Robust Control Toolbox

For Use with MATLAB

Gary Balas
Richard Chiang
Andy Packard
Michael Safonov

Users Guide
Version 3

How to Contact The MathWorks:

Web
Newsgroup
www.mathworks.com/contact_TS.html Technical support
www.mathworks.com

comp.soft-sys.matlab

[email protected]
[email protected]
[email protected]
[email protected]
[email protected]

Product enhancement suggestions

Bug reports
Documentation error reports
Order status, license renewals, passcodes
Sales, pricing, and general information

508-647-7000 (Phone)
508-647-7001 (Fax)
The MathWorks, Inc.
3 Apple Hill Drive
Natick, MA 01760-2098
For contact information about worldwide offices, see the MathWorks Web site.

Robust Control Toolbox Users Guide

COPYRIGHT 2005-2006 by The MathWorks, Inc.
The software described in this document is furnished under a license agreement. The software may be used
or copied only under the terms of the license agreement. No part of this manual may be photocopied or reproduced in any form without prior written consent from The MathWorks, Inc.
FEDERAL ACQUISITION: This provision applies to all acquisitions of the Program and Documentation by,
for, or through the federal government of the United States. By accepting delivery of the Program or
Documentation, the government hereby agrees that this software or documentation qualifies as commercial
computer software or commercial computer software documentation as such terms are used or defined in
FAR 12.212, DFARS Part 227.72, and DFARS 252.227-7014. Accordingly, the terms and conditions of this
Agreement and only those rights specified in this Agreement, shall pertain to and govern the use,
modification, reproduction, release, performance, display, and disclosure of the Program and Documentation
by the federal government (or other entity acquiring for or through the federal government) and shall
supersede any conflicting contractual terms or conditions. If this License fails to meet the government's
needs or is inconsistent in any respect with federal procurement law, the government agrees to return the
Program and Documentation, unused, to The MathWorks, Inc.

Trademarks
MATLAB, Simulink, Stateflow, Handle Graphics, Real-Time Workshop, and xPC TargetBox are
registered trademarks of The MathWorks, Inc. Other product or brand names are trademarks
or registered trademarks of their respective holders.

Patents
The MathWorks products are protected by one or more U.S. patents. Please see
www.mathworks.com/patents for more information.

Revision History
September 2005
March 2006
September 2006

Online only
Online only
Online only

New for Version 3.0.2 (Release 14SP3)

Revised for Version 3.1 (Release 2006a)
Revised for Version 3.1.1 (Release 2006b)

Contents
Introduction

1
What Is the Robust Control Toolbox? . . . . . . . . . . . . . . . . . . . 1-2
Required Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-2
Modeling Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3
Example: ACC Benchmark Problem . . . . . . . . . . . . . . . . . . . . . 1-3
Worst-Case Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-7
Example: ACC Two-Cart Benchmark Problem . . . . . . . . . . . . . 1-7
Synthesis of Robust MIMO Controllers . . . . . . . . . . . . . . . . 1-10
Example: Designing a Controller with LOOPSYN . . . . . . . . . 1-10
Model Reduction and Approximation . . . . . . . . . . . . . . . . . . 1-14
Example: NASA HIMAT Controller Order Reduction . . . . . . 1-14
LMI Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-18
Extends Control System Toolbox . . . . . . . . . . . . . . . . . . . . . . 1-19
About the Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-20
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-21

Multivariable Loop Shaping

2
Tradeoff Between Performance and Robustness . . . . . . . . . 2-2
Norms and Singular Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
Typical Loop Shapes, S and T Design . . . . . . . . . . . . . . . . . . . 2-5
Singular Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-5

Guaranteed Gain/Phase Margins in MIMO Systems . . . . . . . . 2-9

Using LOOPSYN to Do H-Infinity Loop Shaping . . . . . . . .
Example: NASA HiMAT Loop Shaping . . . . . . . . . . . . . . . . . .
Design Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
MATLAB Commands for a LOOPSYN Design . . . . . . . . . . . . .

2-11
2-11
2-13
2-13

Using MIXSYN for H-Infinity Loop Shaping . . . . . . . . . . . . 2-17

Example: NASA HiMAT Design Using MIXSYN . . . . . . . . . . 2-18
Loop-Shaping Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-20

Model Reduction for Robust Control

3
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
Hankel Singular Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
Overview of Model Reduction Techniques . . . . . . . . . . . . . . . 3-4
Approximating Plant Models
Additive Error Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
Approximating Plant Models
Multiplicative Error Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-8
Using Modal Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
Rigid Body Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
Reducing Large-Scale Models . . . . . . . . . . . . . . . . . . . . . . . . . 3-13
Using Normalized Coprime Factor Methods . . . . . . . . . . . . 3-14
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-15

ii

Contents

Robustness Analysis

4
Uncertainty Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Uncertain Models of Dynamic Systems . . . . . . . . . . .
Creating Uncertain Parameters . . . . . . . . . . . . . . . . . . . . . . . . .
Quantifying Unmodeled Dynamics . . . . . . . . . . . . . . . . . . . . . . .

4-2
4-2
4-3
4-5

Robustness Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-8

Multiinput, Multioutput Robustness Analysis . . . . . . . . . . .
Adding Independent Input Uncertainty to Each Channel . . .
Closed-Loop Robustness Analysis . . . . . . . . . . . . . . . . . . . . . . .
Nominal Stability Margins . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Robustness of Stability Model Uncertainty . . . . . . . . . . . . . . .

4-12
4-13
4-15
4-17
4-18

Worst-Case Gain Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-20

Summary of Robustness Analysis Tools . . . . . . . . . . . . . . . . 4-23

H-Infinity and Mu Synthesis

5
H-Infinity Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-2
Performance As Generalized Disturbance Rejection . . . . . . . . . 5-2
Robustness in the H Framework . . . . . . . . . . . . . . . . . . . . . . . 5-8
Application of H and mu to Active Suspension Control .
Quarter Car Suspension Model . . . . . . . . . . . . . . . . . . . . . . . . .
Linear H Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . .
H Control Design 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
H Control Design 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Control Design via m-Synthesis . . . . . . . . . . . . . . . . . . . . . . . .

5-10
5-10
5-12
5-13
5-14
5-20

Functions for Control Design . . . . . . . . . . . . . . . . . . . . . . . . . 5-26

Interpretation of H-Infinity Norm . . . . . . . . . . . . . . . . . . . . . 5-28

iii

Norms of Signals and Systems . . . . . . . . . . . . . . . . . . . . . . . . . 5-28

Using Weighted Norms to Characterize Performance . . . . . . . 5-29
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-34

Building Uncertain Models

6
Introduction to Uncertain Atoms . . . . . . . . . . . . . . . . . . . . . . . 6-2
Uncertain Real Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-3
Uncertain LTI Dynamics Atoms . . . . . . . . . . . . . . . . . . . . . . . . 6-10
Complex Parameter Atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-13
Complex Matrix Atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-15
Unstructured Uncertain Dynamic Systems . . . . . . . . . . . . . . . 6-17

iv

Contents

Uncertain Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Uncertain Matrices from Uncertain Atoms . . . . . . .
Accessing Properties of a umat . . . . . . . . . . . . . . . . . . . . . . . . .
Row and Column Referencing . . . . . . . . . . . . . . . . . . . . . . . . . .
Matrix Operation on umat Objects . . . . . . . . . . . . . . . . . . . . . .
Substituting for Uncertain Atoms . . . . . . . . . . . . . . . . . . . . . . .

6-19
6-19
6-20
6-21
6-22
6-23

Uncertain State-Space Systems (uss) . . . . . . . . . . . . . . . . . . .

Creating Uncertain Systems . . . . . . . . . . . . . . . . . . . . . . . . . . .
Properties of uss Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Sampling Uncertain Systems . . . . . . . . . . . . . . . . . . . . . . . . . .
Feedback around Uncertain Plant . . . . . . . . . . . . . . . . . . . . . .
Interpreting Uncertainty in Discrete Time . . . . . . . . . . . . . . .
Lifting a ss to a uss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Handling Delays in uss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6-25
6-25
6-26
6-27
6-28
6-30
6-31
6-31

Uncertain frd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Uncertain Frequency Response Objects . . . . . . . . . .
Properties of ufrd objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Interpreting Uncertainty in Discrete Time . . . . . . . . . . . . . . .
Lifting an frd to a ufrd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Handling Delays in ufrd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6-33
6-33
6-33
6-36
6-36
6-36

Basic Control System Toolbox and

MATLAB Interconnections . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-37
Simplifying Representation of Uncertain Objects . . . . . . . 6-38
Effect of Autosimplify Property . . . . . . . . . . . . . . . . . . . . . . . . . 6-39
Direct Use of simplify . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-41
Sampling Uncertain Objects . . . . . . . . . . . . . . . . . . . . . . . . . .
Generating One Sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Generating Many Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Sampling ultidyn Atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6-42
6-42
6-42
6-43

Substitution by usubs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-46

Specifying the Substitution with Structures . . . . . . . . . . . . . . 6-47
Nominal and Random Values . . . . . . . . . . . . . . . . . . . . . . . . . . 6-47
Array Management for Uncertain Objects . . . . . . . . . . . . . .
Referencing Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Arrays with stack and cat Functions . . . . . . . . . . . .
Creating Arrays by Assignment . . . . . . . . . . . . . . . . . . . . . . . .
Binary Operations with Arrays . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Arrays with usample . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Arrays with usubs . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Arrays with gridureal . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Arrays with repmat . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating Arrays with repsys . . . . . . . . . . . . . . . . . . . . . . . . . . .
Using permute and ipermute . . . . . . . . . . . . . . . . . . . . . . . . . .

6-49
6-49
6-50
6-52
6-53
6-53
6-55
6-56
6-57
6-58
6-58

Decomposing Uncertain Objects (for Advanced Users) . .

Normalizing Functions for Uncertain Atoms . . . . . . . . . . . . . .
Properties of the Decomposition . . . . . . . . . . . . . . . . . . . . . . . .
Syntax of lftdata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Advanced Syntax of lftdata . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6-60
6-60
6-61
6-62
6-64

Generalized Robustness Analysis

7
Introduction to Generalized Robustness Analysis . . . . . . . . 7-2
Robust Stability Margin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
Robust Performance Margin . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-5
Worst-Case Gain Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-6

Introduction to Linear Matrix Inequalities

8
Linear Matrix Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
LMI Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
LMIs and LMI Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
The Three Generic LMI Problems . . . . . . . . . . . . . . . . . . . . . . . . 8-5
Further Mathematical Background . . . . . . . . . . . . . . . . . . . . . 8-9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-10

The LMI Lab

9
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
Some Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
Overview of the LMI Lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-5
Specifying a System of LMIs . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-7
A Simple Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-8
Initializing the LMI System . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-10

vi

Contents

Specifying the LMI Variables . . . . . . . . . . . . . . . . . . . . . . . . . .

Specifying Individual LMIs . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Specifying LMIs with the LMI Editor . . . . . . . . . . . . . . . . . . . .
How It All Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9-10
9-13
9-15
9-19

Querying the LMI System Description . . . . . . . . . . . . . . . . . 9-20

lmiinfo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-20
lminbr and matnbr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-20
LMI Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-21
From Decision to Matrix Variables and Vice Versa . . . . . . 9-27
Validating Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-28
Modifying a System of LMIs . . . . . . . . . . . . . . . . . . . . . . . . . . .
Deleting an LMI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Deleting a Matrix Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instantiating a Matrix Variable . . . . . . . . . . . . . . . . . . . . . . . .

9-29
9-29
9-29
9-30

Advanced Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Structured Matrix Variables . . . . . . . . . . . . . . . . . . . . . . . . . . .
Complex-Valued LMIs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Specifying cTx Objectives for mincx . . . . . . . . . . . . . . . . . . . . .
Feasibility Radius . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Well-Posedness Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Semi-Definite B(x) in gevp Problems . . . . . . . . . . . . . . . . . . . .
Efficiency and Complexity Issues . . . . . . . . . . . . . . . . . . . . . . .
Solving M + PTXQ + QTXTP < 0 . . . . . . . . . . . . . . . . . . . . . . . .

9-32
9-32
9-34
9-37
9-38
9-39
9-40
9-41
9-41

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-43

Function Reference

10
Functions By Category . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-2
Uncertain Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-4

vii

Uncertain Matrices and Systems . . . . . . . . . . . . . . . . . . . . . . . 10-4

Manipulation of Uncertain Models . . . . . . . . . . . . . . . . . . . . . . 10-4
Interconnection of Uncertain Models . . . . . . . . . . . . . . . . . . . . 10-6
Model Order Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-6
Robustness and Worst-Case Analysis . . . . . . . . . . . . . . . . . . . . 10-7
Robustness Analysis for
Parameter-Dependent Systems (P-Systems) . . . . . . . . . . . . . . 10-8
Controller Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-8
m-Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-10
Sampled-Data Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-10
Gain Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-10
Frequency-Response Data (FRD) Models . . . . . . . . . . . . . . . . 10-10
Supporting Utilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-12
LMIs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-12
Functions Alphabetical List . . . . . . . . . . . . . . . . . . . . . . . 10-15

Index

viii Contents

1
Introduction

What Is the Robust Control Toolbox?

(p. 1-2)

Tools for analysis and design of uncertain control systems

Modeling Uncertainty (p. 1-3)

How to model uncertain parameters, matrices, and

systems

Worst-Case Performance (p. 1-7)

Computing worst-case values for uncertain systems

Synthesis of Robust MIMO Controllers Designing controllers that optimize worst-case

(p. 1-10)
performance and maximize stability margins (H, H2,
loop-shaping, mixed sensitivity, -synthesis, and LMI
optimization)
Model Reduction and Approximation
(p. 1-14)

Simplifying high-order LTI plant and controller models

LMI Solvers (p. 1-18)

Tools for solving linear matrix inequalities

Extends Control System Toolbox

(p. 1-19)

How the Robust Control Toolbox uses LTI objects and

commands from the Control System Toolbox

About the Authors (p. 1-20)

Brief biographies of the toolbox authors

Bibliography (p. 1-21)

A list of sources on robust control theory

Introduction

What Is the Robust Control Toolbox?

The Robust Control Toolbox is a collection of functions and tools that help you
to analyze and design multiinput-multioutput (MIMO) control systems with
uncertain elements. You can build uncertain LTI system models containing
uncertain parameters and uncertain dynamics. You get tools to analyze MIMO
system stability margins and worst case performance.
The Robust Control Toolbox includes a selection of control synthesis tools that
compute controllers that optimize worst-case performance and identify
worst-case parameter values. The toolbox lets you simplify and reduce the
order of complex models with model reduction tools that minimize additive and
multiplicative error bounds. It provides tools for implementing advanced
robust control methods like H, H2, linear matrix inequalities (LMI), and
-synthesis robust control. You can shape MIMO system frequency responses
and design uncertainty tolerant controllers.

Required Software
The Robust Control Toolbox requires that you have installed the Control
System Toolbox.

1-2

Modeling Uncertainty

Modeling Uncertainty
At the heart of robust control is the concept of an uncertain LTI system. Model
uncertainty arises when system gains or other parameters are not precisely
known, or can vary over a given range. Examples of real parameter
uncertainties include uncertain pole and zero locations and uncertain gains.
You can also have unstructured uncertainties, by which is meant complex
parameter variations satisfying given magnitude bounds.
With the Robust Control Toolbox you can create uncertain LTI models as
MATLAB objects specifically designed for robust control applications. You can
build models of complex systems by combining models of subsystems using
addition, multiplication, and division, as well as with Control System Toolbox
commands like feedback and lft.

Example: ACC Benchmark Problem

For instance, consider the two-cart ACC Benchmark system [13] consisting of
two frictionless carts connected by a spring as shown below.

Position
Measurement
- x2 = y1

Position

Control
Force
u1

- x1

m1

e
%e

e%
e

e

Spring k

m2
m

ACC Benchmark Problem

The system has the block diagram model shown below, where the individual
carts have the respective transfer functions
1
G1(s) = -------------- ;
2
m1s

1
G2(s) = -------------- .
2
m2 s

The parameters m1, m2, and k are uncertain, equal to one plus or minus 20%:

1-3

Introduction

m1 = 10.2
m2 = 10.2
k = 10.2
P(s)
+

u1

f1

G1(s)

x1

f2

Cart 1

G2(s)

F(s)

x2

Cart 2
+

u2

Spring
k

ACC Benchmark Two-Cart System Block Diagram y1 = P(s) u1

The upper dashed-line block has transfer function matrix

F(s) =

0
1
0 G2 ( s )
1 1 +
G1 ( s )
1

This code builds the uncertain system model P shown above.

% Create the uncertain real parameters m1, m2, & k
m1 = ureal('m1',1,'percent',20);
m2 = ureal('m2',1,'percent',20);
k = ureal('k',1,'percent',20);
s = zpk('s'); % create the Laplace variable s
G1 = ss(1/s^2)/m1; % Cart 1
G2 = ss(1/s^2)/m2; % Cart 2
% Now build F and P
F = [0;G1]*[1 -1]+[1;-1]*[0,G2];
P = lft(F,k) % close the loop with the spring k

1-4

y1

y2

Modeling Uncertainty

The variable P is a SISO uncertain state-space (USS) object with four states
and three uncertain parameters, m1, m2, and k. You can recover the nominal
plant with the command
zpk(P.nominal)

which returns
Zero/pole/gain:
1
-------------s^2 (s^2 + 2)

If the uncertain model P(s) has LTI negative feedback controller

3

100 ( s + 1 ) C ( s ) = ---------------------------------( 0.001s + 1 ) 3

u1
C

y1

then you can form the controller and the closed-loop system y1 = T(s) u1 and
view the closed-loop system's step response on the time interval from t=0 to
t=0.1 for a Monte Carlo random sample of five combinations of the three
uncertain parameters k, m1, and m2 using this code:
C=100*ss((s+1)/(.001*s+1)^3); % LTI controller
T=feedback(P*C,1); % closed-loop uncertain system
step(usample(T,5),.1);

1-5

Introduction

The resulting plot is shown below.

Step Response

1.4
1.2

Amplitude

1
0.8
0.6
0.4
0.2
0
0

0.02

0.04
0.06
Time (sec)

0.08

0.1

Monte Carlo Sampling of Uncertain System's Step Response

1-6

Worst-Case Performance

Worst-Case Performance
To be robust, your control system should meet your stability and performance
requirements for all possible values of uncertain parameters. Monte Carlo
parameter sampling via usample can be used for this purpose as shown in
Monte Carlo Sampling of Uncertain Systems Step Response on page 1-6, but
Monte Carlo methods are inherently hit or miss. With Monte Carlo methods,
you might need to take an impossibly large number of samples before you hit
upon or near a worst-case parameter combination.
The Robust Control Toolbox gives you a powerful assortment of robustness
analysis commands that let you directly calculate upper and lower bounds on
worst-case performance without random sampling:
Worst-Case Robustness Analysis Commands
loopmargin

Comprehensive analysis of feedback loop

loopsens

Sensitivity functions of feedback loop

ncfmargin

Normalized coprime stability margin of feedback loop

robustperf

Robust performance of uncertain systems

robuststab

Stability margins of uncertain systems

wcgain

Worst-case gain of an uncertain system

wcmargin

Worst-case gain/phase margins for feedback loop

wcsens

Worst-case sensitivity functions of feedback loop

Example: ACC Two-Cart Benchmark Problem

Returning to the Example: ACC Benchmark Problem on page 1-3, the closed
loop system is
T=feedback(P*C,1); % Closed-loop uncertain system

This uncertain state-space model T has three uncertain parameters, k, m1, and
m2, each equal to 120% uncertain variation. To analyze whether the
closed-loop system T is robustly stable for all combinations of values for these
three parameters, you can execute the commands

1-7

Introduction

[StabilityMargin,Udestab,REPORT] = robuststab(T);
REPORT

This displays the REPORT

Uncertain System is robustly stable to modeled uncertainty.
-- It can tolerate up to 311% of modeled uncertainty.
-- A destabilizing combination of 500% the modeled uncertainty exists,
causing an instability at 44.3 rad/s.

The report tells you that the control system is robust for all parameter
variations in the 20% range, and that the smallest destabilizing combination
of real variations in the values k, m1, and m2 has sizes somewhere between
311% and 500% greater than 20%, i.e., between 62.2% and 100%. The value
Udestab returns an estimate of the 500% destabilizing parameter variation
combination:
Udestab
k:
m1:
m2:

=
1.2174e-005
1.2174e-005
2.0000.

Bode Diagram

0.5

Magnitude (dB)

0.5

1.5

Twc worstcase
Trand random samples

10

10
Frequency (rad/sec)

Uncertain System Closed-Loop Bode Plots

1-8

Worst-Case Performance

You have a comfortable safety margin of between 311% to 500% larger than the
anticipated 20% parameter variations before the closed loop goes unstable.
But how much can closed-loop performance deteriorate for parameter
variations constrained to lie strictly within the anticipated 20% range? The
following code computes worst-case peak gain of T, and estimates the frequency
and parameter values at which the peak gain occurs.
[PeakGain,Uwc] = wcgain(T);
Twc=usubs(T,Uwc);
% Worst case closed-loop system T
Trand=usample(T,4);
% 4 random samples of uncertain system T
bodemag(Twc,'r',Trand,'b-.',{.5,50}); % Do bode plot
legend('T_{wc}
- worst-case',...
'T_{rand} - random samples',3);

The resulting plot is shown in Uncertain System Closed-Loop Bode Plots on

page 1-8.

1-9

Introduction

Synthesis of Robust MIMO Controllers

You can design controllers for multiinput-multioutput (MIMO) LTI models
with your Robust Control Toolbox using the following commands:
Robust Control Synthesis Commands
h2hinfsyn

Mixed H2/H controller synthesis

h2syn

H2 controller synthesis

hinfsyn

H controller synthesis

loopsyn

H loop-shaping controller synthesis

ltrsyn

Loop-transfer recovery controller synthesis

mixsyn

H mixed-sensitivity controller synthesis

ncfsyn

H normalized coprime factor controller synthesis

sdhinfsyn

Sampled-data H controller synthesis

Example: Designing a Controller with LOOPSYN

One of the most powerful yet simple controller synthesis tools is loopsyn.
Given an LTI plant, you specify the shape of the open-loop systems frequency
response plot that you want, then loopsyn computes a stabilizing controller
that best approximates your specified loop shape.
For example, consider the 2-by-2 NASA HiMAT aircraft model (Safonov, Laub,
and Hartmann [8]) depicted in the following figure. The control variables are
elevon and canard actuators (e and c). The output variables are angle of
attack () and attitude angle (). The model has six states, viz.

1-10

Synthesis of Robust MIMO Controllers

x =

x1

x2

x3
x4

x5

xe

x6

where x e and x are elevator and canard actuator states.

Aileron
Canard
Flap
Rudder

Elevator
Elevon
x-axis
0

Velocity
Horizontal

Aircraft Configuration and Vertical Plane Geometry

1-11

Introduction

You can enter the state-space matrices for this model with the following code:
% NASA HiMAT model G(s)
ag =[ -2.2567e-02 -3.6617e+01 -1.8897e+01 -3.2090e+01
9.2572e-05 -1.8997e+00
1.2338e-02

1.1720e+01 -2.6316e+00

1.0000e+00

bg = [ 0

0;

0;

0;

0;

30
0

2.2396e+01;

0;

0 -3.0000e+01

0;

0 -3.0000e+01];

0;
30];
1

0;

0];

8.7582e-04 -3.1604e+01
0

cg = [ 0
dg = [ 0

3.2509e+00 -7.6257e-01;

9.8312e-01 -7.2562e-04 -1.7080e-01 -4.9652e-03;

0;
0];

G=ss(ag,bg,cg,dg);

To design a controller to shape the frequency response (sigma) plot so that the
system has approximately a bandwidth of 10 rad/s, you can set as your target
desired loop shape Gd(s)=10/s, then use loopsyn(G,Gd) to find a loop-shaping
controller for G that optimally matches the desired loop shape Gd by typing
s=zpk('s'); w0=10; Gd=w0/(s+.001);
[K,CL,GAM]=loopsyn(G,Gd); % Design a loop-shaping controller K
% Plot the results
sigma(G*K,'r',Gd,'k-.',Gd/GAM,'k:',Gd*GAM,'k:',{.1,30})
figure ;T=feedback(G*K,eye(2));
sigma(T,ss(GAM),'k:',{.1,30});grid

The value of = GAM returned is an indicator of the accuracy to which the

optimal loop shape matches your desired loop shape and is an upper bound on
the resonant peak magnitude of the closed-loop transfer function
T=feedback(G*K,eye(2)). In this case, = 1.6024 = 4 dB see the next figure.

1-12

Synthesis of Robust MIMO Controllers

sigma plot
40

magnitude, db

30

desired loop shape

G , db
d
achieved loop shape
GK

20
10

Gd /, db

, db

10
20 1
10

10

10

frequency rad/s
MIMO Robust Loop Shaping with loopsyn(G,Gd)

The achieved loop shape matches the desired target Gd to within about dB.

1-13

Introduction

Model Reduction and Approximation

Complex models are not always required for good control. Unfortunately,
however, optimization methods (including methods based on H, H2, and
-synthesis optimal control theory) generally tend to produce controllers with
at least as many states as the plant model. For this reason, the Robust Control
Toolbox offers you an assortment of model-order reduction commands that help
you to find less complex low-order approximations to plant and controller
models.
Model Reduction Commands
reduce

Main interface to model approximation algorithms

balancmr

Balanced truncation model reduction

bstmr

Balanced stochastic truncation model reduction

hankelmr

Optimal Hankel norm model approximations

modreal

State-space modal truncation/realization

ncfmr

Balanced normalized coprime factor model reduction

schurmr

Schur balanced truncation model reduction

slowfast

State-space slow-fast decomposition

stabproj

State-space stable/antistable decomposition

imp2ss

Impulse response to state-space approximation

Among the most important types of model reduction methods are minimize
bounds methods on additive, multiplicative, and normalized coprime factor
(NCF) model error. You can access all three of these methods using the
command reduce.

Example: NASA HIMAT Controller Order Reduction

For instance, the NASA HiMAT model considered in the last section has eight
states, and the optimal loop-shaping controller turns out to have 16 states.
Using model reduction, you can remove at least some of the states without
appreciably affecting stability or closed-loop performance. For controller order

1-14

Model Reduction and Approximation

reduction, the NCF model reduction is particularly useful, and it works equally
well with controllers that have poles anywhere in the complex plane.
For the NASA HiMAT design in the last section, you can type
hankelsv(K,'ncf','log');

which displays a logarithmic plot of the NCF Hankel singular values see the
figure below.
Hankel SV of Coprime Factors [Nl Ml]

10

10

10

log

10

10

10

10

12

10

14

10

8
Order

10

12

14

16

Hankel Singular Values of Coprime Factorization of K

Theory says that, without danger of inducing instability, you can confidently
discard at least those controller states that have NCF Hankel singular values
that are much smaller than ncfmargin(G,K).
Compute ncfmargin(G,K) and add it to your Hankel singular values plot.
hankelsv(K,'ncf','log');v=axis;

1-15

Introduction

hold on; plot(v(1:2), ncfmargin(G,K)*[1 1],'--'); hold off

Hankel SV of Coprime Factors [Nl Ml]

10

10

10

log

10

10

10

10

12

10

14

10

8
Order

10

12

14

16

Five of the 16 NCF Hankel Singular Values of HiMAT Controller K Are Small
Compared to ncfmargin(G,K)

In this case, you can safely discard five of the 16 states of K and compute an
11-state reduced controller by typing
K1=reduce(K,11,'errortype','ncf');

The result is plotted in the following figure.

sigma(G*K1,'b',G*K,'r--',{.1,30});

1-16

Model Reduction and Approximation

50

singular values: openloop

GK1 11state reduced K1
GK 16state original K

40

Singular Values (dB)

30

20

10

10

20 1
10

10
Frequency (rad/sec)

10

HiMAT with 11-State Controller K1 vs. Original 16-State Controller K

The picture above shows that low-frequency gain is decreased considerably for
inputs in one vector direction. Although this does not affect stability, it affects
performance. If you wanted to better preserve low-frequency performance, you
would discard fewer than five of the 16 states of K.

1-17

Introduction

LMI Solvers
At the core of many emergent robust control analysis and synthesis routines
are powerful general purpose functions for solving a class of convex nonlinear
programming problems known as linear matrix inequalities. The LMI
capabilities are invoked by Robust Control Toolbox functions that evaluate
worst-case performance, as well as functions like hinfsyn and h2hinfsyn.
Some of the main functions that help you access the LMI capabilities of the
Robust Control Toolbox are shown in the table below.
Specification of LMIs
lmiedit

GUI for LMI specification

setlmis

Initialize the LMI description

lmivar

Define a new matrix variable

lmiterm

Specify the term content of an LMI

newlmi

Attach an identifying tag to new LMIs

getlmis

Get the internal description of the LMI system

LMI Solvers
feasp

Test feasibility of a system of LMIs

gevp

Minimize generalized eigenvalue with LMI constraints

mincx

Minimize a linear objective with LMI constraints

dec2mat

Convert output of the solvers to values of matrix variables

Evaluation of LMIs/Validation of Results

evallmi

Evaluate for given values of the decision variables

showlmi

Return the left and right sides of an evaluated LMI

Complete documentation is available in The LMI Lab on page 9-1.

1-18

Extends Control System Toolbox

Extends Control System Toolbox

The Robust Control Toolbox is designed to work with the Control System
Toolbox. The Robust Control Toolbox extends the capabilities of the Control
System Toolbox and leverages the LTI and plotting capabilities of the Control
System Toolbox. The major analysis and synthesis commands in the Robust
Control Toolbox accept LTI object inputs, e.g., LTI state-space systems
produced by commands such as
G=tf(1,[1 2 3])
G=ss([-1 0; 0 -1], [1;1],[1 1],3)

The uncertain system (USS) objects in the Robust Control Toolbox generalize
the Control System Toolbox LTI SS objects and help ease the task of analyzing
and plotting uncertain systems. You can do many of the same algebraic
operations on uncertain systems that are possible for LTI objects (multiply,
add, invert), and the Robust Control Toolbox provides USS uncertain system
extensions of Control System Toolbox interconnection and plotting functions
like feedback, lft, and bode.

1-19

Introduction

About the Authors

Professor Andy Packard is with the Faculty of Mechanical Engineering at the
University of California, Berkeley. His research interests include robustness
issues in control analysis and design, linear algebra and numerical algorithms
in control problems, applications of system theory to aerospace problems, flight
control, and control of fluid flow.
Professor Gary Balas is with the Faculty of Aerospace Engineering &
Mechanics at the University of Minnesota and is president of MUSYN Inc. His
research interests include aerospace control systems, both experimental and
theoretical.
Dr. Michael Safonov is with the Faculty of Electrical Engineering at the
University of Southern California. His research interests include control and
decision theory.
Dr. Richard Chiang is employed by Boeing Satellite Systems, El Segundo, CA.
He is a Boeing Technical Fellow and has been working in the aerospace
industry over 25 years. In his career, Richard has designed three flight control
laws, 12 spacecraft attitude control laws, and three large space structure
vibration controllers, using modern robust control theory and the tools he built
in this toolbox. His research interests include robust control theory, model
reduction, and in-flight system identification. Working in industry instead of
academia, Richard serves a unique role in our team, bridging the gap between
theory and reality.
The linear matrix inequality (LMI) portion of the Robust Control Toolbox was
developed by these two authors:
Dr. Pascal Gahinet is employed by The MathWorks. His research interests
include robust control theory, linear matrix inequalities, numerical linear
algebra, and numerical software for control.
Professor Arkadi Nemirovski is with the Faculty of Industrial Engineering
and Management at Technion, Haifa, Israel. His research interests include
convex optimization, complexity theory, and nonparametric statistics.

1-20

Bibliography

Bibliography
[1] Boyd, S.P., El Ghaoui, L., Feron, E., and Balakrishnan, V., Linear Matrix
Inequalities in Systems and Control Theory, Philadelphia, PA, SIAM, 1994.
[2] Dorato, P. (editor), Robust Control, New York, IEEE Press, 1987.
[3] Dorato, P., and Yedavalli, R.K. (editors), Recent Advances in Robust
Control, New York, IEEE Press, 1990.
[4] Doyle, J.C., and Stein, G., Multivariable Feedback Design: Concepts for a
Classical/Modern Synthesis, IEEE Trans. on Automat. Contr., 1981, AC-26(1),
pp. 4-16.
[5] El Ghaoui, L., and Niculescu, S., Recent Advances in LMI Theory for
Control, Philadelphia, PA, SIAM, 2000.
[6] Lehtomaki, N.A., Sandell, Jr., N.R., and Athans, M., Robustness Results in
Linear-Quadratic Gaussian Based Multivariable Control Designs, IEEE
Trans. on Automat. Contr., Vol. AC-26, No. 1, Feb. 1981, pp. 75-92.
[7] Safonov, M.G., Stability and Robustness of Multivariable Feedback
Systems, Cambridge, MA, MIT Press, 1980.
[8] Safonov, M.G., Laub, A.J., and Hartmann, G., Feedback Properties of
Multivariable Systems: The Role and Use of Return Difference Matrix, IEEE
Trans. of Automat. Contr., 1981, AC-26(1), pp. 47-65.
[9] Safonov, M.G., Chiang, R.Y., and Flashner, H., H Control Synthesis for a
Large Space Structure, Proc. of American Contr. Conf., Atlanta, GA, June
15-17, 1988.
[10] Safonov, M.G., and Chiang, R.Y., CACSD Using the State-Space L
Theory A Design Example, IEEE Trans. on Automatic Control, 1988,
AC-33(5), pp. 477-479.
[11] Sanchez-Pena, R.S., and Sznaier, M., Robust Systems Theory and
Applications, New York, Wiley, 1998.
[12] Skogestad, S., and Postlethwaite, I., Multivariable Feedback Control, New
York, Wiley, 1996.
[13] Wie, B., and Bernstein, D.S., A Benchmark Problem for Robust Controller
Design, Proc. American Control Conf., San Diego, CA, May 23-25, 1990; also
Boston, MA, June 26-28, 1991.

1-21

Introduction

[14] Zhou, K., Doyle, J.C., and Glover, K., Robust and Optimal Control,
Englewood Cliffs, NJ, Prentice Hall, 1996.

1-22

2
Multivariable Loop
Shaping

Tradeoff Between Performance and

Robustness (p. 2-2)

Singular values, disturbance attenuation, and robustness

Typical Loop Shapes, S and T Design

(p. 2-5)

Sigma plot design specifications (loop shape, sensitivity,

and complementary sensitivity)

Using LOOPSYN to Do H-Infinity Loop Synthesis of MIMO H optimal loop-shaping controllers:

Shaping (p. 2-11)
LOOPSYN method and HiMAT aircraft example
Using MIXSYN for H-Infinity Loop
Shaping (p. 2-17)

Synthesis of MIMO H optimal loop-shaping controllers:

MIXSYN method and HiMAT aircraft example

Loop-Shaping Commands (p. 2-20)

Summary of loop-shaping commands and utilities

Multivariable Loop Shaping

Tradeoff Between Performance and Robustness

When the plant modeling uncertainty is not too big, you can design high-gain,
high-performance feedback controllers. High loop gains significantly larger
than 1 in magnitude can attenuate the effects of plant model uncertainty and
reduce the overall sensitivity of the system to plant noise. But if your plant
model uncertainty is so large that you do not even know the sign of your plant
gain, then you cannot use large feedback gains without the risk that the system
will become unstable. Thus, plant model uncertainty can be a fundamental
limiting factor in determining what can be achieved with feedback.

Multiplicative Uncertainty: Given an approximate model of the plant G0 of

a plant G, the multiplicative uncertainty M of the model G0 is defined
1
as M = G 0 ( G G 0 ) or, equivalently, G = ( I + M )G 0 .
Plant model uncertainty arises from many sources. There might be small
unmodeled time delays or stray electrical capacitance. Imprecisely understood
actuator time constants or, in mechanical systems, high-frequency torsional
bending modes and similar effects can be responsible for plant model
uncertainty. These types of uncertainty are relatively small at lower
frequencies and typically increase at higher frequencies.
In the case of single-input/single-output (SISO) plants, the frequency at which
there are uncertain variations in your plant of size|M|=2 marks a critical
threshold beyond which there is insufficient information about the plant to
reliably design a feedback controller. With such a 200% model uncertainty, the
model provides no indication of the phase angle of the true plant, which means
that the only way you can reliably stabilize your plant is to ensure that the loop
gain is less than 1. Allowing for an additional factor of 2 margin for error, your
control system bandwidth is essentially limited to the frequency range over
which your multiplicative plant uncertainty M has gain magnitude |M|<1.

Norms and Singular Values

For MIMO systems the transfer functions are matrices, and relevant measures
of gain are determined by singular values, H, and H2 norms, which are
defined as follows:

2-2

Tradeoff Between Performance and Robustness

H2 and H norms: The H2-norm is the energy of the impulse response of

plant G. The H-norm is the peak gain of G across all frequencies and all
input directions.

Another important concept is the notion of singular values.

mn

Singular Values: The singular values of a rank r matrix A C

,
denoted i , are the nonnegative square roots of the eigenvalues of A * A
ordered such that 1 2 p > 0, p min { m , n } .
If r < p then there are p r zero singular values, i.e.,
r + 1 = r + 2 = = p = 0
The greatest singular value 1 is sometimes denoted ( A ) = 1 .
When A is a square n-by-n matrix, then the nth singular value (i.e., the least

singular value) is denoted ( A ) = n .

Properties of Singular Values

Some useful properties of singular values are
Ax
( A ) = max x C n -----------x
Ax
( A ) = min x C n -----------x
These properties are especially important because they establish that the
greatest and least singular values of a matrix A are the maximal and minimal
gains of the matrix as the input vector x varies over all possible directions.
For stable continuous-time LTI systems G ( s ) , the H2-norm and the H-norms
are defined terms of the frequency-dependent singular values of G ( j ) :

2-3

Multivariable Loop Shaping

H2-norm:

G 2=

1
-----2

( i ( G ( j ) ) )

i=1

H-norm

2-4

G = sup ( G ( j ) )

( sup: the least upper bound ) :

Typical Loop Shapes, S and T Design

Typical Loop Shapes, S and T Design

Consider the multivariable feedback control system shown in the figure below.
In order to quantify the multivariable stability margins and performance of
such systems, you can use the singular values of the closed-loop transfer
function matrices from r to each of the three outputs e, u, and y, viz.
def

S( s ) = ( I + L ( s ) )

def

def

R( s ) = K ( s ) ( I + L ( s ) )
T( s ) = L ( s ) ( I + L ( s ) )

= I S(s)

where the L(s) is the loop transfer function matrix

L ( s ) = G ( s )K ( s )

command
r

error
e
-

(2-1)

K(s)

control
u

d
plant
disturbance
effects
+
+

G(s)

output
y

"controller"
"plant"

Block Diagram of the Multivariable Feedback Control System

The two matrices S(s) and T(s) are known as the sensitivity function and
complementary sensitivity function, respectively. The matrix R(s) has no
common name. The singular value Bode plots of each of the three transfer
function matrices S(s), R(s), and T(s) play an important role in robust
multivariable control system design. The singular values of the loop transfer
function matrix L(s) are important because L(s) determines the matrices S(s)
and T(s).

Singular Values
The singular values of S(j) determine the disturbance attenuation, because
S(s) is in fact the closed-loop transfer function from disturbance d to plant
output y see Block Diagram of the Multivariable Feedback Control System

2-5

Multivariable Loop Shaping

on page 2-5. Thus a disturbance attenuation performance specification can be

written as
1

( S ( j ) ) W 1 ( j )

(2-2)

where W 1 ( j ) is the desired disturbance attenuation factor. Allowing

W 1 ( j )

to depend on frequency enables you to specify a different

attenuation factor for each frequency .

The singular value Bode plots of R(s) and of T(s) are used to measure the
stability margins of multivariable feedback designs in the face of additive plant
perturbations A and multiplicative plant perturbations M , respectively. See
the following figure.
Consider how the singular value Bode plot of complementary sensitivity T(s)
determines the stability margin for multiplicative perturbations M . The
multiplicative stability margin is, by definition, the size of the smallest stable
M (s) that destabilizes the system in the figure below when A = 0 .

Perturbed Plant

A(s)

K(s)

G(s)

Additive/Multiplicative Uncertainty

2-6

+ A(s)

Typical Loop Shapes, S and T Design

Taking ( M ( j ) ) to be the definition of the size of M ( j ) , you have the

following useful characterization of multiplicative stability robustness:

Multiplicative Robustness The size of the smallest destabilizing

multiplicative uncertainty M (s) is
1
( M ( j ) ) = ----------------------- ( T ( j ) )

The smaller is ( T ( j ) ) , the greater will be the size of the smallest

destabilizing multiplicative perturbation, and hence the greater will be the
stability margin.
A similar result is available for relating the stability margin in the face of
additive plant perturbations A (s) to R(s) if you take ( A ( j ) ) to be the
definition of the size of ( A ( j ) ) at frequency .
Additive Robustness The size of the smallest destabilizing additive
uncertainty A is
1
( A ( j ) ) = ----------------------- ( R ( j ) )

As a consequence of robustness theorems 1 and 2, it is common to specify the

stability margins of control systems via singular value inequalities such as
1

( R { j } ) W 2 ( j )
1

( T { j } ) W 3 ( j )

(2-3)
(2-4)

where W 2 ( j ) and W 3 ( j ) are the respective sizes of the largest

anticipated additive and multiplicative plant perturbations.
It is common practice to lump the effects of all plant uncertainty into a single
fictitious multiplicative perturbation M , so that the control design
requirements can be written

2-7

Multivariable Loop Shaping

1
-------------------------- W 1 ( j ) ;
i ( S ( j ) )

i ( T [ j ] ) W 3 ( j )

as shown in Singular Value Specifications on L, S, and T on page 2-9.

It is interesting to note that in the upper half of the figure (above the 0 dB line),
1
( L ( j ) ) ----------------------- ( S ( j ) )
while in the lower half of Singular Value Specifications on L, S, and T on
page 2-9 (below the 0 dB line),
( L ( j ) ) ( T ( j ) )
This results from the fact that
def

S( s ) = ( I + L ( s ) )
def

T( s ) = L ( s ) ( I + L ( s ) )

2-8

L(s) ,
1

L ( s ),

if ( L ( s ) ) 1
if ( L ( s ) ) 1

Typical Loop Shapes, S and T Design

o(L)
|Gd| DESIRED LOOPSHAPE

|W |
1
0 db

PERFORMANCE
BOUND

o(T)

1
o(S)

ROBUSTNESS
BOUND
-1
|W |
3

o(L)
DESIRED CROSSOVER c
Singular Value Specifications on L, S, and T

Thus, it is not uncommon to see specifications on disturbance attenuation and

multiplicative stability margin expressed directly in terms of forbidden regions
for the Bode plots of i ( L ( j ) ) as singular value loop shaping requirements,
either as specified upper/lower bounds or as a target desired loop shape see
the preceding figure.

Guaranteed Gain/Phase Margins in MIMO Systems

For those who are more comfortable with classical single-loop concepts, there
are the important connections between the multiplicative stability margins
predicted by ( T ) and those predicted by classical M-circles, as found on the
Nichols chart. Indeed in the single-input/single-output case,
L ( j ) ( T ( j ) ) = -----------------------1 + L ( j )
which is precisely the quantity you obtain from Nichols chart M-circles. Thus,
T is a multiloop generalization of the closed-loop resonant peak magnitude
which, as classical control experts will recognize, is closely related to the

2-9

Multivariable Loop Shaping

damping ratio of the dominant closed-loop poles. Also, it turns out that you can
relate T , S to the classical gain margin GM and phase margin M in each
feedback loop of the multivariable feedback system of Block Diagram of the
Multivariable Feedback Control System on page 2-5 via the formulas [6]
1
G M 1 + -----------T
1
G M 1 + ---------------------------11 S
1
1
M 2 sin ----------------
2 T
1
1
M 2 sin ----------------
2 S

These formulas are valid provided S and T are larger than 1, as is

normally the case. The margins apply even when the gain perturbations or
phase perturbations occur simultaneously in several feedback channels.
The infinity norms of S and T also yield gain reduction tolerances. The gain
reduction tolerance gm is defined to be the minimal amount by which the gains
in each loop would have to be decreased in order to destabilize the system.
Upper bounds on gm are
1
g m 1 -----------T
1
g m ----------------------------1+1 S

2-10

Using LOOPSYN to Do H-Infinity Loop Shaping

Using LOOPSYN to Do H-Infinity Loop Shaping

The command loopsyn lets you design a stabilizing feedback controller to
optimally shape the open loop frequency response of a MIMO feedback control
system to match as closely as possible a desired loop shape Gd see the
preceding figure. The basic syntax of the loopsyn loop-shaping controller
synthesis command is
K = loopsyn(G,Gd)

Here G is the LTI transfer function matrix of a MIMO plant model, Gd is the
target desired loop shape for the loop transfer function L=G*K, and K is the
optimal loop-shaping controller. The LTI controller K has the property that it
shapes the loop L=G*K so that it matches the frequency response of Gd as closely
as possible, subject to the constraint that the compensator must stabilize the
plant model G.

Example: NASA HiMAT Loop Shaping

To see how the loopsyn command works in practice to address robustness and
performance tradeoffs, consider again the NASA HiMAT aircraft model taken
from the paper of Safonov, Laub, and Hartmann [8]. The longitudinal dynamics
of the HiMAT aircraft trimmed at 25000 ft and 0.9 Mach are unstable and have
two right-half-plane phugoid modes. The linear model has state-space
1
realization G ( s ) = C ( Is A ) B with six states, with the first four states
representing angle of attack () and attitude angle () and their rates of
change, and the last two representing elevon and canard control actuator
dynamics see Aircraft Configuration and Vertical Plane Geometry on
page 2-12.
ag =[
-2.2567e-02 -3.6617e+01 -1.8897e+01 -3.2090e+01
3.2509e+00 -7.6257e-01;
9.2572e-05 -1.8997e+00
9.8312e-01 -7.2562e-04 -1.7080e-01 -4.9652e-03;
1.2338e-02
1.1720e+01 -2.6316e+00
8.7582e-04 -3.1604e+01
2.2396e+01;
0
0
1.0000e+00
0
0
0;
0
0
0
0 -3.0000e+01
0;
0
0
0
0
0 -3.0000e+01];
bg = [0
0;
0
0;
0
0;
0
0;
30
0;
0
30];
cg = [0
1
0
0
0
0;
0
0
0
1
0
0];
dg = [0
0;

2-11

Multivariable Loop Shaping

0];

G=ss(ag,bg,cg,dg);

The control variables are elevon and canard actuators (e and c). The output
variables are angle of attack () and attitude angle ().
.

Aileron
Canard
Flap
Rudder

Elevator
Elevon
x-axis
0

Velocity
Horizontal

Aircraft Configuration and Vertical Plane Geometry

This model is good at frequencies below 100 rad/s with less than 30% variation
between the true aircraft and the model in this frequency range. However as
noted in [8], it does not reliably capture very high-frequency behaviors, because
it was derived by treating the aircraft as a rigid body and neglecting lightly
damped fuselage bending modes that occur at somewhere between 100 and 300
rad/s. These unmodeled bending modes might cause as much as 20 dB
deviation (i.e., 1000%) between the frequency response of the model and the
actual aircraft for frequency > 100 rad/s. Other effects like control actuator
time delays and fuel sloshing also contribute to model inaccuracy at even
higher frequencies, but the dominant unmodeled effects are the fuselage
bending modes. You can think of these unmodeled bending modes as
multiplicative uncertainty of size 20 dB, and design your controller using
loopsyn, by making sure that the loop has gain less than -20 dB at, and beyond,
the frequency > 100 rad/s.

2-12

Using LOOPSYN to Do H-Infinity Loop Shaping

Design Specifications
The singular value design specifications are
Robustness Spec.: 20 dB/decade roll-off slope and 20 dB loop gain at 100
rad/s
Performance Spec.: Maximize the sensitivity function as much as possible.
Both specs can be accommodated by taking as the desired loop shape
Gd(s)=8/s

MATLAB Commands for a LOOPSYN Design

%% Enter the desired loop shape Gd
s=zpk('s'); % Laplace variable s
Gd=8/s; % desired loop shape
%% Compute the optimal loop shaping controller K
[K,CL,GAM]=loopsyn(G,Gd);
%% Compute the loop L, sensitivity S and
%% complementary sensitivity T:
L=G*K;
I=eye(size(L));
S=feedback(I,L); % S=inv(I+L);
T=I-S;
%% Plot the results:
% step response plots
step(T);title('\alpha and \theta command step responses');
% frequency response plots
figure;
sigma(I+L,'--',T,':',L,'r--',Gd,'k-.',Gd/GAM,'k:',...
Gd*GAM,'k:',{.1,100});grid
legend('1/\sigma(S) performance',...
'\sigma(T) robustness',...
'\sigma(L) loopshape',...
'\sigma(Gd) desired loop',...
'\sigma(Gd) \pm GAM, dB');

2-13

Multivariable Loop Shaping

The plots of the resulting step and frequency response for the NASA HiMAT
loopsyn loop-shaping controller design are shown in the figure below. The
number GAM, dB (i.e., 20log10(GAM) tells you the accuracy with which your
loopsyn control design matches the target desired loop:
( GK ), db G d , db - GAM, db

for < c

( GK ), db G d , db + GAM, db

for > c .

From: In(1)

From: In(2)

To: Out(1)

1.5
1
0.5

To: Out(2)

0
1.5
1
0.5
0
0

10

HiMAT Closed Loop Step Responses

2-14

10

Using LOOPSYN to Do H-Infinity Loop Shaping

Singular Values

50

(L) loopshape
(Gd) desired loop
(Gd) GAM, db

40

Singular Values (dB)

30
20
10
0
10
20
30
40 1
10

10

Singular Values

40

(T) robustness
1/(S) performance

30
Singular Values (dB)

10
10
Frequency (rad/sec)

20
10
0
10
20
30
40 1
10

10
10
Frequency (rad/sec)

10

LOOPSYN Design Results for NASA HiMAT

2-15

Multivariable Loop Shaping

Fine-Tuning the LOOPSYN Target Loop Shape Gd to Meet Design Goals

If your first attempt at loopsyn design does not achieve everything you wanted,
you will need to readjust your target desired loop shape Gd. Here are some basic
design tradeoffs to consider:
Stability Robustness. Your target loop Gd should have low gain (as small
as possible) at high frequencies where typically your plant model is so poor
that its phase angle is completely inaccurate, with errors approaching 180
or more.
Performance. Your Gd loop should have high gain (as great as possible) at
frequencies where your model is good, in order to ensure good control
accuracy and good disturbance attenuation.
Crossover and Roll-Off. Your desired loop shape Gd should have its 0 dB
crossover frequency (denoted c) between the above two frequency ranges,
and below the crossover frequency c it should roll off with a negative slope
of between -20 and -40 dB/decade, which helps to keep phase lag to less than
-180 inside the control loop bandwidth (0 < < c).
Other considerations that might affect your choice of Gd are the
right-half-plane poles and zeros of the plant G, which impose fundamental
limits on your 0 dB crossover frequency c [12]. For instance, your 0 dB
crossover c must be greater than the magnitude of any plant right-half-plane
poles and less than the magnitude of any right-half-plane zeros.

max
Re(p i) > 0

pi < c <

min
Re(z i) > 0

zi

If you do not take care to choose a target loop shape Gd that conforms to these
fundamental constraints, then loopsyn will still compute the optimal
loop-shaping controller K for your Gd, but you should expect that the optimal
loop L=G*K will have a poor fit to the target loop shape Gd, and consequently it
might be impossible to meet your performance goals.

2-16

Using MIXSYN for H-Infinity Loop Shaping

Using MIXSYN for H-Infinity Loop Shaping

A popular alternative approach to loopsyn loop shaping is H mixed-sensitivity
loop shaping, which is implemented by the Robust Control Toolbox command
K=mixsyn(G,W1,[],W3)

With mixsyn controller synthesis, your performance and stability robustness

specifications equations (2-2) and (2-4) are combined into a single infinity norm
specification of the form
T y1u1 1
where (see MIXSYN H Mixed-Sensitivity Loop Shaping Ty1 u1)
def

Ty1 u1 =

W1 S
W3 T

The term T y1u1 is called a mixed-sensitivity cost function because it

penalizes both sensitivity S(s) and complementary sensitivity T(s). Loop
shaping is achieved when you choose W1 to have the target loop shape for
frequencies < c, and you choose 1/W3 to be the target for > c. In choosing
design specifications W1 and W3 for a mixsyn controller design, you need to
ensure that your 0 dB crossover frequency for the Bode plot of W1 is below the
0 dB crossover frequency of 1/W3, as shown in Singular Value Specifications
on L, S, and T on page 2-9, so that there is a gap for the desired loop shape Gd
1
to pass between the performance bound W1 and your robustness bound W 3 .
Otherwise, your performance and robustness requirements will not be
achievable.

2-17

Multivariable Loop Shaping

AUGMENTED PLANT P(s)

WWW111

y
1a

y1
+
u
1
u

e
-

GG G

WW 33 W3

1b

2
y
2

CONTROLLER

K K

F(s)

MIXSYN H Mixed-Sensitivity Loop Shaping Ty1 u1

Example: NASA HiMAT Design Using MIXSYN

To do a mixsyn H mixed-sensitivity synthesis design on the HiMAT model,
start with the plant model G created in Example: NASA HiMAT Loop
Shaping on page 2-11 and type the following commands:
% Set up the performance and robustness bounds W1 & W3
s=zpk('s'); % Laplace variable s
MS=2;AS=.03;WS=5;
W1=(s/MS+WS)/(s+AS*WS);
MT=2;AT=.05;WT=20;
W3=(s+WT/MT)/(AT*s+WT);
% Compute the H-infinity mixed-sensitivity optimal sontroller K1
[K1,CL1,GAM1]=mixsyn(G,W1,[],W3);
% Next compute and plot the closed-loop system.
% Compute the loop L1, sensitivity S1, and comp sensitivity T1:
L1=G*K1;
I=eye(size(L1));
S1=feedback(I,L1); % S=inv(I+L1);
T1=I-S1;
% Plot the results:
% step response plots
step(T1,1.5);

2-18

Using MIXSYN for H-Infinity Loop Shaping

title('\alpha and \theta command step responses');

% frequency response plots
figure;
sigma(I+L1,'--',T1,':',L1,'r--',...
W1/GAM1,'k--',GAM1/W3,'k-.',{.1,100});grid
legend('1/\sigma(S) performance',...
'\sigma(T) robustness',...
'\sigma(L) loopshape',...
'\sigma(W1) performance bound',...
'\sigma(1/W3) robustness bound');

The resulting mixsyn singular value plots for the NASA HiMAT model are
shown below.

Singular Values

40

1/(S) performance
(T) robustness
(L) loopshape
(W1) performance bound
(1/W3) robustnes bound

Singular Values (dB)

30
20
10
0
10
20
30 1
10

10
10
Frequency (rad/sec)

10

MIXSYN Design Results for NASA HiMAT

2-19

Multivariable Loop Shaping

Loop-Shaping Commands
The Robust Control Toolbox gives you several choices for shaping the frequency
response properties of multiinput/multioutput (MIMO) feedback control loops.
Some of the main commands that you are likely to use for loop-shaping design,
and associated utility functions, are listed below:
MIMO Loop-Shaping Commands
loopsyn

H loop-shaping controller synthesis

ltrsyn

LQG loop-transfer recovery

mixsyn

H mixed-sensitivity controller synthesis

ncfsyn

Glover-McFarlane H normalized coprime factor loopshaping controller synthesis

MIMO Loop-Shaping Utility Functions

2-20

augw

Augmented plant for weighted H2 and H mixedsensitivity control synthesis

makeweight

Weights for H mixed sensitivity (mixsyn, augw)

sigma

Singular value plots of LTI feedback loops

3
Model Reduction for
Robust Control

Introduction (p. 3-2)

Motivations, Hankel singular values

Overview of Model Reduction

Techniques (p. 3-4)

Hankel singular value based model reduction routines

and their error bounds

Approximating Plant Models

Additive Error Methods (p. 3-6)

Robust frequency; rigid body dynamics; large size plant;

examples

Approximating Plant Models

Multiplicative Error Method (p. 3-8)

Multiplicative error method, followed by a comparison

between additive and multiplicative methods

Using Modal Algorithms (p. 3-10)

How to keep the j-axis poles of a plant during model

reduction

Reducing Large-Scale Models (p. 3-13) Strategies for reducing very large plant models
Using Normalized Coprime Factor
Methods (p. 3-14)

Model order reduction examples

References (p. 3-15)

Sources for model reduction theory

Model Reduction for Robust Control

Introduction
In the design of robust controllers for complicated systems, model reduction fits
several goals:
1 To simplify the best available model in light of the purpose for which the

model is to be used namely, to design a control system to meet certain

specifications.
2 To speed up the simulation process in the design validation stage, using a

smaller size model with most of the important system dynamics preserved.
3 Finally, if a modern control method such as LQG or H is used for which the

complexity of the control law is not explicitly constrained, the order of the
resultant controller is likely to be considerably greater than is truly needed.
A good model reduction algorithm applied to the control law can sometimes
significantly reduce control law complexity with little change in control
system performance.
Model reduction routines in this toolbox can be put into two categories:
Additive error method The reduced-order model has an additive error
bounded by an error criterion.
Multiplicative error method The reduced-order model has a
multiplicative or relative error bounded by an error criterion.
The error is measured in terms of peak gain across frequency (H norm), and
the error bounds are a function of the neglected Hankel singular values.

Hankel Singular Values

In control theory, eigenvalues define a system stability, whereas Hankel
singular values define the energy of each state in the system. Keeping larger
energy states of a system preserves most of its characteristics in terms of
stability, frequency, and time responses. Model reduction techniques presented
here are all based on the Hankel singular values of a system. They can achieve
a reduced-order model that preserves the majority of the system
characteristics.
Mathematically, given a stable state-space system (A,B,C,D), its Hankel
singular values are defined as [1]

3-2

Introduction

H =

i ( PQ )

where P and Q are controllability and observability grammians satisfying

AP + PA

= BB

A Q + QA = C C
For example,
rand('state',1234); randn('state',5678);
G = rss(30,4,3);
hankelsv(G)

returns a Hankel singular value plot as follows:

Hankel Singular Value

10

abs

10

15

20

25

30

35

Order

which shows that system G has most of its energy stored in states 1 through
15 or so. Later, you will see how to use model reduction routines to keep a
15-state reduced model that preserves most of its dynamic response.

3-3

Model Reduction for Robust Control

Overview of Model Reduction Techniques

The Robust Control Toolbox offers several algorithms for model approximation
and order reduction. These algorithms let you control the absolute or relative
approximation error, and are all based on the Hankel singular values of the
system.
As discussed in previous sections, robust control theory quantifies a system
uncertainty as either additive or multiplicative types. These model reduction
routines are also categorized into two groups: additive error and multiplicative
error types. In other words, some model reduction routines produce a
reduced-order model Gred of the original model G with a bound on the error
G Gred , the peak gain across frequency. Others produce a reduced-order
1

model with a bound on the relative error G ( G Gred ) .

These theoretical bounds are based on the tails of the Hankel singular values
of the model, i.e.,

Additive Error Bound: [1]

n

G Gred 2

i
k+1

where i are denoted the ith Hankel singular value of the original system G.

Multiplicative (Relative) Error Bound: [2]

n
1

G ( G Gred )

( 1 + 2i (

1 + i2 + i ) ) 1

k+1

where i are denoted the ith Hankel singular value of the phase matrix of the
model G (see the bstmr reference page).
Top-Level Model Reduction Command

3-4

Method

Description

reduce

Main interface to model approximation algorithms

Overview of Model Reduction Techniques

Normalized Coprime Balanced Model Reduction Command

Method

Description

ncfmr

Normalized coprime balanced truncation

Additive Error Model Reduction Commands

Method

Description

balancmr

Square-root balanced model truncation

schurmr

Schur balanced model truncation

hankelmr

Hankel minimum degree approximation

Multiplicative Error Model Reduction Command

Method

Description

bstmr

Balanced stochastic truncation

Additional Model Reduction Tools

Method

Description

modreal

Modal realization and truncation

slowfast

Slow and fast state decomposition

stabproj

Stable and antistable state projection

3-5

Model Reduction for Robust Control

Approximating Plant Models Additive Error Methods

Given a system in LTI form, the following commands reduce the system to any
desired order you specify. The judgment call is based on its Hankel singular
values as shown in previous paragraph.
rand('state',1234); randn('state',5678);
G = rss(30,4,3);
% balanced truncation to models with sizes 12:16
[g1,info1] = balancmr(G,12:16); % or use reduce
% Schur balanced truncation by specifying `MaxError'
[g2,info2] = schurmr(G,'MaxError',[1,0.8,0.5,0.2]);
sigma(G,'b-',g1,'r--',g2,'g-.')

shows a comparison plot of the original model G and reduced models g1 and g2.

Singular Values

30

G
gr
g2
25

20

Singular Values (dB)

15

10

10
4
10

10

10

10
10
Frequency (rad/sec)

10

10

10

To determine whether the theoretical error bound is satisfied,

norm(G-g1(:,:,1),'inf') % 2.0123
info1.ErrorBound(1) % 2.8529

or plot the model error vs. error bound via the following commands:

3-6

Approximating Plant Models Additive Error Methods

[sv,w] = sigma(G-g1(:,:,1));
loglog(w,sv,w,info1.ErrorBound(1)*ones(size(w)))
xlabel('rad/sec');ylabel('SV');
title('Error Bound and Model Error')

Error Bound and Model Error

10

10

SV

10

10

10

10

10

10

10

10
rad/sec

10

10

10

3-7

Model Reduction for Robust Control

Approximating Plant Models Multiplicative Error Method

In most cases, multiplicative error model reduction method bstmr tends to
bound the relative error between the original and reduced-order models across
the frequency range of interest, hence producing a more accurate
reduced-order model than the additive error methods. This characteristic is
obvious in system models with low damped poles.
The following commands illustrate the significance of a multiplicative error
model reduction method as compared to any additive error type. Clearly the
phase-matching algorithm using bstmr provides a better fit in the Bode plot.
rand('state',1234); randn('state',5678); G = rss(30,1,1);
[gr,infor] = reduce(G,'algo','balance','order',7);
[gs,infos] = reduce(G,'algo','bst','order',7);
figure(1);bode(G,'b-',gr,'r--');
title('Additive Error Method')
figure(2);bode(G,'b-',gs,'r--');
title('Relative Error Method')

Additive Error Method

10
5
0

Magnitude (dB)

5
10
15
20
25
30
35
40
135
90

Phase (deg)

45
0
45
90
135
2

10

3-8

10

10
Frequency (rad/sec)

10

10

Approximating Plant Models Multiplicative Error Method

Relative Error Method

20

Magnitude (dB)

10

10

20

30

40
180
135

Phase (deg)

90
45
0
45
90
135
2

10

10

10
Frequency (rad/sec)

10

10

Therefore, for some systems with low damped poles/zeros, the balanced
stochastic method (bstmr) produces a better reduced-order model fit in those
frequency ranges to make multiplicative error small. Whereas additive error
methods such as balancmr, schurmr, or hankelmr only care about minimizing
the overall absolute peak error, they can produce a reduced-order model
missing those low damped poles/zeros frequency regions.

3-9

Model Reduction for Robust Control

Using Modal Algorithms

Rigid Body Dynamics
In many cases, a models j-axis poles are important to keep after model
reduction, e.g., rigid body dynamics of a flexible structure plant or integrators
of a controller. A unique routine, modreal, serves the purpose nicely.
modreal puts a system into its modal form, with eigenvalues appearing on the

diagonal of its A-matrix. Real eigenvalues appear in 1-by-1 blocks, and complex
eigenvalues appear in 2-by-2 real blocks. All the blocks are ordered in
ascending order, based on their eigenvalue magnitudes, by default, or
descending order, based on their real parts. Therefore, specifying the number
of j-axis poles splits the model into two systems with one containing only
j-axis dynamics, the other containing the non-j axis dynamics.
rand('state',5678); randn('state',1234); G = rss(30,1,1);
[Gjw,G2] = modreal(G,1); % only one rigid body dynamics
G2.d = Gjw.d; Gjw.d = 0; % put DC gain of G into G2
subplot(211);sigma(Gjw);ylabel('Rigid Body')
subplot(212);sigma(G2);ylabel('Nonrigid Body')

3-10

Using Modal Algorithms

Further model reduction can be done on G2 without any numerical difficulty.

After G2 is further reduced to Gred, the final approximation of the model is
simply Gjw+Gred.
This process of splitting j-axis poles has been built in and automated in all the
model reduction routines (balancmr, schurmr, hankelmr, bstmr, hankelsv) so
that users need not worry about splitting the model.
The following single command creates a size 8 reduced-order model from its
original 30-state model:
rand('state',5678); randn('state',1234); G = rss(30,1,1);
[gr,info] = reduce(G); % choose a size of 8 at prompt
bode(G,'b-',gr,'r--')

Without specifying the size of the reduced-order model, a Hankel singular

value plot is shown below.

3-11

Model Reduction for Robust Control

Hankel Singular Values (1 of jwaxis pole(s) = inf, not shown)

4
3.5
3

abs

2.5
2
1.5
1
0.5
0

10

15

20

25

30

35

Order

Bode Diagram

30
25

Original G (30state)

20

Gred (8state)

Magnitude (dB)

15
10
5
0
5
10
15
20
720
630

Phase (deg)

540
450
360
270
180
2

10

10

10

10

10

10

Frequency (rad/sec)

The default algorithm balancmr of reduce has done a great job of

approximating a 30-state model with just eight states. Again, the rigid body
dynamics are preserved for further controller design.

3-12

Reducing Large-Scale Models

Reducing Large-Scale Models

For some really large size problems (states > 200), modreal turns out to be the
only way to start the model reduction process. Because of the size and
numerical properties associated with those large size, and low damped
dynamics, most Hankel based routines can fail to produce a good reduced-order
model.
modreal puts the large size dynamics into the modal form, then truncates the

dynamic model to an intermediate stage model with a comfortable size of 50 or

so states. From this point on, those more sophisticated Hankel singular value
based routines can further reduce this intermediate stage model, in a much
more accurate fashion, to a smaller size for final controller design.
For a typical 240-state flexible spacecraft model in the spacecraft industry,
applying modreal and bstmr (or any other additive routines) in sequence can
reduce the original 240-state plant dynamics to a seven-state three-axis model
including rigid body dynamics. Any modern robust control design technique
mentioned in this toolbox can then be easily applied to this smaller size plant
for a controller design.

3-13

Model Reduction for Robust Control

Using Normalized Coprime Factor Methods

A special model reduction routine ncfmr produces a reduced-order model by
truncating a balanced coprime set of a given model. It can directly simplify a
modern controller with integrators to a smaller size by balanced truncation of
the normalized coprime factors. It does not need modreal for
pre-/postprocessing as the other routines do. However, the integrators will not
be preserved.
rand('state',5678); randn('state',1234);
K= rss(30,4,3); % The same model G used in the 1st example
[Kred,info2] = ncfmr(K);
sigma(K,Kred)

Again, without specifying the size of the reduced-order model, any model
reduction routine presented here will plot a Hankel singular value bar chart
and prompt you for a reduced model size. In this case, enter 15.

Singular Values

120

Original K (30state)
Kred (15state)
100

Singular Values (dB)

80

60

40

20

20

10

10

10

10
Frequency (rad/sec)

10

10

10

10

If integral control is important, previously mentioned methods (except ncfmr)

can nicely preserve the original integrator(s) in the model.

3-14

References

References
[15] Glover, K., All Optimal Hankel Norm Approximation of Linear
Multivariable Systems, and Their L - Error Bounds, Int. J. Control, Vol. 39,
No. 6, 1984, pp. 1145-1193.
[16] Zhou, K., Doyle, J.C., and Glover, K., Robust and Optimal Control,
Englewood Cliffs, NJ, Prentice Hall, 1996.
[17] Safonov, M.G., and Chiang, R.Y., A Schur Method for Balanced Model
Reduction, IEEE Trans. on Automat. Contr., Vol. 34, No. 7, July 1989, pp.
729-733.
[18] Safonov, M.G., Chiang, R.Y., and Limebeer, D.J.N., Optimal Hankel
Model Reduction for Nonminimal Systems, IEEE Trans. on Automat. Contr.,
Vol. 35, No. 4, April 1990, pp. 496-502.
[19] Safonov, M.G., and Chiang, R.Y., Model Reduction for Robust Control: A
Schur Relative Error Method, International J. of Adaptive Control and Signal
Processing, Vol. 2, 1988, pp. 259-272.
[20] Obinata, G., and Anderson, B.D.O., Model Reduction for Control System
Design, London, Springer-Verlag, 2001.

3-15

Model Reduction for Robust Control

3-16

4
Robustness Analysis

Uncertainty Modeling (p. 4-2)

What is uncertainty modeling? Why is it important?

Robustness Analysis (p. 4-8)

Designing and comparing various controllers

Multiinput, Multioutput Robustness

Analysis (p. 4-12)

Creating and analyzing uncertain MIMO models

Worst-Case Gain Analysis (p. 4-20)

How to perform worst-case gain analysis

Summary of Robustness Analysis Tools Tables of functions available for robust control
(p. 4-23)

Robustness Analysis

Uncertainty Modeling
Dealing with and understanding the effects of uncertainty are important tasks
for the control engineer. Reducing the effects of some forms of uncertainty
(initial conditions, low-frequency disturbances) without catastrophically
increasing the effects of other dominant forms (sensor noise, model
uncertainty) is the primary job of the feedback control system.
Closed-loop stability is the way to deal with the (always present) uncertainty
in initial conditions or arbitrarily small disturbances.
High-gain feedback in low-frequency ranges is a way to deal with the effects of
unknown biases and disturbances acting on the process output. In this case,
you are forced to use roll-off filters in high-frequency ranges to deal with
high-frequency sensor noise in a feedback system.
Finally, notions such as gain and phase margins (and their generalizations)
help quantify the sensitivity of stability and performance in the face of model
uncertainty, which is the imprecise knowledge of how the control input directly
affects the feedback variables.
The Robust Control Toolbox has built-in features allowing you to specify model
uncertainty simply and naturally. The primary building blocks, called
uncertain elements or atoms, are uncertain real parameters and uncertain
linear, time-invariant objects. These can be used to create coarse and simple or
detailed and complex descriptions of the model uncertainty present within your
process models.
Once formulated, high-level system robustness tools can help you analyze the
potential degradation of stability and performance of the closed-loop system
brought on by the system model uncertainty.

Creating Uncertain Models of Dynamic Systems

The two dominant forms of model uncertainty are
Uncertainty in parameters of the underlying differential equation models
Frequency-domain uncertainty, which often quantifies model uncertainty by
describing absolute or relative uncertainty in the processs frequency
response
Using these two basic building blocks, along with conventional system creation
commands (such as ss and tf), you can easily create uncertain system models.

4-2

Uncertainty Modeling

Creating Uncertain Parameters

An uncertain parameter has a name (used to identify it within an uncertain
system with many uncertain parameters) and a nominal value. Being
uncertain, it also has variability, described in one of the following ways:
An additive deviation from the nominal
A range about the nominal
A percentage deviation from the nominal
Create a real parameter, with name 'bw', nominal value 5, and a percentage
uncertainty of 10%.
bw = ureal('bw',5,'Percentage',10)

This creates a ureal object. View its properties using the get command.
Uncertain Real Parameter: Name bw, NominalValue 5, variability =
[-10 10]%
get(bw)
Name: 'bw'
NominalValue: 5
Mode: 'Percentage'
Range: [4.5000 5.5000]
PlusMinus: [-0.5000 0.5000]
Percentage: [-10 10]
AutoSimplify: 'basic'

Note that the range of variation (Range property) and the additive deviation
from nominal (the PlusMinus property) are consistent with the Percentage
property value.
You can create state-space and transfer function models with uncertain real
coefficients using ureal objects. The result is an uncertain state-space object, or
uss. As an example, use the uncertain real parameter bw to model a first-order
system whose bandwidth is between 4.5 and 5.5 rad/s.
H = tf(1,[1/bw 1])
USS: 1 State, 1 Output, 1 Input, Continuous System
bw: real, nominal = 5, variability = [-10 10]%, 1 occurrence

Note that the result H is an uncertain system, called a uss object. The nominal
value of H is a state-space object. Verify that the pole is at -5.

4-3

Robustness Analysis

pole(H.NominalValue)
ans =
-5

Next, use bode and step to examine the behavior of H.

bode(H,{1e-1 1e2});

Bode Diagram

Magnitude (dB)

10

15

20

25

Phase (deg)

30
0

45

90
1

10

10

10
Frequency (rad/sec)

4-4

10

Uncertainty Modeling

step(H)

Step Response

0.9

0.8

0.7

Amplitude

0.6

0.5

0.4

0.3

0.2

0.1

0.5

1.5

Time (sec)

While there are variations in the bandwidth and time constant of H, the
high-frequency rolls off at 20 dB/decade regardless of the value of bw. You can
capture the more complicated uncertain behavior that typically occurs at high
frequencies using the ultidyn uncertain element, which is described next.

Quantifying Unmodeled Dynamics

An informal way to describe the difference between the model of a process and
the actual process behavior is in terms of bandwidth. It is common to hear The
model is good out to 8 radians/second. The precise meaning is not clear, but it
is reasonable to believe that for frequencies lower than, say, 5 rad/s, the model
is accurate, and for frequencies beyond, say, 30 rad/s, the model is not
necessarily representative of the process behavior. In the frequency range
between 5 and 30, the guaranteed accuracy of the model degrades.
The uncertain linear, time-invariant dynamics object ultidyn can be used to
model this type of knowledge. An ultidyn object represents an unknown linear
system whose only known attribute is a uniform magnitude bound on its
frequency response. When coupled with a nominal model and a

4-5

Robustness Analysis

frequency-shaping filter, ultidyn objects can be used to capture uncertainty

associated with the model dynamics.
Suppose that the behavior of the system modeled by H significantly deviates
from its first-order behavior beyond 9 rad/s, for example, about 5% potential
relative error at low frequency, increasing to 1000% at high frequency where H
rolls off. In order to model frequency domain uncertainty as described above
using ultidyn objects, follow the steps below.
1 Create the nominal system Gnom, using tf, ss, or zpk. Gnom itself might

already have parameter uncertainty. In this case Gnom is H, the first-order

system with an uncertain time constant.
2 Create a filter W, called the weight, whose magnitude represents the

relative uncertainty at each frequency. The utility makeweight is useful for

creating first-order weights with specific low- and high-frequency gains, and
specified gain crossover frequency.
3 Create an ultidyn object Delta with magnitude bound equal to 1.

The uncertain model G is formed by G = Gnom*(1+W*Delta).

If the magnitude of W represents an absolute (rather than relative) uncertainty,
use the formula G = Gnom + W*Delta instead.
The commands below carry out these steps.
Gnom = H;
W = makeweight(.05,9,10);
Delta = ultidyn('Delta',[1 1]);
G = Gnom*(1+W*Delta)
USS: 2 States, 1 Output, 1 Input, Continuous System
Delta: 1x1 LTI, max. gain = 1, 1 occurrence
bw: real, nominal = 5, variability = [-10 10]%, 1 occurrence

Note that the result G is also an uncertain system, with dependence on both
Delta and bw. You can use bode to make a Bode plot of 20 random samples of
G's behavior over the frequency range [0.1 100] rad/s.

4-6

Uncertainty Modeling

bode(G,{1e-1 1e2},25)

Bode Diagram

10
5

Magnitude (dB)

0
5
10
15
20
25
30
360

Phase (deg)

180

180
1

10

10

10

10

Frequency (rad/sec)

In the next section, you design and compare two feedback controllers for G.

4-7

Robustness Analysis

Robustness Analysis
Next, design a feedback controller for G. The goals of this design are the usual
ones: good steady-state tracking and disturbance rejection properties. Because
the plant model is nominally a first-order lag, choose a PI control architecture.
Given the desired closed-loop damping ratio and natural frequency n, the
design equations for KI and KP (based on the nominal open-loop time constant
of 0.2) are
2

n
2
K I = -------, K P = -------------2- 1
5
5
Follow these steps to design the controller:
1 In order to study how the uncertain behavior of G affects the achievable

closed-loop bandwidth, design two controllers, both achieving =0.707, with

different n: 3 and 7.5 respectively.
xi
wn
K1
wn
K2

=
=
=
=
=

0.707;
3;
tf([(2*xi*wn/5-1) wn*wn/5],[1 0]);
7.5;
tf([(2*xi*wn/5-1) wn*wn/5],[1 0]);

Note that the nominal closed-loop bandwidth achieved by K2 is in a region

where G has significant model uncertainty. It will not be surprising if the
model variations lead to significant degradations in the closed-loop
performance.
2 Form the closed-loop systems using feedback.

T1 = feedback(G*K1,1);
T2 = feedback(G*K2,1);
3 Plot the step responses of 20 samples of each closed-loop system.

tfinal = 3;
step(T1,'b',T2,'r',tfinal,20)

4-8

Robustness Analysis

Step Response
2

1.5

Amplitude

0.5

0.5

0.5

1.5

2.5

Time (sec)

The step responses for T2 exhibit a faster rise time because K2 sets a higher
closed loop bandwidth. However, the model variations have a greater effect.
You can use robuststab to check the robustness of stability to the model
variations.
[stabmarg1,destabu1,report1] = robuststab(T1);
stabmarg1
stabmarg1 =
ubound: 4.0241
lbound: 4.0241
destabfreq: 3.4959
[stabmarg2,destabu2,report2] = robuststab(T2);
stabmarg2
stabmarg2 =
ubound: 1.2545
lbound: 1.2544
destabfreq: 10.5249

4-9

Robustness Analysis

The stabmarg variable gives lower and upper bounds on the stability margin.
A stability margin greater than 1 means the system is stable for all values of
the modeled uncertainty. A stability margin less than 1 means there are
allowable values of the uncertain elements that make the system unstable. The
report variable briefly summarizes the analysis.
report1
report1 =
Uncertain System is robustly stable to modeled uncertainty.
-- It can tolerate up to 402% of modeled uncertainty.
-- A destabilizing combination of 402% the modeled uncertainty
exists, causing an instability at 3.5 rad/s.
report2
report2 =
Uncertain System is robustly stable to modeled uncertainty.
-- It can tolerate up to 125% of modeled uncertainty.
-- A destabilizing combination of 125% the modeled uncertainty
exists, causing an instability at 10.5 rad/s.

While both systems are stable for all variations, their performance is clearly
affected to different degrees. To determine how the uncertainty affects
closed-loop performance, you can use wcgain to compute the worst-case effect
of the uncertainty on the peak magnitude of the closed-loop sensitivity
(S=1/(1+GK)) function. This peak gain is typically correlated with the amount
of overshoot in a step response.
To do this, form the closed-loop sensitivity functions and call wcgain.
S1 = feedback(1,G*K1);
S2 = feedback(1,G*K2);
[maxgain1,wcu1] = wcgain(S1);
maxgain1
maxgain1 =
lbound: 1.8684
ubound: 1.9025
critfreq: 3.5152
[maxgain2,wcu2] = wcgain(S2);
maxgain2
maxgain2 =
lbound: 4.6031
ubound: 4.6671
critfreq: 11.0231

4-10

Robustness Analysis

The maxgain variable gives lower and upper bounds on the worst-case peak
gain of the sensitivity transfer function, as well as the specific frequency where
the maximum gain occurs. The wcu variable contains specific values of the
uncertain elements that achieve this worst-case behavior.
You can use usubs to substitute these worst-case values for uncertain
elements, and compare the nominal and worst-case behavior. Use bodemag and
step to make the comparison.
bodemag(S1.NominalValue,'b',usubs(S1,wcu1),'b');
hold on
bodemag(S2.NominalValue,'r',usubs(S2,wcu2),'r');
hold off

Bode Diagram
20

10

Magnitude (dB)

10

20

30

40

50
1
10

10

10

10

10

Frequency (rad/sec)

Clearly, while K2 achieves better nominal sensitivity than K1, the nominal
closed-loop bandwidth extends too far into the frequency range where the
process uncertainty is very large. Hence the worst-case performance of K2 is
inferior to K1 for this particular uncertain model.
The next section explores these robustness analysis tools further on a
multiinput, multioutput system.

4-11

Robustness Analysis

Multiinput, Multioutput Robustness Analysis

The previous sections focused on simple uncertainty models of single-input and
single-output systems, predominantly from a transfer function perspective.
You can also create uncertain state-space models made up of uncertain
state-space matrices. Moreover, all the analysis tools covered thus far can be
applied to these systems as well.
Consider, for example, a two-input, two-output, two-state system whose model
has parametric uncertainty in the state-space matrices. First create an
uncertain parameter p. Using the parameter, make uncertain A and C matrices.
The B matrix happens to be not-uncertain, although you will add frequency
domain input uncertainty to the model in the next section.
p
A
B
C
H

=
=
=
=
=

ureal('p',10,'Percentage',10);
[0 p;-p 0];
eye(2);
[1 p;-p 1];
ss(A,B,C,[0 0;0 0]);

You can view the properties of the uncertain system H using the get command.
get(H)
a:
b:
c:
d:
StateName:
Ts:
InputName:
OutputName:
InputGroup:
OutputGroup:
NominalValue:
Uncertainty:
Notes:
UserData:

[2x2
[2x2
[2x2
[2x2
{2x1
0
{2x1
{2x1
[1x1
[1x1
[2x2
[1x1
{}
[]

umat]
double]
umat]
double]
cell}
cell}
cell}
struct]
struct]
ss]
atomlist]

The properties a, b, c, d, and StateName behave in exactly the same manner as

ss objects. The properties InputName, OutputName, InputGroup, and
OutputGroup behave in exactly the same manner as all the system objects (ss,
zpk, tf, and frd). The NominalValue is an ss object.

4-12

Multiinput, Multioutput Robustness Analysis

Adding Independent Input Uncertainty

to Each Channel
The model for H does not include actuator dynamics. Said differently, the
actuator models are unity-gain for all frequencies.
Nevertheless, the behavior of the actuator for channel 1 is modestly uncertain
(say 10%) at low frequencies, and the high-frequency behavior beyond 20 rad/s
is not accurately modeled. Similar statements hold for the actuator in channel
2, with larger modest uncertainty at low frequency (say 20%) but accuracy out
to 45 rad/s.
Use ultidyn objects Delta1 and Delta2 along with shaping filters W1 and W2 to
add this form of frequency domain uncertainty to the model.
W1 = makeweight(.1,20,50);
W2 = makeweight(.2,45,50);
Delta1 = ultidyn('Delta1',[1 1]);
Delta2 = ultidyn('Delta2',[1 1]);
G = H*blkdiag(1+W1*Delta1,1+W2*Delta2)
USS: 4 States, 2 Outputs, 2 Inputs, Continuous System
Delta1: 1x1 LTI, max. gain = 1, 1 occurrence
Delta2: 1x1 LTI, max. gain = 1, 1 occurrence
p: real, nominal = 10, variability = [-10 10]%, 2
occurrences

Note that G is a two-input, two-output uncertain system, with dependence on

three uncertain elements, Delta1, Delta2, and p. It has four states, two from H
and one each from the shaping filters W1 and W2, which are embedded in G.
You can plot a 2-second step response of several samples of G. The 10%
uncertainty in the natural frequency is obvious.

4-13

Robustness Analysis

step(G,2)

Step Response

From: In(1)

From: In(2)

1.5
1

To: Out(1)

0.5
0
0.5
1
1.5

Amplitude

2
2.5
1.5
1

To: Out(2)

0.5
0
0.5
1
1.5
2
2.5

0.5

1.5

20
Time (sec)

0.5

1.5

You can create a Bode plot of 50 samples of G. The high-frequency uncertainty

in the model is also obvious. For clarity, start the Bode plot beyond the
resonance.

4-14

Multiinput, Multioutput Robustness Analysis

bode(G,{13 100},50)

From: In(1)

Bode Diagram

From: In(2)

To: Out(1)

20

20

To: Out(2)

Magnitude (dB) ; Phase (deg)

To: Out(1)

40
360

360
20

20

To: Out(2)

40
360

360
2

10
Frequency (rad/sec)

10

Closed-Loop Robustness Analysis

You need to load the controller and verify that it is two-input and two-output.
load mimoKexample
size(K)
State-space model with 2 outputs, 2 inputs, and 9 states.

You can use the command loopsens to form all the standard plant/controller
feedback configurations, including sensitivity and complementary sensitivity
at both the input and output. Because G is uncertain, all the closed-loop
systems are uncertain as well.

4-15

Robustness Analysis

F = loopsens(G,K)
F =
Poles: [13x1 double]
Stable: 1
Si: [2x2 uss]
Ti: [2x2 uss]
Li: [2x2 uss]
So: [2x2 uss]
To: [2x2 uss]
Lo: [2x2 uss]
PSi: [2x2 uss]
CSo: [2x2 uss]
F is a structure with many fields. The poles of the nominal closed-loop system
are in F.Poles, and F.Stable is 1 if the nominal closed-loop system is stable.
In the remaining 10 fields, S stands for sensitivity, T for complementary
sensitivity, and L for open-loop gain. The suffixes i and o refer to the input and
output of the plant (G). Finally, P and C refer to the plant and controller.

Hence Ti is mathematically the same as

1

K ( I + GK ) G
while Lo is G*K, and CSo is mathematically the same as
K ( I + GK )

You can examine the transmission of disturbances at the plant input to the
plant output using bodemag on F.PSi. Graph 50 samples along with the
nominal.

4-16

Multiinput, Multioutput Robustness Analysis

bodemag(F.PSi,':/-',{1e-1 100},50)

Bode Diagram

From: In(1)

From: In(2)

20
10

To: Out(1)

0
10
20
30

Magnitude (dB)

40
50
60
20
10

To: Out(2)

0
10
20
30
40
50
60
1
10

10

10

1
2

10
Frequency (rad/sec)

10

10

10

Nominal Stability Margins

You can use loopmargin to investigate loop-at-a-time gain and phase margins,
loop-at-a-time disk margins, and simultaneous multivariable margins. They
are computed for the nominal system and do not reflect the uncertainty models
within G.
Explore the simultaneous margins individually at the plant input, individually
at the plant output, and simultaneously at both input and output.
[I,DI,SimI,O,DO,SimO,Sim] = loopmargin(G,K);

The third output argument is the simultaneous gain and phase variations
allowed in all input channels to the plant.
SimI

4-17

Robustness Analysis

SimI =
GainMargin: [0.1180 8.4769]
PhaseMargin: [-76.5441 76.5441]
Frequency: 6.2287

This information implies that the gain at the plant input can vary in both
channels independently by factors between (approximately) 1/8 and 8, as well
as phase variations up to 76 degrees.
The sixth output argument is the simultaneous gain and phase variations
allowed in all output channels to the plant.
SimO
SimO =
GainMargin: [0.1193 8.3836]
PhaseMargin: [-76.3957 76.3957]
Frequency: 18.3522

Note that the simultaneous margins at the plant output are similar to those at
the input. This is not always the case in multiloop feedback systems.
The last output argument is the simultaneous gain and phase variations
allowed in all input and output channels to the plant. As expected, when you
consider all such variations simultaneously, the margins are somewhat
smaller than those at the input or output alone.
Sim
Sim =
GainMargin: [0.5671 1.7635]
PhaseMargin: [-30.8882 30.8882]
Frequency: 18.3522

Nevertheless, these numbers indicate a generally robust closed-loop system,

able to tolerate significant gain (more than +/-50% in each channel) and 30
degree phase variations simultaneously in all input and output channels of the
plant.

Robustness of Stability Model Uncertainty

With loopmargin, you determined various margins of the nominal, multiloop
system. These margins are computed only for the nominal system, and do not
reflect the uncertainty explicitly modeled by the ureal and ultidyn objects.
When you work with detailed, complex uncertain system models, the

4-18

Multiinput, Multioutput Robustness Analysis

conventional margins computed by loopmargin might not always be indicative

of the actual stability margins associated with the uncertain elements. You can
use robuststab to check the stability margin of the system to these specific
modeled variations.
In this example, use robuststab to compute the stability margin of the
closed-loop system represented by Delta1, Delta2, and p.
Use any of the closed-loop systems within F = loopsens(G,K). All of them,
F.Si, F.To, etc., have the same internal dynamics, and hence the stability
properties are the same.
[stabmarg,desgtabu,report] = robuststab(F.So);
stabmarg
stabmarg =
ubound: 2.2175
lbound: 2.2175
destabfreq: 13.7576
report
report =
Uncertain System is robustly stable to modeled uncertainty.
-- It can tolerate up to 222% of modeled uncertainty.
-- A destabilizing combination of 222% the modeled uncertainty
exists, causing an instability at 13.8 rad/s.

This analysis confirms what the loopmargin analysis suggested. The

closed-loop system is quite robust, in terms of stability, to the variations
modeled by the uncertain parameters Delta1, Delta2, and p. In fact, the
system can tolerate more than twice the modeled uncertainty without losing
closed-loop stability.
The next section studies the effects of these variations on the closed-loop output
sensitivity function.

4-19

Robustness Analysis

Worst-Case Gain Analysis

You can plot the Bode magnitude of the nominal output sensitivity function. It
clearly shows decent disturbance rejection in all channels at low frequency.
bodemag(F.So.NominalValue,{1e-1 100})

Bode Diagram

From: In(1)

From: In(2)

10
0

To: Out(1)

10
20
30
40
Magnitude (dB)

50
60
10
0
10
20
To: Out(2)

30
40
50
60
70
80
90
100
1
10

10

10

1
2

10
Frequency (rad/sec)

10

10

10

You can compute the peak value of the maximum singular value of the
frequency response matrix using norm.
[PeakNom,freq] = norm(F.So.NominalValue,'inf')
PeakNom =
1.1317
freq =
7.0483

The peak is about 1.13, occurring at a frequency of 36 rad/s.

4-20

Worst-Case Gain Analysis

What is the maximum output sensitivity gain that is achieved when the
uncertain elements Delta1, Delta2, and p vary over their ranges? You can use
wcgain to answer this.
[maxgain,wcu]
maxgain
maxgain =
lbound:
ubound:
critfreq:

= wcgain(F.So);

2.1017
2.1835
8.5546

The analysis indicates that the worst-case gain is somewhere between 2.1 and
2.2. The frequency where the peak is achieved is about 8.5.
You can replace the values for Delta1, Delta2, and p that achieve the gain of
2.1, using usubs. Make the substitution in the output complementary
sensitivity, and do a step response.
step(F.To.NominalValue,usubs(F.To,wcu),5)
Step Response

From: In(1)

From: In(2)

1.5

To: Out(1)

0.5

Amplitude

1.4
1.2

To: Out(2)

1
0.8
0.6
0.4
0.2
0
0.2

50
Time (sec)

The perturbed response, which is the worst combination of uncertain values in

terms of output sensitivity amplification, does not show significant
degradation of the command response. The settling time is increased by about

4-21

Robustness Analysis

50%, from 2 to 4, and the off-diagonal coupling is increased by about a factor of

about 2, but is still quite small.

