Bompart Vincent

Universite Paul Sabatier Toulouse III
U.F.R.

Ecole Doctorale
Mathematiques, Informatique Mathematiques, Informatique
et Gestion et Telecommunications de Toulouse
Optimisation non lisse pour la
commande
des syst`emes de lAeronautique
TH
`
ESE
presentee et soutenue publiquement le 23 novembre 2007
pour lobtention du
Doctorat de lUniversite Paul Sabatier Toulouse III
(specialite Mathematiques appliquees et Automatique)
par
Vincent BOMPART
Composition du jury
Rapporteurs : M. Samir ADLY Universite de Limoges
M. Pierre-Alexandre BLIMAN INRIA Rocquencourt
Examinateurs : M. Paul ARMAND Universite de Limoges
M. Jean-Philippe HARCAUT MBDA France
M. Jean-Baptiste HIRIART-URRUTY Universite Paul Sabatier
Toulouse III
Directeurs : M. Pierre APKARIAN ONERA Centre de Toulouse
M. Dominikus NOLL Universite Paul Sabatier
Toulouse III
Institut de Mathematiques de Toulouse UMR CNRS 5219
Equipe Mathematiques pour lIndustrie et la Physique
Mise en page L
A
T
E
X base sur la classe thloria.
i
Remerciements
Mes premiers remerciements sadressent Pierre Apkarian et Dominikus Noll, tous
deux lorigine de cette thse et qui lont dirige conjointement durant ces trois annes ;
jai pu bncier de leur expertise scientique et de leurs clairages complmentaires, mais
galement de leur coute et de leur gentillesse.
Je remercie les rapporteurs, Samir Adly et Pierre-Alexandre Bliman : ils ont accept
de prendre connaissance de mon travail de faon approfondie, malgr les dlais rduits
dont ils disposaient, et leurs remarques prcises ont permis damliorer la qualit de ce
document. Merci galement aux autres membres du jury, Paul Armand, Jean-Philippe
Harcaut et Jean-Baptiste Hiriart-Urruty, qui ont bien voulu se rendre disponibles pour
juger mon travail.
Jai pu bncier pendant toute la thse de lenvironnement scientique et humain de
lONERA, ainsi que de son soutien nancier. Jai t accueilli au dpartement de com-
mande des systmes et dynamique du vol (DCSD), Toulouse, dont je souhaite remercier
tout particulirement le directeur, Patrick Fabiani, qui a su me donner les moyens de
mener bien cette thse, ainsi que Philippe Mouyon, responsable de lunit de recherche
commande et intgration (UR CDIN). Merci galement aux ingnieurs de recherche du
dpartement (et en particulier Pierre Vacher, Michel Corrge, et . . .Catherine Tessier
pour ses clats de rire communicatifs). Je noublie pas les membres du groupe de travail
Optimisation et Interactions, au sein de lquipe MIP de lInstitut de Mathmatiques de
Toulouse.
Merci aux doctorants du DCSD : Jean-Baptiste, dont jai partag le bureau pen-
dant la premire anne, et qui ma pass le relais et encourag. Bureau que jai mon
tour partag avec Patrice qui, jen suis sr, trouvera vite un remplaant pour changer
ses rexions scientiques ou philosophiques toujours passionnantes. Merci aussi, dans le
dsordre, Sophie, Damien, Greg et Greg, Manu, Nico, Florian, Andreas, Alex, Julien,
Sylvain B., Clment, Alberto, et aux anciens thsards, Charles, Olivier, Cdric, Florent,
Sylvain D., Sebastian, Fabien et Elodie. Encore des remerciements pour Olivier et Aude
de lquipe MIP.
La thse est un investissement personnel de chaque instant, qui passe par des doutes
et des remises en question parfois trop lourds porter seul ; mes parents ont su mcou-
ter, me parler et me soutenir dans ces moments dlicats, et je dsire ici leur tmoigner
ma profonde reconnaissance. Merci galement David pour ses encouragements constants.
Un dernier petit mot et une pense, enn, pour ma grand-mre, Marie-Louise, qui sen
est alle alors que je rdigeais ma thse.
ii
Table des matires
Introduction 3
I Notions fondamentales et cadre de travail 9
1 Introduction la commande des systmes 11
1.1 Terminologie et dnitions . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.1 Systmes dynamiques . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.2 Normes de signaux . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.1.3 Stabilit des systmes linaires . . . . . . . . . . . . . . . . . . 16
1.1.4 Normes de systmes . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2 Problmes de synthse pour la commande . . . . . . . . . . . . . . . . 19
1.2.1 Forme standard pour la commande . . . . . . . . . . . . . . . . 20
1.2.2 Stabilisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.2.3 Synthse H
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2 Optimisation non lisse 33
2.1 Terminologie et dnitions . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.1 Ingrdients pour lanalyse non lisse . . . . . . . . . . . . . . . . 33
2.1.2 Analyse non lisse : le cas convexe . . . . . . . . . . . . . . . . . 36
2.1.3 Analyse non lisse : le cas lipschitizien . . . . . . . . . . . . . . . 38
2.2 Conditions doptimalit . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.3 Algorithmes pour loptimisation non lisse . . . . . . . . . . . . . . . . . 43
2.3.1 Direction de descente . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3.2 Mthodes de faisceaux . . . . . . . . . . . . . . . . . . . . . . . 45
2.3.3 Mthodes pour les problmes minimax . . . . . . . . . . . . . . 47
iii
iv Table des matires
II Algorithmes dvelopps et applications 55
3 Techniques non lisses pour la stabilisation des systmes linaires 57
3.1 Introduction and notations . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2 Minimizing the spectral abscissa . . . . . . . . . . . . . . . . . . . . . . 61
3.3 Subgradients of the spectral abscissa . . . . . . . . . . . . . . . . . . . 62
3.3.1 Subgradients in state-space . . . . . . . . . . . . . . . . . . . . 62
3.3.2 Subgradients and dynamic controllers . . . . . . . . . . . . . . . 64
3.3.3 Subgradients with structured controllers . . . . . . . . . . . . . 64
3.4 Descent step and optimality function . . . . . . . . . . . . . . . . . . . 65
3.5 Nonsmooth descent algorithms . . . . . . . . . . . . . . . . . . . . . . 66
3.5.1 Variant I (rst-order type) . . . . . . . . . . . . . . . . . . . . . 66
3.5.2 Variant II (second-order type) . . . . . . . . . . . . . . . . . . . 68
3.6 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.6.1 Academic test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.6.2 Transport airplane . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.6.3 VTOL helicopter . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.6.4 B-767 airplane . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.6.5 PID controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4 Optimisation non lisse pour la commande structure 81
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2 Nonsmooth H
design technique . . . . . . . . . . . . . . . . . . . . . 85
4.2.1 Subdierential of the H
map . . . . . . . . . . . . . . . . . . 86
4.2.2 Structured controllers . . . . . . . . . . . . . . . . . . . . . . . 87
4.2.3 PID controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.2.4 Setpoint lter design . . . . . . . . . . . . . . . . . . . . . . . 90
4.2.5 Nonsmooth descent method . . . . . . . . . . . . . . . . . . . . 91
4.3 Nonsmooth loop-shaping design . . . . . . . . . . . . . . . . . . . . . . 94
4.3.1 Loop-shaping design . . . . . . . . . . . . . . . . . . . . . . . . 94
4.3.2 PID Loop-shaping design . . . . . . . . . . . . . . . . . . . . . 95
4.4 Application to a separating tower . . . . . . . . . . . . . . . . . . . . 96
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
v
5 Commande temporelle et frquentielle par techniques non lisses 103
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2 Time- and frequency domain designs . . . . . . . . . . . . . . . . . . . 107
5.3 Nonsmooth descent method . . . . . . . . . . . . . . . . . . . . . . . . 109
5.3.1 Nonsmooth properties . . . . . . . . . . . . . . . . . . . . . . . 111
5.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.4.1 Application to observer-based controller design . . . . . . . . . 115
5.4.2 Application to reliable control . . . . . . . . . . . . . . . . . . 116
5.4.3 Application to PID time response shaping . . . . . . . . . . . . 121
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6 Optimisation non lisse du second ordre pour la synthse H
125
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.2 Problem setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.3 Approach via semi-innite programming . . . . . . . . . . . . . . . . . 129
6.4 Solving with SQP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.4.1 Quadratic tangent subproblem . . . . . . . . . . . . . . . . . . 132
6.4.2 First and second derivative formulas . . . . . . . . . . . . . . . 134
6.4.3 Multiple eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . 136
6.4.4 Globalisation via trust-region . . . . . . . . . . . . . . . . . . . 138
6.5 Technical aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.5.1 Identifying peak frequencies . . . . . . . . . . . . . . . . . . . . 140
6.5.2 Stopping criteria . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.6 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Conclusion 147
Bibliographie 149
vi Table des matires
1
Notations et symboles
Notation Signication
Egal , par dnition
Ensemble vide
R, C Corps des nombres rels, des nombres complexes
Re z, Imz Partie relle, partie imaginaire de z C
[x[, [z[ Valeur absolue de x R, module de z C
R
+
Ensemble des nombre rels positifs ou nuls
x
+
Partie positive de x C
x
+
max x, 0
C
, C
+
Demi-plan ouvert gauche, droit
R
n
, C
n
Espace des vecteurs rels, complexes n composantes
R
pm
Espace vectoriel des matrices relles de taille p m,
soit p lignes et m colonnes
(L
n
(R) Groupe linaire des matrices carres inversibles de taille nn
o
n
(R) Sous-espace vectoriel des matrices symtriques relles de taille
n n
I
n
, I Matrice identit (de taille n, de taille fonction du contexte)
X ~ 0, X 0 X o
n
(R) est dnie positive, dnie ngative
X _ 0, X _ 0 X o
n
(R) est semi-dnie positive, semi-dnie ngative
Tr X Trace de la matrice X C
nn
vec X Vectoris (par colonnes) de X R
pm
X
T
Matrice transpose de X R
pm
X
H
Matrice adjointe de X C
pm
X
1
Matrice inverse de X C
nn
X
+
Pseudo-inverse (ou inverse de Moore-Penrose) de X C
pm
X
1
2
Racine carre de la matrice X _ 0
x, y) Produit scalaire euclidien (hermitien) sur R
n
(C
n
)
x, y) x
H
y
X, Y )) Produit scalaire euclidien (hermitien) sur R
pm
(C
pm
)
X, Y )) Tr (X
H
Y )
|x|
Norme inni du vecteur x R

n
(x C
n
)
|x|
max
1in
[x
i
[
|x|
2
ou |x| Norme euclidienne (hermitienne) du vecteur x R
n
(x C
n
)
|x|
2

x
H
x
B(x, r) Boule de centre x R
n
et de rayon r R
+
pour la norme
euclidienne
B(X, r) y R
n
: |y x| < r
2 Table des matires
Notation Signication
|X|
F
ou |X| Norme hermitienne (ou de Frbenius) de la matrice X C
pm
|X|
F

_
Tr (X
H
X)
|X|
2
Norme 2 de la matrice X C
pm
|X|
2
(X)
u(t) ou u Signal causal rel temps continu u : t R
+
u(t) R
m
U(s) ou U Transforme de Laplace du signal u
U(s)
_
+
0
u(t)e
st
dt pour tout s C tel que Re s x
0
|u|
Norme inni (valeur de crte) du signal u

|u|
2
Norme L
2
(nergie) du signal u L
2
(R
+
)
|U|
2
Norme L
2
ou, suivant le contexte, norme H
2
|U|
Norme H
(X),
(X) Spectre, pseudo-spectre de la matrice X C

nn
i
(X), 1 i n Valeurs propres de la matrice X C
nn
i
(X), 1 i q Valeurs propres (sans rptitions) de la matrice X C
nn
(X),
(X) Abscisse spectrale, pseudo-spectrale de la matrice X C

nn
(X) Distance linstabilit de la matrice X C
nn
(X) Rayon spectral de la matrice X C
nn
(X) max
i
[
i
(M)[
i
(X), 1 i min(m, p) Valeurs singulires (en ordre dcroissant) de la matrice X
C
pm
(X) ou
1
(X) Plus grande valeur singulire de la matrice X C
pm
f : R
n
R Fonction f dnie sur R
n
, valeurs relles
[a, b] Segment dextrmits a, b R
n
[a, b] x R
n
: il existe t, 0 t 1 et x = ta + (1 t)b
co S Enveloppe convexe de S R
n
min S, max S Minimum (plus petit lment), maximum (plus grand l-
ment) dun ensemble non vide S R (sils existent)
min
xE
f(x), max
xE
f(x) Valeur minimum, maximum de f sur E (si elle en admet), ou,
suivant le contexte, problme de minimisation, de maximisa-
tion (locale ou globale) de f sur E
arg min
xE
f(x), Un minimum, un maximum x E de f (sil en existe)
arg max
xE
f(x)
f
t
(x) Direntielle de f en x R
n
f
t
x
(x, t) Direntielle partielle de f par rapport x
f
t
(x, d) Drive directionnelle de f en x R
n
, dans la direction d R
n
f
(x, d) Drive directionnelle gnralise de f en x R

n
, dans la
direction d R
n
c
f(x) Sous-direntiel de f en x R
n
, en analyse convexe
f(x) Sous-direntiel de Clarke de f en x R
n
Introduction
Les processus industriels prsentent une complexit grandissante, quil sagisse dunits
de production (racteurs chimiques, centrales lectriques), de moyens de transport (avions
civils ou militaires, hlicoptres), ou autres. Ils rsultent frquemment de linterconnexion
de plusieurs sous-systmes mis en rseau. Ils intgrent des fonctionnalits en plus grand
nombre quauparavant, et doivent satisfaire des cahiers des charges plus contraignants :
satisfaction de contraintes de qualit ou de scurit, respect de normes environnementales
(bruit, pollution)... Le respect de ces spcications passe par la conception dun systme
de commande, dont laction sur le processus lui confre les proprits souhaites.
Ces processus sont par ailleurs de mieux en mieux modliss : les techniques diden-
tication sont plus performantes et peuvent intgrer de plus grands volumes de donnes
exprimentales. Les moyens de simulation de ces modles suivent laccroissement de leur
complexit.
Ainsi, les mthodes de conception de commande dites classiques, bases sur des tech-
niques frquentielles, essentiellement graphiques, parfois empiriques, ne sont plus adap-
tes. La formulation des objectifs de commande sous la forme dun problme doptimisa-
tion (commande LQ, LQG, H
2
, H
), la base de lAutomatique moderne, a permis de

mieux apprhender la commande des systmes complexes. Les lois de commande optimales
sexpriment alors comme des solutions dquations algbriques (quations de Lyapunov,
de Riccati), ou de problmes de minimisation sous des contraintes matricielles de positi-
vit. Les ingalits linaires matricielles (ou LMI, de langlais linear matrix inequalities)
ont acquis une grande popularit dans le domaine de la commande des systmes ; elles per-
mettent de formuler un grand nombre de problmes danalyse des systmes dynamiques,
et de synthse de lois de commande. Cest avec la thorie de Lyapunov sur la stabilit
des tats dquilibre, plus que centenaire (n du xix
e
), quapparat la premire LMI. Elle
caractrise la stabilit interne du systme dynamique dni par le systme direntiel
x(t) = Ax(t) (A tant une matrice carre relle) par lexistence dune matrice symtrique
dnie positive telle que A
T
P +PA est dnie ngative, ce que lon note
A
T
P +PA 0 . (1)
P, de mme taille que A, est appele matrice de Lyapunov. De nombreux autres problmes
de commande conduisent la rsolution dune ou plusieurs LMI, ou la minimisation
dun critre linaire sous contraintes de type LMI (que lon dnomme programmation
semi-dnie, ou SDP). La conception dalgorithmes ecaces, adapts de tel problmes
a largement popularis les techniques LMI parmi les automaticiens : en eet, si la LMI
(1) peut facilement tre rsolue de faon explicite (il sut dinverser le systme linaire
3
4 Introduction
A
T
P +PA = I), ce nest pas le cas dune LMI quelconque, sous la forme gnrique (ou
standard)
F
0
+
m
i=1
x
i
F
i
0 , (2)
o les matrices F
i
(0 i n) sont symtriques relles. Ce nest quavec lapparition de
mthodes itratives, telles celles dites des points intrieurs, issues de la programmation
linaire il y a une vingtaine dannes, que le domaine de recherche a dpass le cadre de
lAutomatique thorique, pour connatre un essor certain.
Tous les problmes de commande ne sont cependant pas rductibles des LMI/SDP!
Bien souvent, en pratique, et mme pour le critre aussi essentiel quest la stabilit
interne, les expressions matricielles ne sont pas linaires ou anes. Plus exactement, elles
ne sont pas convexes, et les problmes doptimisation sous ce type de contraintes sont
bien plus diciles rsoudre. Quand bien mme la synthse dun correcteur peut se
ramener, aprs dventuelles manipulations algbriques, un problme SDP, la loi de
commande obtenue est bien souvent du mme ordre
1
que le systme commander. Or,
dans les applications embarques, par exemple, les ressources en puissance de calcul et
en mmoire sont limites ; larchitecture du processeur cible peut tre particulire, et
contraint la complexit de la loi de commande implmenter.
De plus, la mise en uvre de la commande sur le systme rel saccompagne dune
tape de validation, ayant ventuellement pour consquence la retouche du correcteur,
rendue dicile par son niveau de complexit. La synthse dun correcteur plus simple
que le systme commander (on parle de correcteur dordre rduit) gure donc parmi les
thmes de recherche actuels de lAutomatique, pour lesquels les applications attendues
sont nombreuses.
Dautre part, les savoir-faire spciques lapplication considre et la connaissance
du systme rel impliquent gnralement des contraintes sur la structure de la loi de
commande : dans le cas de systmes rpartis, ou pour limiter la complexit des commandes,
par exemple, il peut tre impos de ne pas lier la commande dun actionneur une mesure
donne (commande dite dcentralise). Bien souvent, laction du correcteur synthtiser
a une forme prdtermine (cas des correcteurs PID par exemple, trs rpandus, qui
combinent une action proportionnelle, intgrale, et drive), et seuls quelques paramtres
restent ajuster. Ces exemples plaident pour la prise en compte de la structure de la loi
de commande en amont, ds sa conception. Malheureusement, lajout a priori de telles
contraintes nest pas ais avec les mthodes actuelles base de LMI, car les variables
du correcteur napparaissent pas toujours explicitement dans les ingalits matricielles
eectivement rsolues.
Enn, la prsence de variables de Lyapunov dans ces ingalits vient les compliquer
articiellement : dune part, la taille de ces variables ne correspond pas la dimension
des variables du correcteur, mais crot avec le carr de lordre du systme comman-
der. Ainsi, la synthse de lois de commande pour les systmes dynamiques comportant
1
lordre dune loi de commande, en tant que systme dynamique suppos linaire, est, intuitivement,
la dimension minimale de la variable qui permet de dcrire compltement son volution, comme solution
dun systme direntiel linaire. Nous en donnerons une dnition prcise au chapitre 1.
5
quelques dizaines dtats est dicile : un Boeing 767, modlis autour dun point de vol
par un systme linaire dordre 55, ncessite 1540 variables scalaires pour la matrice de
Lyapunov. Les systmes linaires modlisant des quations aux drives partielles (EDP),
obtenus par lments nis, sont encore plus gros (souvent plusieurs milliers dtats) et sont
totalement hors de porte dune synthse par LMI/SDP. Dautre part, lordre de grandeur
des variables de Lyapunov na aucune raison dtre le mme que celui des variables du
correcteur, ce qui peut poser de srieux problmes dordre numrique lors de la rsolution
ou de loptimisation.
Ces trois points fondamentaux constituent les principales limitations actuelles des m-
thodes de synthse de lois de commande. Labondance et la richesse des publications
dans le domaine des LMI ces quinze dernires annes na pas apport de mthode num-
rique satisfaisante pour les problmes de synthse structure ou dordre rduit. Celles qui
sont proposes sont souvent heuristiques, sans preuve de convergence, ou ne sappliquent
qu des classes trs particulires de systmes. Mme les approches les plus rcentes,
qui proposent de rsoudre les problmes doptimisation avec contraintes matricielles non
convexes, restent limites par la taille des variables de Lyapunov. Pourquoi alors ne pas
laisser de ct ces variables auxiliaires pour se concentrer sur les vritables variables de
dcision qui dnissent le correcteur ? Ces variables sont gnralement en petit nombre,
car les lois de commande recherches sont les plus simples possibles (correcteurs dordre
rduit, voire statiques), et les mesures et commandes restent en nombre trs raisonnable
(typiquement, de une quatre sorties mesures, autant dentres commandes). Les formu-
lations directes des problmes de synthse amnent cependant minimiser des fonctions
qui se prsentent comme un maximum de fonctions, de la forme
f
() max
x.
f(, x) . (3)
La variable regroupe tous les paramtres du correcteur, et x permet d indexer les
fonctions sur un ensemble A, qui peut tre discret ou continu. Ce dernier peut tre un
intervalle de frquences dans le cas dun critre portant sur les transferts (norme H
par
exemple), un intervalle de dates pour un critre temporel, ou encore un ensemble ni din-
dices sil sagit dun critre spectral. On voit ici que cette formulation est susamment
gnrale pour permettre dexprimer, voire de combiner entre eux de nombreux critres de
synthse. Elle laisse au concepteur le soin de dnir la forme de loi de commande quil
recherche, partir des paramtres scalaires contenus dans . Dans le cas dun correcteur
statique, par exemple, contient tous les gains scalaires, canal par canal. Pour un correc-
teur PID, regroupe les gains des actions proportionnelle, intgrale et drive. Le prix
payer cependant dune telle approche est le caractre non lisse, cest--dire non partout
direntiable, du critre minimiser. Qui plus est, les points o concident les valeurs de
plusieurs fonctions partielles f(, x) sont gnralement de bons candidats pour constituer
des minima locaux (voir gure 1). Les algorithmes doptimisation direntiable (descente
de gradient, Newton ou quasi-Newton, programmation quadratique successive pour les
problmes sous contraintes, etc.) sont alors mis en dfaut.
Le cadre thorique adapt est celui propos par lanalyse non lisse, qui gnralise les no-
tions de direntielle et de gradient certaines classes de fonctions (essentiellement les
6 Introduction
f
2
()
f
() = max f
1
(), f
2
()
f
1
()

Fig. 1 Fonction max sur R (en trait gras) : f
nest pas drivable en
.
fonction convexes dune part, et les fonctions localement lipschitziennes dautre part).
Cest un cadre riche dun point de vue de lOptimisation numrique, car il permet de
construire des algorithmes de descente et de prouver leur convergence vers des minima
locaux. Ces algorithmes doptimisation non lisse doivent tre bien distingus des m-
thodes doptimisation sans drive (parfois dites de recherche directe, ou dordre zro),
qui nexploitent que les valeurs de la fonction critre pour progresser.
Ainsi le point de vue scientique adopt dans cette thse se situe la croise de
lAutomatique (puisque nous nous intressons des problmes de commande des systmes
linaires), de lOptimisation non lisse, et du dveloppement logiciel. Sur ce dernier point,
lobjectif concret est, en eet, de proposer un outil numrique cohrent, document, et
volutif, directement accessible lautomaticien, et utilisable la faon dune bote
outils pour la synthse de lois de commande. Cet outil permet de formuler et de rsoudre
des problmes diciles, notamment en terme de taille (systme commander dordre
lev), de contraintes sur le compensateur synthtiser (ordre ou structure xs), ou de
nature du problme (synthse multimodle, commande tolrante aux pannes). Il repose
sur le socle thorique et algorithmique mis en place dans la thse. Cest l que rside
lobjectif central de nos travaux : dmontrer le bien-fond et la faisabilit de lapproche
non lisse pour la rsolution de problmes de synthse de lois de commande. Pour ce faire,
nous nous sommes attachs prsenter des applications varies, issues pour partie dune
bibliothque de problmes de commande (bibliothque COMPl
e
ib de F. Leibfritz), mais
aussi dexemples isols, diciles, traits jusqualors par des mthodes heuristiques.
Le document est organis en deux grandes parties : la premire partie regroupe en deux
chapitres (chapitres 1 et 2) lensemble des notions qui nous ont sembl tre fondamen-
tales. Ces deux chapitres sont complmentaires dans leur thmatique : le lecteur familier
des systmes dynamiques linaires et de leur commande en boucle ferme pourra survoler
7
le premier chapitre, dans lequel nous formulons les problmes tudis (stabilisation et
synthse H
) et donnons un aperu des mthodes de rsolution existantes. Il pourra en

revanche sattarder sur la lecture du deuxime chapitre, qui expose les bases de lOptimi-
sation non direntiable, hors du contexte de lAutomatique, pour les fonctions convexes,
puis pour les fonctions localement lipschitziennes. Plusieurs algorithmes doptimisation y
sont prsents (mthodes de faisceaux, mthodes pour des problmes minimax).
La contribution scientique de la thse fait lobjet de la seconde partie, elle-mme
divise en quatre chapitres (chapitres 3 6). Chacun de ces chapitres reprend dans son
contenu une partie des travaux de thse sous la forme dune publication ou dune commu-
nication. Le choix de lordre de prsentation des travaux ne respecte pas ncessairement
la chronologie de ceux-ci, mais a t pens de faon proposer une progression logique au
lecteur. Nous partons du problme de la stabilisation en boucle ferme (chapitre 3), qui
sest rvl fondamental pour la minimisation ultrieure dun critre de performance tel
que la norme H
(dont traitent les chapitres 4 6). Nous avons donc compar direntes
techniques doptimisation permettant de rechercher un correcteur stabilisant, dont une
approche non lisse originale que nous prsentons. Laccent est ensuite mis sur la synthse
H
, travers un exemple de loop-shaping par un correcteur structur (chapitre 4). Lal-

gorithme de minimisation de la norme H
y est dcrit. Il exploite le sous-direntiel de

Clarke, que lon peut ici dterminer de faon complte par des calculs dans le domaine fr-
quentiel. Ce mme algorithme a t gnralis an de lappliquer de faon analogue des
critres formuls dans le domaine temporel ; le chapitre 5 prsente ce point de vue uni,
et propose une srie dapplications en synthse structure. Lexpos se termine (chapitre
6) par une ouverture sur une toute autre approche pour la minimisation du critre H
:
elle consiste reformuler localement la synthse comme un problme direntiable, puis
rsoudre une suite de problmes quadratiques dits tangents. Une convergence plus
rapide en est attendue, et sera eectivement montre.
Les notations utilises dans la premire partie du document sont standard, et un
tableau les rcapitule dans les pages prcdentes. Elles sont gnralement suivies dans
les quatre chapitres de la deuxime partie, sauf lorsque le contexte nous a conduits
prfrer une notation spcique (en cas dambigut, par exemple) ; la nouvelle notation
choisie est alors explicite en dbut de chapitre. Enn, nous proposons au lecteur, en n
de document, les rfrences synthtiques des sources bibliographiques cites dans la thse.
8 Introduction
Premire partie
Notions fondamentales et cadre de
travail
9
Chapitre 1
Introduction la commande des
systmes
Ce premier chapitre prsente les problmes de commande qui sont abords dans le
cadre de nos travaux, prsents dans la deuxime partie de ce document. Nous nous
intressons plus particulirement aux problmes de stabilisation et de synthse H
.
1.1 Terminologie et dnitions
Nous rappelons ici les concepts fondamentaux de lAutomatique indispensables lana-
lyse et la commande des systmes dynamiques, et notamment des systmes linaires.
Nous invitons le lecteur se reporter aux ouvrages [2, 143] sil souhaite en approfondir
certains aspects.
1.1.1 Systmes dynamiques
Nous considrons un systme linaire et invariant dans le temps, en anglais linear and
time-invariant (LTI). Cest un processus liant des signaux dentre et de sortie, dcrit par
un systme dquations direntielles linaires coecients constants. Cette hypothse
conduit gnralement manipuler une approximation du systme rel considr ; les sys-
tmes rencontrs dans les applications industrielles comportent en eet des non-linarits,
ne serait-ce quen raison de contraintes physiques ou de scurit portant sur les action-
neurs ou sur les capteurs, qui se traduisent par des saturations en position ou en vitesse.
Il faut donc garder en mmoire que le domaine de validit dun systme linaire invariant
dans le temps est limit, par exemple, au voisinage dun point de fonctionnement
1
, une
bande de frquence, ou un intervalle de temps.
Dautre part, les algorithmes qui ont t dvelopps dans le cadre de la thse ne
concernent que les systmes LTI temps continu. Il ny a cependant aucun obstacle
thorique leur transposition aux systmes discrets.
1
On parle de point de vol en aronautique, dni par laltitude et la vitesse relative.
11
12 Chapitre 1. Introduction la commande des systmes
Reprsentation des systmes LTI
Matrice de transfert Un systme LTI entre et sortie scalaires (SISO, single-input
/ single-output) est compltement dcrit par sa rponse impulsionnelle g(t), cest--dire
la rponse de ce systme une impulsion de Dirac. En eet, puisque la rponse y(t)
du systme est solution dune quation direntielle linaire coecients constants, le
principe de superposition permet dexprimer y(t) pour tout signal dentre u(t) comme le
produit de convolution
y(t) =
_
+
g(t )u() d . (1.1)

Dans le domaine de Laplace, on obtient le produit simple
Y (s) = G(s).U(s) , (1.2)
o G est une fraction rationnelle, appele fonction de transfert du systme (s dsignant
la variable de Laplace).
Plus gnralement, dans le cas dun systme LTI entre ou sortie vectorielle (MIMO,
multi-input / multi-output), G devient une matrice de fractions rationnelles, appele ma-
trice de transfert. On dira que la matrice de transfert G est propre si
lim
[s[
|G(s)| < , (1.3)
et quelle est strictement propre si
lim
[s[
|G(s)| = 0 . (1.4)
Cette reprsentation des systmes par des matrices de transfert facilite leur inter-
connexion puisquune association en srie se traduit par un produit de matrices, une
association en parallle par une somme. Par ailleurs, elle permet ltude harmonique
des systmes, notamment SISO; en eet, si le signal dentre est de type sinusodal
u(t) = U
0
sin(
0
t), alors la sortie est galement sinusodale, de mme frquence, mais
dphase de = arg [G(j
0
)] et amplie de [G(j
0
)[. Des outils graphiques issus de
la commande classique (diagramme de Bode, lieu de Black/Nichols, lieu de Nyquist)
permettent danalyser les proprits de tels systmes, notamment en prsence dun asser-
vissement. Cependant, la reprsentation dun systme LTI par sa matrice de transfert ne
rend pas compte de la dynamique interne du systme.
Reprsentation dtat Ltude physique de nombreux systmes dynamiques temps
continu permet de dcrire leur comportement par un systme dquations de la forme
_
x(t) = f (x(t), u(t), t)
y(t) = g (x(t), u(t), t) .
(1.5)
Ici encore, le vecteur u(t) R
m
dsigne lentre (ou commande) du systme, et y(t) R
p
la sortie, issue de la seconde quation de (1.5) (dite quation dobservation). Le vecteur
x(t) R
n
, solution de la premire quation (quation direntielle ordinaire, dite qua-
tion dvolution), est appel vecteur dtat du systme. Il dcrit, un instant donn, la
1.1. Terminologie et dnitions 13
conguration interne du systme. Sous les hypothses prcites de linarit et dinvariance
dans le temps, f et g sont linaires et indpendantes de t, et les quations deviennent
_
x(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t) .
(1.6)
A R
nn
est appele matrice dtat du systme, B R
nm
matrice de commande,
C R
pn
matrice dobservation et D R
pm
matrice de transmission directe. Etant
donns la commande u(t) et ltat initial x(0) = x
0
, la sortie y(t) est donne par
y(t) = Ce
At
x
0
+
_
t
0
Ce
A(t)
Bu() d +Du(t) . (1.7)
Une simple transformation de Laplace des quations de (1.6), avec ltat initial x(0) =
0, permet dliminer X(s) et de formuler la matrice de transfert du systme
G(s) = C(sI A)
1
B +D. (1.8)
Le quadruplet (A, B, C, D) est appel ralisation de ce systme LTI. Cette ralisation
dpend du choix du vecteur dtat x(t), et tout changement de variable x(t) = T
1
x(t)
(avec la matrice de passage T (L
n
(R)) dans (1.6) introduit une nouvelle ralisation
(T
1
AT, T
1
B, CT, D) de ce mme systme. La matrice de transfert reste cependant in-
change, et dnit de faon unique le systme, du point de vue entre-sortie. Notons que
la matrice de transfert G est strictement propre si et seulement si D = 0.
Cette multiplicit des reprsentations en espace dtat amne en privilgier certaines.
Les raisons peuvent tre dordre numrique (reprsentation dtat quilibre par exemple,
voir [143]), ou encore parce que certaines formes de ralisations facilitent le passage de
la matrice de transfert une reprsentation dtat. Cest le cas notamment des formes
canoniques (forme diagonale ou de Jordan, et pour les systmes SISO, forme compagne
de commande ou dobservation). La dimension du vecteur dtat x(t) (appele ordre de la
ralisation (A, B, C, D)) nest pas non plus dnie de manire unique ; ainsi, parmi toutes
les ralisations (A, B, C, D) dune matrice de transfert G(s), celles pour lesquelles A a
la plus petite taille possible sont qualies de minimales. Lordre n associ est appel le
degr de McMillan de G(s).
Commandabilit, observabilit Etant donne une ralisation (A, B, C, D) dun
systme LTI, le couple (A, B) est dit commandable si, par dnition, pour tout t
1
> 0,
x
0
, x
1
R
n
, il existe une commande u : [0, t
1
] R
m
, intgrable sur [0, t
1
], qui amne la
variable dtat x(t) de x(0) = x
0
x(t
1
) = x
1
, autrement dit, telle que
x
1
= e
At
x
0
+
_
t
1
0
e
A(t
1
)
Bu() d . (1.9)
Le couple (C, A) est dit observable si, par dnition, pour tout t
1
> 0, ltat initial
x(0) peut tre dtermin de manire unique partir de toute commande u : [0, t
1
] R
m
et de la sortie y : [0, t
1
] R
p
qui en rsulte.
Comme il est dusage, on dira dans la suite que la ralisation (A, B, C, D) dun systme
est commandable (resp. observable) si (A, B) est commandable (resp. si (C, A) est obser-
vable).
Remarques:
On peut, sans amoindrir lnonc, remplacer x
0
par 0 dans la dnition de la com-
mandabilit.
De mme, lobservabilit peut tre dnie de faon quivalente en ne considrant
que la sortie y : [0, t
1
] R
p
rsultant de la commande nulle (y est alors la rponse
libre du systme).
Une fois dtermin lunique x(0) dune ralisation observable (A, B, C, D), toute
valeur prise par le vecteur dtat x(t) sur [0,
t] peut tre calcule avec

x(t) = e
At
x(0) +
_
t
0
e
A(t)
Bu() d . (1.10)
Il existe de nombreuses caractrisations algbriques de ces deux notions :
(A, B) est commandable si et seulement si la matrice de commandabilit
c
_
B AB A
2
B . . . A
n1
B
_
(1.11)
est de plein rang (n, en ligne).
(C, A) est observable si et seulement si la matrice dobservabilit
O
_
_
C
CA
CA
2
.
.
.
CA
n1
_
_
(1.12)
est de plein rang (n, en colonne).
On dispose aussi des caractrisations modales suivantes : (A, B) est commandable (resp.
(C, A) est observable) si et seulement si tout vecteur propre gauche v de A est tel que
v
B ,= 0 (resp. tout vecteur propre droite v de A est tel que Cv ,= 0).

Cela conduit dnir la commandabilit et lobservabilit par mode du systme, cest-
-dire par valeur propre de A, vis--vis de B et C : un mode (A) est commandable
si, par dnition, pour tout vecteur propre gauche (resp. droite) v associ la valeur
propre , on a v
B ,= 0 (resp. Cv ,= 0). Ainsi (A, B) est commandable (resp. (C, A)

est observable) si et seulement si tous les modes du systme sont commandables (resp.
observables).
Le critre de Popov-Belevitch-Hautus (ou PBH) traduit cette interprtation modale
par une condition de rang : (A, B) est commandable si et seulement si, pour toute valeur
propre (A) la matrice
_
A I
n
B
_
(1.13)
est de plein rang (n, en ligne). De mme, (C, A) est observable si et seulement si, pour
toute valeur propre (A) la matrice
_
A I
n
C
_
(1.14)
est de plein rang (n, en colonne).
Nous donnons une dernire caractrisation, qui permet de lier la commandabilit la
notion de placement de valeurs propres. En eet, (A, B) est commandable (resp. (C, A) est
observable) si et seulement si les valeurs propres de A + BF (resp. de A + LC) peuvent
tre places arbitrairement
2
dans C par un choix convenable de F R
mn
(resp. de
L R
np
).
Stabilisabilit, dtectabilit Les notions de stabilisabilit et de dtectabilit sont
voisines, mais moins fortes que les prcdentes : (A, B) est stabilisable si les modes instables
de A sont commandables, cest--dire, avec le critre PBH, si pour toute valeur propre
(A) C
+
la matrice (1.13) est de rang n. De faon duale, (C, A) est dtectable si
les modes instables de A sont observables : pour toute valeur propre (A) C
+
la
matrice (1.14) est de rang n.
On a enn des caractrisations par proprits de placement de valeurs propres : (A, B)
est stabilisable (resp. (C, A) dtectable) sil existe F R
mn
(resp. L R
np
) tel que
A+BF (resp. A+LC) est Hurwitz, cest--dire que ses valeurs propres sont toutes dans
le demi-plan gauche C
.
La commandabilit et lobservabilit, tout comme la stabilisabilit et la dtectabilit,
sont des notions despace dtat. Si elles sont invariantes par changement de variable,
elles dpendent cependant du choix de la dimension de la variable dtat. Ainsi, parmi
toutes les ralisations (A, B, C, D) dune matrice de transfert G(s), les ralisations la
fois commandables et observables sont exactement celles qui sont minimales.
Ples et zros dun systme LTI
Les ples (resp. les zros) dun systme SISO sont les racines du numrateur (resp. du
dnominateur) de sa fonction de transfert G(s), suppose irrductible. Pour un systme
MIMO, les ples sont les racines du polynme det(sI A), cest--dire les valeurs propres
de A, o (A, B, C, D) est une ralisation minimale de G(s). Plus gnralement, les ples
dune ralisation quelconque (A, B, C, D) de G(s) sont les valeurs propres de A. La d-
nition de zros dans le cas MIMO est plus dlicate que dans le cas SISO, et ncessite den
distinguer plusieurs types :
les valeurs s C telles que le rang de la matrice
_
A sI B
C D
_
(1.15)
est dcient sont les zros invariants (ou simplement les zros),
2
Sous rserve que toute valeur propre non relle soit couple avec sa conjugue.
les valeurs s C qui annulent compltement la matrice de transfert G(s) sont les
zros de blocage.
En particulier, les modes non observables ou non commandables sont des zros invariants.
1.1.2 Normes de signaux
On considre un signal vectoriel complexe temps continu u : R
+
C
m
. On dnit
la norme inni ou norme sup de u par
|u|
sup
t0
|u(t)|
. (1.16)
Si u est de carr sommable sur R
+
(cest--dire si u L
2
(R
+
), espace de Hilbert), on
dnit la norme
|u|
2

__

0
|u(t)|
2
2
dt
_1
2
, (1.17)
dnomme norme L
2
. Cest la norme induite par le produit scalaire sur L
2
(R
+
)
u, v)
_

0
u(t)
H
v(t) dt . (1.18)
Elle sinterprte comme lnergie du signal, et sexprime de faon quivalente dans le
domaine frquentiel
|u|
2
=
_
1
2
_

|U(j)|
2
2
d
_
1
2
, (1.19)
grce au thorme de Parseval. Lexpression droite de lgalit dnit une norme (encore
note | |
2
) sur lespace de Hardy H
2
, constitu des fonctions de la variable complexe U,
analytiques dans le demi-plan droit C
+
, et telles que
sup
>0
_
1
2
_

|U( +j)|
2
2
d
_
1
2
< . (1.20)
Cest, l encore, la norme induite par le produits scalaire sur H
2
(encore not , ))
U, V )
1
2
_

U(j)
H
V (j) d . (1.21)
Lespace H
2
est limage de L
2
(R
+
) par la transforme de Fourier, qui ralise ainsi un
isomorphisme despaces de Hilbert.
1.1.3 Stabilit des systmes linaires
La notion de stabilit est fondamentale pour les systmes dynamiques, quil sagisse de
commande en boucle ouverte ou en boucle ferme. Dnir la stabilit pour un systme LTI
dpend du point de vue adopt pour le reprsenter : dans le cas dune approche entre-
sortie, par matrice de transfert, on parle de stabilit BIBO (bounded input / bounded
output) ; si ltat x(t) du systme est connu, au moyen dune ralisation
_
x(t) = Ax(t) +Bu(t)
y(t) = Cx(t) + Du(t) .
(1.22)
on peut introduire la notion (plus forte) de stabilit interne.
Stabilit BIBO
Un systme (LTI ou autre) est BIBO-stable si, par dnition, pour toute entre u(t)
telle que |u|
< , la sortie y(t) vrie |y|
< . Dans le cas des systmes LTI,

la stabilit BIBO est caractrise par la localisation de tous les ples dans le demi-plan
gauche C
. Dun point de vue numrique, des critres algbriques (Routh-Hurwitz no-

tamment) permettent de vrier la stabilit BIBO dun systme partir des coecients
du dnominateur de sa fonction de transfert, sans avoir expliciter ses ples.
Ainsi la stabilit BIBO garantit que, sous laction dune commande borne pour la
norme inni, la rponse force du systme reste nie. Elle ne dit rien de lvolution du
systme lui-mme, dont la dynamique interne peut faire crotre considrablement certaines
grandeurs dtat, et engendrer par la mme laltration voire la destruction du systme.
La notion de stabilit interne permet de combler cette carence.
Stabilit interne
Le systme LTI (1.22), dni par la ralisation (A, B, C, D), est stable de faon interne
si, par dnition, pour tout tat initial x(0) = x
0
, et en labsence de signal de commande
(u(t) = 0, t 0), on a
lim
t
|x (t)| = 0 . (1.23)
Ltat dquilibre x
e
= 0 est dit dans ce cas asymptotiquement stable. On montre aisment
que la stabilit interne est quivalente la proprit de Hurwitz de la matrice dtat A
(toutes ses valeurs propres sont dans le demi-plan gauche C
).
La stabilit interne implique la stabilit BIBO de la matrice de transfert associe
G(s) = C(sI A)
1
B + D, puisque les ples du systme sont des valeurs propres de
A. La rciproque est fausse en gnral, sauf si les modes instables de A sont comman-
dables et observables (cest--dire si (A, B) est stabilisable et (C, A) est dtectable). On a
alors quivalence des deux notions de stabilit. Cest le cas en particulier si la ralisation
(A, B, C, D) est minimale, les ples de G tant alors exactement les valeurs propres de A.
An de caractriser la stabilit interne de faon analytique, on dnit, pour toute
matrice M R
nn
, labscisse spectrale de M, note (M), comme la plus grande partie
relle des valeurs propres de M :
(M) max
(M)
Re () . (1.24)
La stabilit interne du systme (1.22) est donc quivalente lingalit (A) < 0. Si la
matrice dtat dpend de paramtres (A = A()), la recherche de valeurs de paramtres
rendant le systme stable peut ainsi tre ramene un problme de minimisation de
labscisse spectrale.
Une autre approche consiste caractriser la stabilit interne par la condition de
Lyapunov, exprime en terme dingalit linaire matricielle (ou LMI, Linear Matrix In-
equality) :
Thorme 1.1.1 Le systme linaire (1.22) est stable de faon interne si et seulement
si il existe une matrice P o
n
(R), P ~ 0, tel le que
A
T
P +PA 0 . (1.25)
Nous dresserons un inventaire des techniques de stabilisation en boucle ferme dans
la section 1.2.2 ci-aprs, ainsi quau chapitre 3.
1.1.4 Normes de systmes
La conception dune loi de commande doit gnralement rpondre, au del de lexigence
de stabilit, des spcications relatives la taille des signaux ou des matrices de
transfert : on cherchera par exemple rduire limpact sur le comportement du systme
des perturbations pouvant aecter la commande ou la sortie, ou encore minimiser leet
du bruit apparaissant sur les mesures. On pourra aussi exiger de maintenir la puissance
dlivre par les actionneurs en de dune certaine limite. Ces objectifs ou contraintes de
performance sont exprimes au moyen de normes de systmes.
Norme H
Soit G la matrice de transfert dun systme LTI, suppos stable. Cest une fonction
rationnelle relle de la variable complexe s, comme fraction de polynmes coecients
rels. Elle est propre, analytique et borne dans le demi-plan droit C
+
. On dnit sa
norme H
par
|G|
sup
Re s>0
[G(s)] (1.26)
= sup
R
[G(j)]
= sup
R
+
[G(j)] .
o [G(s)] dsigne la plus grande valeur singulire de G(s).
On montre que lespace H
2
(dni en 1.1.2) est invariant par loprateur linaire M
G
:
U GU, et que la norme H
de G concide avec la norme de M

G
en tant quoprateur
linaire H
2
H
2
. |G|
est donc la norme induite par la norme L

2
|G|
= sup
|U|
2
<
U,=0
|GU|
2
|U|
2
(1.27)
et mesure le gain maximal en nergie du systme.
La norme H
nest pas dnie dans le cas dun systme LTI instable (appliquer la
dnition (1.26) aboutirait une valeur innie). Cependant, pour un systme LTI sans
ples sur laxe imaginaire, de matrice de transfert G, la quantit sup [G(j)] / R
est nie. Cest la norme L
de G, qui concide avec la norme H
ds que G est stable.

