Representation and Stability Analysis of

PDE-ODE Coupled Systems

Amritam Das ∗ Sachin Shivakumar ∗∗ Siep Weiland ∗
Matthew Peet ∗∗
∗
Department of Electrical Engineering, Eindhoven University of
Technology, 5600 MB, The Netherlands (e-mail:
{am.das,s.weiland}@tue.nl).
∗∗
School for Engineering of Matter, Transport and Energy, Arizona
arXiv:1812.07186v3 [math.OC] 20 Dec 2018

State University, Tempe, AZ 85292 USA (e-mail: {sshivak8,

mpeet}@asu.edu)

Abstract: In this work, we present a scalable Linear Matrix Inequality (LMI) based framework
to verify the stability of a set of linear Partial Differential Equations (PDEs) in one spatial
dimension coupled with a set of Ordinary Differential Equations (ODEs) via input-output
based interconnection. Our approach extends the newly developed state space representation
and stability analysis of coupled PDEs that allows parametrizing the Lyapunov function on
L2 with multipliers and integral operators using polynomial kernels of semi-separable class.
In particular, under arbitrary well-posed boundary conditions, we define the linear operator
inequalities on Rn × L2 and cast the stability condition as a feasibility problem constrained by
LMIs. In this framework, no discretization or approximation is required to verify the stability
conditions of PDE-ODE coupled systems. The developed algorithm has been implemented in
MATLAB where the stability of example PDE-ODE coupled systems are verified.

Keywords: Partial Differential Equations, Ordinary Differential equations, Stability, Lyapunov

Function, Linear Matrix Inequality, Sum of Squares.

1. INTRODUCTION In this paper, the linear ODEs are of the following form:
ẋ(t) = Ax(t) + Bv(t),
Coupled Ordinary Differential Equations(ODEs)-Partial (1)
w(t) = Cx(t).
Differential Equations(PDEs) are useful to model systems
For s ∈ [a, b], the linear one dimensional PDEs are
that involve states varying over time as well as states
described in the following form:
varying over both space and time. Interaction among
systems that are governed by ODEs and PDEs appear ∂z(s, t)
ż(s, t) = A0 (s)z(s, t) + A1 (s) (2)
in multi-scale modeling of fluid flows, for instance in ∂s
Quarteroni and Veneziani (2003) and Stinner et al. (2014), ∂ 2 z(s, t)
the flow structures are the function of space and time; + A2 (s) + B1 (s)u(t),
∂s2
however, they are driven by global flow pattern that is Z b
∂z(s, t)

only a function of time. They also frequently appear in y(t) = C1 z(t) + Ca (s)z(s, t) + Cb (s) ds,
thermo-mechanical systems (e.g. Tang and Xie (2010)) and a ∂s
 ∂z(s, t) ∂z(s, t) 
chemical reactions (e.g. Christofides (2012)). A large class Bc col z(a, t), z(b, t), |s=a , |s=b = 0.
of these applications involves linear ODEs and PDEs that ∂s ∂s
are mutually coupled via inputs and outputs as illustrated Here, x(t) ∈ Rno and z(s, t) ∈ Rnp are the states of
in Fig. 1. the ODEs and the PDEs respectively. The models (1)-
(2) are connected by the signals col(v(t), w(t)) ∈ Rmo ×
Rpo and col(u(t), y(t)) ∈ Rmp × Rpp . In (1) A, B, C
are constant matrices of suitable dimensions. In (2),
u(t) y(t)
PDE A0 (s), A1 (s), A2 (s), B1 (s), Ca (s), Cb (s) are matrix valued
functions of appropriate dimensions and C1 is a constant
matrix. Bc ∈ R2np ×4np is required to have full row rank.
The interconnection between (1)-(2) is described by a
w(t) v(t) matrix L ∈ Rmo +mp ×po +pp that has full row rank and
ODE relates the interconnection signals with the help of the
following algebraic equation:
   
v(t) w(t)
Fig. 1. Interconnection between PDE-ODE models. =L . (3)
u(t) y(t)
In this paper, we develop a scalable approach to verify the constraints on these operators can be re-formulated as
stability of the coupled system in (1)-(2) that does not scalable Linear Matrix Inequalities(LMIs) that establish
depend on any discretization or approximation. Regard- the stability condition for PDE-ODE coupled system.
ing verifying stability, many methods used in the anal-  h i
x
P{P,Q,Q⊤ ,S,R1 ,R2 } (s) :=
ysis of Distributed Parameter Systems(DPS) rely on ei- z
ther application-specific analytical approaches(e.g. Nicaise  Z b

et al. (2009)) or projection of the entire system onto a  Px + Q(s)z(s)ds

finite-dimensional vector space as shown in Ravindran 

Z s
a Z b

 (4)
(2000), Das et al. (2018) and the references therein. Once Q⊤ (s)x + S(s)z(s) + R1 (s, η)z(η)dη + R2 (s, η)z(η)dη
projected to a finite-dimensional subspace, the reduced a s

