10th IFAC Symposium on Nonlinear Control Systems

10th IFAC Symposium on Nonlinear Control Systems

August 23-25,
10th IFAC 2016. Monterey,
Symposium California,
on Nonlinear USA
Control Systems
August
10th 23-25,
IFAC 2016. Monterey,
Symposium California,
on Nonlinear USA
Control Systems
Available
August 23-25, 2016. Monterey, California, USA online at www.sciencedirect.com
August 23-25, 2016. Monterey, California, USA

ScienceDirect
IFAC-PapersOnLine 49-18 (2016) 499–504
Non
Non Lipschitz
Lipschitz triangular
triangular canonical
canonical form
form for
for
Non
Non Lipschitz
Lipschitz
uniformly triangular
triangular
observable canonical
canonical
controlled form
form
systemsfor
for
uniformly
uniformly observable
observable controlled
controlled systems
systems
uniformly observable
∗
controlled
∗
systems
∗∗
P.
P. Bernard ∗ L. Praly ∗ V. Andrieu ∗∗
P. Bernard
Bernard ∗∗ L. L. Praly
Praly ∗∗ V. V. Andrieu
Andrieu ∗∗
∗ P. Bernard L. Praly V. Andrieu ∗∗
∗ Centre Automatique et Systèmes, MINES ParisTech, PSL Research
∗ Centre Automatique et Systèmes, MINES ParisTech, PSL Research
University,
∗ Centre
University, France
Automatique
France (e-mails:
et Systèmes,
(e-mails: pauline.bernard@mines-paristech.fr,
MINES ParisTech, PSL Research
pauline.bernard@mines-paristech.fr,
Centre
University, Automatique
France et Systèmes,
(e-mails: MINES ParisTech, PSL Research
pauline.bernard@mines-paristech.fr,
University, France laurent.praly@mines-paristech.fr).
(e-mails: pauline.bernard@mines-paristech.fr,
laurent.praly@mines-paristech.fr).
∗∗
∗∗ Université Lyon 1, Villeurbanne,
laurent.praly@mines-paristech.fr). France – CNRS, UMR 5007,
∗∗ Université Lyon 1, Villeurbanne,
laurent.praly@mines-paristech.fr). France – CNRS, UMR 5007,
LAGEP,
∗∗ Université
LAGEP, France
Lyon
France 1, (e-mail:
Villeurbanne,
(e-mail: vincent.andrieu@gmail.com)
France –
vincent.andrieu@gmail.com) CNRS, UMR 5007,
Université
LAGEP, France (e-mail: vincent.andrieu@gmail.com) 5007,
Lyon 1, Villeurbanne, France – CNRS, UMR
LAGEP, France (e-mail: vincent.andrieu@gmail.com)
Abstract:
Abstract: We
We study study
studyand the problem
the differentially
problem of of designing
of designing
designing observers
observers for
for controlled
controlled systems
systems which
which are
are
uniformly
Abstract:
uniformly observable
We
observable the
and problem
differentially observable,
observable, but
observers
but with
withfor an
for order
controlled
an controlled
order largerlarger
larger than
systems
than the the
which
the system
systemare
Abstract:
state
uniformly dimension We :study
observable we the only
have
and problem an
differentially of designing
injection, and
observable, not observers
buta diffeomorphism.
with an order We systems
establish
than which
that are
they
system
state
uniformly
can
state bedimensionobservable
transformed
dimension :: we
we into
have
have only
andaonly an injection,
differentially
triangular
an injection,
canonical and form
observable,
and notbut
not aa but
diffeomorphism.
with with an possibly
diffeomorphism. order larger We
We non establish
than
locally
establish thethat they
system
Lipschitz
that they
can
state be
functions.
can transformed
dimensionSince
be transformed :thewe into
have
classicala triangular
only high
into a triangular an gain canonical
injection,
observerand isform
not no a but with
diffeomorphism.
longer possibly
sufficient,
possiblywe We non locally
establish
review and Lipschitz
that they
propose
functions.
can
other beobservers
functions. Since to
transformed
Since thedeal
the classical
intowith
classicala triangular
such
high gaincanonical
highsystems,
gain observer
canonical
such
observer is
as
is
form
form ano
but with
nocascade
longer
but with
longer sufficient,
of possiblywe
homogeneous
sufficient, wenon
nonreview
review
locally
locally
observers.andLipschitz
and propose
Lipschitz
propose
other
functions. observers
Since tothe deal with
classical such
high systems,
gain such
observer as
is
other observers to deal with such systems, such as a cascade of homogeneous observers. anocascade
longer of homogeneous
sufficient, we review observers.and propose
other
© 2016,observers to deal with
IFAC (International such systems,
Federation of Automaticsuch Control)
as a cascade Hosting of by
homogeneous
Elsevier Ltd. All observers.
rights reserved.
Keywords:
Keywords: uniform
uniform observability,
observability, differential
differential observability,
observability, canonical
canonical observable
observable form,
form,
finite-time
Keywords: observers,
finite-time observers,
observers, homogeneous
uniform observability,
homogeneous observers,
differential exact
observers,observability, differentiators
exact differentiators canonical observable form,
differentiators
Keywords:
finite-time uniform observability,
homogeneous differential
observers, observability,
exact canonical observable form,
finite-time observers, homogeneous observers, exact differentiators
1. INTRODUCTION tain, as targeted form,
1. INTRODUCTION
1. INTRODUCTION tain,
form
tain, as targeted
for
as targeted
which form, aaa so-called
aa high-gain
form,
so-called
so-called
observer
triangular
triangular
can
triangularbe
canonical
canonical
built. More
canonical
1. INTRODUCTION form
tain,
precisely,
form for
as
for which
targeted
as
which detailed
a high-gain
form,below,
high-gain a itobserver
so-called
is known
observer can
triangular
canthat be
be built.
this More
canonical
observer
built. More
1.1 Context
1.1 Context
Context precisely,
form for as
which detailed
a below,
high-gain it is known
observer canthat be this observer
built. More
1.1 can
precisely,
can be be built
be built
built when,
as detailed
when, below, together
together it iswithknown
with the
thethatuniform (in
this observer
uniform (in the the
the
1.1 Context precisely,
can as detailed
when, below,
together it iswithknown the that this observer
uniform (in
control)
can
control) observability
be built when, together
observability of
of the
the system
with the
system (see
(see (Gauthier
uniform
(Gauthier (in andand
the
A
A lot of
lotofof attention
ofnonlinear
attentionobservers.has
has beenbeen dedicated
been Although
dedicated ato to the
togeneral construc-
the construc-
construc- control)
Kupka, observability
2001, Definition of I.2.1.2)
the system or (see (Gauthier
Definition 2 below),and
tion
A lot attention has dedicated the theory control)
Kupka, observability
2001, Definition of I.2.1.2)
the system orthe (see (Gauthier
Definition 2 below),
below),and
tion
A lot ofof nonlinear
attention observers.
has been Although
dedicated ato general
the theory
construc- the
Kupka,
the transformation,
2001,
transformation, Definitionobtained
obtained I.2.1.2) from
from or the strong
Definition
strong differential
2
differential
has
tion been
of obtained
nonlinear for
