5

Functional Unknown Input Observers: Further

Results and Applications

5.1 Introduction
In Section 4.4 of Chapter 4, the design of reduced-order scalar functional observers
for linear systems with unknown inputs has been discussed. The existence condi-
tions have been presented in Proposition 4.3. It has been shown that the design
of scalar functional observers for linear systems with unknown inputs is related to
solving the following three coupled matrix equations

NL2 − L1 A12 − L2 A22 = 0, N is Hurwitz (5.1)

F2 = DL2 (5.2)
L1W1 + L2W2 = 0 (5.3)

where F2 ∈ R1×(n−p) is a row matrix, W1 ∈ R p×l and W2 ∈ R(n−p)×l are known

constant matrices and the pair (A12 , A22 ) is observable. Matrices L1 ∈ Rq×p, L2 ∈
Rq×(n−p), D ∈ R1×q and N ∈ Rq×q (N must be a stable matrix) are unknown. A
solution method based on the parametric approach has been presented in Chapter 4
for solving the above three coupled matrix equations.
This chapter considers multiple functions (i.e., F2 ∈ Rr×(n−p), r ≥ 1) and dis-
cusses another approach [129] for solving the three coupled matrix equations (5.1)-
(5.3). Before that, this chapter first points out that there is indeed a clear and strong
connection between the design of functional observers for linear systems with un-
known inputs and the design of delay-free functional observers for time-delay
systems. As well, there is a strong connection with the design of decentralized
observers/controllers for systems under decentralized information structure and/or
interconnected systems. For this, one will show that the existence conditions of
functional observers for these different types of systems are the same as those con-
ditions presented in Proposition 4.3. This implies that the design of linear func-
tional observers for these systems can be done under the general framework of linear
functional unknown input observers. As a result, the design of linear functional ob-
servers for these systems boils down to solving the three coupled matrix equations
(5.1)-(5.3). This chapter will also present an application of the theory of functional

H. Trinh and T. Fernando: Functional Observers for Dynamical Systems, LNCIS 420, pp. 101–126.
springerlink.com c Springer-Verlag Berlin Heidelberg 2012
102 5 Functional Unknown Input Observers: Further Results and Applications

unknown input observers to the design of a common disturbance decoupled observer

for two linear systems with unknown inputs.

5.2 Delay-Free Functional Observers for Time-Delay Systems

This section shows that there is a strong connection between the design of delay-free
functional observers for time-delay systems and the design of functional observers
for linear systems with unknown inputs. Consider the following class of time-delay
systems

ẋ(t) = Ax(t) + Ad x(t − τ (t)) + Bu(t), t ≥ 0 (5.4)

x(t) = φ (t), t ∈ [−τu , 0) (5.5)
y(t) = Cx(t) (5.6)
z(t) = Fx(t) (5.7)

where φ (t) is a continuous initial function, x(t) ∈ Rn is the state vector, u(t) ∈ Rm is
the control input vector, y(t) ∈ R p is the measurement output vector and z(t) ∈ Rr
is the vector to be estimated. Matrices A ∈ Rn×n , Ad ∈ Rn×n , B ∈ Rn×m , C ∈ R p×n
and F ∈ Rr×n are known real constant. The delay, τ (t), is a time-varying delay in
the range of (0, τu ] where τu > 0. Here, the real-time knowledge of the delay is not
required nor its rate of change is assumed to be bounded. Thus the time-delay can be
a fast time-varying function
  and not necessary to be differentiable. It is assumed that
C
rank(C) = p and rank = (p + r) and also there are no self-convergent states in
F
the system because the asymptotic value of such states are zero, those states play no
part in an asymptotic estimate of z(t), and if present, those states can be removed
from the system and the function z(t).
In the literature, the problem of designing full-order and/or reduced-order ob-
servers to asymptotically estimate the entire state vector, x(t), or a linear function
of the state vector, z(t), has been a subject of extensive studies (see, for exam-
ple, [19, 34, 74, 93, 119, 121, 129] and the references therein). This section discusses
the design of delay-free reduced-order observers to asymptotically estimate z(t).
Consider the following reduced-order, delay-free functional observer for the class
of time-delay systems (5.4)-(5.6)

ẑ(t) = Dw(t) + Ey(t) (5.8)

ẇ(t) = Nw(t) + Jy(t) + LBu(t) (5.9)

where w(t) ∈ Rq is the observer state vector and ẑ(t) denotes the estimate of z(t).
Here, the order of the observer is reduced (i.e., r ≤ q ≤ (n − p)), matrix D =
[Ir 0] ∈ Rr×q and the design objective is to determine matrices E ∈ Rr×p , N ∈ Rq×q
(N is Hurwitz), J ∈ Rq×p and L ∈ Rq×n such that ẑ(t) converges, with any pre-
scribed asymptotic convergence rate, to z(t). Clearly, delay-free and reduced-order
observers of the structure (5.8)-(5.9) are very easy to implement and cost effective.
5.2 Delay-Free Functional Observers for Time-Delay Systems 103

In the following, the necessary and sufficient conditions for the existence of delay-
free observers (5.8)-(5.9) for the time-delay systems (5.4)-(5.6) will be presented.
It will be shown that the existence conditions are the same as those conditions pre-
sented in Proposition 4.3. As a result, the determination of all the observer parame-
ters can be done by solving the three coupled matrix equations (5.1)-(5.3).
Let L ∈ Rq×n be a full-row rank matrix and define the error vectors ε (t) ∈ Rq and
e(t) ∈ Rr to be

ε (t)  w(t) − Lx(t) (5.10)

e(t)  ẑ(t) − z(t). (5.11)

The following theorem provides the necessary and sufficient conditions for en-
suring that ẑ(t) converges asymptotically to z(t).
Theorem 5.1.
For the proposed delay-free observer (5.8)-(5.9), the estimate ẑ(t) will converge
asymptotically to z(t) for any initial condition φ (t), w(0) and any u(t) if and only if

N is Hurwitz (5.12)
NL + JC − LA = 0 (5.13)
LAd = 0 (5.14)
F − DL − EC = 0. (5.15)

Proof: (Sufficiency) From (5.10), the following error dynamics equation is obtained

ε̇ (t) = ẇ(t) − Lẋ(t)

= N ε (t) + (NL + JC − LA)x(t) − LAd x(t − τ (t)), t ≥ 0 (5.16)
ε (t) = −Lφ (t), t ∈ [−τu , 0). (5.17)

From (5.11), the error vector e(t) can be expressed as

e(t) = Dε (t) + (DL + EC − F)x(t). (5.18)

From (5.16)-(5.17), ε (t) → 0 as t → ∞ if (5.12)-(5.14) of Theorem 5.1 are satisfied.

Now, if condition (5.15) of Theorem 5.1 is satisfied, then from (5.18), one has e(t) =
Dε (t). Since ε (t) → 0 as t → ∞, it follows that e(t) → 0 as t → ∞ and hence ẑ(t) →
z(t) as t → ∞.
Necessity: If N is not Hurwitz, then there exists an initial function φ (t) such that
ε (t) = 0, t ∈ [−τu , 0), which makes ε (t)  0 and also e(t)  0 as t → ∞. If
any of the conditions (5.13), (5.14) and (5.15) is not satisfied, then one can find
a u(t) to generate a x(t) to make e(t)  0 as t → ∞. This completes the proof of
Theorem 5.1.
Now, let l be an integer number such that l  rank(Ad ), l ≤ n. By using the
singular value decomposition (SVD), one can always find a full-column rank matrix
W ∈ Rn×l and a full-row rank matrix U ∈ Rl×n such that
104 5 Functional Unknown Input Observers: Further Results and Applications

Ad = WU. (5.19)

Substituting (5.19) into (5.14) gives LWU = 0. Since U is a full-row rank matrix,
it follows that LWU = 0 if and only if

LW = 0. (5.20)

It is now clear that the existence conditions for the delay-free functional observers
are the same as those existence conditions presented in Proposition 4.3 for functional
observers of linear systems with unknown inputs. By using the partitioned matrices
as defined in (4.48), one therefore arrives at the three coupled matrix equations (5.1)-
(5.3). Thus the design of delay-free functional observers for time-delay systems can
be accomplished by solving the matrix equations (5.1)-(5.3). Also, the remaining
observer matrices E and J are determined from (4.49) and (4.50), respectively.