order model is treated as a set of ODEs, and there are The organization of the paper is as follows. Section 2 pro-
many well-studied approaches to test for stability and de- vides a preliminary discussion on the notations, a model
sign controllers. However, discretization techniques for ap- of the PDE-ODE coupled systems and the Lyapunov sta-
proximating an infinite-dimensional system with a finite- bility theory. In Section 3, we define the class of opera-
dimensional model are prone to numerical instabilities tors P{P,Q1 ,Q2 ,S,R1 ,R2 } and derive its positivity condition
and other truncation errors. This results in large system as well as the composition and adjoint operations of these
matrices which often make them computationally demand- operators. In Section 4, the PDE-ODE coupled model has
ing. Moreover, these finite-dimensional approximations of been represented using the same class of operators. In
the DPS do not necessarily represent the behavior of the Section 5, the equivalent LMI conditions have been derived
original system to a required level of accuracy, thus the to verify the stability of PDE-ODE coupled system. In
properties, such as stability, robustness are not exact if Section 6, we numerically illustrate the methodology to
the finite dimensional approximation replaces the original verify the stability of example PDE-ODE coupled systems
model. Another approach frequently used in stabilization in MATLAB. At last, Section 7 provides some conclusions
and control of coupled PDE-ODE systems is the backstep- and directions for future research.
ping method, e.g. Krstic and Smyshlyaev (2008), Krstic
(2009) and Susto and Krstic (2010). The backstepping
technique does not require any discretization or approxi- 2. PRELIMINARIES
mation of the model; however, it does not allow us to utilize
Lyapunov theory to verify stability. Also, these methods 2.1 Notation
are neither generic nor scalable for a general class of PDE-
2
ODE coupled system. For convenience, we denote ∂x ∂ x
∂s := xs and ∂s2 := xss . We
On the other hand, using the extension of Lyapunov the- use Z to denote a field of integers. We use Sm ⊂ Rm×m
ory in infinite-dimensional space by Datko (1970), many to denote the symmetric matrices. We define the space of
attempts were made to use Sum-of-Squares (SOS) op- square integrable Rm -valued functions on X as Lm 2 (X).
timization methods for constructing Lyapunov functions Lm (X) is equipped with the inner product hx, yiL2 =
R b2 ⊤ n
for PDEs. Some of the notable works include but not a x (s)y(s)ds. For denoting the inner product in R ,
limited to Papachristodoulou and Peet (2006), Fridman ⊤
we use hx, yiRn = x y. The Sobolov space is defined by
and Orlov (2009) and Valmorbida et al. (2016). However, k
W q,n (X) := {x ∈ Ln2 (X) | ∂∂skx ∈ Ln2 (X) for all k ≤ q}.
these works cannot deal with a general class of coupled For an inner product space X, operator P : X → X is
PDEs that involve reaction, transport, and diffusion terms. called positive, if for all x ∈ X, we have hx, PxiX ≥ 0. We
From our lab, numerous problems related to the analysis of use P < 0 to indicate that P is a positive operator. We
stability and robustness and controller-observer synthesis say that P : X → X is coercive if there exists some ǫ > 0
are addressed in Meyer and Peet (2015), Gahlawat and such that hx, PxiX ≥ ǫkxk2X for all x ∈ X.
Peet (2017) etc. Recently, Peet (2018), Shivakumar and
Peet (2018) have developed a novel SOS framework for
representation, stability and robustness analysis for cou- 2.2 PDE-ODE Coupled System
pled PDEs with arbitrary boundary condition. These re-
sults provide a unifying framework and a computationally In this section, we elaborate on the description of the cou-
scalable method for the analysis of a large class of PDEs. pled linear PDE-ODE system.The system is represented
The contributions of this paper involve extending the as an interconnection between two systems; one of them is
representation and the stability analysis of coupled PDEs governed by a set of partial differential equations (PDEs)
to the case of linear PDE-ODE coupled systems. To while the other is governed by a set of ordinary differential
this end, we construct a quadratic Lyapunov function equations (ODEs). In the following subsections we briefly
D   E
x x describe the corresponding PDE and ODE model.
of the form V (x, z) = , P{P,Q,Q⊤ ,S,R1 ,R2 } ,
z z
where P{P,Q,Q⊤ ,S,R1 ,R2 } is a class of self-adjoint lin- A. ODE Model: The ODE model has the following state
ear operator on Rno × L2 . The criterion for the sta- space representation:
bility of the PDE-ODE coupled system is to search ẋ(t) = Ax(t) + Bv(t),
for a positive P{P,Q1 ,Q2 ,S,R1 ,R2 } that restricts the time (5)
w(t) = Cx(t),
derivative of the Lyapunov function to be negative.
We parametrize the multipliers and integral operators where x(t) ∈ Rno are the state variables, v(t) ∈ Rmo ,
of P{P,Q1 ,Q2 ,S,R1 ,R2 } with polynomial kernels of semi- w(t) ∈ Rpo are the interconnection signals. A, B, C are
separable class as in (4). We show that the positivity matrices of suitable dimensions.
B. PDE Model: An operator based notation can be and
utilized to represent the PDE model in the following
abstract state space form: *     2
 + *    +  
ż(t) = Az(t) + Bu(t), x x x x  x 
(6) , P Ā + Ā ,P ≤ −γ   ,
y(t) = Cz(t). z z z z  z 
X X X
For functions z : [a, b] × R+ → Rnp , u : R+ → Rmp , (13)
n
y : R+ → Rpp , the system operators A : L2 p [a, b] ⊃ DA →
np n n for all x ∈ Rno , z ∈ DA . Then (10) is exponentially stable.
L2 [a, b], B : Rmp → L2 [a, b] and C : L2 [a, b] → Rpp are
p p

specified by the following definitions:

(Az)(s) := A0 (s)z(s) + A1 (s)zs + A2 (s)zss , 3. PARAMETRIZATION OF THE CLASS OF
(Bu)(s) := B1 (s)u, OPERATOR P{P,Q1 ,Q2 ,S,R1 ,R2 }
Zb !
Definition 2. For a matrix P ∈ Rno ×no , and bounded
Cz := C1 zb + Ca (s)z + Cb (s)zs ds, (7) polynomial functions Q1 : [a, b] → Rno ×np , Q2 : [a, b] →
a Rnp ×no , S : [a, b] → Rnp ×np , and R1 , R2 : [a, b] × [a, b] →
n
with its domain Rnp ×np with x ∈ Rno and z ∈ L2 p [a, b], the class of
DA :={z ∈ W 2,2 : Bc zb = 0}, operators P{P,Q1 ,Q2 ,S,R1 ,R2 } are defined as

(8)   
zb :=col z(a), z(b), zs (a), zs (b) . x
P{P,Q1 ,Q2 ,S,R1 ,R2 } (s) :=
In (7), A0 (s), A1 (s), A2 (s) ∈ Rnp ×np are matrix valued z
functions. B1 (s) ∈ Rnp ×mp , Ca (s), Cb (s) ∈ Rmp ×np are  Z b 
also matrix valued functions. C1 ∈ Rmp ×4np is a constant  Px + Q1 (s)z(s)ds

matrix. Bc ∈ R2np ×4np has full row rank. 

Z s
a Z b
.

Q2 (s)x + S(s)z(s) + R1 (s, η)z(η)dη + R2 (s, η)z(η)dη
C. PDE-ODE Interconnection: A regular interconnec- a s
tion between the ODE model in (5) and the PDE model (14)
in (6) is specified by the following algebraic constraint on
the variable pairs (v, w) and (u, y): In the remainder of this section, we investigate three
     properties of the class of operator P{P,Q1 ,Q2 ,S,R1 ,R2 } . First,
v(t) L1 L2 w(t)
= . (9) we provide a sufficient condition on the positivity of the
u(t) L3 L4 y(t) operator. Subsequently, an efficient method to compute
| {z }
L composition and adjoint of the operators of the form
The matrix L ∈ Rmo +mp ×po +pp has full row rank and P{P,Q1 ,Q2 ,S,R1 ,R2 } has been provided. The composition and
is partitioned to relate interconnection signals of ODEs the adjoint operation on the operators will allow us to
and PDEs individually. Substituting (9) in (5) and (6), we obtain a scalable represention of the PDE-ODE coupled
obtain the following autonomous state-space model. system. With the help of the positivity condition on
     the operator, the Lyapunov stability conditions can be
ẋ(t) A + BL1 C BL2 C x(t)
= , (10) reformulated as LMIs.
ż(t) BL3 C A + BL4 C z(t)
| {z }
Ā
n n 3.1 Positivity of P{P,Q,Q⊤ ,S,R1 ,R2 } Operator
with Ā : Rno × L2 p [a, b] ⊃ Rno × DA → Rno × L2 p [a, b].
n
We define X := Rno × L2 p [a, b] with inner product
*    + Theorem 3. For any functions Z(s) ∈ Rd1 ×n , Z(s, η) ∈
x1 z1
 Rd2 ×n ∀s, η ∈ [a, b], suppose there exists a matrix T < 0
, := x⊤
1 z1 + x2 , z2 Lnp [a,b] , (11) such that
x2 (s) z2 (s) 2
X
and norm P = T11 ,
   2 Z b Z s