On introduit enn la norme H
dite a-shifte (a-shifted H
norm en anglais). Pour

a R x, elle est dnie par
|G|
,a
sup
Re s>a
[G(s)] = |G
a
|
, (1.28)
o G est la matrice de transfert dun systme LTI, et G
a
(s) = G(sa), ce qui a pour eet
de translater de a les ples du systme dans le plan complexe. La norme H
a-shifte est
1.2. Problmes de synthse pour la commande 19
nie si et seulement les ples de G sont partie relle strictement infrieure a. Ainsi
en choisissant a < 0 susamment grand en valeur absolue, la norme shifte permet de
mesurer le gain dun systme instable.
Norme H
2
La norme H
2
est dnie pour toute matrice de transfert stable strictement propre G.
Elle vaut
|G|
2

_
1
2
_

Tr
_
G(j)
H
G(j)
_
d
_
1
2
(1.29)
= sup
|U|
<
U,=0
|GU|
2
|U|
. (1.30)
avec |U|
= sup
Re s>0
|U(s)| = sup
R
|U(j)|.
La norme H
2
mesure lnergie en sortie du systme, lorsquun bruit blanc normalis
(cest--dire de densit spectrale uniforme S
uu
(j) = U(j)U
H
(j) = I) est appliqu en
entre.
1.2 Problmes de synthse pour la commande
La commande
3
des systmes dynamiques ne recouvre pas tout fait les mmes notions
suivant que lon se place du point de vue du mathmaticien ou de celui de lautomaticien,
comme le souligne Bergounioux dans [21]. Le but poursuivi est le mme : trouver une
expression du signal de commande u(t) appliquer en entre du systme an que ce dernier
vrie un ensemble de spcications. Cela peut tre par exemple un critre minimiser
(comme la norme H
2
dun transfert par exemple), on parle alors de commande optimale. Il
peut sagir simplement de trouver une commande qui stabilise le systme, ou, dans le cas
dun systme linaire, qui permet de placer les ples dans le plan complexe (commande
modale). Le mathmaticien va rechercher une telle commande dans un espace fonctionnel
donn, dont les lments ne dpendent pas a priori de la variable dtat x du systme,
ou de la sortie y. Cest ce que lautomaticien nomme commande en boucle ouverte, car le
systme est pilot en aveugle par un signal prcalcul. La commande en boucle ferme
est, au contraire, construite en temps-rel, partir des signaux disponibles en sortie du
systme et mesurs par des capteurs. Les signaux de commande sont alors appliqus en
entre via des actionneurs. Le processus de fabrication du signal de commande dnit un
systme dynamique en tant que tel, que lon appelle correcteur, compensateur, ou encore
loi de commande. Nous nous limiterons dans le cadre de notre tude aux correcteurs de
mme type que les systmes considrs, cest--dire linaires et invariants dans le temps
(LTI).
3
on trouve parfois le terme contrle, ou contrle actif, par opposition au contrle passif qui implique
une modication de la structure du systme.
1.2.1 Forme standard pour la commande
Nous introduisons tout dabord une formulation gnrique dite standard, reprsenta-
tion qui donne un cadre trs gnral pour la synthse de lois de commande. Elle facilite
lexplicitation du cahier des charges en isolant, parmi les entres et sorties du systme,
celles qui interviennent dans les spcications.
Parmi les entres du systme command (not P(s) sur la gure 1.1), on distingue celles
produites par le correcteur K(s) de celles dites exognes. Ces dernires sont regroupes
dans le vecteur w R
m
1
et comprennent les perturbations (rafales de vent pour un
aronef), les bruits, ou encore les signaux de consigne. Finalement, lentre u R
m
2
ne contient que les signaux des actionneurs gnrs par le correcteur, cest--dire les
commandes. De mme, parmi toutes les sorties du systme, celles qui sont accessibles au
correcteur pour llaboration des ces commandes sont dsignes par le vecteur y R
p
2
(ce
sont les mesures), tandis que z R
p
1
contient les sorties rgules, sur lesquelles portent
les spcications.
sorties mesures commandes
sorties rgules entres exognes
y
z w
u
K(s)
P(s)
Fig. 1.1 La forme standard.
La forme standard permet donc une description de direntes congurations de boucles
fermes (boucle de suivi, boucle de rgulation, asservissement un ou plusieurs degrs de
libert) laide dune reprsentation unique.
Transformation linaire fractionnaire
La matrice de transfert P(s), des signaux dentre w et u vers les signaux de sortie z
et y, est partitionne en quatre blocs, associs chacun des quatre canaux du systme,
de la faon suivante :
_
Z(s)
Y (s)
_
=
_
P
11
(s) P
12
(s)
P
21
(s) P
22
(s)
_ _
W(s)
U(s)
_
. (1.31)
Linterconnexion en boucle ferme du systme et du correcteur, reprsents par leurs
matrices de transfert respectives P et K (supposes propres), constitue un nouveau sys-
tme LTI, dentre w et de sortie z. La matrice de transfert en boucle ferme T
wz
est
T
wz
(s) = P
11
(s) + P
12
(s)K(s)(I P
22
(s)K(s))
1
P
21
(s) , (1.32)
appele transformation linaire fractionnaire (LFT) ou produit de Redheer de P et K.
Le systme boucl est dit bien pos si, par dnition, cette matrice de transfert existe
et est propre, ou, de faon quivalente, ds que I P
22
()K() est inversible. Cette
hypothse de travail, qui revient sassurer que la boucle ferme possde bien un sens
physique, sera toujours suppose vraie par la suite.
Reprsentation dtat
La formulation LFT et lapproche par variable dtat orent un cadre mthodologique
riche pour la rsolution de problmes de commande, notamment dans le cas de la synthse
H
. Nous reprenons donc ici les notations usuelles pour la forme standard exprime en
espace dtat. Supposons que les sorties z et y sont lies aux entres w et u par la ralisation
_
_
x(t) = Ax(t) + B
1
w(t) + B
2
u(t)
z(t) = C
1
x(t) + D
11
w(t) + D
12
u(t)
y(t) = C
2
x(t) + D
12
w(t) + D
22
u(t) ,
(1.33)
avec x R
n
, et la condition initiale x(0) = 0, de sorte que
P(s) =
_
P
11
(s) P
12
(s)
P
21
(s) P
22
(s)
_
=
_
C
1
C
2
_
(sI A)
1
_
B
1
B
2
_
+
_
D
11
D
12
D
21
D
22
_
. (1.34)
De mme, soit
_
x
K
(t) = A
K
x
K
(t) + B
K
y(t)
u(t) = C
K
x
K
(t) + D
K
y(t)
(1.35)
une ralisation du correcteur, de variable dtat x
K
R
n
K
(avec x
K
(0) = 0), telle que
K(s) = C
K
(sI A
K
)
1
B
K
+D
K
. (1.36)
An dallger la formulation en espace dtat de la forme standard (gure 1.1), nous
faisons lhypothse que le bloc P
22
de la matrice de transfert P est strictement propre, ce
qui revient poser D
22
= 0 (pas de transmission directe de u vers y). Cette condition est
trs souvent vrie dans les applications, et peut se poser sans perte de gnralit dans
les problmes de synthse de correcteur. En eet, si D
22
,= 0, il sut de dporter le bloc
de transmission directe dans le correcteur, en posant
y(t) C
2
x(t) +D
21
w(t) (1.37)
la place de la dernire galit dans (1.33) et, par suite, de remplacer (1.35) par
_
_
_
x
K
(t) =
_
A
K
+B
K
D
22
(I D
K
D
22
)
1
C
K
_
x
K
(t) + B
K
(I D
22
D
K
)
1
y(t)
u(t) = (I D
K
D
22
)
1
C
K
x
K
(t) + (I D
K
D
22
)
1
D
K
y(t) .
(1.38)
I D
K
D
22
et I D
22
D
K
sont inversibles puisque la boucle ferme est bien pose, par
hypothse. Ainsi un simple changement de variable au niveau du correcteur permet de
se ramener au cas D
22
= 0, sans consquence sur le transfert T
wz
. On obtient alors, en
liminant u et y, une ralisation de la boucle ferme
_
_
x(t) = (A +B
2
D
K
C
2
)x(t) + B
2
C
K
x
K
(t) + (B
1
+B
2
D
K
D
21
)w(t)
x
K
(t) = B
K
C
2
x(t) + A
K
x
K
(t) +B
K
D
21
w(t)
z(t) = (C
1
+D
12
D
K
C
2
)x(t) + D
12
C
K
x
K
(t) + (D
11
+D
12
D
K
D
21
)w(t) ,
(1.39)
de sorte que
T
wz
(s) = C
c
(K) [sI A
c
(K)]
1
B
c
(K) + D
c
(K) , (1.40)
avec
A
c
(K) =
_
A +B
2
D
K
C
2
B
2
C
K
B
K
C
2
A
K
_
, B
c
(K) =
_
B
1
+B
2
D
K
D
21
B
K
D
21
_
,
C
c
(K) =
_
C
1
+D
12
D
K
C
2
D
12
C
K
_
, D
c
(K) =
_
D
11
+D
12
D
K
D
21
_
.
(1.41)
Nous ferons donc toujours lhypothse D
22
= 0 dans notre travail.
Retour statique, retour dynamique
Les contraintes dimplmentation des correcteurs, en termes de puissance de calcul
disponible, de mmoire, mais aussi de cot ou de abilit, font que les lois de commande
simples sont privilgies lors de la conception.
La commande par retour statique de sortie (SOF en anglais, pour static output feed-
back) apparat comme un cas particulier de la commande dnie par (1.35), avec n
K
= 0.
Le correcteur qui la ralise est rduit un simple gain matriciel y(t) = D
K
u(t) = Ku(t).
La synthse dun correcteur dynamique, dordre x n
K
1, permet daccder des lois
de commande plus labores. Lorsque n
K
< n, on parle dordre rduit, et pour n
K
= n
dordre plein. Notons que la commande par retour dynamique de sortie (DOF en anglais),
dordre x, se ramne formellement laction dun correcteur statique

K, constitu des
blocs A
K
, B
K
, C
K
et D
K
, par augmentation de x, u et y (voir par exemple [95]). Plus
prcisment, en posant
x =
_
x
x
K
_
, y =
_
x
K
y
_
, u =
_
x
K
u
_
, (1.42)
nous obtenons la ralisation
A =
_
A 0
0 0
n
K
_
,

B
1
=
_
B
1
0
_
,

B
2
=
_
0 B
2
I
n
K
0
_
,
C
1
=
_
C
1
0
_
,

D
11
= D
11
,

D
12
=
_
0 D
12
_
,
C
2
=
_
0 I
n
K
C
2
0
_
,

D
21
=
_
0
D
21
_
,

D
22
= 0 .
(1.43)
Le systme ainsi augment, command par
u =
_
A
K
B
K
C
K
D
K
_
. .
K
y (1.44)
concide avec la boucle ferme obtenue par retour dynamique. Dans le cas de la commande
par un correcteur dordre x, nous pourrons donc supposer sans perte de gnralit et
aprs augmentation du systme commander que ce correcteur est statique.
1.2.2 Stabilisation
La stabilisation dun systme LTI par retour statique de sortie consiste rechercher,
sil en existe, un correcteur K R
m
2
p
2
tel que la boucle ferme soit stable de manire
interne, cest--dire tel que
(A +B
2
KC
2
) C
. (1.45)
En utilisant labscisse spectrale dnie par (1.24), cela revient chercher K tel que
(A +B
2
KC
2
) < 0.
Lorsque le correcteur recherch est dynamique, il sagit de dterminer, toujours sous
condition dexistence, des matrices A
K
, B
K
, C
K
et D
K
telles que
(A
c
(K)) C
avec A
c
(K) =
_
A +B
2
D
K
C
2
B
2
C
K
B
K
C
2
A
K
_
. (1.46)
La stabilisation par retour de sortie demeure un problme ouvert de la commande
[22]. Il apparat comme le plus souvent cit dans lenqute [23] compilant les points de
vue dautomaticiens sur les enjeux de leur discipline. Il a t, et est toujours, lobjet de pu-
blications trs abondantes, dont le lecteur trouvera une slection de rfrences dans [100].
La plupart des approches existant pour trouver un correcteur K stabilisant sont prsen-
tes dans [131]. Dans certains cas particuliers, lexistence dun correcteur K stabilisant
(la stabilisabilit) peut tre vrie par des critres algbriques bien connus [143] :
Si n
K
= 0 et C = I (retour dtat statique), il faut et il sut que le couple
(A, B
2
) soit stabilisable (par dnition mme), ce qui revient vrier que les modes
instables de A sont commandables, soit encore [A I
n
B
2
] de plein rang (en ligne)
pour toute valeur propre (A) C
+
(critre PBH).
Si n
K
= n (retour de sortie dordre plein), une condition ncessaire et susante est
(A, B
2
) stabilisable et (C
2
, A) dtectable, cest--dire que les modes instables de A
sont commandables et observables. Elle peut tre vrie par le critre PBH.
Si n
K
< n (retour de sortie dordre rduit), cette dernire condition nest que n-
cessaire.
Si n
K
= 0 (retour de sortie statique), la condition suivante dite PIP paire (pour
parity interlacing property) portant sur les ples et les zros de G(s) = C
2
(sI
A)
1
B
2
est une condition ncessaire de stabilisabilit [131] :
le nombre de ples rels de G (compts avec leur multiplicit) entre deux zros
de blocage rels de G positifs ou nuls est pair, et
le nombre de zros de blocage rels de G entre deux ples rels de G est pair.
Ces conditions dexistence sont dun faible intrt pratique car elles ne sappliquent
qu un nombre limit de problmes de stabilisabilit, et surtout parce quelles ne sont
pas constructives. Dautre part, la complexit du problme de stabilisabilit par retour
de sortie statique nest pas tablie, mais des problmes trs voisins (stabilisabilit par
retour dtat avec correcteur born, stabilisation simultane par retour de sortie, . . .) sont
clairement NP-diciles [24].
Des mthodes constructives sont cependant bien connues dans le cas des systmes
SISO, pour lesquels K est un scalaire : ltude graphique du lieu des ples de la fonction de
transfert en boucle ferme (dit lieu dEvans) permet de rpondre la question dexistence
dun correcteur statique stabilisant, et donne le cas chant une valeur numrique pour
K. Il en est de mme pour le critre algbrique de Routh-Hurwitz.
Dans le cas des systmes MIMO, la tche de construction dun correcteur stabilisant
est plus complexe, et direntes mthodes ont t proposes, faisant appel des domaines
des mathmatiques trs divers. Nous en donnons ci-aprs un aperu.
Mthodes de dcision
Il est toujours possible de ramener le problme de stabilit de la boucle ferme un
ensemble dinquations scalaires polynmiales, dont les inconnues sont les lments de
K. Cela peut tre ralis au moyen des dterminants de Hurwitz associs det(sI
A B
2
KC
2
), qui sont tous strictement positifs ds que les ples sont dans le demi-plan
gauche C
(critre de Routh-Hurwitz), ou encore par le test de Linard-Chipart, qui

ncessite environ moiti moins dingalits avec dterminants que Routh-Hurwitz. Lide
matresse prsente dans [3] consiste considrer cet ensemble dingalits, munies de
quanticateurs, comme une proposition logique, puis utiliser une mthode de dcision
(algorithmes de Tarski et de Seidenberg) an de rduire le nombre de quanticateurs et
dcider, au bout dun nombre ni dtapes, si la proposition est vraie ou fausse, cest--dire
si le systme est stabilisable ou non. Si cest eectivement le cas, alors la mthode permet
aussi de construire un correcteur stabilisant. Un exemple avec n = 3, m
2
= 1 et p
2
= 2 est
dtaill et comment dans [3]. Mme avec ces dimensions rduites (le critre de stabilit
ne comporte que trois ingalits), le nombre dtapes et de manipulations ncessaires
la mise en uvre de lalgorithme de Tarski est important. Ainsi, si cette approche des
problmes de stabilit par la dcidabilit est sduisante, elle se heurte cependant la
complexit exponentielle des algorithmes dlimination de quanticateurs existants [20].
Ingalits linaires matricielles
La thorie de Lyapunov permet de ramener le problme de stabilisation par retour
de sortie statique la rsolution dune ingalit matricielle de la forme de (1.25). Ainsi,
le systme en boucle ferme est stable de manire interne si et seulement si il existe une
matrice K R
m
2
p
2
et une matrice P o
n
(R), P ~ 0, telles que
(A +B
2
KC
2
)
T
P +P(A +B
2
KC
2
) 0 . (1.47)
Une telle inquation matricielle est dicile rsoudre, parce que les inconnues (P, K)
ninterviennent pas de faon linaire dans (1.47) ; notons cependant que dans le cas du
retour dtat (C
2
= I
n
), le changement de variables Q = P
1
et Y = KQ permet de
ramener (1.47) la LMI suivante, linaire en (Q, Y )
QA
T
+AQ+Y
T
B
T
2
+B
2
Y 0 (1.48)
avec Q ~ 0. Le correcteur est ensuite reconstruit avec K = Y Q
1
. La complexit des
problmes LMI est polynmiale en le nombre de variables et il existe des mthodes nu-
mriques performantes pour les rsoudre (mthodes de points intrieurs, algorithme de
lellipsode, optimisation non lisse, . . .).
Revenons au cas gnral du retour de sortie (1.47) ; il sagit dune ingalit bilinaire
matricielle (ou BMI, Bilinear Matrix Inequality), dont la rsolution est un problme NP-
dicile [135]. Les mthodes numriques de stabilisation bases sur la thorie de Lyapunov
sappuient donc plutt sur le thorme suivant [75].
Thorme 1.2.1 Le systme (1.33) admet un correcteur statique stabilisant (par retour
de sortie) si et seulement si il existe des matrices X, Y o
n
(R), X ~ 0, Y ~ 0, tel les
que
W
T
B
T
2
(AX +XA
T
)W
B
T
2
0, (1.49)
W
T
C
2
(A
T
Y +Y A)W
C
2
0, (1.50)
XY = I
n
, (1.51)
o W
B
T
2
et W
C
2
sont des matrices dont les colonnes forment respectivement des bases des
noyaux de B
T
2
et de C
2
.
Lensemble des correcteurs statiques stabilisants K associs la matrice de Lyapunov
P = Y , cest--dire qui vrient (A + B
2
KC
2
)
T
Y + Y (A + B
2
KC
2
) 0, est alors dcrit
par les matrices de la forme
K = R
1
B
T
2
Y Q
1
C
T
2
(C
2
Q
1
C
T
2
)
1
+S
1
2
L(C
2
Q
1
C
T
2
)
1
2
, (1.52)
o S o
m
2
(R), S ~ 0, est dnie par
S R
1
R
1
B
T
2
Y Q
1
_
QC
T
2
_
C
2
Q
1
C
T
2
_
1
C
2
_
1
Q
1
Y B
2
R
1
, (1.53)
et R o
m
2
(R), R ~ 0, est choisie tel le que
R
1
~ B
+
2
_
W
B
T
2
_
W
T
B
T
2
W
B
T
2
_
1
W
T
B
T
2

_
(B
+
2
)
T
, (1.54)
= XA
T
+AX. (1.55)
Enn, L R
m
2
p
2
doit vrier |L|
2
< 1, et Q o
n
(R), Q ~ 0, est dtermine par
Q = Y B
2
R
1
B
T
2
Y Y A A
T
Y. (1.56)
Bien que (1.49) et (1.50) soient des LMI, la contrainte de couplage (1.51) est non
convexe et rend la rsolution en (X, Y ) dicile. La complexit en est dailleurs connue :
le problme est NP-dicile [52]. Il a donn lieu de nombreux algorithmes heuristiques,
dont la convergence vers une solution de (1.49)-(1.51) nest pas garantie (min/max [58],
projections alternes [62], par exemple). Les mthodes de type Lagrangien augment par-
tiel [49, 15], avec contraintes LMI explicites, prennent en compte la contrainte non linaire
(1.51) et sont assorties dune preuve de convergence [106] vers une solution locale. Leur
mise en uvre est cependant coteuse en temps de calcul.
Ds quune solution de (1.49)-(1.51) est connue, un correcteur statique stabilisant K
se dduit explicitement avec (1.52), aprs avoir x les paramtres R et L. Une alternative
consiste rsoudre lingalit matricielle (A+B
2
KC
2
)
T
Y +Y (A+B
2
KC
2
) 0, devenue
une LMI en K puisque la variable de Lyapunov Y est maintenant xe.
Optimisation non linaire
Nous lavons dj not, la stabilisation du systme (1.33) par un correcteur statique
peut tre assimile un problme doptimisation, en introduisant labscisse spectrale de la
matrice dtat en boucle ferme A
c
(K) = A+B
2
KC
2
. Nous considrons donc le problme
doptimisation
min
KR
m
2
p
2
(A +B
2
KC
2
). (1.57)
Lapplication K A + B
2
KC
2
est ane, mais na pas de bonnes proprits
pour loptimisation : elle est continue, mais non convexe, et surtout non lisse, et mme
non lipschitzienne en gnral. Une mthode numrique inspire des mthodes de faisceaux
(gradient bundle methods, prsentes dans le chapitre suivant) est introduite dans [35] pour
rsoudre (1.57). Elle comporte cependant, chaque itration, une tape dchantillonnage
autour du correcteur courant, ce qui lui donne un caractre alatoire et ne permet pas la
reproductibilit de la squence doptimisation. Cest pourquoi nous proposons au chapitre
3 dappliquer un algorithme dterministe de minimisation non lisse pour rsoudre (1.57).
Dautres fonctions critres sont proposes dans [36]. Elles sont introduites partir
du pseudo-spectre
, partie ferme de C dnie pour 0 x et pour toute matrice

X C
nn
par
(X)
_
z C : Y C
nn
det (Y zI
n
) = 0 et |Y X|
2

_
. (1.58)
On dnit ainsi
labscisse pseudo-spectrale
(X), plus grande partie relle des lments de
(X) :
(X) max Re z : z
(X) , (1.59)
la distance linstabilit (X), dnie comme la distance de X lensemble (ferm)
des matrices non Hurwitz de C
nn
:
(X) min
_
|Y X|
2
: Y C
nn
, (Y ) 0
_
. (1.60)
Ces deux quantits sont lies entre elles par la relation
(X) < 0 (X) > . (1.61)

De plus, si on pose G(s) (sI X)
1
, et si X est Hurwitz, on a
(X) = |G|
1
. (1.62)
La minimisation de
(A
c
(K)) (avec > 0 x), permet de rechercher un correcteur
stabilisant, avec de surcrot une proprit de robustesse [37]. Le problme demeure non
convexe et non lisse, mais
(avec > 0 x) est plus rgulire que : condition

que les valeurs propres de A
c
(K) telles que Re = (A
c
(K)) (valeurs propres dite
actives) soient non drogatoires
4
, labscisse pseudo-spectrale
est localement lipschit-

zienne en A
c
(K) (voir [37]). Ce rsultat permet demployer directement des algorithmes
doptimisation non direntiable de type faisceau [80]. Les valuations de
(A
c
(K)) ou de
(A
c
(K)) sont toutefois coteuses en temps de calcul : les mthodes numriques proposes
dans [38, 99] sappuient, comme pour la norme H
(voir [28]), sur des calculs rpts de

valeurs propres de matrices hamiltoniennes. Pour labscisse spectrale (A
c
(K)), un simple
calcul de valeurs propres sut.
1.2.3 Synthse H
Etant donn un systme LTI sous forme standard (1.33), le problme H
consiste
rechercher une loi de commande par retour de sortie U(s) = K(s)Y (s), qui stabilise la
boucle ferme et minimise la norme H
du transfert T
wz
(K) :
min
K(s) stabilise (1.33)
|T
wz
(K)|
. (1.63)
Rappelons que la norme H
est la norme induite sur lespace des matrices de trans-

fert stables par la norme de signal L
2
(qui mesure lnergie). Le but recherch est ainsi
de rduire le plus possible leet des entres exognes w sur les sorties rgules z. Le
problme doptimisation (1.63) nest pas convexe, du fait que la transformation linaire
fractionnaire K T
wz
(K) nest pas linaire. Il est de plus non lisse, car la norme
H
cache un double maximum portant sur les valeurs singulires

i
(T
wz
(K, j)), en
i 1, 2, . . . , min(m
1
, p
1
) et en [0, ].
Il existe essentiellement deux approches pour rsoudre le problme H
. Elles ont en
commun de sappuyer sur des caractrisations de lensemble des correcteurs stabilisants -
sous-optimaux, cest--dire tels que |T
wz
(K)|
< pour une performance > 0 xe.

La plus classique sappuie sur des quations algbriques de Riccati ; publie en 1989 dans
[47] et rappele ci-aprs, elle est souvent dsigne par les initiales DGKF, formes partir
des noms des quatre auteurs (Doyle, Glover, Khargonekar et Francis). La seconde, plus
rcente (Gahinet et Apkarian, 1994 [55]), utilise des ingalits linaires matricielles. Les
correcteurs synthtiss par ces deux approches sont gnralement dordre plein, et donc
peu adapts une implmentation directe. Ceci dautant plus que lordre du systme
mis sous forme standard (1.33) est souvent suprieur lordre du systme commander,
car la forme standard peut inclure des matrices de pondration dynamique destines
4
Une valeur propre (X) est dite non drogatoire si le sous-espace propre associ est de
dimension 1, ou, de faon quivalente, si le bloc associ dans la forme de Jordan de X ne comporte
que des 1 sur la sur-diagonale.
intgrer les spcications frquentielles dans le critre H
(loop-shaping). Il est bien sr

possible de procder une rduction dordre, soit a priori, directement sur le systme
sous forme standard, soit a posteriori, sur le correcteur H
dordre plein. Quelle que soit

la mthode de rduction de modle choisie, parmi les nombreuses existantes (troncature
modale, troncature quilibre, . . .voir [112] pour un inventaire non exhaustif), loptimalit
est perdue : rien ne garantit quun autre correcteur du mme ordre que le correcteur rduit
ne ralise pas localement une meilleur performance.
Equations de Riccati
Lapproche DGKF de [47] ore, sous certaines hypothses, une caractrisation et une
paramtrisation en espace dtat de lensemble des correcteurs -sous-optimaux pour le
problme H
, cest--dire tels que, tant donne une performance > 0,

K stabilise (1.33) de manire interne et |T
wz
(K)|
< (1.64)
Les conditions de validit de cette premire approche sont les suivantes :
(i) (A, B
2
) est stabilisable et (C
2
, A) est dtectable,
(ii) D
12
et D
21
sont de plein rang,
(iii) P
12
et P
21
(transferts partiels dnis par 1.31) nont pas de zro sur laxe imaginaire,
(iv) D
T
12
_
C
1
D
12
_
=
_
0 I
_
et
_
B
1
D
21
_
D
T
21
=
_
0
I
_
,
(v) D
11
= 0 et D
22
= 0.
Il est possible de satisfaire les hypothses (iv) et (v), dites de normalisation, par des
changements de variables. Les manipulations, dont une partie a dj t voque en (1.37)
et (1.38) pour annuler D
22
, sont dtailles dans [125]. Notons cependant quil est ncessaire
de supposer (ii) an de pouvoir se ramener D
21
D
T
21
= I et D
T
12
D
12
= I. (ii) et (iii) sont
des hypothses de rgularit du systme. Enn, (i) est une condition ncessaire et susante
pour la stabilisabilit du systme par un correcteur dordre plein.
Thorme 1.2.2 Etant donn > 0, et sous les hypothses (i)-(v), le systme (1.33)
admet un correcteur -sous-optimal pour le problme H
(1.64) si et seulement si il existe

des solutions stabilisantes X
, Y
o
n
(R) aux deux quations algbriques de Riccati
suivantes :
A
T
X
+X
A +X
(
2
B
1
B
T
1
B
2
B
T
2
)X
+C
T
1
C
1
= 0, (1.65)
AY
+Y
A
T
+Y
(
2
C
T
1
C
1
C
T
2
C
2
)Y
+B
1
B
T
1
= 0, (1.66)
et si, de plus, ces solutions vrient
X
_ 0, Y
_ 0, (X
) <
2
. (1.67)
Une ralisation dun correcteur -sous-optimal particulier (appel correcteur central) est
alors donne par
_
_
A
K
= A + (
2
B
1
B
T
1
B
2
B
T
2
)X
(I
2
Y
)
1
Y
C
T
2
C
2
B
K
= (I
2
Y
)
1
Y
C
T
2
C
K
= B
T
2
X
D
K
= 0 .
(1.68)
Cette caractrisation permet, en procdant par dichotomie sur , de trouver un en-
cadrement du gain H
-optimal
opt
, et ceci la prcision souhaite : cest lalgorithme
de -itration. Le correcteur central (strictement propre et dordre plein) est reconstruit
avec (1.68).
Notons daprs [47] que lensemble des correcteurs -sous-optimaux K
, ralisant une
performance |T
wz
(K
)| < , est entirement dcrit sous la forme dune LFT dans

laquelle intervient un paramtre Q(s).
K
(s)
K
0
(s)
w z
y u
Q(s)
P(s)
Fig. 1.2 Paramtrisation des correcteurs sous-optimaux K
.
Le bloc K
0
est bas sur le correcteur central (1.68), augment de la faon suivante :
K
0
(s) =
_
_
A
K
B
K
(I
2
Y
)
1
B
2
C
K
D
K
I
C
2
I 0
_
_ . (1.69)
Le paramtre Q(s), qui doit tre choisi stable et tel que |Q|
< , permet de jouer

sur la nature du correcteur H
: il est ainsi possible, comme indiqu dans [90], de choisir

Q de faon former un correcteur H
dordre n
K
= n p
2
. En dehors de quelques cas
particuliers cependant, la faon dont le paramtre Q modie lordre ou la structure du
correcteur H
nest pas matrise. Dans le cas de lordre rduit, seules des conditions
susantes ont t proposes [63]. Elles font apparatre des quations de Riccati modies
et couples, rsolues par des algorithmes de continuation homotopique.
Ainsi, le problme H
se prte bien une rsolution par quations de Riccati dans le

cas de la synthse dordre plein, sans structure particulire sur le correcteur. Lapproche
est cependant limite par les hypothses de rgularit (ii) et (iii), ainsi que par lordre
potentiellement lev du correcteur central, dfaut de savoir exploiter la diversit des
correcteurs H
. La formulation du problme H
par des ingalits matricielles permet

de lever certaines de ces limitations.
Ingalits linaires matricielles
La caractrisation des correcteurs H
sous-optimaux par des ingalits matricielles

repose sur le lemme born-rel [4], qui est au critre H
ce que le thorme de Lyapunov

(Thorme 1.1.1) est la stabilit interne :
Lemme 1.2.3 Le systme LTI dni par la ralisation (A, B, C, D) est stable de manire
interne et vrie la condition de performance
_
_
_C(sI A)
1
B +D
_
_
_
< (1.70)
si et seulement si il existe une matrice X o
n
(R), X ~ 0, tel le que
_
_
A
T
X +XA XB C
T
B
T
X I D
T
C D I
_
_ 0 . (1.71)
Lapplication du lemme born-rel la ralisation
_
A
c
(K), B
c
(K), C
c
(K), D
c
(K)
_
, qui
dcrit la boucle ferme sous forme standard ((1.41), page 22) donne ainsi une caractrisa-
tion des correcteurs H
sous-optimaux par une ingalit matricielle en X o

n+n
K
(R) et
K R
(n
K
+m
2
)(n
K
+p
2
)
. Les hypothses sont ici beaucoup plus faibles que pour lapproche
DGKF, puisquon supposera seulement :
(i) (A, B
2
) est stabilisable et (C
2
, A) est dtectable,
(ii) D
22
= 0.
Ainsi le problme de synthse H
se ramne un problme de programmation semi-

dnie (SDP), dont lobjet est de minimiser le critre sous les contraintes matricielles
_
_
(A +B
2
KC
2
)
T
X +X(A +B
2
KC
2
) X(B
1
+B
2
KD
21
) (C
1
+D
12
KC
2
)
T
(B
1
+B
2
KD
21
)
T
X I (D
11
+D
12
KD
21
)
T
C
1
+D
12
KC
2
D
11
+D
12
KD
21
I
_
_ 0 ,
(1.72)
P ~ 0 , (1.73)
o les matrices A, B
1
, B
2
, C
1
, C
2
, D
11
, D
12
, D
21
et K sont augmentes conformment
(1.43)-(1.44), dans le cas dun retour dynamique (n
K
> 0). Bien que le critre soit
linaire, la nature des contraintes (une BMI) rend le problme non convexe en (K, X, ).
On trouvera dans la thse [133] (ainsi que dans les rfrences qui y gurent) un tat des
lieux concernant les techniques de rsolution actuelles de tels problmes (points intrieurs,
optimisation non lisse, branch-and-bound, descente coordonne), ainsi que la prsentation
de plusieurs mthodes originales (dont une de type Lagrangien augment).
Comme pour la stabilit (Thorme 1.2.1, page 25), il est cependant possible de rame-
ner la BMI (1.72) un systme de deux LMI couples par une contrainte non linaire. Par
une srie de manipulations algbriques introduites dans [55] (incluant des projections du
lemme born-rel), la variable K du correcteur disparat, et lon aboutit une nouvelle
caractrisation des solution du problme H
sous-optimal.
Thorme 1.2.4 Etant donn > 0, et sous les hypothses (i)-(ii), le systme (1.33)
admet un correcteur -sous-optimal pour le problme H
(1.64) si et seulement si il
existe des matrices X, Y o
n+n
K
(R) vriant les LMI
_
W
[B
T
2
D
T
12
]
0
0 I
_
T
_
_
AX +XA
T
XC
T
1
B
1
C
1
X I D
11
B
T
1
D
T
11
I
_
_
_
W
[B
T
2
D
T
12
]
0
0 I
_
0 , (1.74)
_
W
[C
2
D
21
]
0
0 I
_
T
_
_
A
T
Y +Y A Y B
1
C
T
1
B
T
1
Y I D
T
11
C
1
D
11
I
_
_
_
W
[C
2
D
21
]
0
0 I
_
0 , (1.75)
_
X I
I Y
_
_ 0 , (1.76)
o W
[B
T
2
D
T
12
]
et W
[C
2
D
21
]
sont des matrices dont les colonnes forment des bases des noyaux
de [B
T
2
D
T
12
] et de [C
2
D
21
].
Lensemble des correcteurs -sous-optimaux contient des correcteurs dordre n
K
< n
si et seulement si les matrices X et Y vrient de plus la contrainte de rang
rang (I XY ) n
K
. (1.77)
Grce au thorme 1.2.4, la synthse H
par correcteur dordre plein est accessible

en minimisant sous les contraintes LMI (1.74)-(1.76). Cest un problme doptimisation
semi-dnie convexe, pour lequel les mthodes de points intrieurs sont ecaces. Le cas de
lordre rduit ajoute la contrainte de rang non convexe (1.77) (et celle-ci devient XY = I
dans le cas dun correcteur statique) ; les mthodes de type Lagrangien augment partiel,
prcdemment cites pour la stabilisation (page 26), sont encore applicables ici. Dautre
part, une approche dite spectral quadratic-SDP est propose dans [14] pour rsoudre les
problmes de ce type (minimisation dun critre linaire sous contraintes LMI et galit
non linaire).
Dans tous les cas, une fois calcule une solution (X, Y ) de (1.74)-(1.76) (et ventuel-
lement (1.77)), associe un niveau de performance , un correcteur -sous-optimal K
est reconstruit notamment comme solution dune nouvelle LMI [55]. De plus, il est pos-
sible dimposer des conditions supplmentaires sur X et Y an de garantir au correcteur
reconstruit certaines proprits. Il est ainsi possible de xer un majorant sur son ordre,
cest lobjet de la contrainte (1.77). Une contrainte sur le degr de stabilit
5
de la ma-
trice A
c
(K) peut galement tre pose via deux LMI additionnelles en (X, Y ) (qui ne
sont toutefois que des conditions ncessaires), ceci an dviter la simplication ple/zro
entre le systme et le correcteur, comme illustr dans [55]. La diversit et la lisibilit des
solutions (X, Y ) ore ainsi une exibilit dans la synthse H
que ne permettait pas

lapproche DGKF. Enn, notons que les ingalits matricielles permettent de saranchir
des hypothses de rgularit.
Optimisation non lisse
Des techniques doptimisation non lisse (exploitant le -sous-direntiel dans le cas
non convexe) ont t appliques avec succs pour rsoudre les problmes doptimisation
semi-dnie avec contraintes de type BMI dans la thse [133] dj cite. Plus rcemment,
une technique de rgion de conance applique la rsolution de problmes quadratiques
locaux sous contraintes SDP a t propose [11]. De telles approches permettent a priori
de rsoudre le problme H
, qui se formule sans perte de gnralit dans ce cadre grce

au lemme born-rel (Lemme 1.2.3) (retour dtat, de sortie, correcteur statique, dordre
rduit, ou dordre plein). La dimension de lespace de recherche devient cependant vite
trs importante pour des systmes complexes, en raison de la prsence de la variable de
Lyapunov (
1
2
(n + n
K
)(n + n
K
+ 1) variable relles indpendantes) en plus du correcteur
((n
K
+m
2
)(n
K
+p
2
) variables) et du scalaire .
Quitte faire intervenir des techniques non lisses, pourquoi ne pas les appliquer direc-
tement au problme H
originel (1.63), qui est lui-mme non lisse comme nous lavons
dj soulign ? Cette approche directe, qui fonde les travaux dvelopps dans cette thse,
peut tre exploite pour rsoudre des problme de stabilisation (travaux de Burke, Lewis
et Overton prcdemment cits, et nouvel algorithme prsent au chapitre 3). Dans le cas
de la synthse H
, les travaux rcents prsents dans [8, 9, 7] et dans [94, 32] montrent
clairement les bnces tirer :
la dimension de lespace de recherche est rduite au nombre de paramtres du cor-
recteur ; elle est indpendante de lordre du systme commander,
la lisibilit des variables de dcision fait quil devient ais dimposer des contraintes
au correcteur ; en pratique, il sura de disposer dune paramtrisation diren-
tiable K() du correcteur pour xer la structure souhaite (ordre rduit,
Proportionnel-Intgral-Driv ou PID, commande dcentralise, . . .),
bien que le problme de synthse H
soit non lisse et non convexe, les algorithmes

doptimisation non direntiable donnent, sous certaines hypothses vries en
pratique, un certicat de convergence qui garantit loptimalit locale du correcteur
obtenu.
Nous rappelons dans le chapitre suivant les lments doptimisation non lisse qui per-
mettront de dvelopper en dtail ces points forts dans la deuxime partie de ce document.
5
Le degr de stabilit de la matrice dtat A dun systme LTI est dni par
sd
(A) max Re
i
(A) = (A).
Chapitre 2
Optimisation non lisse
Je me dtourne avec eroi et horreur de
cette plaie lamentable des fonctions
continues qui nont point de drives.
Charles Hermite (1822-1901),
cit par F. H. Clarke [44].
2.1 Terminologie et dnitions
Nous adoptons ici les notations habituelles et gnrales de lAnalyse et de lOptimi-
sation non lisse. En particulier, les variables et symboles choisis ne font jamais rfrence,
dans ce chapitre, aux notions relatives aux systmes dynamiques et la commande pr-
cdemment prsents.
Nous avons choisi une prsentation didactique des notions danalyse non lisse : nous
commenons par quelques pr-requis, puis nous introduisons le sous-direntiel dans le
cas des fonctions convexes, pour lesquelles la gnralisation du gradient est assez intuitive.
Nous prsentons ensuite le sous-direntiel de Clarke, partir des drives directionnelles
gnralises. Le principe gnral des mthodes de faisceaux est expos. Enn, nous dcri-
vons des mthodes doptimisation non lisse spciques aux problmes minimax, du type
de ceux rencontrs dans les problmes de commande.
2.1.1 Ingrdients pour lanalyse non lisse
Lespace R
n
est muni du produit scalaire euclidien x, y) = x
T
y. La norme associe est
note |.|. B(x, r) dsigne la boule (ouverte) de centre x R
n
et de rayon r > 0, dnie
par
B(x, r) y R
n
: |y x| < r .
An de bien distinguer les composantes dun vecteur x R
n
dans la base canonique
dune part, et les termes dune suite (x
i
)
iI
(I ni ou non) dlments de R
n
dautre part,
on notera les composantes x
j
et les termes x
i
. Ainsi x
j
i
dsigne la j-me composante du
i-me terme dune suite.
33
34 Chapitre 2. Optimisation non lisse
Nous ne rappelons pas ici les notions fondamentales du calcul direntiel (direntia-
bilit au sens de Frchet, au sens de Gteaux, gradient, fonctions de classe c
n
), mais nous
xons les quelques dnitions supplmentaires indispensables une gnralisation de ces
notions.
Dnition 2.1.1 Un ensemble C R
n
est dit convexe si, pour tout x, y C et pour
tout rel [0, 1], on a x + (1 )y C.
On note [x, y] z R
n
: z = x+(1)y et 0 1 le segment ferm dextrmits
x et y.
Pour tout k N
,
k
dsigne le simplexe unit de R
k
, dni par
k
R
k
:
k
i=1
i
= 1 et
i
0 pour tout 1 i k . (2.1)
Si C R
n
est convexe, et si (x
1
, . . . , x
k
) C
k
, alors toute combinaison linaire convexe
de la forme

k
i=1
i
x
i
(avec
k
) est encore dans C.
Toute intersection de convexes est encore convexe. Etant donne un ensemble S R
n
,
on appelle enveloppe convexe de S et on note co S lintersection de tous les convexes
contenant S, cest--dire le plus petit convexe contenant S. Il est dcrit par toutes les
combinaisons linaires convexes dlments de S :
co S = x R
n
: x =
k
i=1
i
x
i
,
k
, x
i
S, 1 i k, k 1 . (2.2)
Dnition 2.1.2 Soit C R
n
un ensemble convexe, s R
n
et x