5.3 Observer-Based Control of Multi-Agent Systems under

Decentralized Information Structure
This section shows a link between the design of functional unknown input observers
and the problem of decentralized implementation of a global state feedback con-
troller for multi-agent systems under the constraint of a complete decentralized in-
formation structure [45]. With regard to the later, the task of effectively controlling
multi-agent systems in the absence of communication channels is an interesting
control problem. Centralized control of these systems requires the establishment
of an extensive communication network for the transfer of information between
each subsystem and the central station. This may impose cost and implementation
problems which may prevent the realization of centralized controllers. Moreover,
systems using explicit communication channels have, in most of the cases, a lim-
ited number of interconnected agents due to the communication requirements of
the task set and the bandwidth of the channels. It is therefore desirable to devise
a control strategy for such systems where the processing of information and the
control task are shared among the decentralized controllers, provided that the over-
all system stability and performance are preserved. Obvious advantages of such
control strategies are considerable savings in the cost of the information transfer
network, and also the amenability of controller realization against system com-
plexity. A considerable amount of research work has appeared in the control lit-
erature in an attempt to resolve the problems of decentralized control (see, for ex-
ample, [3, 56, 61, 99, 104, 106, 110, 120, 137]).
This section discusses the problem of decentralized implementation of a global
state feedback controller for multi-agent systems. The system is assumed to be un-
der the constraint of a complete decentralized information structure, where no flow
of information among the control stations takes place. Thus, for a control system
consisting of N local control stations, the control input for the ith station is calcu-
lated from the information contained in its local input and output signals only. Here,
5.3 Observer-Based Control of Multi-Agent Systems 105

the global controller is obtained by using existing design methods and its implemen-
tation is accomplished by using N decentralized reduced-order linear functional ob-
servers to generate the required corresponding control signal for each local control
station.
Consider a linear system comprising N control stations [45] as described by
N
ẋ(t) = Ax(t) + ∑ Bi ui (t) (5.21)
i=1
yi (t) = Ci x(t); i = 1, 2, ..., N (5.22)

where x(t) ∈ Rn is the global state vector, ui (t) ∈ Rmi and yi (t) ∈ R pi (i = 1, 2, ..., N)
are, respectively, the input and output vectors of the ith control station. It is assumed
that the pair (Ci , A) is observable, matrices A ∈ Rn×n , Bi ∈ Rn×mi and Ci ∈ R pi ×n are
known real constants. It is assumed that a satisfactory global state feedback control
law has been found of the form

u(t) = Fx(t) (5.23)

where F ∈ Rm×n by using any standard state feedback control method to obtain the
satisfaction of some system performance index. Thus the control signal at the ith
control station, ui (t), can be written as

ui (t) = Fi x(t); i = 1, 2, ..., N (5.24)

where Fi ∈ Rmi ×n is a known constant matrix ⎡ corresponding

⎤ to the ith block-row of
F1
⎢ F2 ⎥
⎢ ⎥
a given global state-feedback matrix, F = ⎢ . ⎥, ∑Ni=1 mi = m.
⎣ .. ⎦
FN
Let Ii (t) denotes the information available to the ith control station. Under the
constraint of no communication between control stations, the information available
to the ith control station includes only the local output and control of the ith station:

Ii (t) = {yi (t), ui (t)}; i = 1, 2, ..., N. (5.25)

Taking into account the constraint of the decentralized information structure

(5.25), the objective is to design N decentralized controllers of the form

ui (t) = fi {Ii (t),t}; i = 1, 2, ..., N (5.26)

using only information available to local control stations, that is Ii (t), such that
the multi-agent system (5.21)-(5.22) is stable with satisfactory performance as pre-
scribed in the global control law (5.23). Note that (5.26) represents a decentralized
control law for large-scale systems under a decentralized information structure. To
satisfy the information constraint, one seeks to reconstruct the global control (5.23)
dynamically via decentralized linear functional observers that receive only Ii (t) as
106 5 Functional Unknown Input Observers: Further Results and Applications

their inputs. These observers are designed such that the resulting local controls will
reproduce asymptotically the global law (5.23).
Thus, the aim is to implement the global controller in a decentralized manner by
using N decentralized dynamic controllers of the following form

ui (t) = Fi x(t) = (Di Li + EiCi )x(t) = Di wi (t) + Eiyi (t) (5.27)

ẇi (t) = Ni wi (t) + Ji yi (t) + Li Bi ui (t); i = 1, 2, ..., N (5.28)

where in (5.27), matrix Fi is decomposed according to Fi = Di Li + EiCi and wi =

Li x(t) ∈ Rqi is the state vector of the dynamic system (5.28) with a reduced order,
where mi ≤ qi ≤ (n − pi ). Matrices Di , Ei , Ni , Ji and Li are constant matrices of
appropriate dimensions to be determined. The significance of the controller structure
(5.27)-(5.28) is that ui (t) can now be implemented by using the local output yi (t)
and a linear combination of global state wi (t) = Li x(t), which can be generated from
information (5.26), available at the ith control station only.
Let the global control input matrix B ∈ Rn×m be partitioned as
 
B = Bi BR (5.29)

where BR ∈ Rn×(m−mi ) . Accordingly, (5.21)-(5.22) can be expressed as

ẋ(t) = Ax(t) + Bi ui (t) + BRuR (t) (5.30)

yi (t) = Ci x(t); i = 1, 2, ..., N (5.31)

where uR (t) contains (N − 1) input vectors of the remaining (N − 1) remote control

stations.
By following similar lines as in Section 5.2, let Li ∈ Rqi ×n be a full-row rank
matrix and define the error vector εi (t) ∈ Rqi to be

εi (t)  wi (t) − Li x(t) (5.32)

then one can arrive at the following existence conditions for ensuring that εi (t) → 0:

Ni is Hurwitz (5.33)
Ni Li + JiCi − Li A = 0 (5.34)
Li BR = 0. (5.35)

The above matrix equations, together with the matrix equation Fi = Di Li + EiCi ,
are of the same form as those existence conditions presented in Proposition 4.3.
Essentially, the task of implementing a global controller (5.23) in a decentralized
manner (5.27)-(5.28) is boiled down to solving the three coupled matrix equations
(5.1)-(5.3).
It is also shown in the following that the desired separation property of the over-
all closed-loop system under the decentralized dynamic controller can be verified.
Using (5.32), the control signal, ui (t), defined in (5.27) can be expressed as

ui (t) = Di εi (t) + (Di Li + EiCi )x(t) = Di εi (t) + Fix(t). (5.36)

5.4 Multiple Functional Unknown Input Observers 107

Accordingly, ∑Ni=1 Bi ui (t) is written as

N N N
∑ Bi ui(t) = ∑ Bi Di εi (t) + ∑ Bi Fix(t) = BFx(t) + BDε (t) (5.37)
i=1 i=1 i=1
⎡ ⎤
ε1 (t)
⎢ ε2 (t) ⎥
⎢ ⎥
where D = block − diag(Di ), i = 1, 2, ..., N and ε (t) = ⎢ . ⎥. Hence, the follow-
⎣ .. ⎦
εN (t)
ing augmented closed-loop system is obtained
    
ẋ(t) (A + BF) BD x(t)
= (5.38)
ε̇ (t) 0 N ε (t)

where N = block − diag(Ni ), i = 1, 2, ..., N. Equation (5.38) implies that the separa-
tion principle holds.