observers.linear systems,
Although avery few
general general
theory Kupka,
observability
the 2001,(see
transformation, Definition
(Gauthier
obtained I.2.1.2)and
from ortheDefinition
Kupka, 2001,
strong 2 below),
Definition
differential
has
tion been
approaches
has been obtained
of nonlinear
exist
obtained for for
observers.linear Although
nonlinear systems,
systems. very
avery
In few general
general
particular, general
theorythe observability
the transformation, (see (Gauthier
(Gauthier
obtainedand and
from Kupka,
the strong2001, differential
Definition
approaches
has
theorybeen
approachesof existgain
obtained
high
exist for for
for for
(Khalil
linear
nonlinear
linearand
nonlinear
systems,
systems.
systems,
Praly
systems. In
very
(2013)
In
few
particular,
fewand
particular, generalthe observability
refer-
the I.2.4.2)
observability or (see
Definition 11 below),
(see (Gauthier and is aaKupka,
Kupka,
2001,
diffeomorphism. Definition
2001, Definition
theory of of high
high gain (Khalil and and Praly (2013)(2013) and refer-
refer- I.2.4.2) or Definition below), is diffeomorphism.
approaches
theory existgain for nonlinear
(Khalil systems.
Praly In particular,
and the I.2.4.2) or Definition 1 below), is a diffeomorphism.
ences
theory
ences therein)
of highand
therein) gain
and Luenberger
(Khalil and
Luenberger (Andrieu
Praly and
(Andrieu (2013)
and Praly
Praly refer- I.2.4.2)
and(2013);
(2013); In
In this or Definition
paper
this paper
paperand we study
we strong
1 below),
study the the
the caseis awhere
case where
diffeomorphism.
where we have
we have uniform
have uniform
uniform
ences
Andrieu therein)
(2014)) andobservers
Luenberger have (Andrieu
been and Praly for
developed (2013);au- Inobservability
this we study differential
case observability,
we but the
ences
Andrieu therein)
(2014)) andobservers
Luenberger have (Andrieu
been and Praly for
developed (2013);au- Inobservability
latter this
observability paper
with and
an we
andorderstrong
study
stronglarger differential
the case
differentialthan observability,
where
the we
systemhave
observability, but the
uniform
state
but the
di-
tonomous
Andrieu
tonomous non
(2014))
non linear
linear systems
observers
systems have but
but their
been
their extension
developed
extension to
for
to con-
au-
con- latter
observability with an
and order
strong larger
differentialthan the system
observability, state
but di-
the
Andrieu
trolled
tonomous (2014))
systems
non is observers
not
linear have
straightforward.
systems but been
their developed
extension for
to au-
con- mension,
latter with implying
an that
order the
larger transformation
than the systemis at most
state an
di-
trolled
tonomous systems
non is not
linear straightforward.
systems but their extension to con- mension,
latter
injective
mension, with implying
an
immersion,
implying that
order
thatand the
larger
thenot transformation
than
a the
diffeomorphism
transformation systemis
is at
at asmost
state
above.
most an
di-
an
trolled systems is not straightforward. injective
mension, immersion,
implying thatand thenot a diffeomorphism
diffeomorphism
transformation is atasasmost
above. an
trolled
For systemsanisobserver
designing not straightforward.
for a system, a preliminary step We
injective shall see that,
immersion, in this
and case,
not a the system dynamics above.can
For
is
For designing
often required.
designing an observer
an observer
It consists for in
for aa system,
system,
finding aaa preliminary
reversible
preliminary step injective
coor-
step We shall see that,
that,byin
immersion, in andthis case,
not the system form
a diffeomorphism dynamics
as above.can
is
Foroften required.
designing an It consists
observer for in
a finding
system, a a reversible
preliminary coor-
step Westill
still be
shall described
see
be described
described byinaaabe triangular
this case,
triangular thecanonical
system
canonical dynamics
form but with
but leadscan
with
dinate
is often transformation,
required. It allowing
consists in us to
finding rewrite
a the
reversible system
coor- Wefunctions
still shall
be see
which that, may
by this noncase,
triangular the
locally system
Lipschitz.
canonical dynamics
form This
but can
with
dinate
is
dynamics
dinate transformation,
oftentransformation,
required.
in a targeted allowing
It consists form
allowing inmoreus to
us to
finding rewrite
favorable
rewrite the
a reversible
for
the system
coor- still
writing
system functions bestudy which
described may
by abe be non
triangular locally Lipschitz.
canonical form This
but leads
with
dynamics
dinate in a
transformation,targeted form
allowing moreus favorable
to rewrite for
the writing
system us
functions
us to
to study
study whichobservers
may
observers able
able non to cope
locally
topropose
cope with with
with such
Lipschitz.
such an
anThisextreme
leads
extreme
and/or
dynamics analyzing
in a the
targeted observer.
form In
more presence
favorable of control,
for two
writing functions
uscontext to and,whichin may be
particular,
observers able nontoto locally
cope Lipschitz.
a newsuch This
observer
an leads
made
extreme
and/or
dynamics analyzing
in the
a targeted observer.
form on Inwhether
more presence
favorable of consider
control,
for writing two context and, observers
in particular,
particular, totopropose
propose
tracks
and/or
tracks
and/or
are
are possible
analyzing
possible
analyzing
depending
the observer.
depending
the observer.
In
on
In
presence
whether
presence
we
of
we
of
control,
consider
control,
the
two usof
the context
two
tocascade
of aaa cascade
study
and,
cascade of
in
of homogeneous
homogeneous
ableto cope with
observers.
observers. aa new
newsuchobserver
an extreme
observer made
made
input
tracks as
are a simple
possible time function,
depending on making
whether the
we system
consider time
the context
of and, of in particular,
homogeneous to propose
observers. a new observer made
input
tracks as
dependentareaa possible
simple
or as time function,
depending making
on(infinite the
whetherdimensional) system
we system
considertime time
the of a cascade of homogeneous observers.
input
rameter,
as
dependent
input
dependent or
as making
a or as aa
simple
simple
as
more
athetime
more
time
more
involved
function,
involved
function,
systeminvolveda family
making
(infinite
making
of
(infinite
the
dimensional)
the system
dynamical
dimensional)
pa-
pa-
time
system, pa-
rameter,
dependent
indexed
rameter, making
by orthe
making the
as acontrol.
more
the system
system aa family
involved family
Accordingly ofthe
(infinite
of dynamical
dimensional) system,
transformation
dynamical system, pa-
indexed
rameter, by
makingthe control.
control.
the system Accordingly
a family ofthe the transformation
dynamical system, 1.2 Definitions
1.2 Definitions
Definitions
mentioned
indexed
mentioned by above
the
above is
is simply
