Generic Observability of Dynamical Systems
Esfandiar Nava Yazdani
Received: January 18, 2005
Revised: July 3, 2007
Communicated by Ulf Rehmann
Abstract.
Wedealwithaertainobservationmappingdened by meansofweightedmeasurmentsonadynamialsystemandgivene-essaryand suientonditions,under whih thismappingis
generi-allyaninjetiveimmersion.
2000 Mathematis Subjet Classiation: 93B07, 37C20, 37C10,
94A20.
Keywords and Phrases: Observability, generi, dynamial system,
sampling.
1
Introduction
Theobservabilityproblem ofnonlineardynamial systemshasbeenan
inter-esting subjetandativeeldof researhthroughout thelastdeades. Inthe
presentworkweonsidertimeinvariantsystemsoftheform
∗
˙
x
=
f
(
x
)
y
=
h
(
x
)
.The rst equation desribes a real dynami proess. Its state
x
(
t
)
at timet
is assumedto beelementofasmoothseond ountable (hene paraompat)
n-dimensional manifold
M
alled state spae. The dynamis of the systemis given by the vetor eld
f
onM
. The output funtionh
is a mappingfrom the state spae into the reals and stands for a measuring devie. The
seond equation desribes the output, whih ontains partial information
of the state. The output
y
is the only measurable quantity. There is avery broadvariety of systems, whih anbe desribed in this way. We all the triple
(
M, f, h
)
orsimply the pair(
f, h
)
a system. The system is alled∗We
C
r
i
M
,f
andh
areC
r
. Wedenotethesetof
C
r
vetoreldson
M
byX
r
(
M
)
.
In many appliations it is of essential importane to know the state of the
system at any time
t
. But measurement of the entire state (e.g. all itsoordinates)is oftenimpossibleorverydiult. Forinstane beauseofhigh
ostsortehnialreasons. Inmostasesonehasonlypartialinformationfrom
measurementof theoutput
y
, mathematially beingdesribed by theseondequation. Sotheissueistogetthestate(aswellasdistinguishdierentstates)
by using only output measurement
∗
. If this is possible, then the system is
said to be observable. Hene, the issue of the observability problem is the
following. Find riteria on the system, suh that by means of information
from theoutputtrajetory,itispossibleto distinguishdierentstatesaswell
as to reonstrut the states. There is no uniform or anonial denition of observabilityin theliterature. Theweakest andmostnaturaldenitionis the
following.
Definition 1.1
Thesystem(
f, h
)
issaidtobeobservable(ordistinguishable) ifforeah(
x, x
′
)
∈
M
×
M
with
x
6
=
x
′
thereexists atime
t
0
=
t
0
(
x, x
′
)
suh
that
h
(Φ
t0
(
x
))
6
=
h
(Φ
t0
(
x
′
))
.
However, the above denition is not well-suited for the treatment of the
observabilityproblem. Thereforeoneseekstoestablishastongernotionof
ob-servabilityasfollows. Consideramapping
Θ
(inthesequelalled observationmapping), whihmaps thestatespae intosomenite dimensional Eulidian
spae, assigning states to data derived from the output trajetorieson some
observation time interval
J
. Then deide the observability of the system bymeans of injetivity of
Θ
. Moreoverthefollowingnatural questionarises. Isobservabilityin thissense (,i.e. with respetto
Θ
)generi? Thelatteris themainissueofthepresentwork. Intheontroltheorythenoatationofobserver is generallystandingfor anothersystem havingthe output (andinput in the
ontrolledase)oftheoriginalsystemasinputandgeneratinganoutputwhih
isanasymptotiestimateoftheoriginalsystemstate.
Beside the task of reonstrution of the state, observabilityhas also
applia-tionin thetheoryofhaosand turbuleneinthefollowingsense. Suppose an
observablesystemhasaglobalattrator. Thenusing anobservationmapping
one anget a homeomorphi pitureof the attrator orat least information
about some of its harateristi properties. Examples an be found in [RT℄
and [T℄. We onsider a ertainobservation mappingintrodued in [KE℄ and
[E ℄. Wederiveneessaryandsuientonditionsforgeneriityofobservability
and loalobservability withrespet tothis mapping. Basially,there aretwo
otherwell-knwonapproahestotheobservabilitytask: samplingandhigh-gain approah. Inhislassialwork[T℄,Takensprovedgeneriityresultssimilarto
∗It
oursforthese approahes. AboutthesametimeAeyels[Ay℄ahievedsharper
results onerning the sampling approah. Further results on high-gain
ap-proahanbefoundin [GK℄,[GHK℄and[J℄. Partiularly,foraomrehensive
and intensive investigation of deterministi observation theory and
applia-tions inluding many deep resultsonerning high-gain approah we referto
[GK℄. Ourapproah hasthe disadvantagethat a suitablelinear lter hasto
beonstruted. Ontheothersideinontrasttohigh-gainapproahwedonot
needto restritourselvestosmoothsystems. Formoreonourapproahwith
appliationswereferto[N ℄.
2
Main Results
Definition 2.1
Let0
< τ <
∞
,(
A, b
)
be a stable and ontrollablem
-dimensional linearlter and(
M, f, h
)
aC
r
-system with ow
Φ
. Weall themapping
M
∋
x
7→
Θ
f,h
:=
R
0
−
τ
e
−
At
bh
(Φ
t
(
x
))
dt
∈
R
m
theobservationmap-pingandthenumber
m
observationdimension. IfΘ
f,h
isinjetiveweallthesystem
(
M, f, h
) Θ
-observableorsimplyobservable.Furthermoreweallthesystemloallyobservableif
Θ
f,h
isanimmersion.Theindistinguishablesubsetof
M
×
M
isgivenbyΩ
f,h
:=
{
(
x, x
′
)
∈
M
×
M
\
∆
M
: Θ
f,h
(
x
) = Θ
f,h
(
x
′
)
}
.The time interval
I
:= [
−
τ,
0]
has the physial interpretation of observation interval. We treat for simpliity of notation and in view of the physialinterpretation of the observation data ashistory of the output mapping, the
ase
I
⊂
R
≤
0
,0
∈
I
. Forunboundedintervalsmodiationsareneeded,whih wegiveexpliitlyfortheaseI
=
R
≤
0
.The key point in theproof of the following generiity resultsis to showthat
zeroisaregularvalueoftheobservationmapping. Provingthiswethenapply
transversality density and openness theorems to get loally our statements,
whihthenwillbeglobalizedtothewholestatespae. ABaireargumentthen
yieldsthenal results. Partiularlyfor
τ <
∞
in anyC
r
neighborhoodof the
systemthereisasystemwhihis bothobservableandloally observablewith respet to
Θ
. An appropriate statement is also valid forτ
=
∞
in theC
1
topology.
