustness measure, in conjunction
with Equation 4-5 and Equation 4-6, respectively. Leaving aside for the
moment the explanation, the key differences in the algorithm computations
lie solely in the calculation of the grammians P and Q. For type="left
stab", these are given by
P(A – BKR)′ + (A – BKR)P = –BB′
Q(A – KEC) + (A – KEC)′Q = –K′RKR
and for "right stab",
P(A – BKR)′ + (A – BKR)P = –KEKE′
(4-12)
(4-13)
Q(A – KEC) + (A – KEC)′Q = –C′C
© National Instruments Corporation
4-19
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
Additional Background
A discussion of the stability robustness measure can be found in [AnM89]
and [LAL90]. The idea can be understood with reference to the transfer
functions E(s) and Er(s) used in discussing type="right perf". It is
possible to argue (through block diagram manipulation) that
•
C(s) stabilizes P(s) when E(s) stabilizes (as a series compensator) with
ˆ
unity negative feedback
.
P(s) =
P(s) I
•
Er(s) also will stabilize [P(s)I], and then Cr(s) will stabilize P(s),
provided
C( jωI – A + KEC)–1B I – C( jωI – A + KEC)–1KE
∞
Accordingly, it makes sense to try to reduce E by frequency-weighted
balanced truncation. When this is done, the controllability grammian for
Equation 4-5, at least with Qyy = I, and Equation 4-12 are the same while
Equation 4-6 and Equation 4-13 are quite different.) The calculations
leading to Equation 4-13 are set out in [LAL90].
The argument for type="left perf"is dual. Another insight into
Equation 4-14 is provided by relations set out in [NJB84]. There, it is
established (in a somewhat broader context) that
{C( jωI – A + KEC)–1B
Kr(sI – A + BKR)–1KE
I + C( jωI – A + BKR)–1KE
×
= I
The left matrix is the weighting matrix in Equation 4-14; the right matrix is
the numerator of C(jω) stacked on the denominator, or alternatively
0
E(jω) +
I
This formula then suggests the desirability of retaining the weight in the
Xmath Model Reduction Module
4-20
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 4
Frequency-Weighted Error Reduction
The four schemes all produce different HSVs; it follows that it may be
prudent to try all four schemes for a particular controller reduction. Recall
again that their relative sizes are only a guide as to what can be thrown away
without incurring much error. There is no simple rule to indicate which of
the four schemes will be the most effective at controller reduction.
Two rough rules can, however, be formulated.
•
•
Problems with instability through reduction to too low a controller
order are more likely with "left perf"and "right perf"than
"left stab"or "right stab".
If the controller has been designed using the loop transfer recovery
idea, "left stab"will probably be attractive if the input noise
covariance is very large, and "right stab"will probably be
large, [LiA90]. The reduced controllers will then actually be very
similar to those obtained using wtbalance( )with the option
"input stab"in the first case and "output stab"in the second
case.
One example gives the HSVs summarized in Table 4-3 for an eighth order
controller.
Table 4-3. HSVs for an Eighth Order Controller
1
2
3
4
5
6
7
8
right perf
left perf
right stab
left stab
.0339
.0164
4.8742
.7278
1.317
.0128
3.8457
.1123
1.1269
.0102
3.7813
.0783
1.0862
.0040
1.2255
.0242
.9638
.0037
1.1750
.0181
.5846
.0000
.5055
.0107
.5646
.0000
.0413
.0099
.3144
4.9075
3.3081
1.3914
The most attractive candidate for reducing to second order is right stab.
This is because the HSVs being discarded (columns 3 to 8) are smaller
relative to those being retained (columns 1 and 2) for right stabthan for
the other three candidates.
Note The relative values count, not the absolute values.
Related Functions
redschur(), wtbalance()
© National Instruments Corporation
4-21
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
5
Utilities
and compare( ).
The background to hankelsv( ), which calculates Hankel singular
values, was presented in Chapter 1, Introduction. Hankel singular values
are also calculated in other functions, sometimes by other procedures.
A comparison of the procedures is given in the Hankel Singular Values
section. The function compare( )serves to facilitate the comparisons
of an unreduced and a reduced system, from various points of views.
The function stable( )is used to separate (additively) a system into its
stable and unstable parts, that is, given G(s), the function determines Gs(s)
and Gu(s), the first with all poles in Re[s] < 0, the second with all poles in
Re[s] ≥ 0, such that
G(s) = Gs(s) + Gu(s)
The function is used within some of the other functions of the Model
Reduction Module. It should also be used when reduction of an unstable
G(s) is contemplated. The normal reduction functions, for example,
balmoore( )or redschur( ), require stability of the transfer function
matrix G(s) being reduced. If G(s) is unstable, stable( )should be used
to generate Gs(s) and Gu(s); reduction of Gs(s) should be performed, and
then Gu(s) added to the outcome using the +operator, to yield the desired
reduction of Gs(s).
hankelsv( )
[HSV,Wc,Wo] = hankelsv(Sys,{noplot})
The hankelsv( )function computes the Hankel Singular Values of a
stable system (continuous or discrete) and displays them in a bar plot.
