
 

 

\documentclass{article}
%\usepackage{here}
%\usepackage{paper}
%\usepackage{epsf}

% ----- PREAMBLE -----

\pagenumbering{arabic}

\newcommand{\limh}{\mbox{{\rm h-lim}}}
\newcommand{\lime}{\mbox{{\rm e-lim}}}
\newcommand{\pa}        {\parallel}
\newcommand{\ep}        {\varepsilon}
% --------------        {\epsilon}
\newcommand{\da}        {\delta}
\newcommand{\spp}       {\vspace{3ex}}
\newcommand{\text}[1]{\mbox{{\rm #1}}}

\newcommand{\grad}{\nabla}
\newcommand{\directsum}{\oplus}

\newcommand{\dom}{{\rm dom}}
\newcommand{\intt}{{\rm int}}

\newcommand{\e}{e^{i \theta}}
\newcommand{\cross}{{\times}}
\newcommand{\la}{{\langle}}
\newcommand{\ra}{{\rangle}}


\newcommand{\cW}{{\cal W}}
\newcommand{\cV}{{\cal V}}
\newcommand{\cU}{{\cal U}}
\newcommand{\cY}{{\cal Y}}
\newcommand{\cD}{{\cal D}}
\newcommand{\cE}{{\cal E}}
\newcommand{\cX}{{\cal X}}
\newcommand{\cC}{{\cal C}}
\newcommand{\cG}{{\cal G}}
\newcommand{\cN}{{\cal N}}

\newcommand{\bu}{{\bf u}}
\newcommand{\by}{{\bf y}}
\newcommand{\bw}{{\bf w}}
\newcommand{\bU}{{\bf U}}
\newcommand{\tW}{{\tilde{W}}}
\newcommand{\cP}{{\cal P}}
\newcommand{\cB}{{\cal B}}

\newcommand{\iso}{\equiv}
 
\newcommand{\ds}{\displaystyle}


%\def\Bar#1{\raisebox{-.65ex}{\(\stackrel{\mbox{\normalsize$#1$}}%
%                                        {\rule[.1ex]{1.25ex}{.1ex}}\)}}
\def\Bar#1{\raisebox{-.65ex}{\(\stackrel{\mbox{\normalsize$#1$}}%
                                        {\rule[.1ex]{1.25ex}{.05ex}}\)}}
\newcommand{\ud}{u(\cdot)}
\newcommand{\T}{[0,T]}
\newcommand{\be}{\begin{equation}}
\newcommand{\ee}{\end{equation}}
\newcommand{\ba}{\begin{array}}
\newcommand{\ea}{\end{array}}
%\newcommand{\argmin}{{\rm argmin}}
\newcommand{\lip}{{\rm Lip}}
\newcommand{\dist}{{\rm dist}}
\newcommand{\tr}{{\rm tr}}
\newcommand{\er}[1]{\hbox{(\ref{#1})}}

%---------------redefine references commands for MCSS

%\newcommand{\bib}[1]{\hbox{{\bf [#1]}}}
%\newcommand{\bibentry}[1]{\item[\bib{#1}]}
%\renewcommand{\cite}{\bib}
%\renewcommand{\bibitem}{\bibentry}
%---------------------------------1.3 spacing for MCSS
%\renewcommand{\baselinestretch}{1.3}
%-----------------------------------------------------
\font\mitbig=cmmi12
\newcommand{\pdt}[1]{{\frac{\partial {\mitbig #1}}{\partial t}}}


\renewcommand{\Re}      {{\bf R}}
% ----------------      {I \! \! R}


% ----- lengthes

\setlength{\textwidth}          {16cm}  %16cm    %13cm
\setlength{\oddsidemargin}      {0.6cm}
\setlength{\evensidemargin}     {0.6cm}
\setlength{\textheight}         {22cm}  %22cm    %19.5cm
\setlength{\topmargin}          {0cm}
\setlength{\parskip}            {\medskipamount}
\setlength{\arraycolsep}        {2.5pt}          % half the usual length
\setlength{\unitlength}         {0.75cm}