 x1  Q(s) = T12 Z(s) + T13 Z(η, s)dη + T14 Z(η, s)dη,
 x (s)  =|| x1 ||2Rno + || x2 (s) ||2Lnp [a,b] . (12) s a
 2  2
X S(s) = Z ⊤ (s)T22 Z(s),
2.3 Lyapunov Stability Theorem R1 (s, η) = Z ⊤ (s)T23 Z(s, η) + Z ⊤ (η, s)T42 Z ⊤ (η)
Z b Z s
The necessary and sufficient condition for exponential + Z ⊤ (θ, s)T33 Z(θ, η)dθ + Z ⊤ (θ, s)T43 Z(θ, η)dθ
stability of a linear infinite dimensional system has been
Zs η η
provided in Curtain and Zwart (1995). The same can be ⊤
extended as Theorem 1 that provides a condition to verify + Z (θ, s)T44 Z(θ, η)dθ,
a
the exponential stability of PDE-ODE coupled system.
R2 (s, η) = Z ⊤ (s)T32 Z(s, η) + Z ⊤ (η, s)T24 Z ⊤ (η)
Theorem 1. Suppose Ā : X ⊃ Rno × DA → X generates a Z b Z η
strongly continuous semigroup and there exists α, β, γ > 0
+ Z ⊤ (θ, s)T33 Z(θ, η)dθ + Z ⊤ (θ, s)T34 Z(θ, η)dθ
and P : X → X such that η s
   2 *   +    2 Z s
   
 x  x x  x 
α  ≤ ,P ≤ β   , + Z ⊤ (θ, s)T44 Z(θ, η)dθ, (15)
 z  z z  z  a
X X X
with Ŝ(s) = D(s)S(s),
 
T11 T12 T13 T14
T21 T22 T23 T24  R̂1 (s, η) = B2 (s)Q1 (η) + D(s)R1 (s, η) + C1 (s, η)S(η)
T = . Z η Z s
T31 T32 T33 T34  + C1 (s, θ)R2 (θ, η)dθ + C1 (s, θ)R1 (θ, η)dθ
T41 T42 T43 T44 a η
Z b
Then the operator P{P,Q,Q⊤ ,S,R1 ,R2 } as defined in (14) is
+ C2 (s, θ)R1 (θ, η)dθ,
positive, i.e. P{P,Q,Q⊤ ,S,R1 ,R2 } < 0. s

Proof. Since T < 0, we can define a square root of T as R̂2 (s, η) = B2 (s)Q1 (η) + D(s)R2 (s, η) + C2 (s, η)S(η)
Z s Z η
U = [U1 U2 U3 U4 ].
  + C1 (s, θ)R2 (θ, η)dθ + C2 (s, θ)R2 (θ, η)dθ
T11 T12 T13 T14 a
Z b
s
T T T T 
T =  21 22 23 24  = U ⊤ U + C2 (s, θ)R1 (θ, η)dθ. (17)
T31 T32 T33 T34
η
T41 T42 T43 T44
 ⊤ 
U1 U1 U1⊤ U2 U1⊤ U3 U1⊤ U4 Proof. The proof can be found in the Appendix C.
 ∗⊤ U ⊤ U2 U ⊤ U3 U ⊤ U4  Notation 1. For convenience, we say
= 2 2 2     
 ∗⊤ ∗⊤ U3⊤ U3 U3⊤ U4  P̂ , Q̂1 , Q̂2 , Ŝ, R̂1 , R̂2 = A, B1 , B2 , D, C1 , C2 (18)
∗⊤ ∗⊤ ∗⊤ U4⊤ U4  
For x1 ∈ Rm and x2 ∈ Ln2 [a, b], let us define v(s) := U1 x1 + × P, Q1 , Q2 , S, R1 , R2 , (19)
(Ψx2 )(s), where
Z s if P̂ , Q̂1 , Q̂2 , Ŝ, R̂1 , R̂2 are related to A, B1 , B2 , D, C1 , C2
(Ψx2 )(s) = U2 Z(s) + U3 Z(s, η)x2 (η)dη and P, Q1 , Q2 , S, R1 , R2 according to (17) in Lemma 4.
a
Z b 3.3 Adjoint of The Operators
+ U4 Z(s, η)x2 (η)dη.
s Now, we also discuss the adjoint of an operator of the
Then
D E class P{P,Q1 ,Q2 ,S,R1 ,R2 } and it is typically denoted by
∗
v(s), v(s) P{P,Q 1 ,Q2 ,S,R1 ,R2 }
.
Rm ×Ln
2 [a,b]

= hU1 x1 , U1 x1 iRm + hU1 x1 , (Ψx2 )(s)iRm Lemma 5. For any matrices P ∈ Rm×m and bounded
functions Q1 : [a, b] → Rm×n , Q2 : [a, b] → Rn×m ,
+ h(Ψx2 )(s), U1 x1 iLn2 [a,b] + h(Ψx2 )(s), (Ψx2 )(s)iLn2 [a,b] S : [a, b] → Rn×n , R1 , R2 : [a, b] × [a, b] → Rn×n , the
D   E
following identity holds for any x ∈ Rm , y ∈ Ln2 ([a, b]).
x1 x
= P ⊤ , 1 ≥ 0. D E
x2 {P,Q,Q ,S,R1 ,R2 } x2 Rm ×Ln2 [a,b]
x, P{P,Q1 ,Q2 ,S,R1 ,R2 } y m n
Hence P{P,Q,Q⊤ ,S,R1 ,R2 } < 0. D
R ×L2 [a,b]
E
∗
= P{P,Q1 ,Q2 ,S,R1 ,R2 } x, y n
, (20)
Rm ×L2 [a,b]
3.2 Composition of The Operators ∗
where, P{P,Q1 ,Q2 ,S,R1 ,R2 }
= P{P̂ ,Q̂1 ,Q̂2 ,Ŝ,R̂1 ,R̂2 } with
In this subsection, we show that the composition of two P̂ = P ⊤ , Ŝ(s) = S ⊤ (s),
operators of the class P{P,Q1 ,Q2 ,S,R1 ,R2 } also takes the same
structure and provide an efficient method to calculate it. Q̂1 (s) = Q⊤
2 (s), R̂1 (s, η) = R2⊤ (η, s),
Lemma 4. For any matrices A, P ∈ Rm×m and bounded Q̂2 (s) = Q⊤
1 (s), R̂2 (s, η) = R1⊤ (η, s). (21)
functions B1 , Q1 : [a, b] → Rm×n , B2 , Q2 : [a, b] → Rn×m ,
D, S : [a, b] → Rn×n , Ci , Ri : [a, b] × [a, b] → Rn×n with Proof. The proof can be found in the Appendix D.
i ∈ {1, 2}, the following identity holds Notation 2. For convenience, we say
   ∗
P{A,B1 ,B2 ,D,C1 ,C2 } P{P,Q1 ,Q2 ,S,R1 ,R2 } P̂ , Q̂1 , Q̂2 , Ŝ, R̂1 , R̂2 = P, Q1 , Q2 , S, R1 , R2 , (22)
= P{P̂ ,Q̂1 ,Q̂2 ,Ŝ,R̂1 ,R̂2 } , (16)
if P̂ , Q̂1 , Q̂2 , Ŝ, R̂1 , R̂2 are related to P, Q1 , Q2 , S, R1 , R2
where according to (21) in Lemma 5.
Z b
P̂ = AP + B1 (s)Q2 (s)ds, 4. FUNDAMENTAL IDENTITIES
a
Z b
Q̂1 (s) = AQ1 (s) + B1 (s)S(s) + B1 (η)R1 (η, s)dη In this section, we utilize the class of operators defined by
Z s s P{P,Q1 ,Q2 ,S,R1 ,R2 } to uniquely represent the system (10) in
terms of the variables x and zss replacing z.
+ B1 (η)R2 (η, s)dη,
a Lemma 6. Suppose x ∈ Rno and z ∈ W 2,2 [a, b] and
Z s
Bc col(z(a), z(b), zs (a), zs (b)) = 0,
Q̂2 (s) = B2 (s)P + D(s)Q2 (s) + C1 (s, η)Q2 (η)dη
a with Bc ∈ R2np ×4np has full row rank. Then
Z b    
x x
+ C2 (s, η)Q2 (η)dη, = P{I,0,0,0,G1 ,G2 } , (23)
s
z zss
     *   +
x x x x
= P{I,0,0,0,G3 ,G4 } , (24) V (x, z) = , P{P,Q,Q⊤ ,S,R1 ,R2 } , (36)
zs zss z z
with X