C. La direction s est
dite normale C en x si
s, y x) 0 pour tout y C. (2.3)
Lensemble de ces directions s forme le cne normal C en x, not N
C
(x).
Dnition 2.1.3 Une fonction f : R
n
R est dite convexe si, pour tout x, y R
n
et
pour tout rel [0, 1], on a
f(x + (1 )y) f(x) + (1 )f(y) . (2.4)
Si lingalit (2.4) est stricte pour tout x, y R
n
tels que x ,= y, et pour tout ]0, 1[,
f est dite strictement convexe.
Une fonction f : R
n
R est convexe si et seulement si son pigraphe
Epi (f)
__
x
s
_
R
n+1
: s f(x)
_
est une partie convexe de R
n+1
. Enn, toute fonction convexe est continue sur R
n
.
n
R est dite lipschitzienne sur S R
n
sil existe
une constante L > 0 tel le que, pour tout y, z S
[f(y) f(z)[ L|y z| (2.5)
f est dite localement lipschitzienne en x R
n
sil existe un rel positif > 0 tel que f est
lipschitzienne sur B(x, ).
Si f est lipschitzienne sur S (resp. localement lipschitzienne en x), toute valeur de L
qui vrie lingalit (2.5) de la dnition 2.1.4 est appele constante de Lipschitz de f
sur S (resp. en x, ou au voisinage de x).
Le caractre localement lipschitzien dune fonction en x assure que son taux daccrois-
sement peut tre contrl dans un voisinage de x. Une fonction localement lipschitzienne
en un point nest pas ncessairement direntiable en ce point (par exemple x [x[
en 0). Enn, on peut montrer que toute fonction convexe f : R
n
R est localement
lipschitzienne en tout point de R
n
(et donc que f est lipschitzienne sur tout compact de
R
n
).
Terminons par la notion de drive directionnelle :
n
R admet une drive directionnelle (unilatrale)
en x R
n
dans la direction d R
n
si la limite
lim
t0
t>0
f(x +td) f(x)
t
(2.6)
existe et est nie. On la note alors f
t
(x, d).
Si f est direntiable en x (au sens de Frchet ou au sens de Gteaux), alors elle admet
des drives directionnelles dans toutes les directions d, et on a f
t
(x, d) = f
t
(x)(d) =
f(x)
T
d. Rappelons que la rciproque nest pas vraie en gnral, sauf si les drives
directionnelles sont continues.
Dnition 2.1.6 Une fonction : R
n
R est dite sous-linaire si, par dnition, elle
est convexe et positivement homogne, cest--dire
(tx) = t(x) pour tout x R
n
et t > 0 . (2.7)
On montre facilement que, dans cette dnition, lhypothse de convexit peut tre rem-
place par la sous-additivit de :
(x +y) (x) + (y) pour tout x, y R
n
(2.8)
Ainsi une fonction sous-linaire est une fonction sous-additive et positivement homogne.
Toute fonction sous-linaire possde la proprit fondamentale de majorer au moins
une fonction linaire, ce qui permet dnoncer le thorme suivant :
Thorme 2.1.7 Soit : R
n
R une fonction sous-linaire. Alors lensemble
S
s R
n
: s, x) (x) pour tout x R
n
(2.9)
est non vide. Il est de plus compact et convexe. On a enn la relation
(x) = sup s, x) : s S
. (2.10)
Rciproquement, tant donn un ensemble S non vide, compact et convexe, la fonction
support
S
: R
n
R de S, dnie par
S
(x) sup s, x) : s S , (2.11)
est sous-linaire.
Le thorme prcdent permet ainsi dtablir une bection entre lensemble des parties
non vides, compactes et convexes de R
n
dune part, et lensemble des applications sous-
linaires R
n
R dautre part.
2.1.2 Analyse non lisse : le cas convexe
Sous-direntiel dune fonction convexe
Lanalyse convexe permet de gnraliser assez naturellement la notion de direntielle,
en introduisant de nouveaux objets (sous-direntiel et sous-gradient), dnis pour les
fonctions convexes sur R
n
(donc pas ncessairement direntiables). Nous en donnons
ici les principales tapes, et nous invitons le lecteur se rfrer louvrage [68], qui
propose une approche trs pdagogique de lanalyse convexe, ainsi qu [121], rfrence
en la matire.
Dans toute la section 2.1.2, f : R
n
R dsigne une fonction convexe.
Thorme 2.1.8 Pour tout x R
n
, f admet des drives directionnelles dans toutes les
directions d R
n
, et on a
f
t
(x, d) = inf
t>0
f(x +td) f(x)
t
. (2.12)
De plus, pour tout x x, lapplication d R
n
f
t
(x, d) est sous-linaire.
Remarquons que toute constante locale de Lipschitz L de f en x est encore une
constante de Lipschitz de d R
n
f
t
(x, d), et ce de faon globale. On a en particu-
lier
[f
t
(x, d)[ L|d| (2.13)
Le thorme 2.1.7 permet donc de considrer d R
n
f
t
(x, d) comme la fonction
support dun ensemble non vide, compact et convexe de R
n
, dni comme suit :
Dnition 2.1.9 Le sous-direntiel de f en x, not
c
f(x) (avec c pour convexe), est
lensemble non vide, compact et convexe de R
n
dont la fonction support est d R
n
f
t
(x, d), cest--dire
c
f(x) s R
n
: s, d) f
t
(x, d) pour tout d R
n
. (2.14)
Les lments de
c
f(x) sont appels les sous-gradients de f en x.
De plus, connaissant
c
f(x), il est immdiat de retrouver nimporte quelle drive direc-
tionnelle :
f
t
(x, d) = max s, d) : s
c
f(x) . (2.15)
Enn, on a
c
f(x) B(0, L), avec (2.13).
Interprtation gomtrique
Le thorme suivant donne une caractrisation du sous direntiel qui ne fait pas appel
aux drives directionnelles, mais plutt la notion de minoration par une fonction ane :
Thorme 2.1.10 Pour tout x R
n
, on a lgalit suivante
c
f(x) = s R
n
: f(y) f(x) +s, y x) pour tout y R
n
. (2.16)
Cette description du sous-direntiel sinterprte bien en termes gomtriques, dans
R
n+1
: s
c
f(x) si et seulement si [
s
1
] est dans le cne normal lpigraphe de f
en [
x
f(x)
]. La gure 2.1 illustre cette interprtation dans le cas dune fonction convexe
dnie sur R; le cne tangent Epi (f) en [
x
f(x)
] est reprsent par des ches en trait
pointill, le cne normal par des ches en trait continu. Un vecteur [
s
1
] du cne normal
est reprsent en gras.
x +
c
f(x)
x x +s
f(x)
Epi (f)
f(x) 1
Fig. 2.1 Interprtation gomtrique du sous-direntiel dans le cas convexe.
Cas direntiable
Thorme 2.1.11 Si f convexe est direntiable en x, alors
c
f(x) est rduit au single-
ton f(x). Rciproquement, si f est convexe et
c
f(x) = s, alors f est direntiable
en x et s = f(x).
Nous navons volontairement pas indiqu le type de direntiabilit de f en x, car
celle au sens de Frchet et celle au sens de Gteaux concident dans le cas des fonctions
localement lipschitziennes, donc en particulier pour les fonctions convexes.
2.1.3 Analyse non lisse : le cas lipschitizien
Sous-direntiel de Clarke
f : R
n
R est maintenant suppose localement lipschitzienne en x R
n
, de constante
de Lipschitz L, et ce dans toute la section 2.1.3. Lhypothse nest pas susante pour
parler de drives directionnelles de f, contrairement au cas convexe. Clarke [44], dont
notre prsentation reprend ici les principaux rsultats, dnit la drive directionnelle
gnralise de la faon suivante :
Dnition 2.1.12 Pour tout d R
n
, la fonction f admet une drive directionnelle
gnralise en x dans la direction d, note f
(x, d), et dnie par

f
(x, d) limsup
yx
t0
t>0
f(y +td) f(y)
t
. (2.17)
Puisque f est localement lipschitzienne en x R
n
, f
(x, d) est nie, et

[f
(x, d)[ L|d| . (2.18)

Comme dans le cas convexe, lapplication d f
(x, d) hrite du caractre localement

lipschitzien de f, de constante L, et celui-ci est mme global. On montre de mme que
d f
(x, d) est sous-linaire, de sorte quil est permis de dnir, de faon analogue la
dnition 2.1.9, un sous-direntiel pour les fonctions localement lipschitziennes :
Dnition 2.1.13 Le sous-direntiel de Clarke de f en x, not f(x) est lensemble
non vide, compact et convexe de R
n
dont la fonction support est d R
n
f
(x, d),
cest--dire
f(x) s R
n
: s, d) f

n
. (2.19)
Les lments de f(x) sont appels les sous-gradients de Clarke (ou gradients gnraliss)
de f en x.
On a f(x) B(0, L), daprs (2.18). Les drives directionnelles gnralises se
dduisent de f(x) pour toute direction d R
n
:
f
(x, d) = max s, d) : s f(x) . (2.20)

Le sous-direntiel de Clarke f(x) sinterprte gomtriquement dans R
n+1
(voir
gure 2.2), condition toutefois de gnraliser la notion de cne normal une partie
quelconque C ,= (pas ncessairement convexe). Nous invitons le lecteur se reporter
[44], o est dni le cne tangent C en un point x C, puis, par polarit, le cne
normal C en x. Ces dnitions poses, on montre que f(x) est lensemble des vecteurs
s R
n
tels que [
s
1
] est dans le cne normal lpigraphe de f en [
x
f(x)
].
Epi (f)
f(x)
x x +s
f(x) 1
x +f(x)
Fig. 2.2 Interprtation gomtrique du sous-direntiel de Clarke.
Cas convexe, cas direntiable
Ces objets gnralisent ceux introduits la section 2.1.2. En eet, si f est convexe, on
a f
(x, d) = f
t
n
, et par suite f(x) =
c
f(x).
Le lien avec le cas direntiable est moins vident que dans le cas convexe : si f,
toujours suppose localement lipschitzienne en x, est (Frchet- ou Gteaux-) direntiable
en x, alors on a f
t
(x, d) = f(x), d) f

n
, et donc f(x) f(x).
Si on suppose maintenant que f est continment direntiable en x (f est donc en
particulier localement lipschitzienne en x), on a lgalit f(x), d) = f
(x, d) pour tout

d R
n
, par dnition de la drive directionnelle gnralise. Ainsi d f
(x, d) est elle-

mme linaire, et ne peut tre minore sur R
n
par une autre application linaire. Daprs
la dnition 2.1.13, on a donc f(x) = f(x).
Caractrisation du sous-direntiel de Clarke
Si une fonction localement lipschitzienne nest pas ncessairement direntiable, len-
semble de ses points de non-direntiabilit est nanmoins de mesure nulle, comme laf-
rme le thorme suivant, d Rademacher :
Thorme 2.1.14 Soit U R
n
un ouvert, et g : U R une fonction localement lip-
schitzienne en tout point de U. Alors g est direntiable presque-partout sur R
n
(pour la
mesure de Lebesgue).
Toute fonction f : R
n
R localement lipschitzienne sur R
n
est donc direntiable par-
tout, sauf sur
ns
, un ensemble de mesure nulle. f admet donc un gradient f(x) en
un point x arbitrairement proche de x
ns

ns
. On a dautre part la caractrisation du
sous-direntiel de Clarke suivante :
Thorme 2.1.15 Si f : R
n
R est localement lipschitzienne sur R
n
, direntiable sur
R
n

ns
, avec
ns
Lebesgue-ngligeable, alors
f(x) = co s R
n
: il existe (x
i
) R
n

ns
tel le que x
i
x et f(x
i
) s .
(2.21)
Cette nouvelle description permet une construction eective et exhaustive du sous-diff-
rentiel de Clarke, ds que lon sait calculer les gradients l o cela est possible (cest--dire
presque-partout). Elle reste bien sr valide dans le cas convexe.
Rgles de calcul
Nous donnons ici quelques rgles de calcul utiles du sous-direntiel de Clarke. Les
rsultats ne sont souvent que des inclusions, mais il est possible dobtenir des galits en
demandant une rgularit plus forte que le caractre localement lipschitzien, dite rgularit
au sens de Clarke :
Dnition 2.1.16 f : R
n
R localement lipschitzienne en x est dite rgulire au sens de
Clarke si el le admet des drives directionnelles f
t
(x, d) dans toutes les directions d R
n
,
et si, de plus, on a f
t
(x, d) = f

n
.
En particulier, toute fonction convexe sur R
n
et toute fonction continment direntiable
en x est rgulire au sens de Clarke en x.
Dautre part, si f est rgulire au sens de Clarke et direntiable en x, alors f(x) =
f(x), ce qui prcise le rsultat annonc prcdemment.
Proposition 2.1.17 Soit (f
i
)
1ip
une famil le nie de fonctions dnies sur R
n
, va-
leurs relles, et localement lipschitziennes en x R
n
, et soit (
i
)
1ip
R. Alors la
somme pondre

i
f
i
est encore localement lipschitzienne en x, et on a
_
p
i=1
i
f
i
_
(x)
p
i=1
i
f
i
(x) . (2.22)
On a galit si une fonction f
i
au plus nest pas continment direntiable en x. On a
encore galit si les f
i
sont rgulires au sens de Clarke, et si
i
0 pour tout i.
Prcisons que le second membre de (2.22) contient toutes les sommes de la forme
i
s
i
, o s
i
f
i
(x).
Proposition 2.1.18 (Rgle de direntiation en chane)
Soit F : R
n
R
m
une fonction continment direntiable en x R
n
, et g : R
m
R
une fonction localement lipschitzienne en F(x). Alors la compose g F est localement
lipschitzienne en x, et on a
(g F)(x) g (F (x)) F
t
(x), (2.23)
o F
t
(x) dsigne la direntielle de F en x. On a galit si g est rgulire au sens de
Clarke en F(x), et, dans ce cas, la compose g F est aussi rgulire au sens de Clarke
en x.
Le second membre de (2.23) dsigne lensemble des formes linaires sur R
n
de la forme
d s, F
t
(x).d), o s dcrit g (F (x)). On le note encore [F
t
(x)]
g (F (x)), comme
action de lapplication linaire adjointe [F
t
(x)]
sur le sous-direntiel.
Proposition 2.1.19 Soit (f
i
)
1ip
une famil le nie de fonctions dnies sur R
n
, va-
leurs relles, et localement lipschitziennes en x R
n
. Alors le maximum point par point
max
i
f
i
est encore localement lipschitzien en x R
n
, et on a
_
max
1ip
f
i
_
(x) co f
j
(x) : j (x) , (2.24)
o (x) dsigne lensemble des indices j actifs en x, tels que f
j
(x) = max
i
f
i
(x). On a
galit si, pour tout j (x), les f
j
sont rgulires au sens de Clarke en x, et dans ce
cas max
i
f
i
est elle-mme rgulire au sens de Clarke en x.
Cette dernire rgle est fondamentale, car elle permet dobtenir tous les sous gradients
dune fonction dnie comme le maximum point par point dune famille de fonctions
continment direntiables. De telles fonctions apparaissent souvent en pratique, lors-
quun ensemble (ni ou non) de spcications exprimes par des ingalits doivent tre
vries. La minimisation de ces fonctions forme la familles des problmes dits minimax,
sur lesquels nous reviendrons la section 2.3.3.
2.2 Conditions doptimalit
Nous montrons dans cette section en quoi les lments qui viennent dtre prsents
forment un cadre thorique adapt la rsolution de problmes doptimisation.
Considrons le problme doptimisation
min
xR
n
f(x) , (2.25)
o f : R
n
R est suppose localement lipschitzienne en tout point x R
n
. Il sagit dun
problme non lisse, puisque f nest pas ncessairement direntiable sur R
n
en entier.
Rappelons quune solution de (2.25), si elle existe, peut tre globale, ou seulement locale :
2.3. Algorithmes pour loptimisation non lisse 43
Dnition 2.2.1 Un point x
R
n
est un minimum global de f sur R
n
si, pour tout
x R
n
, f(x
) f(x).
Un point x
R
n
est un minimum local de f sil existe > 0 tel que, pour tout
x B(x
, ), f(x
) f(x).
Le sous-direntiel de Clarke permet de formuler une condition ncessaire doptimalit
locale (sans contrainte) pour les fonctions localement lipschitzienne sur R.
Thorme 2.2.2 Si x
est un minimum local de f, alors 0 f(x
).
Remarque: Si x
est un minimum local de f, on a alors

f
(x
, d) = max s, d) : s f(x
) 0 (2.26)
puisque 0 f(x
).
Le thorme 2.2.2 gnralise la condition ncessaire doptimalit locale du premier
ordre des fonctions direntiables (f(x
) = 0). Par analogie, un point x
R
n
tel que
0 f(x
) est appel un point stationnaire de f.

Dans le cas f convexe, on a une caractrisation des minima de f par le sous-direntiel,
et ces minima sont globaux :
Thorme 2.2.3 Si f est convexe, x
est un minimum global de f sur R

n
si et seulement
si 0
c
f(x
).
Le thorme de Rademacher pourrait laisser penser que les conditions doptimalit
des thormes 2.2.2 et 2.2.3 ont un faible intrt pratique, puisquil y a peu de chances
de tomber sur un point x o f nest pas direntiable. Cest en pratique le contraire
qui se produit : un minimum est souvent atteint en un point o la direntiabilit est
perdue, et en lequel les conditions doptimalit pour les fonctions lisses sont videmment
inapplicables.
2.3 Algorithmes pour loptimisation non lisse
2.3.1 Direction de descente
En optimisation non lisse comme en optimisation direntiable, la recherche pratique
des minima locaux se fait rarement par la rsolution directe des conditions doptima-
lit, mais utilise des mthodes itratives pour construire une suite ditrs (x
k
)
kN
R
n
,
qui converge vers un minimum (local ou global) x
lorsque k . Lorsque la suite

(f (x
k
))
kN
R
n
des valeurs de f correspondantes est strictement dcroissante, la m-
thode est dite de descente. Cest gnralement le cas des mthodes classiques doptimisa-
tion direntiable sans contrainte (plus profonde descente, mthode de Newton, de quasi-
Newton, . . .). Dans le cas gnral, quelle que soit la rgularit de la fonction considre,
on a la dnition suivante :
Dnition 2.3.1 Soit f : R
n
R. On dit que la direction d R
n
est une direction de
descente pour f en x si, par dnition, il existe un scalaire > 0 tel que
f(x +td) < f(x) pour tout t ]0, ] . (2.27)
Dans le cas non lisse, on a les caractrisations suivantes :
Thorme 2.3.2 Soit f : R
n
R une fonction localement lipschitzienne en x R
n
.
d R
n
est une direction de descente pour f en x si lune des conditions quivalentes
suivantes est ralise :
(i) f
(x, d) < 0
(ii) s, d) < 0 pour tout s f(x)
Une direction dite de plus profonde descente est une direction de descente pour laquelle
la drive directionnelle gnralise est la plus ngative possible. Etant donn que d
f
(x, d) est positivement homogne, on peut se ramener au problme normalis

min
|d|1
f
(x, d) , (2.28)
ou, de faon quivalente,
min
|d|1
max
sf(x)
s, d) , (2.29)
dont toute solution d
est dite direction normalise de plus profonde descente. Avec la

norme sur R
n
choisie ici (norme euclidienne), on montre que sil existe une direction de
descente pour f en x, alors la direction de plus profonde descente d
est unique (on la

note d
(x)) ; elle est obtenue partir du sous-gradient de norme minimale, projection de

lorigine sur le convexe compact f(x) :
d
(x) =
s
(x)
|s
(x)|
, o s
(x) = arg min |s| : s f(x) . (2.30)

Ainsi la recherche dune direction de plus profonde descente pour une fonction non lisse
ncessite la rsolution dun programme convexe. Une mthode de descente non lisse ba-
se sur la direction de plus profonde descente a cependant toutes les chances de ne pas
converger vers un point stationnaire de f (voir par exemple [68], dans le cas convexe),
en raison dventuelles discontinuits de x d
(x). En eet, lapplication multivalue

x f(x) T(R
n
) est semi-continue suprieurement
1
(voir [44, Proposition 2.1.5])
mais pas infrieurement. Autrement dit, si f(x) ne peut pas grossir brutalement, rien ne
garantit que f(x
0
) ne se contracte pas brusquement, et cest justement ce qui se produit
en un point x
0
o f est non direntiable : f(x
0
) nest pas rduit un singleton, mais
tout voisinage de x
0
contient un point x o f est direntiable (thorme de Radema-
cher 2.1.14) et en lequel f(x) nest plus quun singleton. La direction de plus profonde
descente nest donc pas continue en x
0
.
1
Lapplication multivalue F : R
n
T(R
m
) est dite semi-continue suprieurement (on trouve parfois
semi-continue extrieurement) si, par dnition, pour tout x
0
R
n
, F(x
0
) est ferm et pour tout compact
C R
m
disjoint de F(x
0
), il existe > 0 tel que, pour tout x B(x
0
, ), F(x) et C restent disjoints.
F est dite semi-continue infrieurement (ou intrieurement) si, par dnition, pour tout x
0
R
n
, pour
tout ouvert O R
m
qui coupe F(x
0
), il existe > 0 tel que, pour tout x B(x
0
, ), F(x) coupe encore
O.
2.3.2 Mthodes de faisceaux
Les mthodes de faisceaux (ou bundle methods) viennent corriger ce dfaut de conver-
gence, et sont largement utilises pour loptimisation de fonctions non lisses. Elles sont
prsentes dans [69] dans le cas convexe, et dans [80] dans le cas localement lipschitzien,
et il en existe de nombreuses variantes (voir par exemple [127], et [92] pour une vue
densemble). Nous nen donnons ici que les grandes lignes, dans le cas convexe :
Direction de recherche
On suppose qu litration k 1, le point courant est x
k
R
n
, et on dispose dune
famille nie de points (y
j
)
j7
k
R
n
(avec 1
k
1, . . . , k), constitue au cours des
itrations prcdentes. On suppose de plus que lon connat un sous gradient s
j
f(y
j
)
pour tout j 1
k
. On dnit lapproximation ane par morceaux de f suivante :
f
k
(x) max
j7
k
f(y
j
) +s
j
, x y
j
) , (2.31)
ou encore, en faisant apparatre le point courant x
k
,
f
k
(x) = max
j7
k
_
f(x
k
) +s
j
, x x
k
)
j
k
_
, (2.32)
o le terme
j
k
est lerreur de linarisation, pour tout j 1
k
j
k
f(x
k
) f(y
j
) s
j
, x
k
y
j
) . (2.33)
Comme f est convexe, on a
j
k
0 pour tout j 1
k
, et

f
k
(x) f(x) pour tout x R
n
. La
direction de recherche d
k
est ensuite obtenue comme solution du problme de minimisation
min
dR
n
_
f
k
(x
k
+d) +
1
2
d, G
k
d)
_
, (2.34)
o le terme quadratique (avec G
k
~ 0) assure lexistence et lunicit du minimum, et
pnalise le critre si x
k
+ d est trop loign de x
k
. Il agit ainsi comme une contrainte de
type rgion de conance, permettant de tenir compte du caractre local de

f
k
. Il permet
de plus dinclure une information du second-ordre sur f au voisinage de x
k
.
Le sous-problme (2.34) est encore non lisse, mais peut facilement se reformuler comme
le programme quadratique convexe sous contrainte de type ingalit
min
dR
n
vR
v +
1
2
d, G
k
d)
tel que s
j
, d)
j
k
v pour tout j 1
k
(2.35)
qui est lui-mme quivalent, par dualit lagrangienne (voir par exemple [102]), la re-
cherche du multiplicateur
k
= (
j
k
)
j7
k
solution de
min
I
k
j7
k
j
k
j
k
+
1
2
_
_
j7
k
j
k
s
j
_
_
T
G
1
k
_
_
j7
k
j
k
s
j
_
_
, (2.36)
o
7
k
dsigne le simplexe unit dni par
7
k

_
_
_
(
j
k
)
j7
k
:

j7
k
j
k
= 1 et
j
k
0 pour tout j 1
k
_
_
_
. (2.37)
La direction de recherche d
k
solution de (2.35) est reconstruite partir du multiplica-
teur optimal
k
solution de (2.36), comme combinaison convexe :
d
k
=
j7
k
j
k
G
1
k
s
j
. (2.38)
Enn, la dcroissance de f attendue (base sur le modle

f
k
) est la valeur v
k
de v, solution
du problme (2.35). On a
v
k
=

f
k
(x
k
+d
k
) f(x
k
)
=
j7
k
j
k
j
k
d
k
, G
k
d
k
)
0 .
Recherche linaire
On procde ensuite une recherche linaire dans la direction d
k
, partir du point
courant x
k
. Comme t f(x
k
+td
k
) est non lisse, il faut recourir des mthodes adaptes.
On suppose xs les paramtres m
L
]0,
1
2
[, m
R
]m
L
, 1[ et une taille de pas minimale
t ]0, 1]. On recherche le plus grand pas t
L,k
[t, 1] qui ralise la fraction m
L
de la
dcroissance attendue, cest--dire
f(x
k
+t
L,k
d
k
) f(x
k
) +m
L
t
L,k
v
k
. (2.39)
La condition (2.39) est rapprocher de la condition dArmo, o v
k
remplace ici la drive
directionnelle f
t
(x
k
, d
k
).
Trois situations peuvent se prsenter :
1. Un tel t
L,k
existe ; on ralise un long serious step :
x
k+1
x
k
+t
L,k
d
k
et y
k+1
x
k+1
. (2.40)
2. La condition (2.39) nest vrie pour aucun pas dans [t, 1], mais par t
L,k
]0, t[ ; on
ralise un short serious step :
x
k+1
x
k
+t
L,k
d
k
et y
k+1
x
k+1
+t
R,k
d
k
. (2.41)
3. La condition (2.39) nest vrie sur [0, 1] que pour t
L,k
= 0 ; on ralise un null step :
x
k+1
x
k
et y
k+1
x
k+1
+t
R,k
d
k
. (2.42)
Dans les deux derniers cas, t
R,k
est choisi tel que
t
R,k
> t
L,k
et
k+1
k+1
+s
k+1
, d
k
) m
R
v
K
(2.43)
avec s
k+1
f(y
k+1
), et
k+1
k+1
= f(x
k+1
) f(y
k+1
) + s
k+1
, y
k+1
x
k+1
). La condition
(2.43) garantit que tout nouveau sous-gradient s
k+1
f(y
k+1
) enrichit linformation
contenue dans (s
j
)
j7
k
, car alors x
k+1
et y
k+1
sont localiss de part et dautre dune
discontinuit de x f
t
(x, d
k
).
Une nouvelle direction de recherche d
k+1
est ensuite dtermine, suivie dune nouvelle
recherche linaire, et ce, tant que la condition
v
k
=

f
k
(x
k
+d
k
) f(x
k
) < (2.44)
est vrie, pour > 0 x. Sous certaines hypothses (voir les remarques ci-aprs, ainsi
que [80]), lalgorithme ainsi dni neectue quun nombre ni de nul l steps, et converge
vers un point stationnaire de f.
Remarques:
Si f nest pas convexe mais localement lipschitzienne, lalgorithme doit tre adapt
car lerreur de linarisation (2.33) peut changer de signe (voir par exemple [80]).
Au bout dun grand nombre ditrations, lensemble 1
k
devenant important, le sto-
ckage en mmoire des sous-gradients peut tre problmatique, et leur grand nombre
peut rendre le sous-problme quadratique (2.36) dicile rsoudre. On impose alors
une taille maximale 1
k
, que lon met jour comme une le FIFO (rst in, rst out,
soit premier entr, premier sorti). La convergence de lalgorithme nest cependant
plus assure, et lon doit faire appel une stratgie dagrgation de sous-gradients
an de la garantir (voir [80]).
La famille des matrices G
k
~ 0 doit rester borne an de conserver la convergence
de lalgorithme. La varit des mthodes de faisceaux existantes repose en grande
partie sur la latitude dans le choix de G
k
. Avec G
k
= u
k
I
n
, on obtient par exemple
la mthode dite proximal bundle (voir [81]) ou bundle trust region (voir [127]).
2.3.3 Mthodes pour les problmes minimax
Les problmes doptimisation dits minimax (ou min-max) sans contrainte sont de la
forme
min
xR
n
max
j
p
f
j
(x) , (2.45)
o
p
est lensemble ni dindices
p
1, . . . , p, et les fonctions f
j
: R
n
R sont
soit des fonctions continment direntiables ; on dit dans ce cas que (2.45) est un
problme minimax ni, sans contrainte,
soit dnies elles-mmes comme des maxima de fonctions :
f
j
(x) = max
y
j
Y
j
j
(x, y
j
) , (2.46)
avec, pour tout j
p
, Y
j
une partie compacte de R
m
j
et
j
: R
n
R
m
j
R une
fonction continue et telle que pour tout y R
m
j
, lapplication partielle x
j
(x, y)
est continment direntiable sur R
n
. On parle dans ce cas de problme minimax
semi-inni, car si lespace de recherche est de dimension nie n, la fonction objectif
apparat comme le maximum dune innit de fonctions.
On notera par la suite f = max
j
f
j
.
Sous les hypothses prcdemment nonces, on montre que f est rgulire au sens de
Clarke (elle admet donc en tout point un sous-direntiel et des drives directionnelles
dans toutes les directions, gales aux drives directionnelles gnralises), et pour tout
x, d R
n
on a
f(x) = co
j
p
(x)
f
j
(x) ,
f
t
(x, d) = max
j
p
(x)
f
j
(x), d)
= max
sf(x)
s, d)
(2.47)
dans le cas minimax ni, avec
p
(x) = j
p
: f
j
(x) = f(x) lensemble des indices
dits actifs en x, et
f(x) = co
j
p
(x)
_
co
yY
j
(x)
j
(x, y)
_
,
f
t
(x, d) = max
j
p
(x)
max
yY
j
(x)
j
(x, y), d)
= max
sf(x)
s, d)
(2.48)
dans le cas minimax semi-inni, o Y
j
(x) = y
j
Y
j
:
j
(x, y
j
) = f
j
(x) (ensemble des
valeurs de y actives) et pour tout y Y
j
(x),
x
j
(x, y) dsigne le gradient en x de
lapplication partielle
j
(, y).
Remarque: Le problme de synthse H
(1.63), si on ignore la contrainte de stabilit, est

clairement du type minimax semi-inni, puisquil se prsente sous la forme
min
KR
(m
2
+n
K
)(p
2
+n
K
)
max
1imin(m
1
,p
1
)
max
[0,+]
i
[T
wz
(K)(j)] (2.49)
o les
i
[T
wz
(K)(j)] dsignent les valeurs singulires de la matrice de transfert en
boucle ferme T
wz
(K), value la pulsation .
Les problmes minimax sont largement voqus par Polak dans louvrage [114], ainsi
que dans [113]. Les lments qui suivent, et notamment le choix de la fonction doptimalit,
en sont directement issus. Nous invitons le lecteur sy rfrer pour tout approfondisse-
ment.
Conditions doptimalit
Dans le cas ni, la condition ncessaire doptimalit dcoule de la proposition 2.1.19
et du thorme 2.2.2 :
Thorme 2.3.3 Si x
est un minimum local du problme minimax ni (2.45), alors

0 f(x
) = co
j
p
(x
)
f
j
(x
) , (2.50)
avec
p
(x
) =
_
j
p
: f
j
(x
) = f(x
) = max
i
p
f
i
(x
)
_
.
Elle peut encore snoncer par combinaison convexe : il existe un vecteur
p
tel que
p
i=1
i
f
i
(x
) = 0 , (2.51)
et vriant la condition de complmentarit
p
i=1
i
[f(x
) f
i
(x
)] = 0 . (2.52)
Ce rsultat se gnralise au cas minimax semi-inni (voir [113, Thorme 3.1.4]) :
Thorme 2.3.4 Si x
est un minimum local du problme minimax semi-inni (2.45),

alors
0 f(x
) = co
j
p
(x
)
f
j
(x
) , (2.53)
avec
p
(x
) = j
p
: f
j
(x
) = f(x
), et
f
j
(x
) = co
y
j
Y
j
(x
j
(x
, y
j
) , (2.54)
o Y
j
(x
) = y
j
Y
j
:
j
(x
, y
j
) = f
j
(x
).
Ces conditions ncessaires doptimalit locale se prsentent sous la forme dquations
o apparaissent les gradients des fonctions intervenant dans le max, et des multiplicateurs
convexes. Comme nous lavons dj remarqu dans un cadre plus vaste, la rsolution
directe de ces quations nest gnralement pas envisageable en pratique ; on a alors recours
un algorithme de descente. La direction normalise de plus profonde descente (pour la
norme euclidienne) est, dans le cas dun problme minimax ni :
d
(x) =
s
(x)
|s
(x)|
, o
s
(x) = arg min

_
_
_
|
j
p
(x)
j
f
j
(x)| :
p
(x)
_
_
_
.
avec
p
(x)
le simplexe unit dni par
p
(x)

_
_
_
(
j
k
)
j
p
(x)
:

j
p
(x)
j
k
= 1 et
j
k
0 pour tout j
p
(x)
_
_
_
. (2.55)
Nous lavons dj soulign, lapplication x d
(x) nest pas continue. Cela reste vrai

pour les fonctions de type max (ni ou non) telles que f, et apparat mme clairement
en remarquant que lensemble des indices actifs
p
(x) peut se rduire brutalement au
voisinage dun point o f nest pas direntiable. La consquence sur lalgorithme de plus
profonde descente est, ici encore, une convergence vers un point non stationnaire de f.
Minimax ni : fonction doptimalit et de descente
An de palier le dfaut de continuit de la direction de plus profonde descente, Polak
[114] dnit

f(x
0
, ), une approximation quadratique par morceaux de f en x
0
R
n
, tel
que pour tout x R
n
f(x
0
, x) max
j
p
_
f
j
(x
0
) +f
j
(x
0
), x x
0
) +
1
2
|x x
0
|
2
_
, (2.56)
avec > 0 x. Cest un modle du premier ordre de f en x
0
, strictement convexe. Il
permet de dnir une fonction doptimalit : R
n
R, ainsi quune direction de recherche
h : R
n
R; pour tout x R
n
(x) min
hR
n
_
f(x, x +h) f(x)

_
= min
hR
n
max
j
p
_
f
j
(x) f(x) +f
j
(x), h) +
1
2
|h|
2
_
,
(2.57)
et h(x) est dni comme lunique minimum de (2.57). La forme de ce dernier programme
est rapprocher de celle du problme (2.34) (page 45) introduit pour les mthodes de
faisceaux ; par un raisonnement analogue, on montre que (x) peut tre exprime par un
programme quadratique convexe : pour tout x R
n
(x) = min
p
_
_
q
j=1
j
[f(x) f
j
(x)] +
1
2
_
_
_
_
_
_
q
j=1
j
f
j
(x)
_
_
_
_
_
_
2
_
_
. (2.58)
En eet le max dans (2.57) porte sur une famille ni de scalaires, et reste inchang sil
est tendu lenveloppe convexe de ces scalaires :
(x) = min
hR
n
max
p
_
_
_
q
j=1
j
[f
j
(x) f(x)] +
q
j=1
j
f
j
(x), h) +
1
2
|h|
2
_
_
_
. (2.59)
Par dualit lagrangienne, il est possible dchanger le min et le max sans modier (x),
et pour
p
x, le min interne
min
hR
n
_
_
_
q
j=1
j
[f
j
(x) f(x)] +
q
j=1
j
f
j
(x), h) +
1
2
|h|
2
_
_
_
(2.60)
est un programme quadratique strictement convexe et sans contrainte, dont la solution
est h =
1
q
j=1
j
f
j
(x). En substituant cette dernire expression h dans (2.59), on
obtient nalement (2.58), et, toujours par dualit, h(x) sexprime laide du multiplicateur
(x) solution de (2.58) :
h(x) =
1
j=1
j
(x)f
j
(x) (2.61)
On a, de plus, les proprits suivantes pour tout x R
n
:
(i) (x) 0 et 0 f(x) si et seulement si (x) = 0
(ii) f
t
(x, h(x)) (x)
1
2
|h(x)|
2
(x).
(iii) Lapplication : R
n
R est continue.
(iv) Lapplication h : R
n
R est continue.
La fonction permet de caractriser les solutions du problme minimax ni ; son va-
luation fournit une direction de recherche h(x) qui est une direction de descente pour f en
x. Il est ainsi possible de construire un algorithme de descente, gnralisant aux problmes
minimax nis la mthode de plus profonde descente avec critre dArmo. Cet algorithme,
dit algorithme PPP (pour Pshenichnyi, Pironneau, Polak, voir [114, Algorithme 2.4.1]),
se prsente de la faon suivante :
Algorithme 1 Problmes minimax nis (Algorithme PPP).
Fixer les paramtres ]0, 1], > 0.
Initialiser le compteur k 0 et xer le point initial x
0
R
n
.
Etape 1. Evaluer f
j
(x
k
) et f
j
(x
k
) pour tout j
p
= 1, . . . , p.
En dduire f(x
k
).
Etape 2. Evaluer la fonction doptimalit
k
(x
k
) en rsolvant le sous-problme qua-
dratique convexe (2.58). En dduire la direction de recherche h
k
h(x
k
) avec
(2.61).
Etape 3. Si
k
= 0, arrt de lalgorithme.
Sinon, dterminer une taille de pas t
k
dans la direction h
k
ralisant au moins la
fraction de la dcroissance de f attendue, autrement dit telle que
f(x
k
+t
k
h
k
) f(x
k
) t
k
k
. (2.62)
Etape 4. Mettre jour x
k+1
x
k
+t
k
h
k
.
Incrmenter le compteur k k + 1. Retourner ltape 1.
On montre que les points daccumulation x
de toute suite innie ditrs de lalgo-

rithme 1 vrient (x
) = 0, et sont donc des solutions locales du problme minimax

ni.
Remarques:
Si on suppose les fonctions f
j
strictement convexes (en plus dtre continment
direntiables), f est elle-mme strictement convexe, et dans ce cas x
k
x
lunique
solution du problme minimax.
Si on suppose de plus les fonctions f
j
deux fois continment direntiables, et quil
existe m > 0, M > 0, m M tels que, pour tout j
p
, pour tout x, h R
n
m|h|
2
_
h,
2
xx
f(x)h
_
M|h|
2
; (2.63)
si on suppose enn que [m, M], et que, ltape 3 de lalgorithme 1, la taille de
pas est dtermine par backtracking de paramtre ]0, 1[, cest--dire
t
k
arg max
lN
_
l
: f(x
k
+
l
h
k
) f(x
k
)
l
k
_
, (2.64)
alors la convergence de la suite des itrs est R-linaire : il existe une constante
C > 0 telle que, pour tout k N
|x
k
x
| C
__
1
m
M
_
k
. (2.65)
Notons que sous une hypothse assez forte portant sur x
, minimum local du pro-

blme minimax ni (x
doit tre un point de Haar de ce problme, voir [114,

Dnition 2.4.18]), la convergence de la suite des itrs de lalgorithme 1 est Q-
quadratique : il existe une constante C 0 et un rang k
0
N tels que, pour tout
k k
0
|x
k+1
x
|
|x
k
x
|
2
C . (2.66)
Signalons enn que la mthode de Newton a galement t adapte aux problmes
minimax. La fonction doptimalit propose dans [115, 114] est cependant beaucoup
plus dicile valuer. Elle sexprime sous forme duale par
(x) = min
p
_
q
j=1
j
_
f(x) f
j
(x)
_
+
1
2
_ q
j=1
j
f
j
(x),
_
q
j=1
2
f
j
(x)
_
1
q
j=1
j
f
j
(x)
_
_
, (2.67)
qui nest plus du tout un programme quadratique convexe.
Lalgorithme 1 a t appliqu dans le cadre des problmes de commande un critre
temporel (chapitre 5). Il a galement t utilis pour rsoudre, sous certaines hypothses,
des problme de stabilisation de systmes (chapitre 3).
Minimax semi-inni : fonction doptimalit et de descente
Les problmes minimax semi-innis appartiennent au domaine plus gnral de lOpti-
misation appel programmation semi-innie. On appelle problme semi-inni (linaire ou
non linaire) tout problme doptimisation sous contraintes, de la forme
min
zR
q
g(z)
tel que h(z, y) 0 pour tout y Y
(2.68)
o Y contient une innit dlments. On retrouve le problme minimax en posant z = [
t
x
]
o t est une variable relle auxiliaire, g(z) = t, Y = 1 Y
1
p Y
p
, et
h(z, y) =
y
1
_
_
x,
_
_
y
2
.
.
.
y
m
y
1
_
_
_
_
t.
La rsolution numrique des problmes semi-innis est aborde en dtails dans [67],
nous en donnons ici un bref aperu; les direntes mthodes procdent toutes par rsolu-
tions successives de sous problmes issus de (2.68), pour lesquels les contraintes sont en
nombre ni. Les approches varient dans la faon de rduire lensemble Y un nombre ni
dlments :
Les mthodes dchange sappuient sur une suite de parties nies Y
k
de Y , qui sont
mises jour chaque itration par ajout et suppression dlments.
Les mthodes de discrtisation (ou gridding) utilisent des discrtisations de plus en
plus nes de Y .
Enn, les mthodes de rduction locale exploitent les maxima (locaux et supposs
en nombre ni) de la fonction y h(z
k
, y) (z
k
dsignant litr courant) an de
formuler un sous-problme avec un nombre ni de contraintes.
Un algorithme pour le problme minimax semi-inni bas sur une technique de dis-
crtisation des ensembles Y
j
est propos dans [113, 114]. Il sappuie sur la thorie des
approximations consistantes, prsentes dans [114], et, pour chaque sous-problme mi-
nimax ni gnr, eectue un pas de descente analogue celui de lalgorithme 1. Nous
prfrons dvelopper ici la mthode introduite dans [6] pour la synthse H
et aisment
gnralisable aux problmes minimax semi-innis (algorithme 2 ci-aprs). Son application
au problme de synthse H
est aborde largement aux chapitres 4 et 5. Cet algorithme

a par ailleurs t utilis pour rsoudre des problmes de synthse H
multibande [10]
et multidisque [9]. Enn, nous proposerons au chapitre 6 dappliquer une mthode de
rduction locale pour la rsolution du problme de synthse H
, chaque sous-problme
tant rsolu par programmation quadratique successive (SQP).
Algorithme 2 Problmes minimax semi-innis (Noll et Apkarian).
Fixer les paramtres ]0, 1], ]0, 1[, > 0.
Initialiser le compteur k 0 et xer le point initial x
0
R
n
.
Etape 1. Dterminer pour tout j
p
= 1, . . . , p un ensemble ni Y
j,k
R
m
j
contenant Y
j
(x
k
).
Etape 2. Evaluer
j
(x
k
, y
j
) et
x
j
(x
k
, y
j
) pour tout j
p
et pour tout y
j
Y
j,k
.
En dduire f(x
k
).
Etape 3. Evaluer la fonction doptimalit
k
(x
k
) en rsolvant le sous-problme
(x
k
) = min
hR
n
max
j
p
max
y
j
Y
j,k
_
j
(x
k
, y
j
) f(x
k
) +
x
j
(x
k
, y
j
), h) +
1
2
|h|
2
_
.
(2.69)
En dduire la direction de recherche h
k
h(x
k
) avec
h(x
k
) = arg min
hR
n
max
j
p
max
y
j
Y
j,k
_
j
(x
k
, y
j
) +
x
j
(x
k
, y
j
), h) +
1
2
|h|
2
_
. (2.70)
Etape 4. Si
k
= 0, arrt de lalgorithme.
Sinon, dterminer une taille de pas t
k
dans la direction h
k
telle que la dcrois-
sance du modle max ni vrie la condition de type Armo
k
max
j
p
max
y
j
Y
j,k
j
(x
k
+t
k
h
k
, y
j
) f(x
k
) t
k
k
. (2.71)
Calculer la dcroissance eective
k
f(x
k
+t
k
h
k
, y
j
) f(x
k
) . (2.72)
Etape 5. Si
k

k
, mettre jour x
k+1
x
k
+t
k
h
k
, incrmenter le compteur k k+1,
et retourner ltape 1.
Sinon, pour tout j
p
, mettre jour Y
j,k
par Y
j,k+1
ni tel que Y
j,k
Y
j,k+1