4-22

Summary of Robustness Analysis Tools

Summary of Robustness Analysis Tools

Function

Description

ureal
ultidyn
uss

ufrd
loopsens

Create uncertain real parameter

Create uncertain, linear, time-invariant dynamics
Create uncertain state-space object from uncertain
state-space matrices
Create uncertain frequency response object
Compute all relevant open and closed-loop quantities for a
MIMO feedback connection

loopmargin

Compute loop-at-a-time as well as MIMO gain and phase

margins for a multiloop system, including the simultaneous
gain/phase margins

robustperf

Robustness performance of uncertain systems

robuststab

Compute the robust stability margin of a nominally stable

uncertain system

wcgain

wcmargin

wcsens

Compute the worst-case gain of a nominally stable

uncertain system
Compute worst-case (over uncertainty) loop-at-a-time
disk-based gain and phase margins
Compute worst-case (over uncertainty) sensitivity of
plant-controller feedback loop

4-23

Robustness Analysis

4-24

5
H-Infinity and Mu Synthesis

H-Infinity Performance (p. 5-2)

Discusses a modern approach to characterizing

closed-loop performance

Application of H and mu to Active

Suspension Control (p. 5-10)

A fully worked example of designing a compensator for

active suspension control

Functions for Control Design (p. 5-26)

A discussion of the functions you can use for robust

control design

Interpretation of H-Infinity Norm

(p. 5-28)

Details about norms and their properties

References (p. 5-34)

A list of relevant papers and books about H-infinity and

structured singular value design

H-Infinity and Mu Synthesis

H-Infinity Performance
Performance As Generalized Disturbance Rejection
The modern approach to characterizing closed-loop performance objectives is to
measure the size of certain closed-loop transfer function matrices using various
matrix norms. Matrix norms provide a measure of how large output signals can
get for certain classes of input signals. Optimizing these types of performance
objectives over the set of stabilizing controllers is the main thrust of recent
optimal control theory, such as L1, H2, H, and optimal control. Hence, it is
important to understand how many types of control objectives can be posed as
a minimization of closed-loop transfer functions.
Consider a tracking problem, with disturbance rejection, measurement noise,
and control input signal limitations, as shown in Generalized and Weighted
Performance Block Diagram on page 5-4. K is some controller to be designed
and G is the system you want to control.

reference

control
input

external force
disturbance

- e?; - tracking
error
?

noise

Typical Closed-Loop Performance Objective

A reasonable, though not precise, design objective would be to design K to keep

tracking errors and control input signal small for all reasonable reference
commands, sensor noises, and external force disturbances.
Hence, a natural performance objective is the closed-loop gain from exogenous
influences (reference commands, sensor noise, and external force disturbances)
to regulated variables (tracking errors and control input signal). Specifically,
let T denote the closed-loop mapping from the outside influences to the
regulated variables,

5-2

H-Infinity Performance

reference
tracking error = T
external force
control input
noise
regulated variables
outside influences
You can assess performance by measuring the gain from outside influences to
regulated variables. In other words, good performance is associated with T
being small. Because the closed-loop system is a multiinput, multioutput
(MIMO) dynamic system, there are two different aspects to the gain of T:
Spatial (vector disturbances and vector errors)
Temporal (dynamic relationship between input/output signals)
Hence the performance criterion must account for
Relative magnitude of outside influences
Frequency dependence of signals
Relative importance of the magnitudes of regulated variables
So if the performance objective is in the form of a matrix norm, it should
actually be a weighted norm
||WLTWR||
where the weighting function matrices WL and WR are frequency dependent, to
account for bandwidth constraints and spectral content of exogenous signals. A
natural (mathematical) manner to characterize acceptable performance is in
terms of the MIMO |||| (H) norm. See Interpretation of H-Infinity Norm on
page 5-28 for an interpretation of the H norm and signals.

Interconnection with Typical MIMO Performance Objectives

The closed-loop performance objectives are formulated as weighted closed-loop
transfer functions that are to be made small through feedback. A generic
example, which includes many relevant terms, is shown in block diagram form
in Generalized and Weighted Performance Block Diagram on page 5-4. In the
diagram, G denotes the plant model and K is the feedback controller.

5-3

H-Infinity and Mu Synthesis

Generalized and Weighted Performance Block Diagram

The blocks in this figure might be scalar (SISO) and/or multivariable (MIMO),
depending on the specific example. The mathematical objective of H control
is to make the closed-loop MIMO transfer function Ted satisfy ||Ted|| < 1. The
weighting functions are used to scale the input/output transfer functions such
that when ||Ted|| < 1, the relationship between d and e is suitable.
Performance requirements on the closed-loop system are transformed into the
H framework with the help of weighting or scaling functions. Weights are
selected to account for the relative magnitude of signals, their frequency
dependence, and their relative importance. This is captured in the figure
above, where the weights or scalings [Wcmd, Wdist,Wsnois] are used to transform
and scale the normalized input signals [d1,d2,d3] into physical units defined as
[d1, d2, d3]. Similarly weights or scalings [Wact, Wperf1,Wperf2] transform and

5-4

H-Infinity Performance

scale physical units into normalized output signals [e1, e2, e3]. An
interpretation of the signals, weighting functions, and models follows.
Signal

Meaning

d1

Normalized reference command

d 1

Typical reference command in physical units

d2

Normalized exogenous disturbances

d 2

Typical exogenous disturbances in physical units

d3

Normalized sensor noise

d 3

Typical sensor noise in physical units

e1

Weighted control signals

e 1

Actual control signals in physical units

e2

Weighted tracking errors

e 2

Actual tracking errors in physical units

e3

Weighted plant errors

e 3

Actual plant errors in physical units

Wcmd
Wcmd is included in H control problems that require tracking of a reference
command. Wcmd shapes the normalized reference command signals
(magnitude and frequency) into the actual (or typical) reference signals that
you expect to occur. It describes the magnitude and the frequency dependence
of the reference commands generated by the normalized reference signal.
Normally Wcmd is flat at low frequency and rolls off at high frequency. For
example, in a flight control problem, fighter pilots generate stick input
reference commands up to a bandwidth of about 2 Hz. Suppose that the stick
has a maximum travel of three inches. Pilot commands could be modeled as
normalized signals passed through a first-order filter

5-5

H-Infinity and Mu Synthesis

3
W act = --------------------------1
-------------- s + 1
2 2

Wmodel
Wmodel represents a desired ideal model for the closed-looped system and is
often included in problem formulations with tracking requirements. Inclusion
of an ideal model for tracking is often called a model matching problem, i.e., the
objective of the closed-loop system is to match the defined model. For good
command tracking response, you might want the closed-loop system to respond
like a well-damped second-order system. The ideal model would then be
W model =

------------------------------------2
2
s + 2 +

for specific desired natural frequency and desired damping ratio . Unit
conversions might be necessary to ensure exact correlation between the ideal
model and the closed-loop system. In the fighter pilot example, suppose that
roll-rate is being commanded and 10/second response is desired for each inch
of stick motion. Then, in these units, the appropriate model is
W model =

------------------------------------10
2
2
s + 2 +

Wdist
Wdist shapes the frequency content and magnitude of the exogenous
disturbances affecting the plant. For example, consider an electron microscope
as the plant. The dominant performance objective is to mechanically isolate the
microscope from outside mechanical disturbances, such as ground excitations,
sound (pressure) waves, and air currents. You can capture the spectrum and
relative magnitudes of these disturbances with the transfer function weighting
matrix Wdist.

Wperf1
Wperf1 weights the difference between the response of the closed-loop system
and the ideal model Wmodel. Often you might want accurate matching of the
ideal model at low frequency and require less accurate matching at higher
frequency, in which case Wperf1 is flat at low frequency, rolls off at first or
second order, and flattens out at a small, nonzero value at high frequency. The

5-6

H-Infinity Performance

inverse of the weight is related to the allowable size of tracking errors, when
dealing with the reference commands and disturbances described by Wref and
Wdist.

Wperf2
Wperf2 penalizes variables internal to the process G, such as actuator states
that are internal to G or other variables that are not part of the tracking
objective.

Wact
Wact is used to shape the penalty on control signal use. Wact is a frequency
varying weighting function used to penalize limits on the deflection/position,
deflection rate/velocity, etc., response of the control signals, when dealing with
the tracking and disturbance rejection objectives defined above. Each control
signal is usually penalized independently.

Wsnois
Wsnois represents frequency domain models of sensor noise. Each sensor
measurement feedback to the controller has some noise, which is often higher
in one frequency range than another. The Wsnois weight tries to capture this
information, derived from laboratory experiments or based on manufacturer
measurements, in the control problem. For example, medium grade
accelerometers have substantial noise at low frequency and high frequency.
Therefore the corresponding Wsnois weight would be larger at low and high
frequency and have a smaller magnitude in the mid-frequency range.
Displacement or rotation measurement is often quite accurate at low frequency
and in steady state, but responds poorly as frequency increases. The weighting
function for this sensor would be small at low frequency, gradually increase in
magnitude as a first- or second-order system, and level out at high frequency.

Hsens
Hsens represents a model of the sensor dynamics or an external antialiasing
filter. The transfer functions used to describe Hsens are based on physical
characteristics of the individual components. These models might also be
lumped into the plant model G.
This generic block diagram has tremendous flexibility and many control
performance objectives can be formulated in the H framework using this
block diagram description.

5-7

H-Infinity and Mu Synthesis

Robustness in the H Framework

Performance and robustness tradeoffs in control design were discussed in the
context of multivariable loop shaping in Tradeoff Between Performance and
Robustness on page 2-2. In the H control design framework, you can include
robustness objectives as additional disturbance to error transfer functions
disturbances to be kept small. Consider the following figure of a closed-loop
feedback system with additive and multiplicative uncertainty models.








 







The transfer function matrices are defined as

-1

TF(s) z 1 w 1

TI ( s )

KG ( I + GK ) ,

TF(s) z 2 w 2

KS O ( s )

K ( I + GK )

where TI(s) denotes the input complementary sensitivity function and SO(s)
denotes the output sensitivity function. Theorems 1 and 2 in Chapter 2 give
bounds on the size of the transfer function matrices from z1 to w1 and z2 to w2
to ensure that the closed-loop system is robust to multiplicative uncertainty,
M(s), at the plant input, and additive uncertainty, A(s), around the plant
G(s). In the H control problem formulation, the robustness objectives enter
the synthesis procedure as additional input/output signals to be kept small.
The interconnection with the uncertainty blocks removed follows.



5-8


 









H-Infinity Performance

The H control robustness objective is now in the same format as the

performance objectives, that is, to minimize the H norm of the transfer matrix
from z, [z1,z2], to w, [w1,w2].
Weighting or scaling matrices are often introduced to shape the frequency and
magnitude content of the sensitivity and complementary sensitivity transfer
function matrices. Let WM correspond to the multiplicative uncertainty and WA
correspond to the additive uncertainty model. M(s) and A(s) are assumed to
be a norm bounded by 1, i.e., |M(s)|<1 and |A(s)|<1. Hence as a function of
frequency, |WM(j)| and |WA(j)| are the respective sizes of the largest
anticipated additive and multiplicative plant perturbations.
The multiplicative weighting or scaling WM represents a percentage error in
the model and is often small in magnitude at low frequency, between 0.05 and
0.20 (5% to 20% modeling error), and growing larger in magnitude at high
frequency, 2 to 5 ((200% to 500% modeling error). The weight will transition by
crossing a magnitude value of 1, which corresponds to 100% uncertainty in the
model, at a frequency at least twice the bandwidth of the closed-loop system. A
typical multiplicative weight is

1
--- s + 1
5
= 0.10 ----------------------M
1
---------- s + 1
200

By contrast, the additive weight or scaling WA represents an absolute error

that is often small at low frequency and large in magnitude at high frequency.
The magnitude of this weight depends directly on the magnitude of the plant
model, G(s).

5-9

H-Infinity and Mu Synthesis

Application of H and mu to Active Suspension Control

Conventional passive suspensions employ a spring and damper between the
car body and wheel assembly, representing a tradeoff between conflicting
performance metrics such as passenger comfort, road holding, and suspension
deflection. Active suspensions allow the designer to balance these objectives
using a hydraulic actuator, controlled by feedback, between the chassis and
wheel assembly.
In this section, you design an active suspension system for a quarter car body
and wheel assembly model using the H control design technique. You will see
the tradeoff between passenger comfort, i.e., minimizing car body travel,
versus suspension travel as the performance objective.

Quarter Car Suspension Model

The quarter car model shown is used to design active suspension control laws.









 















 





 

The sprung mass ms represents the car chassis, while the unsprung mass mus
represents the wheel assembly. The spring, ks, and damper, bs, represent a
passive spring and shock absorber that are placed between the car body and
the wheel assembly, while the spring kt serves to model the compressibility of
the pneumatic tire. The variables xs, xus, and r are the car body travel, the
wheel travel, and the road disturbance, respectively. The force fs, kN, applied
between the sprung and unsprung masses, is controlled by feedback and
represents the active component of the suspension system. The dynamics of the

5-10

Application of H and mu to Active Suspension Control

actuator are ignored in this example, and assume that the control signal is the

force fs . Defining x1:=xs, x 2 : = x s , x3:=xus and x 4 : = x us , the following is the

state-space description of the quarter car dynamics.

x 1 = x 2,

1
x 2 = ------- [ k s ( x 1 x 3 ) + b s ( x 2 x 4 ) f s ],
ms

x 3 = x 4,

1
x 4 = ---------- [ k s ( x 1 x 3 ) + b s ( x 2 x 4 ) k t ( x 3 r ) f s ].
m us
The following component values are taken from reference [Lin97].
ms = 290;
mus = 59;
bs = 1000;
ks = 16182 ;
kt = 190000;

%
%
%
%
%

kg
kg
N/m/s
N/m
N/m

A linear, time-invariant model of the quarter car model, qcar, is constructed

from the equations of motion and parameter values. The inputs to the model
are the road disturbance and actuator force, respectively, and the outputs are
the car body deflection, acceleration, and suspension deflection.
A12 = [ 0 1 0 0; [-ks -bs ks bs]/ms ];
A34 = [ 0 0 0 1; [ks bs -ks-kt -bs]/mus];
B12 = [0 0; 0 10000/ms];
B34 = [0 0; [kt -10000]/mus];
C
= [1 0 0 0; A12(2,:); 1 0 -1 0; 0 0 0 0];
D
= [0 0; B12(2,:); 0 0; 0 1];
qcar = ss([A12; A34],[B12; B34],C,D);

It is well known [Hedrick90] that the acceleration transfer function has an

invariant point at the tirehop frequency, 56.7 rad/s. Similarly, the suspension
deflection transfer function has an invariant point at the rattlespace frequency,
23.3 rad/s. The tradeoff between passenger comfort and suspension deflection
is because it is not possible to simultaneously keep both transfer functions
small around the tirehop frequency and in the low frequency range.

5-11

H-Infinity and Mu Synthesis

Linear H Controller Design

The design of linear suspension controllers that emphasize either passenger
comfort or suspension deflection. The controllers in this section are designed
using linear H synthesis [FialBal]. As is standard in the H framework, the
performance objectives are achieved via minimizing weighted transfer function
norms.
Weighting functions serve two purposes in the H framework: They allow the
direct comparison of different performance objectives with the same norm, and
they allow for frequency information to be incorporated into the analysis. For
more details on H control design, refer to [DGKF], [Fran1], [GloD], [SkoP],
and [Zame]. A block diagram of the H control design interconnection for the
active suspension problem is shown below.






 




 

 











The measured output or feedback signal y is the suspension deflection x1x3.

The controller acts on this signal to produce the control input, the hydraulic
actuator force fs. The block Wn serves to model sensor noise in the
measurement channel. Wn is set to a sensor noise value of 0.01 m.
Wn = 0.01;

In a more realistic design, Wn would be frequency dependent and would serve

to model the noise associated with the displacement sensor. The weight Wref is
used to scale the magnitude of the road disturbances. Assume that the
maximum road disturbance is 7 cm and hence choose Wref = 0.07.
Wref = 0.07;

5-12

Application of H and mu to Active Suspension Control

The magnitude and frequency content of the control force fs are limited by the
100 s + 50
weighting function Wact. Choose W act = ---------- ------------------- . The magnitude of the
13 s + 500
weight increases above 50 rad/s in order to limit the closed-loop bandwidth.
Wact = (100/13)*tf([1 50],[1 500]);

H Control Design 1
The purpose of the weighting functions W x 1 and W x 1 x 3 is to keep the car
deflection and the suspension deflection small over the desired frequency
ranges. In the first design, you are designing the controller for passenger
comfort, and hence the car body deflection x1 is penalized.
Wx1 = 8*tf(2*pi*5,[1 2*pi*5]);

The weight magnitude rolls off above 52 rad/s to respect a well-known H

design rule of thumb that requires the performance weights to roll off before an
open-loop zero (56.7 rad/s in this case). The suspension deflection weight
W x x is not included in this control problem formulation.
1

You can construct the weighted H plant model for control design, denoted
qcaric1, using the sysic command. The control signal corresponds to the last
input to qcaric1, fs. The car body acceleration, which is noisy, is the measured
signal and corresponds to the last output of qcaric1.
systemnames = 'qcar Wn Wref Wact Wx1';
inputvar = '[ d1; d2; fs ]';
outputvar = '[ Wact; Wx1; qcar(3)+Wn ]';
input_to_qcar = '[ Wref; fs]';
input_to_Wn = '[ d2 ]';
input_to_Wref = '[ d1 ]';
input_to_Wact = '[ fs ]';
input_to_Wx1 = '[ qcar(1) ]';
qcaric1 = sysic;

An H controller is synthesized with the hinfsyn command. There is one

control input, the hydraulic actuator force, and one measurement signal, the
car body acceleration.
ncont = 1;
nmeas = 1;
[K1,Scl1,gam1] = hinfsyn(qcaric1,nmeas,ncont);

5-13

H-Infinity and Mu Synthesis

CL1 = lft(qcar([1:4 3],1:2),K1);

sprintf('H-infinity controller K1 achieved a norm of %2.5g',gam1)
ans =
H-infinity controller K1 achieved a norm of 0.61043

You can analyze the H controller by constructing the closed-loop feedback

system CL1. Bode magnitude plots of the passive suspension and active
suspension are shown in the following figure.
bodemag(qcar(3,1),'k-.',CL1(3,1),'r-',logspace(0,2.3,140))

H Control Design 2
In the second design, you are designing the controller to keep the suspension
deflection transfer function small. Hence the road disturbance to suspension
deflection x1x3 is penalized via the weighting function Wx1x3. The Wx1x3
weight magnitude rolls off above 10 rad/s to roll off before an open-loop zero
(23.3 rad/s) in the design.
Wx1x3 = 25*tf(1,[1/10 1]);

5-14

Application of H and mu to Active Suspension Control

The car deflection weight W x is not included in this control problem

1
formulation. You can construct the weighted H plant model for control design,
denoted qcaric2, using the sysic command. As an alternative, you can create
qcaric2 using iconnect objects. The same control and measurements are
used as in the first design.
M = iconnect;
d = icsignal(2);
fs = icsignal(1);
ycar = icsignal(size(qcar,1));
M.Equation{1} = equate(ycar,qcar*[Wref*d(1); fs]);
M.Input = [d;fs];
M.Output = [Wact*fs;Wx1x3*ycar(1);ycar(2)+Wn*d(2)];
qcaric2 = M.System;

The second H controller is synthesized with the hinfsyn command.

[K2,Scl2,gam2] = hinfsyn(qcaric2,nmeas,ncont);
CL2 = lft(qcar([1:4 2],1:2),K2);
sprintf('H-infinity controller K2 achieved a norm of %2.5g',gam2)
ans =
H-infinity controller K2 achieved a norm of 0.89949

Recall that this H control design emphasizes minimization of suspension

deflection over passenger comfort, whereas the first H design focused on
passenger comfort.
You can analyze the H controller by constructing the closed-loop feedback
system CL2. Bode magnitude plots of the transfer function from road
disturbance to suspension deflection for both controllers and the passive
suspension system are shown in the following figure.
bodemag(qcar(3,1),'k-.',CL1(3,1),'r-',CL2(3,1),'b.',...

5-15

H-Infinity and Mu Synthesis

logspace(0,2.3,140))

The dotted and solid lines in the figure are the closed-loop frequency responses
that result from the different performance weighting functions selected.
Observe the reduction in suspension deflection in the vicinity of the tirehop
frequency, 1= 56.7 rad/s, and the corresponding increase in the acceleration
frequency response in this vicinity. Also, compared to design 1, a reduction in
suspension deflection has been achieved for frequencies below the rattlespace
frequency, 2= 23.3 rad/s.
The second H control design attenuates both resonance modes, whereas the
first controller focused its efforts on the first mode, the rattlespace frequency
at 23.3 rad/s.
bodemag(qcar(2,1),'k-.',CL1(2,1),'r-',CL2(2,1),'b.',...

5-16

Application of H and mu to Active Suspension Control

logspace(0,2.3,140))

All the analysis till now has been in the frequency domain. Time-domain
performance characteristics are critical to the success of the active suspension
system on the car. Time response plots of the two H controllers are shown in
following figures. The dashed, solid, and dotted lines correspond to the passive
suspension, H controller 1, and controller 2 respectively. All responses
correspond to the road disturbance r(t)
r(t)

sec
= a ( 1 cos 8t ), 0 t 0.25
=

0 otherwise

where a=0.025 corresponds to a road bump of peak magnitude 5 cm. Observe

that the acceleration response of design 1 to the 5 cm bump is very good;
however the suspension deflection is larger than for design 2. This is because
suspension deflection was not penalized in this design. The suspension
deflection response of design 2 to a 5 cm bump is good; however the acceleration
response to the 5 cm bump is much inferior to design 1 (see the figure). Once
again this is because car body displacement and acceleration were not
penalized in design 2.

5-17

H-Infinity and Mu Synthesis

time = 0:0.005:1;
roaddist = 0*time;
roaddist(1:51) = 0.025*(1-cos(8*pi*time(1:51)));
[p1,t] = lsim(qcar(1:4,1),roaddist,time);
[y1,t] = lsim(CL1(1:4,1),roaddist,time);
[y2,t] = lsim(CL2(1:4,1),roaddist,time);
subplot(221)
plot(t,y1(:,1),'b-',t,y2(:,1),'r.',t,p1(:,1),'k--',t,...
roaddist,'g-.')
title('Body Travel')
ylabel('x_1 (m)')
subplot(222)
plot(t,y1(:,2),'b-',t,y2(:,2),'r.',t,p1(:,2),'k--')
title('Body Acceleration')
ylabel('Accel (m/s/s)')
subplot(223)
plot(t,y1(:,3),'b-',t,y2(:,3),'r.',t,p1(:,3),'k--')
title('Suspension Deflection')
xlabel('Time (sec)')
ylabel('x_1 - x_3 (m)')
subplot(224)
plot(t,y1(:,4),'b-',t,y2(:,4),'r.',t,p1(:,4),'k--')
title('Control Force')
xlabel('Time (sec)')

5-18

Application of H and mu to Active Suspension Control

ylabel('fs (10kN)')

Designs 1 and 2 represent extreme ends of the performance tradeoff spectrum.

This section described synthesis of H to achieve the performance objectives on
the active suspension system. Equally, if not more important, is the design of
controllers robust to model error or uncertainty.
The goal of every control design is to achieve the desired performance
specifications on the nominal model as well as other plants that are close to the
nominal model. In other words, you want to achieve the performance objectives
in the presence of model error or uncertainty. This is called robust performance.
In the next section, you will design a controller that achieves robust
performance using the -synthesis control design methodology. The active
suspension system again serves as the example. Instead of assuming a perfect
actuator, a nominal actuator model with modeling error is introduced into the
control problem.

5-19

H-Infinity and Mu Synthesis

Control Design via -Synthesis

The active suspension H controllers designed in the previous section ignored
the hydraulic actuator dynamics. In this section, you will include a first-order
model of the hydraulic actuator dynamics as well as an uncertainty model to
account for differences between the actuator model and the actual actuator
dynamics.
The nominal model for the hydraulic actuator is
1 actnom = -----------------1
------ s + 1
60
actnom = tf(1,[1/60 1]);

The actuator model itself is uncertain. You can describe the actuator model
error as a set of possible models using a weighting function. At low frequency,
below 4 rad/s, it can vary up to 10% from its nominal value. Around 4 rad/s the
percentage variation starts to increase and reaches 400% at approximately 800
rad/s. The model uncertainty is represented by the weight Wunc, which
corresponds to the frequency variation of the model uncertainty and the
uncertain LTI dynamic object unc defined as unc.
Wunc = 0.10*tf([1/4 1],[1/800 1]);
unc = ultidyn('unc',[1 1]);
actmod = actnom*(1+ Wunc*unc)
USS: 2 States, 1 Output, 1 Input, Continuous System
unc: 1x1 LTI, max. gain = 1, 1 occurrence

The actuator model actmod is an uncertain state-space system. The following

Bode plot shows the nominal actuator model, actnom, denoted with a '+'
symbol, and 20 random actuator models described by actmod.

5-20

Application of H and mu to Active Suspension Control

bode(actnom,'r+',actmod,'b',logspace(-1,3,120))

Nominal and 20 random actuator models

Magnitude (dB)

30
20
10
0

Phase (deg)

10
90
45
0
45
90
135
180
1
10

10

10

10

10

Frequency (rad/sec)

The uncertain actuator model actmod represents the model of the hydraulic
actuator used for control. The revised control design interconnection diagram
is






 



 




 

 








5-21

H-Infinity and Mu Synthesis

You are designing the controller for passenger comfort, as in the first H
control design, hence the car body deflection x1 is penalized with Wx1. The
uncertain weighted H plant model for control design, denoted qcaricunc, is
using the sysic command. As previously described, the control signal
corresponds to the last input to qcaric1, fs. The car body acceleration, which
is noisy, is the measured signal and corresponds to the last output of
qcaricunc.
systemnames = 'qcar Wn Wref Wact Wx1 actmod';
inputvar = '[ d1; d2; fs ]';
outputvar = '[ Wact; Wx1; qcar(2)+Wn ]';
input_to_actmod = '[ fs ]';
input_to_qcar = '[ Wref; fs]';
input_to_Wn = '[ d2 ]';
input_to_Wref = '[ d1 ]';
input_to_Wact = '[ fs ]';
input_to_Wx1 = '[ qcar(1) ]';
qcaricunc = sysic;

A -synthesis controller is synthesized using DK iteration with the dksyn

command. The DK iteration procedure is an approximation to -synthesis
that attempts to synthesize a controller that achieves robust performance
[SteD], [BalPac], [PacDB], [SkoP]. There is one control input, the hydraulic
actuator force, and one measurement signal, the car body acceleration.
[Kdk,CLdk,gdk] = dksyn(qcaricunc,nmeas,ncont);
CLdkunc = lft(qcar([1:4 2],1:2)*blkdiag(1,actmod),Kdk);
sprintf('mu-synthesis controller Kdk achieved a norm of
%2.5g',gdk)
ans =
mu-synthesis controller Kdk achieved a norm of 0.53946

You can analyze the performance of the -synthesis controller by constructing

the closed-loop feedback system CLdkunc. Bode magnitude plots of the passive
suspension and active suspension systems on the nominal actuator model with
H design 1 and the -synthesis controller are shown in the following figure.
Note that the -synthesis controller better attenuates the first resonant mode
at the expense of decreased performance below 3 rad/s.
bodemag(qcar(3,1),'k-.',CL1(3,1),'r-',CLmuunc.Nominal(3,1),'b.',
...

5-22

Application of H and mu to Active Suspension Control

logspace(0,2.3,140))
)

Bode Diagram
From: Road Disturbance To: Suspension Deflection
15
10
5

Magnitude (dB)

0
5
10
15
20
25
Passive
Active (K1)
Active (Kmu)

30
35
0
10

10

10

Frequency (rad/sec)

It is important to understand how robust both controllers are in the presence

of model error. You can simulate the active suspension system with the H
design 1 and the -synthesis controller. The uncertain closed-loop systems,
CL1unc and CLdkunc, are formed with K1 and Kdk, respectively. For each
uncertain system, 40 random plant models in the model set are simulated. As
you can see, both controllers are robust and perform well in the presence of
actuator model error. The -synthesis controller Kdk achieves slightly better
performance than H design 1.
CL1unc = lft(qcar([1:4 2],1:2)*blkdiag(1,actmod),K1);
[CLdkunc40,dksamples] = usample(CLdkunc,40);
CL1unc40 = usubs(CL1unc,dksamples);
nsamp = 40;
for i=1:nsamp
[ymusamp,t] = lsim(CLmuunc40(1:4,1,i),roaddist,time);
plot(t,ymusamp(:,1),'b')
hold on
end

5-23

H-Infinity and Mu Synthesis

[ymusamp,t] = lsim(CLmuunc.Nominal(1:4,1),roaddist,time);
plot(t,ymusamp(:,1),'r+',t,roaddist,'m--')

Random sample of 40 plant models: DK Controller

0.07
0.06

Body Travel (m)

0.05
0.04
0.03
0.02
0.01
0
0.01
0.02
0.03

0.2

0.4
0.6
Time (seconds)

0.8

for i=1:nsamp
[y1samp,t] = lsim(CL1unc40(1:4,1,i),roaddist,time);
plot(t,y1samp(:,1),'b')
hold on
end
[y1samp,t] = lsim(CL1unc.Nominal(1:4,1),roaddist,time);

5-24

Application of H and mu to Active Suspension Control

plot(t,y1samp(:,1),'r+',t,roaddist,'m--')
Random sample of 40 plant models: H Design 1

0.07
0.06

Body Travel (m)

0.05
0.04
0.03
0.02
0.01
0
0.01
0.02
0.03

0.2

0.4
0.6
Time (seconds)

0.8

5-25

H-Infinity and Mu Synthesis

Functions for Control Design

The term control system design refers to the process of synthesizing a feedback
control law that meets design specifications in a closed-loop control system.
The design methods are iterative, combining parameter selection with
analysis, simulation, and insight into the dynamics of the plant. The Robust
Control Toolbox provides a set of commands that you can use for a broad range
of multivariable control applications, including
H2 control design
H standard and loop-shaping control design
H normalized coprime factor control design
Mixed H2/H control design
-synthesis via DK iteration
Sampled-data H control design
These functions cover both continuous and discrete-time problems. The
following table summarizes the H2 and H control design commands.

5-26

Function

Description

augw

Augments plant weights for mixed-sensitivity control

design

h2hinfsyn

Mixed H2/H controller synthesis

h2syn

H2 controller synthesis

hinfsyn

H controller synthesis

loopsyn

H loop-shaping controller synthesis

ltrsyn

Loop-transfer recovery controller synthesis

mixsyn

H mixed-sensitivity controller synthesis

ncfsyn

H normalized coprime factor controller synthesis

sdhinfsyn

Sample-data H controller synthesis

Functions for Control Design

The following table summarizes -synthesis via DK iteration control design

commands.
Function

Description

dksyn

Synthesis of a robust controller via -synthesis

dkitopt

Create a dksyn options object

drawmag

Interactive mouse-based sketching and fitting tool

fitfrd

Fit scaling frequency response data with LTI model

fitmagfrd

Fit scaling magnitude data with stable, minimum-phase

model

5-27

H-Infinity and Mu Synthesis

Interpretation of H-Infinity Norm

Norms of Signals and Systems
There are several ways of defining norms of a scalar signal e(t) in the time
domain. We will often use the 2-norm, (L2-norm), for mathematical
convenience, which is defined as

2
e 2 :=
e ( t ) dt

1
--2

If this integral is finite, then the signal e is square integrable, denoted as e

L2. For vector-valued signals,

e1 ( t )
e2 ( t )

e( t) =

en ( t )
the 2-norm is defined as
e 2 := (

= (

e(t)

1
--2
2
dt
)
2

1
--2
T
e ( t )e ( t )dt )

In -tools the dynamic systems we deal with are exclusively linear, with
state-space model
x = A B x
C D d
e
or, in the transfer function form,
e(s) = T(s)d(s),T(s):= C(sI A)1B + D
Two mathematically convenient measures of the transfer matrix T(s) in the
frequency domain are the matrix H2 and H norms,

5-28

Interpretation of H-Infinity Norm

1
--2
1
-----T 2 :=
T ( j ) F d 2
2

T := max [ T ( j ) ]
wR

where the Frobenius norm (see the MATLAB norm command) of a complex
matrix M is
M F :=

trace ( M*M )

Both of these transfer function norms have input/output time-domain

interpretations. If, starting from initial condition x(0) = 0, two signals d and e
are related by
x = A B x
C D d
e
then
For d, a unit intensity, white noise process, the steady-state variance of e is
||T||2.
The L2 (or RMS) gain from d e ,
e 2
max ----------d0 d 2
is equal to ||T||. This is discussed in greater detail in the next section.

Using Weighted Norms to Characterize

Performance
Any performance criterion must also account for
Relative magnitude of outside influences
Frequency dependence of signals
Relative importance of the magnitudes of regulated variables
So, if the performance objective is in the form of a matrix norm, it should
actually be a weighted norm
||WLTWR||

5-29

H-Infinity and Mu Synthesis

where the weighting function matrices WL and WR are frequency dependent, to

account for bandwidth constraints and spectral content of exogenous signals.
The most natural (mathematical) manner to characterize acceptable
performance is in terms of the MIMO |||| (H) norm. For this reason, this
section now discusses some interpretations of the H norm.

e~

d~

Unweighted MIMO System

Suppose T is a MIMO stable linear system, with transfer function matrix T(s).
For a given driving signal d ( t ) , define e as the output, as shown below.
Note that it is more traditional to write the diagram in Unweighted MIMO
System: Vectors from Left to Right on page 5-30 with the arrows going from
left to right as in Weighted MIMO System on page 5-32.

d~

-e~

Unweighted MIMO System: Vectors from Left to Right

The two diagrams shown above represent the exact same system. We prefer to
write these block diagrams with the arrows going right to left to be consistent
with matrix and operator composition.
Assume that the dimensions of T are ne nd. Let > 0 be defined as
:= T := max [ T ( j ) ]
wR

(11-1)

Now consider a response, starting from initial condition equal to 0. In that case,
Parsevals theorem gives that
12

T
[ e ( t )e ( t ) dt ]
e 2
0
----------- = -----------------------------------------------------
12
T
d 2
d ( t )d ( t ) dt

5-30

Interpretation of H-Infinity Norm

e 2
Moreover, there are specific disturbances d that result in the ratio ----------d 2
arbitrarily close to . Because of this, ||T|| is referred to as the L2 (or RMS) gain
of the system.
As you would expect, a sinusoidal, steady-state interpretation of ||T|| is also
possible: For any frequency R , any vector of amplitudes a R n , and any
d

vector of phases R n d , with ||a||2 1, define a time signal

a 1 sin ( t + 1 )

d ( t ) =

a n sin ( t + n )
d

Applying this input to the system T results in a steady-state response e ss of the

form
b 1 sin ( t + 1 )

e ss ( t ) =

b ne sin ( t + ne )
The vector b R n e will satisfy ||b||2 . Moreover, , as defined in Weighted
MIMO System on page 5-32, is the smallest number such that this is true for
every ||a||2 1, , and .
Note that in this interpretation, the vectors of the sinusoidal magnitude
responses are unweighted, and measured in Euclidean norm. If realistic
multivariable performance objectives are to be represented by a single MIMO
|||| objective on a closed-loop transfer function, additional scalings are
necessary. Because many different objectives are being lumped into one matrix
and the associated cost is the norm of the matrix, it is important to use
frequency-dependent weighting functions, so that different requirements can
be meaningfully combined into a single cost function. Diagonal weights are
most easily interpreted.

5-31

H-Infinity and Mu Synthesis

Consider the diagram of Weighted MIMO System on page 5-32, along with
Unweighted MIMO System: Vectors from Left to Right on page 5-30.
Assume that WL and WR are diagonal, stable transfer function matrices, with
diagonal entries denoted Li and Ri.
L1 0 0

e~

WL

WL e~

d~

WL T d~

WR

0 R2 0
..
.
0 0 Rn

WR =

0 L2 0
..
.
0 0 Ln

WL =

R1 0 0

WL T WR d

Weighted MIMO System

Bounds on the quantity ||WLTWR|| will imply bounds about the sinusoidal
steady-state behavior of the signals d and e (= Td ) in the diagram of
Unweighted MIMO System: Vectors from Left to Right on page 5-30.
Specifically, for sinusoidal signal d , the steady-state relationship between
e (= Td ) , d and ||WLTWR|| is as follows. The steady-state solution e ss , denoted
as

e ss ( t ) =

e 1 sin ( t + 1 )
e n e sin ( t + n )
d

satisfies

5-32

ne
i=1

(5-1)

W L i ( jw )e i 1 for all sinusoidal input signals d of the form

Interpretation of H-Infinity Norm

d ( t ) =

d 1 sin ( t + i )
d n d sin ( t + n )
d

(5-2)

satisfying
nd

2
d i
--------------------------1
2
W
(
j
)
Ri
i=1

if and only if ||WLTWR|| 1.

This approximately (very approximately the next statement is not actually
correct) implies that ||WLTWR|| 1 if and only if for every fixed frequency ,
and all sinusoidal disturbances d of the form (5-2) satisfying
d i W R ( j )
i
the steady-state error components will satisfy
1
e i -----------------------W L i ( j )
This shows how one could pick performance weights to reflect the desired
frequency-dependent performance objective. Use WR to represent the relative
magnitude of sinusoids disturbances that might be present, and use

1-------W
L

to

represent the desired upper bound on the subsequent errors that are produced.
Remember, however, that the weighted H norm does not actually give
element-by-element bounds on the components of e based on
element-by-element bounds on the components of d . The precise bound it gives
is in terms of Euclidean norms of the components of e and d (weighted
appropriately by WL(j ) and WR(j )).

5-33

H-Infinity and Mu Synthesis

References
[BalPac:] Balas, G.J., and A.K. Packard, The structured singular value
-framework, CRC Controls Handbook, Section 2.3.6, January, 1996,
pp. 671-688.
[BallC:] Ball, J.A., and N. Cohen, Sensitivity minimization in an H norm:
Parametrization of all suboptimal solutions, International Journal of Control,
Vol. 46, 1987, pp. 785-816.
[BamP:] Bamieh, B.A., and Pearson, J.B., A general framework for linear
periodic systems with applications to H sampled-data control, IEEE
Transactions on Automatic Control, Vol. AC-37, 1992, pp. 418-435.
[DGKF:] Doyle, J.C., Glover, K., Khargonekar, P., and Francis, B., State-space
solutions to standard H2 and H control problems, IEEE Transactions on
Automatic Control, Vol. AC-34, No. 8, August 1989, pp. 831-847.
[FialBal:] Fialho, I., and Balas, G.J., Design of nonlinear controllers for active
vehicle suspensions using parameter-varying control synthesis, Vehicle
Systems Dynamics, Vol. 33, No. 5, May 2000, pp. 351-370.
[Fran1:] Francis, B.A., A course in H control theory, Lecture Notes in Control
and Information Sciences, Vol. 88, Springer-Verlag, Berlin, 1987.
[GloD:] Glover, K., and Doyle, J.C., State-space formulae for all stabilizing
controllers that satisfy an H norm bound and relations to risk sensitivity,
Systems and Control Letters, vol. 11, pp. 167-172, August 1989. International
Journal of Control, Vol. 39, 1984, pp. 1115-1193.
[Hedrick90:] Hedrick, J.K., and Batsuen, T., Invariant Properties of
Automotive Suspensions, Proceedings of The Institution of Mechanical
Engineers, 204 (1990), pp. 21-27.
[Lin971:] Lin, J., and Kanellakopoulos, I., Road Adaptive Nonlinear Design of
Active Suspensions, Proceedings of the American Control Conference, (1997),
pp. 714-718.
[PacDB:] Packard, A.K., Doyle, J.C., and Balas, G.J., Linear, multivariable
robust control with a perspective, ASME Journal of Dynamics,
Measurements and Control: Special Edition on Control, Vol. 115, No. 2b, June,
1993, pp. 426-438.
[SkoP:] Skogestad, S., and Postlethwaite, I., Multivariable Feedback Control:
Analysis & Design, John Wiley & Sons, 1996.

5-34

References

[SteD:] Stein, G., and Doyle, J., Beyond singular values and loopshapes, AIAA
Journal of Guidance and Control, Vol. 14, Num. 1, January, 1991, pp. 5-16.
[Zame:] Zames, G., Feedback and optimal sensitivity: model reference
transformations, multiplicative seminorms, and approximate inverses, IEEE
Transactions on Automatic Control, Vol. AC-26, 1981, pp. 301-320.

5-35

H-Infinity and Mu Synthesis

5-36

6
Building Uncertain
Models

Introduction to Uncertain Atoms

(p. 6-2)

How to build uncertain real, complex, and LTI uncertain

matrices, dynamics, and systems

Uncertain Matrices (p. 6-19)

How to manipulate matrices used in systems with

structured uncertainty

Uncertain State-Space Systems (uss)

(p. 6-25)

Building systems with uncertain state-space matrices

and/or uncertain linear dynamics

Uncertain frd (p. 6-33)

Discusses uncertain frequency response data (frd)

objects

Basic Control System Toolbox and

MATLAB Interconnections (p. 6-37)

A list of Control System Toolbox interconnection

commands that work with uncertain objects

Simplifying Representation of
Uncertain Objects (p. 6-38)

How to simplify representations of uncertain objects in

your models

Sampling Uncertain Objects (p. 6-42)

How to randomly sample uncertain objects

Substitution by usubs (p. 6-46)

How to fix a subset of uncertain objects in your model,

while leaving the rest as uncertain

Array Management for Uncertain

Objects (p. 6-49)

Working with multidimensional arrays containing

uncertain objects

Decomposing Uncertain Objects (for

Advanced Users) (p. 6-60)

Discusses advanced decomposition techniques

Building Uncertain Models

Introduction to Uncertain Atoms

Uncertain atoms are the building blocks used to form uncertain matrix
objects and uncertain system objects. There are 5 classes of uncertain atoms:
Function

Description

ureal

Uncertain real parameter

ultidyn

Uncertain, linear, time-invariant dynamics

ucomplex

Uncertain complex parameter

ucomplexm

Uncertain complex matrix

udyn

Uncertain dynamic system

All of the atoms have properties, which are accessed through get and set
methods. This get and set interface mimics the Control System Toolbox and
MATLAB Handle Graphics behavior. For instance,
get(a,'PropertyName') is the same as a.PropertyName, and
set(b,'PropertyName',Value) is the same as b.PropertyName = value.
Functionality also includes tab-completion and case-insensitive, partial name
property matching.
For ureal, ucomplex and ucomplexm atoms, the syntax is
p1 = ureal(name, NominalValue, Prop1, val1, Prop2, val2,...);
p2 = ucomplex(name, NominalValue, Prop1, val1, Prop2, val2,...);
p3 = ucomplexm(name, NominalValue, Prop1, val1, Prop2,
val2,...);

For ultidyn and udyn, the NominalValue is fixed, so the syntax is

p4 = ultidyn(name, ioSize, Prop1, val1, Prop2, val2,...);
p5 = udyn(name, ioSize, Prop1, val1, Prop2, val2,...);

For ureal, ultidyn, ucomplex and ucomplexm atoms, the command usample
will generate a random instance (i.e., not uncertain) of the atom, within its
modeled range. For example,

6-2

Introduction to Uncertain Atoms

usample(p1)

creates a random instance of the uncertain real parameter p1. With an

integer argument, whole arrays of instances can be created. For instance
usample(p4,100)

generates an array of 100 instances of the ultidyn object p4. See Sampling
Uncertain Objects on page 6-42 to learn more about usample.

Uncertain Real Parameters

An uncertain real parameter is used to represent a real number whose value
is uncertain. Uncertain real parameters have a name (the Name property), and
a nominal value (the NominalValue property). Several other properties
(PlusMinus, Range, Percentage) describe the uncertainty in the parameter's
value.
All properties of a ureal can be accessed through get and set. The properties
are:
Properties

Meaning

Class

Name

Internal Name

char

NominalValue

Nominal value of atom

double

Mode

Signifies which description (from

char

'PlusMinus','Range','Percentage') of

uncertainty is invariant when

NominalValue is changed
PlusMinus

Additive variation

scalar or
1x2 double

Range

Numerical range

1x2 double

Percentage

Additive variation (% of absolute value of

nominal)

scalar or
1x2 double

AutoSimplify

'off' | {'basic'} |'full'

char

6-3

Building Uncertain Models

The properties Range, Percentage and PlusMinus are all automatically

synchronized. If the nominal value is 0, then the Mode cannot be Percentage.
The Mode property controls what aspect of the uncertainty remains
unchanged when NominalValue is changed. Assigning to any of
Range/Percentage/PlusMinus changes the value, but does not change the
mode.
The AutoSimplify property controls how expressions involving the real
parameter are simplified. Its default value is 'basic', which means
elementary methods of simplification are applied as operations are
completed. Other values for AutoSimplify are 'off' (no simplification
performed) and 'full' (model-reduction-like techniques are applied). See
Simplifying Representation of Uncertain Objects on page 6-38 to learn more
about the AutoSimplify property and the command simplify.
If no property/value pairs are specified, default values are used. The default
Mode is PlusMinus, and the default value of PlusMinus is [-1 1]. Some

examples are shown below. In many cases, the full property name is not
specified, taking advantage of the case-insensitive, partial name property
matching.
Create an uncertain real parameter, nominal value 3, with default values for
all unspecified properties (including plus/minus variability of 1). View the
properties and their values, and note that the Range and Percentage
descriptions of variability are automatically maintained.
a = ureal('a',3)
Uncertain Real Parameter: Name a, NominalValue 3, variability =
[-1 1]
get(a)
Name: 'a'
NominalValue: 3
Mode: 'PlusMinus'
Range: [2 4]
PlusMinus: [-1 1]
Percentage: [-33.3333 33.3333]
AutoSimplify: 'basic'

Create an uncertain real parameter, nominal value 2, with 20% variability.

Again, view the properties, and note that the Range and PlusMinus
descriptions of variability are automatically maintained.

6-4

Introduction to Uncertain Atoms

b = ureal('b',2,'percentage',20)
Uncertain Real Parameter: Name b, NominalValue 2, variability =
[-20 20]%
get(b)
Name: 'b'
NominalValue: 2
Mode: 'Percentage'
Range: [1.6000 2.4000]
PlusMinus: [-0.4000 0.4000]
Percentage: [-20.0000 20.0000]
AutoSimplify: 'basic'

Change the range of the parameter. All descriptions of variability are

automatically updated, while the nominal value remains fixed. Although the
change in variability was accomplished by specifying the Range, the Mode is
unaffected, and remains Percentage.
b.Range = [1.9 2.3];
get(b)
Name: 'b'
NominalValue: 2
Mode: 'Percentage'
Range: [1.9000 2.3000]
PlusMinus: [-0.1000 0.3000]
Percentage: [-5.0000 15.0000]
AutoSimplify: 'basic'

As mentioned, the Mode property signifies what aspect of the uncertainty

remains unchanged when NominalValue is modified. Hence, if a real
parameter is in Percentage mode, then the Range and PlusMinus properties
are determined from the Percentage property and NominalValue. Changing
NominalValue preserves the Percentage property, and automatically
updates the Range and PlusMinus properties.
b.NominalValue = 2.2;
get(b)
Name: 'b'
NominalValue: 2.2000
Mode: 'Percentage'
Range: [2.0900 2.5300]
PlusMinus: [-0.1100 0.3300]

6-5

Building Uncertain Models

Percentage: [-5.0000 15.0000]

AutoSimplify: 'basic'

Create an uncertain parameter with an unsymmetric variation about its

nominal value.
c = ureal('c',-5,'per',[-20 30]);
get(c)
Name: 'c'
NominalValue: -5
Mode: 'Percentage'
Range: [-6 -3.5000]
PlusMinus: [-1 1.5000]
Percentage: [-20 30]
AutoSimplify: 'basic'

Create an uncertain parameter, specifying variability with Percentage, but

force the Mode to be Range.
d = ureal('d',-1,'mode','range','perc',[-40 60]);
get(d)
Name: 'd'
NominalValue: -1
Mode: 'Range'
Range: [-1.4000 -0.4000]
PlusMinus: [-0.4000 0.6000]
Percentage: [-40.0000 60]
AutoSimplify: 'basic'

Finally, create an uncertain real parameter, and set the AutoSimplify

property to 'full'.
e = ureal('e',10,'plusminus',[-23],'mode','perce',...
'autosimplify','full')
Uncertain Real Parameter: Name e, NominalValue 10, variability
= [-20 30]%
get(e)
Name: 'e'
NominalValue: 10
Mode: 'Percentage'
Range: [8 13]
PlusMinus: [-2 3]

6-6

Introduction to Uncertain Atoms

Percentage: [-20 30]

AutoSimplify: 'full'

Specifying conflicting values for Range/Percentage/PlusMinus in a multiple

property/value set is not an error. In this case, the last (in list) specified
property is used. This last occurrence also determines the Mode, unless Mode
is explicitly specified, in which case that is used, regardless of the
property/value pairs ordering.
f = ureal('f',3,'plusminus',[-2 1],'perce',40)
Uncertain Real Parameter: Name f, NominalValue 3, variability =
[-40 40]%
g = ureal('g',2,'plusminus',[-2 1],'mode','range','perce',40)
Uncertain Real Parameter: Name g, NominalValue 2, Range [1.2
2.8]
g.Mode
ans =
Range

Create an uncertain real parameter, use usample to generate 1000 instances

(resulting in a 1-by-1-by-1000 array), reshape the array, and plot a
histogram, with 20 bins (within the range of 2-to-4).
h = ureal('h',3);
hsample = usample(h,1000);
hist(reshape(hsample,[1000 1]),20);

6-7

Building Uncertain Models

60

50

40

30

20

10

2.2

2.4

2.6

2.8

3.2

3.4

3.6

3.8

Make the range unsymmetric about the nominal value, and repeat the
sampling, and histogram plot (with 40 bins over the range of 2-to-6)
h.Range = [2 6];
hsample = usample(h,1000);
hist(reshape(hsample,[1000 1]),40);

6-8

Introduction to Uncertain Atoms

80

70

60

50

40

30

20

10

2.5

3.5

4.5

5.5

Note that the distribution is skewed. However, the number of samples less
than the nominal value and the number of samples greater than the nominal
value is equal (on average). Verify this.
length(find(hsample(:)<h.NominalValue))
ans =
509
length(find(hsample(:)>h.NominalValue))
ans =
491

The distribution used in usample is uniform in the normalized description of

the uncertain real parameter. See Decomposing Uncertain Objects (for
Advanced Users) on page 6-60 to learn more about the normalized
description.
There is no notion of an empty ureal (or any other atom, for that matter).
ureal, by itself, creates an unnamed atom, with default property values. The
given name is 'UNNAMED'. This can be observed with get and set.
get(ureal)
Name: 'UNNAMED'
NominalValue: 0
Mode: 'PlusMinus'

6-9

Building Uncertain Models

Range:
PlusMinus:
Percentage:
AutoSimplify:
set(ureal)
Name:
NominalValue:
Mode:
Range:
PlusMinus:
Percentage:
AutoSimplify:

[-1 1]
[-1 1]
[-Inf Inf]
'basic'
'String'
'1x1 real DOUBLE'
'Range | PlusMinus'
'1x2 DOUBLE'
'1x2 or scalar DOUBLE'
'Not settable since Nominal==0'
'['off' | 'basic' | 'full']'

Uncertain LTI Dynamics Atoms

Uncertain linear, time-invariant objects, ultidyn, are used to represent
unknown linear, time-invariant dynamic objects, whose only known
attributes are bounds on their frequency response. Uncertain linear,
time-invariant objects have an internal name (the Name property), and are
created by specifying their size (number of outputs and number of inputs).
The property Type specifies whether the known attributes about the
frequency response are related to gain or phase. The property Type may be
'GainBounded' or 'PositiveReal'. The default value is 'GainBounded'.
The property Bound is a single number, which along with Type, completely
specifies what is known about the uncertain frequency response. Specifically,
if is an ultidyn atom, and if denotes the value of the Bound property, then
the atom represents the set of all stable, linear, time-invariant systems whose
frequency response satisfies certain conditions:

If Type is 'GainBounded', [ ( ) ] for all frequencies. When Type is

'GainBounded', the default value for Bound (i.e., ) is 1. The NominalValue of
is always the 0-matrix.
If Type is 'PositiveReal', ()+*()2 for all frequencies. When Type is
'PositiveReal', the default value for Bound (i.e., ) is 0. The NominalValue
is always (+1+2||)I.

6-10

Introduction to Uncertain Atoms

All properties of a ultidyn are can be accessed with get and set (although
the NominalValue is determined from Type and Bound, and not accessible
with set). The properties are
Properties

Meaning

Class

Name

Internal Name

char

NominalValue

Nominal value of atom

See above

Type

'GainBounded' |'PositiveReal'

char

Bound

Norm bound or minimum real

scalar double

SampleStateDim

State-space dimension of random

samples of this uncertain element

scalar double

AutoSimplify

'off' | {'basic'} |'full'

char

The SampleStateDim property specifies the state dimension of random

samples of the atom when using usample. The default value is 1. The
AutoSimplify property serves the same function as in the uncertain real
parameter.
You can create a 2-by-3 gain-bounded uncertain linear dynamics atom. Verify
its size, and check the properties.
f = ultidyn('f',[2 3]);
size(f)
ans =
2
3
get(f)
Name: 'f'
NominalValue: [2x3 double]
Type: 'GainBounded'
Bound: 1
SampleStateDim: 1
AutoSimplify: 'basic'

You can create a 1-by-1 (scalar) positive-real uncertain linear dynamics atom,
whose frequency response always has real part greater than -0.5. Set the

6-11

Building Uncertain Models

SampleStateDim property to 5. View the properties, and plot a Nyquist plot of

30 instances of the atom.

g = ultidyn('g',[1 1],'type','positivereal','bound',-0.5);
g.SampleStateDim = 5;
get(g)
Name: 'g'
NominalValue: 1.5000
Type: 'PositiveReal'
Bound: -0.5000
SampleStateDim: 5
AutoSimplify: 'basic'
nyquist(usample(g,30))
xlim([-2 10])
ylim([-6 6]);

Nyquist Diagram

Imaginary Axis

6
2

10

Real Axis

Time-domain of ultidyn atoms

On its own, every ultidyn atom is interpreted as a continuous-time, system
with uncertain behavior, quantified by bounds (gain or real-part) on its

6-12

Introduction to Uncertain Atoms

frequency response. To see this, create a ultidyn, and view the sample time
of several random samples of the atom.
h = ultidyn('h',[1 1]);
get(usample(h),'Ts')
ans =
0
get(usample(h),'Ts')
ans =
0
get(usample(h),'Ts')
ans =
0

However, when a ultidyn atom is an uncertain element of an uncertain state

space model (uss), then the time-domain characteristic of the atom is
determined from the time-domain characteristic of the system. The bounds
(gain-bounded or positivity) apply to the frequency-response of the atom. This
is explained and demonstrated in Interpreting Uncertainty in Discrete
Time on page 6-30.

Complex Parameter Atoms

The ucomplex atom represents an uncertain complex number, whose value
lies in a disc, centered at NominalValue, with radius specified by the Radius
property. The size of the disc can also be specified by Percentage, which
means the radius is derived from the absolute value of the NominalValue. The
properties of ucomplex objects are
Properties

Meaning

Class

Name

Internal Name

char

NominalValue

Nominal value of atom

double

Mode

'Range' | 'Percentage'

char

Radius

Radius of disk

double

6-13

Building Uncertain Models

Properties

Meaning

Percentage

Additive variation (percent of Radius)

AutoSimplify

'off' | {'basic'} | 'full'

Class
double
char

The simplest construction requires only a name and nominal value. The
default Mode is Radius, and the default radius is 1.
a = ucomplex('a',2-j)
Uncertain Complex Parameter: Name a, NominalValue 2-1i, Radius 1
get(a)
Name: 'a'
NominalValue: 2.0000- 1.0000i
Mode: 'Radius'
Radius: 1
Percentage: 44.7214
AutoSimplify: 'basic'
set(a)
Name: 'String'
NominalValue: '1x1 DOUBLE'
Mode: 'Radius | Percentage'
Radius: 'scalar DOUBLE'
Percentage: 'scalar DOUBLE'
AutoSimplify: '['off' | 'basic' | 'full']'

Sample the uncertain complex parameter at 400 values, and plot in the
complex plane. Clearly, the samples appear to be from a disc of radius 1,
centered in the complex plane at the value 2j.
asample = usample(a,400);
plot(asample(:),'o'); xlim([0 4]); ylim([-3 1]);

6-14

Introduction to Uncertain Atoms

0.5

0.5

1.5

2.5

3
0

0.5

1.5

2.5

3.5

Complex Matrix Atoms

The uncertain complex matrix class, ucomplexm, represents the set of
matrices given by the formula
N + W L W R
where N,WL,WR are known matrices, and is any complex matrix with

( ) 1 . All properties of a ucomplexm are can be accessed with get and

set. The properties are
Properties

Meaning

Class

Name

Internal Name

char

NominalValue

Nominal value of atom

double

WL

Left weight

double

6-15

Building Uncertain Models

Properties

Meaning

WR

Right weight

AutoSimplify

'off' | {'basic'} | 'full'

Class
double
char

The simplest construction requires only a name and nominal value. The
default left and right weight matrices are identity.
You can create a 4-by-3 ucomplexm element, and view its properties.
m = ucomplexm('m',[1 2 3;4 5 6;7 8 9;10 11 12])
Uncertain Complex Matrix: Name m, 4x3
get(m)
Name: 'm'
NominalValue: [4x3 double]
WL: [4x4 double]
WR: [3x3 double]
AutoSimplify: 'basic'
m.NominalValue
ans =
1
2
3
4
5
6
7
8
9
10
11
12
m.WL
ans =
1
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1

Sample the uncertain matrix, and compare to the nominal value. Note the
element-by-element sizes of the difference are generally equal, indicative of
the default (identity) weighting matrices that are in place.
abs(usample(m)-m.NominalValue)
ans =
0.2948
0.1001
0.2867
0.3028
0.2384
0.2508
0.3376
0.1260
0.2506

6-16

Introduction to Uncertain Atoms

0.2200

0.3472

0.1657

Change the left and right weighting matrices, making the uncertainty larger
as you move down the rows, and across the columns.
m.WL = diag([0.2 0.4 0.8 1.6]);
m.WR = diag([0.1 1 4]);

Sample the uncertain matrix, and compare to the nominal value. Note the
element-by-element sizes of the difference, and the general trend that the
smallest differences are near the (1,1) element, and the largest differences
are near the (4,3) element, which is completely expected by choice of the
diagonal weighting matrices.
abs(usample(m)-m.NominalValue)
ans =
0.0091
0.0860
0.2753
0.0057
0.1717
0.6413
0.0304
0.2756
1.4012
0.0527
0.4099
1.8335

Unstructured Uncertain Dynamic Systems

The unstructured uncertain dynamic system class, udyn, represents
completely unknown multivariable, time-varying nonlinear systems.
For practical purposes, these uncertain elements represent noncommuting
symbolic variables (placeholders). All algebraic operations, such as addition,
subtraction, multiplication (i.e., cascade) operate properly, and substitution
(with usubs) is allowed. However, all of the analysis tools (e.g., robuststab)
do not handle these types of uncertain elements. As such, these elements do
not provide a significant amount of usability, and their role in the user's
guide is small.
You can create a 2-by-3 udyn element. Check its size, and properties.
m = udyn('m',[2 3])
Uncertain Dynamic System: Name m, size 2x3
size(m)
ans =
2
3
get(m)
Name: 'm'

6-17

Building Uncertain Models

NominalValue: [2x3 double]

AutoSimplify: 'basic'

6-18

Uncertain Matrices

Uncertain Matrices
Uncertain matrices (class umat) are built from doubles, and uncertain atoms,
using traditional MATLAB matrix building syntax. Uncertain matrices can
be added, subtracted, multiplied, inverted, transposed, etc., resulting in
uncertain matrices. The rows and columns of an uncertain matrix are
referenced in the same manner that MATLAB references rows and columns
of an array, using parenthesis, and integer indices. The NominalValue of a
uncertain matrix is the result obtained when all uncertain atoms are replaced
with their own NominalValue. The uncertain atoms making up a umat are
accessible through the Uncertainty gateway, and the properties of each atom
within a umat can be changed directly.
Using usubs, specific values may be substituted for any of the uncertain
atoms within a umat. The command usample generates a random sample of
the uncertain matrix, substituting random samples (within their ranges) for
each of the uncertain atoms.
The command wcnorm computes tight bounds on the worst-case (maximum
over the uncertain elements' ranges) norm of the uncertain matrix.
Standard MATLAB numerical matrices (i.e., double) naturally can be viewed
as uncertain matrices without any uncertainty.

Creating Uncertain Matrices from Uncertain Atoms

You can create 2 uncertain real parameters, and then a 3-by-2 uncertain
matrix using these uncertain atoms.
a = ureal('a',3);
b = ureal('b',10,'pe',20);
M = [-a 1/b;b a+1/b;1 3]
UMAT: 3 Rows, 2 Columns
a: real, nominal = 3, variability = [-1 1], 2 occurrences
b: real, nominal = 10, variability = [-20 20]%, 3 occurrences

The size and class of M are as expected

size(M)
ans =
3
class(M)
ans =

6-19

Building Uncertain Models

umat

Accessing Properties of a umat

Use get to view the accessible properties of a umat.
get(M)
NominalValue: [3x2 double]
Uncertainty: [1x1 atomlist]

The NominalValue is a double, obtained by replacing all uncertain elements

with their nominal values.
M.NominalValue
ans =
-3.0000
0.1000
10.0000
3.1000
1.0000
3.0000

The Uncertainty property is a atomlist object, which is simply a gateway

from the umat to the uncertain atoms.
class(M.Uncertainty)
ans =
atomlist
M.Uncertainty
a: [1x1 ureal]
b: [1x1 ureal]

Direct access to the atoms is facilitated through Uncertainty. Check the

Range of the uncertain element named 'a' within M, then change it.
M.Uncertainty.a.Range
ans =
2
4
M.Uncertainty.a.Range = [2.5 5];
M
UMAT: 3 Rows, 2 Columns
a: real, nominal = 3, variability = [-0.5 2], 2 occurrences
b: real, nominal = 10, variability = [-20 20]%, 3 occurrences

6-20

Uncertain Matrices

The change to the uncertain real parameter a only took place within M. Verify
that the variable a in the workspace is no longer the same as the variable a
within M.
isequal(M.Uncertainty.a,a)
ans =
0

Note that combining atoms which have a common internal name, but
different properties leads to an error. For instance, subtracting the two atoms
gives an error, not 0.
M.Uncertainty.a - a
??? Error using ==> ndlft.lftmask

Atoms named 'a' have different properties.

Row and Column Referencing

Standard Row/Column referencing is allowed. Note, however, that
single-indexing is only allowed if the umat is a column or a row.
Reconstruct M (if need be), and make a 2-by-2 selection from M
a = ureal('a',3);
b = ureal('b',10,'pe',20);
M = [-a 1/b;b a+1/b;1 3];
M.Uncertainty.a.Range = [2.5 5];
M(2:3,:)
UMAT: 2 Rows, 2 Columns
a: real, nominal = 3, variability = [-0.5 2], 1 occurrence
b: real, nominal = 10, variability = [-20 20]%, 2 occurrences

Make a single column selection from M, and use single-index references to

access elements of it.
h = M([2 1 2 3],2)
UMAT: 4 Rows, 1 Columns
a: real, nominal = 3, variability = [-0.5 2], 1 occurrence
b: real, nominal = 10, variability = [-20 20]%, 1 occurrence
h(2)
UMAT: 1 Rows, 1 Columns
b: real, nominal = 10, variability = [-20 20]%, 1 occurrence

6-21

Building Uncertain Models

h(3)
UMAT: 1 Rows, 1 Columns
a: real, nominal = 3, variability = [-0.5 2], 1 occurrence
b: real, nominal = 10, variability = [-20 20]%, 1 occurrence

Finally, make the (3,2) entry of M uncertain.

M(3,2) = ureal('c',3,'perc',40)
UMAT: 3 Rows, 2 Columns
a: real, nominal = 3, variability = [-0.5 2], 2 occurrences
b: real, nominal = 10, variability = [-20 20]%, 2 occurrences
c: real, nominal = 3, variability = [-40 40]%, 1 occurrence

Matrix Operation on umat Objects

Many matrix operations are allowed, such as matrix-multiply, transpose, and
inverse. Combinations of certain (i.e., not uncertain) matrices and uncertain
matrices are allowed.
Premultiply M by a 1-by-3 constant matrix, resulting in a 1-by-2 umat.
M1 = [2 3 1]*M
UMAT: 1 Rows, 2 Columns
a: real, nominal = 3, variability = [-0.5 2], 1 occurrence
b: real, nominal = 10, variability = [-20 20]%, 2 occurrences
c: real, nominal = 3, variability = [-40 40]%, 1 occurrence

Verify that the 1st entry of M1 is -2*a + 3*b + 1. Direct subtraction yields a
umat without any dependence on uncertain elements. Simplifying the class
shows that the result is zero as expected.
d = M1(1) - (-2*M.Uncertainty.a + 3*M.Uncertainty.b + 1)
UMAT: 1 Rows, 1 Columns
simplify(d,'class')
ans =
0

Transpose M, form a product, an inverse, and sample the uncertain result. As

expected, the result is the 2-by-2 identity matrix.
H = M.'*M;
K = inv(H);
usample(K*H,3)

6-22

Uncertain Matrices

ans(:,:,1) =
1.0000
-0.0000
-0.0000
1.0000
ans(:,:,2) =
1.0000
-0.0000
-0.0000
1.0000
ans(:,:,3) =
1.0000
-0.0000
-0.0000
1.0000

Substituting for Uncertain Atoms

Uncertain atoms can be substituted for using usubs. This is described in more
detail in the section Substitution by usubs. Here, we illustrate a few special
cases.
Substitute all instances of the uncertain real parameter named a with the
number 4. This results in a umat, with dependence on the uncertain real
parameters b and c.
M2 = usubs(M,'a',4)
UMAT: 3 Rows, 2 Columns
b: real, nominal = 10, variability = [-20 20]%, 2 occurrences
c: real, nominal = 3, variability = [-40 40]%, 1 occurrence

Similarly, we can substitute all instances of the uncertain real parameter

named b with M.Uncertainty.a, resulting in a umat with dependence on the
uncertain real parameters a and c.
M3 = usubs(M,'b', M.Uncertainty.a)
UMAT: 3 Rows, 2 Columns
a: real, nominal = 3, variability = [-0.5 2], 4 occurrences
c: real, nominal = 3, variability = [-40 40]%, 1 occurrence
Nominal and/or random instances can easily be specified.
M4 = usubs(M,'a','nominal','b','random')
UMAT: 3 Rows, 2 Columns
c: real, nominal = 3, variability = [-40 40]%, 1 occurrence

If one value is being substituted for many different atoms, the atom names
can be listed in a cell array, and then the value given. For example, substitute
a and b with the number 4, and c with the number 5.
M5 = usubs(M,{'a' 'b'},4,'c',5)

6-23

Building Uncertain Models

M5 =
-4.0000
4.0000
1.0000

0.2500
4.2500
5.0000

The command usample also generates multiple random instances of a umat

(and uss and ufrd). See Sampling Uncertain Objects on page 6-42 for
details.

Lifting a double matrix to a umat

A not-uncertain matrix may be interpreted as an uncertain matrix that has
no dependence on uncertain atoms. Use the umat command to lift a double
to the umat class.
Md = [1 2 3;4 5 6];
M = umat(Md)
UMAT: 2 Rows, 3 Columns

High dimensional double matrices can also be lifted. Note from the display
that once the matrix is interpreted as a umat, the 3rd dimension and beyond
are interpreted as array dimensions. See Array Management for Uncertain
Objects on page 6-49 for more information about how multidimensional
arrays of uncertain objects are handled.
Md = randn(4,5,6);
M = umat(Md)
UMAT: 4 Rows, 5 Columns [array, 6 x 1]
Md = randn(4,5,6,7);
M = umat(Md)
UMAT: 4 Rows, 5 Columns [array, 6 x 7]
Md = randn(4,5,6,7,8);
M = umat(Md)
UMAT: 4 Rows, 5 Columns [array, 6 x 7 x 8]

6-24

Uncertain State-Space Systems (uss)

Uncertain State-Space Systems (uss)

Uncertain systems (uss) are linear systems with uncertain state-space
matrices and/or uncertain linear dynamics. Like their certain (i.e., not
uncertain) counterpart, the ss object, they are often built from state-space
matrices using the ss command. In the case where some of the state-space
matrices are uncertain, the result will be a uncertain state-space (uss) object.
Combining uncertain systems with uncertain systems (with the feedback
command, for example) usually leads to an uncertain system. Not-uncertain
systems can be combined with uncertain systems. Usually the result is an
uncertain system.
The nominal value of an uncertain system is a ss object, which is familiar to
Control System Toolbox users.

Creating Uncertain Systems

Uncertain systems (class uss) are built from of certain and/or uncertain
state-space matrices, usually using the ss command.
In the example below, the A, B and C matrices are made up of uncertain real
parameters. Packing them together with the ss command results in a
continuous-time uncertain system.
You can create three uncertain real parameters. Then create 3 uncertain
matrices A, B and C, and one double matrix D.
p1 = ureal('p1',10,'pe',50);
p2 = ureal('p2',3,'plusm',[-.5 1.2]);
p3 = ureal('p3',0);
A = [-p1 p2;0 -p1];
B = [-p2;p2+p3];
C = [1 0;1 1-p3];
D = [0;0];

Pack the 4 matrices together using the ss command. This results in a

continuous-time 2-output, 1-input, 2-state uncertain system.
sys = ss(A,B,C,D)
USS: 2 States, 2 Outputs, 1 Input, Continuous System
p1: real, nominal = 10, variability = [-50 50]%, 2 occurrences
p2: real, nominal = 3, variability = [-0.5 1.2], 2 occurrences

6-25

Building Uncertain Models

p3: real, nominal = 0, variability = [-1

1], 2 occurrences

Properties of uss Objects

View the properties with the get command.
get(sys)
a:
b:
c:
d:
StateName:
Ts:
InputName:
OutputName:
InputGroup:
OutputGroup:
NominalValue:
Uncertainty:
Notes:
UserData:

[2x2
[2x1
[2x2
[2x1
{2x1
0
{''}
{2x1
[1x1
[1x1
[2x1
[1x1
{}
[]

umat]
umat]
umat]
double]
cell}

cell}
struct]
struct]
ss]
atomlist]

The properties a, b, c, d, and StateName behave in exactly the same

manner as Control System Toolbox ss objects. The properties InputName,
OutputName, InputGroup and OutputGroup behave in exactly the same
manner as all of the Control System Toolbox system objects (ss, zpk, tf,
and frd).
The NominalValue is a Control System Toolbox ss object, and hence all
methods for ss objects are available. For instance, compute the poles and step
response of the nominal system.
pole(sys.NominalValue)
ans =
-10
-10
step(sys.NominalValue)

6-26

Uncertain State-Space Systems (uss)

Step Response

To: Out(1)

0.05

0.1

0.15

Amplitude

0.2

0.25
0.1
0.09
0.08

To: Out(2)

0.07
0.06
0.05
0.04
0.03
0.02
0.01
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Time (sec)

Just as with the umat class, the Uncertainty property is a atomlist object,
acting as a gateway to the uncertain atoms. Direct access to the atoms is
facilitated through Uncertainty. Check the Range of the uncertain element
named 'p2' within sys, then change its left endpoint.
sys.Uncertainty.p2.range
ans =
2.5000
4.2000
sys.Uncertainty.p2.range(1) = 2;

Sampling Uncertain Systems

The command usample randomly samples the uncertain system at a specified
number of points.
Randomly sample the uncertain system at 20 points in its modeled
uncertainty range. This gives a 20-by-1 ss array. Consequently, all analysis
tools from the Control System Toolbox are available.
manysys = usample(sys,20);
size(manysys)
20x1 array of state-space models
Each model has 2 outputs, 1 input, and 2 states.
step(manysys)

6-27

Building Uncertain Models

Step Response

0
0.05

To: Out(1)

0.1
0.15
0.2
0.25
0.3

Amplitude

0.35
0.4
0.8
0.7
0.6

To: Out(2)

0.5
0.4
0.3
0.2
0.1
0
0.1
0.2

0.5

1.5

2.5

Time (sec)

The command step can be called directly on a uss object. The default
behavior samples the uss object at 20 instances, and plots the step responses
of these 20 models, as well as the nominal value.
The same features are available for bode, bodemag, impulse, nyquist and
step.

Feedback around Uncertain Plant

It is possible to form interconnections of uss objects. A common example is to
form the feedback interconnection of a given controller with an uncertain
plant.
First create the uncertain plant. Start with two uncertain real parameters.
gamma = ureal('gamma',4);
tau = ureal('tau',.5,'Percentage',30);

Next, create an unmodeled dynamics atom, delta, and a 1st order weighting
function, whose DC value is 0.2, high-frequency gain is 10, and whose
crossover frequency is 8 rad/sec.
delta = ultidyn('delta',[1 1],'SampleStateDim',5);
W = makeweight(0.2,6,6);

6-28

Uncertain State-Space Systems (uss)

Finally, create the uncertain plant consisting of the uncertain parameters

and the unmodeled dynamics.
P = tf(gamma,[tau 1])*(1+W*delta);

You can create an integral controller based on nominal plant parameters.

Nominally the closed-loop system will have damping ratio of 0.707 and time
constant of 2*tau.
KI = 1/(2*tau.Nominal*gamma.Nominal);
C = tf(KI,[1 0]);

Create the uncertain closed-loop system using the feedback command.

CLP = feedback(P*C,1);

Using usample and step, plot samples of the open-loop and closed-loop step
responses. As expected the integral controller reduces the variability in the
low frequency response.
subplot(2,1,1); step(P,5,20)
subplot(2,1,2); step(CLP,5,20)

Step Response

6
5

Amplitude

4
3
2
1
0
1
2

0.5

1.5

2.5
Time (sec)

3.5

4.5

3.5

4.5

Step Response

1.5

Amplitude

0.5

0.5

0.5

1.5

2.5
Time (sec)

6-29

Building Uncertain Models

Interpreting Uncertainty in Discrete Time

The interpretation of a ultidyn atom as a continuous-time or discrete-time
system depends on the nature of the uncertain system (uss) within which it
is an uncertain element.
For example, create a scalar ultidyn object. Then, create two 1-input,
1-output uss objects using the ultidyn object as their D matrix. In one case,
create without specifying sample-time, which indicates continuous time. In
the second case, force discrete-time, with a sample time of 0.42.
delta = ultidyn('delta',[1 1]);
sys1 = uss([],[],[],delta)
USS: 0 States, 1 Output, 1 Input, Continuous System
delta: 1x1 LTI, max. gain = 1, 1 occurrence
sys2 = uss([],[],[],delta,0.42)
USS: 0 States, 1 Output, 1 Input, Discrete System, Ts = 0.42
delta: 1x1 LTI, max. gain = 1, 1 occurrence

Next, get a random sample of each system. When obtaining random samples
using usample, the values of the atoms used in the sample are returned in the
2nd argument from usample as a structure.
[sys1s,d1v] = usample(sys1);
[sys2s,d2v] = usample(sys2);

Look at d1v.delta.Ts and d2v.delta.Ts. In the first case, since sys1 is

continuous-time, the system d1v.delta is continuous-time. In the second
case, since sys2 is discrete-time, with sample time 0.42, the system
d2v.delta is discrete-time, with sample time 0.42.
d1v.delta.Ts
ans =
0
d2v.delta.Ts
ans =
0.4200

Finally, in the case of a discrete-time uss object, it is not the case that
ultidyn objects are interpreted as continuous-time uncertainty in feedback
with sampled-data systems. This very interesting hybrid theory has been
studied by many authors, see [DullerudGlover], but it is beyond the scope of
the toolbox.

6-30

Uncertain State-Space Systems (uss)

Lifting a ss to a uss
A not-uncertain state space object may be interpreted as an uncertain state
space object that has no dependence on uncertain atoms. Use the uss
command to lift a ss to the uss class.
sys = rss(3,2,1);
usys = uss(sys)
USS: 3 States, 2 Outputs, 1 Input, Continuous System

Arrays of ss objects can also be lifted. See Array Management for Uncertain
Objects on page 6-49 for more information about how arrays of uncertain
objects are handled.

Handling Delays in uss

In the current implementation, delays are not allowed. Delays are omitted
and a warning is displayed when ss objects are lifted to uss objects.
sys = rss(3,2,1);
sys.inputdelay = 1.3;
usys = uss(sys)
Warning: Omitting DELAYs in conversion to USS
> In uss.uss at 103
USS: 3 States, 2 Outputs, 1 Input, Continuous System

This lifting process happens in the background whenever ss objects are

combined with any uncertain object. Consequently all delays will be lost in
such operations.
Use the command pade to approximately preserve the effect of the time delay
in the ss object. Before operations involving ss objects containing delays and
uncertain objects, use the pade command to convert the ss object to a delay
free object.
For example, consider an uncertain system with a time constant
approximately equal to 1, an extra input delay of 0.3 seconds, second-order
rolloff beyond 20 rad/s, and an uncertain steady-state gain ranging from 4 to
6. This can be approximated using the pade command as follows.
sys = tf(1,[1 1])*tf(1,[0.05 1]);
sys.inputdelay = 0.3;
gain = ureal('gain',5);
usys = gain*pade(sys,4)

6-31

Building Uncertain Models

USS: 6 States, 1 Output, 1 Input, Continuous System

gain: real, nominal = 5, variability = [-1 1], 1 occurrence

If gain is multiplied by sys directly, the time delay is unfortunately omitted,

since this operation involves lifting sys to a uss as described above. The
difference is obvious from the step responses.
step(usys,gain*sys,4,5)
Warning: Omitting DELAYs in conversion to USS
> In uss.uss at 103
In umat.umat at 98
In atom.mtimes at 7

Step Response

Amplitude

6-32

0.5

1.5

2
Time (sec)

2.5

3.5

Uncertain frd

Uncertain frd
Uncertain frequency responses (ufrd) arise naturally when computing the
frequency response of an uncertain state-space (uss). They also arise when
frequency response data (in an frd object) is combined (added, multiplied,
concatenated, etc.) to an uncertain matrix (umat).

Creating Uncertain Frequency Response Objects

The most common manner in which a ufrd arises is taking the frequency
response of a uss. The natural command that would do this is frd (an
overloaded version in the @uss directory).
Reconstruct sys, if necessary.
p1 = ureal('p1',10,'pe',50);
p2 = ureal('p2',3,'plusm',[-.5 1.2]);
p3 = ureal('p3',0);
A = [-p1 p2;0 -p1];
B = [-p2;p2+p3];
C = [1 0;1 1-p3];
D = [0;0];
sys = ss(A,B,C,D)
USS: 2 States, 2 Outputs, 1 Input, Continuous System
p1: real, nominal = 10, variability = [-50 50]%, 2 occurrences
p2: real, nominal = 3, variability = [-0.5 1.2], 2 occurrences
p3: real, nominal = 0, variability = [-1 1], 2 occurrences

Compute the uncertain frequency response of the uncertain system. Use the
frd command, along with a frequency grid containing 100 points. The result
is an uncertain frequency response data object, referred to as a ufrd.
sysg = frd(sys,logspace(-2,2,100))
UFRD: 2 Outputs, 1 Input, Continuous System, 100 Frequency
points
p1: real, nominal = 10, variability = [-50 50]%, 2 occurrences
p2: real, nominal = 3, variability = [-0.5 1.2], 2 occurrences
p3: real, nominal = 0, variability = [-1 1], 2 occurrences

Properties of ufrd objects

View the properties with the get command.

6-33

Building Uncertain Models

get(sysg)
Frequency:
ResponseData:
Units:
Ts:
InputName:
OutputName:
InputGroup:
OutputGroup:
NominalValue:
Uncertainty:
Notes:
UserData:
Version:

[100x1 double]
[2x1x100 umat]
'rad/s'
0
{''}
{2x1 cell}
[1x1 struct]
[1x1 struct]
[2x1 frd]
[1x1 atomlist]
{}
[]
4

The properties ResponseData and Frequency behave in exactly the same

manner as Control System Toolbox frd objects, except that ResponseData is
a umat. The properties InputName, OutputName, InputGroup and
OutputGroup behave in exactly the same manner as all of the Control System
Toolbox system objects (ss, zpk, tf, and frd).
The NominalValue is a Control System Toolbox frd object, and hence all
methods for frd objects are available. For instance, plot the Bode response of
the nominal system.
bode(sysg.nom)

6-34

Uncertain frd

Bode Diagram

To: Out(1)

10
20
30

Magnitude (dB) ; Phase (deg)

To: Out(2)
To: Out(1)

40
180

90
0
20
40
60
80

To: Out(2)

100
0

90

180
3
10

10

10

10

10

10

Frequency (Hz)

Just as with the umat and uss classes, the Uncertainty property is an
atomlist object, acting as a gateway to the uncertain atoms. Direct access to
the atoms is facilitated through Uncertainty. Change the nominal value of
the uncertain element named 'p1' within sysg to 14, and replot the Bode
plot of the (new) nominal system.
sysg.unc.p1.nom = 14
UFRD: 2 Outputs, 1 Input, Continuous System, 100 Frequency
points
p1: real, nominal = 14, variability = [-50 50]%, 2 occurrences
p2: real, nominal = 3, variability = [-0.5 1.2], 2 occurrences
p3: real, nominal = 0, variability = [-1 1], 2 occurrences

6-35

Building Uncertain Models

Interpreting Uncertainty in Discrete Time

See Interpreting Uncertainty in Discrete Time on page 6-30. The issues are
identical.

Lifting an frd to a ufrd

A not-uncertain frequency response object may be interpreted as an
uncertain frequency response object that has no dependence on uncertain
atoms. Use the ufrd command to lift an frd object to the ufrd class.
sys = rss(3,2,1);
sysg = frd(sys,logspace(-2,2,100));
usysg = ufrd(sysg)
UFRD: 2 Outputs, 1 Input, Continuous System, 100 Frequency
points

Arrays of frd objects can also be lifted. See Array Management for Uncertain
Objects on page 6-49 for more information about how arrays of uncertain
objects are handled.

Handling Delays in ufrd

See Handling Delays in uss on page 6-31. The issues are identical.

6-36

Basic Control System Toolbox and MATLAB Interconnections

Basic Control System Toolbox and MATLAB

Interconnections
This list has all of the basic system interconnection functions defined in the
Control System Toolbox or in MATLAB.
append
blkdiag
series
parallel
feedback
lft
star
These functions work with uncertain objects as well. Uncertain objects may
be combined with certain objects, resulting in an uncertain object.

6-37

Building Uncertain Models

Simplifying Representation of Uncertain Objects

A minimal realization of the transfer function matrix
2
---------s+1
H(s) =
3
-----------s+1

4
---------s+1
6
-----------s+1

has only 1 state, obvious from the decomposition

1
H ( s ) = 2 ------------ 1 2
3 s+1
However, a natural construction, formed by
sys11
sys12
sys21
sys22
sys =
a =
x1
x2
x3
x4
b =
x1
x2
x3
x4
c =
y1
y2
d =

6-38

= ss(tf(2,[1 1]));
= ss(tf(4,[1 1]));
= ss(tf(3,[1 1]));
= ss(tf(6,[1 1]));
[sys11 sys12;sys21 sys22]
x1
-1
0
0
0

x2
0
-1
0
0

u1
2
0
2
0

u2
0
2
0
2

x1
1
0

x2
2
0

x3
0
0
-1
0

x4
0
0
0
-1

x3
0
1.5

x4
0
3

Simplifying Representation of Uncertain Objects

u1 u2
y1
0
0
y2
0
0
Continuous-time model

has 4 states, and is nonminimal.

In the same manner, the internal representation of uncertain objects built up
from uncertain atoms can become nonminimal, depending on the sequence of
operations in their construction. The command simplify employs ad-hoc
simplification and reduction schemes to reduce the complexity of the
representation of uncertain objects. There are three levels of simplification:
off, basic and full. Each uncertain atom has an AutoSimplify property
whose value is one of the strings 'off', 'basic' or 'full'. The default
value is 'basic'.
After (nearly) every operation, the command simplify is automatically run
on the uncertain object, cycling through all of the uncertain atoms, and
attempting to simplify (without error) the representation of the effect of that
uncertain object. The AutoSimplify property of each atom dictates the types
of computations that are performed. In the 'off' case, no simplification is
even attempted. In 'basic', fairly simple schemes to detect and eliminate
nonminimal representations are used. Finally, in 'full', numerical based
methods similar to truncated balanced realizations are used, with a very
tight tolerance to minimize error.

Effect of Autosimplify Property

Create an uncertain real parameter, view the AutoSimplify property of a,
and then create a 1-by-2 umat, both of whose entries involve the uncertain
parameter.
a = ureal('a',4);
a.AutoSimplify
ans =
basic
m1 = [a+4 6*a]
UMAT: 1 Rows, 2 Columns
a: real, nominal = 4, variability = [-1

1], 1 occurrence

6-39

Building Uncertain Models

Note that although the uncertain real parameter a appears in both (two)
entries of the matrix, the resulting uncertain matrix m1 only depends on 1
occurrence of a.
Set the AutoSimplify property of a to 'off' (from 'basic'). Recreate the
1-by-2 umat. Now note that the resulting uncertain matrix m2 depends on 2
occurrences of a.
a.AutoSimplify = 'off';
m2 = [a+4 6*a]
UMAT: 1 Rows, 2 Columns
a: real, nominal = 4, variability = [-1

1], 2 occurrences

The 'basic' level of autosimplification often detects (and simplifies)

duplication created by linear terms in the various entries. Higher order
(quadratic, bilinear, etc.) duplication is often not detected by the 'basic'
autosimplify level.
For example, reset the AutoSimplify property of a to 'basic' (from 'off').
Create an uncertain real parameter, and a 1-by-2 umat, both of whose entries
involve the square of the uncertain parameter.
a.AutoSimplify = 'basic';
m3 = [a*(a+4) 6*a*a]
UMAT: 1 Rows, 2 Columns
a: real, nominal = 4, variability = [-1

1], 4 occurrences

Note that the resulting uncertain matrix m3 depends on 4 occurrences of a.

Set the AutoSimplify property of a to 'full' (from 'basic'). Recreate the
1-by-2 umat. Now note that the resulting uncertain matrix m4 depends on 2
occurrences of a.
a.AutoSimplify = 'full';
m4 = [a*(a+4) 6*a*a]
UMAT: 1 Rows, 2 Columns
a: real, nominal = 4, variability = [-1

1], 2 occurrences

Although m4 has a less complex representation (2 occurrences of a rather than

4 as in m3), some numerical variations are seen when both uncertain objects
are evaluated at (say) 0.
usubs(m3,'a',0)
ans =

6-40

Simplifying Representation of Uncertain Objects

0
0
usubs(m4,'a',0)
ans =
1.0e-015 *
-0.4441

Small numerical differences are also noted at other evaluation points. The
example below shows the differences encountered evaluating at a equal to 1.
usubs(m3,'a',1)
ans =
5
6
usubs(m4,'a',1)
ans =
5.0000
6.0000

Direct Use of simplify

The simplify command can be used to override all uncertain element's
AutoSimplify property. The first input to the simplify command is an
uncertain object. The second input is the desired reduction technique, which
can either 'basic' or 'full'.
Again create an uncertain real parameter, and a 1-by-2 umat, both of whose
entries involve the square of the uncertain parameter. Set the AutoSimplify
property of a to 'basic'.
a.AutoSimplify = 'basic';
m3 = [a*(a+4) 6*a*a]
UMAT: 1 Rows, 2 Columns
a: real, nominal = 4, variability = [-1

1], 4 occurrences

Note that the resulting uncertain matrix m3 depends on 4 occurrences of a.

The simplify command can be used to perform a 'full' reduction on the
resulting umat.
m4 = simplify(m3,'full')
UMAT: 1 Rows, 2 Columns
a: real, nominal = 4, variability = [-1

1], 2 occurrences

The resulting uncertain matrix m4 depends on only 2 occurrences of a after

the reduction.

6-41

Building Uncertain Models

Sampling Uncertain Objects

The command usample is used to randomly sample an uncertain object,
giving a not-uncertain instance of the uncertain object.

Generating One Sample

If A is an uncertain object, then usample(A) generates a single sample of A.
For example, a sample of a ureal is a scalar double.
A = ureal('A',6);
B = usample(A)
B =
5.7298

Create a 1-by-3 umat with A and an uncertain complex parameter C. A single

sample of this umat is a 1-by-3 double.
C = ucomplex('C',2+6j);
M = [A C A*A];
usample(M)
ans =
5.9785
1.4375 + 6.0290i

35.7428

Generating Many Samples

If A is an uncertain object, then usample(A,N) generates N samples of A.
For example, 20 samples of a ureal gives a 1-by-1-20 double array.
B = usample(A,20);
size(B)
ans =
1
1
20

Similarly, 30 samples of the 1-by-3 umat M yields a 1-by-3-by-30 array.

size(usample(M,30))
ans =
1
3
30

See Creating Arrays with usample on page 6-53 for more information on
sampling uncertain objects.

6-42

Sampling Uncertain Objects

Sampling ultidyn Atoms

When sampling a ultidyn atom (or an uncertain object that contains a
ultidyn atom in its Uncertainty gateway) the result is always a state-space
(ss) object. The property SampleStateDim of the ultidyn class determines the
state dimension of the samples.
Create a 1-by-1, gain bounded ultidyn object, with gain-bound 3. Verify that
the default state dimension for samples is 1.
del = ultidyn('del',[1 1],'Bound',3);
del.SampleStateDim
ans =
1

Sample the uncertain atom at 30 points. Verify that this creates a 30-by-1 ss
array of 1-input, 1-output, 1-state systems.
delS = usample(del,30);
size(delS)
30x1 array of state-space models
Each model has 1 output, 1 input, and 1 state.