G1 (s, η) = Ba (s, η) + (s − η)I, (25) such that P{P,Q,Q⊤ ,S,R1 ,R2 } is self-adjoint, i.e.,
∗
G2 (s, η) = Ba (s, η), (26) P{P,Q,Q⊤ ,S,R1 ,R2 } = P{P,Q,Q⊤ ,S,R ,R } .
1 2

G3 (s, η) = Bb (s, η) + I, (27)

G4 (s, η) = Bb (s, η), (28) 5.1 Derivative of The Lyapunov Function
Ba (s, η) = − [I (s − a)I] Bf (η), (29)
Theorem 8. Suppose x ∈ Rno and z ∈ W 2,2 [a, b] and
Bb (η) = − [0 I] Bf (η), (30)
  Bc col(z(a), z(b), zs (a), zs (b)) = 0,
0
with Bc ∈ R2np ×4np has full row rank. Moreover, suppose
(b − η)I 
Bf (η) = (Bc Bd )−1 Bc  , (31) there exists a matrix P ∈ Rno ×no , and bounded polyno-
0 
mial functions Q : [a, b] → Rno ×np , S : [a, b] → Rnp ×np ,
I
  R1 , R2 : [a, b] × [a, b] → Rnp ×np , and
I 0  
I (b − a)I  A + BL1 C BL2 C
Bd =  . (32) Ā = . (37)
0 I BL3 C A + BL4 C
0 I Then *   +
Proof. For detailed proof, we refer to (Peet, 2018, p. 2-3). x x
, P{P,Q,Q⊤ ,S,R1 ,R2 } Ā (38)
z z
*   X
Lemma 7. Suppose x ∈ Rno and z ∈ W 2,2 [a, b] and  +
x x
Bc col(z(a), z(b), zs (a), zs (b)) = 0, + Ā , P{P,Q,Q⊤ ,S,R1 ,R2 }
z z
with Bc ∈ R2np ×4np has full row rank. Then *   +
X
    
ẋ A + BL1 C BL2 C x x x
= (33) = , P{K̂,L̂,L̂⊤ ,M̂,N̂1 ,N̂2 }
ż BL3C A + BL4C z zss zss
X
implies     where
ẋ x
= P{U,V1 ,V2 ,Y,W1 ,W2 } , (34) K̂ = K + K̄, M̂ = M + M̄ , L̂ = L1 + L̄1 ,
ż zss
⊤
where L̂ = L2 + L̄2 , N̂1 = N1 + N̄1 , N̂2 = N2 + N̄2 ,
U = Λ1 + Λ2 , Y = Π1 + Π2 + A2 + B1 L4 C3 ,
V1 = Ω1 + Ω2 + BL2 C3 , W1 = Ψ1 + Ψ2 ,

∗
K̄, L̄1 , L̄2 , M̄ , N̄1 , N̄2 = K, L1 , L2 , M, N1 , N2 ,
V2 = Σ1 + Σ2 , W2 = Υ1 + Υ2 ,

∗
K, L1 , L2 , M, N1 , N2 = I, 0, 0, 0, G1 , G2
with

Λ1 , Ω1 , Σ1 , Π1 , Ψ1 , Υ1

× P, Q, Q⊤ , S, R1 , R2



= A + BL1 C, BL2 Ca , B1 L3 C, A0 + B1 L4 Ca , 0, 0 × U1 , V1 , V2 , Y, W1 , W2 . (39)
× I, 0, 0, 0, G1 , G2 ,

Moreover, G1 , G2 are as defined in Lemma 6 and U , V1 ,

V2 , Y , W1 , W2 are as defined in Lemma 7.
Λ2 , Ω2 , Σ2 , Π2 , Ψ2 , Υ2


Proof. The proof of Theorem 8 can be easily shown by
= 0, BL2 Cb , 0, A1 + B1 L4 Cb , 0, 0 × I, 0, 0, 0, G3 , G4 ,
  substituting Lemma 6-7 in (38).
0
(b − s)I  5.2 Conditions of Exponential Stability
C3 (s) = −C1 Bd Bc (s) + C1  . (35)
0 
I In order to verify the exponential stability of the PDE-
Moreover, the definitions of G1 , G2 , G3 , G4 are given ac- ODE coupled system, we search for the operator
cording to Lemma 6.
P{P,Q,Q⊤ ,S,R1 ,R2 } ≻ 0, subject to P{K̂,L̂,L̂⊤ ,M̂,N̂1 ,N̂2 } ≺ 0.
Proof. For detailed proof, see Appendix A. Definition 9. The cone of positive operators P that are
given in the Definition 2 with polynomial multipliers and
The above identities imply that the original state z can be
kernels associated with degree d1 and d2 are specified by
fully reconstructed by the ’fundamental state’ zss (see Peet
(2018)). Moreover, the new representation of the system Φd1 ,d2 := {(P, Q, Q⊤ , S, R1 , R2 ) :
dynamics in terms of zss is independent of the boundary the positivity condition is satisfied
constraints. In the subsequent sections, these identities are according to Theorem 3}. (40)
utilized for deriving stability conditions. Theorem 10. Suppose there exists ǫ > 0, d1 , d2 ∈ Z with
5. LYAPUNOV STABILITY CONDITION P ∈ Rno , and bounded polynomial functions Q : [a, b] →
Rno ×np , S : [a, b] → Rnp ×np , R1 , R2 : [a, b] × [a, b] →
In order to verify the stability of the PDE-ODE coupled Rnp ×np such that


system, we have proposed a Lyapunov function of the form P − ǫI, Q, Q⊤, S − ǫI, R1 , R2 ∈ Φd1 ,d2 ,



− K̂, L̂, L̂⊤ , M, N1 , N2 − ǫ I, 0, 0, 0, T1, T2 ∈ Φd1 ,d2 , In Theorem 3, we can choose Z(s) to be a vector of mono-
(41) mials of degree d1 and Z(s, η) to be a vector of monomials
of degree d2 . Hence, the constraint (P, Q, Q⊤ , S, R1 , R2 ) ∈
with
Φd1 ,d2 can be viewed as an LMI constraint. As a result,
K̂ = K + K̄, M̂ = M + M̄ , L̂ = L1 + L̄1 , the verification of stability for PDE-ODE coupled system
L̂⊤ = L2 + L̄2 , N̂1 = N1 + N̄1 , N̂2 = N2 + N̄2 , amounts to a feasibility test of satisfying the LMI con-
straints related to (41) in Theorem 10.