If thestatespaeis notompat,itis moresuitableto onsidertheso alled
strongorWhitney(
C
r
-)topologyon
C
r
(
M,
R
)
andX
r
(
M
)
. Thereasonisthat
inthistopologyonehasmoreontrolonthebehaviorofthefuntionsand
ve-tor eldsat innity. Note thatdensityin thistopologyisastrongerproperty than in the ompat-open(also alled weak)topology. For instane, roughly
speaking,asequeneofoutputfuntions
h
j
onvergesinWhitney topologytoh
i there exists a ompatsetK
suh thath
j
=
h
outside ofK
exeptforTheaseofsmoothvetoreldsissimilar. Thewellknownfatthat
C
r
(
M,
R
)
and
X
r
(
M
)
areBairespaesin theWhitney topology(proofsanbefoundin
in[H,2.4.4℄and[P ℄)isofbasiimportaneforourresults. Aresidualsubsetof
C
r
(
M,
R
)
orX
r
(
M
)
isdense. Fromnow onthe spaes of
C
r
funtionsaswell
asvetoreldson
M
areequippedwiththeirC
r
Whitneytopology,unless
oth-erwiseindiated. Ifthestatespae
M
isompatandr <
∞
,thenC
r
(
M,
R
k
)
and
X
r
(
M
)
endowedwiththeompat-opentopologyareBanahspaes(while
in general Fréhet spaes), their Whitney and ompat-open topology
oin-ideandtheow
Φ
ofeahvetoreldf
∈
X
r
(
M
)
isdenedgloballyon
M
×
R.
For some pairs in
X
r
(
M
)
×
C
r
(
M,
R
)
there is no possibility to
distin-guish or loally distinguish the states. For generiity results it is an
es-sential fat that the omplement of the set of suh pairs is residual. Let
Sing
(
f
) :=
{
x
∈
M
:
f
(
x
) = 0
x
∈
T
x
M
}
denote the set of singularities ofthevetoreld
f
(equilibria ofthe system),where0
x
is thezeroofT
x
M
. Inthe sequelwe shalloften omit thesubsript
x
and write simplyf
(
x
) = 0
forx
∈
Sing
(
f
)
.Reall that asingularity
x
0
∈
Sing
(
f
)
is alled simplei the prinipialpart of thelinearizationoff
atx
0
, i.e. thelinearmappingd
x0
f
:
T
x0
M
→
T
x0
M
, doesnothavezeroasaneigenvalue.We denote the set of
C
r
vetor elds, whose singularities are all simple, by
X
r
0
(
M
)
. ItiswellknownthatasimplesingularityisisolatedandX
r
0
(
M
)
isanopenanddensesubsetof
X
r
(
M
)
(foraproofwereferto[PD ,3.3℄).
Definition 2.2
We allx
0
∈
M
aΘ
-simple singularity i there exists a otangent vetorv
∈
T
∗
x0
M
suh that the linear system(
T
x0
M, d
x0
f, v
T
)
is
observable,i.e.,thelinearmapping
Z
0
−
τ
e
−
tA
bve
td
x
0f
dt
:
T
x0
M
→
R
m
isinjetive. Inthisasewesaythat
v
T
isa
Θ
-oyliovetorofd
x0
f
.Wedenoteby
X
r
1
(
M
)
thesetofC
r
vetoreldson
M
,whosesingularitiesareΘ
-simple. MoreoverwesetX
r
0
,
1
(
M
) :=
X
r
0
(
M
)
∩
X
r
1
(
M
)
.Inthelimitingase
τ
=
∞
denseorbitsaswellasnontrivialreurreneause diulties and speial onsiderationsare neessary. We investigate this aseundertheassumptionthat
M
isompat. Appropriateresultsinthe nonom-pat ase an be similarly deriven if we further restrit (in order to ahievewell-denedness of theobservation mapping) the systemsto be globally
Lip-shitzian. Furthermore,in order to ensuredierentiability ofthe observation
mapping,if
τ
=
∞
werestritthevetoreldstotheopensetX
r
(
M, a
) :=
{
f
∈
X
r
(
M
) :
sup
x
∈
M
k
d
j
x
f
k
< a
forallj
= 1
, ..., r
}
eld
f
byC
(
f
)
andtheunionofthenegativelimitsets byL
−
(
f
)
,wesetX
r
−
(
M
) :=
{
f
∈
X
r
(
M
) :
L
−
(
f
)
⊂
C
(
f
)
}
.WedenotethatsomeineterestinglassesofvetoreldslikeMorse-Smaleelds
∗
areontainedin
X
r
−
(
M
)
. Partiularly,sinethesetonsistingofMorse-Smalevetor elds is open and nonempty in
X
r
(
M
)
, the interior
int
(
X
r
−
(
M
))
ofX
r
−
(
M
)
isaBairespae(in theinduedtopology). Furthermorenotethat thelimitset ofagradienteld onsistsonlyofthe ritialpointsofthepotential
funtion. Therefore
X
r
−
(
M
)
also ontainstheset ofgradientelds. Moreoverweset
X
r
2
(
M
) :=
X
r
(
M, a
)
∩
int
(
X
r
−
(
M
))
.It is well known that on a ompat manifold
C
1
-generially the
nonwander-ing set of a smooth vetor eld oinides with the losure of the set of its
periodi points. This statement alled general density theorem is a
onse-quene of Pugh's losing lemma, whih ensures that a nonwandering point
an be made periodi by a small
C
1
-petrubation in a neighbourhood of the
point. Seealso [Pu ℄,[AR, 7.3.6℄ andthereferenesgiventhere
∗
. Partiularly
{
f
∈
X
1
(
M
) :
L
−
(
f
)
⊂
C
(
f
)
}
isaresidualsubsetofX
1
(
M
)
.