Plot the Nyquist plot of these samples and add a disk of radius 3. Note that
the gain bound is satisfied and that the Nyquist plots are all circles,
indicative of 1st order systems.
nyquist(delS)
hold on;
theta = linspace(-pi,pi);
plot(del.Bound*exp(sqrt(-1)*theta),'r');
hold off;

6-43

Building Uncertain Models

Nyquist Diagram

1.5

Imaginary Axis

0.5

0.5

1.5

2
3

0
Real Axis

Change the SampleStateDim to 4, and repeat entire procedure. The Nyquist

plots satisfy the gain bound and as expected are more complex than the
circles found in the 1st-order sampling.
del.SampleStateDim = 4;
delS = usample(del,30);
nyquist(delS)
hold on;
theta = linspace(-pi,pi);
plot(del.Bound*exp(sqrt(-1)*theta),'r');
hold off;

6-44

Sampling Uncertain Objects

Nyquist Diagram

Imaginary Axis

3
3

0
Real Axis

6-45

Building Uncertain Models

Substitution by usubs
If an uncertain object (umat, uss, ufrd) has many uncertain parameters, it
is often necessary to freeze some, but not all, of the uncertain parameters to
specific values. The usubs command accomplishes this, and also allows more
complicated substitutions for an atom.
usubs accepts a list of atom names, and respective values to substitute for
them. You can create 3 uncertain real parameters and use them to create a
2-by-2 uncertain matrix A.
delta = ureal('delta',2);
eta = ureal('eta',6);
rho = ureal('rho',-1);
A = [3+delta+eta delta/eta;7+rho rho+delta*eta]
UMAT: 2 Rows, 2 Columns
delta: real, nominal = 2, variability = [-1 1], 2 occurrences
eta: real, nominal = 6, variability = [-1 1], 3 occurrences
rho: real, nominal = -1, variability = [-1 1], 1 occurrence

Use usubs to substitute the uncertain element named delta in A with the
value 2.3, leaving all other uncertain atoms intact. Note that the result, B, is
an uncertain matrix with dependence only on eta and rho.
B = usubs(A,'delta',2.3)
UMAT: 2 Rows, 2 Columns
eta: real, nominal = 6, variability = [-1 1], 3 occurrences
rho: real, nominal = -1, variability = [-1 1], 1 occurrence

To set multiple atoms, list individually, or in cells. The following are the same
B1 = usubs(A,'delta',2.3,'eta',A.Uncertainty.rho);
B2 = usubs(A,{'delta';'eta'},{2.3;A.Uncertainty.rho});

In each case, delta is replaced by 2.3, and eta is replaced by

A.Uncertainty.rho.

If it makes sense, a single replacement value can be used to replace multiple

atoms. So
B3 = usubs(A,{'delta';'eta'},2.3);

replaces both the atoms delta and eta with the real number 2.3. Any
superfluous substitution requests are ignored. Hence

6-46

Substitution by usubs

B4 = usubs(A,'fred',5);

is the same as A, and

B5 = usubs(A,{'delta';'eta'},2.3,{'fred' 'gamma'},0);

is the same as B3.

Specifying the Substitution with Structures

An alternative syntax for usubs is to specify the substituted values in a
structure, whose fieldnames are the names of the atoms being substituted
with values.
Create a structure NV with 2 fields, delta and eta. Set the values of these
fields to be the desired substituted values. Then perform the substitution
with usubs.
NV.delta = 2.3;
NV.eta = A.Uncertainty.rho;
B6 = usubs(A,NV);

Here, B6 is the same as B1 and B2 above. Again, any superfluous fields are
ignored. Therefore, adding an additional field gamma to NV, and substituting
does not alter the result.
NV.gamma = 0;
B7 = usubs(A,NV);

Here, B7 is the same as B6.

The commands wcgain, robuststab and usample all return substitutable
values in this structure format. More discussion can be found in Creating
Arrays with usubs on page 6-55.

Nominal and Random Values

If the replacement value is the (partial and case-independent) string
'Nominal', then the listed atom are replaced with their nominal values.
Therefore
B8 = usubs(A,fieldnames(A.Uncertainty),'nom')
B8 =
11.0000
0.3333

6-47

Building Uncertain Models

6.0000
11.0000
B9 = A.NominalValue
B9 =
11.0000
0.3333
6.0000
11.0000

are the same. It is possible to only set some of the atoms to NominalValues,
and would be the typical use of usubs with the 'nominal' argument.
Within A, set eta to its nominal value, delta to a random value (within its
range) and rho to a specific value, say 6.5
B10 = usubs(A,'eta','nom','delta','rand','rho',6.5)
B10 =
10.5183
0.2531
13.5000
15.6100

Unfortunately, the 'Nominal' and 'Random' specifiers may not be used in

the structure format. However, explicitly setting a field of the structure to an
atom's nominal value, and then following (or preceeding) the call to usubs
with a call to usample (to generate the random samples) is acceptable, and
achieves the same effect.

6-48

Array Management for Uncertain Objects

Array Management for Uncertain Objects

All of the uncertain system classes (uss, ufrd) may be multidimensional
arrays. This is intended to provide the same functionality as the LTI-arrays
of Control System Toolbox. The command size returns a row vector with the
sizes of all dimensions.
The first two dimensions correspond to the outputs and inputs of the system.
Any dimensions beyond are referred to as the array dimensions. Hence, if szM
= size(M), then szM(3:end) are sizes of the array dimensions of M.
For these types of objects, it is clear that the first two dimensions (system
output and input) are interpreted differently from the 3rd, 4th, 5th and
higher dimensions (which often model parametrized variability in the system
input/output behavior).
umat objects are treated in the same manner. The first two dimensions are the
rows and columns of the uncertain matrix. Any dimensions beyond are the
array dimensions.

Referencing Arrays
Suppose M is a umat, uss or ufrd, and that Yidx and Uidx are vectors of
integers. Then
M(Yidx,Uidx)

selects the outputs (rows) referred to by Yidx and the inputs (columns)
referred to by Uidx, preserving all of the array dimensions. For example, if
size(M) equals [4 5 3 6 7], then (for example) the size of M([4 2],[1 2
4]) is [2 3 3 6 7].
If size(M,1)==1 or size(M,2)==1, then single indexing on the inputs or
outputs (rows or columns) is allowed. If Sidx is a vector of integers, then
M(Sidx) selects the corresponding elements. All array dimensions are
preserved.
If there are K array dimensions, and idx1, idx2, ..., idxK are vectors of
integers, then
G = M(Yidx,Uidx,idx1,idx2,...,idxK)

selects the outputs and inputs referred to by Yidx and Uidx, respectively, and
selects from each array dimension the slices referred to by the idx1,

6-49

Building Uncertain Models

idx2,..., idxK index vectors. Consequently, size(G,1) equals

length(Yidx), size(G,2) equals length(Uidx), size(G,3) equals
length(idx1), size(G,4) equals length(idx2), and size(G,K+2) equals
length(idxK).

If M has K array dimensions, and less than K index vectors are used in doing
the array referencing, then the MATLAB convention for single indexing is
followed. For instance, suppose size(M) equals [3 4 6 5 7 4]. The
expression
G = M([1 3],[1 4],[2 3 4],[5 3 1],[8 10 12 2 4 20 18])

is valid. The result has size(G) equals [2 2 3 3 7]. The last index vector
[8 10 12 2 4 20 18] is used to reference into the 7-by-4 array, preserving
the order dictated by MATLAB single indexing (e.g., the 10th element of a
7-by-4 array is the element in the (3,2) position in the array).
Note that if M has either one output (row) or one input (column), and M has
array dimensions, then it is not allowable to combine single indexing in the
output/input dimensions along with indexing in the array dimensions. This
will result in an ambiguity in how to interpret the second index vector in the
expression (i.e., does it correspond to the input/output reference, or does it
correspond to the first array dimension?).

Creating Arrays with stack and cat Functions

An easy manner to create an array is with stack. Create a [4-by-1] umat array
by stacking four 1-by-3 umat objects with the stack command. The first
argument of stack specifies in which array dimension the stacking occurs. In
the example below, the stacking is done is the 1st array dimension, hence the
result is a 1-by-3-by-4-by-1 umat, referred to as a 4-by-1 umat array.
a = ureal('a',4);
b = ureal('b',2);
M = stack(1,[a b 1],[-a
UMAT: 1 Rows, 3 Columns
a: real, nominal = 4,
b: real, nominal = 2,
size(M)
ans =
1
3
4
arraysize(M)

6-50

-b 4+a],[4 5 6],[a 0 0])

[array, 4 x 1]
variability = [-1 1], 1 occurrence
variability = [-1 1], 1 occurrence

Array Management for Uncertain Objects

ans =
4

Check that result is valid. Use referencing to access parts of the [4-by-1] umat
array and compare to the expected values. The first 4 examples should all be
arrays full of 0 (zeros). The last two should be the value 5, and the uncertain
real parameter a, respectively.
simplify(M(:,:,1) - [a b 1])
ans =
0
0
0
simplify(M(:,:,2) - [-a -b 4+a])
ans =
0
0
0
simplify(M(:,:,3) - [4 5 6])
ans =
0
0
0
simplify(M(:,:,4) - [a 0 0])
ans =
0
0
0
simplify(M(1,2,3)) % should be 5
ans =
5
simplify(M(1,3,2)-4)
Uncertain Real Parameter: Name a, NominalValue 4, variability =
[-1 1]

You can create a random 1-by-3-by-4 double matrix and stack this with M
along the second array dimension, creating a 1-by-3-by-4-by-2 umat.
N = randn(1,3,4);
M2 = stack(2,M,N);
size(M2)
ans =
1
3
4
arraysize(M2)
ans =
4
2

As expected, both M and N can be recovered from M2.

d1 = simplify(M2(:,:,:,1)-M);

6-51

Building Uncertain Models

d2 = simplify(M2(:,:,:,2)-N);
[max(abs(d1(:))) max(abs(d2(:)))]
ans =
0
0

It is also possible to stack M and N along the 1st array dimension, creating a
1-by-3-by-8-by-1 umat.
M3 = stack(1,M,N);
size(M3)
ans =
1
3
8
arraysize(M3)
ans =
8
1

As expected, both M and N can be recovered from M3.

d3 = simplify(M3(:,:,1:4)-M);
d4 = simplify(M3(:,:,5:8)-N);
[max(abs(d3(:))) max(abs(d4(:)))]
ans =
0
0

Creating Arrays by Assignment

Arrays can be created by direct assignment. As with other MATLAB classes,
there is no need to preallocate the variable first. Simply assign elements all
resizing is performed automatically.
For instance, an equivalent construction to
a = ureal('a',4);
b = ureal('b',2);
M = stack(1,[a b 1],[-a -b 4+a],[4 5 6],[a 0 0]);
is
Mequiv(1,1,1) = a;
Mequiv(1,2,1) = b;
Mequiv(1,3,1) = 1;
Mequiv(1,:,4) = [a 0 0];
Mequiv(1,:,2:3) = stack(1,[-a -b 4+a],[4 5 6]);

6-52

Array Management for Uncertain Objects

The easiest manner for you to verify that the results are the same is to
subtract and simplify,
d5 = simplify(M-Mequiv);
max(abs(d5(:)))
ans =
0

Binary Operations with Arrays

Most operations simply cycle through the array dimensions, doing pointwise
operations. Assume A and B are umat (or uss, or ufrd) arrays with identical
array dimensions (slot 3 and beyond). The operation C = fcn(A,B) is
equivalent to looping on k1, k2, ..., setting
C(:,:,k1,k2,...) = fcn(A(:,:,k1,k2,...),B(:,:,k1,k2,...))

The result C has the same array dimensions as A and B. The user is required
to manage the extra dimensions (i.e., keep track of what they mean).
Methods such as permute, squeeze and reshape are included to facilitate
this management.
In general, any binary operation requires that the extra-dimensions are
compatible. The umat, uss and ufrd objects allow for slightly more flexible
interpretation of this. For illustrative purposes, consider a binary operation
involving variables A and B. Suppose the array dimensions of A are
n1 nl

and that the array dimensions of B are m 1 m l . By

B
MATLAB convention, the infinite number of singleton (i.e., 1) trailing
dimensions are not listed. The compatibility of the extra dimensions is
determined by the following rule: If lA=lB, then pad the shorter dimension list
with trailing 1's. Now compare the extra dimensions: In the k-th dimension,
it must be that one of 3 conditions hold: nk=mk, or nk=1 or mk=1. In other
words, non-singleton dimensions must exactly match (so that the pointwise
operation can be executed), and singleton dimensions match with anything,
implicitly through a repmat.
A

Creating Arrays with usample

An extremely common manner in which to generate an array is to sample (in
some of the uncertain elements) an uncertain object. Using the ureal
objects a and b from above, create a 2-by-3 umat.

6-53

Building Uncertain Models

M = [a b;b*b a/b;1-b 1+a*b]

UMAT: 3 Rows, 2 Columns
a: real, nominal = 4, variability = [-1
b: real, nominal = 2, variability = [-1
size(M)
ans =
3
2

1], 3 occurrences
1], 6 occurrences

Sample (at 20 random points within its range) the uncertain real parameter
b in the matrix M. This results in a 3-by-2-by-20 umat, with only one uncertain
element, a. The uncertain element b of M has been sampled out, leaving a
new array dimension in its place.
[Ms,bvalues] = usample(M,'b',20);
Ms
UMAT: 3 Rows, 2 Columns [array, 20 x 1]
a: real, nominal = 4, variability = [-1
size(Ms)
ans =
3
2
20

1], 2 occurrences

Continue sampling (at 15 random points within its range) the uncertain real
parameter a in the matrix Ms. This results in a 3-by-2-by-20-by-15 double.
[Mss,avalues] = usample(Ms,'a',15);
size(Mss)
ans =
3
2
20
15
class(Mss)
ans =
double

The above 2-step sequence can be performed in 1 step,

[Mss,values] = usample(M,'b',20,'a',15);
class(Mss)
ans =
double

In this case, values is a 20-by-15 struct array, with 2 fields b and a, whose
values are the values used in the random sampling. It follows that
usubs(M,values) is the same as Mss.

6-54

Array Management for Uncertain Objects

Rather than sampling the each variable (a and b) independently, generating

a 20-by-15 grid in a 2-dimensional space, the two-dimensional space can be
sampled. Sample the 2-dimensional space with 800 points,
[Ms,values] = usample(M,{'a' 'b'},800);
size(Ms)
ans =
3
2
800
size(values)
ans =
800
1

Creating Arrays with usubs

Suppose Values is a struct array, with the following properties: the
dimensions of Value match the array dimensions of M, in the manner
described in Section above; the fieldnames of Values are some (or all) of the
names of the uncertain elements of M; and the dimensions of the contents of
the fields within Values match the sizes of the uncertain elements within M.
Then usubs(M,Values) will substitute the uncertain elements in M with the
contents found in the respective fields of Values.
You can create a 3-by-2 uncertain matrix using 2 uncertain real parameters.
a = ureal('a',4);
b = ureal('b',2);
M = [a b;b*b a/b;1-b 1+a*b];

Create a 5-by-1 struct array with fieldname a. Make its values random
scalars. Create a 1-by-4 struct array with fieldname b.
Avalue = struct('a',num2cell(rand(5,1)));
Bvalue = struct('b',num2cell(rand(1,4)));

Substitute the uncertain real parameter a in M with the values in Avalue,

yielding ma. Similarly substitute the uncertain real parameter b in M with the
values in Avalue, yielding mb.
ma = usubs(M,Avalue)
UMAT: 3 Rows, 2 Columns [array, 5 x 1]
b: real, nominal = 2, variability = [-1
mb = usubs(M,Bvalue)
UMAT: 3 Rows, 2 Columns [array, 1 x 4]

1], 6 occurrences

6-55

Building Uncertain Models

a: real, nominal = 4, variability = [-1

1], 2 occurrences

Continue, substituting the uncertain real parameter b in ma with the values

in Bvalue, yielding mab. Do the analogous operation for mb, yielding mba.
Subtract, and note that the difference is 0, as expected.
mab = usubs(ma,Bvalue);
mba = usubs(mb,Avalue);
thediff = mab-mba;
max(abs(thediff(:)))
ans =
4.4409e-016

Creating Arrays with gridureal

The command gridureal enables uniform sampling of specified uncertain
real parameters within an uncertain object. It is a specialized case of usubs.
gridureal removes a specified uncertain real parameter and adds an array
dimension (to the end of the existing array dimensions). The new array
dimension represents the uniform samples of the uncertain object along the
specified uncertain real parameter's range.

Create a 2-by-2 uncertain matrix with 3 uncertain real parameters.

a = ureal('a',3,'Range',[2.5 4]);
b = ureal('b',4,'Percentage',15);
c = ureal('c',-2,'Plusminus',[-1 .3]);
M = [a b;b c]
UMAT: 2 Rows, 2 Columns
a: real, nominal = 3, range = [2.5 4], 1 occurrence
b: real, nominal = 4, variability = [-15 15]%, 2 occurrences
c: real, nominal = -2, variability = [-1 0.3], 1 occurrence

Grid the uncertain real parameter b in M with 100 points. The result is a umat
array, with dependence on uncertain real parameters a and c.
Mgrid1 = gridureal(M,'b',100)
UMAT: 2 Rows, 2 Columns [array, 100 x 1]
a: real, nominal = 3, range = [2.5 4], 1 occurrence
c: real, nominal = -2, variability = [-1 0.3], 1 occurrence

6-56

Array Management for Uncertain Objects

Operating on the uncertain matrix M, grid the uncertain real parameter a

with 20 points, the uncertain real parameter b with 12 points, and the
uncertain real parameter c with 7 points, The result is a
2-by-2-by20-by-12-by7 double array.
Mgrid3 = gridureal(M,'a',20,'b',12,'c',7);
size(Mgrid3)
ans =
2
2
20
12
7

Creating Arrays with repmat

The MATLAB command repmat is used to replicate and tile arrays. It works
on the built-in objects of MATLAB, namely double, char, as well as the
generalized container objects cell and struct. The identical functionality is
provided for replicating and tiling uncertain elements (ureal, ultidyn, etc.)
and umat objects.
You can create an uncertain real parameter, and replicate it in a 2-by-3
uncertain matrix. Compare to generating the same uncertain matrix through
multiplication.
a = ureal('a',5);
Amat = repmat(a,[2 3])
UMAT: 2 Rows, 3 Columns
a: real, nominal = 5, variability = [-1
Amat2 = a*ones(2,3);
simplify(Amat-Amat2)
ans =
0
0
0
0
0
0

1], 1 occurrence

Create (as in section a [4-by-1] umat array by stacking four 1-by-3 umat
objects with the stack command. Use repmat to tile this 1-by-3-by-4-by-1
umat, into a 2-by-3-by-8-by-5 umat.
a = ureal('a',4);
b = ureal('b',2);
M = stack(1,[a b 1],[-a -b 4+a],[4 5 6],[a 0 0]);
size(M)
ans =
1
3
4

6-57

Building Uncertain Models

Mtiled = repmat(M,[2 1 2 5])

UMAT: 2 Rows, 3 Columns [array, 8 x 5]
a: real, nominal = 4, variability = [-1 1], 1 occurrence
b: real, nominal = 2, variability = [-1 1], 1 occurrence
Verify the equality of M and a few certain tiles of Mtiled.
d1 = simplify(M-Mtiled(2,:,5:8,3));
d2 = simplify(M-Mtiled(1,:,1:4,2));
d3 = simplify(M-Mtiled(2,:,1:4,5));
[max(abs(d1(:))) max(abs(d2(:))) max(abs(d3(:)))]
ans =
0
0
0

Note that repmat never increases the complexity of the representation of an

uncertain object. The number of occurrences of each uncertain element
remains the same, regardless of the extent of the replication and tiling.

Creating Arrays with repsys

Replicating and tiling uncertain state-space systems (uss, and uncertain
frequency response data (ufrd) is done with repsys. The syntax and behavior
are the same as the manner in which repmat is used to replicate and tile
matrices. The syntax and behavior of repsys for uss and ufrd objects are the
same as the traditional repsys which operates on ss and frd object. Just as
in those cases, the uncertain version of repsys also allows for diagonal tiling.

Using permute and ipermute

The commands permute and ipermute are generalizations of transpose,
which exchanges the rows and columns of a two-dimensional matrix.
permute(A,ORDER) rearranges the dimensions of A so that they are in the
order specified by the vector ORDER. The array produced has the same values
of A but the order of the subscripts needed to access any particular element
are rearranged as specified by ORDER. The elements of ORDER must be a
rearrangement of the numbers from 1 to N.

All of the uncertain objects are essentially 2-dimensional (output and input)
operators with array dependence. This means that the first 2 dimensions are
treated differently from dimensions 3 and beyond. It is not permissible to
permute across these groups.

6-58

Array Management for Uncertain Objects

For uss and ufrd, the restriction is built into the syntax. The elements of the
ORDER vector only refer to array dimensions. Therefore, there is no possibility
of permute across these dimensions. In you need to permute the first two
dimensions, use the command transpose instead.
For umat, the restriction is enforced in the software. The elements of the
ORDER vector refer to all dimensions. However, the first two elements of ORDER
must be a rearrangement of the numbers 1 and 2. The remaining elements of
ORDER must be a rearrangement of the numbers 3 through N. If either of those
conditions fail, an error is generated. Hence, for umat arrays, either permute
or transpose can be used to effect the transpose operation.

6-59

Building Uncertain Models

Decomposing Uncertain Objects (for Advanced Users)

Each uncertain object (umat, uss, ufrd) is a generalized feedback
connection (lft) of a not-uncertain object (e.g., double, ss, frd) with a
diagonal augmentation of uncertain atoms (ureal, ultidyn, ucomplex,
ucomplexm, udyn). In robust control jargon, if the uncertain elements are
normalized, this decomposition is often called the M/D form.
The purpose of the uncertain objects (ureal, ultidyn, umat, uss, etc.) is
to hide this underlying decomposition, and allow the user to focus on
modeling and analyzing uncertain systems, rather than the details of
correctly propagating the M/D representation in manipulations.
Nevertheless, advanced users may want access to the familiar M/D form. The
command lftdata accomplishes this decomposition.
Since ureal, ucomplex and ucomplexm do not have their NominalValue
necessarily at zero, and in the case of ureal objects, are not symmetric about
the NominalValue, some details are required in describing the decomposition.

Normalizing Functions for Uncertain Atoms

Associated with each uncertain element is a normalizing function. The
normalizing function maps the uncertain element into a normalized
uncertain element.
If is an uncertain real parameter, with range [L R] and nominal value N,
then the normalizing function F is
A + B
F ( ) = ------------------C + D
with the property that for all satisfying LR, it follows that 1F()1,
moreover, F(L)=1, F(N)=0, and F(R)=1. If the nominal value is centered in
the range, then it is easy to conclude that
R+L
2
A = -------------- , B = -------------- , C=1, D=0
RL
RL
It is left as an algebra exercise for the user to work out the various values for
A, B, C and D when the nominal value is not centered.
If E is an uncertain gain-bounded, linear, time-invariant dynamic
uncertainty, with gain-bound , then the normalizing function F is

6-60

Decomposing Uncertain Objects (for Advanced Users)

1
F ( E ) = --- E

If E is an uncertain positive-real, linear, time-invariant dynamic uncertainty,

with positivity bound , then the normalizing function F is

F ( E ) = I E --- I

1
I + E --- I

where =2||+1.
The normalizing function for an uncertain complex parameter , with
nominal value C and radius is
1
F ( ) = --- ( C )

The normalizing function for uncertain complex matrices H, with nominal

value N and weights WL and WR is
1

F ( H ) = W L ( H N )W R

In each case, as the uncertain atom varies over its range, the absolute value
of the normalizing function (or norm, in the matrix case) varies from 0 and 1.

Properties of the Decomposition

Take an uncertain object A, dependent on uncertain real parameters 1,,N,
uncertain complex parameters 1,,K, uncertain complex matrices
H1,,HB, uncertain gain-bounded linear, time-invariant dynamics E1,,ED,
and uncertain positive-real linear, time-invariant dynamics P1,,PQ.
Write A(,,H,E,P) to indicate this dependence. Using lftdata, A can be
decomposed into two separate pieces: M and (,,H,E,P) with the following
properties:
M is certain (i.e., if A is uss, then M is ss; if A is umat, then M is double; if
A is ufrd, then M is frd).
is always a umat, depending on the same uncertain elements as A, with
ranges, bounds, weights, etc., unaltered.

6-61

Building Uncertain Models

The form of is block diagonal, with elements made up of the normalizing

functions acting on the individual uncertain elements.
F() 0
0
0
0
0 F() 0
0
0
( , , H, E , P ) =
0
0 F(H) 0
0
0
0
0 F(E) 0
0
0
0
0 F(P)
A(,,H,E,P) is given by a linear fractional transformation of M and
(,,H,E,P),
1

A ( , ) = M 22 + M 21 ( , , H, E, P ) [ I M 11 ( , , H, E, P ) ] M 12
The order of the normalized atoms making up A is not the simple order shown
above. It is actually the same order as given by the command
fieldnames(M.Uncertainty). See Advanced Syntax of lftdata on page 6-64
for more information.

Syntax of lftdata
The decomposition is carried out by the command lftdata.
You can create a 2-by-2 umat named A using three uncertain real parameters.
delta = ureal('delta',2);
eta = ureal('eta',6);
rho = ureal('rho',-1);
A = [3+delta+eta delta/eta;7+rho rho+delta*eta]
UMAT: 2 Rows, 2 Columns
delta: real, nominal = 2, variability = [-1 1], 2 occurrences
eta: real, nominal = 6, variability = [-1 1], 3 occurrences
rho: real, nominal = -1, variability = [-1 1], 1 occurrence

Note that A depends on two occurrences of delta, three occurrences of eta

and one occurrence of rho.
Decompose A into M and Delta. Note that M is a double, and Delta has the
same uncertainty dependence as A.
[M,Delta] = lftdata(A);
class(M)

6-62

Decomposing Uncertain Objects (for Advanced Users)

ans =
double
Delta
UMAT: 6 Rows, 6 Columns
delta: real, nominal = 2, variability = [-1 1], 2 occurrences
eta: real, nominal = 6, variability = [-1 1], 3 occurrences
rho: real, nominal = -1, variability = [-1 1], 1 occurrence

Sample Delta at 5 points. Things to note are: it is diagonal; the values range
between -1 and 1; there are 3 independent values, and duplication of the
entries is consistent with the dependence of Delta and A on the 3 uncertain
real parameters.
usample(Delta,5)
ans(:,:,1) =
-0.7106
0
0
-0.7106
0
0
0
0
0
0
0
0
ans(:,:,2) =
-0.5850
0
0
-0.5850
0
0
0
0
0
0
0
0
ans(:,:,3) =
0.7013
0
0
0.7013
0
0
0
0
0
0
0
0
ans(:,:,4) =
0.4262
0
0
0.4262
0
0
0
0
0
0

0
0
0.6374
0
0
0

0
0
0
0.6374
0
0

0
0
0
0
0.6374
0

0
0
0
0
0
-0.1258

0
0
-0.3021
0
0
0

0
0
0
-0.3021
0
0

0
0
0
0
-0.3021
0

0
0
0
0
0
0.0803

0
0
-0.6749
0
0
0

0
0
0
-0.6749
0
0

0
0
0
0
-0.6749
0

0
0
0
0
0
0.3967

0
0
0.0795
0
0

0
0
0
0.0795
0

0
0
0
0
0.0795

0
0
0
0
0

6-63

Building Uncertain Models

0
0
ans(:,:,5) =
-0.8392
0
0
-0.8392
0
0
0
0
0
0
0
0

-0.9959

0
0
0.8467
0
0
0

0
0
0
0.8467
0
0

0
0
0
0
0.8467
0

0
0
0
0
0
0.6732

In fact, verify that the maximum gain of Delta is indeed 1.

maxnorm = wcnorm(Delta)
maxnorm =
LowerBound: 1.0000
UpperBound: 1.0004

Finally, verify that lft(Delta,M) is the same as A. Subtract (and use the
'full' option in simplify)
simplify(lft(Delta,M)-A,'full')
ans =
0
0
0
0

Advanced Syntax of lftdata

Even for the advanced user, the variable Delta will actually not be that
useful, as it is still a complex object. On the other hand, its internal structure
is described completely using a 3rd (and 4th) output argument.
[M,Delta,BlkStruct,NormUnc] = lftdata(A);

The rows of BlkStruct correspond to the uncertain atoms named in

fieldnames(A.Uncertainty). Note that the range/bound information about
each uncertain atom is not included in BlkStruct.
The elements of BlkStruct describe the size, type and number-of-copies of the
uncertain atoms in A, and implicitly delineate the exact block-diagonal
structure of Delta. Note that the range/bound information about each
uncertain atom is not included in BlkStruct.
BlkStruct(1)
ans =

6-64

Decomposing Uncertain Objects (for Advanced Users)

Name:
Size:
Type:
Occurrences:
BlkStruct(2)
ans =
Name:
Size:
Type:
Occurrences:
BlkStruct(3)
ans =
Name:
Size:
Type:
Occurrences:

'delta'
[1 1]
'ureal'
2

'eta'
[1 1]
'ureal'
3

'rho'
[1 1]
'ureal'
1

Together, these mean Delta is a block diagonal augmentation of the

normalized version of 3 uncertain atoms.
The first atom is named 'delta'. It is 1-by-1; it is of class ureal; and there
are 2 copies diagonally augmented.
The second atom is named 'eta'. It is 1-by-1; it is of class ureal; and there
are 3 copies diagonally augmented.
The third atom is named 'rho'. It is 1-by-1; it is of class ureal; and there is
1 copy,
The 4th output argument contains a cell array of normalized uncertain
elements. The cell array contains as many occurrences of each element as
there are occurrences in the original uncertain object A.
size(NormUnc)
ans =
6
1
NormUnc{1}
Uncertain Real Parameter: Name deltaNormalized, NominalValue 0,
variability = [-1 1]
isequal(NormUnc{2},NormUnc{1})

6-65

Building Uncertain Models

ans =
1
NormUnc{3}
Uncertain Real Parameter: Name etaNormalized, NominalValue 0,
variability = [-1 1]
isequal(NormUnc{4},NormUnc{3})
ans =
1
isequal(NormUnc{5},NormUnc{3})
ans =
1
NormUnc{6}
Uncertain Real Parameter: Name rhoNormalized, NominalValue 0,
variability = [-1 1]

Each normalized element has the string 'Normalized' appended to its

original name to avoid confusion. By normalized,
ureal objects have nominal value of 0, and range from -1 to 1.
ultidyn objects are norm bounded, with norm bound of 1.
ucomplex objects have nominal value of 0, and radius 1.
ucomplexm objects have nominal value of 0, and identity matrices for each
of the WL and WR weights.
The possible behaviors of Delta and blkdiag(NormUnc{:}) are the same.
Consequently, the possible behaviors of A and lft(blkdiag(NormUnc{:}),M)
are the same.
Hence, by manipulating M, BlkStruct and NormUnc, a power-user has direct
access to all of the linear fractional transformation details, and can easily
work at the level of the theorems and algorithms that underlie the methods.

6-66

7
Generalized Robustness
Analysis

Introduction to Generalized
Robustness Analysis (p. 7-2)

What is generalized robustness analysis?

Robust Stability Margin (p. 7-4)

A brief discussion of robust stability margins

Robust Performance Margin (p. 7-5)

The definition of robust performance margins

Worst-Case Gain Measure (p. 7-6)

The maximum achievable gain over all uncertain system

objects

Generalized Robustness Analysis

Introduction to Generalized Robustness Analysis

The performance of a nominally stable uncertain system model will generally
degrade for specific values of its uncertain elements. Moreover, the maximum
possible degradation increases as the uncertain elements are allowed to
further and further deviate from their nominal values.

Maximum System Gain over Uncertainty

The graph below shows the typical tradeoff curve between allowable deviation
of uncertain elements from their nominal values and the worst-case
degradation in system performance. Here, system performance is
characterized by system gain (e.g., peak magnitude on Bode plot). Interpreting
the system as the relationship mapping disturbances/commands to errors,
small system gains are desirable, and large gains are undesirable.

3
2.5

Maximum System Gain due to

varying amounts of uncertainty

System performance
degradation curve

1.72

1.5
Nominal System
Gain = 0.8

1
0.5
0
0

System Gain as large as 1.72 if

uncertain elements can deviate from
their nominal values by 1.5 units.

0.5
1
1.5
2
2.5
Bound on Normalized Uncertainty

When all uncertain elements are set to their nominal values (i.e., zero
deviation from their nominal values) the input/output gain of the system is its
nominal value. In the figure, the nominal system gain is about 0.8. As the
uncertainties are allowed to deviate from nominal, the maximum (over the

7-2

Introduction to Generalized Robustness Analysis

uncertain elements) system gain increases. The heavy blue line represents the
maximum system gain due to uncertainty of various sizes (the horizontal axis).
and is called the system performance degradation curve. It is monotonically
increasing.
Determining specific attributes of the system performance degradation curve
are referred to as robustness computations.
Generally, robustness computations refer to determining specific attributes
of the system performance degradation curve. The commands robuststab,
robustperf and wcgain all compute single scalar attributes of the system
performance degradation curve.
Redraw the system performance degradation curve with 3 additional curves: a
hyperbola defined by xy=1; a vertical line drawn at the uncertainty bound = 1;
and a vertical line tangent to the asymptotic behavior of the performance
degradation curve at large uncertainty bounds. These are used to define three
robustness measures, explained next.

Maximum System Gain over Uncertainty

WCGain uses bound of

1 on normalized uncertainty

2.5

Uncertainty level
at which system can
become unstable

1.5
WCGain=1.22
y=1/x curve in
uncertainty size
.vs. performance
tradeoff space

1
PerfMarg=0.88

0.5
StabMarg=1.9

0
0

0.5

1
1.5
2
2.5
Bound on Normalized Uncertainty

7-3

Generalized Robustness Analysis

Robust Stability Margin

The robust stability margin, StabMarg, is the size of the smallest deviation from
nominal of the uncertain elements that leads to system instability.

Maximum System Gain over Uncertainty

Robust Stability Margin

2.5

System performance
degradation curve

Uncertainty level
at which system can
become unstable

1.5

0.5
StabMarg=1.9

0
0

0.5

1
1.5
2
2.5
Bound on Normalized Uncertainty

System instability is equivalent to the system gain becoming arbitrarily large,

and hence characterized by the vertical line tangent to the asymptotic behavior
of the performance degradation curve.

7-4

Robust Performance Margin

Robust Performance Margin

The hyperbola is used to define the performance margin. Systems whose
performance degradation curve intersects high on the hyperbola curve
represent non-robustly performing systems in that very small deviations of
the uncertain elements from their nominal values can result in very large
system gains. Conversely, an intersection low on the hyperbola represent
robustly performing systems.

Maximum System Gain over Uncertainty

Robust Performance Margin

2.5

2
System performance
degradation curve

1.5

1
PerfMarg=0.88

y=1/x curve in
uncertainty size
.vs. performance
tradeoff space

0.5

0
0

0.5

1
1.5
2
2.5
Bound on Normalized Uncertainty

The point where the system performance degradation curve crosses the green
line is used as a scalar measure of the robustness of a system to uncertainty.
The horizontal coordinate of the crossing point is the robust performance
margin, PerfMarg.

7-5

Generalized Robustness Analysis

Worst-Case Gain Measure

The worst-case gain measure is the maximum achievable system gain over all
uncertain elements whose normalized size is bounded by 1.

Maximum System Gain over Uncertainty

WCGain uses bound of

1 on normalized uncertainty

2.5

2
System performance
degradation curve

1.5
WCGain=1.22

0.5

0
0

WorstCase Gain Measure

0.5

1
1.5
2
2.5
Bound on Normalized Uncertainty

On the graph, this is the vertical coordinate of the performance degradation

curve as it crosses the vertical line drawn at the uncertainty bound = 1.
Each measure captures a single scalar attribute of the system performance
degradation curve. Mathematically, they are independent quantities,
answering subtlely different questions. Consequently, for two uncertain
systems, sysA and sysB, it is possible that the StabMarg of sysA is larger than
the StabMarg of sysB, though the PerfMarg of sysA is smaller than the
PerfMarg of sysB. Nevertheless, they are useful metrics for concise description
of the robustness of a system (uss or ufrd) due to various uncertain elements.

7-6

8
Introduction to Linear
Matrix Inequalities

Linear Matrix Inequalities (p. 8-2)

An introduction to the concept of linear matrix

inequalities and what LMI functionality can do for you

LMIs and LMI Problems (p. 8-4)

The basic properties of LMIs

Further Mathematical Background

(p. 8-9)

Detailed mathematical development of LMI theory

References (p. 8-10)

Relevant papers on linear matrix inequalities

Introduction to Linear Matrix Inequalities

Linear Matrix Inequalities

Linear Matrix Inequalities (LMIs) and LMI techniques have emerged as
powerful design tools in areas ranging from control engineering to system
identification and structural design. Three factors make LMI techniques
appealing:
A variety of design specifications and constraints can be expressed as LMIs.
Once formulated in terms of LMIs, a problem can be solved exactly by
efficient convex optimization algorithms (see LMI Solvers on page 1-18).
While most problems with multiple constraints or objectives lack analytical
solutions in terms of matrix equations, they often remain tractable in the
LMI framework. This makes LMI-based design a valuable alternative to
classical analytical methods.
See [9] for a good introduction to LMI concepts. The LMI Control Toolbox is
designed as an easy and progressive gateway to the new and fast-growing field
of LMIs:
For users who occasionally need to solve LMI problems, the LMI Editor and
the tutorial introduction to LMI concepts and LMI solvers provide for quick
and easy problem solving.
For more experienced LMI users, Chapter 9, The LMI Lab, offers a rich,
flexible, and fully programmable environment to develop customized
LMI-based tools.

LMI Features
The Robust Control Toolbox LMI functionality serves two purposes:
Provide state-of-the-art tools for the LMI-based analysis and design of robust
control systems
Offer a flexible and user-friendly environment to specify and solve general
LMI problems (the LMI Lab)
Examples of LMI-based analysis and design tools include
Functions to analyze the robust stability and performance of uncertain
systems with varying parameters (popov, quadstab, quadperf ...)

8-2

Linear Matrix Inequalities

Functions to design robust control with a mix of H2, H, and pole placement
objectives (h2hinfsyn)
Functions for synthesizing robust gain-scheduled H controllers (hinfgs)
For users interested in developing their own applications, the LMI Lab
provides a general-purpose and fully programmable environment to specify
and solve virtually any LMI problem. Note that the scope of this facility is by
no means restricted to control-oriented applications.

Note The LMI Control Toolbox implements state-of-the-art interior-point

LMI solvers. While these solvers are significantly faster than classical convex
optimization algorithms, you should keep in mind that the complexity of LMI
computations can grow quickly with the problem order (number of states). For
example, the number of operations required to solve a Riccati equation is o(n3)
where n is the state dimension, while the cost of solving and equivalent
Riccati inequality LMI is o(n6).

8-3

Introduction to Linear Matrix Inequalities

LMIs and LMI Problems

A linear matrix inequality (LMI) is any constraint of the form
A ( x ) := A 0 + x 1 A 1 + + x N A N < 0
where
x = (x1, . . . , xN) is a vector of unknown scalars (the decision or optimization
variables)
A0, . . . , AN are given symmetric matrices
< 0 stands for negative definite, i.e., the largest eigenvalue of A(x) is
negative
Note that the constraints A(x) > 0 and A(x) < B(x) are special cases of (1-1) since
they can be rewritten as A(x) < 0 and A(x) B(x) < 0, respectively.
The LMI (1-1) is a convex constraint on x since A(y) < 0 and A(z) < 0 imply that
+ z-
A y----------< 0 . As a result,
2

Its solution set, called the feasible set, is a convex subset of RN

Finding a solution x to (1-1), if any, is a convex optimization problem.
Convexity has an important consequence: even though (1-1) has no analytical
solution in general, it can be solved numerically with guarantees of finding a
solution when one exists. Note that a system of LMI constraints can be
regarded as a single LMI since

A1 ( x ) < 0

..

is equivalent to A ( x ) := diag ( A 1 ( x ),, A K ( x ) ) < 0

.

AK ( x ) < 0

where diag (A1(x), . . . , AK(x)) denotes the block-diagonal matrix with

A1(x), . . . , AK(x) on its diagonal. Hence multiple LMI constraints can be
imposed on the vector of decision variables x without destroying convexity.

8-4

LMIs and LMI Problems

In most control applications, LMIs do not naturally arise in the canonical form
(1-1), but rather in the form
L(X1, . . . , Xn) < R(X1, . . . , Xn)
where L(.) and R(.) are affine functions of some structured matrix variables X1,
. . . , Xn. A simple example is the Lyapunov inequality
T

A X + XA < 0

(8-1)

where the unknown X is a symmetric matrix. Defining x1, . . . , xN as the

independent scalar entries of X, this LMI could be rewritten in the form (1-1).
Yet it is more convenient and efficient to describe it in its natural form (1-2),
which is the approach taken in the LMI Lab.

The Three Generic LMI Problems

Finding a solution x to the LMI system
A(x) < 0

(8-2)

is called the feasibility problem. Minimizing a convex objective under LMI

constraints is also a convex problem. In particular, the linear objective
minimization problem
T

Minimize c x subject to A ( x ) < 0

(8-3)

plays an important role in LMI-based design. Finally, the generalized

eigenvalue minimization problem
A ( x ) < B ( x )

Minimize subject to B ( x ) > 0

C(x) < 0

(8-4)

is quasi-convex and can be solved by similar techniques. It owes its name to the
fact that is related to the largest generalized eigenvalue of the pencil
(A(x),B(x)).
Many control problems and design specifications have LMI formulations [9].
This is especially true for Lyapunov-based analysis and design, but also for
optimal LQG control, H control, covariance control, etc. Further applications
of LMIs arise in estimation, identification, optimal design, structural

8-5

Introduction to Linear Matrix Inequalities

design[6], [7], matrix scaling problems, and so on. The main strength of LMI
formulations is the ability to combine various design constraints or objectives
in a numerically tractable manner.
A nonexhaustive list of problems addressed by LMI techniques includes the
following:
Robust stability of systems with LTI uncertainty (-analysis) ([24], [21], [27])
Robust stability in the face of sector-bounded nonlinearities (Popov criterion)
([22], [28], [13], [16])
Quadratic stability of differential inclusions ([15], [8])
Lyapunov stability of parameter-dependent systems ([12])
Input/state/output properties of LTI systems (invariant ellipsoids, decay
rate, etc.) ([9])
Multi-model/multi-objective state feedback design ([4], [17], [3], [9], [10])
Robust pole placement
Optimal LQG control ([9])
Robust H control ([11], [14])
Multi-objective H synthesis ([18], [23], [10], [18])
Design of robust gain-scheduled controllers ([5], [2])
Control of stochastic systems ([9])
Weighted interpolation problems ([9])
To hint at the principles underlying LMI design, lets review the LMI
formulations of a few typical design objectives.

Stability
The stability of the dynamic system
x = Ax
is equivalent to the feasibility of
Find P = PT such that AT P + P A < 0, P > I.
This can be generalized to linear differential inclusions (LDI)
x = A ( t )x

8-6

LMIs and LMI Problems

where A(t) varies in the convex envelope of a set of LTI models:

N
n

A ( t ) Co { A 1 ,,A n } =
a i A i : a i 0,
ai = 1

i=1
i = 1

A sufficient condition for the asymptotic stability of this LDI is the feasibility of
T

Find P = PT such that A i P + PA i < 0, P > I.

RMS gain
The random-mean-squares (RMS) gain of a stable LTI system
x = Ax + Bu

y = Cx + Du
is the largest input/output gain over all bounded inputs u(t). This gain is the
global minimum of the following linear objective minimization problem [1],
[25], [26].
Minimize over X = XT and such that
T
T
A X + XA XB C

T
T
B X
I D < 0

C
D I
X > 0

LQG Performance
For a stable LTI system
x = Ax + Bw
G
y = Cx
where w is a white noise disturbance with unit covariance, the LQG or H2
performance ||G||2 is defined by

8-7

Introduction to Linear Matrix Inequalities

1 T T

2
G 2 : = lim E ---- y ( t )y ( t )dt
T
T
0

1 H
= -----G ( j )G ( j )d
2

It can be shown that

2

G 2 = inf { Trace ( CPC ) : AP + PA + BB < 0 }

2

Hence G 2 is the global minimum of the LMI problem

Minimize Trace (Q) over the symmetric matrices P,Q such that
T

AP + PA + BB < 0
Q CP

>0
PC T P
Again this is a linear objective minimization problem since the objective Trace
(Q) is linear in the decision variables (free entries of P,Q).

8-8

Further Mathematical Background

Further Mathematical Background

Efficient interior-point algorithms are now available to solve the three generic
LMI problems (8-2)(8-4) defined in The Three Generic LMI Problems on
page 8-5. These algorithms have a polynomial-time complexity. That is, the
number N() of flops needed to compute an -accurate solution is bounded by
M N3 log(V/)
where M is the total row size of the LMI system, N is the total number of scalar
decision variables, and V is a data-dependent scaling factor. The Robust
Control Toolbox implements the Projective Algorithm of Nesterov and
Nemirovski [20], [19]. In addition to its polynomial-time complexity, this
algorithm does not require an initial feasible point for the linear objective
minimization problem (8-3) or the generalized eigenvalue minimization
problem (8-4).
Some LMI problems are formulated in terms of inequalities rather than strict
inequalities. For instance, a variant of (8-3) is
Minimize cTx subject to A(x) < 0.
While this distinction is immaterial in general, it matters when A(x) can be
made negative semi-definite but not negative definite. A simple example is

T
Minimize c x subject to x x 0
x x

(8-5)

Such problems cannot be handled directly by interior-point methods which

require strict feasibility of the LMI constraints. A well-posed reformulation of
(8-5) would be
Minimize cTx subject to x 0.
Keeping this subtlety in mind, we always use strict inequalities in this manual.

8-9

Introduction to Linear Matrix Inequalities

References
[1] Anderson, B.D.O., and S. Vongpanitlerd, Network Analysis, Prentice-Hall,
Englewood Cliffs, 1973.
[2] Apkarian, P., P. Gahinet, and G. Becker, Self-Scheduled H Control of
Linear Parameter-Varying Systems, Proc. Amer. Contr. Conf., 1994, pp.
856-860.
[3] Bambang, R., E. Shimemura, and K. Uchida, Mixed H2 /H Control with
Pole Placement, State-Feedback Case, Proc. Amer. Contr. Conf., 1993, pp.
2777-2779.
[4] Barmish, B.R., Stabilization of Uncertain Systems via Linear Control,
IEEE Trans. Aut. Contr., AC28 (1983), pp. 848-850.
[5] Becker, G., and Packard, P., Robust Performance of Linear-Parametrically
Varying Systems Using Parametrically-Dependent Linear Feedback, Systems
and Control Letters, 23 (1994), pp. 205-215.
[6] Bendsoe, M.P., A. Ben-Tal, and J. Zowe, Optimization Methods for Truss
Geometry and Topology Design, to appear in Structural Optimization.
[7] Ben-Tal, A., and A. Nemirovski, Potential Reduction Polynomial-Time
Method for Truss Topology Design, to appear in SIAM J. Contr. Opt.
[8] Boyd, S., and Q. Yang, Structured and Simultaneous Lyapunov Functions
for System Stability Problems, Int. J. Contr., 49 (1989), pp. 2215-2240.
[9] Boyd, S., L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix
Inequalities in Systems and Control Theory, SIAM books, Philadelphia, 1994.
[10] Chilali, M., and P. Gahinet, H Design with Pole Placement Constraints:
an LMI Approach, to appear in IEEE Trans. Aut. Contr. Also in Proc. Conf.
Dec. Contr., 1994, pp. 553-558.
[11] Gahinet, P., and P. Apkarian, A Linear Matrix Inequality Approach to H
Control, Int. J. Robust and Nonlinear Contr., 4 (1994), pp. 421-448.
[12] Gahinet, P., P. Apkarian, and M. Chilali, Affine Parameter-Dependent
Lyapunov Functions for Real Parametric Uncertainty, Proc. Conf. Dec. Contr.,
1994, pp. 2026-2031.
[13] Haddad, W.M., and D.S. Berstein,Parameter-Dependent Lyapunov
Functions, Constant Real Parameter Uncertainty, and the Popov Criterion in

8-10

References

Robust Analysis and Synthesis: Part 1 and 2, Proc. Conf. Dec. Contr., 1991, pp.
2274-2279 and 2632-2633.
[14] Iwasaki, T., and R.E. Skelton, All Controllers for the General H Control
Problem: LMI Existence Conditions and State-Space Formulas, Automatica,
30 (1994), pp. 1307-1317.
[15] Horisberger, H.P., and P.R. Belanger, Regulators for Linear
Time-Varying Plants with Uncertain Parameters, IEEE Trans. Aut. Contr.,
AC21 (1976), pp. 705-708.
[16] How, J.P., and S.R. Hall, Connection between the Popov Stability
Criterion and Bounds for Real Parameter Uncertainty, Proc. Amer. Contr.
Conf., 1993, pp. 1084-1089.
[17] Khargonekar, P.P., and M.A. Rotea,Mixed H2 /H Control: a Convex
Optimization Approach, IEEE Trans. Aut. Contr., 39 (1991), pp. 824-837.
[18] Masubuchi, I., A. Ohara, and N. Suda, LMI-Based Controller Synthesis:
A Unified Formulation and Solution, submitted to Int. J. Robust and
Nonlinear Contr., 1994.
[19] Nemirovski, A., and P. Gahinet, The Projective Method for Solving Linear
Matrix Inequalities, Proc. Amer. Contr. Conf., 1994, pp. 840-844.
[20] Nesterov, Yu, and A. Nemirovski, Interior Point Polynomial Methods in
Convex Programming: Theory and Applications, SIAM Books, Philadelphia,
1994.
[21] Packard, A., and J.C. Doyle, The Complex Structured Singular Value,
Automatica, 29 (1994), pp. 71-109.
[22] Popov, V.M., Absolute Stability of Nonlinear Systems of Automatic
Control, Automation and Remote Control, 22 (1962), pp. 857-875.
[23] Scherer, C., Mixed H2 H Control, to appear in Trends in Control: A
European Perspective, volume of the special contributions to the ECC 1995.
[24] Stein, G., and J.C. Doyle, Beyond Singular Values and Loop Shapes, J.
Guidance, 14 (1991), pp. 5-16.
[25] Vidyasagar, M., Nonlinear System Analysis, Prentice-Hall, Englewood
Cliffs, 1992.

8-11

Introduction to Linear Matrix Inequalities

[26] Willems, J.C., Least-Squares Stationary Optimal Control and the

Algebraic Riccati Equation, IEEE Trans. Aut. Contr., AC16 (1971), pp.
621-634.
[27] Young, P.M., M.P. Newlin, and J.C. Doyle, Lets Get Real, in Robust
Control Theory, Springer Verlag, 1994, pp. 143-174.
[28] Zames, G., On the Input-Output Stability of Time-Varying Nonlinear
Feedback Systems, Part I and II, IEEE Trans. Aut. Contr., AC11 (1966), pp.
228-238 and 465-476.

8-12

9
The LMI Lab

Introduction on page 9-2

A quick look at the LMI Lab, canonical forms of the LMI

problem, and the tools available in the LMI Lab

Specifying a System of LMIs (p. 9-7)

How to set up a system of LMIs

Querying the LMI System Description

(p. 9-20)

How to find out characteristics of your LMI system

LMI Solvers (p. 9-21)

How the LMI solvers work

From Decision to Matrix Variables and The relationships between decision variables and matrix
Vice Versa (p. 9-27)
variables
Validating Results (p. 9-28)

Verifying your solutions

Modifying a System of LMIs (p. 9-29)

Adapting an existing system to a new problem

Advanced Topics (p. 9-32)

Topics for the experienced user

References (p. 9-43)

A list of relevant papers

The LMI Lab

Introduction
The LMI Lab is a high-performance package for solving general LMI problems.
It blends simple tools for the specification and manipulation of LMIs with
powerful LMI solvers for three generic LMI problems. Thanks to a
structure-oriented representation of LMIs, the various LMI constraints can be
described in their natural block-matrix form. Similarly, the optimization
variables are specified directly as matrix variables with some given structure.
Once an LMI problem is specified, it can be solved numerically by calling the
appropriate LMI solver. The three solvers feasp, mincx, and gevp constitute
the computational engine of the LMI Control Toolbox. Their high performance
is achieved through C-MEX implementation and by taking advantage of the
particular structure of each LMI.
The LMI Lab offers tools to
Specify LMI systems either symbolically with the LMI Editor or
incrementally with the lmivar and lmiterm commands
Retrieve information about existing systems of LMIs
Modify existing systems of LMIs
Solve the three generic LMI problems (feasibility problem, linear objective
minimization, and generalized eigenvalue minimization)
Validate results
This chapter gives a tutorial introduction to the LMI Lab as well as more
advanced tips for making the most out of its potential. The tutorial material is
also covered by the demo lmidem.

Some Terminology
Any linear matrix inequality can be expressed in the canonical form
L(x) = L0 + x1L1 + . . . + xNLN < 0
where
L0, L1, . . . , LN are given symmetric matrices
x = (x1, . . . , xN)T RN is the vector of scalar variables to be determined. We
refer to x1, . . . , xN as the decision variables. The names design variables
and optimization variables are also found in the literature.

9-2

Introduction

Even though this canonical expression is generic, LMIs rarely arise in this form
in control applications. Consider for instance the Lyapunov inequality
T

A X + XA < 0

(9-1)

x x

where A = 1 2 and the variable X = 1 2 is a symmetric matrix.

0 2
x2 x3

Here the decision variables are the free entries x1, x2, x3 of X and the canonical
form of this LMI reads

x1 2 2 + x2 0 3 + x3 0 0 < 0
2 0
3 4
0 4

(9-2)

Clearly this expression is less intuitive and transparent than (8-1). Moreover,
the number of matrices involved in (8-2) grows roughly as n2 /2 if n is the size
of the A matrix. Hence, the canonical form is very inefficient from a storage
viewpoint since it requires storing o(n2 /2) matrices of size n when the single
n-by-n matrix A would be sufficient. Finally, working with the canonical form
is also detrimental to the efficiency of the LMI solvers. For these various
reasons, the LMI Lab uses a structured representation of LMIs. For instance,
the expression ATX + XA in the Lyapunov inequality (8-1) is explicitly
described as a function of the matrix variable X, and only the A matrix is
stored.
In general, LMIs assume a block matrix form where each block is an affine
combination of the matrix variables. As a fairly typical illustration, consider
the following LMI drawn from H theory
T
T
A X + XA XC B
T
N
CX
I D

T
T
B
D I

N < 0

(9-3)

where A, B, C, D, N are given matrices and the problem variables are

X = XT Rnn and R. We use the following terminology to describe such
LMIs:

9-3

The LMI Lab

N is called the outer factor, and the block matrix

T
T
A X + XA XC B
L ( X, ) =
CX
I D

T
T
B
D I

is called the inner factor. The outer factor needs not be square and is often
absent.
X and are the matrix variables of the problem. Note that scalars are
considered as 1-by-1 matrices.
The inner factor L(X, ) is a symmetric block matrix, its block structure being
characterized by the sizes of its diagonal blocks. By symmetry, L(X, ) is
entirely specified by the blocks on or above the diagonal.
Each block of L(X, ) is an affine expression in the matrix variables X and .
This expression can be broken down into a sum of elementary terms. For
instance, the block (1,1) contains two elementary terms: ATX and XA.
Terms are either constant or variable. Constant terms are fixed matrices like
B and D above. Variable terms involve one of the matrix variables, like XA,
XCT, and I above.
The LMI (Equation (9-3)) is specified by the list of terms in each block, as is any
LMI regardless of its complexity.
As for the matrix variables X and , they are characterized by their dimensions
and structure. Common structures include rectangular unstructured,
symmetric, skew-symmetric, and scalar. More sophisticated structures are
sometimes encountered in control problems. For instance, the matrix variable
X could be constrained to the block-diagonal structure
x 0 0
1
X = 0 x2 x3

0 x x
3 4

9-4

Introduction

Another possibility is the symmetric Toeplitz structure

x x x
1 2 3
X = x2 x1 x2

x x x
3 2 1

Summing up, structured LMI problems are specified by declaring the matrix
variables and describing the term content of each LMI. This term-oriented
description is systematic and accurately reflects the specific structure of the
LMI constraints. There is no built-in limitation on the number of LMIs that you
can specify or on the number of blocks and terms in any given LMI. LMI
systems of arbitrary complexity can therefore, be defined in the LMI Lab.

Overview of the LMI Lab

The LMI Lab offers tools to specify, manipulate, and numerically solve LMIs.
Its main purpose is to
Allow for straightforward description of LMIs in their natural block-matrix
form
Provide easy access to the LMI solvers (optimization codes)
Facilitate result validation and problem modification
The structure-oriented description of a given LMI system is stored as a single
vector called the internal representation and generically denoted by LMISYS in
the sequel. This vector encodes the structure and dimensions of the LMIs and
matrix variables, a description of all LMI terms, and the related numerical
data. It must be stressed that you need not attempt to read or understand the
content of LMISYS since all manipulations involving this internal
representation can be performed in a transparent manner with LMI-Lab tools.
The LMI Lab supports the following functionalities:

Specification of a System of LMIs

LMI systems can be either specified as symbolic matrix expressions with the
interactive graphical user interface lmiedit, or assembled incrementally with
the two commands lmivar and lmiterm. The first option is more intuitive and
transparent while the second option is more powerful and flexible.

9-5

The LMI Lab

Information Retrieval
The interactive function lmiinfo answers qualitative queries about LMI
systems created with lmiedit or lmivar and lmiterm. You can also use
lmiedit to visualize the LMI system produced by a particular sequence of
lmivar/lmiterm commands.

Solvers for LMI Optimization Problems

General-purpose LMI solvers are provided for the three generic LMI problems
defined in The Three Generic LMI Problems on page 8-5. These solvers can
handle very general LMI systems and matrix variable structures. They return
a feasible or optimal vector of decision variables x*. The corresponding values
X *1, , X *K of the matrix variables are given by the function dec2mat.

Result Validation
The solution x* produced by the LMI solvers is easily validated with the
functions evallmi and showlmi. This allows a fast check and/or analysis of the
results. With evallmi, all variable terms in the LMI system are evaluated for
the value x* of the decision variables. The left- and right-hand sides of each
LMI then become constant matrices that can be displayed with showlmi.

Modification of a System of LMIs

An existing system of LMIs can be modified in two ways:
An LMI can be removed from the system with dellmi.
A matrix variable X can be deleted using delmvar. It can also be instantiated,
that is, set to some given matrix value. This operation is performed by
setmvar and allows, for example, to fix some variables and solve the LMI
problem with respect to the remaining ones.

9-6

Specifying a System of LMIs

Specifying a System of LMIs

The LMI Lab can handle any system of LMIs of the form
NT L(X1, . . . , XK) N < MT R(X1, . . . , XK) M
where
X1, . . . , XK are matrix variables with some prescribed structure
The left and right outer factors N and M are given matrices with identical
dimensions
The left and right inner factors L(.) and R(.) are symmetric block matrices
with identical block structures, each block being an affine combination of X1,
. . . , XK and their transposes.

Note Throughout this chapter, left-hand side refers to what is on the

smaller side of the inequality, and right-hand side to what is on the
larger side. Accordingly, X is called the right-hand side and 0 the left-hand
side of the LMI
0<X
even when this LMI is written as X > 0.

The specification of an LMI system involves two steps:

1 Declare the dimensions and structure of each matrix variable X1, . . . , XK .
2 Describe the term content of each LMI.

This process creates the so-called internal representation of the LMI system.
This computer description of the problem is used by the LMI solvers and in all
subsequent manipulations of the LMI system. It is stored as a single vector
called LMISYS.
There are two ways of generating the internal description of a given LMI
system: (1) by a sequence of lmivar/lmiterm commands that build it
incrementally, or (2) via the LMI Editor lmiedit where LMIs can be specified
directly as symbolic matrix expressions. Though somewhat less flexible and
powerful than the command-based description, the LMI Editor is more
straightforward to use, hence particularly well-suited for beginners. Thanks to

9-7

The LMI Lab

its coding and decoding capabilities, it also constitutes a good tutorial

introduction to lmivar and lmiterm. Accordingly, beginners may elect to skip
the subsections on lmivar and lmiterm and to concentrate on the GUI-based
specification of LMIs with lmiedit.

A Simple Example
The following tutorial example is used to illustrate the specification of LMI
systems with the LMI Lab tools. Run the demo lmidem to see a complete
treatment of this example.

Example 8.1
Consider a stable transfer function
1

G ( s ) = C ( sI A ) B

(9-4)

with four inputs, four outputs, and six states, and consider the set of
input/output scaling matrices D with block-diagonal structure

D =

d 1 0 0 0
0 d 1 0 0
0 0 d 2 d 3
0 0 d 4 d 5

(9-5)

The following problem arises in the robust stability analysis of systems with
time-varying uncertainty [4]:
Find, if any, a scaling D of structure (Equation (9-5)) such that the largest
gain across frequency of D G(s) D1 is less than one.
This problem has a simple LMI formulation: there exists an adequate scaling
D if the following feasibility problem has solutions:
Find two symmetric matrices X R66 and S = DT D R44 such that

A T X + XA + C T SC XB

<0

T
B X
S

9-8

(9-6)

Specifying a System of LMIs

X>0

(9-7)

S>1

(9-8)

The LMI system (Equation (9-6), Equation (9-7), and Equation (9-8)) can be
described with the LMI Editor as outlined below. Alternatively, its internal
description can be generated with lmivar and lmiterm commands as follows:
setlmis([])
X=lmivar(1,[6 1])
S=lmivar(1,[2 0;2 1])
% 1st LMI
lmiterm([1
lmiterm([1
lmiterm([1
lmiterm([1

1
1
1
2

1
1
2
2

X],1,A,'s')
S],C',C)
X],1,B)
S], 1,1)

% 2nd LMI
lmiterm([ 2 1 1 X],1,1)
% 3rd LMI
lmiterm([ 3 1 1 S],1,1)
lmiterm([3 1 1 0],1)
LMISYS = getlmis

Here the lmivar commands define the two matrix variables X and S while the
lmiterm commands describe the various terms in each LMI. Upon completion,
getlmis returns the internal representation LMISYS of this LMI system. The
following subsections give more details on the syntax and usage of these
various commands:
Initializing the LMI System on page 9-10
Specifying the LMI Variables on page 9-10
Specifying Individual LMIs on page 9-13
More information on how the internal representation is updated by
lmivar/lmiterm can also be found in How It All Works on page 9-19.

9-9

The LMI Lab

Initializing the LMI System

The description of an LMI system should begin with setlmis and end with
getlmis. The function setlmis initializes the LMI system description. When
specifying a new system, type
setlmis([])

To add on to an existing LMI system with internal representation LMIS0, type

setlmis(LMIS0)

Specifying the LMI Variables

The matrix variables are declared one at a time with lmivar and are
characterized by their structure. To facilitate the specification of this structure,

9-10

Specifying a System of LMIs

the LMI Lab offers two predefined structure types along with the means to
describe more general structures:
Type 1

Symmetric block diagonal structure. This corresponds to matrix

variables of the form

...
... ...

D 0 0
1

0 D

2
X =

0 0 Dr
where each diagonal block Dj is square and is either zero, a full
symmetric matrix, or a scalar matrix
Dj = d I, d R
This type encompasses ordinary symmetric matrices (single block)
and scalar variables (one block of size one).
Type 2

Rectangular structure. This corresponds to arbitrary rectangular

matrices without any particular structure.

Type 3

General structures. This third type is used to describe more

sophisticated structures and/or correlations between the matrix
variables. The principle is as follows: each entry of X is specified
independently as either 0, xn, or xn where xn denotes the n-th
decision variable in the problem. For details on how to use Type 3,
see Structured Matrix Variables on page 9-32 as well as the
lmivar entry in the reference pages.

In Example 8.1 on page 9-8, the matrix variables X and S are of Type 1.
Indeed, both are symmetric and S inherits the block-diagonal structure (8-5) of
D. Specifically, S is of the form

9-11

The LMI Lab

S =

s 1 0 0 0
0 s 1 0 0
0 0 s 2 s 3
0 0 s 3 s 4

After initializing the description with the command setlmis([]), these two
matrix variables are declared by
lmivar(1,[6 1]) % X
lmivar(1,[2 0;2 1]) % S

In both commands, the first input specifies the structure type and the second
input contains additional information about the structure of the variable:
For a matrix variable X of Type 1, this second input is a matrix with two
columns and as many rows as diagonal blocks in X. The first column lists the
sizes of the diagonal blocks and the second column specifies their nature with
the following convention:
1 full symmetric block
0 scalar block
1 zero block
In the second command, for instance, [2 0;2 1] means that S has two
diagonal blocks, the first one being a 2-by-2 scalar block and the second one
a 22 full block.
For matrix variables of Type 2, the second input of lmivar is a two-entry
vector listing the row and column dimensions of the variable. For instance, a
3-by-5 rectangular matrix variable would be defined by
lmivar(2,[3 5])

For convenience, lmivar also returns a tag that identifies the matrix variable
for subsequent reference. For instance, X and S in Example 8.1 could be
defined by
X = lmivar(1,[6 1])
S = lmivar(1,[2 0;2 1])

9-12

Specifying a System of LMIs

The identifiers X and S are integers corresponding to the ranking of X and S in

the list of matrix variables (in the order of declaration). Here their values
would be X=1 and S=2. Note that these identifiers still point to X and S after
deletion or instantiation of some of the matrix variables. Finally, lmivar can
also return the total number of decision variables allocated so far as well as the
entry-wise dependence of the matrix variable on these decision variables (see
the lmivar entry in the reference pages for more details).

Specifying Individual LMIs

After declaring the matrix variables with lmivar, we are left with specifying
the term content of each LMI. Recall that LMI terms fall into three categories:
The constant terms, i.e., fixed matrices like I in the left-hand side of the LMI
S>I
The variable terms, i.e., terms involving a matrix variable. For instance, ATX
and CTSC in (8-6). Variable terms are of the form PXQ where X is a variable
and P, Q are given matrices called the left and right coefficients, respectively.
The outer factors
The following rule should be kept in mind when describing the term content of
an LMI:

Important: Specify only the terms in the blocks on or above the diagonal. The
inner factors being symmetric, this is sufficient to specify the entire LMI.
Specifying all blocks results in the duplication of off-diagonal terms, hence in
the creation of a different LMI. Alternatively, you can describe the blocks on or
below the diagonal.

LMI terms are specified one at a time with lmiterm. For instance, the LMI

A T X + XA + C T SC XB

<0

T
B X
S

is described by
lmiterm([1 1 1 1],1,A,'s')

9-13

The LMI Lab

lmiterm([1 1 1 2],C',C)
lmiterm([1 1 2 1],1,B)
lmiterm([1 2 2 2], 1,1)

These commands successively declare the terms ATX + XA, CTSC, XB, and S.
In each command, the first argument is a four-entry vector listing the term
characteristics as follows:
The first entry indicates to which LMI the term belongs. The value m means
left-hand side of the m-th LMI, and m means right-hand side of the m-th
LMI.
The second and third entries identify the block to which the term belongs.
For instance, the vector [1 1 2 1] indicates that the term is attached to the
(1, 2) block.
The last entry indicates which matrix variable is involved in the term. This
entry is 0 for constant terms, k for terms involving the k-th matrix variable
T
Xk, and k for terms involving X k (here X and S are first and second
variables in the order of declaration).
Finally, the second and third arguments of lmiterm contain the numerical data
(values of the constant term, outer factor, or matrix coefficients P and Q for
variable terms PXQ or PXTQ). These arguments must refer to existing
MATLAB variables and be real-valued. See Complex-Valued LMIs on
page 9-34 for the specification of LMIs with complex-valued coefficients.
Some shorthand is provided to simplify term specification. First, blocks are
zero by default. Second, in diagonal blocks the extra argument 's' allows you
to specify the conjugated expression AXB + BTXTAT with a single lmiterm
command. For instance, the first command specifies ATX + XA as the
symmetrization of XA. Finally, scalar values are allowed as shorthand for
scalar matrices, i.e., matrices of the form I with scalar. Thus, a constant
term of the form I can be specified as the scalar . This also applies to the
coefficients P and Q of variable terms. The dimensions of scalar matrices are
inferred from the context and set to 1 by default. For instance, the third LMI S
> I in Example 8.3 on page 9-32 is described by
lmiterm([ 3 1 1 2],1,1)
lmiterm([3 1 1 0],1)

% 1*S*1 = S
% 1*I = I

Recall that by convention S is considered as the right-hand side of the

inequality, which justifies the 3 in the first command.

9-14

Specifying a System of LMIs

Finally, to improve readability it is often convenient to attach an identifier

(tag) to each LMI and matrix variable. The variable identifiers are returned by
lmivar and the LMI identifiers are set by the function newlmi. These
identifiers can be used in lmiterm commands to refer to a given LMI or matrix
variable. For the LMI system of Example 8.1, this would look like:
setlmis([])
X = lmivar(1,[6 1])
S = lmivar(1,[2 0;2 1])
BRL = newlmi
lmiterm([BRL
lmiterm([BRL
lmiterm([BRL
lmiterm([BRL

1
1
1
2

1
1
2
2

X],1,A,'s')
S],C',C)
X],1,B)
S], 1,1)

Xpos = newlmi
lmiterm([-Xpos 1 1 X],1,1)
Slmi = newlmi
lmiterm([-Slmi 1 1 S],1,1)
lmiterm([Slmi 1 1 0],1)

When the LMI system is completely specified, type

LMISYS = getlmis

This returns the internal representation LMISYS of this LMI system. This
MATLAB description of the problem can be forwarded to other LMI-Lab
functions for subsequent processing. The command getlmis must be used only
once and after declaring all matrix variables and LMI terms.
Here the identifiers X and S point to the variables X and S while the tags BRL,
Xpos, and Slmi point to the first, second, and third LMI, respectively. Note that
Xpos refers to the right-hand side of the second LMI. Similarly, X would
indicate transposition of the variable X.

Specifying LMIs with the LMI Editor

The LMI Editor lmiedit is a graphical user interface (GUI) to specify LMI
systems in a straightforward symbolic manner. Typing
lmiedit

9-15

The LMI Lab

calls up a window with several editable text areas and various buttons.

Use the view

commands buttons to

Declare each matrix variable

(name and structure) here.

visualize the sequence

of lmivar/lmiterm
commands needed to
describe this LMI
system.

Specify the LMIs as MATLAB

expressions here.

Read a sequence of
lmivar/lmiterm

commands from a file by

clicking Read.

Generate the internal

representation of the LMI
system by clicking
Create.

Save/Load the symbolic

description of the LMI
system as a MATLAB string.

In more detail, to specify your LMI system,

1 Declare each matrix variable (name and structure) in the upper half of the

worksheet. The structure is characterized by its type (S for symmetric block

diagonal, R for unstructured, and G for other structures) and by an additional
structure matrix. This matrix contains specific information about the
structure and corresponds to the second argument of lmivar (see Specifying
the LMI Variables on page 9-10 for details).
Please use one line per matrix variable in the text editing areas.

9-16

Specifying a System of LMIs

2 Specify the LMIs as MATLAB expressions in the lower half of the

worksheet. For instance, the LMI

T

A X + XA XB < 0

I
B X
is entered by typing
[a'*x+x*a x*b; b'*x

1] < 0

if x is the name given to the matrix variable X in the upper half of the
worksheet. The left- and right-hand sides of the LMIs should be valid
MATLAB expressions.
Once the LMI system is fully specified, the following tasks can be performed by
clicking the corresponding button:
Visualize the sequence of lmivar/lmiterm commands needed to describe this
LMI system (view commands button). Conversely, the LMI system defined by
a particular sequence of lmivar/lmiterm commands can be displayed as a
MATLAB expression by clicking on the describe... buttons.
Beginners can use this facility as a tutorial introduction to the lmivar and
lmiterm commands.

Save the symbolic description of the LMI system as a MATLAB string (save
button). This description can be reloaded later on by clicking the load button.
Read a sequence of lmivar/lmiterm commands from a file (read button). You
can then click on describe the matrix variables or describe the LMIs to
visualize the symbolic expression of the LMI system specified by these
commands. The file should describe a single LMI system but may otherwise
contain any sequence of MATLAB commands.
This feature is useful for code validation and debugging.
Write in a file the sequence of lmivar/lmiterm commands needed to describe
a particular LMI system (write button).
This is helpful to develop code and prototype MATLAB functions based on
the LMI Lab.

9-17

The LMI Lab

Generate the internal representation of the LMI system by clicking create.

The result is written in a MATLAB variable named after the LMI system (if
the name of the LMI system is set to mylmi, the internal representation is
written in the MATLAB variable mylmi). Note that all LMI-related data
should be defined in the MATLAB workspace at this stage.
The internal representation can be passed directly to the LMI solvers or any
other LMI Lab function.

Keyboard Shortcuts
As with lmiterm, you can use various shortcuts when entering LMI expressions
at the keyboard. For instance, zero blocks can be entered simply as 0 and need
not be dimensioned. Similarly, the identity matrix can be entered as 1 without
dimensioning. Finally, upper diagonal LMI blocks need not be fully specified.
Rather, you can just type (*) in place of each such block.

Limitations
Though fairly general, lmiedit is not as flexible as lmiterm and the following
limitations should be kept in mind:
Parentheses cannot be used around matrix variables. For instance, the
expression
(a*x+b)'*c + c'*(a*x+b)

is invalid when x is a variable name. By contrast,

(a+b)'*x + x'*(a+b)
is perfectly valid.
Loops and if statements are ignored.
When turning lmiterm commands into a symbolic description of the LMI
system, an error is issued if the first argument of lmiterm cannot be
evaluated. Use the LMI and variable identifiers supplied by newlmi and
lmivar to avoid such difficulties.
Figure 8-1 shows how to specify the feasibility problem of Example 8.1 on
page 9-8 with lmiedit.

9-18

Specifying a System of LMIs

How It All Works

Users familiar with MATLAB may wonder how lmivar and lmiterm physically
update the internal representation LMISYS since LMISYS is not an argument to
these functions. In fact, all updating is performed through global variables for
maximum speed. These global variables are initialized by setlmis, cleared by
getlmis, and are not visible in the workspace. Even though this artifact is
transparent from the user's viewpoint, be sure to
Invoke getlmis only once and after completely specifying the LMI system
Refrain from using the command clear global before the LMI system
description is ended with getlmis

9-19

The LMI Lab

Querying the LMI System Description

Recall that the full description of an LMI system is stored as a single vector
called the internal representation. The user should not attempt to read or
retrieve information directly from this vector. The Robust Control Toolbox
provides three functions called lmiinfo, lminbr, and matnbr to extract and
display all relevant information in a user-readable format.

lmiinfo
lmiinfo is an interactive facility to retrieve qualitative information about LMI

systems. This includes the number of LMIs, the number of matrix variables
and their structure, the term content of each LMI block, etc. To invoke lmiinfo,
enter
lmiinfo(LMISYS)

where LMISYS is the internal representation of the LMI system produced by

getlmis.

lminbr and matnbr

These two functions return the number of LMIs and the number of matrix
variables in the system. To get the number of matrix variables, for instance,
enter
matnbr(LMISYS)

9-20

LMI Solvers

LMI Solvers
LMI solvers are provided for the following three generic optimization problems
(here x denotes the vector of decision variables, i.e., of the free entries of the
matrix variables X1, . . . , XK):
Feasibility problem
Find x RN (or equivalently matrices X1, . . . , XK with prescribed structure)
that satisfies the LMI system
A(x) < B(x)
The corresponding solver is called feasp.
Minimization of a linear objective under LMI constraints
Minimize cTx over x RN subject to A(x) < B(x)
The corresponding solver is called mincx.
Generalized eigenvalue minimization problem
Minimize over x RN subject to
C(x) < D(x)
0 < B(x)
A(x) < B(x).
The corresponding solver is called gevp.
Note that A(x) < B(x) above is a shorthand notation for general structured LMI
systems with decision variables x = (x1, . . . , xN).
The three LMI solvers feasp, mincx, and gevp take as input the internal
representation LMISYS of an LMI system and return a feasible or optimizing
value x* of the decision variables. The corresponding values of the matrix
variables X1, . . . , XK are derived from x* with the function dec2mat. These
solvers are C-MEX implementations of the polynomial-time Projective
Algorithm Projective Algorithm of Nesterov and Nemirovski [3], [2].
For generalized eigenvalue minimization problems, it is necessary to
distinguish between the standard LMI constraints C(x) < D(x) and the
linear-fractional LMIs
A(x) < B(x)

9-21

The LMI Lab

attached to the minimization of the generalized eigenvalue . When using

gevp, you should follow these three rules to ensure proper specification of the
problem:
Specify the LMIs involving as A(x) < B(x) (without the )
Specify them last in the LMI system. gevp systematically assumes that the
last L LMIs are linear-fractional if L is the number of LMIs involving
Add the constraint 0 < B(x) or any other constraint that enforces it. This
positivity constraint is required for well-posedness of the problem and is not
automatically added by gevp (see the reference pages for details).
An initial guess xinit for x can be supplied to mincx or gevp. Use mat2dec to
derive xinit from given values of the matrix variables X1, . . . , XK. Finally,
various options are available to control the optimization process and the solver
behavior. These options are described in detail in the reference pages.
The following example illustrates the use of the mincx solver.

Example 8.2
Consider the optimization problem
Minimize Trace(X) subject to
T

A X + XA + XBB X + Q < 0

(9-9)

with data
1 2 1
1
1 1 0

A = 3 2 1 ; B = 0 ; Q = 1 3 12

1 2 1
1
0 12 36

It can be shown that the minimizer X* is simply the stabilizing solution of the
algebraic Riccati equation
ATX + XA + XBBTX + Q = 0
This solution can be computed directly with the Riccati solver care and
compared to the minimizer returned by mincx.
From an LMI optimization standpoint, problem [9-9] is equivalent to the
following linear objective minimization problem:

9-22

LMI Solvers

Minimize Trace(X) subject to A X + XA + Q XB < 0

B X
I

(9-10)

Since Trace(X) is a linear function of the entries of X, this problem falls within
the scope of the mincx solver and can be numerically solved as follows:
1 Define the LMI constraint [9-9] by the sequence of commands

setlmis([])
X = lmivar(1,[3 1]) % variable X, full symmetric
lmiterm([1
lmiterm([1
lmiterm([1
lmiterm([1

1
1
2
2

1
1
2
1

X],1,a,'s')
0],q)
0],-1)
X],b',1)

LMIs = getlmis
2 Write the objective Trace(X) as cTx where x is the vector of free entries of X.

Since c should select the diagonal entries of X, it is obtained as the decision

vector corresponding to X = I, that is,
c = mat2dec(LMIs,eye(3))

Note that the function defcx provides a more systematic way of specifying
such objectives (see Specifying cTx Objectives for mincx on page 9-37 for
details).
3 Call mincx to compute the minimizer xopt and the global minimum

copt = c'*xopt of the objective:

options = [1e 5,0,0,0,0]
[copt,xopt] = mincx(LMIs,c,options)

Here 1e 5 specifies the desired relative accuracy on copt.

The following trace of the iterative optimization performed by mincx appears
on the screen:
Solver for linear objective minimization under LMI constraints

9-23

The LMI Lab

Iterations : Best objective value so far

1
2

8.511476

13.063640

***
4
***
5
***
6
***
7
***
8
***
9
***
10
***
11
***
12
***
13
***
14
***
15

9-24

new lower bound:

34.023978

15.768450
new lower bound:

25.005604

17.123012
new lower bound:

21.306781

17.882558
new lower bound:

19.819471

18-339853
new lower bound:

19.189417

18.552558
new lower bound:

18.919668

18.646811
new lower bound:

18.803708

18.687324
new lower bound:

18.753903

18.705715
new lower bound:

18.732574

18.712175
new lower bound:

18.723491

18.714880
new lower bound:

18.719624

18.716094
new lower bound:
18.716509

18.717986

LMI Solvers

***
16

new lower bound:

18.717297

18.716695

***

new lower bound:

18.716873

Result: feasible solution of required accuracy

best objective value: 18.716695
guaranteed relative accuracy: 9.50e 06
f-radius saturation: 0.000% of R = 1.00e+09

The iteration number and the best value of cTx at the current iteration
appear in the left and right columns, respectively. Note that no value is
displayed at the first iteration, which means that a feasible x satisfying the
constraint [9-10] was found only at the second iteration. Lower bounds on
the global minimum of cTx are sometimes detected as the optimization
progresses. These lower bounds are reported by the message
*** new lower bound: xxx

Upon termination, mincx reports that the global minimum for the objective
Trace(X) is 18.716695 with relative accuracy of at least 9.5-by-106. This is
the value copt returned by mincx.
4 mincx also returns the optimizing vector of decision variables xopt. The

corresponding optimal value of the matrix variable X is given by

Xopt = dec2mat(LMIs,xopt,X)

which returns
6.3542 5.8895 2.2046

X opt = 5.8895 6.2855 2.2201

2.2046 2.2201 6.0771

This result can be compared with the stabilizing Riccati solution computed
by care:
Xst = care(a,b,q, 1)
norm(Xopt-Xst)

9-25

The LMI Lab

ans =
6.5390e 05

9-26

From Decision to Matrix Variables and Vice Versa

From Decision to Matrix Variables and Vice Versa

While LMIs are specified in terms of their matrix variables X1, . . . , XK , the
LMI solvers optimize the vector x of free scalar entries of these matrices, called
the decision variables. The two functions mat2dec and dec2mat perform the
conversion between these two descriptions of the problem variables.
Consider an LMI system with three matrix variables X1, X2, X3. Given
particular values X1, X2, X3 of these variables, the corresponding value xdec of
the vector of decision variables is returned by mat2dec:
xdec = mat2dec(LMISYS,X1,X2,X3)

An error is issued if the number of arguments following LMISYS differs from the
number of matrix variables in the problem (see matnbr).
Conversely, given a value xdec of the vector of decision variables, the
corresponding value of the k-th matrix is given by dec2mat. For instance, the
value X2 of the second matrix variable is extracted from xdec by
X2 = dec2mat(LMISYS,xdec,2)

The last argument indicates that the second matrix variable is requested. It
could be set to the matrix variable identifier returned by lmivar.
The total numbers of matrix variables and decision variables are returned by
matnbr and decnbr, respectively. In addition, the function decinfo provides
precise information about the mapping between decision variables and matrix
variable entries (see the function reference pages).

9-27

The LMI Lab

Validating Results
The LMI Lab offers two functions to analyze and validate the results of an LMI
optimization. The function evallmi evaluates all variable terms in an LMI
system for a given value of the vector of decision variables, for instance, the
feasible or optimal vector returned by the LMI solvers. Once this evaluation is
performed, the left- and right-hand sides of a particular LMI are returned by
showlmi.
In the LMI problem considered in Example 8.2 on page 9-22, you can verify
that the minimizer xopt returned by mincx satisfies the LMI constraint [9-10]
as follows:
evlmi = evallmi(LMIs,xopt)
[lhs,rhs] = showlmi(evlmi,1)

The first command evaluates the system for the value xopt of the decision
variables, and the second command returns the left- and right-hand sides of
the first (and only) LMI. The negative definiteness of this LMI is checked by
eig(lhs-rhs)
ans =
2.0387e 04
3.9333e 05
1.8917e 07
4.6680e+01

9-28

Modifying a System of LMIs

Modifying a System of LMIs

Once specified, a system of LMIs can be modified in several ways with the
functions dellmi, delmvar, and setmvar.

Deleting an LMI
The first possibility is to remove an entire LMI from the system with dellmi.
For instance, suppose that the LMI system of Example 8.1 on page 9-8 is
described in LMISYS and that we want to remove the positivity constraint on X.
This is done by
NEWSYS = dellmi(LMISYS,2)

where the second argument specifies deletion of the second LMI. The resulting
system of two LMIs is returned in NEWSYS.
The LMI identifiers (initial ranking of the LMI in the LMI system) are not
altered by deletions. As a result, the last LMI
S>I
remains known as the third LMI even though it now ranks second in the
modified system. To avoid confusion, it is safer to refer to LMIs via the
identifiers returned by newlmi. If BRL, Xpos, and Slmi are the identifiers
attached to the three LMIs [9-6][9-8], Slmi keeps pointing to S > I even after
deleting the second LMI by
NEWSYS = dellmi(LMISYS,Xpos)

Deleting a Matrix Variable

Another way of modifying an LMI system is to delete a matrix variable, that is,
to remove all variable terms involving this matrix variable. This operation is
performed by delmvar. For instance, consider the LMI
ATX + XA + BW + WTBT + I < 0
with variables X = XT R44 and W R24. This LMI is defined by
setlmis([])
X = lmivar(1,[4 1]) % X
W = lmivar(2,[2 4]) % W

9-29

The LMI Lab

lmiterm([1 1 1 X],1,A,'s')
lmiterm([1 1 1 W],B,1,'s')
lmiterm([1 1 1 0],1)
LMISYS = getlmis

To delete the variable W, type the command

NEWSYS = delmvar(LMISYS,W)

The resulting NEWSYS now describes the Lyapunov inequality

ATX + XA + I < 0
Note that delmvar automatically removes all LMIs that depended only on the
deleted matrix variable.
The matrix variable identifiers are not affected by deletions and continue to
point to the same matrix variable. For subsequent manipulations, it is
therefore advisable to refer to the remaining variables through their identifier.
Finally, note that deleting a matrix variable is equivalent to setting it to the
zero matrix of the same dimensions with setmvar.

Instantiating a Matrix Variable

The function setmvar is used to set a matrix variable to some given value. As
a result, this variable is removed from the problem and all terms involving it
become constant terms. This is useful, for instance, to fixsetmvar some
variables and optimize with respect to the remaining ones.
Consider again Example 8.1 on page 9-8 and suppose we want to know if the
peak gain of G itself is less than one, that is, if
||G|| < 1
This amounts to setting the scaling matrix D (or equivalently, S = DTD) to a
multiple of the identity matrix. Keeping in mind the constraint S > I, a
legitimate choice is S = 2I. To set S to this value, enter
NEWSYS = setmvar(LMISYS,S,2)

The second argument is the variable identifier S, and the third argument is the
value to which S should be set. Here the value 2 is shorthand for 2byI. The
resulting system NEWSYS reads

9-30

Modifying a System of LMIs

A T X + XA + 2C T C XB

<0

T
2I
B X

X>0
2I > I
Note that the last LMI is now free of variable and trivially satisfied. It could,
therefore, be deleted by
NEWSYS = dellmi(NEWSYS,3)

or
NEWSYS = dellmi(NEWSYS,Slmi)

if Slmi is the identifier returned by newlmi.

9-31

The LMI Lab

Advanced Topics
This last section gives a few hints for making the most out of the LMI Lab. It
is directed toward users who are comfortable with the basics described above.

Structured Matrix Variables

Fairly complex matrix variable structures and interdependencies can be
specified with lmivar. Recall that the symmetric block-diagonal or rectangular
structures are covered by Types 1 and 2 of lmivar provided that the matrix
variables are independent. To describe more complex structures or correlations
between variables, you must use Type 3 and specify each entry of the matrix
variables directly in terms of the free scalar variables of the problem (the
so-called decision variables).
With Type 3, each entry is specified as either 0 or xn where xn is the n-th
decision variable. The following examples illustrate how to specify nontrivial
matrix variable structures with lmivar. We first consider the case of
uncorrelated matrix variables.

Example 8.3
Suppose that the problem variables include a 3-by-3 symmetric matrix X and
a 3-by-3 symmetric Toeplitz matrix
y y y
1 2 3
Y = y2 y1 y2

y y y
3 2 1

The variable Y has three independent entries, hence involves three decision
variables. Since Y is independent of X, these decision variables should be
labeled n + 1, n + 2, n + 3 where n is the number of decision variables involved
in X. To retrieve this number, define the variable X (Type 1) by
setlmis([])
[X,n] = lmivar(1,[3 1])

The second output argument n gives the total number of decision variables
used so far (here n = 6). Given this number, Y can be defined by
Y = lmivar(3,n+[1 2 3;2 1 2;3 2 1])

9-32

Advanced Topics

or equivalently by
Y = lmivar(3,toeplitz(n+[1 2 3]))

where toeplitz is a standard MATLAB function. For verification purposes, we

can visualize the decision variable distributions in X and Y with decinfo:
lmis = getlmis
decinfo(lmis,X)
ans =
1
2
4

2
3
5

4
5
6

decinfo(lmis,Y)
ans =
7
8
9

8
7
8

9
8
7

The next example is a problem with interdependent matrix variables.

Example 8.4
Consider three matrix variables X, Y, Z with structure

X = x 0 ,
0 y

Y = z 0 ,
0 t

Z = 0 x
t 0

where x, y, z, t are independent scalar variables. To specify such a triple, first

define the two independent variables X and Y (both of Type 1) as follows:
setlmis([])
[X,n,sX] = lmivar(1,[1 0;1 0])
[Y,n,sY] = lmivar(1,[1 0;1 0])

The third output of lmivar gives the entry-wise dependence of X and Y on the
decision variables (x1, x2, x3, x4) := (x, y, z, t):
sX =
1
0

0
2

9-33

The LMI Lab

sY =
3
0

0
4

Using Type 3 of lmivar, you can now specify the structure of Z in terms of the
decision variables x1 = x and x4 = t:
[Z,n,sZ] = lmivar(3,[0

sX(1,1);sY(2,2) 0])

Since sX(1,1) refers to x1 while sY(2,2) refers to x4, this defines the variable
0 x

1
Z =
= 0 x
x 0
t 0
4

as confirmed by checking its entry-wise dependence on the decision variables:

sZ =
0

1
4

Complex-Valued LMIs
The LMI solvers are written for real-valued matrices and cannot directly
handle LMI problems involving complex-valued matrices. However,
complex-valued LMIs can be turned into real-valued LMIs by observing that a
complex Hermitian matrix L(x) satisfies
L(x) < 0
if and only if
Re ( L ( x ) ) Im ( L ( x ) )

<0
Im ( L ( x ) ) Re ( L ( x ) )
This suggests the following systematic procedure for turning complex LMIs
into real ones:
Decompose every complex matrix variable X as

9-34

Advanced Topics

X = X1 + jX2
where X1 and X2 are real
Decompose every complex matrix coefficient A as
A = A1 + jA2
where A1 and A2 are real
Carry out all complex matrix products. This yields affine expressions in X1,
X2 for the real and imaginary parts of each LMI, and an equivalent
real-valued LMI is readily derived from the above observation.
For LMIs without outer factor, a streamlined version of this procedure consists
of replacing any occurrence of the matrix variable X = X1 + jX2 by
X
1

X 2
X 1

and any fixed matrix A = A1 + jA2, including real ones, by

A A
1 2

A A

2 1

For instance, the real counterpart of the LMI system

H

M XM < X,

X = X >I

(9-11)

reads (given the decompositions M = M1 + jM2 and X = X1 + jX2 with Mj, Xj real):
T

M M X X M M X X
1
2
1
2
1
2
1
2

<
M M X X M M X X
2
1
2 1
2
1
2 1

X X
1
2

X X < I
2 1

Note that X = XH in turn requires that X 1 = X 1 and X 2 + X 2 = 0 .