5.4 Multiple Functional Unknown Input Observers

This section discusses a solution method for solving the three coupled matrix equa-
tions (5.1)-(5.3) for multiple functions, i.e., r ≥ 1.
Case 1 (q = r): Here, one first starts with r-th order observer (i.e., q = r). This
gives the lowest possible order observer and hence the most simple structure ob-
server. Since q = r, without loss of generality, one can let D = Ir . Accordingly, from
(5.2), one has L2 = F2 and hence (5.1) and (5.3) can be expressed as

NF2 − L1 A12 − F2A22 = 0, N is Hurwitz (5.39)

L1W1 = −F2W2 . (5.40)

Lemma 5.2.
Equation (5.40) has a solution for L1 if and only if
 
W1
rank = rank (W1 ) . (5.41)
F2W2
The proof for Lemma 5.2 is trivial and thus omitted.
Remark 1-A necessary condition: Condition (5.41) is a necessary condition for the
existence of r-th order functional observers for linear systems with unknown inputs.
It is also the same as the necessary condition (4.63) presented in Theorem 4.5 by
using the parametric approach. Since F2 , W1 and W2 are known matrices, (5.41) can
be tested easily. If (5.41) is not satisfied, one can declare that there does not exist
any solution to the matrix equations (5.1)-(5.3). Clearly, when W1 is a full-column
rank matrix (i.e., rank(W1 ) = l), condition (5.41) is satisfied regardless of F2 . Even
when W1 is not a full-column rank matrix, (5.41) can still be satisfied if either:
(i) F2 belongs to an orthogonal basis of the null-space of W2 (i.e., F2W2 = 0);
108 5 Functional Unknown Input Observers: Further Results and Applications

or (ii) F2W2 = 0 is linearly dependent on W1 . Also, as explained in Remark 6

of Section 4.4.1, condition (5.41) is less conservative than the observer matching
condition (2.37).
If (5.41) is satisfied, one can proceed as follows. Let F2⊥ ∈ R(n−p)×(n−p−r) be an
+ (n−p)×r be the Moore-Penrose
 + of F⊥2 and F2 ∈ R
orthogonal basis for the null-space
inverse of F2 . It is clear that F2 F2 is a square and invertible matrix. Also,
let Ω ∈ R p×(n+l−p−r), Φ ∈ Rr×(n+l−p−r) , N1 ∈ Rr×r and N2 ∈ R p×r be defined
according to the following equations
 
Ω = A12 F2⊥ W1 (5.42)
 ⊥

Φ = − F2 A22 F2 F2W2 (5.43)
N1 = (ΦΩ + A12 + F2 A22 )F2+ (5.44)
N2 = (Ω Ω + − I p)A12 F2+ . (5.45)

The following Theorem provides necessary and sufficient conditions for the solution
to the matrix equations (5.39) and (5.40).
Theorem 5.3.
There exists a solution to the matrix equations (5.39) and (5.40) if and only if
 
Φ
rank = rank (Ω ) (5.46)
Ω
the pair (N2 , N1 ) is detectable. (5.47)
 
Proof: Post-multiplying (5.39) by F2+ F2⊥ to give

N = L1 A12 F2+ + F2 A22 F2+ (5.48)

L1 A12 F2⊥ = −F2 A22 F2⊥ . (5.49)

Now, (5.40) and (5.49) can be expressed as

L1 Ω = Φ (5.50)

where Ω and Φ are as defined in (5.42) and (5.43), respectively. It is clear that (5.50)
has a solution for L1 if and only if the condition (5.46) of Theorem 5.3 is satisfied.
Since both Ω and Φ are known matrices, testing the condition (5.46) is a trivial task.
Upon the satisfaction of (5.46), a solution for L1 is given by [97]

L1 = ΦΩ + + Z(I p − Ω Ω + ) (5.51)

where Ω + is the Moore-Penrose inverse of Ω and Z ∈ Rr×p is an arbitrary matrix.

Substituting (5.51) into (5.48) to get

N = N1 − ZN2 (5.52)

where N1 and N2 are as defined in (5.44) and (5.45), respectively. There exists a
stable matrix N if and only if the pair (N2 , N1 ) is detectable. Detectability implies
that the poles of the observer are stable but they can not be arbitrarily assigned.
5.4 Multiple Functional Unknown Input Observers 109

When (N2 , N1 ) is observable, then one can always find a Z to place all the poles of
N at any location. This completes the proof of Theorem 5.3.
Remark 2: One can show that conditions (5.46) and (5.47) of Theorem 5.3 are,
respectively, equivalent to the following conditions
⎛⎡ ⎤⎞
F2 A22 F2W2  
⎝ ⎣ ⎦ ⎠ A12 W1
rank A12 W1 = rank (5.53)
F2 0
F2 0
   
(sF2 − F2A22 ) −F2W2 A12 W1
rank = rank
A12 W1 F2 0
∀s ∈ C, ℜ(s) ≥ 0. (5.54)

Here, the equivalency of (5.46) and (5.53) can be shown by noting that the right-
hand side of (5.53) is the same as
   + ⊥ 
A12 W1 F2 F2 0
rank = r + rank(Ω ). (5.55)
F2 0 0 0 Il

Similarly, the left-hand side of (5.53) is the same as

⎛⎡ ⎤ ⎞
F2 A22 F2W2  + ⊥   
⎝ ⎣ ⎦ F F 0 ⎠ Φ
rank A12 W1 2 2 = r + rank . (5.56)
0 0 Il Ω
F2 0

To establish the equivalency of (5.47) and (5.54), note that the detectability of the
pair (N2 , N1 ), by definition, is
 
sIr − N1
rank = r, ∀s ∈ C, ℜ(s) ≥ 0. (5.57)
N2

It can also be shown that the left-hand side of (5.54) is the same as
 
sIr − N1
rank + rank(Ω ), ∀s ∈ C, ℜ(s) ≥ 0. (5.58)
N2

From (5.58) and (5.55), one can now conclude that the pair (N2 , N1 ) is detectable.
Given that the necessary condition (5.41) is satisfied, the task of solving the ma-
trix equations (5.1)-(5.3) for all the unknown matrices can be performed as follows.
First, condition (5.46) of Theorem 5.3 is tested. As this condition involves only
known, well-defined matrices, it can be easily tested. Once (5.46) is satisfied, the
next step is to check the detectability of the pair (N2 , N1 ). If it is detectable, then
a stable matrix N is obtained. Also, if the pair (N2 , N1 ) is observable, then all the
poles of the observer can be arbitrary assigned.