Wesetfor
y
∈
E
r
a,τ
:=
{
y
∈
C
r
([
−
τ,
0]
,
R
) :
R
0
−
τ
e
at
|
y
(
t
)
|
dt <
∞}
P
τ
y
:=
Z
0
−
τ
e
−
At
by
(
t
)
dt
andforxed
τ
simplywriteP
insteadofP
τ
.Lemma 2.1
Letr, τ
≤ ∞
.Furthermore letq
0
∈
R
m
and
T, δ >
0
. Thenthefollowingshold.
a) Thereexistsafuntion
y
∈
C
r
(
R
,
R
)
beingompatlysupportedin
]
−
δ,
0[
andsatisfyingP
τ
y
=
q
0
.b) There exists a
T
-periodi funtiony
∈
C
r
(
R
,
R
)
with
P
τ
y
=
q
0
andsupp
(
y
|
[
−
(
k
+1)
T,
0]
)
⊂
]
−
(
k
+ 1)
δ,
0[
forallk
∈
Z.
Proof:
Ad a) Letǫ
:=
min
{
δ, τ
}
. The mappingL
1
([
−
ǫ,
0])
∋
y
7→
K
(
y
) :=
R
0
−
ǫ
e
−
At
by
(
t
)
dt
∈
R
m
is linear, ontinuousand beause of the ontrollability
of
(
A, b
)
surjetive.C
r
ǫ
is a dense linear subspae ofL
1
([
−
ǫ,
0])
. Therefore
R
:=
K
(
C
r
ǫ
)
is a dense linear subspae ofR
m
and onsequently
R
=
R
m
.
Therefore there exists a funtion
y
0
∈
C
r
ǫ
having the propertyK
(
y
0
) =
q
0
.Thetrivialextensionof
y
0
onR
isobviouslythedesiredfuntion.Ad b) If
τ
≤
T
, the assertion follows diretly from part a). We prove the result forτ > T
using sampling. Assume rstτ <
∞
and letN
:=
max
{
k
∈
N
:
N T
≤
τ
}
. Due to thestabilityofA
theseriesP
∞
k
=0
e
kT A
∗
ReallthatMorse-Smalevetoreldsarestruturallystable.
∗It
isstillunknownwhetherthe
C
k
onvergesto
(
I
−
e
T A
)
−
1
and
S
N
:=
P
N
k
=0
e
kT A
= (
I
−
e
T A
)
−
1
(
I
−
e
(
N
+1)
T A
)
isinvertible. Letuswriteǫ
:=
min
{
τ
−
T, δ
}
. Aordingtoparta)thereexists˜
y
0
∈
C
ǫ
r
withP
T
y
˜
0
=
S
−
1
N
q
0
(as well as(
I
−
e
T A
)
q
0
in aseτ
=
∞
). Lety
0
denote the trivial extension of
y
˜
0
on[
−
T,
0]
. Then wehaveP
T
y
0
=
S
−
1
N
q
0
.Let
y
denotetheT
-periodiextensionofy
0
onR.
ConsequentlyZ
0
−
τ
e
−
At
by
(
t
)
dt
=
N
X
k
=0
Z
−
kT
−
(1+
k
)
T
e
−
At
by
0
(
t
+
kT
)
dt
=
N
X
k
=0
e
kT A
Z
0
−
T
e
−
At
by
0
(
t
)
dt
=
q
0
,
whihyieldsimmediatelythedesiredonlusion.
Let
π
denotetheanonialprojetionof thetangentbundleT M
:π
:
T M
→
M, π
(
v
) :=
x
forv
∈
T
x
M.
Let
M
beendowedwithaRiemannianmetri. Denotingtheinduednormonthetangentspaesby
|
.
|
,theunit tangentbundleT
1
M
isgivenbyT
1
M
:=
∪
x
∈
M
{
v
∈
T
x
M
:
|
v
|
= 1
}
.Reall that
T
1
M
is a(2
n
−
1)
-dimensionalC
r
−
1
submanifold of
T M
. It isompat,if
M
isompat.Let
K
beasubsetofM
. WesetT
0
K
:=
K
×
K
\
∆
K
anddenotetherestrition ofT
1
M
toK
withT
1
K
,i.e.,T
1
K
:=
{
v
∈
T
1
M
:
π
(
v
)
∈
K
}
. Reallthat ifK
isans
-dimensionalsubmanifold,thenT
1
K
hasdimensionn
+
s
−
1
.Definition 2.3
We dene theτ
-history ofK
by the owΦ
of the ve-toreldf
tobethelosureofΦ(
K
;
τ
) :=
{
Φ
t
(
x
) :
−
τ < t
≤
0
, x
∈
K
}
.Wedenote
∆Θ
f,h
(
x, x
′
) := Θ
f,h
(
x
)
−
Θ
f,h
(
x
′
)
forx, x
′
∈
M
.
Let
V
be an open subset ofM
ontaining theτ
-historyofK
andL
be thelosureof
V
. ThenwedenoteH
0
(
L
;
K
) :=
{
h
∈
C
r
(
L,
R
) :
zeroisaregularvalueof∆Θ
f,h
|
Λ
0
K
}
,and
Inthesequelweset
i
= 0
,
1
aswellasH
i
(
K
) :=
H
i
(
M
;
K
)
andH
i
:=
H
i
(
M
)
.If
K
′
⊂
K
and for an output funtion
h
, zero is a regular value of themapping
∆Θ
f,h
onΛ
0
K
,thenitis alsoaregularvalueoftherestritedmap-ping
∆Θ
f,h
|
Λ
0K
′
,thatis,H
0
(
K
)
⊂
H
0
(
K
′
)
. Similarly
H
1
(
K
)
⊂
H
1
(
K
′
)
holds.