Consequently, X1 and X2 should be declared as symmetric and skewsymmetric matrix variables, respectively.

9-35

The LMI Lab

Assuming, for instance, that M C55, the LMI system [9-11] would be
specified as follows:
M1=real(M), M2=imag(M)
bigM=[M1 M2;-M2 M1]
setlmis([])
% declare bigX=[X1 X2;-X2 X1] with X1=X1' and X2+X2'=0:
[X1,n1,sX1] = lmivar(1,[5 1])
[X2,n2,sX2] = lmivar(3,skewdec(5,n1))
bigX = lmivar(3,[sX1 sX2;-sX2 sX1])
% describe the real counterpart of (1.12):
lmiterm([1 1 1 0],1)
lmiterm([ 1 1 1 bigX],1,1)
lmiterm([2 1 1 bigX],bigM',bigM)
lmiterm([ 2 1 1 bigX],1,1)
lmis = getlmis

Note the three-step declaration of the structured matrix variable bigX,

X X
1
2
bigX =

X
X
2 1

1 Specify X1 as a (real) symmetric matrix variable and save its structure

description sX1 as well as the number n1 of decision variables used in X1.

2 Specify X2 as a skew-symmetric matrix variable using Type 3 of lmivar and

the utility skewdec. The command skewdec(5,n1) creates a 5-by5

skew-symmetric structure depending on the decision variables n1 + 1,
n1 + 2,...
3 Define the structure of bigX in terms of the structures sX1 and sX2 of X1 and

X2 .

9-36

Advanced Topics

See Structured Matrix Variables on page 9-32 for more details on such
structure manipulations.

Specifying cTx Objectives for mincx

The LMI solver mincx minimizes linear objectives of the form cTx where x is the
vector of decision variables. In most control problems, however, such objectives
are expressed in terms of the matrix variables rather than of x. Examples
include Trace(X) where X is a symmetric matrix variable, or uTXu where u is a
given vector.
The function defcx facilitates the derivation of the c vector when the objective
is an affine function of the matrix variables. For the sake of illustration,
consider the linear objective
T

Trace(X) + x 0 Px0
where X and P are two symmetric variables and x0 is a given vector. If lmisys
is the internal representation of the LMI system and if x0, X, P have been
declared by
x0 = [1;1]
setlmis([])
X = lmivar(1,[3 0])
P = lmivar(1,[2 1])
:
:
lmisys = getlmis
T

the c vector such that cTx = Trace(X) + x 0 Px0 can be computed as follows:
n = decnbr(lmisys)
c = zeros(n,1)
for j=1:n,
[Xj,Pj] = defcx(lmisys,j,X,P)
c(j) = trace(Xj) + x0'*Pj*x0
end

The first command returns the number of decision variables in the problem and
the second command dimensions c accordingly. Then the for loop performs the
following operations:

9-37

The LMI Lab

1 Evaluate the matrix variables X and P when all entries of the decision vector

x are set to zero except xj:= 1. This operation is performed by the function
defcx. Apart from lmisys and j, the inputs of defcx are the identifiers X and
P of the variables involved in the objective, and the outputs Xj and Pj are the
corresponding values.
2 Evaluate the objective expression for X:= Xj and P:= Pj. This yields the j-th

entry of c by definition.
In our example the result is
c =
3
1
2
1

Other objectives are handled similarly by editing the following generic

skeleton:
n = decnbr( LMI system )
c = zeros(n,1)
for j=1:n,
[ matrix values ] = defcx( LMI system,j,
matrix identifiers)
c(j) = objective(matrix values)
end

Feasibility Radius
When solving LMI problems with feasp, mincx, or gevp, it is possible to
constrain the solution x to lie in the ball
xTx < R2
where R > 0 is called the feasibility radius. This specifies a maximum
(Euclidean norm) magnitude for x and avoids getting solutions of very large
norm. This may also speed up computations and improve numerical stability.
Finally, the feasibility radius bound regularizes problems with redundant
variable sets. In rough terms, the set of scalar variables is redundant when an
equivalent problem could be formulated with a smaller number of variables.

9-38

Advanced Topics

The feasibility radius R is set by the third entry of the options vector of the LMI
solvers. Its default value is R = 109. Setting R to a negative value means no
rigid bound, in which case the feasibility radius is increased during the
optimization if necessary. This flexible bound mode may yield solutions of
large norms.

Well-Posedness Issues
The LMI solvers used in the LMI Lab are based on interior-point optimization
techniques. To compute feasible solutions, such techniques require that the
system of LMI constraints be strictly feasible, that is, the feasible set has a
nonempty interior. As a result, these solvers may encounter difficulty when the
LMI constraints are feasible but not strictly feasible, that is, when the LMI
L(x) 0
has solutions while
L(x) < 0
has no solution.
For feasibility problems, this difficulty is automatically circumvented by
feasp, which reformulates the problem

Find x such that L ( x ) 0

(9-12)

Minimize t subject to Lx < t I.

(9-13)

as

In this modified problem, the LMI constraint is always strictly feasible in x, t

and the original LMI Equation (9-12) is feasible if and only if the global
minimum tmin of Equation (9-12) satisfies
tmin 0
For feasible but not strictly feasible problems, however, the computational
effort is typically higher as feasp strives to approach the global optimum tmin
= 0 to a high accuracy.
For the LMI problems addressed by mincx and gevp, nonstrict feasibility
generally causes the solvers to fail and to return an infeasibility diagnosis.
Although there is no universal remedy for this difficulty, it is sometimes

9-39

The LMI Lab

possible to eliminate underlying algebraic constraints to obtain a strictly

feasible problem with fewer variables.
Another issue has to do with homogeneous feasibility problems such as
ATP + P A < 0,

P>0

While this problem is technically well-posed, the LMI optimization is likely to

produce solutions close to zero (the trivial solution of the nonstrict problem). To
compute a nontrivial Lyapunov matrix and easily differentiate between
feasibility and infeasibility, replace the constraint P > 0-by-P > I with > 0.
Note that this does not alter the problem due to its homogeneous nature.

Semi-Definite B(x) in gevp Problems

Consider the generalized eigenvalue minimization problem
Minimize subject to A ( x ) < B ( x ), B ( x ) > 0, C ( x ) < 0.

(9-14)

Technically, the positivity of B(x) for some x Rn is required for the

well-posedness of the problem and the applicability of polynomial-time
interior-point methods. Hence problems where
B (x) 0
B(x) = 1
, with B 1 ( x ) > 0 strictly feasible

0 0
cannot be directly solved with gevp. A simple remedy consists of replacing the
constraints
A(x) < B(x),

B(x) > 0

by

A(x) < Y 0 ,
0 0

Y < B 1 ( x ),

B1 ( x ) > 0

where Y is an additional symmetric variable of proper dimensions. The

resulting problem is equivalent to Equation (9-14) and can be solved directly
with gevp.

9-40

Advanced Topics

Efficiency and Complexity Issues

As explained in Introduction on page 9-2, the term-oriented description of
LMIs used in the LMI Lab typically leads to higher efficiency than the
canonical representation
A0 + x1 A1 + + xN AN < 0

(9-15)

This is no longer true, however, when the number of variable terms is nearly
equal to or greater than the number N of decision variables in the problem. If
your LMI problem has few free scalar variables but many terms in each LMI,
it is therefore preferable to rewrite it as (9-15) and to specify it in this form.
Each scalar variable xj is then declared independently and the LMI terms are
of the form xjAj.
If M denotes the total row size of the LMI system and N the total number of
scalar decision variables, the flop count per iteration for the feasp and mincx
solvers is proportional to
N3 when the least-squares problem is solved via. Cholesly factorization of the
Hessian matrix (default) [2]
M-by-N2 when numerical instabilities warrant the use of QR factorization
instead
While the theory guarantees a worst-case iteration count proportional to M, the
number of iterations actually performed grows slowly with M in most
problems. Finally, while feasp and mincx are comparable in complexity, gevp
typically demands more computational effort. Make sure that your LMI
problem cannot be solved with mincx before using gevp.

Solving M + PTXQ + QTXTP < 0

In many output-feedback synthesis problems, the design can be performed in
two steps:
1 Compute a closed-loop Lyapunov function via LMI optimization.

9-41

The LMI Lab

2 Given this Lyapunov function, derive the controller state-space matrices by

solving an LMI of the form

T

M + P XQ + Q X P < 0
where M, P, Q are given matrices and X is an unstructured m-by-n matrix
variable.
It turns out that a particular solution Xc of (9-16) can be computed via simple
linear algebra manipulations [1]. Typically, Xc corresponds to the center of the
ellipsoid of matrices defined by Equation ().
The function basiclmi returns the explicit solution Xc:
Xc = basiclmi(M,P,Q)

Since this central solution sometimes has large norm, basiclmi also offers the
option of computing an approximate least-norm solution of Equation (). This is
done by
X = basiclmi(M,P,Q,'Xmin')

and involves LMI optimization to minimize ||X||.

9-42

References

References
[1] Gahinet, P., and P. Apkarian, A Linear Matrix Inequality Approach to H
Control, Int. J. Robust and Nonlinear Contr., 4 (1994), pp. 421-448.
[2] Nemirovski, A., and P. Gahinet, The Projective Method for Solving Linear
Matrix Inequalities, Proc. Amer. Contr. Conf., 1994, pp. 840-844.
[3] Nesterov, Yu, and A. Nemirovski, Interior Point Polynomial Methods in
Convex Programming: Theory and Applications, SIAM Books, Philadelphia,
1994.
[4] Shamma, J.S., Robustness Analysis for Time-Varying Systems, Proc.
Conf. Dec. Contr., 1992, pp. 3163-3168.

9-43

The LMI Lab

9-44

10
Function Reference

Functions By Category (p. 10-2)

Lists the Robust Control Toolbox functions according to

their purpose.

Functions Alphabetical List

(p. 10-15)

Lists the Robust Control Toolbox functions

alphabetically.

10

Function Reference

Functions By Category
Uncertain Elements on
page 10-4

Functions for building uncertain elements

Uncertain Matrices and

Systems on page 10-4

Functions for building uncertain matrices

and systems

Manipulation of Uncertain
Models on page 10-4

Functions for transforming and analyzing

uncertain models

Interconnection of Uncertain
Models on page 10-6

Functions for connecting uncertain models

Model Order Reduction on

page 10-6

Functions for generating low-order

approximations to plant and controller
models

Robustness and Worst-Case

Analysis on page 10-7

Functions for characterizing system

robustness and worst-case performance

Functions for analyzing P-Systems

Robustness Analysis for
Parameter-Dependent Systems
(P-Systems) on page 10-8

10-2

Controller Synthesis on
page 10-8

H control design functions

m-Synthesis on page 10-10

Structured singular value control design

functions

Sampled-Data Systems on
page 10-10

Functions for analyzing sampled-data

systems

Gain Scheduling on
page 10-10

Functions for synethsizing gain-scheduled

controllers

Frequency-Response Data
(FRD) Models on page 10-10

Functions for operating on FRD models

Functions By Category

Supporting Utilities on
page 10-12

Additional functions for working with

systems containing uncertain elements

LMIs on page 10-12

Functions for building and solving systems

of Linear Matrix Inequalities

10-3

10

Function Reference

Uncertain Elements
ucomplex

Create uncertain complex parameter

ucomplexm

Create uncertain complex matrix

udyn

Create unstructured uncertain dynamic system

object

ultidyn

Create uncertain linear time-invariant object

ureal

Create uncertain real parameter

Uncertain Matrices and Systems

diag

Diagonalize vector of uncertain matrices and

systems

randatom

Generate random uncertain atom objects

randumat

Create random uncertain matrix

randuss

Create random uncertain state-space model

ufrd

Create uncertain frequency response data

(ufrd) object, or convert another model type to
ufrd model

umat

Create uncertain matrix

uss

Specify uncertain state-space models or convert

LTI model to uncertain state-space model

Manipulation of Uncertain Models

10-4

actual2normalized

Calculate normalized distance between

nominal value and given value for uncertain
atom

gridureal

Grid ureal parameters uniformly over range

isuncertain

True for uncertain systems

lftdata

Decompose uncertain object into fixed

normalized and fixed uncertain parts

Functions By Category

normalized2actual

Convert value for atom in normalized

coordinates to corresponding actual value

repmat

Replicate and tile uncertain matrix

simplify

Simplify representation of uncertain object

squeeze

Remove singleton dimensions for umat objects

usample

Generate random samples of uncertain object

usubs

Substitute values for uncertain elements of

uncertain objects

uss/ssbal

Scale state/uncertainty of uncertain system

10-5

10

Function Reference

Interconnection of Uncertain Models

iconnect

Create empty interconnection object

icsignal

Create icsignal object

imp2exp

Convert implicit linear relationship to explicit

input-output relationship

stack

Construct array by stacking uncertain

matrices, models, or arrays

sysic

Build interconnections of certain and uncertain

matrices and systems

Model Order Reduction

10-6

balancmr

Balanced model truncation via square root

method

bstmr

Balanced stochastic model truncation (BST) via

Schur method

hankelmr

Hankel minimum degree approximation (MDA)

without balancing

hankelsv

Compute Hankel singular values for

stable/unstable or continuous/discrete system

imp2ss

Convert impulse response to approximate

state-space realization

modreal

Modal form realization and projection

ncfmr

Balanced normalized coprime factor model

reduction

reduce

Model order reduction using Hankel singular

value based algorithms

schurmr

Balanced truncation via Schur method

slowfast

State-space slow-fast decomposition

stabproj

State-space stable/anti-stable decomposition

Functions By Category

Robustness and Worst-Case Analysis

cpmargin

Coprime stability margin of plant-controller

feedback loop

gapmetric

Compute upper bounds on gap and nugap

distances between two systems

loopmargin

Analyze feedback loop

loopsens

Analyze sensitivity of plant-controller feedback

loop

mussv

Compute bounds on structured singular value

()

mussvextract

Extract muinfo structure returned by mussv

ncfmargin

Calculate normalized coprime stability margin

of feedback loop

popov

Perform Popov robust stability test

robopt

Create options object for use with robuststab

and robustperf

robustperf

Calculate robust performance margin of

uncertain multivariable system

robuststab

Calculate robust stability margin of uncertain

multivariable system

wcgain

Calculate bounds on worst-case gain of

uncertain system

wcgopt

Create options object for use with wcgain,

wcsens, and wcmargin

wcmargin

Calculate worst-case gain/phase margins for

feedback loop

wcnorm

Calculate worst-case norm of uncertain matrix

wcsens

Calculate worst-case sensitivity and

complementary sensitivity functions of
feedback loop

10-7

10

Function Reference

Robustness Analysis for Parameter-Dependent

Systems (P-Systems)
aff2pol

Convert affine parameter-dependent models to

polytopic models

decay

Quadratic decay rate of polytopic or affine

P-systems

ispsys

True for parameter-dependent systems

pdlstab

Assess robust stability of polytopic or

parameter-dependent system

pdsimul

Time response of parameter-dependent system

along parameter trajectory

polydec

Compute polytopic coordinates with respect to

box corners

psinfo

Query characteristics of polytopic or

parameter-dependent systems

psys

Specify polytopic or parameter-dependent

linear system

pvec

Specify vector of uncertain or time-varying

parameters

pvinfo

Query characteristics of parameter vector

quadperf

Compute quadratic H performance of

polytopic or parameter-dependent system

quadstab

Assess quadratic stability of polytopic or affine

parameter-dependent systems

Controller Synthesis

10-8

augw

Create augmented plant model for weighted

mixed-sensitivity loop shaping design

h2hinfsyn

Mixed H2/H synthesis with pole placement

constraints

Functions By Category

h2syn

H2 controller synthesis for LTI plant

hinfsyn

H optimal controller synthesis for LTI plant

loopsyn

H loop shaping controller synthesis

ltrsyn

LQG loop-transfer recovery controller synthesis

mkfilter

Generate Bessel, Butterworth, Chebyshev, or

RC filter

mixsyn
ncfsyn

H mixed-sensitivity controller synthesis

H normalized coprime factor controller
synthesis

10-9

10

Function Reference

-Synthesis
cmsclsyn

Constant matrix -synthesis

dksyn

Synthesis of robust controller via -synthesis

D-K iteration

dkitopt

Create options object for use with dksyn

drawmag

Interactive mouse-based sketching and fitting

tool

fitfrd

Fit D-scaling frequency response data with

state-space model

fitmagfrd

Fit scaling magnitude data with stable,

minimum-phase state-space model

genphase

Fit single-input/single-output magnitude data

with real, rational, minimum-phase transfer
function

Sampled-Data Systems
sdhinfnorm
sdhinfsyn
sdlsim

Compute L2 norm of sampled-data system

Sampled-data H controller synthesis
Time response of sampled-data feedback
system

Gain Scheduling
hinfgs

Gain-scheduled H controller synthesis

Frequency-Response Data (FRD) Models

10-10

frd/loglog

Log-log scale plot of frd objects

frd/semilogx

Semilog scale plot of frd object

frd/semilogy

Semilog scale plot of frd object

frd/rcond

Reciprocal condition estimator of frd object

Functions By Category

frd/schur

Schur decomposition of frd object

frd/svd

Singular value decomposition of frd object

10-11

10

Function Reference

Supporting Utilities
bilin

Multivariable binlinear transform of frequency

(s or z)

dmplot

Interpret disk gain and phase margins

mktito

Partition LTI system into two-input-two-output

system

sectf

State-space sector bilinear transformation

skewdec

Create skew-symmetric matrix

symdec

Form symmetric matrix

LMIs
LMI Systems on page 10-12

Functions for specifying systems of

LMIs

LMI Characteristics on
page 10-14

Functions for extracting information

about systems of LMIs

LMI Solvers on page 10-14

Functions for solving systems of LMIs

Validation of Results on
page 10-14

Functions for analyzing and

validating LMI optimization results

Modification of Systems of
LMIs on page 10-14

Functions for altering systems of

LMIs

LMI Systems

10-12

getlmis

Internal description of LMI system

lmiedit

Specify or display systems of LMIs as MATLAB

expressions

lmiterm

Add new term to existing LMI

lmivar

Specify matrix variables in existing LMI

system

Functions By Category

newlmi

Add new LMI to LMI system

setlmis

Initialize description of LMI system

10-13

10

Function Reference

LMI Characteristics
dec2mat

Extract matrix variable value from vector of

decision variables

decinfo

Describe how entries of matrix variable X

relate to decision variables

decnbr

Number of decision variables in LMI system

lmiinfo

Information about existing system of LMIs

lminbr

Number of LMIs in LMI system

mat2dec

Extract vector of decision variables from matrix

variable values

matnbr

Number of matrix variables in LMI system

LMI Solvers
defcx

Specify cTx objectives of mincx

feasp

Compute solution to given system of LMIs

gevp

Generalized eigenvalue minimization under

LMI constraints

mincx

Minimize linear objective under LMI

constraints

Validation of Results
evallmi

Evaluate LMI for given values of decision

variables

showlmi

Left - and right-hand sides of evalauted LMI

Modification of Systems of LMIs

10-14

dellmi

Remove LMI from system of LMIs

delmvar

Remove matrix variable from LMI problem

setmvar

Instantiate matrix variable and evaluate LMI

terms

Functions Alphabetical List

Functions Alphabetical List

10

The following pages comprise the function reference pages. The functions are
listed alphabetically.

10-15

actual2normalized

Purpose

10actual2normalized

Calculate normalized distance between nominal value and given value for
uncertain atom

Syntax

ndist = actual2normalized(A,V)

Description

ndist = actual2normalized(A,V) is the normalized distance between the

nominal value of the uncertain atom A and the given value V.

If V is an array of values, then ndist is an array of normalized distances.

The robustness margins computed in robuststab and robustperf serve as
bounds for the normalized distances in ndist. For example, if an uncertain
system has a stability margin of 1.4, this system is stable when the normalized
distance of the uncertain element values from the nominal is less than 1.4.

Examples

For uncertain real parameters whose range is symmetric about their nominal
value, the normalized distance is intuitive.
Create uncertain real parameters with a range that is symmetric about the
nominal value, where each end point is 1 unit from the nominal. Points that lie
inside the range are less than 1 unit from the nominal, while points that lie
outside the range are greater than 1 unit from the nominal.
a = ureal('a',3,'range',[1 5]);
actual2normalized(a,[1 3 5])
ans =
-1.0000
-0.0000
1.0000
actual2normalized(a,[2 4])
ans =
-0.5000
0.5000
actual2normalized(a,[0 6])
ans =
-1.5000
1.5000

Plot the normalized distance for several values. Note that the relationship
between a normalized distance and a numerical difference is linear.
values = linspace(-3,9,250);
ndist = actual2normalized(a,values);
plot(values,ndist)

10-16

actual2normalized

3
4

2
Values

Next, create an asymmetric parameter, where each endpoint is 1 unit from the
nominal. Points that lie inside the range are less than 1 unit from the nominal,
while points that lie outside the range are greater than 1 unit from the
nominal.
au = ureal('a',4,'range',[1 5]);
actual2normalized(a,[1 4 5])
ans =
-1.0000
0.5000
1.0000
actual2normalized(a,[2 4.5])
ans =
-0.5000
0.7500
actual2normalized(a,[0 6])
ans =
-1.5000
1.5000

Plot the normalized distance for several values. Note that the relationship
between normalized distance and numerical difference is nonlinear.
ndistu = actual2normalized(au,values);

10-17

actual2normalized

plot(values,ndistu)
xlim([-4 10])
ylim([0 4])

2
3
4
4

2
Values

See Also

10-18

robuststab
robustperf

Calculate robust stability margin

Calculate robust performance margin

aff2pol

Purpose

10aff2pol

Convert affine parameter-dependent models to polytopic models

Syntax

polsys = aff2pol(affsys)

Description

aff2pol derives a polytopic representation polsys of the affine

parameter-dependent system
E ( p )x = A ( p )x + B ( p )u

(10-1)

y = C ( p )x + D ( p )u

(10-2)

where p = (p1, . . ., pn) is a vector of uncertain or time-varying real parameters

taking values in a box or a polytope. The description affsys of this system
should be specified with psys.
The vertex systems of polsys are the instances of (10-1) and (10-2) at the
vertices pex of the parameter range, i.e., the SYSTEM matrices
A ( p ) + jE ( p ) B ( p )
ex
ex
ex

D ( p ex )
C ( p ex )

for all corners pex of the parameter box or all vertices pex of the polytope of
parameter values.

See Also

psys
pvec
uss

Specification of uncertain state-space model

Quantification of uncertainty on physical parameters
Create an uncertain state-space model

10-19

augw

Purpose

10augw

Create augmented plant model for weighted mixed-sensitivity loop shaping

design

Syntax

P = AUGW(G,W1,W2,W3)

Description

P = AUGW(G,W1,W2,W3) computes a state-space model of an augmented LTI

plant P(s) with weighting functions W1(s), W2(s), and W3(s) penalizing the error
signal, control signal and output signal respectively (see block diagram) so that
the closed-loop transfer function matrix is the weighted mixed sensitivity

W 1S

Ty u = W2 R
1 1
W3 T
where S, R and T are given by
S = ( I + GK )

R = K ( I + GK )

T = GK ( I + GK )

The LTI systems S and T are called the sensitivity and complementary
sensitivity, respectively.

10-20

augw

AUGMENTED PLANT P(s)

W1
W2
+
u
1
u

W3

2
y
2

CONTROLLER
F(s)
K(s)

Figure 10-1: Plant Augmentation

For dimensional compatibility, each of the three weights W1, W2 and W3 must
be either empty, a scalar (SISO) or have respective input dimensions NY, NU,
and NY where G is NY-by-NU. If one of the weights is not needed, you may
simply assign an empty matrix [ ]; e.g., P = AUGW(G,W1,[],W3) is P(s) as in
Algorithm, but without the second row (without the row containing W2).

Algorithm

The augmented plant P(s) produced by is

P(s) =

W1

W1G

W2

W3G

Partitioning is embedded via P=mktito(P,NY,NU), which sets the InputGroup

and OutputGroup properties of P as follows
[r,c]=size(P);
P.InputGroup = struct('U1',1:c-NU,'U2',c-NU+1:c);
P.OutputGroup = struct('Y1',1:r-NY,'Y2',r-NY+1:r);

Examples

s=zpk('s'); G=(s-1)/(s+1);
W1=0.1*(s+100)/(100*s+1); W2=0.1; W3=[];

10-21

augw

P=augw(G,W1,W2,W3);
[K,CL,GAM]=hinfsyn(P); [K2,CL2,GAM2]=h2syn(P);
L=G*K; S=inv(1+L); T=1-S;
sigma(S,'k',GAM/W1,'k-.',T,'r',GAM*G/W2,'r-.')
legend('S = 1/(1+L)','GAM/W1', 'T=L/(1+L)','GAM*G/W2',2)
Singular Values

60

40

S = 1/(1+L)
GAM/W1
T=L/(1+L)
GAM*G/W2

Singular Values (dB)

20

20

40

60

80
4
10

10

10

10
Frequency (rad/sec)

10

10

Limitations

The transfer functions G, W1, W2 and W3 must be proper, i.e., bounded as

s or, in the discrete-time case, as z . Additionally, W1, W2 and W3
should be stable. The plant G should be stabilizable and detectable; else, P will
not be stabilizable by any K.

See Also

h2syn
hinfsyn
mixsyn
mktito

10-22

H2 controller synthesis
H controller synthesis
H mixed-sensitivity controller synthesis
Partition LTI system via Input and Output Groups

balancmr

Purpose

10balancmr

Balanced model truncation via square root method

Syntax

GRED = balancmr(G)
GRED = balancmr(G,order)
[GRED,redinfo] = balancmr(G,key1,value1,...)
[GRED,redinfo] = balancmr(G,order,key1,value1,...)

Description

balancmr returns a reduced order model GRED of G and a struct array redinfo

containing the error bound of the reduced model and Hankel singular values of
the original system.
The error bound is computed based onHankel singular values of G. For a stable
system these values indicate the respective state energy of the system. Hence,
reduced order can be directly determined by examining the system Hankel
SVs, .
With only one input argument G, the function will show a Hankel singular
value plot of the original model and prompt for model order number to reduce.
This method guarantees an error bound on the infinity norm of the additive
error || G-GRED || for well-conditioned model reduced problems [1]:
n

G Gred 2

i
k+1

This table describes input arguments for balancmr.

Argument

Description

LTI model to be reduced (without any other inputs will plot

its Hankel singular values and prompt for reduced order)

ORDER

(Optional) an integer for the desired order of the reduced

model, or optionally a vector packed with desired orders for
batch runs

A batch run of a serial of different reduced order models can be generated by

specifying order = x:y, or a vector of positive integers. By default, all the

10-23

balancmr

anti-stable part of a system is kept, because from control stability point of view,
getting rid of unstable state(s) is dangerous to model a system.
'MaxError' can be specified in the same fashion as an alternative for 'Order'.
In this case, reduced order will be determined when the sum of the tails of the
Hankel SV's reaches the 'MaxError'.
This table lists the input arguments 'key' and its 'value'.
Argument

Value

Description

'MaxError'

Real number or
vector of
different errors

Reduce to achieve H error. When

present, 'MaxError'overides ORDER
input.

'Weights'

{Wout,Win} cell

Optimal 1-by-2 cell array of LTI weights

Wout (output) and Win (input). Defaults
are both identity. Weights must be
invertible.

array

'Display'

'on' or 'off'

Display Hankel singular plots (default

'off').

'Order'

Integer, vector
or cell array

Order of reduced model. Use only if not

specified as 2nd argument.

Weights on the original model input and/or output can make the model
reduction algorithm focus on some frequency range of interests. But weights
have to be stable, minimum phase and invertible.

10-24

balancmr

This table describes output arguments.

Argument

Description

GRED

LTI reduced order model. Becomes multidimensional array

when input is a serial of different model order array

REDINFO

A STRUCT array with three fields:

REDINFO.ErrorBound (bound on || G-GRED ||)
REDINFO.StabSV (Hankel SV of stable part of G)
REDINFO.UnstabSV (Hankel SV of unstable part of G)

G can be stable or unstable, continuous or discrete.

Algorithm

Given a state-space (A,B,C,D) of a system and k, the desired reduced order, the
following steps will produce a similarity transformation to truncate the
original state-space system to the kth order reduced model.
1 Find the SVD of the controllability and observability grammians

P = Up p VpT
Q = U q q V qT
2 Find the square root of the grammians (left/right eigenvectors)

Lp = Up p1/2
Lo = Uq q1/2
3 Find the SVD of (LoTLp)

Lo T Lp = U V T
4 Then the left and right transformation for the final kth order reduced model

is
SL,BIG = Lo U(:,1:k) (1;k,1:k))-1/2
SR,BIG = Lp V(:,1:k) (1;k,1:k))-1/2
5 Finally,

10-25

balancmr

S T L, BIG AS R, BIG S L, BIG B

CS R, BIG

The proof of the square root balance truncation algorithm can be found in [2].

Examples

Given a continuous or discrete, stable or unstable system, G, the following

commands can get a set of reduced order models based on your selections:
rand('state',1234); randn('state',5678);G = rss(30,5,4);
[g1, redinfo1] = balancmr(G); % display Hankel SV plot
% and prompt for order (try 15:20)
[g2, redinfo2] = balancmr(G,20);
[g3, redinfo3] = balancmr(G,[10:2:18]);
[g4, redinfo4] = balancmr(G,'MaxError',[0.01, 0.05]);
rand('state',12345); randn('state',6789);
wt1 = rss(6,5,5); wt1.d = eye(5)*2;
wt2 = rss(6,4,4); wt2.d = 2*eye(4);
[g5, redinfo5] = balancmr(G, [10:2:18], 'weight',{wt1,wt2});
for i = 1:5
figure(i); eval(['sigma(G,g' num2str(i) ');']);
end

References

[1] Glover, K., All Optimal Hankel Norm Approximation of Linear

Multivariable Systems, and Their L - error Bounds, Int. J. Control, Vol. 39,
No. 6, 1984, p. 1145-1193
[2] Safonov, M.G., and R.Y. Chiang, A Schur Method for Balanced Model
Reduction, IEEE Trans. on Automat. Contr., Vol. 34, No. 7, July 1989,
p. 729-733

See Also

reduce
schurmr
hankelmr
bstmr
ncfmr
hankelsv

10-26

Top-level model reduction function

Balanced model truncation via Schur method
Hankel minimum degree approximation
Balanced stochastic model truncation via Schur
method
Balanced model truncation for normalized coprime
factors
Hankel singular values

bilin

Purpose

10bilin

Multivariable bilinear transform of frequency (s or z)

Syntax

GT = bilin(G,VERS,METHOD,AUG)

Description

bilin computes the effect on a system of the frequency-variable substitution:

z +
s = ---------------z +
The variable VERS denotes the transformation direction:
VERS= 1, forward transform ( s z ) or ( s s ) .
VERS=-1, reverse transform ( z s ) or ( s s ) .

This transformation maps lines and circles to circles and lines in the complex
plane. People often use this transformation to do sampled-data control system
design [1] or, in general, to do shifting of j modes [2], [3], [4].
Bilin computes several state-space bilinear transformations such as backward
rectangular, etc., based on the METHOD you select.
Bilinear transform types
Method

Type of bilinear transform

'BwdRec'

Backward rectangular:
z1
s = -----------Tz
AUG = T, the sampling period.

'FwdRec'

Forward rectangular:
z1
s = -----------T
AUG = T, the sampling period.

10-27

bilin

Bilinear transform types

Method

Type of bilinear transform

'S_Tust'

Shifted Tustin:

2 z1
s = ---- -------------
T --z- + 1

h
AUG = [T h], is the shift coefficient.

'S_ftjw'

Shifted j-axis, bilinear pole-shifting, continuous-time

to continuous-time:
s + p 1
s = --------------------1 + s p 2
AUG = [p2 p1]

'G_Bilin'

METHOD = 'G_Bilin', general bilinear, continuous-time

to continuous-time:
s +
s = ---------------s +
AUG = [ ]

Examples

Example 1. Tustin continuous s-plane to discrete z-plane transforms

Consider the following continuous-time plant (sampled at 20 Hz):
A = 1 1 , B = 1 0 , C = 1 0 , D = 0 0 ; T s = 0.05
0 2
1 1
0 1
0 0
Following is an example of four common continuous to discrete bilin
transformations for the sampled plant:
A= [-1 1; 0 -2]; B=[1 0; 1 1];
C= [1 0; 0 1];
D=[0 0; 0 0];
sys = ss(A,B,C,D);
Ts=0.05; % sampling time

10-28

% ANALOG

bilin

[syst] = c2d(sys,Ts,'tustin');
% Tustin
[sysp] = c2d(sys,Ts,'prewarp',40);
% Pre-warped Tustin
[sysb] = bilin(sys,1,'BwdRec',Ts);
% Backward Rectangular
[sysf] = bilin(sys,1,'FwdRec',Ts);
% Forward Rectangular
w = logspace(-2,3,50); % frequencies to plot
sigma(sys,syst,sysp,sysb,sysf,w);

Figure 10-2: Comparison of Four Bilinear Transforms from Example 1

Example 2. Bilinear continuous to continuous pole-shifting 'S_ftjw'

Design an H mixed-sensitivity controller for the ACC Benchmark plant
1
G ( s ) = ------------------------2 2
s (s + 2)
such that all closed-loop poles lie inside a circle in the left half of the s-plane
whose diameter lies on between points [p1,p2]=[-12,-2]:
p1=-12; p2=-2; s=zpk('s');
G=ss(1/(s^2*(s^2+2)));
% original unshifted plant
Gt=bilin(G,1,'Sft_jw',[p1 p2]); % bilinear pole shifted plant Gt
Kt=mixsyn(Gt,1,[],1);
% bilinear pole shifted controller

10-29

bilin

K =bilin(Kt,-1,'Sft_jw',[p1 p2]); % final controller K

As shown in Figure 10-3, 'S_ftjw' final closed-loop poles are inside the left
[p1,p2] circle, on page 10-30, closed-loop poles are placed in the left circle [p1
p2]. The shifted plant, which has its non-stable poles shifted to the inside the
right circle, is
4

5
( s 12 )
G t ( s ) = 4.765 10 ---------------------------------------------------------------------------2 2
( s 2 ) ( s 4.274s + 5.918 )

Example of Bilinear Mapping: s~ = (s + p1) / (s/p2 1)

original plant poles (splane)
shifted plant poles (s~plane)
shifted HInf closedloop poles
final HInf closedloop poles

10
8
6
4
2
0

p2

p1

p1

p2

2
4
6
8
10
15

10

5
0
5
(p1 = 2, p2 = 12)

10

15

Figure 10-3: 'S_ftjw' final closed-loop poles are inside the left [p1,p2] circle

Algorithm

bilin employs the state-space formulae in [3]:

1
Ab Bb
A I ) ( I + A ) ----------------- = (-----------------------------------------------------Cb Db
1
C ( I A )

10-30

( ) ( I A ) B
----------------------------------------------------------1
D + C ( I A ) B

bilin

References

[1] Franklin, G.F., and J.D. Powell, Digital Control of Dynamics System,
Addison-Wesley, 1980.
[2] Safonov, M.G., R.Y. Chiang, and H. Flashner, H Control Synthesis for a
Large Space Structure, AIAA J. Guidance, Control and Dynamics, 14, 3, p.
513-520, May/June 1991.
[3] Safonov, M.G., Imaginary-Axis Zeros in Multivariable H Optimal
Control, in R.F. Curtain (editor), Modelling, Robustness and Sensitivity
Reduction in Control Systems, p. 71-81, Springer-Varlet, Berlin, 1987.
[4] Chiang, R.Y., and M.G. Safonov, H Synthesis using a Bilinear Pole
Shifting Transform, AIAA, J. Guidance, Control and Dynamics, vol. 15, no. 5,
p. 1111-1117, September-October 1992.

See Also

c2d
d2c
sectf

Convert from continous- to discrete-time

Convert from continous- to discrete-time
Sector transformation

10-31

bstmr

Purpose

10bstmr

Balanced stochastic model truncation (BST) via Schur method

Syntax

GRED = bstmr(G)
GRED = bstmr(G,order)
[GRED,redinfo] = bstmr(G,key1,value1,...)
[GRED,redinfo] = bstmr(G,order,key1,value1,...)

Description

bstmr returns a reduced order model GRED of G and a struct array redinfo

containing the error bound of the reduced model and Hankel singular values of
the phase matrix of the original system [2].
The error bound is computed based on Hankel singular values of the phase
matrix of G. For a stable system these values indicate the respective state
energy of the system. Hence, reduced order can be directly determined by
examining these values.
With only one input argument G, the function will show a Hankel singular
value plot of the phase matrix of G and prompt for model order number to
reduce.
This method guarantees an error bound on the infinity norm of the
multiplicative || GRED-1(G-GRED) || or relative error || G-1(G-GRED) || for
well-conditioned model reduction problems [1]:
n
1

G ( G Gred )

( 1 + 2i (

1 + i2 + i ) ) 1

k+1

This table describes input arguments for bstmr.

Argument

Description

LTI model to be reduced (without any other inputs will plot

its Hankel singular values and prompt for reduced order)

ORDER

(Optional) an integer for the desired order of the reduced

model, or a vector of desired orders for batch runs

A batch run of a serial of different reduced order models can be generated by

specifying order = x:y, or a vector of integers. By default, all the anti-stable

10-32

bstmr

part of a system is kept, because from control stability point of view, getting rid
of unstable state(s) is dangerous to model a system.
'MaxError' can be specified in the same fashion as an alternative for 'ORDER'.
In this case, reduced order will be determined when the accumulated product
of Hankel SVs shown in the above equation reaches the 'MaxError'.
Argument

Value

Description

'MaxError'

Real number or
vector of
different errors

Reduce to achieve H error.

When present, 'MaxError'overides
ORDER input.

'Display'

'on' or 'off'

Display Hankel singular plots (default

'off').

'Order'

Integer, vector
or cell array

Order of reduced model. Use only if not

specified as 2nd argument.

This table describes output arguments.

Argument

Description

GRED

LTI reduced order model. Become multi-dimension array

when input is a serial of different model order array.

REDINFO

A STRUCT array with three fields:

REDINFO.ErrorBound (bound on || G-1(G-GRED) ||)
REDINFO.StabSV (Hankel SV of stable part of G)
REDINFO.UnstabSV (Hankel SV of unstable part of G)

G can be stable or unstable, continuous or discrete.

Algorithm

Given a state-space (A,B,C,D) of a system and k, the desired reduced order, the
following steps will produce a similarity transformation to truncate the
original state-space system to the kth order reduced model:

10-33

bstmr

1 Find the controllability grammian P and observability grammian Q of the

left spectral factor = ()() = ()() by solving the following

Lyapunov and Riccati equations
AP + PAT + BBT = 0
BW = PCT + BDT
QA + AT Q + (QBW - CT) (-DDT) (QBW - CT)T = 0
2 Find the Schur decomposition for PQ in both ascending and descending

order, respectively,
T
V A PQV A

T
V D PQV D

1
=

0
0 0 n
n

0
0 0 1

3 Find the left/right orthonormal eigen-bases of PQ associated with the kth big

Hankel singular values of the all-pass phase matrix (W*(s))-1G(s).

k
V A = [V R, SMALL, V L, BIG ]

k
V D = [ V R, BIG ,V L, SMALL]
4 Find the SVD of (VTL,BIG VR,BIG) = U
5 Form the left/right transformation for the final kth order reduced model

SL,BIG = V L,BIG U(1:k,1:k)1/2

SR,BIG = VR,BIGV(1:k,1:k)1/2
6 Finally,

10-34

bstmr

S T L, BIG AS R, BIG S L, BIG B

CS R, BIG

The proof of the Schur BST algorithm can be found in [2].

Note The BST model reduction theory requires that the original model D
matrix be full rank, for otherwise the Riccati solver fails. For any problem
with strictly proper model, you can shift the j-axis via bilin such that
BST/REM approximation can be achieved up to a particular frequency range
of interests. Alternatively, you can attach a small but full rank D matrix to the
original problem but remove the D matrix of the reduced order model
afterwards. As long as the size of D matrix is insignificant inside the control
bandwidth, the reduced order model should be fairly close to the true model.
By default, the bstmr program will assign a full rank D matrix scaled by 0.001
of the minimum eigenvalue of the original model, if its D matrix is not full
rank to begin with. This serves the purpose for most problems if the user does
not want to go through the trouble of model pretransformation.

Examples

Given a continuous or discrete, stable or unstable system, G, the following

commands can get a set of reduced order models based on your selections:
rand('state',1234); randn('state',5678);
G = rss(30,5,4); G.d = zeros(5,4);
[g1, redinfo1] = bstmr(G); % display Hankel SV plot
% and prompt for order (try 15:20)
[g2, redinfo2] = bstmr(G,20);
[g3, redinfo3] = bstmr(G,[10:2:18]);
[g4, redinfo4] = bstmr(G,'MaxError',[0.01, 0.05]);
for i = 1:4
figure(i); eval(['sigma(G,g' num2str(i) ');']);
end

References

[1] Zhou, K., Frequency weighted L error bounds, Syst. Contr. Lett., Vol. 21,
115-125, 1993.

10-35

bstmr

[2] Safonov, M.G., and R.Y. Chiang, Model Reduction for Robust Control: A
Schur Relative Error Method, International J. of Adaptive Control and Signal
Processing, Vol. 2, p. 259-272, 1988.

See Also

10-36

reduce
balancmr
hankelmr
schurmr
ncfmr
hankelsv

Top level model reduction function

Balanced truncation via square-root method
Hankel minimum degree approximation
Balanced truncation via Schur method
Balanced truncation for normalized coprime factors
Hankel singular value

cmsclsyn

Purpose

Constant matrix -synthesis

10cmsclsyn

Syntax

[qopt,bnd]
[qopt,bnd]
[qopt,bnd]
[qopt,bnd]

Description

cmsclsyn approximately solves the constant-matrix, upper bound -synthesis

problem by minimization,

min

=
=
=
=

QC

cmsclsyn(R,U,V,BlockStructure);
cmsclsyn(R,U,V,BlockStructure,opt);
cmsclsyn(R,U,V,BlockStructure,opt,qinit);
cmsclsyn(R,U,V,BlockStructure,opt,'random',N)

rt

( R + UQV )

for given matrices R Cnm, U Cnr, V Ctm, and a set Cmn. This
applies to constant matrix data in R, U, and V.
[QOPT,BND] = cmsclsyn(R,U,V,BlockStructure) minimizes, by choice of Q.
QOPT is the optimum value of Q, the upper bound of mussv(R+U*Q*V,BLK), BND.
The matrices R,U and V are constant matrices of the appropriate dimension.
BlockStructure is a matrix specifying the perturbation blockstructure as
defined for mussv.
[QOPT,BND] = cmsclsyn(R,U,V,BlockStructure,OPT) uses the options
specified by OPT in the calls to mussv. See mussv for more information. The
default value for OPT is 'cUsw'.
[QOPT,BND] = cmsclsyn(R,U,V,BlockStructure,OPT,QINIT) initializes the
iterative computation from Q = QINIT. Because of the nonconvexity of the
overall problem, different starting points often yield different final answers. If
QINIT is an N-D array, then the iterative computation is performed multiple
times - the ith optimization is initialized at Q = QINIT(:,:,i). The output
arguments are associated with the best solution obtained in this brute force
approach.
[QOPT,BND] = cmsclsyn(R,U,V,BlockStructure,OPT,'random',N) initializes
the iterative computation from N random instances of QINIT. If NCU is the
number of columns of U, and NRV is the number of rows of V, then the
approximation to solving the constant matrix synthesis problem is two-fold:
only the upper bound for is minimized, and the minimization is not convex,
hence the optimum is generally not found. If U is full column rank, or V is full
row rank, then the problem can (and is) cast as a convex problem, [Packard,

10-37

cmsclsyn
Zhou, Pandey and Becker], and the global optimizer (for the upper bound for )
is calculated.

Algorithm

The cmsclsyn algorithm is iterative, alternatively holding Q fixed, and

computing the mussv upper bound, followed by holding the upper bound
multipliers fixed, and minimizing the bound implied by choice of Q. If U or V is
square and invertible, then the optimization is reformulated (exactly) as an
linear matrix inequality, and solved directly, without resorting to the iteration.

References

Packard, A.K., K. Zhou, P. Pandey, and G. Becker, A collection of robust

control problems leading to LMIs, 30th IEEE Conference on Decision and
Control, Brighton, UK, 1991, p. 12451250.

See Also

dksyn
hinfsyn
mussv
robuststab
robustperf

10-38

Synthesize a robust controller via D-K iteration

Synthesize a H controller
Calculate bounds on the Structured Singular Value ()
Calculate stability margins of uncertain systems
Calculate performance margins of uncertain systems

cpmargin

Purpose

10cpmargin

Coprime stability margin of plant-controller feedback loop

Syntax

[MARG,FREQ] = cpmargin(P,C)
[MARG,FREQ] = cpmargin(P,C,TOL)

Description

[MARG,FREQ] = cpmargin(P,C) calculates the normalized coprime factor/gap

metric robust stability of the multivariable feedback loop consisting of C in
negative feedback with P. C should only be the compensator in the feedback

path, not any reference channels, if it is a two degree-of-freedom (2-DOF)

architecture. The output MARG contains upper and lower bound for the
normalized coprime factor/gap metric robust stability margin. FREQ is the
frequency associated with the upper bound.
[MARG,FREQ] = cpmargin(P,C,TOL) specifies a relative accuracy TOL for

calculating the normalized coprime factor/gap metric robust stability margin.

(TOL=1e-3 by default).

See Also

loopmargin
gapmetric
wcmargin

Comprehensive analysis of feedback loops

Compute the gap and the Vinnicombe gap metric
Worst-case disk gain/phase margins for
plant-controller feedback loop

10-39

dec2mat

Purpose

10dec2mat

Extract matrix variable value from vector of decision variables

Syntax

valX = dec2mat(lmisys,decvars,X)

Description

Given a value decvars of the vector of decision variables, dec2mat computes the
corresponding value valX of the matrix variable with identifier X. This
identifier is returned by lmivar when declaring the matrix variable.
Recall that the decision variables are all free scalar variables in the LMI
problem and correspond to the free entries of the matrix variables X1, . . ., XK.
Since LMI solvers return a feasible or optimal value of the vector of decision
variables, dec2mat is useful to derive the corresponding feasible or optimal
values of the matrix variables.

Example

See the description of feasp.

See Also

mat2dec
decnbr
decinfo

10-40

Return the vector of decision variables corresponding

to particular values of the matrix variables
Give the total number of decision variables in a system
of LMIs
Describe how the entries of a matrix variable X relate
to the decision variables

decay

Purpose

10decay

Quadratic decay rate of polytopic or affine P-systems

Syntax

[drate,P] = decay(ps,options)

Description

For affine parameter-dependent systems

E(p) x = A(p)x, p(t) = (p1(t), . . ., pn(t))
or polytopic systems
E(t) x = A(t)x,

(A, E) Co{(A1, E1), . . ., (An, En)},

decay returns the quadratic decay rate drate, i.e., the smallest R such that

ATQE + EQAT < Q

holds for some Lyapunov matrix Q > 0 and all possible values of (A, E). Two
control parameters can be reset via options(1) and options(2):
If options(1)=0 (default), decay runs in fast mode, using the least expensive
sufficient conditions. Set options(1)=1 to use the least conservative
conditions.
options(2) is a bound on the condition number of the Lyapunov matrix P.
The default is 109.

See Also

quadstab
pdlstab
psys

Quadratic stability of polytopic or

parameter-dependent systems
Robust stability of polytopic or affine
parameter-dependent systems (P-system)
Specification of uncertain state-space models

10-41

decinfo

Purpose

10decinfo

Describe how entries of matrix variable X relate to decision variables

Syntax

decinfo(lmisys)
decX = decinfo(lmisys,X)

Description

The function decinfo expresses the entries of a matrix variable X in terms of

the decision variables x1, . . ., xN. Recall that the decision variables are the free
scalar variables of the problem, or equivalently, the free entries of all matrix
variables described in lmisys. Each entry of X is either a hard zero, some
decision variable xn, or its opposite xn.
If X is the identifier of X supplied by lmivar, the command
decX = decinfo(lmisys,X)

returns an integer matrix decX of the same dimensions as X whose (i, j) entry is
0 if X(i, j) is a hard zero
n if X(i, j) = xn (the n-th decision variable)
n if X(i, j) = xn
decX clarifies the structure of X as well as its entry-wise dependence on
x1, . . ., xN. This is useful to specify matrix variables with atypical structures
(see lmivar).
decinfo can also be used in interactive mode by invoking it with a single

argument. It then prompts the user for a matrix variable and displays in return
the decision variable content of this variable.

Example 1

Consider an LMI with two matrix variables X and Y with structure:

X = x I3 with x scalar
Y rectangular of size 2-by-1
If these variables are defined by
setlmis([])
X = lmivar(1,[3 0])
Y = lmivar(2,[2 1])
:
:

10-42

decinfo

lmis = getlmis

the decision variables in X and Y are given by

dX = decinfo(lmis,X)
dX =
1
0
0

0
1
0

0
0
1

dY = decinfo(lmis,Y)
dY =
2
3

This indicates a total of three decision variables x1, x2, x3 that are related to the
entries of X and Y by
x 0 0
1
X = 0 x1 0

0 0 x
1

x
Y= 2
x
3

Note that the number of decision variables corresponds to the number of free
entries in X and Y when taking structure into account.

Example 2

Suppose that the matrix variable X is symmetric block diagonal with one 2-by-2
full block and one 2-by-2 scalar block, and is declared by
setlmis([])
X = lmivar(1,[2 1;2 0])
:
lmis = getlmis

The decision variable distribution in X can be visualized interactively as

follows:
decinfo(lmis)

10-43

decinfo

There are 4 decision variables labeled x1 to x4 in this problem.

Matrix variable Xk of interest (enter k between 1 and 1, or 0 to
quit):
?> 1
The decision variables involved in X1 are among {-x1,...,x4}.
Their entry-wise distribution in X1 is as follows
(0,j>0,-j<0 stand for 0,xj,-xj, respectively):
X1 :
1
2
0
0

2
3
0
0

0
0
4
0

0
0
0
4

*********
Matrix variable Xk of interest (enter k between 1 and 1, or 0 to
quit):
?> 0

See Also

lmivar
mat2dec
dec2mat

10-44

Specify the matrix variables in an LMI problem

Return the vector of decision variables corresponding
to particular values of the matrix variables
Given values of the decision variables, derive the
corresponding values of the matrix variables decnbr

decnbr

Purpose

10decnbr

Number of decision variables in LMI system

Syntax

ndec = decnbr(lmisys)

Description

The function decnbr returns the number ndec of decision variables (free
scalar variables) in the LMI problem described in lmisys. In other words, ndec
is the length of the vector of decision variables.

Example

For an LMI system lmis with two matrix variables X and Y such that
X is symmetric block diagonal with one 2-by-2 full block, and one 2-by-2
scalar block
Y is 2-by-3 rectangular,
the number of decision variables is
ndec = decnbr(LMIs)
ndec =
10

This is exactly the number of free entries in X and Y when taking structure into
account (see decinfo for more details).

See Also

dec2mat
decinfo
mat2dec

Given values of the decision variables, derive the

corresponding values of the matrix variables decnbr
Describe how the entries of a matrix variable X relate
to the decision variables
Return the vector of decision variables corresponding
to particular values of the matrix variables

10-45

defcx

Purpose

Specify cTx objectives for mincx

10defcx

Syntax

[V1,...,Vk] = defcx(lmisys,n,X1,...,Xk)

Description

defcx is useful to derive the c vector needed by mincx when the objective is

expressed in terms of the matrix variables.

Given the identifiers X1,...,Xk of the matrix variables involved in this
objective, defcx returns the values V1,...,Vk of these variables when the n-th
decision variable is set to one and all others to zero.

See Also

10-46

mincx
decinfo

Minimize a linear objective under LMI constraints

Describe how the entries of a matrix variable X relate
to the decision variables

dellmi

Purpose

10dellmi

Remove LMI from system of LMIs

Syntax

newsys = dellmi(lmisys,n)

Description

dellmi deletes the n-th LMI from the system of LMIs described in lmisys. The
updated system is returned in newsys.

The ranking n is relative to the order in which the LMIs were declared and
corresponds to the identifier returned by newlmi. Since this ranking is not
modified by deletions, it is safer to refer to the remaining LMIs by their
identifiers. Finally, matrix variables that only appeared in the deleted LMI are
removed from the problem.

Example

Suppose that the three LMIs

T

A 1 X1 + X1A1 + Q1 < 0
T

A 2 X2 + X2A2 + Q2 < 0
T

A 3 X3 + X3A3 + Q3 < 0
have been declared in this order, labeled LMI1, LMI2, LMI3 with newlmi, and
stored in lmisys. To delete the second LMI, type
lmis = dellmi(lmisys,LMI2)
lmis now describes the system of LMIs
T

A1 X1 + X1 A1 + Q1 < 0
T

A3 X3 + X3 A3 + Q3 < 0

(10-3)
(10-4)

and the second variable X2 has been removed from the problem since it no
longer appears in the system (10-3)(10-4).
To further delete (10-4), type
lmis = dellmi(lmis,LMI3)

or equivalently

10-47

dellmi

lmis = dellmi(lmis,3)

Note that (10-4) has retained its original ranking after the first deletion.

See Also

newlmi
lmiedit
lmiinfo

10-48

Attach an identifying tag to LMIs

Specify or display systems of LMIs as MATLAB
expressions
Interactively retrieve information about the variables
and term content of LMIs

delmvar

Purpose

10delmvar

Remove matrix variable from LMI problem

Syntax

newsys = delmvar(lmisys,X)

Description

delmvar removes the matrix variable X with identifier X from the list of
variables defined in lmisys. The identifier X should be the second argument
returned by lmivar when declaring X. All terms involving X are automatically
removed from the list of LMI terms. The description of the resulting system of
LMIs is returned in newsys.

Example

Consider the LMI

T
T
0 < A Y + B YA + Q CX + D

T T
T
T

( X + X )
X C +D

involving two variables X and Y with identifiers X and Y. To delete the variable
X, type
lmisys = delmvar(lmisys,X)

Now lmisys describes the LMI

T

T
0 < A YB + B YA + Q D

D
0
with only one variable Y. Note that Y is still identified by the label Y.

See Also

lmivar
setmvar
lmiinfo

Specify the matrix variables in an LMI problem

Instantiate a matrix variable and evaluate all LMI
terms involving this matrix variable
Interactively retrieve information about the variables
and term content of LMIs

10-49

diag

Purpose

10diag

Diagonalize vector of uncertain matrices and systems

Syntax

v = diag(x)

Description

If x is a vector of uncertain system models or matrices, diag(x) puts x on the

main diagonal. If x is a matrix of uncertain system models or matrices, diag(x)
is the main diagonal of v. diag(diag(x)) is a diagonal matrix of uncertain
system models or matrices.

Example

The statement produces a diagonal system mxg of size 4-by-4. Given

multivariable system xx, a vector of the diagonal elements of xxg is found using
diag.
x = rss(3,4,1);
xg = frd(x,logspace(-2,2,80));
size(xg)
FRD model with 4 output(s) and 1 input(s), at 80 frequency
point(s).
mxg = diag(xg);
size(mxg)
FRD model with 4 output(s) and 4 input(s), at 80 frequency
point(s).
xxg = [xg(1:2,1) xg(3:4,1)];
m = diag(xxg);
size(m)
FRD model with 2 output(s) and 1 input(s), at 80 frequency
point(s).

See Also

10-50

append

Group models by appending their inputs and outputs

dkitopt

Purpose

10dkitopt

Create options object for use with dksyn

Syntax

dkitopt
options = dkitopt
options = dkitopt('name1',value1,'name2',value2,...)

Description

dkitopt with no input or output arguments displays a complete list of option

properties and their default values.
options = dkitopt creates an options object with all the properties set to their

default values.
options = dkitopt('name1',value1,'name2',value2,...) creates a
dkitopt object called options with specific values assigned to certain

properties. Any unspecified property is set to its default value. When entering
property names, it is sufficient to only type the leading characters that identify
the name uniquely. Property names are not case sensitive.
If the AutoIter property is set to 'off', the D-K iteration procedure is
interactive. You are prompted to fit the D-Scale data and provide input on the
control design process.
This table lists the dkitopt object properties.
Object Property

Description

FrequencyVector

Frequency vector used for analysis.

Default is an empty matrix ([]) which
results in the frequency range and
number of points chosen automatically.

InitialController

Controller used to initiate first

iteration. Default is an empty SS object.

AutoIter

Automated mu-synthesis mode. Default

is 'on'.

DisplayWhileAutoIter

Displays iteration progress in AutoIter

mode. Default is 'off'.

StartingIterationNumber

Starting iteration number. Default is 1.

10-51

dkitopt

Example

Object Property

Description

NumberOfAutoIterations

Number of D-K iterations to perform.

Default is 10.

AutoScalingOrder

Maximum state order for fitting

D-scaling data. Default is 5.

AutoIterSmartTerminate

Automatic termination of iteration

procedure based on progress of design
iteration. Default is 'on'.

AutoIterSmartTerminateTol

Tolerance used by AutoIterSmartTerminate. Default is 0.005.

Default

A 1-by-1 structure of dkitopt property

default values.

Meaning

A 1-by-1 structure text description of

each property.

This statement creates a dkitopt options object called opt with all default
values.
opt = dkitopt
Property Object Values:
FrequencyVector:
InitialController:
AutoIter:
DisplayWhileAutoIter:
StartingIterationNumber:
NumberOfAutoIterations:
AutoScalingOrder:
AutoIterSmartTerminate:
AutoIterSmartTerminateTol:
Default:
Meaning:

[]
[0x0 ss]
'on'
'off'
1
10
5
'on'
0.0050
[1x1 struct]
[1x1 struct]

The dksyn options object opt is updated with the following statements. The
frequency vector is set to logspace(-2,3,80), the number of D-K iterations to
16, and the maximum state order of the fitted D-scale data to 9.

10-52

dkitopt

opt.FrequencyVector = logspace(-2,3,80);
opt.NumberOfAutoIterations = 16;
opt.AutoScalingOrder = 16;
opt
Property Object Values:
FrequencyVector: [1x80 double]
InitialController: [0x0 ss]
AutoIter: 'on'
DisplayWhileAutoIter: 'off'
StartingIterationNumber: 1
NumberOfAutoIterations: 16
AutoScalingOrder: 16
AutoIterSmartTerminate: 'on'
AutoIterSmartTerminateTol: 0.0050
Default: [1x1 struct]
Meaning: [1x1 struct]

In this statement, the same properties are set with a single call to robopt.
opt = dkitopt('FrequencyVector',logspace(-2,3,80),...
'NumberOfAutoIterations',16,'AutoScalingOrder',9)
Property Object Values:
FrequencyVector: [1x80 double]
InitialController: [0x0 ss]
AutoIter: 'on'
DisplayWhileAutoIter: 'off'
StartingIterationNumber: 1
NumberOfAutoIterations: 16
AutoScalingOrder: 9
AutoIterSmartTerminate: 'on'
AutoIterSmartTerminateTol: 0.0050
Default: [1x1 struct]
Meaning: [1x1 struct]

Algorithm

The dksyn command stops iterating before the total number of automated
iterations ('NumberOfAutoIterations') if 'AutoIterSmartTerminate' is set
to 'on' and a stopping criterion is satisfied. The stopping criterion involves the
m(i) value of the current ith iteration, m(i-1) and m(i-2), the previous two
iterations and the options property 'AutoIterSmartTerminateTol'. The DK
iteration procedure automatically terminates if the difference between each of

10-53

dkitopt
the three values is less than the relative tolerance of
AutoIterSmartTerminateTol (i) or the current value (i) has increased
relative to the value of the previous iteration (i1) by
20AutoIterSmartTerminateTol.

See Also

10-54

dksyn
h2syn
hinfsyn
mussv
robopt
robuststab
robustperf
wcgopt

Synthesize a robust controller via D-K iteration

Synthesize a H2 optimal controller
Synthesize a H optimal controller
Calculate bounds on the Structured Singular Value ()
Create a robustperf/robuststab options object
Calculate stability margins of uncertain systems
Calculate performance margins of uncertain systems
Create a wcgain options object

dksyn

Purpose

Synthesis of robust controller via -synthesis D-K iteration

10dksyn

Syntax

k = dksyn(p,nmeas,ncont)
[k,clp,bnd] = dksyn(p,nmeas,ncont)
[k,clp,bnd] = dksyn(p,nmeas,ncont,opt)
[k,clp,bnd,dkinfo] = dksyn(p,nmeas,ncont,prevdkinfo,opt)
[k,clp,bnd,dkinfo] = dksyn(p)

Description

dksyn synthesizes a robust controller via D-K iteration. The D-K iteration

procedure is an approximation to -synthesis control design. The objective of

-synthesis is to minimize the structure singular value of the corresponding
robust performance problem associated with the uncertain system p. The
uncertain system p is an open-loop interconnection containing known
components including the nominal plant model, uncertain parameters,
ucomplex, and unmodeled LTI dynamics, ultidyn, and performance and
uncertainty weighting functions. You use weighting functions to include
magnitude and frequency shaping information in the optimization. The control
objective is to synthesize a stabilizing controller k that minimizes the robust
performance value, which corresponds to bnd.
The D-K iteration procedure involves a sequence of minimizations, first over
the controller variable K (holding the D variable associated with the scaled
upper bound fixed), and then over the D variable (holding the controller K
variable fixed). The D-K iteration procedure is not guaranteed to converge to
the minimum value, but often works well in practice. dksyn automates the
D-K iteration procedure and the options object dkitopt allows you to customize
its behavior.
[k,clp,bnd] = dksyn(p,nmeas,ncont) synthesizes a robust controller k for
the uncertain open-loop plant model p via the D-K iteration. p is a uss and has
nmeas measurement outputs and ncont control inputs. It is assumed that the
measurement outputs and control inputs correspond to the last nmeas outputs
and ncont inputs. k is the controller, clp is the closed-loop interconnection and
bnd is the robust performance bound for the uncertain closed-loop system clp.
p, k, clp, and bnd are related as follows:
clp = lft(p,k);
bnd = robustperf(clp);

10-55

dksyn

[k,clp,bnd] = dksyn(p,nmeas,ncont,opt) specifies options described in

dkitopt. opt is created via the command opt=dkitopt. See dkitopt for more

details.
[k,clp,bnd,dkinfo] = dksyn(p,nmeas,ncont,prevdkinfo,opt) allows you
to use information from a previous dksyn iteration. prevdkinfo is a structure
from a previous attempt at designing a robust controller using dksyn.
prevdkinfo is used when the dksyn starting iteration is not 1
(opt.StartingIterationNumber = 1) to determine the correct D-scalings to
initiate the iteration procedure.
dkinfo is an N-by-1 cell array where N is the total number of iterations
performed. The ith cell contains a structure with the following fields:
Field

Description

Controller at ith iteration, an ss object

Bnds

Robust performance bound on the closed-loop

system (double)

DL

Left D-scale, an ss object

DR

Right D-scale, an ss object

MussvBnds

Upper and lower bounds, an frd object

MussvInfo

Structure returned from mussv at each iteration.

k = dksyn(p) is a valid calling argument provided p is a uss object and has

two-input/two-output partitioning as defined by mktito.

Examples

The following statements create a robust performance control design for an

unstable, uncertain single-input/single-output plant model. The nominal plant
s
model, G, is an unstable first order system ------------ .
s1
G = tf(1,[1 -1]);

The model itself is uncertain. At low frequency, below 2 rad/s, it can vary up to
25% from its nominal value. Around 2 rad/s the percentage variation starts to

10-56

dksyn

increase and reaches 400% at approximately 32 rad/s. The percentage model

uncertainty is represented by the weight Wu which corresponds to the frequency
variation of the model uncertainty and the uncertain LTI dynamic object
InputUnc.
Wu = 0.25*tf([1/2 1],[1/32 1]);
InputUnc = ultidyn('InputUnc',[1 1]);T

The uncertain plant model Gpert represents the model of the physical system
to be controlled.

Plant model set: Gpert

Input Line
d

Wu
u
K

Wp

Gpert = G*(1+InputUnc*Wu);

The robust stability objective is to synthesize a stabilizing LTI controller for all
the plant models parameterized by the uncertain plant model, Gpert. The
control interconnection structure is shown in the following figure.
The goal is to synthesize a controller that stabilizes and achieves the
closed-loop performance objectives for all possible plant models in Gpert. The
performance objective is defined as a weighted sensitivity minimization
problem. See the preceding figure.
1
The sensitivity function, S, is defined as S = ------------------ where P is the plant
1 + PK
model and K is the controller. A weighted sensitivity minimization problem
selects a weight Wp, which corresponds to the inverse of the desired sensitivity
function of the closed-loop system as a function of frequency. Hence the product
of the sensitivity weight Wp and actual closed-loop sensitivity function is less
than 1 across all frequencies. The sensitivity weight Wp has a gain of 100 at low

10-57

dksyn

frequency, begins to decrease at 0.006 rad/s, and reaches a minimum

magnitude of 0.25 after 2.4 rad/s.
Wp = tf([1/4 0.6],[1 0.006]);

The defined sensitivity weight Wp implies that the desired disturbance rejection
should be at least 100:1 disturbance rejection at DC, rise slowly between 0.006
and 2.4 rad/s, and allow the disturbance rejection to increase above the
open-loop level, 0.25, at high frequency.
When the plant model is uncertain, the closed-loop performance objective is to
achieve the desired sensitivity function for all plant models defined by the
uncertain plant model, Gpert. The performance objective for an uncertain
system is a robust performance objective. A block diagram of this uncertain
closed-loop system illustrating the performance objective (closed-loop transfer
function from de) is shown.
e

d
P
y

u
K

From the definition of the robust performance control objective, the weighted,
uncertain control design interconnection model, which includes the robustness
and performance objectives, can be constructed and is denoted by P. The
robustness and performance weights are selected such that if the robust
performance structure singular value, bnd, of the closed-loop uncertain system,
clp, is less than 1 then the performance objectives have been achieved for all
the plant models in the model set.
You can form the uncertain transfer matrix P from [d; u] to [e; y] using the
following commands.
P = [Wp; 1 ]*[1 Gpert];
[K,clp,bnd] = dksyn(P,1,1);
bnd
bnd =

10-58

dksyn

0.6860
P = [Wp; 1 ]*[1 Gpert];
[K,clp,bnd] = dksyn(P,1,1);
bnd
bnd =
0.6860

The controller K achieves a robust performance value bnd of 0.685. Therefore

you have achieved the robust performance objectives for the given problem.
You can use the robustperf command to analyze the closed-loop robust
performance of clp.
[rpnorm,wcf,wcu,report] = robustperf(clp);
disp(report{1})
Uncertain system, clp, achieves robust performance. The analysis
showed clp can tolerate 146% of the model uncertainty and achieve
the performance and stability objectives. A model uncertainty
exists of size 146% that results in a peak gain performance of
0.686 at 0.569 rad/s.

Algorithm

The D-K iteration procedure is an approximation to -synthesis control design.

It involves a sequence of minimizations, first over the controller variable K
(holding the D variable associated with the scaled upper bound fixed), and
then over the D variable (holding the controller K variable fixed). The D-K
iteration procedure is not guaranteed to converge to the minimum value, but
often works well in practice.
dksyn automates the D-K iteration procedure. Internally, the algorithm works
with the generalized scaled plant model P, which is extracted from a uss object
using the command lftdata.

The following is a list of what occurs during a single, complete step of the D-K
iteration.
1 (In the first iteration, this step is skipped.) The calculation (from the

previous step) provides a frequency-dependent scaling matrix, Df. The

fitting procedure fits these scalings with rational, stable transfer function
matrices. After fitting, plots of

10-59

dksyn

( D f ( j )F L ( P, K ) ( j )D f ( j ) )
and
( j )F ( P, K ) ( j )D
1 ( j ) )
(D
f
f
L
are shown for comparison.
is absorbed into
(In the first iteration, this step is skipped.) The rational D
the open-loop interconnection for the next controller synthesis. Using either
, an estimate
the previous frequency-dependent Ds or the just-fit rational D
of an appropriate value for the H norm is made. This is simply a
conservative value of the scaled closed-loop H norm, using the most recent
controller and either a frequency sweep (using the frequency-dependent Ds)
or a state-space calculation (with the rational Ds).
2 (The first iteration begins at this point.) A controller is designed using

synthesis on the scaled open-loop interconnection. If you set the

DisplayWhileAutoIter field in dkitopt to 'on', the following information
is displayed:
a The progress of the -iteration is displayed.
b The singular values of the closed-loop frequency response are plotted.
c

You are given the option to change the frequency range. If you change it,
all relevant frequency responses are automatically recomputed.

d You are given the option to rerun the H synthesis with a set of modified

parameters if you set the AutoIter field in dkitopt to 'off'. This is

convenient if, for instance, the bisection tolerance was too large, or if
maximum gamma value was too small.
3 The structured singular value of the closed-loop system is calculated and

plotted.
4 An iteration summary is displayed, showing all the controller order, as well

as the peak value of of the closed-loop frequency responses.

5 The choice of stopping or performing another iteration is given.

Subsequent iterations proceed along the same lines without the need to reenter
the iteration number. A summary at the end of each iteration is updated to
reflect data from all previous iterations. This often provides valuable
information about the progress of the robust controller synthesis procedure.

10-60

dksyn

Interactive Fitting of D-Scalings

Setting the AutoIter field in dkitopt to 'off' requires that you interactively
fit the D-scales each iteration. During step 2 of the D-K iteration procedure, you
are prompted to enter your choice of options for fitting the D-scaling data. You
press return after, the following is a list of your options.
Enter Choice (return for list):
Choices:
nd
Move to Next D-Scaling
nb
Move to Next D-Block
i
d
apf
mx 3
at 1.01
0
2
n
e
s

Increment Fit Order

Decrement Fit Order
Auto-PreFit
Change Max-Order to 3
Change Auto-PreFit tol to 1.01
Fit with zeroth order
Fit with second order
Fit with n'th order
Exit with Current Fittings
See Status

nd and nb allow you to move from one D-scale data to another. nd moves to
the next scaling, whereas nb moves to the next scaling block. For scalar
D-scalings, these are identical operations, but for problems with full
D-scalings, (perturbations of the form I) they are different. In the (1,2)
subplot window, the title displays the D-scaling block number, the
row/column of the scaling that is currently being fitted, and the order of the
current fit (with d for data when no fit exists).
You can increment or decrement the order of the current fit (by 1) using i and
d.
apf automatically fits each D-scaling data. The default maximum state order
of individual D-scaling is 5. The mx variable allows you to change the
maximum D-scaling state order used in the automatic prefitting routine. mx
must be a positive, nonzero integer. at allows you to define how close the
rational, scaled upper bound is to approximate the actual upper bound in
a norm sense. Setting at to 1 would require an exact fit of the D-scale data,
and is not allowed. Allowable values for at are greater than 1. This setting

10-61

dksyn

plays a role (mildly unpredictable, unfortunately) in determining where in

the (D,K) space the D-K iteration converges.
Entering a positive integer at the prompt will fit the current D-scale data
with that state order rational transfer function.
e exits the D-scale fitting to continue the D-K iteration.
The variable s displays a status of the current and fits.

Limitations

There are two shortcomings of the D-K iteration control design procedure:
Calculation of the structured singular value () is approximated by its
upper bound. This is not a serious problem because the value of and its
upper bound are often close.
The D-K iteration is not guaranteed to converge to a global, or even local
minimum [SteD]. This is a serious problem, and represents the biggest
limitation of the design procedure.
In spite of these drawbacks, the D-K iteration control design technique appears
to work well on many engineering problems. It has been applied to a number
of real-world applications with success. These applications include vibration
suppression for flexible structures, flight control, chemical process control
problems, and acoustic reverberation suppression in enclosures.

References

[1] Balas, G.J., and J.C. Doyle, Robust control of flexible modes in the
controller crossover region, AIAA Journal of Guidance, Dynamics and
Control, Vol. 17, no. 2, March-April, 1994, p. 370-377.
[2] Balas, G.J., A.K. Packard, and J.T. Harduvel, Application of -synthesis
techniques to momentum management and attitude control of the space
station, AIAA Guidance, Navigation and Control Conference, New Orleans,
August 1991.
[3] Doyle, J.C., K. Lenz, and A. Packard, Design examples using -synthesis:
Space shuttle lateral axis FCS during reentry, NATO ASI Series, Modelling,
Robustness, and Sensitivity Reduction in Control Systems, vol. 34,
Springer-Verlag, Berlin 1987.
[4] Packard, A., J. Doyle, and G. Balas, Linear, multivariable robust control
with a perspective, ASME Journal of Dynamic Systems, Measurement and
Control, 50th Anniversary Issue, Vol. 115, no. 2b, June 1993, p. 310-319.

10-62

dksyn

[5] [SteD:] Stein, G., and J. Doyle, Beyond singular values and loopshapes,
AIAA Journal of Guidance and Control, Vol. 14, No. 1, January, 1991, p. 5-16.

See Also

dkitopt
h2syn
hinfsyn
loopmargin
mktito
mussv
robuststab
robustperf
wcgain
wcsens
wcmargin

Create a dksyn options object

Synthesize a H2 optimal controller
Synthesize a H optimal controller
Comprehensive analysis of feedback loop
Create a two-input/two-output LTI partition
Calculate bounds on the Structured Singular Value ()
Calculate stability margins of uncertain systems
Calculate performance margins of uncertain systems
Calculate worst-case gain of a system
Calculate worst-case sensitivities for feedback loop
Calculate worst-case margins for feedback loop

10-63

dmplot

Purpose

10dmplot

Interpret disk gain and phase margins

Syntax

dmplot
dmplot(diskgm)
[dgm,dpm] = dmplot

Description

dmplot plots disk gain margin (dgm) and disk phase margin (dpm). Both
margins are derived from the largest disk that

Contains the critical point (-1,0)

Does not intersect the Nyquist plot of the open-loop response L
diskgm is the radius of this disk and a lower bound on the classical gain margin.
dmplot(diskgm) plots the maximum allowable phase variation as a function of
the actual gain variation for a given disk gain margin diskgm (the maximum
gain variation being diskgm). The closed-loop system is guaranteed to remain

stable for all combined gain/phase variations inside the plotted ellipse.
[dgm,dpm] = dmplot returns the data used to plot the gain/phase variation

ellipse.

Examples

When you call dmplot (without an argument), the resulting plot shows a
comparison of a disk margin analysis with the classical notations of gain and
phase margins. The Nyquist plot is of the loop transfer function L(s)
s
------ + 1
30
L ( s ) = ----------------------------------------------------------( s + 1 ) ( s 2 + 1.6s + 16 )
The Nyquist plot of L corresponds to the blue line.
The unit disk corresponds to the dotted red line.
GM and PM indicate the location of the classical gain and phase margins for
the system L.
DGM and DPM correspond to the disk gain and phase margins, respectively.
The disk margins provide a lower bound on classical gain and phase margins.

10-64

dmplot

The disk margin circle, represented by the dashed black line, corresponds to
the largest disk centered at (GMD + 1/GMD)/2 that just touches the loop
transfer function L. This location is indicated by the red dot.
Disk gain margin (DGM) and disk phase margin (DPM) in the Nyquist plot
Unit Disk
Disk Margin
Critical point

1
0.8
0.6
Disk Margin

Imaginary Axis

0.4
0.2

DGM

GM

0.2
0.4

DPM

0.6
0.8

PM
1
1.5

0.5

0.5

Real Axis

The x-axis corresponds to the gain variation, in dB, and the y-axis corresponds
to the phase variation allowable, in degrees. For a disk gain margin
corresponding to 3 dB (1.414), the closed-loop system is stable for all phase and
gain variations inside the blue ellipse. For example, the closed-loop system can
simultaneously tolerate +/- 2 dB gain variation and +/- 14 deg phase variations.
dmplot(1.414)

10-65

dmplot

Max. Allowable Phase Variation for a 1.41dB Disk Gain Margin.

(stability is guaranteed for all gain/phase variations inside the ellipse)
20

18

16

Phase Variation (deg)

14

12

10

0
4

References

0
1
Gain Variation (dB)

[1] Barrett, M.F., Conservatism with robustness tests for linear feedback
control systems, Ph.D. Thesis, Control Science and Dynamical Systems,
University of Minnesota, 1980.
[2] Blight, J.D., R.L. Dailey, and Gangsass, D., Practical control law design for
aircraft using multivariable techniques, International Journal of Control, Vol.
59, No. 1, 1994, 93-137.
[3] Bates, D., and I. Postlethwaite, Robust Multivariable Control of Aerospace
Systems, Delft University Press, Delft, The Netherlands, ISBN: 90-407-2317-6,
2002.

See Also

10-66

loopmargin
wcmargin

Comprehensive stability analysis of feedback loops

Calculate worst-case margins for feedback loop

drawmag

Purpose

10drawmag

Interactive mouse-based sketching and fitting tool

Syntax

[sysout,pts] = drawmag(data)
[sysout,pts] = drawmag(data,init_pts)

Description

drawmag interactively uses the mouse in the plot window to create pts (the frd
object) and sysout (a stable minimum-phase ss object), which approximately
fits the frequency response (magnitude) in pts.

Input arguments:
data

Either a frequency response object that is plotted as a

reference, or a constant matrix of the form [xmin xmax ymin
ymax] specifying the plot window on the data.

init_pts

Optional frd objects of initial set of points.

Output arguments:
sysout

Stable, minimum-phase ss object that approximately fits, in

magnitude, the pts data.

pts

Frequency response of points.

While drawmag is running, all interaction with the program is through the
mouse and/or the keyboard. The mouse, if there is one, must be in the plot
window. The program recognizes several commands:
Clicking the mouse button adds a point at the crosshairs. If the crosshairs
are outside the plotting window, the points are plotted when the fitting,
windowing, or replotting mode is invoked. Typing a is the same as clicking
the mouse button.
Typing r removes the point with frequency nearest that of the crosshairs.
Typing any integer between 0 and 9 fits the existing points with a transfer
function of that order. The fitting routine approximately minimizes the
maximum error in a log sense. The new fit is displayed along with the points,
and the most recent previous fit, if it exists.
Typing w uses the crosshair location as the initial point in creating a window.
Moving the crosshairs and clicking the mouse or pressing any key then gives

10-67

drawmag

a second point at the new crosshair location. These two points define a new
window on the data, which is immediately replotted. This is useful in fine
tuning parts of the data. You can call windowing repeatedly.
Typing p simply replots the data using a window that covers all the current
data points as well as whatever was specified in in. Typically used after
windowing to view all the data.
Typing k invokes the keyboard using the keyboard command. You should
exercise caution when using this option, as it can wreak havoc on the
program if variables are changed.

See Also

10-68

ginput
loglog

Graphically input information from mouse

Plotting frequency responses on log-log scale

equate

Purpose

10equate

Equate icsignal expresssions

Syntax

c=equate(a,b)

Description

c = equate(a,b) is the icsignal object capturing the equality of the icsignal

objects a and b. The Equation property within iconnect objects is a cell array,
whose entries are set by equating icsignal objects to define the relationship.
a or b may also be the zero-vector of appropriate dimension (or even just the
scalar 0) to equate an icsignal object to zero.

See Also

iconnect, icsignal

10-69

evallmi

Purpose

10evallmi

Evaluate LMI for given values of decision variables

Syntax

evalsys = evallmi(lmisys,decvars)

Description

evallmi evaluates all LMI constraints for a particular instance decvars of the
vector of decision variables. Recall that decvars fully determines the values of

the matrix variables X1, . . ., XK. The evaluation consists of replacing all terms
involving X1, . . ., XK by their matrix value. The output evalsys is an LMI
system containing only constant terms.
The function evallmi is useful for validation of the LMI solvers output. The
vector returned by these solvers can be fed directly to evallmi to evaluate all
variable terms. The matrix values of the left- and right-hand sides of each LMI
are then returned by showlmi.

Observation

evallmi is meant to operate on the output of the LMI solvers. To evaluate all

LMIs for particular instances of the matrix variables X1, . . ., XK, first form the
corresponding decision vector x with mat2dec and then call evallmi with x as
input.

Examples

Consider the feasibility problem of finding X > 0 such that

ATXA X + I < 0

where A = 0.5 0.2

0.1 0.7

This LMI system is defined by:

setlmis([])
X = lmivar(1,[2 1]) % full symmetric X
lmiterm([1 1 1 X],A',A) % LMI #1: A'*X*A
lmiterm([1 1 1 X],-1,1) % LMI #1: -X
lmiterm([1 1 1 0],1) % LMI #1: I
lmiterm([-2 1 1 X],1,1) % LMI #2: X
lmis = getlmis

To compute a solution xfeas, call feasp by

[tmin,xfeas] = feasp(lmis)

10-70

evallmi

The result is
tmin =
-4.7117e+00
xfeas' =
1.1029e+02

-1.1519e+01

1.1942e+02

The LMI constraints are therefore feasible since tmin < 0. The solution X
corresponding to the feasible decision vector xfeas would be given by
X = dec2mat(lmis,xfeas,X).
To check that xfeas is indeed feasible, evaluate all LMI constraints by typing
evals = evallmi(lmis,xfeas)

The left- and right-hand sides of the first and second LMIs are then given by
[lhs1,rhs1] = showlmi(evals,1)
[lhs2,rhs2] = showlmi(evals,2)

and the test

eig(lhs1-rhs1)
ans =
-8.2229e+01
-5.8163e+01

confirms that the first LMI constraint is satisfied by xfeas.

See Also

showlmi
setmvar
dec2mat
mat2dec

Return the left- and right-hand sides of an LMI after

evaluation of all variable terms
Instantiate a matrix variable and evaluate all LMI
terms involving this matrix variable
Given values of the decision variables, derive the
corresponding values of the matrix variables
Return the vector of decision variables corresponding
to particular values of the matrix variables

10-71

feasp

Purpose

10feasp

Compute solution to given system of LMIs

Syntax

[tmin,xfeas] = feasp(lmisys,options,target)

Description

The function feasp computes a solution xfeas (if any) of the system of LMIs
described by lmisys. The vector xfeas is a particular value of the decision
variables for which all LMIs are satisfied.
Given the LMI system
T
T
N LxN M R ( x )M

(10-5)

xfeas is computed by solving the auxiliary convex program:

T

Minimize t subject to N L ( x )N M R ( x ) M tI
The global minimum of this program is the scalar value tmin returned as first
output argument by feasp. The LMI constraints are feasible if tmin 0 and
strictly feasible if tmin < 0. If the problem is feasible but not strictly feasible,
tmin is positive and very small. Some post-analysis may then be required to
decide whether xfeas is close enough to feasible.
The optional argument target sets a target value for tmin. The optimization
code terminates as soon as a value of t below this target is reached. The default
value is target = 0.
Note that xfeas is a solution in terms of the decision variables and not in terms
of the matrix variables of the problem. Use dec2mat to derive feasible values of
the matrix variables from xfeas.

Control
Parameters

The optional argument options gives access to certain control parameters for
the optimization algorithm. This five-entry vector is organized as follows:
options(1) is not used
options(2) sets the maximum number of iterations allowed to be performed
by the optimization procedure (100 by default)
options(3) resets the feasibility radius. Setting options(3) to a value R >
0 further constrains the decision vector x = (x1, . . ., xN) to lie within the ball

10-72

feasp

xi < R
2

i=1

In other words, the Euclidean norm of xfeas should not exceed R. The
feasibility radius is a simple means of controlling the magnitude of solutions.
Upon termination, feasp displays the f-radius saturation, that is, the norm
of the solution as a percentage of the feasibility radius R.
The default value is R = 109. Setting options(3) to a negative value
activates the flexible bound mode. In this mode, the feasibility radius is
initially set to 108, and increased if necessary during the course of
optimization
options(4) helps speed up termination. When set to an integer value J > 0,
the code terminates if t did not decrease by more than one percent in relative
terms during the last J iterations. The default value is 10. This parameter
trades off speed vs. accuracy. If set to a small value (< 10), the code
terminates quickly but without guarantee of accuracy. On the contrary, a
large value results in natural convergence at the expense of a possibly large
number of iterations.
options(5) = 1 turns off the trace of execution of the optimization
procedure. Resetting options(5) to zero (default value) turns it back on.
Setting option(i) to zero is equivalent to setting the corresponding control
parameter to its default value. Consequently, there is no need to redefine the
entire vector when changing just one control parameter. To set the maximum
number of iterations to 10, for instance, it suffices to type
options=zeros(1,5)
options(2)=10

Memory
Problems

% default value for all parameters

When the least-squares problem solved at each iteration becomes ill

conditioned, the feasp solver switches from Cholesky-based to QR-based linear
algebra (see Memory Problems on page 10-202 for details). Since the QR
mode typically requires much more memory, MATLAB may run out of memory
and display the message
??? Error using ==> feaslv
Out of memory. Type HELP MEMORY for your options.

10-73

feasp

You should then ask your system manager to increase your swap space or, if no
additional swap space is available, set options(4) = 1. This will prevent
switching to QR and feasp will terminate when Cholesky fails due to
numerical instabilities.

Examples

Consider the problem of finding P > I such that

T

A 1 P + PA 1 < 0

(10-6)

A 2 P + PA 2 < 0

(10-7)

A 3 P + PA 3 < 0

(10-8)

with data

A 1 = 1 2 , A 2 = 0.8 1.5 , A 3 = 1.4 0.9

1 3
1.3 2.7
0.7 2.0
This problem arises when studying the quadratic stability of the polytope of
matrices Co{A1, A2, A3}.
To assess feasibility with feasp, first enter the LMIs (9-8)(9-10) by:
setlmis([])
p = lmivar(1,[2 1])
lmiterm([1 1 1 p],1,a1,'s') % LMI
lmiterm([2 1 1 p],1,a2,'s') % LMI
lmiterm([3 1 1 p],1,a3,'s') % LMI
lmiterm([-4 1 1 p],1,1) % LMI #4:
lmiterm([4 1 1 0],1) % LMI #4: I
lmis = getlmis

#1
#2
#3
P

Then call feasp to find a feasible decision vector:

[tmin,xfeas] = feasp(lmis)

This returns tmin = -3.1363. Hence (9-8)(9-10) is feasible and the dynamical
system x = A(t)x is quadratically stable for A(t) Co{A1, A2, A3}.
To obtain a Lyapunov matrix P proving the quadratic stability, type

10-74

feasp

P = dec2mat(lmis,xfeas,p)

This returns

P = 270.8 126.4
126.4 155.1
It is possible to add further constraints on this feasibility problem. For
instance, you can bound the Frobenius norm of P by 10 while asking tmin to be
less than or equal to 1. This is done by
[tmin,xfeas] = feasp(lmis,[0,0,10,0,0],-1)

The third entry 10 of options sets the feasibility radius to 10 while the third
argument -1 sets the target value for tmin. This yields tmin = -1.1745 and a
matrix P with largest eigenvalue max(P) = 9.6912.

References

The feasibility solver feasp is based on Nesterov and Nemirovskis Projective

Method described in
Nesterov, Y., and A. Nemirovski, Interior Point Polynomial Methods in Convex
Programming: Theory and Applications, SIAM, Philadelphia, 1994.
Nemirovski, A., and P. Gahinet, The Projective Method for Solving Linear
Matrix Inequalities, Proc. Amer. Contr. Conf., 1994, Baltimore, Maryland, p.
840844.
The optimization is performed by the C-MEX file feaslv.mex.

See Also

mincx
gevp
dec2mat

Minimize a linear objective under LMI constraints

Generalized eigenvalue minimization under LMI
constraints
Given values of the decision variables, derive the
corresponding values of the matrix variables

10-75

fitfrd

Purpose

10fitfrd

Fit D-scaling frequency response data with state-space model

Syntax

B = fitfrd(A,N)
B = fitfrd(A,N,RD)
B = fitfrd(A,N,RD,WT)

Description

B = fitfrd(A,N) is a state-space object with state dimension N, where A is an

frd object and N is a nonnegative integer. The frequency response of B closely
matches the D-scale frequency response data in A.
A must have either 1 row or 1 column, although it need not be 1-by-1. B will be
the same size as A. In all cases, N should be a nonnegative scalar.
B = fitfrd(A,N,RD) forces the relative degree of B to be RD. RD must be a
nonnegative integer. The default value for RD is 0. If A is a row (or column) then
RD can be a vector of the same size as well, specifying the relative degree of each
entry of B. If RD is a scalar, then it specifies the relative degree for all entries of
B. You can specify the default value for RD by setting RD to an empty matrix.
B = fitfrd(A,N,RD,WT) uses the magnitude of WT to weight the optimization
fit criteria. WT can be a double, ss or frd. If WT is a scalar, then it is used to
weight all entries of the error criteria (A-B). If WT is a vector, it must be the same
size as A, and each individual entry of WT acts as a weighting function on the
corresponding entry of (A-B).

Examples

You can use the fitfrd command to fit D-scale data. For example, create
D-scale frequency response data from a fifth-order system.
sys = tf([1 2 2],[1 2.5 1.5])*tf(1,[1 0.1]);
sys = sys*tf([1 3.75 3.5],[1 2.5 13]);
omeg = logspace(-1,1);
sysg = frd(sys,omeg);
bode(sysg,'r-');

10-76

fitfrd

Original data from 5th order system

10

Magnitude (dB)

10

15

20
0

Phase (deg)

30

60

90

120
1

10

10
Frequency (rad/sec)

10

You can try to fit the frequency response D-scale data sysg with a first-order
system, b1. Similarly, you can fit the D-scale data with a third-order system,
b3.
b1 = fitfrd(sysg,1);
b3 = fitfrd(sysg,3);

Compare the original D-scale data sysg with the frequency responses of the
first and third-order models calculated by fitfrd:
b1g = frd(b1,omeg);
b3g = frd(b3,omeg);
bode(sysg,'r-',b1g,'k:',b3g,'b-.')

10-77

fitfrd

First and third order fit to 5th order system

10

Magnitude (dB)

10

15

20
0

Phase (deg)

45

90

5th order system

1st order fit
3rd order fit

135
1

10

10
Frequency (rad/sec)

10

Limitations

Numerical conditioning problems arise if the state order of the fit N is selected
to be higher than required by the dynamics of A.