Design Algorithm
1. Compute Ω , Φ , N1 and N2 from (5.42)-(5.45).
110 5 Functional Unknown Input Observers: Further Results and Applications

2. Check (5.46) and (5.47) of Theorem 5.3. If either

condition is not satisfied: stop, as a solution does
not exist. Otherwise, go to the next step.
3. Determine Z such that N defined in (5.52) is stable.
4. Compute L1 from (5.51), obtain L2 = F2 and D = Ir .
Remark 3: In practical applications, due to some performance requirement, the rate
of asymptotic convergence of the observer error is normally specified a priori, i.e.,
the eigenvalues of matrix N are given in advance. For the case where matrix N is
specified a priori, the solution to the matrix equations (5.39)-(5.40) can be obtained
according to the following condition.
Corollary 5.4.
Given N, with arbitrary specified eigenvalues, there exists a solution to matrix equa-
tions (5.39)-(5.40) if and only if
 
A12 W1  
rank = rank A12 W1 . (5.59)
(NF2 − F2A22 ) −F2W2
Proof: The proof is trivial by noting that as N is given, (5.39)-(5.40) can be ex-
pressed as    
L1 A12 W1 = (NF2 − F2A22 ) −F2W2 . (5.60)
Thus for the case where N is specified, the solution to the matrix equations (5.39)-
(5.40) amounts to just testing the condition (5.59). Once it is found to be satisfied,
matrix L1 can be obtained from solving (5.60). Also, D = Ir and L2 = F2 .

Case 2 (q > r): When the above Algorithm fails, this means that there does not
exist a solution to the matrix equations (5.1)-(5.3) for the case where q = r. Now, a
question that arises here is whether that there still exists an observer of order higher
than r, i.e., a q-th order (r < q ≤ (n − p)) observer for the system? If so, under what
conditions? and how can it be derived?. By intuition, a higher-order observer would
exist since high order would add more degree of freedom into (5.1)-(5.3). Though,
it will be shown later that this is only true when l < p and that W1 has full-column
rank. The same finding for the scalar case has also been reported in Section 4.4.4.
(q−r)×(n−p)
Now, for a q-th order observer, let  define R ∈ R to be any arbitrary
full-row rank matrix. Also, let D = Ir 0r×(q−r) , then from (5.2), one can choose
L2 to be  
F
L2 = 2 . (5.61)
R
Here, L2 ∈ Rq×(n−p) and since R is any arbitrary full-row rank matrix, one can make
L2 to have full-row rank, i.e., rank(L2 ) = q. Substituting (5.61) into (5.1) and (5.3)
to get
   
F2 F
N − L1 A12 − 2 A22 = 0, N is Hurwitz (5.62)
R R
 
F
L1W1 = − 2 W2 . (5.63)
R
5.4 Multiple Functional Unknown Input Observers 111

where N ∈ Rq×q and L1 ∈ Rq×p . The introduction of an arbitrary full-row rank ma-
trix, R, seems to provide additional degree of freedom into (5.62). Now, by replacing
F2 with L2 in (5.53) and (5.54), necessary and sufficient conditions for the solvabil-
ity of (5.62) and (5.62) are obtained. They are expressed in the following theorem.

Theorem 5.5.
Let R be any arbitrary full-row rank matrix of dimension (q − r) × (n − p). Also, let
L2 be a full-row rank matrix as defined in (5.61). The two coupled matrix equations
(5.62) and (5.63) are solvable if and only if
⎛⎡ ⎤⎞
L2 A22 L2W2  
A12 W1
rank ⎝⎣ A12 W1 ⎦⎠ = rank (5.64)
L2 0
L2 0
   
(sL2 − L2 A22 ) −L2W2 A12 W1
rank = rank
A12 W1 L2 0
∀s ∈ C, ℜ(s) ≥ 0. (5.65)

Theorem 5.5 tells that a higher-order observer may still exist if there exists a full-
row rank matrix R such that both (5.64) and (5.65) hold. This suggests a search
for R. Here, one will show that the search only proceeds when W1 is a full-column
rank matrix. However, for other cases, there is no need to look for any higher-order
observer beyond r as there won’t be any.
Lemma 5.6.
For the case where l ≥ p and rank(W1 ) = p, if (5.54) is not satisfied, then (5.65)
won’t be satisfied for any q > r.
Proof: First, let one proves the case where l = p and rank(W1 ) = p. Now, when a
stable r-th order observer does not exist, this would mean that the pair (N2 , N1 ) is
not detectable. Or equivalently, condition (5.54) does not satisfy. This implies that
the left-hand side of (5.54) is always less than the right-hand side of (5.54), i.e.,
   
(sF2 − F2A22 ) −F2W2 A12 W1
rank < rank
A12 W1 F2 0
∀s ∈ C, ℜ(s) ≥ 0. (5.66)

Since l = p and rank(W1 ) = p, the right-hand side of (5.66) can be determined as

 
A12 W1
rank = rank(W1 ) + rank(F2 ) = (p + r). (5.67)
F2 0

The left-hand side of (5.66) can be determined as

  
(sF2 − F2A22 ) −F2W2 In−p 0
rank
A12 W1 −W1−1 A12 Il
 
= rank sF2 − F2 (A22 − W2W1−1 A12 ) + rank(W1 ). (5.68)
112 5 Functional Unknown Input Observers: Further Results and Applications

Thus, when (5.54) does not hold, this would mean that
 
rank sF2 − F2 (A22 − W2W1−1 A12 ) < r, ∀s ∈ C, ℜ(s) ≥ 0. (5.69)

Now, consider the right-hand side of (5.65): With l = p, rank(W1 ) = p and for
any q within r < q ≤ (n − p), one has
 
A12 W1
rank = (p + q). (5.70)
L2 0

The left-hand side of (5.65) is also obtained as

 
sF2 − F2(A22 − W2W1−1 A12 )
rank + p, ∀s ∈ C, ℜ(s) ≥ 0. (5.71)
sR − R(A22 − W2W1−1 A12 )

Due to (5.69), it is now impossible to find a R to satisfy (5.65) because one will
always have
 
sF2 − F2(A22 − W2W1−1 A12 )
rank < q, ∀s ∈ C, ℜ(s) ≥ 0 (5.72)
sR − R(A22 − W2W1−1 A12 )

which implies that the condition (5.65) is not satisfied for any higher-order q, where
r < q ≤ (n − p).  
F
Also, when L2 = 2 is a square matrix, i.e., q = (n − p), the left-hand side of
R
(5.72) can be expressed as
 
rank L2 (sIn−p − (A22 − W2W1−1 A12 ) )
 
= rank sIn−p − (A22 − W2W1−1 A12 ) , ∀s ∈ C, ℜ(s) ≥ 0 (5.73)

which implies that all the eigenvalues of matrix (A22 −W2W1−1 A12 ) have to be stable
to ensure the satisfaction of (5.65). However, due to (5.69), one can tell that there
must be at least one unstable pole. Even when L2 is a full-column rank matrix, i.e.,
q > (n − p), the left-hand side of (5.72) is the same as (5.73) which still requires
that matrix (A22 − W2W1−1 A12 ) to be stable. This proves the case where l = p and
rank(W1 ) = p.
To prove the case where l > p and rank(W1 ) = p one only needs to prove that
if (5.54) does not hold for q = r then (5.65) won’t hold for any q within the range
r < q < (n − p). The proof for this can be constructed by following similar lines
as above and therefore is omitted here. Finally note that for any q ≥ (n − p), the
necessary condition (5.41) does not hold since the left-hand side of (5.41) is
 
W1
rank = rank(W ) = l (5.74)
F2W2

and the right-hand side of (5.41) is rank(W1 ) = p. As l > p, condition (5.41) is not
satisfied. This completes the proof of Lemma 5.6.
5.4 Multiple Functional Unknown Input Observers 113