In the following, if
f
∈
X
r
0
,
1
(
M
)
, thenH
0
(
L
)
stands for the set of outputfuntions
h
∈
C
r
(
L,
R
)
suhthat
h
(
x
0
)
6
=
h
(
x
′
0
)
forall(
x
0
, x
′
0
)
∈
Λ
0
(
L
∩
Sing
(
f
))
,and
H
1
(
L
)
forthoseh
∈
C
r
+1
(
L,
R
)
suh that
d
x0
h
isaΘ
-oyliovetorofd
x0
f
forallx
0
∈
L
∩
Sing
(
f
)
.Note that if
L
isompat,then thenumberof singularitiesoff
inL
isniteand for nite
r
, as an immediate appliation of the transversality opennessanddensitytheorems,itfollowsthat
H
i
(
L
)
isanopenanddensesubsetoftheBanahspae
C
r
+
i
(
L,
R
)
.Lemma 2.2
Assume thati < r <
∞
andf
∈
X
r
0
,
1
(
M
)
is omplete. LetS
beaC
r
submanifold of
M
suh that theτ
-historyofS
is ontainedin anopen subset
V
ofM
with ompat losureL
:= ¯
V
. Consider the mappingsF
i
:
H
i
(
L
)
×
Λ
i
S
→
R
m
denedbyF
0
(
h, x, x
′
) := ∆Θ
f,h
(
x, x
′
)
and
F
1
(
h, v
) :=
d
π
(
v
)
Θ
f,h
(
v
)
.Thenthefollowingholds.
a) Zeroisaregularvalueof
F
0
.
b)Zeroisaregularvalueof
F
1
.
Supposemoreoverthat
f
∈
X
r
2
(
M
)
. Thentheassertionsalsohold forτ
=
∞
.Proof:
Ad a) LetW
0
:=
{
(
h, x, x
′
)
∈
H
0
(
L
)
×
T
0
S
:
F
0
(
h, x, x
′
) = 0
}
.We have to show that the funtion
F
0
is submersive on
W
0
. SineR
m
is nite dimensional, it suies to prove that the linear mapping
d
(
h,x,x
′
)
F
0
:
T
(
h,x,x
′
)
(
H
0
(
L
)
×
T
0
S
)
→
R
m
is surjetiveforall(
h, x, x
′
)
∈
W
0
.Fix
(
h, x, x
′
)
∈
W
0
andq
0
∈
R
m
. Aordingtotheondition
h
∈
H
0
(
L
)
wesee thatx
andx
′
annotbebothequilibriumpoints. Thereforeweassumewithout
lossofgeneralitythat
x
isnotanequilibriumpoint. SineH
0
(
L
)
isopenandd
ds
|
s
=0
F
0
(
h
+
sg, x, x
′
) = ∆Θ
f,g
(
x, x
′
) =
F
0
(
g, x, x
′
)
, it is suient to showtheexisteneof anoutputfuntion
g
∈
C
r
(
L,
R
)
satisfying
F
0
(
g, x, x
′
) =
q
0
.Weusethefat thattheowthroughapointofthestatespae
M
mapseahlosednitetimeintervalonalosedsubsetof
M
anddeneasuitablemappingLet
γ
andγ
′
denote the
τ
-histories of the pointsx
andx
′
respetively and
Z
:=
γ
∪
γ
′
. Wedene
g
onZ
. Wersttreattheaseτ <
∞
.Case1: Bothorbitsareritialelements. Let
T
denotetheperiodofx
. Inviewof lemma 2.1 there exists a
T
-periodi funtiony
∈
C
k
(
R
,
R
)
, whih satises
theondition
P
τ
y
=
q
0
. Wesetg
(Φ
t
+
kT
(
x
)) =
y
(
t
)
for0
≤
t
≤
T, k
∈
Z
andg
= 0
else.Case2: Oneof theintegralurves,say
Φ(
x
)
, isinjetiveandtheother oneisperiodi oran equilibriumpoint. AordingtoLemma 2.1there isafuntion
y
∈
C
k
([
−
τ,
0]
,
R
)
withompatsupportsuhthat
P
τ
y
=
q
0
. Wedeneg
onγ
by
g
(Φ
t
(
x
)) =
y
(
t
)
for−
τ
≤
t
≤
0
andg
= 0
else. IfΦ(
x
′
)
isinjetiveand
x
isperiodi, wejust set
g
= 0
onγ
and deneg
onγ
′
suhthat
P
τ
(
g
◦
Φ(
x
′
))) =
−
q
0
.Case3: Both integralurvesareinjetive. Inthis asewedene
g
onγ
asinase2andon
γ
′
by
g
= 0
.If
τ
=
∞
,thenbeauseofeventualpreseneofdenseorbitsg
annotbesimply dened on a part ofZ
and then trivially extended. Hene the onstrutionof
g
beomes alittle moredeliate. Assumingf
∈
X
2
ensures thenthat the reurrene is trivial and the previous proedure also works. If both orbitsare ritial elements, i.e. in ase 1, then everythingremains the sameasfor
nite observation time. Problems ould arise in ase 2 or 3 if at least one
of the integral urves is injetive and the past half of one of the orbits, say
the one through
x
′
, belongs to the negative limit set of the other orbit or
itself, i.e.,
{
Φ
t
(
x
′
) :
t
≤
0
} ⊂
α
(
x
)
∪
α
(
x
′
)
. But this an aording to the
assumption
f
∈
X
r
2
(
M
)
onlyourifx
′
isperiodioranequilibriumpoint. We
proeedasin ase2ofnite
τ
andndagaininviewofLemma2.1afuntiony
∈
C
r
(
R
,
R
)
ompatly supported in an interval
[
−
ǫ,
0]
withǫ >
0
and setg
(Φ
t
(
x
)) :=
y
(
t
)
forallt
andg
= 0
else.Inallaseswehavedened a
C
r
funtion onthelosedsubset
Z
ofthestatespae
M
with the property thatZ
⊂
L
andP
τ
(
g
◦
Φ(
x
)
−
g
◦
Φ(
x
′
)) =
q
0
.AordingtothesmoothTietzeextensiontheorem,thereexistsa
C
r
extension
of the funtion
g
toL
. This funtion denoted again byg
is obviously thedesiredfuntion,whihsatises
F
0
(
g, x, x
′
) =
q
0
.Ad b) Denote
W
1
:=
{
(
h, v
)
∈
H
1
(
L
)
×
T
1
S
:
F
1
(
h, v
) = 0
}