See Also

fitmagfrd

10-78

Fit magnitude data with stable LTI model

fitmagfrd

Purpose

10fitmagfrd

Fit frequency response magnitude data with stable, minimum-phase

state-space model

Syntax

B
B
B
B

Description

B = fitmagfrd(A,N) is a stable, minimum-phase ss object, with

state-dimension N, whose frequency response magnitude closely matches the
magnitude data in A. A is a 1-by-1 frd object, and N is a nonnegative integer.

=
=
=
=

fitmagfrd(A,N)
fitmagfrd(A,N,RD)
fitmagfrd(A,N,RD,WT)
fitmagfrd(A,N,RD,WT,C)

B = fitmagfrd(A,N,RD) forces the relative degree of B to be RD. RD must be a

nonnegative integer whose default value is 0. You can specify the default value
for RD by setting RD to an empty matrix.
B = fitmagfrd(A,N,RD,WT) uses the magnitude of WT to weight the
optimization fit criteria. WT can be a double, ss or frd. If WT is a scalar, then it
is used to weight all entries of the error criteria (A-B). If WT is a vector, it must
be the same size as A, and each individual entry of WT acts as a weighting
function on the corresponding entry of (A-B). The default value for WT is 1, and
you can specify it by setting WT to an empty matrix.
B = fitmagfrd(A,N,RD,WT,C) enforces additional magnitude constraints on B,
relative to A as:

if C(w) == -1, then enforce abs(B(w)) <= abs(A(w)

if C(w) == 1, then enforce abs(B(w)) >= abs(A(w))
if C(w) == 0, then no additional constraint
where w denotes the frequency. C should be a scalar, double, or frd, with
C.Frequency equal to A.Frequency.

Examples

You can use the fitmagfrd command to fit frequency response magnitude data.
Create frequency response data from a fifth-order system.
sys = tf([1 2 2],[1 2.5 1.5])*tf(1,[1 0.1]);
sys = sys*tf([1 3.75 3.5],[1 2.5 13]);
omeg = logspace(-1,1);

10-79

fitmagfrd

sysg = abs(frd(sys,omeg));
bodemag(sysg,'r');
10

Original data from 5th order system

Magnitude (dB)

10

15

20
1
10

10
Frequency (rad/sec)

10

Fit the magnitude data with a minimum-phase, stable third-order system:

ord = 3;
b1 = fitmagfrd(sysg,ord);
b1g = frd(b1,omeg);

10-80

fitmagfrd

bodemag(sysg,'r',b1g,'k:');
Third order fit to 5th order system

10

Magnitude (dB)

10

15

5th order system

3rd order fit
20
1
10

10
Frequency (rad/sec)

10

Fit the magnitude data with a third order system constrained to lie below and
above the given data.
b2 = fitmagfrd(sysg,ord,[],[],-1);
b2g = frd(b2,omeg);
b3 = fitmagfrd(sysg,ord,[],[],1);
b3g = frd(b3,omeg);

10-81

fitmagfrd

bodemag(sysg,'r',b1g,'k:',b2g,'b-.',b3g,'m--')
Third order fits to 5th order system

10

Magnitude (dB)

10

15

5th order system

3rd order fit
Bounded below fit
Bounded above fit
20
1
10

Algorithm

10
Frequency (rad/sec)

10

fitmagfrd uses a version of log-Chebychev magnitude design, solving

min f

subject to (at every frequency point in A):

|d|^2 /(1+ f/WT) < |n|^2/A^2 < |d|^2*(1 + f/WT)

plus additional constraints imposed with C. n, d denote the numerator and

denominator, respectively, and B = n/d. n and d have orders (N-RD) and N,
respectively. The problem is solved using bisection on f and linear
programming for fixed f. An alternate approximate method, which cannot
enforce the constraints defined by C, is B = fitfrd(genphase(A),N,RD,WT).

Limitations

This input frd object must be either a scalar 1-by-1 object or, a row, or column
vector.

References

Oppenheim, A.V., and R.W. Schaffer, Digital Signal Processing, Prentice Hall,
New Jersey, 1975, p. 513.

10-82

fitmagfrd

See Also

fitfrd

Fits frequency response data with state-space model

10-83

frd/loglog

Purpose

10frd/loglog

Log-log scale plot of frd objects

Syntax

loglog(sys)
loglog(sys,linetype)

Description

loglog is the same as plot, except a logarithmic (base 10) scale is used for
both the X- and Y-axis.

The argument list consists of (many) x-data, y-data and line type triples. The
x-data and y-data can be specified individually, as doubles, or jointly as an frd.

Examples

Generate three frequency response objects, sys1g, sys2g and sys3g. Plot the
magnitude of these transfer functions versus frequency on a log-log plot.
omega = logspace(-2,2,40);
sys1 = tf(0.9,[.1 1]);
sys1g = frd(sys1,omega);
sys2 = tf(0.9,[1 1]);
sys2g = frd(sys2,omega);
sys3 = tf(0.9,[10 1]);
sys3g = frd(sys3,omega);
loglog(abs(sys1g),abs(sys2g),'r+',sys3g.Freq,...
abs(sys3g.Resp(:)),'--');

10-84

frd/loglog

10

10

10

10

10

10

See Also

10

plot
semilogx, semilogy

10

10

10

Plots on linear axis

Plots semi-log scale plot

10-85

frd/rcond

Purpose

10frd/rcond

Reciprocal condition estimator of frd object

Syntax

r = rcond(x)

Description

rcond(x) is an estimate for the reciprocal of the condition of the frd object x in
the 1-norm obtained by the LAPACK condition estimator. rcond operates on
x.ReponseData of the x frd at each frequency to construct r. If x is well
conditioned, rcond(x) is near 1.0. If x is badly conditioned, rcond(x) is near

EPS.
r=rcond(x) returns r as an frd object.

See Also

10-86

cond
norm
condest
normest

Calculates condition number with respect to inversion

Calculates matrix or vector norm
Calculates a 1-norm condition number estimate
Calculates a matrix 2-norm estimate

frd/schur

Purpose

10frd/schur

Schur decomposition of frd object

Syntax

[u,t]
t
[u,t]
t
[u,t]
t

Description

frd/schur applies the schur command to frd objects. [u,t] = schur(x)

operates on the x.ReponseData of the frd object at each frequency point to
construct u and t. u and t are frd objects. x must be square. See the built-in
schur command for details.

See Also

qz
schur

=
=
=
=
=
=

schur(x)
schur(x)
schur(x,0)
schur(x,0)
schur(x,'econ')
schur(x,'econ')

Creates a QZ factorization for generalized eigenvalues

Calculates a Schur decomposition

10-87

frd/semilogx

Purpose

10frd/semilogx

Syntax

semilogx(...) same as plot

Description

semilogx(...) is the same as plot(...), except a logarithmic (base 10) scale

Semilog scale plot of frd object

is used for the x-axis.

See Also

10-88

loglog
plot
frd/semilog

Plot frd object on a log-log scale

Plot frd object on a linear scale
Plot frd object on a semi-log scale

frd/semilog

Purpose

10frd/semilog

Syntax

semilogy(...) same as plot

Description

semilogy(...) is the same as plot(...), except a logarithmic (base 10) scale

is used for the y-axis.

See Also

loglog
plot
frd/semilogx

Semilog scale plot of frd object

Plot frd object on a log-log scale

Plot frd object on a linear scale
Plot frd object on a semi-log scale

10-89

frd/svd

Purpose

10frd/svd

Singular value decomposition of frd object

Syntax

S = svd(X)
[U,S,V] = svd(X)

Description

S = svd(X) operates on X.ReponseData at each frequency to construct S.

[U,S,V] = svd(X) produces a diagonal frd S that has the same dimensions as
X and includes positive diagonal elements in decreasing order. U and V are
unitary matrices and frd objects, such that X = U*S*V'. For more information,
see the built-in svd command.

See Also

10-90

schur
svd

Constructs a Schur decomposition

Constructs a singular value decomposition

gapmetric

Purpose

10gapmetric

Compute upper bounds on gap and nugap distances between two systems

Syntax

[gap,nugap] = gapmetric(p0,p1)
[gap,nugap] = gapmetric(p0,p1,tol)

Description

[gap,nugap] = gapmetric(p0,p1) calculates upper bounds on the gap and

nugap distances (metric) between systems p0 and p1. The gap and nugap values

lie between 0 and 1. A small value (relative to 1) implies that any controller
that stabilizes p0 will likely stabilize p1, and, moreover, that the closed-loop
gains of the two closed-loop systems will be similar. A gap or nugap of 0 implies
that p0 equals p1, and a value of 1 implies that the plants are far apart. The
input and output dimensions of p0 and p1 must be the same.
[gap,nugap] = gapmetric(p0,p1,tol) specifies a relative accuracy for
calculating the gap metric and nugap metric. The default value for tol is 0.001.
The computed answers are guaranteed to satisfy
gap-tol < gapexact(p0,p1) <=

Examples

gap

Consider two plant models. One plant is an unstable, first-order, with transfer
function 1/(s-0.001) and the other plant is stable first-order with transfer
function 1/(s+0.001).
p1 = tf(1,[1 -0.001]);
p2 = tf(1,[1 0.001]);

Despite the fact that one plant is unstable and the other is stable, these plants
are close in the gap and nugap metrics. Intuitively, this is obvious, because, for
instance, the feedback controller K=1 stabilizes both plants and renders the
closed-loop systems nearly identical.
[g,ng] = gapmetric(p1,p2)
g =
0.0029
ng =
0.0020
K = 1;
H1 = loopsens(p1,K);
H2 = loopsens(p2,K);
subplot(2,2,1); bode(H1.Si,'-',H2.Si,'--');
subplot(2,2,2); bode(H1.Ti,'-',H2.Ti,'--');

10-91

gapmetric

subplot(2,2,3); bode(H1.PSi,'-',H2.PSi,'--');
subplot(2,2,4); bode(H1.CSo,'-',H2.CSo,'--');

Bode Diagram

50

90

0
4

10

45

90
2
10

10

Bode Diagram

10

10

50

100
180
Phase (deg)

50
0

10
10
Frequency (rad/sec)

Bode Diagram

0
Magnitude (dB)

Magnitude (dB)

10
10
Frequency (rad/sec)

50

Phase (deg)

50
0
Phase (deg)

Phase (deg)

100
180

45

90
2
10

Bode Diagram

50
Magnitude (dB)

Magnitude (dB)

90

0
1

10

10
10
Frequency (rad/sec)

10

10

10
10
Frequency (rad/sec)

10

Next, consider two stable plant models that differ by a first-order system. One
plant is the transfer function 50/(s+50) and the other plant is the transfer
function 50/(s+50) * 8/(s+8).
p3 = tf([50],[1 50]);
p4 = tf([8],[1 8])*p3;

Although the two systems have similar high-frequency dynamics and the same
unity gain at low frequency, the plants are modestly far apart in the gap and
nugap metrics.
[g,ng] = gapmetric(p3,p4)
g =
0.6156
ng =
0.6147

10-92

gapmetric

Algorithm

gap and nugap compute the gap and gap metrics between two LTI objects.
Both quantities give a numerical value (p0,p1) between 0 and 1 for the
distance between a nominal system p0 (G0) and a perturbed system p1 (G1).

The gap metric was introduced into the control literature by Zames and
El-Sakkary 1980, and exploited by Georgiou and Smith 1990. The gap metric
was derived by Vinnicombe 1993. For both of these metrics the following robust
performance result holds from Qui and Davidson 1992, and Vinnicombe 1993
arcsin b(G1,K1) arcsin b(G0,K0) arcsin (G0,G1) arcsin (K0,K1)
where
1

b ( G, K ) =

I ( I GK ) 1
G I
K

The interpretation of this result is that if a nominal plant G0 is stabilized by

controller K0, with stability margin b(G0,K0), then the stability margin when
G0 is perturbed to G1 and K0 is perturbed to K1 is degraded by no more than the
above formula. Note that 1/b(G,K) is also the signal gain from disturbances on
the plant input and output to the input and output of the controller. The gap
is always less than or equal to the gap, so its predictions using the above
robustness result are tighter.
To make use of the gap metrics in robust design, weighting functions need to
be introduced. In the above robustness result, G needs to be replaced by W2GW1
1
1
and K by W 1 KW 2 (similarly for G0, G1, K0 and K1). This makes the weighting
functions compatible with the weighting structure in the H loop shaping
control design procedure (see loopsyn and ncfsyn for more details).
The computation of the gap amounts to solving 2-block H problems (Georgiou,
Smith 1988). The particular method used here for solving the H problems is
based on Green et al., 1990. The computation of the nugap uses the method of
Vinnicombe, 1993.

References

Georgiou, T.T., On the computation of the gap metric, Systems Control

Letters, Vol. 11, 1988, p. 253-257
Georgiou, T.T., and M. Smith, Optimal robustness in the gap metric, IEEE
Transactions on Automatic Control, Vol. 35, 1990, p. 673-686

10-93

gapmetric

Green, M., K. Glover, D. Limebeer, and J.C. Doyle, A J-spectral factorization

approach to H control, SIAM J. of Control and Opt., 28(6), 1990, p. 1350-1371
Qiu, L., and E.J. Davison, Feedback stability under simultaneous gap metric
uncertainties in plant and controller, Systems Control Letters, Vol. 18-1, 1992
p. 9-22
Vinnicombe, G., Measuring Robustness of Feedback Systems, PhD
Dissertation, Department of Engineering, University of Cambridge, 1993.
Zames, G., and El-Sakkary, Unstable systems and feedback: The gap metric,
Proceedings of the Allerton Conference, October 1980, p. 380-385

See Also

10-94

loopmargin
loopsyn
ncfsyn
robuststab
wcsens
wcmargin

Comprehensive analysis of feedback loop

H - loop shaping controller synthesis
H - normalized coprime controller synthesis
Calculate stability margins of uncertain systems
Calculate worst-case sensitivities for feedback loop
Calculate worst-case margins for feedback loop

genphase

Purpose

10genphase

Fit single-input/single-output magnitude data with real, rational,

minimum-phase transfer function

Syntax

resp = genphase(d)

Description

genphase uses the complex-cepstrum algorithm to generate a complex

frequency response resp whose magnitude is equal to the real, positive
response d, but whose phase corresponds to a stable, minimum-phase function.
The input, d, and output, resp, are frd objects.

References

Oppenheim, A.V., and R.W. Schaffer, Digital Signal Processing, Prentice Hall,
New Jersey, 1975, p. 513.

See Also

fitmagfrd

Fitting magnitude data with stable LTI model

10-95

getlmis

Purpose

10getlmis

Internal description of LMI system

Syntax

lmisys = getlmis

Description

After completing the description of a given LMI system with lmivar and
lmiterm, its internal representation lmisys is obtained with the command
lmisys = getlmis

This MATLAB representation of the LMI system can be forwarded to the LMI
solvers or any other LMI-Lab function for subsequent processing.

See Also

10-96

setlmis
lmivar
lmiterm
newlmi

Initialize the description of an LMI system

Specify the matrix variables in an LMI problem
Specify the term content of LMIs
Attach an identifying tag to LMIs

gevp

Purpose

10gevp

Generalized eigenvalue minimization under LMI constraints

Syntax

[lopt,xopt] = gevp(lmisys,nlfc,options,linit,xinit,target)

Description

gevp solves the generalized eigenvalue minimization problem

Minimize l subject to:

C(x)

<

D(x)

(10-9)

0 <

B(x)

(10-10)

A(x)

< B(x)

(10-11)

where C(x) < D(x) and A(x) < B(x) denote systems of LMIs. Provided that
(9-11)(9-12) are jointly feasible, gevp returns the global minimum lopt and
the minimizing value xopt of the vector of decision variables x. The
corresponding optimal values of the matrix variables are obtained with
dec2mat.
The argument lmisys describes the system of LMIs (9-11)(9-13) for = 1. The
LMIs involving are called the linear-fractional constraints while (9-11)(9-12)
are regular LMI constraints. The number of linear-fractional constraints (9-13)
is specified by nlfc. All other input arguments are optional. If an initial
feasible pair (0, x0) is available, it can be passed to gevp by setting linit to 0
and xinit to x0. Note that xinit should be of length decnbr(lmisys) (the
number of decision variables). The initial point is ignored when infeasible.
Finally, the last argument target sets some target value for . The code
terminates as soon as it has found a feasible pair (, x) with target.

Caution

When setting up your gevp problem, be cautious to

Always specify the linear-fractional constraints (9-13) last in the LMI
system. gevp systematically assumes that the last nlfc LMI constraints are
linear fractional
Add the constraint B(x) > 0 or any other LMI constraint that enforces it (see
Remark below). This positivity constraint is required for regularity and good
formulation of the optimization problem.

10-97

gevp

Control
Parameters

The optional argument options gives access to certain control parameters of

the optimization code. In gevp, this is a five-entry vector organized as follows:
options(1) sets the desired relative accuracy on the optimal value lopt
(default = 102).
options(2) sets the maximum number of iterations allowed to be performed
by the optimization procedure (100 by default).
options(3) sets the feasibility radius. Its purpose and usage are as for
feasp.
options(4) helps speed up termination. If set to an integer value J > 0, the
code terminates when the progress in over the last J iterations falls below
the desired relative accuracy. By progress, we mean the amount by which
decreases. The default value is 5 iterations.
options(5) = 1 turns off the trace of execution of the optimization
procedure. Resetting options(5) to zero (default value) turns it back on.
Setting option(i) to zero is equivalent to setting the corresponding control
parameter to its default value.

Examples

Given

A 1 = 1 2 , A 2 = 0.8 1.5 , A 3 = 1.4 0.9 ,

1 3
1.3 2.7
0.7 2.0
consider the problem of finding a single Lyapunov function V(x) = xTPx that
proves stability of
x = Aix (i = 1, 2, 3)
( x )- . This is equivalent to minimizing
and maximizes the decay rate dV
--------------dt
subject to
I<P
T

A 1 P + PA 1 < P
T

A 2 P + PA 2 < P

10-98

(10-12)
(10-13)
(10-14)

gevp

A 3 P + PA 3 < P

(10-15)

To set up this problem for gevp, first specify the LMIs (9-15)(9-17) with = 1:
setlmis([]);
p = lmivar(1,[2 1])
lemiterm([1 1 1 0],1) % P > I : I
lemiterm([ 1 1 1 p],1,1) % P > I :
lemiterm([2 1 1 p],1,a1,'s') % LFC
lemiterm([ 2 1 1 p],1,1) % LFC # 1
lemiterm([3 1 1 p],1,a2,'s') % LFC
lemiterm([ 3 1 1 p],1,1) % LFC # 2
lemiterm([4 1 1 p],1,a3,'s') % LFC
lemiterm([ 4 1 1 p],1,1) % LFC # 3
lmis = getlmis

P
# 1 (lhs)
(rhs)
# 2 (lhs)
(rhs)
# 3 (lhs)
(rhs)

Note that the linear fractional constraints are defined last as required. To
minimize subject to (9-15)(9-17), call gevp by
[alpha,popt]=gevp(lmis,3)

This returns alpha = -0.122 as optimal value (the largest decay rate is
therefore 0.122). This value is achieved for

P = 5.58 8.35
8.35 18.64

Remark

Generalized eigenvalue minimization problems involve standard LMI

constraints (9-11) and linear fractional constraints (9-13). For well-posedness,
the positive definiteness of B(x) must be enforced by adding the constraint
B(x) > 0 to the problem. Although this could be done automatically from inside
the code, this is not desirable for efficiency reasons. For instance, the set of
constraints (9-12) may reduce to a single constraint as in the example above.
In this case, the single extra LMI P > I is enough to enforce positivity of all
linear-fractional right-hand sides. It is therefore left to the user to devise the
least costly way of enforcing this positivity requirement.

10-99

gevp

References

The solver gevp is based on Nesterov and Nemirovskis Projective Method

described in
Nesterov, Y., and A. Nemirovski, Interior Point Polynomial Methods in Convex
Programming: Theory and Applications, SIAM, Philadelphia, 1994.
The optimization is performed by the CMEX file fpds.mex.

See Also

dec2mat
decnbr
feasp
mincx

10-100

Given values of the decision variables, derive the

corresponding values of the matrix variables
Give the total number of decision variables in a system
of LMIs
Find a solution to a given system of LMIs
Minimize a linear objective under LMI constraints

gridureal

Purpose

10gridureal

Grid ureal parameters uniformly over range

Syntax

B = gridreal(A,N)
[B,SampleValues] = gridreal(A,N)
[B,SampleValues] = gridreal(A,Names,N)
[B,SampleValues] = gridreal(A,Names1,N1,Names2,N2,...)

Description

B = gridureal(A,N) substitutes N uniformly spaced samples of the uncertain

real parameters in A. The size of array B is equal to [size(A) N]. The N samples
are generated by uniformly gridding each ureal in A across its range. If A
includes uncertain objects other than ureal, then B is an uncertain object.
[B,SampleValues] = gridureal(A,N) additionally returns the specific
sampled values (as a STRUCT whose field names are the names of As uncertain
elements) of the uncertain reals. Hence, B is the same as
usubs(A,SampleValues).
[B,SampleValues] = gridureal(A,NAMES,N) samples only the uncertain
reals listed in the NAMES variable (CELL, or CHAR array). Any entries of NAMES
that are not elements of A are ignored. Note that
gridureal(A,fieldnames(A.Uncertainty),N) is the same as
gridureal(A,N).
[B,SampleValues] = gridureal(A,NAMES1,N1,NAMES2,N2,...) takes N1
samples of the uncertain real parameters listed in NAMES1, and N2 samples of
the uncertain real parameters listed in NAMES2 and so on. size(B) will equal
[size(A) N1 N2 ...].

Examples

Create two uncertain real parameters gamma and tau. The nominal value of
gamma is 4 and its range is 3 to 5. The nominal value of tau is 0.5 and its value
can change by +/- 30 percent.
gamma = ureal('gamma',4);
tau = ureal('tau',.5,'Percentage',30);

These uncertain parameters are used to construct an uncertain transfer

function p. An integral controller, c, is synthesized for the plant p based on the
nominal values of gamma and tau. The uncertain closed-loop system clp is
formed.

10-101

gridureal

p = tf(gamma,[tau 1]);
KI = 1/(2*tau.Nominal*gamma.Nominal);
c = tf(KI,[1 0]);
clp = feedback(p*c,1);

The figure below shows the open-loop unit step response (top plot) and
closed-loop response (bottom plot) for a grid of 20 values of gamma and tau.
subplot(2,1,1); step(gridureal(p,20),6)
title('Open-loop plant step responses')
subplot(2,1,2); step(gridureal(clp,20),6)

Openloop plant step responses

6
5

Amplitude

4
3
2
1
0

3
Time (sec)

Closedloop plant step responses

1.5

Amplitude

0.5

3
Time (sec)

Clearly this illustrates the low-frequency closed-loop insensitivity achieved by

the PI control system.

Multi-Parameter Example
The next example illustrates the different options in gridding
high-dimensional (e.g., n greater than 2) parameter spaces. An uncertain

10-102

gridureal

matrix, m, is constructed from four uncertain real parameters, a, b, c and d,

each making up the individual entries.
a=ureal('a',1); b=ureal('b',2); c=ureal('c',3); d=ureal('d',4);
m = [a b;c d];

In the first case, the (a,b) space is gridded at five places, and the (c,d) space at
3 places. The uncertain matrix m is evaluated at these 15 grid-points resulting
in the matrix m1.
m1 = gridureal(m,{'a';'b'},5,{'c';'d'},3);

In the second case, the (a,b,c,d) space is gridded at 15 places, and the
uncertain matrix m is sampled at these 15 points. The resulting matrix is m2.
m2 = gridureal(m,{'a';'b';'c';'d'},15);

The (2,1) entry of m is just the uncertain real parameter c. Below, you see the
histogram plots of the (2,1) entry of both m1 and m2. The (2,1) entry of m1 only
takes on 3 distinct values, while the (2,1) entry of m2 (which is also c) takes on
15 distinct values uniformly through its range.
subplot(2,1,1)
hist(m1(2,1,:))
title('2,1 entry of m1')
subplot(2,1,2)
hist(m2(2,1,:))

10-103

gridureal

title('2,1 entry of m2')

2,1 entry of m1
5
4
3
2
1
0

2.2

2.4

2.6

2.8

3.2

3.4

3.6

3.8

3.2

3.4

3.6

3.8

2,1 entry of m2
2

1.5

0.5

See Also

10-104

usample
usubs

2.2

2.4

2.6

2.8

Generates random samples of an atom

Substitutes values for atoms

h2hinfsyn

Purpose

10h2hinfsyn

Mixed H2/H synthesis with pole placement constraints

Syntax

[gopt,h2opt,K,R,S] = h2hinfsyn(P,r,obj,region,dkbnd,tol)

Description

h2hinfsyn performs multi-objective output-feedback synthesis. The control

problem is sketched in this figure.

z2

P(s)
y

u
K(s)

Figure 10-4: Mixed H2/H synthesis

If T(s) and T2(s) denote the closed-loop transfer functions from w to z and z2,
respectively, h2hinfsyn computes a suboptimal solution of the following
synthesis problem:
Design an LTI controller K(s) that minimizes the mixed H2/H criterion
2

T + T2 2
subject to
||T|| < 0
||T2||2 < 0
The closed-loop poles lie in some prescribed LMI region D.
Recall that ||.|| and ||.||2 denote the H norm (RMS gain) and H2 norm of transfer
functions.
P is any SS, TF, or ZPK LTI representation of the plant P(s), and r is a

three-entry vector listing the lengths of z2, y, and u. Note that z and/or z2 can
be empty. The four-entry vector obj = [0, 0, , ] specifies the H2/H

10-105

h2hinfsyn

constraints and trade-off criterion, and the remaining input arguments are
optional:
region specifies the LMI region for pole placement (the default region = []
is the open left-half plane). Use lmireg to interactively build the LMI region
description region
dkbnd is a user-specified bound on the norm of the controller feedthrough
matrix DK. The default value is 100. To make the controller K(s) strictly
proper, set dkbnd = 0.
tol is the required relative accuracy on the optimal value of the trade-off
criterion (the default is 102).
The function h2hinfsyn returns guaranteed H and H2 performances gopt and
h2opt as well as the SYSTEM matrix K of the LMI-optimal controller. You can
also access the optimal values of the LMI variables R, S via the extra output
arguments R and S.
A variety of mixed and unmixed problems can be solved with h2hinfsyn. In
particular, you can use h2hinfsyn to perform pure pole placement by setting
obj = [0 0 0 0]. Note that both z and z2 can be empty in such case.

References

Chilali, M., and P. Gahinet, H Design with Pole Placement Constraints: An

LMI Approach, to appear in IEEE Trans. Aut. Contr., 1995.
Scherer, C., Mixed H2 H Control, to appear in Trends in Control: A
European Perspective, volume of the special contributions to the ECC 1995.

See Also

10-106

lmireg
msfsyn

Specify LMI regions for pole placement purposes

Multi-model/multi-objective state-feedback synthesis

h2syn

Purpose

10h2syn

H2 control synthesis for LTI plant

Syntax

[K,CL,GAM,INFO]=H2SYN(P,NMEAS,NCON)

Description

h2syn computes a stabilizing H2 optimal LTI/SS controller K for a partitioned

LTI plant P. The controller, K, stabilizes the plant P and has the same number

A B1 B2
P = C 1 D 11 D 12
C 2 D 21 D 22
of states as P. The LTI system P is partitioned where inputs to B1 are the
disturbances, inputs to B2 are the control inputs, output of C1 are the errors to
be kept small, and outputs of C2 are the output measurements provided to the
controller. B2 has column size (NCON) and C2 has row size (NMEAS).
If P is constructed with mktito, you can omit NMEAS and NCON from the
arguments.
The closed-loop system is returned in CL and the achieved H2 cost in GAM. See
Figure 10-5, H2 control system CL= lft(P,K)=, on page 10-107. INFO is a
STRUCT array that returns additional information about the design.

A B1 B2

u2

P =

C 1P(s)
D 11 D 12
C 2 D 21 D 22

y1
y2

K
F(s)

Figure 10-5:

H2 control system CL=

lft(P,K)= T y u
1 1

10-107

h2syn

Output arguments

Description

LTI controller

CL= lft(P,K)
GAM = norm(CL)

LTI closed-loop system T y u

1 1
T
The H2 optimal cost = y1 u1

INFO

Additional output information

Additional output structure array INFO containing possible additional

information depending on METHOD:
INFO.NORMS

norms of 4 different quantities, full information control

cost (FI), output estimation cost (OEF), direct feedback
cost (DFL) and full control cost
(FC). NORMS = [FI OEF DFL FC];

INFO.KFI

full-information gain matrix (constant feedback)

u (t) = K x(t)
2
FI

INFO.GFI

full-information closed-loop system

GFI=ss(A-B2*KFI,B1,C1-D12*KFI,D11)

Examples

INFO.HAMX

X Hamiltonian matrix (state-feedback)

INFO.HAMY

Y Hamiltonian matrix (Kalman filter)

Example 1
Stabilize 4-by-5 unstable plant with 3-states, NMEAS=2, NCON=2.
rand('seed',0);randn('seed',0);
P=rss(3,4,5)';
[K,CL,GAM]=h2syn(P,2,1);
open_loop_poles=pole(P)
closed_loop_poles=pole(CL)

10-108

h2syn

Example 2
Mixed-Sensitivity H2 loop-shaping. Here the goal is to shape the sigma plots of
sensitivity S:=(I+GK)-1 and complementary sensitivity T:=GK(I+GK)-1, by
choosing a stabilizing K the minimizes the H2 norm of

Ty1 u1 =

W 1S
( W 2 G )T
W3 T

where
0.1 ( s + 1000 )
s1
G ( s ) = ------------ , W 1 = ----------------------------------- , W 2 = 0.1
100s + 1
s2
and there is no W3.
s=zpk('s');
G=10*(s-1)/(s+1)^2;
W1=0.1*(s+100)/(100*s+1); W2=0.1; W3=[];
P=augw(G,W1,W2,W3);
[K,CL,GAM]=h2syn(P);
L=G*K; S=inv(1+L); T=1-S;
sigma(L,'k-.',S,'r',T,'g')

Algorithms

The H2 optimal control theory has its roots in the frequency domain
interpretation the cost function associated with time-domain state-space LQG
control theory [1]. The equations and corresponding nomenclature used here
are taken from the Doyle et al., 1989 [2]-[3].
h2syn solves the H2 optimal control problem by observing that it is equivalent
to a conventional Linear-Quadratic Gaussian (LQG) optimal control problem.
For simplicity, we shall describe the details of algorithm only for the
continuous-time case, in which case the cost function JLQG satisfies

10-109

h2syn

1 T T

J LQG = lim E ---- y 1 y 1 dt

T T 0

1 T T T Q Nc x

= lim E ---- [ x u 2 ]
dt
T
u
T T 0

2
Nc R

T
1 T T T C1

x
--= lim E
[ x u2 ]
[ C 1 D 12 ]
dt
T
u2
T T 0

D 12

with plant noise u1 channel of intensity I, passing through the matrix

[B1;0;D12] to produce equivalent white correlated with plant and white
measurement noise having joint correlation function
Nf

T
(t )
E ( t ) [ ( ) ( ) ] =
T
( t)

Nf
=

B1
D 21

T (t )
B 1T D 21

The H2 optimal controller K(s) is thus realizable in the usual LQG manner as
a full-state feedback KFI and a Kalman filter with residual gain matrix KFC.
1 Kalman Filter

x = A x + B 2 u 2 + K FC ( y 2 C 2 x D 22 u 2 )
T

K FC = ( YC 2 + N f )

T 1

= ( YC 2 + B 1 D 21 ) ( D 21 D 21 )

where Y= YT0 solves the Kalman filter Riccati equation

T

YA + A Y ( YC 2 + N f ) ( C 2 Y + N f ) + = 0
2 Full-State Feedback

10-110

h2syn

u 2 = K FI x
1

K FI = R ( B 2 X + N c ) = ( D 12 D 12 ) ( B 2 X + D 12 C 1 )
where X = XT0 solves the state-feedback Riccati equation
T

A X + X A ( XB 2 + N c ) R ( B 2 X + N c ) + Q = 0
The final positive-feedback H2 optimal controller u 2 = K ( s )y 2 has a familiar
closed-form
K( s ) :=

A K FC C 2 B 2 K FI + K FC D 22 K FI

Kf

K FI

h2syn implements the continuous optimal H2 control design computations

using the formulae described in the Doyle, et al.[2]; for discrete-time plants,
h2syn uses the same controller formula, except that the corresponding discrete
time Riccati solutions (dare) are substituted for X and Y. A Hamiltonian is
formed and solved via a Riccati equation. In the continuous-time case, the
optimal H2-norm is infinite when the plant D11 matrix associated with the
input disturbances and output errors is non-zero; in this case, the optimal H2
controller returned by h2syn is computed by first setting D11 to zero.
3 Optimal Cost GAM

1
---

The full information (FI) cost is given by the equation ( trace ( B 1 X 2 B 1 ) ) 2 . The
1
---

output estimation cost (OEF) is given by ( trace ( F 2 Y 2 F 2 ) ) 2 , where

F 2 =: ( B 2 X 2 + D 12 C 1 ) . The disturbance feedforward cost (DFL) is

1
---

( trace ( L 2 X 2 L 2 ) ) 2 , where L2 is defined by ( Y 2 C 2 + B 1 D 21 ) and the full control

1
---

cost (FC) is given by ( trace ( C 1 Y 2 C 1 ) ) 2 . X2 and Y2 are the solutions to the X and
Y Riccati equations, respectively. For for continuous-time plants with zero
feedthrough term (D11 = 0), and for all discrete-time plants, the optimal H2 cost
= T y 1 u 1 2 is

10-111

h2syn

GAM =sqrt(FI^2 + OEF^2+ trace(D11*D11'));

otherwise, GAM = Inf.

Limitations

(A, B2, C2) must be stabilizable and detectable.

D12 must have full column rank and D21 must have full row rank.

References

[1] Safonov, M.G., A.J. Laub, and G. Hartmann, Feedback Properties of

Multivariable Systems: The Role and Use of Return Difference Matrix, IEEE
Trans. of Automat. Contr., AC-26, pp. 47-65, 1981.
[2] Doyle, J.C., K. Glover, P. Khargonekar, and B. Francis, State-space
solutions to standard H2 and H control problems, IEEE Transactions on
Automatic Control, vol. 34, no. 8, pp. 831847, August 1989.
[3] Glover, K., and J.C. Doyle, State-space formulae for all stabilizing
controllers that satisfy an H norm bound and relations to risk sensitivity,
Systems and Control Letters, 1988. vol. 11, pp. 167172, August 1989.

See Also

10-112

augw
hinfsyn

Augment plant weights for control design

H synthesis controller

hankelmr

Purpose

10hankelmr

Hankel minimum degree approximation (MDA) without balancing

Syntax

GRED = hankelmr(G)
GRED = hankelmr(G,order)
[GRED,redinfo] = hankelmr(G,key1,value1,...)
[GRED,redinfo] = hankelmr(G,order,key1,value1,...)

Description

hankelmr returns a reduced order model GRED of G and a struct array redinfo

containing the error bound of the reduced model and Hankel singular values of
the original system.
The error bound is computed based on Hankel singular values of G. For a stable
system Hankel singular values indicate the respective state energy of the
system. Hence, reduced order can be directly determined by examining the
system Hankel SVs, .
With only one input argument G, the function will show a Hankel singular
value plot of the original model and prompt for model order number to reduce.
This method guarantees an error bound on the infinity norm of the additive
error || G-GRED || for well-conditioned model reduced problems [1]:

G Gred 2

i
k+1

Note It seems this method is similar to the additive model reduction routines
balancmr and schurmr, but actually it can produce more reliable reduced
order model when the desired reduced model has nearly controllable and/or
observable states (has Hankel singular values close to machine accuracy).
hankelmr will then select an optimal reduced system to satisfy the error bound
criterion regardless the order one might naively select at the beginning.

10-113

hankelmr

This table describes input arguments for hankelmr.

Argument

Description

LTI model to be reduced (without any other inputs will plot

its Hankel singular values and prompt for reduced order)

ORDER

(Optional) an integer for the desired order of the reduced

model, or optionally a vector packed with desired orders for
batch runs

A batch run of a serial of different reduced order models can be generated by

specifying order = x:y, or a vector of integers. By default, all the
anti-stable part of a system is kept, because from control stability point of view,
getting rid of unstable state(s) is dangerous to model a system.
'MaxError' can be specified in the same fashion as an alternative for 'ORDER'.

In this case, reduced order will be determined when the sum of the tails of the
Hankel sv's reaches the 'MaxError'.
Argument

Value

Description

'MaxError'

Real number or
vector of
different errors

Reduce to achieve H error.

When present, 'MaxError'overides
ORDER input.

'Weights'

{Wout,Win} cell

Optimal 1x2 cell array of LTI weights

Wout (output) and Win (input). Default
for both is identity. Weights must be
invertible.

array

'Display'

'on' or 'off'

Display Hankel singular plots (default

'off').

'Order'

Integer, vector
or cell array

Order of reduced model. Use only if not

specified as 2nd argument.

Weights on the original model input and/or output can make the model
reduction algorithm focus on some frequency range of interests. But weights
have to be stable, minimum phase and invertible.

10-114

hankelmr

This table describes output arguments.

Argument

Description

GRED

LTI reduced order model. Become multi-dimensional array

when input is a serial of different model order array.

REDINFO

A STRUCT array with four fields:

REDINFO.ErrorBound (bound on || G-GRED ||)
REDINFO.StabSV (Hankel SV of stable part of G)
REDINFO.UnstabSV (Hankel SV of unstable part of G)
REDINFO.Ganticausal (Anti-causal part of Hankel MDA)

G can be stable or unstable, continuous or discrete.

Note If size(GRED) is not equal to the order you specified. The optimal
Hankel MDA algorithm has selected the best Minimum Degree Approximate
it can find within the allowable machine accuracy.

Algorithm

Given a state-space (A,B,C,D) of a system and k, the desired reduced order, the
following steps will produce a similarity transformation to truncate the
original state-space system to the kth order reduced model:
1 Find the controllability and observability grammians P and Q.
2 Form the descriptor
2

E = QP I
where k > k + 1 , and descriptor state-space
Take SVD of descriptor E and partition the result into kth order truncation
form

Es A

= A + QAP QB
CP
D

10-115

hankelmr

T
E 0 V E1

E = [U E1,U E2]

T
0 0 V E2

3 Apply the transformation to the descriptor state-space system above we

have
T
A 11 A 12
U E1
2 T
=
( A + QAP ) V E1 V E2
T
A 21 A 22
U E2

B1
B2

T
U E1
T
U E2

C1 C2 =

QB C

CP
B

V E1 V E2

D1 = D
4 Form the equivalent state-space model.

1 ( A A A A ) 1 ( B A A B )

E
B
11
12 22
21
E
1
12 22
2
=

C 1 C 2 A 22 A 21
D 1 C 2 A 22 B 2
D

The final kth order Hankel MDA is the stable part of the above state-space
realization. Its anticausal part is stored in redinfo.Ganticausal.
The proof of the Hankel MDA algorithm can be found in [2]. The error system
between the original system G and the Zeroth Order Hankel MDA G0 is an
all-pass function [1].

Examples

Given a continuous or discrete, stable or unstable system, G, the following

commands can get a set of reduced order models based on your selections:
rand('state',1234); randn('state',5678);G = rss(30,5,4);
[g1, redinfo1] = hankelmr(G); % display Hankel SV plot
% and prompt for order (try 15:20)
[g2, redinfo2] = hankelmr(G,20);

10-116

hankelmr

[g3, redinfo3] = hankelmr(G,[10:2:18]);

[g4, redinfo4] = hankelmr(G,'MaxError',[0.01, 0.05]);
rand('state',12345); randn('state',6789);
wt1 = rss(6,5,5); wt1.d = eye(5)*2;
wt2 = rss(6,4,4); wt2.d = 2*eye(4);
[g5, redinfo5] = hankelmr(G, [10:2:18], 'weight',{wt1,wt2});
for i = 1:5
figure(i); eval(['sigma(G,g' num2str(i) ');']);
end

Figure 10-6, Singular Value Bode Plot of G (30-state, 5 outputs, 4 inputs), on

page 10-118 shows a singular value Bode plot of a random system G with 20
states, 5 output and 4 inputs. The error system between G and its Zeroth order
Hankel MDA has it infinity norm equals to an all pass function, as shown in
Figure 10-7, All-Pass Error System Between G and Zeroth Order G Anticausal,
on page 10-118 (ref.: [5]).
The Zeroth order Hankel MDA and its error system sigma plot are obtained via
commands
[g0,redinfo0] = hankelmr(G,0);
sigma(G-redinfo0.Ganticausal)

This interesting all-pass property is unique in Hankel MDA model reduction.

10-117

hankelmr

Figure 10-6: Singular Value Bode Plot of G (30-state, 5 outputs, 4 inputs)

Figure 10-7: All-Pass Error System Between G and Zeroth Order G Anticausal

10-118

hankelmr

References

[1] Glover, K., All Optimal Hankel Norm Approximation of Linear

Multivariable Systems, and Their L error Bounds, Int. J. Control, vol. 39,
no. 6, pp. 1145-1193, 1984.
[2] Safonov, M.G., R.Y. Chiang, and D.J.N. Limebeer, Optimal Hankel Model
Reduction for Nonminimal Systems, IEEE Trans. on Automat. Contr., vol. 35,
no. 4, April 1990, pp. 496-502.

See Also

reduce
balancmr
schurmr
bstmr
ncfmr
hankelsv

Top level model reduction routines

Balanced truncation via square-root method
Balanced truncation via Schur method
Balanced stochastic truncation via Schur method
Balanced truncation for normalized coprime factors
Hankel singular value

10-119

hankelsv

Purpose

10hankelsv

Compute Hankel singular values for stable/unstable or continuous/discrete

system

Syntax

hankelsv(G)
hankelsv(G,ErrorType,style)
[sv_stab,sv_unstab]=hankelsv(G,ErrorType,style)

Description

[sv_stab,sv_unstab]=hankelsv(G,ErrorType,style) returns a column

vector SV_STAB containing the Hankel singular values of the stable part of G
and SV_UNSTAB of anti-stable part (if it exists). The Hankel SV's of anti-stable
part ss(a,b,c,d) is computed internally via ss(-a,-b,c,d). Discrete model is

converted to continuous one via the bilinear transform.

hankelsv(G) with no output arguments draws a bar graph of the Hankel

singular values such as the following:

10-120

hankelsv

This table describes optional input arguments for hankelsvd.

Algorithm

Argument

Value

Description

ERRORTYPE

'add'
'mult'
'ncf'

Regular Hankel SVs of G

Hankel SVs of phase matrix
Hankel SVs of coprime factors

STYLE

'abs'
'log'

Absolute value
logarithm scale

For ErrorType = 'add', hankelsv implements the numerically robust square

root method to compute the Hankel singular values [1]. Its algorithm goes as
follows:
Given a stable model G, with controllability and observability grammians P and
Q, compute the SVD of P and Q:
[Up,Sp,Vp] = svd(P);
[Uq,Sq,Vq] = svd(Q);

Then form the square roots of the grammians:

Lr = Up*diag(sqrt(diag(Sp)));
Lo = Uq*diag(sqrt(diag(Sq)));

The Hankel singular values are simply:

H = svd(Lo'*Lr);

This method not only takes the advantages of robust SVD algorithm, but also
ensures the computations stay well within the square root of the machine
accuracy.
For ErrorType = 'mult', hankelsv computes the Hankel singular value of the
phase matrix of G [2].
For ErrorType = 'ncf', hankelsv computes the Hankel singular value of the
normalized coprime factor pair of the model [3].

10-121

hankelsv

References

[1] Safonov, M.G., and R.Y. Chiang, A Schur Method for Balanced Model
Reduction, IEEE Trans. on Automat. Contr., vol. AC-2, no. 7, July 1989, pp.
729-733.
[2] Safonov, M.G., and R.Y. Chiang, Model Reduction for Robust Control: A
Schur Relative Error Method, International J. of Adaptive Control and Signal
Processing, Vol. 2, pp. 259-272, 1988.
[3] Vidyasagar, M., Control System Synthesis - A Factorization Approach.
London: The MIT Press, 1985.

See Also

10-122

reduce
balancmr
schurmr
bstmr
ncfmr
hankelmr

Top-level model reduction routines

Balanced truncation via square-root method
Balanced truncation via Schur method
Balanced stochastic truncation via Schur method
Balanced truncation for normalized coprime factors
Hankel minimum degree approximation

hinfgs

Purpose

10hinfgs

Gain-scheduled H controller synthesis

Syntax

[gopt,pdK,R,S] = hinfgs(pdP,r,gmin,tol,tolred)

Description

Given an affine parameter-dependent plant

x = A ( p )x + B ( p )w + B u
1
2

P z = C 1 ( p )x + D 11 ( p )w + D 12 u
y = C x + D w + D u
2
21
22

where the time-varying parameter vector p(t) ranges in a box and is measured
in real time, hinfgs seeks an affine parameter-dependent controller

= A K ( p ) + B K ( p )y
K
u = C K ( p ) + D K ( p )y
scheduled by the measurements of p(t) and such that
K stabilizes the closed-loop system
z

w
P

u
K

for all admissible parameter trajectories p(t).

K minimizes the closed-loop quadratic H performance from w to z.
The description pdP of the parameter-dependent plant P is specified with psys
and the vector r gives the number of controller inputs and outputs (set
r=[p2,m2] if y Rp2 and u Rm2). Note that hinfgs also accepts the polytopic
model of P returned, e.g., by aff2pol.

10-123

hinfgs

hinfgs returns the optimal closed-loop quadratic performance gopt and a

polytopic description of the gain-scheduled controller pdK. To test if a
closed-loop quadratic performance is achievable, set the third input gmin to .
The arguments tol and tolred control the required relative accuracy on gopt
and the threshold for order reduction. Finally, hinfgs also returns solutions R,
S of the characteristic LMI system.

Controller
Implementation

The gain-scheduled controller pdK is parametrized by p(t) and characterized by

A (p) B (p)
K
at the corners of the parameter box. The
the values Kj of K
j

CK ( p ) DK ( p )

command
Kj = psinfo(pdK,'sys',j)

returns the j-th vertex controller Kj while

pv = psinfo(pdP,'par')
vertx = polydec(pv)
Pj = vertx(:,j)

gives the corresponding corner j of the parameter box (pv is the parameter
vector description).
The controller scheduling should be performed as follows. Given the
measurements p(t) of the parameters at time t,
1 Express p(t) as a convex combination of the j:
N

p(t) = 11 + . . .+ NN,

j 0,

= 1

i=1

This convex decomposition is computed by polydec.

2 Compute the controller state-space matrices at time t as the convex

combination of the vertex controllers Kj:

A (t) B (t)
K
K

C
(
t
)
D
K
K( t )

10-124

=
j K .

i=1

hinfgs

3 Use AK(t), BK(t), CK(t), DK(t) to update the controller state-space equations.

References

Apkarian, P., P. Gahinet, and G. Becker, Self-Scheduled H Control of Linear

Parameter-Varying Systems, submitted to Automatica, October 1995.
Becker, G., Packard, P., Robust Performance of Linear-Parametrically
Varying Systems Using Parametrically-Dependent Linear Feedback, Systems
and Control Letters, 23 (1994), pp. 205-215.
Packard, A., Gain Scheduling via Linear Fractional Transformations, Syst.
Contr. Letters, 22 (1994), pp. 79-92.

See Also

psys
pvec
pdsimul
polydec

Specification of uncertain state-space models

Quantification of uncertainty on physical parameters
Time response of a parameter-dependent system along
a given parameter trajectory
Compute polytopic coordinates wrt. box corners

10-125

hinfsyn

Purpose

10hinfsyn

H optimal controller synthesis for LTI plant

Syntax

[K,CL,GAM,INFO] = hinfsyn(P)
[K,CL,GAM,INFO] = hinfsyn(P,NMEAS,NCON)
[K,CL,GAM,INFO] = hinfsyn(P,NMEAS,NCON,KEY1,VALUE1,KEY2,VALUE2,...)

Description

hinfsyn computes a stabilizing

partitioned LTI plant P.

H optimal LTI/SS controller K for a

A B1 B2
P = C 1 D 11 D 12
C 2 D 21 D 22
The controller, K, stabilizes the P and has the same number of states as P. The
SYSTEM P is partitioned where inputs to B1 are the disturbances, inputs to B2
are the control inputs, output of C1 are the errors to be kept small, and outputs
of C2 are the output measurements provided to the controller. B2 has column
size (NCON) and C2 has row size (NMEAS). The optional KEY and VALUE inputs
determine tolerance, solution method and so forthsee Figure 10-8, H control
system CL= lft(P,K)=, on page 10-126 for details.
The closed-loop system is returned in CL and the achieved H cost in GAM.
INFO is a STRUCT array that returns additional information about the
designsee Figure 10-9, Optional input arguments (KEY, VALUE) pairs, on
page 10-127.

u2

A B1 B2
D 11 D 12
P= C 1 P(s)
C 2 D 21 D 22

y1
y2

K
F(s)
Figure 10-8:

10-126

H control system CL=

lft(P,K)= T y u
1 1

hinfsyn

Key

Value

Meaning

'GMAX'

real

Initial upper bound on GAM (default=Inf)

'GMIN'

real

Initial lower bound on GAM (default=0)

'TOLGAM'

real

Relative error tolerance for GAM

(default=.01)

'S0'

real

(default=Inf) Frequency S0 at which entropy

is evaluated, only applies to METHOD 'maxe'

'METHOD'

'ric'

(default) Standard 2-Riccati solution,

'lmi'

LMI solution

'maxe'

Maximum entropy solution

'off'
'on'

(default) No command window display, or

command window displays synthesis progress
information

'DISPLAY'

Figure 10-9: Optional input arguments (KEY, VALUE) pairs

When DISPLAY='on', the hinfsyn program displays several variables

indicating the progress of the algorithm. For each value being tested, the
minimum magnitude, real part of the eigenvalues of the X and Y Hamiltonian
matrices are displayed along with the minimum eigenvalue of X and Y,
which are the solutions to the X and Y Riccati equations, respectively. The
maximum eigenvalue of XY, scaled by 2,is also displayed. A # sign is placed
to the right of the condition that failed in the printout.
Output arguments

Description

LTI controller

CL= lft(P,K)

LTI closed-loop system T y 1 u1

10-127

hinfsyn

GAM = norm(CL,Inf)

The H cost = T y 1 u 1

INFO

Additional output information

Additional output structure array INFO containing possible additional

information depending on METHOD:
INFO.AS

All solutions controller, LTI two-port LFT

INFO.KFI

Full information gain matrix (constant feedback)

x(t)
u (t) = K
2
FI u ( t )
1

Algorithm

INFO.KFC

Full control gain matrix (constant output-injection;

KFC is the dual of KFI)

INFO.GAMFI

H cost for full information KFI

INFO.GAMFC

H cost for full control KFC

The default 'ric' method uses the two-riccati formulae ([4],[5]) with
loopshifting [6]. In the case of the 'lmi' method, hinfsyn employs the LMI
technique ([7],[8],[9]). With 'METHOD' 'maxe', K returns the max entropy H
controller that minimize an entropy integral relating to the point s0; i.e.,
2
s 02
2

Entropy = -----ln detI T y 1 u1 ( j ) T y 1 u 1 ( j ) --------------------- d
2
s 2 + 2
0

where T y 1 u 1 is the closed-loop transfer function CL. With all methods, hinfsyn
uses a standard -iteration technique to determine the optimal value of .
Starting with high and low estimates of . The -iteration is a bisection
algorithm that iterates on the value of in an effort to approach the optimal H
control design. The stopping criterion for the bisection algorithm requires the
relative difference between the last value that failed and the last value that
passed be less than TOLGAM (default=.01)

10-128

hinfsyn

At each value of , the algorithm employed requires tests to determine whether

a solution exists for a given value. In the case of the 'ric' method, the
conditions checked for the existence of a solution are:
H and J Hamiltonian matrices (which are formed from the state-space data
of P and the level) must have no imaginary-axis eigenvalues.
The stabilizing Riccati solutions X and Y associated with the Hamiltonian
matrices must exist and be positive, semi-definite.
Spectral radius of (X,Y) must be less than or equal to 2.
When, DISPLAY is 'on', the hinfsyn program displays several variables, which
indicate which of the above conditions are satisfied for each value being
tested. In the case of the default 'ric' method, the display includes the
current value of being tested, real part of the eigenvalues of the X and Y
Hamiltonian matrices along with the minimum eigenvalue of X and Y, which
are the solutions to the X and Y Riccati equations, respectively. The maximum
eigenvalue of XY, scaled by 2, is also displayed. A # sign is placed to the
right of the condition that failed in the printout. A similar DISPLAY is
produced with method 'lmi'
The algorithm works best when the following conditions are satisfied by the
plant:
D12 and D21 have full rank.
A jI B 2
has full column rank for all R.
C 1 D 12
A jI B 1
has full row rank for all R.
C 2 D 21
When the above rank conditions do not hold, the controller may have
undesirable properties: If D12 and D21 are not full rank, the H controller K may
have large high-frequency gain. If either of the latter two rank conditions does
not hold at some frequency , the controller may have very lightly damped
poles near that frequency .

10-129

hinfsyn

In general, the solution to the infinity-norm optimal control problem is

non-unique. Whereas the K returned by hinfsyn is only a particular F(s), when
the 'ric' method is selected, the INFO.AS field of INFO give you in addition
the all-solution controller parameterization KAS(s) such that all solutions to
the infinity-norm control problem are parameterized by a free stable
contraction map U(s) constrained by ( U ( s ) < 1 ) (see Figure 10-10,
All-solution KAS(s) returned by INFO.AS, on page 10-130); that is, every

stabilizing controller K(s) that makes T y 1 u1 = sup max ( T y1 u1 ( j ) ) <

K=lft(INFO.AS,U)

where U is a stable LTI system satisfying norm(U,Inf) <1.

u

u2

y1

P(s)

K(s)

y2

KAS(s)

U(s)
F(s)

K(s)

Figure 10-10: All-solution KAS(s) returned by INFO.AS

An important use of the infinity-norm control theory is for direct shaping of

closed-loop singular value Bode plots of control systems. In such cases, the
system P(s) will typically be the plant augmented with suitable loop-shaping
filters see and mixsyn.

10-130

hinfsyn

Examples

Following are three simple problems solved via hinfsyn.

Example 1
A random 4-by-5 plant with 3-states, NMEAS=2, NCON=2
rand('seed',0);randn('seed',0);
P=rss(3,4,5);
[K,CL,GAM]=hinfsyn(P,2,2);

The optimal H cost in this case is GAM=0.2641. You verify that

T y1 u1

= sup max ( T y 1 u 1 ( j ) ) <

with a sigma plot

sigma(CL,ss(GAM));

Example 2
Mixed-Sensitivity,
0.1 ( s + 1000 )
1- W = ----------------------------------G ( s ) = s----------, W 2 = 0.1
, 1
100s + 1
s2
and there is no W3.
s=zpk('s');
G=(s-1)/(s+1);
W1=0.1*(s+100)/(100*s+1); W2=0.1; W3=[];
P=(G,W1,W2,W3);
[K,CL,GAM]=hinfsyn(P);
sigma(CL,ss(GAM));

In this case, GAM = 0.1854 = -14.6386 db

Example 3
Mixed sensitivity with W1 removed.
s=zpk('s');
G=(s-1)/(s+1);
W1=[]; W2=0.1; W3=[];
P=(G,W1,W2,W3)
[K,CL,GAM]=hinfsyn(P)

10-131

hinfsyn

In this case, GAM=0, K=0, and CL=K*(1+G*K)=0.

Limitation

The plant must be stabilizable from the control inputs u2 and detectable from
the measurement output y2:
(A,B2) must be stabilizable and (C2,A) must be detectable.
Otherwise, hinfsyn returns an error.

References

[1] Glover, K., and J.C. Doyle, State-space formulae for all stabilizing
controllers that satisfy an Hnorm bound and relations to risk sensitivity,
Systems and Control Letters, vol. 11, pp. 167172, 1988.
[2] Doyle, J.C., K. Glover, P. Khargonekar, and B. Francis, State-space
solutions to standard H2 and H control problems, IEEE Transactions on
Automatic Control, vol. 34, no. 8, pp. 831847, August 1989
[3] Safonov, M.G., D.J.N. Limebeer, and R.Y. Chiang, Simplifying the H
Theory via Loop Shifting, Matrix Pencil and Descriptor Concepts, Int. J.
Contr., vol. 50, no. 6, pp. 2467-2488, 1989.
[4] Packard, A., K. Zhou, P. Pandey, J. Leonhardson, and G. Balas, Optimal,
constant I/O similarity scaling for full-information and state-feedback
problems, Systems and Control Letters, 19:271280, 1992.
[5] Gahinet, P., and P. Apkarian, A linear matrix inequality approach to
H1-control, Int J. Robust and Nonlinear Control, 4(4):421448, JulyAugust
1994.
[6] Iwasaki, T., and R.E. Skelton, All controllers for the general H -control
problem: LMI existence conditions and state space formulas, Automatica,
30(8):13071317, August 1994.

See Also

10-132

augw
h2syn
loopsyn
mktito
ncfsyn

Augments plant weights for control design

H2 synthesis controller
H - loop shaping controller synthesis
2-input 2-output partition of a USS object
H - normailized coprime controller synthesis

iconnect

Purpose

10iconnect

Create empty interconnection object

Syntax

H = iconnect

Description

Interconnection objects (class iconnect) are an alternative to sysic, and are

used to build complex interconnections of uncertain matrices and systems.
An iconnect object has 3 fields to be set by the user, Input, Output and
Equation. Input and Output are icsignal objects, while Equation.is a
cell-array of equality constraints (using equate) on icsignal objects. Once
these are specified, then the System property is the input/output model,
implied by the constraints in Equation. relating the variables defined in Input
and Output.

Examples

iconnect can be used to create the transfer matrix M as described in the

following figure.

Create three scalar icsignals, r, e and y. Create an empty iconnect object.

Define the output of the interconnection to be [e; y], and the input to be r.
Define two constraints among the variables: e = r-y, and y = (2/s) e. Get
the transfer function representation of the relationship between the input (r)
and the output [e; y].
r = icsignal(1);
e = icsignal(1);
y = icsignal(1);
M = iconnect;
M.Input = r;
M.Output = [e;y];
M.Equation{1} = equate(e,r-y);
M.Equation{2} = equate(y,tf(2,[1 0])*e);
tf(M.System)

10-133

iconnect

The transfer functions from input to outputs are

#1:

s
----s + 2

#2:

2
----s + 2

By not explicitly introducing e, this can be done more concisely with only one
equality constraint.
r = icsignal(1);
y = icsignal(1);
N = iconnect;
N.Input = r;
N.Output = [r-y;y];
N.Equation{1} = equate(y,tf(2,[1 0])*(r-y));
tf(N.System)

You have created the same transfer functions from input to outputs.
#1:

s
----s + 2

#2:

2
----s + 2

You can also specify uncertain, multivariable interconnections using iconnect.

Consider two uncertain motor/generator constraints among 4 variables
[V;I;T;W], V-R*I-K*W=0, and T=K*I. Find the uncertain 2x2 matrix B so that
[V;T] = B*[W;I].
R
K
V
I
T
W

10-134

=
=
=
=
=
=

ureal('R',1,'Percentage',[-10 40]);
ureal('K',2e-3,'Percentage',[-30 30]);
icsignal(1);
icsignal(1);
icsignal(1);
icsignal(1);

iconnect

M = iconnect;
M.Input = [W;I];
M.Output = [V;T];
M.Equation{1} = equate(V-R*I-K*W,iczero(1));
M.Equation{2} = equate(T,K*I);
B = M.System
UMAT: 2 Rows, 2 Columns
K: real, nominal = 0.002, variability = [-30 30]%, 2 occurrences
R: real, nominal = 1, variability = [-10 40]%, 1 occurrence
B.NominalValue
ans =
0.0020
1.0000
0
0.0020

A simple system interconnection, identical to the system illustrated in the

sysic reference pages. Consider a three-input, two-output state-space matrix
T,







 

which has internal structure


y2 
y1


g; g
6+ 6++
noise

57.3

setpoint




wt

deltemp

act

10-135

iconnect

P = rss(3,2,2);
K = rss(1,1,2);
A = rss(1,1,1);
W = rss(1,1,1);
M = iconnect;
noise = icsignal(1);
deltemp = icsignal(1);
setpoint = icsignal(1);
yp = icsignal(2);
rad2deg = 57.3
rad2deg =
57.3000
M.Equation{1} =
equate(yp,P*[W*deltemp;A*K*[noise+yp(2);setpoint]]);
M.Input = [noise;deltemp;setpoint];
M.Output = [rad2deg*yp(1);setpoint-yp(2)];
T = M.System;
size(T)
State-space model with 2 outputs, 3 inputs, and 6 states.

Algorithm

Each equation represents an equality constraint among the variables. You

choose the input and output variables, and the imp2exp function makes the
implicit relationship between them explicit.

Limitations

The syntax for iconnect objects and icsignals is very flexible. Without care,
you can build inefficient (i.e., nonminimal) representations where the state
dimension of the interconnection is greater than the sum of the state
dimensions of the components. This is in contrast to sysic. In sysic, the
syntax used to specify inputs to systems (the input_to_ListedSubSystemName
variable) forces you to include each subsystem of the interconnection only once
in the equations. Hence, interconnections formed with sysic are
componentwise minimal. That is, the state dimension of the interconnection
equals the sum of the state dimensions of the components.

See Also

equate
icsignal
sysic

10-136

Equates expressions for icsignal objects

Constructs an icsignal object
Constructs system interconnection

icsignal

Purpose

10icsignal

Create icsignal object

Syntax

v = icsignal(n);
v = icsignal(n,'name');

Description

icsignal creates an icsignal object, which is a symbolic column vector. The

icsignal object is used with iconnect objects to specify signal constraints

described by the interconnection of components.

v = icsignal(n) creates an icsignal object of vector length n. The value of n
must be a nonnegative integer. icsignal objects are symbolic column vectors,
used in conjunction with iconnect (interconnection) objects to specify the
signal constraints described by an interconnection of components.
v = icsignal(n,'name') creates an icsignal object of dimension n, with
internal name identifier given by the character string argument name.

See Also

iconnect
sysic

Equates expressions for icsignal objects

Constructs system interconnections

10-137

imp2exp

Purpose

10imp2exp

Convert implicit linear relationship to explicit input-output relationship

Syntax

B = imp2exp(A,yidx,uidx)

Description

B = imp2exp(A,yidx,uidx) transforms a linear constraint between variables

Y and U of the form A(:,[yidx;uidx])*[Y;U] = 0 into an explicit input/output
relationship Y = B*U. The vectors yidx and uidx refer to the columns (inputs)
of A as referenced by the explicit relationship for B.

The constraint matrix A can be a double, ss, tf, zpk and frd object as well
as an uncertain object, including umat, uss and ufrd. The result B will be of the
same class.

Examples

Scalar Algebraic Constraint

Consider the constraint 4y + 7u = 0. Solving for y gives y = 1.75u. You form
the equation using imp2exp:
A = [4 7];
Yidx = 1;
Uidx = 2;

and then
B = imp2exp(A,Yidx,Uidx)
B =
-1.7500

yields B equal to -1.75.

Matrix Algebraic Constraint

Consider two motor/generator constraints among 4 variables [V;I;T;W],
namely [1 -1 0 -2e-3;0 -2e-3 1 0]*[V;I;T;W] = 0. You can find the 2-by-2
matrix B so that [V;T] = B*[W;I] using imp2exp.
A = [1 -1 0 -2e-3;0 -2e-3 1 0];
Yidx = [1 3];
Uidx = [4 2];
B = imp2exp(A,Yidx,Uidx)
B =
0.0020
1.0000
0
0.0020

10-138

imp2exp

You can find the 2-by-2 matrix C so that [I;W] = C*[T;V]

Yidx = [2 4];
Uidx = [3 1];
C = imp2exp(A,Yidx,Uidx)
C =
500
0
-250000
500

Uncertain Matrix Algebraic Constraint

Consider two uncertain motor/generator constraints among 4 variables
[V;I;T;W], namely [1 -R 0 -K;0 -K 1 0]*[V;I;T;W] = 0. You can find the
uncertain 2-by-2 matrix B so that [V;T] = B*[W;I].
R = ureal('R',1,'Percentage',[-10 40]);
K = ureal('K',2e-3,'Percentage',[-30 30]);
A = [1 -R 0 -K;0 -K 1 0];
Yidx = [1 3];
Uidx = [4 2];
B = imp2exp(A,Yidx,Uidx)
UMAT: 2 Rows, 2 Columns
K: real, nominal = 0.002, variability = [-30 30]%, 2 occurrences
R: real, nominal = 1, variability = [-10 40]%, 1 occurrence

Scalar Dynamic System Constraint

Consider a standard single-loop feedback connection of controller C and an
uncertain plant P, described by the equations e=r-y, u=Ce; f=d+u; y=Pf.










P = tf([1],[1 0]);
C = tf([2*.707*1 1^2],[1 0]);
A = [1 -1 0 0 0 -1;0 -C 1 0 0 0;0 0 -1 -1 1 0;0 0 0 0 -P 1];
OutputIndex = [6;3;2;5]; % [y;u;e;f]
InputIndex = [1;4];
% [r;d]
Sys = imp2exp(A,OutputIndex,InputIndex);

10-139

imp2exp

Sys.InputName = {'r';'d'};
Sys.OutputName = {'y';'u';'e';'f'};
pole(Sys)
ans =
-0.7070 + 0.7072i
-0.7070 - 0.7072i
step(Sys)
Step Response

From: r

From: d

2
1.5

To: y

1
0.5
0
0.5
2

To: u

1
0

Amplitude

1
2
1

To: e

0.5
0
0.5
1
2

To: f

1
0
1
2

80
1
Time (sec)

Algorithm

The number of rows of A must equal the length of yidx.

See Also

iconnect
inv

10-140

Equates expressions for icsignal objects

Forms the system inverse

imp2ss

Purpose

10imp2ss

Convert impulse response to approximate state-space realization

Syntax

[a,b,c,d,totbnd,hsv] = imp2ss(y)
[a,b,c,d,totbnd,hsv] = imp2ss(y,ts,nu,ny,tol)
[ss,totbnd,hsv] = imp2ss(imp)
[ss,totbnd,hsv] = imp2ss(imp,tol)

Description

The function imp2ss produces an approximate state-space realization of a

given impulse response
imp=mksys(y,t,nu,ny,'imp');

using the Hankel SVD method proposed by S. Kung [2]. A continuous-time

realization is computed via the inverse Tustin transform (using bilin) if t is
positive; otherwise a discrete-time realization is returned. In the SISO case the
variable y is the impulse response vector; in the MIMO case y is an N+1-column
matrix containing N + 1 time samples of the matrix-valued impulse response
H0, , HN of an nu-input, ny-output system stored row-wise:
y = [ H 0 ( : ) ;H 2 ( : ) ;H 3 ( : ) ; ;H N ( : ) ]
The variable tol bounds the H norm of the error between the approximate
realization (a, b, c, d) and an exact realization of y; the order, say n, of the
realization (a, b, c, d) is determined by the infinity norm error bound specified
by the input variable tol. The inputs ts, nu, ny, tol are optional; if not
present they default to the values ts = 0, nu = 1, ny = (number of rows of
y)/nu, tol = 0.01 1. The output hsv = [ 1 , 2 , ] returns the singular
values (arranged in descending order of magnitude) of the Hankel matrix:

H1 H2 H3 HN
H2 H3 H4 0
H3 H4 H5 0
...

...
...

...
...

H N 0 0s
Denoting by GN a high-order exact realization of y, the low-order approximate
model G enjoys the H norm bound

10-141

imp2ss

G G N totbnd
where
N

totbnd = 2

i = n+1

Algorithm

The realization (a, b, c, d) is computed using the Hankel SVD procedure

proposed by Kung [2] as a method for approximately implementing the
classical Hankel factorization realization algorithm. Kungs SVD realization
procedure was subsequently shown to be equivalent to doing balanced
truncation (balmr) on an exact state-space realization of the finite impulse
response {y(1),.y(N)} [3]. The infinity norm error bound for discrete balanced
truncation was later derived by Al-Saggaf and Franklin [1]. The algorithm is
as follows:
1 Form the Hankel matrix from the data y.
2 Perform SVD on the Hankel matrix

= UV = [ U 1 U 2 ]

1 0

V 1

0 2 V 2

= U 1 1 V 1

where 1 has dimension n n and the entries of 2 are nearly zero. U1 and
V1 have ny and nu columns, respectively.
3 Partition the matrices U1 and V1 into three matrix blocks:

U 11

V 11

U 1 = U 12 ; V 12
U 13 V 13
ny n

nu n

where U 11 , U 13 C
and V 11 , V 13 C
4 A discrete state-space realization is computed as

10-142

imp2ss

1 2

A = 1

1 2

B = 1

12

U 1

V 11
1 2

C = U 11 1
D = H0
where
U=

U 11

U 12

U 12

U 13

5 If the sampling time t is greater than zero, then the realization is converted

to continuous time via the inverse of the Tustin transform

2z 1
s = --- ------------ ;
tz+1
Otherwise, this step is omitted and the discrete-time realization calculated
in Step 4 is returned.

References

[1] Al-Saggaf, U.M., and G.F. Franklin, An Error Bound for a Discrete
Reduced Order Model of a Linear Multivariable System, IEEE Trans. on
Autom. Contr., AC-32, 1987, p. 815-819.
[2] Kung, S.Y., A New Identification and Model Reduction Algorithm via
Singular Value Decompositions, Proc.Twelth Asilomar Conf. on Circuits,
Systems and Computers, November 6-8, 1978, p. 705-714.
[3] Silverman, L.M., and M. Bettayeb, Optimal Approximation of Linear
Systems, Proc. American Control Conf., San Francisco, CA, 1980.

10-143

ispsys

Purpose

10ispsys

True for parameter-dependent systems

Syntax

bool = ispsys(sys)

Description

bool = ispsys(sys) returns 1 if sys is a polytopic or parameter-dependent

system.

See Also

10-144

psys
psinfo

Specification of uncertain state-space models

Inquire about polytopic or parameter-dependent
systems created with psys

isuncertain

Purpose

10isuncertain

True for uncertain systems

Syntax

B = isuncertain(A)

Description

Returns true if input argument is uncertain, false otherwise. Uncertain

classes are umat, ufrd, uss, ureal, ultidyn, ucomplex, ucomplexm, and udyn.

Examples

In this example, you verify the correct operation of isuncertain on double,

ureal, ss, and uss objects.
isuncertain(rand(3,4))
ans =
0
isuncertain(ureal('p',4))
ans =
1
isuncertain(rss(4,3,2))
ans =
0
isuncertain(rss(4,3,2)*[ureal('p1',4) 6;0 1])
ans =
1

Limitations

isuncertain only checks the class of the input argument, and does not actually
verify that the input argument is truly uncertain. Create a umat by lifting a
constant (i.e., not-uncertain) matrix to the umat class.
A = umat([2 3;4 5;6 7]);

Note that although A is in class umat, it is not actually uncertain. Nevertheless,

based on class, the result of isuncertain(A) is true.
isuncertain(A)
ans =
1

The result of simplify(A) is a double, and hence not uncertain.

isuncertain(simplify(A))
ans =
0

10-145

isuncertain

The number of rows of A must equal the length of yidx.

10-146

lftdata

Purpose

10lftdata

Decompose uncertain object into fixed normalized and fixed uncertain parts

Syntax

[M,Delta] = lftdata(A);
[M,Delta] = lftdata(A,List);
[M,Delta,Blkstruct] = lftdata(A);
[M,Delta,Blkstruct,Normunc] = lftdata(A);

Description

lftdata decomposes an uncertain object into a fixed certain part and a

normalized uncertain part. lftdata can also partially decompose an uncertain
object into an uncertain part and a normalized uncertain part. Uncertain
objects (umat, ufrd, uss) are represented as certain (i.e., not-uncertain)
objects in feedback with block-diagonal concatenations of uncertain elements.
[M,Delta] = lftdata(A) separates the uncertain object A into a certain object
M and a normalized uncertain matrix Delta such that A is equal to
lft(Delta,M), as shown below.





If A is a umat, then M will be double; if A is a uss, then M will be ss; if A is a ufrd,
then M will be frd. In all cases, Delta is a umat.
[M,Delta] = lftdata(A,List) separates the uncertain object A into an
uncertain object M, in feedback with a normalized uncertain matrix Delta. List
is a cell (or char) array of names of uncertain elements of A that make up Delta.
All other uncertainty in A remains in M.
lftdata(A,fieldnames(A.Uncertainty)) is the same as lftdata(A).
[M,DELTA,BLKSTRUCT] = LFTDATA(A) returns an N-by-1 structure array
BLKSTRUCT, where BLKSTRUCT(i) describes the i-th normalized uncertain

element. This uncertainty description can be passed directly to the low-level

structured singular value analysis function MUSSV.

10-147

lftdata

[M,DELTA,BLKSTRUCT,NORMUNC] = LFTDATA(A) returns the cell array NORMUNC

of normalized uncertain elements. Each normalized element has the string

'Normalized' appended to its original name to avoid confusion. Note that
LFT(BLKDIAG(NORMUNC{:}),M) is equivalent to A.

Examples

Create an uncertain matrix A with 3 uncertain parameters p1, p2 and p3. You
can decompose A into it certain, M, and normalized uncertain parts, Delta.
p1 = ureal('p1',-3,'perc',40);
p2 = ucomplex('p2',2);
A = [p1 p1+p2;1 p2];
[M,Delta] = lftdata(A);

You can inspect the difference between the original uncertain matrix, A, and
the result formed by combining the two results from the decomposition.
simplify(A-lft(Delta,M))
ans =
0
0
0
0
M
M =
0
0
1.0954
0
0
0
1.0954
1.0000
-3.0000
0
1.0000
1.0000

1.0954
1.0000
-1.0000
2.0000

You can check the worst-case norm of the uncertain part using wcnorm.
Compare samples of the uncertain part A with the uncertain matrix A.
wcn = wcnorm(Delta)
wcn =
lbound: 1.0000
ubound: 1.0001
usample(Delta,5)
ans(:,:,1) =
0.8012
0
ans(:,:,2) =
0.4919
0

10-148

0
0.2499 + 0.6946i
0
0.2863 + 0.6033i

lftdata

ans(:,:,3) =
-0.1040
0
ans(:,:,4) =
0.8296
0
ans(:,:,5) =
0.6886
0

0
0.7322 - 0.3752i
0
0.6831 + 0.1124i
0
0.0838 + 0.3562i

Uncertain Systems
Create an uncertain matrix A with 2 uncertain real parameters v1 and v2 and
create an uncertain system G using A as the dynamic matrix and simple
matrices for the input and output.
A = [ureal('p1',-3,'perc',40) 1;1 ureal('p2',-2)];
sys = ss(A,[1;0],[0 1],0);
sys.InputGroup.ActualIn = 1;
sys.OutputGroup.ActualOut = 1;

You can decompose G into a certain system, Msys, and a normalized uncertain
matrix, Delta. You can see from Msys that it is certain and that the input and
output groups have been adjusted.
[Msys,Delta] = lftdata(sys);
Msys
a =
x1
x2

x1
-3
1

x2
1
-2

b =
x1
x2

u1
1.095
0

u2
0
1

x1

x2

u3
1
0

c =

10-149

lftdata

y1
y2
y3

1.095
0
0

0
1
1

d =
y1
y2
y3

u1
0
0
0

u2
0
0
0

Input groups:
Name
ActualIn
p1_NC
p2_NC
Output groups:
Name
ActualOut
p1_NC
p2_NC

u3
0
0
0

Channels
3
1
2

Channels
3
1
2

Continuous-time model.

You can compute the norm on samples of the difference between the original
uncertain matrix and the result formed by combining Msys and Delta.
norm(usample(sys-lft(Delta,Msys),'p1',4,'p2',3),'inf')
ans =
0
0
0
0
0
0
0
0
0
0
0
0

Partial Decomposition
Create an uncertain matrix A and derive an uncertain matrix B using an
implicit-to-explicit conversion, imp2exp. Note that B has 2 uncertain
parameters R and K. You can decompose B into certain, M, and normalized
uncertain parts, Delta.

10-150

lftdata

R = ureal('R',1,'Percentage',[-10 40]);
K = ureal('K',2e-3,'Percentage',[-30 30]);
A = [1 -R 0 -K;0 -K 1 0];
Yidx = [1 3];
Uidx = [4 2];
B = imp2exp(A,Yidx,Uidx);
[M,Delta] = lftdata(B);

The same operation can be performed by defining the uncertain parameters, K

and R, to be extracted.
[MK,DeltaR] = lftdata(B,'R');
MK
UMAT: 3 Rows, 3 Columns
K: real, nominal = 0.002, variability = [-30 30]%, 2 occurrences
[MR,DeltaK] = lftdata(B,'K');
MR
UMAT: 4 Rows, 4 Columns
R: real, nominal = 1, variability = [-10 40]%, 1 occurrence
simplify(B-lft(Delta,M))
ans =
0
0
0
0
simplify(B-lft(DeltaR,MK))
ans =
0
0
0
0
simplify(B-lft(DeltaK,MR))
ans =
0
0
0
0

Sample and inspect the uncertain part as well as the difference between the
original uncertain matrix and the sampled matrix. You can see the result
formed by combining the two results from the decomposition.
[Mall,Deltaall] = lftdata(B,{'K';'R'});
simplify(Mall)-M
ans =
0
0
0
0
0

10-151

lftdata

0
0
0
0

See Also

10-152

lft
ssdata

0
0
0
0

0
0
0
0

0
0
0
0

0
0
0
0

Forms Redheffer star product of systems

Returns uncertain state-space data

lmiedit

Purpose

10lmiedit

Specify or display systems of LMIs as MATLAB expressions

Syntax

lmiedit

Description

lmiedit is a graphical user interface for the symbolic specification of LMI

problems. Typing lmiedit calls up a window with two editable text areas and
various buttons. To specify an LMI system,
1 Give it a name (top of the window).
2 Declare each matrix variable (name and structure) in the upper half of the

window. The structure is characterized by its type (S for symmetric block

diagonal, R for unstructured, and G for other structures) and by an
additional structure matrix similar to the second input argument of lmivar.
Please use one line per matrix variable in the text editing areas.
3 Specify the LMIs as MATLAB expressions in the lower half of the window.

An LMI can stretch over several lines. However, do not specify more than
one LMI per line.
Once the LMI system is fully specified, you can perform the following
operations by pressing the corresponding button:
Visualize the sequence of lmivar/lmiterm commands needed to describe this
LMI system (view commands buttons)
Conversely, display the symbolic expression of the LMI system produced by
a particular sequence of lmivar/lmiterm commands (click the describe...
buttons)
Save the symbolic description of the LMI system as a MATLAB string (save
button). This description can be reloaded later on by pressing the load
button
Read a sequence of lmivar/lmiterm commands from a file (read button). The
matrix expression of the LMI system specified by these commands is then
displayed by clicking describe the LMIs...
Write in a file the sequence of lmivar/lmiterm commands needed to specify
a particular LMI system (write button)
Generate the internal representation of the LMI system by pressing create.
The result is written in a MATLAB variable with the same name as the LMI
system

10-153

lmiedit

Remark

Editable text areas have built-in scrolling capabilities. To activate the scroll
mode, click in the text area, maintain the mouse button down, and move the
mouse up or down. The scroll mode is only active when all visible lines have
been used.

See Also

lmivar
lmiterm
newlmi
lmiinfo

10-154

Specify the matrix variables in an LMI problem

Specify the term content of LMIs
Attach an identifying tag to LMIs
Interactively retrieve information about the variables
and term content of LMIs

lmiinfo

Purpose

10lmiinfo

Information about existing system of LMIs

Syntax

lmiinfo(lmisys)

Description

lmiinfo provides qualitative information about the system of LMIs lmisys.

This includes the type and structure of the matrix variables, the number of
diagonal blocks in the inner factors, and the term content of each block.
lmiinfo is an interactive facility where the user seeks specific pieces of

information. General LMIs are displayed as

N' * L(x) * N < M' * R(x) * M

where N,M denote the outer factors and L,R the left and right inner factors. If
the outer factors are missing, the LMI is simply written as
L(x) < R(x)

If its right-hand side is zero, it is displayed as

N' * L(x) * N < 0

Information on the block structure and term content of L(x) and R(x) is also
available. The term content of a block is symbolically displayed as
C1 + A1*X2*B1 + B1'*X2*A1' + a2*X1 + x3*Q1

with the following conventions:

X1, X2, x3 denote the problem variables. Uppercase X indicates matrix
variables while lowercase x indicates scalar variables. The labels 1,2,3 refer
to the first, second, and third matrix variable in the order of declaration.
Cj refers to constant terms. Special cases are I and I (I = identity matrix).
Aj, Bj denote the left and right coefficients of variable terms. Lowercase
letters such as a2 indicate a scalar coefficient.
Qj is used exclusively with scalar variables as in x3*Q1.
The index j in Aj, Bj, Cj, Qj is a dummy label. Hence C1 may appear in
several blocks or several LMIs without implying any connection between the
corresponding constant terms. Exceptions to this rule are the notations
A1*X2*A1' and A1*X2*B1 + B1'*X2'*A1' which indicate symmetric terms and
symmetric pairs in diagonal blocks.

10-155

lmiinfo

Examples

Consider the LMI

T
T T
0 2X + A YB + B Y A + I XC

C X
zI
where the matrix variables are X of Type 1, Y of Type 2, and z scalar. If this
LMI is described in lmis, information about X and the LMI block structure can
be obtained as follows:
lmiinfo(lmis)
LMI ORACLE
-----This is a system of 1 LMI with 3 variable matrices
Do you want information on
(v) matrix variables (l) LMIs (q) quit
?> v
Which variable matrix (enter its index k between 1 and 3) ? 1
X1 is a 2x2 symmetric block diagonal matrix
its (1,1)-block is a full block of size 2
------This is a system of 1 LMI with 3 variable matrices
Do you want information on
(v) matrix variables (l) LMIs (q) quit
?> l
Which LMI (enter its number k between 1 and 1) ? 1
This LMI is of the form
0 < R(x)
where the inner factor(s) has 2 diagonal block(s)

10-156

lmiinfo

Do you want info on the right inner factor ?

(w) whole factor (b) only one block
(o) other LMI (t) back to top level
?> w
Info about the right inner factor
block (1,1) : I + a1*X1 + A2*X2*B2 + B2'*X2'*A2'
block (2,1) : A3*X1
block (2,2) : x3*A4
(w) whole factor (b) only one block
(o) other LMI (t) back to top level
-----This is a system of 1 LMI with 3 variable matrices
Do you want information on
(v) matrix variables (l) LMIs (q) quit
?> q
It has been a pleasure serving you!

Note that the prompt symbol is ?> and that answers are either indices or
letters. All blocks can be displayed at once with option (w), or you can prompt
for specific blocks with option (b).

Remark

lmiinfo does not provide access to the numerical value of LMI coefficients.

See Also

decinfo
lminbr

Describe how the entries of a matrix variable X relate

to the decision variables
Return the number of LMIs in an LMI system

10-157

lmiinfo

matnbr
decnbr

10-158

Return the number of matrix variables in a system of

LMIs
Give the total number of decision variables in a system
of LMIs

lminbr

Purpose

10lminbr

Number of LMIs in LMI system

Syntax

k = lminbr(lmisys)

Description

lminbr returns the number k of linear matrix inequalities in the LMI problem
described in lmisys.

See Also

lmiinfo
matnbr

Interactively retrieve information about the variables

and term content of LMIs
Return the number of matrix variables in a system of
LMIs

10-159

lmireg

Purpose

10lmireg

Specify LMI regions for pole placement

Syntax

region = lmireg
region = lmireg(reg1,reg2,...)

Description

lmireg is an interactive facility to specify the LMI regions involved in

multi-objective H synthesis with pole placement constraints (see msfsyn).

Recall that an LMI region is any convex subset D of the complex plane that can
be characterized by an LMI in z and z , i.e.,
D = {z C : L + Mz + MT z < 0}
for some fixed real matrices M and L = LT. This class of regions encompasses
half planes, strips, conic sectors, disks, ellipses, and any intersection of the
above.
Calling lmireg without argument starts an interactive query/answer session
where you can specify the region of your choice. The matrix region = [L, M] is
returned upon termination. This matrix description of the LMI region can be
passed directly to msfsyn for synthesis purposes.
The function lmireg can also be used to intersect previously defined LMI
regions reg1, reg2,.... The output region is then the [L, M] description of
the intersection of these regions.

See Also

10-160

msfsyn

Multi-model/multi-objective state-feedback synthesis

lmiterm

Purpose

10lmiterm

Add new term to existing LMI

Syntax

lmiterm(termID,A,B,flag)

Description

lmiterm specifies the term content of an LMI one term at a time. Recall that
LMI term refers to the elementary additive terms involved in the block-matrix
expression of the LMI. Before using lmiterm, the LMI description must be
initialized with setlmis and the matrix variables must be declared with
lmivar. Each lmiterm command adds one extra term to the LMI system
currently described.

LMI terms are one of the following entities:

outer factors
constant terms (fixed matrices)
variable terms AXB or AXTB where X is a matrix variable and A and B are
given matrices called the term coefficients.
When describing an LMI with several blocks, remember to specify only the
terms in the blocks on or below the diagonal (or equivalently, only the terms
in blocks on or above the diagonal). For instance, specify the blocks (1,1), (2,1),
and (2,2) in a two-block LMI.
In the calling of limterm, termID is a four-entry vector of integers specifying
the term location and the matrix variable involved:
+p

termID (1) =

-p

where positive p is for terms on the left-hand side of the p-th LMI and negative
p i s for terms on the right-hand side of the p-th LMI.
Recall that, by convention, the left-hand side always refers to the smaller side
of the LMI. The index p is relative to the order of declaration and corresponds
to the identifier returned by newlmi.
[ 0, 0 ] for outer factors

termID (2:3) = [ i, j ] for terms in the ( i, j )-th

block of the left or right inner factor

10-161

lmiterm

0 for outer factors

termID (4) = x for variable terms AXB

T
x for variable terms AX B
where x is the identifier of the matrix variable X as returned by lmivar.
The arguments A and B contain the numerical data and are set according to:
Type of Term

outer factor N

matrix value of N

omit

constant term C

matrix value of C

omit

variable term
AXB or AXTB

matrix value of A
(1 if A is absent)

matrix value of B
(1 if B is absent)

Note that identity outer factors and zero constant terms need not be specified.
The extra argument flag is optional and concerns only conjugated expressions
of the form
(AXB) + (AXB)T = AXB + BTX(T)AT
in diagonal blocks. Setting flag = 's' allows you to specify such expressions
with a single lmiterm command. For instance,
lmiterm([1 1 1 X],A,1,'s')

adds the symmetrized expression AX + XTAT to the (1,1) block of the first LMI
and summarizes the two commands
lmiterm([1 1 1 X],A,1)
lmiterm([1 1 1 X],1,A')

Aside from being convenient, this shortcut also results in a more efficient
representation of the LMI.

10-162

lmiterm

Examples

Consider the LMI

T
T T
2AX 2 A x 3 E + DD B X 1

X1 B
I

T T
T

T CX 1 C + CX 1 C
0
< M

0
f X2

where X1, X2 are matrix variables of Types 2 and 1, respectively, and x3 is a

scalar variable (Type 1).
After initializing the LMI description with setlmis and declaring the matrix
variables with lmivar, the terms on the left-hand side of this LMI are specified
by
lmiterm([1
lmiterm([1
lmiterm([1
lmiterm([1
lmiterm([1

1
1
1
2
2

1
1
1
1
2

X2],2*A,A') % 2*A*X2*A'
x3],-1,E) % -x3*E
0],D*D') % D*D'
-X1],1,B) % X1'*B
0],-1) % -I

Here X1, X2, X3 should be the variable identifiers returned by lmivar.

Similarly, the term content of the right-hand side is specified by
lmiterm([-1 0 0 0],M) % outer factor M
lmiterm([-1 1 1 X1],C,C','s') % C*X1*C'+C*X1'*C'
lmiterm([-1 2 2 X2],-f,1) % -f*X2

Note that CX1CT + CX1TCT is specified by a single lmiterm command with the
flag 's' to ensure proper symmetrization.

See Also

setlmis
lmivar
getlmis
lmiedit
newlmi

Initialize the description of an LMI system

Specify the matrix variables in an LMI problem
Get the internal description of an LMI system
Specify or display systems of LMIs as MATLAB
expressions
Attach an identifying tag to LMIs

10-163

lmivar

Purpose

10lmivar

Specify matrix variables in existing LMI system

Syntax

X = lmivar(type,struct)
[X,n,sX] = lmivar(type,struct)

Description

lmivar defines a new matrix variable X in the LMI system currently described.
The optional output X is an identifier that can be used for subsequent reference
to this new variable.

The first argument type selects among available types of variables and the
second argument struct gives further information on the structure of X
depending on its type. Available variable types include:
type=1: Symmetric matrices with a block-diagonal structure. Each diagonal
block is either full (arbitrary symmetric matrix), scalar (a multiple of the
identity matrix), or identically zero.
If X has R diagonal blocks, struct is an R-by-2 matrix where
struct(r,1) is the size of the r-th block.
struct(r,2) is the type of the r-th block (1 for full, 0 for scalar, 1 for zero
block).
type=2: Full m-by-n rectangular matrix. Set struct = [m,n] in this case.
type=3: Other structures. With Type 3, each entry of X is specified as zero or
xn where xn is the n-th decision variable.
Accordingly, struct is a matrix of the same dimensions as X such that
struct(i,j)=0 if X(i, j) is a hard zero
struct(i,j)=n if X(i, j) = xn
struct(i,j)= n if X(i, j) = xn
Sophisticated matrix variable structures can be defined with Type 3. To specify
a variable X of Type 3, first identify how many free independent entries are
involved in X. These constitute the set of decision variables associated with X.
If the problem already involves n decision variables, label the new free
variables as xn+1, . . ., xn+p. The structure of X is then defined in terms of
xn+1, . . ., xn+p as indicated above. To help specify matrix variables of Type 3,
lmivar optionally returns two extra outputs: (1) the total number n of scalar

10-164

lmivar

decision variables used so far and (2) a matrix sX showing the entry-wise
dependence of X on the decision variables x1, . . ., xn.

