2017 IEEE 56th Annual Conference on Decision and Control (CDC)

December 12-15, 2017, Melbourne, Australia

Luenberger observers for nonlinear controlled systems

Pauline Bernard1

Abstract— We show how the (original) Luenberger method- in a ”Lipschitz” way in terms of the first N − 1 ones.
ology presented in [1] for autonomous linear systems can In this case, a classical high gain observer can be used.
be used to design causal observers for controlled non linear But a drawback of this transformation is that it involves
systems. Their implementation relies on the resolution of a time-
varying PDE, the solutions of which transform the dynamics the input’s time derivatives (appearing when differentiating
into linear asymptotically stable ones. We prove the existence the output), which may make this solution unsuitable for
and injectivity (after a certain time) of such transformations, practical applications.
under a differential observability assumption. We also show on The Luenberger-like methodology which we develop in
practical examples how this PDE can be solved and how the this paper also relies on an input-dependent transformation.
observability assumption can be checked.
Inspired by [1], [3], [5], the goal is to transform the system
I. I NTRODUCTION into a linear asymptotically stable system via the resolution
of a PDE, but which is now time-varying due to the presence
Online estimation of the state of a dynamical system of inputs. This resolution is not always straightforward,
is crucial in practice, especially for monitoring or control which makes this method less direct and systematic than the
purposes. Originally1 , in [1], Luenberger developed a general high gain methodology, but we will see that the computation
method to build observers for autonomous linear systems. of the derivatives of the input is not necessary, which
Its extension to autonomous nonlinear systems was proposed constitutes a significant advantage.
and analyzed in a general context by [2]. It was rediscovered
later by [3] where a local analysis close to an equilibrium II. P ROBLEM STATEMENT
point was given under conditions relaxed later on in [4].
The localness as well as most of the restrictive assumptions Consider a general system of the form
were then by-passed in [5]. As noticed in [6] and [5], this
nonlinear Luenberger observer is also strongly related to the ẋ = f (x, u) , y = h(x, u) (1)
observer proposed in [7]. where x is the state in Rn , y the measurement in Rp , f
In this paper, we want to extend the use of Luenberger and h sufficiently many times differentiable functions and
observers to controlled systems. First steps in this direction u : [0, +∞) → Rq an input of class C  , with  a sufficiently
were made in [8], [9] for linear time-varying systems, but up large integer. We denote X(x, t; s; u) the value at time s of
to our current knowledge, no result concerning Luenberger the solution to System (1) with input u, initialized at x at
observers for nonlinear time-varying - let alone controlled - time t, and Y (x, t; s; u) the corresponding output function
systems is available (although some preliminary results exist at time s. In all our forthcoming results, the following basic
in [10]). assumption will be made :
In fact, very few general observer designs exist for non lin- Assumption There exist a subset U of {u : [0, +∞) →
ear controlled systems. Some, such as the popular extended
Kalman filters ([11]) rely on linearization methods, but thus Rq | u ∈ C  } and a set S of Rn such that for any u in U ,
provide only local convergence. Others consist in finding a any x in S and any time t in [0, +∞), X(x, t; s; u) is in S
reversible input-independent change of coordinates, which for all s in [0, +∞).
transforms the dynamics into a more favorable form such In this paper, we want to design an observer for System
as state-affine time-varying forms ([12],[13] among others), (1), via the Luenberger-like methodology developed in [1],
for which a Kalman filter can be used, or a triangular form [2], [3], [5]. We assume the inputs and outputs are known
([14], [15]) for which a high gain observer can be used. in a causal way, namely the observer can use only their past
But the existence of such a change of coordinates usually or current values, i.e., at time t, u|[0,t] and y|[0,t] only2 .
requires restrictive assumptions on the system, such as the so- The idea is to find a transformation T : S ×[0, +∞) → Rd
called uniform observability for triangular forms. However, for some strictly positive integer d such that for any u in U ,
in [16], an explicit transformation into a more general time- any x in S and any time t in [0, +∞), the state function
varying triangular form is proposed under the weaker rank- T (X(x, t; s; u), s) follows the dynamics
observability condition. The authors then use a Kalman-
like observer (mixing of Kalman and high gain observers), ż = Az + By (2)
but an additional excitation condition is needed to ensure
convergence. with A a Hurwitz matrix in R and B a vector in R
d×d d×p
.
On the other hand, if we allow the transformation to This is achieved if T is solution to PDE
depend on the input, the range of possibilities widens. For ∂T ∂T
instance, the transformation obtained by considering the (x, t)f (x, u(t)) + (x, t) = A T (x, t) + B h(x, u(t))
measurement and a certain number of its derivatives, trans- ∂x ∂t
(3)
forms the dynamics into the so-called phase-variable form for any x in S and t in [0, +∞). In this case, implementing
([17]) under the ACP(N ) condition ([17]), which roughly the dynamics (2) provides an asymptotically converging es-
says that the N th-derivative of the output can be expressed timate of T (X(x, t; s; u), s). If T (·, s) is besides injective at
1 Pauline Bernard is with the Centre Automatique et Systèmes, MINES-
least after a certain time, one can finally deduce an estimate
for X(x, t; s; u). More precisely, we have the following
ParisTech, PSL Research University, France,
[email protected] result:
1 We write ”originally” insofar as what is nowadays usually called
”Luenberger observer” differs from what is in [1]. 2 Time 0 thus corresponds to the initial time of data recording.