© National Instruments Corporation
5-1
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 5
Utilities
The gramian matrices are defined by solving the equations (in continuous
time)
AWc + WcA′ = –BB′
WoA + A′W0 = –C′C
and, in discrete time
Wc – AWcA′ = BB′
Wo – A′WoA = C′C
The computations are effected with lyapunov( )and stability is checked,
which is time-consuming. The Hankel singular values are the square roots
of the eigenvalues of the product.
Related Functions
lyapunov(), dlyapunov()
stable( )
[SysS,SysU] = stable(Sys,{tol})
The stable( )function decomposes Sysinto its stable (SysS) and
unstable(SysU) parts, such that Sys=SysS+SysU.
Continuous systems have unstable poles if real parts > –tol.
Discrete systems have unstable poles if magnitudes > 1-tol.
•
•
The direct term (D matrix) is included in SysS.
If Syshas poles clustered near -tol(or 1-tol), then SysSand SysU
might be ill-conditioned. To avoid this problem choose tolto a value
that is not close to the majority of poles.
Algorithm
The algorithm begins by transforming the A matrix to Schur form, and
counting the number of stable and unstable eigenvalues, together with
those for which classification is doubtful. Stable eigenvalues are those
•
•
Re[s] < 0 (continuous time)
|z| < 1 (discrete time)
Xmath Model Reduction Module
5-2
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 5
Utilities
Doubtful ones are those for which the real part of the eigenvalue has
magnitude less than or equal to tolfor continuous-time, or eigenvalue
magnitude within the following range for discrete time:
1 – tol, 1 + tol
A warning is given if doubtful eigenvalues exist.
The algorithm then computes a real ordered Schur decomposition of A
so that after transformation
AS ASU
A =
0 AU
where the eigenvalues of AS and AU are respectively stable and unstable.
A matrix X satisfying –ASX + XAU + ASU = 0 is then determined by calling
the algorithm sylvester( ). The eigenvalue properties of AS and AU
guarantee that X exists. If doubtful eigenvalues are present, they are
assigned to the unstable part of Sys. In this circumstance you get the
message,
The system has poles near, or upon, the jw-axis
for continuous systems, and the following for discrete systems:
The system has poles near the unit circle.
Note If A has eigenvalues clustered near -tol(1–tolin discrete-time), then X is likely
to be ill-conditioned and consequently SysSand SysUwill also be ill-conditioned. (For
example, the B matrix of SysScould contain very small values, while the C matrix could
contain large values. In this case, SysSwould be very weakly controllable and very
strongly observable. This will cause problems when gramians and Hankel singular values
are calculated.) To avoid this problem, change tolto a value that is not close to the
majority of eigenvalues.
A further transformation of A is constructed using X:
I X
I –X
A →
A
0 I
0 I
AS
0
=
0 AU
© National Instruments Corporation
5-3
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 5
Utilities
After this last transformation, and with
BS
B =
C = [CSCU]
BU
it follows that
SysS = [AS,AS;CSD]
and
SysU = [AU,BU;CU0]
By combining the transformation yielding the real ordered Schur form for
A with the transformation defined using X, the overall transformation T is
readily identified. In case all eigenvalues of A are stable or all are unstable,
this is flagged, and T = I.
stable( )can be combined with a reduction algorithm such as
redschur( )or balmoore( )to reduce the order of a system with some
unstable and some stable poles. One uses stable( )to separate the stable
and unstable parts, and then, for example, reduces the stable part with
redschur( ); the reduced stable part is added to the original unstable part
to provide the desired system reduction.
Related Functions
sylvester(), schur(), redschur(), balmoore()
compare( )
[respdiff] = compare(Sys,SysRed,FTvec,{Fmin,Fmax,npts,radians,type})
The compare( )function provides a number of different graphical tests
which can be used to compare two state-space system implementations.
compare( )can be used as a tool for evaluating a reduced-order system
by comparing it with the original full-order system from which it was
obtained. However, it can be used for more general comparisons as well,
such as examining the results of different discretization or identification
techniques.
Xmath Model Reduction Module
5-4
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
6
Tutorial
This chapter illustrates a number of the MRM functions and their
underlying ideas. A plant and full-order controller are defined, and then
the effects of various reduction algorithms are examined. The data for this
example is stored in the file mr_disc.xmdin the Xmath demos directory.
