µ-Analysis and

Synthesis Toolbox
For Use with MATLAB ®

Gary J. Balas
John C. Doyle
Keith Glover
Andy Packard
Roy Smith

Computation

Visualization

Programming

User’s Guide
Version 3
How to Contact The MathWorks:

www.mathworks.com Web
comp.soft-sys.matlab Newsgroup

[email protected] Technical support

[email protected] Product enhancement suggestions
[email protected] Bug reports
[email protected] Documentation error reports
[email protected] Order status, license renewals, passcodes
[email protected] Sales, pricing, and general information

508-647-7000 Phone

508-647-7001 Fax

The MathWorks, Inc. Mail

3 Apple Hill Drive
Natick, MA 01760-2098

For contact information about worldwide offices, see the MathWorks Web site.

µ-Analysis and Synthesis Toolbox User’s Guide

 COPYRIGHT 1993 - 2001 by MUSYN Inc. and 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 repro-
duced 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
or for the federal government of the United States. By accepting delivery of the Program, the government
hereby agrees that this software qualifies as "commercial" computer software within the meaning of FAR
Part 12.212, DFARS Part 227.7202-1, DFARS Part 227.7202-3, DFARS Part 252.227-7013, and DFARS Part
252.227-7014. The terms and conditions of The MathWorks, Inc. Software License Agreement shall pertain
to the government’s use and disclosure of the Program and Documentation, and shall supersede any
conflicting contractual terms or conditions. If this license fails to meet the government’s minimum needs or
is inconsistent in any respect with federal procurement law, the government agrees to return the Program
and Documentation, unused, to MathWorks.
MATLAB, Simulink, Stateflow, Handle Graphics, and Real-Time Workshop are registered trademarks, and
Target Language Compiler is a trademark of The MathWorks, Inc.
Other product or brand names are trademarks or registered trademarks of their respective holders.
Printing History: July 1993 First printing
November 1995 Second printing
June 1998 Third printing
January 2001 Fourth Printing for Version 3.05 (Release 12)
June 2001 Online only for Version 3.06 (Release 12.1)
 Contents
Overview of the Toolbox
1
Organization of This Manual . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-4

Working with the Toolbox

2
Command Line Display . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-3

The Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-3

SYSTEM Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-3
VARYING Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-6
CONSTANT Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-8
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-8

Accessing Parts of Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-9

Interconnecting Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-11

Plotting VARYING Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 2-15

VARYING Matrix Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 2-18

More Sophisticated SYSTEM Functions . . . . . . . . . . . . . . . . 2-20

Frequency Domain Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 2-20
Time Domain Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-22
Signal Processing and Identification . . . . . . . . . . . . . . . . . . . . 2-27

Interconnection of SYSTEM Matrices: sysic . . . . . . . . . . . . 2-34

Variable Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-35
Running sysic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-37
HIMAT Design Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-38

i
 H∞ Control and Model Reduction
3
Optimal Feedback Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
Performance as Generalized Disturbance Rejection . . . . . . . . . 3-2
Norms of Signals and Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 3-3
Using Weighted Norms to Characterize Performance . . . . . . . . 3-5

Interconnection with Typical MIMO

Performance Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10

Commands to Calculate the H2 and H∞ Norm . . . . . . . . . . . 3-14

H2 norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-14
H∞ norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-14
Discrete-Time H∞ Norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-15

Commands to Design H∞ Output Feedback Controllers . . 3-16

H∞ Design Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-18

H∞ Optimal Control Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-20

Historical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-22
Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-23
Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-25
Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-27
The Riccati Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-28
Computing the H∞ Norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-30

H∞ Full Information and Full Control Problems . . . . . . . . 3-31

Problem FI: Full Information . . . . . . . . . . . . . . . . . . . . . . . . . . 3-33
Problem FC: Full Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-34

H∞ Output Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3-36
Disturbance Feedforward and Output Estimation . . . . . . . . . 3-38
Converting Output Feedback to Output Estimation . . . . . . . . 3-39
Relaxing Assumptions A1–A4 . . . . . . . . . . . . . . . . . . . . . . . . . . 3-41

Discrete-Time and Sampled-Data H∞ Control . . . . . . . . . . . 3-43

Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-43
Sampled-Data Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-44
Discrete-Time and Sampled-Data Example . . . . . . . . . . . . . . . 3-45

ii Contents
 Loop Shaping Using H∞ Synthesis . . . . . . . . . . . . . . . . . . . . . 3-52

Model Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-55

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-69

Modeling and Analysis of Uncertain Systems

4
Representing Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-3
Linear Fractional Transformations (LFTs) . . . . . . . . . . . . . . . . 4-3
Parametric Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-5

µ-Tools Commands for LFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-10

Interconnections of LFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-13

Parameter Uncertainty in Transfer Functions . . . . . . . . . . . . 4-16
Linear State-Space Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . 4-20
Unmodeled Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-23
Mixed Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-30

Analyzing the Effect of LFT Uncertainty . . . . . . . . . . . . . . . 4-31

Using µ to Analyze Robust Stability . . . . . . . . . . . . . . . . . . . . . 4-31

Using µ to Analyze Robust Performance . . . . . . . . . . . . . . . . 4-40

Using µ to Analyze Worst-Case Performance . . . . . . . . . . . . . . 4-46
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-48

Structured Singular Value Theory . . . . . . . . . . . . . . . . . . . . . 4-49

Complex Structured Singular Value . . . . . . . . . . . . . . . . . . . 4-50

Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-50
Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-54
Computational Exercise with the mu Command . . . . . . . . . . . 4-56

Mixed Real/Complex Structured Singular Value . . . . . . . . 4-61

iii
 Specifics About Using the mu Command with
Mixed Perturbations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-64
Computational Exercise with the mu Command —
Mixed Perturbations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-64

Linear Fractional Transformations . . . . . . . . . . . . . . . . . . . . 4-67

Well Posedness and Performance for Constant LFTs . . . . . . . 4-68

Frequency Domain µ Review . . . . . . . . . . . . . . . . . . . . . . . . . . 4-71

Robust Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-71
Robust Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-73

Real vs. Complex Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 4-74

Block Structures with All Real Blocks . . . . . . . . . . . . . . . . . . . 4-75

Generalized µ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-81

Using the mu Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-83

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-84

Control Design via µ Synthesis

5
Problem Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-3

Replacing µ With Its Upper Bound . . . . . . . . . . . . . . . . . . . . . . 5-6

D-K Iteration: Holding D Fixed . . . . . . . . . . . . . . . . . . . . . . . . 5-9

D-K Iteration: Holding K Fixed . . . . . . . . . . . . . . . . . . . . . . . . 5-10

Two-Step Procedure for Scalar Entries d of D . . . . . . . . . . . . . 5-10
Two-Step Procedure for Full D (Optional Reading) . . . . . . . . . 5-11

Commands for D – K Iteration . . . . . . . . . . . . . . . . . . . . . . . . 5-13

iv Contents
 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-14

Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-14

Auto-Fit for Scalings (Optional Reading) . . . . . . . . . . . . . . . 5-15

Graphical User Interface Tools

6
Workspace User Interface Tool: wsgui . . . . . . . . . . . . . . . . . . 6-3
File Menu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-8
Options Menu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
Export Button . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-11

Spinning Satellite Example: Problem Statement . . . . . . . . 6-14

D-K Iteration User Interface Tool: dkitgui . . . . . . . . . . . . . . 6-19

Drop Box Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-23
Completing Problem Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-24
D-K Iteration Parameters Window . . . . . . . . . . . . . . . . . . . . . . 6-26
D-K Iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-30
Options Menu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-39

LFT Time Simulation User Interface Tool: simgui . . . . . . . 6-40

Spinning Satellite Example: Time Simulation . . . . . . . . . . . . . 6-40
Setting Up simgui . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-41
Plots and Line Types Scroll Table . . . . . . . . . . . . . . . . . . . . . . . 6-46
Plotting Window Setup and Titles . . . . . . . . . . . . . . . . . . . . . . 6-49
Printing Menu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-59
Loading and Saving Plot Information . . . . . . . . . . . . . . . . . . . . 6-59
Simulation Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-61
Simulation Parameter Window . . . . . . . . . . . . . . . . . . . . . . . . . 6-62

Dragging and Dropping Icons . . . . . . . . . . . . . . . . . . . . . . . . . 6-66

v
 Robust Control Examples
7
SISO Gain and Phase Margins . . . . . . . . . . . . . . . . . . . . . . . . . . 7-3

MIMO Loop-at-a-Time Margins . . . . . . . . . . . . . . . . . . . . . . . . . 7-7

Analysis of Controllers for an Unstable System . . . . . . . . . 7-15

Redesign Controllers Using D – K Iteration . . . . . . . . . . . . . . 7-28

MIMO Margins Using µ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-33

Loop-at-a-Time Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-38
Constructing Destabilizing Perturbations . . . . . . . . . . . . . . . . 7-41

Space Shuttle Robustness Analysis . . . . . . . . . . . . . . . . . . . . 7-44

Aircraft Model: Rigid-Body . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-45
Aircraft Model: Aerodynamic Uncertainty . . . . . . . . . . . . . . . . 7-46
Actuator Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-48
Exogenous Disturbances, Noises, and Commands . . . . . . . . . . 7-49
Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-51
LFT Aero-Coefficient Uncertainty . . . . . . . . . . . . . . . . . . . . . . 7-53
Create Open-Loop Interconnection . . . . . . . . . . . . . . . . . . . . . . 7-54
Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-57
Nominal Frequency Responses . . . . . . . . . . . . . . . . . . . . . . . . . 7-58
Robust Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-60
Robust Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-64
Worst-case Perturbations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-65
Time Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-67
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-72
Space Shuttle References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-73
Shuttle Rigid Body Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-73

HIMAT Robust Performance Design Example . . . . . . . . . . 7-75

HIMAT Vehicle Model and Control Objectives . . . . . . . . . . . . 7-76
Closed-Loop Feedback Structure . . . . . . . . . . . . . . . . . . . . . . . 7-77
Uncertainty Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-78
Specifications of Closed-Loop Performance . . . . . . . . . . . . . . . 7-80
Building the Open-Loop Interconnection . . . . . . . . . . . . . . . . . 7-83
µ-synthesis and D – K Iteration . . . . . . . . . . . . . . . . . . . . . . . . 7-85
Loop Shaping Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . 7-86

vi Contents
 H∞ Design on the Open-Loop Interconnection . . . . . . . . . . . . . 7-89
Assessing Robust Performance with µ . . . . . . . . . . . . . . . . . . . . . . 7-93
µ Analysis of H∞ Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-94
µ-Analysis of Loop Shape Design . . . . . . . . . . . . . . . . . . . . . . . 7-97
Recapping Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-99
D – K Iteration for HIMAT Using dkit . . . . . . . . 7-100
H∞ Loop Shaping Design for HIMAT . . . . . . . . . . . . . . . . . . . 7-121
H∞ Loop Shaping Feedback Compensator Design . . . . . . . . . 7-123
Assessing Robust Performance with µ . . . . . . . . . . . . . . . . . . . . . 7-125
Reduced Order Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-126
Introducing a Reference Signal . . . . . . . . . . . . . . . . . . . . . . . . 7-128
HIMAT References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-131

F–14 Lateral-Directional Control Design . . . . . . . . . . . . . . 7-132

Nominal Model and Uncertainty Models . . . . . . . . . . . . . . . . 7-135
Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-142
Analysis of the Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-145

A Process Control Example: Two Tank System . . . . . . . . . 7-151

Experimental Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-151
An Idealized Nonlinear Model . . . . . . . . . . . . . . . . . . . . . . . . . 7-152
Normalization Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-156
Operating Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-157
Actuator Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-157
Experimental Assessment of the Model . . . . . . . . . . . . . . . . . 7-158
Developing the Interconnection Model for Design . . . . . . . . . 7-165
Linearized Nominal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-166
Perturbation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-167
Sensor Noise Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-173
Specifying the Design Requirements . . . . . . . . . . . . . . . . . . . 7-173
Controller Design and Analysis . . . . . . . . . . . . . . . . . . . . . . . 7-176
Design Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-177
Closed-Loop Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-179
Experimental Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-187
Two Tank System References . . . . . . . . . . . . . . . . . . . . . . . . . 7-191

vii
 Reference
8
Summary of Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-3

Commands Grouped by Function . . . . . . . . . . . . . . . . . . . . . . 8-10

viii Contents
 1

Overview of the Toolbox

1 Overview of the Toolbox

The µ-Analysis and Synthesis Toolbox (µ-Tools) is a collection of functions

(commands) developed primarily for the analysis and synthesis of control
systems, with an emphasis on quantifying the effects of uncertainty. µ-Tools
provides a consistent set of data structures for the unified treatment of systems
in either a time domain, frequency domain, or state-space manner. µ-Tools also
gives MATLAB® users access to recent developments in control theory, namely
H∞ optimal control and m analysis and synthesis techniques. This package
allows you to use sophisticated matrix perturbation results and optimal control
techniques to solve control design problems. Control design software, such as
µ-Tools, provides a link between control theory and control engineering.
Computational algorithms for the structured singular value, µ, are main
features of the toolbox. µ is a mathematical object developed to analyze the
effect of uncertainty in linear algebra problems. µ is particularly (though not
exclusively) useful in analyzing the effect of parameter uncertainty and
unmodeled dynamics on the stability and performance of multiloop feedback
systems. µ-Tools is a collection of tools designed to help you analyze the
sensitivity of closed-loop systems to detailed and complex types of modeling
errors. µ-Tools is also suitable to design control systems that are insensitive to
classes of variations that you expect between your model and the actual
physical process which must be controlled.
The µ framework appropriately generalizes notions such as gain margin, phase
margin, disturbance attenuation, tracking, and noise rejection into a common
framework suitable for analysis and design, in both single-loop and multiloop
feedback systems. Even when working with single-loop feedback systems, some
multi-input, multi-output (MIMO) systems arise during the analysis. Hence, a
unified framework to deal with MIMO linear systems is important, with full
support for both the time and frequency domain. µ-Tools provides the
capability to build complex interconnections (such as cascade, parallel, and
feedback connections), compute properties (such as poles and zeros), calculate
time and frequency responses, manipulate these responses (FFT for the time
domain signals, Bode analysis for the frequency domain functions), and plot
results. µ-Tools supports two data types in addition to the standard matrices:
SYSTEM matrices for state-space realizations and VARYING matrices for
time and frequency responses.

1-2
The advanced features of µ-Tools are aimed at:

• Analyzing the effect of uncertain models on achievable closed-loop

performance
• Designing controllers for optimal worst-case performance in the face of the
plant uncertainty

Hence, it is imperative that you understand the following:

• The characterization of “good” closed-loop performance

• How to represent model uncertainty in this framework
• The technical tools available to answer questions about the robustness of a
given closed-loop system to certain forms of model uncertainty
• The technical tools available to design controllers which achieve good
performance in the face of the model uncertainty

The characterization of performance is discussed in Chapter 3. In Chapter 4,

we concentrate on modeling uncertainty, and the effect it can have on the
guaranteed performance level of the closed-loop system. The tools for design
are discussed in the latter part of Chapter 3 and in Chapter 5. Chapter 6
presents graphical user interfaces for the workspace, control design and time
simulation. Chapter 7 contains a number of examples to show how to apply
µ-Tools to robust control problems.

1-3
1 Overview of the Toolbox

Organization of This Manual

The contents of this manual are not intended to be read in sequence. The
chapters are grouped by topic, and coverage within sections varies from syntax
descriptions to theoretical justifications.
After skimming the sections on command use and syntax, we feel that the
examples are the quickest manner in which you can get a feel for the
techniques. However, it may be necessary to refer back to the concept sections
as different topics become relevant.
For basic use of the toolbox (system interconnections, system calculations,
frequency responses, time responses and plotting), the recommended sections
are shown below.

Emphasis Topic Pages

Use Working with the Toolbox Chapter 2
Examples SISO Gain and Phase Margins Chapter 7
MIMO Loop-at-a-Time Margins

GUI Workspace Tool Chapter 6

LFT Time Simulation Tool

For robust stability analysis, additional recommended reading is

Emphasis Topic Pages

Concepts Modeling Uncertainty Chapter 4

Robust Stability Analysis

Examples SISO Gain and Phase Margins Chapter 7

MIMO Loop-at-a-Time Margins
MIMO Margins Using µ
Space Shuttle Robustness Analysis
Two Tank System

GUI LFT Time Simulation Tool Chapter 6

1-4
 Organization of This Manual

For robust performance analysis, additional recommended reading is

Emphasis Topic Pages

Concepts Performance Objectives Chapter 3

Robust Performance Analysis
Chapter 4

Examples Unstable SISO Analysis Chapter 7

HIMAT Robust Performance
F-14 Lateral-directional Control
Space Shuttle Robustness Analysis
Two Tank System

GUI LFT Time Simulation Tool Chapter 6

For robust control design, additional recommended reading is

Emphasis Topic Pages

Concepts H∞ Performance Objectives Chapter 3

H∞ Control Design
H∞ Loop Shaping
Chapter 4
µ Upper Bound
Robust Control Design Chapter 5

Examples HIMAT Robust Performance Design Chapter 7

F-14 Lateral-Directional Control Design
Space Shuttle Robustness Analysis
Two Tank System

GUI D-K Iteration Tool Chapter 6

1-5
1 Overview of the Toolbox

Advanced topics are

Emphasis Topic Pages

Concepts Control Theory Chapter 3

Discrete-time and Sampled-Data Control
Model Reduction
Chapter 4
Structured Singular Value Theory

There are three graphical user interfaces, described in detail in Chapter 6. The
interfaces are:

• wsgui is a Workspace Manager. It allows you to select, save and clear the
workspace variables, based on their type (VARYING, SYSTEM,
CONSTANT) and other more complicated selection rules. This tool is useful
during all MATLAB sessions, and is described in the “Workspace User
Interface Tool: wsgui” section of Chapter 6.
• simgui is a time-domain simulation package for uncertain closed-loop
systems. It is powerful enough to build templates for the complex plotting
requirements of a large MIMO control design report. This tool is described in
“LFT Time Simulation User Interface Tool: simgui” section of Chapter 6.
• dkitgui is a control design program to assist you with the DK iteration. It
aids in understanding the DK iteration process. The flexibility allows you to
easily modify performance objectives and uncertainty models during the
iteration. This tool is described in “DK Iteration User Interface Tool: dkitgui”
section of Chapter 6.

1-6
 2

Working with the Toolbox

2 Working with the Toolbox

This chapter gives a basic introduction, with examples, to the µ-Analysis and
Synthesis Toolbox (µ-Tools) data structure and commands. Introductory
examples are found in the demo programs msdemo1.m and msdemo2.m. Other
demonstration files are introduced in subsequent chapters. You can copy these
files from the mutools/subs source into a local directory and examine the
effects of modifying some of the commands.

2-2
 Command Line Display

Command Line Display

All µ-Tools commands have a built-in use display. Any command called with no
input arguments, or the incorrect number of input arguments results in a brief
description of the correct command line use. For example, at the command line
mu
usage: [bonds,dvec,sens,pvec] = mu(matin,blk,opt)

The Data Structures

µ-Tools represents systems (either in state-space form or as frequency
dependent input/output data) as single data entries. Data structures This
allows you to have all of the information about a system in a single MATLAB
variable. In addition, the µ-Tools functions that return a single variable can be
nested, allowing you to build complex operations out of a few nested operations.
Examples of this are found throughout this chapter.
The layout of the data structure is quite simple. Consider a typical MATLAB
matrix, which is usually made up of real and complex numbers. In addition to
these, MATLAB also allows for a few special values, such as NaN (not a
number), Inf (infinity), and -Inf. Since these are allowable values, but are not
typically found in state-space realizations, frequency responses, or time
responses, they can be used to differentiate more complicated data types from
plain, constant matrices. This is the approach taken by µ-Tools.

SYSTEM Matrices
Consider a linear, finite dimensional system, modeled by the state-space
representation

x· = Ax + Bu
y = Cx + Du

nx × nx
If the system R has nx states, nu inputs, and ny outputs, then A ∈ R ,
B ∈ R n x × n u , C ∈ R n y × n x , and D ∈ R n y × n u . Systems of this type are
represented in µ-Tools by a single MATLAB data structure, referred to as a
SYSTEM matrix.

2-3
2 Working with the Toolbox

Throughout this manual, we use the notation

A B
C D

This is not to be confused with a CONSTANT MATLAB matrix containing the

state-space data. Rather, this notation usually refers to the causal, linear,
dynamical system described by the differential equations

x· = Ax + Bu
y = Cx + Du

However, sometimes the notation stands for the transfer function

D + C(sI – A)-1 B,
and sometimes it pertains to the SYSTEM matrix containing the state-space
data. In any event, the exact meaning is purposely left vague, and easily
determined from context.
The command pck creates SYSTEM matrices from separate state-space data.
The matrices
a = [-.15,.5; -.5, -.15];
b = [.2 4; -.4 0];
c = [5 5];
d = [.1 -.1];

represent the state-space data of a two-state, two-input, single-output system.

The SYSTEM matrix, sys, is created by
sys = pck(a,b,c,d);

Structural information about the matrix sys can be obtained with the
command minfo.
minfo(sys)

system: 2 states 1 outputs 2 inputs

2-4
 The Data Structures

The command unpck extracts the a, b, c, and d matrices from a SYSTEM

matrix. You can also examine the contents of the a, b, c, and d matrices without
explicitly forming them as new variables. Use the command see for this
purpose.
see(sys)

A matrix

-0.1500 0.5000
-0.5000 -0.1500

press any key to move to B matrix

B matrix

-0.2000 4.0000
-0.4000 0

press any key to move to C matrix

C matrix

5 5

press any key to move to D matrix

D matrix

-0.1000 -0.1000

The commands minfo and see work on any of the µ-Tools data structures. The
command pss2sys converts CONSTANT matrix data in packed form,
[A B; C D], into a µ-Tools SYSTEM matrix. The command sys2pss transforms
a SYSTEM matrix in a packed CONSTANT matrix. Alternatively, you can
generate a purely random SYSTEM matrix with the command sysrand by
specifying its number of states, inputs and outputs.
The command spoles finds the eigenvalues of the A matrix of a SYSTEM
matrix. The transmission zeros are calculated using szeros. In this example,
sys has no transmission zeros. A formatted display of the system poles is
produced with the µ-Tools command rifd.

2-5
2 Working with the Toolbox

spoles(sys)
ans =

–0.1500 + 0.5000i
–0.1500 – 0.5000i

rifd(spoles(sys))

real imaginary frequency damping

–1.5000e-01 –5.0000e-01 5.2202e–01 2.8735e–01

–1.5000e-01 5.0000e–01 5.2202e–01 2.8735e–01

SYSTEM matrices can easily be interconnected (cascade, parallel, feedback) to

give new SYSTEM matrices. See “Interconnecting Matrices” for more
information.

VARYING Matrices
Matrix-valued functions of a single, independent real variable are common in
systems theory. VARYING matrices The frequency response of a
multiple-input, multiple-output (MIMO) system is a good example of such a
function. The independent variable is frequency, and at each frequency the
transfer function between the inputs and the outputs is a complex matrix.
represents these types of matrix functions with a data structure called a
VARYING matrix.
In general, suppose G is a matrix-valued function of a single real variable
n×m
G:R → C . One method to store this function on the computer is to evaluate
the function G at N discrete values of x ∈ R , call them x1,x2,. . .,xN and store all
of the evaluations. This is the approach taken by µ-Tools.
Consider a simple example.
mat1 = [.1 -.1;.25.5];
iv1 = 0;
mat2 = 2*mat1;
iv2 = 1;
mat3 = 2*mat2;
iv3 = 2;

2-6
 The Data Structures

The command vpck creates the VARYING matrix data structure from column
stacked matrix and independent variable data. This is done as follows.
matdata = [mat1; mat2; mat3];
ivdata = [iv1; iv2; iv3];
vmat = vpck(matdata,ivdata)

minfo displays structural characteristics of the matrix and displays the data.
minfo(vmat)

varying:3 pts 2 rows 2 cols

see(vmat)

2 rows 2 columns

iv = 0

0.1000 –0.1000
0.2500 0.5000

iv = 1

0.2000 –0.2000
0.5000 1.0000

iv = 2

0.4000 –0.4000
1.0000 2.0000

Note that variable name iv stands for independent variable in the above
display. The command seeiv displays only the independent variable values of
the VARYING matrix vmat.
seeiv(vmat)

Analogous to vpck the command vunpck unpacks the matrix data and
independent variable data from a VARYING matrix.

2-7
2 Working with the Toolbox

Several commands allow manipulation of the matrix and independent variable

data. tackon simply appends one VARYING matrix to another (both must have
the same number of rows and columns in the matrix data). This destroys any
sequential properties of the independent variable data; i.e., the data will not be
in monotonically increasing order. sortiv reorders the matrices within a
VARYING matrix so that the independent variables are increasing (or
decreasing).
All of the information about the structure is contained in a single MATLAB
matrix, hence functions returning single matrices can be nested. Hence,
sophisticated manipulations can be formed as single line commands. For
example, to merge two VARYING matrices and reorder the independent
variable’s values, use the command sortiv(tackon(vmat1,vmat2)).

CONSTANT Matrices
If a MATLAB variable is neither a SYSTEM nor a VARYING matrix it is
treated by µ-Tools as a CONSTANT matrix. CONSTANT matrices
CONSTANT matrices can be arguments to functions that normally expect
VARYING or SYSTEM matrix arguments.
The treatment of CONSTANT matrices is consistent with that of a constant
gain linear system. In operations normally performed on SYSTEM matrices,
the CONSTANT matrix is analogous to a linear system with only a D matrix.
In operations where a VARYING interpretation is required, the CONSTANT
matrix is assumed to be constant across all values of the independent variable.
This is consistent with the frequency response (or step response) of a constant
gain linear system.

Acknowledgments
The data structures used in are based on the following paper.
Stein, Gunter, and Stephen Pratt, “LQG Multivariable Design Tools,” AGARD
Lecture Series No 117, Multi-variable Analysis and Design Techniques,
September 1981.

2-8
 Accessing Parts of Matrices

Accessing Parts of Matrices

The SYSTEM matrix and VARYING matrix data structures allow a large
amount of data to be stored in a single MATLAB entity. It is often necessary to
display or operate on only a part of that data.
The command sel creates subsystems from VARYING, SYSTEM, or
CONSTANT matrices. The subsystem rows (or outputs) and columns (or
inputs) are specified. Conceptually, the options are
lcl subvmat = sel(vmat,rows,columns);
subconst = sel(const,rows,columns);
subsyst = sel(syst,outputs,inputs);

For example,
subvmat = sel(vmat,1:2,2);
minfo(subvmat)

varying:3 pts 2 rows 1 cols

selects rows 1 and 2 and column 2 from each matrix in vmat. You can use the
MATLAB colon notation in the specification of the rows and columns. To select
all rows or columns, use the character string ’:’ in single quotes. When sel is
used on a SYSTEM matrix, only the dimensions of the B, C, and D matrices
change. All the states remain, which may result in a (non)minimal system.
Extra states of the system can be removed by performing a balanced realization
(sysbal) or with the commands strunc and sresid.
For VARYING matrices you can access a portion of the independent variables
with the command xtract. For example,
vmat2 = xtract(vmat,0.5,1.5);

2-9
2 Working with the Toolbox

selects the matrices in vmat with independent variables between 0.5 and 1.5.
In this case it is a VARYING matrix with a single data point.
see(vmat2)

2 rows 2 columns

iv = 1

0.2000 –0.2000
0.5000 1.0000

A VARYING matrix can be converted to a CONSTANT matrix via the

command var2con.
The command xtracti extracts (as a VARYING matrix) the data by
independent variable index, rather than by independent variable value. As in
the case of xtract, xtracti returns a VARYING matrix.

2-10
 Interconnecting Matrices

Interconnecting Matrices
µ-Tools provides several functions for connecting matrices together. All of the
functions described here work with interconnections of SYSTEM and
CONSTANT matrices or VARYING and CONSTANT matrices. If the matrices
represented are consistent, the combination is allowed. The interconnection of
a SYSTEM and a VARYING matrix is not allowed in µ-Tools (actually it is
allowed — see veval for examples of VARYING SYSTEM matrices).
The commands madd, msub, and mmult perform the appropriate arithmetic
operations on the matrices. A block diagram representation is shown in the
following figure.

madd(mat1,mat2) mmult(mat1,mat2)

mat1 
 ?
+h  mat1  mat2 
6
mat2 

Note that for multivariate matrices, the order of the arguments is important.
In the VARYING matrix case, the arithmetic operations are performed
matrix-by-matrix, for each value of the independent variable. The following
example illustrates this:
A two-row and one-column VARYING matrix, vmat3, is constructed with three
independent variables values.
vmat3 = vpck([2 2 4 4 8 8]',[0 1 2]');
minfo(vmat3)

varying: 3 pts 2 rows 1 cols

2-11
2 Working with the Toolbox

The VARYING matrix vmat3 is multiplied by vmat to form vmat4 and the value
of the resulting matrix at the independent variables between 0 and 0.5 are
displayed.
vmat4 = mmult(vmat,vmat3);
minfo(vmat4) 60.23in

varying: 3 pts 2 rows 1 cols

see(xtract(vmat4,0,0.5))

2 rows 1 column

iv = 0
0
1.5000

Each matrix in vmat is 2 × 2. Each matrix in vmat3 is 2 × 1, and so the

multiplication results in each matrix of vmat4 being 2 × 1. Commands madd,
msub, and mmult allow up to nine matrices of compatible dimensions to be
added, subtracted or multiplied simultaneously by including them as input
arguments. When interconnecting VARYING matrices with any of the
commands in this section, a check is made to verify that each matrix has the
same independent variable values. If not, an error is returned.
There are additional commands for combining matrices: abv, daug, and sbs.
These can be interpreted as placing matrices above one another, diagonal
augmentation of matrices, and placing matrices side-by-side, as shown in the
following figure.

abv(top,bot) daug(ul,lr) sbs(left,right)

 top   ul   +h left 

6
 bot   lr  right 

2-12
 Interconnecting Matrices

VARYING matrices are combined as described above. For example,

see(xtract(daug(vmat,vmat3),0.5,1.5))

4 rows 3 columns

iv = 1

0.2000 –0.2000 0
0.5000 1.0000 0
0 0 4.0000
0 0 4.0000

abv, daug, and sbs allow up to nine matrices of compatible dimensions to be

combined simultaneously.
For large interconnections of matrices, it is tedious building them up
piece-by-piece using these commands (even using an M-file to do it). The
program sysic performs general interconnections, using algebraic descriptions
of the relationships between the inputs, outputs, and internal matrices. See
“Interconnection of SYSTEM Matrices: sysic” for more detail.
µ-Tools handles CONSTANT, VARYING, and SYSTEM matrices, such that the
following diagram is commutative.

2-13
2 Working with the Toolbox

SYSTEM, CONSTANT
Matrices
Combined  Interconnect Individual
System Systems

Frequency Frequency
Response Response

? ?
Interconnect
Combined  Individual
System Systems

VARYING Matrix
Consider beginning in the upper-right corner (individual systems of SYSTEM
and CONSTANT matrices) and proceeding to the lower left corner (a single
VARYING matrix). The result will be independent of the path taken. This is
because in linear systems, the frequency response of an interconnection is the
algebraic interconnection of the individual frequency responses. However, due
to numerical roundoff, the calculations actually are not commutative, and
there may be small differences between the two results. For example, it is
sometimes numerically better to interconnect the frequency response of two
systems rather than interconnect the two systems and then take their
frequency response. This can be true when there are a large number of states
in the interconnection structure.
Interconnections involving feedback are performed with sysic, described in
“Interconnection of SYSTEM Matrices: sysic” . The basic feedback loop
interconnection program used by sysic is called starp, and is described in
Chapter 4, “m-Tools Commands for LFTs” on page 4-10.
The commands (sbs, mmult, starp, etc.) to form the interconnection step are
identical whether you are dealing with VARYING or SYSTEM
representations.

2-14
 Plotting VARYING Matrices

Plotting VARYING Matrices

The function vplot plots VARYING matrices. The arguments for vplot are
similar to MATLAB’s plot command, with the exception that it is not
necessary to specify the values for the x-axis. The x-axis data corresponds to the
independent variable’s values, which are already stored within each VARYING
matrix. For example,
vplot(vmat)
tmp1 = 'vplot example: vmat, matrix values';
tmp2 = ' vs. independent variable';
title([ tmp1 tmp2 ]) xlabel('independent variable value')
ylabel('matrix element value')

vplot example: vmat, matrix values vs. independent variable

2

1.5
Matrix element value

0.5

−0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Independent variable value

Note that every element of the matrix is plotted against the appropriate
independent variable. In the above example vmat is a 2 × 2 VARYING matrix,
giving four elements to be plotted. There are only three values (0, 1, and 2) of
the independent variable, and by default MATLAB draws a line between points
on the plot.

2-15
2 Working with the Toolbox

In the MATLAB plot command, different axis types are accessed by different
functions, loglog, semilogx, and others. In vplot the axis type is set by an
optional string argument. The default, used in the above example, is a linear/
linear scale. The generic vplot function call looks like
vplot('axistype',vmat1,'linetype1',vmat2,'linetype2',...)

The axistype argument, a character string, allows the specification of

logarithmic or linear axes as well as: magnitude, log magnitude, and phase.
There are also some control-specific options: bode, nyq, and nic, which specify
Bode, Nyquist, and Nichols plots, respectively. The complete range of choices is
not demonstrated here. Refer to vplot in the Chapter 8, “Reference”, for more
details. Subsequent sections introduce additional control-oriented examples
and demonstrate other options in vplot.
The linetype arguments are optional and are identical to those in the
MATLAB plot command.
An important feature of vplot is its ability to plot multiple VARYING matrices
on the same plot without having to have the same independent variables. A
vmat argument may be a CONSTANT. In this case the value of the
CONSTANT is plotted over all independent variables. This is consistent with
the interpretation given to CONSTANT matrices in the interconnection of
systems. Hence a CONSTANT matrix would appear as a horizontal line on the
plot. This is to be contrasted with a VARYING matrix containing only one data
point, which would appear as a single point on the plot.
Consider the following example where vmat2 is a VARYING matrix with only
one independent variable value (ie., one data point). The constant pi/2 is also
plotted.
vplot(vmat,'r-',vmat2,'g*',pi/2,'b-.')
title('vplot example: VARYING and CONSTANT matrices')
xlabel('Independent variable value')
ylabel('Matrix element value')

2-16
 Plotting VARYING Matrices

vplot example: VARYING and CONSTANT matrices

2

1.5
Matrix element value

1 *

0.5 *

-0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Independent variable value

2-17
2 Working with the Toolbox

VARYING Matrix Functions

Many of the MATLAB matrix functions have analogous µ-Tools functions that
operate on VARYING matrices. These operations are performed on a
matrix-by-matrix basis for every submatrix associated with an independent
variable value within the VARYING matrix. If a CONSTANT matrix is the
argument of these functions, the operation is identical to the corresponding
MATLAB function. These functions are
vabs vceil vdet vdiag veig
veval vexpm vfloor vinv vimag
vnorm vpinv vpoly vrcond vreal
vroots vschur vsvd

The functions veval and vebe perform a named operation on VARYING

matrices. vebe performs MATLAB or user-defined functions on the elements of
a VARYING matrix (for example: sin, tan. . .). veval operates on the entire
VARYING matrix and can perform any function including those with multiple
input and output arguments. vebe and veval allow the evaluation of any
MATLAB matrix function on VARYING matrices. The following example
shows a VARYING matrix with only one independent variable value for
brevity.
minfo(vmat2)
varying: 1 pts 2 rows 2 cols

see(veval('sin',vmat2))
2 rows 2 columns
iv = 1

0.1987 –0.1987
0.4794 0.8415

vmat2dat = var2con(vmat2);
sin(vmat2dat)

ans =

0.1987 –0.1987
0.4794 0.8415

2-18
 VARYING Matrix Functions

vebe('sqrt',sin(vmat2dat))
ans =

0.4458 0 + 0.4458i
0.6924 0.9173

All arithmetic operations can be performed on VARYING matrices, sometimes

with a built-in µ-Tools function, and sometimes resorting to veval. The
following table summarizes some standard operation.

MATLAB Matrix Function µ-Tools VARYING Function

A + B +...+ H madd(A,B,...,H)

A – B –...– H msub(A,B,...,H)

A * B *...* H mmult(A,B,...,H)

A / B vrdiv(A,B)

A \ B vldiv(A,B)

A .* B veval('.*',A,B)

A ./ B veval('./',A,B)

A ^ b veval('^',A,b)

A .^b veval('.^',A,b)

A' cjt(A)

A.' transp(A)

conj(A) cj(A)

sin(A) vebe('sin',A)

In each case, veval could have been used. However, veval can be quite slow,
since it is essentially a for loop of eval commands. For that reason, some
specific commands (madd, mmult, vldiv, etc.) are provided. The complete set of
VARYING operations should allow you to write algorithms more easily using
the data structures in µ-Tools.

2-19
2 Working with the Toolbox

More Sophisticated SYSTEM Functions

Frequency Domain Functions
The command frsp calculates frequency responses of SYSTEM matrices. You
can specify the frequencies at which the response is to be evaluated via the
MATLAB logspace and linspace commands. These become the independent
variable values in the VARYING frequency response output.
Given an input vector of N real frequencies, omega = [ω1,. . .,ωN] and a SYSTEM
matrix sys, the µ-Tools command frsp,
sys_g = frsp(sys,omega)

calculates
C(jwiI – A)–1B + D, i = 1, . . ., N
for each independent variable wi and stores it in sys_g whose independent
variables are the N frequency points. You can specify a discrete-time
evaluation by specifying an optional sampling time, T. In this case each matrix
in the VARYING output is
C(ejω,T I – A) B + D.
Consider a simple second order example. The function nd2sys creates the
SYSTEM representation from numerator and denominator polynomials. In the
following example the system sys1 has the transfer function

– 0.5s + 1
--------------------------------- .
2
s + 0.2s + 1

2-20
 More Sophisticated SYSTEM Functions

The MATLAB command logspace can create a logarithmically spaced

frequency vector.
sys1 = nd2sys([-0.5,1],[1,0.2,1]);
minfo(sys1)

system: 2 states 1 outputs 1 inputs

omega = logspace(-1,1,200);
sys1g = frsp(sys1,omega);
minfo(sys1g)

varying: 200 pts 1 rows 1 cols

vplot('bode',sys1g)

10 1
Log Magnitude

10 0

10 -1

10 -2
10 -1 10 0 10 1
Frequency (radians/sec)

4
Phase (radians)

-2

-4
10 -1 10 0 10 1
Frequency (radians/sec)

2-21
2 Working with the Toolbox

You can transform sys1 to the digital domain via a prewarped Tustin
transformation. The command tustin performs this function. In this example
a sample time of one second is used. The prewarping frequency is chosen as one
radian/second. For control purposes it is often better to choose the crossover
point as the prewarp frequency.
dsys1 = tustin(sys1,1,1);
dsys1z = frsp(dsys1,omega,1);
vplot('bode',sys1g,dsys1z)

10 1

10 0
Log Magnitude

10 -1

10 -2

10 -3

10 -4
10 -1 10 0 10 1
Frequency (radians/sec)

4
Phase (radians)

-2

-4
10 -1 10 0 10 1
Frequency (radians/sec)

Time Domain Functions

Time responses of continuous systems are calculated with the function trsp.
The required input arguments are the SYSTEM matrix and an input matrix.
A discrete SYSTEM matrix is handled with the function dtrsp.
User-specified time functions can be created with the function siggen. siggen
can create signals based on both random and deterministic functions. In this
example, siggen generates an input, u, to the system, sys1. Note how a
saturation, in this case π, is implemented.

2-22
 More Sophisticated SYSTEM Functions

u = siggen('min(pi,sqrt(t)+0.25*rand(size(t)))',[0:.1:40]);
y = trsp(sys1,u);

integration step size: 0.1

vplot(u,y)
title('Response of sys1 (dashed) to input, u (solid)')
xlabel('Time (seconds)')

Response of sys1 (dashed) to input, u (solid)

4

3.5

2.5

1.5

0.5

-0.5
0 5 10 15 20 25 30 35 40

Time (seconds)

trsp calculates a default step-size based on the minimum spacing in the input
vector and the highest frequency eigenvalue of the A matrix. For high order
systems, we recommended you use some form of model reduction (see the
“Model Reduction” section in Chapter 3) to remove high frequency modes
which do not have a significant effect on the output.

2-23
2 Working with the Toolbox

trsp assumes that the input is constant between the values specified in the
input vector. The following example illustrates the consequences of this
assumption.
sys1a = pck(-1,1,1);
minfo(sys1a)

system: 1 states 1 outputs 1 inputs

u1a = vpck([0:10:50]',[0:10:50]');
y1a = trsp(sys1a,u1a,60);

integration step size: 0.1

interpolating input vector (zero under hold)
minfo(y1a)

varying: 601 pts 1 rows 1 cols

vplot(u1a,'-.',y1a,'-')
xlabel('Time: seconds')
text(10,20,'input'), text(25,10,'output')

50

45

40

35

30

25
input
20

15
output
10

0
0 10 20 30 40 50 60

Time: seconds

At first glance the output does not seem to be consistent with the plotted input.
Remember that trsp assumes that the input is held constant between specified
values. The vplot and plot commands display a linear interpolation between
points. This can be seen by displaying the input signal interpolated to at least

2-24
 More Sophisticated SYSTEM Functions

as small a step-size as the default integration step (here 0.1 seconds). The
µ-Tools function vinterp performs zero-order hold or linear interpolation of the
independent variable.
vplot(u1a,'-.',vinterp(u1a,0.1),'--',y1a,'-')
xlabel('Time: seconds')
text(5,44,'dash-dot: input')
text(5,40,'dashed: interpolated input')
text(5,36,'solid: output')

50

45 dash-dot: input
dashed: interpolated input
40
solid: output
35

30

25

20

15

10

0
0 10 20 30 40 50 60

Time: seconds

The staircase nature of the input is now evident. To have a ramp input, you can
use the function vinterp to provide linear interpolation as shown by the
following example.
uramp = vinterp(u1a,0.1,60,1);
minfo(uramp)

varying: 601 pts 1 rows 1 cols

yramp = trsp(sys1a,uramp);

integration step size: 0.1

vplot(uramp,'-.',yramp,'-')
xlabel('Time: seconds')
text(20,15,'output')
text(12,20,'input')

2-25
2 Working with the Toolbox

50

45

40

35

30

25

20 input

15 output

10

0
0 10 20 30 40 50 60

Time: seconds

Note that because the input is regularly spaced, with spacing less than or equal
to the default integration time, trsp does not interpolate the input. No final
time was specified in the trsp argument list. However 60 seconds was specified
to vinterp as the final time, and this became the last time in the input vector
uramp.

To illustrate the use of dtrsp, a bilinear transformation generates a digital

system. The sample time is chosen as 1 second. The output is plotted against a
1 second interpolation of the input.
T = 1;
dsys1a = tustin(sys1a,T);
ydig = dtrsp(dsys1a,u1a,T);
vplot(ydig,'-',vinterp(u1a,1),'-.')
xlabel('Time: seconds')

2-26
 More Sophisticated SYSTEM Functions

50

45

40

35

30

25

20

15

10

0
0 10 20 30 40 50 60

Time: seconds

trsp can easily generate large VARYING matrices. If the independent

variables are more closely spaced than necessary for a good graphical display,
you can use the µ-Tools function vdcmate to select every nth point from the
final output.
Note that both vinterp and vdcmate have analogous Signal Processing Toolbox
functions, interp and decimate, but do not function in the same manner.

Signal Processing and Identification

The µ-Tools VARYING data structure provides a convenient means of storing
large amounts of experimental data in a single MATLAB matrix. To use this in
identification experiments, routines have been provided to perform Fast
Fourier Transforms (FFTs) and inverse FFTs for VARYING matrices. The
functions are vfft and vifft, respectively. These routines call the appropriate
MATLAB fft and ifft routines to perform the FFT calculations and have the
same function arguments as those for fft and ifft.
Note that the FFTs performed are one dimensional, irrespective of the row and
column dimensions of the VARYING matrix. The independent variable is used
as the index in the FFT. Consequently xfreq = vfft(xtime) works if xtime is
a VARYING matrix with row and column dimensions greater than one. The
result, xfreq, is a VARYING matrix of the same size and has the same number

2-27
2 Working with the Toolbox

of independent variables. The independent variables of xtime are assumed to

be time in seconds. xfreq is returned with frequency (in radians/second) as the
independent variable.
In the following example a random signal is passed through a single-input,
two-output system and FFTs of the outputs are performed with a single vfft
function. The sample frequency is 10 Hz, with a foldover frequency of 10π
radians/second. This may be hard to distinguish on the log-log plot. Output
time history plots are shown in Figure 2-1 and their FFTs are shown in
Figure 2-2.
time = [0:0.1:102.3]';
u1 = siggen('rand(size(t)) - 0.5',time);
sys2 = nd2sys(1,[5,1]);
sys3 = abv(sys1,sys2);
y1 = trsp(sys3,u1);
integration step size: 0.1
minfo(y1)
varying: 1024pts 2 rows 1 cols
vplot(y1)
title('Response of sys3 (solid) to input u1 (dashed)')
xlabel('Time: seconds')

y1f = vfft(y1);
minfo(y1f)

varying: 1024pts 2 rows 1 cols

vplot('liv,lm',y1f)
z='Fast Fourier Transform magnitude of sys3 response: y1';
title([ z ])
xlabel('Frequency: radians/second')
ylabel('Magnitude')

Warning: Data includes a number that is negative or zero. The LOG

of hits results in NaN or Infinity and is not shown on plot.

2-28
 More Sophisticated SYSTEM Functions

response of sys3 (solid) to input u1 (dashed)

0.5

0.4

0.3

0.2

0.1

-0.1

-0.2

-0.3

-0.4
0 20 40 60 80 100 120

Time: seconds

Figure 2-1 System Response Time Histories

Fast Fourier Transform magnitude of sys3 response: y1

10 3

10 2

10 1
Magnitude

10 0

10 -1

10 -2

10 -3
10 -2 10 -1 10 0 10 1 10 2

Frequency: radians/second

Figure 2-2: FFTs of the Output

2-29
2 Working with the Toolbox

The µ-Tools function vplot displays this warning message since there is a data
value at zero frequency that cannot be plotted on the log frequency scale.
The Signal Processing Toolbox provides a means of performing spectral
analysis with the function spectrum. The µ-Tools function vspect operates in
a similar manner on VARYING matrices. Given a signal x and a signal y,
vspect can calculate the power spectral density of x (Pxx), the power spectral
density of y (Pyy), the cross spectral density (Pxy), the transfer function from x
to y (Txy), and the coherence (Cxy). The VARYING matrix result will have the
following five columns, [Pxx, Pyy, Pxy, Txy, Cxy]. The command vspect(x,m) will
calculate the power spectral density of each element of the VARYING matrix,
x, using averaged FFTs of length m. The algorithm is exactly that used for
spectrum. See the Signal Processing Toolbox for further information.

The calculation of transfer functions using vspect restricts the input x to be a

one-by-one VARYING matrix and the outputs y to be an nr × 1 VARYING
matrix. This corresponds to being able to do single-input, multiple-output
(SIMO) identification experiments for a system with nr outputs. Each row of
the result then corresponds to an output or its associated SISO transfer
function. The calling sequence for vspect is vspect(x,y,m).
In the following example, vspect estimates the single-input, two-output
transfer function from the data generated in the previous example. A small
amount of random noise is added to the output, y1, to make the problem more
realistic. 512 point FFTs are applied to the data with an overlap of 256 points.
An optional Hamming window, the fifth input argument, is used on the data in
conjunction with the FFTs. The fourth column of the output is the estimated
transfer function, which is displayed in Figure 2-3 using the sel function.
noise1 = siggen('0.05*(rand(2,1)-[0.5;0.5])',time);
y1meas = madd(y1,noise1);
P1 = vspect(u1,y1,512,256,'hamming');
3 hamming windows in averaging calculation
minfo(P1)

varying: 256 pts 2 rows 5 cols

vplot('bode',sel(P1,[1:2],4),'-',frsp(sys3,omega),'-.')
tmp1 = 'sys3 (dash-dot) and estimated';
tmp2 = ' transfer function (solid)';
title([ tmp1 tmp2 ])

2-30
 More Sophisticated SYSTEM Functions

sys3 (dash-dot) and estimated transfer function (solid)

10 1

Log Magnitude
10 0

10 -1

10 -2

10 -3
10 -1 10 0 10 1 10 2
Frequency (radians/sec)

4
Phase (radians)

-2

-4
10 -1 10 0 10 1 10 2
Frequency (radians/sec)

Figure 2-3: Estimated Transfer Functions of the System

The command fitsys can be used to construct a state-space realization of the

estimated transfer function data. The first input argument to fitsys is the
frequency response data as a VARYING matrix. This argument, frdata, can
have dimension of a single-input/multi-output or multiple-input/single-output
VARYING matrix. The second input argument, ord, is the state-order of the
desired fit. The third input argument, weight is a weighting matrix with the
same independent variable values as frdata. The fourth input argument, code,
can be set to restrict the curve fitting algorithm to stable transfer functions.
The default value for weight is 1, and the default value of code is 0, placing no
restriction on the location of the rational fit’s poles. The output of fitsys is the
SYSTEM sys. The state-order of sys is ord.
In this example fit the single-input, two-output estimated transfer function
data with a third order model. Recall that the fourth column of P is the
estimated transfer function data. We will restrict the fitting algorithm to
consider only frequency points between 0.1 rad/sec and 10 rad/sec. A plot of the
estimated transfer function data and the third order model, sysord3, is shown
in Figure 2-4. Notice that the poles of sysord3 are very close to the poles of
sys3, which was used to create the estimated transfer function data.

2-31
2 Working with the Toolbox

estdata = sel(P1,[1:2],4);
sysord3 = fitsys(xtract(estdata,0.1,10),3);
vplot('bode',estdata,'-',frsp(sysord3,estdata),'-.')
tmp1 = 'sysord3 (dashed) and estimated';
tmp2 = ' transfer function (solid)';
title([ tmp1 tmp2 ])
rifd(spoles(sysord3))

real imaginary frequency damping

–3.1841e–01 –0.0000e+00 3.1841e–01 1.0000e+00

–1.1555e-01 –9.8732e–01 9.9406e–01 1.1624e–01
–1.1555e-01 9.8732e–01 9.9406e–01 1.1624e–01

rifd(spoles(sys3))

real imaginary frequency damping

–2.0000e–01 0.0000e+00 2.0000e–01 1.0000e+00

1.0000e–01 –9.9499e–01 1.0000e+00 1.0000e–01
–1.0000e–01 9.9499e–01 1.0000e+00 1.0000e–01

2-32
 More Sophisticated SYSTEM Functions

sysord3 (dashed) and estimated transfer function (solid)

2
Log Magnitude 10

0
10

2
10

4
10 1 0 1 2
10 10 10 10
Frequency (radians/sec)

0
Phase (degrees)

200

400 1 0 1 2
10 10 10 10
Frequency (radians/sec)

Figure 2-4: Estimated Transfer Functions and Third Order Model

2-33
2 Working with the Toolbox

Interconnection of SYSTEM Matrices: sysic

µ-Tools has a simple linear interconnection program called sysic. sysic forms
linear interconnections of CONSTANT and SYSTEM matrices (or CONSTANT
and VARYING matrices) simply by calculating the loop equations of the
interconnection.
Using sysic involves setting up several variables in the MATLAB workspace,
and then running the script-file sysic. The defined variables delineate the
details of the interconnection.
In order to explain the meaning of the sysic commands, consider a three-input,
two-output SYSTEM matrix T,

y1  

noise

y2 
T deltemp
 setpoint

which has internal structure

y1  57.3
  wt  deltemp

y2  g, p
6+ g6++  act 

noise

-
- k
setpoint

2-34
 Interconnection of SYSTEM Matrices: sysic

Variable Descriptions
Following are descriptions of the variables required by sysic.

systemnames
This variable is a character string, which contains the names of the matrices
(ie., the subsystems) used in the interconnection. The names must be separated
by spaces and/or tabs, and there should be no additional punctuation. The
order in which the names appear is not important. Each named system must
exist in the MATLAB workspace at the time the program sysic is run.
For the interconnection shown, with four components, k, p, act, and wt, the
following is an appropriate definition for the variable systemnames.
systemnames = ' k p act wt ';

The name of SYSTEM variables used within the sysic program is limited to 10
characters. This limitation is due to the MATLAB 19 character limitation on
the workspace variable names.

inputvar
This variable is a character string, with names of the various external inputs
that are present in the final interconnection. The input names are separated
by semicolons, and the entire list of input names is enclosed in square brackets
[ ]. Inputs can be multivariable signals, for example a windgust input with
three directions (x, y, and z) is specified by using windgust{3}. This indicates
three-variable input to the interconnection called windgust. Alternatively, this
could be specified as three separate, scalar inputs, say, wingustx, windgusty,
and windgustz. The order that the input names appear in the variable
inputvar is the order that the inputs are placed in the interconnection.

This simple interconnection has three external scalar inputs: sensor noise,
temperature disturbance, and a reference input.
inputvar = '[ noise; deltemp; setpoint]';

outputvar
This variable is a character string, describing the external outputs of the
interconnection, which must be linear combinations of the subsystem outputs
and the external inputs. Semicolons separate the channels of the output
variables. Between semicolons, signals can be added and subtracted, and

2-35
2 Working with the Toolbox

multiplied by scalars. For multivariable subsystems, arguments within

parentheses specify which subsystem outputs to use and in what order. For
instance, plant(2:5,8,1,9:11) specifies outputs 2, 3, 4, 5, 8, 1, 9, 10,
11 from the system plant. If no arguments are specified with a system, then it
is assumed that all outputs are being used, and in the order they appear in that
system.
In this example, the two outputs of the interconnection consist of the first
output of the plant (scaled by 57.3 to change units from radians to degrees)
along with a tracking error, which is the difference between the setpoint input
and the second plant output.
outputvar = '[ 57.3*p(1); setpoint - p(2) ]';

input_to_sys
This variable denotes the inputs to a specific system. Each subsystem named
in the variable systemnames must have a variable set to define the inputs to the
subsystem. If the system name is controller, then call the variable that must
be set using input_to_controller. Specify it in the same manner that the
variable outputvar is set, with inputs consisting of linear combinations of
subsystem outputs and external inputs. Separate channels are separated by
semicolons, and the order of the inputs in the variable should match the order
of the inputs in the system itself.
Corresponding to the systemnames variable set above, there are four input_to_
statements required, which are
input_to_k = '[ noise + p(2); setpoint ]';
input_to_act = '[ k ]';
input_to_wt = '[ deltemp ]';
input_to_p = '[ wt; act ]';

This means that the input to the controller consists of the sensor noise plus the
second output of the plant, and the reference input. The input to the actuator
is the output of the controller. The input to the weighting function is the
temperature disturbance, and the input to the plant consists of the output of
the weighting function, followed by the output of the actuator.

sysoutname
This character string variable is optional. If it exists in the MATLAB
workspace when sysic is run, the interconnection that is created by sysic is

2-36
 Interconnection of SYSTEM Matrices: sysic

placed in a MATLAB variable whose name is given by the string in

sysoutname. If this variable does not exist in the workspace, then the
interconnection is automatically placed in the variable ic_ms.
The command line
sysoutname = 'T';

will cause sysic to store the final interconnection in a SYSTEM matrix called
T.

cleanupsysic
This variable is used to clean up the workspace. After running sysic, all of the
above variables that describe the interconnection are left in the workspace.
These will be automatically cleared if the optional variable cleanupsysic is set
to the character string yes. The default value of the variable is no, which does
not result in any of the user-defined sysic descriptions being cleared. The
MATLAB matrices listed in the variable systemnames are never automatically
cleared.

Running sysic
If the variables systemnames, inputvar, and outputvar are set, and for each
name name_i appearing in systemnames, the variable input_to_name_i is set,
then the interconnection is created by running the M-file sysic. Depending on
the existence/nonexistence of the variable sysoutname, the resulting
interconnection is stored in a user-specified MATLAB variable or the default
MATLAB variable ic_ms.
Within sysic, error-checking of the consistency and availability of subsystem
matrices and their inputs aid in debugging faulty sysic interconnection
descriptions.
The input/output dimensions of the final interconnection are defined by
inputvar and outputvar variables.

Returning to the initial example, the following sysic commands were used to
generate the three-input, two-output SYSTEM matrix clp. (Note that the
dimensions of the variables k, p, act, and wt must be consistent with the
problem description.)

2-37
2 Working with the Toolbox

systemnames = ' k p act wt ';

inputvar = '[ noise; deltemp; setpoint]';
outputvar = '[ 57.3*p(1); setpoint - p(2) ]';
input_to_k = '[ noise + p(2); setpoint ]';
input_to_act = '[ k ]';
input_to_wt = '[ deltemp ]';
input_to_p = '[ wt; act ]';
sysoutname = 'clp'; cleanupsysic = 'yes';
sysic;

The syntax of sysic is limited, and for the most part restricted to what is
shown here. Some additional features are illustrated in the more complicated
demonstration problems.

HIMAT Design Example

The HIMAT example provides another example of how to construct
interconnection systems from block diagram descriptions. The HIMAT plant
model is described in more detail in the section, “HIMAT Robust Performance
Design Example” in Chapter 7. The interconnection diagram shown in
Figure 2-5 corresponds to the HIMAT design example.

Figure 2-5: HIMAT Interconnection

Given that there are four SYSTEM matrices, named himat, wdel, wp, and k, in
the MATLAB workspace, each with two inputs and two outputs, the following
10 lines form the sysic commands to make the interconnection structure
shown below, which is placed in the variable clp. These can be executed at the

2-38
 Interconnection of SYSTEM Matrices: sysic

command line (as shown) or placed in a script file. The command mkhimat needs
to be run initially to create himat, wdel, and wp.
mkhimat
himatic
k = zeros(2,2);
systemnames = ' himat wdel wp k ';
inputvar = '[ pertin(2) ; dist(2) ]';
outputvar = '[ wdel ; wp ]';
input_to_himat = '[ k + pertin ]';
input_to_wp = '[ dist + himat ]';
input_to_wdel = '[ k ]';
input_to_k = '[ -dist - himat ]';
sysoutname = 'clp';
cleanupsysic = 'yes';
sysic;

The final interconnection structure is located in clp with two sets of inputs,
pertin and dist, and two sets of outputs w and e, corresponding to the
perturbation and error outputs.

w1 ; w2   pertin

e1 ; e 2
 clp  dist

2-39
2 Working with the Toolbox

2-40
 3
H Control and
Model Reduction
3 H∞ Control and Model Reduction

This chapter covers an introduction to control H ∞analysis and design,

sampled-data control, and model reduction.

Optimal Feedback Control

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, and H∞, and optimal control. Hence, it
is important to develop a clear understanding of 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 Figure 3-1. K is some
controller to be designed and G is the system we want to control.

control
input 6 external force
disturbance
- -
reference
K - G - e?, - tracking
- error

?
e noise

Figure 3-1 Typical Closed-Loop Performance Objectives

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,

3-2
 Optimal Feedback Control

let T denote the closed-loop mapping from the outside influences to the
regulated variables,

reference
tracking error = T
external force
control input
noise
regulated variables
outside influences

We can assess performance by measuring the gain from outside influences to

regulated variables. In other words, good performance is associated with T
being small. Since the closed-loop system is a multi-input, multi-output
(MIMO) dynamical system, there are two different aspects to the gain of T:

• Spatial (vector disturbances and vector errors)

• Temporal (dynamical relationship between input/output signals)

To quantify the term gain mathematically, we need to define some additional

things.

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

1
---
∞ 2
e 2 :=  ∫ e ( t ) dt
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 )

3-3
3 H∞ Control and Model Reduction

the 2-norm is defined as

1
---
∞ 2
∫
2
e 2 := ( e(t) 2 dt )
–∞ 1
---
∞ T 2
= ( ∫ –∞
e ( t )e ( t )dt )

In µ-Tools the dynamical systems we deal with are exclusively linear, with
state-space model

x· = A B x
e C D d

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,

1
1 ∞ ---
∫
2
T 2 := ------ T ( jω ) F dω 2 T ∞ := maxσ [ T ( jω ) ]
2π – ∞ w∈R

where the Frobenious 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
e C D d

3-4
 Optimal Feedback Control

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 -----------
d≠0 d 2

is equal to ||T||∞. This is discussed in greater detail in the next section.

Using Weighted Norms to Characterize

Performance
In any performance criterion, we 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||
where the weighting function matrices WL and WR are frequency dependent, to
account for bandwidth constraints and spectral content of exogenous signals.
Within the structured singular value setting considered in Chapter 4, the most
natural (mathematical) manner to characterize acceptable performance is in
terms of the MIMO ||⋅||∞ (H∞) norm. For this reason, we discuss some
interpretations of the H∞ norm.

e~ T  d~
Figure 3-2: Unweighted MIMO System

Suppose T is a MIMO stable linear system, with transfer function matrix T(s).
For a given driving signal d̃ ( t ) , define ẽ as the output, as shown in Figure 3-2.

3-5
3 H∞ Control and Model Reduction

Note that it is more traditional to write the diagram in Figure 3-2 with the
arrows going from left to right as in Figure 3-3.

d~ - T -e~
Figure 3-3: Unweighted MIMO System: Vectors from Left to Right
The diagrams in Figure 3-2 and Figure 3-3 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 β > 0 be defined as

β := T ∞ := maxσ [ T ( jω ) ]
w∈R (3-1)

Now consider a response, starting from initial condition equal to 0. In that case,
Parseval’s theorem gives that

∞ T 1⁄2
ẽ 2 ∫
[ ẽ ( t )ẽ ( t ) dt ]
0
----------- = ----------------------------------------------------- ≤ β
∞ T 1⁄2
d̃ 2
∫0 d̃ ( t )d̃ ( t ) dt

ẽ 2
Moreover, there are specific disturbances d that result in the ratio -----------
d̃ 2

arbitrarily close to β. 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 ω ∈ R , any vector of amplitudes a ∈ R n d , and any
vector of phases φ ∈ R n d , with ||a||2 ≤ 1, define a time signal

a 1 sin ( ωt + φ 1 )
d̃ ( t ) =
…

a n d sin ( ωt + φ n d )

3-6
 Optimal Feedback Control

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

form
b 1 sin ( ωt + 1 )
ẽ ss ( t ) =

…
b n e sin ( ωt + n e )

The vector b ∈ R n e will satisfy ||b||2 ≤ β. Moreover, β, as defined in equation

Figure 3-1, is the smallest number such that this fact is true for every ||a||2 ≤ 1,
ω , and φ.
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. Since 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.
Consider the diagram of Figure 3-4, along with Figure 3-2.
Assume that WL and WR are diagonal, stable transfer function matrices, with
diagonal entries denoted Li and Ri.

L1 0 … 0 R1 0 … 0
0 L2 … 0 0 R2 … 0
WL = , WR =
.. ..
…
…

.
…
…
…

.
…

0 0 … Ln 0 0 … Rn
e d

e WL  e~ T  d~ WR d
e = WL e~ = WL T d~ = WL T WR d

Figure 3-4: Weighted MIMO System

3-7
3 H∞ Control and Model Reduction

Bounds on the quantity ||WLTWR||∞ will imply bounds about the sinusoidal
steady-state behavior of the signals d̃ and ẽ(= Td̃) in Figure 3-2. Specifically,
for sinusoidal signal d̃ , the steady-state relationship between ẽ(= Td̃) , d̃ and
||WLTWR||∞ is as follows: The steady-state solution ẽ ss , denoted as

ẽ 1 sin ( ωt + ϕ 1 )
ẽ ss ( t ) =

…
ẽ n sin ( ωt + ϕ n )
e d
(3-2)

∑
ne 2
satisfies W L ( jw )ẽ i ≤ 1 for all sinusoidal input signals d̃ of the form
i=1 i

d̃ 1 sin ( ωt + φ i )
d̃ ( t ) =
…

d̃ n d sin ( ωt + φ n d )
(3-3)

satisfying

nd 2
d̃ i
∑ --------------------------- ≤ 1
W R i ( jω )
2
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 ω ,
and all sinusoidal disturbances d̃ of the form (3-3) satisfying

d̃ i ≤ W R ( jω )
i

the steady-state error components will satisfy

1
ẽ i ≤ ------------------------
W L ( jω )
i

3-8
 Optimal Feedback Control

This shows how one could pick performance weights to reflect the desired
frequency-dependent performance objective. Use WR to represent the relative
1
magnitude of sinusoids disturbances that might be present, and use ---------
W
to
L
represent the desired upper bound on the subsequent errors that are produced.
Remember, though, the weighted H∞ norm does not actually give element-
by-element bounds on the components of ẽ based on element-by-element
bounds on the components of d̃ . The precise bound it gives is in terms of
Euclidean norms of the components of ẽ and d̃ (weighted appropriately by
WL(j ω ) and WR(j ω )).

3-9
3 H∞ Control and Model Reduction

Interconnection with Typical MIMO Performance Objectives

Throughout this manual, we formulate closed-loop performance objectives as
weighted closed-loop transfer functions which are to be made small through
feedback. A generic example, which includes many relevant terms, is shown in
block diagram form in Figure 3-5. In the diagram, G denotes the plant model
and K is the feedback controller.

Figure 3-5: Generalized and Weighted Performance Block Diagram

The blocks in Figure 3-5 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 d̃ and ẽ is suitable.

3-10
 Interconnection with Typical MIMO Performance Objectives

This shows the interpretation of the signals, weighting functions and models.

Signal Meaning

d1 Normalized reference command

d̃ 1 Typical reference command

d2 Normalized exogenous disturbances

d̃ 2 Typical exogenous disturbances

d3 Normalized sensor noise

d̃ 3 Typical sensor noise

e1 Weighted control signals

ẽ 1 Actual control signals

e2 Weighted tracking errors

ẽ 2 Actual tracking errors

e3 Weighted plant errors

ẽ 3 Actual plant errors

Wcmd
Wcmd is used in problems requiring tracking of a reference command. Wcmd
shapes (magnitude and frequency) the normalized reference command signals
into the actual (or typical) reference signals that we 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 can (and will) generate stick input reference commands
up to a bandwidth of about 2Hz. 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

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

3-11
3 H∞ Control and Model Reduction

Wmodel
represents a desired ideal model for the closed-looped system, used for
problems with tracking requirements. For good command tracking response,
we might desire our closed-loop system to respond like a well-damped
second-order system. The ideal model would then be

2
ω
W model = 10 -------------------------------------
2 2
-
s + 2ζω + ω

for specific desired natural frequency ω and desired damping ratio ζ. Unit
conversions might be necessary too. 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
·
2
ω
W model = 10 -------------------------------------
2 2
-
s + 2ζω + ω

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 the ground
excitations, sound (pressure) waves, and air currents. The spectrum and
relative magnitudes of these disturbances are captured in the transfer function
weighting matrix Wdist.

Wperf1
Wperf1 weights the difference between the response of the plant and the
response of the ideal model, Wmodel. Often we desire 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
inverse of the weight should be related to the allowable size of tracking errors,
in the face of the reference commands and disturbances described by Wref and
Wdist.

3-12
 Interconnection with Typical MIMO Performance Objectives

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, in the face of 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 system, and level out at high frequency.

Hsens
Hsens represents a model of the sensor dynamics or an external anti-aliasing
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 using this block diagram description.
In Chapter 4, we see how to incorporate uncertainty into the model of G (and
possibly Hsens as well), and how to analyze the implications on performance due
to uncertainty. Chapter 7 presents a number of examples, which explain in
detail how individual performance weighting functions are selected.

3-13
3 H∞ Control and Model Reduction

Commands to Calculate the H2 and H∞ Norm

There are five µ-Tools functions to calculate the H2 and H∞ norm.
dhfnorm h2norm hinfnorm pkvnorm vnorm

H2 norm
The H2 norm of a stable, strictly proper continuous-time SYSTEM matrix can
be calculated using the command h2norm. Its calling sequence is
out = h2norm(sys)

The output variable, out, is a scalar whose value is the two-norm of the
SYSTEM sys. Given a state-space description of a system as

sys = A B
C D

The H2 norm of the SYSTEM follows from the solution to the Lyapunov
equation
AX + XA´ + BB´ = 0.
with sys 2 = [ tr ( CXC )' ] .

H∞ norm
The H∞ norm of a stable, continuous-time SYSTEM, sys, can be calculated
using the command hinfnorm. Its calling sequence is
out = hinfnorm(sys,tol)

The output from hinfnorm is a 1 × 3 vector, out, which is made up (in order) of
a lower bound for ||sys||∞, an upper bound for ||sys||∞, and a frequency, ωo, at
which the lower bound is achieved.

out(1) = σ ( sys ( j Þ out ( 3 ) ) ) ≤ sys ∞ ≤ out(2)

The||⋅||∞norm calculation is an iterative process and requires a test to stop. See

the hinfnorm manual page in Chapter 8, “Reference” for more details. If the
first input argument to hinfnorm, sys, is a VARYING matrix, then hinfnorm

3-14
 Commands to Calculate the H2 and H∞ Norm

calls pkvnorm to find the maximum singular value of the VARYING matrix
across frequency.
The H∞ norm of a frequency VARYING matrix, sysg, can be calculated using
pkvnorm or vnorm. The calling sequences are
[peak,indv,index] = pkvnorm(matin)
out = vnorm(matin)

pkvnorm sweeps through the independent variable and calculates the largest
singular value of matin. The three output arguments all pertain to the peak
norm across frequency and its location: peak value, peak, the independent
variable’s value, indv, and the independent variable’s index, index.
vnorm is a VARYING matrix version of MATLAB’s norm command. The
operation is identical, except that it also works on CONSTANT and VARYING
matrices, producing a CONSTANT or VARYING output. vnorm returns the
matrix out with its norm at each independent variable value. The default is the
largest singular value of matin at each independent variable value.

Discrete-Time H∞ Norm
The H∞ norm of a discrete-time SYSTEM can be calculated using the command
dhfnorm. Its calling sequence is
out = dhfnorm(sys)

The first input argument, sys, can be either a discrete-time SYSTEM, a

CONSTANT or VARYING matrix. The output of dhfnorm, out, is a 1 × 3 vector
giving a lower bound, upper bound of the discrete-time H∞ norm and the
frequency where the lower bound occurs.

3-15
3 H∞ Control and Model Reduction

Commands to Design H∞ Output Feedback Controllers

Given a linear system P with four types of external variables:

• Exogenous disturbances (d)

• Regulated variables, i.e., errors (e)
• Manipulated variables, i.e., controls (u)
• Sensed variables, i.e., measurements (y)

These are related through the linear state-space equations

A B1 B2
P = C 1 D 11 D 12 = A B = C ( sI – A ) B + D
–1
CD
C 2 D 21 D 22

The H∞ output feedback control design problem is: Does there exist a linear
controller, K, with internal structure

AK BK
K =
CK DK

such that the closed-loop system e = FL(P, K)d,

e d
P 
y
-K u
is stable and the ∞-norm of FL(P, K) is less than γ? Note that the above block
diagram represents a linear fractional transformation (LFT). LFTs are
described in more detail in the “Representing Uncertainty” section in Chapter
4. The LFT equation FL(P,K) is given by P11 + P12K(I – P22K)–1P21.
The standard state-space technique to calculate H∞ output feedback
controllers is to select a value of γ and determine if there exists a controller K
such that ||FL(P,K)||∞ < γ. This value of γ is updated based on a modified

3-16
 Commands to Design H∞ Output Feedback Controllers

bisection algorithm, called γ iteration. This iteration procedure continues until

the magnitude of the difference between the smallest γ value that has passed
and the largest γ value that has failed is small.
Two functions are included in µ-Tools to synthesize continuous-time H∞
controllers. hinfsyn and hinfsyne calculate the H∞ output feedback
(sub)optimal controller. The hinfsyn command constructs the standard
centralized H∞ controller whereas the hinfsyne command constructs the H∞
(sub)optimal controller that minimizes the entropy integral at a specific
frequency. The default in hinfsyne is to minimize the entropy integral at ∞.
Both hinfsyn and hinfsyne commands implement the full general output
feedback equations based on the interconnection structure P. The complete set
of equations and results for the H∞ control problem can be found in the “H•
Output Feedback” section of this chapter and in references [DoyGKF] and
[GloD].
The following assumptions are made about the open-loop system P in the
hinfsyn and hinfsyne commands:

(A1) (A,B2) is stabilizable and (C2,A) is detectable.

(A2) D12 is full column rank and D21 is full row rank.

A – jωI B 2
(A3) has full column rank for all ω.
C 1 D 12

A – jωI B 1
(A4) has full row rank for all ω.
C 2 D 21

hinfsyn and hinfsyne return the H∞ controller, the closed-loop system, and
the γ level achieved.
The hinfsyn and hinfsyne programs provide a γ iteration using a modified
bisection method. You select a high and low value of γ, gamma_max and
gamma_min. The bisection method iterates on the value of γ in an effort to
approach the optimal H∞ control design. If the value gamma_max equals
gamma_min, only one γ value is tested. The bisection algorithm stops when the
difference between the smallest value of γ that has passed and the largest value
of γ that has failed is less than tol.

3-17
3 H∞ Control and Model Reduction

The calling sequence for the hinfsyn command is

[k,clp,gfin] =
hinfsyn(p,nmeas,ncon,gamma_min,gamma_max,tol)

p is the SYSTEM interconnection structure, nmeas is the number of

measurements, ncon is the number of control inputs, gamma_min and gamma_max
are the minimum and maximum γ values, and tol is the difference between
final γ values. The output argument k is an H∞ (sub)optimal controller, clp is
the closed-loop system with H∞ controller implemented (clp = starp(p,k)),
and gfin is the final γ value associated with k and clp. See the manual pages
for hinfsyn in Chapter 8, “” for more information. The calling sequence for the
command is identical to.

H∞ Design Example
The objective is to design an H∞ (sub)optimal control law for SYSTEM
interconnection structure given by the block diagram in Figure 3-6. The
HIMAT plant model and weightings are described in more detail in the
“HIMAT Robust Performance Design Example” section in Chapter 7.

Figure 3-6: H∞ Design Example Interconnection Structure

The SYSTEM interconnection structure is located in himat_ic. It consists of

two sensor measurements, two error signals, two actuator inputs, two
disturbance inputs, and eight states. The range of γ is selected to be between
1.0 and 10.0 with a tolerance, tol, on the relative closeness of the final γ
solution of 0.1. For each iteration, the program prints the current γ value being
tested, and the results of five tests for the existence of a controller achieving
the closed-loop norm objective. At the end of each iteration a (p) or (f) is

3-18
 Commands to Design H∞ Output Feedback Controllers

displayed denoting that the γ value either passed or failed. Upon finishing,
hinfsyn and hinfsyne print out the lowest γ value achieved. A # sign is used in
the printout to denote which of the five conditions for the existence of an H∞
(sub)optimal controller failed.
nmeas = 2;
ncont = 2;
gmn = 1;
gmx = 10;
tol = 0.1;
mkhimat
himatic
minfo(himat_ic)
system: 8 states 6 outputs 6 inputs
p = himat_ic;
[k,clp] = hinfsyn(p,ncont,nmeas,gmn,gmx,tol);
Test bounds: 1.0000 < gamma<=10.0000

gamma hamx_eig xinf_eig hamy–eig yinf–eig nrho–xy p/f

10.000 2.3e–02 1.2e–07 2.3e–02 –1.5e–11 0.0223 p
5.5000 2.3e–02 1.3e–07 2.3e–02 0+0e+00 0.0747 p
3.250 2.3e–02 1.3e–07 2.3e–02 0+0e+00 0.2222 p
2.125 2.3e–02 1.3e–07 2.3e–02 0+0e+00 0.5642 p
1.562 2.3e–02 1.4e–07 2.3e–02 0+0e+00 1.1977# f
1.711 2.3e–02 1.4e–07 2.3e–02 0+0e+00 0.9474 p
1.677 2.3e–02 1.4e–07 2.3e–02 –3.0e–14 0.9973 p
1.654 2.3e–02 1.4e–07 2.3e–02 0+0e+00 1.0328# f

Gamma value achieved:1.6770

We can verify that the closed-loop system achieved an H∞ norm of 1.6770 by

calculating the H∞ norm using the hinfnorm command.
hinfnorm(clp)
norm between 1.667 and 1.679
achieved near 3.814

3-19
3 H∞ Control and Model Reduction

H∞ Optimal Control Theory

This part of the H∞ chapter presents simple state-space formulae for all
controllers solving a standard H∞ problem: For a given number γ ≥ 0, find all
stabilizing controllers such that the H∞ norm of the closed-loop transfer
function is less than γ. Under these conditions, a parametrization of all
controllers solving the problem is given as a linear fractional transformation
(LFT) on a contractive, stable free parameter. The state dimension of the
coefficient matrix for the LFT equals that of the plant, and has a separation
structure reminiscent of classical LQG (i.e., H2) theory. The results are
essentially from reference [GloD2] with proofs removed. This directly
generalizes the results in reference [DoyGKF] and [GloD]. It is assumed that
the reader has at least read [DoyGKF] before attempting to read the material
to follow.
Two popular performance measures in optimal control theory are H2 and H∞
norms. Recall that they defined in the frequency domain for a stable transfer
matrix P(s) as

∞ 1⁄2
P 2 :=  ------ trace [ P ( jω )*P ( jω ) ]dω
1
∫
 2π – ∞ 
P ∞ := sup σ([)P ( jω )] (σ ( : ) = maximum singular value)
ω

The former arises when the exogenous signals either are fixed or have a fixed
power spectrum; the latter arises from (weighted) balls of exogenous signals.
H2-optimal control theory was heavily studied in the 1960’s as the Linear
Quadratic Gaussian (LQG) optimal control problem; H∞-optimal control theory
is continuing to be developed. We assume the reader either is familiar with the
engineering motivation for these problems, or is interested in the results of this
chapter for some other reason.
The basic block diagram used in this chapter is

e d
P 
y
-K u
where P is the generalized plant and K is the controller. Only finite
dimensional linear time-invariant (LTI) systems and controllers will be
considered. The generalized plant P contains what is usually called the plant

3-20
 H∞ Optimal Control Theory

in a control problem plus all weighting functions. The signal d contains all
external inputs, including disturbances, sensor noise, and commands, the
output e is an error signal, y is the measured variables, and u is the control
input. The diagram is also referred to as a linear fractional transformation
(LFT) on K and P is called the coefficient matrix for the LFT. The resulting
closed loop transfer function from d to e is denoted by Ted = FL(P,K).
The main H∞ output feedback results are presented in the “H• Output
Feedback” section. The proofs of these results exploit the separation structure
of the controller. If perfect measurements of the states (x) and the disturbances
(d) are available (this is defined as the Full Information problem), then the
central controller is simply a gain matrix F∞, obtained through finding a
certain stable invariant subspace of a Hamiltonian matrix. Also, the optimal
output estimator is an observer whose gain is obtained in a similar way from a
dual Hamiltonian matrix. These special cases are described in the “H• Full
Information and Full Control Problems” section. In the general output
feedback case the controller can be interpreted as an optimal estimator for F∞x.
Furthermore, the two Hamiltonians involved in this solution can be associated
with full information and output estimation problems.
As mentioned, this material is taken primarily from [GloD2], which is a direct
generalization of [DoyGKF], and contains a substantial repetition of material.
Roughly speaking, [GloD2] proves those results in [GloD] which were stated
without proof, using [DoyGKF] machinery, which considered a less general
problem. An alternative approach in relaxing some of the assumptions in
[DoyGKF] is to use loop-shifting techniques as in [ZhouK], [GloD], and more
completely in [SafLC]. We also consider some aspects of generalizations to the
≤ case, primarily to indicate the problems encountered in the optimal case. A
detailed derivation of the necessity of the generalized conditions for the Full
Information problem is given. In keeping with the style of [GloD] and
[DoyGKF], we don’t present a complete treatment of the ≤ case. Complete
derivations of the optimal output feedback case can be found in [GlovM] using
different techniques.

3-21
3 H∞ Control and Model Reduction

Historical Perspective
This section is not intended as a review of the literature in H∞ theory, but
rather an attempt to outline some of the work that led up to and most closely
touches on [DoyGKF], [GloD], and [GloD2]. Control, history For a more
extensive bibliography and review of earlier literature, the interested reader
might see [Fran1] and [FranD].
Zames’ [Zame] original formulation of H∞ optimal control theory was in an
input-output setting. Most solution techniques available at that time involved
analytic functions (Nevanlinna-Pick interpolation) or operator-theoretic
methods [Sara], [AdAK], and [BallH]. An earlier state-space solution was
presented in [Doy1], in which the steps were as follows: parametrize all
internally stabilizing controllers via Youla [YouJB]; obtain realizations of the
closed-loop transfer matrix; convert the resulting model-matching problem into
the equivalent 2 × 2-block general distance or best approximation problem
involving mixed Hankel-Toeplitz operators; reduce to the Nehari problem
(Hankel only); and solve the Nehari problem by the procedure of [Glo1]. Both
[Fran1] and [FranD] give expositions of this approach, which will be referred
to as the “1984” approach.
In a mathematical sense, the 1984 procedure solved the general rational H∞
optimal control problem and much of the subsequent work in H∞ control theory
focused on the 2 × 2-block problems, either in the model-matching or general
distance forms. Unfortunately, the associated complexity of computation was
substantial, involving several Riccati equations of increasing dimension, and
formulae for the resulting controllers tended to be very complicated and have
high state dimension. Encouragement came from [LimH] who showed, for
problems transformable to 2 × 1-block problems, that a subsequent minimal
realization of the controller has state dimension no greater than that of the
generalized plant G. This suggested the likely existence of similarly low
dimension optimal controllers in the general 2 × 2 case.
Additional progress on the 2 × 2-block problems came from [BallC], who gave a
state-space solution involving three Riccati equations. [JonJ] showed a
connection between the 2 × 1-block problem. [FoisT] developed an interesting
class of operators called skew Toeplitz to study the 2 × 2-block problem. Other
approaches have been derived by [Hung] using an interpolation theory
approach, [Kwak] using a polynomial approach, and [Kim] using a method
based on conjugation.

3-22
 H∞ Optimal Control Theory

In addition to providing controller formulae that are simple and expressed in

terms of plant data, the methods in the present approach are a fundamental
departure from the earlier work described above. In particular, the Youla
parametrization and the resulting 2 × 2-block model-matching problem of the
1984 solution are avoided entirely; replaced by a pair of 2 × 1-block problems
and a separation argument. The entire development uses simple and familiar
tools, in the style of [Will1] relying on state feedback and observer-based
control methods and more straightforward and elegant use of operator theory.
Another strong influence on this work is Redheffer’s work [Red2] on linear
fractional transformations.
Independent encouragement for a simpler approach to the H∞ problem came
from papers by [KhaPZ] and [ZhoK]. They showed that for the state-feedback
H∞ problem one can choose a constant gain as a (sub)optimal controller. In
addition, a formula for the state-feedback gain matrix was given in terms of an
algebraic Riccati equation. These results are similar to those in the “H• Full
Information and Full Control Problems” section, though the proof techniques
are entirely different. Also, these papers established connections between H∞
-optimal control, quadratic stabilization, and linear-quadratic differential
games.
As expected, the results and techniques in [DoyGKF] and [GloD2] have
encouraged greater interest in applications of H∞ methods, in alternative
developments of the theory using other techniques, and in extensions to more
general problems. The state-space theory of H∞ can be carried much further,
by generalizing time-invariant to time-varying, infinite horizon to finite
horizon, and finite dimensional to infinite dimensional. A flourish of activity
has begun on these problems and the already numerous results indicate, not
surprisingly, that many of the results generalize mutatis mutandis, to these
cases.

Notation
⊥
The notation is fairly standard. The Hardy spaces H2 and H 2 consist of
square-integrable functions on the imaginary axis with analytic continuation
into, respectively, the right and left half-plane. The Hardy space H∞ consists
of bounded functions with analytic continuation into the right half-plane. The
Lebesgue spaces L2 = L2(–∞,∞), L2+ = L2[0,∞), and L2– = L2(–∞,0] consist
respectively of square-integrable functions on (–∞,∞), [0,∞), and (–∞,0], and
L∞ consists of bounded functions on (–∞,∞). As interpreted in this chapter, L∞

3-23
3 H∞ Control and Model Reduction

will consist of functions of frequency, L2+ and L2– functions of time, and L2 will
be used for both.
We will make liberal use of the Hilbert space isomorphism, via the Laplace
transform and the Paley-Wiener theorem, of L2 = L2+ ⊕ L2– in the time-domain
⊥
with L2 = H2 ⊕ H 2 in the frequency-domain and of L2+ with H2 and L2– with
⊥
H 2 . In fact, we will normally not make any distinction between a time-domain
signal and its transform. Thus we may write d ∈ L 2+ and then treat d as if
d ∈ H 2 . This style streamlines the development, as well as the notation, but
when any possibility of confusion could arise, we will make it clear whether we
are working in the time- or frequency-domain.
All matrices and vectors will be assumed to be complex. A transfer matrix in
terms of state-space data is denoted

A B := C ( sI – A ) –1 B + D
C D
p×r
For a matrix M ∈ C , M´ denotes its conjugate transpose,
σ ( M ) = ρ ( M′M ) 1 ⁄ 2 denotes its maximum singular value, ρ(M)denotes its
spectral radius (if p = r), and M† denotes the Moore-Penrose pseudo-inverse of
M. Im denotes image, ker denotes kernel, and P ~ ( s ) := P ( – s )′ . For operators,
Γ* denotes the adjoint of Γ. The prefix B denotes the open unit ball and the
prefix Rc denotes complex-rational.
⊥
The orthogonal projections P+ and P– map L2 to, respectively, H2 and H 2 (or
L2+ and L2–). For P ∈ L ∞ , the Laurent or multiplication operator M P : L 2 → L 2
for frequency-domain d ∈ L 2 is defined by MPd = Pd. The norms on L∞ and L2
in the frequency-domain were defined in the “Performance as Generalized
Disturbance Rejection” section. Note that both norms apply to matrix or
vector-valued functions. The unsubscripted norm || • || will denote the standard
Euclidean norm on vectors. We will omit all vector and matrix dimensions
throughout, and assume that all quantities have compatible dimensions.

3-24
 H∞ Optimal Control Theory

Problem Statement
Consider the system described by the block diagram

e d
P 
y
-K u
Both P and K are complex-rational and proper, and K is constrained to provide
internal stability. We will denote the transfer functions from d to e as Ted in
general and for a linear fractional transformation feedback connection as above
we also write Ted = FL(P,K). This section discusses the assumptions on P that
will be used. In our application we have state models of P and K. Then internal
stability will mean that the states of P and K go to zero from all initial values
when d = 0.
Since we will restrict our attention exclusively to proper, complex-rational
controllers that are stabilizable and detectable, these properties will be
assumed throughout. Thus the term controller will be taken to mean a
controller that satisfies these properties. Controllers that have the additional
property of being internally stabilizing will be said to be admissible. Although
we are taking everything to be complex, in the special case where the original
data is real (e.g., P is real-rational) then all of the results (such as K) will also
be real.
The problem to be considered is to find all admissible K(s) such that
||Ted||∞ < γ(≤ γ). The realization of the transfer matrix P is taken to be of the form

A B1 B2
P ( s ) = C 1 D 11 D 12 = A B
C D
C 2 D 21 D 22

p1 p2 m1
compatible with the dimensions e ( t ) ∈ C , y(t) ∈ C , d(t) ∈ C ,
m2 n
u(t) ∈ C , and the state x ( t ) ∈ C . The following assumptions are made:
(A1) (A,B2) is stabilizable and (C2,A) is detectable.

3-25
3 H∞ Control and Model Reduction

(A2) D12 is full column rank with D 12 D ⊥ unitary and D21 is full row rank

D 21
with unitary.
D̃ ⊥

A – jωI B 2
(A3) has full column rank for all ω.
C 1 D 12

A – jωI B 1
(A4) has full row rank for all ω.
C 2 D 21

Assumption (A1) is necessary for the existence of stabilizing controllers. The

assumptions in (A2) mean that the penalty on e = C1x + D12u includes a
nonsingular, normalized penalty on the control u, and that the exogenous
signal ω includes both plant disturbance and sensor noise, and the sensor noise
weighting is normalized and nonsingular. Relaxation of (A2) leads to singular
control problems.
Assumption (A3) relaxes the [DoyGKF] assumptions that (C1,A) is detectable
and D´12C1 = 0, and (A4) relaxes (A,B1) stabilizable and B´1D21 = 0.
Assumptions (A3) and (A4) are made for a technical reason: together with (A1)
it guarantees that the two Hamiltonian matrices in the corresponding H2
problem belong to dom(Ric). It is tempting to suggest that (A3) and (A4) can be
dropped, but they are, in some sense, necessary for the methods in this chapter
to be applicable. A further discussion of the assumptions and their possible
relaxation will be discussed in the “Relaxing Assumptions A1–A4” section.
It will be assumed in the rest of this chapter that D22 = 0. To see how to handle
the general case for D22 ≠ 0, suppose K is a stabilizing controller for D22 set to
zero, and satisfies

FL  P – ,K
0 0
<γ
 0 D 22  ∞

3-26
 H∞ Optimal Control Theory

then
FL(P,K(I + D22K)–1) = P11 + P12K(I + D22K – P22K)–1P21

 0 0 
= FL  P – ,K ⋅
 0 D 22 

Hence a controller K for

 0 0 
P– 
 0 D 22 

–1
yields a controller K̃ = K ( I + D 22 K ) for P. The µ-Tools commands hinfsyn
and hinfsyne handle the nonzero D22 case.
When D22 ≠ 0 there is a possibility of the feedback system becoming ill-posed
due to det(I +D22 K̃ ( ∞ ) ) = 0 (or more stringent conditions if we require
well-posedness in the face of infinitesimal time delays [Will1]). Such
possibilities need to be excluded.
It can be assumed, without loss of generality, that γ = 1 since this is achieved
by the scalings γ–1D11, γ–1/2B1, γ–1/2C1, γ1/2B2, γ1/2C2, and γ–1K. This will be
done implicitly for many of statements of this chapter.

Preliminaries
This section reviews some mathematical preliminaries, in particular the
computation of the various norms of a transfer matrix P. Consider the transfer
matrix

P(s) = A B
C D
(3-4)

with A stable (i.e., all eigenvalues in the left half-plane).

The norm ||P||∞ arises in a number of ways. Suppose that we apply an
input d ∈ L 2 and consider the output e ∈ L 2 . Then a standard result is that

3-27
3 H∞ Control and Model Reduction

||P||∞ is the induced norm of the multiplication operator MP, as well as the
Toeplitz operator P + M P : H 2 → H 2 .

P ∞ = sup e 2 = sup P + e 2 = sup P + M P d 2

d ∈ BL d ∈ BL d ∈ BH
2 2+ 2

The rest of this section involves additional characterizations of the norms in

terms of state-space descriptions. “The Riccati Operator” section collects some
basic material on the Riccati equation and the Riccati operator, which play an
essential role in the development of both theories.

The Riccati Operator

Let A, Q, R be complex n × n matrices with Q and R Hermitian. Define the 2n
× 2n Hamiltonian matrix

H := A R
Q – A′

If we begin by assuming H has no eigenvalues on the imaginary axis, then it

must have n eigenvalues in Re s < 0 and n in Re s > 0. Consider the two
n-dimensional spectral subspaces χ–(H) and χ+(H): the former is the invariant
subspace corresponding to eigenvalues in Re s < 0; the latter, to eigenvalues in
Re s > 0. Finding a basis for χ–(H), stacking the basis vectors up to form a
matrix, and partitioning the matrix, we get

X1
χ – ( H ) = Im
X2
(3-5)
n×n
where X 1 ,X 2 ∈ C , and

X1 X1
H = T X , Re λ i ( Tx ) < 0 ∀i
X2 X2
(3-6)

3-28
 H∞ Optimal Control Theory

If X1 is nonsingular, or equivalently, if the two subspaces

χ– ( H ) , I m 0
I
(3-7)

are complementary, we can set X := X2X1–1. Then X is uniquely determined by

H, i.e., H a X is a function, which will be denoted Ric; thus, X = Ric(H). We will
take the domain of Ric, denoted dom(Ric), to consist of Hamiltonian matrices
H with two properties, namely, H has no eigenvalues on the imaginary axis and
the two subspaces in equation (3-7) are complementary. For ease of reference,
these will be called the stability property and the complementary property,
respectively. The following well-known results give some properties of X as well
as verifiable conditions under which H belongs to dom(Ric). See, for example,
Section 7.2 in [Fran1], Theorem 12.2 in [Wonh], and [Kuc1].

Lemma 3.1. Suppose H Πdom(Ric) and X = Ric(H). Then

a X is Hermitian

b satisfies the algebraic Riccati equation

A´X + XA + XRX – Q = 0

c A + RX is stable

Lemma 3.2. Suppose H has no imaginary eigenvalues, R is either positive

semidefinite or negative semidefinite, and (A,R) is stabilizable. Then
H Πdom(Ric).

Lemma 3.3. Suppose H has the form

H = A – BB’
– C'C – A'

with (A,B) stabilizable and rank A + jωI C’ = n ∀ω. Then H ∈ dom(Ric),

X = Ric(H) ≥ 0, and ker(X) ⊂ χ := stable unobservable subspace.
By stable unobservable subspace we mean the intersection of the stable
invariant subspace of A with the unobservable subspace of (A,C). Note that if
(C,–A) is detectable, then Ric(H) ≥ 0. Also, note that ker(X) ⊂ χ ⊂ ker(C), so that

3-29
3 H∞ Control and Model Reduction

the equation XM = C´ always has a solution for M, for example the

least-squares solution given by X† C′ .

Computing the H∞ Norm

For the transfer matrix P(s) in equation (3-4), with A stable, define the
Hamiltonian matrix

–1 –1
H := A + BR D’C BR B’
–1 –1
– C’ ( I – DD’ ) C – ( A + BR D’C )’
(3-8)

= A 0 + B R–1
D′C B′
– C′C – A′ – C′D
(3-9)

where R = I – D´D. The following lemma is essentially from [And], [Will1], and
[BoyBK].

Lemma 3.4. Let σ ( D ) < 1 , then the following conditions are equivalent:

a ||P||
∞<1
b H has no eigenvalues on the imaginary axis

c H ∈ dom(Ric)
d H ∈ dom(Ric) and Ric(H) ≥ 0 (Ric(H) > 0 if (C,A) is observable)

Lemma 3.4 suggests the following way to compute an H∞ norm: select a

positive number γ; test if ||P||∞ < γ by calculating the eigenvalues of H; increase
or decrease γ accordingly; repeat. Thus H∞ norm computation requires a
search, over either γ or ω. We should not be surprised by similar characteristics
of the H∞-optimal control problem. A somewhat analogous situation occurs for
2
matrices with the norms M 2 = trace ( M*M ) and M ∞ = σ [ M ]. In
2
principle, M 2 can be computed exactly with a finite number of operations, as
can the test for whether σ ( M ) < γ (e.g., γ2I – M*M > 0), but the value of σ ( M )
cannot. To compute σ ( M ) we must use some type of iterative algorithm.

3-30
 H∞ Full Information and Full Control Problems

H∞ Full Information and Full Control Problems

In this section we discuss four problems from which the output feedback
solutions will be constructed via a separation argument. These special
problems are central to the whole approach taken in this chapter, and as we
shall see, they are also important in their own right. All pertain to the standard
block diagram,

e d
P 
y
-K u
but with different structures for P. The problems are labeled

FI Full information
FC Full control
DF Disturbance feedforward (to be considered in the “Disturbance
Feedforward and Output Estimation” section)
OE Output estimation (to be considered in the “Disturbance
Feedforward and Output Estimation” section)

FC and OE are natural duals of FI and DF, respectively. The DF solution can
be easily obtained from the FI solution, as shown in the “Disturbance
Feedforward and Output Estimation” section. The output feedback solutions
will be constructed out of the FI and OE results. A dual derivation could use
the FC and DF results.
The FI and FC problems are not, strictly speaking, special cases of the output
feedback problem, as they do not satisfy all of the assumptions. Each of the four
problems inherit certain of the assumptions A1–A4 from the “Problem
Statement” section as appropriate. The terminology and assumptions will be
discussed in the subsections for each problem. In each of the four cases, the
results are necessary and sufficient conditions for the existence of a controller
such that ||Ted||∞ < γ and the family of all controllers such that ||Ted||∞ < γ. In all
cases, K must be admissible.

3-31
3 H∞ Control and Model Reduction

The H∞ solution involves two Hamiltonian matrices, H∞ and J∞, which are
defined as follows:

2
R := D′ 1• D 1• – γ I m1 0 , where D 1• := D 11 D 12
0 0
2 D 11
R̃ := D •1 D′ •1 – γ I p1 0 , where D •1 :=
0 0 D 21

A 0 B –1
H ∞ := – R D′ 1• C 1 B′
– C′ 1 C 1 A′ – C′ 1 D 1•
(3-10)

A′ 0 C′ –1
J ∞ := – R̃ D •1 B′ 1 C
– B 1 B′ 1 – A – B 1 D′ •1
(3-11)

If H∞ ∈ dom(Ric) then let X1, X2 be any matrices such that

X1 X1
H∞ = TX , X ′1 X 2 = X ′2 X 1, Re λ i ( T X ) ð 0 ∀i
X2 X2
(3-12)

Similarly if J∞ ∈ dom(Ric) then let Y1, Y2 be any matrices such that

Y1 Y1
J∞ = TY , Y ′1 Y 2 = Y ′2 Y 1 Re λ i ( T Y ) ð 0 ∀i
Y2 Y2
(3-13)

Also define

–1 –1
X ∞ := X 2 X 1 , Y ∞ := Y 2 Y 1
(3-14)

Finally define the state feedback and output injection matrices as

3-32
 H∞ Full Information and Full Control Problems

F1 –1
F := := – R [ D ′ 1• C 1 + B ′ X ∞ ]
F2
(3-15)

–1
L := L 1 L 2 := – [ B 1 D ′ •1 + Y ∞ C ′ ]R̃
(3-16)

Problem FI: Full Information

In the full information (FI) special problem P has the following form:

A B1 B2
C 1 D 11 D 12
P( s ) =
I 0 0
0 I 0
(3-17)

It is seen that the controller is provided with full information since

 
y =  x  . In some cases, a suboptimal controller may exist which uses just
 d
the state feedback x, but this will not always be possible. While the state
feedback problem is more traditional, we believe that the Full Information
problem is more fundamental and more natural than the state feedback
problem, once you get outside the pure H2 setting.

The assumptions relevant to the FI problem, which are inherited from the
output feedback problem, are
(A1) (A,B2) is stabilizable.
(A2) D12 is full column rank with D 12 D ⊥ unitary.

A – jωI B 2
(A3) has full column rank for all ω.
C 1 D 12

3-33
3 H∞ Control and Model Reduction

The results for the Full Information case are as follows:

Theorem 3.5. Suppose P is given by (3-17) and satisfies A1–A3. Then,

a ∃K such that ||Ted||

∞ < 1 ⌠ H∞ ∈ dom(Ric), Ric(H∞) ≥ 0
b All admissible K(s) such that ||Ted|| < 1 are given by
∞

T1 0 F1 –I
K ( s ) = –Q ( s ) I
T2 I F2 0

for Q ∈ RcH∞, ||Q||∞ < 1.

Problem FC: Full Control

The full control (FC) problem has P given by,

A B1 I 0
P ( s ) = C 1 D 11 0 I
C 2 D 21 0 0
(3-18)

and is the dual of the Full Information case: the P for the FC problem has the
same form as the transpose of P for the FI problem. The term Full Control is
used because the controller has full access to both the state through output
injection and to the output e. The only restriction on the controller is that it
must work with the measurement y. The assumptions that the FC problem
inherits from the output feedback problem are just the dual of those in the FI
problem:
(A1) (C2,A) is detectable.

D 21
(A2) D21 is full row rank with unitary.
D̃ ⊥
A – jωI B 1
(A4) has full row rank for all ω.
C 2 D 21

3-34
 H∞ Full Information and Full Control Problems

Necessary and sufficient conditions for the FC case are given in the following
corollary. The family of all controllers can be obtained from the dual of
Theorem 3.5 but these controllers will not be required in the sequel and are
hence omitted.

Corollary 3.6. Suppose P is given by (3-18) and satisfies A1, A2 and A4. Then,

∃K such that ||Ted||∞ < 1 ⌠ J∞ ∈ dom(Ric), Ric(J∞) ≥ 0

3-35
3 H∞ Control and Model Reduction

H∞ Output Feedback
The solution to the Full Information problem of the “H• Full Information and
Full Control Problems” section is used in this section to solve the output
feedback problem. First in Theorem 3.8 a so-called disturbance feedforward
problem is solved. In this problem one component of the disturbance, d2, can be
estimated exactly from y using an observer, and the other component of the
disturbance, d1, does not affect the state or the output. The conditions for the
existence of a controller satisfying a closed-loop H∞-norm constraint is then
identical to the FI case.
The solution to the general output feedback problem can then be derived from
the transpose of Theorem 3.7 (Corollary 3.9) by a suitable change of variables
which is based on X∞ and the completion of the squares argument (see
[GloD2]).
The main result is now stated in terms of the matrices defined in the “H• Full
Information and Full Control Problems” section involving the solutions of the
X∞ and Y∞ Riccati equations together with the state feedback and output
injection matrices F and L. Assume that unitary changes of coordinates on ω
and z have been carried out to give the following partitions of D, F1 and L1.