. We x
(
h, v
)
∈
W
1
, setx
0
:=
π
(
v
)
and show thatF
1
is submersive at
(
h, v
)
. Fixq
0
∈
R
m
. Sined
ds
|
s
=0
F
1
(
h
+
sg, v
) =
F
1
(
g, v
)
for arbitrary
g
∈
C
r
(
M,
R
)
,itsues to provethe existene of afuntion
g
∈
C
r
(
L,
R
)
with
F
1
(
g, v
) =
q
0
loally and extend it
L
. Ifx
0
would be an equilibrium point, then in viewof the assumption
h
∈
H
1
(
L
)
, the linearsystem(
T
x0
M, d
x0
f, d
x0
h
)
would beΘ
-observableandonsequentlyF
1
(
h, v
)
6
= 0
inontraditiontotheassumption
(
h, v
)
∈
W
1
. Therefore wemay assume thatx
0
is not an equilibrium point.Hene, in viewof the straightening-out theorem there is a loal hart
(
U, ψ
)
at
x
0
suhthatψ
(
U
) =
U
′
×
]
−
ǫ,
−
ǫ
[
with
ǫ >
0
,U
′
anopen subsetof
R
n
−
1
,
ψ
(
x
0
) = 0
andthevetoreldf
hastheloalrepresentative(
z, t
)
7→
e
n
. Heree
n
denotesthen
thstandardbasevetorinR
n
funtion on
T
1
M
atv
byψ
b
. Sinev
6
= 0
we anand do assumethatv
has the loal representativeη
= (
η
1
, η
2
)
T
with
η
1
∈
R
n
−
1
,
η
2
∈
R,
η
6
= 0
, i.e.v
=
∂
∂z
(
x
0
)
η
1
+
∂
∂t
(
x
0
)
η
2
. Furthermorefort
∈
]
−
ǫ,
−
ǫ
[
,theloalrepresentativeof
d
x0
Φ
t
readsasId
0
0
1
andsubsequentlythat of
d
x0
h
◦
Φ
t
v
as∇
h
(0
, t
)
η
.By shrinking
U
if neessary, we an (in the aseτ
=
∞
on aount of the assumption thatf
∈
X
r
2
(
M
)
if the pointx
0
is reurrent, then it is periodi)anddoassumethattheintersetionof
U
andtheτ
-historyofx
0
isonneted.Aordingtolemma2.1thereexistsafuntion
y
ˆ
∈
C
r
+1
(
R
,
R
)
withderivative
y
being supported (on eah period, ifx
0
has period> τ
) in[
−
ǫ,
0]
suhthat
P
ǫ
y
(
t
)
dt
=
q
0
. Obviouslythere exists afuntiong
∈
C
r
(
U
′
×
]
−
ǫ, ǫ
[
,
R
)
being ompatly supported, with
∇
g
(0
, t
)
η
=
y
(
t
)
. For instane dene˜
g
(
z, t
) =
|
η
|
−
2
(
z
T
η
1
y
(
t
) +
η
2
y
ˆ
(
t
))
. Thetrivial extension ofg
˜
◦
ψ
toL
is thedesiredfuntion
g
.Sometimes in appliations one is interested in or limited to observation
re-stritedtoasubsetofthestatespae. Itanbeforinstanebeauseoftehnial
orphysialreasons,orifithappensthatalltheinformationneededanbe
eval-uatedfrom measurementsonaertainsubset. Thelatterasebeingperhaps
themostimportantone,oursifthesubsetunderobservationisanattrator.
Other subsets invariant under the owan also be of interest. Thereforewe
stateourgeneriityresultsforobservationsofsubsetsofthestatespaeaswell.
Lemma 2.3
Suppose thatK
is a subset of ans
-dimensionalC
r
submani-fold of
M
denoted byS
,τ <
∞
andf
∈
X
r
0
,
1
(
M
)
is omplete. Then thefollowingholds.
a) Assume that
m
≥
n
+
s
−
r
andr
≥
2
. ThenH
1
(
K
)
is residual. IfK
is losed,thenH
1
(
K
)
isalsoopen.b) Assume that
m
≥
n
+
s
+ 1
−
r
. ThenH
0
(
K
)
is residual. IfK
is losed, thenH
0
(
K
)
ontainsanopenset.Suppose moreover that
M
is ompat andf
∈
X
r
2
(
M
)
. Then the assertionsholdalsoin thease
τ
=
∞
.Proof:
Assume rstr <
∞
,K
is ompat,U
is a hart domain ofS
, whih ontainsK
and hasompatlosure. ByompatnessofU
andnite-ness of
τ
in ase ofnite observationtime and beause of ompatnessofM
in ase
τ
=
∞
, theτ
-historyofU
is ompat. By loalompatness thereis an open setV
⊂
M
with ompat losureL
:=
V
suh thatV
ontains theτ
-historyofU
.Loal density: We prove residuality with respet to
H
i
(
L
)
. Sine the latteris open and dense in
C
r
(
L,
R
)
, density of
H
i
(
L
;
U
)
is also then shown withrespetto
C
r
(
L,
R
)
.
map-ping
F
i
:
H
i
(
L
)
×
T
i
U
→
R
m
dened byF
0
(
h, x, x
′
) := ∆Θ
f,h
(
x, x
′
)
and
F
1
(
h, v
) :=
d
π
(
v
)
Θ
f,h
(
v
)
.Thereforeaordingtothetransversalitydensitytheorem(seeforinstane[AR ,
19.1℄)
H
i
(
L
;
U
)
isaresidualsubsetofH
i
(
L
)
,andhenedenseinC
r
(
L,
R
)
.
Loal Openness: Compatnessof
K
impliesthatT
1
K
isalsoompatinT
1
U
.Aordingto the transversalityopennesstheorem
H
1
(
L
;
K
)
is open (withre-spet tothe ompat-open andby ompatnessof
L
alsowith respet to theWhitney topology)in
C
r
(
L,
R
)
. These onlusionsdonotworkfor
H
0
(
L
;
K
)
.Instead let
Λ
be a ompat subset ofT
0
U
. Then in view of transversalityopennesstheorem
H
′
0
(
L
; Λ) :=
{
h
∈
C
r
(
L,
R
) :
zeroisaregularvalueof∆Θ
f,h
onΛ
}
isopenin
C
r
(
L,
R
)
.