Remark 4: The result presented in Lemma 5.6 is intriguing and useful which has
partly answered the question one posed earlier regarding if there is a need to look
for a higher-order observer. Now, one has settled this question for the case where
l ≥ p and rank(W1 ) = p. One has established that there is no need to look for a
higher-order observer because if there does not exist a r-th order observer, then there
won’t be any beyond r. The case where l > p and rank(W1 ) ≤ p can be deduced by
following similar lines as shown in the proof of Lemma 5.6. Finally, note that for
the case where W1 = 0, one can show that there only exists a finite number of linear
functions F2 such that condition (5.41) holds. This is discussed below.
When W1 = 0 and since W has full-column rank, this implies that rank(W2 ) = l.
Now, condition (5.41) only holds for any F2 ⊂ X, where X ∈ R(n−p−l)×(n−p) is an
orthogonal basis for the null-space of W2 (i.e., XW2 = 0). Here, there can only be
a maximum of (n − p − l) linear functions that can satisfy the necessary condition
(5.41). Thus, for any given F2 ⊂ X, one can easily check both conditions (5.46)-
(5.47) (or (5.53)-(5.54)). Now, if (5.53) does not satisfy, then one can show that
(5.64) will not hold as well for any additional row matrix R, R ⊂ X. Alternatively,
one can pick out all the rows of X that would satisfy the existence conditions (5.46)-
(5.47). For this, let these rows be stacked together to form a matrix X1 . Then for any
F2 ⊂ X1 , a linear functional observer exists and can be designed.
Now, one considers the case where l < p and rank(W1 ) = l. Since  W 1 is a full-
S1
column rank matrix, there always exists a nonsingular matrix S = ∈ R p×p,
S2
S1 ∈ Rl×p , S2 ∈ R(p−l)×p such that
 
W11
SW1 = (5.75)
S2W1

where W11 ∈ Rl×l is an invertible matrix. Accordingly, one can derive two simplified
equivalent conditions from (5.64) and (5.65).
Corollary 5.7.
Let l < p and rank(W1 ) = l, then conditions (5.64) and (5.65) are the same as
⎛⎡ ⎤⎞
L̄2  
S̄2
rank ⎝⎣ S̄2 ⎦⎠ = rank (5.76)
L2
L2
   
sL2 − L̄2 S̄
rank = rank 2 , ∀s ∈ C, ℜ(s) ≥ 0 (5.77)
S̄2 L2
−1
where L2 is as defined in (5.61), S̄2 = S2 (A12 − W1W11 S1 A12 ) and L̄2 = L2 (A22 −
−1
W2W11 S1 A12 ).
Proof: The proof for Corollary 5.7 can be easily derived. The following is a sketch
of its proof. The right-hand side of (5.64) can be expressed as
114 5 Functional Unknown Input Observers: Further Results and Applications
⎛⎡ ⎤⎞
   S1 A12 W11
S 0 A12 W1
rank = rank ⎝⎣ S2 A12 S2W1 ⎦⎠
0 Iq L2 0
L2 0
⎛⎡ ⎤⎡ ⎤⎞
Il 0 0 S1 A12 W11
= rank ⎝⎣ −S2W1W11 −1
I p−l 0 ⎦ ⎣ S2 A12 S2W1 ⎦⎠
0 0 Iq L2 0
 
S̄2
= rank + l. (5.78)
L2

By following similar lines as above, the left-hand side of (5.64) is the same as
⎛⎡ ⎤ ⎞
L̄2
rank ⎝⎣ S̄2 ⎦⎠ + l. (5.79)
L2

Thus, from (5.78) and (5.79) one arrives at the condition (5.76). Now, the left-hand
side of (5.65) is the same as
 
sL2 − L̄2
rank + l. (5.80)
S̄2

Thus, from (5.80) and (5.78) one arrives at the condition (5.77). This completes the
proof of Corollary 5.7.
Note that the only unknown matrix in the two conditions (5.76) and (5.77) is R.
A close inspection reveals that the two conditions (5.76) and (5.77) are of similar
form as those conditions (3.65) and (3.66) of Theorem 3.5. Thus the discussion in
Sections 3.6 and 3.7 can be used to find a higher order observer beyond r. In the
following, one discusses two cases: (i) q = (n − p); and (ii) q < (n − p).
Clearly, when q = (n − p), condition (5.76) is satisfied since both matrices in
both sides of (5.76) have full-column rank. Now (5.77) can be satisfied if and only
if  
sIn−p − Ā22
rank = (n − p), ∀s ∈ C, ℜ(s) ≥ 0 (5.81)
S̄2
−1 −1
where Ā22 = (A22 − W2W11 S1 A12 ) and S̄2 = S2 (A12 − W1W11 S1 A12 ). Condition
(5.81) is also the same as the detectability of the pair (S̄2 , Ā2 ). As both S̄2 and Ā22
are two known matrices therefore checking the detectability or observability of the
pair is trivial.
Next, one considers the case where  there
 maybe an observer of order less than
S̄2
(n − p). For this, assume that rank = v, where v > r. Accordingly, one can
F2
always
 pick a full-row rank matrix R of dimension (n − p − v) × (n − p) such that
S̄2
rank = (n − p). Since the right-hand side of (5.76) now reaches the maximum
L2
of (n − p), the left-hand side of (5.76) also reaches the maximum of (n − p). Con-
dition (5.76) is therefore satisfied. Note that q = (n − p + r − v) which is less than
(n − p). Now, an observer of order (n − p + r − v) can be found if the left-hand
5.4 Multiple Functional Unknown Input Observers 115

side of (5.77) equals to (n − p). This can be checked since L2 , L̄2 and S̄2 are known
matrices.
This section thus has provided a comprehensive treatment on the solution to the
three coupled matrix equations (5.1)-(5.3). Thus the design of linear functional ob-
servers for linear systems with unknown inputs, or the design of delay-free func-
tional observers for time-delay systems as well the task of decentralized implemen-
tation of a global state feedback controller for multi-agent systems under the con-
straint of a complete decentralized information structure, can all now be done under
the same framework.

5.4.1 A Numerical Example

The following numerical example is given to illustrate the results discussed in Sec-
tion 5.4. Consider a linear system with unknown inputs, where matrices A, W , C and
F are as given below
⎡ ⎤ ⎡ ⎤
−1 0 0 1 0 0 0 1
⎢ 2 0 1 −1 1 0 0 ⎥ ⎢0⎥
⎢ ⎥ ⎢ ⎥ ⎡ ⎤
⎢ 0 3 0 0 1 1 0⎥ ⎢1⎥ 1000000
⎢ ⎥ ⎢ ⎥
A=⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 −3 0 1 1 ⎥, W = ⎢ 0 ⎥, C =
⎣ 1 1 0 0 0 0 0 ⎦ and
⎢ 0 0 0 0 1 0 −1 ⎥ ⎢1⎥ −1 0 1 0 0 0 0
⎢ ⎥ ⎢ ⎥
⎣ 1 0 0 0 0 −1 0 ⎦ ⎣1⎦
 0 1 0 0 1 0 −2 1
3 −2 −2 1 2 1 0
F= .
2 0 −1 1 1 0 0
The design of a reduced-order linear functional observer to estimate the two lin-
ear functions can now proceed as follow. First, one uses the transformation matrix
P as defined in (2.21) and the following submatrices are obtained
⎡ ⎤
⎡ ⎤ ⎡ ⎤ 0 0 0
−1 0 0 1 0 0 0  
⎢ 0 0 0⎥
A11 = ⎣ 2 0 1 ⎦, A12 = ⎣ 0 1 0 0 ⎦, A21 = ⎢ ⎥, F1 = 3 −2 −2
⎣ 1 0 0⎦ 1 0 −1
−2 3 0 −1 1 1 0
⎡ ⎤ ⎡ ⎤ −1 1 0
−3 0 1 1 ⎡ ⎤ 0
1  
⎢ 0 1 0 −1 ⎥ ⎢1⎥
A22 = ⎢ ⎥, W1 = ⎣ 1 ⎦, W2 = ⎢ ⎥ and F2 = 1 2 1 0 .
⎣ 0 0 −1 0 ⎦ ⎣1⎦ 1 1 0 0
0
0 1 0 −2 1
The solution to the three coupled matrix equations (5.1)-(5.3) starts with the low-
est order, i.e., q = r = 2. For this, one first checks the necessary condition (5.41).
Here, the condition is satisfied and thus one next checks the conditions (5.46)-
(5.47) of Theorem 5.3 to see if they are satisfied or not. If both are satisfied, then
a second-order functional observer exists and can be easily obtained. For this ex-
ample, both conditions (5.46)-(5.47) of Theorem 5.3 are not satisfied and hence a
second-order linear functional observer does not exist. (Alternatively, by using the
condition (5.53), it is also found that it is not satisfied and thus one can quickly
116 5 Functional Unknown Input Observers: Further Results and Applications