978-1-5090-2873-3/17/$31.00 ©2017 IEEE 3676

Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on February 17,2024 at 20:25:02 UTC from IEEE Xplore. Restrictions apply.
 Theorem 1: Fix an input u in U . Suppose there exist a III. E XISTENCE OF SOLUTIONS TO PDE (3)
strictly positive integer d, a Hurwitz matrix A in Rd×d , a We have the following lemma :
vector B in Rd×p , and a function T : S × [0, +∞) → Rd Lemma 1: Consider i in {1, . . . , p} and a strictly positive
such that integer mi . Let Ai be a Hurwitz matrix in Rmi ∗mi , Bi
1) T is a C 1 solution to PDE (3) on S × [0, +∞) ; a vector in Rmi , and u an input function in U . For any
2) there exists a time t ≥ 0 and a concave K∞ function strictly positive real number k, the function Ti0 defined on
ρT such that for all (x1 , x2 ) in S 2 and all t ≥ t S × [0, +∞) by
 t
|x1 − x2 | ≤ ρT (|T (x1 , t) − T (x2 , t)|) 0
Ti (x, t) = ekAi (t−s) Bi Yi (x, t; s; u)ds (5)
i-e T becomes injective uniformly in time after a 0
certain time t. 1
is a C solution to PDE
Then, there exists a function T ∗ : Rd × [t, +∞) → S such
that for any x0 in S, and any z0 in Rd , the (unique) solution ∂Ti ∂Ti
(X(x0 , 0; s; u), Z(x0 , z0 , 0; s; u)) to (x, t)f (x, u(t)) + (x, t) = kAi Ti (x, t)
∂x ∂t
ẋ = f (x, u) , ż = Az + B h(x, u) (4) +Bi hi (x, u(t)) . (6)
verifies Proof: See Appendix II.
∗ It follows that by taking
lim |X(x0 , 0; s; u) − T (Z(x0 , z0 , 0; s; u), s)| = 0 .    
t→+∞
kA1 0 0 B1 0 0
Proof: See Appendix I. 0 kAi 0 0 Bi 0
A= B=
This result shows that it is sufficient to find a uniformly 0 0 kAp 0 0 Bp
(in time) injective solution to PDE (3) to be able to write
an observer. We will first check in Section III that there the function defined by
exist solutions to PDE (3). Then, we will give in Section  
IV sufficient conditions to ensure that any solution becomes T 0 (x, t) = T10 (x, t), . . . , Ti0 (x, t), . . . , Tp0 (x, t) (7)
injective after a certain time. Finally, in Section V, we
will show on practical examples how such solutions can be is a C 1 solution to PDE (3).
computed. Note that Ti0 depends only on the past values of the input
u. Thus, it is theoretically computable online. However, for
Notations each couple (x, t), one would need to integrate backwards
1) Since h (resp Y ) takes values in Rp , we denote hi the dynamics (1) until time 0, which is quite heavy. If the
(resp Yi ) its ith-component. input u is known in advance (for instance u(t) = t) it can
2) We denote U a subset of Rq such that for all u in U , also be computed offline.
u([0, +∞)) ⊂ U . We will see in Section V on practical examples how we
3) For some integer m which we choose later in As- can find a solution to PDE (3) in practice, without relying
sumption 2, any solution x to System (1) with some on the expression Ti0 . But, first, we need to see under which
input u, is such that (x, u, u̇, ..., u(m) ) is solution to conditions those solutions are injective.
the extended system
IV. I NJECTIVITY OF SOLUTIONS TO PDE (3)
ẋ = f (x, ν0 ) Assumptions
ν̇0 = ν1 The sets U and S, and the functions f and h are such that :
... ∂f
1) The quantity Mf = sup (x, ν0 ) is finite.
ν̇m = u(m+1) x∈S ∂x
ν0 ∈ U
2) There exist p integers (m1 , . . . , mp ) such that the
with input u(m+1) . We denote ūm = (u, u̇, ..., u(m) ), functions
ν̄m = (ν0 , ..., νm ), f¯ the extended vector field 
  Hi (x, ν̄m ) = h̄i (x, ν̄m ), Lf¯h̄i (x, ν̄m ), . . . ,
f¯(x, ν̄m , u(m+1) ) = f (x, ν0 ), ν1 , . . . , νm , u(m+1) 
i −1
Lfm
¯ h̄i (x, ν̄m ) (8)
and the extended measurement function
defined on S × Rq(m+1) with m = maxi mi and
h̄i (x, ν̄m ) = hi (x, ν0 ) . 1 ≤ i ≤ p verify :
- for all u in U , Hi (·, ūm (0))) is Lipschitz on S.
Note that while ν̄m is an element Rq(m+1) , ūm is - there exists LH such that the function
a function defined on [0, +∞) such that ūm (s) =