To follow the example, start Xmath, and then select File»Load from the
Xmath Commands menu, or enter the load command with the file
specification appropriate to your operating system from the Xmath
Commands area. For example:
load "$XMATH/demos/mr_disc"
Plant and Full-Order Controller
The plant in question comprises four spinning disks, connected by a
flexible shaft. A motor applies torque to the third disk, and the output
variable of interest is the angular displacement of the first disk. The plant
transfer function, which is nonminimum phase and has a double pole at the
origin, is as follows:
s22ζ0ω0s + ω02 s2ζ1ω1s + ω12
s + a
a
----------------------------------- -------------------------------- -----------
⋅
⋅
ω02
ω12
1
----------------------------------------------------------------------------------------------------------------------------
G(s) =
4s2 s22ζ2ω2s + ω22 s22ζ3ω3s + ω23 s22ζ4ω4s + ω42
----------------------------------- ----------------------------------- -----------------------------------
⋅
⋅
ω22
ω32
ω42
with:
ζ =0.02 ω =1
0
0
ζ =-0.4 ω =5.65
1
1
ζ =0.02 ω =0.765
2
2
ζ =0.02 ω =1.41
3
3
ζ =0.02 ω =1.85
4
4
a=4.84
© National Instruments Corporation
6-1
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
A minimal realization in modal coordinates is C(sI – A)–1B where:
⎧
⎨
⎩
⎫
⎬
⎭
0 1 –0.015 0.765
–0.028 1.410
–0.04 1.85
A = diag
,
,
,
0 0 –0.765 –0.015 –1.410 –0.028 –1.85 –0.04
0.026
–0.251
0.033
–0.996
–0.105
0.261
–0.886
–4.017
0.145
0.009
B =
C′ =
–0.001
–0.043
0.002
3.604
0.280
–0.026
The specifications seek high loop gain at low frequencies (for performance)
and low loop gain at high frequencies (to guarantee stability in the presence
of unstructured uncertainty). More specifically, the loop gain has to lie
outside the shaded region shown in Figure 6-1.
40 dB/decade
Frequency (rad/sec)
0.3
0.07
40 dB/decade
Figure 6-1. Loop Gain Constraints
Xmath Model Reduction Module
6-2
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
With a state weighting matrix,
Q = 1e-3*diag([2,2,80,80,8,8,3,3]);
R = 1;
(and unity control weighting), a state-feedback control-gain is determined
through a linear-quadratic performance index minimization as:
[Kr,ev] = regulator(sys,Q,R);
A – B × Kr is stable. Next, with an input noise variance matrix Q = WtBBWt
where,
Wt = DIAG([0.346, 0.346, 0.024, 0.0240.042, 0.0420.042, 0.042])
ˆ
and measurement noise covariance matrix R =1, an estimation gain Ke
(so that A – KeC is stable) is determined:
Qhat = Wt*b*b'*Wt;
Rhat = 1;
[Ke,ev] = estimator(sys,Qhat,Rhat,{skipChks});
The keyword skipChkscircumvents syntax checking in most functions.
It is used here because we know that Qhatdoes not fulfill positive
semidefiniteness due to numerics).
sysc=lqgcomp(sys,Kr,Ke);
poles(sysc)
ans (a column vector) =
-0.296674 + 0.292246 j
-0.296674 - 0.292246 j
-0.15095 + 0.765357 j
-0.15095 - 0.765357 j
-0.239151 + 1.415
-0.239151 - 1.415
j
j
-0.129808 + 1.84093 j
-0.129808 - 1.84093 j
The compensator itself is open-loop stable. A brief explanation of how Q
and Wtare chosen is as follows. First, Qis chosen to ensure that the loop
gain Kr(jωI – A)–1B (which would be relevant were the state measurable)
meets the constraints as far as possible. However, it is not possible to obtain
a 40 dB per decade roll-off at high frequencies, as LQ design virtually
always yields a 20 dB per decade roll-off. Second, a loop transfer recovery
ˆ
Q
approach to the choice of as ρBB′ for some large ρ is modified through
the introduction of the diagonal matrix Wt. The larger entries of Wt, because
of the modal coordinate system, in effect promote better loop transfer
© National Instruments Corporation
6-3
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
recovery at low frequencies; there is consequently a faster roll-off of the
Figure 6-2 displays the (magnitudes of the) plant transfer function, the
compensator transfer function and the loop gain, as well as the constraints;
evidently the compensated plant meets the constraints.
You can enter the following commands to create a plot equivalent to
Figure 6-2:
sysol=sys*sysc;
svals=svplot(sys,w,{radians});
svalsc=svplot(sysc,w,{radians});
svalsol=svplot(sysol,w,{radians});
plot(svals,{x_log,!grid,!ylab,
line_width=2,hold})
plot(svalsc,{keep})
plot(svalsol,{keep})
f2=plot(wc,constr,{keep,
legend=["plant","compensator",
"compensated plant","constraint"]})
plot({!hold})
Figure 6-2. Frequency Response for Plant, Compensator, and Compensated Plant
Xmath Model Reduction Module
6-4
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Controller Reduction
This section contrasts the effect of unweighted and weighted controller
reduction. Unweighted reduction is at first examined, through
redschur( )(using balance( )or balmoore( )will give similar
results). The Hankel singular values of the controller transfer function are
6.264×10–2 4.901×10–2 2.581×10–2 2.474×10–2
1.545×10–2 1.335×10–2 9.467×10–3 9.466×10–3
A reduction to order 2 is attempted. The ending Hankel singular values, that
is, σ3, σ4, ..., σ8, have a sum that is not particularly small with respect to σ1
and σ2; this is an indication that problems may arise in the reduction.