∗
K̄, L̄1 , L̄2 , M̄ , N̄1 , N̄2 = K, L1 , L2 , M, N1 , N2 , 6. NUMERICAL IMPLEMENTATIONS


∗
K, L1 , L2 , M, N1 , N2 = I, 0, 0, 0, G1 , G2 In this section, we illustrate the developed methodology


× P, Q, Q⊤ , S, R1 , R2 for verifying the stability of PDE-ODE coupled system.


× U1 , V1 , V2 , Y, W1 , W2 , The method has been implemented in MATLAB R
using


∗ an adapted version of SOSTOOLS (see Papachristodoulou
I, 0, 0, 0, T1, T2 = I, 0, 0, 0, G1 , G2 et al. (2013)).


× I, 0, 0, 0, G1, G2 , (42)
where G1 , G2 are as defined in Lemma 6 and U , V1 , V2 , 6.1 ODE Coupled with Diffusion Equation
Y , W1 , W2 are as defined in Lemma 7. Then the system
in (10) is exponentially stable. First, we study the boundary controlled thermo-mechanical
process where a lumped mechanical system is driven by
Proof. We have defined the Lyapunov function as converting thermal energy to mechanical work. In Tang
*   + and Xie (2010), such a system is modeled as a finite
x x dimensional ODE model with the actuator dynamics that
V (x, z) = , P{P,Q,Q⊤ ,S,R1 ,R2 }
z z are governed by the diffusion equation. A closed loop rep-
X
*   + resentation of such controlled PDE-ODE coupled system
x x is given in Tang and Xie (2010) as
= , P{P −ǫI,Q,Q⊤ ,S−ǫI,R1 ,R2 } (43)
z z
X Ẋ(t) = −3X(t) + w(0, t), (44)
   2    2
    wt (x, t) = wxx (xt), (45)
 x   x 
+ ǫ  ≥ ǫ z  . wx (0, t) = 0, (46)
 z   
X X w(1, t) = 0. (47)
This shows the strict positivity of P{P,Q,Q⊤ ,S,R1 ,R2 } . Fur- Based on the definitions in Section 3, we obtain A =
thermore, based on Theorem 8, the exponential stability −3, B = 1, C = [] for the ODE system. For the PDE
in Theorem
*  1 can be reformulated as model, A0 (s) = A1 (s) = [], A2 (s) = 1. The inputs and
 +
x x output operators are B1 (s) = [], C1 = [1 0 0 0] with
, P{K̂,L̂,L̂⊤ ,N̂1 ,N̂2 }
zss zss Ca (s)
  = 0, Cb (s) = []. The interconnection matrix L =
X
*   + 01
x x .
= , P{K̂,L̂,L̂⊤ ,N̂1 ,N̂2 } 10
zss zss
   2    2
X The developed methodology proves the exponential stabil-
    ity of this system which has been verified analytically in
 x   x 
+ ǫ  − ǫ z  Tang and Xie (2010).
 z   
*  X X
 + 6.2 ODE Coupled with Diffusion-Reaction PDE
x x
= , P{K̂,L̂,L̂⊤ ,N̂1 ,N̂2 }
zss zss The next example is taken from chemical processes where
* X
   + due to a clear distinction between microscopic processes
x x and macroscopic processes, models often involve PDEs
+ǫ P{I,0,0,0,G1 ,G2 } , P{I,0,0,0,G1,G2 }
zss zss governing macroscopic dynamics coupled with ODEs gov-
X
   2
  erning microscopic dynamics. For exothermic or endother-
 x  mic reaction processes, we consider a general class of forced
− ǫ  ,
 z  diffusion-reaction in the spatial domain [0, 1] which is
X
*   + represented by the following coupled PDEs.
x x ż(s, t) = λz(s, t) + zss (s, t) + B1 (s)u(t). (48)
= , P{K̂,L̂1 ,L̂2 ,M̂,N̂1 ,N̂2 }
zss zss
X In Valmorbida et al. (2014), it has been shown that if the
*   +    2 boundary conditions are chosen to be z(0, t) = z(1, t) = 0,
 
x x  x  then the unforced system given in (48)(i.e. B1 (s) = 0)
+ǫ , P{I,0,0,0,T1 ,T2 } − ǫ 
zss zss  z  is stable if and only if λ < π 2 = 9.8696. Similarly, in
X X
   2
  Valmorbida et al. (2016) it has been shown that for the
 x  boundary conditions z(0, t) = zs (1, t) = 0, the system is
≤ −ǫ  .
 z  stable if and only if λ ≤ 2.467.
X
This proves the exponential stability according to Theo- Such diffusion-reaction processes in (48) are coupled with
rem 1. a finite dimensional system that typically represents a
molecular transport in the microscopic scale. Specifically, Datko, R. (1970). Extending a theorem of A.M. Liapunov
we consider that the effect of these molecular transport to Hilbert space. Journal of Mathematical analysis and
applies a uniform in-domain input to the PDEs over the applications.
entire spatial domain (i.e. B1 (s) = 1, ∀s ∈ [a, b]). On the Fridman, E. and Orlov, Y. (2009). Exponential stability of
other hand, the output flux of the chemical reaction at linear distributed parameter systems with time-varying
s = 0(i.e. zs (0, t) = 0) drives the molecular transport delays. Automatica.
and is considered to be the input  to the ODE model. Gahlawat, A. and Peet, M.M. (2017). A convex sum-of-
0 I squares approach to analysis, state feedback and out-
The interconnection matrix L = . The ODE model
I 0 put feedback control of parabolic pdes. IEEE Trans-
is given in (49). The stability of (49) can be verified by actions on Automatic Control, 62(4), 1636–1651. doi:
checking the eigenvalues. 10.1109/TAC.2016.2593638.
  Krstic, M. (2009). Compensating a string pde in the
−1.2142 1.9649 0.2232 0.5616
actuation or sensing path of an unstable ode. In
−1.8042 −0.7260 −0.3479 5.4355 
ẋ(t) =  x(t) American Control Conference, 2009. ACC’09. IEEE.
−0.2898 0.7381 −1.7606 0.8294 
Krstic, M. and Smyshlyaev, A. (2008). Backstepping
−0.9417 −5.3399 −1.0704 −0.7590
  boundary control for first-order hyperbolic pdes and
−1.5368 0 application to systems with actuator and sensor delays.
0 0.8871 (49)
+

 v(t), Systems & Control Letters.
1.0656 0 Meyer, E. and Peet, M.M. (2015). Stability analysis of
1.1882 0 parabolic linear pdes with two spatial dimensions using
 