conclude that a second-order observer does not exist). As discussed in Section 5.4,
since l < p and rank(W1 ) = l one can indeed look for a higher-order observer.
Now, by using (5.75), one can easily obtain submatrices, S1 and S2 ,
  0 1 0
where S1 = 1 0 0 and S2 = . Also, one can obtain matrix
0 0 1 
−1 −1 1 0 0
S̄2 = S2 (A12 − W1W11 S1 A12 ), where S̄2 = . Now, by examining
−1 1 1 0
the right-hand side term
 of (5.76)
 reveals that one only needs to choose a row
matrix R, where R = 0 0 0 x (x = 0) to make both sides of (5.76) the same, i.e.,
condition (5.76) is satisfied. Here,⎡one simply ⎤chooses x = −1 and hence matrix L2
  1 2 1 0
F
is obtained, where L2 = 2 = ⎣ 1 1 0 0 ⎦. Now, by replacing F2 by L2 , from
R
0 0 0 −1
(5.44) and (5.45) matrices N1 and N2 are obtained, where
⎡ ⎤ ⎡ ⎤
−1.6667 0.5 1 −0.3333 0.5 0
N1 = ⎣ 0 −1.5 0 ⎦ and N2 = ⎣ 0.3333 −0.5 0 ⎦.
0.3333 −0.5 −2 −0.6667 1 0
Now, the pair (N2 , N1 ) is observable and thus matrix N can be assigned with
any prescribed stability. Let its eigenvalues be assigned at ⎡{−3, −4, −5}.
⎤ Using
−493 0 0
pole placement technique, matrix Z is easily obtained as Z = ⎣ −315 0 0 ⎦. Using
⎡ ⎤ 35 0 0
−166 247 1
(5.52), matrix N is obtained as N = ⎣ −105 156 0 ⎦. Using (5.51), matrix L1 is
⎡ ⎤ 12 −18 −2
−82 79 −166  
obtained as L1 = ⎣ −52 51 −106 ⎦. Matrix D is obtained as D = Ir 0r×(q−r) =
6 −5 12
 
1 0 0
. Finally, from (4.49) and (4.50), matrices E and J are obtained, where
0 1 0 ⎡ ⎤
  −201 24 −1307
85 −81 164
E= and J = ⎣ −132 21 −843 ⎦.
53 −51 105
21 −5 103
The observer design is thus completed and a third-order linear functional ob-
server is obtained.

5.5 A Common Disturbance Decoupled Observer for Linear

Systems with Unknown Inputs
This section presents an application of the theory of functional unknown input ob-
servers to the design of a common disturbance decoupled observer for two linear
systems with unknown inputs. The objective is to design a common linear func-
tional observer that can observe a partial set of the state vector of two linear systems
5.5 A Common Disturbance Decoupled Observer for Linear Systems 117

with unknown inputs. Due to the fact that both systems are subjected to unknown
inputs, a common functional observer is proposed which uses only the available
output information. Existence conditions as well as a design procedure are given for
constructing the proposed common observer. A numerical example is given to show
the design of a common linear functional observer.
It is always intriguing to know whether or not there exists a common observer that
can observe the states for each of a given set of plants. The problem of simultaneous
observation ( [68, 86, 143]) of two or several linear plants is of interest when, for
example, a robust observer is designed that converges despite of failures in some
components in a system. In the observer design of a nonlinear system at various
operating conditions, the nonlinear system is often linearized at various operating
points and some of linearized models can be obtained. If the states of each of these
linearized models can be observed using a common observer, then a fixed observer
can be used for the nonlinear system.
The simultaneous observation problem of linear systems with known inputs was
first studied in [143] using the coprime factorization technique. Here, the authors
presented necessary and sufficient conditions for the existence of a simultaneous
observation. Based on the stable inverse approach [68], a general result for param-
eterization of all simultaneous observers was given. Moreno [86] provided a state-
space characterization for simultaneous observation of several linear plants. For the
case of two linear systems with known inputs, the simultaneous observation problem
was completely solved [86] and the author also presented an algorithm and a struc-
ture for simultaneous observers. When the plants are subjected to unknown inputs
or measurement of the system inputs is either too expensive or perhaps physically
not possible [16, 46], one has to use only the available output information to design
a common observer [130].
In this section, one discusses the problem of designing a common linear func-
tional observer for two linear systems with complete unknown inputs. A structure
for a common linear functional observer which only uses the available output in-
formation is proposed. Here, for the proposed structure, one shows that the simulta-
neous observation problem of two plants is reduced to a problem of designing two
observers: The first is a full-order unknown input observer for one of the two sys-
tems and the second is a reduced-order unknown input linear functional observer
of a system comprising two-connected systems. Existence conditions and a design
procedure are given.
Consider two linear time-invariant systems, Φi (i=1,2), as described by
⎧
⎨ ẋi (t) = Ai xi (t) + Wi di (t)
Φi : yi (t) = Ci xi (t) (5.82)
⎩
zi (t) = Fi xi (t),

where xi (t) ∈ Rni is the state vector, yi (t) ∈ R p is the measured output vector,
di (t) ∈ Rli is treated in this chapter as unknown input vector and zi (t) ∈ Rr is a vector
of signals to be estimated. Matrices Ai , Wi , Ci and Fi are known real constant and of
118 5 Functional Unknown Input Observers: Further Results and Applications

appropriate dimensions. Note that, the dimension of the state and unknown input
vectors of both systems can be different (i.e. n1 and l1 do not have to be equal to n2
and l2 , respectively). For brevity, the dimension of the output vectors is assumed to
be the same for both systems, although they can be different. When the dimensions
are different, (dummy) outputs can be added to one of the two systems [86].
The aim is to use only the available real-time output information, yi (t), to con-
struct a single disturbance decoupled linear functional observer, called a Common
Functional Observer (CFO), such that it is a linear functional observer for every el-
ement of the family of LTI systems described by (5.82). Fig. 5.1 shows the block
diagram of a CFO. In Fig. 5.1, Φi (i=1,2), represents the actual plant, Ωc is the
CFO (to be designed) and ẑi (t) denotes the estimate of zi (t) such that ẑi (t) → zi (t),
i = 1, 2.

CFO

d(t) yi (t) ẑi (t)

Φi Ωc

Fig. 5.1. Block Diagram of a Common Functional Observer (CFO)

Now, a structure for the above CFO is proposed. It consists of three main
components. The first is a decision logic switch that provides an early detection/
identification of the characteristics of an estimation error function. The second com-
ponent is a full-order unknown input observer designed to provide asymptotic track-
ing of the state vector of either Φ1 or Φ2 . The third component is a reduced-order
unknown input linear functional observer of a system comprising two-connected
systems.
In the following sections, one will explain in more detail the development of the
three main components and establish an existence condition for the CFO.