[syscr,hsv] = redschur(sysc,2);
svalsRol = svplot(sys*syscr,w,{radians});
plot(svalsol, {keep})
f3=plot(wc, constr,{keep,!grid,
legend=["reduced","original","constrained"],
title="Open-Loop Gain Using redschur()"})
Figure 6-3. Open-Loop Gain Using redschur
© National Instruments Corporation
6-5
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
gain is shown in Figure 6-3. The error near the unity gain crossover
frequency may not look large, but it is considerably larger than that
obtained through frequency weighted reduction methods, as described
later.
Figure 6-3 also shows the inability to suppress all three plant resonances,
in contrast to the full-order controller. Two are such as to cause violation
between the full-order and reduced-order controller, in the vicinity of
0.1 radians per second. The step response has overshoot of 50% as opposed
to 40% and the ripple persists for longer.
We use the compare( )function (refer to the compare( ) section of
Chapter 5, Utilities) to reproduce Figures 6-4 and 6-5. Calculate the
full-order closed-loop system, then the closed-loop system with the
reduced-order compensator:
syscl = feedback(sysol);
sysolr=sys*syscr;
sysclr=feedback(sysolr);
Xmath Model Reduction Module
6-6
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-4:
compare(syscl,sysclr,w,{radians,type=5})
f4=plot({keep,legend=["original","reduced"]})
Figure 6-4. Closed-Loop Gain with redschur
© National Instruments Corporation
6-7
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
ophank( )
ophank( )is next used to reduce the controller with the results shown in
Figures 6-6, 6-7, and 6-8.
Generate Figure 6-6:
[syscr,sysu,hsv]=ophank(sysc,2);
svalsrol = svplot(sys*syscr,w,{radians});
plot(svalsol, {keep})
f6=plot(wc, constr, {keep,!grid,
title="Open-loop gain using ophank()"})
Figure 6-6. Open-Loop Gain Using ophank
© National Instruments Corporation
6-9
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-7:
syscl = feedback(sysol);
sysolr=sys*syscr;
sysclr=feedback(sysolr);
compare(syscl,sysclr,w,{radians,type=5})
f7=plot({keep,legend=["original","reduced"]})
Figure 6-7. Closed-Loop Gain with ophank
Xmath Model Reduction Module
6-10
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-8:
tvec=0:(140/99):140;
compare(syscl,sysclr,tvec,{type=7})
f8=plot({keep,legend=["original","reduced"]})
Figure 6-8. Step Response with ophank
The open-loop gain, closed-loop gain and step response are all inferior to
those obtained with redschur( ). This emphasizes the point that one
cannot automatically assume that, because the error bound formula for
ophank( )is more attractive than that for redschur( ), the error itself
will be better for ophank( ).
© National Instruments Corporation
6-11
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
wtbalance
The next command examined is wtbalancewith the option "match".
[syscr,ysclr,hsv] = wtbalance(sys,sysc,"match",2)
Recall that this command should promote matching of closed-loop transfer
functions. The weighted Hankel singular values are:
1.486 4.513 × 10–1 8.420 × 10–2 5.869 × 1–2
The relative magnitudes suggest that reduction to order 2 will produce less
of an approximation error here (in the closed-loop transfer function) than a
reduction to this order through redschur( )or ophank( )(where the
implicit criterion is the unweighted error in approximating the controller
transfer function). Examination of Figures 6-9, 6-10, and 6-11 reveals that
far better approximation is now obtained.
Violation of the specification is to be observed in the open-loop gain.
Notice though that:
•
The error measure for wtbalancedoes not reflect the open-loop gain;
it reflects the closed-loop gain.
•
While the error in dB looks large, as an absolute value it is not
extremely so; wtbalanceworks with additive, not multiplicative
error.
Hence, it cannot be concluded that the algorithm is not working. Use of the
option "match spec"with wtbalancemight be conjectured as a device
for reducing the violation of the specification: one could introduce a weight
V(jw) emphasizing frequencies from 0.1 radians per second to 5 radians per
second.
For example,
(s + 0.1)(s + 10)
V(jω) = ----------------------------------------
(s + 1)(s + 1.4)
This would tend to force the closed-loop transfer functions derived from
the full-order and reduced controller to match better over this range;
because their absolute value is small there, they are approximately equal
to the open-loop gains which, accordingly, may be close. The flaw in this
reasoning is that a second-order controller, with four independent
parameters only, can only do so much, and the totality of designer demands
cannot be fully met.
Xmath Model Reduction Module
6-12
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
The following function calls produce Figure 6-9:
svalsrol = svplot(sys*syscr,w,{radians})
plot(svalsol, {keep})
f9=plot(wc, constr, {keep,!grid,
legend=["reduced","original","constrained"],
title="Open-Loop Gain Using wtbalance()"})
Figure 6-9. Open-Loop Gain with wtbalance
© National Instruments Corporation
6-13
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-10:
syscl = feedback(sysol);
sysolr=sys*syscr;
sysclr=feedback(sysolr);
compare(syscl,sysclr,w,{radians,type=5})
f10=plot({keep,legend=["original","reduced"]})
Figure 6-10. Closed-Loop Gain with wtbalance
Xmath Model Reduction Module
6-14
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-11:
tvec=0:(140/99):140;
compare(syscl,sysclr,tvec,{type=7})
f11=plot({keep,legend=["original","reduced"]})
Figure 6-11. Step Response with wtbalance
Figures 6-9, 6-10, and 6-11 are obtained for wtbalancewith the option
"input spec". Evidently, there is little difference between this and the
result with the option "match". One notices marginally better matching in
the region of interest (0.1 to 5 rad per second) at the expense of matching
at other frequencies. The weighted Hankel singular values again indicate
that it is reasonable to seek a second order controller.