F′ 11
F′ 12 F′ 2
F′ = L′ 11 D 1111 D 1112 0
L′ D L′ 12 D 1121 D 1122 I
L′ 2 0 I 0
(3-19)

Theorem 3.7. Suppose P satisfies the assumptions A1–A4 of the “Problem

Statement” section.
a There exists an admissible controller K(s) such that
||FL(P,K)||∞ < γ (i.e., ||Ted||∞ < γ) if and only if

i γ > max ( σ [ D 1111, D 1112 ],σ [ D′ 1111, D′ 1121 ] )

ii H∞ ∈ dom(Ric) with X∞ = Ric(H∞) ≥ 0
iiiJ∞ ∈ dom(Ric) with Y∞ = Ric(J∞) ≥ 0
iv ρ(X Y ) < γ2.
∞ ∞

3-36
 H∞ Output Feedback

e Given that the conditions of part (a) are satisfied, then all rational
internally stabilizing controllers K(s) satisfying ||FL(P,K)||∞ < γ are given
by

K = FL(Ka,Φ)for arbitrary Φ ∈ RcH∞such that||Φ||∞ < γ

where

ˆ ˆ ˆ
A B1 B2
Ka = ˆ ˆ ˆ
C1 D 11 D 12
ˆ ˆ
C2 D 21 0
ˆ 2 –1
D 11 = – D 1121 D ′ 1111 ( γ I – D 1111 D′ 1111 ) D 1112 – D 1122 ,

ˆ m2 × m2 ˆ p2 × p2
D 12 ∈ C and D 21 ∈ C are any matrices (e.g., Cholesky factors)
satisfying

ˆ ˆ 2 –1
D 12 D 12 = I – D 1121 ( γ I – D ′ 1111 D 1111 ) D 1121 ,
ˆ ˆ 2 –1
D 21 D 21 = I – D ′ 1112 ( γ I – D 1111 D′ 1111 ) D 1112 ,

and

ˆ –1 ˆ
B 2 = Z ∞ ( B 2 + L 12 ) D 12 ,
ˆ ˆ
C 2 = – D 21 ( C 2 + F 12 ),
ˆ –1 ˆ ˆ –1
B1 = –Z∞ L2 + B2 D 12 D̂ 11 ,
ˆ ˆ ˆ –1 ˆ
C = F + D D C ,
1 2 11 21 2
ˆ ˆ ˆ –1 ˆ
A = A + BF + B 1 D 21 C 2 ,

where

–2
Z∞ = ( I – γ Y∞ X∞ ) .

(Note that if D11 = 0 then the formulae are considerably simplified.)

3-37
3 H∞ Control and Model Reduction

The proof of this main result is via some special problems that are simpler
special cases of the general problem and can be derived from the FI and FC
problems. A separation type argument can then give the solution to the general
problem from these special problems. It can be assumed, without loss of
generality, that γ = 1 since this is achieved by the scalings γ–1D11, γ–1/2B1,
γ–1/2C1, γ1/2B2, γ1/2C2, γ-1X∞, γ-1Y∞ and γ–1K. All the proofs will be given for the
case γ = 1.

Disturbance Feedforward and Output Estimation

In the Disturbance Feedforward problem one component of the disturbance, d1,
does not affect the state or the output. The other component of the disturbance,
d2 (and hence the state x), can be estimated exactly from y using an observer.
The conditions for the existence of a controller satisfying a closed-loop
H∞-norm constraint is then identical to the Full Information case.
Theorem 3.8. (Disturbance Feedforward)
Theorem 3.7 is true under the additional assumptions that
˜
B 1 D ′ ⊥ = 0, A – B 1 D ′ 21 C 2 is stable.
(3-20)

In this case,

Y ∞ = 0, Z = I, L = – 0 B 1 D′ 21

The transpose of Theorem 3.8 can now be stated to obtain another special case
of Theorem 3.7.

Corollary 3.9. (Output Estimation)

Theorem 3.7 is true under the additional assumptions that

D ′ ⊥ C 1 = 0, A – B 2 D ′ 12 C 1 is stable.

In this case

0
X ∞ = 0, Z = I, F = –
D′ 12 C 1

3-38
 H∞ Output Feedback

Converting Output Feedback to Output Estimation

The output feedback case when the disturbance, d, cannot be estimated from
the output is reduced to the case of Corollary 3.9 by the change of variables

2 2 2 2
e 2– d 2 = v 2– r 2

where

v = u + T2 d – T2 , I Fx
r = T1 ( 2 ) ( d – F1 x )

We will perform the change of variables with v replacing e and r replacing d.

Hence

–1
x· = ( A + B 1 F 1 )x + B 1 T 1 r + B 2 u
–1
v = u + T2 T1 r – F2 x
–1
y = C 2 x + D 21 T 1 r + D 21 F 1 x

and the transfer matrix from   to   is

r v
 u  y

–1
A + B1 F1 B1 T1 B2
P vyru ( s ) := –F2 T2 T1
–1
I
–1
C 2 + D 21 F 1 D 21 T 1 0
(3-21)

Similarly substituting v for u in the equation for P gives that the transfer
function from   to   is H as defined by
d e
 v  r

3-39
3 H∞ Control and Model Reduction

e d 2
AF B1 , B2 T2 B2
3
H  H = 64 C1F1 D D D11 D12 75 0

r v ?

-Q
?

,T1 F1 T1 0
(3-22)

2 2 2 2
It can be shown that H~H = I (since e 2 – d 2 = v 2 – r 2 ) and that AF is
stable.
We can show with a little algebra the equivalence of the first two of the
following block diagrams, with Tvr = FL(Pvyru,K) given by the third one.

e d e d v P r
P  H   vyru

y u r v y u
-K -T vr
-K
Lemma 3.10. Let P satisfy A1–A4, and assume that X∞ exists and X∞ ≥ 0. Then
the following are equivalent:
a K internally stabilizes P and ||FL(P,K)||
∞ < 1,
b K internally stabilizes Pvyru and ||FL(Pvyru,K)|| < 1
∞
c K internally stabilizes Ptmp and ||FL(Ptmp,K)||∞ < 1,

where Pvyru is given by (3-21) and

A + B1 F1 B1 B2
P tmp := – D 12 F 2 D 11 D 12 .
C 2 + D 21 F 1 D 21 0

The importance of the above constructions for Pvyru and Ptmp is that they
satisfy the assumptions for the output estimation problem (Corollary 3.9) since
A + BF is stable.

3-40
 H∞ Output Feedback

Relaxing Assumptions A1–A4

In this section we indicate how the results of the “H• Output Feedback” section
can be extended to more general cases by the relaxation of assumptions A1–A4.
The optimal case is not considered, but the interested reader may refer to
[GloD2].

Relaxing A3 and A4
Suppose that,

0 0 1
P = 0 0 1
1 1 0

which violates both A3 and A4 and corresponds to the robust stabilization of an

integrator. If the controller u = –εx, for ε > 0 is used then
– εs
T ed = -----------, with T ed ∞ = ε
s+ε

Hence the norm can be made arbitrarily small as ε → 0, but ε = 0 is not

admissible since it is not stabilizing. This may be thought of as a case where
the H∞-optimum is not achieved on the set of admissible controllers. Of course,
for this system, H∞ optimal control is a silly problem, although the suboptimal
case is not obviously so.
If you simply drop the requirement that controllers be admissible and remove
assumptions A3 and A4, then the formulae in this chapter will yield u = 0 for
both the optimal controller and the suboptimal controller with Φ = 0. This
illustrates that assumptions A3 and A4 are necessary for the techniques in this
chapter to be directly applicable. An alternative is to develop a theory that
maintains the same notion of admissibility, but relaxes A3 and A4. The easiest
way to do this would be to pursue the suboptimal case introducing ε
perturbations so that A3 and A4 are satisfied.

Relaxing A1
If assumption A1 is violated, then it is obvious that no admissible controllers
exist. Suppose A1 is relaxed to allow unstabilizable and/or undetectable modes
on the jω axis, and internal stability is also relaxed to also allow closed-loop jω
axis poles, but A2–A4 is still satisfied. It can be easily shown that under these

3-41
3 H∞ Control and Model Reduction

conditions the closed-loop H∞ norm cannot be made finite, and in particular,

that the unstabilizable and/or undetectable modes on the jω axis must show up
as poles in the closed-loop system.

Violating A1 and Either or Both of A3 and A4

Sensible control problems can be posed that violate A1 and either or both of A3
and A4. For example, cases when A has modes at s = 0, which are unstabilizable
through B2 and/or undetectable through C2, arise when an integrator is
included in a weight on a disturbance input or an error term. In these cases,
either A3 or A4 is also violated, or the closed-loop H∞ norm cannot be made
finite. In many applications such problems can be reformulated so that the
integrator occurs inside the loop (essentially using the internal model
principle), and is hence detectable and stabilizable.
An alternative approach to such problems, which could potentially avoid the
problem reformulation, would be to pursue the techniques in [GloD2], but relax
internal stability to the requirement that all closed-loop modes be in the closed
left half plane. Clearly, to have finite H∞ norm these closed-loop modes could
not appear as poles in Ted. The formulae given in this chapter will often yield
controllers compatible with these assumptions. You would then have to decide
whether closed-loop poles on the imaginary axis were due to weights and hence
acceptable or due to the problem being poorly posed as in the above example.
A third alternative is to again introduce ε perturbations so that A1, A3 and A4
are satisfied. Roughly speaking, this would produce sensible answers for
sensible problems, but the behavior as ε → 0 could be problematic.

Relaxing A2
In the case that either D12 is not full column rank or D21 is not full row rank,
then improper controllers can give bounded H∞-norm for Ted, although will not
be admissible as defined in the “Problem Statement” section. Such singular
filtering and control problems have been well-studied in H2 theory and many
of the same techniques go over to the H∞ -case (e.g., [Will2], [WilKS] and
[HauS]). In particular the structure algorithm of [Silv] could be used to make
the terms D12 and D21 full rank by the introduction of suitable differentiators
in the controller.

3-42
 Discrete-Time and Sampled-Data H∞ Control

Discrete-Time and Sampled-Data H∞ Control

Discrete-Time Systems
The control of a discrete-time system described by the state difference equation

x k + 1 = Ax k + Bu k
y k = Cx k + Du k

is investigated in this section. The state difference equation corresponds to the

transfer function,
P(z) = D + C(zI – A)–1B
As in the continuous-time case, a controller with transfer function Kd(z) can be
synthesized to make the closed-loop, H(z) = FL(P(z),Kd(z)) internally stable and

jθ
H ∞ := supσ ( H ( e ) ) < γ
θ

for any γ sufficiently large. This can be accomplished either directly in terms of
the original data (A,B,C,D) or via the bilinear transformation,

1 + 1--2- sh
-,
z = ------------------
1 – 1--2- sh

which maps the unit disk in the z-plane into the left half of the s-plane for any
h > 0. The µ-Tools command to synthesize discrete-time H∞ controllers,
dhinfsyn, uses this bilinear transformation and the corresponding
continuous-time µ-Tools commands.
An additional consideration is which of the controllers Kd that make ||H||∞ < γ
should be chosen. The controller that maximizes the entropy integral,

–2
γ π
2 1 – zo
∫
–2 jθ jθ
I = ------ log det ( I – γ H ( e )*H ( e ) ) - dθ
--------------------------
2π –π jθ –1 2
e – zo

for any |zo| > 1 can be calculated. The usual central controller, the default for
hinfsyne, is taken as the one corresponding to zo = ∞ and gives a measure of
how far H(ejθ) is less than γ.”

3-43
3 H∞ Control and Model Reduction

Sampled-Data Systems
Continuous-time systems where measurements are sampled and then the
control signal calculated by a discrete-time controller followed by a hold are
termed sampled-data systems. Two possible H∞-type approaches to sampled
data control law design are available in µ-Tools software.
Consider the system in Figure 3-7, where P is the continuous-time generalized
plant, d is a continuous-time disturbance signal, e is a continuous-time error
signal, y is the measurement to be sampled by the sampler S, with sampling
period h, and u is the control signal, which is the output of the hold device, H,
and is constant between sampling points.

e   d
P
y  u

- S - Kd - H
Figure 3-7: Sampled Data System Block Diagram

If it is assumed that the input d is in fact piecewise constant (synchronized

with u), and that only the sampled values of e are of interest, then there is a
discrete-time equivalent system for P that relates the discrete-time inputs and
outputs. The controller can then be designed using discrete-time H∞
techniques.
This approach has two potential problems; one is that the intersample behavior
of the outputs is ignored and the other is that the inputs are assumed piecewise
constant. An alternative approach is to require that the induced norm between
the inputs d and outputs e be less than γ,

e L
sup ------------2- < γ
w d L 2

This will handle both the above difficulties and has been studied in detail by
Bamieh and Pearson [BamP] along with a number of other researchers

3-44
 Discrete-Time and Sampled-Data H∞ Control

[HarK,KabH,Toi,Yam]. It turns out that if γ is achievable then it can be

achieved by a linear time-invariant controller of the same degree as the
generalized plant, P. The computations for this controller involve calculating
an equivalent discrete-time system and then using discrete-time H∞ methods.

Discrete-Time and Sampled-Data Example

The following example is meant to illustrate the discrete-time and
sampled-data norms involved rather than be representative of controller
design. Consider the system in Figure 3-8 where,

1
F 1 ( s ) = -----------------------
( 1 + τ1 s )
2
ωo
F 2 ( s ) = ----------------------------------------------
2 2
( s + 2cω o s + ω o )
1
F 3 ( s ) = -----------------------
( 1 + τ3 s )

e F1  f F2  d
6 ?
F3
H  Kd  S 
u y
Figure 3-8: Sampled Data System Block Diagram for Simple Example

This gives the generalized plant,

P 11 ( s ) P 12 ( s ) F1 F2 F1
P = ,=
P 21 ( s ) P 22 ( s ) F3 0

3-45
3 H∞ Control and Model Reduction

and the closed-loop is trying to match the output of F2 by the controller output
based on the sampled input to F2 filtered by F3. P can be defined as follows.
h = 0.1;
tau1 = 0.001;
om_o = 2*pi; c = 0.05;
tau3 = 0.01;
F1 = nd2sys(1,[tau1 1]);
F2 = nd2sys(1,[omo_o(-2) 2*c/omo_o 1]);
F3 = nd2sys(1,[tau3 1]);
p_ic = abv(mmult(F1,sbs(F2,1)),sbs(F3,0));
ncon = 1; nmeas = 1;
minfo(p_ic)
system: 4 states 2 outputs 2 inputs

The zero controller will result in a purely continuous-time system with induced
norm given by ||P11(s)||∞.
hinfnorm(sel(p_ic,1,1))
norm between 10.01 and 10.02
achieved near 6.267

Now let us design a controller for this sampled-data system using the
corresponding sample and hold discrete-time system. The variable delay
corresponds to the number of sample delays in the controller.
gmin =.001; delay = 0;
gmax = 1;
tol = 0.001; tol2 = 0.001;
p_ic_sh = samhld(p_ic,h);
if delay>0,
p_ic_sh=mmult(daug(1,nd2sys(1,eye(1,delay+1))),p_ic_sh;
end
[k_d,g_d,gfin_d] = ...
dhfsyn(p_ic_sh,nmeas,ncon,gmin,gmax,tol,h,inf,-1,-2);

3-46
 Discrete-Time and Sampled-Data H∞ Control

Test bounds:0.0010 < gamma <=1.0000

gamma hamx_eig xinf_eig hamy–eig yinf–eig nrho–xy p/f

1.000 3.5e–01 0.0e+00 3.5e–01 –3.9e–16 0.0337 p
0.500 3.5e–01 0.0e+00 3.5e–01 –1.1e–15 0.1345 p
0.251 3.5e–01 –8.4e–14 3.5e–01 –7.1e–14 0.5357 p
0.126 3.5e–01 0.0e+00 3.5e–01 –4.1e–16 2.1257# f
0.188 3.5e–01 –7.5e–14 3.5e–01 0.0e+00 0.9498 p
0.157 3.5e–01 0.0e+00 3.5e–01 –3.6e–16 1.3648# f
0.173 3.5e–01 –1.1e–15 3.5e–01 –8.7e–14 1.1292# f
0.181 3.5e–01 0.0e+00 3.5e–01 0.0e+00 1.0337# f
0.184 3.5e–01 0.0e+00 3.5e–01 0.0e+00 0.9904 p
0.182 3.5e–01 –3.5e–14 3.5e–01 –3.8e–16 1.0117# f
0.183 3.5e–01 –3.0e–16 3.5e–01 0+0e+00 1.0010# f

Gamma value achieved: 0.1844

dhfnorm(g_d,tol2,h)
norm between 0.1835 and 0.1837
achieved near 31.42

The induced norm of the sampled-data system from continuous-time inputs to

continuous-time outputs can now be calculated using the command sdhfnorm.
[gaml_d,gamu_d] = sdhfnorm(p_ic,k_d,h,delay,tol2)
gaml_d =
2.0250e+01
gamu_d =
2.0267e+01

The discrete-time system with piecewise constant inputs and ignoring

intersample behavior has an induced norm of only 0.1837 whereas if the input
is allowed to vary during the sampling period the gain can be made 100 times
larger at 20.2. A suboptimal controller for the sampled-data system can also be
calculated using sdhfsyn.

3-47
3 H∞ Control and Model Reduction

[k_sd,gfin_sd]=sdhfsyn(p_ic,1,1,gmin,gamu_d,tol,h,delay,-2);
Test bounds:0.0010 < gamma <=20.2670

gamma hamx_eig xinf_eig hamy–eig yinf–eig nrho_xy p/f

20.267 3.5e–01 –1.6e–30 3.le–01 1.0e–09 0.0016 p
10.134 3.4e–01 1.3e–31 1.6e–01 1 0e–09 0.0087 p
5.067 3.3e–01 –2.7e–20 8.3e–17# ******* ******* f
7.601 3.4e–01 2.3e–20 3.8e–15# ******* ******* f
8.867 3.4e–01 3.7e–21 1.2e–14# ******* ******* f
9.501 3.4e–01 l.le–l9 1.2e–01 1.0e–09 0.0109 p
9.184 3.4e–01 –5.3e–20 7.6e–02 1.0e–09 0.0128 p
9.026 3.4e–01 5.5e–21 4.2e–02 1.0e–09 0.0145 p
8.947 3.4e–01 1.3e–31 9.4e–14# ******* ******* f
8.986 3.4e–01 –7.7e–20 2.7e–02 1.0e–09 0.0153 p
8.966 3.4e–01 –1.5e–31 1.3e–02 1 0e–09 0.0160 p
8.956 3.4e–01 4.1e–20 3.9e–13# ******* ******* f
8.961 3.4e–01 4.9e-20 6.6e-03 1 0e–09 0.0163 p
8.959 3.4e–01 7.6e-20 8.7e-14# ******* ******* f
8.960 3.4e–01 3.5e-03 3.2e-03 1 0e–09 0.0165 p
8.960 3.4e–01 -2.3e-20 9.0e-14# ******* ******* f

Gamma value achieved:8.9601

and the norm checked by

[gaml_sd,gamu_sd] = sdhfnorm(p_ic,k_sd,h,delay,tol2)
gaml_sd =
8.9445e+00
gamu_sd =
9.0028e+00

The initial design gave very low estimates of the possible gain in the system.
The latter design indicates that no controller can give a low gain with this
sampled-data problem. The main difficulty with this particular problem is that

3-48
 Discrete-Time and Sampled-Data H∞ Control

the filter F3(s) has too high a bandwidth and this gives a high potential gain for
F3 followed by the sampler. In contrast to the continuous-time case, the
calculation of a worst-case disturbance in the sampled-data case is not
straightforward. However the time domain simulation of the system is now
performed to illustrate the reason for its high gain.
tfinal = 10;
t1 = (0:h/100:tfinal)';
[nr,nc] = size(t1);
w1 = zeros(nr,nc);
for i = 1:length(t1)/200,
w1(100*i-20:100*i) = ...
cos(i*pi/5-pi/6)*exp(h*(-20:1:0)/(100*tau3))';...
end
w = vpck(w1,t1);
[z_d,y_d,u_d] = sdtrsp(p_ic,k-d,w,h,tfinal,h/100);
vplot(z_d,'-',y_d,'.')
gain_d = norm(vunpck(z-d),2)/vnorm(vunpck(w),2)
gain_d =
1.2924e+01

The time responses are given in Figure 3-9 and the high gain achieved by the
disturbance being large just before the sampling instant and zero elsewhere,
hence having a relatively low total energy.

3-49
3 H∞ Control and Model Reduction

-1

-2

-3

-4
0 1 2 3 4 5 6 7 8 9 10

Figure 3-9: Discrete-Time Controller Time Response

The suboptimal sampled-data controller can be simulated with the same input
as follows:
[z_sd,y_sd,u_sd] = sdtrsp(p_ic,k_sd,w,h,tfinal,h/100);
vplot(z_sd,'-',y_sd,'.')
gain_sd = norm(vunpck(z_sd),2)/vnorm(vunpck(w),2)
gain_sd =
6.3332e-01

The time responses are given in Figure 3-10. The gain in this instance is much
lower for this particular input. However other inputs can give a gain of close to
nine.
The same example can be repeated for different values of the sampling period,
h, and the controller delay, and for variations in the time constants. The two
controllers often give very similar results, however, the discrete-time results
obtained from samhld and dhfsyn can give optimistic gain estimates when
compared with those obtained by sdhfsyn.

3-50
 Discrete-Time and Sampled-Data H∞ Control

0.6

0.4

0.2

-0.2

-0.4

-0.6
0 1 2 3 4 5 6 7 8 9 10

Figure 3-10: Sampled Data Controller Time Response

3-51
3 H∞ Control and Model Reduction

Loop Shaping Using H∞ Synthesis

A particularly straightforward method of designing controllers is to use a
combination of loop shaping and robust stabilization as proposed in McFarlane
and Glover [McFG1, McFG2]. Given a system with transfer function G(s) the
problem set up is given in Figure 3-11.

d1 d2
6e1 = u
?
i W2  G  W1  ?i

-
e2 = y - K1
reference ( ) r
Figure 3-11: H∞ Loop Shaping Standard Block Diagram
The first step is to design a pre-compensator W1(s), so that the gain of
W2(s)G(s)W1(s) 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 design a feedback controller, K∞, so
that

1 I –1 1
---------------------------------------- := ( I – W 2 GW 1 ,K ∞ ) [ W 2 GW 1, I ] ð ---
b ( W 2 GW 1 ,K ∞ ) K∞ ε
∞

which will also give robust stability of the perturbed weighted plant

–1 ∆1
( N + ∆1 ) ( M + ∆2 ) for < b ( W 2 GW 1 ,K ∞ )
∆2
∞

where NM–1 = W2GW1 is a normalized coprime factorization satisfying

N(jω)*N(jω) + M(jω)*M(jω) = I. This stability margin is always less than 1 and
gives a good indication of robust stability to a wide class unstructured
perturbations. Stability margin values of ε > 0.2 – 0.3 are generally considered

3-52
 Loop Shaping Using H∞ Synthesis

satisfactory. The closed-loop H∞-norm performance 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 crossover. In the regulator setup, the final controller
to be implemented is W1K∞W2 in the feedback configuration shown in
Figure 3-12.

G 

- W2 - K1 - W1
r
-
reference ( )

Figure 3-12: H∞ Loop Shaping Controller Implementation

Closely related to the above problem is the approach to uncertainty using the
gap family of metrics. These metrics, δ(G0,G1), give a numerical value between
0 and 1 for the distance between any two linear systems G0(s) and G1(s) that
have the same number of inputs and outputs. The gap metric was introduced
into the control literature by Zames and El-Sakkary [ZamE] and exploited by
Georgiou and Smith [GeoSm]. It can be computed using an H∞ optimization
problem [Geo]. An interesting new metric, ν-gap (nugap), was derived by
Vinnicombe [Vin] and has a frequency response interpretation. For both of
these metrics, the following robust performance result holds [QiuD, Vin].

arcsinb ( G 1, K 1 ) ≥ arcsinb ( G 0, K 0 ) – arcsinδ ν ( G 0, G 1 ) – arcsinδ ν ( K 0, K 1 )

(3-23)

where
–1
b ( G, K ) = I ( I – GK ) – 1 .
G I
K ∞

The nugap is always less than or equal to the gap, so its predictions using the
above robustness results are tighter. To make use of the gap metrics in robust

3-53
3 H∞ Control and Model Reduction

design, weighting functions need to be introduced. In the above robustness

–1 –1
result G needs to be replaced by W2GW1 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 loop shaping H∞ synthesis.
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. Model
reduction of the system model and controller can be performed by using
balanced truncations or Hankel norm approximation of normalized coprime
factor representation.
If a reference signal (r) is available, then there is a variety of methods to
incorporate the tracking error into the objective. One effective procedure is to
choose a coprime factorization of K∞ = UV–1 and let M = V–1(r +Uy) where r is
the reference and y the measurement. U and V can be obtained from a
particular observer form of K∞ in such a way that with closed-loop y = Nr and
G = NM–1 is a normalized coprime factorization [Vin]. The DC. gain of this
controller could then be adjusted to give zero steady state tracking error.
Alternatively, this nominal command response could be diagonalized and a
further command compensation performed. These ideas are applied to the
“HIMAT Robust Performance Design Example” section in Chapter 7.

3-54
 Model Reduction

Model Reduction
It is often desirable to approximate a state-space representation of a system
with a lower order state-space representation. This procedure is referred to as
a model reduction. The µ-Analysis and Synthesis Toolbox (µ-Tools) provides
several commands to aid in reducing the order of a system. These are discussed
in this section along with an example to illustrate their use.
Given a SYSTEM matrix [A B; C D], the simplest method of model reduction
is to truncate a part of the SYSTEM A matrix and remove the corresponding
columns and rows of the B and C matrices. The command strunc performs this
function. You should be careful to order the modes of the A matrix and truncate
modes that do not significantly affect the system response. The command
strans is useful in this context as it transforms the A matrix into block
diagonal form with 1 × 1 or 2 × 2 blocks corresponding to the respectively real
and complex poles in order of increasing magnitude. This is often done prior to
truncating high frequency modes.
Truncating high frequency modes will also affect the low frequency response of
the various transfer functions. The command sresid can be used to residualize
the truncated modes and compensate for the zero frequency contribution of
each truncated mode with an additional D matrix term in the resulting reduced
order SYSTEM matrix.
More advanced model reduction techniques for stable systems can be
performed with the µ-Tools commands sysbal and hankmr. sysbal performs a
balanced realization on the input SYSTEM matrix, which entails balancing the
observability and controllability Grammians (for a more detailed discussion
see [Enn], [Glo1] and [Moo]). In its simplest form, this command will remove
all unobservable and/or uncontrollable modes. sysbal also returns a vector of
the Hankel singular values of the system, which can be used to further
truncate the modes of the SYSTEM.
The following example illustrates how sysbal can be used to remove
unobservable and uncontrollable modes. Two systems, P and C, are created and
then interconnected with unity gain negative feedback. The closed-loop system,
clp, is given by

PC
clp = ----------------
1 + PC

3-55
3 H∞ Control and Model Reduction

Suppose that
P = nd2sys(1,[1,1,1]);
C = nd2sys(1,[1,1]);

The correct way to form this closed-loop transfer function is to use

clpc = starp(mmult([1;1],mmult(mmult(P,C),[1,-1])),1,1,1);

The closed-loop SYSTEM matrix clp has three states. However, suppose
instead the closed-loop system is formed as follows.

clp = mmult(P,C,minv(madd(1,mmult(P,C))));
minfo(clp)
system: 6 states 1 outputs 1 inputs
rifd(spoles(clp))

real imaginary frequency damping

–5.0000e–01 –8.6603e–01 l.0000e+00 5.0000e–01
–5.0000e–01 8.6603e–01 l.0000e+00 5.0000e–01
–l.0000e+00 0.0000e+00 l.0000e+00 l.0000e+00
–2.2816e–01 –1.1151e+00 1.1382e+00 2.0045e–01
–2.2816e–01 1.1151e+00 1.1382e+00 2.0045e–01
–1.5437e+00 0.0000e+00 1.5437e+00 l.0000e+00

The closed-loop system, clp, contains the open-loop poles of p as well as the
closed-loop poles. Interconnecting systems with the commands mmult and madd
often lead to nonminimal realizations. You can see that the open-loop poles are
unobservable and/or uncontrollable by using sysbal with its second input, the
truncation tolerance, set to zero. The output gives the Hankel singular values
that are strictly greater than this tolerance together with a truncated balanced
realization of this order. (A strictly positive default tolerance is also available.)
Note that only five values are returned, the sixth being calculated as being
identically zero, and the fourth and fifth are both zero to machine accuracy.
strunc is then run to remove these two modes. Finally the H∞-norm of the

3-56
 Model Reduction

error between the system with the nonminimal modes truncated and the
system formed using starp is calculated as being essentially zero.
[clpr,hanksv] = sysbal(clp,0);
disp(hanksv')
6.7732e-014.7580e-014.8484e-028.0919e-17
1.5746e-18
clpr = strunc(clpr,3);
minfo(clpr)
system: 3 states 1 outputs 1 inputs
rifd(spoles(clpr))

real imaginary frequency damping

–2.2816e–01 –1.1151e+00 1.1382e+00 2.0045e–01
–2.2816e–01 1.1151e+00 1.1382e+00 2.0045e–01
–1.5437e+00 0.0000e+00 1.5437e+00 1.0000e+00

hinfnorm(msub(clpc,clpr))
norm between 4.163e-16 and 4.167e-16
achieved near 0

A variety of norms is available to measure the error in reducing the model

order. The Hankel-norm and H∞-norms have relatively complete theories.
Note that the Hankel norm is the maximum of the Hankel singular values and
is hence available from sysbal. The command hankmr calculates the optimal
Hankel norm approximation, of a specified order, to the input system. This
method is presented in detail in [Glo1]. The calling sequence is
[sysb,sig] = sysbal(sys)
[sysh,sysu] = hankmr(sysb,sig,k,'d')

The sysh is stable with k poles, sysu is unstable (or anticausal), and ||sys -
sysh - sysu||∞ = sig(k + 1). This answer is optimal. If the 'd' option of
hankmr is used then the following bound is guaranteed,
sig(k + 1) ≤ ||sys - sysh|| ∞≤ sig(k + 1) + sig(k + 2) + ...

The lower bound holds for any sysh of degree k and for truncated balanced
realizations the upper bound needs to be doubled.

3-57
3 H∞ Control and Model Reduction

The following example creates a 15-state system and examines three, 7-state
reduced-order models generated by sysbal, hankmr, and sresid. First generate
the system, which consists of 10 real poles, a resonant pole pair, and first-order
and second-order high frequency all-pass terms.
a = -diag([.03.05.1.2.3.4 1 3 5 10]);
b = [.03.05.1.2.3.4 1 3 5 10]'; c = ones(1,10);
d = 0.001;
sys1 = pck(a,b,c,d);
sys2 = nd2sys([1.1.4],[1.1.1]);
sys3 = nd2sys([1 -3 1000],[1 3 1000]);
sys4 = nd2sys([1 -20],[1 20]);
sys = mmult(sys1,sys2,sys3,sys4);
minfo(sys)
system: 15 states 1 outputs 1 inputs

The frequency response of sys is calculated and plotted in Figure 3-13.

omega = logspace(-1,3,60);
sys_g = frsp(sys,omega);
vplot('bode',sys_g);
title('Original system to be reduced')

3-58
 Model Reduction

10 2 Original system to be reduced

Log Magnitude
10 1

10 0

10 -1

10 -2
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

200
Phase (radians)

100

-100
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

Figure 3-13: Original System to be Reduced via Model Reduction Techniques

Next the balanced realization is formed and truncated to seven states. Its
H∞-norm error is compared with the upper and lower bounds.

3-59
3 H∞ Control and Model Reduction

The Bode plot of the original system and the seven state, balanced realization
model sysb7 is shown in Figure 3-14.
[sysb,sig] = sysbal(sys);
[mattype,p,m,n] = minfo(sysb);
disp(sig')
Columns 1 through 5
3.6422e+01 2.4906e+01 6.1381e+00 2.0261e+00 7.1689e-01
Columns 6 through 11
6.9421e-01 6.0281e-01 4.2291e-01 1.1884e-01 3.4276e-02
Columns 12 through 15
9.0070e-03 2.4834e-03 4.0909e-04 1.4544e-04 3.9050e-06
k = 7;
sysb7 = strunc(sysb,k);
sysb7_g = frsp(sysb7,omega);
vplot('bode',sys_g,sysb7_g) title(['Model reduction example:
Frequency domain'])
tmp = hinfnorm(msub(sys_g,sysb7_g));
disp(['H-inf error = ' num2str(tmp)])
H-inf error = 0.8553
disp(['lower bound = ' num2str(sig(k+1))])
lower bound = 0.4229
disp(['upper bound = ' num2str(2*sum(sig(k+1:n)))])
upper bound = 1.176

3-60
 Model Reduction

10 2 Model reduction example: Frequency domain

10 1

Log Magnitude
10 0

10 -1

10 -2

10 -3
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

200
Phase (radians)

100

-100

-200
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

Figure 3-14: Balanced Realization Model Reduction: Original System (solid),

Balanced (dashed)

The seven-state optimal Hankel norm approximation to sys is calculated with

hankmr and placed in sysha7. Figure 3-15 contains the Bode plots of original
system, the balanced realization model, and the seven-state Hankel norm
model.
kh = 7;
sysha7 = hankmr(sysb,sig,kh,'d');
sysha7_g = frsp(sysha7,omega);
tmp = hinfnorm(msub(sys_g,sysha7_g));
disp(['H-inf error = ' num2str(tmp)])
H-inf error = 0.4351
disp(['lower bound = ' num2str(sig(kh+1))])
lower bound = 0.4229
disp(['upper bound = ' num2str(sum(sig(kh+1:n)))])
upper bound = 0.5881
vplot('bode',sys_g,sysb7_g,sysha7_g)
title(['Three different model reduction techniques'])

3-61
3 H∞ Control and Model Reduction

Three different model reduction techniques

10 2

10 1
Log Magnitude

10 0

10 -1

10 -2

10 -3
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

200
Phase (radians)

100

-100

-200
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

Figure 3-15: Model Reduction: Original System (solid), Balanced (dashed),

Hankel (dotted)

Now obtain a truncated residualization. It turns out that the first seven poles
with smallest modulus also have the largest H∞-norms and hence no
reordering of the poles after strans is required.
sysr = strans(sys);
sysrt7 = sresid(sysr,7);
sysrt7_g = frsp(sysrt7,omega);
tmp = hinfnorm(msub(sys_g,sysrt7_g));
disp(['H-inf error = ' num2str(tmp)])
H-inf error = 7.782
vplot('bode',sys_g,sysb7_g,sysha7_g,sysrt7_g)
title(['Four different model reduction techniques'])

3-62
 Model Reduction

Four different model reduction techniques

10 2

10 1

Log Magnitude
10 0

10 -1

10 -2

10 -3
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

200
Phase (radians)

100

-100

-200
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

Figure 3-16: Model Reduction: Original System (solid), Balanced (dashed),

Hankel (dotted), Residualized (dashed-dotted)

The four different model Bode plots are shown in Figure 3-16. The output from
hankmr is nearly H∞-optimal and that from sysbal has a similar error and does
not change the D matrix. The frequency responses are plotted below with
sys-solid, sysbal-dashed, hankmr-dotted, and resid-dash/dot. Note that as
expected sresid matches well at low frequency but not at high frequency.
The time responses of the four systems, shown in Figure 3-17, are now
compared in response to a 1 second pulse. The same line types are used for the
display of the time domain responses. The response of sysrt7 does not match
sys well over the first 2 seconds but after 2 seconds the match is good.

3-63
3 H∞ Control and Model Reduction

pulse = siggen('t<1',[0:.1:10]);
ysys = trsp(sys,pulse);
integration step size: 0.003162
interpolating input vector (zero order hold)
ysysb7 = trsp(sysb7,pulse);
integration step size: 0.004127
interpolating input vector (zero order hold)
ysysha7 = trsp(sysha7,pulse);
integration step size: 0.003548
interpolating input vector (zero order hold)
ysysrt7 = trsp(sysrt7,pulse);
integration step size: 0.1
vplot(ysys,'-',ysysb7,':',ysysha7,'--',ysysrt7,'-.');
xlabel('Time: seconds')
title('Model reduction example: time domain')

Model reduction example: time domain

1

-1

-2

-3

-4

-5

-6

-7
0 1 2 3 4 5 6 7 8 9 10

Time: seconds

Figure 3-17: Time Response of the Original System and the Three Reduced
Order Models

3-64
 Model Reduction

Note You should not draw general conclusions from this one example as to
the relative merits of the different schemes.

The H∞ norm of the error is not always appropriate, for example, in the system
above none of the methods accurately matches the Bode diagram at high
frequencies. It is therefore desirable to generate reduced-order models whose
frequency weighted error is small. Two methods are available in µ-Tools to
assist in this. The first method is based on frequency weighted Hankel norm
approximation as proposed in Latham and Anderson [LatA] and finds Ĝ of
~– 1 ~– 1
degree k to minimize the Hankel norm of the stable part of W 1 ( G – Ĝ )W 2 .
Note that W1(s)~ is defined as. W1(–s)´. This is implemented in the functions
sfrwtbal and sfrwtbld. G is required to be stable and the weights need to be
square, stable, and minimum phase. sfrwtbal then finds a balanced
~– 1 ~– 1
realization of the stable part of W 1 GW 2 together with its Hankel singular
values, which in this case also provide lower bounds on the achievable error.
The resulting balanced system is approximated using hankmr (or another
method if preferred) and Ĝ constructed using sfrwtbld. This is illustrated on
the 15-state example with a sixth order approximation and with the relative
error criterion, although it is not restricted to this case.
wt1 = mmult(sys1,sys2); wt2 = 1; k = 6; n = 15;
[sysfrwtbal,sigfrwt] = sfrwtbal(sys,wt1,wt2);
disp(sigfrwt')
Columns 1 through 5
1.0000e+00 1.0000e+00 1.0000e+00 9.9697e-01 9.3961e-01
Columns 6 through 11
9.1022e-01 2.2177e-01 5.9077e-02 1.7448e-02 4.5526e-03
Columns 12 through 15
1.2419e-03 2.1046e-04 8.1357e-05 1.2386e-05 2.9765e-07
sysfrwtk = hankmr(sysfrwtbal,sigfrwt,k,'d');
sysfrwthat = sfrwtbld(sysfrwtk,wt1,wt2);
sysfrwthat_g = frsp(sysfrwthat,omega);
disp(hinfnorm(msub(1,vrdiv(sysfrwthat_g,sys_g))))
3.2401e-01
vplot('bode',sys_g,'-',sysfrwthat_g,':')
title('sys_g (-) and sysfrwthat_g (:)')

3-65
3 H∞ Control and Model Reduction

10 2 sys_g (-) and sysfrwthat_g (:)

Log Magnitude
10 1

10 0

10 -1

10 -2
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

200
Phase (degrees)

100

-100
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

Figure 3-18: Frequency Response of the Original System and Frequency

Weighted Balanced Reduction with sfrwtbld

The frequency weighted error is close to the unachievable lower bound of

sigfrwt(k+1). A plot of this is shown in Figure 3-18. You can verify that this
method is optimal by checking that the Hankel norm of the stable part of the
weighted error is sigfrwt(k+1), using sfrwtbal as follows. (Note that these
routines use sdecomp, which decomposes a system into the sum of two systems
with poles to the left and right of a vertical line in the complex plane.)
[sysfrwterr,sigfrwterr] =...
sfrwtbal(msub(sys,sysfrwthat),wt1,wt2);
disp(max(sigfrwterr));
2.2177e-01

A method that is particularly appropriate for relative error model reduction is

obtained using srelbal. The error criterion is the same as in the general
~ ~
frequency weighted case but with W1 = I and W2 satisfying W 2 W 2 = G G .
This method is related to balanced stochastic truncation as introduced by Desai
and Pal [DesP]. The lower bounds can be obtained as above and upper bounds
have been derived by Green [Gre] and improved by Wang and Safonov [WanS].
There is no requirement that the system be square or minimum phase but it

3-66
 Model Reduction

must not have fewer columns than rows. The resulting realization can be
truncated as before.
[sysrelbal,sigrel] = srelbal(sys);
sysrelbalk = strunc(sysrelbal,k);
sysrelbalk_g = frsp(sysrelbalk,omega);
disp(hinfnorm(msub(1,vrdiv(sysrelbalk_g,sys_g))));
4.7480e-01
vplot('bode',sys_g,'-',sysrelbalk_g,'--',sysfrwthat_g,':')
title('model reduction example: relative error')

model reduction example: relative error

10 2
Log Magnitude

10 1

10 0

10 -1

10 -2
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

200
Phase (degrees)

100

-100
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

Figure 3-19: Frequency Response of the Original System and Reduced Order
Models Using Relative Error Methods

It can be checked that sigrel equals sigfrwt above in this case. Both methods
perform well with results close to the lower bound and similar frequency
responses as seen in Figure 3-19. Glover [Glo2] suggests a combination of
additive and relative error by performing relative error model reduction of the

augmented system, G . Relative error is recovered if α = 0, and additive error

αI
as α → ∞. The resulting approximation can be thought of as satisfying

3-67
3 H∞ Control and Model Reduction

Ĝ = ( I + ∆ r )G + α∆ a where
∆ r ∆ a has been made small. This is easily
∞
implemented using srelbal as follows to give a seventh order fit with a
performance between the two extremes.
k = 7; alpha = 15;
[sysrelbal,sigrel] = srelbal(abv(sys,alpha*eye(1)));
sysrelbalk = sel(strunc(sysrelbal,k),1,1);

You can evaluate the error in the same manner as was done in the frequency
weighted balanced reduction case.

3-68
 References

References
[AdAK:] Adamjan, V.M., D.Z. Arov, and M.G. Krein, “Infinite block Hankel
matrices and related extension problems,” AMS Transl., vol. 111, pp. 133-156,
1978.
[And:] Anderson, B.D.O., “An algebraic solution to the spectral factorization
problem,” IEEE Transactions on Automatic Control, vol. AC-12, pp. 410-414,
1967.
[BallC:] Ball, J.A. and N. Cohen, “Sensitivity minimization in an H∞ norm:
Parametrization of all suboptimal solutions,” International Journal of Control,
vol. 46, pp. 785-816, 1987.
[BallH:] Ball, J.A. and J.W. Helton, “A Beurling-Lax theorem for the Lie group
U(m,n) which contains most classical interpolation theory,” J. Op. Theory, vol.
9, pp. 107-142, 1983.
[BamP:] Bamieh, B.A. and J.B. Pearson, “A general framework for linear
periodic systems with applications to H∞ sampled-data control,” IEEE
Transactions on Automatic Control, vol. AC-37, pp. 418–435, 1992.
[BoyBk:] Boyd, S., V. Balakrishnan, and P. Kabamba, “A bisection method for
computing the H∞ norm of a transfer matrix and related problems,” Math.
Control, Signals, and Systems, vol. 2, no. 3, pp. 207-220, 1989.
[ClemG:] Clements, D.J. and K. Glover, “Spectral Factorization via Hermitian
Pencils,” Linear Algebra and its Applications, Linear Systems Special Issue,
pp. 797-846, Sept. 1989.
[DesP:] Desai, U.B. and D. Pal, “A transformation approach to stochastic model
reduction,” IEEE Transactions on Automatic Control, vol. AC-29, pp.
1097-1100, 1984.
[Doy1:] Doyle, J.C., “Lecture notes in advances in multivariable control,”
ONR/Honeywell Workshop, Minneapolis, 1984.
[DoyFT:] Doyle, J.C. , B.A. Francis, and A.R. Tannenbaum, Feedback Control
Theory, Macmillian Publishing Company, New York, Toronto, 1992.
[DoyGKF:] 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. AC-34, no. 8, pp. 831-847, August
1989.

3-69
3 H∞ Control and Model Reduction

[Enn:] Enns, D., “Model reduction for control system design,” Ph.D.
dissertation, Stanford University, June, 1984.
[FoisT:] Foias, C., and A. Tannenbaum, “On the four-block problem, I,” in
Operator Theory: Advances and Applications, vol. 32, pp. 93-122, 1988 and “On
the four-block problem, II: The singular system,” Int. Equations Operator
Theory, vol. 11, pp. 726-767, 1988.
[Fran1:] Francis, B.A., A course in H∞ control theory, Lecture Notes in Control
and Information Sciences, vol. 88, Springer-Verlag, Berlin, 1987.
[FranD:] Francis, B.A. and J.C. Doyle, “Linear control theory with an H∞
optimality criterion,” SIAM J. Control and Opt., vol. 25, pp. 815-844, 1987.
[GeoS:] T. Georgiou and M. Smith, “Robust stabilization in the gap metric:
Controller design for distributed plants,” IEEE Transactions on Automatic
Control, pp. 1133-1143, vol. AC-37., no. 8, Aug. 1992.
[Glo1:] Glover, K., “All optimal Hankel-norm approximations of linear
multivariable systems and their L∞ error bounds,” International Journal of
Control, vol. 39, pp. 1115-1193, 1984.
[Glo2:] K. Glover, “Multiplicative approximation of linear multivariable
systems with L∞ error bounds,” Proceedings of the American Control
Conference, Seattle, pp. 1705-1709, 1986.
[Glo3:] Glover, K., “Tutorial on Hankel-norm approximation,” in From Data to
Model (J.C. Willems ed.), Springer-Verlag, pp. 26-48, 1989.
[GloD:] 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, vol. 11, pp. 167-172, August 1989. International
Journal of Control, vol. 39, pp. 1115–1193, 1984.
[GloD2:] Glover, K. and J. Doyle, “A state-space approach to H∞ optimal
control,” in Three Decades of Mathematical Systems Theory: A Collection of
Surveys at the Occasion of the 50th Birthday of Jan C. Willems, H. Nijmeijer
and J.M. Schumacher (Eds.), Springer-Verlag Lecture Notes in Control and
Information Sciences vol. 135, 1989.
[GloM:] Glover, K. and D. Mustafa, “Derivation of the Maximum Entropy
H∞-controller and a State-space formula for its Entropy,” International
Journal of Control, vol. 50, no. 3, pp. 899-916, Sept. 1989.

3-70
 References

[GloLDKS:] Glover, K., D.J.N. Limebeer, J.C. Doyle, E.M. Kasenally, and M.G.
Safonov, “A characterization of all solutions to the four block general distance
problem,” SIAM J. Control and Opt., vol. 29, no. 2, pp. 283-324, March, 1991.
[GohLR:] Gohberg, I., P. Lancaster and L. Rodman, “On the Hermitian
solutions of the symmetric algebraic Riccati equation,” SIAM J. Control and
Opt., vol. 24, no. 6, pp. 1323-1334, 1986.
[Gre:] Green, M., “A relative error bound for balanced stochastic truncation,”
IEEE Transactions on Automatic Control, vol. AC-33, pp. 961-965, 1988.
[GGLD:] Green, M., K. Glover, D. Limebeer, and J.C. Doyle, “A J-spectral
factorization approach to H∞ control,” SIAM J. Control and Opt., Vol. 28(6), pp.
1350-1371, 1990.
[HarK:] Hara, S. and P.T. Kabamba, “On computing the induced norm of
sampled data feedback systems,” Proceedings of the American Control
Conference, Green Valley, AZ, pp. 319-320, 1990.
[HauS:] Hautus, M.L.J. and L.M. Silverman, “System structure and singular
control.” Linear Algebra Applications, vol. 50, pp. 369-402, 1983.
[Hung:] Hung, Y.S., “RH∞-optimal control-Part I: Model matching, Part II:
Solution for controllers,” International Journal of Control, vol. 49, pp. 675-684,
1998.
[JonJ:] Jonckheere, E.A., and J.C. Juang, “Fast computation of achievable
performance in mixed sensitivity H∞ design,” IEEE Transactions on Automatic
Control, vol. AC-32, pp. 896-906, 1987.
[KabH:] Kabamba, P.T. and S. Hara, “Worst case analysis and design of
sampled data control systems,” IEEE Transactions on Automatic Control, vol.
AC-38, no. 9, pp. 1337-1357, Sept. 1993.
[KhaPZ:] Kargoneckar, P.P., I.R. Peterson, and M.A. Rotea, “H∞ optimal
control with state-feedback,” IEEE Transactions on Automatic Control, vol.
AC-33, pp. 786-788, 1988.
[Kim:] Kimura, H., “Synthesis of H∞ controllers based on conjugation,”
Proceedings of IEEE Conference on Decision and Control, Austin, TX., pp.
1207-1211, 1988.
[Kwak:] Kwakernaak, H., “A polynomial approach to minimax frequency
domain optimization of mutlivariable feedback systems,” International
Journal of Control, pp. 117-156, 1986.

3-71
3 H∞ Control and Model Reduction

[Kuc1:] Kucera V., “A contribution to matrix quadratic equations,” IEEE

Transactions on Automatic Control, vol. AC-17, no. 3, pp. 344-347, 1972.
[LatA:] Latham, G.A. and B.D.O. Anderson, “Frequency-weighted optimal
Hankel norm approximation of state transfer functions,” Systems and Control
Letters, vol. 5, pp. 229-236, 1985.
[LimH:] Limebeer, D.J.N. and G.D. Halikias, “A controller degree bound for
H∞-optimal control problems of the second kind,” SIAM J. Control and Opt.,
vol. 26, no. 3, pp. 646-677, 1988.
McFG1:] 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.
[McFG2:] McFarlane, D.C. and K. Glover, “A Loop Shaping Design Procedure
using H∞ Synthesis,” IEEE Transactions on Automatic Control, vol. AC-37, no.
6, pp. 759-769, June 1992.
[Moo:] Moore, B.C., “Principal components analysis in linear systems:
controllability, observability and model reduction,” IEEE Transactions on
Automatic Control, vol. AC-26, pp. 17-31, 1981.
[MusG:] Mustafa, D. and K. Glover, “Controllers which satisfy a closed-loop H∞
norm bound and maximize an entropy integral,” Proceedings of the 27th IEEE
Conference on Decision and Control, Austin, Texas, 1988, pp. 1609-1613.
[QiuD:] Qiu, L. and E.J. Davison, “Feedback stability under simultaneous gap
metric uncertainties in plant and controller,” Systems & Control Letters, vol.
18(1), pp. 9-22, 1992.
[Red:] Redheffer, R.M., “Inequalities for a matrix Riccati equation,” Journal of
Mathematics and Mechanics, vol. 8, no. 3, pp. 349-367, 1959.
[Red2:] Redheffer, R.M., “On a certain linear fractional transformation,”
Journal of Mathematics and Physics, vol. 39, pp. 269-286, 1960.
[SafCL:] Safonov, M.G., R.Y. Chiang and D.J.N. Limebeer, “Hankel model
reduction without balancing: a descriptor approach,” Proceedings of the 26th
IEEE Conference on Decision and Control, Los Angeles, 1987.
[SafLC:] Safonov, M.G., D.J.N. Limebeer, and R.Y. Chiang, “Simplifying the
H∞ theory via loop shifting,” Proceedings of the 27th IEEE Conference on
Decision and Control, vol. 2, pp. 1399-1404, 1989.

3-72
 References

[Sara:] Sarason, D., “Generalized interpolation in H∞,” AMS. Trans., vol. 127,
pp. 179-203, 1967.
[Silv:] Silverman, L.M., “Inversion of multivariable linear systems,” IEEE
Transactions on Automatic Control, vol. AC-14, pp. 270-276, 1969.
[Toi:] Toivonen, H.T., “Sampled-data control of continuous-time systems with
an H∞ optimality criterion,” Report 90-1, Dept. of Chemical Engineering, Abo
Akademi, Finland, Jan., 1990.
[VanD:] Van Dooren, P., “A generalized eigenvalue approach for solving Riccati
equations,” SIAM J. Sci. Comput., vol. 2, pp. 121-135, 1981.
[Vin:] Vinnicombe, G., “Measuring Robustness of Feedback Systems,” PhD
dissertation, Department of Engineering, University of Cambridge, 1993.
[WanS:] Wang, W. and M.G. Safonov, “A tighter relative error bound for
balanced stochastic truncation,” Systems and Control Letters , vol. 14, pp.
307-317, 1990.
[Will1:] Willems, J.C., “Least-squares stationary optimal control and the
algebraic Riccati equation,” IEEE Transactions on Automatic Control, vol.
AC-16, pp. 621-634, 1971.
[Will2:] Willems, J.C., “Almost invariant subspaces: an approach to high gain
feedback design – Part I: almost controlled invariant subspaces.” IEEE
Transactions on Automatic Control, vol. AC-26, pp. 235-252, 1981.
[WilKS:] Willems, J.C., A. Kitapci and L.M. Silverman, “Singular optimal
control: a geometric approach.” SIAM J. Control and Opt., vol. 24, pp. 323-337,
1986.
[Wonh:] Wonham, W.M., Linear Multivariable Control: A Geometric Approach,
third edition, Springer-Verlag, New York, 1985.
[YouJB:] Youla, D.C., H.A. Jabr, and J.J. Bongiorno, “Modern Wiener-Hopf
design of optimal controllers: part II,” IEEE Transactions on Automatic
Control, vol. AC-21, pp. 319-338, 1976.
[Yam:] Yamamamoto, Y., “A new approach to sampled-data control systems—
A function space viewpoint with applications to tracking problems,”
Proceedings of the 29th Conference on Decision and Control, pp. 1882-1887,
1990.

3-73
3 H∞ Control and Model Reduction

[Zame:] Zames, G., “Feedback and optimal sensitivity: model reference

transformations, multiplicative seminorms, and approximate inverses,” IEEE
Transactions on Automatic Control, vol. AC-26, pp. 301-320, 1981.
[ZamE:] Zames, G. and El-Sakkary, “Unstable systems and feedback: The gap
metric,” Proceedings of the Allerton Conference, pp. 380-385, Oct., 1980.
[ZhoK:] Zhou, K., and P.P. Khargoneckar, “An algebraic Riccati equation
approach to H∞ optimization,” Systems and Control Letters, vol. 11, pp. 85-92,
1988.

3-74
 4
Modeling and Analysis of
Uncertain Systems
4 Modeling and Analysis of Uncertain Systems

The advanced features of the µ-Analysis and Synthesis Toolbox (µ-Tools) are
aimed at

• Analyzing the effect of uncertain models on achievable closed-loop

performance
• Designing controllers to provide optimal worst-case performance in the face
of the plant uncertainty

Hence, it is imperative that you understand

• How model uncertainty is represented in this framework

• The technical tools available to answer questions about the robustness of a
given closed-loop system to certain forms of model uncertainty

In this chapter, we concentrate on these ideas, through concepts of linear

fractional transformations and the structured singular value (µ). We begin in
the next section with linear fractional transformations (LFTs) and their role in
modeling uncertainty in matrices and systems.

4-2
 Representing Uncertainty

Representing Uncertainty
In this section, we describe the linear fractional representation of uncertainty
that is used in µ-Tools. The basic idea in modeling an uncertain system is to
separate what is known from what is unknown in a feedback-like connection,
and bound the possible values of the unknown elements. This is a direct
generalization of the notion of a state-space realization, where a linear
dynamical system is written as a feedback interconnection of a constant matrix
and a very simple dynamic element made up of a diagonal matrix of delays or
integrators. This realization greatly facilitates manipulation and computation
of linear systems, and linear fractional transformations provide the same
capability for uncertain systems.

Linear Fractional Transformations (LFTs)

Linear Fractional Transformations (LFTs) are a powerful and flexible
approach to represent uncertainty in matrices and systems. Consider first a
complex matrix M, relating vectors r and v.

v M  r v = Mr
If r and v are partitioned into a top part and bottom part, then we can draw the
relationship in more detail, explicitly showing the partitioned matrix M.

v1  M11 M12  r1 v1 = M11r1 + M12r2

v2  M21 M22  r2 v2 = M21r1 + M22r2
Suppose a matrix ∆ relates v2 to r2, as

v2 -
- r2 r2 =
v2
The linear fractional transformation of M by ∆ interconnects these two
elements, as follows,

4-3
4 Modeling and Analysis of Uncertain Systems

v1  M11 M12  r1
M21 M22 
-

Eliminate v2 and r2, leaving the relationship between r1 and v1

–1
v 1 = [ M 11 + M 12 ∆ ( I – M 22 ∆ ) M 21 ]r 1

F ( M, ∆ )
L

= F L ( M, ∆ )r 1

The notation FL indicates that the lower loop of M is closed with ∆. It is more
traditional to write a block diagram with the arrows reversed, as in

r1 - - v1
- M


This still represents the same formula, v1 = FL(M,∆)r1, and the choice of
directions is a matter of taste. We prefer to write as much as is convenient of a
block diagram with the arrows going right to left to be consistent with matrix
and operator composition, which goes the same way. This simple convention
reduces the confusion in going between block diagrams and equations,
particularly when blocks have multiple inputs.
If the upper loop of M is closed with Ω, then we have

 = U (M;
) r2
 
M11 M12 v2 F

v2 M21 M r2
22

4-4
 Representing Uncertainty

where

FU(M,Ω) := [M22 + M21Ω(I – M11Ω)–1M12]

Parametric Uncertainty
How do we use LFTs to represent an uncertain parameter? Suppose c is a
parameter, and it is known to take on values
Write this as c = 2.4+0.4δc where δc ∈ [–1,1]. This is a linear fractional
transformation. Indeed, check that
 
c = F L  2.4 0.4 , δ c
 1 0 

So, everywhere ← c ← appears in a block diagram, simply replace it with

 2:4 0:4 
1 0 
- c

If the gain c–1 also appears, the LFT representation can still be used, because
inverses of LFTs are LFTs (on the same δ). Note that

–1 1
c = ----------------------------
2.4 + 0.4δ c

1  – 1--6- δ c 
= --------  1 + ------------------------
-
2.4  1 – ( – 1--6- )δ c

 1
– 1--6- 
= FL  , δ c
--------
2.4
 1 
 --------
2.4
– 1--6- 

So, everywhere ← ← appears in a block diagram, replace it with

1
---
c

4-5
4 Modeling and Analysis of Uncertain Systems

1
--------
2.4 – 1--6-
1
--------
2.4 – 1--6-

δc

The general case for inverses can be solved with the matrix inversion lemma.
Specifically, given a matrix H, there exists matrices HLI and HUI such that for
all ∆ and Ω
[FL(H,∆)]–1 = FL(HLI,∆), [FU(H,Ω)]–1 = FU(HUI,Ω)

H 11 H 12
In fact, with H = , the formulas for HLI and HUI are just
H 21 H 22

–1 –1
H 11 – H 11 H 12
H LI =
–1 –1
H 21 H 11 H 22 – H 21 H 11 H 12

–1 –1
H 11 – H 12 H 22 H 21 H 12 H 22
H LI =
–1 –1
– H 22 H 21 H 22

Consider a second order system, representing a single degree-of-freedom mass/

damper/spring system with uncertain elements
..
mx + cx· + kx = u

The coefficients are assumed to be uncertain, with a nominal value, and a

range of possible variation

m = m ( 1 + 0.5δ m ), c = c ( 1 + 0.3δ c ), k = k ( 1 + 0.4δ k )

with –1 ≤ δm, δc, δk ≤ 1. Note that this represents 50% uncertainty in m, 30%
uncertainty in c, 40% uncertainty in k. A block diagram is shown in Figure 4-1.

4-6
 Representing Uncertainty

-, c, - 1 x - R x_ - R x -y
6@
I

u m
@ 
@ c

k 
Figure 4-1 Second Order Mass/Damper/Spring System

Define matrices

1
– 0.5 -----
M mi := m
, M c := c 0.3c , M k := k 0.4k
– 0.5 1
----- 1 0 1 0
m

Now, replace → ----

m →, ←
c ←, and ← k ←, respectively, with the block
1
-
diagrams in Figure 4-2.

Figure 4-2: Uncertain Elements as LFTs

Since we will eventually separate what is known (Mmi, Mc, MK, and
integrators) from what is unknown (δm, δc, δk), redraw the original block
diagram with the LFT representation of the uncertain elements, leaving out

4-7
4 Modeling and Analysis of Uncertain Systems

the δs, but label the signals which go to and from the δs. This is shown in
Figure 4-3.

wm - -zm R
u -, c@ M
- mi x - x- R x-y
,
6I@
 _

@ M

zc  c  wc
M

zk  k  wk
Figure 4-3: Known Part of Uncertain System

Let Gmck be the four-input (wm, wc, wk, u), four-output (zm, zc, zk, y), two-state
system shown in Figure 4-3 and depicted in Figure 4-4.

zm   wm
zc   wc
zk  Gmck  wk
y  u
Figure 4-4: Macro View of Known System

Note that Gmck only depends on m, c, k, , 0.5, 0.4, and 0.3 and the original
differential equation which relates u to y. Hence, Gmck is known. Also, the
uncertain behavior of the original system is characterized by an upper linear
fractional transformation, FU, of Gmck with a diagonal uncertainty matrix as
shown in Figure 4-5.

4-8
 Representing Uncertainty

 δm 0 0 
 
y = F U  Gmck , 0 δc 0 u
 
 0 0 δk 
 









 
∆

m 0 0
-0 c 0
0 0 k


y  Gmck
 u

Figure 4-5: Uncertain Mass/Damper/Spring System as LFT

The unknown matrix ∆, referred to as the perturbation, is structured. It has a

block-diagonal structure, and affects the input/output relationship between u
and y in an LFT (feedback) manner. During the course of this chapter, and
beyond, many models of uncertainty will involve structured perturbation
matrices and LFTs. Using sysic, we can easily compute Gmck.
mbar = 3; cbar = 1; kbar = 2;
matmi = [-0.5 1/mbar ; -0.5 1/mbar];
matc = [cbar 0.3*cbar;1 0];
matk = [kbar .4*kbar;1 0];
int1 = nd2sys([1],[1 0]);
int2 = nd2sys([1],[1 0]);
systemnames = 'matmi matc matk int1 int2';
sysoutname = 'Gmck';
inputvar = '[wm;wc;wk;u]';
input_to_matmi = '[wm;u-matc(1)-matk(1)]';
input_to_matc = '[int1;wc]';
input_to_matk = '[int2;wk]';
input_to_int1 = '[matmi(1)]';
input_to_int2 = '[int1]';
outputvar = '[matmi(1);matc(2);matk(2);int2]';
sysic;

4-9
4 Modeling and Analysis of Uncertain Systems

µ-Tools Commands for LFTs

The lower and upper LFT formulas are implemented in µ-Tools with the
command starp. The name starp comes from the star product operation
defined and developed in [Red1]. The star product is a generalization of the
LFT, and includes both the lower and upper LFTs as special cases.
Suppose that T and B are two matrices (CONSTANT, VARYING or SYSTEM)
partitioned as below.

T 11 T 12 B 11 B 12
T= , B =
T 21 T 22 B 21 B 22

such that the matrix product T22B11is well defined, and in fact, square.
If I – T22B11 is invertible, define the star product of T and B to be

–1
F L ( T, B 11 ) T 12 ( I – B 11 T 22 ) B 12
S ( T, B ) :=
–1
B 21 ( I – T 22 B 11 ) T 21 F U ( B, T 22 )

where FL and FU are defined as earlier,

–1
F L ( T, B 11 ) = T 11 + T 12 B 11 ( I – T 22 B 11 ) T 21
–1
F U ( B, T 22 ) = B 22 + B 21 T 22 ( I – B 11 T 22 ) B 12

In a block diagram, the star product appears as

 
n1 X X
T
n2  XX
 B 
S

4-10
 µ-Tools Commands for LFTs

and would be computed

S = starp(T,B,n1,n2)

B 11
Here n1 is the row dimension of [T21 T22] (and the column dimension of ),
B 21
which is the number of signals which are fed from T to B. Similarly, n2 is the

T 12
row dimension of [B11 B12] (and the column dimension of ), which is the
T 22
number of signals which are fed from B to T. The remaining inputs and outputs
appear in the output matrix S in the same order as they are in T and B. If the
dimension arguments n1 and n2 are omitted, then the following takes place:
• n1 = min([ynum(T) unum(B)])
• n2 = min([unum(T) ynum(B)])
so all possible loop closures are made. Hence, LFTs (which are special cases of
star products) are easily computed using starp without dimension arguments.
M = crandn(10,8);
Delta = crandn(3,4);
Omega = crandn(6,3);
flmd = starp(M,Delta);
fumo = starp(Omega,M);

It is possible to form the star product of two dynamical systems, and it is

possible to form the star product of their frequency responses. Hence, starp
works with all µ-Tools data types.
We can illustrate this by considering the frequency response of the uncertain
mass/damper/spring system considered in the previous section entitled
“Parametric Uncertainty”. Compute the frequency response of the four-input,
four-output system Gmck. Then, the frequency response of the uncertain system
is simply the linear fractional transformation of the frequency response of Gmck
with the perturbation matrix (δm, δc, δk).

4-11
4 Modeling and Analysis of Uncertain Systems

Gmck_g = frsp(Gmck,logspace(-1,1,100));
delnom = diag([0;0;0]);
rifd(spoles(starp(delnom,Gmck)))
delnpn = diag([-1;1;-1]);
rifd(spoles(starp(delnpn,Gmck)))
delpnp = diag([1;-1;1]);
rifd(spoles(starp(delpnp,Gmck)))
vplot('bode',starp(delnom,Gmck_g),'-',...
starp(delnpn,Gmck_g),'.-',starp(delpnp,Gmck_g),'--')

Frequency Response Magnitudes of Perturbed MCK System

2
10
Log Magnitude

0
10

2
10

4
10 1 0 1
10 10 10
Frequency (radians/sec)
Frequency Response Phases of Perturbed MCK System
0
Phase (degrees)

100

200 1 0 1
10 10 10
Frequency (radians/sec)

4-12
 Interconnections of LFTs

Interconnections of LFTs
By now, you probably have noticed an extremely important property of LFTs
— typical algebraic operations such as frequency response, cascade
connections, parallel connections, and feedback connections preserve the LFT
structure. This means that normal interconnections of LFTs are still in the
form of an LFT. Hence, the LFT is an excellent choice for a general hierarchical
representation of uncertainty. For illustrative purposes, we consider a few
additional examples in this section.
Consider a cascade connection of FL(M,∆) with FU(G,Ω), so that y =
FL(M,∆)FU(G,Ω)u. This is shown below.

Draw a box around M and G, isolating them from ∆ and Ω, calling the boxed
items Q.

Q is made up of the elements of M and G, and relates the variables (u, w∆, wΩ)
to (y, z∆, zΩ)

4-13
4 Modeling and Analysis of Uncertain Systems

From the diagram, Q is easily calculated as

y M 11 G 22 M 12 M 11 G 21 u
z ∆ = M 21 G 22 M 22 M 21 G 21 w ∆
zΩ G 12 0 G 11 wΩ

z∆ w∆
Since ∆ 0 is the matrix that relates → , it is clear that the cascade
0 Ω zΩ wΩ
connection of FL(M,∆) and FU(G,Ω) is yet another LFT, namely

 
F L ( M, ∆ )F U ( G, Ω ) = F L  Q, ∆ 0 
 0 Ω

as shown below

y  u
Q 

-

-

Similar manipulations can be carried out for parallel connections, as well as

feedback connections, and arbitrary combinations of these. For instance, a
complicated feedback connection with three LFTs

e  d1
 
1
G1 
 G2 
u1
-
1 d3 -
2
-
- y3-
- G3

3 

4-14
 Interconnections of LFTs

can be drawn as a single LFT on the diagonal matrix containing the three
individual perturbations,

Here, P depends only on G1, G2, G3, and the interconnection diagram and is
easy to calculate with the interconnection program sysic. For instance, if every
line in the diagram represents a scalar signal, then correct sysic commands to
create P are as follows:
G1 = sysrand(8,3,4);
G2 = sysrand(7,2,2);
G3 = sysrand(6,3,3);
systemnames = 'G1 G2 G3';
sysoutname = 'P';
inputvar = '[w1;w2;w3;d1;d3;u]';
input_to_G1 = '[d1;G2(1);u1;w1]';
input_to_G2 = '[G3(1);w2]';
input_to_G3 = '[d3;G1(2);w3]';
outputvar = '[G1(3);G2(2);G3(3);G1(1);G3(2)]';
cleanupsysic = 'yes';
sysic

Note The uncertainty matrix affecting P is structured, with a block-diagonal

structure. Many elements of the uncertainty matrix are known to be zero.
This is an extremely important observation. In other words, general
uncertainty at component level becomes structured uncertainty at the
interconnection level.

4-15
4 Modeling and Analysis of Uncertain Systems

Parameter Uncertainty in Transfer Functions

Suppose that we have a process with an uncertain gain, first-order lag with
uncertain time constant, and uncertain delay (modeled with a Pade
approximation). The transfer function for the process is
1 – γs + 1
y(s) = K --------------- ------------------ u ( s )
τs + 1 γs + 1
{
gain

{
{
lag Pade
(4-1)

Assume that each of the terms K, γ, and τ is uncertain, with K ∈ [1 3], γ ∈ [0.05
0.15], and τ ∈ [1 2]. Further assume that K and γ are linearly related, so that
as K takes on values from 1 → 3, γ simultaneously takes on values from 0.05 →
0.15. Represent these variations with two uncertainties, δ1 and δ2, with
K = 2 + δ1, γ = 0.1 + 0.05δ1, τ = 1.5 + 0.5δ2
where –1 ≤ δ1, δ2 ≤ 1.
– γs + 1
A block diagram of -------------------
γs + 1
is

1
Similarly, a block diagram of ---------------
τs + 1
is

Use upper-loop LFTs to model K, γ–1 and τ–1, define matrices

1 1 2
– --- 10 – --- ---
M k := 0 1 , M γI := 2
, M τI := 3 3
1 2 1 1 2
– ---
2
10 – --- ---
3 3

so that K = FU(MK,δ1), γ–1 = FU(MγI,δ1), τ–1 = FU(MτI,δ2).

4-16
 Interconnections of LFTs

The first-order Pade system is of the form FU(GP,δ1), and the first-order lag is
of the form FU(GL,δ2), where GP and GL are known, two-input, two-output,
one-state systems

GP is shown in detail in Figure 4-6 and can be built easily using sysic.

Figure 4-6: LFT System Gp for Uncertain Pade

For instance
mgammai = [-0.5 10; -0.5 10];
int = nd2sys([1],[1 0]);
systemnames = 'mgammai int';
sysoutname = 'GP';
inputvar = '[w_gamma;u1]';
input_to_mgammai = '[w_gamma;2*u1-int]';
input_to_int = '[mgammai(2)]';
outputvar = '[mgammai(1);int-u1]';
sysic;

4-17
4 Modeling and Analysis of Uncertain Systems

Similar calculations are necessary for GL, whose internal structure is shown in
Figure 4-7.
mtaui = [-1/3 2/3; -1/3 2/3];
int = nd2sys([1],[1 0]);
systemnames = 'mtaui int';
sysoutname = 'GL';
inputvar = '[w_tau;u2]';
input_to_mtaui = '[w_tau;u2-int]';
input_to_int = '[mtaui(2)]';
outputvar = '[mtaui(1);int]';
sysic;

Figure 4-7: LFT System GL for Uncertain Lag

The uncertain K is directly represented using FU(MK,δ1). The known part of

the uncertain gain/lag/Pade system is simply a four-input, four-output,
two-state system Gproc shown in Figure 4-8 with internal structure shown in
Figure 4-9.

z   w
zk   wk
z  Gproc  w
y  u
Figure 4-8: Known Part of Uncertain Gain/Lag/Pade

w z 6 w z 6 wk zk 6
? 6 ?- 6 ?- 6
-
GP GL MK
u- - - - - -y
Figure 4-9: Cascade of Pade, Lag, and Gain LFTs

4-18
 Interconnections of LFTs

The uncertain system’s behavior, equation Figure 4-1, is an LFT,

 δ I 0 
y = F U  G proc, 1 2 u
 0 δ 
 2 

The perturbation δ1 is repeated twice, due to the coupled variation in K and γ.

The SYSTEM matrix Gproc can be constructed with sysic as follows:
MK = [0 1;1 2];
systemnames = 'GP GL MK';
sysoutname = 'Gproc';
inputvar = '[w_gamma;w_k;w_tau;u]';
input_to_GP = '[w_gamma;u]';
input_to_GL = '[w_tau;GP(2)]';
input_to_MK = '[w_k;GL(2)]';
outputvar = '[GP(1);MK(1);GL(1);MK(2)]';
sysic;
We can do some time-domain simulations to verify that this model is correct.
Recall that δ1 = –1 translates into K = 1, γ = 0.05 while δ1 = 1 means K = 3,
γ = 0.15. Also, δ2 = –1 corresponds to τ = 1, and δ2 = 1 corresponds to τ = 2.
% pertnom ---> nominal values
pertnom = diag([0,0,0]);
% pert1 ---> low gain, short delay, slow lag
pert1 = diag([-1,-1,-1]);
% pert2 ---> low gain, short delay, fast lag
pert2 = diag([-1,-1,1]);
% pert3 ---> high gain, long delay, slow lag
pert3 = diag([1,1,-1]);
% pert4 ---> high gain, long delay, fast lag
pert4 = diag([1,1,1]);
tfinal = 4;
sysnom = starp(pertnom,Gproc);
ynom = trsp(sysnom,1,tfinal);
y1 = trsp(starp(pert1,Gproc),1,tfinal);
y2 = trsp(starp(pert2,Gproc),1,tfinal);
y3 = trsp(starp(pert3,Gproc),1,tfinal);
y4 = trsp(starp(pert4,Gproc),1,tfinal);
vplot(ynom,'-',y1,':',y2,'-.',y3'--',y4,'+')

4-19
4 Modeling and Analysis of Uncertain Systems

2.5

1.5
Output

0.5

0.5
0 0.5 1 1.5 2 2.5 3 3.5 4
Time

Linear State-Space Uncertainty

For general parametric uncertainty in state-space or transfer function models,
the methods outlined in the previous section are used. In the special case of
linear uncertainty in a state-space model, the uncertainty description can be
built up even more easily. Consider an uncertain state-space model,

m m

A0 + ∑ δi Ai B0 + ∑ δi Bi
x· ( t ) = i=1 i=1 x(t)
y(t) m m u(t)
C0 + ∑ δi Ci D0 + ∑ δi Di
i=1 i=1
m
 A B A B 
=  0 0 + δi i i  x ( t )
∑
 C D C i D i  u ( t )
 0 0 i
(4-2)

where for each i = 1,2,. . .,m

4-20
 Interconnections of LFTs

Ai Bi ( n+n u ) × ( n+n y )
∈R
Ci Di
Let

Ai Bi
r i := rank
Ci Di

and factor each matrix (using svd, for instance) as

Ai Bi Ei
= Gi Hi
Ci Di Fi

where

Ei ( n+n y ) × r i r i × ( n+n y )
∈R , Gi Hi ∈ R
Fi

Now, define a linear system Gss, with extra inputs and outputs via the state
equations as shown in Figure 4-10.


zzy1 

 uw
2 Gss  w12
zm  

wm

Figure 4-10: Linear System: Gss

The uncertain system in equation Figure 4-2 is represented as an LFT around

Gss, namely
y = FL(Gss,∆)u
where ∆ maps z → w, and has the structure given as

4-21
4 Modeling and Analysis of Uncertain Systems

x· A0 B0 E1 … Em x
u C0 D0 F1 … Fm u
z1 = G1 H1 0 … 0 w1
… ..

…
…

…
…
…

zm Gm Hm 0 … 0 wm

∆ = {diag[δ1 I r1 ,. . .,δm I rm ] : δi ∈ R}
This approach, developed in [MorM], has its roots in the Gilbert realization,
which is discussed in [Kai].
As an example, consider a two-state, single-input, single-output system with a
single parameter dependence.

0 1 0 0 0 0
A ( δ ) B ( δ ) := + δ
– 16 – 0.16 1 6.4 0 0
C(δ) D(δ)
16 0 0 0 0 0

The matrix multiplying δ has rank 1, and factors simply as

0 0 0 0
6.4 0 0 = 6.4 1 0 0
0 0 0 0

4-22
 Interconnections of LFTs

so the state equations for Gss are

x· 1 x1
0 1 0 0
x·2 – 16 – .16 1 6.4 x 2
=
u 16 0 0 0 u
z1 1 0 0 0 w
1

Unmodeled Dynamics
Models of uncertainty are not limited to parametric uncertainty. Often, a low
order, nominal model, which suitably describes the low-mid frequency range
behavior of the plant is available, but the high-frequency plant behavior is
uncertain. In this situation, even the dynamic order of the actual plant is not
known, and something richer than parametric uncertainty is needed to
represent this uncertainty. One common approach for this type of uncertainty
is to use a multiplicative uncertainty model. Roughly, this allows you to specify
a frequency-dependent percentage uncertainty in the actual plant behavior.
In order to specify the uncertainty set, we need to choose two things:

• A nominal model, G(s)

• A multiplicative uncertainty weighting function, Wu(s)

Given these, the precise definition of the multiplicative uncertainty set is

 ( jω ) – G ( jω ) 
M ( G, W u ) :=  G̃ : G̃
-------------------------------------- ≤ W u ( jω ) 
 G ( jω ) 

with the additional restriction that the number of right-half plane (RHP) poles
of G̃ be equal to the number of right-half plane poles of G. At each frequency,
|Wu(jω)| represents the maximum potential percentage difference between all
of the plants represented by M(G,Wu) and the nominal plant model G. In that
sense, M(G,Wu) represents a ball of possible plants, centered at G. On a
Nyquist plot, a disk of radius |Wu(jω)G(jω)|, centered at G(jω) is the set of
possible values that G̃ ( jω ) can take on, due to the uncertainty description.

4-23
4 Modeling and Analysis of Uncertain Systems

As an example, suppose that the nominal model of the plant is

1
G := ------------
s–1

and the uncertainty in the plant model is parametrized by the multiplicative

uncertainty, with uncertainty weight

--- ( --- s + 1 )
1 1
4 2
W u := ----------------------
1
------ s + 1
32

G = nd2sys([1],[1 -1]);
Wu = nd2sys([0.5 1],[0.03125 1],0.25);
omega = logspace(-2,2,80);
vplot('liv,lm',frsp(G,omega),frsp(wu,omega))

Magnitude of Nominal Plant and Percentage Uncertainty

1
10

0
10
Magnitude

1
10

2
10 2 1 0 1 2
10 10 10 10 10
Frequency (rad/sec)

4-24
 Interconnections of LFTs

It is instructive to consider sets of models that are similar to the nominal model
G, and see to what extent the sets are contained in M(G,Wu). Consider the
following problem. Determine the smallest β such that

 β 
- : β > β  ⊂ M ( G, W u )
 G -----------
 s + β 

The easiest approach using µ-Tools is just to compute, and plot

G̃ ( jω ) – G ( jω )
--------------------------------------
G ( jω )W u ( jω )

for various values of β, and determine the lower limit β by comparing the plot’s
magnitude relative to 1. Using the command ex_unc, it is easy to carry out this
procedure. For instance,
beta = 1;
Gtilde = mmult(nd2sys(beta,[1 beta]),G)
ex_unc(G,Gtilde,wu,omega); % above 1
beta = 10;
Gtilde = mmult(nd2sys(beta,[1 beta]),G)
ex_unc(G,Gtilde,wu,omega); % below 1

file: ex_unc.m
function ex_unc(G,Gtilde,Wu,omega)
Gg = frsp(G,omega);
Gtildeg = frsp(Gtilde,omega);
Wug = frsp(Wu,omega);
percdiff = vabs(vrdiv(msub(Gtildeg,Gg),mmult(Gg,Wug)));
vplot('liv,m',percdiff,1);

4-25
4 Modeling and Analysis of Uncertain Systems

A few more trials reveal β ≈ 6.1. As an exercise, carry out similar calculations
on the following examples:
• Define δ > 0 and δ > 0 as the smallest and largest numbers such that

 1+δ 
 -------------------- : δ ≤ δ ≤ δ  ⊂ M ( G, W u )
s – 1 – δ 

Exercise: Using µ-Tools, show that δ ≈ 0.33 and δ ≈ 0.425 .

• Define τ > 0 as the largest number such that

 – τs + 1 
 G ------------------- : τ ≤ τ  ⊂ M ( G, W u )
 τs + 1 

Exercise: Using µ-Tools, show τ ≈ 0.07 .

• Define r > 0 and r > 0 as the smallest and largest numbers such that

2
 70 
- : ξ ≤ ξ ≤ ξ  ⊂ M ( G, W u )
 G -------------------------------------------
2 2
 s + 2ξ70s + 70 

Exercise: Using µ-Tools, show that ξ ≈ 0.145 and ξ ≈ 5.7 .

• Define m > 0 as the largest integer such that

  50  m 
 G  s--------------
- : m = 0,1,…,m  ⊂ M ( G, W u )
+ 50 
 

Exercise: Using µ-Tools, show m = 6 .

We return to this example, and these specific extreme plants later in the
“Analysis of Controllers for an Unstable System” section in Chapter 7.
G̃ – G
Now, by defining ∆ := --------------- , each G̃ can be drawn as in Figure 4-11.
GW u

4-26
 Interconnections of LFTs

-
Behavior of
plants in M( )
G; Wu

-+ ?f+-
W u
u 6 G
- y

Figure 4-11: Multiplicative Uncertainty

In order to satisfy the constraint G̃ ∈ M ( G, W u ) , ∆ must be a transfer function

that satisfies max ∆ ( jω ) ≤ 1 , and
ω
#RHP poles [G(1 + Wu∆)] = #RHP poles [G]
Note that ∆ itself may be unstable — reconsider the previous example
involving the parameter δ. There, the unstable pole of G̃ changes location. It is
easy to verify that

( -----
32
1
- s + 1 )δs
∆ := ------------------------------------------------
-
--- ( --- s + 1 ) ( s – 1 – δ )
1 1
4 2
(4-3)

gives (by simple algebraic manipulation)

--- ( --- s + 1 ) 
1 1
1  4 2
G ( 1 + W u ∆ ) = ------------  1 + ----------------------∆
s – 1 ( -----
1
32
-s + 1) 

(s – 1)(1 + δ)
= -------------------------------------------
(s – 1)(s – 1 – δ)

1+δ
= --------------------
s–1–δ

as desired. Moreover, the ∆'s that actually yield

G(1 + Wu∆) ∈ M(G,Wu)
are those in equation (4-3) which also satisfy

max ∆ ( jω ) ≤ 1
ω∈R

4-27
4 Modeling and Analysis of Uncertain Systems

It is important to understand that in this situation, ∆, G, and Wu are not viewed

as three separate physical devices which are interconnected to form G̃ . If they
were, then the interconnection in Figure 4-11 would actually have two
unstable modes, at (1) and (1 + δ), but the mode at 1 would be uncontrollable
from u. Obviously, no controller could internally stabilize the system. Instead,
∆, G, and Wu are just combined to yield a useful manner (through the definition
of M(G,Wu)) to parametrize a set of plants that are similar (in a precise way) to
G.
As usual, we represent the uncertain plant set M(G,Wu) as an LFT by
separating the known elements (G and Wu) from the unknown elements (∆) as
shown in the following figure.

-

z ?w
6
Wu
Hmult
u - 6 -+ f?+- G - -y
Known, 2-input,
2-output, linear
system
The uncertain component is now represented as y = FU(Hmult,∆)u, with Hmult
having the value

z = 0 Wu w
y G G u
H
mult

Hmult is easily calculated using sysic or simpler manipulations.

Hmult = madd(mmult([1;0],wu,[0 1]),mmult([0;1],G,[1 1]));
% or
systemnames = 'wu G';
sysoutname = 'Hmult';
inputvar = '[w;u]';
input_to_wu = '[u]';
input_to_G = '[w+u]';
outputvar = '[wu;G]';
sysic

4-28
 Interconnections of LFTs

In summary, the multiplicative uncertainty model captures a wide variety of

plant variations, and represents not only parameter variations, but also
unmodeled dynamics. In terms of Nyquist plots, it represents disk-uncertainty
at each frequency. It is a coarse, yet simple, approach to putting uncertainty
into models.
Additive (as opposed to multiplicative) uncertainty may also be used. Given a
nominal model G and an additive uncertainty weighting function Wu, the
additive uncertainty set is

A ( G, W u ) := { G̃ : G̃ ( jω ) – G ( jω ) ≤ W u ( jω ) }

with the additional restriction that the number of right-half-plane poles of G̃ be

equal to the number of right-half-plane poles of G. The block diagram to
represent this form of uncertainty is shown in Figure 4-12. It can be expressed
as an LFT in the usual manner.

Behavior of
- Wu -
plants in A( G; W u )
u - G
-?
f - y

Figure 4-12: Additive Uncertainty

4-29
4 Modeling and Analysis of Uncertain Systems

Mixed Uncertainty
Uncertainty may be mixed, including both parametric uncertainty and
unmodeled dynamics. For an example, take the uncertain second order system
from “Parametric Uncertainty” section, with parametric uncertainty in the
parameters m, c, and k and additional high-frequency unmodeled dynamics
using the multiplicative uncertainty model. The block diagram is

m 0 0
- 0 c 0  (s) 
0 0 k

6 ? ? 6
 Wu
Hmix

y  Gmck
 ?
c 6 u

The possible behavior is represented as an LFT

y = FU(Hmix, ∆)u
where the perturbation matrix ∆ has the structure
∆ = diag{δm,δc,δk,δ(s)}
and Hmix is the five-input, five-output system defined by the following sysic
commands.
systemnames = 'Gmck wu';
sysoutname = 'Hmix';
inputvar = '[parmpertin{3} ; unmodpertin ; u]';
input_to_Gmck = '[parmpertin; unmodpertin + u]';
input_to_wu = '[u]';
outputvar = '[Gmck(1:3) ; wu ; Gmck(4)]';
sysic

4-30
 Analyzing the Effect of LFT Uncertainty

Analyzing the Effect of LFT Uncertainty

Given uncertain models, the structured singular value (µ) can be used to
analyze the robustness of the system to the structured uncertainty that enters
in the feedback form.

Using µ to Analyze Robust Stability

At this point, we have learned how to represent uncertain systems as LFTs on
unknown, structured uncertainty matrices. However, there has been no
discussion on how to analyze the robustness of these uncertain systems. This
is done with the structured singular value, µ. In this section, we perform an
analysis using the command mu, which computes upper and lower bounds for µ.
We concentrate on the mechanics of using the command, and the conclusions
that you reach with a µ-analysis.
Every µ-analysis consists of the following steps:

1 Recast (using the interconnection program sysic or Simulink®) the problem

into the familiar feedback loop diagram of Figure 4-13, where M is a known
linear system, and ∆ is a structured perturbation.

-

M
Figure 4-13: General Diagram for Robust Stability Analysis

2 Calculate a frequency response of M.

3 Describe the structure of the perturbations ∆.

4 Run the command mu on the frequency response.

5 Plot the bounds obtained from the µ calculation.

4-31
4 Modeling and Analysis of Uncertain Systems

Usually M arises by connecting an uncertain system with a feedback controller.

Consider the mass/damper/spring system in the “Mixed Uncertainty” section,
with both parametric uncertainty and unmodeled dynamics. Suppose that a
feedback controller K has been designed to improve the damping, by feedback
from y to u, namely u = Ky. Then, the uncertain closed-loop system appears in
Figure 4-14.

-


Hmix 

-K
Figure 4-14: Uncertain Closed-Loop System

By grouping Hmix and K together, namely, M := FL(Hmix,K), Figure 4-15 shows

the transformation of Figure 4-14 to Figure 4-13.

Figure 4-15: Transforming to General Form

4-32
 Analyzing the Effect of LFT Uncertainty

With M constructed, next we must clearly describe the structure of the

uncertain element ∆. The structure of the perturbation matrices must be
passed to the mu command. There are three different aspects to consider about
each block of the perturbation matrices when specifying the uncertainty
structure:

• The type (real parameter vs. unmodeled dynamics) of the perturbation

• The dimension of the perturbation
• The number of independent locations that the particular uncertainty occurs
(that is, does it affect the system in two or more places, as δ1 does in the Gain
and Pade parts of the example from the “Parametric Uncertainty” section?)

As for the software, this uncertainty structure information is stored as an

n × 2 array (called the block structure array) where n is the number of different
perturbation elements in the uncertainty matrix. Each row of this block
structure array has information to describe the uncertain block. In µ-Tools,
these conventions are:

• A scalar real parameter is denoted [-1 1] (or [-1 0]).

• A repeated (ƒ times) real parameter is denoted [-f 0].
• A 1 × 1 (ie., scalar) unmodeled dynamics perturbation (later referred to as
complex in the “Structured Singular Value Theory” section), is denoted [1 1]
or ([1 0]).
• An r × c (ie., full) unmodeled dynamics block is denoted [r c] See the
“HIMAT Robust Performance Design Example” section in Chapter 7 for an
example.

Hence, in the“Parametric Uncertainty” section, the uncertainty set for Gmck is

denoted
deltaset = [-1 1;-1 1;-1 1]; % m/c/k

The uncertainty set for Gproc in the “Parameter Uncertainty in Transfer

Functions” section is denoted
deltaset = [-2 0;-1 0]; % Gain/Pade/Lag

The mixed uncertainty example in the “Mixed Uncertainty” section is

represented as
deltaset = [-1 1;-1 1;-1 1;1 1]; % m/c/k/unmodeled

4-33
4 Modeling and Analysis of Uncertain Systems

Note that the ordering of the uncertainty elements is (and must be) consistent
with the ordering of the input/output channels of the known systems (Gmck,
Gproc, and Hmix).
As a notational convention, once the uncertainty structure has been defined for
any given problem, we will use the symbol ∆ to represent all perturbation
matrices with the appropriate structure. For example, in the mixed
uncertainty example of the “Mixed Uncertainty” section, we have

 
 δ1 0 0 0 
 
 0 δ2 0 0 
∆ :=  : δ1 ∈ R , δ2 ∈ R , δ3 ∈ R , δ4 ( s ) 
 0 0 δ3 0 
 
 0 0 0 δ4 
 

Now that the uncertainty structure has been represented, we can compute the
size of perturbations to which the system is robustly stable. We need to
calculate a frequency response of M, and then compute the structured singular
value (µ) of M with respect to the uncertainty set ∆. At each frequency, the
matrix M(jω) is passed to the µ algorithm, and bounds for µ(M(jω)) are
computed, giving upper and lower bound functions of frequency, which are
plotted. The notation for µ will be µ∆(M(jω)), to emphasize the dependency of
the function not only on M, but also on the uncertainty set ∆.
Suppose the peak (across frequency ω) of the µ∆(M(jω)) is β. This means that
for all perturbation matrices ∆ with the appropriate structure (ie., any ∆ ∈ ∆),
and satisfying maxω
σ [ ∆ ( jω ) ] < 1--β- , the perturbed system is stable. Moreover,
there is a particular perturbation matrix ∆ ∈ ∆ satisfying max σ [ ∆ ( jω ) ] < 1--β-
ω
that causes instability. Hence, we think of

1
---------------------------------------------
max µ ∆ ( M ( jω ) )
ω

as a stability margin with respect to the structured uncertainty set affecting M.

As mentioned, the software does not compute µ exactly, but bounds it from
above and below by several optimization steps. Hence, the conclusions must be
carefully stated in terms of upper and lower bounds. Let βu be the peak (across

4-34
 Analyzing the Effect of LFT Uncertainty

frequency) of the upper bound for µ, and βl be the peak of the lower bound for
µ. Then:
• For all perturbation matrices ∆ ∈ ∆ satisfying

1
max σ [ ∆ ( jω ) ] < ------
ω βu

the perturbed system is stable.

• There is a particular perturbation matrix ∆ ∈ ∆ satisfying

1
max σ [ ∆ ( jω ) ] = -----
ω βl

that causes instability.

Hence the gap between the upper and lower bounds translates into gaps
between the conclusions guaranteed robust stability and not robustly stable.
1
The destabilizing perturbation matrix (of size ----- )
β
can be constructed from the
µ calculation using the command dypert. l

It is instructive to carry out these steps on a simple example. Here, we analyze

the robust stability of a simple single-loop feedback regulation system with two
uncertainties. The plant is a lightly-damped, nominal two-state system with
uncertainty in the (2,1) entry of the A matrix (the frequency-squared
coefficient) and unmodeled dynamics (in the form of multiplicative
uncertainty) at the control input. The overall block diagram of the uncertain
closed-loop system is shown in Figure 4-16. The signals d, n, and e will be used
in the next section when robust performance is discussed.
The two-state system with uncertainty in the A matrix is represented as an
upper linear fractional transformation about a two-input, two-output,
two-state system H, whose realization is

0 1 0 0
AH BH – 16 – 0.16 1 1
:=
CH DH 6.4 0 0 0
16 0 0 0

4-35
4 Modeling and Analysis of Uncertain Systems

Uncertain Plant
- 1
z1 w1

 +w2 2 z2 Wu 
 1  H  ?e+ e
+
6+
e 8 d

n - 1
5
-++ e? - K
y u

Figure 4-16: Robust Stability/Performance Example

The resulting second order system takes the form

16
F U ( H, δ 1 ) = ---------------------------------------------------------------------
2
s + 0.16s + 16 ( 1 + 0.4δ 1 )

If we assume that δ1 is unknown, but satisfies |δ1| ≤ 1, then we interpret the

second order system to have 40% uncertainty in the denominator entry of the
natural frequency-squared coefficient.
The plant is also assumed to have unmodeled dynamics at the input. This could
arise from an unmodeled, or unreliable, actuator, for instance. The uncertainty
is assumed to be about 20% at low frequency, rising to 100% at 6 radians/
second. We model it using the multiplicative uncertainty model, using a
first-order weight

7s + 8.5
W u = ---------------------
s + 42

For this example, the controller is chosen as

2
– 12.56 s + 17.32s + 67.28
K = ------------------------------------------------------------------------------------- .
3 2
s + 20.37s + 136.74s + 179.46

Let M(s) in Figure 4-17 denote the closed-loop transfer function matrix from
Figure 4-16, after omitting δ1 and δ2. The dimensions of M are six states, four
inputs and three outputs.

4-36
 Analyzing the Effect of LFT Uncertainty

zz12 
 Ms 
 ww12
e 
 dn
( )

Figure 4-17: Closed-Loop Interconnection

In the robust stability analysis, we are only concerned about the stability of the
perturbed closed-loop system, and in that case, only the transfer function that

δ1 0
the perturbation matrix ∆ := sees is important. For notational
0 δ2 ( s )
purposes, drop the s from M(s), and partition M into

M 11 M 12
M =
M 21 M 22
(4-4)

where M11 is 2 × 2, and M22 is 1 × 2. The perturbation matrix ∆ enters the

problem as feedback around M11, as shown in Figure 4-18.

-

M11 
Figure 4-18: Example: Robust Stability

Hence, for robust stability calculations we only need the submatrix M11 (for
robust performance calculations, presented in the “Robust Performance”
section, we will need the entire matrix M).
The computational steps for analysis are given below:

1 Construct M and check dimensions.

rsexamp; % creates data
minfo(M)

4-37
4 Modeling and Analysis of Uncertain Systems

2 Calculate frequency response, and select the first two input and output
channels for the robust stability test.
omega = logspace(-2,2,200);
M_g = frsp(M,omega);
M11_g = sel(M_g,1:2,1:2);

3 Create matrix to describe the uncertainty structure (1 real parameter, 1

unmodeled dynamics uncertainty).
deltaset = [-1 0;1 1];

4 Compute µ∆(M11(jω)) using mu.

[mubnds,dvec,sens,pvec,gvec] = mu(M11_g,deltaset);

5 Plot the bounds.

vplot('liv,m',mubnds)
pkvnorm(sel(mubnds,1,1))
[pklow,omegapklow] = pkvnorm(sel(mubnds,1,2))

Robust Stability Mu Plot

0.9

0.8

0.7
Mu upper and lower bounds

0.6

0.5

0.4

0.3

0.2

0.1

0 2 1 0 1 2
10 10 10 10 10
Frequency (rads/sec)

The peak is about 0.89 (upper and lower bounds are very close in this example)
and occurs at ω = 5.4 radians/second. Hence, stability is guaranteed for all
perturbations with appropriate structure, and max σ [ ∆ ( jω ) ] < ----------
- ≈ 1.12 .
1
0.89 ω

4-38
 Analyzing the Effect of LFT Uncertainty

6 Construct the smallest destabilizing perturbation. Check that the

perturbation has the appropriate structure and that the size of the
perturbation is equal to the reciprocal of the peak of the lower bound.
pert=dypert(pvec,deltaset,mubnds);
minfo(pert)
seesys(pert)
spoles(pert)
hinfnorm(sel(pert,1,1))
hinfnorm(sel(pert,2,1))
hinfnorm(sel(pert,1,2))
hinfnorm(sel(pert,2,2))
hinfnorm(pert)

7 Verify that the perturbed system is indeed unstable. Note the location of the
perturbed closed-right-half-plane pole and its relationship with the peak of
the µ-lower bound plot.
pertM = starp(pert,M);
spoles(pertM)

4-39
4 Modeling and Analysis of Uncertain Systems

Using µ to Analyze Robust Performance

We have seen how µ can be used to analyze the robust stability of a LFT. In this
section, we use µ to determine the robust performance of a LFT under
perturbation.
Assume that the problem has been manipulated (using the program sysic or
Simulink) into the familiar feedback loop diagram of Figure 4-19, where M is a
known linear system, ∆ is a structured perturbation from a problem-dependent
allowable uncertainty set ∆, and d and e are the generalized disturbance and
error that characterizes the performance objective.
The transformation from a system block diagram to the generic diagram in
Figure 4-19 is analogous to the transformation described in this section and
Figure 4-15, except that exogenous signals (disturbances and errors) are
simply kept in the diagram.

-


e M  d
Figure 4-19: Robust Performance Formulation

Recall from Chapter 3 that the performance of MIMO control systems is

characterized using H∞ norms. Specifically, we assume that good performance
is equivalent to

T ∞ := max σ ( T ( jω ) ) ≤ 1
ω∈R

where T is some weighted, closed-loop transfer function matrix. In the case of

robust performance of uncertain systems, we’ll take T to be the uncertain
transfer function from d → e, so T = FU(M,∆).
Clearly, the transfer function from d → e is a function of ∆, through the
elements of M, and the linear fractional formula for FU. How large can the
transfer function FU(M,∆) get as ∆ takes on its allowed values? In view of this
question, the robust performance condition is defined as The LFT in Figure

4-40
 Using µ to Analyze Robust Performance

4-19 is said to achieve robust performance if it is stable for all perturbations ∆

∈ ∆ satisfying max σ [ ∆ ( jω ) ] < 1 , and moreover, if ||FU(M,∆)|| ≤ 1 for all
ω ∞
such perturbations.
How can µ be used to assess robust performance? The main idea is very simple
— you must first relate the size of a transfer function to a robust stability test.
Suppose that T is a given stable system with input dimension nd and output
dimension ne. By the Nyquist and small-gain theorem, we know that ||T||∞ ≤ 1
if and only if the feedback loop in Figure 4-20 is stable for every stable ∆F(s) (of
dimension nd × ne) satisfying ||∆F||∞ < 1.
Hence, a transfer function T is small (||T||∞ ≤ β) if and only if T can tolerate all
possible stable feedback perturbations ∆F (with ||∆F||∞ < 1--β- ) without leading to
instability. In other words, the size of a transfer function could be determined
using a robust stability test. This ultimately allows us to pose the robust
performance question as a robust stability question.

T 
,  T  1 1
-
F
Stable for all k
F k1 < 1
Figure 4-20: Performance as Robustness

Consider our situation, where T = FU(M,∆). Following the argument presented,

we have that ||FU(M,∆)||∞ ≤ 1 for all perturbations ∆ ∈ ∆ satisfying
max
ω
σ [ ∆ ( jω ) ] < 1 if and only if the LFT shown in Figure 4-21 is stable for all
∆ ∈ ∆ and all stable ∆F satisfying

max σ ( ∆ ( jω ) ) < 1 and max σ ( ∆ F ( jω ) ) < 1.

ω ω

4-41
4 Modeling and Analysis of Uncertain Systems

-


M 
-
F
Figure 4-21: Robust Stability with Augmented Uncertainty

But this is exactly a robust stability problem for M, subjected to perturbation

matrices of the form

∆ 0
∆P =
∆ 0F

Hence, we use robust stability techniques — on a larger problem, computing

µ ∆ ( M ( jω ) ) — to determine bounds on robust performance for our original
P
problem. Specifically, we use an additional (fictitious) uncertainty element,
and determine the robust stability of the extended system, and finally make
conclusions about the robust performance of the original uncertain system,
FU(M,∆).