Y
j
, poser x
k+1
x
k
, incrmenter le compteur k k + 1, et retourner ltape
2.
Deuxime partie
Algorithmes dvelopps et
applications
55
Chapitre 3
Techniques non lisses pour la
stabilisation des systmes linaires
Nous nous intressons dans ce premier dveloppement au problme de la stabilisation
en boucle ferme des systmes LTI (voir section 1.2.2). Si, dans lordre chronologique, nos
premiers travaux ont port sur lapport de lOptimisation non lisse pour la synthse H
([8] et [27], notamment), il apparaissait ncessaire de pouvoir disposer dun correcteur

stabilisant an dinitialiser les algorithmes dvelopps. Nous verrons en eet dans les
chapitres suivants que le systme en boucle ferme est suppos et maintenu stable chaque
itration, que ce soit pour les algorithmes de descente (chapitres 4 et 5) ou pour la mthode
de rduction locale avec programmation quadratique successive (chapitre 6). Le problme
de stabilisation est donc abord ici comme un problme auxiliaire celui de la synthse
H
, mais dont la rsolution conditionne linitialisation de celui-ci. Il nen demeure pas

moins un problme dicile. Dans le cas de la stabilisation par retour dtat, un calcul de
rang permet de vrier lexistence dune solution, elle-mme obtenue par rsolution dune
ingalit matricielle linaire. En dehors de cette situation trs particulire, les mthodes
existantes (par rsolution dune ingalit matricielle bilinaire notamment) noprent que
sur des systmes de petite taille (quelques dizaines dtats au maximum).
Notre approche prsente ici consiste rechercher un correcteur stabilisant en plaant,
de faon itrative, tous les ples de la boucle ferme dans le demi-plan gauche ; nous
procdons pour ce faire par minimisation de labscisse spectrale de la matrice dtat en
boucle ferme.
Poser ainsi directement le problme de la stabilisation ore deux avantages : dune part,
la dimension de lespace de recherche est indpendante du nombre dtats du systme,
puisquelle est rduite au nombre de paramtres libres dnissant le correcteur recherch ;
dautre part, la structure du correcteur peut tre impose par une simple paramtrisation.
Cependant, comme nous lavons not dans la section 1.2.2, labscisse spectrale nest
pas, en gnral, une fonction localement lipschitzienne. La rgularit de en X C
nn
dpend de la multiplicit des valeurs propres de X dites actives, cest--dire telles
que Re = (X). Si celles-ci sont semi-simples
1
, et seulement dans ce cas, alors est
1
Une valeur propre (X) est dite semi-simple si, par dnition, le bloc associ dans la forme
de Jordan de X est diagonal ou, de faon quivalente, si la dimension du sous-espace propre associ
(la multiplicit gomtrique de ) concide avec la multiplicit de dans le polynme caractristique de
57
58 Chapitre 3. Techniques non lisses pour la stabilisation des systmes linaires
localement lipschitzienne en X (voir par exemple [110]). Si, en particulier, les valeurs
propres actives de X sont simples, alors est mme direntiable en X. Dans tous les
autres cas, nest pas localement lipschitzienne en X, et les outils de lanalyse non lisse
sont plus dlicats dnir (voir [41]).
Lalgorithme non lisse prsent dans ce chapitre vient en complment de celui in-
troduit dans [6], bas sur une mthode doptimisation sans drive (dite de recherche
multi-directionnelle) couple un test rigoureux doptimalit. La convergence vers un
correcteur stabilisant ntant garantie ni par lun, ni par lautre (si tant est quun tel
correcteur existe), il nous a paru important dorir un ventail dalgorithmes, capables
ditrer rapidement, et autorisant ainsi la construction dune heuristique de stabilisation.
Elle pourra tre complte par la minimisation dautres critres (non lisses galement) : la
norme H
shifte |.|
,a
(voir [8, section 7.1], labscisse pseudo-spectrale
et la distance
linstabilit .
X (la multiplicit algbrique de )
3.1. Introduction and notations 59
Nonsmooth techniques for stabilizing
linear systems
V. Bompart, P. Apkarian and D. Noll
Article publi dans les actes de American Control Conference,
New York City, 2007.
Abstract
We discuss closed-loop stabilization of linear time-invariant dynamical systems, a problem
which frequently arises in controller synthesis, either as a stand-alone task, or to initialize
algorithms for H
synthesis or related problems. Classical stabilization methods based

on Lyapunov or Riccati equations appear to be inecient for large systems. Recently,
nonsmooth optimization methods like gradient sampling [35] have been successfully used
to minimize the spectral abscissa of the closed-loop state matrix (the largest real part of
its eigenvalues) to solve the stabilization problem. These methods have to address the
nonsmooth and even non-Lipschitz character of the spectral abscissa function. In this
work, we develop an alternative nonsmooth technique for solving similar problems, with
the option to incorporate second-order elements to speed-up convergence to local minima.
Using several case studies, the proposed technique is compared to more conventional
approaches including direct search methods and techniques where minimizing the spectral
abscissa is recast as a traditional smooth nonlinear mathematical programming problem.
3.1 Introduction and notations
Internal stability is certainly the most fundamental design specication in linear control.
Necessary and sucient conditions leading to an ecient algorithmic solution are still not
known [22]. From an algorithmic point of view, the output feedback stabilization problem
is clearly in the class NP and conjectured to be NP-hard (see [24]). Output feedback
stabilizability can be checked via decidability algorithms, like the Tarski and Seidenberg
reduction methods [3], but the computational cost is exponential in the problem size and
becomes rapidly prohibitive as the number of controller parameters increases.
Internal stability could be expressed in terms of a bilinear matrix inequality (BMI).
Solving BMIs globally is algorithmically dicult, and introducing Lyapunov variables to
symmetrize the problem further increases the number of decision parameters and often
leads to numerical ill-conditioning.
A less ambitious line is to address internal stability as a local optimization problem.
Recent approaches using nonsmooth optimization techniques are [33, 35] for stabilization,
and [8, 10, 9, 26] for H
synthesis. In [35] for instance the authors propose to optimize

the spectral abscissa of the closed-loop matrix via specic nonsmooth techniques until a
strictly negative value is obtained. If the method converges to a local minimum with pos-
itive value, indicating failure to solve the control problem, the method has to be restarted
from a dierent initial guess. Our present contribution is also a local optimization tech-
nique, but our method to generate descent steps is new. In particular, in contrast with
[35], our approach is deterministic. We believe that while local optimization techniques do
not provide the strong certicates of global methods, they oer better chances in practice
to solve the stability problem.
The paper is organized as follows. In section 3.2, we recall dierentiability properties
of the spectral abscissa and discuss strategies to minimize it. In section 3.3, we compute
subgradients of the closed-loop spectral abscissa, and specify to cases of practical interest
with structured controllers like PID or others. The computation of descents steps is
explained in section 3.4. It leads to the presentation of the algorithm and its second order
variant in section 3.5. Finally, section 3.6 presents numerical experiments on miscellaneous
automatic control applications, with an example involving structured controllers. For
some of them, the nal spectral properties of the state matrix are analyzed. Our method
is compared to other design methods based on a) direct search algorithm, b) continuous
smooth optimization, and c) nonsmooth optimization.
Matrix notations
In the following (C
nn
, , ))) is the Hilbert space of complex n n matrices with inner
product M, N)) Re tr M
H
N, where M
H
denotes the transconjugate of M. The
induced Hilbert space norm is the Frobenius norm and noted ||. o
n
is the linear subspace
of n n complex symmetric matrices.
For M C
nn
, M _ 0 means that M S
n
is positive semi-denite. The n eigenvalues
of M C
nn
(repeated with multiplicity) are denoted
1
(M), . . . ,
n
(M) and ordered
lexicographically, that is by decreasing real part rst, and next by decreasing imaginary
part in case of equal real parts. When the eigenvalues of M are considered without their
multiplicities, they are denoted
1
(M), . . . ,
q
(M) and ordered in the same manner as
(
i
(M)), with respective algebraic multiplicities n
1
, . . . , n
q
(with n
1
+ + n
q
= n) and
geometric multiplicities p
1
, . . . , p
q
. We say that an eigenvalue
j
(M) is semisimple if
n
j
= p
j
, and is nonderogatory if p
j
= 1. If n
j
> p
j
,
j
(M) is said defective.
In the sequel, (M) denotes the spectral abscissa of M, dened as
(M) max
1jq
Re (
j
(M)) .
Any eigenvalue of M whose real part attains (M) is said active. The set of all active
eigenvalues of M is denoted /(M)
j
(M) : Re (
j
(M)) = (M), the corresponding
active indices sets are
1(M) i N : 1 i n and
i
(M) /(M) ,
and without multiplicity
(M) j N : 1 j q and
j
(M) /(M) .
3.2. Minimizing the spectral abscissa 61
Plant and controller notations
The open-loop system we wish to stabilize is a continuous linear time-invariant plant,
described without loss of generality by the state-space equations
P(s) :
_
x
y
_
=
_
A B
C 0
_ _
x
u
_
(3.1)
where A R
nn
, B R
nm
and C R
pn
. We consider static or dynamic output
feedback control laws of the form u = K(s)y in order to stabilize (3.1) internally. We
suppose that the order of the controller n
K
N is xed. In the case of static feedback
(n
K
= 0), the controller is denoted by K R
mp
. For dynamic controllers dened by the
state-space equations
K(s) :
_
x
K
u
_
=
_
A
K
B
K
C
K
D
K
_ _
x
K
y
_
(3.2)
we use the following standard substitutions in order to reduce to the static feedback case:
K
_
A
K
B
K
C
K
D
K
_
, A
_
A 0
0 0
n
K
_
, B
_
0 B
I
n
K
0
_
, C
_
0 I
n
K
C 0
_
,
n
K
0, n n +n
K
, m m +n
K
, p m +p.
(3.3)
The ane mapping K A +BKC is denoted as A
c
.
3.2 Minimizing the spectral abscissa
We start by writing the stabilization problem as an unconstrained optimization program
min
K|
(A +BKC) (3.4)
where the search space / is either the whole controller space R
mp
, or a subset of R
mp
in those cases where a stabilizing controller with a xed structure is sought.
Closed-loop stability is satised as soon as (A+BKC) < 0, so that the minimization
process can be stopped before convergence. Convergence to a local minimum is important
only in those cases where the method fails to locate negative values < 0. If the process
converges toward a local minimum K
with positive value 0, we know at least that

the situation cannot be improved in a neighborhood of K
, and that a restart away from

that local minimum is inevitable.
Program (3.4) is dicult to solve for two reasons. Firstly, the minimax formulation
calls for nonsmooth optimization techniques, but more severely, the spectral abscissa M
(M) as a function R
nn
R is not even locally Lipschitz everywhere. The variational
properties of have been analyzed by Burke and Overton [41]. In [40] the authors
show that if the active eigenvalues of M are all semisimple (n
j
= p
j
), is directionally
dierentiable at M and admits a Clarke subdierential (M). This property fails in the
presence of a defective eigenvalue in the active set /(M).
Despite the lack of Clarke subdierentiability of the function , we consider program
(3.4) as practically useful, if the following facts are taken into account:
Starting our nonsmooth optimization from a closed-loop matrix A + BK
0
C with
semisimple active eigenvalues, we expect that iterates A+BK
l
C will remain semisim-
ple. Nonderogatory Jordan blocks will only occur at the limit point A +BK
C. If
a negative value (A + BK
l
C) < 0 is obtained, we can stop the procedure at this
iterate, and the question of optimality is irrelevant.
On the other hand, if iterates are not stabilizing, (A + BK
l
C) 0 for all l, then
we need to check whether the limit point A+BK
C is optimal or not. This requires

a nonsmooth stopping test. If the limit point is locally optimal, we know that a
restart of the method away from K
is inevitable. On the other hand, if K
is a
so-called dead point, that is, a limit point which is not locally optimal, then we
should not do a restart, but use nonsmooth decent techniques and keep optimizing
until a local minimum is found.
Several strategies for addressing the nonsmoothness in (3.4) have been put forward: Burke,
Lewis and Overton have extended the idea of gradient bundle methods (see [69] for the
convex case and [80] for the Lipschitz continuous case) to certain non-Lipschitz functions,
for which the gradient is dened, continuous and computable almost everywhere. The
resulting algorithm, called gradient sampling algorithm, is presented in [35] (in the sta-
bilization context) and analyzed in [34, 39] with convergence results. The outcome of
this research is a package HIFOO, which will be included in our tests, see Section 3.6.
The search direction in HIFOO depends on randomly sampled points which gives the
algorithm a nondeterministic aspect. In contrast, Apkarian and Noll propose a composite
approach in [6], combining a direct search method and a nonsmooth strategy for descent
and stopping test. By construction, direct search algorithms require a lot of functions
evaluations, which is demanding if the number of controller decision variable is large.
In response, the present paper proposes an algorithm which uses exclusively nonsmooth
descent steps, allowing to avoid the drawbacks of both approaches.
3.3 Subgradients of the spectral abscissa
3.3.1 Subgradients in state-space
In this section, we suppose that all active eigenvalues of the closed-loop state matrix
A
c
(K) are semisimple, with r = [(K)[ < q distinct active eigenvalues (s = [1(/)[ < n
if counted with their multiplicity). We have /(K) =
1
(A
c
(K)) , . . . ,
r
(A
c
(K)) =
1
(A
c
(K)) , . . . ,
s
(A
c
(K)). The Jordan form J(K) of A
c
(K) is then partly diagonal,
more precisely :
J(K) = V (K)
1
A
c
(K)V (K)
=
_
_
D(K)
J
r+1
(K)
.
.
.
J
q
(K)
_
_
3.3. Subgradients of the spectral abscissa 63
D(K) = diag [
1
(A
c
(K)) , . . . ,
s
(A
c
(K))] is the diagonal part of active eigenval-
ues,
J
j
(K), for r < j q are n
j
n
j
block-diagonal matrices of Jordan blocks, such as
J
j
(K) =
_
_
J
j,1
(K)
.
.
.
J
j,p
j
(K)
_
_
, with
J
j,k
=
_
j
1
.
.
.
.
.
.
.
.
.
1
j
_
_
V (K) = [v
1
(A
c
(K)) , . . . , v
n
(A
c
(K))], where the rst s columns are right eigen-
vectors of A
c
(K) associated with the active eigenvalues (the other columns are
generalized right eigenvectors),
V (K)
1
=
_
_
u
1
(A
c
(K))
H
.
.
.
u
n
(A
c
(K))
H
_
_
,where the rst s rows are left eigenvectors of A
c
(K)
associated with the active eigenvalues (the other rows are generalized left eigenvec-
tors).
We dene U(K) = V (K)
H
, and for 1 j r, V
j
(K) (resp. U
j
(K)) the nn
j
block from
V (K) (resp. from U(K)) composed of the right eigenvectors (resp. of the transconjugate
of the left eigenvectors) associated with
j
.
The function A
c
is Clarke regular at K, as a composition of the ane mapping
A
c
with , which is locally Lipschitz continuous at K. Let
j
/(A
c
(K)) be an active
eigenvalue of A
c
(K), such that
j
=
i
(A
c
(K)) = =
i+n
j
(A
c
(K)). Then the real
matrix
j
(K) = Re
_
B
T
U
j
Y
j
V
H
j
C
T
_
=
_
Re
_
CV
j
Y
j
U
H
j
B
__
T
is a Clarke subgradient of the composite function A
c
at K, where Y
j
_ 0 and Tr(Y
j
) = 1.
Moreover, the whole subdierential ( A
c
)(K) is described by matrices of the form
Y
(K) =

j(A
c
(K))
_
Re
_
CV
j
Y
j
U
H
j
B
__
T
(3.5)
where Y
j
_ 0 and

j(A
c
(K))
Tr(Y
j
) = 1.
Notice that the complex conjugate paired active eigenvalues
j
and
k
=
j
(k ,= j)
share the same closed-loop spectral abscissa subgradient
j
=
k
. The subdierential is
then kept unchanged if the active set only contains the active eigenvalues whose imaginary
part is nonnegative: /(M) =
j
(M) [ Re (
j
(M)) = (M) and Im(
j
(M)) 0.
Remark 1 If the open-loop plant is not controllable, then every uncontrollable mode
l
(A) persists in the closed-loop: for al l controllers K, there exists j such that
l
(A) =
j
(A
c
(K)). Moreover, if this eigenvalue is semisimple and active for A
c
, the asso-
ciated subgradients are null, because U
H
j
B = 0. The case of unobservable modes leads
to the same conclusion, because CV
j
= 0. In this way, whenever an uncontrollable or
unobservable open-loop mode
l
(A) becomes active for the closed-loop spectral abscissa,
we get 0 (A
c
)(K) and then we have local optimality of K. Moreover, the optimality
is global because Re
l
(A) is a lower bound for A
c
.
3.3.2 Subgradients and dynamic controllers
The problem of stabilizing the plant P by dynamic output feedback reduces formally to
the static case, with the substitutions (3.3). Nevertheless, the dynamic case is slightly
more tricky, because the matrices A
K
, B
K
, C
K
and D
K
have to dene a minimal con-
troller realization, both at the initialization stage and at every subsequent iteration of the
algorithm.
As an illustration, if the n
K
-th order (nonminimal) realization of the initial controller
is chosen with B
K
= 0 and C
K
= 0 (neither observable nor controllable) and with
(A
K
) < (A + BD
K
C), it is straightforward to show that the resulting subgradients of
the closed-loop spectral abscissa are convex linear combinations of matrices of the form
j
(K) =
_
_
0 0
0 Re
_
CV
j
Y
j
U
H
j
B
_
T
_
_
where V
j
(resp. U
H
j
) are blocks of right (resp. left) eigenvectors associated with the active
eigenvalues of A + BD
K
C, and Y
j
_ 0, Tr(Y
j
) = 1. As the successive search directions
have the same structure, see (3.13), this results in unchanged A
K
, B
K
, C
K
blocks among
the new iterates. Put dierently, they all represent static controllers.
In order to initialize the descent algorithm with a minimal n
K
-th order controller, and
to maintain this minimality for all subsequent iterates, we use an explicit parametrization
of minimal, stable and balanced systems [107, 43]. With this canonical structure, the
number of free parameters for a n
K
-th order controller with p inputs and m outputs is
n
K
(m +p) +mp.
3.3.3 Subgradients with structured controllers
Formulation (3.4) is general enough to handle state-space structured controllers, such as
decentralized or PID controllers, minimal realizations (see 3.3.2), and others. Let K :
R
k
R
mp
be a smooth parametrization of an open subset / R
mp
, containing state-
space realizations of a family of controllers of a given structure. Then the stabilization
problem can be written as
min
R
k
(A
c
K ()) .
3.4. Descent step and optimality function 65
The Clarke subgradients R
k
of the composite function A
c
K are derived from
(3.5) with the chain rule (see [44, section 2.3])
() = J
vec(K)
()
T
vec ((K ()))
where J
vec(K)
() R
mpk
is the Jacobian matrix of vec(K) : R
k
vec(K()) R
mp
.
3.4 Descent step and optimality function
In order to derive a descent step from the subdierential ( A
c
)(K), we follow a rst-
order step generation mechanism for minimax problems introduced by Polak in [113, 114].
It was described and applied in the semi-innite context of the H
synthesis in [8]. This

descent scheme is based on the minimization of a local and strictly convex rst-order
model (K), which serves both as a descent step generator and as an optimality function.
We rst make the strong assumption that all the eigenvalues of the closed-loop state
matrix A
c
(K) are semisimple. Then, with > 0 xed, we dene
(K) min
H|
max
1jq
max
Y
j
_ 0
Tr(Y
j
) = 1
Re (
j
(A
c
(K))) (A
c
(K)) +
j
(K), H)) +
1
2
|H|
2
(3.6)
where
j
(K), H)) = Tr
_
Re
_
CV
j
Y
j
U
H
j
B
_
H
_
.
Using Fenchel duality for permuting the min and double max operators, we obtain the
dual form of (3.6), where the inner minimization over H becomes unconstrained and can
be computed explicitly, leading to:
(K) = max
j
0
j

j
= 1
max
Y
j
_ 0
Tr(Y
j
) = 1
_
q
j=1
j
Re (
j
(A
c
(K))) (A
c
(K))
1
2
_
_
_
_
_
_
q
j=1
j
(K)
_
_
_
_
_
_
2 _
(3.7)
and we get the minimizer H(K) of the primal formulation (3.6) from the solution
_
_
j
(K)
_
1jq
,
_
Y
j
(K)
_
1jq
_
of the dual expression (3.7) in the explicit form
H(K) =
1
j=1
j
(K) Re
_
CV
j
Y
j
(K)U
H
j
B
_
T
. (3.8)
We recall from [113] the basic properties of and H :
1. (K) 0 for all K /, and (K) = 0 if and only if 0 ( A
c
)(K).
2. If 0 , ( A
c
)(K), then H(K) is a descent direction for the closed-loop spectral
abscissa at K. More precisely for all K:
( A
c
)
t
(K, H(K)) (K)
1
2
|H(K)|
2
(K).
3. The function is continuous.
4. The operator K H(K) is continuous.
Therefore direction H(K) will be chosen as a search direction in a descent-type algorithm
and combined with a line search. The continuity of H () ensures that every accumulation
point

K in the sequence of iterates satises the necessary optimality condition 0 (
A
c
)(

K) (see [8]). It is not clear whether continuity in this sense is still satised when
the hypothesis of semi-simplicity of the eigenvalues is dropped. Notice that even for
semisimple eigenvalues, continuity fails for the steepest descent direction, dened as the
solution of the program
min
|H|1
( A
c
)
t
(K, H) = min
|H|1
max
j(A
c
(K))
max
Y
j
_ 0
Tr(Y
j
) = 1
j
(K), H)) .
This is why steepest descent steps for nonsmooth functions may fail to converge. In our
case this justies the recourse to the quadratic, rst-order model as a descent function.
Moreover, properties 1) and 3) suggest a stopping test based on the value of (K), because
as soon as (K)
(for a small given
> 0), the controller K is in a neighborhood

of a stationary point.
3.5 Nonsmooth descent algorithms
3.5.1 Variant I (rst-order type)
We discuss details of a descent-type algorithm for minimizing the closed-loop spectral
abscissa, based on the theoretical results from the previous section.
For a given iterate K
l
, we have to address rst the practical computation of the
maximizer of the dual form (3.7) of (K
l
). Without any additional hypothesis, it is a
semidenite program (SDP). The matrix variables come from the possible multiplicity
of some eigenvalues of A
c
(K
l
). Fortunately, this situation is unlikely to happen in a
numerical framework, because the eigenvalues and eigenvectors are computed within a
given tolerance. Except in some academic examples (see for instance section 3.6.1), or
for particular values of the controller, coalescence of eigenvalues will not be observed in
practice. Assuming that all the eigenvalues of A
c
(K) are simple, the SDP (3.7) reduces
to a concave quadratic maximization program:
(K) = max
j
0
j

j
= 1
_
_(A
c
(K)) +
n
j=1
j
Re (
j
(A
c
(K)))
1
2
_
_
_
_
_
_
n
j=1
j
(K)
_
_
_
_
_
_
2
_
_ (3.9)
3.5. Nonsmooth descent algorithms 67
where
j
(K) = Re
_
Cv
j
u
H
j
B
_
T
, with v
j
a right eigenvector associated with
j
(A
c
(K)),
and u
H
j
the corresponding left eigenvector (such that u
H
j
v
j
= 1).
To go one step further, we reduce the dimension of the search space. There are n
scalar variables
j
in (3.9), n the order of the open-loop plant, augmented by n
K
, the
order of the controller, in the case of a dynamic controller. For a given ratio [0, 1],
we dene the following enriched set of active eigenvalues
/
(K)
_
j
(A
c
(K)) : (A
c
(K)) Re (
j
(A
c
(K)))

_
(A
c
(K)) min
1in
Re (
i
(A
c
(K)))
_
_
(3.10)
where
(K) = j N [ 1 j n and
j
(A
c
(K)) /
(K) is the corresponding en-

riched active index set. It is clear that /
(K) is nondecreasing on [0, 1], and that

/(K) = /
0
(K) /
(K) /
1
(K) = spec (A
c
(K)) for all [0, 1]. Hence, we have
locally
(A
c
(K)) = max
j
(K)
Re (
j
(A
c
(K))) . (3.11)
By applying the descent function to this local formulation, we nally get the quadratic
program
(K) = max
j
0
j

j
= 1
_
_(A
c
(K)) +
[
(K)[
j=1
j
Re (
j
(A
c
(K)))
1
2
_
_
_
_
_
_
[
(K)[
j=1
j
(K)
_
_
_
_
_
_
2
_
_.
(3.12)
The descent direction H(K) is obtained from the maximizer
_
j
(K)
_
1j[
(K)[
as
H(K) =
1
(K)[
j=1
j
(K) Re
_
Cv
j
u
H
j
B
_
T
. (3.13)
Notice that for = 0 the QP in (3.12) reduces to the steepest descent nding problem
min
j
0
j

j
= 1
_
_
_
_
_
_
[(K)[
j=1
j
(K)
_
_
_
_
_
_
= min
(A
c
)(K)
||
while = 1 reproduces (3.9). The parameter oers some additional numerical exibility,
and allows the weaker asumption that only eigenvalues in /
(K) are simple.

The additional stopping tests in step 5 allows the algorithm to stop when neither
the controller nor the spectral abscissa updates are satisfactory, whereas the stationarity
criterion on the value of is not met in step 2. The situation can occur if some of the
active eigenvalues coalesce:
Algorithm 1 Variant I descent type algorithm for the closed-loop spectral abscissa.
Set the parameters [0, 1], > 0, K
0
/,
,
K
> 0, ]0, 1[. Set the counter
l 0.
1. Compute (A
c
(K
0
)), the enriched active index set
(K
0
) and the corresponding
subgradients
j
(K
0
).
2. Solve (3.12) for K = K
l
and get the search direction H(K
l
) from (3.13).
If (K
l
)
then stop.
3. Find a step length t
l
> 0 satisfying the Armo line search condition
(A
c
(K
l
+t
l
H(K
l
))) (A
c
(K
l
)) +t
l
(K
l
)
4. Set K
l+1
K
l
+t
l
H(K
l
).
Compute (A
c
(K
l+1
)), the extended active index set
(K
l+1
) and the correspond-
ing subgradients
j
(K
l+1
).
5. If (A
c
(K
l
)) (A
c
(K
l+1
))
(1 + (A
c
(K
l
))) and |K
l
K
l+1
|
K
(1 +|K
l
|)
then stop.
Otherwise set l l + 1 and go back to 2.
In case of a multiple but semisimple eigenvalue, the Clarke subdierential still exists
(see 3.3.1), but is underestimated in the algorithm because we made the hypothesis
that active eigenvalues were simple in order to reduce the SDP to a QP. More pre-
cisely in that case
( A
c
)(K
l
) = co
_
Re
_
CV
j
Y
j
U
H
j
B
_
T
: j (K
l
), Y
j
_ 0 and Tr(Y
j
) = 1
_
co
_
Re
_
Cv
j
u
H
j
B
_
T
: j (K
l
)
_
In case of a defective eigenvalue, the Clarke subdierential no longer exists.
3.5.2 Variant II (second-order type)
In the optimality function (3.6) the parameter acts as an estimate of the average of the
curvatures of Re
j
A
c
. If second order information is available, it may therefore be
attractive to replace the scalar in (3.6) by Hessian matrices. Polak [114] extends the
Newton method to min-max problems, but the corresponding dual expression for (K
l
)
does no longer reduce to a quadratic program like (3.9). We propose a dierent approach
here which is based on a heuristic argument. The quadratic term of

is weighted by a
matrix Q
l
, which is updated at each step using a second-order model of A
c
. We suggest
3.6. Numerical examples 69
a quasi-Newton method based on the new optimality function

at iteration l 1:
(K
l
) min
H|
max
j
(K
l
)
max
Y
j
_ 0
Tr(Y
j
) = 1
_
Re (
j
(A
c
(K
l
))) (A
c
(K
l
)) +
j
(K), H))
+
1
2
vec(H)
T
Q
l
vec(H)
_
. (3.14)
The matrix Q
l
is a positive-denite, symmetric mp mp matrix, updated with the sym-
metric rank-two BFGS update
Q
l+1
Q
l
+
y
l
y
T
l
y
T
l
s
l
Q
l
s
l
s
T
l
Q
l
s
T
l
Q
l
s
l
(3.15)
where s
l
= vec(K
l+1
K
l
), and y
l
= vec(g
l+1
g
l
), with g
l
the subgradient of minimal
norm in ( A
c
)(K
l
), by analogy with the gradient for smooth functions.
The dual form of (3.14) is then the convex QP
(K
l
) = max
j
0
j

j
= 1
_
_
(A
c
(K)) +
[
(K
l
)[
j=1
j
Re (
j
(A
c
(K
l
)))
1
2
vec
_
_
[
(K
l
)[
j=1
j
(K
l
)
_
_
T
Q
1
l
vec
_
_
[
(K
l
)[
j=1
j
(K
l
)
_
_
_
_
(3.16)
and the vectorized descent direction derived from the optimal
_
j
(K
l
)
_
convex coecients
is:
vec
_

H(K
l
)
_
= Q
1
l
[
(K
l
)[
j=1
j
(K
l
) vec (
j
(K
l
)) . (3.17)
The resulting algorithm is given below (Algorithm 2). Notice that the quadratic
subproblem (3.16) invokes the inverse of the BFGS update Q
l
. Inversion can be avoided
by directly updating Q
1
l
at step 5, with the update formula
Q
1
l+1
= Q
1
l
+
(s
l
Q
1
l
y
l
)s
T
l
+s
l
(s
l
Q
1
l
y
l
)
T
s
T
l
y
s
y
T
l
(s
l
Q
1
l
y
l
)s
l
s
T
l
(s
T
l
y
l
)
2
(3.18)
=
_
I
s
l
y
T
l
s
T
l
y
l
_
Q
1
l
_
I
s
l
y
T
l
s
T
l
y
l
_
T
+
s
l
s
T
l
s
T
l
y
l
3.6 Numerical examples
In this section we test our nonsmooth algorithm on a variety of stabilization problems
from the literature. We use variants I and II of the descent algorithm in the following
applications, with the default parameters values (unless other values are specied):
= 0.8, = 0.1,
= 10
5
,
= 10
6
,
K
= 10
6
and = 0.9
Algorithm 2 Variant II descent type algorithm for the closed-loop spectral abscissa.
Set the parameters [0, 1], Q
0
~ 0, K
0
/,
,
K
> 0, ]0, 1[.
Set the counter l 0.
1. Compute (A
c
(K
0
(K
0
), the corresponding sub-
gradients
j
(K
0
), and the subgradient of minimal norm g
0
in ( A
c
)(K
0
).
2. Solve (3.16) for K = K
l
and get the search direction

H(K
l
) from (3.17).
If

(K
l
)
then stop.
3. Find a step length t
l
> 0 satisfying the Armo line search condition
_
A
c
_
K
l
+t
l

H(K
l
)
__
(A
c
(K
l
)) +t
l
(K
l
)
(try t
l
= 1 rst)
4. Set K
l+1
K
l
+t
l

H(K
l
).
Compute (A
c
(K
l+1
(K
l+1
), the corresponding
subgradients
j
(K
l+1
), and the subgradient of minimal norm g
l+1
in (A
c
)(K
l+1
).
5. Compute Q
1
l+1
with the BFGS update formula (3.18).
6. If (A
c
(K
l
)) (A
c
(K
l+1
))
(1 + (A
c
(K
l
))) and |K
l
K
l+1
|
K
(1 +|K
l
|)
then stop.
Otherwise set l l + 1 and go back to 2.
We compare the performance of our method with that of other minimization algorithms,
namely multi-directional search (MDS), two algorithms implemented in the Matlab Op-
timization Toolbox, and the Matlab package HIFOO [32].
Multidirectional search (MDS) belongs to the family of direct search algorithms [136].
This derivative-free method explores the controller space via successive geometric trans-
formations of a simplex (contraction, expansion and reection). Its convergence to a local
minimum is established for C
1
-functions, but nonsmoothness can make it converge to a
nondierentiable and nonoptimal point [137], called a dead point. In [6] we have shown
how to combine MDS with nonsmooth descent steps in order to avoid this diculty and
guarantee convergence. Here we use MDS with two stopping criteria: the rst is based on
the relative size of the simplex:
max
i
|v
i
v
0
|
max(1,|v
0
|)
<
1
where v
i
are the vertices of the simplex,
and v
0
is the vertex of smallest function value of the objective f. The second stopping test
quanties the function value variations over the vertices:
max
i
f(v
i
)f(v
0
)
max(1,[f(v
0
)[)
<
2
. Experiments
were performed with
1
=
2
= 10
6
, and two simplex shapes (MDS 1: right-angled, MDS
2: regular).
Secondly, two Matlab Optimization Toolbox functions have been tested, one designed
for general constrained optimization (fmincon), the second suited for min-max problems
(fminimax). Both functions are essentially based on SQP algorithm with BFGS, line
search and an exact merit function, see [96]. The nite constrained cast of (3.4) passed
to fmincon is
min
t R
K /
t subject to Re (
i
(A +BKC)) t for all 1 i n
Equivalently, the min-max formulation used for fminimax is
min
K|
max
1in
Re (
i
(A +BKC))
Clearly here we make the implicit assumption that all the eigenvalues are simple in order
to work with smooth constraints or maximum of smooth functions, which is required by
SQP. Our testing will show whether the toolbox functions run into diculties in those
cases where this hypothesis is violated.
Finally, we use the Matlab package HIFOO (version 1.0). As discussed in [32], the
underlying algorithm consists in a succession of (at most) three optimization phases:
BFGS, local bundle (LB) and gradient sampling (GS). By virtue of its probabilistic nature,
HIFOO does not return the same nal controller even when started from the same initial
guess. We noticed that for the same reason the number of evaluations of the spectral
abscissa could greatly change from one run to the other, even when all the settings are
unchanged. This probabilistic feature of HIFOO is inherent to the multiple starting point
strategy (by default, 3 random controllers, in addition to the user input), and to the
gradient sampling algorithm itself. The rst stabilizing controller is obtained with the
parameter +, whereas the nal one is with s. The iteration number of each stage is
given as BFGS+LB+GS.
As we have indicated before, nal controllers are not necessarily local minimizers of
A
c
. The optimality is less import in practice than in design problems with performance
criteria. However, we discuss the status of every termination case in terms of active
eigenvalues multiplicity, and of associated eigenspaces dimension.
3.6.1 Academic test
We begin with an academic 2nd order SISO plant from the output feedback literature [89],
cited in [85] (example NN2), and representing a damped linear oscillator. The open-loop
plant is marginally unstable, with two pure imaginary modes (
1
= i and
2
= i). The
closed-loop state matrix is
A
c
(K) = A +BKC =
_
0 1
1 K
_
For K R [2, 2] (resp. K ]2, 2[), A
c
(K) is diagonable with two real eigenvalues
=
K
2

K
2
4 (resp. two complex conjugate eigenvalues =
K
2
i
4 K
2
), and
thus A
c
is Clarke regular. This is not the case at K = 2 (resp. K = 2), where the
eigenvalues of A
c
(K) coalesce into the double root = 1 (resp. = 1). For these two
controller values, A
c
looses Lipschitz continuity, and the Clarke subdierential is not
dened.
Nonsmooth optimization algorithm (variant I)
We have A
c
(K) =
K
2
+
_
[K
2
4]
+
. The global minimum is attained at K = 2, for
which the based optimality certicate is unfortunately useless. Yet for K / 2, 2,
the stabilization algorithm provides a descent direction and allows to compute a descent
step. We have tried various initial controller values K
0
, taken on dierent pieces of the
spectral abscissa curve.
For every given K
0
, the algorithm converges to the global minimum. As soon as
the relative step length and the relative spectral abscissa decrease are small enough, the
algorithm stops.
K
0
-5 0 2 5
rst < 0 2.09 10
1
2.50 10
1
2.50 10
1
5.73 10
1
(iter.) (0) (1) (1) (3)
nal 1.00 1.00 1.00 1.00
(iter.) (30) (17) (17) (17)
fun. evaluations 117 71 72 70
nal 1.25 10
1
1.25 10
1
1.25 10
1
1.25 10
1
nal extended set size 1 1 1 1
nal active set size 1 1 1 1
Table 3.1: Damped linear oscillator stabilization, nonsmooth algorithm variant I, various
initial controllers K
0
.
Surprisingly, even from K
0
= 2, our nonsmooth stabilization algorithm converges to
the global minimum, even though Clarke subgradients are not well dened at the rst
iteration. Figure 3.1 clearly shows the nonsmooth and even non-Lipschitz behavior of the
closed-loop eigenvalues for K = 2.
5 0 5
5
4
3
2
1
0
1
2
3
4
5
K
A
c

(
K
)
closedloop eigenvalues
5 0 5
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Re(
i
A
c
(K))
I
m
(
A
c

(
K
)
)
closedloop eigenvalues locus
Figure 3.1: Closed-loop eigenvalues variation around K = 0.
Remark 2 For K 2, 2, the double eigenvalue of A
c
(K) is defective but nonderoga-
tory. Thus, the closed-loop spectral abscissa is subdierentially regular, and the set of
regular subgradients of A
c
can be obtained explicitly from the Jordan decomposition,
as shown in [41]. We get

( A
c
) (2) =
_
,
1
2
_
and

( A
c
) (2) =
_
1
2
, +
_
. Even
though the regular subdierential could be used for characterizing the sharp local minimiz-
ers of A
c
(see [33]), the detection and the computation of nontrivial Jordan blocks
associated with active eigenvalues in A
c
are numerical chal lenges. Except for simplistic
examples with exact and integer data, the variational informations for the defective and
nonderogatory case are practically out of reach.
Other algorithms
In table 3.2, we give the results obtained with MDS, Matlab Optimization Toolbox, and
HIFOO with the initialization K
0
= 0.
algorithm MDS 1 fmincon fminimax HIFOO
rst < 0 1.00 - 0.25 0.25
(iter.) (1) (-) (1) (1+0+0)
nal 1.00 0 9.95 10
1
1.00
(iter.) (21) (198) (5) (8+2+5)
Table 3.2: Damped linear oscillator stabilization, other algorithms, K
0
= 0.
We observe that fmincon fails to converge after 200 function evaluations, and returns
K = 1.84 10
14
. In a second testing, we try to initialize the algorithms with the
problematic controller value K
0
= 2 (see table 3.3).
algorithm MDS 1 fmincon fminimax HIFOO
rst < 0 1.00 - - 1.49 10
8
(iter.) (2) (-) (-) (1)
nal 1.00 1.00 1.00 1.00
(iter.) (21) (100) (100) (76+2+5)
Table 3.3: Damped linear oscillator stabilization, other algorithms, K
0
= 2.
HIFOO nds the global minimum at the BFGS stage, tries in vain to improve it
with local bundle and gradient sampling iterations, but doesnt give any local optimality
certicate at the end (local optimality measure is 1.5 10
1
).
fmincon and fminimax got stuck at K = 2, even after 200 function evaluations. MDS
converged in both cases to the minimizer K = 2, with fewer function evaluations than
the nonsmooth algorithm: the very small dimension of the search space is favorable to
this direct search technique. In contrast, HIFOO needs many more spectral abscissa calls
in order to complete the optimization, without any optimality certicate, as above.
Remark 3 In the one-dimension case here, right-angled and regular simplices are inter-
vals, the shape of the simplex has no inuence on the MDS algorithm.
3.6.2 Transport airplane
The linearized plant of 9th-order describes the longitudinal motion of a transport airplane
at given ight conditions (system AC8 from [83], see also [56]). The open loop is unstable,
with spectral abscissa = 1.22 10
2
, attained by a simple, real mode: the composite
function A
c
is then dierentiable at K
0
= D
K
0
= 0 (all the other modes are stable,
with real parts between 5.00 10
1
and 3.18 10
2
).
Nonsmooth optimization algorithm (variant I)
In table 3.4, we show the inuence of the ratio (see (3.10)) on the nonsmooth algorithm
(variant I here). Notice that the rst case ( = 0) is steepest descent.
In each of the rst two cases, the nal value of is not reliable for optimality, because
A
c
looses Clarke regularity. This is suggested by the graphical representation of
A
c
around the nal controller, along a chosen direction (see gure 3.2 for the 0.1 %
case # 1 2 3 4
ratio 0 % 0.1 % 1 % 2 %
rst < 0 7.07 10
2
1.07 10
2
1.18 10
2
1.05 10
2
(iter.) (1) (1) (1) (1)
nal 1.15 10
1
1.43 10
1
1.30 10
1
4.45 10
1
(iter.) (20) (27) (16) (9)
nal 1.54 10
2
1.30 10
1
3.10 10
1
5.60 10
17
nal extended set size 1 2 4 4
nal active set size 1 1 1 2
Table 3.4: Transport airplane stabilization, nonsmooth algorithm variant I, various active
set extensions.
enlargement case), and further conrmed by inspection of closed-loop modes: two complex
conjugate eigenvalues with small imaginary parts are active (
1
= 1.4310
1
+1.1510
4
i
and
2
=

1
) and they are very near from coalescence into a defective real eigenvalue.
0.05 0 0.05
0.15
0.14
0.13
0.12
0.11
0.1
0.09
K
A
c

(
K
+
K

H
)
closedloop spectral abscissa
0.25 0.2 0.15 0.1
0.1
0.08
0.06
0.04
0.02
0
0.02
0.04
0.06
0.08
0.1
Re(
i
A
c
(K+K H))
I
m
(
A
c
(
K
+
K