© National Instruments Corporation
6-15
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-12:
vtf=poly([-0.1,-10])/poly([-1,-1.4])
[,sysv]=check(vtf,{ss,convert});
svalsv = svplot(sysv,w,{radians});
Figure 6-12. Frequency Response of the Weight V(jω)
Xmath Model Reduction Module
6-16
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-13:
[syscr,sysclr,hsv] = wtbalance(sys,sysc,
"input spec",2,sysv)
svalsrol = svplot(sys*syscr,w,{radians})
plot(svalsol, {keep})
f13=plot(wc,constr,{keep, !grid,
legend=["reduced","original","constrained"],
title="Open-Loop Gain with wtbal(), \"input spec\""})
Figure 6-13. Open-Loop Gain from wtbalance with "input spec"
© National Instruments Corporation
6-17
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-14:
syscl = feedback(sysol);
sysolr=sys*syscr;
sysclr=feedback(sysolr);
compare(syscl,sysclr,w,{radians,type=5})
f14=plot({keep,legend=["original","reduced"]})
Figure 6-14. System Singular Values of wtbalance with "input spec"
Xmath Model Reduction Module
6-18
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-15:
tvec=0:(140/99):140;
compare(syscl,sysclr,tvec,{type=7})
f15=plot({keep,legend=["original","reduced"]})
Figure 6-15. Step Response of wtbalance with "input spec"
© National Instruments Corporation
6-19
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
fracred
fracred, the next command examined, has four options—"right
The options "left stab", "right perf", and "left perf"all
produce instability. Given the relative magnitudes of the Hankel singular
values, this is perhaps not surprising. Figures 6-16, 6-17, and 6-18
illustrate the results using "right stab".
Generate Figure 6-16:
svalsrol = svplot(sys*syscr,w,{radians})
plot(svalsol, {keep})
f16=plot(wc,constr,{keep,!grid,
legend=["reduced","original","constrained"],
title="Open-Loop Gain Using fracred()"})
Figure 6-16. Open-Loop Gain Using fracred
Xmath Model Reduction Module
6-20
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-17:
syscl = feedback(sysol);
sysolr=sys*syscr;
sysclr=feedback(sysolr);
compare(syscl,sysclr,w,{radians,type=5})
f17=plot({keep,legend=["original","reduced"]})
Figure 6-17. Closed-Loop Response with fracred
© National Instruments Corporation
6-21
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
Generate Figure 6-18:
tvec=0:(140/99):140;
compare(syscl,sysclr,tvec,{type=7})
f18=plot({keep,legend=["original","reduced"]})
Figure 6-18. Step Response with fracred
The end result is comparable to that from wtbalance( )with option
"match".
We can create a table to examine the values of the Hankel singular values
based on different decompositions approaches.
set precision 3 # Optional:
set format fixed # we set a smaller precision here so we
could fit
# the table in the manual.
[syscr, hsvrs] = fracred(sys, Kr, Ke, "right stab",2);
[syscr, hsvls] = fracred(sys, Kr, Ke, "left stab",2);
[syscr, hsvrp] = fracred(sys, Kr, Ke, "right perf",2);
[syscr, hsvlp] = fracred(sys, Kr, Ke, "left perf",2);
Xmath Model Reduction Module
6-22
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Chapter 6
Tutorial
hsvtable = [...
"right stab:", string(hsvrs');
"left stab:", string(hsvls');
"right perf:", string(hsvrp');
"left perf:", string(hsvlp')]?
hsvtable (a rectangular matrix of strings) =
right stab:3.308 0.728 0.112 0.078 0.024 0.018 0.011 0.010
left stab:1.403 1.331 1.133 1.092 0.965 0.549 0.526 0.313
right perf:0.034 0.016 0.013 0.010 0.004 0.004 0.000 0.000
left perf:4.907 4.874 3.846 3.781 1.225 1.175 0.505 0.041
© National Instruments Corporation
6-23
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
A
Bibliography
[AnJ]
BDO Anderson and B. James, “Algorithm for multiplicative approximation of a stable
linear system,” in preparation.
[AnL89]
[AnM89]
[BoD87]
[Enn84]
BDO Anderson and Y. Liu, “Controller reduction: Concepts and approaches,” IEEE
Transactions on Automatic Control, Vol. 34, 1989, pp. 802–812.
BDO Anderson and J. B. Moore, Optimal Control: Linear Quadratic Methods,
Prentice-Hall Inc., Englewood Cliffs, NJ, 1989.