% ----- 

\setlength{\arraycolsep}        {2.5pt}

% ----- theorem-like environments
\newtheorem{theorem}            {Theorem}[section]
\newtheorem{corollary}          [theorem]{Corollary}
\newtheorem{defin}          [theorem]{Definition}
\newtheorem{result}          [theorem]{Result}
\newtheorem{conjecture}          [theorem]{Conjecture}
\newtheorem{cor}                [theorem]{Corollary}
\newtheorem{proposition}        [theorem]{Proposition}
\newtheorem{defn}         [theorem]{Definition}
\newtheorem{definition}         [theorem]{Definition}
%\newtheorem{example}            [theorem]{Example}
%\newtheorem{assertion}            [theorem]{Assertion}
\newtheorem{lemma}              [theorem]{Lemma}
\newtheorem{sideremark}         [theorem]{Remark}
\newtheorem{sidenote}           [theorem]{Note}
\newtheorem{sideeg}           [theorem]{Example}

% ----- environments
\newenvironment{remark}         {\begin{sideremark}\rm}{\end{sideremark}}
\newenvironment{note}         {\begin{sidenote}\rm}{\end{sidenote}}
\newenvironment{example}         {\begin{sideeg}\rm}{\end{sideeg}}

% ----- typewriting commands
\newcommand{\frenchpar}         {\hskip\parindent}

% ----- mathematical commands
\def\argmin                     {\mathop{\rm argmin}}
\def\argmax                     {\mathop{\rm argmax}}

\newcommand{\proof}             {{\sc Proof.}\hspace{0.25cm}} %bug with \sc
%\newcommand{\proof}             {{\bf Proof.}\hspace{0.25cm}}
%\newcommand{\qed}               {\hfill $\Box$}
\def\qed{\hfill\qquad\vrule height4pt width4ptdepth0pt\\ }
\newcommand{\eps}               {\varepsilon}
\newcommand{\thetap}            {{\theta^\prime}}
\newcommand{\R}                 {{\bf R}}
\newcommand{\canon}[1]          {C([0,T]\,;\,\R^{#1})}
\newcommand{\Y}                 {{\cal Y}}
\newcommand{\E}                 {{\bf E}}
\newcommand{\Po}                {P^\dagger}
\newcommand{\Eo}                {\E^\dagger}
\newcommand{\Lambdo}            {\Lambda^\dagger}
\newcommand{\bydef}             {\stackrel{\triangle}{=}}
\newcommand{\inverse}[1]        {{\textstyle\frac{1}{#1}}}
\newcommand{\demi}              {\inverse{2}}

% ----- END OF PREAMBLE -----

\begin{document}

 
\title{ AFOSR PROPOSAL}

\author{J.W.~Helton\thanks{
Department of Mathematics,
University of California,
San Diego, CA 92093, USA.  
}
}
%\and
%M.R. James\thanks{
%Department of Engineering,
%Faculty of Engineering and Information Technology,
%Australian National University,
%Canberra, ACT 0200,
%Australia.
%E-mail: Matthew.James@anu.edu.au 
%WWW: http://spigot.anu.edu.au/people/mat/home.html \newline
%The authors  wish to acknowledge the funding of the activities of the 
%Cooperative Research Centre for Robust and Adaptive Systems by the 
%Australian Commonwealth Government under the Cooperative Research 
%Centers Program. 
%Partially supported by NSF, AFOSR.}