Example 1

Consider an LMI system with three matrix variables X1, X2, X3 such that
X1 is a 3 3 symmetric matrix (unstructured).
X2 is a 2 4 rectangular matrix (unstructured).
X3 =
0 0

0 1 0

0 0 2 I2

where is an arbitrary 5 5 symmetric matrix, 1 and 2 are scalars, and I2

denotes the identity matrix of size 2.
These three variables are defined by
setlmis([])
X1 = lmivar(1,[3 1]) % Type 1
X2 = lmivar(2,[2 4]) % Type 2 of dim. 2x4
X3 = lmivar(1,[5 1;1 0;2 0]) % Type 1

The last command defines X3 as a variable of Type 1 with one full block of size
5 and two scalar blocks of sizes 1 and 2, respectively.

Example 2

Combined with the extra outputs n and sX of lmivar, Type 3 allows you to
specify fairly complex matrix variable structures. For instance, consider a
matrix variable X with structure
X 0
X = 1
0 X
2

where X1 and X2 are 2-by-3 and 3-by-2 rectangular matrices, respectively. You
can specify this structure as follows:
1 Define the rectangular variables X1 and X2 by

setlmis([])

10-165

lmivar

[X1,n,sX1] = lmivar(2,[2 3])

[X2,n,sX2] = lmivar(2,[3 2])

The outputs sX1 and sX2 give the decision variable content of X1 and X2:
sX1
sX1 =
1
4

2
5

3
6

sX2
sX2 =
7
9
11

8
10
12

For instance, sX2(1,1)=7 means that the (1,1) entry of X2 is the seventh
decision variable.
2 Use Type 3 to specify the matrix variable X and define its structure in terms

of those of X1 and X2:

[X,n,sX] = lmivar(3,[sX1,zeros(2);zeros(3),sX2])

The resulting variable X has the prescribed structure as confirmed by

sX
sX =
1
4
0
0
0

See Also

setlmis
lmiterm
getlmis
lmiedit
skewdec

10-166

2
5
0
0
0

3
6
0
0
0

0
0
0
0
7
8
9
10
11 12

Initialize the description of an LMI system

Specify the term content of LMIs
Get the internal description of an LMI system
Specify or display systems of LMIs as MATLAB
expressions
Form a skew-symmetric matrix

lmivar

delmvar
setmvar

Delete one of the matrix variables of an LMI problem

Instantiate a matrix variable and evaluate all LMI
terms involving this matrix variable

10-167

loopmargin

Purpose

10loopmargin

Analyze feedback loop

Syntax

[sm,dm,mm] = loopmargin(L)
[cmi,dmi,mmi,cmo,dmo,mmo,mmio] = loopmargin(P,C)

Description

[sm,dm,mm] = loopmargin(L) analyzes the multivariable feedback loop

consisting of the loop transfer matrix L (size N-by-N) in negative feedback with
an N-by-N identity matrix.
sm, or single-loop margin, is an N-by-1 structure corresponding to

loop-at-a-time gain and phase margins for each channel.

dm is an N-by-1 structure corresponding to loop-at-a-time disk gain and phase
margins for each channel. The disk margin for the i-th feedback channel
defines a circular region centered on the negative real axis at the average
GainMargin (GM), e.g. , (GMlow+GMhigh)/2, such that L(i,i) does not enter that
region. Gain and phase disk margin bounds are derived from the radius of the
circle, calculated based on the balanced sensitivity function (see Algorithm
section).
mm, or multiloop disk margin, is a structure corresponding to simultaneous,
independent, variations in the individual channels of loop transfer matrix L. mm

calculates the largest region such that for all gain and phase variations,
occurring independently in each channel, lie inside the region, that the
closed-loop system is stable. Note that mm is a single structure, independent of
because the number of channels, variations in all channels are handled
simultaneously. As in the case for disk margin, the guaranteed bounds are
calculated based on a balanced sensitivity function.
The wcmargin command is used to analyze single-loop margins in the presence
of the worst-case uncertainty. The disk margin and multiloop disk margin
calculations are performed in the frequency domain. Unless L is an frd object,
the frequency range and number of points used to calculate the dm and mm
margins are chosen automatically (see bode command).
The output arguments can be limited to only those requested using an optional
character string argument. [M1,M2] = loopmargin(L,'m,c') returns the
multiloop disk margin ('m') in M1, and the classical margins ('c') in M2.
Use'd' to specify the disk margin. This optional second argument may be any
combination, in any order, of the 3 characters 'c', 'd', and 'm'.

10-168

loopmargin

[cmi,dmi,mmi,cmo,dmo,mmo,mmio] = loopmargin(P,C) analyzes the

multivariable feedback loop consisting of the controller C in negative feedback
with the plant, P. C should only be the compensator in the feedback path,

without reference channels, if it is a 2-dof architecture. That is, if the

closed-loop system has a 2-dof architecture the reference channel of the
controller should be eliminated, resulting in a 1-dof architecture, as shown.

 

 
   

 
   

cmi,dmi and mmi structures correspond to the loop-at-a-time gain and phase

margins, disk margins and multiloop channel margins at the plant input
respectively. The structures cmo, dmo and mmo have the same fields as described
for cmi, dmi and mmi though they correspond to the plant outputs. mmio, or
multiinput/multioutput margins, is a structure corresponding to simultaneous,
independent, variations in all the individual input and output channels of the
feedback loops. mmio has the same fields as mmi and mmo.
If the closed-loop system is an ss/tf/zpk, the frequency range and number of
points used to calculate cm, dm and mm margins are chosen automatically.
The output arguments can be limited to only those requested using an optional
character string argument. [M1,M2,M3] = loopmargin(P,C,'mo,ci,mm')
returns the multiloop disk margin at the plant output ('mo') in M1, the
classical margins at the plant input ('ci') in M2, and the disk margins for
simultaneous, independent variations in all input and output channels ('mm')
in M3. Use'd' to specify the disk margin. This optional third argument may be
any combination, in any order, of the 7 character pairs 'ci', 'di', mi', 'co',
do', 'mo', and 'mm'.

10-169

loopmargin
Basic Syntax
[sm,dm,mm] = loopmargin(L) sm is calculated using the allmargin command
and has the same fields as allmargin. The sm is a structure with the following

fields:
Field

Description

GMFrequency

All -180 deg crossover frequencies (in rad/s)

GainMargin

Corresponding gain margins (GM = 1/L where L is

the gain at crossover)

PhaseMargin

Corresponding phase margins (in degrees)

PMFrequency

All 0 dB crossover frequencies (in rad/s)

DelayMargin

Delay margins (in seconds for continuous-time

systems, and multiples of the sample time for
discrete-time systems)

Stable

1 if nominal closed loop is stable, 0 otherwise. If L is

a frd or ufrd object, the Stable flag is set to NaN.

dm, or Disk Margin, is a structure with the following fields

10-170

Field

Description

GainMargin

Smallest gain variation (GM) such that a disk

centered at the point -(GM(1) + GM(2))/2 would just
touch the loop transfer function

PhaseMargin

Smallest phase variation, in degrees, corresponding

to the disk described in the GainMargin field
(degrees)

Frequency

Associated with GainMargin/PhaseMargin fields

(rad/s)

loopmargin

mm is a structure with the following fields:

Field

Description

GainMargin

Guaranteed bound on simultaneous, independent,

gain variations allowed in all plant channels

PhaseMargin

Guaranteed bound on simultaneous, independent,

phase variations allowed in all plant channels
(degrees)

Frequency

Associated with GainMargin/PhaseMargin fields

(rad/s)

10-171

loopmargin

Examples

MIMO Loop-at-a-Time Margins

This example is designed to illustrate that loop-at-a-time margins (gain, phase,
and/or distance to 1) can be inaccurate measures of multivariable robustness
margins. You will see that margins of the individual loops can be very sensitive
to small perturbations within other loops.
The nominal closed-loop system considered here is as follows

; 6 d- K
;6

G and K are 2-by-2 multiinput/multioutput (MIMO) systems, defined as

2
1
( s + 1 ) ,K=I
G: = ------------------ s
2
2
2
s + ( s + 1 ) s 2

Set := 10, construct G in state-space form, and compute its frequency

response.
a = [0 10;-10 0];
b = eye(2);
c = [1 8;-10 1];
d = zeros(2,2);
G = ss(a,b,c,d);
K = [1 -2;0 1];
[cmi,dmi,mmi,cmo,dmo,mmo,mmio]=loopmargin(G,K);

Note You can compute a limited set of margins using the syntax
[dmo,mmo,mmio]=loopmargin(G,K,'di,do,mm')

First consider the margins at the input to the plant. The first input channel has
infinite gain margin and 90 degrees of phase margin based on the results from

10-172

loopmargin

the allmargin command, cmi(1). The disk margin analysis, dmi, of the first
channel provides similar results.
cmi(1)
ans =
GMFrequency:
GainMargin:
PMFrequency:
PhaseMargin:
DMFrequency:
DelayMargin:
Stable:
dmi(1)
ans =
GainMargin:
PhaseMargin:
Frequency:

[1x0 double]
[1x0 double]
21
90
21
0.0748
1

[0 Inf]
[-90 90]
1.1168

The second input channel has a gain margin of 2.105 and infinite phase margin
based on the single-loop analysis, cmi(2). The disk margin analysis, dmi(2),
which allows for simultaneous gain and phase variations a loop-at-a-time
results in maximum gain margin variations of 0.475 and 2.105 and phase
margin variations of +/- 39.18 degs.
cmi(2)
ans =
GMFrequency:
GainMargin:
PMFrequency:
PhaseMargin:
DMFrequency:
DelayMargin:
Stable:
dmi(2)
ans =
GainMargin:
PhaseMargin:
Frequency:

0
2.1053
[1x0 double]
[1x0 double]
[1x0 double]
[1x0 double]
1

[0.4749 2.1056]
[-39.1912 39.1912]
0.0200

The multiple margin analysis of the plant inputs corresponds to allowing

simultaneous, independent gain and phase margin variations in each channel.

10-173

loopmargin

Allowing independent variation of the input channels further reduces the

tolerance of the closed-loop system to variations at the input to the plant. The
multivariable margin analysis, mmi, leads to a maximum allowable gain
margin variation of 0.728 and 1.373 and phase margin variations of +/- 17.87
deg. Hence even though the first channel had infinite gain margin and 90
degrees of phase margin, allowing variation in both input channels leads to a
factor of two reduction in the gain and phase margin.
mmi
mmi =
GainMargin: [0.7283 1.3730]
PhaseMargin: [-17.8659 17.8659]
Frequency: 9.5238e-004

The guaranteed region of phase and gain variations for the closed-loop system
can be illustrated graphically. The disk margin analysis, dmi(2), indicates the
closed-loop system will remain stable for simultaneous gain variations of 0.475
and 2.105 ( 6.465 dB) and phase margin variations of 39.18 deg in the second
input channel. This is denoted by the region associated with the large ellipse
in the following figure. The multivariable margin analysis at the input to the
plant, mmi, indicates that the closed-loop system will be stable for independent,
simultaneous, gain margin variation up to 0.728 and 1.373 (2.753 dB) and
phase margin variations up to 17.87 deg (the dark ellipse region) in both
input channels.

10-174

loopmargin

Gain and Phase Margin

Guaranteed Disk Margin: Input 2
Guaranteed Multivariable Input Margin

10

Gain Variation (dB)

10

40

30

20

10
0
10
Phase Variation (degrees)

20

30

40

The output channels have single-loop margins of infinite gain and 90 deg phase
variation. The output multivariable margin analysis, mmo, leads to a maximum
allowable gain margin variation of 0.607 and 1.649 and phase margin
variations of +/- 27.53 degs. Hence even though both output channels had
infinite gain margin and 90 degrees of phase margin, simultaneous variations
in both channels significantly reduce the margins at the plant outputs.
mmo
mmo =
GainMargin: [0.6065 1.6489]
PhaseMargin: [-27.5293 27.5293]
Frequency: 0.2287

If all the input and output channels are allow to vary independently, mmio, the
gain margin variation allow are 0.827 and 1.210 and phase margin variations
allowed are +/- 10.84 deg.
mmio
mmio =
GainMargin: [0.8267 1.2097]
PhaseMargin: [-10.8402 10.8402]

10-175

loopmargin

Frequency: 0.2287

Algorithm

Two well-known loop robustness measures are based on the sensitivity

function S=(IL)1 and the complementary sensitivity function T=L(IL)1
where L is the loop gain matrix associated with the input or output loops
broken simultaneously. In the following figure, S is the transfer matrix from
summing junction input u to summing junction output e. T is the transfer
matrix from u to y. If signals e and y are summed, the transfer matrix from u
to e+y is given by (I+L)(IL)1, the balanced sensitivity function. It can be
shown (Dailey, 1991, Blight, Daily and Gangass, 1994) that each broken-loop
gain can be perturbed by the complex gain (1+)(1) where ||<1/(S+T) or
||<1/max(S+T) at each frequency without causing instability at that
frequency. The peak value of (S+T) or max(S+T) gives a robustness guarantee
for all frequencies, and for (S+T) the guarantee is nonconservative (Blight,
Daily and Gangass, 1994).

  




10-176






 

 
    
 

 
 
    

loopmargin

This figure shows a comparison of a disk margin analysis with the classical
notations of gain and phase margins.

Disk gain margin (DGM) and disk phase margin (DPM) in the Nyquist plot
Unit Disk
Disk Margin
Critical point

1
0.8
0.6
Disk Margin

Imaginary Axis

0.4
0.2

DGM

GM

0.2
0.4

DPM

0.6
0.8

PM
1
1.5

0.5

0.5

Real Axis

10-177

loopmargin

The Nyquist plot is of the loop transfer function L(s)

s
------ + 1
30
L ( s ) = ----------------------------------------------------------( s + 1 ) ( s 2 + 1.6s + 16 )
The Nyquist plot of L corresponds to the blue line.
The unit disk corresponds to the dotted red line.
GM and PM indicate the location of the classical gain and phase margins for
the system L.
DGM and DPM correspond to the disk gain and phase margins. The disk
margins provide a lower bound on classical gain and phase margins.
The disk margin circle corresponds to the dashed black line. The disk margin
corresponds to the largest disk centered at (GMD + 1/GMD)/2 that just
touches the loop transfer function L. This location is indicated by the red dot.
The disk margin and multiple channel margins calculation involve the
balanced sensitivity function S+T. For a given peak value of (S+T), any
simultaneous phase and gain variations applied to each loop independently
will not destabilize the system if the perturbations remain inside the
corresponding circle or disk. This corresponds to the disk margin calculation to
find dmi and dmo.
Similarly, the multiple channel margins calculation involves the balanced
sensitivity function S+T. Instead of calculating (S+T) a single loop at a time,
all the channels are included in the analysis. A - analysis problem is
formulated with each channel perturbed by an independent, complex
perturbation. The peak (S+T) value guarantees that any simultaneous,
independent phase and gain variations applied to each loop simultaneously
will not destabilize the system if they remain inside the corresponding circle or
disk of size (S+T).

References

Barrett, M.F., Conservatism with robustness tests for linear feedback control
systems, Ph.D. Thesis, Control Science and Dynamical Systems, University of
Minnesota, 1980.
Blight, J.D., R.L. Dailey, and D. Gangsass, Practical control law design for
aircraft using multivariable techniques, International Journal of Control, Vol.
59, No. 1, 1994, pp. 93-137.

10-178

loopmargin

Bates, D., and I. Postlethwaite, Robust Multivariable Control of Aerospace

Systems, Delft University Press, Delft, The Netherlands, ISBN:
90-407-2317-6, 2002.

See Also

allmargin
bode
loopsens
mussv
robuststab
wcgain
wcsens
wcmargin

Finds all stability margins and crossover frequencies

Plots Bode frequency response of LTI models
Calculates sensitivity functions of feedback loop
Calculate bounds on the Structured Singular Value ()
Calculate stability margins of uncertain systems
Calculates worst-case gain of a system
Calculate worst-case sensitivities for feedback loop
Calculate worst-case margins for feedback loop

10-179

loopsens

Purpose

10loopsens

Analyze sensitivity of plant-controller feedback loop

Syntax

loops = loopsens(P,C)

Description

loops = loopsens(P,C) creates a Struct, loops, whose fields contain the

multivariable sensitivity, complementary and open-loop transfer functions.
The closed-loop system consists of the controller C in negative feedback with
the plant P. C should only be the compensator in the feedback path, not any
reference channels, if it is a 2-dof controller as seen in the figure below. The
plant and compensator P and C can be constant matrices, double, lti objects,
frd/ss/tf/zpk, or uncertain objects umat/ufrd/uss.

 

 
   

 
   

The loops structure contains the following fields:

10-180

Field

Description

Poles

Closed-loop poles. NaN for frd/ufrd objects

Stable

1 if nominal closed loop is stable, 0 otherwise. NaN for

frd/ufrd objects

Si

Input-to-plant sensitivity function

Ti

Input-to-plant complementary sensitivity function

Li

Input-to-plant loop transfer function

So

Output-to-plant sensitivity function

To

Output-to-plant complementary sensitivity function

Lo

Output-to-plant loop transfer function

loopsens

Field

Description

PSi

Plant times input-to-plant sensitivity function

CSo

Compensator times output-to-plant sensitivity function

The multivariable closed-loop interconnection structure, shown below, defines

the input/output sensitivity, complementary sensitivity, and loop transfer
functions.











Examples

Description

Equation

Input sensitivity ( TF e1 d1 )

(I+CP)1

Input complementary sensitivity ( TF e2 d1 )

CP(I+CP)1

Output sensitivity ( TF e3 d2 )

(I+PC)1

Output complementary sensitivity ( TF e4 d2 )

PC(I+PC)1

Input loop transfer function

CP

Output loop transfer function

PC

Consider a three input and four output plant with three integrators. One of the
integrators, uncint, is uncertain. A four input, three output constant gain
controller c is used to stabilize the closed-loop system. The gain of the
controller (1,4) entry is uncertain, ranging between 0.02 and 1.5. You can use
the loopsens command to construct the closed-loop system sensitivity,
complementary sensitivity, and loop transfer functions.

10-181

loopsens

int = tf(1,[1 0]);

addunc = ultidyn('addunc',[1 1],'Bound',0.2);
uncint = int + addunc;
int3 = blkdiag(int,uncint,int);
p = [1 .2 .5;-.5 0 1;1 0 0;0 1 0]*int3;
gain = ureal('gain',0.9, 'Range',[-3 1.5]);
c = [.02 .3 0 0;1 -.3 .32 gain; 0 1 -.2 .1];
loops = loopsens(p,c)
loops =
Poles: [3x1 double]
Stable: 1
Si: [3x3 uss]
Ti: [3x3 uss]
Li: [3x3 uss]
So: [4x4 uss]
To: [4x4 uss]
Lo: [4x4 uss]
PSi: [4x3 uss]
CSo: [3x4 uss]

Note that the input sensitivity, complementary sensitivity and loop transfer
functions have three inputs and three outputs, whereas the output sensitivity,
complementary sensitivity and loop transfer functions have four inputs and
four outputs. For the nominal plant model, i.e., the additive uncertainty addunc
set to zero, the closed-loop system is stable only for positive values of the gain
parameter.
gainval = reshape([-0.22 -0.215 -0.20 1.5],[1 1 4]);
nomSi = usubs(loops.Si,'gain',gainval);
pole(nomSi.NominalValue)
ans(:,:,1) =
-0.8537
0.0018 + 0.2164i
0.0018 - 0.2164i
ans(:,:,2) =
-0.8541
-0.0004 + 0.2176i
-0.0004 - 0.2176i
ans(:,:,3) =
-0.8556

10-182

loopsens

-0.0072 + 0.2207i
-0.0072 - 0.2207i
ans(:,:,4) =
-0.1663
-0.6505
-1.7533

See Also

loopmargin
robuststab
wcsens
wcmargin

Performs a comprehensive analysis of feedback loop

Calculate stability margins of uncertain systems
Calculate worst-case sensitivities for feedback loop
Calculate worst-case margins for feedback loop

10-183

loopsyn

Purpose

10loopsyn

H loop shaping controller synthesis

Syntax

[K,CL,GAM,INFO]=loopsyn(G,Gd)
[K,CL,GAM,INFO]=loopsyn(G,Gd,RANGE)

Description

loopsyn is an H optimal method for loop shaping control synthesis. It

computes a stabilizing Hcontroller K for plant G to shape the sigma plot of the
loop transfer function GK to have desired loop shape Gd with accuracy =GAM in
the sense that if 0 is the 0db crossover frequency of the sigma plot of Gd(j),
then, roughly,

1
( G ( j )K ( j ) ) --- ( G d ( j ) )

for all < 0

(10-16)

1
( G ( j )K ( j ) ) --- ( G d ( j ) )

for all > 0

(10-17)

The STRUCT array INFO returns additional design information, including a

MIMO stable min-phase shaping pre-filter W, the shaped plant Gs = GW, the
controller for the shaped plant Ks=WK, as well as the frequency range
{min,max} over which the loop shaping is achieved.)

10-184

Input arguments

Description

LTI Plant

Gd

Desired loop-shape (LTI Model)

RANGE

(optional, default {0,Inf}) Desired

frequency range for loop-shaping, a 1-by-2
cell array {min,max}; max should be at
least ten times min

Output arguments

Description

LTI controller

CL= G*K/(I+GK)

LTI closed-loop system

loopsyn

Algorithm

GAM

Loop-shaping accuracy (GAM 1, with GAM=1

being perfect fit

INFO

Additional output information

INFO.W

LTI pre-filter W satisfying (Gd)=(GW) for all ;

W is always minimum-phase

INFO.Gs

LTI shaped plant: Gs = GW

INFO.Ks

LTI controller for the shaped plant: Ks=WK

INFO.range

{min,max} Cell-array containing the approximate

frequency range over which loop-shaping could be
accurately achieved to with accuracy G. The output
INFO.range is either the same as or a subset of the
input range.

Using the GCD formula of Le and Safonov [1], loopsyn first computes a
stable-minimum-phase loop-shaping, squaring-down prefilter W such that the
shaped plant Gs = GW is square, and the desired shape Gd is achieved with
good accuracy in the frequency range {min,max} by the shaped plant; i.e.,
(Gd) (Gs) for all {min,max}.
Then, loopsyn uses the Glover-McFarlane [2] normalized-coprime-factor
control synthesis theory to compute an optimal loop-shaping controller for
the shaped plant via Ks=ncfsyn(Gs), and returns K=W*Ks.
If the plant G is a continuous time LTI and
1 G has a full-rank D-matrix, and
2 no finite zeros on the j-axis, and
3

{min,max}=[0,],

then GW theoretically achieves a perfect accuracy fit (Gd) = (GW) for all
frequency . Otherwise, loopsyn uses a bilinear pole-shifting bilinear
transform [3] of the form
Gshifted=bilin(G,-1,'S_Tust',[min,max]),

10-185

loopsyn

which results in a perfect fit for transformed Gshifted and an approximate fit
over the smaller frequency range [min,max] for the original unshifted G
provided that max >> min. For best results, you should choose max to be at
least 100 times greater than min. In some cases, the computation of the
optimal W for Gshifted may be singular or ill-conditioned for the range
[min,max], as when Gshifted has undamped zeros or, in the continuous-time
case only, Gshifted has a D-matrix that is rank-deficient); in such cases,
loopsyn automatically reduces the frequency range further, and returns the
reduced range [min,max] as a cell array in the output
INFO.range={min,max}.

Examples

The following code generates the optimal loopsyn loop shaping control for the
case of a 5-state, 4-output, 5-input plant with a full-rank non-minimum phase
zero at s=+10. The result in shown in Figure 10-11, LOOPSYN controller, on
page 10-187.
rand('seed',0);randn('seed',0);
s=tf('s'); w0=5; Gd=5/s;
% desired bandwith w0=5
G=((s-10)/(s+100))*rss(3,4,5);
% 4-by-5 non-min-phase plant
[K,CL,GAM,INFO]=loopsyn(G,Gd);
sigma(G*K,'r',Gd*GAM,'k-.',Gd/GAM,'k-.',{.1,100}) % plot result

This figure shows that the LOOPSYN controller K optimally fits

sigma(G*K) = sigma(Gd)GAM

dB

In the above example, GAM = 2.0423 = 6.2026 dB.

10-186

loopsyn

Figure 10-11: LOOPSYN controller

The loopsyn controller K optimally fits sigma(G*K). As shown in Figure 10-11,

it is sandwiched between sigma(Gd/GAM) and sigma(Gd*GAM) in accordance
with the inequalities in (10-16) and (10-17). In the this example, GAM = 2.0423
= 6.2026 dB.

Limitations

The plant G must be stabilizable and detectable, must have at least as many
inputs as outputs, and must be full rank; i.e,
size(G,2) size(G,1)
rank(freqresp(G,w)) = size(G,1) for some frequency w.
The order of the controller K can be large. Generically, when Gd is given as a
SISO LTI, then the order NK of the controller K satisfies
NK = NGs + NW
= NyNGd + NRHP + NW
= NyNGd + NRHP + NG
where

10-187

loopsyn

Ny denotes the number of outputs of the plant G.

NRHP denotes the total number of non-stable poles and non-minimum-phase
zeros of the plant G, including those on the stability boundary and at infinity.
NG, NGs, NGd and NW denote the respective orders of G, Gs, Gd and W.
Model reduction can help reduce the order of K see reduce and ncfmr.

References

[1] Le, V.X., and M.G. Safonov. Rational matrix GCDs and the design of
squaring-down compensatorsa state space theory. IEEE Trans.
Autom.Control, AC-36(3):384392, March 1992.
[2] Glover, K., and D. McFarlane. Robust stabilization of normalized coprime
factor plant descriptions with H-bounded uncertainty. IEEE Trans. Autom.
Control, AC-34(8):821830, August 1992.
[3] Chiang, R.Y., and M.G. Safonov. H synthesis using a bilinear pole-shifting
transform. AIAA J. Guidance, Control and Dynamics, 15(5):11111115,
SeptemberOctober 1992.

See Also

10-188

loopsyn_demo
mixsyn
ncfsyn

A demo of this function

H mixed-sensitivity controller synthesis
H - normailized coprime controller synthesis

ltiarray2uss

Purpose

10ltiarray2uss

Computes uncertain system bounding an LTI array

Syntax

usys = ltiarray2uss(P,Parray,ord)
[usys,wt] = ltiarray2uss(P,Parray,ord)
[usys,wt,diffdata] = ltiarray2uss(P,Parray,ord)
[usys,wt,diffdata] = ltiarray2uss(P,Parray,ord,'InputMult')
[usys,wt,diffdata] = ltiarray2uss(P,Parray,ord,'OutputMult')
[usys,wt,diffdata] = ltiarray2uss(P,Parray,ord,'Additive')

Description

The command ltiarray2uss calculates an uncertain system usys with

nominal value P, and whose range of behavior includes the given array of
systems, Parray.
usys = ltiarray2uss(P,Parray,ord) usys is formulated as an input

multiplicative uncertainty model,

usys = P*(I + wt*ultidyn('IMult',[size(P,2) size(P,2)]))

where wt is a stable scalar system, whose magnitude overbounds the relative

difference,
(P - Parray)/P

The state order of the weighting function used to bound the multiplicative
difference between P and Parray is ord. Both P and Parray must be in the
classes ss/tf/zpk/frd. If P is an frd then usys will be a ufrd object, otherwise
usys will be a uss object. The ultidyn atom is named based on the variable
name of Parray in the calling workspace.
[usys,wt] = ltiarray2uss(P,Parray,ord) returns the weight wt used to
bound the infinity norm of ((P - Parray)/P).
[usys,wt] = ltiarray2uss(P,Parray,ord,'OutputMult') uses

multiplicative uncertainty at the plant output (as opposed to input

multiplicative uncertainty). The formula for usys is
usys

= (I + wt*ultidyn('Name',[size(P,1) size(P,1)])*P).

[usys,wt] = ltiarray2uss(P,Parray,ord,'Additive') uses additive

uncertainty.
usys = P + wt*ultidyn('Name',[size(P,1) size(P,2)]).

10-189

ltiarray2uss

wt is a frequency domain overbound of the infinity norm of (Parray - P).

[usys,wt] = ltiarray2uss(P,Parray,ord,'InputMult') uses multiplicative
uncertainty at the plant input (this is the default). The formula for usys is
usys

= P*(I + wt*ultidyn('Name',[size(P,2) size(P,2)])).

[usys,wt,diffdata] = ltiarray2uss(P,Parray,ord,type) returns the norm

of the difference (absolute difference for additive, and relative difference for
multiplicative uncertainty) between the nominal model P and Parray. wt
satisfies
diffdata(w_i) < |wt(w_i)|

at all frequency points.

Example

Note See the Robust Control Toolbox demo entitled First-cut Robust
Design for a more detailed example of how to use ltiarray2uss.

Consider a third order transfer function with an uncertain gain, filter time
constant and a lightly damped flexible mode. This model is used to represent a
physical system from frequency response data is acquired.
gain = ureal('gain',10,'Perc',20);
tau = ureal('tau',.6,'Range',[.42 .9]);
wn = 40;
zeta = 0.1;
usys = tf(gain,[tau 1])*tf(wn^2,[1 2*zeta*wn wn^2]);
sysnom = usys.NominalValue;
parray = usample(usys,30);
om = logspace(-1,2,80);
parrayg = frd(parray,om);

10-190

ltiarray2uss

bode(parrayg)

Bode Diagram
2

Magnitude (abs)

10

10

10

Phase (deg)

90

180

270
1
10

10
10
Frequency (rad/sec)

10

The frequency response data in parray represents 30 experiments performed

on the system. The command ltiarray2uss is used to generate an uncertain
model, umod, based on the frequency response data. Initially an input
multiplicative uncertain model is used to characterize the collection of 30
frequency responses. First and second order input multiplicative uncertainty
weight are calculated from the data.
[umodIn1,wtIn1,diffdataIn] = ltiarray2uss(sysnom,parrayg,1);
[umodIn2,wtIn2,diffdataIn] = ltiarray2uss(sysnom,parrayg,2);

10-191

ltiarray2uss

bodemag(wtIn1,'b-',wtIn2,'g+',diffdataIn,'r.',om)
Input Multiplicative Uncertainty Model using LTIARRAY2USS

0.3

Magnitude (abs)

10

0.4

10

0.5

10

0.6

10

1st order
2nd order
difference

0.7

10

10

10
10
Frequency (rad/sec)

10

Alternatively, an additive uncertain model is used to characterize the

collection of 30 frequency responses.
[umodAdd1,wtAdd1,diffdataAdd] = ...
ltiarray2uss(sysnom,parrayg,1,'Additive');
[umodAdd2,wtAdd2,diffdataAdd] = ...
ltiarray2uss(sysnom,parrayg,2,'Additive');
bodemag(wtAdd1,'b-',wtAdd2,'g+',diffdataAdd,'r.',om)

10-192

ltiarray2uss

Additive Uncertainty Model using LTIARRAY2USS

2

10

Magnitude (abs)

10

10

10

1st order
2nd order
difference

10

10

See Also

lft
ssdata

10
10
Frequency (rad/sec)

10

Forms Redheffer star product of systems

Returns uncertain state-space data

10-193

ltrsyn

Purpose

10ltrsyn

LQG loop-transfer recovery controller synthesis

Syntax

[K,SVL,W1]
[K,SVL,W1]
[K,SVL,W1]
[K,SVL,W1]

Description

[K,SVL,W1] = ltrsyn(G,F,XI,TH,RHO) computes a reconstructed-state

output-feedback controller K for LTI plant G so that K*G asymptotically
recovers plant-input full-state feedback loop transfer function L(s) =
F(Is-A)-1B+D; that is, at any frequency w>0, max(sigma(K*G-L, w)) 0 as
, where L= ss(A,B,F,D) is the LTI full-state feedback loop transfer
function.

=
=
=
=

ltrsyn(G,F,XI,THETA,RHO)
ltrsyn(G,F,XI,THETA,RHO,W)
ltrsyn(G,F,XI,THETA,RHO,OPT)
ltrsyn(G,F,XI,THETA,RHO,W,OPT)

[K,SVL,W1] = ltrsyn(G,F1,Q,R,RHO,'OUTPUT') computes the solution to the

dual problem of filter loop recovery for LTI plant G where F is a Kalman filter
gain matrix. In this case, the recovery is at the plant output, and
max(sigma(G*K-L, w)) 0 as , where L1 denotes the LTI filter loop
feedback loop transfer function L1= ss(A,F,C,D).

Only the LTI controller K for the final value RHO(end)is returned.
Inputs

10-194

LTI plant

LQ full-state-feedback gain matrix

XI

Plant noise intensity,

or, if OPT='OUTPUT' state-cost matrix XI=Q,

THETA

Sensor noise intensity

or, if OPT='OUTPUT' control-cost matrix THETA=R,

RHO

Vector containing a set of recovery gains

(Optional) Vector of frequencies (to be used for plots); if

input W is not supplied, then a reasonable default is used

ltrsyn

Outputs
K

K(s) LTI LTR (loop-transfer-recovery) output-feedback,

for the last element of RHO (i.e., RHO(end))

SVL

Sigma plot data for the recovered loop transfer function if

G is MIMO or, for SISO G only, Nyquist loci SVL =
[re(1:nr) im(1:nr)]

W1

Algorithm

Frequencies for SVL plots, same as W when present

For each value in the vector RHO, [K,SVL,W1] = ltrsyn(G,F,XI,THETA,RHO)

computes the full-state-feedback (default OPT='INPUT') LTR controller
1

K ( s ) = [ K c ( Is A + B K c + K f C K f DK c ) K f ]
where Kc=F and Kf=lqr(A',C',XI+RHO(i)*B*B',THETA). The fictitious noise
term RHO(i)*B*B' results in loop-transfer recovery as RHO(i) . The
T

Kalman filter gain is K f = C

T

where satisfies the Kalman filter Riccati

T

equation 0 = A + A C C + + BB . See [1] for further details.

Similarly for the dual problem of filter loop recovery case, [K,SVL,W1] =
ltrsyn(G,F,Q,R,RHO,'OUTPUT') computes a filter loop recovery controller of
the same form, but with Kf=F is being the input filter gain matrix and the
control gain matrix Kc computed as Kc=lqr(A,B,Q+RHO(i)*C'*C,R).

10-195

ltrsyn

LQG/LTR @ PLANT OUTPUT

50

SV (DB)

-50

-100

-150

-200

- & -- : Kalman Filter Gain

.. & .-: q = 1
- & -- : q = 1e5
.. & .-: q = 1e10
- & -- : q = 1e15

-250

-300
10 -4

10 -3

10 -2

10 -1

10 0

10 1

10 2

10 3

10 4

Rad/Sec

Figure 10-12: Example of LQG/LTR at Plant Output

Examples

s=tf('s');G=ss(1e4/((s+1)*(s+10)*(s+100)));[A,B,C,D]=ssdata(G);
F=lqr(A,B,C'*C,eye(size(B,2)));
L=ss(A,B,F,0*F*B);
XI=100*C'*C; THETA=eye(size(C,1));
RHO=[1e3,1e6,1e9,1e12];W=logspace(-2,2);
nyquist(L,'k-.');hold;
[K,SVL,W1]=ltrsyn(G,F,XI,THETA,RHO,W);

See also ltrdemo.

Limitations

The ltrsyn procedure may fail for non-minimum phase plants. For full-state
LTR (default OPT='INPUT'), the plant should not have fewer outputs than
inputs. Conversely for filter LTR (when OPT='OUTPUT'), the plant should not
have fewer inputs than outputs. The plant must be strictly proper, i.e., the
D-matrix of the plant should be all zeros. ltrsyn is only for continuous time
plants (Ts==0).

References

[1] Doyle, J., and G. Stein, Multivariable Feedback Design: Concepts for a
Classical/Modern Synthesis, IEEE Trans. on Automat. Contr., AC-26, pp.
4-16, 1981.

See Also

h2syn
hinfsyn

10-196

H2 controller synthesis
H controller synthesis

ltrsyn

lqg
loopsyn
ltrdemo
ncfsyn

Continuous linear-quadratic-Gaussian control

synthesis
H - loop shaping controller synthesis
Demo of LQG/LTR optimal control synthesis
H - normalized coprime controller synthesis

10-197

matnbr

Purpose

10matnbr

Number of matrix variables in LMI system

Syntax

K = matnbr(lmisys)

Description

matnbr returns the number K of matrix variables in the system of LMIs

described by lmisys.

See Also

decnbr
lmiinfo
decinfo

10-198

Give the total number of decision variables in a system

of LMIs
Interactively retrieve information about the variables
and term content of LMIs
Describe how the entries of a matrix variable X relate
to the decision variables

mat2dec

Purpose

10mat2dec

Extract vector of decision variables from matrix variable values

Syntax

decvec = mat2dec(lmisys,X1,X2,X3,...)

Description

Given an LMI system lmisys with matrix variables X1, . . ., XK and given
values X1,...,Xk of X1, . . ., XK, mat2dec returns the corresponding value
decvec of the vector of decision variables. Recall that the decision variables are
the independent entries of the matrices X1, . . ., XK and constitute the free
scalar variables in the LMI problem.
This function is useful, for example, to initialize the LMI solvers mincx or gevp.
Given an initial guess for X1, . . ., XK, mat2dec forms the corresponding vector
of decision variables xinit.
An error occurs if the dimensions and structure of X1,...,Xk are inconsistent
with the description of X1, . . ., XK in lmisys.

Examples

Consider an LMI system with two matrix variables X and Y such that
X is A symmetric block diagonal with one 2-by-2 full block and one 2-by-2
scalar block
Y is a 2-by-3 rectangular matrix
Particular instances of X and Y are

X0 =

1
3
0
0

3
1
0
0

0
0
5
0

0
0
0
5

Y0 = 1 2 3
4 5 6

and the corresponding vector of decision variables is given by

decv = mat2dec(lmisys,X0,Y0)
decv'
ans =
1

-1

3 4 5 6

10-199

mat2dec

Note that decv is of length 10 since Y has 6 free entries while X has 4
independent entries due to its structure. Use decinfo to obtain more
information about the decision variable distribution in X and Y.

See Also

dec2mat
decinfo
decnbr

10-200

Given values of the decision variables, derive the

corresponding values of the matrix variables
Describe how the entries of a matrix variable X relate
to the decision variables
Give the total number of decision variables in a system
of LMIs

mincx

Purpose

10mincx

Minimize linear objective under LMI constraints

Syntax

[copt,xopt] = mincx(lmisys,c,options,xinit,target)

Description

The function mincx solves the convex program

T

minimize c x subject to

N L(x)N M R( x)M

(10-18)

where x denotes the vector of scalar decision variables.

The system of LMIs (9-18) is described by lmisys. The vector c must be of the
same length as x. This length corresponds to the number of decision variables
returned by the function decnbr. For linear objectives expressed in terms of the
matrix variables, the adequate c vector is easily derived with defcx.
The function mincx returns the global minimum copt for the objective cTx, as
well as the minimizing value xopt of the vector of decision variables. The
corresponding values of the matrix variables is derived from xopt with
dec2mat.
The remaining arguments are optional. The vector xinit is an initial guess of
the minimizer xopt. It is ignored when infeasible, but may speed up
computations otherwise. Note that xinit should be of the same length as c. As
for target, it sets some target for the objective value. The code terminates as
soon as this target is achieved, that is, as soon as some feasible x such that
cTx target is found. Set options to [] to use xinit and target with the
default options.

Control
Parameters

The optional argument options gives access to certain control parameters of

the optimization code. In mincx, this is a five-entry vector organized as follows:
options(1) sets the desired relative accuracy on the optimal value lopt
(default = 102).
options(2) sets the maximum number of iterations allowed to be performed
by the optimization procedure (100 by default).
options(3) sets the feasibility radius. Its purpose and usage are as for
feasp.

10-201

mincx

options(4) helps speed up termination. If set to an integer value J > 0, the

code terminates when the objective cTx has not decreased by more than the
desired relative accuracy during the last J iterations.
options(5) = 1 turns off the trace of execution of the optimization
procedure. Resetting options(5) to zero (default value) turns it back on.
Setting option(i) to zero is equivalent to setting the corresponding control
parameter to its default value. See feasp for more detail.

Tip for
Speed-Up

In LMI optimization, the computational overhead per iteration mostly comes

from solving a least-squares problem of the form
min Ax b
x

where x is the vector of decision variables. Two methods are used to solve this
problem: Cholesky factorization of ATA (default), and QR factorization of A
when the normal equation becomes ill conditioned (when close to the solution
typically). The message
* switching to QR

is displayed when the solver has to switch to the QR mode.

Since QR factorization is incrementally more expensive in most problems, it is
sometimes desirable to prevent switching to QR. This is done by setting
options(4) = 1. While not guaranteed to produce the optimal value, this
generally achieves a good trade-off between speed and accuracy.

Memory
Problems

QR-based linear algebra (see above) is not only expensive in terms of

computational overhead, but also in terms of memory requirement. As a result,
the amount of memory required by QR may exceed your swap space for large
problems with numerous LMI constraints. In such case, MATLAB issues the
error
??? Error using ==> pds
Out of memory. Type HELP MEMORY for your options.

You should then ask your system manager to increase your swap space or, if no
additional swap space is available, set options(4) = 1. This will prevent
switching to QR and mincx will terminate when Cholesky fails due to
numerical instabilities.

10-202

mincx

References

The solver mincx implements Nesterov and Nemirovskis Projective Method as

described in
Nesterov, Yu, and A. Nemirovski, Interior Point Polynomial Methods in Convex
Programming: Theory and Applications, SIAM, Philadelphia, 1994.
Nemirovski, A., and P. Gahinet, The Projective Method for Solving Linear
Matrix Inequalities, Proc. Amer. Contr. Conf., 1994, Baltimore, Maryland, pp.
840-844.
The optimization is performed by the C-MEX file pds.mex.

See Also

defcx
dec2mat
decnbr
feasp
gevp

Help specify cTx objectives for the mincx solver

Given values of the decision variables, derive the
corresponding values of the matrix variables
Give the total number of decision variables in a system
of LMIs
Find a solution to a given system of LMIs
Generalized eigenvalue minimization under LMI
constraints

10-203

mixsyn

Purpose

10mixsyn

H mixed-sensitivity controller synthesis

Syntax

[K,CL,GAM,INFO]=mixsyn(G,W1,W2,W3)
[K,CL,GAM,INFO]=mixsyn(G,W1,W2,W3,KEY1,VALUE1,KEY2,VALUE2,...)

Description

[K,CL,GAM,INFO]=mixsyn(G,W1,W2,W3) computes a controller K that

minimizes the H norm of the closed-loop transfer function the weighted mixed
sensitivity

W 1S

T y1 u1 = W 2 R
W3 T
where S and T are called the sensitivity and complementary sensitivity,
respectively and S, R and T are given by
S = ( I + GK )

R = K ( I + GK )

T = GK ( I + GK )

AUGMENTED PLANT P(s)

W1
W2
+
u
1
u

e
-

W3

2
y
2

CONTROLLER
F(s)
K(s)

Figure 10-13: Closed-loop transfer function T y u for mixed sensitivity

1 1
mixsyn.

10-204

mixsyn

The returned values of S, R, and T satisfy the following loop shaping

inequalities:
1

( S ( j ) ) ( W 1 ( j ) )
1

( R ( j ) ) ( W 2 ( j ) )
1

( T ( j ) ) ( W 3 ( j ) )
where =GAM. Thus, W1, W3 determine the shapes of sensitivity S and
complementary sensitivity T. Typically, you would choose W1 to be small inside
the desired control bandwidth to achieve good disturbance attenuation (i.e.,
performance), and choose W3 to be small outside the control bandwidth, which
helps to ensure good stability margin (i.e., robustness).
For dimensional compatibility, each of the three weights W1, W2 and W3 must
be either empty, scalar (SISO) or have respective input dimensions NY, NU,
and NY where G is NY-by-NU. If one of the weights is not needed, you may
simply assign an empty matrix []; e.g., P = AUGW(G,W1,[],W3) is SYS but
without the second row (without the row containing W2).

Algorithm

[K,CL,GAM,INFO]=mixsyn(G,W1,W2,W3,KEY1,VALUE1,KEY2,VALUE2,...)

is equivalent to
[K,CL,GAM,INFO]=...
hinfsyn((G,W1,W2,W3),KEY1,VALUE1,KEY2,VALUE2,...).
mixsyn accepts all the same key value pairs as hinfsyn.

Example

The following code illustrates the use of mixsyn for sensitivity and
complementary sensitivity loop-shaping.
s=zpk('s');
G=(s-1)/(s+1)^2;
W1=0.1*(s+100)/(100*s+1); W2=0.1;
[K,CL,GAM]=mixsyn(G,W1,W2,[]);
L=G*K; S=inv(1+L); T=1-S;
sigma(S,'g',T,'r',GAM/W1,'g-.',GAM*G/ss(W2),'r-.')

10-205

mixsyn

Figure 10-14: mixsyn(G,W1,W2,[ ]) shapes sigma plots of S and T to

conform to /W1 and G/W2, respectively.

Limitations

The transfer functions G, W1, W2 and W3 must be proper, i.e., bounded as

s or, in the discrete-time case, as z . Additionally, W1, W2 and W3
should be stable. The plant G should be stabilizable and detectable; else, P will
not be stabilizable by any K.

See Also

augw
hinfsyn

10-206

Augments plant weights for control design

H controller synthesis

mkfilter

Purpose

10mkfilter

Generate Bessel, Butterworth, Chebyshev, or RC filter

Syntax

sys = mkfilter(fc,ord,type)
sys = mkfilter(fc,ord,type,psbndr)

Description

sys = mkfilter(fc,ord,type) returns a single-input, single-output analog

low pass filter sys as an ss object. The cutoff frequency (Hertz) is fc and the
filter order is ord, a positive integer. The string variable type specifies the type
of filter and can be one of the following:
String variable

Description

'butterw'

Butterworth filter

'cheby'

Chebyshev filter

'bessel'

Bessel filter

'rc'

Series of resistor/capacitor filters

The dc gain of each filter (except even-order Chebyshev) is set to unity.

sys = mkfilter(fc,ord,type,psbndr) contains the input argument psbndr
that specifies the Chebyshev passband ripple (in dB). At the cutoff frequency,
the magnitude is -psbndr dB. For even-order Chebyshev filters, the DC gain is
also -psbndr dB.

Examples

butw = mkfilter(2,4,'butterw');
cheb = mkfilter(4,4,'cheby',0.5);
rc = mkfilter(1,4,'rc');
bode(butw_g,'-',cheb_g,'--',rc_g,'-.')
megend('Butterworth','Chebyshev','RC filter')

10-207

mkfilter

Log Magnitude

Butterworth, RC & Chebyshev Filters

10

10

10

10

10

10

Frequency (radians/sec)

Phase (degrees)

0
100
200
300
1

10

10

10

10

Frequency (radians/sec)

Limitations

The Bessel filters are calculated using the recursive polynomial formula. This
is poorly conditioned for high order filters (order > 8).

See Also

augw

10-208

Augments plant weights for control design

mktito

Purpose

10mktito

Partition LTI system into two-input-two-output system

Syntax

SYS=mktito(SYS,NMEAS,NCONT)

Description

SYS=mktito(SYS,NMEAS,NCONT) adds TITO (two-input-two-output)

partitioning to the LTI system SYS, assigning OutputGroup and InputGroup
properties such that
NMEAS = dim ( y 2 )
NCONT = dim ( u 2 )

y1

u1
P

y2

u2
SYS

Any pre-existing OutputGroup or InputGroup properties of SYS are

overwritten. TITO partitioning simplifies syntax for control synthesis
functions like hinfsyn and h2syn.

Algorithm

Examples

[r,c]=size(SYS);
set(SYS,'InputGroup', struct('U1',1:c-NCONT,'U2',c-NCONT+1:c));
set(SYS,'OutputGroup',struct('Y1',1:r-NMEAS,'Y2',r-NMEAS+1:r));

You can type

P=rss(2,4,5); P=mktito(P,2,2);
disp(P.OutputGroup); disp(P.InputGroup);

to create a 4-by-5 LTI system P with OutputGroup and InputGroup properties

U1:
U2:
Y1:
Y2:

[1
[4
[1
[3

2 3]
5]
2]
4]

10-209

mktito

See also

10-210

augw
hinfsyn
h2syn
ltiprops
sdhinfsyn

Augments plant weights for control design

H synthesis controller
H2 synthesis controller
Help on LTI model properties
H discrete-time controller

modreal

Purpose

10modreal

Modal form realization and projection

Syntax

[G1,G2] = modreal(G,cut)

Description

[G1,G2] = modreal(G,cut) returns a set of state-space LTI objects G1 and G2

in modal form given a state space G and the model size of G1, cut.

The modal form realization has its A matrix in block diagonal form with either
1x1 or 2x2 blocks. The real eigenvalues will be put in 1x1 blocks and complex
eigenvalues will be put in 2x2 blocks. These diagonal blocks are ordered in
ascending order based on eigenvalue magnitudes.
The complex eigenvalue a+bj is appearing as 2x2 block
a b
b a
This table describes input arguments for modreal.
Argument

Description

LTI model to be reduced.

cut

(Optional) an integer to split the realization. Without it, a

complete modal form realization is returned

This table lists output arguments.

Argument

Description

G1,G2

LTI models in modal form

G can be stable or unstable. G 1 = ( A 1, B 1, C 1, D 1 ) , G 2 = ( A 2, B 2, C 2, D 2 ) and

D 1 = D + C 2 ( A 2 ) 1 B 2 is calculated such that the system DC gain is
preserved.

Algorithm

Using a real eigen structure decomposition reig and ordering the eigenvectors
in ascending order according to their eigenvalue magnitudes, we can form a

10-211

modreal

similarity transformation out of these ordered real eigenvectors such that he

resulting systems G1 and/or G2 are in block diagonal modal form.

Note This routine is extremely useful when model has j-axis singularities,
e.g., rigid body dynamics. It has been incorporated inside Hankel based model
reduction routines - hankelmr, balancmr, bstmr, and schurmr to isolate those
j-axis poles from the actual model reduction process.

Examples

Given a continuous stable or unstable system, G, the following commands can

get a set of modal form realizations depending on the split index -- cut:
randn('state',1234); rand('state',5678);
G = rss(50,2,2);
[G1,G2] = modreal(G,2); % cut = 2 for two rigid body modes
G1.d = zeros(2,2); % remove the DC gain of the system from G1
sigma(G,G1,G2)

See Also

10-212

reduce
balancmr
schurmr
bstmr
ncfmr
hankelmr
hankelsv

Top level model reduction routines

Balanced truncation via square-root method
Balanced truncation via Schur method
Balanced stochastic truncation via Schur method
Balanced truncation for normalized coprime factors
Hankel minimum degree approximation
Hankel singular value

msfsyn

Purpose

10msfsyn

Multi-model/multi-objective state-feedback synthesis

Syntax

[gopt,h2opt,K,Pcl,X] = msfsyn(P,r,obj,region,tol)

Description

Given an LTI plant P with state-space equations

Ax + B 1 w + B 2 u

z =

C 1 x + D 11 w + D 12 u

z2 =

C 2 x + D 22 u

msfsyn computes a state-feedback control u = Kx that

Maintains the RMS gain (H norm) of the closed-loop transfer function T

from w to z below some prescribed value 0 > 0
Maintains the H2 norm of the closed-loop transfer function T2 from w to z2
below some prescribed value 0 > 0
Minimizes an H2/H trade-off criterion of the form
2

T + T2 2
Places the closed-loop poles inside the LMI region specified by region (see
lmireg for the specification of such regions). The default is the open left-half
plane.
Set r = size(d22) and obj = [0, 0, , ] to specify the problem dimensions
and the design parameters 0, 0, , and . You can perform pure pole
placement by setting obj = [0 0 0 0]. Note also that z or z2 can be empty.
On output, gopt and h2opt are the guaranteed H and H2 performances, K is
the optimal state-feedback gain, Pcl the closed-loop transfer function from w
z

to , and X the corresponding Lyapunov matrix.

z2

The function msfsyn is also applicable to multi-model problems where P is a

polytopic model of the plant:

10-213

msfsyn

A ( t )x + B 1 ( t )w + B 2 ( t )u

z =

C 1 ( t )x + D 11 ( t )w + D 12 ( t )u

z2 =

C 2 ( t )x + D 22 ( t )u

with time-varying state-space matrices ranging in the polytope

A(t) B (t) B (t)

A k B k B 2k
1
2

C ( t ) D ( t ) D ( t ) Co C D
1k 11k D 12k
11
12
1

C (t)

0
D 22 ( t )
C 2k 0 D 22k
2

: k = 1, , K

In this context, msfsyn seeks a state-feedback gain that robustly enforces the
specifications over the entire polytope of plants. Note that polytopic plants
should be defined with psys and that the closed-loop system Pcl is itself
polytopic in such problems. Affine parameter-dependent plants are also
accepted and automatically converted to polytopic models.

See Also

10-214

lmireg
psys

Specify LMI regions for pole placement purposes

Specification of uncertain state-space models

mussv

Purpose

10mussv

Compute bounds on structured singular value ()

Syntax

bounds = mussv(M,BlockStructure)
[bounds,muinfo] = mussv(M,BlockStructure)
[bounds,muinfo] = mussv(M,BlockStructure,Options)
[ubound,q] = mussv(M,F,BlockStructure)
[ubound,q] = mussv(M,F,BlockStructure,'s')

Description

bounds = mussv(M,BlockStructure) calculates upper and lower bounds on

the structured singular value, or , for a given block structure. M is a double,
or frd object. If M is an N-D array (with N 3), then the computation is
performed pointwise along the third and higher array dimensions. If M is a frd

object, the computations are performed pointwise in frequency (as well as any
array dimensions).
BlockStructure is a matrix specifying the perturbation block structure.
BlockStructure has 2 columns, and as many rows as uncertainty blocks in the
perturbation structure. The i-th row of BlockStructure defines the dimensions

of the ith perturbation block.

If BlockStructure(i,:) = [-r 0], then the i-th block is an r-by-r repeated,
diagonal real scalar perturbation;
if BlockStructure(i,:) = [r 0], then the i-th block is an r-by-r repeated,
diagonal complex scalar perturbation;
if BlockStructure(i,:) = [r c], then the i-th block is an r-by-c complex
full-block perturbation.
If BlockStructure is omitted, its default is ones(size(M,1),2), which
implies a perturbation structure of all 1-by-1 complex blocks. In this case, if
size(M,1) does not equal size(M,2), an error results.
If M is a two-dimensional matrix, then bounds is a 1-by-2 array containing an
upper (first column) and lower (second column) bound of the structured
singular value of M. For all matrices Delta with block-diagonal structure
defined by BlockStructure and with norm less than 1/bounds(1) (upper
bound), the matrix I - M*Delta is not singular. Moreover, there is a matrix
DeltaS with block-diagonal structure defined by BlockStructure and with
norm equal to 1/bounds(2) (lower bound), for which the matrix I - M*DeltaS
is singular.

10-215

mussv

If M is an frd, the computations are always performed pointwise in frequency.

The output argument bounds is a 1-by-2 frd of upper and lower bounds at
each frequency. Note that bounds.Frequency equals M.Frequency.
If M is an N-D array (either double or frd), the upper and lower bounds are
computed pointwise along the 3rd and higher array dimensions (as well as
pointwise in frequency, for frd). For example, suppose that size(M) is
rcd1dF. Then size(bounds) is 12d1dF. Using single index
notation, bounds(1,1,i) is the upper bound for the structured singular value
of M(:,:,i), and bounds(1,2,i) is the lower bound for the structured singular
value of M(:,:,i). Here, any i between 1 and d1d2dF (the product of the dk)
would be valid.
bounds = mussv(M,BlockStructure,Options) specifies computation options.
Options is a character string, containing any combination of the following

characters:

10-216

Option

Meaning

'a'

Upper bound to greatest accuracy, using LMI solver

'an'

Same as 'a', but without automatic prescaling

'd'

Display warnings

'f'

Fast upper bound (typically not as tight as the default)

'i'

Initialize lower bound computation using previous matrix

(only relevant if M is ND array or frd)

'm7'

Randomly initialize lower bound iteration multiple times

(in this case 7 times, use 1-9); larger number typically gives
better lower bound

'U'

Upper-bound only (lower bound uses a fast/cheap

algorithm)

's'

Suppress progress information (silent)

'x'

Decrease iterations in lower bound computation (faster but

not as tight as default)

mussv

[bounds,muinfo] = mussv(M,BlockStructure) returns muinfo, a structure

containing more detailed information. The information within muinfo must be
extracted using mussvextract. See mussvextract for more details.

Generalized Structured Singular Value

ubound = mussv(M,F,BlockStructure) calculates an upper bound on the
generalized structured singular value (generalized ) for a given block
structure. M is a double or frd object. M and BlockStructure are as before. F is
an additional (double or frd).
ubound = mussv(M,F,BlockStructure,'s') adds an option to run silently.
Other options are ignored for generalized problems.

Note that in generalized structured singular value computations, only an

upper bound is calculated. ubound is an upper bound of the generalized
structured singular value of the pair (M,F), with respect to the block-diagonal
uncertainty described by BlockStructure. Consequently ubound is 1-by-1
(with additional array dependence, depending on M and F). For all matrices
Delta with block-diagonal structure defined by BlockStructure and
norm<1/ubound, the matrix [I-Delta*M;F] is guaranteed not to lose column
rank. This is verified by the matrix Q, which satisfies
mussv(M+Q*F,BlockStructure,'C')<=ubound.

Examples

See mussvextract for a detailed example of the structured singular value.

A simple example for generalized structured singular value can be done with
random complex matrices, illustrating the relationship between the upper
bound for and generalized , as well as the fact that the upper bound for
generalized comes from an optimized upper bound.
M is a complex 5-by-5 matrix and F is a complex 2-by-5 matrix. The block
structure BlockStructure is an uncertain real parameter 1, an uncertain real

parameter 2, an uncertain complex parameter 3 and a twice-repeated

uncertain complex parameter 4.
M = randn(5,5) + sqrt(-1)*randn(5,5);
F = randn(2,5) + sqrt(-1)*randn(2,5);
BlockStructure = [-1 0;-1 0;1 1;2 0];
[ubound,Q] = mussv(M,F,BlockStructure);
points completed (of 1) ... 1

10-217

mussv

bounds = mussv(M,BlockStructure);
optbounds = mussv(M+Q*F,BlockStructure,'C5');

The quantities optbounds(1) and ubound should be extremely close, and

significantly lower than bounds(1) and bounds(2).
[optbounds(1) ubound]
ans =
1.5917
1.5925
[bounds(1) bounds(2)]
ans =
3.8184
3.7135

Algorithm

The lower bound is computed using a power method, Young and Doyle, 1990,
and Packard et al. 1988, and the upper bound is computed using the
balanced/AMI technique, Young et al., 1992, for computing the upper bound
from Fan et al., 1991.
Peter Young and Matt Newlin wrote the original M-files.
The lower-bound power algorithm is from Young and Doyle, 1990, and Packard
et al. 1988.
The upper-bound is an implementation of the bound from Fan et al., 1991, and
is described in detail in Young et al., 1992. In the upper bound computation, the
matrix is first balanced using either a variation of Osbornes method (Osborne,
1960) generalized to handle repeated scalar and full blocks, or a Perron
approach. This generates the standard upper bound for the associated complex
problem. The Perron eigenvector method is based on an idea of Safonov,
(Safonov, 1982). It gives the exact computation of for positive matrices with
scalar blocks, but is comparable to Osborne on general matrices. Both the
Perron and Osborne methods have been modified to handle repeated scalar and
full blocks. Perron is faster for small matrices but has a growth rate of n3,
compared with less than n2 for Osborne. This is partly due to the MATLAB
implementation, which greatly favors Perron. The default is to use Perron for
simple block structures and Osborne for more complicated block structures. A
sequence of improvements to the upper bound is then made based on various
equivalent forms of the upper bound. A number of descent techniques are used
that exploit the structure of the problem, concluding with general purpose LMI
optimization (Boyd et al.), 1993, to obtain the final answer.

10-218

mussv

The optimal choice of Q (to minimize the upper bound) in the generalized mu
problem is solved by reformulating the optimization into a semidefinite
program (Packard et al., 1991).

References

Boyd, S. and L. El Ghaoui, Methods of centers for minimizing generalized

eigenvalues, Linear Algebra and Its Applications, Vol. 188189, 1993, pp.
63111.
Fan, M. A. Tits, and J. Doyle, Robustness in the presence of mixed
parametric uncertainty and unmodeled dynamics, IEEE Transactions on
Automatic Control, Vol. AC36, 1991, pp. 2538.
Osborne, E., On preconditioning of matrices, Journal of Associated
Computer Machines, Vol. 7, 1960, pp. 338345.
Packard, A.K., M. Fan and J. Doyle, A power method for the structured
singular value, Proc. of 1988 IEEE Conference on Control and Decision,
December 1988, pp. 21322137.
Safonov, M., Stability margins for diagonally perturbed multivariable
feedback systems, IEEE Proc., Vol. 129, Part D, 1992, pp. 251256.
Young, P. and J. Doyle, Computation of with real and complex
uncertainties, Proceedings of the 29th IEEE Conference on Decision and
Control, 1990, pp. 12301235.
Young, P., M. Newlin, and J. Doyle, Practical computation of the mixed
problem, Proceedings of the American Control Conference, 1992, pp. 2190
2194.

See Also

loopmargin
mussvextract
robuststab
robustperf
wcgain
wcsens
wcmargin

Comprehensive analysis of feedback loop

Extract compressed data returned from mussv
Calculate stability margins of uncertain systems
Calculate performance margins of uncertain systems
Calculate worst-case gain of a system
Calculate worst-case sensitivities for feedback loop
Calculate worst-case margins for feedback loop

10-219

mussvextract

Purpose

10mussvextract

Extract muinfo structure returned by mussv

Syntax

[VDelta,VSigma,VLmi] = mussvextract(muinfo)

Description

A structured singular value computation of the form

[bounds,muinfo] = mussv(M,BlockStructure)

returns detailed information in the structure muinfo. mussvextract is used to

extract the compressed information within muinfo into a readable form.
The most general call to mussvextract extracts three usable quantities:
VDelta, VSigma, and VLmi. VDelta is used to verify the lower bound. VSigma is
used to verify the Newlin/Young upper bound and has fields DLeft, DRight,
GLeft, GMiddle, and GRight. VLmi is used to verify the LMI upper bound and
has fields Dr, Dc, Grc, and Gcr. The relation/interpretation of these quantities
with the numerical results in bounds is described below.

Upper Bound Information

The upper bound is based on a proof that det(I - M*Delta) is nonzero for all
block-structured matrices Delta with norm smaller than 1/bounds(1). The
Newlin/Young method consists of finding a scalar and matrices D and G,
consistent with BlockStructure, such that
1

-- --

2 4 Dr
2 4
- jG m (I + G r ) 1
(I + G l ) --------

Because some uncertainty blocks and M need not be square, the matrices D and
G have a few different manifestations. In fact, in the formula above, there are
a left and right D and G, as well as a middle G. Any such is an upper bound
of mussv(M,BlockStructure).
It is true that if BlockStructure consists only of complex blocks, then all G
matrices will be zero, and the expression above simplifies to
1

( D r ) .
The LMI method consists of finding a scalar and matrices D and G, consistent
with BlockStructure, such that

10-220

mussvextract

M'D rM D c + j* ( G cr M M'G rc )
is negative semidefinite. Again, D and G have a few different manifestations to
match the row and column dimensions of M. Any such is an upper bound of
mussv(M,BlockStructure). If BlockStructure consists only of complex blocks,
then all G matrices will be zero, and negative semidefiniteness of MDrM2Dc
is sufficient to derive an upper bound.

Lower Bound Information

The lower bound of mussv(M,BlockStructure) is based on finding a small
(hopefully the smallest) block-structured matrix VDelta that causes det(I M*VDelta) to equal 0. Equivalently, the matrix M*VDelta has an eigenvalue
equal to 1. It will always be true that the lower bound (bounds(2)) will be the
reciprocal of norm(VDelta).

Examples

Suppose M is a 4-by-4 complex matrix. Take the block structure to be two 1-by-1
complex blocks and one 2-by-2 complex block.
M = randn(4,4) + sqrt(-1)*randn(4,4);
BlockStructure = [1 1;1 1;2 2];

You can calculate bounds on the structured singular value using the mussv
command and extract the scaling matrices using mussvextract.
[bounds,muinfo] = mussv(M,BlockStructure);
[VDelta,VSigma,VLmi] = mussvextract(muinfo);

You can first verify the Newlin/Young upper bound with the information
extracted from muinfo. The corresponding scalings are Dl and Dr.
Dl = VSigma.DLeft
Dl =
1.0000
0
0
0.8393
0
0
0
0
Dr = VSigma.DRight
Dr =
1.0000
0
0
0.8393
0
0

0
0
0.8917
0

0
0
0
0.8917

0
0
0.8917

0
0
0

10-221

mussvextract

0
0
[norm(Dl*M/Dr) bounds(1)]
ans =
4.2013
4.2013

0.8917

You can first verify the LMI upper bound with the information extracted from
muinfo. The corresponding scalings are Dr and Dc.
Dr = VLmi.Dr;
Dc = VLmi.Dc;
eig(M'*Dr*M - bounds(1)^2*Dc)
ans =
-15.2960 - 0.0000i
-0.0000 - 0.0000i
-10.1037 + 0.0000i
-5.9719 + 0.0000i

Note that VDelta matches the structure defined by BlockStructure, and the
norm of VDelta agrees with the lower bound,
VDelta
VDelta =
0.2009 - 0.1277i
0
0
0.1648 - 0.1718i
0
0
- 0.1012i
0
0
+ 0.1460i
[norm(VDelta) 1/bounds(2)]
ans =
0.2380
0.2380

0
0
0
0
-0.0408 + 0.0773i -0.0291
0.0635 - 0.1070i

and that M*VDelta has an eigenvalue exactly at 1.

eig(M*VDelta)
ans =
1.0000 + 0.0000i
0.0000 + 0.0000i
-0.1161 + 0.0225i
-0.0837 + 0.6532i

10-222

0.0341

mussvextract

Keep the matrix the same, but change BlockStructure to be a 2-by-2 repeated,
real scalar block and two complex 1-by-1 blocks. Run mussv with the 'C' option
to tighten the upper bound.
BlockStructure2 = [-2 0; 1 0; 1 0];
[bounds2,muinfo2] = mussv(M,BlockStructure2,'C');

You can compare the computed bounds. Note that bounds2 should be smaller
than bounds, because the uncertainty set defined by BlockStructure2 is a
proper subset of that defined by BlockStructure.
[bounds; bounds2]
ans =
4.2013
4.2009
4.0005
4.0005

You can extract the D, G and Delta from muinfo2 using mussvextract.
[VDelta2,VSigma2,VLmi2] = mussvextract(muinfo2);

As before, you can first verify the Newlin/Young upper bound with the
information extracted from muinfo. The corresponding scalings are Dl, Dr,
Gl, Gm and Gr.
Dl = VSigma2.DLeft;
Dr = VSigma2.DRight;
Gl = VSigma2.GLeft;
Gm = VSigma2.GMiddle;
Gr = VSigma2.GRight;
dmd = Dl*M/Dr/bounds2(1) - sqrt(-1)*Gm;
SL = (eye(4)+Gl*Gl)^-0.25;
SR = (eye(4)+Gr*Gr)^-0.25;
norm(SL*dmd*SR)
ans =
1.0000

You can first verify the LMI upper bound with the information extracted from
muinfo. The corresponding scalings are Dr, Dc, Grc and Gcr.
Dr = VLmi2.Dr;
Dc = VLmi2.Dc;
Grc = VLmi2.Grc;
Gcr = VLmi2.Gcr;

10-223

mussvextract

eig(M'*Dr*M
ans =
-0.0075 -0.0009 -0.0001 -0.0004 -

- bounds(1)^2 *Dc + j*(Gcr*M-M'*Grc))

0.0000i
0.0000i
0.0000i
0.0000i

Note that VDelta2 matches the structure defined by BlockStructure, and the
norm of VDelta2 agrees with the lower bound,
VDelta2
VDelta2 =
0.2500
0
0
0.2500
0
0
0
0
[norm(VDelta2) 1/bounds2(2)]
ans =
0.2500
0.2500

0
0
-0.0130 + 0.2496i
0
0.1402

0
0
0
+ 0.2069i

and that M*VDelta2 has an eigenvalue exactly at 1.

eig(M*VDelta2)
ans =
1.0000 + 0.0000i
-0.3097 + 0.6127i
-0.2287 + 0.0747i
-0.2692 - 0.4897i

See Also

10-224

mussv

Calculate bounds on the Structured Singular Value ()

ncfmargin

Purpose

10ncfmargin

Calculate normalized coprime stability margin of feedback loop

Syntax

[marg,freq] = ncfmargin(P,C)
[marg,freq] = ncfmargin(P,C,tol)

Description

[marg,freq] = ncfmargin(P,C) calculates the normalized coprime factor/gap

metric robust stability margin b(P, C), marg, of the multivariable feedback loop
consisting of C in negative feedback with P. The normalized coprime factor b(P,

C) is defined as
1

b ( P, C ) =

I ( I PC ) 1
PI
C

C should only be the compensator in the feedback path, such as the 1-dof
architecture shown below (on the right). If the compensator has 2-dof
architecture shown below (on the left), you must eliminate the reference
channels before calling ncfmargin. freq is the frequency associated with the
upper bound on marg.

 

 
   

 
   

The normalized coprime factor robust stability margin lies between 0 and 1 and
is used as an indication of robustness to unstructured perturbations. Values of
marg greater than 0.3 generally indicate good robustness margins.
[marg,freq] = ncfmargin(P,C,tol) calculates the normalized coprime

factor/gap metric robust stability of the multivariable feedback loop consisting

of C in negative feedback with P. tol specifies a relative accuracy for
calculating the normalized coprime factor metric and must be between 10-5 and
10-2. tol=0.001 is the default value.

Examples

Consider the plant model 4/(s-0.001) an unstable first order, and two
constant gain controllers, k1 = 1 and k2 = 10. Both controllers stabilize the
closed-loop system

10-225

ncfmargin

x = tf(4,[1 0.001]);
clp1 = feedback(x,1)

The transfer function clp1 is shown as is clp2.

4
--------s + 4.001
clp2 = feedback(x,10)
Transfer function:
4
-----s + 40

The closed-loop system with controller k1, clp1, has a normalized coprime
factor robust stability margin of 0.71 that is achieved at infinite frequency.
This indicates that the closed-loop system is very robust to unstructured
perturbations. The closed-loop system with controller k2, clp2, has a
normalized coprime factor robust stability margin of 0.10. This indicates that
the closed-loop system is not robust to unstructured perturbations.
[marg1,freq1] = ncfmargin(x,1)
marg1 =
0.7071
freq1 =
Inf
[marg2,freq2] = ncfmargin(x,10)
marg2 =
0.0995
freq2 =
Inf

Construct an uncertain system, xu, by adding an 11% unmodeled dynamics to

the nominal system x. Calculate the robust stability of the closed-loop system
with the feedback gain 1 and 10.
xu = x + ultidyn('uncstruc',[1 1],'Bound',0.11);
[stabmarg1, du1, report1] = robuststab(feedback(xu,1));
disp(report1{1})
Uncertain System is robustly stable to modeled uncertainty.

10-226

ncfmargin

-- It can tolerate up to 909% of modeled uncertainty.

-- A destabilizing combination of 909% the modeled uncertainty
exists, causing an instability at 165 rad/s.
[stabmarg10, du10, report10] = robuststab(feedback(xu,10));
disp(report10{1})
Uncertain System is NOT robustly stable to modeled uncertainty.
-- It can tolerate up to 90.9% of modeled uncertainty.
-- A destabilizing combination of 90.9% the modeled uncertainty
exists, causing an instability at 1.64e+003 rad/s.

The closed-loop system with K=1 is robustly stable in the presence of the
unmodeled dynamics based on the robust stability analysis. In fact, the
closed-loop system with K=1 can tolerate 909% (or 9.09*11%) of the unmodeled
LTI dynamics, whereas the closed-loop system is not robustly stable with a
constant gain of 10 controller. The closed-loop system with K=10 implemented
can only tolerate 90.9% (or.909*11%) of the unmodeled LTI dynamics.

Algorithm

The computation of the gap amounts to solving 2-block H problems, Georgiou,

Smith, 1988. The particular method used here for solving the H problems is
based on Green et al., 1990. The computation of the nugap uses the method of
Vinnicombe, 1993.

References

McFarlane, D.C. and K. Glover, Robust Controller Design using Normalised

Coprime Factor Plant Descriptions, Lecture Notes in Control and
Information Sciences, Springer Verlag, Vol. 138, 1989.
McFarlane, D.C., and K. Glover, A Loop Shaping Design Procedure using
Synthesis, IEEE Transactions on Automatic Control, Vol. 37, No. 6, 1992,
pp. 759-769.
Vinnicombe, G., Measuring Robustness of Feedback Systems, Ph.D.
Dissertation, Department of Engineering, University of Cambridge, 1993.

See Also

loopmargin
gapmetric
norm
wcmargin

Performs a comprehensive analysis of feedback loop

Computes the gap and the Vinnicombe gap metric
Computes the norm of a system
Calculate worst-case margins for feedback loop

10-227

ncfmr

Purpose

10ncfmr

Balanced normalized coprime factor model reduction

Syntax

GRED = ncfmr(G)
GRED = ncfmr(G,order)
[GRED,redinfo] = ncfmr(G,key1,value1,...)
[GRED,redinfo] = ncfmr(G,order,key1,value1,...)

Description

ncfmr returns a reduced order model GRED formed by a set of balanced

normalized coprime factors and a struct array redinfo containing the left and
right coprime factors of G and their coprime Hankel singular values.

Hankel singular values of coprime factors of such a stable system indicate the
respective state energy of the system. Hence, reduced order can be directly
determined by examining the system Hankel SVs.
With only one input argument G, the function will show a Hankel singular
value plot of the original model and prompt for model order number to reduce.
The left and right normalized coprime factors are defined as [1]
Left Coprime Factorization: G = M l 1 ( s )N l ( s )
Right Coprime Factorization: G = N r ( s )M r1 ( s )
where there exist stable U r ( s ) , V r ( s ) , U l ( s ) , and V l ( s ) such that
Ur Nr + Vr Mr = I
Nl Ul + Ml Vl = I
The left/right coprime factors are stable, hence implies Mr(s) should contain as
RHP-zeros all the RHP-poles of G(s). The comprimeness also implies that there
should be no common RHP-zeros in Nr(s) and Mr(s), i.e., when forming
G = N r ( s )M r 1 ( s ) , there should be no pole-zero cancellations.

10-228

ncfmr

This table describes input arguments for ncmfr.

Argument

Description

LTI model to be reduced (without any other inputs will plot

its Hankel singular values and prompt for reduced order)

ORDER

(Optional) an integer for the desired order of the reduced

model, or optionally a vector packed with desired orders for
batch runs

A batch run of a serial of different reduced order models can be generated by

specifying order = x:y, or a vector of integers. By default, all the anti-stable
part of a system is kept, because from control stability point of view, getting rid
of unstable state(s) is dangerous to model a system. ncfmr method allows the
original model to have j-axis singularities.
'MaxError' can be specified in the same fashion as an alternative for 'ORDER'.
In this case, reduced order will be determined when the sum of the tails of the
Hankel SVs reaches the 'MaxError'.
Argument

Value

Description

'MaxError'

A real number or
a vector of
different errors

Reduce to achieve H error.

When present, 'MaxError'overides
ORDER input.

'Display'

'on' or 'off'

Display Hankel singular plots (default

'off').

'Order'

integer, vector or
cell array

Order of reduced model. Use only if not

specified as 2nd argument.

Weights on the original model input and/or output can make the model
reduction algorithm focus on some frequency range of interests. But weights
have to be stable, minimum phase and invertible.

10-229

ncfmr

This table describes output arguments.

Argument

Description

GRED

LTI reduced order model, that becomes multi-dimensional

array when input is a serial of different model order array.

REDINFO

A STRUCT array with 3 fields:

REDINFO.GL (left coprime factor)
REDINFO.GR (right coprime factor)
REDINFO.hsv (Hankel singular values)

G can be stable or unstable, continuous or discrete.

Algorithm

Given a state space (A,B,C,D) of a system and k, the desired reduced order, the
following steps will produce a similarity transformation to truncate the
original state-space system to the kth order reduced model.
1 Find the normalized coprime factors of G by solving Hamiltonian described

in [1].
Gl = Nl Ml
Nr

Gr =

Mr

2 Perform kth order square root balanced model truncation on Gl (or Gr) [2].
3 The reduced model GRED is [2]:

Ac Bm Cl Bn Bm Dl
Cl

where
Nl:= (Ac, Bn, Cc, Dn)

10-230

Dl

ncfmr

Ml:= (Ac, Bm, Cc, Dm)

Cl = (Dm)-1Cc
Dl = (Dm)-1Dn

Examples

Given a continuous or discrete, stable or unstable system, G, the following

commands can get a set of reduced order models based on your selections:
rand('state',1234); randn('state',5678);
G = rss(30,5,4); G.d = zeros(5,4);
[g1, redinfo1] = ncfmr(G); % display Hankel SV plot
% and prompt for order (try 15:20)
[g2, redinfo2] = ncfmr(G,20);
[g3, redinfo3] = ncfmr(G,[10:2:18]);
[g4, redinfo4] = ncfmr(G,'MaxError',[0.01, 0.05]);
for i = 1:4
figure(i); eval(['sigma(G,g' num2str(i) ');']);
end

Reference

[1] Vidyasagar, M. Control System Synthesis - A Factorization Approach.

London: The MIT Press, 1985.
[2] M. G. Safonov and R. Y. Chiang, A Schur Method for Balanced Model
Reduction, IEEE Trans. on Automat. Contr., vol. AC-2, no. 7, July 1989, pp.
729-733.

See Also

reduce
balancmr
schurmr
bstmr
hankelmr
hankelsv

Top level model reduction routines

Balanced truncation via square-root method
Balanced truncation via Schur method
Balanced stochastic truncation via Schur method
Hankel minimum degree approximation
Hankel singular value

10-231

ncfsyn

Purpose

10ncfsyn

H normalized coprime factor controller synthesis

Syntax

[K,CL,GAM,INFO]=ncfsyn(G)
[K,CL,GAM,INFO]=ncfsyn(G,W1)
[K,CL,GAM,INFO]=ncfsyn(G,W1,W2)

Description

ncfsyn is a method for designing controllers that uses a combination of loop

shaping and robust stabilization as proposed in McFarlane and Glover [1]-[2].

The first step is for you to select a pre- and post-compensator W1 and W2, so
that the gain of the shaped plant Gs:= W2GW1 is sufficiently high at
frequencies where good disturbance attenuation is required and is sufficiently
low at frequencies where good robust stability is required. The second step is
to use ncfsyn to compute an optimal positive feedback controllers K.
The optimal Ks has the property that the sigma plot of the shaped loop
Ls=W2*G*W1*Ks

matches the target loopshape Gs optimally, roughly to within plus or minus

20*log10(GAM) decibels. The number margin GAM=1/ncfmargin(Gs,K) and is
always greater than 1. GAM gives a good indication of robustness of stability to
a wide class of unstructured plant variations, with values in the range 1<GAM<3
corresponding to satisfactory stability margins for most typical control system
designs.

Algorithm

K=W2*Ks*W1, where Ks =K is an optimal

H controller that simultaneously

minimizes the two H cost functions

:= min
K

I ( I G K ) 1 [ G , I ]
s
s
K

:= min
K

I
1
( I KG s ) [ K, I ]
Gs

Roughly speaking, this means for most plants that

(W2GW1 K), db = (W2GW1), db , db
(KW2GW1), db = (W2GW1), db , db,

10-232

ncfsyn

so you can use the weights W1 and W2 for loop shaping. For a more precise
bounds on loop shaping accuracy, see Theorem 16.12 of Zhou and Glover [1].
Theory ensures that if Gs=NM1is a normalized coprime factorization (NCF) of
the weighted plant model Gs satisfying
Gs=N(jw)*N(jw) + M(jw)*M(jw) = I,
then the control system will remain robustly stable for any perturbation G s to
the weighted plant model Gs that can be written
1
G s = ( N + 1 ) ( M + 2 )

for some stable pair 1, 2 satisfying

1
2

< MARG:=1/GAM

The closed-loop H-norm objective has the standard signal gain interpretation.
Finally it can be shown that the controller, K, does not substantially affect the
loop shape in frequencies where the gain of W2GW1 is either high or low, and
will guarantee satisfactory stability margins in the frequency region of gain
cross-over. In the regulator set-up, the final controller to be implemented is
K=W1KW2.
Input arguments:
G
W1,W2

LTI plant to be controlled

Stable min-phase LTI weights, either SISO or MIMO.
Default is W1=I, W2=I

Output arguments:
K
CL

LTI controller K= W1*Ks*W2

I
1
( I W 2 GW 1 K ) [ W 2 GW 1, I ] , LTI H optimal closed loop
K

10-233

ncfsyn

GAM

1
H optimal cost = ----------------------------------------= hinfnorm(CL) 1
b ( W 2 GW 1, K )

INFO

Structure array containing additional information

Additional output INFO fields:

INFO.emax

nugap robustness
emax=1/GAM=ncfmargin(Gs,-Ks)= b ( W 2 GW 1, K )

INFO.Gs

'shaped plant' Gs=W2*G*W1

INFO.Ks

Ks = K

= NCFSYN(Gs) = NCFSYN(W2*G*W1)

[MARG,FREQ] = ncfmargin(G,K,TOL) calculates the normalized coprime

factor/gap metric robust stability margin assuming negative feedback.

MARG =b ( G, K ) = 1

I ( I + GK ) 1 [ G , I ]
K

where G and K are LTI plant and controller, and TOL (default=.001) is the
tolerance used to compute the H norm. FREQ is the peak frequency i.e., the
frequency at which the infinity norm is reached to within TOL.

Algorithm

See McFarlane and Glover [1], [2] for details.

Examples

The following code shows how ncfsyn can be used for loop-shaping. The
achieved loop G*K has a sigma plot is equal to that of the target loop G*W1 to
within plus or minus 20*log10(GAM) decibels.
s=zpk('s');
G=(s-1)/(s+1)^2;
W1=0.5/s;
[K,CL,GAM]=ncfsyn(G,W1);
sigma(G*K,'r',G*W1,'r-.',G*W1*GAM,'k-.',G*W1/GAM,'k-.')

10-234

ncfsyn

Figure 10-15: Achieved loop G*K and shaped loop Gs, 20log(GAM) dB

References

[1] McFarlane, D.C., and K. Glover, Robust Controller Design using

Normalised Coprime Factor Plant Descriptions, Springer Verlag, Lecture Notes
in Control and Information Sciences, vol. 138, 1989.
[2] McFarlane, D.C., and K. Glover, A Loop Shaping Design Procedure using
Synthesis, IEEE Transactions on Automatic Control, vol. 37, no. 6, pp. 759
769, June 1992.
[3] Vinnicombe, G., Measuring Robustness of Feedback Systems, PhD
dissertation, Department of Engineering, University of Cambridge, 1993.
[4] Zhou, K., and J.C. Doyle, Essentials of Robust Control. NY: Prentice-Hall,
1998.

See Also

gapmetric
hinfsyn
loopsyn
ncfmargin

Computes the gap and the Vinnicombe gap metric

H controller synthesis
H - loop shaping controller synthesis
Normalized coprime stability margin of the
plant-controller feedback loop

10-235

newlmi

Purpose

10newlmi

Add new LMI to LMI system

Syntax

tag = newlmi

Description

newlmi adds a new LMI to the LMI system currently described and returns an
identifier tag for this LMI. This identifier can be used in lmiterm, showlmi, or
dellmi commands to refer to the newly declared LMI. Tagging LMIs is optional
and only meant to facilitate code development and readability.

Identifiers can be given mnemonic names to help keep track of the various
LMIs. Their value is simply the ranking of each LMI in the system (in the order
of declaration). They prove useful when some LMIs are deleted from the LMI
system. In such cases, the identifiers are the safest means of referring to the
remaining LMIs.

See Also

setlmis
lmivar
lmiterm
getlmis
lmiedit
dellmi

10-236

Initialize the description of an LMI system

Specify the matrix variables in an LMI problem
Specify the term content of LMIs
Get the internal description of an LMI system
Specify or display systems of LMIs as MATLAB
expressions
Remove an LMI from a given system of LMIs

normalized2actual

Purpose

10normalized2actual

Convert value for atom in normalized coordinates to corresponding actual

value

Syntax

avalue = normalizedactual2(A,NV)

Description

avalue = normalizedactual2(A,NV) converts the value for atom in

normalized coordinates, NV, to the corresponding actual value.

If NV is an array of values, then avalue will be an array of the same dimension.

Examples

Create uncertain real parameters with a range that is symmetric about the
nominal value, where each endpoint is 1 unit from the nominal. Points that lie
inside the range are less than 1 unit from the nominal, while points that lie
outside the range are greater than 1 unit from the nominal.
a = ureal('a',3,'range',[1 5]);
actual2normalized(a,[1 3 5])
ans =
-1.0000
-0.0000
1.0000
normalized2actual(a,[-1 1])
ans =
1.0000
5.0000
normalized2actual(a,[-1.5 1.5])
ans =
0.0000
6.0000

See Also

actual2normalized
robuststab
robustperf

Calculates of an uncertain atom

Calculates robust stability margin
Calculates robust performance margin

10-237

pdlstab

Purpose

10pdlstab

Assess robust stability of polytopic or parameter-dependent system

Syntax

[tau,Q0,Q1,...] = pdlstab(pds,options)

Description

pdlstab uses parameter-dependent Lyapunov functions to establish the

stability of uncertain state-space models over some parameter range or

polytope of systems. Only sufficient conditions for the existence of such
Lyapunov functions are available in general. Nevertheless, the resulting
robust stability tests are always less conservative than quadratic stability tests
when the parameters are either time-invariant or slowly varying.
For an affine parameter-dependent system
E(p) x = A(p)x + B(p)u
y = C(p)x + D(p)u
with p = (p1, . . ., pn) Rn, pdlstab seeks a Lyapunov function of the form
V(x, p) = xTQ(p)1x, Q(p) = Q0 + p1Q1 + . . .pnQn
such that dV(x, p)/dt < 0 along all admissible parameter trajectories. The
system description pds is specified with psys and contains information about
the range of values and rate of variation of each parameter pi.
For a time-invariant polytopic system
E x = Ax + Bu
y = Cx + Du
with
A + jE B

C D

A + jE B
i i
,
i

C
D i
i

i=1

i 0 ,

i = 1 ,
i=1

pdlstab seeks a Lyapunov function of the form

V(x, ) = xTQ()1x, Q() = 1Q1 + . . .+ nQn

such that dV(x, )/dt < 0 for all polytopic decompositions (10-19).

10-238

(10-19)

pdlstab

Several options and control parameters are accessible through the optional
argument options:
Setting options(1)=0 tests robust stability (default).
When options(2)=0, pdlstab uses simplified sufficient conditions for faster
running times. Set options(2)=1 to use the least conservative conditions.

Remark

For affine parameter-dependent systems with time-invariant parameters,

there is equivalence between the robust stability of
E ( p )x = A ( p )x

(10-20)

and that of the dual system

T
T
E ( p ) z = A ( p ) z

(10-21)

However, the second system may admit an affine parameter-dependent

Lyapunov function while the first does not.
In such case, pdlstab automatically restarts and tests stability on the dual
system (10-21) when it fails on (10-20).

See Also

quadstab

Quadratic stability of polytopic or affine

parameter-dependent systems

10-239

pdsimul

Purpose

10pdsimul

Time response of parameter-dependent system along parameter trajectory

Syntax

pdsimul(pds,'traj',tf,'ut',xi,options)
[t,x,y] = pdsimul(pds,pv,'traj',tf,'ut',xi,options)

Description

pdsimul simulates the time response of an affine parameter-dependent system

E(p) x = A(p)x + B(p)u

y = C(p)x + D(p)u
along a parameter trajectory p(t) and for an input signal u(t). The parameter
trajectory and input signals are specified by two time functions p=traj(t) and
u=ut(t). If 'ut' is omitted, the response to a step input is computed by default.
The affine system pds is specified with psys. The function pdsimul also accepts
the polytopic representation of such systems as returned by aff2pol(pds) or
hinfgs. The final time and initial state vector can be reset through tf and xi
(their respective default values are 5 seconds and 0). Finally, options gives
access to the parameters controlling the ODE integration (type help gear for
details).
When invoked without output arguments, pdsimul plots the output trajectories
y(t). Otherwise, it returns the vector of integration time points t as well as the
state and output trajectories x,y.

See Also

10-240

psys
pvec

Specification of uncertain state-space models

Quantification of uncertainty on physical parameters

polydec

Purpose

10polydec

Compute polytopic coordinates with respect to box corners

Syntax

vertx = polydec(PV)
[C,vertx] = polydec(PV,P)

Description

vertx = polydec(PV) takes an uncertain parameter vector PV taking values

ranging in a box, and returns the corners or vertices of the box as columns of
the matrix vertx.
[C,vertx] = polydec(PV,P) takes an uncertain parameter vector PV and a
value P of the parameter vector PV, and returns the convex decomposition C of
P over the set VERTX of box corners:
P = c1*VERTX(:,1) + ... + cn*VERTX(:,n)
cj >=0 ,
c1 + ... + cn = 1

The list vertx of corners can be obtained directly by typing

vertx = polydec(PV)

See Also

pvec
pvinfo
aff2pol
hinfgs

Quantification of uncertainty on physical parameters

Describe a parameter vector specified with pvec
Convert affine parameter-dependent models to
polytopic ones
Synthesis of gain-scheduled H controllers

10-241

popov

Purpose

10popov

Perform Popov robust stability test

Syntax

[t,P,S,N] = popov(sys,delta,flag)

Description

popov uses the Popov criterion to test the robust stability of dynamical systems
with possibly nonlinear and/or time-varying uncertainty. The uncertain
system must be described as the interconnection of a nominal LTI system sys
and some uncertainty delta.