−2.5575 0 1.0368 0 lyapunov method and sos. In Decision and Control
w(t) = x(t). (CDC), 2015 IEEE 54th Annual Conference on. IEEE.
−1.8067 0.4630 1.3621 0
Nicaise, S., Valein, J., and Fridman, E. (2009). Stability of
The developed methodology is able to verify the stability of the heat and of the wave equations with boundary time-
the coupled system for a) λ < π 2 = 9.8696 with boundary varying delays. Discrete and Continuous Dynamical
conditions z(0, t) = z(1, t) = 0, and b) λ ≤ 2.467 with Systems- Series S 2(3).
boundary conditions z(0, t) = zs (1, t) = 0. Papachristodoulou, A., Anderson, J., Valmorbida, G., Pra-
jna, S., Seiler, P., and Parrilo, P.A. (2013). SOS-
7. CONCLUSIONS TOOLS: Sum of squares optimization toolbox for MAT-
LAB. http://arxiv.org/abs/1310.4716.
In this paper, we have presented a novel methodology to Papachristodoulou, A. and Peet, M.M. (2006). On the
verify stability for a large class of mutually coupled linear analysis of systems described by classes of partial dif-
PDE-ODE systems in one spatial dimension. In particular, ferential equations. In Proceedings of the 45th IEEE
by using positive-definite matrices to parametrize a cone Conference on Decision and Control, 747–752. doi:
of Lyapunov functions, we have shown that the stability 10.1109/CDC.2006.377815.
of linear PDE-ODE coupled systems amounts to satisfy- Peet, M.M. (2018). A new state-space representation
ing a specific Linear Matrix Inequality (LMI) that can of lyapunov stability for coupled pdes and scalable
be efficiently solved using convex optimization. We have stability analysis in the sos framework. arXiv preprint
provided a scalable variation of Sum of Squares (SOS) arXiv:1803.07290.
algorithm that can be utilized to verify the stability a large Quarteroni, A. and Veneziani, A. (2003). Analysis of
class of PDE-ODE coupled system efficiently. a geometrical multiscale model based on the coupling
of ode and pde for blood flow simulations. Multiscale
In future, the results of stability analysis for PDE-ODE Modeling & Simulation.
coupled system can be extended to study the robustness Ravindran, S.S. (2000). A reduced-order approach for
properties and H∞ analysis of PDE-ODE coupled system. optimal control of fluids using proper orthogonal decom-
Extending the current framework for PDEs in more than position. International journal for numerical methods in
one spatial dimension and nonlinear PDEs are a potential fluids.
direction of future research. Furthermore, viewing the Shivakumar, S. and Peet, M.M. (2018). Computing Input-
coupling of ODEs and PDEs as a feedback interconnection, Output Properties of Coupled PDE systems. arXiv e-
the problem of synthesizing finite dimensional controller prints, arXiv:1812.05081.
and observer are of great practical importance. Stinner, C., Surulescu, C., and Winkler, M. (2014). Global
weak solutions in a pde-ode system modeling multiscale
REFERENCES cancer cell invasion. SIAM Journal on Mathematical
Analysis.
Christofides, P.D. (2012). Nonlinear and robust control Susto, G.A. and Krstic, M. (2010). Control of pde–ode
of PDE systems: Methods and applications to transport- cascades with neumann interconnections. Journal of the
reaction processes. Springer Science & Business Media. Franklin Institute.
Curtain, R.F. and Zwart, H.J. (1995). An Introduction to Tang, S. and Xie, C. (2010). Stabilization of a coupled
Infinite-dimensional Linear Systems Theory. Springer- pde-ode system by boundary control. In 49th IEEE
Verlag New York. Conference on Decision and Control (CDC), 4042–4047.
Das, A., Weiland, S., and Iapichino, L. (2018). Model doi:10.1109/CDC.2010.5718141.
approximation of thermo-fluidic diffusion processes in Valmorbida, G., Ahmadi, M., and Papachristodoulou, A.
spatially interconnected structures. In 2018 European (2014). Semi-definite programming and functional in-
Control Conference (ECC). IEEE.
 equalities for distributed parameter systems. In 53rd Rb
IEEE Conference on Decision and Control, 4304–4309. [Cb zs ](s) := Cb (s)zs (s)ds,
a
doi:10.1109/CDC.2014.7040060.
Valmorbida, G., Ahmadi, M., and Papachristodoulou, A. Rb
(2016). Stability analysis for a class of partial differ- and [C3 zss ](s) := C3 (s)zss (s)ds.
a
ential equations via semidefinite programming. IEEE
Transactions on Automatic Control. Rewriting in the form of P{P,Q1 ,Q2 ,S,R1 ,R2 }
   
ẋ x
Appendix A. PROOF OF LEMMA 7 =P{A+BL1 C,BL2 Ca ,B1 L3 C,A0 +B1 L4 Ca ,0,0}
ż z
 
Lemma 7. Suppose x ∈ Rno and z ∈ W 2,2 [a, b] and + P{0,BL2 Cb ,0,A1 +B1 L4 Cb ,0,0}
x
Bc col(z(a), z(b), zs (a), zs (b)) = 0, zs
 
with Bc ∈ R2np ×4np has full row rank. Then x
+ P{0,BL2 C3 ,0,A2 +B1 L4 C3 ,0,0} . (A.7)
     zss
ẋ A + BL1 C BL2 C x
= (A.1) Now, substituting Lemma 6 in (A.8) we obtain
ż BL3C A + BL4C z
   
implies ẋ x
    =P{Λ1 ,Ω1 ,Σ1 ,Π1 ,Ψ1 ,Υ1 }
ẋ x ż zss
= P{U,V1 ,V2 ,Y,W1 ,W2 } , (A.2)  
ż zss x
where + P{Λ2 ,Ω2 ,Σ2 ,Π2 ,Ψ2 ,Υ2 }
zss
U = Λ1 + Λ2 , Y = Π1 + Π2 + A2 + B1 L4 C3 ,  
x
V1 = Ω1 + Ω2 + BL2 C3 , W1 = Ψ1 + Ψ2 , + P{0,BL2 C3 ,0,A2 +B1 L4 C3 ,0,0} , (A.8)
zss
V2 = Σ1 + Σ2 , W2 = Υ1 + Υ2 , where P{Λ1 ,Ω1 ,Σ1 ,Π1 ,Ψ1 ,Υ1 } is defined as
with

P{A+BL1 C,BL2 Ca ,B1 L3 C,A0 +B1 L4 Ca ,0,0} P{I,0,0,0,G1,G2 } .
Λ1 , Ω1 , Σ1 , Π1 , Ψ1 , Υ1

Similarly, P{Λ2 ,Ω2 ,Σ2 ,Π2 ,Ψ2 ,Υ2 } is defined as
= A + BL1 C, BL2 Ca , B1 L3 C, A0 + B1 L4 Ca , 0, 0

P{0,BL2 Cb ,0,A1 +B1 L4 Cb ,0,0} P{I,0,0,0,G3 ,G4 } .
× I, 0, 0, 0, G1 , G2 ,

As a result,
Λ2 , Ω2 , Σ2 , Π2 , Ψ2 , Υ2 P{Λ1 ,Ω1 ,Σ1 ,Π1 ,Ψ1 ,Υ1 } + P{Λ2 ,Ω2 ,Σ2 ,Π2 ,Ψ2 ,Υ2 }



= 0, BL2 Cb , 0, A1 + B1 L4 Cb , 0, 0 × I, 0, 0, 0, G3 , G4 , + P{0,BL2 C3 ,0,A2 +B1 L4 C3 ,0,0} = P{U,V1 ,V2 ,Y,W1 ,W2 } .
 