simply time-varying
Accordingly
time-varying or
or input
transformation
input depen-
depen- 1.2
indexed
dent.
mentioned by
Moreover, the
above control.
along
is the
simply Accordingly
later itself,
time-varying the
with or transformation
the
inputinput seen
depen- 1.2 Definitions
dent. Moreover,
mentioned abovealongalong
is simplythe later
later itself,
time-varying withordependence
the
inputinput seen
depen- Consider
as
as aa parameter,
dent.
dent.
Moreover,
parameter,
Moreover,
the
the
along
strength
the
strength
the later
of
of the
itself,
the
itself,
input
with
input
with
the input
dependence
the input
seen
seen
of Consider aaa controlled
of Consider
controlled
controlled system system
system of of the
of the
the formform
form :::
the
as a transformation
parameter,
theatransformation
transformation the may
strength
may vary. of the input dependence
vary. of the input dependence of Consider aẋcontrolledof system of the form :
as parameter, themay strength = ff (x) + g(x)u
the
the
For transformation
example, in may vary.
(Hammouri
vary.
and Morales (1990); Besançon ẋ = (x) +
ẋ = f (x) + g(x)u g(x)u ,,, y=
yy =
h(x)
= h(x)
h(x) (1)
(1)
(1)
For
For example,
example, in (Hammouri
inthe
(Hammouri and Morales
and Morales (1990); Besançon
(1990);arbitrarily
Besançon ẋ = f (x) + g(x)u , y = h(x) (1)
et
For
et al.
al. (1996)),
example,
(1996)), inthe transformation
(Hammouri
transformation and Morales can
can depend
(1990);arbitrarily
depend Besançon
on
et the
al. input
(1996)), with the objective
the transformation of
canobtaining
depend aarbitrarily
targeted where
where x
x isis the
is the state
theoutput
state in in
in R.R n
,, u
RnnGiven u is
is an input
an input
input in
in R
R
p
p and y
on
et thewhich
al. inputiswith
(1996)), with
the the objective
objective
transformation of
canobtaining
depend aaarbitrarily
targeted where p and y
form
on the input state-affine
the up to input/output
of obtaining injection,
targeted is a measured
x state R
R. , u is an input time
in R function
p and y
form
on
formthewhich
input
which is state-affine
with the
is state-affine up
objective to input/output
of obtaining
up to input/output a injection,
targeted
injection,
is
where
tt →
is a
a measured
x
u(t),
measuredis
we the output
denotestate
output X(x, in
in R
R.
t)
n
a Given
, u is
solution
Given an
an input
input
of (1)
input time
in
going
time R function
and y
through
function
or more
form
or morewhichgenerally as
is state-affine
generally as in
in Besançon (1999),
up to input/output
Besançon (1999), a
a targeted
targeted form
injection,
form is → u(t),
a measured we denote
output X(x, in R. t) aaGiven
solution an of (1)
input going through
time function
t
x →
 u(t), we denote X(x, t) solution of (1) going through
which
or morehas a triangular
generally structure.
as in Besançon The
(1999), dependence
a targeted form may tx →at at timewe0.
u(t), We are
denote X(x,interested
t) a solution in ofestimating
(1) going throughX(x, t)
which
or more
also
which be has
hason aathe
generally triangular
derivative
triangular structure.
as in Besançon of
structure. the The dependence
(1999),
inputs
The dependence
aastargeted
proposed may
form
may in x
knowing at time
timeyy and
0.
0. WeWe
u but
are
areonly interested
interested
as long
in
in
as
estimating
estimating
(X(x, t), u(t))
X(x,
X(x, is
t)
t)
in
also
which be
Gauthier
also be hason
on a
and the
the derivative
triangular
Kupka
derivative (2001) of
structure.
of the
with
the inputs
The
the
inputs so as
as proposed
dependence
called
proposed may
phase- in
in x
knowing at time 0.
and We
u butare onlyinterested
as long in
as estimating
(X(x, t), u(t)) X(x, is t)
in
Gauthier
also be onand the Kupka
derivative (2001) of with
the the
inputs so as called
proposed phase- in a
knowing
a given
given compact
y and
compact u set
but
set C
C ×
only
× U
U .
as
. Let
Letlong S
S be
as
be an
(X(x,
an open
t),
open subset
u(t))
subset is of
in
of
variable
Gauthier representation
and Kupka (2001) as targeted the so called phase- aRngiven
with form. knowing
n
containingy and
compact C.u but
set Cwill
We only
× use as
U . Let long
the S as (X(x,
following
be an open t),
two u(t))
notions
subset ofis in
variable representation
Gauthier representation
and Kupka (2001) as targeted
targeted the so called phase- aRngiven
form.
with form. containing C. We Cwill use the following two notions
variable as R
observability compact
containing C.
definedset in
We will× U . Let
Gauthier
use the S
and beKupka
following an open
two subset
(2001)
notions:: of of
variable representation
Alternatively,
Alternatively, we may
weinput.
may impose
as targeted
impose
impose the
the
form.
transformation
transformation not
not to to R n
observability
to observability containing defined
C. We in
will
defined in Gauthier Gauthier
use the and Kupka
following
and Kupka two(2001)
notions
(2001) : of
depend on
Alternatively, the we may This is the
the context
transformation of uniformly
not Definition
observability 1. (Differential
defined in observability).
Gauthier and Kupka (See (Gauthier
(2001) :
depend
observable
depend on systems.
Alternatively,
on theweinput.
the input.
may For This
impose
Thisexampleis the
is theGauthier
the context and
transformation
context of not to Definition
of uniformly
uniformly
Bornard Definition
and Kupka,
1. (Differential
1. 2001,
(Differential
Definition
observability).
observability).
I.2.4.2).) The
(See
(Seesystem
(Gauthier
(Gauthier
(1) is
observable
depend on systems.
the input. For Thisexampleis the Gauthier
context and
of Bornard
uniformly Definition
and Kupka, 1. 2001,
(Differential
Definition observability).
I.2.4.2).) The(Seesystem(Gauthier
(1) is
(1981);
observable
(1981); Gauthier
systems.
Gauthier et
et al.
For
al. (1992)
example
(1992) propose
Gauthier
propose this
this track
and
track to
Bornard
to ob-
ob- differentially
and Kupka,
differentially observable
2001,
observableDefinition of
of order
order m
I.2.4.2).)
m on
on S
S if
The
if the function
system
the (1)
function ::
is
(1981); Gauthier et al. (1992) propose this track to ob- differentially observable of order m on S if the function is
observable systems. For example Gauthier and Bornard and Kupka, 2001, Definition I.2.4.2).) The system (1) :
(1981); Gauthier et al. (1992) propose this track to ob- differentially observable of order m on S if the function :
Copyright
2405-8963 © 2016 IFAC 511 Hosting by Elsevier Ltd. All rights reserved.
Copyright © 2016, IFAC (International Federation of Automatic Control)
2016 IFAC 511
Copyright
Peer review©under
2016 responsibility
IFAC 511Control.
of International Federation of Automatic
Copyright © 2016 IFAC 511
10.1016/j.ifacol.2016.10.214
IFAC NOLCOS 2016
500 P. Bernard et al. / IFAC-PapersOnLine 49-18 (2016) 499–504
August 23-25, 2016. Monterey, California, USA