(u(s), u̇(s), ..., u(m) (s)) is in Rq(m+1) . We denote
H(x, ν̄m ) = H1 (x, ν̄m ), . . . , Hi (x, ν̄m ),
U m a subset of Rq(m+1) such that for all u in U ,
ūm ([0, +∞)) ⊂ U m . For 1 ≤ i ≤ p, the successive 
time derivatives of Yi are related to the Lie derivatives . . . , Hp (x, ν̄m ) (9)
of h̄i along the vector fields f¯, namely for j ≤ m
∂ j Yi verifies for any (x1 , x2 ) in S 2 and any ν̄m in U m
(x, t; s; u) = Ljf¯h̄i (X(x, t; s; u), ūm (s)) .
∂sj |x1 − x2 | ≤ LH |H(x1 , ν̄m ) − H(x2 , ν̄m )|

3677
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on February 17,2024 at 20:25:02 UTC from IEEE Xplore. Restrictions apply.
 namely H is Lipschitz-injective on S, uniformly is that the latter does not require the computation of the
with respect to ν̄m in U m . derivatives of the input (see examples in Section V).
3) For all 1 ≤ i ≤ p, there exists Li such that for all In order to check the assumptions of Theorem 2 more
(x1 , x2 ) in S 2 and for all ν̄m in U m , easily in practical cases, we have the following result :
Lemma 2: Assume that S is compact and there exist p
|Lm
f¯
i
h̄i (x1 , ν̄m ) − Lm
f¯
i
h̄i (x2 , ν̄m )| ≤ Li |x1 − x2 | integers (m1 , ..., mp ) such that U m with m = maxi mi is
compact, and for any ν̄m in U m , H(·, ν̄m ) defined in (9) is
namely Lm f¯
i
h̄i (·, ν̄m ) is Lipschitz on S, uniformly with an injective immersion4 on S. Then, Assumptions 1, 2, 3 are
respect to ν̄m in U m . satisfied.
Proof: See Appendix IV.
We are now ready to introduce our main result : In other words, since the additional assumption ”T (·, 0)
Theorem 2: Suppose Assumptions 1-2-3 are satisfied. Lipschitz on S” made in Theorem 2 is automatically verified
Consider Hurwitz matrices Ai in Rmi ∗mi , with mi defined when S is compact, the result of Theorem 2 holds under the
in Assumption 2, and vectors Bi in Rmi such that the pairs assumptions of Lemma 2.
(Ai , Bi ) are controllable. There exists a strictly positive real
number k such that for all k ≥ k, for all input u in U , V. E XAMPLES
there exists tk,u such that any C 1 solutions Ti to PDE (6) A. Permanent Magnet Synchronous Motor (PMSM)
on S × [0, +∞), with Ti (·, 0) Lipschitz on S, are such that
T (·, t) defined by A first practical example which falls directly into the scope
of this paper is the Luenberger observer presented in [18] for
T (x, t) = (T1 (x, t), . . . , Ti (x, t), . . . , Tp (x, t)) (10) a PMSM. We reproduce here the minimal information needed
for comprehension, and we add the theoretical arguments
is injective on S for all t ≥ tk,u . More precisely, there exists which are not given in [18]. The system can be modeled by
a constant Lk such that for any (x1 , x2 ) in S 2 , any u in U
and any time t ≥ tk,u ẋ = u − Ri , y = |x − Li|2 − Φ2 (12)
2
|x1 − x2 | ≤ Lk |T (x1 , t) − T (x2 , t)| . where x is in R , the voltages u and currents i are inputs
in R2 , the resistance R, impedance L and flux Φ are known
Besides, for any t ≥ tk,u , T (·, t) is an injective immersion scalar parameters and the measurement y is constantly zero.
on S. Here p = 1, so we can drop the subscript i. Since the
Proof: See Appendix III. dynamics are linear and the measurement quadratic in x,
The additional assumption ”Ti (·, 0) Lipschitz on S” is not one can look for Tλ of the form :
very restrictive because the solution Ti can usually be chosen
arbitrarily at initial time 0 (see examples in Section V). In Tλ (x, t) = |x|2 + aλ (t) x + bλ (t)
particular, the elementary solutions Ti0 found in Lemma 1 where the dynamics of aλ and bλ are to be chosen so that
are zero at time 0 and thus clearly verifies this assumption. Tλ is solution of PDE (11). We can check that the dynamics
Applying successively Lemma 1, Theorem 2 and Theorem
1, we conclude that under Assumptions 1, 2, 3, it is possible ȧλ = −λ aλ − 2(u − Ri) + 2Li
to write an observer for system (1) by choosing any (Ai , Bi )
controllable and k sufficiently large. ḃλ = −λ bλ − a 2 2
λ (u − Ri) + L |i| − Φ
2
(13)
Remark 1: It is important to note that k does not depen- make Tλ follow the dynamics
dent on u, thanks to the fact that LH , Mf and Li given by
Assumptions 1-2-3 are the same for all u in U . However, żλ = −λzλ + y = −λzλ
the time tk,u after which the solution becomes injective a
priori depends on k and u. This is not a problem in practice and a trivial solution is thus zλ = 0. Let us now check
˙
since we only want to be sure that for k sufficiently large, whether the assumptions are verified. We suppose that i, i ,
any solution will become injective after a certain time. If we ¨
want this time tk,u to be uniform in u, the Lipschitz constants i and u, u̇ are bounded, so that the state x also remains
of Hi (·, ūm (0))) and of Ti (·, 0) must be independent from bounded (since y = 0). Choosing m = 3, we have
u in U . ⎛ ⎞
˙ ¨ |x − Li|2 − Φ2
Remark 2: If we choose m = maxi mi sufficiently large
distinct strictly positive real3 numbers λj , and take Ai = H(x, u, i, u̇, i , i ) = ⎝ 2η  (x − Li) ⎠
 