Important Robust ||⋅||∞ Performance is characterized by introducing a

fictitious uncertainty block across the disturbance/error channels and
carrying out a Robust Stability Analysis.

In summary, each robust performance µ-analysis consists of the following

steps:

1 Recast (using the interconnection program sysic or Simulink) the problem

into the familiar feedback loop diagram of Figure 4-19, where M is a known
linear system, and ∆ is a structured perturbation, and d and e are the
generalized disturbance and error that characterizes the performance
objective.

4-42
 Using µ to Analyze Robust Performance

2 Calculate a frequency response of M.

3 Describe the structure of the perturbation set ∆.

4 Use the dimensions of the disturbance/error channels to define the

fictitiously uncertainty block, and augment this with the actual uncertainty
structure of ∆ to obtain an extended uncertainty set ∆P.

5 Compute µ∆p(M(j,ω)) on the frequency response of M, using the augmented

uncertainty set ∆P (which contains the fictitious full, transfer function
block).

6 Plot the bounds obtained from the µ calculation.

Suppose the peak of the µ-plot is β. This means that for all perturbation
matrices ∆ ∈ ∆ satisfying max ω
σ [ ∆ ( jω ) ] < 1--β- , the perturbed system is stable
and ||FU(M,∆)||∞ ≤ β. Moreover, there is a particular perturbation ∆ ∈ ∆
satisfying max σ [ ∆ ( jω ) ] = 1--β- that causes either ||FU(M,∆)||∞ = β, or instability.
ω
However, the software does not compute µ exactly, but upper and lower
bounds. Let βu be the peak (across frequency) of the upper bound for µ, and βl
be the peak (across frequency) of the lower bound for µ. Then:

• For all perturbation matrices ∆ ∈ ∆ satisfying

1
max σ [ ∆ ( jω ) ] < ------ ,
ω βu

the perturbed system is stable and ||FU(M,∆)||∞ ≤ βu.

• There is a particular perturbation matrix ∆ ∈ ∆ satisfying

1
max σ [ ∆ ( jω ) ] = -----
ω βl

that causes either ||FU(M,∆)||∞ ≥ βl, or instability.

Hence the gap between the upper and lower bounds leads to gaps in your
inability to precisely determine the robust performance.
There are two exogenous disturbances: a force disturbance d at the input to the
plant, and sensor noise n at the measurement. There is a single error, the

4-43
4 Modeling and Analysis of Uncertain Systems

output of the plant. The (weighted) open-loop transfer functions satisfy 2.6 ≤
Ted ≤ 4.0 depending on the value of 1 ≥ δ1 ≥ –1, and of course Ten = 0. Hence the
objective of control is to satisfy

1
---
2 2 2
max [ T ed ( jω ) + T en ( jω ) ] ≤ 1
ω

for all allowable perturbations.

In terms of M, we have

 δ 0 
T = F U  M, 1 
 0 δ2 ( s ) 
 

T is a 1-output, 2-input system, so the fictitious perturbation block, ∆F, must

be 2 × 1. Hence, the augmented uncertainty structure has three blocks, and
they are:

• A scalar real parameter

• A scalar, transfer function uncertainty block
• A 2 × 1 full transfer function uncertainty block
deltaset = [-1 0;1 1];
fict_blk = [2 1];
aug_deltaset = [deltaset ; fict_blk];

Now, the nominal value of T is simply FU(M,02×2), which is just M22. Plot the
magnitude (versus frequency) of these elements to see the nominal closed-loop
transfer function. In the next command, we extract M22 in two different but
equivalent manners.
vplot('liv,lm',sel(M_g,3,3:4),starp(zeros(2,2),M_g))

The robust performance µ calculation gives information about how much these
transfer functions are affected by the linear fractional perturbations δ1and δ2.

4-44
 Using µ to Analyze Robust Performance

The calculation is carried out on the entire matrix M, using the augmented
uncertainty structure.
[rpbnds,rpdvec,rpsens,rppvec,rpgvec] = mu(M_g,aug_deltaset);
vplot('liv,m',rpbnds);

Robust Performance Mu Plot

1.2

1
Mu upper and lower bounds

0.8

0.6

0.4

0.2

0 2 1 0 1 2
10 10 10 10 10
Frequency (rads/sec)

The peak value (of both lower and upper bounds) is about 1.02. This implies
that robust performance is not quite achieved. In other words, for every
perturbation ∆ = diag [δ1,δ2(s)] satisfying max σ [ ∆ ( jω ) ] < 1
-----------
1.02
, we are
ω
guaranteed stability and FU(M,∆) ≤ 1.02. Moreover, there is a perturbation ∆ =
diag [δ1,δ2(s)] with max σ [ ∆ ( jω ) ] ≈ ----------
1.02
1
- < 1 such that
ω

||FU(M,∆)||∞ ≈ 1.02 > 1.

4-45
4 Modeling and Analysis of Uncertain Systems

This can be constructed with dypert, and put into the closed-loop system to
verify the degradation of performance.
pert = dypert(rppvec,aug_deltaset,rpbnds,[1;2]);
spoles(pert)
seesys(pert)
hinfnorm(pert)
hinfnorm(starp(pert,M))
vplot('liv,lm',vnorm(starp(pert,M_g)))
vplot('liv,lm',starp(pert,M_g))

Using µ to Analyze Worst-Case Performance

In addition to determining if a system has robust performance to uncertainty,
it is useful to get the worst-case perturbation of a given size. For instance,
using perturbations of a particular structure ∆, and restricting to those of size
≤ α, what is the worst performance possible (as measured in || ⋅ ||∞ norm) and
what is the perturbation that causes the largest degradation of performance?
Precisely given α > 0, the worst-case performance associated with structured
perturbations of size α is defined as

W ( M, ∆, α ) := max F U ( M, ∆ ) ∞
∆∈∆
max σ [ ∆ ( jω ) ] ≤ α
ω

The perturbation matrix which achieves the maximum is denoted ∆bad,α.

The command wcperf assumes that the performance transfer function is an
upper loop LFT
FU(M,∆)
so that the uncertainty structure represented by deltaset is closed around the
upper loops of M. The corresponding worst-case perturbation can be used in
high-fidelity time-domain simulations to further understand the effect of
uncertainty. Plotting W(M,α) versus α yields the worst-case performance
tradeoff curve, which shows the tradeoff between size of uncertainty and
worst-case performance. This can be used to assess the relative merits of two
different controllers. The command wcperf computes the performance
degradation curves. The syntax to compute the worst-case perturbation of size

4-46
 Using µ to Analyze Robust Performance

0.8, and compute the performance degradation curves with at least 10 points is
given below. The degradation curves are shown in Figure 4-22.
alpha = 0.8;
npts = 10;
[deltabad,lowbnd,uppbnd] = wcperf(M_g,deltaset,alpha,npts);
seesys(deltabad)
hinfnorm(deltabad)
hinfnorm(starp(deltabad,M))
vplot(lowbnd,uppbnd)
grid

WorstCase Performance Degradation

2.5

2
Performance Norm Bounds

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1 1.2
Size of DELTA

Figure 4-22: Worst-Case Performance Degradation Curves

Because µ can only be bounded above and below, the worst-case performance
degradation can also only be bounded. The VARYING matrices lowbnd and
uppbnd bound the worst-case performance degradation.

4-47
4 Modeling and Analysis of Uncertain Systems

Summary
At this point, you should be well suited to explore the examples involving
µ-analysis, namely the “Analysis of Controllers for an Unstable System”,
“MIMO Margins Using m” and the “Space Shuttle Robustness Analysis”
sections in Chapter 7. The rest of this chapter is a theoretical development of
the properties of µ, and is not necessary reading at this point. Later, as you
work with the µ software, and get comfortable with the interpretation of µ as a
robust stability and robust performance measure, we encourage you to
complete reading this chapter and learn more about this powerful framework.

4-48
 Structured Singular Value Theory

Structured Singular Value Theory

This section covers a brief description of the mathematical properties of the
structured singular value, µ. The material is mathematical, relying mostly on
linear algebra concepts. Unfortunately, although many of the concepts are
simple, the notation required to state things precisely gets messy and
cumbersome. The notation used in this chapter is standard. It is listed in the
“Notation” section of Chapter 3.

4-49
4 Modeling and Analysis of Uncertain Systems

Complex Structured Singular Value

Definitions
This section is devoted to defining the structured singular value, a matrix
function denoted by µ(⋅). We consider matrices M ∈ Cn×n. In the definition of
µ(M), there is an underlying structure ∆, (a prescribed set of block diagonal
matrices) on which everything in the sequel depends. For each problem, this
structure is in general different; it depends on the uncertainty and
performance objectives of the problem. Defining the structure involves
specifying three things; the type of each block, the total number of blocks, and
their dimensions.
There are two types of blocks: repeated scalar and full blocks. Two nonnegative
integers, S and F, represent the number of repeated scalar blocks and the
number of full blocks, respectively. To keep tracking of their dimensions, we
introduce positive integers r1,. . . ,rS; m1,. . . ,mF. The ith repeated scalar block
is ri × ri, while the jth full block is mj × mj. With those integers given, we define
∆ ⊂ Cn×n as

 mj × mj 
∆ =  diag [ δ 1 I r ,…,δ s I r ,∆ 1 ,…,∆ F ] : δ i ∈ C, ∆ j ∈ C 
 
1 S

(4-5)

For consistency among all the dimensions, we must have

S F

∑ ri + ∑ mj = n.
i=1 j=1

Often, we will need norm bounded subsets of ∆, and we introduce the following
notation

B∆ = { ∆ ∈ ∆ : σ ( ∆ ) ≤ 1 }
(4-6)

Note that in Figure 4-5 all of the repeated scalar blocks appear first. This is just
to keep the notation as simple as possible, in fact they can come in any order.
Also, the full blocks do not have to be square, but restricting them as such saves

4-50
 Complex Structured Singular Value

a great deal in terms of notation. The software will handle nonsquare full
blocks, as well as any order to the blocks.
Definition 4.1: For M ∈ Cn×n, µ∆(M) is defined

1
µ ∆ ( M ) := -------------------------------------------------------------------------------------------------
min { σ ( ∆ ) : ∆ ∈ ∆ , det ( I – M∆ ) = 0 } (4-7)

unless no ∆ ∈ ∆ makes I – M∆ singular, in which case µ∆(M) := 0.

Remark 1: Without loss in generality, the full blocks in the minimal norm ∆
can each be chosen to be dyads (rank = 1). To see this, first consider the case
of only 1 full block, ∆ = Cn×n. Suppose that I – M∆ is singular. Then for some
unit-norm vector x ∈ CnM∆x = x. Define y := ∆x. It follows that y ≠ 0, and ||y||
n×n
≤ σ ( ∆ ) . Hence, define a new perturbation, ∆˜ ∈ C as

∆˜ := yx*

Obviously, σ ( ∆˜ ) = y ≤ σ ( ∆ ) , and y = ∆˜ x , so that I – M∆˜ is also singular.

Hence we have replaced a general perturbation ∆ which satisfies the
singularity condition, with a rank 1 perturbation that is no larger (in the
σ ( . ) sense), but still satisfies the singularity condition. Repeating this on a
block-by-block basis allows for each full block to be a dyad. The software
always carries out this particular construction.

Remark 2: It is instructive to consider a feedback interpretation of µ∆(M)at

this point. Let M ∈ Cn×n be given, and consider the loop shown below.

M 
u v
-

This picture is meant to represent the loop equations

4-51
4 Modeling and Analysis of Uncertain Systems

u = Mv
v = ∆u.
(4-8)

As long as I – M∆ is nonsingular, the only solutions u, v to the loop equations

are u = v = 0. However, if I – M∆ is singular, then there are infinitely many
solutions to Figure 4-8, and the norms ||u||, ||v|| of the solutions can be
arbitrarily large. With an abuse of convention, we call this constant matrix
feedback system unstable. Likewise, the term stable will describe the situation
when the only solutions are identically zero. In this context, then, µ∆(M) is a
measure of the smallest structured ∆ that causes instability of the constant
matrix feedback loop shown above.
An alternative expression for µ∆(M) follows from the definition.
Lemma 4.2:

µ∆ ( M ) = max ρ ( M∆ )
∆ ∈ B∆

In view of this lemma, continuity of the function µ : Cn×n → R is apparent. In

general, though, the function µ : Cn×n → R is not a norm, since it doesn’t satisfy
the triangle inequality. However, for any α ∈ C, µ(αM) = |α|µ(M), so in some
sense, it is related to how big the matrix is in a norm sense.
We can relate µ∆(M) to familiar linear algebra quantities when ∆ is one of two
extreme sets:

• If ∆ = {δI : δ ∈ C} (S = 1, F = 0, r1 = n), then µ∆(M) = ρ(M), the spectral radius

of M.
Proof: The only ∆’s in ∆ which satisfy the det(I – M∆) = 0 constraint are
reciprocals of nonzero eigenvalues of M. The smallest one of these is
associated with the largest (magnitude) eigenvalue, so, µ∆(M) = ρ(M).

• If ∆ = Cn×n (S = 0, F = 1, m1 = n), then µ ∆ M = σ ( M ).

- , then σ ( M∆ ) < 1 , so I – M∆ is nonsingular. Applying
Proof: If σ ( ∆ ) < -------------
1
σ(M)

equation Figure 4-7 implies µ ∆ ( M ) ≤ σ ( M ) . On the other hand, let u and v be

4-52
 Complex Structured Singular Value

unit vectors satisfying Mv = σ ( M )u , and define ∆ := 1

-------------- vu* . Then
σ(M)

σ(∆) = 1
-------------- and I – M∆ is obviously singular. Hence, µ ∆ ( M ) ≥ σ ( M ) .
σ(M)

For a general ∆ as in Figure 4-5 we must have

n×n
{ δI n : δ ∈ C } ⊂ ∆ ⊂ C
(4-9)

From the definition of µ, and the two special cases above, we conclude that

ρ ( M ) ≤ µ∆ ( M ) ≤ σ ( M )
(4-10)

These bounds alone are not sufficient for our purposes, because the gap
between ρ and σ can be arbitrarily large. They are refined by considering
transformations on M that do not affect µ∆(M), but affect ρ and σ . To do this,
define the following two subsets of Cn×n.

Q ∆ = { Q ∈ ∆ : Q*Q = I n }
(4-11)

diag [ D 1 ,…, D S ,d 1 I m ,…, d F – 1 I m ,I m ] : {

D∆ =
{ Di ∈ C
ri × ri
1 F–1 F

, D i = D *i > 0, d j ∈ R, d j > 0
.
(4-12)

Note that for any ∆ ∈ ∆, Q ∈ Q∆, and D ∈ D∆,

Q* ∈ Q ∆ Q∆ ∈ ∆ ∆Q ∈ ∆ σ ( Q∆ ) = σ ( ∆Q ) = σ ( ∆ )
(4-13)

D∆ = ∆D.
(4-14)

4-53
4 Modeling and Analysis of Uncertain Systems

Consequently,
Theorem 4.3: For all Q ∈ Q∆ and D ∈ D∆

–1
µ ∆ ( MQ ) = µ ∆ ( QM ) = µ ∆ ( M ) = µ ∆ ( DMD )
(4-15)

Proof: For all D ∈ D∆ and ∆ ∈ ∆,

det(I – M∆) = det(I – MD–1∆D) = det(I – DMD–1∆)

since D commutes with ∆. Therefore, µ∆(M) = µ∆(DMD–1). Also, for each
Q ∈ Q∆, det(I – M∆) = 0 ⇔ det(I + MQQ*∆) = 0. Since Q*∆ ∈ ∆ and
σ ( Q*∆ ) = σ ( ∆ ) ,we get µ∆(MQ) = µ∆(M) as desired. The argument for QM is
the same.
Therefore, the bounds in Figure 4-10 can be tightened to

–1
max ρ ( QM ) ≤ max ρ ( ∆M ) = µ ∆ ( M ) ≤ inf σ ( DMD )
Q ∈ Q∆ ∆ ∈ B∆ D ∈ D∆
(4-16)

where the equality comes from Lemma 2.2. Note that the last element in the D
matrices in equation Figure 4-12 is normalized to 1 since for any nonzero scalar
γ, DMD–1 = (γD)M(γD)–1.

Bounds
In this section we will concentrate on the bounds

–1
max ρ ( QM ) ≤ µ ∆ ( M ) ≤ inf σ ( DMD )
Q ∈ Q∆ D ∈ D∆

The lower bound is always an equality ([Doy]). Unfortunately, the quantity

ρ(QM) can have multiple local maxima which are not global. Thus local search
cannot be guaranteed to obtain µ, but can only yield a lower bound. The µ
software actually uses a slightly different formulation of the lower bound as a
power algorithm which is reminiscent of power algorithms for eigenvalues and
singular values ([PacD]). While there are open questions about convergence,
the algorithm usually works quite well and has proven to be an effective
method to compute µ.

4-54
 Complex Structured Singular Value

The upper bound can be reformulated as a convex optimization problem, so the

global minimum can, in principle, be found. Unfortunately, the upper bound is
not always equal to µ. For block structures ∆ satisfying 2S + F ≤ 3, the upper
bound is always equal to µ∆(M), and for block structures with 2S + F > 3, there
exist matrices for which µ is less than the infimum. This can be summarized in
the following diagram, which shows for which cases the upper bound is
guaranteed to be equal to µ.

F
0 1 2 3 4

0 yes yes yes no

S 1 yes yes no no no
2 no no no no no

Several of the boxes have connections with standard results:

• S = 0, F = 1 : µ ∆ ( M ) = σ ( M )
–1
• S = 1, F = 0: µ ∆ ( M ) = ρ ( M ) = inf σ ( DMD ) This is a standard result
D ∈ D∆
in linear algebra. It is also equivalent to fact that Lyapunov stability and
exponential stability are equivalent for linear systems.
• S = 0, F = 2: This case was studied by Redheffer ([Red1] and [Red2]).
• S = 1, F = 1: This is equivalent to a standard result on state-space
computation of H∞ norms.
• S = 2, F = 0: This is equivalent to the fact that for multidimensional systems
(2-d, and higher), exponential stability is not equivalent to Lyapunov
stability, [AndAJM], [PacD].
• S = 0, F = 3: For this case, the upper bound is always equal to µ. This
important result is due to Doyle. [Doy]
• S = 0, F ≥ 4: For this case, the upper bound is not always equal to µ. This is
important, as these are the cases that arise most frequently in applications.
Fortunately, the bound seems to be close to µ. The worst known example has
a ratio of µ over the bound of about.85, and most systems are close to 1.

The above bounds are much more than just computational schemes, although
that is their primary role in this toolbox. They are also theoretically rich, and
can unify a number of apparently quite different results in linear systems

4-55
4 Modeling and Analysis of Uncertain Systems

theory. There are several connections with Lyapunov stability, two of which
were hinted at above, but there are further connections between the upper
bound scalings and solutions to Lyapunov and Riccati equations. Indeed, many
major theorems in linear systems theory follow from the upper bounds and
some results for linear fractional transformations (see “Linear Fractional
Transformations”). The lower bound can be viewed as a natural generalization
of the maximum modulus theorem [BoyD]. While a complete exposition of these
ideas is beyond the scope of this tutorial, some of the more elementary concepts
will be explored in later sections.
For the purposes of this toolbox, the most important use of the upper bound is
as a computational scheme when combined with the lower bound. For reliable
use of the µ theory it is essential to have upper and lower bounds. The other
important feature of the upper bound is that it can be combined with H∞
controller synthesis methods to yield an ad-hoc µ-synthesis method. Note that
the upper bound, when applied to transfer functions, and maxed across
frequencies, is simply a scaled H∞ norm. This is exploited in the µ-synthesis
techniques in this toolbox.

Computational Exercise with the mu Command

The calculation of bounds for the structured singular value (µ) is performed
with the µ-Tools command mu. The input arguments to mu include the matrix
on which µ is to be calculated, the block structure associated with the input
matrix and an optional argument defining the type bound calculations to be
performed. The outputs from mu are the lower and upper bound, scaling
matrices used to achieve the upper bound, a sensitivity measure associated
with the upper bound calculation and a perturbation matrix ∆ which makes
det(I – M∆) = 0. A more detailed description of the mu can be found in the
command reference section.
The following exercise uses the command mu to compute upper and lower
bounds for the structured singular value of a given 5 × 5 complex matrix, for a
variety of block structures. It is intended to show the dependence of µ∆(M) on
the particular structure ∆. The user can also verify the correctness of the
bounds produced by the calculation.

4-56
 Complex Structured Singular Value

The set of commands used are:

mu upper and lower bounds for µ∆(M)
unwrapd unwrap the D scaling matrices associated with upper bound
unwrapp unwrap the perturbation ∆ associated with lower bound

Syntax for mu
[bnds,dvec,sens,pvec] = mu(M,deltaset);

Description

M Matrix to calculate µ of, can be a CONSTANT or VARYING

matrix.
deltaset block structure information about the set ∆; the number of
perturbation blocks, their sizes and types. For example: the
block structure

 
 δ1 0 0 0 
 
 0 δ2 0 0 
∆ :=  : δi ∈ C 
 0 0 δ3 0 
 
 0 0 0 δ4 
 

is represented by the MATLAB array

1 1
deltaset =
1 1 .
1 1
1 1

4-57
4 Modeling and Analysis of Uncertain Systems

The block structure

 
 ∆1 0 0 
 3×2 4×5 
∆ :=  0 ∆ 2 0 : ∆ 1 ∈ C ,∆ 2 ∈ C ,δ 3 ∈ C ,
 
 0 0 δ3 
 

is represented by the MATLAB array

3 2
deltaset = 4 5 .
1 1

Finally, the block structure

 
 δ1 I3 0 0 
 2 × 2
∆ :=  0 ∆ 2 0 : δ 1, δ 3 ∈ C,∆ 2 ∈ C 
 
 0 0 δ3 I2 
 

is represented by the MATLAB array

3 0
deltaset = 2 2 .
2 0

Lower and Upper bounds: The output bnds is a 1 × 2 matrix. It is VARYING

if M is VARYING, and CONSTANT if M is CONSTANT. The first entry of bnds
(either bnds(1,1) or sel(bnds,1,1) in the VARYING case) is an upper bound
for µ∆(M), and the second entry is a lower bound for µ∆(M).
Scaling matrices (these give upper bound): The matrix dvec contains the
scaling matrices for the upper bound. As in the case for bnds, it is of the same
type as M. Since the scaling matrices almost always have a number of zero

4-58
 Complex Structured Singular Value

entries, they are stored in dvec as a row vector. They can be unwrapped into
the block diagonal form using unwrapd.
[Dl,Dr] = unwrapd(dvec,deltaset);

For the most part, these two matrices are the same, in fact, if the block
structure deltaset has no nonsquare full blocks, then Dl = Dr. In any event,
the following are always true

D l = D *l > 0 ,D r = D *r > 0, D r ∆ = ∆ D l ∀∆ ∈ ∆

–1
and µ ∆ ( M ) ≤ σ ( D l MD r ) .
Sensitivity (used in µ-synthesis): a sensitivity measure of the maximum
–1
singular value of D l MD r with respect to the values in Dl (and Dr). It is
calculated in an ad-hoc manner, and is mainly used when fitting frequency
varying D’s with rational functions via the routine fitsys. We will not make
use of it in this example.
Perturbation (gives lower bound): The matrix pvec contains the
perturbation matrix ∆ which makes I – M∆ singular. This perturbation
corresponds to the lower bound in bnds. It is of the same type as M. Since the
structured set usually contains many zero elements, the perturbation matrix ∆
is stored efficiently in pvec as a row vector. It can be unwrapped into the block
diagonal form using unwrapp.
delta = unwrapp(pvec,deltaset);

Note that σ ( ∆ ) is equal to the reciprocal of the lower bound (in bnds), and

∆∈∆
det ( I – M∆ ) = 0

Try this on some examples:

simplemu;

This creates a 5 × 5 matrix, M, and the different block structures, deltaseta,

deltasetb,...,deltaseti.

4-59
4 Modeling and Analysis of Uncertain Systems

Consider the block structure defined by the array deltasete. Run the mu
command, and unwrap the ds and the perturbation.
[bnds,dvec,sens,pvec] = mu(M,deltasete);
[Dl,Dr] = unwrapd(dvec,deltasete);
delta = unwrapp(pvec,deltasete);

Verify that:
• ∆ ∈ ∆; print out the matrix delta, and check that its structure corresponds
to that given by the array deltasete.

• Compare the norm of the matrix delta with the lower bound from bnds.
bnds
norm(delta);

• Note that the σ ( ∆ ) should equal the inverse of the lower bound
(1 bnds(1,2)).
• Verify that det(I5 – M∆) = 0.
det(eye(5)-M*delta)
or
eig(M*delta) % should have an eigenvalue at 1

• Look at the scaling matrices Dl and Dr. Note that Dr∆ = ∆Dl for all ∆ ∈ ∆.
Check that the upper bound in bnds comes from these.
bnds(1,1)
norm(Dl*M/Dr)

Try the other examples, and verify the consistency of all aspects of the results.

4-60
 Mixed Real/Complex Structured Singular Value

Mixed Real/Complex Structured Singular Value

Up until this point, this chapter has dealt only with complex-valued
perturbation sets, as in equation Figure 4-5, and the complex structured
singular value to assess stability and performance degradation under these
types of perturbations. In specific instances, we have seen that it may be more
natural to model some of the uncertainty with real perturbations, for instance
when the real coefficients of a linear differential equation are uncertain. While
it is possible to simply treat these perturbations as complex and proceed with
the complex-µ analysis, the results can be expected to be conservative. Hence,
researchers have developed algorithms for robustness tests with both real and
complex perturbation blocks. The command mu also works for these extremely
useful calculations.
Hence, there are three different types of uncertainty blocks allowed in the mu
software:

• Repeated scalar real blocks

• Repeated scalar complex blocks
• Full complex blocks

The general structured singular value theory also includes full real blocks, but
these are difficult to motivate from the physics of real problems, and
convenient upper and lower bounds for structures with these types of blocks
are not well developed. This is an ongoing area of research, and later versions
of µ-Tools may support these types of blocks.
The theory for a mixed real/complex upper bound is more complicated to
describe than the upper bound theory for complex µ. In addition to D matrices,
which exploit the block diagonal structure of the perturbation, there are G
matrices, which exploit the real structure of the perturbations. For illustrative
purposes, consider a specific block structure. Suppose that

 
 δ1 I2 × 2 0 0 
 3 × 3
∆ :=  0 δ 2 I 4 × 4 0 : δ 1 ∈ R,δ 2 ∈ C ,∆ 3 ∈ C 
 
 0 0 ∆3 
 

4-61
4 Modeling and Analysis of Uncertain Systems

In mu, this is denoted by

deltaset = [-2 0;4 0;3 3];

A zero (0) in the second column signifies a repeated scalar block (as before). The
negative sign (–) indicates a real block.
Associated with ∆, define the sets

D1 0 0
2×2
D∆ = 0 D2 0 : D1 ∈ C ,det ( D 1 ) ≠ 0
0 0 d3 I3 × 3
4×4
D2 ∈ C ,det ( D 2 ) ≠ 0,d 3 ∈ C ,d 3 ≠ 0

and

 
 diag [ g i ] i = 1, 2 0 0 
 
G∆ =  0 04 × 4 0 : g i ∈ R 
 
 0 0 02 × 2 
 

If there is a β > 0, D ∈ D∆ and G ∈ G∆ such that

2 – 1--- 1 2 – 1---
σ ( I + G ) 4  --- DMD – jG ( I + G ) 4 ≤ 1
–1
β 
(4-17)

then
µ∆(M) ≤ β
This bound, [YouND1], is a derivative of an earlier bound in [FanTD]. The
smallest β < 0 for which D and G matrices exist which satisfy this constraint is
what µ-Tools calls the mixed µ upper bound for M. Using manipulations that
are now standard in robust control theory, the computation of the best such β
is reformulated into an Affine Matrix Inequality (AMI) and solved with special
purpose convex programming techniques. For perturbation sets with multiple

4-62
 Mixed Real/Complex Structured Singular Value

blocks, the general structure of the sets D∆ and G∆ remains the same, with one
scaling block for each uncertainty block.
In mu, only the complex full blocks can be nonsquare. This causes the D scaling
on the left of M to be slightly different from the D scaling on the right. The
single d variable associated with the full block is repeated a certain number of
times on the left, and a different number of times on the right, leading to
nonsquare D scaling matrices (the Dl and Dr that we have already seen). Of
course, the G scaling comes in different sizes too. Note that for any complex full
blocks, the associated blocks of G are zeros, since G is only nonzero in the blocks
associated with the real uncertainties. However, the dimension of the zero
blocks of G must line up with the correct rows/columns of M. Hence, in equation
Figure 4-17, there are three different Gs, all having exactly the same nonzero
elements, but different sizes of zero blocks associated with any nonsquare full
blocks. The different Gs are denoted Gl, Gm, and Gr. The sufficient condition
for the µ∆(M) < β is rewritten as

2 – 1--- 1 2 – 1---
σ ( I + G l ) 4  --- D l MD r – jG m ( I + G r ) 4 ≤ 1.
–1
β 
(4-18)

The lower bound generated in mu comes from a power iteration which finds
matrices ∆ ∈ ∆ that make I – M∆ singular. The power iteration for mu is a
generalization of the power iteration used in earliest versions of µ-Tools. The
generalization is described in detail in [YouD].
The combination of upper and lower bounds makes the mu software unique. The
upper bounds give a guarantee on the worst-case performance degradation
that can occur under perturbation. The lower bounds actually exhibit a
perturbation that causes significant performance degradation. This
perturbation can then be used in time-domain simulations to better
understand its effect.

4-63
4 Modeling and Analysis of Uncertain Systems

Specifics About Using the mu Command with Mixed

Perturbations
The syntax of mu and muunwrap are
[bnds,dvec,sens,pvec,gvec] = mu(M,deltaset,options);
[dl,dr,gl,gm,gr] = muunwrap(dvec,gvec,deltaset);
[delta] = muunwrap(pvec,deltaset);

Their purposes are:

mu general mixed µ computation
muunwrap extract block-diagonal scalings D and G from row vectors, and
extract from row vector the smallest perturbation ∆ which
causes singularity.

Other functions associated with mu (such as unwrapd and unwrapp, which have
been illustrated already) are detailed in Chapter 8, “Reference” under mu.

Computational Exercise with the mu Command —

Mixed Perturbations
The input argument deltaset is as before, but real repeated scalar blocks are
specified with a negative integer rather than a positive integer. For example,
deltaset = [-3 0; 2 3] is similar to deltaset = [3 0; 2 3] except that the
3 × 3 repeated block is treated as real rather than complex. Also note that
redefining deltaset = abs(deltaset) always gives the complex version of a
mixed µ problem.
For instance, the block structure

 
 δ1 0 0 0 
 
 0 δ2 0 0 
∆ :=  : δ i ∈ C,i = 1, 2, δ j ∈ R ,j = 3, 4. 
 0 0 δ3 0 
 
 0 0 0 δ4 
 

is represented by the array

deltaset = [1 1;1 1;-1 0;-1 0];

4-64
 Mixed Real/Complex Structured Singular Value

The block structure

 3×2 4×5 
∆ =  diag [∆ 1 ∆ 2 δ 3 I 3 × 3 ] : ∆ 1 ∈ C ,∆ 2 ∈ C ,δ 3 ∈ R ,
 

is represented by the array

deltaset = [3 2;4 5;-3 0];

Finally, the block structure

 
 δ1 I3 × 3 0 0 
 2 × 2
∆ :=  0 ∆2 0 : δ 1 ∈ C,δ 3 ∈ R ,∆ 2 ∈ C 
 
 0 0 δ3 I2 × 2 
 

is represented by the array

deltaset = [3 0;2 2;-2 0];

The correctness of the upper bound can easily be checked with the inequality
in equation Figure 4-18. The correctness of the lower bound can be verified by
calculating the perturbation, ∆, that mu returns, verifying its block structure
and norm, checking that the matrix M∆ has an eigenvalue exactly at 1 (which
is equivalent to I – M∆ being singular).
Try some examples on a constant 5 × 5 matrix.
simprmu

This loads a 5 × 5 complex matrix, M, and different block structures

deltaseta, deltasetb,...,deltaseti. Consider the block structure ∆
defined by deltasete, Run the mu command, and unwrap the D, G, and
perturbation matrices.
[bnds,dvec,sens,pvec,gvec] = mu(M,deltasete);
[dl,dr,gl,gm,gr] = muunwrap(dvec,gvec,deltasete);
[delta] = muunwrap(pvec,deltasete);

4-65
4 Modeling and Analysis of Uncertain Systems

Verify that:

• ∆ ∈ ∆; print out delta, and check that its block structure corresponds to that
given by deltasete.
deltasete
delta

• Compare the lower bound from µ with the norm of ∆.

[bnds(1,2) 1/norm(delta)]

• Verify that det(I5 – M∆) = 0

eig(M*delta)

• Verify the upper bound by checking the structure of dl, dr, gl, gm, and gr,
and the inequality in equation Figure 4-18.
deltasete
dl
dr
gl
gm
gr
oobdmdimjg = 1/bnds(1,1)*dl*mat/dr - sqrt(-1)*gm;
gscl_l = inv(sqrtm(sqrtm(eye(5) + gl*gl))));
gscl_r = inv(sqrtm(sqrtm((eye(5) + gr*gr))));
norm(gscl_l*oobdmdimjg*gscl_r)

Try the other block structures, and verify the consistency of all aspects of the
bounds, scaling matrices, and perturbations.

4-66
 Linear Fractional Transformations

Linear Fractional Transformations

Using only the definition of µ, some simple theorems about a class of general
matrix transformations called linear fractional transformations (LFTs) can be
proven. To introduce these, consider a complex matrix M partitioned as

M 11 M 12
M =
M 21 M 22
(4-19)

and suppose there is a defined block structure ∆2 which is compatible in size

with M22 (for any ∆2 ∈ ∆2, M22∆2 is square). For ∆2 ∈ ∆2, consider the following
loop equations,

e = M 11 d + M 12 w
z = M 21 d + M 22 w
w = ∆2 z
(4-20)

This set of equations Figure 4-20 is called well posed if for any vector d, there
exist unique vectors w, z, and e satisfying the loop equations. It is easy to see
that the set of equations is well posed if and only if the inverse of I – M22∆2
exists. If not, then depending on d and M, there is either no solution to the loop
equations, or there are an infinite number of solutions. When the inverse does
indeed exist, the vectors e and d must satisfy e = FL(M,∆2)d, where

–1
F L ( M, ∆ 2 ) = M 11 + M 12 ∆ 2 ( I – M 22 ∆ 2 ) M 21
(4-21)

FL(M,∆2) is a linear fractional transformation on M by ∆2, and in a feedback

diagram appears as

e 
M  d
z w
-
2

4-67
4 Modeling and Analysis of Uncertain Systems

The subscript L on FL pertains to the lower loop of M and is closed by ∆2. An

analogous formula describes FU(M,∆1), which is the resulting matrix obtained
by closing the upper loop of M with a matrix ∆1 ∈ ∆1.
In this formulation, the matrix M11 is assumed to be something nominal, and
∆2 ∈ B2 := { ∆ 2 ∈ ∆ 2 : σ ( ∆ 2 ) ≤ 1 } is viewed as a norm-bounded perturbation
from an allowable perturbation class, ∆2. The matrices M12, M21, and M22 and
the formula FL reflect prior knowledge on how the unknown perturbation
affects the nominal map, M11. This type of uncertainty, called linear fractional,
is natural for many control problems, and encompasses many other special
cases considered by researchers.
The constant matrix problem to solve is:

• Determine whether the LFT is well posed for all ∆2 ∈ B2, and,
• If so, then determine how large FL(M,∆2) can get for this norm-bounded set
of perturbations.

The next section has three simple theorems which answer this problem.

Well Posedness and Performance for Constant LFTs

Let M be a complex matrix partitioned as

M 11 M 12
M =
M 21 M 22
(4-22)

and suppose there are two defined block structures, ∆1 and ∆2, which are
compatible in size with M11 and M22 respectively. Define a third structure ∆ as
 
 ∆ 0 
∆ =  1 : ∆ 1 ∈ ∆ 1, ∆ 2 ∈ ∆ 2 .
 0 ∆2 
  (4-23)

Now there are three structures with which we may compute µ. The notation we
use to keep track of this is as follows: µ1(⋅) is with respect to ∆1, µ2(⋅) is with
respect to ∆2, : µ∆ (⋅) is with respect to ∆. In view of this, µ1(M11), µ2(M22), and
µ∆(M) all make sense, though for instance, µ1(M) does not. Again, define the
norm-bounded perturbation sets as

4-68
 Linear Fractional Transformations

B i : = { ∆ i ∈ ∆ i : σ ( ∆ i ) ≤ 1 }.

Let ∆2 ∈ ∆2. The linear fractional transformation, FL(M,∆2) is well posed if

I - M22∆2 is invertible, and in that case is defined as
–1
F L ( M, ∆ 2 ) = M 11 + M 12 ∆ 2 ( I – M 22 ∆ 2 ) M 21
(4-24)

The first theorem is nothing more than restating the definition of µ.

Theorem 4.4: The linear fractional transformation FL(M,∆2) is well posed for
all ∆2 ∈ B2 if and only if µ2(M22) < 1.
As the perturbation ∆2 deviates from zero, the matrix FL(M,∆2) deviates from
M11. The range of values that µ1(FL(M,∆2))takes on is intimately related to
µ∆(M), as follows:
Theorem 4.5 [Main Loop Theorem]: The following are equivalent:

1 µ∆(M) <1

2 µ2(M22) < 1, and

max µ 1 ( F L ( M, ∆ 2 ) ) < 1
∆ ∈B
2 2

3 µ1(M11) < 1, and

max µ 2 ( F U ( M, ∆ 1 ) ) < 1
∆ ∈B
1 1

Proof: The proof is based on Schur formulae for determinants of block

partitioned matrices, and can be found in [PacD]. The basic idea is simple, and
we present the proof showing 3 → 1.

4-69
4 Modeling and Analysis of Uncertain Systems

Let ∆i ∈ ∆i be arbitrary, with σ ( ∆ i ) ≤ 1 . Define

∆1 0
∆ :=
0 ∆2

Obviously, ∆ ∈ ∆, and σ ( ∆ ) ≤ 1 .
Now

I – M 11 ∆ 1 – M 12 ∆ 2
det ( I – M∆ ) = det
M 21 ∆ 1 I – M 22 ∆ 2

Since µ1(M11) < 1, and σ ( ∆ 1 ) ≤ 1 , it follows that I – M11∆1 is invertible,

giving (using the Schur formula)

det ( I – M∆ ) =
–1
det ( I – M 11 ∆ 1 )det ( I – M 22 ∆ 2 + – M 21 ∆ 1 ( I – M 11 ∆ 1 ) M 12 ∆ 2 )

Collecting the ∆2 terms leaves

det(I – M∆) = det(I – M11∆1)det(I – FU(M,∆1)∆2)
Since each ∆i ∈ Bi, the assumption implies that µ2(FU(M,∆1)) < 1. By
definition, then,
I – FU(M,∆1)∆2
is nonsingular, and so then is I – M∆. Now, this argument holds for any
∆i ∈ ∆i, hence the definition of µ gives
µ∆(M) < 1
Remark: This theorem forms the basis for all uses of µ in linear system
robustness analysis, whether from a state-space, frequency domain, or
Lyapunov approach.

4-70
 Frequency Domain µ Review

Frequency Domain µ Review

Since the frequency domain µ tests play a key role in robustness analysis, we
summarize in this section the computational procedure and its subsequent
interpretations. These involve not only constant, complex structured
perturbations, but also linear, time-invariant, dynamical perturbations
(unmodeled dynamics) as well.

Robust Stability
The most well-known use of µ as a robustness analysis tool is in the frequency
domain. Suppose M(s) is a stable, multi-input, multi-output transfer function
of a linear system, M. For clarity, assume M has nz inputs and nw outputs. Let
∆ be a block structure, as in equation Figure 4-5, and assume that the
n ×n
dimensions are such that ∆ ⊂ C z w . We want to consider feedback
perturbations to M which are themselves dynamical systems, with the
block-diagonal structure of the set ∆.
In this section, we outline the proofs for the situation where the perturbations
are assumed to be stable. This is not a restriction with parametric real
uncertainty, as constant parameters are clearly stable. However, when using a
multiplicative or additive unmodeled dynamics perturbation to model
uncertainty in an unstable component (see the example in the “Unmodeled
Dynamics” section), it is useful to allow unstable perturbations, with the
restriction that the number of right-half-plane poles of the component remains
constant. An alternate approach is to allow a block of the perturbation matrix
to be an unstable transfer function. However, we restrict it to only take on
values that preserve the number of right-half-plane poles of the perturbed
component with which it is associated. In this case, the theorems we state are
still correct, though the proofs must be modified. In fact, even more
sophisticated assumptions about the perturbed systems can be made, including
structured coprime factor uncertainty and gap metric uncertainty, but these
are beyond the scope of this tutorial.
Let S denote the set of real-rational, proper (no poles at s = ∞), stable, transfer
matrices. Associated with any block structure ∆, let S∆ denote the set of all
block diagonal, stable rational transfer functions, with block structure like ∆.

S ∆ := { ∆ ∈ S : ∆ ( s o ) ∈ ∆ for all s o ∈ C + }

4-71
4 Modeling and Analysis of Uncertain Systems

Theorem 4.6: Let β > 0. The loop shown in Figure 4-23 is well-posed and
internally stable for all ∆ ∈ S∆ with ∆ ∞ < 1--β- if and only if

M ∆ := sup µ $∆ ( M ( jω ) ) ≤ β
ω∈R


- M
Figure 4-23: Robust Stability

Explanation of proof: This result is proven using the definition of µ, and

the multivariable Nyquist criterion, [CheD]. Since M(s) and ∆(s) are assumed
stable, the loop is internally stable if and only if the Nyquist plot of
det (I – M(jω)∆(jω)) does not pass through, or encircle the origin. If ||M||∆ ≤ β,
this guarantees that for any ∆ ∈ S∆, with ||∆||∞ < 1--β- , any ω ∈ R, and any
α ∈ [0, 1], the determinant
det[I – M(jω)(α ∆(jω))] ≠ 0
This guarantees that the Nyquist plot of det[I], (the above expression with
α = 0) and the Nyquist plot of det[I – M(jω)∆(jω)], (above expression with
α = 1) encircle the origin the same number of times. It is obvious that the
Nyquist plot of det[I] is just a single point at 1, and hence does not encircle the
origin at all, therefore the Nyquist plot of det[I – M(jω)∆(jω)] does not pass
through, or encircle the origin, and hence the loop is indeed stable. Conversely,
if ||M||∆ > β, then at some particular frequency ω, µ ∆ ( M ( jω ) ) > β . By definition
of µ∆(⋅), this means there is a constant matrix ∆ ∈ ∆ such that σ ( ∆ ) < 1--β- , and
det ( I – M ( jω )∆ ) = 0 . Now, it is possible to find a real-rational, stable block
diagonal transfer matrix ∆ ∈ S∆ such that ∆ ∞ < 1--β- , and ∆ ( jω ) = ∆ . Hence,
using this real-rational ∆, the loop will have a closed-loop pole at s = jω . A
more careful proof of these ideas is found in [CheD]. An alternate proof using
the maximum-modulus property of µ can be found in [PacP].
In summary, the peak value on the µ plot of the frequency response that the
perturbation sees determines the size of perturbations that the loop is robustly
stable against.

4-72
 Frequency Domain µ Review

Robust Performance
Stability is not the only property of a closed-loop system that must be robust to
perturbations. Typically there are exogenous disturbances acting on the
system (wind gusts, sensor noise) which result in tracking and regulation
errors. Under perturbation, the effect that these disturbances have on error
signals can greatly increase. In most cases, long before the onset of instability,
the closed-loop performance will degrade to the point of unacceptability. Hence
the need for a robust performance test. Such a test will indicate the worst-case
level of performance degradation associated with a given level of perturbations.
Assume M is a stable, real-rational, proper transfer function, with nz + nd
inputs, and nw + ne outputs. Partition M in the obvious manner, so that M11
has nz inputs and nw outputs, and so on. Let ∆ ⊂ C n w × n z be a block structure,
as in equation Figure 4-5. Define an augmented block structure

 ∆ 0 nd × ne 
∆ P :=  : ∆ ∈ ∆, ∆ F ∈ C 
 0 ∆F 

The setup is to address theoretically the robust performance questions about

the loop shown in Figure 4-24.
The perturbed transfer function from d to e is denoted by FU(M,∆).
Theorem 4.7: Let β > 0. For all ∆(s) ∈ S∆ with ∆ ∞ < 1--β- , the loop shown above
is well-posed, internally stable, and ||FU(M,∆)||∞ ≤ β if and only if

M ∆ := sup µ ∆ ( M ( jω ) ) ≤ β
P P
ω∈R

-


e  M  d
Figure 4-24: Robust Performance

The proof of this is exactly along the lines of the earlier proof, but also applying
Theorem 4.5. See [DoyWS] and [PacD] for details.

4-73
4 Modeling and Analysis of Uncertain Systems

Real vs. Complex Parameters

There are many approaches to model real, uncertain parameters. Suppose that
a coefficient c, in a particular system, is assumed to be constant, but unknown,
and the value of c is modeled to lie in an interval, say,
c ∈ [0.8 1.6]
This can be modeled effectively with a real perturbation,
c ∈ {1.2 + (0.4)δ : δ ∈ R, |δ| ≤ 1}
Clearly, this set captures the uncertainty in the coefficient c. What is the
correct interpretation of the uncertain set model if δ is taken as complex,
c ∈ {1.2 + (0.4)δ : δ ∈ C, |δ| ≤ 1} ?
In a linear, time-invariant system, robustness to this constant, complex
uncertain parameter c is mathematically equivalent to robustness to all stable,
linear, time-invariant transfer functions, ĉ ( s ) , whose Nyquist plots lie in the

'$
disk shown in Figure 4-25.

Im 6 C

&%
0:8

0:4

-
,0:8 ,0:4 0:4 0:8 1:2 1:6 2:0
Re
,0:4
,0:8

Figure 4-25: Complex Disc Covering Real Interval: Restriction of the Nyquist
Plot of ĉ

Using complex parameters, the uncertain model for c represents a stable linear
system whose characteristics are similar to an uncertain real gain, but deviate
in a manner quantified by the disc-shaped constraint on its frequency response.
In general, using disks instead of intervals leads to more conservative
robustness properties. With mu, it is easy and fast to explore the differences in
the robustness properties as the uncertainty model changes.

4-74
 Real vs. Complex Parameters

Block Structures with All Real Blocks

The function µ is not necessarily a continuous function when all of the
perturbation blocks are real. This mathematical fact is pointed out in
[BarKST], and an example is given where the robustness margin to real
parameter uncertainty changes abruptly for infinitesimal changes in the
problem data. Also, in the [BarKST] example, the structured singular value of
the frequency response exhibits discontinuities across frequency.
What is the significance of these issues on mu? The discontinuities can cause
problems in the convergence of the lower bound power algorithm. For problems
with purely real uncertainty, the lower bound algorithm may converge to a
value which is significantly lower than µ itself, or may not even converge at all.
This could be a serious problem, but usually it is not, because almost all
problems have a full complex block associated with a robust performance
specification. It turns out that if a µ problem has a complex block that counts,
then the function µ will be continuous at the problem data.
Sometimes, though, a robust stability calculation for an uncertain system with
only real uncertainties is needed. Consider the robust stability problem
represented in Figure 4-26.

-
R

M 
Figure 4-26: Robust Stability with Real Uncertainty

While the upper bound from mu will be effective, the lower bound potentially
will have convergence problems, yielding little information in terms of bad
perturbations. A fix, which has both engineering and mathematical
justification, is available.

4-75
4 Modeling and Analysis of Uncertain Systems

In µ-Tools vernacular, let M be the VARYING matrix (usually a frequency

response) of which µ is to be computed. Let blkrs_R be the real block structure,
that is compatible in dimension with M. Recall that abs(blkrs_R) is exactly the
same block structure as blkrs_R, but consists of complex blocks. Define:
pdim = ynum(M); %only real blocks --> square
alpha = 0.1;
fixl = [eye(pdim) ; alpha*eye(pdim)];
fixr = fixl';
% duplicate complex blocks below reals
blkrs_RC = [blkrs_R ; abs(blkrs_R)];
Mmix = mmult(fixl,M,fixr);
[bnd_RC,rowd_RC,sens,rowp_RC] = mu(Mmix,blkrs_RC);

What does all of this mean? The block structure has been expanded (to twice
the original size) by including complex blocks which are exactly the same
dimension as the original real blocks. The matrix has been expanded (to twice
the original size) by the multiplication on the left and right. The scale factors
of α = 0.1 imply that the input/output channels, which the complex uncertainty
affects, are each scaled down by a factor of 10, giving an overall scaling of the
complex blocks of 0.01. The µ calculation determines upper and lower bounds
for robust stability in the uncertain system shown in Figure 4-27.

-
R -
R + 2
C

  M  c?
6  M 

-
C
Figure 4-27: Replacing Real Uncertainty with Real+Complex Uncertainty

In this process, each real parameter δR has been replaced by a real parameter
plus a smaller complex parameter (α2δC). Rather than computing robustness
margins to purely real parameters, the modified problem determines the
robust stability characteristics of the system with respect to predominantly

4-76
 Real vs. Complex Parameters

real uncertainties, though each uncertainty is allowed to have a very small

complex part. This slight variation in the uncertainty model is easy to accept
in engineering problems, since models of uncertainty are rarely fixed, and
small amounts of phase in coefficients of physical models usually can be
explained by some underlying dynamics that have been ignored. It can also be
proven that as the parameter α converges to 0, the calculated robustness
margin, which is a function of α, converges to the robustness margin associated
with just the original real parameters.
In summary, the modified problem contains both real and complex uncertainty;
consequently:

• µ with respect to this structure has guaranteed continuity properties.

• The lower bound from mu generally has better convergence behavior.

Moreover, although the modified problem is not equivalent to the original

problem, it has relevant engineering interpretation and provable convergence
properties as the complex weighting α goes to 0.
As an example of this, consider the 9 × 9 transfer function M(s), with a robust
stability uncertainty block structure of
∆R := {diag[δi]i=1,. . .,9 : δi ∈ R},
from the “Space Shuttle Robustness Analysis” section in Chapter 7. Using mu,
calculate the robust stability of the system in Figure 4-26, and for α = 0.1, 0.2,
0.3 in Figure 4-27.

4-77
4 Modeling and Analysis of Uncertain Systems

load shut_rs
minfo(clp_muRS)
blkrs_R = [-1 0;-1 0;-1 0;-1 0;-1 0;-1 0;-1 0;-1 0;-1 0];
clp_muRSg = frsp(clp_muRS,logspace(-2,2,40));
fix1 = [eye(9) ; 0.1*eye(9)];
fix2 = [eye(9) ; 0.2*eye(9)];
fix3 = [eye(9) ; 0.3*eye(9)];
blk_RC = [blkrs_R ; abs(blkrs_R)];
m1 = mmult(fix1,clp_muRSg,fix1');
m2 = mmult(fix2,clp_muRSg,fix2');
m3 = mmult(fix3,clp_muRSg,fix3');
[rbnd,rp] = mu(clp_muRSg,blkrs_R);
[bnd1,rd1,s1,rp1] = mu(m1,blk_RC);
[bnd2,rd2,s2,rp2] = mu(m2,blk_RC);
[bnd3,rd3,s3,rp3] = mu(m3,blk_RC);
allbnds = abv(rbnd,bnd1,bnd2,bnd3);
vplot('liv,d',sel(allbnds,':',1))
title('UPPER BOUNDS: 0%, 1%, 4%, 9% COMPLEX')
xlabel('FREQUENCY, RAD/SEC')
ylabel('UPPER BOUND')
vplot('liv,d',sel(allbnds,':',2))
title('LOWER BOUNDS: 0%, 1%, 4%, 9% COMPLEX')
xlabel('FREQUENCY, RAD/SEC')
ylabel('LOWER BOUND')

4-78
 Real vs. Complex Parameters

UPPER BOUNDS: 0%, 1%, 4%, 9% COMPLEX

1

0.9

0.8

0.7

0.6
UPPER BOUND

0.5

0.4

0.3 0% − SOLID
1% − DASHED
4% − DOTTED
0.2 9% − DASHED/DOTTED

0.1

0
−2 −1 0 1 2
10 10 10 10 10
FREQUENCY, RAD/SEC

LOWER BOUNDS: 0%, 1%, 4%, 9% COMPLEX

0.9

0% − SOLID
0.8
0% − SOLID
1% − DASHED
4% − DOTTED
0.7 9% − DASHED/DOTTED

0.6
LOWER BOUND

0.5

0.4

0.3

0.2

0.1

0
−2 −1 0 1 2
10 10 10 10 10
FREQUENCY, RAD/SEC

For each case, the upper bound shows a slight increase as the percentage of
allowable complex perturbation is increased. This is expected. The lower bound
behaves similarly, though the introduction of very small complex terms has a
more dramatic effect. For 0%, the lower bound from mu is zero — the program
is simply unable to find purely real perturbations which cause singularity.
However, upon introducing a 1% complex term in each perturbation, the lower

4-79
4 Modeling and Analysis of Uncertain Systems

bound jumps up to 0.76, indicating mu has found a diagonal perturbation

causing singularity, with each term of the perturbation having the form δR +
0.01δC, δR ∈ R, δC ∈ C, and

1 1
δ R ≤ -----------, δ C ≤ -----------
0.76 0.76

Using unwrapp, this perturbation can be constructed. For larger values of

complex contribution, the lower bound is even better behaved, and more closely
matches up with the corresponding upper bound. Of course, the relevance of
these nonreal perturbations on the actual robust stability properties must be
assessed through careful engineering judgement and considerations.
[PacP] contains a more detailed discussion of the continuity properties of mixed
µ, and the general relationship between the robustness bound in Figure 4-26
and Figure 4-27.
Finally, in defense of mu — usually, robustness problems involve a performance
objective, beyond robust stability. This performance objective leads to a full
complex block in the uncertainty block structure (when calculating robust
performance tests, the performance block must be a complex block). In those
cases, the lower bound performance of mu is quite well-behaved, and the small
complex augmentation to the real blocks is unnecessary. This real perturbation
problem is explored in detail in the “Space Shuttle Robustness Analysis”
section in Chapter 7. Also, the robust performance characteristics are
successfully computed using mu, without adding small amounts of complex
uncertainty to each real parameter.

4-80
 Generalized µ

Generalized µ
Generalized µ allows us to put additional constraints on the directions that
I - M∆ becomes singular. Given a matrix M ∈ Cn×n, and C ∈ Cm×n, find the
smallest ∆ ∈ ∆ such that

I – ∆M
C

is not full column rank. In terms of a mathematical expression, we write

1
µ ∆ ( M, C ) : = -----------------------------------------------------------------------------------------------------
 
min  σ ( ∆ ) : ∆ ∈ ∆ ,rank I – ∆M < n 
 C 

This quantity can be bounded above easily, using standard µ ideas. Suppose
µ∆(M,C) ≥ β
Then, there is a ∆ ∈ ∆, σ ( ∆ ) ≤ 1--β- and a nonzero vector x such that
(I – ∆M)x = θn, Cx = θm
Hence, for every matrix Q ∈ Cn×m, it follows that
(I – ∆(M + QC))x = θn
so that for every matrix Q ∈ Cn×m, µ∆(M + QC) ≥ β.
By contrapositive, if there exists a matrix Q such that
µ∆(M + QC) < β
then µ∆(M,C) < β.
Hence, we have

µ ∆ ( M, C ) ≤ min µ ∆ ( M + QC )
n×m
Q∈C

4-81
4 Modeling and Analysis of Uncertain Systems

Now, it is possible to compute the optimal matrix Q which minimizes the

standard upper bound for µ∆(M + QC). The optimization problem can be
reformulated into an affine matrix inequality, [PacZPB], and solved with a
combination of heuristics and general purpose AMI solvers. This is how genmu
computes the upper bound µ∆(M,C).

4-82
 Using the mu Software

Using the mu Software

One of the most important uses of the mu software is to study the sensitivity of
the calculated robustness properties to the uncertainty models themselves.
That is, if a particular perturbation’s model is changed from real to complex,
how significantly does this affect the computed robustness properties? Or, if
the weighting coefficient on a particular perturbation is increased by 40%, how
significantly does this impact the computed robustness properties? If the
change is significant, then the robustness properties are extremely sensitive to
the uncertainty model (real vs. complex, effective size of the perturbation). The
appropriate and relevant representation of the uncertainty becomes an issue of
importance.
From an engineering perspective, one must have confidence in the
appropriateness of the uncertainty model on which decisions are ultimately
made. While µ-Tools cannot resolve these questions, careful use of the mu
software can help you assess the overall robustness of the closed-loop system,
and give a better understanding of the destabilizing mechanisms that are
present.

4-83
4 Modeling and Analysis of Uncertain Systems

References
Additional information on the structured singular value, linear fractional
transformations and their use in robustness analysis on uncertain linear
systems, as well as a historical account of the development of the theory can be
found in the following references.
[AndAJM:] B. Anderson, P. Agathoklis, E. Jury, and M. Mansour, “Stability
and the matrix Lyapunov equation for discrete 2-dimensional systems,” IEEE
Transactions on Circuits and Systems, Vol. 33, no. 3, pp. 261-267, March 1986.
[BarKST:] B. Barmish, P. Khargonekar, Z. Shi, and R. Tempo, “Robustness
margin need not be a continuous function of the problem data,” Systems and
Control Letters, Vol. 15, pp. 91-98, 1989.
[BoyD:] S. Boyd and C. Desoer, “Subharmonic functions and performance
bounds on linear time-invariant feedback systems,” IMA Journal of
Mathematical Control and Information, Vol. 2, pp. 153-170, 1985.
[BoyE:] Boyd, S., and El Ghaoui, L., “Method of centers for minimizing
generalized eigenvalues,” Linear Algebra and Its Applications, special issue on
Numerical Linear Algebra Methods in Control, Signals and Systems, vol. 188,
pp. 63-111, July 1993.
[CheD:] M.J. Chen and C.A. Desoer, “Necessary and sufficient condition for
robust stability of linear distributed feedback systems,” International Journal
of Control, Vol. 35, no. 2, pp. 255-267, 1982.
[Doy:] J.C. Doyle, “Analysis of feedback systems with structured
uncertainties,” IEEE Proceedings, Vol. 129, Part D, no. 6, pp. 242-250, Nov.
1982.
[DoyPZ:] J. Doyle, A. Packard, and K. Zhou, “Review of LFTs, LMIs and µ,”
Proceedings of the 30th IEEE Conference on Decision and Control, pp.
1227-1232, 1991.
[DoyS:] J.C. Doyle and G. Stein, “Multivariable feedback design: Concepts for
a classical/modern synthesis,” IEEE Transactions on Automatic Control, Vol.
AC-26, pp. 4-16, Feb. 1981.
[DoyWS:] J.C. Doyle, J. Wall and G. Stein, “Performance and robustness
analysis for structured uncertainty,” Proceedings of the 21st IEEE Conference
on Decision and Control, pp. 629-636, Dec. 1982.

4-84
 References

[FanT:] M.K.H. Fan and A.L. Tits, “Characterization and Efficient

Computation of the Structured Singular Value,” IEEE Trans. Auto. Control,
Vol. AC-31, no. 8, pp. 734-743, Aug. 1986.
[FanTD:] M. Fan, A. Tits, and J. Doyle, “Robustness in the presence of joint
parametric uncertainty and unmodeled dynamics,” IEEE Trans. Auto. Control,
vol. 36, pp. 25-38, Jan. 1991.
[Kai:] T. Kailath, Linear Systems, Prentice-Hall, 1980.
[MorM:] B. Morton and R. McAfoos, “A Mu-test for real parameter variations,”
Proceedings of the American Control Conference, pp. 135-138, 1985.
[Osb:] E.E. Osborne, “On Preconditioning of Matrices,” J. Assoc. Comp. Math.,
7, pp. 338-345, 1960.
[Ove:] M. Overton, “Large-Scale Optimization of Eigenvalues,” SIAM Journal
of Optimization, Vol. 2, no. 1, pp. 88-120, February 1992.
[PacD:] A. Packard and J. Doyle, “The complex structured singular value,”
Automatica, Vol. 29, pp. 71-109, 1993.
[PacP:] A. Packard and P. Pandey, “Continuity of the real/complex structured
singular value,” IEEE Transactions on Automatic Control, Vol. 38, no. 3, pp.
415-428, March 1993.
[PacZPB:] A. Packard, K. Zhou, P. Pandey, and G. Becker, “A collection of
robust control problems leading to LMI’s,” 30th IEEE Conference on Decision
and Control, pp. 1245-1250, Brighton, UK, 1991.
[Red1:] R. Redheffer, “Inequalities for a matrix Riccati equation,” Journal of
Mathematics and Mechanics, Vol. 8, no. 3, 1959.
[Red2:] Redheffer, “On a certain linear fractional transformation,” Journal of
Mathematical Physics Vol. 39. pp. 269-286, 1960.
[Saf1:] M.G. Safonov, “Tight bounds on the response of multivariable systems
with component uncertainty,” Allerton Conference on Communication, Control
and Computing, pp. 451-460, 1978.
[Saf2:] M.G. Safonov, “Stability and robustness of multivariable feedback
systems,” Ph.D. dissertation, MIT Press, 1980.
[Saf3:] M.G. Safonov, “Stability margins of diagonally perturbed multivariable
feedback systems,” Proceedings of the IEE, Vol. 129, Part D, no. 6, pp. 251-256,
Nov. 1982.

4-85
4 Modeling and Analysis of Uncertain Systems

[YouD:] P. M. Young and J. C. Doyle, “Computation of µ with real and complex

uncertainties,” Proceedings of the 29th IEEE Conference on Decision and
Control, pp. 1230-1235, 1990.
[YouND1:] P. M. Young, M. P. Newlin, and J. C. Doyle, “µ analysis with real
parametric uncertainty,” Proceedings of the 30th IEEE Conference on Decision
and Control, pp. 1251-1256, 1991.
[YouND2:] P. M. Young, M. P. Newlin, and J. C. Doyle, “Practical computation
of the mixed µ problem,” Proceedings of the American Control Conference, pp.
2190-2194, 1992.

4-86
 5
Control Design via
Synthesis
5 Control Design via µ Synthesis

The structured singular value, µ, is the appropriate tool for analyzing the
robustness (both stability and performance) of a system subjected to
structured, LFT perturbations. This is evident from the discussions and
examples in Chapter 4. In this section, we cover the mechanics of a controller
design methodology based on structured singular value objectives. We rely
heavily on the upper bound for µ.

5-2
 Problem Setting

Problem Setting
In order to apply the general structured singular value theory to control system
design, the control problem has been recast into the linear fractional
transformation (LFT) setting as in Figure 5-1.

-
pert
w z

e P 
 d
y u
- K
Figure 5-1 LFT Description of Control Problem

The system labeled P is the open-loop interconnection and contains all of the
known elements including the nominal plant model and performance and
uncertainty weighting functions. The ∆pert block is the uncertain element from
the set ∆pert , which parametrizes all of the assumed model uncertainty in the
problem. The controller is K. Three sets of inputs enter P: perturbation inputs
!w, disturbances d, and controls u. Three sets of outputs are generated:
perturbation outputs z, errors e, and measurements y.
The set of systems to be controlled is described by the LFT

{ F U ( P, ∆ pert ) : ∆ pert ∈ ∆ pert, max σ [ ∆ pert ( jω ) ] ≤ 1 },

ω

The design objective is to find a stabilizing controller K, such that for all such
perturbations ∆pert, the closed-loop system is stable and satisfies

||FL[FU(P,∆pert),K]||∞ ð 1.
perturbed plant

Observing Figure 5-2, it is clear that

FL[FU(P,∆pert)K] = FU[FL(P,K),∆pert].

5-3
5 Control Design via µ Synthesis

U (P;
pert)
w
-
pert z
F
-
pert
w z

 
   
e P
 d e P
 d

y
- u y
- u

L (P; K )
K K
F

Figure 5-2: Two Different Views of the Closed Loop

Therefore, the design objective is to find a nominally stabilizing controller K,

such that for all ∆pert ∈ ∆pert, maxσ [ ∆ pert ( jω ) ] ≤1, the closed-loop system is
ω
stable and satisfies
||FU[FL(P,K),∆pert]||∞ ð 1.
Given any K, this performance objective can be checked utilizing a robust
performance test on the linear fractional transformation FL(P,K). The robust
performance test should be computed with respect to an augmented
uncertainty structure,

 
 ∆ pert 0 nd × ne 
∆ :=  : ∆ pert ∈ ∆ pert, ∆ F ∈ C .
 0 ∆F 
 

The structured singular value provides the correct test for robust performance.
We know from the discussion in Chapter 4 that K achieves robust performance
if and only if

maxµ ∆ ( F L ( P, K ) ( jω ) ) < 1 .
ω

5-4
 Problem Setting

The goal of µ synthesis is to minimize over all stabilizing controllers K, the peak
value of µ∆(⋅) of the closed-loop transfer function FL(P,K). More formally,

min maxµ ∆ ( F L ( P, K ) ( j, ω ) )
K ω
stabilizing
(5-1)

This is shown in Figure 5-3.

   
min max
K !2R

 P 

-K
Figure 5-3: µ Synthesis

5-5
5 Control Design via µ Synthesis

Replacing With Its Upper Bound

For tractability of the µ synthesis problem it is necessary to replace µ∆(⋅) with
the upper bound. In Chapter 4, we saw that for a constant matrix M and an
uncertainty structure ∆, an upper bound for µ∆(M) is an optimally scaled
maximum singular value,

–1
µ∆ ( M ) ≤ inf σ ( DMD )
D ∈ D∆

Recall that Dx∆ is the set of matrices with the property that D∆ = ∆D for every
D ∈ D∆, ∆ ∈ ∆.
Using this upper bound, the optimization in equation Figure 5-1 is
reformulated as

–1
min max min σ [ D ω F L ( P, K ) ( jω )D ω ]
K ω D ∈D
stabilizing ω ∆
(5-2)

Remember, the D minimization is simply an approximation to µ[FL(P,K)(jω)].

Dω is chosen from the set of scalings, D∆, independently at every ω. Hence, we
have

–1
min min max σ [ D ω F L ( P, K ) ( jω )D ω ]
K
stabilizing
D.,D ω ∈ D ∆ ω
(5-3)

By D.,Dω ∈ D∆, we mean a frequency-dependent function D that satisfies

Dω ∈ D∆ for each ω. The general expression maxσ [ f ( ω ) ] is notated as ||ƒ||∞,
ω
giving

–1
min min DF L ( P, K )D ∞
K D.,D ω ∈ D ∆
stabilizing
(5-4)

Consider a single matrix D ∈ D∆, and a complex matrix M. Suppose that U is a

complex matrix with the same structure as D, but satisfying U*U = UU* = I.
Each block of U is a unitary (orthogonal) matrix. Recall that matrix

5-6
 Replacing µ With Its Upper Bound

multiplication by an orthogonal matrix does not affect the maximum singular

value, hence
–1 –1
σ [ ( UD )M ( UD ) ] = σ [ UDMD U* ]
–1
= σ ( DMD )

So, replacing D by UD does not affect the upper bound. Using this freedom in
the phase of each block of D, we can restrict the frequency-dependent scaling
matrix Dω of equation Figure 5-4 to be a real-rational, stable, minimum-phase
ˆ ( s ) , and not affect the value of the minimum.
transfer function, D
Hence the new optimization is

ˆ F ( P, K )D̂ – 1
min min D L ∞
ˆ
K D(s) ∈ D
stabilizing ∆
stable, min-phase
(5-5)

This optimization is currently solved by an iterative approach, referred to as

D – K iteration. A block diagram depicting the optimization is shown in
Figure 5-4.

min  ^  ^ ,1 
 D P D  1

K;D

- K
Figure 5-4: Replacing µ with Upper Bound

A specific example clarifies some of the ideas. Assume for simplicity that the
uncertainty block ∆pert only has full, unmodeled dynamics (ie., complex) blocks,
say, N of them. Then the set ∆pert is of the form

 ri × ci 
∆ pert =  diag [ ∆ 1, ∆ 2, …, ∆ N ] : ∆ i ∈ C 
 
(5-6)

5-7
5 Control Design via µ Synthesis

The set ∆ has the additional fictitious block (for the robust performance
characterization)

 ri × ci nd × ne 
∆ =  diag [ ∆ 1, ∆ 2, …, ∆ N, ∆ F ] : ∆ i ∈ C , ∆F ∈ C 
 
(5-7)

Therefore, the scaling set D∆ is easily seen to be

D ∆ = { diag [ d 1 I, d 2 I, …, d N I, I ] : d i > 0 }
(5-8)

The elements of D∆, which are defined in Figure 5-8 to be real and positive, can
be allowed to take on any nonzero complex values and not change the value of
–1
the upper bound, inf D ∈ D σ ( DMD ) . Using this freedom in the phase of each
∆
entry of D, we can restrict the frequency-dependent scaling matrix Dω of
equation Figure 5-4 to be a real-rational, stable, minimum-phase transfer
ˆ
function, d ( s ) . The optimization is now

–1
ˆ ˆ
d1 I … 0 0 d1 I … 0 0
min F L ( P, K )
K, d̂ ˆ ˆ
0 ….. d N I 0 0 …...d N I 0
…
…

…
…

…
…

.
0 … 0 I 0 … 0 I ∞

5-8
 D-K Iteration: Holding D Fixed

D-K Iteration: Holding D Fixed

To solve equation Figure 5-5, first consider holding D ( s ) fixed at a given,
stable, minimum phase, real-rational D̂ ( s ) . Then, solve the optimization

–1
min D̂F L ( P, K )D̂ ∞
K
stabilizing
(5-9)

Define PD to be the system shown in Figure 5-5.

   ^  ^ ,1 
 PD  =  D P D 
   
Figure 5-5: Absorbing Rational D Scaling

It is clear that the optimization in equation Figure 5-9 is equivalent to

min F L ( P D, K ) ∞
K
stabilizing

Since PD is known at this step, this optimization is precisely an H∞

optimization control problem, as covered in Chapter 3. The solution to the H∞
problem is well known and involves solving algebraic Riccati equations in
terms of a state-space model for PD.

5-9
5 Control Design via µ Synthesis

D-K Iteration: Holding K Fixed

With K held fixed, the optimization over D is carried out in a two-step
procedure:

1 Finding the optimal frequency-dependent scaling matrix D at a large, but

finite set of frequencies (this is the upper bound calculation for µ).

2 Fitting this optimal frequency-dependent scaling with a stable,

minimum-phase, real-rational transfer function D̂ .

The two-step procedure is a viable and reliable approach. The primary reason
for its success is the efficiency with which both of the individual steps are
carried out. The µ upper bound is based on a convex optimization problem. For
this problem, we have developed many heuristics, which when combined with
standard convex minimization tools leads to a fast, and accurate computation
of the upper bound.The fitting algorithm, using genphase and fitsys, is based
on FFT, least squares and again heuristics. It is extremely fast and works well
in most situations.

Two-Step Procedure for Scalar Entries d of D

We explain the two-step procedure for the case of scalar entries d of the D
scaling matrix, as described earlier. The two-step procedure for full-block D
scalings is covered in the next section.
A stabilizing controller, K(s), is given, and the closed-loop FL(P,K) is formed. At
each frequency, we solve the minimization corresponding to the upper bound
for µ.

–1
min σ [ D ω F L ( P, K ) ( jω )D ω ]
D ∈D
ω

This minimization is done over the real, positive Dω from the set D∆ defined in
equation Figure 5-8. This is carried out with µ, in the upper bound
optimization. Recall that the addition of phase to each di (ω) does not affect the
–1
value of σ [ D ω F l ( P, K ) ( jω )D ω ] . In other words, the important aspect of the
scaling di is its magnitude, |di (jω)|.

5-10
 D-K Iteration: Holding K Fixed

Hence, each positive function, di, which is defined on a finite set of frequencies,
is fit (in magnitude) by a proper, stable, minimum-phase transfer function,
dˆ i ( s ) . This is accomplished as follows: Use the Bode integral formulae to
determine the phase θi(ω) of the stable, minimum-phase function Li that
satisfies
|Li(jω)| = di(ω)
for all ω. Then, use the transfer function fitting routine fitsys to construct a
ˆ
real-rational transfer function d i ( s ) such that

ˆ jθ i ( ω )
d i ( jω ) ≈ e d i ( ω ) = L i ( jω )
{
{
phase magnitude

These rational functions are collected together in a diagonal transfer function

ˆ (s) ,
matrix D

ˆ ( s ) = diag [ d̂ ( s )I, d̂ ( s )I,…dˆ F – 1 ( s )I,I ]

D 1 2

and absorbed into the original open-loop generalized plant P (to yield PD, as
described earlier).

Two-Step Procedure for Full D (Optional Reading)

If there are repeated scalar uncertainty blocks, the corresponding blocks of the
D matrices are square, full, positive definite Hermitian matrices. Furthermore,
at frequencies where M(jω) is a real matrix, the optimal scaling matrix D is also
real. Fitting these requires more care than fitting the scalar d scalings
described earlier. We describe the approach used in µ-Tools. Suppose that D :
jR → Cn×n is a continuous function, with:
n×n
1 lim ω → ∞ D ( jω )=:D ∈ R

2 D(0) ∈ Rn×n

3 ∀ω ∈ R, D ( jω ) = D* ( jω ) > 0

4 D = DT > 0

Then, under some smoothness assumptions, we do the following. For each 1 ð i

ð n:

5-11
5 Control Design via µ Synthesis

1 Find a stable minimum phase rational function ĝ i such that for all ω ∈ R,

ĝ i ( jω ) ≈ D ii ( jω )

2 Define a real-valued function φi : R → R such that

jθ i ( ω ) ĝ i ( jω )
e := -------------------
ĝ i ( jω )

3 For each k = 1,2,. . .,i – 1,i + 1,. . .,n, find rational ĝ ik (not necessarily stable
or minimum phase) such that for all ω ∈ R,

j∅ i ( w )
ĝ ik ( jw ) ≈ e D ik ( jw )

Now, put all of these rational functions into a matrix

ĝ 1 ĝ 12 … ĝ 1n
ˆ ĝ 21 ĝ 2 … ĝ 2n
G ( s ) :=
..
…
…

.
ˆ g ˆ … gˆ
g n1 n2 n

and define Φ := diag[φi]. If the rational fitting was done accurately, then, for all
ω

ˆ ( jω ) = e jΦ ( ω ) D ( jω )
G
ˆ
Also, G has no poles on the imaginary axis, and has nonzero determinant
everywhere on the imaginary axis. Hence, by spectral factorization techniques
(and the command extsmp), we can find a stable, minimum-phase D ˆ and a
ˆ
unitary matrix function U such that

ˆ ˆ ( jω )D̂ ( jω )
G ( jω ) = U

ˆ is an appropriate
for all ω. The stable, minimum-phase rational matrix D
scaling for the iteration.

5-12
 Commands for D – K Iteration

Commands for D – K Iteration

Within µ-Tools there are four ways to perform D – K iteration. They are:

1 Use the graphical user interface dkitgui for automated, adjustable, and
visual iterations that allow for easy monitoring of progress.

2 Use the script file dkit (improved syntax, algorithms and ease-of-use from
version 2.0) for automated but adjustable iterations.

3 Use the dkit command in the auto mode to run a specified number of
iterations in an automatic mode (requires no user intervention).

4 Write your own iteration loop, using commands such as hinfsyn, frsp, mu,
and msf. This approach is not recommended.

5-13
5 Control Design via µ Synthesis

Discussion
There are two shortcomings with the D – K iteration control design procedure:

• We have approximated µ∆(⋅) by its upper bound. This is not a serious problem
since the value of µ 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 very 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 with
success to vibration suppression for flexible structures, flight control, chemical
process control problems, and acoustic reverberation suppression in
enclosures.
At this point, we recommend that you proceed to Chapter 7 for examples of the
iteration.

Reference
[SteD:] Stein, G., and J. Doyle, “Beyond singular values and loopshapes,” AIAA
Journal of Guidance and Control, vol. 14, num. 1, pp. 5-16, January, 1991.

5-14
 Auto-Fit for Scalings (Optional Reading)

Auto-Fit for Scalings (Optional Reading)

In dkit and dkitgui, there are routines that automatically choose the order of
the fit. This is done by comparing the effectiveness of the rational fit. For
simplicity of explanation, consider the situation where all of the uncertainty
blocks are full, so that the D-scaling matrix is made up of several scalar
functions, di.
Suppose that the kth scaling dk(ω)is fit with an rth order, stable, minimum
jθ k ( ω )
phase function d̂ k ( s ) ≈ e d k ( ω ) . The suitability of this rational fit is
assessed by first defining a scaling matrix

d1 ( ω ) … 0 0 0 … 0
..
… 0
…

. 0 0 0
0 … dk – 1 ( ω ) 0 0 … 0
D k, r ( ω ) := 0 … 0 d̂ k ( jω ) 0 … 0
0 … 0 0 dk + 1 ( ω ) … 0
..
…
…

…
…

.
0 … 0 0 0 … I

Note that in this matrix we have simply replaced the optimal

frequency-by-frequency scaling dk (obtained in the µ upper bound computation)
with the frequency response of the rational fit d̂ k . Next, we compare the values
of

–1 –1
σ [ D ( ω )F L (P, K) ( jω )D ( ω ) ] and σ [ D [ k, r ] ( ω )F L (P, K) ( jω )D [ k, r ] ( ω ) ]

If these are close, then r is deemed a suitable order for the kth scaling function
dk. The measure of closeness can be chosen. Define

–1
β := maxσ [ D ( ω )F L (P, K) ( jω )D ( ω ) ]
ω∈R

5-15
5 Control Design via µ Synthesis

Our rules are as follows. For frequencies where

–1
σ [ D ( ω )F L (P, K) ( jω )D ( ω ) ] ≥ 0.5β

we require that

–1 –1
σ [ D k, r ( ω )F L (P, K) ( jω )D k, r ( ω ) ] ≤ 1.03σ [ D ( ω )F L (P, K) ( jω )D ( ω ) ]

The quantity 1.03 is the AutoTol parameter in dkit, which can be modified
easily. In dkitgui, the Auto-Fit Tolerance, which is adjustable in the
Parameter window, varies this parameter from 1.01 (tight) to 1.06 (loose).
For other frequencies, we require that

–1 –1
σ [ D k, r ( ω )F L (P, K) ( jω )D k, r ( ω ) ] ≤ σ [ D ( ω )F L (P, K) ( jω )D ( ω ) ] + 0.1β

Note that the order of d̂ k is chosen based on its performance as a scaling while
all of the other scalings are set to their optimal (dk(ω)). You can easily modify
the constants 0.5 and 0.1 to demand tighter tolerance on the auto-fitting
algorithm.

5-16
 6
Graphical User Interface
Tools
6 Graphical User Interface Tools

This chapter describes three µ-Analysis and Synthesis Toolbox (µ-Tools)

graphical user interface tools:

• Workspace tool: wsgui

• D – K Iteration tool: dkitgui
• Simulation tool: simgui

The functionality and application of these tools is illustrated with a

multivariable control problem.

6-2
 Workspace User Interface Tool: wsgui

Workspace User Interface Tool: wsgui

The µ-Tools Workspace Manager, wsgui, is used to view variables in the
workspace, drag them to other µ-Tools graphical user interface (GUI) drop
boxes, and export variables from the µ-Tools user interfaces to the workspace.
In this section we only describe how the tool is used to view the workspace
variables. In other sections we show how the tool integrates with other µ-Tools
GUI tools.
Clear the workspace, create the weighting functions used in the Shuttle
example, and start the Workspace Manager.
clear; mk_wts; wsgui

The wsgui Workspace Manager window appears as in Figure 6-1.

Each time Refresh Variables is pressed, the MATLAB command who is
executed, and minfo is run to determine the variable type and dimension. This
information is displayed in the main scrollable table. The date and time of the
last Refresh are displayed below the button.

6-3
6 Graphical User Interface Tools

Figure 6-1 Initial wsgui Main Window

The scrollable table can be moved up/down one page by pressing above/below
the slider. Pressing the arrows at the end of the slider moves the table one line.
A filter is used to make viewing of a reduced number of selections easy. The
Prefix, Suffix, and matrix type filters are on the bottom of the scrollable
table. The matrix type filter is a pop-up menu to the right of Suffix. For
instance, let’s look at SYSTEM matrices whose names begin with w. Type a w
in the Prefix box, as shown in Figure 6-2. Note that the Apply Selection
button becomes enabled, yet all 20 matrices in the workspace are still displayed
(even those that don’t start with a w). The selection filter is only applied when
you press the Apply Selection button. Move to the right to the pop-up menu,
which currently displays All, and select System, as shown in Figure 6-2. The
Apply Selection remains enabled, and again, matrices that are not SYSTEM
matrices are still displayed. Press Apply Selection to apply the filter (first

6-4
 Workspace User Interface Tool: wsgui

letter = w, matrix type = SYSTEM). The scrollable table refreshes, leaving 7 (of
the original 20) matrices displayed, as seen in Figure 6-3.

Figure 6-2: wsgui Main Window with Selected Options

6-5
6 Graphical User Interface Tools

Figure 6-3: wsgui Main Window After Apply Selection

Now, go into the MATLAB command window, and create a new SYSTEM
matrix, with first letter w
wnew = nd2sys([1 2],[1 2 3 4]);

Note that this does not immediately appear in the scrollable table, even though
it satisfies the selection criteria. This new variable will not appear until the
Refresh Variables button is pushed. Press the Refresh Variables button and
note that when the table is refilled, the system wnew appears as expected,
Figure 6-4.

6-6
 Workspace User Interface Tool: wsgui

Figure 6-4: wsgui Main Window

Often, you want to create a more complicated selection criteria. The Custom
filter can be used to do this. Press the push button marked with a * (to the right
of the matrix type pop-up menu) to switch to the Custom filter. Lets find all
matrices with four or more outputs. In the Custom box, type
ynum(mdata)>=4

and press Apply Selection. The results are shown in Figure 6-5. This will
evaluate the expression in the Custom box, and select those workspace
variables for which the expression is true. In the Custom box, use mdata to
indicate the matrix’s value, and mname to substitute the matrix’s name. Hence,
selections can be based on the name and value of any workspace variable.

6-7
6 Graphical User Interface Tools

Figure 6-5: wsgui Main Window with Custom Selection Open

File Menu
The File menu at the top of the Workspace Manager window has three menu
items as seen in Figure 6-6.

Figure 6-6: File Menu

6-8
 Workspace User Interface Tool: wsgui

The Clear Selected Matrices item allows you to clear the variable names
currently appearing in the Workspace Manager window from the MATLAB
workspace. Upon selecting the Clear Selected Matrices item you are
prompted with the box shown in Figure 6-7.

Figure 6-7: Clear Selected Matrices Box

Pressing the Clear push button clears the selected variables from the
MATLAB workspace. Pressing Cancel cancels this command.
Similarly, the Save Selected Matrices item allows you to save the variable
names currently appearing in the Workspace Manager window to a MATLAB
MAT-file. Upon selecting Save Selected Matrices, you are prompted with the
box shown in Figure 6-8.

Figure 6-8: Save Selected Matrices Box

You must enter the name of the file in editable text frame in which to store
these variables. The default filename for the variables to be saved into is
savefile. Pressing the Save push button saves the selected variables from the
MATLAB workspace to the filename as defined by the editable text string.
Pressing Cancel cancels this command. The Quit item quits the workspace tool
and deletes the Workspace Manager window.

6-9
6 Graphical User Interface Tools

Options Menu
The Options menu at the top of the Workspace Manager window has three
menu items, as seen in Figure 6-9.

Figure 6-9: Options Menu

The CleanUp item redisplays the variable names and data appearing in the
Workspace Manager window. The other two menu items are Font and # of
Lines. These items correspond to the font type and number of items shown in
the workspace window. The Font menu item allows you to select a font size of
7 to 12 to display the data (see Figure 6-10). The # of Lines menu item allows
you to select the number of lines of data displayed in the workspace window.
You can select between [12 20 28 36 44] lines of data. This is especially useful
since the Workspace Manager window is resizable.

Figure 6-10: Font and # of Lines Menu Items

Note You must select the CleanUp item from the Options menu after
resizing.

6-10
 Workspace User Interface Tool: wsgui

Export Button
The Export button and editable text boxes at the bottom of the Workspace
Manager window allow you to export data from other µ-Tools user interfaces
to the MATLAB workspace. The Export button also allows you to copy
workspace variables, although it is just as easy to do this at the MATLAB
command line.
Consider the following example of how to copy variables. Select all the
SYSTEM matrices currently in your MATLAB workspace. To copy the wp
SYSTEM matrix to the variable TEMP, type wp in the editable text box to the
right of the Export button and TEMP in the editable text box to the right of As.
Your Workspace Manager window should correspond to Figure 6-11.
Pressing the Export button copies wp to the variable TEMP in your MATLAB
workspace. The text display in the message bar shows the MATLAB command
and the time and date it was executed. Your Workspace Manager window
should look like Figure 6-12 after the Refresh Variables button is pressed.
The drop box to the right of the Export button provides another manner to
deposit information to be copied into the workspace. For more information
about how to use µ-Tools drop boxes see the “Dragging and Dropping Icons”
section of this chapter.

6-11
6 Graphical User Interface Tools

Figure 6-11: wsgui Main Window with SYSTEMs Selected and Export Data

6-12
 Workspace User Interface Tool: wsgui

Figure 6-12: wsgui Main Window After Refreshing Exported Data

6-13
6 Graphical User Interface Tools

Spinning Satellite Example: Problem Statement

In this section we describe an example application that will be used in the
following two sections to demonstrate µ-analysis and synthesis methods using
µ-Tools graphical user interfaces.
Consider a satellite spinning about its z-axis with a control objective to
minimize the x and y-axes rotations due to x and y-axes disturbances. The
control inputs are torque actuators in the x and y-axis and the feedback
measurements are angular sensors. In this example the sensor measurements
are poorly aligned with the axis of rotation being measured. The poor
alignment of the sensor measurements is introduced on purpose to illustrate
the importance of direction in a multivariable control problem.
The following are the linear, state equations of the satellite.

d Θ Θx u1
------ x = 0 10 + 0 1
dt Θ – 10 0 Θ y 1 0 u2
y

y1
= 1 10 Θ x
y2 – 10 1 Θ y

Therefore, the nominal state-space model for the spinning satellite is defined as

0 10 1 0
AG BG – 10 0 0 1
G = =
CG DG 1 10 0 0
– 10 1 0 0

6-14
 Spinning Satellite Example: Problem Statement

The spinning satellite has the following characteristics that need to be included
in the control problem formulation:

• The model of the channel 1 actuator has uncertainty or error of 20% at low
frequency, below 1 rad/sec. This modeling error reaches 100% uncertainty at
20 rad/sec with very large potential errors in this actuator model above 200
rad/sec. We choose to model this in the µ framework as a multiplicative input
uncertainty. A frequency domain weight is constructed to describe the
percentage modeling error as a function of frequency. The multiplicative
uncertainty weight associated with the first actuator is

10 ( s + 4 )
W del1 := ------------------------
s + 200

A frequency response plot of this weight is shown in Figure 6-13.

W_del1 (solid) and W_del2 (dashed)

1
10
Magnitude

0
10

1
10 2 1 0 1 2 3 4
10 10 10 10 10 10 10
Frequency (rad/sec)

Figure 6-13: Uncertainty Weights: Wdel1 (solid), Wdel2 (dashed)

6-15
6 Graphical User Interface Tools

• The channel 2 actuator has 40% uncertainty at low frequency and the
uncertainty in this model reaches 100% at 58 rad/sec. The multiplicative
uncertainty weight associated with this actuator is

10 ( s + 24 )
W del2 := ---------------------------
3 ( s + 200 )

A frequency response plot of this weight is shown in Figure 6-13.

• The two sensor measurements are assumed to be noisy. A sensor noise

weight, Wn, of the form Wn = wnI2×2 is used to model the sensor noise. The
diagonal structure of the noise weight indicates that there is an equal
amount of noise in each measurement. The scalar transfer wn function is

12 ( s + 25 )
w n := ------------------------------
5 ( s + 6000 )

Based on this weight at low frequency, below 20 rad/sec, the noise signal has
a magnitude of ± 0.01radians, at high frequency the noise level reaches a
magnitude of ± 2.4 radians.
• The desired closed-loop performance is to achieve 100:1 disturbance rejection
at DC. This can also be interpreted as desiring 1% tracking error at DC. The
performance objective is the same in each channel. Hence it is represented
by a diagonal performance weighting function Wp, Wp = wpI2×2. The scalar
transfer function wp is defined as

s+4
w p = -----------------------------
2 ( s + 0.02 )

This weight also defines a desired closed-loop bandwidth of 2 rad/sec and a

limit on the peak of the output sensitivity function to a value of 2. A
frequency response of the performance weighting functions is shown in
Figure 6-14.

6-16
 Spinning Satellite Example: Problem Statement

W_p (solid) and W_n (dashed)

2
10

1
Magnitude 10

0
10

1
10

2
10 2 1 0 1 2 3 4
10 10 10 10 10 10 10
Frequency (rad/sec)

Figure 6-14: Performance Weights: Wp (solid), Wn (dashed)

The control design block diagram for the spinning satellite is shown in
Figure 6-15. The uncertainty weights, Wdel1 and Wdel2, and the performance
weights, Wp and Wn, are design parameters that you, the control engineer, can
manipulate. These weights are used to incorporate information about the
physical system into the control design.
Let P denote this open-loop interconnection. Suppose we order the inputs and
outputs in P, as shown in Figure 6-16 (each signal represents a vectored valued
signal with two components).

6-17
6 Graphical User Interface Tools

d1 d2
w1 z1
6 ? ? ?
Wdel2 Wp
u1 - ?g - -?g - e1
u2 -g - G ?
-g - e2
6
Wdel1 y1  ?
g  n1
6 y2  ?
g Wn
 n2
w?
2 z2
Figure 6-15: Spinning Satellite Interconnection Structure

w  z
e
 d
Pss  n
y  u

Figure 6-16: Spinning Satellite Control Design System

This is a robust, multivariable feedback control problem. The open-loop

interconnection in Figure 6-15 and Figure 6-16 have several features:
• Uncertainty in each Input Channel (z/w)
• Disturbance and Error at Output of Plant (d/e)
• Sensor Noise on Measurements (n)

The uncertainty block structure for the problem is

∆ := {diag[δ1,δ2] : δ1,δ2 ∈ C}

To create and load the P_ss interconnection structure shown in Figure 6-16,
type
ssic

at the command line. dkitgui, a graphical user interface for D – K iteration, is

used to design a robust controller that achieves all of these specifications.
simgui is used to analyze the time response of the resulting controllers.

6-18
 D-K Iteration User Interface Tool: dkitgui

D-K Iteration User Interface Tool: dkitgui

In this section we describe the graphical user interface for µ-synthesis via
D – K iteration, an approach to robust control design. The GUI-tool for D – K
iteration is created by the command dkitgui. There are five windows
associated with this tool:

• Main Iteration window, which is the main interface for the user during the
iteration.
• Setup window, where initial data is entered.
• Parameter window, which is occasionally used to modify properties of the D
– K iteration, such as H∞ parameters, and to select the variables that are
automatically exported to the workspace each iteration.
• Frequency Response window, where the plots of µ and σ of the closed-loop
transfer function matrix are displayed.
• Scaling window, where the rational fits of the frequency-dependent D-scale
data are shown, and can be modified.

The spinning satellite control design example from the previous section is used
to illustrate the basic features of dkitgui. Start the tool by typing
dkitgui

Depending on the computer, dkitgui takes up to a minute to start. The main

interactive window finally displays the message
Press SETUP to begin

in its lower left corner, as shown in Figure 6-17. This is the location of the
message bar where information about the D – K iteration is displayed.

6-19
6 Graphical User Interface Tools

Figure 6-17: Initial dkitgui Main Window

Press the SETUP button as instructed. This brings the Setup window to the
foreground. The Setup window is shown in Figure 6-18.

6-20
 D-K Iteration User Interface Tool: dkitgui

Figure 6-18: Initial dkitgui Setup Window

To start an iteration, you must enter the following items:

• Open-loop interconnection structure

• Number of uncertainty blocks
• Sizes of the uncertainty blocks
• Sizes of the performance transfer functions (errors and disturbances)

6-21
6 Graphical User Interface Tools

• Sizes of the feedback signals (measurements and controls)

• Frequency response range

You can enter this information in any order, except that the dimensions of the
uncertainty blocks can only be specified after entering the number of
uncertainty blocks. All data entry points are uicontrol editable text objects,
and operate in a machine dependent manner with which you should be
familiar. As before, when we instruct you to enter data in an editable text
object, this implicitly means enter the text, and complete the action (by
pressing Return, by pressing the mouse on another object, by moving the
mouse pointer out of the text object, etc.). See the MATLAB Function Reference
online for more details on completing a text entry.
Notice that three items in the Setup window — Controller, Iteration
Suffix, and Iteration Name — are enclosed in < >. This notation denotes a
variable that is optional, and no action is necessary.
Enter P_ss (the open-loop interconnection) in the open-loop editable text frame
(see Figure 6-19). Press the Open-Loop IC checkbox to load the data into the
Open-Loop IC variable. The pointer turns into an hourglass while MATLAB
loads the data. Upon loading the data, the pointer turns back into an arrow,
and the matrix type (S for System, C for Constant, and V for Varying) of the
Open-Loop IC variable and its dimensions are displayed to the right of the
editable text. In this example the variable P_ss is a SYSTEM matrix with six
outputs, eight inputs, and eight states. The <Controller> data is optional.
This allows you to load an initial controller to start the D – K iteration or during
the D – K iteration you may want to load a reduced order controller.

Figure 6-19: Open-Loop/Controller Input Frame

Since there are two uncertain actuators in the spinning satellite problem, the
uncertainty structure has 2, 1-by-1 uncertainty blocks. Enter a 2 in the
Uncertainty Structure # of Blocks text field. This opens a 2-by- 3 editable
matrix, as seen in Figure 6-20, where the dimensions of the uncertainty blocks
are entered. Enter a 1 for each row and column dimension. The third column,
labeled Fac, is used for scaling the size of the uncertainty during the iteration.

6-22
 D-K Iteration User Interface Tool: dkitgui

The value of Fac may be varied from.1 to 10, effectively reducing or increasing
the size of the uncertainty by a factor of 10. Leave these factors as 1 for the time
being.

Figure 6-20: Uncertainty Structure Frame

In the spinning satellite problem (Figure 6-16), there are four exogenous
disturbances (two disturbance torques and two measurement noises) and two
penalized errors (two tracking errors). Enter a 2 in the # of Errors editable text
and a 4 into the # of Disturbances editable text in the Performance
Structure frame. The spinning satellite has two measured variables for
feedback and two control actuators. Enter these in the # of Measurements and
# of Controls editable text, respectively, in the Feedback Structure frame.

Drop Box Data

In the right column of the Setup window are push buttons with the variables
names Uncertainty, Performance, Feedback, and Omega, as seen in
Figure 6-21.

Figure 6-21: Uncertainty, Performance, Feedback, and Omega Frame

Entering data with these push buttons is optional. These push buttons are
used to enter the same data as SIGNAL DIMENSIONS and Frequency Range

6-23
6 Graphical User Interface Tools

data. To the left of each push button is a drop box and to the right of each push
button is an editable text box. Data that is dropped into the drop boxes or
entered in the editable text overwrites the data in the left column of the Setup
window. (See the “Dragging and Dropping Icons” section at the end of this
chapter for more details.) For example, instead of entering a 2 into the # of
Blocks text and 1, 1, 1, and 1 into each block structure, you could type
ones(2,2) in the Uncertainty editable text and press the Uncertainty push
button. This action overloads this data into the Uncertainty Structure by
redrawing the Uncertainty Structure frame and filling out the corresponding
editable text locations. Similarly the Performance data overwrites the
Performance Structure data, Feedback overwrites the Feedback Structure
data, and Omega overwrites the Frequency Range data.

Completing Problem Setup

To complete the problem setup, enter the frequency response information
appropriate for the problem in the Frequency Range frame. Keep the
Logspace option, and enter 0.001 for the Low frequency, 100 for the High
frequency, and 60 in the # Points editable text. Selecting the Custom option
allows you to enter any valid MATLAB expression in the editable text frame.
This completes the input of the data required by dkitgui to start a D – K
iteration. Upon correctly entering the required data, the message
Mu-Synthesis Problem Specification Complete...

appears in the message bar of the Main window.

Enter text in the optional fields Iteration Suffix and Iteration Name.
Iteration Suffix is used for data that is automatically exported to the
workspace. This string is appended to default names for the variables as they
are exported to the workspace. Enter ss in the editable text suffix box. We will
see the precise effect of this later in the example. The Iteration Name modifies
the title of each figure. Enter Spinning Satellite in the editable text frame,
and notice its effect. The Setup window should appear as in Figure 6-22.

6-24
 D-K Iteration User Interface Tool: dkitgui

Figure 6-22: dkitgui Setup Window After Entering Data

To continue with the D – K iteration, first hide the Setup window. This is done
by pulling down the Window menu and selecting Hide Setup from the menu
and returning to the main window. Now select the Parameters window from
the Window menu.

6-25
6 Graphical User Interface Tools

D-K Iteration Parameters Window

The fifth window titled Parameters, shown in Figure 6-23, contains additional
data you may want to enter for the D – K iteration. This window is shown if
Parameters is selected from the Windows menu. All of the parameters in this
window are set to default values. No data entry is required and you can return
to the main Iteration window to initiate the first iteration. The Parameters
window contains five settings you can change. They are:
• HinfSyn Parameters
• Structured Singular Value (Mu)
• Riccati Solver
• D-Scale Prefit
• Each Iteration Export

Figure 6-23: D – K Iteration Parameter Window

6-26
 D-K Iteration User Interface Tool: dkitgui

The HinfSyn Parameters and Riccati Solver settings, shown in the table
below, correspond to the inputs of the H∞ control design program hinfsyn.
Gamma Min and Gamma Max are the minimum and maximum γ values.

The Suboptimal Tol is how close to the optimal γ value is desired. The
Imaginary Tol and Positive Def Tol are epr and epp in the hinfsyn program.
They correspond to the measure of when the real part of an eigenvalue of the
Hamiltonian matrices is zero and determination of the positive definiteness of
the X∞ and Y∞ solutions. The current default value for each parameter is shown
in parentheses, ( ), to the right of each label. The Riccati Solver has a mutually
exclusive set of buttons for selecting either the Schur or Eigenvalue method to
be used to solve the H∞ Riccati equations.

Hinfsyn Parameters Riccati Solver

Gamma Min Schur Method

Gamma Max Eigenvalue Method

Suboptimal Tol

Imaginary Tol

Pos Def Tol

The HinfSyn Parameters frame also allows you to deselect the measurements
and controls used during the control design process. For the spinning satellite
example there are two controls and two measurements. The Measurements
Utilized frame indicates that all measurements are currently being used. You
can input to the Measurements Utilized editable text a standard MATLAB
vector to denote the measurements that are to be used. Similarly, you can
select in the Controls Utilized frame which control inputs are to be used. The
resulting control design will have zeros in the state space B or C matrix of the
controller corresponding to the measurements inputs or control outputs that
have been deselected.

The Structured Singular Value (Mu) settings frame has two sets of options.
These options correspond to calculation of the structured singular value (µ)
using the mu program during D – K iteration. The first set of buttons is
mutually exclusive. You can either select to use greatest accuracy or a less

6-27
6 Graphical User Interface Tools

accurate but faster technique for calculating µ. (The Optimal method calls the
mu program with the option ’c,’ Fast method calls mu with the ’f’ option). You
can also select to calculate only an upper bound for µ which calls mu with the ’U’
option. Selecting this option speeds up the µ calculation, but you will be unable
to see how different the upper and lower µ bounds are when they are calculated
and plotted in the Mu/SVD Plot window.
The D-Scale Prefit frame contains settings for calculation of the rational
D-scales. The Max Auto-Order defines the maximum D-scale state order to be
used to fit an individual D-scaling during the prefitting part of the D-scales
fitting routine. The Max Auto-Order default is five states. The Auto-Fit
Tolerance scroll bar allows you to define how close the rational scaled µ upper
bound is to approximate the actual µ upper bound in a norm sense. Selecting
Loose will result in lower order D-scales being used in the D-scale prefitting,
whereas selecting Tight will likely result in higher order D-scales during the
D-scale prefit computation. This setting can play an important role in
determining which minimum of the D – K iteration is achieved. Currently this
is done by trial and error.
The Each Iteration: Export... frame shows data in the form of radio buttons
that is available to be exported to the MATLAB workspace. In the following
list, a subscript i denotes that the integer iteration number is added to the
variable’s name. These variables are:

• Controller, which is exported as Ki.