 
h(x) 2. IMMERSION CASE (m > n)
 Lf h(x) 
Hm (x) = 
 .. 
 The specificity of the triangular canonical form (2) is not
. so much in its structure but more in the dependence of
Lm−1
f h(x) its functions gi and ϕn . Indeed, for any k, Hk (x) satisfies
is injective on S. If it is also an immersion, the system is always :
called strongly differentially observable.  ˙ 
Definition 2. (Uniform observability). (See (Gauthier and Hk (x)
   
Kupka, 2001, Definition I.2.1.2).) The system (1) is uni- 0 1 0 ... 0 0
. .. .. .. .. ..
formly observable on S if for any pair (xa , xb ) in S 2 with  .. . . . .  . 
 .. .. ..   .. 
xa = xb , there is no C 1 function u : [0, T ) → U such that =. . . 
0 k H (x) +  .  + Lg Hk (x)u
  0 
0 ... ... 0 1
h(X(xa , t)) = h(X(xb , t)) 0 ... ... ... 0 L k
h(x)
f
for all t ≤ T such that (X(xa , s), X(xb , s)) ∈ S 2 for all
s ≤ t. To get (2), we need further the existence of a sufficiently
smooth function ϕk satisfying
In the case where m = n, i.e. Hn is a diffeomorphism, we Lkf h(x) = ϕk (Hk (x)) (4)
have :
Proposition 1. (See Gauthier and Bornard (1981); Gau- and of sufficiently smooth functions gi satisfying
thier et al. (1992)) If the system (1) is uniformly observable
and strongly differentially observable of order m = n, it Lg Li−1 i−1
f (x) = gi (h(x), . . . , Lf h(x)) (5)
can be transformed into the following triangular canonical
form : and this at least for all x in C if not in S.
ż1 = z2 + g1 (z1 ) u Let us illustrate via the following elementary example
.. what can occur.
.
żi = zi+1 + gi (z1 , . . . , zi ) u (2) Example 1. Consider the system
.. ẋ1 = x2 , ẋ2 = x33 , ẋ3 = 1 + u , y = x1
.
żn = ϕn (z) + gn (z) u , It is uniformly observable, differentially observable of order
3 and strongly differentially observable of order 5 since
where the functions gi are locally Lipschitz.
H3 (x) = (h(x), Lf h(x), L2f h(x)) = (x1 , x2 , x33 )
Such a triangular form named Lipschitz triangular form,
with Lipschitz nonlinearities is fortunately the nominal H5 (x) = (H3 (x), L3f h(x), L4f h(x)) = (H3 (x), 3x23 , 6x3 )
case for the high gain paradigm. where H3 is a bijection and H5 is an injective immersion
But as we shall see in Section 2, when the system is on R3 . For H3 , we have
strongly differentially observable of order m > n, trian- L3f h(x) = 3x23 = 3(L2f h(x))2/3
gularity is preserved but Lipschitzness is lost. Hence high
gain observers as those presented in Gauthier et al. (1992) Hence there is no Lipschitz function ϕ3 satisfying (4).
can no longer be used. Similarly, for H5 , we have
We thus present in Section 3 possible designs of observers Lg L2f h(x) = 3x23 = L3f h(x) = 3(L2f h(x))2/3
for the triangular canonical form (2) with non-Lipschitz gi . so there is no locally Lipschitz function g3 satisfying (5).
Everything is finally illustrated with an example in Section
4. Concerning the existence of continuous functions ϕk and
Notations. gi we have the following results given without proof due
to space limitations.
(1) We define the signed power function as Proposition 2. Suppose the system (1) is differentially
b
a = sign(a) |a|b observable of order m on an open set S containing the
given compact set C. There exists a continuous function
where b is a nonnegative real number. In the partic- ϕm : Rm → R satisfying (4) for all x in C. If the system
0
ular case where b = 0, a is actually any number in (1) is strongly differentially observable of order m on S,
the set  the function ϕm can be chosen Lipschitz on Rm .
{1} if a > 0 Proposition 3. Suppose the system (1) is uniformly ob-
S(a) = [−1, 1] if a = 0 servable on an open set S containing the given compact
{−1} if a < 0 set C, then, for all i, there exist continuous functions
0
Namely, writing c = a means c ∈ S(a). Note gi : Ri → R satisfying (5) for all x in C.
that the set valued map a → S(a) is upper semi- Note that the values of ϕm and gi are only imposed on the
continuous with closed and convex values. compact set Hm (C). In particular, their behavior when
(2) For x in Rp with p ≥ i, we denote |z| tends to infinity is free and can be chosen to satisfy
xi = (x1 , ..., xi ) . some extra constraints given by the observer design (see
and, for x in S, Assumption (7) in Section 3).
Note also that no assumption on Hm is needed for Propo-
Hi (x) = (h(x), . . . , Li−1
f h(x)) (3) sition 3 to hold. But it says nothing on the regularity of the