−diag(λ1 , ..., λmi ) and Bi = col(1, ..., 1), then, the only 2η̇ (x − Li) + 2η η
PDEs to solve are ˙
∂Tλ,i ∂Tλ,i where we denote η = u − Ri + L i . Thus, if we suppose
(x, t)f (x, u(t)) + (x, t) = −λ Tλ,i (x, t) besides that there exists c > 0 such that the inputs verify
∂x ∂t | det(η, η̇)| ≥ c, every assumption of Lemma 2 is satisfied. In
+ hi (x, u(t)) (11)
fact, the inputs happen to be such that5 det(η , η̇) = w3 Φ2 ,
for λ in {λ1 , . . . , λp } and for each 1 ≤ i ≤ p. where ω is the rotor angular velocity. We conclude that all the
Remark 3: Under Assumption 2-3, we could also build a conditions are verified when the inputs and their derivatives
high gain observer of dimension d = p × maxi mi because are bounded and the rotor angular velocity is away from zero.
the assumption ACP(d) (see [17]) is verified. If we wanted Applying our theorem, it follows that for any three distinct
to use only mi derivatives for each input and obtain an and sufficiently large strictly positive λj , the function
observer of dimension i mi , each mi -th derivative would T (x, t) = (Tλ1 (x, t), Tλ2 (x, t), Tλ3 (x, t))
have to satisfy an additional triangularity assumption. But in
any case, the crucial difference with the Luenberger observer 4 H(·, ν̄ is injective on S and ∂H
m) (x, ν̄m ) is full-rank for any x in S
  ∂x
5η − sin θ
3 It is also possible to take λj complex with strictly positive real part. = Φω cos θ , with θ the motor angle, and ω = θ̇.