• Mu Analysis exports mubndi, ddatai, dsensi, and perti. The mubndi variable
is selected based on the Structured Singular Value (Mu) frame. This is
normally the upper and lower bounds for µ. The ddatai is the D-scale data,
dsensi is the sensitivity, and perti is the worst-case perturbation as a
function of frequency. These are the standard outputs of the mu function.
• Closed-Loop Freq Response exports the closed-loop system, clpi, and its
frequency response clpig.
• Rational D-scalings exports the left, dli, and right, dri, rational D-scalings.
• Open-Loop Interconnection is exported as olici.

The Iteration Suffix string, which was input in the Setup window, is appended
to the end of all the output variables selected in the Each Iteration: Export...
frame. In this example, the Controller and the Mu Analysis data is selected
for output. Therefore after the Control Design button executes for the first
time, the variable K1ss will be in the workspace.

6-28
 D-K Iteration User Interface Tool: dkitgui

Hide the Parameter window and return to the main D – K iteration window to
continue this example. This can be done by simply pulling down the Window
menu and first selecting the Iteration option. This moves the main Iteration
window into the foreground. The main window is now shown in Figure 6-24. Go
back to the Parameter window, pull-down the Window menu, and select Hide
Parameter to hide the Parameter window.

Figure 6-24: Main D – K Iteration Window After All the Data Is Specified

6-29
6 Graphical User Interface Tools

D-K Iteration
The main window has (at this point) five significant items:
• Five push buttons, whose actions constitute the D – K iteration. Recall, the
D – K iteration pertains to the picture shown below, and is

   D  ,1 
 PD  =   P D 
 
- K - K

- Hold D fixed, design K to minimize H∞ norm.

- Then, hold K fixed, and find the new optimal D scalings.
- Go to Step 1.
• A read-only DK Iteration Summary table.
• A menu bar with File, Iteration, Options and Window menus.
• A scrollable list of linkable variables that can be dragged from this tool into
other tools’ drop boxes.
• A message bar.

The scrollable list of linkable variables includes the D – K iteration variables.

Linkable Variables Meaning

Khinf H∞ controller
Kuse Controller K used in the current
D – K iteration.
Blk Block structure for µ calculation
Ydim Output dimension of the controller
Udim Input dimension of the controller
IC Open-loop interconnection structure

6-30
 D-K Iteration User Interface Tool: dkitgui

Linkable Variables Meaning

Clpg Closed-loop frequency response

Clpg Closed-loop system

These variables can be dragged and dropped at anytime into other µ-Tools user
interface commands.
The Control Design push button is enabled, and is the first step of the
iteration. At this point, the D matrices are set to identity (of appropriate
dimension). Design the first H∞ controller by pressing the Control Design
button. The standard gamma iteration data from hinfsyn is displayed in the
MATLAB command window. The DK Iteration Summary table is updated at
the end of the control design, and the Form Closed-Loop button is enabled.
The next two steps are simple — form the closed-loop system, and calculate
closed-loop frequency response. Press the Form Closed-Loop and Frequency
Response buttons as they are enabled. As the frequency response of the
closed-loop system is calculated, a running tab of the number of frequency
points calculated is shown in the message frame of the main window. The norm
of the closed-loop transfer function as a function of frequency is plotted in the
Frequency Response window as seen in Figure 6-25. (Note that Figure 6-25
includes the µ plot, which is not accurate at this point in the D – K iteration.)
In general, this transfer function has the D-scalings from the previous iteration
(for the first iteration it’s just identity) and the controller which was just
designed. Upon completing the frequency response the Compute Mu button is
enabled.
Now, compute the structured singular value, µ, of the closed-loop frequency
response by pressing the Compute Mu button. The block structure was defined
in the Setup window, in the Block Structure field. Recall (see Chapter 4) that
the upper bound for µ is computed by determining the optimal D-scalings as a
function of frequency. Like the frequency response, a running tab of the µ
calculation is shown in the message frame of the main window. The results are
shown in Figure 6-25.

6-31
6 Graphical User Interface Tools

Figure 6-25: µ Upper Bound and Maximum Singular Value for First D – K
Iteration

You have now completed one D – K iteration. The DK Iteration Summary table
is completely updated for the first iteration, as shown in Figure 6-26. The Next
Iteration button is highlighted. Once the Next Iteration button is pushed you
cannot effectively return to the previous iteration.
Suppose you desire a more refined frequency response for the µ calculation.
You can do this before you move to the second iteration by changing the data
in the Frequency Range frame in the Setup window. Changing the number of
frequency response points from 60 to 80 results in the <Frequency Response>
button being enabled. Note that the sideways carrots < > around the button
name denote that you may select the frequency response or go to the next
iteration. Continuing to the next iteration without pressing the Frequency
button will not change any of the data calculated during the first iteration.

6-32
 D-K Iteration User Interface Tool: dkitgui

Figure 6-26: D – K Iteration Summary After First Iteration

Pressing the Next Iteration button results in the D-scale data output from the
µ calculation being fit with rational D-scales. In the spinning satellite example,
there are two uncertainty blocks and one performance block, therefore, there
are two D-scales to be fit. A table entitled D Scaling Order appears in the main
D – K iteration window, as shown in Figure 6-27. The table contains the scaling
number and the order of each scaling. Each D-scale data is prefit with up to a
maximum state order transfer function in an attempt to minimize the
difference between the scaled µ upper bound and the µ upper bound with the
rational D-scales. The maximum order of the prefit D-scales is specified in the
D-Scale Prefit field, with the Max Auto-Order data in the Parameters
window.

Figure 6-27: D-Scalings Order Table

The D-scaling information for the first D-scaling is shown in graphical form in
the Scaling window, Figure 6-28. Note that the first D-scaling was fit with a
first order transfer function (see Figure 6-27). There are three plots shown in
this window. The top plot shows the µ upper bound, which contains the
D-scaling data, and a plot of the scaled upper bound. The scaled upper bound
is calculated with the rational D-scales wrapped in to the original closed-loop
frequency response. The middle plot shows the D-scale magnitude data and the
rational fit for the first D-scale. The D-scale magnitude data is the variable
being fit. The bottom plot shows the sensitivity of the µ upper bound to changes
in the D-scale. The larger the sensitivity, the more important it is to fit the
D-scale well in that frequency range.

6-33
6 Graphical User Interface Tools

Figure 6-28: D-Scalings for Second Iteration

You can change which D-scale data is shown in the Scaling window by pressing
the ’- -’ or ’++’ buttons to the left of the Scaling title. This will cycle through
each of the D-scalings. The ’- -’ or ’++’ buttons to the left of the Order title

6-34
 D-K Iteration User Interface Tool: dkitgui

decrement or increment the order of the D-scale fit by 1. You can also change
the order of the D-scale fit by editing the D-scale order directly. Changing the
D fit order will affect the middle plot, which shows the magnitude data (solid
line) and the rational fit (dashed line) and will also affect the current scaled
upper bound (dashed line) shown in the top figure of the Scaling window. Note
that the goal of D – K iteration is to reduce the µ upper bound. It is usually
important that the current scaled upper bound, which incorporates the rational
D-scalings, closely matches the calculated µ upper bound. This is especially
true in the frequency range where µ is large.
The D-scale data for this example are both fit with first order transfer
functions. These fits appear to be sufficient, therefore we will go on to the
second control design. Pressing the Control Design button wraps the rational
D-scalings into the original interconnection structure, P_ss, and designs a new
H∞ controller. For this second D – K iteration, a γ value of 7.89 is achieved.
We now desire to run the next three buttons in sequence. This can be done by
pulling down the Iteration menubar in the main Iteration window. Selecting
Auto Steps, dragging the mouse to the right of Auto Steps allows you to choose
the menu item Next 3 Steps, as shown in Figure 6-29. This automatically runs
the next three steps of the D – K iteration. Selecting either the Auto Steps or
Auto Iterate will result in the appearance of a Stop button in the main
Iteration window below the DK Iteration Summary table. Pressing the Stop
button terminates the automated D – K iteration after the current button
running has completed.

Figure 6-29: Main Window Iteration Pull-Down Menu

Under the Iteration menu, the Restart option allows you to restart a D – K
iteration at the very beginning while leaving the Setup window data intact.
This is often useful if the incorrect weights were selected and you would like to
reload the system interconnection structure and start over.

6-35
6 Graphical User Interface Tools

Note that after two complete D – K iterations we have achieved a µ value of 2.11
(Figure 6-30). The objective is to achieve a µ value less than 1. Therefore,
several more D – K iterations may be required. User interaction can be
eliminated from the D – K iteration by selecting the Auto Iterate menu item
from the DK Iteration window, Iteration menubar, as shown in Figure 6-31.
Dragging the mouse to the right of the Auto Iterate menu item allows you to
select up from one to eight automated D – K iterations. During each iteration,
the rational D-scale order is selected automatically by the dkitgui program.
Again, the Stop button allows you to terminate the Auto Iterate option at any
time. After three complete D – K iterations, a µ value of 1.03 is achieved.
Selecting one more automated D – K iteration results in a µ value of 0.91 after
four iterations (Figure 6-32).

Figure 6-30: D – K Iteration Summary After Two Iterations

Figure 6-31: Main Window Iteration Pull-Down Menu

6-36
 D-K Iteration User Interface Tool: dkitgui

Figure 6-32: D – K Iteration Summary After Four Iterations

Based on the setting of the Each Iteration Export radio buttons in the
Parameter window, the controller designed each iteration has been exported
to the MATLAB workspace. Therefore, after four D – K iterations controllers
K1ss, K2ss, K3ss, and K4ss are present in your workspace. We also have data
from the µ analysis, ddataiss, desnsiss, mubndiss, and pertiss in the
workspace. If you have your Workspace Manager open, press the Refresh
Variables button. The results are shown in Figure 6-33.

6-37
6 Graphical User Interface Tools

Figure 6-33: Data in the MATLAB Workspace After Four D – K Iterations

We have successfully designed a controller to achieve our objectives since µ is

less than 1. To quit the D – K iteration graphical user interface, pull down the
File menu from the main Iteration window and highlight the Quit option.

6-38
 D-K Iteration User Interface Tool: dkitgui

Options Menu
The Options menu in the main dkitgui window allows you to perform two
specialized operations. Selecting the Auto_Refresh K item from the Options
menu will refresh the <Controller> editable text in the Setup window after
successful completion of the Control Design button. Any valid MATLAB
expression can be typed in the <Controller> editable text space. This allows
you to perhaps have an automated controller reduction scheme that would get
executed after the design of the H∞ controller.

Figure 6-34: D – K Options Menu

Selecting the Auto_Refresh Olic item from the Options menu will refresh the
Open-Loop IC editable text in the Setup window after successful completion
of the fitting the D-scale data. Any valid expression can be typed in the
Open-Loop IC editable text space. This allows you to have an automated
program that modifies the open-loop interconnection weightings based on the
value of µ from the previous iteration.

6-39
6 Graphical User Interface Tools

LFT Time Simulation User Interface Tool: simgui

The GUI tool for time simulations is created by the command simgui. simgui
has two interface windows and up to six plot windows:

• Main Simulation Tool window, which is the main interface for the
simulation.
• Parameter window, which is used to modify properties of the time
simulation, such as the final time, integration step size, initial conditions,
and variables automatically exported to the workspace.
• Plot windows, where the plots of time responses are displayed. You can open
up to six of these windows.

Spinning Satellite Example: Time Simulation

Several controllers for the spinning satellite example have been designed in the
previous dkitgui section. We will modify the spinning satellite open-loop
interconnection shown in Figure 6-15, to include additional outputs for
simulation purposes. Let Ptss denote this new open-loop interconnection, with
the inputs/outputs in Ptss, as shown in Figure 6-35 (each signal shown below
represents a vectored valued signal with two components).

w
e  z
y  d
d 
Ptss
n
u  u
y

Figure 6-35: Time Simulation Interconnection System: Ptss

6-40
 LFT Time Simulation User Interface Tool: simgui

Two controllers to be analyzed are K1ss and K4ss. K1ss is the controller design
from the first D – K iteration and K4ss was designed in the fourth D – K
iteration. They are either present in your current workspace or they can be
loaded by typing
load ss_cont

at the command line. The function simgui will be used to analyze the time
response of these controllers. Typing the command tssic creates and loads the
Ptss interconnection structure.
tssic

Setting Up simgui
Start the linear fractional transformation (LFT) time simulation tool by typing
simgui

Depending on the computer, this takes up to a minute to start. The message

Setting up simulation GUI....

appears in the lower left corner of the main window. This is the location of the
message bar where information about the time simulation is displayed. Upon
completion of the simgui setup, the main window displays the message
Done with setup

in the message bar, as shown in Figure 6-36.

6-41
6 Graphical User Interface Tools

Figure 6-36: Initial simgui Main Window

The simgui tool simulates linear fractional models and plots their responses.
simgui is based on the standard linear fractional model shown in Figure 6-37.

-

w z

 P 
e  d
y u
- K
Figure 6-37: Standard Linear Fractional Model

6-42
 LFT Time Simulation User Interface Tool: simgui

The P block corresponds to the open-loop plant interconnection model, referred

to as the Plant in simgui. The K block corresponds to a state-space controller,
Controller in simgui. The ∆ block corresponds to the perturbation to the
model, Perturbation in simgui. The Input Signal to the time simulation is
denoted by d. The individual systems are formed using the starp command.
The time simulation outputs available for plotting correspond to the variable e.
Three types of simulation are possible: continuous-time (default) using trsp,
discrete-time using dtrsp and sample-data using sdtrsp. In this example, we
discuss the continuous-time simulation. This requires the open-loop
interconnection, Plant, Controller, and Perturbation to be continuous-time
systems, or constants.
To start the time simulation, you must at least enter the following items:

• Open-loop plant interconnection model, Plant

• Input excitation time signal, Input Signal

If this is the only data entered, the input dimension of the Plant matrix must
match the row dimension of the Input Signal VARYING or CONSTANT
matrix. In addition you can enter a:
• Controller
• Perturbation

Note The dimensions of the respective systems and inputs corresponding to

those in Figure 6-37 must match one another. If they do not match one
another, simgui will display an error message in the message bar.

The Plant, Controller, Perturbation, and Input Signal variables can be

entered in any order. For this example enter P_tss (the open-loop
interconnection) in the Plant text entry, (see Figure 6-36), and press the Plant
push button to load the plant data. The MATLAB pointer turns into an
hourglass while MATLAB loads the data. After the data is loaded, the
MATLAB pointer turns back into an arrow, and the Plant data matrix type (S
for System, C for Constant, and V for Varying) and its dimensions are
displayed to the right of the Plant text entry box. Similarly, enter K1ss in the
Controller text entry and press the Controller push button. K1ss was one of
the controllers loaded with the load ss_cont command. Enter [0.1 0;0

6-43
6 Graphical User Interface Tools

-0.1]) in the Perturbation text entry and press the Perturbation push
button.
Plant, Perturbation, and Controller correspond directly to P, ∆, and K in
Figure 6-37. Based on the dimensions of these matrices, the Input Signal must
have four rows of signals for the system to have the correct dimensions. An
error message will be displayed if the dimensions of the interconnected system
are incorrect. The resulting closed-loop, perturbed system has a total of eight
outputs. They are e, y, d, and u, as shown in Figure 6-35.
Let’s create an input signal for the spinning satellite example. The first two
inputs are disturbances, and the third and fourth inputs are sensor noise. For
this example, input a unit step command into channel 1, zero input to channel
2, and a normally distributed random noise signal into channels 3 and 4 which
is scaled by 0.05. Typing the following commands at the MATLAB prompt will
generate this input signal.
u1 = step_tr(0,1,.01,5);
u2 = mscl(u1,0);
t = getiv(u1);
u34 = siggen('0.05*randn(2,1)',t);
u = abv(u1,u2,u34);

Enter u in the Input Signal text entry and press its push button. The
successful entry of the input signal results in the appearance of the Plots and
Line Types scroll table in the main simulation window, as shown in
Figure 6-38.

6-44
 LFT Time Simulation User Interface Tool: simgui

Figure 6-38: Main Simulation Window with Data

6-45
6 Graphical User Interface Tools

Plots and Line Types Scroll Table

The Plots and Line Types scroll table allows you to select the time responses
and outputs desired. The four checkboxes on the left side of the scroll table
correspond to the four time simulations that can be performed:

• Open-Loop Nominal: response of the open-loop plant, Plant, without a

controller or perturbation included.
• Open-Loop Perturbed: response of the open-loop plant, Plant, with the
perturbation, Perturbation, included but no controller.
• Closed-Loop Nominal: response of the closed-loop plant,
starp(Plant,Controller), without the perturbation included.
• Closed-Loop Perturbed: response of the closed-loop plant,
starp(Plant,Controller),with the perturbation, Perturbation, included.
Currently each checkbox name is enclosed by ( ). This indicates that the time
response data for these four is either not current or does not exist. Also each
button is not highlighted and no line type data appears in the scroll table. This
indicates that currently no time response data is desired. Pressing the
Open-Loop Nominal and Closed-Loop Nominal checkboxes enables the
Compute button and Ready to Simulate appears in the message bar. This
action also displays the checkboxes associated with the Open-Loop and
Closed-Loop Nominal responses.
The scroll table contains information corresponding to the eight outputs. The
scroll bar at the bottom of the table (see Figure 6-39) allows you to scroll the
table to select or deselect output channels. The scroll table contains the four
response types and output channels and line type information corresponding to
these outputs. Each response and output channel has a corresponding
checkbox. The Plots and Line Types scroll table also has a pull-down button
to denote the Grouped or Free Form operation of the checkboxes (see
Figure 6-40). The default is Grouped, which means that only one output
channel, and all four types of responses, are selected when an output checkbox
is pressed. Initially the first output checkboxes of the selected responses are
enabled. Therefore, currently in this example, pressing the Compute button
would result in the first output of the Open-Loop and Closed-Loop Nominal
responses being plotted in the Plot Page #1.

6-46
 LFT Time Simulation User Interface Tool: simgui

Figure 6-39: simgui Plots and Line Types Scroll Table

Figure 6-40: Grouped/Free Form Button

Leaving the Grouped pull-down menu selected, as seen in Figure 6-40, and
pressing the third output of the Closed-Loop Nominal response results in the
third output of all four responses being selected. Therefore, currently in this
example, this would result in the third output of the Open-Loop and
Closed-Loop Nominal responses being plotted in the Plot Page #1. The Free
Form option decouples the responses. Pressing on any output channel
checkbox only selects or deselects that checkbox.
The text in the checkbox corresponds to a MATLAB color type, next to a
MATLAB line type. For example, the default line types and colors for a color
monitor are: Open-Loop Nominal outputs are yellow (y) and solid lines (–).
Closed-Loop Nominal outputs are magenta (m) and dotted lines (:).
The LineStyle menu provides a way of modifying the line color and type
(Figure 6-41). Selecting edit from the LineStyle menu changes the checkboxes
into uicontrol editable text objects, and operates in a machine dependent
manner with which you should be familiar. As before, when we instruct you to
enter data in an editable text object, this implicitly means enter the text, and
complete the action (by pressing Return, by pressing the mouse on another
object, by moving the mouse out of the text object, etc.). See the online
MATLAB Function Reference for more details on completing a text entry.

6-47
6 Graphical User Interface Tools

Figure 6-41: LineStyle Pull-Down Menu

You can modify the first output of the Closed-Loop Nominal channels to be
white (w) and dashed (--), as shown in Figure 6-42.

Figure 6-42: Modified Plots and Line Types Scroll Table

Press the Done Edit button in the top right corner of the Plots and Line Types
table to return the table to checkboxes for output selection. The Default
selection of the LineStyle menu provides four different line type defaults:

• Color
1 Open-Loop Nominal outputs: yellow, solid (y-)
2 Open-Loop Perturbed outputs: red, dashed (r--)
3 Closed-Loop Nominal outputs: magenta, dotted (m:)
4 Closed-Loop Perturbed outputs: white, dashed-dotted (w-.)

• B/W (Black and White)

1 Open-Loop Nominal outputs: white, solid (w-)
2 Open-Loop Perturbed outputs: white, dashed (w--)
3 Closed-Loop Nominal outputs: white, dotted (w:)
4 Closed-Loop Perturbed outputs: white, dashed-dotted (w-.)

6-48
 LFT Time Simulation User Interface Tool: simgui

• Color Symbols
1 Open-Loop Nominal outputs: yellow, x (yx)
2 Open-Loop Perturbed outputs: red, star (r*)
3 Closed-Loop Nominal outputs: magenta, plus (m+)
4 Closed-Loop Perturbed outputs: white, circle (wo)

• B/W (Black and White) Symbols

1 Open-Loop Nominal outputs: white, x (wx)
2 Open-Loop Perturbed outputs: white, star (w*)
3 Closed-Loop Nominal outputs: white, plus (w+)
4 Closed-Loop Perturbed outputs: white, circle (wo)

Plotting Window Setup and Titles

The output data selected in the scroll table is displayed in a single plot
determined by the information in the Current Plotting Figure Information
frame, as shown in Figure 6-44. You can format up to six pages (MATLAB plot
figures) to display output data. The simgui tool initializes to have one plot page
(figure) and one set of plot axes on that page, as shown in Figure 6-43. Setup of
the plotting pages and titles are controlled via the Current Plotting Figure
Information frames in Figure 6-44.

6-49
6 Graphical User Interface Tools

Figure 6-43: Simulation Plot Page #1

Figure 6-44: Plotting Data Frames and Titles

The Plot Figure frame, Plot Fig#, corresponds to the plotting window number
displayed. There can be a maximum of six plotting windows. Pressing the
increment ++ and decrement -- buttons changes the page number and brings
the new plot window to the foreground. If the page number is incremented to a
page that previously didn’t exist, a new page is created with one plot axes. You

6-50
 LFT Time Simulation User Interface Tool: simgui

can also change the Plot Fig# editable text to reflect the plot page desired. The
input to this editable text must be a positive integer between 1 and 6.
The row number, Row# frame, has two editable texts and a decrement -- and
increment ++ button as seen in Figure 6-44. The far right editable text, after
of, corresponds to the number of subplot axes rows desired for the current page
(Plot Fig#). The first editable text, after Row#, indicates the row number of the
corresponding subplot on the given Plot Fig#. The minimum Row# editable
text value is 1, and its maximum values correspond to the value shown in the
second editable text of Row#. The increment and decrement buttons increase
or decrease the row number by 1. This is shown in the first Row# editable text.
The minimum row number is 1 and the maximum is the value of the second
Row# editable text.
Similarly, the column number, Col# frame, has two editable texts and a
decrement -- and increment ++ button (see Figure 6-46). The second editable
text, after of, corresponds to the number of subplot axes columns desired for
the current page (Plot Fig#). The first editable text, after Col#, indicates the
column number of the corresponding subplot on the given Plot Fig#. The
increment and decrement buttons increase or decrease by 1 the column number
shown in the first Col# editable text. The minimum column number is 1 and
the maximum is the value of the second Col# editable text.
For example, selecting the second editable text of Row# to be 2 and Col# to be
3 would result in six subplots, two rows of three columns. Changing either the
second editable text of Row# or Col# will display the result in a push button
label Apply, appearing in the right of these two frames, as shown in
Figure 6-44. The new desired subplot description is applied after this button is
pushed. As before, the inputs to the editable text frames must be positive,
nonzero integers.
The Plot Fig# and the second editable text frames of the Row#/Col# frames
indicate the layout of the current simulation plot page. The Plot Fig#, and first
editable text of the Row#, and Col# frames indicate which plot data is
currently in the Plots and LineTypes scroll table and their corresponding plot
labels and fonts. The first editable text in the Row# and Col# frames indicate
the row and column number of the corresponding subplot in the Plot Fig#. The
minimum first editable text values of Row#, and Col# are 1, and their
maximum values correspond to the values shown in the second editable text
frames in Row# and Col# respectively. In this example, let’s include two plots
on the first page, Plot Fig# 1, one above the other. Enter a 2 in the # of rows
editable text frame and press the Apply button, as seen in Figure 6-45.

6-51
6 Graphical User Interface Tools

Figure 6-45: Current Plotting Figure Information with Apply Button Enabled

Increment Row# to 2 to indicate the bottom plot.

The remaining seven editable text frames and pull down button correspond to
labeling and marking the subplot denoted by the data in the Plot Fig#, and the
first editable text in the Row#, and Col# frames. The editable text string
associated with the Title frame is the same as a string entered using the
MATLAB title command for the given subplot. The font size of the title can
be entered to the right of the Title frame editable text. The default font size for
all the labels is 10. Input to the font editable text must be a positive integer.
Similarly the Xlabel and Ylabel editable text strings correspond to the
MATLAB xlabel and ylabel commands. The axes font editable text allows
you to change the font of the subplot axes fonts. This data is restricted to be a
positive, nonzero integer. The button labeled Grid Off toggles between Grid
Off and Grid On for the current subplot.
For the top subplot in the current example, enter Plant Error Output 1 in the
Title editable text, an x-axis label of Time and a y-axis label of Degrees. Select
the font size to be 13 for all the text and the axes, grid off. Select output 1 of the
Open-Loop Nominal and Closed-Loop Nominal responses in the Plots and
LineType scroll table to be plotted.
Selecting an output from the Plots and LineType scroll table results in the
enabling of the Compute push button in the main simgui simulation window
and the appearance of the string Ready to Simulate in the main window
message bar. After entering this data, the main simulation window should look
like Figure 6-46.

6-52
 LFT Time Simulation User Interface Tool: simgui

Figure 6-46: Main Simulation Window with Subplot(1,1,1) Data Entered

Increment the Row# to 2, to indicate the bottom plot (subplot 212 of Plot Page
#1). Enter the title of Plant Error Output 2, an x-axis label of Time and a
y-axis label of Degrees. Select the font size to be 11 for the text and the axes,
grid off, and select output 2 Open-Loop Nominal and Closed-Loop Nominal
responses and the second output of these two responses to be plotted in the
Plots and LineType scroll table. Your first simulation plot page should look
the same as Figure 6-47.

6-53
6 Graphical User Interface Tools

Figure 6-47: First Simulation Plot Page

Increment the Plot Fig# to 2. This creates a new page with one plot axes and
brings it to the foreground. The Window pull-down menu in any simgui
window shows all the current simgui windows and allows you to hide the
current window or bring any of the other windows to the foreground. For Plot
Fig# 2, enter the title of Closed-loop plant outputs and controls, an x-axis
label of Time and a y-axis label of Degrees - Newtons. Select the font size to be
9 for the title, and 10 for all the text and the axes, grid off, and select the
Closed-Loop Nominal responses. Change the Grouped button in the Plots
and LineType scroll table to Free Form and select the y and u and outputs 3,
4, 5 and 6 be plotted. Change the line types of the output 3 to be white, solid
(w-), output 4 to be white, dashed (w--), output 5 to be white, dashed-dotted
(w-.), and output to be white, dotted (w:). You have now finished setting up the
plotting data and labeling the plots. It’s now time to calculate the
continuous-time simulation.

6-54
 LFT Time Simulation User Interface Tool: simgui

Pull down the Window menu from any simgui window and select the
Parameters window. Select an integration time of 0.01 second by entering
0.01 into the Integration Step Size editable text, as seen in Figure 6-48.

Figure 6-48: Simulation Parameter Window

Pull down the Window menu in the Parameter window and hide the
Parameter window. Press the Compute button in the main window to initiate
the simulations. Immediately the message Computing appears in the message
bar and the label on the Compute button changes to Stop. By pressing the
Stop button, the time simulation will terminate at the next available execution
of a break. After the Computing message appears in the message bar, and
assuming that the Stop button was not pressed, one of two pieces of
information will appear in the message bar. If each simulation is estimated to
take less than 3 minutes to compute then a running tab of simulation as it
progresses is shown. If a simulation is estimated to take more than 3 minutes
to compute the message, then
Simulation will take approx. X seconds:
check Integration Step Size

6-55
6 Graphical User Interface Tools

appears in the message bar. The variable X is an estimate of the number of

seconds it will take to calculate the given time simulation. After 1 second, this
changes into a running tab of simulation as it progresses.
Each time simulation in this example should take less than 3 minutes each to
calculate unless the default Integration Step Size was selected. The results of
these simulations are shown in Figure 6-49. Note that the time simulations are
calculated for all of the system outputs. Therefore other outputs, not initially
selected, can be displayed by selecting their respective checkbox without
having to resimulate the system. Similarly outputs can be deselected without
deleting them. You should try to select other outputs, change the color and
linetype of the plotted lines, and simulate the open-loop and closed-loop
perturbed systems.

6-56
 LFT Time Simulation User Interface Tool: simgui

Figure 6-49: Simulation Plot Page 1 and 2 with Time Response

Let’s compare these results to those with controller K4ss implemented. Return
to the Main Simulation window and enter K4ss in the Controller editable
text frame. Press the Controller push button. This loads K4ss into the
Controller variable and enables the Compute button. The previous
simulation data with K1ss implemented is deleted from the plot windows since

6-57
6 Graphical User Interface Tools

it is no longer current. Simulate the open-loop and closed-loop nominal systems

by pressing the Compute button. The time responses are shown in Figure 6-50.
Comparing this response to the response of K1ss (Figure 6-49), you can see the
improved performance with K4ss implemented.

Figure 6-50: Simulation Plot Page 1 and 2 with Time Response

6-58
 LFT Time Simulation User Interface Tool: simgui

Printing Menu
Each Simulation Plot Page has a Printing menu for sending that figure to a
printer or saving it to a file. Selecting the Print option from the Printing menu
will pop up a Print dialog box, as shown in Figure 6-51. This window has three
editable text boxes across from Device, Options, and Filename, and a Print
and Cancel button. Device, Options, and Filename correspond to the exact
same inputs you would provide to the standard MATLAB print command.
Therefore, the exact string -dps and a filename have to be entered into the
Device and Filename editable text respectively, to output the plot to a
postscript file.
Leaving the three editable text boxes empty and pressing the Print button will
send the current figure to the printer. This is similar to the MATLAB print
command. Pressing the Cancel button will not execute any print command.
After filling in any or all of the three editable text boxes, pressing the Print
button will execute the MATLAB print command. Pressing either Print or
Cancel executes that command and hides this window.

Figure 6-51: Printing Dialog Box

Loading and Saving Plot Information

The File menu of the main simulation window contains a Load Setup, Save
Setup, and Quit as seen in Figure 6-52.

Figure 6-52: File Pull-Down Menu

6-59
6 Graphical User Interface Tools

Selecting the Load Setup from the menu will display a Load Setup dialog box,
as shown in Figure 6-53. The Load Setup option is used to load plot setup
information that was previously saved using the Save Setup option. You can
enter a Variable name, (the default is SAVESET), and a Filename. If the
Filename editable text is left empty, the variable is loaded from the current
workspace. Pressing the Load button loads the data from the location
described by the Variable and Filename data and hides the window. Pressing
Cancel hides the window and loads no data. The data loaded includes all the
Plots and LineTypes Table information, plotting information and labels, final
time, integration time, sample time, initial conditions, export suffix, and
simulation name if available. If an error occurs during the load operation, an
error message will appear in the main window message bar.
Selecting Save Setup from the File menu will display a Save Setup dialog box,
very similar to Figure 6-53 except with the Save button replacing the Load
button. This option saves all the current plot and line type data along with the
labels, final time, integration time, sample time, initial conditions, export
suffix, and simulation name, if available, to the Variable string name.
The Variable and Filename editable text and the Cancel button operate in the
exact same manner as in the Load Setting dialog box. Pressing the Save
button saves the data to the current workspace in the Variable editable text
string. If this is empty, the default variable name is set to SAVESET. The data is
saved to the filename defined by the Filename editable text, or to the filename
SAVESET if there is no Filename string.
The Load Setup and Save Setup options are extremely useful. They allow you
to customize simulation plots for use with many plants, controllers,
perturbations, and input signals.
The Quit button of the File menu exits the simgui tool and deletes all the data
and windows opened by simgui.

Figure 6-53: Load Setup Dialog Box

6-60
 LFT Time Simulation User Interface Tool: simgui

Simulation Type
You can select three different time simulations from the Options menu on the
main window, as seen in Figure 6-54.

Figure 6-54: Options Pull-Down Menu

The default is continuous-time simulation, which assumes that the Plant,

Controller, and Perturbation variables are either continuous-time
SYSTEMs, CONSTANTs, or not entered. Continuous-time simulations are
performed using the trsp command and make use of the Final Time and
Integration Step Size if you have input them in the Parameter window. The
default final time is the last time data in the Input Signal. An integration step
size is calculated by trsp if you do not provided one. See trsp for more
information.

Note The trsp command is conservative in its selection of an integration

time; therefore, not inputting an Integration Step Size may lead to very
time-consuming simulations.

Discrete-time simulations assume that the Plant, Controller, and

Perturbation variables are either discrete-time SYSTEMs with the same
sample rate, CONSTANTs, or not entered. Discrete-time simulations are
performed using the dtrsp command and make use of the Final Time and
Sample Time if you input them in the Parameter window. The default final
time is the last time data in the Input Signal. The default sample time is 1.
See dtrsp for more information.
A sample-data time simulation assumes that the Plant and Perturbation
variables are either continuous-time SYSTEMs or CONSTANTs. The
Controller variable is assumed to be either a discrete-time SYSTEM,
CONSTANT, or not entered. Sample-data time simulations are performed
using the sdtrsp command and make use of the Final Time, Integration
Step Size and Sample Time if you input them in the Parameter window. The

6-61
6 Graphical User Interface Tools

default final time and integration step size are the same as in the
continuous-time case. The default sample time is the same as the discrete-time
case. See sdtrsp for more information.
Progress of all these simulations is shown in the message bar of the main
window during their computation.

Simulation Parameter Window

The initial Parameter window is shown in Figure 6-48. The Simulation Data
frame contains data that is used in the calculation of the time responses. The
Final Time data is used by all the simulations to define the final simulation
time and overrides the final time of the Input Signal data. The default, an
empty Final Time editable text string, is the final time of the Input Signal
data. The Integration Step Size data defines the integration step size used
by the continuous and sampled-data simulations. The default, an empty
Integration Step Size editable text string, is to have the trsp calculate the
integration step size. The Sampled Time is the sample rate used by the
discrete-time and sampled-data simulations. The default Sampled Time is 1.
All three variables must be positive, nonzero numbers.
Note that all the variables with editable text to the right of them are encased
in < >. This notation denotes that these variables are optional and are not
required. The simgui program either calculates the value, Final Time and
Integration Step Size, or assigns them a default value.
The Plant open-loop plant, Controller, and Perturbation initial conditions
may be entered in their respective editable text boxes. These must be a single,
real column vector whose size is the same as the number of states of the
respective variable. You can input any workspace variable or valid MATLAB
expression to define the initial conditions. The default, with the editable text
boxes left empty, is for all the initial conditions to be zero. Incorrect initial
condition data will result in an error message in the main window message bar.
The second column of data in the Parameter window corresponds to the
Export Suffix, Simulation Tool Name and the linkable variables frame.
Export Suffix is used for data that is automatically exported to the MATLAB
workspace upon computation. This string is appended to default names for the
variables as they are exported to the workspace. The Simulation Tool Name
modifies the title of each simgui figure. The linkable variable frame contains a
list of simgui variables that can be dragged, using the mouse button, to other
µ-Tools user interface tools. Table 6-1 contains a list of the linkable variables

6-62
 LFT Time Simulation User Interface Tool: simgui

and their meaning. See “Dragging and Dropping Icons” for more information
on how to drag and drop µ-Tools linkage variables.

Table 6-1: Linkable Variables for simgui

Linkable Variable Meaning

Plant Open-loop interconnection SYSTEM

K The Controller variable
Pert The Perturbation variable
Input The Input Signal variable
YOLN Open-Loop Nominal time response
YOLP Open-Loop Perturbed time response
YCLN Closed-Loop Nominal time response
YCLP Closed-Loop Perturbed time response

The Export to Workspace frame contains variables that you can export to the
MATLAB workspace. Each variable is a radio button, which you can select or
deselect. If a variable is selected, the variable is exported to the MATLAB
workspace each time that simulation is performed. The default setting is to
export YOLN, YOLP, YCLN, and YCLP every time they are calculated. The Export
to Workspace names and variables are:

• Open-Loop Nominal time response, YOLN

• Open-Loop Perturbed time response, YOLP
• Closed-Loop Nominal time response, YCLN
• Closed-Loop Perturbed time response, YCLP
• Plant, the open-loop interconnection SYSTEM
• Controller
• Perturbation
• Input Signal

6-63
6 Graphical User Interface Tools

The Iteration Suffix string is appended to the end of all the output variables
selected. The default is to export the time simulation data YOLN, YOLP, YCLN,
and YCLP as they are computed.

Note These exported variables are overwritten with their new output
responses after the individual time responses are calculated.

Returning to the example, press the Refresh Variables button in the

Workspace Manager if it is currently open. We can see from the Workspace
Manager that the open-loop nominal time response, YOLN, and the closed-loop
nominal time response, YCLN, are saved in the MATLAB workspace (see
Figure 6-56). This is based on the setting of the Export to Workspace radio
buttons. YOLP and YCLP are not in the MATLAB workspace since the open-loop
and closed-loop perturbed responses were not calculated in this example.

6-64
 LFT Time Simulation User Interface Tool: simgui

Figure 6-55: Workspace Manager Window

6-65
6 Graphical User Interface Tools

Dragging and Dropping Icons

All the µ-Tools graphical user interface (GUI) programs, wsgui, dkitgui, and
simgui have a drag and drop facility. The data icons that can be dragged come
from a scroll table of linkable variable names in each program. In wsgui, the
MATLAB workspace variable names currently in the main scroll table (see
Figure 6-57) are the icons that are dragable. In dkitgui, the scroll table shown
in Figure 6-56 contains the linkable variable names that are dragable. The
simgui scroll table shown in Figure 6-56 contains the variables from simgui
that are dragable. The left column shows the names of the link variables and
the right column describes their meaning.

Figure 6-56: dkitgui (left) and smgui (right) Linkable Variables Icon Tables for
Dragging

To drag an icon, which is a variable name, position the mouse arrow over the
variable name to be dragged. Press the left mouse button and hold the button
down. You have now selected the variable name for dragging. For example,
select the idmod variable for dragging from the wsgui scroll table and move the
mouse button slightly. (Note: the wsgui workspace has just loaded the mk_wts
file from the MATLAB workspace.) Your screen should look similar to
Figure 6-57.

6-66
 Dragging and Dropping Icons

Figure 6-57: Dragging the Controller Variable Name in dkitgui

There are several important facts regarding the dragging and dropping of icons
that should be noted:

• Icons being dragged are only visible in windows, and they are not visible
when the icons are over a frame or other MATLAB uicontrol objects.
• Dragging an icon may be slow on networked computers. This is due to the
overhead of transferring the information about the icon being dragged.
• The drag and drop operation is only completed if the icon is deposited in a
drop box.

6-67
6 Graphical User Interface Tools

Items being dragged need to be deposited directly inside a drop box. An

example of a drop box is shown in Figure 6-59; it’s to the left of the Controller
push button in the simgui main window. To do this, position the icon being
dragged in the center of a drop box and release the mouse button.

Figure 6-58: Drag Khinf from dkitgui to simgui

For example, you can select to drag the fourth D – K iteration controller from
the spinning satellite example in the “D-K Iteration User Interface Tool:
dkitgui” section to the LFT time simulation tool, simgui. Select the Khinf link
variable in the dkitgui main window and move the mouse button slightly
(Figure 6-57). Drag the Khinf link variable over to the main simgui window
(Figure 6-58) and drop it in the Controller drop box, as shown in Figure 6-59.
Upon successfully dragging and dropping the data, the string
grabvar(1,'Khinf') appears in the main simgui window Controller editable
text frame.

6-68
 Dragging and Dropping Icons

Figure 6-59: Drop Khinf into simgui Controller Drop Box

6-69
6 Graphical User Interface Tools

Figure 6-60: Evaluated Drop of Khinf

In simgui and dkigui the linkable variables are to be extracted from a large
storage matrix. The program grabvar performs this operation. The inputs to
the grabvar function are the µ-Tools GUI figure number and the variable name
in the originating function. In this example, the variable dragged from dkitgui
to simgui is Kinf. The dkitgui figure number is 1, which corresponds to the 1
in the simgui editable text: grabvar(1,'Khinf'), as seen in Figure 6-60. The
drag and drop operation is responsible for entering the required input data and
requires no additional input from you.

6-70
 7

Robust Control Examples

7 Robust Control Examples

This chapter has a number of examples to show how to apply the µ-Analysis
and Synthesis Toolbox (µ-Tools) to robust control problems. These examples
include:

• SISO Gain and Phase Margins

• MIMO Loop-at-a-Time Margins

• Analysis of Controllers for an Unstable System

• MIMO Margins Using m

• Space Shuttle Robustness Analysis

• HIMAT Robust Performance Design Example

• F–14 Lateral-Directional Control Design

• A Process Control Example: Two Tank System

7-2
 SISO Gain and Phase Margins

SISO Gain and Phase Margins

Gain and phase margins play an important role in control by defining
robustness margins for systems. Unfortunately these margins, by themselves,
may be misleading in some cases.
Consider the plant

s–2
G ( s ) = ---------------
2s – 1

G = nd2sys([1 -2],[2 -1],-1);

minfo(G)

and four different controllers that stabilize the plant

K1 ( s ) = 1
s+β
K 2 ( s ) = ---------------- ( β > 0, as β → 2 )
βs + 1
βs + 1
K 3 ( s ) = ---------------- ( β > 0, as β → 2 )
s+β
2
s + 2.5 1.7s + 1.5s + 1
K 4 ( s ) = --------------------- -----------------------------------------
2.5s + 1 s 2 + 1.5s + 1.7

Construct the controller K1 and form the closed-loop using formloop.

k1 = 1;
T1 = formloop(G,k1,'neg','pos');
spoles(T1)

The command formloop produces a 2-input, 2-output system T1, as shown in

Figure 7-1. The third input argument 'neg' denotes that there is negative
feedback on the y2 channel. The fourth input argument 'pos' denotes that the
feedback forward input u2 is positive. These are the default values for
formloop.

7-3
7 Robust Control Examples

y1   u1
u1-+ g -
u2
?
K -
y +- - y2
y2  T  u2 1 g
G
,
6 +

Figure 7-1 formloop Interconnection Structure

For the problem at hand, form the closed-loop system and calculate the poles of
the closed-loop system.
K2 = nd2sys([1 1.9],[1.9 1]);
clp2 = formloop(G,K2);
spoles(clp2)
k3 = nd2sys([1.9 1],[1 1.9]);
clp3 = formloop(G,k3);
spoles(clp3)
k4 = nd2sys([1 2.5],[2.5 1]);
k4 = mmult(k4,nd2sys([1.7 1.5 1],[1 1.5 1.7]));
clp4 = formloop(G,k4);
spoles(clp4)

For gain and phase margins, we are interested in the loop transfer function L(s)
:= G(s) K(s). Select a frequency range of 0.01 rad/sec to 100 rad/sec with 100
points. Calculate the frequency response of the loop transfer function and plot
the Bode and Nyquist diagrams of L. The command add_disk adds a unit disk
to the Nyquist plot. The results are shown in Figure 7-2.
omega = logspace(-2,2,100);
G_g = frsp(G,omega);
k1_g = frsp(k1,omega);L1 = mmult(G_g,k1_g);
K2_g = frsp(K2,omega);L2 = mmult(G_g,K2_g);
k3_g = frsp(k3,omega);L3 = mmult(G_g,k3_g);
k4_g = frsp(k4,omega);L4 = mmult(G_g,k4_g);
vplot('nyq',L1,'-',L2,'--',L3,'-.',L4,'.'), add_disk
axis([-4 1 -1 1])
vplot('bode',L1,'-',L2,'--',L3,'-.',L4,'.'), grid

7-4
 SISO Gain and Phase Margins

Controller K1 provides a compromise between a large gain and phase margin.

Controller K2 optimizes the gain margin for this example, whereas controller
K3 optimizes the phase margin. Controller K4 tries to optimize both gain and
phase margins, resulting with about 50° phase margin, and a gain margin of 3
(it can tolerate increase or decrease in gain by a factor of 3). These margins,
though, are very sensitive to small changes in the plant (a 50% gain change, to
either 1.5 or 0.67, reduces the phase margin to less than 10°) illustrating that
gain and phase margins are not always an accurate measure of the robustness
of a control design even for single-input/single-output systems.

7-5
7 Robust Control Examples

Nyquist Plot
1

0.8

0.6

0.4

0.2
Imaginary

0.2

0.4

0.6

0.8

1
4 3.5 3 2.5 2 1.5 1 0.5 0 0.5 1
Real

Bode Plot
1
10
Log Magnitude

0
10

1
10 2 1 0 1 2
10 10 10 10 10
Frequency (radians/sec)

100
Phase (degrees)

150

200 2 1 0 1 2
10 10 10 10 10
Frequency (radians/sec)

k1 (solid line) k2 (dashed line) k3 (dash-dotted line) k4 (dotted line)

Figure 7-2: Nyquist and Bode Plots of the Loop Gain with k1, k2, k3, and k4

7-6
 MIMO Loop-at-a-Time Margins

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. We will see that margins of the individual loops may be very sensitive
to small perturbations within other loops.
The nominal closed-loop system considered here is shown in Figure 7-3.

d - -
, 6 d- K - G
,6

Figure 7-3: Nominal Closed-Loop System

G and K are 2 × 2 multi-input/multi-output (MIMO) systems, defined as

2
G := ------------------ s – α α(s + 1) ,
1
K = I2
2 2
s + α –α ( s + 1 ) s – α2

Set α := 10, construct G in state-space form and compute its frequency

response.
a = [0 10;-10 0];
b = eye(2);
c = [1 10;-10 1];
G = pck(a,b,c);
K = eye(2);
omega = logspace(-3,2,50);
Gg = frsp(G,omega);
minfo(Gg);

We want to consider perturbations to both input channels, as shown in

Figure 7-4.

7-7
7 Robust Control Examples

- 1
d - - -d?
, 6 d- K - -d G
,6 - 2 6

Figure 7-4: Perturbed Closed-Loop System

Break the loop at the place where one perturbation is, and compute the
open-loop transfer function. This transfer function will be a function of the
remaining perturbation. For instance, to check the margins in the first channel
with perturbations in the second channel, consider the diagram in Figure 7-5.

d - -
w1
-
z1
, 6 d- K - d - G
,6 w2
- 2 6z2

Figure 7-5: Closed-Loop System with Loop Broken in Channel 1 and

Perturbation in Channel 2

In this particular example, K is the identity, so the loop in Figure 7-5 can be
redrawn, as shown in Figure 7-6. This figure is easily constructed using starp.

Figure 7-6: Redrawn Closed-Loop System with Loop Broken in Channel 1 and
Perturbation in Channel 2

A Bode or Nyquist plot of the transfer function from z1 to w1 reveals the margin
(nearness to +1 point in this case, since the negative sign is embedded in the

7-8
 MIMO Loop-at-a-Time Margins

transfer function from z to w) in the first channel as a function of perturbations

in the second channel. Refer to Figure 7-6 and calculate the nominal transfer
function obtained by breaking the loop in the 1st channel. A Bode plot of the
transfer function is shown in Figure 7-7. Note that this loop is an integrator, so
the single-loop margin is great. In a similar manner, calculate the nominal
transfer function obtained by breaking the loop in the second channel. The
results are shown in Figure 7-7.
nom_ch1_lb = starp(mscl(Gg,-1),1,1,1);
nom_ch2_lb = starp(1,mscl(Gg,-1),1,1);
vplot('bode',nom_ch1_lb,'-',nom_ch2_lb,'--');

Bode Plot: Single Loop Broken in Channel 1 and 2

4
10
Log Magnitude

2
10

0
10

2
10 3 2 1 0 1 2
10 10 10 10 10 10
Frequency (radians/sec)
Phase (degrees)

100

0 3 2 1 0 1 2
10 10 10 10 10 10
Frequency (radians/sec)

Figure 7-7: Bode Plot the Single-Loop Transfer Functions

The previous Bode plots show that each channel can tolerate large,
loop-at-a-time, gain variations. Now though, consider a simultaneous 10%
variation in each loop. Specifically, check the internal stability of the perturbed
closed-loop shown in Figure 7-4, redrawn in Figure 7-8, with δ 1 = – 1 ⁄ 101 ,
and δ 2 = 1 ⁄ 101 .

7-9
7 Robust Control Examples

To do this, note that since K = I2, the perturbed loop in Figure 7-4 is simply the
star product (starp) of –G and the matrix [1+delta1 0; 0 1+delta2].

Figure 7-8: Equivalent Perturbed Closed-Loops

Note When two SYSTEM matrices are connected using starp, and all of the
inputs of the top system are all of the outputs of the bottom system, and all of
the inputs of the bottom system are all of the outputs of the top system, then
the output of starp is the “A” matrix of the interconnection, stored as a
regular MATLAB matrix.

The following code creates two perturbations of size 101 , creates the
perturbation matrix, forms the closed-loop system and calculates the
closed-loop eigenvalues.
delta1 = -1/sqrt(101);
delta2 = 1/sqrt(101);
delta = [1+delta1 0 ; 0 1+delta2];
minfo(G)
minfo(delta)
clpAmat = starp(mscl(G,-1),delta,2,2)
minfo(clpAmat)
eig(clpAmat)

Notice that the eigenvalues of clpAmat are unstable. Hence, this small (10%),
simultaneous perturbation to both channels causes instability (slightly larger
values for δ1 and δ2 push the eigenvalues further into the right-half-plane).

7-10
 MIMO Loop-at-a-Time Margins

We can see the effect of simultaneous perturbations with the help of a Nyquist
plot. Refer back to Figure 7-6 and calculate the transfer function obtained by
breaking the loop in the first channel, with a perturbation of δ2 = 0.001. The
single-loop margin in channel 1 is still quite large, since the Nyquist plot, solid
line in Figure 7-9 intersects the real axis around 11. Try a slightly larger value
for δ2 (again, note that the +1 point is of interest due to the extra minus (–) sign
on G).
delta2 = 0.001;
pert_ch1_lb1 = starp(mscl(Gg,-1),1+delta2,1,1);
vplot('nyq',pert_ch1_lb1);
delta2 = 0.01;
pert_ch1_lb2 = starp(mscl(Gg,-1),1+delta2,1,1);
vplot('nyq',pert_ch1_lb1,'-',pert_ch1_lb2,'--');
add_disk

Note the significant degradation in the margin due to a 1% variation in channel

2.

Nyquist Plot: Loop Broken in Channel 1, Delta2 =.001, .01

6

3
Imaginary

1
2 0 2 4 6 8 10 12
Real

unit disk (solid) δ2 = 0.001 (solid) δ2 = 0.01 (dashed)

Figure 7-9: Nyquist Plot of Simultaneously Perturbed Closed-Loop System

7-11
7 Robust Control Examples

It is easy to write a for loop (as in ex_ml1) that calculates the open-loop
frequency response in channel 1 for several values of δ2. Six δ2 values are
constructed from 0.001 to 0.1. A for loop calculates the perturbed loop transfer
function in channel 1 for the six values of δ2 and their responses are plotted on
a Nyquist diagram.

file:ex_ml1.m
delta2values = logspace(-3,-1,6);
pert_ch1_lb = [];
for i=1:length(delta2values)
delta2 = delta2values(i);
pert_ch1_lb = ...
sbs(pert_ch1_lb,starp(mscl(Gg,-1),1+delta2,1,1));
end
xa = vpck([-2; 12],[1 2]);
ya = vpck([-sqrt(-1); 6*sqrt(-1)],[1 2]);
vplot('nyq',pert_ch1_lb,xa,ya);
title('Loop Broken in Channel 1, Delta2 in [.001 .1]')
xlabel('Real')
ylabel('Imag')

Typing
ex_ml1
at the command line generates Figure 7-10. Note that the margin in channel 1
rapidly disappears due to small perturbations in channel 2.

7-12
 MIMO Loop-at-a-Time Margins

Nyquist Plot: Loop Broken in Channel 1, Delta2 in [.001 .1]

6

3
Imaginary

1
2 0 2 4 6 8 10 12
Real

Figure 7-10: Nyquist Plot of Simultaneously Perturbed Closed-Loop System

from δ2 = 0.001 to δ2 = 0.1

Exactly analogous results hold for breaking the loop in channel 2, and
considering small perturbations in channel 1. This can be verified by closing
the upper loop of –G with (1 + δ1). For example,
delta1 = -0.01;
pert_ch2_lb = starp(1+delta1,mscl(Gg,-1),1,1);
vplot('nyq',pert_ch2_lb);
add_disk

7-13
7 Robust Control Examples

Nyquist Plot: Loop Broken in Channel 2, Delta1 = 0.01

2

1.5

0.5
Imaginary

0.5

1.5

2
2 1.5 1 0.5 0 0.5 1
Real

Figure 7-11: Nyquist Plot of Simultaneously Perturbed Closed-Loop System

for δ1 = –0.01

Hence, in a multivariable system the single loop margins may be good though
simultaneous interaction of perturbations in each loop may lead to instability,
even for small perturbations. The structured singular value, µ, can be used to
detect the nonrobustness in this feedback system. To see this we will reanalyze
this example using µ in the “MIMO Margins Using m” section in this chapter.

7-14
 Analysis of Controllers for an Unstable System

Analysis of Controllers for an Unstable System

Unstable example Recall from the “Unmodeled Dynamics” section in Chapter
4, the example with the nominal plant model

1
G := ------------
s–1

and the multiplicative input uncertainty model

--- ( --- s + 1 )
1 1
4 2
W u := ----------------------
1
------- s + 1
32

G and Wu describe a set M(G, Wu) of plant models defined by

M ( G, W u ) := { G ( I + ∆ G W u ) : max ∆ G ( jω ) ≤ 1 }.
ω
(7-1)

with the restriction that the number of right-half plane poles of perturbed plant
equal the number of right-half plane poles of G. The nominal model G,
weighting function Wu, and the unknown transfer function ∆G are used to
parameterize all the possible models.
We are interested in the stability and performance of the closed-loop system for
all possible plant models in the set M(G, Wu). In this example, we choose the
performance objective to be a stable, closed-loop system, and output
disturbance rejection up to 0.6 rad/sec, with at least 100:1 disturbance rejection
at DC.
This objective can approximately be represented as a weighted H∞ norm
constraint on the sensitivity function,

Wp
------------------ ≤ 1
1 – G̃K ∞

for all G̃ ∈ M(G, Wu), using the weighting function

7-15
7 Robust Control Examples

1
--- s + 0.6
4
W p := -------------------
- = s + 0.006
s

A block diagram of this uncertain closed-loop system illustrating the

performance objective (closed-loop transfer function from d → e) is shown in
Figure 7-12.

Plant model set

-
G
d
Wu
- K u 6 -+ ?g+- G - g? - Wp -e

Figure 7-12: Uncertain Closed-Loop System

Grouping G, Wu, and Wp together (as P), the uncertain closed-loop system is
redrawn in Figure 7-13.

-
G
w z

e P 
 d
y u
- K
Figure 7-13: General Interconnection for Uncertain System

Type
ex_unic

to create the plant model, weighting functions and the open-loop

interconnection P. The commands in ex_unic are as follows.

7-16
 Analysis of Controllers for an Unstable System

file: ex_unic.m
G = nd2sys(1,[1 -1]);
Wu = nd2sys([0.5 1],[0.03125 1],.25);
WP = nd2sys([.25 0.6],[1 0.006]);
systemnames = 'G Wu WP';
sysoutname = 'P';
inputvar = '[ z; d; u ]';
input_to_p = '[ z + u ]';
input_to_Wu = '[ u ]';
input_to_WP = '[ G + d ]';
outputvar = '[ Wu; WP; G + d ]';
cleanupsysic = 'yes';
sysic

Next, consider two controllers, K1 and K2 which stabilize the nominal plant
model G.

0.9s + 1 2.8s + 1
K 1 = – 10 --------------------- K 2 = – 1 ---------------------
s s

For each controller, define M := FL(P,K). M22 is the nominal d → e transfer

function, and hence, nominal performance is characterized by ||M22||∞ ≤ 1. The
magnitude of the M22 transfer function with K1 and K2 implemented is shown
in Figure 7-14. Both controllers easily achieve the nominal performance
objective.
M11 is the transfer function that the perturbation ∆ sees, so that robust
stability over M(G, Wu) is equivalent to ||M11||∞ < 1. The magnitude of the M11
transfer function is shown in Figure 7-15. Both controllers achieve the robust
stability objective.

7-17
7 Robust Control Examples

Nominal Performance (K1 solid, K2 dashed)

0.6

0.5

0.4
M22

0.3

0.2

0.1

0 2 1 0 1 2
10 10 10 10 10
Frequency (rad/sec)

Figure 7-14: Nominal Performance Plots

Robust Stability (K1 solid, K2 dashed)

1

0.9

0.8

0.7

0.6
M11

0.5

0.4

0.3

0.2

0.1 2 1 0 1 2
10 10 10 10 10
Frequency (rad/sec)

Figure 7-15: Robust Stability Plots

7-18
 Analysis of Controllers for an Unstable System

The actual goal is to achieve the performance objective for every plant model
as defined by the dashed-box in Figure 7-12. This objective is defined as robust
performance.We can use the structured singular value, µ, to analyze the
closed-loop system to determine if robust performance is achieved. Achieving
robust performance is mathematically equivalent to
maxµ ∆ ( M ( jω ) ) ≤ 1
ω∈R
P

where ∆P is the extended uncertainty set, consisting of the actual scalar

unmodeled dynamics (i.e., a complex block) uncertainty block (representing
∆G), and the fictitious uncertainty block used to reformulate the performance
criterion as a stability problem (see section “Using m to Analyze Robust
Performance” in Chapter 4 for details). The block structure used to evaluate
robust performance is
∆P := {diag[∆,∆F] : ∆ ∈ C, ∆F ∈ C},

A plot of µ ∆ ( M ( jω ) ) is shown in Figure 7-16.

P

Robust Performance (K1 solid, K2 dashed)

1.2

0.8
mu(M)

0.6

0.4

2 1 0 1 2
10 10 10 10 10
Frequency (rad/sec)

Figure 7-16: Robust Performance Plots

7-19
7 Robust Control Examples

Note that while K1 achieves better nominal performance than K2, the closed-
loop system with K2 has better robust performance properties than with K1. In
fact, the controller K2 does, just barely, achieve robust performance. Hence, for
every plant G̃ ∈ M ( G, W u ) , the closed-loop system, with K2 is stable, and
moreover

Wp
--------------------- < 1
1 – G̃K 2 ∞

This is not true for the closed-loop system with K1, as the robust performance
test, using µ is not satisfied for FL(P,K1).
The script-file
ex_usrp

constructs the controllers, forms the closed-loop systems, calculates the

closed-loop frequency responses, performs the analysis, and generates the
plots. A listing of ex_usrp is as follows.

7-20
 Analysis of Controllers for an Unstable System

file: ex_usrp.m
K1 = nd2sys([.9 1],[1 0],-10);
K2 = nd2sys([2.8 1],[1 0],-1);
om = logspace(-2,2,80);
M1 = starp(P,K1);
M2 = starp(P,K2);
M1g = frsp(M1,om);
M2g = frsp(M2,om);
uncblk = [1 1];
fictblk = [1 1];
deltaset = [uncblk;fictblk];
bnds1 = mu(M1g,deltaset);
bnds2 = mu(M2g,deltaset);
vplot('liv,m',sel(M1g,2,2),'-',sel(M2g,2,2),'--')
xlabel('Frequency (rad/sec)')
ylabel('M22');
title('Nominal Performance (K1 solid, K2 dashed)');
vplot('liv,m',sel(M1g,1,1),'-',sel(M2g,1,1),'--')
xlabel('Frequency (rad/sec)')
ylabel('M11');
title('Robust Stability (K1 solid, K2 dashed)');
vplot('liv,m',sel(bnds1,1,1),'-',sel(bnds2,1,1),'--')
xlabel('Frequency (rad/sec)')
ylabel('mu(M)');
title('Robust Performance (K1 solid, K2 dashed)');

How does the weighted sensitivity transfer function degrade with uncertainty?
This can be answered easily using the worst-case performance analysis, which
shows the worst-case degradation of performance as a function of the size of the
uncertainty.We also construct, for each closed-loop system, the worst-case
perturbation of size (i.e., the worst case plant from M(G, Wu)) and use this later
in time-domain simulations. The script file
ex_wcp

performs these calculations.

7-21
7 Robust Control Examples

file: ex_wcp.m
[deltabad1,wcp_l1,wcp_u1] = wcperf(M1g,uncblk,1,8);
[deltabad2,wcp_l2,wcp_u2] = wcperf(M2g,uncblk,1,8);
vplot(wcp_l1,wcp_u1,wcp_l2,'--',wcp_u2,'--')
xlabel('Size of Delta_G')
ylabel('Weighted Sensitivity')
title('Performance Degradation (K1 solid, K2 dashed)')

The performance degradation curve is shown in Figure 7-17. The fourth

argument to wcperf, in this case 8, means that there will be at least eight
points in the performance degradation curve.

Performance Degradation (K1 solid, K2 dashed)

3.5

2.5
Weighted Sensitivity

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Size of Delta_G

k1 (solid line) k2 (dashed line)

Figure 7-17: Performance Degradation

This clearly shows that for perturbations ∆G satisfying

max ∆ G ( jω ) Š 0.76,
ω

7-22
 Analysis of Controllers for an Unstable System

the robust closed-loop performance using K2 is better than that obtained using
K1. The perturbations deltabad1 and deltabad2 are the worst-case
perturbations ∆G, of size 1.They each cause the weighted sensitivity to degrade
a maximal amount, over all perturbations of size 1. In the case of K1, the
perturbation deltabad1 causes the weighted sensitivity to degrade to
approximately 3.5, while in the case of K2, deltabad2 causes the weighted
sensitivity to degrade to approximately 1.0. We will use these worst-case
perturbations in time-domain simulations.
Based on the robust performance analysis results, K2 will stabilize the nominal
plant and the seven extreme plant models discussed in the “Unmodeled
Dynamics” section in Chapter 4. Also, the performance degradation, in terms
of the weighted sensitivity function, should be small. Using controller K1,
stability is also guaranteed. However, since FL(P,K1)(jω) has a peak µ-value of
about 1.2, it is clear that K1 does not have robust performance over the set
M(G,Wu) and we expect the performance degradation over the seven extreme
plants to be worse than that using K2.
These seven plants, all of which are members of the set M(G,Wu) and the
worst-case plant from M(G,Wu), are

G 1 = ------------  -----------------
1 6.1 1.425
G 2 = -----------------------
s – 1  s + 6.1 s – 1.425

G 4 = ------------  ----------------------------
.67 1 – 0.07s + 1
G 3 = --------------------
s – 0.67 s – 1  0.07s + 1 

1   1  
2 2
70 70
G 5 = ------------  ------------------------------------------------------------- G 6 = ------------  ----------------------------------------------------------
s – 1  s 2 + 2 ⋅ 0.15 ⋅ 70s + 70 2 s – 1  s 2 + 2 ⋅ 5.6 ⋅ 70s + 70 2

50 6
G 7 = ------------  ---------------
1
G wc = [ G ( 1 + ∆ wc W u ) ]
s – 1  s + 50

and are constructed via

ex_mkpl

The contents of ex_mkpl are listed below.

7-23
7 Robust Control Examples

file: ex_mkpl.m
G1 = mmult(G,nd2sys([6.1],[1 6.1]));
G2 = nd2sys([1+0.425],[1 -1-0.425]);
G3 = nd2sys([1-0.33],[1 -1+0.33]);
G4 = mmult(G,nd2sys([-0.07 1],[0.07 1]));
G5 = mmult(G,nd2sys([70^2], [1 2*0.14*70 70^2]));
G6 = mmult(G,nd2sys([70^2], [1 2*5.6*70 70^2]));
Gt = nd2sys([50],[1 50]);
G7 = mmult(G,Gt,Gt,Gt,Gt,Gt,Gt);
Gwc1 = mmult(G,madd(1,mmult(deltabad1,Wu)));
Gwc2 = mmult(G,madd(1,mmult(deltabad2,Wu)));

The closed-loop system shown in Figure 7-18 is used for time simulations.
Twenty time simulations are computed, for all combinations of K1 and K2, with
the nominal plant G, the seven extreme plants G1,. . .,G7, and the two
worst-case plants Gwc1 and Gwc2. Rather than simulate an output disturbance,
we manipulate the diagram, and simulate a unit-step reference command.

Figure 7-18: Simulation Block Diagram

This interconnection is most easily constructed using formloop, as described

earlier. We use positive feedback (as that was the convention used in the design
of the controllers) and invert the reference signal gain, to accomplish the
tracking. The (2,1) entry of the MIMO system that formloop produces is the
SYSTEM matrix from reference input r to plant output y.

7-24
 Analysis of Controllers for an Unstable System

Typing
ex_mkclp

forms the 20 SISO systems, as below.

file: ex_mkclp.m
clpg_k1 = sel(formloop(G,K1,'pos','neg'),2,1);
clpg1_k1 = sel(formloop(G1,K1,'pos','neg'),2,1);
clpg2_k1 = sel(formloop(G2,K1,'pos','neg'),2,1);
clpg3_k1 = sel(formloop(G3,K1,'pos','neg'),2,1);
clpg4_k1 = sel(formloop(G4,K1,'pos','neg'),2,1);
clpg5_k1 = sel(formloop(G5,K1,'pos','neg'),2,1);
clpg6_k1 = sel(formloop(G6,K1,'pos','neg'),2,1);
clpg7_k1 = sel(formloop(G7,K1,'pos','neg'),2,1);
clpgwc1_k1 = sel(formloop(Gwc1,K1,'pos','neg'),2,1);
clpgwc2_k1 = sel(formloop(Gwc2,K1,'pos','neg'),2,1);
clpg_K2 = sel(formloop(G,K2,'pos','neg'),2,1);
clpg1_K2 = sel(formloop(G1,K2,'pos','neg'),2,1);
clpg2_K2 = sel(formloop(G2,K2,'pos','neg'),2,1);
clpg3_K2 = sel(formloop(G3,K2,'pos','neg'),2,1);
clpg4_K2 = sel(formloop(G4,K2,'pos','neg'),2,1);
clpg5_K2 = sel(formloop(G5,K2,'pos','neg'),2,1);
clpg6_K2 = sel(formloop(G6,K2,'pos','neg'),2,1);
clpg7_K2 = sel(formloop(G7,K2,'pos','neg'),2,1);
clpgwc1_K2 = sel(formloop(Gwc1,K2,'pos','neg'),2,1);
clpgwc2_K2 = sel(formloop(Gwc2,K2,'pos','neg'),2,1);

These closed-loop systems are simulated with a unit-step reference input at t =

0.5. The time response with controllers K1 (top) and K2 (bottom) implemented
are shown in Figure 7-19.
Typing
ex_ustr

runs the simulations, and generates the time response plots.

7-25
7 Robust Control Examples

file: ex_ustr.m
stepref = step_tr([0 0.5 10],[0 1 1],.01,10);
Tfinal = 10;
tinc = 0.01;
yg_k1 = trsp(clpg_k1,stepref,Tfinal,tinc);
yg1_k1 = trsp(clpg1_k1,stepref,Tfinal,tinc);
yg7_k1 = trsp(clpg7_k1,stepref,Tfinal,tinc);
ygwc1_k1 = trsp(clpgwc1_k1,stepref,Tfinal,tinc);
ygwc2_k1 = trsp(clpgwc2_k1,stepref,Tfinal,tinc);
yg_k2 = trsp(clpg_k2,stepref,Tfinal,tinc);
yg1_k2 = trsp(clpg1_k2,stepref,Tfinal,tinc);
ygwc2_k2 = trsp(clpgwc2_k2,stepref,Tfinal,tinc);
subplot(211)
vplot(vdcmate(yg_k1,5),'+',yg1_k1,yg2_k1, yg3_k1,yg4_k1,...
yg5_k1,yg6_k1,yg7_k1,ygwc1_k1,ygwc2_k1);
xlabel('Time (seconds)')
title('Closed-loop responses using k1')
subplot(212)
vplot(vdcmate(yg_k2,5),'+',yg1_k2,yg2_k2, yg3_k2,yg4_k2,...
yg5_k2,yg6_k2,yg7_k2,ygwc1_k2,ygwc2_k2);
xlabel('Time (seconds)')
title('Closed-loop responses using K2')

Both controllers stabilize all plant models. The closed-loop performance

provided by K2 is very similar for all the plants. This is expected based on the
robust performance analysis using µ, which showed that every plant in the set
M(G,Wu) is stabilized and effectively controlled with K2. The closed-loop
performance provided by K1 is not very robust. The large overshoot, undershoot
and oscillations for some of the responses reinforces the notion that K1 does not
achieve robust performance, as defined by the uncertain plant model set
M(G,Wu), and the weighted sensitivity transfer function.

7-26
 Analysis of Controllers for an Unstable System

Closedloop responses using K1

3

2.5

1.5

0.5

0.5

1
0 1 2 3 4 5 6 7 8 9 10
Time (seconds)

Closedloop responses using K2

1.8

1.6

1.4

1.2

0.8

0.6

0.4

0.2

0.2
0 1 2 3 4 5 6 7 8 9 10
Time (seconds)

Figure 7-19: Time Response with K1 (top) and K2 (bottom) Implemented

7-27
7 Robust Control Examples

Redesign Controllers Using D – K Iteration

The D – K iteration, using dkitgui, can be used to design a robust controller
for the uncertain plant model M(G,Wu) and weighted sensitivity performance
objective implied by the interconnection structure in Figure 7-12. Start
dkitgui, and enter the data as shown in the dkitgui Setup window,
Figure 7-20.

Figure 7-20: Unstable System dkitgui Setup Window

Return to the main Iteration window, as shown inFigure 7-21, and pull down
the Iteration menu and select 3 from the Auto Iterate submenu. Figure 7-21
shows the main Iteration window after the third iteration has completed.

7-28
 Analysis of Controllers for an Unstable System

Figure 7-21: Main dkitgui Window After Three Iterations

Controller K1dk corresponds to the first D – K iteration H∞ design and

controller K3dk corresponds to the third D – K iteration design.
Controller K1dk stabilizes all seven plant models, though the performance is
poor, with large settling times, and large overshoot. This is not surprising,
based on the robust performance analysis using µ. The value of µ with K1dk
implemented was 1.55(> 1).
Controller K3dk achieves a closed-loop, robust performance µ-value of about
0.7, guaranteeing stability and good closed-loop performance for all seven plant
models (as these are contained in the set M(G,Wu)). The simulations agree with
this conclusion. Note the improvement in overshoot and settling time (and
greater insensitivity) over all of the previous closed-loop simulations (using K1,
K2 and K1dk). The nominal responses are denoted by the “+” symbol linetype.
Typing
ex_ustrd

generates the plots shown inFigure 7-22.

7-29
7 Robust Control Examples

Closedloop responses using K1

3

2.5

1.5

0.5

0.5

1
0 1 2 3 4 5 6 7 8 9 10
Time (seconds)

Closedloop responses using K3dk

1.6

1.4

1.2

0.8

0.6

0.4

0.2

0.2

0.4
0 1 2 3 4 5 6 7 8 9 10
Time (seconds)

Figure 7-22: Time Response with K1dk (top) and K3dk (bottom) Implemented

7-30
 Analysis of Controllers for an Unstable System

Near the H∞ optimal solution, the controller tends to have high bandwidth,
providing some argument against using the H∞ norm as the sole measure of
performance. For this reason, we often back off from optimality, leading to a
lower bandwidth, and more reasonable controller. In the Parameter window,
change Gamma Min to 0.78, and change Gamma Max to 0.78. This fixes the next
control design to be at γ = 0.78, which is approximately 10% above the optimal
value of 0.71. Then, in the Iteration window, press Control Design to produce
K4dk. This controller will sacrifice some (about 10%) H∞ performance, but have
much lower bandwidth. You should repeat the time-domain simulations for
this controller.
You can also do simple model reduction on this controller, using sysbal and
strunc. In the Setup window, type
reducek(

in the <Controller> data entry box. Then, drag, from the Iteration window,
the linkable variable Khinf into the Setup window drop box for the controller.
Finally, type
,3)

in the controller data entry box, which should now read

reducek(grabvar(2,'Khinf'),3)

The 2 corresponds to the MATLAB figure number associated with the

Iteration window. In your case, it may not be 2. However, the dragging action
will put the correct number in this place. Next, press the <Controller>
pushbutton. This will grab the current H∞ controller, balance it, and reduce it
to third order using a simple M-file reducek, and place the reduced order
controller back into Kuse. Kuse is the controller for which all analysis in
dkitgui takes place (in dkitgui, the variable Kuse is automatically set to
Khinf immediately after Khinf is designed). This reenables the Form Closed
Loop pushbutton, which will implement the controller stored in Kuse. Pressing
the Frequency Response and Compute Mu pushbuttons, you can determine
the effectiveness of the reduced order controller, and repeat the process for
different choice of controller order. In this example, a third order controller is
suitable.
Time responses using this suboptimal, reduced order controller are shown in
Figure 7-23.

7-31
7 Robust Control Examples

Closedloop responses using K4reduced

1.8

1.6

1.4

1.2

0.8

0.6

0.4

0.2

0.2
0 1 2 3 4 5 6 7 8 9 10
Time (seconds)

Figure 7-23: Time Response with K4dkreduced Implemented

7-32
 MIMO Margins Using µ

MIMO Margins Using µ

Recall the conclusions of the “MIMO Loop-at-a-Time Margins” section
concerning loop-at-a-time margins in MIMO systems. These margins can be
inaccurate measures of multivariable robustness margins. In this exercise, we
compare two controllers for a weighted robust stability problem similar to that
in the“MIMO Loop-at-a-Time Margins” section, and see that the structured
singular value, µ, accurately predicts their robust stability characteristics.
Consider an uncertainty model consisting of multiplicative uncertainty at the
plant input in each channel. A block diagram of the model of the open-loop
plant, with perturbations, is shown in Figure 7-24.

w1
1  z1
? 6
W 1


y1   c? u1

y2   c
G

6
u2

W2 
6 ?
w2
2  z2

Figure 7-24: Plant Model with Multiplicative Uncertainty

The linear, time-invariant perturbations, represented by transfer functions δ1
and δ2, are used to model the uncertainty in the plant, and are assumed to
satisfy

• δi(s) is stable
• max δ i ( jω ) < 1 ,
ω∈R
but otherwise be unknown. Everything inside the dashed box, in particular, the
transfer functions G, Wδ1 and Wδ2, along with the specific interconnection, are
assumed given and known. Since the uncertainty represented by the δi is

7-33
7 Robust Control Examples

normalized to 1, the transfer functions Wδi reflect the frequency-dependent

size of the uncertainty.
The first input channel is modeled with 10% uncertainty at low frequency,
increasing to 100% uncertainty at approximately 20 rads/sec, and larger
uncertainty for higher frequencies. In channel 2, the low frequency uncertainty
is 20%, rising to 100% at approximately 45 rads/sec. Type in the following data
for the problem.
a = [0 10;-10 0];
b = eye(2);
c = [1 10;-10 1];
G = pck(a,b,c);
wdel1 = nd2sys([1 2],[1/60 20]);
wdel2 = nd2sys([1 9],[1/40 45]);
omega = logspace(-3,2,50);
wdel1_g = frsp(wdel1,omega);
wdel2_g = frsp(wdel2,omega);
vplot('liv,lm',wdel1_g,wdel2_g)
title('Multiplicative Uncertainty Weights')

Multiplicative Uncertainty Weights

2
10

1
10
Magnitude

0
10

1
10 1 0 1 2 3 4
10 10 10 10 10 10
Frequency (rad/sec)

Figure 7-25: Multiplicative Uncertainty Weights

7-34
 MIMO Margins Using µ

To use the µ theory, we need to isolate the perturbations from the known
components. We do this by constructing a four input, four output system, with
two fictitious input/output pairs (w,z), as well as two actual input/output pairs
(u,y). The fictitious inputs and outputs are used to model the multiplicative
uncertainty at the plant input. The system is called P, and is shown below.

The internal structure of P is shown in Figure 7-26.

Figure 7-26: Plant Model with Multiplicative Uncertainty Blocks Pulled Out
The SYSTEM matrix P can be constructed easily from the components, using
the SYSTEM interconnection program, sysic.
systemnames = 'G wdel1 wdel2';
inputvar = '[w{2}; u{2}]';
input_to_G = '[u(1) + wdel1; u(2) + wdel2]';
input_to_wdel1 = '[w(1)]';
input_to_wdel2 = '[w(2)]';
outputvar = '[u; G]';
sysoutname = 'P';
sysic;

7-35
7 Robust Control Examples

The feedback controller for this problem will measure the variables y, and
generate control signals, u. In a block diagram, the perturbed closed-loop
system appears as shown in Figure 7-27. In terms of P, the perturbed
closed-loop system appears as Figure 7-28.

w1
1  z1
? 6
W1 
  ?c  6
 G
 c K

6  6
W2

6 ?
P
w2
2  z2

Figure 7-27: Closed-Loop System with δ1 and δ2 Uncertainty

- 1 0
z1 0 2 w1
z2 w2


P

y1 u1

-
y2 u2
K

Figure 7-28: Closed-Loop System with Uncertainty

Next, we need to construct the transfer function that the 2 × 2 perturbation

matrix ∆ sees, namely M := FL(P,K). This transfer function is computed by

7-36
 MIMO Margins Using µ

closing the measurement/control feedback loop of the system P with the

controller. For this example, we will compare two controllers, Ka, a constant
gain 2 × 2 matrix, and Kb, a four-state, two input, two output SYSTEM. The
command ex_mmmk makes up controllers Ka and Kb.
ex_mmmk;
ka
ka =
-1 0
0 -1
seesys(kb)

–1.9e+01 2.6e+01 –4.1e+01 –2.4e+01 | –5.4e+00 –2.6e+00

–8.9e+00 –4.5e+01 9.4e+00 8.3e+01 | 5.1e+00 –6.5e+00
4.1e+01 2.2e–02 –7.4e+01 –6.7e+01 | 4.1e+00 3.5e+00
1.6e+01 –8.4e+01 1.2e+01 –3.4e+02 | 4.9e+00 –5.0e+00
------------------------------------- | ------------------
–1.1e+00 –7.0e+00 –2.2e+00 5.2e+00 | 0.0e+00 0.0e+00
5.9e+00 –4.5e+00 5.0e+00 4.8e+00 | 0.0e+00 0.0e+00

Next, close the lower loop of P with the controller (Ka or Kb) yielding M, which
casts the robust stability problem as that depicted in Figure 4-13. Use starp to
compute the closed-loop system M. A block diagram of this is shown in
Figure 7-29.

z  w z M  w
P

- K
Figure 7-29: Closed-Loop Interconnection Block Diagrams

7-37
7 Robust Control Examples

Ma = starp(P,ka,2,2);
Mb = starp(P,kb,2,2);
spoles(Ma)

real imaginary frequency damping

–1.8000e+03 0.0000e+00 1.8000e+03 1.0000e+00
–1.2000e+03 0.0000e+00 1.2000e+03 1.0000e+00
–1.0000e+00 0.0000e+00 1.0000e+00 1.0000e+00
–1.0000e+00 0.0000e+00 1.0000e+00 1.0000e+00

spoles(Mb)

real imaginary frequency damping

–1.5959e+01 0.0000e+00 1.5959e+01 l.0000e+00
–2.7899e+01 –1.6622e+01 3.2475e+01 8.5908e–01
–2.7899e+01 1.6622e+01 3.2475e+01 8.5908e–01
–4.7231e+01 –6.8509e+00 4.7725e+01 9.8964e–01
–4.7231e+01 6.8509e+00 4.7725e+01 9.8964e–01
–3.1101e+02 0.0000e+00 3.1101e+02 1.0000e+00
–1.2000e+03 0.0000e+00 1.2000e+03 1.0000e+00
–1.8000e+03 0.0000e+00 1.8000e+03 1.0000e+00

Loop-at-a-Time Robustness
The loop-at-a-time robustness measures are computed by separately plotting
the transfer function that each δi sees. This corresponds to the (1,1) and (2,2)
entries of M. Recall that, if δ2 ≡ 0, then the perturbed closed-loop system is
stable for all δ1(s) satisfying

1
δ 1 ∞ < -------------------
M 11 ∞

and that there is a perturbation δ1(s), with ||δ1||∞ = 1/||M11||∞

7-38
 MIMO Margins Using µ

that causes instability. This is the robustness test for perturbations in channel
1, with no uncertainty in channel 2. Similar comments apply for the robustness
test for perturbations in channel 2, with no uncertainty in channel 1.
The results for Ma_g (solid) and Mb_g (dashed) are computed below, and shown
in Figure 7-30.
Ma_g = frsp(Ma,omega);
Mb_g = frsp(Mb,omega);
vplot('liv,m',sel(Ma_g,1,1),'-',sel(Mb_g,1,1),'--');
vplot('liv,m',sel(Ma_g,2,2),'-',sel(Mb_g,2,2),'--');

With respect to these uncertainty weights, the closed-loop system associated

with controller Ka (solid) has better loop-at-a-time margins than the closed-loop
system associated with controller Kb (dashed) since the peak is lower. Note that
this holds for both channels.
The robustness measure for simultaneous perturbations in channels 1 and 2
requires the structured singular value, µ. Recall from Chapter 4 that the
structured singular value of the matrix that the perturbation sees gives a
nonconservative bound on the allowable simultaneous perturbations.
In this example there are two, scalar, unmodelled dynamics perturbations (i.e.,
complex blocks), so the uncertainty structure is

 
 δ1 0 
∆ :=  : δ 1 ∈ C, δ 2 ∈ C 
 0 δ2 
 

The µ-Tools representation of the uncertainty set, along with the µ calculation
syntax are given as follows.

7-39
7 Robust Control Examples

Channel 1 robustness: ka (solid), kb (dashed) Channel 2 robustness: ka (solid), kb (dashed)

0.45 0.35

0.4
0.3

0.35
0.25
0.3

0.2

Magnitude
Magnitude
0.25

0.2 0.15

0.15
0.1
0.1

0.05
0.05

0 1 0 1 0 1 2 3 4
0 1 2 3 4
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

Figure 7-30: Loop-at-a-Time Robustness Tests

deltaset = [1 1;1 1];

[bnds_a,dvec_a,sens_a,pvec_a] = mu(Ma_g,deltaset);
[bnds_b,dvec_b,sens_b,pvec_b] = mu(Mb_g,deltaset);
vplot('liv,m',bnds_a,'-',bnds_b,'--');

Plots of µ∆(M(jω)) for both closed-loop systems are shown in Figure 7-31. Note
that bnds_a and bnds_b each contain upper and lower bounds for µ, hence there
are a total of four plots in Figure 7-31. In the case of two complex uncertainties,
the upper and lower bound for µ are guaranteed to be equal, so it appears that
there are only 2 plots in the figure.
The peak associated with controller ka is much larger than the peak associated
with controller kb. Again, the size of the smallest block-diagonal perturbation
which causes instability is equal to

1
--------------------------------------------
max ω µ ∆ ( M ( jω ) )

Hence, although the margins to loop-at-a-time perturbations are better with

controller ka, the margins with respect to simultaneous perturbations are
significantly better using controller kb.

7-40
 MIMO Margins Using µ

Structured singular value with: ka(solid), kb (dashed)

1.5

1
mu

0.5

0 1 0 1 2 3 4
10 10 10 10 10 10
Frequency (rad/sec)

Figure 7-31: Structured Singular Value Plot

Constructing Destabilizing Perturbations

Constructing the smallest destabilizing perturbation for a closed-loop system
is quite useful. This perturbation can provide physical insight to the types of
variations for which the closed-loop system is very sensitive. The command
dypert is used to construct the smallest destabilizing perturbation from the
output of mu. The program dypert is a block-structured version of sisorat. The
data required from the mu calculation is the perturbation vector, pvec, the block
structure, deltaset, and the bounds for µ, bnds. Based on this data, dypert
creates a real-rational, block structured perturbation that causes instability.
The ||⋅ ||∞ norm of the block diagonal perturbation is equal to 1/(peak value µ)
from the associated µ plot. From the µ plot in Figure 7-31, the peak lower bound
µ value for Ma_g is about 1.47 and for Mb_g, 0.64. Therefore, the H∞ norm of the
- ≈ 0.68 , while for Mb_g we get
1
smallest destabilizing perturbation for Ma_g is ----------
1.47
----------- ≈ 1.55 . The following commands illustrate this fact.
1
0.64

7-41
7 Robust Control Examples

perta = dypert(pvec_a,deltaset,bnds_a);
pertb = dypert(pvec_b,deltaset,bnds_b);
[pkvnorm(sel(bnds_a,1,2)) 1/pkvnorm(sel(bnds_a,1,2))]
ans =
1.4657 0.6823
[pkvnorm(sel(bnds_b,1,2)) 1/pkvnorm(sel(bnds_b,1,2))]
ans =
0.6435 1.5540
hinfnorm(perta)
norm between 0.6823 and 0.6829
achieved near 0
hinfnorm(pertb)
norm between 1.554 and 1.556
achieved near 2.182

The destabilizing perturbations, perta and pertb, have the same structure as
∆, block-diagonal. Therefore, the diagonal entries of perta and pertb will have
norms of 0.68 and 1.55, respectively, with zero off diagonal entries. The
following commands verify this fact. The output of these functions is not
shown.
hinfnorm(sel(perta,1,1))
hinfnorm(sel(perta,1,2))
hinfnorm(sel(perta,2,1))
hinfnorm(sel(perta,2,2))
hinfnorm(pertb)
hinfnorm(sel(pertb,1,1))
hinfnorm(sel(pertb,1,2))
hinfnorm(sel(pertb,2,1))
hinfnorm(sel(pertb,2,2))

7-42
 MIMO Margins Using µ

The perturbed closed-loop can be formed with starp. Each perturbation results
in a pair of imaginary axis eigenvalues at the frequency associated with the
peak (across frequency) of µ∆(M(jω)).
pertclpa = starp(perta,Ma,2,2);
pertclpb = starp(pertb,Mb,2,2);
rifd(eig(pertclpa))

real imaginary frequency damping

–1.0201e–04 0.0000e+00 1.0201e–04 l.0000e+00
1.0394e–14 –l.0000e–01 l.0000e–01 –1.0394e–13
1.0394e–14 l.0000e–01 l.0000e–01 –1.0394e–13
–1.4313e+00 0.0000e+00 1.4313e+00 l.0000e+00
–1.0808e+03 0.0000e+00 1.0808e+03 l.0000e+00
–1.9334e+03 0.0000e+00 1.9334e+03 l.0000e+00

rifd(eig(pertclpb))

real imaginary frequency damping

-1.0550e+00 0.0000e+00 1.0550e+00 l.0000e+00
-8.0315e+00 -1.1012e+01 1.3630e+01 5.8927e-01
-8.0315e+00 1.1012e+01 1.3630e+01 5.8927e-01
2.2808e-12 -3.5565e+01 3.5565e+01 -6.4132e-14
2.2808e-12 3.5565e+01 3.5565e+01 -6.4132e-14
-9.6571e+01 0.0000e+00 9.6571e+01 l.0000e+00
-3.3746e+02 -9.4865e+01 3.5054e+02 9.6268e-01
-3.3746e+02 9.4865e+01 3.5054e+02 9.6268e-01
-1.2000e+03 0.0000e+00 1.2000e+03 l.0000e+00
-1.8000e+03 0.0000e+00 1.8000e+03 l.0000e+00

7-43
7 Robust Control Examples

Space Shuttle Robustness Analysis

This section outlines a robust stability and robust performance analysis of the
Space Shuttle lateral axis flight control system during re-entry. It serves as a
general illustration of the usefulness of the real and complex µ analysis
methods.
The system is a simplified model of the Space Shuttle, in the final stages of
landing, as it transitions from supersonic to subsonic speeds. The material in
this chapter is based on the paper:
Doyle, J., K. Lenz, and A. Packard, “Design Examples Using µ Synthesis: Space
Shuttle Lateral Axis FCS During Re-entry,” NATO ASI Series, Modelling,
Robustness, and Sensitivity Reduction in Control Systems, vol. 34,
Springer-Verlag, 1987.
The analysis procedure involves several steps:

1 Build uncertain model of plant.

2 Define performance specifications and uncertainty bounds.

3 Construct open-loop interconnection.

4 Close feedback loop with controller.

5 Perform a variety of real and complex µ analysis tests on the closed-loop

system, and explore the impact of the uncertainty model (real vs. complex)
on the robust stability and robust performance requirements.

6 Construct worst-case perturbations, and see their effect on the closed-loop

system in the frequency and time domain.

7-44
 Space Shuttle Robustness Analysis

Aircraft Model: Rigid-Body

The rigid body model for the aircraft at Mach 0.9 is a four-state system, with
states

β sideslip angle (rad)

x = p = roll rate (rad/s)
r yaw rate (rad/s)
φ bank angle (rad)

An input/output block diagram of the aircraft is shown in Figure 7-32.

p  
r  
ele

ny  AC

rud

 dgust

Figure 7-32: Input-Output Description of the Aircraft

The three inputs to the aircraft are denoted by u,

θ ele (rad)
u = θ rud (rad)
d gust (ft/sec)

The first input is the actual angular deflection of the elevon surface. The second
is the actual deflection of the rudder surface. Finally, there is a lateral wind
gust disturbance input, due to the winds that occur at this altitude.
There are four output variables of the aircraft. Three of these are states, while
the fourth is the lateral acceleration at the pilot’s location, denoted ny (units of
ny are ft/sec2).

p
r
y =
ny
φ

7-45
7 Robust Control Examples

All variables in y are measured with inertial devices (gyroscopes and

accelerometers) whose individual noise characteristics are discussed later.

Aircraft Model: Aerodynamic Uncertainty

The major source of uncertainty in the aircraft model (AC) is in the aerodynamic
coefficients. These are standard aerodynamic parameters which express
incremental forces and torques generated by incremental changes in sideslip,
elevon, and rudder angles. This is a linear relationship, expressed as

side force c yβ c ya c yr β
yawing moment = c ηβ c ηa c ηr θ ele
rolling moment c lβ c la c lr θ rud

The coefficients c•• are typically estimated based on theoretical predictions,

numerical calculations, experiments in wind tunnels, and flight tests. At Mach
0.9, the shuttle is in a transonic regime involving a combination of subsonic and
supersonic flows. Theoretical, computational, and wind tunnel techniques are
inaccurate at this flight condition, so with extremely limited flight data (early
in the shuttle program), the coefficient uncertainty for the shuttle model is
unusually large.
Uncertainty in these coefficients is modeled as a nominal value, plus a
perturbation.

c yβ c ya c yr c yβ c ya c yr r yβ δ yβ r ya δ ya r yr δ yr
c ηβ c ηa c ηr = c ηβ c ηa c ηr + r ηβ δ ηβ r ηa δ ηa r ηr δ ηr
c lβ c la c lr c lβ c la c lr r lβ δ lβ r la δ la r lr δ lr

where the values of the r•• are

7-46
 Space Shuttle Robustness Analysis

r yβ r ya r yr 2.19 – 1.33 – 0.37

r ηβ r ηa r ηr = – 1.52 1.35 0.87
r lβ r la r lr – 0.72 0.52 0.24

and the perturbations δ•• are assumed to be fixed, unknown, real parameters,
with each satisfying |δ••| ≤ 1. We use the notation r••. * δ•• to denote the 3 ×
3 perturbation matrix in the model for the aero coefficients, c••.
The aircraft model acnom has the nominal aerodynamic coefficients absorbed
into the state-space data. In addition to the inputs µ and outputs y described
earlier, acnom has three fictitious inputs and outputs such that the uncertain
behavior of the aircraft AC is given by the linear fractional transformation in
Figure 7-33.
The state-space model for acnom is created by the M-file mk_acnom. A listing of
state-space model acnom is given in “Shuttle Rigid Body Model” at the end of
this section.

Figure 7-33: Uncertain Aircraft Model

7-47
7 Robust Control Examples

Actuator Models
The aircraft has two controlled inputs, rudder command, and elevon command.
Each actuator is modeled with a second order transfer function, as well as a
second order delay approximation to model the effects of the digital
implementation.
The model for the rudder is


rud
s2 +2
2
!rud
2
rud !rud s !rud +
vrud Wdelay  urud

Here, urud is the electrical command that the controller will generate to move
the rudder. The transfer function Wdelay is a second order approximation of a
delay, to model the effects of the digital implementation of the control system.
In particular

2 2
1 – 2ξ del ( s ⁄ ω del ) + ( s ⁄ ω del )
W delay ( s ) = ------------------------------------------------------------------------------
2 2
1 + 2ξ del ( s ⁄ ω del ) + ( s ⁄ ω del )

with ωdel = 173 rad/s, and ξdel = 0.866. The transfer function

2
ω rud
------------------------------------------------------------
2 2
s + 2ξ rud ω rud s + ω rud

models the physical devices (motors, inertias, etc.) involved in actually moving
the rudder. The variable θrud is the actual deflection of the rudder surface,
while urud represents the command to the rudder system. The values of the
parameters are ξrud = 0.75, ωrud =21 rads/sec.
A similar model is used for the elevon actuation system. The parameters in
that case are ξele = 0.72, ωele =14 rads/sec, with an identical second order delay
model.
The state-space models for the actuators are created by the M-file mk_act.
Since the closed-loop performance objectives include penalties on the

7-48
 Space Shuttle Robustness Analysis

deflections, rates, and accelerations of the control surfaces, the state-space

models created in mk_act each have three outputs, as shown below.

rud  ele 
rud  actrud  urud ele 
_ actele  uele
rud 
_

 ele 


Exogenous Disturbances, Noises, and Commands

There are three sources of exogenous signals:

• Wind gusts
• Sensor noise
• Pilot bank-angle command

In the H∞ framework, all time domain signals are modeled as the unit ball in
L2, filtered by problem dependent weighting functions which reflect typically
occurring signals in the application. In addition to the L2 gain, the H∞ norm
also has an interpretation in terms of gain from sinusoids to sinusoids. Now,
suppose h represents one of the exogenous signals, and Wh is the associated
stable weighting function. Then, the signal h is assumed to be any signal from
the set
h ∈ {Whηh : ||ηh||2 ≤ 1}
By choosing the form of Wh(s), the spectral content of such signals h can be
shaped.

• Lateral Wind Gusts: The set of lateral wind gusts is modeled as

 1+s⁄2 
d gust ∈  W gust η gust : W gust = 30 -------------------, η gust 2 ≤ 1 .
 1 + s 

The set on the right-hand side of the equation models the typical wind gusts
that the shuttle will encounter at this flight condition.

7-49
7 Robust Control Examples

• Sensor Noise: Each measurement is corrupted with sensor noise which

becomes more severe with increasing frequency. Since p and r are measured
with comparable gyroscopes, their sensor noise weights are identical,

1 + s/0.01
W P = W r = 0.0003 --------------------------
1 + s/0.5

These weighting functions imply a low frequency measurement error in p

and r of 0.0003 rads/sec, and a high frequency error of 0.015 rads/sec. The
model of the measured value of p, denoted pmeas, is given by
pmeas = p + Wpηp
where ηp is an arbitrary signal, with ||ηp||2 ≤ 1. This type of weighted, additive
L2 sensor noise is assumed for each of the 4 measured variables.
The measurement of φ is obtained from a navigation package at a reduced
sample rate, so its weight is chosen to be

1 + s/0.01
W φ = 0.0007 --------------------------
1 + s/2

which is relatively large in the mid-to-high frequency range. The sensor

noise weight on the ny accelerometer is

1 + s/0.05
W n = 0.25 --------------------------
y 1 + s/10

For the variables r, φ, and ny, we have

r meas = r + W r η r
φ meas = φ + W φ η φ
ny = n y + W ny ηny
meas

• Pilot Bank-Angle Command: In this problem, the pilot (or autopilot) takes
the shuttle through a series of sweeping “S” turns to slow the vehicle down.

7-50
 Space Shuttle Robustness Analysis

These require accurate tracking of a bank-angle command. Typical

bank-angle commands are modeled as
φcmd = Wφcmdηφcmd,
where ηφcmd is assumed to be an arbitrary signal with ||ηφcmd||2 ≤ 1. In this
example, the weight on the bank-angle command is chosen as

1 + s/2
W φcmd := 0.5 ----------------------
1 + s/0.5

The particular choice roughly implies that the bank-angle commands are
dominated by low frequency signals, with a maximum magnitude of
approximately 0.5 radians.
The noise weighting functions are denoted by Wnoise = diag{Wp, Wr, Wφ, W n y }
in the control block diagram.

Errors
There are several variables that are to be kept small in the face of the
exogenous signals listed in the previous section. In this context, these variables
will be considered errors.
Actuator signal levels: the angular position, angular rates, and angular
accelerations of the rudder (⋅ rud) and elevon (⋅ ele) surfaces should remain
reasonably small in the face of the exogenous signals. The signals are weighted
to give an actuator error vector of

ee 4 θ ele
·
e e· θ ele
.. ..
ee 0.005 θ ele
e act = :=
er 2 θ rud
·
e r· 0.2 θ rud
.. ..
er 0.009 θ rud

7-51
7 Robust Control Examples

This performance specification can be loosely interpreted as a requirement

that the closed-loop system should, under the excitation of the modeled
·
exogenous
..
signals, maintain θele to below 0.25 radians, θ ele to below 1 rad/
sec, θ ele to below 200 rads/sec2, and so on for the rudder variables. For
notational purposes, let Wact be the 6 × 6 constant matrix so that

θ ele
·
θ ele
..
θ ele
e act = W act
θ rud
·
θ rud
..
θ rud

• Performance variables:
- The ideal bank angle response (φideal) of the shuttle to a bank-angle
command (φcmd) is

1
φ ideal := ---------------------------------------------------------φ cmd
2
1 + 2 ⋅ ξ(s/ω ) + (s/ω )

where ω = 1.2 rad/sec, and ξ = 0.7. The bank-angle tracking error is defined
as φ – φideal.
- Turn coordination: in an ideal turn, the bank angle, and the yaw rate are
related. For this aircraft, a turn coordination error is defined as
rp := r – 0.037φ
- In a turn, it is desired that the pilot feel very little lateral acceleration,
hence, the lateral acceleration variable, ny, is an error.
These error signals are weighted by frequency dependent weights to give
a performance error vector as

7-52
 Space Shuttle Robustness Analysis

1+s
0.8 1----------------------
+ s/0.1
0 0 ny
e perf := 1+s
0 500 -------------------------
1 + s/0.01
- 0 r – 0.037φ
1+s φ – φ ideal
0 0 250 -------------------------
1 + s/0.01
-

For notational purposes, let Wperf be a 3 × 5 transfer function matrix so

that

p
r
e perf = W perf ny
φ
φ ideal

The error weight on the lateral acceleration indicates a tolerance for low
frequency accelerations of 1.25 ft/sec2, which is relaxed at high frequency,
allowing accelerations up to 12.5 ft/sec2. Again, these specifications
correspond to ny errors produced by the exogenous signal set (wind gusts,
measurement noises, and bank angle commands). Similar interpretation
is given to the other performance variables.

LFT Aero-Coefficient Uncertainty

The perturbations in the aero-coefficients can be written as an LFT (linear
fractional transformation) on a structured uncertainty matrix. Define constant
matrices WL ∈ R3×9 and WR ∈ R9×3 such that

r yβ δ yβ r ya δ ya r yr δ yr
W L ⋅ diag[δ yβ ,δ ηβ ,δ lβ ,δ ya ,…,δ lr ] ⋅ W R = r ηβ δ ηβ r ηa δ ηa r ηr δ ηr
r lβ δ lβ r la δ la r lr δ lr

for all δ••. This is easily done with the permutation matrices WL and WR shown
below.

7-53
7 Robust Control Examples

1 1 1 0 0 0 0 0 0
T
WR = 0 0 0 1 1 1 0 0 0
0 0 0 0 0 0 1 1 1

2.19 0 0 – 1.33 0 0 – 0.37 0 0

WL = 0 – 1.52 0 0 1.35 0 0 0.87 0
0 0 – 0.72 0 0 0.52 0 0 0.24

Create Open-Loop Interconnection

The aircraft model, actuator models, and weighting functions discussed in the
previous sections can be constructed from M-files.
mk_acnom;
mk_act;
mk_wts;

The open-loop interconnection structure, which includes the uncertainty model

and the performance objectives, is shown in Figure 7-34.