512
IFAC NOLCOS 2016
P. Bernard et al. / IFAC-PapersOnLine 49-18 (2016) 499–504
August 23-25, 2016. Monterey, California, USA 501

functions gi , besides continuity. As we saw in Example 1, j

even the usual assumption of strong differential observabil-
ity is not sufficient to make it Lipschitz everywhere. Thus, m−1
a triangular canonical form still exists but maybe without m
m−2 m−2
Lipschitzness of the functions it involves. Fortunately, as m m−1
stated in the next result, local Lipschitzness of gi may be .. ..
lost only at the image via Hi of points where the rank of αi ≥ . .
∂Hi 2 2
. . . 23
changes (z3 = x33 = 0 in Example 1). m m−1
∂x 1 1
Proposition 4. Assume there exists a neighborhood of x∗ m m−1 . . . . . . 12
∂Hi 0 0 ... ... ... 0
in S on which has a constant rank q. Then, if
∂x Table 1 : Hölder restrictions on the gi when d0 = −1
the system (1) is uniformly observable, there exists a
neighborhood N∗ of x∗ , a real number L∗ and a function 1−d (m−2)
0
gi defined on Ri , such that ẑ˙ 1 = ẑ2 + g1 (ẑ1 )u − k1 ẑ1 − z1  1−d0 (m−1)
1−d0 (m−3)
Lg Li−1f h(x) = gi (h(x), ..., Lf h(x))
i−1
∀x ∈ N∗ ẑ˙ 2 = ẑ3 + g2 (ẑ1 , ẑ2 )u − k2 ẑ1 − z1  1−d0 (m−1)
.. (6)
and, for all (za , zb ) in R × R ,m m
.
i
 0 1+d
|gi (za,1 , ..., za,i ) − gi (zb,1 , ..., zb,i )| ≤ L∗ |za,j − zb,j | . ẑ˙ m = ϕm (ẑ) + gm (ẑ)u − km ẑ1 − z1  1−d0 (m−1)
j=1
where the ki ’s are gains to be tuned and d0 is a parameter
This shows that, with uniform observability, we should to be chosen in [−1, 0). We refer to Notation (1) for the
get functions gi in (2) which are locally Lipschitz except case d0 = −1.
maybe close to the image of points where the rank of
the Jacobian of Hi changes. For a strongly differentially We have :
observable system of order m = n, this cannot happen and Proposition 5. Assume there exist d0 in [−1, 0) and c0 ≥ 0
we thus recover the result of Proposition 1. such that, for all (zi , ei ) in Ri × Ri , we have :
From Khalil and Praly (2013), it can be expected that a
standard high gain observer would not work for a canonical |gi (z1 + e1 , ..., zi + ei ) − gi (z1 , ..., zi )| (7)
 i
form (2) with non-Lipschitz gi . Hence we need to be able 1−d0 (m−i−1)

to design an observer for the canonical form (2) with ≤ c0 |ej | 1−d0 (m−j)
continuous but non-Lipschitz functions gi . j=1