0
(b − s)I  Appendix B. PROOF OF THEOREM 8
C3 (s) = −C1 Bd Bc (s) + C1  . (A.3)
0 
I Theorem 8. Suppose x ∈ Rno and z ∈ W 2,2 [a, b] and
Moreover, the definitions of G1 , G2 , G3 , G4 are given ac- Bc col(z(a), z(b), zs (a), zs (b)) = 0,
cording to Lemma 6. with Bc ∈ R2np ×4np has full row rank. Moreover, suppose
Proof. First of all, using (8) we can show that there exists a matrix P ∈ Rno ×no , and bounded polyno-
mial functions Q : [a, b] → Rno ×np , S : [a, b] → Rnp ×np ,
Zb R1 , R2 : [a, b] × [a, b] → Rnp ×np , and
C1 zb = C3 (s)zss (s)ds, (A.4) 
A + BL1 C BL2 C

Ā = . (B.1)
a BL3 C A + BL4 C
where
  Then
0 *   +
(b − s)I  x x
C3 (s) = −C1 Bd Bc (s) + C1  . (A.5) , P{P,Q,Q⊤ ,S,R1 ,R2 } Ā (B.2)
0  z z
X
I *    +
Now we use the definitions A, B, C in (7) and re-write (33) x x
+ Ā , P{P,Q,Q⊤ ,S,R1 ,R2 }
as z z
     *  X
ẋ A + BL1 C BL2 Ca x  +
= x x
ż BL3C A0 + BL4 Ca z = , P{K̂,L̂,L̂⊤ ,M̂,N̂1 ,N̂2 }
      zss zss
0 BL2 Cb x 0 BL2 C3 x X
+ + . where
0 A1 + BL4 Cb zs 0 A2 + BL4 C3 zss
(A.6) K̂ = K + K̄, M̂ = M + M̄ , L̂ = L1 + L̄1 ,
⊤
Based on (7), L̂ = L2 + L̄2 , N̂1 = N1 + N̄1 , N̂2 = N2 + N̄2 ,
[A0 z](s) := A0 (s)z(s), [A1 zs ](s) := A1 (s)zs (s),


∗
Rb K̄, L̄1 , L̄2 , M̄ , N̄1 , N̄2 = K, L1 , L2 , M, N1 , N2 ,
[A2 zss ](s) := A2 (s)zss (s), [Ca z](s) := Ca (s)z(s)ds,

∗
a K, L1 , L2 , M, N1 , N2 = I, 0, 0, 0, G1, G2


× P, Q, Q⊤ , S, R1 , R2 P{K̂,L̂,L̂⊤ ,M̂,N̂1 ,N̂2 } =P{K,L1 ,L2 ,M,N1 ,N2 }

∗
× U1 , V1 , V2 , Y, W1 , W2 . + P{K,L 1 ,L2 ,M,N1 ,N2 }
,
(B.3) This completes the proof.
Moreover, G1 , G2 are as defined in Lemma 6 and U , V1 ,
V2 , Y , W1 , W2 are as defined in Lemma 7. Appendix C. COMPOSITION OF THE OPERATORS

Proof. The time derivative of the Lyapunov function is Lemma 4. For any matrices A, P ∈ Rm×m and bounded
*   +
x x functions B1 , Q1 : [a, b] → Rm×n , B2 , Q2 : [a, b] → Rn×m ,
z
, P{P,Q,Q⊤ ,S,R1 ,R2 } Ā
z D, S : [a, b] → Rn×n , Ci , Ri : [a, b] × [a, b] → Rn×n with
*   X i ∈ {1, 2}, the following identity holds
 +
x x P{A,B1 ,B2 ,D,C1 ,C2 } P{P,Q1 ,Q2 ,S,R1 ,R2 }
+ Ā , P{P,Q,Q⊤ ,S,R1 ,R2 }
z z = P{P̂ ,Q̂1 ,Q̂2 ,Ŝ,R̂1 ,R̂2 }
* X
   + where
x x Z b
= P{U1 ,V1 ,V2 ,Y,W1 ,W2 } , P{P,Q,Q⊤ ,S,R1 ,R2 }
zss z P̂ = AP + B1 (s)Q2 (s)ds,
X a
| {z } Z b
:=Γ1
*   + Q̂1 (s) = AQ1 (s) + B1 (s)S(s) + B1 (η)R1 (η, s)dη
x x s
+ , P{P,Q,Q⊤ ,S,R1 ,R2 } P{U1 ,V1 ,V2 ,Y,W1 ,W2 } . Z s
z zss
| {z X
} + B1 (η)R2 (η, s)dη,
a
:=Γ2 Z s
(B.4) Q̂2 (s) = B2 (s)P + D(s)Q2 (s) + C1 (s, η)Q2 (η)dη
Now, taking two individual terms separately and substi- a
Z b
tuting the identity (23) we obtain + C2 (s, η)Q2 (η)dη,
* 
s
x ∗
Γ2 = , P{I,0,0,0,G 1 ,G2 } Ŝ(s) = D(s)S(s),
zss
 + R̂1 (s, η) = B2 (s)Q1 (η) + D(s)R1 (s, η) + C1 (s, η)S(η)
x Z η Z s
P{P,Q,Q⊤ ,S,R1 ,R2 } P{U1 ,V1 ,V2 ,Y,W1 ,W2 } . (B.5)
zss + C1 (s, θ)R2 (θ, η)dθ + C1 (s, θ)R1 (θ, η)dθ
*  X a η
 + Z
x x b
= , P{K,L1 ,L2 ,M,N1 ,N2 } . + C2 (s, θ)R1 (θ, η)dθ,
zss zss
X s
Similarly R̂2 (s, η) = B2 (s)Q1 (η) + D(s)R2 (s, η) + C2 (s, η)S(η)
*   Z s Z η
x + C1 (s, θ)R2 (θ, η)dθ + C2 (s, θ)R2 (θ, η)dθ
Γ1 = P{U1 ,V1 ,V2 ,Y,W1 ,W2 } ,
zss a s
Z b
 +
x + C2 (s, θ)R1 (θ, η)dθ.
P{P,Q,Q⊤ ,S,R1 ,R2 } P{I,0,0,0,G1 ,G2 } . η
zss
X
*  Proof. Suppose
x ∗
  