7-54
 Space Shuttle Robustness Analysis

--
-
pertoutf1-9g pertinf1-9g - -
eact
-- Wact
-
Wr
  Wl 
 ele actele 
acnom

 (p,r,ny ,)  rud actrud 


eperf
Wperf
 2 3  Wgust  gust
p
66 r 77
64 75 - Wnoise - ?h
ny


- noisy(p,r,ny,) elevon cmd

cmd - Wcmd - cmd rudder cmd

Ideal bank
angle response 
model

Figure 7-34: Shuttle Interconnection Structure

The M-file mk_olic uses the sysic command to create a SYSTEM matrix
description of the open-loop interconnection structure. In the workspace, the
open-loop system is denoted by olic, and has 23 states, 23 outputs, and 17
inputs.
mk_olic;
minfo(olic)

A schematic diagram, with the specific input/output ordering for olic, is shown
in Figure 7-35.

7-55
7 Robust Control Examples

pertoutf1g 

1
pertoutf2g

2
pertoutf3g
 
3
pertoutf4g
  pertinf1g
4 1
pertoutf5g
  pertinf2g
5 2
pertoutf6g
  pertinf3g
6 3
pertoutf7g
  pertinf4g
7 4
pertoutf8g
  pertinf5g
8 5
pertoutf9g
  pertinf6g
9 6
weighted ny
  pertinf7g
10 7
eperf weighted r
  pertinf8g
11 8
weighted err
 olic
 pertinf9g
12 9
p
 
weighted elevon acc 13 10
r
 
weighted elevon rate 14 11
ny
 
weighted elevon pos 15 12 exogenous
eact
 disturbances
weighted rudder acc
 
16 13
gust
weighted rudder rate
 
17 14
cmd
cmd  
weighted rudder pos 18 15

noisy p  
19 16 elevon cmd
u
noisy r 
20 17 rudder cmd
y
noisy ny 
21

noisy  
22
23

Figure 7-35: Schematic Diagram of Space Shuttle olic

7-56
 Space Shuttle Robustness Analysis

Controllers
In this section, the robustness properties of three different controllers are
analyzed using µ. The controllers receive four sensor measurements along with
the φ command signal and produce two control signals for the elevon and
rudder commands. The controller block diagram is shown below.

 cmd
elevon cmd   noisy p
 noisy r
rudder cmd  
K
noisy ny
 noisy 

In this example, each controller has different characteristics:

• k_h is designed to optimize H∞ performance, under the assumption that

there is no model uncertainty;
• k_mu is designed with the D – K iteration approach to µ -synthesis;
• k_x is constructed to be a tradeoff between the two controllers

The µ-Tools commands to design the H∞ optimal controller, k_h, are:

olic_h = sel(olic,[10:23],[10:17]);
minfo(olic_h)
k_h = hinfsyn(olic_h,5,2,0,5,0.1);

The first command, sel, removes the aero-coefficient uncertainty channels,

leaving only the exogenous signals and errors, and feedback signals. The third
command, hinfsyn, designs a suboptimal H∞ controller for the open-loop
system olic_h. This controller measures five signals, and generates two
control signals.
It is simple to check some characteristics of the controller and the closed-loop
system
minfo(k_h)
clp_h = starp(olic,k_h,5,2);
rifd(spoles(clp_h))
rifd(spoles(k_h))

7-57
7 Robust Control Examples

The two other controllers have already been designed and stored in the file
shutcont.mat.
load shutcont
minfo(k_x)
minfo(k_mu)

Nominal Frequency Responses

The closed-loop system is constructed using the star product command starp.

In the closed-loop system, there are six exogenous signals (the six η signals:
four sensor noises, wind gust, bank angle command) and nine errors (weighted
performance error vector and the weighted actuator error vector). The nominal
performance objective is that this multivariable transfer function matrix
should have an H∞ norm less than 1. Using µ-Tools, it is easy to evaluate this
performance criterion. Simply form the closed-loop system, calculate its
frequency response, and plot the norm of the appropriate transfer function
versus frequency.

7-58
 Space Shuttle Robustness Analysis

omega = logspace(-2,3,30);
clp_h = starp(olic,k_h,5,2);
clp_hg = frsp(clp_h,omega);
clp_x = starp(olic,k_x,5,2);
minfo(clp_x)
clp_xg = frsp(clp_x,omega);
minfo(clp_xg)
clp_mu = starp(olic,k_mu,5,2);
clp_mug = frsp(clp_mu,omega);

Note that the closed-loop systems have additional inputs and outputs from the
nine aero-perturbation channels. The relevant exogenous signals and errors
are selected (using sel) before calculating the maximum singular value
(vnorm).
np_hg = sel(clp_hg,[10:18],[10:15]);
np_xg = sel(clp_xg,[10:18],[10:15]);
np_mug = sel(clp_mug,[10:18],[10:15]);
vplot('liv,m',vnorm(sel(clp_hg,10:18,10:15)),...
vnorm(sel(clp_xg,10:18,10:15)),...
vnorm(sel(clp_mug,10:18,10:15)))
title('NOMINAL PERFORMANCE: ALL CONTROLLERS')

NOMINAL PERFORMANCE: ALL CONTROLLERS

1

k_h − SOLID
0.9
k_x − DASHED
k_mu − DOTTED
0.8

0.7
H−INFINITY NORM

0.6

0.5

0.4

0.3

0.2

0.1

0
−2 −1 0 1 2 3
10 10 10 10 10 10
FREQUENCY (RAD/SEC)

Figure 7-36: Nominal Performance of k_h, k_x, and k_mu

7-59
7 Robust Control Examples

Note that the best nominal performance is achieved by controller k_h, as seen
in Figure 7-36. This is not surprising, since it was designed specifically with
these disturbances and errors in mind. Relatively, the performance from k_mu
is poor, though it does meet the nominal performance objective. In later
calculations, it will become clear that the degradation in nominal performance
is offset by a much greater insensitivity to variations in the aerodynamic
coefficients.

Robust Stability
Using µ, the robust stability characteristics of each closed-loop system can be
evaluated. The uncertain parameters (δyβ,. . .,δlr) can be assumed to be real,
representing uncertainty in the constant aerodynamic coefficients. However,
the flow around the vehicle is very complex, and the quasi-steady implication
of constant aerodynamic coefficients is somewhat simplistic. Consequently, for
a more conservative analysis, the uncertain parameters can be treated as
complex. In this section, both models of uncertainty will be analyzed, and
compared. Refer to Chapter 4, “Modeling and Analysis of Uncertain Systems”
for more detail on the interpretations.
This motivates two separate representations of the uncertainty set,

∆ C := { diag [ δ 1 ,δ 2 ,…,δ 9 ] : δ i ∈ C }
∆ R := { diag [ δ 1 ,δ 2 ,…,δ 9 ] : δ i ∈ R }

where the perturbations represent uncertainty in the aero-coefficients. Note

that the first set contains all complex perturbations, while the second set
includes only real perturbations.
In the µ-Tools syntax, these are represented as
delsetrs_C = ones(9,2);
delsetrs_R = [-ones(9,1) zeros(9,1)];

Here, the lower case rs refers to robust stability (as opposed to robust
performance, rp, which will be addressed later).
The perturbation inputs/outputs from the frequency responses are selected for
a robust stability µ test. The input/output channels associated with the
performance criterion are not used in the robust stability µ test. A diagram of
the closed-loop system is shown in Figure 7-37.

7-60
 Space Shuttle Robustness Analysis

pertoutf1g   pertinf1g
pertoutf2g  1 1
 pertinf2g
 
2 2
pertoutf3g pertinf3g
 
3 3
pertoutf4g pertinf4g
 
4 4
pertoutf5g clp RS pertinf5g
 
5 5
pertoutf6g pertinf6g
 
6 6
pertoutf7g pertinf7g
 
7 7
pertoutf8g pertinf8g
 
8 8
pertoutf9g 9 9 pertinf9g

Figure 7-37: Schematic Design of clp_RS

clp_hgRS = sel(clp_hg,1:9,1:9);
clp_xgRS = sel(clp_xg,1:9,1:9);
clp_mugRS = sel(clp_mug,1:9,1:9);

Calculate µ across frequency, and look at µ plots. Start with the complex
uncertainty structure.
[bnds_h,dv_h,sens_h,rp_h]=mu(clp_hgRS,delsetrs_C);
[bnds_x,dv_x,sens_x,rp_x]=mu(clp_xgRS,delsetrs_C);
[bnds_mu,dv_mu,sens_mu,rp_mu]=mu(clp_mugRS,delsetrs_C);
vplot('liv,d',bnds_h,'-',bnds_x,'--',bnds_mu,'-.')
title('ROBUST STABILITY OF CLOSED-LOOP: COMPLEX')

7-61
7 Robust Control Examples

ROBUST STABILITY OF CLOSED-LOOP: COMPLEX

3.5
k_h - SOLID
k_x - DASHED
3
k_mu - DOTTED

2.5

2
MU

1.5

1 k_h (solid line)

k_x (dashed line)
0.5 k_mu (dotted line)

0
10 -2 10 -1 10 0 10 1 10 2

FREQUENCY (RAD/SEC)

Figure 7-38: Complex Robust Stability µ Analysis of k_h, k_x, and k_mu

According to Figure 7-38, the k_mu controller has the best robust stability
properties when the perturbations are treated as complex (dynamic). The peak
of the lower bound, 0.9, implies that there is a diagonal complex perturbation
1
of size, -------
0.9
- , that causes instability. The peak of the upper bound, approximately
1
0.99, implies that for diagonal perturbations smaller than ----------
0.99
- , the closed-loop
system remains stable. The gap between the upper and lower bound can be
reduced by using the “c” option in the mu command. Without this option, the
upper bound from mu is a computational approximation to

–1
inf σ ( DMD )
D∈D

that can be refined (option “c”) at the expense of slower execution. Using the “c”
option reduces the upper bound peak to 0.9, so that the complex µ analysis
gives a tight estimate on the size of the smallest destabilizing perturbation.
Similar interpretations are possible for the closed-loop systems with
controllers k_h and k_x, though, since the µ plots have larger peaks, the bound
on allowable perturbations is smaller. Hence, the closed-loop system with the

7-62
 Space Shuttle Robustness Analysis

controller k_mu achieves robust stability to complex perturbations, whereas the

other controllers do not.
The closed-loop system is also analyzed treating the aerodynamic uncertainty
as real perturbations with the µ-Tools mu command.
[rbnds_h,rrp_h] = mu(clp_hgRS,delsetrs_R);
[rbnds_x,rrp_x] = mu(clp_xgRS,delsetrs_R);
[rbnds_mu,rrp_mu] = mu(clp_mugRS,delsetrs_R);
vplot('liv,d',rbnds_h,'-',rbnds_x,'--',rbnds_mu,'-.')
title('ROBUST STABILITY OF CLOSED-LOOP: REAL')

ROBUST STABILITY OF CLOSED-LOOP: REAL

1.4

k_h - SOLID
1.2 k_x - DASHED
k_mu - DOTTED
1

0.8
MU

0.6

0.4
k_h (solid line)
0.2 k_x (dashed line)
k_mu (dotted line)
0
10 -2 10 -1 10 0 10 1 10 2

FREQUENCY (RAD/SEC)

Figure 7-39: Real Robust Stability µ Analysis of k_h, k_x, and k_mu

The k_h controller has the best robust stability properties, when the
perturbations are treated as real, as seen in Figure 7-39. This is in contrast to
the robust stability analysis with complex perturbation where k_mu exhibited
the best properties. The peak of the upper bound, approximately 0.66, implies
1
that for diagonal, real perturbations smaller than ----------
0.66
- , the closed-loop system
remains stable. The lower bound in Figure 7-39 is often 0 and does not converge
for all values of frequency, leading to a large gap between the upper and lower
bound. This gap can be reduced by adding a small amount of complex
perturbation to the pure real perturbation. A detailed discussion of this

7-63
7 Robust Control Examples

approach can be found in the “Mixed Real/Complex Structured Singular Value”

section in Chapter 4.

Robust Performance
Using µ, the robust performance characteristics of each closed-loop system can
be evaluated. The uncertain parameters are treated as real parameters in this
analysis. These parameters can also be treated as complex perturbations,
though this is not done in this section.
The appropriate block structure for the robust performance test is
∆P := {diag[δ1,δ2,. . .,δ9,∆10] : δi ∈ R, ∆10 ∈ C6×9}
which is simply an augmentation of the original real robust stability
uncertainty set, ∆R, with a complex 6 × 9 full block to include the performance
objectives. Recall from the “Using m to Analyze Robust Performance” section
in Chapter 4: H∞ performance objectives are always represented with a full,
complex block. Hence,
delsetrp_R = [delsetrs_R;6 9]

The µ calculations are performed on the entire 18 × 15 closed-loop matrix,

which includes the perturbation channels and the exogenous signals and
errors. The command mu is called with both real and complex blocks.
[bnds_h,ph] = mu(clp_hg,delsetrp_R);
[bnds_x,px] = mu(clp_xg,delsetrp_R);
[bnds_mu,p_mu] = mu(clp_mug,delsetrp_R);
vplot('liv,d',bnds_h,'-',bnds_x,'--',bnds_mu,'-.')
title('ROBUST PERFORMANCE OF ALL CONTROLLERS')

7-64
 Space Shuttle Robustness Analysis

ROBUST PERFORMANCE OF ALL CONTROLLERS

3
k_h - SOLID
k_x - DASHED
2.5
k_mu - DOTTED

MU 2

1.5

1
k_h (solid line)
k_x (dashed line)
0.5
k_mu (dotted line)

0
10 -2 10 -1 10 0 10 1 10 2

FREQUENCY (RAD/SEC)

Figure 7-40: Robust Performance µ Plots of k_h, k_x, and k_mu

The axis is selected in Figure 7-40 to show a comparison of controllers k_x and
k_mu. At low frequency, the closed-loop robust performance with k_h
implemented gets as bad as 14. The closed-loop system using controller k_x
achieves a robust performance µ value of 1.56, while controller k_mu achieves a
robust performance µ value of 1.22.

Worst-case Perturbations
Using a µ calculation, we have seen that all controllers achieve robust-stability
to the 9 × 9 real uncertainty matrix which represents uncertainty in the
aero-coefficients. However, the performance of each closed-loop system
degrades differently under LFT real, diagonal perturbations. We use wcperf to
compute the worst-case performance degradation as well as the worst-case,
norm 1, perturbation. The worst-case perturbation of norm 1 will be used in the
next section for uncertain time-domain simulations.
[deltabadh,wcp_lowh,wcp_upph] = wcperf(clp_hg,delsetrs_R,.05,4);
[deltabadx,wcp_lowh,wcp_upph] = wcperf(clp_xg,delsetrs_R,.05,10);
[deltabadmu,wcp_lowmu,wcp_uppmu]= wcperf(clp_mug,delsetrs_R,.05,10);
vplot(wcp_lowh,wcp_upph,wcp_lowx,wcp_uppx,wcp_lowmu,wcp_uppmu)

7-65
7 Robust Control Examples

Worst-Case Performance
2

1.8 k_h k_x

1.6 k_mu

1.4

1.2

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1 1.2
Size of Uncertainty

Figure 7-41: Performance Degradation of Closed-Loop

Using k_h, it is clear that the closed-loop performance degrades rapidly and
severely. It would not be an acceptable controller in the real aircraft.

Note Optimizing the H∞ norm of some closed-loop transfer function does not,
in any way, guarantee robustness to perturbations at other points in the
feedback loop.

Using k_x and k_mu, reasonable robustness properties (on the order of the
original specifications) are attained. The controller k_x achieves better
nominal performance (i.e., at ||∆|| = 0), at the expense of more rapid potential
performance degradation under uncertainty. Both closed-loop systems
potentially degrade to unacceptable (performance norm > 1) performance with
less than one-half of the original modeled uncertainty. At that level of
uncertainty, the closed-loop system with k_mu degrades more gracefully. This
type of tradeoff curve illustrates some of the differences between the two
controllers, and can be helpful in understanding the tradeoffs involved.

7-66
 Space Shuttle Robustness Analysis

Time Simulations
The open-loop simulation interconnection, Figure 7-42, is similar to olic, but
contains none of the weighting functions. It is used exclusively for nominal and
perturbed time-domain simulations, where unweighted time signals will be
calculated and plotted.
mk_olsim;
minfo(olsim)

f1g 
f2g 
pertout 1

f3g 
pertout 2

f4g 
pertout 3

f5g 
pertout 4

f6g 
pertout 5
pertout
f7g 
6
 f g
 pertin 1
1
pertout
f8g 
7
f g
 pertin 2
2
pertout
f9g 
8
f g
 pertin 3
3
pertout

9
f g
 pertin 4
4


10
f g
 pertin 5
5
ny
r , c 
11
f g
 pertin 6
6

 , ideal 
12
f g
 pertin 7
olsim 7
f g
ele 
13

 pertin 8
8
f g
ele 
 14

 pertin 9
9

ele 
_ 15


10 dgust

rud 
16

 cmd
11

rud 
 17

 elevoncmd
12

rud 
_ 18
13 ruddercmd

cmd 
19

p
20

p
21


22

ny 
23
24

Figure 7-42: Open-Loop Simulation Model

For the purposes of this exercise, the four sensor noises have been eliminated
from the simulation model. It is easy to modify mk_olsim.m to include these if
desired.
The LFT time simulation GUI, simgui, is used to simulate the nominal and
perturbed time response of the three controllers. For more details on simgui

7-67
7 Robust Control Examples

see the “LFT Time Simulation User Interface Tool: simgui” section in Chapter
6.
The main performance objective is bank angle tracking, so the response to a 0.5
radian step input for φcmd is investigated. The gust input is set to zero in these
simulations. This data is entered into the simgui Main window Input Signal
editable text. Note that φcmd is the 11th input of olsim and the second
non-perturbation input. The output signals of interest are φ, ny, r – cφ, and φ –
φideal, which are outputs 10 through 13 of olsim or the first through fourth
outputs after the perturbation has been included. In the simgui Main window,
input olsim into the Plant editable text. Figure 7-43 shows the main
simulation window for the nominal and perturbed response of controller k_x.

Figure 7-43: simgui Main Window for Shuttle Time Simulation

7-68
 Space Shuttle Robustness Analysis

We are interested in the nominal and perturbed closed-loop response in the

presence of a worst-case, real perturbation. This corresponds to the top nine
channels of olsim, the aero-coefficient perturbations. Therefore, worst-case
real perturbations of size 1 in the aero-coefficients for k_x and k_mu are
calculated using the wcperf command. These perturbations are:

badpertx = diag([ 1 1 1 –.87 1 1 –.28 –1 –1 ])

badpertmu = diag([ –1 –1 –1 1 1 1 –1 –1 –1 ])

badpertx is used in the perturbed response for controller k_x and badpertmu is
used as the worst-case real perturbation for controllers k_h and k_mu. This data
is input into the Perturbation editable text in the simgui main window.
The controllers are implemented in discrete-time at a sample-rate of 20Hz on
the shuttle. To replicate the same implementation, a sample-data time
simulation is performed. This simulation is available under the Options menu
in the Main simulation window. Therefore, the continuous-time controllers,
k_x, k_h, and k_mu, must be discretized for the sampled-data time simulation.
The continuous-time plant, olsim, is simulated at 200Hz and the controllers at
20Hz as seen in the simgui parameter window, Figure 7-44.

Figure 7-44: simgui Parameter Window for Shuttle Time Simulation

7-69
7 Robust Control Examples

The nominal and perturbed closed-loop responses with k_x, k_h, and k_mu
implemented are shown in Figure 7-45 and 7-46. As expected, the time domain
simulations reinforce the conclusions that were reached in the frequency
domain analysis. The nominal performance associated with k_h is superb, but
degrades significantly with the aerodynamic uncertainty. In that respect, the
controller k_mu performs the best, nearly achieving all of the robust
performance objectives. The nominal and perturbed time response of other
performance variables can also be easily investigated.

7-70
 Space Shuttle Robustness Analysis

nominal (solid line) perturbed (dashed line)

Figure 7-45: Closed-Loop Nominal and Perturbed Time Response, k_x (top)
and k_h

7-71
7 Robust Control Examples

Nominal (solid line) Perturbed (dashed line)

Figure 7-46: Closed-Loop Nominal and Perturbed Time Response, k_mu

Conclusions
This exercise illustrated the use of the µ-Tools software to analyze the robust
stability and robust performance objectives on a complicated, uncertain plant
model.
There is an important feature of the mu software that cannot be overlooked or
overemphasized. These algorithms calculate both upper and lower bounds for
µ, and produce worst-case perturbations which provide the lower bound. The
perturbations, and their effects, can be analyzed in both the frequency domain
and time domain. In practice, the bad perturbations are also used in high
fidelity, nonlinear simulations of the closed-loop system to discover limitations
and unforeseen problems.

7-72
 Space Shuttle Robustness Analysis

Although this problem did not have repeated uncertain parameters (each δ••
appeared only once), the algorithms and software do handle these cases, and
the reader is referred back to the “Complex Structured Singular Value” section
in Chapter 4 for details.

Space Shuttle References

Doyle, J., K. Lenz, and A. Packard, “Design Examples Using µ Synthesis: Space
Shuttle Lateral Axis FCS During Reentry,” NATO ASI Series, Modelling,
Robustness, and Sensitivity Reduction in Control Systems, vol. 34,
Springer-Verlag, 1987.

Shuttle Rigid Body Model

The perturbed, state-space rigid body model of the aircraft, acnom, is shown in
the following figure.

7-73
7 Robust Control Examples

acnom = A_acnom B_acnom

C_acnom D_acnom

– 9.5e – 2 1.4e – 1 – 9.9e – 1 3.6e – 2

A_acnom =
– 3.6e + 0 – 4.3e – 1 2.8e – 1 0
4.0e – 1 – 1.3e – 2 – 8.1e – 2 0
0 1 – 1.4e – 1 0

1.3e – 2 0 0 – 1.2e – 2 1.0e – 2 – 1.1e – 7

B_acnom =
0 – 3.1e – 2 – 3.1e + 0 6.6e + 0 1.3e + 0 – 4.1e – 6
0 – 1.9e – 1 – 6.4e – 2 3.8e – 1 – 2.6e – 1 4.5e – 7
0 0 0 0 0 0

1 0 0 0
0 0 0 0
0 0 0 0
C_acnom = 0 1 0 0
0 0 1 0
– 6.8e + 1 – 1.7e + 0 – 4.1e + 0 – 3.7e – 5
0 0 0 1.0e + 0
0 0 0 0 0 1.2e – 6
0 0 0 1 0 0
0 0 0 0 1 0
D_acnom = 0 0 0 0 0 0
0 0 0 0 0 0
1.1e + 1 – 1.1e + 1 – 1.1e + 1 2.7e + 1 – 3.0e + 0 – 7.8e – 5
0 0 0 0 0 0

7-74
 HIMAT Robust Performance Design Example

HIMAT Robust Performance Design Example

This section contains an idealized example of µ-synthesis as applied to the
design of a pitch axis controller of an experimental highly maneuverable
airplane, HIMAT. The airplane model is taken from aerodynamic data for the
HIMAT vehicle. The problem is posed as a robust performance problem, with
multiplicative plant uncertainty at the plant input and plant output weighted
sensitivity function as the performance criterion. The design procedure
presented in this section involves several steps:

1 Specification of closed-loop feedback structure.

2 Specification of model uncertainty and performance objectives in terms of

frequency-dependent weighting matrices.

3 Construction of open-loop interconnection for control synthesis routines.

4 Loop shape controller design for the open-loop interconnection.

5 H∞ optimal controller design for the open-loop interconnection.

6 Analysis of robust performance properties of the resulting closed-loop
systems using the structured singular value, µ (µ-analysis).

7 Use of frequency dependent similarity scalings, obtained in the µ-analysis

step, to scale the open-loop interconnection, and redesign H∞ controller
(iterating on steps 5, 6, and 7 constitutes the approach to µ-synthesis called
“D – K iteration,” which is described in detail in Chapter 5).

The main objective of this section is to illustrate µ-synthesis design methods

(steps 1, 2, 3, 5, 6, 7). The loop shape controller (step 4) is included to illustrate
that robust stability and nominal performance do not necessarily imply robust
performance.
Many of the command outputs are not displayed in the text, since it is assumed
that the reader is simultaneously working through the example on a computer.

7-75
7 Robust Control Examples

HIMAT Vehicle Model and Control Objectives

The HIMAT vehicle model and control objectives are taken from a paper by
Safonov et al. (1981). The interested reader should consult this paper as well
as Hartman et al. (1979) and Merkel and Whitmoyer (1976) for more details.
The HIMAT vehicle was a scaled, remotely piloted vehicle (RPV) version of an
advanced fighter, which was flight tested in the late 1970s. The actual HIMAT
vehicle is currently on display in the Smithsonian National Aerospace Museum
in Washington, D.C. The design example will consider only the longitudinal
dynamics of the airplane. These dynamics are assumed to be uncoupled from
the lateral-directional dynamics. Linearized models for a collection of flight
conditions can be found in [HartBG]. The state vector consists of the vehicle’s
basic rigid body variables.
xT = (δv, α, q, θ)
representing the forward velocity, angle-of-attack, pitch rate, and pitch angle,
respectively. The flight path angle (γ) is defined as γ = θ – α. The state variables
used to describe motions in the vertical plane are given below.
δv — perturbations along the velocity vector
α — angle between velocity vector and aircraft's longitudinal axis
q — rate-of-change of aircraft attitude angle
θ — aircraft attitude angle
The control inputs are the elevon (δe) and the canard (δc). The variables to be
measured are α and θ.
There are three longitudinal maneuvers to be considered.
Vertical Translation: Control the vertical velocity at a constant θ (α varies).
This implies that the attitude is held constant as the velocity vector rotates.
Pitch Pointing: Control the attitude at a constant flight path angle (i.e., θ – α
= constant). In this case the velocity vector does not rotate.
Direct Lift: Control the flight path angle at constant angle-of-attack (i.e.,
γ = θ). This maneuver produces a normal acceleration response without
changing the angle-of-attack.
These control objectives are accounted for within the performance
specification.

7-76
 HIMAT Robust Performance Design Example

Closed-Loop Feedback Structure

A diagram for the closed-loop system, which includes the feedback structure of
the plant and controller, and elements associated with the uncertainty models
and performance objectives, is shown inFigure 7-47.

Figure 7-47: HIMAT Closed-Loop Interconnection Structure

The dashed box represents the true airplane, with associated transfer function
G. Inside the box is the nominal model of the airplane dynamics, Gnom, and two
elements, wdel and ∆G, which parametrize the uncertainty in the model. This
type of uncertainty is called multiplicative uncertainty at the plant input, for
obvious reasons. The transfer function wdel is assumed known, and reflects the
amount of uncertainty in the model. The transfer function ∆G is assumed to be
stable and unknown, except for the norm condition, ||∆G||∞ < 1. The performance
objective is that the transfer function from d to e be small, in the ||⋅ ||∞ sense, for
all possible uncertainty transfer functions ∆G. The weighting function WP is
used to reflect the relative importance of various frequency ranges for which
performance is desired.
The control design objective is to design a stabilizing controller K such that for
all stable perturbations ∆G(s), with ||∆G||∞ < 1, the perturbed closed-loop system
remains stable, and the perturbed weighted sensitivity transfer function,
S(∆G) := WP(I + P(I + ∆GWdel)K)–1
has ||S(∆G)||∞ < 1 for all such perturbations. Recall that these mathematical
objectives exactly fit in the structured singular value framework.

7-77
7 Robust Control Examples

Uncertainty Models
The airplane model we consider has two inputs: elevon command (δe) and
canard command (δc); and two measured outputs: angle-of-attack (α) and pitch
angle (θ).
A first principles set of uncertainties about the aircraft model would include:

• Uncertainty in the canard and the elevon actuators. The electrical signals
that command deflections in these surfaces must be converted to actual
mechanical deflections by the electronics and hydraulics of the actuators.
This is not done perfectly in the actual system, unlike the nominal model.
• Uncertainty in the forces and moments generated on the aircraft, due to
specific deflections of the canard and elevon. As a first approximation, this
arises from the uncertainties in the aerodynamic coefficients, which vary
with flight conditions, as well as uncertainty in the exact geometry of the
airplane. An even more detailed view is that surface deflections generate the
forces and moments by changing the flow around the vehicle in very complex
ways. Thus there are uncertainties in the force and moment generation that
go beyond the quasi-steady uncertainties implied by uncertain aerodynamic
coefficients.
• Uncertainty in the linear and angular accelerations produced by the
aerodynamically generated forces and moments. This arises from the
uncertainty in the various inertial parameters of the airplane, in addition to
neglected dynamics such as fuel slosh and airframe flexibility.
• Other forms of uncertainty that are less well understood.

In this example, we choose not to model the uncertainty in this detailed

manner, but rather to lump all of these effects together into one full-block
uncertainty at the input of a four-state, nominal model of the aircraft rigid
body. This nominal model has no (i.e., perfect) actuators and only quasi-steady
dynamics. The nominal model for the airplane is loaded from the mutools/subs
directory.

7-78
 HIMAT Robust Performance Design Example

The simple model of the airplane has four states: forward speed (v),
angle-of-attack (α), pitch rate (q) and pitch angle (θ); two inputs: elevon
command (δe) and canard command (δc); and two measured outputs:
angle-of-attack (α) and pitch angle (θ).
mkhimat;
minfo(himat)
seesys(himat,'%9.le')

-2.3e-02 -3.7e+01 -1.9e+01 -3.2e+01 | 0.0e+00 0.0e+00

0.0e+00 -1.9e+00 9.8e-01 0.0e+00 | -4.1e-01 0.0e+00
1.2e-02 -1.2e+01 -2.6e+00 0.0e+00 | -7.8e+01 2.2e+01
0.0e+00 0.0e+00 1.0e+00 0.0e+00 | 0.0e+00 0.0e+00
------------------------------------ | -----------------
0.0e+00 5.7e+01 0.0e+00 0.0e+00 | 0.0e+00 0.0e+00
0.0e+00 0.0e+00 0.0e+00 5.7e+01 | 0.0e+00 0.0e+00

The partitioned matrix represents the [A B; C D] state space data. Given this
nominal model himat (i.e., Gnom(s)) we also specify a stable, 2 × 2 transfer
matrix Wdel(s), called the uncertainty weight. These two transfer matrices
parametrize an entire set of plants, , which must be suitably controlled by the
robust controller K.
:= {Gnom(I + ∆GWdel) : ∆G stable, ||∆G||∞ ≤ 1}.
All of the uncertainty in modeling the airplane is captured in the normalized,
unknown transfer function ∆G. The unknown transfer function ∆G(s) is used to
parametrize the potential differences between the nominal model Gnom(s), and
the actual behavior of the real airplane, denoted by G. The dependence on
frequency of the uncertainty weight indicates that the level of uncertainty in
the airplane’s behavior depends on frequency.
In this example, the uncertainty weight Wdel is of the form Wdel(s) := wdel(s)I2,
for a given scalar valued function wdel(s). The fact that the uncertainty weight
is diagonal, with equal diagonal entries, indicates that the modeled
uncertainty is in some sense a round ball about the nominal model Gnom. The
scalar weight associated with the multiplicative input uncertainty is

7-79
7 Robust Control Examples

constructed using the command nd2sys. The weight chosen for this problem is
( s + 100 )
w del = 50
------------------------------- .
s + 10000
wdel = nd2sys([1 100],[1 10000],50);

The set of plants that are represented by this uncertainty weight is

 50 ( s + 100 ) 
:=  G nom  I 2 + ------------------------------∆ G ( s ) : ∆ G ( s ) stable, ∆ G ∞ ≤ 1 
 s + 10000 
 

The weighting function is used to normalize the size of the unknown

perturbation ∆G. At any frequency ω, |ωdel(jω)| can be interpreted as the
percentage of uncertainty in the model at that frequency.
omega = logspace(0,5,50);
wdel_g = frsp(wdel,omega);
vplot('liv,lm',wdel_g)
title('Multiplicative Uncertainty Weighting Function')
xlabel('Frequency (rad/s)')

The particular uncertainty weight chosen for this problem indicates that at low
frequency, there is potentially a 50% modeling error, and at a frequency of 173
rad/sec, the uncertainty in the model is up to 100%, and could get larger at
higher and higher frequencies. A frequency response of wdel is shown in
Figure 7-48.

Specifications of Closed-Loop Performance

The performance of the closed loop system will be evaluated using the output
sensitivity transfer function, (I + GK)–1. Good performance will be
characterized in terms of a weighted H∞ norm on this transfer function. Given
a 2 × 2 stable, rational transfer matrix WP, we say that nominal performance
is achieved if ||WP(I + GK)–1||∞ < 1. As in the uncertainty modeling, the
weighting function Wp is used to normalize specifications, in this case, to define
performance as whether a particular norm is less than 1.
In this problem, we choose a simple weight of the form WP(s) = wp(s)I2, where
0.5 ( s + 3 )
w p ( s ) = -------------------------
s + 0.03
-.

7-80
 HIMAT Robust Performance Design Example

Multiplicative Uncertainty Weighting Function

10 2

10 1

10 0

10 -1
10 0 10 1 10 2 10 3 10 4 10 5

Frequency (rad/s)

Figure 7-48: HIMAT Multiplicative Uncertainty Weighting Function

wp = nd2sys([1 3],[1 0.03],0.5);

For performance to be achieved, ||WP(I + GK)–1||∞ < 1, and since WP is a scalar

(times a 2 × 2 identity), the maximum singular value plot of the sensitivity
transfer function (I + GK)–1 must lie below the plot of 1
----------
wp
at every frequency.

That is, ||WP(I + GK)–1||∞ < 1, if and only if at all frequencies,

–1
σ [ ( I + GK ) ( jω ) ] < 1 ⁄ w p ( jω ) .

omega = logspace(-3,2,50);
wp_g = frsp(wp,omega);
vplot('liv,lm',minv(wp_g))
title('Inverse of Performance Weighting function')
xlabel('Frequency (rad/s)')

This sensitivity weight indicates that at low frequency, the closed-loop (both
nominal and perturbed) should reject disturbances at the output by a factor of
50-to–1. Expressed differently, steady-state tracking errors in both channels,
due to reference step-inputs in either channel should be on the order of 0.02 or
smaller. This performance requirement gets less and less stringent at higher
and higher frequencies. The closed-loop system should perform better than

7-81
7 Robust Control Examples

open-loop for frequencies up to 1.73 radians/second, and for higher frequencies,

the closed-loop performance should degrade gracefully, always lying
1
underneath the inverse of the weight, wp. The frequency response of -------
wp
is
shown in Figure 7-49.

Inverse of Performance Weighting function

10 1

10 0

10 -1

10 -2
10 -3 10 -2 10 -1 10 0 10 1 10 2

Frequency (rad/s)

Figure 7-49: Inverse of the HIMAT Performance Weight

The 2 × 2 weighting matrices in the interconnection involve the scalar functions

we have discussed, and identity matrices of dimension 2. We can build these
matrices using the command daug, which stands for diagonal augmentation.
Each new weight has two states, two inputs and two outputs as one can see
using minfo.
wdel = daug(wdel,wdel);
wp = daug(wp,wp);
minfo(wdel)
minfo(wp)

The engineering motivation for a performance specification like this might

come from the desire to have independent tracking of the angle of attack and
pitch angle. This allows the vehicle to be pointed in pitch independently from
vertical motions. We would expect this to be difficult to achieve, given that it is
obviously easier for the vehicle to simultaneously pitch up and accelerate up
than it is to simultaneously pitch down and accelerate up.

7-82
 HIMAT Robust Performance Design Example

Robust Stability, Nominal Performance, Robust Performance

The phrases robust stability, nominal performance, and robust performance
are used in this framework extensively.

Nominal Performance. The closed-loop system achieves nominal performance if

the performance objective is satisfied for the nominal plant model, Gnom.
In this problem, that is equivalent to:
Nominal Performance ⇔ ||WP(I + GnomK)–1||∞ < 1

Robust Stability. The closed-loop system achieves robust stability if the closed
loop system is internally stable for all of the possible plant models G ∈ .
In this problem, that is equivalent to a simple norm test on a particular
nominal closed-loop transfer function.
Robust Stability ⇔ ||WdelKGnom(I + KGnom)–1||∞ < 1

Robust Performance. The closed-loop system achieves robust performance if the

closed-loop system is internally stable for all G ∈ , and in addition to that, the
performance objective,
||WP(I + GK)–1||∞ < 1,
is satisfied for every G ∈ . The property of robust performance is equivalent
to a structured singular value test (a generalization of the two H∞ norm tests
in the previous conditions) on a particular, nominal closed- loop transfer
function. This is discussed further in Chapter 4, “Modeling and Analysis of
Uncertain Systems”.

Building the Open-Loop Interconnection

The command sysic is used to construct the open-loop interconnection. We will
often refer to this open-loop system as the generalized plant. In this particular
example, we store the system in the MATLAB variable himat_ic. The
command sysic will build any specified interconnection of smaller subsystems,
provided the correct information about the interconnection is in the MATLAB
workspace.

7-83
7 Robust Control Examples

A six-input, six-output SYSTEM matrix, himat_ic, (also referred to as P)

z   pertin
e  himat ic  dist
y   control

has internal structure shown in Figure 7-50. The variables control, pertin,
dist, and y are two element vectors.

Figure 7-50: HIMAT Open-Loop Interconnection Structure

This can be produced with nine MATLAB commands, listed below. The first
eight lines describe the various aspects of the interconnection, and may appear
in any order. The last command, sysic, produces the final interconnection. The
commands can be placed in an M-file, or executed at the command line.
systemnames = ' himat wp wdel ';
inputvar = '[ pertin{2} ; dist{2} ; control{2} ]';
outputvar = '[ wdel ; wp ; himat + dist ]';
input_to_himat = '[ control + pertin ]';
input_to_wdel = '[ control ]';
input_to_wp = '[ himat + dist ]';
sysoutname = 'himat_ic';
cleanupsysic = 'yes';
sysic;

7-84
 HIMAT Robust Performance Design Example

Since the system himat_ic is still open-loop, its poles are simply the poles of
the various components that make up the interconnection.
minfo(himat_ic)
spoles(himat_ic)
spoles(himat)
spoles(wdel)
spoles(wp)

µ-synthesis and D – K Iteration

For notational purposes, let P(s) denote the transfer function of the six-input,
six-output open-loop interconnection, himat_ic. Define a block structure ∆ as

 
 ∆1 0 2×2 2 × 2 4×4
∆:=  : ∆1 ∈ C ,∆ 2 ∈ C ⊂C .
 0 ∆2 
 

The first block of this structured set corresponds to the full-block uncertainty
∆G used in section to model the uncertainty in the airplane’s behavior. The
second block, ∆2 is a fictitious uncertainty block, used to incorporate the H∞
performance objectives on the weighted output sensitivity transfer function
into the µ-framework.
Using theorem 4.5 from the “Robust Performance” section in Chapter 4, a
stabilizing controller K achieves closed-loop, robust performance if and only if
for each frequency ω ∈ [0, ∞], the structured singular value
µ∆[FL(P,K)(jω)] < 1
Using the upper bound for µ, (recall that in this case, two full blocks, the upper
bound is exactly equal to µ) we can attempt to minimize the peak closed-loop µ
value by posing the optimization problem

min min d ( s )I 2 0 –1
d ( s )I 2 0
F L ( P, K )
K d(s) 0 I2 0 I2
stabilizing stable,min–phase ∞

7-85
7 Robust Control Examples

Finding the value of this minimum and constructing controllers K that achieve
levels of performance arbitrarily close to the optimal level is called µ-synthesis.
A more detailed discussion of D – K iteration is given in Chapter 5.
Before plunging into the D – K iteration design procedure, we begin with a
controller designed via basic MIMO loop shaping methods.

Loop Shaping Control Design

One approach to control design for the HIMAT model is to synthesize a loop
shaping controller. We want the loop shape controller, Kloop, to make the
open-loop gain act as an integrator at low frequency and at crossover. At high
frequencies, we won't worry too much about the details of the roll-off, provided
that it is at least first order. To achieve this, we'll roughly invert the plant G(s)
(G has only 1 MIMO finite zero, at s ≈ −0.026; it also has zeros at s = ∞, so our
inverse is only approximate) and augment the desired loop gain dynamics to
the controller. The series of commands below constructs one such controller
and plots the open-loop gain (broken at the input to the controller), as seen in
Figure 7-51. The interested reader may want to explore various alternative
schemes for constructing loop shape controllers discussed in Freudenberg and
Looze (1988).
[a,b,c,d] = unpck(himat);
cn = c*a*a + 1000*c*a;
dn = c*a*b + 1000*c*b;
kloop = mscl(minv(pck(a,b,cn,dn)),-9000);
L = mmult(himat,kloop);
omega = logspace(-1,4,50);
Lg = frsp(L,omega);
vplot('liv,lm',vsvd(Lg))
title('Loop Gain Plot with Loop Shape Controller')
xlabel('Frequency (rad/s)')

7-86
 HIMAT Robust Performance Design Example

Loop Gain Plot with Loop Shape Controller

10 2

10 1

10 0

10 -1

10 -2

10 -3
10 -1 10 0 10 1 10 2 10 3 10 4

Frequency (rad/s)

Figure 7-51: Loop Gain of the Loop Shaping Controller

The open-loop gain plot satisfies both the low frequency performance objective
and the high frequency robustness goals. We have only plotted the singular
values of GKloop, but KloopG looks similar. Hence, you would expect the
controller to satisfy the robust stability and nominal performance
requirements.
The two 2 × 2 transfer functions associated with robust stability and nominal
performance can be evaluated for the loop shaping controller. Simply close the
open-loop interconnection P (himat_ic) with the loop shaping controller, Kloop
(kloop) and evaluate the pertinent transfer functions using the command sel.

7-87
7 Robust Control Examples

In using sel, the desired outputs (or rows) are specified first, followed by the
desired inputs (or columns). The results are seen in Figure 7-52.
clp = starp(himat_ic,kloop,2,2);
spoles(clp)
rs_loop = sel(clp,1:2,1:2);
np_loop = sel(clp,3:4,3:4);
rs_loopg = frsp(rs_loop,omega);
np_loopg = frsp(np_loop,omega);
vplot('liv,m',vnorm(rs_loopg),vnorm(np_loopg))
tmp1 = 'ROBUST STABILITY (solid) and';
tmp2 = ' NOMINAL PERFORMANCE (dashed)';
title([tmp1 tmp2])
xlabel('Frequency(rad/s)')

ROBUST STABILITY (solid) and NOMINAL PERFORMANCE (dashed)

0.55

0.5

0.45

0.4

0.35

0.3

0.25

0.2

0.15

0.1

0.05
10 -1 10 0 10 1 10 2 10 3 10 4

Frequency (rad/s)

Robust stability (solid line) Nominal Performance (dashed line)

Figure 7-52: Robust Stability and Nominal Performance Plots for the Loop
Shaping Controller

7-88
 HIMAT Robust Performance Design Example

The interpretation of the plots in Figure 7-52 is as follows:

• The controlled system (with loop shaping controller) achieves nominal

performance. This conclusion follows from the singular value plot of the
nominal weighted output sensitivity function, which has a peak value of
0.50.
• The controlled system (with loop shaping controller) achieves robust
stability. This conclusion stems from the singular value plot of the nominal
weighted input complementary sensitivity function, which has a peak value
of 0.50.

H∞ Design on the Open-Loop Interconnection

In this section, we carry out the first step of the D – K iteration, which is an H∞
(sub)optimal control, design for the open-loop interconnection, himat_ic. In
terms of the iteration, this amounts to holding the d variable fixed (at 1), and
minimizing the ||⋅ ||∞ norm of FL(P,K), over the controller variable K. Recall that
FL(P,K) is the nominal closed loop transfer function from the perturbation
inputs and disturbances (sysic variables pertin and dist) to the perturbation
outputs and errors (z and e), shown in Figure 7-53.

Figure 7-53: Closed-Loop Linear Fractional Transformation

The function hinfsyn designs a (sub)optimal H∞ control law based on the

open-loop interconnection structure provided. Syntax, input and output
arguments of hinfsyn are
[k,clp] = hinfsyn(p,nmeas,ncon,glow,ghigh,tol)

7-89
7 Robust Control Examples

The arguments are as follows.

Inputs

open-loop interconnection (SYSTEM matrix) p

number of measurements nmeas

number of controls ncons

lower bound for bisection glow

upper bound for bisection ghigh

absolute tolerance for bisection method tol

Outputs

controller (SYSTEM matrix) k

closed-loop (SYSTEM matrix) clp

In this example, the open-loop interconnection is himat_ic, with two

measurements, two control inputs, and the bisection algorithm will search for
the optimal achievable closed-loop norm, to an absolute tolerance of 0.06,
between lower and upper limits of 0.8 and 6.0, respectively. Since we are
planning on performing several iterations of the D – K iteration procedure, we
label the resulting controller k1. The resulting closed loop system (4-input,
4-output), from [pertin;dist] to [z;e] is labeled clp1.
[k1,clp1] = hinfsyn(himat_ic,2,2,0.8,6.0,.06);
The controller is stable, and its Bode plot is shown in Figure 7-54.

Properties of Controller
minfo(k1)
omega = logspace(-1,4,50);
spoles(k1)
k1_g = frsp(k1,omega);
vplot('bode',k1_g)
subplot(211), title('Frequency Response of k1')

7-90
 HIMAT Robust Performance Design Example

Frequency Response of k1
2
10

Log Magnitude
0
10

2
10

4
10 1 0 1 2 3 4
10 10 10 10 10 10
Frequency (radians/sec)

200
Phase (degrees)

200

400 1 0 1 2 3 4
10 10 10 10 10 10
Frequency (radians/sec)

Figure 7-54: Bode Plot of k1

Figure 7-55 shows the singular values of the closed-loop system clp1. Although
clp1 is 4 × 4, at each frequency it only has rank equal to 2, hence only two
singular values are nonzero.

Closed-Loop Properties
rifd(spoles(clp1))
clp1g = frsp(clp1,omega);
clp1gs = vsvd(clp1g);
vplot('liv,m',clp1gs)
title('Singular Value Plot of clp1')
xlabel('Frequency (rad/s)')

7-91
7 Robust Control Examples

Singular Value Plot of clp1

1.8

1.6

1.4

1.2

0.8

0.6

0.4

0.2

0
10 -1 10 0 10 1 10 2 10 3 10 4

Frequency (rad/s)

Figure 7-55: Singular Value Plot of the Closed-Loop System with k1

The two 2 × 2 transfer functions associated with robust stability and nominal
performance may be evaluated separately, using the command sel. Recall that
the robust stability test is performed on the upper 2 × 2 transfer function in
clp1, and the nominal performance test is on the lower 2 × 2 transfer function
in clp1. Since a frequency response of clp1 is already available, (in clp1g) we
simply perform the sel on the frequency response, and plot the norms.
rob_stab = sel(clp1g,[1 2],[1 2]);
nom_perf = sel(clp1g,[3 4],[3 4]);
minfo(rob_stab)
minfo(nom_perf)
vplot('liv,m',vnorm(rob_stab),vnorm(nom_perf))
tmp1 = 'ROBUST STABILITY (solid) and';
tmp2 = ' NOMINAL PERFORMANCE (dashed)';
title([tmp1 tmp2])
xlabel('Frequency (rad/s)')

7-92
 HIMAT Robust Performance Design Example

ROBUST STABILITY (solid) and NOMINAL PERFORMANCE (dashed)

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
10 -1 10 0 10 1 10 2 10 3 10 4

Frequency (rad/s)

robust stability (solid line) nominal performance (dashed line)

Figure 7-56: Robust Stability and Nominal Performance Plots Using

Controller k1

The interpretation of the plots in Figure 7-56 is as follows:

• The controlled system achieves nominal performance. This conclusion

follows from the singular value plot of the nominal weighted output
sensitivity function, which has a peak value of 0.92.
• The controlled system achieves robust stability. This conclusion stems from
the singular value plot of the nominal weighted input complementary
sensitivity function, which has a peak value of 0.86.

Assessing Robust Performance with µ

The robust performance, HIMAT example properties of the two different
closed-loop systems can be analyzed using µ-analysis. The closed-loop systems,
clp1 associated with the H∞ controller, and clp, associated with the loop
shaping controller, each have four inputs and four outputs. The first two
inputs/outputs correspond to the two channels across which the perturbation
∆G connects, while the third and fourth inputs/outputs correspond to the

7-93
7 Robust Control Examples

weighted output sensitivity transfer function. Therefore, for a frequency

domain µ-analysis of robust performance properties, the block structure should
consist of a 2 × 2 uncertainty block, and a 2 × 2 performance block.

 
 ∆1 0 2×2 2 × 2
∆:=  : ∆1 ∈ C ,∆ 2 ∈ C 
 0 ∆2 
 

Referring back to the “Robust Performance” section in Chapter 4, robust

performance (with respect to the uncertainty and performance weighting
functions specified above) is achieved if and only if for every frequency, µ∆(⋅) of
the closed-loop frequency response is less than 1.
The syntax of a general µ calculation is:
[bnds,dvec,sens,pvec] = mu(matin,deltaset)
The µ-analysis program, mu, calculates upper and lower bounds for the
structured singular value of the matrix matin, with respect to the block
structure deltaset. The matrix matin can be a CONSTANT MATLAB matrix,
or a VARYING matrix, such as a frequency response matrix of a closed-loop
transfer function. In this example, the frequency response is clp1g and the
block structure is two, 2 × 2 full blocks. mu returns the upper and lower bounds
in 1× 2 VARYING matrix bnds1, the frequency-varying D-scaling matrices in
dvec1, the frequency dependent perturbation associated with the lower bound
in rp1, and the sensitivity of the upper bound to the D-scales in sens1.
The bounds can be calculated by specifying the block structure, and running mu.

µ Analysis of H∞ Design
The H∞ design is analyzed with respect to structured uncertainty using µ.
First, the density of points in the frequency response is increased from 50 to
100 to yield smoother plots. Then the upper and lower bounds for µ are
calculated on the 4 × 4 closed-loop response of the matrix clp_g1. The upperand
lower bounds for µ are plotted (in this example they lie on top of one another)
along with the maximum singular value in Figure 7-57.

7-94
 HIMAT Robust Performance Design Example

deltaset=[2 2; 2 2];
omega1 = logspace(-1,4,100);
clp_g1 = frsp(clp1,omega1);
[bnds1,dvec1,sens1,pvec1] = mu(clp_g1,deltaset);
vplot('liv,m',vnorm(clp_g1),bnds1)
title('Maximum Singular Value and mu Plot')
xlabel('Frequency (rad/s)')
text(.15,.84,'max singular value (solid)','sc')
text(.3,.4,'mu bounds (dashed)','sc')
text(.2,.15,'H-infinity Controller','sc')

Maximum Singular Value and mu Plot

1.8

max singular value (solid)

1.6

1.4

1.2

1
mu bounds (dashed)
0.8

0.6

H-infinity Controller
0.4
10 -1 10 0 10 1 10 2 10 3 10 4

Frequency (rad/s)

FL(P,K1)(jω) (solid line) robust performance µ (dashed line)

Figure 7-57: Maximum Singular Value of the 4 × 4 Closed-Loop Transfer

Function FL(P,K1)(jω) and Robust Performance µ

7-95
7 Robust Control Examples

Hence, the controlled system (from H∞) does not achieve robust performance.
This conclusion follows from the µ plot in Figure 7-57, which peaks to a value
of 1.69, at a frequency of 73.6 rad/sec. This means that there is a perturbation
matrix ∆G, with ||∆G||∞ = ----------
1.69
1
- , for which the perturbed weighted sensitivity gets
large
||WP(I + Gnom(I + Wdel∆G)K–1||∞ = 1.69
This perturbation, ∆G, can be constructed using dypert. The input variables to
the command dypert consist of two outputs from µ, the perturbation matrix
and the bounds, along with the block structure, and the numbers of the blocks
for which the rational matrix construction should be carried out. Often times,
some of the blocks correspond to performance blocks and therefore need not be
constructed. Here, only the first block is an actual perturbation, so the
construction is only done for this 2 × 2 perturbation (fourth argument of
dypert).
delta_G = dypert(pvec1,deltaset,bnds1,1);
minfo(delta_G) % 2 by 2
rifd(spoles(delta_G)) % stable
hinfnorm(delta_G) % 1/1.69
clp_pert = starp(delta_G,clp1,2,2); % close top loop with delta
minfo(clp_pert)
rifd(spoles(clp_pert)) % stable, since RS passed
hinfnorm(clp_pert) % degradation of performance

- delta G

clp1 e  clp pert 
e   d
d

7-96
 HIMAT Robust Performance Design Example

µ-Analysis of Loop Shape Design

Robust performance for the system with the loop shape controller, Kloop, can
also be analyzed using µ. You might think that the loop shaping controller
would exhibit good robust performance properties, based on its excellent
nominal performance and robust stability properties.
clpg = frsp(clp,omega1);
bnds_loop = mu(clpg,deltaset);
vplot('liv,m',bnds_loop)
title('mu Plot of Closed-loop System')
xlabel('Frequency (rad/s)')
text(.6,.85,'Loop Shape Controller','sc')

However, the closed-loop system with the loop shaping controller does not
achieve robust performance. In fact, µ reaches a peak value of 11.7 at a
frequency of 0.202 rad/sec, as seen in Figure 7-58. This means that there is a
perturbation matrix ∆G, with ||∆G||∞ = ----------
11.7
1
- , for which the perturbed weighted
sensitivity gets large
||WP(I + Gnom(I + Wdel∆G)K–1||∞ = 11.7
Notice that this perturbation is 8.2 times smaller than the perturbation
associated with the H∞ control design, but that the subsequent degradation in
closed-loop performance is 8.2 times worse. Therefore, the loop shaping
controller will most likely perform poorly on the real system.

7-97
7 Robust Control Examples

mu Plot of Closed-loop System

12

Loop Shape Controller

10

0
10 -1 10 0 10 1 10 2 10 3 10 4

Frequency (rad/s)

Figure 7-58: Robust Performance µ Plot of the Closed-Loop HIMAT System

with the Loop Shaping Controller

The structured singular value µ is large in the low frequency range due to the
off-diagonal elements of clpg being large. One can see this using the command
blknorm, which outputs the individual norms of the respective blocks. The
coupling between the off-diagonal terms associated with 0.202 rads/sec point to
the problem — the upper right entry is 0.14, somewhat small, but not small

7-98
 HIMAT Robust Performance Design Example

enough to counteract the large (nearly 1000) lower left entry. As expected,
µ ≈ 0.14 * 959 = 11.6 .
blkn_cl = blknorm(clpg,deltaset);
see(xtract(blkn_cl,.15,.3))

2 rows 2 columns

iv = 0.159986
4.9995e-01 1.4127e-01
5.6525e+02 1.6402e-01

iv = 0.202359
4.9991e-01 1.4193e-01
9.5950e+02 1.6520e-01

iv = 0.255955
4.9985e-01 1.4294e-01
7.5635e+02 1.6607e-01

Recapping Results
Let’s summarize what has been done so far:

• The generalized plant, himat_ic, which includes the aircraft model,

uncertainty and performance weighting functions, and the interconnection
of all of these components was built using sysic.
• A controller was designed using hinfsyn.
• The robust performance characteristics of the closed-loop system were
analyzed with a structured singular value frequency domain test using mu.

The structured singular value analysis involved computing µ at each frequency

of this 4 × 4 closed loop response, with respect to a block structure ∆ which is
made up of two 2 × 2 full blocks. The blocks represent, respectively, uncertainty
in the aircraft model, and the performance objectives.
At this stage, the controller which has been designed using H∞ techniques
(nearly) minimizes the H∞ norm of the closed loop transfer function from the
4 × 1 vector of perturbation inputs and disturbance inputs to the 4 × 1 vector of
perturbation outputs and error signals. The structured singular value analysis

7-99
7 Robust Control Examples

shows that the µ analysis improves on the σ ( . ) bound at most frequencies, but
there is no improvement at the frequency of 73.6 rads/sec.
Hence, the peak value of the µ-plot is as high as the peak value on the singular
value plot, the µ analysis seems to have been of little use. However, at most of
the frequencies, µ is smaller than σ , and in the next iteration of synthesis, the
controller can essentially focus its efforts at the problem frequency, and lower
the peak of the µ-plot.

D – K Iteration for HIMAT Using dkit

The µ-Tools M-file dkit automates the µ-synthesis procedure via D – K
iteration. This example is a modified version of the HIMAT problem considered
earlier (see Figure 7-47) and is extended to include a frequency dependent
sensor noise signal, as shown in the closed-loop interconnection diagram in
Figure 7-59. This sensor noise signal is included to represent a more realistic
performance objective.

Figure 7-59: HIMAT Closed-Loop Interconnection Structure

Now, the open-loop interconnection structure is the eight input, six output
linear system, shown below

z   pertin
e himat ic  dist
y   control

with internal structure, as in Figure 7-60.

7-100
 HIMAT Robust Performance Design Example

The M-file mkhicn creates the plant model, weighting functions and the
interconnection structure shown in Figure 7-60. This can be produced with
nine MATLAB commands, listed below, and also in the M-file mkhicn (which
creates the plant and weighting functions).

" #
w1
w2 6
pertin dist(1:2)
wdel
6 -++?f- himat -++f? - wp -
"
e1
#

e2

control
y +?f+ wn  dist(3:4)
Figure 7-60: HIMAT Open-Loop Interconnection Structure

mkhicn

file: mkhicn.m
mkhimat;
wdel = nd2sys([50 5000],[1 10000]);
wp = nd2sys([0.5 0.9],[1 0.018]);
poleloc = 320;
Wn = nd2sys([2 0.008*poleloc],[1 poleloc]);
wdel = daug(wdel,wdel);
wp = daug(wp,wp);
Wn = daug(wn,wn);
systemnames = ' himat wp wdel wn ';
inputvar = '[ pertin{2} ; dist{4} ; control{2} ]';
outputvar = '[ wdel ; wp ; himat + dist(1:2) + wn ]';
input_to_himat = '[ control + pertin ]';
input_to_wdel = '[ control ]';
input_to_wp = '[ himat + dist(1:2) ]';
input_to_wn = '[ dist(3:4) ]';
sysoutname = 'himat_ic';
cleanupsysic = 'yes';
sysic;

7-101
7 Robust Control Examples

The dkit file himat_dk has been set up with the necessary variables to design
robust controllers for HIMAT using D – K iteration. A listing of the himat_dk
file follows. You can copy this file into your directory from the µ-Tools
subroutines directory, mutools/subs, and modify it for other problems, as
appropriate.
% himat_dk
%
% This script file contains the USER DEFINED VARIABLES for the
%mutools DKIT script file. The user MUST define the 5
%variables below.
%------------------------------------------%
REQUIRED USER DEFINED VARIABLES
%------------------------------------------%
% Nominal plant interconnection structure
NOMINAL_DK = himat_ic;

% Number of measurements
NMEAS_DK = 2;

% Number of control inputs

NCONT_DK = 2;

% Block structure for mu calculation

BLK_DK = [2 2;4 2];

% Frequency response range

OMEGA_DK = logspace(-3,3,60);

%----------------------end of himat_dk-------------------------%

7-102
 HIMAT Robust Performance Design Example

After the himat_dk.m file has been set up, you need to let the dkit program
know which setup file to use. This is done by setting the string variable
DK_DEF_NAME in the MATLAB workspace equal to the setup filename. Typing
dkit at the MATLAB prompt will then begin the D – K iteration procedure.
DK_DEF_NAME = 'himat_dk';
dkit
starting mu iteration #: 1

Iteration Number: 1
-------------------

Information about the Interconnection Structure IC_DK:

system10 states 6 outputs8 inputs
Test bounds: 0.0000 < gamma <= 100.0000

gamma hamx_eig xinf_eig hamy_eig yinf_eig nrho_xy p/f

100.000 2.3e-02 0.0e+00 1.8e-02 0.0e+00 0.0003 p
50.000 2.3e-02 0.0e+00 1.8e-02 0.0e+00 0.0011 p
25.000 2.3e-02 0.0e+00 1.8e-02 0.0e+00 0.0046 p
12.500 2.3e-02 0.0e+00 1.8e-02 0.0e+00 0.0183 p
6.250 2.3e-02 0.0e+00 1.8e-02 0.0e+00 0.0742 p
3.125 2.3e-02 0.0e+00 1.8e-02 0.0e+00 0.3117 p
1.562 2.3e-02 0.0e+00 1.7e-02 0.0e+00 1.5583# f
2.152 2.3e-02 0.0e+00 1.8e-02 0.0e+00 0.7100 p

Gamma value achieved: 2.1516

7-103
7 Robust Control Examples

SINGULAR VALUE PLOT: CLOSED-LOOP RESPONSE

2

1.8

1.6

1.4
MAGNITUDE

1.2

0.8

0.6

0.4

0.2

0
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3

FREQUENCY (rad/s)

Singular value plot of closed-loop system in graphics window. Make

sure that the chosen frequency range is appropriate.

7-104
 HIMAT Robust Performance Design Example

Next, we get to change the frequency range, if desired. For illustrative

purposes, we will change the number of logarithmically spaced points from 60
to 70.
Do you want to modify OMEGA_DK? (y/n): y

Current Frequency Variable

---------------------------------------------------------------
(s) Frequency Spacinglog
(n) # Frequency Points60
(b) Frequency - bottom1.00e-03
(h) Frequency - high1.00e+03

Enter (s n b and/or h) to change OMEGA, (e) to exit unchanged: n

------------CHANGING # of Points--------------

Enter desired # of points: 70

Current Frequency Variable

---------------------------------------------------------------
(s) Frequency Spacinglog
(n) # Frequency Points70
(b) Frequency - bottom1.00e-03
(h) Frequency - high1000

Enter (s n b and/or h) to change, (e) to exit: e

By typing e, we exit the frequency range modification, and the closed-loop

singular value frequency response is recalculated and plotted. In this case, the
plot looks exactly the same.

7-105
7 Robust Control Examples

SINGULAR VALUE PLOT: CLOSED-LOOP RESPONSE

2

1.8

1.6

1.4
MAGNITUDE

1.2

0.8

0.6

0.4

0.2

0
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3

FREQUENCY (rad/s)

Singular value plot of closed-loop system in graphics window.

Make sure that chosen Frequency range is appropriate.

Do you want to modify OMEGA_DK? (y/n): n

RERUN H_inf with DIFFERENT bounds/tolerances? (y/n): n

Calculating MU of closed-loop system: g_dk

points completed....
1.2.3.4.5.6.7.8.9.10.11.12.13.14.15.16.17.
18.19.20.21.22.23.24.25.26.27.28.29.30.31.32.33.34.35.
36.37.38.39.40.41.42.43.44.45.46.47.48.49.50.51.52.53 .
54.55.56.57.58.59.60.61.62.63.64.65.66.67.68.69.70.

7-106
 HIMAT Robust Performance Design Example

CLOSED-LOOP MU: CONTROLLER #1

2

1.8

1.6

1.4
MU

1.2

0.8

0.6

0.4
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3

FREQUENCY (rad/s)

MU plot for control design:Press any key to continue

Iteration Summary
------------------------------------
Iteration #1
Controller Order10
Total D-Scale Order0
Gamma Achieved2.152
Peak mu-Value2.075

Another D-K itera;tion? (y/n): y

Proceeding with the D – K iteration, we must fit the D-scaling variable that was
calculated in the µ upper-bound computation. This rational D-scaling will then
be absorbed into the open-loop interconnection.

7-107
7 Robust Control Examples

A plot of the µ upper bound, the first frequency-dependent D-scaling data (this
is the curve we want to fit), and the sensitivity of the µ upper bound. The
sensitivity measure roughly shows (across frequency) the relative importance
of the accuracy of the curve fit. It is used in the curve fit optimization to weight
some frequency ranges differently than others.

MU Upper Bound Magnitude Data: (1;1 1:d)

1
2.5 10

0
1.5 10

1
0.5 5 0 5
10 5 0 5
10 10 10 10 10 10

Sensitivity
1
10

0
10

1
10 5 0 5
10 10 10

7-108
 HIMAT Robust Performance Design Example

You are prompted to enter your choice of options for fitting the D-scaling data.
Press return to see your options.
Enter Choice (return for list):
Choices:

nd Move to Next D-Scaling

nb Move to Next D-Block
i Increment Fit Order
d Decrement Fit Order
apf Auto-PreFit
mx 3 Change Max-Order to 3
at 1.01 Change Auto-PreFit Tol to 1.01
0 Fit with zeroth order
2 Fit with second order
n Fit with n'th order
e Exit with Current Fittings
s 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 δI) 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 fit, and the order of the current
fit (with d for data, when no fit exists).
• The order of the current fit can be incremented or decremented (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 pre-fitting routine. mx
must be a positive, nonzero integer. at allows you to define how close the
rational, scaled µ upper bound is to approximate the actual µ upper bound in
a norm sense. Setting at 1 would require an exact fit of the D-scale data, and
is not allowed. Allowable values are greater than 1, and the meaning is

7-109
7 Robust Control Examples

explained in Chapter 5, “Control Design via m Synthesis”. This setting 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 will display a status of the current and fits.
Select apf to automatically fit the D-scale data. After a few seconds of
calculation, the first D-scale is fit with a fourth order rational curve as shown
in the top-right plot, along with the frequency-dependent magnitude data that
is being fit. Also shown in the top-left portion of the graphics window is a plot
comparing the upper bound of µ (using the frequency dependent D-scalings)
along with the maximum singular value of the closed-loop transfer function
after being scaled by the just-computed rational fit. Note that the second
D-scale data, which corresponds to the performance block, is fit with a
constant. This is expected since one of the D-scalings can always be normalized
to be 1. Enter s after the D-scale fitting is completed to see the status.
Enter Choice (return for list): apf
Starting Auto-PreFit...
Block 1 , Order = 0 1 2 3 4
Block 2 , Order = 0

Done

Enter Choice (return for list): s

Block 1: 4
Block 2: 0
Auto PreFit Fit Tolerance: 1.03
Auto PreFit Maximum Order: 5

Enter Choice (return for list):

7-110
 HIMAT Robust Performance Design Example

Scaled Bound and MU Magnitude Data and Fit: (1;1 1:4)

1
2.5 10

0
1.5 10

1
0.5 5 0 5
10 5 0 5
10 10 10 10 10 10

Sensitivity
1
10

0
10

1
10 5 0 5
10 10 10

In this case, the µ upper bound with the D-scale data is very close to the µ upper
bound with the rational D-scale fit. The fourth order fit is quite adequate in
scaling the closed-loop transfer function. The curve fitting procedure for this
scaling variable is concluded by entering e at the Enter Choice prompt.
Enter Choice (return for list): e

In this problem, the block structure consists of two complex full blocks: the 2 ×
2 block associated with the multiplicative uncertainty model for the aircraft,
and the 4 × 2 performance block. Since there are two blocks, there is only one
D-scaling variable, and we are completely done with the curve fitting in this
iteration.

7-111
7 Robust Control Examples

At the conclusion of the curve fitting procedure, a frequency response plot is

shown, which compares the norm of the rationally scaled, closed-loop system to
the lower and upper bound for µ.

MU bnds .vs. D*M*D^1 : ITERATION 2

1.5

0.5

3 2 1 0 1 2 3
10 10 10 10 10 10 10

Finally, before the next H∞ synthesis procedure, we get the option of changing
the parameters used in the hinfsyn routine. This is useful to change the lower
bound in the γ-iteration. In this example, we make no changes, and simply
continue.

7-112
 HIMAT Robust Performance Design Example

Altering the HINFSYN settings for next synthesis...

HINFSYN Settings Previously Next

------------------------------------------------------------
(u) GAMMA Upper Bound100 2.146
(l) GAMMA Lower Bound 0.00e+000.00e+00
(t) Bisection Tolerance1.0004.29e-02
(p) Riccati PSD epsilonl.00e-06l.00e-06
(j) Riccati j-w epsilonl.00e-08l.00e-08

Enter (u l t p j) to change, (e) to exit: e

The iteration proceeds by computing the H∞ optimal controller for the scaled
(using the rational scalings from the curve fitting) open-loop system.
Iteration Number: 2
--------------------

Information about the Interconnection Structure IC_DK:

system:26 states6 outputs8 inputs
Test bounds: 0.0000 < gamma <= 2.1461

gamma hamx_eig xinf_eig hamy_eig yinf_eig nrho_xy p/f

2.146 2.Oe–02 –5.2e–14 1.8e–02 –1.5e–16 0.1557 p
1.073 1.9e–02 –4.6e–14 1.7e–02 –3.4e–17 0.9958 p
0.537 1.4e–14# ******* 1.2e–02 –1.7e–18 ****** f
0.805 1.8e–02 –7.7e–13 1.6e–02 –5.8e–18 4.9336# f
1.019 1.9e–02 –6.Oe–14 1.7e–02 –2.4e–17 1.2077# f
1.055 1.9e–02 –2.3e–13 1.7e–02 –3.8e–17 1.0589# f

Gamma value achieved:1.0730

7-113
7 Robust Control Examples

SINGULAR VALUE PLOT: CLOSED-LOOP RESPONSE

1.2

0.8
MAGNITUDE

0.6

0.4

0.2

0
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3

FREQUENCY (rad/s)

Singular value plot of closed-loop system in graphics window.

Make sure that chosen Frequency range is appropriate.

Do you want to modify OMEGA_DK? (y/n): n

RERUN H_inf with DIFFERENT bounds/tolerances? (y/n): n

Calculating MU of closed-loop system: g_dk

points completed....
1.2.3.4.5.6.7.8.9.10.11.12.13.14.15.16.17.
18.19.20.21.22.23.24.25.26.27.28.29.30.31.32.33.34.35.
36.37.38.39.40.41.42.43.44.45.46.47.48.49.50.51.52.53.
54.55.56.57.58.59.60.61.62.63.64.65.66.67.68.69.70.

7-114
 HIMAT Robust Performance Design Example

CLOSED-LOOP MU: CONTROLLER #2

1.1

0.9
MU

0.8

0.7

0.6

0.5
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3

FREQUENCY (rad/s)

MU plot for control design:Press any key to continue

Iteration Summary
--------------------------------------------
Iteration #12
Controller Order1026
Total D-Scale Order016
Gamma Achieved2.1521.073
Peak mu-Value2.0751.073

Another D-K iteration? (y/n): y

The third iteration begins by fitting the new frequency-dependent D-scaling.

7-115
7 Robust Control Examples

MU Upper Bound Magnitude Data: (1;1 1:d)

1
1.2 10

0
1 10

1
0.8 10

2
0.6 5 0 5
10 5 0 5
10 10 10 10 10 10

Sensitivity
1
10

0
10

1
10 5 0 5
10 10 10

Again, enter the automatic pre-fitting option apf.

Enter Choice (return for list): apf
Starting Auto-PreFit...
Block 1 , Order = 0 1 2 3 4 5
Block 2 , Order = 0
Done

7-116
 HIMAT Robust Performance Design Example

Scaled Bound and MU Magnitude Data and Fit: (1;1 1:5)

1
1.2 10

0
1 10

1
0.8 10

2
0.6 5 0 5
10 5 0 5
10 10 10 10 10 10

Sensitivity
1
10

0
10

1
10 5 0 5
10 10 10

This fifth order fit works well in scaling the transfer function, so we exit the
curve fitting routine.
Enter Choice (return for list):e

MU bnds .vs. D*M*D^1 : ITERATION 3

1.1

0.9

0.8

0.7

0.6 3 2 1 0 1 2 3
10 10 10 10 10 10 10

7-117
7 Robust Control Examples

Altering the HINFSYN settings for next synthesis...

HINFSYN Settings Previously Next

----------------------------------------------------------------
(u) GAMMA Upper Bound 2.146 1.095
(1) GAMMA Lower Bound O.OOe+OO O.OOe+OO
(t) Bisection Tolerance 4.29e-02 2.19e-02
(p) Riccati PSD epsilon 1.OOe-06 1.OOe-06
(j) Riccati j-w epsilon 1.OOe-08 1.OOe-08

Enter (u l t p j) to change, (e) to exit: e

Iteration Number: 3
--------------------

Information about the Interconnection Structure IC_DK:

system:30 states6 outputs8 inputs
Test bounds: 0.0000 < gamma <= 1.0947

gamma hamx_eig xinf_eig hamy_eig yinf_eig nrho_xy p/f

1.095 2.1e-02 -1.2e-13 1.7e-02 -3.1e-18 0.5970 p
0.547 9.1e-13# ******* 1.2e-02 -4.2e-17 ****** f
0.821 2.Oe-02 -1.2e-11 1.6e-02 -5.5e-16 45.1126# f
1.040 2.1e-02 -7.Oe-13 1.7e-02 -2.4e-16 0.7263 p
0.996 2.1e-02 -8.4e-13 1.6e-02 -2.9e-17 0.8741 p
0.961 2.1e-02 -1.8e-13 1.6e-02 -2.5e-17 1.0433# f
0.970 2.1e-02 -7.9e-13 1.6e-02 -2.2e-16 0.9922 p

Gamma value achieved:0.9704

7-118
 HIMAT Robust Performance Design Example

SINGULAR VALUE PLOT: CLOSED-LOOP RESPONSE

1

0.9

0.8

0.7
MAGNITUDE

0.6

0.5

0.4

0.3

0.2

0.1

0
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3

FREQUENCY (rad/s)

Singular Value plot of closed-loop system in GRAPHICS window. Make sure

that chosen Frequency range is appropriate

Do you want to modify OMEGA_DK? (y/n): n

RERUN H_inf with DIFFERENT bounds/tolerances? (y/n): n

Calculating MU of closed-loop system: g_dk

points completed....
1.2.3.4.5.6.7.8.9.10.11.12.13.14.15.16.17.
18.19.20.21.22.23.24.25.26.27.28.29.30.31.32.33.34.35.
36.37.38.39.40.41.42.43.44.45.46.47.48.49.50.51.52.53.
54.55.56.57.58.59.60.61.62.63.64.65.66.67.68.69.70.

7-119
7 Robust Control Examples

CLOSEDLOOP MU: CONTROLLER #3

1

0.95

0.9

0.85
MU

0.8

0.75

0.7

0.65

0.6 3 2 1 0 1 2 3
10 10 10 10 10 10 10
FREQUENCY (rad/s)

MU plot for control design:Press any key to continue

Iteration Summary
----------------------------------------------------------
Iteration #123
Controller Order102630
Total D-Scale Order01620
Gamma Achieved2.1521.0730.970
Peak mu-Value2.0751.0730.973

Another D-K itera;tion? (y/n): n

echo off
Next MU iteration number: 4

At this point, we have achieved the robust performance objective, and we end
the D – K iteration. We have designed a 30 state controller using D – K
iteration which achieves a µ value less than 1.
In this example, it is also possible to reduce the controller order to 12, using
truncated balanced realizations, and still maintain closed-loop stability and
robust performance.

7-120
 HIMAT Robust Performance Design Example

max(real(spoles(kd_k3)))

ans =
-1.6401e-02

[k_dk3bal,hsv] = sysbal(k_dk3);
[k_dk3red] = strunc(k_dk3bal,12);
clpred_12 = starp(himat_ic,k_dk3red);
max(real(spoles(clpred_12)))

ans =
-6.9102e-03

clpred_12g = frsp(clpred_12,OMEGA_DK);
[bnds] = mu(clpred_12g,[2 2;4 2],'c');
pkvnorm(sel(bnds,1,1))

ans =
9.9910e-01

H∞ Loop Shaping Design for HIMAT

Now consider H∞ loop shaping control design for the HIMAT example discussed
in previous sections. Recall that the objective is to reject disturbances up to
about 1 rad/sec in the presence of substantial plant uncertainty above 100 rad/
sec. A loop-shaping design that gives a bandwidth of approximately 10 rad/sec
and robustness which should be satisfactory.

Design Precompensator
First form the HIMAT system and plot its maximum singular values across
frequency (see Figure 7-61).
mkhimat
[type,p,m,n] = minfo(himat);
om = logspace(-2,4,100);
himatg = frsp(himat,om);
vplot('liv,lm',vsvd(himatg),1);
title('SINGULAR VALUES OF HIMAT')
ylabel('SINGULAR VALUES'); xlabel('FREQUENCY (RAD/SEC)');

7-121
7 Robust Control Examples

SINGULAR VALUES OF HIMAT

10 5

10 4

SINGULAR VALUES 10 3

10 2

10 1

10 0

10 -1

10 -2

10 -3
10 -2 10 -1 10 0 10 1 10 2 10 3 10 4

FREQUENCY (RAD/SEC)

Figure 7-61: Singular Values of HIMAT

The singular values of himat are plotted in Figure 7-61, and although the unity
gain cross over frequency is approximately correct, the low frequency gain is
too low. We therefore introduce a proportional plus integral (P+I)
precompensator with transfer function (1 + s–1)I2×2 to boost the low frequency
gain and give zero steady state errors. The singular values of himat and himat
augmented with the P+I compensator are shown in Figure 7-62.
sysW1 = daug(nd2sys([1 1],[1 0]),nd2sys([1 1],[1 0]));
sysGW = mmult(himat,sysW1);
sysGWg = frsp(sysGW,om);
vplot('liv,lm',vsvd(himatg),'-.',vsvd(sysGWg),'-',1,'--')
title('SINGULAR VALUES OF HIMAT AND AUGMENTED HIMAT')
ylabel('SINGULAR VALUES');
xlabel('FREQUENCY (RAD/SEC)');

7-122
 HIMAT Robust Performance Design Example

SINGULAR VALUES OF HIMAT AND AUGMENTED HIMAT

10 5

10 4

SINGULAR VALUES 10 3

10 2

10 1

10 0

10 -1

10 -2

10 -3
10 -2 10 -1 10 0 10 1 10 2 10 3 10 4

FREQUENCY (RAD/SEC)

Figure 7-62: Singular Values of HIMAT (dashed-dotted) and Augmented Plant

with the H∞ Loop Shaping Controller (solid)

H∞ Loop Shaping Feedback Compensator Design

The optimally robust controller can now be designed for the frequency shaped
plant.
[sysK1,emax] = ncfsyn(sysGW,1);
disp(['emax = ' num2str(emax)]);
emax = 0.436

The value of emax = 0.436 is a very satisfactory stability margin. The

closed-loop norm can be checked by forming the open-loop interconnection of
Figure 7-63, denoted by p_ic, and checking the reciprocal of the H∞ gain. See
the “Loop Shaping Using H• Synthesis” section in Chapter 3 for more details
about this control design technique.

7-123
7 Robust Control Examples

Figure 7-63: H∞ Loop Shaping Standard Block Diagram

systemnames = 'sysGw';
inputvar = '[ w12; w22; u2 ]';
outputvar = '[ u; w1+sysGw; w1+sysGw ]';
input_to_sysGw = '[ w2+u ]';
sysoutname = 'p_ic';
cleanupsysic = 'yes';
sysic;
ncf_cl = starp(p_ic,sysK1);
ncf_cl_nm = hinfnorm(ncf_cl);
1/ncf_cl_nm(1)
ans =
4.3598e-01
The implemented controller involves the pre- and postweighting functions W1
and W2, as shown in Figure 7-64.

Figure 7-64: Actual Implemented H∞ Loop Shaping Controller

7-124
 HIMAT Robust Performance Design Example

In this example W2 = I2×2, therefore, the implemented loop shaping controller

is:
sysKloop = mmult(sysW1,sysK1);

Assessing Robust Performance with µ

We can now assess this controller by testing the original specification by using
a µ-test as in previous designs. First the interconnection structure needs to be
formed.
wdel = nd2sys([50,5000],[1,10000]);
wp = nd2sys([0.5,1.5],[1,0.03]);
wdel = daug(wdel,wdel);
wp = daug(wp,wp);
himatic
clear wp wdel

Now form the closed-loop and evaluate the robust performance µ with the H∞
loop shaping compensator implemented (see Figure 7-65).
clp1 = starp(himat_ic,sysKloop,2,2);
clp_g1 = frsp(clp1,om);
deltaset = [2 2; 2 2];
[bnds1,dvec1,sens1,pvec1] = mu(clp_g1,deltaset);
vplot('liv,m',bnds1);
title('ROBUST PERFORMANCE MU WITH LOOPSHAPE CONTROLLER')
ylabel('MU');
xlabel('FREQUENCY (RAD/SEC)');
disp(['mu value is ' num2str(pkvnorm(sel(bnds1,1,1)))])
mu value is 1.323

7-125
7 Robust Control Examples

ROBUST PERFORMANCE MU WITH LOOPSHAPE CONTROLLER

1.4

1.2

0.8
MU

0.6

0.4

0.2

0
10 -2 10 -1 10 0 10 1 10 2 10 3 10 4

FREQUENCY (RAD/SEC)

Figure 7-65: Robust Performance µ with sysK1

The plot of µ is shown in Figure 7-65 (solid line), the µ-value is close to that
required, giving a satisfactory design without exploiting the details of the
performance and uncertainty weights. This substantiates the claim that this
design method can give a very robust initial design which does not require
detailed trade-offs between weights to be studied.

Reduced Order Designs

The previously designed controller will typically have one less state than the
precompensator plus the plant. It is therefore often desirable to reduce the
number of states in the controller. There are systematic techniques for doing
this based on model reduction in the ν gap metric, δν, which is roughly
equivalent to model reduction of normalized coprime factors of the plant and
controller.

7-126
 HIMAT Robust Performance Design Example

First reduce the weighted plant model order and measure the resulting gap.
[sysGW_cf,sigGW]=sncfbal(sysGW);
sigGW
sigGW =

8.9996e-01
7.1355e-01
3.3542e-01
7.9274e-02
8.5314e-04
2.1532e-04

sysGW_4 = cf2sys(hankmr(sysGW_cf,sigGW,4,'d'));
gapGW_4 = nugap(sysGW,sysGW_4)
gapGW_4 =

8.6871e-04

It is seen that a fourth order model is essentially indistinguishable from the

full order model due to the small value of the ν gap. Now design the controller
for this reduced order system.
[sysK1_3,emax_3] = ncfsyn(sysGW_4,1);
emax_3
emax_3 =

4.3597e-01
This three state controller can be reduced to two states using Hankel model
reduction techniques (hankmr).
[sysK1_3_cf,sigK1_3] = sncfbal(sysK1_3);
sigK1_3
sigK1_3 =

3.1674e-01
2.7851e-01
6.9959e-02

sysK1_2 = cf2sys(hankmr(sysK1_3_cf,sigK1_3,2,'d'));
gapK_2=nugap(sysK1_3,sysK1_2)
gapK_2 =

6.9959e-02

7-127
7 Robust Control Examples

The robustness bound in the “Loop Shaping Using H• Synthesis” section in

Chapter 3, equation 3-23, can now be applied to give a lower bound on
robustness,
e_bound=sin(asin(emax_3)-asin(gapGW_4)-asin(gapK_2))
e_bound =

3.7114e-01

and this can be compared with the actual stability margin with the reduced
order controller as follows.
cl_red = starp(p_ic,sysK1_2);
tmp = hinfnorm(cl_red);
e_act=1/tmp(1)
e_act =

4.0786e-01

It is seen that the actual robustness is about half way between the optimal and
this lower bound. The important use of the bounds is that they indicate what
level of reduction is guaranteed not to degrade robustness significantly.
This gives a third-order controller together with the second-order P+I term.
The µ-value for this controller, not shown here, turns out to have essentially
the same µ -value as the closed-loop system with the full order controller.

Introducing a Reference Signal

A reference signal can be introduced into the loop shaping control design as
follows.
[sysK3,emax] = ncfsyn(sysGW,1.1,'ref');
cl_ref = starp(p_ic,sysK3,2,2);
minfo(cl_ref)
system: 12 states 4 outputs 6 inputs

When the ncfsyn option ref is specified, the controller includes an extra set of
reference inputs. The second input argument to ncfsyn is 1.1. This implies we
are designing a suboptimal controller with 10% less performance than at the
optimal. In practice, a 10% suboptimal design often performs better in terms of
robust performance than the optimal controller on the actual system.

7-128
 HIMAT Robust Performance Design Example

The last two inputs to cl_ref correspond to the reference signals, the first two
outputs are the outputs of the controller and the last two outputs are the inputs
to the controller (plant output plus observation noise). This design makes the
closed-loop transfer function from reference to plant output the numerator of a
normalized coprime factorization of sysGW. An external reference compensator
could also be added to improve the command response and there are many
possibilities. Here we first diagonalize the closed-loop reference to output
transfer function and then insert some phase advance to increase the speed of
response.
cl_ref_yr=sel(cl_ref,3:4,5:6);
P0 = transp(mmult([0 1; -1 0],cl_ref_yr,[0 1; -1 0]));
P1 = nd2sys([10 50],[1 50]);
P2 = daug(P1,P1);
sysQ = mmult(P0,P2);

Now reduce the order of sysQ to four states using the balanced realization
technique (sysbal), and incorporate into the controller.
[sysQ_b,sig_Q] = sysbal(sysQ);
sig_Q
sig_Q =
3.9665e+00
2.9126e+00
7.2360e-01
4.5915e-01
2.3600e-02
1.0016e-02
1.2526e-06
5.2197e-07
sysQ4 = strunc(sysQ_b,4);
sysK_ref = mmult(sysK3,daug(eye(2),sysQ4));

Finally form the closed-loop and calculate the step response.

sys_cl_ref = starp(p_ic,sysK_ref,2,2);
y = trsp(sys_cl_ref,[0;0;0;0;1;0],0.5,.001);
vplot(sel(y,1,1),'-.',sel(y,2,1),'.',sel(y,3,1),'-',...
sel(y,4,1),'--')
title('CLOSED LOOP TIME RESPONSE WITH SYSK1')
xlabel('TIME (SECONDS)')

7-129
7 Robust Control Examples

The step responses are plotted in Figure 7-66. The first output (solid) tracks the
command well with a rise time of less than 0.1 second and no overshoot. The
output of the second channel (dashed) is zero, indicating that there is no cross
coupling between the output channels in the nominal closed-loop system. The
controller output commands (dotted and dashed-dotted lines) are also plotted.
This is just the nominal step response and further tests are needed to check the
sensitivity of the closed-loop to the plant uncertainty.

CLOSED LOOP TIME RESPONSE WITH SYSK1

1

0.5

0 ... ....................................................................................................................................................................................................................................................................................................................................
. ...........................................................................
..........
. ......
....
. .....
-0.5 . ..
..
. ..
..
. ...
.
-1 . ..
..
. ...
. .
-1.5 .
. .
.
. .
.
.
. .
-2 .
. .
.
. .
. .
.
. .
-2.5 . .
.
. .
. ..
. ..
.. ..
-3 .... ...
..

-3.5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

TIME (SECONDS)

Figure 7-66: Time Response of Closed Loop System: sysK1

7-130
 HIMAT Robust Performance Design Example

HIMAT References
Freudenberg, J., and D. Looze, “Frequency domain properties of scalar and
multivariable feedback systems,” Lecture Notes in Control and Information
Sciences, Springer-Verlag, 1988.
Hartman, G.L., M.F. Barrett and C.S. Greene, “Control designs for an unstable
vehicle,” NASA Dryden Flight Research Center, Contract Report NAS 4–2578,
December, 1979.
Merkel, P.A., and R.A. Whitmoyer, “Development and evaluation of precision
control modes for fighter aircraft,” Proceedings of the AIAA Guidance and
Control Conference, San Diego, CA, Paper 76–1950, 1976.
Safonov, M.G., A.J. Laub, and G.L. Hartman, “Feedback properties of
multivariable systems: The role and use of the return difference matrix,” IEEE
Transactions on Automatic Control, Vol. AC–26, No. 1, pp. 47–65, February,
1981.

7-131
7 Robust Control Examples

F–14 Lateral-Directional Control Design

Consider the design of a lateral-directional axis controller for the F–14 aircraft
during powered-approach to landing. The linearized F–14 model is found at an
angle-of-attack (α) of 10.5 degrees and airspeed of 140 knots. The control
problem is posed as a robust performance, model matching problem with
multiplicative plant uncertainty at the plant input and minimization of
weighted output transfer functions as the performance criterion. A diagram for
the closed-loop system, which includes the feedback structure of the plant and
controller, and elements associated with the uncertainty models and
performance objectives, is shown in Figure 7-67.

Figure 7-67: F-14 Control Block Diagram

The performance objective is to have the true airplane, represented by the

dashed box in Figure 7-67, respond effectively to the pilot's lateral stick and
rudder pedal inputs. These performance objectives include:

7-132
 F–14 Lateral-Directional Control Design

• Decoupled response of the lateral stick, δlstk, to roll rate, p, and rudder
pedals, δrudp, to side-slip angle, β. The lateral stick and rudder pedals have
a maximum deflection of ± 1 inch. Therefore, they are represented as
unweighted signals in Figure 7-67.
The aircraft handling quality (HQ) response from the lateral stick to roll rate
2 deg ⁄ sec
should be a first order system, 5 -----------
s + 2 inch
- ---------------------- . The aircraft handling quality

response from the rudder pedals to side-slip angle should be

2
1.25 deg ⁄ sec
– 2.5 --------------------------------------------
2
- ---------------------- .
2 inch
s + 2.5s + 1.25

hqmod_p = nd2sys(5,[1 2]);

hqmod_beta = nd2sys(1.25^2,[1 2.5 1.25^2],-2.5);

The stabilizer actuators have ±20 degs and ±90 degs/sec deflection and
deflection rate limits. The rudder actuators have ±30 degs and ±125 degs/sec
deflection and deflection rate limits.
• The three measurement signals — roll rate, yaw rate, and lateral
acceleration — are passed through second order, anti-aliasing filters prior to
being fed to the controller.

2
ω
Anti-aliasing filter = ------------------------------------
2 2
s + 2ζω + ω

• The natural frequency, ω, and damping, ζ, values for the yaw rate and lateral
acceleration filters are 12.5Hz and 0.5, respectively, and 4.1 Hz and 0.7 for
the roll rate filter. The anti-aliasing filters have unity gain at DC (see
Figure 7-67). These signals are also corrupted by noise prior to entering the
controller.

7-133
7 Robust Control Examples

freq = 12.5;
fr = freq*2*pi;
zeta = 0.5;
antiaf_yaw = nd2sys(fr^2,[1 2*zeta*fr fr^2]);
antiaf_lata = nd2sys(fr^2,[1 2*zeta*fr fr^2]);
freq = 4.1; fr = freq*2*pi;
zeta = 0.7;
antiaf_roll = nd2sys(fr^2,[1 2*zeta*fr fr^2]);
antia_filt = daug(antiaf_lata,antiaf_roll,antiaf_yaw);

The performance objectives are accounted for in this framework via

minimizing weight transfer function norms. Weighting functions serve two
purposes in the H∞ and µ framework: they allow the direct comparison of
different performance objectives with the same norm and they allow
frequency information to be incorporated into the analysis. The F–14
performance weighting functions include:
Limits on the actuator deflection magnitude and rates are included via the
W_act weight. Wact is a 4 × 4 constant, diagonal scaling matrix described by
1 1 1 1
W act = diag( -----
-, ------, ----------, ------ . These weights correspond to the stabilizer and
90 20 125 30 

rudder deflection rate and deflection limits.

Wact = daug(1/90,1/20,1/125,1/30);

• Wn is a 3 × 3 diagonal, frequency varying weight used to model the

magnitude of the sensor noise W n = diag(0.025, 0.025, 0.0125 ------------------
s+1
s + 100
- ) , which
corresponds to the noise levels in the roll rate, yaw rate and lateral
acceleration channels.
W_n = daug(0.025,0.025,nd2sys([1 1],[1 100],0.0125);

• The desired δlstk-to-p and δrudp-to-β responses of the aircraft are formulated
as a model matching problem in the µ-framework. The difference between
the ideal response of the transfer functions, δlstk filtered through the roll rate
HQ model and δrudp filtered through the side-slip angle HQ model, and the
aircraft response, p and β, is used to generate an error that is to be

7-134
 F–14 Lateral-Directional Control Design

minimized. The Wp transfer function (see Figure 7-67) weights the difference
between the idealized roll rate response and the actual aircraft response, p.
4 3 2
0.05s + 2.90s + 105.93s + 6.17s + 0.16
W p = 0.5 * W β = ---------------------------------------------------------------------------------------------------------------
4 3 2
s + 9.2s + 30.80s + 18.33s + 3.95

The magnitude of Wp emphasizes the frequency range between 0.06 and 30

rad/sec. The desired performance frequency range is limited due to a
right-half plane zero in the model at 0.002 rad/sec, therefore accurate
tracking of sinusoids below 0.002 rad/sec isn’t required. Between 0.06 and 30
rad/sec a roll rate tracking error of less than 5% is desired. The performance
weight on the β tracking error, Wβ, is just 2 × Wp. This also corresponds to a
5% tracking error objective.
W_p = nd2sys([.05 2.9 105.93 6.17 .16],...
[1 9.2 30.8 18.83 3.95]);
W_beta = mscl(W_p,2);

All the weighted performance objectives are scaled to have an H∞ norm less
than 1 when they are achieved. The performance of the closed-loop system is
evaluated by calculating the maximum singular value of the weighted
transfer functions from the disturbance and command inputs to the error
outputs, as shown in Figure 7-68.

Figure 7-68: F-14 Weighted Performance Objectives Transfer Matrix

Nominal Model and Uncertainty Models

The pilot has the ability to command the lateral-directional response of the
aircraft with the lateral stick (δlstk) and rudder pedals (δrped). The aircraft has
two control inputs: differential stabilizer deflection (δdstab, degrees) and rudder
deflection (δrud, degrees); three measured outputs: roll rate (p, degs/sec), yaw
rate (r, degs/sec) and lateral acceleration (yac, g’s) and a calculated output,

7-135
7 Robust Control Examples

side-slip angle (β). Note that β is not a measured variable but is used as a
performance measure. The lateral-directional F–14 model, F14nom, has four
states: lateral velocity (v), yaw rate (r), roll rate (p) and roll angle (φ). These
variables are related by the state-space equations

v·
r· v
p· r
· p
φ = A B
β C D φ
p δ dstab
r δ drud
y ac

– 1.16e – 1 – 2.27e + 2 4.30e + 1 3.16e + 1

A = 2.65e – 3 – 2.59e – 1 – 1.45e – 1 0.00e + 0 ,
– 2.11e – 2 6.70e – 1 – 1.36e + 0 0.00e + 0
0.00e + 0 1.85e – 1 1.00e + 0 0.00e + 0

6.22e – 02 1.01e – 1
B = – 5.25e – 03 – 1.12e – 2 ,
– 4.67e – 02 3.64e – 3
0.00e + 00 0.00e + 0

2.47e – 1 0.00e + 0 0.00e + 0 0.00e + 0

C = 0.00e + 0 0.00e + 0 5.73e + 1 0.00e + 0 ,
0.00e + 0 5.73e + 1 0.00e + 0 0.00e + 0
– 2.83e – 3 – 7.88e – 3 5.11e – 2 0.00e + 0

0.00e + 00 0.00e + 0
D = 0.00e + 00 0.00e + 0 ,
0.00e + 00 0.00e + 0
2.89e – 03 2.27e – 3

7-136
 F–14 Lateral-Directional Control Design

Typing
load F14_nom

will load the nominal, F–14 plant model into the workspace. The dashed box in
Figure 7-67 represents the true airplane, which corresponds to a set of F–14
plant models defined by . Inside the box is the nominal model of the airplane
dynamics, F14nom, models of the actuators, AS and AR, and two elements, Win
and ∆G, which parameterize the uncertainty in the model. This type of
uncertainty is called multiplicative plant input uncertainty. The transfer
function Win is assumed known, and reflects the amount of uncertainty in the
model. The transfer function ∆G is assumed to be stable and unknown, except
for the norm condition, ||∆G||∞ ≤ 1. The aircraft uncertainty is modeled as a
complex full-block, multiplicative uncertainty at the input of the rigid body
aircraft nominal model. This is the same type of uncertainty description that
was used in the previous section entitled “HIMAT Robust Performance Design
Example”.
The stabilizer and rudder actuators, AS and AR, are modeled as first order
25
transfer functions, --------------
s + 25
- . The actuator outputs are their respective rates and
deflections.
A_S = pck(-25,25,[-25;1],[25;0]);
A_R = pck(-25,25,[-25;1],[25;0]);

Given the actuator and aircraft nominal models (denoted by Gnom(s)) we also
specify a stable, 2 × 2 transfer function matrix Win(s), called the uncertainty
weight. These transfer matrices parameterize an entire set of plants, , which
must be suitably controlled by the robust controller K.
:= {Gnom(I + ∆GWin) : ∆G stable, ||∆G||∞ ≤ 1}.
All of the uncertainty in modeling the airplane is captured in the normalized,
unknown transfer function ∆G. The unknown transfer function ∆G(s) is used to
parameterize the potential differences between the nominal model, Gnom(s),
and the actual behavior of the real airplane, denoted by .
In this example, the uncertainty weight Win is of the form

w1 ( s ) 0
W in ( s ) :=
0 w2 ( s )

7-137
7 Robust Control Examples

for particular scalar valued functions w1(s) and w2(s). The w1(s) weight
associated with the differential stabilizer input is selected to be
2(s + 4)
w1 ( s ) = ---------------------
s + 160
. The w2(s) weight associated with the differential rudder input
1.5 ( s + 20 )
is selected to be w 1 ( s ) = -----------------------------
s + 200
.

w_1 = nd2sys([1 4],[1 160],2);

w_2 = nd2sys([1 20],[1 200],1.5);
W_in = daug(w_1,w_2);

Hence the set of plants that are represented by this uncertainty weight

 25  2(s + 4) 
 0 I + 0 
∆ G ( s ) 
--------------- ---------------------
s + 25 s + 100
G :=  F14 nom
 2 1.5 ( s + 20 ) 
 0 25
---------------  0 ----------------------------- 
 s + 25 s + 200 

with ∆G(s) stable and ||∆G||∞ ≤ 1. Note that the weighting functions are used to
normalize the size of the unknown perturbation ∆G. At any frequency ω, the
value of |w1(jω)| and |w2(jω)| can be interpreted as the percentage of
uncertainty in the model at that frequency. The dependence on frequency of the
uncertainty weight indicates that the level of uncertainty in the airplane’s
behavior depends on frequency. Frequency response plots of weights w1 and w2
are shown in Figure 7-69.
om = logspace(-1,3,120);
W_ing=frsp(W_in,om);
vplot('liv,lm',sel(W_ing,1,1),'-',sel(W_ing,2,2),'--')
xlabel('Frequency (rad/sec)')
ylabel('Magnitude')

7-138
 F–14 Lateral-Directional Control Design

1
10

0
10
Magnitude

1
10

2
10 1 0 1 2 3
10 10 10 10 10
Frequency (rad/sec)

w1 (solid line) w2 (dashed line)

Figure 7-69: F-14 Uncertainty Weights w1, and w2

7-139
7 Robust Control Examples

The particular uncertainty weights chosen imply that in the differential

stabilizer channel at low frequency, there is potentially a 5% modeling error,
and at a frequency of 93 rad/sec, the uncertainty in channel 1 can be up to
100%, and could get larger at higher frequencies. The rudder channel has more
uncertainty at low frequency, up to 15% modeling error, and at a frequency of
177 rad/sec, the uncertainty is at 100%. To illustrate the variety of plants
represented by the set , type ex_f14tp at the command line to generate the
step responses of different systems from shown in Figure 7-70.
ex_f14tp

Diff. Stabilizer to Roll Rate Diff. Rudder to Beta

0 0.8

0.5 0.6
p (degrees/sec)

Beta (degrees)
1 0.4

1.5 0.2

2 0
0 1 2 3 4 0 1 2 3 4
Time (seconds) Time (seconds)

3 Diff. Stabilizer to Lat. Acceleration Diff. Rudder to Yaw Rate

x 10
8 0

6 0.2
r (degrees/sec)
y_ac (g’s)

4 0.4

2 0.6

0 0.8
0 1 2 3 4 0 1 2 3 4
Time (seconds) Time (seconds)

Figure 7-70: Unit Step Responses of the Nominal Model (+) and 15 Perturbed
Models from

The M-file ex_f14tp generates the family of perturbed time responses shown
in Figure 7-70.

7-140
 F–14 Lateral-Directional Control Design

file: ex_f14tp.m
Gnom = mmult(F14_nom,daug(sel(A_S,2,1),sel(A_R,2,1)));
u = step_tr(0,1,.02,2);
ydstab = trsp(Gnom,abv(u,0),4,.05);
ydrud = trsp(Gnom,abv(0,u),4,.05);
for i=1:15
delta = randn(2,2);
delta = delta/norm(delta);
p = mmult(Gnom,madd(eye(2),mmult(W_in,delta)));
y1 = trsp(p,abv(u,0),4,.05);
y2 = trsp(p,abv(0,u),4,.05);
ydstab = sbs(ydstab,y1);
ydrud = sbs(ydrud,y2);
end
cold = ynum(ydrud);
index = 2:cold;
subplot(221)
vplot(sel(ydstab,2,1),'+',sel(ydstab,2,[index]))
title('Diff. Stabilizer to Roll Rate')
xlabel('Time (seconds)'), ylabel('p (degrees/sec)')
subplot(222)
vplot(sel(ydrud,1,1),'+',sel(ydrud,1,[index]))
title('Diff. Rudder to Beta')
xlabel('Time (seconds)'), ylabel('Beta (degrees)')
subplot(223)
vplot(sel(ydstab,4,1),'+',sel(ydstab,4,[index]))
title('Diff. Stabilizer to Lat. Acceleration')
xlabel('Time (seconds)'), ylabel('ac_y (g''s)')
subplot(224)
vplot(sel(ydrud,3,1),'+',sel(ydrud,3,[index]))
title('Diff. Rudder to Yaw Rate')
xlabel('Time (seconds)'), ylabel('r (degrees/sec)')

The control design objective is to design a stabilizing controller K such that, for
all stable perturbations ∆G(s), with ||∆G||∞ ≤ 1, the perturbed closed-loop system
remains stable, and the perturbed weighted performance transfer function has
an H∞ norm less than 1 for all such perturbations. These mathematical
objectives exactly fit in the structured singular value framework.