Sinetheassertionsareprovedfor
r
niteandsuientlylarge,theyalsoholdfor
r
=
∞
. Heneweassumefromnowonthatr
≤ ∞
.Globalization: This part of the proof is basily standard. Therefore we
give an outline and refer to [H, 2.2℄ for details. Sine
K
⊂
U
, we haveH
1
(
L
;
U
)
⊂
H
1
(
L
;
K
)
. HeneH
1
(
L
;
K
)
isalsodenseinC
r
(
L,
R
)
. Likewiseone
getsdensityof
H
′
0
(
L
; Λ)
inC
r
(
L,
R
)
fromdensityof
H
0
(
L
;
U
)
,sineΛ
⊂
T
0
U
andsubsequentlyH
0
(
L
;
U
)
⊂
H
′
0
(
L
; Λ)
. Usingabump funtionwenowprovethat
H
1
(
K
)
isdenseinC
r
(
M,
R
)
. Fix
g
0
∈
C
r
(
M,
R
)
. Sine
H
1
(
L
;
K
)
isdensein
C
r
(
L,
R
)
, thereisasequene
{
h
j
}
inH
1
(
L
;
K
)
onvergingin the ompat-opentopologytog
0
|
L
. SineK
⊂
U
,thereisaC
r
-funtion
ρ
:
M
→
[0
,
1]
with ompat support inL
, suh thatρ
= 1
onan open neighborhood ofK
. Thesequene
ρh
j
+ (1
−
ρ
)
g
0
onvergestog
0
withrespettotheWhitneytopology. ThereforeH
1
(
K
)
is a dense subset ofC
r
(
M,
R
)
. A similar argumentshows
that
H
′
0
(Λ)
isadensesubsetofC
r
(
M,
R
)
.Wenowdroptheassumptionthat
K
⊂
U
. LetJ
beaountable indexingset,{
U
j
}
withj
∈
J
be aoveringofS
with hartdomainsU
j
. Furthermore let{
K
j
}
beasubordinatefamilyofompatsetssuhthatK
=
∪
J
K
j
and
K
j
⊂
U
j
.Weananddoassumethatthereisaompatoveringof
M
denotedby{
L
j
}
suhthattheinteriorofL
j
ontainstheτ
-historyofU
j
.Opennessstatements: Supposethat
K
islosed. Then itisalso paraompatand theoveringanbeassumedtobeloallynite. Sine
{
L
j
}
oversM
,it holds thatH
1
(
K
) =
{
h
∈
C
r
(
M,
R
) :
h
|
L
j
∈
H
1
(
L
j
;
K
j
)
forallj
∈
J
}
. Heneloalnitenessoftheoveringimpliesthat
H
1
(
K
)
isopen. Simliarilyitfollowsthat theset
{
h
∈
C
r
(
M,
R
) :
zeroisaregularvalueof
∆Θ
f,h
onK
×
K
}
is open. ThelatterisontainedinH
0
(
K
)
.Residualitystatements: Wenowdroptheassumptionthat
K
islosed. Bythepreedingarguments
H
1
(
K
j
)
isopenanddense. ThereforeH
1
(
K
) =
∩
j
isresidual(andinviewoftheBairepropertyof
C
r
(
M,
R
)
alsodense). Taking
a ompat overing
{
Λ
j
}
ofT
0
K
a similar argument yields the results onH
0
(
K
)
.The openness results in the preeding lemma an also be proved (without
applying the transversality openness theorem) as follows. For instane, by
ompatnessof
L
themappingH
0
(
U
)
∋
h
7→
F
0
(
h, .
))
∈
C
r
(
T
0
U,
R
m
)
isontin-uousin theWhitneytopology. Thisfatand opennessof
{
F
∈
C
r
(
T
0
U,
R
m
) :
zeroisaregularvalueof
F
}
(this fat follows, for instane, from the lower semiontinuityofthemappingC
r
(
L,
R
)
×
T
0
U
∋
(
h, x, x
′
)
7→
rank
(
d
(
h,x,x
′
)
F
0
)
and ompatness of
L
) in the Whitney topology implies thatH
0
(
L
;
U
)
isopen in the Whitney topology and by ompatness of
L
also open in theompat-opentopologyof
C
r
(
L,
R
)
. Otherpartsfollowlikewise.
In light of the preeding lemmas wean nowprove the generiity results on
outputfuntions justbyomparingdimensions.
Theorem 2.1
Suppose thatK
is ontained in ans
-dimensionalC
r
sub-manifold of
M
andf
∈
X
r
0
,
1
(
M
)
is omplete. Then the following assertionshold.
a)Assumethat
r
≥
2
andm
≥
n
+
s
. Thenfuntionsh
belongingtoC
r
(
M,
R
)
suhthat
Θ
f,h
isimmersiveateahpointofK
,onstitutearesidualset. This set isalsoopen,ifK
islosed.b)Assumethat
m
≥
2
s
+ 1
. Thenthesetoffuntionsh
belongingtoC
r
(
M,
R
)
suh that
Θ
f,h
is injetive (respetively an injetive immersion) onK
, isresidual(respetively,if
r
≥
2)
. Itontainsanopenset,ifK
is losed. Supposingm
≤
2
s
andr >
2
s
−
m
thesameresultsholdforthesetoffuntionsh
belongingtoC
r
(
M,
R
)
suhthatthe
Θ
-unobservablepointsofK
×
K
belong to asubmanifoldofdimension2
s
−
m
.Suppose moreover that
M
is ompat andf
∈
X
r
2
(
M
)
. Then the assertionsalsoholdfor
τ
=
∞
.Proof:
Ad a) Sinem
≥
n
+
s
, the set of funtionsh
belonging toC
r
(
M,
R
)
suh that
Θ
f,h
isan immersion at eah point ofK
, oinides withH
1
(
K
)
, i.e.,d
Θ
f,h
|
T
1
K
is transversal to{
0
} ∈
R
m
. Therefore the assertion
followsimmediatelyfrom thepreviouslemma.