S. P. Boyd and J. Doyle, “Comparison of peak and RMS gains for discrete-time systems,”
System Control Letters, Vol. 9, No. 1, pp. 1–6, 1987.
D. F. Enns, “Model reduction with balanced realizations: an error bound and a frequency
weighted generalization,” Proceedings for the 23rd IEEE Conference on Decision and
Control, Las Vegas, November 1984, pp. 127–132.
[Enn84a]
[GCP88]
D.F. Enns, “Model reduction for control systems design,” PhD Thesis, Dept of
Aeronautics and Astronautics, Stanford University, CA, USA, 1984.
K. Glover, R. F. Curtain, and J. R. Partington, “Realisation and approximation of linear
infinite-dimensional systems with error bounds,” SIAM J Controls and Optimization,
Vol. 26, 1988, pp. 863–898.
[Glo84]
[Glo86]
[GrA86]
[GrA89]
K. Glover, “All optimal Hankel norm approximations of linear multivariable systems and
their L∞ error bounds,” Int J Controls, Vol. 39, 1984, pp. 1115–1193.
K. Glover, “Multiplicative approximation of linear multivariable systems with L∞ error
bounds,” Proceedings for American Controls Conference, Seattle, 1986, pp. 1705–1709.
M. Green and BDO Anderson, “The approximation of power spectra by phase matching,”
Proceedings for 25th CDC, 1986, pp. 1085–1090.
M. Green and BDO Anderson, “Model reduction by phase matching,” Mathematics of
Control, Signals, and Systems, Vol. 2, 1989, pp. 221–263.
© National Instruments Corporation
A-1
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Appendix A
[GrA90]
[Gre88]
[Gre88a]
[HiP90]
Bibliography
M. Green and BDO Anderson, “Generalized balanced stochastic truncation,”
Proceedings for 29th CDC, 1990.
M. Green, “Balanced stochastic realization,” Linear Algebra and Applications, Vol. 98,
1988, pp. 211–247.
M. Green, “A relative error bound for balanced stochastic truncation,” IEEE Transactions
on Automatic Control, Vol. 33, 1988, pp. 961–965.
D. Hinrichsen and A J Pritchard, “An improved error estimate for reduced-order models
of discrete-time systems,” IEEE Transactions on Automatic Control, Vol. 35, 1990,
pp. 317–320.
[LAL90]
[Lau80]
Y. Liu, BDO Anderson, and U-L Ly, “Coprime factorization controller reduction with
Bezout identity induced frequency weighting,” Automatica, Vol. 26, No. 2, 1990,
pp. 233–249.
A. J. Laub, “On computing ‘balancing’ transformations,” Proceedings on Joint American
Controls Conference, San Francisco, CA, 1980, Section FA8-E.
[LHPW87] A. J. Laub, M. T. Heath, C. C. Paige, and R. C. Ward, “Computation of system balancing
transformations and other applications of simultaneous diagonalizing algorithms,” IEEE
Transactions on Automatic Control, Vol. AC-32, 1987, pp. 115–122.
[LiA86]
[LiA89]
[LiA90]
[Moo81]
Y. Liu and BDO Anderson, “Controller reduction via stable factorization and balancing,”
Int. J. Control, Vol. 44, 1986, pp. 507–531.
Y. Liu and BDO Anderson, “Singular perturbation approximation of balanced systems,”
International Journal of Control, Vol. 50, 1989, pp. 1379–1405.
Y. Liu and BDO Anderson, “Frequency weighted controller reduction methods and loop
transfer recovery,” Automatica, Vol. 26, No. 3, pp. 487–489.
B.C. Moore, “Principal component analysis in linear systems: Controllability,
observability and model reduction,” IEEE Transactions on Automatic Control,
Vol. AC-26, No. 1, 1981, pp. 17–32.
[NJB84]
[PeS82]
C. N. Nett, C. A. Jacobson, and M. J. Balas, “A connection between state-space
and doubly coprime fractional representations,” IEEE Transactions on Automatic
Control, Vol. AC-29, 1984, pp. 831–832.
L. Pernebo, and L. M. Silverman, “Model reduction via balanced state space
pp. 382–387.
Xmath Model Reduction Module
A-2
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Appendix A
Bibliography
[SaC88]
[Saf87]
[SCL90]
M. G. Safonov and R. Y. Chiang, “Model reduction for robust control: a Schur
relative-error method,” Proceedings for the American Controls Conference, 1988,
pp. 1685–1690.
M. G. Safonov, “Imaginary-axis zeros in multivariable H• optimal control,” Modeling,
Robustness, and Sensitivity Reduction in Control, (Ed. R. F. Curtain), Springer Verlag,
Berlin, 1987.
M. G. Safonov, R. Y. Chiang, and DJN Limebeer, “Optimal Hankel model reduction
for nonminimal systems,” IEEE Transactions on Automatic Control, Vol. 35 No. 4,
pp. 496–502, 1990.
[Vid85]
M. Vidyasagar, Control Systems Synthesis: A Factorization Approach, MIT Press,
Cambridge, MA, 1985.
[WaS90]
W. Wang and M. G. Safonov, “A tighter relative error bound for balanced stochastic
truncation,” Systems and Control Letters, Vol. 14, 1990, pp. 307–317.