The command
[t,P,S,N] = popov(sys,delta)

tests the robust stability of this interconnection. Robust stability is guaranteed

if t < 0. Then P determines the quadratic part xTPx of the Lyapunov function
and D and S are the Popov multipliers.
If the uncertainty delta contains real parameter blocks, the conservatism of
the Popov criterion can be reduced by first performing a simple loop
transformation. To use this refined test, call popov with the syntax
[t,P,S,N] = popov(sys,delta,1)

See Also

quadstab
pdlstab

10-242

Quadratic stability of polytopic or affine

parameter-dependent systems
Robust stability of polytopic or affine
parameter-dependent systems (P-system)

psinfo

Purpose

10psinfo

Query characteristics of polytopic or parameter-dependent systems

Syntax

psinfo(ps)
[type,k,ns,ni,no] = psinfo(ps)
pv = psinfo(ps,'par')
sk = psinfo(ps,'sys',k)
sys = psinfo(ps,'eval',p)

Description

psinfo is a multi-usage function for queries about a polytopic or

parameter-dependent system ps created with psys. It performs the following

operations depending on the calling sequence:

psinfo(ps) displays the type of system (affine or polytopic); the number k of
SYSTEM matrices involved in its definition; and the numbers of ns, ni, no of
states, inputs, and outputs of the system. This information can be optionally
stored in MATLAB variables by providing output arguments.
pv = psinfo(ps,'par') returns the parameter vector description (for
parameter-dependent systems only).
sk = psinfo(ps,'sys',k) returns the k-th SYSTEM matrix involved in the
definition of ps. The ranking k is relative to the list of systems syslist used
in psys.
sys = psinfo(ps,'eval',p) instantiates the system for a given vector p of
parameter values or polytopic coordinates.
For affine parameter-dependent systems defined by the SYSTEM matrices S0,
S1, . . ., Sn, the entries of p should be real parameter values p1, . . ., pn and
the result is the LTI system of SYSTEM matrix
S(p) = S0 + p1S1 + . . .+ pnSn
For polytopic systems with SYSTEM matrix ranging in
Co{S1, . . ., Sn},
the entries of p should be polytopic coordinates p1, . . ., pn satisfying pj 0
and the result is the interpolated LTI system of SYSTEM matrix
p1 S1 + + pn S
S = ---------------------------------------------np1 + + pn

See Also

psys

Specification of uncertain state-space models

10-243

psys

Purpose

10psys

Specify polytopic or parameter-dependent linear system

Syntax

pols = psys(syslist)
affs = psys(pv,syslist)

Description

psys specifies state-space models where the state-space matrices can be

uncertain, time-varying, or parameter-dependent.

Two types of uncertain state-space models can be manipulated in the LMI

Control Toolbox:
Polytopic systems
E(t) x = A(t)x + B(t)u
y = C(t)x + D(t)u
whose SYSTEM matrix takes values in a fixed polytope:

A ( t ) + jE ( t ) B ( t ) Co A 1 + jE 1 B 1 , . . ., A k + jE k B k

C(t)
D(t)
C1
D1
Ck
Dk

S(t)
S1
Sk

k given vertex systems

k
and
where S1, . . ., Sk are

Co{S1, . . ., Sk} =
i Si : i 0 ,
i = 1

denotes the convex

1 of S1, . . ., Ski(polytope
=1
i =hull
of matrices with vertices
S1, . . ., Sk)

Affine parameter-dependent systems

E(p) x = A(p)x + B(p)u
y = C(p)x + D(p)u
where A(.); B(.), . . ., E(.) are fixed affine functions of some vector
p = (p1, . . ., pn) of real parameters, i.e.,

10-244

psys

A ( p ) + jE ( p ) B ( p ) =
C(p)
D(p)
S(p)

A 0 + jE 0 B 0
A + jE 1 B 1
A + jE n B n
+ p1 1
+ . . . + pn n
C0
D0
C1
D1
Cn
Dn
S0

S1

Sn

where S0, S1, . . ., Sn are given SYSTEM matrices. The parameters pi can be
time-varying or constant but uncertain.
Both types of models are specified with the function psys. The argument
syslist lists the SYSTEM matrices Si characterizing the polytopic value set or
parameter dependence. In addition, the description pv of the parameter vector
(range of values and rate of variation) is required for affine parameterdependent models (see pvec for details). Thus, a polytopic model with vertex
systems S1, . . ., S4 is created by
pols = psys([s1,s2,s3,s4])

while an affine parameter-dependent model with 4 real parameters is defined

by
affs = psys(pv,[s0,s1,s2,s3,s4])

The output is a structured matrix storing all the relevant information.

See Also

psinfo
pvec
aff2pol

Inquire about polytopic or parameter-dependent

systems created with psys
Quantification of uncertainty on physical parameters
Convert affine parameter-dependent models to
polytopic ones

10-245

pvec

Purpose

10pvec

Specify vector of uncertain or time-varying parameters

Syntax

pv = pvec('box',range,rates)
pv = pvec('pol',vertices)

Description

pvec is used in conjunction with psys to specify parameter-dependent systems.

Such systems are parametrized by a vector p = (p1, . . ., pn) of uncertain or

time-varying real parameters pi. The function pvec defines the range of values
and the rates of variation of these parameters.
The type 'box' corresponds to independent parameters ranging in intervals
pj pj pj
The parameter vector p then takes values in a hyperrectangle of Rn called the
parameter box. The second argument range is an n-by-2 matrix that stacks up
the extremal values p j and p j of each pj. If the third argument rates is
omitted, all parameters are assumed time-invariant. Otherwise, rates is also
an n-by-2 matrix and its j-th row specifies lower and upper bounds j and j
dp j
-:
on -------dt
dp j
j --------- j
dt
Set j = Inf and j = Inf if pj(t) can vary arbitrarily fast or discontinuously.
The type 'pol' corresponds to parameter vectors p ranging in a polytope of the
parameter space Rn. This polytope is defined by a set of vertices V1, . . ., Vn
corresponding to extremal values of the vector p. Such parameter vectors are
declared by the command
pv = pvec('pol',[v1,v2, . . ., vn])

where the second argument is the concatenation of the vectors v1,...,vn.

The output argument pv is a structured matrix storing the parameter vector
description. Use pvinfo to read the contents of pv.

Examples

Consider a problem with two time-invariant parameters

p1 [1, 2], p2 [20, 50]

10-246

pvec

The corresponding parameter vector p = (p1, p2) is specified by

pv = pvec('box',[-1 2;20 50])

Alternatively, this vector can be regarded as taking values in the rectangle

drawn in Figure 9.2. The four corners of this rectangle are the four vectors

v1 = 1 , v2 = 1 ,
20
50

v3 = 2 ,
20

v4 = 2
50

Hence, you could also specify p by

pv = pvec('pol',[v1,v2,v3,v4])

p2
50

20

-1

p1

Figure 10-16: Parameter box

See Also

pvinfo
psys

Describe a parameter vector specified with pvec

Specification of uncertain state-space models

10-247

pvinfo

Purpose

10pvinfo

Query characteristics of parameter vector

Syntax

[typ,k,nv] = pvinfo(pv)
[pmin,pmax,dpmin,dpmax] = pvinfo(pv,'par',j)
vj = pvinfo(pv,'par',j)
p = pvinfo(pv,'eval',c)

Description

pvinfo retrieves information about a vector p = (p1, . . ., pn) of real parameters

declared with pvec and stored in pv. The command pvinfo(pv) displays the
type of parameter vector ('box' or 'pol'), the number n of scalar parameters,
and for the type 'pol', the number of vertices used to specify the parameter

range.
For the type 'box':
[pmin,pmax,dpmin,dpmax] = pvinfo(pv,'par',j)

returns the bounds on the value and rate of variations of the j-th real
parameter pj. Specifically,
dp j

pmin p ( t ) pmax , dpmin --------- dpmax

dt

For the type 'pol':

pvinfo(pv,'par',j)

returns the j-th vertex of the polytope of Rn in which p ranges, while

pvinfo(pv,'eval',c)

returns the value of the parameter vector p given its barycentric coordinates c
with respect to the polytope vertices (V1, . . .,Vk). The vector c must be of length
k and have nonnegative entries. The corresponding value of p is then given by
cV

i=1 i i
---------------------------k
i = 1 c i
k

p =

See Also

10-248

pvec
psys

Quantification of uncertainty on physical parameters

Specification of uncertain state-space models

quadperf

Purpose

Compute quadratic H performance of polytopic or parameter-dependent

system

Syntax

[perf,P] = quadperf(ps,g,options)

Description

The RMS gain of the time-varying system

10quadperf

E ( t )x = A ( t )x + B ( t )u,

y = C ( t )x + D ( t )u

(10-22)

is the smallest > 0 such that

y L u L
2
2

(10-23)

for all input u(t) with bounded energy. A sufficient condition for (10-23) is the
existence of a quadratic Lyapunov function
V(x) = xTPx,

P>0

such that
T
2 T
u L 2 , dV
-------- + y y u u < 0
dt

Minimizing over such quadratic Lyapunov functions yields the quadratic H

performance, an upper bound on the true RMS gain.
The command
[perf,P] = quadperf(ps)

computes the quadratic H performance perf when (10-22) is a polytopic or

affine parameter-dependent system ps (see psys). The Lyapunov matrix P
yielding the performance perf is returned in P.
The optional input options gives access to the following task and control
parameters:
If options(1)=1, perf is the largest portion of the parameter box where the
quadratic RMS gain remains smaller than the positive value g (for affine
parameter-dependent systems only). The default value is 0.
If options(2)=1, quadperf uses the least conservative quadratic
performance test. The default is options(2)=0 (fast mode).

10-249

quadperf

options(3) is a user-specified upper bound on the condition number of P (the

default is 109).

See Also

quadstab
psys

10-250

Quadratic stability of polytopic or affine

parameter-dependent systems
Specification of uncertain state-space models

quadstab

Purpose

10quadstab

Assess quadratic stability of polytopic or affine parameter-dependent systems

Syntax

[tau,P] = quadstab(ps,options)

Description

For affine parameter-dependent systems

E(p) x = A(p)x,

p(t) = (p1(t), . . ., pn(t))

or polytopic systems
E(t) x = A(t)x,

(A, E) Co{(A1, E1), . . ., (An, En)},

quadstab seeks a fixed Lyapunov function V(x) = xTPx with P > 0 that

establishes quadratic stability. The affine or polytopic model is described by ps

(see psys).
The task performed by quadstab is selected by options(1):
if options(1)=0 (default), quadstab assesses quadratic stability by solving
the LMI problem
Minimize over Q = QT such that
ATQE + EQAT < I for all admissible values of (A, E)
Q>I
The global minimum of this problem is returned in tau and the system is
quadratically stable if tau < 0
if options(1)=1, quadstab computes the largest portion of the specified
parameter range where quadratic stability holds (only available for affine
models). Specifically, if each parameter pi varies in the interval
pi [pi0 i, pi0 + i],
quadstab computes the largest > 0 such that quadratic stability holds over

the parameter box

pi [pi0 i, pi0 + i]
This quadratic stability margin is returned in tau and ps is quadratically
stable if tau 1.
Given the solution Qopt of the LMI optimization, the Lyapunov matrix P is
1
given by P = Q opt . This matrix is returned in P.

10-251

quadstab

Other control parameters can be accessed through options(2) and

options(3):
if options(2)=0 (default), quadstab runs in fast mode, using the least
expensive sufficient conditions. Set options(2)=1 to use the least
conservative conditions
options(3) is a bound on the condition number of the Lyapunov matrix P.
The default is 109.

See Also

pdlstab
decay
quadperf
psys

10-252

Robust stability of polytopic or affine

parameter-dependent systems (P-system)
Quadratic decay rate of polytopic or affine P-systems
Compute the quadratic H performance of a polytopic
or parameter-dependent system
Specification of uncertain state-space models

randatom

Purpose

10randatom

Generate random uncertain atom objects

Syntax

A = randatom(Type)
A = randatom(Type,sz)
A = randatom

Description

A = randatom(Type) generates a 1-by-1 type uncertain object. Valid values for

Type include 'ureal', 'ultidyn', 'ucomplex', and 'ucomplexm'.
A = randatom(Type,sz) generates an sz(1)-by-sz(2) uncertain object. Valid
values for Type include 'ultidyn' or 'ucomplexm'. If Type is set to 'ureal' or
'ucomplex', the size variable is ignored and A is a 1-by-1 uncertain object.
A = randatom, where randatom has no input arguments, results in a 1-by-1
uncertain object. The class is of this object is randomly selected between
'ureal','ultidyn' and 'ucomplex'.

In general, both rand and randn are used internally. You can control the result
of randatom by setting seeds for both random number generators before calling
the function.

Examples

The following statement creates the ureal uncertain object xr. Note that your
display can differ because a random seed is used.
xr = randatom('ureal')
Uncertain Real Parameter: Name BMSJA, NominalValue -6.75,
Range [-7.70893 -1.89278]

The following statement creates the variable ultidyn uncertain object xlti
with three inputs and four outputs. You will get the results shown below if both
the random variable seeds are set to 29.
rand(`seed',29);
randn(`seed',29);
xlti = randatom('ultidyn',[4 3])
Uncertain GainBounded LTI Dynamics: Name OOJGS, 4x3,
Gain Bound = 0.646

See Also

rand
randn
randumat

Generates uniformly distributed random numbers

Generates normally distributed random numbers
Creates a random uncertain matrix

10-253

randatom

randuss
ucomplex
ucomplexm
ultidyn

10-254

Creates a random uncertain system

Creates an uncertain complex parameter
Creates an uncertain complex matrix
Creates an uncertain linear time-invariant object

randumat

Purpose

10randumat

Create random uncertain matrix

Syntax

um = randumat(ny,nu)
um = randumat

Description

um = randumat(ny,nu) generates an uncertain matrix of size ny-by-nu.

randumat randomly selects from uncertain atoms of type 'ureal', 'ultidyn',
and 'ucomplex'.
um = randumat results in a 1-by-1 umat uncertain object, including up to four
uncertain objects.

Example

The following statement creates the umat uncertain object x1 of size 2-by-3.
Note that your result can differ because a random seed is used.
x1 = randumat(2,3)
UMAT: 2 Rows, 3 Columns
ROQAW: complex, nominal = 9.92+4.84i, radius = 0.568, 1
occurrence
UEPDY: real, nominal = -5.81, variability = [-1.98681
0.133993], 3 occurrences
VVNHL: complex, nominal = 5.64-6.13i, radius = 1.99, 2
occurrences

The following statement creates the umat uncertain object x2 of size 4-by-2 with
the seed 91.
rand(`seed',91); randn(`seed',91);
x2 = randumat(4,2)
UMAT: 4 Rows, 2 Columns
SSAFF: complex, nominal = -0.366+2.81i, radius = 1.76, 3
occurrences
VDTIH: complex, nominal = -3.03-3i, +/- 27.5%, 2 occurrences
XOLLJ: real, nominal = 0.0628, range = [-3.73202 4.28174], 1
occurrence

See Also

rand
randn
randatom
randuss
ucomplex

Generate uniformly distributed random numbers

Generate normally distributed random numbers
Create a random uncertain atom
Create a random uncertain system
Creates an uncertain complex parameter

10-255

randumat

ultidyn

10-256

Creates an uncertain linear time-invariant object

randuss

Purpose

10randuss

Create random uncertain state-space model

Syntax

usys
usys
usys
usys
usys

Description

usys = randuss(n) generates an nth order single-input/single-output

uncertain continuous-time system. randuss randomly selects from uncertain
atoms of type 'ureal', 'ultidyn', and 'ucomplex'.

=
=
=
=
=

randuss(n)
randuss(n,p)
randuss(n,p,m)
randuss(n,p,m,Ts)
randuss

usys = randuss(n,p) generates an nth order single-input uncertain

continuous-time system with p outputs.
usys = randuss(n,p,m) generates an nth order uncertain continuous-time
system with p outputs and m inputs.
usys = randuss(n,p,m,Ts) generates an nth order uncertain discrete-time
system with p outputs and m inputs. The sample time is Ts.
usys = randuss (without arguments) results in a 1-by-1 uncertain
continuous-time uss object with up to four uncertain objects.

In general, both rand and randn are used internally. You can control the result
of randuss by setting seeds for both random number generators before calling
the function.

Examples

The statement creates a fifth order, continuous-time uncertain system s1 of

size 2-by-3. Note your display can differ because a random seed is used.
s1 = randuss(5,2,3)
USS: 5 States, 2 Outputs, 3 Inputs, Continuous System
CTPQV: 1x1 LTI, max. gain = 2.2, 1 occurrence
IGDHN: real, nominal = -4.03, variability = [-3.74667
22.7816]%, 1 occurrence
MLGCD: complex, nominal = 8.36+3.09i, +/- 7.07%, 1 occurrence
OEDJK: complex, nominal = -0.346-0.296i, radius = 0.895, 1
occurrence

10-257

randuss

See Also

10-258

rand
randn
randatom
randumat
ucomplex
ultidyn

Generates uniformly distributed random numbers

Generates normally distributed random numbers
Creates a random uncertain atom
Creates a random uncertain matrix
Creates an uncertain complex parameter
Creates an uncertain linear time-invariant object

reduce

Purpose

10reduce

Model order reduction using Hankel singular value based algorithms

Syntax

GRED = reduce(G)
GRED = reduce(G,order)
[GRED,redinfo] = reduce(G,'key1','value1',...)
[GRED,redinfo] = reduce(G,order,'key1','value1',...)

Description

reduce returns a reduced order model GRED of G and a struct array redinfo

containing the error bound of the reduced model, Hankel singular values of the
original system and some other relevant model reduction information.
An error bound is a measure of how close GRED is to G and is computed based
on either additive error, || G-GRED ||, multiplicative error, || G-1(G-GRED) ||, or
nugap error (ncfmr) [1],[4],[5].
Hankel singular values of a stable system indicate the respective state energy
of the system. Hence, reduced order can be directly determined by examining
the system Hankel SV's. Model reduction routines, which based on Hankel SV's
are grouped by their error bound types. In many cases, the additive error
method GRED=reduce(G,ORDER) is adequate to provide a good reduced order
model. But for systems with lightly damped poles and/or zeros, a multiplicative
error method (namely, GRED=reduce(G,ORDER,'ErrorType','mult')) that
minimizes the relative error between G and GRED tends to produce a better fit.
This table describes input arguments for reduce.
Argument

Description

LTI model to be reduced (without any other inputs will plot

its Hankel singular values and prompt for reduced order).

ORDER

(Optional) an integer for the desired order of the reduced

model, or optionally a vector packed with desired orders for
batch runs.

A batch run of a serial of different reduced order models can be generated by

specifying order = x:y, or a vector of integers. By default, all the anti-stable
part of a physical system is kept, because from control stability point of view,
getting rid of unstable state(s) is dangerous to model a system.

10-259

reduce

'MaxError' can be specified in the same fashion as an alternative for 'ORDER'

after an 'ErrorType' is selected. In this case, reduced order will be determined
when the sum of the tails of the Hankel SV's reaches the 'MaxError'.
Argument

Value

Description

'Algorithm'

'balance'
'schur'
'hankel'
'bst'
'ncf'

Default for 'add' (balancmr)

Option for 'add' (schurmr)
Option for 'add' (hankelmr)
Default for 'mult' (bstmr)
Default for 'ncf' (ncfmr)

'ErrorType'

'add'
'mult'
'ncf'

Additive error (default)

Multiplicative error at model output
NCF nugap error

'MaxError'

A real number
or a vector of
different
errors

Reduce to achieve H error.

When present, 'MaxError' overrides
ORDER input.

'Weights'

{Wout,Win}

Optimal 1x2 cell array of LTI weights

cell array

Wout (output) and Win (input); default is

both identity; used only with

'ErrorType', 'add'. Weights must be

invertible.
'Display'

'on' or 'off'

Display Hankel singular plots (default

'off').

'Order'

Integer, vector
or cell array

Order of reduced model. Use only if not

specified as 2nd argument.

Weights on the original model input and/or output can make the model
reduction algorithm focus on some frequency range of interests. But weights
have to be stable, minimum phase and invertible.

10-260

reduce

This table describes output arguments.

Argument

Description

GRED

LTI reduced order model. Becomes multi-dimensional array

when input is a serial of different model order array.

REDINFO

A STRUCT array with 3 fields:

REDINFO.ErrorBound
REDINFO.StabSV
REDINFO.UnstabSV
For 'hankel' algorithm, STRUCT array becomes:
REDINFO.ErrorBound
REDINFO.StabSV
REDINFO.UnstabSV
REDINFO.Ganticausal
For 'ncf' option, STRUCT array becomes:
REDINFO.GL
REDINFO.GR
REDINFO.hsv

G can be stable or unstable. G and GRED can be either continuous or discrete.

A successful model reduction with a well-conditioned original model G will

ensure that the reduced model GRED satisfies the infinity norm error bound.

Examples

Given a continuous or discrete, stable or unstable system, G, the following

commands can get a set of reduced order models based on your selections:
rand('state',1234); randn('state',5678);G = rss(30,5,4);
[g1, redinfo1] = reduce(G); % display Hankel SV plot
% and prompt for order
[g2, redinfo2] = reduce(G,20); % default to balancmr
[g3, redinfo3] = reduce(G,[10:2:18],'algorithm','schur');
% select schurmr

10-261

reduce

[g4, redinfo] = reduce(G,'ErrorType','mult','MaxError',[0.01,

0.05]);
rand('state',12345); randn('state',6789);
wt1 = rss(6,5,5); wt1.d = eye(5)*2;
wt2 = rss(6,4,4); wt2.d = 2*eye(4);
[g5, redinfo5] = reduce(G, [10:2:18], 'weight',{wt1,wt2});
[g6, redinfo6] =
reduce(G,'ErrorType','add','algorithm','hankel', ...
'maxerror',[0.01]);
for i = 1:6
figure(i); eval(['sigma(G,g' num2str(i) ');']);
end

References

[1] K. Glover, All Optimal Hankel Norm Approximation of Linear

Multivariable Systems, and Their L error Bounds, Int. J. Control, vol. 39,
no. 6, pp. 1145-1193, 1984.
[2] M. G. Safonov and R. Y. Chiang, A Schur Method for Balanced Model
Reduction, IEEE Trans. on Automat. Contr., vol. AC-2, no. 7, July 1989, pp.
729-733.
[3] M. G. Safonov, R. Y. Chiang and D. J. N. Limebeer, Optimal Hankel Model
Reduction for Nonminimal Systems, IEEE Trans. on Automat. Contr., vol. 35,
No. 4, April, 1990, pp. 496-502.
[4] M. G. Safonov and R. Y. Chiang, Model Reduction for Robust Control: A
Schur Relative-Error Method, International Journal of Adaptive Control and
Signal Processing, vol. 2, pp. 259-272, 1988.
[5] K. Zhou, Frequency weighted L error bounds, Syst. Contr. Lett., Vol. 21,
115-125, 1993.

See Also

10-262

balancmr
schurmr
bstmr
ncfmr
hankelmr
hankelsv

Balanced truncation via square-root method

Balanced truncation via Schur method
Balanced stochastic truncation via Schur method
Balanced truncation for normalized coprime factors
Hankel minimum degree approximation
Hankel singular value

repmat

Purpose
Syntax
Description

10repmat

Replicate and tile uncertain matrix

B = repmat(A,M,N)
B = repmat(A,M,N) creates a large matrix B consisting of an M-by-N tiling of

copies of A.
B = repmat(A,[M N]) accomplishes the same result as repmat(A,M,N).
B = repmat(A,[M N P ...]) tiles the array A to produce an M-by-N-by-P-by-...

block array. A can be N-D.

repmat(A,M,N) for scalar A is commonly used to produce an M-by-N matrix filled
with values of A.

Examples

Simple examples of using repmat are

repmat(randumat(2,2),2,3)
repmat(ureal(`A',6),[4 2])

10-263

robopt

Purpose

10robopt

Create options object for use with robuststab and robustperf

Syntax

opts = robopt
opts = robopt('name1',value1,'name2',value2,...)

Description

opts = robopt (with no input arguments) creates an options object with all

the properties set to their default values.

opts = robopt('name1',value1,'name2',value2,...) creates a robopt

object in which specified properties have the given values. Any unspecified
property is set to its default value. It is sufficient to type only enough leading
characters to define the property name uniquely. Case is ignored for property
names.
robopt with no input or output arguments displays a complete list of option

properties and their default values.

Fields
The following are the robopt object properties:
Object Property

Description

Display

Displays progress of computations. Default is

'off'

10-264

Sensitivity

Computes margin sensitivity to individual

uncertainties. Default is 'on'.

VaryUncertainty

Percentage variation of uncertainty used as a

stepsize in finite-difference calculations to
estimate sensitivity. Default is 25.

Mussv

Option used in internal structured singular

value calculations (when calling mussv).
Default is 'sm9'.

robopt

Examples

Object Property

Description

Default

Structure, field names are robopt properties,

and values are the default values.

Meaning

Structure, field names are robopt properties,

and values are the text description of the property.

You can create a robopt options object called opt with all default values.
opt = robopt
Property Object Values:
Display: 'off'
Sensitivity: 'on'
VaryUncertainty: 25
Mussv: 'sm9'
Default: [1x1 struct]
Meaning: [1x1 struct]

An elementary finite-difference scheme is used in estimating local sensitivities.

The property VaryUncertainty denotes the step size used in estimating the
derivatives necessary in computing sensitivities.
In the following statements, you are requesting that the sensitivity of the
robust stability margin calculation to a 50% variation in individual
uncertainties be calculated. The robopt options properties 'Sensitivity' and
'VaryUncertainty' are set individually.
opt = robopt;
opt.VaryUncertainty = 50;
opt
Property Object Values:
Display: 'off'
Sensitivity: 'on'
VaryUncertainty: 50
Mussv: 'sm9'
Default: [1x1 struct]
Meaning: [1x1 struct]

See Also

dkitopt

Creates an options object for dksyn

10-265

robopt

robuststab
robustperf
wcgopt
wcsens
wcmargin

10-266

Calculates stability margins of uncertain systems

Calculates performance margins of uncertain systems
Creates a wcgain options object
Calculates worst-case sensitivities for a feedback loop
Calculates worst-case margins for a feedback loop

robustperf

Purpose

10robustperf

Calculate robust performance margin of uncertain multivariable system

Syntax

perfmarg = robustperf(sys)
[perfmarg,wcu,report,info] = robustperf(sys)
[perfmarg,wcu,report,info] = robustperf(sys,opt)

Description

The performance of a nominally stable uncertain system model will generally

degrade for specific values of its uncertain elements. robustperf, largely
included for historical purposes, computes the robust performance margin,
which is one measure of the level of degradation brought on by the modeled
uncertainty. The relationship between robustperf and other measures, such
as robuststab and wcgain, is described in Chapter 7, Generalized Robustness
Analysis.
As with other uncertain-system analysis tools, only bounds on the performance
margin are computed. The exact robust performance margin is guaranteed to
lie between these upper and lower bounds.
The computation used in robustperf is a frequency domain calculation. If the
input system sys is a ufrd, then the analysis is performed on the frequency
grid within the ufrd. If the input system sys is a uss, then an appropriate
frequency grid is generated (automatically), and the analysis performed on
that frequency grid. In all discussion that follows, N denotes the number of
points in the frequency grid.

Basic Syntax
Suppose sys is a ufrd or uss with M uncertain elements. The results of
[perfmarg,perfmargunc,Report] = robustperf(sys)

are such that perfmarg is a structure with the following fields:

Field

Description

LowerBound

Lower bound on robust performance margin,

positive scalar.

10-267

robustperf

Field

Description

UpperBound

Upper bound on robust performance margin,

positive scalar.

CriticalFrequency

The value of frequency at which the

performance degradation curve crosses the
y=1/x curve. See Generalized Robustness
Analysis in the online documentation.

perfmargunc is a struct of values of uncertain elements associated with the

intersection of the performance degradation curve and the y=1/x curve. See
Chapter 7, Generalized Robustness Analysis. There are M field names, which
are the names of uncertain elements of sys.
Report is a text description of the robust performance analysis results.

Examples

Create a plant with a nominal model of an integrator, and include additive

unmodeled dynamics uncertainty of a level of 0.4 (this corresponds to 100%
model uncertainty at 2.5 rads/s).
P = tf(1,[1 0]) + ultidyn('delta',[1 1],'bound',0.4);

Design a proportional controller K that puts the nominal closed-loop

bandwidth at 0.8 rad/s. Roll off K at a frequency 25 times the nominal
closed-loop bandwidth. Form the closed-loop sensitivity function.
BW = 0.8;
K = tf(BW,[1/(25*BW) 1]);
S = feedback(1,P*K);

Assess the performance margin of the closed-loop sensitivity function. Because

the nominal gain of the sensitivity function is 1, and the performance
degradation curve is monotonically increasing (see Chapter 7, Generalized
Robustness Analysis), the performance margin should be less than 1.
[perfmargin,punc] = robustperf(S);
perfmargin
perfmargin =
UpperBound: 7.4305e-001
LowerBound: 7.4305e-001
CriticalFrequency: 5.3096e+000

10-268

robustperf

You can verify that the upper bound of the performance margin corresponds to
a point on or above the y=1/x curve. First, compute the normalized size of the
value of the uncertain element, and check that this agrees with the upper
bound.
nsize = actual2normalized(S.Uncertainty.delta, punc.delta)
nsize =
perfmargin.UpperBound
ans =
7.4305e-001

Compute the system gain with that value substituted, and verify that the
product of the normalized size and the system gain is greater than or equal to 1.
gain = norm(usubs(S,punc),inf,.00001);
nsize*gain
ans =
1.0000e+000

Finally, as a sanity check, verify that the robust performance margin is less
than the robust stability margin (it should always be, as described in Chapter
4, Robustness Analysis).
[stabmargin] = robuststab(S);
stabmargin
stabmargin =
UpperBound: 3.1251e+000
LowerBound: 3.1251e+000
DestabilizingFrequency: 4.0862e+000

While the robust stability margin is easy to describe (poles migrating from
stable region into unstable region), describing the robust performance margin
is less elementary. See the diagrams and figures in Chapter 7, Generalized
Robustness Analysis. Rather than finding values for uncertain elements that
lead to instability, the analysis finds values of uncertain elements
corresponding to the intersection point of the performance degradation curve
with a y=1/x hyperbola. This characterization, mentioned above in the
description of perfmarg.CriticalFrequency and perfmargunc, is used often in
the descriptions below.

10-269

robustperf
Basic Syntax with Fourth Output Argument
A fourth output argument yields more specialized information, including
sensitivities and frequency-by-frequency information.
[perfmarg,perfmargunc,Report,Info] = robustperf(sys)

In addition to the first 3 output arguments, described previously, Info is a

structure with the following fields:
Field

Description

Sensitivity

A struct with M fields, field names are names of

uncertain elements of sys. Values of fields are
positive and contain the local sensitivity of the
overall Stability Margin to that elements
uncertainty range. For instance, a value of 25
indicates that if the uncertainty range is enlarged
by 8%, then the stability margin should drop by
about 2% (25% of 8). If the Sensitivity property of
the robopt object is 'off', the values are set to
NaN.

Frequency

N-by-1 frequency vector associated with analysis.

BadUncertainValues

N-by-1 cell array, with one entry for each frequency

point. The kth entry Info.BadUncertainValues{k}
is a struct of values of uncertain elements
resulting from a robust performance analysis at
frequency Info.Frequency(k).

MussvBnds

A 1-by-2 frd, with upper and lower bounds from

mussv. The (1,1) entry is the -upper bound
(corresponds to perfmarg.LowerBound) and the
(1,2) entry is the -lower bound (for
perfmarg.UpperBound).

MussvInfo

Structure of compressed data from mussv.

Options (e.g., controlling what is displayed during the computation, turning

on/off the sensitivity computation, setting the step size in the sensitivity

10-270

robustperf

computation, and controlling the option argument used in the underlying call
to mussv) is specified using the robustness analysis options robopt object. For
instance, you can turn the display on and turn off the sensitivity by executing
opt = robopt('Sensitivity','off','Display','on');
[PerfMarg,Destabunc,Report,Info] = robustperf(sys,opt)

Handling Array Dimensions

If sys has array dimensions (for example, suppose that the size of sys is
rcd1d2dF, refer to the d1d2dF as the array dimensions) then the
margin calculation is performed pointwise (individually, at each and every
array value) and the computed answers all have array dimensions as well.
Details are described below. Again, assume that there are N frequency points
and M uncertain elements.
The results of
[perfmarg,perfmargunc,Report,Info] = robustperf(sys,opt)

are perfmarg a structure with the following fields

Field

Description

LowerBound

d1dF, lower bound on stability margin across

the array dimensions.

UpperBound

d1dF, upper bound on performance margin

across the array dimensions. Using single indexing,
for each i, the upper bound on the performance
margin of sys(:,:,i) is
perfmarg.UpperBound(i).

CriticalFrequency

d1dF, the value of frequency at which the

performance degradation curve crosses the y=1/x
curve. Using single indexing, for each i, the
frequency at which the performance degradation
curve crosses the y=1/x curve in robust performance
analysis of sys(:,:,i) is
perfmarg.CriticalFrequency(i). See Chapter 7,
Generalized Robustness Analysis.

10-271

robustperf
perfmargunc is a d1dF structure array of values of uncertain elements,
associated with the intersection of the performance degradation curve and the
y=1/x curve. See Generalized Robustness Analysis in the online
documentation. Using single indexing, for each i, the struct of values of
uncertain elements for uncertain system sys(:,:,i) is perfmargunc(i).
Report is a character array, dimensions 3, 4, ..., F+2 are d1dF, containing
text description of the robustness analysis results at each grid in the array
dimensions.

In addition to the first 3 output arguments, described previously, Info is a

structure with the following fields

10-272

Field

Description

Sensitivity

A d1dF struct, field names are names of

uncertain elements of sys. Using single indexing
notation, Sensitivity(i) contains the
sensitivities of perfmarg.UpperBound(i) for the
uncertain system sys(:,:,i).

Frequency

N-by-1 frequency vector associated with analysis.

BadUncertainValues

N-by-1 cell array, with one entry for each

frequency point. The kth entry
Info.BadUncertainValues{k} is a d1dF
struct of values of uncertain elements resulting
from a d1dF family of robust performance
computations at frequency Info.Frequency(k).

MussvBnds

12d1dF frd, with upper and lower bounds

from mussv. Using single indexing for the
dimensions associated with the array dimensions,
it follows that the (1,1,i) entry is the -upper
bound (reciprocal of perfmarg.UpperBound(i))
while the (1,2,i) entry is the -lower bound
(reciprocal of perfmarg.UpperBound(i)).

MussvInfo

Structure of compressed data from mussv.

robustperf

The smallest performance margin over all array dimensions is computed

min(perfmarg.UpperBound(:)). Computing
i = find(UpperBound==min(UpperBound(:)))

and then selecting perfmargunc(i) yields values for an uncertainty

corresponding to the smallest performance margin across all array dimensions.

Algorithm

A rigorous robust performance analysis consists of two steps:

1 Verify that the nominal system is stable;
2 Robust performance analysis on an augmented system.

The algorithm in robustperf follows this in spirit, but might require user
attention.
If sys is a uss object, then the first requirement of stability of nominal value is
explicitly checked within robustperf. However, if sys is an ufrd, then the
verification of nominal stability from the nominal frequency response data is
not performed, and is instead assumed.
The exact performance margin is guaranteed to be no larger than UpperBound
(some uncertain elements associated with this magnitude cause instability
one instance is returned in the structure perfmargunc). The instability created
by perfmargunc occurs at the frequency value in CriticalFrequency.
Similarly, the exact performance margin is guaranteed to be no smaller than
LowerBound.

Limitations

Because the calculation is carried out with a frequency gridding, it is possible

(likely) that the true critical frequency is missing from the frequency vector
used in the analysis. This is similar to the problem in robuststab. However, in
comparing to robuststab, the problem in robustperf is less acute. The robust
performance margin, considered a function of problem data and frequency, is
typically a continuous function (unlike the robust stability margin, described
in the Robust Control Toolbox demo called Getting Reliable Estimates of
Robustness Margins in the online documentation). Hence, in robust
performance margin calculations, increasing the density of the frequency grid
will always increase the accuracy of the answers, and in the limit, answers
arbitrarily close to the actual answers are obtainable with finite frequency
grids.

10-273

robustperf

See Also

10-274

loopmargin
mussv
norm
robopt
robuststab
actual2normalized
wcgain
wcsens
wcmargin

Comprehensive analysis of feedback loop.

Calculate bounds on the Structured Singular Value ()
Calculate LTI system norms
Create a robuststab/robustperf options object
Calculates stability margins of uncertain systems
Normalizes range of uncertain atoms
Calculate worst-case gain of uncertain systems
Calculate worst-case sensitivities for feedback loop
Calculate worst-case margins for feedback loop

robuststab

Purpose

10robuststab

Calculate robust stability margin of uncertain multivariable system

Syntax

[stabmarg,destabunc,report,info] = robuststab(sys)
[stabmarg,destabunc,report,info] = robuststab(sys,opt)

Description

A nominally stable uncertain system is generally unstable for specific values of

its uncertain elements. Determining the values of the uncertain elements
closest to their nominal values for which instability occurs is a robust stability
calculation.
If the uncertain system is stable for all values of uncertain elements within
their allowable ranges (ranges for ureal, norm bound or positive-real
constraint for ultidyn, radius for ucomplex, weighted ball for ucomplexm), the
uncertain system is robustly stable. Conversely, if there is a combination of
element values that cause instability, and all lie within their allowable ranges,
then the uncertain system is not robustly stable.
robuststab computes the margin of stability robustness for an uncertain

system. A stability robustness margin greater than 1 means that the uncertain
system is stable for all values of its modeled uncertainty. A stability robustness
margin less than 1 implies that certain allowable values of the uncertain
elements, within their specified ranges, lead to instability.
Numerically, a margin of 0.5 (for example) implies two things: the uncertain
system remains stable for all values of uncertain elements that are less than
0.5 normalized units away from their nominal values and, there is a collection
of uncertain elements that are less than or equal to 0.5 normalized units away
from their nominal values that results in instability. Similarly, a margin of 1.3
implies that the uncertain system remains stable for all values of uncertain
elements up to 30% outside their modeled uncertain ranges. See
actual2normalized for converting between actual and normalized deviations
from the nominal value of an uncertain element.
As with other uncertain-system analysis tools, only bounds on the exact
stability margin are computed. The exact robust stability margin is guaranteed
to lie in between these upper and lower bounds.
The computation used in robuststab is a frequency-domain calculation. If the
input system sys is a ufrd, then the analysis is performed on the frequency
grid within the ufrd. If the input system sys is a uss, then an appropriate

10-275

robuststab

frequency grid is generated (automatically), and the analysis performed on

that frequency grid. In all discussion that follows, N denotes the number of
points in the frequency grid.

Basic Syntax
Suppose sys is a ufrd or uss with M uncertain elements. The results of
[stabmarg,destabunc,Report] = robuststab(sys)

are that stabmarg is a structure with the following fields:

Field

Description

LowerBound

Lower bound on stability margin, positive

scalar. If greater than 1, then the uncertain
system is guaranteed stable for all values of
the modeled uncertainty. If the nominal
value of the uncertain system is unstable,
then stabmarg.UpperBound and
stabmarg.LowerBound will equal .

UpperBound

Upper bound on stability margin, positive

scalar. If less than 1, the uncertain system is
not stable for all values of the modeled
uncertainty.

DestabilizingFrequency

The critical value of frequency at which

instability occurs, with uncertain elements
closest to their nominal values. At a
particular value of uncertain elements (see
destabunc below), the poles migrate across
the stability boundary (imaginary-axis in
continuous-time systems, unit-disk in
discrete-time systems) at the frequency given
by DestabilizingFrequency.

destabunc is a structure of values of uncertain elements, closest to nominal,

that cause instability. There are M field names, which are the names of
uncertain elements of sys. The value of each field is the corresponding value of

10-276

robuststab

the uncertain element, such that when jointly combined, lead to instability.
The command pole(usubs(sys,destabunc)) shows the instability. If A is an
uncertain atom of sys, then
actual2normalized(destabunc.A,sys.Uncertainty.A)

will be less than or equal to UpperBound, and for at least one uncertain element
of sys, this normalized distance will be equal to UpperBound, proving that
UpperBound is indeed an upper bound on the robust stability margin.
Report is a text description of the robustness analysis results.

Examples

Construct a feedback loop with a second-order plant and a PID controller with
approximate differentiation. The second-order plant has frequency-dependent
uncertainty, in the form of additive unmodeled dynamics, introduced with an
ultidyn object and a shaping filter.
robuststab is used to compute the stability margins of the closed-loop system

with respect to the plant model uncertainty.

P = tf(4,[1 .8 4]);
delta = ultidyn('delta',[1 1],'SampleStateDim',5);
Pu = P + 0.25*tf([1],[.15 1])*delta;
C = tf([1 1],[.1 1]) + tf(2,[1 0]);
S = feedback(1,Pu*C);
[stabmarg,destabunc,report,info] = robuststab(S);

You can view the stabmarg variable.

stabmarg
stabmarg =
UpperBound: 0.8181
LowerBound: 0.8181
DestabilizingFrequency: 9.1321

As the margin is less than 1, the closed-loop system is not stable for plant
models covered by the uncertain model Pu. There is a specific plant within the
uncertain behavior modeled by Pu (actually about 82% of the modeled
uncertainty) that leads to closed-loop instability, with the poles migrating
across the stability boundary at 9.1 rads/s.
The report variable is specific, giving a plain-language version of the
conclusion.

10-277

robuststab

report
report =
Uncertain System is NOT robustly stable to modeled uncertainty.
-- It can tolerate up to 81.8% of modeled uncertainty.
-- A destabilizing combination of 81.8% the modeled uncertainty
exists, causing an instability at 9.13 rad/s.
-- Sensitivity with respect to uncertain element ...
'delta' is 100%. Increasing 'delta' by 25% leads to a 25%
decrease in the margin.

Because the problem has only one uncertain element, the stability margin is
completely determined by this element, and hence the margin exhibits 100%
sensitivity to this uncertain element.
You can verify that the destabilizing value of delta is indeed about 0.82
normalized units from its nominal value.
actual2normalized(S.Uncertainty.delta,destabunc.delta)
ans =
0.8181

Use usubs to substitute the specific value into the closed-loop system. Verify
that there is a closed-loop pole near j9.1, and plot the unit-step response of the
nominal closed-loop system, as well as the unstable closed-loop system.
Sbad = usubs(S,destabunc);
pole(Sbad)
ans =
1.0e+002 *
-3.2318
-0.2539
-0.0000 + 0.0913i
-0.0000 - 0.0913i
-0.0203 + 0.0211i
-0.0203 - 0.0211i
-0.0106 + 0.0116i
-0.0106 - 0.0116i
step(S.NominalValue,'r--',Sbad,'g',4);

Finally, as an ad-hoc test, set the gain bound on the uncertain delta to 0.81
(slightly less than the stability margin). Sample the closed-loop system at 100
values, and compute the poles of all these systems.

10-278

robuststab

S.Uncertainty.delta.Bound = 0.81;
S100 = usample(S,100);
p100 = pole(S100);
max(real(p100(:)))
ans =
-6.4647e-007

As expected, all poles have negative real parts.

Basic Syntax with Fourth Output Argument

A fourth output argument yields more specialized information, including
sensitivities and frequency-by-frequency information.
[StabMarg,Destabunc,Report,Info] = robuststab(sys)

In addition to the first 3 output arguments, described previously, Info is a

structure with the following fields:
Field

Description

Sensitivity

A struct with M fields, Field names are names of

uncertain elements of sys. Values of fields are
positive, each the local sensitivity of the overall
stability margin to that elements uncertainty
range. For instance, a value of 25 indicates that if
the uncertainty range is enlarged by 8%, then the
stability margin should drop by about 2% (25% of
8). If the Sensitivity property of the robopt object
is 'off', the values are set to NaN.

Frequency

N-by-1 frequency vector associated with analysis.

10-279

robuststab

Field

Description

BadUncertainValues

N-by-1 cell array, with one entry for each frequency

point. The kth entryInfo.BadUncertainValues{k}
is a struct of values of uncertain elements, closest
to their nominal values, which cause the system
poles to migrate across the stability boundary at
frequency Info.Frequency(k). The command
pole(usubs(sys,Info.BadUncertainValues{k}))

shows the migration. The command

usubs(sys,cat(1,Info.BadUncertainValues{:})
) generates an N-by-1 ss array. Each instance is

unstable, with poles on the stability boundary at

frequencies given by the vector Info.Frequency.
This migration to instability has been achieved
with the smallest normalized deviations in the
uncertain elements from their nominal values.
MussvBnds

A 1-by-2 frd, with upper and lower bounds from

mussv. The (1,1) entry is the -upper bound
(corresponds to stabmarg.LowerBound) and the
(1,2) entry is the -lower bound (for
stabmarg.UpperBound).

MussvInfo

Structure of compressed data from mussv.

Options (e.g., controlling what is displayed during the computation, turning

on/off the sensitivity computation, setting the step-size in the sensitivity
computation, and controlling the option argument used in the underlying call
to mussv) can be specified using the robustness analysis options robopt object.
For instance, you can turn the display on, and the sensitivity calculation off by
executing
opt = robopt('Sensitivity','off','Display','on');
[StabMarg,Destabunc,Report,Info] = robuststab(sys,opt)

Handling Array Dimensions

If sys has array dimensions (for example, suppose that the size of sys is
rcd1d2dF, refer to d1d2dF as the array dimensions) then the

10-280

robuststab

margin calculation is performed pointwise (individually, at each and every

array value) and the computed answers all have array dimensions as well.
Details are described below. Again, assume that there are N frequency points
and M uncertain elements.
The results of
[stabmarg,destabunc,Report,Info] = robuststab(sys,opt)

are stabmarg is a structure with the following fields:

Field

Description

LowerBound

d1dF, lower bound on stability margin

across the array dimensions.

UpperBound

d1dF, upper bound on stability margin

across the array dimensions. Using
single-indexing, for each i, the upper bound
on the stability margin of sys(:,:,i) is
stabmarg.UpperBound(i).

DestabilizingFrequency

d1dF, frequency at which instability

occurs, associated with
stabmarg.UpperBound. Using
single-indexing, for each i, the frequency at
which instability occurs in robust stability
analysis of sys(:,:,i) is
stabmarg.DestabilizingFrequency(i).

destabunc is a d1dF structure array of values of uncertain elements, that

cause instability. Using single-indexing, for each i, the destabilizing values of
uncertain elements for uncertain system sys(:,:,i) is destabunc(i).
Report is a character array, dimensions 3, 4, ... , F+2 are d1dF, containing
text description of the robustness analysis results at each grid in the array
dimensions.

10-281

robuststab

In addition to the first 3 output arguments, described previously, Info is a

structure with the following fields:
Field

Description

Sensitivity

A d1dF struct, field names are names of

uncertain elements of sys. Using single indexing
notation, Sensitivity(i) contains the sensitivities
of stabmarg.UpperBound(i) for the uncertain
system sys(:,:,i).

Frequency

N1 frequency vector associated with analysis.

BadUncertainValues

N-by-1 cell array, with one entry for each frequency

point. The kth entry Info.BadUncertainValues{k}
is a d1dF struct of values of uncertain
elements, closest to their nominal values, that
cause the system poles to migrate across the
stability boundary at frequency
Info.Frequency(k). The command
usubs(sys,Info.BadUncertainValues{k})

produces an ss array of size d1dF with the

substitutions made. Alternatively,
usubs(sys,cat(F+1,Info.BadUncertainValues{:
})) produces an ss array of size d1dFN with

the substitutions made.

MussvBnds

A 12d1df frd, with upper and lower bounds

from mussv. Using single-indexing for the
dimensions associated with the array dimensions,
it follows that the (1,1,i) entry is the -upper
bound (corresponding to stabmarg.LowerBound(i))
while the (1,2,i) entry is the -lower bound
(which corresponds to stabmarg.UpperBound(i)).

MussvInfo

Structure of compressed data from mussv.

You can compute the smallest stability margin over all array dimensions via
min(stabmarg.UpperBound(:)).

10-282

robuststab

Computing i = find(UpperBound==min(UpperBound(:))) and then

destabunc(i) yields values for an uncertainty corresponding to the smallest
stability margin across all array dimensions.

Algorithm

A rigorous robust stability analysis consists of two steps:

1 Verify that the nominal system is stable.
2 Verify that no poles cross the stability boundary as the uncertain elements

vary within their ranges.

Because the stability boundary is also associated with the frequency response,
the second step can be interpreted (and carried out) as a frequency domain
calculation. This amounts to a classical -analysis problem.
The algorithm in robuststab follows this in spirit, but might require your
attention.
If sys is a uss object, then the first requirement of stability of nominal value is
explicitly checked within robuststab. However, if sys is an ufrd, then the
verification of nominal stability from the nominal frequency response data is
not performed, and is instead assumed.
In the second step (monitoring the stability boundary for the migration of
poles), rather than check all points on stability boundary, the algorithm only
detects migration of poles across the stability boundary at the frequencies in
info.Frequency.
See Limitations on page 10-283 about issues related to migration detection.
The exact stability margin is guaranteed to be no larger than UpperBound
(some uncertain elements associated with this magnitude cause instability
one instance is returned in the structure destabunc). The instability created by
destabunc occurs at the frequency value in DestabilizingFrequency.
Similarly, the exact stability margin is guaranteed to be no smaller than
LowerBound. In other words, for all modeled uncertainty with magnitude up to
LowerBound, the system is guaranteed stable. These bounds are derived using
the upper bound for the structured singular value, which is essentially
optimally scaled, small-gain theorem analysis.

Limitations

Under most conditions, the robust stability margin that occurs at each
frequency is a continuous function of the problem data at that frequency.

10-283

robuststab

Because the problem data, in turn, is a continuous function of frequency, it

follows that finite frequency grids are usually adequate in correctly assessing
robust stability bounds, assuming the frequency grid is dense enough.
Nevertheless, there are simple examples that violate this. In some problems,
the migration of poles from stable to unstable only occurs at a finite collection
of specific frequencies (generally unknown to you). Any frequency grid that
excludes these critical frequencies (and almost every grid will exclude them)
will result in undetected migration and misleading results, namely stability
margins of .
See the Robust Control Toolbox demo Getting Reliable Estimates of
Robustness Margins in the online documentation about circumventing the
problem in an engineering-relevant fashion.

See Also

loopmargin
mussv
robopt
robustperf
wcgain
wcsens
wcmargin

10-284

Comprehensive analysis of a feedback loop

Calculates bounds on the Structured Singular Value
()
Creates a robuststab/robustperf options object
Calculates performance margins of uncertain systems
Calculates worst-case gain of uncertain systems
Calculates worst-case sensitivities for a feedback loop
Calculates worst-case margins for a feedback loop

schurmr

Purpose

10schurmr

Balanced truncation via Schur method

Syntax

GRED = schurmr(G)
GRED = schurmr(G,order)
[GRED,redinfo] = schurmr(G,key1,value1,...)
[GRED,redinfo] = schurmr(G,order,key1,value1,...)

Description

schurmr returns a reduced order model GRED of G and a struct array redinfo

containing the error bound of the reduced model and Hankel singular values of
the original system.
The error bound is computed based on Hankel singular values of G. For a stable
system Hankel singular values indicate the respective state energy of the
system. Hence, reduced order can be directly determined by examining the
system Hankel SVs, .
With only one input argument G, the function will show a Hankel singular
value plot of the original model and prompt for model order number to reduce.
This method guarantees an error bound on the infinity norm of the additive
error || G-GRED || for well-conditioned model reduced problems [1]:
n

G Gred 2

i
k+1

This table describes input arguments for schurmr.

Argument

Description

LTI model to be reduced (without any other inputs will plot

its Hankel singular values and prompt for reduced order).

ORDER

(Optional) an integer for the desired order of the reduced

model, or optionally a vector packed with desired orders for
batch runs.

A batch run of a serial of different reduced order models can be generated by

specifying order = x:y, or a vector of integers. By default, all the anti-stable

10-285

schurmr

part of a system is kept, because from control stability point of view, getting rid
of unstable state(s) is dangerous to model a system.
'MaxError' can be specified in the same fashion as an alternative for 'ORDER'.
In this case, reduced order will be determined when the sum of the tails of the
Hankel SVs reaches the 'MaxError'.
Argument

Value

Description

'MaxError'

A real
number or a
vector of
different
errors

Reduce to achieve H error.

When present, 'MaxError'overides ORDER
input.

{Wout,Win}

cell array

Optimal 1x2 cell array of LTI weights Wout

(output) and Win (input); default is both
identity; Weights must be invertible.

'Display'

'on' or
'off'

Display Hankel singular plots (default

'off').

'Order'

Integer,
vector or
cell array

Order of reduced model. Use only if not

specified as 2nd argument.

'Weights'

Weights on the original model input and/or output can make the model
reduction algorithm focus on some frequency range of interests. But weights
have to be stable, minimum phase and invertible.

10-286

schurmr

This table describes output arguments.

Argument

Description

GRED

LTI reduced order model. Becomes multi-dimensional array

when input is a serial of different model order array.

REDINFO

A STRUCT array with 3 fields:

REDINFO.ErrorBound
REDINFO.StabSV
REDINFO.UnstabSV

G can be stable or unstable. G and GRED can be either continuous or discrete.

Given a state space (A,B,C,D) of a system and k, the desired reduced order, the
following steps will produce a similarity transformation to truncate the
original state-space system to the kth order reduced model [16].
1 Find the controllability and observability grammians P and Q.
2 Find the Schur decomposition for PQ in both ascending and descending

order, respectively,
T
V A PQV A

T
V D PQV D

1
=

0
0 0 n
n

0
0 0 1

3 Find the left/right orthonormal eigen-bases of PQ associated with the kth big

Hankel singular values.

Algorithm

V A = [V R, SMALL, V L, BIG ]
4 Find the SVD of (VTL,BIG VR,BIG) = U VT

10-287

schurmr

V D = [ V R, BIG ,V L, SMALL]
5 Form the left/right transformation for the final kth order reduced model

SL,BIG = V L,BIG U(1:k,1:k)-1/2

SR,BIG = VR,BIGV(1:k,1:k)-1/2
6 Finally,

S T L, BIG AS R, BIG S L, BIG B

CS R, BIG

The proof of the Schur balance truncation algorithm can be found in [2].

Examples

Given a continuous or discrete, stable or unstable system, G, the following

commands can get a set of reduced order models based on your selections:
rand('state',1234); randn('state',5678);G = rss(30,5,4);
[g1, redinfo1] = schurmr(G); % display Hankel SV plot
% and prompt for order (try 15:20)
[g2, redinfo2] = schurmr(G,20);
[g3, redinfo3] = schurmr(G,[10:2:18]);
[g4, redinfo4] = schurmr(G,'MaxError',[0.01, 0.05]);
rand('state',12345); randn('state',6789);
wt1 = rss(6,5,5); wt1.d = eye(5)*2;
wt2 = rss(6,4,4); wt2.d = 2*eye(4);
[g5, redinfo5] = schurmr(G, [10:2:18], 'weight',{wt1,wt2});
for i = 1:5
figure(i); eval(['sigma(G,g' num2str(i) ');']);
end

References

[1] K. Glover, All Optimal Hankel Norm Approximation of Linear

Multivariable Systems, and Their L- error Bounds, Int. J. Control, vol. 39,
no. 6, pp. 1145-1193, 1984.
[2] M. G. Safonov and R. Y. Chiang, A Schur Method for Balanced Model
Reduction, IEEE Trans. on Automat. Contr., vol. 34, no. 7, July 1989, pp.
729-733.

See Also

10-288

reduce

Top level model reduction routines

schurmr

balancmr
bstmr
ncfmr
hankelmr
hankelsv

Balanced truncation via square-root method

Balanced stochastic truncation via Schur method
Balanced truncation for normalized coprime factors
Hankel minimum degree approximation
Hankel singular value

10-289

sdhinfnorm

Purpose

10sdhinfnorm

Compute L2 norm of sampled-data system

Syntax

[gaml,gamu] = sdhinfnorm(sdsys,k)
[gaml,gamu] = sdhinfnorm(sdsys,k,delay)
[gaml,gamu] = sdhinfnorm(sdsys,k,delay,tol)

Description

[gaml,gamu] = sdhinfnorm(sdsys,k) computes the L2 induced norm of a

continuous-time LTI plant, sdsys, in feedback with a discrete-time controller,
k, connected through an ideal sampler and a zero-order hold (see figure below).
sdsys must be strictly proper, such that the constant feedback gain must be
zero. The outputs, gamu and gaml, are upper and lower bounds on the induced
L2 norm of the sampled-data closed-loop system.

 

  

[gaml,gamu] = sdhinfnorm(sdsys,k,h,delay) includes the input argument

delay. delay is a nonnegative integer associated with the number of

computational delays of the controller. The default value of the delay is 0.

[gaml,gamu] = sdhinfnorm(sdsys,k,h,delay,tol) includes the input
argument, tol, which defines the difference between upper and lower bounds
when search terminates. The default value of tol is 0.001.

Examples

Consider an open-loop, continuous-time transfer function p = 30/s(s+30) and

a continuous-time controller k = 4/(s+4). The closed-loop continuous-time
system has a peak magnitude across frequency of 1.
p = ss(tf(30,[1 30])*tf([1],[1 0]));
k = ss(tf(4,[1 4]));
cl = feedback(p,k);
norm(cl,'inf')
ans =
1

10-290

sdhinfnorm

Initially the controller is to be implemented at a sample rate of 1.5 Hz. The

sample-data norm of the closed-loop system with the discrete-time controller is
1.0.
kd = c2d(k,0.75,'zoh');
[gu,gl] = sdhinfnorm([1; 1]*p*[1 1],-kd);
[gu gl]
ans =
3.7908
3.7929

Because of the large difference in norm between the continuous-time and

sampled-data closed-loop system, the sample rate of the controller is increased
from 1.5 Hz to 5 Hz. The sample-data norm of the new closed-loop system is
3.79.
kd = c2d(k,0.2,'zoh');
[gu,gl] = sdhinfnorm([1; 1]*p*[1 1],-kd);
[gu gl]
ans =
1.0044
1.0049

Algorithm

sdhinfnorm uses variations of the formulas described in the Bamieh and

Pearson paper to obtain an equivalent discrete-time system. (These variations

are done to improve the numerical conditioning of the algorithms.) A
preliminary step is to determine whether the norm of the continuous-time
system over one sampling period without control is less than the given value.
This requires a search and is, computationally, a relatively expensive step.

References

Bamieh, B.A., and J.B. Pearson, A General Framework for Linear Periodic
Systems with Applications to Sampled-Data Control, IEEE Transactions on
Automatic Control, Vol. AC37, 1992, pp. 418-435.

See Also

gapmetric
hinfsyn
norm
sdhinfsyn
sdlsim

Computes the gap and the Vinnicombe gap metric

Synthesizes a H optimal controller
Calculates the system norm of an LTI object
Synthesizes a sample-data H optimal controller
Simulates response of a sampled-data feedback system

10-291

sdhinfsyn

Purpose

10sdhinfsyn

Sampled-data H controller synthesis

Syntax

[K,GAM]=sdhinfsyn(P,NMEAS,NCON)
[K,GAM]=sdhinfsyn(P,NMEAS,NCON, KEY1,VALUE1,KEY2,VALUE2,...)

Description

sdhinfsyn controls a continuous-time LTI system P with a discrete-time

controller K. The continuous-time LTI plant P has a state-space realization

partitioned as follows:
A B1 B2
P = C1 0 0
C2 0

where the continuous-time disturbance inputs enter through B1, the outputs
from the controller are held constant between sampling instants and enter
through B2, the continuous-time errors (to be kept small) correspond to the C1
partition, and the output measurements that are sampled by the controller
correspond to the C2 partition. B2 has column size ncon and C2 has row size
nmeas. Note that the D matrix must be zero.
sdhinfsyn synthesizes a discrete-time LTI controller K to achieve a given norm
(if possible) or find the minimum possible norm to within tolerance TOLGAM.

y1

u1
P

y2

u2

sampler

delay
Ts

hold

Similar to hinfsyn, the function sdhinfsyn employs a iteration. Given a high

and low value of , GMAX and GMIN, the bisection method is used to iterate on the
value of in an effort to approach the optimal H control design. If GMAX = GMIN,
only one value is tested. The stopping criterion for the bisection algorithm

10-292

sdhinfsyn

requires that the relative difference between the last value that failed and the
last value that passed be less than TOLGAM.
Input arguments:
P

LTI plant

NMEAS

Number of measurements output to controller

NCON

Number of control inputs

Optional input arguments (KEY, VALUE) pairs are similar to hinfsyn, but with
additional KEY values 'Ts' and 'DELAY'.
KEY

VALUE

Meaning

'GMAX'

real

Initial upper bound on GAM (default=Inf)

'GMIN'

real

Initial lower bound on GAM (default=0)

'TOLGAM'

real

Relative error tolerance for GAM (default=.01)

'Ts'

real

(Default=1) sampling period of the controller to

be designed

'DELAY'

integer

(Default=0) a nonnegative integer giving the

number of sample periods delay for the control
computation

'DISPLAY'

'off'
'on'

(Default) no command window display, or the

command window displays synthesis progress
information

Output arguments:

Algorithm

H controller

GAM

Final value of H cost achieved

sdhinfsyn uses a variation of the formulas described in the Bamieh and

Pearson paper [1] to obtain an equivalent discrete-time system. (This is done

to improve the numerical conditioning of the algorithms.) A preliminary step is

10-293

sdhinfsyn

to determine whether the norm of the continuous-time system over one

sampling period without control is less than the given -value. This requires a
search and is computationally a relatively expensive step.

References

[1] Bamieh, B.A., and J.B. Pearson, A General Framework for Linear Periodic
Systems with Applications to Sampled-Data Control, IEEE Transactions on
Automatic Control, Vol. AC37, 1992, pp. 418-435.

See Also

norm
hinfsyn
sdhinfnorm

10-294

System norm of an LTI object

Synthesize an H optimal controller
Calculate norm of sampled-data feedback system

sdlsim

Purpose

10sdlsim

Time response of sampled-data feedback system

Syntax

sdlsim(p,k,w,t,tf)
sdlsim(p,k,w,t,tf,x0,z0)
sdlsim(p,k,w,t,tf,x0,z0,int)
[vt,yt,ut,t] = sdlsim(p,k,w,t,tf)
[vt,yt,ut,t] = sdlsim(p,k,w,t,tf,x0,z0,int)

Description

sdlsim(p,k,w,t,tf) plots the time response of the hybrid feedback system.

lft(p,k), is forced by the continuous input signal described by w and t (values
and times, as in lsim). p must be a continuous-time LTI system, and k must be

discrete-time LTI system with a specified sampling time (the unspecified

sampling time -1 is not allowed). The final time is specified with tf.
sdlsim(p,k,w,t,tf,x0,z0) specifies the initial state vector x0 of p, and z0 of
k, at time t(1).
sdlsim(p,k,w,t,tf,x0,z0,int) specifies the continuous-time integration
step size int. sdlsim forces int = (k.Ts)/N int where N>4 is an integer. If

any of these optional arguments is omitted, or passed as empty matrices, then

default values are used. The default value for x0 and z0 is zero. Nonzero initial
conditions are allowed for p (and/or k) only if p (and/or k) is an ss object.
If p and/or k is an LTI array with consistent array dimensions, then the time
simulation is performed pointwise across the array dimensions.
[vt,yt,ut,t] = sdlsim(p,k,w,t,tf) computes the continuous-time
response of the hybrid feedback system lft(p,k) forced by the continuous
input signal defined by w and t (values and times, as in lsim). p must be a
continuous-time system, and k must be discrete-time, with a specified
sampling time (the unspecified sampling time -1 is not allowed). The final time
is specified with tf. The outputs vt, yt and ut are 2-by-1 cell arrays: in each
the first entry is a time vector, and the second entry is the signal values. Stored
in this manner, the signal vt is plotted by using one of the following commands:
plot(vt{1},vt{2})
plot(vt{:})

Signals yt and ut are respectively the input to k and output of k.

10-295

sdlsim

If p and/or k are LTI arrays with consistent array dimensions, then the time
simulation is performed pointwise across the array dimensions. The outputs
are 2-by-1-by-array dimension cell arrays. All responses can be plotted
simultaneously, for example, plot(vt).
[vt,yt,ut,t] = sdlsim(p,k,w,t,tf,x0,z0,int) The optional arguments
are int (integration step size), x0 (initial condition for p), and z0 (initial
condition for k). sdlsim forces int = (k.Ts)/N, where N>4 is an integer. If any

of these arguments is omitted, or passed as empty matrices, then default

values are used. The default value for x0 and z0 is zero. Nonzero initial
conditions are allowed for p (and/or k) only if p (and/or k) is an ss object.

Examples

To illustrate the use of sdlsim, consider the application of a discrete controller

to an integrator with near integrator. A continuous plant and a discrete
controller are created. A sample and hold equivalent of the plant is formed and
the discrete closed-loop system is calculated. Simulating this with lsim gives
the system response at the sample points. sdlsim is then used to calculate the
intersample behavior.
P = tf(1,[1, 1e-5,0]);
T = 1.0/20;
C = ss([-1.5 T/4; -2/T -.5],[ .5 2;1/T 1/T],...
[-1/T^2 -1.5/T], [1/T^2 0],T);
Pd = c2d(P,T,'zoh');

The closed-loop digital system is now set up. You can use sysic to construct the
interconnected feedback system.
systemnames = 'Pd C';
inputvar = '[ref]';
outputvar = '[Pd]';
input_to_Pd = '[C]';
input_to_C = '[ref ; Pd]';
sysoutname = 'dclp';
cleanupsysic = 'yes';
sysic;
lsim is used to simulate the digital step response.
[yd,td] = step(dclp,20*T);

10-296

sdlsim

The continuous interconnection is set up and the sampled data response is

calculated with sdlsim.
M = [0,1;1,0;0,1]*blkdiag(1,P);
t = [0:.01:1]';
u = ones(size(t));
y1 = sdlsim(M,C,u,t);
plot(td,yd,'r*',y1{:},'b-')
axis([0,1,0,1.5])
xlabel('Time: seconds')
title('Step response: discrete (*), &continuous')

Step response: discrete (*), &continuous

1.5

0.5

0.1

0.2

0.3

0.4

0.5
0.6
Time: seconds

0.7

0.8

0.9

You can see the effect of a nonzero initial condition in the continuous-time
system. Note how examining the system at only the sample points will
underestimate the amplitude of the overshoot.
y2 = sdlsim(M,C,u,t,1,0,[0.25;0]);
plot(td,yd,'r*',y1{:},'b-',y2{:},'g--')

10-297

sdlsim

axis([0,1,0,1.5])
xlabel('Time: seconds')
title('Step response: non zero initial condition')

Step response: non zero initial condition

1.5

0.5

0.1

0.2

0.3

0.4

0.5
0.6
Time: seconds

0.7

0.8

0.9

Finally, you can examine the effect of a sinusoidal disturbance at the

continuous-time plant output. This controller is not designed to reject such a
disturbance and the system does not contain antialiasing filters. Simulating
the effect of antialiasing filters is easily accomplished by including them in the
continuous interconnection structure.
M2 = [0,1,1;1,0,0;0,1,1]*blkdiag(1,1,P);
t = [0:.001:1]';
dist = 0.1*sin(41*t);
u = ones(size(t));
[y3,meas,act] = sdlsim(M2,C,[u dist],t,1);
plot(y3{:},'-',t,dist,'b--',t,u,'g-.')
xlabel('Time: seconds')
title('Step response: disturbance (dashed) &

10-298

output (solid)')

sdlsim

Step response: disturbance (dashed) & output (solid)

1.4

1.2

0.8

0.6

0.4

0.2

0.2

0.1

0.2

0.3

0.4

0.5
0.6
Time: seconds

0.7

0.8

0.9

Algorithm

sdlsim oversamples the continuous-time, N times the sample rate of the

controller k.

See Also

gapmetric
hinfsyn
norm
sdhinfnorm
sdhinfsyn

Computes the gap and the Vinnicombe gap metric

Synthesizes a H optimal controller
Computes the system norm of an LTI object
Calculates the norm of a sampled-data feedback
system
Synthesizes a sample-data H optimal controller

10-299

sectf

Purpose

10sectf

State-space sector bilinear transformation

Syntax

[G,T] = sectf(F,SECF,SECG)

Description

[G,T] = sectf(F,SECF,SECG) computes a linear fractional transform T such

that the system lft(F,K) is in sector SECF if and only if the system lft(G,K)
is in sector SECG where
G=lft(T,F,NU,NY)

where NU and NY are the dimensions of uT2 and yT2, respectivelysee

Figure 10-17.

G
yG1

yT1
yT2

yF1
yG2

uT1
T

uT2

uF1
uF2

yF2

K
Figure 10-17: Sector transform G=lft(T,F,NU,NY)

10-300

uG1

uG2

sectf

sectf are used to transform general conic-sector control system performance

specifications into equivalent H-norm performance specifications.

Input arguments
F

LTI state-space plant

SECG, SECF:

Conic Sector:

[-1,1] or [-1;1]
[0,Inf] or
[0;Inf]

0 Re [ y u ]

[A,B] or [A;B]

0 Re [ ( y Au ) ( y Bu ) ]

[a,b] or [a;b]

0 Re [ ( y diag ( a )u ) ( y diag ( b )u ) ]

0 Re [ ( S 11 u + S 12 y ) ( S 21 u + S 22 y ) ]

0 Re [ ( S 11 u + S 12 y ) ( S 21 u + S 22 y ) ]

where A,B are scalars in [, ] or square matrices; a,b are vectors; S=[S11
S12;S21,S22] is a square matrix whose blocks S11,S12,S21,S22 are either
scalars or square matrices; S is a two-port system S=mksys(a,b1,b2, ,'tss') with
transfer function
S(s) =

S 11 ( s ) S 12 ( s )
S 21 ( s ) S 22 ( s )

Output arguments

Description

Transformed plant G(s)=lftf(T,F)

LFT sector transform, maps conic sector

SECF into conic sector SECG

10-301

sectf

Output variables are:

Examples

The transformed plant G(s)=lftf(T,F):

The linear fractional transformation T(s)=T

The statement G(j) inside sector[1, 1] is equivalent to the H inequality

sup ( G ( j ) ) = G 1

Given a two-port open-loop plant P(s) := P, the command

P1 = sectf(P,[0,Inf],[-1,1]) computes a transformed P1(s):= P1 such that
if lft(G,K) is inside sector[1, 1] if and only if lft(F,K) is inside sector[0, ].
In other words, norm(lft(G,K),inf)<1 if and only if lft(F,K) is strictly
positive real. See Figure 10-18.
u1
u2

P(s)

y1
y2

K(s)
Figure 10-18: Sector Transform Block Diagram

Here is a simple example of the sector transform.

1 - sec tor [ 1, 1 ] P ( s ) = s----------+ 2- sec tor [ 0 , ]
P ( s ) = ----------1
s+1
s
You can compute this by simply executing the following commands:
P = ss(tf(1,[1 1]));
P1 = sectf(P,[-1,1],[0,Inf]);

The Nyquist plots for this transformation are depicted in Figure 10-19. The
condition P1(s) inside [0, ] implies that P1(s) is stable and P1(j) is positive
real, i.e.,

10-302

sectf

P1 ( j ) + P 1 ( j ) 0

P1 = (s+2)/s in SEC[0,inf]
0

-0.05

-0.1

-0.1

-0.2

-0.15

-0.3

-0.2

-0.4

IMAG(P1)

IMAG(P)

P(s) = 1/(s+1) in SEC[-1,1]

0

-0.25

-0.5

-0.3

-0.6

-0.35

-0.7

-0.4

-0.8

-0.45

-0.9

-0.5
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

-1
0

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

REAL(P1)

REAL(P)

Figure 10-19: Example of Sector Transform

Algorithm

sectf uses the generalization of the sector concept of [3] described by [1]. First
the sector input data Sf= SECF and Sg=SECG is converted to two-port
state-space form; non-dynamical sectors are handled with empty a, b1, b2, c1,
c2 matrices. Next the equation

Sg ( s )

ug1
y g1

= Sf ( s )

uf1
y f1

is solved for the two-port transfer function T(s) from u g y f to u f y g . Finally,

1
1
1
1
the function lftf is used to compute G(s) as G=lftf(T,F).

10-303

sectf

Limitations

A well-posed conic sector must have det ( B A ) 0 or det

s 11 s 12 ;
0.
s 21 s 22

Also, you must have dim ( u F1 ) = dim ( y F1 ) since sectors are only defined for
square systems.

References

[1] Safonov, M.G., Stability and Robustness of Multivariable Feedback

Systems. Cambridge, MA: MIT Press, 1980.
[2] Safonov, M.G., E.A. Jonckheere, M. Verma and D.J.N. Limebeer,
Synthesis of Positive Real Multivariable Feedback Systems, Int. J. Control,
vol. 45, no. 3, pp. 817-842, 1987.
[3] Zames, G., On the Input-Output Stability of Time-Varying Nonlinear
Feedback Systems Part I: Conditions Using Concepts of Loop Gain,
Conicity, and Positivity, IEEE Trans. on Automat. Contr., AC-11, pp. 228-238,
1966.

See Also

10-304

lft
hinfsyn

Forms Redheffer star product of systems

H controller synthesis

setlmis

Purpose

10setlmis

Initialize description of LMI system

Syntax

setlmis(lmi0)

Description

Before starting the description of a new LMI system with lmivar and lmiterm,
type
setlmis([])

to initialize its internal representation.

To add on to an existing LMI system, use the syntax
setlmis(lmi0)

where lmi0 is the internal representation of this LMI system. Subsequent

lmivar and lmiterm commands will then add new variables and terms to the
initial LMI system lmi0.

See Also

getlmis
lmivar
lmiterm
newlmi

Get the internal description of an LMI system

Specify the matrix variables in an LMI problem
Specify the term content of LMIs
Attach an identifying tag to LMIs

10-305

setmvar

Purpose

10setmvar

Instantiate matrix variable and evaluate LMI terms

Syntax

newsys = setmvar(lmisys,X,Xval)

Description

setmvar sets the matrix variable X with identifier X to the value Xval. All

terms involving X are evaluated, the constant terms are updated accordingly,
and X is removed from the list of matrix variables. A description of the
resulting LMI system is returned in newsys.
The integer X is the identifier returned by lmivar when X is declared.
Instantiating X with setmvar does not alter the identifiers of the remaining
matrix variables.
The function setmvar is useful to freeze certain matrix variables and optimize
with respect to the remaining ones. It saves time by avoiding partial or
complete redefinition of the set of LMI constraints.

Examples

Consider the system

x = Ax + Bu
and the problem of finding a stabilizing state-feedback law u = Kx where K is
an unknown matrix.
By the Lyapunov Theorem, this is equivalent to finding P > 0 and K such that
(A + BK)P + P(A + BK)T + I < 0.
With the change of variable Y := KP, this condition reduces to the LMI
AP + PAT + BY + YTBT + I < 0.
This LMI is entered by the commands
n = size(A,1) % number of states
ncon = size(B,2) % number of inputs
setlmis([])
P = lmivar(1,[n 1]) % P full symmetric
Y = lmivar(2,[ncon n]) % Y rectangular
lmiterm([1 1 1 P],A,1,'s') % AP+PA'
lmiterm([1 1 1 Y],B,1,'s') % BY+Y'B'

10-306

setmvar

lmiterm([1 1 1 0],1) % I
lmis = getlmis

To find out whether this problem has a solution K for the particular Lyapunov
matrix P = I, set P to I by typing
news = setmvar(lmis,P,1)

The resulting LMI system news has only one variable Y = K. Its feasibility is
assessed by calling feasp:
[tmin,xfeas] = feasp(news)
Y = dec2mat(news,xfeas,Y)

The computed Y is feasible whenever tmin < 0.

See Also

evallmi
delmvar

Given a particular instance of the decision variables,

evaluate all variable terms in the system of LMIs
Delete one of the matrix variables of an LMI problem

10-307

showlmi

Purpose

10showlmi

Left- and right-hand sides of evaluated LMI

Syntax

[lhs,rhs] = showlmi(evalsys,n)

Description

For given values of the decision variables, the function evallmi evaluates all
variable terms in a system of LMIs. The left- and right-hand sides of the n-th
LMI are then constant matrices that can be displayed with showlmi. If evalsys
is the output of evallmi, the values lhs and rhs of these left- and right-hand
sides are given by
[lhs,rhs] = showlmi(evalsys,n)

An error is issued if evalsys still contains variable terms.

Examples

See the description of evallmi.

See Also

evallmi
setmvar

10-308

Given a particular instance of decision variables,

evaluate all variable terms in an LMI system
Instantiate a matrix variable and evaluate all LMI
terms involving this matrix variable

simplify

Purpose

10simplify

Simplify representation of uncertain object

Syntax

B
B
B
B

Description

B = simplify(A) performs model-reduction-like techniques to detect and

eliminate redundant copies of uncertain elements. Depending on the result, the
class of B may be lower than A. The AutoSimplify property of each uncertain
element in A governs what reduction methods are used. After reduction, any
uncertain element which does not actually affect the result is deleted from the
representation.

=
=
=
=

simplify(A)
simplify(A,'full')
simplify(A,'basic')
simplify(A,'off')

B = simplify(A,'full') overrides all uncertain elements AutoSimplify

property, and uses 'full' reduction techniques.
B = simplify(A,'basic') overrides all uncertain elements AutoSimplify
property, and uses 'basic' reduction techniques.
B = simplify(A,'off') does not perform reduction. However, any uncertain
elements in A with zero occurences are eliminated, and the class of B may be
lower than the class of A.

Examples

Create a simple umat with a single uncertain real parameter. Select specific
elements, note that result remains in class umat. Simplify those same elements,
and note that class changes.
p1 = ureal('p1',3,'Range',[2 5]);
L = [2 p1];
L(1)
UMAT: 1 Rows, 1 Columns
L(2)
UMAT: 1 Rows, 1 Columns
p1: real, nominal = 3, range = [2
simplify(L(1))
ans =
2
simplify(L(2))

5], 1 occurrence

10-309

simplify

Uncertain Real Parameter: Name p1, NominalValue 3, Range [2

5]

Create four uncertain real parameters, with a default value of AutoSimplify

('basic'), and define a high order polynomial [1].
m = ureal('m',125000,'Range',[100000 150000]);
xcg = ureal('xcg',.23,'Range',[.15 .31]);
zcg = ureal('zcg',.105,'Range',[0 .21]);
va = ureal('va',80,'Range',[70 90]);
cw = simplify(m/(va*va)*va,'full')
UMAT: 1 Rows, 1 Columns
m: real, nominal = 1.25e+005, range = [100000 150000], 1
occurrence
va: real, nominal = 80, range = [70 90], 1 occurrence
cw = m/va;
fac2 = .16726*xcg*cw*cw*zcg - .17230*xcg*xcg*cw ...
-3.9*xcg*cw*zcg - .28*xcg*xcg*cw*cw*zcg ...
-.07*xcg*xcg*zcg + .29*xcg*xcg*cw*zcg ...
+ 4.9*xcg*cw - 2.7*xcg*cw*cw ...
+.58*cw*cw - 0.25*xcg*xcg - 1.34*cw ...
+100.1*xcg -14.1*zcg - 1.91*cw*cw*zcg ...
+1.12*xcg*zcg + 24.6*cw*zcg ...
+.45*xcg*xcg*cw*cw - 46.85
UMAT: 1 Rows, 1 Columns
m: real, nominal = 1.25e+005, range = [100000 150000], 18
occurrences
va: real, nominal = 80, range = [70 90], 8 occurrences
xcg: real, nominal = 0.23, range = [0.15 0.31], 18 occurrences
zcg: real, nominal = 0.105, range = [0 0.21], 1 occurrence

The result of the high-order polynomial is an inefficient representation

involving 18 copies of m, 8 copies of va, 18 copies of xcg and 1 copy of zcg.
Simplify the expression, using the 'full' simplification algorithm
fac2s = simplify(fac2,'full')
UMAT: 1 Rows, 1 Columns
m: real, nominal = 1.25e+005, range = [100000 150000], 4
occurrences
va: real, nominal = 80, range = [70 90], 4 occurrences
xcg: real, nominal = 0.23, range = [0.15 0.31], 2 occurrences
zcg: real, nominal = 0.105, range = [0 0.21], 1 occurrence

10-310

simplify

which results in a much more economical representation.

Alternatively, change the AutoSimplify property of each parameter to 'full'
before forming the polynomial.
m.AutoSimplify = 'full';
xcg.AutoSimplify = 'full';
zcg.AutoSimplify = 'full';
va.AutoSimplify = 'full';

You can form the polynomial, which immediately gives a low order
representation.
cw = m/va;
fac2f = .16726*xcg*cw*cw*zcg - .17230*xcg*xcg*cw ...
-3.9*xcg*cw*zcg - .28*xcg*xcg*cw*cw*zcg ...
-.07*xcg*xcg*zcg + .29*xcg*xcg*cw*zcg ...
+ 4.9*xcg*cw - 2.7*xcg*cw*cw ...
+.58*cw*cw - 0.25*xcg*xcg - 1.34*cw ...
+100.1*xcg -14.1*zcg - 1.91*cw*cw*zcg ...
+1.12*xcg*zcg + 24.6*cw*zcg ...
+.45*xcg*xcg*cw*cw - 46.85
UMAT: 1 Rows, 1 Columns
m: real, nominal = 1.25e+005, range = [100000 150000], 4
occurrences
va: real, nominal = 80, range = [70 90], 4 occurrences
xcg: real, nominal = 0.23, range = [0.15 0.31], 2 occurrences
zcg: real, nominal = 0.105, range = [0 0.21], 1 occurrence

Create two real parameters, da and dx, and a 2-by-3 matrix, ABmat, involving
polynomial expressions in the two real parameters [2].
da = ureal('da',0,'Range',[-1 1]);
dx = ureal('dx',0,'Range',[-1 1]);
a11 = -.32 + da*(.8089 + da*(-.987 + 3.39*da)) + .15*dx;
a12 = .934 + da*(.0474 - .302*da);
a21 = -1.15 + da*(4.39 + da*(21.97 - 561*da*da)) ...
+ dx*(9.65 - da*(55.7 + da*177));
a22 = -.66 + da*(1.2 - da*2.27) + dx*(2.66 - 5.1*da);
b1 = -0.00071 + da*(0.00175 - da*.00308) + .0011*dx;
b2 = -0.031 + da*(.078 + da*(-.464 + 1.37*da)) + .0072*dx;
ABmat = [a11 a12 b1;a21 a22 b2]

10-311

simplify

UMAT: 2 Rows, 3 Columns

da: real, nominal = 0, range = [-1
dx: real, nominal = 0, range = [-1

1], 19 occurrences
1], 2 occurrences

Use 'full' simplification to reduce the complexity of the description.

ABmatsimp = simplify(ABmat,'full')
UMAT: 2 Rows, 3 Columns
da: real, nominal = 0, range = [-1
dx: real, nominal = 0, range = [-1

1], 7 occurrences
1], 2 occurrences

Alternatively, you can set the parameters AutoSimplify property to 'full'.

da.AutoSimplify = 'full';
dx.AutoSimplify = 'full';

Now you can rebuild the matrix

a11 = -.32 + da*(.8089 + da*(-.987 + 3.39*da)) + .15*dx;
a12 = .934 + da*(.0474 - .302*da);
a21 = -1.15 + da*(4.39 + da*(21.97 - 561*da*da)) ...
+ dx*(9.65 - da*(55.7 + da*177));
a22 = -.66 + da*(1.2 - da*2.27) + dx*(2.66 - 5.1*da);
b1 = -0.00071 + da*(0.00175 - da*.00308) + .0011*dx;
b2 = -0.031 + da*(.078 + da*(-.464 + 1.37*da)) + .0072*dx;
ABmatFull = [a11 a12 b1;a21 a22 b2]
UMAT: 2 Rows, 3 Columns
da: real, nominal = 0, range = [-1 1], 7 occurrences
dx: real, nominal = 0, range = [-1 1], 2 occurrences

Algorithm

simplify uses heuristics along with one-dimensional model reduction

algorithms to partially reduce the dimensionality of the representation of an
uncertain matrix or system.

Limitations

Multidimensional model reduction and realization theory are only partially

complete theories. The heuristics used by simplify are that - heuristics. The
order in which expressions involving uncertain elements are built up, eg.,
distributing across addition and multiplication, can affect the details of the
representation (i.e., the number of occurences of a ureal in an uncertain
matrix). It is possible that simplifys naive methods cannot completely resolve

10-312

simplify

these differences, so one may be forced to work with nonminimal

representations of uncertain systems.

References

[1] Varga, A. and G. Looye, Symbolic and numerical software tools for
LFT-based low order uncertainty modeling, IEEE International Symposium
on Computer Aided Control System Design, 1999, pp. 5-11.
[2] Belcastro, C.M., K.B. Lim and E.A. Morelli, Computer aided uncertainty
modeling for nonlinear parameter-dependent systems Part II: F-16 example,
IEEE International Symposium on Computer Aided Control System Design,
1999, pp. 17-23.

See Also

umat
uss
ucomplex
ureal
uss

Creates an uncertain matrix object

Creates an uncertain system object
Creates an uncertain complex parameter
Creates an uncertain real parameter
Creates an uncertain system

10-313

skewdec

Purpose

10skewdec

Create skew-symmetric matrix

Syntax

x = skewdec(m,n)

Description

skewdec(m,n) forms the m-by-m skew-symmetric matrix

0
( n 1 ) ( n 2 )
(n + 1)
0
( n 3 )
(n + 2) (n + 3)
0

This function is useful to define skew-symmetric matrix variables. In this case,

set n to the number of decision variables already used.

See Also

decinfo
lmivar

10-314

Describe how the entries of a matrix variable X relate

to the decision variables
Specify the matrix variables in an LMI problem

slowfast

Purpose

10slowfast

State-space slow-fast decomposition

Syntax

[G1,G2] = slowfast(G,cut)

Description

slowfast computes the slow and fast modes decompositions of a system G(s)

such that
G ( s ) = [ G ( s ) ]s + [ G ( s ) ]f
,B
,D
,C
) denotes the slow part of G(s), and
where [ G ( s ) ] s : = ( A
11
1
1
1
,B
,D
,C
) denotes the fast part. The variable cut denotes the
[ G ( s ) ] := ( A
22

index where the modes will be split.

stabproj employs the algorithm in [1] as follows:

Find a unitary matrix V via the ordered Schur decomposition routines blksch
or rschur such that
T

A = V AV =

A
11 A 12

0 A
22

) < (A
).
Based on the style of ordered Schur form, you get i ( A
11
22
i
Finally solving the matrix equation for X
X XA
+A
=0
A
11
22
12
you get the state-space projections
,B
,D
,C
)
[ G ( s ) ]s : = ( A
11
1
1
1
and
,B
,D
,C
)
[ G ( s ) ]f : = ( A
22
2
2
2
where

10-315

slowfast

B
1
: = I X VB

0 I
B2
and
C
: = CV T I X
C
1 2
0 I

References

[1] M.G. Safonov, E.A. Jonckheere, M. Verma and D.J.N. Limebeer, Synthesis
of Positive Real Multivariable Feedback Systems, Int. J. Control, vol. 45, no.
3, pp. 817-842, 1987.

See Also

schur
modreal

10-316

Schur decomposition
Modal form realization

squeeze

Purpose

10squeeze

Remove singleton dimensions for umat objects

Syntax

B = squeeze(A)

Description

B = squeeze(A) returns an array B with the same elements as A but with all
the singleton dimensions removed. A singleton is a dimension such that
size(A,dim)==1. 2-D arrays are unaffected by squeeze so that row vectors
remain rows.

See Also

permute
reshape

Permutes array dimensions.

Changes size of matrix

10-317

stabproj

Purpose

10stabproj

State-space stable/anti-stable decomposition

Syntax

[G1,G2,m] = stabproj(G)

Description

stabproj computes the stable and antistable projections of a minimal

Algorithm

stabproj employs the algorithm in [1] as follows:

realization G(s) such that G ( s ) = [ G ( s ) ] - + [ G ( s ) ] + where

,B
,D
,C
) denotes the stable part of G(s), and
[ G ( s ) ]- : = ( A
11
1
1
1

) denotes the antistable part. The variable m

[ G ( s ) ] : = ( A 22, B 2, C 2, D
2
+
returns the number of stable eigenvalues of A.

Find a unitary matrix V via the ordered Schur decomposition routines blksch
or rschur such that
T

A = V AV =

A
11 A 12

0 A
22

Based on the style of ordered Schur form, you can get a stable A
11 and an
. (A
) < (A
) for the case of slowfast.
antistable A
22

11

22

Finally solving the matrix equation for X

X XA
+A
=0
A
11
22
12
,B
,D
,C
)
you get the state-space projections [ G ( s ) ] - or [ G ( s ) ] s : = ( A
11
1
1
1
,B
,D
,C
)
and [ G ( s ) ] + or [ G ( s ) ] us : = ( A
22
2
2
2
where

B
1
: = I X VB

0 I
B2
and

10-318

stabproj

C
: = CV T I X
C
1 2
0 I

References

[1] M.G. Safonov, E.A. Jonckheere, M. Verma and D.J.N. Limebeer, Synthesis
of Positive Real Multivariable Feedback Systems, Int. J. Control, vol. 45, no.
3, pp. 817-842, 1987.

See Also

schur
modreal

Schur decomposition
Modal form realization

10-319

stack

Purpose

10stack

Construct array by stacking uncertain matrices, models, or arrays

Syntax

umatout = stack(arraydim,umat1,umat2,...)
usysout = stack(arraydim,usys1,usys2,...)

Description

stack constructs an uncertain array by stacking uncertain matrices, models,

or arrays along array dimensions of an uncertain array.

umatout = stack(arraydim,umat1,umat2,...) produces an array of
uncertain matrices, umatout, by stacking (concatenating) the umat matrices (or
umat arrays) umat1, umat2,... along the array dimension arraydim. All models
must have the same number of columns and rows. The column/row dimensions
are not counted in the array dimensions.
umatout = stack(arraydim,usys1,usys2,...) produces an array of
uncertain models, ufrd or uss, or usysout, by stacking (concatenating) the
ufrd or uss matrices (or ufrd or uss arrays) usys1, usys2,... along the array
dimension arraydim. All models must have the same number of columns and
rows (the same input/output dimensions). Note that the input/output
dimensions are not considered for arrays.

Examples

Consider usys1 and usys2, two single-input/single-output uss models:

zeta
wn =
P1 =
P2 =

= ureal('zeta',1,'Range',[0.4 4]);
ureal('wn',0.5,'Range',[0.3 0.7]);
tf(1,[1 2*zeta*wn wn^2]);
tf(zeta,[1 10]);

You can stack along the first dimension to produce a 2-by-1 uss array.
stack(1,P1,P1)
USS: 2 States, 1 Output, 1 Input, Continuous System [array, 2 x 1]
wn: real, nominal = 0.5, range = [0.3 0.7], 3 occurrences
zeta: real, nominal = 1, range = [0.4 4], 1 occurrence

You can stack along the second dimension to produce a 1-by-2 uss array.
stack(2,P1,P2)
% produces a 1-by-2 USS array.
USS: 2 States, 1 Output, 1 Input, Continuous System [array, 1 x 2]
wn: real, nominal = 0.5, range = [0.3 0.7], 3 occurrences
zeta: real, nominal = 1, range = [0.4 4], 1 occurrence

10-320

stack

You can stack along the third dimension to produce a 1-by-1-by-2 uss array.
stack(3,P1,P2)
% produces a 1-by-1-by-2 USS array.
USS: 2 States, 1 Output, 1 Input, Continuous System [array, 1 x 1
x 2]
wn: real, nominal = 0.5, range = [0.3 0.7], 3 occurrences
zeta: real, nominal = 1, range = [0.4 4], 1 occurrence

See Also

append
blkdiag
horzcat
vertcat

Groups models by appending their inputs and outputs

Groups models by appending their inputs and outputs
Performs horizontal concatenation
Performs vertical concatenation

10-321

symdec

Purpose

10symdec

Form symmetric matrix

Syntax

x = symdec(m,n)

Description

symdec(m,n) forms an m-by-m symmetric matrix of the form

(n + 1) (n + 2) (n + 4)
(n + 2) (n + 3) (n + 5)
(n + 4) (n + 5) (n + 6)

This function is useful to define symmetric matrix variables. n is the number

of decision variables.