Ad b) Aording to the previous lemma
H
0
(
K
)
is residual and ontains anopen set,if
K
is losed. Ifm
≥
2
s
+ 1
, thentheset of funtionsh
belonging toC
r
(
M,
R
)
suhthattherestritionof
Θ
f,h
toK
isinjetive(respetivelyaninjetiveimmersion), is preisely
H
0
(
K
)
(respetivelyH
0
(
K
)
∩
H
1
(
K
)
). The statementontheindistinguishablesetfollowsfromthepreimagetheorem.Ifthestatespaeisompat,wegetsharperresults. Inpartiularthefollowing
theoremisimportant,sineembeddingofthestatespaegivesinformationon limitbehaviorofthesystem.
Theorem 2.2
Suppose thatm
≥
2
n
+ 1
,f
∈
X
r
0
,
1
(
M
)
andM
isom-pat. Then output funtions
h
∈
C
r
(
M,
R
)
suh that
Θ
f,h
is an embeddingonstituteanopenanddenseset.
If wefurther assumethat
f
∈
X
r
2
(
M
)
, then theassertionremainstrue in thease
τ
=
∞
.Proof:
Reall that by ompatness ofM
an injetive immersion is also an embedding (ase2
≤
r
≤ ∞
in Theorem 2.1) and an injetive mapping is also a topologial embedding (aser
= 1
). Hene density follows fromTheorem2.1. Opennessisaonsequeneofthefatthatbyompatnessof
M
themapping
C
r
(
M,
R
)
×
T
0
M
∋
h
7→
∆Θ
f,h
∈
C
r
(
M,
R
m
)
isontinuousandthe set of embeddings
Emb
r
(
M,
R
m
) :=
{
F
∈
C
r
(
M,
R
m
) :
F
isembedding
}
isopen.
Nextweshallproveresidualityof
X
r
0
,
1
(
M
)
byusingtheharaterizationofsim-pliityand
Θ
-simpliityofasingularity,intermsoftransversalnonintersetion.Lemma 2.4
Assume thatm
≥
n
. ThenX
r
0
,
1
(
M
)
is open and dense inX
r
(
M
)
. Moreover,theassertionholdsalsofor
τ
=
∞
,ifwerestritthevetor elds toX
r
(
M, a
)
.Proof:
We give a proof forτ <
∞
. The arguments forτ
=
∞
are similar. It suies to show thatX
r
1
(
M
)
is open and dense. LetO
resp.U
denote the set ofn
-dimensionalΘ
-observable resp. unobservable linearsystems. Let
f
∈
X
r
1
(
M
)
,x
0
∈
Sing
(
f
)
andd
x0
f
denote theprinipialpartofthelinearizationof
f
.Note that
f
∈
X
r
1
(
M
)
if and only if there exists av
∈
T
x0
M
suh that thesystem
(
d
x0
f, v
)
avoidsthesetofΘ
-unobservablelinearsystemsonT
x0
M
.We now give a loal haraterization. Let
ψ
:
U
→
R
n
be a loal hart
suh that
U
has ompat losure andψ
(
x
0
) = 0
. The tangent mappingdψ
:
T U
→
R
n
×
R
n
denedbydψ
(
v
) = (
ψ
(
x
)
, d
x
ψv
)
withx
=
π
(
v
)
isaloalhartfor
T M
. LetP r
2
denotetheprojetionR
n
×
R
n
∋
(
x, w
)
7→
w
. Consider
themapping
ξ
f
:=
P r
2
dψf
◦
ψ
−
1
on
ψ
(
U
)
. Hened
0
ξ
f
=
d
x0
ψd
x0
f d
0
ψ
−
1
. Thesingularity
x
0
isΘ
-simpleifand onlyif(
d
0
ξ
f
, d
x0
v
)
∈
/
U
forsomev
∈
T
x0
M
. ObviouslyU
is losed, analyti and6
=
End
(
R
n
)
×
R
n
, hene nite union of
losed positive odimensional real analyti submanifold of
End
(
R
n
)
×
R
n
. Givenapair(
G, w
)
∈
O,
obviouslythere existsavetoreldg
∈
X
r
(
M
)
suhthat
g
andf
oinide onM
\
U
,x
0
∈
Sing
(
g
)
andd
x0
ξ
g
=
G
. Animmediate appliation (details are similar to those in the proof of lemma 2.3) of theWeannowprovethemainresultongeneriobservabilitywithrespettothe
mapping
Θ
.Theorem 2.3
Suppose thatm
≥
2
n
+ 1
andM
is ompat. Then pairs(
f, h
)
inX
r
(
M
)
×
C
r
(
M,
R
)
suh that
Θ
f,h
is an embedding onstitute anopen and dense subsetof
X
r
(
M
)
×
C
r
(
M,
R
)
. Restriing thevetorelds to
X
r
2
(
M
)
, theassertionremainsvalidforτ
=
∞
aswell.Proof:
Reall thatX
r
0
,
1
(
M
)
is open and dense. FurthermoreX
r
2
(
M
)
isopen in
X
r
(
M, a
)
. The assertions follow immediately from this fats and
theorem2.2.
Similarily,generiityresultsfornonompatstatespaesanbederiven,ifwe
restritourselfto thesetof ompletevetorelds, replaeopenand denseby
residual andembedding by injetiveimmersion. We remarkalso that
onsid-eringtheobservationmapping
Θ
g,h
(
x
) :=
0
X
k
=
−
N
e
−
kA
bh
(
g
k
(
x
))
for
x
∈
M
and(
g, h
)
∈
Dif f
r
(
M
)
×
C
r
(
M,
R
)
withg
k
:=
g
◦
g
k
−
1
,
orrespond-ingresultsfordisretedynamialsystemsanbeprovedlikewise.
3
Concluding Remarks
Remark 3.1
Sinetheset ofm
-dimensional ontrollablelinearltersisopen anddenseinEnd
(
R
m
)
×
R
m
,allgeneriityresultsofthelastsetionholdalso
generiallywithrespettothelinearlters.