5.5.1 A Decision Logic Switch

Let e(t) be defined as an error function, where

p
e(t) = ∑ |ŷ1j (t) − yij (t)|, (5.83)
j=1

where ŷ1j (t) ( j = 1, 2, ..., p) denotes the estimated j-th output of the system Φ1 , yij (t)
(i = 1, 2) denotes the j-th output of the i-th system, Φi , and |x| denotes the absolute
value of x. Thus for the case where ŷ1 (t) is an asymptotic estimate of y1 (t) and that
5.5 A Common Disturbance Decoupled Observer for Linear Systems 119

the real-time measurement (i.e., yi (t)) is coming from y1 (t) then one would expect
that the error function (5.83) exponentially diminished to zeros (i.e., e(t) → 0 as
t → ∞). Based on the property of e(t), one proposes the following decision logic
switch as shown in Fig. 5.2.

S2

S1

Fig. 5.2. A Decision Logic Switch

Here, the decision logic is modelled as a switch, where its status of being opened
or closed is determined from the asymptotic convergence property of e(t) as defined
in (5.83). Let one define the decision logic as follows

1, e(t) ≥ δ
L (e(t)) = (5.84)
0, e(t) < δ ,

where δ > 0 is a threshold value of the error function e(t). Thus, the switch is open
when it is determined that the error, e(t), is small (i.e., ŷ1 (t) is actually estimating
the measurement signal yi (t)). Otherwise the switch is closed.
Fig. 5.3 shows the proposed structure for the CFO for two linear systems with
complete unknown inputs.
As can be seen from Fig. 5.3, the only information that is fed into the CFO is
the real-time output information, yi (t). The proposed CFO observer for Φi consists
of the decision logic switch as defined in (5.84) and two unknown input functional
observers (UIFOs), namely Ω1 and Ωu . The construction of both Ω1 and Ωu will be
discussed next. In Fig. 5.3, v̂1 (t) is an estimate of v1 (t), ỹ1 (t) and v1 (t) are defined
as

ỹ1 (t)  ŷ1 (t) − yi (t) (5.85)

and

v1 (t)  z1 (t) − zi (t). (5.86)

5.5.2 Construction of Ω1

In Fig. 5.3, Ω1 is an unknown input observer such that when the real-time output
information is coming from system 1, i.e. yi (t) = y1 (t), Ω1 provides asymptotic esti-
mation for both y1 (t) and z1 (t). To achieve this task, let Ω1 be a full-order unknown
input observer of the following form
120 5 Functional Unknown Input Observers: Further Results and Applications

d(t)
Φi

CFO
yi (t)

ŷ1 (t) − ỹ1 (t) v̂1 (t)

Ω1 Ωu
+

S2
− ẑi (t)
ẑ1 (t) S1 +

ẑ1 (t)

Fig. 5.3. Block Diagram of a CFO for Two Linear Systems with Unknown Inputs

⎧
⎨ ẇ1 (t) = N1 w1 (t) + J1yi (t)
Ω1 : ŷ1 (t) = D11 w1 (t) + E11yi (t) (5.87)
⎩
ẑ1 (t) = D12 w1 (t) + E12yi (t),

where w1 (t) ∈ Rn1 , N1 (Hurwitz), J1 , D11 , E11 , D12 and E12 are matrices of appropri-
ate dimensions to be determined such that for i = 1, ẑ1 (t) → z1 (t) and ŷ1 (t) → y1 (t)
for Φ1 .
Here, note that, necessary and sufficient conditions for the existence of Ω1 for
system Φ1 are well-known and have been discussed in Section 2.4. Thus one can
readily derive matrices N1 , J1 , D11 , E11 , D12 and E12 such that ẑ1 (t) → z1 (t) and
ŷ1 (t) → y1 (t).

5.5.3 Construction of Ωu

Assuming that the switch is closed, thus the interconnection of the observer Ω1 and
the system Φ2 is described as follows
      
ẇ1 (t) N1 J1C2 w1 (t) 0
= + d (t) (5.88)
ẋ2 (t) 0 A2 x2 (t) W2 2

   
ỹ1 (t) ŷ1 (t) − y2(t)
ȳ(t) = =
ẑ1 (t) D12 w1 (t) + E12y2 (t)
  
D11 (E11C2 − C2 ) w1 (t)
= (5.89)
D12 E12C2 x2 (t)
5.5 A Common Disturbance Decoupled Observer for Linear Systems 121
 
  w1 (t)
v1 (t) = ẑ1 (t) − z2 (t) = D12 (E12C2 − F2) . (5.90)
x2 (t)

For brevity, let the interconnected system defined by equations (5.88)-(5.90) be

called Φu . Now, let Ωu be a reduced-order unknown input linear functional observer
of Φu , where

ẇ2 (t) = N2 w2 (t) + J2 ȳ(t)

Ωu : (5.91)
v̂1 (t) = D2 w2 (t) + E2 ȳ(t),

w2 (t) ∈ Rq , N2 (Hurwitz), J2 , D2 and E2 are matrices of appropriate dimensions to

be determined such that v̂1 (t) → v1 (t).
Note that, the construction of Ωu for Φu such that v̂1 (t) → v1 (t) can be readily
carried out as this is basically the problem of designing a linear functional observer
for linear systems with unknown inputs (refer to Section 5.4).

5.5.4 Existence Conditions for a Common Functional Observer

The following theorem establishes an existence condition for the proposed CFO of
Fig. 5.3.

Theorem 5.8.
There exists a CFO of the structure depicted in Fig. 5.3 such that ẑi (t) → zi (t) (i =
1, 2) provided that there exist two UIFOs Ω1 and Ωu for Φ1 and Φu , respectively.

Proof: Firstly, suppose that the information that is feeding into Ω1 is actually com-
ing from the first plant, (i.e., i = 1, Φi = Φ1 and yi (t) = y1 (t)). Accordingly, if Ω1
exists, then Ω1 provides both estimates ẑ1 (t) and ŷ1 (t) such that ẑ1 (t) → zi (t) and
ŷ1 (t) → yi (t). Since ŷ1 (t) → yi (t), it follows that ỹ1 (t) → 0. From (5.83), it is also
clear that the error function e(t) is exponentially diminishing to zeros. This in turn
would activate the logic gate and the switch is open. Hence the signal ȳ(t) that is
feeding into Ωu is now ȳ(t) = ỹ1 (t) (i.e., ȳ(t) → 0). Accordingly, provided that there
exists an internally stable Ωu (i.e., matrix N2 is Hurwitz) then v̂1 (t) → 0. From Fig.
5.3, it is clear that ẑi (t) → z1 (t). In this case, the CFO provides the estimate ẑ1 (t) of
system Φ1 . Here, the CFO achieves the task it is supposed to fulfill, i.e., when the
information is from system Φ1 , the CFO produces the estimate ẑ1 (t) of Φ1 .
Secondly, let one now considers the case where the information is coming from
the second plant (i.e., i = 2, Φi = Φ2 and yi (t) = y2 (t)). In this case, Ω1 is not an
actual observer for Φ2 . Therefore ỹ1 (t) = (ŷ1 (t) − y2(t))  0 and e(t)  0. Hence,
the logic switch would be closed and the interconnection of systems Ω1 and Φ2 is
now described by Φu of equations (5.88)-(5.90). Accordingly, if Ωu for Φu exists,
then Ωu provides the estimate v̂1 (t) such that v̂1 (t) → v1 (t). From Fig. 5.3, it is clear
that ẑi (t) = (ẑ1 (t) − v̂1(t)) → z2 (t). In this case, the CFO provides the estimate ẑ2 (t)
of system Φ2 . Again, the proposed CFO has achieved its objective, i.e., when the in-
formation is coming from Φ2 , the CFO produces the estimate ẑ2 (t). This completes
the proof of Theorem 5.8.
122 5 Functional Unknown Input Observers: Further Results and Applications