3. OBSERVERS FOR A NON-LIPSCHITZ
TRIANGULAR FORM
All along this section it is assumed that the system (1)
is differentially observable of order m on an open set S
containing the given compact set C. It is also assumed that,
the system is uniformly observable on S. According to
Propositions 2 and 3, the image by Hm of the dynamics (1)
is contained 1 in the triangular form (2), with m replacing
n and with continuous gi and ϕm functions. The only
observer we are aware of able to cope with gi no more than
continuous is the one presented in Barbot et al. (1996).
Its dynamics are described by a differential inclusion (see
Appendix) :
ẑ˙ ∈ F (ẑ, y, u)
where (ẑ, y, u) → F (ẑ, y, u) is a set valued map. It can
be shown that any absolutely continuous solution gives in
finite time an estimate of z under the only assumption of
boundedness of the gi ’s. But the set valued map F above
does not satisfy the usual basic assumptions (upper semi-
continuous with closed and convex values) (see Filippov
(1988); Smirnov (2001)). It follows that we are not guar-
anteed of the existence of absolutely continuous solutions
nor of possible sequential compactness of such solutions
and therefore of possibilities of approximations of F . So
we do not pursue with this observer and look at other
possible ones.
3.1 Homogeneous observer
Homogeneous observers are extensions of high gain ob-
servers able to cope with some non Lipschitz functions. In
our context they take the form :
1 not equal since we have an immersion and not a diffeomorphism.
and ϕm verifies the same condition as gm . Then, there
exist (k1 , . . . , km ) such that, for all (x, ẑ, u) in Rn × Rm ×
loc (R+ ; U ) such that X(x, t) ∈ C for all t in R+ , any
L∞
absolutely continuous solution Ẑ(ẑ, x, t) of (6) is defined
on R+ and there exists T0 , depending on (ẑ, x), such that :
Ẑ(ẑ, x, t) = Hm (X(x, t)) , ∀t ≥ T0 .
For the case d0 ∈ (−1, 0), the proof of this proposition
follows from the arguments in Andrieu et al. (2008) (see
also Qian (2005)), more specifically from the proof of
(Andrieu et al., 2009, Corollary 7) and (Andrieu et al.,
2008, Theorem 5.1) and from the fact that the function
Hm is injective.
For the case, d0 = −1, we refer the reader to (Levant,
2001 b, Theorem 1). Although the arguments there are
not sufficient. Actually it is possible to prove this result
using a Lyapunov argument which we omit here due to
space limitations.
Assumption (7) imposes that, for each i, gi is Hölder with
order αi larger than the values given in Table 1 below for
the case d0 = −1, where i is the index of gi and j is the
index of ej .
3.2 Cascade of homogeneous observers
In the case where Assumption (7) is not satisfied, we can
still take advantage of the fact that the homogeneous ob-
server dot not ask for any restriction besides boundedness
on the last function gm when d0 = −1.
From the remark that such an observer
(1) can be used for the system

513
IFAC NOLCOS 2016
502 P. Bernard et al. / IFAC-PapersOnLine 49-18 (2016) 499–504
August 23-25, 2016. Monterey, California, USA

ż1 = z2 + r1 (t) gi satisfy (an appropriate version of) (7). For instance, in
.. the example below the first block has dimension 3, leading
. to an observer of dimension 12 and not 15.
żk−1 = zk + rk−1 (t) 4. EXAMPLE
żk = Φk (t)
provided the functions ri are known and the function Consider the system
Φk is unknown but bounded, with known bounds. ẋ1 = x2 , ẋ2 = −x1 + x33 , ẋ3 = −x2 + v , y = x1 (8)
(2) gives estimates of the zi ’s in finite time, with v as input. When v = 0, it admits 2x21 + 2x22 + x43 as
we see that it can be used as a preliminary step to deal positive definite and radially unbounded first integral. So
with the system we can expect the existence of solutions remaining in the
compact set
ż1 = z2 + r1 (t)  
C = x ∈ R3 : 2x21 + 2x22 + x43 ≤ c
..
. for instance when v is a small periodic time function,
except maybe for some particular pair of input v and initial
żk−1 = zk + rk−1 (t) condition (x1 , x2 , x3 ) for which resonance occurs. Also due
żk = zk+1 + gk (z1 , . . . , zk ) u to their periodic behavior, such solutions are likely to have
żk+1 = Φk+1 (t) + gk+1 (z1 , . . . , zk+1 ) u their x3 component recurrently crossing zero.
The system (8) is uniformly observable since, whatever v
Indeed, thanks to the above observer we know in finite time
the values of z1 , . . . , zk , so that the function gk (z1 , . . . , zk )u is, the knowledge of the function t → y(t) = X1 (x, t) and
therefore of its derivatives gives us right away x1 , x2 and
becomes a known signal rk (t). x3 . So it satisfies the first basic condition needed for an
From this, we can propose the following observer made of observer to be designed following the techniques described
a cascade of homogeneous observers : above. This is done in the following although we are aware
of the fact that, by exploiting the monotonicity of x3 → x33 ,
ẑ˙ 11 ∈ −k11 S(ẑ11 − y) if we know a compact set in which the solutions remain,
....................................................
1 then we can find k2 and k3 such that the following very
ẑ˙ 21 = ẑ22 + sat1 (g1 (ẑ11 ))u − k21 ẑ21 − y 2 simple reduced order observer is exponentially convergent
ẑ˙ 22 ∈ −k22 S(ẑ21 − y) ˙
ξˆ2 = −y + x̂33 − k2 x̂2 , x̂2 = ξˆ2 + k2 y
....................................................
˙ˆ
.. ξ = −(1 + k3 )x̂2 + u ,
3 x̂3 = ξˆ3 + k3 y
.
....................................................
m−1 4.1 Strong differential observability
ẑ˙ m1 = ẑm2 + sat1 (g1 (ẑ11 ))u − km1 ẑm1 − y m
.. After letting u = v − u0 , where u0 is at our disposal,
.   computations give successively
ẑ˙ m(m−1) = ẑmm + satm−1 gm (ẑm1 , . . . , ẑm(m−1) u  
1 x1
−km(m−1) ẑm1 − y m H3 (x) = x2
ẑ˙ mm ∈ −kmm S(ẑm1 − y) −x1 + x33
where the kij are positive real numbers to be tuned and  
where we have used the notations H3 (x)
sati (gi (z1 , ..., zi )) = min{Ḡi , max{−Ḡi , gi (z1 , ..., zi )}} . H5 (x) =  −x2 − 3(x2 − u0 )x23  .
(1 + 3x3 )(x1 − x33 ) + 6(x2 − u0 )2 x3
2
with the positive real numbers Ḡi given as
   H3 is a bijection on R3 but not a diffeomorphism even
 