7-141
7 Robust Control Examples

Controller Design
The control design block diagram shown in Figure 7-67 is redrawn as F14IC
shown in Figure 7-71. F14IC is the 25-state, six-input, six-output open-loop
transfer matrix used for control design. The M-file ex_f14ic contains the sysic
commands to generate the F14IC interconnection structure. The M-file
ex_f14wt, called from ex_f14ic, creates the weighting functions (Wact, Win, Wn,
Wp, and Wβ), the handling qualities models (hqmod_beta, hqmod_p),
anti-aliasing filters (anti_filt), the actuator models (AS and AR) and loads the
nominal F–14 plant model.
ex_f14ic

file: ex_f14ic.m
ex_f14wt
systemnames = 'W_in A_S A_R antia_filt hqmod_p hqmod_beta';
systemnames = [systemnames 'F14_nom W_act W_n W_P W_beta'];
inputvar = '[in_unc{2}; sn_nois{3}; roll_cmd; beta_cmd; ';
inputvar = [inputvar ' delta_dstab; delta_rud]' ];
outputvar = '[ W_in; W_P; W_beta; W_act; roll_cmd; ';
outputvar = [outputvar 'beta_cmd; antia_filt + W_n ]' ];
input_to_W_in = '[ delta_dstab; delta_rud ]';
input_to_A_S = '[ delta_dstab + in_unc(1) ]';
input_to_A_R = '[ delta_rud + in_unc(2) ]';
input_to_W_act = '[ A_S; A_R ]';
input_to_F14_nom = '[ A_S(1); A_R(1) ]';
input_to_antia_filt = '[ F14_nom(4) F14_nom(3) F14_nom(2)]';
input_to_hqmod_beta = '[ beta_cmd ]';
input_to_hqmod_p = '[ roll_cmd ]';
input_to_W_beta = '[ hqmod_beta - F14_nom(1) ]';
input_to_W_P = '[ hqmod_p - F14_nom(3) ]';
input_to_W_n = '[ sn_nois ]';
sysoutname = 'F14IC';
cleanupsysic = 'yes';
sysic

7-142
 F–14 Lateral-Directional Control Design

The first step in the D – K iteration control design procedure is to design an H∞

(sub)optimal controller for the open-loop interconnection, F14IC. In terms of
the D – K iteration, this amounts to holding the d variable fixed (at 1), and
minimizing the ||⋅ ||∞ norm of FL(F14IC, K), over the controller variable K. The
resulting controller is labeled K1.
The second step in D – K iteration involves solving a µ analysis corresponding
to the closed-loop system, FL(F14IC, K1). This calculation produces a frequency
dependent scaling variable dω, the (1,1) entry in the scaling matrix. In a
general problem (with more than two blocks), there would be several d
variables, and the overall matrix is referred to as the D-scales. The varying
variables in the D-scales are fit (in magnitude) with proper, stable, minimum
phase rational functions and absorbed into the generalized plant for additional
iterations. These scalings are used to trick the H∞ minimization to concentrate
more on minimizing µ rather than σ across frequency. For the first iteration in
this example, the d scale data is fit with a first order transfer function.

Figure 7-71: F-14 Generalized Plant

The new generalized plant used in the second iteration has 29 states, 4 more
states than the original 25-state generalized plant, F14IC. These extra states
are due to the D-scale data being fit with a rational function, and absorbed into
the generalized plant for the next iteration. Four D – K iterations are
performed until µ reaches a value of 1.02. Information about the D – K
iterations is shown in Table 7-1.

7-143
7 Robust Control Examples

Table 7-1 F-14 D – K Iteration Information

Iteration Number 1 2 3 4

Total D-Scale Order 0 4 4 4

Controller Order 25 29 29 29

H∞ Norm Achieved 1.562 1.079 1.025 1.017

Peak µ Value 1.443 1.079 1.025 1.017

To replicate these results using D – K iteration, start up dkitgui and press the
SETUP button in the main window. The data required in the DK Iteration
Setup window should be filled in to duplicate the Setup window shown in
Figure 7-72. The message “Mu-Synthesis Problem Specification
Complete...” will appear in the message bar upon correctly entering the
required data. Return to the main dkitgui window and press the Control
Design button. This will synthesize controller K1. To run 4 automated D – K
iterations, pull down the Iteration menu and select the number 4 from the
Auto Iterate menu.

7-144
 F–14 Lateral-Directional Control Design

Figure 7-72: dkitgui F-14 D – K Iteration Setup Window

Analysis of the Controllers

The robust performance properties of the controllers can be analyzed using
µ-analysis methods. Recall that robust performance is achieved if and only if
for every frequency, µ∆(FL(F14IC,K)(jω)) is less than 1. Plots of robust
performance µ of the closed-loop system with K1 and K4 implemented are
shown in Figure 7-73. The M-file ex_f14mu generates Figure 7-73 given that

7-145
7 Robust Control Examples

the original generalized plant, F14IC, and the controller SYSTEM matrices, K1
and K4, are in your MATLAB workspace.
ex_f14mu

file: ex_f14mu.m
om = logspace(-2,2,60);
clp1 = starp(F14IC,K1);
clp4 = starp(F14IC,K4);
clp1g = frsp(clp1,om);
clp4g = frsp(clp4,om);
deltaset = [2 2; 5 6];
mubnds1 = mu(clp1g,deltaset);
mubnds4 = mu(clp4g,deltaset);
vplot('liv,lm',mubnds1,'-',mubnds4,'--')
xlabel('Frequency (rad/sec)')

The controlled system with K1 implemented does not achieve robust

performance. This conclusion follows from the µ plot, which peaks to a value of
1.44, at a frequency of 7 rad/sec. Since µ is 1.44, there is a perturbation matrix
∆G, such that ||∆G||∞ = ----------
1.44
1
- , and the perturbed weighted performance transfer
functions gets large. After four D – K iterations the peak robust performance µ
value is reduced to 1.02 (Figure 7-73), thereby nearly achieving all of our robust
performance objectives.
To illustrate the robustness of the closed-loop system in the time-domain, time
responses of the ideal model, the nominal closed-loop system and the
worst-case closed-loop system from (using perturbations of size 1) are shown
in Figure 7-74. Controller K4 is implemented in the closed-loop simulations. A
1-inch lateral stick command is given at nine seconds, held at 1-inch till 12
seconds, and then returns to zero. The rudder is commanded at one second with
a positive 1-inch pedal deflection and held at 1-inch till four seconds. At four
seconds a –1-inch pedal deflection is commanded and held to seven seconds and
then returned to zero. One can see from the time responses that the closed-loop
response is nearly identical for the nominal closed-loop system and the
worst-case closed-loop system.

7-146
 F–14 Lateral-Directional Control Design

mu after 1 iteration (solid) and mu after 4 iterations (dashed)

1.6

1.4

1.2

1
mu

0.8

0.6

0.4

0.2 2 1 0 1 2
10 10 10 10 10
Frequency (rad/sec)

Figure 7-73: F-14 Robust Performance µ Plots with K1 and K4 Implemented

7-147
7 Robust Control Examples

beta: ideal (solid), actual (dasheddot), perturbed (dashed)

2

1.5

0.5
Sideslip angle (degrees)

0.5

1.5

2.5
0 2 4 6 8 10 12 14
Time (seconds)

rollrate: ideal (solid), actual (dasheddot), perturbed (dashed)

5

3
p (degrees/sec)

1
0 2 4 6 8 10 12 14
Time (seconds)

ideal (solid line) nominal (dash-dotted line) perturbed (dashed line)

Figure 7-74: Time Response Plots of the F-14 Lateral-Directional Control

System: Ideal, Nominal, and Perturbed

7-148
 F–14 Lateral-Directional Control Design

The M-file ex_f14s1 contains the sysic commands to generate the F14IC
simulation interconnection structure. ex_f14s2 contains the commands to
calculate the worst-case perturbation of size 1 for K4, the closed-loop time
response of the nominal and perturbed systems, and plots the results.
ex_f14s1
ex_f14s2

Figure 7-74 validates the frequency domain results showing that the controller
synthesized via D – K iteration, K4, is insensitive to changes in the model. You
will notice that the roll-rate response of the F–14 tracks the roll-rate command
well initially and then departs from the command. This is due to a right-half
plane zero in the aircraft model at 0.024 rad/sec.

file: ex_f14s1.m
systemnames = 'Win A_S A_R F14_nom antia_filt hqmod_p ';
systemnames = [systemnames ' hqmod_beta '];
inputvar = '[in_unc{2}; roll_cmd; beta_cmd; ';
inputvar = [inputvar ' delta_dstab; delta_rud]' ];
outputvar = '[ W_in; hqmod_p; F14_nom(2); hqmod_beta; '
outputvar = [outputvar ' F14_nom(1); roll_cmd; beta_cmd; '];
outputvar = [outputvar 'antia_filt ]' ];
input_to_W_in = '[ delta_dstab ; delta_rud ]';
input_to_A_S = '[ delta_dstab + in_unc(1) ]';
input_to_A_R = '[ delta_rud + in_unc(2) ]';
input_to_F14_nom = '[ AS(1); AR(1) ]';
input_to_antia_filt = '[F14_nom(4); F14_nom(3); F14_nom(2)]';
input_to_hqmod_beta = '[ beta_cmd ]';
input_to_hqmod_p = '[ roll_cmd ]';
sysoutname = 'F14SIM';
cleanupsysic = 'yes';
sysic

7-149
7 Robust Control Examples

file: ex_f14s2.m
om = logspace(-2,2,40);
delta = wcperf(frsp(clp4,om),deltaset,1,1);
sclp4_nom = starp(zeros(2,2),starp(F14SIM,K4));
sclp4_pert = starp(delta,starp(F14SIM,K4));
ustk = step_tr([0 1 4],[0 1 0],.02,10);
upedal = step_tr([0 1 4 7 ],[0 1 -1 0],.02,10);
input = abv(ustk,upedal);
y4nom = trsp(sclp4_nom,input,14,0.02);
y4pert = trsp(sclp4_pert,input,14,0.02);
subplot(211), vplot(sel(y4nom,3,1),'-',sel(y4nom,4,1),...
'-.',sel(y4pert,4,1),'--')
xlabel('Time (seconds)')
ylabel('Side-slip angle (degrees)')
title('beta: ideal (solid), actual (dashed-dot),...
perturbed (dashed)')
subplot(212), vplot(sel(y4nom,1,1),'-',sel(y4nom,2,1),...
'-.',sel(y4pert,2,1),'--')
xlabel('Time (seconds)')
ylabel('Roll rate (degrees/sec)')
title('roll-rate: ideal (solid), actual (dashed-dot),...
perturbed (dashed)')

7-150
 A Process Control Example: Two Tank System

A Process Control Example: Two Tank System

In this example we provide a relatively complete description of the entire
modeling, identification, and design process. The experiment is a simple two
tank system at Caltech. Other experimental work relating to this system is
described by Smith et al. [SmDMS, SmD1, SmD2].

Experimental Description
The system consists of two water tanks in cascade and is shown schematically
in Figure 7-75. The upper tank (tank 1) is fed by hot and cold water via
computer controllable valves. The lower tank (tank 2) is fed by water from an
exit at the bottom of tank 1. A constant level is maintained in tank 2 by means
of an overflow. A cold water bias stream also feeds tank 2 and enables the tanks
to have different steady-state temperatures.
Tank 1 is 5 3--8- inches in diameter and 30 inches in height. Tank 2 is 7 1--2- inches
in diameter and the overflow maintains the water level at 7 1--4- inches. This
configuration maintains the water level in tank 2 at 4 3--4- inches below the base
of tank 1. Flow control is obtained via linear electropneumatic actuators with
a CV of 1.0. One hundred inches of 1--2- -inch piping runs from each valve to the
top of tank 1. Approximately 36 inches of pipe connect the tanks, from the base
of tank 1 to the base of tank 2. The tank 2 cold water bias stream is manually
adjustable between 0.015 and 0.3 gpm. Thermocouples are mounted 1--4- inch
above the base of each tank. A pressure sensor (0 to 5 psig) provides a
measurement of the water level in tank 1.
All measured signals are filtered with fourth order Butterworth filters, each
with a cutoff frequency of 2.25 Hz. Twelve bit resolution is used for the A/D and
D/A conversions. In digital implementations of the controllers a sample period
of 0.1 seconds has been used.

7-151
7 Robust Control Examples

fh command
hot(th) supply
Actuator

fc command
fl+fd
cold (tc) supply
Actuator

supply at temp td
Tank1
fd

fl
fl+fd

Tank2 h2

Figure 7-75: Schematic Diagram of the Two Tank System

An Idealized Nonlinear Model

The Two tank example, nonlinear model system is first considered without the
actuators or sensors and an ideal nonlinear model is derived. Suitable units are
added to give a basis for the subsequent discussion of the experimental data.
The model of the actuators and sensors (including noise) is based on
experimental data and is included later.

7-152
 A Process Control Example: Two Tank System

We make the following unrealistic assumptions:

• There are no thermal losses in the system.

• Perfect mixing occurs in both tanks.
• The flow out of tank 1 is related only to the height of tank 1.
• There are no thermal or flow delays.

The system variables are given the following designations.

Variable Physical Quantity

ƒhc Command to hot flow actuator

ƒh Hot water flow into tank 1

ƒcc Command to cold flow actuator

ƒc Cold water flow into tank 1

ƒ1 Total flow out of tank 1

A1 Cross-sectional area of tank 1

h1 Tank 1 water level

t1 Temperature of tank 1

t2 Temperature of tank 2

A2 Cross-sectional area of tank 2

h2 Tank 2 water level

ƒb Flow rate of tank 2 bias stream

tb Temperature of tank 2 bias stream

th Hot water supply temperature

tc Cold water supply temperature

Tank 1 is considered first. Conservation of mass gives,

7-153
7 Robust Control Examples

d
----- ( A 1 h 1 ) = f h + f c – f 1 .
dt
(7-2)

It is assumed that the flow out of tank 1 (ƒ1) is a memoryless function of the
height (h1). As the exit from tank 1 is a pipe with a large length to diameter
ratio, the flow is proportional to the pressure drop across the pipe and thus to
the height in the tank. With a constant correction term for the flow behavior at
low tank levels the height and flow can be reasonably approximated by an
affine function,

h 1 = α f 1 – β where α, β > 0 and f 1 ≥ β ⁄ α .

(7-3)

Defining ƒ1 as a state variable leads to a linear state equation and an affine

output equation for h1 (in the allowable range of ƒ1).

· –1 1 1
f 1 = ----------
- f 1 + ----------- f h + ----------- f c .
A1 α A1 α A1 α
(7-4)

h 1 = α f 1 – β.
(7-5)

Conservation of energy leads to a model for the temperature of tank 1 (t1). It is

useful to define a variable,

E1 = h1 t1 ,
(7-6)

which can loosely be thought of as the energy in tank 1. Defining E1 as a state

variable gives a nonlinear state equation and a nonlinear output equation for
t 1.

β th tc
E 1 = -----------  1 + ------ E 1 + ------- f h + ------- f c .
· –1
A1 α  h1  A1 A1
(7-7)

1
t 1 = ------ E 1 .
h1
(7-8)

7-154
 A Process Control Example: Two Tank System

Note that for a fixed h1 the above equations are linear. This will aid in the
identification.
In tank 2 the height (h2) is fixed and the input flow from tank 1 (ƒ1) is of
temperature t1. This gives only one equation.

d
----- ( A 2 h 2 t 2 ) = f 1 t 1 + f b t b – ( f 1 + f b )t 2 .
dt
(7-9)

We will develop a model which has t1 and h1 as inputs and t2 as an output. This
will allow us to concatenate the tank 1 and tank 2 models to give a model for
the full system. A more physically motivated model might have t1 and ƒ1 as
inputs. The linearizations will differ only by the factor α, so the difference is
not significant.
We can rearrange equation 7-9 to give,

 h 1 + β + α f b h1 + β fb tb
t·2 = –  ------------------------------- t 2 +  ------------------ t 1 + --------------
 αA h
2 2 
 αA 2 2h  A2 h2
(7-10)

The output equation for tank 2 is trivial.

Equations 7-4, 7-5, 7-7, 7-8, and 7-10 provide a simple nonlinear model for the
two tanks. To complete the description of the system a model must be obtained
for the actuators and sensors. The uncertainty associated with elements of the
model must also be identified. First, we define a suitable set of units for our
model.

7-155
7 Robust Control Examples

Normalization Units
To quantify the model we define a system of normalized units as follows.

Physical Unit 0 Unit 1 Unit

Variable Name Definition Definition

temperature tunit tc = 1.0 tunit th = 0.0 tunit

height hunit tank 1 empty tank 1 full

flow funit zero input flow maximum (2.0 gpm)

input flow

The first two definitions are sufficient to define all of the other units in the
problem. The input flows range from 0 to 2.0 gallons/minute and it is
convenient to define a flow unit at the input by 2.0 gpm = 1.0 funit. Using the
above units the system dimensions are now,

Variable Value Units

A1 0.0256 hunits2

A2 0.0477 hunits2

h2 0.241 hunits

ƒb 7.4 × 10–5 hunits3/sec

tb 0.0 tunits

ƒs 0.00028 hunits3/sec/funit

The variable ƒs is a flow scaling factor which converts the input (0 to 1 funits)
to flow in hunits3/second. This is used in the tankdemo script.

7-156
 A Process Control Example: Two Tank System

Operating Range
The physical system imposes constraints on the operating region. The most
obvious of these is that as the bias stream is cold, the temperature of tank 2 (t2)
must be less than that of tank 1 (t1). Saturation in the actuators prevents tank
1 from being completely full of either hot or cold water. The relationship
between ƒ1 and h1 can only be modeled by equation 7-3 for h1 in the range:

0.15 ≤ h 1 ≤ 0.75.

The numerical model given below applies only to this range.

In the linear design example the operating range of tank 1 is further reduced
to:

0.25 ≤ h 1 ≤ 0.75.
0.25 ≤ t 1 ≤ 0.75.

Both sets of constraints severely limit the operating regions.

Actuator Model
There are significant dynamics and saturations associated with the actuators
and a model of these is included. In the frequency range of interest the
actuators can be modeled as a single pole system with rate and magnitude
saturations. The rate saturation has been estimated from observing the effect
of triangle waves of different frequencies and magnitudes. The following model
will be used for the actuators.

1
f h = ----------------------------- fhc
( 1 + 0.05s )
(7-11)

with a magnitude limit of 1.0 funits and a rate limit of 3.5 funits/sec. It is the
rate limit, rather than the pole location, that limits the actuator performance
for most signals. For a linear model some of the effects of rate limiting can be
included in a perturbation model.

7-157
7 Robust Control Examples

Experimental Assessment of the Model

Where possible, the system has been broken into subsystems, which can be
identified independently. For example, the actuators are considered
independently and tank 1 is identified separately from tank 2. This approach
is not necessarily general.
Open-Loop Experiments
Static measurements of h1 and ƒ1 can be used to obtain estimates of α and β
from the theoretical relationship: h1 = αƒ1 – β. Note that α appears as a gain in
the nominal model and can also be easily estimated by dynamic measurements.
Using such experiments gives the estimates, α = 4876 and β = 0.59.
Open loop experiments have been conducted to test the applicability of the
model and obtain some feel for the level of uncertainty that will be appropriate.
Band limited white noise, in several frequency bands, was used for the input
signals. Data records were 8192 samples in length with sample rates of 1.0 Hz
and 10.0 Hz. The transfer function estimates, presented below, have been
obtained by the Welch method, using Hanning windows on sections of the data.
The data plotted in the figures comes from several window lengths, typically
1024 and 4096. Only the points with good coherency are plotted.
Equation 7-76 shows the estimated transfer function between ƒhc + ƒcc and h1,
and the transfer function predicted from the model (equations 7-4 and 7-5),
including the nominal actuator and Butterworth filter. The experimental data
comes from five experiments at three different levels. For frequencies greater
than 0.2 Hz, the plotted data comes from an experiment at h1 = 0.47.
Note that the theoretical model is a very close match to the experimental data
over a wide range of frequencies. This suggests that a small perturbation
weight is suitable for modeling the discrepancies between the nominal model
and physical system behavior. Note also that the h1 model is independent of
both h1 and t1. In other words it applies to the entire range of operating points.
For h1 fixed, the E1 state variable equation 7-7 and the t1 output equation 7-8
are linear. Experiments have been performed at h1 = 0.15, 0.25, 0.47, and 0.75.
The input waveforms were generated such that ƒhc = –ƒcc which maintains a
constant h1. Figure 7-77 shows the transfer function between ƒhc – ƒcc(= 2ƒhc)
and t1 calculated from the experimental data and estimated from the model
(equations 7-7 and 7-8. For the data shown h1 = 0.15 and h1 = 0.75. The other
cases lie between the two curves shown.

7-158
 A Process Control Example: Two Tank System

10 1

10 0

10 -1
magnitude

10 -2

10 -3

10 -4

10 -5
10 -4 10 -3 10 -2 10 -1 10 0 10 1

frequency: Hz

-100

-200
phase - degrees

-300

-400

-500
experimental data
and theoretical
-600
10 -4 10 -3 10 -2 10 -1 10 0 10 1 model (solid line)
frequency: Hz

Figure 7-76: Transfer Function Between ƒhc + ƒcc and h1. Experimental Data
and Theoretical Model

7-159
7 Robust Control Examples

10 1

10 0

10 -1 h1 = 0.15
magnitude

10 -2

h1 = 0.75
10 -3

10 -4
10 -4 10 -3 10 -2 10 -1 10 0 10 1

frequency: Hz

-50

-100

-150
phase - degrees

h1 = 0.15
-200

-250

-300

-350 h1 = 0.75

-400 experimental data

-450
and models
10 -4 10 -3 10 -2 10 -1 10 0 10 1 (solid lines)
frequency: Hz

Figure 7-77: Transfer Function Between ƒhc – ƒcc and t1. Experimental Data
and Models. h1 = 0.15 and h1 = 0.75.

7-160
 A Process Control Example: Two Tank System

The nominal model/physical system discrepancies are far more significant for
the t1 case. A larger t1 perturbation weight is required to cover these
discrepancies. A more complete discussion on this issue is given in the
“Perturbation Model” section.
Similar analyses have been performed on tank 2. The open loop data is less
conclusive due to observing the system through tank 1 and a higher noise level.
The available data does confirm the theoretical model but to a lower frequency
than that for tank 1. More definitive results have been obtained for tank 2 with
the closed-loop experiments.

Closed-Loop Experiments
We will look at a closed-loop method of estimating a suitable uncertainty level
for the model. This involves using a relay to induce limit cycling and is based
on an auto-tuning method proposed by Åström and Hägglund [AstH]. More
detail on using such approaches for estimating uncertainty levels is described
by Smith and Doyle [SmD2].
Applying a relay in a feedback loop may drive the closed-loop system into stable
limit cycles. This technique works for a large class of systems including the two
tank system. In the experiments performed here a decoupled controller (into
height and temperature loops) was used with a relay in the temperature
control loop. This allowed stable control of the tank height (h1) and produced
limit cycles in t1. These experiments were performed at fixed heights (h1 = 0.15,
0.25, 0.47, and 0.75).
With a simple relay the closed-loop system will limit cycle at the frequency
where the response has a phase of 180 degrees. The gain at this frequency can
also be estimated from the input/output data. This experiment will identify the
system at a single point. Using this information, a new controller is designed
to introduce some lead into the closed-loop system. This new closed-loop system
has limit cycles at a higher frequency giving an additional point at which the
plant can be identified. In practice this technique can be repeated until the
nonlinear and/or inconsistent effects dominate and the closed-loop system no
longer limit cycles consistently. This also provides information on the
frequency at which uncertainty should dominate in the model.
Details of the application of this approach to tank 1 are given in [SmD2]. To
illustrate the concept, the configuration used to induce limit cycles in t1 is
shown in Figure 7-78.

7-161
7 Robust Control Examples

Height, h, is controlled by the proportional controller, Kh. R denotes the relay and
Ci denotes one of a series of lead controllers used to adjust the limit cycle frequency.

Figure 7-78: Closed-Loop Configuration for Relay Limit Cycle Experiments

of t1.

A series of controllers, Ci in Figure 7-78, was applied. These were used to

introduce differing amounts of lead into the closed-loop system, giving differing
limit cycle frequencies and amplitudes. A typical time response for such an
experiment is shown in Figure 7-79. In this manner limit cycles were induced
in the t1 loop. The highest frequency at which a limit cycle could be induced was
0.023 Hz. We will subsequently see that this can be used as a heuristic for
determining an appropriate perturbation weight for t1.
The tank 2 model was similarly studied. Figure 7-80 presents the limit cycle
data obtained for tank 2. The E1 to t2 theoretical transfer function (for fixed h1)
is shown for h1 = 0.15, 0.25, 0.47, and 0.75. Also shown are the points identified
from each experiment. It was not possible to induce consistent limit cycles at
frequencies above 0.03 Hz, indicating that this is a frequency at which the
uncertainty should dominate in the model.

7-162
 A Process Control Example: Two Tank System

0.6

0.55

0.5
level & temp

0.45

0.4

0.35

0.3
0 50 100 150 200 250 300 350 400

seconds

0.9

0.8

0.7
hot & cold flows

0.6

0.5

0.4

0.3

0.2

0.1

0
0 50 100 150 200 250 300 350 400

seconds

Figure 7-79: Closed-Loop Relay Limit Cycle Experiment Controller: C1

7-163
7 Robust Control Examples

10 1

+
+
+
10 0 x
o
x * +
*+
*+

h1 = 0.15
10 -1
magnitude

h1 = 0.75
10 -2

10 -3

10 -4
10 -4 10 -3 10 -2 10 -1 10 0 10 1

frequency: Hz

-50 o
x +
x
+
* ++ +
+
-100 **

-150
magnitude

-200

-250

-300

-350

-400

-450
10 -4 10 -3 10 -2 10 -1 10 0 10 1

frequency: Hz

Figure 7-80: Limit Cycle Identification Experiments for Tank 2. Tank 1 Height
Held Fixed at h1 = 0.15, 0.25, 0.47, and 0.75

7-164
 A Process Control Example: Two Tank System

Developing the Interconnection Model for Design

The problem considered is the control of t1 and t2 for command response. The
desired closed-loop configuration is illustrated in Figure 7-81. The controller
has access to both the reference inputs and the temperature measurements. It
would be possible (although more restrictive) to design a controller which used
only the error between the temperature measurements and the setpoints.

The t1 and t2 reference signals are denoted by t1cmd and t2cmd, respectively.
Noise is assumed to enter within the tank system block.

Figure 7-81: Closed-Loop Design Problem

The layout of the interconnection structure is illustrated in Figure 7-82. The

subsequent sections will develop a perturbed model and the necessary design
weights for the design problem.

7-165
7 Robust Control Examples

Figure 7-82: Interconnection Structure for the Design Problem

Linearized Nominal Model

The top tank model, given in equations 7-4, 7-5, 7-7, and 7-8, can be linearized
to give the following model.

–1
----------- 0
αA 1
f·1 = f1
·
E1 βt̂ 1 – ( β + ĥ 1 ) E
-------------- ------------------------ 1
A 1 ĥ 1 αA 1 ĥ 1

1 1
----------- -----------
αA 1 αA 1 fh
+
th tc f
----------- ----------- c
αA 1 αA 1
(7-12)

7-166
 A Process Control Example: Two Tank System

α 0
h1 1 f
= – αt̂ 1 1
t1 ------------ ------ E 1
ĥ 1 ĥ 1
(7-13)

The steady state values of h1 and t1 are given by ĥ 1 and t̂ 1 , respectively.

A linearized model for tank 2 is given by,

– ( ĥ 1 + β + α f b ) t̂ 1 + t̂ 2 ĥ 1 + β h 1
t·2 = ---------------------------------------t 2 + -----------------
- ------------------ .
αA 2 h 2 αA 2 h 2 αA 2 h 2 t 1
(7-14)

As above, t̂ 2 denotes the steady state value of t2, and can be calculated by,

f̂ 1 t̂ 1 + f b t b
t̂ 2 = --------------------------
f̂ 1 + f b

where the steady state flow out of tank 1 is,

f̂ 1 = αĥ 1 – β .

To complete the linearized nominal model we must select a nominal operating

point. The demo script, tankdemo, develops the equations in the form given
above and uses the values hˆ 1 = 0.75 and t̂ 1 = 0.75 for the actual design.

Perturbation Model
We must now select a perturbation model structure for our robust control
model. This involves determining the manner in which the perturbations enter
the model, and selecting appropriate frequency dependent weights to
normalize the perturbations. The guiding approach is to make the model

7-167
7 Robust Control Examples

perturbations match our estimate of uncertainty in the physical system as

closely as possible.
For example, it is reasonable to attribute some dynamic uncertainty to the
actuators, which suggests that uncertainty at the inputs, ƒh and ƒc, is
appropriate. It is also reasonable to consider this uncertainty as being relative
with respect to the input. A potential model may have uncertainty modeled by
multiplicative perturbations on ƒh and ƒc. This is particularly appropriate for
high condition number plants as it prevents the controller from inverting the
plant from the input in a robust design.
We already know from the open-loop experiments that there is a significant
amount of dynamic uncertainty in the t1 response. This is due primarily to
mixing and heat loss. This suggests that a multiplicative perturbation on the
t1 output is also an appropriate model.
There are many choices available and we have chosen to capture the
uncertainty with multiplicative perturbations on h1, t1, and t2. One could also
add perturbations on ƒh and ƒc, although it is doubtful that the additional
model complexity (it would now have five perturbations), would be worthwhile.
Arguably, we do not have such a detailed model of the uncertainty in this case.
Figure 7-83 illustrates the perturbed linear model of the two tank system. The
three perturbations, ∆1, ∆2, and ∆3, model the uncertainty in h1, t1, and t2,
respectively. Note that they are multiplicative perturbations and that t2 is
influenced by all three perturbations. The nominal tank 1 model is described
by equations 7-12 and 7-13, and the nominal tank 2 model is given by equation
7-14.

Figure 7-83: Schematic Representation of the Perturbed, Linear, Two Tank

Model

7-168
 A Process Control Example: Two Tank System

To complete this model we must specify the perturbation weights, Wh1, Wt1,
and Wt2. As we might expect, it is not possible to precisely determine the
amount of uncertainty in the system. At best, we look for a rough frequency
dependent bound. We discuss some of the ad-hoc means for getting suitable
weights below:

• Accuracy limits on actuators and sensors. Equipment manufacturers will

often specify accuracy limits on such equipment and this serves as a useful
starting point for at least the steady-state uncertainty of some system
components.
Known rate limits can also be viewed as limiting the high frequency behavior
of an actuator. Rate limits are nonlinear and it will not be possible to
completely capture their effects with even a linear perturbation model.
However, these limits may give an approximate estimate of the bandwidth
of an actuator.

• Linearization constants. If a nonlinear model is available, then examining

the variation in the linearization constants across the expected range of
operation will give an estimate of appropriate levels of uncertainty for
elements of the model. For example, we could examine how the time constant
of the tank 1 t1 response varied across the h1 operating range.
Again, this is only a crude approximation. The full nonlinear behavior of the
plant is not captured by considering a series of frozen linear systems at
various operating points.

• Open-loop identification experiments. Comparing experimentally estimated

transfer functions to those predicted by a nominal model gives a good
estimate of the model/system discrepancy at each frequency.

• Closed-loop identification experiments. These are preferred over open-loop

identification experiments for a number of reasons:
Closing the loop automatically biases the plant to the correct operating point.
This provides appropriate signals for identification as they are close to those
signals that will occur in final closed-loop operation;
The achievable performance (at least in terms of closed-loop bandwidth) is
limited by the uncertainty at frequencies close to the loop cross-over

7-169
7 Robust Control Examples

frequency. Open-loop experiments tend to emphasize the frequencies where

the plant gain is high – typically steady-state in many process control
systems. This is simply due to the significantly larger response on the plant
at those frequencies. On the other hand, using closed-loop identification
experiments places the emphasis on the loop cross-over frequencies, which is
more important for the ultimate closed-loop design.
The analysis of closed-loop identification experiments needs to be performed
carefully. The measurement noise is no longer uncorrelated with the input
and output signals and its effect cannot be removed simply by averaging.
Van den Hof et al. [VdHSB] give a suitable means of dealing with this
situation.

• Closed-loop relay experiments. The “Closed-Loop Experiments” section

briefly discussed using relays to induce closed-loop limit cycle operation. A
large class of systems (including the two tank system) will limit cycle under
these circumstances. The frequency at which the system limit cycles is
simply the loop cross-over frequency and we can investigate the achievable
closed-loop bandwidth by using lead controllers to make the system limit
cycle at successively higher frequencies.
Beyond some frequency the system no longer has a limit cycle; the dynamics
are no longer dominated by linear time-invariant behavior. The heuristic is
that at such frequencies the effect of the perturbations should dominate the
model. This is easy to apply when a multiplicative perturbation is used;
when the magnitude of the weight is greater than one, the perturbation
dominates the model.
This approach was used for determining the frequencies at which |Wt1| and
|Wt2| exceeded one.

• Systematic design, implementation, and evaluation. In other words, try it

and see. Clearly, closed-loop implementation of the controller design should
be the final arbiter of a suitable model. One approach is to design a series of
controllers, each with a smaller perturbation weight (i.e., assuming less
uncertainty), and implement these controllers on the physical system.
Setting the weight too high will give a conservative result as performance is
unnecessarily traded off against robust stability. A weight that is too small
will lead to a controller design that exhibits too little robustness, or is even

7-170
 A Process Control Example: Two Tank System

unstable, in practice. Balas and Doyle [BalDoy1, BalDoy2] give details of a

similar approach to a flexible structure problem.
At first it seems that this approach does not save any effort as we must
design and implement all potential controllers. However, it does provide
information about the appropriate level of uncertainty, and we may be able
to conduct simpler experiments (for example designing SISO controllers to
evaluate Wh1) to gain information about components within the system. This
involves less effort than attempting to tune all perturbation weights at once.
These issues have been considered in developing the perturbation weights. The
most important aspects of the two tank problem were considered to be:

• The nominal model for h1 is very accurate up to at least 0.3 Hz.

• Limit cycle experiments in the t1 loop suggest that uncertainty should
dominate above 0.02 Hz.
• There is about 180 degrees of additional phase lag in the t1 model
(at h1 = 0.75) at about 0.02 Hz. There is also a significant gain loss at this
frequency. These effects are the result of the unmodeled mixing dynamics.
• Limit cycle experiments in the t2 loop suggest that uncertainty should
dominate above 0.03 Hz.

It is not surprising that there is less uncertainty associated with tank 2; most
of the temperature uncertainty arises from the mixing dynamics and tank 2 is
somewhat smaller than tank 1. The perturbation weights are illustrated
graphically in Figure 7-84.

7-171
7 Robust Control Examples

Perturbation Weights for the two tank design

3
10

10
2 Wh1
Wt1
Wt2

1
10
Magnitude

0
10

1
10

2
10

3
10 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

Figure 7-84: Perturbation Weights for the Robust Control Design Problem

The output uncertainties Wh1 and Wt1 are given by (note that here we express
the t1 perturbation weight as a function of the steady state height),

0.5s
W h1 = 0.01 + ------------------------
0.25s + 1

20ĥ 1 s
W t1 = 0.1 + --------------------- .
0.2s + 1

The t2 perturbation weight is chosen as,

100s
W t2 = 0.1 + ---------------- .
s + 21

7-172
 A Process Control Example: Two Tank System

Sensor Noise Weights

The sensor dynamics are insignificant relative to the dynamics of the rest of the
system. This will not be true of the sensor noise. The potential sources of noise
include electronic noise in thermocouple compensators, amplifiers, and filters,
radiated noise from the stirrers, and poor grounding (this is, after all, an
inexpensive laboratory experiment).
We will model the noise by adding a weighted unknown input to h1, t1, and t2.
Smoothed FFT analysis has been used to estimate the noise on a quiescent
system. This gives the following weights.

Wh1noise = 0.01
Wt1noise = 0.03
Wt2noise = 0.03

The design considered in tankdemo.m uses measurements of only t1 and t2. The
weight Wh1noise is unnecessary for this case but is included so that you can
investigate different control configurations.
There are additional disturbances or noises associated with the measurements
that are not, strictly speaking, sensor noise. For example, in tank 2 the
imperfect mixing of the bias stream causes variations in the temperature
measurement. The inclusion of such noises here does not affect the
performance requirements of the controller.

Specifying the Design Requirements

It now remains to include in the interconnection structure (refer to Figure 7-82)
the weighting functions that specify performance. These are the reference
weights, error weights, and actuator weights. Arguably, noise weights and
disturbance weights also fall into those categories although we have chosen to
present the noise weights as part of the system model above. In a more general
setting weighted disturbances or more complex error specifications might also
be considered.
The weights should be considered as frequency dependent, relative weightings.
For example, it is the relative size of the reference weights and noise weights
that determines the extent to which the final controller design emphasizes
reference tracking error over sensor noise rejection.

7-173
7 Robust Control Examples

The error weights are first order low pass filters, with Wt1perƒ weighting the t1
tracking error and Wt2perƒ weighting the t2 tracking error. These are given by,

100
W t1perf = ----------------------
400s + 1
50
W t2perf = ----------------------.
800s + 1

We have selected a higher weight (better tracking performance) for t1 because

physical considerations lead us to believe that t1 will be easier to control than
t2.
The majority of the water flowing into tank 2 comes from tank 1, which means
that changes in t2 are dominated by changes in t1. It makes more sense to
express the reference weighting in terms of t1 and t2– t1. This allows us to
express the fact that t2 is normally commanded to a value close to t1. This is
done by using the following weighting approach.

t 1cmd W t 1cmd 0 w1
= 1 0 .
t 2cmd 1 1 0 W t diffcmd w 2

In this case,

W t 1cmd = 0.1

W t diffcmd = 0.01

Figure 7-85 graphically illustrates the error and command weighting

functions.

7-174
 A Process Control Example: Two Tank System

Perturbation Weights for the two tank design

3
10

Wt1perf
2
10 Wt2perf
Wt1cmd
Wtdiffcmd
1
10
Magnitude

0
10

1
10

2
10

3
10 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

Figure 7-85: Performance Weights for the Robust Control Design Problem

In the case of the actuation weights we would like to weight both the amplitude
and the rate of the actuator. This can be done by weighting ƒhc (and ƒcc) with a
function that rolls up at high frequencies. An alternative approach can be used
when we have a first order actuator model.
Using ƒhc as an example; the approach is to create an actuator model with ƒh
and dƒh/dt as outputs. These can then be separately weighted with constant
weights. Note that this approach has the advantage of reducing the number of
states in the interconnection structure, and hence in the final design.
Figure 7-86 illustrates the form of such a weighted actuator model.

7-175
7 Robust Control Examples

Figure 7-86: Model of the Flow Valve Actuator Including Magnitude and Rate
Weightings

The actuator bandwidth, denoted by BW in Figure 7-86, is,

BW = 20 radians/sec.
The weights selected for the tankdemo.m design were,

W hact = 0.01
W cact = 0.01
W hrate = 50
W crate = 50

Note that each weighted actuator model contributes only one state to the
·
interconnection structure, and allows independent weighting of ƒh and f h (and
·
ƒc and f c ).

Controller Design and Analysis

The tankdemo.m script runs through the design for the problem outlined above.
The script file is not discussed in detail as it roughly follows the layout of the
above material.

7-176
 A Process Control Example: Two Tank System

Design Issues
One issue regarding the design approach is worth noting. We begin by
designing for the nominal case. This is done by selecting out of the
interconnection structure the perturbation inputs and outputs (v and z). This
design achieved a closed-loop H∞ norm of γ = 0.9082. This is a lower bound on
the achievable value of µ for robust performance.
The response of the nominal controller (k0 in the script) is simulated, and is
shown in Figure 7-87. This is a check on the applicability of our performance
weights. In general, the inclusion of the robustness in the design problem will
have the effect of trading off this nominal performance in order to improve the
stability robustness. Simulating the nominal design gives a rough idea of what
the time domain performance implied by the weights.

7-177
7 Robust Control Examples

Design k0, step response

0.8

0.75

0.7
h1
t1
Measurements

0.65
t2

0.6

0.55

0.5

0.45
0 100 200 300 400 500 600 700 800
Time: seconds

Design k0, step response

0.8

0.7
fhc
fcc

0.6
Actuators

0.5

0.4

0.3

0.2
0 100 200 300 400 500 600 700 800
Time: seconds

The reference signal ramps (from 80 to 100 seconds) t1 from 0.75 to 0.57,
and t2 from 0.67 to 0.47.

Figure 7-87: Time Response for the Nominal H• Controller

7-178
 A Process Control Example: Two Tank System

The demo file, tankdemo, uses this interconnection structure for the robust
performance design. In the demo example three D – K iterations are performed.
In each of the D-scale fitting calls (performed with the function musynfit),
second order D-scale fits were selected. The final value of µ achieved is 1.73.
Further iterations may give further improvements; for this example it was
decided to stop at this point.

Closed-Loop Analysis
We will run through a typical series of analyses for our design and briefly
discuss the aspects of interest. The procedure guarantees closed-loop stability
(although it is still worth checking to make sure numerical problems have not
invalidated the design) but the controller need not be stable. It often is stable
and this is preferable for implementation purposes.
The frequency response of the controller is given in Figure 7-88. We note that
the controller rolls off around the frequency range where the uncertainties
start to become larger. This is what we would expect from a classical point of
view.
Note that there is not a great deal of low frequency gain. We will subsequently
see that this is because the noise level is high relative to the error weighting.
An identical simulation is run to give an idea of the loss of nominal
performance in the time domain. This is not intended as a comparison between
k0 and kmu. A more appropriate comparison would also include a simulation of
a perturbed system. The function dypert can be used to select the worst case
perturbation. The nominal response of kmu is shown in Figure 7-89. There is
little deterioration in the nominal response for controller kmu.

7-179
7 Robust Control Examples

Controller: kmu
1
10

0
10

1
10
Magnitude

2
10

3
10

4
10

5
10 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

Controller: kmu
300

200

100
Phase (degrees)

100

200

300

400 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

Figure 7-88: Frequency Response of the µ Synthesis Controller

7-180
 A Process Control Example: Two Tank System

Design kmu, step response

0.8

0.75

0.7
h1
t1
Measurements

0.65 t2

0.6

0.55

0.5

0.45
0 100 200 300 400 500 600 700 800
Time: seconds

Design kmu, step response

0.8

0.7
fhc
fcc

0.6
Actuators

0.5

0.4

0.3

0.2
0 100 200 300 400 500 600 700 800
Time: seconds

The reference signal ramps (from 80 to 100 seconds) t1 from 0.75 to 0.57,
and t2 from 0.67 to 0.47

Figure 7-89: Nominal Time Response for the kmu Controller

7-181
7 Robust Control Examples

The tradeoff between robustness and performance can be further studied by

independently calculating µ for robust stability, and the closed-loop nominal
performance. Note that robust stability is a µ test on the v to z block of the
closed-loop transfer function in Figure 7-82. Three 1 × 1 blocks are specified for
the robust stability block structure. The robust performance is given by the
maximum singular value of the w to e block of the closed-loop transfer function.
Figure 7-90 compares these with robust performance over the frequency range.

mu analysis: controller kmu

1.8

1.6 rob perf

rob stab
1.4 nom perf

1.2

0.8

0.6

0.4

0.2

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

Figure 7-90: Robust Performance, Robust Stability, and Nominal

Performance for the µ Synthesis Controller

Figure 7-90 illustrates that the nominal performance limits the low frequency
robust performance. Robust stability does not become an issue except around
0.01 Hz. This is not surprising as the perturbation weights begin to increase
around this frequency, yet the controller has only begun to roll off.
These issues can be further investigated by examining smaller subblocks of the
D-scaled closed-loop system. This form of analysis has been referred to as
M-analysis in the aerospace robust control community, simply because the
closed-loop system was habitually designated by M. It allows us to determine
which inputs and outputs are limiting the achievable robust performance. This

7-182
 A Process Control Example: Two Tank System

analysis is performed by calculating the µ upper bound, DMD–1, at each

frequency. We then look at the maximum singular values of the appropriate
subblocks. The example will serve as clarification.
The performance inputs, w, are divided into two groups:
• Reference commands
• Noise

There are three groups of performance outputs:

• Tracking errors
• Actuator penalties
• Actuator rate penalties

This give six 2 × 2 transfer function blocks, and for each of these we calculate
the maximum singular value as a function of frequency.
The results for the temperature inputs are shown in Figure 7-91. The
calculated nominal performance (maximum singular value of the w to e
transfer function) is an upper bound for each of the subblock calculations and
serves as a comparison point.
The temperature reference command to tracking error transfer function
dominates those shown in Figure 7-91. However it is only really significant,
with respect to the nominal performance, in the frequency range 0.001 to 0.01
Hz. It also has some contribution at lower frequencies. By contrast, the
actuator penalty has almost no effect on the design. We could change this
weight significantly without affecting the design. The actuator rate penalty
starts to influence the design at frequencies above 0.01 Hz.
Figure 7-92 illustrates the subblocks corresponding to the noise inputs. It is
immediately clear that the noise to tracking error transfer function dominates
the design at low frequencies. To improve the low frequency nominal
performance (and the low frequency robust performance) we must reduce the
noise weights. In practical terms this means buying higher quality sensors.

7-183
7 Robust Control Examples

DMDinv analysis for kmu: temp cmds > errors

2
cmds > errors
rob perf
nom perf
1

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

DMDinv analysis for kmu: temp cmds > actuator

2
cmds > actuator
rob perf
nom perf
1

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

DMDinv analysis for kmu: temp cmds > act rate

2
cmds > act rate
rob perf
nom perf
1

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

*robust performance and nominal performance are included for comparison

Figure 7-91: DMD–1 Analysis for Temperature Reference Command Inputs

7-184
 A Process Control Example: Two Tank System

DMDinv analysis for kmu: noise > errors

2
noise > errors
rob perf
nom perf
1

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

DMDinv analysis for kmu: noise > actuator

2
noise > actuator
rob perf
nom perf
1

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

DMDinv analysis for kmu: noise > act rate

2
noise > act rate
rob perf
nom perf
1

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

*robust performance and nominal performance are included for comparison

Figure 7-92: DMD–1 Analysis for Noise Inputs

7-185
7 Robust Control Examples

A similar analysis can be used to determine which aspects of the design are
limiting robust stability. In this case, there are three transfer functions to
examine; v1 to z1, etc. These are illustrated in Figure 7-93 and compared to
both robust performance and robust stability.

DMDinv analysis for kmu: h1 perturbation

2
h1 pert tf
rob perf
rob stab
1

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

DMDinv analysis for kmu: t1 perturbation

2
t1 perf tf
rob perf
rob stab
1

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz

DMDinv analysis for kmu: t2 perturbation

2
t2 pert tf
rob perf
rob stab
1

0 5 4 3 2 1 0 1
10 10 10 10 10 10 10
Frequency: Hertz
*robust performance and nominal performance are included for comparison

Figure 7-93: DMD–1 Analysis for the Perturbations

As we would expect, the h1 perturbation has little influence on the design.

Recall that the uncertainty associated with the h1 output is small compared to
the temperature outputs. We can also see that the t1 perturbation weight
dominates the robust stability. This is still not large compared with robust

7-186
 A Process Control Example: Two Tank System

performance, but if we wish to improve the robust stability, we must reduce

Wt1. In practical terms this means that we must invest more effort in modeling
and identifying some of the tank 1 mixing dynamics.
We can also apply this type of analysis at a single frequency. To do this we
would simply examine the mu upper bound for the closed-loop system at a
particular frequency. A column with large numbers would indicate that the
corresponding input was dominating the design. Similarly, a row of large
numbers would indicate that the corresponding output was dominating the
design.
This form of analysis is very useful in giving engineering information about the
limiting factors in the design. As we have illustrated above, it can be used to
indicate where additional effort (in terms of further modeling/identification, or
higher quality sensing) is required to improve robust performance.

Experimental Evaluation
We present an experimental comparison between a µ synthesis design and a
more standard loopshaping controller on the two tank system. The design
shown here is not identical to that given above as it was calculated and
implemented several years before µ-Tools was written. However, the method
and relative weightings were similar.
A simple SISO style loopshaping technique has been used to design controllers
(Kloop) for comparison purposes. One of the worst case plant conditions, (h1 =
0.75, t1 = 0.25), was selected as a nominal design point. The controller simply
inverts the plant to a diagonal loopshape. It is now well known that this
technique will not work well for high condition number plants, particularly
those with uncertainty at the input. Refer to the work of Skogestad et al.
[SkoMD] for further details on this point.
The loopshape chosen was

100
t 1 loop = ------------------------------
( 1 + 3000s )

Figure 7-94 shows the magnitude of the frequency response of Kloop.

100
t 2 loop = ------------------------------
( 1 + 5000s )

7-187
7 Robust Control Examples

10 3

10 2

10 1

10 0

10 -1

10 -2
10 -5 10 -4 10 -3 10 -2 10 -1 10 0 10 1

frequency: Hz

Figure 7-94: Magnitude Response of the Loopshaping Controller: Kloop

Both controllers have been tested over a wide range of commands. Figure 7-95
shows a typical command tracking response for the Kloop and Figure 7-96 shows
the same response for Kmu. The oscillatory behavior of the loopshaping
controller cannot be alleviated by selecting loopshapes that have a lower
crossover frequency. These merely produce lower frequency and higher
amplitude oscillations. It is the method of inverting the plant that leads to this
problem.

7-188
 A Process Control Example: Two Tank System

Kloop command response: plant outputs

1

0.9
h1
0.8

0.7
t1
0.6

0.5
t2
0.4

0.3

0.2

0.1

0
0 100 200 300 400 500 600 700 800

time: seconds

Kloop command response: actuators

1
fc command
0.9
fh
command
0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 100 200 300 400 500 600 700 800

time: seconds

Figure 7-95: Experimental Command Response: Controller Kloop with Input

Temperature Ramps Going from t1 = 0.75, t2 = 0.67 at 80 Seconds to
t1 = 0.55, t2 = 0.47 at 100 Seconds

7-189
7 Robust Control Examples

Kmu command response: plant outputs

1

0.9

0.8

0.7
t1
0.6
t2
0.5
h1
0.4

0.3

0.2

0.1

0
0 100 200 300 400 500 600 700 800

time: seconds

Kmu command response: actuators

1

0.9

0.8

0.7

0.6
fh command
0.5

0.4

0.3

0.2 fc command
0.1

0
0 100 200 300 400 500 600 700 800

time: seconds

Figure 7-96: Experimental Command Response: Controller Kmu with Input

Temperature Ramps Going from t1 = 0.75, t2 = 0.67 at 80 Seconds to t1 =
0.55, t2 = 0.47 at 100 Seconds

7-190
 A Process Control Example: Two Tank System

Two Tank System References

[SmDMS:] Smith, R.S., J. Doyle, M. Morari, and A. Skjellum, “A case study
using µ: Laboratory process control problem,” in Proc. Int. Fed. Auto. Control,
vol. 8, pp. 403–415, 1987.
[SmD1:] Smith, R.S, and J. Doyle, “The two tank experiment: A benchmark
control problem,” in Proc. Amer. Control Conf., vol. 3, pp. 403–415, 1988.
[Smith:] Smith, R.S., “Model validation for robust control: an experimental
process control application,” Automatica, Nov. 1995.
[AstH:] Åström, K.J., and T. Hägglund, “Automatic tuning of simple regulators
with specifications on phase and amplitude margins,” Automatica, vol. 20, no.
5, pp. 645–651, 1984.
[SmD2:] Smith, R.S., and J. C. Doyle, “Closed loop relay estimation of
uncertainty bounds for robust control models,” in Proc. of the 12th IFAC World
Congress, vol. 9, pp. 57–60, July 1993.
[VdHSB:] Van den Hof, P.M., R. J. Schrama, and O. H. Bosgra, “An indirect
method for transfer function estimation from closed loop data,” in Proc. IEEE
Control Decision Conf., pp. 1702–1706, 1992.
[BalD1:] Balas, G.J., and J. C. Doyle, “Identification of flexible structures for
robust control,” IEEE Control Sys. Magazine, vol. 10, pp. 51–58, June 1990.
[BalD2:] Balas, G.J., and J. Doyle, “Robustness and performance trade-offs in
control design for flexible structures,” IEEE Trans. Control Syst. Tech.,” vol. 2,
no. 4, pp. 352–361, 1994.
[SkoMD:] Skogestad, S., M. Morari, and J. C. Doyle, “Robust control of
ill-conditioned plants: High-purity distillation,” IEEE Trans. Auto. Control,
vol. 33, pp. 1092–1105, December 1988.

7-191
7 Robust Control Examples

7-192
 8

Reference
8 Reference

This chapter contains a detailed description of all µ-Analysis and Synthesis

Toolbox (µ-Tools) functions. It begins with a list of functions in alphabetical
order, followed by a list of functions grouped by subject area, and continues
with a detailed description of each command. Information on each function is
also available through the MATLAB on-line help facility.
In the summary of commands index and tables, the following abbreviations are
used:

• CONSTANT matrices are denoted by C

• SYSTEM matrices are denoted by S
• VARYING matrices are denoted by V

Therefore, if a command can be used with either a CONSTANT, SYSTEM or

VARYING matrix, it may be denoted by CSV.

8-2
 Summary of Commands

Summary of Commands

Command Description

abv Stack CSV matrices above one another

blknorm Block norm of a CV matrix
cf2sys Create a S from a normalized coprime factorization
cjt Conjugate transpose of a CSV matrix
cmmusyn Constant matrix µ synthesis
cos_tr Generate a cosine signal as a V matrix
crand Generate a complex random C matrix (uniform
distribution)
crandn Generate a complex random C matrix (normal
distribution)
csord Ordered complex Schur form
daug Diagonal augmentation of CSV matrices
dhfnorm Calculate discrete-time ∞-norm of a stable S matrix
dhfsyn Discrete-time H∞ control design
dkit Automated D – K iteration for µ synthesis
dkitgui Automated D – K iteration GUI for µ synthesis
drawmag Interactive moused-based sketch and fitting tool
dtrsp Discrete-time response of a linear S matrix
dypert Create a rational perturbation from frequency mu output
data
emargin Calculate normalized coprime factor robust stability
margin
fitmag Fit magnitude data with a real, rational, transfer function

8-3
8 Reference

Command Description

fitsys Fit frequency response data with a transfer function

frsp Frequency response of a S matrix
gap Calculate the gap metric for S matrices
genmu Generalized µ-analysis of CV matrices
genphase Generate a minimum phase function from magnitude
data
getiv Get the independent variable of a V matrix
h2norm Calculate 2-norm of a stable, strictly proper S matrix
h2syn H2 control design
hankmr Optimal Hankel norm approximation of a S matrix
hinffi H∞ full information control design
hinfnorm Calculate ∞-norm of a stable, proper S matrix
hinfsyn H∞ control design
hinfsyne H∞ control design, minimum entropy
indvcmp Compare independent variable data of two V matrices
madd Addition of CSV matrices
magfit Fit frequency response data with a transfer function
(batch)
massign Assign a portion of a matrix
mfilter Construct a Bessel, Butterworth, Chebychev, or RC filter
minfo Information on a CSV matrix
minv Inverse of CSV matrices
mmult Multiplication of CSV matrices

8-4
 Summary of Commands

Command Description

mprintf Formatted printing of a matrix

mscl Scale (by a scalar) a S or V matrix
msf An interactive D-scaling rational fit routine
msub Subtraction of CSV matrices
mu µ-analysis of CV matrices
muftbtch A batch D-scaling rational fit routine
musynfit An interactive D-scaling rational fit routine
musynflp An interactive D-scaling rational fit routine (linear
programming)
muunwrap Construct D-scaling and perturbation from mu
ncfsyn H∞ loopshaping control design
nd2sys Convert a SISO transfer function into a S matrix
negangle Calculate angle of CV matrices elements between 0 and –
2π
nugap Calculate the ν (nu) gap for S matrices
pck Create a S matrix from state-space data (A,B,C,D)
pkvnorm Peak norm of a V matrix
pss2sys Convert an [A B;C D] matrix into a µ-Tools S matrix
randel Generate a random perturbation for mu block structure
reordsys Reorder states of a S matrix
ric_eig Solve a Riccati equation via eigenvalue decomposition
ric_schr Solve a Riccati equation via real Schur decomposition
rifd Display real, imaginary, frequency, and damping data

8-5
8 Reference

Command Description

samhld Sample-hold approximation of a continuous S matrix

sbs Stack CSV matrices next to one another
sclin Scale S matrix input
scliv Scale the independent variable of V matrix
sclout Scale S matrix output
sdecomp Decompose a S matrix into two S matrices
sdhfnorm Calculate sample-data ∞-norm of a stable S matrix
sdhfsyn Sample-data H∞ control design
sdtrsp Time response of a sample-data S matrix
see Display SV matrices
seeiv Print independent variables of a V matrix
seesys Formatted SV display
sel Select rows/columns or outputs/inputs of CSV matrices
sfrwtbal Weighted balanced realization of a S matrix
sfrwtbld Stable weighted balanced realization of a S matrix
siggen Generate a signal as a V matrix
simgui A GUI for time simulations of LFTs
sin_tr Generate a sine signal as a V matrix
sisorat Fit a frequency point with a first order, all-pass, stable
function
sncfbal Balanced realization of coprime factors of a S matrix
sortiv Sort independent variable of a V matrix
spoles Poles of a S matrix

8-6
 Summary of Commands

Command Description

srelbal Stochastic balanced realization of a S matrix

sresid Residualize states of a S matrix
starp Redheffer star product
statecc Apply a coordinate transformation to S matrices
step_tr Generate a step signal as a V matrix
strans Bidiagonal coordinate transformation of S matrices
strunc Truncate states of a S matrix
sys2pss Extract the state-space matrix [A B; C D] from a S
matrix
sysbal Balanced realization of a S matrix
sysic System interconnection program
sysrand Generate a random S matrix
szeros Transmission zeros of a S matrix
tackon String together V matrices
transp Transpose of SV matrices
trsp Time response of a linear S matrix
tustin Prewarped continuous-time to discrete-time S
transformation
unum Input or column dimension of CSV matrix
unpck Extract state-space data (A,B,C,D) from a S matrix
unwrapd Construct D-scaling from mu
unwrapp Construct ∆ perturbation from mu
vabs Absolute value of a CV matrix

8-7
8 Reference

Command Description

var2con Convert a V matrix to a C matrix

varyrand Generate a random V matrix
vceil Round elements of CV matrices towards ∞
vcjt Conjugate transpace of CV matrices
vdcmate Decimate a V matrix
vdet Determinant of CV matrices
vdiag Diagonal of CV matrices
vebe Element-by-element operations on V matrices
veig Eigenvalue decomposition of CV matrices
veval Evaluate general functions of V matrices
vexpm Exponential of CV matrices
vfind Find individual elements of a V matrix
vfft FFT for V matrices
vfloor Round elements of CV matrices towards –∞
vifft Inverse FFT for V matrices
vimag Imaginary part of a CV matrix
vinterp Interpolate V matrices
vinv Inverse of a CV matrix
vldiv Left division of CV matrices
vnorm Norm of CV matrices
vpck Pack a V matrix
vpinv Pseudoinverse of a CV matrix
vplot Plot CV matrices

8-8
 Summary of Commands

Command Description

vpoly Characteristic polynomial of CV matrices

vrcond Condition number of a CV matrix
vrdiv Right division of CV matrices
vreal Real part of a CV matrix
vrho Spectral radius of a CV matrix
vroots Polynomial roots of CV matrices
vschur Schur form of a CV matrix
vspect Signal processing spectrum command for V matrices
vsvd Singular value decomposition of a CV matrix
vtp Transpose of CV matrices
vunpck Unpack a V matrix
vzoom Mouse-driven axis selection of plot window
wcperf Worst-case performance for given uncertainty level
wsgui A MATLAB workspace GUI
xnum State dimension of a S matrix
xtract Extract portions of a V matrix using independent
variables
xtracti Extract portions of a V matrix
ynum Output or row dimension of a CSV matrix
zp2sys Convert transfer function poles and zeros into a S matrix

8-9
8 Reference

Commands Grouped by Function

Standard Operations/Basic Functions

abv Stack CSV matrices above one another

cjt Conjugate transpose of SV matrices
daug Diagonal augmentation of CSV matrices
madd Addition of CSV matrices
minv Inverse of CSV matrices
mmult Multiplication of CSV matrices
mscl Scale (by a scalar) a SV matrix
msub Subtraction of CSV matrices
sbs Stack CSV matrices next to one another
sclin Scale S matrix input
sclout Scale S matrix output
sel Select CSV matrix rows/columns or outputs/inputs
starp Redheffer star product
transp Transpose of SV matrices

8-10
 Commands Grouped by Function

Matrix Information, Display and Plotting

drawmag Interactive mouse-based sketch and fitting tool

minfo Information on a matrix
mprintf Formatted printing of a matrix
rifd Display real, imaginary, frequency, and damping data
see Display SV matrices
seeiv Display independent variables of a V matrix
seesys Formatted SV display
unum Input or column dimension of a CSV matrix
vplot Plotting CV matrices
vzoom Mouse-driven axis selection of plot window
wsgui A MATLAB workspace GUI
xnum State dimension of a S matrix
ynum Output or row dimension of a CSV matrix

8-11
8 Reference

Modeling Functions

mfilter Construct a Bessel, Butterworth, Chebychev, or RC filter

nd2sys Convert a SISO transfer function into a µ-Tools S matrix
pck Create a S matrix from state-space data (A, B, C, D)
pss2sys Convert an [A B;C D] matrix into a µ-Tools S matrix
sys2pss Extract state-space matrix [A B; C D] from a S matrix
sysic System interconnection program
unpck Extract state-space data (A,B,C,D) from a S matrix
zp2sys Convert transfer function poles and zeros to a S matrix

SYSTEM Matrix Functions

reordsys Reorder states in a S matrix

samhld Sample-hold approximation of a continuous S matrix
spoles Poles of a S matrix
statecc Apply a coordinate transformation to S matrices
strans Bidiagonal coordinate transformation of S matrices
sysrand Generate a random S matrix
szeros Transmission zeros of a S matrix
tustin Prewarped continuous to discrete S transformation

8-12
 Commands Grouped by Function

Model Reduction Functions

cf2sys Create a S from a normalized coprime factorization

hankmr Optimal Hankel norm approximation of a S matrix
sdecomp Decompose a S matrix into two S matrices
sfrwtbal Frequency weighted balanced realization of a S matrix
sfrwtbld Stable frequency weighted realization of a S matrix
sncfbal Balanced realization of coprime factors of a S matrix
srelbal Stochastic balanced realization of a S matrix
sresid Residualize states of a S matrix
strunc Truncate states of a S matrix
sysbal Balanced realization of a S matrix

SYSTEM Response Functions

cos_tr Generate a cosine signal as a V matrix

dtrsp Discrete-time response of a linear S matrix
frsp Frequency response of a S matrix
sdtrsp Sample data time response of a linear S matrix
siggen Generate a signal as a V matrix
simgui A GUI for time simulations of LFTs
sin_tr Generate a sine signal as a V matrix
step_tr Generate a step signal as a V matrix
trsp Time response of a linear S matrix

8-13
8 Reference

H2 and H∞ Analysis and Synthesis Functions

dhfnorm Calculate discrete-time ∞-norm of a stable S matrix
dhfsyn Discrete-time H∞ control design
emargin Normalized coprime factor robust stability margin
gap Calculate the gap metric between S matrices
h2norm Calculate 2-norm of a stable, strictly proper S matrix
h2syn H2 control design
hinffi H∞ full information control design
hinfnorm Calculate ∞-norm of a stable, proper S matrix
hinfsyn H∞ control design
hinfsyne H∞ minimum entropy control design
ncfsyn H∞ loopshaping control design
nugap Calculate the ν (nu) gap between S matrices
pkvnorm Peak norm of a V matrix
sdhfnorm Sample-data ∞-norm of a stable S matrix
sdhfsyn Sample-data H∞ control design

8-14
 Commands Grouped by Function

Structured Singular Value (µ) Analysis and Synthesis

blknorm Block norm of CV matrices

cmmusyn Constant matrix µ synthesis
dkit Automated D – K iteration for µ synthesis
dkitgui Automated D – K iteration GUI for µ synthesis
dypert Create a rational perturbation from frequency mu data
fitmag Fit magnitude data with real, rational, transfer function
fitmaglp Fit magnitude data with real, rational, transfer function
fitsys Fit frequency response data with transfer function
genphase Generate a minimum phase frequency response to magnitude
data
genmu Real and complex generalized µ-analysis of CV matrices
magfit Fit magnitude data with real, rational, transfer function (a
batch process)
mu Real and complex µ-analysis of CV matrices
msf Interactive D-scaling rational fit routine
muftbtch Batch D-scaling rational fit routine
musynfit Interactive D-scaling rational fit routine
musynflp Interactive D-scaling rational fit routine (linear program)
muunwrap Construct D-scaling and ∆ perturbation from mu
randel Generate a random perturbation
sisorat Fit a frequency point with first order, all-pass, stable transfer
function
unwrapd Construct D-scaling from mu

8-15
8 Reference

Structured Singular Value (µ) Analysis and Synthesis

unwrapp Construct ∆ perturbation from mu

wcperf Worst-case performance for a given ∆

VARYING Matrix Manipulation

getiv Get the independent variable of a V matrix

indvcmp Compare the independent variable data of two V matrices
scliv Scale the independent variable of a V matrix
sortiv Sort the independent variable of a V matrix
tackon String together V matrices
var2con Convert a V matrix to a C matrix
varyrand Generate a random V matrix
vfind Find individual elements of a V matrix
vpck Pack a V matrix
vunpck Unpack a V matrix
xtract Extract portions of a V matrix
xtracti Extract portions of a V matrix using independent variable

8-16
 Commands Grouped by Function

Standard MATLAB Commands for VARYING Matrices

vabs Absolute value of a CV matrix

vceil Round elements of CV matrices towards ∞
vdet Determinant of CV matrices
vdiag Diagonal of CV matrices
veig Eigenvalue decomposition of CV matrices
vexpm Exponential of CV matrices
vfft FFT for V matrices
vfloor Round elements of CV matrices towards –∞
vifft Inverse FFT for V matrices
vimag Imaginary part of a CV matrix
vinv Inverse of a CV matrix
vnorm Norm of CV matrices
vpinv Pseudoinverse of a CV matrix
vpoly Characteristic polynomial of CV matrices
vrcond Condition number of a CV matrix
vreal Real part of a CV matrix
vroots Polynomial roots of CV matrices
vschur Schur form of a CV matrix
vspect Signal processing spectrum command for V matrices
vsvd Singular value decomposition of a CV matrix

8-17
8 Reference

Additional VARYING Matrix Functions

vcjt Conjugate transpose of CV matrices

vdcmate Decimate V matrices
vebe Element-by-element operations on V matrices
veval Evaluate general functions of V matrices
vinterp Interpolate V matrices
vldiv Left division of CV matrices
vrdiv Right division of CV matrices
vrho Spectral radius of a CV matrix
vtp Transpose of CV matrices

Utilities and Miscellaneous Functions

crand Complex random matrix generator (uniform distribution)

crandn Complex random matrix generator (normal distribution)
csord Order complex Schur form matrices
massign Assign a portion of a matrix to another
negangle Calculate angle of matrix elements between 0 and –2π
ric_eig Solve a Riccati equation via eigenvalue decomposition
ric_schr Solve a Riccati equation via real Schur decomposition

8-18
 abv, daug, sbs

Purpose 8abv, daug, sbs

Augment CONSTANT, SYSTEM and VARYING matrices

Syntax out = abv(mat1,mat2,...,matN)

out = daug(mat1,mat2,...,matN)
out = sbs(mat1,mat2,...,matN)

Description abv places the matrix mat1 above the matrix mat2. daug places the input
matrices on the diagonal of the output matrix. sbs places the input matrices
next to one another. All these commands, abv, daug, and sbs, allow the use of
multiple input arguments inputs (up to nine). CONSTANT, SYSTEM, and
VARYING matrices can be placed by one another based on the following table.
mat2
CONSTANT SYSTEM VARYING

CONSTANT yes yes yes

mat1
SYSTEM yes yes no
VARYING yes no yes

The input matrices must be compatible in the respective dimension in order for
the function to be performed. abv requires the same number of columns (inputs
for SYSTEM matrices) and sbs requires the input matrices to have the same
number of rows (outputs for SYSTEM matrices).

Pictorial Representation of Functions

8-19
abv, daug, sbs

Examples Create two CONSTANT matrices a and b along with two SYSTEM matrices p1
and p2. Examples of manipulation using abv, daug, and sbs are shown in the
following examples.
a = [1 2 3; 4 5 6];
b = [7 7 7; 8 8 8 ];
pl = pck(-10,1,10,0);
p2 = pck(-3,2,4,.1),
seesys(abv(p1,p2))

-1.0e+01 0.0e+00 | 1.0e+00

0.0e+00 -3.0e+00 | 2.0e+00
---------------------------------
0.0e+01 0.0e+00 | 0.0e+00
0.0e+01 4.0e+00 | 1.0e-00

out = abv(a,b,b)
out =

1 2 3
4 5 6
7 7 7
8 8 8
7 7 7
8 8 8

out = daug(a,b)
out =

1 2 3 0 0 0
4 5 6 0 0 0
0 0 0 7 7 7
0 0 0 8 8 8

out = sbs(p1,p2);
seesys = out

-1.0e+01 0.0e+00 | 1.0e+00 0.0e+00

0.0e+00 -3.0e+00 | 0.0e+00 2.0e+00
--------------------------------------------
1.0e+01 4.0e+00 | 0.0e+00 1.0e-01

8-20
 abv, daug, sbs

minfo(out)
system:2states1 outputs2 inputs
out = sbs(a,b,a)
out =
1 2 3 7 7 7 1 2 3
4 5 6 8 8 8 4 5 6

See Also madd, mmult, sel, vdiag

8-21
blknorm

Purpose 8blknorm
Create a matrix which is made up of norms of subblocks. Used in conjunction
with mu (µ)

Syntax matout = blknorm(matin,blk)

Description blknorm computes the maximum singular value of the subblocks of matin,
using the information in the perturbation block structure, blk. The output of
blknorm is matout whose entries are the maximum singular value of the
subblocks of matin with these norms as elements. This helps to show which
parts of the matrix are contributing to making µ large. A more complete
description of the perturbation block structure, blk, can be found with the
command mu and in Chapter 4. blknorm is best used on scaled matrices from
the upper bound for µ. Repeated δI blocks are treated the same way as full
blocks. The function blknorm can be applied to both CONSTANT and
VARYING matrices.

Examples Create a 4 × 3 random matrix and determine its subblock norms for two
different block structures. The first block structure consists of a two element
repeated block and a 1 × 2 full block. The second block structure consists of a 1
× 1 block, a 1 × 2 full block, and a 1 × 1 block.
m = crand(4,3);
disp(m)

0.7012 + 0.9826i 0.0475 + 0.0727i 0.7564 + 0.4364i

0.9103 + 0.7227i 0.7361 + 0.6316i 0.9910 + 0.7665i
0.7622 + 0.7534i 0.3282 + 0.8847i 0.3653 + 0.4777i
0.2625 + 0.6515i 0.6326 + 0.2727i 0.2470 + 0.2378i

disp(blknorm(m,[2 0; 1 2]))
1.8498 1.5272
1.6656 0.6923

mprintf(blknorm(m,[1 1; 1 2; 1 1]), '%6.2f')

1.21 0.09 0.87
1.58 1.35 1.39
0.70 0.69 0.34

8-22
 blknorm

Algorithm The maximum singular value of each block associated with the blk structure
is calculated via the MATLAB norm.

See Also mu, norm, vnorm

8-23
cjt, transp, vcjt, vtp

Purpose 8cjt, transp, vcjt, vtp

Transpose and conjugate transpose of CONSTANT, SYSTEM and VARYING
matrices

Syntax out = cjt(mat)

out = transp(mat)
out = vcjt(mat)
out = vtp(mat)

Description cjt forms the complex conjugate transpose of the input matrix mat and transp
forms the transpose of mat. transp outputs similar results to the MATLAB
command .′. These commands also work on SYSTEM and VARYING matrices.
For consistency in our naming convention, vcjt and vtp are the same
commands as cjt and transp, but work on just CONSTANT and VARYING
matrices.
For a SYSTEM matrix mat, transp, and cjt are defined as

mat =
A B
C D

transp(mat) =
A.’ C.’ , –A ’ –C ’
cjt(mat) =
B.’ D.’ –B ’ –D ’

8-24
 cjt, transp, vcjt, vtp

Examples Create a SYSTEM matrix and calculate its transpose and conjugate transpose
using cjt and transp.
A = [-10 0; 0 3];
B = [1 0 3; 0 2 -9];
C = [10 0; 0 4];
D = [0 -.2 -45; .82 0 .1];
out = pck(A,B,C,D);
seesys (out, '%5 .2g')

-10 0 | 1 0 3
0 3 | 0 2 -9
--------------------------------------
10 0 | 0 -.2 -45
0 4 | .82 0 .1

x = transp (out);
seesys (x, '%5 .2g')

-10 0 | 10 0
0 3 | 0 4
---------------------------
1 0 | 0 .82
0 2 | -.2 0
3 -9 | -45 .1

x = cjt(out);
seesys (x, '%5 .2g')

10 0 | -10 0
0 3 | 0 -4
---------------------------
1 0 | 0 .82
0 2 | -.2 0
3 -9 | -45 .1

8-25
cjt, transp, vcjt, vtp

Algorithm These functions call the MATLAB commands ′ and .′ consistent with the type
of input matrices.

See Also vdiag

8-26
 cmmusyn

Purpose Approximately solves the constant-matrix µ-synthesis problem, via the

8cmmusyn

minimization

r×t
minQ ∈ C µ ∆ ( R + UQV )

for given matrices R ∈ Cn×m, U ∈ Cn×r,V ∈ Ct×m, and a set ∆ ⊂ Cm×n.

Syntax [bnd,dvec,gvec,qopt] = cmmusyn(R,U,V,blk);

Algorithm This works for CONSTANT or VARYING data in R, U, and V. If two or more
matrices are VARYING, the independent variable values of these matrices
must be the same.
The approximation to solving the constant matrix µ synthesis problem is
two-fold: only the upper bound for µ is minimized, and the minimzation 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,
[PacZPB], and the global optimizer (for the upper bound for µ) is calculated.
The upper bound is returned in bnd, and the optimizing Q is returned in qopt.
The scaling matrices associated with the upper bound are in dvec and gvec and
may be unwrapped into block diagonal form using muunwrap.

Reference Packard, A.K., K. Zhou, P. Pandey, and G. Becker, “A collection of robust

control problems leading to LMI’s,” 30th IEEE Conference on Decision and
Control, pp. 1245–1250, Brighton, UK, 1991.

See Also genmu, mu

8-27
cos_tr, sin_tr, step_tr

Purpose 8cos_tr, sin_tr, step_tr

Generate a VARYING matrix containing a cosine, sine, or step signal at an
evenly spaced time interval

Syntax out = cos_tr(freq,mag,tinc,lastt)

out = sin_tr(freq,mag,tinc,lastt)
out = step_tr(timedata,stepdata,tinc,lastt)

Description cos_tr, sin_tr, and step_tr generate time signals for use with the trsp (time
response) command. The following are the input variables provided to cos_tr
and sin_tr.
Inputs to cos_tr and sin_tr:

freq frequency of the cosine (sine) signal (radians/second)

mag magnitude of signal
tinc time step (increment)
lastt final time (the signal starts at time t=0)

Output:

out time varying matrix of the cosine (sine) signal

Inputs to step_tr:

timedata time at which step occurs (vector)

stepdata magnitude of step (vector)
tinc time step (increment)
lastt final time (the signal starts at time t=0)

Output:

out time varying matrix of steps

8-28
 cos_tr, sin_tr, step_tr

Examples Synthesize a cosine signal with a 1 Hz frequency and of magnitude 1.5, 5

seconds in length. The time increment is .05 second.
out=cos_tr(2*pi,1.5,.05,5);
vplt('iv,d',out)
grid
title('Generate a cosine signal via cos_tr')
xlabel('time (seconds)')
ylabel('Magnitude')

Generate a cosine signal via cos_tr

1.5

0.5
Magnitude

-0.5

-1

-1.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

time (seconds)

Algorithm cos_tr and sin_tr call the MATLAB commands cos and sin, respectively.

See Also minfo, siggen, trsp, vpck

8-29
crand, crandn, sysrand, varyrand

Purpose 8crand, crandn, sysrand, varyrand

Generate a random complex CONSTANT, SYSTEM or VARYING matrix

Syntax out = crand(n,m)

out = crandn(n,m)
sys = sysrand(nstates,ninputs,noutputs,stabflag)
vary = varyrand(rdim,cdim,nindv,ivflg)

Description crand and crandn generate random complex matrices of dimensions n × m,

using the MATLAB rand and randn commands, respectively. crand elements
are uniformly distributed and crandn elements are normally distributed.
Inputs to crand and crandn:

n number of rows of the output matrix

m number of columns of the output matrix

Output from crand and crandn:

out an n × m complex matrix of random elements

sysrand generates a random SYSTEM matrix with nstates states, ninputs

inputs, and noutputs outputs. Setting the stabflag to 1 will result in a stable,
random SYSTEM. The default for stabflag is 0.
varyrand creates a random VARYING matrix with a specified number of rows
(rdim), columns (cdim), and independent variable values (nindv). The optional
argument ivflg sorts the independent variables to be monotonically
increasing if it is set to 0 (default). Otherwise, if ivflg is set to a nonzero value,
no sorting is done.

8-30
 crand, crandn, sysrand, varyrand

Examples Create a CONSTANT, complex random matrix and a random SYSTEM matrix
using the commands crand and sysrand
crand(4,3)
ans =
0.4764 + 0.1622i 0.9017 + 0.1351i 0.4103 + 0.4523i
0.3893 + 0.0711i 0.4265 + 0.7832i 0.1312 + 0.8089i
0.2033 + 0.3653i 0.1420 + 0.4553i 0.8856 + 0.9317i
0.0284 + 0.2531i 0.9475 + 0.3495i 0.0922 + 0.6516i

sys=sysrand(2,4,1);
seesys(sys)

0.2190 0.6789 | 0.9347 0.5194 0.0346 0.5297

0.0470 0.6793 | 0.3835 0.8310 0.0535 0.6711
------------------|----------------------------------------
0.0077 0.3834 | 0.0668 0.4175 0.6868 0.5890

See Also minfo, pck, pss2sys, rand, randel, sys2pss, unpck, vpck, vunpck

8-31
csord

Purpose 8csord
Compute an ordered, complex Schur form matrix

Syntax [v,t,flgout,reig_min] = csord(m,epp,flgord,flgjw,flgeig)

Description The csord function produces an ordered, complex Schur form matrix of the
input CONSTANT square matrix m with

t 11 t 12
v′ * m * v = t =
0 t 22

The MATLAB function schur is called, which results in an unordered Schur

form matrix. The subroutine cgivens forms a complex Givens rotation matrix,
which orders the t matrix as you define it. The v matrix is the transformation
matrix. A series of optional input flags can be set.

epp user-supplied zero tolerance (default epp = 0)

flgord=0 order eigenvalues in ascending real part (default)

flgord=1 partial real part ordering, with real parts less than zero first,
then the jω axis eigenvalues and finally the real parts greater
than zero
flgjw=0 no exit condition on eigenvalue location (default)

flgjw=1 exit if abs(real(eigenvalues(i)))< epp

flgeig=0 no exit condition on half-plane eigenvalue distribution (default)

flgeig=1 exit if length(real(eigenvalues)>0) = length(real(eigenvalues)<0)

The output flag flgout is nominally 0. flout is set to 1 if there are jω-axis
eigenvalues, set to 2 if there are an unequal number of positive and negative
eigenvalues, or set to 3 if both conditions occur. The fourth output argument,
reig_min, is the minimum, magnitude real part of the eigenvalues of m.

8-32
 csord

The ric_schr routine calls csord to solve for a stabilizing solution to a matrix
Riccati equation. In this case, the m matrix has a special structure, and failure
modes are flagged to avoid extra, unnecessary computations.

Algorithm The eigenvalues are reordered by iterating through each of them and
interchanging them via a bubble sort based on the input flag, flgord. The
subroutine cgivens exchanges the out of order eigenvalues.

Reference Golub, G.H. and C.F. Van Loan, Matrix Computations, The Johns Hopkins
University Press, 1983.

See Also cgivens, ric_schr, rsf2csf, schur

8-33
dhfnorm

Purpose 8dhfnorm
dhfnorm computes the H∞ gain of a stable, discrete-time SYSTEM matrix

Syntax out = dhfnorm(sys,ttol,h,iiloc)

Description dhfnorm computes the H∞ norm of a stable, discrete-time SYSTEM. It converts

the SYSTEM to continuous-time via a bilinear transformation and then calls
the routine hinfnorm.
The method uses the bilinear transformation from the z-plane to the s-plane
given by,

1 + sh ⁄ 2 2z – 1
z = ----------------------- , s = ----------------
1 – sh ⁄ 2 hz + 1

where h is the time between samples. This is a transformation between the

unit disk and the left half plane. If the system p has a pole close to z = –1, then
a preliminary output feedback is used to move such poles before the bilinear
transformation is performed.
Input arguments:
sys SYSTEM matrix (discrete-time), CONSTANT or VARYING
matrix
ttol relative tolerance of accuracy for SYSTEM matrices only,(default
= 0.001)
h time between samples (default 1)
iiloc initial estimate of worst case frequency (optional)

Output arguments:
out a 1 × 3 vector giving a lower bound, upper bound and
the frequency where the lower bound occurs.

Reference 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. 831–847, August 1989.

See Also hinfsyne, hinffi, hinfnorm, hinfsyn, h2syn, h2norm, ric_eig, ric_schr,
sdhfnorm, sdhfsyn

8-34
 dhfsyn

Purpose dhfsyn computes an H∞ controller for a discrete-time SYSTEM interconnection

8dhfsyn

matrix

Syntax [k,g,gfin,ax,ay,hamx,hamy] = dhfsyn(p,nmeas,ncon,...

gmin,gmax,tol,h,z0,quiet,ricmethd,epr,epp)

Description dhfsyn calculates a discrete-time H∞ controller that achieves the infinity norm
gfin for the interconnection structure p. The controller, k, stablizes the
discrete-time SYSTEM matrix p and has the same number of states as p. The
SYSTEM p is partitioned

A B1 B2
p = C 1 D 11 D 12
C 2 D 21 D 22

where B1 are the disturbance inputs, B2 are the control inputs, C1 are the
errors to be kept small, and C2 are the output measurements provided to the
controller. B2 has column size (ncon) and C2 has row size (nmeas).
The closed-loop system is returned in g. The same bilinear transformation
method described for dhfnorm is used. The controller k is returned that
minimizes the entropy integral,

gfin π
2 –2
1 – z0
∫
–2 jθ jθ
I = – ----------- log det ( I – gfin g ( e )′ g ( e ) ) ------------------------------- dθ
2π – π jθ –1 2
e – z0

The program calls the continuous-time routine hinfsyne and the

corresponding conditions and tests need to be satisfied.

8-35
dhfsyn

Input arguments
p SYSTEM interconnection structure matrix, (stable, discrete
time)
nmeas number of measurements output to controller
ncon number of control inputs
gmin lower bound on γ
gmax upper bound on γ
tol relative difference between final γ values
h time between samples (optional)
z0 point at which entropy is evaluated (default ∞)
quiet controls printing on the screen
1. no printing
0. header not printed
–1. full printing (default)
ricmethod 1.Eigenvalue decomposition (with balancing)
–1. Eigenvalue decomposition (without balancing)
2. Schur decomposition (with balancing, default)
–2. Schur decomposition (without balancing)
epr measure of when a real part of an eigenvalue of the
Hamiltonian matrix is zero (default epr = 1e–10)
epp positive definite determination of the X∞ and Y∞ solution
(default epp = 1e–6)

Output arguments:

k H∞ (sub) optimal controller (discrete time)

g closed-loop system with H∞ controller (discrete time)
gfin final γ value associated with k and g
ax X∞ Riccati solution as a VARYING matrix with independent
variable γ

8-36
 dhfsyn

ay Y∞ Riccati solution as a VARYING matrix with independent

variable γ
hamx X∞ Hamiltonian matrix as a VARYING matrix with
independent variable γ
hamy Y∞ Hamiltonian matrix as a VARYING matrix with
independent variable γ

Note that the outputs ax, ay, hamx, and hamy correspond to the equivalent
continuous-time problems and can also be scaled and/or balanced.
The dhfsyn program outputs several variables, which can be checked to ensure
that the above conditions are being met. For each γ value the minimum
magnitude, real part of the eigenvalues of the X Hamiltonian matrices is
displayed along with the minimum eigenvalue of X∞, which is the solution to
the X Riccati equation. A # sign is placed to the right of the condition that failed
in the printout. This additional information can aid you in the control design
process.

Algorithm dhfsyn uses the above bilinear transformation to continuous-time and then the
formulae described in the Glover and Doyle paper for solution to the optimal
H∞ control design problem.
Subroutines called. hinfsyne, hinf_st, hinf_gam, hinfe_c:
hinf_gam calls ric_eig, ric_schr, csord, and cgivens

Reference 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. 831–847, August 1989.
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, vol. 11, pp. 167–172, 1988.

See Also hinfsyne, hinffi, hinfnorm, hinfsyn, h2syn, h2norm, ric_eig, ric_schr,
sdhfnorm, sdhfsyn

8-37
dkit

Purpose A script file for µ synthesis via D–K iteration

8dkit

Syntax dkit

Description dkit is a µ-Tools script file for D–K iteration. The D–K iteration procedure is
an approximation to µ synthesis control design. It involves a sequence of
minimizations, first over the controller variable K (holding the D variable
associated with the scaled µ 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 µ value, but often works well in
practice. A detailed description of the D–K iteration can be found in Chapter 5.
dkit automates the D–K iteration procedure but requires the initialization of
several variables. The file dk_defin.m is an example of the information
required by dkit. You can copy this file from the µ-Tools subroutine directory
mutools/subs and modify it for your application. This file can also be renamed.
After renaming, assign the variable DK_DEF_NAME in the MATLAB workspace
to the (character string) name of the new file containing the user-defined
variables for dkit. For example, if the filename containing the setup data is
himat_def.m, then

DK_DEF_NAME = 'himat_def';

should be executed in the MATLAB workspace.

The following is a list of what occurs during a single, complete step of the D–K
iteration.

1 Upon running dkit, the program prompts you for starting D–K iteration
number.
Starting mu iteration #:
Type 1 to indicate the first D–K iteration.
2 (In the 1st iteration, this step is skipped.) The µ calculation (from the
previous step) provides a frequency-dependent scaling matrices, Df. The
fitting procedure is interactive (msf), and fits these scalings with rational,
stable transfer function matrices, D̂ ( s ) .
After fitting, plots of

–1
σ ( D f ( jω )F L ( P, K ) ( jω )D f ( jω ) )

8-38
 dkit

and

–1
σ ( D̂ ( jω )F L ( P, K ) ( jω )D̂ f ( jω ) )

are shown for comparison.

3 (In the 1st iteration, this step is skipped.) The rational D̂ is absorbed into
the open-loop interconnnection for the next controller synthesis. Using
either (based on GMAX_DK_PLAN, see below) the previous
frequency-dependent D’s or the just-fit rational D̂ , an estimate of an
appropriate value for GMAX_DK 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 D’s).
4 (In the 1st iteration, this step is skipped.) The parameters that will be used
in the upcoming hinfsyn are displayed. It is your option to change any/all/
none of these.
5 (The 1st iteration begins at this point.) A controller is designed using H∞
synthesis on the scaled open-loop interconnection.
a The progress of the γ-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 (OMEGA_DK). If you
change it, all relevant frequency responses are automatically
recomputed.
d You are given the option to rerun the H∞ synthesis with modified hinfsyn
parameters. This is convienient if, for instance, the bisection tolerance
was too large, or if GMAX_DK was too small.
6 Using the block structure BLK_DK, bounds for the structured singular value
of the closed-loop system are calculated and plotted.
7 An iteration summary is displayed, showing all of the controller order, as
well as the peak value of µ of the closed-loop frequency responses.
8 The choice of stopping, or performing another iteration is given.

Subsequent iterations proceed along the same lines without the need to
re-enter the iteration number. A summary at the end of each iteration is
updated to reflect data from all previous iterations. This often provides

8-39
dkit

valuable information about the progress of the robust controller synthesis

procedure.
To continue iterating on a problem that was started earlier, make sure the
relevent data is in the workspace, run dkit and type the number of the
iteration you would like to begin.
The following is a list of the variables that you must set (in either dk_defin file
or the file defined by DK_DEF_NAME) and their meanings.
NOMINAL_DK Nominal plant interconnection structure, a µ-Tools
SYSTEM matrix.
NMEAS_DK Number of sensor measurements.
NCONT_DK Number of control inputs.
BLK_DK Block structure forµ calculation, used by mu.
OMEGA_DK Frequency response range.

The following is a list of the optional variables that may be set (in either
dk_defin file or the file defined by DK_DEF_NAME) and their meanings.

GMIN_DK Lower bound for first H∞ controller design

GMAX_DK Upper bound for first H∞ controller design
GTOL_DK H∞ bisection tolerance for first iteration
GMAX_DK_PLAN Estimates for GMAX_DK in subsequent iterations are made
in three manners. The value of GMAX_DK_PLAN determines
which formula is used:
0 Use previous GMAX_DK
1 use 1.2*pkvnorm(peak_mu_value)
2 use hinfnorm(mmult(dsysL,clp,minv(dsysR)))

8-40
 dkit

A number of variables are saved in the workspace after each iteration. Some of
these variables are required every iteration, hence, it doesn’t make sense to
recompute them. The other variables are outputs from the D–K iteration
procedure. The variables saved after each iteration are
bnds_dk(i) Frequency domain upper and lower bounds for µ
associated with the ith iteration. The (i) denotes the ith
iteration which is augmented to the name by the program
dkit.
dl_dk(i) Left state-space D-scale associated with ith iteration. The
(i) denotes the ith iteration which is augmented to the
name by the program dkit. Hence, dl_dk5 would be the
left state-space D-scale from the fifth iteration.
dr_dk(i) Right state-space D-scale associated with ith iteration.
Same notation.
Dscale_dk(i) D-scaling data output from mu associated with ith
iteration. Dscale_dk(i) is in compressed form. Same
notation.
gf_dk(i) The H∞ norm of the ith iteration closed-loop system. Same
notation.
k_dk(i) Controller from the ith iteration. Same notation.
nom_dk_g Frequency response of NOMINAL_DK using OMEGA_DK.
sens_dk(i) Sensitivity data output from the ith iteration mu
calculation. Same notation.

Fitting D-Scalings
The D-scale fitting procedure is interactive and uses the µ-Tools command msf.
During step 2 of the D–K iteration procedure, you are prompted to enter your
choice of options for fitting the D-scaling data. After pressing return, 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

8-41
dkit

i Increment Fit Order

d Decrement Fit Order
apf Auto-PreFit
mx 3 Change Max-Order to 3
at 1.01 Change Auto-PreFit tol to 1.01
0 Fit with zeroth order
2 Fit with second order
n Fit with n'th order
e Exit with Current Fittings
s 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 δI) 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 fit, and the order of the current
fit (with d for data, when no fit exists).
• The order of the current fit can be incremented or decremented (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 µ upper bound is to approximate the actual µ upper bound in
a norm sense. Setting at 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 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 will display a status of the current and fits.

8-42
 dkit

Examples An example of using dkit for D–K iteration is provided in the “HIMAT Robust
Performance Design Example” in Chapter 7.

Reference 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, pp. 370–377, March-April, 1994.
Balas, G.J., A.K. Packard and J.T. Harduvel, “Application of µ-synthesis
techniques to momentum management and attitude control of the space
station,” AIAA Guidance, Navigation and Control Conference, New Orleans,
August 1991.
Doyle, J.C., Doyle, K. Lenz, and A. Packard, “Design examples using
µ-synthesis: Space shuttle lateral axis FCS during reentry,” NATO ASI Series,
Modelling, Robustness, and Sensitivity Reduction in Control Systems, vol. 34,
Springer-Verlag, 1987.
Packard, A., J. Doyle, and G. Balas, “Linear, multivariable robust control with
a µ perspective,” ASME Journal of Dynamic Systems, Measurement and
Control, 50th Anniversary Issue, vol. 115, no. 2b, pp. 310–319, June 1993.
Stein, G., and J. Doyle, “Beyond singular values and loopshapes,” AIAA
Journal of Guidance and Control, vol. 14, no. 1, pp. 5–16, January, 1991.

See Also hinfsyn, hinfnorm, msf, mu

8-43
dkitgui

Purpose A graphical user interface for µ synthesis via D–K iteration

8dkitgui

Syntax dkitgui

Description dkitgui is a graphical user interface (GUI) for D–K iteration. The D–K
iteration procedure is an approximation to µ synthesis control design. It
involves a sequence of minimizations, first over the controller variable K
(holding the D variable associated with the scaled µ 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 µ value, but
often works well in practice. A more detailed description of the D–K iteration
can be found in Chapter 5.
The GUI tool, dkitgui, has five windows. They are:

• Main Iteration window, which is the main interface for the user during the
iteration.
• Setup window, where initial data is entered.
• Parameter window, which is occasionally used to modify properties of the
D–K iteration, such as H∞ parameters, and to select the variables that are
automatically exported to the workspace each iteration.
• Frequency Response window, where the plots of µ and σ of the closed-loop
transfer function matrix are displayed.
• Scaling window, where the rational fits of the frequency-dependent D-scale
data are shown, and can be modified.

dkitgui completely automates the D–K iteration procedure. A detailed

description of dkitgui and its use can be found in Chapter 5.

Examples An example of using dkitgui for D–K iteration is provided in the “HIMAT
Robust Performance Design Example” section in Chapter 7.

Reference 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, pp. 370–377, March-April, 1994.
Balas, G.J., A.K. Packard and J.T. Harduvel, “Application of µ-synthesis
techniques to momentum management and attitude control of the space

8-44
 dkitgui

station,” AIAA Guidance, Navigation and Control Conference, New Orleans,

pp. 1204–1213, August 1991.
Doyle, J.C., Doyle, 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, 1987.
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, pp. 310–319, June 1993.
Stein, G., and J. Doyle, “Beyond singular values and loopshapes,” AIAA
Journal of Guidance and Control, vol. 14, no. 1, pp. 5–16, January, 1991.

See Also dkit, hinfsyn, hinfnorm, mu, pkvnorm

8-45
drawmag

Purpose 8drawmag
Provide an interactive mouse-based log-log sketch and fitting tool

Syntax [sysout,pts] = drawmag(in,init_pts)

Description drawmag interactively uses the mouse in the plot window to create a VARYING
matrix pts and a stable, minimum-phase SYSTEM sysout, which
approximately fits, in magnitude, the frequency VARYING matrix in pts.
Input arguments:

in either a VARYING matrix which is plotted each time as a

reference or a CONSTANT matrix of the form [xmin xmax ymin
ymax] specifying the plot window on the data.
init_pts optional VARYING matrix of initial set of points

Output arguments:
sysout SYSTEM matrix fitting pts.
pts VARYING matrix 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 will be plotted when the fitting,
windowing, or replotting modes are 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 0-9 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
a second point at the new crosshair location. These two points define a new

8-46
 drawmag

window on the data, which is immediately replotted. This is useful in fine

tuning parts of the data. Windowing may be called 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. Caution
should be exercised when using this option, as it can wreak havoc on the
program if variables are changed.

See Also axis, ginput, magfitlp, musynfit, vplot

8-47
dypert, sisorat

Purpose 8dypert, sisorat

dypert operates on output from a VARYING mu (µ) calculation to construct a
worst-case, real-rational, stable perturbation
sisorat constructs a first-order, real-rational, stable, all-pass transfer
function that interpolates a particular complex number at a given, positive
frequency.

Syntax pert = dypert(pvec,blk,bnds);

pert = dypert(pvec,blk,bnds,blkindex);
ratfit = sisorat(value);

Description The input to dypert consists of the perturbation pvec, the block structure blk,
and lower and upper bounds, bnds, produced from a VARYING matrix µ
calculation. (See the µ-Tools command mu for a more complete description of
these variables.) By searching in bnds, dypert finds the peak value of the lower
bound — for example γ, occurring at frequency ω0. dypert then extracts the
perturbation (call this matrix ∆0) from pvec at the frequency ω0. dypert
constructs a SYSTEM matrix, pert, which is stable, and has the block-diagonal
structure associated with blk, and also satisfies the equations

1
pert ( jω 0 ) = ∆ 0 and pert ∞ = ---
γ

The command dypert can also be called with four input arguments. In this
case, the last argument is a vector of integers. This vector, blkindex, specifies
which blocks in the perturbation structure specified by the blk vector, are to be
used to construct the rational perturbation. For instance, if the block structure
specified by blk is

 2×3 
 diag [ ∆ 1, δ 2 I 4 × 4, δ 3 I 2 × 2, ∆ 4, ∆ 5, δ 6 I 3 × 3, ∆ 7 ] : ∆ 1 ∈ C ,
∆ :=  
 δ ∈ C, δ ∈ C, ∆ ∈ C 4 × 2, ∆ ∈ C 3 × 3, δ ∈ C, ∆ ∈ C 2 × 1 
 2 3 4 5 6 7 

8-48
 dypert, sisorat

and the fourth argument to dypert is [3 6 4 1] (indicating first, third, fourth,

and sixth blocks are desired), then the output SYSTEM matrix produced by
dypert will be a SYSTEM matrix that looks like

∆1 ( s ) 02 × 2 02 × 2 02 × 3
0 2 × 3 δ 3 ( s )I 2 × 2 0 2 × 2 02 × 3
04 × 3 04 × 3 ∆4 ( s ) 04 × 3
03 × 3 03 × 3 0 3 × 2 δ 6 ( s )I 3 × 3

This is useful when some of the perturbations are not physical perturbations,
but correspond to performance objectives, and hence do not need to be
constructed. Note that regardless of the order in which the numbers occur in
the fourth argument, the perturbations in the output SYSTEM pert are in
ascending order. When dypert is called with three arguments, all the
perturbations are constructed.
The command sisorat is essentially a scalar version of dypert, and is the main
subroutine for dypert. The input to sisorat is value, a 1 × 1, VARYING
matrix, with one independent variable value. The independent variable is
interpreted as a frequency, ω0, and the numerical value of value at that
frequency is denoted by γ. The output of sisorat is a single-input/single-output
stable, real, SYSTEM matrix, ratfit, satisfying

ratfit ( jω 0 ) = γ and ratfit ( jω ) = γ ∀ω

Algorithm The main subroutine of dypert is sisorat, which operates on the following
fact. For any complex number γ, and any real frequency ω0 > 0, there is a real
number β > 0 such that by proper choice of sign, the equality

s–β
± γ ------------ = γ
s + β s = jω
0

holds. This is exactly how sisorat works.

In order to understand how dypert works, recall in Chapter 4, the section
“Complex Structured Singular Value,” Remark 1 after Definition 2.1. There it
was shown that each full block of a perturbation that causes singularity can in

8-49
dypert, sisorat

fact be chosen to be a dyad, and the program mu does this at each independent
variable value. Hence, the matrix ∆0, as described above, is a diagonal
augmentation of scalar blocks, and dyads. Given that the ith block is a dyad,
write it as ∆ 0 i = y i x *i for complex vectors yi and xi. Using several calls to
sisorat, it is possible to create two stable, rational vectors h(s) (column) and
r(s) (row), such that each element of the vectors is of the form generated by
sisorat (stable, and flat across frequency), and for each k and l

h k ( jω 0 ) = y i ( jω 0 ) , r l ( jω 0 ) = x i ( jω 0 )
k l

ˆ
If we define ∆ i ( s ) := h ( s )r ( s ) then it is stable, and

ˆ ˆ
∆ i ∞ = σ ( ∆0i ) , ∆ i ( jω 0 ) = ∆ 0 i

Performing this on a block-by-block basis produces the entire rational

perturbation. For the case of a repeated scalar block, that part of the
perturbation is constructed directly with one call to sisorat and diagonally
augmented as many times as needed.