[WaS90a] W. Wang and M. G. Safonov, “Comparison between continuous and discrete-time model
truncation,” Proceedings for the 29th CDC, 1990.
[BBK88]
S. Boyd, V. Balakrishnan, and P. Kabamba, “A bisection method for computing the
L∞ norm of a transfer matrix and related problems,” Mathematical Controls, Signals,
and Systems, Vol. 2, No. 3, pp. 207–219, 1989.
[BeP79]
[BoB90]
A. Berman and R. J. Plemmons, Nonnegative Matrices in the Mathematical Sciences.
Computer Science and Applied Mathematics series, Academic Press, 1979.
S. Boyd and V. Balakrishnan. “A regularity result for the singular values of a transfer
matrix and a quadratically convergent algorithm for computing its L∞ norm.” Systems
Control Letters Vol. 15, pp. 1–7, 1990.
[BoB91]
[BH69]
S. Boyd and C. Barratt, Linear Controller Design: Limits of Performance, Prentice-Hall,
1991.
A. E. Bryson and Y. C. Ho, Applied Optimal Control, p. 149, Blaisdell Publishing Co.,
1969.
[DoS79]
[DoS81]
J. C. Doyle and G. Stein. “Robustness with Observers,” IEEE Transactions on Automatic
Control, August 1979.
J. C. Doyle and G. Stein. “Multivariable Feedback Design: Concepts for a
February 1981, pp 4–16.
© National Instruments Corporation
A-3
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Appendix A
[Doy82]
[DWS82]
Bibliography
J. C. Doyle. “Analysis of Feedback Systems with Structured Uncertainties.” IEEE
Proceedings, November 1982.
J. C. Doyle, J. E. Wall, and G. Stein. “Performance and Robustness Analysis for Structure
Uncertainties,” Proceedings IEEE Conference on Decision and Control, pp. 629–636,
1982.
[FaT88]
[FaT86]
[Fr87]
M. K. Fan and A. L. Tits, “m-form Numerical Range and the Computation of the
Structured Singular Value.” IEEE Transactions on Automatic Control, Vol. 33,
pp. 284–289, March 1988.
M. K. Fan and A. L. Tits. “Characterization and Efficient Computation of the Structured
Singular Value,” IEEE Transactions on Automatic Control, Vol. AC-31, pp. 734–743,
August 1986.
B. Francis, A Course in L• Control Theory, Springer-Verlag, Berlin-New York, 1987.
[FPGM87] D. S. Flamm, S. Boyd, G. Stein, and S. K. Mitter, “Tutorial Workshop on L∞ Control
Theory,” pre-conference workshop, Proceedings 26th IEEE Conference on Decision and
Control, December 1988.
[GD88]
K. Glover and J. C. Doyle, “State-space formulae for all stabilizing controllers that
satisfy an L∞ norm bound and relations to risk sensitivity,” Systems and Control Letters,
Vol. 11, pp. 167–172, 1988.
[DGKF89] J. C. Doyle, K. Glover, P. K. Khargonekar, and B. Francis, “State-space solutions to
standard H2 and L∞ control problems,” IEEE Transactions on Automatic Control,
Vol. AC-34, No. 8, pp. 831–847, August 1989.
[Gu80]
N. K. Gupta, “Frequency Shaping of Cost Functionals: An extension of LQG Design
Methods,” AIAA Journal of Guidance and Control, Vol. 3, No. 6, December 1980.
[ONR84]
ONR/Honeywell Workshop on Advances in Multivariable Control, Lecture Notes,
Minneapolis, MN, 1984.
[Osb60]
[Saf82]
E. E. Osborne, “On Preconditioning of Matrices,” JACM, 7:338–345, 1960.
M.G. Safonov, “Stability Margins of Diagonally Perturbed Multivariable Feedback
Systems,” IEEE Proceedings, 129-D:251–256, November 1982.
[SD83]
[SD84]
M. G. Safonov and J. C. Doyle. “Optimal Scaling for Multivariable Stability Margin
Singular Value Computation,” Proceedings of MECO/EES 1983, Symposium, 1983.
M. G. Safonov and J. C. Doyle, “Minimizing Conservativeness of Robust Singular
Values,” Multivariable Control, pp. 197–207, S. G. Tzafestas, ed. D. Reidel Publishing
Company, 1984.
Xmath Model Reduction Module
A-4
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Appendix A
Bibliography
[SLH81]
[SA88]
[Za81]
M. G. Safonov, A. J. Laub, and G. L. Hartmann, “Feedback Properties of Multivariable
Systems: The Role and Use of the Return Difference Matrix,” IEEE Transactions on
Automatic Control, Vol. AC-26, February 1981.
G. Stein and M. Athans. “The LQG/LTR Procedure for Multivariable Control Design,”
IEEE Transactions on Automatic Control, Vol. AC-32, No. 2, February 1987, pp.
105–114.
G. Zames, “Feedback and optimal sensitivity: model reference transformations,
multiplicative semi-norms, and approximate inverses,” IEEE Transactions on Automatic
Control, Vol. AC-26, pp. 301–320, 1981.