Remark 5: The CFO discussed in this chapter caters for both known/unknown in-
puts and it only uses the available output information yi (t). In [86], it is required
that the input/output pair (d1 (t), yi (t)) be available. Note that even when d1 (t) is a
known input vector, sometimes it is not possible to have access to d1 (t) when i = 1.
The performance of the CFO of this chapter, however, depending on the timely de-
tection, is dictated by the logic decision (5.84). An alternative for this will be either
to refine the logic decision (5.84) or to incorporate a microprocessor based detection
system that can take more measurements of the error function (5.83) at more fre-
quent intervals in order to provide a more timely identification of the error function
(5.83). This would be a topic for further research. In the numerical example, it will
be shown that early detection can lead to improvement in the convergence rate of
the estimates.
The design procedure of a CFO can now proceed as follows. In the first step,
one designs a full-order unknown input observer, Ω1 for system Φ1 . For this, one
can easily adopt a design method of [24] (or the discussion in Section 2.4) to obtain
matrices N1 , J1 , D11 , E11 , D12 and E12 such that ẑ1 (t) → z1 (t) and ŷ1 (t) → y1 (t).
In the second step, one first obtains the interconnected system Φu and proceeds to
design a reduced-order observer Ωu for Φu . For this step, one can use the results
discussed in Section 5.4 to obtain matrices N2 (Hurwitz), J2 , D2 and E2 such that
v̂1 (t) → v1 (t).

5.5.5 A Numerical Example

Consider a family of systems Φi for i = 1, 2, where

   
0 1 1    
System 1, Φ1 : A1 = , W1 = , C1 = 2 7 and F1 = 1 −1 .
−1 −5 0
   
0 1 1    
System 2, Φ2 : A2 = , W2 = , C2 = 2 7 and F2 = 1 −1 .
−2 −10 0
In the following, let one design a CFO for the above systems based on the pro-
posed structure and design procedure of this chapter. First, one designs an observer
Ω1 of the form (5.87) for system 1, Φ1 , such that ẑ1 (t) → z1 (t) and ŷ1 (t) → y1 (t).
By using the design procedure presented in Section 2.4, the following observer pa-
rameters for Ω1 are obtained
   
−7.9 −22.4 1.75  
N1 = , eig(N1 ) = {−1.5, −3}, J1 = , D11 = 2 7 ,
1.4 3.4 −0.5
E11 = 1, D12 = 1 −1 and E12 = 0.5.

The next step in the design of a CFO is to design an observer Ωu for Φu of (5.88)-
(5.90). By substituting the above observer parameters into (5.88)-(5.90), the system
Φu is obtained, where
5.5 A Common Disturbance Decoupled Observer for Linear Systems 123
⎡ ⎤ ⎡ ⎤
  −7.9 −22.4 3.5 12.25   0
ẇ1 (t) ⎢ 1.4 3.4 −1 −3.5 ⎥ ⎢ ⎥
=⎣⎢ ⎥ w1 (t) + ⎢ 0 ⎥ d2 (t),
ẋ2 (t) 0 0 0 1 ⎦ x2 (t) ⎣1⎦
0 0 −2 −10 0
  
2 7 0 0 w1 (t)
ȳ(t) = ,
1 −1 1 3.5 x2 (t)
 
  w1 (t)
v1 (t) = 1 −1 0 4.5 .
x2 (t)

One can now easily design an observer, Ωu , for Φu such that v̂1 (t) → v1 (t). Here,
by adopting the design method discussed in Section 5.4, a reduced-order observer
of second-order can be designed to estimate v1 (t), where
   
−15.2344 −15.7299 −4.4075 −4.5
N2 = , eig(N2 ) = {−3, −6}, J2 = ,
7.1823 6.2344 4.8189 1.7835
   
D2 = 1 0 and E2 = 0.016 0 .

5.5.6 Simulation Results

In the simulation studies, one first considers the case where the output signal that
feeds into the CFO is from system 1 (i.e., yi (t) = y1 (t)). Fig. 5.4 shows the unknown
input signal, d1 (t), arbitrarily generated. For the case when the switch is activated
(open) within one second delay, Fig. 5.5 shows the simulated  responses z1 (t) and
−20
ẑ1 (t), with initial conditions x1 (0) = 0, w1 (0) = and w2 (0) = 0.
20
Fig. 5.6 shows the simulated responses z1 (t) and ẑ1 (t) for the case where the
switch is activated (open) within five seconds delay. Observe that one still has
ẑ1 (t) → z1 (t). It is also clear that the earlier the detection time, the better the
tracking.
Next in the simulation studies, one considers the case where after thirty seconds,
the output signal that feeds into the CFO is coming from system 2 (i.e., yi (t) =
y2 (t)). Fig. 5.7 shows the unknown input signals, d1 (t) for the first thirty seconds
and d2 (t) for the last thirty seconds.
Fig. 5.8 shows the simulated responses zi (t) and ẑi (t) for the case where the
switch is activated (open) within five seconds delay and the switch is closed after
thirty four seconds. It is clear from Fig. 5.8 that ẑi (t) → zi (t) (i = 1, 2).
This section has shown that the design of a common disturbance decoupled lin-
ear functional observer for two linear systems with unknown inputs can be achieved
based on the theory of functional unknown input observers. Future work may in-
clude design feedback controller using the proposed observer. It is also desirable to
extend current work to more than two plants and how to (best) design a decision
logic to provide early detection which in turn improves the convergence rate of the
estimates.
124 5 Functional Unknown Input Observers: Further Results and Applications

80

60

40

20
d1(t)

−20

−40

−60

−80
0 2 4 6 8 10 12 14
Time

Fig. 5.4. Unknown input d1 (t)

250

200

150

100

50

Actual
0 Estimated

−50
0 5 10 15 20 25 30
Time

Fig. 5.5. Responses of z1 (t) and ẑ1 (t) (one second delay)
5.5 A Common Disturbance Decoupled Observer for Linear Systems 125

250

200

150

100

50
Actual
Estimated
0

−50
0 5 10 15 20 25 30
Time

Fig. 5.6. Responses of z1 (t) and ẑ1 (t) (five seconds delay)

80

60

40

20

−20

−40

−60

−80
0 10 20 30 40 50 60
Time

Fig. 5.7. Unknown input di (t)

126 5 Functional Unknown Input Observers: Further Results and Applications

250

200

150

100

50

−50

−100

−150
Actual
Estimated
−200

−250
0 10 20 30 40 50 60
Time

Fig. 5.8. Responses of zi (t) and ẑi (t)

5.6 Conclusion
This chapter has provided a detailed treatment on the solution to the three coupled
matrix equations (5.1)-(5.3) for multiple functions. This chapter has also shown that
the design of linear functional observers for linear systems with unknown inputs,
and the design of delay-free functional observers for time-delay systems as well
the task of decentralized implementation of a global state feedback controller for
multi-agent systems under the constraint of a complete decentralized information
structure, can all be done under the same framework of linear functional unknown
inputs observer theory. This chapter has also discussed the design of a common dis-
turbance decoupled linear functional observer for two linear systems with unknown
inputs based on the theory of functional unknown input observers. It would be de-
sirable to extend current work to more than two plants and how to better design a
decision logic to provide early detection which in turn improves the convergence
rate of the estimates. In the next chapter, one will look at the design of functional
observers for nonlinear systems and the problem of simultaneous estimation of both
state and unknown inputs of uncertain systems.