%\date{December 1, 1995}
\date{\today}

\maketitle

%\begin{abstract}

 
%\spp

%\noindent
%{\bf Key words:}  


%\end{abstract}

% ----- TABLE OF CONTENTS -----

%\newpage
%\pagestyle{plain}
%\thispagestyle{plain}
%\pagenumbering{roman}

\pagenumbering{arabic}

\def\nin{\noindent}



\bigskip
\bigskip

\noindent
{\bf \Large ABSTRACT}
\bigskip
\bigskip

The design of a system or circuit in which stability is a key constraint
frequently leads to an optimization problem over the space of functions
analytic on the right half plane (R.H.P.)   Mathematical techniques for solving
such optimization problems for mean square error ($L^2$ error) criteria have
been widespread in engineering since the time of Wiener. Much of my effort goes
to developing techniques for handling worst case error ($L^\infty$ error)
criteria. These occur naturally in design of control systems and amplifiers.
While $L^\infty$ (bang-bang) control in the time domain is heavily studied, work
on frequency domain $L^\infty$ control is more recent. Practically speaking\ there
is evidence that frequency domain $L^\infty$ criteria control system designs
have desirable robustness properties. Furthermore, ``classical control'' as
opposed to LQG control, is very naturally formulated in $L^\infty$ terms, so
there is a substantial hope of codifying classical control and thereby putting
it into a computer implementable form which will extend to many input many
output systems (MIMO). The ultimate objective is to develop a new CAD approach
to MIMO control design which has the flavor of classical control as well as a
systematic approach to worst case frequency domain design as it occurs in many
areas. The promise of this approach is sufficient to have attracted many
investigators and it is currently the focus of much attention.

This is a proposal to 
extend $H^\infty$ control in new directions and to 
continue work on many aspects of the problem ranging
from the development of new computer algorithms to new theoretical methods.
One major effort goes to control of non-linear plants. 
Another concerns computer algebra methods appropriate for systems which are 
presented as interconnections of many subsystems.

\newpage




\noindent
{\Large \bf STATEMENT OF WORK}
\bigskip

The design of a system or circuit in which stability is a key constraint
frequently leads to an optimization problem over the space of functions
analytic on the right half plant, R.H.P.  Mathematical techniques for solving
such optimization problems for mean square error ($L^2$ error) criteria have
been widespread in engineering since the time of Wiener. Much of my effort goes
to developing techniques for handling worst case error ($L^\infty$ error)
criteria. These occur naturally in design of control systems and amplifiers.
While $L^\infty$ (bang-bang) control in the time domain is heavily studied, work
on frequency domain $L^\infty$ control is recent. Practically speaking\ there
is evidence that frequency domain $L^\infty$ criteria control system designs
have desirable robustness properties. Furthermore, ``classical control'' as
opposed to LQG control, is very naturally formulated in $L^\infty$ terms, so
there is a substantial hope of codifying classical control and thereby putting
it into a computer implementable form which will extend to many input many
output systems (MIMO). The ultimate objective is to develop a new CAD approach
to MIMO control design which has the flavor of classical control as well as a
systematic theory.

  The promise of this approach has attracted an enormous number 
of investigators and much of the field is devoted to case studies, 
many of which report successs. Most commercial software and most 
researchers deal with what I call disk or quasicircular problems. These 
have come a very long way since the first solutions were introduced 
to the frequency domain design community for Single Input Single Output
(SISO) systems by Youla 
(mid 60's), for Multiple Input Multiple Output (MIMO) systems by myself (mid 70's) and to control by 
Zames-Francis and Tannenbaum $(\approx 1980)$. 
While work remains to be done
with disk  problems, most of my attention goes to developing mathematics for
what I think will be the next generation of worst case engineering 
design. 
This splits into several parts.

\medskip

\noindent {\bf Nonlinear Plants}
       
Industry on many fronts is running into highly nonlinear 
problems which it often treats with linear techniques. For example, in 
the aircraft industry gain scheduling is common. 
At high angles of attack linearization eventually fails to 
keep up with the highly nonlinear behavior of the airplane.
On a jet engine 
there are sometimes several hundred operating points so this linear 
method is being pushed to extremes. 
The objective is to extend the $H^\infty$ control theory which
has been so successsful for linear plants to nonlinear plants.


Five year old discrete time 
work of Ball and Helton indicated that measurement feedback
control of stable plants 
requires only one HJBI (non-linear Riccati) 
equation. This is of comparable difficulty
to state feedback control where there was
was considerable progress
by Byrnes- Isidori, Isidori and coworkers, by Van der Schaft
and by the group consisting of
Ball, Helton, Walker, and Zhan.
Recently, the measurement feedback problem has made a big advance
due to work of James-Baras- Helton, Didinsky-Basar,  
and also van der Schaft.

 
Extending  $H^\infty$ control
to  a nonlinear system 
at this time has been converted to solving two 
particular PDE's.