See Also

10-322

decinfo

Show how matrix variables depend on decision

variables

sysic

Purpose

10sysic

Build interconnections of certain and uncertain matrices and systems

Syntax

sysout = sysic

Description

sysic requires that 3 variables with fixed names be present in the calling
workspace: systemnames, inputvar and outputvar.
systemnames is a char containing the names of the subsystems (double, tf,
zpk, ss, uss, frd, ufrd, etc) that make up the interconnection. The names

must be separated by spaces with no additional punctuation. Each named

variable must exist in the calling workspace.
inputvar is a char, defining the names of the external inputs to the
interconnection. The names are separated by semicolons, and the entire list is
enclosed in square brackets [ ]. Inputs can be scalar or multivariate. For
instance, a 3-component (x,y,z) force input can be specified with 3 separate
names, Fx, Fy, Fz. Alternatively, a single name with a defined integer
dimension can be specified, as in F{3}. The order of names in inputvar
determines the order of inputs in the interconnection.
outputvar is a char, describing the outputs of the interconnection. Outputs do
not have names-they are simply linear combinations of individual subsystems
outputs and external inputs. Semicolons delineate separate components of the
interconnections outputs. Between semicolons, signals can be added and
subtracted, and multiplied by scalars. For multivariable subsystems,
arguments within parentheses specify which subsystem outputs are to be used
and in what order. For instance, plant(2:4,1,9:11) specifies outputs
2,3,4,1,9,10,11 from the subsystem plant. If a subsystem is listed in
outputvar without arguments, then all outputs from that subsystem are used.
sysic also requires that for every subsystem name listed in systemnames, a
corresponding variable, input_to_ListedSubSystemName must exist in the
calling workspace. This variable is similar to outputvar it defines the input
signals to this particular subsystem as linear combinations of individual
subsystems outputs and external inputs.
sysout = sysic will perform the interconnection described by the variables
above, using the subsystem data in the names found in systemnames. The

resulting interconnection is returned in the output argument, listed above as

sysout.

10-323

sysic

Examples

A simple system interconnection, identical to the system illustrated in the

iconnect description. Consider a three-input, two-output LTI matrix T,

y1

noise
deltemp
setpoint

y2

which has internal structure


y2 
y1


g; g
6+ 6++
noise

57.3

setpoint




wt

deltemp

act

P = rss(3,2,2);
K = rss(1,1,2);
A = rss(1,1,1);
W = rss(1,1,1);
systemnames = 'W A K P';
inputvar = '[noise;deltemp;setpoint]';
outputvar = '[57.3*P(1);setpoint-P(2)]';
input_to_W = '[deltemp]';
input_to_A = '[K]';
input_to_K = '[P(2)+noise;setpoint]';
input_to_P = '[W;A]';
T = sysic;

Limitations

10-324

The syntax of sysic is limited, and for the most part is restricted to what is
shown here. The iconnect interconnection object can also be used to define
complex interconnections, and has a more flexible syntax.

sysic

Within sysic, error-checking routines monitor the consistency and availability

of the subsystems and their inputs. These routines provide a basic level of error
detection to aid the user in debugging.

See Also

iconnect

Equates expressions for icsignal objects

10-325

ucomplex

Purpose

10ucomplex

Create uncertain complex parameter

Syntax

A = ucomplex('NAME',nominalvalue)
A = ucomplex('NAME',nominalvalue,'Property1',Value1,...
'Property2',Value2,...)

Description

An uncertain complex parameter is used to represent a complex number whose

value is uncertain. Uncertain complex parameters have a name (the Name
property), and a nominal value (the NominalValue property).
The uncertainty (potential deviation from the nominal value) is described in
two different manners:
Radius (radius of disc centered at NominalValue)
Percentage (disc size is percentage of magnitude of NominalValue)
The Mode property determines which description remains invariant if the
NominalValue is changed (the other is derived). The default Mode is 'Radius'
and the default radius is 1.
Property/Value pairs can also be specified at creation. For instance,
B = ucomplex('B',6-j,'Percentage',25)

sets the nominal value to 6-j, the percentage uncertainty to 25 and, implicitly,
the Mode to 'Percentage'.

Examples

Create an uncertain complex parameter with internal name A. The uncertain

parameters possible values are a complex disc of radius 1, centered at 4+3j.
The value of A.percentage is 20 (radius is 1/5 of the magnitude of the nominal
value).
A = ucomplex('A',4+3*j)
Uncertain Complex Parameter: Name A, NominalValue 4+3i, Radius 1

You can visualize the uncertain complex parameter by sampling and plotting
the data.
sa = usample(A,400);
w = linspace(0,2*pi,200);
circ = sin(w) + j*cos(w);
rc = real(A.NominalValue+circ);

10-326

ucomplex

ic = imag(A.NominalValue+circ);
plot(real(sa(:)),imag(sa(:)),'o',rc,ic,'k-')
xlim([2.5 5.5])
ylim([1.5 4.5])
axis equal

Sampled complex parameter A

4.5

3.5

2.5

1.5

See Also

get
umat
ucomplexm
ultidyn
ureal

2.5

3.5

4.5

5.5

Gets object properties

Creates an uncertain matrix object
Creates an uncertain complex matrix
Creates an uncertain LTI dynamic object
Creates an uncertain real parameter

10-327

ucomplexm

Purpose

10ucomplexm

Create uncertain complex matrix

Syntax

M = ucomplexm('Name',NominalValue)
M = ucomplexm('Name',NominalValue,'WL',WLvalue,'WR',WRvalue)
M = ucomplexm('Name',NominalValue,'Property',Value)

Description

M = ucomplexm('Name',NominalValue) creates an uncertain complex matrix

representing a ball of complex-valued matrices, centered at a NominalValue
and named Name.
M = ucomplexm('Name',NominalValue,'WL',WLvalue,'WR',WRvalue)
creates an uncertain complex matrix with weights WL and WR. Specifically, the
values represented by M are all matrices H that satisfy norm(inv(M.WL)*(H M.NominalValue)*inv(M.WR)) <= 1. WL and WR are square, invertible, and

weighting matrices that quantify the size and shape of the ball of matrices
represented by this object. The default values for WL and WR are identity
matrices of appropriate dimensions.
Trailing Property/Value pairs are allowed, as in
M = ucomplexm('NAME',nominalvalue,'P1',V1,'P2',V2,...)

The property AutoSimplify controls how expressions involving the uncertain

matrix are simplified. Its default value is 'basic', which means elementary
methods of simplification are applied as operations are completed. Other
values for AutoSimplify are 'off'', no simplification performed, and 'full'
which applies model-reduction-like techniques to the uncertain object.

Examples

Create a ucomplexm with the name 'F', nominal value [1 2 3; 4 5 6], and
weighting matrices WL = diag([.1.3]), WR = diag([.4 .8 1.2]).
F = ucomplexm('F',[1 2 3;4 5 6],'WL',diag([.1 .3]),...
'WR',diag([.4 .8 1.2]));

Sample the difference between the uncertain matrix and its nominal value at
80 points, yielding a 2-by-3-by-80 matrix typicaldev.
typicaldev = usample(F-F.NominalValue,40);

Plot histograms of the deviations in the (1,1) entry as well as the deviations in
the (2,3) entry.

10-328

ucomplexm

The absolute values of the (1,1) entry and the (2,3) entry are shown by
histogram plots. Typical deviations in the (1,1) entry should be about 10 times
smaller than the typical deviations in the (2,3) entry.
subplot(2,1,1);
hist(abs(typicaldev(1,1,:)));xlim([0 .25])
title('Sampled F(1,1) - F(1,1).NominalValue')
subplot(2,1,2);
hist(abs(typicaldev(2,3,:)));xlim([0 .25])
title('Sampled F(2,3) - F(2,3).NominalValue')

Sampled F(1,1) F(1,1).NominalValue

12
10
8
6
4
2
0

0.05

0.1

0.15

0.2

0.25

0.2

0.25

Sampled F(2,3) F(2,3).NominalValue

20

15

10

See Also

get
umat
ucomplex
ultidyn
ureal

0.05

0.1

0.15

Gets object properties

Creates an uncertain matrix object
Creates an uncertain complex parameter
Creates an uncertain LTI dynamic object
Creates an uncertain real parameter

10-329

udyn

Purpose

10udyn

Create unstructured uncertain dynamic system object

Syntax

n = udyn('name',iosize);

Description

n = udyn('name',iosize) creates an unstructured uncertain dynamic system

class, with input/output dimension specified by iosize. This object represents
the class of completely unknown multivariable, time-varying nonlinear
systems.

For practical purposes, these uncertain elements represent noncommuting

symbolic variables (placeholders). All algebraic operations, such as addition,
subtraction, and multiplication (i.e., cascade) operate properly, and
substitution (with usubs) is allowed.
The analysis tools (e.g., robuststab) do not currently handle these types of
uncertain elements. Therefore, these elements do not provide a significant
amount of usability, and their role in the toolbox is small.

Examples

You can create a 2-by-3 udyn element and check its size and properties.
N = udyn('N',[2 3])
Uncertain Dynamic System: Name N, size 2x3
size(N)
ans =
2
3
get(N)
Name: 'N'
NominalValue: [2x3 double]
AutoSimplify: 'basic'

See Also

10-330

ureal
ultidyn
ucomplex
ucomplexm

Creates an uncertain real parameter

Creates an uncertain linear time-invariant object
Creates an uncertain complex parameter
Creates an uncertain complex matrix

ufrd

Purpose

Create uncertain frequency response data (ufrd) object, or convert another

model type to ufrd model

Syntax

usysfrd = ufrd(usys,frequency)
usysfrd = ufrd(usys,frequency,'Units',units)
usysfrd = ufrd(sysfrd)
usys = ufrd(response,frequency)
usys = ufrd(response,frequency,Ts)
usys = ufrd(response,frequency,RefSys)
usys = ufrd(response,frequency,'Units',units,Ts)
usys = ufrd(response,frequency,'Units',units,Ts,RefSys)

Description

Uncertain frequency response data (ufrd) models result from the

conversion of an uncertain state-space (uss) system to its uncertain frequency
response. ufrd models also result when frequency response data models (frd)
are combined with uncertain matrices (umat).

10ufrd

usysfrd = ufrd(usys,frequency,'Units',units) converts a uss model usys

to a ufrd model usysfrd by frequency response.'Units' specifies the units of
the frequencies in frequency, which can be 'rad/s' or 'Hz'. If the last two
arguments are omitted, the default for frequency units is 'rad/s'.

Any of the previous syntaxes can be followed by property name/property value

pairs.
usysfrd = ufrd(usys,frequency,'Units',units,'P1',V1,'P2',V2,...)
sets the properties P1, P2, ... to the values V1, V2, ...
usys = ufrd(response,frequency) creates a ufrd from the response and
frequency arguments. response should be a umat array, whose first array
dimension (i.e., size(response,3)) aligns with the frequency. Note that you

are unlikely to use this option.

usysfrd = ufrd(sysfrd) converts an frd model sysfrd to a ufrd model
usysfrd with no uncertain elements.

Examples

Example 1
You create a continuous-time uncertain system with both parametric
uncertainty and unmodeled dynamics uncertainty. Compute the uncertain

10-331

ufrd

frequency response and plot the Bode plot, using 20 random samples, with a
color choice of red for random samples, and blue for nominal.
p1 = ureal('p1',5,'Range',[2 6]);
p2 = ureal('p2',3,'Plusminus',0.4);
p3 = ultidyn('p3',[1 1]);
Wt = makeweight(.15,30,10);
A = [-p1 0;p2 -p1];
B = [0;p2];
C = [1 1];
usys = uss(A,B,C,0)*(1+Wt*p3);
usysfrd = ufrd(usys,logspace(-2,2,60));
bode(usysfrd,'r',usysfrd.NominalValue,'b+')

Example 2
You convert a not-uncertain frd model to ufrd without uncertainties. You can
verify the equality of the nominal value of the ufrd and simplified
representation to the original system.
G = frd(tf([1 2 3],[1 2 3 4]),logspace(-2,2,40));
usys = ufrd(G)
UFRD: 1 Output, 1 Input, Continuous System, 40 Frequency points
isequal(usys.NominalValue,G)
ans =
1
isequal(simplify(usys,'class'),G)
ans =
1

See Also

10-332

frd
ss

Creates or converts to frequency response data model

Creates or converts to state-space model

ultidyn

Purpose

10ultidyn

Create uncertain linear time-invariant object

Syntax

H = ultidyn('Name',iosize)
H =
ultidyn('Name',iosize,'Property1',Value1,'Property2',Value2,...)

Description

H = ultidyn('Name',iosize) creates an uncertain linear, time-invariant

objects are used to represent unknown dynamic objects whose only known
attributes are bounds on their frequency response. Uncertain linear,
time-invariant objects have a name (the Name property), and an input/output
size (ioSize property).
The property Type is 'GainBounded' (default) or 'PositiveReal', and
describes in what form the knowledge about the objects frequency response is
specified.
If Type is 'GainBounded', then the knowledge is an upper bound on the
magnitude (i.e., absolute value), namely abs(H)<= Bound at all frequencies.
The matrix generalization of this is ||H||<= Bound.
If Type is 'PositiveReal' then the knowledge is a lower bound on the real
part, namely Real(H) >= Bound at all frequencies. The matrix generalization
of this is H+H' >= 2*Bound
The property Bound is a real, scalar that quantifies the bound on the frequency
response of the uncertain object as described above.
Trailing Property/Value pairs are allowed in the construction.
H=ultidyn('name',iosize,'Property1',Value1,'Property2',Value2,...)

The property SampleStateDim is a positive integer, defining the state

dimension of random samples of the uncertain object when sampled with
usample. The default value is 1.
The property AutoSimplify controls how expressions involving the uncertain
matrix are simplified. Its default value is 'basic', which means elementary
methods of simplification are applied as operations are completed. Other
values for AutoSimplify are 'off'', no simplification performed, and 'full'
which applies model-reduction-like techniques to the uncertain object.

10-333

ultidyn

Examples

Example 1
Create an ultidyn object with internal name 'H', dimensions 2-by-3, norm
bounded by 7.
H = ultidyn('H',[2 3],'Bound',7)
Uncertain GainBounded LTI Dynamics: Name H, 2x3, Gain Bound = 7

Example 2
Create a scalar ultidyn object with an internal name 'B', whose frequency
response has a real part greater than 2.5. Change the SampleStateDim to 5,
and plot the Nyquist plot of 30 random samples.
B = ultidyn('B',[1 1],'Type','PositiveReal','Bound',2.5)
Uncertain PositiveReal LTI Dynamics: Name B, 1x1, M+M' >= 2*(2.5)
B.SampleStateDim = 5;
nyquist(usample(B,30))

Nyquist Diagram

30

20

Imaginary Axis

10

10

20

30
10

10

20

30

40

50

60

Real Axis

See Also

10-334

get
ureal
uss

Gets object properties

Creates an uncertain real parameter
Creates an uncertain LTI system object

umat

Purpose

10umat

Create uncertain matrix

Syntax

h = umat(m)

Description

Uncertain matrices are usually created by manipulation of uncertain atoms

(ureal, ucomplex, ultidyn, etc.), double matrices, and other uncertain
matrices. Most standard matrix manipulations are valid, including addition,
multiplication, inverse, horizontal and vertical concatenation. Specific
rows/columns of an uncertain matrix can be referenced and assigned also.
The command umat is rarely used. There are two situations where it may be
useful. If M is a double, then H = umat(M) recasts M as an uncertain matrix
(umat object) without any uncertainties. Similarly, if M is an uncertain atom,
then H = umat(M) recasts M as an uncertain matrix (umat object) whose value
is merely the uncertain atom. In both cases, simplify(H,'class') is the same
as M.
If M is a umat, then M.NominalValue is the result obtained by replacing each
uncertain atom in M with its own nominal value.
If M is a umat, then M.Uncertainty is an object describing all the uncertain
atoms in M. All atoms can be referenced and their properties modified with this
Uncertainty gateway. For instance, if B is an uncertain real parameter in M,
then M.Uncertainty.B accesses the uncertain atom B in M.

Examples

Create 3 uncertain atoms and then a 3-by-2 umat.

a = ureal('a',5,'Range',[2 6]);
b = ucomplex('b',1+j,'Radius',0.5);
c = ureal('c',3,'Plusminus',0.4);
M = [a b;b*a 7;c-a b^2]
UMAT: 3 Rows, 2 Columns
a: real, nominal = 5, range = [2 6], 1 occurrence
b: complex, nominal = 1+1i, radius = 0.5, 4 occurrences
c: real, nominal = 3, variability = [-0.4 0.4], 1 occurrence

View the properties of M with get

get(M)
NominalValue: [3x2 double]
Uncertainty: [1x1 atomlist]

10-335

umat

The nominal value of M is the result when all atoms are replaced by their
nominal values. View the properties of M with get
M.NominalValue
ans =
5.0000
5.0000 + 5.0000i
-2.0000

1.0000 + 1.0000i
7.0000
0 + 2.0000i

Change the nominal value of a within M to 4. The nominal value of M reflects

this change.
M.Uncertainty.a.NominalValue = 4;
M.NominalValue
ans =
4.0000
1.0000 + 1.0000i
4.0000 + 4.0000i
7.0000
-1.0000
0 + 2.0000i

Get a random sample of M, obtained by taking random samples of the uncertain

atoms within M.
usample(M)
ans =
2.0072
1.7358 + 2.7808i
1.3829

0.8647 + 1.3854i
7.0000
-1.1715 + 2.3960i

Select the 1st and 3rd rows, and the 2nd column of M. The result is a 2-by-1
umat, whose dependence is only on b.
M([1 3],2)
UMAT: 2 Rows, 1 Columns
b: complex, nominal = 1+1i, radius = 0.5, 3 occurrences

See Also

10-336

ureal
ultidyn
ucomplex
ucomplexm
usample

Creates an uncertain real parameter

Creates an uncertain linear time-invariant object
Creates an uncertain complex parameter
Creates an uncertain complex matrix
Generates random samples of an uncertain object

uplot

Purpose

10uplot

Plot multiple frequency response objects and doubles on same graph

Syntax

uplot(G1)
uplot(G1,G2)
uplot(G1,Xdata,Ydata)
uplot(G1,Xdata,Ydata,...)
uplot(G1,linetype)
uplot(G1,linetype,G2,...)
uplot(G1,linetype,Xdata,Ydata,linetype)
uplot(type,G1,linetype,Xdata,Ydata,linetype)
H = uplot(G1)
H = uplot(G1,G2)
H = uplot(G1,Xdata,Ydata)
H = uplot(G1,Xdata,Ydata,...)
H = uplot(G1,linetype)
H = uplot(G1,linetype,G2,...)
H = uplot(G1,linetype,Xdata,Ydata,linetype)

Description

uplot plots double and frd objects. The syntax is the same as the MATLAB
plot command except that all data is contained in frd objects, and the axes
are specified by type.

The (optional) type argument must be one of

Type

Description

'iv,d'

Data versus independent variable (default)

'iv,m'

Magnitude versus independent variable

'iv,lm'

log(magnitude) versus independent variable

'iv,p'

Phase versus independent variable

'liv,m'

Magnitude versus log(independent variable)

'liv,d'

Data versus log(independent variable)

'liv,m'

Magnitude versus log(independent variable)

10-337

uplot

Type

Description

'liv,lm'

log(magnitude) versus log(independent variable)

'liv,p'

Phase versus log(independent variable)

'r,i'

Real versus imaginary (parametrize by independent

variable)

'nyq'

Real versus imaginary (parametrize by independent

variable)

'nic'

Nicholas plot

'bode'

Bode magnitude and phase plot

The remaining arguments of uplot take the same form as the MATLAB plot
command. Line types (for example,'+', 'g-.', or '*r') can be optionally
specified after any frequency response argument.
There is a subtle distinction between constants and frd objects with only one
independent variable. A constant is treated as such across all frequencies, and
consequently shows up as a line on any graph with the independent variable as
an axis. A frd object with only one frequency point always shows up as a point.
You might need to specify one of the more obvious point types in order to see it
(e.g., '+', 'x', etc.).

Examples

Two SISO second-order systems are created, and their frequency responses are
calculated over different frequency ranges.
a1 = [-1,1;-1,-0.5];
b1 = [0;2]; c1 = [1,0]; d1 = 0;
sys1 = ss(a1,b1,c1,d1);
a2 = [-.1,1;-1,-0.05];
b2 = [1;1]; c2 = [-0.5,0]; d2 = 0.1;
sys2 = ss(a2,b2,c2,d2);
omega = logspace(-2,2,100);
sys1g = frd(sys1,omega);
omega2 = [ [0.05:0.1:1.5] [1.6:.5:20] [0.9:0.01:1.1] ];
omega2 = sort(omega2);
sys2g = frd(sys2,omega2);

10-338

uplot

An frd object with a single frequency is also created. Note the distinction
between the frd object and the constant matrix in the subsequent plots.
sys3 = rss(1,1,1);
rspot = frd(sys3,2);

The following plot uses the 'liv,lm' plot_type specification.

uplot('liv,lm',sys1g,'b-.',rspot,'r*',sys2g);
xlabel('log independent variable')
ylabel('log magnitude')
title('axis specification: liv,lm')
axis specification: liv,lm

10

log magnitude

10

10

10

10

10

10

10

10

10

10

log independent variable

See Also

bode
plot
nichols
nyquist
semilogx, semilogy
sigma

Plots Bode frequency response

Plots on linear axis
Plots Nichols frequency response
Plots Nyquist frequency response
Plots semi-log scale plot
Plots singular values of a LTI system

10-339

ureal

Purpose

10ureal

Create uncertain real parameter

Syntax

p = ureal('name',nominalvalue)
p = ureal('name',nominalvalue,'Property1',Value1,...
'Property2',Value2,...)

Description

An uncertain real parameter is used to represent a real number whose value is

uncertain. Uncertain real parameters have a name (the Name property), and a
nominal value (NominalValue property).
The uncertainty (potential deviation from NominalValue) is described
(equivalently) in 3 different properties:
PlusMinus: the additive deviation from NominalValue
Range: the interval containing NominalValue
Percentage: the percentage deviation from NominalValue
The Mode property specifies which one of these three descriptions remains
unchanged if the NominalValue is changed (the other two descriptions are
derived). The possible values for the Mode property are 'Range',
'Percentage' and 'PlusMinus'.
The default Mode is 'PlusMinus', and [-1 1] is the default value for the
'PlusMinus' property. The range of uncertainty need not be symmetric about
NominalValue.
The property AutoSimplify controls how expressions involving the uncertain
matrix are simplified. Its default value is 'basic', which means elementary
methods of simplification are applied as operations are completed. Other
values for AutoSimplify are 'off'', no simplification performed, and
'full', which applies model-reduction-like techniques to the uncertain object.

Examples

Example 1
Create an uncertain real parameter and use get to display the properties and
their values. Create uncertain real parameter object a with the internal name
'a' and nominal value 5.
a = ureal('a',5)
Uncertain Real Parameter: Name a, NominalValue 5, variability =
[-1 1]

10-340

ureal

get(a)
Name:
NominalValue:
Mode:
Range:
PlusMinus:
Percentage:
AutoSimplify:

'a'
5
'PlusMinus'
[4 6]
[-1 1]
[-20 20]
'basic'

Note that the Mode is 'PlusMinus', and that the value of PlusMinus is indeed
[-1 1]. As expected, the range description of uncertainty is [4 6], while the
percentage description of uncertainty is [-20 20].
Set the range to [3 9]. This leaves Mode and NominalValue unchanged, but all
three descriptions of uncertainty have been modified.
a.Range = [3 9];
get(a)
Name:
NominalValue:
Mode:
Range:
PlusMinus:
Percentage:
AutoSimplify:

'a'
5
'PlusMinus'
[3 9]
[-2 4]
[-40 80]
'basic'

Example 2
Property/Value pairs can also be specified at creation.
b = ureal('b',6,'Percentage',[-30 40],'AutoSimplify','full');
get(b)
Name: 'b'
NominalValue: 6
Mode: 'Percentage'
Range: [4.2000 8.4000]
PlusMinus: [-1.8000 2.4000]
Percentage: [-30.0000 40.0000]
AutoSimplify: 'full'

Note that Mode is automatically set to 'Percentage'.

10-341

ureal
Example 3
Specify the uncertainty in terms of percentage, but force Mode to 'Range'.
c = ureal('c',4,'Mode','Range','Percentage',25);
get(c)
Name: 'c'
NominalValue: 4
Mode: 'Range'
Range: [3 5]
PlusMinus: [-1 1]
Percentage: [-25 25]
AutoSimplify: 'basic'

See Also

10-342

ucomplex
umat
uss

Creates an uncertain complex parameter

Creates an uncertain matrix
Creates an uncertain, linear dynamic object

usample

Purpose

10usample

Generate random samples of uncertain object

Syntax

B = usample(A);
B = usample(A,N)
[B,SampleValues] = usample(A,N)
[B,SampleValues] = usample(A,Names,N)
[B,SampleValues] = usample(A,Names1,N1,Names2,N2,...)

Description

B = usample(A) substitutes a random sample of the uncertain objects in A,

returning a certain (i.e., not uncertain) array of size [size(A)].
B = usample(A,N) substitutes N random samples of the uncertain objects in A,
returning a certain (i.e., not uncertain) array of size [size(A) N].
[B,SampleValues] = usample(A,N) additionally returns the specific sampled
values (as a Struct whose field names are the names of As uncertain elements)
of the uncertain elements. Hence, B is the same as usubs(A,SampleValues).
[B,SampleValues] = usample(A,Names,N) samples only the uncertain
elements listed in the Names variable (cell, or char array). If Names does not
include all the uncertain objects in A, then B will be an uncertain object. Any
entries of Names that are not elements of A are simply ignored. Note that
usample(A,fieldnames(A.Uncertainty),N) is the same as usample(A,N).
[B,SampleValues] = usample(A,Names1,N1,Names2,N2,...) takes N1
samples of the uncertain elements listed in Names1, and N2 samples of the
uncertain elements listed in Names2, and so on. size(B) will equal [size(A) N1
N2 ...].

Examples

Sample a real parameter and plot a histogram.

A = ureal('A',5);
Asample = usample(A,500);
size(A)
ans =
1
1
size(Asample)
ans =
1
1
500

10-343

usample

class(Asample)
ans =
double
hist(Asample(:))

The second example illustrates the open and closed-loop response of an

uncertain plant model. You can create two uncertain real parameters and an
uncertain plant.
gamma = ureal('gamma',4);
tau = ureal('tau',.5,'Percentage',30);
P = tf(gamma,[tau 1]);

Create an integral controller based on nominal plant parameter.

KI = 1/(2*tau.Nominal*gamma.Nominal);
C = tf(KI,[1 0]);

Now create an uncertain closed-loop system.

CLP = feedback(P*C,1);

10-344

usample

You can sample the plant at 20 values (distributed uniformly about the tau and
gamma parameter cube).
[Psample1D,Values1D] = usample(P,20);
size(Psample1D)
20x1 array of state-space models
Each model has 1 output, 1 input, and 1 state.

You can sample the plant P at 10 values in the tau parameter and 15 values in
the gamma parameter.
[Psample2D,Values2D] = usample(P,'tau',10,'gamma',15);
size(Psample2D)
10x15 array of state-space models
Each model has 1 output, 1 input, and 1 state.

You can plot the 1-D sampled plant step responses

subplot(2,1,1); step(Psample1D)

You can also evaluate the uncertain closed-loop at the same values, and plot
the step response using usubs.
subplot(2,1,2); step(usubs(CLP,Values1D))

10-345

usample

Step Response

6
5

Amplitude

4
3
2
1
0

0.5

1.5

2
Time (sec)

2.5

3.5

Step Response

1.5

Amplitude

0.5

See Also

10-346

usubs

4
Time (sec)

Substitutes values for uncertain atoms

uss

Purpose

10uss

Specify uncertain state-space models or convert LTI model to uncertain

state-space model

Syntax

usys
usys
usys
usys
usys
usys

Description

uss creates uncertain state-space models (uss objects) or to convert LTI models
to the uss class.

=
=
=
=
=
=

uss(a,b,c,d)
uss(a,b,c,d,Ts)
uss(d)
uss(a,b,c,d,Property,Value,...)
uss(a,b,c,d,Ts,Property,Value,...)
uss(sys)

usys = uss(a,b,c,d) creates a continuous-time uncertain state-space object.

The matrices a, b, c and d can be umat and/or double and/or uncertain
atoms. These are the 4 matrices associated with the linear differential equation
model to describe the system.
usys = uss(a,b,c,d,Ts) creates a discrete-time uncertain state-space object

with sampling time Ts.

usys = uss(d) specifies a static gain matrix and is equivalent to usys =
uss([],[],[],d).

Any of these syntaxes can be followed by property name/property value pairs.

usys = uss(a,b,c,d,'P1',V1,'P2',V2,...) set the properties P1, P2, ...

to the values V1, V2, ...

usys = uss(sys) converts an arbitrary ss, tf or zpk model sys to an uncertain
state-space object without uncertainties. Both usys.NominalValue and
simplify(usys,'class') are the same as ss(sys).

Examples

You can first create two uncertain atoms and use them to create two uncertain
matrices. These four matrices can be packed together to form a 1-output,
1-input, 2-state continuous-time uncertain state-space system.
p1 = ureal('p1',5,'Range',[2 6]);
p2 = ureal('p2',3,'Plusminus',0.4);
A = [-p1 0;p2 -p1];

10-347

uss

B = [0;p2];
C = [1 1];
usys = uss(A,B,C,0);

In the second example, you can convert a not-uncertain tf model to an

uncertain state-space model without uncertainties. You can verify the equality
of the nominal value of the usys object and simplified representation to the
original system.
G = tf([1 2 3],[1 2 3 4]);
usys = uss(G)
USS: 3 States, 1 Output, 1 Input, Continuous System
isequal(usys.NominalValue,ss(G))
ans =
1
isequal(simplify(usys,'class'),ss(G))
ans =
1

See Also

10-348

frd
ss

Creates or converts to frequency response data model

Creates or converts to state-space model

uss/ssbal

Purpose

10uss/ssbal

Scale state/uncertainty of uncertain system

Syntax

usysout
usysout
usysout
usysout

Description

usysout = ssbal(usys) yields a system whose input/output and uncertain

properties are the same as usys, a uss object. The numerical conditioning of
usysout is usually better than that of usys, improving the accuracy of
additional computations performed with usysout. usysout is a uss object. The
balancing algorithm uses mussv to balance the constant uncertain state-space
matrices in discrete time. If usys is a continuous-time uncertain system, the

=
=
=
=

ssbal(usys)
ssbal(usys,Wc)
ssbal(usys,Wc,FSflag)
ssbal(usys,Wc,FSflag,BLTflag)

uncertain state-space is mapped by using a bilinear transformation into

discrete time for balancing.
usysout = ssbal(usys,wc) defines the critical frequency wc for the bilinear

prewarp transformation from continuous time to discrete time. The default

value of wc is 1 when the nominal uncertain system is stable and 1.25*mxeig
when it is unstable. mxeig corresponds to the value of the real, most positive
pole of usys.
usysout = ssbal(usys,wc,FSflag) sets the scaling flag FSflag to handle
repeated uncertain parameters. Setting FSflag=1 uses full matrix scalings to
balance the repeated uncertain parameter blocks. FSflag=0, the default, uses
a single, positive scalar to balance the repeated uncertain parameter blocks.
usysout = ssbal(usys,wc,FSflag,BLTflag) sets the bilinear transformation
flag, BLTflag. By default, BLTflag=1 and transforms the continuous-time
system usys to a discrete-time system for balancing. BLTflag=0 results in
balancing the continuous-time state-space data from usys. Note that if usys is

a discrete-time system, no bilinear transformation is performed.

ssbal does not work on an array of uncertain systems. An error message is

generated to alert you to this.

Examples

Consider a two-input, two-output, two-state uncertain system with two real

parameter uncertainties, p1 and p2.

10-349

uss/ssbal

p2=ureal('p2',-17,'Range',[-19 -11]);
p1=ureal('p1',3.2,'Percentage',0.43);
A = [-12 p1;.001 p2];
B = [120 -809;503 24];
C = [.034 .0076; .00019 2];
usys = ss(A,B,C,zeros(2,2))
USS: 2 States, 2 Outputs, 2 Inputs, Continuous System
p1: real, nominal = 3.2, variability = [-0.43 0.43]%, 1
occurrence
p2: real, nominal = -17, range = [-19 -11], 1 occurrence
usys.NominalValue
a =
x1
x2
x1
-12
3.2
x2 0.001
-17
b =
x1
x2

u1
120
503

u2
-809
24

c =
y1
y2

x1
0.034
0.00019

x2
0.0076
2

d =
y1
y2

u1
0
0

u2
0
0

Continuous-time model.
ssbal is used to balance the uncertain system usys.
usysout = ssbal(usys)
USS: 2 States, 2 Outputs, 2 Inputs, Continuous System
p1: real, nominal = 3.2, variability = [-0.43 0.43]%, 1
occurrence
p2: real, nominal = -17, range = [-19 -11], 1 occurrence

10-350

uss/ssbal

usysout.NominalValue
a =
x1
x2
x1
-12
0.3302
x2 0.009692
-17
b =
x1
x2

u1
0.7802
31.7

u2
-5.26
1.512

c =
y1
y2

x1
5.229
0.02922

x2
0.1206
31.74

d =
y1
y2

u1
0
0

u2
0
0

Continuous-time model.

See Also

canon
c2d
d2c
mussv
mussvextract
ss2ss

Forms canonical state-space realizations

Converts continuous-time models to discrete-time
Converts discrete-time models to continuous-time
Sets bounds on the Structure Singular Value ()
Extracts compressed data returned from mussv
Changes state coordinates for state-space models

10-351

usubs

Purpose

10usubs

Substitute values for uncertain elements of uncertain objects

Syntax

B = usubs(M,atomname1,value1,atomname2,value2,...)
B = usubs(M,{atomname1;atomname2;...},{value1;value2;...})
B = usubs(M,StrucArray)

Description

usubs is used to substitute a specific value for an uncertain element of an

uncertain object. The value can itself be uncertain. It needs to be the correct
size, but otherwise can be of any class, and can be an array. Hence, the result
can be of any class. In this manner, uncertain elements act as symbolic
placeholders, for which specific values (which can also contain other
placeholders too) can be substituted.
B = usubs(M,atomname1,value1,atomname2,value2,...) sets the atoms in
M, identified by atomname1, atomname2, etc., to the values in value1, value2,

etc., respectively.
B = usubs(M,ElementName1,value1,ElementName2,value2,...) sets the
elements in M, identified by ElementName1, ElementName2, etc., to the values in
value1, value2, etc. respectively.

The names and values can also be grouped in cell arrays, as

B = usubs(M,atomname1,value1,...)

In this case, if the value cell is 1-by-1, then that value is substituted for all the
listed atoms. For this situation, it is not required that the value be in a cell
array.
value can also be the string 'NominalValue' or 'Random' (or only partially
specified) in which case the nominal value, or a random instance of the atom is
used.

Combinations of the above syntaxes are also allowed, so that

B = usubs(M,{atomname1;atomname2},{value1;value2})

is allowed.
The names and values can also be grouped in a structure, with its field names
constituting the Names, and the field values constituting the Values. In the
following function call, StrucArray is a structure with field names and values.

10-352

usubs

B = usubs(M,StrucArray)

Robustness analysis commands such as wcnorm, wcgain and robuststab

return the offending uncertain element values in this manner. usample, which
randomly samples uncertain objects, also returns the sample points in this
manner.

Example

Create an uncertain matrix and perform identical substitution in two different

manners.
p = ureal('p',5);
m = [1 p;p^2 4];
size(m)
ans =
2
2
m1 = usubs(m,'p',5)
m1 =
1
5
25
4
NamesValues.p = 5;
m2 = usubs(m,NamesValues)
m2 =
1
5
25
4
m1 - m2
ans =
0
0
0
0

You can make an array-valued substitution using the structure-based syntax,

NamesValues.p = rand(1,1,6);
m3 = usubs(m,NamesValues); % 2-by-2-by-6
size(m3)
ans =
2
2
6

You can use usubs to substitute for individual uncertainties. Create three
uncertain real parameters, and form a simple 2-by-2 uncertain matrix with
the parameters
a = ureal('a',5); b = ureal('b',3); c = ureal('c',1);

10-353

usubs

m = [a b;c a*b*c];

You can perform a single parameter substitution and check the results
m1 = usubs(m,'a',10);
simplify(m1(1,1))
ans =
10
simplify(10*m1(1,2)*m1(2,1) - m1(2,2))
ans =
0

You can replace one real parameter with a transfer function, and other
parameters with doubles. You can do this using two different forms of the
syntax and check that the results are identical.
m2 = usubs(m,'a',tf([5],[1 1]),'b',2.6,'c',1.3);
nv.a = tf([5],[1 1]);
nv.b = 2.6;
nv.c = 1.3;
m3 = usubs(m,nv);
norm(m2-m3,'inf')
ans =
0

In m, replace 'a' with 'b', obtaining 'b' directly from m :

m4 = usubs(m,'a',m.Uncertainty.b);

See Also

10-354

gridureal
usample
simplify

Grids uncertain real parameters over their range

Generates random samples of an atom
Simplify representation of uncertain objects

wcgain

Purpose

10wcgain

Calculate bounds on worst-case gain of uncertain system

Syntax

[maxgain,wcu,info] = wcgain(sys)
[maxgain,wcu,info] = wcgain(sys,opts)

Description

The gain of an uncertain system generally depends on the values of its

uncertain elements. Here gain refers to the frequency response magnitude.
Determining the maximum gain over all allowable values of the uncertain
elements is referred to as a worst-case gain analysis. This maximum gain is
called the worst-case gain.
The following figure shows the frequency response magnitude of many samples
of an uncertain system model.

10

Various sample
responses

Magnitude

10

10

Nominal Response

10

10

10

10

10

10

Frequency
wcgain can perform two types of analysis on uncertain systems.

A pointwise-in-frequency worst-case gain analysis yields the

frequency-dependent curve of maximum gain, shown in the following figure.

10-355

wcgain

10

Maximum Gain
pointwise
across frequency

Magnitude

10

WorstCase gain
degradation from nominal

10

10

10

10

10

10

10

Frequency

This plot shows the maximum frequency-response magnitude at each

frequency due to the uncertain elements within the model.
A peak-over-frequency worst-case gain analysis only aims to compute the
largest value of the frequency-response magnitude across all frequencies.
During such an analysis, large frequency ranges can be quickly eliminated
from consideration, thus reducing the computation time.

10-356

wcgain

10

WorstCase Gain, peakacrossfrequency

Magnitude

10

Nominal

10

10

10

10

10

10

10

Frequency

The default analysis performed by wcgain is peak-over-frequency. You can

control which analysis is performed by using the wcgopt options object. For
multi-input, multi-output systems, the gain is the maximum singular value of
the frequency response matrix.
As with other uncertain-system analysis tools, only bounds on the worst-case
gain are computed. The exact value of the worst-case gain is guaranteed to lie
between these upper and lower bounds.
The computation used in wcgain is a frequency-domain calculation. If the input
system sys is an uncertain frequency response object (ufrd), then the analysis
is performed on the frequency grid within the ufrd. If the input system sys is
an uncertain state-space object (uss), then an appropriate frequency grid is
generated (automatically), and the analysis performed on that frequency grid.
In all descriptions below, N denotes the number of points in the frequency grid.

Basic Syntax
Suppose sys is an ufrd or uss with M uncertain elements. Calculate the
worst-case gain of
[maxgain,maxgainunc] = wcgain(sys)

10-357

wcgain

maxgain is a structure with the following fields

Field

Description

LowerBound

Lower bound on worst-case gain, positive scalar.

UpperBound

Upper bound on worst-case gain, positive scalar. If

the nominal value of the uncertain system is
unstable, then maxgain.LowerBound and
maxgain.UpperBound equal .

CriticalFrequency

The critical value of frequency at which maximum

gain occurs (this is associated with
maxgain.LowerBound).

maxgainunc is a structure containing values of uncertain elements that

maximize the system gain. There are M field names, which are the names of
uncertain elements of sys. The value of each field is the corresponding value of
the uncertain element, such that when combined lead to the gain value in
maxgain.LowerBound. The command
norm(usubs(sys,maxgainunc),'inf')

shows the gain.

Examples

Create a plant with nominal model of an integrator, and include additive

unmodeled dynamics uncertainty of a level of 0.4 (this corresponds to 100%
model uncertainty at 2.5 rad/s).
Design a proportional controller K1 that puts the nominal closed-loop
bandwidth at 0.8 rad/s. Roll off K1 at a frequency 25 times the nominal
closed-loop bandwidth. Repeat the design for a controller K2 that puts the
nominal closed-loop bandwidth at 2.0 rad/s. In each case, form the closed-loop
sensitivity function.
P = tf(1,[1 0]) + ultidyn('delta',[1 1],'bound',0.4);
BW1 = 0.8;
K1 = tf(BW1,[1/(25*BW1) 1]);
S1 = feedback(1,P*K1);
BW2 = 2.0;
K2 = tf(BW2,[1/(25*BW2) 1]);

10-358

wcgain

S2 = feedback(1,P*K2);

Assess the worst-case gain of the closed-loop sensitivity function.

[maxgain1,wcunc1] = wcgain(S1);
[maxgain2,wcunc2] = wcgain(S2);
maxgain1
maxgain1 =
LowerBound: 1.5070e+000
UpperBound: 1.5080e+000
CriticalFrequency: 5.3096e+000
maxgain2
maxgain2 =
LowerBound: 5.1024e+000
UpperBound: 5.1034e+000
CriticalFrequency: 1.0215e+001

The maxgain variables indicate that controller K1 achieves better worst-case

performance than K2. Plot Bode magnitude plots of the nominal closed-loop
sensitivity functions, as well as the worst instances, using usubs to replace the
uncertain element with the worst value returned by wcgain.
bodemag(S1.Nom,'r--',usubs(S1,wcunc1),'r',...
S2.Nom,'b--',usubs(S2,wcunc2),'b')

10-359

wcgain

Bode Diagram

10

Magnitude (abs)

10

10

Nominal S1
WorstCase S1
Nominal S2
WorstCase S2
2

10

10

10

10
Frequency (rad/sec)

10

10

Note that although the nominal closed-loop sensitivity resulting from K2 is

superior to that with K1, the worst-case behavior is much worse.

Basic Syntax with Third Output Argument

A third output argument yields more specialized information, including
sensitivities of the worst-case gain to the uncertain elements ranges and
frequency-by-frequency information.
[maxgain,maxgainunc,info] = wcgain(sys)

10-360

wcgain

The third output argument info is a structure with the following fields
Field

Description

Sensitivity

A struct with M fields. Field names are names of

uncertain elements of sys. Values of fields are positive
numbers, each entry indicating the local sensitivity of the
worst-case gain in maxgain.LowerBound to all the
individual uncertain elements uncertainty ranges. For
instance, a value of 25 indicates that if the uncertainty
range is enlarged by 8%, then the worst-case gain should
increase by about 2%. If the Sensitivity property of the
wcgopt object is 'off', the values are NaN.

Frequency

N-by-1 frequency vector associated with analysis.

ArrayIndex

1-by-1 scalar matrix whose value is 1. In more

complicated situations (described later) the value of this

field is dependent on the input data.
Options (e.g., turning on/off the sensitivity computation, setting the step-size
in the sensitivity computation, adjusting the stopping criteria, and controlling
behavior across frequency and array dimensions) can be specified using the
worst-case gain analysis options wcgopt object. For instance, you can turn the
sensitivity calculation off by executing
opt = wcgopt('Sensitivity','off');
[maxgain,maxgainunc,info] = wcgain(sys,opt)

Advanced Options: Pointwise-in-Frequency Calculations

It is also possible to perform the computation pointwise-in-frequency,
determining the worst-case gain at each and every frequency point. To do this,
the wcgopt options object must be used.
opt = wcgopt('FreqPtWise',1);
[maxgain,maxgainunc,info] = wcgain(sys);

10-361

wcgain

Because the calculation is pointwise-in-frequency, many results are N-by-1

cell arrays, often containing scalar information relevant to each particular
frequency. maxgain is a structure with the following fields:
Field

Description

LowerBound

Lower bound on worst-case gain, positive scalar

(frd with N frequency points).

UpperBound

Upper bound on worst-case gain, positive scalar. If

the nominal value of the uncertain system is
unstable, then maxgain.LowerBound and
maxgain.UpperBound equal (frd with N
frequency points).

CriticalFrequency

Scalar. The critical value of frequency at which

maximum gain occurs (this is associated with
norm(maxgain.LowerBound,inf)).

maxgainunc is a N-by-1 cell array of values of uncertain elements that

maximize the system gain. Each entry of the cell array is a struct whose M
field names are the names of uncertain elements of sys. The maximum
singular value of usubs(sys,maxgainunc{k}) at the kth frequency (in
info.Frequency(k)) is equal to maxgain.LowerBound{k}.
info is a structure with the following fields:

10-362

Field

Description

Sensitivity

N-by-1 cell. Each entry is a struct corresponding to the

sensitivities in the worst-case gain at each individual
frequency.

Frequency

N-by-1 frequency vector associated with analysis.

ArrayIndex

N-by-1 cell array. Each value is the 1-by-1 matrix whose

numerical value is 1. In more complicated situations
(described later) the value of this field is dependent on the
input data.

wcgain

Advanced Options: Handling Array Dimensions

If sys has array dimensions, the default behavior is to maximize over all of
these dimensions as well. This can be controlled however, and it is also possible
to perform the computation pointwise-in-the-array-dimensions, determining
the worst-case gain at each and every grid point. Moreover, any combination of
peak-over and pointwise-over is allowed. To specify the desired
computation, you must use wcgopt. For concreteness, suppose that sys is a
rc758 uncertain system (i.e., a 7-by-5-by-8 array of uncertain r output, c
input systems). In order to perform the worst-case gain calculation pointwise
over the 2nd and 3rd array dimensions (the slot with 5 points and the slot with
8 points), set the ArrayDimPtWise property as follows:
opt = wcgopt('ArrayDimPtWise',[2 3]);

In this case, the worst-case gain calculation is performed pointwise on the

5-by-8 grid, but only the peak value over the first array dimension (the slot
with 7 points) is kept track of. For that reason, many of the results are of
dimension 1-by-5-by-8.
In general, suppose that the array dimensions of sys are d1dF (758 in
the above example). Furthermore, assume that the ArrayDimPtWise property
of the wcgopt object is set to some of the integers between 1 and F. Let
e1,e2,,eF denote the dimensions of the array on which the results are
computed. By definition, if j is an integer listed in ArrayDimPtWise, then ej=dj
(all grid points in slot j are computed); otherwise, ej=1 (only the maximum in
slot j is computed). In the above example, with ArrayDimPtWise set to [2 3],
it follows that e1=1,e2=5,e3=8.
Assume FreqPtWise is set to 'off' (you will return to that case later). In this
case, the results of
[maxgain,maxgainunc,info] = wcgain(sys,opt)

10-363

wcgain

are that maxgain is a structure with the following fields:

Field

Description

LowerBound

1-by-1 frd, with array dimensions e1eF, lower

bound on worst-case gain, computed pointwise over
all array dimensions listed in ArrayDimPtWise
property, and peaked over all others.

UpperBound

Upper bound, analogous to LowerBound

CriticalFrequency

e1eF array with the critical value of the

frequency at which maximum gain occurs (this is
associated with maxgain.LowerBound).

maxgainunc is a e1eF struct, containing values of uncertain elements

that maximize the system gain. There are M field names, which are the names
of uncertain elements of sys. The value of each field is the corresponding value
of the uncertain element, such that when combined lead to the gain value in
maxgain.LowerBound. The command norm(usubs(sys,maxgainunc),'inf')
shows the gain, and should be identical to maxgain.LowerBound (to within the
tolerance used in norm).
info is a structure with the following fields:

10-364

Field

Description

Sensitivity

e1eF struct array. Each entry is the local sensitivity of

the worst-case gain in maxgain.LowerBound to all the
individual uncertain elements uncertainty ranges.

Frequency

Nby-1 frequency vector associated with analysis.

ArrayIndex

At each value in the e1eF grid, there is a corresponding

value in the d1dF grid where the maximum occurs.
The variable info.ArrayIndex is an e1eF matrix
whose value is the single-index representation of the
maximizing location in the d1dF grid.

wcgain

Advanced Options:
Array Dimension Handling with FreqPtWise Set to'on'

The final case involves array dimensions and pointwise-in-frequency

calculations. Again, suppose that the array dimensions of sys are d1dF.
Furthermore, assume that the ArrayDimPtWise property of the wcgopt object
is set to some of the integers between 1 and F. Let e1,e2,,eF denote the
dimensions of the array on which the results are computed.
Because the calculation is pointwise-in-frequency, many results are N-by-1 cell
arrays, often containing e1eF arrays in each cell.
maxgain is a structure with the following fields:
Field

Description

LowerBound

N-by-1 cell array, maxgain.LowerBound{k} is a

1-by-1 frd with array dimensions e1eF, and is a
lower bound on worst-case gain at frequency
info.Frequency(k), computed pointwise over all
array dimensions listed in the ArrayDimPtWise
property, and peaked over all others.

UpperBound

Upper bound on worst-case gain, analogous to

maxgain.LowerBound.

CriticalFrequency

e1eF array with the critical value of the

frequency at which maximum gain (pointwise over
all array dimensions listed in ArrayDimPtWise
property, and peaked over all others) occurs.

maxgain.CriticalFrequency e1eF array with the critical value of the

frequency at which maximum gain (pointwise over all array dimensions listed
in the ArrayDimPtWise property, and peaked over all others) occurs.
maxgainunc is a N-by-1 cell array, kth entry is a e1eF struct, containing
values of uncertain elements that maximize the system gain at frequency
info.Frequency(k).

10-365

wcgain

info is a structure with the following fields:

Field

Description

Sensitivity

N-by-1 cell. Each entry is the e1eF struct array

which holds the local sensitivity of the worst-case gain at

one frequency to all of the individual uncertain elements
uncertainty ranges.
Frequency

N-by-1 frequency vector associated with analysis.

ArrayIndex

N-by-1 cell array, kth e1eF, at each value in the

e1eF grid, there is a corresponding value in the
d1dF grid where the maximum occurs. The variable
info.ArrayIndex is a e1eF matrix whose value is the
single-index representation of the maximizing location in
the d1dF grid.

Behavior on Not-Uncertain Systems

wcgain can also be used on not-uncertain systems (e.g., ss and frd). If sys is
a single ss or frd, then the worst-case gain is simply the gain of the system
(identical to norm(sys,'inf')). However, if sys has array dimensions, then

the possible combinations of peak-over and pointwise-over can be used to

customize the computation.

Algorithm

The worst-case gain is guaranteed to be at least as large as LowerBound (some

value of allowable uncertain elements yield this gain one instance is returned
in the structure maxgainunc. The frequency at which the gain in LowerBound
occurs is in CriticalFrequency. Lower bounds for wcgain are computed using
a power iteration on ultidyn, ucomplex and ucomplexm uncertain atoms,
(holding uncertain real parameters fixed) and a coordinate aligned search on
the uncertain real parameters (while holding the complex blocks fixed).
Similarly, the worst-case gain is guaranteed to be no larger than UpperBound.
In other words, for all allowable modeled uncertainty, the gain is provably less
than or equal to UpperBound. These bounds are derived using the upper bound
for the structured singular value, which is essentially optimally-scaled,
small-gain theorem analysis. Upper bounds are obtained by solving a

10-366

wcgain

semidefinite program. wcgain uses branch and bound on the uncertain real
parameters to tighten the lower and upper bounds.

Limitations

Because the calculation is carried out with a frequency grid, it is possible

(likely) that the true critical frequency is missing from the frequency vector
used in the analysis. This is similar to the problem in robuststab. However,
compared with robuststab, the problem in wcgain is less acute. Thought of as
a function of problem data and frequency, the worst-case gain is a continuous
function (unlike the robust stability margin, which in special cases is not; see
Regularizing Robust Stability calculations with only ureal uncertain
elements in the online documentation). Hence, in worst-case gain calculations,
increasing the density of the frequency grid will always increase the accuracy
of the answers and in the limit, answers arbitrarily close to the actual answers
are obtainable with finite frequency grids.

See Also

loopmargin
mussv
norm
robuststab
wcgopt
wcsens
wcmargin

Comprehensive analysis of feedback loops

Calculate bounds on the Structured Singular Value ()
System norm of an LTI object
Calculates stability margins of uncertain systems
Creates a wcgain options object
Calculates worst-case sensitivities for a feedback loop
Calculates worst-case margins for a feedback loop

10-367

wcgopt

Purpose

10wcgopt

Create options object for use with wcgain, wcsens, and wcmargin

Syntax

options = wcgopt
options = wcgopt('name1',value1,'name2',value2,...)
wcgopt

Description

options = wcgopt (with no input arguments) creates an options object with all
the properties set to their default values.
options = wcgopt('name1',value1,'name2',value2,...) creates a wcgain,
wcsens and wcmargin options object called options in which specified properties

have specific values. Any unspecified property is set to its default value. It is
sufficient to type only enough leading characters to define the property name
uniquely. Case is ignored for property names.
wcgopt with no input or output arguments displays a complete list of option

properties and their default values.

The following are the wcgopt object properties:

10-368

Object Property

Description

Sensitivity

Computes margin sensitivity to individual

uncertainties. Default is 'on'

LowerBoundOnly

If LowerBoundOnly is 'on', then only the lower bound

computation is performed. The default value is
'off', which implies that both upper and lower
bounds for worst-case gain are computed.

FreqPtWise

Apply stopping criteria based on upper/lower bounds

(described below) at every frequency point (as
opposed to just the peak value). FreqPtWise=1
activates the pointwise criteria. In order to only
compute the peak value to within tolerance, use 0.
Default = 0.

wcgopt

Object Property

Description

ArrayDimPtWise

Relevant for uss/ufrd/ss/frd arrays. For indices

specified in ArrayDimPtWise, the stopping criteria
based on upper/lower bounds (described below) is
used at every point in array dimensions specified in
ArrayDimPtWise, being applied to the peak value
over all other array dimensions. Default = [].
If FreqPtWise==1, the computation terminates when
at least one of the following four conditions is true at
every frequency:
UpperBound - LowerBound <= AbsTol
UpperBound - LowerBound <= Reltol*UpperBound
UpperBound <= AGoodThreshold +
MGoodThreshold*Norm(NominalValue)

LowerBound >= ABadThreshold +

MBadThreshold*Norm(NominalValue)

If FreqPtWise==0, the computation terminates when

any one of the following four conditions is true:
PeakUpperBound - PeakLowerBound <= AbsTol
PeakUpperBound - PeakLowerBound <=
Reltol*PeakUpperBound at every frequency
UpperBound <= AGoodThreshold +
MGoodThreshold*Norm(NominalValue)) at some

frequency
LowerBound >= ABadThreshold +
MBadThreshold*Norm(NominalValue)

In both situations above, the stopping condition is

applied at every point in array dimensions specified
in ArrayDimPtWise. UpperBound and LowerBound are
the peak values over all other array dimensions.
Default

Structure with default values of all wcgopt

properties.

10-369

wcgopt

10-370

Object Property

Description

Meaning

Structure. Field names are wcgopt properties, and

values are the text description of property.

VaryUncertainty

Percentage variation of uncertainty used as a step

size in finite-difference calculations to estimate
sensitivity. Default is 25.

AbsTol

Upper and lower absolute stopping tolerance.

Default=0.02.

RelTol

Upper and lower absolute stopping tolerance.

Default = 0.02.

AbsTol

Upper and lower relative stopping tolerance.

Default=0.05.

MGoodThreshold

Multiplicative (UpperBound) stopping threshold.

Default = 1.04.

AGoodThreshold

Additive (UpperBound) stopping threshold. Default =

0.05.

MBadThreshold

Multiplicative (LowerBound) stopping threshold,

Default = 5.

AGoodThreshold

Additive (LowerBound) stopping threshold. Default =

20.

NTimes

Number of restarts in lower bound search (positive

integer).

MaxCnt

Number of cycles in lower bound search (positive

integer). Default = 3.

MaxTime

Maximum computation time allowed (in seconds).

The computation is prematurely terminated if this
much real time elapses before the computation is
complete. All quantities that have been computed are
returned. Default = 720.

wcgopt

Example

You can create a wcgopt options object called opt with all default values.
opt = wcgopt
Property Object Values:
Sensitivity: 'on'
LowerBoundOnly: 'off'
FreqPtWise: 0
ArrayDimPtWise: []
VaryUncertainty: 25
Default: [1x1 struct]
Meaning: [1x1 struct]
AbsTol: 0.0200
RelTol: 0.0500
MGoodThreshold: 1.0400
AGoodThreshold: 0.0500
MBadThreshold: 20
ABadThreshold: 5
NTimes: 2
MaxCnt: 3
MaxTime: 720

The following statements change the absolute tolerance stopping criterion from
0.02 to 0.04 and the point wise over frequency test from the peak worst-case
value, opt.FreqPtWise=0, to the worst-case value at every frequency.
opt = wcgopt;
opt.AbsTol = 0.04;
opt.FreqPtWise = 1;
opt
Property Object Values:
Sensitivity: 'on'
LowerBoundOnly: 'off'
FreqPtWise: 1
ArrayDimPtWise: []
Default: [1x1 struct]
Meaning: [1x1 struct]
VaryUncertainty: 25
AbsTol: 0.0400
RelTol: 0.0500
MGoodThreshold: 1.0400
AGoodThreshold: 0.0500

10-371

wcgopt

MBadThreshold:
ABadThreshold:
NTimes:
MaxCnt:
MaxTime:

20
5
2
3
720

This statement makes a single call to wcgopt to set the maximum computation
time to 10000 seconds and disables the Sensitivity calculation.
opt = wcgopt('MaxTime',10000,'Sensitivity','off');

See Also

10-372

dkitopt
robopt
wcgain
wcnorm
wcsens
wcmargin

Creates a dksyn options object

Creates a robustperf/robuststab options object
Calculates worst-case gain of a system
Calculates worst-case gain of a matrix
Calculates worst-case sensitivities for a feedback loop
Calculates worst-case margins for a feedback loop

wcmargin

Purpose

10wcmargin

Calculate worst-case gain/phase margins for feedback loop

Syntax

wcmargi = wcmargin(p,c)
[wcmargi,wcmargo] = wcmargin(p,c)
wcmargi = wcmargin(p,c,opt)
[wcmargi,wcmargo] = wcmargin(p,c,opt)

Description

Classical gain and phase margins define the loop-at-a-time allowable,

independent variations in the nominal system gain and phase for which the
feedback loop retains stability. An alternative to classical gain and phase
margins is the disk margin. The disk margin calculates the largest region for
each channel such that for all gain and phase variations inside the region the
nominal closed-loop system is stable. The guaranteed bound is calculated based
on the balanced sensitivity function. See the dmplot and loopmargin
Algorithm sections for more information.
Consider a system with uncertain elements. It is of interest to determine the
gain and phase margins of each individual channel in the presence of
uncertainty. These margins are called worst-case margins. Worst-case margin,
wcmargin calculates the largest disk margin such that for values of the
uncertainty and all gain and phase variations inside the disk, the closed-loop
system is stable. The worst-case gain and phase margin bounds are defined
based on the balanced sensitivity function. Hence, results from the worst-case
margin calculation imply that the closed-loop system is stable for a given
uncertainty set and would remain stable in the presence of an additional gain
and phase margin variation in the specified input/output channel.
[wcmargi,wcmargo] = wcmargin(L) calculates the combined worst-case input
and output loop-at-a-time gain/phase margins of the feedback loop consisting
of the loop transfer matrix L in negative feedback with an identity matrix. L
must be an uncertain system, uss or ufrd object. If L is a uss object, the
frequency range and number of points used to calculate wcmargi and wcmargo
are chosen automatically. Note that in this case, the worst-case margins at the
input and output are equal because an identity matrix is used in feedback.
[wcmargi,wcmargo] = wcmargin(P,C) calculates the combined worst-case
input and output loop-at-a-time gain/phase margins of the feedback loop
consisting of C in negative feedback with P. C should only be the compensator
in the feedback path, without reference channels, if it is a 2-DOF architecture.

10-373

wcmargin

That is, if the closed-loop system has a 2-DOF architecture the reference
channel of the controller should be eliminated resulting in a 1-DOF
architecture as shown in the following figure. Either P or C must be an
uncertain system, uss or ufrd, or an uncertain matrix, umat. If P and C are
ss/tf/zpk or uss objects, the frequency range and number of points used to
calculate wcmargi and wcmargo are chosen automatically.

 

 
   

 
   

Basic Syntax
[wcmargi,wcmargo] = wcmargin(L)
[wcmargi,wcmargo] = wcmargin(P,C)
wcmargi and wcmargo are structures corresponding to the loop-at-a-time

worst-case, single-loop gain and phase margin of the channel. For the
single-loop transfer matrix L of size N-by-N, wcmargi is a N-by-1 structure. For
the case with two input arguments, the plant model P will have NY outputs and
NU inputs and hence the controller C must have NU outputs and NY inputs.
wcmargi is a NU-by-1 structure with the following fields:

10-374

Field

Description

GainMargin

Guaranteed bound on worst-case, single-loop gain

margin at plant inputs. Loop-at-a-time analysis.

PhaseMargin

Loop-at-a-time worst-case phase margin at plant inputs.

Units are degrees.

wcmargin

Field

Description

Frequency

Frequency associated with the worst-case margin

(rad/s).

Sensitivity

Struct with M fields. Field names are names of

uncertain elements of P and C. Values of fields are
positive numbers, which each entry indicating the local
sensitivity of the worst-case margins to all the
individual uncertain elements uncertainty ranges. For
instance, a value of 50 indicates that if the uncertainty
range is enlarged by 8%, then the worst-case gain
should increase by about 4%. If the Sensitivity
property of the wcgopt object is 'off', the values are
NaN.

wcmargo is an N-by-1 structure for the single loop transfer matrix input and
wcmargo is an NY-by-1 structure when the plant and controller are input. In
both these cases, wcmargo has the same fields as wcmargi. The worst-case

bound on the gain and phase margins are calculated based on a balanced
sensitivity function.
[wcmargi,wcmargo] = wcmargin(L,opt) and
[wcmargi,wcmargo] = wcmargin(p,c,opt) specify options described in opt.
(See wcgopt for more details on the options for wcmargin.)

The sensitivity of the worst-case margin calculations to the individual

uncertain elements is selected using the options object opt. To compute
sensitivities, create a wcgopt options object, and set the Sensitivity property
to 'on'.
opt = wcgopt('Sensitivity','on');
[maxgain,maxgainunc,info] = wcgain(sys,opt)

Examples

MIMO Loop-at-a-Time Margins

This example is designed to illustrate that loop-at-a-time margins (gain, phase,
and/or distance to 1) can be inaccurate measures of multivariable robustness
margins. Margins of the individual loops can be very sensitive to small
perturbations within other loops.

10-375

wcmargin

The nominal closed-loop system considered here is shown as follows

; 6 d- K
;6

G and K are 2-by-2 multi-input/multi-output (MIMO) systems, defined as

2
1
(s + 1) , K = I
G := ------------------ s
2
2
2
s + ( s + 1 ) s 2

Set := 10, construct the nominal model G in state-space form, and compute
its frequency response.
a
b
c
d
G
K

=
=
=
=
=
=

[0 10;-10 0];
eye(2);
[1 8;-10 1];
zeros(2,2);
ss(a,b,c,d);
[1 -2;0 1];

The nominal plant was analyzed previously using the loopmargin command.
Based on experimental data, the gain of the first input channel, b(1,1), is
found to vary between 0.97 and 1.06. The following statement generates the
updated uncertain model.
ingain1 = ureal('ingain1',1,'Range',[0.97 1.06]);
b = [ingain1 0;0 1];
Gunc = ss(a,b,c,d);

Because of differences between measured data and the plant model an 8%

unmodeled dynamic uncertainty is added to the plant outputs.
unmod = ultidyn('unmod',[2 2],'Bound',0.08);
Gmod = (eye(2)+unmod)*Gunc;
Gmodg = ufrd(Gmod,logspace(-1,3,60));

10-376

wcmargin

You can use the command wcmargin to determine the worst-case gain and
phase margins in the presences of the uncertainty.
[wcmi,wcmo] = wcmargin(Gmodg,K);

The worst-case analysis corresponds to maximum allowable disk margin for all
possible defined uncertainty ranges. The worst-case single-loop margin
analysis performed using wcmargin results in a maximum allowable gain
margin variation of 1.31 and phase margin variations of 15.6 degs in the
second input channel in the presence of the uncertainties 'unmod' and
'ingain1'. wcmi(1)
ans =
GainMargin: [0.3613 2.7681]
PhaseMargin: [-50.2745 50.2745]
Frequency: 0.1000
Sensitivity: [1x1 struct]
wcmi(2)
ans =
GainMargin: [0.7585 1.3185]
PhaseMargin: [-15.6426 15.6426]
Frequency: 0.1000
Sensitivity: [1x1 struct]

Hence even though the second channel had infinite gain margin and 90 degrees
of phase margin, allowing variation in both uncertainties, 'unmod' and
'ingain1' leads to a dramatic reduction in the gain and phase margin. See the
loopmargin command page example for more details
You can display the sensitivity of the worst-case margin in the second input
channel to 'unmod' and 'ingain1' as follows:
wcmi(2).Sensitivity
ans =
ingain1: 12.1865
unmod: 290.4557

The results indicate that the worst-case margins are not very sensitive to the
gain variation in the first input channel, 'ingain1', but very sensitive to the
LTI dynamic uncertainty at the output of the plant.

10-377

wcmargin

The worst-case single-loop margin at the output results in a maximum

allowable gain margin variation of 1.46 and phase margin variation of 21.3
degs in the second output channel in the presence of the uncertainties 'unmod'
and 'ingain1'.
wcmo(1)
ans =
GainMargin: [0.2521 3.9664]
PhaseMargin: [-61.6995 61.6995]
Frequency: 0.1000
Sensitivity: [1x1 struct]
wcmo(2)
ans =
GainMargin: [0.6835 1.4632]
PhaseMargin: [-21.2984 21.2984]
Frequency: 0.1000
Sensitivity: [1x1 struct]

You can display the sensitivity of the worst-case margin in the second output
channel to 'unmod' and 'ingain1' as follows:
wcmo(2).Sensitivity
ans =
ingain1: 16.3435
unmod: 392.1320

The results are similar to the worst-case margins at the input. However, the
worst-case margins at the second output channel are even more sensitive to the
LTI dynamic uncertainty than the input channel margins.

See Also

10-378

dmplot
loopsens
loopmargin
robuststab
usubs
wcgain
wcgopt
wcsens

Interprets disk gain and phase margins

Calculates sensitivity functions of feedback loops
Performs comprehensive analysis of feedback loops
Calculates stability margins of uncertain systems
Substitutes values for uncertain atoms
Calculates worst-case gain of a system
Creates a worst-case options object
Calculates worst-case sensitivity functions

wcnorm

Purpose

10wcnorm

Calculate worst-case norm of uncertain matrix

Syntax

maxnorm = wcnorm(m)
[maxnorm,wcu] = wcnorm(m)
[maxnorm,wcu] = wcnorm(m,opts)
[maxnorm,wcu,info] = wcnorm(m)
[maxnorm,wcu,info] = wcnorm(m,opts)

Description

The norm of an uncertain matrix generally depends on the values of its

uncertain elements. Determining the maximum norm over all allowable values
of the uncertain elements is referred to as a worst-case norm analysis. The
maximum norm is called the worst-case norm.
As with other uncertain-system analysis tools, only bounds on the worst-case
norm are computed. The exact value of the worst-case norm is guaranteed to
lie between these upper and lower bounds.

Basic syntax
Suppose mat is a umat or a uss with M uncertain elements. The results of
[maxnorm,maxnormunc] = wcnorm(mat)
maxnorm is a structure with the following fields
Field

Description

LowerBound

Lower bound on worst-case norm, positive scalar.

UpperBound

Upper bound on worst-case norm, positive scalar.

maxnormunc is a structure that includes values of uncertain elements and

maximizes the matrix norm. There are M fieldnames, which are the names of
uncertain elements of mat. The value of each field is the corresponding value of
the uncertain element, such that when jointly combined, lead to the norm value
in maxnorm.LowerBound. The following command shows the norm:
norm(usubs(mat,maxnormunc))

10-379

wcnorm
Basic syntax with third output argument
A third output argument provides information about sensitivities of the
worst-case norm to the uncertain elements ranges.
[maxnorm,maxnormunc,info] = wcgain(mat)

The third output argument info is a structure with the following fields:
Field

Description

Sensitivity

A struct with M fields. Fieldnames are names of

uncertain elements of sys. Field values are positive
numbers, each entry indicating the local sensitivity of the
worst-case norm in maxnorm.LowerBound to all of the
individual uncertain elements uncertainty ranges. For
instance, a value of 25 indicates that if the uncertainty
range is increased by 8%, then the worst-case norm
should increase by about 2%. If the Sensitivity
property of the wcgopt object is 'off', the values are NaN.

ArrayIndex

1-by-1 scalar matrix with the value of 1. In more

complicated situations (described later) the value of this
field depends on the input data.

Advanced options: Handling array dimensions

If mat has array dimensions, the default behavior is to maximize over all
dimensions. It is also possible to perform the computation
pointwise-in-the-array-dimensions to determine the worst-case norm at each
grid point. Any combination of peak-over and pointwise-over is allowed.
To specify the desired computation, the wcgopt must be used. For concreteness,
suppose that mat is an rc758 uncertain system (i.e., a 7-by-5-by-8 array of
uncertain r output, c input systems). To perform the worst-case gain
calculation pointwise over the second and third array dimensions (the slots
with 5 points and 8 points, respectively), set the ArrayDimPtWise property:
opt = wcgopt('ArrayDimPtWise',[2 3]);

In this case, the worst-case norm calculation is performed pointwise on the

5-by-8 grid. Only the peak value in the first array dimension (the slot with 7

10-380

wcnorm

points) is tracked. For that reason, many of the results will be of dimension
1-by-5-by-8.
In general, suppose that the array dimensions of sys are d1dF (758 in
the above example). Furthermore, assume that the ArrayDimPtWise property
of the wcgopt object has been set to some of the integers between 1 and F. Let
e1,e2,,eF denote the dimensions of the array on which the results are
computed. By definition, if j is an integer listed in ArrayDimPtWise, then ej=dj
(all grid points in slot j are computed), otherwise ej=1 (only the maximum in
slot j is computed). In the above example, with ArrayDimPtWise set to [2 3],
it follows that e1=1, e2=5, e3=8.
In this case, the following command
[maxgain,maxgainunc,info] = wcgain(sys,opt)

produces the maxgain a structure with the following fields

Field

Description

LowerBound

e1eF matrix of lower bounds on worst-case

norm, computed pointwise over all array
dimensions listed in ArrayDimPtWise property and
peaked over all others.

UpperBound

Upper bound, analogous to LowerBound

maxgainunc is a e1eF struct, containing values of uncertain elements

which maximize the system norm. There are M fieldnames, which are the
names of uncertain elements of mat. The value of each field is the corresponding
value of the uncertain element, which lead to the gain value in
maxnorm.LowerBound when jointly combined.

10-381

wcnorm

info is a structure with the following fields

Examples

Field

Description

Sensitivity

e1eF struct array, where each entry is the local

sensitivity of the worst-case norm in maxnorm.LowerBound
to the uncertainty range of each uncertain element.

ArrayIndex

At each value in the e1eF grid, there is a corresponding

value in the d1dF grid where the maximum occurs.
The variable info.ArrayIndex is an e1eF matrix,
where the value is the single-index representation of the
maximizing location in the d1dF grid.

You can construct an uncertain matrix and compute the worst-case norm of the
matrix, as well as its inverse. Your objective is to accurately estimate the
worst-case, or the largest, value of the condition number of the matrix.
a=ureal('a',5,'Range',[4
b=ureal('b',2,'Range',[1
b=ureal('b',3,'Range',[2
c=ureal('c',9,'Range',[8
d=ureal('d',1,'Range',[0
M = [a b;c d];
Mi = inv(M);
[maxnormM] = wcnorm(M)
maxnormM =
LowerBound: 14.7199
UpperBound: 14.7327
[maxnormMi] = wcnorm(Mi)
maxnormMi =
LowerBound: 2.5963
UpperBound: 2.5979

6]);
3]);
10]);
11]);
2]);

The condition number of M must be less than the product of the two upper
bounds for all values of the uncertain elements making up M. Conversely, the
largest value of M condition number must be at least equal to the condition
number of the nominal value of M. Compute these crude bounds on the
worst-case value of the condition number.

10-382

wcnorm

condUpperBound = maxnormM.UpperBound*maxnormMi.UpperBound;
condLowerBound = cond(M.NominalValue);
[condLowerBound condUpperBound]
ans =
5.0757
38.2743

How can you get a more accurate estimate? Recall that the condition number
of an nxm matrix M can be expressed as an optimization, where a free
norm-bounded matrix tries to align the gains of M and M1
( M ) = max

m m

C
max ( ) 1

( max ( MM ) )

If M is itself uncertain, then the worst-case condition number involves further

maximization over the possible values of M. Therefore, you can compute the
worst-case condition number of an uncertain matrix by using a ucomplexm
uncertain element, and then by using wcnorm to carry out the maximization.
Create a 2-by-2 ucomplexm object, with nominal value equal to zero.
Delta = ucomplexm('Delta',zeros(2,2));

The range of values represented by Delta includes 2-by-2 matrices with the
maximum singular value less than or equal to 1.
You can create the expression involving M, Delta and inv(M).
H = M*Delta*Mi;

Finally, consider the stopping criteria and call wcnorm. One stopping criteria
for wcnorm(H) is based on the norm of the nominal value of H. During the
computation, if wcnorm determines that the worst-case norm is at least
ABadThreshold + MBadThreshold*norm ( H.NominalValue )

then the calculation is terminated. In our case, since H.NominalValue equals 0,

the stopping criteria is governed by ABadThreshold. The default value of
ABadThreshold is 5. To keep wcnorm from prematurely stopping, set
ABadThreshold to 38 (based on our crude upper bound above).
opt = wcgopt('ABadThreshold',38);
[maxKappa,wcu,info] = wcnorm(H,opt);

10-383

wcnorm

maxKappa
maxKappa =
LowerBound: 26.9629
UpperBound: 27.9926

You can verify that wcu makes the condition number as large as
maxKappa.LowerBound.
cond(usubs(M,wcu))
ans =
26.9629

Algorithm

See wcgain

See Also

lti/norm
svd
wcgain
wcgopt

10-384

Calculates LTI system norms

Calculates singular value decomposition
Calculates worst-case gain of a system
Creates a wcgain options object

wcsens

Purpose

10wcsens

Calculate worst-case sensitivity and complementary sensitivity functions of

feedback loop

Syntax

wcst
wcst
wcst
wcst
wcst
wcst
wcst
wcst
wcst
wcst

Description

The sensitivity function, S=(I+L)1, and the complementary sensitivity

function, T=L(I+L)1, where L is the loop gain matrix associated with the input,
CP, or output, PC, are two transfer functions related to the robustness and
performance of the closed-loop system. The multivariable closed-loop
interconnection structure, shown below, defines the input/output sensitivity,
complementary sensitivity and loop transfer functions.

=
=
=
=
=
=
=
=
=
=

wcsens(L)
wcsens(L,type)
wcsens(L,opt)
wcsens(L,type,scaling)
wcsens(L,type,scaling,opt)
wcsens(P,C)
wcsens(P,C,type)
wcsens(P,C,opt)
wcsens(P,C,type,scaling)
wcsens(P,C,type,scaling,opt)










Description

Equation

Si: Input Sensitivity

(I+CP)1

Ti: Input Complementary Sensitivity

CP(I+CP)1

So: Output Sensitivity

(I+PC)1

10-385

wcsens

Description

Equation

To: Output Complementary Sensitivity

PC(I+PC)1

Li: Input Loop Transfer Function

CP

Lo: Output Loop Transfer Function

PC

wcst = wcsens(L) calculates the worst-case sensitivity and complementary

sensitivity functions for the loop transfer matrix L in feedback in negative
feedback with an identity matrix. If L is a uss object, the frequency range
and number of points are chosen automatically.
wcst = wcsens(P,C) calculates the worst-case sensitivity and complementary
sensitivity functions for the feedback loop C in negative feedback with P. C

should only be the compensator in the feedback path, not any reference
channels, if it is a 2-dof architecture (see loopsens). If P and C are ss/tf/zpk
or uss objects, the frequency range and number of points are chosen
automatically. wcst is a structure with the following substructures:
Table 10-1: Fields of wcst

10-386

Field

Description

Si

Worst-case input-to-plant sensitivity function

Ti

Worst-case input-to-plant complementary sensitivity

function

So

Worst-case output-to-plant sensitivity function

To

Worst-case output-to-plant complementary sensitivity

function

PSi

Worst-case plant times input-to-plant sensitivity

function

CSo

Worst-case compensator times output-to-plant

sensitivity function

Stable

1 if nominal closed loop is stable, 0 otherwise. NaN for

frd/ufrd objects.

wcsens

Each sensitivity substructure is a structures with five fields MaximumGain,

BadUncertainValues, System, BadSystem, Sensitivity derived from the
outputs of wcgain.
Table 10-2: Fields of Si, So, Ti, To, PSi, CSo
Field

Description

MaximumGain

struct with fields LowerBound, UpperBound and

CriticalFrequency. LowerBound and UpperBound
are bounds on the unweighted maximum gain of
the uncertain sensitivity function.
CriticalFrequency is the frequency at which the
maximum gain occurs.

BadUncertainValues

Struct, containing values of uncertain elements

which maximize the sensitivity gain. There are M

fluidness, which are the names of uncertain

elements of sensitivity function. The value of each
field is the corresponding value of the uncertain
element, such that when jointly combined, lead to
the gain value in MaximumGain.LowerBound.
System

Uncertain sensitivity function (ufrd or uss).

BadSystem

Worst-case system based on the uncertain object

values in BadUncertainValues. BadSystem is
defined as BadSystem=usubs(System,
BadUncertainValues).

Sensitivity

Struct with M fields, fieldnames are names of

uncertain elements of system. Values of fields are

positive numbers, each entry indicating the local
sensitivity of the maximum gain to all of the
individual uncertain elements uncertainty
ranges. For instance, a value of 50 indicates that
if the uncertainty range is enlarged by 8%, then
the maximum gain should increase by about 4%.
If the 'Sensitivity' property of the wcgopt
object is 'off', the values are NaN.

10-387

wcsens

wcst = wcsens(L,type) and wcst = wcsens(P,C,type) allows selection of

individual Sensitivity and Complementary Sensitivity functions, type, as
'Si','Ti','So','To','PSi','CSo' corresponding to the sensitivity and
complementary sensitivity functions. Setting type to 'S' or 'T' selects all
sensitivity functions ('Si','So','PSi','CSo') or all complementary
sensitivity functions ('Ti','To'). Similarly, setting type to 'Input' or
'Output' selects all input Sensitivity functions ('Si','Ti','PSi') or all
output sensitivity functions ('So,'To','CSo'). 'All' selects all six Sensitivity
functions for analysis (default). type may also be a cell containing a collection
of strings, i.e. 'Si','To', as well as a comma separated list.
wcst = wcsens(L,type,scaling) and wcst = wcsens(P,C,type,scaling)
adds a scaling to the worst-case sensitivity analysis. scaling is either the
character strings 'Absolute' (default), 'Relative' or a ss/tf/zpk/frd
object. The default scaling 'Absolute' calculates bounds on the maximum gain
of the uncertain sensitivity function. The 'Relative' scaling finds bounds on

the maximum relative gain of the uncertain sensitivity function. That is, the
maximum relative gain is the largest ratio of the worst-case gain and the
nominal gain evaluated at each frequency point in the analysis, Similarly if
scaling is a ss/tf/zpk/frd object, bounds on the maximum scaled gain of the
uncertain sensitivity function are found. If scaling is 'Relative'or a
ss/tf/zpk/frd object, the worst-case analysis peaks over frequency. If
scaling is an object, its input/output dimensions should be 1-by-1 or
dimensions compatible with P and C. type and scaling can also be combined
in a cell array, e.g.
wcst = wcsens(P,C,{'Ti','So'},'Abs','Si','Rel','PSi',wt)
wcst = wcsens(P,C,opt) or wcst = wcsens(P,C,type,scaling,opt)

specifies options for the worst-case gain calculation as defined by opt. (See
wcgopt for more details on the options for wcsens.)
The sensitivity of the worst-case sensitivity calculations to the individual
uncertain components can be determined using the options object opt. To
compute the sensitivities to the individual uncertain components, create a
wcgopt options object, and set the Sensitivity property to 'on'.
opt = wcgopt('Sensitivity','on');
wcst = wcsens(P,C,opt)

10-388

wcsens

Examples

The following constructs a feedback loop with a first order plant and a
proportional-integral controller. The time constant is uncertain and the model
also includes an multiplicative uncertainty. The nominal (input) sensitivity
function has a peak of 1.09 at omega = 1.55 rad/sec. Since the plant and
controller are single-input / single-output, the input/output sensitivity
functions are the same.
delta = ultidyn('delta',[1 1]);
tau = ureal('tau',5,'range',[4 6]);
P = tf(1,[tau 1])*(1+0.25*delta);
C=tf([4 4],[1 0]);
looptransfer = loopsens(P,C);
Snom = looptransfer.Si.NominalValue;
norm(Snom,inf)
ans =
1.0864
wcsens is then used to compute the worst-case sensitivity function as the

uncertainty ranges over its possible values. More information about the fields
in wcst.Si can be found in the wcgain help. The badsystem field of wcst.Si
contains the worst case sensitivity function. This worst case sensitivity has a
peak of 1.52 at omega = 1.02 rad/sec. The maxgainunc field of wcst.Si contains
the perturbation that corresponds to this worst case sensitivity function.
wcst = wcsens(P,C)
wcst =
Si: [1x1 struct]
Ti: [1x1 struct]
So: [1x1 struct]
To: [1x1 struct]
PSi: [1x1 struct]
CSo: [1x1 struct]
Stable: 1
Swc = wcst.Si.BadSystem;
omega = logspace(-1,1,50);
bodemag(Snom,'-',Swc,'-.',omega);
legend('Nominal Sensitivity','Worst-Case Sensitivity',...
'Location','SouthEast')
norm(Swc,inf)
ans =
1.5075

10-389

wcsens

For multi-input/multi-output systems the various input/output sensitivity

functions will, in general, be different.

References

J. Shin, G.J. Balas, and A.K. Packard, Worst case analysis of the X-38 crew
return vehicle flight control system, AIAA Journal of Guidance, Dynamics and
Control, vol. 24, no. 2, March-April 2001, pp. 261-269.

See Also

loopsens
loopmargin
robuststab
usubs
wcgain
wcgopt
wcmargin

10-390

Calculate sensitivity functions of feedback loops

Comprehensive analysis of feedback loops
Calculate stability margins of uncertain systems
Substitutes values for uncertain atoms
Calculate worst-case gain of a system
Create a worst-case options object
Calculate worst-case margins for feedback loop

wcsens

10-391

wcsens

10-392

Index
A
ACC Benchmark plant 10-29
actmod actuator model 5-20
actnom nominal actuator model 5-20
additive error 10-113
hankelmr 10-113

additive error bound 3-4

all-pass phase matrix 10-34
atoms 4-2
augmented plant 10-20

C
Chiang, R. Y.
bilinear pole shifting 10-31
cmsclsyn 10-37

complementary sensitivity 2-5

complementary sensitivity T 10-20
conic-sector 10-301
crossover frequency c 2-16

Delta1 object 4-13

Balanced model truncation 10-23

balancmr 10-23
Schur method 10-285
schurmr 10-285
square root method 10-23
Balanced stochastic truncation 10-32
BST 10-32
bstmr 10-32

balancmr 10-23
balancmr 3-6
Bamieh, B.A. 10-293
bilinear transform, frequency
continuous to continuous
pole-shifting transform 10-29
continuous to discrete
backward rectangular 10-27
forward rectangular 10-27
shifted Tustin 10-28
general bilinear 10-28
reverse transform 10-27
bisection algorithm 10-128
block diagrams
direction of arrows 5-30
bstmr 3-9

Delta2 object 4-13

design goals
crossover 2-16
performance 2-16
roll-off 2-16
stability robustness 2-16
disturbance attenuation 2-6
D-K iteration control design commands
summary 5-27
dksyn 5-22
Doyle, J. C.
state-space H 10-112
D-scalings
automatic prefitting 10-61

E
Euclidean norms 5-33

F
feedback 4-8
fitmag 10-95
fitmaglp 10-95

Index-1

Index

forbidden regions 2-9

Franklin, G. F. 10-31
frequency domain uncertainty
adding to the model 4-13
fundamental limits
right-half-plane poles 2-16
right-half-plane zeros 2-16

mixed-sensitivity 10-8, 10-20

H2 and H control design commands
summary 5-26
h2syn 10-107
Hankel Minimum Degree Approximation 10-113
Hankel minimum degree approximation
hankelmr 10-113
MDA 10-6, 10-113
Zeroth Order Hankel MDA 10-116
Hankel singular value

gain reduction tolerance 2-10

gain/phase margins
MIMO system 2-9
gap metric 10-234

bstmr 10-32
hankelmr 10-113

hankelsv 10-120
NCF 1-15

genphase 10-95
get

viewing the properties of the uncertain system

4-12
-iteration 10-128
Glover, K.
loop shaping 10-188
state-space H2 10-112

schurmr 10-285

Hankel singular value based model reduction

reduce 10-259
HiMAT aircraft model 1-10
hinfsyn 5-13

I
InputGroup property 4-12
InputName property 4-12

H
H
loop shaping 1-10
mixed-sensitivity 1-10
mixsyn 2-17
norm 2-3
norm, sectf 10-301
sampled-data 1-10
H control
performance objectives 5-3
H2
norm 2-3
H2 control synthesis 10-107

Index-2

L
L2-norm 5-28
Le,V. X. 10-188
linear matrix inequalities
LMI solvers 1-18
LMI solvers 1-18
loop shaping 1-10
loopsyn 2-5
loop transfer function matrix 2-5
loopmargin 4-17
loopsens 4-15

Index

loop-shaping synthesis
loopsyn H optimal method 10-184
LTR loop transfer recovery method 10-194
loopsyn 10-184
LQG loop tranfer-function recovery. See ltrsyn
LQG optimal control 10-109
ltrsyn 10-194

with uncertain real coefficients 4-3

modreal 3-13

Monte Carlo random sample 1-5

-synthesis control design 5-19
multiplicative (relative) error bound 3-4
multiplicative error bound 10-32
multiplicative uncertainty 2-2
multivariable margins 10-172, 10-376

magfit 10-95

makeweight utility 4-6

ncfmargin 10-232

max entropy 10-128

maxgain variable 4-11
McFarlane, D. C., 10-235

ncfmr 3-14

mfilter 10-207

norm 4-20
Normalized coprime factor 10-228
Left Compie Factorization 10-228
Right Coprime Facotrization 10-228
Normalized coprme factor
ncfmr 10-228
Normallized coprime factor
balanced model truncation 10-228
norms 5-28
H 2-2
H2 2-2
performance 5-29

Mixed H /H2
lcontroller synthesis 10-8
mixed H/H2
controller synthesis 1-10
mixed-sensitivity cost function 2-17
mixed-sensitivity loop shaping 2-17
mixed-sensitivity synthesis
augw 10-20

plant augmentation for 10-8

mixsyn 10-204
mktito 10-209

Modal form realization 10-211

modreal 10-211
model reduction 1-14
additive error methods 3-6
balanced stochastic method 3-9
large-scale models 3-13
multiplicative error method 3-8
normalized coprime factor truncation 3-14
model reduction routines 3-4
models

ncfsyn 10-232
NominalValue property 4-12

O
OutputGroup property 4-12
OutputName property 4-12

P
Percentage property 4-3

performance weights 5-33

Index-3

Index

perturbation
additive 2-6
multiplicative 2-6
PlusMinus property 4-3
proper system 10-22

Slow and fadt modes decomposition

slowfast 10-315
slowfast 10-315

Range property 4-3

spectral factor 10-34

square root method 10-23
squaring-down prefilter 10-185
Stable and antistable projection 10-318
stabproj 10-318
stabmarg variable 4-10

reduce 3-6

stabproj 10-318

relative error 10-32

report variable 4-10
robust performance
defined 5-19
robustness
of stability 2-16
robustness analysis
Monte Carlo 1-7
worst case 1-7
robustperf 4-23
robuststab 4-9
roll-off 2-16

SVD system realization 10-141

sysic 5-13
System realization 10-141

S
Safonov, M. G.
imaginy axis zeros H 10-31
return difference matrix 10-112
stabiltiy and robustness 10-304
schurmr 3-6
sdhfsyn 10-292
sectf 10-300
sector bilinear transformation 10-12, 10-300
sensitivity 2-5
sensitivity S 10-20
singular values 2-3
properties of 2-3

Index-4

T
2-norm
definition 5-28

U
ultidyn 4-5
uncertain elements 4-2
uncertain LTI system 1-3
uncertain parameters 4-3
uncertain state-space object
See USS object
uncertainty
capturing 4-6
ureal 4-3
USS object 1-5
usubs

substituting worst-case values for uncertain

elements 4-11

Index

W
W1 and W2 shaping filters 4-13
wcgain

computing worst-case peak gain 4-10

wcsens 4-23
wcu variable 4-11
weighted mixed-sensitivity 10-8
augw 10-20

weighted norms 5-29

worst-case
peak gain 1-9

Z
Zames, G. 10-304
Zhou, K. 10-235

Index-5