= , P{U (B.6) x
zss 1 ,V1 ,V2 ,Y,W1 ,W2 }
P{A,B1 ,B2 ,D,C1 ,C2 } P{P,Q1 ,Q2 ,S,R1 ,R2 } 1 (s)
x2
+
   
x x
P{P,Q,Q⊤ ,S,R1 ,R2 } P{I,0,0,0,G1 ,G2 }
zss
. = P{P̂ ,Q̂1 ,Q̂2 ,Ŝ,R̂1 ,R̂2 } 1 (s)
x2
X
*   + Let
x ∗ x   
x

y1

= , P{K,L 1 ,L2 ,M,N1 ,N2 } z P{P,Q1 ,Q2 ,S,R1 ,R2 } 1 (s) =
zss ss x2 y2 (s)
X
Now, using the adjoint operation of the operators, we Rb
where y1 = P x1 + a Q1 (s)x2 (s)ds and y2 (s) = Q2 (s)x1 +
obtain* Rs Rb
   + S(s)x2 (s) + a R1 (s, η)x2 (η)dη + s R2 (s, η)x2 (η)dη.
x x
, P{K,L1 ,L2 ,M,N1 ,N2 } Then,
zss zss
X  
*   + x1
x x P{A,B1 ,B2 ,D,C1 ,C2 } P{P,Q1 ,Q2 ,S,R1 ,R2 }
+ ∗
, P{K,L (B.7) x2
zss 1 ,L2 ,M,N1 ,N2 } z     
ss
X y1 z1
*   + = P{A,B1 ,B2 ,D,C1 ,C2 } (s) =
x x y2 z2 (s)
= , P{K̂,L̂,L̂⊤ ,M̂ ,N̂1 ,N̂2 } . Rb
zss zss where z1 = Ay1 + a B1 (s)y2 (s)ds and z2 (s) = B2 (s)y1 +
X Rs Rb
with D(s)y2 (s) + a C1 (s, η)y2 (η)dη + s C2 (s, η)y2 (η)dη.
 D E
Finding the composition is a straight-forward algebraic ∗
= P{P,Q x, y , (D.1)
1 ,Q2 ,S,R1 ,R2 }
operation. We do this by expanding each term separately. Rm ×Ln
2 [a,b]
Firstly, ∗
where, P{P,Q = P{P̂ ,Q̂1 ,Q̂2 ,Ŝ,R̂1 ,R̂2 } with
Z b Z b 1 ,Q2 ,S,R1 ,R2 }

B1 (s)y2 (s)ds = B1 (s) Q2 (s)x1 + S(s)x2 (s) P̂ = P ⊤ , Ŝ(s) = S ⊤ (s),
a a
Z s Z b  Q̂1 (s) = Q⊤
2 (s), R̂1 (s, η) = R2⊤ (η, s),
+ R1 (s, η)x2 (η)dη + R2 (s, η)x2 (η)dη ds. Q̂2 (s) = Q⊤ R̂2 (s, η) = R1⊤ (η, s).
a s 1 (s), (D.2)
Seperating the terms involving x1 and x2 (s), we obtain Proof. We use  the fact
Z b ! Z b  that for any scalar
 a we have
⊤ x1 y1
z1 = AP + B1 (s)Q2 (s) x1 + Q̂1 (s)x2 (s)ds a = a . Let x = and y = .
x2 (s) y2 (s)
a a
| {z } Then
P̂ D E
x, P{P,Q1 ,Q2 ,S,R1 ,R2 } y
where Rm ×Ln
2
[a,b]
Z b Z b Z b
Q̂1 (s) = AQ1 (s) + B1 (s)S(s) + B1 (η)R1 (η, s)dη = x⊤1 P y1 + x⊤1 Q1 (s)y2 (s)ds + y2⊤ (s)Q2 (s)x1 ds
Z s s a a
Z b Z bZ s
+ B1 (η)R2 (η, s)dη + x2 (s)⊤ S(s)y2 (s)ds + x⊤2 (s)R1 (s, η)y2 (η)dηds
a a a a
Z bZ b
Next, we expand the terms of z2 (s) and specifically collect + x⊤2 (s)R2 (s, η)y2 (η)dηds
the terms having x1 . We obtain a s
Z b Z b
 Z s 
= y1⊤ P ⊤ x1 + x⊤ ⊤
1 Q2 (s)y2 (s)ds + y2 (s)Q⊤ 1 (s)x1 ds
B2 (s)P + D(s)Q2 (s) + C1 (s, η)Q2 (η)dη x1 a a
a Z b Z bZ s
= Q̂2 (s)x1 + y2⊤ (s)S ⊤ (s)x2 (s)ds + y2⊤ (s)R2⊤ (η, s)x2 (η)dηds
a a a
Next, grouping the terms having x2 (s) we obtain Z bZ b
+ y2⊤ (s)R1⊤ (η, s)x2 (η)dηds
D(s)S(s)x2 (s) = Ŝ(s)x2 (s). a s
Z b Z b
⊤
Similarly, we can group the terms involving x2 (θ) as = y1 P̂ x1 + x⊤1 Q̂1 (s)y2 (s)ds + y2⊤ (s)Q̂2 (s)x1 ds
Z s a a
Z b Z bZ s
B2 (s)Q1 (η) + D(s)R1 (s, η) + C1 (s, η)S(η) + y2⊤ (s)Ŝ(s)x2 (s)ds + y2⊤ (s)R̂1 (s, η)x2 (η)dηds
a a a a
Z s Z η
Z bZ b
+ C1 (s, θ)R1 (θ, η)dθ + C1 (s, θ)R2 (θ, η)dθ + y2⊤ (s)R̂2 (s, η)x2 (η)dηds
a s
η a D E
!
Z b = P{P̂ ,Q̂1 ,Q̂2 ,Ŝ,R̂1 ,R̂2 } x, y
Rm ×Ln [a,b]
+ C2 (s, θ)R1 (θ, η)dθ x2 (η)dη D E 2
∗
s = P{P,Q1 ,Q2 ,S,R1 ,R2 } x, y n
.
Z s Rm ×L2 [a,b]

= R̂1 (s, η)x2 (η)dη. where,

a
and P̂ = P ⊤ , Ŝ(s) = S ⊤ (s),
Q̂1 (s) = Q⊤ R̂1 (s, η) = R2⊤ (η, s),
Z b
2 (s),
B2 (s)Q1 (η) + D(s)R2 (s, η) + C2 (s, η)S(η)
s Q̂2 (s) = Q⊤
1 (s), R̂2 (s, η) = R1⊤ (η, s).
Z η Z s This completes the proof.
+ C2 (s, θ)R2 (θ, η)dθ + C1 (s, θ)R2 (θ, η)dθ
s ! a
Z b
+ C2 (s, θ)R1 (θ, η)dθ x2 (η)dη
θ
Z b
= Rˆ2 (s, η)x2 (η)dη.
s
This completes the proof.

Appendix D. ADJOINT OF THE OPERATORS

Lemma 5. For any matrices P ∈ Rm×m and bounded

functions Q1 : [a, b] → Rm×n , Q2 : [a, b] → Rn×m ,
S : [a, b] → Rn×n , R1 , R2 : [a, b] × [a, b] → Rn×n , the
following identity holds for any x ∈ Rm , y ∈ Ln2 ([a, b]).
D E
x, P{P,Q1 ,Q2 ,S,R1 ,R2 } y m n
R ×L2 [a,b]