3678
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on February 17,2024 at 20:25:02 UTC from IEEE Xplore. Restrictions apply.
becomes injective after a certain time (once the filters (13) VI. C ONCLUSION
have sufficiently converged). Implementing (13) for each λj , We have shown how a Luenberger methodology could
one can obtain after a certain time an estimate x̂ of x(t) for be applied on nonlinear controlled systems. It is based
instance by : on the resolution of a PDE, the solutions of which exist,
 −1   transform the system into a linear asymptotically stable one,
aλ1 (t) − aλ3 (t) bλ1 (t) − bλ3 (t) and become injective after a certain time. This is done under
x̂(t) = − bλ2 (t) − bλ3 (t) .
aλ2 (t) − aλ3 (t) the quite strong assumptions that :
- the function made of the output and a certain number
Note that for this system, a classical gradient observer of its derivatives is Lipschitz-injective (Assumption 2)
of smaller dimension exists ([19], [20]). The Luenberger - the set S is forward invariant.
observer proposed here offers the advantage of depending A possible improvement would be to relax those conditions
only on filtered versions of u and i, which can be useful maybe using tools from [5] for autonomous systems.
in presence of significant noise. On the other hand, no high
gain design would have been possible for this system without A PPENDIX I
computing the derivatives of i, which is not desirable in P ROOF OF T HEOREM 1
practice. For any t ≥ t, T (·, t) is injective on S, thus there exists
a function Tt−1 : T (S, t) → S such that for all x in S,
B. Non-holomic vehicle
Tt−1 (T (x, t)) = x. Besides, for all (z1 , z2 ) in T (S, t)2 ,
Another appropriate example is the celebrated non-
holomic vehicle with dynamics |Tt−1 (z1 ) − Tt−1 (z2 )| ≤ ρT (|z1 − z2 |) . (17)
ẋ1 = u1 cos(x3 ) , ẋ2 = u1 sin(x3 ) , ẋ3 = u1 u2 , According to [21, Theorem 2], there exists an extension6 of
Tt−1 on Rd verifying (17) for all (z1 , z2 ) in Rd ×Rd . Defining
y = (x1 , x2 ) (14) T ∗ as T ∗ (z, t) = Tt−1 (z) for (z, t) in Rd ×[0, +∞), we have
with u1 and u2 inputs corresponding to the average wheels’ for all t ≥ tk,u and all (x, z) in S × Rd ,
velocity and their orientation respectively. A large literature
is available on this system and our goal here is not to propose |x − T ∗ (z, t)| ≤ ρT (|T (x, t) − z|) .
yet another observer, but to show on another example how to Besides, for all s in [0, +∞),
solve PDE (11) for each component of the measurement. The ˙
dynamics and measurements being linear in x1 , x2 , cos(x3 ), T (X(x0 , 0; s; u), s) = AT (X(x0 , 0; s; u), s)
sin(x3 ), it is quite natural to look for a function T linear in +Bh(X(x0 , 0; s; u), u(s))
those quantities. Besides, x1 and x2 are independent so we
look for Tλ,1 and Tλ,2 , associated to measurement x1 and according to PDE (3). Thus,
x2 respectively, of the form : ˙
T (X(x0 , 0; s; u), s) − Z(x0 , z0 , 0; s; u)
Tλ,1 (x, t) = aλ (t) x1 + bλ (t) cos(x3 ) + cλ (t) sin(x3 ) = A (T (X(x0 , 0; s; u), s) − Z(x0 , z0 , 0; s; u))
Tλ,2 (x, t) = ãλ (t) x2 + b̃λ (t) cos(x3 ) + c̃λ (t) sin(x3 ) . and since A is Hurwitz,
By straightforward computations, we conclude that to satisfy lim T (X(x0 , 0; s; u), s) − Z(x0 , z0 , 0; s; u) = 0 .
t→+∞
PDE (11), we can take :
1 Besides, X(x0 , 0; s; u), s) is in S for all s. Therefore, for all
ãλ = aλ = , b̃λ = −cλ , c̃λ = dλ s ≥ t,
λ 1
ḃλ = −λ bλ − u1 u2 cλ − u1 |X(x0 , 0; s; u) − T ∗ (Z(x0 , z0 , 0; s; u), s)|
λ
ċλ = −λ cλ + u1 u2 bλ . (15) ≤ ρT (|T (X(x0 , 0; s; u), s) − Z(x0 , z0 , 0; s; u)|)
Then, Tλ,1 and Tλ,2 are solutions of and we have the required result.
A PPENDIX II
żλ,1 = −λ zλ,1 + x1 P ROOF OF L EMMA 1
żλ,2 = −λ zλ,2 + x2 (16) First, for any i in {1, . . . , p}, any u in U , and any s in
respectively. Besides, computing the successive derivatives [0, +∞), (x, t) → Yi (x, t; s; u) = h(X(x, t; s; u), s) is C 1 ,
of the measurements (x1 , x2 ), we can see that we have an thus Ti0 is C 1 .
injective immersion as soon as at least one derivative of u1 Take x in S and t in [0, +∞). For any τ in R,
is nonzero. Therefore, if the state, the inputs u1 , u2 and X(X(x, t, t + τ ; u), t + τ ; s; u) = X(x, t; s; u).
their derivatives remain in compact sets, and if there exist
an integer m ≥ 2 and a real number c > 0 such that for all Therefore,
t and all considered input u, Ti0 (X(x, t; t + τ ; u), t + τ )
(m−2)
 t+τ
u1 (t)2 + u̇1 (t)2 + ... + u1 (t)2 ≥ c , = ekAi (t+τ −s) Bi hi (X(x, t, s; u), u(s))ds
0
then all the assumptions in Lemma 2 are verified with m1 =
m2 = m. Therefore, by choosing m strictly positive distinct = ekAi τ Ti0 (x, t)
 t+τ
real numbers λj , the function
+ekAi τ ekAi (t−s) Bi hi (X(x, t; s; u), u(s))ds
T (x, t) = (Tλ1 ,1 (x, t), ..., Tλm ,1 (x, t), t
Tλ1 ,2 (x, t), ..., Tλm ,2 (x, t)) −1
6 Denoting Tt,j the jth component of Tt−1 , take Tt,j
−1
(z) =
becomes injective after a certain time. Implementing (15) − −1
minz̃∈T (S,t) {Tt,j (z̃) + ρT (|z̃ − z|)} −1
or equivalently Tt,j (z) =
(16) for each λj , we thus get an observer of dimension 4m. minx∈S {xj + ρT (|T (x, t) − z|)}