One P.D.E. , {\it the state feedback HJBI} ,
corresponds to the $CTRL^\infty$ problem  where the
controller can access the full state of the plant.
Indeed solving this PDE
gives an optimal solution to the state feedback
$CTRL^\infty$ problem and can be solved off line.
The second PDE,  the {\it information state PDE},
gives the dynamics of the controller and so it must be solved 
ON LINE. 

Unfortunately, these are P.D.E's on the state space of the original plant
(often a high dimensional space)
so a numerical solution faces what is called the curse of dimensionality.
Getting around this is the main challenge to the field and
it requires basic theory as well as numerical analysis.
 
Collaborators and I have ongoing work on both of these equations.
Much of the effort goes to the information state PDE since 
at first glance, computation (on line) seems 
impossible in general situations with state-space higher than two.
However,
recent work with M . James indicates that, for the mixed sensitivity problem,
the controller dynamics in practice might NOT suffer prohibitively from
the curse of dimensionality.
Formulas are now in place and our main effort is checking their
range of validity and robustness.
Work on th state feedback HJBI focuses on numerical solutions.



\medskip

%---------------------------------------------------------

\noindent $H^\infty$ {\bf Design}

$H^\infty$ optimization is critical to design of linear systems
meeting realistic performance conditions.
The fundamental $H^\infty$ problem of control
is that of finding the stable frequency response function
that best fits worst case frequency domain specifications.
This is a non smooth optimization problem that underlies
the frequency domain formulation of the
$H^\infty$ problem of control.

Textbook designs give ``disk'' problems. Competing constraints 
give intersections of disks. Plant uncertainty gives highly complicated 
$H^\infty$ non-disk problems. I have work which 
address most aspects of these problems.
Collaborators (primarily  Orlando Merino) and I develop
qualitative theory,  algorithms, theory of the algorithms, 
and computer programs to check the  algorithms. 
We recently showed  how
the fundamental $H^\infty$ optimization problem of control
can be naturally treated with modern
primal-dual interior point (PDIP) methods.
Our  theory generalizes and unifies
approaches to solving large classes of optimization problems
involving matrix valued functions,
a subclass of which are commonly treated with
Linear Matrix Inequalities techniques (LMI).
Our work relies on the study of 
optimality conditions for
$H^\infty$ optimization problems over matrix valued functions.
In experiments we find the algorithms exhibit
(local) quadratic convergence rate in many instances.

One of the most intriguing and important questions 
to which we are starting to devote considerable attention 
is 
the development of algorithms and of optimality tests
for solving plant uncertainty problems in general.  
We also are having success with 
$H^\infty$ Optimization with time domain constraints.
%

Ours is the only
computational effort on general $H^\infty$  optimization found in a mathematics 
department; all others are in  
engineering departments and this gives our work a
distinctive flavor. 

%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%


\iffalse
DELETE ??


??
The fundamental $H^\infty$ problem of control
is that of finding the stable frequency response function
that best fits worst case frequency domain specifications.
This is a non smooth optimization problem that underlies
the frequency domain formulation of the
$H^\infty$ problem of control;
it is the main optimization problem in Qualitative Feedback Theory
for example.
It is shown in this article how
the fundamental $H^\infty$ optimization problem of control
can be naturally treated with modern
primal-dual interior point (PDIP) methods.
The theory introduced here generalizes and unifies
approaches to solving large classes of optimization problems
involving matrix valued functions,
a subclass of which are commonly treated with
Linear Matrix Inequalities techniques.
Also, in this article new
optimality conditions for
$H^\infty$ optimization problems over matrix valued functions
are proved, and
numerical experience on  natural (PDIP) algorithms
for these problems is reported.
In experiments we find the algorithms exhibit
(local) quadratic convergence rate in many instances.
Finally, $H^\infty$ optimization problems with an
uncertainty parameter are considered.  It is shown how
to apply the theory developed here to obtain
optimality conditions and derive algorithms.
Numerical tests on simple examples are reported.
 