H
)
)
(3 largest real parts eigenvalues)
Figure 3.2: Closed-loop eigenvalues variation around nal K along H = [1, 0, 0, 0, 0]
(0.1 % enlargement).
The fourth case is more favorable. The enlargement of /
(K) generates better descent

directions for A
c
and allows longer descent steps and fewer iterations. The nal value
of is close to zero, indicating local optimality. There are three active eigenvalues at
the last iteration: two of them are complex conjugate (
1
= 4.45 10
1
+ 4.40 10
3
i
and
2
=

1
), the other one is real (
3
= 4.45 10
1
). We notice that these three
modes come directly from the plant, and are not controllable. This is conrmed by the
associated closed-loop subgradients,
1
=
2
0 and
3
0, leading to a singleton
subdierential ( A
c
)(K
9
) = 0. The nal point is then smooth, in spite of multiple
active eigenvalues, and the uncontrollability of the active modes gives a global optimality
certicate (see Remark 1, section 3.3.1). As an illustration, see gure 3.3.
0.05 0 0.05
0.445
0.44
0.435
0.43
0.425
0.42
0.415
0.41
0.405
0.4
0.395
K
A
c

(
K
+
K

H
)
0.55 0.5 0.45 0.4
0.25
0.2
0.15
0.1
0.05
0
0.05
0.1
0.15
0.2
0.25
Re(
i
A
c
(K+K H))
I
m
(
A
c
(
K
+
K

H
)
)
(6 largest real parts eigenvalues)
3 uncontrollable modes
Figure 3.3: Closed-loop eigenvalues variation around nal K along H = [1, 0, 0, 0, 0]
(2 % enlargement).
Other algorithms
Here the algorithms were initialized with K
0
= 0, and the multidirectional search (MDS)
algorithm was applied with two dierent shapes of simplices in the controller space.
algorithm MDS 1 MDS 2 fmincon fminimax HIFOO
rst < 0 1.09 10
1
1.04 10
2
4.45 10
1
1.13 10
2
2.62 10
2
(iter.) (3) (7) (13) (1) (1+0+0)
nal 4.21 10
1
1.57 10
1
4.45 10
1
4.45 10
1
2.31 10
1
(iter.) (36) (37) (13) (25) (396+3+2)
fun. eval. 366 376 32 131 1140
Table 3.5: Transport airplane stabilization, other algorithms.
MDS becomes very greedy in function evaluations, as the controller size increases in
this example. Moreover, the global minimum is not found, either because of an unsuc-
cessful local minimum, or a dead point.
Both Matlab functions return the global minimum, after very few iterations for fmincon.
HIFOO terminates far from the global minimum, because slow convergence occurs:
numerous BFGS iterations (99) are needed for each of the four initial controllers (K
0
= 0
and three perturbed K
0
). The nal optimality measure is 5.28 10
4
.
3.6.3 VTOL helicopter
This model (HE1 from [85]) with four states, one measurement and two control variables,
describes the longitudinal motion of a VTOL (Vertical Take-O and Landing) helicopter,
at given ight conditions. The open-loop spectral abscissa is = 2.76 10
1
, attained by
two complex conjugate eigenvalues. All the open-loop eigenvalues are simple.
Nonsmooth optimization algorithm (variants I and II)
Algorithm variant I II (with BFGS)
rst < 0 6.16 10
2
6.16 10
2
(iter.) (1) (1)
nal 2.39 10
1
2.47 10
1
(iter.) (216) (26)
fun. eval. 796 90
nal 9.77 10
6
1.70 10
6
Table 3.6: VTOL helicopter stabilization, nonsmooth algorithm, variants I and II.
Using variant I, the closed-loop becomes stable after the rst iteration, and the spectral
abscissa decreases slowly until satisfaction of the local optimality stopping test. This slow
convergence strongly calls for variant II of our nonsmooth descent algorithm, which nds
a lower closed-loop spectral abscissa with much less iterations. For both cases, the value
close to 0 of indicates local optimality.
The nal closed-loop spectrum at convergence obtained by algorithm variant I is
=
_
2.39 10
1
5.76 10
1
i, 2.39 10
1
, 7.91 10
1
_
,
and the subgradients associated with
1
= 2.39 10
1
+ 5.76 10
1
i and with
3
=
2.39 10
1
are, respectively,
1
=
2
=
_
1.26 10
1
+2.92 10
2
_
,
3
=
_
+6.26 10
2
1.55 10
2
_
.
Convergence analysis is favorable for our algorithm, because the nonsmoothness comes
from several simple active eigenvalues for the closed-loop spectral abscissa: the Clarke
subdierential is then well dened and the value of (K) is reliable as an optimality
criterion.
A plot in the neighborhood of the convergence point of variant I (see gure 3.4) clearly
shows the nonsmoothness at optimality.
2 2.5 3 3.5 4
10.5 11 11.5 12 12.5
0.24
0.22
0.2
0.18
0.16
0.14
0.12
0.1
0.08
0.06
K
1
around variant I local minimum
K
2
(
A
+
B
[
K
1
;
K
2
]
C
)
Figure 3.4: Closed-loop spectral abscissa around optimal K of variant I algorithm.
Other algorithms
The same closed-loop spectral abscissa is found by MDS (with regular simplex shape),
the Matlab routines, which are very ecient in this example, and by HIFOO (with nal
local optimality measure 9.87 10
4
). Notice that iterates of fmincon become feasible
only at the last iteration, a classical feature of SQP algorithms.
rst < 0 1.23 10
1
9.95 10
2
2.47 10
1
5.09 10
2
6.16 10
2
(iter.) (1) (1) (36) (1) (1+0+0)
nal 1.35 10
1
2.47 10
1
2.47 10
1
2.47 10
1
2.46 10
1
(iter.) (15) (58) (36) (36) (116+0+0)
fun. eval. 63 235 73 73 216
Table 3.7: VTOL helicopter stabilization, other algorithms.
The same closed-loop spectral abscissa is found by MDS (with regular simplex shape),
the Matlab routines, which are very ecient in this example, and by HIFOO (with nal
local optimality measure 9.87 10
4
). Notice that iterates of fmincon become feasible
only at the last iteration, a classical feature of SQP algorithms.
3.6.4 B-767 airplane
Our last example taken from [85] (system AC10) is of higher order (55 states), with two
controlled inputs and two measured outputs. The state-space data describe a modied
Boeing B-767 at utter condition (see [46]). The open-loop is unstable, but the only active
eigenvalues of A for the spectral abscissa are
1
= 1.015 10
1
and
1
, with multiplicity
one.
Nonsmooth optimization algorithm (variants I and II)
We use algorithm variants I and II, with all gains of the initial controller set to 0. The two
versions of our algorithm stabilize the plant after a single iteration. If the optimization
is continued, variant II gives fast convergence to a local minimum (certied by the small
value of ). Variant I is slower here.
Algorithm variant I II (with BFGS)
rst < 0 2.37 10
2
2.36 10
2
(iter.) (1) (1)
nal 7.99 10
2
3.50 10
2
(iter.) (99) (29)
fun. eval. 387 111
nal 8.00 10
4
8.70 10
6
Table 3.8: B-767 airplane stabilization, nonsmooth algorithm, variants I and II.
Other algorithms
See table 3.9.
rst < 0 3.59 10
3
2.35 10
2
2.47 10
2
2.36 10
2
2.34 10
2
(iter.) (2) (13) (3) (1) (1+0+0)
nal 3.23 10
2
3.54 10
2
3.44 10
2
5.24 10
2
3.62 10
2
(iter.) (60) (100) (9) (15) (132+3+15)
fun. eval. 485 805 21 31 1858
Table 3.9: B-767 airplane stabilization, other algorithms.
3.6.5 PID controllers
As our algorithm can handle controller structure (see 3.3.3), it oers an interesting frame-
work for PID controller design, particularly attractive for MIMO plants where very few
generic tuning techniques are available. In this example, we seek a 3 input, 3 output
stabilizing PID controller for an open-loop marginally unstable aircraft model (AC2 from
[85]), given as:
K(s) = K
P
+
1
s
K
I
+
s
1 + s
K
D
where K
P
, K
I
, K
D
R
22
and > 0. The algorithm is initialized with K
P
= 0,
K
I
= K
D
= I
3
and = 10
3
. The resulting closed-loop is unstable ((A
c
(K)) = 8.06).
The algorithm (variant I) nds a stabilizing PID controller after 2 iterations. It stops
after 42 iterations, with:
K
P
=
_
_
72.23 4.62 28.69
17.89 3.93 33.70
33.65 13.17 114.77
_
_ ,
K
I
=
_
_
41.38 3.80 102.45
23.93 3.82 42.75
11.79 18.42 20.37
_
_ ,
K
D
=
_
_
0.91 0 0
0 0.91 0
0 0 0.91
_
_, = 9.55 10
4
The nal closed-loop spectral abscissa is = 6.03 10
1
. Notice that K
D
and are
nearly unchanged.
3.7 Conclusion
Formulated as an optimization program, static or xed-structure output feedback sta-
bilization has been solved for several case studies from the literature. The proposed
nonsmooth algorithm addresses the nonsmoothness of the spectral abscissa and generates
successive descent steps. Even if the theoretical assumption of semisimple active eigen-
values may seem restrictive, the experimental results show that very few nonsmooth steps
will generally yield a stabilizing static controller. Our framework is generic enough to
handle realistic stabilization problems with structured compensators. The two proposed
variants are deterministic and numerically ecient, with signicantly fewer evaluations
of the spectral abscissa than MDS or HIFOO. Finally, although BFGS is designed for
smooth optimization, we noticed that it performs quite well for static output feedback
stabilization (in variant II, in the initial phase of HIFOO and in the Matlab optimization
tools), agreeing with [32].
Chapitre 4
Optimisation non lisse pour la
commande structure, application au
loop-shaping par correcteur de type
PID
Lalgorithme non lisse prsent dans le chapitre prcdent pour la stabilisation a t
conu originellement dans le but de rsoudre le problme de la synthse H
(voir [8]). Il
est plus dlicat mettre en uvre dans ce contexte, car le problme minimax sous-jacent
est semi-inni. Il est aussi plus adapt, car, sans hypothse particulire, la norme H
du
transfert en boucle ferme est localement lipschitzienne, et admet donc en tout point un
sous-direntiel au sens de Clarke. Comme pour la stabilisation, notre approche par op-
timisation locale permet de prendre en compte en amont, cest--dire dans toute la phase
de synthse du correcteur, une structure particulire de la boucle ferme ou de la loi de
commande elle-mme. Elle peut tre dordre x, ou suivre un certain formalisme (PID,
ples/zros, etc.). Ces spcications peuvent gnralement se traduire par une paramtri-
sation continment direntiable du correcteur qui, grce la rgle de direntiation en
chane pour le sous-direntiel de Clarke, sintgre dans le processus doptimisation.
Nous proposons dans larticle qui suit dillustrer la exibilit de cette approche tra-
vers la commande par loop-shaping dun racteur chimique, laide dun correcteur de
type PID. Dans ce type dapplication, des spcications frquentielles portent sur le prol
des valeurs singulires en boucle ouverte. Ces spcications sont introduites par lajout
de fonctions de pondration dynamiques en entre et en sortie, ce qui a pour eet de
faire crotre articiellement lordre du systme commander. La synthse de correcteurs
dordre rduit (et, plus gnralement, de structure xe) est donc particulirement sou-
haitable dans le cadre du loop-shaping. Grce lapport thorique de lOptimisation non
direntiable, nous traitons directement ce problme de commande structure dans les-
pace des paramtres du correcteur PID. A la convergence de lalgorithme de descente,
loptimalit locale vis vis de la norme H
est garantie.
Nous montrons aussi quil est possible de tester loptimalit locale dun correcteur
PID donn, pour le critre H
. Nous vrions ainsi, en particulier, que les mthodes

heuristiques procdant par des rsolutions alternes de programmes semi-dnis convexes
81
82 Chapitre 4. Optimisation non lisse pour la commande structure
(mthodes dites de descentes coordonnes) sarrtent en des points non optimaux. Notre
algorithme de descente non lisse permet den poursuivre loptimisation, et converge avec
une norme H
localement minimale (cest--dire une marge de stabilit =

1
locale-
ment maximale).
83
Nonsmooth structured control design
with application to PID loop-shaping
of a process
P. Apkarian, V. Bompart and D. Noll
Article publi dans International Journal of Robust and Nonlinear Control,
17(14) :1320-1342, 2007.
Abstract
Feedback controllers with specic structure arise frequently in applications because they
are easily apprehended by design engineers and facilitate on-board implementations and
re-tuning. This work is dedicated to H
synthesis with structured controllers. In this

context, straightforward application of traditional synthesis techniques fails, which ex-
plains why only a few ad-hoc methods have been developed over the years. In response,
we propose a more systematic way to design H
-optimal controllers with xed struc-

ture using local optimization techniques. Our approach addresses in principle all those
controller structures which can be built into mathematical programming contraints. We
apply nonsmooth optimization techniques to compute locally optimal solutions, and pro-
vide practical tests for descent and optimality. In the experimental part we apply our
technique to H
loop-shaping PID controllers for MIMO systems and demonstrate its

use for PID control of a chemical process.
Notation
Let R
nm
be the space of nm matrices, equipped with the corresponding scalar product
X, Y ) = Tr(X
T
Y ), where X
T
is the transpose of the matrix X, Tr (X) its trace. For
complex matrices X
H
stands for its conjugate transpose. For Hermitian or symmetric
matrices, X ~ Y means that X Y is positive denite, X _ Y that X Y is positive
semi-denite. We use the symbol
1
to denote the maximum eigenvalue of a symmetric or
Hermitian matrix. Given an operator T, T
is used to denote its adjoint operator on the

appropriate space. The notation vec applied to a matrix stands for the usual column-wise
vectorization of a matrix. The operator diag applied to a matrix produces a vector whose
entries are the diagonal elements of the matrix. We use concepts from nonsmooth analysis
covered by [44]. For a locally Lipschitz function f : R
n
R, f(x) denotes its Clarke
subdierential at x.
4.1 Introduction
Considerable eorts have been made over the past two decades to develop new and pow-
erful control methodologies. H
synthesis [47] is certainly the most prominent outcome

of this search. In spite of its theoretical success, it turns out that structured controllers
such as PID, lead-lag, observed-based, and others, are still preferred in industrial control.
The reason is that controllers designed with modern control techniques are usually of high
order, dicult to implement and often impossible to re-tune in case of model changes.
But those are precisely the properties which make structured controllers so popular for
practitioners. Easy to implement and to understand, and easy to re-tune whenever per-
formance or stability specications change. The trade-o may be roughly described as
high performance combined with low exibility versus lower performance combined with
high exibility. The question we ask here is how the performance properties of structured
controllers may be improved.
Structured control design is generally a dicult problem. Even the simple static output
feedback stabilization problem is known to be NP-hard [24]. Due to their importance for
practice, a number of innovative techniques and heuristics for structured control have been
proposed in the literature. Some authors use branch-and-bound techniques to construct
globally optimal solution to the design problem [17]. In the same vein, Wong and Bigras
[139] propose evolutionary optimization to reduce the computational overhead, while still
aiming at globally optimal solutions. These approaches are certainly of interest for small
problems, but quickly succumb when problems get sizable.
A fairly disparate set of heuristic techniques for structured control design was de-
veloped in the realm of linear matrix inequalities (LMIs) [30]. Alternating projection
techniques were proposed in [62] for static controller and for the more specic structured
control design problem in [48]. In the same vein, coordinate descent schemes were pro-
posed in [60, 74, 64]. In [65], the authors suggest using a BMI formulation for solving
a variety of problems including structured control. These techniques may be useful in
practice, but they bear the risk of missing a local solution, because convergence to a local
minimum is rarely ensured. Iterative solving of SDPs based on successive linearizations
is yet another idea, but often leads to prohibitive running times. In [19], 2 hours cputime
were necessary to compute a decentralized PID controller for a 22 process on a Pentium
II 333 MHz computer. Even longer cputimes are reported in [57] for medium size PID
design problems.
A relatively rich literature addresses specic controller structures such as decentralized
or PIDs. In [101], Miyamoto and Vinnicombe discuss a coordinate scheme for H
loop-
shaping with decentralized constraints. In [132], again in the loop shaping context, the
authors adopt a truncation procedure to reduce a full-order controller to a PID controller.
Those are heuristic procedures, because closed-loop performance is not necessarily inher-
ited by the nal controller. In [123], Saeki addresses sucient conditions under which
PID synthesis reduces to solving LMIs.
In a recent interesting work, Rotkowitz and Lall [122] fully characterize a class of
problems for which structured controller design can be solved using convex programming.
They introduce the concept of quadratic invariance and show that for problems having
this property, optimal structured controllers may be eciently synthesized. In a dierent
4.2. Nonsmooth H
design technique 85
but related work [140], the authors identify various control structures that are amenable
to convex optimization by an appropriate choice of coprime factors in the Youla-Kucera
parametrization [29]. A similar analysis is made by Scherer in [126] both for structured
controller design and multi-objective H
control. Unfortunately, these concepts and tools

only apply to very particular problem classes and controller patterns and do not easily
lend themselves to generalization when ner controller structures are required.
T
wz
(K) :=
G
K
y u
w
z
Figure 4.1: Standard interconnection.
In our opinion local optimization is the approach best suited for these dicult design
problems. It should whenever possible be used in tandem with heuristic methods, as
those may be useful to compute good initial points for the optimization. We mention
that early approaches to structured design based on tailored optimization techniques can
be traced back to the work of Mkil and Toivonen [93] for parametric LQ problems, or
Polak and Wardi [117] for problems with frequency domain singular value inequalities.
In the latter reference, many design problems are recognized as nondierentiable, and in
consequence, techniques from nonsmooth analysis are employed. More recently, we have
used nonsmooth analysis to fully caracterize the subdierential properties of closed-loop
mappings of the form |.|
T
wz
acting on the controller space, where T
wz
(K) denotes
the closed-loop transfer function from w to z at a given controller K, gure 4.1. These
results are used to develop nonsmooth descent algorithms for various design problems
[8, 9, 10]. Here we extend our results to structured controller design and elaborate the
case of MIMO PID controllers. We then demonstrate how the proposed technique can be
used in the context of PID loop-shaping as introduced in [98, 57]. We conclude with an
application to chemical process control.
4.2 Nonsmooth H
design technique
Consider a plant P in state-space form
P(s) :
_
_
x
z
y
_
_ =
_
_
A B
1
B
2
C
1
D
11
D
12
C
2
D
21
D
22
_
_
_
_
x
w
u
_
_ , (4.1)
where x R
n
is the state vector of P, u R
m
2
the vector of control inputs, w R
m
1
the vector of exogenous inputs, y R
p
2
the vector of measurements and z R
p
1
the
controlled or performance vector. Without loss, it is assumed throughout that D
22
= 0.
The focus is on H
synthesis with structured controllers, which consists in designing

a dynamic output feedback controller K(s) with feedback law u = K(s)y for the plant in
(4.1) having the following properties:
Controller structure: K(s) has a prescribed structure.
Internal stability: K(s) stabilizes the original plant P(s) in closed-loop.
Performance: Among all stabilizing controllers with that structure, K(s) mini-
mizes the H
norm |T
wz
(K)|
. Here T
wz
(K) denotes the closed-loop transfer
function from w to z.
4.2.1 Subdierential of the H
map
For the time being we leave apart structural constraints and assume that K(s) has the
frequency domain representation:
K(s) = C
K
(sI A
K
)
1
B
K
+D
K
, A
K
R
kk
, (4.2)
where k is the order of the controller, and where the case k = 0 of a static controller
K(s) = D
K
is included. A further simplication is obtained if we assume that preliminary
dynamic augmentation of the plant P(s) has been performed:
A
_
A 0
0 0
k
_
, B
1

_
B
1
0
_
, etc.
so that manipulations will involve a static matrix
/
_
A
K
B
K
C
K
D
K
_
. (4.3)
With this proviso, the following closed-loop notations will be useful:
_
/(/) B(/)
c(/) T(/)
_
_
A B
1
C
1
D
11
_
+
_
B
2
D
12
_
/
_
C
2
D
21
_
. (4.4)
Owing to its special composite structure, the function f = |.|
T
wz
, which maps
the set T R
(m
2
+k)(p
2
+k)
of stabilizing controllers into R
+
, is Clarke subdierentiable
[104, 8, 7]. Its Clarke subdierential can be described as follows. Introduce the set of
active frequencies at a given /
(/) [0, +] : (T
wz
(/, j)) = f(/). (4.5)
We assume throughout that (/) is a nite set and we refer the reader to [27] for a
justication of this hypothesis. We shall also need the notation:
_
T
wz
(/, s) G
12
(/, s)
G
21
(/, s)
_

_
c(/)
C
2
_
(sI /(/))
1
_
B(/) B
2
_
+
_
T(/) D
12
D
21

_
.
(4.6)
This leads to the following result
4.2. Nonsmooth H
design technique 87
Theorem 4.2.1 Assume the controller K(s) stabilizes P(s) in (4.1), that is, / T.
With the notations (4.5) and (4.6), let Q
be a matrix whose columns form an orthonormal

basis of the eigenspace of T
wz
(/, j)T
wz
(/, j)
H
associated with the largest eigenvalue
1
_
T
wz
(/, j)T
wz
(/, j)
H
_
= (T
wz
(/, j))
2
. Then, the Clarke subdierential of
the mapping f at / T is the compact and convex set f(/) =
Y
: Y o(/), where
Y
= f(/)
1

(|)
Re
_
G
21
(/, j) T
wz
(/, j)
H
Q
(Q
)
H
G
12
(/, j)
_
T
, (4.7)
and S(/) is the spectraplex set
o(/) Y = (Y
)
(|)
: Y
= (Y
)
H
_ 0,
(|)
Tr Y
= 1. (4.8)
Proof: The proof is based on the chain rule for the Clarke gradient of the composite
mapping f = |.|
T
wz
and we refer the reader to [44] and [116, 8, 7] for a proof and
further details.
In geometric terms, the subdierential of f is a linear image of the spectraplex set
S(/). We shall see later that it reduces to a more familiar geometric set under additional
assumptions on the multiplicity of the singular values.
4.2.2 Structured controllers
Note that we have assumed so far that controllers have no specic structure. We now
extend the results in section 4.2.1 to structured controllers using chain rules.
Assume / dened in (4.3) depends smoothly on a free parameter R
q
, that is,
/ = /(), where /() is smooth. Then the subgradients with respect to of the mapping
g = |.|
T
wz
(.) /(.) at are obtained as /
t
()
f(/), where f(/) is given in

Theorem 4.2.1, /
t
() is the derivative of /() at , and where /
t
()
is its adjoint. This

is a direct application of the chain rule in [44]. Note that the adjoint /
t
()
acts on
elements F R
(m
2
+k)(p
2
+k)
via
/
t
()
F =
_
Tr (
|()
1
T
F), . . . , Tr (
|()
q
T
F)
_
T
.
We infer the following
Corollary 4.2.2 Assume the controller /() stabilizes P(s) in (4.1), that is, /() T.
With the notations of Theorem 4.2.1, the Clarke subdierential of the mapping g = |.|
T
wz
(.) /(.) at R
q
is the compact and convex set
g() =
_
_
Tr (
|()
1
T
Y
), . . . , Tr (
|()
q
T
Y
)
_
T
:
Y
f(/())
_
. (4.9)
Using vectorization, the subgradients in (4.9) can be rewritten as
_
vec
|()
1
, . . . , vec
|()
q
_
T
vec
Y
. (4.10)
An important special case in practice is when the maximum singular values
(T
wz
(/(), j))
have multiplicity one for every (/()). Then the subgradients
Y
reduce in vector
form to vec
Y
= where

(|())
= 1,
0, (/()) and matrix is

constructed columnwise as

_
vec Re
_
G
21
(/, j) T
wz
(/, j)
H
Q
(Q
)
H
G
12
(/, j)
_
T
_
(|())
.
Combining this expression with (4.10), the subdierential g() at admits a simpler
representation in the form of a linear image of a simplex
g() =
_
_
_
_
vec
|()
1
, . . . , vec
|()
q
_
T
:

(|())
= 1,
0, (/())
_
_
_
.
4.2.3 PID controllers
If this section we specialize the above results to PID controllers. A common representation
of MIMO PID controllers is
K(s) = K
p
+
K
i
s
+
K
d
s
1 + s
, (4.11)
where K
p
, K
i
and K
d
are the proportional, the integral and the derivative gains, re-
spectively. The PID gains K
p
, K
i
and K
d
all belong to R
m
2
m
2
for a square plant with
m
2
inputs and outputs. is a small scalar which determines how close the last term in
(4.11) comes to a pure derivative action. Using partial fraction expansion, an alternative
representation can be obtained in the form
K(s) = D
K
+
R
i
s
+
R
d
s +
, (4.12)
with the correspondence
D
K
K
p
+
K
d
, R
i
K
i
, R
d

K
d
2
,
1
.
Note that these two representations are in one-to-one correspondence via
K
d
=
2
R
d
, K
p
= D
K
+R
d
, K
i
= R
i
, =
1
.
From (4.12) we obtain a linearly parameterized state-space representation of a MIMO
PID controller
/ =
_
A
K
B
K
C
K
D
K
_
=
_
_
0 0 R
i
0 I R
d
I I D
K
_
_, A
K
R
2m
2
2m
2
. (4.13)
4.2. Nonsmooth H
design technique 89
Since the state-space representation of the PID controller is ane in the parameters ,
R
i
, R
d
and D
k
, the same is true for its vectorized form and we can write
vec
_
A
K
B
K
C
K
D
K
_
= vec
_
_
0 0 0
0 0 0
I I 0
_
_ +T
..
_
vec R
i
vec R
d
vec D
K
_
_
,
for a suitable matrix T R
(k+m
2
)(k+p
2
)(3m
2
2
+1)
. The linear part of (4.13) can be expanded
as
_
_
_
0
I
0
_
_
_I
_
_
_
0
I
0
_
_
_
T
+
_
_
_
I
0
0
_
_
_R
i
_
_
_
0
0
I
_
_
_
T
+
_
_
_
0
I
0
_
_
_R
d
_
_
_
0
0
I
_
_
_
T
+
_
_
_
0
0
I
_
_
_D
K
_
_
_
0
0
I
_
_
_
T
.
In consequence, using the Kronecker product to convert matrix products into vectors, we
obtain the transformation
T =
_
_
_
_
_
0
I
0
_
_
_
_
_
_
0
I
0
_
_
_vec I
_
_
_
0
0
I
_
_
_
_
_
_
I
0
0
_
_
_
_
_
_
0
0
I
_
_
_
_
_
_
0
I
0
_
_
_
_
_
_
0
0
I
_
_
_
_
_
_
0
0
I
_
_
_
_
_ .
Finally, the subdierential of the mapping g = |.|
T
wz
(.) /(.) at , where /()
describes a MIMO PID controller (4.11) or (4.12) above, is the compact and convex set
of subgradients
g() =
_
T
T
vec
Y
:
Y
f(/())
_
. (4.14)
For a decentralized MIMO PID controller, R
i
, R
d
and D
K
reduce to diagonal matrices.
Introducing the linear transformation
J
_
e
1
e
1
. . . e
m
2
e
m
2
_
,
where (e
i
)
i=1,...,m
2
is the canonical basis of R
m
2
, it is easily veried that
vec M = J [M
11
, M
22
, . . . , M
m
2
m
2
]
T
for any square matrix M of size m
2
. This leads to
_
vec R
i
vec R
d
vec D
K
_
_
=
L
..
_
_
1 0 0 0
0 J 0 0
0 0 J 0
0 0 0 J
_
..
_
vec diag R
i
vec diag R
d
vec diag D
K
_
_
with the new parameter vector as indicated above. Again by the chain rule for Clarke
subdierentials, we obtain that the subdierential of the mapping g = |.|
T
wz
(.)/(.)
at , where /() describes a MIMO decentralized PID controller stabilizing (4.1), is the
compact and convex set of subgradients
g() =
_
L
T
T
T
vec
Y
:
Y
f(/())
_
.
We emphasize that the outlined procedure to determine subdierentials of various
types of PID controllers is general and encompasses most controller structures. In partic-
ular, this includes all structures / = /() with a dierentiable parametrization /(). In
addition, in some cases the parameter R
q
may be restricted to a contraint subset of
R
q
.
4.2.4 Setpoint lter design
Despite the improvement in performance achieved by our new technique, using PID feed-
back alone may not be sucient to meet suitable time-domain constraints. Traditionally,
this diculty is overcome by using a two-degree of freedom strategy including feedback
and preltering or setpoint ltering. Setpoint lters operate on the reference signals to
improve responses properties such as settling-time, overshoot and decouplings. In gure
4.2, a typical model following strategy is shown. The setpoint lter F(s) is used in such a
way that the responses of the feedback controlled plant G(s) match as closely as possible
those of a reference model G
ref
(s). Finding such a lter could also be cast as an H
synthesis problem, where the transfer function from the reference signal r to the tracking
error e is minimized:
minimize
F(s)
|T
re
(F)|
. (4.15)
+
e
r
F(s)
G
ref
(s)
K(s)
+
G(s)
Figure 4.2: Setpoint lter design.
To solve the setpoint lter design problem, we suggest once again the use of nonsmooth
optimization methods. In order to illustrate the construction, consider the case of a two-
input two-output system. To achieve decoupling and good quality responses, the setpoint
4.2. Nonsmooth H
design technique 91
lter is sought in the form [132],
F(s) =
_
1
1
s+1
a
1
s
b
1
s+1
a
2
s
b
2
s+1
1
2
s+1
_
. (4.16)
Setting
1
=
1
1
,
2
=
1
b
1
,
3
=
a
1
b
1
,
4
=
1
2
,
5
=
1
b
2
,
6
=
a
2
b
2
,
a state-space representation of the lter is obtained as
T()
_
A
F
B
F
C
F
D
F
_
=
_
1
0 0 0
1
0
0
2
0 0 0
3
0 0
4
0 0
4
0 0 0
5

6
0
1
2
0 0 0
3
0 0 1
5

6
0
_
_
.
This means there exists a matrix U such that
vec T() = vec T(0) + U, R
6
.
We immediately deduce the relevant subgradient formulas for program (4.15). With
v |.|
T
re
(.) T(.), the subdierential of v at , where T() is a setpoint lter, is
the compact and convex set of subgradients
v() =
_
U
T
vec
Y
:
Y
(|.|
T
re
) (T())
_
. (4.17)
The remaining expression for the subdierential is directly obtained from Theorem 4.2.1.
4.2.5 Nonsmooth descent method
In this section we briey present our nonsmooth optimization technique for composite
functions of the H
-norm. For a more detailed discussion we refer the reader to [8, 9].
We start by representing the composite functions f = | |
T
wz
or more generally
g = | |
T
wz
/() under the form
g() = max
[0,+]
g(, ),
where each g(, ) is a composite maximum singular value function
g(, ) = (((, j)) .
Here ((, j) = T
wz
(/(), j). At a given parameter , we can compute the set
() (/()) of active frequencies, which is either nite, or coincides with [0, +] in
those rare cases where the closed-loop system is all-pass. Excluding this case, we assume
() nite and construct a nite extension
e
() by adding frequencies according to the
strategy presented in [8, 9]. See gure 4.3 for a typical choice.
Following the general trend of Polak [114], we now dene the optimality function
e
() min
hR
q
max
e
()
max
Y
0
Tr(Y
)=1
g() + g(, ) + h
T
+
1
2
h
T
Qh, (4.18)
where for every xed ,
Y
is a subgradient of g(, ) at obtained as

_
Tr (
|()
1
T
), . . . , Tr (
|()
q
T
)
_
T
,
where
= g(, )
1
Re G
21
(/, j) T
wz
(/, j)
H
Q
(Q
)
H
G
12
(/, j) , Y
_ 0, Tr Y
= 1.
The model of the objective function represented by
e
is in principle of rst order, but
the quadratic term h
T
Qh may in some cases be used to include second order information.
In [8, 9] we had worked with the basic choice Q = I ~ 0, but we shall propose a more
sophisticated choice here using BFGS updates.
Notice that independently of the choices of Q ~ 0 and the nite extension
e
() of
() used, the optimality function has the following property:
e
() 0, and
e
() = 0
if and only if 0 g(), that is, is a critical point of g. In order to use
e
to compute
descent steps, it is convenient to obtain a dual representation of
e
. To do this we use
Fenchel duality to swap the max and min operators in (4.18). This means that we rst
replace the rst inner supremum by a supremum over a convex hull which does not alter
the value of
e
. Then, after swapping max and min, the now inner inmum over h R
q
becomes unconstrained and can be computed explicitely. Namely, for xed Y
and
in
the outer program, we obtain the solution of the form
h(Y, ) = Q
1
_
_

e
()
_
_
. (4.19)
Substituting this back we obtain the dual expression
e
() = max
e
()
=1
max
Y
0
Tr(Y
)=1
_
_

e
()
(g(, ) g())
1
2
_

e
()
_
T
Q
1
_

e
()
_
_
_
. (4.20)
Notice that in its dual form, computing
e
() leads to a semidenite program. Indeed,
substituting Z
, program (4.20) becomes
e
() = max
Z
e
()
Tr(Z
)=1
_
_

e
()
Tr(Z
) (g(, ) g())
1
2
_

e
()
_
T
Q
1
_

e
()
_
_
_
. (4.21)
4.2. Nonsmooth H
design technique 93
The latter program is converted to an LMI problem using a Schur complement argument.
As a byproduct we see that
e
() 0 and that
e
() = 0 implies is critical that is,
0 g().
What is important is that the direction h(Y, ) = h(Z) in (4.19) is a descent direction
of g at in the sense that the directional derivative satises the decrease condition
g
t
(; h(Z))
e
()
1
2
_

e
()
_
Q
1
_

e
()
_
,
where Z is the dual optimal solution. See [9, Lemma 4.3] for a proof. In conclusion, we
obtain the following algorithmic scheme:
Nonsmooth descent method for min
g()
Parameters 0 < < 1, 0 < < 1, 0 < 1.
1. Initialization. Find a structured closed-loop stabilizing controller /().
2. Active frequencies. Compute g() using the algorithm of [28] in its rened
version [27] and obtain set of active frequencies ().
3. Add frequencies. Build nite extension
e
() of () as proposed in [8, 9],
and choose Q _ I.
4. Step computation. Compute
e
() by the dual SDP (4.20) and thereby
obtain direction h(Z) in (4.19). If
e
() = 0 stop. Otherwise:
5. Line search. Find largest t =
k
such that g( +th(Z)) < g() t
e
()
and such that /( +th(Z)) remains stabilizing.
6. Step. Replace by +th(Z) and go back to step 2.
Notice that the line search in step 5 is successful because t
1
(g( +th(Z)) g())
g
t
(; h(Z)) as t 0
+
, and because
e
() < 0 and 0 < < 1. Choosing t under the form
t =
k
with the largest possible k comes down to doing a backtracking line search, which
safeguards against taking too small steps.
The nal elements to be provided is computation of the matrix Q
1
. Since Q is sup-
posed to carry second order information on the objective function, it may seem appropriate
to do a BFGS update
Q
+
= Q+
yy
T
y
T
s

Qss
T
Q
s
T
Qs
,
where s =
+
and y =
+
, where is the subgradient of minimal norm in g(),
g
+
the subgradient of minimal norm in g(
+
). Here the notation x
+
and x stands for
current and past iterates, respectively. Since the inverse Q
1
is required, an alternative
is to use the inverse BFGS update. Here we maintain the matrix P Q
1
in step 4 of
the algorithm through
P
+
= P +
(s Py)s
T
+s(s Py)
T
s
T
y

y
T
(s Py)ss
T
(s
T
y)
2
.
As is well-known, P ~ 0 is guaranteed as long as s
T
y > 0. If this fails, or if P ,~ 0 for
numerical reasons, we modify P
+
or restart the procedure.
Note that computing minimal norm elements g() amounts to computing mini-
mal norm elements in the LMI set (4.9) and therefore reduces to an LMI problem. Finally,
we emphasize the important fact that when singular values (((, j)) are simple on
e
(), which is the rule in practice, we have Z
= Tr Z
so that SDP (4.21) simplies to

a much faster convex QP. A fact that can also be exploited for computing minimal norm
elements.
cut
active frequencies
extended set

(T
wz
(K)) (T
wz
(K))
Figure 4.3: Selection of frequencies: 1)lhs active only, 2)rhs active and secondary peaks.
4.3 Nonsmooth loop-shaping design
In this section we recall some key facts from the loop-shaping design technique introduced
by McFarlane and Glover [98], and we discuss how it can be merged with our nonsmooth
framework to arrive at a new practical PID loop-shaping design method.
4.3.1 Loop-shaping design
Loop-shaping design is an ecient and practical technique which has been applied success-
fully to a variety of dicult design problems, see [42, 53, 120] to cite a few. It proceeds
as presented in gure 4.4. Firstly, the open-loop plant G is altered by pre- and post-
compensators W
1
and W
2
, respectively, to achieve desired open-loop shapes. Roughly
speaking, the shaped plant W
2
GW
1
should have large gains at low frequencies for per-
formance and small gains at high frequencies for noise attenuation. Also, the roll-o
rate should not be too large in the mid-frequency range. Secondly, an H
synthesis is
4.3. Nonsmooth loop-shaping design 95
performed to minimize the objective
|T
(w
1
,w
2
)(z
1
,z
2
)
(K)|
|
_
K
I
_
(I W
2
GW
1
)
1
_
G I
_
|
(4.22)
which yields a controller K. The nal controller is then implemented in the form W
1
KW
2
and has no specic structure. Notice that the H
norm in (4.22) is greater than one for

the optimal controller K. Put dierently, 1/ is bounded above by unity. The scalar
is therefore an indicator of success of the procedure. Usually values greater than 0.2 or
0.3 are deemed acceptable in the sense that the controller K does not signicantly alter
the desired loop shape W
2
GW
1
[98]. Moreover, in this situation, the closed-loop system
will be robust against coprime factor uncertainties [98]. A smaller , on the contrary,
tells us that the desired loop shape is incompatible with robustness requirements, and
the pre- and post-compensators must be modied to reect less ambitious performance
requirements. A central element of the design technique is therefore to select appropriate
compensators for the trade-o between performance and robustness. In practice this is
generally accomplished by trials-and-errors.
z
2
z
1
w
2 w
1
W
2
W
1
G
K
Figure 4.4: Loop-shaping H
design.
4.3.2 PID Loop-shaping design
For loop-shaping design with a PID controller we have adopted the strategy introduced
in [57]. In this approach the controller K is sought in the form K = W
1
1
K
PID
, assuming
that the pre-compensator W
1
is invertible and K
PID
is a PID controller. The synthesis
problem is then of the form
minimize
K
PID
stabilizing
|
_
W
1
1
K
PID
I
_
(I W
2
GK
PID
)
1
_
W
2
GW
1
I
_
|
. (4.23)
Note that this scheme retains all the benets of the H
loop-shaping design technique of

section 4.3.1. As before 1/, where is the optimal H
cost in (4.23), will serve to

diagnose whether performance and robustness requirements have been achieved. The nal
controller is obtained as the series connection of the post-compensator W
2
and the PID
controller in the form K
PID
W
2
. An immediate consequence is that the nal controller has
better noise attenuation in the high frequency range than a pure PID controller, whenever
W
2
is strictly proper. To sum up, we have to solve program (4.23), see also gure 4.5, for
a PID controller structure and the nonsmooth technique and tools discussed in sections
4.2 will be used to that purpose.
z
1
W
1
W
1
1
K
pid
w
1
z
2
w
2
G
W
2
Figure 4.5: PID loop-shaping H
design.
4.4 Application to a separating tower
In this section, we consider the application of the nonsmooth technique to the control
design for a chemical process from the literature [57]. It consists of a 24-tray tower for
separating methanol and water. The transfer function for controlling the temperature on
the 4th and 17th trays is given as
_
t
17
t
4
_
=
_
2.2e
s
7s+1
1.3e
0.3s
7s+1
2.8e
1.8s
9.5s+1
4.3e
0.35s
9.2s+1
_ _
u
1
u
2
_
. (4.24)
Settling times of about 10 seconds are desired for the closed-loop process in response to
step inputs, as well as good decoupling between the temperatures t
17
and t
4
. A good
robustness margin is also required to account for process model uncertainties. The latter
will be assessed using the coprime factor uncertainty margin as dened in section 4.3.
The actual plant in (4.24) is approximated by a rational model using 2nd-order Pade ap-
proximation of the delays. This leads to a 12th-order model. Pre- and post-compensators
are taken from [57]:
W
1
(s) =
_
5s+2
s+0.001
0
0
5s+2
s+0.001
_
, W
2
(s) =
_
10
s+10
0
0
10
s+10
_
.
4.4. Application to a separating tower 97
The standard form of gure 4.5 incorporating the compensators is therefore of 18th-
order. For comparison, we have synthesized an unstructured full-order controller, whose
associated step responses and singular value frequency responses are displayed in gures
4.6 and 4.7, respectively. For full order the robustness was found to be
1
= 0.361.
This is very satisfactory in terms of stability and gives good agreement with the prescribed
open-loop shapes. This is conrmed by the fast and well-damped step responses in gure
4.6. We also observe short term couplings between t
17
and t
4
, an unpleasant behavior
which we try to reduce later when dealing with PID controllers. For future use, we
keep in mind that the the optimal = 2.77 achieved at this stage is globally optimal if
controller structural constraints are disregarded.
In reference [57] a state-space BMI formulation is used to characterize PID solutions
of the H
synthesis problem in (4.23). The proposed algorithmic strategy is a coordinate

descent scheme, which alternates between minimizing over Lyapunov variables and over
PID controller parameters. Such techniques are often referred to as D-K or V -K itera-
tions, see [60] and references therein. Unfortunately, coordinate descent does not come
with any convergence guarantee [66] and breakdown is often observed in practice. In our
case, reference [57] reported 38 minutes of cputime to obtain the following solution with
such a technique:
= 0.060, K
p
=
_
2.4719 1.2098
1.1667 2.4766
_
,
K
i
=
_
0.4657 0.31
0.2329 0.487
_
, K
d
=
_
0.0534 0.0072
0.015 0.0434
_
.
The corresponding robustness margin is = 1/4.02 = 0.249. Time domain and frequency
responses are shown in gures 4.8 and 4.9, respectively. Using the tools from section 4.2.5,
we now show that this PID controller is not a local minimum of program (4.23). Our
local optimality certicate evaluated at the PID controller above takes a negative value
= 0.119, revealing failure of the D-K iteration scheme to reach local optimality. This
also indicates that further progress can be achieved by running our nonsmooth method
initialized at the point of failure. Ultimately the following PID controller is obtained
= 0.1527, K
p
=
_
2.6047 0.6543
1.1253 2.3226
_
,
K
i
=
_
0.8527 0.2591
0.0701 0.9362
_
, K
d
=
_
0.7414 0.2551
1.5610 0.0331
_
.
This represents 38% improvement of the robustness margin over the value = 1/2.91 =
0.343. Better time responses are also observed, see gure 4.10. The optimal PID controller
also exhibits higher gains in the low and medium frequency ranges, gure 4.11. On the
other hand, it can be seen in gure 4.10 that step responses of t
17
and t
4
are strongly
coupled, which is undesirable in this application. To reduce the coupling we added a
setpoint lter using the model reference approach discussed in section 4.2.4, with the
lter structure in (4.16). The reference model was selected as
G
ref
(s)
_
1
s+1
0
0
1
s+1
_
Solving program (4.15) using our nonsmooth technique produced the following setpoint
lter:
F(s) =
_
1.045
s+1.045
0.3428s
s+2.25
0.3666s
s+0.6147
3.675
s+3.675
_
.
Step responses with PID controller in tandem with the setpoint lter are presented in
gure 4.12.
In a second experiment, we do not use the result in [57] and initialize the algorithm
from scratch, by rst computing a stabilizing PID controller, using an extension of our
method to minimize the spectral abscissa [25]. Note that with a dierent initial seed,
there is no reason why we would reach the same local solution. And indeed, a new local
minimum is reached:
= 0.3780, K
p
=
_
2.4798 0.5153
0.8850 2.1737
_
,
K
i
=
_
0.7428 0.1873
0.0455 0.7698
_
, K
d
=
_
0.2640 0.2347
2.1424 0.0723
_
.
The robustness margin is now = 1/3.05 = 0.3279, which is only marginally worse than
the previously synthesized PID controller and again improves the solution given in [57]
by 24%. Time- and frequency-domain evaluations are given in gures 4.13 and 4.14. As
before, a setpoint lter was computed using the reference model
G
ref
(s)
_
0.25
s
2
+0.4s+0.25
0
0
1
s+1
_
.
The setpoint lter is described as
F(s) =
_
0.7333
s+0.7333
35.81s
s+297.7
0.5452s
s+0.31
215.1
s+215.1
_
.
and step responses are given in gure 4.15.
As one might expect for a local optimization technique, our experiment underlines that
dierent local solutions are generally obtained when dierent initial guesses are used. Such
guesses may be obtained by heuristic techniques developed in PID synthesis. If many local
optima are observed, using semi-global techniques may be indicated. Such methods try
to improve on a given set of local solutions [130].
A strong point of our method is that it practically always nds local optimal solutions,
as theoretically expected. Also, the running times are very fast. In fact, in our test we
never exceed a minute cputime. This is in strong contrast with state-space sequential LMI
or BMI approaches, which require substantial numerical eorts. In those cases where a
local minimum is not satisfactory, we have to do restarts with dierent initial seeds, to
explore whether further improvements is possible. How this can be organized in general
will be investigated in the future.
4.4. Application to a separating tower 99
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)

o
u
t
p
u
t

#
1
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)
o
u
t
p
u
t

#
2
Figure 4.6: Time-domain simulations of
H
controller.
10
5
10
4
10
3
10
2
10
1
10
0
10
1
10
2
10
3
80
60
40
20
0
20
40
60
Singular Values
Frequency (rad/sec)
S
in
g
u
la
r