3679
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on February 17,2024 at 20:25:02 UTC from IEEE Xplore. Restrictions apply.
and Thanks to Assumption 1, for any (x1 , x2 ) in S, and (t, τ )
in [0, +∞)2 , we have (see for instance [22])
Ti0 (X(x, t; t+ τ ; u), t + τ ) − Ti0 (x, t)
τ |X(x1 , t; τ ; u) − X(x2 , t; τ ; u)| ≤ eMf |τ −t| |x1 − x2 | . (18)
ekAi τ − I 0
= Ti (x, t) By assumption Ti (·, 0) and Hi (·, ūm (0))) are Lipschitz on
τ  S, thus there exists LΨi such that
ekAi τ t+τ kAi (t−s)
+ e Bi hi (X(x, t, s; u), u(s))ds .|Ψi (X(x1 , t; 0; u), 0) − Ψi (X(x1 , t; 0; u), 0)|
τ t
≤ LΨi eMf t |x1 − x2 | .
Making τ tend to 0, we get PDE (6).
Then, Ai being Hurwitz, there exists strictly positive num-
A PPENDIX III bers ai and γi (see [22]) such that for all τ in [0, t]
P ROOF OF T HEOREM 2
Take u in U , i in {1, . . . , p}, x in S and t in [0, +∞). ekAi (t−s) ≤ γi e−kai (t−s) . (19)
According to PDE (6), Ti satisfies for all s in [0, +∞),
Using Assumption 3 and inequalities (18) and (19), we
d M
deduce that if k > aif ,
Ti (X(x, t; s; u), s) = kAi Ti (X(x, t; s; u), s)
ds
+Bi Yi (x, t; s; u) . |Ri (x1 , t)−Ri (x2 , t)|
 t
Integrating between t and s, it follows that
≤ Li |Bi |γi e−(kai −Mf )(t−τ ) dτ |x1 − x2 |
kAi (s−t) 0
Ti (X(x, t; s; u), s) = e Ti (X(x, t; t; u), t)
  Li |Bi |γi
Ti (x,t) ≤ |x1 − x2 | .
 s kai − Mf
+ ekAi (s−τ ) Bi Yi (x, t; τ ; u)dτ
t We finally deduce that
and thus, |Ti (x1 , t) − Ti (x2 , t)| 
Ti (x, t) = ekAi (t−s) Ti (X(x,t; s; u), s) ≥ |A−m
i
i
Ci K i | |ΔHi |
t  
kAi (t−τ ) 1
+ e BYi (x, t; τ ; u)dτ . −|Ki−1 Ci−1 | ekAi t |ΔΨi | + mi |ΔRi |
s
 k
applying this inequality at s = 0, we get |A−m i
C i |
≥ i mi |ΔHi |
k
Ti (x, t) = ekAi t Ti (X(x, t; 0; u), 0) + Ti0 (x, t)
−k |Ci−1 | γi LΨi e−(kai −Mf )t |x1 − x2 | 
mi

with Ti0 defined in (5). Li |Bi |

−|Ci−1 | γi |x1 − x2 | .
But after mi successive integration by parts in (5), we get, kai − Mf
Ti0 (x, t) = −A−m
i
i
Ci Ki Hi (x, ūm (t)) where ΔHi , ΔΨi and ΔRi denote the difference between
+A−m
i
i kAi t
e Ci Ki Hi (X(x, t; 0; u), ūm (0)) x1 and x2 of the functions Hi (·, ūm (t)), Ψi (X(·, t; 0; u), 0)
1 and Ri (·, t) respectively. It follows (by norm equivalence),
+ mi A−mi
i
Ri (x, t) that there exists a constant c such that
k
  |T (x1 , t) − T (x2 , t)|
where Ki = diag k1 , . . . , km 1
, Ci is the invertible 
controllability matrix
i
mini (|A−m i
Ci |)
≥c i
m
|H(x1 , ūm (t)) − H(x2 , ūm (t))|
i −1
k p 
Ci = [Am
i B i , . . . , Ai B i , Bi ] , 
−1
− k γi |Ci | LΨi e−(ka−Mf )t |x1 − x2 |
mi
Hi (x, ν̄m ) is defined in (8), and Ri is the remainder : i=1  p  
 t −1
i=1 Li γi |Ci ||Bi |
Ri (x, t) = ekAi (t−τ ) Bi Lm i
hi (X(x, t; τ ; u), ūm (τ )) dτ − |x1 − x2 |
0
f¯ ka − Mf
mini (|A−m i
Ci |)
We finally deduce that ≥c i
|x1 − x2 |×
k m 
Ti (x, t) = A−m i
Ci K i × 1 m −(ka−Mf )t 1

i × − c1 k e − c2
LH ka − Mf
× −Hi (x, ūm (t)) + Ki−1 Ci−1 ×
  where m a, c1 , c2 are constants independent from k and t
1 defined by
× ekAi t Ψi (X(x, t; 0; u), 0) + mi Ri (x, t)
k
m = max mi , a = min ai
i i
with Ψi (x, t) = Am i Ti (x, t) + Ci Ki Hi (x, ūm (t)) .
i