The fundamental $H^\infty$ problem of control
is that of finding the stable frequency response function
that best fits worst case frequency domain specifications.
This is a non smooth optimization problem that underlies
the frequency domain formulation of the
$H^\infty$ problem of control;
it is the main optimization problem in Qualitative Feedback Theory
for example.
It is shown in this article how
the fundamental $H^\infty$ optimization problem of control
can be naturally treated with modern
primal-dual interior point (PDIP) methods.
The theory introduced here generalizes and unifies
approaches to solving large classes of optimization problems
involving matrix valued functions,
a subclass of which are commonly treated with
Linear Matrix Inequalities techniques.


Also, in this article new
optimality conditions for
$H^\infty$ optimization problems over matrix valued functions
are proved, and
numerical experience on  natural (PDIP) algorithms
for these problems is reported.
In experiments we find the algorithms exhibit
(local) quadratic convergence rate in many instances.
Finally, $H^\infty$ optimization problems with an
uncertainty parameter are considered.  It is shown how
to apply the theory developed here to obtain
optimality conditions and derive algorithms.
Numerical tests on simple examples are reported.
 

The core of our
$H^\infty$ optimization theory is a set of equations which locally characterize
optimal functions.
Indeed there are many such characterizations
most of which tend to underdetermine the answer.
This prevents Newton's method from being effective.
In a previous article [HMW] we described 2 ways to cut down
a characterization when the performance function is scalar valued
to produce a set
of optimality conditions on which Newton's method is effective.
Unfortunately this does not extend to the more general situations
treated in this article.
 
In \S 3  we generalize
the optimality conditions from scalar
valued to matrix valued performance functions.
This adds the complication of yet further types of
nonsmoothness to the $H^\infty$ optimization problem and
the difficulty that local optima typically are not isolated.
This latter occurrence is a traditional bane to
the second order convergence of Newton's method,
however an accomplishment of this article is that we
present methods which appear to overcome this second
difficulty in many cases.
The added generality of the matrix problems
is physically important
since this contains engineering problems with competing constraints
and plant uncertainty.
Among the numerical schemes we propose and examine (in \S 4 and \S 5) are
the so called "interior point methods".
These of course iterate Newton steps.
We find on several tests of local (rather than global behavior)
that interior point methods do not offer great advantages.
This is discussed in \S 6.
An important conclusion of our tests in \S 6 is that
the order of convergence seems to be intimately related to
the characteristics of the optimal dual function.
Roughly, numerical experiments indicate that
quadratic convergence rate holds when the
optimal dual function is smooth, linear convergence rate
holds when the optimal dual function is continuous but not smooth,
and very slow convergence rate holds when the optimal dual function
is not continuous.
 
Applications of our theory to $H^\infty$ optimization
with an uncertainty parameter are considered in \S 7 and \S 8.
There, we apply the theory developed in previous sections to obtain
optimality conditions and derive algorithms.
Numerical tests on simple examples are reported.


\newpage

\noindent $H^\infty$ {\bf Design ATTEMPT}

In my view three main subjects within $H^\infty$ optimization
are fundamental and my proposed research intends to 
address all of them.
%
%
\begin{itemize}
\item[1.] {\sl General Theory of $H^\infty$ Optimization.}
We recently showed  how
the fundamental $H^\infty$ optimization problem of control
can be naturally treated with modern
primal-dual interior point (PDIP) methods.
Our  theory generalizes and unifies
approaches to solving large classes of optimization problems
involving matrix valued functions,
a subclass of which are commonly treated with
Linear Matrix Inequalities techniques (LMI).
Our work relies on the study of 
optimality conditions for
$H^\infty$ optimization problems over matrix valued functions.
In experiments we find the algorithms exhibit
(local) quadratic convergence rate in many instances.
%
\item[2.] {\sl $H^\infty$ Optimization with time domain constraints.}
Our initial efforts in this area yielded qualitative results in the form
of optimality conditions, as well as algorithms. 
%
\item[3.] {\sl $H^\infty$ Optimization with plant uncertainty.}
One of the most intriguing and important questions is 
the development of algorithms and of optimality tests
for solving plant uncertainty problems in general.  
%
\end{itemize}





Collaborators (primarily  Orlando Merino) and I develop
qualitative theory,  algorithms, theory of the algorithms, and computer programs to check the  algorithms. 
Ours is the only
computational effort on general $H^\infty$  optimization found in a mathematics 
department; all others are in  
engineering departments and this gives our work a
distinctive flavor. 