Thefollowingexamplesshowthattheonditionsontheobservationdimension
arealsoneessary,thus
m
≥
2
n
forgeneriloalobservabilityandm
≥
2
n
+ 1
forgeneriobservabilityannotbeweakend.Example 3.1
LetM
=
S
1
and
f
(
ϕ
) = 1
, whereϕ
denotes the standardangular oordinate of the irle. Furthermore onsider the pair
(
λ, b
)
withλ <
0
andb
6
= 0
. Takingτ
= 2
π
andtheoutputfuntionh
(
ϕ
) =
cos ϕ
leadstoΘ
f,h
(
ϕ
) =
1
−
e
2
λπ
1+
λ
2
(
λcos ϕ
−
sin ϕ
)
, whihis notan immersion. Moreoverthezeroof
d
Θ
f,h
atϕ
0
=
−
arctan
1
λ
istransversal. Hene thenonimmersivityofΘ
f,h
ispreservedundersmallperturbationsoftheoutputfuntion,thevetoreld andthelinearlter.
Example 3.2
LetM
,f
,ϕ
andτ
beasinthepreedingexample. Furthermore letA
=
diag
(
−
1
,
−
2)
,b
= (1
−
e
−
2
π
)
−
1
(1
,
1)
T
Thenastraightforwardomputationyields
Θ
f,h
(
ϕ
) =
cos ϕ
+
sin ϕ
+
cos
2
ϕ
+ 2
sin
2
ϕ
(
e
−
2
π
+ 1)(
2
5
(2
cos ϕ
+
sin ϕ
) +
5
4
(
cos
2
ϕ
+
sin
2
ϕ
))
.
Thefollowinggureshowstheimageof
S
1
by
Θ
f,h
.−3
−2
−1
0
1
2
3
4
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
3
X
Y
Imageof
S
1
bytheontinuouslinearltermappinggivenby
ϕ
7→
(cosϕ+sinϕ+cos2ϕ+2sin2ϕ,
(e
−
2
π
+1)(
2
5
(2cosϕ+sinϕ)+
5
4
(cos2ϕ+sin2ϕ))
T
inthe
XY
-planeTheselntersetion oftheimageis transversal. Hene thenoninjetivenessof
Θ
f,h
ispersistentundersmallperturbationsoftheoutputfuntion,thevetoreld andthelinearlter.
Forinstane, smallperturbationsof
h
donotresultininjetivity, i.e.,thereisan
ǫ >
0
suhthat foreahoutputfuntion˜
h
,whihisC
r
nearto
h
withinǫ
,themapping
Θ
f,
˜
h
isnotinjetive.Notethattheonsideredsystemisalsounobservablewithrespettothe high-gainmappinggivenby
ϕ
7→
(2
cos ϕ
+ 5
cos
2
ϕ,
2
sin ϕ
−
10
sin
2
ϕ
)
T
aswellas
samplingmapping
ϕ
7→
(2
cos
(
ϕ
+
t
1
) + 5
cos
(2
ϕ
+ 2
t
1
,
2
cos
(
ϕ
+
t
2
) + 5
cos
(2
ϕ
+
2
t
2
))
T
withsamplingtimest
1
, t
2
. Thefollowinggureshowstheimage ofthe−6
−4
−2
0
2
4
6
8
−6
−4
−2
0
2
4
6
8
X
Y
Imageof
S
1
bythesamplingmappinggivenby
ϕ
7→
(2cos ϕ
+ 5cos
2ϕ,
−
2sin ϕ
−
5cos
2ϕ)
T
inthe
XY
-planeReferences
[AM℄ R. Abraham and J. Marsden: Foundations of Mehanis,
Ben-jamin/Cummings PublishingCompany,1978
[AR℄ R. Abraham and T. Ratio: Transversal Mappings and Flows,
Ben-jamin/Cummings Publishing,1967
[Ay℄ D. Aeyels: Generi Observabilityofnonlinearsystems, SIAMJournalof
ControlandOptimization,19,1981a,595-603
[E℄ R.Engel: ZustandsbeobahterfürNihtlineareSysteme,Dissertation
Uni-versitätKassel,2002
[GG℄ M.GolubitskiandV.Guillemin: StablemappingsandtheirSingularities,
GraduateTextsinMathematis,Vol.14,Springer-Verlag,1974
[GHK℄ J. P. Gauthier, G. Hammouri and I. Kupka: Obervers for nonlinear
systems, Pro. of the 30th Conf. on Deision and Control in Brighton,
england,1483-1489,1991
[GK℄ J. P. Gauthier and I. Kupka: Deterministi Observation Theory and
Appliations,CambridgeUniversityPress,2001
[H℄ M.W. Hirsh: DierentialTopology,GraduateTextsin Mathematis,Vol
33, Springer-Verlag,1976
[J℄ Ph. Jouan: Observability of real analyti vetor elds on ompat
[KE℄ G.KreisselmeierandR.Engel: Nonlinearobserversforautonomous
Lip-shitz ontinous systems, IEEE Trans. Autom. Control., Vol. 48, No. 3,
Marh2003
[KH℄ A.KatokandB.Hasselblatt: Introdutionto themodern theoryof
dy-namialsystems,CambridgeUniversityPress,1995
[N℄ E. Nava yazdani: Generi Observability of Smooth Dynamial Systems,
DissertationUniversitätKassel,2005
[P℄ M.M.Peixoto: OnanapproximationtheoremofKupkaandSmale,Journal
of DierentialEquations3,214-227,1966
[PD℄ J. Palis and W. De Melo: Geometri Theory of Dynamial Systems,
Springer-Verlag,1982
[Pu℄ C.C. Pugh: An improvedlosing lemma and ageneraldensitytheorem,
Amer. J.Math.,89,1967,1010-1021
[RT℄ D. Ruelle and F. Takens: On the Nature of Turbulene, Comm. Math.
Phys., 20,1971,167-192
[T℄ F.Takens: Detetingstrangeattrators inturbulane,Proeedings ofthe
SymposiumonDynamialSystemsandTurbulane,LetureNotesin
Math-ematis898,366-381,Springer-Verlag,1981
EsfandiarNavaYazdani