See Also mu, randel, svd, vsvd

8-50
 fitmag, genphase, fitmaglp, magfit

Purpose 8fitmag, genphase, fitmaglp, magfit

Fit single-input/single-output magnitude data with a real, rational, minimum
phase transfer function

Syntax sys = fitmag(magdata,weight,heading,oldfit,...

dmdi,upbd,blk,blknum)
resp = genphase(d)
sys = fitmaglp(magdata,weight,heading,oldfit,...
dmdi,upbd,blk,blknum)
[sys,fit] = magfit(magdata,dim,weight)

Description fitmag fits a stable, minimum phase transfer function to magnitude data,
magdata, with a supplied frequency domain weighting function, weight. Both
of these are VARYING matrices, with identical independent variable values.
fitmag uses genphase to generate phase data, and fitsys to do the fit.

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.
fitmaglp has the same inputs, outputs, and user interaction as fitmag , but
uses a linear programming approach to do the fitting instead of fitsys and
genphase.

fitmag and fitmaglp have the additional input arguments dmdi, upbd, blk,
and blknum. These arguments are used exclusively with D–K iteration when
called by musynfit and musynflp. In this case, the magdata is the dvec output
of the mu program and corresponds to the blknum’th frequency varying D scale
to be fit. weight corresponds to a measure of the sensitivity of mu to changes in
the D scales at each frequency. This is the sens output from mu. heading is a
string variable denoting the title of the plot and oldfit is usually the D scalings
from the previous D–K iteration.
dmdi represents the VARYING matrix analyzed using mu. upbd is the upper
bound calculated using mu of dmdi with the perturbation block structure blk.
The last argument, blknum, corresponds to the current D scale in the block
structure being fit with fitmag or fitmaglp. Upon fitting the magnitude data,
magdata, the resulting transfer function sys is absorbed into the original
matrix dmdi and plotted along with the mu upper bound on the lower graph.

8-51
fitmag, genphase, fitmaglp, magfit

magfit is a batch version of fitmaglp that eliminates the user interaction. The
weight is optional, but dim is a required argument of parameters for the linear
program. dim has the form [hmax htol nmin nmax] where

• hmax is a measure of the allowable error in the fit.

• htol is a measure of the accuracy with which the optimization is carried out.
• nmin and nmax are the minimum and maximum orders considered for the
curve fit.

Examples Create a second-order transfer function sys to test fitmag. Fit its magnitude
data with a first- and second-order transfer function via fitmag.
sys = nd2sys([1 -5 12],[1 2 7]);
w = logspace(-2,2,200);
sysg = frsp(sys,w);
wgt = 0.2;
wgtg = frsp(wgt,w);
sysfit = fitmag(vabs(sysg),wgtg);
ENTER ORDER OF CURVE FIT or 'drawmag' 1

CURVE FITTING, W/ORDER = 1

10 1

10 0

10 -1
10 -2 10 -1 10 0 10 1 10 2
1) data 2) newfit

10 0 weight for fit

10 -1
10 -2 10 -1 10 0 10 1 10 2
ENTER NEW ORDER, ’drawmag’, or NEGATIVE NUMBER TO STOP

8-52
 fitmag, genphase, fitmaglp, magfit

A first-order fit does not accurately respresent the frequency data as shown in
the above figure. The solid line represents the curve fit and the dashed line
represents the original frequency data. You can try and fit the data again with
a second-order system.
ENTER ORDER, 'drawmag', or NEGATIVE TO STOP 2
CURVE FITTING, W/ORDER = 2
10 1

10 0
10 -2 10 -1 10 0 10 1 10 2
1) data 2) newfit

10 0 weight for fit

10 -1
10 -2 10 -1 10 0 10 1 10 2
ENTER NEW ORDER, ’drawmag’, or NEGATIVE NUMBER TO STOP

ENTER ORDER, 'drawmag', or NEGATIVE TO STOP -1

The second-order fit lies directly on top of the original data, hence it is difficult
to distinguish the two plots. A -1 is entered at the end of this iteration
procedure to indicate satisfaction with the results.

Algorithm The algorithm for fitmag is as follows. On a log-log scale, the magnitude data
is interpolated linearly, with a very fine discretization. Then, using the
complex cepstrum algorithm, the phase, associated with a stable, minimum
phase, real, rational transfer function with the same magnitude as the magdata
variable is generated. This involves two fft’s, and logarithmic/exponential
conversions. With the new phase data, and the input magnitude data, the
MATLAB function invfreqs is used to find a real, rational transfer function
that fits the data. heading is an optional title and oldfit is an optional

8-53
fitmag, genphase, fitmaglp, magfit

previous fit that can be added to the graphs if they are included. These options
are used in the program musynfit.
The algorithm for magfit is as follows. The system sys is derived by solving a
linear program and searching over a parameter h according to the following
specification. Let m be the given magnitude data, g the transfer function of sys
and w the values of weight; then h is found such that at each frequency,
g < r,
1 ⁄ r < ----
-
m

where
2
r = 1+h⁄w .

The order of sys is increased until an h less than hmax is obtained or nmax is
reached. The minimum value of h at this order is then determined to an
accuracy of htol.

Problems For problems with very coarse data, fitmag may give incorrect answers, even
if the data was generated by taking the frequency response of a linear system.
The inaccuracy arises in the log-log interpolation step, which is used in the
phase calculation. This step is avoided in magfit. Hence, for coarse data sets,
magfit should be used. fitmaglp and magfit appear to be sometimes slower
than fitmag.

Reference Oppenheim, A.V., and R.W. Schaffer, Digital Signal Processing, Prentice Hall,
New Jersey, 1975, pp. 513.

See Also fitsys, invfreqs, musynfit, musynflp, muftbtch

8-54
 fitsys

Purpose 8fitsys
Fits single-input/single-output (SISO), single-input/multi-output (SIMO) and
multi-input/single-output (MISO) frequency response data with SYSTEM
matrix

Syntax sys = fitsys(resp,ord,wt,code)

Description fitsys fits frequency response (VARYING) data in resp with a transfer
function of order ord, using a frequency dependent weight in wt (optional). The
frequency response data may be either a row (SIMO) or column (MISO). The
optional frequency dependent weight is a VARYING matrix. This weight may
be a scalar (1 row, 1 column), or may be the same shape as resp.
The fourth argument, code, is optional. If set to 0 (default), then the fit is as
described. If code = 1, as in the µ-synthesis routines, it forces the fit to be stable,
minimum phase, simply by reflecting the poles and zeros if necessary. In this
case, the response resp comes from the program genphase and already
corresponds to a stable, minimum phase transfer function. fitsys is called by
fitmag and msf.

Examples An example of how to use fitsys to derive a SIMO transfer function model is
provided in the “More Sophisticated SYSTEM Functions” section in Chapter 2.

Algorithm The author of fitsys is Xin Hua Yang.

See Also fitmag, genphase, msf, vspect

8-55
frsp

Purpose 8frsp
Calculate the complex frequency response of a linear system

Syntax out = frsp(sys,omega,T,balflg)

Description frsp calculates the complex frequency response of a given SYSTEM matrix
(sys) for a vector of frequency points (omega). The output matrix out is a
frequency dependent VARYING matrix containing the frequency response of
the input system sys at the frequency values contained in the vector omega .
For systems with multiple inputs and outputs, a multivariable frequency
response is returned.
Input arguments:
sys SYSTEM matrix
response calculated at these frequencies. If another VARYING
matrix is input here, then its independent variables are used
T 0 (default) indicates a continuous system. A nonzero value forces
discrete system evaluation with sample time T (optional)
balflg 0 (default) balances the SYSTEM A matrix prior to evaluation. A
nonzero value for balflg leaves the state-space data unchanged
(optional)

Output arguments
out VARYING frequency response matrix

The vector of frequency points is assumed to be real and can be generated from
the MATLAB command logspace or linspace. Given a continuous system sys,
of the form

sys =
A B
CD

and an input vector, omega , with N frequencies, [ω1, ω2,. . .,ωN], frsp evaluates
the following equation

8-56
 frsp

C(jωiI – A)–1B + D, i = 1,. . .,N

You can specify a discrete time evaluation by specifying an optional sampling
time, T. For the discrete time case each matrix in the VARYING output is given
by

jω i T –1
C(e – A ) B + D , i = 1, … , N

Note that setting T = 0 implies that a continuous frequency response is to be

performed and not to evaluate

jω i 0 –1
C(e – A) B + D

Examples The SYSTEM matrix sys is constructed to have two inputs and two outputs
with poles at –2 and –10. A frequency vector omega is constructed with 30
points log spaced between .1 and 100 rad/s. The complex frequency response of
sys is calculated and its values between 3.5 and 4.6 rad/s are displayed.

a = [-2 0;0 -10];b = [.2 .12; -.3 .4];c = [.3 .7; 2 -1];
sys = pck(a,b,c);
omega = logspace(-1,2,30);
sysg = frsp(sys,omega);
see(xtract(sysg,3.5,4.6))
2 rows 2 columns

iv = 3.56225

-0.0114 - 0.0062i 0.0292 - 0.0165i

0.0746 - 0.0949i -0.0067 - 0.0386i

iv = 4.52035

-0.0125 - 0.0032i 0.0262 - 0.0172i

0.0577 - 0.0853i -0.0136 - 0.0294i

8-57
frsp

A frequency response is performed using frsp with the default variables set. A
plot of the frequency response is shown with the four line types corresponding
to the sysg(1,1), sysg(1,2), sysg(2,1), and the sysg(2,2) elements.
vplot('bode',sysg);
title('Complex frequency response example - continuous time')

Complex frequency response example - continuous time

10 0
Log Magnitude

10 -1

10 -2

10 -3
10 -1 10 0 10 1 10 2
Frequency (radians/sec)

4
Phase (radians)

-2

-4
10 -1 10 0 10 1 10 2
Frequency (radians/sec)

To demonstrate the calculation of a discrete frequency response, convert this

system into a digital system via the bilinear transformation. The sample
frequency is chosen as 100 radians/second.
T = 2*pi/100;
dsys = tustin(sys,T);
omega = sort([omega,logspace(0,2,60)]);
dsysg = frsp(dsys,,T);
vplot('bode',sysg);
title('Complex frequency response example - discrete time')

For digital filter design, you can examine the transfer function from 0 to π by
specifying T = 1. A Chebyshev type II filter is designed and its magnitude is
plotted to demonstrate this feature.

8-58
 frsp

Complex frequency response example - discrete time

10 0

10 -1
Log Magnitude

10 -2

10 -3

10 -4

10 -5
10 -1 10 0 10 1 10 2
Frequency (radians/sec)

4
Phase (radians)

-2

-4
10 -1 10 0 10 1 10 2
Frequency (radians/sec)

[a,b,c,d] = cheby2(11,30,0.3);
dfilt = pck(a,b,c,d);
omega = [0:pi/100:pi*99/100];
dsysg = frsp(dfilt, omega,1);
vplot('iv,lm',dsysg);
xlabel('frequency on unit circle')
ylabel('Magnitude')
title('Complex frequency response on the unit circle')

Algorithm The algorithm to calculate the complex frequency response involves an matrix
inverse problem, which is solved via a Hessenberg matrix. If balflg is set to 0,
the frequency response balances the SYSTEM A matrix (using the MATLAB
balance command) prior to calculation of the Hessenberg form.

Note Balancing the system may cause errors in the frequency response. If
the output of frsp is questioned, compare the results with balancing and
without balancing the SYSTEM prior to calculating the frequency response.

8-59
frsp

Complex frequency response on the unit circle

10 0

10 -1
Magnitude

10 -2

10 -3

10 -4
0 0.5 1 1.5 2 2.5 3 3.5

frequency on unit circle

Reference Laub, A.J., “Efficient Multivariable Frequency Response Computations,” IEEE

Transactions on Automatic Control, vol. AC–26, No. 2, pp. 407–408, April,
1981.

See Also balance, hess, samhld, tustin, vplot

8-60
 gap, nugap

Purpose 8gap, nugap

gap calculates the gap metric between two SYSTEM matrices

nugap calculates the ν gap between two SYSTEM matrices.

Syntax gap = gap(sys1,sys2, ttol)

nugap = nugap (sys1, sys2, ttol)

Description gap and nugap compute the gap and ν gap metrics between two SYSTEM
matrices. Both quantities give a numerical value δ(G0,G1) between 0 and 1 for
the distance between a nominal system sys1 (G0) and a perturbed system sys2
(G1). The gap metric was introduced into the control literature by Zames and
El-Sakkary, 1980, and exploited by Georgiou and Smith, 1990. The ν 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 δ(G0,G1)– arcsin δ(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 ν 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
–1 –1
needs to be replaced by W2GW1 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 “Loop Shaping Using H• Synthesis” section in Chapter 3.
Model reduction of the system model and controller can be performed by using
balanced truncations or Hankel norm approximation of normalized coprime
factor representations.
ttol defines the tolerance to which the gap is computed. The default is 0.001.

8-61
gap, nugap

Algorithm Tryphon Georgiou and Malcolm Smith wrote the gap program.
The computation of the gap amounts to solving 2-block H∞-problems, Georgiou,
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.

Reference Georgiou, T.T., On the computation of the gap metric, Systems Control Letters,
vol. 11, pp. 253–257, 1988.
Georgiou, T.T., and M. Smith, “Optimal robustness in the gap metric,” IEEE
Transactions on Automatic Control, vol. 35, pp. 673–686, 1990.
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), pp. 1350–1371,
1990.
Qiu, L., and E.J. Davison, “Feedback stability under simultaneous gap metric
uncertainties in plant and controller,” Systems Control Letters, vol. 18–1, pp.
9–22, 1992.
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, pp. 380–385, Oct., 1980.

See Also dhfnorm, emargin, hinfnorm, ncfsyn, mu

8-62
 genmu

Purpose 8genmu
Compute upper bounds for the mixed (real and complex) generalized
structured singular value (referred to as generalized mixed µ) of a VARYING/
CONSTANT matrix

Syntax [bnd,dvec,gvec,qmat] = genmu(M,C,blk);

Description Generalized µ allows us to put additional constraints on the directions that I –

M∆ becomes singular. Given a matrix M ∈ Cn×n, and C ∈ Cm×n, find the
smallest ∆ ∈ ∆ (described by blk) such that

I – ∆M
C

is not full column rank. Specifically, define

1
µ ∆ ( M, C ) := -------------------------------------------------------------------------------------------------------
 
min  σ ( ∆ ) : δ ∈ ∆, rank I – ∆M < n 
 C 

This quantity can be bounded above, using standard µ ideas. If there exists a
matrix Q such that
µ∆(M + QC) < β
then µ∆(M,C) < β. Hence,

µ ∆ ( M, C ) min µ ∆ ( M + QC )
n×m
Q∈C

It is possible to compute the optimal matrix Q which minimizes the standard µ

upper bound for µ∆(M + QC). The optimization problem can be reformulated
into an affine matrix inequality, [PacZPB], and solved with a combination of
heuristics and general purpose AMI solvers. This is how genmu computes the
upper bound µ∆(M,C).
The upper bound for µ∆(M,C) is returned in bnd. The scaling matrices D and G
associated with the upper bound are in packed format, in the matrics dvec and
gvec, and can be unwrapped with muunwrap. The Q matrix which mimimizes

8-63
genmu

the bound is returned in qmat. If either M or C are VARYING matrices, then

the bound is computed at each value of the independent variable, and the
output matrices (bnd, dvec, gvec and qmat) are also VARYING matrices.

Reference Pachard, A., K. Zhou, P. Pandey, and G. Becker, “A collection of robust control
problems leading to LMI’s,” 30th IEEE Conference on Decision and Control, pp.
1245–1250, Brighton, UK, 1991.

See Also cmmusyn, mu

8-64
 getiv, sortiv, tackon

Purpose 8getiv, sortiv, tackon

Return, sort, or append the independent variable values of a VARYING
matrix

Syntax [indv,err] = getiv(mat)

getiv(mat)
[out,err] = sortiv(mat,sortflg,nored,epp)
out = tackon(mat1,mat2)

Description getiv returns the independent variable values of the VARYING matrix mat. If
the input matrix mat is a VARYING matrix, the independent variable is
returned as a column vector, indv, and the output err is set to 0. If mat is not
a VARYING matrix, then indv is set to empty, and err is set to 1.
sortiv will reorder the independent variable and associated VARYING matrix
to be monotonically increasing or decreasing. The optional sortflg is set to 0
(default) for monotonically increasing sorting or nonzero for monotonically
decreasing sorting. sortiv can be used in conjunction with tackon to mesh
together two different VARYING matrices. The optional third input argument,
nored, is set to 0 (default) which does not reduce the number of independent
variables even if there are repeated ones. Setting nored to a nonzero value
causes repeated independent variables to be collapsed down if their
corresponding matrices are the same. If they are not, an error message is
displayed and only the first independent variable and corresponding matrix is
kept. The output argument err, which is nominally 0, is set to 1 if an error
message is displayed. The optional fourth input argument, epp, is a vector used
for checking closeness of two variables. If two independent variables are within
epp(1), and the norm of the difference between the two matrices at these points
is within epp(2), sortiv collapses these two independent variable values down
to one. If the two independent variables are within epp(1), and the norm
condition is not satisfied, an error message is displayed and out is set to the
null matrix. When nored is nonzero, the default value for epp is [1e – 9;1e – 9].
tackon strings together two VARYING matrices placing mat1 on top of mat2.
mat1 and mat2 must have the same row and column dimensions.

8-65
getiv, sortiv, tackon

Examples The frequency response VARYING matrix from the frsp example is a
two-input/two-output matrix containing 30 points. These independent
variables vary from 0.1 rad/sec to 100 rad/sec.
minfo(sysg)
varying: 30 pts2 rows2 cols
seeiv(sysg)

1.000e-01 1.269e-01 1.610e-01 2.043e-01 2.593e-01

3.290e-01 4.175e-01 5.298e-01 6.723e-01 8.532e-01
1.083e+00 1.374e+00 1.743e+00 2.212e+00 2.807e+00
3.562e+00 4.520e+00 5.736e+00 7.279e+00 9.237e+00
1.172e+01 1.487e+01 1.887e+01 2.395e+01 3.039e+01
3.857e+01 4.894e+01 6.210e+01 7.880e+01 1.000e+02

Typing getiv without any arguments outputs a brief description of its calling
sequence. All µ-Tools commands have this feature. The xtract command
selects the independent variables between 1 and 5 rad/sec and the getiv
command returns these independent variables from sysg and stores them in
indv.

getiv
usage: [indv,err] = getiv(mat)
[indv] = getiv(xtract(sysg,l,5));
indv'
ans =
1.0826 1.3738 1.7433 2.2122 2.8072
3.5622 4.5204
The sortiv command (with an optional second argument) resorts the
independent variable of the frequency response of sys in decreasing order.
syslg = sortiv (xtract ( sysg ,1 , 5 ) , 1 ) ;
seeiv(syslg)

4.520e+00 3.562e+00 2.807e+00 2.212e+00 1.743e+00

1.374e+00 1.083e+00

Algorithm getiv and sortiv manipulate VARYING matrices.

See Also indvcmp, sort, xtract, xtracti

8-66
 h2norm, hinfnorm

Purpose Calculate the H2, H∞ norms of a SYSTEM matrix

8h2norm, hinfnorm

Syntax out = h2norm(sys)

out = hinfnorm(sys,tol)

Description h2norm calculates the 2-norm of a stable, strictly proper SYSTEM matrix. The
output is a scalar, whose value is the 2-norm of the system.
The output from hinfnorm is a 1 × 3 vector, out, which is made up (in order) of
a lower bound for ||sys||∞, an upper bound for ||sys||∞, and a frequency, ωo, at
which the lower bound is achieved.

out(1) = σ(sys ( j ⋅ out ( 3 ) ) ) ≤ sys ∞ ≤ out(2)

The ||⋅||∞ norm calculation is an iterative process and requires a test to stop. The
variable tol specifies the tolerance used to calculate the ||sys||∞. The iteration
stops when
(the current upper bound) ≤ (1 + tol) × (the current lower bound).
The default value of tol is 0.001.

Algorithm The H2 norm of a SYSTEM follows from the solution to the Lyapunov equation.
AX + XA′ + BB′ = 0,
with ||sys||2 = trace (CXC′).
Calculation of the H∞ norm requires checking for jω axis eigenvalues of a
Hamiltonian matrix, Hα, which depends on a parameter α. If Hα has no jω axis
eigenvalues, then the ||⋅||∞ norm of the SYSTEM matrix is less than α. If the
matrix Hα does have jω axis eigenvalues, then these occur at the frequencies
where the transfer matrix has a singular value (not necessarily the maximum)
equal to α. By iterating, the value of the ||⋅||∞ norm can be obtained.

8-67
h2norm, hinfnorm

Reference Boyd, S., K. Balakrishnan and P. Kabamba, “A bisection method for computing
the H∞ norm of a transfer matrix and related problems,” Math Control Signals
and Systems, 2(3), pp. 207–219, 1989.
Boyd, S., and K. Balakrishnan, “A regularity result for the singular values of a
transfer matrix and a quadratically convergent algorithm for computing its H∞
norm,” Systems and Control Letters, vol. 15–1, 1990.
Bruinsma, O., and M. Steinbuch, “A fast algorithm to compute theH∞ norm of
a transfer function matrix,” Systems and Control Letters, vol. 14, pp. 287–293,
1990.

See Also hinfsyn, h2syn, ric_eig, ric_schr

8-68
 h2syn

Purpose Compute the optimal H2 controller given a SYSTEM interconnection matrix

8h2syn

Syntax [k,g,norms,kfi,gfi,hamx,hamy] =h2syn(p,nmeas,ncon,ricmethod)

Description h2syn calculates the H2 optimal controller k and the closed-loop system g for
the linear fractional interconnection structure p. nmeas and ncon are the
dimensions of the measurement outputs from p and the controller inputs to p.
The optional fourth argument, ricmethod, determines the method used to solve
the Riccati equations. The interconnection structure, p, is defined by

A B1 B2
p = C 1 D 11 D 12
C 2 D 21 D 22

Input arguments:

p SYSTEM interconnection structure matrix

nmeas number of measurements output to controller
ncon number of control inputs
ricmethd 1 Eigenvalue decomposition with balancing
–1 Eigenvalue decomposition with no balancing
2 Schur decomposition with balancing. (default)
–2 Schur decomposition with no balancing.

Output arguments:

k H2 optimal controller
g closed-loop system with optimal controller
norms norms of four different quantities, full information control cost
(FI), output estimation cost (OEF), disturbance feedforward cost
(DFL) and full control cost (FC), norms = [FI OEF DFL FC];
kfi full information/state feedback control law
gfi full information/state feedback closed-loop system

8-69
h2syn

hamx H2 Hamiltonian matrix

hamy H2 Hamiltonian matrix

The equations and corresponding nomenclature are taken from the Doyle, et
al., 1989, reference. The full information cost is given by the
1
′ ---
equation ( trace ( B 1 X 2 B 1 ) ) 2 . The output estimation cost is given
1
′ --- ′ ′
by ( trace ( F 2 Y 2 F 2 ) ) 2 , where F 2 =: – ( B 2 X 2 + D 12 C 1 ) . The disturbance
1
′ --- ′ ′
feedforward cost is ( trace ( L 2 X 2 L 2 ) ) 2 , where L2 is defined by – ( Y 2 C 2 + B 1 D 21 )
1
′ ---
and the full control cost 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.
The H2 solution provides an upper bound on γ for use in the hinfsyn program.

Examples Design an H2 optimal controller for a system matrix, himat_icn, with two
sensor measurements (nmeas), two error signals, two actuator inputs (ncont),
and eight states. himat_icn differs from the SYSTEM interconnection
structure himat_ic by the fact that the D11 term of himat_ic is set to be zero.
The Schur decompostion method, ricmethd = 2, will be used for solution of the
Riccati equations. The program outputs the minimum eigenvalue of X2 and Y2
during the computation.
nmeas = 2;
ncont = 2;
ricmethd = 2;
minfo(himat_icn)
system:8 states6 outputs6 inputs
[k,g] = h2syn(himat_icn,nmeas,ncont,ricmethd);
minimum eigenvalue of X2: 2.260000e-02
minimum eigenvalue of Y2: 2.251670e-02

8-70
 h2syn

The H∞ and H2 norm of the resulting closed-loop system g can be calculated via
the commands hinfnorm and h2norm.
hinfnorm(g)
norm between 2.787 and 2.79
achieved near 29.9
h2norm(g)
1.594e+01

Algorithm h2syn is an M-file in µ-Tools that uses the formulae described in the Doyle, et
al., 1989, reference for solution to the optimal H2 control design problem. A
Hamiltonian is formed and solved via a Riccati equation (ric_eig and
ric_schr). The D matrix associated with the input disturbances and output
errors is restricted to be zero.

Subroutines called. ric_eig, ric_schr, csord, and cgivens.

Reference 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. 831–847, August 1989.
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. 167–172, August 1989.

See Also hinfsyn, hinffi, h2norm, hinfnorm, ric_eig, ric_schr

8-71
hinffi

Purpose Compute H∞ full-information controller for a SYSTEM interconnection matrix

8hinffi

Syntax [k,g,gfin,ax,hamx] =
hinffi(p,ncon,gmin,gmax,tol,ricmethd,epr,epp)

Description hinffi calculates an H∞ full information controller that achieves the infinity
norm gfin for the interconnection structure p. The controller, k, stablizes the
SYSTEM matrix p and is constant gain. The system p is partitioned

A B1 B2
p =
C 1 D 11 D 12

where B1 are the disturbance inputs, B2 are the control inputs, and C1 are the
errors to be kept small. B2 has the column size ncon. Within the hinffi
program, the SYSTEM matrix p is augmented with state and disturbance
measurements; i.e., the identity matrix with size equal to the number of states
of p and the identity matrix with size equal to the number of disturbances. Be
careful when closing the loop with the full information controller since the
extra measurements are only augmented inside the command hinffi. The
internal system used for control design is

A B1 B2
C 1 D 11 D 12

I 0 0
0 I 0

The controller is returned in k and the closed-loop system is returned in g. The

program provides a γ iteration using the bisection method. Given a high and
low value of γ, gmax and gmin, the bisection method is used to iterate on the
value of γ in an effort to approach the optimal full information H∞ control
design. If gmax = gmin, only one γ value is tested. The stopping criteria for the
bisection algorithm requires the relative difference between the last γ value
that failed and the last γ value that passed be less than tol. You can select
either the eigenvalue or Schur method with or without balancing for solving

8-72
 hinffi

the Riccati equations. The eigenvalue method is faster but can have numerical
problems, while the Schur method is slower but generally more reliable.
The algorithm employed requires tests to determine whether a solution exists
for a given γ value. epr is used as a measure of when the Hamiltonian matrix
has imaginary eigenvalues and epp is used to determine whether the Riccati
solution is positive semi-definite. The selection of epr and epp should be based
on your knowledge of the numerical conditioning of the interconnection
structure p. The conditions checked for the existence of a solution are

• H Hamiltonian matrix must have no jω-axis eigenvalues

• the stabilizing solution, X∞, of the associated Riccati equation must exist,
and be must be positive, semi-definite.

Input arguments:
p SYSTEM interconnection structure matrix
ncon number of controller outputs
gmin lower bound on γ
gmax upper bound on γ
tol relative difference between final γ values, iteration stopping
criteria
ricmethd 1 Eigenvalue decomposition with balancing.
–1 Eigenvalue decomposition with no balancing
2 Schur decomposition with balancing (default)
–2 Schur decomposition with no balancing
epr measure of when a real eigenvalue of the Hamiltonian matrix
is zero (default epr = 1e–10, optional)
epp positive definite determination of the X∞ solution (default
epp = 1e–6, optional)

8-73
hinffi

Output arguments:
k H∞ full information controller
g closed-loop system with H∞ full information controller
gfin final γ achieved
ax Riccati solution as a VARYING matrix with independent
variable γ
hamx Hamiltonian matrix as a VARYING matrix with independent
variable γ

Note that the outputs ax, ay, hamx, and hamy may correspond to scaled or
balanced data. The following assumptions are made in the implementation of
the hinfsyn algorithm and must be satisfied.
(i) (A,B2) is stabilizable
(ii) D12 is full column rank

A – jωI B 2
(iii) has full column rank for all ω.
C 1 D 12

where ω denotes the frequency variable.

On return, there must be no jω-axis eigenvalues associated with the H
Hamiltonian matrices and the eigenvalues of the Riccati solution, X∞, must all
be ≥ 0, for the closed-loop system to be stable and to have an H∞ norm less than
γ. The bisection algorithm iterates on the value of γ to approach the optimal H∞
full information controller.
The hinffi program outputs several variables, which can be checked to ensure
that the above conditions are being met. For each γ value the minimum
magnitude, real part of the eigenvalues of the H Hamiltonian matrices is
displayed along with the minimum eigenvalue of X∞, which is the solution to
the Riccati equation. A # sign is placed to the right of the condition that failed
in the printout.

Examples Given an interconnection structure sys with one control input, it is desired to
synthesize a full information controller. The upper bound on γ is 1.0 and the
lower bound is specified as 0.1. A tolerance of 0.02 is selected for the stopping

8-74
 hinffi

condition for the γ iteration and the Schur method is used to solve the Riccati
equations. The command hinffi outputs the display shown for each value of γ.
The final γ value achieved is 0.2547.
ncont = 1;% number of control inputs
gmin = .1;% minimum gamma value to be tested
gmax = 1;% maximum gamma value to be tested
tol = .02;% tolerance on the gamma stopping value
ricmethd = 2;% Riccati solution via the Schur method
seesys(p)% plant interconnection structure
p=

3 1 | 4
0 0 | 1
-------------
1 0 | 0

[k,g,gf,ax,hx] = hinffi(p,1,.1,1,.02,2);
Test bounds:0.1000< gamma<= 1.0000

gamma ham_eig x_eig pass/fail

1.000 4.90e+00 5.27e-01 pass
0.550 4.66e+00 6.03e-01 pass
0.325 3.94e+00 1.06e+00 pass
0.213 1.69e+00 -7.63e-01# fail
0.269 3.34e+00 2.94e+00 pass
0.241 2.78e+00 -4.55e+00# fail
0.255 3.10e+00 1.04e+01 pass

Gamma value achieved:0.2547

Algorithm hinffi uses the formulas similar to the ones described in the Glover and Doyle,
1988 paper for solution to the H∞ control design problem. See the hinfsyn
command for more information.

Subroutines called. hinffi_t, hinffi_p, hinffi_c, and hinffi_g

hinffi_g calls: ric_eig, ric_schr, csord, and cgivens

8-75
hinffi

Reference 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. 831–847, August 1989.
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, vol. 11, pp. 167–172, 1988.

See Also h2syn, h2norm, hinfsyn, hinfsyne, hinfnorm, ric_eig, ric_schr

8-76
 hinfsyn

Purpose Compute an H∞ controller for a SYSTEM interconnection matrix

8hinfsyn

Syntax [k,g,gfin,ax,ay,hamx,hamy] =
hinfsyn(p,nmeas,ncon,gmin,gmax,tol,ricmethd,epr,epp)

Description hinfsyn calculates an H∞ controller, which achieves the infinity norm gfin for
the interconnection structure p. The controller, k, stabilizes the SYSTEM
matrix p and has the same number of states as p. The SYSTEM p is partitioned

A B1 B2
p = C 1 D 11 D 12
C 2 D 21 D 22

where B1are the disturbance inputs, B2 are the control inputs, C1 are the errors
to be kept small, and C2 are the output measurements provided to the
controller. B2 has column size (ncon) and C2 has row size (nmeas).
The closed-loop system is returned in g. The program provides a γ iteration
using the bisection method. Given a high and low value of γ, gmax and gmin, the
bisection method is used to iterate on the value of γ in an effort to approach the
optimal H∞ control design. If the value of gmax is equal to gmin, only one γ value
is tested. The stopping criteria for the bisection algorithm requires the relative
difference between the last γ value that failed and the last γ value that passed
be less than tol. You can select either the eigenvalue or Schur method the
Riccati equations. The eigenvalue method is faster but can have numerical
problems, while the Schur method is slower but generally more reliable.
The algorithm employed requires tests to determine whether a solution exists
for a given γ value. epr is used as a measure of when the Hamiltonian matrix
has imaginary eigenvalues and epp is used to determine whether the Riccati
solutions are positive semi-definite. The conditions checked for the existence of
a solution are:

• H and J Hamiltionian matrices (which are formed from the state-space data
of P and the γ level) must have no imaginary-axis eigenvalues.
• the stabilizing Ricatti solutions X∞ and Y∞ associated with the Hamiltionian
matrices must exist and be positive, semi-definite.
• spectral radius of (X∞,Y∞) must be less than or equal to γ2.

8-77
hinfsyn

The selection of epr and epp should be based on your knowledge of the
numerical conditioning of the interconnection structure p. The following
assumptions are made in the implementation of the hinfsyn algorithm and
must be satisfied.
(i) (A,B2) is stabilizable and (C2,A) detectable.
(ii) D12 and D21 have full rank.
A – jωI B 2
(iii) has full column rank for all ω ∈ R.
C 1 D 12

A – jωI B 1
(iv) has full row rank for all ω ∈ R.
C 2 D 21

Inputs arguments:

p SYSTEM interconnection structure matrix

nmeas number of measurements output to controller
ncon number of control inputs
gmin lower bound on γ
gmax upper bound on γ
tol relative difference between final γ values
ricmethod 1 Eigenvalue decomposition with balancing
2 Schur decomposition with balancing (default)
–2 Schur decomposition with no balancing
epr measure of when a real part of an eigenvalue of the
Hamiltonian matrix is zero (default epr = 1e–10)
epp positive definite determination of the X∞ and Y∞ solution
(default epp = 1e–6)

8-78
 hinfsyn

Output arguments:
k H∞ (sub) optimal controller
g closed-loop system with H∞ controller
gfin final γ achieved
ax X∞ Riccati solution as a VARYING matrix with independent
variable γ
ay Y∞ Riccati solution as a VARYING matrix with independent
variable γ
hamx H∞ Hamiltonian matrix as a VARYING matrix with
independent variable γ
hamy J∞ Hamiltonian matrix as a VARYING matrix with
independent variable γ

Note that the outputs ax, ay, hamx, and hamy correspond to scaled or balanced
data.
The hinfsyn program displays several variables, which can be checked to
ensure that the above conditions are being satisfied. For each γ 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 X∞Y∞, scaled by γ–2,is also displayed. A # sign is
placed to the right of the condition that failed in the printout.

Note When a Hamiltonian has repeated eigenvalues, solving the Riccati

equation via the eigenvalue method (ric_eig) may have problems. This is due
to the MATLAB command eig incorrectly selecting the eigenvectors
associated with the repeated roots.

8-79
hinfsyn

Examples This example is taken from the “HIMAT Robust Performance Design Example”
section in Chapter 7. himat_ic contains the open-loop interconnection
structure. Design an H∞ (sub)optimal controller for the SYSTEM matrix,
himat_ic, with two sensor measurements, two error signals, two actuator
inputs, two disturbances, and eight states. The range of γ is selected to be
between 1.0 and 10.0 with a tolerance, tol, on the relative closeness of the final
γ solution of 0.1. The Schur decompostion method, ric_schr, is used for
solution of the Riccati equations. The program outputs at each iteration the
current γ value being tested, and eigenvalue information about the H and J
Hamiltonian matrices and X∞ and Y∞ Riccati solutions. At the end of each
iteration a (p) denoting the tested γ value passed or an (f) denoting a failure is
displayed. Upon finishing, hinfsyn prints out the γ value achieved.
nmeas = 2;% number of sensor measurements
ncont = 2;% number of control inputs
gmin = 1;% minimum gamma value to be tested
gmax = 10;% maximum gamma value to be tested
tol = .1;% tolerance on the gamma stopping value
ric = 2;% Riccati equation solved via the Schur method
minfo(himat_ic)% SYSTEM interconnection structure
system: 8 states6 outputs6 inputs

[k,g] = hinfsyn(himat_ic,nmeas,ncon,gmin,gmax,tol,ric);
Test bounds: 1.0000 < gamma <=10.0000

gamma hamx_eig xinf_eig hamy_eig yinf_eig nrho_xy p/f

10.000 2.3e-02 2.1e-10 2.3e-02 -3.7e-11 0.022 p
5.500 2.3e-02 2.1e-10 2.3e-02 -0.0e+00 0.075 p
3.250 2.3e-02 2.2e-10 2.3e-02 -0.0e+00 0.222 p
2.125 2.3e-02 2.2e-10 2.3e-02 -0.0e+00 0.564 p
1.562 2.3e-02 2.4e-10 2.3e-02 -0.0e+00 1.198# f
1.844 2.3e-02 2.3e-10 2.3e-02 -0.0e+00 0.789 p
1.703 2.3e-02 2.3e-10 2.3e-02 -2.1e-11 0.959 p
1.633 2.3e-02 2.3e-10 2.3e-02 -0.0e+00 1.068# f

Gamma value achieved:1.7031

8-80
 hinfsyn

Algorithm hinfsyn uses the formulae described in the Glover and Doyle, 1988, paper for
solution to the optimal H∞ control design problem. There are a number of
research issues that need to be addressed for the “best” solution of the Riccati
equations but only two of the standard methods are included.

Subroutines called. hinf_st, hinf_sp, hinf_c, and hinf_gam

hinf_gam calls: ric_eig, ric_schr, csord, cgivens

Reference 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. 831–847, August 1989.
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, vol. 11, pp. 167–172, 1988.

See Also hinffi, hinfnorm, hinfsyne, h2syn, h2norm, ric_eig, ric_schr

8-81
hinfsyne

Purpose Compute an H∞ controller for a SYSTEM interconnection matrix that

8hinfsyne

minimizes the entropy integral at a specific frequency

Syntax [k,g,gfin,ax,ay,hamx,hamy] = hinfsyne(p,nmeas,ncon,...

gmin,gmax,tol,s0,quiet,ricmethd,epr,epp)

Description hinfsyne is a variation of hinfsyn and calculates an H∞ controller that

achieves the infinity norm gfin for the interconnection structure p. The
controller, k, stablizes the SYSTEM matrix p and has the same number of
states as p. Of the controllers achieving this norm bound, k is chosen to
minimize an entropy integral relating to the point s0; i.e.,

2 2
∞
∫
gfin –2 s0
I = – --------------- ln det ( I – γ g ( jω )′ g ( jω ) ) ----------------------- dω
2π – ∞ s0 + ω
2 2

where g is the closed-loop transfer function. In addition, the amount of printing

on the screen can be controlled.
Input arguments:
p SYSTEM interconnection structure matrix
nmeas number of measurements output to controller
ncon number of control inputs
gmin lower bound on γ
gmax upper bound on γ
tol relative difference between final γ values
s0 point at which entropy is evaluated (default ∞)
quiet controls printing on the screen
1 no printing
1 header not printed
–1 full printing (default)
ricmethod 1 Eigenvalue decomposition with balancing
–1 Eigenvalue decomposition with no balancing
2 Schur decomposition with balancing (default)
–2 Schur decomposition with no balancing

8-82
 hinfsyne

epr measure of when a real part of an eigenvalue of the

Hamiltonian matrix is zero (default epr = 1e–10)
epp positive definite determination of the X∞ and Y∞ solution
(default epp = 1e–6)

Output arguments:
k H∞ (sub) optimal controller
g closed-loop system with H∞ controller
gfin final γ value achieved
ax X∞ Riccati solution as a VARYING matrix with independent
variable γ
ay Y∞ Riccati solution as a VARYING matrix with independent
variable γ
hamx H∞ Hamiltonian matrix as a VARYING matrix with
independent variable γ
hamy J∞ Hamiltonian matrix as a VARYING matrix with
independent variable γ

Note that the outputs ax, ay, hamx, and hamy correspond to scaled or balanced
data.
The hinfsyne program outputs several variables, which can be checked to
ensure that the above conditions are being met. For each γ value the minimum
magnitude, real part of the eigenvalues of the X Hamiltonian matrices is
displayed along with the minimum eigenvalue of X∞, which is the solution to
the X Riccati equation. A # sign is placed to the right of the condition that failed
in the printout. This additional information can aid you in the control design
process.

8-83
hinfsyne

Algorithm hinfsyne uses the formulas similar to the ones described in the Glover and
Doyle paper for solution to the H∞ control design problem. See the hinfsyn
command for more information.

Subroutines called. hinf_st, hinf_gam, hinfe_c, hinf_gam calls: ric_eig,

ric_schr, csord, and cgivens

Reference 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. 831–847, August 1989.
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, vol. 11, pp. 167–172, 1988.

See Also dhfsyn, hinfsyn, hinffi, hinfnorm, h2syn, h2norm, ric_eig, ric_schr,
sdhfsyn

8-84
 indvcmp

Purpose 8indvcmp
Compare the independent variable data of two VARYING matrices

Syntax code = indvcmp(mat1,mat2,errcrit)

indvcmp(mat1,mat2,errcrit)

Description indvcmp compares the data for two VARYING matrices. If the two sets of
independent variables are within a specified tolerance of one another, then the
VARYING matrices are assumed to have identical independent variables, and
the VARYING matrices can be combined (i.e., added, subtracted, multiplied,
etc.). The results are displayed if an output argument is not provided.
Input arguments:
mat1, mat2= matrices to be compared
errcrit= 1 x 2 optional matrix containing the relative error and
absolute error bounds. The relative error is used to test the
error in independent variables whose magnitude is greater
than 1e-9, while the absolute error bound used for smaller
independent variable values. Default values are 1e-6, and
1e-13, respectively.

Output arguments:
code=0 independent variable data is different
code=1 independent variable data is identical
code=2 different number of points
code=3 at least one matrix isn’t a VARYING matrix

Examples Compare the two frequency response matrices, mat has its independent
variable at 0.01 and 0.1 and mat2 has its independent variable at 0.011 and
0.1. Given the default comparison criteria, the independent variable is
different. Changing the tolerance leads to the command checking different
indvcmp variations in the independent variable.

8-85
indvcmp

see(mat)
2 rows 3 columns
indep variable 0.01
1 2 3
4 5 6
indep variable 0.1
7 8 9
10 11 12

see(mat2)

2 rows 3 columns
indep variable 0.011
10 20 30
40 50 60
indep variable 0.1
70 80 90
100 110 120

indvcmp(mat,mat2)

varying data is DIFFERENT

code = indvcmp(mat,mat2)

code =
0

Changing the relative and absolute error bounds in indvcmp leads these two
independent variables to be deemed the same.
indvcmp(mat,mat2,[1 1])
varying data is the same

Algorithm indvcmp uses standard MATLAB commands.

See Also getiv, sortiv, vunpck, xtract, xtracti

8-86
 madd, msub

Purpose 8madd, msub

Add and subtract CONSTANT, SYSTEM, and VARYING matrices

Syntax out = madd(mat1,mat2,,matN)

out = msub(mat1,mat2,,matN)

Description madd (msub) allows the addition or subtraction of matrices, regardless of their
type, as long as their dimensions are compatible. CONSTANT, SYSTEM, and
VARYING matrices can be added to or subtracted from one another based on
the following table.

mat2 CONSTANT SYSTEM VARYING

mat1

CONSTANT yes yes yes

SYSTEM yes yes no

VARYING yes no yes

For compatibility, the number of rows and columns of mat1 must equal the
number of rows and columns of mat2. In the case of SYSTEM matrices, the
number of inputs and outputs of mat1 must equal the number of inputs and
outputs of mat2. The same is true for VARYING matrices and in addition, the
independent variables of the VARYING matrices must be identical. Up to nine
matrices of compatible dimension can be added or subtracted by including
them as input arguments.

8-87
madd, msub

Pictorial Representationof Functions

Examples Create two SYSTEM matrices p and p1 and a CONSTANT matrix.

a = -10;b = 3;c = 10;d = 0;
p = pck(a,b,c,d);
minfo(p)
system: 1 states1 outputs1 inputs
a1 = -2; b1 = 3; c1 = 1; d1 = .1;
p1 = pck(a1,b1,c1,d1);
minfo(p1)
seesys(p,'3.2g')

-10 | 3
-----|----
10 | 0

seesys(p1,'3.2g')

-2 | 3
----|----
1 | .1

8-88
 madd, msub

Adding two SYSTEM matrices returns a SYSTEM matrix with the same
number of inputs and outputs as p and p1.
out = madd(p,p1);
seesys(out,'3.2g')

-10 0 | 3
0 -2 | 3
---------|------
10 1 | .1

minfo(out)
system: 2 states 1 outputs 1 inputs

Adding a SYSTEM matrix and a CONSTANT matrix returns a SYSTEM

matrix with the CONSTANT term added to the D-term of the state-space
system.
out = madd(p,a);
seesys(out,'%3.2g')

-10 | 3
------|-----
10 | -10

Subtracting a SYSTEM matrix and a CONSTANT returns a SYSTEM matrix

with the CONSTANT term added to the D-term of the state-space system.
out = msub(p,a);
seesys(out,'%3.2g')

-10 | 3
------|-----
10 | -10

Algorithm madd and msub call the MATLAB + and – commands consistent with the type of
matrices.

See Also +, -, mmult, mscl

8-89
massign

Purpose 8massign
Matrix assignment for VARYING and SYSTEM matrices

Syntax out = massign(matin,rowindex,colindex,data)

Description Performs a matrix assignment like operation on VARYING and SYSTEM

matrices. It is functionally equivalent to
matin(rowindex,colindex) = data
where rowindex and colindex are vectors specifying the rows and columns (or
outputs and inputs if matin is a SYSTEM) to be changed.
data must either be a constant or of the same type as matin. The dimensions of
data must be consistent with the lengths of rowindex and colindex.

Note When applied to a SYSTEM, the result will almost always be

nonminimal.

Examples In the first example a VARYING matrix with two independent variables is
formed with identical (and obvious) data for each matrix.
tl = [11,12,13,14;
21,22,23,24;
31,32,33,34;
41,42,43,44];

vmat = vpck([tl;tl],[0.1,0.2]);

Now make a 2 × 2 data matrix and insert it into the VARYING matrix.
Changing the order of the row and column indices has the effect of permuting
the result. This is identical to the constant matrix case.
ri = [1,3];
ci = [4,2];
data = [0.001, 0.002; 0.003, 0.004];
vmatl = massign(vmat,ri,ci,data);
see(vmatl)
4 rows 4 columns

8-90
 massign

iv = 0.1

11.0000 0.0020 13.0000 0.0010

21.0000 22.0000 23.0000 24.0000
31.0000 0.0040 33.0000 0.0030
41.0000 42.0000 43.0000 44.0000

iv = 0.2

11.0000 0.0020 13.0000 0.0010

21.0000 22.0000 23.0000 24.0000
31.0000 0.0040 33.0000 0.0030
41.0000 42.0000 43.0000 44.0000

In the following example part of a system is replaced by the massign function.

The initial system is a diagonal system of first order lags. The 2,2 element is
replaced by a lightly damped system and the result plotted. Note that the
modified system is no longer minimal.
a = diag([1,2,-3,-4]);
b = eye(4);
c = diag([1,2,3,4]);
sys = pck(a,b,c);
subsys = nd2sys(1,[1,0.1,1]);
sys = massign(sys,2,2,subsys);
minfo(sys)

system: 10 states 4 outputs 4 inputs

omega = logspace(-1,2,100);
sys_g = frsp(sys,omega);
vplot('liv,lm',sys_g,'-')

8-91
massign

10 1

10 0

10 -1

10 -2

10 -3

10 -4
10 -1 10 0 10 1 10 2

See Also sel

8-92
 mfilter

Purpose 8mfilter
Generate SYSTEM representations of a Bessel, Butterworth, Chebyshev or RC
filter

Syntax sys = mfilter(fc,ord,type,psbndr)

Description Calculates, as a SYSTEM, a single-input, single-output analog low-pass filter.

These filters are often encountered in experimental arrangements and must be
accounted for in experimental data processing and control design. For more
sophisticated filters see the Signal Processing Toolbox functions.
The cutoff frequency (Hertz) is fc and the filter order is ord. The string
variable, type, specifies the type of filter and can be one of the following.
butterw Butterworth
cheby Chebyshev
bessel Bessel
rc series of resistor/capacitor filters

The dc gain of each filter (except even order Chebyshev) is set to unity. The
argument psbndr 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.
The Bessel filters are calculated using the recursive polynomial formula. This
is poorly conditioned for high order filters (order > 8).

Examples butw = mfilter(2,4,'butterw');

cheb = mfilter(4,4,'cheby',0.5);
rc = mfilter(1,4,'rc');
omega = logspace(-1,2,100);
butw_g = frsp(butw,omega);
cheb_g = frsp(cheb,omega);
rc_g = frsp(rc,omega);
vplot('bode_gl',[.1 100 .001 10],[.1 100 -180 180],...
butw_g,'-',cheb_g,'-',rc_g,'-.')
subplot(211),title('Butterworth, RC Chebyshev Filters')

8-93
mfilter

Butterworth, RC & Chebyshev Filters

0
10
Log Magnitude

2
10

1 0 1 2
10 10 10 10
Frequency (radians/sec)

0
Phase (degrees)

100

200

300

1 0 1 2
10 10 10 10
Frequency (radians/sec)

8-94
 minfo

Purpose 8minfo
Provide matrix information

Syntax minfo(matin)
[systype,rowdata,coldata,pointdata] = minfo(matin)

Description minfo returns information about the data type, and size of the matin. With no
output assignment, minfo returns text output to the screen. The information is
determined from the data structure as defined in the “The Data Structures”
section in Chapter 2.
With output arguments, minfo returns four arguments. The first argument,
systype, is a string variable that can take one of four values. The
interpretation of the three additional output arguments is based on the
variable systype.
systype == 'vary' matin is a VARYING matrix. pointdata tells how
many independent variable values there are,
rowdata is the row dimension of a matrix, and
coldata is the column dimension
systype == 'syst' matin is a SYSTEM matrix. pointdata is the
number of states, rowdata is the number of outputs,
and coldata is the number of inputs.
systype == 'cons' matin is a regular MATLAB matrix, pointdata is set
to NaN, rowdata is the number of rows, and coldata
is the number of columns.
systype == 'empt' matin is an empty MATLAB matrix; also pointdata,
rowdata, and coldata are set to empty.

8-95
minfo

Examples minfo identifies the type of matrix being manipulated. Compare the displays
for a CONSTANT, SYSTEM, and VARYING matrix.
a = rand(2,2);b = rand(2,3);c = rand(1,2);
minfo(a)
constant: 2 rows 2 cols
sys = pck(a,b,c);
minfo(sys)
system: 2 states1 outputs3 inputs
sys_g = frsp(sys,[.1 .5 .9 1.4]);
minfo(sysg)
varying: 4 pts 1 rows 3 cols
4 pts between 0.1 and 0.4
sys = sysrand(2,3,1);
[mtype,mrows,mcols,mnum] = minfo(sys);
mtype
mtype =
syst
[mrows,mcols,mnum]
ans =
1 3 2

See Also pck, pss2sys, sys2pss, unpck, vpck, vunpck

8-96
 minv, vinv

Purpose 8minv, vinv

Calculate the inverse of CONSTANT, SYSTEM, and VARYING matrices. (The
matrices must be square.)

Syntax out = minv(mat)

out = vinv(mat)

Description minv calculates the inverse of the input matrix. For VARYING matrices, minv
returns a VARYING matrix with the inverse of each independent variable
matrix. For SYSTEM matrices, the inverse is defined as

–1 –1
mat =
A B, out = ( mat )
–1
= A – BD C – BD C
CD –1 –1
D C D

vinv is the same command as minv, but works only on CONSTANT and
VARYING matrices.

Examples Determine the inverse of a SYSTEM matrix and a VARYING matrix.

sys = pck(-2,3,1,.1);
omega = [1 10];
sysg = frsp(sys,omega);
sysi = minv(sys);
seesys(sysi)

-3.2e+01 | -3.0e+01
----------|-----------
1.0e+01 | 1.0e+01

sysig = frsp(sysi,omega);
see(sbs(sysg,sysig,minv(sysig)))
1 row 3 columns
iv =
1.3000 - 0.6000i 0.6341 + 0.2927i 1.3000 - 0.6000i
iv = 10
0.1577 - 0.2885i 1.4591 + 2.6690i 0.1577 - 0.2885i

8-97
minv, vinv

Algorithm minv and vinv call the MATLAB inv command.

See Also inv, vdet

8-98
 mmult

Purpose 8mmult
Multiply CONSTANT, SYSTEM, and VARYING matrices

Syntax out = mmult(mat1,mat2,,matN)

Description mmult allows the multiplication of matrices, mat1 and mat2 regardless of their
type, provided their dimensions are compatible. CONSTANT, SYSTEM and
VARYING matrices can be multiplied by one another based on the following
table.

mat2 CONSTANT SYSTEM VARYING

mat1

CONSTANT yes yes yes

SYSTEM yes yes no

VARYING yes no yes

For compatibility, the number of columns of mat1 must equal the number of
rows of mat2. In the case of SYSTEM matrices, the number of inputs of mat1
must equal the number of outputs of mat2. (An alternative term for the
multiplication of two SYSTEM matrices is cascade.) Similarly restrictions
apply for VARYING matrices. Up to nine matrices of compatible dimension can
be multiplied via the same command by including them as input arguments.

Pictorial
Representation
of Function

8-99
mmult

Examples The multiplication (cascade) of two SYSTEM matrices is shown below.

seesys(p1)

-1.0e+01 | 1.0e+00
----------|----------
1.0e+01 | 1.0e+00

minfo(p1)
system: 1 states 1 outputs 1 inputs
seesys(p2)

-3.0e+00 | 2.0e+00
----------|----------
4.0e+00 | 1.0e-01

minfo(p2)
system: 1 states 1 outputs 1 inputs
out = mmult(p1,p2)
seesys(out,%5.2g')

-10 4 | 0.1
0 -3 | 2
---------|------
10 0 | 0

8-100
 mmult

Algorithm mmult uses the MATLAB “ * ” command when the multiplication does not
involve two SYSTEM matrices. The equation for the multiplication of two
subsystems is given by

A1 B1 A2 B2
sys1 = , sys2 = ,
C1 D1 C2 D2

A1 B1 C2 B1 D2
→ mmult(sys1,sys2) = 0 A2 B2
C1 D1 C2 D1 D2

See Also *, abv, madd, mscl, msub

8-101
mprintf

Purpose 8mprintf
Format output of matrix to screen

Syntax mprintf(matin,'format','end of line characters')

Description mprintf displays a matrix in formatted form. The optional 'format' specifies
the format exactly as in the MATLAB function sprintf. If no 'format' is
specified the default is '%.1e'. This routine is primarily for use in seesys and
does not work well for f format when the minimum field width is too small.
There is no input checking, so you can wreak havoc if you use mprintf
incorrectly. See sprintf for more details. The optional 'end of line
characters' is exactly what it says. The default is the newline C escape
sequence (\n). To get no newline at the end of each line use
mprintf(matin,'format',[]).

Examples The mprintf command displays any type of matrix. An example of its use for
SYSTEM and VARYING matrices follows.
mprintf(m)

1.7e+01 8.7e+00 -6.4e+00 -1.3e+01

5.9e-01 -1.4e+01 5.8e+00 -1.3e+01
1.8e+01 -7.0e+00 -3.6e+00 9.8e+00
2.6e+00 1.2e+01 -1.4e+00 -4.5e-01

mprintf(m,'%6.2f ')

16.96 8.72 -6.39 -13.49

0.59 -14.46 5.77 -12.70
17.97 -7.01 -3.60 9.85
2.64 12.46 -1.36 -0.45

mprintf (m, '%3.f ')

17 9 -6 -13
1 -14 6 -13
18 -7 -4 10
3 12 -1 -O

See Also fprintf, rifd, see, seesys, sprintf

8-102
 msf, msfbatch

Purpose Interactive D-scaling rational fit routine used in µ-synthesis

8msf, msfbatch

Syntax [dsysL,dsysR] = msf(Mg,bnds,dvec,sens,blk)

[dsysL,dsysR] = msfbatch(Mg,bnds,dvec,sens,blk,maxord)

Description msf fits the block diagonal, frequency-dependent matrices DL(ω) and DR(ω)
(contained in the VARYING matrix dvec, with block structure implied by the
entries of blk) with rational, stable, minimum-phase D̂ L ( s ) and D̂ R ( s ) such
that

–1 –1
maxσ [ D L ( ω )M ( jω )D R ( ω ) ] ≈ D̂ L MD̂ R ∞

msf returns the stable, minimum phase system matrices dsysL and dsysR.

Note Typically, there is no need to call msf directly. The standard use of msf
is a subroutine within µ-synthesis. The programs dkit and/or dkitgui are
fully functional µ-synthesis routines.

Input arguments:
Mg is the frequency response upon which the µ calculation was
performed.
bnds is the upper bound from the µ calculation.
dvec is a frequency varying vector containing the Ds (obtained from
mu).
sens is the sensitivity of the upper bound in the µ calculation on the
Ds. The sensitivity sens is a frequency domain weight calculated
by mu.
blk is the uncertainty block structure. This should correspond with
the block strcucture used in the µ calculation (and produced bnds,
dvec and sens).

8-103
msf, msfbatch

Output arguments:
dsysL is the left (i.e., output) block diagonal scaling matrix. It is a
SYSTEM matrix (it may be CONSTANT)
dsysR is the right (i.e., input) block diagonal scaling matrix. It is the
same type as dsysL

msftbatch is a batch version of msf with no user interaction.

See Also dkit, dkitgui, fitmag, fitmaglp, fitsys, genphase, magfit

8-104
 mscl, sclin, sclout

Purpose 8mscl, sclin, sclout

Scale CONSTANT, SYSTEM and VARYING matrices

Syntax matout = mscl(matin,fac); matout = sclin(matin,inputs,fac); matout

= sclout(matin,outputs,fac);

Description The mscl command scales the individual elements of a CONSTANT or

VARYING matrix by the scalar fac. mscl produces a CONSTANT (VARYING)
matrix, matout, given matin is a CONSTANT (VARYING) matrix, and fac is a
scalar. When the output matrix, matout, is VARYING, it has the same
independent variable values as matin, and is equal to matin multiplied by the
scalar value fac. If matin is a SYSTEM matrix, then mscl produces a SYSTEM
matrix, matout, in which the A and C matrices of matin are unchanged, and the
B and D matrices have been scaled by fac. fac must be a scalar.
The command sclin scales the input channels to the matrix matin (which can
be a CONSTANT, SYSTEM or VARYING matrix) defined by the variable
inputs. inputs is a vector of integers, which selects the inputs channels
(columns) to be scaled by fac. The command sclout performs the similar
scaling on the outputs (or rows). fac can be a scalar, single-input/single-output
(SISO) SYSTEM or a VARYING, 1 × 1, matrix. If matin and fac are both
VARYING matrices, they must have the same independent variables to be
used with sclin and sclout. The following options can be performed using
sclin and sclout:

• Scale select inputs and outputs of a SYSTEM or CONSTANT (matin) matrix

by a scalar or a SISO SYSTEM (fac).
• Scale select inputs and outputs of a VARYING matrix by a scalar.
• Scale select inputs and outputs of a VARYING matrix by a VARYING
(matin), 1 × 1, matrix (fac).

8-105
mscl, sclin, sclout

Examples mscl scales the three input, two output VARYING matrix, matin, by –2.5.

minfo(matin)
varying- 2 pts2 rows3 cols
see(matin)
2 rows 3 columns
indep variable0.2
3 13 23
4 14 24
indep variable0.3
4 14 24
5 15 25
matout = mscl(matin,-2.5);
see(matout)
2 rows 3 columns
indep variable0.2
-7.5000 -32.5000-57.5000
-10.0000-35.0000-60.0000
indep variable 0.3
-10.0000-35.0000-60.0000
-12.5000-37.5000-62.5000
Use the sclin command to scale the first input of a SYSTEM matrix by –3.
sys = pck(ones(3,3),2*ones(3,2),3*ones(1,3),4*ones(1,2));
seesys(sys)

1.0e+00 1.0e+00 1.0e+00 | 2.0e+00 2.0e+00

1.0e+00 1.0e+00 1.0e+00 | 2.0e+00 2.0e+00
1.0e+00 1.0e+00 1.0e+00 | 2.0e+00 2.0e+00
------------------------------------------------------
3.0e+00 3.0e+00 3.0e+00 | 4.0e+00 4.0e+00

sysout = sclin(sys,1,-3);

8-106
 mscl, sclin, sclout

seesys(sysout)

1.0e+00 1.0e+00 1.0e+00 | -6.0e+00 2.0e+00

1.0e+00 1.0e+00 1.0e+00 | -6.0e+00 2.0e+00
1.0e+00 1.0e+00 1.0e+00 | -6.0e+00 2.0e+00
-------------------------------------------------------
3.0e+00 3.0e+00 3.0e+00 | -1.2e+01 4.0e+00

The sclout command can be used to scale the first and third outputs
of a SYSTEM matrix by the first order transfer function 10/(s +
10).

sys = sysrand(2,3,1);
seesys(sys,'%11.2e)

2.19e-01 6.79e-01 | 9.35e-01

4.70e-02 6.79e-01 | 3.84e-01
----------------------------------
5.19e-01 5.35e-02 | 7.70e-03
8.31e-01 5.30e-01 | 3.83e-01
3.46e-02 6.71e-01 | 6.68e-02

sysout=sclout(sys,[1 3],nd2sys(10,[1 10]));

seesys(sysout,'%11.2e')

-1.00e+01 0.00e+00 -1.64e+00 -1.69e-01 | -2.43e-02

0.00e+00 -1.00e+01 -1.09e-01 -2.12e+00 | -2.11e-01
0.00e+00 0.00e+00 2.19e-01 6.79e-01 | 9.35e-01
0.00e+00 0.00e+00 4.70e-0 6.79e-01 | 3.84e-01
---------------------------------------------|------------
-3.16e+00 0.00e+00 0.00e+00 0.00e+00 | 0.00e+00
0.00e+00 0.00e+00 8.31e-01 5.30e-01 | 3.83e-01
0.00e+00 -3.16e+00 0.00e+00 0.00e+00 | 0.00e+00

8-107
mscl, sclin, sclout

matin = vpck([ones(3,2);2*ones(3,2);3*ones(3,2)],[1;2;3]);
seesys(matin)
3 rows 2 columns

iv = 1
l.0e+00 l.0e+00
l.0e+00 1.0e+00
l.0e+00 l.0e+00

iv = 2
2.0e+00 2.0e+00
2.0e+00 2.0e+00
2.0e+00 2.0e+00

iv = 3
3.0e+00 3.0e+00
3.0e+00 3.0e+00
3.0e+00 3.0e+00

fac = vpck([1;2;3],[1;2;3]);
sysout = sclout(matin,l,fac)i
seesys(matin)
3 rows2 columns

iv = 1
l.0e+00 l.0e+00
l.0e+00 l.0e+00
l.0e+00 l.0e+00

iv = 2
4.0e+00 4.0e+00
2.0e+00 2.0e+00
2.0e+00 2.0e+00

iv = 3
9.0e+00 9.0e+00
3.0e+00 3.0e+00
3.0e+00 3.0e+00

See Also *, mmult, scliv, and sel

8-108
 mu, muunwrap, randel, unwrapd, unwrapp

Purpose 8mu, muunwrap, randel, unwrapd, unwrapp

Compute upper and lower bounds for the complex and mixed (real and
complex) structured singular value (referred to as mixed µ) of a VARYING/
CONSTANT matrix

Syntax [bnds,dvec,sens,pvec,gvec] = mu(matin,blk,options);

[dl,dr,gl,gm,gr,pert] = muunwrap(dvec,gvec,pvec,blk);
[dl,dr,gl,gm,gr] = muunwrap(dvec,gvec,blk);
[dl,dr] = muunwrap(dvec,blk);
pert = randel(blk,nrm,opt);
[dl,dr] = unwrapd(dvec,blk);
pert = unwrapp(pvec,blk);
The functions associated with mixed µ are
mu general complex and mixed computation
muunwrap extract block-diagonal D- and G- scalings from row vectors
containing the scalings
randel create random block structured matrix
unwrapd extract block-diagonal D-scalings from row vector
containing the scalings
unwrapp extract block-diagonal perturbation from row vector
containing the perturbation

8-109
mu, muunwrap, randel, unwrapd, unwrapp

Description Input arguments:

matin A CONSTANT or VARYING matrix
blk An array that describes the perturbation block structure. Its
size is nblk × 2, where nblk is the total number of blocks in
the perturbation structure. The ith row of blk defines the
dimensions of the ith perturbation block. If blk(i,:) = [-r
0], then the ith block is an r × r repeated, diagonal real
scalar perturbation, while if blk(i,:) = [r 0], then the ith
block is an r × r repeated, diagonal complex scalar
perturbation, and if is blk(i,:) = [r c], then the ith block
is an r × c complex full-block perturbation. If blk is omitted,
the default is all 1 × 1 complex blocks, and results in an
error if matin is not square.
options An optional character string describing the desired
computations. It can consist of the following characters:
'l' - compute lower bound using a power iteration
't' - use more iterations in the lower bound
'R' - start power iteration with RANDOM vectors
'Rj' - restart lower bound j times with RANDOM vectors
where j is an integer between 1 and 9
'u' - compute upper bound using a balanced/LMI technique
'c' - compute upper bound to greater accuracy'C'
'C' - compute tighest upper bound (may be slow)
'f' - compute a fast but crude upper bound
'r' - restart computation at EACH independent variable
's' - suppress progress information
'w' - suppress warnings
'L' - compute only the lower bound
'U' - compute only the upper bound

8-110
 mu, muunwrap, randel, unwrapd, unwrapp

The default value of options is 'lu', meaning that a lower bound will be
computed using the power method, Young and Doyle 1990 and Packard et al.
1988, and an upper bound will be computed, using the balanced/AMI
technique, Young et al., 1992, for computing the upper bound from Fan et al.,
1991.
Output arguments:
bnds A 1 × 2 vector. If matin is VARYING, so is bnds, whereas if
matin is a CONSTANT matrix, then bnds is CONSTANT.
The first column of bnds contains an upper bound to mixed
µ of matin, and the second column contains a lower bound
to mixed µ.
dvec and gvec Row vectors which contain the D and G scaling matrices
that have produced the upper bound in bnds. dvec and gvec
are the same data type as bnds and are stored as vectors to
save memory. They can be unwrapped into the appropriate
D and G matrices by using the command, muunwrap:

[dl,dr,gl,gm,gr] = muunwrap(dvec,gvec,blk);
The upper bound in bnds for a matrix M is a number β > 0 such that there are
scaling matrices Dl, Dr, Gl, Gm, Gr (see Young et al., 1992, for details) satisfying

–1
 2 – --1-  D l MD r  2 – --1-
σ  ( I + G l ) 4  ----------------------- – jG m ( I + G r ) 4 ≤ 1
  β  

sens is a row vector which contains the sensitivity of

–1
σ ( D l MD r )

with respect to the values in Dl (and Dr). It is calculated in an ad-hoc manner,

and is mainly used when fitting frequency varying Ds with rational functions
via the routines in dkit, dkitgui, and autodkit.
pvec is a row vector containing a perturbation matrix that has the structure
defined by blk. As with dvec and gvec, pvec is the same data type as matin and
is stored as a vector. It can be unwrapped into the actual perturbation with the
command, unwrapp.

8-111
mu, muunwrap, randel, unwrapd, unwrapp

pert = unwrapp(pvec,blk);

After being unwrapped, the perturbation matrix pert satisfies three

conditions:
• It has the block structure defined by blk;
• The maximum singular value of pert is equal to the reciprocal of the lower
bound in bnds (when the lower bound is not zero);
• The matrix mmult(matin,pert) has an eigenvalue equal to 1 at each
independent variable.

Examples Suppose sys is a system matrix with four inputs and four outputs, and that it
is stable. sys_g is a frequency response of sys.
% ∆ is 4 1 × 1 perturbation blocks
blk = ones(4,2);

% Calculate µ on frequency response

[bnds,dvec,sens,pvec] = mu(sys_g,blk);

% Unwrap the D scaling matrices

[dl,dr] = unwrapd(dvec,blk);

% Generate scaled matrix

dmdi = mmult(dl,sys_g,minv(dr));

% Verify the upper bound

vplot('liv,m',vnorm(dmdi),sel(bnds,1,1))

% Unwrap the perturbation

actpert = unwrapp(pvec,blk);

% Check that perturbation is correct structure

see(xtracti(actpert,1))

% Check lower bound and perturbation

vplot('liv,lm',sel(bnds,1,2),vnorm(actpert))

8-112
 mu, muunwrap, randel, unwrapd, unwrapp

% Form M∆
mdel = mmult(sys_g,actpert);

% Verify M∆ has an eigenvalue at 1

veig(mdel)

Looking at the same frequency response sys_g with a mixed real/complex block
structure.
% ∆ is 4 1 × 1 perturbation blocks
% with the first 2 real, and the last 2 complex
blk = [-1 0; -1 0; 1 0; 1 0];

% Calculate mixed µ on frequency response

[bnds,dvec,sens,pvec,gvec] = mu(sys_g,blk);

% Unwrap the D and G scaling matrices

[dl,dr,gl,gm,gr] = muunwrap(dvec,gvec,blk);

% Generate scaled matrix

dmd = mmult(dl,sys_g,minv(dr));
oobdmd = veval('*',dmd,minv(sel(bnds,1,1)));
oobdmdjg = msub(oobdmd,mscl(gm,j));
scall = madd(eye(4),mmult(gl,gl));
scalr = madd(eye(4),mmult(gr,gr));
scall = veval(,scall,-0.25);
scalr = veval(”,scalr,-0.25);
scaledmat = mmult(scall,oobdmdjg,scalr);

% Verify the upper bound (scaledmat should have norm ≤1)

vplot('liv,m',vnorm(scaledmat))

% Unwrap the perturbation

pert = unwrapp(pvec,blk);

% Check that perturbation is correct structure

see(xtracti(pert,1))

% Check lower bound and perturbation

vplot('liv,m',sel(bnds,1,2),minv(vnorm(pert)))

8-113
mu, muunwrap, randel, unwrapd, unwrapp

% Form M∆
mdel = mmult(sys_g,pert);

% Verify M∆ has an eigenvalue at 1

see(veig(mdel))

For more examples of computing bounds for µ, please refer to the

“Computational Exercise with the mu Command” and “Computational
Exercise with the mu Command — Mixed Perturbations” sections in Chapter
4 as well as the robust multivariable control design examples in Chapters 6
and 7.

Algorithm Peter Young and Matt Newlin helped write the mu program and supporting
routines.
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 Osborne’s 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
µ problem. The Perron eigenvector method is based on an idea of Safonov,
(Safonov, 1982). It gives the exact computation of µ 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
which exploit the structure of the problem, concluding with general purpose
AMI optimization (Boyd et al.), 1993, to obtain the final answer.

8-114
 mu, muunwrap, randel, unwrapd, unwrapp

Reference Boyd, S. and L. El Ghaoui, “Methods of centers for minimizing generalized

eigenvalues,” Linear Algebra and Its Applications, vol. 188–189, pp. 63–111,
1993.
Fan, M. A. Tits, and J. Doyle, “Robustness in the presence of mixed parametric
uncertainty and unmodeled dynamics,” IEEE Transactions on Automatic
Control, vol. AC–36, pp. 25–38, 1991.
Osborne, E. “On preconditioning of matrices,” Journal of Associated Computer
Machines, vol. 7, pp. 338–345, 1960.
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, pp.
2132–2137, December 1988.
Safonov, M. “Stability margins for diagonally perturbed multivariable
feedback systems,” IEEE Proc., vol. 129, Part D, pp. 251–256, 1982.
Young, P. and J. Doyle, “Computation of with real and complex uncertainties,”
Proceedings of the 29th IEEE Conference on Decision and Control, pp. 1230–
1235, 1990.
Young, P., M. Newlin, and J. Doyle, “Practical computation of the mixed
problem,” Proceedings of the American Control Conference, pp. 2190–2194,
June, 1992.

See Also blknorm, dypert, genmu, norm, vnorm, vrho, vsvd

8-115
musynfit, musynflp, muftbtch

Purpose Interactive D-scaling rational fit routines used in old µ-synthesis routines
8musynfit, musynflp, muftbtch

Note These routines are included only for backwards compatibility with
versions 1.0 and 2.0. They will not be supported in future versions. They
should not be used by new users of the toolbox. All new routines should be
based on the new routine, msf, which is described on page ???.

Syntax [dsysL,dsysR] = musynfit(pre_dsysL,dvec,sens,blk,...

nmeas,ncntrl,clpg,upbd,wt)
[dsysL,dsysR] = musynflp(pre_dsysL,dvec,sens,blk,...
nmeas,ncntrl,clpg,upbd,wt)
[dsysL,dsysR] = muftbtch(pre_dsysL,dvec,sens,blk,...
nmeas,ncntrl,dim)

Description musynfit fits the magnitude curve obtained by multiplying the old D frequency
response (from pre_dsysl) with the dvec data. musynfit returns stable,
minimum phase system matrices dsysL and dsysR, which can be absorbed into
the original interconnection structure. Once absorbed, a H∞ design is
performed with hinfsyn completing another D–K iteration of µ-synthesis.
For the first µ-synthesis iteration, set the variable pre_dsysl to the string
first. In subsequent iterations, pre_dsysl should be the previous (left)
rational D-scaling system matrix, dsysL. Essentially, the element-by-element
magnitudes of the matrices
mmult(unwrapd(dvec,blk),frsp(pre_dsysL,getiv(dvec))), and
frsp(dsysL,getiv(dvec)) are equal.

The (optional) variable clpg is the VARYING matrix that produced the dvec,
sens, and upbd data output from µ. The fitting procedure is interactive
(musynfit or musynflp), and fits (in magnitude) these scalings with rational,
stable transfer function matrices, D̂ ( s ) . After fitting the dvec data, plots of

–1
σ ( D f ( jω ) ⋅ clpg ( jω ) ⋅ D f ( jω ) )

and

8-116
 musynfit, musynflp, muftbtch

–1
σ ( D̂ ( jω ) ⋅ clpg ( jω ) ⋅ D̂ ( jω ) )

are shown in the lower graph window for comparison. At this point, you have
the option of refitting the D data. If clpg and upbd are not provided, the default
is to plot the sens variable in the the lower graph.

Note You are strongly discouraged from calling musynfit and musynflp
directly and are encouraged to use dkit or dkitgui to perform µ-synthesis
calculations.

Input arguments:
pre_dsysl is set to the character string 'first' for the first iteration.
As the iteration proceeds, it should be the previous dsysL.
sens is the sensitivity of the upper bound in the µ calculation on
the Ds. The sensitivity sens is a frequency domain weight,
which is obtained from mu.
dvec is a frequency varying row vector containing the Ds (from mu).
blk is the block structure, same block structure used in mu.
nmeas is the number of measurements in control problem.
ncntrl is the number of controls in control problem.
clpg is the frequency response upon which the calculation was
performed (optional).
upbd is the upper bound from the µ calculation (optional).
wt is a weight used to influence the frequency range in which the
data is to be fit more accurately (optional).
dim is the highest order fit to be used (only used in muftbtch).

8-117
musynfit, musynflp, muftbtch

Output arguments:
dsysL is the output (left) block diagonal, SYSTEM scaling.
dsysR is the input (right) block diagonal, SYSTEM scaling (needs to
be inverted before being absorbed into the interconnection
structure).
musynflp has the same inputs and outputs and user interaction as
musynfit but uses a linear programing routine to do the
fitting. muftbtch is a batch version of musynflp that has no
user interaction. The extra argument dim is a required
argument of parameters for the linear program. dim has the
form [hmax htol nmin nmax] where
• hmax is a measure of the allowable error in the fit.
• htol is a measure of the accuracy with which the optimiza-
tion is carried out.
• nmin and nmax are the minimum and maximum orders con-
sidered in the problem

For more detail about the role of hmax and htol, see the reference pages for
fitmaglp and magfit. Reasonable choices are hmax = .26 and htol = .1.

The musynflp and musynfit commands provide the option of fitting the
frequency varying D-scale data by hand using the µ-Tools drawmag command.
You can invoke this option with the string ’drawmag’ in response to the prompt
ENTER ORDER OF CURVE FIT or 'drawmag'

The mouse is used in the plot window to identify the data to be fit with a stable,
minimum-phase system. See the drawmag command for more information.

Examples musynfit is used within a D–K iteration (µ-synthesis) to fit the D-scales, which
are output from the mu command. The first step in the D–Kiteration is to design
an H∞ control law. The closed-loop system is analyzed with mu based on the
block structure blk defined. The optimal D-scaling output from mu, which are
real coeeficients, are fit with real, rational, minimum-phase stable transfer
functions via musynfit. These fitted D-scales are wrapped back around the
orginal interconnection structure P. After absorbing the D-scales, another D–K
iteration is performed, starting with the design of an H∞ control law for the

8-118
 musynfit, musynflp, muftbtch

modified plant. This process usually continues until the value of µ doesn’t
change significantly between control design iterations.
This example is taken from the “HIMAT Robust Performance Design Example”
section in Chapter 7. himat_ic contains the open-loop interconnection
structure. It has one multiplicative input perturbation, which is two by two,
and has two error signals, and two external disturbances. There are two
measurements, and two control inputs to the system. The block structure for
the µ-analysis problem is given by blk=[2 2; 2 2].
First step in a D–K iteration is to design an H∞ controller and analyze the
closed loop system with µ.
mkhic
omega = logspace(0,4,40);
blk = [2 2; 2 2];
[k1,g1,gf1] = hinfsyn(himat_ic,2,2,0.8,6,0.05,2);
g1_g = frsp(g1,omega);
[bnds1,dvec1,sens1,rp1] = mu(g1_g,blk);

The D-scalings output from the µ-analysis problem needs to be fit with real,
rational, stable, minimum-phase transfer functions. This is done with
musynfit. The first D-scale is fit with both a first and third-order transfer
function, with the third order transfer function selected.

8-119
musynfit, musynflp, muftbtch

[dsysL1,dsysR1] = musynfit('first',dvec1,sens1,blk,2,2);
FITTING D SCALING #1 of 1
10 1

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
data and old fit

10 1 wt for fit

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
NOTE APPROXIMATE ORDER NECESSARY FOR FIT.....

ENTER ORDER OF CURVE FIT or 'drawmag' 1

FITTING D SCALING #1 of 1, W/ORDER = 1

10 1

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
1) data 2) newfit 3) oldfit

10 1 weight for fit

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
ENTER NEW ORDER, ’drawmag’, or NEGATIVE NUMBER TO STOP

8-120
 musynfit, musynflp, muftbtch

ENTER ORDER OF CURVE FIT or 'drawmag' 3

FITTING D SCALING #1 of 1, W/ORDER = 3

10 1

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
1) data 2) newfit 3) oldfit

10 1 weight for fit

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
ENTER NEW ORDER, ’drawmag’, or NEGATIVE NUMBER TO STOP

ENTER ORDER OF CURVE FIT or 'drawmag' -1

Now the fitted D-scales are absorbed into the interconnection structure,
himat_ic, to generate himat_ic2.

mu_ic1 = mmult(dsysL1,himat_ic,minv(dsysR1));

The modified interconnection structure, mu_ic1, is used in the second iteration

of µ-synthesis. An H∞ control law is designed for the modified system, and then
analyzed again using mu.
[k2,g2] = hinfsyn(mu_ic1,2,2,.9,1.3,.04,2);
g2_g = frsp(g2,omega);
[bnds2,dvec2,sens2,pvec2] = mu(g2_g,blk);

The new D-scales are fit again using the previous D-scale information.
[dsysL2,dsysR2] = musynfit(dsysL1,dvec2,sens2,blk,2,2);

8-121
musynfit, musynflp, muftbtch

The graphs displayed with musynfit for this iteration are not included. Wrap
the new fitted D-scales around the original plant interconnection structure and
start D–K iteration again.
mu_ic2 = mmult(dsysL2,himat_ic,minv(dsysR2));
[k3,g3] = hinfsyn(mu_ic2,2,2,.9,1.3,.05,2);

musynfit can be called as before with the frequency response of the closed-loop
system analyzed using mu, g1_g, and the mu upper bound, sel(bnds1,1,1)
passed. The first D-scale is fit with both a first- and third-order transfer
function and the first order transfer function selected. As you can see from the
scaled upper bound plots (the lower graph), the first-order fit does a better job
minimizing the scaled upper bound.
[dsysL1,dsysR1] = ...
musynfit('first',dvec1,sens1,blk,2,2,gl_g,sel(bnds1,1,1)

FITTING D SCALING #1 of 1
10 1

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
data and old fit

wt for fit
10 1

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
NOTE APPROXIMATE ORDER NECESSARY FOR FIT.....

8-122
 musynfit, musynflp, muftbtch

ENTER ORDER OF CURVE FIT or 'drawmag' 3

10 1 FITTING D SCALING #1 of 1, W/ORDER = 3

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
1) mag data 2) newfit 3) previous D-K

2.5 Scaled transfer function: optimal and rational upper bound

1.5

0.5
10 0 10 1 10 2 10 3 10 4
1) mu upper bnd 2) upper bnd with rational fit

ENTER ORDER OF CURVE FIT or 'drawmag' 1

10 1 FITTING D SCALING #1 of 1, W/ORDER = 1

10 0

10 -1

10 -2
10 0 10 1 10 2 10 3 10 4
1) mag data 2) newfit 3) previous D-K
Scaled transfer function: optimal and rational upper bound
2

1.5

0.5
10 0 10 1 10 2 10 3 10 4
1) mu upper bnd 2) upper bnd with rational fit
ENTER ORDER OF CURVE FIT or 'drawmag' -1

8-123
musynfit, musynflp, muftbtch

As before you would aborb the fitted D-scales into the interconnection
structure, himat_ic, to generate himat_ic2.

Algorithm A frequency response is done on the previous rational D-scaling matrix. This is
multiplied by the current data in dvec, to produce the frequency varying
scaling that needs to be fit. The fit is only in magnitude, and the freedom in the
phase allows the rational function to be defined as stable, and minimum phase.
musynfit calls fitsys, which calls fitmag, flatten, and genphase. The curve
fitting is done the fitsys command.
musynflp is an alternative program that uses linear programming to do the fit.
musynflp fits the data very well within the frequency response window at the
expense of perhaps large variations outside the data window. This may lead to
problems in D–K iteration. muftbtch is a batch version of musynflp.

Reference Doyle, J.C., K. Lenz, and A.K. Packard, “Design examples using µ-synthesis:
Space shuttle lateral axis FCS during reentry,” NATO ASI Series, Modelling,
Robustness and Sensitivity Reduction in Control Systems, vol. F34 R.F. Curtin,
Editor, Springer-Verlag, Berlin-Heidelberg, 1987.

See Also drawmag, fitmag, fitmaglp, fitsys, flatten, genphase, invfreqs, magfit, msf

8-124
 ncfsyn, cf2sys, emargin

Purpose ncfsyn synthesizes an H∞ controller to robustly stabilize a family of systems

8ncfsyn, cf2sys, emargin

given by a ball of uncertainty in the normalized coprime factors of the system

description
cf2sys calculates a SYSTEM/CONSTANT/VARYING matrix from a coprime
factorization.
emargin calculates the normalized coprime factor/gap metric robust stability
margin b(P,K) as defined in the “Loop Shaping Using H• Synthesis” section in
Chapter 3.

Syntax [sysk,emax,sysobs] = ncfsyn(sysgw,factor,opt)

sysout = cf2sys(sysrcf)
emarg = emargin(sys_g,sysw,tol)

Description A method of designing controllers is to use a combination of loop shaping and

robust stabilization as proposed in McFarlane and Glover. The first step is to
design a pre- and post-compensator W1(s) and W2, so that the gain of
W2(s)P(s)W1(s) 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 design a feedback controller, K∞, so
that

1 I –1 1
----------------------------------------- := ( I – W 2 PW 1, K ∞ ) [ W 2 PW 1, I ] ð ---
b ( W 2 PW 1, K ∞ ) K∞ ε
∞

which will also give robust stability of the perturbed weighted plant

–1 ∆1
( N + ∆1 ) ( M + ∆2 ) for < b ( W 2 PW 1, K ∞ )
∆2
∞

where NM–1 = W2PW1is a normalized coprime factorization satisfying

N(jw)*N(jw) + M(jw)*M(jw) = I. This stability margin is always less than 1 and
gives a good indication of robust stability to a wide class of unstructured
perturbations, with values of ε > 0.2 – 0.3 generally satisfactory.
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

8-125
ncfsyn, cf2sys, emargin

loop shape in frequencies where the gain of W2PW1 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
W1K∞W2.
When the option 'ref' is specified, the controller includes an extra set of
reference inputs as proposed in Vinnicombe, 1993, and should be implemented

as u = sysk * y where y = sysgw * u and r is a reference input. The

r
closed-loop response will then be y = Nr, where N is the numerator of a
normalized right coprime factor of sysgw.
Input arguments
sysgw the weighted system to be controlled
factor =1 implies that an optimal controller is required.
>1 implies that a suboptimal controller is required achieving a
performance FACTOR less than optimal.
opt 'ref' the controller includes an extra set of reference input and
should be implemented as in Figure 3–12 on page 3-? (optional).

Output arguments
sysk H∞ loopshaping controller
emax Stability margin as an indication robustness to unstructured
perturbations. emax is always less than 1 and values of emax
greater than 0.3 generally indicate good robustness margins.
sysobs H∞ loopshaping observer controller. This variable is created only
if factor>1 and opt = 'ref'

cf2sys creates a SYSTEM/CONSTANT/VARYING matrix from a coprime

factorization.
emargin calculates the normalized coprime factor/gap metric robust stability
margin b(P,K) as defined in the “Loop Shaping Using H• Synthesis” section in

8-126
 ncfsyn, cf2sys, emargin

Chapter 3. The default value for the tolerance tol supplied to the hinfnorm
computation is 0.001.

Algorithm See the McFarlane and Glover reference for details.

Reference 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.
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.
Vinnicombe, G., “Measuring Robustness of Feedback Systems,” PhD
dissertation, Department of Engineering, University of Cambridge, 1993.

See Also gap, hinfsyn, hinfnorm, nugap

8-127
nd2sys, zp2sys

Purpose 8nd2sys, zp2sys

Convert single-input/single-output (SISO) transfer functions to SYSTEM
matrices

Syntax sys = nd2sys(num,den,gain)

sys = zp2sy(zeros,poles,gain)

Description nd2sys converts a numerator/denominator (num,den) SISO transfer function

into a SYSTEM matrix, sys. This function uses the MATLAB command tf2ss
for the conversion. zp2sys converts zeros and poles of a SISO transfer function
into a SYSTEM matrix, sys. This function uses the MATLAB command zp2ss
to do the conversion. An optional argument gain can be supplied to scale the
transfer function. The default value of gain is 1. The output SYSTEM matrix,
sys, is balanced with a call to the µ-Tools command sysbal prior to being
return if the real parts of the poles are all less than zero.

Examples Convert the single-input/single-output transfer function sys1 into a SYSTEM

matrix.

2
4s + 5s + 1
sys = -----------------------------------------------------------------
4 3 2
7s + 3s + 6s + 2s + 8

sys = nd2sys([4 5 1],[7 3 6 2 8]);

minfo(sys)
system:4 states1 outputs1 inputs
see(sys)
see (sys)

A matrix

-0.4286 -0.8571 -0.2857 -1.1429

1.0000 0 0 0
0 1.0000 0 0
0 0 1.0000 0

8-128
 nd2sys, zp2sys

B matrix
1
0
0
0

C matrix
0 0.5714 0.7143 0.1429

D matrix
0

Note zp2sys requires the Signal Processing or Control Toolbox.

Algorithm nd2sys and zp2sys realize the transfer functions using the MATLAB
commands tf2ss and zp2ss.

See Also pss2sys, pck, tf2ss

8-129
negangle

Purpose 8negangle
Calculate the angle of elements of a matrix, which is always between [–2π,0]

Syntax y = negangle(x)

Description negangle returns the phase angles, in radians, of a matrix with complex valued
elements. The returned value is always in the range 0 to –2π radians.

Examples a = [1+i,1-i];
see(angle(a))
0.7854-0.7854
see(negangle(a))
-5.4978-0.7854

See Also angle

8-130
 pck, pss2sys, sys2pss, unpck

Purpose 8pck, pss2sys, sys2pss, unpck

Convert to and from CONSTANT and SYSTEM matrices

Syntax sys = pss2sys(mat,n)

sys = pck(A,B,C,D)
sys = pck(A,B,C)
[A,B,C,D] = unpck(sys)
mat = sys2pss(sys)

Description pss2sys translates a regular MATLAB matrix that is in packed form into a
SYSTEM matrix. mat contains [A B; C D] which describes the individual
components of a SYSTEM matrix with n being the number of states (size of A).
sys2pss returns the CONSTANT matrix, mat = [A B; C D], from the input
SYSTEM matrix sys.
pck takes consistent state-space data and forms a SYSTEM matrix with the
data structure defined. Consistent state-space data requires a square A matrix,
a B matrix with the same number of rows as A, a C matrix with the same
number of columns as A, and a D matrix with same number of columns of B and
rows of C. If the fourth input argument is omitted, then the D matrix is assumed
to be identically zero, of appropriate dimensions. unpck is the inverse operation
of pck, taking a SYSTEM matrix sys and converting it to A, B, C and D
CONSTANT matrices.
Note that based on the data structure definition, a -Inf in the bottom right
corner of a matrix denotes a SYSTEM matrix, with the top, right corner
element of the matrix containing the number of states.

8-131
pck, pss2sys, sys2pss, unpck

Examples Create a SYSTEM matrix from MATLAB CONSTANT matrices via pss2sys.
Define matrices A, B, C, and D as follows.
A = [-1 1; -1 -3]; B =[2 2; 2 2]i C = [3 3]; D = [4 4]
mat = [A B; C D];
sys = pss2sys(mat,2);
minfo(mat)
3 rows4 cols: regular MATLAB matrix
mat =
-1 1 22
-1 -3 22
3 3 44
minfo(sys)
system: 2 states1 outputs2 inputs
seesys(sys)

-1.0e+00 1.0e+00 | 2.0e+00 2.0e+00

-1.0e+00 -3.0e+00 | 2.0e+00 2.0e+00
----------------------|----------------------
3.0e+00 3.0e+00 | 4.0e+00 4.0e+00

The same SYSTEM matrix can be constructed using the pck command.
sys = pck(A,B,C,D);
minfo(sys)
system: 2 states1 outputs2 inputs
seesys(sys)

-1.0e+00 1.0e+00 | 2.0e+00 2.0e+00

-1.0e+00 -3.0e+00 | 2.0e+00 2.0e+00
----------------------|----------------------
3.0e+00 3.0e+00 | 4.0e+00 4.0e+00

8-132
 pck, pss2sys, sys2pss, unpck

Matrices A, B, C, and D can be constructed from sys using unpck.

[A,B,C,D] = unpck(sys);
A
A =
-1 1
-1 3
B
B =
2 2
2 2
C
C =
3 3
D
D =
4 4
mat = sys2pss(sys);
mat
mat =
-1 12 2
-1 -32 2
3 34 4

See Also minfo, nd2sys, vpck, vunpck, zp2sys

8-133
pkvnorm, vnorm

Purpose 8pkvnorm, vnorm

Find the norm or peak norm of a VARYING matrix

Syntax [peak,indv,index] = pkvnorm(matin,p)

out = (matin)
out = (matin,p)

Description pkvnorm sweeps through the independent variable, calculating the norm of
each matrix as specified by the input argument p, following the convention
from MATLAB’s norm command. The default for p is the largest singular value
of matin. The three output arguments all pertain to the peak and its location:
peak value, peak, the independent variable’s value, indv, and the independent
variable’s index, index.
vnorm is a VARYING matrix version of MATLAB’s norm command. The
operation of the norm command is identical to vnorm, except that vnorm also
works on CONSTANT and VARYING matrices, which produces a CONSTANT
or VARYING output. vnorm returns the matrix out with its norm at each
independent variable value.

Examples The two-input/two-output VARYING matrix matin has its independent

variable at 0.2 and 0.6. The vnorm command finds the largest singular value of
the VARYING matrix at each value of the indepedent variable. pkvnorm
returns the largest singular value of the VARYING matrix, its independent
variable value, and the index of the independent variable.
minfo(matin)
varying:2 pts2 rows2 cols
see(matin)
2 rows2 columns
indep variable 0.2
1.0000 0
0 0.5000
indep variable 0.6
4 0
-5 1
nrm = vnorm(matin,2);
see(nrm)

8-134
 pkvnorm, vnorm

1 row 1 column
indep variable 0.2
1
indep variable 0.6
6.4510
[peak,indv,index] = pkvnorm(matin);
peak
peak =
6.4510
indv
indv =
0.6000
index
index ==
2

Algorithm pkvnorm and vnorm call the MATLAB norm command.

See Also norm, vsvd

8-135
ric_eig

Purpose 8ric_eig
Solution of a Riccati equation via eigenvalue decomposition

Syntax [x1,x2,fail,reig_min] = ric_eig(ham,epp)

x = x2/x1;

Description ric_eig (along with a call to x=x2/x1) solves the Riccati equation,

A′X + XA + XRX – Q = 0
with the constraint that the matrix A + RX has all of its eigenvalues in open
left-half plane. The data matrices A, R and Q come from the input Hamiltonian
matrix, ham, in the form

ham =
A R
Q – A′

and it is assumed that R = R′, Q = Q′.

If ham has no jω axis eigenvalues, then there exists n × n matrices x1 and x2
such that [x1; x2] spans the n-dimensional stable, invariant subspace of ham.
If the matrix x1 is indeed invertible, then X := x2 * x1-1 satisfies the Riccati
equation and results in A + RX being stable. It is the only such matrix with
these properties.
ric_eig has internal error checking and returns a fail value of 1 if jω axis
eigenvalues of ham are found. If this occurs, there is no n-dimensional, stable
invariant subspace, and hence no stabilizing Riccati solution. An eigenvalue is
considered to be purely imaginary if the magnitude of the real part is less than
epp. The minimum real part of the eigenvalues is returned in reig_min. epp is
an optional argument and its default value is 1e – 10.

Note When a Hamiltonian has repeated eigenvalues, solving the Riccati

equation via the eigenvalue method may have problems. This is due to the
MATLAB command eig incorrectly selecting the eigenvectors associated with
the repeated roots.

8-136
 ric_eig

Algorithm Under the assumption that the Hamiltonian matrix has a full set of
eigenvectors, the stable-invariant subspace is spanned by the eigenvectors
associated with the stable eigenvalues. Hence, an eigenvalue-eigenvector
decomposition can obtain the stable invariant subspace of the Hamiltonian
matrix, ham. Assuming there are no jω axis eigenvalues, and that there is a full
set of eigenvectors, the two components, x1 and x2, can be generated by
choosing the eigenvectors associated with the stable eigenvalues. The ric_eig
subroutine operates on the assumption that the Jordan form of the
Hamiltonian is diagonal, and returns the stable invariant subspace, as
spanned by the eigenvectors, in the two block form described above.

See Also eig, h2syn, hinfsyn, hinffi, ric_schr

8-137
ric_schr

Purpose 8ric_schr
Solve a Riccati equation via Schur decomposition

Syntax [x1,x2,fail,reig_min] = ric_schr(ham,epp)

x = x2/x1

Description ric_schr (along with a call to x=x2/x1) solves the Riccati equation,

A′X + XA + XRX – Q = 0
such that A + RX is stable. A real Schur decomposition can obtain the stable
invariant subspace of the Hamiltonian matrix, ham. The data matrices A, R,
and Q come from the input Hamiltonian matrix in the form

ham =
A R
Q – A′

and it is assumed that and it is assumed that R = R′, Q = Q′.

If ham has no jω axis eigenvalues, then there exists n × n matrices x1 and x2
such that [x1; x2] spans the n-dimensional stable, invariant subspace of ham.
If x1 is invertible, then X := x2 * x1-1 satisfies the Riccati equation and
results in A + RX being stable. The output flag fail is nominally 0. If there are
jω-axis eigenvalues, fail is set to 1. If there are an unequal number of positive
and negative eigenvalues, fail is set to 2, and if both conditions occur, fail =
3.

Algorithm ric_schr calls csord to produce an ordered complex Schur form, which is
converted to a real Schur form, and yields a stable, invariant subspace of the
Hamiltonian. The csord command orders the solution with negative real
eigenvalues in the top half of the matrix and the positive real eigenvalues on
the bottom, and returns the stable solution. The input matrix is assumed to be
a Hamiltonian matrix of size 2n with n stable eigenvalues and n unstable
eigenvalues. The minimum real part of the eigenvalues is output to reig_min.
epp is an optional argument and its default value is 1e-10.

See Also csord, h2syn, hinfsyn, hinffi, ric_eig, schur

8-138
 rifd

Purpose 8rifd
Display the real, imaginary, frequency and damping ratios of a CONSTANT
input vector

Syntax rifd(vec)

Description rifd displays the real, imaginary, frequency, and damping ratios of a
CONSTANT input vector. The ith frequency is given by the

1
2 2 ---
( real(sys(i)) + imag(sys(i) ) 2

and its damping ratio is defined as real(sys(i))/ω(i).

Examples Display the poles of the SYSTEM matrix, sys.

sys = pck(rand(4,4),rand(4,2),rand(1,4));
rifd(spoles(sys))

real imaginary frequency damping

-7.6287e-02 0.0000e+00 7.6287e-02 l.0000e+00
1.0820e–01 –4.6815e–01 4.8049e–01 –2.2519e–01
1.0820e–01 4.6815e–01 4.8049e–01 –2.2519e–01
1.4095e+00 0.0000e+00 1.4095e+00 -l.0000e+00

See Also eig, imag, spoles, real

8-139
samhld

Purpose 8samhld
Create a sample-hold approximation of a continuous system

Syntax discout = samhld(sys,T)

Description samhld applies a sample-hold to the input of the continuous-time systems sys,
and samples the output, to produce a discrete-time system, discout. The
sampling time is the same at the input and output, and is specified by T.

Examples Construct the system sys via the nd2sys command and verify that all of its
poles are in the closed left-half plane. Perform a sample hold of the system for
a 200 Hz sample rate. All the poles for the discretized system, discout, are
within the unit disk.
sys = nd2sys([1 2 4 5],[2 7 9 2]);
spoles(sys)

-1.6114 + 1.0049i
-1.6114 - 1.0049i
-0.2773
discout = samhld (sys ,1 / 2 0 0)
seesys (discout, '%8 .4f ')

0.9826 -0.0223 -0.0050 | 0.0050

o.0050 0.9999 -0.0000 | 0.0000
0.0000 0.0050 1.0000 | 0.0000
-----------------------------|---------
-0.7500 -0.2500 2.0000 | 0.5000

spoles(discout)

0.9920 + 0.0050i
0.9920 - 0.0050i
0.9986

Algorithm Suppose that the continuous-time system has state-space representation

sys = A B .
C D

8-140
 samhld

If u(t) is held constant over the interval [kT,(k + 1)T], then over that interval,
the state evolution is governed by the differential equation

x· = A B x , x ( kT ) = x k ,
u· 0 nu × n 0nu × n u u ( kT ) uk

which captures the behavior of the continuous-time system, over one sample
period, while the input u(t) is held constant.

A B
Let à :=
0nu × n 0nu × n

( n + nu ) × ( n + nu ) ÃT
Define W ∈ R as W := e . Then W appears as

W 11 W 12
W =
0 I

Define Adisc := W11, and Bdisc := W12, and define the discretized system as

A disc B disc
discout =
C D

See Also exp, expm, tustin

8-141
scliv

Purpose 8scliv
Scale the independent variable values of a VARYING matrix with an affine
transformation

Syntax vout = scliv(vin,factor,offset)

Description scliv scales the independent variable of a VARYING matrix in the following
manner. Let indvi and newindvi denote the independent variable’s ith value,
before and after applying scliv. Then, for each i, they are related as
newindvi = (factor × indvi) + offset

The default value for offset is zero.

Examples Scale the independent value of vin by a factor of 3 and offset it from its original
value by 0.5.
seeiv(vin)
l.000e+002.000e+003.000e+00 4.000e+00 5.000e+00
vout = scliv(vin,3,0.5)
seeiv(vout)
3.500e+006.500e+009.500e+00 1.250e+01 1.550e+01

See Also getiv, seeiv

8-142
 sdhfnorm

Purpose 8sdhfnorm
sdhfnorm calculates the induced norm of a sampled-data system

Syntax [gaml,gamu]=sdhfsyn(p,k,h,delay,tol)

Description sdhfnorm is concerned with the control of a continuous-time system by a

discrete-time controller. The continuous-time interconnection structure
structure, p of type SYSTEM, has state-space realization partitioned as
follows.