gi h(x), . . . , Li−1
f h(x)  ≤ Ḡi ∀x ∈ C on C because of a singularity of its Jacobian at x3 = 0.
So the system is differentially observable of order 3 on
C but not strongly. According to Propositions 2 and 3, it
As a direct consequence of (Levant, 2001 b, Theorem 1) admits a triangular canonical form of dimension 3 but with
or Proposition 5, we have functions ϕ3 and {gi }i=1,3 maybe non Lipschitz.
Proposition 6. If, in (2), the functions gi are locally
bounded, then we can find positive real numbers (kij ) such We have the same result with the order 4 and H4 .
that, for all (x, u) in C × L∞
loc (R+ ; U ) such that X(x, t), For the order 5, H5 is injective on R3 and an immersion at
solution in C for all t in R+ and, for all ẑ in R 2 ,
m(m+1)
least on C when we select u0 large enough. So the system
is strongly differentially observable of order 5 on C and it
there exist absolutely continuous solutions Ẑ(ẑ, x, t) of the admits a triangular canonical form of dimension 5 with
observer above which are defined on R+ and for which ϕ5 locally Lipschitz but functions {gi }i=1,5 maybe non
there exists T0 such that, by denoting Ẑm the last block Lipschitz at points where the rank of Hi is smaller.
of m components of Ẑ, we have
4.2 An homogeneous observer of order 3
Ẑm (ẑ, x, t) = Hm (X(x, t)) ∀t ≥ T0 .
The triangular canonical form of dimension 3 mentioned
A drawback of the cascade of homogeneous observers is above is
that it gives an observer with dimension m(m+1)
2 in general. ż1 = z2
However, it is possible to reduce this dimension since, ż2 = z3
for each new block, one may increase the dimension by ż3 = ϕ3 (z) + g3 (z)u (9)
more than one, when the corresponding added functions y = z1 .

514
IFAC NOLCOS 2016
P. Bernard et al. / IFAC-PapersOnLine 49-18 (2016) 499–504
August 23-25, 2016. Monterey, California, USA 503

Fig. 1. Errors between the state x and its estimate given Fig. 2. Errors between the state x and its estimate given
by a high gain observer. by a homogeneous observer.
Note that in this case, the third-order exact differentiator
where proposed in Levant (2001 b), obtained by picking d0 = −1,
ϕ3 (z) = −z2 − 3(z2 − u0 )(z3 + z1 ) 3 ,
2
would also work.
2
g3 (z) = 3(z3 + z1 ) 3 . 4.3 A cascade of homogeneous observers
These functions are not Lipschitz at the points on the
hyperplane z3 = −z1 (image by H3 of points where x3 = 0) The function H3 is a bijection but its inverse H−1 3 is not
known to be visited possibly recurrently along solutions. locally Lipschitz around the points such that z3 = −z1
(i.e. H3 ({x3 = 0})). This may lead to too high sensitivity
As expected a high gain observer has problems dealing to disturbances on the estimate of x3 . Therefore, it may
with this system. Figure 1 shows the results of a simulation be preferable to implement the observer in dimension 5.
done with such an observer initialized at the origin, for Indeed, H5 is an injective immersion and has a locally Lip-
a system solution issued from x = (1, 1, 0) with v = schitz inverse. The triangular canonical form of dimension
5 sin(t/10). Each time the system solution crosses the 5 given by H5 is
singularity hyperplane x3 = 0, the error jumps likely
because the observer does not manage to compensate for ż1 = z2
the infinite Lipschitz constant of ϕ3 and g3 at this point. ż2 = z3
ż3 = z4 + g3 (z1 , z2 , z3 )u
ż4 = z5 + g4 (z1 , z2 , z3 , z4 )u
On another hand, because g3 and ϕ3 are Hölder of order ż5 = ϕ5 (z) + g5 (z)u
2 y = z1 .
3 , we can consider an homogeneous observer as discussed where the functions g3 , g4 , g5 and ϕ5 , not expressed here
in Proposition 5. It is
due to space limitations are Hölder of order 2/3, 1/3,
1−d
1, and 1 respectively. Hence Assumption (7) is satisfied
0
ẑ˙ 1 = ẑ2 − k1 ẑ1 − z1  1−2d0 with d0 = −1. So an homogeneous observer can be
1
used. However, we choose here to illustrate the cascaded
ẑ˙ 2 = ẑ3 − k2 ẑ1 − z1  1−2d0 observer proposed in Section 3.2. It takes the following
1+d 0 form :
ẑ˙ 3 = ψ(ẑ1 , ẑ2 , ẑ3 , u) − k3 ẑ1 − z1  1−2d0 2
ẑ˙ 11 = ẑ12 − k11 ẑ11 − y 3
1
with d0 to be chosen in (−1, 0) and where ψ = ϕ3 + g3 u ẑ˙ 12 = ẑ13 − k12 ẑ11 − y 3
is given in (9). It converges if the ki are sufficiently large ẑ˙ 13 ∈ −k13 S(ẑ11 − y)
and the function ψ verify inequality (7), i.e. : ..................................................
3
˙ẑ 21 = ẑ22 − k21 ẑ21 − y 4
|ψ(z1 + e1 , z2 + e2 , z3 + e3 , u) − ψ(z1 , z2 , z3 , u)| ≤ 1
L [|e1 |a1 + |e2 |a2 + |e3 |a3 ] , ẑ˙ 22 = ẑ23 − k22 ẑ21 − y 2
1

with ẑ˙ 23 = ẑ24 + sat3 (g3 (ẑ11 , ẑ12 , ẑ13 ))u − k23 ẑ21 − y 4
1 + d0 1 + d0 ẑ˙ 24 ∈ −k24 S(ẑ21 − y)
a1 ≥ , a2 ≥ , a3 ≥ 1 + d 0 . ..................................................
1 − 2d0 1 − d0 4
ẑ˙ 31 = ẑ32 − k31 ẑ31 − y 5
In our case, we have a1 = 23 = a3 and a2 = 1, so that the 3
ẑ˙ 32 = ẑ33 − k32 ẑ31 − y 5
condition (7) is satisfied with d0 = − 31 . The results of a 2

simulation with this d0 and k1 = 8, k2 = 35, k3 = 75 are ẑ˙ 33 = ẑ34 + sat3 (g3 (ẑ11 , ẑ12 , ẑ13 ))u − k33 ẑ31 − y 5
1
presented in Figure 2. ẑ˙ 34 = ẑ35 + sat4 (g4 (ẑ11 , ẑ12 , ẑ13 ))u − k34 ẑ31 − y 5
ẑ˙ 35 ∈ −k34 S(ẑ31 − y)