V
a
lu
e
s

(
d
B
)
Figure 4.7: Frequency response of H
controller.
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
time (sec.)

o
u
t
p
u
t

#
1
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)
o
u
t
p
u
t

#
2
PID controller in [57].
10
2
10
1
10
0
10
1
10
2
10
3
30
20
10
0
10
20
30
40
Singular Values
Frequency (rad/sec)
S
i
n
g
u
l
a
r

V
a
l
u
e
s

(
d
B
)
Figure 4.9: Frequency response of PID
controller in [57].
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)

o
u
t
p
u
t

#
1
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)
o
u
t
p
u
t

#
2
1st PID controller.
10
3
10
2
10
1
10
0
10
1
10
2
10
3
30
20
10
0
10
20
30
40
50
60
70
Singular Values
Frequency (rad/sec)
S
i
n
g
u
l
a
r

V
a
l
u
e
s

(
d
B
)
Figure 4.11: Frequency response of 1st
PID controller.
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)

o
u
t
p
u
t

#
1
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)
o
u
t
p
u
t

#
2
Figure 4.12: Time-domain simulations of 1st PID controller with setpoint lter.
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)

o
u
t
p
u
t

#
1
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)
o
u
t
p
u
t

#
2
2nd PID controller.
10
3
10
2
10
1
10
0
10
1
10
2
10
3
40
20
0
20
40
60
80
Singular Values
Frequency (rad/sec)
S
in
g
u
la
r

V
a
lu
e
s

(
d
B
)
Figure 4.14: Frequency response of 2nd
PID controller.
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)

o
u
t
p
u
t

#
1
0 10 20 30
0.2
0
0.2
0.4
0.6
0.8
1
1.2
time (sec.)
o
u
t
p
u
t

#
2
Figure 4.15: Time-domain simulations of 2nd PID controller with setpoint lter.
4.5. Conclusion 101
4.5 Conclusion
We have presented and discussed a nonsmooth optimization technique for the synthesis of
nely structured controllers with an H
objective. Our approach is general and encom-

passes most controller structures of practical interest. The proposed technique is endowed
by an easily tractable convergence certicate, which may either serve to validate a given
controller, or to drive the iterative descent method to termination. A specialization to
loop-shaping design with MIMO PID controllers is also introduced. Application to a
process system indicates that the technique is a practical and eective numerical tool in
structured controller design.
Ackowledgements
The authors acknowledge nancial support from the Fondation dentreprise EADS (under
contract Solving chal lenging problems...) and from the Agence Nationale de Recherche
(ANR) (project NT05 Guidage).
Chapitre 5
Commande temporelle et
frquentielle par techniques non
lisses
Nous illustrons dans larticle qui suit la exibilit et la richesse de lapproche non lisse
pour rsoudre dirents problmes pratiques de commande, travers trois exemples de
synthse structure.
Le premier est une application immdiate de la synthse H
qui permet lobtention

directe dune loi de commande optimale sous forme observateur (forme dite observer based
ou LQG). Un tel correcteur, bien que dordre plein, est particulirement lisible, puisque
ses variables dtat sont les tats estims du systme command.
Le deuxime exemple montre comment structurer la forme standard an de traiter
par notre approche des problmes de commande tolrante aux pannes (reliable ou fault-
tolerant control en anglais). La loi de commande est alors synthtise de faon garantir
une performance en boucle ferme pour le systme nominal ainsi que pour un ensemble ni
de systmes dgrads (cas de pannes de capteurs, par exemple). Le correcteur lui-mme
peut tre structur ; une action PID a t choisie ici.
Enn, nous montrons que lalgorithme de descente non lisse appliqu jusquici un
critre frquentiel (synthse H
) est transposable dans le domaine temporel. Les spci-

cations temporelles usuelles (dpassement, temps de monte, temps de rponse,. . .) font
en eet intervenir des maxima sur des intervalles de temps, et leur satisfaction sexprime
donc comme un problme minimax semi-inni, tout comme la synthse H
. Seul le calcul
des lments direntiels dire, et fait ici appel des simulations temporelles.
La mthode de descente non lisse pour la commande structure est ici rappele dans
un contexte commun ces trois applications. Cette unit dapproche nest pas seulement
satisfaisante dun point de vue thorique : nous pensons quelle facilitera, dans les travaux
futurs, la combinaison de critres et de contraintes de nature frquentielle ou temporelle
dans un mme problme doptimisation non lisse.
103
104 Chapitre 5. Commande temporelle et frquentielle par techniques non lisses
5.1. Introduction 105
Control design in the time- and frequency
domains using nonsmooth techniques
V. Bompart, P. Apkarian and D. Noll
Article paratre dans Systems & Control Letters (sous presse).
Abstract
Signicant progress in control design has been achieved by the use of nonsmooth and semi-
innite mathematical programming techniques. In contrast with LMI or BMI approaches,
these new methods avoid the use of Lyapunov variables, which gives them two major
strategic advances over matrix inequality methods. Due to the much smaller number of
decision variables, they do not suer from size restrictions, and they are much easier to
adapt to structural constraints on the controller. In this paper, we further develop this
line and address both frequency- and time-domain design specications by means of a
nonsmooth algorithm general enough to handle both cases. Numerical experiments are
presented for reliable or fault-tolerant control, and for time response shaping.
Keywords: Nonsmooth optimization, H
synthesis, structured controllers, PID tuning,

time-domain constraints, fault tolerance control, time response shaping.
Notation
Let R
nm
be the space of nm matrices, equipped with the corresponding scalar product
X, Y ) = Tr(X
T
Y ), where X
T
is the transpose of the matrix X, Tr (X) its trace. The
notation H
r
stands for the set of Hermitian matrices of size r. For complex matrices X
H
stands for its conjugate transpose. For Hermitian or symmetric matrices, X ~ Y means
that X Y is positive denite, X _ Y that X Y is positive semi-denite. We use
the symbol
1
to denote the maximum eigenvalue of a symmetric or Hermitian matrix.
Given an operator T, T
is used to denote its adjoint operator on the appropriate space.

The notation vec applied to a matrix stands for the usual column-wise vectorization of a
matrix. We use concepts from nonsmooth analysis covered by [44]. For a locally Lipschitz
function f : R
n
R, f(x) denotes its Clarke subdierential at x while the notation
f
t
(x; h) stand for its directional derivative at x in the direction h.
5.1 Introduction
Interesting new methods in nonsmooth optimization for the synthesis of controllers have
recently been proposed. See [33, 27] for stabilization problems, [8, 9, 94, 7, 32] for H

synthesis, and [7, 11] for design with IQCs. These techniques are in our opinion a valuable
addition to the designers toolkit:
They avoid expensive state-space characterizations, which suer the curse of di-
mension, because the number of Lyapunov variables grows quadratically with the
system size.
The preponderant computational load of these new methods is transferred to the
frequency domain and consist mainly in the computation of spectra and eigenspaces,
and of frequency domain quantities, for which ecient algorithms exist. This key
feature is the result of the idea of the diligent use of nonsmooth criteria of the form
f(K) = max
[0,]
1
(F(K, )), which are composite functions of a smooth but
nonlinear operator F, and a nonsmooth but convex function
1
.
The new approach is highly exible, as it allows to address, with almost no additional
cost, structured synthesis problems of the form f() = max
[0,]
1
(F(/(), )),
where /() denes a mapping from the space of controller parameters to the space
of state-space representations K. From a practical viewpoint, structured controllers
are better apprehended by designers and facilitate implementation and re-tuning
whenever performance or stability specications change. This may be the major
advantage of the new approach over matrix inequality methods.
The new approach is general and encompasses a wide range of problems beyond
pure stabilization and H
synthesis. A number of important problems in control

theory can be regarded as structured control problems. Striking examples are si-
multaneous stabilization, reliable and decentralized control, multi frequency band
design, multidisk synthesis, and much else.
Finally, the new methods are supported by mathematical convergence theory, which
certies global convergence under practically useful hypotheses in the sense that
iterates converge to critical points from arbitrary starting points.
In this paper, we expand on the nonsmooth technique previously introduced in [8], and
explore its applicability to structured controller design in the presence of frequency- and
time-domain specications. We show that the same nonsmooth minimization technique
can be used to handle these seemingly dierent specications. We address implementation
details of the proposed technique and highlight dierences between frequency and time
domain. Finally, we present numerical results in observer-based control design and reliable
control, as well as for time-response shaping.
We refer the reader to the articles cited above for references on controller synthesis
using nonsmooth optimization. General concepts in nonsmooth analysis can be found in
[44], and optimization of max functions is covered by [114]. Time response shaping is
addressed at length in [59, 71, 73]. These techniques are often referred to as the Iterative
Feedback Tuning (IFT) approach, mainly developed by M. Gevers, H. Hjalmarsson and
co-workers.
5.2. Time- and frequency domain designs 107
5.2 Time- and frequency domain designs
P
K
y u
w
z
T
wz
(K) :=
Figure 5.1: Standard interconnection.
Consider a plant P in state-space form
P(s) :
_
_
x
z
y
_
_ =
_
_
A B
1
B
2
C
1
D
11
D
12
C
2
D
21
D
22
_
_
_
_
x
w
u
_
_ , (5.1)
where x R
n
is the state vector of P, u R
m
2
the vector of control inputs, w R
m
1
the vector of exogenous inputs or a test signal, y R
p
2
the vector of measurements and
z R
p
1
the controlled or performance vector. Without loss, it is assumed throughout
that D
22
= 0.
The focus is on time- or frequency domain synthesis with structured controllers, which
consists in designing a dynamic output feedback controller K(s) with feedback law u =
K(s)y for the plant in (5.1), having the following properties:
Controller structure: K(s) has a prescribed structure.
Internal stability: K(s) stabilizes the original plant P(s) in closed-loop.
Performance: Among all stabilizing controllers with that structure, K(s) is such
that either the closed-loop time response z(t) to a test signal w(t) satises prescribed
constraints, or the H
norm of transfer function |T

wz
(K)|
is minimized. Here
T
wz
(K) denotes the closed-loop transfer function from w to z, see gure 5.1.
For the time being we leave apart structural constraints and assume that K(s) has
the frequency domain representation:
K(s) = C
K
(sI A
K
)
1
B
K
+D
K
, A
K
R
kk
, (5.2)
where k is the order of the controller, and where the case k = 0 of a static controller
K(s) = D
K
is included. A further simplication is obtained if we assume that preliminary
dynamic augmentation of the plant P(s) has been performed:
A
_
A 0
0 0
k
_
, B
1

_
B
1
0
_
, etc.
so that manipulations will involve a static matrix
/
_
A
K
B
K
C
K
D
K
_
R
(k+m
2
)(k+p
2
)
. (5.3)
With this proviso, the following closed-loop notations will be useful:
_
/(/) B(/)
c(/) T(/)
_
_
A B
1
C
1
D
11
_
+
_
B
2
D
12
_
/
_
C
2
D
21
_
. (5.4)
Structural constraints on the controller will be dened by a matrix-valued mapping
/(.) from R
q
to R
(k+m
2
)(k+p
2
)
, that is / = /(), where vector R
q
denotes the
independent variables in the controller parameter space R
q
. For the time being we will
consider free variation R
q
, but the reader will be easily convinced that adding param-
eter restriction by means of mathematical programming constraints g
I
() 0, g
E
() = 0
could be added if need be. We will assume throughout that the mapping /(.) is continu-
ously dierentiable, but otherwise arbitrary. As a typical example, consider MIMO PID
controllers, given as
K(s) = K
p
+
K
i
s
+
K
d
s
1 + s
, (5.5)
where K
p
, K
i
and K
d
are the proportional, the integral and the derivative gains, respec-
tively, are alternatively represented in the form
K(s) = D
K
+
R
i
s
+
R
d
s +
, (5.6)
with the relations
D
K
K
p
+
K
d
, R
i
K
i
, R
d

K
d
2
,
1
, (5.7)
and a linearly parameterized state-space representation is readily derived as
/() =
_
A
K
B
K
C
K
D
K
_
=
_
_
0 0 R
i
0 I R
d
I I D
K
_
_, A
K
R
2m
2
2m
2
. (5.8)
Free parameters in this representation can be gathered in the vector obtained as

_
vec R
i
vec R
d
vec D
K
_
_
R
3m
2
2
+1
.
We stress that the above construction is general and encompasses most controller struc-
tures of practical interest. We shall see later that interesting control problems such as
reliable control are also special cases of the general structured design problem.
With the introduced notation, time-domain design is the optimization program
minimize
R
q
_
f
() max
t[0,T]
f(, t)
_
5.3. Nonsmooth descent method 109
where the case T = is allowed. See section 5.3.1.0 for further details and other practical
options.
Frequency-domain design is the standard H
problem
minimize
R
q
_
f
() sup
[0,]
(T
wz
(/(), j)) = [[T
wz
(/())[[
_
.
5.3 Nonsmooth descent method
In this section we briey present our nonsmooth optimization technique for time- and
frequency-domain max functions. For a detailed discussion of the H
norm, we refer the

reader to [8, 9]. The setting under investigation is
minimize
max
xX
f(, x) , (5.9)
where the semi-innite variable x = t or x = is restricted to a one-dimensional set X.
Here X may be the haline [0, ], or a limited band [
1
,
2
], or a union of such bands in
the frequency domain, and similarly in the time domain. The symbol denotes the design
variable involved in the controller parametrization /(), and we introduce the objective
or cost function
f
() max
xX
f(, x) .
At a given parameter , we assume that we can compute the set () of active times
or frequencies, which we assume nite for the time being:
() x X : f(, x) = f
() . (5.10)
For future use we construct a nite extension
e
() of () by adding times or frequencies
to the nite active set (). An ecient strategy to construct this set for x = has been
discussed in [8, 9]. For x = t some ideas will be given in section 5.4.3.
For the ease of presentation we assume that the cost function f is dierentiable with
respect to for xed x
e
(), so that gradients
x
=
f(, x) are available. Extensions

to the general case are easily obtained by passing to subgradients, since f(., x) has a Clarke
gradient with respect to for every x X [44]. Following the line in Polak [114], see also
[8], we introduce the optimality function
e
() min
hR
q
max
x
e
()
f
() + f(, x) + h
T
x
+
1
2
h
T
Qh, (5.11)
Notice that
e
is a rst-order model of the objective function f
() in (5.9) in a neigh-
borhood of the current iterate . The model oers the possibility to include second-order
information [5] via the term h
T
Qh, but Q ~ 0 has to be assured. For simplicity, we will
assume Q = I with > 0 in our tests.
Notice that independently of the choices of Q ~ 0 and the nite extension
e
() of
() used, the optimality function has the following property:
e
() 0, and
e
() = 0 if
and only if 0 f
(), that is, is a critical point of f
. In order to use
e
to compute
descent steps, it is convenient to obtain a dual representation of
e
. To this aim, we rst
replace the inner maximum over
e
() in (5.11) by a maximum over its convex hull and
we use Fenchel duality to swap the max and min operators. This leads to
e
() max
x
0
x
e
()
x
=1
min
hR
q
x
e
()
x
(f(, x) f
() + h
T
x
) +
1
2
h
T
Qh.
These operations do not alter the value of
e
. The now inner inmum over h R
q
is now
unconstrained and can be computed explicitly. Namely, for xed
x
in the outer program,
we obtain the solution of the form
h() = Q
1
_
_

x
e
()
x
_
_
. (5.12)
Substituting this back in the primal program (5.11) we obtain the dual expression
e
() = max
x
0
x
e
()
x
=1
_
_

x
e
()
x
(f(, x) f
())
1
2
_

x
e
()
x
_
T
Q
1
_

x
e
()
x
_
_
_
. (5.13)
Notice that in its dual form, computing
e
() is a convex quadratic program (QP). As
a byproduct we see that
e
() 0 and that
e
() = 0 implies is critical that is,
0 f
().
What is important is that as long as
e
() < 0, the direction h() in (5.12) is a
descent direction of f
at in the sense that the directional derivative satises the

decrease condition
f
t
(; h())
e
()
1
2
_

x
e
()
x
_
T
Q
1
_

x
e
()
x
_

e
() < 0,
where is the dual optimal solution of program (5.13). See [9, Lemma 4.3] for a proof.
In conclusion, we obtain the following algorithmic scheme:
Nonsmooth descent method for min
()
Parameters 0 < < 1, 0 < < 1.
1. Initialize. Find a structured closed-loop stabilizing controller /().
2. Active times or frequencies. Compute f
() and obtain the set of active

times or frequencies ().
3. Add times or frequencies. Build nite extension
e
() of ().
4. Compute step. Calculate
e
() by the dual QP (5.13) and thereby
obtain direction h() in (5.12). If
e
() = 0 stop. Otherwise:
5. Line search. Find largest b =
k
such that f
( +bh()) < f
() b
e
()
and such that /( +b h()) remains closed-loop stabilizing.
6. Step. Replace by +b h() and go back to step 2.
Finally, we mention that the above algorithm is guaranteed to converge to a critical
point [8, 9], a local minimum in practice.
5.3.1 Nonsmooth properties
In order to make our conceptual algorithm more concrete, we need to clarify how subdif-
ferential information can be obtained for both time- and frequency-domain design.
Frequency-domain design
In the frequency domain we have x = . The function f
() becomes f
() = [[.[[

T
wz
(.) /(), which maps R
q
into R
+
, and is Clarke subdierentiable as a composite
function [104, 8, 7]. Its Clarke gradient is obtained as /
t
()
(/), where /
t
() is
the derivative of /(.) at , /
t
()
its adjoint, and where g
is dened as g
[[.[[

T
wz
(.) and maps the set T R
(m
2
+k)(p
2
+k)
of closed-loop stabilizing controllers into
R
+
. Introducing the notation
_
T
wz
(/, s) G
12
(/, s)
G
21
(/, s)
_

_
c(/)
C
2
_
(sI /(/))
1
_
B(/) B
2
_
+
_
T(/) D
12
D
21

_
.
(5.14)
the Clarke subdierential of g
at / is the compact and convex set of subgradients

g
(/)
Y
: Y o(/) where
Y
= g
(/)
1

(|)
Re
_
G
21
(/, j) T
wz
(/, j)
H
Q
(Q
)
H
G
12
(/, j)
_
T
, (5.15)
and where S(/) is the spectraplex
o(/) = Y = (Y
)
(|)
: Y
= (Y
)
H
_ 0,
(|)
Tr (Y
) = 1, Y
H
r
.
In the above expressions, Q
is a matrix whose columns span the eigenspace of

T
wz
(/, j)T
wz
(/, j)
H
associated with its largest eigenvalue
1
_
T
wz
(/, j)T
wz
(/, j)
H
_
of multiplicity r
. We also deduce from expression (5.15) the form of the subgradients of

f(, ) (T
wz
(/(), j))
at with xed , which are used in the primal and dual programs (5.11) and (5.13),
respectively
x
=
Y
= /
t
()
f(, )
1
Re
_
G
21
(/, j) T
wz
(/, j)
H
Q
(Q
)
H
G
12
(/, j)
_
T
where Q
is as before and Y
H
r
, Y
_ 0, Tr (Y
) = 1. Finally, we note that all

subgradient formulas are made implementable by expliciting the action of the adjoint
operator /
t
()
on elements F R
(m
2
+k)(p
2
+k)
. Namely, we have
/
t
()
F =
_
Tr (
|()
1
T
F), . . . , Tr (
|()
q
T
F)
_
T
.
In the general case, where some of the maximum eigenvalues at some of the frequencies
in the extended set
e
() has multiplicity > 1, the formulas above should be used, and
the dual program in (5.13) becomes a linear SDP [8, 9]. This is more expensive than a
QP, but the size of the SDP remains small, so that the method is functional even for large
systems. When maximum eigenvalues are simple, which seems to be the rule in practice,
matrices Y
are scalars, and the primal and dual subproblems become much faster convex
QPs. This feature, taken together with the fact that Lyapunov variables are never used,
explains the eciency of the proposed technique.
Time-domain design
We now specialize the objective function f
to time-domain specications. For simplicity

of the exposition, we assume the performance channel w z is SISO, that is m
1
= p
1
= 1,
while the controller channel y u remains unrestricted.
As noted in [29], most specications are in fact envelope constraints:
z
min
(t) z(, t) z
max
(t) for all t 0 (5.16)
where z(, .) is the closed-loop time response to the input signal w (typically a unit step
command), when controller / = /() is used, and where z
min
(t) z
max
(t) +
for all t 0. This formulation oers sucient exibility to cover basic step response
specications such as rise and settling times, overshoot and undershoot, or steady-state
tracking. Several constraints of this type can be combined using piecewise constant en-
velope functions z
min
and z
max
. A model following specication is easily incorporated by
setting z
min
= z
max
= z
ref
, where z
ref
is the desired closed-loop response.
For a stabilizing controller / = /(), the maximum constraint violation
f
() = max
t0
max [z(, t) z
max
(t)]
+
, [z
min
(t) z(, t)]
+
(5.17)
is well dened, f
() 0, and f
() = 0 if and only if z(, .) satises the constraint

(5.16). Minimizing f
is therefore equivalent to reducing constraint violation, and will as

a rule lead to a controller /( ) achieving the stated time-domain specications. In the
case of failure, this approach converges at least to a local minimum of constraint violation.
The objective function f
is a composite function with a double max operator. The

outer max on t 0 makes the program in (5.17) semi-innite, while the inner max, for all
t 0, is taken over f
1
(, t) = z(, t) z
max
(t), f
2
(, t) = z
min
(t) z(, t) and f
3
(, t) = 0.
Assuming that the time response z(., t) is continuously dierentiable, f
is Clarke
regular and its subdierential is
f
() = co
t()
_
co
i7(,t)
f
i
(, t)
_
, (5.18)
where () is the set of active times dened by (5.10), and
1(, t) = i 1, 2, 3 : f(, t) = f
i
(, t) .
More precisely, for all t (),
co
i7(,t)
f
i
(, t) =
_
z(, t) if z(, t) > z

max
(t)
z(, t) if z(, t) > z

min
(t)
0 if z
min
(t) < z(, t) < z
max
(t)
[
z(, t), 0] if z(, t) = z

max
(t) > z
min
(t)
[
z(, t), 0] if z(, t) = z

min
(t) < z
max
(t)
[
z(, t),
z(, t)] if z(, t) = z

min
(t) = z
max
(t)
(5.19)
Clearly, as soon as the envelope constraint is satised for one active time t (), either
one of the last four alternatives in (5.19) is met, we have f
() = 0 for all t 0 so that

0 f
() and is a global minimum of program (5.9). The computation of the descent

step only makes sense in the rst two cases, i.e., when f
() > 0. Notice then that the

active times set () can be partitioned into
1
() t : t (), f
1
(, t) = f
()
2
() t : t (), f
2
(, t) = f
()
(5.20)
and the Clarke subdierential g
(/) is completely described by the subgradients
Y
(/) =

t
1
(|)
Y
t
|
z(/, t)
t
2
(|)
Y
t
|
z(/, t) (5.21)
where Y
t
0 for all t (/), and

t(|)
Y
t
= 1.
Remark. The hypothesis of a nite set () may be unrealistic in the time domain
case, because the step response trajectory z(, t) is not necessarily analytic or piecewise
analytic, and may therefore attain the maximum value on one or several contact intervals
[t
, t
+
], where t
is the entry time, t

+
the exit time, and where it is reasonable to assume
that there are only nitely many such contact intervals. In that case, our method is
easily adapted, and (5.11) remains correct in so far as the full contact interval can be
represented by three pieces of information: the gradients
x
of the trajectory at x = t
,
x = t
+
, and one additional element
x
= 0 for say x = (t
+ t
+
)/2 on the interior of the
contact interval. (This is a dierence with the frequency domain case, where the functions
f(, ) are analytic, so that the phenomenon of a contact interval could not occur).
A more systematic approach to problems of this form with innite active sets would
consist in allowing choices of nite sets
e
(), where () ,
e
() is allowed. This leads
to a variation of the present algorithm discussed in [11, 119, 12], where a trust region
strategy replaces the present line search method.
Gradient computation By dierentiating the state-space equations (5.1) with respect
to /
ij
, we get
_
_
d
dt
x
|
ij
(/, t) = A
x
|
ij
(/, t) + B
2
u
|
ij
(/, t)
z
|
ij
(/, t) = C
1
x
|
ij
(/, t) +D
12
u
|
ij
(/, t)
y
|
ij
(/, t) = C
2
x
|
ij
(/, t)
(5.22)
controlled by
u
|
ij
(/, t) =
|
|
ij
(/, t)y(/, t) +/
y
|
ij
(/, t)
= y
j
(/, t)e
i
+/
y
|
ij
(/, t)
(5.23)
where e
i
stands for the i-th vector of the canonical basis of R
m
2
. It follows that the partial
derivative of the output signal
z
|
ij
(/, t) is the simulated output of the interconnection
in gure 5.2, where the exogenous input w is held at 0, and the vector y
j
(/, t)e
i
is added
to the controller output signal. We readily infer that n
u
n
y
simulations are required in
order to form the sought gradients.
/
y
j
e
i
0
+
+
y
/
ij
u
/
ij
z
/
ij
P
Figure 5.2: Interconnection for gradient computation.
This way of computing output signal gradients by performing closed-loop simulations
is at the root of the Iterative Feedback Tuning (IFT) method, intially proposed in [73] for
SISO systems and controllers. This optimization technique has originated an extensive
bibliography (see [72, 71, 59] and references therein) and was extended to multivariable
controllers [70]. Most of these papers illustrate the IFT with a smooth quadratic objective
function, minimized with the Gauss-Newton algorithm. In [78], the nonsmooth absolute
error is used, but a dierentiable optimization algorithm (DFP) is applied. Our approach
here diers both in the choice of the nonsmooth optimization criterion f
, and in the
design of a tailored nonsmooth algorithm as outlined in section 5.3.
Practical aspects The active time sets
1
(/) and
2
(/) are computed via numerical
simulation of the closed-loop system in response to the input signal w, see gure 5.1.
This rst simulation determines the time samples (t
l
)
0lN
that will be used throughout
the optimization phase. Measured output values
_
y(t
l
)
_
must be stored for subsequent
gradient computation. The extension
e
(/) is built from (/) by adding time samples
with largest envelope constraint violation (5.16), up to n
elements in all are retained.

According to our experiments the set extension generally provides a better model of
the original problem as captured by the optimality function
e
(5.11) and thus descent
directions (5.12) with better quality are obtained. The gradients
|
z(/, t
l
) (for t
e
(/)) result from n
u
n
y
additional simulations of the closed-loop (gure 5.2) at the
same time samples (t
l
)
0lN
.
5.4. Applications 115
5.4 Applications
In this section, we illustrate our nonsmooth technique on a variety of examples including
structured H
control, reliable control and time response shaping. We insist on the high
exibility of the proposed nonsmooth techniques and its excellent success in solving hard
practical problems.
5.4.1 Application to observer-based controller design
Observer-based controllers suit many realistic control problems, whenever the system state
variables are not directly available for feedback synthesis. Unlike unstructured dynamic
controllers, they have a straightforward physical interpretation, since controller states are
simply estimates of the plant states. In addition, observer-based controllers are easier to
implement and to re-tune.
Let x be an asymptotic estimate of the state vector x of the plant P (5.1). We assume
that (C
2
, A) is detectable, and that (A, B
2
) is stabilizable. As is well known, the estimate
x is obtained through the full-order Luenberger observer:
x = A x +B
2
u +K
f
(y C
2
x), (5.24)
where K
f
R
np
2
is the state estimator gain. The controlled input is u = K
c
x, where
K
c
R
m
2
n
is the state feedback gain. With the notations in (5.3), we get
/() =
_
A
K
B
K
C
K
D
K
_
=
_
A B
2
K
c
K
f
C
2
K
f
K
c
0
_
(5.25)
where the controller parameters are described as

_
vec K
f
vec K
c
_
R
n(p
2
+m
2
)
.
This state-space realization of the standard full-order observer-based controller is linear
in . It follows that the Jacobian matrix /
t
() only depends on the plant data A, B
2
and
C
2
.
Nonsmooth H
synthesis with observer-based controllers has been tested on the fol-

lowing problem discussed in [1], taken from the Matlab library of mu-synthesis examples
for Simulink. The state-space equations of the controlled plant are:
G(s) :
_
x
G
y
G
_
=
_
_
36.6 18.923 0.414
1.9 0.983 77.8
0 57.3 0
_
_
_
x
G
u
G
_
(5.26)
The exogenous inputs are the disturbances w
1
(on the actuator signal) and w
2
(on the
measurement). Two frequency domain weights W
u
(s) and W
y
(s) are applied to u =
u
G
w
1
respectively to y = y +w
1
, dening the regulated outputs z
1
and z
2
. They are
W
u
(s) =
50s + 5000
s + 10000
, and W
y
(s) =
0.5s + 1.5
s + 10000
. (5.27)
The resulting augmented model (gure 5.3) has order 4, with state vector x = [x
G
, x
u
, x
y
]
T
,
and the realization used in our computations is the following:
P(s) :
_
_
x
z
y
_
_ =
_
_
36.6 18.923 0 0 0.414 0 0.414
1.9 0.983 0 0 77.8 0 77.8
0 0 10000 0 0 0 512
0 3667.2 0 10000 0 64 0
0 0 966.8 0 0 0 50
0 28.65 0 78.102 0 0.5 0
0 57.3 0 0 0 1 0
_
_
_
_
x
w
u
_
_
(5.28)
Notice that the static gains K
f
and K
c
depend on the state-space realization chosen for
P, since the controller state variable x estimates x.
z
1
z
2
y
+
+
w
1
+
+
W
u
W
y
G
u
K()
w
2
Figure 5.3: Observer-based feedback with augmented model.
Running our nonsmooth algorithm to minimize |T
wz
(/())|
, yields the gains

K
f
= [3886, 7203, 163.3, 29.22]
T
K
c
= [484, 219.2, 1876, 152.9]
. (5.29)
Correspondingly, the achieved H
norm is 0.5168, which is about the same as the H
performance of the standard full-order H
controller = 0.5109.
5.4.2 Application to reliable control
Reliable or fault-tolerant control has to guarantee satisfactory stability and performance
in nominal conditions as well as in scenarios, where some system components turn faulty
or deviate from nominal conditions. The active approach uses a fault detection scheme
or identication procedure to adjust the controller in real-time to ensure safe opera-
tion. Here the focus is on the passive approach, where a single controller is required
to handle most plausible default scenarios. See [129] and the survey [111] for a com-
prehensive view. We reformulate this problem as a doubly structured design problem
and use our nonsmooth method to derive a local solution. In gure (5.4) we show on
the left side various synthesis scenarios, where a single (unstructured or already struc-
tured) controller /() must simultaneously minimize the closed-loop performance chan-
nels [[T
w
i
z
i
(/())[[
for i = 1, . . . , N. This is readily formulated as a single structured

H
synthesis problem, where the sought controller K() has a repeated block-diagonal
structure K() = 1 /(). See the right-hand side of gure 5.4. A state-space represen-
tation of the plant P(s) is readily derived as
_
_
A B
1
B
2
C
1
D
11
D
11
C
2
D
21
D
22
_
_ =
_
_
diag
i=1,...,N
A
i
diag
i=1,...,N
B
i
1
diag
i=1,...,N
B
i
2
diag
i=1,...,N
C
i
1
diag
i=1,...,N
D
i
11
diag
i=1,...,N
D
i
12
diag
i=1,...,N
C
i
2
diag
i=1,...,N
D
i
21
diag
i=1,...,N
D
i
22
_
_
.
where the superscript i refers to the ith scenario and where appropriate dynamic augmen-
tations have been performed beforehand for dynamic controllers. Note that the controller
is now doubly-structured in the sense that the already structured data /() are repeated
in the block-diagonal structure I
N
/(), which we describe by the diagonal block rep-
etition operator 1 above (see gure 5.4). This ts nicely into the general framework
presented in sections 5.2 and 5.3, because the chain rule can obviously be used to cope
with a diagonal augmentation 1 / of controller parameterizations.
P
w
N
z
N
y
N
P
N
w
1
z
1
y
1
P
1
w
z
y
u
N
u
1
u
.
.
.
.
.
.
/()
T
w
N
z
N
(/())
T
w
1
z
1
(/())
/()
/()
T
wz
(/())
/()
Figure 5.4: Reliable control based on N scenarios.
We illustrate the synthesis technique by means of the following example from [142].
It consists of 10 plants G
0
to G
9
with G
0
(s) =
1
s1
, the nominal model and
G
1
(s) =
6.1
s
2
+5.1s6.1
, G
2
(s) =
1.425
s1.425
G
3
(s) =
0.67
s0.67
,
G
4
(s) =
(s14.29)
(s+14.29)(s1)
, G
5
(s) =
4900
(s1)(s
2
+21s+4900)
, G
6
(s) =
4900
(s+777.7)(s+6.301)(s1)
,
G
7
(s) =
1562510
6
(s+50)
6
(s1)
, G
8
(s) =
2.9621(s9.837)(s+0.7689)
(s+32)(s1)(s+0.5612)
, G
9
(s) =
4.991(s
2
+3.672s+34.85)
(s+32)(s+7.241)(s1)
,
as faulty modes. The design objective is to minimize
max
i=0,...,9
[[W(s)T
w
i
z
i (K)[[
, where T
w
i
z
i (I +G
i
(s)K(s))
1
(5.30)
and W(s) =
0.25s+0.6
s+0.006
is a weighting function penalizing the low frequency range to achieve
good tracking properties in response to a step command. Dierent controller structures
/() will be considered in the sequel. The author in [142] investigated various options
including a controller with good nominal performance K
1
(s) =
10(0.9s+1)
s
, a robust con-
troller K
2
(s) =
2.8s+1
s
, and a generalized internal model control (GIMC) implementation
based on both K
1
(s) and K
2
(s), which we shall denote K
gimc
(s). The latter is a 6th
order controller with Youla parameter
Q
0.68889s(s + 1.452)(s + 1)
(s + 1.111)(s
2
+ 1.8s + 1)
.
Step responses of these controllers for nominal and faulty modes of the plant are
displayed in gures 5.6 to 5.8. Clearly, K
1
has good nominal performance, but is not
satisfactory on faulty modes. On the other hand, controller K
2
shows very robust per-
formance, but responses are slower and the controller performs poorly in the nominal
case. The GIMC implementation dramatically improves the situation, since good nomi-
nal performance is maintained and the controller also gives good performance on faulty
modes.
We have used the proposed nonsmooth technique to minimize the worst case objective
in (5.30) over the set of PID feedback controllers with parametrization given in (5.6). The
following parameter values were obtained:
= 0.0100, R
i
= 1.1589, R
d
= 0.4883, D
K
= 4.0263 .
The achieved max H
norm is 0.4886, and step responses with this controller are shown
in gure 5.9. This controller shows clearly better performance than K
1
(s) or K
2
(s), but
performs slightly worse than K
gimc
(s). Note that the GIMC controller can also be viewed
as a two-degree of freedom controller [141], which suggests using a two-degree of freedom
architecture to improve the PID feedback controller. We have used the model following
synthesis interconnection shown in gure 5.5, where G
ref
is a reference model to compute
a second-order feedforward controller F(s) with xed feedback controller K(s) = K
PID
.
As before, the feedforward controller is required to minimize the worst-case tracking error
from r to e over simultaneously the nominal model and the faulty modes
minimize
F(s)
max
i=0,...,9
[[T
i
re
(F(s))[[
.
With a slow reference model
G
ref
=
11.11
s
2
+ 6s + 11.11
,
the following feedforward controller was obtained
F(s) =
3.0642(s + 1.976)(s + 1.273)
(s + 3.796)(s + 1.049)
.
With a faster reference model
G
ref
=
28.7
s
2
+ 7.5s + 28.7
,
we have obtained
F(s) =
1.1684(s
2
+ 1.46s + 18.51)
s
2
+ 5.015s + 18.39
.
Note that the overall controller has order 4 since it consists in a PID controller in the
feedback path and a second-order controller for feedforward action. Steps responses are
given in gures 5.10 and 5.11, respectively, for slow and fast feedforward controllers. Note
that the latter controller appears to be the best obtained so far in terms of settling time
and overshoot.
+
G
ref
(s)
e
K(s)
G(s)
+
+
+
F(s)
r
Figure 5.5: Feedforward design with reference model.
0 1 2 3 4 5 6 7 8 9 10
1
0.5
0
0.5
1
1.5
2
2.5
3
time (sec.)

o
u
t
p
u
t

Figure 5.6: Step response with perfor-
mance controller.
0 1 2 3 4 5 6 7 8 9 10
0.5
0
0.5
1
1.5
2
time (sec.)

o
u
t
p
u
t

Figure 5.7: Step response with robust
controller.
0 1 2 3 4 5 6 7 8 9 10
0.5
0
0.5
1
1.5
2
time (sec.)

o
u
t
p
u
t

Figure 5.8: Step response with general-
ized internal model controller.
0 1 2 3 4 5 6 7 8 9 10
0.5
0
0.5
1
1.5
2
time (sec.)

o
u
t
p
u
t

Figure 5.9: Step response with nons-
mooth PID controller.
0 1 2 3 4 5 6 7 8 9 10
0.5
0
0.5
1
1.5
2
time (sec.)

o
u
t
p
u
t

Figure 5.10: Step response with PID
+ feedforward control, slow reference
model.
0 1 2 3 4 5 6 7 8 9 10
0.5
0
0.5
1
1.5
2
time (sec.)

o
u
t
p
u
t

Figure 5.11: Step response with PID +
feedforward control, fast reference model.
5.4.3 Application to PID time response shaping
In order to illustrate the ability of the proposed nonsmooth technique to handle time-
domain constraints and controller structure, the following test examples from [88] have
been used
G
1
(s) =
1
20s+1
e
5s
, G
2
(s) =
1
20s+1
e
20s
G
3
(s) =
1
(10s+1)
8
, G
4
(s) =
5s+1
(10s+1)(20s+1)
.
(5.31)
The time delays are replaced by a third-order Pad approximation, which generates non-
minimum phase zeros in the transfer functions G
1
and G
2
. Each of these plants is regu-
lated by a two-degree-of-freedom (2-DOF) PID controller, as shown in gure 5.12, with
K
r
(s) = K
p
+
K
i
s
and K
y
(s) = K
p
+
K
i
s
+
K
d
s
1+s
.
K
r
(s) G(s)
K
y
(s)
u y + r
Figure 5.12: 2-DOF with PID.
/()
r
_
y P
G
w = r z = y
y
u
Figure 5.13: Equivalent standard form to
2-DOF with PID.
This 2-DOF controller structure is equivalent to the following state-space representa-
tion, in the framework shown in gure 5.13, with the notations from (5.7):
/() =
_
_
0 0 0 R
i
0
0 0 R
d
0
0 0 0 0 R
i
1 1 1 D
K
D
K
+
R
d
_
, (5.32)
where the search space is = [, R
i
, R
d
, D
K
]
T
R
4
.
For each of the four systems (5.31), the envelope functions were chosen piecewise
constant, in order to constrain the overshoot z
os
and the settling time at 2% of the
steady state. Using the indicatrix function 1
[a,b[
to restrict to intervals, we dene
z
max
= (1 + z
os
) 1
[0,t
s
[
+ 1.02 1
[t
s
,+[
z
min
=
_
on [0, t
s
[
0.98 on [t
s
, +[
(5.33)
The vector was initialized with the PID parameters obtained with the Ziegler-Nichols
(ZN) tuning rules in [88], a heuristic which gives unsatisfactory step responses for the
plants (5.31), with high overshoot and slow settling times. The results are summarized in
table 5.1, where NS indicates the controller parameters obtained from the nonsmooth
optimization. They are compared to those from [88], denoted by IFT. The corresponding
step responses are drawn in gures 5.14 to 5.17.
Obviously, arbitrary low z
os
or t
s
are not necessarily achievable with the structured
controller /(). If the chosen time envelope is too restrictive, the algorithm will converge
to a minimizer of f
(i.e.
e
( ) 0) with f
> 0.
Conversely, as the algorithm returns a local minimizer , f
( ) > 0 does not necessary

mean that the time constraints could not be satised. A restart with another initial
controller may very well lead to f
= 0. This diculty appeared for plant G

2
(resp.
G
3
): with the ZN controller as starting point, we reached f
= 3.4033 10
2
(resp.
f
= 1.3907 10
2
) at convergence, which was an unsatisfactory local minimum. We got
around this diculty by restarting the algorithm from a new point, denoted by NEW
in table 5.1.
plant contr. K
p
K
i
K
d
t
s
(s) z
os
(%)
G
1
ZN (init) 10
3
4.0588 0.4388 9.3860 46.22 46.92
G
1
IFT 10
3
3.6717 0.1324 7.7311 21.34 5.37
G
1
NS 10
3
3.3315 0.1221 9.3874 20.94 1.95
G
2
NEW (init) 10
3
0.5000 0.0500 5.0000 319.45 45.72
G
2
IFT 10
3
0.9303 0.0309 5.6332 50.15 0.90
G
2
NS 9.9992 10
4
0.9361 0.0305 4.9920 49.35 1.73
G
3
NEW 10
3
0.5000 0.0100 10.0000 235.54 0.11
G
3
IFT 10
3
0.6641 0.0123 12.0959 131.99 1.00
G
3
NS 10
3
0.6590 0.0121 9.9997 129.91 1.97
G
4
ZN 10
3
3.5294 0.2101 14.8235 69.22 53.82
G
4
IFT 10
3
3.0279 0.0654 18.4075 28.31 0.53
G
4
NS 10
3
2.8947 0.0615 14.8247 25.58 1.67
Table 5.1: PID parameters, settling time and overshoot.
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
1.5
t
z
(
t
)

ZN (initial)
IFT
NS
z
max
z
min
Figure 5.14: G
1
step responses with
PIDs.
0 50 100 150 200 250 300
0
0.5
1
1.5
t
z
(
t
)

initial
IFT
NS
z
max
z
min
Figure 5.15: G
2
step responses with
PIDs.
0 50 100 150 200 250 300 350 400 450 500
0
0.2
0.4
0.6
0.8
1
1.2
t
z
(
t
)

initial
IFT
NS
z
max
z
min
Figure 5.16: G
3
step responses with
PIDs.
0 20 40 60 80 100 120 140 160 180 200
0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
t
z
(
t
)

ZN (initial)
IFT
NS
z
max
z
min
Figure 5.17: G
4
step responses with
PIDs.
Finally, we mention that we did not use any temporal mask as is classically done in
the IFT approach to avoid spurious solutions.
5.5 Conclusion
We have described a general and very exible nonsmooth algorithm to compute locally
optimal solutions to synthesis problems subject to frequency- or time-domain constraints.
Our method oers the new and appealing possibility to integrate controller structures of
practical interest in the design. We have now several encouraging reports of successful
experiments, which advocate the use of nonsmooth mathematical programming techniques
when it comes to solving dicult (often NP-hard) design problems. The results obtained
in this paper corroborate previous studies on dierent problem classes. Extension of our
nonsmooth technique to problems involving a mixture of frequency- and time-domain
constraints seems a natural next step, which is near at hand. For time-domain design, we
have noticed that the proposed technique assumes very little about the system nature,
except the access to simulated responses. A more ambitious goal would therefore consider
extensions to nonlinear systems.
Ackowledgements
The authors acknowledge nancial support from the Fondation dentreprise EADS (under
contract Solving challenging problems in feedback control), and from the Agence Nationale
de Recherche (ANR) (under project NT05 Guidage).
Chapitre 6
Optimisation non lisse du second
ordre pour la synthse H
Les applications dveloppes dans les chapitres prcdents sappuient toutes sur le
mme schma algorithmique doptimisation non lisse, prsent dans la premire partie
(section 2.3.3), rappel et gnralis plusieurs reprises (sections 3.5.1, 4.2.5 et 5.3). Il
sagit dune mthode du premier ordre, convergence R-linaire en gnral. Nous avons
dj propos daner le modle quadratique servant la construction de la direction de
descente, laide dune mise jour de type BFGS (voir section 3.5.2). Sa mise en uvre
dans un cadre non direntiable est cependant tout fait heuristique, et, si un gain en
terme de nombre ditrations a pu tre observ sur certains problmes de stabilisation
(chapitre 3), nous ne pouvons cependant conclure de faon gnrale.
Ce dernier chapitre dcrit un algorithme du second ordre pour la synthse H
, dont
le mcanisme dire fortement de lapproche prcdente : le problme non lisse est tout
dabord reformul comme un problme semi-inni doptimisation non linaire. Celui-ci est
ensuite ramen par rduction locale un nombre ni de contraintes, et peut tre rsolu par
la mthode SQP (Programmation Quadratique Successive), dont la convergence locale
1
est superlinaire. Lalgorithme propos couple la mthode SQP lapproche du premier
ordre prcdente, an de garantir la proprit de convergence globale.
Les exprimentations numriques montrent un apport indiscutable de cette mthode en
terme de nombre ditrations ncessaires pour la convergence. Le gain en temps de calcul
est cependant moins net : trs important dans certains cas, inexistant dans dautres. En
eet, la globalisation de lalgorithme sappuie sur un recours la mthode du premier
ordre, et ncessite donc chaque itration de calculer deux directions : celle de Cauchy
(premier ordre) et celle de Newton (SQP), pour lesquelles les gradients et hessiens sont
calculs de manire exacte.
Enn, ces travaux tant chronologiquement antrieurs aux applications prsentes dans
les trois chapitres prcdents, nous navons pas envisag le cas, pourtant essentiel en pra-
tique, des correcteurs structurs. La description de ces travaux dans les pages qui suivent
est cependant motive par leur originalit en regard des dveloppements prcdents, et
1
Un algorithme doptimisation converge localement vers un minimum x
sil existe r > 0 tel que pour

tout x
0
B(x
, r), la suite des itrs (x

k
)
k0
converge vers x
. La convergence est dite globale si elle a

lieu pour tout x
0
R
n
.
125
126 Chapitre 6. Optimisation non lisse du second ordre pour la synthse H
par le gain en temps de calcul constat sur certains exemples.