%%%%%%%%%%%%%%%%%%%%%%%%%%

The fundamental $H^\infty$ problem of control
is that of finding the stable frequency response function
that best fits worst case frequency domain specifications.
This is a non smooth optimization problem that underlies
the frequency domain formulation of the
$H^\infty$ problem of control;
it is the main optimization problem in Qualitative Feedback Theory
for example.

At the core of our
$H^\infty$ optimization theory is a set of equations which locally characterize
optimal functions.
Indeed there are many such characterizations
most of which tend to underdetermine the answer.
This prevents Newton's method from being effective.

We recently generalized 
the optimality conditions obtained by us in the late 80's  from scalar
valued to matrix valued performance functions.
This adds the complication of yet further types of
nonsmoothness to the $H^\infty$ optimization problem and
the difficulty that local optima typically are not isolated.
This latter occurrence is a traditional bane to
the second order convergence of Newton's method,
however an accomplishment of our approach is that we
present methods which appear to overcome this second
difficulty in many cases.
The added generality of the matrix problems
is physically important
since this contains engineering problems with competing constraints
and plant uncertainty.
Among the numerical schemes we propose and are
the so called "interior point methods".
These of course iterate Newton steps.
We find on several tests of local (rather than global behavior)
that interior point methods do not offer great advantages.
This is discussed in \S 6.
An important conclusion of our testsis that
the order of convergence seems to be intimately related to
the characteristics of the optimal dual function.





END DELETE
------------------------------
%%%%%%%%------------------------------
%%%%%%%%------------------------------
\fi

\noindent

\medskip

\noindent 
{\bf Symbolic Algebra for Systems Research}

%We are the providers to Mathematica of general Noncommuting
%algebra capability.

%We have a major computer package,
%NCAlgebra, which is a collection of "functions" for Mathematica 
%designed to facilitate manipulation and reduction of noncommutative 
%algebraic expressions in operator theory and engineering systems.
%
Noncommutative algebra is a cornerstone of matrix theory
and, consequently, of research in linear engineering systems.
For example, a typical issue of the control transactions
contains many articles full of long calculations in noncommuting 
algebra.
Also large scale systems are usually presented
as interconnections of many subsystems. Mathematically,
each subsystem corresponds to a matrix or a matrix valued
transfer function. Thus, dealing with large scale
systems forces one immediately into equations
in noncommuting elements. 


Our effort should not be confused with the business of doing
standard systems calculations symbolically. This fairly extensive
area manipulates commutative expressions only.



Our work has 3 phases which run from bread and butter to ambitious research.

\begin{description}
\item[1.]
{NCAlgebra: The basic commands of Mathematica, Maple etc do NOT apply to
noncommuting expressions. 
Our group wrote a 
package NCAlgebra which does noncommutative algebra in Mathematica 
and which we view as a reasonable beginning.
To our knowledge this is the only interactive package
which does general non-commuting algebra. 
(There are packages devoted to special noncommuting objects
such as differential forms and tensors.)
Mathematica recommends it widely, and it differentiates them
from Maple which has no such capability.


Running under NCAlgebra
is a collection of files called SYSTEMS.
This software aims toward the specialized objective of doing 
Hamiltonian (energy balance) calculations
of the type found in $H^\infty$ control for linear  systems.
We have used this extensively in the nonlinear work,
since it saves  a lot of time.}
\item[2.]{ Simplifying expressions: Our simplification 
algorithm has been 
demonstrated to our satisfaction 
at least in certain restricted classes of (control) problems 
where it proved valuable.
It is based on 
noncommutative
Gr\"obner Basis Algorithms (10 years old).
}
\item[3.]{ Eliminate unknowns from collections of equations.
The goal is to give ``nice forms" for the equations and 
help one discover elegant formulas. 
Our method uses Non Commutative GB Algorithms's plus 
algorithms we have developed for 
removing redundant equations.}
\end{description}

Current research focuses on developing this last area 
to help an engineer discover formulas which are critical to his
particular problem.





\noindent {\bf Other Directions}

I play a role in various local engineering projects.

%\end{document}

\end{document}