Let us now consider x1 and x2 in S, and t in [0, +∞). 

p 
p
We are interested in the quantity |T (x1 , t) − T (x2 , t)|, and c1 = γi |Ci−1 | LΨi , c2 = Li γi |Ci−1 ||Bi | .
thus in |Ti (x1 , t) − Ti (x2 , t)|. i=1 i=1

3680
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on February 17,2024 at 20:25:02 UTC from IEEE Xplore. Restrictions apply.
We deduce that for which is a closed subset of the compact set S 2 and therefore
1 ln(4k m c1 LH ) compact. The function (x1 , x2 , ν̄m ) → |H(x1 ,ν̄|x 1 −x2 |
m )−H(x2 ,ν̄m )|
k ≥ (Mf + 4c2 LH ) , t≥ , is defined and continuous on Ω × U m since H(·, ν̄m ) is
a ka − Mf
injective for any ν̄m in U m . Thus, it admits a maximum
we have M on the compact set Ω × U m .
mini (|A−m i
Ci |) 1 Finally, take any (x1 , x2 ) in S 2 and any ν̄m in U m . There
|T (x1 , t) − T (x2 , t)| ≥ c i
m
|x1 − x2 | are two cases :
k 2LH
• either (x1 , x2 ) ∈
/ Ω, i-e |x1 − x2 | < 2r, and
i-e
Pm
km |x1 − x2 | ≤ |H(x1 , ν̄m ) − H(x2 , ν̄m )| .
|x1 − x2 | ≤ 2LH |T (x1 , t) − T (x2 , t)| 2
c mini (|A−m
i
i
Ci |) • or (x1 , x2 ) ∈ Ω, and
(20)
and Tk (·, t) is injective on S, uniformly in time. We conclude |x1 − x2 | ≤ M |H(x1 , ν̄m ) − H(x2 , ν̄m )| .
that the result holds with
We conclude that Assumptions 2 is verified with LH =
1 km max{ P2m , M }.
k = (Mf + 4c2 LH ) , Lk = 2LH
a c mini (|A−m
i
i
Ci |) R EFERENCES
 m
 [1] D. Luenberger, “Observing the state of a linear system,” IEEE Trans-
ln(4k c1 LH ) actions on Military Electronics, vol. 8, pp. 74–80, 1964.
tk,u = max ,0
ka − Mf [2] A. Shoshitaishvili, “On control branching systems with degenerate lin-
earization,” IFAC Symposium on Nonlinear Control Systems, pp. 495–
Since Mf , LH and Li (and thus c2 ) are independent from 500, 1992.
[3] N. Kazantzis and C. Kravaris, “Nonlinear observer design using
u, k and Lk are the same for all u in U , while tk,u depends Lyapunov’s auxiliary theorem,” Systems and Control Letters, vol. 34,
on u through LΨi . pp. 241–247, 1998.
[4] A. Krener and M. Xiao, “Nonlinear observer design in the siegel
Now, take any x in S and t ≥ tk,u . For any v and any h domain,” SIAM Journal on Control and Optimization, vol. 41, no. 3,
such that x + hv is in S, we have pp. 932–953, 2003.
[5] V. Andrieu and L. Praly, “On the existence of a Kazantzis–Kravaris
|T (x + hv, t) − T (x, t)| / Luenberger observer,” SIAM Journal on Control and Optimization,
Lk |v| ≤ vol. 45, no. 2, pp. 432–456, 2006.
|h| [6] A. Krener and M. Xiao, “Nonlinear observer design for smooth
systems,” Chaos in Automatic Control, W. Perruguetti and J.-P. Barbot,
and by letting h go to 0, we get Lk |v| ≤ ∂T
∂x (x, t)v . Hence, Eds., Taylor and Francis, pp. 411–422, 2006.
[7] G. Kreisselmeier and R. Engel, “Nonlinear observers for autonomous
T (·, t) is an immersion on S. lipshitz continues systems,” IEEE Transactions on Automatic Control,
vol. 48, no. 3, pp. 451–464, 2003.
A PPENDIX IV [8] F. Rotella and I. Zambettakis, “On functional observers for linear time-
varying systems,” IEEE Transactions on Automatic Control, vol. 58,
P ROOF OF L EMMA 2 no. 5, 2013.
First, S and U m being compact, Assumptions 1 and 3 [9] J. Trumpf, “Observers for linear time-varying systems,” Linear Alge-
bra and its Applications, vol. 425, pp. 303–312, 2007.
are satisfied. Besides, Hi (·, ūm (0)) and Ti (·, 0) are clearly [10] Y. Hamami, “Observateur de Kazantzis-Kravaris/luenberger dans le