6.1. Introduction 127
Second-order nonsmooth optimization for
H
synthesis
V. Bompart, D. Noll and P. Apkarian
Article publi dans Numerische Mathematik,
107(3) :433-454, 2007.
Abstract
The standard way to compute H
feedback controllers uses algebraic Riccati equations

and is therefore of limited applicability. Here we present a new approach to the H
output
feedback control design problem, which is based on nonlinear and nonsmooth mathemat-
ical programming techniques. Our approach avoids the use of Lyapunov variables, and is
therefore exible in many practical situations.
6.1 Introduction
Well designed feedback control systems in high technology elds are expected to respond
favorably to a list of concurring performance specications such as stability, good reg-
ulation against disturbances, desirable responses to commands, robustness, control law
specications, system security, and much else. In addition, controllers should be hardware
implementable at low cost, and should allow a exible adaption to strategic changes dur-
ing the model building process. These ever growing demands have shown the limitations
of currently used mathematical tools for synthesis, which are based on solving algebraic
Riccati equations (AREs). The rise of linear matrix inequalities (LMIs) since the early
1990s has certainly improved the situation, but the main limitations of these approaches
persist.
In response, various new design methods have been proposed since the late 1990s,
including methods from global optimization, [16, 18, 138, 124], matrix inequality con-
strained nonlinear programming, [50, 86, 87, 14, 134], augmented Lagrangian methods,
[49, 106], eigenvalue optimization, [6, 8, 9, 76], and others. Presently we discuss a new
strategy for H
output feedback control synthesis, which has several advantages over

existing approaches, because it avoids the use of Lyapunov variables. We shall show that
this leads to much smaller and better manageable problems, at the price that controller
synthesis becomes a nonsmooth optimization program. We develop local nonsmooth opti-
mization strategies suited for this new context, and show that they improve the situation
considerably.
6.2 Problem setting

We consider a linear time-invariant plant described in standard form by the state-space
equations
P(s) :
_
_
x
z
y
_
_ =
_
_
A B
1
B
2
C
1
D
11
D
12
C
2
D
21
0
p
2
m
2
_
_
_
_
x
w
u
_
_ ,
where s is the Laplace variable, x R
n
is the state of the system, u R
m
2
the control,
w R
m
1
an exogenous signal (noise, commands), z R
p
1
the regulated variable (response
of the system to the input w), and y R
p
2
the measured output. The goal of feedback
control synthesis is to design an output feedback controller
K(s) :
_
x
K
u
_
=
_
A
K
B
K
C
K
D
K
_ _
x
K
y
_
where x
K
R
n
K
, A
K
R
n
K
n
K
, B
K
R
n
K
p
2
, C
K
R
m
2
n
K
and D
K
R
m
2
p
2
, such
that the following three criteria are met:
Internal stability: The closed-loop system obtained by substituting K into P is expo-
nentially stable.
Performance: Among all internally stabilizing controllers, K minimizes the H
-norm
of the performance channel w z.
Control law specications: The controller K may be subject to additional structural
constraints. For instance, one may look for decentralized controllers, reduced-order con-
trollers, PID, feed-forward, lead lag structures, etc.
Let us explain the performance requirement for the case u = Ky of a static output
feedback controller, where D
K
= K R
m
2
p
2
. Here the closed-loop transfer function for
the performance channel w z is
T
wz
(K)(s) = T(K) +c(K)
_
sI
n
/(K)
_
1
B(K) (6.1)
where the closed loop data are
/(K) = A +B
2
KC
2
B(K) = B
1
+B
2
KD
21
c(K) = C
1
+D
12
KC
2
T(K) = D
11
+D
12
KD
21
For brevity, we shall subsequently write T(K, ) T
wz
(K)(j). Denoting by
1
the
maximum eigenvalue function on the space of Hermitian matrices, and by the maximum
singular value of m
1
p
1
matrices, we dene
f(K, ) =
1
(T(K, )
H
T(K, )) = [ (T(K, ))]
2
,
where Z
H
stands for the conjugate transpose of the complex matrix Z. The square of the
H
norm is then dened as

|T
wz
(K)|
2
= max
[0,]
f(K, ). (6.2)
6.3. Approach via semi-innite programming 129
Optimizing the performance of the controller therefore leads to minimizing the function
f(K) = max
[0,]
f(K, ) over the set o of all closed-loop exponentially stabilizing
controllers K R
m
2
p
2
. Notice that f is nonsmooth with two possible sources of nons-
moothness: (a) the innite max-operator, and (b) the nonsmoothness of
1
, which may
lead to nonsmoothness of f(, ) for xed .
Remark. A diculty is that the set of exponentially stabilizing controllers o is open and
not a constraint set in the usual sense of nonlinear programming. We have observed that
it is usually possible to ignore the constraint K o, and start the minimization (6.3)
with an initial closed-loop stabilizing controller K o. The fact that |T
wz
(K)|
is
minimized, and therefore stays nite, very often implies closed-loop stability of the limit
point

K. That is why we shall consider program
min
KR
m
2
p
2
f(K) = |T
wz
(K)|
2
.
(6.3)
under the implicit assumption that iterates K
k
and the limit

K are closed-loop stabilizing.
In those cases where |T
wz
(K)|
< does not imply closed-loop stability, a problem

occurs. Here a minimizing sequence K
k
o for (6.3) will converge to a controller

K on
the boundary of the domain o. Such a

K is not a valid solution of the control problem.
Fortunately, the phenomenon is easily detected, and a practical way out is to consider an
additional closed-loop transfer channel T
stab
(K, s) = (sI /(K))
1
, called the stabilizing
channel. For a small parameter > 0, we may then minimize the objective
f(K) = max
_
|T
wz
(K)|
2
, |T
stab
(K)|
2
_
, (6.4)
and we are now sure that f(K) < implies closed-loop stability of K. Notice that f is
again of the form (6.3) if a transfer matrix with the two diagonal blocks T
wz
and T
stab
is introduced.
Remark. Our approach to H
controller synthesis based on program (6.3) diers sub-

stantially from standard ways to solve such problems, where the H
norm is avoided
and replaced by a matrix inequality using the bounded real lemma [30]. In solving (6.3)
directly we dispense with Lyapunov variables, which leads to a spectacular reduction in
the number of unknowns, especially when small controllers for large systems are sought.
In exchange we have to deal with nonsmoothness and with semi-innity, and those are
addressed during the next sections.
First-order nonsmooth optimization methods for (6.3) have for the rst time been pro-
posed in [6], and have been further developed in [8, 10, 9]. Here we propose a second-order
approach, which when combined with the rst-order technique speeds up the minimization
process at the end, improves the nal precision, and serves as an additional convergence
certicate.
6.3 Approach via semi-innite programming
In the previous section we have introduced program (6.3) for the case of static controllers,
and we shall continue to do so because the general case of dynamic K may be reduced to
the static case by prior dynamic augmentation of the system. More precisely, the following
substitutions are performed:
K
_
A
K
B
K
C
K
D
K
_
, A
_
A 0
0 0
k
_
B
1

_
B
1
0
_
, C
1

_
C
1
0
_
B
2

_
0 B
2
I
k
0
_
, C
2

_
0 I
k
C
2
0
_
, D
12

_
0 D
12
_
, D
21

_
0
D
21
_
.
(6.5)
Now we have the following
Lemma 6.3.1. For a closed-loop stabilizing controller K, the set of active frequencies
(K) [0, ] : f(K) = f(K, ) is either nite, or (K) = [0, ], that is,
f(K, ) = f(K) for all .
Proof. Notice that as K is xed, we are concerned with a one parameter family
T(K, )
H
T(K, ) of m
1
m
1
Hermitian matrices. Here the bundle of the m
1
eigenvalues
i
_
T(K, )
H
T(K, )
_
as functions of has a very specic structure: it consists of m
1
real analytic functions
i
(), R; see e.g. [79, p. 82, p. 138], or [91]. More precisely,
there are m
1
1
, . . . ,
m
1
such that
1
(T(K, )), . . . ,
m
1
(T(K, )) = [
1
()[, . . . , [
m
1
()[, (6.6)
where
i
(T(K, )) are the singular values of T(K, ). In consequence, the maximum
eigenvalue function (6.2) is the maximum of m
1
2
i
(). These
2
i
are also analytic at , because we are dealing with a stable (proper) transfer matrix
T(K, ), which is rational as a function of . (The statement above may be made more
precise: There exist functions
i
which are analytic on a strip B on the Riemann sphere
S
2
= C , such that B contains the meridian S
1
= R passing through the
north pole , with
i
taking real values on S
1
, such that (6.6) is satised for all S
1
.)
Suppose now (K) is innite. Then one of these m
1
2
i
has an
innity of maxima on S
1
with the same value f(K). Since S
1
is compact, these maxima
have an accumulation point . In terms of the analytic extensions on B, that means the
Taylor expansion of the
2
i
in question at is the same as the Taylor expansion of the
function with constant value f(K). This implies indeed
2
i
() = f(K) for all .
Remark. 1) The stable transfer function : s
s1
s+1
satises (j)(j) = 1, so it is
theoretically possible that (K) = [0, ]. This may even arrive in practice: It is known
that the so-called central (full order) optimal H
-controller K
renders the frequency

plot f(K
, ) at or all-pass.
2) Notice also that we cannot omit the frequency = in the discussion, as witnessed
by the stable transfer function (s) : s
s+1
s+2
. Here max
[0,]
(j)(j) = 1 is
attained at = .
3) The frequencies (K) will be called the peaks of the frequency curve at
K. Local maxima of the frequency curve f(K, ) which are not peaks will be
called secondary peaks. Notice that the above argument also shows that the number of
secondary peaks is nite.
6.3. Approach via semi-innite programming 131
For the technical formulas we shall concentrate on those K, where the set of active
frequencies or peaks (K) = [0, ] : f(K) = f(K, ) is nite. In practice we have
never observed at frequency curves when the order of the controller is smaller than the
order of the plant.
During the following we will analyze the case where the multiplicity of
1
(T(K, )
H
T(K, ))
is one at all active frequencies . This is motivated by practical considerations, because
nonsmoothness (b) never occurred in our tests. The necessary changes required for the
general case will be discussed in Section 6.4.3.
It is useful to cast program (6.3) as an equivalent constrained semi-innite program
minimize t
subject to f(K, ) t 0, [0, ]
(6.7)
with decision variable (K, t). In [67] three approaches to semi-innite programming are
discussed: exchange of constraints, discretization, and local reduction. Here we use a local
reduction method. The main ideas are recalled below; see also [77] for this approach.
Let (

K,
t), with

t = f(

K), be a local solution of (6.7). Indexing the active frequencies
(

K) =
1
, . . . ,
p
at

K, we suppose that the following conditions are satised
(i) f
t
(

K,
i
) = 0, i = 1, . . . , p.
(ii) f
tt
(

K,
i
) < 0, i = 1, . . . , p.
(iii) f(

K, ) < f(

K), for every / (

K) =
1
, . . . ,
p
.
These assumptions dene the setting denoted as the standard case in semi-innite
programming [67]. The three conditions express the fact that the frequencies
i
(

K)
are the strict global maximizers of f(

K, ). Notice that condition (iii) is the niteness
hypothesis already mentioned, justied by Lemma 6.3.1, while condition (ii) is slightly
conservative, because the necessary optimality condition only tells us that f
tt
(

K,
i
) 0.
Remark. Notice that f(

K, ) is twice dierentiable at each
i
even without the hypothesis
that the maximum singular values of the T(K,
i
) have multiplicity 1. This is because the
maximum singular value function (G(j)) of a stable transfer matrix G is always
twice continuously dierentiable at its local maxima [28].
Conditions (i) and (ii) allow the implicit function theorem, according to which we can
nd a neighborhood U of

K, and neighborhoods V
i
of
i
(i = 1, . . . , p), together with
C
1
-functions
i
: U V
i
, such that the following conditions are satised
(iv)
i
(

K) =
i
, i = 1, . . . , p.
(v) f
t
(K,
i
(K)) = 0, i = 1, . . . , p.
(vi) Whenever K U and V
i
satisfy f
t
(K, ) = 0, then =
i
(K).
Condition (ii) says f

tt
(

K,
i
) < 0, so by shrinking U if required, we may arrange that
f
tt
(K,
i
(K)) < 0 for K U. Then
i
(K) are local maxima of f(K, ). Moreover, by
(vi),
i
(K) is the only critical point of f(K, ) in V
i
. We then have the following
Lemma 6.3.2. Under conditions (i) (iii) the neighborhood U of

K may be chosen
such that max
[0,]
f(K, ) = max
i=1,...,p
f(K,
i
(K)) for every K U. In particular, (K)
1
(K), . . . ,
p
(K) for every K U.
Proof. Choose U and V
i
such that conditions (iv) - (vi) are satised. Suppose that
contrary to the statement there exists a sequence K
r

K such that
m
r
max
i=1,...,p
f(K
r
,
i
(K
r
)) < max
[0,]
f(K
r
, ) =: M
r
.
Pick
r
where the maximum M
r
is attained. Passing to a subsequence, we may assume
for some
[0, ]. Since m
r
f(

K) and M
r
f(

K), we have f(

K) =
f(

K,
). By axiom (iii), that means
=
i
for some i = 1, . . . , p. Then (K
r
,
r
) UV
i
from some index r
0
onwards.
But f
t
(K
r
,
r
) = 0, because
r
is a maximum at K
r
, so condition (vi) implies
r
=
i
(K
r
). That shows m
r
= M
r
for r r
0
, a contradiction.
Altogether, we have shown that program (6.3) is locally equivalent to the standard
constrained nonlinear program
minimize t
subject to f (K,
i
(K)) t 0, i = 1, . . . , p
(6.8)
which we may solve via a SQP method. In the next section we will discuss how this
should be organized, and that the jet information may be computed eciently. Local
convergence of this approach will be assured under the following hypothesis
(vii) f
t
K
(

K,
1
), . . . , f
t
K
(

K,
p
) are linearly independent
because this guarantees that program (6.8) satises the linear independence constraint
qualication hypothesis (LICQ).
6.4 Solving with SQP
6.4.1 Quadratic tangent subproblem
In this section we assume that for K in a neighborhood of

K, the set
1
(K), . . . ,
p
(K)
is available. The computation of this set will be discussed in Section 6.5.1.
In order to derive the tangent quadratic program for (6.8), let us write G
i
(K, t) =
f (K,
i
(K)) t, F(K, t) = t. The Lagrangian of (6.8) is then
L(K, t; ) = F(K, t) +
p
i=1
i
G
i
(K, t),
6.4. Solving with SQP 133
so that
L
t
K
(K, t; ) =
p
i=1
i
(f
t
K
(K,
i
(K)) +f
t
(K,
i
(K))
t
i
(K))
=
p
i=1
i
f
t
K
(K,
i
(K))
using condition (v) above. Similarly,
L
t
t
(K, t; ) = 1
p
i=1
i
.
The second-order elements are L
tt
Kt
(K, t; ) = 0, L
tt
tt
(K, t; ) = 0, and
L
tt
KK
(K, t; ) =
p
i=1
i
[f
tt
KK
(K,
i
(K))
+f
t
(K,
i
(K))
tt
i
(K)
+f
tt
K
(K,
i
(K))
t
i
(K)
]
=
p
i=1
i
_
f
tt
KK
(K,
i
(K)) +f
tt
K
(K,
i
(K))
t
i
(K)
_
,
again using condition (v). Dierentiating (v) gives
0 = f
tt
K
(K,
i
(K)) +f
tt
(K,
i
(K))
t
i
(K),
which allows us to express
t
i
(K) through derivatives of f. Altogether,
L
tt
(K, t; ) =
_
L
tt
KK
(K, t; ) 0
0 0
_
with
L
tt
KK
(K, t; ) =
p
i=1
i
[f
tt
KK
(K,
i
(K))
f
tt
K
(K,
i
(K)) f
tt
(K,
i
(K))
1
f
tt
K
(K,
i
(K))
].
The tangent quadratic program is now
minimize t +
1
2
K
L
tt
KK
(K, t; )K
subject to f (K,
i
(K)) +f
t
K
(K,
i
(K))
K t t 0,
i = 1, . . . , p
t R, K R
m
2
p
2
(6.9)
6.4.2 First and second derivative formulas

We rst recall some useful results from matrix perturbation theory, concerning the rst
derivatives of a nondegenerate eigenvalue and an associated eigenvector. General formulae
may be found in [61]. We specialize to the case of a Hermitian matrix depending on real
parameters: given A(u), a nn complex Hermitian matrix depending smoothly on u R
p
,
we denote by
1
(u)
n
(u) the eigenvalues of A(u), and by
_
q
1
(u), . . . , q
n
(u)
_
an
orthonormal basis of eigenvectors, such that
_
_
A(u) = Q(u)(u)Q(u)
H
,
(u) = diag
_
1
(u), . . . ,
n
(u)
_
and
Q(u) =
_
q
1
(u), . . . , q
n
(u)
_
.
For a xed k and a given u
0
R
p
, under the assumption that
k
(u
0
) is a simple (nonde-
generate) eigenvalue of A(u
0
),
k
and q
k
are dierentiable at u
0
, and for all u in R
p
, we
have
t
k
(u
0
).u = q
k
(u
0
)
H
_
A
t
(u
0
).u
_
q
k
(u
0
)
q
t
k
(u
0
).u =
n
i=1
i,=k
q
i
(u
0
)
H
_
A
t
(u
0
).u
_
q
k
(u
0
)
k
(u
0
)
i
(u
0
)
q
i
(u
0
)
Notice from the second equality that q
t
k
(u
0
).u is orthogonal to the eigenvector q
k
(u
0
).
Furthermore we derive from these expressions that
k
is twice dierentiable at u
0
, and
for all u
1
, u
2
in R
p
, dropping the dependency on u
0
for the right hand side terms, we
get
tt
k
(u
0
).(u
1
, u
2
) = q
H
k
(A
tt
.(u
1
, u
2
)) q
k
+ 2 Re
_
_
_
_
n
i=1
i,=k
q
H
k
(A
t
.u
1
) q
i
q
H
i
(A
t
.u
2
) q
k
i
_
_
_
_
.
We specialize to the case of f(K, ), the squared maximum singular value of the
transfer matrix T(K, ). Then u = (K, ) and A(u) = A(K, ) = T(K, )
H
T(K, ), and
f =
1
A.
We note eigenvalues of T(K, )
H
T(K, ) as (
i
)
1im
1
and an orthonormal basis of
associated eigenvectors as (q
i
)
1im
1
. We drop the dependency on K and for ease of
notation. We assume that
1
>
2

3

m
1
. Applying the above formulae gives:
f
t
K
(K, ).K = 2 Re
_
q
H
1
T
H
(T
t
K
.K)q
1
_
f
t
(K, ) = 2 Re
_
q
H
1
T
H
T
t
q
1
_
f
tt
KK
(K, ).(K
1
, K
2
) = 2 Re
_
_
q
H
1
_
(T
t
K
.K
2
)
H
T
t
K
.K
1
+T
H
T
tt
KK
.(K
1
, K
2
)
_
q
1
+
m
1
i=2
q
H
1
_
T
H
T
t
K
.K
1
+ (T
t
K
.K
1
)
H
T
_
q
i
q
H
i
_
T
H
T
t
K
.K
2
+ (T
t
K
.K
2
)
H
T
_
q
1
i
_
_
f
tt
(K, ) = 2 Re
_
_
q
H
1
_
T
t
H
T
t
+T
H
T
tt
_
q
1
_
_
+ 2
m
1
i=2
q
H
1
_
T
H
T
t
+T
t
H
T
_
q
i
i
f
tt
K
(K, ).K = 2 Re
_
_
q
H
1
_
(T
t
K
.K)
H
T
t
+T
H
T
tt
K
.K
_
q
1
+
m
1
i=2
q
H
1
_
T
H
T
t
+T
t
H
T
_
q
i
q
H
i
_
T
H
T
t
K
.K + (T
t
K
.K)
H
T
_
q
1
i
_
_
The rst and second partial derivatives of the transfer function T can be expanded
with the closed-loop state-space matrices /, B and c. Starting out with formula (6.1),
we derive
T
t
K
.K = T
t
K
(K, ).K
=
_
D
12
+c
_
jI
n
/
_
1
B
2
_
K
_
C
2
_
jI
n
/
_
1
B +D
21
_
T
tt
KK
.(K
1
, K
2
) = T
tt
KK
(K, ).(K
1
, K
2
)
=
_
D
12
+c
_
jI
n
/
_
1
B
2
_
_
K
2
C
2
_
jI
n
/
_
1
B
2
K
1
+K
1
C
2
_
jI
n
/
_
1
B
2
K
2
_
_
D
21
+C
2
_
jI
n
/
_
1
B
_
T
t
= T
t
(K, ) T
tt
= T
tt
(K, )
= j c
_
jI
n
/
_
2
B = 2 c
_
jI
n
/
_
3
B
T
tt
K
.K = T
tt
K
(K, ).K
= j
_
_
D
12
+c
_
jI
n
/
_
1
B
2
_
K
_
C
2
_
jI
n
/
_
2
B
_
+
_
c
_
jI
n
/
_
2
B
2
_
K
_
D
21
+C
2
_
jI
n
/
_
1
B
_
_
Introducing the following transfer matrices

F(K, ) =
_
jI
n
/(K)
_
1
G
12
(K, ) = D
12
+c(K)F(K, )B
2
G
21
(K, ) = D
21
+C
2
F(K, )B(K) G
22
(K, ) = C
2
F(K, )B
2
.
we nally get
T
t
K
.K = G
12
KG
21
T
tt
KK
.(K
1
, K
2
) = G
12
_
K
2
G
22
K
1
+K
1
G
22
K
2
_
G
21
T
t
= jcF
2
B
T
tt
= 2cF
3
B
T
tt
K
.K = j
_
G
12
KC
2
F
2
B +cF
2
B
2
KG
21
_
.
Remark. Notice that the transfer matrices F, T, G
12
, G
21
and G
22
can be reliably and
eciently evaluated at the active frequencies (K) with a Hessenberg method, as
explained in [82].
Altogether, we have proved the following
Proposition 6.4.1. Suppose
1
_
T(K,
i
)
H
T(K,
i
)
_
has multiplicity 1 for every
i

1
(K), . . . ,
p
(K). Then the frequency response (w
1
, w
2
) (z
1
, z
2
) of the plant
_
_
x
z
1
z
2
_
_ =
_
_
/(K) B(K) B
2
c(K) T(K) D
12
C
2
D
21
0
p
2
m
2
_
_
_
_
x
w
1
w
2
_
_,
can be used to compute all the jet elements of tangent program (6.9). Indeed, it suces
to partition in order to get T, G
12
, G
21
and G
22
:
T(K, ) = T
w
1
z
1
(K, j) G
12
(K, ) = T
w
2
z
1
(K, j)
G
21
(K, ) = T
w
1
z
2
(K, j) G
22
(K, ) = T
w
2
z
2
(K, j)
Similar formulae for the derivatives of the transfer function T are obtained for dynamic
controllers, using the substitutions (6.5). Proposition 6.4.1 has the practical consequence
that the jet information for (6.9) may be computed using the MATLAB control tool
environment. Our experience shows that this works eciently even for systems with
several hundreds of states.
6.4.3 Multiple eigenvalues
The working hypothesis of the previous section was that leading eigenvalues
1
_
T(K,
i
(K))
H
T(K,
i
(K)
_
had multiplicity 1 for all frequencies in the set
1
(K), . . . ,
p
(K) and for all K in a
neighborhood of

K. This hypothesis is motivated by our numerical experience, where we
have never encountered multiple eigenvalues. This is clearly in contrast with experience
in pure eigenvalue optimization problems, where multiplicity of limiting elements is > 1
as a rule, causing failure of those numerical methods which assume smoothness of the
maximum eigenvalue function.
Nonetheless, our approach is still functional if the hypothesis of single eigenvalues at
peak frequencies is dropped. What is needed is the weaker assumption that the eigenvalue
multiplicities r
i
at the limit point

K are known for all active frequencies
i
, i = 1, . . . , p.
More precisely,
1
_
T(

K,
i
)
H
T(

K,
i
_
has multiplicity r
i
, i = 1, . . . , p, and for K su-
ciently close to

K, we have a mechanism to reliably guess r
i
, based solely on the informa-
tion at the current iterate K.
This situation has been discussed by several authors, see e.g. [51, 108, 109, 105].
With r
i
known, SQP can still be applied as follows. Consider X S
m
1
where
1
(X) has
multiplicity r. We replace the maximum eigenvalue function
1
by the average of the rst
r eigenvalues
r
(X) =
1
r
r
=1
(X) .
This function is smooth in a neighborhood of the smooth manifold
/
r
= X S
m
1
:
1
(X) = =
r
(X) >
r+1
(X)
of Hermitian matrices X S
m
1
with eigenvalue multiplicity r, and
1
=

r
on /
r
. The
manifold has co-dimension d
r(r+1)
2
1 in S
m
1
, and in a neighborhood of

X may be
described by d equations h
1
(X) = 0, . . . , h
d
(X) = 0. The tangent space of /
r
can be
computed, see [128, 109], and an SQP approach may be derived, if hypotheses (i) and
(ii) are suitably modied. For pure eigenvalue optimization, this has been discussed in
[105]. The extension to the semi-innite case is clear under the niteness assumption
(iii). Introducing natural analogues of conditions (i), (ii), see for instance [67], we may
then approach minimization of the H
-norm along the same lines and obtain the nite

program
minimize t
subject to

r
i
_
T(K,
i
(K))
H
T(K,
i
(K))
_
t
T(K,
i
(K))
H
T(K,
i
(K)) /
r
i
i = 1, . . . , p
with decision variables (K, t). Its tangent quadratic program uses the explicit form of the
tangent space T(/
r
, X) given in [128].
The case of multiple eigenvalues therefore requires two separate estimation steps. As
before we have to select p primary and secondary peaks. Then we have to estimate the
eigenvalue multiplicity for each T(K,
i
)
H
T(K,
i
). There exist indeed practically useful
ways to estimate the limiting multiplicity r = r
i
of
1
for each frequency
i
. For instance,
with a small threshold > 0, a guess r of r could be obtained by letting
r
< max1, [
1
[,
1
r+1
max1, [
1
[.
We leave the details of this extension to the care of the reader.
6.4.4 Globalisation via trust-region

Independently of whether or not the eigenvalue multiplicity at the active and nearly active
frequency peaks is 1, the local program (6.8) remains dependent on the characteristics of
the local solution (

K,
t) of (6.7), because the number p of primary and secondary peaks

may vary. Using Helly type theorems, an upper bound for p may be derived, see e.g. [67,
Thm. 4.2]. For a controller K of order k we have p (m
1
+k)(p
1
+k) peaks, but this is
pessimistic as a rule. In order to stabilize our method, we will therefore use a trust region
strategy.
In order to control the step quality computed at a given primal-dual pair (K, t; ),
we use the following
1
-merit functions
1
and
1
, respectively associated with the local
program (6.8) and the tangent program (6.9);
1
(K, t; ) = t +
1
i=1
[f (K,
i
(K)) t]
+
1
(K, t; ) = t +
1
2
K
L
tt
KK
(K, t; )K+
1
i=1
[K
f (K,
i
(K)) t
+f (K,
i
(K)) t]
+
.
The agreement between the actual reduction and the predicted reduction is measured by
the ratio
=

1
(K, t; )
1
(K +K, t +t; )
1
(0; )
1
(K, t; )
. (6.10)
Then the trust-region radius is managed according to the following algorithm, based on
a basic trust-region algorithm from [45].
Fix 0 <
1

2
< 1, 0 < < 1.
1. Initialize. Find an initial closed-loop stabilizing controller K
0
. Fix t
0
=
f(K
0
), p
0
and
0
R
p
0
+
, and set counter k = 0.
2. Peak estimation. Given current K
k
and
k
R
p
k1
+
, estimate number
p
k
and positions of primary and secondary peaks
1
(K
k
), . . . ,
p
k
(K
k
).
3. Model inconsistency. If p
k
diers from p
k1
, or if last step taken was
Cauchy step, then modify old
k
R
p
k1
+
or create a new consistent
multiplier vector
k
R
p
k
+
.
4. Initialize Newton method. Put K
k,0
= K
k
,
k,0
=
k
.
Set
0
= 1 if k = 0 or if k 1 and K
k
= K
C
Otherwise, set
0
=
j+1
.
Set counter j = 0.
5. Newton step. Use current iterate K
k,j
, multiplier estimate
k,j
and trust
region radius
j
to solve (6.9) and generate Newton trial step K
k,j+1
, with
associated multiplier estimate
k,j+1
R
p
k
+
.
6. Decision. Compare predicted progress in the local quadratic model
1
to
progress in
1
using progress ratio
j+1
in (6.10). There are two cases:
Step accepted:
j+1

1
. If even
j+1

2
and |K
k,j
K
k,j+1
|
=
j
, then double radius
j+1
= 2
j
. Otherwise keep
j+1
=
j
.
Put K
N
= K
k,j+1
and p
N
f(K
k
) f(K
N
). Goto step 7.
Step refused:
j+1
<
1
. Put
j+1
=
j
/2, increase counter j, and
go back to step 5.
7. Cauchy step. Given the current iterate K
k
, compute a Cauchy step K
C
away from K
k
using the rst-order method described in [8]. Let p
C
f(K
k
) f(K
C
) 0 be rst-order progress. If p
N
p
C
let K
k+1
= K
N
,
and
k+1
=
k,j+1
, otherwise put K
k+1
= K
C
.
8. Stopping test. If accepted step K
k+1
K
N
, K
C
oers no progress
over K
k
, stop. Otherwise increase counter k and go back to step 2.
In the inner loop j, the penalty parameter
j
for the expressions
1
,
1
is initialized at
j = 0 with
0
= (|
1
|
+ )
1
and updated according to the following rule for j 1,
with a chosen constant > 0
j
=
_

j1
if
1
j1
|
j+1
|
+,
(|
j+1
|
+ 2)
1
otherwise,
In this way,
j
< |
j+1
|
1
and the
1
merit function
1
is exact (see [103]).
Notice that the trust region procedure in the inner loop j between steps 4 and 6 follows
standard lines, but is based on the guess p and
1
(K), . . . ,
p
(K) of model (6.8), so a
few comments are in order here. Namely, since the model may be incorrect, the following
phenomenon may be observed. The Newton step may be successful with regard to (6.8),
i.e., with regard to the inner loop, but may nevertheless fail when matched with reality
in step 7. This is when the rst order step K
C
takes over. In the worst case, our method
therefore converges with the speed of the underlying rst-order technique [8]. Alternative
rst-order methods which could be used instead are for instance [117, 97, 10, 9].
A second phenomenon, which also arises due to the necessity to guess p, is addressed
in step 3. It may happen that the new K
k+1
K
N
, K
C
is no longer consistent with
the old model used in the previous step, because the number p had to undergo a change,
or because a rst-order step K
C
had to be taken. The multiplier estimate
k+1
=
k,j+1
from the last instance of step 6 is then of little use. We then restart multipliers afresh, or
we recycle the old ones.
Finally, when p and
i
(K) have been estimated correctly, the quadratic model will
ultimately produce steps with quadratic progress. This means the test in step 8 will
ultimately accept the Newton step, showing that our method has a fair chance to give
local quadratic convergence.
Notice that two kinds of stopping tests are needed. We have to apply a standard second
order test in order to halt the inner loop j if too small a trust region radius (indicating
failure) arises. But even when the inner loop j is stopped with a successful step K
N
, the
rst-order step K
C
may override this decision in cases where the model was incorrect. It
may even happen that the Newton model alerts a local minimum (K
N
= K
k
), but K
C
allows to decrease the function value further. These cases are covered by the test in step
8.
6.5 Technical aspects
6.5.1 Identifying peak frequencies
It is important to observe that our approach diers substantially from standard semi-
innite programming methods in so far as we have a highly ecient procedure to evaluate
the H
norm of the transfer channel at a xed closed-loop stabilizing controller K

R
(m
2
+k)(p
2
+k)
.
Computing peak frequencies can be based on a classical algorithm for estimating the
L
norm of a transfer matrix G(s) = D + C(sI A)

1
B explained in detail in [28];
see also the variations in [28, 31, 54]. This algorithm detects in the rst place the peak
frequencies (K), but may also be used to estimate secondary peaks. Its basic version is
the following:
Computing |G|
1. Initialization. Find initial such that (D) |G|
.
Fix tolerance > 0.
2. Step. For current , nd the frequency intervals I
1
, . . . , I
where (G(j)) > . If none is found, return the current .

3. Update. Otherwise let
k
be the midpoint of I
k
. Update
+
= (1 + ) max
k=1,...,
G(j
k
). Go back to step 2.
The returned value satises |G|
< (1 + ) on exit, which guarantees a relative

tolerance for |G|
as well as for the positions of the peak frequencies. Secondary peaks

may be estimated in several ways.
6.5. Technical aspects 141
A rst idea is to detected them on the way. Each interval I
k
occurring in step 2
above contains one or several primary or secondary peaks. If an interval I
k
at level
does not have a successor at level
+
, we conclude that a secondary peak
i
I
i
with
f(K,
i
) [,
+
) has been left behind. Fixing a threshold (0, 1), we keep those
secondary peaks
i
with value f(K,
i
) > |G|
. Since the values and slopes of the

curve f(K, ) at the endpoints of the I
k
are known, and since the peak value is in
the range [,
+
), a polynomial estimation of the
i
is usually quite accurate.
A second more direct way to estimate secondary peaks is the following: As soon as
the above algorithm is stopped at the nal , the imaginary eigenvalues of H() indicate
primary peaks. After xing the threshold , we keep those complex eigenvalues of
H() which lie in the strip Re < , and choose = Re as our set of secondary
peaks. This method is elegant as it also allows to estimate p very reliably.
It is clear that the above methods can be combined, and more sophisticated ways to
estimate secondary peaks can be imagined. However, it should be maintained that know-
ing
i
(K) to a high precision if f(K,
i
(K)) < f(K) is not necessary, as the constraint is
currently inactive. The closer the secondary peaks come to being active, the higher the
precision to which they are computed.
Detecting intervals I
k
where (G(j)) > is based on the Hamiltonian matrix
H() =
_
A 0
0 A
_
0 B
C
0
__
I D
D
I
_
1
_
C 0
0 B
_
.
We use the fact that
Lemma 6.5.1. is a singular value of G(j) if and only if j is an eigenvalue of H().
Remark. Our approach is robust with respect to the estimation of the cardinality p of
the unknown set (

K). If we overestimate p, we create a program (6.8), where some
constraints remain inactive near

K, a situation which is automatically dealt with by the
SQP solver. What is needed to justify this situation theoretically is an implicit function
for any of these additional secondary peaks. That may be assured by extending conditions
(i) and (ii) to secondary peaks.
6.5.2 Stopping criteria
We implemented two stopping tests in order to check convergence. The rst of them
checks criticality of the iterate (K
k
, t
k
) with multipliers
k
, through the absolute test
|L
t
(K,t)
(K
k
, t
k
;
k
)| <
1
,
where | | is a norm on R
m
2
p
2
+1
.
The second stopping condition checks the relative step length on the decision variables
|(K
k
, t
k
) (K
k1
, t
k1
)| = |(K, t)| <
2
(1 +|(K
k1
, t
k1
)|)
As some SQP iterates may become infeasible for problem (6.8), we also check if
f(K,
i
(K)) t < for all i = 1, . . . , p. The algorithm stops as soon as the current
iterate is feasible and one of the stopping tests is satised.
6.6 Numerical results

The results presented here were obtained with the algorithm described in section 6.4.4.
The plants are taken from [84], and the acronyms refer to that reference and allow to
retrieve the data from the literature.
At every iteration, Newton and Cauchy steps are generated, and the one giving best
progress is performed. In order to initialize the method, a closed-loop stabilizing controller
is computed in the initial phase (step 1) of the algorithm. This rst stage uses rst-order
nonsmooth techniques for minimizing the closed-loop spectral abscissa, as described in
[25]. Alternatively, derivative-free methods have been successfully used for stabilization
(see [6]) if the controller size is not too large. If no prior information on K is available,
we start this preliminary stabilization phase at K = 0. Then the initial multipliers for
SQP are set to
i
0
=
1
p
for i = 1, . . . , p.
Test example AC7 (transport airplane from [84]) has been given special attention,
because it is possible to visualize iterates in controller space (Figure 6.2). This exam-
ple highlights the benets of the second order algorithm (denoted by NS2) in terms
of convergence speed, compared to the nonsmooth rst-order method from [8] for H
synthesis (denoted by NS1). As required in step 1, a closed-loop stabilizing static con-

troller K
0
= [4.5931 1.2164] is rst computed. The initial closed-loop H
performance
is |T
wz
(K
0
)|
= 1.4614, which after optimization is reduced to the optimal value

|T
wz
(

K)|
= 6.5091 10
2
.
Table 6.1 shows that the gain of the second-order algorithm NS2 over the existing method
NS1 is signicant in terms of the number of iterations and CPU.
Alg. iterations CPU H
optimality nal
Cauchy Newton time (s) norm criteria [[
NS1 434 0 95.42 6.5098 10
2
9.7008 10
6
2
NS2 2 11 12.23 6.5091 10
2
9.4168 10
6
2
Table 6.1: H
synthesis: transport airplane (AC7).