515
IFAC NOLCOS 2016
504 P. Bernard et al. / IFAC-PapersOnLine 49-18 (2016) 499–504
August 23-25, 2016. Monterey, California, USA
Fig. 3. Errors between the state x and its estimate given
by a cascaded homogeneous observer.
The results of a simulation with the gains
k11 = 14, k12 = 56, k13 = 110
k21 = 28, k22 = 277, k23 = 790, k24 = 1100
k31 = 50, k32 = 947, k33 = 6232, k34 = 11842, k35 = 11000
and a discretization time step of 10−6 are given in Figure
3 (see Levant (2003) for information on the gains settings)
5. CONCLUSION
It is known that observers can be designed for systems
which can be transformed into a triangular canonical
form, the functions of which are locally Lipschitz. We
have established that, when omitting the requirement of
local Lipschitzness, such systems include those which are
uniformly observable and differentially observable with
an order maybe larger than the system state dimension.
Hence, in this case, the usual assumption of strong dif-
ferential observability of order equal to the system state
dimension is too restrictive. But then we have to deal
with an observer design for a non-Lipschitz triangular
canonical form. Homogeneous observers give an answer to
this problem for some classes of Hölder functions. We have
proposed to combine them in a cascade way to go beyond
this Hölderness, imposing only continuity.
All this is only a preliminary work. Many topics remain to
be addressed : sensitivity to uncertainties in the dynamics,
sensitivity to measurement noise, gain tuning procedure.
APPENDIX : BARBOT ET AL’S OBSERVER
The set valued map proposed in Barbot et al. (1996) to
obtain an observer for a triangular canonical form where
the functions are only locally bounded is defined as follows.
Given (ẑ, y, u), (v1 , . . . , vm ) is in F (ẑ, y, u) if there exists
(z̃2 , ..., z̃m ) in Rm−1 such that:
v1 = z̃2 + g1 (y) u
z̃2 ∈ sat(ẑ2 ) − k1 S(y − ẑ1 )
..
.
vi = z̃i+1 + gi (y, z̃2 , . . . , z̃i ) u
516
z̃i+1 ∈ sat(ẑi+1 ) − ki S(ẑi − z̃i )
..
.
vm ∈ ϕm (y, z̃2 , . . . , z̃m )
+gm (y, z̃2 , . . . , z̃m ) u − km S(ẑm − z̃m )
REFERENCES
Andrieu, V. (2014). Convergence speed of nonlinear
Luenberger observers. SIAM Journal on Control and
Optimization, 52(5), 2831–2856.
Andrieu, V. and Praly, L. (2013). On the existence of
a Kazantzis–Kravaris / Luenberger observer. SIAM
Journal on Control and Optimization, 45(2), 432–456.
Andrieu, V., Praly, L., and Astolfi, A. (2008). Homo-
geneous approximation, recursive observer design, and
output feedback. SIAM Journal on Control and Opti-
mization, 47(4), 1814–1850.
Andrieu, V., Praly, L., and Astolfi, A. (2009). Homogene-
ity in the bi-limit as a tool for observer and feedback
design. IEEE Conference on Decision and Control.
Barbot, J., Boukhobza, T., and Djemai, M. (1996). Sliding
mode observer for triangular input form. IEEE Confer-
ence on Decision and Control, 2, 1489 – 1490.
Besançon, G. (1999). Further results on high gain ob-
servers for nonlinear suystems. IEEE Conference on
Decision and Control, 3, 2904–2909.
Besançon, G., Bornard, G., and Hammouri, H. (1996). Ob-
server synthesis for a class of nonlinear control systems.
European Journal of Control, 3(1), 176–193.
Filippov, A. (1988). Differential equations with discontinu-
ous right-hand sides. Mathematics and its Applications
Kluwer Academic Publishers Group.
Gauthier, J.P. and Bornard, G. (1981). Observability
for any u(t) of a class of nonlinear systems. IEEE
Transactions on Automatic Control, 26, 922 – 926.
Gauthier, J.P., Hammouri, H., and Othman, S. (1992).
A simple observer for nonlinear systems application to
bioreactors. IEEE Transactions on Automatic Control,
37(6), 875–880.
Gauthier, J.P. and Kupka, I. (2001). Deterministic obser-
vation theory and applications. Cambridge University
Press.
Hammouri, H. and Morales, J.D.L. (1990). Observer
synthesis for state-affine systems. IEEE Conference on
Decision and Control, 784–785.
Khalil, H.K. and Praly, L. (2013). High-gain observers in
nonlinear feedback control. Int. J. Robust. Nonlinear
Control, 24.
Levant, A. (2001 b). Higher-order sliding modes and
arbitrary-order exact robust differentiation. Proceedings
of the European Control Conference, 996–1001.
Levant, A. (2003). Higher-order sliding modes, differentia-
tion and output-feedback control. International Journal
of Control, 76(9-10), 924–941.
Qian, C. (2005). A homogeneous domination approach
for global output feedback stabilization of a class of
nonlinear systems. IEEE American Control Conference,
4708–4715.
Smirnov, G. (2001). Introduction to the theory of dif-
ferential inclusions, volume 41. Graduate studies in
Mathematics.