A B1 B2
p = C1 0 0
C2 0 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 is assumed to be zero.
sdhfnorm calculates the maximum gain from the L2 norm of the disturbance
inputs to the L2 norm of the error outputs.
Input arguments:
p SYSTEM interconnection structure matrix, (continuous-time)
k discrete-time controller
h sampling period
delay number of samples computational delay (default = 0)
(integer ≥ 0 with default =0)
tol required relative accuracy

Output arguments:
gaml lower bound on the norm
gamu upper bound on the norm

8-143
sdhfnorm

Examples An illustrative example is given in the “Discrete-Time and Sampled-Data H•

Control” section in Chapter 3.

Algorithm sdhfnorm uses variations of the formulae 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 γ-value. This requires a
search and is, computationally, a relatively expensive step.

Subroutines called. dhfsyn, ham2schr, and compnorm

Reference 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. AC–37, pp. 418-–435, 1992.

See Also dhfsyn, hinfsyne, hinffi, hinfnorm, hinfsyn, h2syn, h2norm, ric_eig,
ric_schr

8-144
 sdhfsyn

Purpose sdhfsyn computes an H∞ controller for a sampled-data SYSTEM

8sdhfsyn

interconnection matrix

Syntax [k,gfin] = sdhfsyn(p,nmeas,ncon,gmin,gmax,tol,h,...

delay,ricmethd,epr,epp)

Description sdhfsyn is concerned with the control of a continuous-time system by a

discrete-time controller. The continuous-time interconnection structure
structure, p of type SYSTEM, has state-space realization partitioned as follows

A B1 B2
p = C1 0 0
C2 0 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 is assumed to be zero.
sdhfsyn synthesizes a discrete-time controller to achieve a given norm (if
possible) or find the minimum possible norm to within some tolerance.
sdhfsyn provides a γ iteration using the bisection method. Given a high and low
value of γ, gmax and gmin, the bisection method is used to iterate on the value
of γ in an effort to approach the optimal H∞ control design. If gmax = gmin, only
one γ value is tested. The stopping criteria for the bisection algorithm requires
the relative difference between the last γ value that failed and the last γ value
that passed be less than tol. You can select either the eigenvalue or Schur
method for solution of the Riccati equations with and without balancing. The
eigenvalue method is faster but can have numerical problems, while the Schur
method is slower but generally more reliable.
The algorithm employed calculates an equivalent purely discrete-time problem
for each value of γ and then calls dhfsyn with γ = 1. The screen printing is then
derived from the tests performed by dhfsyn.

8-145
sdhfsyn

Input arguments
p SYSTEM interconnection structure matrix
nmeas number of measurements output to controller
ncon number of control inputs
gmin lower bound on γ
gmax upper bound on γ
tol relative difference between final γ values
delay number of samples computational delay (default = 0)
h time between samples
ricmethod 1 Eigenvalue decomposition with balancing
–1 Eigenvalue decomposition with no balancing
2 Schur decomposition with balancing (default)
–2 Schur decomposition with no balancing
epr measure of when a real part of an eigenvalue of the
Hamiltonian matrix is zero (default epr = 1e–10)
epp positive definite determination of the X∞ and Y∞ solution
(default epp = 1e–6)

Output arguments
k H∞ (sub) optimal controller
gfin final γ value achieved

You might design a first controller using the dhfsyn function on the SYSTEM
(samhld(p,h)), followed by sdhfnorm to determine an upper bound gmax to use
for the start of this sampled data control design iterative process.

8-146
 sdhfsyn

The sdhfsyn program outputs several variables, which can be checked to

ensure that the above conditions are being met. For each γ value the minimum
magnitude, real part of the eigenvalues of the H Hamiltonian matrices is
displayed along with the minimum eigenvalue of X∞, which is the solution to
the X Riccati equation. A # sign is placed to the right of the condition that failed
in the printout. This additional information can aid you in the control design
process.

Examples An illustrative example is given in the “Discrete-Time and Sampled-Data H•

Control” section in Chapter 3.

Algorithm sdhfsyn uses variations of the formulae 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 γ-value, this requires a
search and is computationally a relatively expensive step.

Subroutines called. dhfsyn, ham2schr, compnorm

Reference 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. AC–37, pp. 418–435, 1992.

See Also dhfsyn, hinfsyne, hinffi, hinfnorm, hinfsyn, h2syn, h2norm, ric_eig,
ric_schr

8-147
see, seeiv

Purpose 8see, seeiv

Display a SYSTEM or VARYING matrix

Syntax see(mat,iv_low,iv_high)
see(matin)
seeiv(mat)

Description see displays the A, B, C, and D matrices of matin for a SYSTEM matrix or the
independent variable and the matrix at that variable if matin is a VARYING
matrix. iv_low and iv_high are the optional range of the independent
variables to be displayed. see displays the matrix itself if the input is
CONSTANT.
seeiv displays only the independent variable of the input VARYING matrix
mat. An error message is displayed if the input matrix is not a VARYING
matrix.

Examples The see command displays any type of matrix. An example of its use for
SYSTEM and VARYING matrices follows.
see(sys)
A matrix
1 1
1 1
press any key to move to B matrix
B matrix
2 2
2 2
press any key to move to C matrix
C matrix
3 3
press any key to move to D matrix
D matrix
4 4
sysg = frsp(sys,[0.4 0.9]);
see(sysg)
1 row2 columns
iv = 0.4
-1.7692 - 1.1538i-1.7692 - 1.1538i
iv = 0.9
-0.9896 - 2.2453i-0.9896 - 2.2453i

8-148
 see, seeiv

See Also getiv, mprintf, rifd, seesys, sortiv

8-149
seesys

Purpose 8seesys
Display a SYSTEM or VARYING matrix with sprintf formatting

Syntax seesys(matin,'format')

Description seesys displays a SYSTEM matrix in packed form if matin is a SYSTEM

matrix, or the independent variable and the matrix at that variable if matin is
a VARYING matrix. The optional string format specifies the format exactly as
in the MATLAB function sprintf. seesys is similar to see, but prints a
SYSTEM matrix in packed form and gives more control over formatting. The
default format is %.1e. An incorrect format string can cause erroneous output
displays. See sprintf and mprintf for more details.

Examples The seesys command displays any type of matrix. An example of its use for
SYSTEM and VARYING matrices follows.
seesys(sys)

l.0e+00 l.0e+00 | 2.0e+00 2.0e+00

l.0e+00 l.0e+00 | 2.0e+00 2.0e+00
--------------------|---------------------
3.0e+00 3.0e+00 | 4.0e+00 4.0e+00

seesys(sys,'%1.0f')

1 1 | 2 2
1 1 | 2 2
-------|--------
3 3 | 4 4

sysg = frsp(sys, [0.4 0.9]);

seesys (sysg)
1 row 2 columns
iv = .4
-1. 8e+00 -1. 8e+00
iv = .9
-9.9e-01 -9.9e-01

See Also getiv, mprintf, rifd, see, sortiv

8-150
 sel, reordsys

Purpose 8sel, reordsys

Select rows/outputs and columns/inputs from a CONSTANT, SYSTEM or
VARYING matrix or reorder the SYSTEM states

Syntax out = sel(mat,rows,cols)

out = sel(sys,outputs,inputs)
sysout = reordsys(sys,index)

Description sel selects desired rows and columns from a CONSTANT/VARYING matrix,
or outputs and inputs from a SYSTEM matrix. For CONSTANT and VARYING
matrices, the rows and cols input arguments are row vectors with the desired
rows/columns of mat specified. For SYSTEM matrices, outputs and inputs are
row vectors with the desired inputs/outputs specified. Use the string ':' to
specify all rows (inputs) and/or columns (outputs).
reordsys reorders the states of SYSTEM matrix sys as defined by the vector
of position variables, index. The index variable is restricted to be the same
length as the number of states of sys. This command can be used in conjunction
with strans and sresid to reduce the states of a SYSTEM matrix.

Examples You can use the sel command with any matrix type. First, construct and
display a one state, two output, three input SYSTEM matrix.
minfo(sys)
system:1 states2 outputs3 inputs
seesys(sys)

1.0e+00 | 2.0e+00 3.0e+00 4.0e+00

---------|------------------------------
5.0e+00 | 6.0e+00 7.0e+00 8.0e+00
9.0e+00 | 1.0e+01 1.1e+01 1.2e+01

8-151
sel, reordsys

Reorder the outputs of sys to be output 2, 1 and repeat output 2; also reorder
the inputs to be input 3, 1 and 2.
sys2 = sel(sys,[2 1 2],[3 1 2]);
minfo(sys2)
system:1 states3 outputs3 inputs
seesys(sys2)

1.0e+00 | 4.0e+00 2.0e+00 3.0e+00

---------|------------------------------
9.0e+00 | 1.2e+01 1.0e+01 1.1e+01
5.0e+00 | 8.0e+00 6.0e+00 7.0e+00
9.0e+00 | 1.2e+01 1.0e+01 1.1e+01

The same manipulations can be done on a VARYING matrix.

sysg =frsp(sys,logspace(-1,0,2));
see(sysg)
2 rows3 columns
indep variable 0.1

–3.9010 – 0.9901i –7.8515 – 1.4851i –11.8020 – 1.9802i

–7.8218 – 1.7822i –15.7327 – 2.6733i –23.6436 – 3.5644i

indep variable 1

1.0000 – 5.0000i –0.5000 – 7.5000i –2.0000 – 10.0000i

1.0000 – 9.0000i –2.5000 – 13.5000i –6.0000 – 18.0000i

Select the second and first outputs and the third and second inputs and display
them.
part = sel(sysg,[2 1],[3 2]);
see(part)
2 rows2 columns
indep variable 0.1

–23.6436 – 3.5644i –15.7327 – 2.6733i

–11.8020 – 1.9802i –7.8515 – 1.4851i

8-152
 sel, reordsys

indep variable 1

–6.0000 – 18.0000i –2.5000 – 13.5000i

–2.0000 – 10.0000i –0.5000 – 7.5000i

The command reordsys interchanges states in a SYSTEM matrix. The matrix

(sys) has four states, two inputs, and one output and it is to be reordered so
that states three and four interchange with states one and two in the
state-space system.
minfo(sys)
system: 4 states1 outputs2 inputs
seesys(sys)

1.5e-01 -1.3e-01 0.0e+00 0.0e+00 | 2.4e-01 2.3e-01

1.3e-01 1.5e-01 0.0e+00 0.0e+00 | -4.4e-01 -2.2e-01
0.0e+00 0.0e+00 -6.2e-01 0.0e+00 | -6.5e-01 -3.9e-01
0.0e+00 0.0e+00 0.0e+00 2.0e+00 | 1.2e+00 7.2e-01
----------------------------------------|--------------------
-9.7e-02 -1.2e-01 -1.3e-01 1.1e+00 | 0.0e+00 0.0e+00

sysout = reordsys(sys,[3 4 1 2]);

minfo(sysout)
system: 4 states1 outputs2 inputs
seesys(sysout)

-6.2e-01 0.0e+00 0.0e+00 0.0e+00 | -6.5e-01 -3.9e-01

0.0e+00 2.0e+00 0.0e+00 0.0e+00 | 1.2e+00 7.2e-01
0.0e+00 0.0e+00 1.5e-01 -1.3e-01 | 2.4e-01 2.3e-01
0.0e+00 0.0e+00 1.3e-01 1.5e-01 | -4.4e-01 -2.2e-01
---------------------------------------|---------------------
-1.3e-01 1.1e+00 -9.7e-02 -1.2e-01 | 0.0e+00 0.0e+00

See Also getiv, sresid, strunc, unpck, vunpck

8-153
siggen

Purpose 8siggen
Generate VARYING matrix functions

Syntax y = siggen('function(t)',t)

Description siggen is a general purpose signal generator. You can provide a timebase with
the argument t, and the function to be evaluated with the first argument (a
string), function(t). The output, y, will be a VARYING matrix. t could also be
VARYING, in which case the timebase is the independent variables contained
in t.
function(t) is not necessarily dependent on t. In the cases where it doesn’t
depend on t, siggen can be slow. This is because function is evaluated with a
MATLAB eval call for every element in t. For example, consider generating a
random vector. The command
u = siggen('rand(size(t))',[0:100]);

generates a 1 × 1 VARYING matrix with 101 independent variables. This is

very fast because it depends on t. To a 2 × 1 VARYING matrix of random
values, you can use the command
u = siggen('rand(2,1)',[0:100]);

This is slow because 101 evaluations of rand(2,1) are performed in a MATLAB

for loop. For vectors of random signals, a much faster alternative is to use vpck
with a random matrix of the appropriate size. In the above example this would
be
u = vpck(rand(202,1),[0:100]);

Examples The first example illustrates what is perhaps the most common use of siggen.
A single-input single-output signal is created from MATLAB mathematical
functions. It is important to use t as the independent variable in the function
string.
timebase = [0:0.05:10];
y1 = siggen('exp(0.1*t) - sin(3*t)',timebase);
minfo(y1)
varying:201 pts1 rows1 cols
vplot(y1)
title('siggen example: function depends on t')

8-154
 siggen

Siggen example: function depends on t

4

3.5

2.5

1.5

0.5

0
0 1 2 3 4 5 6 7 8 9 10

The second example illustrates that the second argument can make use of the
independent values of a VARYING matrix. Note also that the specified function
is independent of t, and is executed at each instance of t. This example is
included to illustrate that the function string need not depend on t. In practice
the string rand(size(t)) is orders of magnitude faster than rand.
y2 = siggen('rand',y1);
minfo(y2)
varying:201 pts1 rows1 cols
vplot(y2)
title('siggen example: function independent of t')

8-155
siggen

Siggen example: function independent of t

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7 8 9 10

The use of siggen is not restricted to single-input, single-output signals. The

following example creates a 2 × 2 VARYING matrix. In this example the matrix
elements are all functions of t and the results plotted in the figure. They could
equally well have all been independent of t, for example 'rand(3,2)' could be
the function string. (Note that they must all be one type or the other (this
restriction does not apply if t is a single number).
func = '[t/max(t) 2*cos(3*t+0.2); 2+0.1*sin(2*t) ';
func = [func ' sqrt(t)+0.3*rand(size(t))]'];
y3 = siggen(func,timebase);
minfo(y3)
varying:201 pts2 rows2 cols
vplot(y3)
title('Siggen example: 2 x 2 varying matrix')

8-156
 siggen

Siggen example: 2 x 2 varying matrix

4

-1

-2
0 1 2 3 4 5 6 7 8 9 10

siggen cannot generate stair-step signals with user-specified values. You can
do this however (for single-input/single-output signals) using the command
vpck or the µ-Tools commmand step_tr. Vectors of signals can be created with
vpck and abv. The following example demonstrates the use of vpck. vinterp is
used to plot a meaningful representation of the signal.
y4 = vpck([0:10]',[0:2:20]');
minfo(y4)
varying:11 pts1 rows1 cols
vplot(vinterp(y4,0.1))
title('Siggen example: step function')

8-157
siggen

Siggen example: step function

10

0
0 2 4 6 8 10 12 14 16 18 20

See Also cos_tr, sin_tr, step_tr, vpck, vinterp

8-158
 simgui

Purpose 8simgui
A graphical user interface for time simulations of linear fractional
transformations

Syntax simgui

Description simgui provides the ability to simulate linear fractional models and plot their
responses. The standard linear fractional model considered is shown below.

Standard Linear Fractional Model

The P block corresponds to the open-loop plant interconnection model, referred

to as Plant in simgui. The K block corresponds to a state-space controller,
Controller in simgui. The ∆ block corresponds to the perturbation to the model,
Perturbation in simgui. The Input Signal to the time simulation is denoted by
U in the figure. The individual systems are formed using the starp command.
The time simulation outputs available for plotting correspond to the variable
Y. Three types of simulation are possible: continuous-time using trsp,
discrete-time using dtrsp, and sample-data using sdtrsp.
simgui has two interface windows and up to six plot windows:

• Main Simulation window, which is the main interface for the user.
• Parameter window, which is used to modify properties of the time
simulation, such as the final time, integration step size, initial conditions,
and which variables are automatically exported to the workspace.
• Plot windows, where the plots of time responses are displayed. You can open
up to six of these windows.

8-159
simgui

A detailed description of the use and application of simgui is provided in the

“LFT Time Simulation User Interface Tool: simgui” section in Chapter 6.

Examples An example of using simgui for simulation of a linear fractional transformation

is shown in the “LFT Time Simulation User Interface Tool: simgui” section in
Chapter 6.

See Also dtrsp, sdtrsp, starp, trsp, vplot

8-160
 spoles

Purpose 8spoles
Calculate the eigenvalues of a SYSTEM A matrix

Syntax out = spoles(sys)

Description spoles returns the eigenvalues of the A matrix from the SYSTEM matrix sys.

Examples Find the poles of the two input, one output, three state SYSTEM matrx sys.
A = [1 1 1; 3 1 1; 1 1 -2];
B = 2*ones(3,2);
C = 3*ones(1,3);
D = 4*ones(1,2);
sys = pck(A,B,C,D);
minfo(sys)
system: 3 states1 outputs2 inputs
spoles(sys)
3.1474
-0.8186
-2.3289
eig(A)
ans =
3.1474
-0.8186
-2.3289

Algorithm spoles uses the MATLAB command schur to find the eigenvalues of the
SYSTEM A matrix. This is a more numerically reliable method than using the
eig function.

See Also eig, rifd, schur, szeros

8-161
srelbal, sfrwtbal, sfrwtbld, sncfbal, sdecomp

Purpose 8srelbal, sfrwtbal, sfrwtbld, sncfbal, sdecomp

srelbal calculates the stochastically balanced realization of a SYSTEM
matrix, sfrwtbal, the frequency weighted balanced realization for performing
relative error, and sfrwtbld, the frequency weighted model order reduction.
sncfbal calculates the normalized coprime factorizations. sdecomp decomposes
a SYSTEM matrix as the sum of stable and unstable systems.

Syntax [sysb,relsv,sysfact] = srelbal(sys,tol)

[sys1,sig1] = sfrwtbal(sys,wt1,wt2)
syshat = sfrwtbld(sys1hat,wt1,wt2)
[sysnlcf,signcf,sysnrcf] = sncfbal(sys,tol)
[sysst,sysun] = sdecomp(sys,bord,fl)

Description srelbal performs a stochastically balanced realization of the input SYSTEM

matrix. sys must be stable and be of full column rank at infinite frequency, but
not necessarily square or minimum phase. Difficulties may occur if sys has
zeros on the imaginary axis. sysb will have the same transfer function as sys,
and sysfact gives the stable minimum phase system such that sys~ sys =
sysfact sysfact~. Note that sys~ denotes cjt(sys). If [a,b,c,d] =
unpck(sysb) and [af,bf,cf,df] = unpck(sysfact), then the realization
[a,bf,c,0] will be balanced with Hankel singular values relsv, and will also
equal the stable part of sys*sysfact~–1. A reduced-order system can be
obtained by strunc(sysb,k) that will have guaranteed performance in the
relative error.
sfrwtbal performs a frequency-weighted balancing. It calculates the stable
part of (wt1) ~–1 * sys* (wt2)~–1 and sys1 is a balanced realization of this,
with Hankel singular values sig1. wt1 and wt2 must be stable and minimum
phase, square and of compatible dimensions with sys. wt2 has the identity as
default value. sys must be stable. The resulting system sys1 can then be
approximated by sys1hat of order k using, for example, hankmr, and an
approximation syshat to sys is obtained by sfrwtbld, which calculates the
stable part of (wt1)~ *sys1hat*(wt2)~ using sdecomp.
A general lower bound on the frequency weighted approximation error is given
by
(wt1) ~–1(sys - syshat)(wt2)~–1||∞ ≥ sig1(k + 1)

where in the relative error case wt1 is the identity and wt2 = sysfact.

8-162
 srelbal, sfrwtbal, sfrwtbld, sncfbal, sdecomp

sncfbal produces balanced realizations of the normalized left and right

coprime factorizations of a SYSTEM matrix. That is for a transfer function G,
Nr
balanced state-space realizations are calculated for N l M l and , where
Mr
~ ~ ~ ~ –1 –1
N l N l + M l M l = I , N r N r + M r M r = I , G = M l N l = N r M r , and

N l, M l, N r , and M r are all stable. The Hankel singular values of both N l M l

Nr
and are given by the column vector signcf. Model reduction for these
Mr
systems can then be performed using strunc or hankmr. The method is well
suited to plant or controller reduction in feedback systems.
sdecomp decomposes a system into the sum of two systems, sys =
madd(sysst,sysun). sysst has the real parts of all its poles < bord and sysun
has the real parts of all its poles ≥ bord. bord has default value 0. The D matrix
for sysun is zero unless fl = 'd' when that for sysst is zero.
srelbal, sfrwtbal , sfrwtbld, sncfbal, and sdecomp are restricted to be used
on continuous-time SYSTEM matrices.
( s + 1 ) ( s + 10 ) ( s + 90 )
Examples Given the system sys = -------------------------------------------------------------- reduce the system to two and
( s + 2 ) ( s + 91 ) ( s + 100 )
one states, respectively. An approximate system of order 1 or 2 can be obtained
as follows.
sys = zp2sys([-1 -10 -90],[-2 -91 -100]);
[sysb,relsv,sysfact] = srelbal(sys);
disp(relsv')
8.5985e-012.0777e-012.1769e-04
sysrel1 = strunc(sysb,1);
sysrel2 = strunc(sysb,2);

The relative error in the second-order model will be negligible since relsv(3) is
very small; however, with a first-order model, it will be substantial.

8-163
srelbal, sfrwtbal, sfrwtbld, sncfbal, sdecomp

The reduced-order models of order k can be obtained in the frequency weighted

case as follows.
wt1 = nd2sys([1 10],[1 1]);
[sys1,sig1] = sfrwtbal(sys,wt1);
disp(sig1');
4.1873e-014.6472e-031.0280e-04
sys1hat = hankmr(sys1,sig1,1,'d');
syshat = sfrwtbld(sys1hat,wt1);
disp(hinfnorm(mmult(msub(sys,syshat),minv(wt1))));
4.6471e-034.6517e-03Inf

In this example the method nearly reaches the lower bound, but this cannot be
claimed in general.
Now consider approximating the unstable third order system,

10
sys = ----------------------------------------
s ( s – 1 ) ( s + 10 )

using sncfbal. First the balanced realization of the normalized left coprime
factors is calculated, then this is truncated to two states and the reduced-order
system recovered from these normalized coprime factors using starp.
sys = zp2sys([],[0 1 -10],10);
[sysnlcf,signcf] = sncfbal(sys);
disp(signcf')
9.6700e-015.2382e-012.3538e-02
sysnlcfr = strunc(sysnlcf,2);
sysr = starp(mmult([1;1],msub(sysnlcfr,[0 1])),-1,1,1)

If this is the transfer function of a plant to be controlled, then signcf(1) can

be used to predict the possible robust stability to perturbations in the coprime
factors, and the potential for model-order reduction of the controller is given by
signcf(2:3), McFarlane and Glover (1989). In this example the maximum
stablizable perturbations in the coprime factors is given by
1 – signcf ( 1 ) 2 = 0.20770 . Furthermore if a controller is designed to be
optimal for the second-order reduced model, then its stability margin will be at
least
0.25477 - 2 * signcf(3) = 0.20770

8-164
 srelbal, sfrwtbal, sfrwtbld, sncfbal, sdecomp

Algorithm The algorithms are based on the results in the following papers.

Reference Anderson, B.D.O., and Y. Liu, “Controller reduction: Concepts and

Approaches,” IEEE Transactions on Automatic Control, vol. AC–34, pp. 802–
812, 1989.
Desai, U.B., and D. Pal, “A transformation approach to stochastic model
reduction,” IEEE Transactions on Automatic Control, vol. AC–29, pp. 1097–
1100, 1984.
Glover, K., “Multiplicative approximation of linear multivariable systems with
error bounds,” Proceedings of the American Control Conference, Seattle, pp.
1705–1709, 1986.
Latham, G.A., and B.D.O. Anderson, “Frequency-weighted optimal Hankel
norm approximation of state transfer functions,” Systems and Control Letters,
vol. 5, pp. 229–236, 1985.
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.
Wang, W., and M.G. Safonov, “A tighter relative error bound for balanced
stochastic truncation,” Systems and Control Letters, vol. 14, pp. 307–317, 1990.

See Also hankmr, sysbal, sresid, strunc

8-165
sresid, strunc

Purpose 8sresid, strunc

Reduce the state dimension of a SYSTEM matrix

Syntax sysout = sresid(sys,ord)

sysout = strunc(sys,ord)

Description sresid residualizes the last states of a SYSTEM matrix sys. sresid accounts
for the DC contribution of the last columns and rows of the SYSTEM A matrix
and the corresponding rows and columns of B and C. sresid assumes that the
SYSTEM matrix is ordered so that the last states are to be residualized. If the
orignal SYSTEM matrix is partitioned as

A 11 A 12 B 1
p = A 21 A 22 B 2
C1 C2 D

with A11 of size ord × ord, then the command

sysout = sresid(sys,ord)

results in

 A B A 
sysout = pss2sys  – 12 A 22 A 21 B 2 , ord
11 1 –1
 C D C2 
 1 

strunc truncates the states of the input system matrix sys, to a system with
state dimension equal to ord. strunc can be used in conjunction with the model
reduction routines sysbal and hankmr.
The resulting SYSTEM output matrix is
sysout = pss2sys ([A_11 B_1; C_1 D]);

8-166
 sresid, strunc

Examples A two input, one output, four state SYSTEM is reduced down to a two input,
one output, two state SYSTEM via sresid and strunc. The only difference
between the two reduced-order systems is the value of their D matrices.
seesys(sys)
-1.2e-01 0.0e+00 0.0e+00 0.0e+00 | 9.le-01 5.2e-01
0.0e+00 -3.2e-01 0.0e+00 0.0e+00 | 6.le-02 3.2e-01
0.0e+00 0.0e+00 -4.3e+00 0.0e+00 | 9.le-01 9.9e-01
0.0e+00 0.0e+00 0.0e+00 -9.9e+01 | 5.le-01 4.9e-01
---------------------------------------|-------------------
2.7e-01 9.le-02 9.5e-01 7.4e-02 | 0.0e+00 0.0e+00

sys_strunc = strunc(sys,3);
seesys(sys_strunc)
-1.2e-01 0.0e+00 0.0e+00 | 9.le-01 5.2e-01
0.0e+00 -3.2e-01 0.0e+00 | 6.le-02 3.2e-01
0.0e+00 0.0e+00 -4.3e+00 | 9.le-01 9.9e-01
-------------------------------|---------------------
2.7e-01 9.le-02 9.5e-01 | 0.0e+00 0.0e+00

sys_resid = sresid(sys,3)
seesys(sys_resid)

-1.2e-01 0.0e+00 0.0e+00 | 9.le-01 5.2e-01

0.0e+00 -3.2e-01 0.0e+00 | 6.le-02 3.2e-01
0.0e+00 0.0e+00 -4.3e+00 | 9.le-01 9.9e-01
------------------------------|--------------------
2.7e-01 9.le-02 9.5e-01 | 3.8e-04 3.7e-04

sysstrunc = strunc(sys,3); seesys(sysstrunc)

sysresid = sresid(sys,3); seesys(sysresid)

See Also rifd, statecc, strans

8-167
starp

Purpose 8starp
Form the Redheffer star product of two VARYING/SYSTEM/CONSTANT
matrices. The star product is a generalization of a linear fractional
transformation

Syntax sysout = starp(top,bot,dim1,dim2)

Description Connects the two matrices top and bot in the star product loop shown below.

The last dim1 outputs of top are fed to the first dim1 inputs of bot, and the first
dim2 outputs of bot are fed into the last dim2 inputs of top. The remaining
inputs and outputs constitute sysout. By this description, the dimensions must
satisfy
min(dim_out(top),dim_in(bot)) ≥ dim1
min(dim_out(bot),dim_in(top)) ≥ dim2
Further restrictions also arise
IF dim1 = dim_out(top) & dim2 = dim_out(bot)
THEN there are no outputs remaining in the interconnection
IF dim1 = dim_in(bot) & dim2 = dim_in(top)
THEN there are no inputs remaining in the interconnection

8-168
 starp

In either case, it is unclear what to return as sysout, so it is returned empty.

There is one exception to this situation. If either top or bot is a SYSTEM
matrix with a nonzero number of states, and all of the equalities in the both of
the above IF conditions hold (hence there are no inputs or outputs in the
interconnection), then sysout will be a CONSTANT matrix, and will be the A
matrix governing the internal dynamics of the loop.
As usual, the only types of matrices that cannot be combined are SYSTEM
matrices with VARYING matrices.
If only two arguments are given (i.e., no dimensions specified),
out = starp(top,bot);

then the operation is equivalent to

dim1 = min(ynum(top),unum(bot));
dim2 = min(unum(top),ynum(bot));
out = starp(top,bot,dim1,dim2);

Algorithm The “m-Tools Commands for LFTs” section in Chapter 4 provides details of the
star product formulae.

Reference Redheffer, R., “Inequalities for a matrix Riccati equation,” Journal of

Mathematics and Mechanics, vol. 8, no. 3, 1959.

See Also sysic

8-169
statecc, strans

Purpose 8statecc, strans

Apply state-coordinate transformation to a SYSTEM matrix

Syntax sysout = statecc(sysin,t)

[sysout,v] = strans(sys)

Description statecc applies a state coordinate transformation to the matrix, yielding a new
SYSTEM matrix with
sysout = pck(tA*t,tB,C*t,D)

where A, B, C, and D are the state-space entries of the matrix sysin. t is

restricted to be square and have the same dimensions as the A matrix.
strans transforms the A matrix of sys in bidiagonal form with the complex
conjugate roots in real 2 × 2 form. sysout contains the transformed SYSTEM
matrix and v is the transformation matrix. The A matrix is ordered by
increasing magnitude of its eigenvalues. strans calls the MATLAB eig
command to do the reordering.

Note strans may be inaccurate when a SYSTEM/CONSTANT matrix has

repeated eigenvalues. This is due to the potential defective eigensystem, and
the lack of a full set of eigenvectors.

Examples The strans command shows the individual contributions of the modes of the
SYSTEM matrix. In this example sys, which has four states, two inputs and
one output is transformed into bidiagonal form.
see(sys)
A matrix

0.2190 0.9347 0.0346 0.0077

0.0470 0.3835 0.0535 0.3834
0.6789 0.5194 0.5297 0.0668
0.6793 0.8310 0.6711 0.4175

8-170
 statecc, strans

B matrix

0.6868 0.5269
0.5890 0.0920
0.9304 0.6539
0.8462 0.4160

C matrix

0.7012 0.9103 0.7622 0.2625

D matrix

O O

sys=strc
see(sys)

A matrix

-0.0763 0 0 0
0 0.1082 –0.4681 0
0 0 0 1.4095

B matrix

-0.4731 -0.1839
0.5971 0.3199
0.2869 0.5542
-1.7033 -0.8132
C matrix

–0.1150 0.0298 0.3214 –1.0477

D matrix

O O

See Also eig, sclin, sclout, veig

8-171
sysbal, hankmr

Purpose 8sysbal, hankmr

Calculate the balanced realization and optimal Hankel norm approximation of
a SYSTEM matrix. sysbal and hankmr are restricted to be used on
continuous-time SYSTEM matrices whose poles have negative real part

Syntax [sysb,hanksv] = sysbal(sys,tol)

[sysred,sysanti,siganti] = hankmr(sysb,hanksv,k,opt)

Description sysbal performs a truncated balanced realization of the input SYSTEM

matrix. The result is truncated to retain all Hankel singular values greater
than tol. If tol is omitted then it is set to
max(hanksv(1) * 1.0–12,1.0–16)
The second output argument of sysbal is the vector hanksv, which contains the
Hankel singular values of the input system, sys. One method to get a
reduced-order model is to truncate the balanced system sysb using strunc.
hankmr returns sysred, the optimal Hankel norm approximation of order k to
the SYSTEM matrix, sysb, which is a balanced realization with Hankel
singular values hanksv and is of order n (n > k). The fourth optional input
argument, opt, may be omitted in which case sysanti contains the anti-causal
term such that the L∞ norm of (sysb - sysred - sysanti) is hanksv(k+1) or
set to
'a' when sysout also includes the anti-casual term, and sysanti=0
'd' when sysout includes a D matrix to reduce the H∞ error norm of
(sys – sysout)
If the d option is set, the third output argument of hankmr is the vector siganti,
which contains the Hankel singular values of the system, sysanti~. In this
case
hanksv(k+1)≤||sys - sysred||∞≤hanksv(k+1)+sum(siganti)

8-172
 sysbal, hankmr

( s + 10 ) ( s + 90 )
Examples Given the system sys = ----------------------------------------------------------- , reduce the system to two and
( s + 2 ) ( s + 91 ) ( s + 10 )

one states, respectively. First reduce the system to two states.

sys = zp2sys([-10 -90],[-2 -91 -100]);
w = logspace(-1,3,100);
sysg = frsp(sys,w);
[syssb,sv]=sysbal(sys);
sv
ans =
2.0613e-024.1136e-031.2663e-06
sys2s = strunc(syssb,2);
sys2sg = frsp(sys2s,w);
sys2h = hankmr(syssb,sv,2);
sys2hg = frsp(sys2h,w);
vplot('bode',sys_g,sys2sg,sys2hg)
tmp = 'Original 3 state system, 2 state Balanced ';
tmp1 = 'and Hankel Model Reduction';
title([tmp tmp1])

Original 3 state system, 2 state Balanced and Hankel Model Reduction

10 -1
Log Magnitude

10 -2

10 -3

10 -4
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

-20
Phase (radians)

-40

-60

-80

-100
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

8-173
sysbal, hankmr

Notice that there is virtually no difference between the three systems. Now we
will reduce the original system down to one state with sysbal and hankmr.
sys1s = strunc(syssb,1);
sys1sg = frsp(sys1s,w);
sys1h = hankmr(syssb,sv,1);
sys1hg = frsp(sys1h,w);
[syssb,sv] = sysbal(sys);
vplot('bode',sys_g,sys1sg,sys1hg)
tmp = 'Original 3 state system, 1 state Balanced '
tmp1 = 'and Hankel Model Reduction')
title([tmp tmp1])

Original 3 state system, 1 state Balanced and Hankel Model Reduction

10 -1
Log Magnitude

10 -2

10 -3

10 -4
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

-20
Phase (radians)

-40

-60

-80

-100
10 -1 10 0 10 1 10 2 10 3
Frequency (radians/sec)

The original three state system corresponds to the solid line, the one state
balanced realization system corresponds to the dashed line, and the one state
Hankel model reduced system corresponds to the dotted line. There is
significant differences between the models and the two model reduction
techniques. Depending on the model reduction objectives, the one state models
may be inappropriate for use.

8-174
 sysbal, hankmr

Reference Glover, K. “All optimal Hankel-norm approximations of linear multivariable

systems and their error bounds,” International Journal of Control, vol. 39, pp.
1115–1193, 1984.

See Also sdecomp, sfrwtbal, sfrwtbld, sresid, srelbal, sresid

8-175
sysic

Purpose 8sysic
Form linear interconnections of CONSTANT and SYSTEM matrices (or
CONSTANT and VARYING matrices)

Syntax sysic

Description µ-Tools provides a simple linear system interconnection program called sysic.
It forms linear interconnections of CONSTANT and SYSTEM matrices (or
CONSTANT and VARYING matrices, though this can require a lot of memory),
by writing the loop equations of the interconnection.
Using sysic involves setting up several variables in the MATLAB workspace,
and then running the M-file sysic. The variables that are defined delineate the
details of the interconnection.

Variable Descriptions
A list and description of the variables required by sysic follow.

systemnames. This variable is a character string, which contains the names of

the matrices used in the interconnection. The names must be separated by
spaces and/or tabs, and there should be no additional punctuation. Each named
system must exist in the MATLAB workspace at the time the program sysic
is run. The SYSTEM matrices names used within the sysic program are
limited to 10 characters. This limitation is due to the MATLAB 19 character
limitation on the workspace variable names. That is, a SYSTEM matrix named
andygaryjohnkeithroy would be invalid.

inputvar. This variable is a character string, with names of the various external
inputs that are present in the final interconnection. The input names are
separated by semicolons, and the entire list of input names is enclosed in
square brackets [ ]. Inputs can be multivariable signals; for instance a
windgust input, with three directions (x, y, and z) that can be specified by using
windgust{3}. This means that there is a three variable input to the
interconnection called windgust. Alternatively, this could be specified as three
separate, scalar inputs, say wingustx, windgusty, and windgustz. The order
that the input names appear in the variable inputvar is the order that the
inputs will be placed in the interconnection.

8-176
 sysic

outputvar. This variable is a character string, describing the outputs of the

interconnection, which must be linear combinations of the subsystem outputs
and the external inputs. Semicolons are used to separate the channels of the
output variables. 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 (2:5,8,1,9:11) specifies outputs 2,3,4,5,8,1,9,10,11
from the system plant. If no arguments are specified with a system, then it is
assumed that all outputs are being used, and in the order they appear in that
system.

input_to_sys. Each subsystem named in the variable systemnames must have a

variable set to define the inputs to the subsystem. If the system name is
controller, then the variable that must be set should be called
input_to_controller. It is specified in the same manner that the variable
outputvar is set, with inputs consisting of linear combinations of subsystem
outputs and external inputs. Separate channels are separated by semicolons,
and the order of the inputs in the variable should match the order of the inputs
in the system itself.

sysoutname. This character string variable is optional. If it exists in the

MATLAB workspace when sysic is run, then the interconnection that is
created by running sysic is placed in a MATLAB variable whose name is given
by the string in sysoutname. If this variable does not exist in the workspace,
then the interconnection is automatically placed in the variable ic_ms.

cleanupsysic. After running sysic, all of the above variables, which describe the
interconnection, are left in the workspace. These will be automatically cleared
if the optional variable cleanupsysic is set to the character string yes. The
default value of the variable is 'no' which does not result in any of the sysic
descriptions you defined to be cleared. The MATLAB matrices listed in the
variable systemnames are never automatically cleared.

Running sysic
If the variables systemnames, inputvar, and outputvar are set, and for each
name name_i appearing in systemnames, the variable input_to_name_i is set,
then the interconnection is created by running the M-file sysic. Depending on
the existence/nonexistence of the variable sysoutname, the resulting

8-177
sysic

interconnection is stored in your specified MATLAB variable, or the default

MATLAB variable ic_ms.
Within sysic, a series of error-checking routines monitor the consistency and
availability of system matrices and their inputs. These routines provide a basic
level of error detection to aid you in debugging. The input/output dimensions of
the final interconnection are defined by inputvar and outputvar variables.
The syntax of sysic is limited, and for the most part is restricted to what is
shown here. Some additional features are illustrated in the more complicated
demonstration problems. Note that you must keep track of input/output
variables defined for the final interconnection structure.

Examples The HIMAT example provides another example of how to construct

interconnection systems from block diagram descriptions. The interconnection
diagram below corresponds to the HIMAT design example.
Given that there are four SYSTEM matrices, named himat, wdel, wp, and k, in
the MATLAB workspace, each with two inputs, and two outputs. The following
10 lines form the sysic commands to make the interconnection structure
shown below, which is placed in the variable clp. You can execute these at the
command line (as shown) or type them into an M-file. (Note that to run this
example you must create the variables himat, wdel, wp and k.)
systemnames = ' himat wdel wp k ';
inputvar = '[ pertin{2};dis{2}]';
outputvar = '[ wdel ;wp ]';
input_to_himat = '[ k + pertin ]';
input_to_wp = '[ dist + himat ]';
input_to_wdel = '[ k ]';
input_to_k = '[ -dist - himat ]';
sysoutname = 'clp';
cleanupsysic = 'yes';
sysic;

8-178
 sysic

The final interconnection structure is located in clp with two sets of inputs,
pertin and dist, and two sets of outputs w and e, corresponding to the
perturbation and error outputs.

See Also abv, madd, daug, mmult, sbs, sel, starp

8-179
szeros

Purpose 8szeros
Transmission zeros of a SYSTEM matrix

Syntax veczeros = szeros(sys,epp)

Description szeros calculates the transmission zeros of the input SYSTEM matrix, sys.
The output veczeros contains the vector of transmission zeros.
epp is an optional input argument which is used to test the closeness of the
generalized eigenvalues of the randomly perturbed matrices. Its default value
is the machine epsilon. Occasionally zeros at infinity are displayed as very
large values due to numerical accuracy problems.
For a square SYSTEM matrix, [A B; C D], the generalized eigenvalue test
consists of finding the roots of

 
det  A – λI B  = 0
 C D

Algorithm For a square system, the transmission zeros are found via the generalized
eigenvalue problem described above. To solve for the transmission zeros of a
nonsquare SYSTEM matrix, additional random rows or columns are
augmented to the SYSTEM matrix to make it square and the corresponding
zeros are found. This is done twice, and the unchanged generalized
eigenvalues, where the difference between the eigenvalues is less than epp, are
considered to be the transmission zeros of the SYSTEM matrix.

Reference Laub, A.J., and B.C. Moore, “Calculation of transmission zeros using QZ
techniques,” Automatica, vol. 14, pp. 557–563, 1978.

See Also spoles

8-180
 trsp, dtrsp, sdtrsp

Purpose 8trsp, dtrsp, sdtrsp

Compute the time response of linear system

Syntax y = trsp(sys,u,tfinal,int,x0)
y = dtrsp(dsys,u,T,tfinal,x0)
[output,y,u] = sdtrsp(sys,k,input,T,tfinal,int,x0,z0)

Description trsp computes the time response of the continuous-time system, sys, with the
input, u. The input, u, is a VARYING matrix, which contains the input signal
vector at certain points in time. The input can be irregularly spaced in the
independent variable or a constant, in which case it is assumed to occur at t =
0.
The final time, tfinal, is an optional argument. If omitted, it defaults to the
maximum time in u. The time response is calculated as though the input is a
constant value between the points specified in u. If tfinal is greater than the
largest independent variable in u, the input is held at the last value in u.
For continuous-time evaluation (trsp), you can optionally specify an
integration time with the variable int. If this is omitted, or is equal to zero, an
appropriate value is calculated and displayed. The calculated integration time
depends on the minimum spacing in the input and the fastest dynamics in sys.
int will also be the independent variable step size in the regularly spaced
output, y. If a coarser output is adequate, it can be obtained with the function
vdcmate.

Initial conditions can optionally be specified with the argument, x0. This
specifies the state vector at the first time point in the input vector. If x0 is
omitted, or is a zero scalar, then it is assumed to be a zero vector.
trsp interpolates the input with a zero-order hold of step size equal to int,
discretizes the output at this same step size, and calculates the response from
the initial time to tfinal in steps of int.
dtrsp calculates the response for a discrete-time system, dsys. The time (for
the independent variable) between discrete indices is T. If the input is not
regularly spaced at intervals of time T, it is interpolated. tfinal and x0 behave
in the same manner as for trsp.
sdtrsp calculates a sampled-data time response for a closed-loop system with
a continuous generalized plant (sys) and a discrete controller (K). The
interconnection is illustrated below.

8-181
trsp, dtrsp, sdtrsp

The signals, output, y, and u are calculated. T is the sampled-data controller

sample time, and you have the same input and tfinal options as the trsp
function. Similarly, an integration step size, int, can optionally be specified for
the continous part of the simulation. Initial conditions can be specified for sys
(x0) and K (z0).

Examples A simple SISO system illustrates the use of trsp. This example shows the
consequences of the input being assumed to be constant between time points.
sys = pck(-1,1,1);
minfo(sys)
system:1 states1 outputs1 inputs
u = vpck([0:10:50]',[0:10:50]');
y = trsp(sys,u,60);
integration step size: 0.1
interpolating input vector (zero order hold)
minfo(y)
varying:601 pts1 rows1 cols
vplot(u,'-.',y,'-')
xlabel('time: seconds')
text(10,20,'input')
text(25,10,'output')

8-182
 trsp, dtrsp, sdtrsp

50

45

40

35

30

25
input
20

15
output
10

0
0 10 20 30 40 50 60

Time: seconds

At first glance the output does not seem to be consistent with the plotted input.
Remember that trsp assumes that the input is held constant between specified
values. The vplot and plot commands display a linear interpolation between
points. This can be more clearly seen by displaying the input signal
interpolated to at least as small a step size as the default integration step (here
0.1 seconds).
vplot(u,'-.',vinterp(u,0.1),'--',y,'-')
xlabel('time: seconds')
text(5,44,'dash-dot: input')
text(5,40,'dashed: interpolated input')
text(5,36,'solid: output')

8-183
trsp, dtrsp, sdtrsp

50

45 dash-dot: input
dashed: interpolated input
40
solid: output
35

30

25

20

15

10

0
0 10 20 30 40 50 60

Time: seconds

The staircase nature of the input is now evident. If you really want to have a
ramp input, the function vinterp also provides linear interpolation. A linearly
interpolated input is used in the following example.
uramp = vinterp(u,0.1,60,1);
minfo(uramp)
varying:601 pts1 rows1 cols
yramp = trsp(sys,uramp);
integration step size: 0.1
vplot(uramp,'-.',yramp,'-')
xlabel('time: seconds')
text(20,15,'output')
text(12,20,'input')

8-184
 trsp, dtrsp, sdtrsp

50

45

40

35

30

25
input
20
output
15

10

0
0 10 20 30 40 50 60

Time: seconds

Note that because the input is regularly spaced, with spacing less than or equal
to the default integration time, the input is not interpolated by trsp. Since no
final time was specified in the trsp argument list, and 60 seconds was specified
to vinterp as the final time, this becomes the last time in the input vector
uramp.
To illustrate the use of dtrsp, a bilinear transformation generates a digital
system. The sample time is chosen as 1 second. The output is plotted against a
1 second interpolation of the input.
T = 1;
dsys = tustin(sys,T);
ydig = dtrsp(dsys,u,T);
vplot(ydig,'-',vinterp(u,1),'-.')
xlabel('time: seconds')

8-185
trsp, dtrsp, sdtrsp

50

45

40

35

30

25

20

15

10

0
0 10 20 30 40 50 60

Time: seconds

To illustrate the use of sdtrsp, consider the application of a discrete controller

to a double 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 dtrsp gives the system
response at the sample points. sdtrsp is then used to calculate the intersample
behavior.
P = nd2sys(1,[1,0,0]);
T = 1.0/20;
C=pck([-1.5 T/4; -2/T -.5],[ .5 2;1/T 1/T], ...
[-1/T2 -1.5/T], [1/T2 0]);

The closed-loop digital system is now set up.

Pd = samhld(P,T);
systemnames = 'Pd C';
inputvar = '[ref]';
outputvar = '[Pd]';
input_to_Pd = '[C]';
input_to_C = '[ref ; Pd]';
sysoutname = 'dclp';
cleanupsysic = 'yes';
sysic;

8-186
 trsp, dtrsp, sdtrsp

dtrsp is used to simulate the digital step response.

ustep = step_tr(0,1,T,20*T);
y = dtrsp(dclp,ustep,T);

The continuous interconnection is set up and the sampled data response is

calculated with sdtrsp.
M = mmult([0,1;1,0;0,1],daug(1,P));
y1 = sdtrsp(M,C,ustep,T);
vplot(y,'*',y1,'-')
axis([0,1,0,1.5])
xlabel('Time: seconds')
title('Step response: discrete (*), &continuous')

Step response: discrete (*), & continuous

1.4

1.2

1 * * * * * * * * * * * * * * * * * * *

0.8

0.6
*

0.4

0.2

0*
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Time: seconds

8-187
trsp, dtrsp, sdtrsp

Now we look at the effect of a nonzero initial condition in the continuous

system. Note how examining the system at only the sample points will
underestimate the amplitude of the overshoot.
y2 = sdtrsp(M,C,vpck(1,0),T,1,0,[1;0]);
vplot(y1,'--',y2,'-')
axis([0,1,0,1.5])
xlabel('Time: seconds')
title('Step response: non zero initial condition')
Step response: non zero initial condition

1.4

1.2

0.8

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Time: seconds

Finally, we will examine the effect of a sinusoidal disturbance at the

continuous plant output. This controller has not been designed to reject such a
disturbance and the system does not contain anti-aliasing filters. Simulating
the effect of anti-aliasing filters is easily accomplished by including them in the
continuous interconnection structure.
M2 = mmult([0,1,1;1,0,0;0,1,1],daug(1,1,P));
dist = sin_tr(41,0.1,0.001,1);
[dat,datptr,t] = vunpck(dist);
input = abv(vpck(ones(t),t),dist);
[y3,meas,act] = sdtrsp(M2,C,input,T);
vplot(y3,'-',input,'--')
xlabel('Time: seconds')
title('Step response: disturbance (dashed) & output (solid)')

8-188
 trsp, dtrsp, sdtrsp

Step response: disturbance (dashed) & output (solid)

1.4

1.2

0.8

0.6

0.4

0.2

-0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Time: seconds

Algorithm trsp first calculates an integration time (or uses the specified integration time)
to determine the sample time at which to discretize the continuous-time
system. The integration time is taken to be the inverse of 10 times the fastest
mode of the input system. The input vector is interpolated at each sample time
via a zero-order hold, and then a sample-hold of the input continous system is
performed. Finally the time response of the system is performed via a for loop
at each integration time step. dtrsp is provided a discrete time system and a
sample time. dtrsp first interpolates the input vector via a zero-order hold and
then determines the time response via a for loop at each sample time.

Caution Systems with fast dynamics lead to very small integration times.
This is both time consuming and requires a significant amount of storage. We
recommend you residualize the fastest modes of the system, which does not
affect the time response. This can be done with the µ-Tools command sresid.

See Also cos_tr, siggen, sin_tr, step_tr, sysbal, vdcmate, vinterp

8-189
tustin

Purpose 8tustin
Create a discrete-time version of a continuous-time SYSTEM matrix using a
bilinear or prewarped tustin transformation

Syntax dsys = tustin(csys,T,prewarpf)

Description The packed continuous SYSTEM matrix, csys, is converted into a discrete-time
SYSTEM matrix, dsys, using a bilinear transformation with prewarping. The
argument T is the sample time, in seconds. prewarpf is the prewarp frequency
in rads/sec. prewarpf is an optional argument, and if omitted, or equal to zero,
a bilinear transformation is performed instead.
The resulting discrete system, dsys, has the same transfer function at the
continuous system, csys, at the prewarp frequency. Choosing a prewarp
frequency close to the crossover frequency is often appropriate for a control
system. Choosing a prewarp frequency too close to the Nyquist frequency (1/2T)
can result in severe distortion at the lower frequencies. In the extreme, if
prewarp is greater than or equal to π/T, the discrete system can be unstable.
Note that the transfer function is preserved at zero frequency with a bilinear
transformation, hence having the input variable prewarpf equal to zero to
indicate a bilinear transformation is therefore consistent.

Examples Create a second-order system with a resonance at 1 rad/sec.

a = [-.1,1;-1,-0.05];
b = [1;1]; c = [-0.5,0]
sys = pck(a,b,c);
minfo(sys)
system:2states1 outpus1 inputs
omega = logspace(-2,2,100);
omega2 =[ [0.05:0.1:1.5] [1.6:.5:20] [0.9:0.01:1.1] ];
omega = sort([omega omega2]);
sys_g = frsp(sys,omega);

8-190
 tustin

Choose a sample frequency of 20 rads/sec and discretize the system with a

bilinear transformation.
T = 2*pi/20;
dsys = tustin(sys,T);
dsys_g = frsp(dsys,omega,T);
vplot('bode',sys_g,dsys_g,'-.');
title('Continuous system (solid), bilinear equivalent
(dot-dash)')

Continuous system (solid), bilinear equivalent (dot-dash)

10 1
Log Magnitude

10 0

10 -1

10 -2

10 -3
10 -2 10 -1 10 0 10 1 10 2
Frequency (radians/sec)

4
Phase (radians)

-2

-4
10 -2 10 -1 10 0 10 1 10 2
Frequency (radians/sec)

The bilinear approximation is accurate up to about 2 rads/sec. This example

shows the effect of choosing a higher prewarping frequency, specifically 5 rads/
sec.
prewarpf = 5;
dsys2 = tustin(sys,T,prewarpf);
dsys2g = frsp(dsys2,omega,T);
vplot('bode',sys_g,dsys2_g,'-.');
title('Continuous system (solid), tustin equivalent (dot-dash)')

8-191
tustin

Continuous system (solid), Tustin equivalent (dot-dash)

10 1
Log Magnitude

10 0

10 -1

10 -2

10 -3
10 -2 10 -1 10 0 10 1 10 2
Frequency (radians/sec)

4
Phase (radians)

-2

-4
10 -2 10 -1 10 0 10 1 10 2
Frequency (radians/sec)

Note the distortion in the frequency of the lightly damped peak. At 5 rads/sec
both the continuous and discrete systems have the same transfer function.
sys_5 = vunpck(frsp(sys,5));
dsys2_5 = vunpck(frsp(dsys2,5,T));
err = abs(dsys2_5 - sys_5);
fprintf('error at %g rad/sec is : %g ',prewarpf,err);
error at 5 rad/sec is : 1.155158e-17

This example highlights the distortion possible. If the frequency of the

resonance had been critical to the design, a prewarp frequency of 1 rad/sec
would have been more appropriate.
As an alternative, you can generate a filter/controller design using a warped
frequency scale in the continuous domain. Then the transformation to the
discrete domain would result in the correct transfer function at the frequencies
of interest.

Algorithm The prewarped tustin transformation is based on the equation:

8-192
 tustin

csys =
A B , using the prewarped Tustin transform
C D
A disc B disc
dsys = ,
C disc D disc

where

prewrapf
α = --------------------------------
tan ( T prewarpf/2 )
*
–1
A disc = ( I + --α1- A ) * ( I – ---α1 A )
1 –1
---
B disc = ( --α2- ) 2 * ( I – --α1- A ) *B
1 –1
---
C disc = ( --α2- ) 2 * C * ( I – ---α1 A )
–1
D disc = D + ---α1 * C * ( I – ---α1 A ) * B

Reference Oppenheim, A.V., and R.W. Schafer, Digital Signal Processing, Prentice Hall,
New Jersey, 1975.

See Also dtrsp, frsp, samhld, tustin

8-193
unum, xnum, ynum

Purpose 8unum, xnum, ynum

unum returns the SYSTEM matrix input dimension

xnum returns the SYSTEM matrix state dimension

ynum returns the SYSTEM matrix output dimension

Syntax numinputs = unum(sys)

numstates = xnum(sys)
numoutputs = ynum(mat)

Description unum returns the input (column) dimension of SYSTEM, CONSTANT, and
VARYING matrices.
xnum returns the state dimension of SYSTEM matrices.

ynum returns the output (row) dimension of SYSTEM, CONSTANT, and

VARYING matrices.

Examples sys = sysrand(3,4,5);

xnum(sys)
ans =
3
unum(sys)
ans =
5
ynum(sys)
ans =
4
mat = crand(17,9);
xnum(mat)
ans =
0
unum(mat)
ans =
9
ynum(mat)
ans =
17

See Also find, minfo, xtract, xtracti

8-194
 vabs, vimag, vreal, vfloor, vceil

Purpose 8vabs, vimag, vreal, vfloor, vceil

Perform element-by-element operations on CONSTANT and VARYING
matrices

Syntax matout = vabs(matin)

matout = vimag(matin)
matout = vreal(matin)
matout = vfloor(matin)
matout = vceil(matin)

Description This set of commands allows standard element-by-element operations on

CONSTANT/VARYING matrices. These commands are identical to abs, imag,
real, floor, and ceil but also work on VARYING matrices.

vabs element-by-element absolute value of a VARYING matrix

vimag element-by-element imaginary part of a VARYING matrix
vreal element-by-element real part of a VARYING matrix
vfloor element-by-element floor of a VARYING matrix
vceil element-by-element ceiling of a VARYING matrix

A general element-by-element command, vebe, allows all standard arithmetic

MATLAB element commands that have only one input argument.

Examples Construct a complex VARYING matrix and find the magnitude of the entries
and the real parts.
see(matin)
1 row 2 columns
iv = .2
0.2190 - 0.4379i0.6789 - 1.3577i
iv = .7
0.0470 - 0.0941i0.6793 - 1.3586i

8-195
vabs, vimag, vreal, vfloor, vceil

see(vabs(matin))
1 row 2 columns
iv = .2
0iv = .7
0.10521.5190
see(vreal(matin))
1 row2 columns
iv = .2
0.21900.6789
iv = .7
0.04700.6793

Use pkvnorm to find the maximum element magnitude of the VARYING

matrix.
pkvnorm(matin,inf)
1 row 2 columns
ans =
1.5190
This agrees with the maximum magnitude of matin(1,1) associated with the
second independent variable.

See Also vdet, vdiag, vebe, veval

8-196
 vdet, vdiag, vexpm, vrcond

Purpose 8vdet, vdiag, vexpm, vrcond

Calculate determinant, diagonal, matrix exponential and estimate of condition
number of a CONSTANT or VARYING matrix

Syntax out = vdet(matin)

out = vdiag(matin)
out = vexpm(matin)
out = vrcond(matin)

Description These commands operate on square, CONSTANT and VARYING matrices and
they are identical to the MATLAB commands det, diag, exp, and rcond on
CONSTANT matrices.
vdet of a square, VARYING matrix, returns matout, which is a VARYING
1 × 1 matrix, containing the value of the determinant of matin at each
independent variable value.
vdiag of a square, VARYING matrix, returns matout, which is a VARYING
matrix of size min(size(matin)) × 1, containing the diagonal elements of
matin at each independent variable.

vexpm of a square, VARYING matrix, returns matout, which is a VARYING

matrix of the same size as matin, containing the matrix exponential of matin.
The MATLAB command expm, which is called, uses a Pade expansion after
scaling matin to calculate the exponential.
vrcond of a square, VARYING matrix, returns matout, which is an estimate of
the condition number of a matrix.

Examples vdet and vrcond work similarly to their MATLAB counterparts, det and
rcond, but on square VARYING matrices as shown below.

see(matin)
2 rows 2 columns
iv = 2.3
0.04750.3282
0.73610.6326
iv = 5.6
0.75640.3653
0.99100.2470
matout = vdet(matin);

8-197
vdet, vdiag, vexpm, vrcond

see(matout)
1 row 1 column
iv = 2.3
-0.2116
iv = 5.6
-0.1752
see(vrcond(matin))
1 row 1 column
iv = 2.3
0.1907
iv = 5.6
0.0848

See Also det, diag, expm, rcond

8-198
 vebe

Purpose 8vebe
Perform element-by-element operations on CONSTANT and VARYING
matrices

Syntax out = vebe('oper',matin)

Description The vebe function allows any single argument MATLAB element-by-element
arithmetic command to operate on a VARYING matrix. The first input
argument, oper, is the character string defining the MATLAB
element-by-element command and matin is the VARYING matrix on which the
command is applied. vebe calls the MATLAB eval command to execute the
string command. Some standard MATLAB comands compatible with vebe are
sin, abs, real, imag, and gamma.

Examples In this example of vebe, the real part of a matrix is found along with gamma of
each matrix element.
see(matin)
3 rows 3 columns
iv = 4.2

1.0000 + 2.0000i 1.0000 + 2.0000i 1.0000 + 2.0000i

2.0000 + 4.0000i 2.0000 + 4.0000i 2.0000 + 4.0000i
3.0000 + 6.0000i 3.0000 + 6.0000i 3.0000 + 6.0000i

iv = 11.01

4.0000 + 8.0000i 4.0000 + 8.0000i 4.0000 + 8.0000i

5.0000 + 10.0000i 5.0000 + 10.0000i 5.0000 + 10.0000i
6.0000 + 12.0000i 6.0000 + 12.0000i 6.0000 + 12.0000i

matout = vebe('real',matin);
see(matout)
3 rows 3 columns

8-199
vebe

iv = 4.2
1 1 1
2 2 2
3 3 3

iv = 11.01
4 4 4
5 5 5
6 6 6
see (vebe ('gamma ', matin))
3 rows 3 columns

iv = 4.2
1 1 1
1 1 1
2 2 2

iv = 11.01
6 6 6
24 24 24
120 120 120

See Also eval, vabs, vceil, veval, vfloor, vimag, vreal

8-200
 veig

Purpose 8veig
Calculate eigenvalues and eigenvectors of CONSTANT and VARYING
matrices

Syntax evals = veig(matin)

[evecs,evals] = veig(matin)
[evecs,evals] = veig(matin,'nonbalance')
[evecs,evals] = veig(mat1,mat2)

Description veig is identical to MATLAB’s command eig, but works on VARYING

matrices. veig solves for the eigenvalues (evals) and optionally the
eigenvectors (evecs) of the input matrix matin. veig works only on square
CONSTANT or VARYING matrices. Depending on the input arguments, the
following operations are performed:
evals = veig(matin) returns a VARYING vector evals containing the
eigenvalues of the VARYING matrix matin for each independent variable.
[evecs,evals] = veig(matin) returns the VARYING diagonal matrix evals
and a square VARYING matrix evecs whose columns are the corresponding
eigenvectors for each independent variable.
[evecs,evals] = veig(matin,'nobalance') is the same as the above
command without performing a preliminary balancing step. Balancing a
matrix, which has very small entries due to round off error, can lead to
incorrect eigenvectors.
[evecs,evals] = veig(mat1,mat2) returns a VARYING diagonal matrix of
generalized eigenvalues evals and the corresponding values of generalized
right eigenvectors evecs associated with each independent variable. mat1 and
mat2 can be CONSTANT or VARYING matrices. If either is CONSTANT, then
that same matrix is used in the generalized eigenvalue solution for each
independent variable. If they are both VARYING matrices, then they must
have the same independent variables.

8-201
veig

Examples Create a 2 × 2 random VARYING matrix and find its eigenvalues.

see(matin)
2 rows 2 columns

iv = 0.1
0.9304 0.5269
0.8462 0.0920
iv = 0.4
0.6539 0.7012
0.4160 0.9103
evals = veig(matin);
see(evals)
2 rows1 column

iv = 0.1
1.2996
-0.2772

iv = 0.4
0.2270
1.3372

Using the same matrix and creating another two by two VARYING matrix,
solve the generalized eigenvalue problem with these two matrices.
mat1 = matin;
mat2 = vpck([4*eye(2);3*eye(2)],[.1 .4]);
[evecs,evals] = veig(mat1,mat2);
see(evecs)
2 rows 2 columns

iv = 0.1
0.8190 -0.3999
0.5738 0.9166

iv = 0.4
0.8542 0.7162
-0.5200 0.6979
see(evals)

8-202
 veig

2 rows 2 columns

iv = 0.1
0.3249 0
0 -0.0693
iv = 0.4
0.0757 0
0 0.4457

Algorithm veig calls the MATLAB eig command.

See Also eig, indvcmp, svd, vsvd, vpoly, vroots

8-203
veval

Purpose 8veval
Evaluate general functions of CONSTANT, SYSTEM, and VARYING matrices

Syntax [out1,out2,out3,...]= veval('oper',in1,in2,in3,...)

Description The veval function evaluates the command oper on the input matrices. veval
works like feval but on collections of VARYING, CONSTANT, and SYSTEM
matrices. 'oper' is a character string with the name of a MATLAB function
(user written, or MATLAB supplied). The function is applied to each input
argument at the independent variable’s values. Any CONSTANT or SYSTEM
matrix is held at its value while the sweep through the independent variable is
performed. veval is currently limited to 10 output arguments, and 13 input
arguments. These are both easily changeable. veval can be used to generate
and manipulate VARYING, SYSTEM matrices or VARYING matrices whose
elements are themselves VARYING matrices. Arbitrary nesting of VARYING
matrices is possible.
The veval function is very useful for rapid prototyping of customized
commands employing VARYING matrices.

Examples To show the flexibility of veval, two random SYSTEM matrices are
constructed. The poles of each SYSTEM are determined with the spoles
command.
sys1 = sysrand(2,1,1);
sys2 = sysrand(2,1,1);
spoles(sysl)
ans =
0.1577
0.7405
spoles(sys2)
ans =
0.6273
-0.5661

8-204
 veval

These two SYSTEM matrices are combined to form a VARYING matrix, vsys.
The veval command is used to find poles of the VARYING matrix, which
consists of the two SYSTEM matrices. A SYSTEM matrix is associated with
each independent variable.
vsys = vpck([sysl;sys2],[1 2]);
vsyspoles = veval('spoles',vsys);
see(vsyspoles)
2 rows1 column

iv = 1
0.1577
0.7405

iv = 2
0.6273
-0.5661

See Also eval, feval, vebe

8-205
vfft, vifft, vspect

Purpose 8vfft, vifft, vspect

Calculate FFTs, inverse FFTs, and perform spectral analysis on VARYING
matrices

Syntax yfreq = vfft(ytime,n)

ytime = vifft(yfreq)
P = vspect(x,m,noverlap,'window')
P = vspect(x,y,m,noverlap,'window')

Description vfft implements the MATLAB fft command on VARYING matrix structures.
A one-dimensional FFT of length n is performed on each element of the
VARYING matrix ytime. It is assumed that the independent variable is in
units of seconds. The independent variables are regularly spaced — only the
first interval is used to determine the frequency scale. yfreq is returned with
the independent variable, frequency, in radians/second.
vifft performs the inverse FFT. This is done with the MATLAB command
ifft(yfreq) for each element of the VARYING matrix.

vspect is the VARYING matrix structure equivalent of the Signal Processing

Toolbox command, spectrum. For algorithmic details, see the spectrum
command. Note that vspect gives you the option of specifying a window for the
data. For example, using the string hamming as the fifth argument generates a
window with the command window = hamming(n);. hamming is an M-file in the
Signal Processing Toolbox. You can use custom windows by specifying the
name as the window argument.
In the case of the spectrum of a single signal, the command
P = vspect(x,m,noverlap,'window');

will return a VARYING matrix, P, containing the power spectral density of x.

Note that x, and therefore P, can be a matrix of signals. In the case of the
spectrum, and cross-spectrum, of two signals, the command
P = vspect(x,y,m,noverlap,'window');
will return a VARYING matrix, P, with the following five columns.
[Pxx Pyy Pxy Txy Cxy]

8-206
 vfft, vifft, vspect

These are given below.

Pxx Power spectral density of x
Pyy Power spectral density of y
Pxy Cross spectral density
Txy Complex transfer function between x and y
Cxy Coherence function between x and y

The signal x, must be scalar (i.e., a one row, one column, VARYING matrix). y
can be a vector signal. The row dimension of p is the same as that of y. vspect
can do single-input, multiple-output (SIMO) identification. This is illustrated
in the following example. Refer also to the example in the Tutorial chapter.
vfft, vifft, and vspect have not been optimized for speed. The appropriate
row and column data is extracted from the VARYING matrices with the µ-Tools
commands, sel and xtract. sbs and abv are used to create the final output.

Examples A single-input two-output system is generated as an identification example.

This example is only a simple illustration of some of the frequency domain
techniques available.
a1 = [-.1,1;-1,-0.05];
b1 = [1;1]; c1 = [-0.5,0];
sys1 = pck(a1,b1,c1);
a2 = [-.1,0.5;-0.5,-0.1];
b2 = [1;1]; c2 = [-0.5,0];
sys2 = pck(a2,b2,c2);
sys = abv(sys1,sys2);
minfo(sys)
system:4 states2 outputs1 inputs

Now a random input signal is created.

t = [0:0.05:102.35];
u = siggen('0.5-rand(size(t))',t);
minfo(u)
varying:2048 pts1 rows1 cols

8-207
vfft, vifft, vspect

The signal u is the input to the system. siggen is used to generate some random
noise on the output signal, y.
y =
madd(trsp(sys,u),siggen('[0.01*rand(size(t));0.025*rand(size(t))
]',t));
integration step size: 0.05
vplot(y)
title('vspect example: output waveform with noise')
xlabel('time: seconds')

vspect example: output waveform with noise

0.5

0.4

0.3

0.2

0.1

-0.1

-0.2

-0.3
0 20 40 60 80 100 120

time: seconds

The vspect command specifies a 1024 point window, with 512 points of
overlap. A Hamming window is applied to the data.
P = vspect(u,y,1024,512,'hamming');
3 hamming windows in averaging calculation

Column 4 in P contains the complex transfer function estimate. Its magnitude

is compared to the actual system transfer function.
omega = logspace(-2,2,100);

8-208
 vfft, vifft, vspect

omega2 =[ [0.05:0.1:1.5] [1.6:.5:20] [0.9:0.01:1.1] ];

omega = sort([omega omega2]);
sys_g = frsp(sys,omega);
vplot('liv,lm',sel(P,1:2,4),sys_g);
Warning: Data includes a number that is negative or zero.
The LOG of this results in NaN or Infinity and is not shown on
plot.
title('vspect example: transfer function estimation ')
ylabel('magnitude')
xlabel('frequency: rad/sec')

vspect example: transfer function estimation

10 1

10 0
magnitude

10 -1

10 -2

10 -3
10 -2 10 -1 10 0 10 1 10 2

frequency: rad/sec

Algorithm vfft, vifft, and vspectrum call the MATLAB commands fft and ifft.

Reference Ljung, L., System Identification: Theory for the User, Prentice Hall, New
Jersey, 1987.
Oppenheim, A.V., and R.W. Schafer, Digital Signal Processing, Prentice Hall,
New Jersey, 1975.

See Also fft, ifft, spectrum

8-209
vfind

Purpose 8vfind
Unary find function across independent variable

Syntax [iv_value,iv_index] = vfind(condition,mat)

Description vfind is a unary find function that searches across independent variable
values. The condition to be tested can be any valid MATLAB conditional
statement, using the string mat to identify the matrix, and iv as the
independent variable’s value. Both the values and indices of the applicable
independent variables are returned.

Examples Suppose that matin is a VARYING matrix. In order to find those entries for
which the product of the norm of the matrix, and the independent variable is
greater than 2, use vfind as follows.
[iv_value,iv_index] = vfind('iv*norm(mat)>2',matin);
matpropv = xtract(matin,iv_value); % extract by value
matpropi = xtracti(matin,iv_index); % extract by index
pkvnorm(msub(matpropv,matpropi)) % compare - both are the same

See Also find, xtract, xtracti

8-210
 vinterp, vdcmate

Purpose 8vinterp, vdcmate

Interpolate or decimate VARYING matrices

Syntax vout = vinterp(vin,stepsize,finaliv,order)

vout = vinterp(vin,varymat,order)
vout = vdcmate(vin,spacing)

Description In the first form, vinterp produces a regularly spaced interpolated version of
the input VARYING matrix. The input arguments are
stepsize independent variable stepsize
finaliv end value for independent variable (Optional: the default is
the final independent variable in the input)
order type of interpolation (optional, default = 0)
0 zero-order hold
1 linear interpolation

The end value for the independent variable may or may not be in the actual
output. This is consistent with the usual MATLAB treatment of regularly
spaced vectors. For example, consider
iv = [1:2:6];
disp(iv)
1 3 5
Note that the value of 6 does not appear in the vector.
In the second form, vinterp produces a VARYING matrix vout that is an
interpolated version of vin. The independent variables of vout are the same as
the independent variables of varymat. The input arguments are
varymat VARYING matrix with desired independent variables
order type of interpolation (optional, default = 0)
0 zero-order hold
1 linear interpolation

vdcmate decimates the VARYING matrix vin, whose independent variable

must be linearly spaced and in ascending order. If spacing has a value of n,
then the output contains only the matrices corresponding to every nth
independent variable of the input. If no spacing is specified, the default is 10.

8-211
vinterp, vdcmate

Examples siggen creates a sinewave. This is effectively sampled by vdcmate and then
interpolated by vinterp. Note that the default interpolation is a zero-order
hold, giving a stair-step output, yi. If a linearly interpolated output were
specified, it would look identical to yd since the MATLAB plot command
displays a linear interpolation.
timebase = [0:0.005:20];
y = siggen('sin(2*pi*t)',timebase);
minfo(y)
varying:4001 pts1 rows1 cols
yd = vdcmate(y,210);
minfo(yd)
varying:20 pts1 rows1 cols
yi = vinterp(yd,0.005,20,0);
minfo(yi)
varying:4001 pts1 rows1 cols
axis([0,20,-1.5,1.5])
vplot(y,yd,yi)
title('vdcmate/vinterp example: undersampled sine wave')
xlabel('time: seconds')

vdcmate/vinterp example: undersampled sine wave

1.5

0.5

-0.5

-1

-1.5
0 2 4 6 8 10 12 14 16 18 20

time: seconds

See Also dtrsp, sort, sortiv, tackon, trsp

8-212
 vldiv, vpinv, vrdiv

Purpose 8vldiv, vpinv, vrdiv

Determine left division, pseudo-inverse and right division of a CONSTANT or
VARYING matrix

Syntax out = vldiv(mat1,mat2)

out = vpinv(mat,tol)
out = vrdiv(mat1,mat2)

Description These commands operate on VARYING matrices and they are identical to the
MATLAB commands \, pinv, and / on CONSTANT matrices.
vldiv of a VARYING matrix returns out, which is a VARYING matrix,
containing the value of the left division (mat1(i)mat2(i)) at each independent
variable value. vldiv is identical to the MATLAB command \ for CONSTANT
matrices.
vpinv of a VARYING matrix returns out, which is the pseudo-inverse of mat at
each independent variable. out is of the same dimension as vcjt(out), and
satisfies mat = mmult(mat,out,mat). tol is used within the svd routine to
determine zero singular values. The default value of tol is 1e–12. vpinv is
identical to the MATLAB command pinv for CONSTANT matrices.
vrdiv of a VARYING matrix returns out, which is a VARYING matrix,
containing the value of the right division (mat1(i)/mat2(i)) at each
independent variable value. vrdiv is identical to the MATLAB command / for
CONSTANT matrices.

See Also \, /, pinv, vinv, vsvd

8-213
vpck, vunpck, var2con

Purpose 8vpck, vunpck, var2con

Pack and unpack a VARYING matrix and convert from a VARYING matrix to
a CONSTANT

Syntax matout = vpck(matin,indv)

[varydata,rowpoint,indv,err] = vunpck(mat)
[matout,ivval] = var2con(mat,desiv)

Description The data structure for a VARYING matrix consists of the sampled matrix
values stacked one upon each other, and the particular independent variable
values. vpck places the stacked data from the input variable, matin, and the
vector, indv, which represents the independent variable values, into a new
matrix, matout, with the correct structure and data structure of a VARYING
matrix.
The command vunpck performs the inverse operation; unpacking a VARYING
matrix into stacked data varydata, row pointers rowpoint, a vector of
independent variables indv, and an error flag err. The value of rowpoint(i)
points to the row of data that corresponds to the first row of the ith value of
matin. indv is a column vector with the independent variable values. The error
flag is normally 0 but it is set to 1 if the input matrix is a SYSTEM.
var2con converts VARYING matrices to CONSTANT matrices. If there is one
input argument, mat, and it is a VARYING matrix, then the output matout is
the CONSTANT matrix in mat associated with the independent variable’s first
value. The optional second output argument is this independent variable’s
value. If two input arguments are used, then the first is a VARYING matrix,
and the second is a desired independent variable’s value. The command finds
the matrix in mat whose independent variable’s value is closest to desiv, and
returns this matrix as a CONSTANT matrix.

8-214
 vpck, vunpck, var2con

Examples Construct a VARYING matrix from a CONSTANT matrix and a vector of

independent variables.
disp(matin)
1 1 1
2 2 2
3 3 3
4 4 4
5 5 5
6 6 6
time = [.1 2.3 5.6]';
disp(time)
1.0000e-01
2.3000e+00
5.6000e+00
matout = vpck(matin,time);
see(matout)
2 rows 3 columns

iv = 0.1
1 1 1
2 2 2

iv = 2.3
3 3 3
4 4 4

iv = 5.6
5 5 5
6 6 6

See Also pck, unpck, xtract, xtracti

8-215
vplot

Purpose 8vplot
Plot multiple VARYING matrices on the same graph

Syntax vplot('plot_type',vmat1,vmat2,...)
vplot('plot_type',vmat1,'linetype1',...)
vplot('bode_l',top_axis_limits,bottom_axis_limits,vmat1,vmat2,...)

Description The vplot command calls the standard MATLAB plot command for plotting.
The optional plot_type argument specifies the type of graph, and selects
between the various logarithmic or linear graph types. The plot_type
specification choices are
iv,d matrix (decimal) vs. independent variable
iv,m magnitude vs. independent variable
iv,lm log(magnitude) vs. independent variable
iv,p phase vs. independent variable
liv,d matrix vs. log(independent variable)
liv,m magnitude vs. log(independent variable)
liv,lm log(magnitude) vs. log(independent variable)
liv,p phase vs. log(independent variable)
ri real vs. imaginary (parametrized by independent variable)
nyq real vs. imaginary (parametrized by independent variable)
nic Nichols chart
bode Bode magnitude and phase plots
bode_g Bode magnitude and phase plots with grids
bode_l Bode magnitude and phase plots with axis limits
bode_gl Bode magnitude and phase plots with grids and axis limits

If no plot_type specification is given the default is 'iv,d'.

The bode_l and bode_gl plot_type specifications require that the second and
third arguments are the desired axis limits for the top and bottom plots. These
are simply the 1× 4 vectors to be used as arguments for the axis command.

8-216
 vplot

The remaining arguments of vplot take the same form as the MATLAB plot
command. Line types (for example,'+', 'g-.', or '*r') can be optionally
specified after any VARYING matrix argument.
There is a subtle distinction between CONSTANT and VARYING matrices
with only one independent variable. A CONSTANT is treated as such across all
independent variables, and consequently shows up as a line on any graph with
the independent variable as an axis. A VARYING matrix with only one
independent variable will always show up as a point. You may 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 for each over different frequency ranges.
a1 = [-1,1;-1,-0.5];
b1 = [0;2]; c1 = [1,0]; d1 = 0;
sys1 = pck(a1,b1,c1,d1);
minfo(sys1)
system:2 states1 outputs1 inputs
a2 = [-.1,1;-1,-0.05];
b2 = [1;1]; c2 = [-0.5,0]; d2 = 0.1;
sys2 = pck(a2,b2,c2,d2);
minfo(sys2)
system:2 states1 outputs1 inputs
omega = logspace(-2,2,100);
sys1_g = frsp(sys1,omega);
omega2 = [ [0.05:0.1:1.5] [1.6:.5:20] [0.9:0.01:1.1] ];
omega2 = sort(omega2);
sys2_g2 = frsp(sys2,omega2);

A VARYING matrix with a single independent variable is also created. Note

the distinction between this and the CONSTANT matrix in the subsequent
plots.
rspot = vpck(sqrt(2)-sqrt(2)*i,2);
minfo(rspot)
varying:1 pts1 rows1 cols

The following plot uses the 'liv,lm' plot_type specification. Note that the
CONSTANT matrix is seen over all values of the independent variable. This is
only true because it is displayed as a line type. If it were displayed as a point,

8-217
vplot

then one would see points only on each of the side axes. The single valued
VARYING matrix (rspot) is shown only at the appropriate independent
variable value.
vplot('liv,lm',sys1_g,'b-.',[1+i;0.5-0.707*i],'g--',...
rspot,'r*',sys2_g2);
xlabel('log independent variable')
ylabel('log magnitude')
title('plot_type specification: liv,lm')

axis specification: liv,lm

10 1

10 0

10 -1
log magnitude

10 -2

10 -3

10 -4
10 -2 10 -1 10 0 10 1 10 2

log independent variable

You can customize vplot to select the type of axis uses for log magnitude and
phase plots. The default is to plot the log magnitude on a base 10 scale and plot
phase in radians. It is a simple modification to select a dB scale and phase in
degrees. Documentation of the modification is provided in the M-file vplot. You
can copy the command vplot to a private directory (for example, matlab/
toolboxes/mu_cmds on UNIX systems) and make the appropriate
modifications.

8-218
 vplot

Several control design plot functions are also provided. These are bode, nic,
and nyq, for Bode, Nichols, and Nyquist, respectively. The following three plots
demonstrate each of these commands.
vplot('bode',sys1_g,'b',sys2_g2,'g+');
title('plot_type specification: bode')

axis specification: bode

10 1
++++++++
++ +++
10 0 ++ ++
+ + +
Log Magnitude

+ +
+ + + +
+ + + ++
+++++++++++++++++++++++++++++
10 -1

10 -2

10 -3

10 -4
10 -2 10 -1 10 0 10 1 10 2
Frequency (radians/sec)

4
++
+
+
+
++
Phase (radians)

2 +
+
++
+
++++
+ + + + + ++
++++++++++
++++++++++++++++++++
0

-2
+ + + + + + + + ++
+
-4
10 -2 10 -1 10 0 10 1 10 2
Frequency (radians/sec)

The log magnitude and phase axes are labeled automatically. You can change
these labels. Documentation for doing this is in the Help facility for vplot.
vplot('nic',sys1_g,'b-.',[1+i;0.5-0.707*i],'go',rspot,...
rspot,'r*',sys2_g2);
title('plot_type specification: nic')
xlabel('phase (degrees)')
ylabel('log magnitude (dB)')
title('plot_type specification: nic (Nichols Chart)')

8-219
vplot

axis specification: nic (Nichols Chart)

20

10
*
o
0 o

-10
log magnitude (dB)

-20

-30

-40

-50

-60

-70

-80
-350 -300 -250 -200 -150 -100 -50 0

phase (degrees)

The default axis scale selection for the Nichols plot is dB versus phase in
degrees. This corresponds to the usual choice for this plot and can be different
from the axis scale selection for bode, liv, lm, liv, p, etc. Again you can change
this if required.
vplot('nyq',sys1_g,'b-.',[1+i;0.5-0.707*i],'go',rspot,...
xlabel('nyquist diagram (real)')
vplot('liv,lm',sys1_g,'b-.',[1+i;0.5-0.707*i],'g--',...
rspot,'r*',sys2_g2);
ylabel('imaginary')
title('plot_type specification: nyq')

8-220
 vplot

axis specification: nyq

5

2
imaginary

1 o

o
-1
*

-2
-5 -4 -3 -2 -1 0 1 2

nyquist diagram (real)

See Also plot

8-221
vpoly, vroots

Purpose 8vpoly, vroots

Find coefficients and roots of a characteristic polynomial from a CONSTANT
or VARYING matrix

Syntax matout = vpoly(matin)

vecout = vpoly(vecin)
vecout = vroots(vecin)

Description vpoly forms an n + 1 element VARYING row vector whose elements form the
coefficients of the characteristic polynomial, det(sI – matin(i)), if matin is an
n × n VARYING matrix. The coefficients are ordered in descending powers of s.
If the input is a column vector vecin containing the roots of a polynomial,
vpoly(vecin) returns a VARYING row vector whose elements are the
coefficients of the corresponding characteristic polynomial.
vroots returns as a VARYING column vector vecout whose elements are the
roots of the polynomial at each independent variable, if vecin is a VARYING
row vector containing the coefficients of a polynomial. vpoly and vroots are
identical to the MATLAB poly and roots commands, but also work on
VARYING matrices.

Examples Given a 3 × 3 VARYING matrix, find the characteristic polynomial and its
roots. Compare this to finding the eigenvalues of the input matrix via veig.
see(matin)
3 rows 3 columns
iv = 0.1
1 2 3
4 5 6
7 8 9
iv = 0.4
10 11 12
13 14 15
16 17 18
matout = vpoly(matin);
see(matout)
1 row 4 columns
iv = 0.1
l.0000e+00 -1.5000e+01 -1.8000e+01 -1.4483e-14
iv = 0.4
l.0000e+00 -4.2000e+01 -1.8000e+01 1.2818e-14

8-222
 vpoly, vroots

vecout = vroots(matout);
see(vecout)
3 rows 1 column
iv = 0.1
1.6117e+01
-1.1168e+00
-8.0463e-16

iv = 0.4
4.2424e+01
-4.2429e-01
7.1212e-16
evals = veig(matin);
see(evals)
3 rows 1 column

iv = 0.1
1.6117e+01
-1.1168e+00
-8.0463e-16
iv = 0.4
4.2424e+01
-4.2429e-01
7.1212e-16

Algorithm vpoly and vroots call the MATLAB poly and roots commands.

See Also eig, poly, roots, veig

8-223
vsvd, vrho, vschur

Purpose 8vsvd, vrho, vschur

Perform a singular value decomposition, spectral radius and Schur
decomposition of a CONSTANT or VARYING matrix

Syntax s = vsvd(matin)
[u,s,v] = vsvd(matin)
out = vrho(matin)
t = vschur(matin)
[u,t] = vschur(matin)

Description vsvd performs a singular value decomposition on a VARYING matrix. It is

identical to MATLAB’s svd routine, and will work on CONSTANT matrices as
well. If there is one output argument, the output is a VARYING matrix with
the singular values of matin at each point. If there are three output arguments,
[u,s,v], then u is a VARYING matrix with the left singular vectors, s is a
VARYING matrix with the singular values, and v is a VARYING matrix with
the right singular vectors.
vrho finds the spectral radius, max(abs(eig(xtracti(matin,i)))), at each
independent variable of a VARYING matrix.
vschur computes the Schur form of a VARYING matrix for each independent
variable. It is identical to the MATLAB schur command, but also works on
VARYING matrices. Given two output arguments, vschur returns two
VARYING matrices u and t. t corresponds to the Schur form matrix and u is a
VARYING unitary matrix such that
matin = mmult(u,t,vcjt(u))

Examples Construct a random VARYING matrix and find its singular values.
see(matin)
2 rows 2 columns

iv = 0.1
0.93040.5269
0.84620.0920
iv = 0.4
0.65390.7012
0.41600.9103
[u,s,v] = vsvd(matin);

8-224
 vsvd, vrho, vschur

see(u)
2 rows 2 columns
iv = 0.1
0.7884 0.6152
0.6152 -0.7884

iv = 0.4
0.6909 -0.7229
0.7229 0.6909
see(s)

iv = 0.1
1.3400 0.2689

iv = 0.4
1.3681 0.2219
see(v)
2 rows 2 columns

iv = 0.1
0.9359 -0.3522
0.3522 0.9359

iv = 0.4
0.5501 -0.8351
0.8351 0.5501

Algorithm vrho, vschur, and vsvd call the MATLAB commands svd, eig, and schur

See Also eig, hess, pkvnorm, mu, qz, schur, svd, veig, vnorm

8-225
vzoom

Purpose 8vzoom
Freeze plot axes by clicking mouse twice in plot window

Syntax vzoom('axis')

Description vzoom uses the MATLAB functions ginput and axis to freeze the axes by
clicking the mouse twice in the plot window that defines minimum and
maximum values for x and y. The clicking may be done in any order.
The axis argument specifies the type of graph, and can select between the
various logarithmic or linear graph types, just as in vplot. Unlike vplot, the
axis argument is not optional. The axis specification choices are

'iv,d', 'iv,m','iv,p' decimal, magnitude, or phase vs. independent

variable
'liv,d', 'liv,m','liv,p' decimal, magnitude, or phase vs.
'iv,lm' log(magnitude) vs. independent variable
'liv,lm' log(magnitude) vs. log(independent variable)
'ri', 'nyq' real vs. imaginary (parametrized by
independent variable)
'nic' Nichols chart
'ss' standard decimal
'll' log-log
'ls' semilogx
'sl' semilogy

Note that the axis specification is the same as for vplot, with the addition of
the last four possibilities. The function is not defined for 'bode'.

Examples The command vzoom(’liv,m’) would be equivalent to

[x,y] = ginput(2);
axis([log10([min(x) max(x)]) min(y) max(y)]);
An example of the use of vzoom is
tf = frsp(nd2sys([ 1 .1],[.1 1]),logspace(-2,2,100));
vplot('nic',tf); vzoom('nic'); vplot('nic',tf); axis;

See Also axis, ginput, vplot

8-226
 wsgui

Purpose 8wsgui
A graphical user interface for the MATLAB workspace

Syntax wsgui

Description wsgui is a graphical user interface (GUI) for the MATLAB workspace. It allows
you to view, delete, and save variables in the workspace, drag these variables
to other µ-Tools GUIs, dkitgui and simgui, drop boxes, and export variables
from the µ-Tools GUI interfaces to the MATLAB workspace.
The wsgui Workspace Manager window appears as shown on the following
page.

Each time Refresh Variables is pressed, the MATLAB command who is

executed, and minfo is run to determine the variable type and dimension. This
information is displayed in the main scrollable table. The date and time of the
last refresh are displayed below the button.

8-227
wsgui

The scrollable table can be moved up/down one page by pressing above/below
the slider. Pressing the arrows at the end of the slider moves the table one line.
A filter is used to make viewing of a reduced number of selections easy. The
Prefix, Suffix and matrix type filters are on the bottom of the scrollable
table. The matrix type filter is a pop-up menu to the right of Suffix. The
Custom filter, which is shown if * is pressed, allows you to create a more
complicated selection criteria. Press the pushbutton marked with an *; this
pushbutton is to the right of the pop-up menu, to switch to the custom filter. A
detailed description of wsgui is provided in the “Workspace User Interface Tool:
wsgui” section in Chapter 6.

Examples An example of using wsgui is shown in the “Workspace User Interface Tool:
wsgui” section in Chapter 6.

See Also clear, save, who

8-228
 wcperf

Purpose 8wcperf
Computes upper and lower bounds for the worst-case gain of a linear system
subjected to structured, bounded, LFT perturbations. Also computes
worst-case structured perturbation of a specified H∞

Syntax [deltawc,lowbnd,uppbnd] = wcperf(Mg,uncblk,alpha,npts);

Description The command wcperf is associated with the block diagram

where ∆ has block structure as defined by ∆, which is described via the matrix
uncblk. The worst-case performance curve, ƒ(α), is defined as

f ( α ) := max F U ( M, ∆ ) ∞
∆ ∈ $∆ , max w σ ( ∆ ( jω ) )ðα

Both lower and upper bounds for ƒ are returned as VARYING matrices in
lowbnd and uppbnd. Each VARYING matrix is guaranteed to have at least npts
values of the independent variable α, spread uniformly between 0 and the
stability limit.
The first output argument, delta_wc, is the “worst-case” perturbation from ∆
with norm equal to the value of alpha. delta_wc has the block-diagonal
structure associated with uncblk, and causes the LFT FU (M,∆wc) to have norm
equal to the value of lowbnd associated with the independent variable value α
= alpha.

See Also dypert, mu

8-229
xtract, xtracti

Purpose 8xtract, xtracti

Extract a specified portion of a VARYING matrix

Syntax [matout,err] = xtract(mat,iv_low,iv_high)

[matout,err] = xtract(mat,ivdes)
[matout,err] = xtracti(mat,indvindex)

Description xtract extracts a portion of a VARYING matrix. The independent variable

associated with a VARYING matrix monotonically increases (like frequency or
time). In the first form, xtract is called with three input arguments: a
VARYING matrix, a lower bound iv_low, and an upper bound iv_high. The
matrix values associated with any independent variables between iv_low and
iv_high are extracted, and returned as a VARYING matrix. In the second
form, xtract is called with two arguments. The second argument is a vector of
desired independent variable values. For each desired value, the matrix from
mat with the closest independent variable value (in absolute value) is
extracted.
xtracti extracts the value of the VARYING matrix at the specific indices
indvindex of the independent variable. Hence indvindex should be an array of
positive integers. The extracted matrix is returned as a VARYING matrix.

Examples Extract ranges of independent variable from a VARYING matrix.

see(mat)
2 rows2 columns

iv = 0.1
2.9703e+00 - 2.9703e-0li5.9406e+00 - 5.9406e-0li
3.9604e+00 - 3.9604e-0li7.9208e+00 - 7.9208e-0li

iv = 0.4
2.5862e+00 - 1.0345e+00i5.1724e+00 - 2.0690e+00i
3.4483e+00 - 1.3793e+00i6.8966e+00 - 2.7586e+00i

iv = 0.9
1.6575e+00 - 1.4917e+00i3.3149e+00 - 2.9834e+00i
2.2099e+00 - 1.9890e+00i4.4199e+00 - 3.9779e+00i

8-230
 xtract, xtracti

matl = xtract(mat,.3,.8);
see(matl)
2 rows2 columns

iv = 0.4
2.5862e+00 - 1.0345e+00i5.1724e+00 - 2.0690e+00i
3.4483e+00 - 1.3793e+00i6.8966e+00 - 2.7586e+00i

matl = xtracti(mat,2);
see(matl)
2 rows2 columns

iv = 0.4
2.5862e+00 - 1.0345e+00i5.1724e+00 - 2.0690e+00i
3.4483e+00 - 1.3793e+00i6.8966e+00 - 2.7586e+00i

See Also sel, var2con, vpck, vunpck, vfind

8-231
8-232
 Index

Symbols
8-109 dhfsyn 8-35
example 3-46
dhinfsyn
A example 3-43
abv 8-19 D-K iteration 5-3
example 2-12 commands 5-13
discussion 5-14
D-scalings 5-7
B full D-scaling 5-10
blknorm 8-22 HIMAT dkit example 7-100
HIMAT example 7-98 HIMAT example 7-85
holding Dfixed 5-9
holding Kfixed 5-10
C µ upper bound 5-6
cf2sys 8-125
problem formulation 5-3
HIMAT example 7-127
D-K iteration user interface 6-19
cjt 8-24
auto iteration 6-36
cmmusyn 8-27
D-K iteration 6-30
CONSTANT matrices 2-8
example 6-19
diagram 2-14
main window 6-20
control design
options 6-39
spinning satellite example 6-14
parameter window 6-26
cos_tr 8-28
setup window 6-20
crand 8-30
dkit 8-38
crandn 8-30
HIMAT example 7-100
csord 8-32
HIMAT setup file 7-102
dkitgui 6-2, 6-18, 8-44
unstable example 7-28
D
dragging and dropping icons 6-66
data structures 2-3
dkitgui 6-66
daug 8-19
example 6-68
Destabilizing perturbations
simgui 6-66
constructing 7-41
wsgui 6-66
dhfnorm 8-34
drawmag 8-46
dhfnormBody text>
D-scalings
example 3-15
automatic prefitting 7-109, 8-42

I-1
 Index

dtrsp 8-181 discrete-time 3-43

example 2-26 full control 3-31
dypert 8-48 full information 3-31
constructing perturbations 7-41 generalizations 3-41
history 3-22
loop shaping 3-52
E norm calculation 3-41
emargin 8-125 notation 3-23
output feedback 3-36
performance objectives 3-10
F problem definition 3-25
F–14 lateral-directional control design 7-132 sampled data 3-43
fft theory 3-20
example 2-27 H∞ control 3-1
fitmag 8-51 H∞ loop shaping
fitmaglp 8-51 HIMAT example 7-121
fitsys 8-55 reference signal 7-128
frsp 8-56 h2norm 8-67
example 2-20 example 3-14
h2syn 8-69
hankmr 8-172
G example 3-55
gain and phase margins 7-3
HIMAT design example 7-75
gap 8-61
hinffi 8-72
genmu 8-63
hinfnorm 8-67
genphase 8-51
example 3-14
getiv 8-65
hinfsyn 8-77
GUI
example 3-17
D-K iteration tool 6-19
HIMAT example 7-89
D-K workspace tool 6-2
shuttle example 7-55
LFT time simulation tool 6-41
hinfsyne 8-82
example 3-17

H
H∞ Control
commands 3-16
I
ifft
design example 3-18
example 2-27

I-2
 Index

indvcmp 8-85 mscl 8-105

msf 8-103
msfbatch 8-103
L msub 8-87
LFT time simulation user interface 6-40 example 2-11
continuous-time 6-61 mu
discrete-time 6-61 example 4-31
example 6-40 mixed perturbations 4-64
integration step size 6-55 mu 8-109
loading and saving 6-59 HIMAT example 7-93
main window 6-41 MIMO example 7-38
parameter window 6-55 muftbtch 8-116
plot titles 6-49 multivariable margins 7-7
plots and line type table 6-46 musynfit 8-116
printing 6-59 musynflp 8-116
sampled data 6-62 muunwrap 8-109
linear fractional transformations (LFTs) 4-3
loop shaping 7-86
N
ncfsyn 8-125
M HIMAT example 7-123
madd 8-87 nd2sys 8-128
example 2-11 negangle 8-130
magfit 8-51 nominal performance 7-83
massign 8-90 HIMAT example 7-87
mfilter 8-93 unstable example 7-15
minfo 8-95 norms 3-3
example 2-4 commands 3-14
minv 8-97 performance 3-5
mmult 8-99 nugap 8-61
example 2-11 HIMAT example 7-127
model reduction 3-55
balanced realizations 3-59
Hankel-norm 3-57 P
relative error 3-66 parametric uncertainty 4-5
weighted 3-65 pck 8-131
mprintf 8-102 performance objectives

I-3
 Index

F-14 example 7-134 example 3-47

unstable example 7-15 sdhfsyn 8-145
pkvnorm 8-134 example 3-47
hinfnorm 3-14 sdtrsp 8-181
plot 2-15 see 8-148
pss2sys 8-131 seeiv 8-148
example 2-5 example 2-7
seesys 8-150
sel 8-151
R example 2-9
randel 8-109 sensitivity 7-80
reordsys 8-151 sfrwtbal 8-162
representing uncertainty 4-3 sfrwtbld 8-162
ric_eig 8-136 example 3-65
ric_schr 8-138 siggen 8-154
Riccati operator 3-28 example 2-22
rifd 8-139 simgui 8-159
example 2-5 shuttle example 7-68
robust control design 5-2 sin_tr 8-28
robust performance 4-73, 7-83 sisorat 8-48
F-14 example 7-146 constructing perturbations 7-41
HIMAT example 7-85, 7-93 sncfbal 8-162
shuttle example 7-64 HIMAT example 7-126
unstable example 7-19 sortiv 8-65
robust stability 4-31, 4-71 example 2-8
space shuttle
time simulations example 7-67
S space shuttle FCS analysis 7-44
samhld 8-140 spectrum
sbs 8-19 example 2-30
example 2-12 spoles 8-161
sclin 8-105 example 2-5
scliv 8-142 srelbal 8-162
sclout 8-105 example 3-66
sdecomp 8-162 sresid 8-166
example 3-66 example 2-9
sdhfnorm 8-143 starp 8-168

I-4
 Index

example 7-8 SYSTEM matrices 2-3

statecc 8-170 diagram 2-14
state-space uncertainty 4-20 szeros 8-180
step_tr 8-28 example 2-5
strans 8-170
example 3-55
structured singular value (µ) T
block structure 4-56 tackon 8-65
bounds 4-54 transp 8-24
complex µ exercise 4-56 trsp 8-181
complex µ theory 4-50 example 2-22
D-scales 4-54 tustin 8-190
D-scalings 4-58 example 2-22
frequency domain 4-70 two tank example 7-151
generalized µ 4-81 closed-loop analysis 7-179
mixed µ exercise 4-64 closed-loop experiments 7-161
mixed µ theory 4-61 linearized model 7-166
mixed upper bound 4-61 nonlinear model 7-152
perturbations 4-59 open-loop experiments 7-158
real and complex blocks 4-61 problem formulation 7-165
real blocks 4-74 two tank experiment
robust performance 4-73 uncertainty models 7-168
robust stability 4-71
theory 4-49
strunc 8-166 U
example 2-9 uncertainty
sys2pss 8-131 aero coefficients 7-53
example 2-5 shuttle example 7-45
sysbal 8-172 uncertainty model
example 2-9 F-14 example 7-135
sysic 8-176 HIMAT 7-78
description 2-34 unstable example 7-15
F-14 example 7-142 unpck 8-131

HIMAT example 2-38, 7-83 example 2-5

unstable example 7-15 unstable example 7-15
sysrand 8-30 unum 8-194

example 2-5 unwrapp 8-109

I-5
 Index

V vpoly 8-222
vabs 8-195 vrcond 8-197
var2con 8-214 vrdiv 8-213
example 2-10 vreal 8-195
VARYING matrices 2-6 vrho 8-224
diagram 2-14 vroots 8-222
varyrand 8-30 vschur 8-224
vceil 8-195 vspect 8-206
vcjt 8-24 example 2-30
vdcmate 8-211 vsvd 8-224
example 2-27 vtp 8-24
vdet 8-197 vunpck 8-214
vdiag 8-197 example 2-7
vebe 8-199 vzoom 8-226
example 2-18
veig 8-201
veval 8-204 W
example 2-18 wcperf 8-229
vexpm 8-197 workspace user interface 6-3
vfft 8-206 clearing and saving 6-9
example 2-25, 2-27 customize 6-7
vfind 8-210 workspace user interface filter 6-4
vfloor 8-195 font and lines 6-10
vifft 8-206 worst-case performance
example 2-27 unstable example 7-146
vimag 8-195 worst-case perturbations
vinterp 8-211 shuttle example 7-65
example 2-25 wsgui 8-227
vinv 8-97
vldiv 8-213
vnorm 8-134 X
example 3-15 xnum 8-194

vpck 8-214 xtract 8-230

example 2-7 example 2-9

vpinv 8-213 xtracti 8-230

vplot 2-15, 8-216 example 2-10

example 2-15

I-6
 Index

Y
ynum 8-194

Z
zp2sys 8-128

I-7
 Index

I-8