[KS72]
H. Kwakernaak and R. Sivan, Linear Optimal Control Systems, Wiley, 1972.
© National Instruments Corporation
A-5
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
B
Technical Support and
Professional Services
Visit the following sections of the National Instruments Web site at
ni.comfor technical support and professional services:
•
Support—Online technical support resources at ni.com/support
include the following:
–
Self-Help Resources—For answers and solutions, visit the
award-winning National Instruments Web site for software drivers
and updates, a searchable KnowledgeBase, product manuals,
step-by-step troubleshooting wizards, thousands of example
programs, tutorials, application notes, instrument drivers, and
so on.
–
Free Technical Support—All registered users receive free Basic
Service, which includes access to hundreds of Application
Engineers worldwide in the NI Discussion Forums at
ni.com/forums. National Instruments Application Engineers
make sure every question receives an answer.
For information about other technical support options in your
area, visit ni.com/servicesor contact your local office at
ni.com/contact.
•
•
Training and Certification—Visit ni.com/trainingfor
self-paced training, eLearning virtual classrooms, interactive CDs,
and Certification program information. You also can register for
instructor-led, hands-on courses at locations around the world.
System Integration—If you have time constraints, limited in-house
technical resources, or other project challenges, National Instruments
Alliance Partner members can help. To learn more, call your local
NI office or visit ni.com/alliance.
If you searched ni.comand could not find the answers you need, contact
your local office or NI corporate headquarters. Phone numbers for our
the Worldwide Offices section of ni.com/niglobalto access the branch
office Web sites, which provide up-to-date contact information, support
phone numbers, email addresses, and current events.
© National Instruments Corporation
B-1
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Index
controller robustness, 4-2
Symbols
*, 1-6
´, 1-6
diagnostic tools (NI resources), B-1
documentation
A
additive error, reduction, 2-1
algorithm
balanced stochastic truncation (bst), 3-4
fractional representation reduction, 4-18
Hankel multi-pass, 2-20
optimal Hankel norm reduction, 2-15
E
equality bounds, tight, 1-7
error bound, 2-7
all-pass transfer function, 1-6, 2-4
for balanced stochastic truncation, 3-8
for impulse responses, 2-3
for multiplicative Hankel reduction, 3-16
for stochastic truncation, 3-9
error formulas, ophank, 2-22
B
balance, 1-5, 2-4
algorithm, 1-11
balanced realization
definition, 1-10
additive, 2-1
internally balanced, 3-9
singular perturbation, 2-5
truncation, 2-2, 2-4
balanced stochastic truncation, 3-3
See also bst
balmoore, 1-5, 2-4
algorithm, 1-10
bst, 1-5, 1-14, 3-3
for Hankel norm approximation, 2-7
fracred, 1-5, 4-15
reduction, 4-18
frequency weighted error reduction, 1-1, 4-1
controller reduction, 4-2
C
compare, 1-5, 5-4
controller reduction, 4-2
with fractional representations, 4-5
© National Instruments Corporation
I-1
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
Index
G
N
grammians
National Instruments support and
services, B-1
controllability, 1-7
NI support and services, B-1
description of, 1-7
H
Hankel matrix, 1-9
O
Hankel norm approximation, 2-6
Hankel singular values, 1-8, 3-9, 5-1
hankelsv, 1-5, 5-1
algorithm, multipass, 2-20
help, technical support, B-1
ophank, 1-5, 2-14
discrete-time systems, 2-21
error formulas, 2-22
impulse response error, 2-22
multipass, 2-20
onepass, 2-18
unstable system approximation, 2-23
I
instrument drivers (NI resources), B-1
internal balancing, 1-10
P
Padé approximation, 1-5
K
KnowledgeBase, B-1
of balanced realization, 2-5
singular, 1-11, 2-6
phase function, 3-5, 3-15
L
lmax(A), 1-6
lyapunov, 1-8
M
MATRIXx Help, 1-3
Reli(A), 1-6
minimality requirements, 1-5
model reduction, schur, 2-5
Moore-Penrose pseudo-inverse, 1-6
mreduce, 1-5, 1-13, 2-10
S
mulhank, 1-5, 1-14, 3-14
multiplicative error, 1-1, 3-1
singular perturbation, 1-11
skipChks, 6-3
software (NI resources), B-1
Xmath Model Reduction Module
I-2
ni.com
Download from Www.Somanuals.com. All Manuals Search And Download.
Index
spectral factorization, 1-13
stability requirements, 1-5
stable, 1-5, 5-2
U
unstable zeros, 2-3
sup, 1-6
support, technical, B-1
Web resources, B-1
wtbalance, 1-5, 4-10
T
technical support, B-1
tight equality bounds, 1-7
training and certification (NI resources), B-1
transfer function, allpass, 1-6
troubleshooting (NI resources), B-1
truncate, 1-5, 2-4, 2-11
© National Instruments Corporation
I-3
Xmath Model Reduction Module
Download from Www.Somanuals.com. All Manuals Search And Download.
|