Notice that the optimality criteria for the rst and second order methods are dierent:
a nonpositive optimality function for NS1 (see [8] for details), and the norm of the
Lagrangian gradient for NS2 (as dened in section 6.5.2).
The resulting controllers are respectively K
NS1
= [2.0331 1.8446 10
3
] and K
NS2
=
[2.0330 1.9655 10
3
]. Both K
NS1
and K
NS2
have two active frequencies. In the case
of K
NS2
these are
1
= 1.9067 and
2
= 1.3057 10
1
, in rad.s
1
, with almost linearly
dependent gradients
f
t
K
(K
NS2
,
1
) =
_
3.2855 10
3
4.6984 10
3
_
,
f
t
K
(K
NS2
,
2
) =
_
3.9785 10
3
5.6738 10
3
_
.
Figure 6.1 quanties the convergence rates. Our second-order algorithm (NS2) starts
with two Cauchy steps, then SQP takes over and converges superlinearly within only
11 Newton iterations. In contrast, with Cauchy steps only (NS1), a slow rst-order
6.6. Numerical results 143
convergence occurs. The iterates and contour lines of the H
norm are drawn in Figure

6.2.
50 100 150 200 250 300 350 400
10
1
10
0
10
1
10
2
10
3
10
4
10
5
iteration
AC7 : convergence rate estimation for NS1

||K
k+1
K
NS1
|| / ||K
k
K
NS1
||
||K
k+1
K
NS1
|| / ||K
k
K
NS1
||
2
2 4 6 8 10 12
10
3
10
2
10
1
10
0
10
1
iteration
AC7 : convergence rate estimation for NS2

||K
k+1
K
NS2
|| / ||K
k
K
NS2
||
||K
k+1
K
NS2
|| / ||K
k
K
NS2
||
2
Figure 6.1: Convergence rates of NS1 and NS2 for AC7. NS1 shows linear convergence
with rate close to 1 (left). NS2 gives quadratic convergence (right).
AC7 : H
contour levels and NS1 iterates

K
1
1 2 3 4 5
0
0.2
0.4
0.6
0.8
1
1.2
AC7 : H
contour levels and NS2 iterates

K
1
K
2
1 2 3 4 5
0
0.2
0.4
0.6
0.8
1
1.2
K
0
K
NS1
K
0
K
NS2
Figure 6.2: Iterates in controller space for NS1 (left) and NS2 (right). NS2 is signi-
cantly faster despite the change of peaks caused by the nonsmoothness.
Table 6.2 shows various other tests from [84], illustrating that the second-order method
can speed up convergence by decreasing the number of iterations. The extra work needed
for a second-order iteration, that is, dening the tangent quadratic program (6.9), locat-
ing peak frequencies, and computing rst and second derivatives to solve it, is usually
compensated by the reduction of the number of steps. For plants where the rst-order
method gives little progress and iterates many times (like AC7 or HE2), the gain is dra-
matic. Other examples like REA3 show less signicant gain or no gain at all, (AC6, AC8)
in cases where no superlinear convergence occurs. Nonetheless, attempting second order
steps is generally benecial, because the nal precision is improved (AC10).
plant n m
2
p
2
alg. iter CPU (s) H
[(K)[
AC6 7 2 4 NS1 79 30.28 4.1140 2
NS2 25 41.45 4.1140 2
AC7 7 1 2 NS1 434 95.42 6.5098 10
2
2
NS2 13 12.23 6.5091 10
2
2
AC8 9 1 5 NS1 14 23.50 2.0050 4
NS2 9 37.20 2.0050 4
AC10 55 2 2 NS1 74 206.16 1.3244 10
1
2
NS2 35 187.47 1.3236 10
1
5
AC10
55 2 2 NS1 82 762.65 1.3251 10

1
1
NS2 71 877.18 1.3237 10
1
5
HE2 4 2 2 NS1 746 196.20 4.2495 2
NS2 12 14.67 4.2492 2
REA3 12 1 3 NS1 23 14.18 7.4251 10
1
1
NS2 9 11.06 7.4251 10
1
1
Table 6.2: H
synthesis: comparison of NS1 and NS2 on various plants. AC stands for

aircraft models, HE for a helicopter model, and REA for a chemical reactor model.
An interesting case highlighting the importance of the nal precision is the bench-
mark study Boeing 767 under utter condition, which we have tested with two dierent
scenarios, labeled AC10 and AC10
. In a rst test, AC10, model (6.3) is used. Here

the rst-order method NS1 converges after 74 iterates. The second order method NS2
Alg. iterations CPU H
optimality nal
Cauchy Newton time (s) norm criteria dist
NS1 74 0 206.16 1.3244 10
1
5.4287 10
2
3.4000 10
5
NS2 28 7 187.47 1.3236 10
1
1.1926 10
6
3.3987 10
5
Table 6.3: Comparison of NS1 and NS2 for large scale example Boeing 747 under utter
condition (AC10). Column dist shows nal distance to instability.
needs 35 steps and converges quadratically. Despite the higher cost for setting up and
computing second order steps, this leads to a better CPU. What is more important in
this example, however, is that the second order stopping test is much more stringent
(second column from the right in Table 6.3) than for the rst order method. This is
particularly signicant for the Boeing 767 example, because iterates of both NS1 and
NS2 get extremely close to the hidden boundary of the stability region. The rightmost
column of Table 6.3 gives the distance to instability of the two solutions, and iterates on
the way have approximately the same distance. For NS1 it is therefore impossible to
decide whether the optimal solution is strictly inside or on the boundary of the stability
region, because the nal precision (the size of the steps before stopping) is comparable
in magnitude to the distance to instability. In contrast, for NS2, the higher precision
of the stopping test allows to conclude that the local minimum of NS2 is indeed inside
the stability region. In this example the NS1 solution has only two active peaks, while
6.7. Conclusion 145
NS2 nds 5 peaks, which gives further evidence that NS1 indeed fails to converge, (even
though the solution is still reasonable from a control point of view).
In this study the stability region turns out extremely narrow, which explains numerical
diculties with this example in the past. We have therefore performed a second test,
labeled AC10
, where a stabilizing channel as in (6.4) is added to the objective (6.3).

Not surprisingly, as the transfer channel is now much larger, this almost doubles CPU
times, even though the number of steps is practically the same. The controllers obtained
are fairly similar to those obtained for AC10, but with AC10
closed-loop stability of the

limit

K can now be certied. In the numerical testing the parameter in (6.4) was chosen
such that for the initial iterate K
0
, |T
wz
(K
0
)|
2
= 4|(sI /(K
0
))
1
|
2
, which gave
= 2.3150 10
4
.
We mention that adding the stabilizing channel does not resolve the intrinsic diculty
of this example, caused by the fact that the objective function in (6.3) or (6.4) is only
dened on a narrow region.
6.7 Conclusion
We have developed and tested a second-order method to compute static and reduced-
order H
-controllers and more general synthesis problems with structural constraints on

the controller. Our approach avoids the use of Lyapunov variables and therefore leads to
small or medium size optimization programs even for large systems (AC10 with 55 states).
We use a methodology from semi-innite programming to obtain a local nonlinear pro-
gramming model and apply a trust region SQP method. A rst-order nonsmooth spectral
bundle method is used to initialize the second-order algorithm in the neighbourhood of a
local solution. Our numerical testing indicates that speed of convergence and numerical
reliability may be improved by using the second-order method.
Conclusion
Les problmes de synthse par retour de sortie, avec un correcteur structur ou dordre
rduit, sont des problmes diciles de lAutomatique, pour lesquels les mthodes de rso-
lutions actuelles ne donnent pas toujours satisfaction, quand elles ne sont pas inoprantes.
Les travaux de cette thse se sont fonds sur un parti-pris original : formuler et rsoudre
ces problmes dans le contexte gnral de lOptimisation, en dpit de leurs mauvaises
proprits (ils sont non linaires bien videmment, non convexes, et surtout non lisses).
Les dicults sous-jacentes ont t leves en proposant deux approches distinctes.
La premire dentre elles, base sur les travaux de Polak, est une mthode de descente
non lisse, avec recherche linaire. Cette approche sest avre riche, exible et performante.
Riche, parce que nous avons pu la mettre en uvre pour des critres de synthse de
natures direntes (chapitres 3 5). Flexible, ensuite, parce que les contraintes portant
sur la structure ou lordre du correcteur peuvent tre directement et simplement prises en
compte. Performante, enn, comme cela est soulign dans la varit des exemples traits :
pour certains dentre eux (chapitres 4 et 5), nous avons amlior des rsultats obtenus par
les mthodes existantes, avec des temps de calcul beaucoup plus courts.
La seconde approche (chapitre 6) a t tudie pour la synthse H
. Elle sappuie sur la

rsolution dun sous-problme local, direntiable, par SQP avec rgion de conance. Elle
ore une convergence rapide, mais cette convergence nest que locale, ce qui ncessite de
recourir lapproche prcdente tant quun bassin dattraction nest pas atteint. Elle sest
avre assez dlicate mettre en uvre, plus sensible la taille des problmes, et fortement
dpendante de son initialisation. Cette seconde approche vient donc en complment de
la premire, dans les cas o apparat le phnomne de convergence lente, typique des
mthodes de descente du premier ordre.
Nous souhaitons enn souligner que la plupart des problmes de commande abords
dans cette thse ne sont pas convexes a priori. Si loptimisation non lisse permet dutiliser
une information de nature direntielle sur les fonctions minimiser, elle reste cepen-
dant tributaire de lexistence de plusieurs minima locaux. Ainsi, pour une initialisation
donne, chacune des approches proposes converge vers un minimum local ; une nouvelle
initialisation avec un autre correcteur peut conduire aprs optimisation un tout autre
minimum local, associ une valeur dirente du critre de synthse. Dans de nom-
breux cas, et lorsque le systme nest pas trop complexe, la connaissance ingnieur du
problme de commande rsoudre permet dinitialiser loptimisation avec des valeurs de
gains empiriques. Lorsque cela nest pas possible, il incombe lutilisateur de diversier
les initialisations an dexplorer lespace de recherche. Nos approches restent cependant
dterministes : un mme point initial donnera toujours le mme minimum local, ce qui
147
148 Conclusion
semble raisonnable pour des problmes de conception.
Les travaux mens au cours de cette thse se sont concentrs sur la minimisation de la
norme H
par des techniques non lisses. Nous avons montr que cette approche permet-
tait de rsoudre ecacement des problmes de commande considrs comme diciles, et
quelle pouvait tre applique dautres types de critres de synthse (stabilit, critre
temporel). Il ne sagit l, notre avis, que dun chantillon des problmes de commande
pour lesquels le recours loptimisation non lisse est pertinent. Nous avons dj fait rf-
rence au problme H
multidisque (voir [9]), dans lequel interviennent plusieurs canaux

de performance :
min
Kstabilisant
max
1iN
|T
w
i
z
i
(K)|
,
et au problme multibande (voir [10]), encore plus gnral :
min
Kstabilisant
max
1iN
max
I
i
[T
w
i
z
i
(K, j)] ,
o les I
i
sont des intervalles ou des runions dintervalles de frquences. Dautre part, la
norme H
permettant de quantier la robustesse du systme vis--vis de perturbations

ou dincertitudes, elle peut tre considre en tant que contrainte plutt que fonction
objectif. Cest le cas pour le problme de synthse mixte H
2
/H
, pour lequel la norme

H
2
, qui mesure la perfomance nominale du systme, est minimise sous une contrainte
dingalit portant sur la norme H
:
min
Kstabilisant
_
_
_T
w
H
2
z
H
2
(K)
_
_
_
2
t.q.
_
_
_T
w
H
z
H
(K)
_
_
_
.
La norme H
2
est parfaitement direntiable, cependant lapproche non lisse permet de
prendre en compte la contrainte (non lisse) en suivant la dmarche algorithmique de
Polak dcrite dans [114]. Nous renvoyons le lecteur intress par ces dveloppements aux
travaux de Apkarian, Noll et Rondepierre [12, 13]. Dautres formes de contraintes non
lisses sont envisageables, telles que le critre temporel prsent au chapitre 5, dont nous
avons dtaill le calcul des sous-gradients, et qui pourrait mme tre tendu sans dicult
aux systmes non linaires.
Enn, les problmes de synthse robuste qui se prtent au formalisme IQC (Contrainte
Quadratique Intgrale) peuvent tre rsolus en considrant une formulation non lisse dans
le domaine frquentiel, donc sans matrice de Lyapunov (voir [7, 119, 118]).
Bibliographie
[1] D. Alazard and P. Apkarian. Exact observer-based structures for arbitrary com-
pensators. Int. J. of Robust and Nonlinear Control, 9(2) :101118, 1999.
[2] D. Alazard, C. Cumer, P. Apkarian, M. Gauvrit, and G. Ferreres. Robustesse et
Commande Optimale. Cpadus, 1999.
[3] B. D. O. Anderson, N. K. Bose, and E. I. Jury. Output feedback stabilization
and related problems Solution via decision methods. IEEE Transactions on
Automatic Control, 20(1) :5366, 1975.
[4] B. D. O. Anderson and S. Vongpanitlerd. Network analysis and synthesis : a modern
systems theory approach. Prentice-Hall, 1973.
[5] P. Apkarian, V. Bompart, and D. Noll. Nonsmooth structured control design with
application to PID loop-shaping of a process. Int. J. of Robust and Nonlinear
Control, 17(14) :13201342, 2007.
[6] P. Apkarian and D. Noll. Controller design via nonsmooth multi-directional search.
SIAM Journal on Control and Optimization, 44(6) :19231949, 2006.
[7] P. Apkarian and D. Noll. IQC analysis and synthesis via nonsmooth optimization.
Systems and Control Letters, 55(12) :971981, 2006.
[8] P. Apkarian and D. Noll. Nonsmooth H
synthesis. IEEE Transactions on Auto-

matic Control, 51(1) :7186, 2006.
[9] P. Apkarian and D. Noll. Nonsmooth optimization for multidisk H
synthesis.
European Journal of Control, 12(3) :229244, 2006.
[10] P. Apkarian and D. Noll. Nonsmooth optimization for multiband frequency domain
control design. Automatica, 43(4) :724731, 2007.
[11] P. Apkarian, D. Noll, and O. Prot. Trust region spectral bundle method for non-
convex eigenvalue optimization. submitted, 2007.
[12] P. Apkarian, D. Noll, and A. Rondepierre. Mixed H
2
/H
control via nonsmooth

optimization. submitted, 2007.
[13] P. Apkarian, D. Noll, and A. Rondepierre. Nonsmooth optimization algorithm
for mixed H
2
/H
synthesis. In Proc. IEEE Conf. on Decision and Control, New

Orleans, LA, December 2007.
[14] P. Apkarian, D. Noll, J.-B. Thevenet, and H. D. Tuan. A spectral quadratic SDP
method with applications to xed-order H
2
and H
synthesis. European Journal

of Control, 10(6) :527538, 2004.
149
150 Bibliographie
[15] P. Apkarian, D. Noll, and H. D. Tuan. Fixed-order H
control design via a par-

tially augmented Lagrangian method. Int. J. of Robust and Nonlinear Control,
13(12) :11371148, 2003.
[16] P. Apkarian and H. D. Tuan. Robust control via concave minimization Local
and global algorithms. IEEE Transactions on Automatic Control, 45(2) :299305,
2000.
[17] V. Balakrishnan and S. Boyd. Global optimization in control system analysis and
design. In C. T. Leondes, editor, Control and Dynamic Systems : Advances in
Theory and Applications, volume 53. Academic Press, New York, 1992.
[18] V. Balakrishnan, S. Boyd, and S. Balemi. Branch and bound algorithm for compu-
ting the minimum stability degree of parameter-dependent linear systems. Int. J.
of Robust and Nonlinear Control, 1(4) :295317, 1991.
[19] J. Bao, J. F. Forbes, and P. J. McLellan. Robust multiloop PID controller de-
sign : A successive semidenite programming approach. Ind. & Eng. Chem. Res.,
38(9) :34073413, 1999.
[20] S. Basu, R. Pollack, and M.-F. Roy. On the combinatorial and algebraic complexity
of quantier elimination. In Proc. 35th Symp. on Foundations of Computer Science,
pages 623641, Santa Fe, NM, November 1994.
[21] M. Bergounioux. Mathmatique et automatique : de la boucle ouverte la boucle
ferme. J3eA, 2(HS1), 2003.
[22] D. S. Bernstein. Some open problems in matrix theory arising in linear systems and
control. Linear Algebra and its Applications, 162-164 :409432, 1992.
[23] V. Blondel, M. Gevers, and A. Lindquist. Survey on the state of systems and control.
European Journal of Control, 1(1) :523, 1995.
[24] V. Blondel and J. Tsitsiklis. NP-hardness of some linear control design problems.
SIAM Journal on Control and Optimization, 35(6) :21182127, 1997.
[25] V. Bompart, P. Apkarian, and D. Noll. Nonsmooth techniques for stabilizing linear
systems. In Proc. American Control Conf., New York, NY, July 2007.
[26] V. Bompart, D. Noll, and P. Apkarian. Second-order nonsmooth optimization for
H
synthesis. In Proc. 5th IFAC Symposium on Robust Control Design, Toulouse,

France, July 2006.
[27] V. Bompart, D. Noll, and P. Apkarian. Second-order nonsmooth optimization for
H
synthesis. Numerische Mathematik, 107(3) :433454, 2007.

[28] S. Boyd and V. Balakrishnan. A regularity result for the singular values of a trans-
fer matrix and a quadratically convergent algorithm for computing its L
-norm.
Systems and Control Letters, 15(1) :17, 1990.
[29] S. Boyd and C. Barratt. Linear Controller Design : Limits of Performance. Prentice-
Hall, 1991.
[30] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear Matrix Inequalities
in System and Control Theory, volume 15 of SIAM Studies in Applied Mathematics.
SIAM, Philadelphia, PA, 1994.
151
[31] N. A. Bruinsma and M. Steinbuch. A fast algorithm to compute the H
-norm of a
transfer function matrix. Systems and Control Letters, 14(5) :287293, 1990.
[32] J. V. Burke, D. Henrion, A. S. Lewis, and M. L. Overton. HIFOO A MATLAB
package for xed-order controller design and H
optimization. In Proc. 5th IFAC

Symposium on Robust Control Design, Toulouse, France, July 2006.
[33] J. V. Burke, A. S. Lewis, and M. L. Overton. Optimizing matrix stability. Proc.
American Mathematical Society, 129(6) :16351642, 2001.
[34] J. V. Burke, A. S. Lewis, and M. L. Overton. Approximating subdierentials by
random sampling of gradients. Mathematics of Operations Research, 27(3) :567584,
2002.
[35] J. V. Burke, A. S. Lewis, and M. L. Overton. Two numerical methods for optimizing
matrix stability. Linear Algebra and its Applications, 351-352 :117145, 2002.
[36] J. V. Burke, A. S. Lewis, and M. L. Overton. A nonsmooth, nonconvex optimization
approach to robust stabilization by static output feedback and low-order controllers.
In Proc. 4th IFAC Symposium on Robust Control Design, volume 5, pages 175181,
Milan, Italy, June 2003.
[37] J. V. Burke, A. S. Lewis, and M. L. Overton. Optimization and pseudospectra, with
applications to robust stability. SIAM Journal on Matrix Analysis and Applications,
25(1) :80104, 2003.
[38] J. V. Burke, A. S. Lewis, and M. L. Overton. Robust stability and a criss-cross
algorithm for pseudospectra. IMA Journal of Numerical Analysis, 23(3) :359375,
2003.
[39] J. V. Burke, A. S. Lewis, and M. L. Overton. A robust gradient sampling algo-
rithm for nonsmooth, nonconvex optimization. SIAM Journal on Optimization,
15(3) :751779, 2005.
[40] J. V. Burke and M. L. Overton. Dierential properties of the spectral abscissa
and the spectral radius for analytic matrix-valued mappings. Nonlinear Analysis :
Theory, Methods & Applications, 23(4) :467488, 1994.
[41] J. V. Burke and M. L. Overton. Variational analysis of non-Lipschitz spectral
functions. Mathematical Programming, 90(2) :317351, 2001.
[42] S. U. Cheang and W. J. Chen. Stabilizing control of an inverted pendulum system
based on H
loop shaping design procedure. In Proc. 3rd World Congress on

Intelligent Control and Automation, volume 5, pages 33853388, Hefei, China, June
2000.
[43] C. T. Chou and J. M. Maciejowski. System identication using balanced parame-
trizations. IEEE Transactions on Automatic Control, 42(7) :956974, 1997.
[44] F. H. Clarke. Optimization and Nonsmooth Analysis. Canadian Math. Soc. Series.
John Wiley & Sons, New York, NY, 1983.
[45] A. Conn, N. Gould, and Ph. Toint. Trust-Region Methods. MPS-SIAM Series on
Optimization, Philadelphia, PA, 2000.
152 Bibliographie
[46] E. J. Davison. Benchmark problems for control system design. Technical report,
International Federation of Automatic Control, 1990.
[47] J. C. Doyle, K. Glover, P. Khargonekar, and B. A. Francis. State-space solutions to
standard H
2
and H
control problems. IEEE Transactions on Automatic Control,

34(8) :831847, 1989.
[48] Y. Ebihara, K. Tokuyama, and T. Hagiwara. Structured controller synthesis using
LMI and alternating projection method. Int. J. Control, 77(12) :11371147, 2004.
[49] B. Fares, P. Apkarian, and D. Noll. An augmented Lagrangian method for a class of
LMI-constrained problems in robust control. Int. J. Control, 74(4) :348360, 2001.
[50] B. Fares, D. Noll, and P. Apkarian. Robust control via sequential semidenite
programming. SIAM Journal on Control and Optimization, 40(6) :17911820, 2002.
[51] R. Fletcher. Semi-denite matrix constraints in optimization. SIAM Journal on
Control and Optimization, 23(4) :493513, 1985.
[52] M. Fu and Z.-Q. Luo. Computational complexity of a problem arising in xed order
output feedback design. Systems and Control Letters, 30(5) :209215, 1997.
[53] M. Fujita, K. Hatake, and F. Matsumura. Loop shaping based robust control of a
magnetic bearing. IEEE Control Systems Magazine, 13(4) :5765, 1993.
[54] P. Gahinet and P. Apkarian. Numerical computation of the L
-norm revisited. In
Proc. IEEE Conf. on Decision and Control, pages 22572258, Tucson, AZ, December
1992.
[55] P. Gahinet and P. Apkarian. A linear matrix inequality approach to H
control.
Int. J. of Robust and Nonlinear Control, 4(4) :421448, 1994.
[56] D. Gangsaas, K. R. Bruce, J. D. Blight, and U.-L. Ly. Application of modern
synthesis to aircraft control : Three case studies. IEEE Transactions on Automatic
Control, 31(11) :9951014, 1986.
[57] A. U. Genc. A state-space algorithm for designing H
loop shaping PID controllers.

Technical report, Cambridge University, Cambridge, UK, 2000.
[58] J. C. Geromel, C. C. de Souza, and R. E. Skelton. Static output feedback controllers :
stability and convexity. IEEE Transactions on Automatic Control, 43(1) :120125,
1998.
[59] M. Gevers. A decade of progress in iterative process control design : from theory to
practice. J. Process Control, 12(4) :519531, 2002.
[60] L. El Ghaoui and V. Balakrishnan. Synthesis of xed-structure controllers via
numerical optimization. In Proc. IEEE Conf. on Decision and Control, pages 2678
2683, Lake Buena Vista, FL, December 1994.
[61] G. Golub and C. Van Loan. Matrix Computations. Johns Hopkins University Press,
Baltimore and London, second edition, 1989.
[62] K. M. Grigoriadis and R. E. Skelton. Low-order control design for LMI problems
using alternating projection methods. Automatica, 32(8) :11171125, 1996.
153
[63] W. M. Haddad and D. S. Bernstein. Generalized Riccati equations for the full-
and reduced-order mixed-norm H
2
/H
standard problem. In Proc. IEEE Conf. on

Decision and Control, pages 397402, Tampa, FL, December 1989.
[64] J. Han and R. E. Skelton. An LMI optimization approach for structured linear
controllers. In Proc. IEEE Conf. on Decision and Control, pages 51435148, Maui,
HI, December 2003.
[65] A. Hassibi, J. How, and S. Boyd. A pathfollowing method for solving BMI problems
in control. In Proc. American Control Conf., pages 13851389, San Diego, CA, June
1999.
[66] J. W. Helton and O. Merino. Coordinate optimization for bi-convex matrix inequa-
litites. In Proc. IEEE Conf. on Decision and Control, pages 36093613, San Diego,
CA, December 1997.
[67] R. Hettich and K. Kortanek. Semi-innite programming : theory, methods, and
applications. SIAM Review, 35(3) :380429, 1993.
[68] J.-B. Hiriart-Urruty and C. Lemarchal. Convex Analysis and Minimization Algo-
rithms I. Springer-Verlag, 1993.
[69] J.-B. Hiriart-Urruty and C. Lemarchal. Convex Analysis and Minimization Algo-
rithms II. Springer-Verlag, 1993.
[70] H. Hjalmarsson. Ecient tuning of linear multivariable controllers using iterative
feedback tuning. Int. J. Adaptive Contr. and Sig. Process., 13(7) :553572, 1999.
[71] H. Hjalmarsson. Iterative feedback tuning An overview. Int. J. Adaptive Contr.
and Sig. Process., 16(5) :373395, 2002.
[72] H. Hjalmarsson, M. Gevers, S. Gunnarsson, and O. Lequin. Iterative feedback
tuning : theory and applications. IEEE Control Syst. Mag., 18(4) :2641, 1998.
[73] H. Hjalmarsson, S. Gunnarsson, and M. Gevers. A convergent iterative restricted
complexity control design scheme. In Proc. IEEE Conf. on Decision and Control,
pages 17351740, Orlando, FL, December 1994.
[74] T. Iwasaki. The dual iteration for xed-order control. IEEE Transactions on Auto-
matic Control, 44(4) :783788, 1999.
[75] T. Iwasaki and R. E. Skelton. Parametrization of all stabilizing controllers via
quadratic Lyapunov functions. Journal of Optimization Theory and Applications,
85(2) :291307, 1995.
[76] F. Jarre. An interior-point method for minimizing the maximum eigenvalue of
a linear combination of matrices. SIAM Journal on Control and Optimization,
31(5) :13601377, 1993.
[77] H. Th. Jongen, K. Meer, and E. Triesch. Optimization theory. Kluwer Academic
Publisher, 2004.
[78] L. C. Kammer, F. De Bruyne, and R. R. Bitmead. Iterative feedback tuning via
minimization of the absolute error. In Proc. IEEE Conf. on Decision and Control,
pages 46194624, Phoenix, AZ, December 1999.
154 Bibliographie
[79] T. Kato. Perturbation theory for linear operators. Springer-Verlag, New York,
second edition, 1980.
[80] K. C. Kiwiel. Methods of Descent for Nondierentiable Optimization, volume 1133
of Lecture Notes in Mathematics. Springer-Verlag, 1985.
[81] K. C. Kiwiel. Methods of descent for nondierentiable optimization. Mathematical
Programming, 46(1-3) :105122, 1990.
[82] A. J. Laub. Ecient multivariable frequency response computations. IEEE Tran-
sactions on Automatic Control, 26(2), 1981.
[83] F. Leibfritz. Computational design of stabilizing static output feedback controllers.
Technical Report 99-01, University of Trier, 1999.
[84] F. Leibfritz. COMPl
e
ib : COnstrained Matrix-optimization Problem library A
collection of test examples for nonlinear semidenite programs, control system de-
sign and related problems. Technical report, University of Trier, 2003.
[85] F. Leibfritz and W. Lipinski. Description of the benchmark examples in COMPl
e
ib
1.0. Technical report, University of Trier, 2003.
[86] F. Leibfritz and E. M. E. Mostafa. An interior-point constrained trust region method
for a special class of nonlinear semi-denite programming problems. SIAM Journal
on Optimization, 12(4) :10481074, 2002.
[87] F. Leibfritz and E. M. E. Mostafa. Trust region methods for solving the optimal
output feedback design problem. Int. J. Control, 76(5) :501519, 2003.
[88] O. Lequin, M. Gevers, M. Mossberg, E. Bosmans, and L. Triest. Iterative feed-
back tuning of PID parameters : comparison with classical tuning rules. Control
Engineering Practice, 11(9) :10231033, 2003.
[89] W. S. Levine and M. Athans. On the determination of the optimal constant output-
feedback gains for linear multivariable systems. IEEE Transactions on Automatic
Control, 15(1) :4448, 1970.
[90] X. P. Li and B. C. Chang. A parametrization approach to reduced-order H
control-
ler design. In Proc. IEEE Conf. on Decision and Control, pages 29092912, San
Antonio, TX, December 1993.
[91] A. G. J. MacFarlane and Y. S. Hung. Analytic properties of the singular value of
a rational matrix. Int. J. Control, 37(2) :221234, 1983.
[92] M. Mkel. Survey of bundle methods for nonsmooth optimization. Optimization
methods and software, 17(1) :129, 2002.
[93] P. M. Mkil and H. T. Toivonen. Computational methods for parametric LQ
minimization A survey. IEEE Transactions on Automatic Control, 32(8) :658
671, 1987.
[94] M. Mammadov and R. Orsi. H
synthesis via a nonsmooth, nonconvex optimization

approach. Pacic Journal of Optimization, 1(2) :405420, 2005.
[95] B. Mrtensson. The order of any stabilizing regulator is sucient a priori informa-
tion for adaptive stabilization. Systems and Control Letters, 6(2) :8791, 1985.
155
[96] The MathWorks. Optimization Toolbox Users Guide, march 2006. Version 3.
[97] D. Q. Mayne and E. Polak. Algorithms for the design of control systems sub-
ject to singular value inequalities. IEEE Transactions on Automatic Control, AC-
20(4) :546548, 1975.
[98] D. McFarlane and K. Glover. A loop shaping design procedure using H
synthesis.
IEEE Transactions on Automatic Control, 37(6) :759769, 1992.
[99] E. Mengi and M. L. Overton. Algorithms for the computation of the pseudospectral
radius and the numerical radius of a matrix. IMA Journal of Numerical Analysis,
25(4) :648669, 2005.
[100] M. Mesbahi. Fixed-order output feedback synthesis via semidenite programming.
http://www.aa.washington.edu/faculty/mesbahi/papers/ofp.pdf, 2004.
[101] S. Miyamoto and G. Vinnicombe. Robust decentralized control design based on
coprime factorization representations and LMI optimization. In Proc. SICE Annual
Conf., pages 10071012, Tokushima, Japan, July 1997.
[102] S. G. Nash and A. Sofer. Linear and nonlinear programming. McGraw-Hill, New
York, 1996.
[103] J. Nocedal and S. Wright. Numerical optimization. Springer-Verlag, New York,
1999.
[104] D. Noll and P. Apkarian. Spectral bundle methods for nonconvex maximum ei-
genvalue functions : rst-order methods. Mathematical Programming Series B,
104(2) :701727, 2005.
[105] D. Noll and P. Apkarian. Spectral bundle methods for nonconvex maximum ei-
genvalue functions : second-order methods. Mathematical Programming Series B,
104(2) :729747, 2005.
[106] D. Noll, M. Torki, and P. Apkarian. Partially augmented Lagrangian method for
matrix inequality constraints. SIAM Journal on Optimization, 15(1) :161184, 2004.
[107] R. J. Ober. Balanced parametrizations of classes of linear systems. SIAM Journal
on Control and Optimization, 29(6) :12511287, 1991.
[108] M. L. Overton. A quadratically convergent method for minimizing a sum of Eucli-
dean norms. Mathematical Programming, 27 :3463, 1983.
[109] M. L. Overton. On minimizing the maximum eigenvalue of a symmetric matrix.
SIAM Journal on Matrix Analysis and Applications, 9(2) :256268, 1988.
[110] M. L. Overton and R. S. Womersley. On minimizing the special radius of a nonsym-
metric matrix function : Optimality conditions and duality theory. SIAM Journal
on Matrix Analysis and Applications, 9(2) :473498, 1988.
[111] R. J. Patton. Fault-tolerant control : the 1997 situation. In Proc. IFAC Symp. on
Fault Detection, Supervision and Safety, pages 10331055, Hull, UK, August 1997.
[112] T. Penzl. Algorithms for model reduction of large dynamical systems. Linear Algebra
and its Applications, 415(2-3) :322343, 2006.
[113] E. Polak. On the mathematical foundations of nondierentiable optimization in
engineering design. SIAM Review, 29(1) :2189, 1987.
156 Bibliographie
[114] E. Polak. Optimization : Algorithms and Consistent Approximations. Applied Ma-
thematical Sciences. Springer-Verlag, 1997.
[115] E. Polak, D. Q. Mayne, and J. E. Higgins. On the extension of Newtons method
to semi-innite minimax problems. SIAM Journal on Control and Optimization,
30(2) :367389, 1992.
[116] E. Polak and S. Salcudean. On the design of linear multivariable feedback systems
via constrained nondierentiable optimization in H
spaces. IEEE Transactions

on Automatic Control, 34(3) :268276, 1989.
[117] E. Polak and Y. Wardi. A nondierentiable optimization algorithm for the design
of control systems subject to singular value inequalities over a frequency range.
Automatica, 18(3) :267283, 1982.
[118] O. Prot, P. Apkarian, and D. Noll. A nonsmooth IQC method for robust synthesis.
In Proc. IEEE Conf. on Decision and Control, New Orleans, LA, December 2007.
[119] O. Prot, P. Apkarian, and D. Noll. Nonsmooth methods for control design with
integral quadratic constraints. submitted, 2007.
[120] J. Raisch, L. Lang, and M. Groebel. Loop shaping controller design for a binary
distillation column. In Proc. Control91, volume 2, pages 12711276, Edinburgh,
UK, March 1991.
[121] R. T. Rockafellar. Convex Analysis. Princeton University Press, Princeton, NJ,
1970.
[122] M. Rotkowitz and S. Lall. A characterization of convex problems in decentralized
control. IEEE Transactions on Automatic Control, 51(2) :274286, 2006.
[123] M. Saeki. Fixed structure PID controller design for standard H
control problem.
Automatica, 42(1) :93100, 2006.
[124] M. G. Safonov, K. C. Goh, and J. H. Ly. Control system synthesis via bilinear
matrix inequalities. In Proc. American Control Conf., pages 4549, Baltimore, MD,
June-July 1994.
[125] M. G. Safonov and D. J. N. Limebeer. Simplifying the H
theory via loop shif-

ting. In Proc. IEEE Conf. on Decision and Control, pages 13991404, Austin, TX,
December 1988.
[126] C. Scherer. Structured nite-dimensional controller design by convex optimization.
Linear Algebra and its Applications, 351-352 :639669, 2002.
[127] H. Schramm and J. Zowe. A version of the bundle idea for minimizing a nonsmooth
function : conceptual idea, convergence analysis, numerical results. SIAM Journal
on Optimization, 2(1) :121152, 1992.
[128] A. Shapiro. First and second order analysis of nonlinear semidenite programs.
Mathematical Programming, 77(2) :301320, 1997.
[129] J. Stoustrup and V. D. Blondel. Fault tolerant control : A simultaneous stabilization
result. IEEE Transactions on Automatic Control, 49(2) :305310, 2001.
[130] J. Suykens, K. Vandewalle, and B. De Moor. Intelligence and cooperative search
by coupled local minimizers. International Journal of Bifurcation and Chaos,
11(8) :21332144, 2001.
157
[131] V. L. Syrmos, C. T. Abdallah, P. Dorato, and K. Grigoriadis. Static output feedback
A survey. Automatica, 33(2) :125137, 1997.
[132] W. Tan, T. Chen, and H. J. Marquez. Robust controller design and PID tuning for
multivariable processes. Asian Journal of Control, 4(4) :439451, 2002.
[133] J.-B. Thevenet. Techniques doptimisation avances pour la synthse de lois de
commande. PhD thesis, Universit Toulouse III - Paul Sabatier, Toulouse, France,
Dcembre 2005.
[134] J.-B. Thevenet, D. Noll, and P. Apkarian. Nonlinear spectral SDP method for
BMI-constrained problems : applications to control design. In J. Braz, H. Arajo,
A. Vieira, and B. Encarnao, editors, Informatics in Control, Automation and
Robotics I, pages 6172. Springer Verlag, 2006.
[135] O. Toker and H. zbay. On the NP-hardness of solving bilinear matrix inequali-
ties andsimultaneous stabilization with static output feedback. In Proc. American
Control Conf., pages 25252526, Seattle, WA, June 1995.
[136] V. Torczon. Multi-Directional Search : A Direct Search Algorithm for Parallel ma-
chines. PhD thesis, Department of Mathematical Sciences, Rice University, Houston,
Texas, May 1989.
[137] V. Torczon. On the convergence of the multidirectional search algorithm. SIAM
Journal on Optimization, 1(1) :123145, 1991.
[138] H. D. Tuan, P. Apkarian, and Y. Nakashima. A new Lagrangian dual global op-
timization algorithm for solving bilinear matrix inequalities. Int. J. of Robust and
Nonlinear Control, 10(7) :561578, 2000.
[139] T. Wong and P. Bigras. Structured controller design with evolutionary optimization.
In Proc. AIA2003, pages 163168, Benalmdena, Spain, September 2003.
[140] Q Xin, M. V. Salapaka, P. G. Voulgaris., and M. Khammash. Structured optimal
and robust control with multiple criteria : a convex solution. IEEE Transactions on
Automatic Control, 49(10) :16231640, 2004.
[141] K. Zhou. A new controller architecture for high performance, robust, and fault-
tolerant control. IEEE Transactions on Automatic Control, 46(10) :16131618, 2001.
[142] K. Zhou. A natural approach to high performance robust control : another look
at Youla parameterization. In Proc. SICE Annual Conf., pages 869874, Sapporo,
Japan, August 2004.
[143] K. Zhou, J. C. Doyle, and K. Glover. Robust and Optimal Control. Prentice-Hall,
1995.
Rsum
Les techniques modernes de synthse pour la commande en boucle ferme des systmes dynamiques
linaires reposent sur la rsolution dquations matricielles (Lyapunov, Riccati) ou de problmes dopti-
misation sous contraintes dingalits matricielles (LMI, BMI). La mise en uvre de ces techniques est
rendue dicile, voire impossible, par la prsence de variables auxiliaires (dites variables de Lyapunov),
dont le nombre crot avec le carr de lordre du systme commander. En outre, elles sont dicilement
exploitables lorsque la loi de commande recherche est statique ou dordre rduit, ou encore lorsquune
structure particulire est impose au correcteur, autant de contraintes pourtant frquentes dans les ap-
plications.
An de rsoudre ecacement de tels problmes, la dmarche originale adopte dans cette thse
consiste les formuler dans le cadre thorique et gnral de la programmation non linaire, non convexe,
et plus spciquement ici, non lisse. Les problmes de stabilisation, de synthse H
, ou de satisfaction
de contraintes temporelles se posent naturellement avec ce formalisme et sont tudis dans la thse. Une
premire approche exploitant la description exhaustive du sous-direntiel de Clarke est prsente, puis
mise en uvre pour chacun de ces trois problmes. Laccent est port sur son ecacit numrique et
sa exibilit, travers un large choix dapplications en synthse structure ou dordre x. Une seconde
approche, dveloppe spciquement pour la synthse H
, sappuie sur une reformulation locale lisse

du problme, rsolue par programmation quadratique successive (SQP). Sa convergence plus rapide est
illustre sur des problmes de commande par retour de sortie statique.
Mots-cls: Optimisation non lisse et non convexe, sous-direntiel de Clarke, optimisation semi-innie,
systmes dynamiques, commande robuste, synthse H
, commande structure, commande dordre rduit.

Abstract
Modern controller design methods for linear dynamical systems are based on solving matrix
equations (Lyapunov, Riccati) or on optimization programs with matrix inequality constraints (LMI,
BMI). The execution times of these techniques are often prohibitive, because of auxiliary variables (called
Lyapunov variables), whose number grows quadratically with the plant order. Furthermore, they lead to
analytical and computational diculties when the sought controller is static, reduced-order, or when it
is nely structured, although such constraints very often arise in a realistic setting.
In order to solve such challenging problems eciently, the approach adopted in this thesis consists in
formulating them in the theoretical and general framework of nonlinear, nonconvex, and more specically
here, nonsmooth programming. The problems of closed-loop stabilization, H
synthesis, or satisfaction of
time-domain constraints arise naturally in this formalism and are studied in this thesis. Firstly, we propose
a nonsmooth descent algorithm which exploits an exhaustive description of the Clarke subdierential. It
is applied to each of these three fundamental control problems. Its numerical eciency and versatility
are stressed through a vast array of applications in structured or xed-order controller synthesis. A
second approach, specically developed for H
synthesis, is based on a smooth local reformulation of the

problem and solved with the sequential quadratic programming (SQP) algorithm. Its faster convergence
is illustrated with static output feedback control examples.
Keywords: Nonconvex and nonsmooth optimization, Clarke subdierential, semi-innite optimization,
dynamical systems, robust control, H
synthesis, structured controller, reduced-order controller.

Bompart Vincent

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Bompart Vincent

Diunggah oleh

Hak Cipta:

Format Tersedia

Universite Paul Sabatier Toulouse III

Norme inni du vecteur x R

Norme inni (valeur de crte) du signal u

(X) Spectre, pseudo-spectre de la matrice X C

(X) Abscisse spectrale, pseudo-spectrale de la matrice X C

(x, d) Drive directionnelle gnralise de f en x R

), la base de lAutomatique moderne, a permis de

Fig. 1 Fonction max sur R (en trait gras) : f

nest pas drivable en

) et donnons un aperu des mthodes de rsolution existantes. Il pourra en

, travers un exemple de loop-shaping par un correcteur structur (chapitre 4). Lal-

y est dcrit. Il exploite le sous-direntiel de

g(t )u() d . (1.1)

t] peut tre calcule avec

B ,= 0 (resp. tout vecteur propre droite v de A est tel que Cv ,= 0).

B ,= 0 (resp. Cv ,= 0). Ainsi (A, B) est commandable (resp. (C, A)

< , la sortie y(t) vrie |y|

< . Dans le cas des systmes LTI,

. Dun point de vue numrique, des critres algbriques (Routh-Hurwitz no-

de G concide avec la norme de M

est donc la norme induite par la norme L

de G, qui concide avec la norme H

ds que G est stable.

dite a-shifte (a-shifted H

norm en anglais). Pour

(critre de Routh-Hurwitz), ou encore par le test de Linard-Chipart, qui

, partie ferme de C dnie pour 0 x et pour toute matrice

(X), plus grande partie relle des lments de

(X) < 0 (X) > . (1.61)

(avec > 0 x) est plus rgulire que : condition

est localement lipschit-

(voir [28]), sur des calculs rpts de

Etant donn un systme LTI sous forme standard (1.33), le problme H

est la norme induite sur lespace des matrices de trans-

cache un double maximum portant sur les valeurs singulires

< pour une performance > 0 xe.

(loop-shaping). Il est bien sr

dordre plein. Quelle que soit

, cest--dire tels que, tant donne une performance > 0,

(1.64) si et seulement si il existe

)| < , est entirement dcrit sous la forme dune LFT dans

< , permet de jouer

: il est ainsi possible, comme indiqu dans [90], de choisir

se prte bien une rsolution par quations de Riccati dans le

par des ingalits matricielles permet

sous-optimaux par des ingalits matricielles

ce que le thorme de Lyapunov

sous-optimaux par une ingalit matricielle en X o

se ramne un problme de programmation semi-

par correcteur dordre plein est accessible

que ne permettait pas

, qui se formule sans perte de gnralit dans ce cadre grce

soit non lisse et non convexe, les algorithmes

(x, d), et dnie par

(x, d) est nie, et

(x, d)[ L|d| . (2.18)

(x, d) hrite du caractre localement

(x, d) pour tout d R

(x, d) = max s, d) : s f(x) . (2.20)

(x, d) pour tout d R

(x, d) pour tout

(x, d) est elle-

(x, d) pour tout d R

est un minimum local de f, alors 0 f(x