Lipschitz on S. The only thing to prove is the uniform cas d’un système instationnaire,” tech. rep., MINES ParisTech, 2008.
Lipschitz-injectivity of H. The following proof generalizes [11] A. H. Jazwinski, Stochastic processes and filtering theory. Academic
Press, 1970.
that of [23, Lemma 3.2]. Let Δ be the function defined on [12] D. Bossane, D. Rakotopara, and J. P. Gauthier, “Local and global
S × S × U m by immersion into linear systems up to output injection,” Conference on
Decision and Control, pp. 2000–2004, 1989.
Δ(x1 , x2 , ν̄m ) = H(x1 , ν̄m ) − H(x2 , ν̄m ) [13] P. Jouan, “Immersion of nonlinear systems into linear systems modulo
output injection,” SIAM Journal on Control and Optimization, vol. 41,
∂H no. 6, pp. 1756–1778, 2003.
− (x2 , ν̄m )(x1 − x2 ) . [14] J.-P. Gauthier and G. Bornard, “Observability for any u(t) of a class of
∂x nonlinear systems,” IEEE Transactions on Automatic Control, vol. 26,
∂H pp. 922 – 926, 1981.
Since ∂x (x, ν̄m ) is full-rank by assumption, the function [15] J.-P. Gauthier, H. Hammouri, and S. Othman, “A simple observer for
 −1 nonlinear systems application to bioreactors,” IEEE Transactions on
∂H ∂H ∂H Automatic Control, vol. 37, no. 6, pp. 875–880, 1992.
P (x, ν̄m ) = (x, ν̄m ) (x, ν̄m ) (x, ν̄m ) [16] G. Besançon and A. Ticlea, “An immersion-based observer design for
∂x ∂x ∂x rank-observable nonlinear systems,” IEEE Transactions on Automatic
Control, vol. 52, no. 1, pp. 83–88, 2007.
[17] J.-P. Gauthier and I. Kupka, Deterministic observation theory and
is well-defined and continuous on S × U m , and for any applications. Cambridge University Press, 2001.
(x1 , x2 , ν̄m ) in S × S × U m , we have [18] N. Henwood, J. Malaizé, and L. Praly, “A robust nonlinear luenberger
observer for the sensorless control of SM-PMSM : Rotor position
|x1 − x2 | ≤ Pm (|H(x1 , ν̄m ) − H(x2 , ν̄m | + |Δ(x1 , x2 , ν̄m )|) and magnets flux estimation,” IECON Conference on IEEE Industrial
Electronics Society, 2012.
[19] J. Lee, J. Hong, K. Nam, R. Ortega, L. Praly, and A. Astolfi,
with Pm = maxS×U m |P (x, ν̄m )|. Besides, the function “Sensorless control of surface-mount permanent-magnet synchronous
|Δ(x1 ,x2 ,ν̄m )|
|x1 −x2 |2 is defined and continuous on S × S × U m , motors based on a nonlinear observer,” IEEE Transactions on Power
Electronics, vol. 25, no. 2, pp. 290–297, 2010.
thus there exists LΔ > 0 such that [20] J. Malaizé, L. Praly, and N. Henwood, “Globally convergent nonlinear
observer for the sensorless control of surface-mount permanent magnet
1 synchronous machines,” IEEE Conference on Decision and Control,
|Δ(x1 , x2 , ν̄m )| ≤ LΔ |x1 − x2 |2 ≤ |x1 − x2 | 2012.
2Pm [21] E. J. McShane, “Extension of range of functions,” Bull. Amer. Math.
Soc., vol. 40, no. 12, pp. 837–842, 1934.
for any (x1 , x2 ) in S 2 such that |x1 − x2 | ≤ 2r with r = [22] A. M. R. Miller, Ordinary Differential Equations. Academic Press,
1 1982.
4Pm LΔ , and for any ν̄m in U m . Now, define the set [23] V. Andrieu, “Convergence speed of nonlinear Luenberger observers,”
SIAM Journal on Control and Optimization, vol. 52, no. 5, pp. 2831–
Ω = {(x1 , x2 ) ∈ S 2 | |x1 − x2 | ≥ 2r} 2856, 2014.

3681
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on February 17,2024 at 20:25:02 UTC from IEEE Xplore. Restrictions apply.